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/param.h> 48 #include <sys/systm.h> 49 #include <sys/asan.h> 50 #include <sys/bio.h> 51 #include <sys/bitset.h> 52 #include <sys/boottrace.h> 53 #include <sys/buf.h> 54 #include <sys/conf.h> 55 #include <sys/counter.h> 56 #include <sys/devicestat.h> 57 #include <sys/eventhandler.h> 58 #include <sys/fail.h> 59 #include <sys/ktr.h> 60 #include <sys/limits.h> 61 #include <sys/lock.h> 62 #include <sys/malloc.h> 63 #include <sys/memdesc.h> 64 #include <sys/mount.h> 65 #include <sys/mutex.h> 66 #include <sys/kernel.h> 67 #include <sys/kthread.h> 68 #include <sys/pctrie.h> 69 #include <sys/proc.h> 70 #include <sys/racct.h> 71 #include <sys/refcount.h> 72 #include <sys/resourcevar.h> 73 #include <sys/rwlock.h> 74 #include <sys/sched.h> 75 #include <sys/smp.h> 76 #include <sys/sysctl.h> 77 #include <sys/syscallsubr.h> 78 #include <sys/vmem.h> 79 #include <sys/vmmeter.h> 80 #include <sys/vnode.h> 81 #include <sys/watchdog.h> 82 #include <geom/geom.h> 83 #include <vm/vm.h> 84 #include <vm/vm_param.h> 85 #include <vm/vm_kern.h> 86 #include <vm/vm_object.h> 87 #include <vm/vm_page.h> 88 #include <vm/vm_pageout.h> 89 #include <vm/vm_pager.h> 90 #include <vm/vm_extern.h> 91 #include <vm/vm_map.h> 92 #include <vm/swap_pager.h> 93 94 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 95 96 struct bio_ops bioops; /* I/O operation notification */ 97 98 struct buf_ops buf_ops_bio = { 99 .bop_name = "buf_ops_bio", 100 .bop_write = bufwrite, 101 .bop_strategy = bufstrategy, 102 .bop_sync = bufsync, 103 .bop_bdflush = bufbdflush, 104 }; 105 106 struct bufqueue { 107 struct mtx_padalign bq_lock; 108 TAILQ_HEAD(, buf) bq_queue; 109 uint8_t bq_index; 110 uint16_t bq_subqueue; 111 int bq_len; 112 } __aligned(CACHE_LINE_SIZE); 113 114 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock) 115 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq))) 116 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq))) 117 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED) 118 119 struct bufdomain { 120 struct bufqueue *bd_subq; 121 struct bufqueue bd_dirtyq; 122 struct bufqueue *bd_cleanq; 123 struct mtx_padalign bd_run_lock; 124 /* Constants */ 125 long bd_maxbufspace; 126 long bd_hibufspace; 127 long bd_lobufspace; 128 long bd_bufspacethresh; 129 int bd_hifreebuffers; 130 int bd_lofreebuffers; 131 int bd_hidirtybuffers; 132 int bd_lodirtybuffers; 133 int bd_dirtybufthresh; 134 int bd_lim; 135 /* atomics */ 136 int bd_wanted; 137 bool bd_shutdown; 138 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers; 139 int __aligned(CACHE_LINE_SIZE) bd_running; 140 long __aligned(CACHE_LINE_SIZE) bd_bufspace; 141 int __aligned(CACHE_LINE_SIZE) bd_freebuffers; 142 } __aligned(CACHE_LINE_SIZE); 143 144 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock) 145 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd))) 146 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd))) 147 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED) 148 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock) 149 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd))) 150 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd))) 151 #define BD_DOMAIN(bd) (bd - bdomain) 152 153 static char *buf; /* buffer header pool */ 154 static struct buf * 155 nbufp(unsigned i) 156 { 157 return ((struct buf *)(buf + (sizeof(struct buf) + 158 sizeof(vm_page_t) * atop(maxbcachebuf)) * i)); 159 } 160 161 caddr_t __read_mostly unmapped_buf; 162 #ifdef INVARIANTS 163 caddr_t poisoned_buf = (void *)-1; 164 #endif 165 166 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */ 167 struct proc *bufdaemonproc; 168 169 static void vm_hold_free_pages(struct buf *bp, int newbsize); 170 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 171 vm_offset_t to); 172 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 173 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 174 vm_page_t m); 175 static void vfs_clean_pages_dirty_buf(struct buf *bp); 176 static void vfs_setdirty_range(struct buf *bp); 177 static void vfs_vmio_invalidate(struct buf *bp); 178 static void vfs_vmio_truncate(struct buf *bp, int npages); 179 static void vfs_vmio_extend(struct buf *bp, int npages, int size); 180 static int vfs_bio_clcheck(struct vnode *vp, int size, 181 daddr_t lblkno, daddr_t blkno); 182 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int, 183 void (*)(struct buf *)); 184 static int buf_flush(struct vnode *vp, struct bufdomain *, int); 185 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int); 186 static void buf_daemon(void); 187 static __inline void bd_wakeup(void); 188 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS); 189 static void bufkva_reclaim(vmem_t *, int); 190 static void bufkva_free(struct buf *); 191 static int buf_import(void *, void **, int, int, int); 192 static void buf_release(void *, void **, int); 193 static void maxbcachebuf_adjust(void); 194 static inline struct bufdomain *bufdomain(struct buf *); 195 static void bq_remove(struct bufqueue *bq, struct buf *bp); 196 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock); 197 static int buf_recycle(struct bufdomain *, bool kva); 198 static void bq_init(struct bufqueue *bq, int qindex, int cpu, 199 const char *lockname); 200 static void bd_init(struct bufdomain *bd); 201 static int bd_flushall(struct bufdomain *bd); 202 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS); 203 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS); 204 205 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 206 int vmiodirenable = TRUE; 207 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 208 "Use the VM system for directory writes"); 209 long runningbufspace; 210 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 211 "Amount of presently outstanding async buffer io"); 212 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 213 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers"); 214 static counter_u64_t bufkvaspace; 215 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, 216 "Kernel virtual memory used for buffers"); 217 static long maxbufspace; 218 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace, 219 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace, 220 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L", 221 "Maximum allowed value of bufspace (including metadata)"); 222 static long bufmallocspace; 223 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 224 "Amount of malloced memory for buffers"); 225 static long maxbufmallocspace; 226 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 227 0, "Maximum amount of malloced memory for buffers"); 228 static long lobufspace; 229 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace, 230 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace, 231 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L", 232 "Minimum amount of buffers we want to have"); 233 long hibufspace; 234 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace, 235 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace, 236 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L", 237 "Maximum allowed value of bufspace (excluding metadata)"); 238 long bufspacethresh; 239 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh, 240 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh, 241 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L", 242 "Bufspace consumed before waking the daemon to free some"); 243 static counter_u64_t buffreekvacnt; 244 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 245 "Number of times we have freed the KVA space from some buffer"); 246 static counter_u64_t bufdefragcnt; 247 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 248 "Number of times we have had to repeat buffer allocation to defragment"); 249 static long lorunningspace; 250 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 251 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L", 252 "Minimum preferred space used for in-progress I/O"); 253 static long hirunningspace; 254 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 255 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L", 256 "Maximum amount of space to use for in-progress I/O"); 257 int dirtybufferflushes; 258 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 259 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 260 int bdwriteskip; 261 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 262 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 263 int altbufferflushes; 264 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS, 265 &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers"); 266 static int recursiveflushes; 267 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS, 268 &recursiveflushes, 0, "Number of flushes skipped due to being recursive"); 269 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS); 270 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers, 271 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I", 272 "Number of buffers that are dirty (has unwritten changes) at the moment"); 273 static int lodirtybuffers; 274 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers, 275 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers, 276 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I", 277 "How many buffers we want to have free before bufdaemon can sleep"); 278 static int hidirtybuffers; 279 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers, 280 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers, 281 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I", 282 "When the number of dirty buffers is considered severe"); 283 int dirtybufthresh; 284 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh, 285 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh, 286 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I", 287 "Number of bdwrite to bawrite conversions to clear dirty buffers"); 288 static int numfreebuffers; 289 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 290 "Number of free buffers"); 291 static int lofreebuffers; 292 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers, 293 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers, 294 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I", 295 "Target number of free buffers"); 296 static int hifreebuffers; 297 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers, 298 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers, 299 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I", 300 "Threshold for clean buffer recycling"); 301 static counter_u64_t getnewbufcalls; 302 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, 303 &getnewbufcalls, "Number of calls to getnewbuf"); 304 static counter_u64_t getnewbufrestarts; 305 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, 306 &getnewbufrestarts, 307 "Number of times getnewbuf has had to restart a buffer acquisition"); 308 static counter_u64_t mappingrestarts; 309 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD, 310 &mappingrestarts, 311 "Number of times getblk has had to restart a buffer mapping for " 312 "unmapped buffer"); 313 static counter_u64_t numbufallocfails; 314 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW, 315 &numbufallocfails, "Number of times buffer allocations failed"); 316 static int flushbufqtarget = 100; 317 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 318 "Amount of work to do in flushbufqueues when helping bufdaemon"); 319 static counter_u64_t notbufdflushes; 320 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 321 "Number of dirty buffer flushes done by the bufdaemon helpers"); 322 static long barrierwrites; 323 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS, 324 &barrierwrites, 0, "Number of barrier writes"); 325 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, 326 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 327 &unmapped_buf_allowed, 0, 328 "Permit the use of the unmapped i/o"); 329 int maxbcachebuf = MAXBCACHEBUF; 330 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0, 331 "Maximum size of a buffer cache block"); 332 333 /* 334 * This lock synchronizes access to bd_request. 335 */ 336 static struct mtx_padalign __exclusive_cache_line bdlock; 337 338 /* 339 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 340 * waitrunningbufspace(). 341 */ 342 static struct mtx_padalign __exclusive_cache_line rbreqlock; 343 344 /* 345 * Lock that protects bdirtywait. 346 */ 347 static struct mtx_padalign __exclusive_cache_line bdirtylock; 348 349 /* 350 * bufdaemon shutdown request and sleep channel. 351 */ 352 static bool bd_shutdown; 353 354 /* 355 * Wakeup point for bufdaemon, as well as indicator of whether it is already 356 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 357 * is idling. 358 */ 359 static int bd_request; 360 361 /* 362 * Request for the buf daemon to write more buffers than is indicated by 363 * lodirtybuf. This may be necessary to push out excess dependencies or 364 * defragment the address space where a simple count of the number of dirty 365 * buffers is insufficient to characterize the demand for flushing them. 366 */ 367 static int bd_speedupreq; 368 369 /* 370 * Synchronization (sleep/wakeup) variable for active buffer space requests. 371 * Set when wait starts, cleared prior to wakeup(). 372 * Used in runningbufwakeup() and waitrunningbufspace(). 373 */ 374 static int runningbufreq; 375 376 /* 377 * Synchronization for bwillwrite() waiters. 378 */ 379 static int bdirtywait; 380 381 /* 382 * Definitions for the buffer free lists. 383 */ 384 #define QUEUE_NONE 0 /* on no queue */ 385 #define QUEUE_EMPTY 1 /* empty buffer headers */ 386 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 387 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */ 388 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */ 389 390 /* Maximum number of buffer domains. */ 391 #define BUF_DOMAINS 8 392 393 struct bufdomainset bdlodirty; /* Domains > lodirty */ 394 struct bufdomainset bdhidirty; /* Domains > hidirty */ 395 396 /* Configured number of clean queues. */ 397 static int __read_mostly buf_domains; 398 399 BITSET_DEFINE(bufdomainset, BUF_DOMAINS); 400 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS]; 401 struct bufqueue __exclusive_cache_line bqempty; 402 403 /* 404 * per-cpu empty buffer cache. 405 */ 406 uma_zone_t buf_zone; 407 408 static int 409 sysctl_runningspace(SYSCTL_HANDLER_ARGS) 410 { 411 long value; 412 int error; 413 414 value = *(long *)arg1; 415 error = sysctl_handle_long(oidp, &value, 0, req); 416 if (error != 0 || req->newptr == NULL) 417 return (error); 418 mtx_lock(&rbreqlock); 419 if (arg1 == &hirunningspace) { 420 if (value < lorunningspace) 421 error = EINVAL; 422 else 423 hirunningspace = value; 424 } else { 425 KASSERT(arg1 == &lorunningspace, 426 ("%s: unknown arg1", __func__)); 427 if (value > hirunningspace) 428 error = EINVAL; 429 else 430 lorunningspace = value; 431 } 432 mtx_unlock(&rbreqlock); 433 return (error); 434 } 435 436 static int 437 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS) 438 { 439 int error; 440 int value; 441 int i; 442 443 value = *(int *)arg1; 444 error = sysctl_handle_int(oidp, &value, 0, req); 445 if (error != 0 || req->newptr == NULL) 446 return (error); 447 *(int *)arg1 = value; 448 for (i = 0; i < buf_domains; i++) 449 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = 450 value / buf_domains; 451 452 return (error); 453 } 454 455 static int 456 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS) 457 { 458 long value; 459 int error; 460 int i; 461 462 value = *(long *)arg1; 463 error = sysctl_handle_long(oidp, &value, 0, req); 464 if (error != 0 || req->newptr == NULL) 465 return (error); 466 *(long *)arg1 = value; 467 for (i = 0; i < buf_domains; i++) 468 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = 469 value / buf_domains; 470 471 return (error); 472 } 473 474 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 475 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 476 static int 477 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 478 { 479 long lvalue; 480 int ivalue; 481 int i; 482 483 lvalue = 0; 484 for (i = 0; i < buf_domains; i++) 485 lvalue += bdomain[i].bd_bufspace; 486 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 487 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 488 if (lvalue > INT_MAX) 489 /* On overflow, still write out a long to trigger ENOMEM. */ 490 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 491 ivalue = lvalue; 492 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 493 } 494 #else 495 static int 496 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 497 { 498 long lvalue; 499 int i; 500 501 lvalue = 0; 502 for (i = 0; i < buf_domains; i++) 503 lvalue += bdomain[i].bd_bufspace; 504 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 505 } 506 #endif 507 508 static int 509 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS) 510 { 511 int value; 512 int i; 513 514 value = 0; 515 for (i = 0; i < buf_domains; i++) 516 value += bdomain[i].bd_numdirtybuffers; 517 return (sysctl_handle_int(oidp, &value, 0, req)); 518 } 519 520 /* 521 * bdirtywakeup: 522 * 523 * Wakeup any bwillwrite() waiters. 524 */ 525 static void 526 bdirtywakeup(void) 527 { 528 mtx_lock(&bdirtylock); 529 if (bdirtywait) { 530 bdirtywait = 0; 531 wakeup(&bdirtywait); 532 } 533 mtx_unlock(&bdirtylock); 534 } 535 536 /* 537 * bd_clear: 538 * 539 * Clear a domain from the appropriate bitsets when dirtybuffers 540 * is decremented. 541 */ 542 static void 543 bd_clear(struct bufdomain *bd) 544 { 545 546 mtx_lock(&bdirtylock); 547 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers) 548 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); 549 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers) 550 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); 551 mtx_unlock(&bdirtylock); 552 } 553 554 /* 555 * bd_set: 556 * 557 * Set a domain in the appropriate bitsets when dirtybuffers 558 * is incremented. 559 */ 560 static void 561 bd_set(struct bufdomain *bd) 562 { 563 564 mtx_lock(&bdirtylock); 565 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers) 566 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); 567 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers) 568 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); 569 mtx_unlock(&bdirtylock); 570 } 571 572 /* 573 * bdirtysub: 574 * 575 * Decrement the numdirtybuffers count by one and wakeup any 576 * threads blocked in bwillwrite(). 577 */ 578 static void 579 bdirtysub(struct buf *bp) 580 { 581 struct bufdomain *bd; 582 int num; 583 584 bd = bufdomain(bp); 585 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1); 586 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) 587 bdirtywakeup(); 588 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) 589 bd_clear(bd); 590 } 591 592 /* 593 * bdirtyadd: 594 * 595 * Increment the numdirtybuffers count by one and wakeup the buf 596 * daemon if needed. 597 */ 598 static void 599 bdirtyadd(struct buf *bp) 600 { 601 struct bufdomain *bd; 602 int num; 603 604 /* 605 * Only do the wakeup once as we cross the boundary. The 606 * buf daemon will keep running until the condition clears. 607 */ 608 bd = bufdomain(bp); 609 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1); 610 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) 611 bd_wakeup(); 612 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) 613 bd_set(bd); 614 } 615 616 /* 617 * bufspace_daemon_wakeup: 618 * 619 * Wakeup the daemons responsible for freeing clean bufs. 620 */ 621 static void 622 bufspace_daemon_wakeup(struct bufdomain *bd) 623 { 624 625 /* 626 * avoid the lock if the daemon is running. 627 */ 628 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) { 629 BD_RUN_LOCK(bd); 630 atomic_store_int(&bd->bd_running, 1); 631 wakeup(&bd->bd_running); 632 BD_RUN_UNLOCK(bd); 633 } 634 } 635 636 /* 637 * bufspace_adjust: 638 * 639 * Adjust the reported bufspace for a KVA managed buffer, possibly 640 * waking any waiters. 