1 /** @file mdb.c
2 * @brief Lightning memory-mapped database library
3 *
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
6 */
7 /*
8 * Copyright 2011-2016 Howard Chu, Symas Corp.
9 * All rights reserved.
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
13 * Public License.
14 *
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
18 *
19 * This code is derived from btree.c written by Martin Hedenfalk.
20 *
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
22 *
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
26 *
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
34 */
35 #ifndef _GNU_SOURCE
36 #define _GNU_SOURCE 1
37 #endif
38 #if defined(MDB_VL32) || defined(__WIN64__)
39 #define _FILE_OFFSET_BITS 64
40 #endif
41 #ifdef _WIN32
42 #include <malloc.h>
43 #include <windows.h>
44
45 /* We use native NT APIs to setup the memory map, so that we can
46 * let the DB file grow incrementally instead of always preallocating
47 * the full size. These APIs are defined in <wdm.h> and <ntifs.h>
48 * but those headers are meant for driver-level development and
49 * conflict with the regular user-level headers, so we explicitly
50 * declare them here. Using these APIs also means we must link to
51 * ntdll.dll, which is not linked by default in user code.
52 */
53 NTSTATUS WINAPI
54 NtCreateSection(OUT PHANDLE sh, IN ACCESS_MASK acc,
55 IN void * oa OPTIONAL,
56 IN PLARGE_INTEGER ms OPTIONAL,
57 IN ULONG pp, IN ULONG aa, IN HANDLE fh OPTIONAL);
58
59 typedef enum _SECTION_INHERIT {
60 ViewShare = 1,
61 ViewUnmap = 2
62 } SECTION_INHERIT;
63
64 NTSTATUS WINAPI
65 NtMapViewOfSection(IN PHANDLE sh, IN HANDLE ph,
66 IN OUT PVOID *addr, IN ULONG_PTR zbits,
67 IN SIZE_T cs, IN OUT PLARGE_INTEGER off OPTIONAL,
68 IN OUT PSIZE_T vs, IN SECTION_INHERIT ih,
69 IN ULONG at, IN ULONG pp);
70
71 NTSTATUS WINAPI
72 NtClose(HANDLE h);
73
74 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
75 * as int64 which is wrong. MSVC doesn't define it at all, so just
76 * don't use it.
77 */
78 #define MDB_PID_T int
79 #define MDB_THR_T DWORD
80 #include <sys/types.h>
81 #include <sys/stat.h>
82 #ifdef __GNUC__
83 # include <sys/param.h>
84 #else
85 # define LITTLE_ENDIAN 1234
86 # define BIG_ENDIAN 4321
87 # define BYTE_ORDER LITTLE_ENDIAN
88 # ifndef SSIZE_MAX
89 # define SSIZE_MAX INT_MAX
90 # endif
91 #endif
92 #else
93 #include <sys/types.h>
94 #include <sys/stat.h>
95 #define MDB_PID_T pid_t
96 #define MDB_THR_T pthread_t
97 #include <sys/param.h>
98 #include <sys/uio.h>
99 #include <sys/mman.h>
100 #ifdef HAVE_SYS_FILE_H
101 #include <sys/file.h>
102 #endif
103 #include <fcntl.h>
104 #endif
105
106 #if defined(__mips) && defined(__linux)
107 /* MIPS has cache coherency issues, requires explicit cache control */
108 #include <asm/cachectl.h>
109 extern int cacheflush(char *addr, int nbytes, int cache);
110 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
111 #else
112 #define CACHEFLUSH(addr, bytes, cache)
113 #endif
114
115 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
116 /** fdatasync is broken on ext3/ext4fs on older kernels, see
117 * description in #mdb_env_open2 comments. You can safely
118 * define MDB_FDATASYNC_WORKS if this code will only be run
119 * on kernels 3.6 and newer.
120 */
121 #define BROKEN_FDATASYNC
122 #endif
123
124 #include <errno.h>
125 #include <limits.h>
126 #include <stddef.h>
127 #include <inttypes.h>
128 #include <stdio.h>
129 #include <stdlib.h>
130 #include <string.h>
131 #include <time.h>
132
133 #ifdef _MSC_VER
134 #include <io.h>
135 typedef SSIZE_T ssize_t;
136 #else
137 #include <unistd.h>
138 #endif
139
140 #if defined(__sun) || defined(ANDROID)
141 /* Most platforms have posix_memalign, older may only have memalign */
142 #define HAVE_MEMALIGN 1
143 #include <malloc.h>
144 #endif
145
146 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
147 #include <netinet/in.h>
148 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
149 #endif
150
151 #if defined(__APPLE__) || defined (BSD)
152 # if !(defined(MDB_USE_POSIX_MUTEX) || defined(MDB_USE_POSIX_SEM))
153 # define MDB_USE_SYSV_SEM 1
154 # endif
155 # define MDB_FDATASYNC fsync
156 #elif defined(ANDROID)
157 # define MDB_FDATASYNC fsync
158 #endif
159
160 #ifndef _WIN32
161 #include <pthread.h>
162 #ifdef MDB_USE_POSIX_SEM
163 # define MDB_USE_HASH 1
164 #include <semaphore.h>
165 #elif defined(MDB_USE_SYSV_SEM)
166 #include <sys/ipc.h>
167 #include <sys/sem.h>
168 #ifdef _SEM_SEMUN_UNDEFINED
169 union semun {
170 int val;
171 struct semid_ds *buf;
172 unsigned short *array;
173 };
174 #endif /* _SEM_SEMUN_UNDEFINED */
175 #else
176 #define MDB_USE_POSIX_MUTEX 1
177 #endif /* MDB_USE_POSIX_SEM */
178 #endif /* !_WIN32 */
179
180 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) + defined(MDB_USE_SYSV_SEM) \
181 + defined(MDB_USE_POSIX_MUTEX) != 1
182 # error "Ambiguous shared-lock implementation"
183 #endif
184
185 #if defined(__SUNPRO_C)
186 #include "mbarrier.h"
187 #endif
188
189 #ifdef USE_VALGRIND
190 #include <valgrind/memcheck.h>
191 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
192 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
193 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
194 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
195 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
196 #else
197 #define VGMEMP_CREATE(h,r,z)
198 #define VGMEMP_ALLOC(h,a,s)
199 #define VGMEMP_FREE(h,a)
200 #define VGMEMP_DESTROY(h)
201 #define VGMEMP_DEFINED(a,s)
202 #endif
203
204 #ifndef BYTE_ORDER
205 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
206 /* Solaris just defines one or the other */
207 # define LITTLE_ENDIAN 1234
208 # define BIG_ENDIAN 4321
209 # ifdef _LITTLE_ENDIAN
210 # define BYTE_ORDER LITTLE_ENDIAN
211 # else
212 # define BYTE_ORDER BIG_ENDIAN
213 # endif
214 # else
215 # define BYTE_ORDER __BYTE_ORDER
216 # endif
217 #endif
218
219 #ifndef LITTLE_ENDIAN
220 #define LITTLE_ENDIAN __LITTLE_ENDIAN
221 #endif
222 #ifndef BIG_ENDIAN
223 #define BIG_ENDIAN __BIG_ENDIAN
224 #endif
225
226 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
227 #define MISALIGNED_OK 1
228 #endif
229
230 #include "lmdb.h"
231 #include "midl.h"
232
233 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
234 # error "Unknown or unsupported endianness (BYTE_ORDER)"
235 #elif (-6 & 5) || CHAR_BIT!=8 || UINT_MAX!=0xffffffff || MDB_SIZE_MAX%UINT_MAX
236 # error "Two's complement, reasonably sized integer types, please"
237 #endif
238
239 #ifdef __GNUC__
240 /** Put infrequently used env functions in separate section */
241 # ifdef __APPLE__
242 # define ESECT __attribute__ ((section("__TEXT,text_env")))
243 # else
244 # define ESECT __attribute__ ((section("text_env")))
245 # endif
246 #else
247 #define ESECT
248 #endif
249
250 #ifdef _WIN32
251 #define CALL_CONV WINAPI
252 #else
253 #define CALL_CONV
254 #endif
255
256 /** @defgroup internal LMDB Internals
257 * @{
258 */
259 /** @defgroup compat Compatibility Macros
260 * A bunch of macros to minimize the amount of platform-specific ifdefs
261 * needed throughout the rest of the code. When the features this library
262 * needs are similar enough to POSIX to be hidden in a one-or-two line
263 * replacement, this macro approach is used.
264 * @{
265 */
266
267 /** Features under development */
268 #ifndef MDB_DEVEL
269 #define MDB_DEVEL 0
270 #endif
271
272 /** Wrapper around __func__, which is a C99 feature */
273 #if __STDC_VERSION__ >= 199901L
274 # define mdb_func_ __func__
275 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
276 # define mdb_func_ __FUNCTION__
277 #else
278 /* If a debug message says <mdb_unknown>(), update the #if statements above */
279 # define mdb_func_ "<mdb_unknown>"
280 #endif
281
282 /* Internal error codes, not exposed outside liblmdb */
283 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
284 #ifdef _WIN32
285 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
286 #elif defined MDB_USE_SYSV_SEM
287 #define MDB_OWNERDEAD (MDB_LAST_ERRCODE + 11)
288 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
289 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
290 #endif
291
292 #ifdef __GLIBC__
293 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__)
294 #endif
295 /** Some platforms define the EOWNERDEAD error code
296 * even though they don't support Robust Mutexes.
297 * Compile with -DMDB_USE_ROBUST=0, or use some other
298 * mechanism like -DMDB_USE_SYSV_SEM instead of
299 * -DMDB_USE_POSIX_MUTEX. (SysV semaphores are
300 * also Robust, but some systems don't support them
301 * either.)
302 */
303 #ifndef MDB_USE_ROBUST
304 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */
305 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \
306 (defined(__GLIBC__) && GLIBC_VER < 0x020004))
307 # define MDB_USE_ROBUST 0
308 # else
309 # define MDB_USE_ROBUST 1
310 /* glibc < 2.12 only provided _np API */
311 # if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \
312 (defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST))
313 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP
314 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag)
315 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex)
316 # endif
317 # endif
318 #endif /* MDB_USE_ROBUST */
319
320 #if defined(MDB_OWNERDEAD) && MDB_USE_ROBUST
321 #define MDB_ROBUST_SUPPORTED 1
322 #endif
323
324 #ifdef _WIN32
325 #define MDB_USE_HASH 1
326 #define MDB_PIDLOCK 0
327 #define THREAD_RET DWORD
328 #define pthread_t HANDLE
329 #define pthread_mutex_t HANDLE
330 #define pthread_cond_t HANDLE
331 typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
332 #define pthread_key_t DWORD
333 #define pthread_self() GetCurrentThreadId()
334 #define pthread_key_create(x,y) \
335 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
336 #define pthread_key_delete(x) TlsFree(x)
337 #define pthread_getspecific(x) TlsGetValue(x)
338 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
339 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
340 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
341 #define pthread_cond_signal(x) SetEvent(*x)
342 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
343 #define THREAD_CREATE(thr,start,arg) \
344 (((thr) = CreateThread(NULL, 0, start, arg, 0, NULL)) ? 0 : ErrCode())
345 #define THREAD_FINISH(thr) \
346 (WaitForSingleObject(thr, INFINITE) ? ErrCode() : 0)
347 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
348 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
349 #define mdb_mutex_consistent(mutex) 0
350 #define getpid() GetCurrentProcessId()
351 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
352 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
353 #define ErrCode() GetLastError()
354 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
355 #define close(fd) (CloseHandle(fd) ? 0 : -1)
356 #define munmap(ptr,len) UnmapViewOfFile(ptr)
357 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
358 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
359 #else
360 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
361 #endif
362 #else
363 #define THREAD_RET void *
364 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
365 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
366
367 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
368 #define MDB_PIDLOCK 1
369
370 #ifdef MDB_USE_POSIX_SEM
371
372 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
373 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
374 #define UNLOCK_MUTEX(mutex) sem_post(mutex)
375
376 static int
mdb_sem_wait(sem_t * sem)377 mdb_sem_wait(sem_t *sem)
378 {
379 int rc;
380 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
381 return rc;
382 }
383
384 #elif defined MDB_USE_SYSV_SEM
385
386 typedef struct mdb_mutex {
387 int semid;
388 int semnum;
389 int *locked;
390 } mdb_mutex_t[1], *mdb_mutexref_t;
391
392 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
393 #define UNLOCK_MUTEX(mutex) do { \
394 struct sembuf sb = { 0, 1, SEM_UNDO }; \
395 sb.sem_num = (mutex)->semnum; \
396 *(mutex)->locked = 0; \
397 semop((mutex)->semid, &sb, 1); \
398 } while(0)
399
400 static int
mdb_sem_wait(mdb_mutexref_t sem)401 mdb_sem_wait(mdb_mutexref_t sem)
402 {
403 int rc, *locked = sem->locked;
404 struct sembuf sb = { 0, -1, SEM_UNDO };
405 sb.sem_num = sem->semnum;
406 do {
407 if (!semop(sem->semid, &sb, 1)) {
408 rc = *locked ? MDB_OWNERDEAD : MDB_SUCCESS;
409 *locked = 1;
410 break;
411 }
412 } while ((rc = errno) == EINTR);
413 return rc;
414 }
415
416 #define mdb_mutex_consistent(mutex) 0
417
418 #else /* MDB_USE_POSIX_MUTEX: */
419 /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
420 * local variables keep it (mdb_mutexref_t).
421 *
422 * An mdb_mutex_t can be assigned to an mdb_mutexref_t. They can
423 * be the same, or an array[size 1] and a pointer.
424 * @{
425 */
426 typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
427 /* @} */
428 /** Lock the reader or writer mutex.
429 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
430 */
431 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
432 /** Unlock the reader or writer mutex.
433 */
434 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
435 /** Mark mutex-protected data as repaired, after death of previous owner.
436 */
437 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
438 #endif /* MDB_USE_POSIX_SEM || MDB_USE_SYSV_SEM */
439
440 /** Get the error code for the last failed system function.
441 */
442 #define ErrCode() errno
443
444 /** An abstraction for a file handle.
445 * On POSIX systems file handles are small integers. On Windows
446 * they're opaque pointers.
447 */
448 #define HANDLE int
449
450 /** A value for an invalid file handle.
451 * Mainly used to initialize file variables and signify that they are
452 * unused.
453 */
454 #define INVALID_HANDLE_VALUE (-1)
455
456 /** Get the size of a memory page for the system.
457 * This is the basic size that the platform's memory manager uses, and is
458 * fundamental to the use of memory-mapped files.
459 */
460 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
461 #endif
462
463 #define Z MDB_FMT_Z /**< printf/scanf format modifier for size_t */
464 #define Yu MDB_PRIy(u) /**< printf format for #mdb_size_t */
465 #define Yd MDB_PRIy(d) /**< printf format for "signed #mdb_size_t" */
466
467 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
468 #define MNAME_LEN 32
469 #elif defined(MDB_USE_SYSV_SEM)
470 #define MNAME_LEN (sizeof(int))
471 #else
472 #define MNAME_LEN (sizeof(pthread_mutex_t))
473 #endif
474
475 #ifdef MDB_USE_SYSV_SEM
476 #define SYSV_SEM_FLAG 1 /**< SysV sems in lockfile format */
477 #else
478 #define SYSV_SEM_FLAG 0
479 #endif
480
481 /** @} */
482
483 #ifdef MDB_ROBUST_SUPPORTED
484 /** Lock mutex, handle any error, set rc = result.
485 * Return 0 on success, nonzero (not rc) on error.
486 */
487 #define LOCK_MUTEX(rc, env, mutex) \
488 (((rc) = LOCK_MUTEX0(mutex)) && \
489 ((rc) = mdb_mutex_failed(env, mutex, rc)))
490 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
491 #else
492 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
493 #define mdb_mutex_failed(env, mutex, rc) (rc)
494 #endif
495
496 #ifndef _WIN32
497 /** A flag for opening a file and requesting synchronous data writes.
498 * This is only used when writing a meta page. It's not strictly needed;
499 * we could just do a normal write and then immediately perform a flush.
500 * But if this flag is available it saves us an extra system call.
501 *
502 * @note If O_DSYNC is undefined but exists in /usr/include,
503 * preferably set some compiler flag to get the definition.
504 */
505 #ifndef MDB_DSYNC
506 # ifdef O_DSYNC
507 # define MDB_DSYNC O_DSYNC
508 # else
509 # define MDB_DSYNC O_SYNC
510 # endif
511 #endif
512 #endif
513
514 /** Function for flushing the data of a file. Define this to fsync
515 * if fdatasync() is not supported.
516 */
517 #ifndef MDB_FDATASYNC
518 # define MDB_FDATASYNC fdatasync
519 #endif
520
521 #ifndef MDB_MSYNC
522 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
523 #endif
524
525 #ifndef MS_SYNC
526 #define MS_SYNC 1
527 #endif
528
529 #ifndef MS_ASYNC
530 #define MS_ASYNC 0
531 #endif
532
533 /** A page number in the database.
534 * Note that 64 bit page numbers are overkill, since pages themselves
535 * already represent 12-13 bits of addressable memory, and the OS will
536 * always limit applications to a maximum of 63 bits of address space.
537 *
538 * @note In the #MDB_node structure, we only store 48 bits of this value,
539 * which thus limits us to only 60 bits of addressable data.
540 */
541 typedef MDB_ID pgno_t;
542
543 /** A transaction ID.
544 * See struct MDB_txn.mt_txnid for details.
545 */
546 typedef MDB_ID txnid_t;
547
548 /** @defgroup debug Debug Macros
549 * @{
550 */
551 #ifndef MDB_DEBUG
552 /** Enable debug output. Needs variable argument macros (a C99 feature).
553 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
554 * read from and written to the database (used for free space management).
555 */
556 #define MDB_DEBUG 0
557 #endif
558
559 #if MDB_DEBUG
560 static int mdb_debug;
561 static txnid_t mdb_debug_start;
562
563 /** Print a debug message with printf formatting.
564 * Requires double parenthesis around 2 or more args.
565 */
566 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
567 # define DPRINTF0(fmt, ...) \
568 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
569 #else
570 # define DPRINTF(args) ((void) 0)
571 #endif
572 /** Print a debug string.
573 * The string is printed literally, with no format processing.
574 */
575 #define DPUTS(arg) DPRINTF(("%s", arg))
576 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
577 #define DDBI(mc) \
578 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
579 /** @} */
580
581 /** @brief The maximum size of a database page.
582 *
583 * It is 32k or 64k, since value-PAGEBASE must fit in
584 * #MDB_page.%mp_upper.
585 *
586 * LMDB will use database pages < OS pages if needed.
587 * That causes more I/O in write transactions: The OS must
588 * know (read) the whole page before writing a partial page.
589 *
590 * Note that we don't currently support Huge pages. On Linux,
591 * regular data files cannot use Huge pages, and in general
592 * Huge pages aren't actually pageable. We rely on the OS
593 * demand-pager to read our data and page it out when memory
594 * pressure from other processes is high. So until OSs have
595 * actual paging support for Huge pages, they're not viable.
596 */
597 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
598
599 /** The minimum number of keys required in a database page.
600 * Setting this to a larger value will place a smaller bound on the
601 * maximum size of a data item. Data items larger than this size will
602 * be pushed into overflow pages instead of being stored directly in
603 * the B-tree node. This value used to default to 4. With a page size
604 * of 4096 bytes that meant that any item larger than 1024 bytes would
605 * go into an overflow page. That also meant that on average 2-3KB of
606 * each overflow page was wasted space. The value cannot be lower than
607 * 2 because then there would no longer be a tree structure. With this
608 * value, items larger than 2KB will go into overflow pages, and on
609 * average only 1KB will be wasted.
610 */
611 #define MDB_MINKEYS 2
612
613 /** A stamp that identifies a file as an LMDB file.
614 * There's nothing special about this value other than that it is easily
615 * recognizable, and it will reflect any byte order mismatches.
616 */
617 #define MDB_MAGIC 0xBEEFC0DE
618
619 /** The version number for a database's datafile format. */
620 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
621 /** The version number for a database's lockfile format. */
622 #define MDB_LOCK_VERSION ((MDB_DEVEL) ? 999 : 1)
623
624 /** @brief The max size of a key we can write, or 0 for computed max.
625 *
626 * This macro should normally be left alone or set to 0.
627 * Note that a database with big keys or dupsort data cannot be
628 * reliably modified by a liblmdb which uses a smaller max.
629 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
630 *
631 * Other values are allowed, for backwards compat. However:
632 * A value bigger than the computed max can break if you do not
633 * know what you are doing, and liblmdb <= 0.9.10 can break when
634 * modifying a DB with keys/dupsort data bigger than its max.
635 *
636 * Data items in an #MDB_DUPSORT database are also limited to
637 * this size, since they're actually keys of a sub-DB. Keys and
638 * #MDB_DUPSORT data items must fit on a node in a regular page.
639 */
640 #ifndef MDB_MAXKEYSIZE
641 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
642 #endif
643
644 /** The maximum size of a key we can write to the environment. */
645 #if MDB_MAXKEYSIZE
646 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
647 #else
648 #define ENV_MAXKEY(env) ((env)->me_maxkey)
649 #endif
650
651 /** @brief The maximum size of a data item.
652 *
653 * We only store a 32 bit value for node sizes.
654 */
655 #define MAXDATASIZE 0xffffffffUL
656
657 #if MDB_DEBUG
658 /** Key size which fits in a #DKBUF.
659 * @ingroup debug
660 */
661 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
662 /** A key buffer.
663 * @ingroup debug
664 * This is used for printing a hex dump of a key's contents.
665 */
666 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
667 /** Display a key in hex.
668 * @ingroup debug
669 * Invoke a function to display a key in hex.
670 */
671 #define DKEY(x) mdb_dkey(x, kbuf)
672 #else
673 #define DKBUF
674 #define DKEY(x) 0
675 #endif
676
677 /** An invalid page number.
678 * Mainly used to denote an empty tree.
679 */
680 #define P_INVALID (~(pgno_t)0)
681
682 /** Test if the flags \b f are set in a flag word \b w. */
683 #define F_ISSET(w, f) (((w) & (f)) == (f))
684
685 /** Round \b n up to an even number. */
686 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
687
688 /** Used for offsets within a single page.
689 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
690 * this is plenty.
691 */
692 typedef uint16_t indx_t;
693
694 /** Default size of memory map.
695 * This is certainly too small for any actual applications. Apps should always set
696 * the size explicitly using #mdb_env_set_mapsize().
697 */
698 #define DEFAULT_MAPSIZE 1048576
699
700 /** @defgroup readers Reader Lock Table
701 * Readers don't acquire any locks for their data access. Instead, they
702 * simply record their transaction ID in the reader table. The reader
703 * mutex is needed just to find an empty slot in the reader table. The
704 * slot's address is saved in thread-specific data so that subsequent read
705 * transactions started by the same thread need no further locking to proceed.
706 *
707 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
708 *
709 * No reader table is used if the database is on a read-only filesystem, or
710 * if #MDB_NOLOCK is set.
711 *
712 * Since the database uses multi-version concurrency control, readers don't
713 * actually need any locking. This table is used to keep track of which
714 * readers are using data from which old transactions, so that we'll know
715 * when a particular old transaction is no longer in use. Old transactions
716 * that have discarded any data pages can then have those pages reclaimed
717 * for use by a later write transaction.
718 *
719 * The lock table is constructed such that reader slots are aligned with the
720 * processor's cache line size. Any slot is only ever used by one thread.
721 * This alignment guarantees that there will be no contention or cache
722 * thrashing as threads update their own slot info, and also eliminates
723 * any need for locking when accessing a slot.
724 *
725 * A writer thread will scan every slot in the table to determine the oldest
726 * outstanding reader transaction. Any freed pages older than this will be
727 * reclaimed by the writer. The writer doesn't use any locks when scanning
728 * this table. This means that there's no guarantee that the writer will
729 * see the most up-to-date reader info, but that's not required for correct
730 * operation - all we need is to know the upper bound on the oldest reader,
731 * we don't care at all about the newest reader. So the only consequence of
732 * reading stale information here is that old pages might hang around a
733 * while longer before being reclaimed. That's actually good anyway, because
734 * the longer we delay reclaiming old pages, the more likely it is that a
735 * string of contiguous pages can be found after coalescing old pages from
736 * many old transactions together.
737 * @{
738 */
739 /** Number of slots in the reader table.
740 * This value was chosen somewhat arbitrarily. 126 readers plus a
741 * couple mutexes fit exactly into 8KB on my development machine.
742 * Applications should set the table size using #mdb_env_set_maxreaders().
743 */
744 #define DEFAULT_READERS 126
745
746 /** The size of a CPU cache line in bytes. We want our lock structures
747 * aligned to this size to avoid false cache line sharing in the
748 * lock table.
749 * This value works for most CPUs. For Itanium this should be 128.
750 */
751 #ifndef CACHELINE
752 #define CACHELINE 64
753 #endif
754
755 /** The information we store in a single slot of the reader table.
756 * In addition to a transaction ID, we also record the process and
757 * thread ID that owns a slot, so that we can detect stale information,
758 * e.g. threads or processes that went away without cleaning up.
759 * @note We currently don't check for stale records. We simply re-init
760 * the table when we know that we're the only process opening the
761 * lock file.
762 */
763 typedef struct MDB_rxbody {
764 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
765 * Multiple readers that start at the same time will probably have the
766 * same ID here. Again, it's not important to exclude them from
767 * anything; all we need to know is which version of the DB they
768 * started from so we can avoid overwriting any data used in that
769 * particular version.
770 */
771 volatile txnid_t mrb_txnid;
772 /** The process ID of the process owning this reader txn. */
773 volatile MDB_PID_T mrb_pid;
774 /** The thread ID of the thread owning this txn. */
775 volatile MDB_THR_T mrb_tid;
776 } MDB_rxbody;
777
778 /** The actual reader record, with cacheline padding. */
779 typedef struct MDB_reader {
780 union {
781 MDB_rxbody mrx;
782 /** shorthand for mrb_txnid */
783 #define mr_txnid mru.mrx.mrb_txnid
784 #define mr_pid mru.mrx.mrb_pid
785 #define mr_tid mru.mrx.mrb_tid
786 /** cache line alignment */
787 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
788 } mru;
789 } MDB_reader;
790
791 /** The header for the reader table.
792 * The table resides in a memory-mapped file. (This is a different file
793 * than is used for the main database.)
794 *
795 * For POSIX the actual mutexes reside in the shared memory of this
796 * mapped file. On Windows, mutexes are named objects allocated by the
797 * kernel; we store the mutex names in this mapped file so that other
798 * processes can grab them. This same approach is also used on
799 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
800 * process-shared POSIX mutexes. For these cases where a named object
801 * is used, the object name is derived from a 64 bit FNV hash of the
802 * environment pathname. As such, naming collisions are extremely
803 * unlikely. If a collision occurs, the results are unpredictable.
804 */
805 typedef struct MDB_txbody {
806 /** Stamp identifying this as an LMDB file. It must be set
807 * to #MDB_MAGIC. */
808 uint32_t mtb_magic;
809 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
810 uint32_t mtb_format;
811 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
812 char mtb_rmname[MNAME_LEN];
813 #elif defined(MDB_USE_SYSV_SEM)
814 int mtb_semid;
815 int mtb_rlocked;
816 #else
817 /** Mutex protecting access to this table.
818 * This is the reader table lock used with LOCK_MUTEX().
819 */
820 mdb_mutex_t mtb_rmutex;
821 #endif
822 /** The ID of the last transaction committed to the database.
823 * This is recorded here only for convenience; the value can always
824 * be determined by reading the main database meta pages.
825 */
826 volatile txnid_t mtb_txnid;
827 /** The number of slots that have been used in the reader table.
828 * This always records the maximum count, it is not decremented
829 * when readers release their slots.
830 */
831 volatile unsigned mtb_numreaders;
832 } MDB_txbody;
833
834 /** The actual reader table definition. */
835 typedef struct MDB_txninfo {
836 union {
837 MDB_txbody mtb;
838 #define mti_magic mt1.mtb.mtb_magic
839 #define mti_format mt1.mtb.mtb_format
840 #define mti_rmutex mt1.mtb.mtb_rmutex
841 #define mti_rmname mt1.mtb.mtb_rmname
842 #define mti_txnid mt1.mtb.mtb_txnid
843 #define mti_numreaders mt1.mtb.mtb_numreaders
844 #ifdef MDB_USE_SYSV_SEM
845 #define mti_semid mt1.mtb.mtb_semid
846 #define mti_rlocked mt1.mtb.mtb_rlocked
847 #endif
848 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
849 } mt1;
850 union {
851 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
852 char mt2_wmname[MNAME_LEN];
853 #define mti_wmname mt2.mt2_wmname
854 #elif defined MDB_USE_SYSV_SEM
855 int mt2_wlocked;
856 #define mti_wlocked mt2.mt2_wlocked
857 #else
858 mdb_mutex_t mt2_wmutex;
859 #define mti_wmutex mt2.mt2_wmutex
860 #endif
861 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
862 } mt2;
863 MDB_reader mti_readers[1];
864 } MDB_txninfo;
865
866 /** Lockfile format signature: version, features and field layout */
867 #define MDB_LOCK_FORMAT \
868 ((uint32_t) \
869 ((MDB_LOCK_VERSION) \
870 /* Flags which describe functionality */ \
871 + (SYSV_SEM_FLAG << 18) \
872 + (((MDB_PIDLOCK) != 0) << 16)))
873 /** @} */
874
875 /** Common header for all page types. The page type depends on #mp_flags.
876 *
877 * #P_BRANCH and #P_LEAF pages have unsorted '#MDB_node's at the end, with
878 * sorted #mp_ptrs[] entries referring to them. Exception: #P_LEAF2 pages
879 * omit mp_ptrs and pack sorted #MDB_DUPFIXED values after the page header.
880 *
881 * #P_OVERFLOW records occupy one or more contiguous pages where only the
882 * first has a page header. They hold the real data of #F_BIGDATA nodes.
883 *
884 * #P_SUBP sub-pages are small leaf "pages" with duplicate data.
885 * A node with flag #F_DUPDATA but not #F_SUBDATA contains a sub-page.
886 * (Duplicate data can also go in sub-databases, which use normal pages.)
887 *
888 * #P_META pages contain #MDB_meta, the start point of an LMDB snapshot.
889 *
890 * Each non-metapage up to #MDB_meta.%mm_last_pg is reachable exactly once
891 * in the snapshot: Either used by a database or listed in a freeDB record.
892 */
893 typedef struct MDB_page {
894 #define mp_pgno mp_p.p_pgno
895 #define mp_next mp_p.p_next
896 union {
897 pgno_t p_pgno; /**< page number */
898 struct MDB_page *p_next; /**< for in-memory list of freed pages */
899 } mp_p;
900 uint16_t mp_pad; /**< key size if this is a LEAF2 page */
901 /** @defgroup mdb_page Page Flags
902 * @ingroup internal
903 * Flags for the page headers.
904 * @{
905 */
906 #define P_BRANCH 0x01 /**< branch page */
907 #define P_LEAF 0x02 /**< leaf page */
908 #define P_OVERFLOW 0x04 /**< overflow page */
909 #define P_META 0x08 /**< meta page */
910 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
911 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
912 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
913 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
914 #define P_KEEP 0x8000 /**< leave this page alone during spill */
915 /** @} */
916 uint16_t mp_flags; /**< @ref mdb_page */
917 #define mp_lower mp_pb.pb.pb_lower
918 #define mp_upper mp_pb.pb.pb_upper
919 #define mp_pages mp_pb.pb_pages
920 union {
921 struct {
922 indx_t pb_lower; /**< lower bound of free space */
923 indx_t pb_upper; /**< upper bound of free space */
924 } pb;
925 uint32_t pb_pages; /**< number of overflow pages */
926 } mp_pb;
927 indx_t mp_ptrs[1]; /**< dynamic size */
928 } MDB_page;
929
930 /** Size of the page header, excluding dynamic data at the end */
931 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
932
933 /** Address of first usable data byte in a page, after the header */
934 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
935
936 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
937 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
938
939 /** Number of nodes on a page */
940 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
941
942 /** The amount of space remaining in the page */
943 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
944
945 /** The percentage of space used in the page, in tenths of a percent. */
946 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
947 ((env)->me_psize - PAGEHDRSZ))
948 /** The minimum page fill factor, in tenths of a percent.
949 * Pages emptier than this are candidates for merging.
950 */
951 #define FILL_THRESHOLD 250
952
953 /** Test if a page is a leaf page */
954 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
955 /** Test if a page is a LEAF2 page */
956 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
957 /** Test if a page is a branch page */
958 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
959 /** Test if a page is an overflow page */
960 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
961 /** Test if a page is a sub page */
962 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
963
964 /** The number of overflow pages needed to store the given size. */
965 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
966
967 /** Link in #MDB_txn.%mt_loose_pgs list.
968 * Kept outside the page header, which is needed when reusing the page.
969 */
970 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
971
972 /** Header for a single key/data pair within a page.
973 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
974 * We guarantee 2-byte alignment for 'MDB_node's.
975 */
976 typedef struct MDB_node {
977 /** lo and hi are used for data size on leaf nodes and for
978 * child pgno on branch nodes. On 64 bit platforms, flags
979 * is also used for pgno. (Branch nodes have no flags).
980 * They are in host byte order in case that lets some
981 * accesses be optimized into a 32-bit word access.
982 */
983 #if BYTE_ORDER == LITTLE_ENDIAN
984 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
985 #else
986 unsigned short mn_hi, mn_lo;
987 #endif
988 /** @defgroup mdb_node Node Flags
989 * @ingroup internal
990 * Flags for node headers.
991 * @{
992 */
993 #define F_BIGDATA 0x01 /**< data put on overflow page */
994 #define F_SUBDATA 0x02 /**< data is a sub-database */
995 #define F_DUPDATA 0x04 /**< data has duplicates */
996
997 /** valid flags for #mdb_node_add() */
998 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
999
1000 /** @} */
1001 unsigned short mn_flags; /**< @ref mdb_node */
1002 unsigned short mn_ksize; /**< key size */
1003 char mn_data[1]; /**< key and data are appended here */
1004 } MDB_node;
1005
1006 /** Size of the node header, excluding dynamic data at the end */
1007 #define NODESIZE offsetof(MDB_node, mn_data)
1008
1009 /** Bit position of top word in page number, for shifting mn_flags */
1010 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
1011
1012 /** Size of a node in a branch page with a given key.
1013 * This is just the node header plus the key, there is no data.
1014 */
1015 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
1016
1017 /** Size of a node in a leaf page with a given key and data.
1018 * This is node header plus key plus data size.
1019 */
1020 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
1021
1022 /** Address of node \b i in page \b p */
1023 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
1024
1025 /** Address of the key for the node */
1026 #define NODEKEY(node) (void *)((node)->mn_data)
1027
1028 /** Address of the data for a node */
1029 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
1030
1031 /** Get the page number pointed to by a branch node */
1032 #define NODEPGNO(node) \
1033 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
1034 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
1035 /** Set the page number in a branch node */
1036 #define SETPGNO(node,pgno) do { \
1037 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
1038 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
1039
1040 /** Get the size of the data in a leaf node */
1041 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
1042 /** Set the size of the data for a leaf node */
1043 #define SETDSZ(node,size) do { \
1044 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
1045 /** The size of a key in a node */
1046 #define NODEKSZ(node) ((node)->mn_ksize)
1047
1048 /** Copy a page number from src to dst */
1049 #ifdef MISALIGNED_OK
1050 #define COPY_PGNO(dst,src) dst = src
1051 #else
1052 #if MDB_SIZE_MAX > 0xffffffffU
1053 #define COPY_PGNO(dst,src) do { \
1054 unsigned short *s, *d; \
1055 s = (unsigned short *)&(src); \
1056 d = (unsigned short *)&(dst); \
1057 *d++ = *s++; \
1058 *d++ = *s++; \
1059 *d++ = *s++; \
1060 *d = *s; \
1061 } while (0)
1062 #else
1063 #define COPY_PGNO(dst,src) do { \
1064 unsigned short *s, *d; \
1065 s = (unsigned short *)&(src); \
1066 d = (unsigned short *)&(dst); \
1067 *d++ = *s++; \
1068 *d = *s; \
1069 } while (0)
1070 #endif
1071 #endif
1072 /** The address of a key in a LEAF2 page.
1073 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
1074 * There are no node headers, keys are stored contiguously.
1075 */
1076 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
1077
1078 /** Set the \b node's key into \b keyptr, if requested. */
1079 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
1080 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
1081
1082 /** Set the \b node's key into \b key. */
1083 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
1084
1085 /** Information about a single database in the environment. */
1086 typedef struct MDB_db {
1087 uint32_t md_pad; /**< also ksize for LEAF2 pages */
1088 uint16_t md_flags; /**< @ref mdb_dbi_open */
1089 uint16_t md_depth; /**< depth of this tree */
1090 pgno_t md_branch_pages; /**< number of internal pages */
1091 pgno_t md_leaf_pages; /**< number of leaf pages */
1092 pgno_t md_overflow_pages; /**< number of overflow pages */
1093 mdb_size_t md_entries; /**< number of data items */
1094 pgno_t md_root; /**< the root page of this tree */
1095 } MDB_db;
1096
1097 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
1098 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
1099 /** #mdb_dbi_open() flags */
1100 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
1101 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
1102
1103 /** Handle for the DB used to track free pages. */
1104 #define FREE_DBI 0
1105 /** Handle for the default DB. */
1106 #define MAIN_DBI 1
1107 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */
1108 #define CORE_DBS 2
1109
1110 /** Number of meta pages - also hardcoded elsewhere */
1111 #define NUM_METAS 2
1112
1113 /** Meta page content.
1114 * A meta page is the start point for accessing a database snapshot.
1115 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
1116 */
1117 typedef struct MDB_meta {
1118 /** Stamp identifying this as an LMDB file. It must be set
1119 * to #MDB_MAGIC. */
1120 uint32_t mm_magic;
1121 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
1122 uint32_t mm_version;
1123 #ifdef MDB_VL32
1124 union { /* always zero since we don't support fixed mapping in MDB_VL32 */
1125 MDB_ID mmun_ull;
1126 void *mmun_address;
1127 } mm_un;
1128 #define mm_address mm_un.mmun_address
1129 #else
1130 void *mm_address; /**< address for fixed mapping */
1131 #endif
1132 pgno_t mm_mapsize; /**< size of mmap region */
1133 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */
1134 /** The size of pages used in this DB */
1135 #define mm_psize mm_dbs[FREE_DBI].md_pad
1136 /** Any persistent environment flags. @ref mdb_env */
1137 #define mm_flags mm_dbs[FREE_DBI].md_flags
1138 /** Last used page in the datafile.
1139 * Actually the file may be shorter if the freeDB lists the final pages.
1140 */
1141 pgno_t mm_last_pg;
1142 volatile txnid_t mm_txnid; /**< txnid that committed this page */
1143 } MDB_meta;
1144
1145 /** Buffer for a stack-allocated meta page.
1146 * The members define size and alignment, and silence type
1147 * aliasing warnings. They are not used directly; that could
1148 * mean incorrectly using several union members in parallel.
1149 */
1150 typedef union MDB_metabuf {
1151 MDB_page mb_page;
1152 struct {
1153 char mm_pad[PAGEHDRSZ];
1154 MDB_meta mm_meta;
1155 } mb_metabuf;
1156 } MDB_metabuf;
1157
1158 /** Auxiliary DB info.
1159 * The information here is mostly static/read-only. There is
1160 * only a single copy of this record in the environment.
1161 */
1162 typedef struct MDB_dbx {
1163 MDB_val md_name; /**< name of the database */
1164 MDB_cmp_func *md_cmp; /**< function for comparing keys */
1165 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
1166 MDB_rel_func *md_rel; /**< user relocate function */
1167 void *md_relctx; /**< user-provided context for md_rel */
1168 } MDB_dbx;
1169
1170 /** A database transaction.
1171 * Every operation requires a transaction handle.
1172 */
1173 struct MDB_txn {
1174 MDB_txn *mt_parent; /**< parent of a nested txn */
1175 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
1176 MDB_txn *mt_child;
1177 pgno_t mt_next_pgno; /**< next unallocated page */
1178 #ifdef MDB_VL32
1179 pgno_t mt_last_pgno; /**< last written page */
1180 #endif
1181 /** The ID of this transaction. IDs are integers incrementing from 1.
1182 * Only committed write transactions increment the ID. If a transaction
1183 * aborts, the ID may be re-used by the next writer.
1184 */
1185 txnid_t mt_txnid;
1186 MDB_env *mt_env; /**< the DB environment */
1187 /** The list of pages that became unused during this transaction.
1188 */
1189 MDB_IDL mt_free_pgs;
1190 /** The list of loose pages that became unused and may be reused
1191 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
1192 */
1193 MDB_page *mt_loose_pgs;
1194 /** Number of loose pages (#mt_loose_pgs) */
1195 int mt_loose_count;
1196 /** The sorted list of dirty pages we temporarily wrote to disk
1197 * because the dirty list was full. page numbers in here are
1198 * shifted left by 1, deleted slots have the LSB set.
1199 */
1200 MDB_IDL mt_spill_pgs;
1201 union {
1202 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
1203 MDB_ID2L dirty_list;
1204 /** For read txns: This thread/txn's reader table slot, or NULL. */
1205 MDB_reader *reader;
1206 } mt_u;
1207 /** Array of records for each DB known in the environment. */
1208 MDB_dbx *mt_dbxs;
1209 /** Array of MDB_db records for each known DB */
1210 MDB_db *mt_dbs;
1211 /** Array of sequence numbers for each DB handle */
1212 unsigned int *mt_dbiseqs;
1213 /** @defgroup mt_dbflag Transaction DB Flags
1214 * @ingroup internal
1215 * @{
1216 */
1217 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1218 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1219 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1220 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1221 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */
1222 /** @} */
1223 /** In write txns, array of cursors for each DB */
1224 MDB_cursor **mt_cursors;
1225 /** Array of flags for each DB */
1226 unsigned char *mt_dbflags;
1227 #ifdef MDB_VL32
1228 /** List of read-only pages (actually chunks) */
1229 MDB_ID3L mt_rpages;
1230 /** We map chunks of 16 pages. Even though Windows uses 4KB pages, all
1231 * mappings must begin on 64KB boundaries. So we round off all pgnos to
1232 * a chunk boundary. We do the same on Linux for symmetry, and also to
1233 * reduce the frequency of mmap/munmap calls.
1234 */
1235 #define MDB_RPAGE_CHUNK 16
1236 #define MDB_TRPAGE_SIZE 4096 /**< size of #mt_rpages array of chunks */
1237 #define MDB_TRPAGE_MAX (MDB_TRPAGE_SIZE-1) /**< maximum chunk index */
1238 unsigned int mt_rpcheck; /**< threshold for reclaiming unref'd chunks */
1239 #endif
1240 /** Number of DB records in use, or 0 when the txn is finished.
1241 * This number only ever increments until the txn finishes; we
1242 * don't decrement it when individual DB handles are closed.
1243 */
1244 MDB_dbi mt_numdbs;
1245
1246 /** @defgroup mdb_txn Transaction Flags
1247 * @ingroup internal
1248 * @{
1249 */
1250 /** #mdb_txn_begin() flags */
1251 #define MDB_TXN_BEGIN_FLAGS (MDB_NOMETASYNC|MDB_NOSYNC|MDB_RDONLY)
1252 #define MDB_TXN_NOMETASYNC MDB_NOMETASYNC /**< don't sync meta for this txn on commit */
1253 #define MDB_TXN_NOSYNC MDB_NOSYNC /**< don't sync this txn on commit */
1254 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1255 /* internal txn flags */
1256 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1257 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */
1258 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1259 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1260 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1261 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */
1262 /** most operations on the txn are currently illegal */
1263 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD)
1264 /** @} */
1265 unsigned int mt_flags; /**< @ref mdb_txn */
1266 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1267 * Includes ancestor txns' dirty pages not hidden by other txns'
1268 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1269 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1270 */
1271 unsigned int mt_dirty_room;
1272 };
1273
1274 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1275 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1276 * raise this on a 64 bit machine.
1277 */
1278 #define CURSOR_STACK 32
1279
1280 struct MDB_xcursor;
1281
1282 /** Cursors are used for all DB operations.
1283 * A cursor holds a path of (page pointer, key index) from the DB
1284 * root to a position in the DB, plus other state. #MDB_DUPSORT
1285 * cursors include an xcursor to the current data item. Write txns
1286 * track their cursors and keep them up to date when data moves.
1287 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1288 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1289 */
1290 struct MDB_cursor {
1291 /** Next cursor on this DB in this txn */
1292 MDB_cursor *mc_next;
1293 /** Backup of the original cursor if this cursor is a shadow */
1294 MDB_cursor *mc_backup;
1295 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1296 struct MDB_xcursor *mc_xcursor;
1297 /** The transaction that owns this cursor */
1298 MDB_txn *mc_txn;
1299 /** The database handle this cursor operates on */
1300 MDB_dbi mc_dbi;
1301 /** The database record for this cursor */
1302 MDB_db *mc_db;
1303 /** The database auxiliary record for this cursor */
1304 MDB_dbx *mc_dbx;
1305 /** The @ref mt_dbflag for this database */
1306 unsigned char *mc_dbflag;
1307 unsigned short mc_snum; /**< number of pushed pages */
1308 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1309 /** @defgroup mdb_cursor Cursor Flags
1310 * @ingroup internal
1311 * Cursor state flags.
1312 * @{
1313 */
1314 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1315 #define C_EOF 0x02 /**< No more data */
1316 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1317 #define C_DEL 0x08 /**< last op was a cursor_del */
1318 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1319 #define C_WRITEMAP MDB_TXN_WRITEMAP /**< Copy of txn flag */
1320 /** Read-only cursor into the txn's original snapshot in the map.
1321 * Set for read-only txns, and in #mdb_page_alloc() for #FREE_DBI when
1322 * #MDB_DEVEL & 2. Only implements code which is necessary for this.
1323 */
1324 #define C_ORIG_RDONLY MDB_TXN_RDONLY
1325 /** @} */
1326 unsigned int mc_flags; /**< @ref mdb_cursor */
1327 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1328 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1329 #ifdef MDB_VL32
1330 MDB_page *mc_ovpg; /**< a referenced overflow page */
1331 # define MC_OVPG(mc) ((mc)->mc_ovpg)
1332 # define MC_SET_OVPG(mc, pg) ((mc)->mc_ovpg = (pg))
1333 #else
1334 # define MC_OVPG(mc) ((MDB_page *)0)
1335 # define MC_SET_OVPG(mc, pg) ((void)0)
1336 #endif
1337 };
1338
1339 /** Context for sorted-dup records.
