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 /* 784 * bufspace_daemon: 785 * 786 * buffer space management daemon. Tries to maintain some marginal 787 * amount of free buffer space so that requesting processes neither 788 * block nor work to reclaim buffers. 789 */ 790 static void 791 bufspace_daemon(void *arg) 792 { 793 struct bufdomain *bd; 794 795 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread, 796 SHUTDOWN_PRI_LAST + 100); 797 798 bd = arg; 799 for (;;) { 800 kthread_suspend_check(); 801 802 /* 803 * Free buffers from the clean queue until we meet our 804 * targets. 805 * 806 * Theory of operation: The buffer cache is most efficient 807 * when some free buffer headers and space are always 808 * available to getnewbuf(). This daemon attempts to prevent 809 * the excessive blocking and synchronization associated 810 * with shortfall. It goes through three phases according 811 * demand: 812 * 813 * 1) The daemon wakes up voluntarily once per-second 814 * during idle periods when the counters are below 815 * the wakeup thresholds (bufspacethresh, lofreebuffers). 816 * 817 * 2) The daemon wakes up as we cross the thresholds 818 * ahead of any potential blocking. This may bounce 819 * slightly according to the rate of consumption and 820 * release. 821 * 822 * 3) The daemon and consumers are starved for working 823 * clean buffers. This is the 'bufspace' sleep below 824 * which will inefficiently trade bufs with bqrelse 825 * until we return to condition 2. 826 */ 827 while (bd->bd_bufspace > bd->bd_lobufspace || 828 bd->bd_freebuffers < bd->bd_hifreebuffers) { 829 if (buf_recycle(bd, false) != 0) { 830 if (bd_flushall(bd)) 831 continue; 832 /* 833 * Speedup dirty if we've run out of clean 834 * buffers. This is possible in particular 835 * because softdep may held many bufs locked 836 * pending writes to other bufs which are 837 * marked for delayed write, exhausting 838 * clean space until they are written. 839 */ 840 bd_speedup(); 841 BD_LOCK(bd); 842 if (bd->bd_wanted) { 843 msleep(&bd->bd_wanted, BD_LOCKPTR(bd), 844 PRIBIO|PDROP, "bufspace", hz/10); 845 } else 846 BD_UNLOCK(bd); 847 } 848 maybe_yield(); 849 } 850 bufspace_daemon_wait(bd); 851 } 852 } 853 854 /* 855 * bufmallocadjust: 856 * 857 * Adjust the reported bufspace for a malloc managed buffer, possibly 858 * waking any waiters. 859 */ 860 static void 861 bufmallocadjust(struct buf *bp, int bufsize) 862 { 863 int diff; 864 865 KASSERT((bp->b_flags & B_MALLOC) != 0, 866 ("bufmallocadjust: non-malloc buf %p", bp)); 867 diff = bufsize - bp->b_bufsize; 868 if (diff < 0) 869 atomic_subtract_long(&bufmallocspace, -diff); 870 else 871 atomic_add_long(&bufmallocspace, diff); 872 bp->b_bufsize = bufsize; 873 } 874 875 /* 876 * runningwakeup: 877 * 878 * Wake up processes that are waiting on asynchronous writes to fall 879 * below lorunningspace. 880 */ 881 static void 882 runningwakeup(void) 883 { 884 885 mtx_lock(&rbreqlock); 886 if (runningbufreq) { 887 runningbufreq = 0; 888 wakeup(&runningbufreq); 889 } 890 mtx_unlock(&rbreqlock); 891 } 892 893 /* 894 * runningbufwakeup: 895 * 896 * Decrement the outstanding write count according. 897 */ 898 void 899 runningbufwakeup(struct buf *bp) 900 { 901 long space, bspace; 902 903 bspace = bp->b_runningbufspace; 904 if (bspace == 0) 905 return; 906 space = atomic_fetchadd_long(&runningbufspace, -bspace); 907 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", 908 space, bspace)); 909 bp->b_runningbufspace = 0; 910 /* 911 * Only acquire the lock and wakeup on the transition from exceeding 912 * the threshold to falling below it. 913 */ 914 if (space < lorunningspace) 915 return; 916 if (space - bspace > lorunningspace) 917 return; 918 runningwakeup(); 919 } 920 921 /* 922 * waitrunningbufspace() 923 * 924 * runningbufspace is a measure of the amount of I/O currently 925 * running. This routine is used in async-write situations to 926 * prevent creating huge backups of pending writes to a device. 927 * Only asynchronous writes are governed by this function. 928 * 929 * This does NOT turn an async write into a sync write. It waits 930 * for earlier writes to complete and generally returns before the 931 * caller's write has reached the device. 932 */ 933 void 934 waitrunningbufspace(void) 935 { 936 937 mtx_lock(&rbreqlock); 938 while (runningbufspace > hirunningspace) { 939 runningbufreq = 1; 940 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 941 } 942 mtx_unlock(&rbreqlock); 943 } 944 945 946 /* 947 * vfs_buf_test_cache: 948 * 949 * Called when a buffer is extended. This function clears the B_CACHE 950 * bit if the newly extended portion of the buffer does not contain 951 * valid data. 952 */ 953 static __inline void 954 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off, 955 vm_offset_t size, vm_page_t m) 956 { 957 958 /* 959 * This function and its results are protected by higher level 960 * synchronization requiring vnode and buf locks to page in and 961 * validate pages. 962 */ 963 if (bp->b_flags & B_CACHE) { 964 int base = (foff + off) & PAGE_MASK; 965 if (vm_page_is_valid(m, base, size) == 0) 966 bp->b_flags &= ~B_CACHE; 967 } 968 } 969 970 /* Wake up the buffer daemon if necessary */ 971 static void 972 bd_wakeup(void) 973 { 974 975 mtx_lock(&bdlock); 976 if (bd_request == 0) { 977 bd_request = 1; 978 wakeup(&bd_request); 979 } 980 mtx_unlock(&bdlock); 981 } 982 983 /* 984 * Adjust the maxbcachbuf tunable. 985 */ 986 static void 987 maxbcachebuf_adjust(void) 988 { 989 int i; 990 991 /* 992 * maxbcachebuf must be a power of 2 >= MAXBSIZE. 993 */ 994 i = 2; 995 while (i * 2 <= maxbcachebuf) 996 i *= 2; 997 maxbcachebuf = i; 998 if (maxbcachebuf < MAXBSIZE) 999 maxbcachebuf = MAXBSIZE; 1000 if (maxbcachebuf > MAXPHYS) 1001 maxbcachebuf = MAXPHYS; 1002 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF) 1003 printf("maxbcachebuf=%d\n", maxbcachebuf); 1004 } 1005 1006 /* 1007 * bd_speedup - speedup the buffer cache flushing code 1008 */ 1009 void 1010 bd_speedup(void) 1011 { 1012 int needwake; 1013 1014 mtx_lock(&bdlock); 1015 needwake = 0; 1016 if (bd_speedupreq == 0 || bd_request == 0) 1017 needwake = 1; 1018 bd_speedupreq = 1; 1019 bd_request = 1; 1020 if (needwake) 1021 wakeup(&bd_request); 1022 mtx_unlock(&bdlock); 1023 } 1024 1025 #ifdef __i386__ 1026 #define TRANSIENT_DENOM 5 1027 #else 1028 #define TRANSIENT_DENOM 10 1029 #endif 1030 1031 /* 1032 * Calculating buffer cache scaling values and reserve space for buffer 1033 * headers. This is called during low level kernel initialization and 1034 * may be called more then once. We CANNOT write to the memory area 1035 * being reserved at this time. 1036 */ 1037 caddr_t 1038 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 1039 { 1040 int tuned_nbuf; 1041 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 1042 1043 /* 1044 * physmem_est is in pages. Convert it to kilobytes (assumes 1045 * PAGE_SIZE is >= 1K) 1046 */ 1047 physmem_est = physmem_est * (PAGE_SIZE / 1024); 1048 1049 maxbcachebuf_adjust(); 1050 /* 1051 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 1052 * For the first 64MB of ram nominally allocate sufficient buffers to 1053 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 1054 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 1055 * the buffer cache we limit the eventual kva reservation to 1056 * maxbcache bytes. 1057 * 1058 * factor represents the 1/4 x ram conversion. 1059 */ 1060 if (nbuf == 0) { 1061 int factor = 4 * BKVASIZE / 1024; 1062 1063 nbuf = 50; 1064 if (physmem_est > 4096) 1065 nbuf += min((physmem_est - 4096) / factor, 1066 65536 / factor); 1067 if (physmem_est > 65536) 1068 nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 1069 32 * 1024 * 1024 / (factor * 5)); 1070 1071 if (maxbcache && nbuf > maxbcache / BKVASIZE) 1072 nbuf = maxbcache / BKVASIZE; 1073 tuned_nbuf = 1; 1074 } else 1075 tuned_nbuf = 0; 1076 1077 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 1078 maxbuf = (LONG_MAX / 3) / BKVASIZE; 1079 if (nbuf > maxbuf) { 1080 if (!tuned_nbuf) 1081 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 1082 maxbuf); 1083 nbuf = maxbuf; 1084 } 1085 1086 /* 1087 * Ideal allocation size for the transient bio submap is 10% 1088 * of the maximal space buffer map. This roughly corresponds 1089 * to the amount of the buffer mapped for typical UFS load. 1090 * 1091 * Clip the buffer map to reserve space for the transient 1092 * BIOs, if its extent is bigger than 90% (80% on i386) of the 1093 * maximum buffer map extent on the platform. 1094 * 1095 * The fall-back to the maxbuf in case of maxbcache unset, 1096 * allows to not trim the buffer KVA for the architectures 1097 * with ample KVA space. 1098 */ 1099 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 1100 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 1101 buf_sz = (long)nbuf * BKVASIZE; 1102 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 1103 (TRANSIENT_DENOM - 1)) { 1104 /* 1105 * There is more KVA than memory. Do not 1106 * adjust buffer map size, and assign the rest 1107 * of maxbuf to transient map. 1108 */ 1109 biotmap_sz = maxbuf_sz - buf_sz; 1110 } else { 1111 /* 1112 * Buffer map spans all KVA we could afford on 1113 * this platform. Give 10% (20% on i386) of 1114 * the buffer map to the transient bio map. 1115 */ 1116 biotmap_sz = buf_sz / TRANSIENT_DENOM; 1117 buf_sz -= biotmap_sz; 1118 } 1119 if (biotmap_sz / INT_MAX > MAXPHYS) 1120 bio_transient_maxcnt = INT_MAX; 1121 else 1122 bio_transient_maxcnt = biotmap_sz / MAXPHYS; 1123 /* 1124 * Artificially limit to 1024 simultaneous in-flight I/Os 1125 * using the transient mapping. 1126 */ 1127 if (bio_transient_maxcnt > 1024) 1128 bio_transient_maxcnt = 1024; 1129 if (tuned_nbuf) 1130 nbuf = buf_sz / BKVASIZE; 1131 } 1132 1133 if (nswbuf == 0) { 1134 nswbuf = min(nbuf / 4, 256); 1135 if (nswbuf < NSWBUF_MIN) 1136 nswbuf = NSWBUF_MIN; 1137 } 1138 1139 /* 1140 * Reserve space for the buffer cache buffers 1141 */ 1142 buf = (void *)v; 1143 v = (caddr_t)(buf + nbuf); 1144 1145 return(v); 1146 } 1147 1148 /* Initialize the buffer subsystem. Called before use of any buffers. */ 1149 void 1150 bufinit(void) 1151 { 1152 struct buf *bp; 1153 int i; 1154 1155 KASSERT(maxbcachebuf >= MAXBSIZE, 1156 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf, 1157 MAXBSIZE)); 1158 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock"); 1159 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 1160 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 1161 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); 1162 1163 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS); 1164 1165 /* finally, initialize each buffer header and stick on empty q */ 1166 for (i = 0; i < nbuf; i++) { 1167 bp = &buf[i]; 1168 bzero(bp, sizeof *bp); 1169 bp->b_flags = B_INVAL; 1170 bp->b_rcred = NOCRED; 1171 bp->b_wcred = NOCRED; 1172 bp->b_qindex = QUEUE_NONE; 1173 bp->b_domain = -1; 1174 bp->b_subqueue = mp_maxid + 1; 1175 bp->b_xflags = 0; 1176 bp->b_data = bp->b_kvabase = unmapped_buf; 1177 LIST_INIT(&bp->b_dep); 1178 BUF_LOCKINIT(bp); 1179 bq_insert(&bqempty, bp, false); 1180 } 1181 1182 /* 1183 * maxbufspace is the absolute maximum amount of buffer space we are 1184 * allowed to reserve in KVM and in real terms. The absolute maximum 1185 * is nominally used by metadata. hibufspace is the nominal maximum 1186 * used by most other requests. The differential is required to 1187 * ensure that metadata deadlocks don't occur. 1188 * 1189 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 1190 * this may result in KVM fragmentation which is not handled optimally 1191 * by the system. XXX This is less true with vmem. We could use 1192 * PAGE_SIZE. 1193 */ 1194 maxbufspace = (long)nbuf * BKVASIZE; 1195 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10); 1196 lobufspace = (hibufspace / 20) * 19; /* 95% */ 1197 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2; 1198 1199 /* 1200 * Note: The 16 MiB upper limit for hirunningspace was chosen 1201 * arbitrarily and may need further tuning. It corresponds to 1202 * 128 outstanding write IO requests (if IO size is 128 KiB), 1203 * which fits with many RAID controllers' tagged queuing limits. 1204 * The lower 1 MiB limit is the historical upper limit for 1205 * hirunningspace. 1206 */ 1207 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf), 1208 16 * 1024 * 1024), 1024 * 1024); 1209 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf); 1210 1211 /* 1212 * Limit the amount of malloc memory since it is wired permanently into 1213 * the kernel space. Even though this is accounted for in the buffer 1214 * allocation, we don't want the malloced region to grow uncontrolled. 1215 * The malloc scheme improves memory utilization significantly on 1216 * average (small) directories. 1217 */ 1218 maxbufmallocspace = hibufspace / 20; 1219 1220 /* 1221 * Reduce the chance of a deadlock occurring by limiting the number 1222 * of delayed-write dirty buffers we allow to stack up. 1223 */ 1224 hidirtybuffers = nbuf / 4 + 20; 1225 dirtybufthresh = hidirtybuffers * 9 / 10; 1226 /* 1227 * To support extreme low-memory systems, make sure hidirtybuffers 1228 * cannot eat up all available buffer space. This occurs when our 1229 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our 1230 * buffer space assuming BKVASIZE'd buffers. 1231 */ 1232 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 1233 hidirtybuffers >>= 1; 1234 } 1235 lodirtybuffers = hidirtybuffers / 2; 1236 1237 /* 1238 * lofreebuffers should be sufficient to avoid stalling waiting on 1239 * buf headers under heavy utilization. The bufs in per-cpu caches 1240 * are counted as free but will be unavailable to threads executing 1241 * on other cpus. 1242 * 1243 * hifreebuffers is the free target for the bufspace daemon. This 1244 * should be set appropriately to limit work per-iteration. 1245 */ 1246 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus); 1247 hifreebuffers = (3 * lofreebuffers) / 2; 1248 numfreebuffers = nbuf; 1249 1250 /* Setup the kva and free list allocators. */ 1251 vmem_set_reclaim(buffer_arena, bufkva_reclaim); 1252 buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf), 1253 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0); 1254 1255 /* 1256 * Size the clean queue according to the amount of buffer space. 