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/boottrace.h> 56 #include <sys/buf.h> 57 #include <sys/conf.h> 58 #include <sys/counter.h> 59 #include <sys/devicestat.h> 60 #include <sys/eventhandler.h> 61 #include <sys/fail.h> 62 #include <sys/ktr.h> 63 #include <sys/limits.h> 64 #include <sys/lock.h> 65 #include <sys/malloc.h> 66 #include <sys/mount.h> 67 #include <sys/mutex.h> 68 #include <sys/kernel.h> 69 #include <sys/kthread.h> 70 #include <sys/proc.h> 71 #include <sys/racct.h> 72 #include <sys/refcount.h> 73 #include <sys/resourcevar.h> 74 #include <sys/rwlock.h> 75 #include <sys/sched.h> 76 #include <sys/smp.h> 77 #include <sys/sysctl.h> 78 #include <sys/syscallsubr.h> 79 #include <sys/vmem.h> 80 #include <sys/vmmeter.h> 81 #include <sys/vnode.h> 82 #include <sys/watchdog.h> 83 #include <geom/geom.h> 84 #include <vm/vm.h> 85 #include <vm/vm_param.h> 86 #include <vm/vm_kern.h> 87 #include <vm/vm_object.h> 88 #include <vm/vm_page.h> 89 #include <vm/vm_pageout.h> 90 #include <vm/vm_pager.h> 91 #include <vm/vm_extern.h> 92 #include <vm/vm_map.h> 93 #include <vm/swap_pager.h> 94 95 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 96 97 struct bio_ops bioops; /* I/O operation notification */ 98 99 struct buf_ops buf_ops_bio = { 100 .bop_name = "buf_ops_bio", 101 .bop_write = bufwrite, 102 .bop_strategy = bufstrategy, 103 .bop_sync = bufsync, 104 .bop_bdflush = bufbdflush, 105 }; 106 107 struct bufqueue { 108 struct mtx_padalign bq_lock; 109 TAILQ_HEAD(, buf) bq_queue; 110 uint8_t bq_index; 111 uint16_t bq_subqueue; 112 int bq_len; 113 } __aligned(CACHE_LINE_SIZE); 114 115 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock) 116 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq))) 117 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq))) 118 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED) 119 120 struct bufdomain { 121 struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */ 122 struct bufqueue bd_dirtyq; 123 struct bufqueue *bd_cleanq; 124 struct mtx_padalign bd_run_lock; 125 /* Constants */ 126 long bd_maxbufspace; 127 long bd_hibufspace; 128 long bd_lobufspace; 129 long bd_bufspacethresh; 130 int bd_hifreebuffers; 131 int bd_lofreebuffers; 132 int bd_hidirtybuffers; 133 int bd_lodirtybuffers; 134 int bd_dirtybufthresh; 135 int bd_lim; 136 /* atomics */ 137 int bd_wanted; 138 bool bd_shutdown; 139 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers; 140 int __aligned(CACHE_LINE_SIZE) bd_running; 141 long __aligned(CACHE_LINE_SIZE) bd_bufspace; 142 int __aligned(CACHE_LINE_SIZE) bd_freebuffers; 143 } __aligned(CACHE_LINE_SIZE); 144 145 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock) 146 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd))) 147 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd))) 148 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED) 149 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock) 150 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd))) 151 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd))) 152 #define BD_DOMAIN(bd) (bd - bdomain) 153 154 static char *buf; /* buffer header pool */ 155 static struct buf * 156 nbufp(unsigned i) 157 { 158 return ((struct buf *)(buf + (sizeof(struct buf) + 159 sizeof(vm_page_t) * atop(maxbcachebuf)) * i)); 160 } 161 162 caddr_t __read_mostly unmapped_buf; 163 164 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */ 165 struct proc *bufdaemonproc; 166 167 static void vm_hold_free_pages(struct buf *bp, int newbsize); 168 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 169 vm_offset_t to); 170 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 171 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 172 vm_page_t m); 173 static void vfs_clean_pages_dirty_buf(struct buf *bp); 174 static void vfs_setdirty_range(struct buf *bp); 175 static void vfs_vmio_invalidate(struct buf *bp); 176 static void vfs_vmio_truncate(struct buf *bp, int npages); 177 static void vfs_vmio_extend(struct buf *bp, int npages, int size); 178 static int vfs_bio_clcheck(struct vnode *vp, int size, 179 daddr_t lblkno, daddr_t blkno); 180 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int, 181 void (*)(struct buf *)); 182 static int buf_flush(struct vnode *vp, struct bufdomain *, int); 183 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int); 184 static void buf_daemon(void); 185 static __inline void bd_wakeup(void); 186 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS); 187 static void bufkva_reclaim(vmem_t *, int); 188 static void bufkva_free(struct buf *); 189 static int buf_import(void *, void **, int, int, int); 190 static void buf_release(void *, void **, int); 191 static void maxbcachebuf_adjust(void); 192 static inline struct bufdomain *bufdomain(struct buf *); 193 static void bq_remove(struct bufqueue *bq, struct buf *bp); 194 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock); 195 static int buf_recycle(struct bufdomain *, bool kva); 196 static void bq_init(struct bufqueue *bq, int qindex, int cpu, 197 const char *lockname); 198 static void bd_init(struct bufdomain *bd); 199 static int bd_flushall(struct bufdomain *bd); 200 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS); 201 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS); 202 203 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 204 int vmiodirenable = TRUE; 205 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 206 "Use the VM system for directory writes"); 207 long runningbufspace; 208 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 209 "Amount of presently outstanding async buffer io"); 210 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 211 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers"); 212 static counter_u64_t bufkvaspace; 213 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, 214 "Kernel virtual memory used for buffers"); 215 static long maxbufspace; 216 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace, 217 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace, 218 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L", 219 "Maximum allowed value of bufspace (including metadata)"); 220 static long bufmallocspace; 221 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 222 "Amount of malloced memory for buffers"); 223 static long maxbufmallocspace; 224 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 225 0, "Maximum amount of malloced memory for buffers"); 226 static long lobufspace; 227 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace, 228 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace, 229 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L", 230 "Minimum amount of buffers we want to have"); 231 long hibufspace; 232 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace, 233 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace, 234 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L", 235 "Maximum allowed value of bufspace (excluding metadata)"); 236 long bufspacethresh; 237 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh, 238 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh, 239 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L", 240 "Bufspace consumed before waking the daemon to free some"); 241 static counter_u64_t buffreekvacnt; 242 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 243 "Number of times we have freed the KVA space from some buffer"); 244 static counter_u64_t bufdefragcnt; 245 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 246 "Number of times we have had to repeat buffer allocation to defragment"); 247 static long lorunningspace; 248 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 249 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L", 250 "Minimum preferred space used for in-progress I/O"); 251 static long hirunningspace; 252 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 253 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L", 254 "Maximum amount of space to use for in-progress I/O"); 255 int dirtybufferflushes; 256 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 257 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 258 int bdwriteskip; 259 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 260 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 261 int altbufferflushes; 262 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS, 263 &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers"); 264 static int recursiveflushes; 265 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS, 266 &recursiveflushes, 0, "Number of flushes skipped due to being recursive"); 267 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS); 268 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers, 269 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I", 270 "Number of buffers that are dirty (has unwritten changes) at the moment"); 271 static int lodirtybuffers; 272 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers, 273 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers, 274 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I", 275 "How many buffers we want to have free before bufdaemon can sleep"); 276 static int hidirtybuffers; 277 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers, 278 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers, 279 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I", 280 "When the number of dirty buffers is considered severe"); 281 int dirtybufthresh; 282 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh, 283 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh, 284 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I", 285 "Number of bdwrite to bawrite conversions to clear dirty buffers"); 286 static int numfreebuffers; 287 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 288 "Number of free buffers"); 289 static int lofreebuffers; 290 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers, 291 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers, 292 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I", 293 "Target number of free buffers"); 294 static int hifreebuffers; 295 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers, 296 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers, 297 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I", 298 "Threshold for clean buffer recycling"); 299 static counter_u64_t getnewbufcalls; 300 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, 301 &getnewbufcalls, "Number of calls to getnewbuf"); 302 static counter_u64_t getnewbufrestarts; 303 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, 304 &getnewbufrestarts, 305 "Number of times getnewbuf has had to restart a buffer acquisition"); 306 static counter_u64_t mappingrestarts; 307 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD, 308 &mappingrestarts, 309 "Number of times getblk has had to restart a buffer mapping for " 310 "unmapped buffer"); 311 static counter_u64_t numbufallocfails; 312 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW, 313 &numbufallocfails, "Number of times buffer allocations failed"); 314 static int flushbufqtarget = 100; 315 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 316 "Amount of work to do in flushbufqueues when helping bufdaemon"); 317 static counter_u64_t notbufdflushes; 318 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 319 "Number of dirty buffer flushes done by the bufdaemon helpers"); 320 static long barrierwrites; 321 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS, 322 &barrierwrites, 0, "Number of barrier writes"); 323 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, 324 &unmapped_buf_allowed, 0, 325 "Permit the use of the unmapped i/o"); 326 int maxbcachebuf = MAXBCACHEBUF; 327 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0, 328 "Maximum size of a buffer cache block"); 329 330 /* 331 * This lock synchronizes access to bd_request. 332 */ 333 static struct mtx_padalign __exclusive_cache_line bdlock; 334 335 /* 336 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 337 * waitrunningbufspace(). 338 */ 339 static struct mtx_padalign __exclusive_cache_line rbreqlock; 340 341 /* 342 * Lock that protects bdirtywait. 343 */ 344 static struct mtx_padalign __exclusive_cache_line bdirtylock; 345 346 /* 347 * bufdaemon shutdown request and sleep channel. 348 */ 349 static bool bd_shutdown; 350 351 /* 352 * Wakeup point for bufdaemon, as well as indicator of whether it is already 353 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 354 * is idling. 355 */ 356 static int bd_request; 357 358 /* 359 * Request for the buf daemon to write more buffers than is indicated by 360 * lodirtybuf. This may be necessary to push out excess dependencies or 361 * defragment the address space where a simple count of the number of dirty 362 * buffers is insufficient to characterize the demand for flushing them. 363 */ 364 static int bd_speedupreq; 365 366 /* 367 * Synchronization (sleep/wakeup) variable for active buffer space requests. 368 * Set when wait starts, cleared prior to wakeup(). 369 * Used in runningbufwakeup() and waitrunningbufspace(). 370 */ 371 static int runningbufreq; 372 373 /* 374 * Synchronization for bwillwrite() waiters. 375 */ 376 static int bdirtywait; 377 378 /* 379 * Definitions for the buffer free lists. 380 */ 381 #define QUEUE_NONE 0 /* on no queue */ 382 #define QUEUE_EMPTY 1 /* empty buffer headers */ 383 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 384 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */ 385 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */ 386 387 /* Maximum number of buffer domains. */ 388 #define BUF_DOMAINS 8 389 390 struct bufdomainset bdlodirty; /* Domains > lodirty */ 391 struct bufdomainset bdhidirty; /* Domains > hidirty */ 392 393 /* Configured number of clean queues. */ 394 static int __read_mostly buf_domains; 395 396 BITSET_DEFINE(bufdomainset, BUF_DOMAINS); 397 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS]; 398 struct bufqueue __exclusive_cache_line bqempty; 399 400 /* 401 * per-cpu empty buffer cache. 402 */ 403 uma_zone_t buf_zone; 404 405 static int 406 sysctl_runningspace(SYSCTL_HANDLER_ARGS) 407 { 408 long value; 409 int error; 410 411 value = *(long *)arg1; 412 error = sysctl_handle_long(oidp, &value, 0, req); 413 if (error != 0 || req->newptr == NULL) 414 return (error); 415 mtx_lock(&rbreqlock); 416 if (arg1 == &hirunningspace) { 417 if (value < lorunningspace) 418 error = EINVAL; 419 else 420 hirunningspace = value; 421 } else { 422 KASSERT(arg1 == &lorunningspace, 423 ("%s: unknown arg1", __func__)); 424 if (value > hirunningspace) 425 error = EINVAL; 426 else 427 lorunningspace = value; 428 } 429 mtx_unlock(&rbreqlock); 430 return (error); 431 } 432 433 static int 434 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS) 435 { 436 int error; 437 int value; 438 int i; 439 440 value = *(int *)arg1; 441 error = sysctl_handle_int(oidp, &value, 0, req); 442 if (error != 0 || req->newptr == NULL) 443 return (error); 444 *(int *)arg1 = value; 445 for (i = 0; i < buf_domains; i++) 446 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = 447 value / buf_domains; 448 449 return (error); 450 } 451 452 static int 453 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS) 454 { 455 long value; 456 int error; 457 int i; 458 459 value = *(long *)arg1; 460 error = sysctl_handle_long(oidp, &value, 0, req); 461 if (error != 0 || req->newptr == NULL) 462 return (error); 463 *(long *)arg1 = value; 464 for (i = 0; i < buf_domains; i++) 465 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = 466 value / buf_domains; 467 468 return (error); 469 } 470 471 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 472 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 473 static int 474 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 475 { 476 long lvalue; 477 int ivalue; 478 int i; 479 480 lvalue = 0; 481 for (i = 0; i < buf_domains; i++) 482 lvalue += bdomain[i].bd_bufspace; 483 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 484 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 485 if (lvalue > INT_MAX) 486 /* On overflow, still write out a long to trigger ENOMEM. */ 487 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 488 ivalue = lvalue; 489 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 490 } 491 #else 492 static int 493 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 494 { 495 long lvalue; 496 int i; 497 498 lvalue = 0; 499 for (i = 0; i < buf_domains; i++) 500 lvalue += bdomain[i].bd_bufspace; 501 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 502 } 503 #endif 504 505 static int 506 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS) 507 { 508 int value; 509 int i; 510 511 value = 0; 512 for (i = 0; i < buf_domains; i++) 513 value += bdomain[i].bd_numdirtybuffers; 514 return (sysctl_handle_int(oidp, &value, 0, req)); 515 } 516 517 /* 518 * bdirtywakeup: 519 * 520 * Wakeup any bwillwrite() waiters. 521 */ 522 static void 523 bdirtywakeup(void) 524 { 525 mtx_lock(&bdirtylock); 526 if (bdirtywait) { 527 bdirtywait = 0; 528 wakeup(&bdirtywait); 529 } 530 mtx_unlock(&bdirtylock); 531 } 532 533 /* 534 * bd_clear: 535 * 536 * Clear a domain from the appropriate bitsets when dirtybuffers 537 * is decremented. 538 */ 539 static void 540 bd_clear(struct bufdomain *bd) 541 { 542 543 mtx_lock(&bdirtylock); 544 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers) 545 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); 546 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers) 547 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); 548 mtx_unlock(&bdirtylock); 549 } 550 551 /* 552 * bd_set: 553 * 554 * Set a domain in the appropriate bitsets when dirtybuffers 555 * is incremented. 556 */ 557 static void 558 bd_set(struct bufdomain *bd) 559 { 560 561 mtx_lock(&bdirtylock); 562 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers) 563 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); 564 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers) 565 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); 566 mtx_unlock(&bdirtylock); 567 } 568 569 /* 570 * bdirtysub: 571 * 572 * Decrement the numdirtybuffers count by one and wakeup any 573 * threads blocked in bwillwrite(). 574 */ 575 static void 576 bdirtysub(struct buf *bp) 577 { 578 struct bufdomain *bd; 579 int num; 580 581 bd = bufdomain(bp); 582 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1); 583 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) 584 bdirtywakeup(); 585 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) 586 bd_clear(bd); 587 } 588 589 /* 590 * bdirtyadd: 591 * 592 * Increment the numdirtybuffers count by one and wakeup the buf 593 * daemon if needed. 594 */ 595 static void 596 bdirtyadd(struct buf *bp) 597 { 598 struct bufdomain *bd; 599 int num; 600 601 /* 602 * Only do the wakeup once as we cross the boundary. The 603 * buf daemon will keep running until the condition clears. 604 */ 605 bd = bufdomain(bp); 606 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1); 607 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) 608 bd_wakeup(); 609 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) 610 bd_set(bd); 611 } 612 613 /* 614 * bufspace_daemon_wakeup: 615 * 616 * Wakeup the daemons responsible for freeing clean bufs. 617 */ 618 static void 619 bufspace_daemon_wakeup(struct bufdomain *bd) 620 { 621 622 /* 623 * avoid the lock if the daemon is running. 624 */ 625 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) { 626 BD_RUN_LOCK(bd); 627 atomic_store_int(&bd->bd_running, 1); 628 wakeup(&bd->bd_running); 629 BD_RUN_UNLOCK(bd); 630 } 631 } 632 633 /* 634 * bufspace_adjust: 635 * 636 * Adjust the reported bufspace for a KVA managed buffer, possibly 637 * waking any waiters. 