1 /*- 2 * Copyright (c) 2004 Poul-Henning Kamp 3 * Copyright (c) 1994,1997 John S. Dyson 4 * All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28 /* 29 * this file contains a new buffer I/O scheme implementing a coherent 30 * VM object and buffer cache scheme. Pains have been taken to make 31 * sure that the performance degradation associated with schemes such 32 * as this is not realized. 33 * 34 * Author: John S. Dyson 35 * Significant help during the development and debugging phases 36 * had been provided by David Greenman, also of the FreeBSD core team. 37 * 38 * see man buf(9) for more info. 39 */ 40 41 #include <sys/cdefs.h> 42 __FBSDID("$FreeBSD$"); 43 44 #include <sys/param.h> 45 #include <sys/systm.h> 46 #include <sys/bio.h> 47 #include <sys/conf.h> 48 #include <sys/buf.h> 49 #include <sys/devicestat.h> 50 #include <sys/eventhandler.h> 51 #include <sys/lock.h> 52 #include <sys/malloc.h> 53 #include <sys/mount.h> 54 #include <sys/mutex.h> 55 #include <sys/kernel.h> 56 #include <sys/kthread.h> 57 #include <sys/proc.h> 58 #include <sys/resourcevar.h> 59 #include <sys/sysctl.h> 60 #include <sys/vmmeter.h> 61 #include <sys/vnode.h> 62 #include <geom/geom.h> 63 #include <vm/vm.h> 64 #include <vm/vm_param.h> 65 #include <vm/vm_kern.h> 66 #include <vm/vm_pageout.h> 67 #include <vm/vm_page.h> 68 #include <vm/vm_object.h> 69 #include <vm/vm_extern.h> 70 #include <vm/vm_map.h> 71 #include "opt_directio.h" 72 #include "opt_swap.h" 73 74 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 75 76 struct bio_ops bioops; /* I/O operation notification */ 77 78 struct buf_ops buf_ops_bio = { 79 .bop_name = "buf_ops_bio", 80 .bop_write = bufwrite, 81 .bop_strategy = bufstrategy, 82 .bop_sync = bufsync, 83 .bop_bdflush = bufbdflush, 84 }; 85 86 /* 87 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 88 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 89 */ 90 struct buf *buf; /* buffer header pool */ 91 92 static struct proc *bufdaemonproc; 93 94 static int inmem(struct vnode *vp, daddr_t blkno); 95 static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 96 vm_offset_t to); 97 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 98 vm_offset_t to); 99 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 100 int pageno, vm_page_t m); 101 static void vfs_clean_pages(struct buf *bp); 102 static void vfs_setdirty(struct buf *bp); 103 static void vfs_setdirty_locked_object(struct buf *bp); 104 static void vfs_vmio_release(struct buf *bp); 105 static int vfs_bio_clcheck(struct vnode *vp, int size, 106 daddr_t lblkno, daddr_t blkno); 107 static int flushbufqueues(int, int); 108 static void buf_daemon(void); 109 static void bremfreel(struct buf *bp); 110 111 int vmiodirenable = TRUE; 112 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 113 "Use the VM system for directory writes"); 114 int runningbufspace; 115 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 116 "Amount of presently outstanding async buffer io"); 117 static int bufspace; 118 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 119 "KVA memory used for bufs"); 120 static int maxbufspace; 121 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 122 "Maximum allowed value of bufspace (including buf_daemon)"); 123 static int bufmallocspace; 124 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 125 "Amount of malloced memory for buffers"); 126 static int maxbufmallocspace; 127 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 128 "Maximum amount of malloced memory for buffers"); 129 static int lobufspace; 130 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 131 "Minimum amount of buffers we want to have"); 132 int hibufspace; 133 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 134 "Maximum allowed value of bufspace (excluding buf_daemon)"); 135 static int bufreusecnt; 136 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 137 "Number of times we have reused a buffer"); 138 static int buffreekvacnt; 139 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 140 "Number of times we have freed the KVA space from some buffer"); 141 static int bufdefragcnt; 142 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 143 "Number of times we have had to repeat buffer allocation to defragment"); 144 static int lorunningspace; 145 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 146 "Minimum preferred space used for in-progress I/O"); 147 static int hirunningspace; 148 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 149 "Maximum amount of space to use for in-progress I/O"); 150 int dirtybufferflushes; 151 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 152 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 153 int bdwriteskip; 154 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 155 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 156 int altbufferflushes; 157 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 158 0, "Number of fsync flushes to limit dirty buffers"); 159 static int recursiveflushes; 160 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 161 0, "Number of flushes skipped due to being recursive"); 162 static int numdirtybuffers; 163 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 164 "Number of buffers that are dirty (has unwritten changes) at the moment"); 165 static int lodirtybuffers; 166 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 167 "How many buffers we want to have free before bufdaemon can sleep"); 168 static int hidirtybuffers; 169 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 170 "When the number of dirty buffers is considered severe"); 171 int dirtybufthresh; 172 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 173 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 174 static int numfreebuffers; 175 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 176 "Number of free buffers"); 177 static int lofreebuffers; 178 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 179 "XXX Unused"); 180 static int hifreebuffers; 181 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 182 "XXX Complicatedly unused"); 183 static int getnewbufcalls; 184 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 185 "Number of calls to getnewbuf"); 186 static int getnewbufrestarts; 187 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 188 "Number of times getnewbuf has had to restart a buffer aquisition"); 189 190 /* 191 * Wakeup point for bufdaemon, as well as indicator of whether it is already 192 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 193 * is idling. 