1257 * One queue per-256mb up to the max. More queues gives better 1258 * concurrency but less accurate LRU. 1259 */ 1260 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS); 1261 for (i = 0 ; i < buf_domains; i++) { 1262 struct bufdomain *bd; 1263 1264 bd = &bdomain[i]; 1265 bd_init(bd); 1266 bd->bd_freebuffers = nbuf / buf_domains; 1267 bd->bd_hifreebuffers = hifreebuffers / buf_domains; 1268 bd->bd_lofreebuffers = lofreebuffers / buf_domains; 1269 bd->bd_bufspace = 0; 1270 bd->bd_maxbufspace = maxbufspace / buf_domains; 1271 bd->bd_hibufspace = hibufspace / buf_domains; 1272 bd->bd_lobufspace = lobufspace / buf_domains; 1273 bd->bd_bufspacethresh = bufspacethresh / buf_domains; 1274 bd->bd_numdirtybuffers = 0; 1275 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains; 1276 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains; 1277 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains; 1278 /* Don't allow more than 2% of bufs in the per-cpu caches. */ 1279 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus; 1280 } 1281 getnewbufcalls = counter_u64_alloc(M_WAITOK); 1282 getnewbufrestarts = counter_u64_alloc(M_WAITOK); 1283 mappingrestarts = counter_u64_alloc(M_WAITOK); 1284 numbufallocfails = counter_u64_alloc(M_WAITOK); 1285 notbufdflushes = counter_u64_alloc(M_WAITOK); 1286 buffreekvacnt = counter_u64_alloc(M_WAITOK); 1287 bufdefragcnt = counter_u64_alloc(M_WAITOK); 1288 bufkvaspace = counter_u64_alloc(M_WAITOK); 1289 } 1290 1291 #ifdef INVARIANTS 1292 static inline void 1293 vfs_buf_check_mapped(struct buf *bp) 1294 { 1295 1296 KASSERT(bp->b_kvabase != unmapped_buf, 1297 ("mapped buf: b_kvabase was not updated %p", bp)); 1298 KASSERT(bp->b_data != unmapped_buf, 1299 ("mapped buf: b_data was not updated %p", bp)); 1300 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf + 1301 MAXPHYS, ("b_data + b_offset unmapped %p", bp)); 1302 } 1303 1304 static inline void 1305 vfs_buf_check_unmapped(struct buf *bp) 1306 { 1307 1308 KASSERT(bp->b_data == unmapped_buf, 1309 ("unmapped buf: corrupted b_data %p", bp)); 1310 } 1311 1312 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 1313 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 1314 #else 1315 #define BUF_CHECK_MAPPED(bp) do {} while (0) 1316 #define BUF_CHECK_UNMAPPED(bp) do {} while (0) 1317 #endif 1318 1319 static int 1320 isbufbusy(struct buf *bp) 1321 { 1322 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) || 1323 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI)) 1324 return (1); 1325 return (0); 1326 } 1327 1328 /* 1329 * Shutdown the system cleanly to prepare for reboot, halt, or power off. 1330 */ 1331 void 1332 bufshutdown(int show_busybufs) 1333 { 1334 static int first_buf_printf = 1; 1335 struct buf *bp; 1336 int iter, nbusy, pbusy; 1337 #ifndef PREEMPTION 1338 int subiter; 1339 #endif 1340 1341 /* 1342 * Sync filesystems for shutdown 1343 */ 1344 wdog_kern_pat(WD_LASTVAL); 1345 kern_sync(curthread); 1346 1347 /* 1348 * With soft updates, some buffers that are 1349 * written will be remarked as dirty until other 1350 * buffers are written. 1351 */ 1352 for (iter = pbusy = 0; iter < 20; iter++) { 1353 nbusy = 0; 1354 for (bp = &buf[nbuf]; --bp >= buf; ) 1355 if (isbufbusy(bp)) 1356 nbusy++; 1357 if (nbusy == 0) { 1358 if (first_buf_printf) 1359 printf("All buffers synced."); 1360 break; 1361 } 1362 if (first_buf_printf) { 1363 printf("Syncing disks, buffers remaining... "); 1364 first_buf_printf = 0; 1365 } 1366 printf("%d ", nbusy); 1367 if (nbusy < pbusy) 1368 iter = 0; 1369 pbusy = nbusy; 1370 1371 wdog_kern_pat(WD_LASTVAL); 1372 kern_sync(curthread); 1373 1374 #ifdef PREEMPTION 1375 /* 1376 * Spin for a while to allow interrupt threads to run. 1377 */ 1378 DELAY(50000 * iter); 1379 #else 1380 /* 1381 * Context switch several times to allow interrupt 1382 * threads to run. 1383 */ 1384 for (subiter = 0; subiter < 50 * iter; subiter++) { 1385 thread_lock(curthread); 1386 mi_switch(SW_VOL); 1387 DELAY(1000); 1388 } 1389 #endif 1390 } 1391 printf("\n"); 1392 /* 1393 * Count only busy local buffers to prevent forcing 1394 * a fsck if we're just a client of a wedged NFS server 1395 */ 1396 nbusy = 0; 1397 for (bp = &buf[nbuf]; --bp >= buf; ) { 1398 if (isbufbusy(bp)) { 1399 #if 0 1400 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */ 1401 if (bp->b_dev == NULL) { 1402 TAILQ_REMOVE(&mountlist, 1403 bp->b_vp->v_mount, mnt_list); 1404 continue; 1405 } 1406 #endif 1407 nbusy++; 1408 if (show_busybufs > 0) { 1409 printf( 1410 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:", 1411 nbusy, bp, bp->b_vp, bp->b_flags, 1412 (intmax_t)bp->b_blkno, 1413 (intmax_t)bp->b_lblkno); 1414 BUF_LOCKPRINTINFO(bp); 1415 if (show_busybufs > 1) 1416 vn_printf(bp->b_vp, 1417 "vnode content: "); 1418 } 1419 } 1420 } 1421 if (nbusy) { 1422 /* 1423 * Failed to sync all blocks. Indicate this and don't 1424 * unmount filesystems (thus forcing an fsck on reboot). 1425 */ 1426 printf("Giving up on %d buffers\n", nbusy); 1427 DELAY(5000000); /* 5 seconds */ 1428 } else { 1429 if (!first_buf_printf) 1430 printf("Final sync complete\n"); 1431 /* 1432 * Unmount filesystems 1433 */ 1434 if (!KERNEL_PANICKED()) 1435 vfs_unmountall(); 1436 } 1437 swapoff_all(); 1438 DELAY(100000); /* wait for console output to finish */ 1439 } 1440 1441 static void 1442 bpmap_qenter(struct buf *bp) 1443 { 1444 1445 BUF_CHECK_MAPPED(bp); 1446 1447 /* 1448 * bp->b_data is relative to bp->b_offset, but 1449 * bp->b_offset may be offset into the first page. 1450 */ 1451 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 1452 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 1453 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 1454 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 1455 } 1456 1457 static inline struct bufdomain * 1458 bufdomain(struct buf *bp) 1459 { 1460 1461 return (&bdomain[bp->b_domain]); 1462 } 1463 1464 static struct bufqueue * 1465 bufqueue(struct buf *bp) 1466 { 1467 1468 switch (bp->b_qindex) { 1469 case QUEUE_NONE: 1470 /* FALLTHROUGH */ 1471 case QUEUE_SENTINEL: 1472 return (NULL); 1473 case QUEUE_EMPTY: 1474 return (&bqempty); 1475 case QUEUE_DIRTY: 1476 return (&bufdomain(bp)->bd_dirtyq); 1477 case QUEUE_CLEAN: 1478 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]); 1479 default: 1480 break; 1481 } 1482 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex); 1483 } 1484 1485 /* 1486 * Return the locked bufqueue that bp is a member of. 1487 */ 1488 static struct bufqueue * 1489 bufqueue_acquire(struct buf *bp) 1490 { 1491 struct bufqueue *bq, *nbq; 1492 1493 /* 1494 * bp can be pushed from a per-cpu queue to the 1495 * cleanq while we're waiting on the lock. Retry 1496 * if the queues don't match. 1497 */ 1498 bq = bufqueue(bp); 1499 BQ_LOCK(bq); 1500 for (;;) { 1501 nbq = bufqueue(bp); 1502 if (bq == nbq) 1503 break; 1504 BQ_UNLOCK(bq); 1505 BQ_LOCK(nbq); 1506 bq = nbq; 1507 } 1508 return (bq); 1509 } 1510 1511 /* 1512 * binsfree: 1513 * 1514 * Insert the buffer into the appropriate free list. Requires a 1515 * locked buffer on entry and buffer is unlocked before return. 1516 */ 1517 static void 1518 binsfree(struct buf *bp, int qindex) 1519 { 1520 struct bufdomain *bd; 1521 struct bufqueue *bq; 1522 1523 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY, 1524 ("binsfree: Invalid qindex %d", qindex)); 1525 BUF_ASSERT_XLOCKED(bp); 1526 1527 /* 1528 * Handle delayed bremfree() processing. 1529 */ 1530 if (bp->b_flags & B_REMFREE) { 1531 if (bp->b_qindex == qindex) { 1532 bp->b_flags |= B_REUSE; 1533 bp->b_flags &= ~B_REMFREE; 1534 BUF_UNLOCK(bp); 1535 return; 1536 } 1537 bq = bufqueue_acquire(bp); 1538 bq_remove(bq, bp); 1539 BQ_UNLOCK(bq); 1540 } 1541 bd = bufdomain(bp); 1542 if (qindex == QUEUE_CLEAN) { 1543 if (bd->bd_lim != 0) 1544 bq = &bd->bd_subq[PCPU_GET(cpuid)]; 1545 else 1546 bq = bd->bd_cleanq; 1547 } else 1548 bq = &bd->bd_dirtyq; 1549 bq_insert(bq, bp, true); 1550 } 1551 1552 /* 1553 * buf_free: 1554 * 1555 * Free a buffer to the buf zone once it no longer has valid contents. 1556 */ 1557 static void 1558 buf_free(struct buf *bp) 1559 { 1560 1561 if (bp->b_flags & B_REMFREE) 1562 bremfreef(bp); 1563 if (bp->b_vflags & BV_BKGRDINPROG) 1564 panic("losing buffer 1"); 1565 if (bp->b_rcred != NOCRED) { 1566 crfree(bp->b_rcred); 1567 bp->b_rcred = NOCRED; 1568 } 1569 if (bp->b_wcred != NOCRED) { 1570 crfree(bp->b_wcred); 1571 bp->b_wcred = NOCRED; 1572 } 1573 if (!LIST_EMPTY(&bp->b_dep)) 1574 buf_deallocate(bp); 1575 bufkva_free(bp); 1576 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1); 1577 BUF_UNLOCK(bp); 1578 uma_zfree(buf_zone, bp); 1579 } 1580 1581 /* 1582 * buf_import: 1583 * 1584 * Import bufs into the uma cache from the buf list. The system still 1585 * expects a static array of bufs and much of the synchronization 1586 * around bufs assumes type stable storage. As a result, UMA is used 1587 * only as a per-cpu cache of bufs still maintained on a global list. 1588 */ 1589 static int 1590 buf_import(void *arg, void **store, int cnt, int domain, int flags) 1591 { 1592 struct buf *bp; 1593 int i; 1594 1595 BQ_LOCK(&bqempty); 1596 for (i = 0; i < cnt; i++) { 1597 bp = TAILQ_FIRST(&bqempty.bq_queue); 1598 if (bp == NULL) 1599 break; 1600 bq_remove(&bqempty, bp); 1601 store[i] = bp; 1602 } 1603 BQ_UNLOCK(&bqempty); 1604 1605 return (i); 1606 } 1607 1608 /* 1609 * buf_release: 1610 * 1611 * Release bufs from the uma cache back to the buffer queues. 1612 */ 1613 static void 1614 buf_release(void *arg, void **store, int cnt) 1615 { 1616 struct bufqueue *bq; 1617 struct buf *bp; 1618 int i; 1619 1620 bq = &bqempty; 1621 BQ_LOCK(bq); 1622 for (i = 0; i < cnt; i++) { 1623 bp = store[i]; 1624 /* Inline bq_insert() to batch locking. */ 1625 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1626 bp->b_flags &= ~(B_AGE | B_REUSE); 1627 bq->bq_len++; 1628 bp->b_qindex = bq->bq_index; 1629 } 1630 BQ_UNLOCK(bq); 1631 } 1632 1633 /* 1634 * buf_alloc: 1635 * 1636 * Allocate an empty buffer header. 1637 */ 1638 static struct buf * 1639 buf_alloc(struct bufdomain *bd) 1640 { 1641 struct buf *bp; 1642 int freebufs; 1643 1644 /* 1645 * We can only run out of bufs in the buf zone if the average buf 1646 * is less than BKVASIZE. In this case the actual wait/block will 1647 * come from buf_reycle() failing to flush one of these small bufs. 1648 */ 1649 bp = NULL; 1650 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1); 1651 if (freebufs > 0) 1652 bp = uma_zalloc(buf_zone, M_NOWAIT); 1653 if (bp == NULL) { 1654 atomic_add_int(&bd->bd_freebuffers, 1); 1655 bufspace_daemon_wakeup(bd); 1656 counter_u64_add(numbufallocfails, 1); 1657 return (NULL); 1658 } 1659 /* 1660 * Wake-up the bufspace daemon on transition below threshold. 1661 */ 1662 if (freebufs == bd->bd_lofreebuffers) 1663 bufspace_daemon_wakeup(bd); 1664 1665 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1666 panic("getnewbuf_empty: Locked buf %p on free queue.", bp); 1667 1668 KASSERT(bp->b_vp == NULL, 1669 ("bp: %p still has vnode %p.", bp, bp->b_vp)); 1670 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0, 1671 ("invalid buffer %p flags %#x", bp, bp->b_flags)); 1672 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1673 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 1674 KASSERT(bp->b_npages == 0, 1675 ("bp: %p still has %d vm pages\n", bp, bp->b_npages)); 1676 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp)); 1677 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp)); 1678 1679 bp->b_domain = BD_DOMAIN(bd); 1680 bp->b_flags = 0; 1681 bp->b_ioflags = 0; 1682 bp->b_xflags = 0; 1683 bp->b_vflags = 0; 1684 bp->b_vp = NULL; 1685 bp->b_blkno = bp->b_lblkno = 0; 1686 bp->b_offset = NOOFFSET; 1687 bp->b_iodone = 0; 1688 bp->b_error = 0; 1689 bp->b_resid = 0; 1690 bp->b_bcount = 0; 1691 bp->b_npages = 0; 1692 bp->b_dirtyoff = bp->b_dirtyend = 0; 1693 bp->b_bufobj = NULL; 1694 bp->b_data = bp->b_kvabase = unmapped_buf; 1695 bp->b_fsprivate1 = NULL; 1696 bp->b_fsprivate2 = NULL; 1697 bp->b_fsprivate3 = NULL; 1698 LIST_INIT(&bp->b_dep); 1699 1700 return (bp); 1701 } 1702 1703 /* 1704 * buf_recycle: 1705 * 1706 * Free a buffer from the given bufqueue. kva controls whether the 1707 * freed buf must own some kva resources. This is used for 1708 * defragmenting. 1709 */ 1710 static int 1711 buf_recycle(struct bufdomain *bd, bool kva) 1712 { 1713 struct bufqueue *bq; 1714 struct buf *bp, *nbp; 1715 1716 if (kva) 1717 counter_u64_add(bufdefragcnt, 1); 1718 nbp = NULL; 1719 bq = bd->bd_cleanq; 1720 BQ_LOCK(bq); 1721 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd), 1722 ("buf_recycle: Locks don't match")); 1723 nbp = TAILQ_FIRST(&bq->bq_queue); 1724 1725 /* 1726 * Run scan, possibly freeing data and/or kva mappings on the fly 1727 * depending. 1728 */ 1729 while ((bp = nbp) != NULL) { 1730 /* 1731 * Calculate next bp (we can only use it if we do not 1732 * release the bqlock). 1733 */ 1734 nbp = TAILQ_NEXT(bp, b_freelist); 1735 1736 /* 1737 * If we are defragging then we need a buffer with 1738 * some kva to reclaim. 1739 */ 1740 if (kva && bp->b_kvasize == 0) 1741 continue; 1742 1743 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1744 continue; 1745 1746 /* 1747 * Implement a second chance algorithm for frequently 1748 * accessed buffers. 1749 */ 1750 if ((bp->b_flags & B_REUSE) != 0) { 1751 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1752 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1753 bp->b_flags &= ~B_REUSE; 1754 BUF_UNLOCK(bp); 1755 continue; 1756 } 1757 1758 /* 1759 * Skip buffers with background writes in progress. 1760 */ 1761 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 1762 BUF_UNLOCK(bp); 1763 continue; 1764 } 1765 1766 KASSERT(bp->b_qindex == QUEUE_CLEAN, 1767 ("buf_recycle: inconsistent queue %d bp %p", 1768 bp->b_qindex, bp)); 1769 KASSERT(bp->b_domain == BD_DOMAIN(bd), 1770 ("getnewbuf: queue domain %d doesn't match request %d", 1771 bp->b_domain, (int)BD_DOMAIN(bd))); 1772 /* 1773 * NOTE: nbp is now entirely invalid. We can only restart 1774 * the scan from this point on. 1775 */ 1776 bq_remove(bq, bp); 1777 BQ_UNLOCK(bq); 1778 1779 /* 1780 * Requeue the background write buffer with error and 1781 * restart the scan. 1782 */ 1783 if ((bp->b_vflags & BV_BKGRDERR) != 0) { 1784 bqrelse(bp); 1785 BQ_LOCK(bq); 1786 nbp = TAILQ_FIRST(&bq->bq_queue); 1787 continue; 1788 } 1789 bp->b_flags |= B_INVAL; 1790 brelse(bp); 1791 return (0); 1792 } 1793 bd->bd_wanted = 1; 1794 BQ_UNLOCK(bq); 1795 1796 return (ENOBUFS); 1797 } 1798 1799 /* 1800 * bremfree: 1801 * 1802 * Mark the buffer for removal from the appropriate free list. 1803 * 1804 */ 1805 void 1806 bremfree(struct buf *bp) 1807 { 1808 1809 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1810 KASSERT((bp->b_flags & B_REMFREE) == 0, 1811 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1812 KASSERT(bp->b_qindex != QUEUE_NONE, 1813 ("bremfree: buffer %p not on a queue.", bp)); 1814 BUF_ASSERT_XLOCKED(bp); 1815 1816 bp->b_flags |= B_REMFREE; 1817 } 1818 1819 /* 1820 * bremfreef: 1821 * 1822 * Force an immediate removal from a free list. Used only in nfs when 1823 * it abuses the b_freelist pointer. 