1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 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; 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 /* 1172 * Pager buffers are allocated for short periods, so scale the 1173 * number of reserved buffers based on the number of CPUs rather 1174 * than amount of memory. 1175 */ 1176 nswbuf = min(nbuf / 4, 32 * mp_ncpus); 1177 if (nswbuf < NSWBUF_MIN) 1178 nswbuf = NSWBUF_MIN; 1179 } 1180 1181 /* 1182 * Reserve space for the buffer cache buffers 1183 */ 1184 buf = (char *)v; 1185 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) * 1186 atop(maxbcachebuf)) * nbuf; 1187 1188 return (v); 1189 } 1190 1191 /* 1192 * Single global constant for BUF_WMESG, to avoid getting multiple 1193 * references. 1194 */ 1195 static const char buf_wmesg[] = "bufwait"; 1196 1197 /* Initialize the buffer subsystem. Called before use of any buffers. */ 1198 void 1199 bufinit(void) 1200 { 1201 struct buf *bp; 1202 int i; 1203 1204 KASSERT(maxbcachebuf >= MAXBSIZE, 1205 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf, 1206 MAXBSIZE)); 1207 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock"); 1208 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 1209 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 1210 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); 1211 1212 unmapped_buf = (caddr_t)kva_alloc(maxphys); 1213 1214 /* finally, initialize each buffer header and stick on empty q */ 1215 for (i = 0; i < nbuf; i++) { 1216 bp = nbufp(i); 1217 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf)); 1218 bp->b_flags = B_INVAL; 1219 bp->b_rcred = NOCRED; 1220 bp->b_wcred = NOCRED; 1221 bp->b_qindex = QUEUE_NONE; 1222 bp->b_domain = -1; 1223 bp->b_subqueue = mp_maxid + 1; 1224 bp->b_xflags = 0; 1225 bp->b_data = bp->b_kvabase = unmapped_buf; 1226 LIST_INIT(&bp->b_dep); 1227 BUF_LOCKINIT(bp, buf_wmesg); 1228 bq_insert(&bqempty, bp, false); 1229 } 1230 1231 /* 1232 * maxbufspace is the absolute maximum amount of buffer space we are 1233 * allowed to reserve in KVM and in real terms. The absolute maximum 1234 * is nominally used by metadata. hibufspace is the nominal maximum 1235 * used by most other requests. The differential is required to 1236 * ensure that metadata deadlocks don't occur. 1237 * 1238 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 1239 * this may result in KVM fragmentation which is not handled optimally 1240 * by the system. XXX This is less true with vmem. We could use 1241 * PAGE_SIZE. 1242 */ 1243 maxbufspace = (long)nbuf * BKVASIZE; 1244 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10); 1245 lobufspace = (hibufspace / 20) * 19; /* 95% */ 1246 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2; 1247 1248 /* 1249 * Note: The 16 MiB upper limit for hirunningspace was chosen 1250 * arbitrarily and may need further tuning. It corresponds to 1251 * 128 outstanding write IO requests (if IO size is 128 KiB), 1252 * which fits with many RAID controllers' tagged queuing limits. 1253 * The lower 1 MiB limit is the historical upper limit for 1254 * hirunningspace. 1255 */ 1256 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf), 1257 16 * 1024 * 1024), 1024 * 1024); 1258 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf); 1259 1260 /* 1261 * Limit the amount of malloc memory since it is wired permanently into 1262 * the kernel space. Even though this is accounted for in the buffer 1263 * allocation, we don't want the malloced region to grow uncontrolled. 1264 * The malloc scheme improves memory utilization significantly on 1265 * average (small) directories. 1266 */ 1267 maxbufmallocspace = hibufspace / 20; 1268 1269 /* 1270 * Reduce the chance of a deadlock occurring by limiting the number 1271 * of delayed-write dirty buffers we allow to stack up. 1272 */ 1273 hidirtybuffers = nbuf / 4 + 20; 1274 dirtybufthresh = hidirtybuffers * 9 / 10; 1275 /* 1276 * To support extreme low-memory systems, make sure hidirtybuffers 1277 * cannot eat up all available buffer space. This occurs when our 1278 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our 1279 * buffer space assuming BKVASIZE'd buffers. 1280 */ 1281 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 1282 hidirtybuffers >>= 1; 1283 } 1284 lodirtybuffers = hidirtybuffers / 2; 1285 1286 /* 1287 * lofreebuffers should be sufficient to avoid stalling waiting on 1288 * buf headers under heavy utilization. The bufs in per-cpu caches 1289 * are counted as free but will be unavailable to threads executing 1290 * on other cpus. 1291 * 1292 * hifreebuffers is the free target for the bufspace daemon. This 1293 * should be set appropriately to limit work per-iteration. 1294 */ 1295 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus); 1296 hifreebuffers = (3 * lofreebuffers) / 2; 1297 numfreebuffers = nbuf; 1298 1299 /* Setup the kva and free list allocators. */ 1300 vmem_set_reclaim(buffer_arena, bufkva_reclaim); 1301 buf_zone = uma_zcache_create("buf free cache", 1302 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf), 1303 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0); 1304 1305 /* 1306 * Size the clean queue according to the amount of buffer space. 1307 * One queue per-256mb up to the max. More queues gives better 1308 * concurrency but less accurate LRU. 1309 */ 1310 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS); 1311 for (i = 0 ; i < buf_domains; i++) { 1312 struct bufdomain *bd; 1313 1314 bd = &bdomain[i]; 1315 bd_init(bd); 1316 bd->bd_freebuffers = nbuf / buf_domains; 1317 bd->bd_hifreebuffers = hifreebuffers / buf_domains; 1318 bd->bd_lofreebuffers = lofreebuffers / buf_domains; 1319 bd->bd_bufspace = 0; 1320 bd->bd_maxbufspace = maxbufspace / buf_domains; 1321 bd->bd_hibufspace = hibufspace / buf_domains; 1322 bd->bd_lobufspace = lobufspace / buf_domains; 1323 bd->bd_bufspacethresh = bufspacethresh / buf_domains; 1324 bd->bd_numdirtybuffers = 0; 1325 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains; 1326 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains; 1327 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains; 1328 /* Don't allow more than 2% of bufs in the per-cpu caches. */ 1329 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus; 1330 } 1331 getnewbufcalls = counter_u64_alloc(M_WAITOK); 1332 getnewbufrestarts = counter_u64_alloc(M_WAITOK); 1333 mappingrestarts = counter_u64_alloc(M_WAITOK); 1334 numbufallocfails = counter_u64_alloc(M_WAITOK); 1335 notbufdflushes = counter_u64_alloc(M_WAITOK); 1336 buffreekvacnt = counter_u64_alloc(M_WAITOK); 1337 bufdefragcnt = counter_u64_alloc(M_WAITOK); 1338 bufkvaspace = counter_u64_alloc(M_WAITOK); 1339 } 1340 1341 #ifdef INVARIANTS 1342 static inline void 1343 vfs_buf_check_mapped(struct buf *bp) 1344 { 1345 1346 KASSERT(bp->b_kvabase != unmapped_buf, 1347 ("mapped buf: b_kvabase was not updated %p", bp)); 1348 KASSERT(bp->b_data != unmapped_buf, 1349 ("mapped buf: b_data was not updated %p", bp)); 1350 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf + 1351 maxphys, ("b_data + b_offset unmapped %p", bp)); 1352 } 1353 1354 static inline void 1355 vfs_buf_check_unmapped(struct buf *bp) 1356 { 1357 1358 KASSERT(bp->b_data == unmapped_buf, 1359 ("unmapped buf: corrupted b_data %p", bp)); 1360 } 1361 1362 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 1363 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 1364 #else 1365 #define BUF_CHECK_MAPPED(bp) do {} while (0) 1366 #define BUF_CHECK_UNMAPPED(bp) do {} while (0) 1367 #endif 1368 1369 static int 1370 isbufbusy(struct buf *bp) 1371 { 1372 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) || 1373 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI)) 1374 return (1); 1375 return (0); 1376 } 1377 1378 /* 1379 * Shutdown the system cleanly to prepare for reboot, halt, or power off. 1380 */ 1381 void 1382 bufshutdown(int show_busybufs) 1383 { 1384 static int first_buf_printf = 1; 1385 struct buf *bp; 1386 int i, iter, nbusy, pbusy; 1387 #ifndef PREEMPTION 1388 int subiter; 1389 #endif 1390 1391 /* 1392 * Sync filesystems for shutdown 1393 */ 1394 wdog_kern_pat(WD_LASTVAL); 1395 kern_sync(curthread); 1396 1397 /* 1398 * With soft updates, some buffers that are 1399 * written will be remarked as dirty until other 1400 * buffers are written. 1401 */ 1402 for (iter = pbusy = 0; iter < 20; iter++) { 1403 nbusy = 0; 1404 for (i = nbuf - 1; i >= 0; i--) { 1405 bp = nbufp(i); 1406 if (isbufbusy(bp)) 1407 nbusy++; 1408 } 1409 if (nbusy == 0) { 1410 if (first_buf_printf) 1411 printf("All buffers synced."); 1412 break; 1413 } 1414 if (first_buf_printf) { 1415 printf("Syncing disks, buffers remaining... "); 1416 first_buf_printf = 0; 1417 } 1418 printf("%d ", nbusy); 1419 if (nbusy < pbusy) 1420 iter = 0; 1421 pbusy = nbusy; 1422 1423 wdog_kern_pat(WD_LASTVAL); 1424 kern_sync(curthread); 1425 1426 #ifdef PREEMPTION 1427 /* 1428 * Spin for a while to allow interrupt threads to run. 1429 */ 1430 DELAY(50000 * iter); 1431 #else 1432 /* 1433 * Context switch several times to allow interrupt 1434 * threads to run. 1435 */ 1436 for (subiter = 0; subiter < 50 * iter; subiter++) { 1437 sched_relinquish(curthread); 1438 DELAY(1000); 1439 } 1440 #endif 1441 } 1442 printf("\n"); 1443 /* 1444 * Count only busy local buffers to prevent forcing 1445 * a fsck if we're just a client of a wedged NFS server 1446 */ 1447 nbusy = 0; 1448 for (i = nbuf - 1; i >= 0; i--) { 1449 bp = nbufp(i); 1450 if (isbufbusy(bp)) { 1451 #if 0 1452 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */ 1453 if (bp->b_dev == NULL) { 1454 TAILQ_REMOVE(&mountlist, 1455 bp->b_vp->v_mount, mnt_list); 1456 continue; 1457 } 1458 #endif 1459 nbusy++; 1460 if (show_busybufs > 0) { 1461 printf( 1462 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:", 1463 nbusy, bp, bp->b_vp, bp->b_flags, 1464 (intmax_t)bp->b_blkno, 1465 (intmax_t)bp->b_lblkno); 1466 BUF_LOCKPRINTINFO(bp); 1467 if (show_busybufs > 1) 1468 vn_printf(bp->b_vp, 1469 "vnode content: "); 1470 } 1471 } 1472 } 1473 if (nbusy) { 1474 /* 1475 * Failed to sync all blocks. Indicate this and don't 1476 * unmount filesystems (thus forcing an fsck on reboot). 1477 */ 1478 BOOTTRACE("shutdown failed to sync buffers"); 1479 printf("Giving up on %d buffers\n", nbusy); 1480 DELAY(5000000); /* 5 seconds */ 1481 swapoff_all(); 1482 } else { 1483 BOOTTRACE("shutdown sync complete"); 1484 if (!first_buf_printf) 1485 printf("Final sync complete\n"); 1486 1487 /* 1488 * Unmount filesystems and perform swapoff, to quiesce 1489 * the system as much as possible. In particular, no 1490 * I/O should be initiated from top levels since it 1491 * might be abruptly terminated by reset, or otherwise 1492 * erronously handled because other parts of the 1493 * system are disabled. 1494 * 1495 * Swapoff before unmount, because file-backed swap is 1496 * non-operational after unmount of the underlying 1497 * filesystem. 1498 */ 1499 if (!KERNEL_PANICKED()) { 1500 swapoff_all(); 1501 vfs_unmountall(); 1502 } 1503 BOOTTRACE("shutdown unmounted all filesystems"); 1504 } 1505 DELAY(100000); /* wait for console output to finish */ 1506 } 1507 1508 static void 1509 bpmap_qenter(struct buf *bp) 1510 { 1511 1512 BUF_CHECK_MAPPED(bp); 1513 1514 /* 1515 * bp->b_data is relative to bp->b_offset, but 1516 * bp->b_offset may be offset into the first page. 1517 */ 1518 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 1519 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 1520 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 1521 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 1522 } 1523 1524 static inline struct bufdomain * 1525 bufdomain(struct buf *bp) 1526 { 1527 1528 return (&bdomain[bp->b_domain]); 1529 } 1530 1531 static struct bufqueue * 1532 bufqueue(struct buf *bp) 1533 { 1534 1535 switch (bp->b_qindex) { 1536 case QUEUE_NONE: 1537 /* FALLTHROUGH */ 1538 case QUEUE_SENTINEL: 1539 return (NULL); 1540 case QUEUE_EMPTY: 1541 return (&bqempty); 1542 case QUEUE_DIRTY: 1543 return (&bufdomain(bp)->bd_dirtyq); 1544 case QUEUE_CLEAN: 1545 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]); 1546 default: 1547 break; 1548 } 1549 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex); 1550 } 1551 1552 /* 1553 * Return the locked bufqueue that bp is a member of. 1554 */ 1555 static struct bufqueue * 1556 bufqueue_acquire(struct buf *bp) 1557 { 1558 struct bufqueue *bq, *nbq; 1559 1560 /* 1561 * bp can be pushed from a per-cpu queue to the 1562 * cleanq while we're waiting on the lock. Retry 1563 * if the queues don't match. 1564 */ 1565 bq = bufqueue(bp); 1566 BQ_LOCK(bq); 1567 for (;;) { 1568 nbq = bufqueue(bp); 1569 if (bq == nbq) 1570 break; 1571 BQ_UNLOCK(bq); 1572 BQ_LOCK(nbq); 1573 bq = nbq; 1574 } 1575 return (bq); 1576 } 1577 1578 /* 1579 * binsfree: 1580 * 1581 * Insert the buffer into the appropriate free list. Requires a 1582 * locked buffer on entry and buffer is unlocked before return. 1583 */ 1584 static void 1585 binsfree(struct buf *bp, int qindex) 1586 { 1587 struct bufdomain *bd; 1588 struct bufqueue *bq; 1589 1590 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY, 1591 ("binsfree: Invalid qindex %d", qindex)); 1592 BUF_ASSERT_XLOCKED(bp); 1593 1594 /* 1595 * Handle delayed bremfree() processing. 1596 */ 1597 if (bp->b_flags & B_REMFREE) { 1598 if (bp->b_qindex == qindex) { 1599 bp->b_flags |= B_REUSE; 1600 bp->b_flags &= ~B_REMFREE; 1601 BUF_UNLOCK(bp); 1602 return; 1603 } 1604 bq = bufqueue_acquire(bp); 1605 bq_remove(bq, bp); 1606 BQ_UNLOCK(bq); 1607 } 1608 bd = bufdomain(bp); 1609 if (qindex == QUEUE_CLEAN) { 1610 if (bd->bd_lim != 0) 1611 bq = &bd->bd_subq[PCPU_GET(cpuid)]; 1612 else 1613 bq = bd->bd_cleanq; 1614 } else 1615 bq = &bd->bd_dirtyq; 1616 bq_insert(bq, bp, true); 1617 } 1618 1619 /* 1620 * buf_free: 1621 * 1622 * Free a buffer to the buf zone once it no longer has valid contents. 1623 */ 1624 static void 1625 buf_free(struct buf *bp) 1626 { 1627 1628 if (bp->b_flags & B_REMFREE) 1629 bremfreef(bp); 1630 if (bp->b_vflags & BV_BKGRDINPROG) 1631 panic("losing buffer 1"); 1632 if (bp->b_rcred != NOCRED) { 1633 crfree(bp->b_rcred); 1634 bp->b_rcred = NOCRED; 1635 } 1636 if (bp->b_wcred != NOCRED) { 1637 crfree(bp->b_wcred); 1638 bp->b_wcred = NOCRED; 1639 } 1640 if (!LIST_EMPTY(&bp->b_dep)) 1641 buf_deallocate(bp); 1642 bufkva_free(bp); 1643 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1); 1644 MPASS((bp->b_flags & B_MAXPHYS) == 0); 1645 BUF_UNLOCK(bp); 1646 uma_zfree(buf_zone, bp); 1647 } 1648 1649 /* 1650 * buf_import: 1651 * 1652 * Import bufs into the uma cache from the buf list. The system still 1653 * expects a static array of bufs and much of the synchronization 1654 * around bufs assumes type stable storage. As a result, UMA is used 1655 * only as a per-cpu cache of bufs still maintained on a global list. 1656 */ 1657 static int 1658 buf_import(void *arg, void **store, int cnt, int domain, int flags) 1659 { 1660 struct buf *bp; 1661 int i; 1662 1663 BQ_LOCK(&bqempty); 1664 for (i = 0; i < cnt; i++) { 1665 bp = TAILQ_FIRST(&bqempty.bq_queue); 1666 if (bp == NULL) 1667 break; 1668 bq_remove(&bqempty, bp); 1669 store[i] = bp; 1670 } 1671 BQ_UNLOCK(&bqempty); 1672 1673 return (i); 1674 } 1675 1676 /* 1677 * buf_release: 1678 * 1679 * Release bufs from the uma cache back to the buffer queues. 1680 */ 1681 static void 1682 buf_release(void *arg, void **store, int cnt) 1683 { 1684 struct bufqueue *bq; 1685 struct buf *bp; 1686 int i; 1687 1688 bq = &bqempty; 1689 BQ_LOCK(bq); 1690 for (i = 0; i < cnt; i++) { 1691 bp = store[i]; 1692 /* Inline bq_insert() to batch locking. */ 1693 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1694 bp->b_flags &= ~(B_AGE | B_REUSE); 1695 bq->bq_len++; 1696 bp->b_qindex = bq->bq_index; 1697 } 1698 BQ_UNLOCK(bq); 1699 } 1700 1701 /* 1702 * buf_alloc: 1703 * 1704 * Allocate an empty buffer header. 1705 */ 1706 static struct buf * 1707 buf_alloc(struct bufdomain *bd) 1708 { 1709 struct buf *bp; 1710 int freebufs, error; 1711 1712 /* 1713 * We can only run out of bufs in the buf zone if the average buf 1714 * is less than BKVASIZE. In this case the actual wait/block will 1715 * come from buf_reycle() failing to flush one of these small bufs. 1716 */ 1717 bp = NULL; 1718 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1); 1719 if (freebufs > 0) 1720 bp = uma_zalloc(buf_zone, M_NOWAIT); 1721 if (bp == NULL) { 1722 atomic_add_int(&bd->bd_freebuffers, 1); 1723 bufspace_daemon_wakeup(bd); 1724 counter_u64_add(numbufallocfails, 1); 1725 return (NULL); 1726 } 1727 /* 1728 * Wake-up the bufspace daemon on transition below threshold. 