194 */ 195 static int bd_request; 196 197 /* 198 * This lock synchronizes access to bd_request. 199 */ 200 static struct mtx bdlock; 201 202 /* 203 * bogus page -- for I/O to/from partially complete buffers 204 * this is a temporary solution to the problem, but it is not 205 * really that bad. it would be better to split the buffer 206 * for input in the case of buffers partially already in memory, 207 * but the code is intricate enough already. 208 */ 209 vm_page_t bogus_page; 210 211 /* 212 * Synchronization (sleep/wakeup) variable for active buffer space requests. 213 * Set when wait starts, cleared prior to wakeup(). 214 * Used in runningbufwakeup() and waitrunningbufspace(). 215 */ 216 static int runningbufreq; 217 218 /* 219 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 220 * waitrunningbufspace(). 221 */ 222 static struct mtx rbreqlock; 223 224 /* 225 * Synchronization (sleep/wakeup) variable for buffer requests. 226 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 227 * by and/or. 228 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 229 * getnewbuf(), and getblk(). 230 */ 231 static int needsbuffer; 232 233 /* 234 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 235 */ 236 static struct mtx nblock; 237 238 /* 239 * Lock that protects against bwait()/bdone()/B_DONE races. 240 */ 241 242 static struct mtx bdonelock; 243 244 /* 245 * Lock that protects against bwait()/bdone()/B_DONE races. 246 */ 247 static struct mtx bpinlock; 248 249 /* 250 * Definitions for the buffer free lists. 251 */ 252 #define BUFFER_QUEUES 6 /* number of free buffer queues */ 253 254 #define QUEUE_NONE 0 /* on no queue */ 255 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 256 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 257 #define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ 258 #define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 259 #define QUEUE_EMPTY 5 /* empty buffer headers */ 260 261 /* Queues for free buffers with various properties */ 262 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 263 264 /* Lock for the bufqueues */ 265 static struct mtx bqlock; 266 267 /* 268 * Single global constant for BUF_WMESG, to avoid getting multiple references. 269 * buf_wmesg is referred from macros. 270 */ 271 const char *buf_wmesg = BUF_WMESG; 272 273 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 274 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 275 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 276 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 277 278 #ifdef DIRECTIO 279 extern void ffs_rawread_setup(void); 280 #endif /* DIRECTIO */ 281 /* 282 * numdirtywakeup: 283 * 284 * If someone is blocked due to there being too many dirty buffers, 285 * and numdirtybuffers is now reasonable, wake them up. 286 */ 287 288 static __inline void 289 numdirtywakeup(int level) 290 { 291 292 if (numdirtybuffers <= level) { 293 mtx_lock(&nblock); 294 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 295 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 296 wakeup(&needsbuffer); 297 } 298 mtx_unlock(&nblock); 299 } 300 } 301 302 /* 303 * bufspacewakeup: 304 * 305 * Called when buffer space is potentially available for recovery. 306 * getnewbuf() will block on this flag when it is unable to free 307 * sufficient buffer space. Buffer space becomes recoverable when 308 * bp's get placed back in the queues. 309 */ 310 311 static __inline void 312 bufspacewakeup(void) 313 { 314 315 /* 316 * If someone is waiting for BUF space, wake them up. Even 317 * though we haven't freed the kva space yet, the waiting 318 * process will be able to now. 319 */ 320 mtx_lock(&nblock); 321 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 322 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 323 wakeup(&needsbuffer); 324 } 325 mtx_unlock(&nblock); 326 } 327 328 /* 329 * runningbufwakeup() - in-progress I/O accounting. 330 * 331 */ 332 void 333 runningbufwakeup(struct buf *bp) 334 { 335 336 if (bp->b_runningbufspace) { 337 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 338 bp->b_runningbufspace = 0; 339 mtx_lock(&rbreqlock); 340 if (runningbufreq && runningbufspace <= lorunningspace) { 341 runningbufreq = 0; 342 wakeup(&runningbufreq); 343 } 344 mtx_unlock(&rbreqlock); 345 } 346 } 347 348 /* 349 * bufcountwakeup: 350 * 351 * Called when a buffer has been added to one of the free queues to 352 * account for the buffer and to wakeup anyone waiting for free buffers. 353 * This typically occurs when large amounts of metadata are being handled 354 * by the buffer cache ( else buffer space runs out first, usually ). 355 */ 356 357 static __inline void 358 bufcountwakeup(void) 359 { 360 361 atomic_add_int(&numfreebuffers, 1); 362 mtx_lock(&nblock); 363 if (needsbuffer) { 364 needsbuffer &= ~VFS_BIO_NEED_ANY; 365 if (numfreebuffers >= hifreebuffers) 366 needsbuffer &= ~VFS_BIO_NEED_FREE; 367 wakeup(&needsbuffer); 368 } 369 mtx_unlock(&nblock); 370 } 371 372 /* 373 * waitrunningbufspace() 374 * 375 * runningbufspace is a measure of the amount of I/O currently 376 * running. This routine is used in async-write situations to 377 * prevent creating huge backups of pending writes to a device. 378 * Only asynchronous writes are governed by this function. 379 * 380 * Reads will adjust runningbufspace, but will not block based on it. 381 * The read load has a side effect of reducing the allowed write load. 382 * 383 * This does NOT turn an async write into a sync write. It waits 384 * for earlier writes to complete and generally returns before the 385 * caller's write has reached the device. 