1824 */ 1825 void 1826 bremfreef(struct buf *bp) 1827 { 1828 struct bufqueue *bq; 1829 1830 bq = bufqueue_acquire(bp); 1831 bq_remove(bq, bp); 1832 BQ_UNLOCK(bq); 1833 } 1834 1835 static void 1836 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname) 1837 { 1838 1839 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF); 1840 TAILQ_INIT(&bq->bq_queue); 1841 bq->bq_len = 0; 1842 bq->bq_index = qindex; 1843 bq->bq_subqueue = subqueue; 1844 } 1845 1846 static void 1847 bd_init(struct bufdomain *bd) 1848 { 1849 int i; 1850 1851 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1]; 1852 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock"); 1853 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock"); 1854 for (i = 0; i <= mp_maxid; i++) 1855 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i, 1856 "bufq clean subqueue lock"); 1857 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF); 1858 } 1859 1860 /* 1861 * bq_remove: 1862 * 1863 * Removes a buffer from the free list, must be called with the 1864 * correct qlock held. 1865 */ 1866 static void 1867 bq_remove(struct bufqueue *bq, struct buf *bp) 1868 { 1869 1870 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X", 1871 bp, bp->b_vp, bp->b_flags); 1872 KASSERT(bp->b_qindex != QUEUE_NONE, 1873 ("bq_remove: buffer %p not on a queue.", bp)); 1874 KASSERT(bufqueue(bp) == bq, 1875 ("bq_remove: Remove buffer %p from wrong queue.", bp)); 1876 1877 BQ_ASSERT_LOCKED(bq); 1878 if (bp->b_qindex != QUEUE_EMPTY) { 1879 BUF_ASSERT_XLOCKED(bp); 1880 } 1881 KASSERT(bq->bq_len >= 1, 1882 ("queue %d underflow", bp->b_qindex)); 1883 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1884 bq->bq_len--; 1885 bp->b_qindex = QUEUE_NONE; 1886 bp->b_flags &= ~(B_REMFREE | B_REUSE); 1887 } 1888 1889 static void 1890 bd_flush(struct bufdomain *bd, struct bufqueue *bq) 1891 { 1892 struct buf *bp; 1893 1894 BQ_ASSERT_LOCKED(bq); 1895 if (bq != bd->bd_cleanq) { 1896 BD_LOCK(bd); 1897 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) { 1898 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1899 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp, 1900 b_freelist); 1901 bp->b_subqueue = bd->bd_cleanq->bq_subqueue; 1902 } 1903 bd->bd_cleanq->bq_len += bq->bq_len; 1904 bq->bq_len = 0; 1905 } 1906 if (bd->bd_wanted) { 1907 bd->bd_wanted = 0; 1908 wakeup(&bd->bd_wanted); 1909 } 1910 if (bq != bd->bd_cleanq) 1911 BD_UNLOCK(bd); 1912 } 1913 1914 static int 1915 bd_flushall(struct bufdomain *bd) 1916 { 1917 struct bufqueue *bq; 1918 int flushed; 1919 int i; 1920 1921 if (bd->bd_lim == 0) 1922 return (0); 1923 flushed = 0; 1924 for (i = 0; i <= mp_maxid; i++) { 1925 bq = &bd->bd_subq[i]; 1926 if (bq->bq_len == 0) 1927 continue; 1928 BQ_LOCK(bq); 1929 bd_flush(bd, bq); 1930 BQ_UNLOCK(bq); 1931 flushed++; 1932 } 1933 1934 return (flushed); 1935 } 1936 1937 static void 1938 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock) 1939 { 1940 struct bufdomain *bd; 1941 1942 if (bp->b_qindex != QUEUE_NONE) 1943 panic("bq_insert: free buffer %p onto another queue?", bp); 1944 1945 bd = bufdomain(bp); 1946 if (bp->b_flags & B_AGE) { 1947 /* Place this buf directly on the real queue. */ 1948 if (bq->bq_index == QUEUE_CLEAN) 1949 bq = bd->bd_cleanq; 1950 BQ_LOCK(bq); 1951 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist); 1952 } else { 1953 BQ_LOCK(bq); 1954 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1955 } 1956 bp->b_flags &= ~(B_AGE | B_REUSE); 1957 bq->bq_len++; 1958 bp->b_qindex = bq->bq_index; 1959 bp->b_subqueue = bq->bq_subqueue; 1960 1961 /* 1962 * Unlock before we notify so that we don't wakeup a waiter that 1963 * fails a trylock on the buf and sleeps again. 1964 */ 1965 if (unlock) 1966 BUF_UNLOCK(bp); 1967 1968 if (bp->b_qindex == QUEUE_CLEAN) { 1969 /* 1970 * Flush the per-cpu queue and notify any waiters. 1971 */ 1972 if (bd->bd_wanted || (bq != bd->bd_cleanq && 1973 bq->bq_len >= bd->bd_lim)) 1974 bd_flush(bd, bq); 1975 } 1976 BQ_UNLOCK(bq); 1977 } 1978 1979 /* 1980 * bufkva_free: 1981 * 1982 * Free the kva allocation for a buffer. 1983 * 1984 */ 1985 static void 1986 bufkva_free(struct buf *bp) 1987 { 1988 1989 #ifdef INVARIANTS 1990 if (bp->b_kvasize == 0) { 1991 KASSERT(bp->b_kvabase == unmapped_buf && 1992 bp->b_data == unmapped_buf, 1993 ("Leaked KVA space on %p", bp)); 1994 } else if (buf_mapped(bp)) 1995 BUF_CHECK_MAPPED(bp); 1996 else 1997 BUF_CHECK_UNMAPPED(bp); 1998 #endif 1999 if (bp->b_kvasize == 0) 2000 return; 2001 2002 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize); 2003 counter_u64_add(bufkvaspace, -bp->b_kvasize); 2004 counter_u64_add(buffreekvacnt, 1); 2005 bp->b_data = bp->b_kvabase = unmapped_buf; 2006 bp->b_kvasize = 0; 2007 } 2008 2009 /* 2010 * bufkva_alloc: 2011 * 2012 * Allocate the buffer KVA and set b_kvasize and b_kvabase. 2013 */ 2014 static int 2015 bufkva_alloc(struct buf *bp, int maxsize, int gbflags) 2016 { 2017 vm_offset_t addr; 2018 int error; 2019 2020 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0, 2021 ("Invalid gbflags 0x%x in %s", gbflags, __func__)); 2022 2023 bufkva_free(bp); 2024 2025 addr = 0; 2026 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr); 2027 if (error != 0) { 2028 /* 2029 * Buffer map is too fragmented. Request the caller 2030 * to defragment the map. 2031 */ 2032 return (error); 2033 } 2034 bp->b_kvabase = (caddr_t)addr; 2035 bp->b_kvasize = maxsize; 2036 counter_u64_add(bufkvaspace, bp->b_kvasize); 2037 if ((gbflags & GB_UNMAPPED) != 0) { 2038 bp->b_data = unmapped_buf; 2039 BUF_CHECK_UNMAPPED(bp); 2040 } else { 2041 bp->b_data = bp->b_kvabase; 2042 BUF_CHECK_MAPPED(bp); 2043 } 2044 return (0); 2045 } 2046 2047 /* 2048 * bufkva_reclaim: 2049 * 2050 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem 2051 * callback that fires to avoid returning failure. 2052 */ 2053 static void 2054 bufkva_reclaim(vmem_t *vmem, int flags) 2055 { 2056 bool done; 2057 int q; 2058 int i; 2059 2060 done = false; 2061 for (i = 0; i < 5; i++) { 2062 for (q = 0; q < buf_domains; q++) 2063 if (buf_recycle(&bdomain[q], true) != 0) 2064 done = true; 2065 if (done) 2066 break; 2067 } 2068 return; 2069 } 2070 2071 /* 2072 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 2073 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 2074 * the buffer is valid and we do not have to do anything. 2075 */ 2076 static void 2077 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt, 2078 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *)) 2079 { 2080 struct buf *rabp; 2081 struct thread *td; 2082 int i; 2083 2084 td = curthread; 2085 2086 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 2087 if (inmem(vp, *rablkno)) 2088 continue; 2089 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 2090 if ((rabp->b_flags & B_CACHE) != 0) { 2091 brelse(rabp); 2092 continue; 2093 } 2094 #ifdef RACCT 2095 if (racct_enable) { 2096 PROC_LOCK(curproc); 2097 racct_add_buf(curproc, rabp, 0); 2098 PROC_UNLOCK(curproc); 2099 } 2100 #endif /* RACCT */ 2101 td->td_ru.ru_inblock++; 2102 rabp->b_flags |= B_ASYNC; 2103 rabp->b_flags &= ~B_INVAL; 2104 if ((flags & GB_CKHASH) != 0) { 2105 rabp->b_flags |= B_CKHASH; 2106 rabp->b_ckhashcalc = ckhashfunc; 2107 } 2108 rabp->b_ioflags &= ~BIO_ERROR; 2109 rabp->b_iocmd = BIO_READ; 2110 if (rabp->b_rcred == NOCRED && cred != NOCRED) 2111 rabp->b_rcred = crhold(cred); 2112 vfs_busy_pages(rabp, 0); 2113 BUF_KERNPROC(rabp); 2114 rabp->b_iooffset = dbtob(rabp->b_blkno); 2115 bstrategy(rabp); 2116 } 2117 } 2118 2119 /* 2120 * Entry point for bread() and breadn() via #defines in sys/buf.h. 2121 * 2122 * Get a buffer with the specified data. Look in the cache first. We 2123 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 2124 * is set, the buffer is valid and we do not have to do anything, see 2125 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 2126 * 2127 * Always return a NULL buffer pointer (in bpp) when returning an error. 2128 * 2129 * The blkno parameter is the logical block being requested. Normally 2130 * the mapping of logical block number to disk block address is done 2131 * by calling VOP_BMAP(). However, if the mapping is already known, the 2132 * disk block address can be passed using the dblkno parameter. If the 2133 * disk block address is not known, then the same value should be passed 2134 * for blkno and dblkno. 2135 */ 2136 int 2137 breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, 2138 daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags, 2139 void (*ckhashfunc)(struct buf *), struct buf **bpp) 2140 { 2141 struct buf *bp; 2142 struct thread *td; 2143 int error, readwait, rv; 2144 2145 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 2146 td = curthread; 2147 /* 2148 * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags 2149 * are specified. 2150 */ 2151 error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp); 2152 if (error != 0) { 2153 *bpp = NULL; 2154 return (error); 2155 } 2156 KASSERT(blkno == bp->b_lblkno, 2157 ("getblkx returned buffer for blkno %jd instead of blkno %jd", 2158 (intmax_t)bp->b_lblkno, (intmax_t)blkno)); 2159 flags &= ~GB_NOSPARSE; 2160 *bpp = bp; 2161 2162 /* 2163 * If not found in cache, do some I/O 2164 */ 2165 readwait = 0; 2166 if ((bp->b_flags & B_CACHE) == 0) { 2167 #ifdef RACCT 2168 if (racct_enable) { 2169 PROC_LOCK(td->td_proc); 2170 racct_add_buf(td->td_proc, bp, 0); 2171 PROC_UNLOCK(td->td_proc); 2172 } 2173 #endif /* RACCT */ 2174 td->td_ru.ru_inblock++; 2175 bp->b_iocmd = BIO_READ; 2176 bp->b_flags &= ~B_INVAL; 2177 if ((flags & GB_CKHASH) != 0) { 2178 bp->b_flags |= B_CKHASH; 2179 bp->b_ckhashcalc = ckhashfunc; 2180 } 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 /* binsfree unlocks bp. */ 2779 binsfree(bp, qindex); 2780 } 2781 2782 /* 2783 * Release a buffer back to the appropriate queue but do not try to free 2784 * it. The buffer is expected to be used again soon. 2785 * 2786 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 2787 * biodone() to requeue an async I/O on completion. It is also used when 2788 * known good buffers need to be requeued but we think we may need the data 2789 * again soon. 2790 * 2791 * XXX we should be able to leave the B_RELBUF hint set on completion. 2792 */ 2793 void 2794 bqrelse(struct buf *bp) 2795 { 2796 int qindex; 2797 2798 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2799 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2800 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2801 2802 qindex = QUEUE_NONE; 2803 if (BUF_LOCKRECURSED(bp)) { 2804 /* do not release to free list */ 2805 BUF_UNLOCK(bp); 2806 return; 2807 } 2808 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 2809 2810 if (bp->b_flags & B_MANAGED) { 2811 if (bp->b_flags & B_REMFREE) 2812 bremfreef(bp); 2813 goto out; 2814 } 2815 2816 /* buffers with stale but valid contents */ 2817 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | 2818 BV_BKGRDERR)) == BV_BKGRDERR) { 2819 BO_LOCK(bp->b_bufobj); 2820 bp->b_vflags &= ~BV_BKGRDERR; 2821 BO_UNLOCK(bp->b_bufobj); 2822 qindex = QUEUE_DIRTY; 2823 } else { 2824 if ((bp->b_flags & B_DELWRI) == 0 && 2825 (bp->b_xflags & BX_VNDIRTY)) 2826 panic("bqrelse: not dirty"); 2827 if ((bp->b_flags & B_NOREUSE) != 0) { 2828 brelse(bp); 2829 return; 2830 } 2831 qindex = QUEUE_CLEAN; 2832 } 2833 buf_track(bp, __func__); 2834 /* binsfree unlocks bp. */ 2835 binsfree(bp, qindex); 2836 return; 2837 2838 out: 2839 buf_track(bp, __func__); 2840 /* unlock */ 2841 BUF_UNLOCK(bp); 2842 } 2843 2844 /* 2845 * Complete I/O to a VMIO backed page. Validate the pages as appropriate, 2846 * restore bogus pages. 2847 */ 2848 static void 2849 vfs_vmio_iodone(struct buf *bp) 2850 { 2851 vm_ooffset_t foff; 2852 vm_page_t m; 2853 vm_object_t obj; 2854 struct vnode *vp __unused; 2855 int i, iosize, resid; 2856 bool bogus; 2857 2858 obj = bp->b_bufobj->bo_object; 2859 KASSERT(REFCOUNT_COUNT(obj->paging_in_progress) >= bp->b_npages, 2860 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)", 2861 REFCOUNT_COUNT(obj->paging_in_progress), bp->b_npages)); 2862 2863 vp = bp->b_vp; 2864 KASSERT(vp->v_holdcnt > 0, 2865 ("vfs_vmio_iodone: vnode %p has zero hold count", vp)); 2866 KASSERT(vp->v_object != NULL, 2867 ("vfs_vmio_iodone: vnode %p has no vm_object", 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 if (bogus == false) { 2886 bogus = true; 2887 VM_OBJECT_RLOCK(obj); 2888 } 2889 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2890 if (m == NULL) 2891 panic("biodone: page disappeared!"); 2892 bp->b_pages[i] = m; 2893 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) { 2894 /* 2895 * In the write case, the valid and clean bits are 2896 * already changed correctly ( see bdwrite() ), so we 2897 * only need to do this here in the read case. 2898 */ 2899 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK, 2900 resid)) == 0, ("vfs_vmio_iodone: page %p " 2901 "has unexpected dirty bits", m)); 2902 vfs_page_set_valid(bp, foff, m); 2903 } 2904 KASSERT(OFF_TO_IDX(foff) == m->pindex, 2905 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch", 2906 (intmax_t)foff, (uintmax_t)m->pindex)); 2907 2908 vm_page_sunbusy(m); 2909 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2910 iosize -= resid; 2911 } 2912 if (bogus) 2913 VM_OBJECT_RUNLOCK(obj); 2914 vm_object_pip_wakeupn(obj, bp->b_npages); 2915 if (bogus && buf_mapped(bp)) { 2916 BUF_CHECK_MAPPED(bp); 2917 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 2918 bp->b_pages, bp->b_npages); 2919 } 2920 } 2921 2922 /* 2923 * Perform page invalidation when a buffer is released. The fully invalid 2924 * pages will be reclaimed later in vfs_vmio_truncate(). 2925 */ 2926 static void 2927 vfs_vmio_invalidate(struct buf *bp) 2928 { 2929 vm_object_t obj; 2930 vm_page_t m; 2931 int flags, i, resid, poffset, presid; 2932 2933 if (buf_mapped(bp)) { 2934 BUF_CHECK_MAPPED(bp); 2935 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 2936 } else 2937 BUF_CHECK_UNMAPPED(bp); 2938 /* 2939 * Get the base offset and length of the buffer. Note that 2940 * in the VMIO case if the buffer block size is not 2941 * page-aligned then b_data pointer may not be page-aligned. 2942 * But our b_pages[] array *IS* page aligned. 2943 * 2944 * block sizes less then DEV_BSIZE (usually 512) are not 2945 * supported due to the page granularity bits (m->valid, 2946 * m->dirty, etc...). 2947 * 2948 * See man buf(9) for more information 2949 */ 2950 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; 2951 obj = bp->b_bufobj->bo_object; 2952 resid = bp->b_bufsize; 2953 poffset = bp->b_offset & PAGE_MASK; 2954 VM_OBJECT_WLOCK(obj); 2955 for (i = 0; i < bp->b_npages; i++) { 2956 m = bp->b_pages[i]; 2957 if (m == bogus_page) 2958 panic("vfs_vmio_invalidate: Unexpected bogus page."); 2959 bp->b_pages[i] = NULL; 2960 2961 presid = resid > (PAGE_SIZE - poffset) ? 