1729 */ 1730 if (freebufs == bd->bd_lofreebuffers) 1731 bufspace_daemon_wakeup(bd); 1732 1733 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWITNESS, NULL); 1734 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp, 1735 error)); 1736 (void)error; 1737 1738 KASSERT(bp->b_vp == NULL, 1739 ("bp: %p still has vnode %p.", bp, bp->b_vp)); 1740 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0, 1741 ("invalid buffer %p flags %#x", bp, bp->b_flags)); 1742 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1743 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 1744 KASSERT(bp->b_npages == 0, 1745 ("bp: %p still has %d vm pages\n", bp, bp->b_npages)); 1746 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp)); 1747 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp)); 1748 MPASS((bp->b_flags & B_MAXPHYS) == 0); 1749 1750 bp->b_domain = BD_DOMAIN(bd); 1751 bp->b_flags = 0; 1752 bp->b_ioflags = 0; 1753 bp->b_xflags = 0; 1754 bp->b_vflags = 0; 1755 bp->b_vp = NULL; 1756 bp->b_blkno = bp->b_lblkno = 0; 1757 bp->b_offset = NOOFFSET; 1758 bp->b_iodone = 0; 1759 bp->b_error = 0; 1760 bp->b_resid = 0; 1761 bp->b_bcount = 0; 1762 bp->b_npages = 0; 1763 bp->b_dirtyoff = bp->b_dirtyend = 0; 1764 bp->b_bufobj = NULL; 1765 bp->b_data = bp->b_kvabase = unmapped_buf; 1766 bp->b_fsprivate1 = NULL; 1767 bp->b_fsprivate2 = NULL; 1768 bp->b_fsprivate3 = NULL; 1769 LIST_INIT(&bp->b_dep); 1770 1771 return (bp); 1772 } 1773 1774 /* 1775 * buf_recycle: 1776 * 1777 * Free a buffer from the given bufqueue. kva controls whether the 1778 * freed buf must own some kva resources. This is used for 1779 * defragmenting. 1780 */ 1781 static int 1782 buf_recycle(struct bufdomain *bd, bool kva) 1783 { 1784 struct bufqueue *bq; 1785 struct buf *bp, *nbp; 1786 1787 if (kva) 1788 counter_u64_add(bufdefragcnt, 1); 1789 nbp = NULL; 1790 bq = bd->bd_cleanq; 1791 BQ_LOCK(bq); 1792 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd), 1793 ("buf_recycle: Locks don't match")); 1794 nbp = TAILQ_FIRST(&bq->bq_queue); 1795 1796 /* 1797 * Run scan, possibly freeing data and/or kva mappings on the fly 1798 * depending. 1799 */ 1800 while ((bp = nbp) != NULL) { 1801 /* 1802 * Calculate next bp (we can only use it if we do not 1803 * release the bqlock). 1804 */ 1805 nbp = TAILQ_NEXT(bp, b_freelist); 1806 1807 /* 1808 * If we are defragging then we need a buffer with 1809 * some kva to reclaim. 1810 */ 1811 if (kva && bp->b_kvasize == 0) 1812 continue; 1813 1814 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1815 continue; 1816 1817 /* 1818 * Implement a second chance algorithm for frequently 1819 * accessed buffers. 1820 */ 1821 if ((bp->b_flags & B_REUSE) != 0) { 1822 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1823 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1824 bp->b_flags &= ~B_REUSE; 1825 BUF_UNLOCK(bp); 1826 continue; 1827 } 1828 1829 /* 1830 * Skip buffers with background writes in progress. 1831 */ 1832 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 1833 BUF_UNLOCK(bp); 1834 continue; 1835 } 1836 1837 KASSERT(bp->b_qindex == QUEUE_CLEAN, 1838 ("buf_recycle: inconsistent queue %d bp %p", 1839 bp->b_qindex, bp)); 1840 KASSERT(bp->b_domain == BD_DOMAIN(bd), 1841 ("getnewbuf: queue domain %d doesn't match request %d", 1842 bp->b_domain, (int)BD_DOMAIN(bd))); 1843 /* 1844 * NOTE: nbp is now entirely invalid. We can only restart 1845 * the scan from this point on. 1846 */ 1847 bq_remove(bq, bp); 1848 BQ_UNLOCK(bq); 1849 1850 /* 1851 * Requeue the background write buffer with error and 1852 * restart the scan. 1853 */ 1854 if ((bp->b_vflags & BV_BKGRDERR) != 0) { 1855 bqrelse(bp); 1856 BQ_LOCK(bq); 1857 nbp = TAILQ_FIRST(&bq->bq_queue); 1858 continue; 1859 } 1860 bp->b_flags |= B_INVAL; 1861 brelse(bp); 1862 return (0); 1863 } 1864 bd->bd_wanted = 1; 1865 BQ_UNLOCK(bq); 1866 1867 return (ENOBUFS); 1868 } 1869 1870 /* 1871 * bremfree: 1872 * 1873 * Mark the buffer for removal from the appropriate free list. 1874 * 1875 */ 1876 void 1877 bremfree(struct buf *bp) 1878 { 1879 1880 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1881 KASSERT((bp->b_flags & B_REMFREE) == 0, 1882 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1883 KASSERT(bp->b_qindex != QUEUE_NONE, 1884 ("bremfree: buffer %p not on a queue.", bp)); 1885 BUF_ASSERT_XLOCKED(bp); 1886 1887 bp->b_flags |= B_REMFREE; 1888 } 1889 1890 /* 1891 * bremfreef: 1892 * 1893 * Force an immediate removal from a free list. Used only in nfs when 1894 * it abuses the b_freelist pointer. 1895 */ 1896 void 1897 bremfreef(struct buf *bp) 1898 { 1899 struct bufqueue *bq; 1900 1901 bq = bufqueue_acquire(bp); 1902 bq_remove(bq, bp); 1903 BQ_UNLOCK(bq); 1904 } 1905 1906 static void 1907 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname) 1908 { 1909 1910 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF); 1911 TAILQ_INIT(&bq->bq_queue); 1912 bq->bq_len = 0; 1913 bq->bq_index = qindex; 1914 bq->bq_subqueue = subqueue; 1915 } 1916 1917 static void 1918 bd_init(struct bufdomain *bd) 1919 { 1920 int i; 1921 1922 /* Per-CPU clean buf queues, plus one global queue. */ 1923 bd->bd_subq = mallocarray(mp_maxid + 2, sizeof(struct bufqueue), 1924 M_BIOBUF, M_WAITOK | M_ZERO); 1925 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1]; 1926 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock"); 1927 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock"); 1928 for (i = 0; i <= mp_maxid; i++) 1929 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i, 1930 "bufq clean subqueue lock"); 1931 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF); 1932 } 1933 1934 /* 1935 * bq_remove: 1936 * 1937 * Removes a buffer from the free list, must be called with the 1938 * correct qlock held. 1939 */ 1940 static void 1941 bq_remove(struct bufqueue *bq, struct buf *bp) 1942 { 1943 1944 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X", 1945 bp, bp->b_vp, bp->b_flags); 1946 KASSERT(bp->b_qindex != QUEUE_NONE, 1947 ("bq_remove: buffer %p not on a queue.", bp)); 1948 KASSERT(bufqueue(bp) == bq, 1949 ("bq_remove: Remove buffer %p from wrong queue.", bp)); 1950 1951 BQ_ASSERT_LOCKED(bq); 1952 if (bp->b_qindex != QUEUE_EMPTY) { 1953 BUF_ASSERT_XLOCKED(bp); 1954 } 1955 KASSERT(bq->bq_len >= 1, 1956 ("queue %d underflow", bp->b_qindex)); 1957 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1958 bq->bq_len--; 1959 bp->b_qindex = QUEUE_NONE; 1960 bp->b_flags &= ~(B_REMFREE | B_REUSE); 1961 } 1962 1963 static void 1964 bd_flush(struct bufdomain *bd, struct bufqueue *bq) 1965 { 1966 struct buf *bp; 1967 1968 BQ_ASSERT_LOCKED(bq); 1969 if (bq != bd->bd_cleanq) { 1970 BD_LOCK(bd); 1971 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) { 1972 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1973 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp, 1974 b_freelist); 1975 bp->b_subqueue = bd->bd_cleanq->bq_subqueue; 1976 } 1977 bd->bd_cleanq->bq_len += bq->bq_len; 1978 bq->bq_len = 0; 1979 } 1980 if (bd->bd_wanted) { 1981 bd->bd_wanted = 0; 1982 wakeup(&bd->bd_wanted); 1983 } 1984 if (bq != bd->bd_cleanq) 1985 BD_UNLOCK(bd); 1986 } 1987 1988 static int 1989 bd_flushall(struct bufdomain *bd) 1990 { 1991 struct bufqueue *bq; 1992 int flushed; 1993 int i; 1994 1995 if (bd->bd_lim == 0) 1996 return (0); 1997 flushed = 0; 1998 for (i = 0; i <= mp_maxid; i++) { 1999 bq = &bd->bd_subq[i]; 2000 if (bq->bq_len == 0) 2001 continue; 2002 BQ_LOCK(bq); 2003 bd_flush(bd, bq); 2004 BQ_UNLOCK(bq); 2005 flushed++; 2006 } 2007 2008 return (flushed); 2009 } 2010 2011 static void 2012 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock) 2013 { 2014 struct bufdomain *bd; 2015 2016 if (bp->b_qindex != QUEUE_NONE) 2017 panic("bq_insert: free buffer %p onto another queue?", bp); 2018 2019 bd = bufdomain(bp); 2020 if (bp->b_flags & B_AGE) { 2021 /* Place this buf directly on the real queue. */ 2022 if (bq->bq_index == QUEUE_CLEAN) 2023 bq = bd->bd_cleanq; 2024 BQ_LOCK(bq); 2025 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist); 2026 } else { 2027 BQ_LOCK(bq); 2028 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 2029 } 2030 bp->b_flags &= ~(B_AGE | B_REUSE); 2031 bq->bq_len++; 2032 bp->b_qindex = bq->bq_index; 2033 bp->b_subqueue = bq->bq_subqueue; 2034 2035 /* 2036 * Unlock before we notify so that we don't wakeup a waiter that 2037 * fails a trylock on the buf and sleeps again. 2038 */ 2039 if (unlock) 2040 BUF_UNLOCK(bp); 2041 2042 if (bp->b_qindex == QUEUE_CLEAN) { 2043 /* 2044 * Flush the per-cpu queue and notify any waiters. 2045 */ 2046 if (bd->bd_wanted || (bq != bd->bd_cleanq && 2047 bq->bq_len >= bd->bd_lim)) 2048 bd_flush(bd, bq); 2049 } 2050 BQ_UNLOCK(bq); 2051 } 2052 2053 /* 2054 * bufkva_free: 2055 * 2056 * Free the kva allocation for a buffer. 2057 * 2058 */ 2059 static void 2060 bufkva_free(struct buf *bp) 2061 { 2062 2063 #ifdef INVARIANTS 2064 if (bp->b_kvasize == 0) { 2065 KASSERT(bp->b_kvabase == unmapped_buf && 2066 bp->b_data == unmapped_buf, 2067 ("Leaked KVA space on %p", bp)); 2068 } else if (buf_mapped(bp)) 2069 BUF_CHECK_MAPPED(bp); 2070 else 2071 BUF_CHECK_UNMAPPED(bp); 2072 #endif 2073 if (bp->b_kvasize == 0) 2074 return; 2075 2076 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize); 2077 counter_u64_add(bufkvaspace, -bp->b_kvasize); 2078 counter_u64_add(buffreekvacnt, 1); 2079 bp->b_data = bp->b_kvabase = unmapped_buf; 2080 bp->b_kvasize = 0; 2081 } 2082 2083 /* 2084 * bufkva_alloc: 2085 * 2086 * Allocate the buffer KVA and set b_kvasize and b_kvabase. 2087 */ 2088 static int 2089 bufkva_alloc(struct buf *bp, int maxsize, int gbflags) 2090 { 2091 vm_offset_t addr; 2092 int error; 2093 2094 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0, 2095 ("Invalid gbflags 0x%x in %s", gbflags, __func__)); 2096 MPASS((bp->b_flags & B_MAXPHYS) == 0); 2097 KASSERT(maxsize <= maxbcachebuf, 2098 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf)); 2099 2100 bufkva_free(bp); 2101 2102 addr = 0; 2103 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr); 2104 if (error != 0) { 2105 /* 2106 * Buffer map is too fragmented. Request the caller 2107 * to defragment the map. 2108 */ 2109 return (error); 2110 } 2111 bp->b_kvabase = (caddr_t)addr; 2112 bp->b_kvasize = maxsize; 2113 counter_u64_add(bufkvaspace, bp->b_kvasize); 2114 if ((gbflags & GB_UNMAPPED) != 0) { 2115 bp->b_data = unmapped_buf; 2116 BUF_CHECK_UNMAPPED(bp); 2117 } else { 2118 bp->b_data = bp->b_kvabase; 2119 BUF_CHECK_MAPPED(bp); 2120 } 2121 return (0); 2122 } 2123 2124 /* 2125 * bufkva_reclaim: 2126 * 2127 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem 2128 * callback that fires to avoid returning failure. 2129 */ 2130 static void 2131 bufkva_reclaim(vmem_t *vmem, int flags) 2132 { 2133 bool done; 2134 int q; 2135 int i; 2136 2137 done = false; 2138 for (i = 0; i < 5; i++) { 2139 for (q = 0; q < buf_domains; q++) 2140 if (buf_recycle(&bdomain[q], true) != 0) 2141 done = true; 2142 if (done) 2143 break; 2144 } 2145 return; 2146 } 2147 2148 /* 2149 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 2150 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 2151 * the buffer is valid and we do not have to do anything. 2152 */ 2153 static void 2154 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt, 2155 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *)) 2156 { 2157 struct buf *rabp; 2158 struct thread *td; 2159 int i; 2160 2161 td = curthread; 2162 2163 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 2164 if (inmem(vp, *rablkno)) 2165 continue; 2166 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 2167 if ((rabp->b_flags & B_CACHE) != 0) { 2168 brelse(rabp); 2169 continue; 2170 } 2171 #ifdef RACCT 2172 if (racct_enable) { 2173 PROC_LOCK(curproc); 2174 racct_add_buf(curproc, rabp, 0); 2175 PROC_UNLOCK(curproc); 2176 } 2177 #endif /* RACCT */ 2178 td->td_ru.ru_inblock++; 2179 rabp->b_flags |= B_ASYNC; 2180 rabp->b_flags &= ~B_INVAL; 2181 if ((flags & GB_CKHASH) != 0) { 2182 rabp->b_flags |= B_CKHASH; 2183 rabp->b_ckhashcalc = ckhashfunc; 2184 } 2185 rabp->b_ioflags &= ~BIO_ERROR; 2186 rabp->b_iocmd = BIO_READ; 2187 if (rabp->b_rcred == NOCRED && cred != NOCRED) 2188 rabp->b_rcred = crhold(cred); 2189 vfs_busy_pages(rabp, 0); 2190 BUF_KERNPROC(rabp); 2191 rabp->b_iooffset = dbtob(rabp->b_blkno); 2192 bstrategy(rabp); 2193 } 2194 } 2195 2196 /* 2197 * Entry point for bread() and breadn() via #defines in sys/buf.h. 2198 * 2199 * Get a buffer with the specified data. Look in the cache first. We 2200 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 2201 * is set, the buffer is valid and we do not have to do anything, see 2202 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 2203 * 2204 * Always return a NULL buffer pointer (in bpp) when returning an error. 2205 * 2206 * The blkno parameter is the logical block being requested. Normally 2207 * the mapping of logical block number to disk block address is done 2208 * by calling VOP_BMAP(). However, if the mapping is already known, the 2209 * disk block address can be passed using the dblkno parameter. If the 2210 * disk block address is not known, then the same value should be passed 2211 * for blkno and dblkno. 2212 */ 2213 int 2214 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, 2215 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags, 2216 void (*ckhashfunc)(struct buf *), struct buf **bpp) 2217 { 2218 struct buf *bp; 2219 struct thread *td; 2220 int error, readwait, rv; 2221 2222 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 2223 td = curthread; 2224 /* 2225 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags 2226 * are specified. 2227 */ 2228 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp); 2229 if (error != 0) { 2230 *bpp = NULL; 2231 return (error); 2232 } 2233 KASSERT(blkno == bp->b_lblkno, 2234 ("getblkx returned buffer for blkno %jd instead of blkno %jd", 2235 (intmax_t)bp->b_lblkno, (intmax_t)blkno)); 2236 flags &= ~GB_NOSPARSE; 2237 *bpp = bp; 2238 2239 /* 2240 * If not found in cache, do some I/O 2241 */ 2242 readwait = 0; 2243 if ((bp->b_flags & B_CACHE) == 0) { 2244 #ifdef RACCT 2245 if (racct_enable) { 2246 PROC_LOCK(td->td_proc); 2247 racct_add_buf(td->td_proc, bp, 0); 2248 PROC_UNLOCK(td->td_proc); 2249 } 2250 #endif /* RACCT */ 2251 td->td_ru.ru_inblock++; 2252 bp->b_iocmd = BIO_READ; 2253 bp->b_flags &= ~B_INVAL; 2254 if ((flags & GB_CKHASH) != 0) { 2255 bp->b_flags |= B_CKHASH; 2256 bp->b_ckhashcalc = ckhashfunc; 2257 } 2258 if ((flags & GB_CVTENXIO) != 0) 2259 bp->b_xflags |= BX_CVTENXIO; 2260 bp->b_ioflags &= ~BIO_ERROR; 2261 if (bp->b_rcred == NOCRED && cred != NOCRED) 2262 bp->b_rcred = crhold(cred); 2263 vfs_busy_pages(bp, 0); 2264 bp->b_iooffset = dbtob(bp->b_blkno); 2265 bstrategy(bp); 2266 ++readwait; 2267 } 2268 2269 /* 2270 * Attempt to initiate asynchronous I/O on read-ahead blocks. 2271 */ 2272 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc); 2273 2274 rv = 0; 2275 if (readwait) { 2276 rv = bufwait(bp); 2277 if (rv != 0) { 2278 brelse(bp); 2279 *bpp = NULL; 2280 } 2281 } 2282 return (rv); 2283 } 2284 2285 /* 2286 * Write, release buffer on completion. (Done by iodone 2287 * if async). Do not bother writing anything if the buffer 2288 * is invalid. 2289 * 2290 * Note that we set B_CACHE here, indicating that buffer is 2291 * fully valid and thus cacheable. This is true even of NFS 2292 * now so we set it generally. This could be set either here 2293 * or in biodone() since the I/O is synchronous. We put it 2294 * here. 2295 */ 2296 int 2297 bufwrite(struct buf *bp) 2298 { 2299 int oldflags; 2300 struct vnode *vp; 2301 long space; 2302 int vp_md; 2303 2304 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2305 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) { 2306 bp->b_flags |= B_INVAL | B_RELBUF; 2307 bp->b_flags &= ~B_CACHE; 2308 brelse(bp); 2309 return (ENXIO); 2310 } 2311 if (bp->b_flags & B_INVAL) { 2312 brelse(bp); 2313 return (0); 2314 } 2315 2316 if (bp->b_flags & B_BARRIER) 2317 atomic_add_long(&barrierwrites, 1); 2318 2319 oldflags = bp->b_flags; 2320 2321 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 2322 ("FFS background buffer should not get here %p", bp)); 2323 2324 vp = bp->b_vp; 2325 if (vp) 2326 vp_md = vp->v_vflag & VV_MD; 2327 else 2328 vp_md = 0; 2329 2330 /* 2331 * Mark the buffer clean. Increment the bufobj write count 2332 * before bundirty() call, to prevent other thread from seeing 2333 * empty dirty list and zero counter for writes in progress, 2334 * falsely indicating that the bufobj is clean. 