1 /*- 2 * Copyright (c) 2004 Poul-Henning Kamp 3 * Copyright (c) 1994,1997 John S. Dyson 4 * Copyright (c) 2013 The FreeBSD Foundation 5 * All rights reserved. 6 * 7 * Portions of this software were developed by Konstantin Belousov 8 * under sponsorship from the FreeBSD Foundation. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 */ 31 32 /* 33 * this file contains a new buffer I/O scheme implementing a coherent 34 * VM object and buffer cache scheme. Pains have been taken to make 35 * sure that the performance degradation associated with schemes such 36 * as this is not realized. 37 * 38 * Author: John S. Dyson 39 * Significant help during the development and debugging phases 40 * had been provided by David Greenman, also of the FreeBSD core team. 41 * 42 * see man buf(9) for more info. 43 */ 44 45 #include <sys/cdefs.h> 46 __FBSDID("$FreeBSD$"); 47 48 #include <sys/param.h> 49 #include <sys/systm.h> 50 #include <sys/bio.h> 51 #include <sys/conf.h> 52 #include <sys/buf.h> 53 #include <sys/devicestat.h> 54 #include <sys/eventhandler.h> 55 #include <sys/fail.h> 56 #include <sys/limits.h> 57 #include <sys/lock.h> 58 #include <sys/malloc.h> 59 #include <sys/mount.h> 60 #include <sys/mutex.h> 61 #include <sys/kernel.h> 62 #include <sys/kthread.h> 63 #include <sys/proc.h> 64 #include <sys/resourcevar.h> 65 #include <sys/rwlock.h> 66 #include <sys/sysctl.h> 67 #include <sys/vmem.h> 68 #include <sys/vmmeter.h> 69 #include <sys/vnode.h> 70 #include <geom/geom.h> 71 #include <vm/vm.h> 72 #include <vm/vm_param.h> 73 #include <vm/vm_kern.h> 74 #include <vm/vm_pageout.h> 75 #include <vm/vm_page.h> 76 #include <vm/vm_object.h> 77 #include <vm/vm_extern.h> 78 #include <vm/vm_map.h> 79 #include "opt_compat.h" 80 #include "opt_swap.h" 81 82 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 83 84 struct bio_ops bioops; /* I/O operation notification */ 85 86 struct buf_ops buf_ops_bio = { 87 .bop_name = "buf_ops_bio", 88 .bop_write = bufwrite, 89 .bop_strategy = bufstrategy, 90 .bop_sync = bufsync, 91 .bop_bdflush = bufbdflush, 92 }; 93 94 /* 95 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 96 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 97 */ 98 struct buf *buf; /* buffer header pool */ 99 caddr_t unmapped_buf; 100 101 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */ 102 struct proc *bufdaemonproc; 103 104 static int inmem(struct vnode *vp, daddr_t blkno); 105 static void vm_hold_free_pages(struct buf *bp, int newbsize); 106 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 107 vm_offset_t to); 108 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 109 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 110 vm_page_t m); 111 static void vfs_clean_pages_dirty_buf(struct buf *bp); 112 static void vfs_setdirty_locked_object(struct buf *bp); 113 static void vfs_vmio_release(struct buf *bp); 114 static int vfs_bio_clcheck(struct vnode *vp, int size, 115 daddr_t lblkno, daddr_t blkno); 116 static int buf_flush(struct vnode *vp, int); 117 static int flushbufqueues(struct vnode *, int, int); 118 static void buf_daemon(void); 119 static void bremfreel(struct buf *bp); 120 static __inline void bd_wakeup(void); 121 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS); 122 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 123 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 124 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 125 #endif 126 127 int vmiodirenable = TRUE; 128 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 129 "Use the VM system for directory writes"); 130 long runningbufspace; 131 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 132 "Amount of presently outstanding async buffer io"); 133 static long bufspace; 134 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 135 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 136 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 137 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 138 #else 139 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 140 "Physical memory used for buffers"); 141 #endif 142 static long bufkvaspace; 143 SYSCTL_LONG(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, 0, 144 "Kernel virtual memory used for buffers"); 145 static long maxbufspace; 146 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 147 "Maximum allowed value of bufspace (including buf_daemon)"); 148 static long bufmallocspace; 149 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 150 "Amount of malloced memory for buffers"); 151 static long maxbufmallocspace; 152 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 153 "Maximum amount of malloced memory for buffers"); 154 static long lobufspace; 155 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 156 "Minimum amount of buffers we want to have"); 157 long hibufspace; 158 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 159 "Maximum allowed value of bufspace (excluding buf_daemon)"); 160 static int bufreusecnt; 161 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 162 "Number of times we have reused a buffer"); 163 static int buffreekvacnt; 164 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 165 "Number of times we have freed the KVA space from some buffer"); 166 static int bufdefragcnt; 167 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 168 "Number of times we have had to repeat buffer allocation to defragment"); 169 static long lorunningspace; 170 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 171 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L", 172 "Minimum preferred space used for in-progress I/O"); 173 static long hirunningspace; 174 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 175 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L", 176 "Maximum amount of space to use for in-progress I/O"); 177 int dirtybufferflushes; 178 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 179 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 180 int bdwriteskip; 181 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 182 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 183 int altbufferflushes; 184 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 185 0, "Number of fsync flushes to limit dirty buffers"); 186 static int recursiveflushes; 187 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 188 0, "Number of flushes skipped due to being recursive"); 189 static int numdirtybuffers; 190 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 191 "Number of buffers that are dirty (has unwritten changes) at the moment"); 192 static int lodirtybuffers; 193 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 194 "How many buffers we want to have free before bufdaemon can sleep"); 195 static int hidirtybuffers; 196 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 197 "When the number of dirty buffers is considered severe"); 198 int dirtybufthresh; 199 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 200 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 201 static int numfreebuffers; 202 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 203 "Number of free buffers"); 204 static int lofreebuffers; 205 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 206 "XXX Unused"); 207 static int hifreebuffers; 208 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 209 "XXX Complicatedly unused"); 210 static int getnewbufcalls; 211 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 212 "Number of calls to getnewbuf"); 213 static int getnewbufrestarts; 214 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 215 "Number of times getnewbuf has had to restart a buffer aquisition"); 216 static int mappingrestarts; 217 SYSCTL_INT(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RW, &mappingrestarts, 0, 218 "Number of times getblk has had to restart a buffer mapping for " 219 "unmapped buffer"); 220 static int flushbufqtarget = 100; 221 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 222 "Amount of work to do in flushbufqueues when helping bufdaemon"); 223 static long notbufdflushes; 224 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 0, 225 "Number of dirty buffer flushes done by the bufdaemon helpers"); 226 static long barrierwrites; 227 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0, 228 "Number of barrier writes"); 229 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, 230 &unmapped_buf_allowed, 0, 231 "Permit the use of the unmapped i/o"); 232 233 /* 234 * Lock for the non-dirty bufqueues 235 */ 236 static struct mtx_padalign bqclean; 237 238 /* 239 * Lock for the dirty queue. 240 */ 241 static struct mtx_padalign bqdirty; 242 243 /* 244 * This lock synchronizes access to bd_request. 245 */ 246 static struct mtx_padalign bdlock; 247 248 /* 249 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 250 * waitrunningbufspace(). 251 */ 252 static struct mtx_padalign rbreqlock; 253 254 /* 255 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 256 */ 257 static struct rwlock_padalign nblock; 258 259 /* 260 * Lock that protects bdirtywait. 261 */ 262 static struct mtx_padalign bdirtylock; 263 264 /* 265 * Wakeup point for bufdaemon, as well as indicator of whether it is already 266 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 267 * is idling. 268 */ 269 static int bd_request; 270 271 /* 272 * Request for the buf daemon to write more buffers than is indicated by 273 * lodirtybuf. This may be necessary to push out excess dependencies or 274 * defragment the address space where a simple count of the number of dirty 275 * buffers is insufficient to characterize the demand for flushing them. 276 */ 277 static int bd_speedupreq; 278 279 /* 280 * bogus page -- for I/O to/from partially complete buffers 281 * this is a temporary solution to the problem, but it is not 282 * really that bad. it would be better to split the buffer 283 * for input in the case of buffers partially already in memory, 284 * but the code is intricate enough already. 285 */ 286 vm_page_t bogus_page; 287 288 /* 289 * Synchronization (sleep/wakeup) variable for active buffer space requests. 290 * Set when wait starts, cleared prior to wakeup(). 291 * Used in runningbufwakeup() and waitrunningbufspace(). 292 */ 293 static int runningbufreq; 294 295 /* 296 * Synchronization (sleep/wakeup) variable for buffer requests. 297 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 298 * by and/or. 299 * Used in numdirtywakeup(), bufspacewakeup(), bufcountadd(), bwillwrite(), 300 * getnewbuf(), and getblk(). 301 */ 302 static volatile int needsbuffer; 303 304 /* 305 * Synchronization for bwillwrite() waiters. 306 */ 307 static int bdirtywait; 308 309 /* 310 * Definitions for the buffer free lists. 311 */ 312 #define BUFFER_QUEUES 5 /* number of free buffer queues */ 313 314 #define QUEUE_NONE 0 /* on no queue */ 315 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 316 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 317 #define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 318 #define QUEUE_EMPTY 4 /* empty buffer headers */ 319 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 320 321 /* Queues for free buffers with various properties */ 322 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 323 #ifdef INVARIANTS 324 static int bq_len[BUFFER_QUEUES]; 325 #endif 326 327 /* 328 * Single global constant for BUF_WMESG, to avoid getting multiple references. 329 * buf_wmesg is referred from macros. 330 */ 331 const char *buf_wmesg = BUF_WMESG; 332 333 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 334 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 335 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 336 337 static int 338 sysctl_runningspace(SYSCTL_HANDLER_ARGS) 339 { 340 long value; 341 int error; 342 343 value = *(long *)arg1; 344 error = sysctl_handle_long(oidp, &value, 0, req); 345 if (error != 0 || req->newptr == NULL) 346 return (error); 347 mtx_lock(&rbreqlock); 348 if (arg1 == &hirunningspace) { 349 if (value < lorunningspace) 350 error = EINVAL; 351 else 352 hirunningspace = value; 353 } else { 354 KASSERT(arg1 == &lorunningspace, 355 ("%s: unknown arg1", __func__)); 356 if (value > hirunningspace) 357 error = EINVAL; 358 else 359 lorunningspace = value; 360 } 361 mtx_unlock(&rbreqlock); 362 return (error); 363 } 364 365 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 366 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 367 static int 368 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 369 { 370 long lvalue; 371 int ivalue; 372 373 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 374 return (sysctl_handle_long(oidp, arg1, arg2, req)); 375 lvalue = *(long *)arg1; 376 if (lvalue > INT_MAX) 377 /* On overflow, still write out a long to trigger ENOMEM. */ 378 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 379 ivalue = lvalue; 380 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 381 } 382 #endif 383 384 /* 385 * bqlock: 386 * 387 * Return the appropriate queue lock based on the index. 388 */ 389 static inline struct mtx * 390 bqlock(int qindex) 391 { 392 393 if (qindex == QUEUE_DIRTY) 394 return (struct mtx *)(&bqdirty); 395 return (struct mtx *)(&bqclean); 396 } 397 398 /* 399 * bdirtywakeup: 400 * 401 * Wakeup any bwillwrite() waiters. 402 */ 403 static void 404 bdirtywakeup(void) 405 { 406 mtx_lock(&bdirtylock); 407 if (bdirtywait) { 408 bdirtywait = 0; 409 wakeup(&bdirtywait); 410 } 411 mtx_unlock(&bdirtylock); 412 } 413 414 /* 415 * bdirtysub: 416 * 417 * Decrement the numdirtybuffers count by one and wakeup any 418 * threads blocked in bwillwrite(). 419 */ 420 static void 421 bdirtysub(void) 422 { 423 424 if (atomic_fetchadd_int(&numdirtybuffers, -1) == 425 (lodirtybuffers + hidirtybuffers) / 2) 426 bdirtywakeup(); 427 } 428 429 /* 430 * bdirtyadd: 431 * 432 * Increment the numdirtybuffers count by one and wakeup the buf 433 * daemon if needed. 434 */ 435 static void 436 bdirtyadd(void) 437 { 438 439 /* 440 * Only do the wakeup once as we cross the boundary. The 441 * buf daemon will keep running until the condition clears. 442 */ 443 if (atomic_fetchadd_int(&numdirtybuffers, 1) == 444 (lodirtybuffers + hidirtybuffers) / 2) 445 bd_wakeup(); 446 } 447 448 /* 449 * bufspacewakeup: 450 * 451 * Called when buffer space is potentially available for recovery. 452 * getnewbuf() will block on this flag when it is unable to free 453 * sufficient buffer space. Buffer space becomes recoverable when 454 * bp's get placed back in the queues. 455 */ 456 static __inline void 457 bufspacewakeup(void) 458 { 459 int need_wakeup, on; 460 461 /* 462 * If someone is waiting for bufspace, wake them up. Even 463 * though we may not have freed the kva space yet, the waiting 464 * process will be able to now. 465 */ 466 rw_rlock(&nblock); 467 for (;;) { 468 need_wakeup = 0; 469 on = needsbuffer; 470 if ((on & VFS_BIO_NEED_BUFSPACE) == 0) 471 break; 472 need_wakeup = 1; 473 if (atomic_cmpset_rel_int(&needsbuffer, on, 474 on & ~VFS_BIO_NEED_BUFSPACE)) 475 break; 476 } 477 if (need_wakeup) 478 wakeup(__DEVOLATILE(void *, &needsbuffer)); 479 rw_runlock(&nblock); 480 } 481 482 /* 483 * bufspaceadjust: 484 * 485 * Adjust the reported bufspace for a KVA managed buffer, possibly 486 * waking any waiters. 487 */ 488 static void 489 bufspaceadjust(struct buf *bp, int bufsize) 490 { 491 int diff; 492 493 KASSERT((bp->b_flags & B_MALLOC) == 0, 494 ("bufspaceadjust: malloc buf %p", bp)); 495 diff = bufsize - bp->b_bufsize; 496 if (diff < 0) { 497 atomic_subtract_long(&bufspace, -diff); 498 bufspacewakeup(); 499 } else 500 atomic_add_long(&bufspace, diff); 501 bp->b_bufsize = bufsize; 502 } 503 504 /* 505 * bufmallocadjust: 506 * 507 * Adjust the reported bufspace for a malloc managed buffer, possibly 508 * waking any waiters. 509 */ 510 static void 511 bufmallocadjust(struct buf *bp, int bufsize) 512 { 513 int diff; 514 515 KASSERT((bp->b_flags & B_MALLOC) != 0, 516 ("bufmallocadjust: non-malloc buf %p", bp)); 517 diff = bufsize - bp->b_bufsize; 518 if (diff < 0) { 519 atomic_subtract_long(&bufmallocspace, -diff); 520 bufspacewakeup(); 521 } else 522 atomic_add_long(&bufmallocspace, diff); 523 bp->b_bufsize = bufsize; 524 } 525 526 /* 527 * runningwakeup: 528 * 529 * Wake up processes that are waiting on asynchronous writes to fall 530 * below lorunningspace. 531 */ 532 static void 533 runningwakeup(void) 534 { 535 536 mtx_lock(&rbreqlock); 537 if (runningbufreq) { 538 runningbufreq = 0; 539 wakeup(&runningbufreq); 540 } 541 mtx_unlock(&rbreqlock); 542 } 543 544 /* 545 * runningbufwakeup: 546 * 547 * Decrement the outstanding write count according. 