641 */ 642 static void 643 bufspace_adjust(struct buf *bp, int bufsize) 644 { 645 struct bufdomain *bd; 646 long space; 647 int diff; 648 649 KASSERT((bp->b_flags & B_MALLOC) == 0, 650 ("bufspace_adjust: malloc buf %p", bp)); 651 bd = bufdomain(bp); 652 diff = bufsize - bp->b_bufsize; 653 if (diff < 0) { 654 atomic_subtract_long(&bd->bd_bufspace, -diff); 655 } else if (diff > 0) { 656 space = atomic_fetchadd_long(&bd->bd_bufspace, diff); 657 /* Wake up the daemon on the transition. */ 658 if (space < bd->bd_bufspacethresh && 659 space + diff >= bd->bd_bufspacethresh) 660 bufspace_daemon_wakeup(bd); 661 } 662 bp->b_bufsize = bufsize; 663 } 664 665 /* 666 * bufspace_reserve: 667 * 668 * Reserve bufspace before calling allocbuf(). metadata has a 669 * different space limit than data. 670 */ 671 static int 672 bufspace_reserve(struct bufdomain *bd, int size, bool metadata) 673 { 674 long limit, new; 675 long space; 676 677 if (metadata) 678 limit = bd->bd_maxbufspace; 679 else 680 limit = bd->bd_hibufspace; 681 space = atomic_fetchadd_long(&bd->bd_bufspace, size); 682 new = space + size; 683 if (new > limit) { 684 atomic_subtract_long(&bd->bd_bufspace, size); 685 return (ENOSPC); 686 } 687 688 /* Wake up the daemon on the transition. */ 689 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh) 690 bufspace_daemon_wakeup(bd); 691 692 return (0); 693 } 694 695 /* 696 * bufspace_release: 697 * 698 * Release reserved bufspace after bufspace_adjust() has consumed it. 699 */ 700 static void 701 bufspace_release(struct bufdomain *bd, int size) 702 { 703 704 atomic_subtract_long(&bd->bd_bufspace, size); 705 } 706 707 /* 708 * bufspace_wait: 709 * 710 * Wait for bufspace, acting as the buf daemon if a locked vnode is 711 * supplied. bd_wanted must be set prior to polling for space. The 712 * operation must be re-tried on return. 713 */ 714 static void 715 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags, 716 int slpflag, int slptimeo) 717 { 718 struct thread *td; 719 int error, fl, norunbuf; 720 721 if ((gbflags & GB_NOWAIT_BD) != 0) 722 return; 723 724 td = curthread; 725 BD_LOCK(bd); 726 while (bd->bd_wanted) { 727 if (vp != NULL && vp->v_type != VCHR && 728 (td->td_pflags & TDP_BUFNEED) == 0) { 729 BD_UNLOCK(bd); 730 /* 731 * getblk() is called with a vnode locked, and 732 * some majority of the dirty buffers may as 733 * well belong to the vnode. Flushing the 734 * buffers there would make a progress that 735 * cannot be achieved by the buf_daemon, that 736 * cannot lock the vnode. 737 */ 738 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 739 (td->td_pflags & TDP_NORUNNINGBUF); 740 741 /* 742 * Play bufdaemon. The getnewbuf() function 743 * may be called while the thread owns lock 744 * for another dirty buffer for the same 745 * vnode, which makes it impossible to use 746 * VOP_FSYNC() there, due to the buffer lock 747 * recursion. 748 */ 749 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 750 fl = buf_flush(vp, bd, flushbufqtarget); 751 td->td_pflags &= norunbuf; 752 BD_LOCK(bd); 753 if (fl != 0) 754 continue; 755 if (bd->bd_wanted == 0) 756 break; 757 } 758 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd), 759 (PRIBIO + 4) | slpflag, "newbuf", slptimeo); 760 if (error != 0) 761 break; 762 } 763 BD_UNLOCK(bd); 764 } 765 766 static void 767 bufspace_daemon_shutdown(void *arg, int howto __unused) 768 { 769 struct bufdomain *bd = arg; 770 int error; 771 772 if (KERNEL_PANICKED()) 773 return; 774 775 BD_RUN_LOCK(bd); 776 bd->bd_shutdown = true; 777 wakeup(&bd->bd_running); 778 error = msleep(&bd->bd_shutdown, BD_RUN_LOCKPTR(bd), 0, 779 "bufspace_shutdown", 60 * hz); 780 BD_RUN_UNLOCK(bd); 781 if (error != 0) 782 printf("bufspacedaemon wait error: %d\n", error); 783 } 784 785 /* 786 * bufspace_daemon: 787 * 788 * buffer space management daemon. Tries to maintain some marginal 789 * amount of free buffer space so that requesting processes neither 790 * block nor work to reclaim buffers. 791 */ 792 static void 793 bufspace_daemon(void *arg) 794 { 795 struct bufdomain *bd = arg; 796 797 EVENTHANDLER_REGISTER(shutdown_pre_sync, bufspace_daemon_shutdown, bd, 798 SHUTDOWN_PRI_LAST + 100); 799 800 BD_RUN_LOCK(bd); 801 while (!bd->bd_shutdown) { 802 BD_RUN_UNLOCK(bd); 803 804 /* 805 * Free buffers from the clean queue until we meet our 806 * targets. 807 * 808 * Theory of operation: The buffer cache is most efficient 809 * when some free buffer headers and space are always 810 * available to getnewbuf(). This daemon attempts to prevent 811 * the excessive blocking and synchronization associated 812 * with shortfall. It goes through three phases according 813 * demand: 814 * 815 * 1) The daemon wakes up voluntarily once per-second 816 * during idle periods when the counters are below 817 * the wakeup thresholds (bufspacethresh, lofreebuffers). 818 * 819 * 2) The daemon wakes up as we cross the thresholds 820 * ahead of any potential blocking. This may bounce 821 * slightly according to the rate of consumption and 822 * release. 823 * 824 * 3) The daemon and consumers are starved for working 825 * clean buffers. This is the 'bufspace' sleep below 826 * which will inefficiently trade bufs with bqrelse 827 * until we return to condition 2. 828 */ 829 while (bd->bd_bufspace > bd->bd_lobufspace || 830 bd->bd_freebuffers < bd->bd_hifreebuffers) { 831 if (buf_recycle(bd, false) != 0) { 832 if (bd_flushall(bd)) 833 continue; 834 /* 835 * Speedup dirty if we've run out of clean 836 * buffers. This is possible in particular 837 * because softdep may held many bufs locked 838 * pending writes to other bufs which are 839 * marked for delayed write, exhausting 840 * clean space until they are written. 841 */ 842 bd_speedup(); 843 BD_LOCK(bd); 844 if (bd->bd_wanted) { 845 msleep(&bd->bd_wanted, BD_LOCKPTR(bd), 846 PRIBIO|PDROP, "bufspace", hz/10); 847 } else 848 BD_UNLOCK(bd); 849 } 850 maybe_yield(); 851 } 852 853 /* 854 * Re-check our limits and sleep. bd_running must be 855 * cleared prior to checking the limits to avoid missed 856 * wakeups. The waker will adjust one of bufspace or 857 * freebuffers prior to checking bd_running. 858 */ 859 BD_RUN_LOCK(bd); 860 if (bd->bd_shutdown) 861 break; 862 atomic_store_int(&bd->bd_running, 0); 863 if (bd->bd_bufspace < bd->bd_bufspacethresh && 864 bd->bd_freebuffers > bd->bd_lofreebuffers) { 865 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), 866 PRIBIO, "-", hz); 867 } else { 868 /* Avoid spurious wakeups while running. */ 869 atomic_store_int(&bd->bd_running, 1); 870 } 871 } 872 wakeup(&bd->bd_shutdown); 873 BD_RUN_UNLOCK(bd); 874 kthread_exit(); 875 } 876 877 /* 878 * bufmallocadjust: 879 * 880 * Adjust the reported bufspace for a malloc managed buffer, possibly 881 * waking any waiters. 882 */ 883 static void 884 bufmallocadjust(struct buf *bp, int bufsize) 885 { 886 int diff; 887 888 KASSERT((bp->b_flags & B_MALLOC) != 0, 889 ("bufmallocadjust: non-malloc buf %p", bp)); 890 diff = bufsize - bp->b_bufsize; 891 if (diff < 0) 892 atomic_subtract_long(&bufmallocspace, -diff); 893 else 894 atomic_add_long(&bufmallocspace, diff); 895 bp->b_bufsize = bufsize; 896 } 897 898 /* 899 * runningwakeup: 900 * 901 * Wake up processes that are waiting on asynchronous writes to fall 902 * below lorunningspace. 903 */ 904 static void 905 runningwakeup(void) 906 { 907 908 mtx_lock(&rbreqlock); 909 if (runningbufreq) { 910 runningbufreq = 0; 911 wakeup(&runningbufreq); 912 } 913 mtx_unlock(&rbreqlock); 914 } 915 916 /* 917 * runningbufwakeup: 918 * 919 * Decrement the outstanding write count according. 920 */ 921 void 922 runningbufwakeup(struct buf *bp) 923 { 924 long space, bspace; 925 926 bspace = bp->b_runningbufspace; 927 if (bspace == 0) 928 return; 929 space = atomic_fetchadd_long(&runningbufspace, -bspace); 930 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", 931 space, bspace)); 932 bp->b_runningbufspace = 0; 933 /* 934 * Only acquire the lock and wakeup on the transition from exceeding 935 * the threshold to falling below it. 936 */ 937 if (space < lorunningspace) 938 return; 939 if (space - bspace > lorunningspace) 940 return; 941 runningwakeup(); 942 } 943 944 /* 945 * waitrunningbufspace() 946 * 947 * runningbufspace is a measure of the amount of I/O currently 948 * running. This routine is used in async-write situations to 949 * prevent creating huge backups of pending writes to a device. 950 * Only asynchronous writes are governed by this function. 951 * 952 * This does NOT turn an async write into a sync write. It waits 953 * for earlier writes to complete and generally returns before the 954 * caller's write has reached the device. 955 */ 956 void 957 waitrunningbufspace(void) 958 { 959 960 mtx_lock(&rbreqlock); 961 while (runningbufspace > hirunningspace) { 962 runningbufreq = 1; 963 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 964 } 965 mtx_unlock(&rbreqlock); 966 } 967 968 /* 969 * vfs_buf_test_cache: 970 * 971 * Called when a buffer is extended. This function clears the B_CACHE 972 * bit if the newly extended portion of the buffer does not contain 973 * valid data. 974 */ 975 static __inline void 976 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off, 977 vm_offset_t size, vm_page_t m) 978 { 979 980 /* 981 * This function and its results are protected by higher level 982 * synchronization requiring vnode and buf locks to page in and 983 * validate pages. 984 */ 985 if (bp->b_flags & B_CACHE) { 986 int base = (foff + off) & PAGE_MASK; 987 if (vm_page_is_valid(m, base, size) == 0) 988 bp->b_flags &= ~B_CACHE; 989 } 990 } 991 992 /* Wake up the buffer daemon if necessary */ 993 static void 994 bd_wakeup(void) 995 { 996 997 mtx_lock(&bdlock); 998 if (bd_request == 0) { 999 bd_request = 1; 1000 wakeup(&bd_request); 1001 } 1002 mtx_unlock(&bdlock); 1003 } 1004 1005 /* 1006 * Adjust the maxbcachbuf tunable. 1007 */ 1008 static void 1009 maxbcachebuf_adjust(void) 1010 { 1011 int i; 1012 1013 /* 1014 * maxbcachebuf must be a power of 2 >= MAXBSIZE. 1015 */ 1016 i = 2; 1017 while (i * 2 <= maxbcachebuf) 1018 i *= 2; 1019 maxbcachebuf = i; 1020 if (maxbcachebuf < MAXBSIZE) 1021 maxbcachebuf = MAXBSIZE; 1022 if (maxbcachebuf > maxphys) 1023 maxbcachebuf = maxphys; 1024 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF) 1025 printf("maxbcachebuf=%d\n", maxbcachebuf); 1026 } 1027 1028 /* 1029 * bd_speedup - speedup the buffer cache flushing code 1030 */ 1031 void 1032 bd_speedup(void) 1033 { 1034 int needwake; 1035 1036 mtx_lock(&bdlock); 1037 needwake = 0; 1038 if (bd_speedupreq == 0 || bd_request == 0) 1039 needwake = 1; 1040 bd_speedupreq = 1; 1041 bd_request = 1; 1042 if (needwake) 1043 wakeup(&bd_request); 1044 mtx_unlock(&bdlock); 1045 } 1046 1047 #ifdef __i386__ 1048 #define TRANSIENT_DENOM 5 1049 #else 1050 #define TRANSIENT_DENOM 10 1051 #endif 1052 1053 /* 1054 * Calculating buffer cache scaling values and reserve space for buffer 1055 * headers. This is called during low level kernel initialization and 1056 * may be called more then once. We CANNOT write to the memory area 1057 * being reserved at this time. 1058 */ 1059 caddr_t 1060 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 1061 { 1062 int tuned_nbuf; 1063 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 1064 1065 /* 1066 * With KASAN or KMSAN enabled, the kernel map is shadowed. Account for 1067 * this when sizing maps based on the amount of physical memory 1068 * available. 1069 */ 1070 #if defined(KASAN) 1071 physmem_est = (physmem_est * KASAN_SHADOW_SCALE) / 1072 (KASAN_SHADOW_SCALE + 1); 1073 #elif defined(KMSAN) 1074 physmem_est /= 3; 1075 1076 /* 1077 * KMSAN cannot reliably determine whether buffer data is initialized 1078 * unless it is updated through a KVA mapping. 1079 */ 1080 unmapped_buf_allowed = 0; 1081 #endif 1082 1083 /* 1084 * physmem_est is in pages. Convert it to kilobytes (assumes 1085 * PAGE_SIZE is >= 1K) 1086 */ 1087 physmem_est = physmem_est * (PAGE_SIZE / 1024); 1088 1089 maxbcachebuf_adjust(); 1090 /* 1091 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 1092 * For the first 64MB of ram nominally allocate sufficient buffers to 1093 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 1094 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 1095 * the buffer cache we limit the eventual kva reservation to 1096 * maxbcache bytes. 1097 * 1098 * factor represents the 1/4 x ram conversion. 1099 */ 1100 if (nbuf == 0) { 1101 int factor = 4 * BKVASIZE / 1024; 1102 1103 nbuf = 50; 1104 if (physmem_est > 4096) 1105 nbuf += min((physmem_est - 4096) / factor, 1106 65536 / factor); 1107 if (physmem_est > 65536) 1108 nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 1109 32 * 1024 * 1024 / (factor * 5)); 1110 1111 if (maxbcache && nbuf > maxbcache / BKVASIZE) 1112 nbuf = maxbcache / BKVASIZE; 1113 tuned_nbuf = 1; 1114 } else 1115 tuned_nbuf = 0; 1116 1117 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 1118 maxbuf = (LONG_MAX / 3) / BKVASIZE; 1119 if (nbuf > maxbuf) { 1120 if (!tuned_nbuf) 1121 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 1122 maxbuf); 1123 nbuf = maxbuf; 1124 } 1125 1126 /* 1127 * Ideal allocation size for the transient bio submap is 10% 1128 * of the maximal space buffer map. This roughly corresponds 1129 * to the amount of the buffer mapped for typical UFS load. 1130 * 1131 * Clip the buffer map to reserve space for the transient 1132 * BIOs, if its extent is bigger than 90% (80% on i386) of the 1133 * maximum buffer map extent on the platform. 1134 * 1135 * The fall-back to the maxbuf in case of maxbcache unset, 1136 * allows to not trim the buffer KVA for the architectures 1137 * with ample KVA space. 1138 */ 1139 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 1140 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 1141 buf_sz = (long)nbuf * BKVASIZE; 1142 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 1143 (TRANSIENT_DENOM - 1)) { 1144 /* 1145 * There is more KVA than memory. Do not 1146 * adjust buffer map size, and assign the rest 1147 * of maxbuf to transient map. 1148 */ 1149 biotmap_sz = maxbuf_sz - buf_sz; 1150 } else { 1151 /* 1152 * Buffer map spans all KVA we could afford on 1153 * this platform. Give 10% (20% on i386) of 1154 * the buffer map to the transient bio map. 1155 */ 1156 biotmap_sz = buf_sz / TRANSIENT_DENOM; 1157 buf_sz -= biotmap_sz; 1158 } 1159 if (biotmap_sz / INT_MAX > maxphys) 1160 bio_transient_maxcnt = INT_MAX; 1161 else 1162 bio_transient_maxcnt = biotmap_sz / maxphys; 1163 /* 1164 * Artificially limit to 1024 simultaneous in-flight I/Os 1165 * using the transient mapping. 1166 */ 1167 if (bio_transient_maxcnt > 1024) 1168 bio_transient_maxcnt = 1024; 1169 if (tuned_nbuf) 1170 nbuf = buf_sz / BKVASIZE; 1171 } 1172 1173 if (nswbuf == 0) { 1174 /* 1175 * Pager buffers are allocated for short periods, so scale the 1176 * number of reserved buffers based on the number of CPUs rather 1177 * than amount of memory. 1178 */ 1179 nswbuf = min(nbuf / 4, 32 * mp_ncpus); 1180 if (nswbuf < NSWBUF_MIN) 1181 nswbuf = NSWBUF_MIN; 1182 } 1183 1184 /* 1185 * Reserve space for the buffer cache buffers 1186 */ 1187 buf = (char *)v; 1188 v = (caddr_t)buf + (sizeof(struct buf) + sizeof(vm_page_t) * 1189 atop(maxbcachebuf)) * nbuf; 1190 1191 return (v); 1192 } 1193 1194 /* 1195 * Single global constant for BUF_WMESG, to avoid getting multiple 1196 * references. 1197 */ 1198 static const char buf_wmesg[] = "bufwait"; 1199 1200 /* Initialize the buffer subsystem. Called before use of any buffers. */ 1201 void 1202 bufinit(void) 1203 { 1204 struct buf *bp; 1205 int i; 1206 1207 TSENTER(); 1208 KASSERT(maxbcachebuf >= MAXBSIZE, 1209 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf, 1210 MAXBSIZE)); 1211 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock"); 1212 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 1213 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 1214 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); 1215 1216 unmapped_buf = (caddr_t)kva_alloc(maxphys); 1217 #ifdef INVARIANTS 1218 poisoned_buf = unmapped_buf; 1219 #endif 1220 1221 /* finally, initialize each buffer header and stick on empty q */ 1222 for (i = 0; i < nbuf; i++) { 1223 bp = nbufp(i); 1224 bzero(bp, sizeof(*bp) + sizeof(vm_page_t) * atop(maxbcachebuf)); 1225 bp->b_flags = B_INVAL; 1226 bp->b_rcred = NOCRED; 1227 bp->b_wcred = NOCRED; 1228 bp->b_qindex = QUEUE_NONE; 1229 bp->b_domain = -1; 1230 bp->b_subqueue = mp_maxid + 1; 1231 bp->b_xflags = 0; 1232 bp->b_data = bp->b_kvabase = unmapped_buf; 1233 LIST_INIT(&bp->b_dep); 1234 BUF_LOCKINIT(bp, buf_wmesg); 1235 bq_insert(&bqempty, bp, false); 1236 } 1237 1238 /* 1239 * maxbufspace is the absolute maximum amount of buffer space we are 1240 * allowed to reserve in KVM and in real terms. The absolute maximum 1241 * is nominally used by metadata. hibufspace is the nominal maximum 1242 * used by most other requests. The differential is required to 1243 * ensure that metadata deadlocks don't occur. 1244 * 1245 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 1246 * this may result in KVM fragmentation which is not handled optimally 1247 * by the system. XXX This is less true with vmem. We could use 1248 * PAGE_SIZE. 1249 */ 1250 maxbufspace = (long)nbuf * BKVASIZE; 1251 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10); 1252 lobufspace = (hibufspace / 20) * 19; /* 95% */ 1253 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2; 1254 1255 /* 1256 * Note: The 16 MiB upper limit for hirunningspace was chosen 1257 * arbitrarily and may need further tuning. It corresponds to 1258 * 128 outstanding write IO requests (if IO size is 128 KiB), 1259 * which fits with many RAID controllers' tagged queuing limits. 1260 * The lower 1 MiB limit is the historical upper limit for 1261 * hirunningspace. 1262 */ 1263 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf), 1264 16 * 1024 * 1024), 1024 * 1024); 1265 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf); 1266 1267 /* 1268 * Limit the amount of malloc memory since it is wired permanently into 1269 * the kernel space. Even though this is accounted for in the buffer 1270 * allocation, we don't want the malloced region to grow uncontrolled. 1271 * The malloc scheme improves memory utilization significantly on 1272 * average (small) directories. 1273 */ 1274 maxbufmallocspace = hibufspace / 20; 1275 1276 /* 1277 * Reduce the chance of a deadlock occurring by limiting the number 1278 * of delayed-write dirty buffers we allow to stack up. 1279 */ 1280 hidirtybuffers = nbuf / 4 + 20; 1281 dirtybufthresh = hidirtybuffers * 9 / 10; 1282 /* 1283 * To support extreme low-memory systems, make sure hidirtybuffers 1284 * cannot eat up all available buffer space. This occurs when our 1285 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our 1286 * buffer space assuming BKVASIZE'd buffers. 1287 */ 1288 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 1289 hidirtybuffers >>= 1; 1290 } 1291 lodirtybuffers = hidirtybuffers / 2; 1292 1293 /* 1294 * lofreebuffers should be sufficient to avoid stalling waiting on 1295 * buf headers under heavy utilization. The bufs in per-cpu caches 1296 * are counted as free but will be unavailable to threads executing 1297 * on other cpus. 1298 * 1299 * hifreebuffers is the free target for the bufspace daemon. This 1300 * should be set appropriately to limit work per-iteration. 1301 */ 1302 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus); 1303 hifreebuffers = (3 * lofreebuffers) / 2; 1304 numfreebuffers = nbuf; 1305 1306 /* Setup the kva and free list allocators. */ 1307 vmem_set_reclaim(buffer_arena, bufkva_reclaim); 1308 buf_zone = uma_zcache_create("buf free cache", 1309 sizeof(struct buf) + sizeof(vm_page_t) * atop(maxbcachebuf), 1310 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0); 1311 1312 /* 1313 * Size the clean queue according to the amount of buffer space. 1314 * One queue per-256mb up to the max. More queues gives better 1315 * concurrency but less accurate LRU. 1316 */ 1317 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS); 1318 for (i = 0 ; i < buf_domains; i++) { 1319 struct bufdomain *bd; 1320 1321 bd = &bdomain[i]; 1322 bd_init(bd); 1323 bd->bd_freebuffers = nbuf / buf_domains; 1324 bd->bd_hifreebuffers = hifreebuffers / buf_domains; 1325 bd->bd_lofreebuffers = lofreebuffers / buf_domains; 1326 bd->bd_bufspace = 0; 1327 bd->bd_maxbufspace = maxbufspace / buf_domains; 1328 bd->bd_hibufspace = hibufspace / buf_domains; 1329 bd->bd_lobufspace = lobufspace / buf_domains; 1330 bd->bd_bufspacethresh = bufspacethresh / buf_domains; 1331 bd->bd_numdirtybuffers = 0; 1332 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains; 1333 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains; 1334 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains; 1335 /* Don't allow more than 2% of bufs in the per-cpu caches. */ 1336 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus; 1337 } 1338 getnewbufcalls = counter_u64_alloc(M_WAITOK); 1339 getnewbufrestarts = counter_u64_alloc(M_WAITOK); 1340 mappingrestarts = counter_u64_alloc(M_WAITOK); 1341 numbufallocfails = counter_u64_alloc(M_WAITOK); 1342 notbufdflushes = counter_u64_alloc(M_WAITOK); 1343 buffreekvacnt = counter_u64_alloc(M_WAITOK); 1344 bufdefragcnt = counter_u64_alloc(M_WAITOK); 1345 bufkvaspace = counter_u64_alloc(M_WAITOK); 1346 TSEXIT(); 1347 } 1348 1349 #ifdef INVARIANTS 1350 static inline void 1351 vfs_buf_check_mapped(struct buf *bp) 1352 { 1353 1354 KASSERT(bp->b_kvabase != unmapped_buf, 1355 ("mapped buf: b_kvabase was not updated %p", bp)); 1356 KASSERT(bp->b_data != unmapped_buf, 1357 ("mapped buf: b_data was not updated %p", bp)); 1358 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf + 1359 maxphys, ("b_data + b_offset unmapped %p", bp)); 1360 } 1361 1362 static inline void 1363 vfs_buf_check_unmapped(struct buf *bp) 1364 { 1365 1366 KASSERT(bp->b_data == unmapped_buf, 1367 ("unmapped buf: corrupted b_data %p", bp)); 1368 } 1369 1370 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 1371 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 1372 #else 1373 #define BUF_CHECK_MAPPED(bp) do {} while (0) 1374 #define BUF_CHECK_UNMAPPED(bp) do {} while (0) 1375 #endif 1376 1377 static int 1378 isbufbusy(struct buf *bp) 1379 { 1380 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) || 1381 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI)) 1382 return (1); 1383 return (0); 1384 } 1385 1386 /* 1387 * Shutdown the system cleanly to prepare for reboot, halt, or power off. 1388 */ 1389 void 1390 bufshutdown(int show_busybufs) 1391 { 1392 static int first_buf_printf = 1; 1393 struct buf *bp; 1394 int i, iter, nbusy, pbusy; 1395 #ifndef PREEMPTION 1396 int subiter; 1397 #endif 1398 1399 /* 1400 * Sync filesystems for shutdown 1401 */ 1402 wdog_kern_pat(WD_LASTVAL); 1403 kern_sync(curthread); 1404 1405 /* 1406 * With soft updates, some buffers that are 1407 * written will be remarked as dirty until other 1408 * buffers are written. 