1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2004 Poul-Henning Kamp 5 * Copyright (c) 1994,1997 John S. Dyson 6 * Copyright (c) 2013 The FreeBSD Foundation 7 * All rights reserved. 8 * 9 * Portions of this software were developed by Konstantin Belousov 10 * under sponsorship from the FreeBSD Foundation. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 /* 35 * this file contains a new buffer I/O scheme implementing a coherent 36 * VM object and buffer cache scheme. Pains have been taken to make 37 * sure that the performance degradation associated with schemes such 38 * as this is not realized. 39 * 40 * Author: John S. Dyson 41 * Significant help during the development and debugging phases 42 * had been provided by David Greenman, also of the FreeBSD core team. 43 * 44 * see man buf(9) for more info. 45 */ 46 47 #include <sys/cdefs.h> 48 __FBSDID("$FreeBSD$"); 49 50 #include <sys/param.h> 51 #include <sys/systm.h> 52 #include <sys/asan.h> 53 #include <sys/bio.h> 54 #include <sys/bitset.h> 55 #include <sys/conf.h> 56 #include <sys/counter.h> 57 #include <sys/buf.h> 58 #include <sys/devicestat.h> 59 #include <sys/eventhandler.h> 60 #include <sys/fail.h> 61 #include <sys/ktr.h> 62 #include <sys/limits.h> 63 #include <sys/lock.h> 64 #include <sys/malloc.h> 65 #include <sys/mount.h> 66 #include <sys/mutex.h> 67 #include <sys/kernel.h> 68 #include <sys/kthread.h> 69 #include <sys/proc.h> 70 #include <sys/racct.h> 71 #include <sys/refcount.h> 72 #include <sys/resourcevar.h> 73 #include <sys/rwlock.h> 74 #include <sys/smp.h> 75 #include <sys/sysctl.h> 76 #include <sys/syscallsubr.h> 77 #include <sys/vmem.h> 78 #include <sys/vmmeter.h> 79 #include <sys/vnode.h> 80 #include <sys/watchdog.h> 81 #include <geom/geom.h> 82 #include <vm/vm.h> 83 #include <vm/vm_param.h> 84 #include <vm/vm_kern.h> 85 #include <vm/vm_object.h> 86 #include <vm/vm_page.h> 87 #include <vm/vm_pageout.h> 88 #include <vm/vm_pager.h> 89 #include <vm/vm_extern.h> 90 #include <vm/vm_map.h> 91 #include <vm/swap_pager.h> 92 93 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 94 95 struct bio_ops bioops; /* I/O operation notification */ 96 97 struct buf_ops buf_ops_bio = { 98 .bop_name = "buf_ops_bio", 99 .bop_write = bufwrite, 100 .bop_strategy = bufstrategy, 101 .bop_sync = bufsync, 102 .bop_bdflush = bufbdflush, 103 }; 104 105 struct bufqueue { 106 struct mtx_padalign bq_lock; 107 TAILQ_HEAD(, buf) bq_queue; 108 uint8_t bq_index; 109 uint16_t bq_subqueue; 110 int bq_len; 111 } __aligned(CACHE_LINE_SIZE); 112 113 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock) 114 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq))) 115 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq))) 116 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED) 117 118 struct bufdomain { 119 struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */ 120 struct bufqueue bd_dirtyq; 121 struct bufqueue *bd_cleanq; 122 struct mtx_padalign bd_run_lock; 123 /* Constants */ 124 long bd_maxbufspace; 125 long bd_hibufspace; 126 long bd_lobufspace; 127 long bd_bufspacethresh; 128 int bd_hifreebuffers; 129 int bd_lofreebuffers; 130 int bd_hidirtybuffers; 131 int bd_lodirtybuffers; 132 int bd_dirtybufthresh; 133 int bd_lim; 134 /* atomics */ 135 int bd_wanted; 136 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers; 137 int __aligned(CACHE_LINE_SIZE) bd_running; 138 long __aligned(CACHE_LINE_SIZE) bd_bufspace; 139 int __aligned(CACHE_LINE_SIZE) bd_freebuffers; 140 } __aligned(CACHE_LINE_SIZE); 141 142 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock) 143 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd))) 144 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd))) 145 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED) 146 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock) 147 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd))) 148 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd))) 149 #define BD_DOMAIN(bd) (bd - bdomain) 150 151 static char *buf; /* buffer header pool */ 152 static struct buf * 153 nbufp(unsigned i) 154 { 155 return ((struct buf *)(buf + (sizeof(struct buf) + 156 sizeof(vm_page_t) * atop(maxbcachebuf)) * i)); 157 } 158 159 caddr_t __read_mostly unmapped_buf; 160 161 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */ 162 struct proc *bufdaemonproc; 163 164 static void vm_hold_free_pages(struct buf *bp, int newbsize); 165 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 166 vm_offset_t to); 167 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 168 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 169 vm_page_t m); 170 static void vfs_clean_pages_dirty_buf(struct buf *bp); 171 static void vfs_setdirty_range(struct buf *bp); 172 static void vfs_vmio_invalidate(struct buf *bp); 173 static void vfs_vmio_truncate(struct buf *bp, int npages); 174 static void vfs_vmio_extend(struct buf *bp, int npages, int size); 175 static int vfs_bio_clcheck(struct vnode *vp, int size, 176 daddr_t lblkno, daddr_t blkno); 177 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int, 178 void (*)(struct buf *)); 179 static int buf_flush(struct vnode *vp, struct bufdomain *, int); 180 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int); 181 static void buf_daemon(void); 182 static __inline void bd_wakeup(void); 183 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS); 184 static void bufkva_reclaim(vmem_t *, int); 185 static void bufkva_free(struct buf *); 186 static int buf_import(void *, void **, int, int, int); 187 static void buf_release(void *, void **, int); 188 static void maxbcachebuf_adjust(void); 189 static inline struct bufdomain *bufdomain(struct buf *); 190 static void bq_remove(struct bufqueue *bq, struct buf *bp); 191 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock); 192 static int buf_recycle(struct bufdomain *, bool kva); 193 static void bq_init(struct bufqueue *bq, int qindex, int cpu, 194 const char *lockname); 195 static void bd_init(struct bufdomain *bd); 196 static int bd_flushall(struct bufdomain *bd); 197 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS); 198 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS); 199 200 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 201 int vmiodirenable = TRUE; 202 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 203 "Use the VM system for directory writes"); 204 long runningbufspace; 205 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 206 "Amount of presently outstanding async buffer io"); 207 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 208 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers"); 209 static counter_u64_t bufkvaspace; 210 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, 211 "Kernel virtual memory used for buffers"); 212 static long maxbufspace; 213 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace, 214 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace, 215 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L", 216 "Maximum allowed value of bufspace (including metadata)"); 217 static long bufmallocspace; 218 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 219 "Amount of malloced memory for buffers"); 220 static long maxbufmallocspace; 221 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 222 0, "Maximum amount of malloced memory for buffers"); 223 static long lobufspace; 224 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace, 225 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace, 226 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L", 227 "Minimum amount of buffers we want to have"); 228 long hibufspace; 229 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace, 230 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace, 231 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L", 232 "Maximum allowed value of bufspace (excluding metadata)"); 233 long bufspacethresh; 234 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh, 235 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh, 236 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L", 237 "Bufspace consumed before waking the daemon to free some"); 238 static counter_u64_t buffreekvacnt; 239 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 240 "Number of times we have freed the KVA space from some buffer"); 241 static counter_u64_t bufdefragcnt; 242 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 243 "Number of times we have had to repeat buffer allocation to defragment"); 244 static long lorunningspace; 245 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 246 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L", 247 "Minimum preferred space used for in-progress I/O"); 248 static long hirunningspace; 249 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 250 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L", 251 "Maximum amount of space to use for in-progress I/O"); 252 int dirtybufferflushes; 253 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 254 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 255 int bdwriteskip; 256 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 257 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 258 int altbufferflushes; 259 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS, 260 &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers"); 261 static int recursiveflushes; 262 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS, 263 &recursiveflushes, 0, "Number of flushes skipped due to being recursive"); 264 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS); 265 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers, 266 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I", 267 "Number of buffers that are dirty (has unwritten changes) at the moment"); 268 static int lodirtybuffers; 269 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers, 270 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers, 271 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I", 272 "How many buffers we want to have free before bufdaemon can sleep"); 273 static int hidirtybuffers; 274 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers, 275 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers, 276 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I", 277 "When the number of dirty buffers is considered severe"); 278 int dirtybufthresh; 279 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh, 280 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh, 281 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I", 282 "Number of bdwrite to bawrite conversions to clear dirty buffers"); 283 static int numfreebuffers; 284 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 285 "Number of free buffers"); 286 static int lofreebuffers; 287 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers, 288 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers, 289 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I", 290 "Target number of free buffers"); 291 static int hifreebuffers; 292 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers, 293 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers, 294 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I", 295 "Threshold for clean buffer recycling"); 296 static counter_u64_t getnewbufcalls; 297 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, 298 &getnewbufcalls, "Number of calls to getnewbuf"); 299 static counter_u64_t getnewbufrestarts; 300 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, 301 &getnewbufrestarts, 302 "Number of times getnewbuf has had to restart a buffer acquisition"); 303 static counter_u64_t mappingrestarts; 304 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD, 305 &mappingrestarts, 306 "Number of times getblk has had to restart a buffer mapping for " 307 "unmapped buffer"); 308 static counter_u64_t numbufallocfails; 309 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW, 310 &numbufallocfails, "Number of times buffer allocations failed"); 311 static int flushbufqtarget = 100; 312 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 313 "Amount of work to do in flushbufqueues when helping bufdaemon"); 314 static counter_u64_t notbufdflushes; 315 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 316 "Number of dirty buffer flushes done by the bufdaemon helpers"); 317 static long barrierwrites; 318 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS, 319 &barrierwrites, 0, "Number of barrier writes"); 320 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, 321 &unmapped_buf_allowed, 0, 322 "Permit the use of the unmapped i/o"); 323 int maxbcachebuf = MAXBCACHEBUF; 324 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0, 325 "Maximum size of a buffer cache block"); 326 327 /* 328 * This lock synchronizes access to bd_request. 329 */ 330 static struct mtx_padalign __exclusive_cache_line bdlock; 331 332 /* 333 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 334 * waitrunningbufspace(). 335 */ 336 static struct mtx_padalign __exclusive_cache_line rbreqlock; 337 338 /* 339 * Lock that protects bdirtywait. 340 */ 341 static struct mtx_padalign __exclusive_cache_line bdirtylock; 342 343 /* 344 * Wakeup point for bufdaemon, as well as indicator of whether it is already 345 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 346 * is idling. 347 */ 348 static int bd_request; 349 350 /* 351 * Request for the buf daemon to write more buffers than is indicated by 352 * lodirtybuf. This may be necessary to push out excess dependencies or 353 * defragment the address space where a simple count of the number of dirty 354 * buffers is insufficient to characterize the demand for flushing them. 355 */ 356 static int bd_speedupreq; 357 358 /* 359 * Synchronization (sleep/wakeup) variable for active buffer space requests. 360 * Set when wait starts, cleared prior to wakeup(). 361 * Used in runningbufwakeup() and waitrunningbufspace(). 362 */ 363 static int runningbufreq; 364 365 /* 366 * Synchronization for bwillwrite() waiters. 367 */ 368 static int bdirtywait; 369 370 /* 371 * Definitions for the buffer free lists. 372 */ 373 #define QUEUE_NONE 0 /* on no queue */ 374 #define QUEUE_EMPTY 1 /* empty buffer headers */ 375 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 376 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */ 377 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */ 378 379 /* Maximum number of buffer domains. */ 380 #define BUF_DOMAINS 8 381 382 struct bufdomainset bdlodirty; /* Domains > lodirty */ 383 struct bufdomainset bdhidirty; /* Domains > hidirty */ 384 385 /* Configured number of clean queues. */ 386 static int __read_mostly buf_domains; 387 388 BITSET_DEFINE(bufdomainset, BUF_DOMAINS); 389 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS]; 390 struct bufqueue __exclusive_cache_line bqempty; 391 392 /* 393 * per-cpu empty buffer cache. 394 */ 395 uma_zone_t buf_zone; 396 397 /* 398 * Single global constant for BUF_WMESG, to avoid getting multiple references. 399 * buf_wmesg is referred from macros. 400 */ 401 const char *buf_wmesg = BUF_WMESG; 402 403 static int 404 sysctl_runningspace(SYSCTL_HANDLER_ARGS) 405 { 406 long value; 407 int error; 408 409 value = *(long *)arg1; 410 error = sysctl_handle_long(oidp, &value, 0, req); 411 if (error != 0 || req->newptr == NULL) 412 return (error); 413 mtx_lock(&rbreqlock); 414 if (arg1 == &hirunningspace) { 415 if (value < lorunningspace) 416 error = EINVAL; 417 else 418 hirunningspace = value; 419 } else { 420 KASSERT(arg1 == &lorunningspace, 421 ("%s: unknown arg1", __func__)); 422 if (value > hirunningspace) 423 error = EINVAL; 424 else 425 lorunningspace = value; 426 } 427 mtx_unlock(&rbreqlock); 428 return (error); 429 } 430 431 static int 432 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS) 433 { 434 int error; 435 int value; 436 int i; 437 438 value = *(int *)arg1; 439 error = sysctl_handle_int(oidp, &value, 0, req); 440 if (error != 0 || req->newptr == NULL) 441 return (error); 442 *(int *)arg1 = value; 443 for (i = 0; i < buf_domains; i++) 444 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = 445 value / buf_domains; 446 447 return (error); 448 } 449 450 static int 451 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS) 452 { 453 long value; 454 int error; 455 int i; 456 457 value = *(long *)arg1; 458 error = sysctl_handle_long(oidp, &value, 0, req); 459 if (error != 0 || req->newptr == NULL) 460 return (error); 461 *(long *)arg1 = value; 462 for (i = 0; i < buf_domains; i++) 463 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = 464 value / buf_domains; 465 466 return (error); 467 } 468 469 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 470 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 471 static int 472 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 473 { 474 long lvalue; 475 int ivalue; 476 int i; 477 478 lvalue = 0; 479 for (i = 0; i < buf_domains; i++) 480 lvalue += bdomain[i].bd_bufspace; 481 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 482 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 483 if (lvalue > INT_MAX) 484 /* On overflow, still write out a long to trigger ENOMEM. */ 485 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 486 ivalue = lvalue; 487 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 488 } 489 #else 490 static int 491 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 492 { 493 long lvalue; 494 int i; 495 496 lvalue = 0; 497 for (i = 0; i < buf_domains; i++) 498 lvalue += bdomain[i].bd_bufspace; 499 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 500 } 501 #endif 502 503 static int 504 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS) 505 { 506 int value; 507 int i; 508 509 value = 0; 510 for (i = 0; i < buf_domains; i++) 511 value += bdomain[i].bd_numdirtybuffers; 512 return (sysctl_handle_int(oidp, &value, 0, req)); 513 } 514 515 /* 516 * bdirtywakeup: 517 * 518 * Wakeup any bwillwrite() waiters. 519 */ 520 static void 521 bdirtywakeup(void) 522 { 523 mtx_lock(&bdirtylock); 524 if (bdirtywait) { 525 bdirtywait = 0; 526 wakeup(&bdirtywait); 527 } 528 mtx_unlock(&bdirtylock); 529 } 530 531 /* 532 * bd_clear: 533 * 534 * Clear a domain from the appropriate bitsets when dirtybuffers 535 * is decremented. 536 */ 537 static void 538 bd_clear(struct bufdomain *bd) 539 { 540 541 mtx_lock(&bdirtylock); 542 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers) 543 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); 544 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers) 545 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); 546 mtx_unlock(&bdirtylock); 547 } 548 549 /* 550 * bd_set: 551 * 552 * Set a domain in the appropriate bitsets when dirtybuffers 553 * is incremented. 554 */ 555 static void 556 bd_set(struct bufdomain *bd) 557 { 558 559 mtx_lock(&bdirtylock); 560 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers) 561 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); 562 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers) 563 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); 564 mtx_unlock(&bdirtylock); 565 } 566 567 /* 568 * bdirtysub: 569 * 570 * Decrement the numdirtybuffers count by one and wakeup any 571 * threads blocked in bwillwrite(). 572 */ 573 static void 574 bdirtysub(struct buf *bp) 575 { 576 struct bufdomain *bd; 577 int num; 578 579 bd = bufdomain(bp); 580 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1); 581 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) 582 bdirtywakeup(); 583 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) 584 bd_clear(bd); 585 } 586 587 /* 588 * bdirtyadd: 589 * 590 * Increment the numdirtybuffers count by one and wakeup the buf 591 * daemon if needed. 592 */ 593 static void 594 bdirtyadd(struct buf *bp) 595 { 596 struct bufdomain *bd; 597 int num; 598 599 /* 600 * Only do the wakeup once as we cross the boundary. The 601 * buf daemon will keep running until the condition clears. 602 */ 603 bd = bufdomain(bp); 604 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1); 605 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) 606 bd_wakeup(); 607 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) 608 bd_set(bd); 609 } 610 611 /* 612 * bufspace_daemon_wakeup: 613 * 614 * Wakeup the daemons responsible for freeing clean bufs. 615 */ 616 static void 617 bufspace_daemon_wakeup(struct bufdomain *bd) 618 { 619 620 /* 621 * avoid the lock if the daemon is running. 