1340 * We could have gone to a fully recursive design, with arbitrarily
1341 * deep nesting of sub-databases. But for now we only handle these
1342 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1343 */
1344 typedef struct MDB_xcursor {
1345 /** A sub-cursor for traversing the Dup DB */
1346 MDB_cursor mx_cursor;
1347 /** The database record for this Dup DB */
1348 MDB_db mx_db;
1349 /** The auxiliary DB record for this Dup DB */
1350 MDB_dbx mx_dbx;
1351 /** The @ref mt_dbflag for this Dup DB */
1352 unsigned char mx_dbflag;
1353 } MDB_xcursor;
1354
1355 /** State of FreeDB old pages, stored in the MDB_env */
1356 typedef struct MDB_pgstate {
1357 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1358 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1359 } MDB_pgstate;
1360
1361 /** The database environment. */
1362 struct MDB_env {
1363 HANDLE me_fd; /**< The main data file */
1364 HANDLE me_lfd; /**< The lock file */
1365 HANDLE me_mfd; /**< just for writing the meta pages */
1366 #if defined(MDB_VL32) && defined(_WIN32)
1367 HANDLE me_fmh; /**< File Mapping handle */
1368 #endif
1369 /** Failed to update the meta page. Probably an I/O error. */
1370 #define MDB_FATAL_ERROR 0x80000000U
1371 /** Some fields are initialized. */
1372 #define MDB_ENV_ACTIVE 0x20000000U
1373 /** me_txkey is set */
1374 #define MDB_ENV_TXKEY 0x10000000U
1375 /** fdatasync is unreliable */
1376 #define MDB_FSYNCONLY 0x08000000U
1377 uint32_t me_flags; /**< @ref mdb_env */
1378 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1379 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1380 unsigned int me_maxreaders; /**< size of the reader table */
1381 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1382 volatile int me_close_readers;
1383 MDB_dbi me_numdbs; /**< number of DBs opened */
1384 MDB_dbi me_maxdbs; /**< size of the DB table */
1385 MDB_PID_T me_pid; /**< process ID of this env */
1386 char *me_path; /**< path to the DB files */
1387 char *me_map; /**< the memory map of the data file */
1388 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1389 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */
1390 void *me_pbuf; /**< scratch area for DUPSORT put() */
1391 MDB_txn *me_txn; /**< current write transaction */
1392 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1393 mdb_size_t me_mapsize; /**< size of the data memory map */
1394 off_t me_size; /**< current file size */
1395 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1396 MDB_dbx *me_dbxs; /**< array of static DB info */
1397 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1398 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1399 pthread_key_t me_txkey; /**< thread-key for readers */
1400 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1401 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1402 # define me_pglast me_pgstate.mf_pglast
1403 # define me_pghead me_pgstate.mf_pghead
1404 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1405 /** IDL of pages that became unused in a write txn */
1406 MDB_IDL me_free_pgs;
1407 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1408 MDB_ID2L me_dirty_list;
1409 /** Max number of freelist items that can fit in a single overflow page */
1410 int me_maxfree_1pg;
1411 /** Max size of a node on a page */
1412 unsigned int me_nodemax;
1413 #if !(MDB_MAXKEYSIZE)
1414 unsigned int me_maxkey; /**< max size of a key */
1415 #endif
1416 int me_live_reader; /**< have liveness lock in reader table */
1417 #ifdef _WIN32
1418 int me_pidquery; /**< Used in OpenProcess */
1419 #endif
1420 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
1421 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
1422 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
1423 #else
1424 mdb_mutex_t me_rmutex;
1425 mdb_mutex_t me_wmutex;
1426 #endif
1427 #ifdef MDB_VL32
1428 MDB_ID3L me_rpages; /**< like #mt_rpages, but global to env */
1429 pthread_mutex_t me_rpmutex; /**< control access to #me_rpages */
1430 #define MDB_ERPAGE_SIZE 16384
1431 #define MDB_ERPAGE_MAX (MDB_ERPAGE_SIZE-1)
1432 unsigned int me_rpcheck;
1433 #endif
1434 void *me_userctx; /**< User-settable context */
1435 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1436 };
1437
1438 /** Nested transaction */
1439 typedef struct MDB_ntxn {
1440 MDB_txn mnt_txn; /**< the transaction */
1441 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1442 } MDB_ntxn;
1443
1444 /** max number of pages to commit in one writev() call */
1445 #define MDB_COMMIT_PAGES 64
1446 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1447 #undef MDB_COMMIT_PAGES
1448 #define MDB_COMMIT_PAGES IOV_MAX
1449 #endif
1450
1451 /** max bytes to write in one call */
1452 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4))
1453
1454 /** Check \b txn and \b dbi arguments to a function */
1455 #define TXN_DBI_EXIST(txn, dbi, validity) \
1456 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity)))
1457
1458 /** Check for misused \b dbi handles */
1459 #define TXN_DBI_CHANGED(txn, dbi) \
1460 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1461
1462 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1463 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1464 static int mdb_page_touch(MDB_cursor *mc);
1465
1466 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \
1467 "reset-tmp", "fail-begin", "fail-beginchild"}
1468 enum {
1469 /* mdb_txn_end operation number, for logging */
1470 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET,
1471 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD
1472 };
1473 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */
1474 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */
1475 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */
1476 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */
1477 static void mdb_txn_end(MDB_txn *txn, unsigned mode);
1478
1479 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl);
1480 static int mdb_page_search_root(MDB_cursor *mc,
1481 MDB_val *key, int modify);
1482 #define MDB_PS_MODIFY 1
1483 #define MDB_PS_ROOTONLY 2
1484 #define MDB_PS_FIRST 4
1485 #define MDB_PS_LAST 8
1486 static int mdb_page_search(MDB_cursor *mc,
1487 MDB_val *key, int flags);
1488 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1489
1490 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1491 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1492 pgno_t newpgno, unsigned int nflags);
1493
1494 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1495 static MDB_meta *mdb_env_pick_meta(const MDB_env *env);
1496 static int mdb_env_write_meta(MDB_txn *txn);
1497 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */
1498 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1499 #endif
1500 static void mdb_env_close0(MDB_env *env, int excl);
1501
1502 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1503 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1504 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1505 static void mdb_node_del(MDB_cursor *mc, int ksize);
1506 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1507 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft);
1508 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data);
1509 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1510 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1511
1512 static int mdb_rebalance(MDB_cursor *mc);
1513 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1514
1515 static void mdb_cursor_pop(MDB_cursor *mc);
1516 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1517
1518 static int mdb_cursor_del0(MDB_cursor *mc);
1519 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1520 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1521 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1522 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1523 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1524 int *exactp);
1525 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1526 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1527
1528 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1529 static void mdb_xcursor_init0(MDB_cursor *mc);
1530 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1531 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1532
1533 static int mdb_drop0(MDB_cursor *mc, int subs);
1534 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1535 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
1536
1537 /** @cond */
1538 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1539 /** @endcond */
1540
1541 /** Compare two items pointing at '#mdb_size_t's of unknown alignment. */
1542 #ifdef MISALIGNED_OK
1543 # define mdb_cmp_clong mdb_cmp_long
1544 #else
1545 # define mdb_cmp_clong mdb_cmp_cint
1546 #endif
1547
1548 /** True if we need #mdb_cmp_clong() instead of \b cmp for #MDB_INTEGERDUP */
1549 #define NEED_CMP_CLONG(cmp, ksize) \
1550 (UINT_MAX < MDB_SIZE_MAX && \
1551 (cmp) == mdb_cmp_int && (ksize) == sizeof(mdb_size_t))
1552
1553 #ifdef _WIN32
1554 static SECURITY_DESCRIPTOR mdb_null_sd;
1555 static SECURITY_ATTRIBUTES mdb_all_sa;
1556 static int mdb_sec_inited;
1557
1558 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize);
1559 #endif
1560
1561 /** Return the library version info. */
1562 char * ESECT
mdb_version(int * major,int * minor,int * patch)1563 mdb_version(int *major, int *minor, int *patch)
1564 {
1565 if (major) *major = MDB_VERSION_MAJOR;
1566 if (minor) *minor = MDB_VERSION_MINOR;
1567 if (patch) *patch = MDB_VERSION_PATCH;
1568 return MDB_VERSION_STRING;
1569 }
1570
1571 /** Table of descriptions for LMDB @ref errors */
1572 static char *const mdb_errstr[] = {
1573 "MDB_KEYEXIST: Key/data pair already exists",
1574 "MDB_NOTFOUND: No matching key/data pair found",
1575 "MDB_PAGE_NOTFOUND: Requested page not found",
1576 "MDB_CORRUPTED: Located page was wrong type",
1577 "MDB_PANIC: Update of meta page failed or environment had fatal error",
1578 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1579 "MDB_INVALID: File is not an LMDB file",
1580 "MDB_MAP_FULL: Environment mapsize limit reached",
1581 "MDB_DBS_FULL: Environment maxdbs limit reached",
1582 "MDB_READERS_FULL: Environment maxreaders limit reached",
1583 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1584 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1585 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1586 "MDB_PAGE_FULL: Internal error - page has no more space",
1587 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1588 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1589 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1590 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid",
1591 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1592 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1593 "MDB_PROBLEM: Unexpected problem - txn should abort",
1594 };
1595
1596 char *
mdb_strerror(int err)1597 mdb_strerror(int err)
1598 {
1599 #ifdef _WIN32
1600 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1601 * This works as long as no function between the call to mdb_strerror
1602 * and the actual use of the message uses more than 4K of stack.
1603 */
1604 #define MSGSIZE 1024
1605 #define PADSIZE 4096
1606 char buf[MSGSIZE+PADSIZE], *ptr = buf;
1607 #endif
1608 int i;
1609 if (!err)
1610 return ("Successful return: 0");
1611
1612 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1613 i = err - MDB_KEYEXIST;
1614 return mdb_errstr[i];
1615 }
1616
1617 #ifdef _WIN32
1618 /* These are the C-runtime error codes we use. The comment indicates
1619 * their numeric value, and the Win32 error they would correspond to
1620 * if the error actually came from a Win32 API. A major mess, we should
1621 * have used LMDB-specific error codes for everything.
1622 */
1623 switch(err) {
1624 case ENOENT: /* 2, FILE_NOT_FOUND */
1625 case EIO: /* 5, ACCESS_DENIED */
1626 case ENOMEM: /* 12, INVALID_ACCESS */
1627 case EACCES: /* 13, INVALID_DATA */
1628 case EBUSY: /* 16, CURRENT_DIRECTORY */
1629 case EINVAL: /* 22, BAD_COMMAND */
1630 case ENOSPC: /* 28, OUT_OF_PAPER */
1631 return strerror(err);
1632 default:
1633 ;
1634 }
1635 buf[0] = 0;
1636 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
1637 FORMAT_MESSAGE_IGNORE_INSERTS,
1638 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE);
1639 return ptr;
1640 #else
1641 return strerror(err);
1642 #endif
1643 }
1644
1645 /** assert(3) variant in cursor context */
1646 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1647 /** assert(3) variant in transaction context */
1648 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1649 /** assert(3) variant in environment context */
1650 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1651
1652 #ifndef NDEBUG
1653 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1654 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1655
1656 static void ESECT
mdb_assert_fail(MDB_env * env,const char * expr_txt,const char * func,const char * file,int line)1657 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1658 const char *func, const char *file, int line)
1659 {
1660 char buf[400];
1661 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1662 file, line, expr_txt, func);
1663 if (env->me_assert_func)
1664 env->me_assert_func(env, buf);
1665 fprintf(stderr, "%s\n", buf);
1666 abort();
1667 }
1668 #else
1669 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1670 #endif /* NDEBUG */
1671
1672 #if MDB_DEBUG
1673 /** Return the page number of \b mp which may be sub-page, for debug output */
1674 static pgno_t
mdb_dbg_pgno(MDB_page * mp)1675 mdb_dbg_pgno(MDB_page *mp)
1676 {
1677 pgno_t ret;
1678 COPY_PGNO(ret, mp->mp_pgno);
1679 return ret;
1680 }
1681
1682 /** Display a key in hexadecimal and return the address of the result.
1683 * @param[in] key the key to display
1684 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1685 * @return The key in hexadecimal form.
1686 */
1687 char *
mdb_dkey(MDB_val * key,char * buf)1688 mdb_dkey(MDB_val *key, char *buf)
1689 {
1690 char *ptr = buf;
1691 unsigned char *c = key->mv_data;
1692 unsigned int i;
1693
1694 if (!key)
1695 return "";
1696
1697 if (key->mv_size > DKBUF_MAXKEYSIZE)
1698 return "MDB_MAXKEYSIZE";
1699 /* may want to make this a dynamic check: if the key is mostly
1700 * printable characters, print it as-is instead of converting to hex.
1701 */
1702 #if 1
1703 buf[0] = '\0';
1704 for (i=0; i<key->mv_size; i++)
1705 ptr += sprintf(ptr, "%02x", *c++);
1706 #else
1707 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1708 #endif
1709 return buf;
1710 }
1711
1712 static const char *
mdb_leafnode_type(MDB_node * n)1713 mdb_leafnode_type(MDB_node *n)
1714 {
1715 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1716 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1717 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1718 }
1719
1720 /** Display all the keys in the page. */
1721 void
mdb_page_list(MDB_page * mp)1722 mdb_page_list(MDB_page *mp)
1723 {
1724 pgno_t pgno = mdb_dbg_pgno(mp);
1725 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1726 MDB_node *node;
1727 unsigned int i, nkeys, nsize, total = 0;
1728 MDB_val key;
1729 DKBUF;
1730
1731 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1732 case P_BRANCH: type = "Branch page"; break;
1733 case P_LEAF: type = "Leaf page"; break;
1734 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1735 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1736 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1737 case P_OVERFLOW:
1738 fprintf(stderr, "Overflow page %"Yu" pages %u%s\n",
1739 pgno, mp->mp_pages, state);
1740 return;
1741 case P_META:
1742 fprintf(stderr, "Meta-page %"Yu" txnid %"Yu"\n",
1743 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1744 return;
1745 default:
1746 fprintf(stderr, "Bad page %"Yu" flags 0x%u\n", pgno, mp->mp_flags);
1747 return;
1748 }
1749
1750 nkeys = NUMKEYS(mp);
1751 fprintf(stderr, "%s %"Yu" numkeys %d%s\n", type, pgno, nkeys, state);
1752
1753 for (i=0; i<nkeys; i++) {
1754 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1755 key.mv_size = nsize = mp->mp_pad;
1756 key.mv_data = LEAF2KEY(mp, i, nsize);
1757 total += nsize;
1758 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1759 continue;
1760 }
1761 node = NODEPTR(mp, i);
1762 key.mv_size = node->mn_ksize;
1763 key.mv_data = node->mn_data;
1764 nsize = NODESIZE + key.mv_size;
1765 if (IS_BRANCH(mp)) {
1766 fprintf(stderr, "key %d: page %"Yu", %s\n", i, NODEPGNO(node),
1767 DKEY(&key));
1768 total += nsize;
1769 } else {
1770 if (F_ISSET(node->mn_flags, F_BIGDATA))
1771 nsize += sizeof(pgno_t);
1772 else
1773 nsize += NODEDSZ(node);
1774 total += nsize;
1775 nsize += sizeof(indx_t);
1776 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1777 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1778 }
1779 total = EVEN(total);
1780 }
1781 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1782 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1783 }
1784
1785 void
mdb_cursor_chk(MDB_cursor * mc)1786 mdb_cursor_chk(MDB_cursor *mc)
1787 {
1788 unsigned int i;
1789 MDB_node *node;
1790 MDB_page *mp;
1791
1792 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return;
1793 for (i=0; i<mc->mc_top; i++) {
1794 mp = mc->mc_pg[i];
1795 node = NODEPTR(mp, mc->mc_ki[i]);
1796 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1797 printf("oops!\n");
1798 }
1799 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1800 printf("ack!\n");
1801 if (mc->mc_xcursor && (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
1802 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
1803 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) &&
1804 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) {
1805 printf("blah!\n");
1806 }
1807 }
1808 }
1809 #endif
1810
1811 #if (MDB_DEBUG) > 2
1812 /** Count all the pages in each DB and in the freelist
1813 * and make sure it matches the actual number of pages
1814 * being used.
1815 * All named DBs must be open for a correct count.
1816 */
mdb_audit(MDB_txn * txn)1817 static void mdb_audit(MDB_txn *txn)
1818 {
1819 MDB_cursor mc;
1820 MDB_val key, data;
1821 MDB_ID freecount, count;
1822 MDB_dbi i;
1823 int rc;
1824
1825 freecount = 0;
1826 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1827 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1828 freecount += *(MDB_ID *)data.mv_data;
1829 mdb_tassert(txn, rc == MDB_NOTFOUND);
1830
1831 count = 0;
1832 for (i = 0; i<txn->mt_numdbs; i++) {
1833 MDB_xcursor mx;
1834 if (!(txn->mt_dbflags[i] & DB_VALID))
1835 continue;
1836 mdb_cursor_init(&mc, txn, i, &mx);
1837 if (txn->mt_dbs[i].md_root == P_INVALID)
1838 continue;
1839 count += txn->mt_dbs[i].md_branch_pages +
1840 txn->mt_dbs[i].md_leaf_pages +
1841 txn->mt_dbs[i].md_overflow_pages;
1842 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1843 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1844 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1845 unsigned j;
1846 MDB_page *mp;
1847 mp = mc.mc_pg[mc.mc_top];
1848 for (j=0; j<NUMKEYS(mp); j++) {
1849 MDB_node *leaf = NODEPTR(mp, j);
1850 if (leaf->mn_flags & F_SUBDATA) {
1851 MDB_db db;
1852 memcpy(&db, NODEDATA(leaf), sizeof(db));
1853 count += db.md_branch_pages + db.md_leaf_pages +
1854 db.md_overflow_pages;
1855 }
1856 }
1857 }
1858 mdb_tassert(txn, rc == MDB_NOTFOUND);
1859 }
1860 }
1861 if (freecount + count + NUM_METAS != txn->mt_next_pgno) {
1862 fprintf(stderr, "audit: %"Yu" freecount: %"Yu" count: %"Yu" total: %"Yu" next_pgno: %"Yu"\n",
1863 txn->mt_txnid, freecount, count+NUM_METAS,
1864 freecount+count+NUM_METAS, txn->mt_next_pgno);
1865 }
1866 }
1867 #endif
1868
1869 int
mdb_cmp(MDB_txn * txn,MDB_dbi dbi,const MDB_val * a,const MDB_val * b)1870 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1871 {
1872 return txn->mt_dbxs[dbi].md_cmp(a, b);
1873 }
1874
1875 int
mdb_dcmp(MDB_txn * txn,MDB_dbi dbi,const MDB_val * a,const MDB_val * b)1876 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1877 {
1878 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1879 if (NEED_CMP_CLONG(dcmp, a->mv_size))
1880 dcmp = mdb_cmp_clong;
1881 return dcmp(a, b);
1882 }
1883
1884 /** Allocate memory for a page.
1885 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1886 */
1887 static MDB_page *
mdb_page_malloc(MDB_txn * txn,unsigned num)1888 mdb_page_malloc(MDB_txn *txn, unsigned num)
1889 {
1890 MDB_env *env = txn->mt_env;
1891 MDB_page *ret = env->me_dpages;
1892 size_t psize = env->me_psize, sz = psize, off;
1893 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1894 * For a single page alloc, we init everything after the page header.
1895 * For multi-page, we init the final page; if the caller needed that
1896 * many pages they will be filling in at least up to the last page.
1897 */
1898 if (num == 1) {
1899 if (ret) {
1900 VGMEMP_ALLOC(env, ret, sz);
1901 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1902 env->me_dpages = ret->mp_next;
1903 return ret;
1904 }
1905 psize -= off = PAGEHDRSZ;
1906 } else {
1907 sz *= num;
1908 off = sz - psize;
1909 }
1910 if ((ret = malloc(sz)) != NULL) {
1911 VGMEMP_ALLOC(env, ret, sz);
1912 if (!(env->me_flags & MDB_NOMEMINIT)) {
1913 memset((char *)ret + off, 0, psize);
1914 ret->mp_pad = 0;
1915 }
1916 } else {
1917 txn->mt_flags |= MDB_TXN_ERROR;
1918 }
1919 return ret;
1920 }
1921 /** Free a single page.
1922 * Saves single pages to a list, for future reuse.
1923 * (This is not used for multi-page overflow pages.)
1924 */
1925 static void
mdb_page_free(MDB_env * env,MDB_page * mp)1926 mdb_page_free(MDB_env *env, MDB_page *mp)
1927 {
1928 mp->mp_next = env->me_dpages;
1929 VGMEMP_FREE(env, mp);
1930 env->me_dpages = mp;
1931 }
1932
1933 /** Free a dirty page */
1934 static void
mdb_dpage_free(MDB_env * env,MDB_page * dp)1935 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1936 {
1937 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1938 mdb_page_free(env, dp);
1939 } else {
1940 /* large pages just get freed directly */
1941 VGMEMP_FREE(env, dp);
1942 free(dp);
1943 }
1944 }
1945
1946 /** Return all dirty pages to dpage list */
1947 static void
mdb_dlist_free(MDB_txn * txn)1948 mdb_dlist_free(MDB_txn *txn)
1949 {
1950 MDB_env *env = txn->mt_env;
1951 MDB_ID2L dl = txn->mt_u.dirty_list;
1952 unsigned i, n = dl[0].mid;
1953
1954 for (i = 1; i <= n; i++) {
1955 mdb_dpage_free(env, dl[i].mptr);
1956 }
1957 dl[0].mid = 0;
1958 }
1959
1960 #ifdef MDB_VL32
1961 static void
mdb_page_unref(MDB_txn * txn,MDB_page * mp)1962 mdb_page_unref(MDB_txn *txn, MDB_page *mp)
1963 {
1964 pgno_t pgno;
1965 MDB_ID3L tl = txn->mt_rpages;
1966 unsigned x, rem;
1967 if (mp->mp_flags & (P_SUBP|P_DIRTY))
1968 return;
1969 rem = mp->mp_pgno & (MDB_RPAGE_CHUNK-1);
1970 pgno = mp->mp_pgno ^ rem;
1971 x = mdb_mid3l_search(tl, pgno);
1972 if (x != tl[0].mid && tl[x+1].mid == mp->mp_pgno)
1973 x++;
1974 if (tl[x].mref)
1975 tl[x].mref--;
1976 }
1977 #define MDB_PAGE_UNREF(txn, mp) mdb_page_unref(txn, mp)
1978
1979 static void
mdb_cursor_unref(MDB_cursor * mc)1980 mdb_cursor_unref(MDB_cursor *mc)
1981 {
1982 int i;
1983 if (!mc->mc_snum || !mc->mc_pg[0] || IS_SUBP(mc->mc_pg[0]))
1984 return;
1985 for (i=0; i<mc->mc_snum; i++)
1986 mdb_page_unref(mc->mc_txn, mc->mc_pg[i]);
1987 if (mc->mc_ovpg) {
1988 mdb_page_unref(mc->mc_txn, mc->mc_ovpg);
1989 mc->mc_ovpg = 0;
1990 }
1991 mc->mc_snum = mc->mc_top = 0;
1992 mc->mc_pg[0] = NULL;
1993 mc->mc_flags &= ~C_INITIALIZED;
1994 }
1995 #define MDB_CURSOR_UNREF(mc, force) \
1996 (((force) || ((mc)->mc_flags & C_INITIALIZED)) \
1997 ? mdb_cursor_unref(mc) \
1998 : (void)0)
1999
2000 #else
2001 #define MDB_PAGE_UNREF(txn, mp)
2002 #define MDB_CURSOR_UNREF(mc, force) ((void)0)
2003 #endif /* MDB_VL32 */
2004
2005 /** Loosen or free a single page.
2006 * Saves single pages to a list for future reuse
2007 * in this same txn. It has been pulled from the freeDB
2008 * and already resides on the dirty list, but has been
2009 * deleted. Use these pages first before pulling again
2010 * from the freeDB.
2011 *
2012 * If the page wasn't dirtied in this txn, just add it
2013 * to this txn's free list.
2014 */
2015 static int
mdb_page_loose(MDB_cursor * mc,MDB_page * mp)2016 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
2017 {
2018 int loose = 0;
2019 pgno_t pgno = mp->mp_pgno;
2020 MDB_txn *txn = mc->mc_txn;
2021
2022 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
2023 if (txn->mt_parent) {
2024 MDB_ID2 *dl = txn->mt_u.dirty_list;
2025 /* If txn has a parent, make sure the page is in our
2026 * dirty list.
2027 */
2028 if (dl[0].mid) {
2029 unsigned x = mdb_mid2l_search(dl, pgno);
2030 if (x <= dl[0].mid && dl[x].mid == pgno) {
2031 if (mp != dl[x].mptr) { /* bad cursor? */
2032 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2033 txn->mt_flags |= MDB_TXN_ERROR;
2034 return MDB_PROBLEM;
2035 }
2036 /* ok, it's ours */
2037 loose = 1;
2038 }
2039 }
2040 } else {
2041 /* no parent txn, so it's just ours */
2042 loose = 1;
2043 }
2044 }
2045 if (loose) {
2046 DPRINTF(("loosen db %d page %"Yu, DDBI(mc), mp->mp_pgno));
2047 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
2048 txn->mt_loose_pgs = mp;
2049 txn->mt_loose_count++;
2050 mp->mp_flags |= P_LOOSE;
2051 } else {
2052 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
2053 if (rc)
2054 return rc;
2055 }
2056
2057 return MDB_SUCCESS;
2058 }
2059
2060 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
2061 * @param[in] mc A cursor handle for the current operation.
2062 * @param[in] pflags Flags of the pages to update:
2063 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
2064 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
2065 * @return 0 on success, non-zero on failure.
2066 */
2067 static int
mdb_pages_xkeep(MDB_cursor * mc,unsigned pflags,int all)2068 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
2069 {
2070 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
2071 MDB_txn *txn = mc->mc_txn;
2072 MDB_cursor *m3, *m0 = mc;
2073 MDB_xcursor *mx;
2074 MDB_page *dp, *mp;
2075 MDB_node *leaf;
2076 unsigned i, j;
2077 int rc = MDB_SUCCESS, level;
2078
2079 /* Mark pages seen by cursors */
2080 if (mc->mc_flags & C_UNTRACK)
2081 mc = NULL; /* will find mc in mt_cursors */
2082 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
2083 for (; mc; mc=mc->mc_next) {
2084 if (!(mc->mc_flags & C_INITIALIZED))
2085 continue;
2086 for (m3 = mc;; m3 = &mx->mx_cursor) {
2087 mp = NULL;
2088 for (j=0; j<m3->mc_snum; j++) {
2089 mp = m3->mc_pg[j];
2090 if ((mp->mp_flags & Mask) == pflags)
2091 mp->mp_flags ^= P_KEEP;
2092 }
2093 mx = m3->mc_xcursor;
2094 /* Proceed to mx if it is at a sub-database */
2095 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
2096 break;
2097 if (! (mp && (mp->mp_flags & P_LEAF)))
2098 break;
2099 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
2100 if (!(leaf->mn_flags & F_SUBDATA))
2101 break;
2102 }
2103 }
2104 if (i == 0)
2105 break;
2106 }
2107
2108 if (all) {
2109 /* Mark dirty root pages */
2110 for (i=0; i<txn->mt_numdbs; i++) {
2111 if (txn->mt_dbflags[i] & DB_DIRTY) {
2112 pgno_t pgno = txn->mt_dbs[i].md_root;
2113 if (pgno == P_INVALID)
2114 continue;
2115 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS)
2116 break;
2117 if ((dp->mp_flags & Mask) == pflags && level <= 1)
2118 dp->mp_flags ^= P_KEEP;
2119 }
2120 }
2121 }
2122
2123 return rc;
2124 }
2125
2126 static int mdb_page_flush(MDB_txn *txn, int keep);
2127
2128 /** Spill pages from the dirty list back to disk.
2129 * This is intended to prevent running into #MDB_TXN_FULL situations,
2130 * but note that they may still occur in a few cases:
2131 * 1) our estimate of the txn size could be too small. Currently this
2132 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
2133 * 2) child txns may run out of space if their parents dirtied a
2134 * lot of pages and never spilled them. TODO: we probably should do
2135 * a preemptive spill during #mdb_txn_begin() of a child txn, if
2136 * the parent's dirty_room is below a given threshold.
2137 *
2138 * Otherwise, if not using nested txns, it is expected that apps will
2139 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
2140 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
2141 * If the txn never references them again, they can be left alone.
2142 * If the txn only reads them, they can be used without any fuss.
2143 * If the txn writes them again, they can be dirtied immediately without
2144 * going thru all of the work of #mdb_page_touch(). Such references are
2145 * handled by #mdb_page_unspill().
2146 *
2147 * Also note, we never spill DB root pages, nor pages of active cursors,
2148 * because we'll need these back again soon anyway. And in nested txns,
2149 * we can't spill a page in a child txn if it was already spilled in a
2150 * parent txn. That would alter the parent txns' data even though
2151 * the child hasn't committed yet, and we'd have no way to undo it if
2152 * the child aborted.
2153 *
2154 * @param[in] m0 cursor A cursor handle identifying the transaction and
2155 * database for which we are checking space.
2156 * @param[in] key For a put operation, the key being stored.
2157 * @param[in] data For a put operation, the data being stored.
2158 * @return 0 on success, non-zero on failure.
2159 */
2160 static int
mdb_page_spill(MDB_cursor * m0,MDB_val * key,MDB_val * data)2161 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
2162 {
2163 MDB_txn *txn = m0->mc_txn;
2164 MDB_page *dp;
2165 MDB_ID2L dl = txn->mt_u.dirty_list;
2166 unsigned int i, j, need;
2167 int rc;
2168
2169 if (m0->mc_flags & C_SUB)
2170 return MDB_SUCCESS;
2171
2172 /* Estimate how much space this op will take */
2173 i = m0->mc_db->md_depth;
2174 /* Named DBs also dirty the main DB */
2175 if (m0->mc_dbi >= CORE_DBS)
2176 i += txn->mt_dbs[MAIN_DBI].md_depth;
2177 /* For puts, roughly factor in the key+data size */
2178 if (key)
2179 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
2180 i += i; /* double it for good measure */
2181 need = i;
2182
2183 if (txn->mt_dirty_room > i)
2184 return MDB_SUCCESS;
2185
2186 if (!txn->mt_spill_pgs) {
2187 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
2188 if (!txn->mt_spill_pgs)
2189 return ENOMEM;
2190 } else {
2191 /* purge deleted slots */
2192 MDB_IDL sl = txn->mt_spill_pgs;
2193 unsigned int num = sl[0];
2194 j=0;
2195 for (i=1; i<=num; i++) {
2196 if (!(sl[i] & 1))
2197 sl[++j] = sl[i];
2198 }
2199 sl[0] = j;
2200 }
2201
2202 /* Preserve pages which may soon be dirtied again */
2203 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
2204 goto done;
2205
2206 /* Less aggressive spill - we originally spilled the entire dirty list,
2207 * with a few exceptions for cursor pages and DB root pages. But this
2208 * turns out to be a lot of wasted effort because in a large txn many
2209 * of those pages will need to be used again. So now we spill only 1/8th
2210 * of the dirty pages. Testing revealed this to be a good tradeoff,
2211 * better than 1/2, 1/4, or 1/10.
2212 */
2213 if (need < MDB_IDL_UM_MAX / 8)
2214 need = MDB_IDL_UM_MAX / 8;
2215
2216 /* Save the page IDs of all the pages we're flushing */
2217 /* flush from the tail forward, this saves a lot of shifting later on. */
2218 for (i=dl[0].mid; i && need; i--) {
2219 MDB_ID pn = dl[i].mid << 1;
2220 dp = dl[i].mptr;
2221 if (dp->mp_flags & (P_LOOSE|P_KEEP))
2222 continue;
2223 /* Can't spill twice, make sure it's not already in a parent's
2224 * spill list.
2225 */
2226 if (txn->mt_parent) {
2227 MDB_txn *tx2;
2228 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
2229 if (tx2->mt_spill_pgs) {
2230 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
2231 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
2232 dp->mp_flags |= P_KEEP;
2233 break;
2234 }
2235 }
2236 }
2237 if (tx2)
2238 continue;
2239 }
2240 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
2241 goto done;
2242 need--;
2243 }
2244 mdb_midl_sort(txn->mt_spill_pgs);
2245
2246 /* Flush the spilled part of dirty list */
2247 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
2248 goto done;
2249
2250 /* Reset any dirty pages we kept that page_flush didn't see */
2251 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
2252
2253 done:
2254 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
2255 return rc;
2256 }
2257
2258 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
2259 static txnid_t
mdb_find_oldest(MDB_txn * txn)2260 mdb_find_oldest(MDB_txn *txn)
2261 {
2262 int i;
2263 txnid_t mr, oldest = txn->mt_txnid - 1;
2264 if (txn->mt_env->me_txns) {
2265 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
2266 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
2267 if (r[i].mr_pid) {
2268 mr = r[i].mr_txnid;
2269 if (oldest > mr)
2270 oldest = mr;
2271 }
2272 }
2273 }
2274 return oldest;
2275 }
2276
2277 /** Add a page to the txn's dirty list */
2278 static void
mdb_page_dirty(MDB_txn * txn,MDB_page * mp)2279 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
2280 {
2281 MDB_ID2 mid;
2282 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
2283
2284 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
2285 insert = mdb_mid2l_append;
2286 } else {
2287 insert = mdb_mid2l_insert;
2288 }
2289 mid.mid = mp->mp_pgno;
2290 mid.mptr = mp;
2291 rc = insert(txn->mt_u.dirty_list, &mid);
2292 mdb_tassert(txn, rc == 0);
2293 txn->mt_dirty_room--;
2294 }
2295
2296 /** Allocate page numbers and memory for writing. Maintain me_pglast,
2297 * me_pghead and mt_next_pgno.
2298 *
2299 * If there are free pages available from older transactions, they
2300 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
2301 * Do not modify the freedB, just merge freeDB records into me_pghead[]
2302 * and move me_pglast to say which records were consumed. Only this
2303 * function can create me_pghead and move me_pglast/mt_next_pgno.
2304 * When #MDB_DEVEL & 2, it is not affected by #mdb_freelist_save(): it
2305 * then uses the transaction's original snapshot of the freeDB.
2306 * @param[in] mc cursor A cursor handle identifying the transaction and
2307 * database for which we are allocating.
2308 * @param[in] num the number of pages to allocate.
2309 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
2310 * will always be satisfied by a single contiguous chunk of memory.
2311 * @return 0 on success, non-zero on failure.
2312 */
2313 static int
mdb_page_alloc(MDB_cursor * mc,int num,MDB_page ** mp)2314 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
2315 {
2316 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
2317 /* Get at most <Max_retries> more freeDB records once me_pghead
2318 * has enough pages. If not enough, use new pages from the map.
2319 * If <Paranoid> and mc is updating the freeDB, only get new
2320 * records if me_pghead is empty. Then the freelist cannot play
2321 * catch-up with itself by growing while trying to save it.
2322 */
2323 enum { Paranoid = 1, Max_retries = 500 };
2324 #else
2325 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
2326 #endif
2327 int rc, retry = num * 60;
2328 MDB_txn *txn = mc->mc_txn;
2329 MDB_env *env = txn->mt_env;
2330 pgno_t pgno, *mop = env->me_pghead;
2331 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
2332 MDB_page *np;
2333 txnid_t oldest = 0, last;
2334 MDB_cursor_op op;
2335 MDB_cursor m2;
2336 int found_old = 0;
2337
2338 /* If there are any loose pages, just use them */
2339 if (num == 1 && txn->mt_loose_pgs) {
2340 np = txn->mt_loose_pgs;
2341 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2342 txn->mt_loose_count--;
2343 DPRINTF(("db %d use loose page %"Yu, DDBI(mc), np->mp_pgno));
2344 *mp = np;
2345 return MDB_SUCCESS;
2346 }
2347
2348 *mp = NULL;
2349
2350 /* If our dirty list is already full, we can't do anything */
2351 if (txn->mt_dirty_room == 0) {
2352 rc = MDB_TXN_FULL;
2353 goto fail;
2354 }
2355
2356 for (op = MDB_FIRST;; op = MDB_NEXT) {
2357 MDB_val key, data;
2358 MDB_node *leaf;
2359 pgno_t *idl;
2360
2361 /* Seek a big enough contiguous page range. Prefer
2362 * pages at the tail, just truncating the list.
2363 */
2364 if (mop_len > n2) {
2365 i = mop_len;
2366 do {
2367 pgno = mop[i];
2368 if (mop[i-n2] == pgno+n2)
2369 goto search_done;
2370 } while (--i > n2);
2371 if (--retry < 0)
2372 break;
2373 }
2374
2375 if (op == MDB_FIRST) { /* 1st iteration */
2376 /* Prepare to fetch more and coalesce */
2377 last = env->me_pglast;
2378 oldest = env->me_pgoldest;
2379 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2380 #if (MDB_DEVEL) & 2 /* "& 2" so MDB_DEVEL=1 won't hide bugs breaking freeDB */
2381 /* Use original snapshot. TODO: Should need less care in code
2382 * which modifies the database. Maybe we can delete some code?
2383 */
2384 m2.mc_flags |= C_ORIG_RDONLY;
2385 m2.mc_db = &env->me_metas[(txn->mt_txnid-1) & 1]->mm_dbs[FREE_DBI];
2386 m2.mc_dbflag = (unsigned char *)""; /* probably unnecessary */
2387 #endif
2388 if (last) {
2389 op = MDB_SET_RANGE;
2390 key.mv_data = &last; /* will look up last+1 */
2391 key.mv_size = sizeof(last);
2392 }
2393 if (Paranoid && mc->mc_dbi == FREE_DBI)
2394 retry = -1;
2395 }
2396 if (Paranoid && retry < 0 && mop_len)
2397 break;
2398
2399 last++;
2400 /* Do not fetch more if the record will be too recent */
2401 if (oldest <= last) {
2402 if (!found_old) {
2403 oldest = mdb_find_oldest(txn);
2404 env->me_pgoldest = oldest;
2405 found_old = 1;
2406 }
2407 if (oldest <= last)
2408 break;
2409 }
2410 rc = mdb_cursor_get(&m2, &key, NULL, op);
2411 if (rc) {
2412 if (rc == MDB_NOTFOUND)
2413 break;
2414 goto fail;
2415 }
2416 last = *(txnid_t*)key.mv_data;
2417 if (oldest <= last) {
2418 if (!found_old) {
2419 oldest = mdb_find_oldest(txn);
2420 env->me_pgoldest = oldest;
2421 found_old = 1;
2422 }
2423 if (oldest <= last)
2424 break;
2425 }
2426 np = m2.mc_pg[m2.mc_top];
2427 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2428 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
2429 return rc;
2430
2431 idl = (MDB_ID *) data.mv_data;
2432 i = idl[0];
2433 if (!mop) {
2434 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2435 rc = ENOMEM;
2436 goto fail;
2437 }
2438 } else {
2439 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2440 goto fail;
2441 mop = env->me_pghead;
2442 }
2443 env->me_pglast = last;
2444 #if (MDB_DEBUG) > 1
2445 DPRINTF(("IDL read txn %"Yu" root %"Yu" num %u",
2446 last, txn->mt_dbs[FREE_DBI].md_root, i));
2447 for (j = i; j; j--)
2448 DPRINTF(("IDL %"Yu, idl[j]));
2449 #endif
2450 /* Merge in descending sorted order */
2451 mdb_midl_xmerge(mop, idl);
2452 mop_len = mop[0];
2453 }
2454
2455 /* Use new pages from the map when nothing suitable in the freeDB */
2456 i = 0;
2457 pgno = txn->mt_next_pgno;
2458 if (pgno + num >= env->me_maxpg) {
2459 DPUTS("DB size maxed out");
2460 rc = MDB_MAP_FULL;
2461 goto fail;
2462 }
2463 #if defined(_WIN32) && !defined(MDB_VL32)
2464 if (!(env->me_flags & MDB_RDONLY)) {
2465 void *p;
2466 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
2467 p = VirtualAlloc(p, env->me_psize * num, MEM_COMMIT,
2468 (env->me_flags & MDB_WRITEMAP) ? PAGE_READWRITE:
2469 PAGE_READONLY);
2470 if (!p) {
2471 DPUTS("VirtualAlloc failed");
2472 rc = ErrCode();
2473 goto fail;
2474 }
2475 }
2476 #endif
2477
2478 search_done:
2479 if (env->me_flags & MDB_WRITEMAP) {
2480 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2481 } else {
2482 if (!(np = mdb_page_malloc(txn, num))) {
2483 rc = ENOMEM;
2484 goto fail;
2485 }
2486 }
2487 if (i) {
2488 mop[0] = mop_len -= num;
2489 /* Move any stragglers down */
2490 for (j = i-num; j < mop_len; )
2491 mop[++j] = mop[++i];
2492 } else {
2493 txn->mt_next_pgno = pgno + num;
2494 }
2495 np->mp_pgno = pgno;
2496 mdb_page_dirty(txn, np);
2497 *mp = np;
2498
2499 return MDB_SUCCESS;
2500
2501 fail:
2502 txn->mt_flags |= MDB_TXN_ERROR;
2503 return rc;
2504 }
2505
2506 /** Copy the used portions of a non-overflow page.
2507 * @param[in] dst page to copy into
2508 * @param[in] src page to copy from
2509 * @param[in] psize size of a page
2510 */
2511 static void
mdb_page_copy(MDB_page * dst,MDB_page * src,unsigned int psize)2512 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2513 {
2514 enum { Align = sizeof(pgno_t) };
2515 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2516
2517 /* If page isn't full, just copy the used portion. Adjust
2518 * alignment so memcpy may copy words instead of bytes.
2519 */
2520 if ((unused &= -Align) && !IS_LEAF2(src)) {
2521 upper = (upper + PAGEBASE) & -Align;
2522 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2523 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2524 psize - upper);
2525 } else {
2526 memcpy(dst, src, psize - unused);
2527 }
2528 }
2529
2530 /** Pull a page off the txn's spill list, if present.
2531 * If a page being referenced was spilled to disk in this txn, bring
2532 * it back and make it dirty/writable again.
2533 * @param[in] txn the transaction handle.
2534 * @param[in] mp the page being referenced. It must not be dirty.
2535 * @param[out] ret the writable page, if any. ret is unchanged if
2536 * mp wasn't spilled.
2537 */
2538 static int
mdb_page_unspill(MDB_txn * txn,MDB_page * mp,MDB_page ** ret)2539 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2540 {
2541 MDB_env *env = txn->mt_env;
2542 const MDB_txn *tx2;
2543 unsigned x;
2544 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2545
2546 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2547 if (!tx2->mt_spill_pgs)
2548 continue;
2549 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2550 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2551 MDB_page *np;
2552 int num;
2553 if (txn->mt_dirty_room == 0)
2554 return MDB_TXN_FULL;
2555 if (IS_OVERFLOW(mp))
2556 num = mp->mp_pages;
2557 else
2558 num = 1;
2559 if (env->me_flags & MDB_WRITEMAP) {
2560 np = mp;
2561 } else {
2562 np = mdb_page_malloc(txn, num);
2563 if (!np)
2564 return ENOMEM;
2565 if (num > 1)
2566 memcpy(np, mp, num * env->me_psize);
2567 else
2568 mdb_page_copy(np, mp, env->me_psize);
2569 }
2570 if (tx2 == txn) {
2571 /* If in current txn, this page is no longer spilled.
2572 * If it happens to be the last page, truncate the spill list.
2573 * Otherwise mark it as deleted by setting the LSB.
2574 */
2575 if (x == txn->mt_spill_pgs[0])
2576 txn->mt_spill_pgs[0]--;
2577 else
2578 txn->mt_spill_pgs[x] |= 1;
2579 } /* otherwise, if belonging to a parent txn, the
2580 * page remains spilled until child commits
2581 */
2582
2583 mdb_page_dirty(txn, np);
2584 np->mp_flags |= P_DIRTY;
2585 *ret = np;
2586 break;
2587 }
2588 }
2589 return MDB_SUCCESS;
2590 }
2591
2592 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2593 * @param[in] mc cursor pointing to the page to be touched
2594 * @return 0 on success, non-zero on failure.
2595 */
2596 static int
mdb_page_touch(MDB_cursor * mc)2597 mdb_page_touch(MDB_cursor *mc)
2598 {
2599 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2600 MDB_txn *txn = mc->mc_txn;
2601 MDB_cursor *m2, *m3;
2602 pgno_t pgno;
2603 int rc;
2604
2605 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2606 if (txn->mt_flags & MDB_TXN_SPILLS) {
2607 np = NULL;
2608 rc = mdb_page_unspill(txn, mp, &np);
2609 if (rc)
2610 goto fail;
2611 if (np)
2612 goto done;
2613 }
2614 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2615 (rc = mdb_page_alloc(mc, 1, &np)))
2616 goto fail;
2617 pgno = np->mp_pgno;
2618 DPRINTF(("touched db %d page %"Yu" -> %"Yu, DDBI(mc),
2619 mp->mp_pgno, pgno));
2620 mdb_cassert(mc, mp->mp_pgno != pgno);
2621 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2622 /* Update the parent page, if any, to point to the new page */
2623 if (mc->mc_top) {
2624 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2625 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2626 SETPGNO(node, pgno);
2627 } else {
2628 mc->mc_db->md_root = pgno;
2629 }
2630 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2631 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2632 pgno = mp->mp_pgno;
2633 /* If txn has a parent, make sure the page is in our
2634 * dirty list.