2335 */ 2336 bufobj_wref(bp->b_bufobj); 2337 bundirty(bp); 2338 2339 bp->b_flags &= ~B_DONE; 2340 bp->b_ioflags &= ~BIO_ERROR; 2341 bp->b_flags |= B_CACHE; 2342 bp->b_iocmd = BIO_WRITE; 2343 2344 vfs_busy_pages(bp, 1); 2345 2346 /* 2347 * Normal bwrites pipeline writes 2348 */ 2349 bp->b_runningbufspace = bp->b_bufsize; 2350 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); 2351 2352 #ifdef RACCT 2353 if (racct_enable) { 2354 PROC_LOCK(curproc); 2355 racct_add_buf(curproc, bp, 1); 2356 PROC_UNLOCK(curproc); 2357 } 2358 #endif /* RACCT */ 2359 curthread->td_ru.ru_oublock++; 2360 if (oldflags & B_ASYNC) 2361 BUF_KERNPROC(bp); 2362 bp->b_iooffset = dbtob(bp->b_blkno); 2363 buf_track(bp, __func__); 2364 bstrategy(bp); 2365 2366 if ((oldflags & B_ASYNC) == 0) { 2367 int rtval = bufwait(bp); 2368 brelse(bp); 2369 return (rtval); 2370 } else if (space > hirunningspace) { 2371 /* 2372 * don't allow the async write to saturate the I/O 2373 * system. We will not deadlock here because 2374 * we are blocking waiting for I/O that is already in-progress 2375 * to complete. We do not block here if it is the update 2376 * or syncer daemon trying to clean up as that can lead 2377 * to deadlock. 2378 */ 2379 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 2380 waitrunningbufspace(); 2381 } 2382 2383 return (0); 2384 } 2385 2386 void 2387 bufbdflush(struct bufobj *bo, struct buf *bp) 2388 { 2389 struct buf *nbp; 2390 struct bufdomain *bd; 2391 2392 bd = &bdomain[bo->bo_domain]; 2393 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) { 2394 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 2395 altbufferflushes++; 2396 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) { 2397 BO_LOCK(bo); 2398 /* 2399 * Try to find a buffer to flush. 2400 */ 2401 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 2402 if ((nbp->b_vflags & BV_BKGRDINPROG) || 2403 BUF_LOCK(nbp, 2404 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 2405 continue; 2406 if (bp == nbp) 2407 panic("bdwrite: found ourselves"); 2408 BO_UNLOCK(bo); 2409 /* Don't countdeps with the bo lock held. */ 2410 if (buf_countdeps(nbp, 0)) { 2411 BO_LOCK(bo); 2412 BUF_UNLOCK(nbp); 2413 continue; 2414 } 2415 if (nbp->b_flags & B_CLUSTEROK) { 2416 vfs_bio_awrite(nbp); 2417 } else { 2418 bremfree(nbp); 2419 bawrite(nbp); 2420 } 2421 dirtybufferflushes++; 2422 break; 2423 } 2424 if (nbp == NULL) 2425 BO_UNLOCK(bo); 2426 } 2427 } 2428 2429 /* 2430 * Delayed write. (Buffer is marked dirty). Do not bother writing 2431 * anything if the buffer is marked invalid. 2432 * 2433 * Note that since the buffer must be completely valid, we can safely 2434 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 2435 * biodone() in order to prevent getblk from writing the buffer 2436 * out synchronously. 2437 */ 2438 void 2439 bdwrite(struct buf *bp) 2440 { 2441 struct thread *td = curthread; 2442 struct vnode *vp; 2443 struct bufobj *bo; 2444 2445 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2446 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2447 KASSERT((bp->b_flags & B_BARRIER) == 0, 2448 ("Barrier request in delayed write %p", bp)); 2449 2450 if (bp->b_flags & B_INVAL) { 2451 brelse(bp); 2452 return; 2453 } 2454 2455 /* 2456 * If we have too many dirty buffers, don't create any more. 2457 * If we are wildly over our limit, then force a complete 2458 * cleanup. Otherwise, just keep the situation from getting 2459 * out of control. Note that we have to avoid a recursive 2460 * disaster and not try to clean up after our own cleanup! 2461 */ 2462 vp = bp->b_vp; 2463 bo = bp->b_bufobj; 2464 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 2465 td->td_pflags |= TDP_INBDFLUSH; 2466 BO_BDFLUSH(bo, bp); 2467 td->td_pflags &= ~TDP_INBDFLUSH; 2468 } else 2469 recursiveflushes++; 2470 2471 bdirty(bp); 2472 /* 2473 * Set B_CACHE, indicating that the buffer is fully valid. This is 2474 * true even of NFS now. 2475 */ 2476 bp->b_flags |= B_CACHE; 2477 2478 /* 2479 * This bmap keeps the system from needing to do the bmap later, 2480 * perhaps when the system is attempting to do a sync. Since it 2481 * is likely that the indirect block -- or whatever other datastructure 2482 * that the filesystem needs is still in memory now, it is a good 2483 * thing to do this. Note also, that if the pageout daemon is 2484 * requesting a sync -- there might not be enough memory to do 2485 * the bmap then... So, this is important to do. 2486 */ 2487 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 2488 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 2489 } 2490 2491 buf_track(bp, __func__); 2492 2493 /* 2494 * Set the *dirty* buffer range based upon the VM system dirty 2495 * pages. 2496 * 2497 * Mark the buffer pages as clean. We need to do this here to 2498 * satisfy the vnode_pager and the pageout daemon, so that it 2499 * thinks that the pages have been "cleaned". Note that since 2500 * the pages are in a delayed write buffer -- the VFS layer 2501 * "will" see that the pages get written out on the next sync, 2502 * or perhaps the cluster will be completed. 2503 */ 2504 vfs_clean_pages_dirty_buf(bp); 2505 bqrelse(bp); 2506 2507 /* 2508 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 2509 * due to the softdep code. 2510 */ 2511 } 2512 2513 /* 2514 * bdirty: 2515 * 2516 * Turn buffer into delayed write request. We must clear BIO_READ and 2517 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 2518 * itself to properly update it in the dirty/clean lists. We mark it 2519 * B_DONE to ensure that any asynchronization of the buffer properly 2520 * clears B_DONE ( else a panic will occur later ). 2521 * 2522 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 2523 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 2524 * should only be called if the buffer is known-good. 2525 * 2526 * Since the buffer is not on a queue, we do not update the numfreebuffers 2527 * count. 2528 * 2529 * The buffer must be on QUEUE_NONE. 2530 */ 2531 void 2532 bdirty(struct buf *bp) 2533 { 2534 2535 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 2536 bp, bp->b_vp, bp->b_flags); 2537 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2538 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2539 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2540 bp->b_flags &= ~(B_RELBUF); 2541 bp->b_iocmd = BIO_WRITE; 2542 2543 if ((bp->b_flags & B_DELWRI) == 0) { 2544 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 2545 reassignbuf(bp); 2546 bdirtyadd(bp); 2547 } 2548 } 2549 2550 /* 2551 * bundirty: 2552 * 2553 * Clear B_DELWRI for buffer. 2554 * 2555 * Since the buffer is not on a queue, we do not update the numfreebuffers 2556 * count. 2557 * 2558 * The buffer must be on QUEUE_NONE. 2559 */ 2560 2561 void 2562 bundirty(struct buf *bp) 2563 { 2564 2565 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2566 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2567 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2568 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2569 2570 if (bp->b_flags & B_DELWRI) { 2571 bp->b_flags &= ~B_DELWRI; 2572 reassignbuf(bp); 2573 bdirtysub(bp); 2574 } 2575 /* 2576 * Since it is now being written, we can clear its deferred write flag. 2577 */ 2578 bp->b_flags &= ~B_DEFERRED; 2579 } 2580 2581 /* 2582 * bawrite: 2583 * 2584 * Asynchronous write. Start output on a buffer, but do not wait for 2585 * it to complete. The buffer is released when the output completes. 2586 * 2587 * bwrite() ( or the VOP routine anyway ) is responsible for handling 2588 * B_INVAL buffers. Not us. 2589 */ 2590 void 2591 bawrite(struct buf *bp) 2592 { 2593 2594 bp->b_flags |= B_ASYNC; 2595 (void) bwrite(bp); 2596 } 2597 2598 /* 2599 * babarrierwrite: 2600 * 2601 * Asynchronous barrier write. Start output on a buffer, but do not 2602 * wait for it to complete. Place a write barrier after this write so 2603 * that this buffer and all buffers written before it are committed to 2604 * the disk before any buffers written after this write are committed 2605 * to the disk. The buffer is released when the output completes. 2606 */ 2607 void 2608 babarrierwrite(struct buf *bp) 2609 { 2610 2611 bp->b_flags |= B_ASYNC | B_BARRIER; 2612 (void) bwrite(bp); 2613 } 2614 2615 /* 2616 * bbarrierwrite: 2617 * 2618 * Synchronous barrier write. Start output on a buffer and wait for 2619 * it to complete. Place a write barrier after this write so that 2620 * this buffer and all buffers written before it are committed to 2621 * the disk before any buffers written after this write are committed 2622 * to the disk. The buffer is released when the output completes. 2623 */ 2624 int 2625 bbarrierwrite(struct buf *bp) 2626 { 2627 2628 bp->b_flags |= B_BARRIER; 2629 return (bwrite(bp)); 2630 } 2631 2632 /* 2633 * bwillwrite: 2634 * 2635 * Called prior to the locking of any vnodes when we are expecting to 2636 * write. We do not want to starve the buffer cache with too many 2637 * dirty buffers so we block here. By blocking prior to the locking 2638 * of any vnodes we attempt to avoid the situation where a locked vnode 2639 * prevents the various system daemons from flushing related buffers. 2640 */ 2641 void 2642 bwillwrite(void) 2643 { 2644 2645 if (buf_dirty_count_severe()) { 2646 mtx_lock(&bdirtylock); 2647 while (buf_dirty_count_severe()) { 2648 bdirtywait = 1; 2649 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), 2650 "flswai", 0); 2651 } 2652 mtx_unlock(&bdirtylock); 2653 } 2654 } 2655 2656 /* 2657 * Return true if we have too many dirty buffers. 2658 */ 2659 int 2660 buf_dirty_count_severe(void) 2661 { 2662 2663 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty)); 2664 } 2665 2666 /* 2667 * brelse: 2668 * 2669 * Release a busy buffer and, if requested, free its resources. The 2670 * buffer will be stashed in the appropriate bufqueue[] allowing it 2671 * to be accessed later as a cache entity or reused for other purposes. 2672 */ 2673 void 2674 brelse(struct buf *bp) 2675 { 2676 struct mount *v_mnt; 2677 int qindex; 2678 2679 /* 2680 * Many functions erroneously call brelse with a NULL bp under rare 2681 * error conditions. Simply return when called with a NULL bp. 2682 */ 2683 if (bp == NULL) 2684 return; 2685 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 2686 bp, bp->b_vp, bp->b_flags); 2687 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2688 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2689 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0, 2690 ("brelse: non-VMIO buffer marked NOREUSE")); 2691 2692 if (BUF_LOCKRECURSED(bp)) { 2693 /* 2694 * Do not process, in particular, do not handle the 2695 * B_INVAL/B_RELBUF and do not release to free list. 2696 */ 2697 BUF_UNLOCK(bp); 2698 return; 2699 } 2700 2701 if (bp->b_flags & B_MANAGED) { 2702 bqrelse(bp); 2703 return; 2704 } 2705 2706 if (LIST_EMPTY(&bp->b_dep)) { 2707 bp->b_flags &= ~B_IOSTARTED; 2708 } else { 2709 KASSERT((bp->b_flags & B_IOSTARTED) == 0, 2710 ("brelse: SU io not finished bp %p", bp)); 2711 } 2712 2713 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { 2714 BO_LOCK(bp->b_bufobj); 2715 bp->b_vflags &= ~BV_BKGRDERR; 2716 BO_UNLOCK(bp->b_bufobj); 2717 bdirty(bp); 2718 } 2719 2720 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 2721 (bp->b_flags & B_INVALONERR)) { 2722 /* 2723 * Forced invalidation of dirty buffer contents, to be used 2724 * after a failed write in the rare case that the loss of the 2725 * contents is acceptable. The buffer is invalidated and 2726 * freed. 2727 */ 2728 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE; 2729 bp->b_flags &= ~(B_ASYNC | B_CACHE); 2730 } 2731 2732 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 2733 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) && 2734 !(bp->b_flags & B_INVAL)) { 2735 /* 2736 * Failed write, redirty. All errors except ENXIO (which 2737 * means the device is gone) are treated as being 2738 * transient. 2739 * 2740 * XXX Treating EIO as transient is not correct; the 2741 * contract with the local storage device drivers is that 2742 * they will only return EIO once the I/O is no longer 2743 * retriable. Network I/O also respects this through the 2744 * guarantees of TCP and/or the internal retries of NFS. 2745 * ENOMEM might be transient, but we also have no way of 2746 * knowing when its ok to retry/reschedule. In general, 2747 * this entire case should be made obsolete through better 2748 * error handling/recovery and resource scheduling. 2749 * 2750 * Do this also for buffers that failed with ENXIO, but have 2751 * non-empty dependencies - the soft updates code might need 2752 * to access the buffer to untangle them. 2753 * 2754 * Must clear BIO_ERROR to prevent pages from being scrapped. 2755 */ 2756 bp->b_ioflags &= ~BIO_ERROR; 2757 bdirty(bp); 2758 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 2759 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 2760 /* 2761 * Either a failed read I/O, or we were asked to free or not 2762 * cache the buffer, or we failed to write to a device that's 2763 * no longer present. 2764 */ 2765 bp->b_flags |= B_INVAL; 2766 if (!LIST_EMPTY(&bp->b_dep)) 2767 buf_deallocate(bp); 2768 if (bp->b_flags & B_DELWRI) 2769 bdirtysub(bp); 2770 bp->b_flags &= ~(B_DELWRI | B_CACHE); 2771 if ((bp->b_flags & B_VMIO) == 0) { 2772 allocbuf(bp, 0); 2773 if (bp->b_vp) 2774 brelvp(bp); 2775 } 2776 } 2777 2778 /* 2779 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate() 2780 * is called with B_DELWRI set, the underlying pages may wind up 2781 * getting freed causing a previous write (bdwrite()) to get 'lost' 2782 * because pages associated with a B_DELWRI bp are marked clean. 2783 * 2784 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even 2785 * if B_DELWRI is set. 2786 */ 2787 if (bp->b_flags & B_DELWRI) 2788 bp->b_flags &= ~B_RELBUF; 2789 2790 /* 2791 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 2792 * constituted, not even NFS buffers now. Two flags effect this. If 2793 * B_INVAL, the struct buf is invalidated but the VM object is kept 2794 * around ( i.e. so it is trivial to reconstitute the buffer later ). 2795 * 2796 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 2797 * invalidated. BIO_ERROR cannot be set for a failed write unless the 2798 * buffer is also B_INVAL because it hits the re-dirtying code above. 2799 * 2800 * Normally we can do this whether a buffer is B_DELWRI or not. If 2801 * the buffer is an NFS buffer, it is tracking piecemeal writes or 2802 * the commit state and we cannot afford to lose the buffer. If the 2803 * buffer has a background write in progress, we need to keep it 2804 * around to prevent it from being reconstituted and starting a second 2805 * background write. 