638 */ 639 static void 640 bufspace_adjust(struct buf *bp, int bufsize) 641 { 642 struct bufdomain *bd; 643 long space; 644 int diff; 645 646 KASSERT((bp->b_flags & B_MALLOC) == 0, 647 ("bufspace_adjust: malloc buf %p", bp)); 648 bd = bufdomain(bp); 649 diff = bufsize - bp->b_bufsize; 650 if (diff < 0) { 651 atomic_subtract_long(&bd->bd_bufspace, -diff); 652 } else if (diff > 0) { 653 space = atomic_fetchadd_long(&bd->bd_bufspace, diff); 654 /* Wake up the daemon on the transition. */ 655 if (space < bd->bd_bufspacethresh && 656 space + diff >= bd->bd_bufspacethresh) 657 bufspace_daemon_wakeup(bd); 658 } 659 bp->b_bufsize = bufsize; 660 } 661 662 /* 663 * bufspace_reserve: 664 * 665 * Reserve bufspace before calling allocbuf(). metadata has a 666 * different space limit than data. 667 */ 668 static int 669 bufspace_reserve(struct bufdomain *bd, int size, bool metadata) 670 { 671 long limit, new; 672 long space; 673 674 if (metadata) 675 limit = bd->bd_maxbufspace; 676 else 677 limit = bd->bd_hibufspace; 678 space = atomic_fetchadd_long(&bd->bd_bufspace, size); 679 new = space + size; 680 if (new > limit) { 681 atomic_subtract_long(&bd->bd_bufspace, size); 682 return (ENOSPC); 683 } 684 685 /* Wake up the daemon on the transition. */ 686 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh) 687 bufspace_daemon_wakeup(bd); 688 689 return (0); 690 } 691 692 /* 693 * bufspace_release: 694 * 695 * Release reserved bufspace after bufspace_adjust() has consumed it. 696 */ 697 static void 698 bufspace_release(struct bufdomain *bd, int size) 699 { 700 701 atomic_subtract_long(&bd->bd_bufspace, size); 702 } 703 704 /* 705 * bufspace_wait: 706 * 707 * Wait for bufspace, acting as the buf daemon if a locked vnode is 708 * supplied. bd_wanted must be set prior to polling for space. The 709 * operation must be re-tried on return. 710 */ 711 static void 712 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags, 713 int slpflag, int slptimeo) 714 { 715 struct thread *td; 716 int error, fl, norunbuf; 717 718 if ((gbflags & GB_NOWAIT_BD) != 0) 719 return; 720 721 td = curthread; 722 BD_LOCK(bd); 723 while (bd->bd_wanted) { 724 if (vp != NULL && vp->v_type != VCHR && 725 (td->td_pflags & TDP_BUFNEED) == 0) { 726 BD_UNLOCK(bd); 727 /* 728 * getblk() is called with a vnode locked, and 729 * some majority of the dirty buffers may as 730 * well belong to the vnode. Flushing the 731 * buffers there would make a progress that 732 * cannot be achieved by the buf_daemon, that 733 * cannot lock the vnode. 734 */ 735 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 736 (td->td_pflags & TDP_NORUNNINGBUF); 737 738 /* 739 * Play bufdaemon. The getnewbuf() function 740 * may be called while the thread owns lock 741 * for another dirty buffer for the same 742 * vnode, which makes it impossible to use 743 * VOP_FSYNC() there, due to the buffer lock 744 * recursion. 745 */ 746 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 747 fl = buf_flush(vp, bd, flushbufqtarget); 748 td->td_pflags &= norunbuf; 749 BD_LOCK(bd); 750 if (fl != 0) 751 continue; 752 if (bd->bd_wanted == 0) 753 break; 754 } 755 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd), 756 (PRIBIO + 4) | slpflag, "newbuf", slptimeo); 757 if (error != 0) 758 break; 759 } 760 BD_UNLOCK(bd); 761 } 762 763 static void 764 bufspace_daemon_shutdown(void *arg, int howto __unused) 765 { 766 struct bufdomain *bd = arg; 767 int error; 768 769 if (KERNEL_PANICKED()) 770 return; 771 772 BD_RUN_LOCK(bd); 773 bd->bd_shutdown = true; 774 wakeup(&bd->bd_running); 775 error = msleep(&bd->bd_shutdown, BD_RUN_LOCKPTR(bd), 0, 776 "bufspace_shutdown", 60 * hz); 777 BD_RUN_UNLOCK(bd); 778 if (error != 0) 779 printf("bufspacedaemon wait error: %d\n", error); 780 } 781 782 /* 783 * bufspace_daemon: 784 * 785 * buffer space management daemon. Tries to maintain some marginal 786 * amount of free buffer space so that requesting processes neither 787 * block nor work to reclaim buffers. 788 */ 789 static void 790 bufspace_daemon(void *arg) 791 { 792 struct bufdomain *bd = arg; 793 794 EVENTHANDLER_REGISTER(shutdown_pre_sync, bufspace_daemon_shutdown, bd, 795 SHUTDOWN_PRI_LAST + 100); 796 797 BD_RUN_LOCK(bd); 798 while (!bd->bd_shutdown) { 799 BD_RUN_UNLOCK(bd); 800 801 /* 802 * Free buffers from the clean queue until we meet our 803 * targets. 804 * 805 * Theory of operation: The buffer cache is most efficient 806 * when some free buffer headers and space are always 807 * available to getnewbuf(). This daemon attempts to prevent 808 * the excessive blocking and synchronization associated 809 * with shortfall. It goes through three phases according 810 * demand: 811 * 812 * 1) The daemon wakes up voluntarily once per-second 813 * during idle periods when the counters are below 814 * the wakeup thresholds (bufspacethresh, lofreebuffers). 815 * 816 * 2) The daemon wakes up as we cross the thresholds 817 * ahead of any potential blocking. This may bounce 818 * slightly according to the rate of consumption and 819 * release. 820 * 821 * 3) The daemon and consumers are starved for working 822 * clean buffers. This is the 'bufspace' sleep below 823 * which will inefficiently trade bufs with bqrelse 824 * until we return to condition 2. 825 */ 826 while (bd->bd_bufspace > bd->bd_lobufspace || 827 bd->bd_freebuffers < bd->bd_hifreebuffers) { 828 if (buf_recycle(bd, false) != 0) { 829 if (bd_flushall(bd)) 830 continue; 831 /* 832 * Speedup dirty if we've run out of clean 833 * buffers. This is possible in particular 834 * because softdep may held many bufs locked 835 * pending writes to other bufs which are 836 * marked for delayed write, exhausting 837 * clean space until they are written. 838 */ 839 bd_speedup(); 840 BD_LOCK(bd); 841 if (bd->bd_wanted) { 842 msleep(&bd->bd_wanted, BD_LOCKPTR(bd), 843 PRIBIO|PDROP, "bufspace", hz/10); 844 } else 845 BD_UNLOCK(bd); 846 } 847 maybe_yield(); 848 } 849 850 /* 851 * Re-check our limits and sleep. bd_running must be 852 * cleared prior to checking the limits to avoid missed 853 * wakeups. The waker will adjust one of bufspace or 854 * freebuffers prior to checking bd_running. 855 */ 856 BD_RUN_LOCK(bd); 857 if (bd->bd_shutdown) 858 break; 859 atomic_store_int(&bd->bd_running, 0); 860 if (bd->bd_bufspace < bd->bd_bufspacethresh && 861 bd->bd_freebuffers > bd->bd_lofreebuffers) { 862 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), 863 PRIBIO, "-", hz); 864 } else { 865 /* Avoid spurious wakeups while running. */ 866 atomic_store_int(&bd->bd_running, 1); 867 } 868 } 869 wakeup(&bd->bd_shutdown); 870 BD_RUN_UNLOCK(bd); 871 kthread_exit(); 872 } 873 874 /* 875 * bufmallocadjust: 876 * 877 * Adjust the reported bufspace for a malloc managed buffer, possibly 878 * waking any waiters. 879 */ 880 static void 881 bufmallocadjust(struct buf *bp, int bufsize) 882 { 883 int diff; 884 885 KASSERT((bp->b_flags & B_MALLOC) != 0, 886 ("bufmallocadjust: non-malloc buf %p", bp)); 887 diff = bufsize - bp->b_bufsize; 888 if (diff < 0) 889 atomic_subtract_long(&bufmallocspace, -diff); 890 else 891 atomic_add_long(&bufmallocspace, diff); 892 bp->b_bufsize = bufsize; 893 } 894 895 /* 896 * runningwakeup: 897 * 898 * Wake up processes that are waiting on asynchronous writes to fall 899 * below lorunningspace. 900 */ 901 static void 902 runningwakeup(void) 903 { 904 905 mtx_lock(&rbreqlock); 906 if (runningbufreq) { 907 runningbufreq = 0; 908 wakeup(&runningbufreq); 909 } 910 mtx_unlock(&rbreqlock); 911 } 912 913 /* 914 * runningbufwakeup: 915 * 916 * Decrement the outstanding write count according. 917 */ 918 void 919 runningbufwakeup(struct buf *bp) 920 { 921 long space, bspace; 922 923 bspace = bp->b_runningbufspace; 924 if (bspace == 0) 925 return; 926 space = atomic_fetchadd_long(&runningbufspace, -bspace); 927 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", 928 space, bspace)); 929 bp->b_runningbufspace = 0; 930 /* 931 * Only acquire the lock and wakeup on the transition from exceeding 932 * the threshold to falling below it. 933 */ 934 if (space < lorunningspace) 935 return; 936 if (space - bspace > lorunningspace) 937 return; 938 runningwakeup(); 939 } 940 941 /* 942 * waitrunningbufspace() 943 * 944 * runningbufspace is a measure of the amount of I/O currently 945 * running. This routine is used in async-write situations to 946 * prevent creating huge backups of pending writes to a device. 947 * Only asynchronous writes are governed by this function. 948 * 949 * This does NOT turn an async write into a sync write. It waits 950 * for earlier writes to complete and generally returns before the 951 * caller's write has reached the device. 952 */ 953 void 954 waitrunningbufspace(void) 955 { 956 957 mtx_lock(&rbreqlock); 958 while (runningbufspace > hirunningspace) { 959 runningbufreq = 1; 960 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 961 } 962 mtx_unlock(&rbreqlock); 963 } 964 965 /* 966 * vfs_buf_test_cache: 967 * 968 * Called when a buffer is extended. This function clears the B_CACHE 969 * bit if the newly extended portion of the buffer does not contain 970 * valid data. 971 */ 972 static __inline void 973 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off, 974 vm_offset_t size, vm_page_t m) 975 { 976 977 /* 978 * This function and its results are protected by higher level 979 * synchronization requiring vnode and buf locks to page in and 980 * validate pages. 981 */ 982 if (bp->b_flags & B_CACHE) { 983 int base = (foff + off) & PAGE_MASK; 984 if (vm_page_is_valid(m, base, size) == 0) 985 bp->b_flags &= ~B_CACHE; 986 } 987 } 988 989 /* Wake up the buffer daemon if necessary */ 990 static void 991 bd_wakeup(void) 992 { 993 994 mtx_lock(&bdlock); 995 if (bd_request == 0) { 996 bd_request = 1; 997 wakeup(&bd_request); 998 } 999 mtx_unlock(&bdlock); 1000 } 1001 1002 /* 1003 * Adjust the maxbcachbuf tunable. 1004 */ 1005 static void 1006 maxbcachebuf_adjust(void) 1007 { 1008 int i; 1009 1010 /* 1011 * maxbcachebuf must be a power of 2 >= MAXBSIZE. 1012 */ 1013 i = 2; 1014 while (i * 2 <= maxbcachebuf) 1015 i *= 2; 1016 maxbcachebuf = i; 1017 if (maxbcachebuf < MAXBSIZE) 1018 maxbcachebuf = MAXBSIZE; 1019 if (maxbcachebuf > maxphys) 1020 maxbcachebuf = maxphys; 1021 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF) 1022 printf("maxbcachebuf=%d\n", maxbcachebuf); 1023 } 1024 1025 /* 1026 * bd_speedup - speedup the buffer cache flushing code 1027 */ 1028 void 1029 bd_speedup(void) 1030 { 1031 int needwake; 1032 1033 mtx_lock(&bdlock); 1034 needwake = 0; 1035 if (bd_speedupreq == 0 || bd_request == 0) 1036 needwake = 1; 1037 bd_speedupreq = 1; 1038 bd_request = 1; 1039 if (needwake) 1040 wakeup(&bd_request); 1041 mtx_unlock(&bdlock); 1042 } 1043 1044 #ifdef __i386__ 1045 #define TRANSIENT_DENOM 5 1046 #else 1047 #define TRANSIENT_DENOM 10 1048 #endif 1049 1050 /* 1051 * Calculating buffer cache scaling values and reserve space for buffer 1052 * headers. This is called during low level kernel initialization and 1053 * may be called more then once. We CANNOT write to the memory area 1054 * being reserved at this time. 1055 */ 1056 caddr_t 1057 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 1058 { 1059 int tuned_nbuf; 1060 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 1061 1062 /* 1063 * With KASAN or KMSAN enabled, the kernel map is shadowed. Account for 1064 * this when sizing maps based on the amount of physical memory 1065 * available. 1066 */ 1067 #if defined(KASAN) 1068 physmem_est = (physmem_est * KASAN_SHADOW_SCALE) / 1069 (KASAN_SHADOW_SCALE + 1); 1070 #elif defined(KMSAN) 1071 physmem_est /= 3; 1072 1073 /* 1074 * KMSAN cannot reliably determine whether buffer data is initialized 1075 * unless it is updated through a KVA mapping. 1076 */ 1077 unmapped_buf_allowed = 0; 1078 #endif 1079 1080 /* 1081 * physmem_est is in pages. Convert it to kilobytes (assumes 1082 * PAGE_SIZE is >= 1K) 1083 */ 1084 physmem_est = physmem_est * (PAGE_SIZE / 1024); 1085 1086 maxbcachebuf_adjust(); 1087 /* 1088 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 1089 * For the first 64MB of ram nominally allocate sufficient buffers to 1090 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 1091 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 1092 * the buffer cache we limit the eventual kva reservation to 1093 * maxbcache bytes. 1094 * 1095 * factor represents the 1/4 x ram conversion. 1096 */ 1097 if (nbuf == 0) { 1098 int factor = 4 * BKVASIZE / 1024; 1099 1100 nbuf = 50; 1101 if (physmem_est > 4096) 1102 nbuf += min((physmem_est - 4096) / factor, 1103 65536 / factor); 1104 if (physmem_est > 65536) 1105 nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 1106 32 * 1024 * 1024 / (factor * 5)); 1107 1108 if (maxbcache && nbuf > maxbcache / BKVASIZE) 1109 nbuf = maxbcache / BKVASIZE; 1110 tuned_nbuf = 1; 1111 } else 1112 tuned_nbuf = 0; 1113 1114 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 1115 maxbuf = (LONG_MAX / 3) / BKVASIZE; 1116 if (nbuf > maxbuf) { 1117 if (!tuned_nbuf) 1118 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 1119 maxbuf); 1120 nbuf = maxbuf; 1121 } 1122 1123 /* 1124 * Ideal allocation size for the transient bio submap is 10% 1125 * of the maximal space buffer map. This roughly corresponds 1126 * to the amount of the buffer mapped for typical UFS load. 1127 * 1128 * Clip the buffer map to reserve space for the transient 1129 * BIOs, if its extent is bigger than 90% (80% on i386) of the 1130 * maximum buffer map extent on the platform. 1131 * 1132 * The fall-back to the maxbuf in case of maxbcache unset, 1133 * allows to not trim the buffer KVA for the architectures 1134 * with ample KVA space. 1135 */ 1136 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 1137 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 1138 buf_sz = (long)nbuf * BKVASIZE; 1139 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 1140 (TRANSIENT_DENOM - 1)) { 1141 /* 1142 * There is more KVA than memory. Do not 1143 * adjust buffer map size, and assign the rest 1144 * of maxbuf to transient map. 1145 */ 1146 biotmap_sz = maxbuf_sz - buf_sz; 1147 } else { 1148 /* 1149 * Buffer map spans all KVA we could afford on 1150 * this platform. Give 10% (20% on i386) of 1151 * the buffer map to the transient bio map. 1152 */ 1153 biotmap_sz = buf_sz / TRANSIENT_DENOM; 1154 buf_sz -= biotmap_sz; 1155 } 1156 if (biotmap_sz / INT_MAX > maxphys) 1157 bio_transient_maxcnt = INT_MAX; 1158 else 1159 bio_transient_maxcnt = biotmap_sz / maxphys; 1160 /* 1161 * Artificially limit to 1024 simultaneous in-flight I/Os 1162 * using the transient mapping. 1163 */ 1164 if (bio_transient_maxcnt > 1024) 1165 bio_transient_maxcnt = 1024; 1166 if (tuned_nbuf) 1167 nbuf = buf_sz / BKVASIZE; 1168 } 1169 1170 if (nswbuf == 0) { 1171 nswbuf = min(nbuf / 4, 256); 1172 if (nswbuf < NSWBUF_MIN) 1173 nswbuf = NSWBUF_MIN; 1174 } 1175 1176 /* 1177 * Reserve space for the buffer cache buffers 1178 */ 1179 buf = (char *)v; 1180 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) * 1181 atop(maxbcachebuf)) * nbuf; 1182 1183 return (v); 1184 } 1185 1186 /* 1187 * Single global constant for BUF_WMESG, to avoid getting multiple 1188 * references. 1189 */ 1190 static const char buf_wmesg[] = "bufwait"; 1191 1192 /* Initialize the buffer subsystem. Called before use of any buffers. */ 1193 void 1194 bufinit(void) 1195 { 1196 struct buf *bp; 1197 int i; 1198 1199 KASSERT(maxbcachebuf >= MAXBSIZE, 1200 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf, 1201 MAXBSIZE)); 1202 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock"); 1203 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 1204 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 1205 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); 1206 1207 unmapped_buf = (caddr_t)kva_alloc(maxphys); 1208 1209 /* finally, initialize each buffer header and stick on empty q */ 1210 for (i = 0; i < nbuf; i++) { 1211 bp = nbufp(i); 1212 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf)); 1213 bp->b_flags = B_INVAL; 1214 bp->b_rcred = NOCRED; 1215 bp->b_wcred = NOCRED; 1216 bp->b_qindex = QUEUE_NONE; 1217 bp->b_domain = -1; 1218 bp->b_subqueue = mp_maxid + 1; 1219 bp->b_xflags = 0; 1220 bp->b_data = bp->b_kvabase = unmapped_buf; 1221 LIST_INIT(&bp->b_dep); 1222 BUF_LOCKINIT(bp, buf_wmesg); 1223 bq_insert(&bqempty, bp, false); 1224 } 1225 1226 /* 1227 * maxbufspace is the absolute maximum amount of buffer space we are 1228 * allowed to reserve in KVM and in real terms. The absolute maximum 1229 * is nominally used by metadata. hibufspace is the nominal maximum 1230 * used by most other requests. The differential is required to 1231 * ensure that metadata deadlocks don't occur. 1232 * 1233 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 1234 * this may result in KVM fragmentation which is not handled optimally 1235 * by the system. XXX This is less true with vmem. We could use 1236 * PAGE_SIZE. 1237 */ 1238 maxbufspace = (long)nbuf * BKVASIZE; 1239 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10); 1240 lobufspace = (hibufspace / 20) * 19; /* 95% */ 1241 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2; 1242 1243 /* 1244 * Note: The 16 MiB upper limit for hirunningspace was chosen 1245 * arbitrarily and may need further tuning. It corresponds to 1246 * 128 outstanding write IO requests (if IO size is 128 KiB), 1247 * which fits with many RAID controllers' tagged queuing limits. 1248 * The lower 1 MiB limit is the historical upper limit for 1249 * hirunningspace. 1250 */ 1251 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf), 1252 16 * 1024 * 1024), 1024 * 1024); 1253 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf); 1254 1255 /* 1256 * Limit the amount of malloc memory since it is wired permanently into 1257 * the kernel space. Even though this is accounted for in the buffer 1258 * allocation, we don't want the malloced region to grow uncontrolled. 1259 * The malloc scheme improves memory utilization significantly on 1260 * average (small) directories. 1261 */ 1262 maxbufmallocspace = hibufspace / 20; 1263 1264 /* 1265 * Reduce the chance of a deadlock occurring by limiting the number 1266 * of delayed-write dirty buffers we allow to stack up. 1267 */ 1268 hidirtybuffers = nbuf / 4 + 20; 1269 dirtybufthresh = hidirtybuffers * 9 / 10; 1270 /* 1271 * To support extreme low-memory systems, make sure hidirtybuffers 1272 * cannot eat up all available buffer space. This occurs when our 1273 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our 1274 * buffer space assuming BKVASIZE'd buffers. 1275 */ 1276 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 1277 hidirtybuffers >>= 1; 1278 } 1279 lodirtybuffers = hidirtybuffers / 2; 1280 1281 /* 1282 * lofreebuffers should be sufficient to avoid stalling waiting on 1283 * buf headers under heavy utilization. The bufs in per-cpu caches 1284 * are counted as free but will be unavailable to threads executing 1285 * on other cpus. 1286 * 1287 * hifreebuffers is the free target for the bufspace daemon. This 1288 * should be set appropriately to limit work per-iteration. 1289 */ 1290 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus); 1291 hifreebuffers = (3 * lofreebuffers) / 2; 1292 numfreebuffers = nbuf; 1293 1294 /* Setup the kva and free list allocators. */ 1295 vmem_set_reclaim(buffer_arena, bufkva_reclaim); 1296 buf_zone = uma_zcache_create("buf free cache", 1297 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf), 1298 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0); 1299 1300 /* 1301 * Size the clean queue according to the amount of buffer space. 1302 * One queue per-256mb up to the max. More queues gives better 1303 * concurrency but less accurate LRU. 1304 */ 1305 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS); 1306 for (i = 0 ; i < buf_domains; i++) { 1307 struct bufdomain *bd; 1308 1309 bd = &bdomain[i]; 1310 bd_init(bd); 1311 bd->bd_freebuffers = nbuf / buf_domains; 1312 bd->bd_hifreebuffers = hifreebuffers / buf_domains; 1313 bd->bd_lofreebuffers = lofreebuffers / buf_domains; 1314 bd->bd_bufspace = 0; 1315 bd->bd_maxbufspace = maxbufspace / buf_domains; 1316 bd->bd_hibufspace = hibufspace / buf_domains; 1317 bd->bd_lobufspace = lobufspace / buf_domains; 1318 bd->bd_bufspacethresh = bufspacethresh / buf_domains; 1319 bd->bd_numdirtybuffers = 0; 1320 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains; 1321 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains; 1322 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains; 1323 /* Don't allow more than 2% of bufs in the per-cpu caches. */ 1324 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus; 1325 } 1326 getnewbufcalls = counter_u64_alloc(M_WAITOK); 1327 getnewbufrestarts = counter_u64_alloc(M_WAITOK); 1328 mappingrestarts = counter_u64_alloc(M_WAITOK); 1329 numbufallocfails = counter_u64_alloc(M_WAITOK); 1330 notbufdflushes = counter_u64_alloc(M_WAITOK); 1331 buffreekvacnt = counter_u64_alloc(M_WAITOK); 1332 bufdefragcnt = counter_u64_alloc(M_WAITOK); 1333 bufkvaspace = counter_u64_alloc(M_WAITOK); 1334 } 1335 1336 #ifdef INVARIANTS 1337 static inline void 1338 vfs_buf_check_mapped(struct buf *bp) 1339 { 1340 1341 KASSERT(bp->b_kvabase != unmapped_buf, 1342 ("mapped buf: b_kvabase was not updated %p", bp)); 1343 KASSERT(bp->b_data != unmapped_buf, 1344 ("mapped buf: b_data was not updated %p", bp)); 1345 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf + 1346 maxphys, ("b_data + b_offset unmapped %p", bp)); 1347 } 1348 1349 static inline void 1350 vfs_buf_check_unmapped(struct buf *bp) 1351 { 1352 1353 KASSERT(bp->b_data == unmapped_buf, 1354 ("unmapped buf: corrupted b_data %p", bp)); 1355 } 1356 1357 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 1358 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 1359 #else 1360 #define BUF_CHECK_MAPPED(bp) do {} while (0) 1361 #define BUF_CHECK_UNMAPPED(bp) do {} while (0) 1362 #endif 1363 1364 static int 1365 isbufbusy(struct buf *bp) 1366 { 1367 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) || 1368 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI)) 1369 return (1); 1370 return (0); 1371 } 1372 1373 /* 1374 * Shutdown the system cleanly to prepare for reboot, halt, or power off. 1375 */ 1376 void 1377 bufshutdown(int show_busybufs) 1378 { 1379 static int first_buf_printf = 1; 1380 struct buf *bp; 1381 int i, iter, nbusy, pbusy; 1382 #ifndef PREEMPTION 1383 int subiter; 1384 #endif 1385 1386 /* 1387 * Sync filesystems for shutdown 1388 */ 1389 wdog_kern_pat(WD_LASTVAL); 1390 kern_sync(curthread); 1391 1392 /* 1393 * With soft updates, some buffers that are 1394 * written will be remarked as dirty until other 1395 * buffers are written. 