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