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