548 */ 549 void 550 runningbufwakeup(struct buf *bp) 551 { 552 long space, bspace; 553 554 bspace = bp->b_runningbufspace; 555 if (bspace == 0) 556 return; 557 space = atomic_fetchadd_long(&runningbufspace, -bspace); 558 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", 559 space, bspace)); 560 bp->b_runningbufspace = 0; 561 /* 562 * Only acquire the lock and wakeup on the transition from exceeding 563 * the threshold to falling below it. 564 */ 565 if (space < lorunningspace) 566 return; 567 if (space - bspace > lorunningspace) 568 return; 569 runningwakeup(); 570 } 571 572 /* 573 * bufcountadd: 574 * 575 * Called when a buffer has been added to one of the free queues to 576 * account for the buffer and to wakeup anyone waiting for free buffers. 577 * This typically occurs when large amounts of metadata are being handled 578 * by the buffer cache ( else buffer space runs out first, usually ). 579 */ 580 static __inline void 581 bufcountadd(struct buf *bp) 582 { 583 int mask, need_wakeup, old, on; 584 585 KASSERT((bp->b_flags & B_INFREECNT) == 0, 586 ("buf %p already counted as free", bp)); 587 bp->b_flags |= B_INFREECNT; 588 old = atomic_fetchadd_int(&numfreebuffers, 1); 589 KASSERT(old >= 0 && old < nbuf, 590 ("numfreebuffers climbed to %d", old + 1)); 591 mask = VFS_BIO_NEED_ANY; 592 if (numfreebuffers >= hifreebuffers) 593 mask |= VFS_BIO_NEED_FREE; 594 rw_rlock(&nblock); 595 for (;;) { 596 need_wakeup = 0; 597 on = needsbuffer; 598 if (on == 0) 599 break; 600 need_wakeup = 1; 601 if (atomic_cmpset_rel_int(&needsbuffer, on, on & ~mask)) 602 break; 603 } 604 if (need_wakeup) 605 wakeup(__DEVOLATILE(void *, &needsbuffer)); 606 rw_runlock(&nblock); 607 } 608 609 /* 610 * bufcountsub: 611 * 612 * Decrement the numfreebuffers count as needed. 613 */ 614 static void 615 bufcountsub(struct buf *bp) 616 { 617 int old; 618 619 /* 620 * Fixup numfreebuffers count. If the buffer is invalid or not 621 * delayed-write, the buffer was free and we must decrement 622 * numfreebuffers. 623 */ 624 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 625 KASSERT((bp->b_flags & B_INFREECNT) != 0, 626 ("buf %p not counted in numfreebuffers", bp)); 627 bp->b_flags &= ~B_INFREECNT; 628 old = atomic_fetchadd_int(&numfreebuffers, -1); 629 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 630 } 631 } 632 633 /* 634 * waitrunningbufspace() 635 * 636 * runningbufspace is a measure of the amount of I/O currently 637 * running. This routine is used in async-write situations to 638 * prevent creating huge backups of pending writes to a device. 639 * Only asynchronous writes are governed by this function. 640 * 641 * This does NOT turn an async write into a sync write. It waits 642 * for earlier writes to complete and generally returns before the 643 * caller's write has reached the device. 644 */ 645 void 646 waitrunningbufspace(void) 647 { 648 649 mtx_lock(&rbreqlock); 650 while (runningbufspace > hirunningspace) { 651 runningbufreq = 1; 652 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 653 } 654 mtx_unlock(&rbreqlock); 655 } 656 657 658 /* 659 * vfs_buf_test_cache: 660 * 661 * Called when a buffer is extended. This function clears the B_CACHE 662 * bit if the newly extended portion of the buffer does not contain 663 * valid data. 664 */ 665 static __inline 666 void 667 vfs_buf_test_cache(struct buf *bp, 668 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 669 vm_page_t m) 670 { 671 672 VM_OBJECT_ASSERT_LOCKED(m->object); 673 if (bp->b_flags & B_CACHE) { 674 int base = (foff + off) & PAGE_MASK; 675 if (vm_page_is_valid(m, base, size) == 0) 676 bp->b_flags &= ~B_CACHE; 677 } 678 } 679 680 /* Wake up the buffer daemon if necessary */ 681 static __inline void 682 bd_wakeup(void) 683 { 684 685 mtx_lock(&bdlock); 686 if (bd_request == 0) { 687 bd_request = 1; 688 wakeup(&bd_request); 689 } 690 mtx_unlock(&bdlock); 691 } 692 693 /* 694 * bd_speedup - speedup the buffer cache flushing code 695 */ 696 void 697 bd_speedup(void) 698 { 699 int needwake; 700 701 mtx_lock(&bdlock); 702 needwake = 0; 703 if (bd_speedupreq == 0 || bd_request == 0) 704 needwake = 1; 705 bd_speedupreq = 1; 706 bd_request = 1; 707 if (needwake) 708 wakeup(&bd_request); 709 mtx_unlock(&bdlock); 710 } 711 712 #ifndef NSWBUF_MIN 713 #define NSWBUF_MIN 16 714 #endif 715 716 #ifdef __i386__ 717 #define TRANSIENT_DENOM 5 718 #else 719 #define TRANSIENT_DENOM 10 720 #endif 721 722 /* 723 * Calculating buffer cache scaling values and reserve space for buffer 724 * headers. This is called during low level kernel initialization and 725 * may be called more then once. We CANNOT write to the memory area 726 * being reserved at this time. 727 */ 728 caddr_t 729 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 730 { 731 int tuned_nbuf; 732 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 733 734 /* 735 * physmem_est is in pages. Convert it to kilobytes (assumes 736 * PAGE_SIZE is >= 1K) 737 */ 738 physmem_est = physmem_est * (PAGE_SIZE / 1024); 739 740 /* 741 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 742 * For the first 64MB of ram nominally allocate sufficient buffers to 743 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 744 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 745 * the buffer cache we limit the eventual kva reservation to 746 * maxbcache bytes. 747 * 748 * factor represents the 1/4 x ram conversion. 749 */ 750 if (nbuf == 0) { 751 int factor = 4 * BKVASIZE / 1024; 752 753 nbuf = 50; 754 if (physmem_est > 4096) 755 nbuf += min((physmem_est - 4096) / factor, 756 65536 / factor); 757 if (physmem_est > 65536) 758 nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 759 32 * 1024 * 1024 / (factor * 5)); 760 761 if (maxbcache && nbuf > maxbcache / BKVASIZE) 762 nbuf = maxbcache / BKVASIZE; 763 tuned_nbuf = 1; 764 } else 765 tuned_nbuf = 0; 766 767 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 768 maxbuf = (LONG_MAX / 3) / BKVASIZE; 769 if (nbuf > maxbuf) { 770 if (!tuned_nbuf) 771 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 772 maxbuf); 773 nbuf = maxbuf; 774 } 775 776 /* 777 * Ideal allocation size for the transient bio submap is 10% 778 * of the maximal space buffer map. This roughly corresponds 779 * to the amount of the buffer mapped for typical UFS load. 780 * 781 * Clip the buffer map to reserve space for the transient 782 * BIOs, if its extent is bigger than 90% (80% on i386) of the 783 * maximum buffer map extent on the platform. 784 * 785 * The fall-back to the maxbuf in case of maxbcache unset, 786 * allows to not trim the buffer KVA for the architectures 787 * with ample KVA space. 788 */ 789 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 790 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 791 buf_sz = (long)nbuf * BKVASIZE; 792 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 793 (TRANSIENT_DENOM - 1)) { 794 /* 795 * There is more KVA than memory. Do not 796 * adjust buffer map size, and assign the rest 797 * of maxbuf to transient map. 798 */ 799 biotmap_sz = maxbuf_sz - buf_sz; 800 } else { 801 /* 802 * Buffer map spans all KVA we could afford on 803 * this platform. Give 10% (20% on i386) of 804 * the buffer map to the transient bio map. 805 */ 806 biotmap_sz = buf_sz / TRANSIENT_DENOM; 807 buf_sz -= biotmap_sz; 808 } 809 if (biotmap_sz / INT_MAX > MAXPHYS) 810 bio_transient_maxcnt = INT_MAX; 811 else 812 bio_transient_maxcnt = biotmap_sz / MAXPHYS; 813 /* 814 * Artifically limit to 1024 simultaneous in-flight I/Os 815 * using the transient mapping. 816 */ 817 if (bio_transient_maxcnt > 1024) 818 bio_transient_maxcnt = 1024; 819 if (tuned_nbuf) 820 nbuf = buf_sz / BKVASIZE; 821 } 822 823 /* 824 * swbufs are used as temporary holders for I/O, such as paging I/O. 825 * We have no less then 16 and no more then 256. 826 */ 827 nswbuf = min(nbuf / 4, 256); 828 TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf); 829 if (nswbuf < NSWBUF_MIN) 830 nswbuf = NSWBUF_MIN; 831 832 /* 833 * Reserve space for the buffer cache buffers 834 */ 835 swbuf = (void *)v; 836 v = (caddr_t)(swbuf + nswbuf); 837 buf = (void *)v; 838 v = (caddr_t)(buf + nbuf); 839 840 return(v); 841 } 842 843 /* Initialize the buffer subsystem. Called before use of any buffers. */ 844 void 845 bufinit(void) 846 { 847 struct buf *bp; 848 int i; 849 850 CTASSERT(MAXBCACHEBUF >= MAXBSIZE); 851 mtx_init(&bqclean, "bufq clean lock", NULL, MTX_DEF); 852 mtx_init(&bqdirty, "bufq dirty lock", NULL, MTX_DEF); 853 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 854 rw_init(&nblock, "needsbuffer lock"); 855 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 856 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); 857 858 /* next, make a null set of free lists */ 859 for (i = 0; i < BUFFER_QUEUES; i++) 860 TAILQ_INIT(&bufqueues[i]); 861 862 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS); 863 864 /* finally, initialize each buffer header and stick on empty q */ 865 for (i = 0; i < nbuf; i++) { 866 bp = &buf[i]; 867 bzero(bp, sizeof *bp); 868 bp->b_flags = B_INVAL | B_INFREECNT; 869 bp->b_rcred = NOCRED; 870 bp->b_wcred = NOCRED; 871 bp->b_qindex = QUEUE_EMPTY; 872 bp->b_xflags = 0; 873 bp->b_data = bp->b_kvabase = unmapped_buf; 874 LIST_INIT(&bp->b_dep); 875 BUF_LOCKINIT(bp); 876 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 877 #ifdef INVARIANTS 878 bq_len[QUEUE_EMPTY]++; 879 #endif 880 } 881 882 /* 883 * maxbufspace is the absolute maximum amount of buffer space we are 884 * allowed to reserve in KVM and in real terms. The absolute maximum 885 * is nominally used by buf_daemon. hibufspace is the nominal maximum 886 * used by most other processes. The differential is required to 887 * ensure that buf_daemon is able to run when other processes might 888 * be blocked waiting for buffer space. 889 * 890 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 891 * this may result in KVM fragmentation which is not handled optimally 892 * by the system. 893 */ 894 maxbufspace = (long)nbuf * BKVASIZE; 895 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBCACHEBUF * 10); 896 lobufspace = hibufspace - MAXBCACHEBUF; 897 898 /* 899 * Note: The 16 MiB upper limit for hirunningspace was chosen 900 * arbitrarily and may need further tuning. It corresponds to 901 * 128 outstanding write IO requests (if IO size is 128 KiB), 902 * which fits with many RAID controllers' tagged queuing limits. 903 * The lower 1 MiB limit is the historical upper limit for 904 * hirunningspace. 905 */ 906 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBCACHEBUF), 907 16 * 1024 * 1024), 1024 * 1024); 908 lorunningspace = roundup((hirunningspace * 2) / 3, MAXBCACHEBUF); 909 910 /* 911 * Limit the amount of malloc memory since it is wired permanently into 912 * the kernel space. Even though this is accounted for in the buffer 913 * allocation, we don't want the malloced region to grow uncontrolled. 914 * The malloc scheme improves memory utilization significantly on average 915 * (small) directories. 916 */ 917 maxbufmallocspace = hibufspace / 20; 918 919 /* 920 * Reduce the chance of a deadlock occuring by limiting the number 921 * of delayed-write dirty buffers we allow to stack up. 922 */ 923 hidirtybuffers = nbuf / 4 + 20; 924 dirtybufthresh = hidirtybuffers * 9 / 10; 925 numdirtybuffers = 0; 926 /* 927 * To support extreme low-memory systems, make sure hidirtybuffers cannot 928 * eat up all available buffer space. This occurs when our minimum cannot 929 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 930 * BKVASIZE'd buffers. 931 */ 932 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 933 hidirtybuffers >>= 1; 934 } 935 lodirtybuffers = hidirtybuffers / 2; 936 937 /* 938 * Try to keep the number of free buffers in the specified range, 939 * and give special processes (e.g. like buf_daemon) access to an 940 * emergency reserve. 941 */ 942 lofreebuffers = nbuf / 18 + 5; 943 hifreebuffers = 2 * lofreebuffers; 944 numfreebuffers = nbuf; 945 946 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 947 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 948 } 949 950 #ifdef INVARIANTS 951 static inline void 952 vfs_buf_check_mapped(struct buf *bp) 953 { 954 955 KASSERT(bp->b_kvabase != unmapped_buf, 956 ("mapped buf: b_kvabase was not updated %p", bp)); 957 KASSERT(bp->b_data != unmapped_buf, 958 ("mapped buf: b_data was not updated %p", bp)); 959 } 960 961 static inline void 962 vfs_buf_check_unmapped(struct buf *bp) 963 { 964 965 KASSERT(bp->b_data == unmapped_buf, 966 ("unmapped buf: corrupted b_data %p", bp)); 967 } 968 969 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 970 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 971 #else 972 #define BUF_CHECK_MAPPED(bp) do {} while (0) 973 #define BUF_CHECK_UNMAPPED(bp) do {} while (0) 974 #endif 975 976 static void 977 bpmap_qenter(struct buf *bp) 978 { 979 980 BUF_CHECK_MAPPED(bp); 981 982 /* 983 * bp->b_data is relative to bp->b_offset, but 984 * bp->b_offset may be offset into the first page. 985 */ 986 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 987 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 988 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 989 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 990 } 991 992 /* 993 * binsfree: 994 * 995 * Insert the buffer into the appropriate free list. 996 */ 997 static void 998 binsfree(struct buf *bp, int qindex) 999 { 1000 struct mtx *olock, *nlock; 1001 1002 BUF_ASSERT_XLOCKED(bp); 1003 1004 nlock = bqlock(qindex); 1005 /* Handle delayed bremfree() processing. */ 1006 if (bp->b_flags & B_REMFREE) { 1007 olock = bqlock(bp->b_qindex); 1008 mtx_lock(olock); 1009 bremfreel(bp); 1010 if (olock != nlock) { 1011 mtx_unlock(olock); 1012 mtx_lock(nlock); 1013 } 1014 } else 1015 mtx_lock(nlock); 1016 1017 if (bp->b_qindex != QUEUE_NONE) 1018 panic("binsfree: free buffer onto another queue???"); 1019 1020 bp->b_qindex = qindex; 1021 if (bp->b_flags & B_AGE) 1022 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1023 else 1024 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1025 #ifdef INVARIANTS 1026 bq_len[bp->b_qindex]++; 1027 #endif 1028 mtx_unlock(nlock); 1029 1030 /* 1031 * Something we can maybe free or reuse. 1032 */ 1033 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1034 bufspacewakeup(); 1035 1036 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1037 bufcountadd(bp); 1038 } 1039 1040 /* 1041 * bremfree: 1042 * 1043 * Mark the buffer for removal from the appropriate free list. 1044 * 1045 */ 1046 void 1047 bremfree(struct buf *bp) 1048 { 1049 1050 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1051 KASSERT((bp->b_flags & B_REMFREE) == 0, 1052 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1053 KASSERT(bp->b_qindex != QUEUE_NONE, 1054 ("bremfree: buffer %p not on a queue.", bp)); 1055 BUF_ASSERT_XLOCKED(bp); 1056 1057 bp->b_flags |= B_REMFREE; 1058 bufcountsub(bp); 1059 } 1060 1061 /* 1062 * bremfreef: 1063 * 1064 * Force an immediate removal from a free list. Used only in nfs when 1065 * it abuses the b_freelist pointer. 1066 */ 1067 void 1068 bremfreef(struct buf *bp) 1069 { 1070 struct mtx *qlock; 1071 1072 qlock = bqlock(bp->b_qindex); 1073 mtx_lock(qlock); 1074 bremfreel(bp); 1075 mtx_unlock(qlock); 1076 } 1077 1078 /* 1079 * bremfreel: 1080 * 1081 * Removes a buffer from the free list, must be called with the 1082 * correct qlock held. 1083 */ 1084 static void 1085 bremfreel(struct buf *bp) 1086 { 1087 1088 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 1089 bp, bp->b_vp, bp->b_flags); 1090 KASSERT(bp->b_qindex != QUEUE_NONE, 1091 ("bremfreel: buffer %p not on a queue.", bp)); 1092 BUF_ASSERT_XLOCKED(bp); 1093 mtx_assert(bqlock(bp->b_qindex), MA_OWNED); 1094 1095 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 1096 #ifdef INVARIANTS 1097 KASSERT(bq_len[bp->b_qindex] >= 1, ("queue %d underflow", 1098 bp->b_qindex)); 1099 bq_len[bp->b_qindex]--; 1100 #endif 1101 bp->b_qindex = QUEUE_NONE; 1102 /* 1103 * If this was a delayed bremfree() we only need to remove the buffer 1104 * from the queue and return the stats are already done. 1105 */ 1106 if (bp->b_flags & B_REMFREE) { 1107 bp->b_flags &= ~B_REMFREE; 1108 return; 1109 } 1110 bufcountsub(bp); 1111 } 1112 1113 /* 1114 * bufkvafree: 1115 * 1116 * Free the kva allocation for a buffer. 1117 * 1118 */ 1119 static void 1120 bufkvafree(struct buf *bp) 1121 { 1122 1123 #ifdef INVARIANTS 1124 if (bp->b_kvasize == 0) { 1125 KASSERT(bp->b_kvabase == unmapped_buf && 1126 bp->b_data == unmapped_buf, 1127 ("Leaked KVA space on %p", bp)); 1128 } else if (buf_mapped(bp)) 1129 BUF_CHECK_MAPPED(bp); 1130 else 1131 BUF_CHECK_UNMAPPED(bp); 1132 #endif 1133 if (bp->b_kvasize == 0) 1134 return; 1135 1136 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize); 1137 atomic_subtract_long(&bufkvaspace, bp->b_kvasize); 1138 atomic_add_int(&buffreekvacnt, 1); 1139 bp->b_data = bp->b_kvabase = unmapped_buf; 1140 bp->b_kvasize = 0; 1141 } 1142 1143 /* 1144 * bufkvaalloc: 1145 * 1146 * Allocate the buffer KVA and set b_kvasize and b_kvabase. 