386 */ 387 void 388 waitrunningbufspace(void) 389 { 390 391 mtx_lock(&rbreqlock); 392 while (runningbufspace > hirunningspace) { 393 ++runningbufreq; 394 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 395 } 396 mtx_unlock(&rbreqlock); 397 } 398 399 400 /* 401 * vfs_buf_test_cache: 402 * 403 * Called when a buffer is extended. This function clears the B_CACHE 404 * bit if the newly extended portion of the buffer does not contain 405 * valid data. 406 */ 407 static __inline 408 void 409 vfs_buf_test_cache(struct buf *bp, 410 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 411 vm_page_t m) 412 { 413 414 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 415 if (bp->b_flags & B_CACHE) { 416 int base = (foff + off) & PAGE_MASK; 417 if (vm_page_is_valid(m, base, size) == 0) 418 bp->b_flags &= ~B_CACHE; 419 } 420 } 421 422 /* Wake up the buffer daemon if necessary */ 423 static __inline 424 void 425 bd_wakeup(int dirtybuflevel) 426 { 427 428 mtx_lock(&bdlock); 429 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 430 bd_request = 1; 431 wakeup(&bd_request); 432 } 433 mtx_unlock(&bdlock); 434 } 435 436 /* 437 * bd_speedup - speedup the buffer cache flushing code 438 */ 439 440 static __inline 441 void 442 bd_speedup(void) 443 { 444 445 bd_wakeup(1); 446 } 447 448 /* 449 * Calculating buffer cache scaling values and reserve space for buffer 450 * headers. This is called during low level kernel initialization and 451 * may be called more then once. We CANNOT write to the memory area 452 * being reserved at this time. 453 */ 454 caddr_t 455 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 456 { 457 458 /* 459 * physmem_est is in pages. Convert it to kilobytes (assumes 460 * PAGE_SIZE is >= 1K) 461 */ 462 physmem_est = physmem_est * (PAGE_SIZE / 1024); 463 464 /* 465 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 466 * For the first 64MB of ram nominally allocate sufficient buffers to 467 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 468 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 469 * the buffer cache we limit the eventual kva reservation to 470 * maxbcache bytes. 471 * 472 * factor represents the 1/4 x ram conversion. 473 */ 474 if (nbuf == 0) { 475 int factor = 4 * BKVASIZE / 1024; 476 477 nbuf = 50; 478 if (physmem_est > 4096) 479 nbuf += min((physmem_est - 4096) / factor, 480 65536 / factor); 481 if (physmem_est > 65536) 482 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 483 484 if (maxbcache && nbuf > maxbcache / BKVASIZE) 485 nbuf = maxbcache / BKVASIZE; 486 } 487 488 #if 0 489 /* 490 * Do not allow the buffer_map to be more then 1/2 the size of the 491 * kernel_map. 492 */ 493 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 494 (BKVASIZE * 2)) { 495 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 496 (BKVASIZE * 2); 497 printf("Warning: nbufs capped at %d\n", nbuf); 498 } 499 #endif 500 501 /* 502 * swbufs are used as temporary holders for I/O, such as paging I/O. 503 * We have no less then 16 and no more then 256. 504 */ 505 nswbuf = max(min(nbuf/4, 256), 16); 506 #ifdef NSWBUF_MIN 507 if (nswbuf < NSWBUF_MIN) 508 nswbuf = NSWBUF_MIN; 509 #endif 510 #ifdef DIRECTIO 511 ffs_rawread_setup(); 512 #endif 513 514 /* 515 * Reserve space for the buffer cache buffers 516 */ 517 swbuf = (void *)v; 518 v = (caddr_t)(swbuf + nswbuf); 519 buf = (void *)v; 520 v = (caddr_t)(buf + nbuf); 521 522 return(v); 523 } 524 525 /* Initialize the buffer subsystem. Called before use of any buffers. */ 526 void 527 bufinit(void) 528 { 529 struct buf *bp; 530 int i; 531 532 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 533 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 534 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 535 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 536 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 537 mtx_init(&bpinlock, "bpin lock", NULL, MTX_DEF); 538 539 /* next, make a null set of free lists */ 540 for (i = 0; i < BUFFER_QUEUES; i++) 541 TAILQ_INIT(&bufqueues[i]); 542 543 /* finally, initialize each buffer header and stick on empty q */ 544 for (i = 0; i < nbuf; i++) { 545 bp = &buf[i]; 546 bzero(bp, sizeof *bp); 547 bp->b_flags = B_INVAL; /* we're just an empty header */ 548 bp->b_rcred = NOCRED; 549 bp->b_wcred = NOCRED; 550 bp->b_qindex = QUEUE_EMPTY; 551 bp->b_vflags = 0; 552 bp->b_xflags = 0; 553 LIST_INIT(&bp->b_dep); 554 BUF_LOCKINIT(bp); 555 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 556 } 557 558 /* 559 * maxbufspace is the absolute maximum amount of buffer space we are 560 * allowed to reserve in KVM and in real terms. The absolute maximum 561 * is nominally used by buf_daemon. hibufspace is the nominal maximum 562 * used by most other processes. The differential is required to 563 * ensure that buf_daemon is able to run when other processes might 564 * be blocked waiting for buffer space. 565 * 566 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 567 * this may result in KVM fragmentation which is not handled optimally 568 * by the system. 569 */ 570 maxbufspace = nbuf * BKVASIZE; 571 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 572 lobufspace = hibufspace - MAXBSIZE; 573 574 lorunningspace = 512 * 1024; 575 hirunningspace = 1024 * 1024; 576 577 /* 578 * Limit the amount of malloc memory since it is wired permanently into 579 * the kernel space. Even though this is accounted for in the buffer 580 * allocation, we don't want the malloced region to grow uncontrolled. 581 * The malloc scheme improves memory utilization significantly on average 582 * (small) directories. 583 */ 584 maxbufmallocspace = hibufspace / 20; 585 586 /* 587 * Reduce the chance of a deadlock occuring by limiting the number 588 * of delayed-write dirty buffers we allow to stack up. 589 */ 590 hidirtybuffers = nbuf / 4 + 20; 591 dirtybufthresh = hidirtybuffers * 9 / 10; 592 numdirtybuffers = 0; 593 /* 594 * To support extreme low-memory systems, make sure hidirtybuffers cannot 595 * eat up all available buffer space. This occurs when our minimum cannot 596 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 597 * BKVASIZE'd (8K) buffers. 