2806 */ 2807 2808 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL; 2809 2810 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE || 2811 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) && 2812 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 || 2813 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) { 2814 vfs_vmio_invalidate(bp); 2815 allocbuf(bp, 0); 2816 } 2817 2818 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 || 2819 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) { 2820 allocbuf(bp, 0); 2821 bp->b_flags &= ~B_NOREUSE; 2822 if (bp->b_vp != NULL) 2823 brelvp(bp); 2824 } 2825 2826 /* 2827 * If the buffer has junk contents signal it and eventually 2828 * clean up B_DELWRI and diassociate the vnode so that gbincore() 2829 * doesn't find it. 2830 */ 2831 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 2832 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 2833 bp->b_flags |= B_INVAL; 2834 if (bp->b_flags & B_INVAL) { 2835 if (bp->b_flags & B_DELWRI) 2836 bundirty(bp); 2837 if (bp->b_vp) 2838 brelvp(bp); 2839 } 2840 2841 buf_track(bp, __func__); 2842 2843 /* buffers with no memory */ 2844 if (bp->b_bufsize == 0) { 2845 buf_free(bp); 2846 return; 2847 } 2848 /* buffers with junk contents */ 2849 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 2850 (bp->b_ioflags & BIO_ERROR)) { 2851 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 2852 if (bp->b_vflags & BV_BKGRDINPROG) 2853 panic("losing buffer 2"); 2854 qindex = QUEUE_CLEAN; 2855 bp->b_flags |= B_AGE; 2856 /* remaining buffers */ 2857 } else if (bp->b_flags & B_DELWRI) 2858 qindex = QUEUE_DIRTY; 2859 else 2860 qindex = QUEUE_CLEAN; 2861 2862 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 2863 panic("brelse: not dirty"); 2864 2865 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT); 2866 bp->b_xflags &= ~(BX_CVTENXIO); 2867 /* binsfree unlocks bp. */ 2868 binsfree(bp, qindex); 2869 } 2870 2871 /* 2872 * Release a buffer back to the appropriate queue but do not try to free 2873 * it. The buffer is expected to be used again soon. 2874 * 2875 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 2876 * biodone() to requeue an async I/O on completion. It is also used when 2877 * known good buffers need to be requeued but we think we may need the data 2878 * again soon. 2879 * 2880 * XXX we should be able to leave the B_RELBUF hint set on completion. 2881 */ 2882 void 2883 bqrelse(struct buf *bp) 2884 { 2885 int qindex; 2886 2887 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2888 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2889 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2890 2891 qindex = QUEUE_NONE; 2892 if (BUF_LOCKRECURSED(bp)) { 2893 /* do not release to free list */ 2894 BUF_UNLOCK(bp); 2895 return; 2896 } 2897 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 2898 bp->b_xflags &= ~(BX_CVTENXIO); 2899 2900 if (LIST_EMPTY(&bp->b_dep)) { 2901 bp->b_flags &= ~B_IOSTARTED; 2902 } else { 2903 KASSERT((bp->b_flags & B_IOSTARTED) == 0, 2904 ("bqrelse: SU io not finished bp %p", bp)); 2905 } 2906 2907 if (bp->b_flags & B_MANAGED) { 2908 if (bp->b_flags & B_REMFREE) 2909 bremfreef(bp); 2910 goto out; 2911 } 2912 2913 /* buffers with stale but valid contents */ 2914 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | 2915 BV_BKGRDERR)) == BV_BKGRDERR) { 2916 BO_LOCK(bp->b_bufobj); 2917 bp->b_vflags &= ~BV_BKGRDERR; 2918 BO_UNLOCK(bp->b_bufobj); 2919 qindex = QUEUE_DIRTY; 2920 } else { 2921 if ((bp->b_flags & B_DELWRI) == 0 && 2922 (bp->b_xflags & BX_VNDIRTY)) 2923 panic("bqrelse: not dirty"); 2924 if ((bp->b_flags & B_NOREUSE) != 0) { 2925 brelse(bp); 2926 return; 2927 } 2928 qindex = QUEUE_CLEAN; 2929 } 2930 buf_track(bp, __func__); 2931 /* binsfree unlocks bp. */ 2932 binsfree(bp, qindex); 2933 return; 2934 2935 out: 2936 buf_track(bp, __func__); 2937 /* unlock */ 2938 BUF_UNLOCK(bp); 2939 } 2940 2941 /* 2942 * Complete I/O to a VMIO backed page. Validate the pages as appropriate, 2943 * restore bogus pages. 2944 */ 2945 static void 2946 vfs_vmio_iodone(struct buf *bp) 2947 { 2948 vm_ooffset_t foff; 2949 vm_page_t m; 2950 vm_object_t obj; 2951 struct vnode *vp __unused; 2952 int i, iosize, resid; 2953 bool bogus; 2954 2955 obj = bp->b_bufobj->bo_object; 2956 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages, 2957 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)", 2958 blockcount_read(&obj->paging_in_progress), bp->b_npages)); 2959 2960 vp = bp->b_vp; 2961 VNPASS(vp->v_holdcnt > 0, vp); 2962 VNPASS(vp->v_object != NULL, vp); 2963 2964 foff = bp->b_offset; 2965 KASSERT(bp->b_offset != NOOFFSET, 2966 ("vfs_vmio_iodone: bp %p has no buffer offset", bp)); 2967 2968 bogus = false; 2969 iosize = bp->b_bcount - bp->b_resid; 2970 for (i = 0; i < bp->b_npages; i++) { 2971 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2972 if (resid > iosize) 2973 resid = iosize; 2974 2975 /* 2976 * cleanup bogus pages, restoring the originals 2977 */ 2978 m = bp->b_pages[i]; 2979 if (m == bogus_page) { 2980 bogus = true; 2981 m = vm_page_relookup(obj, OFF_TO_IDX(foff)); 2982 if (m == NULL) 2983 panic("biodone: page disappeared!"); 2984 bp->b_pages[i] = m; 2985 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) { 2986 /* 2987 * In the write case, the valid and clean bits are 2988 * already changed correctly ( see bdwrite() ), so we 2989 * only need to do this here in the read case. 2990 */ 2991 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK, 2992 resid)) == 0, ("vfs_vmio_iodone: page %p " 2993 "has unexpected dirty bits", m)); 2994 vfs_page_set_valid(bp, foff, m); 2995 } 2996 KASSERT(OFF_TO_IDX(foff) == m->pindex, 2997 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch", 2998 (intmax_t)foff, (uintmax_t)m->pindex)); 2999 3000 vm_page_sunbusy(m); 3001 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3002 iosize -= resid; 3003 } 3004 vm_object_pip_wakeupn(obj, bp->b_npages); 3005 if (bogus && buf_mapped(bp)) { 3006 BUF_CHECK_MAPPED(bp); 3007 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3008 bp->b_pages, bp->b_npages); 3009 } 3010 } 3011 3012 /* 3013 * Perform page invalidation when a buffer is released. The fully invalid 3014 * pages will be reclaimed later in vfs_vmio_truncate(). 3015 */ 3016 static void 3017 vfs_vmio_invalidate(struct buf *bp) 3018 { 3019 vm_object_t obj; 3020 vm_page_t m; 3021 int flags, i, resid, poffset, presid; 3022 3023 if (buf_mapped(bp)) { 3024 BUF_CHECK_MAPPED(bp); 3025 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 3026 } else 3027 BUF_CHECK_UNMAPPED(bp); 3028 /* 3029 * Get the base offset and length of the buffer. Note that 3030 * in the VMIO case if the buffer block size is not 3031 * page-aligned then b_data pointer may not be page-aligned. 3032 * But our b_pages[] array *IS* page aligned. 3033 * 3034 * block sizes less then DEV_BSIZE (usually 512) are not 3035 * supported due to the page granularity bits (m->valid, 3036 * m->dirty, etc...). 3037 * 3038 * See man buf(9) for more information 3039 */ 3040 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; 3041 obj = bp->b_bufobj->bo_object; 3042 resid = bp->b_bufsize; 3043 poffset = bp->b_offset & PAGE_MASK; 3044 VM_OBJECT_WLOCK(obj); 3045 for (i = 0; i < bp->b_npages; i++) { 3046 m = bp->b_pages[i]; 3047 if (m == bogus_page) 3048 panic("vfs_vmio_invalidate: Unexpected bogus page."); 3049 bp->b_pages[i] = NULL; 3050 3051 presid = resid > (PAGE_SIZE - poffset) ? 3052 (PAGE_SIZE - poffset) : resid; 3053 KASSERT(presid >= 0, ("brelse: extra page")); 3054 vm_page_busy_acquire(m, VM_ALLOC_SBUSY); 3055 if (pmap_page_wired_mappings(m) == 0) 3056 vm_page_set_invalid(m, poffset, presid); 3057 vm_page_sunbusy(m); 3058 vm_page_release_locked(m, flags); 3059 resid -= presid; 3060 poffset = 0; 3061 } 3062 VM_OBJECT_WUNLOCK(obj); 3063 bp->b_npages = 0; 3064 } 3065 3066 /* 3067 * Page-granular truncation of an existing VMIO buffer. 3068 */ 3069 static void 3070 vfs_vmio_truncate(struct buf *bp, int desiredpages) 3071 { 3072 vm_object_t obj; 3073 vm_page_t m; 3074 int flags, i; 3075 3076 if (bp->b_npages == desiredpages) 3077 return; 3078 3079 if (buf_mapped(bp)) { 3080 BUF_CHECK_MAPPED(bp); 3081 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) + 3082 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages); 3083 } else 3084 BUF_CHECK_UNMAPPED(bp); 3085 3086 /* 3087 * The object lock is needed only if we will attempt to free pages. 3088 */ 3089 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; 3090 if ((bp->b_flags & B_DIRECT) != 0) { 3091 flags |= VPR_TRYFREE; 3092 obj = bp->b_bufobj->bo_object; 3093 VM_OBJECT_WLOCK(obj); 3094 } else { 3095 obj = NULL; 3096 } 3097 for (i = desiredpages; i < bp->b_npages; i++) { 3098 m = bp->b_pages[i]; 3099 KASSERT(m != bogus_page, ("allocbuf: bogus page found")); 3100 bp->b_pages[i] = NULL; 3101 if (obj != NULL) 3102 vm_page_release_locked(m, flags); 3103 else 3104 vm_page_release(m, flags); 3105 } 3106 if (obj != NULL) 3107 VM_OBJECT_WUNLOCK(obj); 3108 bp->b_npages = desiredpages; 3109 } 3110 3111 /* 3112 * Byte granular extension of VMIO buffers. 3113 */ 3114 static void 3115 vfs_vmio_extend(struct buf *bp, int desiredpages, int size) 3116 { 3117 /* 3118 * We are growing the buffer, possibly in a 3119 * byte-granular fashion. 3120 */ 3121 vm_object_t obj; 3122 vm_offset_t toff; 3123 vm_offset_t tinc; 3124 vm_page_t m; 3125 3126 /* 3127 * Step 1, bring in the VM pages from the object, allocating 3128 * them if necessary. We must clear B_CACHE if these pages 3129 * are not valid for the range covered by the buffer. 3130 */ 3131 obj = bp->b_bufobj->bo_object; 3132 if (bp->b_npages < desiredpages) { 3133 KASSERT(desiredpages <= atop(maxbcachebuf), 3134 ("vfs_vmio_extend past maxbcachebuf %p %d %u", 3135 bp, desiredpages, maxbcachebuf)); 3136 3137 /* 3138 * We must allocate system pages since blocking 3139 * here could interfere with paging I/O, no 3140 * matter which process we are. 3141 * 3142 * Only exclusive busy can be tested here. 3143 * Blocking on shared busy might lead to 3144 * deadlocks once allocbuf() is called after 3145 * pages are vfs_busy_pages(). 3146 */ 3147 (void)vm_page_grab_pages_unlocked(obj, 3148 OFF_TO_IDX(bp->b_offset) + bp->b_npages, 3149 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY | 3150 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED, 3151 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages); 3152 bp->b_npages = desiredpages; 3153 } 3154 3155 /* 3156 * Step 2. We've loaded the pages into the buffer, 3157 * we have to figure out if we can still have B_CACHE 3158 * set. Note that B_CACHE is set according to the 3159 * byte-granular range ( bcount and size ), not the 3160 * aligned range ( newbsize ). 3161 * 3162 * The VM test is against m->valid, which is DEV_BSIZE 3163 * aligned. Needless to say, the validity of the data 3164 * needs to also be DEV_BSIZE aligned. Note that this 3165 * fails with NFS if the server or some other client 3166 * extends the file's EOF. If our buffer is resized, 3167 * B_CACHE may remain set! XXX 3168 */ 3169 toff = bp->b_bcount; 3170 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3171 while ((bp->b_flags & B_CACHE) && toff < size) { 3172 vm_pindex_t pi; 3173 3174 if (tinc > (size - toff)) 3175 tinc = size - toff; 3176 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; 3177 m = bp->b_pages[pi]; 3178 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m); 3179 toff += tinc; 3180 tinc = PAGE_SIZE; 3181 } 3182 3183 /* 3184 * Step 3, fixup the KVA pmap. 3185 */ 3186 if (buf_mapped(bp)) 3187 bpmap_qenter(bp); 3188 else 3189 BUF_CHECK_UNMAPPED(bp); 3190 } 3191 3192 /* 3193 * Check to see if a block at a particular lbn is available for a clustered 3194 * write. 3195 */ 3196 static int 3197 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 3198 { 3199 struct buf *bpa; 3200 int match; 3201 3202 match = 0; 3203 3204 /* If the buf isn't in core skip it */ 3205 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 3206 return (0); 3207 3208 /* If the buf is busy we don't want to wait for it */ 3209 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 3210 return (0); 3211 3212 /* Only cluster with valid clusterable delayed write buffers */ 3213 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 3214 (B_DELWRI | B_CLUSTEROK)) 3215 goto done; 3216 3217 if (bpa->b_bufsize != size) 3218 goto done; 3219 3220 /* 3221 * Check to see if it is in the expected place on disk and that the 3222 * block has been mapped. 3223 */ 3224 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 3225 match = 1; 3226 done: 3227 BUF_UNLOCK(bpa); 3228 return (match); 3229 } 3230 3231 /* 3232 * vfs_bio_awrite: 3233 * 3234 * Implement clustered async writes for clearing out B_DELWRI buffers. 3235 * This is much better then the old way of writing only one buffer at 3236 * a time. Note that we may not be presented with the buffers in the 3237 * correct order, so we search for the cluster in both directions. 3238 */ 3239 int 3240 vfs_bio_awrite(struct buf *bp) 3241 { 3242 struct bufobj *bo; 3243 int i; 3244 int j; 3245 daddr_t lblkno = bp->b_lblkno; 3246 struct vnode *vp = bp->b_vp; 3247 int ncl; 3248 int nwritten; 3249 int size; 3250 int maxcl; 3251 int gbflags; 3252 3253 bo = &vp->v_bufobj; 3254 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0; 3255 /* 3256 * right now we support clustered writing only to regular files. If 3257 * we find a clusterable block we could be in the middle of a cluster 3258 * rather then at the beginning. 3259 */ 3260 if ((vp->v_type == VREG) && 3261 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 3262 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 3263 size = vp->v_mount->mnt_stat.f_iosize; 3264 maxcl = maxphys / size; 3265 3266 BO_RLOCK(bo); 3267 for (i = 1; i < maxcl; i++) 3268 if (vfs_bio_clcheck(vp, size, lblkno + i, 3269 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 3270 break; 3271 3272 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 3273 if (vfs_bio_clcheck(vp, size, lblkno - j, 3274 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 3275 break; 3276 BO_RUNLOCK(bo); 3277 --j; 3278 ncl = i + j; 3279 /* 3280 * this is a possible cluster write 3281 */ 3282 if (ncl != 1) { 3283 BUF_UNLOCK(bp); 3284 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 3285 gbflags); 3286 return (nwritten); 3287 } 3288 } 3289 bremfree(bp); 3290 bp->b_flags |= B_ASYNC; 3291 /* 3292 * default (old) behavior, writing out only one block 3293 * 3294 * XXX returns b_bufsize instead of b_bcount for nwritten? 3295 */ 3296 nwritten = bp->b_bufsize; 3297 (void) bwrite(bp); 3298 3299 return (nwritten); 3300 } 3301 3302 /* 3303 * getnewbuf_kva: 3304 * 3305 * Allocate KVA for an empty buf header according to gbflags. 3306 */ 3307 static int 3308 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize) 3309 { 3310 3311 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) { 3312 /* 3313 * In order to keep fragmentation sane we only allocate kva 3314 * in BKVASIZE chunks. XXX with vmem we can do page size. 3315 */ 3316 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 3317 3318 if (maxsize != bp->b_kvasize && 3319 bufkva_alloc(bp, maxsize, gbflags)) 3320 return (ENOSPC); 3321 } 3322 return (0); 3323 } 3324 3325 /* 3326 * getnewbuf: 3327 * 3328 * Find and initialize a new buffer header, freeing up existing buffers 3329 * in the bufqueues as necessary. The new buffer is returned locked. 3330 * 3331 * We block if: 3332 * We have insufficient buffer headers 3333 * We have insufficient buffer space 3334 * buffer_arena is too fragmented ( space reservation fails ) 3335 * If we have to flush dirty buffers ( but we try to avoid this ) 3336 * 3337 * The caller is responsible for releasing the reserved bufspace after 3338 * allocbuf() is called. 3339 */ 3340 static struct buf * 3341 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags) 3342 { 3343 struct bufdomain *bd; 3344 struct buf *bp; 3345 bool metadata, reserved; 3346 3347 bp = NULL; 3348 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3349 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3350 if (!unmapped_buf_allowed) 3351 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3352 3353 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 3354 vp->v_type == VCHR) 3355 metadata = true; 3356 else 3357 metadata = false; 3358 if (vp == NULL) 3359 bd = &bdomain[0]; 3360 else 3361 bd = &bdomain[vp->v_bufobj.bo_domain]; 3362 3363 counter_u64_add(getnewbufcalls, 1); 3364 reserved = false; 3365 do { 3366 if (reserved == false && 3367 bufspace_reserve(bd, maxsize, metadata) != 0) { 3368 counter_u64_add(getnewbufrestarts, 1); 3369 continue; 3370 } 3371 reserved = true; 3372 if ((bp = buf_alloc(bd)) == NULL) { 3373 counter_u64_add(getnewbufrestarts, 1); 3374 continue; 3375 } 3376 if (getnewbuf_kva(bp, gbflags, maxsize) == 0) 3377 return (bp); 3378 break; 3379 } while (buf_recycle(bd, false) == 0); 3380 3381 if (reserved) 3382 bufspace_release(bd, maxsize); 3383 if (bp != NULL) { 3384 bp->b_flags |= B_INVAL; 3385 brelse(bp); 3386 } 3387 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo); 3388 3389 return (NULL); 3390 } 3391 3392 /* 3393 * buf_daemon: 3394 * 3395 * buffer flushing daemon. Buffers are normally flushed by the 3396 * update daemon but if it cannot keep up this process starts to 3397 * take the load in an attempt to prevent getnewbuf() from blocking. 