1396 */ 1397 for (iter = pbusy = 0; iter < 20; iter++) { 1398 nbusy = 0; 1399 for (i = nbuf - 1; i >= 0; i--) { 1400 bp = nbufp(i); 1401 if (isbufbusy(bp)) 1402 nbusy++; 1403 } 1404 if (nbusy == 0) { 1405 if (first_buf_printf) 1406 printf("All buffers synced."); 1407 break; 1408 } 1409 if (first_buf_printf) { 1410 printf("Syncing disks, buffers remaining... "); 1411 first_buf_printf = 0; 1412 } 1413 printf("%d ", nbusy); 1414 if (nbusy < pbusy) 1415 iter = 0; 1416 pbusy = nbusy; 1417 1418 wdog_kern_pat(WD_LASTVAL); 1419 kern_sync(curthread); 1420 1421 #ifdef PREEMPTION 1422 /* 1423 * Spin for a while to allow interrupt threads to run. 1424 */ 1425 DELAY(50000 * iter); 1426 #else 1427 /* 1428 * Context switch several times to allow interrupt 1429 * threads to run. 1430 */ 1431 for (subiter = 0; subiter < 50 * iter; subiter++) { 1432 sched_relinquish(curthread); 1433 DELAY(1000); 1434 } 1435 #endif 1436 } 1437 printf("\n"); 1438 /* 1439 * Count only busy local buffers to prevent forcing 1440 * a fsck if we're just a client of a wedged NFS server 1441 */ 1442 nbusy = 0; 1443 for (i = nbuf - 1; i >= 0; i--) { 1444 bp = nbufp(i); 1445 if (isbufbusy(bp)) { 1446 #if 0 1447 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */ 1448 if (bp->b_dev == NULL) { 1449 TAILQ_REMOVE(&mountlist, 1450 bp->b_vp->v_mount, mnt_list); 1451 continue; 1452 } 1453 #endif 1454 nbusy++; 1455 if (show_busybufs > 0) { 1456 printf( 1457 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:", 1458 nbusy, bp, bp->b_vp, bp->b_flags, 1459 (intmax_t)bp->b_blkno, 1460 (intmax_t)bp->b_lblkno); 1461 BUF_LOCKPRINTINFO(bp); 1462 if (show_busybufs > 1) 1463 vn_printf(bp->b_vp, 1464 "vnode content: "); 1465 } 1466 } 1467 } 1468 if (nbusy) { 1469 /* 1470 * Failed to sync all blocks. Indicate this and don't 1471 * unmount filesystems (thus forcing an fsck on reboot). 1472 */ 1473 BOOTTRACE("shutdown failed to sync buffers"); 1474 printf("Giving up on %d buffers\n", nbusy); 1475 DELAY(5000000); /* 5 seconds */ 1476 swapoff_all(); 1477 } else { 1478 BOOTTRACE("shutdown sync complete"); 1479 if (!first_buf_printf) 1480 printf("Final sync complete\n"); 1481 1482 /* 1483 * Unmount filesystems and perform swapoff, to quiesce 1484 * the system as much as possible. In particular, no 1485 * I/O should be initiated from top levels since it 1486 * might be abruptly terminated by reset, or otherwise 1487 * erronously handled because other parts of the 1488 * system are disabled. 1489 * 1490 * Swapoff before unmount, because file-backed swap is 1491 * non-operational after unmount of the underlying 1492 * filesystem. 1493 */ 1494 if (!KERNEL_PANICKED()) { 1495 swapoff_all(); 1496 vfs_unmountall(); 1497 } 1498 BOOTTRACE("shutdown unmounted all filesystems"); 1499 } 1500 DELAY(100000); /* wait for console output to finish */ 1501 } 1502 1503 static void 1504 bpmap_qenter(struct buf *bp) 1505 { 1506 1507 BUF_CHECK_MAPPED(bp); 1508 1509 /* 1510 * bp->b_data is relative to bp->b_offset, but 1511 * bp->b_offset may be offset into the first page. 1512 */ 1513 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 1514 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 1515 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 1516 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 1517 } 1518 1519 static inline struct bufdomain * 1520 bufdomain(struct buf *bp) 1521 { 1522 1523 return (&bdomain[bp->b_domain]); 1524 } 1525 1526 static struct bufqueue * 1527 bufqueue(struct buf *bp) 1528 { 1529 1530 switch (bp->b_qindex) { 1531 case QUEUE_NONE: 1532 /* FALLTHROUGH */ 1533 case QUEUE_SENTINEL: 1534 return (NULL); 1535 case QUEUE_EMPTY: 1536 return (&bqempty); 1537 case QUEUE_DIRTY: 1538 return (&bufdomain(bp)->bd_dirtyq); 1539 case QUEUE_CLEAN: 1540 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]); 1541 default: 1542 break; 1543 } 1544 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex); 1545 } 1546 1547 /* 1548 * Return the locked bufqueue that bp is a member of. 1549 */ 1550 static struct bufqueue * 1551 bufqueue_acquire(struct buf *bp) 1552 { 1553 struct bufqueue *bq, *nbq; 1554 1555 /* 1556 * bp can be pushed from a per-cpu queue to the 1557 * cleanq while we're waiting on the lock. Retry 1558 * if the queues don't match. 1559 */ 1560 bq = bufqueue(bp); 1561 BQ_LOCK(bq); 1562 for (;;) { 1563 nbq = bufqueue(bp); 1564 if (bq == nbq) 1565 break; 1566 BQ_UNLOCK(bq); 1567 BQ_LOCK(nbq); 1568 bq = nbq; 1569 } 1570 return (bq); 1571 } 1572 1573 /* 1574 * binsfree: 1575 * 1576 * Insert the buffer into the appropriate free list. Requires a 1577 * locked buffer on entry and buffer is unlocked before return. 1578 */ 1579 static void 1580 binsfree(struct buf *bp, int qindex) 1581 { 1582 struct bufdomain *bd; 1583 struct bufqueue *bq; 1584 1585 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY, 1586 ("binsfree: Invalid qindex %d", qindex)); 1587 BUF_ASSERT_XLOCKED(bp); 1588 1589 /* 1590 * Handle delayed bremfree() processing. 1591 */ 1592 if (bp->b_flags & B_REMFREE) { 1593 if (bp->b_qindex == qindex) { 1594 bp->b_flags |= B_REUSE; 1595 bp->b_flags &= ~B_REMFREE; 1596 BUF_UNLOCK(bp); 1597 return; 1598 } 1599 bq = bufqueue_acquire(bp); 1600 bq_remove(bq, bp); 1601 BQ_UNLOCK(bq); 1602 } 1603 bd = bufdomain(bp); 1604 if (qindex == QUEUE_CLEAN) { 1605 if (bd->bd_lim != 0) 1606 bq = &bd->bd_subq[PCPU_GET(cpuid)]; 1607 else 1608 bq = bd->bd_cleanq; 1609 } else 1610 bq = &bd->bd_dirtyq; 1611 bq_insert(bq, bp, true); 1612 } 1613 1614 /* 1615 * buf_free: 1616 * 1617 * Free a buffer to the buf zone once it no longer has valid contents. 1618 */ 1619 static void 1620 buf_free(struct buf *bp) 1621 { 1622 1623 if (bp->b_flags & B_REMFREE) 1624 bremfreef(bp); 1625 if (bp->b_vflags & BV_BKGRDINPROG) 1626 panic("losing buffer 1"); 1627 if (bp->b_rcred != NOCRED) { 1628 crfree(bp->b_rcred); 1629 bp->b_rcred = NOCRED; 1630 } 1631 if (bp->b_wcred != NOCRED) { 1632 crfree(bp->b_wcred); 1633 bp->b_wcred = NOCRED; 1634 } 1635 if (!LIST_EMPTY(&bp->b_dep)) 1636 buf_deallocate(bp); 1637 bufkva_free(bp); 1638 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1); 1639 MPASS((bp->b_flags & B_MAXPHYS) == 0); 1640 BUF_UNLOCK(bp); 1641 uma_zfree(buf_zone, bp); 1642 } 1643 1644 /* 1645 * buf_import: 1646 * 1647 * Import bufs into the uma cache from the buf list. The system still 1648 * expects a static array of bufs and much of the synchronization 1649 * around bufs assumes type stable storage. As a result, UMA is used 1650 * only as a per-cpu cache of bufs still maintained on a global list. 1651 */ 1652 static int 1653 buf_import(void *arg, void **store, int cnt, int domain, int flags) 1654 { 1655 struct buf *bp; 1656 int i; 1657 1658 BQ_LOCK(&bqempty); 1659 for (i = 0; i < cnt; i++) { 1660 bp = TAILQ_FIRST(&bqempty.bq_queue); 1661 if (bp == NULL) 1662 break; 1663 bq_remove(&bqempty, bp); 1664 store[i] = bp; 1665 } 1666 BQ_UNLOCK(&bqempty); 1667 1668 return (i); 1669 } 1670 1671 /* 1672 * buf_release: 1673 * 1674 * Release bufs from the uma cache back to the buffer queues. 1675 */ 1676 static void 1677 buf_release(void *arg, void **store, int cnt) 1678 { 1679 struct bufqueue *bq; 1680 struct buf *bp; 1681 int i; 1682 1683 bq = &bqempty; 1684 BQ_LOCK(bq); 1685 for (i = 0; i < cnt; i++) { 1686 bp = store[i]; 1687 /* Inline bq_insert() to batch locking. */ 1688 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1689 bp->b_flags &= ~(B_AGE | B_REUSE); 1690 bq->bq_len++; 1691 bp->b_qindex = bq->bq_index; 1692 } 1693 BQ_UNLOCK(bq); 1694 } 1695 1696 /* 1697 * buf_alloc: 1698 * 1699 * Allocate an empty buffer header. 1700 */ 1701 static struct buf * 1702 buf_alloc(struct bufdomain *bd) 1703 { 1704 struct buf *bp; 1705 int freebufs, error; 1706 1707 /* 1708 * We can only run out of bufs in the buf zone if the average buf 1709 * is less than BKVASIZE. In this case the actual wait/block will 1710 * come from buf_reycle() failing to flush one of these small bufs. 1711 */ 1712 bp = NULL; 1713 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1); 1714 if (freebufs > 0) 1715 bp = uma_zalloc(buf_zone, M_NOWAIT); 1716 if (bp == NULL) { 1717 atomic_add_int(&bd->bd_freebuffers, 1); 1718 bufspace_daemon_wakeup(bd); 1719 counter_u64_add(numbufallocfails, 1); 1720 return (NULL); 1721 } 1722 /* 1723 * Wake-up the bufspace daemon on transition below threshold. 1724 */ 1725 if (freebufs == bd->bd_lofreebuffers) 1726 bufspace_daemon_wakeup(bd); 1727 1728 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWITNESS, NULL); 1729 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp, 1730 error)); 1731 (void)error; 1732 1733 KASSERT(bp->b_vp == NULL, 1734 ("bp: %p still has vnode %p.", bp, bp->b_vp)); 1735 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0, 1736 ("invalid buffer %p flags %#x", bp, bp->b_flags)); 1737 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1738 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 1739 KASSERT(bp->b_npages == 0, 1740 ("bp: %p still has %d vm pages\n", bp, bp->b_npages)); 1741 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp)); 1742 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp)); 1743 MPASS((bp->b_flags & B_MAXPHYS) == 0); 1744 1745 bp->b_domain = BD_DOMAIN(bd); 1746 bp->b_flags = 0; 1747 bp->b_ioflags = 0; 1748 bp->b_xflags = 0; 1749 bp->b_vflags = 0; 1750 bp->b_vp = NULL; 1751 bp->b_blkno = bp->b_lblkno = 0; 1752 bp->b_offset = NOOFFSET; 1753 bp->b_iodone = 0; 1754 bp->b_error = 0; 1755 bp->b_resid = 0; 1756 bp->b_bcount = 0; 1757 bp->b_npages = 0; 1758 bp->b_dirtyoff = bp->b_dirtyend = 0; 1759 bp->b_bufobj = NULL; 1760 bp->b_data = bp->b_kvabase = unmapped_buf; 1761 bp->b_fsprivate1 = NULL; 1762 bp->b_fsprivate2 = NULL; 1763 bp->b_fsprivate3 = NULL; 1764 LIST_INIT(&bp->b_dep); 1765 1766 return (bp); 1767 } 1768 1769 /* 1770 * buf_recycle: 1771 * 1772 * Free a buffer from the given bufqueue. kva controls whether the 1773 * freed buf must own some kva resources. This is used for 1774 * defragmenting. 1775 */ 1776 static int 1777 buf_recycle(struct bufdomain *bd, bool kva) 1778 { 1779 struct bufqueue *bq; 1780 struct buf *bp, *nbp; 1781 1782 if (kva) 1783 counter_u64_add(bufdefragcnt, 1); 1784 nbp = NULL; 1785 bq = bd->bd_cleanq; 1786 BQ_LOCK(bq); 1787 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd), 1788 ("buf_recycle: Locks don't match")); 1789 nbp = TAILQ_FIRST(&bq->bq_queue); 1790 1791 /* 1792 * Run scan, possibly freeing data and/or kva mappings on the fly 1793 * depending. 1794 */ 1795 while ((bp = nbp) != NULL) { 1796 /* 1797 * Calculate next bp (we can only use it if we do not 1798 * release the bqlock). 1799 */ 1800 nbp = TAILQ_NEXT(bp, b_freelist); 1801 1802 /* 1803 * If we are defragging then we need a buffer with 1804 * some kva to reclaim. 1805 */ 1806 if (kva && bp->b_kvasize == 0) 1807 continue; 1808 1809 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1810 continue; 1811 1812 /* 1813 * Implement a second chance algorithm for frequently 1814 * accessed buffers. 1815 */ 1816 if ((bp->b_flags & B_REUSE) != 0) { 1817 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1818 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1819 bp->b_flags &= ~B_REUSE; 1820 BUF_UNLOCK(bp); 1821 continue; 1822 } 1823 1824 /* 1825 * Skip buffers with background writes in progress. 1826 */ 1827 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 1828 BUF_UNLOCK(bp); 1829 continue; 1830 } 1831 1832 KASSERT(bp->b_qindex == QUEUE_CLEAN, 1833 ("buf_recycle: inconsistent queue %d bp %p", 1834 bp->b_qindex, bp)); 1835 KASSERT(bp->b_domain == BD_DOMAIN(bd), 1836 ("getnewbuf: queue domain %d doesn't match request %d", 1837 bp->b_domain, (int)BD_DOMAIN(bd))); 1838 /* 1839 * NOTE: nbp is now entirely invalid. We can only restart 1840 * the scan from this point on. 1841 */ 1842 bq_remove(bq, bp); 1843 BQ_UNLOCK(bq); 1844 1845 /* 1846 * Requeue the background write buffer with error and 1847 * restart the scan. 1848 */ 1849 if ((bp->b_vflags & BV_BKGRDERR) != 0) { 1850 bqrelse(bp); 1851 BQ_LOCK(bq); 1852 nbp = TAILQ_FIRST(&bq->bq_queue); 1853 continue; 1854 } 1855 bp->b_flags |= B_INVAL; 1856 brelse(bp); 1857 return (0); 1858 } 1859 bd->bd_wanted = 1; 1860 BQ_UNLOCK(bq); 1861 1862 return (ENOBUFS); 1863 } 1864 1865 /* 1866 * bremfree: 1867 * 1868 * Mark the buffer for removal from the appropriate free list. 1869 * 1870 */ 1871 void 1872 bremfree(struct buf *bp) 1873 { 1874 1875 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1876 KASSERT((bp->b_flags & B_REMFREE) == 0, 1877 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1878 KASSERT(bp->b_qindex != QUEUE_NONE, 1879 ("bremfree: buffer %p not on a queue.", bp)); 1880 BUF_ASSERT_XLOCKED(bp); 1881 1882 bp->b_flags |= B_REMFREE; 1883 } 1884 1885 /* 1886 * bremfreef: 1887 * 1888 * Force an immediate removal from a free list. Used only in nfs when 1889 * it abuses the b_freelist pointer. 1890 */ 1891 void 1892 bremfreef(struct buf *bp) 1893 { 1894 struct bufqueue *bq; 1895 1896 bq = bufqueue_acquire(bp); 1897 bq_remove(bq, bp); 1898 BQ_UNLOCK(bq); 1899 } 1900 1901 static void 1902 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname) 1903 { 1904 1905 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF); 1906 TAILQ_INIT(&bq->bq_queue); 1907 bq->bq_len = 0; 1908 bq->bq_index = qindex; 1909 bq->bq_subqueue = subqueue; 1910 } 1911 1912 static void 1913 bd_init(struct bufdomain *bd) 1914 { 1915 int i; 1916 1917 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1]; 1918 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock"); 1919 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock"); 1920 for (i = 0; i <= mp_maxid; i++) 1921 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i, 1922 "bufq clean subqueue lock"); 1923 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF); 1924 } 1925 1926 /* 1927 * bq_remove: 1928 * 1929 * Removes a buffer from the free list, must be called with the 1930 * correct qlock held. 1931 */ 1932 static void 1933 bq_remove(struct bufqueue *bq, struct buf *bp) 1934 { 1935 1936 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X", 1937 bp, bp->b_vp, bp->b_flags); 1938 KASSERT(bp->b_qindex != QUEUE_NONE, 1939 ("bq_remove: buffer %p not on a queue.", bp)); 1940 KASSERT(bufqueue(bp) == bq, 1941 ("bq_remove: Remove buffer %p from wrong queue.", bp)); 1942 1943 BQ_ASSERT_LOCKED(bq); 1944 if (bp->b_qindex != QUEUE_EMPTY) { 1945 BUF_ASSERT_XLOCKED(bp); 1946 } 1947 KASSERT(bq->bq_len >= 1, 1948 ("queue %d underflow", bp->b_qindex)); 1949 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1950 bq->bq_len--; 1951 bp->b_qindex = QUEUE_NONE; 1952 bp->b_flags &= ~(B_REMFREE | B_REUSE); 1953 } 1954 1955 static void 1956 bd_flush(struct bufdomain *bd, struct bufqueue *bq) 1957 { 1958 struct buf *bp; 1959 1960 BQ_ASSERT_LOCKED(bq); 1961 if (bq != bd->bd_cleanq) { 1962 BD_LOCK(bd); 1963 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) { 1964 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1965 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp, 1966 b_freelist); 1967 bp->b_subqueue = bd->bd_cleanq->bq_subqueue; 1968 } 1969 bd->bd_cleanq->bq_len += bq->bq_len; 1970 bq->bq_len = 0; 1971 } 1972 if (bd->bd_wanted) { 1973 bd->bd_wanted = 0; 1974 wakeup(&bd->bd_wanted); 1975 } 1976 if (bq != bd->bd_cleanq) 1977 BD_UNLOCK(bd); 1978 } 1979 1980 static int 1981 bd_flushall(struct bufdomain *bd) 1982 { 1983 struct bufqueue *bq; 1984 int flushed; 1985 int i; 1986 1987 if (bd->bd_lim == 0) 1988 return (0); 1989 flushed = 0; 1990 for (i = 0; i <= mp_maxid; i++) { 1991 bq = &bd->bd_subq[i]; 1992 if (bq->bq_len == 0) 1993 continue; 1994 BQ_LOCK(bq); 1995 bd_flush(bd, bq); 1996 BQ_UNLOCK(bq); 1997 flushed++; 1998 } 1999 2000 return (flushed); 2001 } 2002 2003 static void 2004 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock) 2005 { 2006 struct bufdomain *bd; 2007 2008 if (bp->b_qindex != QUEUE_NONE) 2009 panic("bq_insert: free buffer %p onto another queue?", bp); 2010 2011 bd = bufdomain(bp); 2012 if (bp->b_flags & B_AGE) { 2013 /* Place this buf directly on the real queue. */ 2014 if (bq->bq_index == QUEUE_CLEAN) 2015 bq = bd->bd_cleanq; 2016 BQ_LOCK(bq); 2017 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist); 2018 } else { 2019 BQ_LOCK(bq); 2020 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 2021 } 2022 bp->b_flags &= ~(B_AGE | B_REUSE); 2023 bq->bq_len++; 2024 bp->b_qindex = bq->bq_index; 2025 bp->b_subqueue = bq->bq_subqueue; 2026 2027 /* 2028 * Unlock before we notify so that we don't wakeup a waiter that 2029 * fails a trylock on the buf and sleeps again. 2030 */ 2031 if (unlock) 2032 BUF_UNLOCK(bp); 2033 2034 if (bp->b_qindex == QUEUE_CLEAN) { 2035 /* 2036 * Flush the per-cpu queue and notify any waiters. 2037 */ 2038 if (bd->bd_wanted || (bq != bd->bd_cleanq && 2039 bq->bq_len >= bd->bd_lim)) 2040 bd_flush(bd, bq); 2041 } 2042 BQ_UNLOCK(bq); 2043 } 2044 2045 /* 2046 * bufkva_free: 2047 * 2048 * Free the kva allocation for a buffer. 2049 * 2050 */ 2051 static void 2052 bufkva_free(struct buf *bp) 2053 { 2054 2055 #ifdef INVARIANTS 2056 if (bp->b_kvasize == 0) { 2057 KASSERT(bp->b_kvabase == unmapped_buf && 2058 bp->b_data == unmapped_buf, 2059 ("Leaked KVA space on %p", bp)); 2060 } else if (buf_mapped(bp)) 2061 BUF_CHECK_MAPPED(bp); 2062 else 2063 BUF_CHECK_UNMAPPED(bp); 2064 #endif 2065 if (bp->b_kvasize == 0) 2066 return; 2067 2068 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize); 2069 counter_u64_add(bufkvaspace, -bp->b_kvasize); 2070 counter_u64_add(buffreekvacnt, 1); 2071 bp->b_data = bp->b_kvabase = unmapped_buf; 2072 bp->b_kvasize = 0; 2073 } 2074 2075 /* 2076 * bufkva_alloc: 2077 * 2078 * Allocate the buffer KVA and set b_kvasize and b_kvabase. 2079 */ 2080 static int 2081 bufkva_alloc(struct buf *bp, int maxsize, int gbflags) 2082 { 2083 vm_offset_t addr; 2084 int error; 2085 2086 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0, 2087 ("Invalid gbflags 0x%x in %s", gbflags, __func__)); 2088 MPASS((bp->b_flags & B_MAXPHYS) == 0); 2089 KASSERT(maxsize <= maxbcachebuf, 2090 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf)); 2091 2092 bufkva_free(bp); 2093 2094 addr = 0; 2095 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr); 2096 if (error != 0) { 2097 /* 2098 * Buffer map is too fragmented. Request the caller 2099 * to defragment the map. 2100 */ 2101 return (error); 2102 } 2103 bp->b_kvabase = (caddr_t)addr; 2104 bp->b_kvasize = maxsize; 2105 counter_u64_add(bufkvaspace, bp->b_kvasize); 2106 if ((gbflags & GB_UNMAPPED) != 0) { 2107 bp->b_data = unmapped_buf; 2108 BUF_CHECK_UNMAPPED(bp); 2109 } else { 2110 bp->b_data = bp->b_kvabase; 2111 BUF_CHECK_MAPPED(bp); 2112 } 2113 return (0); 2114 } 2115 2116 /* 2117 * bufkva_reclaim: 2118 * 2119 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem 2120 * callback that fires to avoid returning failure. 2121 */ 2122 static void 2123 bufkva_reclaim(vmem_t *vmem, int flags) 2124 { 2125 bool done; 2126 int q; 2127 int i; 2128 2129 done = false; 2130 for (i = 0; i < 5; i++) { 2131 for (q = 0; q < buf_domains; q++) 2132 if (buf_recycle(&bdomain[q], true) != 0) 2133 done = true; 2134 if (done) 2135 break; 2136 } 2137 return; 2138 } 2139 2140 /* 2141 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 2142 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 2143 * the buffer is valid and we do not have to do anything. 