598 */ 599 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 600 hidirtybuffers >>= 1; 601 } 602 lodirtybuffers = hidirtybuffers / 2; 603 604 /* 605 * Try to keep the number of free buffers in the specified range, 606 * and give special processes (e.g. like buf_daemon) access to an 607 * emergency reserve. 608 */ 609 lofreebuffers = nbuf / 18 + 5; 610 hifreebuffers = 2 * lofreebuffers; 611 numfreebuffers = nbuf; 612 613 /* 614 * Maximum number of async ops initiated per buf_daemon loop. This is 615 * somewhat of a hack at the moment, we really need to limit ourselves 616 * based on the number of bytes of I/O in-transit that were initiated 617 * from buf_daemon. 618 */ 619 620 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 621 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 622 } 623 624 /* 625 * bfreekva() - free the kva allocation for a buffer. 626 * 627 * Since this call frees up buffer space, we call bufspacewakeup(). 628 */ 629 static void 630 bfreekva(struct buf *bp) 631 { 632 633 if (bp->b_kvasize) { 634 atomic_add_int(&buffreekvacnt, 1); 635 atomic_subtract_int(&bufspace, bp->b_kvasize); 636 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 637 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 638 bp->b_kvasize = 0; 639 bufspacewakeup(); 640 } 641 } 642 643 /* 644 * bremfree: 645 * 646 * Mark the buffer for removal from the appropriate free list in brelse. 647 * 648 */ 649 void 650 bremfree(struct buf *bp) 651 { 652 653 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 654 KASSERT(BUF_REFCNT(bp), ("bremfree: buf must be locked.")); 655 KASSERT((bp->b_flags & B_REMFREE) == 0, 656 ("bremfree: buffer %p already marked for delayed removal.", bp)); 657 KASSERT(bp->b_qindex != QUEUE_NONE, 658 ("bremfree: buffer %p not on a queue.", bp)); 659 660 bp->b_flags |= B_REMFREE; 661 /* Fixup numfreebuffers count. */ 662 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 663 atomic_subtract_int(&numfreebuffers, 1); 664 } 665 666 /* 667 * bremfreef: 668 * 669 * Force an immediate removal from a free list. Used only in nfs when 670 * it abuses the b_freelist pointer. 671 */ 672 void 673 bremfreef(struct buf *bp) 674 { 675 mtx_lock(&bqlock); 676 bremfreel(bp); 677 mtx_unlock(&bqlock); 678 } 679 680 /* 681 * bremfreel: 682 * 683 * Removes a buffer from the free list, must be called with the 684 * bqlock held. 685 */ 686 static void 687 bremfreel(struct buf *bp) 688 { 689 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 690 bp, bp->b_vp, bp->b_flags); 691 KASSERT(BUF_REFCNT(bp), ("bremfreel: buffer %p not locked.", bp)); 692 KASSERT(bp->b_qindex != QUEUE_NONE, 693 ("bremfreel: buffer %p not on a queue.", bp)); 694 mtx_assert(&bqlock, MA_OWNED); 695 696 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 697 bp->b_qindex = QUEUE_NONE; 698 /* 699 * If this was a delayed bremfree() we only need to remove the buffer 700 * from the queue and return the stats are already done. 701 */ 702 if (bp->b_flags & B_REMFREE) { 703 bp->b_flags &= ~B_REMFREE; 704 return; 705 } 706 /* 707 * Fixup numfreebuffers count. If the buffer is invalid or not 708 * delayed-write, the buffer was free and we must decrement 709 * numfreebuffers. 710 */ 711 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 712 atomic_subtract_int(&numfreebuffers, 1); 713 } 714 715 716 /* 717 * Get a buffer with the specified data. Look in the cache first. We 718 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 719 * is set, the buffer is valid and we do not have to do anything ( see 720 * getblk() ). This is really just a special case of breadn(). 721 */ 722 int 723 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 724 struct buf **bpp) 725 { 726 727 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 728 } 729 730 /* 731 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 732 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 733 * the buffer is valid and we do not have to do anything. 734 */ 735 void 736 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 737 int cnt, struct ucred * cred) 738 { 739 struct buf *rabp; 740 int i; 741 742 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 743 if (inmem(vp, *rablkno)) 744 continue; 745 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 746 747 if ((rabp->b_flags & B_CACHE) == 0) { 748 if (!TD_IS_IDLETHREAD(curthread)) 749 curthread->td_proc->p_stats->p_ru.ru_inblock++; 750 rabp->b_flags |= B_ASYNC; 751 rabp->b_flags &= ~B_INVAL; 752 rabp->b_ioflags &= ~BIO_ERROR; 753 rabp->b_iocmd = BIO_READ; 754 if (rabp->b_rcred == NOCRED && cred != NOCRED) 755 rabp->b_rcred = crhold(cred); 756 vfs_busy_pages(rabp, 0); 757 BUF_KERNPROC(rabp); 758 rabp->b_iooffset = dbtob(rabp->b_blkno); 759 bstrategy(rabp); 760 } else { 761 brelse(rabp); 762 } 763 } 764 } 765 766 /* 767 * Operates like bread, but also starts asynchronous I/O on 768 * read-ahead blocks. 769 */ 770 int 771 breadn(struct vnode * vp, daddr_t blkno, int size, 772 daddr_t * rablkno, int *rabsize, 773 int cnt, struct ucred * cred, struct buf **bpp) 774 { 775 struct buf *bp; 776 int rv = 0, readwait = 0; 777 778 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 779 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 780 781 /* if not found in cache, do some I/O */ 782 if ((bp->b_flags & B_CACHE) == 0) { 783 if (!TD_IS_IDLETHREAD(curthread)) 784 curthread->td_proc->p_stats->p_ru.ru_inblock++; 785 bp->b_iocmd = BIO_READ; 786 bp->b_flags &= ~B_INVAL; 787 bp->b_ioflags &= ~BIO_ERROR; 788 if (bp->b_rcred == NOCRED && cred != NOCRED) 789 bp->b_rcred = crhold(cred); 790 vfs_busy_pages(bp, 0); 791 bp->b_iooffset = dbtob(bp->b_blkno); 792 bstrategy(bp); 793 ++readwait; 794 } 795 796 breada(vp, rablkno, rabsize, cnt, cred); 797 798 if (readwait) { 799 rv = bufwait(bp); 800 } 801 return (rv); 802 } 803 804 /* 805 * Write, release buffer on completion. (Done by iodone 806 * if async). Do not bother writing anything if the buffer 807 * is invalid. 