622 */ 623 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) { 624 BD_RUN_LOCK(bd); 625 atomic_store_int(&bd->bd_running, 1); 626 wakeup(&bd->bd_running); 627 BD_RUN_UNLOCK(bd); 628 } 629 } 630 631 /* 632 * bufspace_daemon_wait: 633 * 634 * Sleep until the domain falls below a limit or one second passes. 635 */ 636 static void 637 bufspace_daemon_wait(struct bufdomain *bd) 638 { 639 /* 640 * Re-check our limits and sleep. bd_running must be 641 * cleared prior to checking the limits to avoid missed 642 * wakeups. The waker will adjust one of bufspace or 643 * freebuffers prior to checking bd_running. 644 */ 645 BD_RUN_LOCK(bd); 646 atomic_store_int(&bd->bd_running, 0); 647 if (bd->bd_bufspace < bd->bd_bufspacethresh && 648 bd->bd_freebuffers > bd->bd_lofreebuffers) { 649 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP, 650 "-", hz); 651 } else { 652 /* Avoid spurious wakeups while running. */ 653 atomic_store_int(&bd->bd_running, 1); 654 BD_RUN_UNLOCK(bd); 655 } 656 } 657 658 /* 659 * bufspace_adjust: 660 * 661 * Adjust the reported bufspace for a KVA managed buffer, possibly 662 * waking any waiters. 663 */ 664 static void 665 bufspace_adjust(struct buf *bp, int bufsize) 666 { 667 struct bufdomain *bd; 668 long space; 669 int diff; 670 671 KASSERT((bp->b_flags & B_MALLOC) == 0, 672 ("bufspace_adjust: malloc buf %p", bp)); 673 bd = bufdomain(bp); 674 diff = bufsize - bp->b_bufsize; 675 if (diff < 0) { 676 atomic_subtract_long(&bd->bd_bufspace, -diff); 677 } else if (diff > 0) { 678 space = atomic_fetchadd_long(&bd->bd_bufspace, diff); 679 /* Wake up the daemon on the transition. */ 680 if (space < bd->bd_bufspacethresh && 681 space + diff >= bd->bd_bufspacethresh) 682 bufspace_daemon_wakeup(bd); 683 } 684 bp->b_bufsize = bufsize; 685 } 686 687 /* 688 * bufspace_reserve: 689 * 690 * Reserve bufspace before calling allocbuf(). metadata has a 691 * different space limit than data. 692 */ 693 static int 694 bufspace_reserve(struct bufdomain *bd, int size, bool metadata) 695 { 696 long limit, new; 697 long space; 698 699 if (metadata) 700 limit = bd->bd_maxbufspace; 701 else 702 limit = bd->bd_hibufspace; 703 space = atomic_fetchadd_long(&bd->bd_bufspace, size); 704 new = space + size; 705 if (new > limit) { 706 atomic_subtract_long(&bd->bd_bufspace, size); 707 return (ENOSPC); 708 } 709 710 /* Wake up the daemon on the transition. */ 711 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh) 712 bufspace_daemon_wakeup(bd); 713 714 return (0); 715 } 716 717 /* 718 * bufspace_release: 719 * 720 * Release reserved bufspace after bufspace_adjust() has consumed it. 721 */ 722 static void 723 bufspace_release(struct bufdomain *bd, int size) 724 { 725 726 atomic_subtract_long(&bd->bd_bufspace, size); 727 } 728 729 /* 730 * bufspace_wait: 731 * 732 * Wait for bufspace, acting as the buf daemon if a locked vnode is 733 * supplied. bd_wanted must be set prior to polling for space. The 734 * operation must be re-tried on return. 735 */ 736 static void 737 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags, 738 int slpflag, int slptimeo) 739 { 740 struct thread *td; 741 int error, fl, norunbuf; 742 743 if ((gbflags & GB_NOWAIT_BD) != 0) 744 return; 745 746 td = curthread; 747 BD_LOCK(bd); 748 while (bd->bd_wanted) { 749 if (vp != NULL && vp->v_type != VCHR && 750 (td->td_pflags & TDP_BUFNEED) == 0) { 751 BD_UNLOCK(bd); 752 /* 753 * getblk() is called with a vnode locked, and 754 * some majority of the dirty buffers may as 755 * well belong to the vnode. Flushing the 756 * buffers there would make a progress that 757 * cannot be achieved by the buf_daemon, that 758 * cannot lock the vnode. 759 */ 760 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 761 (td->td_pflags & TDP_NORUNNINGBUF); 762 763 /* 764 * Play bufdaemon. The getnewbuf() function 765 * may be called while the thread owns lock 766 * for another dirty buffer for the same 767 * vnode, which makes it impossible to use 768 * VOP_FSYNC() there, due to the buffer lock 769 * recursion. 770 */ 771 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 772 fl = buf_flush(vp, bd, flushbufqtarget); 773 td->td_pflags &= norunbuf; 774 BD_LOCK(bd); 775 if (fl != 0) 776 continue; 777 if (bd->bd_wanted == 0) 778 break; 779 } 780 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd), 781 (PRIBIO + 4) | slpflag, "newbuf", slptimeo); 782 if (error != 0) 783 break; 784 } 785 BD_UNLOCK(bd); 786 } 787 788 /* 789 * bufspace_daemon: 790 * 791 * buffer space management daemon. Tries to maintain some marginal 792 * amount of free buffer space so that requesting processes neither 793 * block nor work to reclaim buffers. 794 */ 795 static void 796 bufspace_daemon(void *arg) 797 { 798 struct bufdomain *bd; 799 800 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread, 801 SHUTDOWN_PRI_LAST + 100); 802 803 bd = arg; 804 for (;;) { 805 kthread_suspend_check(); 806 807 /* 808 * Free buffers from the clean queue until we meet our 809 * targets. 810 * 811 * Theory of operation: The buffer cache is most efficient 812 * when some free buffer headers and space are always 813 * available to getnewbuf(). This daemon attempts to prevent 814 * the excessive blocking and synchronization associated 815 * with shortfall. It goes through three phases according 816 * demand: 817 * 818 * 1) The daemon wakes up voluntarily once per-second 819 * during idle periods when the counters are below 820 * the wakeup thresholds (bufspacethresh, lofreebuffers). 821 * 822 * 2) The daemon wakes up as we cross the thresholds 823 * ahead of any potential blocking. This may bounce 824 * slightly according to the rate of consumption and 825 * release. 826 * 827 * 3) The daemon and consumers are starved for working 828 * clean buffers. This is the 'bufspace' sleep below 829 * which will inefficiently trade bufs with bqrelse 830 * until we return to condition 2. 831 */ 832 while (bd->bd_bufspace > bd->bd_lobufspace || 833 bd->bd_freebuffers < bd->bd_hifreebuffers) { 834 if (buf_recycle(bd, false) != 0) { 835 if (bd_flushall(bd)) 836 continue; 837 /* 838 * Speedup dirty if we've run out of clean 839 * buffers. This is possible in particular 840 * because softdep may held many bufs locked 841 * pending writes to other bufs which are 842 * marked for delayed write, exhausting 843 * clean space until they are written. 844 */ 845 bd_speedup(); 846 BD_LOCK(bd); 847 if (bd->bd_wanted) { 848 msleep(&bd->bd_wanted, BD_LOCKPTR(bd), 849 PRIBIO|PDROP, "bufspace", hz/10); 850 } else 851 BD_UNLOCK(bd); 852 } 853 maybe_yield(); 854 } 855 bufspace_daemon_wait(bd); 856 } 857 } 858 859 /* 860 * bufmallocadjust: 861 * 862 * Adjust the reported bufspace for a malloc managed buffer, possibly 863 * waking any waiters. 864 */ 865 static void 866 bufmallocadjust(struct buf *bp, int bufsize) 867 { 868 int diff; 869 870 KASSERT((bp->b_flags & B_MALLOC) != 0, 871 ("bufmallocadjust: non-malloc buf %p", bp)); 872 diff = bufsize - bp->b_bufsize; 873 if (diff < 0) 874 atomic_subtract_long(&bufmallocspace, -diff); 875 else 876 atomic_add_long(&bufmallocspace, diff); 877 bp->b_bufsize = bufsize; 878 } 879 880 /* 881 * runningwakeup: 882 * 883 * Wake up processes that are waiting on asynchronous writes to fall 884 * below lorunningspace. 885 */ 886 static void 887 runningwakeup(void) 888 { 889 890 mtx_lock(&rbreqlock); 891 if (runningbufreq) { 892 runningbufreq = 0; 893 wakeup(&runningbufreq); 894 } 895 mtx_unlock(&rbreqlock); 896 } 897 898 /* 899 * runningbufwakeup: 900 * 901 * Decrement the outstanding write count according. 902 */ 903 void 904 runningbufwakeup(struct buf *bp) 905 { 906 long space, bspace; 907 908 bspace = bp->b_runningbufspace; 909 if (bspace == 0) 910 return; 911 space = atomic_fetchadd_long(&runningbufspace, -bspace); 912 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", 913 space, bspace)); 914 bp->b_runningbufspace = 0; 915 /* 916 * Only acquire the lock and wakeup on the transition from exceeding 917 * the threshold to falling below it. 918 */ 919 if (space < lorunningspace) 920 return; 921 if (space - bspace > lorunningspace) 922 return; 923 runningwakeup(); 924 } 925 926 /* 927 * waitrunningbufspace() 928 * 929 * runningbufspace is a measure of the amount of I/O currently 930 * running. This routine is used in async-write situations to 931 * prevent creating huge backups of pending writes to a device. 932 * Only asynchronous writes are governed by this function. 933 * 934 * This does NOT turn an async write into a sync write. It waits 935 * for earlier writes to complete and generally returns before the 936 * caller's write has reached the device. 937 */ 938 void 939 waitrunningbufspace(void) 940 { 941 942 mtx_lock(&rbreqlock); 943 while (runningbufspace > hirunningspace) { 944 runningbufreq = 1; 945 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 946 } 947 mtx_unlock(&rbreqlock); 948 } 949 950 /* 951 * vfs_buf_test_cache: 952 * 953 * Called when a buffer is extended. This function clears the B_CACHE 954 * bit if the newly extended portion of the buffer does not contain 955 * valid data. 956 */ 957 static __inline void 958 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off, 959 vm_offset_t size, vm_page_t m) 960 { 961 962 /* 963 * This function and its results are protected by higher level 964 * synchronization requiring vnode and buf locks to page in and 965 * validate pages. 966 */ 967 if (bp->b_flags & B_CACHE) { 968 int base = (foff + off) & PAGE_MASK; 969 if (vm_page_is_valid(m, base, size) == 0) 970 bp->b_flags &= ~B_CACHE; 971 } 972 } 973 974 /* Wake up the buffer daemon if necessary */ 975 static void 976 bd_wakeup(void) 977 { 978 979 mtx_lock(&bdlock); 980 if (bd_request == 0) { 981 bd_request = 1; 982 wakeup(&bd_request); 983 } 984 mtx_unlock(&bdlock); 985 } 986 987 /* 988 * Adjust the maxbcachbuf tunable. 989 */ 990 static void 991 maxbcachebuf_adjust(void) 992 { 993 int i; 994 995 /* 996 * maxbcachebuf must be a power of 2 >= MAXBSIZE. 997 */ 998 i = 2; 999 while (i * 2 <= maxbcachebuf) 1000 i *= 2; 1001 maxbcachebuf = i; 1002 if (maxbcachebuf < MAXBSIZE) 1003 maxbcachebuf = MAXBSIZE; 1004 if (maxbcachebuf > maxphys) 1005 maxbcachebuf = maxphys; 1006 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF) 1007 printf("maxbcachebuf=%d\n", maxbcachebuf); 1008 } 1009 1010 /* 1011 * bd_speedup - speedup the buffer cache flushing code 1012 */ 1013 void 1014 bd_speedup(void) 1015 { 1016 int needwake; 1017 1018 mtx_lock(&bdlock); 1019 needwake = 0; 1020 if (bd_speedupreq == 0 || bd_request == 0) 1021 needwake = 1; 1022 bd_speedupreq = 1; 1023 bd_request = 1; 1024 if (needwake) 1025 wakeup(&bd_request); 1026 mtx_unlock(&bdlock); 1027 } 1028 1029 #ifdef __i386__ 1030 #define TRANSIENT_DENOM 5 1031 #else 1032 #define TRANSIENT_DENOM 10 1033 #endif 1034 1035 /* 1036 * Calculating buffer cache scaling values and reserve space for buffer 1037 * headers. This is called during low level kernel initialization and 1038 * may be called more then once. We CANNOT write to the memory area 1039 * being reserved at this time. 1040 */ 1041 caddr_t 1042 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 1043 { 1044 int tuned_nbuf; 1045 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 1046 1047 /* 1048 * With KASAN or KMSAN enabled, the kernel map is shadowed. Account for 1049 * this when sizing maps based on the amount of physical memory 1050 * available. 1051 */ 1052 #if defined(KASAN) 1053 physmem_est = (physmem_est * KASAN_SHADOW_SCALE) / 1054 (KASAN_SHADOW_SCALE + 1); 1055 #elif defined(KMSAN) 1056 physmem_est /= 3; 1057 1058 /* 1059 * KMSAN cannot reliably determine whether buffer data is initialized 1060 * unless it is updated through a KVA mapping. 1061 */ 1062 unmapped_buf_allowed = 0; 1063 #endif 1064 1065 /* 1066 * physmem_est is in pages. Convert it to kilobytes (assumes 1067 * PAGE_SIZE is >= 1K) 1068 */ 1069 physmem_est = physmem_est * (PAGE_SIZE / 1024); 1070 1071 maxbcachebuf_adjust(); 1072 /* 1073 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 1074 * For the first 64MB of ram nominally allocate sufficient buffers to 1075 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 1076 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 1077 * the buffer cache we limit the eventual kva reservation to 1078 * maxbcache bytes. 1079 * 1080 * factor represents the 1/4 x ram conversion. 1081 */ 1082 if (nbuf == 0) { 1083 int factor = 4 * BKVASIZE / 1024; 1084 1085 nbuf = 50; 1086 if (physmem_est > 4096) 1087 nbuf += min((physmem_est - 4096) / factor, 1088 65536 / factor); 1089 if (physmem_est > 65536) 1090 nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 1091 32 * 1024 * 1024 / (factor * 5)); 1092 1093 if (maxbcache && nbuf > maxbcache / BKVASIZE) 1094 nbuf = maxbcache / BKVASIZE; 1095 tuned_nbuf = 1; 1096 } else 1097 tuned_nbuf = 0; 1098 1099 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 1100 maxbuf = (LONG_MAX / 3) / BKVASIZE; 1101 if (nbuf > maxbuf) { 1102 if (!tuned_nbuf) 1103 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 1104 maxbuf); 1105 nbuf = maxbuf; 1106 } 1107 1108 /* 1109 * Ideal allocation size for the transient bio submap is 10% 1110 * of the maximal space buffer map. This roughly corresponds 1111 * to the amount of the buffer mapped for typical UFS load. 1112 * 1113 * Clip the buffer map to reserve space for the transient 1114 * BIOs, if its extent is bigger than 90% (80% on i386) of the 1115 * maximum buffer map extent on the platform. 1116 * 1117 * The fall-back to the maxbuf in case of maxbcache unset, 1118 * allows to not trim the buffer KVA for the architectures 1119 * with ample KVA space. 1120 */ 1121 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 1122 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 1123 buf_sz = (long)nbuf * BKVASIZE; 1124 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 1125 (TRANSIENT_DENOM - 1)) { 1126 /* 1127 * There is more KVA than memory. Do not 1128 * adjust buffer map size, and assign the rest 1129 * of maxbuf to transient map. 1130 */ 1131 biotmap_sz = maxbuf_sz - buf_sz; 1132 } else { 1133 /* 1134 * Buffer map spans all KVA we could afford on 1135 * this platform. Give 10% (20% on i386) of 1136 * the buffer map to the transient bio map. 1137 */ 1138 biotmap_sz = buf_sz / TRANSIENT_DENOM; 1139 buf_sz -= biotmap_sz; 1140 } 1141 if (biotmap_sz / INT_MAX > maxphys) 1142 bio_transient_maxcnt = INT_MAX; 1143 else 1144 bio_transient_maxcnt = biotmap_sz / maxphys; 1145 /* 1146 * Artificially limit to 1024 simultaneous in-flight I/Os 1147 * using the transient mapping. 1148 */ 1149 if (bio_transient_maxcnt > 1024) 1150 bio_transient_maxcnt = 1024; 1151 if (tuned_nbuf) 1152 nbuf = buf_sz / BKVASIZE; 1153 } 1154 1155 if (nswbuf == 0) { 1156 nswbuf = min(nbuf / 4, 256); 1157 if (nswbuf < NSWBUF_MIN) 1158 nswbuf = NSWBUF_MIN; 1159 } 1160 1161 /* 1162 * Reserve space for the buffer cache buffers 1163 */ 1164 buf = (char *)v; 1165 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) * 1166 atop(maxbcachebuf)) * nbuf; 1167 1168 return (v); 1169 } 1170 1171 /* Initialize the buffer subsystem. Called before use of any buffers. */ 1172 void 1173 bufinit(void) 1174 { 1175 struct buf *bp; 1176 int i; 1177 1178 KASSERT(maxbcachebuf >= MAXBSIZE, 1179 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf, 1180 MAXBSIZE)); 1181 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock"); 1182 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 1183 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 1184 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); 1185 1186 unmapped_buf = (caddr_t)kva_alloc(maxphys); 1187 1188 /* finally, initialize each buffer header and stick on empty q */ 1189 for (i = 0; i < nbuf; i++) { 1190 bp = nbufp(i); 1191 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf)); 1192 bp->b_flags = B_INVAL; 1193 bp->b_rcred = NOCRED; 1194 bp->b_wcred = NOCRED; 1195 bp->b_qindex = QUEUE_NONE; 1196 bp->b_domain = -1; 1197 bp->b_subqueue = mp_maxid + 1; 1198 bp->b_xflags = 0; 1199 bp->b_data = bp->b_kvabase = unmapped_buf; 1200 LIST_INIT(&bp->b_dep); 1201 BUF_LOCKINIT(bp); 1202 bq_insert(&bqempty, bp, false); 1203 } 1204 1205 /* 1206 * maxbufspace is the absolute maximum amount of buffer space we are 1207 * allowed to reserve in KVM and in real terms. The absolute maximum 1208 * is nominally used by metadata. hibufspace is the nominal maximum 1209 * used by most other requests. The differential is required to 1210 * ensure that metadata deadlocks don't occur. 1211 * 1212 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 1213 * this may result in KVM fragmentation which is not handled optimally 1214 * by the system. XXX This is less true with vmem. We could use 1215 * PAGE_SIZE. 1216 */ 1217 maxbufspace = (long)nbuf * BKVASIZE; 1218 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10); 1219 lobufspace = (hibufspace / 20) * 19; /* 95% */ 1220 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2; 1221 1222 /* 1223 * Note: The 16 MiB upper limit for hirunningspace was chosen 1224 * arbitrarily and may need further tuning. It corresponds to 1225 * 128 outstanding write IO requests (if IO size is 128 KiB), 1226 * which fits with many RAID controllers' tagged queuing limits. 1227 * The lower 1 MiB limit is the historical upper limit for 1228 * hirunningspace. 1229 */ 1230 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf), 1231 16 * 1024 * 1024), 1024 * 1024); 1232 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf); 1233 1234 /* 1235 * Limit the amount of malloc memory since it is wired permanently into 1236 * the kernel space. Even though this is accounted for in the buffer 1237 * allocation, we don't want the malloced region to grow uncontrolled. 1238 * The malloc scheme improves memory utilization significantly on 1239 * average (small) directories. 1240 */ 1241 maxbufmallocspace = hibufspace / 20; 1242 1243 /* 1244 * Reduce the chance of a deadlock occurring by limiting the number 1245 * of delayed-write dirty buffers we allow to stack up. 1246 */ 1247 hidirtybuffers = nbuf / 4 + 20; 1248 dirtybufthresh = hidirtybuffers * 9 / 10; 1249 /* 1250 * To support extreme low-memory systems, make sure hidirtybuffers 1251 * cannot eat up all available buffer space. This occurs when our 1252 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our 1253 * buffer space assuming BKVASIZE'd buffers. 1254 */ 1255 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 1256 hidirtybuffers >>= 1; 1257 } 1258 lodirtybuffers = hidirtybuffers / 2; 1259 1260 /* 1261 * lofreebuffers should be sufficient to avoid stalling waiting on 1262 * buf headers under heavy utilization. The bufs in per-cpu caches 1263 * are counted as free but will be unavailable to threads executing 1264 * on other cpus. 1265 * 1266 * hifreebuffers is the free target for the bufspace daemon. This 1267 * should be set appropriately to limit work per-iteration. 1268 */ 1269 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus); 1270 hifreebuffers = (3 * lofreebuffers) / 2; 1271 numfreebuffers = nbuf; 1272 1273 /* Setup the kva and free list allocators. */ 1274 vmem_set_reclaim(buffer_arena, bufkva_reclaim); 1275 buf_zone = uma_zcache_create("buf free cache", 1276 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf), 1277 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0); 1278 1279 /* 1280 * Size the clean queue according to the amount of buffer space. 