1147 */ 1148 static int 1149 bufkvaalloc(struct buf *bp, int maxsize, int gbflags) 1150 { 1151 vm_offset_t addr; 1152 int error; 1153 1154 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0, 1155 ("Invalid gbflags 0x%x in %s", gbflags, __func__)); 1156 1157 bufkvafree(bp); 1158 1159 addr = 0; 1160 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr); 1161 if (error != 0) { 1162 /* 1163 * Buffer map is too fragmented. Request the caller 1164 * to defragment the map. 1165 */ 1166 atomic_add_int(&bufdefragcnt, 1); 1167 return (error); 1168 } 1169 bp->b_kvabase = (caddr_t)addr; 1170 bp->b_kvasize = maxsize; 1171 atomic_add_long(&bufkvaspace, bp->b_kvasize); 1172 if ((gbflags & GB_UNMAPPED) != 0) { 1173 bp->b_data = unmapped_buf; 1174 BUF_CHECK_UNMAPPED(bp); 1175 } else { 1176 bp->b_data = bp->b_kvabase; 1177 BUF_CHECK_MAPPED(bp); 1178 } 1179 return (0); 1180 } 1181 1182 /* 1183 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 1184 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 1185 * the buffer is valid and we do not have to do anything. 1186 */ 1187 void 1188 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 1189 int cnt, struct ucred * cred) 1190 { 1191 struct buf *rabp; 1192 int i; 1193 1194 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 1195 if (inmem(vp, *rablkno)) 1196 continue; 1197 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 1198 1199 if ((rabp->b_flags & B_CACHE) == 0) { 1200 if (!TD_IS_IDLETHREAD(curthread)) 1201 curthread->td_ru.ru_inblock++; 1202 rabp->b_flags |= B_ASYNC; 1203 rabp->b_flags &= ~B_INVAL; 1204 rabp->b_ioflags &= ~BIO_ERROR; 1205 rabp->b_iocmd = BIO_READ; 1206 if (rabp->b_rcred == NOCRED && cred != NOCRED) 1207 rabp->b_rcred = crhold(cred); 1208 vfs_busy_pages(rabp, 0); 1209 BUF_KERNPROC(rabp); 1210 rabp->b_iooffset = dbtob(rabp->b_blkno); 1211 bstrategy(rabp); 1212 } else { 1213 brelse(rabp); 1214 } 1215 } 1216 } 1217 1218 /* 1219 * Entry point for bread() and breadn() via #defines in sys/buf.h. 1220 * 1221 * Get a buffer with the specified data. Look in the cache first. We 1222 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 1223 * is set, the buffer is valid and we do not have to do anything, see 1224 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 1225 */ 1226 int 1227 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, 1228 int *rabsize, int cnt, struct ucred *cred, int flags, struct buf **bpp) 1229 { 1230 struct buf *bp; 1231 int rv = 0, readwait = 0; 1232 1233 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 1234 /* 1235 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 1236 */ 1237 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 1238 if (bp == NULL) 1239 return (EBUSY); 1240 1241 /* if not found in cache, do some I/O */ 1242 if ((bp->b_flags & B_CACHE) == 0) { 1243 if (!TD_IS_IDLETHREAD(curthread)) 1244 curthread->td_ru.ru_inblock++; 1245 bp->b_iocmd = BIO_READ; 1246 bp->b_flags &= ~B_INVAL; 1247 bp->b_ioflags &= ~BIO_ERROR; 1248 if (bp->b_rcred == NOCRED && cred != NOCRED) 1249 bp->b_rcred = crhold(cred); 1250 vfs_busy_pages(bp, 0); 1251 bp->b_iooffset = dbtob(bp->b_blkno); 1252 bstrategy(bp); 1253 ++readwait; 1254 } 1255 1256 breada(vp, rablkno, rabsize, cnt, cred); 1257 1258 if (readwait) { 1259 rv = bufwait(bp); 1260 } 1261 return (rv); 1262 } 1263 1264 /* 1265 * Write, release buffer on completion. (Done by iodone 1266 * if async). Do not bother writing anything if the buffer 1267 * is invalid. 1268 * 1269 * Note that we set B_CACHE here, indicating that buffer is 1270 * fully valid and thus cacheable. This is true even of NFS 1271 * now so we set it generally. This could be set either here 1272 * or in biodone() since the I/O is synchronous. We put it 1273 * here. 1274 */ 1275 int 1276 bufwrite(struct buf *bp) 1277 { 1278 int oldflags; 1279 struct vnode *vp; 1280 long space; 1281 int vp_md; 1282 1283 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1284 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) { 1285 bp->b_flags |= B_INVAL | B_RELBUF; 1286 bp->b_flags &= ~B_CACHE; 1287 brelse(bp); 1288 return (ENXIO); 1289 } 1290 if (bp->b_flags & B_INVAL) { 1291 brelse(bp); 1292 return (0); 1293 } 1294 1295 if (bp->b_flags & B_BARRIER) 1296 barrierwrites++; 1297 1298 oldflags = bp->b_flags; 1299 1300 BUF_ASSERT_HELD(bp); 1301 1302 if (bp->b_pin_count > 0) 1303 bunpin_wait(bp); 1304 1305 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 1306 ("FFS background buffer should not get here %p", bp)); 1307 1308 vp = bp->b_vp; 1309 if (vp) 1310 vp_md = vp->v_vflag & VV_MD; 1311 else 1312 vp_md = 0; 1313 1314 /* 1315 * Mark the buffer clean. Increment the bufobj write count 1316 * before bundirty() call, to prevent other thread from seeing 1317 * empty dirty list and zero counter for writes in progress, 1318 * falsely indicating that the bufobj is clean. 1319 */ 1320 bufobj_wref(bp->b_bufobj); 1321 bundirty(bp); 1322 1323 bp->b_flags &= ~B_DONE; 1324 bp->b_ioflags &= ~BIO_ERROR; 1325 bp->b_flags |= B_CACHE; 1326 bp->b_iocmd = BIO_WRITE; 1327 1328 vfs_busy_pages(bp, 1); 1329 1330 /* 1331 * Normal bwrites pipeline writes 1332 */ 1333 bp->b_runningbufspace = bp->b_bufsize; 1334 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); 1335 1336 if (!TD_IS_IDLETHREAD(curthread)) 1337 curthread->td_ru.ru_oublock++; 1338 if (oldflags & B_ASYNC) 1339 BUF_KERNPROC(bp); 1340 bp->b_iooffset = dbtob(bp->b_blkno); 1341 bstrategy(bp); 1342 1343 if ((oldflags & B_ASYNC) == 0) { 1344 int rtval = bufwait(bp); 1345 brelse(bp); 1346 return (rtval); 1347 } else if (space > hirunningspace) { 1348 /* 1349 * don't allow the async write to saturate the I/O 1350 * system. We will not deadlock here because 1351 * we are blocking waiting for I/O that is already in-progress 1352 * to complete. We do not block here if it is the update 1353 * or syncer daemon trying to clean up as that can lead 1354 * to deadlock. 1355 */ 1356 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 1357 waitrunningbufspace(); 1358 } 1359 1360 return (0); 1361 } 1362 1363 void 1364 bufbdflush(struct bufobj *bo, struct buf *bp) 1365 { 1366 struct buf *nbp; 1367 1368 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1369 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 1370 altbufferflushes++; 1371 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1372 BO_LOCK(bo); 1373 /* 1374 * Try to find a buffer to flush. 1375 */ 1376 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1377 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1378 BUF_LOCK(nbp, 1379 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1380 continue; 1381 if (bp == nbp) 1382 panic("bdwrite: found ourselves"); 1383 BO_UNLOCK(bo); 1384 /* Don't countdeps with the bo lock held. */ 1385 if (buf_countdeps(nbp, 0)) { 1386 BO_LOCK(bo); 1387 BUF_UNLOCK(nbp); 1388 continue; 1389 } 1390 if (nbp->b_flags & B_CLUSTEROK) { 1391 vfs_bio_awrite(nbp); 1392 } else { 1393 bremfree(nbp); 1394 bawrite(nbp); 1395 } 1396 dirtybufferflushes++; 1397 break; 1398 } 1399 if (nbp == NULL) 1400 BO_UNLOCK(bo); 1401 } 1402 } 1403 1404 /* 1405 * Delayed write. (Buffer is marked dirty). Do not bother writing 1406 * anything if the buffer is marked invalid. 1407 * 1408 * Note that since the buffer must be completely valid, we can safely 1409 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1410 * biodone() in order to prevent getblk from writing the buffer 1411 * out synchronously. 1412 */ 1413 void 1414 bdwrite(struct buf *bp) 1415 { 1416 struct thread *td = curthread; 1417 struct vnode *vp; 1418 struct bufobj *bo; 1419 1420 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1421 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1422 KASSERT((bp->b_flags & B_BARRIER) == 0, 1423 ("Barrier request in delayed write %p", bp)); 1424 BUF_ASSERT_HELD(bp); 1425 1426 if (bp->b_flags & B_INVAL) { 1427 brelse(bp); 1428 return; 1429 } 1430 1431 /* 1432 * If we have too many dirty buffers, don't create any more. 1433 * If we are wildly over our limit, then force a complete 1434 * cleanup. Otherwise, just keep the situation from getting 1435 * out of control. Note that we have to avoid a recursive 1436 * disaster and not try to clean up after our own cleanup! 1437 */ 1438 vp = bp->b_vp; 1439 bo = bp->b_bufobj; 1440 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1441 td->td_pflags |= TDP_INBDFLUSH; 1442 BO_BDFLUSH(bo, bp); 1443 td->td_pflags &= ~TDP_INBDFLUSH; 1444 } else 1445 recursiveflushes++; 1446 1447 bdirty(bp); 1448 /* 1449 * Set B_CACHE, indicating that the buffer is fully valid. This is 1450 * true even of NFS now. 1451 */ 1452 bp->b_flags |= B_CACHE; 1453 1454 /* 1455 * This bmap keeps the system from needing to do the bmap later, 1456 * perhaps when the system is attempting to do a sync. Since it 1457 * is likely that the indirect block -- or whatever other datastructure 1458 * that the filesystem needs is still in memory now, it is a good 1459 * thing to do this. Note also, that if the pageout daemon is 1460 * requesting a sync -- there might not be enough memory to do 1461 * the bmap then... So, this is important to do. 1462 */ 1463 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1464 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1465 } 1466 1467 /* 1468 * Set the *dirty* buffer range based upon the VM system dirty 1469 * pages. 1470 * 1471 * Mark the buffer pages as clean. We need to do this here to 1472 * satisfy the vnode_pager and the pageout daemon, so that it 1473 * thinks that the pages have been "cleaned". Note that since 1474 * the pages are in a delayed write buffer -- the VFS layer 1475 * "will" see that the pages get written out on the next sync, 1476 * or perhaps the cluster will be completed. 1477 */ 1478 vfs_clean_pages_dirty_buf(bp); 1479 bqrelse(bp); 1480 1481 /* 1482 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1483 * due to the softdep code. 1484 */ 1485 } 1486 1487 /* 1488 * bdirty: 1489 * 1490 * Turn buffer into delayed write request. We must clear BIO_READ and 1491 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1492 * itself to properly update it in the dirty/clean lists. We mark it 1493 * B_DONE to ensure that any asynchronization of the buffer properly 1494 * clears B_DONE ( else a panic will occur later ). 1495 * 1496 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1497 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1498 * should only be called if the buffer is known-good. 1499 * 1500 * Since the buffer is not on a queue, we do not update the numfreebuffers 1501 * count. 1502 * 1503 * The buffer must be on QUEUE_NONE. 1504 */ 1505 void 1506 bdirty(struct buf *bp) 1507 { 1508 1509 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1510 bp, bp->b_vp, bp->b_flags); 1511 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1512 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1513 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1514 BUF_ASSERT_HELD(bp); 1515 bp->b_flags &= ~(B_RELBUF); 1516 bp->b_iocmd = BIO_WRITE; 1517 1518 if ((bp->b_flags & B_DELWRI) == 0) { 1519 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1520 reassignbuf(bp); 1521 bdirtyadd(); 1522 } 1523 } 1524 1525 /* 1526 * bundirty: 1527 * 1528 * Clear B_DELWRI for buffer. 1529 * 1530 * Since the buffer is not on a queue, we do not update the numfreebuffers 1531 * count. 1532 * 1533 * The buffer must be on QUEUE_NONE. 1534 */ 1535 1536 void 1537 bundirty(struct buf *bp) 1538 { 1539 1540 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1541 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1542 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1543 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1544 BUF_ASSERT_HELD(bp); 1545 1546 if (bp->b_flags & B_DELWRI) { 1547 bp->b_flags &= ~B_DELWRI; 1548 reassignbuf(bp); 1549 bdirtysub(); 1550 } 1551 /* 1552 * Since it is now being written, we can clear its deferred write flag. 1553 */ 1554 bp->b_flags &= ~B_DEFERRED; 1555 } 1556 1557 /* 1558 * bawrite: 1559 * 1560 * Asynchronous write. Start output on a buffer, but do not wait for 1561 * it to complete. The buffer is released when the output completes. 1562 * 1563 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1564 * B_INVAL buffers. Not us. 1565 */ 1566 void 1567 bawrite(struct buf *bp) 1568 { 1569 1570 bp->b_flags |= B_ASYNC; 1571 (void) bwrite(bp); 1572 } 1573 1574 /* 1575 * babarrierwrite: 1576 * 1577 * Asynchronous barrier write. Start output on a buffer, but do not 1578 * wait for it to complete. Place a write barrier after this write so 1579 * that this buffer and all buffers written before it are committed to 1580 * the disk before any buffers written after this write are committed 1581 * to the disk. The buffer is released when the output completes. 1582 */ 1583 void 1584 babarrierwrite(struct buf *bp) 1585 { 1586 1587 bp->b_flags |= B_ASYNC | B_BARRIER; 1588 (void) bwrite(bp); 1589 } 1590 1591 /* 1592 * bbarrierwrite: 1593 * 1594 * Synchronous barrier write. Start output on a buffer and wait for 1595 * it to complete. Place a write barrier after this write so that 1596 * this buffer and all buffers written before it are committed to 1597 * the disk before any buffers written after this write are committed 1598 * to the disk. The buffer is released when the output completes. 1599 */ 1600 int 1601 bbarrierwrite(struct buf *bp) 1602 { 1603 1604 bp->b_flags |= B_BARRIER; 1605 return (bwrite(bp)); 1606 } 1607 1608 /* 1609 * bwillwrite: 1610 * 1611 * Called prior to the locking of any vnodes when we are expecting to 1612 * write. We do not want to starve the buffer cache with too many 1613 * dirty buffers so we block here. By blocking prior to the locking 1614 * of any vnodes we attempt to avoid the situation where a locked vnode 1615 * prevents the various system daemons from flushing related buffers. 1616 */ 1617 void 1618 bwillwrite(void) 1619 { 1620 1621 if (numdirtybuffers >= hidirtybuffers) { 1622 mtx_lock(&bdirtylock); 1623 while (numdirtybuffers >= hidirtybuffers) { 1624 bdirtywait = 1; 1625 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), 1626 "flswai", 0); 1627 } 1628 mtx_unlock(&bdirtylock); 1629 } 1630 } 1631 1632 /* 1633 * Return true if we have too many dirty buffers. 1634 */ 1635 int 1636 buf_dirty_count_severe(void) 1637 { 1638 1639 return(numdirtybuffers >= hidirtybuffers); 1640 } 1641 1642 /* 1643 * brelse: 1644 * 1645 * Release a busy buffer and, if requested, free its resources. The 1646 * buffer will be stashed in the appropriate bufqueue[] allowing it 1647 * to be accessed later as a cache entity or reused for other purposes. 1648 */ 1649 void 1650 brelse(struct buf *bp) 1651 { 1652 int qindex; 1653 1654 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1655 bp, bp->b_vp, bp->b_flags); 1656 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1657 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1658 1659 if (BUF_LOCKRECURSED(bp)) { 1660 /* 1661 * Do not process, in particular, do not handle the 1662 * B_INVAL/B_RELBUF and do not release to free list. 1663 */ 1664 BUF_UNLOCK(bp); 1665 return; 1666 } 1667 1668 if (bp->b_flags & B_MANAGED) { 1669 bqrelse(bp); 1670 return; 1671 } 1672 1673 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { 1674 BO_LOCK(bp->b_bufobj); 1675 bp->b_vflags &= ~BV_BKGRDERR; 1676 BO_UNLOCK(bp->b_bufobj); 1677 bdirty(bp); 1678 } 1679 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1680 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1681 /* 1682 * Failed write, redirty. Must clear BIO_ERROR to prevent 1683 * pages from being scrapped. If the error is anything 1684 * other than an I/O error (EIO), assume that retrying 1685 * is futile. 1686 */ 1687 bp->b_ioflags &= ~BIO_ERROR; 1688 bdirty(bp); 1689 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1690 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1691 /* 1692 * Either a failed I/O or we were asked to free or not 1693 * cache the buffer. 1694 */ 1695 bp->b_flags |= B_INVAL; 1696 if (!LIST_EMPTY(&bp->b_dep)) 1697 buf_deallocate(bp); 1698 if (bp->b_flags & B_DELWRI) 1699 bdirtysub(); 1700 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1701 if ((bp->b_flags & B_VMIO) == 0) { 1702 if (bp->b_bufsize) 1703 allocbuf(bp, 0); 1704 if (bp->b_vp) 1705 brelvp(bp); 1706 } 1707 } 1708 1709 /* 1710 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1711 * is called with B_DELWRI set, the underlying pages may wind up 1712 * getting freed causing a previous write (bdwrite()) to get 'lost' 1713 * because pages associated with a B_DELWRI bp are marked clean. 