808 * 809 * Note that we set B_CACHE here, indicating that buffer is 810 * fully valid and thus cacheable. This is true even of NFS 811 * now so we set it generally. This could be set either here 812 * or in biodone() since the I/O is synchronous. We put it 813 * here. 814 */ 815 int 816 bufwrite(struct buf *bp) 817 { 818 int oldflags; 819 struct vnode *vp; 820 int vp_md; 821 822 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 823 if (bp->b_flags & B_INVAL) { 824 brelse(bp); 825 return (0); 826 } 827 828 oldflags = bp->b_flags; 829 830 if (BUF_REFCNT(bp) == 0) 831 panic("bufwrite: buffer is not busy???"); 832 833 if (bp->b_pin_count > 0) 834 bunpin_wait(bp); 835 836 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 837 ("FFS background buffer should not get here %p", bp)); 838 839 vp = bp->b_vp; 840 if (vp) 841 vp_md = vp->v_vflag & VV_MD; 842 else 843 vp_md = 0; 844 845 /* Mark the buffer clean */ 846 bundirty(bp); 847 848 bp->b_flags &= ~B_DONE; 849 bp->b_ioflags &= ~BIO_ERROR; 850 bp->b_flags |= B_CACHE; 851 bp->b_iocmd = BIO_WRITE; 852 853 bufobj_wref(bp->b_bufobj); 854 vfs_busy_pages(bp, 1); 855 856 /* 857 * Normal bwrites pipeline writes 858 */ 859 bp->b_runningbufspace = bp->b_bufsize; 860 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 861 862 if (!TD_IS_IDLETHREAD(curthread)) 863 curthread->td_proc->p_stats->p_ru.ru_oublock++; 864 if (oldflags & B_ASYNC) 865 BUF_KERNPROC(bp); 866 bp->b_iooffset = dbtob(bp->b_blkno); 867 bstrategy(bp); 868 869 if ((oldflags & B_ASYNC) == 0) { 870 int rtval = bufwait(bp); 871 brelse(bp); 872 return (rtval); 873 } else { 874 /* 875 * don't allow the async write to saturate the I/O 876 * system. We will not deadlock here because 877 * we are blocking waiting for I/O that is already in-progress 878 * to complete. We do not block here if it is the update 879 * or syncer daemon trying to clean up as that can lead 880 * to deadlock. 881 */ 882 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 883 waitrunningbufspace(); 884 } 885 886 return (0); 887 } 888 889 void 890 bufbdflush(struct bufobj *bo, struct buf *bp) 891 { 892 struct buf *nbp; 893 894 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 895 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 896 altbufferflushes++; 897 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 898 BO_LOCK(bo); 899 /* 900 * Try to find a buffer to flush. 901 */ 902 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 903 if ((nbp->b_vflags & BV_BKGRDINPROG) || 904 BUF_LOCK(nbp, 905 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 906 continue; 907 if (bp == nbp) 908 panic("bdwrite: found ourselves"); 909 BO_UNLOCK(bo); 910 /* Don't countdeps with the bo lock held. */ 911 if (buf_countdeps(nbp, 0)) { 912 BO_LOCK(bo); 913 BUF_UNLOCK(nbp); 914 continue; 915 } 916 if (nbp->b_flags & B_CLUSTEROK) { 917 vfs_bio_awrite(nbp); 918 } else { 919 bremfree(nbp); 920 bawrite(nbp); 921 } 922 dirtybufferflushes++; 923 break; 924 } 925 if (nbp == NULL) 926 BO_UNLOCK(bo); 927 } 928 } 929 930 /* 931 * Delayed write. (Buffer is marked dirty). Do not bother writing 932 * anything if the buffer is marked invalid. 933 * 934 * Note that since the buffer must be completely valid, we can safely 935 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 936 * biodone() in order to prevent getblk from writing the buffer 937 * out synchronously. 938 */ 939 void 940 bdwrite(struct buf *bp) 941 { 942 struct thread *td = curthread; 943 struct vnode *vp; 944 struct bufobj *bo; 945 946 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 947 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 948 KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy")); 949 950 if (bp->b_flags & B_INVAL) { 951 brelse(bp); 952 return; 953 } 954 955 /* 956 * If we have too many dirty buffers, don't create any more. 957 * If we are wildly over our limit, then force a complete 958 * cleanup. Otherwise, just keep the situation from getting 959 * out of control. Note that we have to avoid a recursive 960 * disaster and not try to clean up after our own cleanup! 961 */ 962 vp = bp->b_vp; 963 bo = bp->b_bufobj; 964 if ((td->td_pflags & TDP_COWINPROGRESS) == 0) 965 BO_BDFLUSH(bo, bp); 966 else 967 recursiveflushes++; 968 969 bdirty(bp); 970 /* 971 * Set B_CACHE, indicating that the buffer is fully valid. This is 972 * true even of NFS now. 973 */ 974 bp->b_flags |= B_CACHE; 975 976 /* 977 * This bmap keeps the system from needing to do the bmap later, 978 * perhaps when the system is attempting to do a sync. Since it 979 * is likely that the indirect block -- or whatever other datastructure 980 * that the filesystem needs is still in memory now, it is a good 981 * thing to do this. Note also, that if the pageout daemon is 982 * requesting a sync -- there might not be enough memory to do 983 * the bmap then... So, this is important to do. 984 */ 985 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 986 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 987 } 988 989 /* 990 * Set the *dirty* buffer range based upon the VM system dirty pages. 991 */ 992 vfs_setdirty(bp); 993 994 /* 995 * We need to do this here to satisfy the vnode_pager and the 996 * pageout daemon, so that it thinks that the pages have been 997 * "cleaned". Note that since the pages are in a delayed write 998 * buffer -- the VFS layer "will" see that the pages get written 999 * out on the next sync, or perhaps the cluster will be completed. 1000 */ 1001 vfs_clean_pages(bp); 1002 bqrelse(bp); 1003 1004 /* 1005 * Wakeup the buffer flushing daemon if we have a lot of dirty 1006 * buffers (midpoint between our recovery point and our stall 1007 * point). 1008 */ 1009 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1010 1011 /* 1012 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1013 * due to the softdep code. 1014 */ 1015 } 1016 1017 /* 1018 * bdirty: 1019 * 1020 * Turn buffer into delayed write request. We must clear BIO_READ and 1021 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1022 * itself to properly update it in the dirty/clean lists. We mark it 1023 * B_DONE to ensure that any asynchronization of the buffer properly 1024 * clears B_DONE ( else a panic will occur later ). 