3398 */ 3399 static struct kproc_desc buf_kp = { 3400 "bufdaemon", 3401 buf_daemon, 3402 &bufdaemonproc 3403 }; 3404 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 3405 3406 static int 3407 buf_flush(struct vnode *vp, struct bufdomain *bd, int target) 3408 { 3409 int flushed; 3410 3411 flushed = flushbufqueues(vp, bd, target, 0); 3412 if (flushed == 0) { 3413 /* 3414 * Could not find any buffers without rollback 3415 * dependencies, so just write the first one 3416 * in the hopes of eventually making progress. 3417 */ 3418 if (vp != NULL && target > 2) 3419 target /= 2; 3420 flushbufqueues(vp, bd, target, 1); 3421 } 3422 return (flushed); 3423 } 3424 3425 static void 3426 buf_daemon_shutdown(void *arg __unused, int howto __unused) 3427 { 3428 int error; 3429 3430 if (KERNEL_PANICKED()) 3431 return; 3432 3433 mtx_lock(&bdlock); 3434 bd_shutdown = true; 3435 wakeup(&bd_request); 3436 error = msleep(&bd_shutdown, &bdlock, 0, "buf_daemon_shutdown", 3437 60 * hz); 3438 mtx_unlock(&bdlock); 3439 if (error != 0) 3440 printf("bufdaemon wait error: %d\n", error); 3441 } 3442 3443 static void 3444 buf_daemon(void) 3445 { 3446 struct bufdomain *bd; 3447 int speedupreq; 3448 int lodirty; 3449 int i; 3450 3451 /* 3452 * This process needs to be suspended prior to shutdown sync. 3453 */ 3454 EVENTHANDLER_REGISTER(shutdown_pre_sync, buf_daemon_shutdown, NULL, 3455 SHUTDOWN_PRI_LAST + 100); 3456 3457 /* 3458 * Start the buf clean daemons as children threads. 3459 */ 3460 for (i = 0 ; i < buf_domains; i++) { 3461 int error; 3462 3463 error = kthread_add((void (*)(void *))bufspace_daemon, 3464 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i); 3465 if (error) 3466 panic("error %d spawning bufspace daemon", error); 3467 } 3468 3469 /* 3470 * This process is allowed to take the buffer cache to the limit 3471 */ 3472 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 3473 mtx_lock(&bdlock); 3474 while (!bd_shutdown) { 3475 bd_request = 0; 3476 mtx_unlock(&bdlock); 3477 3478 /* 3479 * Save speedupreq for this pass and reset to capture new 3480 * requests. 3481 */ 3482 speedupreq = bd_speedupreq; 3483 bd_speedupreq = 0; 3484 3485 /* 3486 * Flush each domain sequentially according to its level and 3487 * the speedup request. 3488 */ 3489 for (i = 0; i < buf_domains; i++) { 3490 bd = &bdomain[i]; 3491 if (speedupreq) 3492 lodirty = bd->bd_numdirtybuffers / 2; 3493 else 3494 lodirty = bd->bd_lodirtybuffers; 3495 while (bd->bd_numdirtybuffers > lodirty) { 3496 if (buf_flush(NULL, bd, 3497 bd->bd_numdirtybuffers - lodirty) == 0) 3498 break; 3499 kern_yield(PRI_USER); 3500 } 3501 } 3502 3503 /* 3504 * Only clear bd_request if we have reached our low water 3505 * mark. The buf_daemon normally waits 1 second and 3506 * then incrementally flushes any dirty buffers that have 3507 * built up, within reason. 3508 * 3509 * If we were unable to hit our low water mark and couldn't 3510 * find any flushable buffers, we sleep for a short period 3511 * to avoid endless loops on unlockable buffers. 3512 */ 3513 mtx_lock(&bdlock); 3514 if (bd_shutdown) 3515 break; 3516 if (BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) { 3517 /* 3518 * We reached our low water mark, reset the 3519 * request and sleep until we are needed again. 3520 * The sleep is just so the suspend code works. 3521 */ 3522 bd_request = 0; 3523 /* 3524 * Do an extra wakeup in case dirty threshold 3525 * changed via sysctl and the explicit transition 3526 * out of shortfall was missed. 3527 */ 3528 bdirtywakeup(); 3529 if (runningbufspace <= lorunningspace) 3530 runningwakeup(); 3531 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 3532 } else { 3533 /* 3534 * We couldn't find any flushable dirty buffers but 3535 * still have too many dirty buffers, we 3536 * have to sleep and try again. (rare) 3537 */ 3538 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 3539 } 3540 } 3541 wakeup(&bd_shutdown); 3542 mtx_unlock(&bdlock); 3543 kthread_exit(); 3544 } 3545 3546 /* 3547 * flushbufqueues: 3548 * 3549 * Try to flush a buffer in the dirty queue. We must be careful to 3550 * free up B_INVAL buffers instead of write them, which NFS is 3551 * particularly sensitive to. 3552 */ 3553 static int flushwithdeps = 0; 3554 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS, 3555 &flushwithdeps, 0, 3556 "Number of buffers flushed with dependencies that require rollbacks"); 3557 3558 static int 3559 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target, 3560 int flushdeps) 3561 { 3562 struct bufqueue *bq; 3563 struct buf *sentinel; 3564 struct vnode *vp; 3565 struct mount *mp; 3566 struct buf *bp; 3567 int hasdeps; 3568 int flushed; 3569 int error; 3570 bool unlock; 3571 3572 flushed = 0; 3573 bq = &bd->bd_dirtyq; 3574 bp = NULL; 3575 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 3576 sentinel->b_qindex = QUEUE_SENTINEL; 3577 BQ_LOCK(bq); 3578 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist); 3579 BQ_UNLOCK(bq); 3580 while (flushed != target) { 3581 maybe_yield(); 3582 BQ_LOCK(bq); 3583 bp = TAILQ_NEXT(sentinel, b_freelist); 3584 if (bp != NULL) { 3585 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); 3586 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel, 3587 b_freelist); 3588 } else { 3589 BQ_UNLOCK(bq); 3590 break; 3591 } 3592 /* 3593 * Skip sentinels inserted by other invocations of the 3594 * flushbufqueues(), taking care to not reorder them. 3595 * 3596 * Only flush the buffers that belong to the 3597 * vnode locked by the curthread. 3598 */ 3599 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && 3600 bp->b_vp != lvp)) { 3601 BQ_UNLOCK(bq); 3602 continue; 3603 } 3604 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 3605 BQ_UNLOCK(bq); 3606 if (error != 0) 3607 continue; 3608 3609 /* 3610 * BKGRDINPROG can only be set with the buf and bufobj 3611 * locks both held. We tolerate a race to clear it here. 3612 */ 3613 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 3614 (bp->b_flags & B_DELWRI) == 0) { 3615 BUF_UNLOCK(bp); 3616 continue; 3617 } 3618 if (bp->b_flags & B_INVAL) { 3619 bremfreef(bp); 3620 brelse(bp); 3621 flushed++; 3622 continue; 3623 } 3624 3625 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 3626 if (flushdeps == 0) { 3627 BUF_UNLOCK(bp); 3628 continue; 3629 } 3630 hasdeps = 1; 3631 } else 3632 hasdeps = 0; 3633 /* 3634 * We must hold the lock on a vnode before writing 3635 * one of its buffers. Otherwise we may confuse, or 3636 * in the case of a snapshot vnode, deadlock the 3637 * system. 3638 * 3639 * The lock order here is the reverse of the normal 3640 * of vnode followed by buf lock. This is ok because 3641 * the NOWAIT will prevent deadlock. 3642 */ 3643 vp = bp->b_vp; 3644 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 3645 BUF_UNLOCK(bp); 3646 continue; 3647 } 3648 if (lvp == NULL) { 3649 unlock = true; 3650 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 3651 } else { 3652 ASSERT_VOP_LOCKED(vp, "getbuf"); 3653 unlock = false; 3654 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 3655 vn_lock(vp, LK_TRYUPGRADE); 3656 } 3657 if (error == 0) { 3658 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 3659 bp, bp->b_vp, bp->b_flags); 3660 if (curproc == bufdaemonproc) { 3661 vfs_bio_awrite(bp); 3662 } else { 3663 bremfree(bp); 3664 bwrite(bp); 3665 counter_u64_add(notbufdflushes, 1); 3666 } 3667 vn_finished_write(mp); 3668 if (unlock) 3669 VOP_UNLOCK(vp); 3670 flushwithdeps += hasdeps; 3671 flushed++; 3672 3673 /* 3674 * Sleeping on runningbufspace while holding 3675 * vnode lock leads to deadlock. 3676 */ 3677 if (curproc == bufdaemonproc && 3678 runningbufspace > hirunningspace) 3679 waitrunningbufspace(); 3680 continue; 3681 } 3682 vn_finished_write(mp); 3683 BUF_UNLOCK(bp); 3684 } 3685 BQ_LOCK(bq); 3686 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); 3687 BQ_UNLOCK(bq); 3688 free(sentinel, M_TEMP); 3689 return (flushed); 3690 } 3691 3692 /* 3693 * Check to see if a block is currently memory resident. 3694 */ 3695 struct buf * 3696 incore(struct bufobj *bo, daddr_t blkno) 3697 { 3698 return (gbincore_unlocked(bo, blkno)); 3699 } 3700 3701 /* 3702 * Returns true if no I/O is needed to access the 3703 * associated VM object. This is like incore except 3704 * it also hunts around in the VM system for the data. 3705 */ 3706 bool 3707 inmem(struct vnode * vp, daddr_t blkno) 3708 { 3709 vm_object_t obj; 3710 vm_offset_t toff, tinc, size; 3711 vm_page_t m, n; 3712 vm_ooffset_t off; 3713 int valid; 3714 3715 ASSERT_VOP_LOCKED(vp, "inmem"); 3716 3717 if (incore(&vp->v_bufobj, blkno)) 3718 return (true); 3719 if (vp->v_mount == NULL) 3720 return (false); 3721 obj = vp->v_object; 3722 if (obj == NULL) 3723 return (false); 3724 3725 size = PAGE_SIZE; 3726 if (size > vp->v_mount->mnt_stat.f_iosize) 3727 size = vp->v_mount->mnt_stat.f_iosize; 3728 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 3729 3730 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 3731 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff)); 3732 recheck: 3733 if (m == NULL) 3734 return (false); 3735 3736 tinc = size; 3737 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 3738 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 3739 /* 3740 * Consider page validity only if page mapping didn't change 3741 * during the check. 3742 */ 3743 valid = vm_page_is_valid(m, 3744 (vm_offset_t)((toff + off) & PAGE_MASK), tinc); 3745 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff)); 3746 if (m != n) { 3747 m = n; 3748 goto recheck; 3749 } 3750 if (!valid) 3751 return (false); 3752 } 3753 return (true); 3754 } 3755 3756 /* 3757 * Set the dirty range for a buffer based on the status of the dirty 3758 * bits in the pages comprising the buffer. The range is limited 3759 * to the size of the buffer. 3760 * 3761 * Tell the VM system that the pages associated with this buffer 3762 * are clean. This is used for delayed writes where the data is 3763 * going to go to disk eventually without additional VM intevention. 3764 * 3765 * Note that while we only really need to clean through to b_bcount, we 3766 * just go ahead and clean through to b_bufsize. 3767 */ 3768 static void 3769 vfs_clean_pages_dirty_buf(struct buf *bp) 3770 { 3771 vm_ooffset_t foff, noff, eoff; 3772 vm_page_t m; 3773 int i; 3774 3775 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 3776 return; 3777 3778 foff = bp->b_offset; 3779 KASSERT(bp->b_offset != NOOFFSET, 3780 ("vfs_clean_pages_dirty_buf: no buffer offset")); 3781 3782 vfs_busy_pages_acquire(bp); 3783 vfs_setdirty_range(bp); 3784 for (i = 0; i < bp->b_npages; i++) { 3785 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3786 eoff = noff; 3787 if (eoff > bp->b_offset + bp->b_bufsize) 3788 eoff = bp->b_offset + bp->b_bufsize; 3789 m = bp->b_pages[i]; 3790 vfs_page_set_validclean(bp, foff, m); 3791 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3792 foff = noff; 3793 } 3794 vfs_busy_pages_release(bp); 3795 } 3796 3797 static void 3798 vfs_setdirty_range(struct buf *bp) 3799 { 3800 vm_offset_t boffset; 3801 vm_offset_t eoffset; 3802 int i; 3803 3804 /* 3805 * test the pages to see if they have been modified directly 3806 * by users through the VM system. 3807 */ 3808 for (i = 0; i < bp->b_npages; i++) 3809 vm_page_test_dirty(bp->b_pages[i]); 3810 3811 /* 3812 * Calculate the encompassing dirty range, boffset and eoffset, 3813 * (eoffset - boffset) bytes. 3814 */ 3815 3816 for (i = 0; i < bp->b_npages; i++) { 3817 if (bp->b_pages[i]->dirty) 3818 break; 3819 } 3820 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3821 3822 for (i = bp->b_npages - 1; i >= 0; --i) { 3823 if (bp->b_pages[i]->dirty) { 3824 break; 3825 } 3826 } 3827 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3828 3829 /* 3830 * Fit it to the buffer. 3831 */ 3832 3833 if (eoffset > bp->b_bcount) 3834 eoffset = bp->b_bcount; 3835 3836 /* 3837 * If we have a good dirty range, merge with the existing 3838 * dirty range. 3839 */ 3840 3841 if (boffset < eoffset) { 3842 if (bp->b_dirtyoff > boffset) 3843 bp->b_dirtyoff = boffset; 3844 if (bp->b_dirtyend < eoffset) 3845 bp->b_dirtyend = eoffset; 3846 } 3847 } 3848 3849 /* 3850 * Allocate the KVA mapping for an existing buffer. 3851 * If an unmapped buffer is provided but a mapped buffer is requested, take 3852 * also care to properly setup mappings between pages and KVA. 3853 */ 3854 static void 3855 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 3856 { 3857 int bsize, maxsize, need_mapping, need_kva; 3858 off_t offset; 3859 3860 need_mapping = bp->b_data == unmapped_buf && 3861 (gbflags & GB_UNMAPPED) == 0; 3862 need_kva = bp->b_kvabase == unmapped_buf && 3863 bp->b_data == unmapped_buf && 3864 (gbflags & GB_KVAALLOC) != 0; 3865 if (!need_mapping && !need_kva) 3866 return; 3867 3868 BUF_CHECK_UNMAPPED(bp); 3869 3870 if (need_mapping && bp->b_kvabase != unmapped_buf) { 3871 /* 3872 * Buffer is not mapped, but the KVA was already 3873 * reserved at the time of the instantiation. Use the 3874 * allocated space. 3875 */ 3876 goto has_addr; 3877 } 3878 3879 /* 3880 * Calculate the amount of the address space we would reserve 3881 * if the buffer was mapped. 3882 */ 3883 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 3884 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3885 offset = blkno * bsize; 3886 maxsize = size + (offset & PAGE_MASK); 3887 maxsize = imax(maxsize, bsize); 3888 3889 while (bufkva_alloc(bp, maxsize, gbflags) != 0) { 3890 if ((gbflags & GB_NOWAIT_BD) != 0) { 3891 /* 3892 * XXXKIB: defragmentation cannot 3893 * succeed, not sure what else to do. 3894 */ 3895 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp); 3896 } 3897 counter_u64_add(mappingrestarts, 1); 3898 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0); 3899 } 3900 has_addr: 3901 if (need_mapping) { 3902 /* b_offset is handled by bpmap_qenter. */ 3903 bp->b_data = bp->b_kvabase; 3904 BUF_CHECK_MAPPED(bp); 3905 bpmap_qenter(bp); 3906 } 3907 } 3908 3909 struct buf * 3910 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3911 int flags) 3912 { 3913 struct buf *bp; 3914 int error; 3915 3916 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp); 3917 if (error != 0) 3918 return (NULL); 3919 return (bp); 3920 } 3921 3922 /* 3923 * getblkx: 3924 * 3925 * Get a block given a specified block and offset into a file/device. 3926 * The buffers B_DONE bit will be cleared on return, making it almost 3927 * ready for an I/O initiation. B_INVAL may or may not be set on 3928 * return. The caller should clear B_INVAL prior to initiating a 3929 * READ. 3930 * 3931 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3932 * an existing buffer. 3933 * 3934 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3935 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3936 * and then cleared based on the backing VM. If the previous buffer is 3937 * non-0-sized but invalid, B_CACHE will be cleared. 3938 * 3939 * If getblk() must create a new buffer, the new buffer is returned with 3940 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3941 * case it is returned with B_INVAL clear and B_CACHE set based on the 3942 * backing VM. 3943 * 3944 * getblk() also forces a bwrite() for any B_DELWRI buffer whose 3945 * B_CACHE bit is clear. 3946 * 3947 * What this means, basically, is that the caller should use B_CACHE to 3948 * determine whether the buffer is fully valid or not and should clear 3949 * B_INVAL prior to issuing a read. If the caller intends to validate 3950 * the buffer by loading its data area with something, the caller needs 3951 * to clear B_INVAL. If the caller does this without issuing an I/O, 3952 * the caller should set B_CACHE ( as an optimization ), else the caller 3953 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3954 * a write attempt or if it was a successful read. If the caller 3955 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3956 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3957 * 3958 * The blkno parameter is the logical block being requested. Normally 3959 * the mapping of logical block number to disk block address is done 3960 * by calling VOP_BMAP(). However, if the mapping is already known, the 3961 * disk block address can be passed using the dblkno parameter. If the 3962 * disk block address is not known, then the same value should be passed 3963 * for blkno and dblkno. 