2635 */
2636 if (dl[0].mid) {
2637 unsigned x = mdb_mid2l_search(dl, pgno);
2638 if (x <= dl[0].mid && dl[x].mid == pgno) {
2639 if (mp != dl[x].mptr) { /* bad cursor? */
2640 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2641 txn->mt_flags |= MDB_TXN_ERROR;
2642 return MDB_PROBLEM;
2643 }
2644 return 0;
2645 }
2646 }
2647 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2648 /* No - copy it */
2649 np = mdb_page_malloc(txn, 1);
2650 if (!np)
2651 return ENOMEM;
2652 mid.mid = pgno;
2653 mid.mptr = np;
2654 rc = mdb_mid2l_insert(dl, &mid);
2655 mdb_cassert(mc, rc == 0);
2656 } else {
2657 return 0;
2658 }
2659
2660 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2661 np->mp_pgno = pgno;
2662 np->mp_flags |= P_DIRTY;
2663
2664 done:
2665 /* Adjust cursors pointing to mp */
2666 mc->mc_pg[mc->mc_top] = np;
2667 m2 = txn->mt_cursors[mc->mc_dbi];
2668 if (mc->mc_flags & C_SUB) {
2669 for (; m2; m2=m2->mc_next) {
2670 m3 = &m2->mc_xcursor->mx_cursor;
2671 if (m3->mc_snum < mc->mc_snum) continue;
2672 if (m3->mc_pg[mc->mc_top] == mp)
2673 m3->mc_pg[mc->mc_top] = np;
2674 }
2675 } else {
2676 for (; m2; m2=m2->mc_next) {
2677 if (m2->mc_snum < mc->mc_snum) continue;
2678 if (m2 == mc) continue;
2679 if (m2->mc_pg[mc->mc_top] == mp) {
2680 m2->mc_pg[mc->mc_top] = np;
2681 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2682 IS_LEAF(np) &&
2683 (m2->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
2684 {
2685 MDB_node *leaf = NODEPTR(np, m2->mc_ki[mc->mc_top]);
2686 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
2687 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2688 }
2689 }
2690 }
2691 }
2692 MDB_PAGE_UNREF(mc->mc_txn, mp);
2693 return 0;
2694
2695 fail:
2696 txn->mt_flags |= MDB_TXN_ERROR;
2697 return rc;
2698 }
2699
2700 int
mdb_env_sync0(MDB_env * env,int force,pgno_t numpgs)2701 mdb_env_sync0(MDB_env *env, int force, pgno_t numpgs)
2702 {
2703 int rc = 0;
2704 if (env->me_flags & MDB_RDONLY)
2705 return EACCES;
2706 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2707 if (env->me_flags & MDB_WRITEMAP) {
2708 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2709 ? MS_ASYNC : MS_SYNC;
2710 if (MDB_MSYNC(env->me_map, env->me_psize * numpgs, flags))
2711 rc = ErrCode();
2712 #ifdef _WIN32
2713 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2714 rc = ErrCode();
2715 #endif
2716 } else {
2717 #ifdef BROKEN_FDATASYNC
2718 if (env->me_flags & MDB_FSYNCONLY) {
2719 if (fsync(env->me_fd))
2720 rc = ErrCode();
2721 } else
2722 #endif
2723 if (MDB_FDATASYNC(env->me_fd))
2724 rc = ErrCode();
2725 }
2726 }
2727 return rc;
2728 }
2729
2730 int
mdb_env_sync(MDB_env * env,int force)2731 mdb_env_sync(MDB_env *env, int force)
2732 {
2733 MDB_meta *m = mdb_env_pick_meta(env);
2734 return mdb_env_sync0(env, force, m->mm_last_pg+1);
2735 }
2736
2737 /** Back up parent txn's cursors, then grab the originals for tracking */
2738 static int
mdb_cursor_shadow(MDB_txn * src,MDB_txn * dst)2739 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2740 {
2741 MDB_cursor *mc, *bk;
2742 MDB_xcursor *mx;
2743 size_t size;
2744 int i;
2745
2746 for (i = src->mt_numdbs; --i >= 0; ) {
2747 if ((mc = src->mt_cursors[i]) != NULL) {
2748 size = sizeof(MDB_cursor);
2749 if (mc->mc_xcursor)
2750 size += sizeof(MDB_xcursor);
2751 for (; mc; mc = bk->mc_next) {
2752 bk = malloc(size);
2753 if (!bk)
2754 return ENOMEM;
2755 *bk = *mc;
2756 mc->mc_backup = bk;
2757 mc->mc_db = &dst->mt_dbs[i];
2758 /* Kill pointers into src to reduce abuse: The
2759 * user may not use mc until dst ends. But we need a valid
2760 * txn pointer here for cursor fixups to keep working.
2761 */
2762 mc->mc_txn = dst;
2763 mc->mc_dbflag = &dst->mt_dbflags[i];
2764 if ((mx = mc->mc_xcursor) != NULL) {
2765 *(MDB_xcursor *)(bk+1) = *mx;
2766 mx->mx_cursor.mc_txn = dst;
2767 }
2768 mc->mc_next = dst->mt_cursors[i];
2769 dst->mt_cursors[i] = mc;
2770 }
2771 }
2772 }
2773 return MDB_SUCCESS;
2774 }
2775
2776 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2777 * @param[in] txn the transaction handle.
2778 * @param[in] merge true to keep changes to parent cursors, false to revert.
2779 * @return 0 on success, non-zero on failure.
2780 */
2781 static void
mdb_cursors_close(MDB_txn * txn,unsigned merge)2782 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2783 {
2784 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2785 MDB_xcursor *mx;
2786 int i;
2787
2788 for (i = txn->mt_numdbs; --i >= 0; ) {
2789 for (mc = cursors[i]; mc; mc = next) {
2790 next = mc->mc_next;
2791 if ((bk = mc->mc_backup) != NULL) {
2792 if (merge) {
2793 /* Commit changes to parent txn */
2794 mc->mc_next = bk->mc_next;
2795 mc->mc_backup = bk->mc_backup;
2796 mc->mc_txn = bk->mc_txn;
2797 mc->mc_db = bk->mc_db;
2798 mc->mc_dbflag = bk->mc_dbflag;
2799 if ((mx = mc->mc_xcursor) != NULL)
2800 mx->mx_cursor.mc_txn = bk->mc_txn;
2801 } else {
2802 /* Abort nested txn */
2803 *mc = *bk;
2804 if ((mx = mc->mc_xcursor) != NULL)
2805 *mx = *(MDB_xcursor *)(bk+1);
2806 }
2807 mc = bk;
2808 }
2809 /* Only malloced cursors are permanently tracked. */
2810 free(mc);
2811 }
2812 cursors[i] = NULL;
2813 }
2814 }
2815
2816 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2817 enum Pidlock_op {
2818 Pidset, Pidcheck
2819 };
2820 #else
2821 enum Pidlock_op {
2822 Pidset = F_SETLK, Pidcheck = F_GETLK
2823 };
2824 #endif
2825
2826 /** Set or check a pid lock. Set returns 0 on success.
2827 * Check returns 0 if the process is certainly dead, nonzero if it may
2828 * be alive (the lock exists or an error happened so we do not know).
2829 *
2830 * On Windows Pidset is a no-op, we merely check for the existence
2831 * of the process with the given pid. On POSIX we use a single byte
2832 * lock on the lockfile, set at an offset equal to the pid.
2833 */
2834 static int
mdb_reader_pid(MDB_env * env,enum Pidlock_op op,MDB_PID_T pid)2835 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2836 {
2837 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2838 int ret = 0;
2839 HANDLE h;
2840 if (op == Pidcheck) {
2841 h = OpenProcess(env->me_pidquery, FALSE, pid);
2842 /* No documented "no such process" code, but other program use this: */
2843 if (!h)
2844 return ErrCode() != ERROR_INVALID_PARAMETER;
2845 /* A process exists until all handles to it close. Has it exited? */
2846 ret = WaitForSingleObject(h, 0) != 0;
2847 CloseHandle(h);
2848 }
2849 return ret;
2850 #else
2851 for (;;) {
2852 int rc;
2853 struct flock lock_info;
2854 memset(&lock_info, 0, sizeof(lock_info));
2855 lock_info.l_type = F_WRLCK;
2856 lock_info.l_whence = SEEK_SET;
2857 lock_info.l_start = pid;
2858 lock_info.l_len = 1;
2859 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2860 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2861 rc = -1;
2862 } else if ((rc = ErrCode()) == EINTR) {
2863 continue;
2864 }
2865 return rc;
2866 }
2867 #endif
2868 }
2869
2870 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2871 * @param[in] txn the transaction handle to initialize
2872 * @return 0 on success, non-zero on failure.
2873 */
2874 static int
mdb_txn_renew0(MDB_txn * txn)2875 mdb_txn_renew0(MDB_txn *txn)
2876 {
2877 MDB_env *env = txn->mt_env;
2878 MDB_txninfo *ti = env->me_txns;
2879 MDB_meta *meta;
2880 unsigned int i, nr, flags = txn->mt_flags;
2881 uint16_t x;
2882 int rc, new_notls = 0;
2883
2884 if ((flags &= MDB_TXN_RDONLY) != 0) {
2885 if (!ti) {
2886 meta = mdb_env_pick_meta(env);
2887 txn->mt_txnid = meta->mm_txnid;
2888 txn->mt_u.reader = NULL;
2889 } else {
2890 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2891 pthread_getspecific(env->me_txkey);
2892 if (r) {
2893 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2894 return MDB_BAD_RSLOT;
2895 } else {
2896 MDB_PID_T pid = env->me_pid;
2897 MDB_THR_T tid = pthread_self();
2898 mdb_mutexref_t rmutex = env->me_rmutex;
2899
2900 if (!env->me_live_reader) {
2901 rc = mdb_reader_pid(env, Pidset, pid);
2902 if (rc)
2903 return rc;
2904 env->me_live_reader = 1;
2905 }
2906
2907 if (LOCK_MUTEX(rc, env, rmutex))
2908 return rc;
2909 nr = ti->mti_numreaders;
2910 for (i=0; i<nr; i++)
2911 if (ti->mti_readers[i].mr_pid == 0)
2912 break;
2913 if (i == env->me_maxreaders) {
2914 UNLOCK_MUTEX(rmutex);
2915 return MDB_READERS_FULL;
2916 }
2917 r = &ti->mti_readers[i];
2918 /* Claim the reader slot, carefully since other code
2919 * uses the reader table un-mutexed: First reset the
2920 * slot, next publish it in mti_numreaders. After
2921 * that, it is safe for mdb_env_close() to touch it.
2922 * When it will be closed, we can finally claim it.
2923 */
2924 r->mr_pid = 0;
2925 r->mr_txnid = (txnid_t)-1;
2926 r->mr_tid = tid;
2927 if (i == nr)
2928 ti->mti_numreaders = ++nr;
2929 env->me_close_readers = nr;
2930 r->mr_pid = pid;
2931 UNLOCK_MUTEX(rmutex);
2932
2933 new_notls = (env->me_flags & MDB_NOTLS);
2934 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2935 r->mr_pid = 0;
2936 return rc;
2937 }
2938 }
2939 do /* LY: Retry on a race, ITS#7970. */
2940 r->mr_txnid = ti->mti_txnid;
2941 while(r->mr_txnid != ti->mti_txnid);
2942 txn->mt_txnid = r->mr_txnid;
2943 txn->mt_u.reader = r;
2944 meta = env->me_metas[txn->mt_txnid & 1];
2945 }
2946
2947 } else {
2948 /* Not yet touching txn == env->me_txn0, it may be active */
2949 if (ti) {
2950 if (LOCK_MUTEX(rc, env, env->me_wmutex))
2951 return rc;
2952 txn->mt_txnid = ti->mti_txnid;
2953 meta = env->me_metas[txn->mt_txnid & 1];
2954 } else {
2955 meta = mdb_env_pick_meta(env);
2956 txn->mt_txnid = meta->mm_txnid;
2957 }
2958 txn->mt_txnid++;
2959 #if MDB_DEBUG
2960 if (txn->mt_txnid == mdb_debug_start)
2961 mdb_debug = 1;
2962 #endif
2963 txn->mt_child = NULL;
2964 txn->mt_loose_pgs = NULL;
2965 txn->mt_loose_count = 0;
2966 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2967 txn->mt_u.dirty_list = env->me_dirty_list;
2968 txn->mt_u.dirty_list[0].mid = 0;
2969 txn->mt_free_pgs = env->me_free_pgs;
2970 txn->mt_free_pgs[0] = 0;
2971 txn->mt_spill_pgs = NULL;
2972 env->me_txn = txn;
2973 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2974 }
2975
2976 /* Copy the DB info and flags */
2977 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
2978
2979 /* Moved to here to avoid a data race in read TXNs */
2980 txn->mt_next_pgno = meta->mm_last_pg+1;
2981 #ifdef MDB_VL32
2982 txn->mt_last_pgno = txn->mt_next_pgno - 1;
2983 #endif
2984
2985 txn->mt_flags = flags;
2986
2987 /* Setup db info */
2988 txn->mt_numdbs = env->me_numdbs;
2989 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
2990 x = env->me_dbflags[i];
2991 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2992 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0;
2993 }
2994 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID;
2995 txn->mt_dbflags[FREE_DBI] = DB_VALID;
2996
2997 if (env->me_flags & MDB_FATAL_ERROR) {
2998 DPUTS("environment had fatal error, must shutdown!");
2999 rc = MDB_PANIC;
3000 } else if (env->me_maxpg < txn->mt_next_pgno) {
3001 rc = MDB_MAP_RESIZED;
3002 } else {
3003 return MDB_SUCCESS;
3004 }
3005 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN);
3006 return rc;
3007 }
3008
3009 int
mdb_txn_renew(MDB_txn * txn)3010 mdb_txn_renew(MDB_txn *txn)
3011 {
3012 int rc;
3013
3014 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED))
3015 return EINVAL;
3016
3017 rc = mdb_txn_renew0(txn);
3018 if (rc == MDB_SUCCESS) {
3019 DPRINTF(("renew txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3020 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3021 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
3022 }
3023 return rc;
3024 }
3025
3026 int
mdb_txn_begin(MDB_env * env,MDB_txn * parent,unsigned int flags,MDB_txn ** ret)3027 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
3028 {
3029 MDB_txn *txn;
3030 MDB_ntxn *ntxn;
3031 int rc, size, tsize;
3032
3033 flags &= MDB_TXN_BEGIN_FLAGS;
3034 flags |= env->me_flags & MDB_WRITEMAP;
3035
3036 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
3037 return EACCES;
3038
3039 if (parent) {
3040 /* Nested transactions: Max 1 child, write txns only, no writemap */
3041 flags |= parent->mt_flags;
3042 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) {
3043 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
3044 }
3045 /* Child txns save MDB_pgstate and use own copy of cursors */
3046 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1);
3047 size += tsize = sizeof(MDB_ntxn);
3048 } else if (flags & MDB_RDONLY) {
3049 size = env->me_maxdbs * (sizeof(MDB_db)+1);
3050 size += tsize = sizeof(MDB_txn);
3051 } else {
3052 /* Reuse preallocated write txn. However, do not touch it until
3053 * mdb_txn_renew0() succeeds, since it currently may be active.
3054 */
3055 txn = env->me_txn0;
3056 goto renew;
3057 }
3058 if ((txn = calloc(1, size)) == NULL) {
3059 DPRINTF(("calloc: %s", strerror(errno)));
3060 return ENOMEM;
3061 }
3062 #ifdef MDB_VL32
3063 if (!parent) {
3064 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
3065 if (!txn->mt_rpages) {
3066 free(txn);
3067 return ENOMEM;
3068 }
3069 txn->mt_rpages[0].mid = 0;
3070 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
3071 }
3072 #endif
3073 txn->mt_dbxs = env->me_dbxs; /* static */
3074 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
3075 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
3076 txn->mt_flags = flags;
3077 txn->mt_env = env;
3078
3079 if (parent) {
3080 unsigned int i;
3081 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
3082 txn->mt_dbiseqs = parent->mt_dbiseqs;
3083 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
3084 if (!txn->mt_u.dirty_list ||
3085 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
3086 {
3087 free(txn->mt_u.dirty_list);
3088 free(txn);
3089 return ENOMEM;
3090 }
3091 txn->mt_txnid = parent->mt_txnid;
3092 txn->mt_dirty_room = parent->mt_dirty_room;
3093 txn->mt_u.dirty_list[0].mid = 0;
3094 txn->mt_spill_pgs = NULL;
3095 txn->mt_next_pgno = parent->mt_next_pgno;
3096 parent->mt_flags |= MDB_TXN_HAS_CHILD;
3097 parent->mt_child = txn;
3098 txn->mt_parent = parent;
3099 txn->mt_numdbs = parent->mt_numdbs;
3100 #ifdef MDB_VL32
3101 txn->mt_rpages = parent->mt_rpages;
3102 #endif
3103 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3104 /* Copy parent's mt_dbflags, but clear DB_NEW */
3105 for (i=0; i<txn->mt_numdbs; i++)
3106 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
3107 rc = 0;
3108 ntxn = (MDB_ntxn *)txn;
3109 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
3110 if (env->me_pghead) {
3111 size = MDB_IDL_SIZEOF(env->me_pghead);
3112 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
3113 if (env->me_pghead)
3114 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
3115 else
3116 rc = ENOMEM;
3117 }
3118 if (!rc)
3119 rc = mdb_cursor_shadow(parent, txn);
3120 if (rc)
3121 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD);
3122 } else { /* MDB_RDONLY */
3123 txn->mt_dbiseqs = env->me_dbiseqs;
3124 renew:
3125 rc = mdb_txn_renew0(txn);
3126 }
3127 if (rc) {
3128 if (txn != env->me_txn0) {
3129 #ifdef MDB_VL32
3130 free(txn->mt_rpages);
3131 #endif
3132 free(txn);
3133 }
3134 } else {
3135 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
3136 *ret = txn;
3137 DPRINTF(("begin txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3138 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
3139 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
3140 }
3141
3142 return rc;
3143 }
3144
3145 MDB_env *
mdb_txn_env(MDB_txn * txn)3146 mdb_txn_env(MDB_txn *txn)
3147 {
3148 if(!txn) return NULL;
3149 return txn->mt_env;
3150 }
3151
3152 mdb_size_t
mdb_txn_id(MDB_txn * txn)3153 mdb_txn_id(MDB_txn *txn)
3154 {
3155 if(!txn) return 0;
3156 return txn->mt_txnid;
3157 }
3158
3159 /** Export or close DBI handles opened in this txn. */
3160 static void
mdb_dbis_update(MDB_txn * txn,int keep)3161 mdb_dbis_update(MDB_txn *txn, int keep)
3162 {
3163 int i;
3164 MDB_dbi n = txn->mt_numdbs;
3165 MDB_env *env = txn->mt_env;
3166 unsigned char *tdbflags = txn->mt_dbflags;
3167
3168 for (i = n; --i >= CORE_DBS;) {
3169 if (tdbflags[i] & DB_NEW) {
3170 if (keep) {
3171 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
3172 } else {
3173 char *ptr = env->me_dbxs[i].md_name.mv_data;
3174 if (ptr) {
3175 env->me_dbxs[i].md_name.mv_data = NULL;
3176 env->me_dbxs[i].md_name.mv_size = 0;
3177 env->me_dbflags[i] = 0;
3178 env->me_dbiseqs[i]++;
3179 free(ptr);
3180 }
3181 }
3182 }
3183 }
3184 if (keep && env->me_numdbs < n)
3185 env->me_numdbs = n;
3186 }
3187
3188 /** End a transaction, except successful commit of a nested transaction.
3189 * May be called twice for readonly txns: First reset it, then abort.
3190 * @param[in] txn the transaction handle to end
3191 * @param[in] mode why and how to end the transaction
3192 */
3193 static void
mdb_txn_end(MDB_txn * txn,unsigned mode)3194 mdb_txn_end(MDB_txn *txn, unsigned mode)
3195 {
3196 MDB_env *env = txn->mt_env;
3197 #if MDB_DEBUG
3198 static const char *const names[] = MDB_END_NAMES;
3199 #endif
3200
3201 /* Export or close DBI handles opened in this txn */
3202 mdb_dbis_update(txn, mode & MDB_END_UPDATE);
3203
3204 DPRINTF(("%s txn %"Yu"%c %p on mdbenv %p, root page %"Yu,
3205 names[mode & MDB_END_OPMASK],
3206 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
3207 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
3208
3209 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3210 if (txn->mt_u.reader) {
3211 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
3212 if (!(env->me_flags & MDB_NOTLS)) {
3213 txn->mt_u.reader = NULL; /* txn does not own reader */
3214 } else if (mode & MDB_END_SLOT) {
3215 txn->mt_u.reader->mr_pid = 0;
3216 txn->mt_u.reader = NULL;
3217 } /* else txn owns the slot until it does MDB_END_SLOT */
3218 }
3219 txn->mt_numdbs = 0; /* prevent further DBI activity */
3220 txn->mt_flags |= MDB_TXN_FINISHED;
3221
3222 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) {
3223 pgno_t *pghead = env->me_pghead;
3224
3225 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */
3226 mdb_cursors_close(txn, 0);
3227 if (!(env->me_flags & MDB_WRITEMAP)) {
3228 mdb_dlist_free(txn);
3229 }
3230
3231 txn->mt_numdbs = 0;
3232 txn->mt_flags = MDB_TXN_FINISHED;
3233
3234 if (!txn->mt_parent) {
3235 mdb_midl_shrink(&txn->mt_free_pgs);
3236 env->me_free_pgs = txn->mt_free_pgs;
3237 /* me_pgstate: */
3238 env->me_pghead = NULL;
3239 env->me_pglast = 0;
3240
3241 env->me_txn = NULL;
3242 mode = 0; /* txn == env->me_txn0, do not free() it */
3243
3244 /* The writer mutex was locked in mdb_txn_begin. */
3245 if (env->me_txns)
3246 UNLOCK_MUTEX(env->me_wmutex);
3247 } else {
3248 txn->mt_parent->mt_child = NULL;
3249 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD;
3250 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
3251 mdb_midl_free(txn->mt_free_pgs);
3252 mdb_midl_free(txn->mt_spill_pgs);
3253 free(txn->mt_u.dirty_list);
3254 }
3255
3256 mdb_midl_free(pghead);
3257 }
3258 #ifdef MDB_VL32
3259 if (!txn->mt_parent) {
3260 MDB_ID3L el = env->me_rpages, tl = txn->mt_rpages;
3261 unsigned i, x, n = tl[0].mid;
3262 pthread_mutex_lock(&env->me_rpmutex);
3263 for (i = 1; i <= n; i++) {
3264 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
3265 /* tmp overflow pages that we didn't share in env */
3266 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3267 } else {
3268 x = mdb_mid3l_search(el, tl[i].mid);
3269 if (tl[i].mptr == el[x].mptr) {
3270 el[x].mref--;
3271 } else {
3272 /* another tmp overflow page */
3273 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
3274 }
3275 }
3276 }
3277 pthread_mutex_unlock(&env->me_rpmutex);
3278 tl[0].mid = 0;
3279 if (mode & MDB_END_FREE)
3280 free(tl);
3281 }
3282 #endif
3283 if (mode & MDB_END_FREE)
3284 free(txn);
3285 }
3286
3287 void
mdb_txn_reset(MDB_txn * txn)3288 mdb_txn_reset(MDB_txn *txn)
3289 {
3290 if (txn == NULL)
3291 return;
3292
3293 /* This call is only valid for read-only txns */
3294 if (!(txn->mt_flags & MDB_TXN_RDONLY))
3295 return;
3296
3297 mdb_txn_end(txn, MDB_END_RESET);
3298 }
3299
3300 void
mdb_txn_abort(MDB_txn * txn)3301 mdb_txn_abort(MDB_txn *txn)
3302 {
3303 if (txn == NULL)
3304 return;
3305
3306 if (txn->mt_child)
3307 mdb_txn_abort(txn->mt_child);
3308
3309 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE);
3310 }
3311
3312 /** Save the freelist as of this transaction to the freeDB.
3313 * This changes the freelist. Keep trying until it stabilizes.
3314 *
3315 * When (MDB_DEVEL) & 2, the changes do not affect #mdb_page_alloc(),
3316 * it then uses the transaction's original snapshot of the freeDB.
3317 */
3318 static int
mdb_freelist_save(MDB_txn * txn)3319 mdb_freelist_save(MDB_txn *txn)
3320 {
3321 /* env->me_pghead[] can grow and shrink during this call.
3322 * env->me_pglast and txn->mt_free_pgs[] can only grow.
3323 * Page numbers cannot disappear from txn->mt_free_pgs[].
3324 */
3325 MDB_cursor mc;
3326 MDB_env *env = txn->mt_env;
3327 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
3328 txnid_t pglast = 0, head_id = 0;
3329 pgno_t freecnt = 0, *free_pgs, *mop;
3330 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
3331
3332 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
3333
3334 if (env->me_pghead) {
3335 /* Make sure first page of freeDB is touched and on freelist */
3336 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
3337 if (rc && rc != MDB_NOTFOUND)
3338 return rc;
3339 }
3340
3341 if (!env->me_pghead && txn->mt_loose_pgs) {
3342 /* Put loose page numbers in mt_free_pgs, since
3343 * we may be unable to return them to me_pghead.
3344 */
3345 MDB_page *mp = txn->mt_loose_pgs;
3346 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
3347 return rc;
3348 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
3349 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
3350 txn->mt_loose_pgs = NULL;
3351 txn->mt_loose_count = 0;
3352 }
3353
3354 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
3355 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
3356 ? SSIZE_MAX : maxfree_1pg;
3357
3358 for (;;) {
3359 /* Come back here after each Put() in case freelist changed */
3360 MDB_val key, data;
3361 pgno_t *pgs;
3362 ssize_t j;
3363
3364 /* If using records from freeDB which we have not yet
3365 * deleted, delete them and any we reserved for me_pghead.
3366 */
3367 while (pglast < env->me_pglast) {
3368 rc = mdb_cursor_first(&mc, &key, NULL);
3369 if (rc)
3370 return rc;
3371 pglast = head_id = *(txnid_t *)key.mv_data;
3372 total_room = head_room = 0;
3373 mdb_tassert(txn, pglast <= env->me_pglast);
3374 rc = mdb_cursor_del(&mc, 0);
3375 if (rc)
3376 return rc;
3377 }
3378
3379 /* Save the IDL of pages freed by this txn, to a single record */
3380 if (freecnt < txn->mt_free_pgs[0]) {
3381 if (!freecnt) {
3382 /* Make sure last page of freeDB is touched and on freelist */
3383 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
3384 if (rc && rc != MDB_NOTFOUND)
3385 return rc;
3386 }
3387 free_pgs = txn->mt_free_pgs;
3388 /* Write to last page of freeDB */
3389 key.mv_size = sizeof(txn->mt_txnid);
3390 key.mv_data = &txn->mt_txnid;
3391 do {
3392 freecnt = free_pgs[0];
3393 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
3394 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3395 if (rc)
3396 return rc;
3397 /* Retry if mt_free_pgs[] grew during the Put() */
3398 free_pgs = txn->mt_free_pgs;
3399 } while (freecnt < free_pgs[0]);
3400 mdb_midl_sort(free_pgs);
3401 memcpy(data.mv_data, free_pgs, data.mv_size);
3402 #if (MDB_DEBUG) > 1
3403 {
3404 unsigned int i = free_pgs[0];
3405 DPRINTF(("IDL write txn %"Yu" root %"Yu" num %u",
3406 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3407 for (; i; i--)
3408 DPRINTF(("IDL %"Yu, free_pgs[i]));
3409 }
3410 #endif
3411 continue;
3412 }
3413
3414 mop = env->me_pghead;
3415 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3416
3417 /* Reserve records for me_pghead[]. Split it if multi-page,
3418 * to avoid searching freeDB for a page range. Use keys in
3419 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3420 */
3421 if (total_room >= mop_len) {
3422 if (total_room == mop_len || --more < 0)
3423 break;
3424 } else if (head_room >= maxfree_1pg && head_id > 1) {
3425 /* Keep current record (overflow page), add a new one */
3426 head_id--;
3427 head_room = 0;
3428 }
3429 /* (Re)write {key = head_id, IDL length = head_room} */
3430 total_room -= head_room;
3431 head_room = mop_len - total_room;
3432 if (head_room > maxfree_1pg && head_id > 1) {
3433 /* Overflow multi-page for part of me_pghead */
3434 head_room /= head_id; /* amortize page sizes */
3435 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3436 } else if (head_room < 0) {
3437 /* Rare case, not bothering to delete this record */
3438 head_room = 0;
3439 }
3440 key.mv_size = sizeof(head_id);
3441 key.mv_data = &head_id;
3442 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3443 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3444 if (rc)
3445 return rc;
3446 /* IDL is initially empty, zero out at least the length */
3447 pgs = (pgno_t *)data.mv_data;
3448 j = head_room > clean_limit ? head_room : 0;
3449 do {
3450 pgs[j] = 0;
3451 } while (--j >= 0);
3452 total_room += head_room;
3453 }
3454
3455 /* Return loose page numbers to me_pghead, though usually none are
3456 * left at this point. The pages themselves remain in dirty_list.
3457 */
3458 if (txn->mt_loose_pgs) {
3459 MDB_page *mp = txn->mt_loose_pgs;
3460 unsigned count = txn->mt_loose_count;
3461 MDB_IDL loose;
3462 /* Room for loose pages + temp IDL with same */
3463 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3464 return rc;
3465 mop = env->me_pghead;
3466 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3467 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3468 loose[ ++count ] = mp->mp_pgno;
3469 loose[0] = count;
3470 mdb_midl_sort(loose);
3471 mdb_midl_xmerge(mop, loose);
3472 txn->mt_loose_pgs = NULL;
3473 txn->mt_loose_count = 0;
3474 mop_len = mop[0];
3475 }
3476
3477 /* Fill in the reserved me_pghead records */
3478 rc = MDB_SUCCESS;
3479 if (mop_len) {
3480 MDB_val key, data;
3481
3482 mop += mop_len;
3483 rc = mdb_cursor_first(&mc, &key, &data);
3484 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3485 txnid_t id = *(txnid_t *)key.mv_data;
3486 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3487 MDB_ID save;
3488
3489 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3490 key.mv_data = &id;
3491 if (len > mop_len) {
3492 len = mop_len;
3493 data.mv_size = (len + 1) * sizeof(MDB_ID);
3494 }
3495 data.mv_data = mop -= len;
3496 save = mop[0];
3497 mop[0] = len;
3498 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3499 mop[0] = save;
3500 if (rc || !(mop_len -= len))
3501 break;
3502 }
3503 }
3504 return rc;
3505 }
3506
3507 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3508 * @param[in] txn the transaction that's being committed
3509 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3510 * @return 0 on success, non-zero on failure.
3511 */
3512 static int
mdb_page_flush(MDB_txn * txn,int keep)3513 mdb_page_flush(MDB_txn *txn, int keep)
3514 {
3515 MDB_env *env = txn->mt_env;
3516 MDB_ID2L dl = txn->mt_u.dirty_list;
3517 unsigned psize = env->me_psize, j;
3518 int i, pagecount = dl[0].mid, rc;
3519 size_t size = 0;
3520 off_t pos = 0;
3521 pgno_t pgno = 0;
3522 MDB_page *dp = NULL;
3523 #ifdef _WIN32
3524 OVERLAPPED ov;
3525 #else
3526 struct iovec iov[MDB_COMMIT_PAGES];
3527 ssize_t wsize = 0, wres;
3528 off_t wpos = 0, next_pos = 1; /* impossible pos, so pos != next_pos */
3529 int n = 0;
3530 #endif
3531
3532 j = i = keep;
3533
3534 if (env->me_flags & MDB_WRITEMAP) {
3535 /* Clear dirty flags */
3536 while (++i <= pagecount) {
3537 dp = dl[i].mptr;
3538 /* Don't flush this page yet */
3539 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3540 dp->mp_flags &= ~P_KEEP;
3541 dl[++j] = dl[i];
3542 continue;
3543 }
3544 dp->mp_flags &= ~P_DIRTY;
3545 }
3546 goto done;
3547 }
3548
3549 /* Write the pages */
3550 for (;;) {
3551 if (++i <= pagecount) {
3552 dp = dl[i].mptr;
3553 /* Don't flush this page yet */
3554 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3555 dp->mp_flags &= ~P_KEEP;
3556 dl[i].mid = 0;
3557 continue;
3558 }
3559 pgno = dl[i].mid;
3560 /* clear dirty flag */
3561 dp->mp_flags &= ~P_DIRTY;
3562 pos = pgno * psize;
3563 size = psize;
3564 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3565 }
3566 #ifdef _WIN32
3567 else break;
3568
3569 /* Windows actually supports scatter/gather I/O, but only on
3570 * unbuffered file handles. Since we're relying on the OS page
3571 * cache for all our data, that's self-defeating. So we just
3572 * write pages one at a time. We use the ov structure to set
3573 * the write offset, to at least save the overhead of a Seek
3574 * system call.
3575 */
3576 DPRINTF(("committing page %"Yu, pgno));
3577 memset(&ov, 0, sizeof(ov));
3578 ov.Offset = pos & 0xffffffff;
3579 ov.OffsetHigh = pos >> 16 >> 16;
3580 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3581 rc = ErrCode();
3582 DPRINTF(("WriteFile: %d", rc));
3583 return rc;
3584 }
3585 #else
3586 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3587 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3588 if (n) {
3589 retry_write:
3590 /* Write previous page(s) */
3591 #ifdef MDB_USE_PWRITEV
3592 wres = pwritev(env->me_fd, iov, n, wpos);
3593 #else
3594 if (n == 1) {
3595 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3596 } else {
3597 retry_seek:
3598 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3599 rc = ErrCode();
3600 if (rc == EINTR)
3601 goto retry_seek;
3602 DPRINTF(("lseek: %s", strerror(rc)));
3603 return rc;
3604 }
3605 wres = writev(env->me_fd, iov, n);
3606 }
3607 #endif
3608 if (wres != wsize) {
3609 if (wres < 0) {
3610 rc = ErrCode();
3611 if (rc == EINTR)
3612 goto retry_write;
3613 DPRINTF(("Write error: %s", strerror(rc)));
3614 } else {
3615 rc = EIO; /* TODO: Use which error code? */
3616 DPUTS("short write, filesystem full?");
3617 }
3618 return rc;
3619 }
3620 n = 0;
3621 }
3622 if (i > pagecount)
3623 break;
3624 wpos = pos;
3625 wsize = 0;
3626 }
3627 DPRINTF(("committing page %"Yu, pgno));
3628 next_pos = pos + size;
3629 iov[n].iov_len = size;
3630 iov[n].iov_base = (char *)dp;
3631 wsize += size;
3632 n++;
3633 #endif /* _WIN32 */
3634 }
3635 #ifdef MDB_VL32
3636 if (pgno > txn->mt_last_pgno)
3637 txn->mt_last_pgno = pgno;
3638 #endif
3639
3640 /* MIPS has cache coherency issues, this is a no-op everywhere else
3641 * Note: for any size >= on-chip cache size, entire on-chip cache is
3642 * flushed.
3643 */
3644 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3645
3646 for (i = keep; ++i <= pagecount; ) {
3647 dp = dl[i].mptr;
3648 /* This is a page we skipped above */
3649 if (!dl[i].mid) {
3650 dl[++j] = dl[i];
3651 dl[j].mid = dp->mp_pgno;
3652 continue;
3653 }
3654 mdb_dpage_free(env, dp);
3655 }
3656
3657 done:
3658 i--;
3659 txn->mt_dirty_room += i - j;
3660 dl[0].mid = j;
3661 return MDB_SUCCESS;
3662 }
3663
3664 int
mdb_txn_commit(MDB_txn * txn)3665 mdb_txn_commit(MDB_txn *txn)
3666 {
3667 int rc;
3668 unsigned int i, end_mode;
3669 MDB_env *env;
3670
3671 if (txn == NULL)
3672 return EINVAL;
3673
3674 /* mdb_txn_end() mode for a commit which writes nothing */
3675 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE;
3676
3677 if (txn->mt_child) {
3678 rc = mdb_txn_commit(txn->mt_child);
3679 if (rc)
3680 goto fail;
3681 }
3682
3683 env = txn->mt_env;
3684
3685 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3686 goto done;
3687 }
3688
3689 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) {
3690 DPUTS("txn has failed/finished, can't commit");
3691 if (txn->mt_parent)
3692 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3693 rc = MDB_BAD_TXN;
3694 goto fail;
3695 }
3696
3697 if (txn->mt_parent) {
3698 MDB_txn *parent = txn->mt_parent;
3699 MDB_page **lp;
3700 MDB_ID2L dst, src;
3701 MDB_IDL pspill;
3702 unsigned x, y, len, ps_len;
3703
3704 /* Append our free list to parent's */
3705 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3706 if (rc)
3707 goto fail;
3708 mdb_midl_free(txn->mt_free_pgs);
3709 /* Failures after this must either undo the changes
3710 * to the parent or set MDB_TXN_ERROR in the parent.
3711 */
3712
3713 parent->mt_next_pgno = txn->mt_next_pgno;
3714 parent->mt_flags = txn->mt_flags;
3715
3716 /* Merge our cursors into parent's and close them */
3717 mdb_cursors_close(txn, 1);
3718
3719 /* Update parent's DB table. */
3720 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3721 parent->mt_numdbs = txn->mt_numdbs;
3722 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI];
3723 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI];
3724 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
3725 /* preserve parent's DB_NEW status */
3726 x = parent->mt_dbflags[i] & DB_NEW;
3727 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3728 }
3729
3730 dst = parent->mt_u.dirty_list;
3731 src = txn->mt_u.dirty_list;
3732 /* Remove anything in our dirty list from parent's spill list */
3733 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3734 x = y = ps_len;
3735 pspill[0] = (pgno_t)-1;
3736 /* Mark our dirty pages as deleted in parent spill list */
3737 for (i=0, len=src[0].mid; ++i <= len; ) {
3738 MDB_ID pn = src[i].mid << 1;
3739 while (pn > pspill[x])
3740 x--;
3741 if (pn == pspill[x]) {
3742 pspill[x] = 1;
3743 y = --x;
3744 }
3745 }
3746 /* Squash deleted pagenums if we deleted any */
3747 for (x=y; ++x <= ps_len; )
3748 if (!(pspill[x] & 1))
3749 pspill[++y] = pspill[x];
3750 pspill[0] = y;
3751 }
3752
3753 /* Remove anything in our spill list from parent's dirty list */
3754 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) {
3755 for (i=1; i<=txn->mt_spill_pgs[0]; i++) {
3756 MDB_ID pn = txn->mt_spill_pgs[i];
3757 if (pn & 1)
3758 continue; /* deleted spillpg */
3759 pn >>= 1;
3760 y = mdb_mid2l_search(dst, pn);
3761 if (y <= dst[0].mid && dst[y].mid == pn) {
3762 free(dst[y].mptr);
3763 while (y < dst[0].mid) {
3764 dst[y] = dst[y+1];
3765 y++;
3766 }
3767 dst[0].mid--;
3768 }
3769 }
3770 }
3771
3772 /* Find len = length of merging our dirty list with parent's */
3773 x = dst[0].mid;
3774 dst[0].mid = 0; /* simplify loops */
3775 if (parent->mt_parent) {
3776 len = x + src[0].mid;
3777 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3778 for (i = x; y && i; y--) {
3779 pgno_t yp = src[y].mid;
3780 while (yp < dst[i].mid)
3781 i--;
3782 if (yp == dst[i].mid) {
3783 i--;
3784 len--;
3785 }
3786 }
3787 } else { /* Simplify the above for single-ancestor case */
3788 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3789 }
3790 /* Merge our dirty list with parent's */
3791 y = src[0].mid;
3792 for (i = len; y; dst[i--] = src[y--]) {
3793 pgno_t yp = src[y].mid;
3794 while (yp < dst[x].mid)
3795 dst[i--] = dst[x--];
3796 if (yp == dst[x].mid)
3797 free(dst[x--].mptr);
3798 }
3799 mdb_tassert(txn, i == x);
3800 dst[0].mid = len;
3801 free(txn->mt_u.dirty_list);
3802 parent->mt_dirty_room = txn->mt_dirty_room;
3803 if (txn->mt_spill_pgs) {
3804 if (parent->mt_spill_pgs) {
3805 /* TODO: Prevent failure here, so parent does not fail */
3806 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3807 if (rc)
3808 parent->mt_flags |= MDB_TXN_ERROR;
3809 mdb_midl_free(txn->mt_spill_pgs);
3810 mdb_midl_sort(parent->mt_spill_pgs);
3811 } else {
3812 parent->mt_spill_pgs = txn->mt_spill_pgs;
3813 }
3814 }
3815
3816 /* Append our loose page list to parent's */
3817 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp))
3818 ;
3819 *lp = txn->mt_loose_pgs;
3820 parent->mt_loose_count += txn->mt_loose_count;
3821
3822 parent->mt_child = NULL;
3823 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3824 free(txn);
3825 return rc;
3826 }
3827
3828 if (txn != env->me_txn) {
3829 DPUTS("attempt to commit unknown transaction");
3830 rc = EINVAL;
3831 goto fail;
3832 }
3833
3834 mdb_cursors_close(txn, 0);
3835
3836 if (!txn->mt_u.dirty_list[0].mid &&
3837 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3838 goto done;
3839
3840 DPRINTF(("committing txn %"Yu" %p on mdbenv %p, root page %"Yu,
3841 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3842
3843 /* Update DB root pointers */
3844 if (txn->mt_numdbs > CORE_DBS) {
3845 MDB_cursor mc;
3846 MDB_dbi i;
3847 MDB_val data;
3848 data.mv_size = sizeof(MDB_db);
3849
3850 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3851 for (i = CORE_DBS; i < txn->mt_numdbs; i++) {
3852 if (txn->mt_dbflags[i] & DB_DIRTY) {
3853 if (TXN_DBI_CHANGED(txn, i)) {
3854 rc = MDB_BAD_DBI;
3855 goto fail;
3856 }
3857 data.mv_data = &txn->mt_dbs[i];
3858 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3859 F_SUBDATA);
3860 if (rc)
3861 goto fail;
3862 }
3863 }
3864 }
3865
3866 rc = mdb_freelist_save(txn);
3867 if (rc)
3868 goto fail;
3869
3870 mdb_midl_free(env->me_pghead);
3871 env->me_pghead = NULL;
3872 mdb_midl_shrink(&txn->mt_free_pgs);
3873
3874 #if (MDB_DEBUG) > 2
3875 mdb_audit(txn);
3876 #endif
3877
3878 if ((rc = mdb_page_flush(txn, 0)))
3879 goto fail;
3880 if (!F_ISSET(txn->mt_flags, MDB_TXN_NOSYNC) &&
3881 (rc = mdb_env_sync0(env, 0, txn->mt_next_pgno)))
3882 goto fail;
3883 if ((rc = mdb_env_write_meta(txn)))
3884 goto fail;
3885 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE;
3886
3887 done:
3888 mdb_txn_end(txn, end_mode);
3889 return MDB_SUCCESS;
3890
3891 fail:
3892 mdb_txn_abort(txn);
3893 return rc;
3894 }
3895
3896 /** Read the environment parameters of a DB environment before
3897 * mapping it into memory.
3898 * @param[in] env the environment handle
3899 * @param[out] meta address of where to store the meta information
3900 * @return 0 on success, non-zero on failure.
3901 */
3902 static int ESECT
mdb_env_read_header(MDB_env * env,MDB_meta * meta)3903 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3904 {
3905 MDB_metabuf pbuf;
3906 MDB_page *p;
3907 MDB_meta *m;
3908 int i, rc, off;
3909 enum { Size = sizeof(pbuf) };
3910
3911 /* We don't know the page size yet, so use a minimum value.
3912 * Read both meta pages so we can use the latest one.
3913 */
3914
3915 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) {
3916 #ifdef _WIN32
3917 DWORD len;
3918 OVERLAPPED ov;
3919 memset(&ov, 0, sizeof(ov));
3920 ov.Offset = off;
3921 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3922 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3923 rc = 0;
3924 #else
3925 rc = pread(env->me_fd, &pbuf, Size, off);
3926 #endif
3927 if (rc != Size) {
3928 if (rc == 0 && off == 0)
3929 return ENOENT;
3930 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3931 DPRINTF(("read: %s", mdb_strerror(rc)));
3932 return rc;
3933 }
3934
3935 p = (MDB_page *)&pbuf;
3936
3937 if (!F_ISSET(p->mp_flags, P_META)) {
3938 DPRINTF(("page %"Yu" not a meta page", p->mp_pgno));
3939 return MDB_INVALID;
3940 }
3941
3942 m = METADATA(p);
3943 if (m->mm_magic != MDB_MAGIC) {
3944 DPUTS("meta has invalid magic");
3945 return MDB_INVALID;
3946 }
3947
3948 if (m->mm_version != MDB_DATA_VERSION) {
3949 DPRINTF(("database is version %u, expected version %u",
3950 m->mm_version, MDB_DATA_VERSION));
3951 return MDB_VERSION_MISMATCH;
3952 }
3953
3954 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3955 *meta = *m;
3956 }
3957 return 0;
3958 }
3959
3960 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3961 static void ESECT
mdb_env_init_meta0(MDB_env * env,MDB_meta * meta)3962 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3963 {
3964 meta->mm_magic = MDB_MAGIC;
3965 meta->mm_version = MDB_DATA_VERSION;
3966 meta->mm_mapsize = env->me_mapsize;
3967 meta->mm_psize = env->me_psize;
3968 meta->mm_last_pg = NUM_METAS-1;
3969 meta->mm_flags = env->me_flags & 0xffff;
3970 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */
3971 meta->mm_dbs[FREE_DBI].md_root = P_INVALID;
3972 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID;
3973 }
3974
3975 /** Write the environment parameters of a freshly created DB environment.
3976 * @param[in] env the environment handle
3977 * @param[in] meta the #MDB_meta to write
3978 * @return 0 on success, non-zero on failure.
3979 */
3980 static int ESECT
mdb_env_init_meta(MDB_env * env,MDB_meta * meta)3981 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3982 {
3983 MDB_page *p, *q;
3984 int rc;
3985 unsigned int psize;
3986 #ifdef _WIN32
3987 DWORD len;
3988 OVERLAPPED ov;
3989 memset(&ov, 0, sizeof(ov));
3990 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3991 ov.Offset = pos; \
3992 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3993 #else
3994 int len;
3995 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3996 len = pwrite(fd, ptr, size, pos); \
3997 if (len == -1 && ErrCode() == EINTR) continue; \
3998 rc = (len >= 0); break; } while(1)
3999 #endif
4000
4001 DPUTS("writing new meta page");
4002
4003 psize = env->me_psize;
4004
4005 p = calloc(NUM_METAS, psize);
4006 if (!p)
4007 return ENOMEM;
4008 p->mp_pgno = 0;
4009 p->mp_flags = P_META;
4010 *(MDB_meta *)METADATA(p) = *meta;
4011
4012 q = (MDB_page *)((char *)p + psize);
4013 q->mp_pgno = 1;
4014 q->mp_flags = P_META;
4015 *(MDB_meta *)METADATA(q) = *meta;
4016
4017 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0);
4018 if (!rc)
4019 rc = ErrCode();
4020 else if ((unsigned) len == psize * NUM_METAS)
4021 rc = MDB_SUCCESS;
4022 else
4023 rc = ENOSPC;
4024 free(p);
4025 return rc;
4026 }
4027
4028 /** Update the environment info to commit a transaction.
