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