3964 */ 3965 int 3966 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag, 3967 int slptimeo, int flags, struct buf **bpp) 3968 { 3969 struct buf *bp; 3970 struct bufobj *bo; 3971 daddr_t d_blkno; 3972 int bsize, error, maxsize, vmio; 3973 off_t offset; 3974 3975 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3976 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3977 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3978 if (vp->v_type != VCHR) 3979 ASSERT_VOP_LOCKED(vp, "getblk"); 3980 if (size > maxbcachebuf) 3981 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size, 3982 maxbcachebuf); 3983 if (!unmapped_buf_allowed) 3984 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3985 3986 bo = &vp->v_bufobj; 3987 d_blkno = dblkno; 3988 3989 /* Attempt lockless lookup first. */ 3990 bp = gbincore_unlocked(bo, blkno); 3991 if (bp == NULL) { 3992 /* 3993 * With GB_NOCREAT we must be sure about not finding the buffer 3994 * as it may have been reassigned during unlocked lookup. 3995 */ 3996 if ((flags & GB_NOCREAT) != 0) 3997 goto loop; 3998 goto newbuf_unlocked; 3999 } 4000 4001 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0, 4002 0); 4003 if (error != 0) 4004 goto loop; 4005 4006 /* Verify buf identify has not changed since lookup. */ 4007 if (bp->b_bufobj == bo && bp->b_lblkno == blkno) 4008 goto foundbuf_fastpath; 4009 4010 /* It changed, fallback to locked lookup. */ 4011 BUF_UNLOCK_RAW(bp); 4012 4013 loop: 4014 BO_RLOCK(bo); 4015 bp = gbincore(bo, blkno); 4016 if (bp != NULL) { 4017 int lockflags; 4018 4019 /* 4020 * Buffer is in-core. If the buffer is not busy nor managed, 4021 * it must be on a queue. 4022 */ 4023 lockflags = LK_EXCLUSIVE | LK_INTERLOCK | 4024 ((flags & GB_LOCK_NOWAIT) != 0 ? LK_NOWAIT : LK_SLEEPFAIL); 4025 #ifdef WITNESS 4026 lockflags |= (flags & GB_NOWITNESS) != 0 ? LK_NOWITNESS : 0; 4027 #endif 4028 4029 error = BUF_TIMELOCK(bp, lockflags, 4030 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 4031 4032 /* 4033 * If we slept and got the lock we have to restart in case 4034 * the buffer changed identities. 4035 */ 4036 if (error == ENOLCK) 4037 goto loop; 4038 /* We timed out or were interrupted. */ 4039 else if (error != 0) 4040 return (error); 4041 4042 foundbuf_fastpath: 4043 /* If recursed, assume caller knows the rules. */ 4044 if (BUF_LOCKRECURSED(bp)) 4045 goto end; 4046 4047 /* 4048 * The buffer is locked. B_CACHE is cleared if the buffer is 4049 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 4050 * and for a VMIO buffer B_CACHE is adjusted according to the 4051 * backing VM cache. 4052 */ 4053 if (bp->b_flags & B_INVAL) 4054 bp->b_flags &= ~B_CACHE; 4055 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 4056 bp->b_flags |= B_CACHE; 4057 if (bp->b_flags & B_MANAGED) 4058 MPASS(bp->b_qindex == QUEUE_NONE); 4059 else 4060 bremfree(bp); 4061 4062 /* 4063 * check for size inconsistencies for non-VMIO case. 4064 */ 4065 if (bp->b_bcount != size) { 4066 if ((bp->b_flags & B_VMIO) == 0 || 4067 (size > bp->b_kvasize)) { 4068 if (bp->b_flags & B_DELWRI) { 4069 bp->b_flags |= B_NOCACHE; 4070 bwrite(bp); 4071 } else { 4072 if (LIST_EMPTY(&bp->b_dep)) { 4073 bp->b_flags |= B_RELBUF; 4074 brelse(bp); 4075 } else { 4076 bp->b_flags |= B_NOCACHE; 4077 bwrite(bp); 4078 } 4079 } 4080 goto loop; 4081 } 4082 } 4083 4084 /* 4085 * Handle the case of unmapped buffer which should 4086 * become mapped, or the buffer for which KVA 4087 * reservation is requested. 4088 */ 4089 bp_unmapped_get_kva(bp, blkno, size, flags); 4090 4091 /* 4092 * If the size is inconsistent in the VMIO case, we can resize 4093 * the buffer. This might lead to B_CACHE getting set or 4094 * cleared. If the size has not changed, B_CACHE remains 4095 * unchanged from its previous state. 4096 */ 4097 allocbuf(bp, size); 4098 4099 KASSERT(bp->b_offset != NOOFFSET, 4100 ("getblk: no buffer offset")); 4101 4102 /* 4103 * A buffer with B_DELWRI set and B_CACHE clear must 4104 * be committed before we can return the buffer in 4105 * order to prevent the caller from issuing a read 4106 * ( due to B_CACHE not being set ) and overwriting 4107 * it. 4108 * 4109 * Most callers, including NFS and FFS, need this to 4110 * operate properly either because they assume they 4111 * can issue a read if B_CACHE is not set, or because 4112 * ( for example ) an uncached B_DELWRI might loop due 4113 * to softupdates re-dirtying the buffer. In the latter 4114 * case, B_CACHE is set after the first write completes, 4115 * preventing further loops. 4116 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 4117 * above while extending the buffer, we cannot allow the 4118 * buffer to remain with B_CACHE set after the write 4119 * completes or it will represent a corrupt state. To 4120 * deal with this we set B_NOCACHE to scrap the buffer 4121 * after the write. 4122 * 4123 * We might be able to do something fancy, like setting 4124 * B_CACHE in bwrite() except if B_DELWRI is already set, 4125 * so the below call doesn't set B_CACHE, but that gets real 4126 * confusing. This is much easier. 4127 */ 4128 4129 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 4130 bp->b_flags |= B_NOCACHE; 4131 bwrite(bp); 4132 goto loop; 4133 } 4134 bp->b_flags &= ~B_DONE; 4135 } else { 4136 /* 4137 * Buffer is not in-core, create new buffer. The buffer 4138 * returned by getnewbuf() is locked. Note that the returned 4139 * buffer is also considered valid (not marked B_INVAL). 4140 */ 4141 BO_RUNLOCK(bo); 4142 newbuf_unlocked: 4143 /* 4144 * If the user does not want us to create the buffer, bail out 4145 * here. 4146 */ 4147 if (flags & GB_NOCREAT) 4148 return (EEXIST); 4149 4150 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize; 4151 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 4152 offset = blkno * bsize; 4153 vmio = vp->v_object != NULL; 4154 if (vmio) { 4155 maxsize = size + (offset & PAGE_MASK); 4156 } else { 4157 maxsize = size; 4158 /* Do not allow non-VMIO notmapped buffers. */ 4159 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 4160 } 4161 maxsize = imax(maxsize, bsize); 4162 if ((flags & GB_NOSPARSE) != 0 && vmio && 4163 !vn_isdisk(vp)) { 4164 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0); 4165 KASSERT(error != EOPNOTSUPP, 4166 ("GB_NOSPARSE from fs not supporting bmap, vp %p", 4167 vp)); 4168 if (error != 0) 4169 return (error); 4170 if (d_blkno == -1) 4171 return (EJUSTRETURN); 4172 } 4173 4174 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags); 4175 if (bp == NULL) { 4176 if (slpflag || slptimeo) 4177 return (ETIMEDOUT); 4178 /* 4179 * XXX This is here until the sleep path is diagnosed 4180 * enough to work under very low memory conditions. 4181 * 4182 * There's an issue on low memory, 4BSD+non-preempt 4183 * systems (eg MIPS routers with 32MB RAM) where buffer 4184 * exhaustion occurs without sleeping for buffer 4185 * reclaimation. This just sticks in a loop and 4186 * constantly attempts to allocate a buffer, which 4187 * hits exhaustion and tries to wakeup bufdaemon. 4188 * This never happens because we never yield. 4189 * 4190 * The real solution is to identify and fix these cases 4191 * so we aren't effectively busy-waiting in a loop 4192 * until the reclaimation path has cycles to run. 4193 */ 4194 kern_yield(PRI_USER); 4195 goto loop; 4196 } 4197 4198 /* 4199 * This code is used to make sure that a buffer is not 4200 * created while the getnewbuf routine is blocked. 4201 * This can be a problem whether the vnode is locked or not. 4202 * If the buffer is created out from under us, we have to 4203 * throw away the one we just created. 4204 * 4205 * Note: this must occur before we associate the buffer 4206 * with the vp especially considering limitations in 4207 * the splay tree implementation when dealing with duplicate 4208 * lblkno's. 4209 */ 4210 BO_LOCK(bo); 4211 if (gbincore(bo, blkno)) { 4212 BO_UNLOCK(bo); 4213 bp->b_flags |= B_INVAL; 4214 bufspace_release(bufdomain(bp), maxsize); 4215 brelse(bp); 4216 goto loop; 4217 } 4218 4219 /* 4220 * Insert the buffer into the hash, so that it can 4221 * be found by incore. 4222 */ 4223 bp->b_lblkno = blkno; 4224 bp->b_blkno = d_blkno; 4225 bp->b_offset = offset; 4226 bgetvp(vp, bp); 4227 BO_UNLOCK(bo); 4228 4229 /* 4230 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 4231 * buffer size starts out as 0, B_CACHE will be set by 4232 * allocbuf() for the VMIO case prior to it testing the 4233 * backing store for validity. 4234 */ 4235 4236 if (vmio) { 4237 bp->b_flags |= B_VMIO; 4238 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 4239 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 4240 bp, vp->v_object, bp->b_bufobj->bo_object)); 4241 } else { 4242 bp->b_flags &= ~B_VMIO; 4243 KASSERT(bp->b_bufobj->bo_object == NULL, 4244 ("ARGH! has b_bufobj->bo_object %p %p\n", 4245 bp, bp->b_bufobj->bo_object)); 4246 BUF_CHECK_MAPPED(bp); 4247 } 4248 4249 allocbuf(bp, size); 4250 bufspace_release(bufdomain(bp), maxsize); 4251 bp->b_flags &= ~B_DONE; 4252 } 4253 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 4254 end: 4255 buf_track(bp, __func__); 4256 KASSERT(bp->b_bufobj == bo, 4257 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 4258 *bpp = bp; 4259 return (0); 4260 } 4261 4262 /* 4263 * Get an empty, disassociated buffer of given size. The buffer is initially 4264 * set to B_INVAL. 4265 */ 4266 struct buf * 4267 geteblk(int size, int flags) 4268 { 4269 struct buf *bp; 4270 int maxsize; 4271 4272 maxsize = (size + BKVAMASK) & ~BKVAMASK; 4273 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) { 4274 if ((flags & GB_NOWAIT_BD) && 4275 (curthread->td_pflags & TDP_BUFNEED) != 0) 4276 return (NULL); 4277 } 4278 allocbuf(bp, size); 4279 bufspace_release(bufdomain(bp), maxsize); 4280 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 4281 return (bp); 4282 } 4283 4284 /* 4285 * Truncate the backing store for a non-vmio buffer. 4286 */ 4287 static void 4288 vfs_nonvmio_truncate(struct buf *bp, int newbsize) 4289 { 4290 4291 if (bp->b_flags & B_MALLOC) { 4292 /* 4293 * malloced buffers are not shrunk 4294 */ 4295 if (newbsize == 0) { 4296 bufmallocadjust(bp, 0); 4297 free(bp->b_data, M_BIOBUF); 4298 bp->b_data = bp->b_kvabase; 4299 bp->b_flags &= ~B_MALLOC; 4300 } 4301 return; 4302 } 4303 vm_hold_free_pages(bp, newbsize); 4304 bufspace_adjust(bp, newbsize); 4305 } 4306 4307 /* 4308 * Extend the backing for a non-VMIO buffer. 4309 */ 4310 static void 4311 vfs_nonvmio_extend(struct buf *bp, int newbsize) 4312 { 4313 caddr_t origbuf; 4314 int origbufsize; 4315 4316 /* 4317 * We only use malloced memory on the first allocation. 4318 * and revert to page-allocated memory when the buffer 4319 * grows. 4320 * 4321 * There is a potential smp race here that could lead 4322 * to bufmallocspace slightly passing the max. It 4323 * is probably extremely rare and not worth worrying 4324 * over. 4325 */ 4326 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 && 4327 bufmallocspace < maxbufmallocspace) { 4328 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK); 4329 bp->b_flags |= B_MALLOC; 4330 bufmallocadjust(bp, newbsize); 4331 return; 4332 } 4333 4334 /* 4335 * If the buffer is growing on its other-than-first 4336 * allocation then we revert to the page-allocation 4337 * scheme. 4338 */ 4339 origbuf = NULL; 4340 origbufsize = 0; 4341 if (bp->b_flags & B_MALLOC) { 4342 origbuf = bp->b_data; 4343 origbufsize = bp->b_bufsize; 4344 bp->b_data = bp->b_kvabase; 4345 bufmallocadjust(bp, 0); 4346 bp->b_flags &= ~B_MALLOC; 4347 newbsize = round_page(newbsize); 4348 } 4349 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize, 4350 (vm_offset_t) bp->b_data + newbsize); 4351 if (origbuf != NULL) { 4352 bcopy(origbuf, bp->b_data, origbufsize); 4353 free(origbuf, M_BIOBUF); 4354 } 4355 bufspace_adjust(bp, newbsize); 4356 } 4357 4358 /* 4359 * This code constitutes the buffer memory from either anonymous system 4360 * memory (in the case of non-VMIO operations) or from an associated 4361 * VM object (in the case of VMIO operations). This code is able to 4362 * resize a buffer up or down. 4363 * 4364 * Note that this code is tricky, and has many complications to resolve 4365 * deadlock or inconsistent data situations. Tread lightly!!! 4366 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 4367 * the caller. Calling this code willy nilly can result in the loss of data. 4368 * 4369 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 4370 * B_CACHE for the non-VMIO case. 4371 */ 4372 int 4373 allocbuf(struct buf *bp, int size) 4374 { 4375 int newbsize; 4376 4377 if (bp->b_bcount == size) 4378 return (1); 4379 4380 KASSERT(bp->b_kvasize == 0 || bp->b_kvasize >= size, 4381 ("allocbuf: buffer too small %p %#x %#x", 4382 bp, bp->b_kvasize, size)); 4383 4384 newbsize = roundup2(size, DEV_BSIZE); 4385 if ((bp->b_flags & B_VMIO) == 0) { 4386 if ((bp->b_flags & B_MALLOC) == 0) 4387 newbsize = round_page(newbsize); 4388 /* 4389 * Just get anonymous memory from the kernel. Don't 4390 * mess with B_CACHE. 4391 */ 4392 if (newbsize < bp->b_bufsize) 4393 vfs_nonvmio_truncate(bp, newbsize); 4394 else if (newbsize > bp->b_bufsize) 4395 vfs_nonvmio_extend(bp, newbsize); 4396 } else { 4397 int desiredpages; 4398 4399 desiredpages = size == 0 ? 0 : 4400 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 4401 4402 KASSERT((bp->b_flags & B_MALLOC) == 0, 4403 ("allocbuf: VMIO buffer can't be malloced %p", bp)); 4404 4405 /* 4406 * Set B_CACHE initially if buffer is 0 length or will become 4407 * 0-length. 4408 */ 4409 if (size == 0 || bp->b_bufsize == 0) 4410 bp->b_flags |= B_CACHE; 4411 4412 if (newbsize < bp->b_bufsize) 4413 vfs_vmio_truncate(bp, desiredpages); 4414 /* XXX This looks as if it should be newbsize > b_bufsize */ 4415 else if (size > bp->b_bcount) 4416 vfs_vmio_extend(bp, desiredpages, size); 4417 bufspace_adjust(bp, newbsize); 4418 } 4419 bp->b_bcount = size; /* requested buffer size. */ 4420 return (1); 4421 } 4422 4423 extern int inflight_transient_maps; 4424 4425 static struct bio_queue nondump_bios; 4426 4427 void 4428 biodone(struct bio *bp) 4429 { 4430 struct mtx *mtxp; 4431 void (*done)(struct bio *); 4432 vm_offset_t start, end; 4433 4434 biotrack(bp, __func__); 4435 4436 /* 4437 * Avoid completing I/O when dumping after a panic since that may 4438 * result in a deadlock in the filesystem or pager code. Note that 4439 * this doesn't affect dumps that were started manually since we aim 4440 * to keep the system usable after it has been resumed. 4441 */ 4442 if (__predict_false(dumping && SCHEDULER_STOPPED())) { 4443 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue); 4444 return; 4445 } 4446 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 4447 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 4448 bp->bio_flags |= BIO_UNMAPPED; 4449 start = trunc_page((vm_offset_t)bp->bio_data); 4450 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 4451 bp->bio_data = unmapped_buf; 4452 pmap_qremove(start, atop(end - start)); 4453 vmem_free(transient_arena, start, end - start); 4454 atomic_add_int(&inflight_transient_maps, -1); 4455 } 4456 done = bp->bio_done; 4457 /* 4458 * The check for done == biodone is to allow biodone to be 4459 * used as a bio_done routine. 4460 */ 4461 if (done == NULL || done == biodone) { 4462 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4463 mtx_lock(mtxp); 4464 bp->bio_flags |= BIO_DONE; 4465 wakeup(bp); 4466 mtx_unlock(mtxp); 4467 } else 4468 done(bp); 4469 } 4470 4471 /* 4472 * Wait for a BIO to finish. 4473 */ 4474 int 4475 biowait(struct bio *bp, const char *wmesg) 4476 { 4477 struct mtx *mtxp; 4478 4479 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4480 mtx_lock(mtxp); 4481 while ((bp->bio_flags & BIO_DONE) == 0) 4482 msleep(bp, mtxp, PRIBIO, wmesg, 0); 4483 mtx_unlock(mtxp); 4484 if (bp->bio_error != 0) 4485 return (bp->bio_error); 4486 if (!(bp->bio_flags & BIO_ERROR)) 4487 return (0); 4488 return (EIO); 4489 } 4490 4491 void 4492 biofinish(struct bio *bp, struct devstat *stat, int error) 4493 { 4494 4495 if (error) { 4496 bp->bio_error = error; 4497 bp->bio_flags |= BIO_ERROR; 4498 } 4499 if (stat != NULL) 4500 devstat_end_transaction_bio(stat, bp); 4501 biodone(bp); 4502 } 4503 4504 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING) 4505 void 4506 biotrack_buf(struct bio *bp, const char *location) 4507 { 4508 4509 buf_track(bp->bio_track_bp, location); 4510 } 4511 #endif 4512 4513 /* 4514 * bufwait: 4515 * 4516 * Wait for buffer I/O completion, returning error status. The buffer 4517 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 4518 * error and cleared. 