1281 * One queue per-256mb up to the max. More queues gives better 1282 * concurrency but less accurate LRU. 1283 */ 1284 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS); 1285 for (i = 0 ; i < buf_domains; i++) { 1286 struct bufdomain *bd; 1287 1288 bd = &bdomain[i]; 1289 bd_init(bd); 1290 bd->bd_freebuffers = nbuf / buf_domains; 1291 bd->bd_hifreebuffers = hifreebuffers / buf_domains; 1292 bd->bd_lofreebuffers = lofreebuffers / buf_domains; 1293 bd->bd_bufspace = 0; 1294 bd->bd_maxbufspace = maxbufspace / buf_domains; 1295 bd->bd_hibufspace = hibufspace / buf_domains; 1296 bd->bd_lobufspace = lobufspace / buf_domains; 1297 bd->bd_bufspacethresh = bufspacethresh / buf_domains; 1298 bd->bd_numdirtybuffers = 0; 1299 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains; 1300 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains; 1301 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains; 1302 /* Don't allow more than 2% of bufs in the per-cpu caches. */ 1303 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus; 1304 } 1305 getnewbufcalls = counter_u64_alloc(M_WAITOK); 1306 getnewbufrestarts = counter_u64_alloc(M_WAITOK); 1307 mappingrestarts = counter_u64_alloc(M_WAITOK); 1308 numbufallocfails = counter_u64_alloc(M_WAITOK); 1309 notbufdflushes = counter_u64_alloc(M_WAITOK); 1310 buffreekvacnt = counter_u64_alloc(M_WAITOK); 1311 bufdefragcnt = counter_u64_alloc(M_WAITOK); 1312 bufkvaspace = counter_u64_alloc(M_WAITOK); 1313 } 1314 1315 #ifdef INVARIANTS 1316 static inline void 1317 vfs_buf_check_mapped(struct buf *bp) 1318 { 1319 1320 KASSERT(bp->b_kvabase != unmapped_buf, 1321 ("mapped buf: b_kvabase was not updated %p", bp)); 1322 KASSERT(bp->b_data != unmapped_buf, 1323 ("mapped buf: b_data was not updated %p", bp)); 1324 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf + 1325 maxphys, ("b_data + b_offset unmapped %p", bp)); 1326 } 1327 1328 static inline void 1329 vfs_buf_check_unmapped(struct buf *bp) 1330 { 1331 1332 KASSERT(bp->b_data == unmapped_buf, 1333 ("unmapped buf: corrupted b_data %p", bp)); 1334 } 1335 1336 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 1337 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 1338 #else 1339 #define BUF_CHECK_MAPPED(bp) do {} while (0) 1340 #define BUF_CHECK_UNMAPPED(bp) do {} while (0) 1341 #endif 1342 1343 static int 1344 isbufbusy(struct buf *bp) 1345 { 1346 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) || 1347 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI)) 1348 return (1); 1349 return (0); 1350 } 1351 1352 /* 1353 * Shutdown the system cleanly to prepare for reboot, halt, or power off. 1354 */ 1355 void 1356 bufshutdown(int show_busybufs) 1357 { 1358 static int first_buf_printf = 1; 1359 struct buf *bp; 1360 int i, iter, nbusy, pbusy; 1361 #ifndef PREEMPTION 1362 int subiter; 1363 #endif 1364 1365 /* 1366 * Sync filesystems for shutdown 1367 */ 1368 wdog_kern_pat(WD_LASTVAL); 1369 kern_sync(curthread); 1370 1371 /* 1372 * With soft updates, some buffers that are 1373 * written will be remarked as dirty until other 1374 * buffers are written. 1375 */ 1376 for (iter = pbusy = 0; iter < 20; iter++) { 1377 nbusy = 0; 1378 for (i = nbuf - 1; i >= 0; i--) { 1379 bp = nbufp(i); 1380 if (isbufbusy(bp)) 1381 nbusy++; 1382 } 1383 if (nbusy == 0) { 1384 if (first_buf_printf) 1385 printf("All buffers synced."); 1386 break; 1387 } 1388 if (first_buf_printf) { 1389 printf("Syncing disks, buffers remaining... "); 1390 first_buf_printf = 0; 1391 } 1392 printf("%d ", nbusy); 1393 if (nbusy < pbusy) 1394 iter = 0; 1395 pbusy = nbusy; 1396 1397 wdog_kern_pat(WD_LASTVAL); 1398 kern_sync(curthread); 1399 1400 #ifdef PREEMPTION 1401 /* 1402 * Spin for a while to allow interrupt threads to run. 1403 */ 1404 DELAY(50000 * iter); 1405 #else 1406 /* 1407 * Context switch several times to allow interrupt 1408 * threads to run. 1409 */ 1410 for (subiter = 0; subiter < 50 * iter; subiter++) { 1411 thread_lock(curthread); 1412 mi_switch(SW_VOL); 1413 DELAY(1000); 1414 } 1415 #endif 1416 } 1417 printf("\n"); 1418 /* 1419 * Count only busy local buffers to prevent forcing 1420 * a fsck if we're just a client of a wedged NFS server 1421 */ 1422 nbusy = 0; 1423 for (i = nbuf - 1; i >= 0; i--) { 1424 bp = nbufp(i); 1425 if (isbufbusy(bp)) { 1426 #if 0 1427 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */ 1428 if (bp->b_dev == NULL) { 1429 TAILQ_REMOVE(&mountlist, 1430 bp->b_vp->v_mount, mnt_list); 1431 continue; 1432 } 1433 #endif 1434 nbusy++; 1435 if (show_busybufs > 0) { 1436 printf( 1437 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:", 1438 nbusy, bp, bp->b_vp, bp->b_flags, 1439 (intmax_t)bp->b_blkno, 1440 (intmax_t)bp->b_lblkno); 1441 BUF_LOCKPRINTINFO(bp); 1442 if (show_busybufs > 1) 1443 vn_printf(bp->b_vp, 1444 "vnode content: "); 1445 } 1446 } 1447 } 1448 if (nbusy) { 1449 /* 1450 * Failed to sync all blocks. Indicate this and don't 1451 * unmount filesystems (thus forcing an fsck on reboot). 1452 */ 1453 printf("Giving up on %d buffers\n", nbusy); 1454 DELAY(5000000); /* 5 seconds */ 1455 swapoff_all(); 1456 } else { 1457 if (!first_buf_printf) 1458 printf("Final sync complete\n"); 1459 1460 /* 1461 * Unmount filesystems. Swapoff before unmount, 1462 * because file-backed swap is non-operational after unmount 1463 * of the underlying filesystem. 1464 */ 1465 if (!KERNEL_PANICKED()) { 1466 swapoff_all(); 1467 vfs_unmountall(); 1468 } 1469 } 1470 DELAY(100000); /* wait for console output to finish */ 1471 } 1472 1473 static void 1474 bpmap_qenter(struct buf *bp) 1475 { 1476 1477 BUF_CHECK_MAPPED(bp); 1478 1479 /* 1480 * bp->b_data is relative to bp->b_offset, but 1481 * bp->b_offset may be offset into the first page. 1482 */ 1483 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 1484 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 1485 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 1486 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 1487 } 1488 1489 static inline struct bufdomain * 1490 bufdomain(struct buf *bp) 1491 { 1492 1493 return (&bdomain[bp->b_domain]); 1494 } 1495 1496 static struct bufqueue * 1497 bufqueue(struct buf *bp) 1498 { 1499 1500 switch (bp->b_qindex) { 1501 case QUEUE_NONE: 1502 /* FALLTHROUGH */ 1503 case QUEUE_SENTINEL: 1504 return (NULL); 1505 case QUEUE_EMPTY: 1506 return (&bqempty); 1507 case QUEUE_DIRTY: 1508 return (&bufdomain(bp)->bd_dirtyq); 1509 case QUEUE_CLEAN: 1510 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]); 1511 default: 1512 break; 1513 } 1514 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex); 1515 } 1516 1517 /* 1518 * Return the locked bufqueue that bp is a member of. 1519 */ 1520 static struct bufqueue * 1521 bufqueue_acquire(struct buf *bp) 1522 { 1523 struct bufqueue *bq, *nbq; 1524 1525 /* 1526 * bp can be pushed from a per-cpu queue to the 1527 * cleanq while we're waiting on the lock. Retry 1528 * if the queues don't match. 1529 */ 1530 bq = bufqueue(bp); 1531 BQ_LOCK(bq); 1532 for (;;) { 1533 nbq = bufqueue(bp); 1534 if (bq == nbq) 1535 break; 1536 BQ_UNLOCK(bq); 1537 BQ_LOCK(nbq); 1538 bq = nbq; 1539 } 1540 return (bq); 1541 } 1542 1543 /* 1544 * binsfree: 1545 * 1546 * Insert the buffer into the appropriate free list. Requires a 1547 * locked buffer on entry and buffer is unlocked before return. 1548 */ 1549 static void 1550 binsfree(struct buf *bp, int qindex) 1551 { 1552 struct bufdomain *bd; 1553 struct bufqueue *bq; 1554 1555 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY, 1556 ("binsfree: Invalid qindex %d", qindex)); 1557 BUF_ASSERT_XLOCKED(bp); 1558 1559 /* 1560 * Handle delayed bremfree() processing. 1561 */ 1562 if (bp->b_flags & B_REMFREE) { 1563 if (bp->b_qindex == qindex) { 1564 bp->b_flags |= B_REUSE; 1565 bp->b_flags &= ~B_REMFREE; 1566 BUF_UNLOCK(bp); 1567 return; 1568 } 1569 bq = bufqueue_acquire(bp); 1570 bq_remove(bq, bp); 1571 BQ_UNLOCK(bq); 1572 } 1573 bd = bufdomain(bp); 1574 if (qindex == QUEUE_CLEAN) { 1575 if (bd->bd_lim != 0) 1576 bq = &bd->bd_subq[PCPU_GET(cpuid)]; 1577 else 1578 bq = bd->bd_cleanq; 1579 } else 1580 bq = &bd->bd_dirtyq; 1581 bq_insert(bq, bp, true); 1582 } 1583 1584 /* 1585 * buf_free: 1586 * 1587 * Free a buffer to the buf zone once it no longer has valid contents. 1588 */ 1589 static void 1590 buf_free(struct buf *bp) 1591 { 1592 1593 if (bp->b_flags & B_REMFREE) 1594 bremfreef(bp); 1595 if (bp->b_vflags & BV_BKGRDINPROG) 1596 panic("losing buffer 1"); 1597 if (bp->b_rcred != NOCRED) { 1598 crfree(bp->b_rcred); 1599 bp->b_rcred = NOCRED; 1600 } 1601 if (bp->b_wcred != NOCRED) { 1602 crfree(bp->b_wcred); 1603 bp->b_wcred = NOCRED; 1604 } 1605 if (!LIST_EMPTY(&bp->b_dep)) 1606 buf_deallocate(bp); 1607 bufkva_free(bp); 1608 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1); 1609 MPASS((bp->b_flags & B_MAXPHYS) == 0); 1610 BUF_UNLOCK(bp); 1611 uma_zfree(buf_zone, bp); 1612 } 1613 1614 /* 1615 * buf_import: 1616 * 1617 * Import bufs into the uma cache from the buf list. The system still 1618 * expects a static array of bufs and much of the synchronization 1619 * around bufs assumes type stable storage. As a result, UMA is used 1620 * only as a per-cpu cache of bufs still maintained on a global list. 1621 */ 1622 static int 1623 buf_import(void *arg, void **store, int cnt, int domain, int flags) 1624 { 1625 struct buf *bp; 1626 int i; 1627 1628 BQ_LOCK(&bqempty); 1629 for (i = 0; i < cnt; i++) { 1630 bp = TAILQ_FIRST(&bqempty.bq_queue); 1631 if (bp == NULL) 1632 break; 1633 bq_remove(&bqempty, bp); 1634 store[i] = bp; 1635 } 1636 BQ_UNLOCK(&bqempty); 1637 1638 return (i); 1639 } 1640 1641 /* 1642 * buf_release: 1643 * 1644 * Release bufs from the uma cache back to the buffer queues. 1645 */ 1646 static void 1647 buf_release(void *arg, void **store, int cnt) 1648 { 1649 struct bufqueue *bq; 1650 struct buf *bp; 1651 int i; 1652 1653 bq = &bqempty; 1654 BQ_LOCK(bq); 1655 for (i = 0; i < cnt; i++) { 1656 bp = store[i]; 1657 /* Inline bq_insert() to batch locking. */ 1658 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1659 bp->b_flags &= ~(B_AGE | B_REUSE); 1660 bq->bq_len++; 1661 bp->b_qindex = bq->bq_index; 1662 } 1663 BQ_UNLOCK(bq); 1664 } 1665 1666 /* 1667 * buf_alloc: 1668 * 1669 * Allocate an empty buffer header. 1670 */ 1671 static struct buf * 1672 buf_alloc(struct bufdomain *bd) 1673 { 1674 struct buf *bp; 1675 int freebufs, error; 1676 1677 /* 1678 * We can only run out of bufs in the buf zone if the average buf 1679 * is less than BKVASIZE. In this case the actual wait/block will 1680 * come from buf_reycle() failing to flush one of these small bufs. 1681 */ 1682 bp = NULL; 1683 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1); 1684 if (freebufs > 0) 1685 bp = uma_zalloc(buf_zone, M_NOWAIT); 1686 if (bp == NULL) { 1687 atomic_add_int(&bd->bd_freebuffers, 1); 1688 bufspace_daemon_wakeup(bd); 1689 counter_u64_add(numbufallocfails, 1); 1690 return (NULL); 1691 } 1692 /* 1693 * Wake-up the bufspace daemon on transition below threshold. 1694 */ 1695 if (freebufs == bd->bd_lofreebuffers) 1696 bufspace_daemon_wakeup(bd); 1697 1698 error = BUF_LOCK(bp, LK_EXCLUSIVE, NULL); 1699 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp, 1700 error)); 1701 (void)error; 1702 1703 KASSERT(bp->b_vp == NULL, 1704 ("bp: %p still has vnode %p.", bp, bp->b_vp)); 1705 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0, 1706 ("invalid buffer %p flags %#x", bp, bp->b_flags)); 1707 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1708 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 1709 KASSERT(bp->b_npages == 0, 1710 ("bp: %p still has %d vm pages\n", bp, bp->b_npages)); 1711 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp)); 1712 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp)); 1713 MPASS((bp->b_flags & B_MAXPHYS) == 0); 1714 1715 bp->b_domain = BD_DOMAIN(bd); 1716 bp->b_flags = 0; 1717 bp->b_ioflags = 0; 1718 bp->b_xflags = 0; 1719 bp->b_vflags = 0; 1720 bp->b_vp = NULL; 1721 bp->b_blkno = bp->b_lblkno = 0; 1722 bp->b_offset = NOOFFSET; 1723 bp->b_iodone = 0; 1724 bp->b_error = 0; 1725 bp->b_resid = 0; 1726 bp->b_bcount = 0; 1727 bp->b_npages = 0; 1728 bp->b_dirtyoff = bp->b_dirtyend = 0; 1729 bp->b_bufobj = NULL; 1730 bp->b_data = bp->b_kvabase = unmapped_buf; 1731 bp->b_fsprivate1 = NULL; 1732 bp->b_fsprivate2 = NULL; 1733 bp->b_fsprivate3 = NULL; 1734 LIST_INIT(&bp->b_dep); 1735 1736 return (bp); 1737 } 1738 1739 /* 1740 * buf_recycle: 1741 * 1742 * Free a buffer from the given bufqueue. kva controls whether the 1743 * freed buf must own some kva resources. This is used for 1744 * defragmenting. 1745 */ 1746 static int 1747 buf_recycle(struct bufdomain *bd, bool kva) 1748 { 1749 struct bufqueue *bq; 1750 struct buf *bp, *nbp; 1751 1752 if (kva) 1753 counter_u64_add(bufdefragcnt, 1); 1754 nbp = NULL; 1755 bq = bd->bd_cleanq; 1756 BQ_LOCK(bq); 1757 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd), 1758 ("buf_recycle: Locks don't match")); 1759 nbp = TAILQ_FIRST(&bq->bq_queue); 1760 1761 /* 1762 * Run scan, possibly freeing data and/or kva mappings on the fly 1763 * depending. 1764 */ 1765 while ((bp = nbp) != NULL) { 1766 /* 1767 * Calculate next bp (we can only use it if we do not 1768 * release the bqlock). 1769 */ 1770 nbp = TAILQ_NEXT(bp, b_freelist); 1771 1772 /* 1773 * If we are defragging then we need a buffer with 1774 * some kva to reclaim. 1775 */ 1776 if (kva && bp->b_kvasize == 0) 1777 continue; 1778 1779 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1780 continue; 1781 1782 /* 1783 * Implement a second chance algorithm for frequently 1784 * accessed buffers. 1785 */ 1786 if ((bp->b_flags & B_REUSE) != 0) { 1787 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1788 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1789 bp->b_flags &= ~B_REUSE; 1790 BUF_UNLOCK(bp); 1791 continue; 1792 } 1793 1794 /* 1795 * Skip buffers with background writes in progress. 1796 */ 1797 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 1798 BUF_UNLOCK(bp); 1799 continue; 1800 } 1801 1802 KASSERT(bp->b_qindex == QUEUE_CLEAN, 1803 ("buf_recycle: inconsistent queue %d bp %p", 1804 bp->b_qindex, bp)); 1805 KASSERT(bp->b_domain == BD_DOMAIN(bd), 1806 ("getnewbuf: queue domain %d doesn't match request %d", 1807 bp->b_domain, (int)BD_DOMAIN(bd))); 1808 /* 1809 * NOTE: nbp is now entirely invalid. We can only restart 1810 * the scan from this point on. 1811 */ 1812 bq_remove(bq, bp); 1813 BQ_UNLOCK(bq); 1814 1815 /* 1816 * Requeue the background write buffer with error and 1817 * restart the scan. 1818 */ 1819 if ((bp->b_vflags & BV_BKGRDERR) != 0) { 1820 bqrelse(bp); 1821 BQ_LOCK(bq); 1822 nbp = TAILQ_FIRST(&bq->bq_queue); 1823 continue; 1824 } 1825 bp->b_flags |= B_INVAL; 1826 brelse(bp); 1827 return (0); 1828 } 1829 bd->bd_wanted = 1; 1830 BQ_UNLOCK(bq); 1831 1832 return (ENOBUFS); 1833 } 1834 1835 /* 1836 * bremfree: 1837 * 1838 * Mark the buffer for removal from the appropriate free list. 1839 * 1840 */ 1841 void 1842 bremfree(struct buf *bp) 1843 { 1844 1845 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1846 KASSERT((bp->b_flags & B_REMFREE) == 0, 1847 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1848 KASSERT(bp->b_qindex != QUEUE_NONE, 1849 ("bremfree: buffer %p not on a queue.", bp)); 1850 BUF_ASSERT_XLOCKED(bp); 1851 1852 bp->b_flags |= B_REMFREE; 1853 } 1854 1855 /* 1856 * bremfreef: 1857 * 1858 * Force an immediate removal from a free list. Used only in nfs when 1859 * it abuses the b_freelist pointer. 1860 */ 1861 void 1862 bremfreef(struct buf *bp) 1863 { 1864 struct bufqueue *bq; 1865 1866 bq = bufqueue_acquire(bp); 1867 bq_remove(bq, bp); 1868 BQ_UNLOCK(bq); 1869 } 1870 1871 static void 1872 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname) 1873 { 1874 1875 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF); 1876 TAILQ_INIT(&bq->bq_queue); 1877 bq->bq_len = 0; 1878 bq->bq_index = qindex; 1879 bq->bq_subqueue = subqueue; 1880 } 1881 1882 static void 1883 bd_init(struct bufdomain *bd) 1884 { 1885 int i; 1886 1887 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1]; 1888 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock"); 1889 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock"); 1890 for (i = 0; i <= mp_maxid; i++) 1891 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i, 1892 "bufq clean subqueue lock"); 1893 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF); 1894 } 1895 1896 /* 1897 * bq_remove: 1898 * 1899 * Removes a buffer from the free list, must be called with the 1900 * correct qlock held. 1901 */ 1902 static void 1903 bq_remove(struct bufqueue *bq, struct buf *bp) 1904 { 1905 1906 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X", 1907 bp, bp->b_vp, bp->b_flags); 1908 KASSERT(bp->b_qindex != QUEUE_NONE, 1909 ("bq_remove: buffer %p not on a queue.", bp)); 1910 KASSERT(bufqueue(bp) == bq, 1911 ("bq_remove: Remove buffer %p from wrong queue.", bp)); 1912 1913 BQ_ASSERT_LOCKED(bq); 1914 if (bp->b_qindex != QUEUE_EMPTY) { 1915 BUF_ASSERT_XLOCKED(bp); 1916 } 1917 KASSERT(bq->bq_len >= 1, 1918 ("queue %d underflow", bp->b_qindex)); 1919 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1920 bq->bq_len--; 1921 bp->b_qindex = QUEUE_NONE; 1922 bp->b_flags &= ~(B_REMFREE | B_REUSE); 1923 } 1924 1925 static void 1926 bd_flush(struct bufdomain *bd, struct bufqueue *bq) 1927 { 1928 struct buf *bp; 1929 1930 BQ_ASSERT_LOCKED(bq); 1931 if (bq != bd->bd_cleanq) { 1932 BD_LOCK(bd); 1933 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) { 1934 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1935 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp, 1936 b_freelist); 1937 bp->b_subqueue = bd->bd_cleanq->bq_subqueue; 1938 } 1939 bd->bd_cleanq->bq_len += bq->bq_len; 1940 bq->bq_len = 0; 1941 } 1942 if (bd->bd_wanted) { 1943 bd->bd_wanted = 0; 1944 wakeup(&bd->bd_wanted); 1945 } 1946 if (bq != bd->bd_cleanq) 1947 BD_UNLOCK(bd); 1948 } 1949 1950 static int 1951 bd_flushall(struct bufdomain *bd) 1952 { 1953 struct bufqueue *bq; 1954 int flushed; 1955 int i; 1956 1957 if (bd->bd_lim == 0) 1958 return (0); 1959 flushed = 0; 1960 for (i = 0; i <= mp_maxid; i++) { 1961 bq = &bd->bd_subq[i]; 1962 if (bq->bq_len == 0) 1963 continue; 1964 BQ_LOCK(bq); 1965 bd_flush(bd, bq); 1966 BQ_UNLOCK(bq); 1967 flushed++; 1968 } 1969 1970 return (flushed); 1971 } 1972 1973 static void 1974 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock) 1975 { 1976 struct bufdomain *bd; 1977 1978 if (bp->b_qindex != QUEUE_NONE) 1979 panic("bq_insert: free buffer %p onto another queue?", bp); 1980 1981 bd = bufdomain(bp); 1982 if (bp->b_flags & B_AGE) { 1983 /* Place this buf directly on the real queue. */ 1984 if (bq->bq_index == QUEUE_CLEAN) 1985 bq = bd->bd_cleanq; 1986 BQ_LOCK(bq); 1987 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist); 1988 } else { 1989 BQ_LOCK(bq); 1990 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1991 } 1992 bp->b_flags &= ~(B_AGE | B_REUSE); 1993 bq->bq_len++; 1994 bp->b_qindex = bq->bq_index; 1995 bp->b_subqueue = bq->bq_subqueue; 1996 1997 /* 1998 * Unlock before we notify so that we don't wakeup a waiter that 1999 * fails a trylock on the buf and sleeps again. 2000 */ 2001 if (unlock) 2002 BUF_UNLOCK(bp); 2003 2004 if (bp->b_qindex == QUEUE_CLEAN) { 2005 /* 2006 * Flush the per-cpu queue and notify any waiters. 2007 */ 2008 if (bd->bd_wanted || (bq != bd->bd_cleanq && 2009 bq->bq_len >= bd->bd_lim)) 2010 bd_flush(bd, bq); 2011 } 2012 BQ_UNLOCK(bq); 2013 } 2014 2015 /* 2016 * bufkva_free: 2017 * 2018 * Free the kva allocation for a buffer. 2019 * 2020 */ 2021 static void 2022 bufkva_free(struct buf *bp) 2023 { 2024 2025 #ifdef INVARIANTS 2026 if (bp->b_kvasize == 0) { 2027 KASSERT(bp->b_kvabase == unmapped_buf && 2028 bp->b_data == unmapped_buf, 2029 ("Leaked KVA space on %p", bp)); 2030 } else if (buf_mapped(bp)) 2031 BUF_CHECK_MAPPED(bp); 2032 else 2033 BUF_CHECK_UNMAPPED(bp); 2034 #endif 2035 if (bp->b_kvasize == 0) 2036 return; 2037 2038 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize); 2039 counter_u64_add(bufkvaspace, -bp->b_kvasize); 2040 counter_u64_add(buffreekvacnt, 1); 2041 bp->b_data = bp->b_kvabase = unmapped_buf; 2042 bp->b_kvasize = 0; 2043 } 2044 2045 /* 2046 * bufkva_alloc: 2047 * 2048 * Allocate the buffer KVA and set b_kvasize and b_kvabase. 2049 */ 2050 static int 2051 bufkva_alloc(struct buf *bp, int maxsize, int gbflags) 2052 { 2053 vm_offset_t addr; 2054 int error; 2055 2056 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0, 2057 ("Invalid gbflags 0x%x in %s", gbflags, __func__)); 2058 MPASS((bp->b_flags & B_MAXPHYS) == 0); 2059 KASSERT(maxsize <= maxbcachebuf, 2060 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf)); 2061 2062 bufkva_free(bp); 2063 2064 addr = 0; 2065 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr); 2066 if (error != 0) { 2067 /* 2068 * Buffer map is too fragmented. Request the caller 2069 * to defragment the map. 2070 */ 2071 return (error); 2072 } 2073 bp->b_kvabase = (caddr_t)addr; 2074 bp->b_kvasize = maxsize; 2075 counter_u64_add(bufkvaspace, bp->b_kvasize); 2076 if ((gbflags & GB_UNMAPPED) != 0) { 2077 bp->b_data = unmapped_buf; 2078 BUF_CHECK_UNMAPPED(bp); 2079 } else { 2080 bp->b_data = bp->b_kvabase; 2081 BUF_CHECK_MAPPED(bp); 2082 } 2083 return (0); 2084 } 2085 2086 /* 2087 * bufkva_reclaim: 2088 * 2089 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem 2090 * callback that fires to avoid returning failure. 2091 */ 2092 static void 2093 bufkva_reclaim(vmem_t *vmem, int flags) 2094 { 2095 bool done; 2096 int q; 2097 int i; 2098 2099 done = false; 2100 for (i = 0; i < 5; i++) { 2101 for (q = 0; q < buf_domains; q++) 2102 if (buf_recycle(&bdomain[q], true) != 0) 2103 done = true; 2104 if (done) 2105 break; 2106 } 2107 return; 2108 } 2109 2110 /* 2111 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 2112 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 2113 * the buffer is valid and we do not have to do anything. 