2962 (PAGE_SIZE - poffset) : resid; 2963 KASSERT(presid >= 0, ("brelse: extra page")); 2964 vm_page_busy_acquire(m, VM_ALLOC_SBUSY); 2965 if (pmap_page_wired_mappings(m) == 0) 2966 vm_page_set_invalid(m, poffset, presid); 2967 vm_page_sunbusy(m); 2968 vm_page_release_locked(m, flags); 2969 resid -= presid; 2970 poffset = 0; 2971 } 2972 VM_OBJECT_WUNLOCK(obj); 2973 bp->b_npages = 0; 2974 } 2975 2976 /* 2977 * Page-granular truncation of an existing VMIO buffer. 2978 */ 2979 static void 2980 vfs_vmio_truncate(struct buf *bp, int desiredpages) 2981 { 2982 vm_object_t obj; 2983 vm_page_t m; 2984 int flags, i; 2985 2986 if (bp->b_npages == desiredpages) 2987 return; 2988 2989 if (buf_mapped(bp)) { 2990 BUF_CHECK_MAPPED(bp); 2991 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) + 2992 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages); 2993 } else 2994 BUF_CHECK_UNMAPPED(bp); 2995 2996 /* 2997 * The object lock is needed only if we will attempt to free pages. 2998 */ 2999 flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; 3000 if ((bp->b_flags & B_DIRECT) != 0) { 3001 flags |= VPR_TRYFREE; 3002 obj = bp->b_bufobj->bo_object; 3003 VM_OBJECT_WLOCK(obj); 3004 } else { 3005 obj = NULL; 3006 } 3007 for (i = desiredpages; i < bp->b_npages; i++) { 3008 m = bp->b_pages[i]; 3009 KASSERT(m != bogus_page, ("allocbuf: bogus page found")); 3010 bp->b_pages[i] = NULL; 3011 if (obj != NULL) 3012 vm_page_release_locked(m, flags); 3013 else 3014 vm_page_release(m, flags); 3015 } 3016 if (obj != NULL) 3017 VM_OBJECT_WUNLOCK(obj); 3018 bp->b_npages = desiredpages; 3019 } 3020 3021 /* 3022 * Byte granular extension of VMIO buffers. 3023 */ 3024 static void 3025 vfs_vmio_extend(struct buf *bp, int desiredpages, int size) 3026 { 3027 /* 3028 * We are growing the buffer, possibly in a 3029 * byte-granular fashion. 3030 */ 3031 vm_object_t obj; 3032 vm_offset_t toff; 3033 vm_offset_t tinc; 3034 vm_page_t m; 3035 3036 /* 3037 * Step 1, bring in the VM pages from the object, allocating 3038 * them if necessary. We must clear B_CACHE if these pages 3039 * are not valid for the range covered by the buffer. 3040 */ 3041 obj = bp->b_bufobj->bo_object; 3042 if (bp->b_npages < desiredpages) { 3043 /* 3044 * We must allocate system pages since blocking 3045 * here could interfere with paging I/O, no 3046 * matter which process we are. 3047 * 3048 * Only exclusive busy can be tested here. 3049 * Blocking on shared busy might lead to 3050 * deadlocks once allocbuf() is called after 3051 * pages are vfs_busy_pages(). 3052 */ 3053 VM_OBJECT_WLOCK(obj); 3054 (void)vm_page_grab_pages(obj, 3055 OFF_TO_IDX(bp->b_offset) + bp->b_npages, 3056 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY | 3057 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED, 3058 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages); 3059 VM_OBJECT_WUNLOCK(obj); 3060 bp->b_npages = desiredpages; 3061 } 3062 3063 /* 3064 * Step 2. We've loaded the pages into the buffer, 3065 * we have to figure out if we can still have B_CACHE 3066 * set. Note that B_CACHE is set according to the 3067 * byte-granular range ( bcount and size ), not the 3068 * aligned range ( newbsize ). 3069 * 3070 * The VM test is against m->valid, which is DEV_BSIZE 3071 * aligned. Needless to say, the validity of the data 3072 * needs to also be DEV_BSIZE aligned. Note that this 3073 * fails with NFS if the server or some other client 3074 * extends the file's EOF. If our buffer is resized, 3075 * B_CACHE may remain set! XXX 3076 */ 3077 toff = bp->b_bcount; 3078 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3079 while ((bp->b_flags & B_CACHE) && toff < size) { 3080 vm_pindex_t pi; 3081 3082 if (tinc > (size - toff)) 3083 tinc = size - toff; 3084 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; 3085 m = bp->b_pages[pi]; 3086 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m); 3087 toff += tinc; 3088 tinc = PAGE_SIZE; 3089 } 3090 3091 /* 3092 * Step 3, fixup the KVA pmap. 3093 */ 3094 if (buf_mapped(bp)) 3095 bpmap_qenter(bp); 3096 else 3097 BUF_CHECK_UNMAPPED(bp); 3098 } 3099 3100 /* 3101 * Check to see if a block at a particular lbn is available for a clustered 3102 * write. 3103 */ 3104 static int 3105 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 3106 { 3107 struct buf *bpa; 3108 int match; 3109 3110 match = 0; 3111 3112 /* If the buf isn't in core skip it */ 3113 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 3114 return (0); 3115 3116 /* If the buf is busy we don't want to wait for it */ 3117 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 3118 return (0); 3119 3120 /* Only cluster with valid clusterable delayed write buffers */ 3121 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 3122 (B_DELWRI | B_CLUSTEROK)) 3123 goto done; 3124 3125 if (bpa->b_bufsize != size) 3126 goto done; 3127 3128 /* 3129 * Check to see if it is in the expected place on disk and that the 3130 * block has been mapped. 3131 */ 3132 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 3133 match = 1; 3134 done: 3135 BUF_UNLOCK(bpa); 3136 return (match); 3137 } 3138 3139 /* 3140 * vfs_bio_awrite: 3141 * 3142 * Implement clustered async writes for clearing out B_DELWRI buffers. 3143 * This is much better then the old way of writing only one buffer at 3144 * a time. Note that we may not be presented with the buffers in the 3145 * correct order, so we search for the cluster in both directions. 3146 */ 3147 int 3148 vfs_bio_awrite(struct buf *bp) 3149 { 3150 struct bufobj *bo; 3151 int i; 3152 int j; 3153 daddr_t lblkno = bp->b_lblkno; 3154 struct vnode *vp = bp->b_vp; 3155 int ncl; 3156 int nwritten; 3157 int size; 3158 int maxcl; 3159 int gbflags; 3160 3161 bo = &vp->v_bufobj; 3162 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0; 3163 /* 3164 * right now we support clustered writing only to regular files. If 3165 * we find a clusterable block we could be in the middle of a cluster 3166 * rather then at the beginning. 3167 */ 3168 if ((vp->v_type == VREG) && 3169 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 3170 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 3171 3172 size = vp->v_mount->mnt_stat.f_iosize; 3173 maxcl = MAXPHYS / size; 3174 3175 BO_RLOCK(bo); 3176 for (i = 1; i < maxcl; i++) 3177 if (vfs_bio_clcheck(vp, size, lblkno + i, 3178 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 3179 break; 3180 3181 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 3182 if (vfs_bio_clcheck(vp, size, lblkno - j, 3183 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 3184 break; 3185 BO_RUNLOCK(bo); 3186 --j; 3187 ncl = i + j; 3188 /* 3189 * this is a possible cluster write 3190 */ 3191 if (ncl != 1) { 3192 BUF_UNLOCK(bp); 3193 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 3194 gbflags); 3195 return (nwritten); 3196 } 3197 } 3198 bremfree(bp); 3199 bp->b_flags |= B_ASYNC; 3200 /* 3201 * default (old) behavior, writing out only one block 3202 * 3203 * XXX returns b_bufsize instead of b_bcount for nwritten? 3204 */ 3205 nwritten = bp->b_bufsize; 3206 (void) bwrite(bp); 3207 3208 return (nwritten); 3209 } 3210 3211 /* 3212 * getnewbuf_kva: 3213 * 3214 * Allocate KVA for an empty buf header according to gbflags. 3215 */ 3216 static int 3217 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize) 3218 { 3219 3220 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) { 3221 /* 3222 * In order to keep fragmentation sane we only allocate kva 3223 * in BKVASIZE chunks. XXX with vmem we can do page size. 3224 */ 3225 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 3226 3227 if (maxsize != bp->b_kvasize && 3228 bufkva_alloc(bp, maxsize, gbflags)) 3229 return (ENOSPC); 3230 } 3231 return (0); 3232 } 3233 3234 /* 3235 * getnewbuf: 3236 * 3237 * Find and initialize a new buffer header, freeing up existing buffers 3238 * in the bufqueues as necessary. The new buffer is returned locked. 3239 * 3240 * We block if: 3241 * We have insufficient buffer headers 3242 * We have insufficient buffer space 3243 * buffer_arena is too fragmented ( space reservation fails ) 3244 * If we have to flush dirty buffers ( but we try to avoid this ) 3245 * 3246 * The caller is responsible for releasing the reserved bufspace after 3247 * allocbuf() is called. 3248 */ 3249 static struct buf * 3250 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags) 3251 { 3252 struct bufdomain *bd; 3253 struct buf *bp; 3254 bool metadata, reserved; 3255 3256 bp = NULL; 3257 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3258 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3259 if (!unmapped_buf_allowed) 3260 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3261 3262 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 3263 vp->v_type == VCHR) 3264 metadata = true; 3265 else 3266 metadata = false; 3267 if (vp == NULL) 3268 bd = &bdomain[0]; 3269 else 3270 bd = &bdomain[vp->v_bufobj.bo_domain]; 3271 3272 counter_u64_add(getnewbufcalls, 1); 3273 reserved = false; 3274 do { 3275 if (reserved == false && 3276 bufspace_reserve(bd, maxsize, metadata) != 0) { 3277 counter_u64_add(getnewbufrestarts, 1); 3278 continue; 3279 } 3280 reserved = true; 3281 if ((bp = buf_alloc(bd)) == NULL) { 3282 counter_u64_add(getnewbufrestarts, 1); 3283 continue; 3284 } 3285 if (getnewbuf_kva(bp, gbflags, maxsize) == 0) 3286 return (bp); 3287 break; 3288 } while (buf_recycle(bd, false) == 0); 3289 3290 if (reserved) 3291 bufspace_release(bd, maxsize); 3292 if (bp != NULL) { 3293 bp->b_flags |= B_INVAL; 3294 brelse(bp); 3295 } 3296 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo); 3297 3298 return (NULL); 3299 } 3300 3301 /* 3302 * buf_daemon: 3303 * 3304 * buffer flushing daemon. Buffers are normally flushed by the 3305 * update daemon but if it cannot keep up this process starts to 3306 * take the load in an attempt to prevent getnewbuf() from blocking. 3307 */ 3308 static struct kproc_desc buf_kp = { 3309 "bufdaemon", 3310 buf_daemon, 3311 &bufdaemonproc 3312 }; 3313 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 3314 3315 static int 3316 buf_flush(struct vnode *vp, struct bufdomain *bd, int target) 3317 { 3318 int flushed; 3319 3320 flushed = flushbufqueues(vp, bd, target, 0); 3321 if (flushed == 0) { 3322 /* 3323 * Could not find any buffers without rollback 3324 * dependencies, so just write the first one 3325 * in the hopes of eventually making progress. 3326 */ 3327 if (vp != NULL && target > 2) 3328 target /= 2; 3329 flushbufqueues(vp, bd, target, 1); 3330 } 3331 return (flushed); 3332 } 3333 3334 static void 3335 buf_daemon() 3336 { 3337 struct bufdomain *bd; 3338 int speedupreq; 3339 int lodirty; 3340 int i; 3341 3342 /* 3343 * This process needs to be suspended prior to shutdown sync. 3344 */ 3345 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread, 3346 SHUTDOWN_PRI_LAST + 100); 3347 3348 /* 3349 * Start the buf clean daemons as children threads. 3350 */ 3351 for (i = 0 ; i < buf_domains; i++) { 3352 int error; 3353 3354 error = kthread_add((void (*)(void *))bufspace_daemon, 3355 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i); 3356 if (error) 3357 panic("error %d spawning bufspace daemon", error); 3358 } 3359 3360 /* 3361 * This process is allowed to take the buffer cache to the limit 3362 */ 3363 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 3364 mtx_lock(&bdlock); 3365 for (;;) { 3366 bd_request = 0; 3367 mtx_unlock(&bdlock); 3368 3369 kthread_suspend_check(); 3370 3371 /* 3372 * Save speedupreq for this pass and reset to capture new 3373 * requests. 3374 */ 3375 speedupreq = bd_speedupreq; 3376 bd_speedupreq = 0; 3377 3378 /* 3379 * Flush each domain sequentially according to its level and 3380 * the speedup request. 3381 */ 3382 for (i = 0; i < buf_domains; i++) { 3383 bd = &bdomain[i]; 3384 if (speedupreq) 3385 lodirty = bd->bd_numdirtybuffers / 2; 3386 else 3387 lodirty = bd->bd_lodirtybuffers; 3388 while (bd->bd_numdirtybuffers > lodirty) { 3389 if (buf_flush(NULL, bd, 3390 bd->bd_numdirtybuffers - lodirty) == 0) 3391 break; 3392 kern_yield(PRI_USER); 3393 } 3394 } 3395 3396 /* 3397 * Only clear bd_request if we have reached our low water 3398 * mark. The buf_daemon normally waits 1 second and 3399 * then incrementally flushes any dirty buffers that have 3400 * built up, within reason. 3401 * 3402 * If we were unable to hit our low water mark and couldn't 3403 * find any flushable buffers, we sleep for a short period 3404 * to avoid endless loops on unlockable buffers. 3405 */ 3406 mtx_lock(&bdlock); 3407 if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) { 3408 /* 3409 * We reached our low water mark, reset the 3410 * request and sleep until we are needed again. 3411 * The sleep is just so the suspend code works. 3412 */ 3413 bd_request = 0; 3414 /* 3415 * Do an extra wakeup in case dirty threshold 3416 * changed via sysctl and the explicit transition 3417 * out of shortfall was missed. 3418 */ 3419 bdirtywakeup(); 3420 if (runningbufspace <= lorunningspace) 3421 runningwakeup(); 3422 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 3423 } else { 3424 /* 3425 * We couldn't find any flushable dirty buffers but 3426 * still have too many dirty buffers, we 3427 * have to sleep and try again. (rare) 3428 */ 3429 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 3430 } 3431 } 3432 } 3433 3434 /* 3435 * flushbufqueues: 3436 * 3437 * Try to flush a buffer in the dirty queue. We must be careful to 3438 * free up B_INVAL buffers instead of write them, which NFS is 3439 * particularly sensitive to. 3440 */ 3441 static int flushwithdeps = 0; 3442 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS, 3443 &flushwithdeps, 0, 3444 "Number of buffers flushed with dependecies that require rollbacks"); 3445 3446 static int 3447 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target, 3448 int flushdeps) 3449 { 3450 struct bufqueue *bq; 3451 struct buf *sentinel; 3452 struct vnode *vp; 3453 struct mount *mp; 3454 struct buf *bp; 3455 int hasdeps; 3456 int flushed; 3457 int error; 3458 bool unlock; 3459 3460 flushed = 0; 3461 bq = &bd->bd_dirtyq; 3462 bp = NULL; 3463 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 3464 sentinel->b_qindex = QUEUE_SENTINEL; 3465 BQ_LOCK(bq); 3466 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist); 3467 BQ_UNLOCK(bq); 3468 while (flushed != target) { 3469 maybe_yield(); 3470 BQ_LOCK(bq); 3471 bp = TAILQ_NEXT(sentinel, b_freelist); 3472 if (bp != NULL) { 3473 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); 3474 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel, 3475 b_freelist); 3476 } else { 3477 BQ_UNLOCK(bq); 3478 break; 3479 } 3480 /* 3481 * Skip sentinels inserted by other invocations of the 3482 * flushbufqueues(), taking care to not reorder them. 3483 * 3484 * Only flush the buffers that belong to the 3485 * vnode locked by the curthread. 