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