4029 * @param[in] txn the transaction that's being committed
4030 * @return 0 on success, non-zero on failure.
4031 */
4032 static int
mdb_env_write_meta(MDB_txn * txn)4033 mdb_env_write_meta(MDB_txn *txn)
4034 {
4035 MDB_env *env;
4036 MDB_meta meta, metab, *mp;
4037 unsigned flags;
4038 mdb_size_t mapsize;
4039 off_t off;
4040 int rc, len, toggle;
4041 char *ptr;
4042 HANDLE mfd;
4043 #ifdef _WIN32
4044 OVERLAPPED ov;
4045 #else
4046 int r2;
4047 #endif
4048
4049 toggle = txn->mt_txnid & 1;
4050 DPRINTF(("writing meta page %d for root page %"Yu,
4051 toggle, txn->mt_dbs[MAIN_DBI].md_root));
4052
4053 env = txn->mt_env;
4054 flags = txn->mt_flags | env->me_flags;
4055 mp = env->me_metas[toggle];
4056 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
4057 /* Persist any increases of mapsize config */
4058 if (mapsize < env->me_mapsize)
4059 mapsize = env->me_mapsize;
4060
4061 if (flags & MDB_WRITEMAP) {
4062 mp->mm_mapsize = mapsize;
4063 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4064 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4065 mp->mm_last_pg = txn->mt_next_pgno - 1;
4066 #if defined(__SUNPRO_C)
4067 __machine_rw_barrier();
4068 #elif defined(__GNUC__)
4069 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
4070 !(defined(__i386__) || defined(__x86_64__))
4071 /* LY: issue a memory barrier, if not x86. ITS#7969 */
4072 __sync_synchronize();
4073 #endif
4074 #endif
4075 mp->mm_txnid = txn->mt_txnid;
4076 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
4077 unsigned meta_size = env->me_psize;
4078 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
4079 ptr = (char *)mp - PAGEHDRSZ;
4080 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
4081 r2 = (ptr - env->me_map) & (env->me_os_psize - 1);
4082 ptr -= r2;
4083 meta_size += r2;
4084 #endif
4085 if (MDB_MSYNC(ptr, meta_size, rc)) {
4086 rc = ErrCode();
4087 goto fail;
4088 }
4089 }
4090 goto done;
4091 }
4092 metab.mm_txnid = mp->mm_txnid;
4093 metab.mm_last_pg = mp->mm_last_pg;
4094
4095 meta.mm_mapsize = mapsize;
4096 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI];
4097 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
4098 meta.mm_last_pg = txn->mt_next_pgno - 1;
4099 meta.mm_txnid = txn->mt_txnid;
4100
4101 off = offsetof(MDB_meta, mm_mapsize);
4102 ptr = (char *)&meta + off;
4103 len = sizeof(MDB_meta) - off;
4104 off += (char *)mp - env->me_map;
4105
4106 /* Write to the SYNC fd */
4107 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
4108 #ifdef _WIN32
4109 {
4110 memset(&ov, 0, sizeof(ov));
4111 ov.Offset = off;
4112 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
4113 rc = -1;
4114 }
4115 #else
4116 retry_write:
4117 rc = pwrite(mfd, ptr, len, off);
4118 #endif
4119 if (rc != len) {
4120 rc = rc < 0 ? ErrCode() : EIO;
4121 #ifndef _WIN32
4122 if (rc == EINTR)
4123 goto retry_write;
4124 #endif
4125 DPUTS("write failed, disk error?");
4126 /* On a failure, the pagecache still contains the new data.
4127 * Write some old data back, to prevent it from being used.
4128 * Use the non-SYNC fd; we know it will fail anyway.
4129 */
4130 meta.mm_last_pg = metab.mm_last_pg;
4131 meta.mm_txnid = metab.mm_txnid;
4132 #ifdef _WIN32
4133 memset(&ov, 0, sizeof(ov));
4134 ov.Offset = off;
4135 WriteFile(env->me_fd, ptr, len, NULL, &ov);
4136 #else
4137 r2 = pwrite(env->me_fd, ptr, len, off);
4138 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
4139 #endif
4140 fail:
4141 env->me_flags |= MDB_FATAL_ERROR;
4142 return rc;
4143 }
4144 /* MIPS has cache coherency issues, this is a no-op everywhere else */
4145 CACHEFLUSH(env->me_map + off, len, DCACHE);
4146 done:
4147 /* Memory ordering issues are irrelevant; since the entire writer
4148 * is wrapped by wmutex, all of these changes will become visible
4149 * after the wmutex is unlocked. Since the DB is multi-version,
4150 * readers will get consistent data regardless of how fresh or
4151 * how stale their view of these values is.
4152 */
4153 if (env->me_txns)
4154 env->me_txns->mti_txnid = txn->mt_txnid;
4155
4156 return MDB_SUCCESS;
4157 }
4158
4159 /** Check both meta pages to see which one is newer.
4160 * @param[in] env the environment handle
4161 * @return newest #MDB_meta.
4162 */
4163 static MDB_meta *
mdb_env_pick_meta(const MDB_env * env)4164 mdb_env_pick_meta(const MDB_env *env)
4165 {
4166 MDB_meta *const *metas = env->me_metas;
4167 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ];
4168 }
4169
4170 int ESECT
mdb_env_create(MDB_env ** env)4171 mdb_env_create(MDB_env **env)
4172 {
4173 MDB_env *e;
4174
4175 e = calloc(1, sizeof(MDB_env));
4176 if (!e)
4177 return ENOMEM;
4178
4179 e->me_maxreaders = DEFAULT_READERS;
4180 e->me_maxdbs = e->me_numdbs = CORE_DBS;
4181 e->me_fd = INVALID_HANDLE_VALUE;
4182 e->me_lfd = INVALID_HANDLE_VALUE;
4183 e->me_mfd = INVALID_HANDLE_VALUE;
4184 #ifdef MDB_USE_POSIX_SEM
4185 e->me_rmutex = SEM_FAILED;
4186 e->me_wmutex = SEM_FAILED;
4187 #elif defined MDB_USE_SYSV_SEM
4188 e->me_rmutex->semid = -1;
4189 e->me_wmutex->semid = -1;
4190 #endif
4191 e->me_pid = getpid();
4192 GET_PAGESIZE(e->me_os_psize);
4193 VGMEMP_CREATE(e,0,0);
4194 *env = e;
4195 return MDB_SUCCESS;
4196 }
4197
4198 #ifdef _WIN32
4199 /** @brief Map a result from an NTAPI call to WIN32. */
4200 static DWORD
mdb_nt2win32(NTSTATUS st)4201 mdb_nt2win32(NTSTATUS st)
4202 {
4203 OVERLAPPED o = {0};
4204 DWORD br;
4205 o.Internal = st;
4206 GetOverlappedResult(NULL, &o, &br, FALSE);
4207 return GetLastError();
4208 }
4209 #endif
4210
4211 static int ESECT
mdb_env_map(MDB_env * env,void * addr)4212 mdb_env_map(MDB_env *env, void *addr)
4213 {
4214 MDB_page *p;
4215 unsigned int flags = env->me_flags;
4216 #ifdef _WIN32
4217 int rc;
4218 int access = SECTION_MAP_READ;
4219 HANDLE mh;
4220 void *map;
4221 SIZE_T msize;
4222 ULONG pageprot = PAGE_READONLY, secprot, alloctype;
4223
4224 if (flags & MDB_WRITEMAP) {
4225 access |= SECTION_MAP_WRITE;
4226 pageprot = PAGE_READWRITE;
4227 }
4228 if (flags & MDB_RDONLY) {
4229 secprot = PAGE_READONLY;
4230 msize = 0;
4231 alloctype = 0;
4232 } else {
4233 secprot = PAGE_READWRITE;
4234 msize = env->me_mapsize;
4235 alloctype = MEM_RESERVE;
4236 }
4237
4238 rc = NtCreateSection(&mh, access, NULL, NULL, secprot, SEC_RESERVE, env->me_fd);
4239 if (rc)
4240 return mdb_nt2win32(rc);
4241 map = addr;
4242 #ifdef MDB_VL32
4243 msize = NUM_METAS * env->me_psize;
4244 #endif
4245 rc = NtMapViewOfSection(mh, GetCurrentProcess(), &map, 0, 0, NULL, &msize, ViewUnmap, alloctype, pageprot);
4246 #ifdef MDB_VL32
4247 env->me_fmh = mh;
4248 #else
4249 NtClose(mh);
4250 #endif
4251 if (rc)
4252 return mdb_nt2win32(rc);
4253 env->me_map = map;
4254 #else
4255 #ifdef MDB_VL32
4256 (void) flags;
4257 env->me_map = mmap(addr, NUM_METAS * env->me_psize, PROT_READ, MAP_SHARED,
4258 env->me_fd, 0);
4259 if (env->me_map == MAP_FAILED) {
4260 env->me_map = NULL;
4261 return ErrCode();
4262 }
4263 #else
4264 int prot = PROT_READ;
4265 if (flags & MDB_WRITEMAP) {
4266 prot |= PROT_WRITE;
4267 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
4268 return ErrCode();
4269 }
4270 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
4271 env->me_fd, 0);
4272 if (env->me_map == MAP_FAILED) {
4273 env->me_map = NULL;
4274 return ErrCode();
4275 }
4276
4277 if (flags & MDB_NORDAHEAD) {
4278 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
4279 #ifdef MADV_RANDOM
4280 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
4281 #else
4282 #ifdef POSIX_MADV_RANDOM
4283 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
4284 #endif /* POSIX_MADV_RANDOM */
4285 #endif /* MADV_RANDOM */
4286 }
4287 #endif /* _WIN32 */
4288
4289 /* Can happen because the address argument to mmap() is just a
4290 * hint. mmap() can pick another, e.g. if the range is in use.
4291 * The MAP_FIXED flag would prevent that, but then mmap could
4292 * instead unmap existing pages to make room for the new map.
4293 */
4294 if (addr && env->me_map != addr)
4295 return EBUSY; /* TODO: Make a new MDB_* error code? */
4296 #endif
4297
4298 p = (MDB_page *)env->me_map;
4299 env->me_metas[0] = METADATA(p);
4300 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
4301
4302 return MDB_SUCCESS;
4303 }
4304
4305 int ESECT
mdb_env_set_mapsize(MDB_env * env,mdb_size_t size)4306 mdb_env_set_mapsize(MDB_env *env, mdb_size_t size)
4307 {
4308 /* If env is already open, caller is responsible for making
4309 * sure there are no active txns.
4310 */
4311 if (env->me_map) {
4312 MDB_meta *meta;
4313 #ifndef MDB_VL32
4314 void *old;
4315 int rc;
4316 #endif
4317 if (env->me_txn)
4318 return EINVAL;
4319 meta = mdb_env_pick_meta(env);
4320 if (!size)
4321 size = meta->mm_mapsize;
4322 {
4323 /* Silently round up to minimum if the size is too small */
4324 mdb_size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
4325 if (size < minsize)
4326 size = minsize;
4327 }
4328 #ifndef MDB_VL32
4329 /* For MDB_VL32 this bit is a noop since we dynamically remap
4330 * chunks of the DB anyway.
4331 */
4332 munmap(env->me_map, env->me_mapsize);
4333 env->me_mapsize = size;
4334 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
4335 rc = mdb_env_map(env, old);
4336 if (rc)
4337 return rc;
4338 #endif /* !MDB_VL32 */
4339 }
4340 env->me_mapsize = size;
4341 if (env->me_psize)
4342 env->me_maxpg = env->me_mapsize / env->me_psize;
4343 return MDB_SUCCESS;
4344 }
4345
4346 int ESECT
mdb_env_set_maxdbs(MDB_env * env,MDB_dbi dbs)4347 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
4348 {
4349 if (env->me_map)
4350 return EINVAL;
4351 env->me_maxdbs = dbs + CORE_DBS;
4352 return MDB_SUCCESS;
4353 }
4354
4355 int ESECT
mdb_env_set_maxreaders(MDB_env * env,unsigned int readers)4356 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
4357 {
4358 if (env->me_map || readers < 1)
4359 return EINVAL;
4360 env->me_maxreaders = readers;
4361 return MDB_SUCCESS;
4362 }
4363
4364 int ESECT
mdb_env_get_maxreaders(MDB_env * env,unsigned int * readers)4365 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
4366 {
4367 if (!env || !readers)
4368 return EINVAL;
4369 *readers = env->me_maxreaders;
4370 return MDB_SUCCESS;
4371 }
4372
4373 static int ESECT
mdb_fsize(HANDLE fd,mdb_size_t * size)4374 mdb_fsize(HANDLE fd, mdb_size_t *size)
4375 {
4376 #ifdef _WIN32
4377 LARGE_INTEGER fsize;
4378
4379 if (!GetFileSizeEx(fd, &fsize))
4380 return ErrCode();
4381
4382 *size = fsize.QuadPart;
4383 #else
4384 struct stat st;
4385
4386 if (fstat(fd, &st))
4387 return ErrCode();
4388
4389 *size = st.st_size;
4390 #endif
4391 return MDB_SUCCESS;
4392 }
4393
4394 #ifdef BROKEN_FDATASYNC
4395 #include <sys/utsname.h>
4396 #include <sys/vfs.h>
4397 #endif
4398
4399 /** Further setup required for opening an LMDB environment
4400 */
4401 static int ESECT
mdb_env_open2(MDB_env * env)4402 mdb_env_open2(MDB_env *env)
4403 {
4404 unsigned int flags = env->me_flags;
4405 int i, newenv = 0, rc;
4406 MDB_meta meta;
4407
4408 #ifdef _WIN32
4409 /* See if we should use QueryLimited */
4410 rc = GetVersion();
4411 if ((rc & 0xff) > 5)
4412 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
4413 else
4414 env->me_pidquery = PROCESS_QUERY_INFORMATION;
4415 #endif /* _WIN32 */
4416
4417 #ifdef BROKEN_FDATASYNC
4418 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
4419 * https://lkml.org/lkml/2012/9/3/83
4420 * Kernels after 3.6-rc6 are known good.
4421 * https://lkml.org/lkml/2012/9/10/556
4422 * See if the DB is on ext3/ext4, then check for new enough kernel
4423 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
4424 * to be patched.
4425 */
4426 {
4427 struct statfs st;
4428 fstatfs(env->me_fd, &st);
4429 while (st.f_type == 0xEF53) {
4430 struct utsname uts;
4431 int i;
4432 uname(&uts);
4433 if (uts.release[0] < '3') {
4434 if (!strncmp(uts.release, "2.6.32.", 7)) {
4435 i = atoi(uts.release+7);
4436 if (i >= 60)
4437 break; /* 2.6.32.60 and newer is OK */
4438 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
4439 i = atoi(uts.release+7);
4440 if (i >= 15)
4441 break; /* 2.6.34.15 and newer is OK */
4442 }
4443 } else if (uts.release[0] == '3') {
4444 i = atoi(uts.release+2);
4445 if (i > 5)
4446 break; /* 3.6 and newer is OK */
4447 if (i == 5) {
4448 i = atoi(uts.release+4);
4449 if (i >= 4)
4450 break; /* 3.5.4 and newer is OK */
4451 } else if (i == 2) {
4452 i = atoi(uts.release+4);
4453 if (i >= 30)
4454 break; /* 3.2.30 and newer is OK */
4455 }
4456 } else { /* 4.x and newer is OK */
4457 break;
4458 }
4459 env->me_flags |= MDB_FSYNCONLY;
4460 break;
4461 }
4462 }
4463 #endif
4464
4465 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4466 if (i != ENOENT)
4467 return i;
4468 DPUTS("new mdbenv");
4469 newenv = 1;
4470 env->me_psize = env->me_os_psize;
4471 if (env->me_psize > MAX_PAGESIZE)
4472 env->me_psize = MAX_PAGESIZE;
4473 memset(&meta, 0, sizeof(meta));
4474 mdb_env_init_meta0(env, &meta);
4475 meta.mm_mapsize = DEFAULT_MAPSIZE;
4476 } else {
4477 env->me_psize = meta.mm_psize;
4478 }
4479
4480 /* Was a mapsize configured? */
4481 if (!env->me_mapsize) {
4482 env->me_mapsize = meta.mm_mapsize;
4483 }
4484 {
4485 /* Make sure mapsize >= committed data size. Even when using
4486 * mm_mapsize, which could be broken in old files (ITS#7789).
4487 */
4488 mdb_size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4489 if (env->me_mapsize < minsize)
4490 env->me_mapsize = minsize;
4491 }
4492 meta.mm_mapsize = env->me_mapsize;
4493
4494 if (newenv && !(flags & MDB_FIXEDMAP)) {
4495 /* mdb_env_map() may grow the datafile. Write the metapages
4496 * first, so the file will be valid if initialization fails.
4497 * Except with FIXEDMAP, since we do not yet know mm_address.
4498 * We could fill in mm_address later, but then a different
4499 * program might end up doing that - one with a memory layout
4500 * and map address which does not suit the main program.
4501 */
4502 rc = mdb_env_init_meta(env, &meta);
4503 if (rc)
4504 return rc;
4505 newenv = 0;
4506 }
4507 #ifdef _WIN32
4508 /* For FIXEDMAP, make sure the file is non-empty before we attempt to map it */
4509 if (newenv) {
4510 char dummy = 0;
4511 DWORD len;
4512 rc = WriteFile(env->me_fd, &dummy, 1, &len, NULL);
4513 if (!rc) {
4514 rc = ErrCode();
4515 return rc;
4516 }
4517 }
4518 #endif
4519
4520 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4521 if (rc)
4522 return rc;
4523
4524 if (newenv) {
4525 if (flags & MDB_FIXEDMAP)
4526 meta.mm_address = env->me_map;
4527 i = mdb_env_init_meta(env, &meta);
4528 if (i != MDB_SUCCESS) {
4529 return i;
4530 }
4531 }
4532
4533 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4534 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4535 - sizeof(indx_t);
4536 #if !(MDB_MAXKEYSIZE)
4537 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4538 #endif
4539 env->me_maxpg = env->me_mapsize / env->me_psize;
4540
4541 #if MDB_DEBUG
4542 {
4543 MDB_meta *meta = mdb_env_pick_meta(env);
4544 MDB_db *db = &meta->mm_dbs[MAIN_DBI];
4545
4546 DPRINTF(("opened database version %u, pagesize %u",
4547 meta->mm_version, env->me_psize));
4548 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1)));
4549 DPRINTF(("depth: %u", db->md_depth));
4550 DPRINTF(("entries: %"Yu, db->md_entries));
4551 DPRINTF(("branch pages: %"Yu, db->md_branch_pages));
4552 DPRINTF(("leaf pages: %"Yu, db->md_leaf_pages));
4553 DPRINTF(("overflow pages: %"Yu, db->md_overflow_pages));
4554 DPRINTF(("root: %"Yu, db->md_root));
4555 }
4556 #endif
4557
4558 return MDB_SUCCESS;
4559 }
4560
4561
4562 /** Release a reader thread's slot in the reader lock table.
4563 * This function is called automatically when a thread exits.
4564 * @param[in] ptr This points to the slot in the reader lock table.
4565 */
4566 static void
mdb_env_reader_dest(void * ptr)4567 mdb_env_reader_dest(void *ptr)
4568 {
4569 MDB_reader *reader = ptr;
4570
4571 reader->mr_pid = 0;
4572 }
4573
4574 #ifdef _WIN32
4575 /** Junk for arranging thread-specific callbacks on Windows. This is
4576 * necessarily platform and compiler-specific. Windows supports up
4577 * to 1088 keys. Let's assume nobody opens more than 64 environments
4578 * in a single process, for now. They can override this if needed.
4579 */
4580 #ifndef MAX_TLS_KEYS
4581 #define MAX_TLS_KEYS 64
4582 #endif
4583 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4584 static int mdb_tls_nkeys;
4585
mdb_tls_callback(PVOID module,DWORD reason,PVOID ptr)4586 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4587 {
4588 int i;
4589 switch(reason) {
4590 case DLL_PROCESS_ATTACH: break;
4591 case DLL_THREAD_ATTACH: break;
4592 case DLL_THREAD_DETACH:
4593 for (i=0; i<mdb_tls_nkeys; i++) {
4594 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4595 if (r) {
4596 mdb_env_reader_dest(r);
4597 }
4598 }
4599 break;
4600 case DLL_PROCESS_DETACH: break;
4601 }
4602 }
4603 #ifdef __GNUC__
4604 #ifdef _WIN64
4605 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4606 #else
4607 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4608 #endif
4609 #else
4610 #ifdef _WIN64
4611 /* Force some symbol references.
4612 * _tls_used forces the linker to create the TLS directory if not already done
4613 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4614 */
4615 #pragma comment(linker, "/INCLUDE:_tls_used")
4616 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4617 #pragma const_seg(".CRT$XLB")
4618 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4619 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4620 #pragma const_seg()
4621 #else /* _WIN32 */
4622 #pragma comment(linker, "/INCLUDE:__tls_used")
4623 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4624 #pragma data_seg(".CRT$XLB")
4625 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4626 #pragma data_seg()
4627 #endif /* WIN 32/64 */
4628 #endif /* !__GNUC__ */
4629 #endif
4630
4631 /** Downgrade the exclusive lock on the region back to shared */
4632 static int ESECT
mdb_env_share_locks(MDB_env * env,int * excl)4633 mdb_env_share_locks(MDB_env *env, int *excl)
4634 {
4635 int rc = 0;
4636 MDB_meta *meta = mdb_env_pick_meta(env);
4637
4638 env->me_txns->mti_txnid = meta->mm_txnid;
4639
4640 #ifdef _WIN32
4641 {
4642 OVERLAPPED ov;
4643 /* First acquire a shared lock. The Unlock will
4644 * then release the existing exclusive lock.
4645 */
4646 memset(&ov, 0, sizeof(ov));
4647 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4648 rc = ErrCode();
4649 } else {
4650 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4651 *excl = 0;
4652 }
4653 }
4654 #else
4655 {
4656 struct flock lock_info;
4657 /* The shared lock replaces the existing lock */
4658 memset((void *)&lock_info, 0, sizeof(lock_info));
4659 lock_info.l_type = F_RDLCK;
4660 lock_info.l_whence = SEEK_SET;
4661 lock_info.l_start = 0;
4662 lock_info.l_len = 1;
4663 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4664 (rc = ErrCode()) == EINTR) ;
4665 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4666 }
4667 #endif
4668
4669 return rc;
4670 }
4671
4672 /** Try to get exclusive lock, otherwise shared.
4673 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4674 */
4675 static int ESECT
mdb_env_excl_lock(MDB_env * env,int * excl)4676 mdb_env_excl_lock(MDB_env *env, int *excl)
4677 {
4678 int rc = 0;
4679 #ifdef _WIN32
4680 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4681 *excl = 1;
4682 } else {
4683 OVERLAPPED ov;
4684 memset(&ov, 0, sizeof(ov));
4685 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4686 *excl = 0;
4687 } else {
4688 rc = ErrCode();
4689 }
4690 }
4691 #else
4692 struct flock lock_info;
4693 memset((void *)&lock_info, 0, sizeof(lock_info));
4694 lock_info.l_type = F_WRLCK;
4695 lock_info.l_whence = SEEK_SET;
4696 lock_info.l_start = 0;
4697 lock_info.l_len = 1;
4698 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4699 (rc = ErrCode()) == EINTR) ;
4700 if (!rc) {
4701 *excl = 1;
4702 } else
4703 # ifndef MDB_USE_POSIX_MUTEX
4704 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */
4705 # endif
4706 {
4707 lock_info.l_type = F_RDLCK;
4708 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4709 (rc = ErrCode()) == EINTR) ;
4710 if (rc == 0)
4711 *excl = 0;
4712 }
4713 #endif
4714 return rc;
4715 }
4716
4717 #ifdef MDB_USE_HASH
4718 /*
4719 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4720 *
4721 * @(#) $Revision: 5.1 $
4722 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4723 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4724 *
4725 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4726 *
4727 ***
4728 *
4729 * Please do not copyright this code. This code is in the public domain.
4730 *
4731 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4732 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4733 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4734 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4735 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4736 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4737 * PERFORMANCE OF THIS SOFTWARE.
4738 *
4739 * By:
4740 * chongo <Landon Curt Noll> /\oo/\
4741 * http://www.isthe.com/chongo/
4742 *
4743 * Share and Enjoy! :-)
4744 */
4745
4746 typedef unsigned long long mdb_hash_t;
4747 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4748
4749 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4750 * @param[in] val value to hash
4751 * @param[in] hval initial value for hash
4752 * @return 64 bit hash
4753 *
4754 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4755 * hval arg on the first call.
4756 */
4757 static mdb_hash_t
mdb_hash_val(MDB_val * val,mdb_hash_t hval)4758 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4759 {
4760 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4761 unsigned char *end = s + val->mv_size;
4762 /*
4763 * FNV-1a hash each octet of the string
4764 */
4765 while (s < end) {
4766 /* xor the bottom with the current octet */
4767 hval ^= (mdb_hash_t)*s++;
4768
4769 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4770 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4771 (hval << 7) + (hval << 8) + (hval << 40);
4772 }
4773 /* return our new hash value */
4774 return hval;
4775 }
4776
4777 /** Hash the string and output the encoded hash.
4778 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4779 * very short name limits. We don't care about the encoding being reversible,
4780 * we just want to preserve as many bits of the input as possible in a
4781 * small printable string.
4782 * @param[in] str string to hash
4783 * @param[out] encbuf an array of 11 chars to hold the hash
4784 */
4785 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4786
4787 static void ESECT
mdb_pack85(unsigned long l,char * out)4788 mdb_pack85(unsigned long l, char *out)
4789 {
4790 int i;
4791
4792 for (i=0; i<5; i++) {
4793 *out++ = mdb_a85[l % 85];
4794 l /= 85;
4795 }
4796 }
4797
4798 static void ESECT
mdb_hash_enc(MDB_val * val,char * encbuf)4799 mdb_hash_enc(MDB_val *val, char *encbuf)
4800 {
4801 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4802
4803 mdb_pack85(h, encbuf);
4804 mdb_pack85(h>>32, encbuf+5);
4805 encbuf[10] = '\0';
4806 }
4807 #endif
4808
4809 /** Open and/or initialize the lock region for the environment.
4810 * @param[in] env The LMDB environment.
4811 * @param[in] lpath The pathname of the file used for the lock region.
4812 * @param[in] mode The Unix permissions for the file, if we create it.
4813 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4814 * @return 0 on success, non-zero on failure.
4815 */
4816 static int ESECT
mdb_env_setup_locks(MDB_env * env,char * lpath,int mode,int * excl)4817 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4818 {
4819 #ifdef _WIN32
4820 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4821 #else
4822 # define MDB_ERRCODE_ROFS EROFS
4823 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4824 # define MDB_CLOEXEC O_CLOEXEC
4825 #else
4826 int fdflags;
4827 # define MDB_CLOEXEC 0
4828 #endif
4829 #endif
4830 #ifdef MDB_USE_SYSV_SEM
4831 int semid;
4832 union semun semu;
4833 #endif
4834 int rc;
4835 off_t size, rsize;
4836
4837 #ifdef _WIN32
4838 wchar_t *wlpath;
4839 rc = utf8_to_utf16(lpath, -1, &wlpath, NULL);
4840 if (rc)
4841 return rc;
4842 env->me_lfd = CreateFileW(wlpath, GENERIC_READ|GENERIC_WRITE,
4843 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4844 FILE_ATTRIBUTE_NORMAL, NULL);
4845 free(wlpath);
4846 #else
4847 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4848 #endif
4849 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4850 rc = ErrCode();
4851 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4852 return MDB_SUCCESS;
4853 }
4854 goto fail_errno;
4855 }
4856 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4857 /* Lose record locks when exec*() */
4858 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4859 fcntl(env->me_lfd, F_SETFD, fdflags);
4860 #endif
4861
4862 if (!(env->me_flags & MDB_NOTLS)) {
4863 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4864 if (rc)
4865 goto fail;
4866 env->me_flags |= MDB_ENV_TXKEY;
4867 #ifdef _WIN32
4868 /* Windows TLS callbacks need help finding their TLS info. */
4869 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4870 rc = MDB_TLS_FULL;
4871 goto fail;
4872 }
4873 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4874 #endif
4875 }
4876
4877 /* Try to get exclusive lock. If we succeed, then
4878 * nobody is using the lock region and we should initialize it.
4879 */
4880 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4881
4882 #ifdef _WIN32
4883 size = GetFileSize(env->me_lfd, NULL);
4884 #else
4885 size = lseek(env->me_lfd, 0, SEEK_END);
4886 if (size == -1) goto fail_errno;
4887 #endif
4888 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4889 if (size < rsize && *excl > 0) {
4890 #ifdef _WIN32
4891 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4892 || !SetEndOfFile(env->me_lfd))
4893 goto fail_errno;
4894 #else
4895 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4896 #endif
4897 } else {
4898 rsize = size;
4899 size = rsize - sizeof(MDB_txninfo);
4900 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4901 }
4902 {
4903 #ifdef _WIN32
4904 HANDLE mh;
4905 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4906 0, 0, NULL);
4907 if (!mh) goto fail_errno;
4908 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4909 CloseHandle(mh);
4910 if (!env->me_txns) goto fail_errno;
4911 #else
4912 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4913 env->me_lfd, 0);
4914 if (m == MAP_FAILED) goto fail_errno;
4915 env->me_txns = m;
4916 #endif
4917 }
4918 if (*excl > 0) {
4919 #ifdef _WIN32
4920 BY_HANDLE_FILE_INFORMATION stbuf;
4921 struct {
4922 DWORD volume;
4923 DWORD nhigh;
4924 DWORD nlow;
4925 } idbuf;
4926 MDB_val val;
4927 char encbuf[11];
4928
4929 if (!mdb_sec_inited) {
4930 InitializeSecurityDescriptor(&mdb_null_sd,
4931 SECURITY_DESCRIPTOR_REVISION);
4932 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4933 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4934 mdb_all_sa.bInheritHandle = FALSE;
4935 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4936 mdb_sec_inited = 1;
4937 }
4938 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4939 idbuf.volume = stbuf.dwVolumeSerialNumber;
4940 idbuf.nhigh = stbuf.nFileIndexHigh;
4941 idbuf.nlow = stbuf.nFileIndexLow;
4942 val.mv_data = &idbuf;
4943 val.mv_size = sizeof(idbuf);
4944 mdb_hash_enc(&val, encbuf);
4945 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4946 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4947 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4948 if (!env->me_rmutex) goto fail_errno;
4949 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4950 if (!env->me_wmutex) goto fail_errno;
4951 #elif defined(MDB_USE_POSIX_SEM)
4952 struct stat stbuf;
4953 struct {
4954 dev_t dev;
4955 ino_t ino;
4956 } idbuf;
4957 MDB_val val;
4958 char encbuf[11];
4959
4960 #if defined(__NetBSD__)
4961 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4962 #endif
4963 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4964 idbuf.dev = stbuf.st_dev;
4965 idbuf.ino = stbuf.st_ino;
4966 val.mv_data = &idbuf;
4967 val.mv_size = sizeof(idbuf);
4968 mdb_hash_enc(&val, encbuf);
4969 #ifdef MDB_SHORT_SEMNAMES
4970 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4971 #endif
4972 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4973 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4974 /* Clean up after a previous run, if needed: Try to
4975 * remove both semaphores before doing anything else.
4976 */
4977 sem_unlink(env->me_txns->mti_rmname);
4978 sem_unlink(env->me_txns->mti_wmname);
4979 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4980 O_CREAT|O_EXCL, mode, 1);
4981 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4982 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4983 O_CREAT|O_EXCL, mode, 1);
4984 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4985 #elif defined(MDB_USE_SYSV_SEM)
4986 unsigned short vals[2] = {1, 1};
4987 key_t key = ftok(lpath, 'M');
4988 if (key == -1)
4989 goto fail_errno;
4990 semid = semget(key, 2, (mode & 0777) | IPC_CREAT);
4991 if (semid < 0)
4992 goto fail_errno;
4993 semu.array = vals;
4994 if (semctl(semid, 0, SETALL, semu) < 0)
4995 goto fail_errno;
4996 env->me_txns->mti_semid = semid;
4997 env->me_txns->mti_rlocked = 0;
4998 env->me_txns->mti_wlocked = 0;
4999 #else /* MDB_USE_POSIX_MUTEX: */
5000 pthread_mutexattr_t mattr;
5001
5002 /* Solaris needs this before initing a robust mutex. Otherwise
5003 * it may skip the init and return EBUSY "seems someone already
5004 * inited" or EINVAL "it was inited differently".
5005 */
5006 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex));
5007 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex));
5008
5009 if ((rc = pthread_mutexattr_init(&mattr)) != 0)
5010 goto fail;
5011 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
5012 #ifdef MDB_ROBUST_SUPPORTED
5013 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
5014 #endif
5015 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr);
5016 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr);
5017 pthread_mutexattr_destroy(&mattr);
5018 if (rc)
5019 goto fail;
5020 #endif /* _WIN32 || ... */
5021
5022 env->me_txns->mti_magic = MDB_MAGIC;
5023 env->me_txns->mti_format = MDB_LOCK_FORMAT;
5024 env->me_txns->mti_txnid = 0;
5025 env->me_txns->mti_numreaders = 0;
5026
5027 } else {
5028 #ifdef MDB_USE_SYSV_SEM
5029 struct semid_ds buf;
5030 #endif
5031 if (env->me_txns->mti_magic != MDB_MAGIC) {
5032 DPUTS("lock region has invalid magic");
5033 rc = MDB_INVALID;
5034 goto fail;
5035 }
5036 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
5037 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
5038 env->me_txns->mti_format, MDB_LOCK_FORMAT));
5039 rc = MDB_VERSION_MISMATCH;
5040 goto fail;
5041 }
5042 rc = ErrCode();
5043 if (rc && rc != EACCES && rc != EAGAIN) {
5044 goto fail;
5045 }
5046 #ifdef _WIN32
5047 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
5048 if (!env->me_rmutex) goto fail_errno;
5049 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
5050 if (!env->me_wmutex) goto fail_errno;
5051 #elif defined(MDB_USE_POSIX_SEM)
5052 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
5053 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
5054 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
5055 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
5056 #elif defined(MDB_USE_SYSV_SEM)
5057 semid = env->me_txns->mti_semid;
5058 semu.buf = &buf;
5059 /* check for read access */
5060 if (semctl(semid, 0, IPC_STAT, semu) < 0)
5061 goto fail_errno;
5062 /* check for write access */
5063 if (semctl(semid, 0, IPC_SET, semu) < 0)
5064 goto fail_errno;
5065 #endif
5066 }
5067 #ifdef MDB_USE_SYSV_SEM
5068 env->me_rmutex->semid = semid;
5069 env->me_wmutex->semid = semid;
5070 env->me_rmutex->semnum = 0;
5071 env->me_wmutex->semnum = 1;
5072 env->me_rmutex->locked = &env->me_txns->mti_rlocked;
5073 env->me_wmutex->locked = &env->me_txns->mti_wlocked;
5074 #endif
5075 #ifdef MDB_VL32
5076 #ifdef _WIN32
5077 env->me_rpmutex = CreateMutex(NULL, FALSE, NULL);
5078 #else
5079 pthread_mutex_init(&env->me_rpmutex, NULL);
5080 #endif
5081 #endif
5082
5083 return MDB_SUCCESS;
5084
5085 fail_errno:
5086 rc = ErrCode();
5087 fail:
5088 return rc;
5089 }
5090
5091 /** The name of the lock file in the DB environment */
5092 #define LOCKNAME "/lock.mdb"
5093 /** The name of the data file in the DB environment */
5094 #define DATANAME "/data.mdb"
5095 /** The suffix of the lock file when no subdir is used */
5096 #define LOCKSUFF "-lock"
5097 /** Only a subset of the @ref mdb_env flags can be changed
5098 * at runtime. Changing other flags requires closing the
5099 * environment and re-opening it with the new flags.
5100 */
5101 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
5102 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
5103 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
5104
5105 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
5106 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
5107 #endif
5108
5109 int ESECT
mdb_env_open(MDB_env * env,const char * path,unsigned int flags,mdb_mode_t mode)5110 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
5111 {
5112 int oflags, rc, len, excl = -1;
5113 char *lpath, *dpath;
5114 #ifdef _WIN32
5115 wchar_t *wpath;
5116 #endif
5117
5118 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
5119 return EINVAL;
5120
5121 #ifdef MDB_VL32
5122 if (flags & MDB_WRITEMAP) {
5123 /* silently ignore WRITEMAP in 32 bit mode */
5124 flags ^= MDB_WRITEMAP;
5125 }
5126 if (flags & MDB_FIXEDMAP) {
5127 /* cannot support FIXEDMAP */
5128 return EINVAL;
5129 }
5130 #endif
5131
5132 len = strlen(path);
5133 if (flags & MDB_NOSUBDIR) {
5134 rc = len + sizeof(LOCKSUFF) + len + 1;
5135 } else {
5136 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
5137 }
5138 lpath = malloc(rc);
5139 if (!lpath)
5140 return ENOMEM;
5141 if (flags & MDB_NOSUBDIR) {
5142 dpath = lpath + len + sizeof(LOCKSUFF);
5143 sprintf(lpath, "%s" LOCKSUFF, path);
5144 strcpy(dpath, path);
5145 } else {
5146 dpath = lpath + len + sizeof(LOCKNAME);
5147 sprintf(lpath, "%s" LOCKNAME, path);
5148 sprintf(dpath, "%s" DATANAME, path);
5149 }
5150
5151 rc = MDB_SUCCESS;
5152 flags |= env->me_flags;
5153 if (flags & MDB_RDONLY) {
5154 /* silently ignore WRITEMAP when we're only getting read access */
5155 flags &= ~MDB_WRITEMAP;
5156 } else {
5157 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
5158 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
5159 rc = ENOMEM;
5160 }
5161 #ifdef MDB_VL32
5162 if (!rc) {
5163 env->me_rpages = malloc(MDB_ERPAGE_SIZE * sizeof(MDB_ID3));
5164 if (!env->me_rpages) {
5165 rc = ENOMEM;
5166 goto leave;
5167 }
5168 env->me_rpages[0].mid = 0;
5169 env->me_rpcheck = MDB_ERPAGE_SIZE/2;
5170 }
5171 #endif
5172 env->me_flags = flags |= MDB_ENV_ACTIVE;
5173 if (rc)
5174 goto leave;
5175
5176 env->me_path = strdup(path);
5177 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
5178 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
5179 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
5180 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
5181 rc = ENOMEM;
5182 goto leave;
5183 }
5184 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */
5185
5186 /* For RDONLY, get lockfile after we know datafile exists */
5187 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
5188 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
5189 if (rc)
5190 goto leave;
5191 }
5192
5193 #ifdef _WIN32
5194 if (F_ISSET(flags, MDB_RDONLY)) {
5195 oflags = GENERIC_READ;
5196 len = OPEN_EXISTING;
5197 } else {
5198 oflags = GENERIC_READ|GENERIC_WRITE;
5199 len = OPEN_ALWAYS;
5200 }
5201 mode = FILE_ATTRIBUTE_NORMAL;
5202 rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
5203 if (rc)
5204 goto leave;
5205 env->me_fd = CreateFileW(wpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
5206 NULL, len, mode, NULL);
5207 free(wpath);
5208 #else
5209 if (F_ISSET(flags, MDB_RDONLY))
5210 oflags = O_RDONLY;
5211 else
5212 oflags = O_RDWR | O_CREAT;
5213
5214 env->me_fd = open(dpath, oflags, mode);
5215 #endif
5216 if (env->me_fd == INVALID_HANDLE_VALUE) {
5217 rc = ErrCode();
5218 goto leave;
5219 }
5220
5221 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
5222 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
5223 if (rc)
5224 goto leave;
5225 }
5226
5227 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
5228 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
5229 env->me_mfd = env->me_fd;
5230 } else {
5231 /* Synchronous fd for meta writes. Needed even with
5232 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
5233 */
5234 #ifdef _WIN32
5235 len = OPEN_EXISTING;
5236 rc = utf8_to_utf16(dpath, -1, &wpath, NULL);
5237 if (rc)
5238 goto leave;
5239 env->me_mfd = CreateFileW(wpath, oflags,
5240 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
5241 mode | FILE_FLAG_WRITE_THROUGH, NULL);
5242 free(wpath);
5243 #else
5244 oflags &= ~O_CREAT;
5245 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
5246 #endif
5247 if (env->me_mfd == INVALID_HANDLE_VALUE) {
5248 rc = ErrCode();
5249 goto leave;
5250 }
5251 }
5252 DPRINTF(("opened dbenv %p", (void *) env));
5253 if (excl > 0) {
5254 rc = mdb_env_share_locks(env, &excl);
5255 if (rc)
5256 goto leave;
5257 }
5258 if (!(flags & MDB_RDONLY)) {
5259 MDB_txn *txn;
5260 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
5261 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
5262 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
5263 (txn = calloc(1, size)))
5264 {
5265 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
5266 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
5267 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
5268 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
5269 txn->mt_env = env;
5270 #ifdef MDB_VL32
5271 txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
5272 if (!txn->mt_rpages) {
5273 free(txn);
5274 rc = ENOMEM;
5275 goto leave;
5276 }
5277 txn->mt_rpages[0].mid = 0;
5278 txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
5279 #endif
5280 txn->mt_dbxs = env->me_dbxs;
5281 txn->mt_flags = MDB_TXN_FINISHED;
5282 env->me_txn0 = txn;
5283 } else {
5284 rc = ENOMEM;
5285 }
5286 }
5287 }
5288
5289 leave:
5290 if (rc) {
5291 mdb_env_close0(env, excl);
5292 }
5293 free(lpath);
5294 return rc;
5295 }
5296
5297 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
5298 static void ESECT
mdb_env_close0(MDB_env * env,int excl)5299 mdb_env_close0(MDB_env *env, int excl)
5300 {
5301 int i;
5302
5303 if (!(env->me_flags & MDB_ENV_ACTIVE))
5304 return;
5305
5306 /* Doing this here since me_dbxs may not exist during mdb_env_close */
5307 if (env->me_dbxs) {
5308 for (i = env->me_maxdbs; --i >= CORE_DBS; )
5309 free(env->me_dbxs[i].md_name.mv_data);
5310 free(env->me_dbxs);
5311 }
5312
5313 free(env->me_pbuf);
5314 free(env->me_dbiseqs);
5315 free(env->me_dbflags);
5316 free(env->me_path);
5317 free(env->me_dirty_list);
5318 #ifdef MDB_VL32
5319 if (env->me_txn0 && env->me_txn0->mt_rpages)
5320 free(env->me_txn0->mt_rpages);
5321 { unsigned int x;
5322 for (x=1; x<=env->me_rpages[0].mid; x++)
5323 munmap(env->me_rpages[x].mptr, env->me_rpages[x].mcnt * env->me_psize);
5324 }
5325 free(env->me_rpages);
5326 #endif
5327 free(env->me_txn0);
5328 mdb_midl_free(env->me_free_pgs);
5329
5330 if (env->me_flags & MDB_ENV_TXKEY) {
5331 pthread_key_delete(env->me_txkey);
5332 #ifdef _WIN32
5333 /* Delete our key from the global list */
5334 for (i=0; i<mdb_tls_nkeys; i++)
5335 if (mdb_tls_keys[i] == env->me_txkey) {
5336 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
5337 mdb_tls_nkeys--;
5338 break;
5339 }
5340 #endif
5341 }
5342
5343 if (env->me_map) {
5344 #ifdef MDB_VL32
5345 munmap(env->me_map, NUM_METAS*env->me_psize);
5346 #else
5347 munmap(env->me_map, env->me_mapsize);
5348 #endif
5349 }
5350 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
5351 (void) close(env->me_mfd);
5352 if (env->me_fd != INVALID_HANDLE_VALUE)
5353 (void) close(env->me_fd);
5354 if (env->me_txns) {
5355 MDB_PID_T pid = env->me_pid;
5356 /* Clearing readers is done in this function because
5357 * me_txkey with its destructor must be disabled first.
5358 *
5359 * We skip the the reader mutex, so we touch only
5360 * data owned by this process (me_close_readers and
5361 * our readers), and clear each reader atomically.
5362 */
5363 for (i = env->me_close_readers; --i >= 0; )
5364 if (env->me_txns->mti_readers[i].mr_pid == pid)
5365 env->me_txns->mti_readers[i].mr_pid = 0;
5366 #ifdef _WIN32
5367 if (env->me_rmutex) {
5368 CloseHandle(env->me_rmutex);
5369 if (env->me_wmutex) CloseHandle(env->me_wmutex);
5370 }
5371 /* Windows automatically destroys the mutexes when
5372 * the last handle closes.
5373 */
5374 #elif defined(MDB_USE_POSIX_SEM)
5375 if (env->me_rmutex != SEM_FAILED) {
5376 sem_close(env->me_rmutex);
5377 if (env->me_wmutex != SEM_FAILED)
5378 sem_close(env->me_wmutex);
5379 /* If we have the filelock: If we are the
5380 * only remaining user, clean up semaphores.
5381 */
5382 if (excl == 0)
5383 mdb_env_excl_lock(env, &excl);
5384 if (excl > 0) {
5385 sem_unlink(env->me_txns->mti_rmname);
5386 sem_unlink(env->me_txns->mti_wmname);
5387 }
5388 }
5389 #elif defined(MDB_USE_SYSV_SEM)
5390 if (env->me_rmutex->semid != -1) {
5391 /* If we have the filelock: If we are the
5392 * only remaining user, clean up semaphores.
5393 */
5394 if (excl == 0)
5395 mdb_env_excl_lock(env, &excl);
5396 if (excl > 0)
5397 semctl(env->me_rmutex->semid, 0, IPC_RMID);
5398 }
5399 #endif
5400 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
5401 }
5402 if (env->me_lfd != INVALID_HANDLE_VALUE) {
5403 #ifdef _WIN32
5404 if (excl >= 0) {
5405 /* Unlock the lockfile. Windows would have unlocked it
5406 * after closing anyway, but not necessarily at once.