3486 */ 3487 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && 3488 bp->b_vp != lvp)) { 3489 BQ_UNLOCK(bq); 3490 continue; 3491 } 3492 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 3493 BQ_UNLOCK(bq); 3494 if (error != 0) 3495 continue; 3496 3497 /* 3498 * BKGRDINPROG can only be set with the buf and bufobj 3499 * locks both held. We tolerate a race to clear it here. 3500 */ 3501 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 3502 (bp->b_flags & B_DELWRI) == 0) { 3503 BUF_UNLOCK(bp); 3504 continue; 3505 } 3506 if (bp->b_flags & B_INVAL) { 3507 bremfreef(bp); 3508 brelse(bp); 3509 flushed++; 3510 continue; 3511 } 3512 3513 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 3514 if (flushdeps == 0) { 3515 BUF_UNLOCK(bp); 3516 continue; 3517 } 3518 hasdeps = 1; 3519 } else 3520 hasdeps = 0; 3521 /* 3522 * We must hold the lock on a vnode before writing 3523 * one of its buffers. Otherwise we may confuse, or 3524 * in the case of a snapshot vnode, deadlock the 3525 * system. 3526 * 3527 * The lock order here is the reverse of the normal 3528 * of vnode followed by buf lock. This is ok because 3529 * the NOWAIT will prevent deadlock. 3530 */ 3531 vp = bp->b_vp; 3532 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 3533 BUF_UNLOCK(bp); 3534 continue; 3535 } 3536 if (lvp == NULL) { 3537 unlock = true; 3538 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 3539 } else { 3540 ASSERT_VOP_LOCKED(vp, "getbuf"); 3541 unlock = false; 3542 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 3543 vn_lock(vp, LK_TRYUPGRADE); 3544 } 3545 if (error == 0) { 3546 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 3547 bp, bp->b_vp, bp->b_flags); 3548 if (curproc == bufdaemonproc) { 3549 vfs_bio_awrite(bp); 3550 } else { 3551 bremfree(bp); 3552 bwrite(bp); 3553 counter_u64_add(notbufdflushes, 1); 3554 } 3555 vn_finished_write(mp); 3556 if (unlock) 3557 VOP_UNLOCK(vp); 3558 flushwithdeps += hasdeps; 3559 flushed++; 3560 3561 /* 3562 * Sleeping on runningbufspace while holding 3563 * vnode lock leads to deadlock. 3564 */ 3565 if (curproc == bufdaemonproc && 3566 runningbufspace > hirunningspace) 3567 waitrunningbufspace(); 3568 continue; 3569 } 3570 vn_finished_write(mp); 3571 BUF_UNLOCK(bp); 3572 } 3573 BQ_LOCK(bq); 3574 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); 3575 BQ_UNLOCK(bq); 3576 free(sentinel, M_TEMP); 3577 return (flushed); 3578 } 3579 3580 /* 3581 * Check to see if a block is currently memory resident. 3582 */ 3583 struct buf * 3584 incore(struct bufobj *bo, daddr_t blkno) 3585 { 3586 struct buf *bp; 3587 3588 BO_RLOCK(bo); 3589 bp = gbincore(bo, blkno); 3590 BO_RUNLOCK(bo); 3591 return (bp); 3592 } 3593 3594 /* 3595 * Returns true if no I/O is needed to access the 3596 * associated VM object. This is like incore except 3597 * it also hunts around in the VM system for the data. 3598 */ 3599 3600 static int 3601 inmem(struct vnode * vp, daddr_t blkno) 3602 { 3603 vm_object_t obj; 3604 vm_offset_t toff, tinc, size; 3605 vm_page_t m; 3606 vm_ooffset_t off; 3607 3608 ASSERT_VOP_LOCKED(vp, "inmem"); 3609 3610 if (incore(&vp->v_bufobj, blkno)) 3611 return 1; 3612 if (vp->v_mount == NULL) 3613 return 0; 3614 obj = vp->v_object; 3615 if (obj == NULL) 3616 return (0); 3617 3618 size = PAGE_SIZE; 3619 if (size > vp->v_mount->mnt_stat.f_iosize) 3620 size = vp->v_mount->mnt_stat.f_iosize; 3621 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 3622 3623 VM_OBJECT_RLOCK(obj); 3624 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 3625 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 3626 if (!m) 3627 goto notinmem; 3628 tinc = size; 3629 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 3630 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 3631 if (vm_page_is_valid(m, 3632 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 3633 goto notinmem; 3634 } 3635 VM_OBJECT_RUNLOCK(obj); 3636 return 1; 3637 3638 notinmem: 3639 VM_OBJECT_RUNLOCK(obj); 3640 return (0); 3641 } 3642 3643 /* 3644 * Set the dirty range for a buffer based on the status of the dirty 3645 * bits in the pages comprising the buffer. The range is limited 3646 * to the size of the buffer. 3647 * 3648 * Tell the VM system that the pages associated with this buffer 3649 * are clean. This is used for delayed writes where the data is 3650 * going to go to disk eventually without additional VM intevention. 3651 * 3652 * Note that while we only really need to clean through to b_bcount, we 3653 * just go ahead and clean through to b_bufsize. 3654 */ 3655 static void 3656 vfs_clean_pages_dirty_buf(struct buf *bp) 3657 { 3658 vm_ooffset_t foff, noff, eoff; 3659 vm_page_t m; 3660 int i; 3661 3662 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 3663 return; 3664 3665 foff = bp->b_offset; 3666 KASSERT(bp->b_offset != NOOFFSET, 3667 ("vfs_clean_pages_dirty_buf: no buffer offset")); 3668 3669 vfs_busy_pages_acquire(bp); 3670 vfs_setdirty_range(bp); 3671 for (i = 0; i < bp->b_npages; i++) { 3672 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3673 eoff = noff; 3674 if (eoff > bp->b_offset + bp->b_bufsize) 3675 eoff = bp->b_offset + bp->b_bufsize; 3676 m = bp->b_pages[i]; 3677 vfs_page_set_validclean(bp, foff, m); 3678 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3679 foff = noff; 3680 } 3681 vfs_busy_pages_release(bp); 3682 } 3683 3684 static void 3685 vfs_setdirty_range(struct buf *bp) 3686 { 3687 vm_offset_t boffset; 3688 vm_offset_t eoffset; 3689 int i; 3690 3691 /* 3692 * test the pages to see if they have been modified directly 3693 * by users through the VM system. 3694 */ 3695 for (i = 0; i < bp->b_npages; i++) 3696 vm_page_test_dirty(bp->b_pages[i]); 3697 3698 /* 3699 * Calculate the encompassing dirty range, boffset and eoffset, 3700 * (eoffset - boffset) bytes. 3701 */ 3702 3703 for (i = 0; i < bp->b_npages; i++) { 3704 if (bp->b_pages[i]->dirty) 3705 break; 3706 } 3707 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3708 3709 for (i = bp->b_npages - 1; i >= 0; --i) { 3710 if (bp->b_pages[i]->dirty) { 3711 break; 3712 } 3713 } 3714 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3715 3716 /* 3717 * Fit it to the buffer. 3718 */ 3719 3720 if (eoffset > bp->b_bcount) 3721 eoffset = bp->b_bcount; 3722 3723 /* 3724 * If we have a good dirty range, merge with the existing 3725 * dirty range. 3726 */ 3727 3728 if (boffset < eoffset) { 3729 if (bp->b_dirtyoff > boffset) 3730 bp->b_dirtyoff = boffset; 3731 if (bp->b_dirtyend < eoffset) 3732 bp->b_dirtyend = eoffset; 3733 } 3734 } 3735 3736 /* 3737 * Allocate the KVA mapping for an existing buffer. 3738 * If an unmapped buffer is provided but a mapped buffer is requested, take 3739 * also care to properly setup mappings between pages and KVA. 3740 */ 3741 static void 3742 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 3743 { 3744 int bsize, maxsize, need_mapping, need_kva; 3745 off_t offset; 3746 3747 need_mapping = bp->b_data == unmapped_buf && 3748 (gbflags & GB_UNMAPPED) == 0; 3749 need_kva = bp->b_kvabase == unmapped_buf && 3750 bp->b_data == unmapped_buf && 3751 (gbflags & GB_KVAALLOC) != 0; 3752 if (!need_mapping && !need_kva) 3753 return; 3754 3755 BUF_CHECK_UNMAPPED(bp); 3756 3757 if (need_mapping && bp->b_kvabase != unmapped_buf) { 3758 /* 3759 * Buffer is not mapped, but the KVA was already 3760 * reserved at the time of the instantiation. Use the 3761 * allocated space. 3762 */ 3763 goto has_addr; 3764 } 3765 3766 /* 3767 * Calculate the amount of the address space we would reserve 3768 * if the buffer was mapped. 3769 */ 3770 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 3771 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3772 offset = blkno * bsize; 3773 maxsize = size + (offset & PAGE_MASK); 3774 maxsize = imax(maxsize, bsize); 3775 3776 while (bufkva_alloc(bp, maxsize, gbflags) != 0) { 3777 if ((gbflags & GB_NOWAIT_BD) != 0) { 3778 /* 3779 * XXXKIB: defragmentation cannot 3780 * succeed, not sure what else to do. 3781 */ 3782 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp); 3783 } 3784 counter_u64_add(mappingrestarts, 1); 3785 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0); 3786 } 3787 has_addr: 3788 if (need_mapping) { 3789 /* b_offset is handled by bpmap_qenter. */ 3790 bp->b_data = bp->b_kvabase; 3791 BUF_CHECK_MAPPED(bp); 3792 bpmap_qenter(bp); 3793 } 3794 } 3795 3796 struct buf * 3797 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3798 int flags) 3799 { 3800 struct buf *bp; 3801 int error; 3802 3803 error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp); 3804 if (error != 0) 3805 return (NULL); 3806 return (bp); 3807 } 3808 3809 /* 3810 * getblkx: 3811 * 3812 * Get a block given a specified block and offset into a file/device. 3813 * The buffers B_DONE bit will be cleared on return, making it almost 3814 * ready for an I/O initiation. B_INVAL may or may not be set on 3815 * return. The caller should clear B_INVAL prior to initiating a 3816 * READ. 3817 * 3818 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3819 * an existing buffer. 3820 * 3821 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3822 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3823 * and then cleared based on the backing VM. If the previous buffer is 3824 * non-0-sized but invalid, B_CACHE will be cleared. 3825 * 3826 * If getblk() must create a new buffer, the new buffer is returned with 3827 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3828 * case it is returned with B_INVAL clear and B_CACHE set based on the 3829 * backing VM. 3830 * 3831 * getblk() also forces a bwrite() for any B_DELWRI buffer whose 3832 * B_CACHE bit is clear. 3833 * 3834 * What this means, basically, is that the caller should use B_CACHE to 3835 * determine whether the buffer is fully valid or not and should clear 3836 * B_INVAL prior to issuing a read. If the caller intends to validate 3837 * the buffer by loading its data area with something, the caller needs 3838 * to clear B_INVAL. If the caller does this without issuing an I/O, 3839 * the caller should set B_CACHE ( as an optimization ), else the caller 3840 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3841 * a write attempt or if it was a successful read. If the caller 3842 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3843 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3844 * 3845 * The blkno parameter is the logical block being requested. Normally 3846 * the mapping of logical block number to disk block address is done 3847 * by calling VOP_BMAP(). However, if the mapping is already known, the 3848 * disk block address can be passed using the dblkno parameter. If the 3849 * disk block address is not known, then the same value should be passed 3850 * for blkno and dblkno. 3851 */ 3852 int 3853 getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag, 3854 int slptimeo, int flags, struct buf **bpp) 3855 { 3856 struct buf *bp; 3857 struct bufobj *bo; 3858 daddr_t d_blkno; 3859 int bsize, error, maxsize, vmio; 3860 off_t offset; 3861 3862 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3863 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3864 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3865 ASSERT_VOP_LOCKED(vp, "getblk"); 3866 if (size > maxbcachebuf) 3867 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size, 3868 maxbcachebuf); 3869 if (!unmapped_buf_allowed) 3870 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3871 3872 bo = &vp->v_bufobj; 3873 d_blkno = dblkno; 3874 loop: 3875 BO_RLOCK(bo); 3876 bp = gbincore(bo, blkno); 3877 if (bp != NULL) { 3878 int lockflags; 3879 /* 3880 * Buffer is in-core. If the buffer is not busy nor managed, 3881 * it must be on a queue. 3882 */ 3883 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3884 3885 if ((flags & GB_LOCK_NOWAIT) != 0) 3886 lockflags |= LK_NOWAIT; 3887 3888 error = BUF_TIMELOCK(bp, lockflags, 3889 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3890 3891 /* 3892 * If we slept and got the lock we have to restart in case 3893 * the buffer changed identities. 3894 */ 3895 if (error == ENOLCK) 3896 goto loop; 3897 /* We timed out or were interrupted. */ 3898 else if (error != 0) 3899 return (error); 3900 /* If recursed, assume caller knows the rules. */ 3901 else if (BUF_LOCKRECURSED(bp)) 3902 goto end; 3903 3904 /* 3905 * The buffer is locked. B_CACHE is cleared if the buffer is 3906 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3907 * and for a VMIO buffer B_CACHE is adjusted according to the 3908 * backing VM cache. 3909 */ 3910 if (bp->b_flags & B_INVAL) 3911 bp->b_flags &= ~B_CACHE; 3912 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3913 bp->b_flags |= B_CACHE; 3914 if (bp->b_flags & B_MANAGED) 3915 MPASS(bp->b_qindex == QUEUE_NONE); 3916 else 3917 bremfree(bp); 3918 3919 /* 3920 * check for size inconsistencies for non-VMIO case. 3921 */ 3922 if (bp->b_bcount != size) { 3923 if ((bp->b_flags & B_VMIO) == 0 || 3924 (size > bp->b_kvasize)) { 3925 if (bp->b_flags & B_DELWRI) { 3926 bp->b_flags |= B_NOCACHE; 3927 bwrite(bp); 3928 } else { 3929 if (LIST_EMPTY(&bp->b_dep)) { 3930 bp->b_flags |= B_RELBUF; 3931 brelse(bp); 3932 } else { 3933 bp->b_flags |= B_NOCACHE; 3934 bwrite(bp); 3935 } 3936 } 3937 goto loop; 3938 } 3939 } 3940 3941 /* 3942 * Handle the case of unmapped buffer which should 3943 * become mapped, or the buffer for which KVA 3944 * reservation is requested. 3945 */ 3946 bp_unmapped_get_kva(bp, blkno, size, flags); 3947 3948 /* 3949 * If the size is inconsistent in the VMIO case, we can resize 3950 * the buffer. This might lead to B_CACHE getting set or 3951 * cleared. If the size has not changed, B_CACHE remains 3952 * unchanged from its previous state. 3953 */ 3954 allocbuf(bp, size); 3955 3956 KASSERT(bp->b_offset != NOOFFSET, 3957 ("getblk: no buffer offset")); 3958 3959 /* 3960 * A buffer with B_DELWRI set and B_CACHE clear must 3961 * be committed before we can return the buffer in 3962 * order to prevent the caller from issuing a read 3963 * ( due to B_CACHE not being set ) and overwriting 3964 * it. 3965 * 3966 * Most callers, including NFS and FFS, need this to 3967 * operate properly either because they assume they 3968 * can issue a read if B_CACHE is not set, or because 3969 * ( for example ) an uncached B_DELWRI might loop due 3970 * to softupdates re-dirtying the buffer. In the latter 3971 * case, B_CACHE is set after the first write completes, 3972 * preventing further loops. 