1409 */ 1410 for (iter = pbusy = 0; iter < 20; iter++) { 1411 nbusy = 0; 1412 for (i = nbuf - 1; i >= 0; i--) { 1413 bp = nbufp(i); 1414 if (isbufbusy(bp)) 1415 nbusy++; 1416 } 1417 if (nbusy == 0) { 1418 if (first_buf_printf) 1419 printf("All buffers synced."); 1420 break; 1421 } 1422 if (first_buf_printf) { 1423 printf("Syncing disks, buffers remaining... "); 1424 first_buf_printf = 0; 1425 } 1426 printf("%d ", nbusy); 1427 if (nbusy < pbusy) 1428 iter = 0; 1429 pbusy = nbusy; 1430 1431 wdog_kern_pat(WD_LASTVAL); 1432 kern_sync(curthread); 1433 1434 #ifdef PREEMPTION 1435 /* 1436 * Spin for a while to allow interrupt threads to run. 1437 */ 1438 DELAY(50000 * iter); 1439 #else 1440 /* 1441 * Context switch several times to allow interrupt 1442 * threads to run. 1443 */ 1444 for (subiter = 0; subiter < 50 * iter; subiter++) { 1445 sched_relinquish(curthread); 1446 DELAY(1000); 1447 } 1448 #endif 1449 } 1450 printf("\n"); 1451 /* 1452 * Count only busy local buffers to prevent forcing 1453 * a fsck if we're just a client of a wedged NFS server 1454 */ 1455 nbusy = 0; 1456 for (i = nbuf - 1; i >= 0; i--) { 1457 bp = nbufp(i); 1458 if (isbufbusy(bp)) { 1459 #if 0 1460 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */ 1461 if (bp->b_dev == NULL) { 1462 TAILQ_REMOVE(&mountlist, 1463 bp->b_vp->v_mount, mnt_list); 1464 continue; 1465 } 1466 #endif 1467 nbusy++; 1468 if (show_busybufs > 0) { 1469 printf( 1470 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:", 1471 nbusy, bp, bp->b_vp, bp->b_flags, 1472 (intmax_t)bp->b_blkno, 1473 (intmax_t)bp->b_lblkno); 1474 BUF_LOCKPRINTINFO(bp); 1475 if (show_busybufs > 1) 1476 vn_printf(bp->b_vp, 1477 "vnode content: "); 1478 } 1479 } 1480 } 1481 if (nbusy) { 1482 /* 1483 * Failed to sync all blocks. Indicate this and don't 1484 * unmount filesystems (thus forcing an fsck on reboot). 1485 */ 1486 BOOTTRACE("shutdown failed to sync buffers"); 1487 printf("Giving up on %d buffers\n", nbusy); 1488 DELAY(5000000); /* 5 seconds */ 1489 swapoff_all(); 1490 } else { 1491 BOOTTRACE("shutdown sync complete"); 1492 if (!first_buf_printf) 1493 printf("Final sync complete\n"); 1494 1495 /* 1496 * Unmount filesystems and perform swapoff, to quiesce 1497 * the system as much as possible. In particular, no 1498 * I/O should be initiated from top levels since it 1499 * might be abruptly terminated by reset, or otherwise 1500 * erronously handled because other parts of the 1501 * system are disabled. 1502 * 1503 * Swapoff before unmount, because file-backed swap is 1504 * non-operational after unmount of the underlying 1505 * filesystem. 1506 */ 1507 if (!KERNEL_PANICKED()) { 1508 swapoff_all(); 1509 vfs_unmountall(); 1510 } 1511 BOOTTRACE("shutdown unmounted all filesystems"); 1512 } 1513 DELAY(100000); /* wait for console output to finish */ 1514 } 1515 1516 static void 1517 bpmap_qenter(struct buf *bp) 1518 { 1519 1520 BUF_CHECK_MAPPED(bp); 1521 1522 /* 1523 * bp->b_data is relative to bp->b_offset, but 1524 * bp->b_offset may be offset into the first page. 1525 */ 1526 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 1527 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 1528 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 1529 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 1530 } 1531 1532 static inline struct bufdomain * 1533 bufdomain(struct buf *bp) 1534 { 1535 1536 return (&bdomain[bp->b_domain]); 1537 } 1538 1539 static struct bufqueue * 1540 bufqueue(struct buf *bp) 1541 { 1542 1543 switch (bp->b_qindex) { 1544 case QUEUE_NONE: 1545 /* FALLTHROUGH */ 1546 case QUEUE_SENTINEL: 1547 return (NULL); 1548 case QUEUE_EMPTY: 1549 return (&bqempty); 1550 case QUEUE_DIRTY: 1551 return (&bufdomain(bp)->bd_dirtyq); 1552 case QUEUE_CLEAN: 1553 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]); 1554 default: 1555 break; 1556 } 1557 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex); 1558 } 1559 1560 /* 1561 * Return the locked bufqueue that bp is a member of. 1562 */ 1563 static struct bufqueue * 1564 bufqueue_acquire(struct buf *bp) 1565 { 1566 struct bufqueue *bq, *nbq; 1567 1568 /* 1569 * bp can be pushed from a per-cpu queue to the 1570 * cleanq while we're waiting on the lock. Retry 1571 * if the queues don't match. 1572 */ 1573 bq = bufqueue(bp); 1574 BQ_LOCK(bq); 1575 for (;;) { 1576 nbq = bufqueue(bp); 1577 if (bq == nbq) 1578 break; 1579 BQ_UNLOCK(bq); 1580 BQ_LOCK(nbq); 1581 bq = nbq; 1582 } 1583 return (bq); 1584 } 1585 1586 /* 1587 * binsfree: 1588 * 1589 * Insert the buffer into the appropriate free list. Requires a 1590 * locked buffer on entry and buffer is unlocked before return. 1591 */ 1592 static void 1593 binsfree(struct buf *bp, int qindex) 1594 { 1595 struct bufdomain *bd; 1596 struct bufqueue *bq; 1597 1598 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY, 1599 ("binsfree: Invalid qindex %d", qindex)); 1600 BUF_ASSERT_XLOCKED(bp); 1601 1602 /* 1603 * Handle delayed bremfree() processing. 1604 */ 1605 if (bp->b_flags & B_REMFREE) { 1606 if (bp->b_qindex == qindex) { 1607 bp->b_flags |= B_REUSE; 1608 bp->b_flags &= ~B_REMFREE; 1609 BUF_UNLOCK(bp); 1610 return; 1611 } 1612 bq = bufqueue_acquire(bp); 1613 bq_remove(bq, bp); 1614 BQ_UNLOCK(bq); 1615 } 1616 bd = bufdomain(bp); 1617 if (qindex == QUEUE_CLEAN) { 1618 if (bd->bd_lim != 0) 1619 bq = &bd->bd_subq[PCPU_GET(cpuid)]; 1620 else 1621 bq = bd->bd_cleanq; 1622 } else 1623 bq = &bd->bd_dirtyq; 1624 bq_insert(bq, bp, true); 1625 } 1626 1627 /* 1628 * buf_free: 1629 * 1630 * Free a buffer to the buf zone once it no longer has valid contents. 1631 */ 1632 static void 1633 buf_free(struct buf *bp) 1634 { 1635 1636 if (bp->b_flags & B_REMFREE) 1637 bremfreef(bp); 1638 if (bp->b_vflags & BV_BKGRDINPROG) 1639 panic("losing buffer 1"); 1640 if (bp->b_rcred != NOCRED) { 1641 crfree(bp->b_rcred); 1642 bp->b_rcred = NOCRED; 1643 } 1644 if (bp->b_wcred != NOCRED) { 1645 crfree(bp->b_wcred); 1646 bp->b_wcred = NOCRED; 1647 } 1648 if (!LIST_EMPTY(&bp->b_dep)) 1649 buf_deallocate(bp); 1650 bufkva_free(bp); 1651 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1); 1652 MPASS((bp->b_flags & B_MAXPHYS) == 0); 1653 BUF_UNLOCK(bp); 1654 uma_zfree(buf_zone, bp); 1655 } 1656 1657 /* 1658 * buf_import: 1659 * 1660 * Import bufs into the uma cache from the buf list. The system still 1661 * expects a static array of bufs and much of the synchronization 1662 * around bufs assumes type stable storage. As a result, UMA is used 1663 * only as a per-cpu cache of bufs still maintained on a global list. 1664 */ 1665 static int 1666 buf_import(void *arg, void **store, int cnt, int domain, int flags) 1667 { 1668 struct buf *bp; 1669 int i; 1670 1671 BQ_LOCK(&bqempty); 1672 for (i = 0; i < cnt; i++) { 1673 bp = TAILQ_FIRST(&bqempty.bq_queue); 1674 if (bp == NULL) 1675 break; 1676 bq_remove(&bqempty, bp); 1677 store[i] = bp; 1678 } 1679 BQ_UNLOCK(&bqempty); 1680 1681 return (i); 1682 } 1683 1684 /* 1685 * buf_release: 1686 * 1687 * Release bufs from the uma cache back to the buffer queues. 1688 */ 1689 static void 1690 buf_release(void *arg, void **store, int cnt) 1691 { 1692 struct bufqueue *bq; 1693 struct buf *bp; 1694 int i; 1695 1696 bq = &bqempty; 1697 BQ_LOCK(bq); 1698 for (i = 0; i < cnt; i++) { 1699 bp = store[i]; 1700 /* Inline bq_insert() to batch locking. */ 1701 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1702 bp->b_flags &= ~(B_AGE | B_REUSE); 1703 bq->bq_len++; 1704 bp->b_qindex = bq->bq_index; 1705 } 1706 BQ_UNLOCK(bq); 1707 } 1708 1709 /* 1710 * buf_alloc: 1711 * 1712 * Allocate an empty buffer header. 1713 */ 1714 static struct buf * 1715 buf_alloc(struct bufdomain *bd) 1716 { 1717 struct buf *bp; 1718 int freebufs, error; 1719 1720 /* 1721 * We can only run out of bufs in the buf zone if the average buf 1722 * is less than BKVASIZE. In this case the actual wait/block will 1723 * come from buf_reycle() failing to flush one of these small bufs. 1724 */ 1725 bp = NULL; 1726 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1); 1727 if (freebufs > 0) 1728 bp = uma_zalloc(buf_zone, M_NOWAIT); 1729 if (bp == NULL) { 1730 atomic_add_int(&bd->bd_freebuffers, 1); 1731 bufspace_daemon_wakeup(bd); 1732 counter_u64_add(numbufallocfails, 1); 1733 return (NULL); 1734 } 1735 /* 1736 * Wake-up the bufspace daemon on transition below threshold. 1737 */ 1738 if (freebufs == bd->bd_lofreebuffers) 1739 bufspace_daemon_wakeup(bd); 1740 1741 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWITNESS, NULL); 1742 KASSERT(error == 0, ("%s: BUF_LOCK on free buf %p: %d.", __func__, bp, 1743 error)); 1744 (void)error; 1745 1746 KASSERT(bp->b_vp == NULL, 1747 ("bp: %p still has vnode %p.", bp, bp->b_vp)); 1748 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0, 1749 ("invalid buffer %p flags %#x", bp, bp->b_flags)); 1750 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1751 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 1752 KASSERT(bp->b_npages == 0, 1753 ("bp: %p still has %d vm pages\n", bp, bp->b_npages)); 1754 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp)); 1755 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp)); 1756 MPASS((bp->b_flags & B_MAXPHYS) == 0); 1757 1758 bp->b_domain = BD_DOMAIN(bd); 1759 bp->b_flags = 0; 1760 bp->b_ioflags = 0; 1761 bp->b_xflags = 0; 1762 bp->b_vflags = 0; 1763 bp->b_vp = NULL; 1764 bp->b_blkno = bp->b_lblkno = 0; 1765 bp->b_offset = NOOFFSET; 1766 bp->b_iodone = 0; 1767 bp->b_error = 0; 1768 bp->b_resid = 0; 1769 bp->b_bcount = 0; 1770 bp->b_npages = 0; 1771 bp->b_dirtyoff = bp->b_dirtyend = 0; 1772 bp->b_bufobj = NULL; 1773 bp->b_data = bp->b_kvabase = unmapped_buf; 1774 bp->b_fsprivate1 = NULL; 1775 bp->b_fsprivate2 = NULL; 1776 bp->b_fsprivate3 = NULL; 1777 LIST_INIT(&bp->b_dep); 1778 1779 return (bp); 1780 } 1781 1782 /* 1783 * buf_recycle: 1784 * 1785 * Free a buffer from the given bufqueue. kva controls whether the 1786 * freed buf must own some kva resources. This is used for 1787 * defragmenting. 1788 */ 1789 static int 1790 buf_recycle(struct bufdomain *bd, bool kva) 1791 { 1792 struct bufqueue *bq; 1793 struct buf *bp, *nbp; 1794 1795 if (kva) 1796 counter_u64_add(bufdefragcnt, 1); 1797 nbp = NULL; 1798 bq = bd->bd_cleanq; 1799 BQ_LOCK(bq); 1800 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd), 1801 ("buf_recycle: Locks don't match")); 1802 nbp = TAILQ_FIRST(&bq->bq_queue); 1803 1804 /* 1805 * Run scan, possibly freeing data and/or kva mappings on the fly 1806 * depending. 1807 */ 1808 while ((bp = nbp) != NULL) { 1809 /* 1810 * Calculate next bp (we can only use it if we do not 1811 * release the bqlock). 1812 */ 1813 nbp = TAILQ_NEXT(bp, b_freelist); 1814 1815 /* 1816 * If we are defragging then we need a buffer with 1817 * some kva to reclaim. 1818 */ 1819 if (kva && bp->b_kvasize == 0) 1820 continue; 1821 1822 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1823 continue; 1824 1825 /* 1826 * Implement a second chance algorithm for frequently 1827 * accessed buffers. 1828 */ 1829 if ((bp->b_flags & B_REUSE) != 0) { 1830 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1831 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1832 bp->b_flags &= ~B_REUSE; 1833 BUF_UNLOCK(bp); 1834 continue; 1835 } 1836 1837 /* 1838 * Skip buffers with background writes in progress. 1839 */ 1840 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 1841 BUF_UNLOCK(bp); 1842 continue; 1843 } 1844 1845 KASSERT(bp->b_qindex == QUEUE_CLEAN, 1846 ("buf_recycle: inconsistent queue %d bp %p", 1847 bp->b_qindex, bp)); 1848 KASSERT(bp->b_domain == BD_DOMAIN(bd), 1849 ("getnewbuf: queue domain %d doesn't match request %d", 1850 bp->b_domain, (int)BD_DOMAIN(bd))); 1851 /* 1852 * NOTE: nbp is now entirely invalid. We can only restart 1853 * the scan from this point on. 1854 */ 1855 bq_remove(bq, bp); 1856 BQ_UNLOCK(bq); 1857 1858 /* 1859 * Requeue the background write buffer with error and 1860 * restart the scan. 1861 */ 1862 if ((bp->b_vflags & BV_BKGRDERR) != 0) { 1863 bqrelse(bp); 1864 BQ_LOCK(bq); 1865 nbp = TAILQ_FIRST(&bq->bq_queue); 1866 continue; 1867 } 1868 bp->b_flags |= B_INVAL; 1869 brelse(bp); 1870 return (0); 1871 } 1872 bd->bd_wanted = 1; 1873 BQ_UNLOCK(bq); 1874 1875 return (ENOBUFS); 1876 } 1877 1878 /* 1879 * bremfree: 1880 * 1881 * Mark the buffer for removal from the appropriate free list. 1882 * 1883 */ 1884 void 1885 bremfree(struct buf *bp) 1886 { 1887 1888 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1889 KASSERT((bp->b_flags & B_REMFREE) == 0, 1890 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1891 KASSERT(bp->b_qindex != QUEUE_NONE, 1892 ("bremfree: buffer %p not on a queue.", bp)); 1893 BUF_ASSERT_XLOCKED(bp); 1894 1895 bp->b_flags |= B_REMFREE; 1896 } 1897 1898 /* 1899 * bremfreef: 1900 * 1901 * Force an immediate removal from a free list. Used only in nfs when 1902 * it abuses the b_freelist pointer. 1903 */ 1904 void 1905 bremfreef(struct buf *bp) 1906 { 1907 struct bufqueue *bq; 1908 1909 bq = bufqueue_acquire(bp); 1910 bq_remove(bq, bp); 1911 BQ_UNLOCK(bq); 1912 } 1913 1914 static void 1915 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname) 1916 { 1917 1918 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF); 1919 TAILQ_INIT(&bq->bq_queue); 1920 bq->bq_len = 0; 1921 bq->bq_index = qindex; 1922 bq->bq_subqueue = subqueue; 1923 } 1924 1925 static void 1926 bd_init(struct bufdomain *bd) 1927 { 1928 int i; 1929 1930 /* Per-CPU clean buf queues, plus one global queue. */ 1931 bd->bd_subq = mallocarray(mp_maxid + 2, sizeof(struct bufqueue), 1932 M_BIOBUF, M_WAITOK | M_ZERO); 1933 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1]; 1934 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock"); 1935 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock"); 1936 for (i = 0; i <= mp_maxid; i++) 1937 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i, 1938 "bufq clean subqueue lock"); 1939 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF); 1940 } 1941 1942 /* 1943 * bq_remove: 1944 * 1945 * Removes a buffer from the free list, must be called with the 1946 * correct qlock held. 1947 */ 1948 static void 1949 bq_remove(struct bufqueue *bq, struct buf *bp) 1950 { 1951 1952 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X", 1953 bp, bp->b_vp, bp->b_flags); 1954 KASSERT(bp->b_qindex != QUEUE_NONE, 1955 ("bq_remove: buffer %p not on a queue.", bp)); 1956 KASSERT(bufqueue(bp) == bq, 1957 ("bq_remove: Remove buffer %p from wrong queue.", bp)); 1958 1959 BQ_ASSERT_LOCKED(bq); 1960 if (bp->b_qindex != QUEUE_EMPTY) { 1961 BUF_ASSERT_XLOCKED(bp); 1962 } 1963 KASSERT(bq->bq_len >= 1, 1964 ("queue %d underflow", bp->b_qindex)); 1965 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1966 bq->bq_len--; 1967 bp->b_qindex = QUEUE_NONE; 1968 bp->b_flags &= ~(B_REMFREE | B_REUSE); 1969 } 1970 1971 static void 1972 bd_flush(struct bufdomain *bd, struct bufqueue *bq) 1973 { 1974 struct buf *bp; 1975 1976 BQ_ASSERT_LOCKED(bq); 1977 if (bq != bd->bd_cleanq) { 1978 BD_LOCK(bd); 1979 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) { 1980 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1981 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp, 1982 b_freelist); 1983 bp->b_subqueue = bd->bd_cleanq->bq_subqueue; 1984 } 1985 bd->bd_cleanq->bq_len += bq->bq_len; 1986 bq->bq_len = 0; 1987 } 1988 if (bd->bd_wanted) { 1989 bd->bd_wanted = 0; 1990 wakeup(&bd->bd_wanted); 1991 } 1992 if (bq != bd->bd_cleanq) 1993 BD_UNLOCK(bd); 1994 } 1995 1996 static int 1997 bd_flushall(struct bufdomain *bd) 1998 { 1999 struct bufqueue *bq; 2000 int flushed; 2001 int i; 2002 2003 if (bd->bd_lim == 0) 2004 return (0); 2005 flushed = 0; 2006 for (i = 0; i <= mp_maxid; i++) { 2007 bq = &bd->bd_subq[i]; 2008 if (bq->bq_len == 0) 2009 continue; 2010 BQ_LOCK(bq); 2011 bd_flush(bd, bq); 2012 BQ_UNLOCK(bq); 2013 flushed++; 2014 } 2015 2016 return (flushed); 2017 } 2018 2019 static void 2020 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock) 2021 { 2022 struct bufdomain *bd; 2023 2024 if (bp->b_qindex != QUEUE_NONE) 2025 panic("bq_insert: free buffer %p onto another queue?", bp); 2026 2027 bd = bufdomain(bp); 2028 if (bp->b_flags & B_AGE) { 2029 /* Place this buf directly on the real queue. */ 2030 if (bq->bq_index == QUEUE_CLEAN) 2031 bq = bd->bd_cleanq; 2032 BQ_LOCK(bq); 2033 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist); 2034 } else { 2035 BQ_LOCK(bq); 2036 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 2037 } 2038 bp->b_flags &= ~(B_AGE | B_REUSE); 2039 bq->bq_len++; 2040 bp->b_qindex = bq->bq_index; 2041 bp->b_subqueue = bq->bq_subqueue; 2042 2043 /* 2044 * Unlock before we notify so that we don't wakeup a waiter that 2045 * fails a trylock on the buf and sleeps again. 2046 */ 2047 if (unlock) 2048 BUF_UNLOCK(bp); 2049 2050 if (bp->b_qindex == QUEUE_CLEAN) { 2051 /* 2052 * Flush the per-cpu queue and notify any waiters. 2053 */ 2054 if (bd->bd_wanted || (bq != bd->bd_cleanq && 2055 bq->bq_len >= bd->bd_lim)) 2056 bd_flush(bd, bq); 2057 } 2058 BQ_UNLOCK(bq); 2059 } 2060 2061 /* 2062 * bufkva_free: 2063 * 2064 * Free the kva allocation for a buffer. 2065 * 2066 */ 2067 static void 2068 bufkva_free(struct buf *bp) 2069 { 2070 2071 #ifdef INVARIANTS 2072 if (bp->b_kvasize == 0) { 2073 KASSERT(bp->b_kvabase == unmapped_buf && 2074 bp->b_data == unmapped_buf, 2075 ("Leaked KVA space on %p", bp)); 2076 } else if (buf_mapped(bp)) 2077 BUF_CHECK_MAPPED(bp); 2078 else 2079 BUF_CHECK_UNMAPPED(bp); 2080 #endif 2081 if (bp->b_kvasize == 0) 2082 return; 2083 2084 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize); 2085 counter_u64_add(bufkvaspace, -bp->b_kvasize); 2086 counter_u64_add(buffreekvacnt, 1); 2087 bp->b_data = bp->b_kvabase = unmapped_buf; 2088 bp->b_kvasize = 0; 2089 } 2090 2091 /* 2092 * bufkva_alloc: 2093 * 2094 * Allocate the buffer KVA and set b_kvasize and b_kvabase. 2095 */ 2096 static int 2097 bufkva_alloc(struct buf *bp, int maxsize, int gbflags) 2098 { 2099 vm_offset_t addr; 2100 int error; 2101 2102 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0, 2103 ("Invalid gbflags 0x%x in %s", gbflags, __func__)); 2104 MPASS((bp->b_flags & B_MAXPHYS) == 0); 2105 KASSERT(maxsize <= maxbcachebuf, 2106 ("bufkva_alloc kva too large %d %u", maxsize, maxbcachebuf)); 2107 2108 bufkva_free(bp); 2109 2110 addr = 0; 2111 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr); 2112 if (error != 0) { 2113 /* 2114 * Buffer map is too fragmented. Request the caller 2115 * to defragment the map. 2116 */ 2117 return (error); 2118 } 2119 bp->b_kvabase = (caddr_t)addr; 2120 bp->b_kvasize = maxsize; 2121 counter_u64_add(bufkvaspace, bp->b_kvasize); 2122 if ((gbflags & GB_UNMAPPED) != 0) { 2123 bp->b_data = unmapped_buf; 2124 BUF_CHECK_UNMAPPED(bp); 2125 } else { 2126 bp->b_data = bp->b_kvabase; 2127 BUF_CHECK_MAPPED(bp); 2128 } 2129 return (0); 2130 } 2131 2132 /* 2133 * bufkva_reclaim: 2134 * 2135 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem 2136 * callback that fires to avoid returning failure. 2137 */ 2138 static void 2139 bufkva_reclaim(vmem_t *vmem, int flags) 2140 { 2141 bool done; 2142 int q; 2143 int i; 2144 2145 done = false; 2146 for (i = 0; i < 5; i++) { 2147 for (q = 0; q < buf_domains; q++) 2148 if (buf_recycle(&bdomain[q], true) != 0) 2149 done = true; 2150 if (done) 2151 break; 2152 } 2153 return; 2154 } 2155 2156 /* 2157 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 2158 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 2159 * the buffer is valid and we do not have to do anything. 