5407 */
5408 UnlockFile(env->me_lfd, 0, 0, 1, 0);
5409 }
5410 #endif
5411 (void) close(env->me_lfd);
5412 }
5413 #ifdef MDB_VL32
5414 #ifdef _WIN32
5415 if (env->me_fmh) CloseHandle(env->me_fmh);
5416 if (env->me_rpmutex) CloseHandle(env->me_rpmutex);
5417 #else
5418 pthread_mutex_destroy(&env->me_rpmutex);
5419 #endif
5420 #endif
5421
5422 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
5423 }
5424
5425 void ESECT
mdb_env_close(MDB_env * env)5426 mdb_env_close(MDB_env *env)
5427 {
5428 MDB_page *dp;
5429
5430 if (env == NULL)
5431 return;
5432
5433 VGMEMP_DESTROY(env);
5434 while ((dp = env->me_dpages) != NULL) {
5435 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
5436 env->me_dpages = dp->mp_next;
5437 free(dp);
5438 }
5439
5440 mdb_env_close0(env, 0);
5441 free(env);
5442 }
5443
5444 /** Compare two items pointing at aligned #mdb_size_t's */
5445 static int
mdb_cmp_long(const MDB_val * a,const MDB_val * b)5446 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
5447 {
5448 return (*(mdb_size_t *)a->mv_data < *(mdb_size_t *)b->mv_data) ? -1 :
5449 *(mdb_size_t *)a->mv_data > *(mdb_size_t *)b->mv_data;
5450 }
5451
5452 /** Compare two items pointing at aligned unsigned int's.
5453 *
5454 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
5455 * but #mdb_cmp_clong() is called instead if the data type is #mdb_size_t.
5456 */
5457 static int
mdb_cmp_int(const MDB_val * a,const MDB_val * b)5458 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
5459 {
5460 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
5461 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
5462 }
5463
5464 /** Compare two items pointing at unsigned ints of unknown alignment.
5465 * Nodes and keys are guaranteed to be 2-byte aligned.
5466 */
5467 static int
mdb_cmp_cint(const MDB_val * a,const MDB_val * b)5468 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
5469 {
5470 #if BYTE_ORDER == LITTLE_ENDIAN
5471 unsigned short *u, *c;
5472 int x;
5473
5474 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5475 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
5476 do {
5477 x = *--u - *--c;
5478 } while(!x && u > (unsigned short *)a->mv_data);
5479 return x;
5480 #else
5481 unsigned short *u, *c, *end;
5482 int x;
5483
5484 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
5485 u = (unsigned short *)a->mv_data;
5486 c = (unsigned short *)b->mv_data;
5487 do {
5488 x = *u++ - *c++;
5489 } while(!x && u < end);
5490 return x;
5491 #endif
5492 }
5493
5494 /** Compare two items lexically */
5495 static int
mdb_cmp_memn(const MDB_val * a,const MDB_val * b)5496 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
5497 {
5498 int diff;
5499 ssize_t len_diff;
5500 unsigned int len;
5501
5502 len = a->mv_size;
5503 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5504 if (len_diff > 0) {
5505 len = b->mv_size;
5506 len_diff = 1;
5507 }
5508
5509 diff = memcmp(a->mv_data, b->mv_data, len);
5510 return diff ? diff : len_diff<0 ? -1 : len_diff;
5511 }
5512
5513 /** Compare two items in reverse byte order */
5514 static int
mdb_cmp_memnr(const MDB_val * a,const MDB_val * b)5515 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
5516 {
5517 const unsigned char *p1, *p2, *p1_lim;
5518 ssize_t len_diff;
5519 int diff;
5520
5521 p1_lim = (const unsigned char *)a->mv_data;
5522 p1 = (const unsigned char *)a->mv_data + a->mv_size;
5523 p2 = (const unsigned char *)b->mv_data + b->mv_size;
5524
5525 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
5526 if (len_diff > 0) {
5527 p1_lim += len_diff;
5528 len_diff = 1;
5529 }
5530
5531 while (p1 > p1_lim) {
5532 diff = *--p1 - *--p2;
5533 if (diff)
5534 return diff;
5535 }
5536 return len_diff<0 ? -1 : len_diff;
5537 }
5538
5539 /** Search for key within a page, using binary search.
5540 * Returns the smallest entry larger or equal to the key.
5541 * If exactp is non-null, stores whether the found entry was an exact match
5542 * in *exactp (1 or 0).
5543 * Updates the cursor index with the index of the found entry.
5544 * If no entry larger or equal to the key is found, returns NULL.
5545 */
5546 static MDB_node *
mdb_node_search(MDB_cursor * mc,MDB_val * key,int * exactp)5547 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
5548 {
5549 unsigned int i = 0, nkeys;
5550 int low, high;
5551 int rc = 0;
5552 MDB_page *mp = mc->mc_pg[mc->mc_top];
5553 MDB_node *node = NULL;
5554 MDB_val nodekey;
5555 MDB_cmp_func *cmp;
5556 DKBUF;
5557
5558 nkeys = NUMKEYS(mp);
5559
5560 DPRINTF(("searching %u keys in %s %spage %"Yu,
5561 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
5562 mdb_dbg_pgno(mp)));
5563
5564 low = IS_LEAF(mp) ? 0 : 1;
5565 high = nkeys - 1;
5566 cmp = mc->mc_dbx->md_cmp;
5567
5568 /* Branch pages have no data, so if using integer keys,
5569 * alignment is guaranteed. Use faster mdb_cmp_int.
5570 */
5571 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
5572 if (NODEPTR(mp, 1)->mn_ksize == sizeof(mdb_size_t))
5573 cmp = mdb_cmp_long;
5574 else
5575 cmp = mdb_cmp_int;
5576 }
5577
5578 if (IS_LEAF2(mp)) {
5579 nodekey.mv_size = mc->mc_db->md_pad;
5580 node = NODEPTR(mp, 0); /* fake */
5581 while (low <= high) {
5582 i = (low + high) >> 1;
5583 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
5584 rc = cmp(key, &nodekey);
5585 DPRINTF(("found leaf index %u [%s], rc = %i",
5586 i, DKEY(&nodekey), rc));
5587 if (rc == 0)
5588 break;
5589 if (rc > 0)
5590 low = i + 1;
5591 else
5592 high = i - 1;
5593 }
5594 } else {
5595 while (low <= high) {
5596 i = (low + high) >> 1;
5597
5598 node = NODEPTR(mp, i);
5599 nodekey.mv_size = NODEKSZ(node);
5600 nodekey.mv_data = NODEKEY(node);
5601
5602 rc = cmp(key, &nodekey);
5603 #if MDB_DEBUG
5604 if (IS_LEAF(mp))
5605 DPRINTF(("found leaf index %u [%s], rc = %i",
5606 i, DKEY(&nodekey), rc));
5607 else
5608 DPRINTF(("found branch index %u [%s -> %"Yu"], rc = %i",
5609 i, DKEY(&nodekey), NODEPGNO(node), rc));
5610 #endif
5611 if (rc == 0)
5612 break;
5613 if (rc > 0)
5614 low = i + 1;
5615 else
5616 high = i - 1;
5617 }
5618 }
5619
5620 if (rc > 0) { /* Found entry is less than the key. */
5621 i++; /* Skip to get the smallest entry larger than key. */
5622 if (!IS_LEAF2(mp))
5623 node = NODEPTR(mp, i);
5624 }
5625 if (exactp)
5626 *exactp = (rc == 0 && nkeys > 0);
5627 /* store the key index */
5628 mc->mc_ki[mc->mc_top] = i;
5629 if (i >= nkeys)
5630 /* There is no entry larger or equal to the key. */
5631 return NULL;
5632
5633 /* nodeptr is fake for LEAF2 */
5634 return node;
5635 }
5636
5637 #if 0
5638 static void
5639 mdb_cursor_adjust(MDB_cursor *mc, func)
5640 {
5641 MDB_cursor *m2;
5642
5643 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5644 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5645 func(mc, m2);
5646 }
5647 }
5648 }
5649 #endif
5650
5651 /** Pop a page off the top of the cursor's stack. */
5652 static void
mdb_cursor_pop(MDB_cursor * mc)5653 mdb_cursor_pop(MDB_cursor *mc)
5654 {
5655 if (mc->mc_snum) {
5656 DPRINTF(("popping page %"Yu" off db %d cursor %p",
5657 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc));
5658
5659 mc->mc_snum--;
5660 if (mc->mc_snum) {
5661 mc->mc_top--;
5662 } else {
5663 mc->mc_flags &= ~C_INITIALIZED;
5664 }
5665 }
5666 }
5667
5668 /** Push a page onto the top of the cursor's stack. */
5669 static int
mdb_cursor_push(MDB_cursor * mc,MDB_page * mp)5670 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5671 {
5672 DPRINTF(("pushing page %"Yu" on db %d cursor %p", mp->mp_pgno,
5673 DDBI(mc), (void *) mc));
5674
5675 if (mc->mc_snum >= CURSOR_STACK) {
5676 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5677 return MDB_CURSOR_FULL;
5678 }
5679
5680 mc->mc_top = mc->mc_snum++;
5681 mc->mc_pg[mc->mc_top] = mp;
5682 mc->mc_ki[mc->mc_top] = 0;
5683
5684 return MDB_SUCCESS;
5685 }
5686
5687 #ifdef MDB_VL32
5688 /** Map a read-only page.
5689 * There are two levels of tracking in use, a per-txn list and a per-env list.
5690 * ref'ing and unref'ing the per-txn list is faster since it requires no
5691 * locking. Pages are cached in the per-env list for global reuse, and a lock
5692 * is required. Pages are not immediately unmapped when their refcnt goes to
5693 * zero; they hang around in case they will be reused again soon.
5694 *
5695 * When the per-txn list gets full, all pages with refcnt=0 are purged from the
5696 * list and their refcnts in the per-env list are decremented.
5697 *
5698 * When the per-env list gets full, all pages with refcnt=0 are purged from the
5699 * list and their pages are unmapped.
5700 *
5701 * @note "full" means the list has reached its respective rpcheck threshold.
5702 * This threshold slowly raises if no pages could be purged on a given check,
5703 * and returns to its original value when enough pages were purged.
5704 *
5705 * If purging doesn't free any slots, filling the per-txn list will return
5706 * MDB_TXN_FULL, and filling the per-env list returns MDB_MAP_FULL.
5707 *
5708 * Reference tracking in a txn is imperfect, pages can linger with non-zero
5709 * refcnt even without active references. It was deemed to be too invasive
5710 * to add unrefs in every required location. However, all pages are unref'd
5711 * at the end of the transaction. This guarantees that no stale references
5712 * linger in the per-env list.
5713 *
5714 * Usually we map chunks of 16 pages at a time, but if an overflow page begins
5715 * at the tail of the chunk we extend the chunk to include the entire overflow
5716 * page. Unfortunately, pages can be turned into overflow pages after their
5717 * chunk was already mapped. In that case we must remap the chunk if the
5718 * overflow page is referenced. If the chunk's refcnt is 0 we can just remap
5719 * it, otherwise we temporarily map a new chunk just for the overflow page.
5720 *
5721 * @note this chunk handling means we cannot guarantee that a data item
5722 * returned from the DB will stay alive for the duration of the transaction:
5723 * We unref pages as soon as a cursor moves away from the page
5724 * A subsequent op may cause a purge, which may unmap any unref'd chunks
5725 * The caller must copy the data if it must be used later in the same txn.
5726 *
5727 * Also - our reference counting revolves around cursors, but overflow pages
5728 * aren't pointed to by a cursor's page stack. We have to remember them
5729 * explicitly, in the added mc_ovpg field. A single cursor can only hold a
5730 * reference to one overflow page at a time.
5731 *
5732 * @param[in] txn the transaction for this access.
5733 * @param[in] pgno the page number for the page to retrieve.
5734 * @param[out] ret address of a pointer where the page's address will be stored.
5735 * @return 0 on success, non-zero on failure.
5736 */
5737 static int
mdb_rpage_get(MDB_txn * txn,pgno_t pg0,MDB_page ** ret)5738 mdb_rpage_get(MDB_txn *txn, pgno_t pg0, MDB_page **ret)
5739 {
5740 MDB_env *env = txn->mt_env;
5741 MDB_page *p;
5742 MDB_ID3L tl = txn->mt_rpages;
5743 MDB_ID3L el = env->me_rpages;
5744 MDB_ID3 id3;
5745 unsigned x, rem;
5746 pgno_t pgno;
5747 int rc, retries = 1;
5748 #ifdef _WIN32
5749 LARGE_INTEGER off;
5750 SIZE_T len;
5751 #define SET_OFF(off,val) off.QuadPart = val
5752 #define MAP(rc,env,addr,len,off) \
5753 addr = NULL; \
5754 rc = NtMapViewOfSection(env->me_fmh, GetCurrentProcess(), &addr, 0, \
5755 len, &off, &len, ViewUnmap, (env->me_flags & MDB_RDONLY) ? 0 : MEM_RESERVE, PAGE_READONLY); \
5756 if (rc) rc = mdb_nt2win32(rc)
5757 #else
5758 off_t off;
5759 size_t len;
5760 #define SET_OFF(off,val) off = val
5761 #define MAP(rc,env,addr,len,off) \
5762 addr = mmap(NULL, len, PROT_READ, MAP_SHARED, env->me_fd, off); \
5763 rc = (addr == MAP_FAILED) ? errno : 0
5764 #endif
5765
5766 /* remember the offset of the actual page number, so we can
5767 * return the correct pointer at the end.
5768 */
5769 rem = pg0 & (MDB_RPAGE_CHUNK-1);
5770 pgno = pg0 ^ rem;
5771
5772 id3.mid = 0;
5773 x = mdb_mid3l_search(tl, pgno);
5774 if (x <= tl[0].mid && tl[x].mid == pgno) {
5775 if (x != tl[0].mid && tl[x+1].mid == pg0)
5776 x++;
5777 /* check for overflow size */
5778 p = (MDB_page *)((char *)tl[x].mptr + rem * env->me_psize);
5779 if (IS_OVERFLOW(p) && p->mp_pages + rem > tl[x].mcnt) {
5780 id3.mcnt = p->mp_pages + rem;
5781 len = id3.mcnt * env->me_psize;
5782 SET_OFF(off, pgno * env->me_psize);
5783 MAP(rc, env, id3.mptr, len, off);
5784 if (rc)
5785 return rc;
5786 /* check for local-only page */
5787 if (rem) {
5788 mdb_tassert(txn, tl[x].mid != pg0);
5789 /* hope there's room to insert this locally.
5790 * setting mid here tells later code to just insert
5791 * this id3 instead of searching for a match.
5792 */
5793 id3.mid = pg0;
5794 goto notlocal;
5795 } else {
5796 /* ignore the mapping we got from env, use new one */
5797 tl[x].mptr = id3.mptr;
5798 tl[x].mcnt = id3.mcnt;
5799 /* if no active ref, see if we can replace in env */
5800 if (!tl[x].mref) {
5801 unsigned i;
5802 pthread_mutex_lock(&env->me_rpmutex);
5803 i = mdb_mid3l_search(el, tl[x].mid);
5804 if (el[i].mref == 1) {
5805 /* just us, replace it */
5806 munmap(el[i].mptr, el[i].mcnt * env->me_psize);
5807 el[i].mptr = tl[x].mptr;
5808 el[i].mcnt = tl[x].mcnt;
5809 } else {
5810 /* there are others, remove ourself */
5811 el[i].mref--;
5812 }
5813 pthread_mutex_unlock(&env->me_rpmutex);
5814 }
5815 }
5816 }
5817 id3.mptr = tl[x].mptr;
5818 id3.mcnt = tl[x].mcnt;
5819 tl[x].mref++;
5820 goto ok;
5821 }
5822
5823 notlocal:
5824 if (tl[0].mid >= MDB_TRPAGE_MAX - txn->mt_rpcheck) {
5825 unsigned i, y;
5826 /* purge unref'd pages from our list and unref in env */
5827 pthread_mutex_lock(&env->me_rpmutex);
5828 retry:
5829 y = 0;
5830 for (i=1; i<=tl[0].mid; i++) {
5831 if (!tl[i].mref) {
5832 if (!y) y = i;
5833 /* tmp overflow pages don't go to env */
5834 if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
5835 munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
5836 continue;
5837 }
5838 x = mdb_mid3l_search(el, tl[i].mid);
5839 el[x].mref--;
5840 }
5841 }
5842 pthread_mutex_unlock(&env->me_rpmutex);
5843 if (!y) {
5844 /* we didn't find any unref'd chunks.
5845 * if we're out of room, fail.
5846 */
5847 if (tl[0].mid >= MDB_TRPAGE_MAX)
5848 return MDB_TXN_FULL;
5849 /* otherwise, raise threshold for next time around
5850 * and let this go.
5851 */
5852 txn->mt_rpcheck /= 2;
5853 } else {
5854 /* we found some unused; consolidate the list */
5855 for (i=y+1; i<= tl[0].mid; i++)
5856 if (tl[i].mref)
5857 tl[y++] = tl[i];
5858 tl[0].mid = y-1;
5859 /* decrease the check threshold toward its original value */
5860 if (!txn->mt_rpcheck)
5861 txn->mt_rpcheck = 1;
5862 while (txn->mt_rpcheck < tl[0].mid && txn->mt_rpcheck < MDB_TRPAGE_SIZE/2)
5863 txn->mt_rpcheck *= 2;
5864 }
5865 }
5866 if (tl[0].mid < MDB_TRPAGE_SIZE) {
5867 id3.mref = 1;
5868 if (id3.mid)
5869 goto found;
5870 /* don't map past last written page in read-only envs */
5871 if ((env->me_flags & MDB_RDONLY) && pgno + MDB_RPAGE_CHUNK-1 > txn->mt_last_pgno)
5872 id3.mcnt = txn->mt_last_pgno + 1 - pgno;
5873 else
5874 id3.mcnt = MDB_RPAGE_CHUNK;
5875 len = id3.mcnt * env->me_psize;
5876 id3.mid = pgno;
5877
5878 /* search for page in env */
5879 pthread_mutex_lock(&env->me_rpmutex);
5880 x = mdb_mid3l_search(el, pgno);
5881 if (x <= el[0].mid && el[x].mid == pgno) {
5882 id3.mptr = el[x].mptr;
5883 id3.mcnt = el[x].mcnt;
5884 /* check for overflow size */
5885 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
5886 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
5887 id3.mcnt = p->mp_pages + rem;
5888 len = id3.mcnt * env->me_psize;
5889 SET_OFF(off, pgno * env->me_psize);
5890 MAP(rc, env, id3.mptr, len, off);
5891 if (rc)
5892 goto fail;
5893 if (!el[x].mref) {
5894 munmap(el[x].mptr, env->me_psize * el[x].mcnt);
5895 el[x].mptr = id3.mptr;
5896 el[x].mcnt = id3.mcnt;
5897 } else {
5898 id3.mid = pg0;
5899 pthread_mutex_unlock(&env->me_rpmutex);
5900 goto found;
5901 }
5902 }
5903 el[x].mref++;
5904 pthread_mutex_unlock(&env->me_rpmutex);
5905 goto found;
5906 }
5907 if (el[0].mid >= MDB_ERPAGE_MAX - env->me_rpcheck) {
5908 /* purge unref'd pages */
5909 unsigned i, y = 0;
5910 for (i=1; i<=el[0].mid; i++) {
5911 if (!el[i].mref) {
5912 if (!y) y = i;
5913 munmap(el[i].mptr, env->me_psize * el[i].mcnt);
5914 }
5915 }
5916 if (!y) {
5917 if (retries) {
5918 /* see if we can unref some local pages */
5919 retries--;
5920 id3.mid = 0;
5921 goto retry;
5922 }
5923 if (el[0].mid >= MDB_ERPAGE_MAX) {
5924 pthread_mutex_unlock(&env->me_rpmutex);
5925 return MDB_MAP_FULL;
5926 }
5927 env->me_rpcheck /= 2;
5928 } else {
5929 for (i=y+1; i<= el[0].mid; i++)
5930 if (el[i].mref)
5931 el[y++] = el[i];
5932 el[0].mid = y-1;
5933 if (!env->me_rpcheck)
5934 env->me_rpcheck = 1;
5935 while (env->me_rpcheck < el[0].mid && env->me_rpcheck < MDB_ERPAGE_SIZE/2)
5936 env->me_rpcheck *= 2;
5937 }
5938 }
5939 SET_OFF(off, pgno * env->me_psize);
5940 MAP(rc, env, id3.mptr, len, off);
5941 if (rc) {
5942 fail:
5943 pthread_mutex_unlock(&env->me_rpmutex);
5944 return rc;
5945 }
5946 /* check for overflow size */
5947 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
5948 if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
5949 id3.mcnt = p->mp_pages + rem;
5950 munmap(id3.mptr, len);
5951 len = id3.mcnt * env->me_psize;
5952 MAP(rc, env, id3.mptr, len, off);
5953 if (rc)
5954 goto fail;
5955 }
5956 mdb_mid3l_insert(el, &id3);
5957 pthread_mutex_unlock(&env->me_rpmutex);
5958 found:
5959 mdb_mid3l_insert(tl, &id3);
5960 } else {
5961 return MDB_TXN_FULL;
5962 }
5963 ok:
5964 p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
5965 #if MDB_DEBUG /* we don't need this check any more */
5966 if (IS_OVERFLOW(p)) {
5967 mdb_tassert(txn, p->mp_pages + rem <= id3.mcnt);
5968 }
5969 #endif
5970 *ret = p;
5971 return MDB_SUCCESS;
5972 }
5973 #endif
5974
5975 /** Find the address of the page corresponding to a given page number.
5976 * @param[in] mc the cursor accessing the page.
5977 * @param[in] pgno the page number for the page to retrieve.
5978 * @param[out] ret address of a pointer where the page's address will be stored.
5979 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5980 * @return 0 on success, non-zero on failure.
5981 */
5982 static int
mdb_page_get(MDB_cursor * mc,pgno_t pgno,MDB_page ** ret,int * lvl)5983 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl)
5984 {
5985 MDB_txn *txn = mc->mc_txn;
5986 MDB_page *p = NULL;
5987 int level;
5988
5989 if (! (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP))) {
5990 MDB_txn *tx2 = txn;
5991 level = 1;
5992 do {
5993 MDB_ID2L dl = tx2->mt_u.dirty_list;
5994 unsigned x;
5995 /* Spilled pages were dirtied in this txn and flushed
5996 * because the dirty list got full. Bring this page
5997 * back in from the map (but don't unspill it here,
5998 * leave that unless page_touch happens again).
5999 */
6000 if (tx2->mt_spill_pgs) {
6001 MDB_ID pn = pgno << 1;
6002 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
6003 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
6004 goto mapped;
6005 }
6006 }
6007 if (dl[0].mid) {
6008 unsigned x = mdb_mid2l_search(dl, pgno);
6009 if (x <= dl[0].mid && dl[x].mid == pgno) {
6010 p = dl[x].mptr;
6011 goto done;
6012 }
6013 }
6014 level++;
6015 } while ((tx2 = tx2->mt_parent) != NULL);
6016 }
6017
6018 if (pgno >= txn->mt_next_pgno) {
6019 DPRINTF(("page %"Yu" not found", pgno));
6020 txn->mt_flags |= MDB_TXN_ERROR;
6021 return MDB_PAGE_NOTFOUND;
6022 }
6023
6024 level = 0;
6025
6026 mapped:
6027 {
6028 #ifdef MDB_VL32
6029 int rc = mdb_rpage_get(txn, pgno, &p);
6030 if (rc)
6031 return rc;
6032 #else
6033 MDB_env *env = txn->mt_env;
6034 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
6035 #endif
6036 }
6037
6038 done:
6039 *ret = p;
6040 if (lvl)
6041 *lvl = level;
6042 return MDB_SUCCESS;
6043 }
6044
6045 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
6046 * The cursor is at the root page, set up the rest of it.
6047 */
6048 static int
mdb_page_search_root(MDB_cursor * mc,MDB_val * key,int flags)6049 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
6050 {
6051 MDB_page *mp = mc->mc_pg[mc->mc_top];
6052 int rc;
6053 DKBUF;
6054
6055 while (IS_BRANCH(mp)) {
6056 MDB_node *node;
6057 indx_t i;
6058
6059 DPRINTF(("branch page %"Yu" has %u keys", mp->mp_pgno, NUMKEYS(mp)));
6060 /* Don't assert on branch pages in the FreeDB. We can get here
6061 * while in the process of rebalancing a FreeDB branch page; we must
6062 * let that proceed. ITS#8336
6063 */
6064 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1);
6065 DPRINTF(("found index 0 to page %"Yu, NODEPGNO(NODEPTR(mp, 0))));
6066
6067 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
6068 i = 0;
6069 if (flags & MDB_PS_LAST)
6070 i = NUMKEYS(mp) - 1;
6071 } else {
6072 int exact;
6073 node = mdb_node_search(mc, key, &exact);
6074 if (node == NULL)
6075 i = NUMKEYS(mp) - 1;
6076 else {
6077 i = mc->mc_ki[mc->mc_top];
6078 if (!exact) {
6079 mdb_cassert(mc, i > 0);
6080 i--;
6081 }
6082 }
6083 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
6084 }
6085
6086 mdb_cassert(mc, i < NUMKEYS(mp));
6087 node = NODEPTR(mp, i);
6088
6089 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6090 return rc;
6091
6092 mc->mc_ki[mc->mc_top] = i;
6093 if ((rc = mdb_cursor_push(mc, mp)))
6094 return rc;
6095
6096 if (flags & MDB_PS_MODIFY) {
6097 if ((rc = mdb_page_touch(mc)) != 0)
6098 return rc;
6099 mp = mc->mc_pg[mc->mc_top];
6100 }
6101 }
6102
6103 if (!IS_LEAF(mp)) {
6104 DPRINTF(("internal error, index points to a %02X page!?",
6105 mp->mp_flags));
6106 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6107 return MDB_CORRUPTED;
6108 }
6109
6110 DPRINTF(("found leaf page %"Yu" for key [%s]", mp->mp_pgno,
6111 key ? DKEY(key) : "null"));
6112 mc->mc_flags |= C_INITIALIZED;
6113 mc->mc_flags &= ~C_EOF;
6114
6115 return MDB_SUCCESS;
6116 }
6117
6118 /** Search for the lowest key under the current branch page.
6119 * This just bypasses a NUMKEYS check in the current page
6120 * before calling mdb_page_search_root(), because the callers
6121 * are all in situations where the current page is known to
6122 * be underfilled.
6123 */
6124 static int
mdb_page_search_lowest(MDB_cursor * mc)6125 mdb_page_search_lowest(MDB_cursor *mc)
6126 {
6127 MDB_page *mp = mc->mc_pg[mc->mc_top];
6128 MDB_node *node = NODEPTR(mp, 0);
6129 int rc;
6130
6131 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0)
6132 return rc;
6133
6134 mc->mc_ki[mc->mc_top] = 0;
6135 if ((rc = mdb_cursor_push(mc, mp)))
6136 return rc;
6137 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
6138 }
6139
6140 /** Search for the page a given key should be in.
6141 * Push it and its parent pages on the cursor stack.
6142 * @param[in,out] mc the cursor for this operation.
6143 * @param[in] key the key to search for, or NULL for first/last page.
6144 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
6145 * are touched (updated with new page numbers).
6146 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
6147 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
6148 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
6149 * @return 0 on success, non-zero on failure.
6150 */
6151 static int
mdb_page_search(MDB_cursor * mc,MDB_val * key,int flags)6152 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
6153 {
6154 int rc;
6155 pgno_t root;
6156
6157 /* Make sure the txn is still viable, then find the root from
6158 * the txn's db table and set it as the root of the cursor's stack.
6159 */
6160 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) {
6161 DPUTS("transaction may not be used now");
6162 return MDB_BAD_TXN;
6163 } else {
6164 /* Make sure we're using an up-to-date root */
6165 if (*mc->mc_dbflag & DB_STALE) {
6166 MDB_cursor mc2;
6167 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6168 return MDB_BAD_DBI;
6169 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
6170 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
6171 if (rc)
6172 return rc;
6173 {
6174 MDB_val data;
6175 int exact = 0;
6176 uint16_t flags;
6177 MDB_node *leaf = mdb_node_search(&mc2,
6178 &mc->mc_dbx->md_name, &exact);
6179 if (!exact)
6180 return MDB_NOTFOUND;
6181 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
6182 return MDB_INCOMPATIBLE; /* not a named DB */
6183 rc = mdb_node_read(&mc2, leaf, &data);
6184 if (rc)
6185 return rc;
6186 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
6187 sizeof(uint16_t));
6188 /* The txn may not know this DBI, or another process may
6189 * have dropped and recreated the DB with other flags.
6190 */
6191 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
6192 return MDB_INCOMPATIBLE;
6193 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
6194 }
6195 *mc->mc_dbflag &= ~DB_STALE;
6196 }
6197 root = mc->mc_db->md_root;
6198
6199 if (root == P_INVALID) { /* Tree is empty. */
6200 DPUTS("tree is empty");
6201 return MDB_NOTFOUND;
6202 }
6203 }
6204
6205 mdb_cassert(mc, root > 1);
6206 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root) {
6207 #ifdef MDB_VL32
6208 if (mc->mc_pg[0])
6209 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[0]);
6210 #endif
6211 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0)
6212 return rc;
6213 }
6214
6215 #ifdef MDB_VL32
6216 {
6217 int i;
6218 for (i=1; i<mc->mc_snum; i++)
6219 MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[i]);
6220 }
6221 #endif
6222 mc->mc_snum = 1;
6223 mc->mc_top = 0;
6224
6225 DPRINTF(("db %d root page %"Yu" has flags 0x%X",
6226 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
6227
6228 if (flags & MDB_PS_MODIFY) {
6229 if ((rc = mdb_page_touch(mc)))
6230 return rc;
6231 }
6232
6233 if (flags & MDB_PS_ROOTONLY)
6234 return MDB_SUCCESS;
6235
6236 return mdb_page_search_root(mc, key, flags);
6237 }
6238
6239 static int
mdb_ovpage_free(MDB_cursor * mc,MDB_page * mp)6240 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
6241 {
6242 MDB_txn *txn = mc->mc_txn;
6243 pgno_t pg = mp->mp_pgno;
6244 unsigned x = 0, ovpages = mp->mp_pages;
6245 MDB_env *env = txn->mt_env;
6246 MDB_IDL sl = txn->mt_spill_pgs;
6247 MDB_ID pn = pg << 1;
6248 int rc;
6249
6250 DPRINTF(("free ov page %"Yu" (%d)", pg, ovpages));
6251 /* If the page is dirty or on the spill list we just acquired it,
6252 * so we should give it back to our current free list, if any.
6253 * Otherwise put it onto the list of pages we freed in this txn.
6254 *
6255 * Won't create me_pghead: me_pglast must be inited along with it.
6256 * Unsupported in nested txns: They would need to hide the page
6257 * range in ancestor txns' dirty and spilled lists.
6258 */
6259 if (env->me_pghead &&
6260 !txn->mt_parent &&
6261 ((mp->mp_flags & P_DIRTY) ||
6262 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
6263 {
6264 unsigned i, j;
6265 pgno_t *mop;
6266 MDB_ID2 *dl, ix, iy;
6267 rc = mdb_midl_need(&env->me_pghead, ovpages);
6268 if (rc)
6269 return rc;
6270 if (!(mp->mp_flags & P_DIRTY)) {
6271 /* This page is no longer spilled */
6272 if (x == sl[0])
6273 sl[0]--;
6274 else
6275 sl[x] |= 1;
6276 goto release;
6277 }
6278 /* Remove from dirty list */
6279 dl = txn->mt_u.dirty_list;
6280 x = dl[0].mid--;
6281 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
6282 if (x > 1) {
6283 x--;
6284 iy = dl[x];
6285 dl[x] = ix;
6286 } else {
6287 mdb_cassert(mc, x > 1);
6288 j = ++(dl[0].mid);
6289 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
6290 txn->mt_flags |= MDB_TXN_ERROR;
6291 return MDB_PROBLEM;
6292 }
6293 }
6294 txn->mt_dirty_room++;
6295 if (!(env->me_flags & MDB_WRITEMAP))
6296 mdb_dpage_free(env, mp);
6297 release:
6298 /* Insert in me_pghead */
6299 mop = env->me_pghead;
6300 j = mop[0] + ovpages;
6301 for (i = mop[0]; i && mop[i] < pg; i--)
6302 mop[j--] = mop[i];
6303 while (j>i)
6304 mop[j--] = pg++;
6305 mop[0] += ovpages;
6306 } else {
6307 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
6308 if (rc)
6309 return rc;
6310 }
6311 mc->mc_db->md_overflow_pages -= ovpages;
6312 return 0;
6313 }
6314
6315 /** Return the data associated with a given node.
6316 * @param[in] mc The cursor for this operation.
6317 * @param[in] leaf The node being read.
6318 * @param[out] data Updated to point to the node's data.
6319 * @return 0 on success, non-zero on failure.
6320 */
6321 static int
mdb_node_read(MDB_cursor * mc,MDB_node * leaf,MDB_val * data)6322 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data)
6323 {
6324 MDB_page *omp; /* overflow page */
6325 pgno_t pgno;
6326 int rc;
6327
6328 if (MC_OVPG(mc)) {
6329 MDB_PAGE_UNREF(mc->mc_txn, MC_OVPG(mc));
6330 MC_SET_OVPG(mc, NULL);
6331 }
6332 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6333 data->mv_size = NODEDSZ(leaf);
6334 data->mv_data = NODEDATA(leaf);
6335 return MDB_SUCCESS;
6336 }
6337
6338 /* Read overflow data.
6339 */
6340 data->mv_size = NODEDSZ(leaf);
6341 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
6342 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) {
6343 DPRINTF(("read overflow page %"Yu" failed", pgno));
6344 return rc;
6345 }
6346 data->mv_data = METADATA(omp);
6347 MC_SET_OVPG(mc, omp);
6348
6349 return MDB_SUCCESS;
6350 }
6351
6352 int
mdb_get(MDB_txn * txn,MDB_dbi dbi,MDB_val * key,MDB_val * data)6353 mdb_get(MDB_txn *txn, MDB_dbi dbi,
6354 MDB_val *key, MDB_val *data)
6355 {
6356 MDB_cursor mc;
6357 MDB_xcursor mx;
6358 int exact = 0, rc;
6359 DKBUF;
6360
6361 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
6362
6363 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
6364 return EINVAL;
6365
6366 if (txn->mt_flags & MDB_TXN_BLOCKED)
6367 return MDB_BAD_TXN;
6368
6369 mdb_cursor_init(&mc, txn, dbi, &mx);
6370 rc = mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
6371 /* unref all the pages when MDB_VL32 - caller must copy the data
6372 * before doing anything else
6373 */
6374 MDB_CURSOR_UNREF(&mc, 1);
6375 return rc;
6376 }
6377
6378 /** Find a sibling for a page.
6379 * Replaces the page at the top of the cursor's stack with the
6380 * specified sibling, if one exists.
6381 * @param[in] mc The cursor for this operation.
6382 * @param[in] move_right Non-zero if the right sibling is requested,
6383 * otherwise the left sibling.
6384 * @return 0 on success, non-zero on failure.
6385 */
6386 static int
mdb_cursor_sibling(MDB_cursor * mc,int move_right)6387 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
6388 {
6389 int rc;
6390 MDB_node *indx;
6391 MDB_page *mp;
6392 #ifdef MDB_VL32
6393 MDB_page *op;
6394 #endif
6395
6396 if (mc->mc_snum < 2) {
6397 return MDB_NOTFOUND; /* root has no siblings */
6398 }
6399
6400 #ifdef MDB_VL32
6401 op = mc->mc_pg[mc->mc_top];
6402 #endif
6403 mdb_cursor_pop(mc);
6404 DPRINTF(("parent page is page %"Yu", index %u",
6405 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
6406
6407 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6408 : (mc->mc_ki[mc->mc_top] == 0)) {
6409 DPRINTF(("no more keys left, moving to %s sibling",
6410 move_right ? "right" : "left"));
6411 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
6412 /* undo cursor_pop before returning */
6413 mc->mc_top++;
6414 mc->mc_snum++;
6415 return rc;
6416 }
6417 } else {
6418 if (move_right)
6419 mc->mc_ki[mc->mc_top]++;
6420 else
6421 mc->mc_ki[mc->mc_top]--;
6422 DPRINTF(("just moving to %s index key %u",
6423 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
6424 }
6425 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
6426
6427 MDB_PAGE_UNREF(mc->mc_txn, op);
6428
6429 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6430 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) {
6431 /* mc will be inconsistent if caller does mc_snum++ as above */
6432 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
6433 return rc;
6434 }
6435
6436 mdb_cursor_push(mc, mp);
6437 if (!move_right)
6438 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
6439
6440 return MDB_SUCCESS;
6441 }
6442
6443 /** Move the cursor to the next data item. */
6444 static int
mdb_cursor_next(MDB_cursor * mc,MDB_val * key,MDB_val * data,MDB_cursor_op op)6445 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6446 {
6447 MDB_page *mp;
6448 MDB_node *leaf;
6449 int rc;
6450
6451 if ((mc->mc_flags & C_EOF) ||
6452 ((mc->mc_flags & C_DEL) && op == MDB_NEXT_DUP)) {
6453 return MDB_NOTFOUND;
6454 }
6455 if (!(mc->mc_flags & C_INITIALIZED))
6456 return mdb_cursor_first(mc, key, data);
6457
6458 mp = mc->mc_pg[mc->mc_top];
6459
6460 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6461 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6462 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6463 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
6464 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
6465 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
6466 if (rc == MDB_SUCCESS)
6467 MDB_GET_KEY(leaf, key);
6468 return rc;
6469 }
6470 }
6471 else {
6472 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6473 }
6474 } else {
6475 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6476 if (op == MDB_NEXT_DUP)
6477 return MDB_NOTFOUND;
6478 }
6479 }
6480
6481 DPRINTF(("cursor_next: top page is %"Yu" in cursor %p",
6482 mdb_dbg_pgno(mp), (void *) mc));
6483 if (mc->mc_flags & C_DEL) {
6484 mc->mc_flags ^= C_DEL;
6485 goto skip;
6486 }
6487
6488 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
6489 DPUTS("=====> move to next sibling page");
6490 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6491 mc->mc_flags |= C_EOF;
6492 return rc;
6493 }
6494 mp = mc->mc_pg[mc->mc_top];
6495 DPRINTF(("next page is %"Yu", key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6496 } else
6497 mc->mc_ki[mc->mc_top]++;
6498
6499 skip:
6500 DPRINTF(("==> cursor points to page %"Yu" with %u keys, key index %u",
6501 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6502
6503 if (IS_LEAF2(mp)) {
6504 key->mv_size = mc->mc_db->md_pad;
6505 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6506 return MDB_SUCCESS;
6507 }
6508
6509 mdb_cassert(mc, IS_LEAF(mp));
6510 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6511
6512 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6513 mdb_xcursor_init1(mc, leaf);
6514 }
6515 if (data) {
6516 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6517 return rc;
6518
6519 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6520 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6521 if (rc != MDB_SUCCESS)
6522 return rc;
6523 }
6524 }
6525
6526 MDB_GET_KEY(leaf, key);
6527 return MDB_SUCCESS;
6528 }
6529
6530 /** Move the cursor to the previous data item. */
6531 static int
mdb_cursor_prev(MDB_cursor * mc,MDB_val * key,MDB_val * data,MDB_cursor_op op)6532 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
6533 {
6534 MDB_page *mp;
6535 MDB_node *leaf;
6536 int rc;
6537
6538 if (!(mc->mc_flags & C_INITIALIZED)) {
6539 rc = mdb_cursor_last(mc, key, data);
6540 if (rc)
6541 return rc;
6542 mc->mc_ki[mc->mc_top]++;
6543 }
6544
6545 mp = mc->mc_pg[mc->mc_top];
6546
6547 if (mc->mc_db->md_flags & MDB_DUPSORT) {
6548 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6549 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6550 if (op == MDB_PREV || op == MDB_PREV_DUP) {
6551 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
6552 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
6553 if (rc == MDB_SUCCESS) {
6554 MDB_GET_KEY(leaf, key);
6555 mc->mc_flags &= ~C_EOF;
6556 }
6557 return rc;
6558 }
6559 }
6560 else {
6561 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6562 }
6563 } else {
6564 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6565 if (op == MDB_PREV_DUP)
6566 return MDB_NOTFOUND;
6567 }
6568 }
6569
6570 DPRINTF(("cursor_prev: top page is %"Yu" in cursor %p",
6571 mdb_dbg_pgno(mp), (void *) mc));
6572
6573 mc->mc_flags &= ~(C_EOF|C_DEL);
6574
6575 if (mc->mc_ki[mc->mc_top] == 0) {
6576 DPUTS("=====> move to prev sibling page");
6577 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
6578 return rc;
6579 }
6580 mp = mc->mc_pg[mc->mc_top];
6581 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
6582 DPRINTF(("prev page is %"Yu", key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
6583 } else
6584 mc->mc_ki[mc->mc_top]--;
6585
6586 mc->mc_flags &= ~C_EOF;
6587
6588 DPRINTF(("==> cursor points to page %"Yu" with %u keys, key index %u",
6589 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
6590
6591 if (IS_LEAF2(mp)) {
6592 key->mv_size = mc->mc_db->md_pad;
6593 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6594 return MDB_SUCCESS;
6595 }
6596
6597 mdb_cassert(mc, IS_LEAF(mp));
6598 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6599
6600 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6601 mdb_xcursor_init1(mc, leaf);
6602 }
6603 if (data) {
6604 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6605 return rc;
6606
6607 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6608 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6609 if (rc != MDB_SUCCESS)
6610 return rc;
6611 }
6612 }
6613
6614 MDB_GET_KEY(leaf, key);
6615 return MDB_SUCCESS;
6616 }
6617
6618 /** Set the cursor on a specific data item. */
6619 static int
mdb_cursor_set(MDB_cursor * mc,MDB_val * key,MDB_val * data,MDB_cursor_op op,int * exactp)6620 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6621 MDB_cursor_op op, int *exactp)
6622 {
6623 int rc;
6624 MDB_page *mp;
6625 MDB_node *leaf = NULL;
6626 DKBUF;
6627
6628 if (key->mv_size == 0)
6629 return MDB_BAD_VALSIZE;
6630
6631 if (mc->mc_xcursor) {
6632 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6633 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6634 }
6635
6636 /* See if we're already on the right page */
6637 if (mc->mc_flags & C_INITIALIZED) {
6638 MDB_val nodekey;
6639
6640 mp = mc->mc_pg[mc->mc_top];
6641 if (!NUMKEYS(mp)) {
6642 mc->mc_ki[mc->mc_top] = 0;
6643 return MDB_NOTFOUND;
6644 }
6645 if (mp->mp_flags & P_LEAF2) {
6646 nodekey.mv_size = mc->mc_db->md_pad;
6647 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
6648 } else {
6649 leaf = NODEPTR(mp, 0);
6650 MDB_GET_KEY2(leaf, nodekey);
6651 }
6652 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6653 if (rc == 0) {
6654 /* Probably happens rarely, but first node on the page
6655 * was the one we wanted.
6656 */
6657 mc->mc_ki[mc->mc_top] = 0;
6658 if (exactp)
6659 *exactp = 1;
6660 goto set1;
6661 }
6662 if (rc > 0) {
6663 unsigned int i;
6664 unsigned int nkeys = NUMKEYS(mp);
6665 if (nkeys > 1) {
6666 if (mp->mp_flags & P_LEAF2) {
6667 nodekey.mv_data = LEAF2KEY(mp,
6668 nkeys-1, nodekey.mv_size);
6669 } else {
6670 leaf = NODEPTR(mp, nkeys-1);
6671 MDB_GET_KEY2(leaf, nodekey);
6672 }
6673 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6674 if (rc == 0) {
6675 /* last node was the one we wanted */
6676 mc->mc_ki[mc->mc_top] = nkeys-1;
6677 if (exactp)
6678 *exactp = 1;
6679 goto set1;
6680 }
6681 if (rc < 0) {
6682 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
6683 /* This is definitely the right page, skip search_page */
6684 if (mp->mp_flags & P_LEAF2) {
6685 nodekey.mv_data = LEAF2KEY(mp,
6686 mc->mc_ki[mc->mc_top], nodekey.mv_size);
6687 } else {
6688 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6689 MDB_GET_KEY2(leaf, nodekey);
6690 }
6691 rc = mc->mc_dbx->md_cmp(key, &nodekey);
6692 if (rc == 0) {
6693 /* current node was the one we wanted */
6694 if (exactp)
6695 *exactp = 1;
6696 goto set1;
6697 }
6698 }
6699 rc = 0;
6700 goto set2;
6701 }
6702 }
6703 /* If any parents have right-sibs, search.
6704 * Otherwise, there's nothing further.