3973 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3974 * above while extending the buffer, we cannot allow the 3975 * buffer to remain with B_CACHE set after the write 3976 * completes or it will represent a corrupt state. To 3977 * deal with this we set B_NOCACHE to scrap the buffer 3978 * after the write. 3979 * 3980 * We might be able to do something fancy, like setting 3981 * B_CACHE in bwrite() except if B_DELWRI is already set, 3982 * so the below call doesn't set B_CACHE, but that gets real 3983 * confusing. This is much easier. 3984 */ 3985 3986 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3987 bp->b_flags |= B_NOCACHE; 3988 bwrite(bp); 3989 goto loop; 3990 } 3991 bp->b_flags &= ~B_DONE; 3992 } else { 3993 /* 3994 * Buffer is not in-core, create new buffer. The buffer 3995 * returned by getnewbuf() is locked. Note that the returned 3996 * buffer is also considered valid (not marked B_INVAL). 3997 */ 3998 BO_RUNLOCK(bo); 3999 /* 4000 * If the user does not want us to create the buffer, bail out 4001 * here. 4002 */ 4003 if (flags & GB_NOCREAT) 4004 return (EEXIST); 4005 4006 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 4007 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 4008 offset = blkno * bsize; 4009 vmio = vp->v_object != NULL; 4010 if (vmio) { 4011 maxsize = size + (offset & PAGE_MASK); 4012 } else { 4013 maxsize = size; 4014 /* Do not allow non-VMIO notmapped buffers. */ 4015 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 4016 } 4017 maxsize = imax(maxsize, bsize); 4018 if ((flags & GB_NOSPARSE) != 0 && vmio && 4019 !vn_isdisk(vp, NULL)) { 4020 error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0); 4021 KASSERT(error != EOPNOTSUPP, 4022 ("GB_NOSPARSE from fs not supporting bmap, vp %p", 4023 vp)); 4024 if (error != 0) 4025 return (error); 4026 if (d_blkno == -1) 4027 return (EJUSTRETURN); 4028 } 4029 4030 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags); 4031 if (bp == NULL) { 4032 if (slpflag || slptimeo) 4033 return (ETIMEDOUT); 4034 /* 4035 * XXX This is here until the sleep path is diagnosed 4036 * enough to work under very low memory conditions. 4037 * 4038 * There's an issue on low memory, 4BSD+non-preempt 4039 * systems (eg MIPS routers with 32MB RAM) where buffer 4040 * exhaustion occurs without sleeping for buffer 4041 * reclaimation. This just sticks in a loop and 4042 * constantly attempts to allocate a buffer, which 4043 * hits exhaustion and tries to wakeup bufdaemon. 4044 * This never happens because we never yield. 4045 * 4046 * The real solution is to identify and fix these cases 4047 * so we aren't effectively busy-waiting in a loop 4048 * until the reclaimation path has cycles to run. 4049 */ 4050 kern_yield(PRI_USER); 4051 goto loop; 4052 } 4053 4054 /* 4055 * This code is used to make sure that a buffer is not 4056 * created while the getnewbuf routine is blocked. 4057 * This can be a problem whether the vnode is locked or not. 4058 * If the buffer is created out from under us, we have to 4059 * throw away the one we just created. 4060 * 4061 * Note: this must occur before we associate the buffer 4062 * with the vp especially considering limitations in 4063 * the splay tree implementation when dealing with duplicate 4064 * lblkno's. 4065 */ 4066 BO_LOCK(bo); 4067 if (gbincore(bo, blkno)) { 4068 BO_UNLOCK(bo); 4069 bp->b_flags |= B_INVAL; 4070 bufspace_release(bufdomain(bp), maxsize); 4071 brelse(bp); 4072 goto loop; 4073 } 4074 4075 /* 4076 * Insert the buffer into the hash, so that it can 4077 * be found by incore. 4078 */ 4079 bp->b_lblkno = blkno; 4080 bp->b_blkno = d_blkno; 4081 bp->b_offset = offset; 4082 bgetvp(vp, bp); 4083 BO_UNLOCK(bo); 4084 4085 /* 4086 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 4087 * buffer size starts out as 0, B_CACHE will be set by 4088 * allocbuf() for the VMIO case prior to it testing the 4089 * backing store for validity. 4090 */ 4091 4092 if (vmio) { 4093 bp->b_flags |= B_VMIO; 4094 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 4095 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 4096 bp, vp->v_object, bp->b_bufobj->bo_object)); 4097 } else { 4098 bp->b_flags &= ~B_VMIO; 4099 KASSERT(bp->b_bufobj->bo_object == NULL, 4100 ("ARGH! has b_bufobj->bo_object %p %p\n", 4101 bp, bp->b_bufobj->bo_object)); 4102 BUF_CHECK_MAPPED(bp); 4103 } 4104 4105 allocbuf(bp, size); 4106 bufspace_release(bufdomain(bp), maxsize); 4107 bp->b_flags &= ~B_DONE; 4108 } 4109 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 4110 end: 4111 buf_track(bp, __func__); 4112 KASSERT(bp->b_bufobj == bo, 4113 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 4114 *bpp = bp; 4115 return (0); 4116 } 4117 4118 /* 4119 * Get an empty, disassociated buffer of given size. The buffer is initially 4120 * set to B_INVAL. 4121 */ 4122 struct buf * 4123 geteblk(int size, int flags) 4124 { 4125 struct buf *bp; 4126 int maxsize; 4127 4128 maxsize = (size + BKVAMASK) & ~BKVAMASK; 4129 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) { 4130 if ((flags & GB_NOWAIT_BD) && 4131 (curthread->td_pflags & TDP_BUFNEED) != 0) 4132 return (NULL); 4133 } 4134 allocbuf(bp, size); 4135 bufspace_release(bufdomain(bp), maxsize); 4136 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 4137 return (bp); 4138 } 4139 4140 /* 4141 * Truncate the backing store for a non-vmio buffer. 4142 */ 4143 static void 4144 vfs_nonvmio_truncate(struct buf *bp, int newbsize) 4145 { 4146 4147 if (bp->b_flags & B_MALLOC) { 4148 /* 4149 * malloced buffers are not shrunk 4150 */ 4151 if (newbsize == 0) { 4152 bufmallocadjust(bp, 0); 4153 free(bp->b_data, M_BIOBUF); 4154 bp->b_data = bp->b_kvabase; 4155 bp->b_flags &= ~B_MALLOC; 4156 } 4157 return; 4158 } 4159 vm_hold_free_pages(bp, newbsize); 4160 bufspace_adjust(bp, newbsize); 4161 } 4162 4163 /* 4164 * Extend the backing for a non-VMIO buffer. 4165 */ 4166 static void 4167 vfs_nonvmio_extend(struct buf *bp, int newbsize) 4168 { 4169 caddr_t origbuf; 4170 int origbufsize; 4171 4172 /* 4173 * We only use malloced memory on the first allocation. 4174 * and revert to page-allocated memory when the buffer 4175 * grows. 4176 * 4177 * There is a potential smp race here that could lead 4178 * to bufmallocspace slightly passing the max. It 4179 * is probably extremely rare and not worth worrying 4180 * over. 4181 */ 4182 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 && 4183 bufmallocspace < maxbufmallocspace) { 4184 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK); 4185 bp->b_flags |= B_MALLOC; 4186 bufmallocadjust(bp, newbsize); 4187 return; 4188 } 4189 4190 /* 4191 * If the buffer is growing on its other-than-first 4192 * allocation then we revert to the page-allocation 4193 * scheme. 4194 */ 4195 origbuf = NULL; 4196 origbufsize = 0; 4197 if (bp->b_flags & B_MALLOC) { 4198 origbuf = bp->b_data; 4199 origbufsize = bp->b_bufsize; 4200 bp->b_data = bp->b_kvabase; 4201 bufmallocadjust(bp, 0); 4202 bp->b_flags &= ~B_MALLOC; 4203 newbsize = round_page(newbsize); 4204 } 4205 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize, 4206 (vm_offset_t) bp->b_data + newbsize); 4207 if (origbuf != NULL) { 4208 bcopy(origbuf, bp->b_data, origbufsize); 4209 free(origbuf, M_BIOBUF); 4210 } 4211 bufspace_adjust(bp, newbsize); 4212 } 4213 4214 /* 4215 * This code constitutes the buffer memory from either anonymous system 4216 * memory (in the case of non-VMIO operations) or from an associated 4217 * VM object (in the case of VMIO operations). This code is able to 4218 * resize a buffer up or down. 4219 * 4220 * Note that this code is tricky, and has many complications to resolve 4221 * deadlock or inconsistent data situations. Tread lightly!!! 4222 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 4223 * the caller. Calling this code willy nilly can result in the loss of data. 4224 * 4225 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 4226 * B_CACHE for the non-VMIO case. 4227 */ 4228 int 4229 allocbuf(struct buf *bp, int size) 4230 { 4231 int newbsize; 4232 4233 if (bp->b_bcount == size) 4234 return (1); 4235 4236 if (bp->b_kvasize != 0 && bp->b_kvasize < size) 4237 panic("allocbuf: buffer too small"); 4238 4239 newbsize = roundup2(size, DEV_BSIZE); 4240 if ((bp->b_flags & B_VMIO) == 0) { 4241 if ((bp->b_flags & B_MALLOC) == 0) 4242 newbsize = round_page(newbsize); 4243 /* 4244 * Just get anonymous memory from the kernel. Don't 4245 * mess with B_CACHE. 4246 */ 4247 if (newbsize < bp->b_bufsize) 4248 vfs_nonvmio_truncate(bp, newbsize); 4249 else if (newbsize > bp->b_bufsize) 4250 vfs_nonvmio_extend(bp, newbsize); 4251 } else { 4252 int desiredpages; 4253 4254 desiredpages = (size == 0) ? 0 : 4255 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 4256 4257 if (bp->b_flags & B_MALLOC) 4258 panic("allocbuf: VMIO buffer can't be malloced"); 4259 /* 4260 * Set B_CACHE initially if buffer is 0 length or will become 4261 * 0-length. 4262 */ 4263 if (size == 0 || bp->b_bufsize == 0) 4264 bp->b_flags |= B_CACHE; 4265 4266 if (newbsize < bp->b_bufsize) 4267 vfs_vmio_truncate(bp, desiredpages); 4268 /* XXX This looks as if it should be newbsize > b_bufsize */ 4269 else if (size > bp->b_bcount) 4270 vfs_vmio_extend(bp, desiredpages, size); 4271 bufspace_adjust(bp, newbsize); 4272 } 4273 bp->b_bcount = size; /* requested buffer size. */ 4274 return (1); 4275 } 4276 4277 extern int inflight_transient_maps; 4278 4279 static struct bio_queue nondump_bios; 4280 4281 void 4282 biodone(struct bio *bp) 4283 { 4284 struct mtx *mtxp; 4285 void (*done)(struct bio *); 4286 vm_offset_t start, end; 4287 4288 biotrack(bp, __func__); 4289 4290 /* 4291 * Avoid completing I/O when dumping after a panic since that may 4292 * result in a deadlock in the filesystem or pager code. Note that 4293 * this doesn't affect dumps that were started manually since we aim 4294 * to keep the system usable after it has been resumed. 4295 */ 4296 if (__predict_false(dumping && SCHEDULER_STOPPED())) { 4297 TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue); 4298 return; 4299 } 4300 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 4301 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 4302 bp->bio_flags |= BIO_UNMAPPED; 4303 start = trunc_page((vm_offset_t)bp->bio_data); 4304 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 4305 bp->bio_data = unmapped_buf; 4306 pmap_qremove(start, atop(end - start)); 4307 vmem_free(transient_arena, start, end - start); 4308 atomic_add_int(&inflight_transient_maps, -1); 4309 } 4310 done = bp->bio_done; 4311 if (done == NULL) { 4312 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4313 mtx_lock(mtxp); 4314 bp->bio_flags |= BIO_DONE; 4315 wakeup(bp); 4316 mtx_unlock(mtxp); 4317 } else 4318 done(bp); 4319 } 4320 4321 /* 4322 * Wait for a BIO to finish. 4323 */ 4324 int 4325 biowait(struct bio *bp, const char *wchan) 4326 { 4327 struct mtx *mtxp; 4328 4329 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4330 mtx_lock(mtxp); 4331 while ((bp->bio_flags & BIO_DONE) == 0) 4332 msleep(bp, mtxp, PRIBIO, wchan, 0); 4333 mtx_unlock(mtxp); 4334 if (bp->bio_error != 0) 4335 return (bp->bio_error); 4336 if (!(bp->bio_flags & BIO_ERROR)) 4337 return (0); 4338 return (EIO); 4339 } 4340 4341 void 4342 biofinish(struct bio *bp, struct devstat *stat, int error) 4343 { 4344 4345 if (error) { 4346 bp->bio_error = error; 4347 bp->bio_flags |= BIO_ERROR; 4348 } 4349 if (stat != NULL) 4350 devstat_end_transaction_bio(stat, bp); 4351 biodone(bp); 4352 } 4353 4354 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING) 4355 void 4356 biotrack_buf(struct bio *bp, const char *location) 4357 { 4358 4359 buf_track(bp->bio_track_bp, location); 4360 } 4361 #endif 4362 4363 /* 4364 * bufwait: 4365 * 4366 * Wait for buffer I/O completion, returning error status. The buffer 4367 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 4368 * error and cleared. 4369 */ 4370 int 4371 bufwait(struct buf *bp) 4372 { 4373 if (bp->b_iocmd == BIO_READ) 4374 bwait(bp, PRIBIO, "biord"); 4375 else 4376 bwait(bp, PRIBIO, "biowr"); 4377 if (bp->b_flags & B_EINTR) { 4378 bp->b_flags &= ~B_EINTR; 4379 return (EINTR); 4380 } 4381 if (bp->b_ioflags & BIO_ERROR) { 4382 return (bp->b_error ? bp->b_error : EIO); 4383 } else { 4384 return (0); 4385 } 4386 } 4387 4388 /* 4389 * bufdone: 4390 * 4391 * Finish I/O on a buffer, optionally calling a completion function. 4392 * This is usually called from an interrupt so process blocking is 4393 * not allowed. 4394 * 4395 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 4396 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 4397 * assuming B_INVAL is clear. 4398 * 4399 * For the VMIO case, we set B_CACHE if the op was a read and no 4400 * read error occurred, or if the op was a write. B_CACHE is never 4401 * set if the buffer is invalid or otherwise uncacheable. 4402 * 4403 * bufdone does not mess with B_INVAL, allowing the I/O routine or the 4404 * initiator to leave B_INVAL set to brelse the buffer out of existence 4405 * in the biodone routine. 4406 */ 4407 void 4408 bufdone(struct buf *bp) 4409 { 4410 struct bufobj *dropobj; 4411 void (*biodone)(struct buf *); 4412 4413 buf_track(bp, __func__); 4414 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 4415 dropobj = NULL; 4416 4417 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 4418 4419 runningbufwakeup(bp); 4420 if (bp->b_iocmd == BIO_WRITE) 4421 dropobj = bp->b_bufobj; 4422 /* call optional completion function if requested */ 4423 if (bp->b_iodone != NULL) { 4424 biodone = bp->b_iodone; 4425 bp->b_iodone = NULL; 4426 (*biodone) (bp); 4427 if (dropobj) 4428 bufobj_wdrop(dropobj); 4429 return; 4430 } 4431 if (bp->b_flags & B_VMIO) { 4432 /* 4433 * Set B_CACHE if the op was a normal read and no error 4434 * occurred. B_CACHE is set for writes in the b*write() 4435 * routines. 4436 */ 4437 if (bp->b_iocmd == BIO_READ && 4438 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 4439 !(bp->b_ioflags & BIO_ERROR)) 4440 bp->b_flags |= B_CACHE; 4441 vfs_vmio_iodone(bp); 4442 } 4443 if (!LIST_EMPTY(&bp->b_dep)) 4444 buf_complete(bp); 4445 if ((bp->b_flags & B_CKHASH) != 0) { 4446 KASSERT(bp->b_iocmd == BIO_READ, 4447 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd)); 4448 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp)); 4449 (*bp->b_ckhashcalc)(bp); 4450 } 4451 /* 4452 * For asynchronous completions, release the buffer now. The brelse 4453 * will do a wakeup there if necessary - so no need to do a wakeup 4454 * here in the async case. The sync case always needs to do a wakeup. 