2144 */ 2145 static void 2146 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt, 2147 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *)) 2148 { 2149 struct buf *rabp; 2150 struct thread *td; 2151 int i; 2152 2153 td = curthread; 2154 2155 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 2156 if (inmem(vp, *rablkno)) 2157 continue; 2158 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 2159 if ((rabp->b_flags & B_CACHE) != 0) { 2160 brelse(rabp); 2161 continue; 2162 } 2163 #ifdef RACCT 2164 if (racct_enable) { 2165 PROC_LOCK(curproc); 2166 racct_add_buf(curproc, rabp, 0); 2167 PROC_UNLOCK(curproc); 2168 } 2169 #endif /* RACCT */ 2170 td->td_ru.ru_inblock++; 2171 rabp->b_flags |= B_ASYNC; 2172 rabp->b_flags &= ~B_INVAL; 2173 if ((flags & GB_CKHASH) != 0) { 2174 rabp->b_flags |= B_CKHASH; 2175 rabp->b_ckhashcalc = ckhashfunc; 2176 } 2177 rabp->b_ioflags &= ~BIO_ERROR; 2178 rabp->b_iocmd = BIO_READ; 2179 if (rabp->b_rcred == NOCRED && cred != NOCRED) 2180 rabp->b_rcred = crhold(cred); 2181 vfs_busy_pages(rabp, 0); 2182 BUF_KERNPROC(rabp); 2183 rabp->b_iooffset = dbtob(rabp->b_blkno); 2184 bstrategy(rabp); 2185 } 2186 } 2187 2188 /* 2189 * Entry point for bread() and breadn() via #defines in sys/buf.h. 2190 * 2191 * Get a buffer with the specified data. Look in the cache first. We 2192 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 2193 * is set, the buffer is valid and we do not have to do anything, see 2194 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 2195 * 2196 * Always return a NULL buffer pointer (in bpp) when returning an error. 2197 * 2198 * The blkno parameter is the logical block being requested. Normally 2199 * the mapping of logical block number to disk block address is done 2200 * by calling VOP_BMAP(). However, if the mapping is already known, the 2201 * disk block address can be passed using the dblkno parameter. If the 2202 * disk block address is not known, then the same value should be passed 2203 * for blkno and dblkno. 2204 */ 2205 int 2206 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, 2207 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags, 2208 void (*ckhashfunc)(struct buf *), struct buf **bpp) 2209 { 2210 struct buf *bp; 2211 struct thread *td; 2212 int error, readwait, rv; 2213 2214 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 2215 td = curthread; 2216 /* 2217 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags 2218 * are specified. 2219 */ 2220 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp); 2221 if (error != 0) { 2222 *bpp = NULL; 2223 return (error); 2224 } 2225 KASSERT(blkno == bp->b_lblkno, 2226 ("getblkx returned buffer for blkno %jd instead of blkno %jd", 2227 (intmax_t)bp->b_lblkno, (intmax_t)blkno)); 2228 flags &= ~GB_NOSPARSE; 2229 *bpp = bp; 2230 2231 /* 2232 * If not found in cache, do some I/O 2233 */ 2234 readwait = 0; 2235 if ((bp->b_flags & B_CACHE) == 0) { 2236 #ifdef RACCT 2237 if (racct_enable) { 2238 PROC_LOCK(td->td_proc); 2239 racct_add_buf(td->td_proc, bp, 0); 2240 PROC_UNLOCK(td->td_proc); 2241 } 2242 #endif /* RACCT */ 2243 td->td_ru.ru_inblock++; 2244 bp->b_iocmd = BIO_READ; 2245 bp->b_flags &= ~B_INVAL; 2246 if ((flags & GB_CKHASH) != 0) { 2247 bp->b_flags |= B_CKHASH; 2248 bp->b_ckhashcalc = ckhashfunc; 2249 } 2250 if ((flags & GB_CVTENXIO) != 0) 2251 bp->b_xflags |= BX_CVTENXIO; 2252 bp->b_ioflags &= ~BIO_ERROR; 2253 if (bp->b_rcred == NOCRED && cred != NOCRED) 2254 bp->b_rcred = crhold(cred); 2255 vfs_busy_pages(bp, 0); 2256 bp->b_iooffset = dbtob(bp->b_blkno); 2257 bstrategy(bp); 2258 ++readwait; 2259 } 2260 2261 /* 2262 * Attempt to initiate asynchronous I/O on read-ahead blocks. 2263 */ 2264 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc); 2265 2266 rv = 0; 2267 if (readwait) { 2268 rv = bufwait(bp); 2269 if (rv != 0) { 2270 brelse(bp); 2271 *bpp = NULL; 2272 } 2273 } 2274 return (rv); 2275 } 2276 2277 /* 2278 * Write, release buffer on completion. (Done by iodone 2279 * if async). Do not bother writing anything if the buffer 2280 * is invalid. 2281 * 2282 * Note that we set B_CACHE here, indicating that buffer is 2283 * fully valid and thus cacheable. This is true even of NFS 2284 * now so we set it generally. This could be set either here 2285 * or in biodone() since the I/O is synchronous. We put it 2286 * here. 2287 */ 2288 int 2289 bufwrite(struct buf *bp) 2290 { 2291 int oldflags; 2292 struct vnode *vp; 2293 long space; 2294 int vp_md; 2295 2296 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2297 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) { 2298 bp->b_flags |= B_INVAL | B_RELBUF; 2299 bp->b_flags &= ~B_CACHE; 2300 brelse(bp); 2301 return (ENXIO); 2302 } 2303 if (bp->b_flags & B_INVAL) { 2304 brelse(bp); 2305 return (0); 2306 } 2307 2308 if (bp->b_flags & B_BARRIER) 2309 atomic_add_long(&barrierwrites, 1); 2310 2311 oldflags = bp->b_flags; 2312 2313 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 2314 ("FFS background buffer should not get here %p", bp)); 2315 2316 vp = bp->b_vp; 2317 if (vp) 2318 vp_md = vp->v_vflag & VV_MD; 2319 else 2320 vp_md = 0; 2321 2322 /* 2323 * Mark the buffer clean. Increment the bufobj write count 2324 * before bundirty() call, to prevent other thread from seeing 2325 * empty dirty list and zero counter for writes in progress, 2326 * falsely indicating that the bufobj is clean. 2327 */ 2328 bufobj_wref(bp->b_bufobj); 2329 bundirty(bp); 2330 2331 bp->b_flags &= ~B_DONE; 2332 bp->b_ioflags &= ~BIO_ERROR; 2333 bp->b_flags |= B_CACHE; 2334 bp->b_iocmd = BIO_WRITE; 2335 2336 vfs_busy_pages(bp, 1); 2337 2338 /* 2339 * Normal bwrites pipeline writes 2340 */ 2341 bp->b_runningbufspace = bp->b_bufsize; 2342 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); 2343 2344 #ifdef RACCT 2345 if (racct_enable) { 2346 PROC_LOCK(curproc); 2347 racct_add_buf(curproc, bp, 1); 2348 PROC_UNLOCK(curproc); 2349 } 2350 #endif /* RACCT */ 2351 curthread->td_ru.ru_oublock++; 2352 if (oldflags & B_ASYNC) 2353 BUF_KERNPROC(bp); 2354 bp->b_iooffset = dbtob(bp->b_blkno); 2355 buf_track(bp, __func__); 2356 bstrategy(bp); 2357 2358 if ((oldflags & B_ASYNC) == 0) { 2359 int rtval = bufwait(bp); 2360 brelse(bp); 2361 return (rtval); 2362 } else if (space > hirunningspace) { 2363 /* 2364 * don't allow the async write to saturate the I/O 2365 * system. We will not deadlock here because 2366 * we are blocking waiting for I/O that is already in-progress 2367 * to complete. We do not block here if it is the update 2368 * or syncer daemon trying to clean up as that can lead 2369 * to deadlock. 2370 */ 2371 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 2372 waitrunningbufspace(); 2373 } 2374 2375 return (0); 2376 } 2377 2378 void 2379 bufbdflush(struct bufobj *bo, struct buf *bp) 2380 { 2381 struct buf *nbp; 2382 struct bufdomain *bd; 2383 2384 bd = &bdomain[bo->bo_domain]; 2385 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) { 2386 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 2387 altbufferflushes++; 2388 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) { 2389 BO_LOCK(bo); 2390 /* 2391 * Try to find a buffer to flush. 2392 */ 2393 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 2394 if ((nbp->b_vflags & BV_BKGRDINPROG) || 2395 BUF_LOCK(nbp, 2396 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 2397 continue; 2398 if (bp == nbp) 2399 panic("bdwrite: found ourselves"); 2400 BO_UNLOCK(bo); 2401 /* Don't countdeps with the bo lock held. */ 2402 if (buf_countdeps(nbp, 0)) { 2403 BO_LOCK(bo); 2404 BUF_UNLOCK(nbp); 2405 continue; 2406 } 2407 if (nbp->b_flags & B_CLUSTEROK) { 2408 vfs_bio_awrite(nbp); 2409 } else { 2410 bremfree(nbp); 2411 bawrite(nbp); 2412 } 2413 dirtybufferflushes++; 2414 break; 2415 } 2416 if (nbp == NULL) 2417 BO_UNLOCK(bo); 2418 } 2419 } 2420 2421 /* 2422 * Delayed write. (Buffer is marked dirty). Do not bother writing 2423 * anything if the buffer is marked invalid. 2424 * 2425 * Note that since the buffer must be completely valid, we can safely 2426 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 2427 * biodone() in order to prevent getblk from writing the buffer 2428 * out synchronously. 2429 */ 2430 void 2431 bdwrite(struct buf *bp) 2432 { 2433 struct thread *td = curthread; 2434 struct vnode *vp; 2435 struct bufobj *bo; 2436 2437 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2438 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2439 KASSERT((bp->b_flags & B_BARRIER) == 0, 2440 ("Barrier request in delayed write %p", bp)); 2441 2442 if (bp->b_flags & B_INVAL) { 2443 brelse(bp); 2444 return; 2445 } 2446 2447 /* 2448 * If we have too many dirty buffers, don't create any more. 2449 * If we are wildly over our limit, then force a complete 2450 * cleanup. Otherwise, just keep the situation from getting 2451 * out of control. Note that we have to avoid a recursive 2452 * disaster and not try to clean up after our own cleanup! 2453 */ 2454 vp = bp->b_vp; 2455 bo = bp->b_bufobj; 2456 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 2457 td->td_pflags |= TDP_INBDFLUSH; 2458 BO_BDFLUSH(bo, bp); 2459 td->td_pflags &= ~TDP_INBDFLUSH; 2460 } else 2461 recursiveflushes++; 2462 2463 bdirty(bp); 2464 /* 2465 * Set B_CACHE, indicating that the buffer is fully valid. This is 2466 * true even of NFS now. 2467 */ 2468 bp->b_flags |= B_CACHE; 2469 2470 /* 2471 * This bmap keeps the system from needing to do the bmap later, 2472 * perhaps when the system is attempting to do a sync. Since it 2473 * is likely that the indirect block -- or whatever other datastructure 2474 * that the filesystem needs is still in memory now, it is a good 2475 * thing to do this. Note also, that if the pageout daemon is 2476 * requesting a sync -- there might not be enough memory to do 2477 * the bmap then... So, this is important to do. 2478 */ 2479 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 2480 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 2481 } 2482 2483 buf_track(bp, __func__); 2484 2485 /* 2486 * Set the *dirty* buffer range based upon the VM system dirty 2487 * pages. 2488 * 2489 * Mark the buffer pages as clean. We need to do this here to 2490 * satisfy the vnode_pager and the pageout daemon, so that it 2491 * thinks that the pages have been "cleaned". Note that since 2492 * the pages are in a delayed write buffer -- the VFS layer 2493 * "will" see that the pages get written out on the next sync, 2494 * or perhaps the cluster will be completed. 2495 */ 2496 vfs_clean_pages_dirty_buf(bp); 2497 bqrelse(bp); 2498 2499 /* 2500 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 2501 * due to the softdep code. 2502 */ 2503 } 2504 2505 /* 2506 * bdirty: 2507 * 2508 * Turn buffer into delayed write request. We must clear BIO_READ and 2509 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 2510 * itself to properly update it in the dirty/clean lists. We mark it 2511 * B_DONE to ensure that any asynchronization of the buffer properly 2512 * clears B_DONE ( else a panic will occur later ). 2513 * 2514 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 2515 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 2516 * should only be called if the buffer is known-good. 2517 * 2518 * Since the buffer is not on a queue, we do not update the numfreebuffers 2519 * count. 2520 * 2521 * The buffer must be on QUEUE_NONE. 2522 */ 2523 void 2524 bdirty(struct buf *bp) 2525 { 2526 2527 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 2528 bp, bp->b_vp, bp->b_flags); 2529 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2530 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2531 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2532 bp->b_flags &= ~(B_RELBUF); 2533 bp->b_iocmd = BIO_WRITE; 2534 2535 if ((bp->b_flags & B_DELWRI) == 0) { 2536 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 2537 reassignbuf(bp); 2538 bdirtyadd(bp); 2539 } 2540 } 2541 2542 /* 2543 * bundirty: 2544 * 2545 * Clear B_DELWRI for buffer. 2546 * 2547 * Since the buffer is not on a queue, we do not update the numfreebuffers 2548 * count. 2549 * 2550 * The buffer must be on QUEUE_NONE. 2551 */ 2552 2553 void 2554 bundirty(struct buf *bp) 2555 { 2556 2557 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2558 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2559 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2560 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2561 2562 if (bp->b_flags & B_DELWRI) { 2563 bp->b_flags &= ~B_DELWRI; 2564 reassignbuf(bp); 2565 bdirtysub(bp); 2566 } 2567 /* 2568 * Since it is now being written, we can clear its deferred write flag. 2569 */ 2570 bp->b_flags &= ~B_DEFERRED; 2571 } 2572 2573 /* 2574 * bawrite: 2575 * 2576 * Asynchronous write. Start output on a buffer, but do not wait for 2577 * it to complete. The buffer is released when the output completes. 2578 * 2579 * bwrite() ( or the VOP routine anyway ) is responsible for handling 2580 * B_INVAL buffers. Not us. 2581 */ 2582 void 2583 bawrite(struct buf *bp) 2584 { 2585 2586 bp->b_flags |= B_ASYNC; 2587 (void) bwrite(bp); 2588 } 2589 2590 /* 2591 * babarrierwrite: 2592 * 2593 * Asynchronous barrier write. Start output on a buffer, but do not 2594 * wait for it to complete. Place a write barrier after this write so 2595 * that this buffer and all buffers written before it are committed to 2596 * the disk before any buffers written after this write are committed 2597 * to the disk. The buffer is released when the output completes. 2598 */ 2599 void 2600 babarrierwrite(struct buf *bp) 2601 { 2602 2603 bp->b_flags |= B_ASYNC | B_BARRIER; 2604 (void) bwrite(bp); 2605 } 2606 2607 /* 2608 * bbarrierwrite: 2609 * 2610 * Synchronous barrier write. Start output on a buffer and wait for 2611 * it to complete. Place a write barrier after this write so that 2612 * this buffer and all buffers written before it are committed to 2613 * the disk before any buffers written after this write are committed 2614 * to the disk. The buffer is released when the output completes. 2615 */ 2616 int 2617 bbarrierwrite(struct buf *bp) 2618 { 2619 2620 bp->b_flags |= B_BARRIER; 2621 return (bwrite(bp)); 2622 } 2623 2624 /* 2625 * bwillwrite: 2626 * 2627 * Called prior to the locking of any vnodes when we are expecting to 2628 * write. We do not want to starve the buffer cache with too many 2629 * dirty buffers so we block here. By blocking prior to the locking 2630 * of any vnodes we attempt to avoid the situation where a locked vnode 2631 * prevents the various system daemons from flushing related buffers. 2632 */ 2633 void 2634 bwillwrite(void) 2635 { 2636 2637 if (buf_dirty_count_severe()) { 2638 mtx_lock(&bdirtylock); 2639 while (buf_dirty_count_severe()) { 2640 bdirtywait = 1; 2641 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), 2642 "flswai", 0); 2643 } 2644 mtx_unlock(&bdirtylock); 2645 } 2646 } 2647 2648 /* 2649 * Return true if we have too many dirty buffers. 2650 */ 2651 int 2652 buf_dirty_count_severe(void) 2653 { 2654 2655 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty)); 2656 } 2657 2658 /* 2659 * brelse: 2660 * 2661 * Release a busy buffer and, if requested, free its resources. The 2662 * buffer will be stashed in the appropriate bufqueue[] allowing it 2663 * to be accessed later as a cache entity or reused for other purposes. 2664 */ 2665 void 2666 brelse(struct buf *bp) 2667 { 2668 struct mount *v_mnt; 2669 int qindex; 2670 2671 /* 2672 * Many functions erroneously call brelse with a NULL bp under rare 2673 * error conditions. Simply return when called with a NULL bp. 2674 */ 2675 if (bp == NULL) 2676 return; 2677 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 2678 bp, bp->b_vp, bp->b_flags); 2679 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2680 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2681 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0, 2682 ("brelse: non-VMIO buffer marked NOREUSE")); 2683 2684 if (BUF_LOCKRECURSED(bp)) { 2685 /* 2686 * Do not process, in particular, do not handle the 2687 * B_INVAL/B_RELBUF and do not release to free list. 2688 */ 2689 BUF_UNLOCK(bp); 2690 return; 2691 } 2692 2693 if (bp->b_flags & B_MANAGED) { 2694 bqrelse(bp); 2695 return; 2696 } 2697 2698 if (LIST_EMPTY(&bp->b_dep)) { 2699 bp->b_flags &= ~B_IOSTARTED; 2700 } else { 2701 KASSERT((bp->b_flags & B_IOSTARTED) == 0, 2702 ("brelse: SU io not finished bp %p", bp)); 2703 } 2704 2705 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { 2706 BO_LOCK(bp->b_bufobj); 2707 bp->b_vflags &= ~BV_BKGRDERR; 2708 BO_UNLOCK(bp->b_bufobj); 2709 bdirty(bp); 2710 } 2711 2712 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 2713 (bp->b_flags & B_INVALONERR)) { 2714 /* 2715 * Forced invalidation of dirty buffer contents, to be used 2716 * after a failed write in the rare case that the loss of the 2717 * contents is acceptable. The buffer is invalidated and 2718 * freed. 2719 */ 2720 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE; 2721 bp->b_flags &= ~(B_ASYNC | B_CACHE); 2722 } 2723 2724 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 2725 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) && 2726 !(bp->b_flags & B_INVAL)) { 2727 /* 2728 * Failed write, redirty. All errors except ENXIO (which 2729 * means the device is gone) are treated as being 2730 * transient. 2731 * 2732 * XXX Treating EIO as transient is not correct; the 2733 * contract with the local storage device drivers is that 2734 * they will only return EIO once the I/O is no longer 2735 * retriable. Network I/O also respects this through the 2736 * guarantees of TCP and/or the internal retries of NFS. 2737 * ENOMEM might be transient, but we also have no way of 2738 * knowing when its ok to retry/reschedule. In general, 2739 * this entire case should be made obsolete through better 2740 * error handling/recovery and resource scheduling. 2741 * 2742 * Do this also for buffers that failed with ENXIO, but have 2743 * non-empty dependencies - the soft updates code might need 2744 * to access the buffer to untangle them. 2745 * 2746 * Must clear BIO_ERROR to prevent pages from being scrapped. 2747 */ 2748 bp->b_ioflags &= ~BIO_ERROR; 2749 bdirty(bp); 2750 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 2751 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 2752 /* 2753 * Either a failed read I/O, or we were asked to free or not 2754 * cache the buffer, or we failed to write to a device that's 2755 * no longer present. 2756 */ 2757 bp->b_flags |= B_INVAL; 2758 if (!LIST_EMPTY(&bp->b_dep)) 2759 buf_deallocate(bp); 2760 if (bp->b_flags & B_DELWRI) 2761 bdirtysub(bp); 2762 bp->b_flags &= ~(B_DELWRI | B_CACHE); 2763 if ((bp->b_flags & B_VMIO) == 0) { 2764 allocbuf(bp, 0); 2765 if (bp->b_vp) 2766 brelvp(bp); 2767 } 2768 } 2769 2770 /* 2771 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate() 2772 * is called with B_DELWRI set, the underlying pages may wind up 2773 * getting freed causing a previous write (bdwrite()) to get 'lost' 2774 * because pages associated with a B_DELWRI bp are marked clean. 2775 * 2776 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even 2777 * if B_DELWRI is set. 2778 */ 2779 if (bp->b_flags & B_DELWRI) 2780 bp->b_flags &= ~B_RELBUF; 2781 2782 /* 2783 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 2784 * constituted, not even NFS buffers now. Two flags effect this. If 2785 * B_INVAL, the struct buf is invalidated but the VM object is kept 2786 * around ( i.e. so it is trivial to reconstitute the buffer later ). 2787 * 2788 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 2789 * invalidated. BIO_ERROR cannot be set for a failed write unless the 2790 * buffer is also B_INVAL because it hits the re-dirtying code above. 2791 * 2792 * Normally we can do this whether a buffer is B_DELWRI or not. If 2793 * the buffer is an NFS buffer, it is tracking piecemeal writes or 2794 * the commit state and we cannot afford to lose the buffer. If the 2795 * buffer has a background write in progress, we need to keep it 2796 * around to prevent it from being reconstituted and starting a second 2797 * background write. 