2160 */ 2161 static void 2162 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt, 2163 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *)) 2164 { 2165 struct buf *rabp; 2166 struct thread *td; 2167 int i; 2168 2169 td = curthread; 2170 2171 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 2172 if (inmem(vp, *rablkno)) 2173 continue; 2174 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 2175 if ((rabp->b_flags & B_CACHE) != 0) { 2176 brelse(rabp); 2177 continue; 2178 } 2179 #ifdef RACCT 2180 if (racct_enable) { 2181 PROC_LOCK(curproc); 2182 racct_add_buf(curproc, rabp, 0); 2183 PROC_UNLOCK(curproc); 2184 } 2185 #endif /* RACCT */ 2186 td->td_ru.ru_inblock++; 2187 rabp->b_flags |= B_ASYNC; 2188 rabp->b_flags &= ~B_INVAL; 2189 if ((flags & GB_CKHASH) != 0) { 2190 rabp->b_flags |= B_CKHASH; 2191 rabp->b_ckhashcalc = ckhashfunc; 2192 } 2193 rabp->b_ioflags &= ~BIO_ERROR; 2194 rabp->b_iocmd = BIO_READ; 2195 if (rabp->b_rcred == NOCRED && cred != NOCRED) 2196 rabp->b_rcred = crhold(cred); 2197 vfs_busy_pages(rabp, 0); 2198 BUF_KERNPROC(rabp); 2199 rabp->b_iooffset = dbtob(rabp->b_blkno); 2200 bstrategy(rabp); 2201 } 2202 } 2203 2204 /* 2205 * Entry point for bread() and breadn() via #defines in sys/buf.h. 2206 * 2207 * Get a buffer with the specified data. Look in the cache first. We 2208 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 2209 * is set, the buffer is valid and we do not have to do anything, see 2210 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 2211 * 2212 * Always return a NULL buffer pointer (in bpp) when returning an error. 2213 * 2214 * The blkno parameter is the logical block being requested. Normally 2215 * the mapping of logical block number to disk block address is done 2216 * by calling VOP_BMAP(). However, if the mapping is already known, the 2217 * disk block address can be passed using the dblkno parameter. If the 2218 * disk block address is not known, then the same value should be passed 2219 * for blkno and dblkno. 2220 */ 2221 int 2222 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, 2223 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags, 2224 void (*ckhashfunc)(struct buf *), struct buf **bpp) 2225 { 2226 struct buf *bp; 2227 struct thread *td; 2228 int error, readwait, rv; 2229 2230 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 2231 td = curthread; 2232 /* 2233 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags 2234 * are specified. 2235 */ 2236 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp); 2237 if (error != 0) { 2238 *bpp = NULL; 2239 return (error); 2240 } 2241 KASSERT(blkno == bp->b_lblkno, 2242 ("getblkx returned buffer for blkno %jd instead of blkno %jd", 2243 (intmax_t)bp->b_lblkno, (intmax_t)blkno)); 2244 flags &= ~GB_NOSPARSE; 2245 *bpp = bp; 2246 2247 /* 2248 * If not found in cache, do some I/O 2249 */ 2250 readwait = 0; 2251 if ((bp->b_flags & B_CACHE) == 0) { 2252 #ifdef RACCT 2253 if (racct_enable) { 2254 PROC_LOCK(td->td_proc); 2255 racct_add_buf(td->td_proc, bp, 0); 2256 PROC_UNLOCK(td->td_proc); 2257 } 2258 #endif /* RACCT */ 2259 td->td_ru.ru_inblock++; 2260 bp->b_iocmd = BIO_READ; 2261 bp->b_flags &= ~B_INVAL; 2262 if ((flags & GB_CKHASH) != 0) { 2263 bp->b_flags |= B_CKHASH; 2264 bp->b_ckhashcalc = ckhashfunc; 2265 } 2266 if ((flags & GB_CVTENXIO) != 0) 2267 bp->b_xflags |= BX_CVTENXIO; 2268 bp->b_ioflags &= ~BIO_ERROR; 2269 if (bp->b_rcred == NOCRED && cred != NOCRED) 2270 bp->b_rcred = crhold(cred); 2271 vfs_busy_pages(bp, 0); 2272 bp->b_iooffset = dbtob(bp->b_blkno); 2273 bstrategy(bp); 2274 ++readwait; 2275 } 2276 2277 /* 2278 * Attempt to initiate asynchronous I/O on read-ahead blocks. 2279 */ 2280 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc); 2281 2282 rv = 0; 2283 if (readwait) { 2284 rv = bufwait(bp); 2285 if (rv != 0) { 2286 brelse(bp); 2287 *bpp = NULL; 2288 } 2289 } 2290 return (rv); 2291 } 2292 2293 /* 2294 * Write, release buffer on completion. (Done by iodone 2295 * if async). Do not bother writing anything if the buffer 2296 * is invalid. 2297 * 2298 * Note that we set B_CACHE here, indicating that buffer is 2299 * fully valid and thus cacheable. This is true even of NFS 2300 * now so we set it generally. This could be set either here 2301 * or in biodone() since the I/O is synchronous. We put it 2302 * here. 2303 */ 2304 int 2305 bufwrite(struct buf *bp) 2306 { 2307 struct vnode *vp; 2308 long space; 2309 int oldflags, retval; 2310 bool vp_md; 2311 2312 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2313 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) { 2314 bp->b_flags |= B_INVAL | B_RELBUF; 2315 bp->b_flags &= ~B_CACHE; 2316 brelse(bp); 2317 return (ENXIO); 2318 } 2319 if ((bp->b_flags & B_INVAL) != 0) { 2320 brelse(bp); 2321 return (0); 2322 } 2323 2324 if ((bp->b_flags & B_BARRIER) != 0) 2325 atomic_add_long(&barrierwrites, 1); 2326 2327 oldflags = bp->b_flags; 2328 2329 KASSERT((bp->b_vflags & BV_BKGRDINPROG) == 0, 2330 ("FFS background buffer should not get here %p", bp)); 2331 2332 vp = bp->b_vp; 2333 vp_md = vp != NULL && (vp->v_vflag & VV_MD) != 0; 2334 2335 /* 2336 * Mark the buffer clean. Increment the bufobj write count 2337 * before bundirty() call, to prevent other thread from seeing 2338 * empty dirty list and zero counter for writes in progress, 2339 * falsely indicating that the bufobj is clean. 2340 */ 2341 bufobj_wref(bp->b_bufobj); 2342 bundirty(bp); 2343 2344 bp->b_flags &= ~B_DONE; 2345 bp->b_ioflags &= ~BIO_ERROR; 2346 bp->b_flags |= B_CACHE; 2347 bp->b_iocmd = BIO_WRITE; 2348 2349 vfs_busy_pages(bp, 1); 2350 2351 /* 2352 * Normal bwrites pipeline writes 2353 */ 2354 bp->b_runningbufspace = bp->b_bufsize; 2355 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); 2356 2357 #ifdef RACCT 2358 if (racct_enable) { 2359 PROC_LOCK(curproc); 2360 racct_add_buf(curproc, bp, 1); 2361 PROC_UNLOCK(curproc); 2362 } 2363 #endif /* RACCT */ 2364 curthread->td_ru.ru_oublock++; 2365 if ((oldflags & B_ASYNC) != 0) 2366 BUF_KERNPROC(bp); 2367 bp->b_iooffset = dbtob(bp->b_blkno); 2368 buf_track(bp, __func__); 2369 bstrategy(bp); 2370 2371 if ((oldflags & B_ASYNC) == 0) { 2372 retval = bufwait(bp); 2373 brelse(bp); 2374 return (retval); 2375 } else if (space > hirunningspace) { 2376 /* 2377 * Don't allow the async write to saturate the I/O 2378 * system. We do not block here if it is the update 2379 * or syncer daemon trying to clean up as that can 2380 * lead to deadlock. 2381 */ 2382 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 2383 waitrunningbufspace(); 2384 } 2385 2386 return (0); 2387 } 2388 2389 void 2390 bufbdflush(struct bufobj *bo, struct buf *bp) 2391 { 2392 struct buf *nbp; 2393 struct bufdomain *bd; 2394 2395 bd = &bdomain[bo->bo_domain]; 2396 if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh + 10) { 2397 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 2398 altbufferflushes++; 2399 } else if (bo->bo_dirty.bv_cnt > bd->bd_dirtybufthresh) { 2400 BO_LOCK(bo); 2401 /* 2402 * Try to find a buffer to flush. 2403 */ 2404 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 2405 if ((nbp->b_vflags & BV_BKGRDINPROG) || 2406 BUF_LOCK(nbp, 2407 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 2408 continue; 2409 if (bp == nbp) 2410 panic("bdwrite: found ourselves"); 2411 BO_UNLOCK(bo); 2412 /* Don't countdeps with the bo lock held. */ 2413 if (buf_countdeps(nbp, 0)) { 2414 BO_LOCK(bo); 2415 BUF_UNLOCK(nbp); 2416 continue; 2417 } 2418 if (nbp->b_flags & B_CLUSTEROK) { 2419 vfs_bio_awrite(nbp); 2420 } else { 2421 bremfree(nbp); 2422 bawrite(nbp); 2423 } 2424 dirtybufferflushes++; 2425 break; 2426 } 2427 if (nbp == NULL) 2428 BO_UNLOCK(bo); 2429 } 2430 } 2431 2432 /* 2433 * Delayed write. (Buffer is marked dirty). Do not bother writing 2434 * anything if the buffer is marked invalid. 2435 * 2436 * Note that since the buffer must be completely valid, we can safely 2437 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 2438 * biodone() in order to prevent getblk from writing the buffer 2439 * out synchronously. 2440 */ 2441 void 2442 bdwrite(struct buf *bp) 2443 { 2444 struct thread *td = curthread; 2445 struct vnode *vp; 2446 struct bufobj *bo; 2447 2448 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2449 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2450 KASSERT((bp->b_flags & B_BARRIER) == 0, 2451 ("Barrier request in delayed write %p", bp)); 2452 2453 if (bp->b_flags & B_INVAL) { 2454 brelse(bp); 2455 return; 2456 } 2457 2458 /* 2459 * If we have too many dirty buffers, don't create any more. 2460 * If we are wildly over our limit, then force a complete 2461 * cleanup. Otherwise, just keep the situation from getting 2462 * out of control. Note that we have to avoid a recursive 2463 * disaster and not try to clean up after our own cleanup! 2464 */ 2465 vp = bp->b_vp; 2466 bo = bp->b_bufobj; 2467 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 2468 td->td_pflags |= TDP_INBDFLUSH; 2469 BO_BDFLUSH(bo, bp); 2470 td->td_pflags &= ~TDP_INBDFLUSH; 2471 } else 2472 recursiveflushes++; 2473 2474 bdirty(bp); 2475 /* 2476 * Set B_CACHE, indicating that the buffer is fully valid. This is 2477 * true even of NFS now. 2478 */ 2479 bp->b_flags |= B_CACHE; 2480 2481 /* 2482 * This bmap keeps the system from needing to do the bmap later, 2483 * perhaps when the system is attempting to do a sync. Since it 2484 * is likely that the indirect block -- or whatever other datastructure 2485 * that the filesystem needs is still in memory now, it is a good 2486 * thing to do this. Note also, that if the pageout daemon is 2487 * requesting a sync -- there might not be enough memory to do 2488 * the bmap then... So, this is important to do. 2489 */ 2490 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 2491 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 2492 } 2493 2494 buf_track(bp, __func__); 2495 2496 /* 2497 * Set the *dirty* buffer range based upon the VM system dirty 2498 * pages. 2499 * 2500 * Mark the buffer pages as clean. We need to do this here to 2501 * satisfy the vnode_pager and the pageout daemon, so that it 2502 * thinks that the pages have been "cleaned". Note that since 2503 * the pages are in a delayed write buffer -- the VFS layer 2504 * "will" see that the pages get written out on the next sync, 2505 * or perhaps the cluster will be completed. 2506 */ 2507 vfs_clean_pages_dirty_buf(bp); 2508 bqrelse(bp); 2509 2510 /* 2511 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 2512 * due to the softdep code. 2513 */ 2514 } 2515 2516 /* 2517 * bdirty: 2518 * 2519 * Turn buffer into delayed write request. We must clear BIO_READ and 2520 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 2521 * itself to properly update it in the dirty/clean lists. We mark it 2522 * B_DONE to ensure that any asynchronization of the buffer properly 2523 * clears B_DONE ( else a panic will occur later ). 2524 * 2525 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 2526 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 2527 * should only be called if the buffer is known-good. 2528 * 2529 * Since the buffer is not on a queue, we do not update the numfreebuffers 2530 * count. 2531 * 2532 * The buffer must be on QUEUE_NONE. 2533 */ 2534 void 2535 bdirty(struct buf *bp) 2536 { 2537 2538 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 2539 bp, bp->b_vp, bp->b_flags); 2540 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2541 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2542 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2543 bp->b_flags &= ~(B_RELBUF); 2544 bp->b_iocmd = BIO_WRITE; 2545 2546 if ((bp->b_flags & B_DELWRI) == 0) { 2547 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 2548 reassignbuf(bp); 2549 bdirtyadd(bp); 2550 } 2551 } 2552 2553 /* 2554 * bundirty: 2555 * 2556 * Clear B_DELWRI for buffer. 2557 * 2558 * Since the buffer is not on a queue, we do not update the numfreebuffers 2559 * count. 2560 * 2561 * The buffer must be on QUEUE_NONE. 2562 */ 2563 2564 void 2565 bundirty(struct buf *bp) 2566 { 2567 2568 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2569 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2570 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2571 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2572 2573 if (bp->b_flags & B_DELWRI) { 2574 bp->b_flags &= ~B_DELWRI; 2575 reassignbuf(bp); 2576 bdirtysub(bp); 2577 } 2578 /* 2579 * Since it is now being written, we can clear its deferred write flag. 2580 */ 2581 bp->b_flags &= ~B_DEFERRED; 2582 } 2583 2584 /* 2585 * bawrite: 2586 * 2587 * Asynchronous write. Start output on a buffer, but do not wait for 2588 * it to complete. The buffer is released when the output completes. 2589 * 2590 * bwrite() ( or the VOP routine anyway ) is responsible for handling 2591 * B_INVAL buffers. Not us. 2592 */ 2593 void 2594 bawrite(struct buf *bp) 2595 { 2596 2597 bp->b_flags |= B_ASYNC; 2598 (void) bwrite(bp); 2599 } 2600 2601 /* 2602 * babarrierwrite: 2603 * 2604 * Asynchronous barrier write. Start output on a buffer, but do not 2605 * wait for it to complete. Place a write barrier after this write so 2606 * that this buffer and all buffers written before it are committed to 2607 * the disk before any buffers written after this write are committed 2608 * to the disk. The buffer is released when the output completes. 2609 */ 2610 void 2611 babarrierwrite(struct buf *bp) 2612 { 2613 2614 bp->b_flags |= B_ASYNC | B_BARRIER; 2615 (void) bwrite(bp); 2616 } 2617 2618 /* 2619 * bbarrierwrite: 2620 * 2621 * Synchronous barrier write. Start output on a buffer and wait for 2622 * it to complete. Place a write barrier after this write so that 2623 * this buffer and all buffers written before it are committed to 2624 * the disk before any buffers written after this write are committed 2625 * to the disk. The buffer is released when the output completes. 2626 */ 2627 int 2628 bbarrierwrite(struct buf *bp) 2629 { 2630 2631 bp->b_flags |= B_BARRIER; 2632 return (bwrite(bp)); 2633 } 2634 2635 /* 2636 * bwillwrite: 2637 * 2638 * Called prior to the locking of any vnodes when we are expecting to 2639 * write. We do not want to starve the buffer cache with too many 2640 * dirty buffers so we block here. By blocking prior to the locking 2641 * of any vnodes we attempt to avoid the situation where a locked vnode 2642 * prevents the various system daemons from flushing related buffers. 2643 */ 2644 void 2645 bwillwrite(void) 2646 { 2647 2648 if (buf_dirty_count_severe()) { 2649 mtx_lock(&bdirtylock); 2650 while (buf_dirty_count_severe()) { 2651 bdirtywait = 1; 2652 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), 2653 "flswai", 0); 2654 } 2655 mtx_unlock(&bdirtylock); 2656 } 2657 } 2658 2659 /* 2660 * Return true if we have too many dirty buffers. 2661 */ 2662 int 2663 buf_dirty_count_severe(void) 2664 { 2665 2666 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty)); 2667 } 2668 2669 /* 2670 * brelse: 2671 * 2672 * Release a busy buffer and, if requested, free its resources. The 2673 * buffer will be stashed in the appropriate bufqueue[] allowing it 2674 * to be accessed later as a cache entity or reused for other purposes. 2675 */ 2676 void 2677 brelse(struct buf *bp) 2678 { 2679 struct mount *v_mnt; 2680 int qindex; 2681 2682 /* 2683 * Many functions erroneously call brelse with a NULL bp under rare 2684 * error conditions. Simply return when called with a NULL bp. 2685 */ 2686 if (bp == NULL) 2687 return; 2688 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 2689 bp, bp->b_vp, bp->b_flags); 2690 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2691 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2692 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0, 2693 ("brelse: non-VMIO buffer marked NOREUSE")); 2694 2695 if (BUF_LOCKRECURSED(bp)) { 2696 /* 2697 * Do not process, in particular, do not handle the 2698 * B_INVAL/B_RELBUF and do not release to free list. 2699 */ 2700 BUF_UNLOCK(bp); 2701 return; 2702 } 2703 2704 if (bp->b_flags & B_MANAGED) { 2705 bqrelse(bp); 2706 return; 2707 } 2708 2709 if (LIST_EMPTY(&bp->b_dep)) { 2710 bp->b_flags &= ~B_IOSTARTED; 2711 } else { 2712 KASSERT((bp->b_flags & B_IOSTARTED) == 0, 2713 ("brelse: SU io not finished bp %p", bp)); 2714 } 2715 2716 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { 2717 BO_LOCK(bp->b_bufobj); 2718 bp->b_vflags &= ~BV_BKGRDERR; 2719 BO_UNLOCK(bp->b_bufobj); 2720 bdirty(bp); 2721 } 2722 2723 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 2724 (bp->b_flags & B_INVALONERR)) { 2725 /* 2726 * Forced invalidation of dirty buffer contents, to be used 2727 * after a failed write in the rare case that the loss of the 2728 * contents is acceptable. The buffer is invalidated and 2729 * freed. 2730 */ 2731 bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE; 2732 bp->b_flags &= ~(B_ASYNC | B_CACHE); 2733 } 2734 2735 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 2736 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) && 2737 !(bp->b_flags & B_INVAL)) { 2738 /* 2739 * Failed write, redirty. All errors except ENXIO (which 2740 * means the device is gone) are treated as being 2741 * transient. 2742 * 2743 * XXX Treating EIO as transient is not correct; the 2744 * contract with the local storage device drivers is that 2745 * they will only return EIO once the I/O is no longer 2746 * retriable. Network I/O also respects this through the 2747 * guarantees of TCP and/or the internal retries of NFS. 2748 * ENOMEM might be transient, but we also have no way of 2749 * knowing when its ok to retry/reschedule. In general, 2750 * this entire case should be made obsolete through better 2751 * error handling/recovery and resource scheduling. 2752 * 2753 * Do this also for buffers that failed with ENXIO, but have 2754 * non-empty dependencies - the soft updates code might need 2755 * to access the buffer to untangle them. 2756 * 2757 * Must clear BIO_ERROR to prevent pages from being scrapped. 2758 */ 2759 bp->b_ioflags &= ~BIO_ERROR; 2760 bdirty(bp); 2761 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 2762 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 2763 /* 2764 * Either a failed read I/O, or we were asked to free or not 2765 * cache the buffer, or we failed to write to a device that's 2766 * no longer present. 2767 */ 2768 bp->b_flags |= B_INVAL; 2769 if (!LIST_EMPTY(&bp->b_dep)) 2770 buf_deallocate(bp); 2771 if (bp->b_flags & B_DELWRI) 2772 bdirtysub(bp); 2773 bp->b_flags &= ~(B_DELWRI | B_CACHE); 2774 if ((bp->b_flags & B_VMIO) == 0) { 2775 allocbuf(bp, 0); 2776 if (bp->b_vp) 2777 brelvp(bp); 2778 } 2779 } 2780 2781 /* 2782 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate() 2783 * is called with B_DELWRI set, the underlying pages may wind up 2784 * getting freed causing a previous write (bdwrite()) to get 'lost' 2785 * because pages associated with a B_DELWRI bp are marked clean. 2786 * 2787 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even 2788 * if B_DELWRI is set. 2789 */ 2790 if (bp->b_flags & B_DELWRI) 2791 bp->b_flags &= ~B_RELBUF; 2792 2793 /* 2794 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 2795 * constituted, not even NFS buffers now. Two flags effect this. If 2796 * B_INVAL, the struct buf is invalidated but the VM object is kept 2797 * around ( i.e. so it is trivial to reconstitute the buffer later ). 2798 * 2799 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 2800 * invalidated. BIO_ERROR cannot be set for a failed write unless the 2801 * buffer is also B_INVAL because it hits the re-dirtying code above. 2802 * 2803 * Normally we can do this whether a buffer is B_DELWRI or not. If 2804 * the buffer is an NFS buffer, it is tracking piecemeal writes or 2805 * the commit state and we cannot afford to lose the buffer. If the 2806 * buffer has a background write in progress, we need to keep it 2807 * around to prevent it from being reconstituted and starting a second 2808 * background write. 