4455 */ 4456 if (bp->b_flags & B_ASYNC) { 4457 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || 4458 (bp->b_ioflags & BIO_ERROR)) 4459 brelse(bp); 4460 else 4461 bqrelse(bp); 4462 } else 4463 bdone(bp); 4464 if (dropobj) 4465 bufobj_wdrop(dropobj); 4466 } 4467 4468 /* 4469 * This routine is called in lieu of iodone in the case of 4470 * incomplete I/O. This keeps the busy status for pages 4471 * consistent. 4472 */ 4473 void 4474 vfs_unbusy_pages(struct buf *bp) 4475 { 4476 int i; 4477 vm_object_t obj; 4478 vm_page_t m; 4479 bool bogus; 4480 4481 runningbufwakeup(bp); 4482 if (!(bp->b_flags & B_VMIO)) 4483 return; 4484 4485 obj = bp->b_bufobj->bo_object; 4486 bogus = false; 4487 for (i = 0; i < bp->b_npages; i++) { 4488 m = bp->b_pages[i]; 4489 if (m == bogus_page) { 4490 if (bogus == false) { 4491 bogus = true; 4492 VM_OBJECT_RLOCK(obj); 4493 } 4494 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 4495 if (!m) 4496 panic("vfs_unbusy_pages: page missing\n"); 4497 bp->b_pages[i] = m; 4498 if (buf_mapped(bp)) { 4499 BUF_CHECK_MAPPED(bp); 4500 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4501 bp->b_pages, bp->b_npages); 4502 } else 4503 BUF_CHECK_UNMAPPED(bp); 4504 } 4505 vm_page_sunbusy(m); 4506 } 4507 if (bogus) 4508 VM_OBJECT_RUNLOCK(obj); 4509 vm_object_pip_wakeupn(obj, bp->b_npages); 4510 } 4511 4512 /* 4513 * vfs_page_set_valid: 4514 * 4515 * Set the valid bits in a page based on the supplied offset. The 4516 * range is restricted to the buffer's size. 4517 * 4518 * This routine is typically called after a read completes. 4519 */ 4520 static void 4521 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4522 { 4523 vm_ooffset_t eoff; 4524 4525 /* 4526 * Compute the end offset, eoff, such that [off, eoff) does not span a 4527 * page boundary and eoff is not greater than the end of the buffer. 4528 * The end of the buffer, in this case, is our file EOF, not the 4529 * allocation size of the buffer. 4530 */ 4531 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 4532 if (eoff > bp->b_offset + bp->b_bcount) 4533 eoff = bp->b_offset + bp->b_bcount; 4534 4535 /* 4536 * Set valid range. This is typically the entire buffer and thus the 4537 * entire page. 4538 */ 4539 if (eoff > off) 4540 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 4541 } 4542 4543 /* 4544 * vfs_page_set_validclean: 4545 * 4546 * Set the valid bits and clear the dirty bits in a page based on the 4547 * supplied offset. The range is restricted to the buffer's size. 4548 */ 4549 static void 4550 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4551 { 4552 vm_ooffset_t soff, eoff; 4553 4554 /* 4555 * Start and end offsets in buffer. eoff - soff may not cross a 4556 * page boundary or cross the end of the buffer. The end of the 4557 * buffer, in this case, is our file EOF, not the allocation size 4558 * of the buffer. 4559 */ 4560 soff = off; 4561 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4562 if (eoff > bp->b_offset + bp->b_bcount) 4563 eoff = bp->b_offset + bp->b_bcount; 4564 4565 /* 4566 * Set valid range. This is typically the entire buffer and thus the 4567 * entire page. 4568 */ 4569 if (eoff > soff) { 4570 vm_page_set_validclean( 4571 m, 4572 (vm_offset_t) (soff & PAGE_MASK), 4573 (vm_offset_t) (eoff - soff) 4574 ); 4575 } 4576 } 4577 4578 /* 4579 * Acquire a shared busy on all pages in the buf. 4580 */ 4581 void 4582 vfs_busy_pages_acquire(struct buf *bp) 4583 { 4584 int i; 4585 4586 for (i = 0; i < bp->b_npages; i++) 4587 vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY); 4588 } 4589 4590 void 4591 vfs_busy_pages_release(struct buf *bp) 4592 { 4593 int i; 4594 4595 for (i = 0; i < bp->b_npages; i++) 4596 vm_page_sunbusy(bp->b_pages[i]); 4597 } 4598 4599 /* 4600 * This routine is called before a device strategy routine. 4601 * It is used to tell the VM system that paging I/O is in 4602 * progress, and treat the pages associated with the buffer 4603 * almost as being exclusive busy. Also the object paging_in_progress 4604 * flag is handled to make sure that the object doesn't become 4605 * inconsistent. 4606 * 4607 * Since I/O has not been initiated yet, certain buffer flags 4608 * such as BIO_ERROR or B_INVAL may be in an inconsistent state 4609 * and should be ignored. 4610 */ 4611 void 4612 vfs_busy_pages(struct buf *bp, int clear_modify) 4613 { 4614 vm_object_t obj; 4615 vm_ooffset_t foff; 4616 vm_page_t m; 4617 int i; 4618 bool bogus; 4619 4620 if (!(bp->b_flags & B_VMIO)) 4621 return; 4622 4623 obj = bp->b_bufobj->bo_object; 4624 foff = bp->b_offset; 4625 KASSERT(bp->b_offset != NOOFFSET, 4626 ("vfs_busy_pages: no buffer offset")); 4627 if ((bp->b_flags & B_CLUSTER) == 0) { 4628 vm_object_pip_add(obj, bp->b_npages); 4629 vfs_busy_pages_acquire(bp); 4630 } 4631 if (bp->b_bufsize != 0) 4632 vfs_setdirty_range(bp); 4633 bogus = false; 4634 for (i = 0; i < bp->b_npages; i++) { 4635 m = bp->b_pages[i]; 4636 vm_page_assert_sbusied(m); 4637 4638 /* 4639 * When readying a buffer for a read ( i.e 4640 * clear_modify == 0 ), it is important to do 4641 * bogus_page replacement for valid pages in 4642 * partially instantiated buffers. Partially 4643 * instantiated buffers can, in turn, occur when 4644 * reconstituting a buffer from its VM backing store 4645 * base. We only have to do this if B_CACHE is 4646 * clear ( which causes the I/O to occur in the 4647 * first place ). The replacement prevents the read 4648 * I/O from overwriting potentially dirty VM-backed 4649 * pages. XXX bogus page replacement is, uh, bogus. 4650 * It may not work properly with small-block devices. 4651 * We need to find a better way. 4652 */ 4653 if (clear_modify) { 4654 pmap_remove_write(m); 4655 vfs_page_set_validclean(bp, foff, m); 4656 } else if (vm_page_all_valid(m) && 4657 (bp->b_flags & B_CACHE) == 0) { 4658 bp->b_pages[i] = bogus_page; 4659 bogus = true; 4660 } 4661 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4662 } 4663 if (bogus && buf_mapped(bp)) { 4664 BUF_CHECK_MAPPED(bp); 4665 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4666 bp->b_pages, bp->b_npages); 4667 } 4668 } 4669 4670 /* 4671 * vfs_bio_set_valid: 4672 * 4673 * Set the range within the buffer to valid. The range is 4674 * relative to the beginning of the buffer, b_offset. Note that 4675 * b_offset itself may be offset from the beginning of the first 4676 * page. 4677 */ 4678 void 4679 vfs_bio_set_valid(struct buf *bp, int base, int size) 4680 { 4681 int i, n; 4682 vm_page_t m; 4683 4684 if (!(bp->b_flags & B_VMIO)) 4685 return; 4686 4687 /* 4688 * Fixup base to be relative to beginning of first page. 4689 * Set initial n to be the maximum number of bytes in the 4690 * first page that can be validated. 4691 */ 4692 base += (bp->b_offset & PAGE_MASK); 4693 n = PAGE_SIZE - (base & PAGE_MASK); 4694 4695 /* 4696 * Busy may not be strictly necessary here because the pages are 4697 * unlikely to be fully valid and the vnode lock will synchronize 4698 * their access via getpages. It is grabbed for consistency with 4699 * other page validation. 4700 */ 4701 vfs_busy_pages_acquire(bp); 4702 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4703 m = bp->b_pages[i]; 4704 if (n > size) 4705 n = size; 4706 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4707 base += n; 4708 size -= n; 4709 n = PAGE_SIZE; 4710 } 4711 vfs_busy_pages_release(bp); 4712 } 4713 4714 /* 4715 * vfs_bio_clrbuf: 4716 * 4717 * If the specified buffer is a non-VMIO buffer, clear the entire 4718 * buffer. If the specified buffer is a VMIO buffer, clear and 4719 * validate only the previously invalid portions of the buffer. 4720 * This routine essentially fakes an I/O, so we need to clear 4721 * BIO_ERROR and B_INVAL. 4722 * 4723 * Note that while we only theoretically need to clear through b_bcount, 4724 * we go ahead and clear through b_bufsize. 4725 */ 4726 void 4727 vfs_bio_clrbuf(struct buf *bp) 4728 { 4729 int i, j, mask, sa, ea, slide; 4730 4731 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4732 clrbuf(bp); 4733 return; 4734 } 4735 bp->b_flags &= ~B_INVAL; 4736 bp->b_ioflags &= ~BIO_ERROR; 4737 vfs_busy_pages_acquire(bp); 4738 sa = bp->b_offset & PAGE_MASK; 4739 slide = 0; 4740 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4741 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4742 ea = slide & PAGE_MASK; 4743 if (ea == 0) 4744 ea = PAGE_SIZE; 4745 if (bp->b_pages[i] == bogus_page) 4746 continue; 4747 j = sa / DEV_BSIZE; 4748 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4749 if ((bp->b_pages[i]->valid & mask) == mask) 4750 continue; 4751 if ((bp->b_pages[i]->valid & mask) == 0) 4752 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4753 else { 4754 for (; sa < ea; sa += DEV_BSIZE, j++) { 4755 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4756 pmap_zero_page_area(bp->b_pages[i], 4757 sa, DEV_BSIZE); 4758 } 4759 } 4760 } 4761 vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE, 4762 roundup2(ea - sa, DEV_BSIZE)); 4763 } 4764 vfs_busy_pages_release(bp); 4765 bp->b_resid = 0; 4766 } 4767 4768 void 4769 vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4770 { 4771 vm_page_t m; 4772 int i, n; 4773 4774 if (buf_mapped(bp)) { 4775 BUF_CHECK_MAPPED(bp); 4776 bzero(bp->b_data + base, size); 4777 } else { 4778 BUF_CHECK_UNMAPPED(bp); 4779 n = PAGE_SIZE - (base & PAGE_MASK); 4780 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4781 m = bp->b_pages[i]; 4782 if (n > size) 4783 n = size; 4784 pmap_zero_page_area(m, base & PAGE_MASK, n); 4785 base += n; 4786 size -= n; 4787 n = PAGE_SIZE; 4788 } 4789 } 4790 } 4791 4792 /* 4793 * Update buffer flags based on I/O request parameters, optionally releasing the 4794 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM, 4795 * where they may be placed on a page queue (VMIO) or freed immediately (direct 4796 * I/O). Otherwise the buffer is released to the cache. 4797 */ 4798 static void 4799 b_io_dismiss(struct buf *bp, int ioflag, bool release) 4800 { 4801 4802 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0, 4803 ("buf %p non-VMIO noreuse", bp)); 4804 4805 if ((ioflag & IO_DIRECT) != 0) 4806 bp->b_flags |= B_DIRECT; 4807 if ((ioflag & IO_EXT) != 0) 4808 bp->b_xflags |= BX_ALTDATA; 4809 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) { 4810 bp->b_flags |= B_RELBUF; 4811 if ((ioflag & IO_NOREUSE) != 0) 4812 bp->b_flags |= B_NOREUSE; 4813 if (release) 4814 brelse(bp); 4815 } else if (release) 4816 bqrelse(bp); 4817 } 4818 4819 void 4820 vfs_bio_brelse(struct buf *bp, int ioflag) 4821 { 4822 4823 b_io_dismiss(bp, ioflag, true); 4824 } 4825 4826 void 4827 vfs_bio_set_flags(struct buf *bp, int ioflag) 4828 { 4829 4830 b_io_dismiss(bp, ioflag, false); 4831 } 4832 4833 /* 4834 * vm_hold_load_pages and vm_hold_free_pages get pages into 4835 * a buffers address space. The pages are anonymous and are 4836 * not associated with a file object. 4837 */ 4838 static void 4839 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4840 { 4841 vm_offset_t pg; 4842 vm_page_t p; 4843 int index; 4844 4845 BUF_CHECK_MAPPED(bp); 4846 4847 to = round_page(to); 4848 from = round_page(from); 4849 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4850 4851 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4852 /* 4853 * note: must allocate system pages since blocking here 4854 * could interfere with paging I/O, no matter which 4855 * process we are. 4856 */ 4857 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4858 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | 4859 VM_ALLOC_WAITOK); 4860 pmap_qenter(pg, &p, 1); 4861 bp->b_pages[index] = p; 4862 } 4863 bp->b_npages = index; 4864 } 4865 4866 /* Return pages associated with this buf to the vm system */ 4867 static void 4868 vm_hold_free_pages(struct buf *bp, int newbsize) 4869 { 4870 vm_offset_t from; 4871 vm_page_t p; 4872 int index, newnpages; 4873 4874 BUF_CHECK_MAPPED(bp); 4875 4876 from = round_page((vm_offset_t)bp->b_data + newbsize); 4877 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4878 if (bp->b_npages > newnpages) 4879 pmap_qremove(from, bp->b_npages - newnpages); 4880 for (index = newnpages; index < bp->b_npages; index++) { 4881 p = bp->b_pages[index]; 4882 bp->b_pages[index] = NULL; 4883 vm_page_unwire_noq(p); 4884 vm_page_free(p); 4885 } 4886 bp->b_npages = newnpages; 4887 } 4888 4889 /* 4890 * Map an IO request into kernel virtual address space. 4891 * 4892 * All requests are (re)mapped into kernel VA space. 4893 * Notice that we use b_bufsize for the size of the buffer 4894 * to be mapped. b_bcount might be modified by the driver. 4895 * 4896 * Note that even if the caller determines that the address space should 4897 * be valid, a race or a smaller-file mapped into a larger space may 4898 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4899 * check the return value. 4900 * 4901 * This function only works with pager buffers. 4902 */ 4903 int 4904 vmapbuf(struct buf *bp, int mapbuf) 4905 { 4906 vm_prot_t prot; 4907 int pidx; 4908 4909 if (bp->b_bufsize < 0) 4910 return (-1); 4911 prot = VM_PROT_READ; 4912 if (bp->b_iocmd == BIO_READ) 4913 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4914 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4915 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4916 btoc(MAXPHYS))) < 0) 4917 return (-1); 4918 bp->b_npages = pidx; 4919 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4920 if (mapbuf || !unmapped_buf_allowed) { 4921 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx); 4922 bp->b_data = bp->b_kvabase + bp->b_offset; 4923 } else 4924 bp->b_data = unmapped_buf; 4925 return(0); 4926 } 4927 4928 /* 4929 * Free the io map PTEs associated with this IO operation. 4930 * We also invalidate the TLB entries and restore the original b_addr. 4931 * 4932 * This function only works with pager buffers. 4933 */ 4934 void 4935 vunmapbuf(struct buf *bp) 4936 { 4937 int npages; 4938 4939 npages = bp->b_npages; 4940 if (buf_mapped(bp)) 4941 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4942 vm_page_unhold_pages(bp->b_pages, npages); 4943 4944 bp->b_data = unmapped_buf; 4945 } 4946 4947 void 4948 bdone(struct buf *bp) 4949 { 4950 struct mtx *mtxp; 4951 4952 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4953 mtx_lock(mtxp); 4954 bp->b_flags |= B_DONE; 4955 wakeup(bp); 4956 mtx_unlock(mtxp); 4957 } 4958 4959 void 4960 bwait(struct buf *bp, u_char pri, const char *wchan) 4961 { 4962 struct mtx *mtxp; 4963 4964 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4965 mtx_lock(mtxp); 4966 while ((bp->b_flags & B_DONE) == 0) 4967 msleep(bp, mtxp, pri, wchan, 0); 4968 mtx_unlock(mtxp); 4969 } 4970 4971 int 4972 bufsync(struct bufobj *bo, int waitfor) 4973 { 4974 4975 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread)); 4976 } 4977 4978 void 4979 bufstrategy(struct bufobj *bo, struct buf *bp) 4980 { 4981 int i __unused; 4982 struct vnode *vp; 4983 4984 vp = bp->b_vp; 4985 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4986 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4987 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4988 i = VOP_STRATEGY(vp, bp); 4989 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4990 } 4991 4992 /* 4993 * Initialize a struct bufobj before use. Memory is assumed zero filled. 