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