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