4994 */ 4995 void 4996 bufobj_init(struct bufobj *bo, void *private) 4997 { 4998 static volatile int bufobj_cleanq; 4999 5000 bo->bo_domain = 5001 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains; 5002 rw_init(BO_LOCKPTR(bo), "bufobj interlock"); 5003 bo->bo_private = private; 5004 TAILQ_INIT(&bo->bo_clean.bv_hd); 5005 TAILQ_INIT(&bo->bo_dirty.bv_hd); 5006 } 5007 5008 void 5009 bufobj_wrefl(struct bufobj *bo) 5010 { 5011 5012 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 5013 ASSERT_BO_WLOCKED(bo); 5014 bo->bo_numoutput++; 5015 } 5016 5017 void 5018 bufobj_wref(struct bufobj *bo) 5019 { 5020 5021 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 5022 BO_LOCK(bo); 5023 bo->bo_numoutput++; 5024 BO_UNLOCK(bo); 5025 } 5026 5027 void 5028 bufobj_wdrop(struct bufobj *bo) 5029 { 5030 5031 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 5032 BO_LOCK(bo); 5033 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 5034 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 5035 bo->bo_flag &= ~BO_WWAIT; 5036 wakeup(&bo->bo_numoutput); 5037 } 5038 BO_UNLOCK(bo); 5039 } 5040 5041 int 5042 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 5043 { 5044 int error; 5045 5046 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 5047 ASSERT_BO_WLOCKED(bo); 5048 error = 0; 5049 while (bo->bo_numoutput) { 5050 bo->bo_flag |= BO_WWAIT; 5051 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 5052 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 5053 if (error) 5054 break; 5055 } 5056 return (error); 5057 } 5058 5059 /* 5060 * Set bio_data or bio_ma for struct bio from the struct buf. 5061 */ 5062 void 5063 bdata2bio(struct buf *bp, struct bio *bip) 5064 { 5065 5066 if (!buf_mapped(bp)) { 5067 KASSERT(unmapped_buf_allowed, ("unmapped")); 5068 bip->bio_ma = bp->b_pages; 5069 bip->bio_ma_n = bp->b_npages; 5070 bip->bio_data = unmapped_buf; 5071 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 5072 bip->bio_flags |= BIO_UNMAPPED; 5073 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 5074 PAGE_SIZE == bp->b_npages, 5075 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 5076 (long long)bip->bio_length, bip->bio_ma_n)); 5077 } else { 5078 bip->bio_data = bp->b_data; 5079 bip->bio_ma = NULL; 5080 } 5081 } 5082 5083 /* 5084 * The MIPS pmap code currently doesn't handle aliased pages. 5085 * The VIPT caches may not handle page aliasing themselves, leading 5086 * to data corruption. 5087 * 5088 * As such, this code makes a system extremely unhappy if said 5089 * system doesn't support unaliasing the above situation in hardware. 5090 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable 5091 * this feature at build time, so it has to be handled in software. 5092 * 5093 * Once the MIPS pmap/cache code grows to support this function on 5094 * earlier chips, it should be flipped back off. 5095 */ 5096 #ifdef __mips__ 5097 static int buf_pager_relbuf = 1; 5098 #else 5099 static int buf_pager_relbuf = 0; 5100 #endif 5101 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN, 5102 &buf_pager_relbuf, 0, 5103 "Make buffer pager release buffers after reading"); 5104 5105 /* 5106 * The buffer pager. It uses buffer reads to validate pages. 5107 * 5108 * In contrast to the generic local pager from vm/vnode_pager.c, this 5109 * pager correctly and easily handles volumes where the underlying 5110 * device block size is greater than the machine page size. The 5111 * buffer cache transparently extends the requested page run to be 5112 * aligned at the block boundary, and does the necessary bogus page 5113 * replacements in the addends to avoid obliterating already valid 5114 * pages. 5115 * 5116 * The only non-trivial issue is that the exclusive busy state for 5117 * pages, which is assumed by the vm_pager_getpages() interface, is 5118 * incompatible with the VMIO buffer cache's desire to share-busy the 5119 * pages. This function performs a trivial downgrade of the pages' 5120 * state before reading buffers, and a less trivial upgrade from the 5121 * shared-busy to excl-busy state after the read. 5122 */ 5123 int 5124 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count, 5125 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno, 5126 vbg_get_blksize_t get_blksize) 5127 { 5128 vm_page_t m; 5129 vm_object_t object; 5130 struct buf *bp; 5131 struct mount *mp; 5132 daddr_t lbn, lbnp; 5133 vm_ooffset_t la, lb, poff, poffe; 5134 long bsize; 5135 int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b; 5136 bool redo, lpart; 5137 5138 object = vp->v_object; 5139 mp = vp->v_mount; 5140 error = 0; 5141 la = IDX_TO_OFF(ma[count - 1]->pindex); 5142 if (la >= object->un_pager.vnp.vnp_size) 5143 return (VM_PAGER_BAD); 5144 5145 /* 5146 * Change the meaning of la from where the last requested page starts 5147 * to where it ends, because that's the end of the requested region 5148 * and the start of the potential read-ahead region. 5149 */ 5150 la += PAGE_SIZE; 5151 lpart = la > object->un_pager.vnp.vnp_size; 5152 bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex))); 5153 5154 /* 5155 * Calculate read-ahead, behind and total pages. 5156 */ 5157 pgsin = count; 5158 lb = IDX_TO_OFF(ma[0]->pindex); 5159 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs)); 5160 pgsin += pgsin_b; 5161 if (rbehind != NULL) 5162 *rbehind = pgsin_b; 5163 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la); 5164 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size) 5165 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size, 5166 PAGE_SIZE) - la); 5167 pgsin += pgsin_a; 5168 if (rahead != NULL) 5169 *rahead = pgsin_a; 5170 VM_CNT_INC(v_vnodein); 5171 VM_CNT_ADD(v_vnodepgsin, pgsin); 5172 5173 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) 5174 != 0) ? GB_UNMAPPED : 0; 5175 again: 5176 for (i = 0; i < count; i++) 5177 vm_page_busy_downgrade(ma[i]); 5178 5179 lbnp = -1; 5180 for (i = 0; i < count; i++) { 5181 m = ma[i]; 5182 5183 /* 5184 * Pages are shared busy and the object lock is not 5185 * owned, which together allow for the pages' 5186 * invalidation. The racy test for validity avoids 5187 * useless creation of the buffer for the most typical 5188 * case when invalidation is not used in redo or for 5189 * parallel read. The shared->excl upgrade loop at 5190 * the end of the function catches the race in a 5191 * reliable way (protected by the object lock). 5192 */ 5193 if (vm_page_all_valid(m)) 5194 continue; 5195 5196 poff = IDX_TO_OFF(m->pindex); 5197 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size); 5198 for (; poff < poffe; poff += bsize) { 5199 lbn = get_lblkno(vp, poff); 5200 if (lbn == lbnp) 5201 goto next_page; 5202 lbnp = lbn; 5203 5204 bsize = get_blksize(vp, lbn); 5205 error = bread_gb(vp, lbn, bsize, curthread->td_ucred, 5206 br_flags, &bp); 5207 if (error != 0) 5208 goto end_pages; 5209 if (LIST_EMPTY(&bp->b_dep)) { 5210 /* 5211 * Invalidation clears m->valid, but 5212 * may leave B_CACHE flag if the 5213 * buffer existed at the invalidation 5214 * time. In this case, recycle the 5215 * buffer to do real read on next 5216 * bread() after redo. 5217 * 5218 * Otherwise B_RELBUF is not strictly 5219 * necessary, enable to reduce buf 5220 * cache pressure. 5221 */ 5222 if (buf_pager_relbuf || 5223 !vm_page_all_valid(m)) 5224 bp->b_flags |= B_RELBUF; 5225 5226 bp->b_flags &= ~B_NOCACHE; 5227 brelse(bp); 5228 } else { 5229 bqrelse(bp); 5230 } 5231 } 5232 KASSERT(1 /* racy, enable for debugging */ || 5233 vm_page_all_valid(m) || i == count - 1, 5234 ("buf %d %p invalid", i, m)); 5235 if (i == count - 1 && lpart) { 5236 if (!vm_page_none_valid(m) && 5237 !vm_page_all_valid(m)) 5238 vm_page_zero_invalid(m, TRUE); 5239 } 5240 next_page:; 5241 } 5242 end_pages: 5243 5244 VM_OBJECT_WLOCK(object); 5245 redo = false; 5246 for (i = 0; i < count; i++) { 5247 vm_page_sunbusy(ma[i]); 5248 ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL); 5249 5250 /* 5251 * Since the pages were only sbusy while neither the 5252 * buffer nor the object lock was held by us, or 5253 * reallocated while vm_page_grab() slept for busy 5254 * relinguish, they could have been invalidated. 5255 * Recheck the valid bits and re-read as needed. 5256 * 5257 * Note that the last page is made fully valid in the 5258 * read loop, and partial validity for the page at 5259 * index count - 1 could mean that the page was 5260 * invalidated or removed, so we must restart for 5261 * safety as well. 5262 */ 5263 if (!vm_page_all_valid(ma[i])) 5264 redo = true; 5265 } 5266 VM_OBJECT_WUNLOCK(object); 5267 if (redo && error == 0) 5268 goto again; 5269 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); 5270 } 5271 5272 #include "opt_ddb.h" 5273 #ifdef DDB 5274 #include <ddb/ddb.h> 5275 5276 /* DDB command to show buffer data */ 5277 DB_SHOW_COMMAND(buffer, db_show_buffer) 5278 { 5279 /* get args */ 5280 struct buf *bp = (struct buf *)addr; 5281 #ifdef FULL_BUF_TRACKING 5282 uint32_t i, j; 5283 #endif 5284 5285 if (!have_addr) { 5286 db_printf("usage: show buffer <addr>\n"); 5287 return; 5288 } 5289 5290 db_printf("buf at %p\n", bp); 5291 db_printf("b_flags = 0x%b, b_xflags=0x%b\n", 5292 (u_int)bp->b_flags, PRINT_BUF_FLAGS, 5293 (u_int)bp->b_xflags, PRINT_BUF_XFLAGS); 5294 db_printf("b_vflags=0x%b b_ioflags0x%b\n", 5295 (u_int)bp->b_vflags, PRINT_BUF_VFLAGS, 5296 (u_int)bp->b_ioflags, PRINT_BIO_FLAGS); 5297 db_printf( 5298 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 5299 "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, " 5300 "b_vp = %p, b_dep = %p\n", 5301 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 5302 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 5303 (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first); 5304 db_printf("b_kvabase = %p, b_kvasize = %d\n", 5305 bp->b_kvabase, bp->b_kvasize); 5306 if (bp->b_npages) { 5307 int i; 5308 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 5309 for (i = 0; i < bp->b_npages; i++) { 5310 vm_page_t m; 5311 m = bp->b_pages[i]; 5312 if (m != NULL) 5313 db_printf("(%p, 0x%lx, 0x%lx)", m->object, 5314 (u_long)m->pindex, 5315 (u_long)VM_PAGE_TO_PHYS(m)); 5316 else 5317 db_printf("( ??? )"); 5318 if ((i + 1) < bp->b_npages) 5319 db_printf(","); 5320 } 5321 db_printf("\n"); 5322 } 5323 BUF_LOCKPRINTINFO(bp); 5324 #if defined(FULL_BUF_TRACKING) 5325 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt); 5326 5327 i = bp->b_io_tcnt % BUF_TRACKING_SIZE; 5328 for (j = 1; j <= BUF_TRACKING_SIZE; j++) { 5329 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL) 5330 continue; 5331 db_printf(" %2u: %s\n", j, 5332 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]); 5333 } 5334 #elif defined(BUF_TRACKING) 5335 db_printf("b_io_tracking: %s\n", bp->b_io_tracking); 5336 #endif 5337 db_printf(" "); 5338 } 5339 5340 DB_SHOW_COMMAND(bufqueues, bufqueues) 5341 { 5342 struct bufdomain *bd; 5343 struct buf *bp; 5344 long total; 5345 int i, j, cnt; 5346 5347 db_printf("bqempty: %d\n", bqempty.bq_len); 5348 5349 for (i = 0; i < buf_domains; i++) { 5350 bd = &bdomain[i]; 5351 db_printf("Buf domain %d\n", i); 5352 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers); 5353 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers); 5354 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers); 5355 db_printf("\n"); 5356 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace); 5357 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace); 5358 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace); 5359 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace); 5360 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh); 5361 db_printf("\n"); 5362 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers); 5363 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers); 5364 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers); 5365 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh); 5366 db_printf("\n"); 5367 total = 0; 5368 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist) 5369 total += bp->b_bufsize; 5370 db_printf("\tcleanq count\t%d (%ld)\n", 5371 bd->bd_cleanq->bq_len, total); 5372 total = 0; 5373 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist) 5374 total += bp->b_bufsize; 5375 db_printf("\tdirtyq count\t%d (%ld)\n", 5376 bd->bd_dirtyq.bq_len, total); 5377 db_printf("\twakeup\t\t%d\n", bd->bd_wanted); 5378 db_printf("\tlim\t\t%d\n", bd->bd_lim); 5379 db_printf("\tCPU "); 5380 for (j = 0; j <= mp_maxid; j++) 5381 db_printf("%d, ", bd->bd_subq[j].bq_len); 5382 db_printf("\n"); 5383 cnt = 0; 5384 total = 0; 5385 for (j = 0; j < nbuf; j++) 5386 if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) { 5387 cnt++; 5388 total += buf[j].b_bufsize; 5389 } 5390 db_printf("\tLocked buffers: %d space %ld\n", cnt, total); 5391 cnt = 0; 5392 total = 0; 5393 for (j = 0; j < nbuf; j++) 5394 if (buf[j].b_domain == i) { 5395 cnt++; 5396 total += buf[j].b_bufsize; 5397 } 5398 db_printf("\tTotal buffers: %d space %ld\n", cnt, total); 5399 } 5400 } 5401 5402 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 5403 { 5404 struct buf *bp; 5405 int i; 5406 5407 for (i = 0; i < nbuf; i++) { 5408 bp = &buf[i]; 5409 if (BUF_ISLOCKED(bp)) { 5410 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5411 db_printf("\n"); 5412 if (db_pager_quit) 5413 break; 5414 } 5415 } 5416 } 5417 5418 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 5419 { 5420 struct vnode *vp; 5421 struct buf *bp; 5422 5423 if (!have_addr) { 5424 db_printf("usage: show vnodebufs <addr>\n"); 5425 return; 5426 } 5427 vp = (struct vnode *)addr; 5428 db_printf("Clean buffers:\n"); 5429 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 5430 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5431 db_printf("\n"); 5432 } 5433 db_printf("Dirty buffers:\n"); 5434 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 5435 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5436 db_printf("\n"); 5437 } 5438 } 5439 5440 DB_COMMAND(countfreebufs, db_coundfreebufs) 5441 { 5442 struct buf *bp; 5443 int i, used = 0, nfree = 0; 5444 5445 if (have_addr) { 5446 db_printf("usage: countfreebufs\n"); 5447 return; 5448 } 5449 5450 for (i = 0; i < nbuf; i++) { 5451 bp = &buf[i]; 5452 if (bp->b_qindex == QUEUE_EMPTY) 5453 nfree++; 5454 else 5455 used++; 5456 } 5457 5458 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 5459 nfree + used); 5460 db_printf("numfreebuffers is %d\n", numfreebuffers); 5461 } 5462 #endif /* DDB */ 5463