2114 */ 2115 static void 2116 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt, 2117 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *)) 2118 { 2119 struct buf *rabp; 2120 struct thread *td; 2121 int i; 2122 2123 td = curthread; 2124 2125 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 2126 if (inmem(vp, *rablkno)) 2127 continue; 2128 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 2129 if ((rabp->b_flags & B_CACHE) != 0) { 2130 brelse(rabp); 2131 continue; 2132 } 2133 #ifdef RACCT 2134 if (racct_enable) { 2135 PROC_LOCK(curproc); 2136 racct_add_buf(curproc, rabp, 0); 2137 PROC_UNLOCK(curproc); 2138 } 2139 #endif /* RACCT */ 2140 td->td_ru.ru_inblock++; 2141 rabp->b_flags |= B_ASYNC; 2142 rabp->b_flags &= ~B_INVAL; 2143 if ((flags & GB_CKHASH) != 0) { 2144 rabp->b_flags |= B_CKHASH; 2145 rabp->b_ckhashcalc = ckhashfunc; 2146 } 2147 rabp->b_ioflags &= ~BIO_ERROR; 2148 rabp->b_iocmd = BIO_READ; 2149 if (rabp->b_rcred == NOCRED && cred != NOCRED) 2150 rabp->b_rcred = crhold(cred); 2151 vfs_busy_pages(rabp, 0); 2152 BUF_KERNPROC(rabp); 2153 rabp->b_iooffset = dbtob(rabp->b_blkno); 2154 bstrategy(rabp); 2155 } 2156 } 2157 2158 /* 2159 * Entry point for bread() and breadn() via #defines in sys/buf.h. 2160 * 2161 * Get a buffer with the specified data. Look in the cache first. We 2162 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 2163 * is set, the buffer is valid and we do not have to do anything, see 2164 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 2165 * 2166 * Always return a NULL buffer pointer (in bpp) when returning an error. 2167 * 2168 * The blkno parameter is the logical block being requested. Normally 2169 * the mapping of logical block number to disk block address is done 2170 * by calling VOP_BMAP(). However, if the mapping is already known, the 2171 * disk block address can be passed using the dblkno parameter. If the 2172 * disk block address is not known, then the same value should be passed 2173 * for blkno and dblkno. 2174 */ 2175 int 2176 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, 2177 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags, 2178 void (*ckhashfunc)(struct buf *), struct buf **bpp) 2179 { 2180 struct buf *bp; 2181 struct thread *td; 2182 int error, readwait, rv; 2183 2184 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 2185 td = curthread; 2186 /* 2187 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags 2188 * are specified. 2189 */ 2190 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp); 2191 if (error != 0) { 2192 *bpp = NULL; 2193 return (error); 2194 } 2195 KASSERT(blkno == bp->b_lblkno, 2196 ("getblkx returned buffer for blkno %jd instead of blkno %jd", 2197 (intmax_t)bp->b_lblkno, (intmax_t)blkno)); 2198 flags &= ~GB_NOSPARSE; 2199 *bpp = bp; 2200 2201 /* 2202 * If not found in cache, do some I/O 2203 */ 2204 readwait = 0; 2205 if ((bp->b_flags & B_CACHE) == 0) { 2206 #ifdef RACCT 2207 if (racct_enable) { 2208 PROC_LOCK(td->td_proc); 2209 racct_add_buf(td->td_proc, bp, 0); 2210 PROC_UNLOCK(td->td_proc); 2211 } 2212 #endif /* RACCT */ 2213 td->td_ru.ru_inblock++; 2214 bp->b_iocmd = BIO_READ; 2215 bp->b_flags &= ~B_INVAL; 2216 if ((flags & GB_CKHASH) != 0) { 2217 bp->b_flags |= B_CKHASH; 2218 bp->b_ckhashcalc = ckhashfunc; 2219 } 2220 if ((flags & GB_CVTENXIO) != 0) 2221 bp->b_xflags |= BX_CVTENXIO; 2222 bp->b_ioflags &= ~BIO_ERROR; 2223 if (bp->b_rcred == NOCRED && cred != NOCRED) 2224 bp->b_rcred = crhold(cred); 2225 vfs_busy_pages(bp, 0); 2226 bp->b_iooffset = dbtob(bp->b_blkno); 2227 bstrategy(bp); 2228 ++readwait; 2229 } 2230 2231 /* 2232 * Attempt to initiate asynchronous I/O on read-ahead blocks. 2233 */ 2234 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc); 2235 2236 rv = 0; 2237 if (readwait) { 2238 rv = bufwait(bp); 2239 if (rv != 0) { 2240 brelse(bp); 2241 *bpp = NULL; 2242 } 2243 } 2244 return (rv); 2245 } 2246 2247 /* 2248 * Write, release buffer on completion. (Done by iodone 2249 * if async). Do not bother writing anything if the buffer 2250 * is invalid. 2251 * 2252 * Note that we set B_CACHE here, indicating that buffer is 2253 * fully valid and thus cacheable. This is true even of NFS 2254 * now so we set it generally. This could be set either here 2255 * or in biodone() since the I/O is synchronous. We put it 2256 * here. 2257 */ 2258 int 2259 bufwrite(struct buf *bp) 2260 { 2261 int oldflags; 2262 struct vnode *vp; 2263 long space; 2264 int vp_md; 2265 2266 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2267 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) { 2268 bp->b_flags |= B_INVAL | B_RELBUF; 2269 bp->b_flags &= ~B_CACHE; 2270 brelse(bp); 2271 return (ENXIO); 2272 } 2273 if (bp->b_flags & B_INVAL) { 2274 brelse(bp); 2275 return (0); 2276 } 2277 2278 if (bp->b_flags & B_BARRIER) 2279 atomic_add_long(&barrierwrites, 1); 2280 2281 oldflags = bp->b_flags; 2282 2283 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 2284 ("FFS background buffer should not get here %p", bp)); 2285 2286 vp = bp->b_vp; 2287 if (vp) 2288 vp_md = vp->v_vflag & VV_MD; 2289 else 2290 vp_md = 0; 2291 2292 /* 2293 * Mark the buffer clean. Increment the bufobj write count 2294 * before bundirty() call, to prevent other thread from seeing 2295 * empty dirty list and zero counter for writes in progress, 2296 * falsely indicating that the bufobj is clean. 2297 */ 2298 bufobj_wref(bp->b_bufobj); 2299 bundirty(bp); 2300 2301 bp->b_flags &= ~B_DONE; 2302 bp->b_ioflags &= ~BIO_ERROR; 2303 bp->b_flags |= B_CACHE; 2304 bp->b_iocmd = BIO_WRITE; 2305 2306 vfs_busy_pages(bp, 1); 2307 2308 /* 2309 * Normal bwrites pipeline writes 2310 */ 2311 bp->b_runningbufspace = bp->b_bufsize; 2312 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); 2313 2314 #ifdef RACCT 2315 if (racct_enable) { 2316 PROC_LOCK(curproc); 2317 racct_add_buf(curproc, bp, 1); 2318 PROC_UNLOCK(curproc); 2319 } 2320 #endif /* RACCT */ 2321 curthread->td_ru.ru_oublock++; 2322 if (oldflags & B_ASYNC) 2323 BUF_KERNPROC(bp); 2324 bp->b_iooffset = dbtob(bp->b_blkno); 2325 buf_track(bp, __func__); 2326 bstrategy(bp); 2327 2328 if ((oldflags & B_ASYNC) == 0) { 2329 int rtval = bufwait(bp); 2330 brelse(bp); 2331 return (rtval); 2332 } else if (space > hirunningspace) { 2333 /* 2334 * don't allow the async write to saturate the I/O 2335 * system. We will not deadlock here because 2336 * we are blocking waiting for I/O that is already in-progress 2337 * to complete. We do not block here if it is the update 2338 * or syncer daemon trying to clean up as that can lead 2339 * to deadlock. 2340 */ 2341 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 2342 waitrunningbufspace(); 2343 } 2344 2345 return (0); 2346 } 2347 2348 void 2349 bufbdflush(struct bufobj *bo, struct buf *bp) 2350 { 2351 struct buf *nbp; 2352 struct bufdomain *bd; 2353 2354 bd = &bdomain[bo->bo_domain]; 2355 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) { 2356 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 2357 altbufferflushes++; 2358 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) { 2359 BO_LOCK(bo); 2360 /* 2361 * Try to find a buffer to flush. 2362 */ 2363 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 2364 if ((nbp->b_vflags & BV_BKGRDINPROG) || 2365 BUF_LOCK(nbp, 2366 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 2367 continue; 2368 if (bp == nbp) 2369 panic("bdwrite: found ourselves"); 2370 BO_UNLOCK(bo); 2371 /* Don't countdeps with the bo lock held. */ 2372 if (buf_countdeps(nbp, 0)) { 2373 BO_LOCK(bo); 2374 BUF_UNLOCK(nbp); 2375 continue; 2376 } 2377 if (nbp->b_flags & B_CLUSTEROK) { 2378 vfs_bio_awrite(nbp); 2379 } else { 2380 bremfree(nbp); 2381 bawrite(nbp); 2382 } 2383 dirtybufferflushes++; 2384 break; 2385 } 2386 if (nbp == NULL) 2387 BO_UNLOCK(bo); 2388 } 2389 } 2390 2391 /* 2392 * Delayed write. (Buffer is marked dirty). Do not bother writing 2393 * anything if the buffer is marked invalid. 2394 * 2395 * Note that since the buffer must be completely valid, we can safely 2396 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 2397 * biodone() in order to prevent getblk from writing the buffer 2398 * out synchronously. 2399 */ 2400 void 2401 bdwrite(struct buf *bp) 2402 { 2403 struct thread *td = curthread; 2404 struct vnode *vp; 2405 struct bufobj *bo; 2406 2407 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2408 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2409 KASSERT((bp->b_flags & B_BARRIER) == 0, 2410 ("Barrier request in delayed write %p", bp)); 2411 2412 if (bp->b_flags & B_INVAL) { 2413 brelse(bp); 2414 return; 2415 } 2416 2417 /* 2418 * If we have too many dirty buffers, don't create any more. 2419 * If we are wildly over our limit, then force a complete 2420 * cleanup. Otherwise, just keep the situation from getting 2421 * out of control. Note that we have to avoid a recursive 2422 * disaster and not try to clean up after our own cleanup! 2423 */ 2424 vp = bp->b_vp; 2425 bo = bp->b_bufobj; 2426 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 2427 td->td_pflags |= TDP_INBDFLUSH; 2428 BO_BDFLUSH(bo, bp); 2429 td->td_pflags &= ~TDP_INBDFLUSH; 2430 } else 2431 recursiveflushes++; 2432 2433 bdirty(bp); 2434 /* 2435 * Set B_CACHE, indicating that the buffer is fully valid. This is 2436 * true even of NFS now. 2437 */ 2438 bp->b_flags |= B_CACHE; 2439 2440 /* 2441 * This bmap keeps the system from needing to do the bmap later, 2442 * perhaps when the system is attempting to do a sync. Since it 2443 * is likely that the indirect block -- or whatever other datastructure 2444 * that the filesystem needs is still in memory now, it is a good 2445 * thing to do this. Note also, that if the pageout daemon is 2446 * requesting a sync -- there might not be enough memory to do 2447 * the bmap then... So, this is important to do. 2448 */ 2449 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 2450 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 2451 } 2452 2453 buf_track(bp, __func__); 2454 2455 /* 2456 * Set the *dirty* buffer range based upon the VM system dirty 2457 * pages. 2458 * 2459 * Mark the buffer pages as clean. We need to do this here to 2460 * satisfy the vnode_pager and the pageout daemon, so that it 2461 * thinks that the pages have been "cleaned". Note that since 2462 * the pages are in a delayed write buffer -- the VFS layer 2463 * "will" see that the pages get written out on the next sync, 2464 * or perhaps the cluster will be completed. 2465 */ 2466 vfs_clean_pages_dirty_buf(bp); 2467 bqrelse(bp); 2468 2469 /* 2470 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 2471 * due to the softdep code. 2472 */ 2473 } 2474 2475 /* 2476 * bdirty: 2477 * 2478 * Turn buffer into delayed write request. We must clear BIO_READ and 2479 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 2480 * itself to properly update it in the dirty/clean lists. We mark it 2481 * B_DONE to ensure that any asynchronization of the buffer properly 2482 * clears B_DONE ( else a panic will occur later ). 2483 * 2484 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 2485 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 2486 * should only be called if the buffer is known-good. 2487 * 2488 * Since the buffer is not on a queue, we do not update the numfreebuffers 2489 * count. 2490 * 2491 * The buffer must be on QUEUE_NONE. 2492 */ 2493 void 2494 bdirty(struct buf *bp) 2495 { 2496 2497 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 2498 bp, bp->b_vp, bp->b_flags); 2499 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2500 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2501 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2502 bp->b_flags &= ~(B_RELBUF); 2503 bp->b_iocmd = BIO_WRITE; 2504 2505 if ((bp->b_flags & B_DELWRI) == 0) { 2506 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 2507 reassignbuf(bp); 2508 bdirtyadd(bp); 2509 } 2510 } 2511 2512 /* 2513 * bundirty: 2514 * 2515 * Clear B_DELWRI for buffer. 2516 * 2517 * Since the buffer is not on a queue, we do not update the numfreebuffers 2518 * count. 2519 * 2520 * The buffer must be on QUEUE_NONE. 2521 */ 2522 2523 void 2524 bundirty(struct buf *bp) 2525 { 2526 2527 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2528 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2529 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2530 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2531 2532 if (bp->b_flags & B_DELWRI) { 2533 bp->b_flags &= ~B_DELWRI; 2534 reassignbuf(bp); 2535 bdirtysub(bp); 2536 } 2537 /* 2538 * Since it is now being written, we can clear its deferred write flag. 2539 */ 2540 bp->b_flags &= ~B_DEFERRED; 2541 } 2542 2543 /* 2544 * bawrite: 2545 * 2546 * Asynchronous write. Start output on a buffer, but do not wait for 2547 * it to complete. The buffer is released when the output completes. 2548 * 2549 * bwrite() ( or the VOP routine anyway ) is responsible for handling 2550 * B_INVAL buffers. Not us. 2551 */ 2552 void 2553 bawrite(struct buf *bp) 2554 { 2555 2556 bp->b_flags |= B_ASYNC; 2557 (void) bwrite(bp); 2558 } 2559 2560 /* 2561 * babarrierwrite: 2562 * 2563 * Asynchronous barrier write. Start output on a buffer, but do not 2564 * wait for it to complete. Place a write barrier after this write so 2565 * that this buffer and all buffers written before it are committed to 2566 * the disk before any buffers written after this write are committed 2567 * to the disk. The buffer is released when the output completes. 2568 */ 2569 void 2570 babarrierwrite(struct buf *bp) 2571 { 2572 2573 bp->b_flags |= B_ASYNC | B_BARRIER; 2574 (void) bwrite(bp); 2575 } 2576 2577 /* 2578 * bbarrierwrite: 2579 * 2580 * Synchronous barrier write. Start output on a buffer and wait for 2581 * it to complete. Place a write barrier after this write so that 2582 * this buffer and all buffers written before it are committed to 2583 * the disk before any buffers written after this write are committed 2584 * to the disk. The buffer is released when the output completes. 2585 */ 2586 int 2587 bbarrierwrite(struct buf *bp) 2588 { 2589 2590 bp->b_flags |= B_BARRIER; 2591 return (bwrite(bp)); 2592 } 2593 2594 /* 2595 * bwillwrite: 2596 * 2597 * Called prior to the locking of any vnodes when we are expecting to 2598 * write. We do not want to starve the buffer cache with too many 2599 * dirty buffers so we block here. By blocking prior to the locking 2600 * of any vnodes we attempt to avoid the situation where a locked vnode 2601 * prevents the various system daemons from flushing related buffers. 2602 */ 2603 void 2604 bwillwrite(void) 2605 { 2606 2607 if (buf_dirty_count_severe()) { 2608 mtx_lock(&bdirtylock); 2609 while (buf_dirty_count_severe()) { 2610 bdirtywait = 1; 2611 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), 2612 "flswai", 0); 2613 } 2614 mtx_unlock(&bdirtylock); 2615 } 2616 } 2617 2618 /* 2619 * Return true if we have too many dirty buffers. 2620 */ 2621 int 2622 buf_dirty_count_severe(void) 2623 { 2624 2625 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty)); 2626 } 2627 2628 /* 2629 * brelse: 2630 * 2631 * Release a busy buffer and, if requested, free its resources. The 2632 * buffer will be stashed in the appropriate bufqueue[] allowing it 2633 * to be accessed later as a cache entity or reused for other purposes. 2634 */ 2635 void 2636 brelse(struct buf *bp) 2637 { 2638 struct mount *v_mnt; 2639 int qindex; 2640 2641 /* 2642 * Many functions erroneously call brelse with a NULL bp under rare 2643 * error conditions. Simply return when called with a NULL bp. 2644 */ 2645 if (bp == NULL) 2646 return; 2647 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 2648 bp, bp->b_vp, bp->b_flags); 2649 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2650 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2651 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0, 2652 ("brelse: non-VMIO buffer marked NOREUSE")); 2653 2654 if (BUF_LOCKRECURSED(bp)) { 2655 /* 2656 * Do not process, in particular, do not handle the 2657 * B_INVAL/B_RELBUF and do not release to free list. 2658 */ 2659 BUF_UNLOCK(bp); 2660 return; 2661 } 2662 2663 if (bp->b_flags & B_MANAGED) { 2664 bqrelse(bp); 2665 return; 2666 } 2667 2668 if (LIST_EMPTY(&bp->b_dep)) { 2669 bp->b_flags &= ~B_IOSTARTED; 2670 } else { 2671 KASSERT((bp->b_flags & B_IOSTARTED) == 0, 2672 ("brelse: SU io not finished bp %p", bp)); 2673 } 2674 2675 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { 2676 BO_LOCK(bp->b_bufobj); 2677 bp->b_vflags &= ~BV_BKGRDERR; 2678 BO_UNLOCK(bp->b_bufobj); 2679 bdirty(bp); 2680 } 2681 2682 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 2683 (bp->b_flags & B_INVALONERR)) { 2684 /* 2685 * Forced invalidation of dirty buffer contents, to be used 2686 * after a failed write in the rare case that the loss of the 2687 * contents is acceptable. The buffer is invalidated and 2688 * freed. 2689 */ 2690 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE; 2691 bp->b_flags &= ~(B_ASYNC | B_CACHE); 2692 } 2693 2694 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 2695 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) && 2696 !(bp->b_flags & B_INVAL)) { 2697 /* 2698 * Failed write, redirty. All errors except ENXIO (which 2699 * means the device is gone) are treated as being 2700 * transient. 2701 * 2702 * XXX Treating EIO as transient is not correct; the 2703 * contract with the local storage device drivers is that 2704 * they will only return EIO once the I/O is no longer 2705 * retriable. Network I/O also respects this through the 2706 * guarantees of TCP and/or the internal retries of NFS. 2707 * ENOMEM might be transient, but we also have no way of 2708 * knowing when its ok to retry/reschedule. In general, 2709 * this entire case should be made obsolete through better 2710 * error handling/recovery and resource scheduling. 2711 * 2712 * Do this also for buffers that failed with ENXIO, but have 2713 * non-empty dependencies - the soft updates code might need 2714 * to access the buffer to untangle them. 2715 * 2716 * Must clear BIO_ERROR to prevent pages from being scrapped. 2717 */ 2718 bp->b_ioflags &= ~BIO_ERROR; 2719 bdirty(bp); 2720 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 2721 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 2722 /* 2723 * Either a failed read I/O, or we were asked to free or not 2724 * cache the buffer, or we failed to write to a device that's 2725 * no longer present. 2726 */ 2727 bp->b_flags |= B_INVAL; 2728 if (!LIST_EMPTY(&bp->b_dep)) 2729 buf_deallocate(bp); 2730 if (bp->b_flags & B_DELWRI) 2731 bdirtysub(bp); 2732 bp->b_flags &= ~(B_DELWRI | B_CACHE); 2733 if ((bp->b_flags & B_VMIO) == 0) { 2734 allocbuf(bp, 0); 2735 if (bp->b_vp) 2736 brelvp(bp); 2737 } 2738 } 2739 2740 /* 2741 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate() 2742 * is called with B_DELWRI set, the underlying pages may wind up 2743 * getting freed causing a previous write (bdwrite()) to get 'lost' 2744 * because pages associated with a B_DELWRI bp are marked clean. 2745 * 2746 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even 2747 * if B_DELWRI is set. 2748 */ 2749 if (bp->b_flags & B_DELWRI) 2750 bp->b_flags &= ~B_RELBUF; 2751 2752 /* 2753 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 2754 * constituted, not even NFS buffers now. Two flags effect this. If 2755 * B_INVAL, the struct buf is invalidated but the VM object is kept 2756 * around ( i.e. so it is trivial to reconstitute the buffer later ). 2757 * 2758 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 2759 * invalidated. BIO_ERROR cannot be set for a failed write unless the 2760 * buffer is also B_INVAL because it hits the re-dirtying code above. 2761 * 2762 * Normally we can do this whether a buffer is B_DELWRI or not. If 2763 * the buffer is an NFS buffer, it is tracking piecemeal writes or 2764 * the commit state and we cannot afford to lose the buffer. If the 2765 * buffer has a background write in progress, we need to keep it 2766 * around to prevent it from being reconstituted and starting a second 2767 * background write. 