6705 */
6706 for (i=0; i<mc->mc_top; i++)
6707 if (mc->mc_ki[i] <
6708 NUMKEYS(mc->mc_pg[i])-1)
6709 break;
6710 if (i == mc->mc_top) {
6711 /* There are no other pages */
6712 mc->mc_ki[mc->mc_top] = nkeys;
6713 return MDB_NOTFOUND;
6714 }
6715 }
6716 if (!mc->mc_top) {
6717 /* There are no other pages */
6718 mc->mc_ki[mc->mc_top] = 0;
6719 if (op == MDB_SET_RANGE && !exactp) {
6720 rc = 0;
6721 goto set1;
6722 } else
6723 return MDB_NOTFOUND;
6724 }
6725 } else {
6726 mc->mc_pg[0] = 0;
6727 }
6728
6729 rc = mdb_page_search(mc, key, 0);
6730 if (rc != MDB_SUCCESS)
6731 return rc;
6732
6733 mp = mc->mc_pg[mc->mc_top];
6734 mdb_cassert(mc, IS_LEAF(mp));
6735
6736 set2:
6737 leaf = mdb_node_search(mc, key, exactp);
6738 if (exactp != NULL && !*exactp) {
6739 /* MDB_SET specified and not an exact match. */
6740 return MDB_NOTFOUND;
6741 }
6742
6743 if (leaf == NULL) {
6744 DPUTS("===> inexact leaf not found, goto sibling");
6745 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
6746 mc->mc_flags |= C_EOF;
6747 return rc; /* no entries matched */
6748 }
6749 mp = mc->mc_pg[mc->mc_top];
6750 mdb_cassert(mc, IS_LEAF(mp));
6751 leaf = NODEPTR(mp, 0);
6752 }
6753
6754 set1:
6755 mc->mc_flags |= C_INITIALIZED;
6756 mc->mc_flags &= ~C_EOF;
6757
6758 if (IS_LEAF2(mp)) {
6759 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
6760 key->mv_size = mc->mc_db->md_pad;
6761 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6762 }
6763 return MDB_SUCCESS;
6764 }
6765
6766 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6767 mdb_xcursor_init1(mc, leaf);
6768 }
6769 if (data) {
6770 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6771 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
6772 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6773 } else {
6774 int ex2, *ex2p;
6775 if (op == MDB_GET_BOTH) {
6776 ex2p = &ex2;
6777 ex2 = 0;
6778 } else {
6779 ex2p = NULL;
6780 }
6781 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
6782 if (rc != MDB_SUCCESS)
6783 return rc;
6784 }
6785 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
6786 MDB_val olddata;
6787 MDB_cmp_func *dcmp;
6788 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS)
6789 return rc;
6790 dcmp = mc->mc_dbx->md_dcmp;
6791 if (NEED_CMP_CLONG(dcmp, olddata.mv_size))
6792 dcmp = mdb_cmp_clong;
6793 rc = dcmp(data, &olddata);
6794 if (rc) {
6795 if (op == MDB_GET_BOTH || rc > 0)
6796 return MDB_NOTFOUND;
6797 rc = 0;
6798 }
6799 *data = olddata;
6800
6801 } else {
6802 if (mc->mc_xcursor)
6803 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6804 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6805 return rc;
6806 }
6807 }
6808
6809 /* The key already matches in all other cases */
6810 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
6811 MDB_GET_KEY(leaf, key);
6812 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
6813
6814 return rc;
6815 }
6816
6817 /** Move the cursor to the first item in the database. */
6818 static int
mdb_cursor_first(MDB_cursor * mc,MDB_val * key,MDB_val * data)6819 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6820 {
6821 int rc;
6822 MDB_node *leaf;
6823
6824 if (mc->mc_xcursor) {
6825 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6826 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6827 }
6828
6829 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6830 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
6831 if (rc != MDB_SUCCESS)
6832 return rc;
6833 }
6834 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6835
6836 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
6837 mc->mc_flags |= C_INITIALIZED;
6838 mc->mc_flags &= ~C_EOF;
6839
6840 mc->mc_ki[mc->mc_top] = 0;
6841
6842 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6843 key->mv_size = mc->mc_db->md_pad;
6844 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
6845 return MDB_SUCCESS;
6846 }
6847
6848 if (data) {
6849 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6850 mdb_xcursor_init1(mc, leaf);
6851 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
6852 if (rc)
6853 return rc;
6854 } else {
6855 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6856 return rc;
6857 }
6858 }
6859 MDB_GET_KEY(leaf, key);
6860 return MDB_SUCCESS;
6861 }
6862
6863 /** Move the cursor to the last item in the database. */
6864 static int
mdb_cursor_last(MDB_cursor * mc,MDB_val * key,MDB_val * data)6865 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
6866 {
6867 int rc;
6868 MDB_node *leaf;
6869
6870 if (mc->mc_xcursor) {
6871 MDB_CURSOR_UNREF(&mc->mc_xcursor->mx_cursor, 0);
6872 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
6873 }
6874
6875 if (!(mc->mc_flags & C_EOF)) {
6876
6877 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
6878 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
6879 if (rc != MDB_SUCCESS)
6880 return rc;
6881 }
6882 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
6883
6884 }
6885 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
6886 mc->mc_flags |= C_INITIALIZED|C_EOF;
6887 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6888
6889 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6890 key->mv_size = mc->mc_db->md_pad;
6891 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
6892 return MDB_SUCCESS;
6893 }
6894
6895 if (data) {
6896 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6897 mdb_xcursor_init1(mc, leaf);
6898 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
6899 if (rc)
6900 return rc;
6901 } else {
6902 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS)
6903 return rc;
6904 }
6905 }
6906
6907 MDB_GET_KEY(leaf, key);
6908 return MDB_SUCCESS;
6909 }
6910
6911 int
mdb_cursor_get(MDB_cursor * mc,MDB_val * key,MDB_val * data,MDB_cursor_op op)6912 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6913 MDB_cursor_op op)
6914 {
6915 int rc;
6916 int exact = 0;
6917 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
6918
6919 if (mc == NULL)
6920 return EINVAL;
6921
6922 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
6923 return MDB_BAD_TXN;
6924
6925 switch (op) {
6926 case MDB_GET_CURRENT:
6927 if (!(mc->mc_flags & C_INITIALIZED)) {
6928 rc = EINVAL;
6929 } else {
6930 MDB_page *mp = mc->mc_pg[mc->mc_top];
6931 int nkeys = NUMKEYS(mp);
6932 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
6933 mc->mc_ki[mc->mc_top] = nkeys;
6934 rc = MDB_NOTFOUND;
6935 break;
6936 }
6937 rc = MDB_SUCCESS;
6938 if (IS_LEAF2(mp)) {
6939 key->mv_size = mc->mc_db->md_pad;
6940 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
6941 } else {
6942 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6943 MDB_GET_KEY(leaf, key);
6944 if (data) {
6945 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6946 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
6947 } else {
6948 rc = mdb_node_read(mc, leaf, data);
6949 }
6950 }
6951 }
6952 }
6953 break;
6954 case MDB_GET_BOTH:
6955 case MDB_GET_BOTH_RANGE:
6956 if (data == NULL) {
6957 rc = EINVAL;
6958 break;
6959 }
6960 if (mc->mc_xcursor == NULL) {
6961 rc = MDB_INCOMPATIBLE;
6962 break;
6963 }
6964 /* FALLTHRU */
6965 case MDB_SET:
6966 case MDB_SET_KEY:
6967 case MDB_SET_RANGE:
6968 if (key == NULL) {
6969 rc = EINVAL;
6970 } else {
6971 rc = mdb_cursor_set(mc, key, data, op,
6972 op == MDB_SET_RANGE ? NULL : &exact);
6973 }
6974 break;
6975 case MDB_GET_MULTIPLE:
6976 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6977 rc = EINVAL;
6978 break;
6979 }
6980 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6981 rc = MDB_INCOMPATIBLE;
6982 break;
6983 }
6984 rc = MDB_SUCCESS;
6985 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6986 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6987 break;
6988 goto fetchm;
6989 case MDB_NEXT_MULTIPLE:
6990 if (data == NULL) {
6991 rc = EINVAL;
6992 break;
6993 }
6994 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6995 rc = MDB_INCOMPATIBLE;
6996 break;
6997 }
6998 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6999 if (rc == MDB_SUCCESS) {
7000 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
7001 MDB_cursor *mx;
7002 fetchm:
7003 mx = &mc->mc_xcursor->mx_cursor;
7004 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
7005 mx->mc_db->md_pad;
7006 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
7007 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
7008 } else {
7009 rc = MDB_NOTFOUND;
7010 }
7011 }
7012 break;
7013 case MDB_PREV_MULTIPLE:
7014 if (data == NULL) {
7015 rc = EINVAL;
7016 break;
7017 }
7018 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7019 rc = MDB_INCOMPATIBLE;
7020 break;
7021 }
7022 if (!(mc->mc_flags & C_INITIALIZED))
7023 rc = mdb_cursor_last(mc, key, data);
7024 else
7025 rc = MDB_SUCCESS;
7026 if (rc == MDB_SUCCESS) {
7027 MDB_cursor *mx = &mc->mc_xcursor->mx_cursor;
7028 if (mx->mc_flags & C_INITIALIZED) {
7029 rc = mdb_cursor_sibling(mx, 0);
7030 if (rc == MDB_SUCCESS)
7031 goto fetchm;
7032 } else {
7033 rc = MDB_NOTFOUND;
7034 }
7035 }
7036 break;
7037 case MDB_NEXT:
7038 case MDB_NEXT_DUP:
7039 case MDB_NEXT_NODUP:
7040 rc = mdb_cursor_next(mc, key, data, op);
7041 break;
7042 case MDB_PREV:
7043 case MDB_PREV_DUP:
7044 case MDB_PREV_NODUP:
7045 rc = mdb_cursor_prev(mc, key, data, op);
7046 break;
7047 case MDB_FIRST:
7048 rc = mdb_cursor_first(mc, key, data);
7049 break;
7050 case MDB_FIRST_DUP:
7051 mfunc = mdb_cursor_first;
7052 mmove:
7053 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
7054 rc = EINVAL;
7055 break;
7056 }
7057 if (mc->mc_xcursor == NULL) {
7058 rc = MDB_INCOMPATIBLE;
7059 break;
7060 }
7061 {
7062 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7063 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7064 MDB_GET_KEY(leaf, key);
7065 rc = mdb_node_read(mc, leaf, data);
7066 break;
7067 }
7068 }
7069 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
7070 rc = EINVAL;
7071 break;
7072 }
7073 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
7074 break;
7075 case MDB_LAST:
7076 rc = mdb_cursor_last(mc, key, data);
7077 break;
7078 case MDB_LAST_DUP:
7079 mfunc = mdb_cursor_last;
7080 goto mmove;
7081 default:
7082 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
7083 rc = EINVAL;
7084 break;
7085 }
7086
7087 if (mc->mc_flags & C_DEL)
7088 mc->mc_flags ^= C_DEL;
7089
7090 return rc;
7091 }
7092
7093 /** Touch all the pages in the cursor stack. Set mc_top.
7094 * Makes sure all the pages are writable, before attempting a write operation.
7095 * @param[in] mc The cursor to operate on.
7096 */
7097 static int
mdb_cursor_touch(MDB_cursor * mc)7098 mdb_cursor_touch(MDB_cursor *mc)
7099 {
7100 int rc = MDB_SUCCESS;
7101
7102 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & DB_DIRTY)) {
7103 MDB_cursor mc2;
7104 MDB_xcursor mcx;
7105 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
7106 return MDB_BAD_DBI;
7107 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
7108 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
7109 if (rc)
7110 return rc;
7111 *mc->mc_dbflag |= DB_DIRTY;
7112 }
7113 mc->mc_top = 0;
7114 if (mc->mc_snum) {
7115 do {
7116 rc = mdb_page_touch(mc);
7117 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
7118 mc->mc_top = mc->mc_snum-1;
7119 }
7120 return rc;
7121 }
7122
7123 /** Do not spill pages to disk if txn is getting full, may fail instead */
7124 #define MDB_NOSPILL 0x8000
7125
7126 int
mdb_cursor_put(MDB_cursor * mc,MDB_val * key,MDB_val * data,unsigned int flags)7127 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
7128 unsigned int flags)
7129 {
7130 MDB_env *env;
7131 MDB_node *leaf = NULL;
7132 MDB_page *fp, *mp, *sub_root = NULL;
7133 uint16_t fp_flags;
7134 MDB_val xdata, *rdata, dkey, olddata;
7135 MDB_db dummy;
7136 int do_sub = 0, insert_key, insert_data;
7137 unsigned int mcount = 0, dcount = 0, nospill;
7138 size_t nsize;
7139 int rc, rc2;
7140 unsigned int nflags;
7141 DKBUF;
7142
7143 if (mc == NULL || key == NULL)
7144 return EINVAL;
7145
7146 env = mc->mc_txn->mt_env;
7147
7148 /* Check this first so counter will always be zero on any
7149 * early failures.
7150 */
7151 if (flags & MDB_MULTIPLE) {
7152 dcount = data[1].mv_size;
7153 data[1].mv_size = 0;
7154 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
7155 return MDB_INCOMPATIBLE;
7156 }
7157
7158 nospill = flags & MDB_NOSPILL;
7159 flags &= ~MDB_NOSPILL;
7160
7161 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7162 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7163
7164 if (key->mv_size-1 >= ENV_MAXKEY(env))
7165 return MDB_BAD_VALSIZE;
7166
7167 #if SIZE_MAX > MAXDATASIZE
7168 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
7169 return MDB_BAD_VALSIZE;
7170 #else
7171 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
7172 return MDB_BAD_VALSIZE;
7173 #endif
7174
7175 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
7176 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
7177
7178 dkey.mv_size = 0;
7179
7180 if (flags == MDB_CURRENT) {
7181 if (!(mc->mc_flags & C_INITIALIZED))
7182 return EINVAL;
7183 rc = MDB_SUCCESS;
7184 } else if (mc->mc_db->md_root == P_INVALID) {
7185 /* new database, cursor has nothing to point to */
7186 mc->mc_snum = 0;
7187 mc->mc_top = 0;
7188 mc->mc_flags &= ~C_INITIALIZED;
7189 rc = MDB_NO_ROOT;
7190 } else {
7191 int exact = 0;
7192 MDB_val d2;
7193 if (flags & MDB_APPEND) {
7194 MDB_val k2;
7195 rc = mdb_cursor_last(mc, &k2, &d2);
7196 if (rc == 0) {
7197 rc = mc->mc_dbx->md_cmp(key, &k2);
7198 if (rc > 0) {
7199 rc = MDB_NOTFOUND;
7200 mc->mc_ki[mc->mc_top]++;
7201 } else {
7202 /* new key is <= last key */
7203 rc = MDB_KEYEXIST;
7204 }
7205 }
7206 } else {
7207 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
7208 }
7209 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
7210 DPRINTF(("duplicate key [%s]", DKEY(key)));
7211 *data = d2;
7212 return MDB_KEYEXIST;
7213 }
7214 if (rc && rc != MDB_NOTFOUND)
7215 return rc;
7216 }
7217
7218 if (mc->mc_flags & C_DEL)
7219 mc->mc_flags ^= C_DEL;
7220
7221 /* Cursor is positioned, check for room in the dirty list */
7222 if (!nospill) {
7223 if (flags & MDB_MULTIPLE) {
7224 rdata = &xdata;
7225 xdata.mv_size = data->mv_size * dcount;
7226 } else {
7227 rdata = data;
7228 }
7229 if ((rc2 = mdb_page_spill(mc, key, rdata)))
7230 return rc2;
7231 }
7232
7233 if (rc == MDB_NO_ROOT) {
7234 MDB_page *np;
7235 /* new database, write a root leaf page */
7236 DPUTS("allocating new root leaf page");
7237 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
7238 return rc2;
7239 }
7240 mdb_cursor_push(mc, np);
7241 mc->mc_db->md_root = np->mp_pgno;
7242 mc->mc_db->md_depth++;
7243 *mc->mc_dbflag |= DB_DIRTY;
7244 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
7245 == MDB_DUPFIXED)
7246 np->mp_flags |= P_LEAF2;
7247 mc->mc_flags |= C_INITIALIZED;
7248 } else {
7249 /* make sure all cursor pages are writable */
7250 rc2 = mdb_cursor_touch(mc);
7251 if (rc2)
7252 return rc2;
7253 }
7254
7255 insert_key = insert_data = rc;
7256 if (insert_key) {
7257 /* The key does not exist */
7258 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
7259 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
7260 LEAFSIZE(key, data) > env->me_nodemax)
7261 {
7262 /* Too big for a node, insert in sub-DB. Set up an empty
7263 * "old sub-page" for prep_subDB to expand to a full page.
7264 */
7265 fp_flags = P_LEAF|P_DIRTY;
7266 fp = env->me_pbuf;
7267 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
7268 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
7269 olddata.mv_size = PAGEHDRSZ;
7270 goto prep_subDB;
7271 }
7272 } else {
7273 /* there's only a key anyway, so this is a no-op */
7274 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
7275 char *ptr;
7276 unsigned int ksize = mc->mc_db->md_pad;
7277 if (key->mv_size != ksize)
7278 return MDB_BAD_VALSIZE;
7279 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
7280 memcpy(ptr, key->mv_data, ksize);
7281 fix_parent:
7282 /* if overwriting slot 0 of leaf, need to
7283 * update branch key if there is a parent page
7284 */
7285 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
7286 unsigned short dtop = 1;
7287 mc->mc_top--;
7288 /* slot 0 is always an empty key, find real slot */
7289 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
7290 mc->mc_top--;
7291 dtop++;
7292 }
7293 if (mc->mc_ki[mc->mc_top])
7294 rc2 = mdb_update_key(mc, key);
7295 else
7296 rc2 = MDB_SUCCESS;
7297 mc->mc_top += dtop;
7298 if (rc2)
7299 return rc2;
7300 }
7301 return MDB_SUCCESS;
7302 }
7303
7304 more:
7305 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7306 olddata.mv_size = NODEDSZ(leaf);
7307 olddata.mv_data = NODEDATA(leaf);
7308
7309 /* DB has dups? */
7310 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
7311 /* Prepare (sub-)page/sub-DB to accept the new item,
7312 * if needed. fp: old sub-page or a header faking
7313 * it. mp: new (sub-)page. offset: growth in page
7314 * size. xdata: node data with new page or DB.
7315 */
7316 unsigned i, offset = 0;
7317 mp = fp = xdata.mv_data = env->me_pbuf;
7318 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
7319
7320 /* Was a single item before, must convert now */
7321 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7322 MDB_cmp_func *dcmp;
7323 /* Just overwrite the current item */
7324 if (flags == MDB_CURRENT)
7325 goto current;
7326 dcmp = mc->mc_dbx->md_dcmp;
7327 if (NEED_CMP_CLONG(dcmp, olddata.mv_size))
7328 dcmp = mdb_cmp_clong;
7329 /* does data match? */
7330 if (!dcmp(data, &olddata)) {
7331 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP))
7332 return MDB_KEYEXIST;
7333 /* overwrite it */
7334 goto current;
7335 }
7336
7337 /* Back up original data item */
7338 dkey.mv_size = olddata.mv_size;
7339 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
7340
7341 /* Make sub-page header for the dup items, with dummy body */
7342 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
7343 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
7344 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
7345 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7346 fp->mp_flags |= P_LEAF2;
7347 fp->mp_pad = data->mv_size;
7348 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
7349 } else {
7350 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
7351 (dkey.mv_size & 1) + (data->mv_size & 1);
7352 }
7353 fp->mp_upper = xdata.mv_size - PAGEBASE;
7354 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
7355 } else if (leaf->mn_flags & F_SUBDATA) {
7356 /* Data is on sub-DB, just store it */
7357 flags |= F_DUPDATA|F_SUBDATA;
7358 goto put_sub;
7359 } else {
7360 /* Data is on sub-page */
7361 fp = olddata.mv_data;
7362 switch (flags) {
7363 default:
7364 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
7365 offset = EVEN(NODESIZE + sizeof(indx_t) +
7366 data->mv_size);
7367 break;
7368 }
7369 offset = fp->mp_pad;
7370 if (SIZELEFT(fp) < offset) {
7371 offset *= 4; /* space for 4 more */
7372 break;
7373 }
7374 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
7375 case MDB_CURRENT:
7376 fp->mp_flags |= P_DIRTY;
7377 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
7378 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
7379 flags |= F_DUPDATA;
7380 goto put_sub;
7381 }
7382 xdata.mv_size = olddata.mv_size + offset;
7383 }
7384
7385 fp_flags = fp->mp_flags;
7386 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
7387 /* Too big for a sub-page, convert to sub-DB */
7388 fp_flags &= ~P_SUBP;
7389 prep_subDB:
7390 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7391 fp_flags |= P_LEAF2;
7392 dummy.md_pad = fp->mp_pad;
7393 dummy.md_flags = MDB_DUPFIXED;
7394 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7395 dummy.md_flags |= MDB_INTEGERKEY;
7396 } else {
7397 dummy.md_pad = 0;
7398 dummy.md_flags = 0;
7399 }
7400 dummy.md_depth = 1;
7401 dummy.md_branch_pages = 0;
7402 dummy.md_leaf_pages = 1;
7403 dummy.md_overflow_pages = 0;
7404 dummy.md_entries = NUMKEYS(fp);
7405 xdata.mv_size = sizeof(MDB_db);
7406 xdata.mv_data = &dummy;
7407 if ((rc = mdb_page_alloc(mc, 1, &mp)))
7408 return rc;
7409 offset = env->me_psize - olddata.mv_size;
7410 flags |= F_DUPDATA|F_SUBDATA;
7411 dummy.md_root = mp->mp_pgno;
7412 sub_root = mp;
7413 }
7414 if (mp != fp) {
7415 mp->mp_flags = fp_flags | P_DIRTY;
7416 mp->mp_pad = fp->mp_pad;
7417 mp->mp_lower = fp->mp_lower;
7418 mp->mp_upper = fp->mp_upper + offset;
7419 if (fp_flags & P_LEAF2) {
7420 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
7421 } else {
7422 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
7423 olddata.mv_size - fp->mp_upper - PAGEBASE);
7424 for (i=0; i<NUMKEYS(fp); i++)
7425 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
7426 }
7427 }
7428
7429 rdata = &xdata;
7430 flags |= F_DUPDATA;
7431 do_sub = 1;
7432 if (!insert_key)
7433 mdb_node_del(mc, 0);
7434 goto new_sub;
7435 }
7436 current:
7437 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
7438 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
7439 return MDB_INCOMPATIBLE;
7440 /* overflow page overwrites need special handling */
7441 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7442 MDB_page *omp;
7443 pgno_t pg;
7444 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
7445
7446 memcpy(&pg, olddata.mv_data, sizeof(pg));
7447 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0)
7448 return rc2;
7449 ovpages = omp->mp_pages;
7450
7451 /* Is the ov page large enough? */
7452 if (ovpages >= dpages) {
7453 if (!(omp->mp_flags & P_DIRTY) &&
7454 (level || (env->me_flags & MDB_WRITEMAP)))
7455 {
7456 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
7457 if (rc)
7458 return rc;
7459 level = 0; /* dirty in this txn or clean */
7460 }
7461 /* Is it dirty? */
7462 if (omp->mp_flags & P_DIRTY) {
7463 /* yes, overwrite it. Note in this case we don't
7464 * bother to try shrinking the page if the new data
7465 * is smaller than the overflow threshold.
7466 */
7467 if (level > 1) {
7468 /* It is writable only in a parent txn */
7469 size_t sz = (size_t) env->me_psize * ovpages, off;
7470 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
7471 MDB_ID2 id2;
7472 if (!np)
7473 return ENOMEM;
7474 id2.mid = pg;
7475 id2.mptr = np;
7476 /* Note - this page is already counted in parent's dirty_room */
7477 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
7478 mdb_cassert(mc, rc2 == 0);
7479 /* Currently we make the page look as with put() in the
7480 * parent txn, in case the user peeks at MDB_RESERVEd
7481 * or unused parts. Some users treat ovpages specially.
7482 */
7483 if (!(flags & MDB_RESERVE)) {
7484 /* Skip the part where LMDB will put *data.
7485 * Copy end of page, adjusting alignment so
7486 * compiler may copy words instead of bytes.
7487 */
7488 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
7489 memcpy((size_t *)((char *)np + off),
7490 (size_t *)((char *)omp + off), sz - off);
7491 sz = PAGEHDRSZ;
7492 }
7493 memcpy(np, omp, sz); /* Copy beginning of page */
7494 omp = np;
7495 }
7496 SETDSZ(leaf, data->mv_size);
7497 if (F_ISSET(flags, MDB_RESERVE))
7498 data->mv_data = METADATA(omp);
7499 else
7500 memcpy(METADATA(omp), data->mv_data, data->mv_size);
7501 return MDB_SUCCESS;
7502 }
7503 }
7504 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
7505 return rc2;
7506 } else if (data->mv_size == olddata.mv_size) {
7507 /* same size, just replace it. Note that we could
7508 * also reuse this node if the new data is smaller,
7509 * but instead we opt to shrink the node in that case.
7510 */
7511 if (F_ISSET(flags, MDB_RESERVE))
7512 data->mv_data = olddata.mv_data;
7513 else if (!(mc->mc_flags & C_SUB))
7514 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
7515 else {
7516 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
7517 goto fix_parent;
7518 }
7519 return MDB_SUCCESS;
7520 }
7521 mdb_node_del(mc, 0);
7522 }
7523
7524 rdata = data;
7525
7526 new_sub:
7527 nflags = flags & NODE_ADD_FLAGS;
7528 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
7529 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
7530 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
7531 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
7532 if (!insert_key)
7533 nflags |= MDB_SPLIT_REPLACE;
7534 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
7535 } else {
7536 /* There is room already in this leaf page. */
7537 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
7538 if (rc == 0) {
7539 /* Adjust other cursors pointing to mp */
7540 MDB_cursor *m2, *m3;
7541 MDB_dbi dbi = mc->mc_dbi;
7542 unsigned i = mc->mc_top;
7543 MDB_page *mp = mc->mc_pg[i];
7544
7545 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7546 if (mc->mc_flags & C_SUB)
7547 m3 = &m2->mc_xcursor->mx_cursor;
7548 else
7549 m3 = m2;
7550 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue;
7551 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) {
7552 m3->mc_ki[i]++;
7553 }
7554 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
7555 MDB_node *n2 = NODEPTR(mp, m3->mc_ki[i]);
7556 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
7557 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7558 }
7559 }
7560 }
7561 }
7562
7563 if (rc == MDB_SUCCESS) {
7564 /* Now store the actual data in the child DB. Note that we're
7565 * storing the user data in the keys field, so there are strict
7566 * size limits on dupdata. The actual data fields of the child
7567 * DB are all zero size.
7568 */
7569 if (do_sub) {
7570 int xflags, new_dupdata;
7571 mdb_size_t ecount;
7572 put_sub:
7573 xdata.mv_size = 0;
7574 xdata.mv_data = "";
7575 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7576 if (flags & MDB_CURRENT) {
7577 xflags = MDB_CURRENT|MDB_NOSPILL;
7578 } else {
7579 mdb_xcursor_init1(mc, leaf);
7580 xflags = (flags & MDB_NODUPDATA) ?
7581 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
7582 }
7583 if (sub_root)
7584 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
7585 new_dupdata = (int)dkey.mv_size;
7586 /* converted, write the original data first */
7587 if (dkey.mv_size) {
7588 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
7589 if (rc)
7590 goto bad_sub;
7591 /* we've done our job */
7592 dkey.mv_size = 0;
7593 }
7594 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
7595 /* Adjust other cursors pointing to mp */
7596 MDB_cursor *m2;
7597 MDB_xcursor *mx = mc->mc_xcursor;
7598 unsigned i = mc->mc_top;
7599 MDB_page *mp = mc->mc_pg[i];
7600 int nkeys = NUMKEYS(mp);
7601
7602 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7603 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7604 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7605 if (m2->mc_pg[i] == mp) {
7606 if (m2->mc_ki[i] == mc->mc_ki[i]) {
7607 mdb_xcursor_init2(m2, mx, new_dupdata);
7608 } else if (!insert_key && m2->mc_ki[i] < nkeys) {
7609 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
7610 if ((n2->mn_flags & (F_SUBDATA|F_DUPDATA)) == F_DUPDATA)
7611 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7612 }
7613 }
7614 }
7615 }
7616 ecount = mc->mc_xcursor->mx_db.md_entries;
7617 if (flags & MDB_APPENDDUP)
7618 xflags |= MDB_APPEND;
7619 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
7620 if (flags & F_SUBDATA) {
7621 void *db = NODEDATA(leaf);
7622 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7623 }
7624 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
7625 }
7626 /* Increment count unless we just replaced an existing item. */
7627 if (insert_data)
7628 mc->mc_db->md_entries++;
7629 if (insert_key) {
7630 /* Invalidate txn if we created an empty sub-DB */
7631 if (rc)
7632 goto bad_sub;
7633 /* If we succeeded and the key didn't exist before,
7634 * make sure the cursor is marked valid.
7635 */
7636 mc->mc_flags |= C_INITIALIZED;
7637 }
7638 if (flags & MDB_MULTIPLE) {
7639 if (!rc) {
7640 mcount++;
7641 /* let caller know how many succeeded, if any */
7642 data[1].mv_size = mcount;
7643 if (mcount < dcount) {
7644 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
7645 insert_key = insert_data = 0;
7646 goto more;
7647 }
7648 }
7649 }
7650 return rc;
7651 bad_sub:
7652 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
7653 rc = MDB_PROBLEM;
7654 }
7655 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7656 return rc;
7657 }
7658
7659 int
mdb_cursor_del(MDB_cursor * mc,unsigned int flags)7660 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
7661 {
7662 MDB_node *leaf;
7663 MDB_page *mp;
7664 int rc;
7665
7666 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
7667 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7668
7669 if (!(mc->mc_flags & C_INITIALIZED))
7670 return EINVAL;
7671
7672 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7673 return MDB_NOTFOUND;
7674
7675 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
7676 return rc;
7677
7678 rc = mdb_cursor_touch(mc);
7679 if (rc)
7680 return rc;
7681
7682 mp = mc->mc_pg[mc->mc_top];
7683 if (IS_LEAF2(mp))
7684 goto del_key;
7685 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7686
7687 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7688 if (flags & MDB_NODUPDATA) {
7689 /* mdb_cursor_del0() will subtract the final entry */
7690 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
7691 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7692 } else {
7693 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
7694 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7695 }
7696 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
7697 if (rc)
7698 return rc;
7699 /* If sub-DB still has entries, we're done */
7700 if (mc->mc_xcursor->mx_db.md_entries) {
7701 if (leaf->mn_flags & F_SUBDATA) {
7702 /* update subDB info */
7703 void *db = NODEDATA(leaf);
7704 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
7705 } else {
7706 MDB_cursor *m2;
7707 /* shrink fake page */
7708 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
7709 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7710 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7711 /* fix other sub-DB cursors pointed at fake pages on this page */
7712 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
7713 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
7714 if (!(m2->mc_flags & C_INITIALIZED)) continue;
7715 if (m2->mc_pg[mc->mc_top] == mp) {
7716 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
7717 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
7718 } else {
7719 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
7720 if (!(n2->mn_flags & F_SUBDATA))
7721 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
7722 }
7723 }
7724 }
7725 }
7726 mc->mc_db->md_entries--;
7727 return rc;
7728 } else {
7729 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
7730 }
7731 /* otherwise fall thru and delete the sub-DB */
7732 }
7733
7734 if (leaf->mn_flags & F_SUBDATA) {
7735 /* add all the child DB's pages to the free list */
7736 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7737 if (rc)
7738 goto fail;
7739 }
7740 }
7741 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
7742 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
7743 rc = MDB_INCOMPATIBLE;
7744 goto fail;
7745 }
7746
7747 /* add overflow pages to free list */
7748 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7749 MDB_page *omp;
7750 pgno_t pg;
7751
7752 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7753 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) ||
7754 (rc = mdb_ovpage_free(mc, omp)))
7755 goto fail;
7756 }
7757
7758 del_key:
7759 return mdb_cursor_del0(mc);
7760
7761 fail:
7762 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7763 return rc;
7764 }
7765
7766 /** Allocate and initialize new pages for a database.
7767 * @param[in] mc a cursor on the database being added to.
7768 * @param[in] flags flags defining what type of page is being allocated.
7769 * @param[in] num the number of pages to allocate. This is usually 1,
7770 * unless allocating overflow pages for a large record.
7771 * @param[out] mp Address of a page, or NULL on failure.
7772 * @return 0 on success, non-zero on failure.
7773 */
7774 static int
mdb_page_new(MDB_cursor * mc,uint32_t flags,int num,MDB_page ** mp)7775 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
7776 {
7777 MDB_page *np;
7778 int rc;
7779
7780 if ((rc = mdb_page_alloc(mc, num, &np)))
7781 return rc;
7782 DPRINTF(("allocated new mpage %"Yu", page size %u",
7783 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
7784 np->mp_flags = flags | P_DIRTY;
7785 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
7786 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
7787
7788 if (IS_BRANCH(np))
7789 mc->mc_db->md_branch_pages++;
7790 else if (IS_LEAF(np))
7791 mc->mc_db->md_leaf_pages++;
7792 else if (IS_OVERFLOW(np)) {
7793 mc->mc_db->md_overflow_pages += num;
7794 np->mp_pages = num;
7795 }
7796 *mp = np;
7797
7798 return 0;
7799 }
7800
7801 /** Calculate the size of a leaf node.
7802 * The size depends on the environment's page size; if a data item
7803 * is too large it will be put onto an overflow page and the node
7804 * size will only include the key and not the data. Sizes are always
7805 * rounded up to an even number of bytes, to guarantee 2-byte alignment
7806 * of the #MDB_node headers.
7807 * @param[in] env The environment handle.
7808 * @param[in] key The key for the node.
7809 * @param[in] data The data for the node.
7810 * @return The number of bytes needed to store the node.
7811 */
7812 static size_t
mdb_leaf_size(MDB_env * env,MDB_val * key,MDB_val * data)7813 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
7814 {
7815 size_t sz;
7816
7817 sz = LEAFSIZE(key, data);
7818 if (sz > env->me_nodemax) {
7819 /* put on overflow page */
7820 sz -= data->mv_size - sizeof(pgno_t);
7821 }
7822
7823 return EVEN(sz + sizeof(indx_t));
7824 }
7825
7826 /** Calculate the size of a branch node.
7827 * The size should depend on the environment's page size but since
7828 * we currently don't support spilling large keys onto overflow
7829 * pages, it's simply the size of the #MDB_node header plus the
7830 * size of the key. Sizes are always rounded up to an even number
7831 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
7832 * @param[in] env The environment handle.
7833 * @param[in] key The key for the node.
7834 * @return The number of bytes needed to store the node.
7835 */
7836 static size_t
mdb_branch_size(MDB_env * env,MDB_val * key)7837 mdb_branch_size(MDB_env *env, MDB_val *key)
7838 {
7839 size_t sz;
7840
7841 sz = INDXSIZE(key);
7842 if (sz > env->me_nodemax) {
7843 /* put on overflow page */
7844 /* not implemented */
7845 /* sz -= key->size - sizeof(pgno_t); */
7846 }
7847
7848 return sz + sizeof(indx_t);
7849 }
7850
7851 /** Add a node to the page pointed to by the cursor.
7852 * @param[in] mc The cursor for this operation.
7853 * @param[in] indx The index on the page where the new node should be added.
7854 * @param[in] key The key for the new node.
7855 * @param[in] data The data for the new node, if any.
7856 * @param[in] pgno The page number, if adding a branch node.
7857 * @param[in] flags Flags for the node.
7858 * @return 0 on success, non-zero on failure. Possible errors are:
7859 * <ul>
7860 * <li>ENOMEM - failed to allocate overflow pages for the node.
7861 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
7862 * should never happen since all callers already calculate the
7863 * page's free space before calling this function.
7864 * </ul>
7865 */
7866 static int
mdb_node_add(MDB_cursor * mc,indx_t indx,MDB_val * key,MDB_val * data,pgno_t pgno,unsigned int flags)7867 mdb_node_add(MDB_cursor *mc, indx_t indx,
7868 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
7869 {
7870 unsigned int i;
7871 size_t node_size = NODESIZE;
7872 ssize_t room;
7873 indx_t ofs;
7874 MDB_node *node;
7875 MDB_page *mp = mc->mc_pg[mc->mc_top];
7876 MDB_page *ofp = NULL; /* overflow page */
7877 void *ndata;
7878 DKBUF;
7879
7880 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
7881
7882 DPRINTF(("add to %s %spage %"Yu" index %i, data size %"Z"u key size %"Z"u [%s]",
7883 IS_LEAF(mp) ? "leaf" : "branch",
7884 IS_SUBP(mp) ? "sub-" : "",
7885 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
7886 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
7887
7888 if (IS_LEAF2(mp)) {
7889 /* Move higher keys up one slot. */
7890 int ksize = mc->mc_db->md_pad, dif;
7891 char *ptr = LEAF2KEY(mp, indx, ksize);
7892 dif = NUMKEYS(mp) - indx;
7893 if (dif > 0)
7894 memmove(ptr+ksize, ptr, dif*ksize);
7895 /* insert new key */
7896 memcpy(ptr, key->mv_data, ksize);
7897
7898 /* Just using these for counting */
7899 mp->mp_lower += sizeof(indx_t);
7900 mp->mp_upper -= ksize - sizeof(indx_t);
7901 return MDB_SUCCESS;
7902 }
7903
7904 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
7905 if (key != NULL)
7906 node_size += key->mv_size;
7907 if (IS_LEAF(mp)) {
7908 mdb_cassert(mc, key && data);
7909 if (F_ISSET(flags, F_BIGDATA)) {
7910 /* Data already on overflow page. */
7911 node_size += sizeof(pgno_t);
7912 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
7913 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
7914 int rc;
7915 /* Put data on overflow page. */
7916 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
7917 data->mv_size, node_size+data->mv_size));
7918 node_size = EVEN(node_size + sizeof(pgno_t));
7919 if ((ssize_t)node_size > room)
7920 goto full;
7921 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
7922 return rc;
7923 DPRINTF(("allocated overflow page %"Yu, ofp->mp_pgno));
7924 flags |= F_BIGDATA;
7925 goto update;
7926 } else {
7927 node_size += data->mv_size;
7928 }
7929 }
7930 node_size = EVEN(node_size);
7931 if ((ssize_t)node_size > room)
7932 goto full;
7933
7934 update:
7935 /* Move higher pointers up one slot. */
7936 for (i = NUMKEYS(mp); i > indx; i--)
7937 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
7938
7939 /* Adjust free space offsets. */
7940 ofs = mp->mp_upper - node_size;
7941 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
7942 mp->mp_ptrs[indx] = ofs;
7943 mp->mp_upper = ofs;
7944 mp->mp_lower += sizeof(indx_t);
7945
7946 /* Write the node data. */
7947 node = NODEPTR(mp, indx);
7948 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
7949 node->mn_flags = flags;
7950 if (IS_LEAF(mp))
7951 SETDSZ(node,data->mv_size);
7952 else
7953 SETPGNO(node,pgno);
7954
7955 if (key)
7956 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7957
7958 if (IS_LEAF(mp)) {
7959 ndata = NODEDATA(node);
7960 if (ofp == NULL) {
7961 if (F_ISSET(flags, F_BIGDATA))
7962 memcpy(ndata, data->mv_data, sizeof(pgno_t));
7963 else if (F_ISSET(flags, MDB_RESERVE))
7964 data->mv_data = ndata;
7965 else
7966 memcpy(ndata, data->mv_data, data->mv_size);
7967 } else {
7968 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t));
7969 ndata = METADATA(ofp);
7970 if (F_ISSET(flags, MDB_RESERVE))
7971 data->mv_data = ndata;
7972 else
7973 memcpy(ndata, data->mv_data, data->mv_size);
7974 }
7975 }
7976
7977 return MDB_SUCCESS;
7978
7979 full:
7980 DPRINTF(("not enough room in page %"Yu", got %u ptrs",
7981 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7982 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7983 DPRINTF(("node size = %"Z"u", node_size));
7984 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7985 return MDB_PAGE_FULL;
7986 }
7987
7988 /** Delete the specified node from a page.
7989 * @param[in] mc Cursor pointing to the node to delete.
7990 * @param[in] ksize The size of a node. Only used if the page is
7991 * part of a #MDB_DUPFIXED database.
7992 */
7993 static void
mdb_node_del(MDB_cursor * mc,int ksize)7994 mdb_node_del(MDB_cursor *mc, int ksize)
7995 {
7996 MDB_page *mp = mc->mc_pg[mc->mc_top];
7997 indx_t indx = mc->mc_ki[mc->mc_top];
7998 unsigned int sz;
7999 indx_t i, j, numkeys, ptr;
8000 MDB_node *node;
8001 char *base;
8002
8003 DPRINTF(("delete node %u on %s page %"Yu, indx,
8004 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
8005 numkeys = NUMKEYS(mp);
8006 mdb_cassert(mc, indx < numkeys);
8007
8008 if (IS_LEAF2(mp)) {
8009 int x = numkeys - 1 - indx;
8010 base = LEAF2KEY(mp, indx, ksize);
8011 if (x)
8012 memmove(base, base + ksize, x * ksize);
8013 mp->mp_lower -= sizeof(indx_t);
8014 mp->mp_upper += ksize - sizeof(indx_t);
8015 return;
8016 }
8017
8018 node = NODEPTR(mp, indx);
8019 sz = NODESIZE + node->mn_ksize;
8020 if (IS_LEAF(mp)) {
8021 if (F_ISSET(node->mn_flags, F_BIGDATA))
8022 sz += sizeof(pgno_t);
8023 else
8024 sz += NODEDSZ(node);
8025 }
8026 sz = EVEN(sz);
8027
8028 ptr = mp->mp_ptrs[indx];
8029 for (i = j = 0; i < numkeys; i++) {
8030 if (i != indx) {
8031 mp->mp_ptrs[j] = mp->mp_ptrs[i];
8032 if (mp->mp_ptrs[i] < ptr)
8033 mp->mp_ptrs[j] += sz;
8034 j++;
8035 }
8036 }
8037
8038 base = (char *)mp + mp->mp_upper + PAGEBASE;
8039 memmove(base + sz, base, ptr - mp->mp_upper);
8040
8041 mp->mp_lower -= sizeof(indx_t);
8042 mp->mp_upper += sz;
8043 }
8044
8045 /** Compact the main page after deleting a node on a subpage.
8046 * @param[in] mp The main page to operate on.
8047 * @param[in] indx The index of the subpage on the main page.
8048 */
8049 static void
mdb_node_shrink(MDB_page * mp,indx_t indx)8050 mdb_node_shrink(MDB_page *mp, indx_t indx)
8051 {
8052 MDB_node *node;
8053 MDB_page *sp, *xp;
8054 char *base;
8055 indx_t delta, nsize, len, ptr;
8056 int i;
8057
8058 node = NODEPTR(mp, indx);
8059 sp = (MDB_page *)NODEDATA(node);
8060 delta = SIZELEFT(sp);
8061 nsize = NODEDSZ(node) - delta;
8062
8063 /* Prepare to shift upward, set len = length(subpage part to shift) */
8064 if (IS_LEAF2(sp)) {
8065 len = nsize;
8066 if (nsize & 1)
8067 return; /* do not make the node uneven-sized */
8068 } else {
8069 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */
8070 for (i = NUMKEYS(sp); --i >= 0; )
8071 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
8072 len = PAGEHDRSZ;
8073 }
8074 sp->mp_upper = sp->mp_lower;
8075 COPY_PGNO(sp->mp_pgno, mp->mp_pgno);
8076 SETDSZ(node, nsize);
8077
8078 /* Shift <lower nodes...initial part of subpage> upward */
8079 base = (char *)mp + mp->mp_upper + PAGEBASE;
8080 memmove(base + delta, base, (char *)sp + len - base);
8081
8082 ptr = mp->mp_ptrs[indx];
8083 for (i = NUMKEYS(mp); --i >= 0; ) {
8084 if (mp->mp_ptrs[i] <= ptr)
8085 mp->mp_ptrs[i] += delta;
8086 }
8087 mp->mp_upper += delta;
8088 }
8089
8090 /** Initial setup of a sorted-dups cursor.
8091 * Sorted duplicates are implemented as a sub-database for the given key.
8092 * The duplicate data items are actually keys of the sub-database.
8093 * Operations on the duplicate data items are performed using a sub-cursor
8094 * initialized when the sub-database is first accessed. This function does
8095 * the preliminary setup of the sub-cursor, filling in the fields that
8096 * depend only on the parent DB.
8097 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8098 */
8099 static void
mdb_xcursor_init0(MDB_cursor * mc)8100 mdb_xcursor_init0(MDB_cursor *mc)
8101 {
8102 MDB_xcursor *mx = mc->mc_xcursor;
8103
8104 mx->mx_cursor.mc_xcursor = NULL;
8105 mx->mx_cursor.mc_txn = mc->mc_txn;
8106 mx->mx_cursor.mc_db = &mx->mx_db;
8107 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
8108 mx->mx_cursor.mc_dbi = mc->mc_dbi;
8109 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
8110 mx->mx_cursor.mc_snum = 0;
8111 mx->mx_cursor.mc_top = 0;
8112 MC_SET_OVPG(&mx->mx_cursor, NULL);
8113 mx->mx_cursor.mc_flags = C_SUB | (mc->mc_flags & (C_ORIG_RDONLY|C_WRITEMAP));
8114 mx->mx_dbx.md_name.mv_size = 0;
8115 mx->mx_dbx.md_name.mv_data = NULL;
8116 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
8117 mx->mx_dbx.md_dcmp = NULL;
8118 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
8119 }
8120
8121 /** Final setup of a sorted-dups cursor.
8122 * Sets up the fields that depend on the data from the main cursor.
8123 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
8124 * @param[in] node The data containing the #MDB_db record for the
8125 * sorted-dup database.
8126 */
8127 static void
mdb_xcursor_init1(MDB_cursor * mc,MDB_node * node)8128 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
8129 {
8130 MDB_xcursor *mx = mc->mc_xcursor;
8131
8132 mx->mx_cursor.mc_flags &= C_SUB|C_ORIG_RDONLY|C_WRITEMAP;
8133 if (node->mn_flags & F_SUBDATA) {
8134 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
8135 mx->mx_cursor.mc_pg[0] = 0;
8136 mx->mx_cursor.mc_snum = 0;
8137 mx->mx_cursor.mc_top = 0;
8138 } else {
8139 MDB_page *fp = NODEDATA(node);
8140 mx->mx_db.md_pad = 0;
8141 mx->mx_db.md_flags = 0;
8142 mx->mx_db.md_depth = 1;
8143 mx->mx_db.md_branch_pages = 0;
8144 mx->mx_db.md_leaf_pages = 1;
8145 mx->mx_db.md_overflow_pages = 0;
8146 mx->mx_db.md_entries = NUMKEYS(fp);
8147 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
8148 mx->mx_cursor.mc_snum = 1;
8149 mx->mx_cursor.mc_top = 0;
8150 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8151 mx->mx_cursor.mc_pg[0] = fp;
8152 mx->mx_cursor.mc_ki[0] = 0;
8153 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
8154 mx->mx_db.md_flags = MDB_DUPFIXED;
8155 mx->mx_db.md_pad = fp->mp_pad;
8156 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
8157 mx->mx_db.md_flags |= MDB_INTEGERKEY;
8158 }
8159 }
8160 DPRINTF(("Sub-db -%u root page %"Yu, mx->mx_cursor.mc_dbi,
8161 mx->mx_db.md_root));
8162 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
8163 if (NEED_CMP_CLONG(mx->mx_dbx.md_cmp, mx->mx_db.md_pad))
8164 mx->mx_dbx.md_cmp = mdb_cmp_clong;
8165 }
8166
8167
8168 /** Fixup a sorted-dups cursor due to underlying update.
8169 * Sets up some fields that depend on the data from the main cursor.
8170 * Almost the same as init1, but skips initialization steps if the
8171 * xcursor had already been used.