1025 * 1026 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1027 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1028 * should only be called if the buffer is known-good. 1029 * 1030 * Since the buffer is not on a queue, we do not update the numfreebuffers 1031 * count. 1032 * 1033 * The buffer must be on QUEUE_NONE. 1034 */ 1035 void 1036 bdirty(struct buf *bp) 1037 { 1038 1039 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1040 bp, bp->b_vp, bp->b_flags); 1041 KASSERT(BUF_REFCNT(bp) == 1, ("bdirty: bp %p not locked",bp)); 1042 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1043 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1044 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1045 bp->b_flags &= ~(B_RELBUF); 1046 bp->b_iocmd = BIO_WRITE; 1047 1048 if ((bp->b_flags & B_DELWRI) == 0) { 1049 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1050 reassignbuf(bp); 1051 atomic_add_int(&numdirtybuffers, 1); 1052 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1053 } 1054 } 1055 1056 /* 1057 * bundirty: 1058 * 1059 * Clear B_DELWRI for buffer. 1060 * 1061 * Since the buffer is not on a queue, we do not update the numfreebuffers 1062 * count. 1063 * 1064 * The buffer must be on QUEUE_NONE. 1065 */ 1066 1067 void 1068 bundirty(struct buf *bp) 1069 { 1070 1071 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1072 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1073 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1074 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1075 KASSERT(BUF_REFCNT(bp) == 1, ("bundirty: bp %p not locked",bp)); 1076 1077 if (bp->b_flags & B_DELWRI) { 1078 bp->b_flags &= ~B_DELWRI; 1079 reassignbuf(bp); 1080 atomic_subtract_int(&numdirtybuffers, 1); 1081 numdirtywakeup(lodirtybuffers); 1082 } 1083 /* 1084 * Since it is now being written, we can clear its deferred write flag. 1085 */ 1086 bp->b_flags &= ~B_DEFERRED; 1087 } 1088 1089 /* 1090 * bawrite: 1091 * 1092 * Asynchronous write. Start output on a buffer, but do not wait for 1093 * it to complete. The buffer is released when the output completes. 1094 * 1095 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1096 * B_INVAL buffers. Not us. 1097 */ 1098 void 1099 bawrite(struct buf *bp) 1100 { 1101 1102 bp->b_flags |= B_ASYNC; 1103 (void) bwrite(bp); 1104 } 1105 1106 /* 1107 * bwillwrite: 1108 * 1109 * Called prior to the locking of any vnodes when we are expecting to 1110 * write. We do not want to starve the buffer cache with too many 1111 * dirty buffers so we block here. By blocking prior to the locking 1112 * of any vnodes we attempt to avoid the situation where a locked vnode 1113 * prevents the various system daemons from flushing related buffers. 1114 */ 1115 1116 void 1117 bwillwrite(void) 1118 { 1119 1120 if (numdirtybuffers >= hidirtybuffers) { 1121 mtx_lock(&nblock); 1122 while (numdirtybuffers >= hidirtybuffers) { 1123 bd_wakeup(1); 1124 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1125 msleep(&needsbuffer, &nblock, 1126 (PRIBIO + 4), "flswai", 0); 1127 } 1128 mtx_unlock(&nblock); 1129 } 1130 } 1131 1132 /* 1133 * Return true if we have too many dirty buffers. 1134 */ 1135 int 1136 buf_dirty_count_severe(void) 1137 { 1138 1139 return(numdirtybuffers >= hidirtybuffers); 1140 } 1141 1142 /* 1143 * brelse: 1144 * 1145 * Release a busy buffer and, if requested, free its resources. The 1146 * buffer will be stashed in the appropriate bufqueue[] allowing it 1147 * to be accessed later as a cache entity or reused for other purposes. 1148 */ 1149 void 1150 brelse(struct buf *bp) 1151 { 1152 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1153 bp, bp->b_vp, bp->b_flags); 1154 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1155 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1156 1157 if (bp->b_flags & B_MANAGED) { 1158 bqrelse(bp); 1159 return; 1160 } 1161 1162 if (bp->b_iocmd == BIO_WRITE && 1163 (bp->b_ioflags & BIO_ERROR) && 1164 !(bp->b_flags & B_INVAL)) { 1165 /* 1166 * Failed write, redirty. Must clear BIO_ERROR to prevent 1167 * pages from being scrapped. If B_INVAL is set then 1168 * this case is not run and the next case is run to 1169 * destroy the buffer. B_INVAL can occur if the buffer 1170 * is outside the range supported by the underlying device. 1171 */ 1172 bp->b_ioflags &= ~BIO_ERROR; 1173 bdirty(bp); 1174 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1175 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1176 /* 1177 * Either a failed I/O or we were asked to free or not 1178 * cache the buffer. 1179 */ 1180 bp->b_flags |= B_INVAL; 1181 if (!LIST_EMPTY(&bp->b_dep)) 1182 buf_deallocate(bp); 1183 if (bp->b_flags & B_DELWRI) { 1184 atomic_subtract_int(&numdirtybuffers, 1); 1185 numdirtywakeup(lodirtybuffers); 1186 } 1187 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1188 if ((bp->b_flags & B_VMIO) == 0) { 1189 if (bp->b_bufsize) 1190 allocbuf(bp, 0); 1191 if (bp->b_vp) 1192 brelvp(bp); 1193 } 1194 } 1195 1196 /* 1197 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1198 * is called with B_DELWRI set, the underlying pages may wind up 1199 * getting freed causing a previous write (bdwrite()) to get 'lost' 1200 * because pages associated with a B_DELWRI bp are marked clean. 1201 * 1202 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1203 * if B_DELWRI is set. 1204 * 1205 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1206 * on pages to return pages to the VM page queues. 1207 */ 1208 if (bp->b_flags & B_DELWRI) 1209 bp->b_flags &= ~B_RELBUF; 1210 else if (vm_page_count_severe()) { 1211 /* 1212 * XXX This lock may not be necessary since BKGRDINPROG 1213 * cannot be set while we hold the buf lock, it can only be 1214 * cleared if it is already pending. 1215 */ 1216 if (bp->b_vp) { 1217 BO_LOCK(bp->b_bufobj); 1218 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1219 bp->b_flags |= B_RELBUF; 1220 BO_UNLOCK(bp->b_bufobj); 1221 } else 1222 bp->b_flags |= B_RELBUF; 1223 } 1224 1225 /* 1226 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1227 * constituted, not even NFS buffers now. Two flags effect this. If 1228 * B_INVAL, the struct buf is invalidated but the VM object is kept 1229 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1230 * 1231 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1232 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1233 * buffer is also B_INVAL because it hits the re-dirtying code above. 