2798 */ 2799 2800 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL; 2801 2802 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE || 2803 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) && 2804 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 || 2805 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) { 2806 vfs_vmio_invalidate(bp); 2807 allocbuf(bp, 0); 2808 } 2809 2810 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 || 2811 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) { 2812 allocbuf(bp, 0); 2813 bp->b_flags &= ~B_NOREUSE; 2814 if (bp->b_vp != NULL) 2815 brelvp(bp); 2816 } 2817 2818 /* 2819 * If the buffer has junk contents signal it and eventually 2820 * clean up B_DELWRI and diassociate the vnode so that gbincore() 2821 * doesn't find it. 2822 */ 2823 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 2824 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 2825 bp->b_flags |= B_INVAL; 2826 if (bp->b_flags & B_INVAL) { 2827 if (bp->b_flags & B_DELWRI) 2828 bundirty(bp); 2829 if (bp->b_vp) 2830 brelvp(bp); 2831 } 2832 2833 buf_track(bp, __func__); 2834 2835 /* buffers with no memory */ 2836 if (bp->b_bufsize == 0) { 2837 buf_free(bp); 2838 return; 2839 } 2840 /* buffers with junk contents */ 2841 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 2842 (bp->b_ioflags & BIO_ERROR)) { 2843 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 2844 if (bp->b_vflags & BV_BKGRDINPROG) 2845 panic("losing buffer 2"); 2846 qindex = QUEUE_CLEAN; 2847 bp->b_flags |= B_AGE; 2848 /* remaining buffers */ 2849 } else if (bp->b_flags & B_DELWRI) 2850 qindex = QUEUE_DIRTY; 2851 else 2852 qindex = QUEUE_CLEAN; 2853 2854 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 2855 panic("brelse: not dirty"); 2856 2857 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT); 2858 bp->b_xflags &= ~(BX_CVTENXIO); 2859 /* binsfree unlocks bp. */ 2860 binsfree(bp, qindex); 2861 } 2862 2863 /* 2864 * Release a buffer back to the appropriate queue but do not try to free 2865 * it. The buffer is expected to be used again soon. 2866 * 2867 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 2868 * biodone() to requeue an async I/O on completion. It is also used when 2869 * known good buffers need to be requeued but we think we may need the data 2870 * again soon. 2871 * 2872 * XXX we should be able to leave the B_RELBUF hint set on completion. 2873 */ 2874 void 2875 bqrelse(struct buf *bp) 2876 { 2877 int qindex; 2878 2879 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2880 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2881 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2882 2883 qindex = QUEUE_NONE; 2884 if (BUF_LOCKRECURSED(bp)) { 2885 /* do not release to free list */ 2886 BUF_UNLOCK(bp); 2887 return; 2888 } 2889 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 2890 bp->b_xflags &= ~(BX_CVTENXIO); 2891 2892 if (LIST_EMPTY(&bp->b_dep)) { 2893 bp->b_flags &= ~B_IOSTARTED; 2894 } else { 2895 KASSERT((bp->b_flags & B_IOSTARTED) == 0, 2896 ("bqrelse: SU io not finished bp %p", bp)); 2897 } 2898 2899 if (bp->b_flags & B_MANAGED) { 2900 if (bp->b_flags & B_REMFREE) 2901 bremfreef(bp); 2902 goto out; 2903 } 2904 2905 /* buffers with stale but valid contents */ 2906 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | 2907 BV_BKGRDERR)) == BV_BKGRDERR) { 2908 BO_LOCK(bp->b_bufobj); 2909 bp->b_vflags &= ~BV_BKGRDERR; 2910 BO_UNLOCK(bp->b_bufobj); 2911 qindex = QUEUE_DIRTY; 2912 } else { 2913 if ((bp->b_flags & B_DELWRI) == 0 && 2914 (bp->b_xflags & BX_VNDIRTY)) 2915 panic("bqrelse: not dirty"); 2916 if ((bp->b_flags & B_NOREUSE) != 0) { 2917 brelse(bp); 2918 return; 2919 } 2920 qindex = QUEUE_CLEAN; 2921 } 2922 buf_track(bp, __func__); 2923 /* binsfree unlocks bp. */ 2924 binsfree(bp, qindex); 2925 return; 2926 2927 out: 2928 buf_track(bp, __func__); 2929 /* unlock */ 2930 BUF_UNLOCK(bp); 2931 } 2932 2933 /* 2934 * Complete I/O to a VMIO backed page. Validate the pages as appropriate, 2935 * restore bogus pages. 2936 */ 2937 static void 2938 vfs_vmio_iodone(struct buf *bp) 2939 { 2940 vm_ooffset_t foff; 2941 vm_page_t m; 2942 vm_object_t obj; 2943 struct vnode *vp __unused; 2944 int i, iosize, resid; 2945 bool bogus; 2946 2947 obj = bp->b_bufobj->bo_object; 2948 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages, 2949 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)", 2950 blockcount_read(&obj->paging_in_progress), bp->b_npages)); 2951 2952 vp = bp->b_vp; 2953 VNPASS(vp->v_holdcnt > 0, vp); 2954 VNPASS(vp->v_object != NULL, vp); 2955 2956 foff = bp->b_offset; 2957 KASSERT(bp->b_offset != NOOFFSET, 2958 ("vfs_vmio_iodone: bp %p has no buffer offset", bp)); 2959 2960 bogus = false; 2961 iosize = bp->b_bcount - bp->b_resid; 2962 for (i = 0; i < bp->b_npages; i++) { 2963 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2964 if (resid > iosize) 2965 resid = iosize; 2966 2967 /* 2968 * cleanup bogus pages, restoring the originals 2969 */ 2970 m = bp->b_pages[i]; 2971 if (m == bogus_page) { 2972 bogus = true; 2973 m = vm_page_relookup(obj, OFF_TO_IDX(foff)); 2974 if (m == NULL) 2975 panic("biodone: page disappeared!"); 2976 bp->b_pages[i] = m; 2977 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) { 2978 /* 2979 * In the write case, the valid and clean bits are 2980 * already changed correctly ( see bdwrite() ), so we 2981 * only need to do this here in the read case. 2982 */ 2983 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK, 2984 resid)) == 0, ("vfs_vmio_iodone: page %p " 2985 "has unexpected dirty bits", m)); 2986 vfs_page_set_valid(bp, foff, m); 2987 } 2988 KASSERT(OFF_TO_IDX(foff) == m->pindex, 2989 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch", 2990 (intmax_t)foff, (uintmax_t)m->pindex)); 2991 2992 vm_page_sunbusy(m); 2993 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2994 iosize -= resid; 2995 } 2996 vm_object_pip_wakeupn(obj, bp->b_npages); 2997 if (bogus && buf_mapped(bp)) { 2998 BUF_CHECK_MAPPED(bp); 2999 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3000 bp->b_pages, bp->b_npages); 3001 } 3002 } 3003 3004 /* 3005 * Perform page invalidation when a buffer is released. The fully invalid 3006 * pages will be reclaimed later in vfs_vmio_truncate(). 3007 */ 3008 static void 3009 vfs_vmio_invalidate(struct buf *bp) 3010 { 3011 vm_object_t obj; 3012 vm_page_t m; 3013 int flags, i, resid, poffset, presid; 3014 3015 if (buf_mapped(bp)) { 3016 BUF_CHECK_MAPPED(bp); 3017 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 3018 } else 3019 BUF_CHECK_UNMAPPED(bp); 3020 /* 3021 * Get the base offset and length of the buffer. Note that 3022 * in the VMIO case if the buffer block size is not 3023 * page-aligned then b_data pointer may not be page-aligned. 3024 * But our b_pages[] array *IS* page aligned. 3025 * 3026 * block sizes less then DEV_BSIZE (usually 512) are not 3027 * supported due to the page granularity bits (m->valid, 3028 * m->dirty, etc...). 3029 * 3030 * See man buf(9) for more information 3031 */ 3032 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; 3033 obj = bp->b_bufobj->bo_object; 3034 resid = bp->b_bufsize; 3035 poffset = bp->b_offset & PAGE_MASK; 3036 VM_OBJECT_WLOCK(obj); 3037 for (i = 0; i < bp->b_npages; i++) { 3038 m = bp->b_pages[i]; 3039 if (m == bogus_page) 3040 panic("vfs_vmio_invalidate: Unexpected bogus page."); 3041 bp->b_pages[i] = NULL; 3042 3043 presid = resid > (PAGE_SIZE - poffset) ? 3044 (PAGE_SIZE - poffset) : resid; 3045 KASSERT(presid >= 0, ("brelse: extra page")); 3046 vm_page_busy_acquire(m, VM_ALLOC_SBUSY); 3047 if (pmap_page_wired_mappings(m) == 0) 3048 vm_page_set_invalid(m, poffset, presid); 3049 vm_page_sunbusy(m); 3050 vm_page_release_locked(m, flags); 3051 resid -= presid; 3052 poffset = 0; 3053 } 3054 VM_OBJECT_WUNLOCK(obj); 3055 bp->b_npages = 0; 3056 } 3057 3058 /* 3059 * Page-granular truncation of an existing VMIO buffer. 3060 */ 3061 static void 3062 vfs_vmio_truncate(struct buf *bp, int desiredpages) 3063 { 3064 vm_object_t obj; 3065 vm_page_t m; 3066 int flags, i; 3067 3068 if (bp->b_npages == desiredpages) 3069 return; 3070 3071 if (buf_mapped(bp)) { 3072 BUF_CHECK_MAPPED(bp); 3073 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) + 3074 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages); 3075 } else 3076 BUF_CHECK_UNMAPPED(bp); 3077 3078 /* 3079 * The object lock is needed only if we will attempt to free pages. 3080 */ 3081 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; 3082 if ((bp->b_flags & B_DIRECT) != 0) { 3083 flags |= VPR_TRYFREE; 3084 obj = bp->b_bufobj->bo_object; 3085 VM_OBJECT_WLOCK(obj); 3086 } else { 3087 obj = NULL; 3088 } 3089 for (i = desiredpages; i < bp->b_npages; i++) { 3090 m = bp->b_pages[i]; 3091 KASSERT(m != bogus_page, ("allocbuf: bogus page found")); 3092 bp->b_pages[i] = NULL; 3093 if (obj != NULL) 3094 vm_page_release_locked(m, flags); 3095 else 3096 vm_page_release(m, flags); 3097 } 3098 if (obj != NULL) 3099 VM_OBJECT_WUNLOCK(obj); 3100 bp->b_npages = desiredpages; 3101 } 3102 3103 /* 3104 * Byte granular extension of VMIO buffers. 3105 */ 3106 static void 3107 vfs_vmio_extend(struct buf *bp, int desiredpages, int size) 3108 { 3109 /* 3110 * We are growing the buffer, possibly in a 3111 * byte-granular fashion. 3112 */ 3113 vm_object_t obj; 3114 vm_offset_t toff; 3115 vm_offset_t tinc; 3116 vm_page_t m; 3117 3118 /* 3119 * Step 1, bring in the VM pages from the object, allocating 3120 * them if necessary. We must clear B_CACHE if these pages 3121 * are not valid for the range covered by the buffer. 3122 */ 3123 obj = bp->b_bufobj->bo_object; 3124 if (bp->b_npages < desiredpages) { 3125 KASSERT(desiredpages <= atop(maxbcachebuf), 3126 ("vfs_vmio_extend past maxbcachebuf %p %d %u", 3127 bp, desiredpages, maxbcachebuf)); 3128 3129 /* 3130 * We must allocate system pages since blocking 3131 * here could interfere with paging I/O, no 3132 * matter which process we are. 3133 * 3134 * Only exclusive busy can be tested here. 3135 * Blocking on shared busy might lead to 3136 * deadlocks once allocbuf() is called after 3137 * pages are vfs_busy_pages(). 3138 */ 3139 (void)vm_page_grab_pages_unlocked(obj, 3140 OFF_TO_IDX(bp->b_offset) + bp->b_npages, 3141 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY | 3142 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED, 3143 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages); 3144 bp->b_npages = desiredpages; 3145 } 3146 3147 /* 3148 * Step 2. We've loaded the pages into the buffer, 3149 * we have to figure out if we can still have B_CACHE 3150 * set. Note that B_CACHE is set according to the 3151 * byte-granular range ( bcount and size ), not the 3152 * aligned range ( newbsize ). 3153 * 3154 * The VM test is against m->valid, which is DEV_BSIZE 3155 * aligned. Needless to say, the validity of the data 3156 * needs to also be DEV_BSIZE aligned. Note that this 3157 * fails with NFS if the server or some other client 3158 * extends the file's EOF. If our buffer is resized, 3159 * B_CACHE may remain set! XXX 3160 */ 3161 toff = bp->b_bcount; 3162 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3163 while ((bp->b_flags & B_CACHE) && toff < size) { 3164 vm_pindex_t pi; 3165 3166 if (tinc > (size - toff)) 3167 tinc = size - toff; 3168 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; 3169 m = bp->b_pages[pi]; 3170 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m); 3171 toff += tinc; 3172 tinc = PAGE_SIZE; 3173 } 3174 3175 /* 3176 * Step 3, fixup the KVA pmap. 3177 */ 3178 if (buf_mapped(bp)) 3179 bpmap_qenter(bp); 3180 else 3181 BUF_CHECK_UNMAPPED(bp); 3182 } 3183 3184 /* 3185 * Check to see if a block at a particular lbn is available for a clustered 3186 * write. 3187 */ 3188 static int 3189 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 3190 { 3191 struct buf *bpa; 3192 int match; 3193 3194 match = 0; 3195 3196 /* If the buf isn't in core skip it */ 3197 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 3198 return (0); 3199 3200 /* If the buf is busy we don't want to wait for it */ 3201 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 3202 return (0); 3203 3204 /* Only cluster with valid clusterable delayed write buffers */ 3205 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 3206 (B_DELWRI | B_CLUSTEROK)) 3207 goto done; 3208 3209 if (bpa->b_bufsize != size) 3210 goto done; 3211 3212 /* 3213 * Check to see if it is in the expected place on disk and that the 3214 * block has been mapped. 3215 */ 3216 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 3217 match = 1; 3218 done: 3219 BUF_UNLOCK(bpa); 3220 return (match); 3221 } 3222 3223 /* 3224 * vfs_bio_awrite: 3225 * 3226 * Implement clustered async writes for clearing out B_DELWRI buffers. 3227 * This is much better then the old way of writing only one buffer at 3228 * a time. Note that we may not be presented with the buffers in the 3229 * correct order, so we search for the cluster in both directions. 3230 */ 3231 int 3232 vfs_bio_awrite(struct buf *bp) 3233 { 3234 struct bufobj *bo; 3235 int i; 3236 int j; 3237 daddr_t lblkno = bp->b_lblkno; 3238 struct vnode *vp = bp->b_vp; 3239 int ncl; 3240 int nwritten; 3241 int size; 3242 int maxcl; 3243 int gbflags; 3244 3245 bo = &vp->v_bufobj; 3246 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0; 3247 /* 3248 * right now we support clustered writing only to regular files. If 3249 * we find a clusterable block we could be in the middle of a cluster 3250 * rather then at the beginning. 3251 */ 3252 if ((vp->v_type == VREG) && 3253 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 3254 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 3255 size = vp->v_mount->mnt_stat.f_iosize; 3256 maxcl = maxphys / size; 3257 3258 BO_RLOCK(bo); 3259 for (i = 1; i < maxcl; i++) 3260 if (vfs_bio_clcheck(vp, size, lblkno + i, 3261 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 3262 break; 3263 3264 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 3265 if (vfs_bio_clcheck(vp, size, lblkno - j, 3266 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 3267 break; 3268 BO_RUNLOCK(bo); 3269 --j; 3270 ncl = i + j; 3271 /* 3272 * this is a possible cluster write 3273 */ 3274 if (ncl != 1) { 3275 BUF_UNLOCK(bp); 3276 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 3277 gbflags); 3278 return (nwritten); 3279 } 3280 } 3281 bremfree(bp); 3282 bp->b_flags |= B_ASYNC; 3283 /* 3284 * default (old) behavior, writing out only one block 3285 * 3286 * XXX returns b_bufsize instead of b_bcount for nwritten? 3287 */ 3288 nwritten = bp->b_bufsize; 3289 (void) bwrite(bp); 3290 3291 return (nwritten); 3292 } 3293 3294 /* 3295 * getnewbuf_kva: 3296 * 3297 * Allocate KVA for an empty buf header according to gbflags. 3298 */ 3299 static int 3300 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize) 3301 { 3302 3303 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) { 3304 /* 3305 * In order to keep fragmentation sane we only allocate kva 3306 * in BKVASIZE chunks. XXX with vmem we can do page size. 3307 */ 3308 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 3309 3310 if (maxsize != bp->b_kvasize && 3311 bufkva_alloc(bp, maxsize, gbflags)) 3312 return (ENOSPC); 3313 } 3314 return (0); 3315 } 3316 3317 /* 3318 * getnewbuf: 3319 * 3320 * Find and initialize a new buffer header, freeing up existing buffers 3321 * in the bufqueues as necessary. The new buffer is returned locked. 3322 * 3323 * We block if: 3324 * We have insufficient buffer headers 3325 * We have insufficient buffer space 3326 * buffer_arena is too fragmented ( space reservation fails ) 3327 * If we have to flush dirty buffers ( but we try to avoid this ) 3328 * 3329 * The caller is responsible for releasing the reserved bufspace after 3330 * allocbuf() is called. 3331 */ 3332 static struct buf * 3333 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags) 3334 { 3335 struct bufdomain *bd; 3336 struct buf *bp; 3337 bool metadata, reserved; 3338 3339 bp = NULL; 3340 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3341 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3342 if (!unmapped_buf_allowed) 3343 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3344 3345 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 3346 vp->v_type == VCHR) 3347 metadata = true; 3348 else 3349 metadata = false; 3350 if (vp == NULL) 3351 bd = &bdomain[0]; 3352 else 3353 bd = &bdomain[vp->v_bufobj.bo_domain]; 3354 3355 counter_u64_add(getnewbufcalls, 1); 3356 reserved = false; 3357 do { 3358 if (reserved == false && 3359 bufspace_reserve(bd, maxsize, metadata) != 0) { 3360 counter_u64_add(getnewbufrestarts, 1); 3361 continue; 3362 } 3363 reserved = true; 3364 if ((bp = buf_alloc(bd)) == NULL) { 3365 counter_u64_add(getnewbufrestarts, 1); 3366 continue; 3367 } 3368 if (getnewbuf_kva(bp, gbflags, maxsize) == 0) 3369 return (bp); 3370 break; 3371 } while (buf_recycle(bd, false) == 0); 3372 3373 if (reserved) 3374 bufspace_release(bd, maxsize); 3375 if (bp != NULL) { 3376 bp->b_flags |= B_INVAL; 3377 brelse(bp); 3378 } 3379 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo); 3380 3381 return (NULL); 3382 } 3383 3384 /* 3385 * buf_daemon: 3386 * 3387 * buffer flushing daemon. Buffers are normally flushed by the 3388 * update daemon but if it cannot keep up this process starts to 3389 * take the load in an attempt to prevent getnewbuf() from blocking. 3390 */ 3391 static struct kproc_desc buf_kp = { 3392 "bufdaemon", 3393 buf_daemon, 3394 &bufdaemonproc 3395 }; 3396 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 3397 3398 static int 3399 buf_flush(struct vnode *vp, struct bufdomain *bd, int target) 3400 { 3401 int flushed; 3402 3403 flushed = flushbufqueues(vp, bd, target, 0); 3404 if (flushed == 0) { 3405 /* 3406 * Could not find any buffers without rollback 3407 * dependencies, so just write the first one 3408 * in the hopes of eventually making progress. 3409 */ 3410 if (vp != NULL && target > 2) 3411 target /= 2; 3412 flushbufqueues(vp, bd, target, 1); 3413 } 3414 return (flushed); 3415 } 3416 3417 static void 3418 buf_daemon_shutdown(void *arg __unused, int howto __unused) 3419 { 3420 int error; 3421 3422 mtx_lock(&bdlock); 3423 bd_shutdown = true; 3424 wakeup(&bd_request); 3425 error = msleep(&bd_shutdown, &bdlock, 0, "buf_daemon_shutdown", 3426 60 * hz); 3427 mtx_unlock(&bdlock); 3428 if (error != 0) 3429 printf("bufdaemon wait error: %d\n", error); 3430 } 3431 3432 static void 3433 buf_daemon(void) 3434 { 3435 struct bufdomain *bd; 3436 int speedupreq; 3437 int lodirty; 3438 int i; 3439 3440 /* 3441 * This process needs to be suspended prior to shutdown sync. 3442 */ 3443 EVENTHANDLER_REGISTER(shutdown_pre_sync, buf_daemon_shutdown, NULL, 3444 SHUTDOWN_PRI_LAST + 100); 3445 3446 /* 3447 * Start the buf clean daemons as children threads. 3448 */ 3449 for (i = 0 ; i < buf_domains; i++) { 3450 int error; 3451 3452 error = kthread_add((void (*)(void *))bufspace_daemon, 3453 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i); 3454 if (error) 3455 panic("error %d spawning bufspace daemon", error); 3456 } 3457 3458 /* 3459 * This process is allowed to take the buffer cache to the limit 3460 */ 3461 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 3462 mtx_lock(&bdlock); 3463 while (!bd_shutdown) { 3464 bd_request = 0; 3465 mtx_unlock(&bdlock); 3466 3467 /* 3468 * Save speedupreq for this pass and reset to capture new 3469 * requests. 3470 */ 3471 speedupreq = bd_speedupreq; 3472 bd_speedupreq = 0; 3473 3474 /* 3475 * Flush each domain sequentially according to its level and 3476 * the speedup request. 3477 */ 3478 for (i = 0; i < buf_domains; i++) { 3479 bd = &bdomain[i]; 3480 if (speedupreq) 3481 lodirty = bd->bd_numdirtybuffers / 2; 3482 else 3483 lodirty = bd->bd_lodirtybuffers; 3484 while (bd->bd_numdirtybuffers > lodirty) { 3485 if (buf_flush(NULL, bd, 3486 bd->bd_numdirtybuffers - lodirty) == 0) 3487 break; 3488 kern_yield(PRI_USER); 3489 } 3490 } 3491 3492 /* 3493 * Only clear bd_request if we have reached our low water 3494 * mark. The buf_daemon normally waits 1 second and 3495 * then incrementally flushes any dirty buffers that have 3496 * built up, within reason. 3497 * 3498 * If we were unable to hit our low water mark and couldn't 3499 * find any flushable buffers, we sleep for a short period 3500 * to avoid endless loops on unlockable buffers. 3501 */ 3502 mtx_lock(&bdlock); 3503 if (bd_shutdown) 3504 break; 3505 if (BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) { 3506 /* 3507 * We reached our low water mark, reset the 3508 * request and sleep until we are needed again. 3509 * The sleep is just so the suspend code works. 3510 */ 3511 bd_request = 0; 3512 /* 3513 * Do an extra wakeup in case dirty threshold 3514 * changed via sysctl and the explicit transition 3515 * out of shortfall was missed. 