8172 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
8173 * @param[in] src_mx The xcursor of an up-to-date cursor.
8174 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
8175 */
8176 static void
mdb_xcursor_init2(MDB_cursor * mc,MDB_xcursor * src_mx,int new_dupdata)8177 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
8178 {
8179 MDB_xcursor *mx = mc->mc_xcursor;
8180
8181 if (new_dupdata) {
8182 mx->mx_cursor.mc_snum = 1;
8183 mx->mx_cursor.mc_top = 0;
8184 mx->mx_cursor.mc_flags |= C_INITIALIZED;
8185 mx->mx_cursor.mc_ki[0] = 0;
8186 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
8187 #if UINT_MAX < MDB_SIZE_MAX /* matches mdb_xcursor_init1:NEED_CMP_CLONG() */
8188 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
8189 #endif
8190 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
8191 return;
8192 }
8193 mx->mx_db = src_mx->mx_db;
8194 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
8195 DPRINTF(("Sub-db -%u root page %"Yu, mx->mx_cursor.mc_dbi,
8196 mx->mx_db.md_root));
8197 }
8198
8199 /** Initialize a cursor for a given transaction and database. */
8200 static void
mdb_cursor_init(MDB_cursor * mc,MDB_txn * txn,MDB_dbi dbi,MDB_xcursor * mx)8201 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
8202 {
8203 mc->mc_next = NULL;
8204 mc->mc_backup = NULL;
8205 mc->mc_dbi = dbi;
8206 mc->mc_txn = txn;
8207 mc->mc_db = &txn->mt_dbs[dbi];
8208 mc->mc_dbx = &txn->mt_dbxs[dbi];
8209 mc->mc_dbflag = &txn->mt_dbflags[dbi];
8210 mc->mc_snum = 0;
8211 mc->mc_top = 0;
8212 mc->mc_pg[0] = 0;
8213 mc->mc_ki[0] = 0;
8214 MC_SET_OVPG(mc, NULL);
8215 mc->mc_flags = txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
8216 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
8217 mdb_tassert(txn, mx != NULL);
8218 mc->mc_xcursor = mx;
8219 mdb_xcursor_init0(mc);
8220 } else {
8221 mc->mc_xcursor = NULL;
8222 }
8223 if (*mc->mc_dbflag & DB_STALE) {
8224 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
8225 }
8226 }
8227
8228 int
mdb_cursor_open(MDB_txn * txn,MDB_dbi dbi,MDB_cursor ** ret)8229 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
8230 {
8231 MDB_cursor *mc;
8232 size_t size = sizeof(MDB_cursor);
8233
8234 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
8235 return EINVAL;
8236
8237 if (txn->mt_flags & MDB_TXN_BLOCKED)
8238 return MDB_BAD_TXN;
8239
8240 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8241 return EINVAL;
8242
8243 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
8244 size += sizeof(MDB_xcursor);
8245
8246 if ((mc = malloc(size)) != NULL) {
8247 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
8248 if (txn->mt_cursors) {
8249 mc->mc_next = txn->mt_cursors[dbi];
8250 txn->mt_cursors[dbi] = mc;
8251 mc->mc_flags |= C_UNTRACK;
8252 }
8253 } else {
8254 return ENOMEM;
8255 }
8256
8257 *ret = mc;
8258
8259 return MDB_SUCCESS;
8260 }
8261
8262 int
mdb_cursor_renew(MDB_txn * txn,MDB_cursor * mc)8263 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
8264 {
8265 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID))
8266 return EINVAL;
8267
8268 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
8269 return EINVAL;
8270
8271 if (txn->mt_flags & MDB_TXN_BLOCKED)
8272 return MDB_BAD_TXN;
8273
8274 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
8275 return MDB_SUCCESS;
8276 }
8277
8278 /* Return the count of duplicate data items for the current key */
8279 int
mdb_cursor_count(MDB_cursor * mc,mdb_size_t * countp)8280 mdb_cursor_count(MDB_cursor *mc, mdb_size_t *countp)
8281 {
8282 MDB_node *leaf;
8283
8284 if (mc == NULL || countp == NULL)
8285 return EINVAL;
8286
8287 if (mc->mc_xcursor == NULL)
8288 return MDB_INCOMPATIBLE;
8289
8290 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED)
8291 return MDB_BAD_TXN;
8292
8293 if (!(mc->mc_flags & C_INITIALIZED))
8294 return EINVAL;
8295
8296 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
8297 return MDB_NOTFOUND;
8298
8299 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
8300 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
8301 *countp = 1;
8302 } else {
8303 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
8304 return EINVAL;
8305
8306 *countp = mc->mc_xcursor->mx_db.md_entries;
8307 }
8308 return MDB_SUCCESS;
8309 }
8310
8311 void
mdb_cursor_close(MDB_cursor * mc)8312 mdb_cursor_close(MDB_cursor *mc)
8313 {
8314 if (mc) {
8315 MDB_CURSOR_UNREF(mc, 0);
8316 }
8317 if (mc && !mc->mc_backup) {
8318 /* remove from txn, if tracked */
8319 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
8320 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
8321 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
8322 if (*prev == mc)
8323 *prev = mc->mc_next;
8324 }
8325 free(mc);
8326 }
8327 }
8328
8329 MDB_txn *
mdb_cursor_txn(MDB_cursor * mc)8330 mdb_cursor_txn(MDB_cursor *mc)
8331 {
8332 if (!mc) return NULL;
8333 return mc->mc_txn;
8334 }
8335
8336 MDB_dbi
mdb_cursor_dbi(MDB_cursor * mc)8337 mdb_cursor_dbi(MDB_cursor *mc)
8338 {
8339 return mc->mc_dbi;
8340 }
8341
8342 /** Replace the key for a branch node with a new key.
8343 * @param[in] mc Cursor pointing to the node to operate on.
8344 * @param[in] key The new key to use.
8345 * @return 0 on success, non-zero on failure.
8346 */
8347 static int
mdb_update_key(MDB_cursor * mc,MDB_val * key)8348 mdb_update_key(MDB_cursor *mc, MDB_val *key)
8349 {
8350 MDB_page *mp;
8351 MDB_node *node;
8352 char *base;
8353 size_t len;
8354 int delta, ksize, oksize;
8355 indx_t ptr, i, numkeys, indx;
8356 DKBUF;
8357
8358 indx = mc->mc_ki[mc->mc_top];
8359 mp = mc->mc_pg[mc->mc_top];
8360 node = NODEPTR(mp, indx);
8361 ptr = mp->mp_ptrs[indx];
8362 #if MDB_DEBUG
8363 {
8364 MDB_val k2;
8365 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
8366 k2.mv_data = NODEKEY(node);
8367 k2.mv_size = node->mn_ksize;
8368 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Yu,
8369 indx, ptr,
8370 mdb_dkey(&k2, kbuf2),
8371 DKEY(key),
8372 mp->mp_pgno));
8373 }
8374 #endif
8375
8376 /* Sizes must be 2-byte aligned. */
8377 ksize = EVEN(key->mv_size);
8378 oksize = EVEN(node->mn_ksize);
8379 delta = ksize - oksize;
8380
8381 /* Shift node contents if EVEN(key length) changed. */
8382 if (delta) {
8383 if (delta > 0 && SIZELEFT(mp) < delta) {
8384 pgno_t pgno;
8385 /* not enough space left, do a delete and split */
8386 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
8387 pgno = NODEPGNO(node);
8388 mdb_node_del(mc, 0);
8389 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
8390 }
8391
8392 numkeys = NUMKEYS(mp);
8393 for (i = 0; i < numkeys; i++) {
8394 if (mp->mp_ptrs[i] <= ptr)
8395 mp->mp_ptrs[i] -= delta;
8396 }
8397
8398 base = (char *)mp + mp->mp_upper + PAGEBASE;
8399 len = ptr - mp->mp_upper + NODESIZE;
8400 memmove(base - delta, base, len);
8401 mp->mp_upper -= delta;
8402
8403 node = NODEPTR(mp, indx);
8404 }
8405
8406 /* But even if no shift was needed, update ksize */
8407 if (node->mn_ksize != key->mv_size)
8408 node->mn_ksize = key->mv_size;
8409
8410 if (key->mv_size)
8411 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
8412
8413 return MDB_SUCCESS;
8414 }
8415
8416 static void
8417 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
8418
8419 /** Perform \b act while tracking temporary cursor \b mn */
8420 #define WITH_CURSOR_TRACKING(mn, act) do { \
8421 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \
8422 if ((mn).mc_flags & C_SUB) { \
8423 dummy.mc_flags = C_INITIALIZED; \
8424 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \
8425 tracked = &dummy; \
8426 } else { \
8427 tracked = &(mn); \
8428 } \
8429 tracked->mc_next = *tp; \
8430 *tp = tracked; \
8431 { act; } \
8432 *tp = tracked->mc_next; \
8433 } while (0)
8434
8435 /** Move a node from csrc to cdst.
8436 */
8437 static int
mdb_node_move(MDB_cursor * csrc,MDB_cursor * cdst,int fromleft)8438 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft)
8439 {
8440 MDB_node *srcnode;
8441 MDB_val key, data;
8442 pgno_t srcpg;
8443 MDB_cursor mn;
8444 int rc;
8445 unsigned short flags;
8446
8447 DKBUF;
8448
8449 /* Mark src and dst as dirty. */
8450 if ((rc = mdb_page_touch(csrc)) ||
8451 (rc = mdb_page_touch(cdst)))
8452 return rc;
8453
8454 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8455 key.mv_size = csrc->mc_db->md_pad;
8456 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
8457 data.mv_size = 0;
8458 data.mv_data = NULL;
8459 srcpg = 0;
8460 flags = 0;
8461 } else {
8462 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
8463 mdb_cassert(csrc, !((size_t)srcnode & 1));
8464 srcpg = NODEPGNO(srcnode);
8465 flags = srcnode->mn_flags;
8466 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8467 unsigned int snum = csrc->mc_snum;
8468 MDB_node *s2;
8469 /* must find the lowest key below src */
8470 rc = mdb_page_search_lowest(csrc);
8471 if (rc)
8472 return rc;
8473 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8474 key.mv_size = csrc->mc_db->md_pad;
8475 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8476 } else {
8477 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8478 key.mv_size = NODEKSZ(s2);
8479 key.mv_data = NODEKEY(s2);
8480 }
8481 csrc->mc_snum = snum--;
8482 csrc->mc_top = snum;
8483 } else {
8484 key.mv_size = NODEKSZ(srcnode);
8485 key.mv_data = NODEKEY(srcnode);
8486 }
8487 data.mv_size = NODEDSZ(srcnode);
8488 data.mv_data = NODEDATA(srcnode);
8489 }
8490 mn.mc_xcursor = NULL;
8491 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
8492 unsigned int snum = cdst->mc_snum;
8493 MDB_node *s2;
8494 MDB_val bkey;
8495 /* must find the lowest key below dst */
8496 mdb_cursor_copy(cdst, &mn);
8497 rc = mdb_page_search_lowest(&mn);
8498 if (rc)
8499 return rc;
8500 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8501 bkey.mv_size = mn.mc_db->md_pad;
8502 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
8503 } else {
8504 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8505 bkey.mv_size = NODEKSZ(s2);
8506 bkey.mv_data = NODEKEY(s2);
8507 }
8508 mn.mc_snum = snum--;
8509 mn.mc_top = snum;
8510 mn.mc_ki[snum] = 0;
8511 rc = mdb_update_key(&mn, &bkey);
8512 if (rc)
8513 return rc;
8514 }
8515
8516 DPRINTF(("moving %s node %u [%s] on page %"Yu" to node %u on page %"Yu,
8517 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
8518 csrc->mc_ki[csrc->mc_top],
8519 DKEY(&key),
8520 csrc->mc_pg[csrc->mc_top]->mp_pgno,
8521 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
8522
8523 /* Add the node to the destination page.
8524 */
8525 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
8526 if (rc != MDB_SUCCESS)
8527 return rc;
8528
8529 /* Delete the node from the source page.
8530 */
8531 mdb_node_del(csrc, key.mv_size);
8532
8533 {
8534 /* Adjust other cursors pointing to mp */
8535 MDB_cursor *m2, *m3;
8536 MDB_dbi dbi = csrc->mc_dbi;
8537 MDB_page *mpd, *mps;
8538
8539 mps = csrc->mc_pg[csrc->mc_top];
8540 /* If we're adding on the left, bump others up */
8541 if (fromleft) {
8542 mpd = cdst->mc_pg[csrc->mc_top];
8543 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8544 if (csrc->mc_flags & C_SUB)
8545 m3 = &m2->mc_xcursor->mx_cursor;
8546 else
8547 m3 = m2;
8548 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8549 continue;
8550 if (m3 != cdst &&
8551 m3->mc_pg[csrc->mc_top] == mpd &&
8552 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) {
8553 m3->mc_ki[csrc->mc_top]++;
8554 }
8555 if (m3 !=csrc &&
8556 m3->mc_pg[csrc->mc_top] == mps &&
8557 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) {
8558 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8559 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8560 m3->mc_ki[csrc->mc_top-1]++;
8561 }
8562 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8563 IS_LEAF(mps)) {
8564 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
8565 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8566 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8567 }
8568 }
8569 } else
8570 /* Adding on the right, bump others down */
8571 {
8572 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8573 if (csrc->mc_flags & C_SUB)
8574 m3 = &m2->mc_xcursor->mx_cursor;
8575 else
8576 m3 = m2;
8577 if (m3 == csrc) continue;
8578 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top)
8579 continue;
8580 if (m3->mc_pg[csrc->mc_top] == mps) {
8581 if (!m3->mc_ki[csrc->mc_top]) {
8582 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
8583 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
8584 m3->mc_ki[csrc->mc_top-1]--;
8585 } else {
8586 m3->mc_ki[csrc->mc_top]--;
8587 }
8588 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8589 IS_LEAF(mps)) {
8590 MDB_node *node = NODEPTR(m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]);
8591 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8592 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8593 }
8594 }
8595 }
8596 }
8597 }
8598
8599 /* Update the parent separators.
8600 */
8601 if (csrc->mc_ki[csrc->mc_top] == 0) {
8602 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
8603 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8604 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
8605 } else {
8606 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
8607 key.mv_size = NODEKSZ(srcnode);
8608 key.mv_data = NODEKEY(srcnode);
8609 }
8610 DPRINTF(("update separator for source page %"Yu" to [%s]",
8611 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
8612 mdb_cursor_copy(csrc, &mn);
8613 mn.mc_snum--;
8614 mn.mc_top--;
8615 /* We want mdb_rebalance to find mn when doing fixups */
8616 WITH_CURSOR_TRACKING(mn,
8617 rc = mdb_update_key(&mn, &key));
8618 if (rc)
8619 return rc;
8620 }
8621 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
8622 MDB_val nullkey;
8623 indx_t ix = csrc->mc_ki[csrc->mc_top];
8624 nullkey.mv_size = 0;
8625 csrc->mc_ki[csrc->mc_top] = 0;
8626 rc = mdb_update_key(csrc, &nullkey);
8627 csrc->mc_ki[csrc->mc_top] = ix;
8628 mdb_cassert(csrc, rc == MDB_SUCCESS);
8629 }
8630 }
8631
8632 if (cdst->mc_ki[cdst->mc_top] == 0) {
8633 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
8634 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
8635 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
8636 } else {
8637 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
8638 key.mv_size = NODEKSZ(srcnode);
8639 key.mv_data = NODEKEY(srcnode);
8640 }
8641 DPRINTF(("update separator for destination page %"Yu" to [%s]",
8642 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
8643 mdb_cursor_copy(cdst, &mn);
8644 mn.mc_snum--;
8645 mn.mc_top--;
8646 /* We want mdb_rebalance to find mn when doing fixups */
8647 WITH_CURSOR_TRACKING(mn,
8648 rc = mdb_update_key(&mn, &key));
8649 if (rc)
8650 return rc;
8651 }
8652 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
8653 MDB_val nullkey;
8654 indx_t ix = cdst->mc_ki[cdst->mc_top];
8655 nullkey.mv_size = 0;
8656 cdst->mc_ki[cdst->mc_top] = 0;
8657 rc = mdb_update_key(cdst, &nullkey);
8658 cdst->mc_ki[cdst->mc_top] = ix;
8659 mdb_cassert(cdst, rc == MDB_SUCCESS);
8660 }
8661 }
8662
8663 return MDB_SUCCESS;
8664 }
8665
8666 /** Merge one page into another.
8667 * The nodes from the page pointed to by \b csrc will
8668 * be copied to the page pointed to by \b cdst and then
8669 * the \b csrc page will be freed.
8670 * @param[in] csrc Cursor pointing to the source page.
8671 * @param[in] cdst Cursor pointing to the destination page.
8672 * @return 0 on success, non-zero on failure.
8673 */
8674 static int
mdb_page_merge(MDB_cursor * csrc,MDB_cursor * cdst)8675 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
8676 {
8677 MDB_page *psrc, *pdst;
8678 MDB_node *srcnode;
8679 MDB_val key, data;
8680 unsigned nkeys;
8681 int rc;
8682 indx_t i, j;
8683
8684 psrc = csrc->mc_pg[csrc->mc_top];
8685 pdst = cdst->mc_pg[cdst->mc_top];
8686
8687 DPRINTF(("merging page %"Yu" into %"Yu, psrc->mp_pgno, pdst->mp_pgno));
8688
8689 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
8690 mdb_cassert(csrc, cdst->mc_snum > 1);
8691
8692 /* Mark dst as dirty. */
8693 if ((rc = mdb_page_touch(cdst)))
8694 return rc;
8695
8696 /* get dst page again now that we've touched it. */
8697 pdst = cdst->mc_pg[cdst->mc_top];
8698
8699 /* Move all nodes from src to dst.
8700 */
8701 j = nkeys = NUMKEYS(pdst);
8702 if (IS_LEAF2(psrc)) {
8703 key.mv_size = csrc->mc_db->md_pad;
8704 key.mv_data = METADATA(psrc);
8705 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8706 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
8707 if (rc != MDB_SUCCESS)
8708 return rc;
8709 key.mv_data = (char *)key.mv_data + key.mv_size;
8710 }
8711 } else {
8712 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
8713 srcnode = NODEPTR(psrc, i);
8714 if (i == 0 && IS_BRANCH(psrc)) {
8715 MDB_cursor mn;
8716 MDB_node *s2;
8717 mdb_cursor_copy(csrc, &mn);
8718 mn.mc_xcursor = NULL;
8719 /* must find the lowest key below src */
8720 rc = mdb_page_search_lowest(&mn);
8721 if (rc)
8722 return rc;
8723 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
8724 key.mv_size = mn.mc_db->md_pad;
8725 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
8726 } else {
8727 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
8728 key.mv_size = NODEKSZ(s2);
8729 key.mv_data = NODEKEY(s2);
8730 }
8731 } else {
8732 key.mv_size = srcnode->mn_ksize;
8733 key.mv_data = NODEKEY(srcnode);
8734 }
8735
8736 data.mv_size = NODEDSZ(srcnode);
8737 data.mv_data = NODEDATA(srcnode);
8738 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
8739 if (rc != MDB_SUCCESS)
8740 return rc;
8741 }
8742 }
8743
8744 DPRINTF(("dst page %"Yu" now has %u keys (%.1f%% filled)",
8745 pdst->mp_pgno, NUMKEYS(pdst),
8746 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
8747
8748 /* Unlink the src page from parent and add to free list.
8749 */
8750 csrc->mc_top--;
8751 mdb_node_del(csrc, 0);
8752 if (csrc->mc_ki[csrc->mc_top] == 0) {
8753 key.mv_size = 0;
8754 rc = mdb_update_key(csrc, &key);
8755 if (rc) {
8756 csrc->mc_top++;
8757 return rc;
8758 }
8759 }
8760 csrc->mc_top++;
8761
8762 psrc = csrc->mc_pg[csrc->mc_top];
8763 /* If not operating on FreeDB, allow this page to be reused
8764 * in this txn. Otherwise just add to free list.
8765 */
8766 rc = mdb_page_loose(csrc, psrc);
8767 if (rc)
8768 return rc;
8769 if (IS_LEAF(psrc))
8770 csrc->mc_db->md_leaf_pages--;
8771 else
8772 csrc->mc_db->md_branch_pages--;
8773 {
8774 /* Adjust other cursors pointing to mp */
8775 MDB_cursor *m2, *m3;
8776 MDB_dbi dbi = csrc->mc_dbi;
8777 unsigned int top = csrc->mc_top;
8778
8779 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8780 if (csrc->mc_flags & C_SUB)
8781 m3 = &m2->mc_xcursor->mx_cursor;
8782 else
8783 m3 = m2;
8784 if (m3 == csrc) continue;
8785 if (m3->mc_snum < csrc->mc_snum) continue;
8786 if (m3->mc_pg[top] == psrc) {
8787 m3->mc_pg[top] = pdst;
8788 m3->mc_ki[top] += nkeys;
8789 m3->mc_ki[top-1] = cdst->mc_ki[top-1];
8790 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] &&
8791 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) {
8792 m3->mc_ki[top-1]--;
8793 }
8794 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
8795 IS_LEAF(psrc)) {
8796 MDB_node *node = NODEPTR(m3->mc_pg[top], m3->mc_ki[top]);
8797 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
8798 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
8799 }
8800 }
8801 }
8802 {
8803 unsigned int snum = cdst->mc_snum;
8804 uint16_t depth = cdst->mc_db->md_depth;
8805 mdb_cursor_pop(cdst);
8806 rc = mdb_rebalance(cdst);
8807 /* Did the tree height change? */
8808 if (depth != cdst->mc_db->md_depth)
8809 snum += cdst->mc_db->md_depth - depth;
8810 cdst->mc_snum = snum;
8811 cdst->mc_top = snum-1;
8812 }
8813 return rc;
8814 }
8815
8816 /** Copy the contents of a cursor.
8817 * @param[in] csrc The cursor to copy from.
8818 * @param[out] cdst The cursor to copy to.
8819 */
8820 static void
mdb_cursor_copy(const MDB_cursor * csrc,MDB_cursor * cdst)8821 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
8822 {
8823 unsigned int i;
8824
8825 cdst->mc_txn = csrc->mc_txn;
8826 cdst->mc_dbi = csrc->mc_dbi;
8827 cdst->mc_db = csrc->mc_db;
8828 cdst->mc_dbx = csrc->mc_dbx;
8829 cdst->mc_snum = csrc->mc_snum;
8830 cdst->mc_top = csrc->mc_top;
8831 cdst->mc_flags = csrc->mc_flags;
8832 MC_SET_OVPG(cdst, MC_OVPG(csrc));
8833
8834 for (i=0; i<csrc->mc_snum; i++) {
8835 cdst->mc_pg[i] = csrc->mc_pg[i];
8836 cdst->mc_ki[i] = csrc->mc_ki[i];
8837 }
8838 }
8839
8840 /** Rebalance the tree after a delete operation.
8841 * @param[in] mc Cursor pointing to the page where rebalancing
8842 * should begin.
8843 * @return 0 on success, non-zero on failure.
8844 */
8845 static int
mdb_rebalance(MDB_cursor * mc)8846 mdb_rebalance(MDB_cursor *mc)
8847 {
8848 MDB_node *node;
8849 int rc, fromleft;
8850 unsigned int ptop, minkeys, thresh;
8851 MDB_cursor mn;
8852 indx_t oldki;
8853
8854 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
8855 minkeys = 2;
8856 thresh = 1;
8857 } else {
8858 minkeys = 1;
8859 thresh = FILL_THRESHOLD;
8860 }
8861 DPRINTF(("rebalancing %s page %"Yu" (has %u keys, %.1f%% full)",
8862 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
8863 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
8864 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
8865
8866 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
8867 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
8868 DPRINTF(("no need to rebalance page %"Yu", above fill threshold",
8869 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
8870 return MDB_SUCCESS;
8871 }
8872
8873 if (mc->mc_snum < 2) {
8874 MDB_page *mp = mc->mc_pg[0];
8875 if (IS_SUBP(mp)) {
8876 DPUTS("Can't rebalance a subpage, ignoring");
8877 return MDB_SUCCESS;
8878 }
8879 if (NUMKEYS(mp) == 0) {
8880 DPUTS("tree is completely empty");
8881 mc->mc_db->md_root = P_INVALID;
8882 mc->mc_db->md_depth = 0;
8883 mc->mc_db->md_leaf_pages = 0;
8884 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8885 if (rc)
8886 return rc;
8887 /* Adjust cursors pointing to mp */
8888 mc->mc_snum = 0;
8889 mc->mc_top = 0;
8890 mc->mc_flags &= ~C_INITIALIZED;
8891 {
8892 MDB_cursor *m2, *m3;
8893 MDB_dbi dbi = mc->mc_dbi;
8894
8895 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8896 if (mc->mc_flags & C_SUB)
8897 m3 = &m2->mc_xcursor->mx_cursor;
8898 else
8899 m3 = m2;
8900 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum))
8901 continue;
8902 if (m3->mc_pg[0] == mp) {
8903 m3->mc_snum = 0;
8904 m3->mc_top = 0;
8905 m3->mc_flags &= ~C_INITIALIZED;
8906 }
8907 }
8908 }
8909 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
8910 int i;
8911 DPUTS("collapsing root page!");
8912 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
8913 if (rc)
8914 return rc;
8915 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
8916 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL);
8917 if (rc)
8918 return rc;
8919 mc->mc_db->md_depth--;
8920 mc->mc_db->md_branch_pages--;
8921 mc->mc_ki[0] = mc->mc_ki[1];
8922 for (i = 1; i<mc->mc_db->md_depth; i++) {
8923 mc->mc_pg[i] = mc->mc_pg[i+1];
8924 mc->mc_ki[i] = mc->mc_ki[i+1];
8925 }
8926 {
8927 /* Adjust other cursors pointing to mp */
8928 MDB_cursor *m2, *m3;
8929 MDB_dbi dbi = mc->mc_dbi;
8930
8931 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8932 if (mc->mc_flags & C_SUB)
8933 m3 = &m2->mc_xcursor->mx_cursor;
8934 else
8935 m3 = m2;
8936 if (m3 == mc) continue;
8937 if (!(m3->mc_flags & C_INITIALIZED))
8938 continue;
8939 if (m3->mc_pg[0] == mp) {
8940 for (i=0; i<mc->mc_db->md_depth; i++) {
8941 m3->mc_pg[i] = m3->mc_pg[i+1];
8942 m3->mc_ki[i] = m3->mc_ki[i+1];
8943 }
8944 m3->mc_snum--;
8945 m3->mc_top--;
8946 }
8947 }
8948 }
8949 } else
8950 DPUTS("root page doesn't need rebalancing");
8951 return MDB_SUCCESS;
8952 }
8953
8954 /* The parent (branch page) must have at least 2 pointers,
8955 * otherwise the tree is invalid.
8956 */
8957 ptop = mc->mc_top-1;
8958 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
8959
8960 /* Leaf page fill factor is below the threshold.
8961 * Try to move keys from left or right neighbor, or
8962 * merge with a neighbor page.
8963 */
8964
8965 /* Find neighbors.
8966 */
8967 mdb_cursor_copy(mc, &mn);
8968 mn.mc_xcursor = NULL;
8969
8970 oldki = mc->mc_ki[mc->mc_top];
8971 if (mc->mc_ki[ptop] == 0) {
8972 /* We're the leftmost leaf in our parent.
8973 */
8974 DPUTS("reading right neighbor");
8975 mn.mc_ki[ptop]++;
8976 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8977 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8978 if (rc)
8979 return rc;
8980 mn.mc_ki[mn.mc_top] = 0;
8981 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
8982 fromleft = 0;
8983 } else {
8984 /* There is at least one neighbor to the left.
8985 */
8986 DPUTS("reading left neighbor");
8987 mn.mc_ki[ptop]--;
8988 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
8989 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL);
8990 if (rc)
8991 return rc;
8992 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
8993 mc->mc_ki[mc->mc_top] = 0;
8994 fromleft = 1;
8995 }
8996
8997 DPRINTF(("found neighbor page %"Yu" (%u keys, %.1f%% full)",
8998 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
8999 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
9000
9001 /* If the neighbor page is above threshold and has enough keys,
9002 * move one key from it. Otherwise we should try to merge them.
9003 * (A branch page must never have less than 2 keys.)
9004 */
9005 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
9006 rc = mdb_node_move(&mn, mc, fromleft);
9007 if (fromleft) {
9008 /* if we inserted on left, bump position up */
9009 oldki++;
9010 }
9011 } else {
9012 if (!fromleft) {
9013 rc = mdb_page_merge(&mn, mc);
9014 } else {
9015 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
9016 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
9017 /* We want mdb_rebalance to find mn when doing fixups */
9018 WITH_CURSOR_TRACKING(mn,
9019 rc = mdb_page_merge(mc, &mn));
9020 mdb_cursor_copy(&mn, mc);
9021 }
9022 mc->mc_flags &= ~C_EOF;
9023 }
9024 mc->mc_ki[mc->mc_top] = oldki;
9025 return rc;
9026 }
9027
9028 /** Complete a delete operation started by #mdb_cursor_del(). */
9029 static int
mdb_cursor_del0(MDB_cursor * mc)9030 mdb_cursor_del0(MDB_cursor *mc)
9031 {
9032 int rc;
9033 MDB_page *mp;
9034 indx_t ki;
9035 unsigned int nkeys;
9036 MDB_cursor *m2, *m3;
9037 MDB_dbi dbi = mc->mc_dbi;
9038
9039 ki = mc->mc_ki[mc->mc_top];
9040 mp = mc->mc_pg[mc->mc_top];
9041 mdb_node_del(mc, mc->mc_db->md_pad);
9042 mc->mc_db->md_entries--;
9043 {
9044 /* Adjust other cursors pointing to mp */
9045 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9046 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9047 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9048 continue;
9049 if (m3 == mc || m3->mc_snum < mc->mc_snum)
9050 continue;
9051 if (m3->mc_pg[mc->mc_top] == mp) {
9052 if (m3->mc_ki[mc->mc_top] == ki) {
9053 m3->mc_flags |= C_DEL;
9054 } else if (m3->mc_ki[mc->mc_top] > ki) {
9055 m3->mc_ki[mc->mc_top]--;
9056 }
9057 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
9058 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
9059 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
9060 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
9061 }
9062 }
9063 }
9064 }
9065 rc = mdb_rebalance(mc);
9066
9067 if (rc == MDB_SUCCESS) {
9068 /* DB is totally empty now, just bail out.
9069 * Other cursors adjustments were already done
9070 * by mdb_rebalance and aren't needed here.
9071 */
9072 if (!mc->mc_snum)
9073 return rc;
9074
9075 mp = mc->mc_pg[mc->mc_top];
9076 nkeys = NUMKEYS(mp);
9077
9078 /* Adjust other cursors pointing to mp */
9079 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
9080 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
9081 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9082 continue;
9083 if (m3->mc_snum < mc->mc_snum)
9084 continue;
9085 if (m3->mc_pg[mc->mc_top] == mp) {
9086 /* if m3 points past last node in page, find next sibling */
9087 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) {
9088 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9089 rc = mdb_cursor_sibling(m3, 1);
9090 if (rc == MDB_NOTFOUND) {
9091 m3->mc_flags |= C_EOF;
9092 rc = MDB_SUCCESS;
9093 continue;
9094 }
9095 }
9096 if (mc->mc_db->md_flags & MDB_DUPSORT) {
9097 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]);
9098 if (node->mn_flags & F_DUPDATA) {
9099 mdb_xcursor_init1(m3, node);
9100 m3->mc_xcursor->mx_cursor.mc_flags |= C_DEL;
9101 }
9102 }
9103 }
9104 }
9105 }
9106 mc->mc_flags |= C_DEL;
9107 }
9108
9109 if (rc)
9110 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9111 return rc;
9112 }
9113
9114 int
mdb_del(MDB_txn * txn,MDB_dbi dbi,MDB_val * key,MDB_val * data)9115 mdb_del(MDB_txn *txn, MDB_dbi dbi,
9116 MDB_val *key, MDB_val *data)
9117 {
9118 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9119 return EINVAL;
9120
9121 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9122 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9123
9124 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
9125 /* must ignore any data */
9126 data = NULL;
9127 }
9128
9129 return mdb_del0(txn, dbi, key, data, 0);
9130 }
9131
9132 static int
mdb_del0(MDB_txn * txn,MDB_dbi dbi,MDB_val * key,MDB_val * data,unsigned flags)9133 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
9134 MDB_val *key, MDB_val *data, unsigned flags)
9135 {
9136 MDB_cursor mc;
9137 MDB_xcursor mx;
9138 MDB_cursor_op op;
9139 MDB_val rdata, *xdata;
9140 int rc, exact = 0;
9141 DKBUF;
9142
9143 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
9144
9145 mdb_cursor_init(&mc, txn, dbi, &mx);
9146
9147 if (data) {
9148 op = MDB_GET_BOTH;
9149 rdata = *data;
9150 xdata = &rdata;
9151 } else {
9152 op = MDB_SET;
9153 xdata = NULL;
9154 flags |= MDB_NODUPDATA;
9155 }
9156 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
9157 if (rc == 0) {
9158 /* let mdb_page_split know about this cursor if needed:
9159 * delete will trigger a rebalance; if it needs to move
9160 * a node from one page to another, it will have to
9161 * update the parent's separator key(s). If the new sepkey
9162 * is larger than the current one, the parent page may
9163 * run out of space, triggering a split. We need this
9164 * cursor to be consistent until the end of the rebalance.
9165 */
9166 mc.mc_flags |= C_UNTRACK;
9167 mc.mc_next = txn->mt_cursors[dbi];
9168 txn->mt_cursors[dbi] = &mc;
9169 rc = mdb_cursor_del(&mc, flags);
9170 txn->mt_cursors[dbi] = mc.mc_next;
9171 }
9172 return rc;
9173 }
9174
9175 /** Split a page and insert a new node.
9176 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
9177 * The cursor will be updated to point to the actual page and index where
9178 * the node got inserted after the split.
9179 * @param[in] newkey The key for the newly inserted node.
9180 * @param[in] newdata The data for the newly inserted node.
9181 * @param[in] newpgno The page number, if the new node is a branch node.
9182 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
9183 * @return 0 on success, non-zero on failure.
9184 */
9185 static int
mdb_page_split(MDB_cursor * mc,MDB_val * newkey,MDB_val * newdata,pgno_t newpgno,unsigned int nflags)9186 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
9187 unsigned int nflags)
9188 {
9189 unsigned int flags;
9190 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
9191 indx_t newindx;
9192 pgno_t pgno = 0;
9193 int i, j, split_indx, nkeys, pmax;
9194 MDB_env *env = mc->mc_txn->mt_env;
9195 MDB_node *node;
9196 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
9197 MDB_page *copy = NULL;
9198 MDB_page *mp, *rp, *pp;
9199 int ptop;
9200 MDB_cursor mn;
9201 DKBUF;
9202
9203 mp = mc->mc_pg[mc->mc_top];
9204 newindx = mc->mc_ki[mc->mc_top];
9205 nkeys = NUMKEYS(mp);
9206
9207 DPRINTF(("-----> splitting %s page %"Yu" and adding [%s] at index %i/%i",
9208 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
9209 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
9210
9211 /* Create a right sibling. */
9212 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
9213 return rc;
9214 rp->mp_pad = mp->mp_pad;
9215 DPRINTF(("new right sibling: page %"Yu, rp->mp_pgno));
9216
9217 /* Usually when splitting the root page, the cursor
9218 * height is 1. But when called from mdb_update_key,
9219 * the cursor height may be greater because it walks
9220 * up the stack while finding the branch slot to update.
9221 */
9222 if (mc->mc_top < 1) {
9223 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
9224 goto done;
9225 /* shift current top to make room for new parent */
9226 for (i=mc->mc_snum; i>0; i--) {
9227 mc->mc_pg[i] = mc->mc_pg[i-1];
9228 mc->mc_ki[i] = mc->mc_ki[i-1];
9229 }
9230 mc->mc_pg[0] = pp;
9231 mc->mc_ki[0] = 0;
9232 mc->mc_db->md_root = pp->mp_pgno;
9233 DPRINTF(("root split! new root = %"Yu, pp->mp_pgno));
9234 new_root = mc->mc_db->md_depth++;
9235
9236 /* Add left (implicit) pointer. */
9237 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
9238 /* undo the pre-push */
9239 mc->mc_pg[0] = mc->mc_pg[1];
9240 mc->mc_ki[0] = mc->mc_ki[1];
9241 mc->mc_db->md_root = mp->mp_pgno;
9242 mc->mc_db->md_depth--;
9243 goto done;
9244 }
9245 mc->mc_snum++;
9246 mc->mc_top++;
9247 ptop = 0;
9248 } else {
9249 ptop = mc->mc_top-1;
9250 DPRINTF(("parent branch page is %"Yu, mc->mc_pg[ptop]->mp_pgno));
9251 }
9252
9253 mdb_cursor_copy(mc, &mn);
9254 mn.mc_xcursor = NULL;
9255 mn.mc_pg[mn.mc_top] = rp;
9256 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
9257
9258 if (nflags & MDB_APPEND) {
9259 mn.mc_ki[mn.mc_top] = 0;
9260 sepkey = *newkey;
9261 split_indx = newindx;
9262 nkeys = 0;
9263 } else {
9264
9265 split_indx = (nkeys+1) / 2;
9266
9267 if (IS_LEAF2(rp)) {
9268 char *split, *ins;
9269 int x;
9270 unsigned int lsize, rsize, ksize;
9271 /* Move half of the keys to the right sibling */
9272 x = mc->mc_ki[mc->mc_top] - split_indx;
9273 ksize = mc->mc_db->md_pad;
9274 split = LEAF2KEY(mp, split_indx, ksize);
9275 rsize = (nkeys - split_indx) * ksize;
9276 lsize = (nkeys - split_indx) * sizeof(indx_t);
9277 mp->mp_lower -= lsize;
9278 rp->mp_lower += lsize;
9279 mp->mp_upper += rsize - lsize;
9280 rp->mp_upper -= rsize - lsize;
9281 sepkey.mv_size = ksize;
9282 if (newindx == split_indx) {
9283 sepkey.mv_data = newkey->mv_data;
9284 } else {
9285 sepkey.mv_data = split;
9286 }
9287 if (x<0) {
9288 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
9289 memcpy(rp->mp_ptrs, split, rsize);
9290 sepkey.mv_data = rp->mp_ptrs;
9291 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
9292 memcpy(ins, newkey->mv_data, ksize);
9293 mp->mp_lower += sizeof(indx_t);
9294 mp->mp_upper -= ksize - sizeof(indx_t);
9295 } else {
9296 if (x)
9297 memcpy(rp->mp_ptrs, split, x * ksize);
9298 ins = LEAF2KEY(rp, x, ksize);
9299 memcpy(ins, newkey->mv_data, ksize);
9300 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
9301 rp->mp_lower += sizeof(indx_t);
9302 rp->mp_upper -= ksize - sizeof(indx_t);
9303 mc->mc_ki[mc->mc_top] = x;
9304 }
9305 } else {
9306 int psize, nsize, k;
9307 /* Maximum free space in an empty page */
9308 pmax = env->me_psize - PAGEHDRSZ;
9309 if (IS_LEAF(mp))
9310 nsize = mdb_leaf_size(env, newkey, newdata);
9311 else
9312 nsize = mdb_branch_size(env, newkey);
9313 nsize = EVEN(nsize);
9314
9315 /* grab a page to hold a temporary copy */
9316 copy = mdb_page_malloc(mc->mc_txn, 1);
9317 if (copy == NULL) {
9318 rc = ENOMEM;
9319 goto done;
9320 }
9321 copy->mp_pgno = mp->mp_pgno;
9322 copy->mp_flags = mp->mp_flags;
9323 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
9324 copy->mp_upper = env->me_psize - PAGEBASE;
9325
9326 /* prepare to insert */
9327 for (i=0, j=0; i<nkeys; i++) {
9328 if (i == newindx) {
9329 copy->mp_ptrs[j++] = 0;
9330 }
9331 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
9332 }
9333
9334 /* When items are relatively large the split point needs
9335 * to be checked, because being off-by-one will make the
9336 * difference between success or failure in mdb_node_add.
9337 *
9338 * It's also relevant if a page happens to be laid out
9339 * such that one half of its nodes are all "small" and
9340 * the other half of its nodes are "large." If the new
9341 * item is also "large" and falls on the half with
9342 * "large" nodes, it also may not fit.
9343 *
9344 * As a final tweak, if the new item goes on the last
9345 * spot on the page (and thus, onto the new page), bias
9346 * the split so the new page is emptier than the old page.
9347 * This yields better packing during sequential inserts.
9348 */
9349 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
9350 /* Find split point */
9351 psize = 0;
9352 if (newindx <= split_indx || newindx >= nkeys) {
9353 i = 0; j = 1;
9354 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
9355 } else {
9356 i = nkeys; j = -1;
9357 k = split_indx-1;
9358 }
9359 for (; i!=k; i+=j) {
9360 if (i == newindx) {
9361 psize += nsize;
9362 node = NULL;
9363 } else {
9364 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9365 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
9366 if (IS_LEAF(mp)) {
9367 if (F_ISSET(node->mn_flags, F_BIGDATA))
9368 psize += sizeof(pgno_t);
9369 else
9370 psize += NODEDSZ(node);
9371 }
9372 psize = EVEN(psize);
9373 }
9374 if (psize > pmax || i == k-j) {
9375 split_indx = i + (j<0);
9376 break;
9377 }
9378 }
9379 }
9380 if (split_indx == newindx) {
9381 sepkey.mv_size = newkey->mv_size;
9382 sepkey.mv_data = newkey->mv_data;
9383 } else {
9384 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
9385 sepkey.mv_size = node->mn_ksize;
9386 sepkey.mv_data = NODEKEY(node);
9387 }
9388 }
9389 }
9390
9391 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
9392
9393 /* Copy separator key to the parent.
9394 */
9395 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
9396 int snum = mc->mc_snum;
9397 mn.mc_snum--;
9398 mn.mc_top--;
9399 did_split = 1;
9400 /* We want other splits to find mn when doing fixups */
9401 WITH_CURSOR_TRACKING(mn,
9402 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0));
9403 if (rc)
9404 goto done;
9405
9406 /* root split? */
9407 if (mc->mc_snum > snum) {
9408 ptop++;
9409 }
9410 /* Right page might now have changed parent.
9411 * Check if left page also changed parent.
9412 */
9413 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9414 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9415 for (i=0; i<ptop; i++) {
9416 mc->mc_pg[i] = mn.mc_pg[i];
9417 mc->mc_ki[i] = mn.mc_ki[i];
9418 }
9419 mc->mc_pg[ptop] = mn.mc_pg[ptop];
9420 if (mn.mc_ki[ptop]) {
9421 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
9422 } else {
9423 /* find right page's left sibling */
9424 mc->mc_ki[ptop] = mn.mc_ki[ptop];
9425 mdb_cursor_sibling(mc, 0);
9426 }
9427 }
9428 } else {
9429 mn.mc_top--;
9430 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
9431 mn.mc_top++;
9432 }
9433 if (rc != MDB_SUCCESS) {
9434 goto done;
9435 }
9436 if (nflags & MDB_APPEND) {
9437 mc->mc_pg[mc->mc_top] = rp;
9438 mc->mc_ki[mc->mc_top] = 0;
9439 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
9440 if (rc)
9441 goto done;
9442 for (i=0; i<mc->mc_top; i++)
9443 mc->mc_ki[i] = mn.mc_ki[i];
9444 } else if (!IS_LEAF2(mp)) {
9445 /* Move nodes */
9446 mc->mc_pg[mc->mc_top] = rp;
9447 i = split_indx;
9448 j = 0;
9449 do {
9450 if (i == newindx) {
9451 rkey.mv_data = newkey->mv_data;
9452 rkey.mv_size = newkey->mv_size;
9453 if (IS_LEAF(mp)) {
9454 rdata = newdata;
9455 } else
9456 pgno = newpgno;
9457 flags = nflags;
9458 /* Update index for the new key. */
9459 mc->mc_ki[mc->mc_top] = j;
9460 } else {
9461 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
9462 rkey.mv_data = NODEKEY(node);
9463 rkey.mv_size = node->mn_ksize;
9464 if (IS_LEAF(mp)) {
9465 xdata.mv_data = NODEDATA(node);
9466 xdata.mv_size = NODEDSZ(node);
9467 rdata = &xdata;
9468 } else
9469 pgno = NODEPGNO(node);
9470 flags = node->mn_flags;
9471 }
9472
9473 if (!IS_LEAF(mp) && j == 0) {
9474 /* First branch index doesn't need key data. */
9475 rkey.mv_size = 0;
9476 }
9477
9478 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
9479 if (rc)
9480 goto done;
9481 if (i == nkeys) {
9482 i = 0;
9483 j = 0;
9484 mc->mc_pg[mc->mc_top] = copy;
9485 } else {
9486 i++;
9487 j++;
9488 }
9489 } while (i != split_indx);
9490
9491 nkeys = NUMKEYS(copy);
9492 for (i=0; i<nkeys; i++)
9493 mp->mp_ptrs[i] = copy->mp_ptrs[i];
9494 mp->mp_lower = copy->mp_lower;
9495 mp->mp_upper = copy->mp_upper;
9496 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
9497 env->me_psize - copy->mp_upper - PAGEBASE);
9498
9499 /* reset back to original page */
9500 if (newindx < split_indx) {
9501 mc->mc_pg[mc->mc_top] = mp;
9502 } else {
9503 mc->mc_pg[mc->mc_top] = rp;
9504 mc->mc_ki[ptop]++;
9505 /* Make sure mc_ki is still valid.
9506 */
9507 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9508 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9509 for (i=0; i<=ptop; i++) {
9510 mc->mc_pg[i] = mn.mc_pg[i];
9511 mc->mc_ki[i] = mn.mc_ki[i];
9512 }
9513 }
9514 }
9515 if (nflags & MDB_RESERVE) {
9516 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
9517 if (!(node->mn_flags & F_BIGDATA))
9518 newdata->mv_data = NODEDATA(node);
9519 }
9520 } else {
9521 if (newindx >= split_indx) {
9522 mc->mc_pg[mc->mc_top] = rp;
9523 mc->mc_ki[ptop]++;
9524 /* Make sure mc_ki is still valid.