2768 */ 2769 2770 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL; 2771 2772 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE || 2773 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) && 2774 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 || 2775 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) { 2776 vfs_vmio_invalidate(bp); 2777 allocbuf(bp, 0); 2778 } 2779 2780 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 || 2781 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) { 2782 allocbuf(bp, 0); 2783 bp->b_flags &= ~B_NOREUSE; 2784 if (bp->b_vp != NULL) 2785 brelvp(bp); 2786 } 2787 2788 /* 2789 * If the buffer has junk contents signal it and eventually 2790 * clean up B_DELWRI and diassociate the vnode so that gbincore() 2791 * doesn't find it. 2792 */ 2793 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 2794 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 2795 bp->b_flags |= B_INVAL; 2796 if (bp->b_flags & B_INVAL) { 2797 if (bp->b_flags & B_DELWRI) 2798 bundirty(bp); 2799 if (bp->b_vp) 2800 brelvp(bp); 2801 } 2802 2803 buf_track(bp, __func__); 2804 2805 /* buffers with no memory */ 2806 if (bp->b_bufsize == 0) { 2807 buf_free(bp); 2808 return; 2809 } 2810 /* buffers with junk contents */ 2811 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 2812 (bp->b_ioflags & BIO_ERROR)) { 2813 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 2814 if (bp->b_vflags & BV_BKGRDINPROG) 2815 panic("losing buffer 2"); 2816 qindex = QUEUE_CLEAN; 2817 bp->b_flags |= B_AGE; 2818 /* remaining buffers */ 2819 } else if (bp->b_flags & B_DELWRI) 2820 qindex = QUEUE_DIRTY; 2821 else 2822 qindex = QUEUE_CLEAN; 2823 2824 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 2825 panic("brelse: not dirty"); 2826 2827 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT); 2828 bp->b_xflags &= ~(BX_CVTENXIO); 2829 /* binsfree unlocks bp. */ 2830 binsfree(bp, qindex); 2831 } 2832 2833 /* 2834 * Release a buffer back to the appropriate queue but do not try to free 2835 * it. The buffer is expected to be used again soon. 2836 * 2837 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 2838 * biodone() to requeue an async I/O on completion. It is also used when 2839 * known good buffers need to be requeued but we think we may need the data 2840 * again soon. 2841 * 2842 * XXX we should be able to leave the B_RELBUF hint set on completion. 2843 */ 2844 void 2845 bqrelse(struct buf *bp) 2846 { 2847 int qindex; 2848 2849 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2850 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2851 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2852 2853 qindex = QUEUE_NONE; 2854 if (BUF_LOCKRECURSED(bp)) { 2855 /* do not release to free list */ 2856 BUF_UNLOCK(bp); 2857 return; 2858 } 2859 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 2860 bp->b_xflags &= ~(BX_CVTENXIO); 2861 2862 if (LIST_EMPTY(&bp->b_dep)) { 2863 bp->b_flags &= ~B_IOSTARTED; 2864 } else { 2865 KASSERT((bp->b_flags & B_IOSTARTED) == 0, 2866 ("bqrelse: SU io not finished bp %p", bp)); 2867 } 2868 2869 if (bp->b_flags & B_MANAGED) { 2870 if (bp->b_flags & B_REMFREE) 2871 bremfreef(bp); 2872 goto out; 2873 } 2874 2875 /* buffers with stale but valid contents */ 2876 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | 2877 BV_BKGRDERR)) == BV_BKGRDERR) { 2878 BO_LOCK(bp->b_bufobj); 2879 bp->b_vflags &= ~BV_BKGRDERR; 2880 BO_UNLOCK(bp->b_bufobj); 2881 qindex = QUEUE_DIRTY; 2882 } else { 2883 if ((bp->b_flags & B_DELWRI) == 0 && 2884 (bp->b_xflags & BX_VNDIRTY)) 2885 panic("bqrelse: not dirty"); 2886 if ((bp->b_flags & B_NOREUSE) != 0) { 2887 brelse(bp); 2888 return; 2889 } 2890 qindex = QUEUE_CLEAN; 2891 } 2892 buf_track(bp, __func__); 2893 /* binsfree unlocks bp. */ 2894 binsfree(bp, qindex); 2895 return; 2896 2897 out: 2898 buf_track(bp, __func__); 2899 /* unlock */ 2900 BUF_UNLOCK(bp); 2901 } 2902 2903 /* 2904 * Complete I/O to a VMIO backed page. Validate the pages as appropriate, 2905 * restore bogus pages. 2906 */ 2907 static void 2908 vfs_vmio_iodone(struct buf *bp) 2909 { 2910 vm_ooffset_t foff; 2911 vm_page_t m; 2912 vm_object_t obj; 2913 struct vnode *vp __unused; 2914 int i, iosize, resid; 2915 bool bogus; 2916 2917 obj = bp->b_bufobj->bo_object; 2918 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages, 2919 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)", 2920 blockcount_read(&obj->paging_in_progress), bp->b_npages)); 2921 2922 vp = bp->b_vp; 2923 VNPASS(vp->v_holdcnt > 0, vp); 2924 VNPASS(vp->v_object != NULL, vp); 2925 2926 foff = bp->b_offset; 2927 KASSERT(bp->b_offset != NOOFFSET, 2928 ("vfs_vmio_iodone: bp %p has no buffer offset", bp)); 2929 2930 bogus = false; 2931 iosize = bp->b_bcount - bp->b_resid; 2932 for (i = 0; i < bp->b_npages; i++) { 2933 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2934 if (resid > iosize) 2935 resid = iosize; 2936 2937 /* 2938 * cleanup bogus pages, restoring the originals 2939 */ 2940 m = bp->b_pages[i]; 2941 if (m == bogus_page) { 2942 bogus = true; 2943 m = vm_page_relookup(obj, OFF_TO_IDX(foff)); 2944 if (m == NULL) 2945 panic("biodone: page disappeared!"); 2946 bp->b_pages[i] = m; 2947 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) { 2948 /* 2949 * In the write case, the valid and clean bits are 2950 * already changed correctly ( see bdwrite() ), so we 2951 * only need to do this here in the read case. 2952 */ 2953 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK, 2954 resid)) == 0, ("vfs_vmio_iodone: page %p " 2955 "has unexpected dirty bits", m)); 2956 vfs_page_set_valid(bp, foff, m); 2957 } 2958 KASSERT(OFF_TO_IDX(foff) == m->pindex, 2959 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch", 2960 (intmax_t)foff, (uintmax_t)m->pindex)); 2961 2962 vm_page_sunbusy(m); 2963 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2964 iosize -= resid; 2965 } 2966 vm_object_pip_wakeupn(obj, bp->b_npages); 2967 if (bogus && buf_mapped(bp)) { 2968 BUF_CHECK_MAPPED(bp); 2969 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 2970 bp->b_pages, bp->b_npages); 2971 } 2972 } 2973 2974 /* 2975 * Perform page invalidation when a buffer is released. The fully invalid 2976 * pages will be reclaimed later in vfs_vmio_truncate(). 2977 */ 2978 static void 2979 vfs_vmio_invalidate(struct buf *bp) 2980 { 2981 vm_object_t obj; 2982 vm_page_t m; 2983 int flags, i, resid, poffset, presid; 2984 2985 if (buf_mapped(bp)) { 2986 BUF_CHECK_MAPPED(bp); 2987 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 2988 } else 2989 BUF_CHECK_UNMAPPED(bp); 2990 /* 2991 * Get the base offset and length of the buffer. Note that 2992 * in the VMIO case if the buffer block size is not 2993 * page-aligned then b_data pointer may not be page-aligned. 2994 * But our b_pages[] array *IS* page aligned. 2995 * 2996 * block sizes less then DEV_BSIZE (usually 512) are not 2997 * supported due to the page granularity bits (m->valid, 2998 * m->dirty, etc...). 2999 * 3000 * See man buf(9) for more information 3001 */ 3002 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; 3003 obj = bp->b_bufobj->bo_object; 3004 resid = bp->b_bufsize; 3005 poffset = bp->b_offset & PAGE_MASK; 3006 VM_OBJECT_WLOCK(obj); 3007 for (i = 0; i < bp->b_npages; i++) { 3008 m = bp->b_pages[i]; 3009 if (m == bogus_page) 3010 panic("vfs_vmio_invalidate: Unexpected bogus page."); 3011 bp->b_pages[i] = NULL; 3012 3013 presid = resid > (PAGE_SIZE - poffset) ? 3014 (PAGE_SIZE - poffset) : resid; 3015 KASSERT(presid >= 0, ("brelse: extra page")); 3016 vm_page_busy_acquire(m, VM_ALLOC_SBUSY); 3017 if (pmap_page_wired_mappings(m) == 0) 3018 vm_page_set_invalid(m, poffset, presid); 3019 vm_page_sunbusy(m); 3020 vm_page_release_locked(m, flags); 3021 resid -= presid; 3022 poffset = 0; 3023 } 3024 VM_OBJECT_WUNLOCK(obj); 3025 bp->b_npages = 0; 3026 } 3027 3028 /* 3029 * Page-granular truncation of an existing VMIO buffer. 3030 */ 3031 static void 3032 vfs_vmio_truncate(struct buf *bp, int desiredpages) 3033 { 3034 vm_object_t obj; 3035 vm_page_t m; 3036 int flags, i; 3037 3038 if (bp->b_npages == desiredpages) 3039 return; 3040 3041 if (buf_mapped(bp)) { 3042 BUF_CHECK_MAPPED(bp); 3043 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) + 3044 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages); 3045 } else 3046 BUF_CHECK_UNMAPPED(bp); 3047 3048 /* 3049 * The object lock is needed only if we will attempt to free pages. 3050 */ 3051 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; 3052 if ((bp->b_flags & B_DIRECT) != 0) { 3053 flags |= VPR_TRYFREE; 3054 obj = bp->b_bufobj->bo_object; 3055 VM_OBJECT_WLOCK(obj); 3056 } else { 3057 obj = NULL; 3058 } 3059 for (i = desiredpages; i < bp->b_npages; i++) { 3060 m = bp->b_pages[i]; 3061 KASSERT(m != bogus_page, ("allocbuf: bogus page found")); 3062 bp->b_pages[i] = NULL; 3063 if (obj != NULL) 3064 vm_page_release_locked(m, flags); 3065 else 3066 vm_page_release(m, flags); 3067 } 3068 if (obj != NULL) 3069 VM_OBJECT_WUNLOCK(obj); 3070 bp->b_npages = desiredpages; 3071 } 3072 3073 /* 3074 * Byte granular extension of VMIO buffers. 3075 */ 3076 static void 3077 vfs_vmio_extend(struct buf *bp, int desiredpages, int size) 3078 { 3079 /* 3080 * We are growing the buffer, possibly in a 3081 * byte-granular fashion. 3082 */ 3083 vm_object_t obj; 3084 vm_offset_t toff; 3085 vm_offset_t tinc; 3086 vm_page_t m; 3087 3088 /* 3089 * Step 1, bring in the VM pages from the object, allocating 3090 * them if necessary. We must clear B_CACHE if these pages 3091 * are not valid for the range covered by the buffer. 3092 */ 3093 obj = bp->b_bufobj->bo_object; 3094 if (bp->b_npages < desiredpages) { 3095 KASSERT(desiredpages <= atop(maxbcachebuf), 3096 ("vfs_vmio_extend past maxbcachebuf %p %d %u", 3097 bp, desiredpages, maxbcachebuf)); 3098 3099 /* 3100 * We must allocate system pages since blocking 3101 * here could interfere with paging I/O, no 3102 * matter which process we are. 3103 * 3104 * Only exclusive busy can be tested here. 3105 * Blocking on shared busy might lead to 3106 * deadlocks once allocbuf() is called after 3107 * pages are vfs_busy_pages(). 3108 */ 3109 (void)vm_page_grab_pages_unlocked(obj, 3110 OFF_TO_IDX(bp->b_offset) + bp->b_npages, 3111 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY | 3112 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED, 3113 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages); 3114 bp->b_npages = desiredpages; 3115 } 3116 3117 /* 3118 * Step 2. We've loaded the pages into the buffer, 3119 * we have to figure out if we can still have B_CACHE 3120 * set. Note that B_CACHE is set according to the 3121 * byte-granular range ( bcount and size ), not the 3122 * aligned range ( newbsize ). 3123 * 3124 * The VM test is against m->valid, which is DEV_BSIZE 3125 * aligned. Needless to say, the validity of the data 3126 * needs to also be DEV_BSIZE aligned. Note that this 3127 * fails with NFS if the server or some other client 3128 * extends the file's EOF. If our buffer is resized, 3129 * B_CACHE may remain set! XXX 3130 */ 3131 toff = bp->b_bcount; 3132 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3133 while ((bp->b_flags & B_CACHE) && toff < size) { 3134 vm_pindex_t pi; 3135 3136 if (tinc > (size - toff)) 3137 tinc = size - toff; 3138 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; 3139 m = bp->b_pages[pi]; 3140 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m); 3141 toff += tinc; 3142 tinc = PAGE_SIZE; 3143 } 3144 3145 /* 3146 * Step 3, fixup the KVA pmap. 3147 */ 3148 if (buf_mapped(bp)) 3149 bpmap_qenter(bp); 3150 else 3151 BUF_CHECK_UNMAPPED(bp); 3152 } 3153 3154 /* 3155 * Check to see if a block at a particular lbn is available for a clustered 3156 * write. 3157 */ 3158 static int 3159 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 3160 { 3161 struct buf *bpa; 3162 int match; 3163 3164 match = 0; 3165 3166 /* If the buf isn't in core skip it */ 3167 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 3168 return (0); 3169 3170 /* If the buf is busy we don't want to wait for it */ 3171 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 3172 return (0); 3173 3174 /* Only cluster with valid clusterable delayed write buffers */ 3175 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 3176 (B_DELWRI | B_CLUSTEROK)) 3177 goto done; 3178 3179 if (bpa->b_bufsize != size) 3180 goto done; 3181 3182 /* 3183 * Check to see if it is in the expected place on disk and that the 3184 * block has been mapped. 3185 */ 3186 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 3187 match = 1; 3188 done: 3189 BUF_UNLOCK(bpa); 3190 return (match); 3191 } 3192 3193 /* 3194 * vfs_bio_awrite: 3195 * 3196 * Implement clustered async writes for clearing out B_DELWRI buffers. 3197 * This is much better then the old way of writing only one buffer at 3198 * a time. Note that we may not be presented with the buffers in the 3199 * correct order, so we search for the cluster in both directions. 3200 */ 3201 int 3202 vfs_bio_awrite(struct buf *bp) 3203 { 3204 struct bufobj *bo; 3205 int i; 3206 int j; 3207 daddr_t lblkno = bp->b_lblkno; 3208 struct vnode *vp = bp->b_vp; 3209 int ncl; 3210 int nwritten; 3211 int size; 3212 int maxcl; 3213 int gbflags; 3214 3215 bo = &vp->v_bufobj; 3216 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0; 3217 /* 3218 * right now we support clustered writing only to regular files. If 3219 * we find a clusterable block we could be in the middle of a cluster 3220 * rather then at the beginning. 3221 */ 3222 if ((vp->v_type == VREG) && 3223 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 3224 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 3225 size = vp->v_mount->mnt_stat.f_iosize; 3226 maxcl = maxphys / size; 3227 3228 BO_RLOCK(bo); 3229 for (i = 1; i < maxcl; i++) 3230 if (vfs_bio_clcheck(vp, size, lblkno + i, 3231 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 3232 break; 3233 3234 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 3235 if (vfs_bio_clcheck(vp, size, lblkno - j, 3236 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 3237 break; 3238 BO_RUNLOCK(bo); 3239 --j; 3240 ncl = i + j; 3241 /* 3242 * this is a possible cluster write 3243 */ 3244 if (ncl != 1) { 3245 BUF_UNLOCK(bp); 3246 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 3247 gbflags); 3248 return (nwritten); 3249 } 3250 } 3251 bremfree(bp); 3252 bp->b_flags |= B_ASYNC; 3253 /* 3254 * default (old) behavior, writing out only one block 3255 * 3256 * XXX returns b_bufsize instead of b_bcount for nwritten? 3257 */ 3258 nwritten = bp->b_bufsize; 3259 (void) bwrite(bp); 3260 3261 return (nwritten); 3262 } 3263 3264 /* 3265 * getnewbuf_kva: 3266 * 3267 * Allocate KVA for an empty buf header according to gbflags. 3268 */ 3269 static int 3270 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize) 3271 { 3272 3273 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) { 3274 /* 3275 * In order to keep fragmentation sane we only allocate kva 3276 * in BKVASIZE chunks. XXX with vmem we can do page size. 3277 */ 3278 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 3279 3280 if (maxsize != bp->b_kvasize && 3281 bufkva_alloc(bp, maxsize, gbflags)) 3282 return (ENOSPC); 3283 } 3284 return (0); 3285 } 3286 3287 /* 3288 * getnewbuf: 3289 * 3290 * Find and initialize a new buffer header, freeing up existing buffers 3291 * in the bufqueues as necessary. The new buffer is returned locked. 3292 * 3293 * We block if: 3294 * We have insufficient buffer headers 3295 * We have insufficient buffer space 3296 * buffer_arena is too fragmented ( space reservation fails ) 3297 * If we have to flush dirty buffers ( but we try to avoid this ) 3298 * 3299 * The caller is responsible for releasing the reserved bufspace after 3300 * allocbuf() is called. 3301 */ 3302 static struct buf * 3303 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags) 3304 { 3305 struct bufdomain *bd; 3306 struct buf *bp; 3307 bool metadata, reserved; 3308 3309 bp = NULL; 3310 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3311 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3312 if (!unmapped_buf_allowed) 3313 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3314 3315 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 3316 vp->v_type == VCHR) 3317 metadata = true; 3318 else 3319 metadata = false; 3320 if (vp == NULL) 3321 bd = &bdomain[0]; 3322 else 3323 bd = &bdomain[vp->v_bufobj.bo_domain]; 3324 3325 counter_u64_add(getnewbufcalls, 1); 3326 reserved = false; 3327 do { 3328 if (reserved == false && 3329 bufspace_reserve(bd, maxsize, metadata) != 0) { 3330 counter_u64_add(getnewbufrestarts, 1); 3331 continue; 3332 } 3333 reserved = true; 3334 if ((bp = buf_alloc(bd)) == NULL) { 3335 counter_u64_add(getnewbufrestarts, 1); 3336 continue; 3337 } 3338 if (getnewbuf_kva(bp, gbflags, maxsize) == 0) 3339 return (bp); 3340 break; 3341 } while (buf_recycle(bd, false) == 0); 3342 3343 if (reserved) 3344 bufspace_release(bd, maxsize); 3345 if (bp != NULL) { 3346 bp->b_flags |= B_INVAL; 3347 brelse(bp); 3348 } 3349 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo); 3350 3351 return (NULL); 3352 } 3353 3354 /* 3355 * buf_daemon: 3356 * 3357 * buffer flushing daemon. Buffers are normally flushed by the 3358 * update daemon but if it cannot keep up this process starts to 3359 * take the load in an attempt to prevent getnewbuf() from blocking. 3360 */ 3361 static struct kproc_desc buf_kp = { 3362 "bufdaemon", 3363 buf_daemon, 3364 &bufdaemonproc 3365 }; 3366 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 3367 3368 static int 3369 buf_flush(struct vnode *vp, struct bufdomain *bd, int target) 3370 { 3371 int flushed; 3372 3373 flushed = flushbufqueues(vp, bd, target, 0); 3374 if (flushed == 0) { 3375 /* 3376 * Could not find any buffers without rollback 3377 * dependencies, so just write the first one 3378 * in the hopes of eventually making progress. 3379 */ 3380 if (vp != NULL && target > 2) 3381 target /= 2; 3382 flushbufqueues(vp, bd, target, 1); 3383 } 3384 return (flushed); 3385 } 3386 3387 static void 3388 buf_daemon() 3389 { 3390 struct bufdomain *bd; 3391 int speedupreq; 3392 int lodirty; 3393 int i; 3394 3395 /* 3396 * This process needs to be suspended prior to shutdown sync. 3397 */ 3398 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread, 3399 SHUTDOWN_PRI_LAST + 100); 3400 3401 /* 3402 * Start the buf clean daemons as children threads. 3403 */ 3404 for (i = 0 ; i < buf_domains; i++) { 3405 int error; 3406 3407 error = kthread_add((void (*)(void *))bufspace_daemon, 3408 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i); 3409 if (error) 3410 panic("error %d spawning bufspace daemon", error); 3411 } 3412 3413 /* 3414 * This process is allowed to take the buffer cache to the limit 3415 */ 3416 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 3417 mtx_lock(&bdlock); 3418 for (;;) { 3419 bd_request = 0; 3420 mtx_unlock(&bdlock); 3421 3422 kthread_suspend_check(); 3423 3424 /* 3425 * Save speedupreq for this pass and reset to capture new 3426 * requests. 3427 */ 3428 speedupreq = bd_speedupreq; 3429 bd_speedupreq = 0; 3430 3431 /* 3432 * Flush each domain sequentially according to its level and 3433 * the speedup request. 3434 */ 3435 for (i = 0; i < buf_domains; i++) { 3436 bd = &bdomain[i]; 3437 if (speedupreq) 3438 lodirty = bd->bd_numdirtybuffers / 2; 3439 else 3440 lodirty = bd->bd_lodirtybuffers; 3441 while (bd->bd_numdirtybuffers > lodirty) { 3442 if (buf_flush(NULL, bd, 3443 bd->bd_numdirtybuffers - lodirty) == 0) 3444 break; 3445 kern_yield(PRI_USER); 3446 } 3447 } 3448 3449 /* 3450 * Only clear bd_request if we have reached our low water 3451 * mark. The buf_daemon normally waits 1 second and 3452 * then incrementally flushes any dirty buffers that have 3453 * built up, within reason. 3454 * 3455 * If we were unable to hit our low water mark and couldn't 3456 * find any flushable buffers, we sleep for a short period 3457 * to avoid endless loops on unlockable buffers. 3458 */ 3459 mtx_lock(&bdlock); 3460 if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) { 3461 /* 3462 * We reached our low water mark, reset the 3463 * request and sleep until we are needed again. 3464 * The sleep is just so the suspend code works. 3465 */ 3466 bd_request = 0; 3467 /* 3468 * Do an extra wakeup in case dirty threshold 3469 * changed via sysctl and the explicit transition 3470 * out of shortfall was missed. 