3516 */ 3517 bdirtywakeup(); 3518 if (runningbufspace <= lorunningspace) 3519 runningwakeup(); 3520 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 3521 } else { 3522 /* 3523 * We couldn't find any flushable dirty buffers but 3524 * still have too many dirty buffers, we 3525 * have to sleep and try again. (rare) 3526 */ 3527 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 3528 } 3529 } 3530 wakeup(&bd_shutdown); 3531 mtx_unlock(&bdlock); 3532 kthread_exit(); 3533 } 3534 3535 /* 3536 * flushbufqueues: 3537 * 3538 * Try to flush a buffer in the dirty queue. We must be careful to 3539 * free up B_INVAL buffers instead of write them, which NFS is 3540 * particularly sensitive to. 3541 */ 3542 static int flushwithdeps = 0; 3543 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS, 3544 &flushwithdeps, 0, 3545 "Number of buffers flushed with dependencies that require rollbacks"); 3546 3547 static int 3548 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target, 3549 int flushdeps) 3550 { 3551 struct bufqueue *bq; 3552 struct buf *sentinel; 3553 struct vnode *vp; 3554 struct mount *mp; 3555 struct buf *bp; 3556 int hasdeps; 3557 int flushed; 3558 int error; 3559 bool unlock; 3560 3561 flushed = 0; 3562 bq = &bd->bd_dirtyq; 3563 bp = NULL; 3564 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 3565 sentinel->b_qindex = QUEUE_SENTINEL; 3566 BQ_LOCK(bq); 3567 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist); 3568 BQ_UNLOCK(bq); 3569 while (flushed != target) { 3570 maybe_yield(); 3571 BQ_LOCK(bq); 3572 bp = TAILQ_NEXT(sentinel, b_freelist); 3573 if (bp != NULL) { 3574 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); 3575 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel, 3576 b_freelist); 3577 } else { 3578 BQ_UNLOCK(bq); 3579 break; 3580 } 3581 /* 3582 * Skip sentinels inserted by other invocations of the 3583 * flushbufqueues(), taking care to not reorder them. 3584 * 3585 * Only flush the buffers that belong to the 3586 * vnode locked by the curthread. 3587 */ 3588 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && 3589 bp->b_vp != lvp)) { 3590 BQ_UNLOCK(bq); 3591 continue; 3592 } 3593 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 3594 BQ_UNLOCK(bq); 3595 if (error != 0) 3596 continue; 3597 3598 /* 3599 * BKGRDINPROG can only be set with the buf and bufobj 3600 * locks both held. We tolerate a race to clear it here. 3601 */ 3602 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 3603 (bp->b_flags & B_DELWRI) == 0) { 3604 BUF_UNLOCK(bp); 3605 continue; 3606 } 3607 if (bp->b_flags & B_INVAL) { 3608 bremfreef(bp); 3609 brelse(bp); 3610 flushed++; 3611 continue; 3612 } 3613 3614 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 3615 if (flushdeps == 0) { 3616 BUF_UNLOCK(bp); 3617 continue; 3618 } 3619 hasdeps = 1; 3620 } else 3621 hasdeps = 0; 3622 /* 3623 * We must hold the lock on a vnode before writing 3624 * one of its buffers. Otherwise we may confuse, or 3625 * in the case of a snapshot vnode, deadlock the 3626 * system. 3627 * 3628 * The lock order here is the reverse of the normal 3629 * of vnode followed by buf lock. This is ok because 3630 * the NOWAIT will prevent deadlock. 3631 */ 3632 vp = bp->b_vp; 3633 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 3634 BUF_UNLOCK(bp); 3635 continue; 3636 } 3637 if (lvp == NULL) { 3638 unlock = true; 3639 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 3640 } else { 3641 ASSERT_VOP_LOCKED(vp, "getbuf"); 3642 unlock = false; 3643 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 3644 vn_lock(vp, LK_TRYUPGRADE); 3645 } 3646 if (error == 0) { 3647 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 3648 bp, bp->b_vp, bp->b_flags); 3649 if (curproc == bufdaemonproc) { 3650 vfs_bio_awrite(bp); 3651 } else { 3652 bremfree(bp); 3653 bwrite(bp); 3654 counter_u64_add(notbufdflushes, 1); 3655 } 3656 vn_finished_write(mp); 3657 if (unlock) 3658 VOP_UNLOCK(vp); 3659 flushwithdeps += hasdeps; 3660 flushed++; 3661 3662 /* 3663 * Sleeping on runningbufspace while holding 3664 * vnode lock leads to deadlock. 3665 */ 3666 if (curproc == bufdaemonproc && 3667 runningbufspace > hirunningspace) 3668 waitrunningbufspace(); 3669 continue; 3670 } 3671 vn_finished_write(mp); 3672 BUF_UNLOCK(bp); 3673 } 3674 BQ_LOCK(bq); 3675 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); 3676 BQ_UNLOCK(bq); 3677 free(sentinel, M_TEMP); 3678 return (flushed); 3679 } 3680 3681 /* 3682 * Check to see if a block is currently memory resident. 3683 */ 3684 struct buf * 3685 incore(struct bufobj *bo, daddr_t blkno) 3686 { 3687 return (gbincore_unlocked(bo, blkno)); 3688 } 3689 3690 /* 3691 * Returns true if no I/O is needed to access the 3692 * associated VM object. This is like incore except 3693 * it also hunts around in the VM system for the data. 3694 */ 3695 bool 3696 inmem(struct vnode * vp, daddr_t blkno) 3697 { 3698 vm_object_t obj; 3699 vm_offset_t toff, tinc, size; 3700 vm_page_t m, n; 3701 vm_ooffset_t off; 3702 int valid; 3703 3704 ASSERT_VOP_LOCKED(vp, "inmem"); 3705 3706 if (incore(&vp->v_bufobj, blkno)) 3707 return (true); 3708 if (vp->v_mount == NULL) 3709 return (false); 3710 obj = vp->v_object; 3711 if (obj == NULL) 3712 return (false); 3713 3714 size = PAGE_SIZE; 3715 if (size > vp->v_mount->mnt_stat.f_iosize) 3716 size = vp->v_mount->mnt_stat.f_iosize; 3717 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 3718 3719 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 3720 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff)); 3721 recheck: 3722 if (m == NULL) 3723 return (false); 3724 3725 tinc = size; 3726 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 3727 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 3728 /* 3729 * Consider page validity only if page mapping didn't change 3730 * during the check. 3731 */ 3732 valid = vm_page_is_valid(m, 3733 (vm_offset_t)((toff + off) & PAGE_MASK), tinc); 3734 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff)); 3735 if (m != n) { 3736 m = n; 3737 goto recheck; 3738 } 3739 if (!valid) 3740 return (false); 3741 } 3742 return (true); 3743 } 3744 3745 /* 3746 * Set the dirty range for a buffer based on the status of the dirty 3747 * bits in the pages comprising the buffer. The range is limited 3748 * to the size of the buffer. 3749 * 3750 * Tell the VM system that the pages associated with this buffer 3751 * are clean. This is used for delayed writes where the data is 3752 * going to go to disk eventually without additional VM intevention. 3753 * 3754 * Note that while we only really need to clean through to b_bcount, we 3755 * just go ahead and clean through to b_bufsize. 3756 */ 3757 static void 3758 vfs_clean_pages_dirty_buf(struct buf *bp) 3759 { 3760 vm_ooffset_t foff, noff, eoff; 3761 vm_page_t m; 3762 int i; 3763 3764 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 3765 return; 3766 3767 foff = bp->b_offset; 3768 KASSERT(bp->b_offset != NOOFFSET, 3769 ("vfs_clean_pages_dirty_buf: no buffer offset")); 3770 3771 vfs_busy_pages_acquire(bp); 3772 vfs_setdirty_range(bp); 3773 for (i = 0; i < bp->b_npages; i++) { 3774 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3775 eoff = noff; 3776 if (eoff > bp->b_offset + bp->b_bufsize) 3777 eoff = bp->b_offset + bp->b_bufsize; 3778 m = bp->b_pages[i]; 3779 vfs_page_set_validclean(bp, foff, m); 3780 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3781 foff = noff; 3782 } 3783 vfs_busy_pages_release(bp); 3784 } 3785 3786 static void 3787 vfs_setdirty_range(struct buf *bp) 3788 { 3789 vm_offset_t boffset; 3790 vm_offset_t eoffset; 3791 int i; 3792 3793 /* 3794 * test the pages to see if they have been modified directly 3795 * by users through the VM system. 3796 */ 3797 for (i = 0; i < bp->b_npages; i++) 3798 vm_page_test_dirty(bp->b_pages[i]); 3799 3800 /* 3801 * Calculate the encompassing dirty range, boffset and eoffset, 3802 * (eoffset - boffset) bytes. 3803 */ 3804 3805 for (i = 0; i < bp->b_npages; i++) { 3806 if (bp->b_pages[i]->dirty) 3807 break; 3808 } 3809 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3810 3811 for (i = bp->b_npages - 1; i >= 0; --i) { 3812 if (bp->b_pages[i]->dirty) { 3813 break; 3814 } 3815 } 3816 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3817 3818 /* 3819 * Fit it to the buffer. 3820 */ 3821 3822 if (eoffset > bp->b_bcount) 3823 eoffset = bp->b_bcount; 3824 3825 /* 3826 * If we have a good dirty range, merge with the existing 3827 * dirty range. 3828 */ 3829 3830 if (boffset < eoffset) { 3831 if (bp->b_dirtyoff > boffset) 3832 bp->b_dirtyoff = boffset; 3833 if (bp->b_dirtyend < eoffset) 3834 bp->b_dirtyend = eoffset; 3835 } 3836 } 3837 3838 /* 3839 * Allocate the KVA mapping for an existing buffer. 3840 * If an unmapped buffer is provided but a mapped buffer is requested, take 3841 * also care to properly setup mappings between pages and KVA. 3842 */ 3843 static void 3844 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 3845 { 3846 int bsize, maxsize, need_mapping, need_kva; 3847 off_t offset; 3848 3849 need_mapping = bp->b_data == unmapped_buf && 3850 (gbflags & GB_UNMAPPED) == 0; 3851 need_kva = bp->b_kvabase == unmapped_buf && 3852 bp->b_data == unmapped_buf && 3853 (gbflags & GB_KVAALLOC) != 0; 3854 if (!need_mapping && !need_kva) 3855 return; 3856 3857 BUF_CHECK_UNMAPPED(bp); 3858 3859 if (need_mapping && bp->b_kvabase != unmapped_buf) { 3860 /* 3861 * Buffer is not mapped, but the KVA was already 3862 * reserved at the time of the instantiation. Use the 3863 * allocated space. 3864 */ 3865 goto has_addr; 3866 } 3867 3868 /* 3869 * Calculate the amount of the address space we would reserve 3870 * if the buffer was mapped. 3871 */ 3872 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 3873 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3874 offset = blkno * bsize; 3875 maxsize = size + (offset & PAGE_MASK); 3876 maxsize = imax(maxsize, bsize); 3877 3878 while (bufkva_alloc(bp, maxsize, gbflags) != 0) { 3879 if ((gbflags & GB_NOWAIT_BD) != 0) { 3880 /* 3881 * XXXKIB: defragmentation cannot 3882 * succeed, not sure what else to do. 3883 */ 3884 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp); 3885 } 3886 counter_u64_add(mappingrestarts, 1); 3887 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0); 3888 } 3889 has_addr: 3890 if (need_mapping) { 3891 /* b_offset is handled by bpmap_qenter. */ 3892 bp->b_data = bp->b_kvabase; 3893 BUF_CHECK_MAPPED(bp); 3894 bpmap_qenter(bp); 3895 } 3896 } 3897 3898 struct buf * 3899 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3900 int flags) 3901 { 3902 struct buf *bp; 3903 int error; 3904 3905 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp); 3906 if (error != 0) 3907 return (NULL); 3908 return (bp); 3909 } 3910 3911 /* 3912 * getblkx: 3913 * 3914 * Get a block given a specified block and offset into a file/device. 3915 * The buffers B_DONE bit will be cleared on return, making it almost 3916 * ready for an I/O initiation. B_INVAL may or may not be set on 3917 * return. The caller should clear B_INVAL prior to initiating a 3918 * READ. 3919 * 3920 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3921 * an existing buffer. 3922 * 3923 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3924 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3925 * and then cleared based on the backing VM. If the previous buffer is 3926 * non-0-sized but invalid, B_CACHE will be cleared. 3927 * 3928 * If getblk() must create a new buffer, the new buffer is returned with 3929 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3930 * case it is returned with B_INVAL clear and B_CACHE set based on the 3931 * backing VM. 3932 * 3933 * getblk() also forces a bwrite() for any B_DELWRI buffer whose 3934 * B_CACHE bit is clear. 3935 * 3936 * What this means, basically, is that the caller should use B_CACHE to 3937 * determine whether the buffer is fully valid or not and should clear 3938 * B_INVAL prior to issuing a read. If the caller intends to validate 3939 * the buffer by loading its data area with something, the caller needs 3940 * to clear B_INVAL. If the caller does this without issuing an I/O, 3941 * the caller should set B_CACHE ( as an optimization ), else the caller 3942 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3943 * a write attempt or if it was a successful read. If the caller 3944 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3945 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3946 * 3947 * The blkno parameter is the logical block being requested. Normally 3948 * the mapping of logical block number to disk block address is done 3949 * by calling VOP_BMAP(). However, if the mapping is already known, the 3950 * disk block address can be passed using the dblkno parameter. If the 3951 * disk block address is not known, then the same value should be passed 3952 * for blkno and dblkno. 3953 */ 3954 int 3955 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag, 3956 int slptimeo, int flags, struct buf **bpp) 3957 { 3958 struct buf *bp; 3959 struct bufobj *bo; 3960 daddr_t d_blkno; 3961 int bsize, error, maxsize, vmio; 3962 off_t offset; 3963 3964 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3965 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3966 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3967 if (vp->v_type != VCHR) 3968 ASSERT_VOP_LOCKED(vp, "getblk"); 3969 if (size > maxbcachebuf) 3970 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size, 3971 maxbcachebuf); 3972 if (!unmapped_buf_allowed) 3973 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3974 3975 bo = &vp->v_bufobj; 3976 d_blkno = dblkno; 3977 3978 /* Attempt lockless lookup first. */ 3979 bp = gbincore_unlocked(bo, blkno); 3980 if (bp == NULL) { 3981 /* 3982 * With GB_NOCREAT we must be sure about not finding the buffer 3983 * as it may have been reassigned during unlocked lookup. 3984 */ 3985 if ((flags & GB_NOCREAT) != 0) 3986 goto loop; 3987 goto newbuf_unlocked; 3988 } 3989 3990 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0, 3991 0); 3992 if (error != 0) 3993 goto loop; 3994 3995 /* Verify buf identify has not changed since lookup. */ 3996 if (bp->b_bufobj == bo && bp->b_lblkno == blkno) 3997 goto foundbuf_fastpath; 3998 3999 /* It changed, fallback to locked lookup. */ 4000 BUF_UNLOCK_RAW(bp); 4001 4002 loop: 4003 BO_RLOCK(bo); 4004 bp = gbincore(bo, blkno); 4005 if (bp != NULL) { 4006 int lockflags; 4007 4008 /* 4009 * Buffer is in-core. If the buffer is not busy nor managed, 4010 * it must be on a queue. 4011 */ 4012 lockflags = LK_EXCLUSIVE | LK_INTERLOCK | 4013 ((flags & GB_LOCK_NOWAIT) != 0 ? LK_NOWAIT : LK_SLEEPFAIL); 4014 #ifdef WITNESS 4015 lockflags |= (flags & GB_NOWITNESS) != 0 ? LK_NOWITNESS : 0; 4016 #endif 4017 4018 error = BUF_TIMELOCK(bp, lockflags, 4019 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 4020 4021 /* 4022 * If we slept and got the lock we have to restart in case 4023 * the buffer changed identities. 4024 */ 4025 if (error == ENOLCK) 4026 goto loop; 4027 /* We timed out or were interrupted. */ 4028 else if (error != 0) 4029 return (error); 4030 4031 foundbuf_fastpath: 4032 /* If recursed, assume caller knows the rules. */ 4033 if (BUF_LOCKRECURSED(bp)) 4034 goto end; 4035 4036 /* 4037 * The buffer is locked. B_CACHE is cleared if the buffer is 4038 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 4039 * and for a VMIO buffer B_CACHE is adjusted according to the 4040 * backing VM cache. 4041 */ 4042 if (bp->b_flags & B_INVAL) 4043 bp->b_flags &= ~B_CACHE; 4044 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 4045 bp->b_flags |= B_CACHE; 4046 if (bp->b_flags & B_MANAGED) 4047 MPASS(bp->b_qindex == QUEUE_NONE); 4048 else 4049 bremfree(bp); 4050 4051 /* 4052 * check for size inconsistencies for non-VMIO case. 4053 */ 4054 if (bp->b_bcount != size) { 4055 if ((bp->b_flags & B_VMIO) == 0 || 4056 (size > bp->b_kvasize)) { 4057 if (bp->b_flags & B_DELWRI) { 4058 bp->b_flags |= B_NOCACHE; 4059 bwrite(bp); 4060 } else { 4061 if (LIST_EMPTY(&bp->b_dep)) { 4062 bp->b_flags |= B_RELBUF; 4063 brelse(bp); 4064 } else { 4065 bp->b_flags |= B_NOCACHE; 4066 bwrite(bp); 4067 } 4068 } 4069 goto loop; 4070 } 4071 } 4072 4073 /* 4074 * Handle the case of unmapped buffer which should 4075 * become mapped, or the buffer for which KVA 4076 * reservation is requested. 4077 */ 4078 bp_unmapped_get_kva(bp, blkno, size, flags); 4079 4080 /* 4081 * If the size is inconsistent in the VMIO case, we can resize 4082 * the buffer. This might lead to B_CACHE getting set or 4083 * cleared. If the size has not changed, B_CACHE remains 4084 * unchanged from its previous state. 4085 */ 4086 allocbuf(bp, size); 4087 4088 KASSERT(bp->b_offset != NOOFFSET, 4089 ("getblk: no buffer offset")); 4090 4091 /* 4092 * A buffer with B_DELWRI set and B_CACHE clear must 4093 * be committed before we can return the buffer in 4094 * order to prevent the caller from issuing a read 4095 * ( due to B_CACHE not being set ) and overwriting 4096 * it. 4097 * 4098 * Most callers, including NFS and FFS, need this to 4099 * operate properly either because they assume they 4100 * can issue a read if B_CACHE is not set, or because 4101 * ( for example ) an uncached B_DELWRI might loop due 4102 * to softupdates re-dirtying the buffer. In the latter 4103 * case, B_CACHE is set after the first write completes, 4104 * preventing further loops. 4105 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 4106 * above while extending the buffer, we cannot allow the 4107 * buffer to remain with B_CACHE set after the write 4108 * completes or it will represent a corrupt state. To 4109 * deal with this we set B_NOCACHE to scrap the buffer 4110 * after the write. 4111 * 4112 * We might be able to do something fancy, like setting 4113 * B_CACHE in bwrite() except if B_DELWRI is already set, 4114 * so the below call doesn't set B_CACHE, but that gets real 4115 * confusing. This is much easier. 4116 */ 4117 4118 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 4119 bp->b_flags |= B_NOCACHE; 4120 bwrite(bp); 4121 goto loop; 4122 } 4123 bp->b_flags &= ~B_DONE; 4124 } else { 4125 /* 4126 * Buffer is not in-core, create new buffer. The buffer 4127 * returned by getnewbuf() is locked. Note that the returned 4128 * buffer is also considered valid (not marked B_INVAL). 4129 */ 4130 BO_RUNLOCK(bo); 4131 newbuf_unlocked: 4132 /* 4133 * If the user does not want us to create the buffer, bail out 4134 * here. 4135 */ 4136 if (flags & GB_NOCREAT) 4137 return (EEXIST); 4138 4139 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize; 4140 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 4141 offset = blkno * bsize; 4142 vmio = vp->v_object != NULL; 4143 if (vmio) { 4144 maxsize = size + (offset & PAGE_MASK); 4145 } else { 4146 maxsize = size; 4147 /* Do not allow non-VMIO notmapped buffers. */ 4148 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 4149 } 4150 maxsize = imax(maxsize, bsize); 4151 if ((flags & GB_NOSPARSE) != 0 && vmio && 4152 !vn_isdisk(vp)) { 4153 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0); 4154 KASSERT(error != EOPNOTSUPP, 4155 ("GB_NOSPARSE from fs not supporting bmap, vp %p", 4156 vp)); 4157 if (error != 0) 4158 return (error); 4159 if (d_blkno == -1) 4160 return (EJUSTRETURN); 4161 } 4162 4163 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags); 4164 if (bp == NULL) { 4165 if (slpflag || slptimeo) 4166 return (ETIMEDOUT); 4167 /* 4168 * XXX This is here until the sleep path is diagnosed 4169 * enough to work under very low memory conditions. 4170 * 4171 * There's an issue on low memory, 4BSD+non-preempt 4172 * systems (eg MIPS routers with 32MB RAM) where buffer 4173 * exhaustion occurs without sleeping for buffer 4174 * reclaimation. This just sticks in a loop and 4175 * constantly attempts to allocate a buffer, which 4176 * hits exhaustion and tries to wakeup bufdaemon. 4177 * This never happens because we never yield. 4178 * 4179 * The real solution is to identify and fix these cases 4180 * so we aren't effectively busy-waiting in a loop 4181 * until the reclaimation path has cycles to run. 4182 */ 4183 kern_yield(PRI_USER); 4184 goto loop; 4185 } 4186 4187 /* 4188 * This code is used to make sure that a buffer is not 4189 * created while the getnewbuf routine is blocked. 4190 * This can be a problem whether the vnode is locked or not. 4191 * If the buffer is created out from under us, we have to 4192 * throw away the one we just created. 