9525 */
9526 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
9527 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
9528 for (i=0; i<=ptop; i++) {
9529 mc->mc_pg[i] = mn.mc_pg[i];
9530 mc->mc_ki[i] = mn.mc_ki[i];
9531 }
9532 }
9533 }
9534 }
9535
9536 {
9537 /* Adjust other cursors pointing to mp */
9538 MDB_cursor *m2, *m3;
9539 MDB_dbi dbi = mc->mc_dbi;
9540 nkeys = NUMKEYS(mp);
9541
9542 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
9543 if (mc->mc_flags & C_SUB)
9544 m3 = &m2->mc_xcursor->mx_cursor;
9545 else
9546 m3 = m2;
9547 if (m3 == mc)
9548 continue;
9549 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
9550 continue;
9551 if (new_root) {
9552 int k;
9553 /* sub cursors may be on different DB */
9554 if (m3->mc_pg[0] != mp)
9555 continue;
9556 /* root split */
9557 for (k=new_root; k>=0; k--) {
9558 m3->mc_ki[k+1] = m3->mc_ki[k];
9559 m3->mc_pg[k+1] = m3->mc_pg[k];
9560 }
9561 if (m3->mc_ki[0] >= nkeys) {
9562 m3->mc_ki[0] = 1;
9563 } else {
9564 m3->mc_ki[0] = 0;
9565 }
9566 m3->mc_pg[0] = mc->mc_pg[0];
9567 m3->mc_snum++;
9568 m3->mc_top++;
9569 }
9570 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
9571 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
9572 m3->mc_ki[mc->mc_top]++;
9573 if (m3->mc_ki[mc->mc_top] >= nkeys) {
9574 m3->mc_pg[mc->mc_top] = rp;
9575 m3->mc_ki[mc->mc_top] -= nkeys;
9576 for (i=0; i<mc->mc_top; i++) {
9577 m3->mc_ki[i] = mn.mc_ki[i];
9578 m3->mc_pg[i] = mn.mc_pg[i];
9579 }
9580 }
9581 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
9582 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
9583 m3->mc_ki[ptop]++;
9584 }
9585 if (m3->mc_xcursor && (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) &&
9586 IS_LEAF(mp)) {
9587 MDB_node *node = NODEPTR(m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]);
9588 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA)
9589 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node);
9590 }
9591 }
9592 }
9593 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
9594
9595 done:
9596 if (copy) /* tmp page */
9597 mdb_page_free(env, copy);
9598 if (rc)
9599 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
9600 return rc;
9601 }
9602
9603 int
mdb_put(MDB_txn * txn,MDB_dbi dbi,MDB_val * key,MDB_val * data,unsigned int flags)9604 mdb_put(MDB_txn *txn, MDB_dbi dbi,
9605 MDB_val *key, MDB_val *data, unsigned int flags)
9606 {
9607 MDB_cursor mc;
9608 MDB_xcursor mx;
9609 int rc;
9610
9611 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
9612 return EINVAL;
9613
9614 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP))
9615 return EINVAL;
9616
9617 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED))
9618 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
9619
9620 mdb_cursor_init(&mc, txn, dbi, &mx);
9621 mc.mc_next = txn->mt_cursors[dbi];
9622 txn->mt_cursors[dbi] = &mc;
9623 rc = mdb_cursor_put(&mc, key, data, flags);
9624 txn->mt_cursors[dbi] = mc.mc_next;
9625 return rc;
9626 }
9627
9628 #ifndef MDB_WBUF
9629 #define MDB_WBUF (1024*1024)
9630 #endif
9631 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */
9632
9633 /** State needed for a double-buffering compacting copy. */
9634 typedef struct mdb_copy {
9635 pthread_mutex_t mc_mutex;
9636 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */
9637 char *mc_wbuf[2];
9638 char *mc_over[2];
9639 MDB_env *mc_env;
9640 MDB_txn *mc_txn;
9641 int mc_wlen[2];
9642 int mc_olen[2];
9643 pgno_t mc_next_pgno;
9644 HANDLE mc_fd;
9645 int mc_toggle; /**< Buffer number in provider */
9646 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */
9647 volatile int mc_error; /**< Error code, never cleared if set */
9648 } mdb_copy;
9649
9650 /** Dedicated writer thread for compacting copy. */
9651 static THREAD_RET ESECT CALL_CONV
mdb_env_copythr(void * arg)9652 mdb_env_copythr(void *arg)
9653 {
9654 mdb_copy *my = arg;
9655 char *ptr;
9656 int toggle = 0, wsize, rc;
9657 #ifdef _WIN32
9658 DWORD len;
9659 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
9660 #else
9661 int len;
9662 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
9663 #endif
9664
9665 pthread_mutex_lock(&my->mc_mutex);
9666 for(;;) {
9667 while (!my->mc_new)
9668 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9669 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */
9670 break;
9671 wsize = my->mc_wlen[toggle];
9672 ptr = my->mc_wbuf[toggle];
9673 again:
9674 rc = MDB_SUCCESS;
9675 while (wsize > 0 && !my->mc_error) {
9676 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
9677 if (!rc) {
9678 rc = ErrCode();
9679 break;
9680 } else if (len > 0) {
9681 rc = MDB_SUCCESS;
9682 ptr += len;
9683 wsize -= len;
9684 continue;
9685 } else {
9686 rc = EIO;
9687 break;
9688 }
9689 }
9690 if (rc) {
9691 my->mc_error = rc;
9692 }
9693 /* If there's an overflow page tail, write it too */
9694 if (my->mc_olen[toggle]) {
9695 wsize = my->mc_olen[toggle];
9696 ptr = my->mc_over[toggle];
9697 my->mc_olen[toggle] = 0;
9698 goto again;
9699 }
9700 my->mc_wlen[toggle] = 0;
9701 toggle ^= 1;
9702 /* Return the empty buffer to provider */
9703 my->mc_new--;
9704 pthread_cond_signal(&my->mc_cond);
9705 }
9706 pthread_mutex_unlock(&my->mc_mutex);
9707 return (THREAD_RET)0;
9708 #undef DO_WRITE
9709 }
9710
9711 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer.
9712 *
9713 * @param[in] my control structure.
9714 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending).
9715 */
9716 static int ESECT
mdb_env_cthr_toggle(mdb_copy * my,int adjust)9717 mdb_env_cthr_toggle(mdb_copy *my, int adjust)
9718 {
9719 pthread_mutex_lock(&my->mc_mutex);
9720 my->mc_new += adjust;
9721 pthread_cond_signal(&my->mc_cond);
9722 while (my->mc_new & 2) /* both buffers in use */
9723 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
9724 pthread_mutex_unlock(&my->mc_mutex);
9725
9726 my->mc_toggle ^= (adjust & 1);
9727 /* Both threads reset mc_wlen, to be safe from threading errors */
9728 my->mc_wlen[my->mc_toggle] = 0;
9729 return my->mc_error;
9730 }
9731
9732 /** Depth-first tree traversal for compacting copy. */
9733 static int ESECT
mdb_env_cwalk(mdb_copy * my,pgno_t * pg,int flags)9734 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
9735 {
9736 MDB_cursor mc = {0};
9737 MDB_node *ni;
9738 MDB_page *mo, *mp, *leaf;
9739 char *buf, *ptr;
9740 int rc, toggle;
9741 unsigned int i;
9742
9743 /* Empty DB, nothing to do */
9744 if (*pg == P_INVALID)
9745 return MDB_SUCCESS;
9746
9747 mc.mc_snum = 1;
9748 mc.mc_txn = my->mc_txn;
9749 mc.mc_flags = my->mc_txn->mt_flags & (C_ORIG_RDONLY|C_WRITEMAP);
9750
9751 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL);
9752 if (rc)
9753 return rc;
9754 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
9755 if (rc)
9756 return rc;
9757
9758 /* Make cursor pages writable */
9759 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
9760 if (buf == NULL)
9761 return ENOMEM;
9762
9763 for (i=0; i<mc.mc_top; i++) {
9764 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
9765 mc.mc_pg[i] = (MDB_page *)ptr;
9766 ptr += my->mc_env->me_psize;
9767 }
9768
9769 /* This is writable space for a leaf page. Usually not needed. */
9770 leaf = (MDB_page *)ptr;
9771
9772 toggle = my->mc_toggle;
9773 while (mc.mc_snum > 0) {
9774 unsigned n;
9775 mp = mc.mc_pg[mc.mc_top];
9776 n = NUMKEYS(mp);
9777
9778 if (IS_LEAF(mp)) {
9779 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
9780 for (i=0; i<n; i++) {
9781 ni = NODEPTR(mp, i);
9782 if (ni->mn_flags & F_BIGDATA) {
9783 MDB_page *omp;
9784 pgno_t pg;
9785
9786 /* Need writable leaf */
9787 if (mp != leaf) {
9788 mc.mc_pg[mc.mc_top] = leaf;
9789 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9790 mp = leaf;
9791 ni = NODEPTR(mp, i);
9792 }
9793
9794 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9795 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t));
9796 rc = mdb_page_get(&mc, pg, &omp, NULL);
9797 if (rc)
9798 goto done;
9799 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9800 rc = mdb_env_cthr_toggle(my, 1);
9801 if (rc)
9802 goto done;
9803 toggle = my->mc_toggle;
9804 }
9805 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9806 memcpy(mo, omp, my->mc_env->me_psize);
9807 mo->mp_pgno = my->mc_next_pgno;
9808 my->mc_next_pgno += omp->mp_pages;
9809 my->mc_wlen[toggle] += my->mc_env->me_psize;
9810 if (omp->mp_pages > 1) {
9811 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
9812 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
9813 rc = mdb_env_cthr_toggle(my, 1);
9814 if (rc)
9815 goto done;
9816 toggle = my->mc_toggle;
9817 }
9818 } else if (ni->mn_flags & F_SUBDATA) {
9819 MDB_db db;
9820
9821 /* Need writable leaf */
9822 if (mp != leaf) {
9823 mc.mc_pg[mc.mc_top] = leaf;
9824 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
9825 mp = leaf;
9826 ni = NODEPTR(mp, i);
9827 }
9828
9829 memcpy(&db, NODEDATA(ni), sizeof(db));
9830 my->mc_toggle = toggle;
9831 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
9832 if (rc)
9833 goto done;
9834 toggle = my->mc_toggle;
9835 memcpy(NODEDATA(ni), &db, sizeof(db));
9836 }
9837 }
9838 }
9839 } else {
9840 mc.mc_ki[mc.mc_top]++;
9841 if (mc.mc_ki[mc.mc_top] < n) {
9842 pgno_t pg;
9843 again:
9844 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
9845 pg = NODEPGNO(ni);
9846 rc = mdb_page_get(&mc, pg, &mp, NULL);
9847 if (rc)
9848 goto done;
9849 mc.mc_top++;
9850 mc.mc_snum++;
9851 mc.mc_ki[mc.mc_top] = 0;
9852 if (IS_BRANCH(mp)) {
9853 /* Whenever we advance to a sibling branch page,
9854 * we must proceed all the way down to its first leaf.
9855 */
9856 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
9857 goto again;
9858 } else
9859 mc.mc_pg[mc.mc_top] = mp;
9860 continue;
9861 }
9862 }
9863 if (my->mc_wlen[toggle] >= MDB_WBUF) {
9864 rc = mdb_env_cthr_toggle(my, 1);
9865 if (rc)
9866 goto done;
9867 toggle = my->mc_toggle;
9868 }
9869 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
9870 mdb_page_copy(mo, mp, my->mc_env->me_psize);
9871 mo->mp_pgno = my->mc_next_pgno++;
9872 my->mc_wlen[toggle] += my->mc_env->me_psize;
9873 if (mc.mc_top) {
9874 /* Update parent if there is one */
9875 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
9876 SETPGNO(ni, mo->mp_pgno);
9877 mdb_cursor_pop(&mc);
9878 } else {
9879 /* Otherwise we're done */
9880 *pg = mo->mp_pgno;
9881 break;
9882 }
9883 }
9884 done:
9885 free(buf);
9886 return rc;
9887 }
9888
9889 /** Copy environment with compaction. */
9890 static int ESECT
mdb_env_copyfd1(MDB_env * env,HANDLE fd)9891 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
9892 {
9893 MDB_meta *mm;
9894 MDB_page *mp;
9895 mdb_copy my = {0};
9896 MDB_txn *txn = NULL;
9897 pthread_t thr;
9898 pgno_t root, new_root;
9899 int rc = MDB_SUCCESS;
9900
9901 #ifdef _WIN32
9902 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) ||
9903 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) {
9904 rc = ErrCode();
9905 goto done;
9906 }
9907 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
9908 if (my.mc_wbuf[0] == NULL) {
9909 /* _aligned_malloc() sets errno, but we use Windows error codes */
9910 rc = ERROR_NOT_ENOUGH_MEMORY;
9911 goto done;
9912 }
9913 #else
9914 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0)
9915 return rc;
9916 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0)
9917 goto done2;
9918 #ifdef HAVE_MEMALIGN
9919 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
9920 if (my.mc_wbuf[0] == NULL) {
9921 rc = errno;
9922 goto done;
9923 }
9924 #else
9925 {
9926 void *p;
9927 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0)
9928 goto done;
9929 my.mc_wbuf[0] = p;
9930 }
9931 #endif
9932 #endif
9933 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
9934 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
9935 my.mc_next_pgno = NUM_METAS;
9936 my.mc_env = env;
9937 my.mc_fd = fd;
9938 rc = THREAD_CREATE(thr, mdb_env_copythr, &my);
9939 if (rc)
9940 goto done;
9941
9942 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
9943 if (rc)
9944 goto finish;
9945
9946 mp = (MDB_page *)my.mc_wbuf[0];
9947 memset(mp, 0, NUM_METAS * env->me_psize);
9948 mp->mp_pgno = 0;
9949 mp->mp_flags = P_META;
9950 mm = (MDB_meta *)METADATA(mp);
9951 mdb_env_init_meta0(env, mm);
9952 mm->mm_address = env->me_metas[0]->mm_address;
9953
9954 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
9955 mp->mp_pgno = 1;
9956 mp->mp_flags = P_META;
9957 *(MDB_meta *)METADATA(mp) = *mm;
9958 mm = (MDB_meta *)METADATA(mp);
9959
9960 /* Set metapage 1 with current main DB */
9961 root = new_root = txn->mt_dbs[MAIN_DBI].md_root;
9962 if (root != P_INVALID) {
9963 /* Count free pages + freeDB pages. Subtract from last_pg
9964 * to find the new last_pg, which also becomes the new root.
9965 */
9966 MDB_ID freecount = 0;
9967 MDB_cursor mc;
9968 MDB_val key, data;
9969 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
9970 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
9971 freecount += *(MDB_ID *)data.mv_data;
9972 if (rc != MDB_NOTFOUND)
9973 goto finish;
9974 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages +
9975 txn->mt_dbs[FREE_DBI].md_leaf_pages +
9976 txn->mt_dbs[FREE_DBI].md_overflow_pages;
9977
9978 new_root = txn->mt_next_pgno - 1 - freecount;
9979 mm->mm_last_pg = new_root;
9980 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI];
9981 mm->mm_dbs[MAIN_DBI].md_root = new_root;
9982 } else {
9983 /* When the DB is empty, handle it specially to
9984 * fix any breakage like page leaks from ITS#8174.
9985 */
9986 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags;
9987 }
9988 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) {
9989 mm->mm_txnid = 1; /* use metapage 1 */
9990 }
9991
9992 my.mc_wlen[0] = env->me_psize * NUM_METAS;
9993 my.mc_txn = txn;
9994 rc = mdb_env_cwalk(&my, &root, 0);
9995 if (rc == MDB_SUCCESS && root != new_root) {
9996 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */
9997 }
9998
9999 finish:
10000 if (rc)
10001 my.mc_error = rc;
10002 mdb_env_cthr_toggle(&my, 1 | MDB_EOF);
10003 rc = THREAD_FINISH(thr);
10004 mdb_txn_abort(txn);
10005
10006 done:
10007 #ifdef _WIN32
10008 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]);
10009 if (my.mc_cond) CloseHandle(my.mc_cond);
10010 if (my.mc_mutex) CloseHandle(my.mc_mutex);
10011 #else
10012 free(my.mc_wbuf[0]);
10013 pthread_cond_destroy(&my.mc_cond);
10014 done2:
10015 pthread_mutex_destroy(&my.mc_mutex);
10016 #endif
10017 return rc ? rc : my.mc_error;
10018 }
10019
10020 /** Copy environment as-is. */
10021 static int ESECT
mdb_env_copyfd0(MDB_env * env,HANDLE fd)10022 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
10023 {
10024 MDB_txn *txn = NULL;
10025 mdb_mutexref_t wmutex = NULL;
10026 int rc;
10027 mdb_size_t wsize, w3;
10028 char *ptr;
10029 #ifdef _WIN32
10030 DWORD len, w2;
10031 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
10032 #else
10033 ssize_t len;
10034 size_t w2;
10035 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
10036 #endif
10037
10038 /* Do the lock/unlock of the reader mutex before starting the
10039 * write txn. Otherwise other read txns could block writers.
10040 */
10041 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
10042 if (rc)
10043 return rc;
10044
10045 if (env->me_txns) {
10046 /* We must start the actual read txn after blocking writers */
10047 mdb_txn_end(txn, MDB_END_RESET_TMP);
10048
10049 /* Temporarily block writers until we snapshot the meta pages */
10050 wmutex = env->me_wmutex;
10051 if (LOCK_MUTEX(rc, env, wmutex))
10052 goto leave;
10053
10054 rc = mdb_txn_renew0(txn);
10055 if (rc) {
10056 UNLOCK_MUTEX(wmutex);
10057 goto leave;
10058 }
10059 }
10060
10061 wsize = env->me_psize * NUM_METAS;
10062 ptr = env->me_map;
10063 w2 = wsize;
10064 while (w2 > 0) {
10065 DO_WRITE(rc, fd, ptr, w2, len);
10066 if (!rc) {
10067 rc = ErrCode();
10068 break;
10069 } else if (len > 0) {
10070 rc = MDB_SUCCESS;
10071 ptr += len;
10072 w2 -= len;
10073 continue;
10074 } else {
10075 /* Non-blocking or async handles are not supported */
10076 rc = EIO;
10077 break;
10078 }
10079 }
10080 if (wmutex)
10081 UNLOCK_MUTEX(wmutex);
10082
10083 if (rc)
10084 goto leave;
10085
10086 w3 = txn->mt_next_pgno * env->me_psize;
10087 {
10088 mdb_size_t fsize = 0;
10089 if ((rc = mdb_fsize(env->me_fd, &fsize)))
10090 goto leave;
10091 if (w3 > fsize)
10092 w3 = fsize;
10093 }
10094 wsize = w3 - wsize;
10095 while (wsize > 0) {
10096 if (wsize > MAX_WRITE)
10097 w2 = MAX_WRITE;
10098 else
10099 w2 = wsize;
10100 DO_WRITE(rc, fd, ptr, w2, len);
10101 if (!rc) {
10102 rc = ErrCode();
10103 break;
10104 } else if (len > 0) {
10105 rc = MDB_SUCCESS;
10106 ptr += len;
10107 wsize -= len;
10108 continue;
10109 } else {
10110 rc = EIO;
10111 break;
10112 }
10113 }
10114
10115 leave:
10116 mdb_txn_abort(txn);
10117 return rc;
10118 }
10119
10120 int ESECT
mdb_env_copyfd2(MDB_env * env,HANDLE fd,unsigned int flags)10121 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
10122 {
10123 if (flags & MDB_CP_COMPACT)
10124 return mdb_env_copyfd1(env, fd);
10125 else
10126 return mdb_env_copyfd0(env, fd);
10127 }
10128
10129 int ESECT
mdb_env_copyfd(MDB_env * env,HANDLE fd)10130 mdb_env_copyfd(MDB_env *env, HANDLE fd)
10131 {
10132 return mdb_env_copyfd2(env, fd, 0);
10133 }
10134
10135 int ESECT
mdb_env_copy2(MDB_env * env,const char * path,unsigned int flags)10136 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
10137 {
10138 int rc, len;
10139 char *lpath;
10140 HANDLE newfd = INVALID_HANDLE_VALUE;
10141 #ifdef _WIN32
10142 wchar_t *wpath;
10143 #endif
10144
10145 if (env->me_flags & MDB_NOSUBDIR) {
10146 lpath = (char *)path;
10147 } else {
10148 len = strlen(path);
10149 len += sizeof(DATANAME);
10150 lpath = malloc(len);
10151 if (!lpath)
10152 return ENOMEM;
10153 sprintf(lpath, "%s" DATANAME, path);
10154 }
10155
10156 /* The destination path must exist, but the destination file must not.
10157 * We don't want the OS to cache the writes, since the source data is
10158 * already in the OS cache.
10159 */
10160 #ifdef _WIN32
10161 rc = utf8_to_utf16(lpath, -1, &wpath, NULL);
10162 if (rc)
10163 goto leave;
10164 newfd = CreateFileW(wpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
10165 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
10166 free(wpath);
10167 #else
10168 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
10169 #endif
10170 if (newfd == INVALID_HANDLE_VALUE) {
10171 rc = ErrCode();
10172 goto leave;
10173 }
10174
10175 if (env->me_psize >= env->me_os_psize) {
10176 #ifdef O_DIRECT
10177 /* Set O_DIRECT if the file system supports it */
10178 if ((rc = fcntl(newfd, F_GETFL)) != -1)
10179 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
10180 #endif
10181 #ifdef F_NOCACHE /* __APPLE__ */
10182 rc = fcntl(newfd, F_NOCACHE, 1);
10183 if (rc) {
10184 rc = ErrCode();
10185 goto leave;
10186 }
10187 #endif
10188 }
10189
10190 rc = mdb_env_copyfd2(env, newfd, flags);
10191
10192 leave:
10193 if (!(env->me_flags & MDB_NOSUBDIR))
10194 free(lpath);
10195 if (newfd != INVALID_HANDLE_VALUE)
10196 if (close(newfd) < 0 && rc == MDB_SUCCESS)
10197 rc = ErrCode();
10198
10199 return rc;
10200 }
10201
10202 int ESECT
mdb_env_copy(MDB_env * env,const char * path)10203 mdb_env_copy(MDB_env *env, const char *path)
10204 {
10205 return mdb_env_copy2(env, path, 0);
10206 }
10207
10208 int ESECT
mdb_env_set_flags(MDB_env * env,unsigned int flag,int onoff)10209 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
10210 {
10211 if (flag & ~CHANGEABLE)
10212 return EINVAL;
10213 if (onoff)
10214 env->me_flags |= flag;
10215 else
10216 env->me_flags &= ~flag;
10217 return MDB_SUCCESS;
10218 }
10219
10220 int ESECT
mdb_env_get_flags(MDB_env * env,unsigned int * arg)10221 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
10222 {
10223 if (!env || !arg)
10224 return EINVAL;
10225
10226 *arg = env->me_flags & (CHANGEABLE|CHANGELESS);
10227 return MDB_SUCCESS;
10228 }
10229
10230 int ESECT
mdb_env_set_userctx(MDB_env * env,void * ctx)10231 mdb_env_set_userctx(MDB_env *env, void *ctx)
10232 {
10233 if (!env)
10234 return EINVAL;
10235 env->me_userctx = ctx;
10236 return MDB_SUCCESS;
10237 }
10238
10239 void * ESECT
mdb_env_get_userctx(MDB_env * env)10240 mdb_env_get_userctx(MDB_env *env)
10241 {
10242 return env ? env->me_userctx : NULL;
10243 }
10244
10245 int ESECT
mdb_env_set_assert(MDB_env * env,MDB_assert_func * func)10246 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
10247 {
10248 if (!env)
10249 return EINVAL;
10250 #ifndef NDEBUG
10251 env->me_assert_func = func;
10252 #endif
10253 return MDB_SUCCESS;
10254 }
10255
10256 int ESECT
mdb_env_get_path(MDB_env * env,const char ** arg)10257 mdb_env_get_path(MDB_env *env, const char **arg)
10258 {
10259 if (!env || !arg)
10260 return EINVAL;
10261
10262 *arg = env->me_path;
10263 return MDB_SUCCESS;
10264 }
10265
10266 int ESECT
mdb_env_get_fd(MDB_env * env,mdb_filehandle_t * arg)10267 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
10268 {
10269 if (!env || !arg)
10270 return EINVAL;
10271
10272 *arg = env->me_fd;
10273 return MDB_SUCCESS;
10274 }
10275
10276 /** Common code for #mdb_stat() and #mdb_env_stat().
10277 * @param[in] env the environment to operate in.
10278 * @param[in] db the #MDB_db record containing the stats to return.
10279 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
10280 * @return 0, this function always succeeds.
10281 */
10282 static int ESECT
mdb_stat0(MDB_env * env,MDB_db * db,MDB_stat * arg)10283 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
10284 {
10285 arg->ms_psize = env->me_psize;
10286 arg->ms_depth = db->md_depth;
10287 arg->ms_branch_pages = db->md_branch_pages;
10288 arg->ms_leaf_pages = db->md_leaf_pages;
10289 arg->ms_overflow_pages = db->md_overflow_pages;
10290 arg->ms_entries = db->md_entries;
10291
10292 return MDB_SUCCESS;
10293 }
10294
10295 int ESECT
mdb_env_stat(MDB_env * env,MDB_stat * arg)10296 mdb_env_stat(MDB_env *env, MDB_stat *arg)
10297 {
10298 MDB_meta *meta;
10299
10300 if (env == NULL || arg == NULL)
10301 return EINVAL;
10302
10303 meta = mdb_env_pick_meta(env);
10304
10305 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg);
10306 }
10307
10308 int ESECT
mdb_env_info(MDB_env * env,MDB_envinfo * arg)10309 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
10310 {
10311 MDB_meta *meta;
10312
10313 if (env == NULL || arg == NULL)
10314 return EINVAL;
10315
10316 meta = mdb_env_pick_meta(env);
10317 arg->me_mapaddr = meta->mm_address;
10318 arg->me_last_pgno = meta->mm_last_pg;
10319 arg->me_last_txnid = meta->mm_txnid;
10320
10321 arg->me_mapsize = env->me_mapsize;
10322 arg->me_maxreaders = env->me_maxreaders;
10323 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
10324 return MDB_SUCCESS;
10325 }
10326
10327 /** Set the default comparison functions for a database.
10328 * Called immediately after a database is opened to set the defaults.
10329 * The user can then override them with #mdb_set_compare() or
10330 * #mdb_set_dupsort().
10331 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
10332 * @param[in] dbi A database handle returned by #mdb_dbi_open()
10333 */
10334 static void
mdb_default_cmp(MDB_txn * txn,MDB_dbi dbi)10335 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
10336 {
10337 uint16_t f = txn->mt_dbs[dbi].md_flags;
10338
10339 txn->mt_dbxs[dbi].md_cmp =
10340 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
10341 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
10342
10343 txn->mt_dbxs[dbi].md_dcmp =
10344 !(f & MDB_DUPSORT) ? 0 :
10345 ((f & MDB_INTEGERDUP)
10346 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
10347 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
10348 }
10349
mdb_dbi_open(MDB_txn * txn,const char * name,unsigned int flags,MDB_dbi * dbi)10350 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
10351 {
10352 MDB_val key, data;
10353 MDB_dbi i;
10354 MDB_cursor mc;
10355 MDB_db dummy;
10356 int rc, dbflag, exact;
10357 unsigned int unused = 0, seq;
10358 char *namedup;
10359 size_t len;
10360
10361 if (flags & ~VALID_FLAGS)
10362 return EINVAL;
10363 if (txn->mt_flags & MDB_TXN_BLOCKED)
10364 return MDB_BAD_TXN;
10365
10366 /* main DB? */
10367 if (!name) {
10368 *dbi = MAIN_DBI;
10369 if (flags & PERSISTENT_FLAGS) {
10370 uint16_t f2 = flags & PERSISTENT_FLAGS;
10371 /* make sure flag changes get committed */
10372 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
10373 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
10374 txn->mt_flags |= MDB_TXN_DIRTY;
10375 }
10376 }
10377 mdb_default_cmp(txn, MAIN_DBI);
10378 return MDB_SUCCESS;
10379 }
10380
10381 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
10382 mdb_default_cmp(txn, MAIN_DBI);
10383 }
10384
10385 /* Is the DB already open? */
10386 len = strlen(name);
10387 for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
10388 if (!txn->mt_dbxs[i].md_name.mv_size) {
10389 /* Remember this free slot */
10390 if (!unused) unused = i;
10391 continue;
10392 }
10393 if (len == txn->mt_dbxs[i].md_name.mv_size &&
10394 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
10395 *dbi = i;
10396 return MDB_SUCCESS;
10397 }
10398 }
10399
10400 /* If no free slot and max hit, fail */
10401 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
10402 return MDB_DBS_FULL;
10403
10404 /* Cannot mix named databases with some mainDB flags */
10405 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
10406 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
10407
10408 /* Find the DB info */
10409 dbflag = DB_NEW|DB_VALID|DB_USRVALID;
10410 exact = 0;
10411 key.mv_size = len;
10412 key.mv_data = (void *)name;
10413 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
10414 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
10415 if (rc == MDB_SUCCESS) {
10416 /* make sure this is actually a DB */
10417 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
10418 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
10419 return MDB_INCOMPATIBLE;
10420 } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) {
10421 return rc;
10422 }
10423
10424 /* Done here so we cannot fail after creating a new DB */
10425 if ((namedup = strdup(name)) == NULL)
10426 return ENOMEM;
10427
10428 if (rc) {
10429 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */
10430 data.mv_size = sizeof(MDB_db);
10431 data.mv_data = &dummy;
10432 memset(&dummy, 0, sizeof(dummy));
10433 dummy.md_root = P_INVALID;
10434 dummy.md_flags = flags & PERSISTENT_FLAGS;
10435 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
10436 dbflag |= DB_DIRTY;
10437 }
10438
10439 if (rc) {
10440 free(namedup);
10441 } else {
10442 /* Got info, register DBI in this txn */
10443 unsigned int slot = unused ? unused : txn->mt_numdbs;
10444 txn->mt_dbxs[slot].md_name.mv_data = namedup;
10445 txn->mt_dbxs[slot].md_name.mv_size = len;
10446 txn->mt_dbxs[slot].md_rel = NULL;
10447 txn->mt_dbflags[slot] = dbflag;
10448 /* txn-> and env-> are the same in read txns, use
10449 * tmp variable to avoid undefined assignment
10450 */
10451 seq = ++txn->mt_env->me_dbiseqs[slot];
10452 txn->mt_dbiseqs[slot] = seq;
10453
10454 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
10455 *dbi = slot;
10456 mdb_default_cmp(txn, slot);
10457 if (!unused) {
10458 txn->mt_numdbs++;
10459 }
10460 }
10461
10462 return rc;
10463 }
10464
10465 int ESECT
mdb_stat(MDB_txn * txn,MDB_dbi dbi,MDB_stat * arg)10466 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
10467 {
10468 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID))
10469 return EINVAL;
10470
10471 if (txn->mt_flags & MDB_TXN_BLOCKED)
10472 return MDB_BAD_TXN;
10473
10474 if (txn->mt_dbflags[dbi] & DB_STALE) {
10475 MDB_cursor mc;
10476 MDB_xcursor mx;
10477 /* Stale, must read the DB's root. cursor_init does it for us. */
10478 mdb_cursor_init(&mc, txn, dbi, &mx);
10479 }
10480 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
10481 }
10482
mdb_dbi_close(MDB_env * env,MDB_dbi dbi)10483 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
10484 {
10485 char *ptr;
10486 if (dbi < CORE_DBS || dbi >= env->me_maxdbs)
10487 return;
10488 ptr = env->me_dbxs[dbi].md_name.mv_data;
10489 /* If there was no name, this was already closed */
10490 if (ptr) {
10491 env->me_dbxs[dbi].md_name.mv_data = NULL;
10492 env->me_dbxs[dbi].md_name.mv_size = 0;
10493 env->me_dbflags[dbi] = 0;
10494 env->me_dbiseqs[dbi]++;
10495 free(ptr);
10496 }
10497 }
10498
mdb_dbi_flags(MDB_txn * txn,MDB_dbi dbi,unsigned int * flags)10499 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
10500 {
10501 /* We could return the flags for the FREE_DBI too but what's the point? */
10502 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10503 return EINVAL;
10504 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
10505 return MDB_SUCCESS;
10506 }
10507
10508 /** Add all the DB's pages to the free list.
10509 * @param[in] mc Cursor on the DB to free.
10510 * @param[in] subs non-Zero to check for sub-DBs in this DB.
10511 * @return 0 on success, non-zero on failure.
10512 */
10513 static int
mdb_drop0(MDB_cursor * mc,int subs)10514 mdb_drop0(MDB_cursor *mc, int subs)
10515 {
10516 int rc;
10517
10518 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
10519 if (rc == MDB_SUCCESS) {
10520 MDB_txn *txn = mc->mc_txn;
10521 MDB_node *ni;
10522 MDB_cursor mx;
10523 unsigned int i;
10524
10525 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves.
10526 * This also avoids any P_LEAF2 pages, which have no nodes.
10527 * Also if the DB doesn't have sub-DBs and has no overflow
10528 * pages, omit scanning leaves.
10529 */
10530 if ((mc->mc_flags & C_SUB) ||
10531 (!subs && !mc->mc_db->md_overflow_pages))
10532 mdb_cursor_pop(mc);
10533
10534 mdb_cursor_copy(mc, &mx);
10535 #ifdef MDB_VL32
10536 /* bump refcount for mx's pages */
10537 for (i=0; i<mc->mc_snum; i++)
10538 mdb_page_get(&mx, mc->mc_pg[i]->mp_pgno, &mx.mc_pg[i], NULL);
10539 #endif
10540 while (mc->mc_snum > 0) {
10541 MDB_page *mp = mc->mc_pg[mc->mc_top];
10542 unsigned n = NUMKEYS(mp);
10543 if (IS_LEAF(mp)) {
10544 for (i=0; i<n; i++) {
10545 ni = NODEPTR(mp, i);
10546 if (ni->mn_flags & F_BIGDATA) {
10547 MDB_page *omp;
10548 pgno_t pg;
10549 memcpy(&pg, NODEDATA(ni), sizeof(pg));
10550 rc = mdb_page_get(mc, pg, &omp, NULL);
10551 if (rc != 0)
10552 goto done;
10553 mdb_cassert(mc, IS_OVERFLOW(omp));
10554 rc = mdb_midl_append_range(&txn->mt_free_pgs,
10555 pg, omp->mp_pages);
10556 if (rc)
10557 goto done;
10558 mc->mc_db->md_overflow_pages -= omp->mp_pages;
10559 if (!mc->mc_db->md_overflow_pages && !subs)
10560 break;
10561 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
10562 mdb_xcursor_init1(mc, ni);
10563 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
10564 if (rc)
10565 goto done;
10566 }
10567 }
10568 if (!subs && !mc->mc_db->md_overflow_pages)
10569 goto pop;
10570 } else {
10571 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
10572 goto done;
10573 for (i=0; i<n; i++) {
10574 pgno_t pg;
10575 ni = NODEPTR(mp, i);
10576 pg = NODEPGNO(ni);
10577 /* free it */
10578 mdb_midl_xappend(txn->mt_free_pgs, pg);
10579 }
10580 }
10581 if (!mc->mc_top)
10582 break;
10583 mc->mc_ki[mc->mc_top] = i;
10584 rc = mdb_cursor_sibling(mc, 1);
10585 if (rc) {
10586 if (rc != MDB_NOTFOUND)
10587 goto done;
10588 /* no more siblings, go back to beginning
10589 * of previous level.
10590 */
10591 pop:
10592 mdb_cursor_pop(mc);
10593 mc->mc_ki[0] = 0;
10594 for (i=1; i<mc->mc_snum; i++) {
10595 mc->mc_ki[i] = 0;
10596 mc->mc_pg[i] = mx.mc_pg[i];
10597 }
10598 }
10599 }
10600 /* free it */
10601 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
10602 done:
10603 if (rc)
10604 txn->mt_flags |= MDB_TXN_ERROR;
10605 /* drop refcount for mx's pages */
10606 MDB_CURSOR_UNREF(&mx, 0);
10607 } else if (rc == MDB_NOTFOUND) {
10608 rc = MDB_SUCCESS;
10609 }
10610 mc->mc_flags &= ~C_INITIALIZED;
10611 return rc;
10612 }
10613
mdb_drop(MDB_txn * txn,MDB_dbi dbi,int del)10614 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
10615 {
10616 MDB_cursor *mc, *m2;
10617 int rc;
10618
10619 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10620 return EINVAL;
10621
10622 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
10623 return EACCES;
10624
10625 if (TXN_DBI_CHANGED(txn, dbi))
10626 return MDB_BAD_DBI;
10627
10628 rc = mdb_cursor_open(txn, dbi, &mc);
10629 if (rc)
10630 return rc;
10631
10632 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
10633 /* Invalidate the dropped DB's cursors */
10634 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
10635 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
10636 if (rc)
10637 goto leave;
10638
10639 /* Can't delete the main DB */
10640 if (del && dbi >= CORE_DBS) {
10641 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
10642 if (!rc) {
10643 txn->mt_dbflags[dbi] = DB_STALE;
10644 mdb_dbi_close(txn->mt_env, dbi);
10645 } else {
10646 txn->mt_flags |= MDB_TXN_ERROR;
10647 }
10648 } else {
10649 /* reset the DB record, mark it dirty */
10650 txn->mt_dbflags[dbi] |= DB_DIRTY;
10651 txn->mt_dbs[dbi].md_depth = 0;
10652 txn->mt_dbs[dbi].md_branch_pages = 0;
10653 txn->mt_dbs[dbi].md_leaf_pages = 0;
10654 txn->mt_dbs[dbi].md_overflow_pages = 0;
10655 txn->mt_dbs[dbi].md_entries = 0;
10656 txn->mt_dbs[dbi].md_root = P_INVALID;
10657
10658 txn->mt_flags |= MDB_TXN_DIRTY;
10659 }
10660 leave:
10661 mdb_cursor_close(mc);
10662 return rc;
10663 }
10664
mdb_set_compare(MDB_txn * txn,MDB_dbi dbi,MDB_cmp_func * cmp)10665 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10666 {
10667 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10668 return EINVAL;
10669
10670 txn->mt_dbxs[dbi].md_cmp = cmp;
10671 return MDB_SUCCESS;
10672 }
10673
mdb_set_dupsort(MDB_txn * txn,MDB_dbi dbi,MDB_cmp_func * cmp)10674 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
10675 {
10676 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10677 return EINVAL;
10678
10679 txn->mt_dbxs[dbi].md_dcmp = cmp;
10680 return MDB_SUCCESS;
10681 }
10682
mdb_set_relfunc(MDB_txn * txn,MDB_dbi dbi,MDB_rel_func * rel)10683 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
10684 {
10685 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10686 return EINVAL;
10687
10688 txn->mt_dbxs[dbi].md_rel = rel;
10689 return MDB_SUCCESS;
10690 }
10691
mdb_set_relctx(MDB_txn * txn,MDB_dbi dbi,void * ctx)10692 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
10693 {
10694 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID))
10695 return EINVAL;
10696
10697 txn->mt_dbxs[dbi].md_relctx = ctx;
10698 return MDB_SUCCESS;
10699 }
10700
10701 int ESECT
mdb_env_get_maxkeysize(MDB_env * env)10702 mdb_env_get_maxkeysize(MDB_env *env)
10703 {
10704 return ENV_MAXKEY(env);
10705 }
10706
10707 int ESECT
mdb_reader_list(MDB_env * env,MDB_msg_func * func,void * ctx)10708 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
10709 {
10710 unsigned int i, rdrs;
10711 MDB_reader *mr;
10712 char buf[64];
10713 int rc = 0, first = 1;
10714
10715 if (!env || !func)
10716 return -1;
10717 if (!env->me_txns) {
10718 return func("(no reader locks)\n", ctx);
10719 }
10720 rdrs = env->me_txns->mti_numreaders;
10721 mr = env->me_txns->mti_readers;
10722 for (i=0; i<rdrs; i++) {
10723 if (mr[i].mr_pid) {
10724 txnid_t txnid = mr[i].mr_txnid;
10725 sprintf(buf, txnid == (txnid_t)-1 ?
10726 "%10d %"Z"x -\n" : "%10d %"Z"x %"Yu"\n",
10727 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
10728 if (first) {
10729 first = 0;
10730 rc = func(" pid thread txnid\n", ctx);
10731 if (rc < 0)
10732 break;
10733 }
10734 rc = func(buf, ctx);
10735 if (rc < 0)
10736 break;
10737 }
10738 }
10739 if (first) {
10740 rc = func("(no active readers)\n", ctx);
10741 }
10742 return rc;
10743 }
10744
10745 /** Insert pid into list if not already present.
10746 * return -1 if already present.
10747 */
10748 static int ESECT
mdb_pid_insert(MDB_PID_T * ids,MDB_PID_T pid)10749 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
10750 {
10751 /* binary search of pid in list */
10752 unsigned base = 0;
10753 unsigned cursor = 1;
10754 int val = 0;
10755 unsigned n = ids[0];
10756
10757 while( 0 < n ) {
10758 unsigned pivot = n >> 1;
10759 cursor = base + pivot + 1;
10760 val = pid - ids[cursor];
10761
10762 if( val < 0 ) {
10763 n = pivot;
10764
10765 } else if ( val > 0 ) {
10766 base = cursor;
10767 n -= pivot + 1;
10768
10769 } else {
10770 /* found, so it's a duplicate */
10771 return -1;
10772 }
10773 }
10774
10775 if( val > 0 ) {
10776 ++cursor;
10777 }
10778 ids[0]++;
10779 for (n = ids[0]; n > cursor; n--)
10780 ids[n] = ids[n-1];
10781 ids[n] = pid;
10782 return 0;
10783 }
10784
10785 int ESECT
mdb_reader_check(MDB_env * env,int * dead)10786 mdb_reader_check(MDB_env *env, int *dead)
10787 {
10788 if (!env)
10789 return EINVAL;
10790 if (dead)
10791 *dead = 0;
10792 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
10793 }
10794
10795 /** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
10796 static int ESECT
mdb_reader_check0(MDB_env * env,int rlocked,int * dead)10797 mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
10798 {
10799 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
10800 unsigned int i, j, rdrs;
10801 MDB_reader *mr;
10802 MDB_PID_T *pids, pid;
10803 int rc = MDB_SUCCESS, count = 0;
10804
10805 rdrs = env->me_txns->mti_numreaders;
10806 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
10807 if (!pids)
10808 return ENOMEM;
10809 pids[0] = 0;
10810 mr = env->me_txns->mti_readers;
10811 for (i=0; i<rdrs; i++) {
10812 pid = mr[i].mr_pid;
10813 if (pid && pid != env->me_pid) {
10814 if (mdb_pid_insert(pids, pid) == 0) {
10815 if (!mdb_reader_pid(env, Pidcheck, pid)) {
10816 /* Stale reader found */
10817 j = i;
10818 if (rmutex) {
10819 if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
10820 if ((rc = mdb_mutex_failed(env, rmutex, rc)))
10821 break;
10822 rdrs = 0; /* the above checked all readers */
10823 } else {
10824 /* Recheck, a new process may have reused pid */
10825 if (mdb_reader_pid(env, Pidcheck, pid))
10826 j = rdrs;
10827 }
10828 }
10829 for (; j<rdrs; j++)
10830 if (mr[j].mr_pid == pid) {
10831 DPRINTF(("clear stale reader pid %u txn %"Yd,
10832 (unsigned) pid, mr[j].mr_txnid));
10833 mr[j].mr_pid = 0;
10834 count++;
10835 }
10836 if (rmutex)
10837 UNLOCK_MUTEX(rmutex);
10838 }
10839 }
10840 }
10841 }
10842 free(pids);
10843 if (dead)
10844 *dead = count;
10845 return rc;
10846 }
10847
10848 #ifdef MDB_ROBUST_SUPPORTED
10849 /** Handle #LOCK_MUTEX0() failure.
10850 * Try to repair the lock file if the mutex owner died.
10851 * @param[in] env the environment handle
10852 * @param[in] mutex LOCK_MUTEX0() mutex
10853 * @param[in] rc LOCK_MUTEX0() error (nonzero)
10854 * @return 0 on success with the mutex locked, or an error code on failure.
10855 */
10856 static int ESECT
mdb_mutex_failed(MDB_env * env,mdb_mutexref_t mutex,int rc)10857 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
10858 {
10859 int rlocked, rc2;
10860 MDB_meta *meta;
10861
10862 if (rc == MDB_OWNERDEAD) {
10863 /* We own the mutex. Clean up after dead previous owner. */
10864 rc = MDB_SUCCESS;
10865 rlocked = (mutex == env->me_rmutex);
10866 if (!rlocked) {
10867 /* Keep mti_txnid updated, otherwise next writer can
10868 * overwrite data which latest meta page refers to.
10869 */
10870 meta = mdb_env_pick_meta(env);
10871 env->me_txns->mti_txnid = meta->mm_txnid;
10872 /* env is hosed if the dead thread was ours */
10873 if (env->me_txn) {
10874 env->me_flags |= MDB_FATAL_ERROR;
10875 env->me_txn = NULL;
10876 rc = MDB_PANIC;
10877 }
10878 }
10879 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
10880 (rc ? "this process' env is hosed" : "recovering")));
10881 rc2 = mdb_reader_check0(env, rlocked, NULL);
10882 if (rc2 == 0)
10883 rc2 = mdb_mutex_consistent(mutex);
10884 if (rc || (rc = rc2)) {
10885 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
10886 UNLOCK_MUTEX(mutex);
10887 }
10888 } else {
10889 #ifdef _WIN32
10890 rc = ErrCode();
10891 #endif
10892 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
10893 }
10894
10895 return rc;
10896 }
10897 #endif /* MDB_ROBUST_SUPPORTED */
10898 /** @} */
10899
10900 #if defined(_WIN32)
utf8_to_utf16(const char * src,int srcsize,wchar_t ** dst,int * dstsize)10901 static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize)
10902 {
10903 int need;
10904 wchar_t *result;
10905 need = MultiByteToWideChar(CP_UTF8, 0, src, srcsize, NULL, 0);
10906 if (need == 0xFFFD)
10907 return EILSEQ;
10908 if (need == 0)
10909 return EINVAL;
10910 result = malloc(sizeof(wchar_t) * need);
10911 if (!result)
10912 return ENOMEM;
10913 MultiByteToWideChar(CP_UTF8, 0, src, srcsize, result, need);
10914 if (dstsize)
10915 *dstsize = need;
10916 *dst = result;
10917 return 0;
10918 }
10919 #endif /* defined(_WIN32) */
10920