1714 * 1715 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1716 * if B_DELWRI is set. 1717 */ 1718 if (bp->b_flags & B_DELWRI) 1719 bp->b_flags &= ~B_RELBUF; 1720 1721 /* 1722 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1723 * constituted, not even NFS buffers now. Two flags effect this. If 1724 * B_INVAL, the struct buf is invalidated but the VM object is kept 1725 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1726 * 1727 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1728 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1729 * buffer is also B_INVAL because it hits the re-dirtying code above. 1730 * 1731 * Normally we can do this whether a buffer is B_DELWRI or not. If 1732 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1733 * the commit state and we cannot afford to lose the buffer. If the 1734 * buffer has a background write in progress, we need to keep it 1735 * around to prevent it from being reconstituted and starting a second 1736 * background write. 1737 */ 1738 if ((bp->b_flags & B_VMIO) 1739 && !(bp->b_vp->v_mount != NULL && 1740 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1741 !vn_isdisk(bp->b_vp, NULL) && 1742 (bp->b_flags & B_DELWRI)) 1743 ) { 1744 1745 int i, j, resid; 1746 vm_page_t m; 1747 off_t foff; 1748 vm_pindex_t poff; 1749 vm_object_t obj; 1750 1751 obj = bp->b_bufobj->bo_object; 1752 1753 /* 1754 * Get the base offset and length of the buffer. Note that 1755 * in the VMIO case if the buffer block size is not 1756 * page-aligned then b_data pointer may not be page-aligned. 1757 * But our b_pages[] array *IS* page aligned. 1758 * 1759 * block sizes less then DEV_BSIZE (usually 512) are not 1760 * supported due to the page granularity bits (m->valid, 1761 * m->dirty, etc...). 1762 * 1763 * See man buf(9) for more information 1764 */ 1765 resid = bp->b_bufsize; 1766 foff = bp->b_offset; 1767 for (i = 0; i < bp->b_npages; i++) { 1768 int had_bogus = 0; 1769 1770 m = bp->b_pages[i]; 1771 1772 /* 1773 * If we hit a bogus page, fixup *all* the bogus pages 1774 * now. 1775 */ 1776 if (m == bogus_page) { 1777 poff = OFF_TO_IDX(bp->b_offset); 1778 had_bogus = 1; 1779 1780 VM_OBJECT_RLOCK(obj); 1781 for (j = i; j < bp->b_npages; j++) { 1782 vm_page_t mtmp; 1783 mtmp = bp->b_pages[j]; 1784 if (mtmp == bogus_page) { 1785 mtmp = vm_page_lookup(obj, poff + j); 1786 if (!mtmp) { 1787 panic("brelse: page missing\n"); 1788 } 1789 bp->b_pages[j] = mtmp; 1790 } 1791 } 1792 VM_OBJECT_RUNLOCK(obj); 1793 1794 if ((bp->b_flags & B_INVAL) == 0 && 1795 buf_mapped(bp)) { 1796 BUF_CHECK_MAPPED(bp); 1797 pmap_qenter( 1798 trunc_page((vm_offset_t)bp->b_data), 1799 bp->b_pages, bp->b_npages); 1800 } 1801 m = bp->b_pages[i]; 1802 } 1803 if ((bp->b_flags & B_NOCACHE) || 1804 (bp->b_ioflags & BIO_ERROR && 1805 bp->b_iocmd == BIO_READ)) { 1806 int poffset = foff & PAGE_MASK; 1807 int presid = resid > (PAGE_SIZE - poffset) ? 1808 (PAGE_SIZE - poffset) : resid; 1809 1810 KASSERT(presid >= 0, ("brelse: extra page")); 1811 VM_OBJECT_WLOCK(obj); 1812 while (vm_page_xbusied(m)) { 1813 vm_page_lock(m); 1814 VM_OBJECT_WUNLOCK(obj); 1815 vm_page_busy_sleep(m, "mbncsh"); 1816 VM_OBJECT_WLOCK(obj); 1817 } 1818 if (pmap_page_wired_mappings(m) == 0) 1819 vm_page_set_invalid(m, poffset, presid); 1820 VM_OBJECT_WUNLOCK(obj); 1821 if (had_bogus) 1822 printf("avoided corruption bug in bogus_page/brelse code\n"); 1823 } 1824 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1825 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1826 } 1827 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1828 vfs_vmio_release(bp); 1829 1830 } else if (bp->b_flags & B_VMIO) { 1831 1832 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1833 vfs_vmio_release(bp); 1834 } 1835 1836 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1837 if (bp->b_bufsize != 0) 1838 allocbuf(bp, 0); 1839 if (bp->b_vp != NULL) 1840 brelvp(bp); 1841 } 1842 1843 /* 1844 * If the buffer has junk contents signal it and eventually 1845 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1846 * doesn't find it. 1847 */ 1848 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1849 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1850 bp->b_flags |= B_INVAL; 1851 if (bp->b_flags & B_INVAL) { 1852 if (bp->b_flags & B_DELWRI) 1853 bundirty(bp); 1854 if (bp->b_vp) 1855 brelvp(bp); 1856 } 1857 1858 /* buffers with no memory */ 1859 if (bp->b_bufsize == 0) { 1860 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1861 if (bp->b_vflags & BV_BKGRDINPROG) 1862 panic("losing buffer 1"); 1863 if (bp->b_kvasize) 1864 qindex = QUEUE_EMPTYKVA; 1865 else 1866 qindex = QUEUE_EMPTY; 1867 bp->b_flags |= B_AGE; 1868 /* buffers with junk contents */ 1869 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1870 (bp->b_ioflags & BIO_ERROR)) { 1871 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1872 if (bp->b_vflags & BV_BKGRDINPROG) 1873 panic("losing buffer 2"); 1874 qindex = QUEUE_CLEAN; 1875 bp->b_flags |= B_AGE; 1876 /* remaining buffers */ 1877 } else if (bp->b_flags & B_DELWRI) 1878 qindex = QUEUE_DIRTY; 1879 else 1880 qindex = QUEUE_CLEAN; 1881 1882 binsfree(bp, qindex); 1883 1884 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1885 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1886 panic("brelse: not dirty"); 1887 /* unlock */ 1888 BUF_UNLOCK(bp); 1889 } 1890 1891 /* 1892 * Release a buffer back to the appropriate queue but do not try to free 1893 * it. The buffer is expected to be used again soon. 1894 * 1895 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1896 * biodone() to requeue an async I/O on completion. It is also used when 1897 * known good buffers need to be requeued but we think we may need the data 1898 * again soon. 1899 * 1900 * XXX we should be able to leave the B_RELBUF hint set on completion. 1901 */ 1902 void 1903 bqrelse(struct buf *bp) 1904 { 1905 int qindex; 1906 1907 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1908 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1909 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1910 1911 if (BUF_LOCKRECURSED(bp)) { 1912 /* do not release to free list */ 1913 BUF_UNLOCK(bp); 1914 return; 1915 } 1916 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1917 1918 if (bp->b_flags & B_MANAGED) { 1919 if (bp->b_flags & B_REMFREE) 1920 bremfreef(bp); 1921 goto out; 1922 } 1923 1924 /* buffers with stale but valid contents */ 1925 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | 1926 BV_BKGRDERR)) == BV_BKGRDERR) { 1927 BO_LOCK(bp->b_bufobj); 1928 bp->b_vflags &= ~BV_BKGRDERR; 1929 BO_UNLOCK(bp->b_bufobj); 1930 qindex = QUEUE_DIRTY; 1931 } else { 1932 if ((bp->b_flags & B_DELWRI) == 0 && 1933 (bp->b_xflags & BX_VNDIRTY)) 1934 panic("bqrelse: not dirty"); 1935 qindex = QUEUE_CLEAN; 1936 } 1937 binsfree(bp, qindex); 1938 1939 out: 1940 /* unlock */ 1941 BUF_UNLOCK(bp); 1942 } 1943 1944 /* Give pages used by the bp back to the VM system (where possible) */ 1945 static void 1946 vfs_vmio_release(struct buf *bp) 1947 { 1948 vm_object_t obj; 1949 vm_page_t m; 1950 int i; 1951 1952 if (buf_mapped(bp)) { 1953 BUF_CHECK_MAPPED(bp); 1954 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 1955 } else 1956 BUF_CHECK_UNMAPPED(bp); 1957 obj = bp->b_bufobj->bo_object; 1958 if (obj != NULL) 1959 VM_OBJECT_WLOCK(obj); 1960 for (i = 0; i < bp->b_npages; i++) { 1961 m = bp->b_pages[i]; 1962 bp->b_pages[i] = NULL; 1963 /* 1964 * In order to keep page LRU ordering consistent, put 1965 * everything on the inactive queue. 1966 */ 1967 vm_page_lock(m); 1968 vm_page_unwire(m, PQ_INACTIVE); 1969 1970 /* 1971 * Might as well free the page if we can and it has 1972 * no valid data. We also free the page if the 1973 * buffer was used for direct I/O 1974 */ 1975 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) { 1976 if (m->wire_count == 0 && !vm_page_busied(m)) 1977 vm_page_free(m); 1978 } else if (bp->b_flags & B_DIRECT) 1979 vm_page_try_to_free(m); 1980 vm_page_unlock(m); 1981 } 1982 if (obj != NULL) 1983 VM_OBJECT_WUNLOCK(obj); 1984 1985 if (bp->b_bufsize) 1986 bufspaceadjust(bp, 0); 1987 bp->b_npages = 0; 1988 bp->b_flags &= ~B_VMIO; 1989 if (bp->b_vp) 1990 brelvp(bp); 1991 } 1992 1993 /* 1994 * Check to see if a block at a particular lbn is available for a clustered 1995 * write. 1996 */ 1997 static int 1998 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1999 { 2000 struct buf *bpa; 2001 int match; 2002 2003 match = 0; 2004 2005 /* If the buf isn't in core skip it */ 2006 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 2007 return (0); 2008 2009 /* If the buf is busy we don't want to wait for it */ 2010 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2011 return (0); 2012 2013 /* Only cluster with valid clusterable delayed write buffers */ 2014 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 2015 (B_DELWRI | B_CLUSTEROK)) 2016 goto done; 2017 2018 if (bpa->b_bufsize != size) 2019 goto done; 2020 2021 /* 2022 * Check to see if it is in the expected place on disk and that the 2023 * block has been mapped. 2024 */ 2025 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 2026 match = 1; 2027 done: 2028 BUF_UNLOCK(bpa); 2029 return (match); 2030 } 2031 2032 /* 2033 * vfs_bio_awrite: 2034 * 2035 * Implement clustered async writes for clearing out B_DELWRI buffers. 2036 * This is much better then the old way of writing only one buffer at 2037 * a time. Note that we may not be presented with the buffers in the 2038 * correct order, so we search for the cluster in both directions. 2039 */ 2040 int 2041 vfs_bio_awrite(struct buf *bp) 2042 { 2043 struct bufobj *bo; 2044 int i; 2045 int j; 2046 daddr_t lblkno = bp->b_lblkno; 2047 struct vnode *vp = bp->b_vp; 2048 int ncl; 2049 int nwritten; 2050 int size; 2051 int maxcl; 2052 int gbflags; 2053 2054 bo = &vp->v_bufobj; 2055 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0; 2056 /* 2057 * right now we support clustered writing only to regular files. If 2058 * we find a clusterable block we could be in the middle of a cluster 2059 * rather then at the beginning. 2060 */ 2061 if ((vp->v_type == VREG) && 2062 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 2063 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 2064 2065 size = vp->v_mount->mnt_stat.f_iosize; 2066 maxcl = MAXPHYS / size; 2067 2068 BO_RLOCK(bo); 2069 for (i = 1; i < maxcl; i++) 2070 if (vfs_bio_clcheck(vp, size, lblkno + i, 2071 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 2072 break; 2073 2074 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 2075 if (vfs_bio_clcheck(vp, size, lblkno - j, 2076 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 2077 break; 2078 BO_RUNLOCK(bo); 2079 --j; 2080 ncl = i + j; 2081 /* 2082 * this is a possible cluster write 2083 */ 2084 if (ncl != 1) { 2085 BUF_UNLOCK(bp); 2086 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 2087 gbflags); 2088 return (nwritten); 2089 } 2090 } 2091 bremfree(bp); 2092 bp->b_flags |= B_ASYNC; 2093 /* 2094 * default (old) behavior, writing out only one block 2095 * 2096 * XXX returns b_bufsize instead of b_bcount for nwritten? 2097 */ 2098 nwritten = bp->b_bufsize; 2099 (void) bwrite(bp); 2100 2101 return (nwritten); 2102 } 2103 2104 /* 2105 * Ask the bufdaemon for help, or act as bufdaemon itself, when a 2106 * locked vnode is supplied. 2107 */ 2108 static void 2109 getnewbuf_bufd_help(struct vnode *vp, int gbflags, int slpflag, int slptimeo, 2110 int defrag) 2111 { 2112 struct thread *td; 2113 char *waitmsg; 2114 int error, fl, flags, norunbuf; 2115 2116 mtx_assert(&bqclean, MA_OWNED); 2117 2118 if (defrag) { 2119 flags = VFS_BIO_NEED_BUFSPACE; 2120 waitmsg = "nbufkv"; 2121 } else if (bufspace >= hibufspace) { 2122 waitmsg = "nbufbs"; 2123 flags = VFS_BIO_NEED_BUFSPACE; 2124 } else { 2125 waitmsg = "newbuf"; 2126 flags = VFS_BIO_NEED_ANY; 2127 } 2128 atomic_set_int(&needsbuffer, flags); 2129 mtx_unlock(&bqclean); 2130 2131 bd_speedup(); /* heeeelp */ 2132 if ((gbflags & GB_NOWAIT_BD) != 0) 2133 return; 2134 2135 td = curthread; 2136 rw_wlock(&nblock); 2137 while ((needsbuffer & flags) != 0) { 2138 if (vp != NULL && vp->v_type != VCHR && 2139 (td->td_pflags & TDP_BUFNEED) == 0) { 2140 rw_wunlock(&nblock); 2141 /* 2142 * getblk() is called with a vnode locked, and 2143 * some majority of the dirty buffers may as 2144 * well belong to the vnode. Flushing the 2145 * buffers there would make a progress that 2146 * cannot be achieved by the buf_daemon, that 2147 * cannot lock the vnode. 2148 */ 2149 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2150 (td->td_pflags & TDP_NORUNNINGBUF); 2151 2152 /* 2153 * Play bufdaemon. The getnewbuf() function 2154 * may be called while the thread owns lock 2155 * for another dirty buffer for the same 2156 * vnode, which makes it impossible to use 2157 * VOP_FSYNC() there, due to the buffer lock 2158 * recursion. 2159 */ 2160 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2161 fl = buf_flush(vp, flushbufqtarget); 2162 td->td_pflags &= norunbuf; 2163 rw_wlock(&nblock); 2164 if (fl != 0) 2165 continue; 2166 if ((needsbuffer & flags) == 0) 2167 break; 2168 } 2169 error = rw_sleep(__DEVOLATILE(void *, &needsbuffer), &nblock, 2170 (PRIBIO + 4) | slpflag, waitmsg, slptimeo); 2171 if (error != 0) 2172 break; 2173 } 2174 rw_wunlock(&nblock); 2175 } 2176 2177 static void 2178 getnewbuf_reuse_bp(struct buf *bp, int qindex) 2179 { 2180 2181 CTR6(KTR_BUF, "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 2182 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 2183 bp->b_kvasize, bp->b_bufsize, qindex); 2184 mtx_assert(&bqclean, MA_NOTOWNED); 2185 2186 /* 2187 * Note: we no longer distinguish between VMIO and non-VMIO 2188 * buffers. 2189 */ 2190 KASSERT((bp->b_flags & B_DELWRI) == 0, 2191 ("delwri buffer %p found in queue %d", bp, qindex)); 2192 2193 if (qindex == QUEUE_CLEAN) { 2194 if (bp->b_flags & B_VMIO) { 2195 bp->b_flags &= ~B_ASYNC; 2196 vfs_vmio_release(bp); 2197 } 2198 if (bp->b_vp != NULL) 2199 brelvp(bp); 2200 } 2201 2202 /* 2203 * Get the rest of the buffer freed up. b_kva* is still valid 2204 * after this operation. 2205 */ 2206 2207 if (bp->b_rcred != NOCRED) { 2208 crfree(bp->b_rcred); 2209 bp->b_rcred = NOCRED; 2210 } 2211 if (bp->b_wcred != NOCRED) { 2212 crfree(bp->b_wcred); 2213 bp->b_wcred = NOCRED; 2214 } 2215 if (!LIST_EMPTY(&bp->b_dep)) 2216 buf_deallocate(bp); 2217 if (bp->b_vflags & BV_BKGRDINPROG) 2218 panic("losing buffer 3"); 2219 KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p. qindex: %d", 2220 bp, bp->b_vp, qindex)); 2221 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 2222 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 2223 2224 if (bp->b_bufsize) 2225 allocbuf(bp, 0); 2226 2227 bp->b_flags = 0; 2228 bp->b_ioflags = 0; 2229 bp->b_xflags = 0; 2230 KASSERT((bp->b_flags & B_INFREECNT) == 0, 2231 ("buf %p still counted as free?", bp)); 2232 bp->b_vflags = 0; 2233 bp->b_vp = NULL; 2234 bp->b_blkno = bp->b_lblkno = 0; 2235 bp->b_offset = NOOFFSET; 2236 bp->b_iodone = 0; 2237 bp->b_error = 0; 2238 bp->b_resid = 0; 2239 bp->b_bcount = 0; 2240 bp->b_npages = 0; 2241 bp->b_dirtyoff = bp->b_dirtyend = 0; 2242 bp->b_bufobj = NULL; 2243 bp->b_pin_count = 0; 2244 bp->b_fsprivate1 = NULL; 2245 bp->b_fsprivate2 = NULL; 2246 bp->b_fsprivate3 = NULL; 2247 2248 LIST_INIT(&bp->b_dep); 2249 } 2250 2251 static int flushingbufs; 2252 2253 static struct buf * 2254 getnewbuf_scan(int maxsize, int defrag, int unmapped, int metadata) 2255 { 2256 struct buf *bp, *nbp; 2257 int nqindex, qindex, pass; 2258 2259 KASSERT(!unmapped || !defrag, ("both unmapped and defrag")); 2260 2261 pass = 1; 2262 restart: 2263 atomic_add_int(&getnewbufrestarts, 1); 2264 2265 /* 2266 * Setup for scan. If we do not have enough free buffers, 2267 * we setup a degenerate case that immediately fails. Note 2268 * that if we are specially marked process, we are allowed to 2269 * dip into our reserves. 2270 * 2271 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 2272 * for the allocation of the mapped buffer. For unmapped, the 2273 * easiest is to start with EMPTY outright. 2274 * 2275 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 2276 * However, there are a number of cases (defragging, reusing, ...) 2277 * where we cannot backup. 2278 */ 2279 nbp = NULL; 2280 mtx_lock(&bqclean); 2281 if (!defrag && unmapped) { 2282 nqindex = QUEUE_EMPTY; 2283 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2284 } 2285 if (nbp == NULL) { 2286 nqindex = QUEUE_EMPTYKVA; 2287 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2288 } 2289 2290 /* 2291 * If no EMPTYKVA buffers and we are either defragging or 2292 * reusing, locate a CLEAN buffer to free or reuse. If 2293 * bufspace useage is low skip this step so we can allocate a 2294 * new buffer. 