2809 */ 2810 2811 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL; 2812 2813 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE || 2814 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) && 2815 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 || 2816 vn_isdisk(bp->b_vp) || (bp->b_flags & B_DELWRI) == 0)) { 2817 vfs_vmio_invalidate(bp); 2818 allocbuf(bp, 0); 2819 } 2820 2821 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 || 2822 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) { 2823 allocbuf(bp, 0); 2824 bp->b_flags &= ~B_NOREUSE; 2825 if (bp->b_vp != NULL) 2826 brelvp(bp); 2827 } 2828 2829 /* 2830 * If the buffer has junk contents signal it and eventually 2831 * clean up B_DELWRI and diassociate the vnode so that gbincore() 2832 * doesn't find it. 2833 */ 2834 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 2835 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 2836 bp->b_flags |= B_INVAL; 2837 if (bp->b_flags & B_INVAL) { 2838 if (bp->b_flags & B_DELWRI) 2839 bundirty(bp); 2840 if (bp->b_vp) 2841 brelvp(bp); 2842 } 2843 2844 buf_track(bp, __func__); 2845 2846 /* buffers with no memory */ 2847 if (bp->b_bufsize == 0) { 2848 buf_free(bp); 2849 return; 2850 } 2851 /* buffers with junk contents */ 2852 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 2853 (bp->b_ioflags & BIO_ERROR)) { 2854 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 2855 if (bp->b_vflags & BV_BKGRDINPROG) 2856 panic("losing buffer 2"); 2857 qindex = QUEUE_CLEAN; 2858 bp->b_flags |= B_AGE; 2859 /* remaining buffers */ 2860 } else if (bp->b_flags & B_DELWRI) 2861 qindex = QUEUE_DIRTY; 2862 else 2863 qindex = QUEUE_CLEAN; 2864 2865 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 2866 panic("brelse: not dirty"); 2867 2868 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT); 2869 bp->b_xflags &= ~(BX_CVTENXIO); 2870 /* binsfree unlocks bp. */ 2871 binsfree(bp, qindex); 2872 } 2873 2874 /* 2875 * Release a buffer back to the appropriate queue but do not try to free 2876 * it. The buffer is expected to be used again soon. 2877 * 2878 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 2879 * biodone() to requeue an async I/O on completion. It is also used when 2880 * known good buffers need to be requeued but we think we may need the data 2881 * again soon. 2882 * 2883 * XXX we should be able to leave the B_RELBUF hint set on completion. 2884 */ 2885 void 2886 bqrelse(struct buf *bp) 2887 { 2888 int qindex; 2889 2890 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2891 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2892 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2893 2894 qindex = QUEUE_NONE; 2895 if (BUF_LOCKRECURSED(bp)) { 2896 /* do not release to free list */ 2897 BUF_UNLOCK(bp); 2898 return; 2899 } 2900 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 2901 bp->b_xflags &= ~(BX_CVTENXIO); 2902 2903 if (LIST_EMPTY(&bp->b_dep)) { 2904 bp->b_flags &= ~B_IOSTARTED; 2905 } else { 2906 KASSERT((bp->b_flags & B_IOSTARTED) == 0, 2907 ("bqrelse: SU io not finished bp %p", bp)); 2908 } 2909 2910 if (bp->b_flags & B_MANAGED) { 2911 if (bp->b_flags & B_REMFREE) 2912 bremfreef(bp); 2913 goto out; 2914 } 2915 2916 /* buffers with stale but valid contents */ 2917 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | 2918 BV_BKGRDERR)) == BV_BKGRDERR) { 2919 BO_LOCK(bp->b_bufobj); 2920 bp->b_vflags &= ~BV_BKGRDERR; 2921 BO_UNLOCK(bp->b_bufobj); 2922 qindex = QUEUE_DIRTY; 2923 } else { 2924 if ((bp->b_flags & B_DELWRI) == 0 && 2925 (bp->b_xflags & BX_VNDIRTY)) 2926 panic("bqrelse: not dirty"); 2927 if ((bp->b_flags & B_NOREUSE) != 0) { 2928 brelse(bp); 2929 return; 2930 } 2931 qindex = QUEUE_CLEAN; 2932 } 2933 buf_track(bp, __func__); 2934 /* binsfree unlocks bp. */ 2935 binsfree(bp, qindex); 2936 return; 2937 2938 out: 2939 buf_track(bp, __func__); 2940 /* unlock */ 2941 BUF_UNLOCK(bp); 2942 } 2943 2944 /* 2945 * Complete I/O to a VMIO backed page. Validate the pages as appropriate, 2946 * restore bogus pages. 2947 */ 2948 static void 2949 vfs_vmio_iodone(struct buf *bp) 2950 { 2951 vm_ooffset_t foff; 2952 vm_page_t m; 2953 vm_object_t obj; 2954 struct vnode *vp __unused; 2955 int i, iosize, resid; 2956 bool bogus; 2957 2958 obj = bp->b_bufobj->bo_object; 2959 KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages, 2960 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)", 2961 blockcount_read(&obj->paging_in_progress), bp->b_npages)); 2962 2963 vp = bp->b_vp; 2964 VNPASS(vp->v_holdcnt > 0, vp); 2965 VNPASS(vp->v_object != NULL, vp); 2966 2967 foff = bp->b_offset; 2968 KASSERT(bp->b_offset != NOOFFSET, 2969 ("vfs_vmio_iodone: bp %p has no buffer offset", bp)); 2970 2971 bogus = false; 2972 iosize = bp->b_bcount - bp->b_resid; 2973 for (i = 0; i < bp->b_npages; i++) { 2974 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2975 if (resid > iosize) 2976 resid = iosize; 2977 2978 /* 2979 * cleanup bogus pages, restoring the originals 2980 */ 2981 m = bp->b_pages[i]; 2982 if (m == bogus_page) { 2983 bogus = true; 2984 m = vm_page_relookup(obj, OFF_TO_IDX(foff)); 2985 if (m == NULL) 2986 panic("biodone: page disappeared!"); 2987 bp->b_pages[i] = m; 2988 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) { 2989 /* 2990 * In the write case, the valid and clean bits are 2991 * already changed correctly ( see bdwrite() ), so we 2992 * only need to do this here in the read case. 2993 */ 2994 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK, 2995 resid)) == 0, ("vfs_vmio_iodone: page %p " 2996 "has unexpected dirty bits", m)); 2997 vfs_page_set_valid(bp, foff, m); 2998 } 2999 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3000 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch", 3001 (intmax_t)foff, (uintmax_t)m->pindex)); 3002 3003 vm_page_sunbusy(m); 3004 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3005 iosize -= resid; 3006 } 3007 vm_object_pip_wakeupn(obj, bp->b_npages); 3008 if (bogus && buf_mapped(bp)) { 3009 BUF_CHECK_MAPPED(bp); 3010 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3011 bp->b_pages, bp->b_npages); 3012 } 3013 } 3014 3015 /* 3016 * Perform page invalidation when a buffer is released. The fully invalid 3017 * pages will be reclaimed later in vfs_vmio_truncate(). 3018 */ 3019 static void 3020 vfs_vmio_invalidate(struct buf *bp) 3021 { 3022 vm_object_t obj; 3023 vm_page_t m; 3024 int flags, i, resid, poffset, presid; 3025 3026 if (buf_mapped(bp)) { 3027 BUF_CHECK_MAPPED(bp); 3028 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 3029 } else 3030 BUF_CHECK_UNMAPPED(bp); 3031 /* 3032 * Get the base offset and length of the buffer. Note that 3033 * in the VMIO case if the buffer block size is not 3034 * page-aligned then b_data pointer may not be page-aligned. 3035 * But our b_pages[] array *IS* page aligned. 3036 * 3037 * block sizes less then DEV_BSIZE (usually 512) are not 3038 * supported due to the page granularity bits (m->valid, 3039 * m->dirty, etc...). 3040 * 3041 * See man buf(9) for more information 3042 */ 3043 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; 3044 obj = bp->b_bufobj->bo_object; 3045 resid = bp->b_bufsize; 3046 poffset = bp->b_offset & PAGE_MASK; 3047 VM_OBJECT_WLOCK(obj); 3048 for (i = 0; i < bp->b_npages; i++) { 3049 m = bp->b_pages[i]; 3050 if (m == bogus_page) 3051 panic("vfs_vmio_invalidate: Unexpected bogus page."); 3052 bp->b_pages[i] = NULL; 3053 3054 presid = resid > (PAGE_SIZE - poffset) ? 3055 (PAGE_SIZE - poffset) : resid; 3056 KASSERT(presid >= 0, ("brelse: extra page")); 3057 vm_page_busy_acquire(m, VM_ALLOC_SBUSY); 3058 if (pmap_page_wired_mappings(m) == 0) 3059 vm_page_set_invalid(m, poffset, presid); 3060 vm_page_sunbusy(m); 3061 vm_page_release_locked(m, flags); 3062 resid -= presid; 3063 poffset = 0; 3064 } 3065 VM_OBJECT_WUNLOCK(obj); 3066 bp->b_npages = 0; 3067 } 3068 3069 /* 3070 * Page-granular truncation of an existing VMIO buffer. 3071 */ 3072 static void 3073 vfs_vmio_truncate(struct buf *bp, int desiredpages) 3074 { 3075 vm_object_t obj; 3076 vm_page_t m; 3077 int flags, i; 3078 3079 if (bp->b_npages == desiredpages) 3080 return; 3081 3082 if (buf_mapped(bp)) { 3083 BUF_CHECK_MAPPED(bp); 3084 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) + 3085 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages); 3086 } else 3087 BUF_CHECK_UNMAPPED(bp); 3088 3089 /* 3090 * The object lock is needed only if we will attempt to free pages. 3091 */ 3092 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; 3093 if ((bp->b_flags & B_DIRECT) != 0) { 3094 flags |= VPR_TRYFREE; 3095 obj = bp->b_bufobj->bo_object; 3096 VM_OBJECT_WLOCK(obj); 3097 } else { 3098 obj = NULL; 3099 } 3100 for (i = desiredpages; i < bp->b_npages; i++) { 3101 m = bp->b_pages[i]; 3102 KASSERT(m != bogus_page, ("allocbuf: bogus page found")); 3103 bp->b_pages[i] = NULL; 3104 if (obj != NULL) 3105 vm_page_release_locked(m, flags); 3106 else 3107 vm_page_release(m, flags); 3108 } 3109 if (obj != NULL) 3110 VM_OBJECT_WUNLOCK(obj); 3111 bp->b_npages = desiredpages; 3112 } 3113 3114 /* 3115 * Byte granular extension of VMIO buffers. 3116 */ 3117 static void 3118 vfs_vmio_extend(struct buf *bp, int desiredpages, int size) 3119 { 3120 /* 3121 * We are growing the buffer, possibly in a 3122 * byte-granular fashion. 3123 */ 3124 vm_object_t obj; 3125 vm_offset_t toff; 3126 vm_offset_t tinc; 3127 vm_page_t m; 3128 3129 /* 3130 * Step 1, bring in the VM pages from the object, allocating 3131 * them if necessary. We must clear B_CACHE if these pages 3132 * are not valid for the range covered by the buffer. 3133 */ 3134 obj = bp->b_bufobj->bo_object; 3135 if (bp->b_npages < desiredpages) { 3136 KASSERT(desiredpages <= atop(maxbcachebuf), 3137 ("vfs_vmio_extend past maxbcachebuf %p %d %u", 3138 bp, desiredpages, maxbcachebuf)); 3139 3140 /* 3141 * We must allocate system pages since blocking 3142 * here could interfere with paging I/O, no 3143 * matter which process we are. 3144 * 3145 * Only exclusive busy can be tested here. 3146 * Blocking on shared busy might lead to 3147 * deadlocks once allocbuf() is called after 3148 * pages are vfs_busy_pages(). 3149 */ 3150 (void)vm_page_grab_pages_unlocked(obj, 3151 OFF_TO_IDX(bp->b_offset) + bp->b_npages, 3152 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY | 3153 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED, 3154 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages); 3155 bp->b_npages = desiredpages; 3156 } 3157 3158 /* 3159 * Step 2. We've loaded the pages into the buffer, 3160 * we have to figure out if we can still have B_CACHE 3161 * set. Note that B_CACHE is set according to the 3162 * byte-granular range ( bcount and size ), not the 3163 * aligned range ( newbsize ). 3164 * 3165 * The VM test is against m->valid, which is DEV_BSIZE 3166 * aligned. Needless to say, the validity of the data 3167 * needs to also be DEV_BSIZE aligned. Note that this 3168 * fails with NFS if the server or some other client 3169 * extends the file's EOF. If our buffer is resized, 3170 * B_CACHE may remain set! XXX 3171 */ 3172 toff = bp->b_bcount; 3173 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3174 while ((bp->b_flags & B_CACHE) && toff < size) { 3175 vm_pindex_t pi; 3176 3177 if (tinc > (size - toff)) 3178 tinc = size - toff; 3179 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; 3180 m = bp->b_pages[pi]; 3181 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m); 3182 toff += tinc; 3183 tinc = PAGE_SIZE; 3184 } 3185 3186 /* 3187 * Step 3, fixup the KVA pmap. 3188 */ 3189 if (buf_mapped(bp)) 3190 bpmap_qenter(bp); 3191 else 3192 BUF_CHECK_UNMAPPED(bp); 3193 } 3194 3195 /* 3196 * Check to see if a block at a particular lbn is available for a clustered 3197 * write. 3198 */ 3199 static int 3200 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 3201 { 3202 struct buf *bpa; 3203 int match; 3204 3205 match = 0; 3206 3207 /* If the buf isn't in core skip it */ 3208 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 3209 return (0); 3210 3211 /* If the buf is busy we don't want to wait for it */ 3212 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 3213 return (0); 3214 3215 /* Only cluster with valid clusterable delayed write buffers */ 3216 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 3217 (B_DELWRI | B_CLUSTEROK)) 3218 goto done; 3219 3220 if (bpa->b_bufsize != size) 3221 goto done; 3222 3223 /* 3224 * Check to see if it is in the expected place on disk and that the 3225 * block has been mapped. 3226 */ 3227 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 3228 match = 1; 3229 done: 3230 BUF_UNLOCK(bpa); 3231 return (match); 3232 } 3233 3234 /* 3235 * vfs_bio_awrite: 3236 * 3237 * Implement clustered async writes for clearing out B_DELWRI buffers. 3238 * This is much better then the old way of writing only one buffer at 3239 * a time. Note that we may not be presented with the buffers in the 3240 * correct order, so we search for the cluster in both directions. 3241 */ 3242 int 3243 vfs_bio_awrite(struct buf *bp) 3244 { 3245 struct bufobj *bo; 3246 int i; 3247 int j; 3248 daddr_t lblkno = bp->b_lblkno; 3249 struct vnode *vp = bp->b_vp; 3250 int ncl; 3251 int nwritten; 3252 int size; 3253 int maxcl; 3254 int gbflags; 3255 3256 bo = &vp->v_bufobj; 3257 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0; 3258 /* 3259 * right now we support clustered writing only to regular files. If 3260 * we find a clusterable block we could be in the middle of a cluster 3261 * rather then at the beginning. 3262 */ 3263 if ((vp->v_type == VREG) && 3264 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 3265 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 3266 size = vp->v_mount->mnt_stat.f_iosize; 3267 maxcl = maxphys / size; 3268 3269 BO_RLOCK(bo); 3270 for (i = 1; i < maxcl; i++) 3271 if (vfs_bio_clcheck(vp, size, lblkno + i, 3272 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 3273 break; 3274 3275 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 3276 if (vfs_bio_clcheck(vp, size, lblkno - j, 3277 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 3278 break; 3279 BO_RUNLOCK(bo); 3280 --j; 3281 ncl = i + j; 3282 /* 3283 * this is a possible cluster write 3284 */ 3285 if (ncl != 1) { 3286 BUF_UNLOCK(bp); 3287 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 3288 gbflags); 3289 return (nwritten); 3290 } 3291 } 3292 bremfree(bp); 3293 bp->b_flags |= B_ASYNC; 3294 /* 3295 * default (old) behavior, writing out only one block 3296 * 3297 * XXX returns b_bufsize instead of b_bcount for nwritten? 3298 */ 3299 nwritten = bp->b_bufsize; 3300 (void) bwrite(bp); 3301 3302 return (nwritten); 3303 } 3304 3305 /* 3306 * getnewbuf_kva: 3307 * 3308 * Allocate KVA for an empty buf header according to gbflags. 3309 */ 3310 static int 3311 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize) 3312 { 3313 3314 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) { 3315 /* 3316 * In order to keep fragmentation sane we only allocate kva 3317 * in BKVASIZE chunks. XXX with vmem we can do page size. 3318 */ 3319 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 3320 3321 if (maxsize != bp->b_kvasize && 3322 bufkva_alloc(bp, maxsize, gbflags)) 3323 return (ENOSPC); 3324 } 3325 return (0); 3326 } 3327 3328 /* 3329 * getnewbuf: 3330 * 3331 * Find and initialize a new buffer header, freeing up existing buffers 3332 * in the bufqueues as necessary. The new buffer is returned locked. 3333 * 3334 * We block if: 3335 * We have insufficient buffer headers 3336 * We have insufficient buffer space 3337 * buffer_arena is too fragmented ( space reservation fails ) 3338 * If we have to flush dirty buffers ( but we try to avoid this ) 3339 * 3340 * The caller is responsible for releasing the reserved bufspace after 3341 * allocbuf() is called. 3342 */ 3343 static struct buf * 3344 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags) 3345 { 3346 struct bufdomain *bd; 3347 struct buf *bp; 3348 bool metadata, reserved; 3349 3350 bp = NULL; 3351 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3352 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3353 if (!unmapped_buf_allowed) 3354 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3355 3356 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 3357 vp->v_type == VCHR) 3358 metadata = true; 3359 else 3360 metadata = false; 3361 if (vp == NULL) 3362 bd = &bdomain[0]; 3363 else 3364 bd = &bdomain[vp->v_bufobj.bo_domain]; 3365 3366 counter_u64_add(getnewbufcalls, 1); 3367 reserved = false; 3368 do { 3369 if (reserved == false && 3370 bufspace_reserve(bd, maxsize, metadata) != 0) { 3371 counter_u64_add(getnewbufrestarts, 1); 3372 continue; 3373 } 3374 reserved = true; 3375 if ((bp = buf_alloc(bd)) == NULL) { 3376 counter_u64_add(getnewbufrestarts, 1); 3377 continue; 3378 } 3379 if (getnewbuf_kva(bp, gbflags, maxsize) == 0) 3380 return (bp); 3381 break; 3382 } while (buf_recycle(bd, false) == 0); 3383 3384 if (reserved) 3385 bufspace_release(bd, maxsize); 3386 if (bp != NULL) { 3387 bp->b_flags |= B_INVAL; 3388 brelse(bp); 3389 } 3390 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo); 3391 3392 return (NULL); 3393 } 3394 3395 /* 3396 * buf_daemon: 3397 * 3398 * buffer flushing daemon. Buffers are normally flushed by the 3399 * update daemon but if it cannot keep up this process starts to 3400 * take the load in an attempt to prevent getnewbuf() from blocking. 3401 */ 3402 static struct kproc_desc buf_kp = { 3403 "bufdaemon", 3404 buf_daemon, 3405 &bufdaemonproc 3406 }; 3407 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 3408 3409 static int 3410 buf_flush(struct vnode *vp, struct bufdomain *bd, int target) 3411 { 3412 int flushed; 3413 3414 flushed = flushbufqueues(vp, bd, target, 0); 3415 if (flushed == 0) { 3416 /* 3417 * Could not find any buffers without rollback 3418 * dependencies, so just write the first one 3419 * in the hopes of eventually making progress. 3420 */ 3421 if (vp != NULL && target > 2) 3422 target /= 2; 3423 flushbufqueues(vp, bd, target, 1); 3424 } 3425 return (flushed); 3426 } 3427 3428 static void 3429 buf_daemon_shutdown(void *arg __unused, int howto __unused) 3430 { 3431 int error; 3432 3433 if (KERNEL_PANICKED()) 3434 return; 3435 3436 mtx_lock(&bdlock); 3437 bd_shutdown = true; 3438 wakeup(&bd_request); 3439 error = msleep(&bd_shutdown, &bdlock, 0, "buf_daemon_shutdown", 3440 60 * hz); 3441 mtx_unlock(&bdlock); 3442 if (error != 0) 3443 printf("bufdaemon wait error: %d\n", error); 3444 } 3445 3446 static void 3447 buf_daemon(void) 3448 { 3449 struct bufdomain *bd; 3450 int speedupreq; 3451 int lodirty; 3452 int i; 3453 3454 /* 3455 * This process needs to be suspended prior to shutdown sync. 3456 */ 3457 EVENTHANDLER_REGISTER(shutdown_pre_sync, buf_daemon_shutdown, NULL, 3458 SHUTDOWN_PRI_LAST + 100); 3459 3460 /* 3461 * Start the buf clean daemons as children threads. 3462 */ 3463 for (i = 0 ; i < buf_domains; i++) { 3464 int error; 3465 3466 error = kthread_add((void (*)(void *))bufspace_daemon, 3467 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i); 3468 if (error) 3469 panic("error %d spawning bufspace daemon", error); 3470 } 3471 3472 /* 3473 * This process is allowed to take the buffer cache to the limit 3474 */ 3475 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 3476 mtx_lock(&bdlock); 3477 while (!bd_shutdown) { 3478 bd_request = 0; 3479 mtx_unlock(&bdlock); 3480 3481 /* 3482 * Save speedupreq for this pass and reset to capture new 3483 * requests. 3484 */ 3485 speedupreq = bd_speedupreq; 3486 bd_speedupreq = 0; 3487 3488 /* 3489 * Flush each domain sequentially according to its level and 3490 * the speedup request. 3491 */ 3492 for (i = 0; i < buf_domains; i++) { 3493 bd = &bdomain[i]; 3494 if (speedupreq) 3495 lodirty = bd->bd_numdirtybuffers / 2; 3496 else 3497 lodirty = bd->bd_lodirtybuffers; 3498 while (bd->bd_numdirtybuffers > lodirty) { 3499 if (buf_flush(NULL, bd, 3500 bd->bd_numdirtybuffers - lodirty) == 0) 3501 break; 3502 kern_yield(PRI_USER); 3503 } 3504 } 3505 3506 /* 3507 * Only clear bd_request if we have reached our low water 3508 * mark. The buf_daemon normally waits 1 second and 3509 * then incrementally flushes any dirty buffers that have 3510 * built up, within reason. 3511 * 3512 * If we were unable to hit our low water mark and couldn't 3513 * find any flushable buffers, we sleep for a short period 3514 * to avoid endless loops on unlockable buffers. 3515 */ 3516 mtx_lock(&bdlock); 3517 if (bd_shutdown) 3518 break; 3519 if (BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) { 3520 /* 3521 * We reached our low water mark, reset the 3522 * request and sleep until we are needed again. 3523 * The sleep is just so the suspend code works. 3524 */ 3525 bd_request = 0; 3526 /* 3527 * Do an extra wakeup in case dirty threshold 3528 * changed via sysctl and the explicit transition 3529 * out of shortfall was missed. 