4519 */ 4520 int 4521 bufwait(struct buf *bp) 4522 { 4523 if (bp->b_iocmd == BIO_READ) 4524 bwait(bp, PRIBIO, "biord"); 4525 else 4526 bwait(bp, PRIBIO, "biowr"); 4527 if (bp->b_flags & B_EINTR) { 4528 bp->b_flags &= ~B_EINTR; 4529 return (EINTR); 4530 } 4531 if (bp->b_ioflags & BIO_ERROR) { 4532 return (bp->b_error ? bp->b_error : EIO); 4533 } else { 4534 return (0); 4535 } 4536 } 4537 4538 /* 4539 * bufdone: 4540 * 4541 * Finish I/O on a buffer, optionally calling a completion function. 4542 * This is usually called from an interrupt so process blocking is 4543 * not allowed. 4544 * 4545 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 4546 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 4547 * assuming B_INVAL is clear. 4548 * 4549 * For the VMIO case, we set B_CACHE if the op was a read and no 4550 * read error occurred, or if the op was a write. B_CACHE is never 4551 * set if the buffer is invalid or otherwise uncacheable. 4552 * 4553 * bufdone does not mess with B_INVAL, allowing the I/O routine or the 4554 * initiator to leave B_INVAL set to brelse the buffer out of existence 4555 * in the biodone routine. 4556 */ 4557 void 4558 bufdone(struct buf *bp) 4559 { 4560 struct bufobj *dropobj; 4561 void (*biodone)(struct buf *); 4562 4563 buf_track(bp, __func__); 4564 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 4565 dropobj = NULL; 4566 4567 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 4568 4569 runningbufwakeup(bp); 4570 if (bp->b_iocmd == BIO_WRITE) 4571 dropobj = bp->b_bufobj; 4572 /* call optional completion function if requested */ 4573 if (bp->b_iodone != NULL) { 4574 biodone = bp->b_iodone; 4575 bp->b_iodone = NULL; 4576 (*biodone) (bp); 4577 if (dropobj) 4578 bufobj_wdrop(dropobj); 4579 return; 4580 } 4581 if (bp->b_flags & B_VMIO) { 4582 /* 4583 * Set B_CACHE if the op was a normal read and no error 4584 * occurred. B_CACHE is set for writes in the b*write() 4585 * routines. 4586 */ 4587 if (bp->b_iocmd == BIO_READ && 4588 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 4589 !(bp->b_ioflags & BIO_ERROR)) 4590 bp->b_flags |= B_CACHE; 4591 vfs_vmio_iodone(bp); 4592 } 4593 if (!LIST_EMPTY(&bp->b_dep)) 4594 buf_complete(bp); 4595 if ((bp->b_flags & B_CKHASH) != 0) { 4596 KASSERT(bp->b_iocmd == BIO_READ, 4597 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd)); 4598 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp)); 4599 (*bp->b_ckhashcalc)(bp); 4600 } 4601 /* 4602 * For asynchronous completions, release the buffer now. The brelse 4603 * will do a wakeup there if necessary - so no need to do a wakeup 4604 * here in the async case. The sync case always needs to do a wakeup. 4605 */ 4606 if (bp->b_flags & B_ASYNC) { 4607 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || 4608 (bp->b_ioflags & BIO_ERROR)) 4609 brelse(bp); 4610 else 4611 bqrelse(bp); 4612 } else 4613 bdone(bp); 4614 if (dropobj) 4615 bufobj_wdrop(dropobj); 4616 } 4617 4618 /* 4619 * This routine is called in lieu of iodone in the case of 4620 * incomplete I/O. This keeps the busy status for pages 4621 * consistent. 4622 */ 4623 void 4624 vfs_unbusy_pages(struct buf *bp) 4625 { 4626 int i; 4627 vm_object_t obj; 4628 vm_page_t m; 4629 4630 runningbufwakeup(bp); 4631 if (!(bp->b_flags & B_VMIO)) 4632 return; 4633 4634 obj = bp->b_bufobj->bo_object; 4635 for (i = 0; i < bp->b_npages; i++) { 4636 m = bp->b_pages[i]; 4637 if (m == bogus_page) { 4638 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i); 4639 if (!m) 4640 panic("vfs_unbusy_pages: page missing\n"); 4641 bp->b_pages[i] = m; 4642 if (buf_mapped(bp)) { 4643 BUF_CHECK_MAPPED(bp); 4644 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4645 bp->b_pages, bp->b_npages); 4646 } else 4647 BUF_CHECK_UNMAPPED(bp); 4648 } 4649 vm_page_sunbusy(m); 4650 } 4651 vm_object_pip_wakeupn(obj, bp->b_npages); 4652 } 4653 4654 /* 4655 * vfs_page_set_valid: 4656 * 4657 * Set the valid bits in a page based on the supplied offset. The 4658 * range is restricted to the buffer's size. 4659 * 4660 * This routine is typically called after a read completes. 4661 */ 4662 static void 4663 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4664 { 4665 vm_ooffset_t eoff; 4666 4667 /* 4668 * Compute the end offset, eoff, such that [off, eoff) does not span a 4669 * page boundary and eoff is not greater than the end of the buffer. 4670 * The end of the buffer, in this case, is our file EOF, not the 4671 * allocation size of the buffer. 4672 */ 4673 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 4674 if (eoff > bp->b_offset + bp->b_bcount) 4675 eoff = bp->b_offset + bp->b_bcount; 4676 4677 /* 4678 * Set valid range. This is typically the entire buffer and thus the 4679 * entire page. 4680 */ 4681 if (eoff > off) 4682 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 4683 } 4684 4685 /* 4686 * vfs_page_set_validclean: 4687 * 4688 * Set the valid bits and clear the dirty bits in a page based on the 4689 * supplied offset. The range is restricted to the buffer's size. 4690 */ 4691 static void 4692 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4693 { 4694 vm_ooffset_t soff, eoff; 4695 4696 /* 4697 * Start and end offsets in buffer. eoff - soff may not cross a 4698 * page boundary or cross the end of the buffer. The end of the 4699 * buffer, in this case, is our file EOF, not the allocation size 4700 * of the buffer. 4701 */ 4702 soff = off; 4703 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4704 if (eoff > bp->b_offset + bp->b_bcount) 4705 eoff = bp->b_offset + bp->b_bcount; 4706 4707 /* 4708 * Set valid range. This is typically the entire buffer and thus the 4709 * entire page. 4710 */ 4711 if (eoff > soff) { 4712 vm_page_set_validclean( 4713 m, 4714 (vm_offset_t) (soff & PAGE_MASK), 4715 (vm_offset_t) (eoff - soff) 4716 ); 4717 } 4718 } 4719 4720 /* 4721 * Acquire a shared busy on all pages in the buf. 4722 */ 4723 void 4724 vfs_busy_pages_acquire(struct buf *bp) 4725 { 4726 int i; 4727 4728 for (i = 0; i < bp->b_npages; i++) 4729 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY); 4730 } 4731 4732 void 4733 vfs_busy_pages_release(struct buf *bp) 4734 { 4735 int i; 4736 4737 for (i = 0; i < bp->b_npages; i++) 4738 vm_page_sunbusy(bp->b_pages[i]); 4739 } 4740 4741 /* 4742 * This routine is called before a device strategy routine. 4743 * It is used to tell the VM system that paging I/O is in 4744 * progress, and treat the pages associated with the buffer 4745 * almost as being exclusive busy. Also the object paging_in_progress 4746 * flag is handled to make sure that the object doesn't become 4747 * inconsistent. 4748 * 4749 * Since I/O has not been initiated yet, certain buffer flags 4750 * such as BIO_ERROR or B_INVAL may be in an inconsistent state 4751 * and should be ignored. 4752 */ 4753 void 4754 vfs_busy_pages(struct buf *bp, int clear_modify) 4755 { 4756 vm_object_t obj; 4757 vm_ooffset_t foff; 4758 vm_page_t m; 4759 int i; 4760 bool bogus; 4761 4762 if (!(bp->b_flags & B_VMIO)) 4763 return; 4764 4765 obj = bp->b_bufobj->bo_object; 4766 foff = bp->b_offset; 4767 KASSERT(bp->b_offset != NOOFFSET, 4768 ("vfs_busy_pages: no buffer offset")); 4769 if ((bp->b_flags & B_CLUSTER) == 0) { 4770 vm_object_pip_add(obj, bp->b_npages); 4771 vfs_busy_pages_acquire(bp); 4772 } 4773 if (bp->b_bufsize != 0) 4774 vfs_setdirty_range(bp); 4775 bogus = false; 4776 for (i = 0; i < bp->b_npages; i++) { 4777 m = bp->b_pages[i]; 4778 vm_page_assert_sbusied(m); 4779 4780 /* 4781 * When readying a buffer for a read ( i.e 4782 * clear_modify == 0 ), it is important to do 4783 * bogus_page replacement for valid pages in 4784 * partially instantiated buffers. Partially 4785 * instantiated buffers can, in turn, occur when 4786 * reconstituting a buffer from its VM backing store 4787 * base. We only have to do this if B_CACHE is 4788 * clear ( which causes the I/O to occur in the 4789 * first place ). The replacement prevents the read 4790 * I/O from overwriting potentially dirty VM-backed 4791 * pages. XXX bogus page replacement is, uh, bogus. 4792 * It may not work properly with small-block devices. 4793 * We need to find a better way. 4794 */ 4795 if (clear_modify) { 4796 pmap_remove_write(m); 4797 vfs_page_set_validclean(bp, foff, m); 4798 } else if (vm_page_all_valid(m) && 4799 (bp->b_flags & B_CACHE) == 0) { 4800 bp->b_pages[i] = bogus_page; 4801 bogus = true; 4802 } 4803 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4804 } 4805 if (bogus && buf_mapped(bp)) { 4806 BUF_CHECK_MAPPED(bp); 4807 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4808 bp->b_pages, bp->b_npages); 4809 } 4810 } 4811 4812 /* 4813 * vfs_bio_set_valid: 4814 * 4815 * Set the range within the buffer to valid. The range is 4816 * relative to the beginning of the buffer, b_offset. Note that 4817 * b_offset itself may be offset from the beginning of the first 4818 * page. 4819 */ 4820 void 4821 vfs_bio_set_valid(struct buf *bp, int base, int size) 4822 { 4823 int i, n; 4824 vm_page_t m; 4825 4826 if (!(bp->b_flags & B_VMIO)) 4827 return; 4828 4829 /* 4830 * Fixup base to be relative to beginning of first page. 4831 * Set initial n to be the maximum number of bytes in the 4832 * first page that can be validated. 4833 */ 4834 base += (bp->b_offset & PAGE_MASK); 4835 n = PAGE_SIZE - (base & PAGE_MASK); 4836 4837 /* 4838 * Busy may not be strictly necessary here because the pages are 4839 * unlikely to be fully valid and the vnode lock will synchronize 4840 * their access via getpages. It is grabbed for consistency with 4841 * other page validation. 4842 */ 4843 vfs_busy_pages_acquire(bp); 4844 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4845 m = bp->b_pages[i]; 4846 if (n > size) 4847 n = size; 4848 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4849 base += n; 4850 size -= n; 4851 n = PAGE_SIZE; 4852 } 4853 vfs_busy_pages_release(bp); 4854 } 4855 4856 /* 4857 * vfs_bio_clrbuf: 4858 * 4859 * If the specified buffer is a non-VMIO buffer, clear the entire 4860 * buffer. If the specified buffer is a VMIO buffer, clear and 4861 * validate only the previously invalid portions of the buffer. 4862 * This routine essentially fakes an I/O, so we need to clear 4863 * BIO_ERROR and B_INVAL. 4864 * 4865 * Note that while we only theoretically need to clear through b_bcount, 4866 * we go ahead and clear through b_bufsize. 4867 */ 4868 void 4869 vfs_bio_clrbuf(struct buf *bp) 4870 { 4871 int i, j, sa, ea, slide, zbits; 4872 vm_page_bits_t mask; 4873 4874 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4875 clrbuf(bp); 4876 return; 4877 } 4878 bp->b_flags &= ~B_INVAL; 4879 bp->b_ioflags &= ~BIO_ERROR; 4880 vfs_busy_pages_acquire(bp); 4881 sa = bp->b_offset & PAGE_MASK; 4882 slide = 0; 4883 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4884 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4885 ea = slide & PAGE_MASK; 4886 if (ea == 0) 4887 ea = PAGE_SIZE; 4888 if (bp->b_pages[i] == bogus_page) 4889 continue; 4890 j = sa / DEV_BSIZE; 4891 zbits = (sizeof(vm_page_bits_t) * NBBY) - 4892 (ea - sa) / DEV_BSIZE; 4893 mask = (VM_PAGE_BITS_ALL >> zbits) << j; 4894 if ((bp->b_pages[i]->valid & mask) == mask) 4895 continue; 4896 if ((bp->b_pages[i]->valid & mask) == 0) 4897 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4898 else { 4899 for (; sa < ea; sa += DEV_BSIZE, j++) { 4900 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4901 pmap_zero_page_area(bp->b_pages[i], 4902 sa, DEV_BSIZE); 4903 } 4904 } 4905 } 4906 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE, 4907 roundup2(ea - sa, DEV_BSIZE)); 4908 } 4909 vfs_busy_pages_release(bp); 4910 bp->b_resid = 0; 4911 } 4912 4913 void 4914 vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4915 { 4916 vm_page_t m; 4917 int i, n; 4918 4919 if (buf_mapped(bp)) { 4920 BUF_CHECK_MAPPED(bp); 4921 bzero(bp->b_data + base, size); 4922 } else { 4923 BUF_CHECK_UNMAPPED(bp); 4924 n = PAGE_SIZE - (base & PAGE_MASK); 4925 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4926 m = bp->b_pages[i]; 4927 if (n > size) 4928 n = size; 4929 pmap_zero_page_area(m, base & PAGE_MASK, n); 4930 base += n; 4931 size -= n; 4932 n = PAGE_SIZE; 4933 } 4934 } 4935 } 4936 4937 /* 4938 * Update buffer flags based on I/O request parameters, optionally releasing the 4939 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM, 4940 * where they may be placed on a page queue (VMIO) or freed immediately (direct 4941 * I/O). Otherwise the buffer is released to the cache. 4942 */ 4943 static void 4944 b_io_dismiss(struct buf *bp, int ioflag, bool release) 4945 { 4946 4947 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0, 4948 ("buf %p non-VMIO noreuse", bp)); 4949 4950 if ((ioflag & IO_DIRECT) != 0) 4951 bp->b_flags |= B_DIRECT; 4952 if ((ioflag & IO_EXT) != 0) 4953 bp->b_xflags |= BX_ALTDATA; 4954 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) { 4955 bp->b_flags |= B_RELBUF; 4956 if ((ioflag & IO_NOREUSE) != 0) 4957 bp->b_flags |= B_NOREUSE; 4958 if (release) 4959 brelse(bp); 4960 } else if (release) 4961 bqrelse(bp); 4962 } 4963 4964 void 4965 vfs_bio_brelse(struct buf *bp, int ioflag) 4966 { 4967 4968 b_io_dismiss(bp, ioflag, true); 4969 } 4970 4971 void 4972 vfs_bio_set_flags(struct buf *bp, int ioflag) 4973 { 4974 4975 b_io_dismiss(bp, ioflag, false); 4976 } 4977 4978 /* 4979 * vm_hold_load_pages and vm_hold_free_pages get pages into 4980 * a buffers address space. The pages are anonymous and are 4981 * not associated with a file object. 4982 */ 4983 static void 4984 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4985 { 4986 vm_offset_t pg; 4987 vm_page_t p; 4988 int index; 4989 4990 BUF_CHECK_MAPPED(bp); 4991 4992 to = round_page(to); 4993 from = round_page(from); 4994 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4995 MPASS((bp->b_flags & B_MAXPHYS) == 0); 4996 KASSERT(to - from <= maxbcachebuf, 4997 ("vm_hold_load_pages too large %p %#jx %#jx %u", 4998 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf)); 4999 5000 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 5001 /* 5002 * note: must allocate system pages since blocking here 5003 * could interfere with paging I/O, no matter which 5004 * process we are. 5005 */ 5006 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 5007 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK); 5008 pmap_qenter(pg, &p, 1); 5009 bp->b_pages[index] = p; 5010 } 5011 bp->b_npages = index; 5012 } 5013 5014 /* Return pages associated with this buf to the vm system */ 5015 static void 5016 vm_hold_free_pages(struct buf *bp, int newbsize) 5017 { 5018 vm_offset_t from; 5019 vm_page_t p; 5020 int index, newnpages; 5021 5022 BUF_CHECK_MAPPED(bp); 5023 5024 from = round_page((vm_offset_t)bp->b_data + newbsize); 5025 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 5026 if (bp->b_npages > newnpages) 5027 pmap_qremove(from, bp->b_npages - newnpages); 5028 for (index = newnpages; index < bp->b_npages; index++) { 5029 p = bp->b_pages[index]; 5030 bp->b_pages[index] = NULL; 5031 vm_page_unwire_noq(p); 5032 vm_page_free(p); 5033 } 5034 bp->b_npages = newnpages; 5035 } 5036 5037 /* 5038 * Map an IO request into kernel virtual address space. 5039 * 5040 * All requests are (re)mapped into kernel VA space. 5041 * Notice that we use b_bufsize for the size of the buffer 5042 * to be mapped. b_bcount might be modified by the driver. 5043 * 5044 * Note that even if the caller determines that the address space should 5045 * be valid, a race or a smaller-file mapped into a larger space may 5046 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 5047 * check the return value. 5048 * 5049 * This function only works with pager buffers. 