2295 */ 2296 if (nbp == NULL && (defrag || bufspace >= lobufspace)) { 2297 nqindex = QUEUE_CLEAN; 2298 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2299 } 2300 2301 /* 2302 * If we could not find or were not allowed to reuse a CLEAN 2303 * buffer, check to see if it is ok to use an EMPTY buffer. 2304 * We can only use an EMPTY buffer if allocating its KVA would 2305 * not otherwise run us out of buffer space. No KVA is needed 2306 * for the unmapped allocation. 2307 */ 2308 if (nbp == NULL && defrag == 0 && (bufspace + maxsize < hibufspace || 2309 metadata)) { 2310 nqindex = QUEUE_EMPTY; 2311 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 2312 } 2313 2314 /* 2315 * All available buffers might be clean, retry ignoring the 2316 * lobufspace as the last resort. 2317 */ 2318 if (nbp == NULL && !TAILQ_EMPTY(&bufqueues[QUEUE_CLEAN])) { 2319 nqindex = QUEUE_CLEAN; 2320 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2321 } 2322 2323 /* 2324 * Run scan, possibly freeing data and/or kva mappings on the fly 2325 * depending. 2326 */ 2327 while ((bp = nbp) != NULL) { 2328 qindex = nqindex; 2329 2330 /* 2331 * Calculate next bp (we can only use it if we do not 2332 * block or do other fancy things). 2333 */ 2334 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 2335 switch (qindex) { 2336 case QUEUE_EMPTY: 2337 nqindex = QUEUE_EMPTYKVA; 2338 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 2339 if (nbp != NULL) 2340 break; 2341 /* FALLTHROUGH */ 2342 case QUEUE_EMPTYKVA: 2343 nqindex = QUEUE_CLEAN; 2344 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 2345 if (nbp != NULL) 2346 break; 2347 /* FALLTHROUGH */ 2348 case QUEUE_CLEAN: 2349 if (metadata && pass == 1) { 2350 pass = 2; 2351 nqindex = QUEUE_EMPTY; 2352 nbp = TAILQ_FIRST( 2353 &bufqueues[QUEUE_EMPTY]); 2354 } 2355 /* 2356 * nbp is NULL. 2357 */ 2358 break; 2359 } 2360 } 2361 /* 2362 * If we are defragging then we need a buffer with 2363 * b_kvasize != 0. This situation occurs when we 2364 * have many unmapped bufs. 2365 */ 2366 if (defrag && bp->b_kvasize == 0) 2367 continue; 2368 2369 /* 2370 * Start freeing the bp. This is somewhat involved. nbp 2371 * remains valid only for QUEUE_EMPTY[KVA] bp's. 2372 */ 2373 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2374 continue; 2375 /* 2376 * BKGRDINPROG can only be set with the buf and bufobj 2377 * locks both held. We tolerate a race to clear it here. 2378 */ 2379 if (bp->b_vflags & BV_BKGRDINPROG) { 2380 BUF_UNLOCK(bp); 2381 continue; 2382 } 2383 2384 /* 2385 * Requeue the background write buffer with error. 2386 */ 2387 if ((bp->b_vflags & BV_BKGRDERR) != 0) { 2388 bremfreel(bp); 2389 mtx_unlock(&bqclean); 2390 bqrelse(bp); 2391 continue; 2392 } 2393 2394 KASSERT(bp->b_qindex == qindex, 2395 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp)); 2396 2397 bremfreel(bp); 2398 mtx_unlock(&bqclean); 2399 /* 2400 * NOTE: nbp is now entirely invalid. We can only restart 2401 * the scan from this point on. 2402 */ 2403 2404 getnewbuf_reuse_bp(bp, qindex); 2405 mtx_assert(&bqclean, MA_NOTOWNED); 2406 2407 /* 2408 * If we are defragging then free the buffer. 2409 */ 2410 if (defrag) { 2411 bp->b_flags |= B_INVAL; 2412 bufkvafree(bp); 2413 brelse(bp); 2414 defrag = 0; 2415 goto restart; 2416 } 2417 2418 /* 2419 * Notify any waiters for the buffer lock about 2420 * identity change by freeing the buffer. 2421 */ 2422 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2423 bp->b_flags |= B_INVAL; 2424 bufkvafree(bp); 2425 brelse(bp); 2426 goto restart; 2427 } 2428 2429 if (metadata) 2430 break; 2431 2432 /* 2433 * If we are overcomitted then recover the buffer and its 2434 * KVM space. This occurs in rare situations when multiple 2435 * processes are blocked in getnewbuf() or allocbuf(). 2436 */ 2437 if (bufspace >= hibufspace) 2438 flushingbufs = 1; 2439 if (flushingbufs && bp->b_kvasize != 0) { 2440 bp->b_flags |= B_INVAL; 2441 bufkvafree(bp); 2442 brelse(bp); 2443 goto restart; 2444 } 2445 if (bufspace < lobufspace) 2446 flushingbufs = 0; 2447 break; 2448 } 2449 return (bp); 2450 } 2451 2452 /* 2453 * getnewbuf: 2454 * 2455 * Find and initialize a new buffer header, freeing up existing buffers 2456 * in the bufqueues as necessary. The new buffer is returned locked. 2457 * 2458 * Important: B_INVAL is not set. If the caller wishes to throw the 2459 * buffer away, the caller must set B_INVAL prior to calling brelse(). 2460 * 2461 * We block if: 2462 * We have insufficient buffer headers 2463 * We have insufficient buffer space 2464 * buffer_arena is too fragmented ( space reservation fails ) 2465 * If we have to flush dirty buffers ( but we try to avoid this ) 2466 */ 2467 static struct buf * 2468 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 2469 int gbflags) 2470 { 2471 struct buf *bp; 2472 int defrag, metadata; 2473 2474 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 2475 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 2476 if (!unmapped_buf_allowed) 2477 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 2478 2479 defrag = 0; 2480 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 2481 vp->v_type == VCHR) 2482 metadata = 1; 2483 else 2484 metadata = 0; 2485 /* 2486 * We can't afford to block since we might be holding a vnode lock, 2487 * which may prevent system daemons from running. We deal with 2488 * low-memory situations by proactively returning memory and running 2489 * async I/O rather then sync I/O. 2490 */ 2491 atomic_add_int(&getnewbufcalls, 1); 2492 atomic_subtract_int(&getnewbufrestarts, 1); 2493 restart: 2494 bp = getnewbuf_scan(maxsize, defrag, (gbflags & (GB_UNMAPPED | 2495 GB_KVAALLOC)) == GB_UNMAPPED, metadata); 2496 if (bp != NULL) 2497 defrag = 0; 2498 2499 /* 2500 * If we exhausted our list, sleep as appropriate. We may have to 2501 * wakeup various daemons and write out some dirty buffers. 2502 * 2503 * Generally we are sleeping due to insufficient buffer space. 2504 */ 2505 if (bp == NULL) { 2506 mtx_assert(&bqclean, MA_OWNED); 2507 getnewbuf_bufd_help(vp, gbflags, slpflag, slptimeo, defrag); 2508 mtx_assert(&bqclean, MA_NOTOWNED); 2509 } else if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) == GB_UNMAPPED) { 2510 mtx_assert(&bqclean, MA_NOTOWNED); 2511 2512 bufkvafree(bp); 2513 atomic_add_int(&bufreusecnt, 1); 2514 } else { 2515 mtx_assert(&bqclean, MA_NOTOWNED); 2516 2517 /* 2518 * We finally have a valid bp. We aren't quite out of the 2519 * woods, we still have to reserve kva space. In order to 2520 * keep fragmentation sane we only allocate kva in BKVASIZE 2521 * chunks. 2522 */ 2523 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2524 2525 if (maxsize != bp->b_kvasize && 2526 bufkvaalloc(bp, maxsize, gbflags)) { 2527 defrag = 1; 2528 bp->b_flags |= B_INVAL; 2529 brelse(bp); 2530 goto restart; 2531 } 2532 atomic_add_int(&bufreusecnt, 1); 2533 } 2534 return (bp); 2535 } 2536 2537 /* 2538 * buf_daemon: 2539 * 2540 * buffer flushing daemon. Buffers are normally flushed by the 2541 * update daemon but if it cannot keep up this process starts to 2542 * take the load in an attempt to prevent getnewbuf() from blocking. 2543 */ 2544 2545 static struct kproc_desc buf_kp = { 2546 "bufdaemon", 2547 buf_daemon, 2548 &bufdaemonproc 2549 }; 2550 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2551 2552 static int 2553 buf_flush(struct vnode *vp, int target) 2554 { 2555 int flushed; 2556 2557 flushed = flushbufqueues(vp, target, 0); 2558 if (flushed == 0) { 2559 /* 2560 * Could not find any buffers without rollback 2561 * dependencies, so just write the first one 2562 * in the hopes of eventually making progress. 2563 */ 2564 if (vp != NULL && target > 2) 2565 target /= 2; 2566 flushbufqueues(vp, target, 1); 2567 } 2568 return (flushed); 2569 } 2570 2571 static void 2572 buf_daemon() 2573 { 2574 int lodirty; 2575 2576 /* 2577 * This process needs to be suspended prior to shutdown sync. 2578 */ 2579 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2580 SHUTDOWN_PRI_LAST); 2581 2582 /* 2583 * This process is allowed to take the buffer cache to the limit 2584 */ 2585 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2586 mtx_lock(&bdlock); 2587 for (;;) { 2588 bd_request = 0; 2589 mtx_unlock(&bdlock); 2590 2591 kproc_suspend_check(bufdaemonproc); 2592 lodirty = lodirtybuffers; 2593 if (bd_speedupreq) { 2594 lodirty = numdirtybuffers / 2; 2595 bd_speedupreq = 0; 2596 } 2597 /* 2598 * Do the flush. Limit the amount of in-transit I/O we 2599 * allow to build up, otherwise we would completely saturate 2600 * the I/O system. 2601 */ 2602 while (numdirtybuffers > lodirty) { 2603 if (buf_flush(NULL, numdirtybuffers - lodirty) == 0) 2604 break; 2605 kern_yield(PRI_USER); 2606 } 2607 2608 /* 2609 * Only clear bd_request if we have reached our low water 2610 * mark. The buf_daemon normally waits 1 second and 2611 * then incrementally flushes any dirty buffers that have 2612 * built up, within reason. 2613 * 2614 * If we were unable to hit our low water mark and couldn't 2615 * find any flushable buffers, we sleep for a short period 2616 * to avoid endless loops on unlockable buffers. 2617 */ 2618 mtx_lock(&bdlock); 2619 if (numdirtybuffers <= lodirtybuffers) { 2620 /* 2621 * We reached our low water mark, reset the 2622 * request and sleep until we are needed again. 2623 * The sleep is just so the suspend code works. 2624 */ 2625 bd_request = 0; 2626 /* 2627 * Do an extra wakeup in case dirty threshold 2628 * changed via sysctl and the explicit transition 2629 * out of shortfall was missed. 2630 */ 2631 bdirtywakeup(); 2632 if (runningbufspace <= lorunningspace) 2633 runningwakeup(); 2634 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2635 } else { 2636 /* 2637 * We couldn't find any flushable dirty buffers but 2638 * still have too many dirty buffers, we 2639 * have to sleep and try again. (rare) 2640 */ 2641 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2642 } 2643 } 2644 } 2645 2646 /* 2647 * flushbufqueues: 2648 * 2649 * Try to flush a buffer in the dirty queue. We must be careful to 2650 * free up B_INVAL buffers instead of write them, which NFS is 2651 * particularly sensitive to. 2652 */ 2653 static int flushwithdeps = 0; 2654 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2655 0, "Number of buffers flushed with dependecies that require rollbacks"); 2656 2657 static int 2658 flushbufqueues(struct vnode *lvp, int target, int flushdeps) 2659 { 2660 struct buf *sentinel; 2661 struct vnode *vp; 2662 struct mount *mp; 2663 struct buf *bp; 2664 int hasdeps; 2665 int flushed; 2666 int queue; 2667 int error; 2668 bool unlock; 2669 2670 flushed = 0; 2671 queue = QUEUE_DIRTY; 2672 bp = NULL; 2673 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2674 sentinel->b_qindex = QUEUE_SENTINEL; 2675 mtx_lock(&bqdirty); 2676 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2677 mtx_unlock(&bqdirty); 2678 while (flushed != target) { 2679 maybe_yield(); 2680 mtx_lock(&bqdirty); 2681 bp = TAILQ_NEXT(sentinel, b_freelist); 2682 if (bp != NULL) { 2683 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2684 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2685 b_freelist); 2686 } else { 2687 mtx_unlock(&bqdirty); 2688 break; 2689 } 2690 /* 2691 * Skip sentinels inserted by other invocations of the 2692 * flushbufqueues(), taking care to not reorder them. 2693 * 2694 * Only flush the buffers that belong to the 2695 * vnode locked by the curthread. 2696 */ 2697 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && 2698 bp->b_vp != lvp)) { 2699 mtx_unlock(&bqdirty); 2700 continue; 2701 } 2702 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 2703 mtx_unlock(&bqdirty); 2704 if (error != 0) 2705 continue; 2706 if (bp->b_pin_count > 0) { 2707 BUF_UNLOCK(bp); 2708 continue; 2709 } 2710 /* 2711 * BKGRDINPROG can only be set with the buf and bufobj 2712 * locks both held. We tolerate a race to clear it here. 2713 */ 2714 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2715 (bp->b_flags & B_DELWRI) == 0) { 2716 BUF_UNLOCK(bp); 2717 continue; 2718 } 2719 if (bp->b_flags & B_INVAL) { 2720 bremfreef(bp); 2721 brelse(bp); 2722 flushed++; 2723 continue; 2724 } 2725 2726 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2727 if (flushdeps == 0) { 2728 BUF_UNLOCK(bp); 2729 continue; 2730 } 2731 hasdeps = 1; 2732 } else 2733 hasdeps = 0; 2734 /* 2735 * We must hold the lock on a vnode before writing 2736 * one of its buffers. Otherwise we may confuse, or 2737 * in the case of a snapshot vnode, deadlock the 2738 * system. 2739 * 2740 * The lock order here is the reverse of the normal 2741 * of vnode followed by buf lock. This is ok because 2742 * the NOWAIT will prevent deadlock. 2743 */ 2744 vp = bp->b_vp; 2745 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2746 BUF_UNLOCK(bp); 2747 continue; 2748 } 2749 if (lvp == NULL) { 2750 unlock = true; 2751 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 2752 } else { 2753 ASSERT_VOP_LOCKED(vp, "getbuf"); 2754 unlock = false; 2755 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 2756 vn_lock(vp, LK_TRYUPGRADE); 2757 } 2758 if (error == 0) { 2759 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2760 bp, bp->b_vp, bp->b_flags); 2761 if (curproc == bufdaemonproc) { 2762 vfs_bio_awrite(bp); 2763 } else { 2764 bremfree(bp); 2765 bwrite(bp); 2766 notbufdflushes++; 2767 } 2768 vn_finished_write(mp); 2769 if (unlock) 2770 VOP_UNLOCK(vp, 0); 2771 flushwithdeps += hasdeps; 2772 flushed++; 2773 2774 /* 2775 * Sleeping on runningbufspace while holding 2776 * vnode lock leads to deadlock. 2777 */ 2778 if (curproc == bufdaemonproc && 2779 runningbufspace > hirunningspace) 2780 waitrunningbufspace(); 2781 continue; 2782 } 2783 vn_finished_write(mp); 2784 BUF_UNLOCK(bp); 2785 } 2786 mtx_lock(&bqdirty); 2787 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2788 mtx_unlock(&bqdirty); 2789 free(sentinel, M_TEMP); 2790 return (flushed); 2791 } 2792 2793 /* 2794 * Check to see if a block is currently memory resident. 2795 */ 2796 struct buf * 2797 incore(struct bufobj *bo, daddr_t blkno) 2798 { 2799 struct buf *bp; 2800 2801 BO_RLOCK(bo); 2802 bp = gbincore(bo, blkno); 2803 BO_RUNLOCK(bo); 2804 return (bp); 2805 } 2806 2807 /* 2808 * Returns true if no I/O is needed to access the 2809 * associated VM object. This is like incore except 2810 * it also hunts around in the VM system for the data. 2811 */ 2812 2813 static int 2814 inmem(struct vnode * vp, daddr_t blkno) 2815 { 2816 vm_object_t obj; 2817 vm_offset_t toff, tinc, size; 2818 vm_page_t m; 2819 vm_ooffset_t off; 2820 2821 ASSERT_VOP_LOCKED(vp, "inmem"); 2822 2823 if (incore(&vp->v_bufobj, blkno)) 2824 return 1; 2825 if (vp->v_mount == NULL) 2826 return 0; 2827 obj = vp->v_object; 2828 if (obj == NULL) 2829 return (0); 2830 2831 size = PAGE_SIZE; 2832 if (size > vp->v_mount->mnt_stat.f_iosize) 2833 size = vp->v_mount->mnt_stat.f_iosize; 2834 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2835 2836 VM_OBJECT_RLOCK(obj); 2837 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2838 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2839 if (!m) 2840 goto notinmem; 2841 tinc = size; 2842 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2843 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2844 if (vm_page_is_valid(m, 2845 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2846 goto notinmem; 2847 } 2848 VM_OBJECT_RUNLOCK(obj); 2849 return 1; 2850 2851 notinmem: 2852 VM_OBJECT_RUNLOCK(obj); 2853 return (0); 2854 } 2855 2856 /* 2857 * Set the dirty range for a buffer based on the status of the dirty 2858 * bits in the pages comprising the buffer. The range is limited 2859 * to the size of the buffer. 2860 * 2861 * Tell the VM system that the pages associated with this buffer 2862 * are clean. This is used for delayed writes where the data is 2863 * going to go to disk eventually without additional VM intevention. 2864 * 2865 * Note that while we only really need to clean through to b_bcount, we 2866 * just go ahead and clean through to b_bufsize. 