3530 */ 3531 bdirtywakeup(); 3532 if (runningbufspace <= lorunningspace) 3533 runningwakeup(); 3534 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 3535 } else { 3536 /* 3537 * We couldn't find any flushable dirty buffers but 3538 * still have too many dirty buffers, we 3539 * have to sleep and try again. (rare) 3540 */ 3541 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 3542 } 3543 } 3544 wakeup(&bd_shutdown); 3545 mtx_unlock(&bdlock); 3546 kthread_exit(); 3547 } 3548 3549 /* 3550 * flushbufqueues: 3551 * 3552 * Try to flush a buffer in the dirty queue. We must be careful to 3553 * free up B_INVAL buffers instead of write them, which NFS is 3554 * particularly sensitive to. 3555 */ 3556 static int flushwithdeps = 0; 3557 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS, 3558 &flushwithdeps, 0, 3559 "Number of buffers flushed with dependencies that require rollbacks"); 3560 3561 static int 3562 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target, 3563 int flushdeps) 3564 { 3565 struct bufqueue *bq; 3566 struct buf *sentinel; 3567 struct vnode *vp; 3568 struct mount *mp; 3569 struct buf *bp; 3570 int hasdeps; 3571 int flushed; 3572 int error; 3573 bool unlock; 3574 3575 flushed = 0; 3576 bq = &bd->bd_dirtyq; 3577 bp = NULL; 3578 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 3579 sentinel->b_qindex = QUEUE_SENTINEL; 3580 BQ_LOCK(bq); 3581 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist); 3582 BQ_UNLOCK(bq); 3583 while (flushed != target) { 3584 maybe_yield(); 3585 BQ_LOCK(bq); 3586 bp = TAILQ_NEXT(sentinel, b_freelist); 3587 if (bp != NULL) { 3588 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); 3589 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel, 3590 b_freelist); 3591 } else { 3592 BQ_UNLOCK(bq); 3593 break; 3594 } 3595 /* 3596 * Skip sentinels inserted by other invocations of the 3597 * flushbufqueues(), taking care to not reorder them. 3598 * 3599 * Only flush the buffers that belong to the 3600 * vnode locked by the curthread. 3601 */ 3602 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && 3603 bp->b_vp != lvp)) { 3604 BQ_UNLOCK(bq); 3605 continue; 3606 } 3607 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 3608 BQ_UNLOCK(bq); 3609 if (error != 0) 3610 continue; 3611 3612 /* 3613 * BKGRDINPROG can only be set with the buf and bufobj 3614 * locks both held. We tolerate a race to clear it here. 3615 */ 3616 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 3617 (bp->b_flags & B_DELWRI) == 0) { 3618 BUF_UNLOCK(bp); 3619 continue; 3620 } 3621 if (bp->b_flags & B_INVAL) { 3622 bremfreef(bp); 3623 brelse(bp); 3624 flushed++; 3625 continue; 3626 } 3627 3628 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 3629 if (flushdeps == 0) { 3630 BUF_UNLOCK(bp); 3631 continue; 3632 } 3633 hasdeps = 1; 3634 } else 3635 hasdeps = 0; 3636 /* 3637 * We must hold the lock on a vnode before writing 3638 * one of its buffers. Otherwise we may confuse, or 3639 * in the case of a snapshot vnode, deadlock the 3640 * system. 3641 * 3642 * The lock order here is the reverse of the normal 3643 * of vnode followed by buf lock. This is ok because 3644 * the NOWAIT will prevent deadlock. 3645 */ 3646 vp = bp->b_vp; 3647 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 3648 BUF_UNLOCK(bp); 3649 continue; 3650 } 3651 if (lvp == NULL) { 3652 unlock = true; 3653 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 3654 } else { 3655 ASSERT_VOP_LOCKED(vp, "getbuf"); 3656 unlock = false; 3657 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 3658 vn_lock(vp, LK_TRYUPGRADE); 3659 } 3660 if (error == 0) { 3661 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 3662 bp, bp->b_vp, bp->b_flags); 3663 if (curproc == bufdaemonproc) { 3664 vfs_bio_awrite(bp); 3665 } else { 3666 bremfree(bp); 3667 bwrite(bp); 3668 counter_u64_add(notbufdflushes, 1); 3669 } 3670 vn_finished_write(mp); 3671 if (unlock) 3672 VOP_UNLOCK(vp); 3673 flushwithdeps += hasdeps; 3674 flushed++; 3675 3676 /* 3677 * Sleeping on runningbufspace while holding 3678 * vnode lock leads to deadlock. 3679 */ 3680 if (curproc == bufdaemonproc && 3681 runningbufspace > hirunningspace) 3682 waitrunningbufspace(); 3683 continue; 3684 } 3685 vn_finished_write(mp); 3686 BUF_UNLOCK(bp); 3687 } 3688 BQ_LOCK(bq); 3689 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); 3690 BQ_UNLOCK(bq); 3691 free(sentinel, M_TEMP); 3692 return (flushed); 3693 } 3694 3695 /* 3696 * Check to see if a block is currently memory resident. 3697 */ 3698 struct buf * 3699 incore(struct bufobj *bo, daddr_t blkno) 3700 { 3701 return (gbincore_unlocked(bo, blkno)); 3702 } 3703 3704 /* 3705 * Returns true if no I/O is needed to access the 3706 * associated VM object. This is like incore except 3707 * it also hunts around in the VM system for the data. 3708 */ 3709 bool 3710 inmem(struct vnode * vp, daddr_t blkno) 3711 { 3712 vm_object_t obj; 3713 vm_offset_t toff, tinc, size; 3714 vm_page_t m, n; 3715 vm_ooffset_t off; 3716 int valid; 3717 3718 ASSERT_VOP_LOCKED(vp, "inmem"); 3719 3720 if (incore(&vp->v_bufobj, blkno)) 3721 return (true); 3722 if (vp->v_mount == NULL) 3723 return (false); 3724 obj = vp->v_object; 3725 if (obj == NULL) 3726 return (false); 3727 3728 size = PAGE_SIZE; 3729 if (size > vp->v_mount->mnt_stat.f_iosize) 3730 size = vp->v_mount->mnt_stat.f_iosize; 3731 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 3732 3733 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 3734 m = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff)); 3735 recheck: 3736 if (m == NULL) 3737 return (false); 3738 3739 tinc = size; 3740 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 3741 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 3742 /* 3743 * Consider page validity only if page mapping didn't change 3744 * during the check. 3745 */ 3746 valid = vm_page_is_valid(m, 3747 (vm_offset_t)((toff + off) & PAGE_MASK), tinc); 3748 n = vm_page_lookup_unlocked(obj, OFF_TO_IDX(off + toff)); 3749 if (m != n) { 3750 m = n; 3751 goto recheck; 3752 } 3753 if (!valid) 3754 return (false); 3755 } 3756 return (true); 3757 } 3758 3759 /* 3760 * Set the dirty range for a buffer based on the status of the dirty 3761 * bits in the pages comprising the buffer. The range is limited 3762 * to the size of the buffer. 3763 * 3764 * Tell the VM system that the pages associated with this buffer 3765 * are clean. This is used for delayed writes where the data is 3766 * going to go to disk eventually without additional VM intevention. 3767 * 3768 * Note that while we only really need to clean through to b_bcount, we 3769 * just go ahead and clean through to b_bufsize. 3770 */ 3771 static void 3772 vfs_clean_pages_dirty_buf(struct buf *bp) 3773 { 3774 vm_ooffset_t foff, noff, eoff; 3775 vm_page_t m; 3776 int i; 3777 3778 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 3779 return; 3780 3781 foff = bp->b_offset; 3782 KASSERT(bp->b_offset != NOOFFSET, 3783 ("vfs_clean_pages_dirty_buf: no buffer offset")); 3784 3785 vfs_busy_pages_acquire(bp); 3786 vfs_setdirty_range(bp); 3787 for (i = 0; i < bp->b_npages; i++) { 3788 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3789 eoff = noff; 3790 if (eoff > bp->b_offset + bp->b_bufsize) 3791 eoff = bp->b_offset + bp->b_bufsize; 3792 m = bp->b_pages[i]; 3793 vfs_page_set_validclean(bp, foff, m); 3794 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3795 foff = noff; 3796 } 3797 vfs_busy_pages_release(bp); 3798 } 3799 3800 static void 3801 vfs_setdirty_range(struct buf *bp) 3802 { 3803 vm_offset_t boffset; 3804 vm_offset_t eoffset; 3805 int i; 3806 3807 /* 3808 * test the pages to see if they have been modified directly 3809 * by users through the VM system. 3810 */ 3811 for (i = 0; i < bp->b_npages; i++) 3812 vm_page_test_dirty(bp->b_pages[i]); 3813 3814 /* 3815 * Calculate the encompassing dirty range, boffset and eoffset, 3816 * (eoffset - boffset) bytes. 3817 */ 3818 3819 for (i = 0; i < bp->b_npages; i++) { 3820 if (bp->b_pages[i]->dirty) 3821 break; 3822 } 3823 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3824 3825 for (i = bp->b_npages - 1; i >= 0; --i) { 3826 if (bp->b_pages[i]->dirty) { 3827 break; 3828 } 3829 } 3830 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3831 3832 /* 3833 * Fit it to the buffer. 3834 */ 3835 3836 if (eoffset > bp->b_bcount) 3837 eoffset = bp->b_bcount; 3838 3839 /* 3840 * If we have a good dirty range, merge with the existing 3841 * dirty range. 3842 */ 3843 3844 if (boffset < eoffset) { 3845 if (bp->b_dirtyoff > boffset) 3846 bp->b_dirtyoff = boffset; 3847 if (bp->b_dirtyend < eoffset) 3848 bp->b_dirtyend = eoffset; 3849 } 3850 } 3851 3852 /* 3853 * Allocate the KVA mapping for an existing buffer. 3854 * If an unmapped buffer is provided but a mapped buffer is requested, take 3855 * also care to properly setup mappings between pages and KVA. 3856 */ 3857 static void 3858 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 3859 { 3860 int bsize, maxsize, need_mapping, need_kva; 3861 off_t offset; 3862 3863 need_mapping = bp->b_data == unmapped_buf && 3864 (gbflags & GB_UNMAPPED) == 0; 3865 need_kva = bp->b_kvabase == unmapped_buf && 3866 bp->b_data == unmapped_buf && 3867 (gbflags & GB_KVAALLOC) != 0; 3868 if (!need_mapping && !need_kva) 3869 return; 3870 3871 BUF_CHECK_UNMAPPED(bp); 3872 3873 if (need_mapping && bp->b_kvabase != unmapped_buf) { 3874 /* 3875 * Buffer is not mapped, but the KVA was already 3876 * reserved at the time of the instantiation. Use the 3877 * allocated space. 3878 */ 3879 goto has_addr; 3880 } 3881 3882 /* 3883 * Calculate the amount of the address space we would reserve 3884 * if the buffer was mapped. 3885 */ 3886 bsize = vn_isdisk(bp->b_vp) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 3887 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3888 offset = blkno * bsize; 3889 maxsize = size + (offset & PAGE_MASK); 3890 maxsize = imax(maxsize, bsize); 3891 3892 while (bufkva_alloc(bp, maxsize, gbflags) != 0) { 3893 if ((gbflags & GB_NOWAIT_BD) != 0) { 3894 /* 3895 * XXXKIB: defragmentation cannot 3896 * succeed, not sure what else to do. 3897 */ 3898 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp); 3899 } 3900 counter_u64_add(mappingrestarts, 1); 3901 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0); 3902 } 3903 has_addr: 3904 if (need_mapping) { 3905 /* b_offset is handled by bpmap_qenter. */ 3906 bp->b_data = bp->b_kvabase; 3907 BUF_CHECK_MAPPED(bp); 3908 bpmap_qenter(bp); 3909 } 3910 } 3911 3912 struct buf * 3913 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3914 int flags) 3915 { 3916 struct buf *bp; 3917 int error; 3918 3919 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp); 3920 if (error != 0) 3921 return (NULL); 3922 return (bp); 3923 } 3924 3925 /* 3926 * getblkx: 3927 * 3928 * Get a block given a specified block and offset into a file/device. 3929 * The buffers B_DONE bit will be cleared on return, making it almost 3930 * ready for an I/O initiation. B_INVAL may or may not be set on 3931 * return. The caller should clear B_INVAL prior to initiating a 3932 * READ. 3933 * 3934 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3935 * an existing buffer. 3936 * 3937 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3938 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3939 * and then cleared based on the backing VM. If the previous buffer is 3940 * non-0-sized but invalid, B_CACHE will be cleared. 3941 * 3942 * If getblk() must create a new buffer, the new buffer is returned with 3943 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3944 * case it is returned with B_INVAL clear and B_CACHE set based on the 3945 * backing VM. 3946 * 3947 * getblk() also forces a bwrite() for any B_DELWRI buffer whose 3948 * B_CACHE bit is clear. 3949 * 3950 * What this means, basically, is that the caller should use B_CACHE to 3951 * determine whether the buffer is fully valid or not and should clear 3952 * B_INVAL prior to issuing a read. If the caller intends to validate 3953 * the buffer by loading its data area with something, the caller needs 3954 * to clear B_INVAL. If the caller does this without issuing an I/O, 3955 * the caller should set B_CACHE ( as an optimization ), else the caller 3956 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3957 * a write attempt or if it was a successful read. If the caller 3958 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3959 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3960 * 3961 * The blkno parameter is the logical block being requested. Normally 3962 * the mapping of logical block number to disk block address is done 3963 * by calling VOP_BMAP(). However, if the mapping is already known, the 3964 * disk block address can be passed using the dblkno parameter. If the 3965 * disk block address is not known, then the same value should be passed 3966 * for blkno and dblkno. 3967 */ 3968 int 3969 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag, 3970 int slptimeo, int flags, struct buf **bpp) 3971 { 3972 struct buf *bp; 3973 struct bufobj *bo; 3974 daddr_t d_blkno; 3975 int bsize, error, maxsize, vmio; 3976 off_t offset; 3977 3978 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3979 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3980 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3981 if (vp->v_type != VCHR) 3982 ASSERT_VOP_LOCKED(vp, "getblk"); 3983 if (size > maxbcachebuf) { 3984 printf("getblkx: size(%d) > maxbcachebuf(%d)\n", size, 3985 maxbcachebuf); 3986 return (EIO); 3987 } 3988 if (!unmapped_buf_allowed) 3989 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3990 3991 bo = &vp->v_bufobj; 3992 d_blkno = dblkno; 3993 3994 /* Attempt lockless lookup first. */ 3995 bp = gbincore_unlocked(bo, blkno); 3996 if (bp == NULL) { 3997 /* 3998 * With GB_NOCREAT we must be sure about not finding the buffer 3999 * as it may have been reassigned during unlocked lookup. 4000 * If BO_NONSTERILE is still unset, no reassign has occurred. 4001 */ 4002 if ((flags & GB_NOCREAT) != 0) { 4003 /* Ensure bo_flag is loaded after gbincore_unlocked. */ 4004 atomic_thread_fence_acq(); 4005 if ((bo->bo_flag & BO_NONSTERILE) == 0) 4006 return (EEXIST); 4007 goto loop; 4008 } 4009 goto newbuf_unlocked; 4010 } 4011 4012 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL, "getblku", 0, 4013 0); 4014 if (error != 0) { 4015 KASSERT(error == EBUSY, 4016 ("getblk: unexpected error %d from buf try-lock", error)); 4017 /* 4018 * We failed a buf try-lock. 4019 * 4020 * With GB_LOCK_NOWAIT, just return, rather than taking the 4021 * bufobj interlock and trying again, since we would probably 4022 * fail again anyway. This is okay even if the buf's identity 4023 * changed and we contended on the wrong lock, as changing 4024 * identity itself requires the buf lock, and we could have 4025 * contended on the right lock. 4026 */ 4027 if ((flags & GB_LOCK_NOWAIT) != 0) 4028 return (error); 4029 goto loop; 4030 } 4031 4032 /* Verify buf identify has not changed since lookup. */ 4033 if (bp->b_bufobj == bo && bp->b_lblkno == blkno) 4034 goto foundbuf_fastpath; 4035 4036 /* It changed, fallback to locked lookup. */ 4037 BUF_UNLOCK_RAW(bp); 4038 4039 /* As above, with GB_LOCK_NOWAIT, just return. */ 4040 if ((flags & GB_LOCK_NOWAIT) != 0) 4041 return (EBUSY); 4042 4043 loop: 4044 BO_RLOCK(bo); 4045 bp = gbincore(bo, blkno); 4046 if (bp != NULL) { 4047 int lockflags; 4048 4049 /* 4050 * Buffer is in-core. If the buffer is not busy nor managed, 4051 * it must be on a queue. 4052 */ 4053 lockflags = LK_EXCLUSIVE | LK_INTERLOCK | 4054 ((flags & GB_LOCK_NOWAIT) != 0 ? LK_NOWAIT : LK_SLEEPFAIL); 4055 #ifdef WITNESS 4056 lockflags |= (flags & GB_NOWITNESS) != 0 ? LK_NOWITNESS : 0; 4057 #endif 4058 4059 error = BUF_TIMELOCK(bp, lockflags, 4060 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 4061 4062 /* 4063 * If we slept and got the lock we have to restart in case 4064 * the buffer changed identities. 4065 */ 4066 if (error == ENOLCK) 4067 goto loop; 4068 /* We timed out or were interrupted. */ 4069 else if (error != 0) 4070 return (error); 4071 4072 foundbuf_fastpath: 4073 /* If recursed, assume caller knows the rules. */ 4074 if (BUF_LOCKRECURSED(bp)) 4075 goto end; 4076 4077 /* 4078 * The buffer is locked. B_CACHE is cleared if the buffer is 4079 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 4080 * and for a VMIO buffer B_CACHE is adjusted according to the 4081 * backing VM cache. 4082 */ 4083 if (bp->b_flags & B_INVAL) 4084 bp->b_flags &= ~B_CACHE; 4085 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 4086 bp->b_flags |= B_CACHE; 4087 if (bp->b_flags & B_MANAGED) 4088 MPASS(bp->b_qindex == QUEUE_NONE); 4089 else 4090 bremfree(bp); 4091 4092 /* 4093 * check for size inconsistencies for non-VMIO case. 4094 */ 4095 if (bp->b_bcount != size) { 4096 if ((bp->b_flags & B_VMIO) == 0 || 4097 (size > bp->b_kvasize)) { 4098 if (bp->b_flags & B_DELWRI) { 4099 bp->b_flags |= B_NOCACHE; 4100 bwrite(bp); 4101 } else { 4102 if (LIST_EMPTY(&bp->b_dep)) { 4103 bp->b_flags |= B_RELBUF; 4104 brelse(bp); 4105 } else { 4106 bp->b_flags |= B_NOCACHE; 4107 bwrite(bp); 4108 } 4109 } 4110 goto loop; 4111 } 4112 } 4113 4114 /* 4115 * Handle the case of unmapped buffer which should 4116 * become mapped, or the buffer for which KVA 4117 * reservation is requested. 4118 */ 4119 bp_unmapped_get_kva(bp, blkno, size, flags); 4120 4121 /* 4122 * If the size is inconsistent in the VMIO case, we can resize 4123 * the buffer. This might lead to B_CACHE getting set or 4124 * cleared. If the size has not changed, B_CACHE remains 4125 * unchanged from its previous state. 4126 */ 4127 allocbuf(bp, size); 4128 4129 KASSERT(bp->b_offset != NOOFFSET, 4130 ("getblk: no buffer offset")); 4131 4132 /* 4133 * A buffer with B_DELWRI set and B_CACHE clear must 4134 * be committed before we can return the buffer in 4135 * order to prevent the caller from issuing a read 4136 * ( due to B_CACHE not being set ) and overwriting 4137 * it. 4138 * 4139 * Most callers, including NFS and FFS, need this to 4140 * operate properly either because they assume they 4141 * can issue a read if B_CACHE is not set, or because 4142 * ( for example ) an uncached B_DELWRI might loop due 4143 * to softupdates re-dirtying the buffer. In the latter 4144 * case, B_CACHE is set after the first write completes, 4145 * preventing further loops. 4146 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 4147 * above while extending the buffer, we cannot allow the 4148 * buffer to remain with B_CACHE set after the write 4149 * completes or it will represent a corrupt state. To 4150 * deal with this we set B_NOCACHE to scrap the buffer 4151 * after the write. 4152 * 4153 * We might be able to do something fancy, like setting 4154 * B_CACHE in bwrite() except if B_DELWRI is already set, 4155 * so the below call doesn't set B_CACHE, but that gets real 4156 * confusing. This is much easier. 4157 */ 4158 4159 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 4160 bp->b_flags |= B_NOCACHE; 4161 bwrite(bp); 4162 goto loop; 4163 } 4164 bp->b_flags &= ~B_DONE; 4165 } else { 4166 /* 4167 * Buffer is not in-core, create new buffer. The buffer 4168 * returned by getnewbuf() is locked. Note that the returned 4169 * buffer is also considered valid (not marked B_INVAL). 4170 */ 4171 BO_RUNLOCK(bo); 4172 newbuf_unlocked: 4173 /* 4174 * If the user does not want us to create the buffer, bail out 4175 * here. 4176 */ 4177 if (flags & GB_NOCREAT) 4178 return (EEXIST); 4179 4180 bsize = vn_isdisk(vp) ? DEV_BSIZE : bo->bo_bsize; 4181 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 4182 offset = blkno * bsize; 4183 vmio = vp->v_object != NULL; 4184 if (vmio) { 4185 maxsize = size + (offset & PAGE_MASK); 4186 if (maxsize > maxbcachebuf) { 4187 printf( 4188 "getblkx: maxsize(%d) > maxbcachebuf(%d)\n", 4189 maxsize, maxbcachebuf); 4190 return (EIO); 4191 } 4192 } else { 4193 maxsize = size; 4194 /* Do not allow non-VMIO notmapped buffers. */ 4195 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 4196 } 4197 maxsize = imax(maxsize, bsize); 4198 if ((flags & GB_NOSPARSE) != 0 && vmio && 4199 !vn_isdisk(vp)) { 4200 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0); 4201 KASSERT(error != EOPNOTSUPP, 4202 ("GB_NOSPARSE from fs not supporting bmap, vp %p", 4203 vp)); 4204 if (error != 0) 4205 return (error); 4206 if (d_blkno == -1) 4207 return (EJUSTRETURN); 4208 } 4209 4210 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags); 4211 if (bp == NULL) { 4212 if (slpflag || slptimeo) 4213 return (ETIMEDOUT); 4214 /* 4215 * XXX This is here until the sleep path is diagnosed 4216 * enough to work under very low memory conditions. 4217 * 4218 * There's an issue on low memory, 4BSD+non-preempt 4219 * systems (eg MIPS routers with 32MB RAM) where buffer 4220 * exhaustion occurs without sleeping for buffer 4221 * reclaimation. This just sticks in a loop and 4222 * constantly attempts to allocate a buffer, which 4223 * hits exhaustion and tries to wakeup bufdaemon. 4224 * This never happens because we never yield. 4225 * 4226 * The real solution is to identify and fix these cases 4227 * so we aren't effectively busy-waiting in a loop 4228 * until the reclaimation path has cycles to run. 4229 */ 4230 kern_yield(PRI_USER); 4231 goto loop; 4232 } 4233 4234 /* 4235 * 4236 * Insert the buffer into the hash, so that it can 4237 * be found by incore. 