4193 * 4194 * Note: this must occur before we associate the buffer 4195 * with the vp especially considering limitations in 4196 * the splay tree implementation when dealing with duplicate 4197 * lblkno's. 4198 */ 4199 BO_LOCK(bo); 4200 if (gbincore(bo, blkno)) { 4201 BO_UNLOCK(bo); 4202 bp->b_flags |= B_INVAL; 4203 bufspace_release(bufdomain(bp), maxsize); 4204 brelse(bp); 4205 goto loop; 4206 } 4207 4208 /* 4209 * Insert the buffer into the hash, so that it can 4210 * be found by incore. 4211 */ 4212 bp->b_lblkno = blkno; 4213 bp->b_blkno = d_blkno; 4214 bp->b_offset = offset; 4215 bgetvp(vp, bp); 4216 BO_UNLOCK(bo); 4217 4218 /* 4219 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 4220 * buffer size starts out as 0, B_CACHE will be set by 4221 * allocbuf() for the VMIO case prior to it testing the 4222 * backing store for validity. 4223 */ 4224 4225 if (vmio) { 4226 bp->b_flags |= B_VMIO; 4227 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 4228 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 4229 bp, vp->v_object, bp->b_bufobj->bo_object)); 4230 } else { 4231 bp->b_flags &= ~B_VMIO; 4232 KASSERT(bp->b_bufobj->bo_object == NULL, 4233 ("ARGH! has b_bufobj->bo_object %p %p\n", 4234 bp, bp->b_bufobj->bo_object)); 4235 BUF_CHECK_MAPPED(bp); 4236 } 4237 4238 allocbuf(bp, size); 4239 bufspace_release(bufdomain(bp), maxsize); 4240 bp->b_flags &= ~B_DONE; 4241 } 4242 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 4243 end: 4244 buf_track(bp, __func__); 4245 KASSERT(bp->b_bufobj == bo, 4246 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 4247 *bpp = bp; 4248 return (0); 4249 } 4250 4251 /* 4252 * Get an empty, disassociated buffer of given size. The buffer is initially 4253 * set to B_INVAL. 4254 */ 4255 struct buf * 4256 geteblk(int size, int flags) 4257 { 4258 struct buf *bp; 4259 int maxsize; 4260 4261 maxsize = (size + BKVAMASK) & ~BKVAMASK; 4262 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) { 4263 if ((flags & GB_NOWAIT_BD) && 4264 (curthread->td_pflags & TDP_BUFNEED) != 0) 4265 return (NULL); 4266 } 4267 allocbuf(bp, size); 4268 bufspace_release(bufdomain(bp), maxsize); 4269 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 4270 return (bp); 4271 } 4272 4273 /* 4274 * Truncate the backing store for a non-vmio buffer. 4275 */ 4276 static void 4277 vfs_nonvmio_truncate(struct buf *bp, int newbsize) 4278 { 4279 4280 if (bp->b_flags & B_MALLOC) { 4281 /* 4282 * malloced buffers are not shrunk 4283 */ 4284 if (newbsize == 0) { 4285 bufmallocadjust(bp, 0); 4286 free(bp->b_data, M_BIOBUF); 4287 bp->b_data = bp->b_kvabase; 4288 bp->b_flags &= ~B_MALLOC; 4289 } 4290 return; 4291 } 4292 vm_hold_free_pages(bp, newbsize); 4293 bufspace_adjust(bp, newbsize); 4294 } 4295 4296 /* 4297 * Extend the backing for a non-VMIO buffer. 4298 */ 4299 static void 4300 vfs_nonvmio_extend(struct buf *bp, int newbsize) 4301 { 4302 caddr_t origbuf; 4303 int origbufsize; 4304 4305 /* 4306 * We only use malloced memory on the first allocation. 4307 * and revert to page-allocated memory when the buffer 4308 * grows. 4309 * 4310 * There is a potential smp race here that could lead 4311 * to bufmallocspace slightly passing the max. It 4312 * is probably extremely rare and not worth worrying 4313 * over. 4314 */ 4315 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 && 4316 bufmallocspace < maxbufmallocspace) { 4317 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK); 4318 bp->b_flags |= B_MALLOC; 4319 bufmallocadjust(bp, newbsize); 4320 return; 4321 } 4322 4323 /* 4324 * If the buffer is growing on its other-than-first 4325 * allocation then we revert to the page-allocation 4326 * scheme. 4327 */ 4328 origbuf = NULL; 4329 origbufsize = 0; 4330 if (bp->b_flags & B_MALLOC) { 4331 origbuf = bp->b_data; 4332 origbufsize = bp->b_bufsize; 4333 bp->b_data = bp->b_kvabase; 4334 bufmallocadjust(bp, 0); 4335 bp->b_flags &= ~B_MALLOC; 4336 newbsize = round_page(newbsize); 4337 } 4338 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize, 4339 (vm_offset_t) bp->b_data + newbsize); 4340 if (origbuf != NULL) { 4341 bcopy(origbuf, bp->b_data, origbufsize); 4342 free(origbuf, M_BIOBUF); 4343 } 4344 bufspace_adjust(bp, newbsize); 4345 } 4346 4347 /* 4348 * This code constitutes the buffer memory from either anonymous system 4349 * memory (in the case of non-VMIO operations) or from an associated 4350 * VM object (in the case of VMIO operations). This code is able to 4351 * resize a buffer up or down. 4352 * 4353 * Note that this code is tricky, and has many complications to resolve 4354 * deadlock or inconsistent data situations. Tread lightly!!! 4355 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 4356 * the caller. Calling this code willy nilly can result in the loss of data. 4357 * 4358 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 4359 * B_CACHE for the non-VMIO case. 4360 */ 4361 int 4362 allocbuf(struct buf *bp, int size) 4363 { 4364 int newbsize; 4365 4366 if (bp->b_bcount == size) 4367 return (1); 4368 4369 KASSERT(bp->b_kvasize == 0 || bp->b_kvasize >= size, 4370 ("allocbuf: buffer too small %p %#x %#x", 4371 bp, bp->b_kvasize, size)); 4372 4373 newbsize = roundup2(size, DEV_BSIZE); 4374 if ((bp->b_flags & B_VMIO) == 0) { 4375 if ((bp->b_flags & B_MALLOC) == 0) 4376 newbsize = round_page(newbsize); 4377 /* 4378 * Just get anonymous memory from the kernel. Don't 4379 * mess with B_CACHE. 4380 */ 4381 if (newbsize < bp->b_bufsize) 4382 vfs_nonvmio_truncate(bp, newbsize); 4383 else if (newbsize > bp->b_bufsize) 4384 vfs_nonvmio_extend(bp, newbsize); 4385 } else { 4386 int desiredpages; 4387 4388 desiredpages = size == 0 ? 0 : 4389 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 4390 4391 KASSERT((bp->b_flags & B_MALLOC) == 0, 4392 ("allocbuf: VMIO buffer can't be malloced %p", bp)); 4393 4394 /* 4395 * Set B_CACHE initially if buffer is 0 length or will become 4396 * 0-length. 4397 */ 4398 if (size == 0 || bp->b_bufsize == 0) 4399 bp->b_flags |= B_CACHE; 4400 4401 if (newbsize < bp->b_bufsize) 4402 vfs_vmio_truncate(bp, desiredpages); 4403 /* XXX This looks as if it should be newbsize > b_bufsize */ 4404 else if (size > bp->b_bcount) 4405 vfs_vmio_extend(bp, desiredpages, size); 4406 bufspace_adjust(bp, newbsize); 4407 } 4408 bp->b_bcount = size; /* requested buffer size. */ 4409 return (1); 4410 } 4411 4412 extern int inflight_transient_maps; 4413 4414 static struct bio_queue nondump_bios; 4415 4416 void 4417 biodone(struct bio *bp) 4418 { 4419 struct mtx *mtxp; 4420 void (*done)(struct bio *); 4421 vm_offset_t start, end; 4422 4423 biotrack(bp, __func__); 4424 4425 /* 4426 * Avoid completing I/O when dumping after a panic since that may 4427 * result in a deadlock in the filesystem or pager code. Note that 4428 * this doesn't affect dumps that were started manually since we aim 4429 * to keep the system usable after it has been resumed. 4430 */ 4431 if (__predict_false(dumping && SCHEDULER_STOPPED())) { 4432 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue); 4433 return; 4434 } 4435 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 4436 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 4437 bp->bio_flags |= BIO_UNMAPPED; 4438 start = trunc_page((vm_offset_t)bp->bio_data); 4439 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 4440 bp->bio_data = unmapped_buf; 4441 pmap_qremove(start, atop(end - start)); 4442 vmem_free(transient_arena, start, end - start); 4443 atomic_add_int(&inflight_transient_maps, -1); 4444 } 4445 done = bp->bio_done; 4446 /* 4447 * The check for done == biodone is to allow biodone to be 4448 * used as a bio_done routine. 4449 */ 4450 if (done == NULL || done == biodone) { 4451 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4452 mtx_lock(mtxp); 4453 bp->bio_flags |= BIO_DONE; 4454 wakeup(bp); 4455 mtx_unlock(mtxp); 4456 } else 4457 done(bp); 4458 } 4459 4460 /* 4461 * Wait for a BIO to finish. 4462 */ 4463 int 4464 biowait(struct bio *bp, const char *wmesg) 4465 { 4466 struct mtx *mtxp; 4467 4468 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4469 mtx_lock(mtxp); 4470 while ((bp->bio_flags & BIO_DONE) == 0) 4471 msleep(bp, mtxp, PRIBIO, wmesg, 0); 4472 mtx_unlock(mtxp); 4473 if (bp->bio_error != 0) 4474 return (bp->bio_error); 4475 if (!(bp->bio_flags & BIO_ERROR)) 4476 return (0); 4477 return (EIO); 4478 } 4479 4480 void 4481 biofinish(struct bio *bp, struct devstat *stat, int error) 4482 { 4483 4484 if (error) { 4485 bp->bio_error = error; 4486 bp->bio_flags |= BIO_ERROR; 4487 } 4488 if (stat != NULL) 4489 devstat_end_transaction_bio(stat, bp); 4490 biodone(bp); 4491 } 4492 4493 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING) 4494 void 4495 biotrack_buf(struct bio *bp, const char *location) 4496 { 4497 4498 buf_track(bp->bio_track_bp, location); 4499 } 4500 #endif 4501 4502 /* 4503 * bufwait: 4504 * 4505 * Wait for buffer I/O completion, returning error status. The buffer 4506 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 4507 * error and cleared. 4508 */ 4509 int 4510 bufwait(struct buf *bp) 4511 { 4512 if (bp->b_iocmd == BIO_READ) 4513 bwait(bp, PRIBIO, "biord"); 4514 else 4515 bwait(bp, PRIBIO, "biowr"); 4516 if (bp->b_flags & B_EINTR) { 4517 bp->b_flags &= ~B_EINTR; 4518 return (EINTR); 4519 } 4520 if (bp->b_ioflags & BIO_ERROR) { 4521 return (bp->b_error ? bp->b_error : EIO); 4522 } else { 4523 return (0); 4524 } 4525 } 4526 4527 /* 4528 * bufdone: 4529 * 4530 * Finish I/O on a buffer, optionally calling a completion function. 4531 * This is usually called from an interrupt so process blocking is 4532 * not allowed. 4533 * 4534 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 4535 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 4536 * assuming B_INVAL is clear. 4537 * 4538 * For the VMIO case, we set B_CACHE if the op was a read and no 4539 * read error occurred, or if the op was a write. B_CACHE is never 4540 * set if the buffer is invalid or otherwise uncacheable. 4541 * 4542 * bufdone does not mess with B_INVAL, allowing the I/O routine or the 4543 * initiator to leave B_INVAL set to brelse the buffer out of existence 4544 * in the biodone routine. 4545 */ 4546 void 4547 bufdone(struct buf *bp) 4548 { 4549 struct bufobj *dropobj; 4550 void (*biodone)(struct buf *); 4551 4552 buf_track(bp, __func__); 4553 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 4554 dropobj = NULL; 4555 4556 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 4557 4558 runningbufwakeup(bp); 4559 if (bp->b_iocmd == BIO_WRITE) 4560 dropobj = bp->b_bufobj; 4561 /* call optional completion function if requested */ 4562 if (bp->b_iodone != NULL) { 4563 biodone = bp->b_iodone; 4564 bp->b_iodone = NULL; 4565 (*biodone) (bp); 4566 if (dropobj) 4567 bufobj_wdrop(dropobj); 4568 return; 4569 } 4570 if (bp->b_flags & B_VMIO) { 4571 /* 4572 * Set B_CACHE if the op was a normal read and no error 4573 * occurred. B_CACHE is set for writes in the b*write() 4574 * routines. 4575 */ 4576 if (bp->b_iocmd == BIO_READ && 4577 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 4578 !(bp->b_ioflags & BIO_ERROR)) 4579 bp->b_flags |= B_CACHE; 4580 vfs_vmio_iodone(bp); 4581 } 4582 if (!LIST_EMPTY(&bp->b_dep)) 4583 buf_complete(bp); 4584 if ((bp->b_flags & B_CKHASH) != 0) { 4585 KASSERT(bp->b_iocmd == BIO_READ, 4586 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd)); 4587 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp)); 4588 (*bp->b_ckhashcalc)(bp); 4589 } 4590 /* 4591 * For asynchronous completions, release the buffer now. The brelse 4592 * will do a wakeup there if necessary - so no need to do a wakeup 4593 * here in the async case. The sync case always needs to do a wakeup. 4594 */ 4595 if (bp->b_flags & B_ASYNC) { 4596 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || 4597 (bp->b_ioflags & BIO_ERROR)) 4598 brelse(bp); 4599 else 4600 bqrelse(bp); 4601 } else 4602 bdone(bp); 4603 if (dropobj) 4604 bufobj_wdrop(dropobj); 4605 } 4606 4607 /* 4608 * This routine is called in lieu of iodone in the case of 4609 * incomplete I/O. This keeps the busy status for pages 4610 * consistent. 4611 */ 4612 void 4613 vfs_unbusy_pages(struct buf *bp) 4614 { 4615 int i; 4616 vm_object_t obj; 4617 vm_page_t m; 4618 4619 runningbufwakeup(bp); 4620 if (!(bp->b_flags & B_VMIO)) 4621 return; 4622 4623 obj = bp->b_bufobj->bo_object; 4624 for (i = 0; i < bp->b_npages; i++) { 4625 m = bp->b_pages[i]; 4626 if (m == bogus_page) { 4627 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i); 4628 if (!m) 4629 panic("vfs_unbusy_pages: page missing\n"); 4630 bp->b_pages[i] = m; 4631 if (buf_mapped(bp)) { 4632 BUF_CHECK_MAPPED(bp); 4633 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4634 bp->b_pages, bp->b_npages); 4635 } else 4636 BUF_CHECK_UNMAPPED(bp); 4637 } 4638 vm_page_sunbusy(m); 4639 } 4640 vm_object_pip_wakeupn(obj, bp->b_npages); 4641 } 4642 4643 /* 4644 * vfs_page_set_valid: 4645 * 4646 * Set the valid bits in a page based on the supplied offset. The 4647 * range is restricted to the buffer's size. 4648 * 4649 * This routine is typically called after a read completes. 4650 */ 4651 static void 4652 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4653 { 4654 vm_ooffset_t eoff; 4655 4656 /* 4657 * Compute the end offset, eoff, such that [off, eoff) does not span a 4658 * page boundary and eoff is not greater than the end of the buffer. 4659 * The end of the buffer, in this case, is our file EOF, not the 4660 * allocation size of the buffer. 4661 */ 4662 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 4663 if (eoff > bp->b_offset + bp->b_bcount) 4664 eoff = bp->b_offset + bp->b_bcount; 4665 4666 /* 4667 * Set valid range. This is typically the entire buffer and thus the 4668 * entire page. 4669 */ 4670 if (eoff > off) 4671 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 4672 } 4673 4674 /* 4675 * vfs_page_set_validclean: 4676 * 4677 * Set the valid bits and clear the dirty bits in a page based on the 4678 * supplied offset. The range is restricted to the buffer's size. 4679 */ 4680 static void 4681 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4682 { 4683 vm_ooffset_t soff, eoff; 4684 4685 /* 4686 * Start and end offsets in buffer. eoff - soff may not cross a 4687 * page boundary or cross the end of the buffer. The end of the 4688 * buffer, in this case, is our file EOF, not the allocation size 4689 * of the buffer. 4690 */ 4691 soff = off; 4692 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4693 if (eoff > bp->b_offset + bp->b_bcount) 4694 eoff = bp->b_offset + bp->b_bcount; 4695 4696 /* 4697 * Set valid range. This is typically the entire buffer and thus the 4698 * entire page. 4699 */ 4700 if (eoff > soff) { 4701 vm_page_set_validclean( 4702 m, 4703 (vm_offset_t) (soff & PAGE_MASK), 4704 (vm_offset_t) (eoff - soff) 4705 ); 4706 } 4707 } 4708 4709 /* 4710 * Acquire a shared busy on all pages in the buf. 4711 */ 4712 void 4713 vfs_busy_pages_acquire(struct buf *bp) 4714 { 4715 int i; 4716 4717 for (i = 0; i < bp->b_npages; i++) 4718 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY); 4719 } 4720 4721 void 4722 vfs_busy_pages_release(struct buf *bp) 4723 { 4724 int i; 4725 4726 for (i = 0; i < bp->b_npages; i++) 4727 vm_page_sunbusy(bp->b_pages[i]); 4728 } 4729 4730 /* 4731 * This routine is called before a device strategy routine. 4732 * It is used to tell the VM system that paging I/O is in 4733 * progress, and treat the pages associated with the buffer 4734 * almost as being exclusive busy. Also the object paging_in_progress 4735 * flag is handled to make sure that the object doesn't become 4736 * inconsistent. 4737 * 4738 * Since I/O has not been initiated yet, certain buffer flags 4739 * such as BIO_ERROR or B_INVAL may be in an inconsistent state 4740 * and should be ignored. 4741 */ 4742 void 4743 vfs_busy_pages(struct buf *bp, int clear_modify) 4744 { 4745 vm_object_t obj; 4746 vm_ooffset_t foff; 4747 vm_page_t m; 4748 int i; 4749 bool bogus; 4750 4751 if (!(bp->b_flags & B_VMIO)) 4752 return; 4753 4754 obj = bp->b_bufobj->bo_object; 4755 foff = bp->b_offset; 4756 KASSERT(bp->b_offset != NOOFFSET, 4757 ("vfs_busy_pages: no buffer offset")); 4758 if ((bp->b_flags & B_CLUSTER) == 0) { 4759 vm_object_pip_add(obj, bp->b_npages); 4760 vfs_busy_pages_acquire(bp); 4761 } 4762 if (bp->b_bufsize != 0) 4763 vfs_setdirty_range(bp); 4764 bogus = false; 4765 for (i = 0; i < bp->b_npages; i++) { 4766 m = bp->b_pages[i]; 4767 vm_page_assert_sbusied(m); 4768 4769 /* 4770 * When readying a buffer for a read ( i.e 4771 * clear_modify == 0 ), it is important to do 4772 * bogus_page replacement for valid pages in 4773 * partially instantiated buffers. Partially 4774 * instantiated buffers can, in turn, occur when 4775 * reconstituting a buffer from its VM backing store 4776 * base. We only have to do this if B_CACHE is 4777 * clear ( which causes the I/O to occur in the 4778 * first place ). The replacement prevents the read 4779 * I/O from overwriting potentially dirty VM-backed 4780 * pages. XXX bogus page replacement is, uh, bogus. 4781 * It may not work properly with small-block devices. 4782 * We need to find a better way. 4783 */ 4784 if (clear_modify) { 4785 pmap_remove_write(m); 4786 vfs_page_set_validclean(bp, foff, m); 4787 } else if (vm_page_all_valid(m) && 4788 (bp->b_flags & B_CACHE) == 0) { 4789 bp->b_pages[i] = bogus_page; 4790 bogus = true; 4791 } 4792 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4793 } 4794 if (bogus && buf_mapped(bp)) { 4795 BUF_CHECK_MAPPED(bp); 4796 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4797 bp->b_pages, bp->b_npages); 4798 } 4799 } 4800 4801 /* 4802 * vfs_bio_set_valid: 4803 * 4804 * Set the range within the buffer to valid. The range is 4805 * relative to the beginning of the buffer, b_offset. Note that 4806 * b_offset itself may be offset from the beginning of the first 4807 * page. 4808 */ 4809 void 4810 vfs_bio_set_valid(struct buf *bp, int base, int size) 4811 { 4812 int i, n; 4813 vm_page_t m; 4814 4815 if (!(bp->b_flags & B_VMIO)) 4816 return; 4817 4818 /* 4819 * Fixup base to be relative to beginning of first page. 4820 * Set initial n to be the maximum number of bytes in the 4821 * first page that can be validated. 4822 */ 4823 base += (bp->b_offset & PAGE_MASK); 4824 n = PAGE_SIZE - (base & PAGE_MASK); 4825 4826 /* 4827 * Busy may not be strictly necessary here because the pages are 4828 * unlikely to be fully valid and the vnode lock will synchronize 4829 * their access via getpages. It is grabbed for consistency with 4830 * other page validation. 4831 */ 4832 vfs_busy_pages_acquire(bp); 4833 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4834 m = bp->b_pages[i]; 4835 if (n > size) 4836 n = size; 4837 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4838 base += n; 4839 size -= n; 4840 n = PAGE_SIZE; 4841 } 4842 vfs_busy_pages_release(bp); 4843 } 4844 4845 /* 4846 * vfs_bio_clrbuf: 4847 * 4848 * If the specified buffer is a non-VMIO buffer, clear the entire 4849 * buffer. If the specified buffer is a VMIO buffer, clear and 4850 * validate only the previously invalid portions of the buffer. 4851 * This routine essentially fakes an I/O, so we need to clear 4852 * BIO_ERROR and B_INVAL. 4853 * 4854 * Note that while we only theoretically need to clear through b_bcount, 4855 * we go ahead and clear through b_bufsize. 4856 */ 4857 void 4858 vfs_bio_clrbuf(struct buf *bp) 4859 { 4860 int i, j, sa, ea, slide, zbits; 4861 vm_page_bits_t mask; 4862 4863 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4864 clrbuf(bp); 4865 return; 4866 } 4867 bp->b_flags &= ~B_INVAL; 4868 bp->b_ioflags &= ~BIO_ERROR; 4869 vfs_busy_pages_acquire(bp); 4870 sa = bp->b_offset & PAGE_MASK; 4871 slide = 0; 4872 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4873 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4874 ea = slide & PAGE_MASK; 4875 if (ea == 0) 4876 ea = PAGE_SIZE; 4877 if (bp->b_pages[i] == bogus_page) 4878 continue; 4879 j = sa / DEV_BSIZE; 4880 zbits = (sizeof(vm_page_bits_t) * NBBY) - 4881 (ea - sa) / DEV_BSIZE; 4882 mask = (VM_PAGE_BITS_ALL >> zbits) << j; 4883 if ((bp->b_pages[i]->valid & mask) == mask) 4884 continue; 4885 if ((bp->b_pages[i]->valid & mask) == 0) 4886 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4887 else { 4888 for (; sa < ea; sa += DEV_BSIZE, j++) { 4889 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4890 pmap_zero_page_area(bp->b_pages[i], 4891 sa, DEV_BSIZE); 4892 } 4893 } 4894 } 4895 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE, 4896 roundup2(ea - sa, DEV_BSIZE)); 4897 } 4898 vfs_busy_pages_release(bp); 4899 bp->b_resid = 0; 4900 } 4901 4902 void 4903 vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4904 { 4905 vm_page_t m; 4906 int i, n; 4907 4908 if (buf_mapped(bp)) { 4909 BUF_CHECK_MAPPED(bp); 4910 bzero(bp->b_data + base, size); 4911 } else { 4912 BUF_CHECK_UNMAPPED(bp); 4913 n = PAGE_SIZE - (base & PAGE_MASK); 4914 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4915 m = bp->b_pages[i]; 4916 if (n > size) 4917 n = size; 4918 pmap_zero_page_area(m, base & PAGE_MASK, n); 4919 base += n; 4920 size -= n; 4921 n = PAGE_SIZE; 4922 } 4923 } 4924 } 4925 4926 /* 4927 * Update buffer flags based on I/O request parameters, optionally releasing the 4928 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM, 4929 * where they may be placed on a page queue (VMIO) or freed immediately (direct 4930 * I/O). Otherwise the buffer is released to the cache. 