2867 */ 2868 static void 2869 vfs_clean_pages_dirty_buf(struct buf *bp) 2870 { 2871 vm_ooffset_t foff, noff, eoff; 2872 vm_page_t m; 2873 int i; 2874 2875 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2876 return; 2877 2878 foff = bp->b_offset; 2879 KASSERT(bp->b_offset != NOOFFSET, 2880 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2881 2882 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 2883 vfs_drain_busy_pages(bp); 2884 vfs_setdirty_locked_object(bp); 2885 for (i = 0; i < bp->b_npages; i++) { 2886 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2887 eoff = noff; 2888 if (eoff > bp->b_offset + bp->b_bufsize) 2889 eoff = bp->b_offset + bp->b_bufsize; 2890 m = bp->b_pages[i]; 2891 vfs_page_set_validclean(bp, foff, m); 2892 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2893 foff = noff; 2894 } 2895 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 2896 } 2897 2898 static void 2899 vfs_setdirty_locked_object(struct buf *bp) 2900 { 2901 vm_object_t object; 2902 int i; 2903 2904 object = bp->b_bufobj->bo_object; 2905 VM_OBJECT_ASSERT_WLOCKED(object); 2906 2907 /* 2908 * We qualify the scan for modified pages on whether the 2909 * object has been flushed yet. 2910 */ 2911 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2912 vm_offset_t boffset; 2913 vm_offset_t eoffset; 2914 2915 /* 2916 * test the pages to see if they have been modified directly 2917 * by users through the VM system. 2918 */ 2919 for (i = 0; i < bp->b_npages; i++) 2920 vm_page_test_dirty(bp->b_pages[i]); 2921 2922 /* 2923 * Calculate the encompassing dirty range, boffset and eoffset, 2924 * (eoffset - boffset) bytes. 2925 */ 2926 2927 for (i = 0; i < bp->b_npages; i++) { 2928 if (bp->b_pages[i]->dirty) 2929 break; 2930 } 2931 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2932 2933 for (i = bp->b_npages - 1; i >= 0; --i) { 2934 if (bp->b_pages[i]->dirty) { 2935 break; 2936 } 2937 } 2938 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2939 2940 /* 2941 * Fit it to the buffer. 2942 */ 2943 2944 if (eoffset > bp->b_bcount) 2945 eoffset = bp->b_bcount; 2946 2947 /* 2948 * If we have a good dirty range, merge with the existing 2949 * dirty range. 2950 */ 2951 2952 if (boffset < eoffset) { 2953 if (bp->b_dirtyoff > boffset) 2954 bp->b_dirtyoff = boffset; 2955 if (bp->b_dirtyend < eoffset) 2956 bp->b_dirtyend = eoffset; 2957 } 2958 } 2959 } 2960 2961 /* 2962 * Allocate the KVA mapping for an existing buffer. 2963 * If an unmapped buffer is provided but a mapped buffer is requested, take 2964 * also care to properly setup mappings between pages and KVA. 2965 */ 2966 static void 2967 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 2968 { 2969 struct buf *scratch_bp; 2970 int bsize, maxsize, need_mapping, need_kva; 2971 off_t offset; 2972 2973 need_mapping = bp->b_data == unmapped_buf && 2974 (gbflags & GB_UNMAPPED) == 0; 2975 need_kva = bp->b_kvabase == unmapped_buf && 2976 bp->b_data == unmapped_buf && 2977 (gbflags & GB_KVAALLOC) != 0; 2978 if (!need_mapping && !need_kva) 2979 return; 2980 2981 BUF_CHECK_UNMAPPED(bp); 2982 2983 if (need_mapping && bp->b_kvabase != unmapped_buf) { 2984 /* 2985 * Buffer is not mapped, but the KVA was already 2986 * reserved at the time of the instantiation. Use the 2987 * allocated space. 2988 */ 2989 goto has_addr; 2990 } 2991 2992 /* 2993 * Calculate the amount of the address space we would reserve 2994 * if the buffer was mapped. 2995 */ 2996 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 2997 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 2998 offset = blkno * bsize; 2999 maxsize = size + (offset & PAGE_MASK); 3000 maxsize = imax(maxsize, bsize); 3001 3002 mapping_loop: 3003 if (bufkvaalloc(bp, maxsize, gbflags)) { 3004 /* 3005 * Request defragmentation. getnewbuf() returns us the 3006 * allocated space by the scratch buffer KVA. 3007 */ 3008 scratch_bp = getnewbuf(bp->b_vp, 0, 0, size, maxsize, gbflags | 3009 (GB_UNMAPPED | GB_KVAALLOC)); 3010 if (scratch_bp == NULL) { 3011 if ((gbflags & GB_NOWAIT_BD) != 0) { 3012 /* 3013 * XXXKIB: defragmentation cannot 3014 * succeed, not sure what else to do. 3015 */ 3016 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp); 3017 } 3018 atomic_add_int(&mappingrestarts, 1); 3019 goto mapping_loop; 3020 } 3021 KASSERT(scratch_bp->b_kvabase != unmapped_buf, 3022 ("scratch bp has no KVA %p", scratch_bp)); 3023 /* Grab pointers. */ 3024 bp->b_kvabase = scratch_bp->b_kvabase; 3025 bp->b_kvasize = scratch_bp->b_kvasize; 3026 bp->b_data = scratch_bp->b_data; 3027 3028 /* Get rid of the scratch buffer. */ 3029 scratch_bp->b_kvasize = 0; 3030 scratch_bp->b_flags |= B_INVAL; 3031 scratch_bp->b_data = scratch_bp->b_kvabase = unmapped_buf; 3032 brelse(scratch_bp); 3033 } 3034 has_addr: 3035 if (need_mapping) { 3036 /* b_offset is handled by bpmap_qenter. */ 3037 bp->b_data = bp->b_kvabase; 3038 BUF_CHECK_MAPPED(bp); 3039 bpmap_qenter(bp); 3040 } 3041 } 3042 3043 /* 3044 * getblk: 3045 * 3046 * Get a block given a specified block and offset into a file/device. 3047 * The buffers B_DONE bit will be cleared on return, making it almost 3048 * ready for an I/O initiation. B_INVAL may or may not be set on 3049 * return. The caller should clear B_INVAL prior to initiating a 3050 * READ. 3051 * 3052 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3053 * an existing buffer. 3054 * 3055 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3056 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3057 * and then cleared based on the backing VM. If the previous buffer is 3058 * non-0-sized but invalid, B_CACHE will be cleared. 3059 * 3060 * If getblk() must create a new buffer, the new buffer is returned with 3061 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3062 * case it is returned with B_INVAL clear and B_CACHE set based on the 3063 * backing VM. 3064 * 3065 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 3066 * B_CACHE bit is clear. 3067 * 3068 * What this means, basically, is that the caller should use B_CACHE to 3069 * determine whether the buffer is fully valid or not and should clear 3070 * B_INVAL prior to issuing a read. If the caller intends to validate 3071 * the buffer by loading its data area with something, the caller needs 3072 * to clear B_INVAL. If the caller does this without issuing an I/O, 3073 * the caller should set B_CACHE ( as an optimization ), else the caller 3074 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3075 * a write attempt or if it was a successfull read. If the caller 3076 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3077 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3078 */ 3079 struct buf * 3080 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3081 int flags) 3082 { 3083 struct buf *bp; 3084 struct bufobj *bo; 3085 int bsize, error, maxsize, vmio; 3086 off_t offset; 3087 3088 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3089 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3090 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3091 ASSERT_VOP_LOCKED(vp, "getblk"); 3092 if (size > MAXBCACHEBUF) 3093 panic("getblk: size(%d) > MAXBCACHEBUF(%d)\n", size, 3094 MAXBCACHEBUF); 3095 if (!unmapped_buf_allowed) 3096 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3097 3098 bo = &vp->v_bufobj; 3099 loop: 3100 BO_RLOCK(bo); 3101 bp = gbincore(bo, blkno); 3102 if (bp != NULL) { 3103 int lockflags; 3104 /* 3105 * Buffer is in-core. If the buffer is not busy nor managed, 3106 * it must be on a queue. 3107 */ 3108 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3109 3110 if (flags & GB_LOCK_NOWAIT) 3111 lockflags |= LK_NOWAIT; 3112 3113 error = BUF_TIMELOCK(bp, lockflags, 3114 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3115 3116 /* 3117 * If we slept and got the lock we have to restart in case 3118 * the buffer changed identities. 3119 */ 3120 if (error == ENOLCK) 3121 goto loop; 3122 /* We timed out or were interrupted. */ 3123 else if (error) 3124 return (NULL); 3125 /* If recursed, assume caller knows the rules. */ 3126 else if (BUF_LOCKRECURSED(bp)) 3127 goto end; 3128 3129 /* 3130 * The buffer is locked. B_CACHE is cleared if the buffer is 3131 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3132 * and for a VMIO buffer B_CACHE is adjusted according to the 3133 * backing VM cache. 3134 */ 3135 if (bp->b_flags & B_INVAL) 3136 bp->b_flags &= ~B_CACHE; 3137 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3138 bp->b_flags |= B_CACHE; 3139 if (bp->b_flags & B_MANAGED) 3140 MPASS(bp->b_qindex == QUEUE_NONE); 3141 else 3142 bremfree(bp); 3143 3144 /* 3145 * check for size inconsistencies for non-VMIO case. 3146 */ 3147 if (bp->b_bcount != size) { 3148 if ((bp->b_flags & B_VMIO) == 0 || 3149 (size > bp->b_kvasize)) { 3150 if (bp->b_flags & B_DELWRI) { 3151 /* 3152 * If buffer is pinned and caller does 3153 * not want sleep waiting for it to be 3154 * unpinned, bail out 3155 * */ 3156 if (bp->b_pin_count > 0) { 3157 if (flags & GB_LOCK_NOWAIT) { 3158 bqrelse(bp); 3159 return (NULL); 3160 } else { 3161 bunpin_wait(bp); 3162 } 3163 } 3164 bp->b_flags |= B_NOCACHE; 3165 bwrite(bp); 3166 } else { 3167 if (LIST_EMPTY(&bp->b_dep)) { 3168 bp->b_flags |= B_RELBUF; 3169 brelse(bp); 3170 } else { 3171 bp->b_flags |= B_NOCACHE; 3172 bwrite(bp); 3173 } 3174 } 3175 goto loop; 3176 } 3177 } 3178 3179 /* 3180 * Handle the case of unmapped buffer which should 3181 * become mapped, or the buffer for which KVA 3182 * reservation is requested. 3183 */ 3184 bp_unmapped_get_kva(bp, blkno, size, flags); 3185 3186 /* 3187 * If the size is inconsistant in the VMIO case, we can resize 3188 * the buffer. This might lead to B_CACHE getting set or 3189 * cleared. If the size has not changed, B_CACHE remains 3190 * unchanged from its previous state. 3191 */ 3192 if (bp->b_bcount != size) 3193 allocbuf(bp, size); 3194 3195 KASSERT(bp->b_offset != NOOFFSET, 3196 ("getblk: no buffer offset")); 3197 3198 /* 3199 * A buffer with B_DELWRI set and B_CACHE clear must 3200 * be committed before we can return the buffer in 3201 * order to prevent the caller from issuing a read 3202 * ( due to B_CACHE not being set ) and overwriting 3203 * it. 3204 * 3205 * Most callers, including NFS and FFS, need this to 3206 * operate properly either because they assume they 3207 * can issue a read if B_CACHE is not set, or because 3208 * ( for example ) an uncached B_DELWRI might loop due 3209 * to softupdates re-dirtying the buffer. In the latter 3210 * case, B_CACHE is set after the first write completes, 3211 * preventing further loops. 3212 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3213 * above while extending the buffer, we cannot allow the 3214 * buffer to remain with B_CACHE set after the write 3215 * completes or it will represent a corrupt state. To 3216 * deal with this we set B_NOCACHE to scrap the buffer 3217 * after the write. 3218 * 3219 * We might be able to do something fancy, like setting 3220 * B_CACHE in bwrite() except if B_DELWRI is already set, 3221 * so the below call doesn't set B_CACHE, but that gets real 3222 * confusing. This is much easier. 3223 */ 3224 3225 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3226 bp->b_flags |= B_NOCACHE; 3227 bwrite(bp); 3228 goto loop; 3229 } 3230 bp->b_flags &= ~B_DONE; 3231 } else { 3232 /* 3233 * Buffer is not in-core, create new buffer. The buffer 3234 * returned by getnewbuf() is locked. Note that the returned 3235 * buffer is also considered valid (not marked B_INVAL). 3236 */ 3237 BO_RUNLOCK(bo); 3238 /* 3239 * If the user does not want us to create the buffer, bail out 3240 * here. 3241 */ 3242 if (flags & GB_NOCREAT) 3243 return NULL; 3244 if (numfreebuffers == 0 && TD_IS_IDLETHREAD(curthread)) 3245 return NULL; 3246 3247 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 3248 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3249 offset = blkno * bsize; 3250 vmio = vp->v_object != NULL; 3251 if (vmio) { 3252 maxsize = size + (offset & PAGE_MASK); 3253 } else { 3254 maxsize = size; 3255 /* Do not allow non-VMIO notmapped buffers. */ 3256 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3257 } 3258 maxsize = imax(maxsize, bsize); 3259 3260 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 3261 if (bp == NULL) { 3262 if (slpflag || slptimeo) 3263 return NULL; 3264 goto loop; 3265 } 3266 3267 /* 3268 * This code is used to make sure that a buffer is not 3269 * created while the getnewbuf routine is blocked. 3270 * This can be a problem whether the vnode is locked or not. 3271 * If the buffer is created out from under us, we have to 3272 * throw away the one we just created. 3273 * 3274 * Note: this must occur before we associate the buffer 3275 * with the vp especially considering limitations in 3276 * the splay tree implementation when dealing with duplicate 3277 * lblkno's. 3278 */ 3279 BO_LOCK(bo); 3280 if (gbincore(bo, blkno)) { 3281 BO_UNLOCK(bo); 3282 bp->b_flags |= B_INVAL; 3283 brelse(bp); 3284 goto loop; 3285 } 3286 3287 /* 3288 * Insert the buffer into the hash, so that it can 3289 * be found by incore. 3290 */ 3291 bp->b_blkno = bp->b_lblkno = blkno; 3292 bp->b_offset = offset; 3293 bgetvp(vp, bp); 3294 BO_UNLOCK(bo); 3295 3296 /* 3297 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 3298 * buffer size starts out as 0, B_CACHE will be set by 3299 * allocbuf() for the VMIO case prior to it testing the 3300 * backing store for validity. 3301 */ 3302 3303 if (vmio) { 3304 bp->b_flags |= B_VMIO; 3305 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 3306 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 3307 bp, vp->v_object, bp->b_bufobj->bo_object)); 3308 } else { 3309 bp->b_flags &= ~B_VMIO; 3310 KASSERT(bp->b_bufobj->bo_object == NULL, 3311 ("ARGH! has b_bufobj->bo_object %p %p\n", 3312 bp, bp->b_bufobj->bo_object)); 3313 BUF_CHECK_MAPPED(bp); 3314 } 3315 3316 allocbuf(bp, size); 3317 bp->b_flags &= ~B_DONE; 3318 } 3319 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 3320 BUF_ASSERT_HELD(bp); 3321 end: 3322 KASSERT(bp->b_bufobj == bo, 3323 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 3324 return (bp); 3325 } 3326 3327 /* 3328 * Get an empty, disassociated buffer of given size. The buffer is initially 3329 * set to B_INVAL. 3330 */ 3331 struct buf * 3332 geteblk(int size, int flags) 3333 { 3334 struct buf *bp; 3335 int maxsize; 3336 3337 maxsize = (size + BKVAMASK) & ~BKVAMASK; 3338 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 3339 if ((flags & GB_NOWAIT_BD) && 3340 (curthread->td_pflags & TDP_BUFNEED) != 0) 3341 return (NULL); 3342 } 3343 allocbuf(bp, size); 3344 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 3345 BUF_ASSERT_HELD(bp); 3346 return (bp); 3347 } 3348 3349 /* 3350 * This code constitutes the buffer memory from either anonymous system 3351 * memory (in the case of non-VMIO operations) or from an associated 3352 * VM object (in the case of VMIO operations). This code is able to 3353 * resize a buffer up or down. 3354 * 3355 * Note that this code is tricky, and has many complications to resolve 3356 * deadlock or inconsistant data situations. Tread lightly!!! 3357 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 3358 * the caller. Calling this code willy nilly can result in the loss of data. 3359 * 3360 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 3361 * B_CACHE for the non-VMIO case. 3362 */ 3363 3364 int 3365 allocbuf(struct buf *bp, int size) 3366 { 3367 int newbsize, mbsize; 3368 int i; 3369 3370 BUF_ASSERT_HELD(bp); 3371 3372 if (bp->b_kvasize != 0 && bp->b_kvasize < size) 3373 panic("allocbuf: buffer too small"); 3374 3375 if ((bp->b_flags & B_VMIO) == 0) { 3376 caddr_t origbuf; 3377 int origbufsize; 3378 /* 3379 * Just get anonymous memory from the kernel. Don't 3380 * mess with B_CACHE. 3381 */ 3382 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3383 if (bp->b_flags & B_MALLOC) 3384 newbsize = mbsize; 3385 else 3386 newbsize = round_page(size); 3387 3388 if (newbsize < bp->b_bufsize) { 3389 /* 3390 * malloced buffers are not shrunk 3391 */ 3392 if (bp->b_flags & B_MALLOC) { 3393 if (newbsize) { 3394 bp->b_bcount = size; 3395 } else { 3396 free(bp->b_data, M_BIOBUF); 3397 bufmallocadjust(bp, 0); 3398 bp->b_data = bp->b_kvabase; 3399 bp->b_bcount = 0; 3400 bp->b_flags &= ~B_MALLOC; 3401 } 3402 return 1; 3403 } 3404 vm_hold_free_pages(bp, newbsize); 3405 } else if (newbsize > bp->b_bufsize) { 3406 /* 3407 * We only use malloced memory on the first allocation. 3408 * and revert to page-allocated memory when the buffer 3409 * grows. 3410 */ 3411 /* 3412 * There is a potential smp race here that could lead 3413 * to bufmallocspace slightly passing the max. It 3414 * is probably extremely rare and not worth worrying 3415 * over. 3416 */ 3417 if ((bufmallocspace < maxbufmallocspace) && 3418 (bp->b_bufsize == 0) && 3419 (mbsize <= PAGE_SIZE/2)) { 3420 3421 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 3422 bp->b_bcount = size; 3423 bp->b_flags |= B_MALLOC; 3424 bufmallocadjust(bp, mbsize); 3425 return 1; 3426 } 3427 origbuf = NULL; 3428 origbufsize = 0; 3429 /* 3430 * If the buffer is growing on its other-than-first 3431 * allocation then we revert to the page-allocation 3432 * scheme. 