3471 */ 3472 bdirtywakeup(); 3473 if (runningbufspace <= lorunningspace) 3474 runningwakeup(); 3475 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 3476 } else { 3477 /* 3478 * We couldn't find any flushable dirty buffers but 3479 * still have too many dirty buffers, we 3480 * have to sleep and try again. (rare) 3481 */ 3482 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 3483 } 3484 } 3485 } 3486 3487 /* 3488 * flushbufqueues: 3489 * 3490 * Try to flush a buffer in the dirty queue. We must be careful to 3491 * free up B_INVAL buffers instead of write them, which NFS is 3492 * particularly sensitive to. 3493 */ 3494 static int flushwithdeps = 0; 3495 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS, 3496 &flushwithdeps, 0, 3497 "Number of buffers flushed with dependencies that require rollbacks"); 3498 3499 static int 3500 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target, 3501 int flushdeps) 3502 { 3503 struct bufqueue *bq; 3504 struct buf *sentinel; 3505 struct vnode *vp; 3506 struct mount *mp; 3507 struct buf *bp; 3508 int hasdeps; 3509 int flushed; 3510 int error; 3511 bool unlock; 3512 3513 flushed = 0; 3514 bq = &bd->bd_dirtyq; 3515 bp = NULL; 3516 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 3517 sentinel->b_qindex = QUEUE_SENTINEL; 3518 BQ_LOCK(bq); 3519 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist); 3520 BQ_UNLOCK(bq); 3521 while (flushed != target) { 3522 maybe_yield(); 3523 BQ_LOCK(bq); 3524 bp = TAILQ_NEXT(sentinel, b_freelist); 3525 if (bp != NULL) { 3526 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); 3527 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel, 3528 b_freelist); 3529 } else { 3530 BQ_UNLOCK(bq); 3531 break; 3532 } 3533 /* 3534 * Skip sentinels inserted by other invocations of the 3535 * flushbufqueues(), taking care to not reorder them. 3536 * 3537 * Only flush the buffers that belong to the 3538 * vnode locked by the curthread. 3539 */ 3540 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && 3541 bp->b_vp != lvp)) { 3542 BQ_UNLOCK(bq); 3543 continue; 3544 } 3545 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 3546 BQ_UNLOCK(bq); 3547 if (error != 0) 3548 continue; 3549 3550 /* 3551 * BKGRDINPROG can only be set with the buf and bufobj 3552 * locks both held. We tolerate a race to clear it here. 3553 */ 3554 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 3555 (bp->b_flags & B_DELWRI) == 0) { 3556 BUF_UNLOCK(bp); 3557 continue; 3558 } 3559 if (bp->b_flags & B_INVAL) { 3560 bremfreef(bp); 3561 brelse(bp); 3562 flushed++; 3563 continue; 3564 } 3565 3566 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 3567 if (flushdeps == 0) { 3568 BUF_UNLOCK(bp); 3569 continue; 3570 } 3571 hasdeps = 1; 3572 } else 3573 hasdeps = 0; 3574 /* 3575 * We must hold the lock on a vnode before writing 3576 * one of its buffers. Otherwise we may confuse, or 3577 * in the case of a snapshot vnode, deadlock the 3578 * system. 3579 * 3580 * The lock order here is the reverse of the normal 3581 * of vnode followed by buf lock. This is ok because 3582 * the NOWAIT will prevent deadlock. 3583 */ 3584 vp = bp->b_vp; 3585 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 3586 BUF_UNLOCK(bp); 3587 continue; 3588 } 3589 if (lvp == NULL) { 3590 unlock = true; 3591 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 3592 } else { 3593 ASSERT_VOP_LOCKED(vp, "getbuf"); 3594 unlock = false; 3595 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 3596 vn_lock(vp, LK_TRYUPGRADE); 3597 } 3598 if (error == 0) { 3599 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 3600 bp, bp->b_vp, bp->b_flags); 3601 if (curproc == bufdaemonproc) { 3602 vfs_bio_awrite(bp); 3603 } else { 3604 bremfree(bp); 3605 bwrite(bp); 3606 counter_u64_add(notbufdflushes, 1); 3607 } 3608 vn_finished_write(mp); 3609 if (unlock) 3610 VOP_UNLOCK(vp); 3611 flushwithdeps += hasdeps; 3612 flushed++; 3613 3614 /* 3615 * Sleeping on runningbufspace while holding 3616 * vnode lock leads to deadlock. 3617 */ 3618 if (curproc == bufdaemonproc && 3619 runningbufspace > hirunningspace) 3620 waitrunningbufspace(); 3621 continue; 3622 } 3623 vn_finished_write(mp); 3624 BUF_UNLOCK(bp); 3625 } 3626 BQ_LOCK(bq); 3627 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); 3628 BQ_UNLOCK(bq); 3629 free(sentinel, M_TEMP); 3630 return (flushed); 3631 } 3632 3633 /* 3634 * Check to see if a block is currently memory resident. 3635 */ 3636 struct buf * 3637 incore(struct bufobj *bo, daddr_t blkno) 3638 { 3639 return (gbincore_unlocked(bo, blkno)); 3640 } 3641 3642 /* 3643 * Returns true if no I/O is needed to access the 3644 * associated VM object. This is like incore except 3645 * it also hunts around in the VM system for the data. 3646 */ 3647 bool 3648 inmem(struct vnode * vp, daddr_t blkno) 3649 { 3650 vm_object_t obj; 3651 vm_offset_t toff, tinc, size; 3652 vm_page_t m, n; 3653 vm_ooffset_t off; 3654 int valid; 3655 3656 ASSERT_VOP_LOCKED(vp, "inmem"); 3657 3658 if (incore(&vp->v_bufobj, blkno)) 3659 return (true); 3660 if (vp->v_mount == NULL) 3661 return (false); 3662 obj = vp->v_object; 3663 if (obj == NULL) 3664 return (false); 3665 3666 size = PAGE_SIZE; 3667 if (size > vp->v_mount->mnt_stat.f_iosize) 3668 size = vp->v_mount->mnt_stat.f_iosize; 3669 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 3670 3671 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 3672 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff)); 3673 recheck: 3674 if (m == NULL) 3675 return (false); 3676 3677 tinc = size; 3678 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 3679 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 3680 /* 3681 * Consider page validity only if page mapping didn't change 3682 * during the check. 3683 */ 3684 valid = vm_page_is_valid(m, 3685 (vm_offset_t)((toff + off) & PAGE_MASK), tinc); 3686 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff)); 3687 if (m != n) { 3688 m = n; 3689 goto recheck; 3690 } 3691 if (!valid) 3692 return (false); 3693 } 3694 return (true); 3695 } 3696 3697 /* 3698 * Set the dirty range for a buffer based on the status of the dirty 3699 * bits in the pages comprising the buffer. The range is limited 3700 * to the size of the buffer. 3701 * 3702 * Tell the VM system that the pages associated with this buffer 3703 * are clean. This is used for delayed writes where the data is 3704 * going to go to disk eventually without additional VM intevention. 3705 * 3706 * Note that while we only really need to clean through to b_bcount, we 3707 * just go ahead and clean through to b_bufsize. 3708 */ 3709 static void 3710 vfs_clean_pages_dirty_buf(struct buf *bp) 3711 { 3712 vm_ooffset_t foff, noff, eoff; 3713 vm_page_t m; 3714 int i; 3715 3716 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 3717 return; 3718 3719 foff = bp->b_offset; 3720 KASSERT(bp->b_offset != NOOFFSET, 3721 ("vfs_clean_pages_dirty_buf: no buffer offset")); 3722 3723 vfs_busy_pages_acquire(bp); 3724 vfs_setdirty_range(bp); 3725 for (i = 0; i < bp->b_npages; i++) { 3726 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3727 eoff = noff; 3728 if (eoff > bp->b_offset + bp->b_bufsize) 3729 eoff = bp->b_offset + bp->b_bufsize; 3730 m = bp->b_pages[i]; 3731 vfs_page_set_validclean(bp, foff, m); 3732 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3733 foff = noff; 3734 } 3735 vfs_busy_pages_release(bp); 3736 } 3737 3738 static void 3739 vfs_setdirty_range(struct buf *bp) 3740 { 3741 vm_offset_t boffset; 3742 vm_offset_t eoffset; 3743 int i; 3744 3745 /* 3746 * test the pages to see if they have been modified directly 3747 * by users through the VM system. 3748 */ 3749 for (i = 0; i < bp->b_npages; i++) 3750 vm_page_test_dirty(bp->b_pages[i]); 3751 3752 /* 3753 * Calculate the encompassing dirty range, boffset and eoffset, 3754 * (eoffset - boffset) bytes. 3755 */ 3756 3757 for (i = 0; i < bp->b_npages; i++) { 3758 if (bp->b_pages[i]->dirty) 3759 break; 3760 } 3761 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3762 3763 for (i = bp->b_npages - 1; i >= 0; --i) { 3764 if (bp->b_pages[i]->dirty) { 3765 break; 3766 } 3767 } 3768 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3769 3770 /* 3771 * Fit it to the buffer. 3772 */ 3773 3774 if (eoffset > bp->b_bcount) 3775 eoffset = bp->b_bcount; 3776 3777 /* 3778 * If we have a good dirty range, merge with the existing 3779 * dirty range. 3780 */ 3781 3782 if (boffset < eoffset) { 3783 if (bp->b_dirtyoff > boffset) 3784 bp->b_dirtyoff = boffset; 3785 if (bp->b_dirtyend < eoffset) 3786 bp->b_dirtyend = eoffset; 3787 } 3788 } 3789 3790 /* 3791 * Allocate the KVA mapping for an existing buffer. 3792 * If an unmapped buffer is provided but a mapped buffer is requested, take 3793 * also care to properly setup mappings between pages and KVA. 3794 */ 3795 static void 3796 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 3797 { 3798 int bsize, maxsize, need_mapping, need_kva; 3799 off_t offset; 3800 3801 need_mapping = bp->b_data == unmapped_buf && 3802 (gbflags & GB_UNMAPPED) == 0; 3803 need_kva = bp->b_kvabase == unmapped_buf && 3804 bp->b_data == unmapped_buf && 3805 (gbflags & GB_KVAALLOC) != 0; 3806 if (!need_mapping && !need_kva) 3807 return; 3808 3809 BUF_CHECK_UNMAPPED(bp); 3810 3811 if (need_mapping && bp->b_kvabase != unmapped_buf) { 3812 /* 3813 * Buffer is not mapped, but the KVA was already 3814 * reserved at the time of the instantiation. Use the 3815 * allocated space. 3816 */ 3817 goto has_addr; 3818 } 3819 3820 /* 3821 * Calculate the amount of the address space we would reserve 3822 * if the buffer was mapped. 3823 */ 3824 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 3825 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3826 offset = blkno * bsize; 3827 maxsize = size + (offset & PAGE_MASK); 3828 maxsize = imax(maxsize, bsize); 3829 3830 while (bufkva_alloc(bp, maxsize, gbflags) != 0) { 3831 if ((gbflags & GB_NOWAIT_BD) != 0) { 3832 /* 3833 * XXXKIB: defragmentation cannot 3834 * succeed, not sure what else to do. 3835 */ 3836 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp); 3837 } 3838 counter_u64_add(mappingrestarts, 1); 3839 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0); 3840 } 3841 has_addr: 3842 if (need_mapping) { 3843 /* b_offset is handled by bpmap_qenter. */ 3844 bp->b_data = bp->b_kvabase; 3845 BUF_CHECK_MAPPED(bp); 3846 bpmap_qenter(bp); 3847 } 3848 } 3849 3850 struct buf * 3851 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3852 int flags) 3853 { 3854 struct buf *bp; 3855 int error; 3856 3857 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp); 3858 if (error != 0) 3859 return (NULL); 3860 return (bp); 3861 } 3862 3863 /* 3864 * getblkx: 3865 * 3866 * Get a block given a specified block and offset into a file/device. 3867 * The buffers B_DONE bit will be cleared on return, making it almost 3868 * ready for an I/O initiation. B_INVAL may or may not be set on 3869 * return. The caller should clear B_INVAL prior to initiating a 3870 * READ. 3871 * 3872 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3873 * an existing buffer. 3874 * 3875 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3876 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3877 * and then cleared based on the backing VM. If the previous buffer is 3878 * non-0-sized but invalid, B_CACHE will be cleared. 3879 * 3880 * If getblk() must create a new buffer, the new buffer is returned with 3881 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3882 * case it is returned with B_INVAL clear and B_CACHE set based on the 3883 * backing VM. 3884 * 3885 * getblk() also forces a bwrite() for any B_DELWRI buffer whose 3886 * B_CACHE bit is clear. 3887 * 3888 * What this means, basically, is that the caller should use B_CACHE to 3889 * determine whether the buffer is fully valid or not and should clear 3890 * B_INVAL prior to issuing a read. If the caller intends to validate 3891 * the buffer by loading its data area with something, the caller needs 3892 * to clear B_INVAL. If the caller does this without issuing an I/O, 3893 * the caller should set B_CACHE ( as an optimization ), else the caller 3894 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3895 * a write attempt or if it was a successful read. If the caller 3896 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3897 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3898 * 3899 * The blkno parameter is the logical block being requested. Normally 3900 * the mapping of logical block number to disk block address is done 3901 * by calling VOP_BMAP(). However, if the mapping is already known, the 3902 * disk block address can be passed using the dblkno parameter. If the 3903 * disk block address is not known, then the same value should be passed 3904 * for blkno and dblkno. 3905 */ 3906 int 3907 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag, 3908 int slptimeo, int flags, struct buf **bpp) 3909 { 3910 struct buf *bp; 3911 struct bufobj *bo; 3912 daddr_t d_blkno; 3913 int bsize, error, maxsize, vmio; 3914 off_t offset; 3915 3916 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3917 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3918 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3919 if (vp->v_type != VCHR) 3920 ASSERT_VOP_LOCKED(vp, "getblk"); 3921 if (size > maxbcachebuf) 3922 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size, 3923 maxbcachebuf); 3924 if (!unmapped_buf_allowed) 3925 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3926 3927 bo = &vp->v_bufobj; 3928 d_blkno = dblkno; 3929 3930 /* Attempt lockless lookup first. */ 3931 bp = gbincore_unlocked(bo, blkno); 3932 if (bp == NULL) { 3933 /* 3934 * With GB_NOCREAT we must be sure about not finding the buffer 3935 * as it may have been reassigned during unlocked lookup. 3936 */ 3937 if ((flags & GB_NOCREAT) != 0) 3938 goto loop; 3939 goto newbuf_unlocked; 3940 } 3941 3942 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0, 3943 0); 3944 if (error != 0) 3945 goto loop; 3946 3947 /* Verify buf identify has not changed since lookup. */ 3948 if (bp->b_bufobj == bo && bp->b_lblkno == blkno) 3949 goto foundbuf_fastpath; 3950 3951 /* It changed, fallback to locked lookup. */ 3952 BUF_UNLOCK_RAW(bp); 3953 3954 loop: 3955 BO_RLOCK(bo); 3956 bp = gbincore(bo, blkno); 3957 if (bp != NULL) { 3958 int lockflags; 3959 3960 /* 3961 * Buffer is in-core. If the buffer is not busy nor managed, 3962 * it must be on a queue. 3963 */ 3964 lockflags = LK_EXCLUSIVE | LK_INTERLOCK | 3965 ((flags & GB_LOCK_NOWAIT) ? LK_NOWAIT : LK_SLEEPFAIL); 3966 3967 error = BUF_TIMELOCK(bp, lockflags, 3968 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3969 3970 /* 3971 * If we slept and got the lock we have to restart in case 3972 * the buffer changed identities. 3973 */ 3974 if (error == ENOLCK) 3975 goto loop; 3976 /* We timed out or were interrupted. */ 3977 else if (error != 0) 3978 return (error); 3979 3980 foundbuf_fastpath: 3981 /* If recursed, assume caller knows the rules. */ 3982 if (BUF_LOCKRECURSED(bp)) 3983 goto end; 3984 3985 /* 3986 * The buffer is locked. B_CACHE is cleared if the buffer is 3987 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3988 * and for a VMIO buffer B_CACHE is adjusted according to the 3989 * backing VM cache. 3990 */ 3991 if (bp->b_flags & B_INVAL) 3992 bp->b_flags &= ~B_CACHE; 3993 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3994 bp->b_flags |= B_CACHE; 3995 if (bp->b_flags & B_MANAGED) 3996 MPASS(bp->b_qindex == QUEUE_NONE); 3997 else 3998 bremfree(bp); 3999 4000 /* 4001 * check for size inconsistencies for non-VMIO case. 4002 */ 4003 if (bp->b_bcount != size) { 4004 if ((bp->b_flags & B_VMIO) == 0 || 4005 (size > bp->b_kvasize)) { 4006 if (bp->b_flags & B_DELWRI) { 4007 bp->b_flags |= B_NOCACHE; 4008 bwrite(bp); 4009 } else { 4010 if (LIST_EMPTY(&bp->b_dep)) { 4011 bp->b_flags |= B_RELBUF; 4012 brelse(bp); 4013 } else { 4014 bp->b_flags |= B_NOCACHE; 4015 bwrite(bp); 4016 } 4017 } 4018 goto loop; 4019 } 4020 } 4021 4022 /* 4023 * Handle the case of unmapped buffer which should 4024 * become mapped, or the buffer for which KVA 4025 * reservation is requested. 4026 */ 4027 bp_unmapped_get_kva(bp, blkno, size, flags); 4028 4029 /* 4030 * If the size is inconsistent in the VMIO case, we can resize 4031 * the buffer. This might lead to B_CACHE getting set or 4032 * cleared. If the size has not changed, B_CACHE remains 4033 * unchanged from its previous state. 4034 */ 4035 allocbuf(bp, size); 4036 4037 KASSERT(bp->b_offset != NOOFFSET, 4038 ("getblk: no buffer offset")); 4039 4040 /* 4041 * A buffer with B_DELWRI set and B_CACHE clear must 4042 * be committed before we can return the buffer in 4043 * order to prevent the caller from issuing a read 4044 * ( due to B_CACHE not being set ) and overwriting 4045 * it. 4046 * 4047 * Most callers, including NFS and FFS, need this to 4048 * operate properly either because they assume they 4049 * can issue a read if B_CACHE is not set, or because 4050 * ( for example ) an uncached B_DELWRI might loop due 4051 * to softupdates re-dirtying the buffer. In the latter 4052 * case, B_CACHE is set after the first write completes, 4053 * preventing further loops. 4054 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 4055 * above while extending the buffer, we cannot allow the 4056 * buffer to remain with B_CACHE set after the write 4057 * completes or it will represent a corrupt state. To 4058 * deal with this we set B_NOCACHE to scrap the buffer 4059 * after the write. 4060 * 4061 * We might be able to do something fancy, like setting 4062 * B_CACHE in bwrite() except if B_DELWRI is already set, 4063 * so the below call doesn't set B_CACHE, but that gets real 4064 * confusing. This is much easier. 4065 */ 4066 4067 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 4068 bp->b_flags |= B_NOCACHE; 4069 bwrite(bp); 4070 goto loop; 4071 } 4072 bp->b_flags &= ~B_DONE; 4073 } else { 4074 /* 4075 * Buffer is not in-core, create new buffer. The buffer 4076 * returned by getnewbuf() is locked. Note that the returned 4077 * buffer is also considered valid (not marked B_INVAL). 4078 */ 4079 BO_RUNLOCK(bo); 4080 newbuf_unlocked: 4081 /* 4082 * If the user does not want us to create the buffer, bail out 4083 * here. 4084 */ 4085 if (flags & GB_NOCREAT) 4086 return (EEXIST); 4087 4088 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize; 4089 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 4090 offset = blkno * bsize; 4091 vmio = vp->v_object != NULL; 4092 if (vmio) { 4093 maxsize = size + (offset & PAGE_MASK); 4094 } else { 4095 maxsize = size; 4096 /* Do not allow non-VMIO notmapped buffers. */ 4097 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 4098 } 4099 maxsize = imax(maxsize, bsize); 4100 if ((flags & GB_NOSPARSE) != 0 && vmio && 4101 !vn_isdisk(vp)) { 4102 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0); 4103 KASSERT(error != EOPNOTSUPP, 4104 ("GB_NOSPARSE from fs not supporting bmap, vp %p", 4105 vp)); 4106 if (error != 0) 4107 return (error); 4108 if (d_blkno == -1) 4109 return (EJUSTRETURN); 4110 } 4111 4112 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags); 4113 if (bp == NULL) { 4114 if (slpflag || slptimeo) 4115 return (ETIMEDOUT); 4116 /* 4117 * XXX This is here until the sleep path is diagnosed 4118 * enough to work under very low memory conditions. 4119 * 4120 * There's an issue on low memory, 4BSD+non-preempt 4121 * systems (eg MIPS routers with 32MB RAM) where buffer 4122 * exhaustion occurs without sleeping for buffer 4123 * reclaimation. This just sticks in a loop and 4124 * constantly attempts to allocate a buffer, which 4125 * hits exhaustion and tries to wakeup bufdaemon. 4126 * This never happens because we never yield. 4127 * 4128 * The real solution is to identify and fix these cases 4129 * so we aren't effectively busy-waiting in a loop 4130 * until the reclaimation path has cycles to run. 4131 */ 4132 kern_yield(PRI_USER); 4133 goto loop; 4134 } 4135 4136 /* 4137 * This code is used to make sure that a buffer is not 4138 * created while the getnewbuf routine is blocked. 4139 * This can be a problem whether the vnode is locked or not. 4140 * If the buffer is created out from under us, we have to 4141 * throw away the one we just created. 4142 * 4143 * Note: this must occur before we associate the buffer 4144 * with the vp especially considering limitations in 4145 * the splay tree implementation when dealing with duplicate 4146 * lblkno's. 4147 */ 4148 BO_LOCK(bo); 4149 if (gbincore(bo, blkno)) { 4150 BO_UNLOCK(bo); 4151 bp->b_flags |= B_INVAL; 4152 bufspace_release(bufdomain(bp), maxsize); 4153 brelse(bp); 4154 goto loop; 4155 } 4156 4157 /* 4158 * Insert the buffer into the hash, so that it can 4159 * be found by incore. 