4931 */ 4932 static void 4933 b_io_dismiss(struct buf *bp, int ioflag, bool release) 4934 { 4935 4936 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0, 4937 ("buf %p non-VMIO noreuse", bp)); 4938 4939 if ((ioflag & IO_DIRECT) != 0) 4940 bp->b_flags |= B_DIRECT; 4941 if ((ioflag & IO_EXT) != 0) 4942 bp->b_xflags |= BX_ALTDATA; 4943 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) { 4944 bp->b_flags |= B_RELBUF; 4945 if ((ioflag & IO_NOREUSE) != 0) 4946 bp->b_flags |= B_NOREUSE; 4947 if (release) 4948 brelse(bp); 4949 } else if (release) 4950 bqrelse(bp); 4951 } 4952 4953 void 4954 vfs_bio_brelse(struct buf *bp, int ioflag) 4955 { 4956 4957 b_io_dismiss(bp, ioflag, true); 4958 } 4959 4960 void 4961 vfs_bio_set_flags(struct buf *bp, int ioflag) 4962 { 4963 4964 b_io_dismiss(bp, ioflag, false); 4965 } 4966 4967 /* 4968 * vm_hold_load_pages and vm_hold_free_pages get pages into 4969 * a buffers address space. The pages are anonymous and are 4970 * not associated with a file object. 4971 */ 4972 static void 4973 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4974 { 4975 vm_offset_t pg; 4976 vm_page_t p; 4977 int index; 4978 4979 BUF_CHECK_MAPPED(bp); 4980 4981 to = round_page(to); 4982 from = round_page(from); 4983 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4984 MPASS((bp->b_flags & B_MAXPHYS) == 0); 4985 KASSERT(to - from <= maxbcachebuf, 4986 ("vm_hold_load_pages too large %p %#jx %#jx %u", 4987 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf)); 4988 4989 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4990 /* 4991 * note: must allocate system pages since blocking here 4992 * could interfere with paging I/O, no matter which 4993 * process we are. 4994 */ 4995 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 4996 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK); 4997 pmap_qenter(pg, &p, 1); 4998 bp->b_pages[index] = p; 4999 } 5000 bp->b_npages = index; 5001 } 5002 5003 /* Return pages associated with this buf to the vm system */ 5004 static void 5005 vm_hold_free_pages(struct buf *bp, int newbsize) 5006 { 5007 vm_offset_t from; 5008 vm_page_t p; 5009 int index, newnpages; 5010 5011 BUF_CHECK_MAPPED(bp); 5012 5013 from = round_page((vm_offset_t)bp->b_data + newbsize); 5014 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 5015 if (bp->b_npages > newnpages) 5016 pmap_qremove(from, bp->b_npages - newnpages); 5017 for (index = newnpages; index < bp->b_npages; index++) { 5018 p = bp->b_pages[index]; 5019 bp->b_pages[index] = NULL; 5020 vm_page_unwire_noq(p); 5021 vm_page_free(p); 5022 } 5023 bp->b_npages = newnpages; 5024 } 5025 5026 /* 5027 * Map an IO request into kernel virtual address space. 5028 * 5029 * All requests are (re)mapped into kernel VA space. 5030 * Notice that we use b_bufsize for the size of the buffer 5031 * to be mapped. b_bcount might be modified by the driver. 5032 * 5033 * Note that even if the caller determines that the address space should 5034 * be valid, a race or a smaller-file mapped into a larger space may 5035 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 5036 * check the return value. 5037 * 5038 * This function only works with pager buffers. 5039 */ 5040 int 5041 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf) 5042 { 5043 vm_prot_t prot; 5044 int pidx; 5045 5046 MPASS((bp->b_flags & B_MAXPHYS) != 0); 5047 prot = VM_PROT_READ; 5048 if (bp->b_iocmd == BIO_READ) 5049 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 5050 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 5051 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES); 5052 if (pidx < 0) 5053 return (-1); 5054 bp->b_bufsize = len; 5055 bp->b_npages = pidx; 5056 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK; 5057 if (mapbuf || !unmapped_buf_allowed) { 5058 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx); 5059 bp->b_data = bp->b_kvabase + bp->b_offset; 5060 } else 5061 bp->b_data = unmapped_buf; 5062 return (0); 5063 } 5064 5065 /* 5066 * Free the io map PTEs associated with this IO operation. 5067 * We also invalidate the TLB entries and restore the original b_addr. 5068 * 5069 * This function only works with pager buffers. 5070 */ 5071 void 5072 vunmapbuf(struct buf *bp) 5073 { 5074 int npages; 5075 5076 npages = bp->b_npages; 5077 if (buf_mapped(bp)) 5078 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 5079 vm_page_unhold_pages(bp->b_pages, npages); 5080 5081 bp->b_data = unmapped_buf; 5082 } 5083 5084 void 5085 bdone(struct buf *bp) 5086 { 5087 struct mtx *mtxp; 5088 5089 mtxp = mtx_pool_find(mtxpool_sleep, bp); 5090 mtx_lock(mtxp); 5091 bp->b_flags |= B_DONE; 5092 wakeup(bp); 5093 mtx_unlock(mtxp); 5094 } 5095 5096 void 5097 bwait(struct buf *bp, u_char pri, const char *wchan) 5098 { 5099 struct mtx *mtxp; 5100 5101 mtxp = mtx_pool_find(mtxpool_sleep, bp); 5102 mtx_lock(mtxp); 5103 while ((bp->b_flags & B_DONE) == 0) 5104 msleep(bp, mtxp, pri, wchan, 0); 5105 mtx_unlock(mtxp); 5106 } 5107 5108 int 5109 bufsync(struct bufobj *bo, int waitfor) 5110 { 5111 5112 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread)); 5113 } 5114 5115 void 5116 bufstrategy(struct bufobj *bo, struct buf *bp) 5117 { 5118 int i __unused; 5119 struct vnode *vp; 5120 5121 vp = bp->b_vp; 5122 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 5123 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 5124 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 5125 i = VOP_STRATEGY(vp, bp); 5126 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 5127 } 5128 5129 /* 5130 * Initialize a struct bufobj before use. Memory is assumed zero filled. 5131 */ 5132 void 5133 bufobj_init(struct bufobj *bo, void *private) 5134 { 5135 static volatile int bufobj_cleanq; 5136 5137 bo->bo_domain = 5138 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains; 5139 rw_init(BO_LOCKPTR(bo), "bufobj interlock"); 5140 bo->bo_private = private; 5141 TAILQ_INIT(&bo->bo_clean.bv_hd); 5142 TAILQ_INIT(&bo->bo_dirty.bv_hd); 5143 } 5144 5145 void 5146 bufobj_wrefl(struct bufobj *bo) 5147 { 5148 5149 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 5150 ASSERT_BO_WLOCKED(bo); 5151 bo->bo_numoutput++; 5152 } 5153 5154 void 5155 bufobj_wref(struct bufobj *bo) 5156 { 5157 5158 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 5159 BO_LOCK(bo); 5160 bo->bo_numoutput++; 5161 BO_UNLOCK(bo); 5162 } 5163 5164 void 5165 bufobj_wdrop(struct bufobj *bo) 5166 { 5167 5168 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 5169 BO_LOCK(bo); 5170 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 5171 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 5172 bo->bo_flag &= ~BO_WWAIT; 5173 wakeup(&bo->bo_numoutput); 5174 } 5175 BO_UNLOCK(bo); 5176 } 5177 5178 int 5179 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 5180 { 5181 int error; 5182 5183 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 5184 ASSERT_BO_WLOCKED(bo); 5185 error = 0; 5186 while (bo->bo_numoutput) { 5187 bo->bo_flag |= BO_WWAIT; 5188 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 5189 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 5190 if (error) 5191 break; 5192 } 5193 return (error); 5194 } 5195 5196 /* 5197 * Set bio_data or bio_ma for struct bio from the struct buf. 5198 */ 5199 void 5200 bdata2bio(struct buf *bp, struct bio *bip) 5201 { 5202 5203 if (!buf_mapped(bp)) { 5204 KASSERT(unmapped_buf_allowed, ("unmapped")); 5205 bip->bio_ma = bp->b_pages; 5206 bip->bio_ma_n = bp->b_npages; 5207 bip->bio_data = unmapped_buf; 5208 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 5209 bip->bio_flags |= BIO_UNMAPPED; 5210 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 5211 PAGE_SIZE == bp->b_npages, 5212 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 5213 (long long)bip->bio_length, bip->bio_ma_n)); 5214 } else { 5215 bip->bio_data = bp->b_data; 5216 bip->bio_ma = NULL; 5217 } 5218 } 5219 5220 static int buf_pager_relbuf; 5221 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN, 5222 &buf_pager_relbuf, 0, 5223 "Make buffer pager release buffers after reading"); 5224 5225 /* 5226 * The buffer pager. It uses buffer reads to validate pages. 5227 * 5228 * In contrast to the generic local pager from vm/vnode_pager.c, this 5229 * pager correctly and easily handles volumes where the underlying 5230 * device block size is greater than the machine page size. The 5231 * buffer cache transparently extends the requested page run to be 5232 * aligned at the block boundary, and does the necessary bogus page 5233 * replacements in the addends to avoid obliterating already valid 5234 * pages. 5235 * 5236 * The only non-trivial issue is that the exclusive busy state for 5237 * pages, which is assumed by the vm_pager_getpages() interface, is 5238 * incompatible with the VMIO buffer cache's desire to share-busy the 5239 * pages. This function performs a trivial downgrade of the pages' 5240 * state before reading buffers, and a less trivial upgrade from the 5241 * shared-busy to excl-busy state after the read. 5242 */ 5243 int 5244 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count, 5245 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno, 5246 vbg_get_blksize_t get_blksize) 5247 { 5248 vm_page_t m; 5249 vm_object_t object; 5250 struct buf *bp; 5251 struct mount *mp; 5252 daddr_t lbn, lbnp; 5253 vm_ooffset_t la, lb, poff, poffe; 5254 long bo_bs, bsize; 5255 int br_flags, error, i, pgsin, pgsin_a, pgsin_b; 5256 bool redo, lpart; 5257 5258 object = vp->v_object; 5259 mp = vp->v_mount; 5260 error = 0; 5261 la = IDX_TO_OFF(ma[count - 1]->pindex); 5262 if (la >= object->un_pager.vnp.vnp_size) 5263 return (VM_PAGER_BAD); 5264 5265 /* 5266 * Change the meaning of la from where the last requested page starts 5267 * to where it ends, because that's the end of the requested region 5268 * and the start of the potential read-ahead region. 5269 */ 5270 la += PAGE_SIZE; 5271 lpart = la > object->un_pager.vnp.vnp_size; 5272 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)), 5273 &bo_bs); 5274 if (error != 0) 5275 return (VM_PAGER_ERROR); 5276 5277 /* 5278 * Calculate read-ahead, behind and total pages. 5279 */ 5280 pgsin = count; 5281 lb = IDX_TO_OFF(ma[0]->pindex); 5282 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs)); 5283 pgsin += pgsin_b; 5284 if (rbehind != NULL) 5285 *rbehind = pgsin_b; 5286 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la); 5287 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size) 5288 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size, 5289 PAGE_SIZE) - la); 5290 pgsin += pgsin_a; 5291 if (rahead != NULL) 5292 *rahead = pgsin_a; 5293 VM_CNT_INC(v_vnodein); 5294 VM_CNT_ADD(v_vnodepgsin, pgsin); 5295 5296 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) 5297 != 0) ? GB_UNMAPPED : 0; 5298 again: 5299 for (i = 0; i < count; i++) { 5300 if (ma[i] != bogus_page) 5301 vm_page_busy_downgrade(ma[i]); 5302 } 5303 5304 lbnp = -1; 5305 for (i = 0; i < count; i++) { 5306 m = ma[i]; 5307 if (m == bogus_page) 5308 continue; 5309 5310 /* 5311 * Pages are shared busy and the object lock is not 5312 * owned, which together allow for the pages' 5313 * invalidation. The racy test for validity avoids 5314 * useless creation of the buffer for the most typical 5315 * case when invalidation is not used in redo or for 5316 * parallel read. The shared->excl upgrade loop at 5317 * the end of the function catches the race in a 5318 * reliable way (protected by the object lock). 5319 */ 5320 if (vm_page_all_valid(m)) 5321 continue; 5322 5323 poff = IDX_TO_OFF(m->pindex); 5324 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size); 5325 for (; poff < poffe; poff += bsize) { 5326 lbn = get_lblkno(vp, poff); 5327 if (lbn == lbnp) 5328 goto next_page; 5329 lbnp = lbn; 5330 5331 error = get_blksize(vp, lbn, &bsize); 5332 if (error == 0) 5333 error = bread_gb(vp, lbn, bsize, 5334 curthread->td_ucred, br_flags, &bp); 5335 if (error != 0) 5336 goto end_pages; 5337 if (bp->b_rcred == curthread->td_ucred) { 5338 crfree(bp->b_rcred); 5339 bp->b_rcred = NOCRED; 5340 } 5341 if (LIST_EMPTY(&bp->b_dep)) { 5342 /* 5343 * Invalidation clears m->valid, but 5344 * may leave B_CACHE flag if the 5345 * buffer existed at the invalidation 5346 * time. In this case, recycle the 5347 * buffer to do real read on next 5348 * bread() after redo. 5349 * 5350 * Otherwise B_RELBUF is not strictly 5351 * necessary, enable to reduce buf 5352 * cache pressure. 5353 */ 5354 if (buf_pager_relbuf || 5355 !vm_page_all_valid(m)) 5356 bp->b_flags |= B_RELBUF; 5357 5358 bp->b_flags &= ~B_NOCACHE; 5359 brelse(bp); 5360 } else { 5361 bqrelse(bp); 5362 } 5363 } 5364 KASSERT(1 /* racy, enable for debugging */ || 5365 vm_page_all_valid(m) || i == count - 1, 5366 ("buf %d %p invalid", i, m)); 5367 if (i == count - 1 && lpart) { 5368 if (!vm_page_none_valid(m) && 5369 !vm_page_all_valid(m)) 5370 vm_page_zero_invalid(m, TRUE); 5371 } 5372 next_page:; 5373 } 5374 end_pages: 5375 5376 redo = false; 5377 for (i = 0; i < count; i++) { 5378 if (ma[i] == bogus_page) 5379 continue; 5380 if (vm_page_busy_tryupgrade(ma[i]) == 0) { 5381 vm_page_sunbusy(ma[i]); 5382 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex, 5383 VM_ALLOC_NORMAL); 5384 } 5385 5386 /* 5387 * Since the pages were only sbusy while neither the 5388 * buffer nor the object lock was held by us, or 5389 * reallocated while vm_page_grab() slept for busy 5390 * relinguish, they could have been invalidated. 5391 * Recheck the valid bits and re-read as needed. 5392 * 5393 * Note that the last page is made fully valid in the 5394 * read loop, and partial validity for the page at 5395 * index count - 1 could mean that the page was 5396 * invalidated or removed, so we must restart for 5397 * safety as well. 5398 */ 5399 if (!vm_page_all_valid(ma[i])) 5400 redo = true; 5401 } 5402 if (redo && error == 0) 5403 goto again; 5404 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); 5405 } 5406 5407 #include "opt_ddb.h" 5408 #ifdef DDB 5409 #include <ddb/ddb.h> 5410 5411 /* DDB command to show buffer data */ 5412 DB_SHOW_COMMAND(buffer, db_show_buffer) 5413 { 5414 /* get args */ 5415 struct buf *bp = (struct buf *)addr; 5416 #ifdef FULL_BUF_TRACKING 5417 uint32_t i, j; 5418 #endif 5419 5420 if (!have_addr) { 5421 db_printf("usage: show buffer <addr>\n"); 5422 return; 5423 } 5424 5425 db_printf("buf at %p\n", bp); 5426 db_printf("b_flags = 0x%b, b_xflags=0x%b\n", 5427 (u_int)bp->b_flags, PRINT_BUF_FLAGS, 5428 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS); 5429 db_printf("b_vflags=0x%b b_ioflags0x%b\n", 5430 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS, 5431 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS); 5432 db_printf( 5433 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 5434 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, " 5435 "b_vp = %p, b_dep = %p\n", 5436 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 5437 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 5438 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first); 5439 db_printf("b_kvabase = %p, b_kvasize = %d\n", 5440 bp->b_kvabase, bp->b_kvasize); 5441 if (bp->b_npages) { 5442 int i; 5443 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 5444 for (i = 0; i < bp->b_npages; i++) { 5445 vm_page_t m; 5446 m = bp->b_pages[i]; 5447 if (m != NULL) 5448 db_printf("(%p, 0x%lx, 0x%lx)", m->object, 5449 (u_long)m->pindex, 5450 (u_long)VM_PAGE_TO_PHYS(m)); 5451 else 5452 db_printf("( ??? )"); 5453 if ((i + 1) < bp->b_npages) 5454 db_printf(","); 5455 } 5456 db_printf("\n"); 5457 } 5458 BUF_LOCKPRINTINFO(bp); 5459 #if defined(FULL_BUF_TRACKING) 5460 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt); 5461 5462 i = bp->b_io_tcnt % BUF_TRACKING_SIZE; 5463 for (j = 1; j <= BUF_TRACKING_SIZE; j++) { 5464 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL) 5465 continue; 5466 db_printf(" %2u: %s\n", j, 5467 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]); 5468 } 5469 #elif defined(BUF_TRACKING) 5470 db_printf("b_io_tracking: %s\n", bp->b_io_tracking); 5471 #endif 5472 db_printf(" "); 5473 } 5474 5475 DB_SHOW_COMMAND_FLAGS(bufqueues, bufqueues, DB_CMD_MEMSAFE) 5476 { 5477 struct bufdomain *bd; 5478 struct buf *bp; 5479 long total; 5480 int i, j, cnt; 5481 5482 db_printf("bqempty: %d\n", bqempty.bq_len); 5483 5484 for (i = 0; i < buf_domains; i++) { 5485 bd = &bdomain[i]; 5486 db_printf("Buf domain %d\n", i); 5487 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers); 5488 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers); 5489 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers); 5490 db_printf("\n"); 5491 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace); 5492 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace); 5493 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace); 5494 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace); 5495 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh); 5496 db_printf("\n"); 5497 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers); 5498 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers); 5499 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers); 5500 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh); 5501 db_printf("\n"); 5502 total = 0; 5503 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist) 5504 total += bp->b_bufsize; 5505 db_printf("\tcleanq count\t%d (%ld)\n", 5506 bd->bd_cleanq->bq_len, total); 5507 total = 0; 5508 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist) 5509 total += bp->b_bufsize; 5510 db_printf("\tdirtyq count\t%d (%ld)\n", 5511 bd->bd_dirtyq.bq_len, total); 5512 db_printf("\twakeup\t\t%d\n", bd->bd_wanted); 5513 db_printf("\tlim\t\t%d\n", bd->bd_lim); 5514 db_printf("\tCPU "); 5515 for (j = 0; j <= mp_maxid; j++) 5516 db_printf("%d, ", bd->bd_subq[j].bq_len); 5517 db_printf("\n"); 5518 cnt = 0; 5519 total = 0; 5520 for (j = 0; j < nbuf; j++) { 5521 bp = nbufp(j); 5522 if (bp->b_domain == i && BUF_ISLOCKED(bp)) { 5523 cnt++; 5524 total += bp->b_bufsize; 5525 } 5526 } 5527 db_printf("\tLocked buffers: %d space %ld\n", cnt, total); 5528 cnt = 0; 5529 total = 0; 5530 for (j = 0; j < nbuf; j++) { 5531 bp = nbufp(j); 5532 if (bp->b_domain == i) { 5533 cnt++; 5534 total += bp->b_bufsize; 5535 } 5536 } 5537 db_printf("\tTotal buffers: %d space %ld\n", cnt, total); 5538 } 5539 } 5540 5541 DB_SHOW_COMMAND_FLAGS(lockedbufs, lockedbufs, DB_CMD_MEMSAFE) 5542 { 5543 struct buf *bp; 5544 int i; 5545 5546 for (i = 0; i < nbuf; i++) { 5547 bp = nbufp(i); 5548 if (BUF_ISLOCKED(bp)) { 5549 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5550 db_printf("\n"); 5551 if (db_pager_quit) 5552 break; 5553 } 5554 } 5555 } 5556 5557 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 5558 { 5559 struct vnode *vp; 5560 struct buf *bp; 5561 5562 if (!have_addr) { 5563 db_printf("usage: show vnodebufs <addr>\n"); 5564 return; 5565 } 5566 vp = (struct vnode *)addr; 5567 db_printf("Clean buffers:\n"); 5568 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 5569 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5570 db_printf("\n"); 5571 } 5572 db_printf("Dirty buffers:\n"); 5573 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 5574 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5575 db_printf("\n"); 5576 } 5577 } 5578 5579 DB_COMMAND_FLAGS(countfreebufs, db_coundfreebufs, DB_CMD_MEMSAFE) 5580 { 5581 struct buf *bp; 5582 int i, used = 0, nfree = 0; 5583 5584 if (have_addr) { 5585 db_printf("usage: countfreebufs\n"); 5586 return; 5587 } 5588 5589 for (i = 0; i < nbuf; i++) { 5590 bp = nbufp(i); 5591 if (bp->b_qindex == QUEUE_EMPTY) 5592 nfree++; 5593 else 5594 used++; 5595 } 5596 5597 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 5598 nfree + used); 5599 db_printf("numfreebuffers is %d\n", numfreebuffers); 5600 } 5601 #endif /* DDB */ 5602