4238 * 4239 * We don't hold the bufobj interlock while allocating the new 4240 * buffer. Consequently, we can race on buffer creation. This 4241 * can be a problem whether the vnode is locked or not. If the 4242 * buffer is created out from under us, we have to throw away 4243 * the one we just created. 4244 */ 4245 bp->b_lblkno = blkno; 4246 bp->b_blkno = d_blkno; 4247 bp->b_offset = offset; 4248 error = bgetvp(vp, bp); 4249 if (error != 0) { 4250 KASSERT(error == EEXIST, 4251 ("getblk: unexpected error %d from bgetvp", 4252 error)); 4253 bp->b_flags |= B_INVAL; 4254 bufspace_release(bufdomain(bp), maxsize); 4255 brelse(bp); 4256 goto loop; 4257 } 4258 4259 /* 4260 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 4261 * buffer size starts out as 0, B_CACHE will be set by 4262 * allocbuf() for the VMIO case prior to it testing the 4263 * backing store for validity. 4264 */ 4265 4266 if (vmio) { 4267 bp->b_flags |= B_VMIO; 4268 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 4269 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 4270 bp, vp->v_object, bp->b_bufobj->bo_object)); 4271 } else { 4272 bp->b_flags &= ~B_VMIO; 4273 KASSERT(bp->b_bufobj->bo_object == NULL, 4274 ("ARGH! has b_bufobj->bo_object %p %p\n", 4275 bp, bp->b_bufobj->bo_object)); 4276 BUF_CHECK_MAPPED(bp); 4277 } 4278 4279 allocbuf(bp, size); 4280 bufspace_release(bufdomain(bp), maxsize); 4281 bp->b_flags &= ~B_DONE; 4282 } 4283 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 4284 end: 4285 buf_track(bp, __func__); 4286 KASSERT(bp->b_bufobj == bo, 4287 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 4288 *bpp = bp; 4289 return (0); 4290 } 4291 4292 /* 4293 * Get an empty, disassociated buffer of given size. The buffer is initially 4294 * set to B_INVAL. 4295 */ 4296 struct buf * 4297 geteblk(int size, int flags) 4298 { 4299 struct buf *bp; 4300 int maxsize; 4301 4302 maxsize = (size + BKVAMASK) & ~BKVAMASK; 4303 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) { 4304 if ((flags & GB_NOWAIT_BD) && 4305 (curthread->td_pflags & TDP_BUFNEED) != 0) 4306 return (NULL); 4307 } 4308 allocbuf(bp, size); 4309 bufspace_release(bufdomain(bp), maxsize); 4310 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 4311 return (bp); 4312 } 4313 4314 /* 4315 * Truncate the backing store for a non-vmio buffer. 4316 */ 4317 static void 4318 vfs_nonvmio_truncate(struct buf *bp, int newbsize) 4319 { 4320 4321 if (bp->b_flags & B_MALLOC) { 4322 /* 4323 * malloced buffers are not shrunk 4324 */ 4325 if (newbsize == 0) { 4326 bufmallocadjust(bp, 0); 4327 free(bp->b_data, M_BIOBUF); 4328 bp->b_data = bp->b_kvabase; 4329 bp->b_flags &= ~B_MALLOC; 4330 } 4331 return; 4332 } 4333 vm_hold_free_pages(bp, newbsize); 4334 bufspace_adjust(bp, newbsize); 4335 } 4336 4337 /* 4338 * Extend the backing for a non-VMIO buffer. 4339 */ 4340 static void 4341 vfs_nonvmio_extend(struct buf *bp, int newbsize) 4342 { 4343 caddr_t origbuf; 4344 int origbufsize; 4345 4346 /* 4347 * We only use malloced memory on the first allocation. 4348 * and revert to page-allocated memory when the buffer 4349 * grows. 4350 * 4351 * There is a potential smp race here that could lead 4352 * to bufmallocspace slightly passing the max. It 4353 * is probably extremely rare and not worth worrying 4354 * over. 4355 */ 4356 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 && 4357 bufmallocspace < maxbufmallocspace) { 4358 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK); 4359 bp->b_flags |= B_MALLOC; 4360 bufmallocadjust(bp, newbsize); 4361 return; 4362 } 4363 4364 /* 4365 * If the buffer is growing on its other-than-first 4366 * allocation then we revert to the page-allocation 4367 * scheme. 4368 */ 4369 origbuf = NULL; 4370 origbufsize = 0; 4371 if (bp->b_flags & B_MALLOC) { 4372 origbuf = bp->b_data; 4373 origbufsize = bp->b_bufsize; 4374 bp->b_data = bp->b_kvabase; 4375 bufmallocadjust(bp, 0); 4376 bp->b_flags &= ~B_MALLOC; 4377 newbsize = round_page(newbsize); 4378 } 4379 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize, 4380 (vm_offset_t) bp->b_data + newbsize); 4381 if (origbuf != NULL) { 4382 bcopy(origbuf, bp->b_data, origbufsize); 4383 free(origbuf, M_BIOBUF); 4384 } 4385 bufspace_adjust(bp, newbsize); 4386 } 4387 4388 /* 4389 * This code constitutes the buffer memory from either anonymous system 4390 * memory (in the case of non-VMIO operations) or from an associated 4391 * VM object (in the case of VMIO operations). This code is able to 4392 * resize a buffer up or down. 4393 * 4394 * Note that this code is tricky, and has many complications to resolve 4395 * deadlock or inconsistent data situations. Tread lightly!!! 4396 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 4397 * the caller. Calling this code willy nilly can result in the loss of data. 4398 * 4399 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 4400 * B_CACHE for the non-VMIO case. 4401 */ 4402 int 4403 allocbuf(struct buf *bp, int size) 4404 { 4405 int newbsize; 4406 4407 if (bp->b_bcount == size) 4408 return (1); 4409 4410 KASSERT(bp->b_kvasize == 0 || bp->b_kvasize >= size, 4411 ("allocbuf: buffer too small %p %#x %#x", 4412 bp, bp->b_kvasize, size)); 4413 4414 newbsize = roundup2(size, DEV_BSIZE); 4415 if ((bp->b_flags & B_VMIO) == 0) { 4416 if ((bp->b_flags & B_MALLOC) == 0) 4417 newbsize = round_page(newbsize); 4418 /* 4419 * Just get anonymous memory from the kernel. Don't 4420 * mess with B_CACHE. 4421 */ 4422 if (newbsize < bp->b_bufsize) 4423 vfs_nonvmio_truncate(bp, newbsize); 4424 else if (newbsize > bp->b_bufsize) 4425 vfs_nonvmio_extend(bp, newbsize); 4426 } else { 4427 int desiredpages; 4428 4429 desiredpages = size == 0 ? 0 : 4430 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 4431 4432 KASSERT((bp->b_flags & B_MALLOC) == 0, 4433 ("allocbuf: VMIO buffer can't be malloced %p", bp)); 4434 4435 /* 4436 * Set B_CACHE initially if buffer is 0 length or will become 4437 * 0-length. 4438 */ 4439 if (size == 0 || bp->b_bufsize == 0) 4440 bp->b_flags |= B_CACHE; 4441 4442 if (newbsize < bp->b_bufsize) 4443 vfs_vmio_truncate(bp, desiredpages); 4444 /* XXX This looks as if it should be newbsize > b_bufsize */ 4445 else if (size > bp->b_bcount) 4446 vfs_vmio_extend(bp, desiredpages, size); 4447 bufspace_adjust(bp, newbsize); 4448 } 4449 bp->b_bcount = size; /* requested buffer size. */ 4450 return (1); 4451 } 4452 4453 extern int inflight_transient_maps; 4454 4455 static struct bio_queue nondump_bios; 4456 4457 void 4458 biodone(struct bio *bp) 4459 { 4460 struct mtx *mtxp; 4461 void (*done)(struct bio *); 4462 vm_offset_t start, end; 4463 4464 biotrack(bp, __func__); 4465 4466 /* 4467 * Avoid completing I/O when dumping after a panic since that may 4468 * result in a deadlock in the filesystem or pager code. Note that 4469 * this doesn't affect dumps that were started manually since we aim 4470 * to keep the system usable after it has been resumed. 4471 */ 4472 if (__predict_false(dumping && SCHEDULER_STOPPED())) { 4473 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue); 4474 return; 4475 } 4476 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 4477 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 4478 bp->bio_flags |= BIO_UNMAPPED; 4479 start = trunc_page((vm_offset_t)bp->bio_data); 4480 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 4481 bp->bio_data = unmapped_buf; 4482 pmap_qremove(start, atop(end - start)); 4483 vmem_free(transient_arena, start, end - start); 4484 atomic_add_int(&inflight_transient_maps, -1); 4485 } 4486 done = bp->bio_done; 4487 /* 4488 * The check for done == biodone is to allow biodone to be 4489 * used as a bio_done routine. 4490 */ 4491 if (done == NULL || done == biodone) { 4492 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4493 mtx_lock(mtxp); 4494 bp->bio_flags |= BIO_DONE; 4495 wakeup(bp); 4496 mtx_unlock(mtxp); 4497 } else 4498 done(bp); 4499 } 4500 4501 /* 4502 * Wait for a BIO to finish. 4503 */ 4504 int 4505 biowait(struct bio *bp, const char *wmesg) 4506 { 4507 struct mtx *mtxp; 4508 4509 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4510 mtx_lock(mtxp); 4511 while ((bp->bio_flags & BIO_DONE) == 0) 4512 msleep(bp, mtxp, PRIBIO, wmesg, 0); 4513 mtx_unlock(mtxp); 4514 if (bp->bio_error != 0) 4515 return (bp->bio_error); 4516 if (!(bp->bio_flags & BIO_ERROR)) 4517 return (0); 4518 return (EIO); 4519 } 4520 4521 void 4522 biofinish(struct bio *bp, struct devstat *stat, int error) 4523 { 4524 4525 if (error) { 4526 bp->bio_error = error; 4527 bp->bio_flags |= BIO_ERROR; 4528 } 4529 if (stat != NULL) 4530 devstat_end_transaction_bio(stat, bp); 4531 biodone(bp); 4532 } 4533 4534 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING) 4535 void 4536 biotrack_buf(struct bio *bp, const char *location) 4537 { 4538 4539 buf_track(bp->bio_track_bp, location); 4540 } 4541 #endif 4542 4543 /* 4544 * bufwait: 4545 * 4546 * Wait for buffer I/O completion, returning error status. The buffer 4547 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 4548 * error and cleared. 4549 */ 4550 int 4551 bufwait(struct buf *bp) 4552 { 4553 if (bp->b_iocmd == BIO_READ) 4554 bwait(bp, PRIBIO, "biord"); 4555 else 4556 bwait(bp, PRIBIO, "biowr"); 4557 if (bp->b_flags & B_EINTR) { 4558 bp->b_flags &= ~B_EINTR; 4559 return (EINTR); 4560 } 4561 if (bp->b_ioflags & BIO_ERROR) { 4562 return (bp->b_error ? bp->b_error : EIO); 4563 } else { 4564 return (0); 4565 } 4566 } 4567 4568 /* 4569 * bufdone: 4570 * 4571 * Finish I/O on a buffer, optionally calling a completion function. 4572 * This is usually called from an interrupt so process blocking is 4573 * not allowed. 4574 * 4575 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 4576 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 4577 * assuming B_INVAL is clear. 4578 * 4579 * For the VMIO case, we set B_CACHE if the op was a read and no 4580 * read error occurred, or if the op was a write. B_CACHE is never 4581 * set if the buffer is invalid or otherwise uncacheable. 4582 * 4583 * bufdone does not mess with B_INVAL, allowing the I/O routine or the 4584 * initiator to leave B_INVAL set to brelse the buffer out of existence 4585 * in the biodone routine. 4586 */ 4587 void 4588 bufdone(struct buf *bp) 4589 { 4590 struct bufobj *dropobj; 4591 void (*biodone)(struct buf *); 4592 4593 buf_track(bp, __func__); 4594 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 4595 dropobj = NULL; 4596 4597 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 4598 4599 runningbufwakeup(bp); 4600 if (bp->b_iocmd == BIO_WRITE) 4601 dropobj = bp->b_bufobj; 4602 /* call optional completion function if requested */ 4603 if (bp->b_iodone != NULL) { 4604 biodone = bp->b_iodone; 4605 bp->b_iodone = NULL; 4606 (*biodone) (bp); 4607 if (dropobj) 4608 bufobj_wdrop(dropobj); 4609 return; 4610 } 4611 if (bp->b_flags & B_VMIO) { 4612 /* 4613 * Set B_CACHE if the op was a normal read and no error 4614 * occurred. B_CACHE is set for writes in the b*write() 4615 * routines. 4616 */ 4617 if (bp->b_iocmd == BIO_READ && 4618 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 4619 !(bp->b_ioflags & BIO_ERROR)) 4620 bp->b_flags |= B_CACHE; 4621 vfs_vmio_iodone(bp); 4622 } 4623 if (!LIST_EMPTY(&bp->b_dep)) 4624 buf_complete(bp); 4625 if ((bp->b_flags & B_CKHASH) != 0) { 4626 KASSERT(bp->b_iocmd == BIO_READ, 4627 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd)); 4628 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp)); 4629 (*bp->b_ckhashcalc)(bp); 4630 } 4631 /* 4632 * For asynchronous completions, release the buffer now. The brelse 4633 * will do a wakeup there if necessary - so no need to do a wakeup 4634 * here in the async case. The sync case always needs to do a wakeup. 4635 */ 4636 if (bp->b_flags & B_ASYNC) { 4637 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || 4638 (bp->b_ioflags & BIO_ERROR)) 4639 brelse(bp); 4640 else 4641 bqrelse(bp); 4642 } else 4643 bdone(bp); 4644 if (dropobj) 4645 bufobj_wdrop(dropobj); 4646 } 4647 4648 /* 4649 * This routine is called in lieu of iodone in the case of 4650 * incomplete I/O. This keeps the busy status for pages 4651 * consistent. 4652 */ 4653 void 4654 vfs_unbusy_pages(struct buf *bp) 4655 { 4656 int i; 4657 vm_object_t obj; 4658 vm_page_t m; 4659 4660 runningbufwakeup(bp); 4661 if (!(bp->b_flags & B_VMIO)) 4662 return; 4663 4664 obj = bp->b_bufobj->bo_object; 4665 for (i = 0; i < bp->b_npages; i++) { 4666 m = bp->b_pages[i]; 4667 if (m == bogus_page) { 4668 m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i); 4669 if (!m) 4670 panic("vfs_unbusy_pages: page missing\n"); 4671 bp->b_pages[i] = m; 4672 if (buf_mapped(bp)) { 4673 BUF_CHECK_MAPPED(bp); 4674 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4675 bp->b_pages, bp->b_npages); 4676 } else 4677 BUF_CHECK_UNMAPPED(bp); 4678 } 4679 vm_page_sunbusy(m); 4680 } 4681 vm_object_pip_wakeupn(obj, bp->b_npages); 4682 } 4683 4684 /* 4685 * vfs_page_set_valid: 4686 * 4687 * Set the valid bits in a page based on the supplied offset. The 4688 * range is restricted to the buffer's size. 4689 * 4690 * This routine is typically called after a read completes. 4691 */ 4692 static void 4693 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4694 { 4695 vm_ooffset_t eoff; 4696 4697 /* 4698 * Compute the end offset, eoff, such that [off, eoff) does not span a 4699 * page boundary and eoff is not greater than the end of the buffer. 4700 * The end of the buffer, in this case, is our file EOF, not the 4701 * allocation size of the buffer. 4702 */ 4703 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_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 > off) 4712 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 4713 } 4714 4715 /* 4716 * vfs_page_set_validclean: 4717 * 4718 * Set the valid bits and clear the dirty bits in a page based on the 4719 * supplied offset. The range is restricted to the buffer's size. 4720 */ 4721 static void 4722 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4723 { 4724 vm_ooffset_t soff, eoff; 4725 4726 /* 4727 * Start and end offsets in buffer. eoff - soff may not cross a 4728 * page boundary or cross the end of the buffer. The end of the 4729 * buffer, in this case, is our file EOF, not the allocation size 4730 * of the buffer. 4731 */ 4732 soff = off; 4733 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4734 if (eoff > bp->b_offset + bp->b_bcount) 4735 eoff = bp->b_offset + bp->b_bcount; 4736 4737 /* 4738 * Set valid range. This is typically the entire buffer and thus the 4739 * entire page. 4740 */ 4741 if (eoff > soff) { 4742 vm_page_set_validclean( 4743 m, 4744 (vm_offset_t) (soff & PAGE_MASK), 4745 (vm_offset_t) (eoff - soff) 4746 ); 4747 } 4748 } 4749 4750 /* 4751 * Acquire a shared busy on all pages in the buf. 4752 */ 4753 void 4754 vfs_busy_pages_acquire(struct buf *bp) 4755 { 4756 int i; 4757 4758 for (i = 0; i < bp->b_npages; i++) 4759 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY); 4760 } 4761 4762 void 4763 vfs_busy_pages_release(struct buf *bp) 4764 { 4765 int i; 4766 4767 for (i = 0; i < bp->b_npages; i++) 4768 vm_page_sunbusy(bp->b_pages[i]); 4769 } 4770 4771 /* 4772 * This routine is called before a device strategy routine. 4773 * It is used to tell the VM system that paging I/O is in 4774 * progress, and treat the pages associated with the buffer 4775 * almost as being exclusive busy. Also the object paging_in_progress 4776 * flag is handled to make sure that the object doesn't become 4777 * inconsistent. 4778 * 4779 * Since I/O has not been initiated yet, certain buffer flags 4780 * such as BIO_ERROR or B_INVAL may be in an inconsistent state 4781 * and should be ignored. 4782 */ 4783 void 4784 vfs_busy_pages(struct buf *bp, int clear_modify) 4785 { 4786 vm_object_t obj; 4787 vm_ooffset_t foff; 4788 vm_page_t m; 4789 int i; 4790 bool bogus; 4791 4792 if (!(bp->b_flags & B_VMIO)) 4793 return; 4794 4795 obj = bp->b_bufobj->bo_object; 4796 foff = bp->b_offset; 4797 KASSERT(bp->b_offset != NOOFFSET, 4798 ("vfs_busy_pages: no buffer offset")); 4799 if ((bp->b_flags & B_CLUSTER) == 0) { 4800 vm_object_pip_add(obj, bp->b_npages); 4801 vfs_busy_pages_acquire(bp); 4802 } 4803 if (bp->b_bufsize != 0) 4804 vfs_setdirty_range(bp); 4805 bogus = false; 4806 for (i = 0; i < bp->b_npages; i++) { 4807 m = bp->b_pages[i]; 4808 vm_page_assert_sbusied(m); 4809 4810 /* 4811 * When readying a buffer for a read ( i.e 4812 * clear_modify == 0 ), it is important to do 4813 * bogus_page replacement for valid pages in 4814 * partially instantiated buffers. Partially 4815 * instantiated buffers can, in turn, occur when 4816 * reconstituting a buffer from its VM backing store 4817 * base. We only have to do this if B_CACHE is 4818 * clear ( which causes the I/O to occur in the 4819 * first place ). The replacement prevents the read 4820 * I/O from overwriting potentially dirty VM-backed 4821 * pages. XXX bogus page replacement is, uh, bogus. 4822 * It may not work properly with small-block devices. 4823 * We need to find a better way. 4824 */ 4825 if (clear_modify) { 4826 pmap_remove_write(m); 4827 vfs_page_set_validclean(bp, foff, m); 4828 } else if (vm_page_all_valid(m) && 4829 (bp->b_flags & B_CACHE) == 0) { 4830 bp->b_pages[i] = bogus_page; 4831 bogus = true; 4832 } 4833 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4834 } 4835 if (bogus && buf_mapped(bp)) { 4836 BUF_CHECK_MAPPED(bp); 4837 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4838 bp->b_pages, bp->b_npages); 4839 } 4840 } 4841 4842 /* 4843 * vfs_bio_set_valid: 4844 * 4845 * Set the range within the buffer to valid. The range is 4846 * relative to the beginning of the buffer, b_offset. Note that 4847 * b_offset itself may be offset from the beginning of the first 4848 * page. 4849 */ 4850 void 4851 vfs_bio_set_valid(struct buf *bp, int base, int size) 4852 { 4853 int i, n; 4854 vm_page_t m; 4855 4856 if (!(bp->b_flags & B_VMIO)) 4857 return; 4858 4859 /* 4860 * Fixup base to be relative to beginning of first page. 4861 * Set initial n to be the maximum number of bytes in the 4862 * first page that can be validated. 4863 */ 4864 base += (bp->b_offset & PAGE_MASK); 4865 n = PAGE_SIZE - (base & PAGE_MASK); 4866 4867 /* 4868 * Busy may not be strictly necessary here because the pages are 4869 * unlikely to be fully valid and the vnode lock will synchronize 4870 * their access via getpages. It is grabbed for consistency with 4871 * other page validation. 4872 */ 4873 vfs_busy_pages_acquire(bp); 4874 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4875 m = bp->b_pages[i]; 4876 if (n > size) 4877 n = size; 4878 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4879 base += n; 4880 size -= n; 4881 n = PAGE_SIZE; 4882 } 4883 vfs_busy_pages_release(bp); 4884 } 4885 4886 /* 4887 * vfs_bio_clrbuf: 4888 * 4889 * If the specified buffer is a non-VMIO buffer, clear the entire 4890 * buffer. If the specified buffer is a VMIO buffer, clear and 4891 * validate only the previously invalid portions of the buffer. 4892 * This routine essentially fakes an I/O, so we need to clear 4893 * BIO_ERROR and B_INVAL. 4894 * 4895 * Note that while we only theoretically need to clear through b_bcount, 4896 * we go ahead and clear through b_bufsize. 4897 */ 4898 void 4899 vfs_bio_clrbuf(struct buf *bp) 4900 { 4901 int i, j, sa, ea, slide, zbits; 4902 vm_page_bits_t mask; 4903 4904 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4905 clrbuf(bp); 4906 return; 4907 } 4908 bp->b_flags &= ~B_INVAL; 4909 bp->b_ioflags &= ~BIO_ERROR; 4910 vfs_busy_pages_acquire(bp); 4911 sa = bp->b_offset & PAGE_MASK; 4912 slide = 0; 4913 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4914 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4915 ea = slide & PAGE_MASK; 4916 if (ea == 0) 4917 ea = PAGE_SIZE; 4918 if (bp->b_pages[i] == bogus_page) 4919 continue; 4920 j = sa / DEV_BSIZE; 4921 zbits = (sizeof(vm_page_bits_t) * NBBY) - 4922 (ea - sa) / DEV_BSIZE; 4923 mask = (VM_PAGE_BITS_ALL >> zbits) << j; 4924 if ((bp->b_pages[i]->valid & mask) == mask) 4925 continue; 4926 if ((bp->b_pages[i]->valid & mask) == 0) 4927 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4928 else { 4929 for (; sa < ea; sa += DEV_BSIZE, j++) { 4930 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4931 pmap_zero_page_area(bp->b_pages[i], 4932 sa, DEV_BSIZE); 4933 } 4934 } 4935 } 4936 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE, 4937 roundup2(ea - sa, DEV_BSIZE)); 4938 } 4939 vfs_busy_pages_release(bp); 4940 bp->b_resid = 0; 4941 } 4942 4943 void 4944 vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4945 { 4946 vm_page_t m; 4947 int i, n; 4948 4949 if (buf_mapped(bp)) { 4950 BUF_CHECK_MAPPED(bp); 4951 bzero(bp->b_data + base, size); 4952 } else { 4953 BUF_CHECK_UNMAPPED(bp); 4954 n = PAGE_SIZE - (base & PAGE_MASK); 4955 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4956 m = bp->b_pages[i]; 4957 if (n > size) 4958 n = size; 4959 pmap_zero_page_area(m, base & PAGE_MASK, n); 4960 base += n; 4961 size -= n; 4962 n = PAGE_SIZE; 4963 } 4964 } 4965 } 4966 4967 /* 4968 * Update buffer flags based on I/O request parameters, optionally releasing the 4969 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM, 4970 * where they may be placed on a page queue (VMIO) or freed immediately (direct 4971 * I/O). Otherwise the buffer is released to the cache. 