1234 * 1235 * Normally we can do this whether a buffer is B_DELWRI or not. If 1236 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1237 * the commit state and we cannot afford to lose the buffer. If the 1238 * buffer has a background write in progress, we need to keep it 1239 * around to prevent it from being reconstituted and starting a second 1240 * background write. 1241 */ 1242 if ((bp->b_flags & B_VMIO) 1243 && !(bp->b_vp->v_mount != NULL && 1244 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1245 !vn_isdisk(bp->b_vp, NULL) && 1246 (bp->b_flags & B_DELWRI)) 1247 ) { 1248 1249 int i, j, resid; 1250 vm_page_t m; 1251 off_t foff; 1252 vm_pindex_t poff; 1253 vm_object_t obj; 1254 1255 obj = bp->b_bufobj->bo_object; 1256 1257 /* 1258 * Get the base offset and length of the buffer. Note that 1259 * in the VMIO case if the buffer block size is not 1260 * page-aligned then b_data pointer may not be page-aligned. 1261 * But our b_pages[] array *IS* page aligned. 1262 * 1263 * block sizes less then DEV_BSIZE (usually 512) are not 1264 * supported due to the page granularity bits (m->valid, 1265 * m->dirty, etc...). 1266 * 1267 * See man buf(9) for more information 1268 */ 1269 resid = bp->b_bufsize; 1270 foff = bp->b_offset; 1271 VM_OBJECT_LOCK(obj); 1272 for (i = 0; i < bp->b_npages; i++) { 1273 int had_bogus = 0; 1274 1275 m = bp->b_pages[i]; 1276 1277 /* 1278 * If we hit a bogus page, fixup *all* the bogus pages 1279 * now. 1280 */ 1281 if (m == bogus_page) { 1282 poff = OFF_TO_IDX(bp->b_offset); 1283 had_bogus = 1; 1284 1285 for (j = i; j < bp->b_npages; j++) { 1286 vm_page_t mtmp; 1287 mtmp = bp->b_pages[j]; 1288 if (mtmp == bogus_page) { 1289 mtmp = vm_page_lookup(obj, poff + j); 1290 if (!mtmp) { 1291 panic("brelse: page missing\n"); 1292 } 1293 bp->b_pages[j] = mtmp; 1294 } 1295 } 1296 1297 if ((bp->b_flags & B_INVAL) == 0) { 1298 pmap_qenter( 1299 trunc_page((vm_offset_t)bp->b_data), 1300 bp->b_pages, bp->b_npages); 1301 } 1302 m = bp->b_pages[i]; 1303 } 1304 if ((bp->b_flags & B_NOCACHE) || 1305 (bp->b_ioflags & BIO_ERROR)) { 1306 int poffset = foff & PAGE_MASK; 1307 int presid = resid > (PAGE_SIZE - poffset) ? 1308 (PAGE_SIZE - poffset) : resid; 1309 1310 KASSERT(presid >= 0, ("brelse: extra page")); 1311 vm_page_lock_queues(); 1312 vm_page_set_invalid(m, poffset, presid); 1313 vm_page_unlock_queues(); 1314 if (had_bogus) 1315 printf("avoided corruption bug in bogus_page/brelse code\n"); 1316 } 1317 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1318 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1319 } 1320 VM_OBJECT_UNLOCK(obj); 1321 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1322 vfs_vmio_release(bp); 1323 1324 } else if (bp->b_flags & B_VMIO) { 1325 1326 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1327 vfs_vmio_release(bp); 1328 } 1329 1330 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1331 if (bp->b_bufsize != 0) 1332 allocbuf(bp, 0); 1333 if (bp->b_vp != NULL) 1334 brelvp(bp); 1335 } 1336 1337 if (BUF_REFCNT(bp) > 1) { 1338 /* do not release to free list */ 1339 BUF_UNLOCK(bp); 1340 return; 1341 } 1342 1343 /* enqueue */ 1344 mtx_lock(&bqlock); 1345 /* Handle delayed bremfree() processing. */ 1346 if (bp->b_flags & B_REMFREE) 1347 bremfreel(bp); 1348 if (bp->b_qindex != QUEUE_NONE) 1349 panic("brelse: free buffer onto another queue???"); 1350 1351 /* buffers with no memory */ 1352 if (bp->b_bufsize == 0) { 1353 bp->b_flags |= B_INVAL; 1354 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1355 if (bp->b_vflags & BV_BKGRDINPROG) 1356 panic("losing buffer 1"); 1357 if (bp->b_kvasize) { 1358 bp->b_qindex = QUEUE_EMPTYKVA; 1359 } else { 1360 bp->b_qindex = QUEUE_EMPTY; 1361 } 1362 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1363 /* buffers with junk contents */ 1364 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1365 (bp->b_ioflags & BIO_ERROR)) { 1366 bp->b_flags |= B_INVAL; 1367 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1368 if (bp->b_vflags & BV_BKGRDINPROG) 1369 panic("losing buffer 2"); 1370 bp->b_qindex = QUEUE_CLEAN; 1371 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1372 /* remaining buffers */ 1373 } else { 1374 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1375 (B_DELWRI|B_NEEDSGIANT)) 1376 bp->b_qindex = QUEUE_DIRTY_GIANT; 1377 if (bp->b_flags & B_DELWRI) 1378 bp->b_qindex = QUEUE_DIRTY; 1379 else 1380 bp->b_qindex = QUEUE_CLEAN; 1381 if (bp->b_flags & B_AGE) 1382 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1383 else 1384 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1385 } 1386 mtx_unlock(&bqlock); 1387 1388 /* 1389 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1390 * placed the buffer on the correct queue. We must also disassociate 1391 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1392 * find it. 1393 */ 1394 if (bp->b_flags & B_INVAL) { 1395 if (bp->b_flags & B_DELWRI) 1396 bundirty(bp); 1397 if (bp->b_vp) 1398 brelvp(bp); 1399 } 1400 1401 /* 1402 * Fixup numfreebuffers count. The bp is on an appropriate queue 1403 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1404 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1405 * if B_INVAL is set ). 1406 */ 1407 1408 if (!(bp->b_flags & B_DELWRI)) 1409 bufcountwakeup(); 1410 1411 /* 1412 * Something we can maybe free or reuse 1413 */ 1414 if (bp->b_bufsize || bp->b_kvasize) 1415 bufspacewakeup(); 1416 1417 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1418 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1419 panic("brelse: not dirty"); 1420 /* unlock */ 1421 BUF_UNLOCK(bp); 1422 } 1423 1424 /* 1425 * Release a buffer back to the appropriate queue but do not try to free 1426 * it. The buffer is expected to be used again soon. 1427 * 1428 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1429 * biodone() to requeue an async I/O on completion. It is also used when 1430 * known good buffers need to be requeued but we think we may need the data 1431 * again soon. 1432 * 1433 * XXX we should be able to leave the B_RELBUF hint set on completion. 