4160 */ 4161 bp->b_lblkno = blkno; 4162 bp->b_blkno = d_blkno; 4163 bp->b_offset = offset; 4164 bgetvp(vp, bp); 4165 BO_UNLOCK(bo); 4166 4167 /* 4168 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 4169 * buffer size starts out as 0, B_CACHE will be set by 4170 * allocbuf() for the VMIO case prior to it testing the 4171 * backing store for validity. 4172 */ 4173 4174 if (vmio) { 4175 bp->b_flags |= B_VMIO; 4176 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 4177 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 4178 bp, vp->v_object, bp->b_bufobj->bo_object)); 4179 } else { 4180 bp->b_flags &= ~B_VMIO; 4181 KASSERT(bp->b_bufobj->bo_object == NULL, 4182 ("ARGH! has b_bufobj->bo_object %p %p\n", 4183 bp, bp->b_bufobj->bo_object)); 4184 BUF_CHECK_MAPPED(bp); 4185 } 4186 4187 allocbuf(bp, size); 4188 bufspace_release(bufdomain(bp), maxsize); 4189 bp->b_flags &= ~B_DONE; 4190 } 4191 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 4192 end: 4193 buf_track(bp, __func__); 4194 KASSERT(bp->b_bufobj == bo, 4195 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 4196 *bpp = bp; 4197 return (0); 4198 } 4199 4200 /* 4201 * Get an empty, disassociated buffer of given size. The buffer is initially 4202 * set to B_INVAL. 4203 */ 4204 struct buf * 4205 geteblk(int size, int flags) 4206 { 4207 struct buf *bp; 4208 int maxsize; 4209 4210 maxsize = (size + BKVAMASK) & ~BKVAMASK; 4211 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) { 4212 if ((flags & GB_NOWAIT_BD) && 4213 (curthread->td_pflags & TDP_BUFNEED) != 0) 4214 return (NULL); 4215 } 4216 allocbuf(bp, size); 4217 bufspace_release(bufdomain(bp), maxsize); 4218 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 4219 return (bp); 4220 } 4221 4222 /* 4223 * Truncate the backing store for a non-vmio buffer. 4224 */ 4225 static void 4226 vfs_nonvmio_truncate(struct buf *bp, int newbsize) 4227 { 4228 4229 if (bp->b_flags & B_MALLOC) { 4230 /* 4231 * malloced buffers are not shrunk 4232 */ 4233 if (newbsize == 0) { 4234 bufmallocadjust(bp, 0); 4235 free(bp->b_data, M_BIOBUF); 4236 bp->b_data = bp->b_kvabase; 4237 bp->b_flags &= ~B_MALLOC; 4238 } 4239 return; 4240 } 4241 vm_hold_free_pages(bp, newbsize); 4242 bufspace_adjust(bp, newbsize); 4243 } 4244 4245 /* 4246 * Extend the backing for a non-VMIO buffer. 4247 */ 4248 static void 4249 vfs_nonvmio_extend(struct buf *bp, int newbsize) 4250 { 4251 caddr_t origbuf; 4252 int origbufsize; 4253 4254 /* 4255 * We only use malloced memory on the first allocation. 4256 * and revert to page-allocated memory when the buffer 4257 * grows. 4258 * 4259 * There is a potential smp race here that could lead 4260 * to bufmallocspace slightly passing the max. It 4261 * is probably extremely rare and not worth worrying 4262 * over. 4263 */ 4264 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 && 4265 bufmallocspace < maxbufmallocspace) { 4266 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK); 4267 bp->b_flags |= B_MALLOC; 4268 bufmallocadjust(bp, newbsize); 4269 return; 4270 } 4271 4272 /* 4273 * If the buffer is growing on its other-than-first 4274 * allocation then we revert to the page-allocation 4275 * scheme. 4276 */ 4277 origbuf = NULL; 4278 origbufsize = 0; 4279 if (bp->b_flags & B_MALLOC) { 4280 origbuf = bp->b_data; 4281 origbufsize = bp->b_bufsize; 4282 bp->b_data = bp->b_kvabase; 4283 bufmallocadjust(bp, 0); 4284 bp->b_flags &= ~B_MALLOC; 4285 newbsize = round_page(newbsize); 4286 } 4287 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize, 4288 (vm_offset_t) bp->b_data + newbsize); 4289 if (origbuf != NULL) { 4290 bcopy(origbuf, bp->b_data, origbufsize); 4291 free(origbuf, M_BIOBUF); 4292 } 4293 bufspace_adjust(bp, newbsize); 4294 } 4295 4296 /* 4297 * This code constitutes the buffer memory from either anonymous system 4298 * memory (in the case of non-VMIO operations) or from an associated 4299 * VM object (in the case of VMIO operations). This code is able to 4300 * resize a buffer up or down. 4301 * 4302 * Note that this code is tricky, and has many complications to resolve 4303 * deadlock or inconsistent data situations. Tread lightly!!! 4304 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 4305 * the caller. Calling this code willy nilly can result in the loss of data. 4306 * 4307 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 4308 * B_CACHE for the non-VMIO case. 4309 */ 4310 int 4311 allocbuf(struct buf *bp, int size) 4312 { 4313 int newbsize; 4314 4315 if (bp->b_bcount == size) 4316 return (1); 4317 4318 if (bp->b_kvasize != 0 && bp->b_kvasize < size) 4319 panic("allocbuf: buffer too small"); 4320 4321 newbsize = roundup2(size, DEV_BSIZE); 4322 if ((bp->b_flags & B_VMIO) == 0) { 4323 if ((bp->b_flags & B_MALLOC) == 0) 4324 newbsize = round_page(newbsize); 4325 /* 4326 * Just get anonymous memory from the kernel. Don't 4327 * mess with B_CACHE. 4328 */ 4329 if (newbsize < bp->b_bufsize) 4330 vfs_nonvmio_truncate(bp, newbsize); 4331 else if (newbsize > bp->b_bufsize) 4332 vfs_nonvmio_extend(bp, newbsize); 4333 } else { 4334 int desiredpages; 4335 4336 desiredpages = (size == 0) ? 0 : 4337 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 4338 4339 if (bp->b_flags & B_MALLOC) 4340 panic("allocbuf: VMIO buffer can't be malloced"); 4341 /* 4342 * Set B_CACHE initially if buffer is 0 length or will become 4343 * 0-length. 4344 */ 4345 if (size == 0 || bp->b_bufsize == 0) 4346 bp->b_flags |= B_CACHE; 4347 4348 if (newbsize < bp->b_bufsize) 4349 vfs_vmio_truncate(bp, desiredpages); 4350 /* XXX This looks as if it should be newbsize > b_bufsize */ 4351 else if (size > bp->b_bcount) 4352 vfs_vmio_extend(bp, desiredpages, size); 4353 bufspace_adjust(bp, newbsize); 4354 } 4355 bp->b_bcount = size; /* requested buffer size. */ 4356 return (1); 4357 } 4358 4359 extern int inflight_transient_maps; 4360 4361 static struct bio_queue nondump_bios; 4362 4363 void 4364 biodone(struct bio *bp) 4365 { 4366 struct mtx *mtxp; 4367 void (*done)(struct bio *); 4368 vm_offset_t start, end; 4369 4370 biotrack(bp, __func__); 4371 4372 /* 4373 * Avoid completing I/O when dumping after a panic since that may 4374 * result in a deadlock in the filesystem or pager code. Note that 4375 * this doesn't affect dumps that were started manually since we aim 4376 * to keep the system usable after it has been resumed. 4377 */ 4378 if (__predict_false(dumping && SCHEDULER_STOPPED())) { 4379 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue); 4380 return; 4381 } 4382 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 4383 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 4384 bp->bio_flags |= BIO_UNMAPPED; 4385 start = trunc_page((vm_offset_t)bp->bio_data); 4386 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 4387 bp->bio_data = unmapped_buf; 4388 pmap_qremove(start, atop(end - start)); 4389 vmem_free(transient_arena, start, end - start); 4390 atomic_add_int(&inflight_transient_maps, -1); 4391 } 4392 done = bp->bio_done; 4393 /* 4394 * The check for done == biodone is to allow biodone to be 4395 * used as a bio_done routine. 4396 */ 4397 if (done == NULL || done == biodone) { 4398 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4399 mtx_lock(mtxp); 4400 bp->bio_flags |= BIO_DONE; 4401 wakeup(bp); 4402 mtx_unlock(mtxp); 4403 } else 4404 done(bp); 4405 } 4406 4407 /* 4408 * Wait for a BIO to finish. 4409 */ 4410 int 4411 biowait(struct bio *bp, const char *wmesg) 4412 { 4413 struct mtx *mtxp; 4414 4415 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4416 mtx_lock(mtxp); 4417 while ((bp->bio_flags & BIO_DONE) == 0) 4418 msleep(bp, mtxp, PRIBIO, wmesg, 0); 4419 mtx_unlock(mtxp); 4420 if (bp->bio_error != 0) 4421 return (bp->bio_error); 4422 if (!(bp->bio_flags & BIO_ERROR)) 4423 return (0); 4424 return (EIO); 4425 } 4426 4427 void 4428 biofinish(struct bio *bp, struct devstat *stat, int error) 4429 { 4430 4431 if (error) { 4432 bp->bio_error = error; 4433 bp->bio_flags |= BIO_ERROR; 4434 } 4435 if (stat != NULL) 4436 devstat_end_transaction_bio(stat, bp); 4437 biodone(bp); 4438 } 4439 4440 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING) 4441 void 4442 biotrack_buf(struct bio *bp, const char *location) 4443 { 4444 4445 buf_track(bp->bio_track_bp, location); 4446 } 4447 #endif 4448 4449 /* 4450 * bufwait: 4451 * 4452 * Wait for buffer I/O completion, returning error status. The buffer 4453 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 4454 * error and cleared. 4455 */ 4456 int 4457 bufwait(struct buf *bp) 4458 { 4459 if (bp->b_iocmd == BIO_READ) 4460 bwait(bp, PRIBIO, "biord"); 4461 else 4462 bwait(bp, PRIBIO, "biowr"); 4463 if (bp->b_flags & B_EINTR) { 4464 bp->b_flags &= ~B_EINTR; 4465 return (EINTR); 4466 } 4467 if (bp->b_ioflags & BIO_ERROR) { 4468 return (bp->b_error ? bp->b_error : EIO); 4469 } else { 4470 return (0); 4471 } 4472 } 4473 4474 /* 4475 * bufdone: 4476 * 4477 * Finish I/O on a buffer, optionally calling a completion function. 4478 * This is usually called from an interrupt so process blocking is 4479 * not allowed. 4480 * 4481 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 4482 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 4483 * assuming B_INVAL is clear. 4484 * 4485 * For the VMIO case, we set B_CACHE if the op was a read and no 4486 * read error occurred, or if the op was a write. B_CACHE is never 4487 * set if the buffer is invalid or otherwise uncacheable. 4488 * 4489 * bufdone does not mess with B_INVAL, allowing the I/O routine or the 4490 * initiator to leave B_INVAL set to brelse the buffer out of existence 4491 * in the biodone routine. 4492 */ 4493 void 4494 bufdone(struct buf *bp) 4495 { 4496 struct bufobj *dropobj; 4497 void (*biodone)(struct buf *); 4498 4499 buf_track(bp, __func__); 4500 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 4501 dropobj = NULL; 4502 4503 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 4504 4505 runningbufwakeup(bp); 4506 if (bp->b_iocmd == BIO_WRITE) 4507 dropobj = bp->b_bufobj; 4508 /* call optional completion function if requested */ 4509 if (bp->b_iodone != NULL) { 4510 biodone = bp->b_iodone; 4511 bp->b_iodone = NULL; 4512 (*biodone) (bp); 4513 if (dropobj) 4514 bufobj_wdrop(dropobj); 4515 return; 4516 } 4517 if (bp->b_flags & B_VMIO) { 4518 /* 4519 * Set B_CACHE if the op was a normal read and no error 4520 * occurred. B_CACHE is set for writes in the b*write() 4521 * routines. 4522 */ 4523 if (bp->b_iocmd == BIO_READ && 4524 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 4525 !(bp->b_ioflags & BIO_ERROR)) 4526 bp->b_flags |= B_CACHE; 4527 vfs_vmio_iodone(bp); 4528 } 4529 if (!LIST_EMPTY(&bp->b_dep)) 4530 buf_complete(bp); 4531 if ((bp->b_flags & B_CKHASH) != 0) { 4532 KASSERT(bp->b_iocmd == BIO_READ, 4533 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd)); 4534 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp)); 4535 (*bp->b_ckhashcalc)(bp); 4536 } 4537 /* 4538 * For asynchronous completions, release the buffer now. The brelse 4539 * will do a wakeup there if necessary - so no need to do a wakeup 4540 * here in the async case. The sync case always needs to do a wakeup. 4541 */ 4542 if (bp->b_flags & B_ASYNC) { 4543 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || 4544 (bp->b_ioflags & BIO_ERROR)) 4545 brelse(bp); 4546 else 4547 bqrelse(bp); 4548 } else 4549 bdone(bp); 4550 if (dropobj) 4551 bufobj_wdrop(dropobj); 4552 } 4553 4554 /* 4555 * This routine is called in lieu of iodone in the case of 4556 * incomplete I/O. This keeps the busy status for pages 4557 * consistent. 4558 */ 4559 void 4560 vfs_unbusy_pages(struct buf *bp) 4561 { 4562 int i; 4563 vm_object_t obj; 4564 vm_page_t m; 4565 4566 runningbufwakeup(bp); 4567 if (!(bp->b_flags & B_VMIO)) 4568 return; 4569 4570 obj = bp->b_bufobj->bo_object; 4571 for (i = 0; i < bp->b_npages; i++) { 4572 m = bp->b_pages[i]; 4573 if (m == bogus_page) { 4574 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i); 4575 if (!m) 4576 panic("vfs_unbusy_pages: page missing\n"); 4577 bp->b_pages[i] = m; 4578 if (buf_mapped(bp)) { 4579 BUF_CHECK_MAPPED(bp); 4580 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4581 bp->b_pages, bp->b_npages); 4582 } else 4583 BUF_CHECK_UNMAPPED(bp); 4584 } 4585 vm_page_sunbusy(m); 4586 } 4587 vm_object_pip_wakeupn(obj, bp->b_npages); 4588 } 4589 4590 /* 4591 * vfs_page_set_valid: 4592 * 4593 * Set the valid bits in a page based on the supplied offset. The 4594 * range is restricted to the buffer's size. 4595 * 4596 * This routine is typically called after a read completes. 4597 */ 4598 static void 4599 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4600 { 4601 vm_ooffset_t eoff; 4602 4603 /* 4604 * Compute the end offset, eoff, such that [off, eoff) does not span a 4605 * page boundary and eoff is not greater than the end of the buffer. 4606 * The end of the buffer, in this case, is our file EOF, not the 4607 * allocation size of the buffer. 4608 */ 4609 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 4610 if (eoff > bp->b_offset + bp->b_bcount) 4611 eoff = bp->b_offset + bp->b_bcount; 4612 4613 /* 4614 * Set valid range. This is typically the entire buffer and thus the 4615 * entire page. 4616 */ 4617 if (eoff > off) 4618 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 4619 } 4620 4621 /* 4622 * vfs_page_set_validclean: 4623 * 4624 * Set the valid bits and clear the dirty bits in a page based on the 4625 * supplied offset. The range is restricted to the buffer's size. 4626 */ 4627 static void 4628 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4629 { 4630 vm_ooffset_t soff, eoff; 4631 4632 /* 4633 * Start and end offsets in buffer. eoff - soff may not cross a 4634 * page boundary or cross the end of the buffer. The end of the 4635 * buffer, in this case, is our file EOF, not the allocation size 4636 * of the buffer. 4637 */ 4638 soff = off; 4639 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4640 if (eoff > bp->b_offset + bp->b_bcount) 4641 eoff = bp->b_offset + bp->b_bcount; 4642 4643 /* 4644 * Set valid range. This is typically the entire buffer and thus the 4645 * entire page. 4646 */ 4647 if (eoff > soff) { 4648 vm_page_set_validclean( 4649 m, 4650 (vm_offset_t) (soff & PAGE_MASK), 4651 (vm_offset_t) (eoff - soff) 4652 ); 4653 } 4654 } 4655 4656 /* 4657 * Acquire a shared busy on all pages in the buf. 4658 */ 4659 void 4660 vfs_busy_pages_acquire(struct buf *bp) 4661 { 4662 int i; 4663 4664 for (i = 0; i < bp->b_npages; i++) 4665 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY); 4666 } 4667 4668 void 4669 vfs_busy_pages_release(struct buf *bp) 4670 { 4671 int i; 4672 4673 for (i = 0; i < bp->b_npages; i++) 4674 vm_page_sunbusy(bp->b_pages[i]); 4675 } 4676 4677 /* 4678 * This routine is called before a device strategy routine. 4679 * It is used to tell the VM system that paging I/O is in 4680 * progress, and treat the pages associated with the buffer 4681 * almost as being exclusive busy. Also the object paging_in_progress 4682 * flag is handled to make sure that the object doesn't become 4683 * inconsistent. 4684 * 4685 * Since I/O has not been initiated yet, certain buffer flags 4686 * such as BIO_ERROR or B_INVAL may be in an inconsistent state 4687 * and should be ignored. 4688 */ 4689 void 4690 vfs_busy_pages(struct buf *bp, int clear_modify) 4691 { 4692 vm_object_t obj; 4693 vm_ooffset_t foff; 4694 vm_page_t m; 4695 int i; 4696 bool bogus; 4697 4698 if (!(bp->b_flags & B_VMIO)) 4699 return; 4700 4701 obj = bp->b_bufobj->bo_object; 4702 foff = bp->b_offset; 4703 KASSERT(bp->b_offset != NOOFFSET, 4704 ("vfs_busy_pages: no buffer offset")); 4705 if ((bp->b_flags & B_CLUSTER) == 0) { 4706 vm_object_pip_add(obj, bp->b_npages); 4707 vfs_busy_pages_acquire(bp); 4708 } 4709 if (bp->b_bufsize != 0) 4710 vfs_setdirty_range(bp); 4711 bogus = false; 4712 for (i = 0; i < bp->b_npages; i++) { 4713 m = bp->b_pages[i]; 4714 vm_page_assert_sbusied(m); 4715 4716 /* 4717 * When readying a buffer for a read ( i.e 4718 * clear_modify == 0 ), it is important to do 4719 * bogus_page replacement for valid pages in 4720 * partially instantiated buffers. Partially 4721 * instantiated buffers can, in turn, occur when 4722 * reconstituting a buffer from its VM backing store 4723 * base. We only have to do this if B_CACHE is 4724 * clear ( which causes the I/O to occur in the 4725 * first place ). The replacement prevents the read 4726 * I/O from overwriting potentially dirty VM-backed 4727 * pages. XXX bogus page replacement is, uh, bogus. 4728 * It may not work properly with small-block devices. 4729 * We need to find a better way. 4730 */ 4731 if (clear_modify) { 4732 pmap_remove_write(m); 4733 vfs_page_set_validclean(bp, foff, m); 4734 } else if (vm_page_all_valid(m) && 4735 (bp->b_flags & B_CACHE) == 0) { 4736 bp->b_pages[i] = bogus_page; 4737 bogus = true; 4738 } 4739 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4740 } 4741 if (bogus && buf_mapped(bp)) { 4742 BUF_CHECK_MAPPED(bp); 4743 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4744 bp->b_pages, bp->b_npages); 4745 } 4746 } 4747 4748 /* 4749 * vfs_bio_set_valid: 4750 * 4751 * Set the range within the buffer to valid. The range is 4752 * relative to the beginning of the buffer, b_offset. Note that 4753 * b_offset itself may be offset from the beginning of the first 4754 * page. 4755 */ 4756 void 4757 vfs_bio_set_valid(struct buf *bp, int base, int size) 4758 { 4759 int i, n; 4760 vm_page_t m; 4761 4762 if (!(bp->b_flags & B_VMIO)) 4763 return; 4764 4765 /* 4766 * Fixup base to be relative to beginning of first page. 4767 * Set initial n to be the maximum number of bytes in the 4768 * first page that can be validated. 4769 */ 4770 base += (bp->b_offset & PAGE_MASK); 4771 n = PAGE_SIZE - (base & PAGE_MASK); 4772 4773 /* 4774 * Busy may not be strictly necessary here because the pages are 4775 * unlikely to be fully valid and the vnode lock will synchronize 4776 * their access via getpages. It is grabbed for consistency with 4777 * other page validation. 4778 */ 4779 vfs_busy_pages_acquire(bp); 4780 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4781 m = bp->b_pages[i]; 4782 if (n > size) 4783 n = size; 4784 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4785 base += n; 4786 size -= n; 4787 n = PAGE_SIZE; 4788 } 4789 vfs_busy_pages_release(bp); 4790 } 4791 4792 /* 4793 * vfs_bio_clrbuf: 4794 * 4795 * If the specified buffer is a non-VMIO buffer, clear the entire 4796 * buffer. If the specified buffer is a VMIO buffer, clear and 4797 * validate only the previously invalid portions of the buffer. 4798 * This routine essentially fakes an I/O, so we need to clear 4799 * BIO_ERROR and B_INVAL. 4800 * 4801 * Note that while we only theoretically need to clear through b_bcount, 4802 * we go ahead and clear through b_bufsize. 4803 */ 4804 void 4805 vfs_bio_clrbuf(struct buf *bp) 4806 { 4807 int i, j, mask, sa, ea, slide; 4808 4809 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4810 clrbuf(bp); 4811 return; 4812 } 4813 bp->b_flags &= ~B_INVAL; 4814 bp->b_ioflags &= ~BIO_ERROR; 4815 vfs_busy_pages_acquire(bp); 4816 sa = bp->b_offset & PAGE_MASK; 4817 slide = 0; 4818 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4819 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4820 ea = slide & PAGE_MASK; 4821 if (ea == 0) 4822 ea = PAGE_SIZE; 4823 if (bp->b_pages[i] == bogus_page) 4824 continue; 4825 j = sa / DEV_BSIZE; 4826 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4827 if ((bp->b_pages[i]->valid & mask) == mask) 4828 continue; 4829 if ((bp->b_pages[i]->valid & mask) == 0) 4830 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4831 else { 4832 for (; sa < ea; sa += DEV_BSIZE, j++) { 4833 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4834 pmap_zero_page_area(bp->b_pages[i], 4835 sa, DEV_BSIZE); 4836 } 4837 } 4838 } 4839 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE, 4840 roundup2(ea - sa, DEV_BSIZE)); 4841 } 4842 vfs_busy_pages_release(bp); 4843 bp->b_resid = 0; 4844 } 4845 4846 void 4847 vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4848 { 4849 vm_page_t m; 4850 int i, n; 4851 4852 if (buf_mapped(bp)) { 4853 BUF_CHECK_MAPPED(bp); 4854 bzero(bp->b_data + base, size); 4855 } else { 4856 BUF_CHECK_UNMAPPED(bp); 4857 n = PAGE_SIZE - (base & PAGE_MASK); 4858 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4859 m = bp->b_pages[i]; 4860 if (n > size) 4861 n = size; 4862 pmap_zero_page_area(m, base & PAGE_MASK, n); 4863 base += n; 4864 size -= n; 4865 n = PAGE_SIZE; 4866 } 4867 } 4868 } 4869 4870 /* 4871 * Update buffer flags based on I/O request parameters, optionally releasing the 4872 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM, 4873 * where they may be placed on a page queue (VMIO) or freed immediately (direct 4874 * I/O). Otherwise the buffer is released to the cache. 