3433 */ 3434 if (bp->b_flags & B_MALLOC) { 3435 origbuf = bp->b_data; 3436 origbufsize = bp->b_bufsize; 3437 bp->b_data = bp->b_kvabase; 3438 bufmallocadjust(bp, 0); 3439 bp->b_flags &= ~B_MALLOC; 3440 newbsize = round_page(newbsize); 3441 } 3442 vm_hold_load_pages( 3443 bp, 3444 (vm_offset_t) bp->b_data + bp->b_bufsize, 3445 (vm_offset_t) bp->b_data + newbsize); 3446 if (origbuf) { 3447 bcopy(origbuf, bp->b_data, origbufsize); 3448 free(origbuf, M_BIOBUF); 3449 } 3450 } 3451 } else { 3452 int desiredpages; 3453 3454 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 3455 desiredpages = (size == 0) ? 0 : 3456 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 3457 3458 if (bp->b_flags & B_MALLOC) 3459 panic("allocbuf: VMIO buffer can't be malloced"); 3460 /* 3461 * Set B_CACHE initially if buffer is 0 length or will become 3462 * 0-length. 3463 */ 3464 if (size == 0 || bp->b_bufsize == 0) 3465 bp->b_flags |= B_CACHE; 3466 3467 if (newbsize < bp->b_bufsize) { 3468 /* 3469 * DEV_BSIZE aligned new buffer size is less then the 3470 * DEV_BSIZE aligned existing buffer size. Figure out 3471 * if we have to remove any pages. 3472 */ 3473 if (desiredpages < bp->b_npages) { 3474 vm_page_t m; 3475 3476 if (buf_mapped(bp)) { 3477 BUF_CHECK_MAPPED(bp); 3478 pmap_qremove((vm_offset_t)trunc_page( 3479 (vm_offset_t)bp->b_data) + 3480 (desiredpages << PAGE_SHIFT), 3481 (bp->b_npages - desiredpages)); 3482 } else 3483 BUF_CHECK_UNMAPPED(bp); 3484 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3485 for (i = desiredpages; i < bp->b_npages; i++) { 3486 /* 3487 * the page is not freed here -- it 3488 * is the responsibility of 3489 * vnode_pager_setsize 3490 */ 3491 m = bp->b_pages[i]; 3492 KASSERT(m != bogus_page, 3493 ("allocbuf: bogus page found")); 3494 while (vm_page_sleep_if_busy(m, 3495 "biodep")) 3496 continue; 3497 3498 bp->b_pages[i] = NULL; 3499 vm_page_lock(m); 3500 vm_page_unwire(m, PQ_INACTIVE); 3501 vm_page_unlock(m); 3502 } 3503 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3504 bp->b_npages = desiredpages; 3505 } 3506 } else if (size > bp->b_bcount) { 3507 /* 3508 * We are growing the buffer, possibly in a 3509 * byte-granular fashion. 3510 */ 3511 vm_object_t obj; 3512 vm_offset_t toff; 3513 vm_offset_t tinc; 3514 3515 /* 3516 * Step 1, bring in the VM pages from the object, 3517 * allocating them if necessary. We must clear 3518 * B_CACHE if these pages are not valid for the 3519 * range covered by the buffer. 3520 */ 3521 3522 obj = bp->b_bufobj->bo_object; 3523 3524 VM_OBJECT_WLOCK(obj); 3525 while (bp->b_npages < desiredpages) { 3526 vm_page_t m; 3527 3528 /* 3529 * We must allocate system pages since blocking 3530 * here could interfere with paging I/O, no 3531 * matter which process we are. 3532 * 3533 * Only exclusive busy can be tested here. 3534 * Blocking on shared busy might lead to 3535 * deadlocks once allocbuf() is called after 3536 * pages are vfs_busy_pages(). 3537 */ 3538 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3539 bp->b_npages, VM_ALLOC_NOBUSY | 3540 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3541 VM_ALLOC_IGN_SBUSY | 3542 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3543 if (m->valid == 0) 3544 bp->b_flags &= ~B_CACHE; 3545 bp->b_pages[bp->b_npages] = m; 3546 ++bp->b_npages; 3547 } 3548 3549 /* 3550 * Step 2. We've loaded the pages into the buffer, 3551 * we have to figure out if we can still have B_CACHE 3552 * set. Note that B_CACHE is set according to the 3553 * byte-granular range ( bcount and size ), new the 3554 * aligned range ( newbsize ). 3555 * 3556 * The VM test is against m->valid, which is DEV_BSIZE 3557 * aligned. Needless to say, the validity of the data 3558 * needs to also be DEV_BSIZE aligned. Note that this 3559 * fails with NFS if the server or some other client 3560 * extends the file's EOF. If our buffer is resized, 3561 * B_CACHE may remain set! XXX 3562 */ 3563 3564 toff = bp->b_bcount; 3565 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3566 3567 while ((bp->b_flags & B_CACHE) && toff < size) { 3568 vm_pindex_t pi; 3569 3570 if (tinc > (size - toff)) 3571 tinc = size - toff; 3572 3573 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3574 PAGE_SHIFT; 3575 3576 vfs_buf_test_cache( 3577 bp, 3578 bp->b_offset, 3579 toff, 3580 tinc, 3581 bp->b_pages[pi] 3582 ); 3583 toff += tinc; 3584 tinc = PAGE_SIZE; 3585 } 3586 VM_OBJECT_WUNLOCK(obj); 3587 3588 /* 3589 * Step 3, fixup the KVA pmap. 3590 */ 3591 if (buf_mapped(bp)) 3592 bpmap_qenter(bp); 3593 else 3594 BUF_CHECK_UNMAPPED(bp); 3595 } 3596 } 3597 /* Record changes in allocation size. */ 3598 if (bp->b_bufsize != newbsize) 3599 bufspaceadjust(bp, newbsize); 3600 bp->b_bcount = size; /* requested buffer size. */ 3601 return 1; 3602 } 3603 3604 extern int inflight_transient_maps; 3605 3606 void 3607 biodone(struct bio *bp) 3608 { 3609 struct mtx *mtxp; 3610 void (*done)(struct bio *); 3611 vm_offset_t start, end; 3612 3613 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 3614 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 3615 bp->bio_flags |= BIO_UNMAPPED; 3616 start = trunc_page((vm_offset_t)bp->bio_data); 3617 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 3618 bp->bio_data = unmapped_buf; 3619 pmap_qremove(start, OFF_TO_IDX(end - start)); 3620 vmem_free(transient_arena, start, end - start); 3621 atomic_add_int(&inflight_transient_maps, -1); 3622 } 3623 done = bp->bio_done; 3624 if (done == NULL) { 3625 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3626 mtx_lock(mtxp); 3627 bp->bio_flags |= BIO_DONE; 3628 wakeup(bp); 3629 mtx_unlock(mtxp); 3630 } else { 3631 bp->bio_flags |= BIO_DONE; 3632 done(bp); 3633 } 3634 } 3635 3636 /* 3637 * Wait for a BIO to finish. 3638 */ 3639 int 3640 biowait(struct bio *bp, const char *wchan) 3641 { 3642 struct mtx *mtxp; 3643 3644 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3645 mtx_lock(mtxp); 3646 while ((bp->bio_flags & BIO_DONE) == 0) 3647 msleep(bp, mtxp, PRIBIO, wchan, 0); 3648 mtx_unlock(mtxp); 3649 if (bp->bio_error != 0) 3650 return (bp->bio_error); 3651 if (!(bp->bio_flags & BIO_ERROR)) 3652 return (0); 3653 return (EIO); 3654 } 3655 3656 void 3657 biofinish(struct bio *bp, struct devstat *stat, int error) 3658 { 3659 3660 if (error) { 3661 bp->bio_error = error; 3662 bp->bio_flags |= BIO_ERROR; 3663 } 3664 if (stat != NULL) 3665 devstat_end_transaction_bio(stat, bp); 3666 biodone(bp); 3667 } 3668 3669 /* 3670 * bufwait: 3671 * 3672 * Wait for buffer I/O completion, returning error status. The buffer 3673 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3674 * error and cleared. 3675 */ 3676 int 3677 bufwait(struct buf *bp) 3678 { 3679 if (bp->b_iocmd == BIO_READ) 3680 bwait(bp, PRIBIO, "biord"); 3681 else 3682 bwait(bp, PRIBIO, "biowr"); 3683 if (bp->b_flags & B_EINTR) { 3684 bp->b_flags &= ~B_EINTR; 3685 return (EINTR); 3686 } 3687 if (bp->b_ioflags & BIO_ERROR) { 3688 return (bp->b_error ? bp->b_error : EIO); 3689 } else { 3690 return (0); 3691 } 3692 } 3693 3694 /* 3695 * Call back function from struct bio back up to struct buf. 3696 */ 3697 static void 3698 bufdonebio(struct bio *bip) 3699 { 3700 struct buf *bp; 3701 3702 bp = bip->bio_caller2; 3703 bp->b_resid = bip->bio_resid; 3704 bp->b_ioflags = bip->bio_flags; 3705 bp->b_error = bip->bio_error; 3706 if (bp->b_error) 3707 bp->b_ioflags |= BIO_ERROR; 3708 bufdone(bp); 3709 g_destroy_bio(bip); 3710 } 3711 3712 void 3713 dev_strategy(struct cdev *dev, struct buf *bp) 3714 { 3715 struct cdevsw *csw; 3716 int ref; 3717 3718 KASSERT(dev->si_refcount > 0, 3719 ("dev_strategy on un-referenced struct cdev *(%s) %p", 3720 devtoname(dev), dev)); 3721 3722 csw = dev_refthread(dev, &ref); 3723 dev_strategy_csw(dev, csw, bp); 3724 dev_relthread(dev, ref); 3725 } 3726 3727 void 3728 dev_strategy_csw(struct cdev *dev, struct cdevsw *csw, struct buf *bp) 3729 { 3730 struct bio *bip; 3731 3732 KASSERT(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE, 3733 ("b_iocmd botch")); 3734 KASSERT(((dev->si_flags & SI_ETERNAL) != 0 && csw != NULL) || 3735 dev->si_threadcount > 0, 3736 ("dev_strategy_csw threadcount cdev *(%s) %p", devtoname(dev), 3737 dev)); 3738 if (csw == NULL) { 3739 bp->b_error = ENXIO; 3740 bp->b_ioflags = BIO_ERROR; 3741 bufdone(bp); 3742 return; 3743 } 3744 for (;;) { 3745 bip = g_new_bio(); 3746 if (bip != NULL) 3747 break; 3748 /* Try again later */ 3749 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3750 } 3751 bip->bio_cmd = bp->b_iocmd; 3752 bip->bio_offset = bp->b_iooffset; 3753 bip->bio_length = bp->b_bcount; 3754 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3755 bdata2bio(bp, bip); 3756 bip->bio_done = bufdonebio; 3757 bip->bio_caller2 = bp; 3758 bip->bio_dev = dev; 3759 (*csw->d_strategy)(bip); 3760 } 3761 3762 /* 3763 * bufdone: 3764 * 3765 * Finish I/O on a buffer, optionally calling a completion function. 3766 * This is usually called from an interrupt so process blocking is 3767 * not allowed. 3768 * 3769 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3770 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3771 * assuming B_INVAL is clear. 3772 * 3773 * For the VMIO case, we set B_CACHE if the op was a read and no 3774 * read error occured, or if the op was a write. B_CACHE is never 3775 * set if the buffer is invalid or otherwise uncacheable. 3776 * 3777 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3778 * initiator to leave B_INVAL set to brelse the buffer out of existance 3779 * in the biodone routine. 3780 */ 3781 void 3782 bufdone(struct buf *bp) 3783 { 3784 struct bufobj *dropobj; 3785 void (*biodone)(struct buf *); 3786 3787 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3788 dropobj = NULL; 3789 3790 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3791 BUF_ASSERT_HELD(bp); 3792 3793 runningbufwakeup(bp); 3794 if (bp->b_iocmd == BIO_WRITE) 3795 dropobj = bp->b_bufobj; 3796 /* call optional completion function if requested */ 3797 if (bp->b_iodone != NULL) { 3798 biodone = bp->b_iodone; 3799 bp->b_iodone = NULL; 3800 (*biodone) (bp); 3801 if (dropobj) 3802 bufobj_wdrop(dropobj); 3803 return; 3804 } 3805 3806 bufdone_finish(bp); 3807 3808 if (dropobj) 3809 bufobj_wdrop(dropobj); 3810 } 3811 3812 void 3813 bufdone_finish(struct buf *bp) 3814 { 3815 BUF_ASSERT_HELD(bp); 3816 3817 if (!LIST_EMPTY(&bp->b_dep)) 3818 buf_complete(bp); 3819 3820 if (bp->b_flags & B_VMIO) { 3821 vm_ooffset_t foff; 3822 vm_page_t m; 3823 vm_object_t obj; 3824 struct vnode *vp; 3825 int bogus, i, iosize; 3826 3827 obj = bp->b_bufobj->bo_object; 3828 KASSERT(obj->paging_in_progress >= bp->b_npages, 3829 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3830 obj->paging_in_progress, bp->b_npages)); 3831 3832 vp = bp->b_vp; 3833 KASSERT(vp->v_holdcnt > 0, 3834 ("biodone_finish: vnode %p has zero hold count", vp)); 3835 KASSERT(vp->v_object != NULL, 3836 ("biodone_finish: vnode %p has no vm_object", vp)); 3837 3838 foff = bp->b_offset; 3839 KASSERT(bp->b_offset != NOOFFSET, 3840 ("biodone_finish: bp %p has no buffer offset", bp)); 3841 3842 /* 3843 * Set B_CACHE if the op was a normal read and no error 3844 * occured. B_CACHE is set for writes in the b*write() 3845 * routines. 3846 */ 3847 iosize = bp->b_bcount - bp->b_resid; 3848 if (bp->b_iocmd == BIO_READ && 3849 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3850 !(bp->b_ioflags & BIO_ERROR)) { 3851 bp->b_flags |= B_CACHE; 3852 } 3853 bogus = 0; 3854 VM_OBJECT_WLOCK(obj); 3855 for (i = 0; i < bp->b_npages; i++) { 3856 int bogusflag = 0; 3857 int resid; 3858 3859 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3860 if (resid > iosize) 3861 resid = iosize; 3862 3863 /* 3864 * cleanup bogus pages, restoring the originals 3865 */ 3866 m = bp->b_pages[i]; 3867 if (m == bogus_page) { 3868 bogus = bogusflag = 1; 3869 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3870 if (m == NULL) 3871 panic("biodone: page disappeared!"); 3872 bp->b_pages[i] = m; 3873 } 3874 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3875 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3876 (intmax_t)foff, (uintmax_t)m->pindex)); 3877 3878 /* 3879 * In the write case, the valid and clean bits are 3880 * already changed correctly ( see bdwrite() ), so we 3881 * only need to do this here in the read case. 3882 */ 3883 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3884 KASSERT((m->dirty & vm_page_bits(foff & 3885 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3886 " page %p has unexpected dirty bits", m)); 3887 vfs_page_set_valid(bp, foff, m); 3888 } 3889 3890 vm_page_sunbusy(m); 3891 vm_object_pip_subtract(obj, 1); 3892 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3893 iosize -= resid; 3894 } 3895 vm_object_pip_wakeupn(obj, 0); 3896 VM_OBJECT_WUNLOCK(obj); 3897 if (bogus && buf_mapped(bp)) { 3898 BUF_CHECK_MAPPED(bp); 3899 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3900 bp->b_pages, bp->b_npages); 3901 } 3902 } 3903 3904 /* 3905 * For asynchronous completions, release the buffer now. The brelse 3906 * will do a wakeup there if necessary - so no need to do a wakeup 3907 * here in the async case. The sync case always needs to do a wakeup. 3908 */ 3909 3910 if (bp->b_flags & B_ASYNC) { 3911 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3912 brelse(bp); 3913 else 3914 bqrelse(bp); 3915 } else 3916 bdone(bp); 3917 } 3918 3919 /* 3920 * This routine is called in lieu of iodone in the case of 3921 * incomplete I/O. This keeps the busy status for pages 3922 * consistant. 3923 */ 3924 void 3925 vfs_unbusy_pages(struct buf *bp) 3926 { 3927 int i; 3928 vm_object_t obj; 3929 vm_page_t m; 3930 3931 runningbufwakeup(bp); 3932 if (!(bp->b_flags & B_VMIO)) 3933 return; 3934 3935 obj = bp->b_bufobj->bo_object; 3936 VM_OBJECT_WLOCK(obj); 3937 for (i = 0; i < bp->b_npages; i++) { 3938 m = bp->b_pages[i]; 3939 if (m == bogus_page) { 3940 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3941 if (!m) 3942 panic("vfs_unbusy_pages: page missing\n"); 3943 bp->b_pages[i] = m; 3944 if (buf_mapped(bp)) { 3945 BUF_CHECK_MAPPED(bp); 3946 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3947 bp->b_pages, bp->b_npages); 3948 } else 3949 BUF_CHECK_UNMAPPED(bp); 3950 } 3951 vm_object_pip_subtract(obj, 1); 3952 vm_page_sunbusy(m); 3953 } 3954 vm_object_pip_wakeupn(obj, 0); 3955 VM_OBJECT_WUNLOCK(obj); 3956 } 3957 3958 /* 3959 * vfs_page_set_valid: 3960 * 3961 * Set the valid bits in a page based on the supplied offset. The 3962 * range is restricted to the buffer's size. 3963 * 3964 * This routine is typically called after a read completes. 3965 */ 3966 static void 3967 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3968 { 3969 vm_ooffset_t eoff; 3970 3971 /* 3972 * Compute the end offset, eoff, such that [off, eoff) does not span a 3973 * page boundary and eoff is not greater than the end of the buffer. 3974 * The end of the buffer, in this case, is our file EOF, not the 3975 * allocation size of the buffer. 3976 */ 3977 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3978 if (eoff > bp->b_offset + bp->b_bcount) 3979 eoff = bp->b_offset + bp->b_bcount; 3980 3981 /* 3982 * Set valid range. This is typically the entire buffer and thus the 3983 * entire page. 3984 */ 3985 if (eoff > off) 3986 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 3987 } 3988 3989 /* 3990 * vfs_page_set_validclean: 3991 * 3992 * Set the valid bits and clear the dirty bits in a page based on the 3993 * supplied offset. The range is restricted to the buffer's size. 3994 */ 3995 static void 3996 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3997 { 3998 vm_ooffset_t soff, eoff; 3999 4000 /* 4001 * Start and end offsets in buffer. eoff - soff may not cross a 4002 * page boundry or cross the end of the buffer. The end of the 4003 * buffer, in this case, is our file EOF, not the allocation size 4004 * of the buffer. 4005 */ 4006 soff = off; 4007 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4008 if (eoff > bp->b_offset + bp->b_bcount) 4009 eoff = bp->b_offset + bp->b_bcount; 4010 4011 /* 4012 * Set valid range. This is typically the entire buffer and thus the 4013 * entire page. 4014 */ 4015 if (eoff > soff) { 4016 vm_page_set_validclean( 4017 m, 4018 (vm_offset_t) (soff & PAGE_MASK), 4019 (vm_offset_t) (eoff - soff) 4020 ); 4021 } 4022 } 4023 4024 /* 4025 * Ensure that all buffer pages are not exclusive busied. If any page is 4026 * exclusive busy, drain it. 4027 */ 4028 void 4029 vfs_drain_busy_pages(struct buf *bp) 4030 { 4031 vm_page_t m; 4032 int i, last_busied; 4033 4034 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 4035 last_busied = 0; 4036 for (i = 0; i < bp->b_npages; i++) { 4037 m = bp->b_pages[i]; 4038 if (vm_page_xbusied(m)) { 4039 for (; last_busied < i; last_busied++) 4040 vm_page_sbusy(bp->b_pages[last_busied]); 4041 while (vm_page_xbusied(m)) { 4042 vm_page_lock(m); 4043 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4044 vm_page_busy_sleep(m, "vbpage"); 4045 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4046 } 4047 } 4048 } 4049 for (i = 0; i < last_busied; i++) 4050 vm_page_sunbusy(bp->b_pages[i]); 4051 } 4052 4053 /* 4054 * This routine is called before a device strategy routine. 