4972 */ 4973 static void 4974 b_io_dismiss(struct buf *bp, int ioflag, bool release) 4975 { 4976 4977 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0, 4978 ("buf %p non-VMIO noreuse", bp)); 4979 4980 if ((ioflag & IO_DIRECT) != 0) 4981 bp->b_flags |= B_DIRECT; 4982 if ((ioflag & IO_EXT) != 0) 4983 bp->b_xflags |= BX_ALTDATA; 4984 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) { 4985 bp->b_flags |= B_RELBUF; 4986 if ((ioflag & IO_NOREUSE) != 0) 4987 bp->b_flags |= B_NOREUSE; 4988 if (release) 4989 brelse(bp); 4990 } else if (release) 4991 bqrelse(bp); 4992 } 4993 4994 void 4995 vfs_bio_brelse(struct buf *bp, int ioflag) 4996 { 4997 4998 b_io_dismiss(bp, ioflag, true); 4999 } 5000 5001 void 5002 vfs_bio_set_flags(struct buf *bp, int ioflag) 5003 { 5004 5005 b_io_dismiss(bp, ioflag, false); 5006 } 5007 5008 /* 5009 * vm_hold_load_pages and vm_hold_free_pages get pages into 5010 * a buffers address space. The pages are anonymous and are 5011 * not associated with a file object. 5012 */ 5013 static void 5014 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 5015 { 5016 vm_offset_t pg; 5017 vm_page_t p; 5018 int index; 5019 5020 BUF_CHECK_MAPPED(bp); 5021 5022 to = round_page(to); 5023 from = round_page(from); 5024 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 5025 MPASS((bp->b_flags & B_MAXPHYS) == 0); 5026 KASSERT(to - from <= maxbcachebuf, 5027 ("vm_hold_load_pages too large %p %#jx %#jx %u", 5028 bp, (uintmax_t)from, (uintmax_t)to, maxbcachebuf)); 5029 5030 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 5031 /* 5032 * note: must allocate system pages since blocking here 5033 * could interfere with paging I/O, no matter which 5034 * process we are. 5035 */ 5036 p = vm_page_alloc_noobj(VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 5037 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK); 5038 pmap_qenter(pg, &p, 1); 5039 bp->b_pages[index] = p; 5040 } 5041 bp->b_npages = index; 5042 } 5043 5044 /* Return pages associated with this buf to the vm system */ 5045 static void 5046 vm_hold_free_pages(struct buf *bp, int newbsize) 5047 { 5048 vm_offset_t from; 5049 vm_page_t p; 5050 int index, newnpages; 5051 5052 BUF_CHECK_MAPPED(bp); 5053 5054 from = round_page((vm_offset_t)bp->b_data + newbsize); 5055 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 5056 if (bp->b_npages > newnpages) 5057 pmap_qremove(from, bp->b_npages - newnpages); 5058 for (index = newnpages; index < bp->b_npages; index++) { 5059 p = bp->b_pages[index]; 5060 bp->b_pages[index] = NULL; 5061 vm_page_unwire_noq(p); 5062 vm_page_free(p); 5063 } 5064 bp->b_npages = newnpages; 5065 } 5066 5067 /* 5068 * Map an IO request into kernel virtual address space. 5069 * 5070 * All requests are (re)mapped into kernel VA space. 5071 * Notice that we use b_bufsize for the size of the buffer 5072 * to be mapped. b_bcount might be modified by the driver. 5073 * 5074 * Note that even if the caller determines that the address space should 5075 * be valid, a race or a smaller-file mapped into a larger space may 5076 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 5077 * check the return value. 5078 * 5079 * This function only works with pager buffers. 5080 */ 5081 int 5082 vmapbuf(struct buf *bp, void *uaddr, size_t len, int mapbuf) 5083 { 5084 vm_prot_t prot; 5085 int pidx; 5086 5087 MPASS((bp->b_flags & B_MAXPHYS) != 0); 5088 prot = VM_PROT_READ; 5089 if (bp->b_iocmd == BIO_READ) 5090 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 5091 pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 5092 (vm_offset_t)uaddr, len, prot, bp->b_pages, PBUF_PAGES); 5093 if (pidx < 0) 5094 return (-1); 5095 bp->b_bufsize = len; 5096 bp->b_npages = pidx; 5097 bp->b_offset = ((vm_offset_t)uaddr) & PAGE_MASK; 5098 if (mapbuf || !unmapped_buf_allowed) { 5099 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx); 5100 bp->b_data = bp->b_kvabase + bp->b_offset; 5101 } else 5102 bp->b_data = unmapped_buf; 5103 return (0); 5104 } 5105 5106 /* 5107 * Free the io map PTEs associated with this IO operation. 5108 * We also invalidate the TLB entries and restore the original b_addr. 5109 * 5110 * This function only works with pager buffers. 5111 */ 5112 void 5113 vunmapbuf(struct buf *bp) 5114 { 5115 int npages; 5116 5117 npages = bp->b_npages; 5118 if (buf_mapped(bp)) 5119 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 5120 vm_page_unhold_pages(bp->b_pages, npages); 5121 5122 bp->b_data = unmapped_buf; 5123 } 5124 5125 void 5126 bdone(struct buf *bp) 5127 { 5128 struct mtx *mtxp; 5129 5130 mtxp = mtx_pool_find(mtxpool_sleep, bp); 5131 mtx_lock(mtxp); 5132 bp->b_flags |= B_DONE; 5133 wakeup(bp); 5134 mtx_unlock(mtxp); 5135 } 5136 5137 void 5138 bwait(struct buf *bp, u_char pri, const char *wchan) 5139 { 5140 struct mtx *mtxp; 5141 5142 mtxp = mtx_pool_find(mtxpool_sleep, bp); 5143 mtx_lock(mtxp); 5144 while ((bp->b_flags & B_DONE) == 0) 5145 msleep(bp, mtxp, pri, wchan, 0); 5146 mtx_unlock(mtxp); 5147 } 5148 5149 int 5150 bufsync(struct bufobj *bo, int waitfor) 5151 { 5152 5153 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread)); 5154 } 5155 5156 void 5157 bufstrategy(struct bufobj *bo, struct buf *bp) 5158 { 5159 int i __unused; 5160 struct vnode *vp; 5161 5162 vp = bp->b_vp; 5163 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 5164 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 5165 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 5166 i = VOP_STRATEGY(vp, bp); 5167 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 5168 } 5169 5170 /* 5171 * Initialize a struct bufobj before use. Memory is assumed zero filled. 5172 */ 5173 void 5174 bufobj_init(struct bufobj *bo, void *private) 5175 { 5176 static volatile int bufobj_cleanq; 5177 5178 bo->bo_domain = 5179 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains; 5180 rw_init(BO_LOCKPTR(bo), "bufobj interlock"); 5181 bo->bo_private = private; 5182 TAILQ_INIT(&bo->bo_clean.bv_hd); 5183 pctrie_init(&bo->bo_clean.bv_root); 5184 TAILQ_INIT(&bo->bo_dirty.bv_hd); 5185 pctrie_init(&bo->bo_dirty.bv_root); 5186 } 5187 5188 void 5189 bufobj_wrefl(struct bufobj *bo) 5190 { 5191 5192 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 5193 ASSERT_BO_WLOCKED(bo); 5194 bo->bo_numoutput++; 5195 } 5196 5197 void 5198 bufobj_wref(struct bufobj *bo) 5199 { 5200 5201 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 5202 BO_LOCK(bo); 5203 bo->bo_numoutput++; 5204 BO_UNLOCK(bo); 5205 } 5206 5207 void 5208 bufobj_wdrop(struct bufobj *bo) 5209 { 5210 5211 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 5212 BO_LOCK(bo); 5213 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 5214 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 5215 bo->bo_flag &= ~BO_WWAIT; 5216 wakeup(&bo->bo_numoutput); 5217 } 5218 BO_UNLOCK(bo); 5219 } 5220 5221 int 5222 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 5223 { 5224 int error; 5225 5226 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 5227 ASSERT_BO_WLOCKED(bo); 5228 error = 0; 5229 while (bo->bo_numoutput) { 5230 bo->bo_flag |= BO_WWAIT; 5231 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 5232 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 5233 if (error) 5234 break; 5235 } 5236 return (error); 5237 } 5238 5239 /* 5240 * Set bio_data or bio_ma for struct bio from the struct buf. 5241 */ 5242 void 5243 bdata2bio(struct buf *bp, struct bio *bip) 5244 { 5245 5246 if (!buf_mapped(bp)) { 5247 KASSERT(unmapped_buf_allowed, ("unmapped")); 5248 bip->bio_ma = bp->b_pages; 5249 bip->bio_ma_n = bp->b_npages; 5250 bip->bio_data = unmapped_buf; 5251 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 5252 bip->bio_flags |= BIO_UNMAPPED; 5253 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 5254 PAGE_SIZE == bp->b_npages, 5255 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 5256 (long long)bip->bio_length, bip->bio_ma_n)); 5257 } else { 5258 bip->bio_data = bp->b_data; 5259 bip->bio_ma = NULL; 5260 } 5261 } 5262 5263 struct memdesc 5264 memdesc_bio(struct bio *bio) 5265 { 5266 if ((bio->bio_flags & BIO_VLIST) != 0) 5267 return (memdesc_vlist((struct bus_dma_segment *)bio->bio_data, 5268 bio->bio_ma_n)); 5269 5270 if ((bio->bio_flags & BIO_UNMAPPED) != 0) 5271 return (memdesc_vmpages(bio->bio_ma, bio->bio_bcount, 5272 bio->bio_ma_offset)); 5273 5274 return (memdesc_vaddr(bio->bio_data, bio->bio_bcount)); 5275 } 5276 5277 static int buf_pager_relbuf; 5278 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN, 5279 &buf_pager_relbuf, 0, 5280 "Make buffer pager release buffers after reading"); 5281 5282 /* 5283 * The buffer pager. It uses buffer reads to validate pages. 5284 * 5285 * In contrast to the generic local pager from vm/vnode_pager.c, this 5286 * pager correctly and easily handles volumes where the underlying 5287 * device block size is greater than the machine page size. The 5288 * buffer cache transparently extends the requested page run to be 5289 * aligned at the block boundary, and does the necessary bogus page 5290 * replacements in the addends to avoid obliterating already valid 5291 * pages. 5292 * 5293 * The only non-trivial issue is that the exclusive busy state for 5294 * pages, which is assumed by the vm_pager_getpages() interface, is 5295 * incompatible with the VMIO buffer cache's desire to share-busy the 5296 * pages. This function performs a trivial downgrade of the pages' 5297 * state before reading buffers, and a less trivial upgrade from the 5298 * shared-busy to excl-busy state after the read. 5299 */ 5300 int 5301 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count, 5302 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno, 5303 vbg_get_blksize_t get_blksize) 5304 { 5305 vm_page_t m; 5306 vm_object_t object; 5307 struct buf *bp; 5308 struct mount *mp; 5309 daddr_t lbn, lbnp; 5310 vm_ooffset_t la, lb, poff, poffe; 5311 long bo_bs, bsize; 5312 int br_flags, error, i, pgsin, pgsin_a, pgsin_b; 5313 bool redo, lpart; 5314 5315 object = vp->v_object; 5316 mp = vp->v_mount; 5317 error = 0; 5318 la = IDX_TO_OFF(ma[count - 1]->pindex); 5319 if (la >= object->un_pager.vnp.vnp_size) 5320 return (VM_PAGER_BAD); 5321 5322 /* 5323 * Change the meaning of la from where the last requested page starts 5324 * to where it ends, because that's the end of the requested region 5325 * and the start of the potential read-ahead region. 5326 */ 5327 la += PAGE_SIZE; 5328 lpart = la > object->un_pager.vnp.vnp_size; 5329 error = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex)), 5330 &bo_bs); 5331 if (error != 0) 5332 return (VM_PAGER_ERROR); 5333 5334 /* 5335 * Calculate read-ahead, behind and total pages. 5336 */ 5337 pgsin = count; 5338 lb = IDX_TO_OFF(ma[0]->pindex); 5339 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs)); 5340 pgsin += pgsin_b; 5341 if (rbehind != NULL) 5342 *rbehind = pgsin_b; 5343 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la); 5344 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size) 5345 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size, 5346 PAGE_SIZE) - la); 5347 pgsin += pgsin_a; 5348 if (rahead != NULL) 5349 *rahead = pgsin_a; 5350 VM_CNT_INC(v_vnodein); 5351 VM_CNT_ADD(v_vnodepgsin, pgsin); 5352 5353 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) 5354 != 0) ? GB_UNMAPPED : 0; 5355 again: 5356 for (i = 0; i < count; i++) { 5357 if (ma[i] != bogus_page) 5358 vm_page_busy_downgrade(ma[i]); 5359 } 5360 5361 lbnp = -1; 5362 for (i = 0; i < count; i++) { 5363 m = ma[i]; 5364 if (m == bogus_page) 5365 continue; 5366 5367 /* 5368 * Pages are shared busy and the object lock is not 5369 * owned, which together allow for the pages' 5370 * invalidation. The racy test for validity avoids 5371 * useless creation of the buffer for the most typical 5372 * case when invalidation is not used in redo or for 5373 * parallel read. The shared->excl upgrade loop at 5374 * the end of the function catches the race in a 5375 * reliable way (protected by the object lock). 5376 */ 5377 if (vm_page_all_valid(m)) 5378 continue; 5379 5380 poff = IDX_TO_OFF(m->pindex); 5381 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size); 5382 for (; poff < poffe; poff += bsize) { 5383 lbn = get_lblkno(vp, poff); 5384 if (lbn == lbnp) 5385 goto next_page; 5386 lbnp = lbn; 5387 5388 error = get_blksize(vp, lbn, &bsize); 5389 if (error == 0) 5390 error = bread_gb(vp, lbn, bsize, 5391 curthread->td_ucred, br_flags, &bp); 5392 if (error != 0) 5393 goto end_pages; 5394 if (bp->b_rcred == curthread->td_ucred) { 5395 crfree(bp->b_rcred); 5396 bp->b_rcred = NOCRED; 5397 } 5398 if (LIST_EMPTY(&bp->b_dep)) { 5399 /* 5400 * Invalidation clears m->valid, but 5401 * may leave B_CACHE flag if the 5402 * buffer existed at the invalidation 5403 * time. In this case, recycle the 5404 * buffer to do real read on next 5405 * bread() after redo. 5406 * 5407 * Otherwise B_RELBUF is not strictly 5408 * necessary, enable to reduce buf 5409 * cache pressure. 5410 */ 5411 if (buf_pager_relbuf || 5412 !vm_page_all_valid(m)) 5413 bp->b_flags |= B_RELBUF; 5414 5415 bp->b_flags &= ~B_NOCACHE; 5416 brelse(bp); 5417 } else { 5418 bqrelse(bp); 5419 } 5420 } 5421 KASSERT(1 /* racy, enable for debugging */ || 5422 vm_page_all_valid(m) || i == count - 1, 5423 ("buf %d %p invalid", i, m)); 5424 if (i == count - 1 && lpart) { 5425 if (!vm_page_none_valid(m) && 5426 !vm_page_all_valid(m)) 5427 vm_page_zero_invalid(m, TRUE); 5428 } 5429 next_page:; 5430 } 5431 end_pages: 5432 5433 redo = false; 5434 for (i = 0; i < count; i++) { 5435 if (ma[i] == bogus_page) 5436 continue; 5437 if (vm_page_busy_tryupgrade(ma[i]) == 0) { 5438 vm_page_sunbusy(ma[i]); 5439 ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex, 5440 VM_ALLOC_NORMAL); 5441 } 5442 5443 /* 5444 * Since the pages were only sbusy while neither the 5445 * buffer nor the object lock was held by us, or 5446 * reallocated while vm_page_grab() slept for busy 5447 * relinguish, they could have been invalidated. 5448 * Recheck the valid bits and re-read as needed. 5449 * 5450 * Note that the last page is made fully valid in the 5451 * read loop, and partial validity for the page at 5452 * index count - 1 could mean that the page was 5453 * invalidated or removed, so we must restart for 5454 * safety as well. 5455 */ 5456 if (!vm_page_all_valid(ma[i])) 5457 redo = true; 5458 } 5459 if (redo && error == 0) 5460 goto again; 5461 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); 5462 } 5463 5464 #include "opt_ddb.h" 5465 #ifdef DDB 5466 #include <ddb/ddb.h> 5467 5468 /* DDB command to show buffer data */ 5469 DB_SHOW_COMMAND(buffer, db_show_buffer) 5470 { 5471 /* get args */ 5472 struct buf *bp = (struct buf *)addr; 5473 #ifdef FULL_BUF_TRACKING 5474 uint32_t i, j; 5475 #endif 5476 5477 if (!have_addr) { 5478 db_printf("usage: show buffer <addr>\n"); 5479 return; 5480 } 5481 5482 db_printf("buf at %p\n", bp); 5483 db_printf("b_flags = 0x%b, b_xflags = 0x%b\n", 5484 (u_int)bp->b_flags, PRINT_BUF_FLAGS, 5485 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS); 5486 db_printf("b_vflags = 0x%b, b_ioflags = 0x%b\n", 5487 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS, 5488 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS); 5489 db_printf( 5490 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 5491 "b_bufobj = %p, b_data = %p\n" 5492 "b_blkno = %jd, b_lblkno = %jd, b_vp = %p, b_dep = %p\n", 5493 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 5494 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 5495 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first); 5496 db_printf("b_kvabase = %p, b_kvasize = %d\n", 5497 bp->b_kvabase, bp->b_kvasize); 5498 if (bp->b_npages) { 5499 int i; 5500 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 5501 for (i = 0; i < bp->b_npages; i++) { 5502 vm_page_t m; 5503 m = bp->b_pages[i]; 5504 if (m != NULL) 5505 db_printf("(%p, 0x%lx, 0x%lx)", m->object, 5506 (u_long)m->pindex, 5507 (u_long)VM_PAGE_TO_PHYS(m)); 5508 else 5509 db_printf("( ??? )"); 5510 if ((i + 1) < bp->b_npages) 5511 db_printf(","); 5512 } 5513 db_printf("\n"); 5514 } 5515 BUF_LOCKPRINTINFO(bp); 5516 #if defined(FULL_BUF_TRACKING) 5517 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt); 5518 5519 i = bp->b_io_tcnt % BUF_TRACKING_SIZE; 5520 for (j = 1; j <= BUF_TRACKING_SIZE; j++) { 5521 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL) 5522 continue; 5523 db_printf(" %2u: %s\n", j, 5524 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]); 5525 } 5526 #elif defined(BUF_TRACKING) 5527 db_printf("b_io_tracking: %s\n", bp->b_io_tracking); 5528 #endif 5529 } 5530 5531 DB_SHOW_COMMAND_FLAGS(bufqueues, bufqueues, DB_CMD_MEMSAFE) 5532 { 5533 struct bufdomain *bd; 5534 struct buf *bp; 5535 long total; 5536 int i, j, cnt; 5537 5538 db_printf("bqempty: %d\n", bqempty.bq_len); 5539 5540 for (i = 0; i < buf_domains; i++) { 5541 bd = &bdomain[i]; 5542 db_printf("Buf domain %d\n", i); 5543 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers); 5544 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers); 5545 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers); 5546 db_printf("\n"); 5547 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace); 5548 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace); 5549 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace); 5550 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace); 5551 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh); 5552 db_printf("\n"); 5553 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers); 5554 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers); 5555 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers); 5556 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh); 5557 db_printf("\n"); 5558 total = 0; 5559 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist) 5560 total += bp->b_bufsize; 5561 db_printf("\tcleanq count\t%d (%ld)\n", 5562 bd->bd_cleanq->bq_len, total); 5563 total = 0; 5564 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist) 5565 total += bp->b_bufsize; 5566 db_printf("\tdirtyq count\t%d (%ld)\n", 5567 bd->bd_dirtyq.bq_len, total); 5568 db_printf("\twakeup\t\t%d\n", bd->bd_wanted); 5569 db_printf("\tlim\t\t%d\n", bd->bd_lim); 5570 db_printf("\tCPU "); 5571 for (j = 0; j <= mp_maxid; j++) 5572 db_printf("%d, ", bd->bd_subq[j].bq_len); 5573 db_printf("\n"); 5574 cnt = 0; 5575 total = 0; 5576 for (j = 0; j < nbuf; j++) { 5577 bp = nbufp(j); 5578 if (bp->b_domain == i && BUF_ISLOCKED(bp)) { 5579 cnt++; 5580 total += bp->b_bufsize; 5581 } 5582 } 5583 db_printf("\tLocked buffers: %d space %ld\n", cnt, total); 5584 cnt = 0; 5585 total = 0; 5586 for (j = 0; j < nbuf; j++) { 5587 bp = nbufp(j); 5588 if (bp->b_domain == i) { 5589 cnt++; 5590 total += bp->b_bufsize; 5591 } 5592 } 5593 db_printf("\tTotal buffers: %d space %ld\n", cnt, total); 5594 } 5595 } 5596 5597 DB_SHOW_COMMAND_FLAGS(lockedbufs, lockedbufs, DB_CMD_MEMSAFE) 5598 { 5599 struct buf *bp; 5600 int i; 5601 5602 for (i = 0; i < nbuf; i++) { 5603 bp = nbufp(i); 5604 if (BUF_ISLOCKED(bp)) { 5605 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5606 db_printf("\n"); 5607 if (db_pager_quit) 5608 break; 5609 } 5610 } 5611 } 5612 5613 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 5614 { 5615 struct vnode *vp; 5616 struct buf *bp; 5617 5618 if (!have_addr) { 5619 db_printf("usage: show vnodebufs <addr>\n"); 5620 return; 5621 } 5622 vp = (struct vnode *)addr; 5623 db_printf("Clean buffers:\n"); 5624 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 5625 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5626 db_printf("\n"); 5627 } 5628 db_printf("Dirty buffers:\n"); 5629 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 5630 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5631 db_printf("\n"); 5632 } 5633 } 5634 5635 DB_COMMAND_FLAGS(countfreebufs, db_coundfreebufs, DB_CMD_MEMSAFE) 5636 { 5637 struct buf *bp; 5638 int i, used = 0, nfree = 0; 5639 5640 if (have_addr) { 5641 db_printf("usage: countfreebufs\n"); 5642 return; 5643 } 5644 5645 for (i = 0; i < nbuf; i++) { 5646 bp = nbufp(i); 5647 if (bp->b_qindex == QUEUE_EMPTY) 5648 nfree++; 5649 else 5650 used++; 5651 } 5652 5653 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 5654 nfree + used); 5655 db_printf("numfreebuffers is %d\n", numfreebuffers); 5656 } 5657 #endif /* DDB */ 5658