4875 */ 4876 static void 4877 b_io_dismiss(struct buf *bp, int ioflag, bool release) 4878 { 4879 4880 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0, 4881 ("buf %p non-VMIO noreuse", bp)); 4882 4883 if ((ioflag & IO_DIRECT) != 0) 4884 bp->b_flags |= B_DIRECT; 4885 if ((ioflag & IO_EXT) != 0) 4886 bp->b_xflags |= BX_ALTDATA; 4887 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) { 4888 bp->b_flags |= B_RELBUF; 4889 if ((ioflag & IO_NOREUSE) != 0) 4890 bp->b_flags |= B_NOREUSE; 4891 if (release) 4892 brelse(bp); 4893 } else if (release) 4894 bqrelse(bp); 4895 } 4896 4897 void 4898 vfs_bio_brelse(struct buf *bp, int ioflag) 4899 { 4900 4901 b_io_dismiss(bp, ioflag, true); 4902 } 4903 4904 void 4905 vfs_bio_set_flags(struct buf *bp, int ioflag) 4906 { 4907 4908 b_io_dismiss(bp, ioflag, false); 4909 } 4910 4911 /* 4912 * vm_hold_load_pages and vm_hold_free_pages get pages into 4913 * a buffers address space. The pages are anonymous and are 4914 * not associated with a file object. 4915 */ 4916 static void 4917 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4918 { 4919 vm_offset_t pg; 4920 vm_page_t p; 4921 int index; 4922 4923 BUF_CHECK_MAPPED(bp); 4924 4925 to = round_page(to); 4926 from = round_page(from); 4927 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4928 MPASS((bp->b_flags & B_MAXPHYS) == 0); 4929 KASSERT(to - from <= maxbcachebuf, 4930 ("vm_hold_load_pages too large %p %#jx %#jx %u", 4931 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf)); 4932 4933 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4934 /* 4935 * note: must allocate system pages since blocking here 4936 * could interfere with paging I/O, no matter which 4937 * process we are. 4938 */ 4939 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 4940 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK); 4941 pmap_qenter(pg, &p, 1); 4942 bp->b_pages[index] = p; 4943 } 4944 bp->b_npages = index; 4945 } 4946 4947 /* Return pages associated with this buf to the vm system */ 4948 static void 4949 vm_hold_free_pages(struct buf *bp, int newbsize) 4950 { 4951 vm_offset_t from; 4952 vm_page_t p; 4953 int index, newnpages; 4954 4955 BUF_CHECK_MAPPED(bp); 4956 4957 from = round_page((vm_offset_t)bp->b_data + newbsize); 4958 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4959 if (bp->b_npages > newnpages) 4960 pmap_qremove(from, bp->b_npages - newnpages); 4961 for (index = newnpages; index < bp->b_npages; index++) { 4962 p = bp->b_pages[index]; 4963 bp->b_pages[index] = NULL; 4964 vm_page_unwire_noq(p); 4965 vm_page_free(p); 4966 } 4967 bp->b_npages = newnpages; 4968 } 4969 4970 /* 4971 * Map an IO request into kernel virtual address space. 4972 * 4973 * All requests are (re)mapped into kernel VA space. 4974 * Notice that we use b_bufsize for the size of the buffer 4975 * to be mapped. b_bcount might be modified by the driver. 4976 * 4977 * Note that even if the caller determines that the address space should 4978 * be valid, a race or a smaller-file mapped into a larger space may 4979 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4980 * check the return value. 4981 * 4982 * This function only works with pager buffers. 4983 */ 4984 int 4985 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf) 4986 { 4987 vm_prot_t prot; 4988 int pidx; 4989 4990 MPASS((bp->b_flags & B_MAXPHYS) != 0); 4991 prot = VM_PROT_READ; 4992 if (bp->b_iocmd == BIO_READ) 4993 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4994 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4995 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES); 4996 if (pidx < 0) 4997 return (-1); 4998 bp->b_bufsize = len; 4999 bp->b_npages = pidx; 5000 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK; 5001 if (mapbuf || !unmapped_buf_allowed) { 5002 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx); 5003 bp->b_data = bp->b_kvabase + bp->b_offset; 5004 } else 5005 bp->b_data = unmapped_buf; 5006 return (0); 5007 } 5008 5009 /* 5010 * Free the io map PTEs associated with this IO operation. 5011 * We also invalidate the TLB entries and restore the original b_addr. 5012 * 5013 * This function only works with pager buffers. 5014 */ 5015 void 5016 vunmapbuf(struct buf *bp) 5017 { 5018 int npages; 5019 5020 npages = bp->b_npages; 5021 if (buf_mapped(bp)) 5022 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 5023 vm_page_unhold_pages(bp->b_pages, npages); 5024 5025 bp->b_data = unmapped_buf; 5026 } 5027 5028 void 5029 bdone(struct buf *bp) 5030 { 5031 struct mtx *mtxp; 5032 5033 mtxp = mtx_pool_find(mtxpool_sleep, bp); 5034 mtx_lock(mtxp); 5035 bp->b_flags |= B_DONE; 5036 wakeup(bp); 5037 mtx_unlock(mtxp); 5038 } 5039 5040 void 5041 bwait(struct buf *bp, u_char pri, const char *wchan) 5042 { 5043 struct mtx *mtxp; 5044 5045 mtxp = mtx_pool_find(mtxpool_sleep, bp); 5046 mtx_lock(mtxp); 5047 while ((bp->b_flags & B_DONE) == 0) 5048 msleep(bp, mtxp, pri, wchan, 0); 5049 mtx_unlock(mtxp); 5050 } 5051 5052 int 5053 bufsync(struct bufobj *bo, int waitfor) 5054 { 5055 5056 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread)); 5057 } 5058 5059 void 5060 bufstrategy(struct bufobj *bo, struct buf *bp) 5061 { 5062 int i __unused; 5063 struct vnode *vp; 5064 5065 vp = bp->b_vp; 5066 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 5067 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 5068 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 5069 i = VOP_STRATEGY(vp, bp); 5070 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 5071 } 5072 5073 /* 5074 * Initialize a struct bufobj before use. Memory is assumed zero filled. 5075 */ 5076 void 5077 bufobj_init(struct bufobj *bo, void *private) 5078 { 5079 static volatile int bufobj_cleanq; 5080 5081 bo->bo_domain = 5082 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains; 5083 rw_init(BO_LOCKPTR(bo), "bufobj interlock"); 5084 bo->bo_private = private; 5085 TAILQ_INIT(&bo->bo_clean.bv_hd); 5086 TAILQ_INIT(&bo->bo_dirty.bv_hd); 5087 } 5088 5089 void 5090 bufobj_wrefl(struct bufobj *bo) 5091 { 5092 5093 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 5094 ASSERT_BO_WLOCKED(bo); 5095 bo->bo_numoutput++; 5096 } 5097 5098 void 5099 bufobj_wref(struct bufobj *bo) 5100 { 5101 5102 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 5103 BO_LOCK(bo); 5104 bo->bo_numoutput++; 5105 BO_UNLOCK(bo); 5106 } 5107 5108 void 5109 bufobj_wdrop(struct bufobj *bo) 5110 { 5111 5112 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 5113 BO_LOCK(bo); 5114 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 5115 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 5116 bo->bo_flag &= ~BO_WWAIT; 5117 wakeup(&bo->bo_numoutput); 5118 } 5119 BO_UNLOCK(bo); 5120 } 5121 5122 int 5123 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 5124 { 5125 int error; 5126 5127 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 5128 ASSERT_BO_WLOCKED(bo); 5129 error = 0; 5130 while (bo->bo_numoutput) { 5131 bo->bo_flag |= BO_WWAIT; 5132 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 5133 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 5134 if (error) 5135 break; 5136 } 5137 return (error); 5138 } 5139 5140 /* 5141 * Set bio_data or bio_ma for struct bio from the struct buf. 5142 */ 5143 void 5144 bdata2bio(struct buf *bp, struct bio *bip) 5145 { 5146 5147 if (!buf_mapped(bp)) { 5148 KASSERT(unmapped_buf_allowed, ("unmapped")); 5149 bip->bio_ma = bp->b_pages; 5150 bip->bio_ma_n = bp->b_npages; 5151 bip->bio_data = unmapped_buf; 5152 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 5153 bip->bio_flags |= BIO_UNMAPPED; 5154 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 5155 PAGE_SIZE == bp->b_npages, 5156 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 5157 (long long)bip->bio_length, bip->bio_ma_n)); 5158 } else { 5159 bip->bio_data = bp->b_data; 5160 bip->bio_ma = NULL; 5161 } 5162 } 5163 5164 /* 5165 * The MIPS pmap code currently doesn't handle aliased pages. 5166 * The VIPT caches may not handle page aliasing themselves, leading 5167 * to data corruption. 5168 * 5169 * As such, this code makes a system extremely unhappy if said 5170 * system doesn't support unaliasing the above situation in hardware. 5171 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable 5172 * this feature at build time, so it has to be handled in software. 5173 * 5174 * Once the MIPS pmap/cache code grows to support this function on 5175 * earlier chips, it should be flipped back off. 5176 */ 5177 #ifdef __mips__ 5178 static int buf_pager_relbuf = 1; 5179 #else 5180 static int buf_pager_relbuf = 0; 5181 #endif 5182 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN, 5183 &buf_pager_relbuf, 0, 5184 "Make buffer pager release buffers after reading"); 5185 5186 /* 5187 * The buffer pager. It uses buffer reads to validate pages. 5188 * 5189 * In contrast to the generic local pager from vm/vnode_pager.c, this 5190 * pager correctly and easily handles volumes where the underlying 5191 * device block size is greater than the machine page size. The 5192 * buffer cache transparently extends the requested page run to be 5193 * aligned at the block boundary, and does the necessary bogus page 5194 * replacements in the addends to avoid obliterating already valid 5195 * pages. 5196 * 5197 * The only non-trivial issue is that the exclusive busy state for 5198 * pages, which is assumed by the vm_pager_getpages() interface, is 5199 * incompatible with the VMIO buffer cache's desire to share-busy the 5200 * pages. This function performs a trivial downgrade of the pages' 5201 * state before reading buffers, and a less trivial upgrade from the 5202 * shared-busy to excl-busy state after the read. 5203 */ 5204 int 5205 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count, 5206 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno, 5207 vbg_get_blksize_t get_blksize) 5208 { 5209 vm_page_t m; 5210 vm_object_t object; 5211 struct buf *bp; 5212 struct mount *mp; 5213 daddr_t lbn, lbnp; 5214 vm_ooffset_t la, lb, poff, poffe; 5215 long bo_bs, bsize; 5216 int br_flags, error, i, pgsin, pgsin_a, pgsin_b; 5217 bool redo, lpart; 5218 5219 object = vp->v_object; 5220 mp = vp->v_mount; 5221 error = 0; 5222 la = IDX_TO_OFF(ma[count - 1]->pindex); 5223 if (la >= object->un_pager.vnp.vnp_size) 5224 return (VM_PAGER_BAD); 5225 5226 /* 5227 * Change the meaning of la from where the last requested page starts 5228 * to where it ends, because that's the end of the requested region 5229 * and the start of the potential read-ahead region. 5230 */ 5231 la += PAGE_SIZE; 5232 lpart = la > object->un_pager.vnp.vnp_size; 5233 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)), 5234 &bo_bs); 5235 if (error != 0) 5236 return (VM_PAGER_ERROR); 5237 5238 /* 5239 * Calculate read-ahead, behind and total pages. 5240 */ 5241 pgsin = count; 5242 lb = IDX_TO_OFF(ma[0]->pindex); 5243 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs)); 5244 pgsin += pgsin_b; 5245 if (rbehind != NULL) 5246 *rbehind = pgsin_b; 5247 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la); 5248 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size) 5249 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size, 5250 PAGE_SIZE) - la); 5251 pgsin += pgsin_a; 5252 if (rahead != NULL) 5253 *rahead = pgsin_a; 5254 VM_CNT_INC(v_vnodein); 5255 VM_CNT_ADD(v_vnodepgsin, pgsin); 5256 5257 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) 5258 != 0) ? GB_UNMAPPED : 0; 5259 again: 5260 for (i = 0; i < count; i++) { 5261 if (ma[i] != bogus_page) 5262 vm_page_busy_downgrade(ma[i]); 5263 } 5264 5265 lbnp = -1; 5266 for (i = 0; i < count; i++) { 5267 m = ma[i]; 5268 if (m == bogus_page) 5269 continue; 5270 5271 /* 5272 * Pages are shared busy and the object lock is not 5273 * owned, which together allow for the pages' 5274 * invalidation. The racy test for validity avoids 5275 * useless creation of the buffer for the most typical 5276 * case when invalidation is not used in redo or for 5277 * parallel read. The shared->excl upgrade loop at 5278 * the end of the function catches the race in a 5279 * reliable way (protected by the object lock). 5280 */ 5281 if (vm_page_all_valid(m)) 5282 continue; 5283 5284 poff = IDX_TO_OFF(m->pindex); 5285 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size); 5286 for (; poff < poffe; poff += bsize) { 5287 lbn = get_lblkno(vp, poff); 5288 if (lbn == lbnp) 5289 goto next_page; 5290 lbnp = lbn; 5291 5292 error = get_blksize(vp, lbn, &bsize); 5293 if (error == 0) 5294 error = bread_gb(vp, lbn, bsize, 5295 curthread->td_ucred, br_flags, &bp); 5296 if (error != 0) 5297 goto end_pages; 5298 if (bp->b_rcred == curthread->td_ucred) { 5299 crfree(bp->b_rcred); 5300 bp->b_rcred = NOCRED; 5301 } 5302 if (LIST_EMPTY(&bp->b_dep)) { 5303 /* 5304 * Invalidation clears m->valid, but 5305 * may leave B_CACHE flag if the 5306 * buffer existed at the invalidation 5307 * time. In this case, recycle the 5308 * buffer to do real read on next 5309 * bread() after redo. 5310 * 5311 * Otherwise B_RELBUF is not strictly 5312 * necessary, enable to reduce buf 5313 * cache pressure. 5314 */ 5315 if (buf_pager_relbuf || 5316 !vm_page_all_valid(m)) 5317 bp->b_flags |= B_RELBUF; 5318 5319 bp->b_flags &= ~B_NOCACHE; 5320 brelse(bp); 5321 } else { 5322 bqrelse(bp); 5323 } 5324 } 5325 KASSERT(1 /* racy, enable for debugging */ || 5326 vm_page_all_valid(m) || i == count - 1, 5327 ("buf %d %p invalid", i, m)); 5328 if (i == count - 1 && lpart) { 5329 if (!vm_page_none_valid(m) && 5330 !vm_page_all_valid(m)) 5331 vm_page_zero_invalid(m, TRUE); 5332 } 5333 next_page:; 5334 } 5335 end_pages: 5336 5337 redo = false; 5338 for (i = 0; i < count; i++) { 5339 if (ma[i] == bogus_page) 5340 continue; 5341 if (vm_page_busy_tryupgrade(ma[i]) == 0) { 5342 vm_page_sunbusy(ma[i]); 5343 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex, 5344 VM_ALLOC_NORMAL); 5345 } 5346 5347 /* 5348 * Since the pages were only sbusy while neither the 5349 * buffer nor the object lock was held by us, or 5350 * reallocated while vm_page_grab() slept for busy 5351 * relinguish, they could have been invalidated. 5352 * Recheck the valid bits and re-read as needed. 5353 * 5354 * Note that the last page is made fully valid in the 5355 * read loop, and partial validity for the page at 5356 * index count - 1 could mean that the page was 5357 * invalidated or removed, so we must restart for 5358 * safety as well. 5359 */ 5360 if (!vm_page_all_valid(ma[i])) 5361 redo = true; 5362 } 5363 if (redo && error == 0) 5364 goto again; 5365 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); 5366 } 5367 5368 #include "opt_ddb.h" 5369 #ifdef DDB 5370 #include <ddb/ddb.h> 5371 5372 /* DDB command to show buffer data */ 5373 DB_SHOW_COMMAND(buffer, db_show_buffer) 5374 { 5375 /* get args */ 5376 struct buf *bp = (struct buf *)addr; 5377 #ifdef FULL_BUF_TRACKING 5378 uint32_t i, j; 5379 #endif 5380 5381 if (!have_addr) { 5382 db_printf("usage: show buffer <addr>\n"); 5383 return; 5384 } 5385 5386 db_printf("buf at %p\n", bp); 5387 db_printf("b_flags = 0x%b, b_xflags=0x%b\n", 5388 (u_int)bp->b_flags, PRINT_BUF_FLAGS, 5389 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS); 5390 db_printf("b_vflags=0x%b b_ioflags0x%b\n", 5391 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS, 5392 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS); 5393 db_printf( 5394 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 5395 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, " 5396 "b_vp = %p, b_dep = %p\n", 5397 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 5398 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 5399 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first); 5400 db_printf("b_kvabase = %p, b_kvasize = %d\n", 5401 bp->b_kvabase, bp->b_kvasize); 5402 if (bp->b_npages) { 5403 int i; 5404 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 5405 for (i = 0; i < bp->b_npages; i++) { 5406 vm_page_t m; 5407 m = bp->b_pages[i]; 5408 if (m != NULL) 5409 db_printf("(%p, 0x%lx, 0x%lx)", m->object, 5410 (u_long)m->pindex, 5411 (u_long)VM_PAGE_TO_PHYS(m)); 5412 else 5413 db_printf("( ??? )"); 5414 if ((i + 1) < bp->b_npages) 5415 db_printf(","); 5416 } 5417 db_printf("\n"); 5418 } 5419 BUF_LOCKPRINTINFO(bp); 5420 #if defined(FULL_BUF_TRACKING) 5421 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt); 5422 5423 i = bp->b_io_tcnt % BUF_TRACKING_SIZE; 5424 for (j = 1; j <= BUF_TRACKING_SIZE; j++) { 5425 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL) 5426 continue; 5427 db_printf(" %2u: %s\n", j, 5428 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]); 5429 } 5430 #elif defined(BUF_TRACKING) 5431 db_printf("b_io_tracking: %s\n", bp->b_io_tracking); 5432 #endif 5433 db_printf(" "); 5434 } 5435 5436 DB_SHOW_COMMAND(bufqueues, bufqueues) 5437 { 5438 struct bufdomain *bd; 5439 struct buf *bp; 5440 long total; 5441 int i, j, cnt; 5442 5443 db_printf("bqempty: %d\n", bqempty.bq_len); 5444 5445 for (i = 0; i < buf_domains; i++) { 5446 bd = &bdomain[i]; 5447 db_printf("Buf domain %d\n", i); 5448 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers); 5449 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers); 5450 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers); 5451 db_printf("\n"); 5452 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace); 5453 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace); 5454 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace); 5455 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace); 5456 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh); 5457 db_printf("\n"); 5458 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers); 5459 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers); 5460 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers); 5461 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh); 5462 db_printf("\n"); 5463 total = 0; 5464 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist) 5465 total += bp->b_bufsize; 5466 db_printf("\tcleanq count\t%d (%ld)\n", 5467 bd->bd_cleanq->bq_len, total); 5468 total = 0; 5469 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist) 5470 total += bp->b_bufsize; 5471 db_printf("\tdirtyq count\t%d (%ld)\n", 5472 bd->bd_dirtyq.bq_len, total); 5473 db_printf("\twakeup\t\t%d\n", bd->bd_wanted); 5474 db_printf("\tlim\t\t%d\n", bd->bd_lim); 5475 db_printf("\tCPU "); 5476 for (j = 0; j <= mp_maxid; j++) 5477 db_printf("%d, ", bd->bd_subq[j].bq_len); 5478 db_printf("\n"); 5479 cnt = 0; 5480 total = 0; 5481 for (j = 0; j < nbuf; j++) { 5482 bp = nbufp(j); 5483 if (bp->b_domain == i && BUF_ISLOCKED(bp)) { 5484 cnt++; 5485 total += bp->b_bufsize; 5486 } 5487 } 5488 db_printf("\tLocked buffers: %d space %ld\n", cnt, total); 5489 cnt = 0; 5490 total = 0; 5491 for (j = 0; j < nbuf; j++) { 5492 bp = nbufp(j); 5493 if (bp->b_domain == i) { 5494 cnt++; 5495 total += bp->b_bufsize; 5496 } 5497 } 5498 db_printf("\tTotal buffers: %d space %ld\n", cnt, total); 5499 } 5500 } 5501 5502 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 5503 { 5504 struct buf *bp; 5505 int i; 5506 5507 for (i = 0; i < nbuf; i++) { 5508 bp = nbufp(i); 5509 if (BUF_ISLOCKED(bp)) { 5510 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5511 db_printf("\n"); 5512 if (db_pager_quit) 5513 break; 5514 } 5515 } 5516 } 5517 5518 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 5519 { 5520 struct vnode *vp; 5521 struct buf *bp; 5522 5523 if (!have_addr) { 5524 db_printf("usage: show vnodebufs <addr>\n"); 5525 return; 5526 } 5527 vp = (struct vnode *)addr; 5528 db_printf("Clean buffers:\n"); 5529 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 5530 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5531 db_printf("\n"); 5532 } 5533 db_printf("Dirty buffers:\n"); 5534 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 5535 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5536 db_printf("\n"); 5537 } 5538 } 5539 5540 DB_COMMAND(countfreebufs, db_coundfreebufs) 5541 { 5542 struct buf *bp; 5543 int i, used = 0, nfree = 0; 5544 5545 if (have_addr) { 5546 db_printf("usage: countfreebufs\n"); 5547 return; 5548 } 5549 5550 for (i = 0; i < nbuf; i++) { 5551 bp = nbufp(i); 5552 if (bp->b_qindex == QUEUE_EMPTY) 5553 nfree++; 5554 else 5555 used++; 5556 } 5557 5558 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 5559 nfree + used); 5560 db_printf("numfreebuffers is %d\n", numfreebuffers); 5561 } 5562 #endif /* DDB */ 5563