4055 * It is used to tell the VM system that paging I/O is in 4056 * progress, and treat the pages associated with the buffer 4057 * almost as being exclusive busy. Also the object paging_in_progress 4058 * flag is handled to make sure that the object doesn't become 4059 * inconsistant. 4060 * 4061 * Since I/O has not been initiated yet, certain buffer flags 4062 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 4063 * and should be ignored. 4064 */ 4065 void 4066 vfs_busy_pages(struct buf *bp, int clear_modify) 4067 { 4068 int i, bogus; 4069 vm_object_t obj; 4070 vm_ooffset_t foff; 4071 vm_page_t m; 4072 4073 if (!(bp->b_flags & B_VMIO)) 4074 return; 4075 4076 obj = bp->b_bufobj->bo_object; 4077 foff = bp->b_offset; 4078 KASSERT(bp->b_offset != NOOFFSET, 4079 ("vfs_busy_pages: no buffer offset")); 4080 VM_OBJECT_WLOCK(obj); 4081 vfs_drain_busy_pages(bp); 4082 if (bp->b_bufsize != 0) 4083 vfs_setdirty_locked_object(bp); 4084 bogus = 0; 4085 for (i = 0; i < bp->b_npages; i++) { 4086 m = bp->b_pages[i]; 4087 4088 if ((bp->b_flags & B_CLUSTER) == 0) { 4089 vm_object_pip_add(obj, 1); 4090 vm_page_sbusy(m); 4091 } 4092 /* 4093 * When readying a buffer for a read ( i.e 4094 * clear_modify == 0 ), it is important to do 4095 * bogus_page replacement for valid pages in 4096 * partially instantiated buffers. Partially 4097 * instantiated buffers can, in turn, occur when 4098 * reconstituting a buffer from its VM backing store 4099 * base. We only have to do this if B_CACHE is 4100 * clear ( which causes the I/O to occur in the 4101 * first place ). The replacement prevents the read 4102 * I/O from overwriting potentially dirty VM-backed 4103 * pages. XXX bogus page replacement is, uh, bogus. 4104 * It may not work properly with small-block devices. 4105 * We need to find a better way. 4106 */ 4107 if (clear_modify) { 4108 pmap_remove_write(m); 4109 vfs_page_set_validclean(bp, foff, m); 4110 } else if (m->valid == VM_PAGE_BITS_ALL && 4111 (bp->b_flags & B_CACHE) == 0) { 4112 bp->b_pages[i] = bogus_page; 4113 bogus++; 4114 } 4115 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4116 } 4117 VM_OBJECT_WUNLOCK(obj); 4118 if (bogus && buf_mapped(bp)) { 4119 BUF_CHECK_MAPPED(bp); 4120 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4121 bp->b_pages, bp->b_npages); 4122 } 4123 } 4124 4125 /* 4126 * vfs_bio_set_valid: 4127 * 4128 * Set the range within the buffer to valid. The range is 4129 * relative to the beginning of the buffer, b_offset. Note that 4130 * b_offset itself may be offset from the beginning of the first 4131 * page. 4132 */ 4133 void 4134 vfs_bio_set_valid(struct buf *bp, int base, int size) 4135 { 4136 int i, n; 4137 vm_page_t m; 4138 4139 if (!(bp->b_flags & B_VMIO)) 4140 return; 4141 4142 /* 4143 * Fixup base to be relative to beginning of first page. 4144 * Set initial n to be the maximum number of bytes in the 4145 * first page that can be validated. 4146 */ 4147 base += (bp->b_offset & PAGE_MASK); 4148 n = PAGE_SIZE - (base & PAGE_MASK); 4149 4150 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4151 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4152 m = bp->b_pages[i]; 4153 if (n > size) 4154 n = size; 4155 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4156 base += n; 4157 size -= n; 4158 n = PAGE_SIZE; 4159 } 4160 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4161 } 4162 4163 /* 4164 * vfs_bio_clrbuf: 4165 * 4166 * If the specified buffer is a non-VMIO buffer, clear the entire 4167 * buffer. If the specified buffer is a VMIO buffer, clear and 4168 * validate only the previously invalid portions of the buffer. 4169 * This routine essentially fakes an I/O, so we need to clear 4170 * BIO_ERROR and B_INVAL. 4171 * 4172 * Note that while we only theoretically need to clear through b_bcount, 4173 * we go ahead and clear through b_bufsize. 4174 */ 4175 void 4176 vfs_bio_clrbuf(struct buf *bp) 4177 { 4178 int i, j, mask, sa, ea, slide; 4179 4180 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4181 clrbuf(bp); 4182 return; 4183 } 4184 bp->b_flags &= ~B_INVAL; 4185 bp->b_ioflags &= ~BIO_ERROR; 4186 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4187 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4188 (bp->b_offset & PAGE_MASK) == 0) { 4189 if (bp->b_pages[0] == bogus_page) 4190 goto unlock; 4191 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4192 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4193 if ((bp->b_pages[0]->valid & mask) == mask) 4194 goto unlock; 4195 if ((bp->b_pages[0]->valid & mask) == 0) { 4196 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4197 bp->b_pages[0]->valid |= mask; 4198 goto unlock; 4199 } 4200 } 4201 sa = bp->b_offset & PAGE_MASK; 4202 slide = 0; 4203 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4204 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4205 ea = slide & PAGE_MASK; 4206 if (ea == 0) 4207 ea = PAGE_SIZE; 4208 if (bp->b_pages[i] == bogus_page) 4209 continue; 4210 j = sa / DEV_BSIZE; 4211 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4212 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4213 if ((bp->b_pages[i]->valid & mask) == mask) 4214 continue; 4215 if ((bp->b_pages[i]->valid & mask) == 0) 4216 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4217 else { 4218 for (; sa < ea; sa += DEV_BSIZE, j++) { 4219 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4220 pmap_zero_page_area(bp->b_pages[i], 4221 sa, DEV_BSIZE); 4222 } 4223 } 4224 } 4225 bp->b_pages[i]->valid |= mask; 4226 } 4227 unlock: 4228 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4229 bp->b_resid = 0; 4230 } 4231 4232 void 4233 vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4234 { 4235 vm_page_t m; 4236 int i, n; 4237 4238 if (buf_mapped(bp)) { 4239 BUF_CHECK_MAPPED(bp); 4240 bzero(bp->b_data + base, size); 4241 } else { 4242 BUF_CHECK_UNMAPPED(bp); 4243 n = PAGE_SIZE - (base & PAGE_MASK); 4244 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4245 m = bp->b_pages[i]; 4246 if (n > size) 4247 n = size; 4248 pmap_zero_page_area(m, base & PAGE_MASK, n); 4249 base += n; 4250 size -= n; 4251 n = PAGE_SIZE; 4252 } 4253 } 4254 } 4255 4256 /* 4257 * vm_hold_load_pages and vm_hold_free_pages get pages into 4258 * a buffers address space. The pages are anonymous and are 4259 * not associated with a file object. 4260 */ 4261 static void 4262 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4263 { 4264 vm_offset_t pg; 4265 vm_page_t p; 4266 int index; 4267 4268 BUF_CHECK_MAPPED(bp); 4269 4270 to = round_page(to); 4271 from = round_page(from); 4272 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4273 4274 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4275 tryagain: 4276 /* 4277 * note: must allocate system pages since blocking here 4278 * could interfere with paging I/O, no matter which 4279 * process we are. 4280 */ 4281 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4282 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 4283 if (p == NULL) { 4284 VM_WAIT; 4285 goto tryagain; 4286 } 4287 pmap_qenter(pg, &p, 1); 4288 bp->b_pages[index] = p; 4289 } 4290 bp->b_npages = index; 4291 } 4292 4293 /* Return pages associated with this buf to the vm system */ 4294 static void 4295 vm_hold_free_pages(struct buf *bp, int newbsize) 4296 { 4297 vm_offset_t from; 4298 vm_page_t p; 4299 int index, newnpages; 4300 4301 BUF_CHECK_MAPPED(bp); 4302 4303 from = round_page((vm_offset_t)bp->b_data + newbsize); 4304 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4305 if (bp->b_npages > newnpages) 4306 pmap_qremove(from, bp->b_npages - newnpages); 4307 for (index = newnpages; index < bp->b_npages; index++) { 4308 p = bp->b_pages[index]; 4309 bp->b_pages[index] = NULL; 4310 if (vm_page_sbusied(p)) 4311 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 4312 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 4313 p->wire_count--; 4314 vm_page_free(p); 4315 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 4316 } 4317 bp->b_npages = newnpages; 4318 } 4319 4320 /* 4321 * Map an IO request into kernel virtual address space. 4322 * 4323 * All requests are (re)mapped into kernel VA space. 4324 * Notice that we use b_bufsize for the size of the buffer 4325 * to be mapped. b_bcount might be modified by the driver. 4326 * 4327 * Note that even if the caller determines that the address space should 4328 * be valid, a race or a smaller-file mapped into a larger space may 4329 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4330 * check the return value. 4331 * 4332 * This function only works with pager buffers. 4333 */ 4334 int 4335 vmapbuf(struct buf *bp, int mapbuf) 4336 { 4337 vm_prot_t prot; 4338 int pidx; 4339 4340 if (bp->b_bufsize < 0) 4341 return (-1); 4342 prot = VM_PROT_READ; 4343 if (bp->b_iocmd == BIO_READ) 4344 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4345 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4346 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4347 btoc(MAXPHYS))) < 0) 4348 return (-1); 4349 bp->b_npages = pidx; 4350 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4351 if (mapbuf || !unmapped_buf_allowed) { 4352 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx); 4353 bp->b_data = bp->b_kvabase + bp->b_offset; 4354 } else 4355 bp->b_data = unmapped_buf; 4356 return(0); 4357 } 4358 4359 /* 4360 * Free the io map PTEs associated with this IO operation. 4361 * We also invalidate the TLB entries and restore the original b_addr. 4362 * 4363 * This function only works with pager buffers. 4364 */ 4365 void 4366 vunmapbuf(struct buf *bp) 4367 { 4368 int npages; 4369 4370 npages = bp->b_npages; 4371 if (buf_mapped(bp)) 4372 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4373 vm_page_unhold_pages(bp->b_pages, npages); 4374 4375 bp->b_data = unmapped_buf; 4376 } 4377 4378 void 4379 bdone(struct buf *bp) 4380 { 4381 struct mtx *mtxp; 4382 4383 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4384 mtx_lock(mtxp); 4385 bp->b_flags |= B_DONE; 4386 wakeup(bp); 4387 mtx_unlock(mtxp); 4388 } 4389 4390 void 4391 bwait(struct buf *bp, u_char pri, const char *wchan) 4392 { 4393 struct mtx *mtxp; 4394 4395 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4396 mtx_lock(mtxp); 4397 while ((bp->b_flags & B_DONE) == 0) 4398 msleep(bp, mtxp, pri, wchan, 0); 4399 mtx_unlock(mtxp); 4400 } 4401 4402 int 4403 bufsync(struct bufobj *bo, int waitfor) 4404 { 4405 4406 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 4407 } 4408 4409 void 4410 bufstrategy(struct bufobj *bo, struct buf *bp) 4411 { 4412 int i = 0; 4413 struct vnode *vp; 4414 4415 vp = bp->b_vp; 4416 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4417 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4418 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4419 i = VOP_STRATEGY(vp, bp); 4420 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4421 } 4422 4423 void 4424 bufobj_wrefl(struct bufobj *bo) 4425 { 4426 4427 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4428 ASSERT_BO_WLOCKED(bo); 4429 bo->bo_numoutput++; 4430 } 4431 4432 void 4433 bufobj_wref(struct bufobj *bo) 4434 { 4435 4436 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4437 BO_LOCK(bo); 4438 bo->bo_numoutput++; 4439 BO_UNLOCK(bo); 4440 } 4441 4442 void 4443 bufobj_wdrop(struct bufobj *bo) 4444 { 4445 4446 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4447 BO_LOCK(bo); 4448 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4449 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4450 bo->bo_flag &= ~BO_WWAIT; 4451 wakeup(&bo->bo_numoutput); 4452 } 4453 BO_UNLOCK(bo); 4454 } 4455 4456 int 4457 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4458 { 4459 int error; 4460 4461 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4462 ASSERT_BO_WLOCKED(bo); 4463 error = 0; 4464 while (bo->bo_numoutput) { 4465 bo->bo_flag |= BO_WWAIT; 4466 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 4467 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4468 if (error) 4469 break; 4470 } 4471 return (error); 4472 } 4473 4474 void 4475 bpin(struct buf *bp) 4476 { 4477 struct mtx *mtxp; 4478 4479 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4480 mtx_lock(mtxp); 4481 bp->b_pin_count++; 4482 mtx_unlock(mtxp); 4483 } 4484 4485 void 4486 bunpin(struct buf *bp) 4487 { 4488 struct mtx *mtxp; 4489 4490 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4491 mtx_lock(mtxp); 4492 if (--bp->b_pin_count == 0) 4493 wakeup(bp); 4494 mtx_unlock(mtxp); 4495 } 4496 4497 void 4498 bunpin_wait(struct buf *bp) 4499 { 4500 struct mtx *mtxp; 4501 4502 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4503 mtx_lock(mtxp); 4504 while (bp->b_pin_count > 0) 4505 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4506 mtx_unlock(mtxp); 4507 } 4508 4509 /* 4510 * Set bio_data or bio_ma for struct bio from the struct buf. 4511 */ 4512 void 4513 bdata2bio(struct buf *bp, struct bio *bip) 4514 { 4515 4516 if (!buf_mapped(bp)) { 4517 KASSERT(unmapped_buf_allowed, ("unmapped")); 4518 bip->bio_ma = bp->b_pages; 4519 bip->bio_ma_n = bp->b_npages; 4520 bip->bio_data = unmapped_buf; 4521 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 4522 bip->bio_flags |= BIO_UNMAPPED; 4523 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 4524 PAGE_SIZE == bp->b_npages, 4525 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 4526 (long long)bip->bio_length, bip->bio_ma_n)); 4527 } else { 4528 bip->bio_data = bp->b_data; 4529 bip->bio_ma = NULL; 4530 } 4531 } 4532 4533 #include "opt_ddb.h" 4534 #ifdef DDB 4535 #include <ddb/ddb.h> 4536 4537 /* DDB command to show buffer data */ 4538 DB_SHOW_COMMAND(buffer, db_show_buffer) 4539 { 4540 /* get args */ 4541 struct buf *bp = (struct buf *)addr; 4542 4543 if (!have_addr) { 4544 db_printf("usage: show buffer <addr>\n"); 4545 return; 4546 } 4547 4548 db_printf("buf at %p\n", bp); 4549 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 4550 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 4551 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 4552 db_printf( 4553 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4554 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 4555 "b_dep = %p\n", 4556 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4557 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4558 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 4559 db_printf("b_kvabase = %p, b_kvasize = %d\n", 4560 bp->b_kvabase, bp->b_kvasize); 4561 if (bp->b_npages) { 4562 int i; 4563 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4564 for (i = 0; i < bp->b_npages; i++) { 4565 vm_page_t m; 4566 m = bp->b_pages[i]; 4567 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4568 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4569 if ((i + 1) < bp->b_npages) 4570 db_printf(","); 4571 } 4572 db_printf("\n"); 4573 } 4574 db_printf(" "); 4575 BUF_LOCKPRINTINFO(bp); 4576 } 4577 4578 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4579 { 4580 struct buf *bp; 4581 int i; 4582 4583 for (i = 0; i < nbuf; i++) { 4584 bp = &buf[i]; 4585 if (BUF_ISLOCKED(bp)) { 4586 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4587 db_printf("\n"); 4588 } 4589 } 4590 } 4591 4592 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4593 { 4594 struct vnode *vp; 4595 struct buf *bp; 4596 4597 if (!have_addr) { 4598 db_printf("usage: show vnodebufs <addr>\n"); 4599 return; 4600 } 4601 vp = (struct vnode *)addr; 4602 db_printf("Clean buffers:\n"); 4603 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4604 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4605 db_printf("\n"); 4606 } 4607 db_printf("Dirty buffers:\n"); 4608 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4609 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4610 db_printf("\n"); 4611 } 4612 } 4613 4614 DB_COMMAND(countfreebufs, db_coundfreebufs) 4615 { 4616 struct buf *bp; 4617 int i, used = 0, nfree = 0; 4618 4619 if (have_addr) { 4620 db_printf("usage: countfreebufs\n"); 4621 return; 4622 } 4623 4624 for (i = 0; i < nbuf; i++) { 4625 bp = &buf[i]; 4626 if ((bp->b_flags & B_INFREECNT) != 0) 4627 nfree++; 4628 else 4629 used++; 4630 } 4631 4632 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4633 nfree + used); 4634 db_printf("numfreebuffers is %d\n", numfreebuffers); 4635 } 4636 #endif /* DDB */ 4637