5050 */ 5051 int 5052 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf) 5053 { 5054 vm_prot_t prot; 5055 int pidx; 5056 5057 MPASS((bp->b_flags & B_MAXPHYS) != 0); 5058 prot = VM_PROT_READ; 5059 if (bp->b_iocmd == BIO_READ) 5060 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 5061 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 5062 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES); 5063 if (pidx < 0) 5064 return (-1); 5065 bp->b_bufsize = len; 5066 bp->b_npages = pidx; 5067 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK; 5068 if (mapbuf || !unmapped_buf_allowed) { 5069 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx); 5070 bp->b_data = bp->b_kvabase + bp->b_offset; 5071 } else 5072 bp->b_data = unmapped_buf; 5073 return (0); 5074 } 5075 5076 /* 5077 * Free the io map PTEs associated with this IO operation. 5078 * We also invalidate the TLB entries and restore the original b_addr. 5079 * 5080 * This function only works with pager buffers. 5081 */ 5082 void 5083 vunmapbuf(struct buf *bp) 5084 { 5085 int npages; 5086 5087 npages = bp->b_npages; 5088 if (buf_mapped(bp)) 5089 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 5090 vm_page_unhold_pages(bp->b_pages, npages); 5091 5092 bp->b_data = unmapped_buf; 5093 } 5094 5095 void 5096 bdone(struct buf *bp) 5097 { 5098 struct mtx *mtxp; 5099 5100 mtxp = mtx_pool_find(mtxpool_sleep, bp); 5101 mtx_lock(mtxp); 5102 bp->b_flags |= B_DONE; 5103 wakeup(bp); 5104 mtx_unlock(mtxp); 5105 } 5106 5107 void 5108 bwait(struct buf *bp, u_char pri, const char *wchan) 5109 { 5110 struct mtx *mtxp; 5111 5112 mtxp = mtx_pool_find(mtxpool_sleep, bp); 5113 mtx_lock(mtxp); 5114 while ((bp->b_flags & B_DONE) == 0) 5115 msleep(bp, mtxp, pri, wchan, 0); 5116 mtx_unlock(mtxp); 5117 } 5118 5119 int 5120 bufsync(struct bufobj *bo, int waitfor) 5121 { 5122 5123 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread)); 5124 } 5125 5126 void 5127 bufstrategy(struct bufobj *bo, struct buf *bp) 5128 { 5129 int i __unused; 5130 struct vnode *vp; 5131 5132 vp = bp->b_vp; 5133 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 5134 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 5135 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 5136 i = VOP_STRATEGY(vp, bp); 5137 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 5138 } 5139 5140 /* 5141 * Initialize a struct bufobj before use. Memory is assumed zero filled. 5142 */ 5143 void 5144 bufobj_init(struct bufobj *bo, void *private) 5145 { 5146 static volatile int bufobj_cleanq; 5147 5148 bo->bo_domain = 5149 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains; 5150 rw_init(BO_LOCKPTR(bo), "bufobj interlock"); 5151 bo->bo_private = private; 5152 TAILQ_INIT(&bo->bo_clean.bv_hd); 5153 TAILQ_INIT(&bo->bo_dirty.bv_hd); 5154 } 5155 5156 void 5157 bufobj_wrefl(struct bufobj *bo) 5158 { 5159 5160 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 5161 ASSERT_BO_WLOCKED(bo); 5162 bo->bo_numoutput++; 5163 } 5164 5165 void 5166 bufobj_wref(struct bufobj *bo) 5167 { 5168 5169 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 5170 BO_LOCK(bo); 5171 bo->bo_numoutput++; 5172 BO_UNLOCK(bo); 5173 } 5174 5175 void 5176 bufobj_wdrop(struct bufobj *bo) 5177 { 5178 5179 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 5180 BO_LOCK(bo); 5181 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 5182 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 5183 bo->bo_flag &= ~BO_WWAIT; 5184 wakeup(&bo->bo_numoutput); 5185 } 5186 BO_UNLOCK(bo); 5187 } 5188 5189 int 5190 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 5191 { 5192 int error; 5193 5194 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 5195 ASSERT_BO_WLOCKED(bo); 5196 error = 0; 5197 while (bo->bo_numoutput) { 5198 bo->bo_flag |= BO_WWAIT; 5199 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 5200 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 5201 if (error) 5202 break; 5203 } 5204 return (error); 5205 } 5206 5207 /* 5208 * Set bio_data or bio_ma for struct bio from the struct buf. 5209 */ 5210 void 5211 bdata2bio(struct buf *bp, struct bio *bip) 5212 { 5213 5214 if (!buf_mapped(bp)) { 5215 KASSERT(unmapped_buf_allowed, ("unmapped")); 5216 bip->bio_ma = bp->b_pages; 5217 bip->bio_ma_n = bp->b_npages; 5218 bip->bio_data = unmapped_buf; 5219 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 5220 bip->bio_flags |= BIO_UNMAPPED; 5221 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 5222 PAGE_SIZE == bp->b_npages, 5223 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 5224 (long long)bip->bio_length, bip->bio_ma_n)); 5225 } else { 5226 bip->bio_data = bp->b_data; 5227 bip->bio_ma = NULL; 5228 } 5229 } 5230 5231 static int buf_pager_relbuf; 5232 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN, 5233 &buf_pager_relbuf, 0, 5234 "Make buffer pager release buffers after reading"); 5235 5236 /* 5237 * The buffer pager. It uses buffer reads to validate pages. 5238 * 5239 * In contrast to the generic local pager from vm/vnode_pager.c, this 5240 * pager correctly and easily handles volumes where the underlying 5241 * device block size is greater than the machine page size. The 5242 * buffer cache transparently extends the requested page run to be 5243 * aligned at the block boundary, and does the necessary bogus page 5244 * replacements in the addends to avoid obliterating already valid 5245 * pages. 5246 * 5247 * The only non-trivial issue is that the exclusive busy state for 5248 * pages, which is assumed by the vm_pager_getpages() interface, is 5249 * incompatible with the VMIO buffer cache's desire to share-busy the 5250 * pages. This function performs a trivial downgrade of the pages' 5251 * state before reading buffers, and a less trivial upgrade from the 5252 * shared-busy to excl-busy state after the read. 5253 */ 5254 int 5255 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count, 5256 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno, 5257 vbg_get_blksize_t get_blksize) 5258 { 5259 vm_page_t m; 5260 vm_object_t object; 5261 struct buf *bp; 5262 struct mount *mp; 5263 daddr_t lbn, lbnp; 5264 vm_ooffset_t la, lb, poff, poffe; 5265 long bo_bs, bsize; 5266 int br_flags, error, i, pgsin, pgsin_a, pgsin_b; 5267 bool redo, lpart; 5268 5269 object = vp->v_object; 5270 mp = vp->v_mount; 5271 error = 0; 5272 la = IDX_TO_OFF(ma[count - 1]->pindex); 5273 if (la >= object->un_pager.vnp.vnp_size) 5274 return (VM_PAGER_BAD); 5275 5276 /* 5277 * Change the meaning of la from where the last requested page starts 5278 * to where it ends, because that's the end of the requested region 5279 * and the start of the potential read-ahead region. 5280 */ 5281 la += PAGE_SIZE; 5282 lpart = la > object->un_pager.vnp.vnp_size; 5283 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)), 5284 &bo_bs); 5285 if (error != 0) 5286 return (VM_PAGER_ERROR); 5287 5288 /* 5289 * Calculate read-ahead, behind and total pages. 5290 */ 5291 pgsin = count; 5292 lb = IDX_TO_OFF(ma[0]->pindex); 5293 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs)); 5294 pgsin += pgsin_b; 5295 if (rbehind != NULL) 5296 *rbehind = pgsin_b; 5297 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la); 5298 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size) 5299 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size, 5300 PAGE_SIZE) - la); 5301 pgsin += pgsin_a; 5302 if (rahead != NULL) 5303 *rahead = pgsin_a; 5304 VM_CNT_INC(v_vnodein); 5305 VM_CNT_ADD(v_vnodepgsin, pgsin); 5306 5307 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) 5308 != 0) ? GB_UNMAPPED : 0; 5309 again: 5310 for (i = 0; i < count; i++) { 5311 if (ma[i] != bogus_page) 5312 vm_page_busy_downgrade(ma[i]); 5313 } 5314 5315 lbnp = -1; 5316 for (i = 0; i < count; i++) { 5317 m = ma[i]; 5318 if (m == bogus_page) 5319 continue; 5320 5321 /* 5322 * Pages are shared busy and the object lock is not 5323 * owned, which together allow for the pages' 5324 * invalidation. The racy test for validity avoids 5325 * useless creation of the buffer for the most typical 5326 * case when invalidation is not used in redo or for 5327 * parallel read. The shared->excl upgrade loop at 5328 * the end of the function catches the race in a 5329 * reliable way (protected by the object lock). 5330 */ 5331 if (vm_page_all_valid(m)) 5332 continue; 5333 5334 poff = IDX_TO_OFF(m->pindex); 5335 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size); 5336 for (; poff < poffe; poff += bsize) { 5337 lbn = get_lblkno(vp, poff); 5338 if (lbn == lbnp) 5339 goto next_page; 5340 lbnp = lbn; 5341 5342 error = get_blksize(vp, lbn, &bsize); 5343 if (error == 0) 5344 error = bread_gb(vp, lbn, bsize, 5345 curthread->td_ucred, br_flags, &bp); 5346 if (error != 0) 5347 goto end_pages; 5348 if (bp->b_rcred == curthread->td_ucred) { 5349 crfree(bp->b_rcred); 5350 bp->b_rcred = NOCRED; 5351 } 5352 if (LIST_EMPTY(&bp->b_dep)) { 5353 /* 5354 * Invalidation clears m->valid, but 5355 * may leave B_CACHE flag if the 5356 * buffer existed at the invalidation 5357 * time. In this case, recycle the 5358 * buffer to do real read on next 5359 * bread() after redo. 5360 * 5361 * Otherwise B_RELBUF is not strictly 5362 * necessary, enable to reduce buf 5363 * cache pressure. 5364 */ 5365 if (buf_pager_relbuf || 5366 !vm_page_all_valid(m)) 5367 bp->b_flags |= B_RELBUF; 5368 5369 bp->b_flags &= ~B_NOCACHE; 5370 brelse(bp); 5371 } else { 5372 bqrelse(bp); 5373 } 5374 } 5375 KASSERT(1 /* racy, enable for debugging */ || 5376 vm_page_all_valid(m) || i == count - 1, 5377 ("buf %d %p invalid", i, m)); 5378 if (i == count - 1 && lpart) { 5379 if (!vm_page_none_valid(m) && 5380 !vm_page_all_valid(m)) 5381 vm_page_zero_invalid(m, TRUE); 5382 } 5383 next_page:; 5384 } 5385 end_pages: 5386 5387 redo = false; 5388 for (i = 0; i < count; i++) { 5389 if (ma[i] == bogus_page) 5390 continue; 5391 if (vm_page_busy_tryupgrade(ma[i]) == 0) { 5392 vm_page_sunbusy(ma[i]); 5393 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex, 5394 VM_ALLOC_NORMAL); 5395 } 5396 5397 /* 5398 * Since the pages were only sbusy while neither the 5399 * buffer nor the object lock was held by us, or 5400 * reallocated while vm_page_grab() slept for busy 5401 * relinguish, they could have been invalidated. 5402 * Recheck the valid bits and re-read as needed. 5403 * 5404 * Note that the last page is made fully valid in the 5405 * read loop, and partial validity for the page at 5406 * index count - 1 could mean that the page was 5407 * invalidated or removed, so we must restart for 5408 * safety as well. 5409 */ 5410 if (!vm_page_all_valid(ma[i])) 5411 redo = true; 5412 } 5413 if (redo && error == 0) 5414 goto again; 5415 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); 5416 } 5417 5418 #include "opt_ddb.h" 5419 #ifdef DDB 5420 #include <ddb/ddb.h> 5421 5422 /* DDB command to show buffer data */ 5423 DB_SHOW_COMMAND(buffer, db_show_buffer) 5424 { 5425 /* get args */ 5426 struct buf *bp = (struct buf *)addr; 5427 #ifdef FULL_BUF_TRACKING 5428 uint32_t i, j; 5429 #endif 5430 5431 if (!have_addr) { 5432 db_printf("usage: show buffer <addr>\n"); 5433 return; 5434 } 5435 5436 db_printf("buf at %p\n", bp); 5437 db_printf("b_flags = 0x%b, b_xflags=0x%b\n", 5438 (u_int)bp->b_flags, PRINT_BUF_FLAGS, 5439 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS); 5440 db_printf("b_vflags=0x%b b_ioflags0x%b\n", 5441 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS, 5442 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS); 5443 db_printf( 5444 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 5445 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, " 5446 "b_vp = %p, b_dep = %p\n", 5447 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 5448 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 5449 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first); 5450 db_printf("b_kvabase = %p, b_kvasize = %d\n", 5451 bp->b_kvabase, bp->b_kvasize); 5452 if (bp->b_npages) { 5453 int i; 5454 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 5455 for (i = 0; i < bp->b_npages; i++) { 5456 vm_page_t m; 5457 m = bp->b_pages[i]; 5458 if (m != NULL) 5459 db_printf("(%p, 0x%lx, 0x%lx)", m->object, 5460 (u_long)m->pindex, 5461 (u_long)VM_PAGE_TO_PHYS(m)); 5462 else 5463 db_printf("( ??? )"); 5464 if ((i + 1) < bp->b_npages) 5465 db_printf(","); 5466 } 5467 db_printf("\n"); 5468 } 5469 BUF_LOCKPRINTINFO(bp); 5470 #if defined(FULL_BUF_TRACKING) 5471 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt); 5472 5473 i = bp->b_io_tcnt % BUF_TRACKING_SIZE; 5474 for (j = 1; j <= BUF_TRACKING_SIZE; j++) { 5475 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL) 5476 continue; 5477 db_printf(" %2u: %s\n", j, 5478 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]); 5479 } 5480 #elif defined(BUF_TRACKING) 5481 db_printf("b_io_tracking: %s\n", bp->b_io_tracking); 5482 #endif 5483 db_printf(" "); 5484 } 5485 5486 DB_SHOW_COMMAND_FLAGS(bufqueues, bufqueues, DB_CMD_MEMSAFE) 5487 { 5488 struct bufdomain *bd; 5489 struct buf *bp; 5490 long total; 5491 int i, j, cnt; 5492 5493 db_printf("bqempty: %d\n", bqempty.bq_len); 5494 5495 for (i = 0; i < buf_domains; i++) { 5496 bd = &bdomain[i]; 5497 db_printf("Buf domain %d\n", i); 5498 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers); 5499 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers); 5500 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers); 5501 db_printf("\n"); 5502 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace); 5503 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace); 5504 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace); 5505 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace); 5506 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh); 5507 db_printf("\n"); 5508 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers); 5509 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers); 5510 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers); 5511 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh); 5512 db_printf("\n"); 5513 total = 0; 5514 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist) 5515 total += bp->b_bufsize; 5516 db_printf("\tcleanq count\t%d (%ld)\n", 5517 bd->bd_cleanq->bq_len, total); 5518 total = 0; 5519 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist) 5520 total += bp->b_bufsize; 5521 db_printf("\tdirtyq count\t%d (%ld)\n", 5522 bd->bd_dirtyq.bq_len, total); 5523 db_printf("\twakeup\t\t%d\n", bd->bd_wanted); 5524 db_printf("\tlim\t\t%d\n", bd->bd_lim); 5525 db_printf("\tCPU "); 5526 for (j = 0; j <= mp_maxid; j++) 5527 db_printf("%d, ", bd->bd_subq[j].bq_len); 5528 db_printf("\n"); 5529 cnt = 0; 5530 total = 0; 5531 for (j = 0; j < nbuf; j++) { 5532 bp = nbufp(j); 5533 if (bp->b_domain == i && BUF_ISLOCKED(bp)) { 5534 cnt++; 5535 total += bp->b_bufsize; 5536 } 5537 } 5538 db_printf("\tLocked buffers: %d space %ld\n", cnt, total); 5539 cnt = 0; 5540 total = 0; 5541 for (j = 0; j < nbuf; j++) { 5542 bp = nbufp(j); 5543 if (bp->b_domain == i) { 5544 cnt++; 5545 total += bp->b_bufsize; 5546 } 5547 } 5548 db_printf("\tTotal buffers: %d space %ld\n", cnt, total); 5549 } 5550 } 5551 5552 DB_SHOW_COMMAND_FLAGS(lockedbufs, lockedbufs, DB_CMD_MEMSAFE) 5553 { 5554 struct buf *bp; 5555 int i; 5556 5557 for (i = 0; i < nbuf; i++) { 5558 bp = nbufp(i); 5559 if (BUF_ISLOCKED(bp)) { 5560 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5561 db_printf("\n"); 5562 if (db_pager_quit) 5563 break; 5564 } 5565 } 5566 } 5567 5568 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 5569 { 5570 struct vnode *vp; 5571 struct buf *bp; 5572 5573 if (!have_addr) { 5574 db_printf("usage: show vnodebufs <addr>\n"); 5575 return; 5576 } 5577 vp = (struct vnode *)addr; 5578 db_printf("Clean buffers:\n"); 5579 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 5580 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5581 db_printf("\n"); 5582 } 5583 db_printf("Dirty buffers:\n"); 5584 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 5585 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5586 db_printf("\n"); 5587 } 5588 } 5589 5590 DB_COMMAND_FLAGS(countfreebufs, db_coundfreebufs, DB_CMD_MEMSAFE) 5591 { 5592 struct buf *bp; 5593 int i, used = 0, nfree = 0; 5594 5595 if (have_addr) { 5596 db_printf("usage: countfreebufs\n"); 5597 return; 5598 } 5599 5600 for (i = 0; i < nbuf; i++) { 5601 bp = nbufp(i); 5602 if (bp->b_qindex == QUEUE_EMPTY) 5603 nfree++; 5604 else 5605 used++; 5606 } 5607 5608 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 5609 nfree + used); 5610 db_printf("numfreebuffers is %d\n", numfreebuffers); 5611 } 5612 #endif /* DDB */ 5613