1434 */ 1435 void 1436 bqrelse(struct buf *bp) 1437 { 1438 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1439 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1440 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1441 1442 if (BUF_REFCNT(bp) > 1) { 1443 /* do not release to free list */ 1444 BUF_UNLOCK(bp); 1445 return; 1446 } 1447 1448 if (bp->b_flags & B_MANAGED) { 1449 if (bp->b_flags & B_REMFREE) { 1450 mtx_lock(&bqlock); 1451 bremfreel(bp); 1452 mtx_unlock(&bqlock); 1453 } 1454 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1455 BUF_UNLOCK(bp); 1456 return; 1457 } 1458 1459 mtx_lock(&bqlock); 1460 /* Handle delayed bremfree() processing. */ 1461 if (bp->b_flags & B_REMFREE) 1462 bremfreel(bp); 1463 if (bp->b_qindex != QUEUE_NONE) 1464 panic("bqrelse: free buffer onto another queue???"); 1465 /* buffers with stale but valid contents */ 1466 if (bp->b_flags & B_DELWRI) { 1467 if (bp->b_flags & B_NEEDSGIANT) 1468 bp->b_qindex = QUEUE_DIRTY_GIANT; 1469 else 1470 bp->b_qindex = QUEUE_DIRTY; 1471 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1472 } else { 1473 /* 1474 * XXX This lock may not be necessary since BKGRDINPROG 1475 * cannot be set while we hold the buf lock, it can only be 1476 * cleared if it is already pending. 1477 */ 1478 BO_LOCK(bp->b_bufobj); 1479 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1480 BO_UNLOCK(bp->b_bufobj); 1481 bp->b_qindex = QUEUE_CLEAN; 1482 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1483 b_freelist); 1484 } else { 1485 /* 1486 * We are too low on memory, we have to try to free 1487 * the buffer (most importantly: the wired pages 1488 * making up its backing store) *now*. 1489 */ 1490 BO_UNLOCK(bp->b_bufobj); 1491 mtx_unlock(&bqlock); 1492 brelse(bp); 1493 return; 1494 } 1495 } 1496 mtx_unlock(&bqlock); 1497 1498 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1499 bufcountwakeup(); 1500 1501 /* 1502 * Something we can maybe free or reuse. 1503 */ 1504 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1505 bufspacewakeup(); 1506 1507 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1508 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1509 panic("bqrelse: not dirty"); 1510 /* unlock */ 1511 BUF_UNLOCK(bp); 1512 } 1513 1514 /* Give pages used by the bp back to the VM system (where possible) */ 1515 static void 1516 vfs_vmio_release(struct buf *bp) 1517 { 1518 int i; 1519 vm_page_t m; 1520 1521 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1522 vm_page_lock_queues(); 1523 for (i = 0; i < bp->b_npages; i++) { 1524 m = bp->b_pages[i]; 1525 bp->b_pages[i] = NULL; 1526 /* 1527 * In order to keep page LRU ordering consistent, put 1528 * everything on the inactive queue. 1529 */ 1530 vm_page_unwire(m, 0); 1531 /* 1532 * We don't mess with busy pages, it is 1533 * the responsibility of the process that 1534 * busied the pages to deal with them. 1535 */ 1536 if ((m->oflags & VPO_BUSY) || (m->busy != 0)) 1537 continue; 1538 1539 if (m->wire_count == 0) { 1540 /* 1541 * Might as well free the page if we can and it has 1542 * no valid data. We also free the page if the 1543 * buffer was used for direct I/O 1544 */ 1545 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1546 m->hold_count == 0) { 1547 vm_page_free(m); 1548 } else if (bp->b_flags & B_DIRECT) { 1549 vm_page_try_to_free(m); 1550 } else if (vm_page_count_severe()) { 1551 vm_page_try_to_cache(m); 1552 } 1553 } 1554 } 1555 vm_page_unlock_queues(); 1556 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1557 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1558 1559 if (bp->b_bufsize) { 1560 bufspacewakeup(); 1561 bp->b_bufsize = 0; 1562 } 1563 bp->b_npages = 0; 1564 bp->b_flags &= ~B_VMIO; 1565 if (bp->b_vp) 1566 brelvp(bp); 1567 } 1568 1569 /* 1570 * Check to see if a block at a particular lbn is available for a clustered 1571 * write. 1572 */ 1573 static int 1574 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1575 { 1576 struct buf *bpa; 1577 int match; 1578 1579 match = 0; 1580 1581 /* If the buf isn't in core skip it */ 1582 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1583 return (0); 1584 1585 /* If the buf is busy we don't want to wait for it */ 1586 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1587 return (0); 1588 1589 /* Only cluster with valid clusterable delayed write buffers */ 1590 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1591 (B_DELWRI | B_CLUSTEROK)) 1592 goto done; 1593 1594 if (bpa->b_bufsize != size) 1595 goto done; 1596 1597 /* 1598 * Check to see if it is in the expected place on disk and that the 1599 * block has been mapped. 1600 */ 1601 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1602 match = 1; 1603 done: 1604 BUF_UNLOCK(bpa); 1605 return (match); 1606 } 1607 1608 /* 1609 * vfs_bio_awrite: 1610 * 1611 * Implement clustered async writes for clearing out B_DELWRI buffers. 1612 * This is much better then the old way of writing only one buffer at 1613 * a time. Note that we may not be presented with the buffers in the 1614 * correct order, so we search for the cluster in both directions. 1615 */ 1616 int 1617 vfs_bio_awrite(struct buf *bp) 1618 { 1619 int i; 1620 int j; 1621 daddr_t lblkno = bp->b_lblkno; 1622 struct vnode *vp = bp->b_vp; 1623 int ncl; 1624 int nwritten; 1625 int size; 1626 int maxcl; 1627 1628 /* 1629 * right now we support clustered writing only to regular files. If 1630 * we find a clusterable block we could be in the middle of a cluster 1631 * rather then at the beginning. 1632 */ 1633 if ((vp->v_type == VREG) && 1634 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1635 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1636 1637 size = vp->v_mount->mnt_stat.f_iosize; 1638 maxcl = MAXPHYS / size; 1639 1640 VI_LOCK(vp); 1641 for (i = 1; i < maxcl; i++) 1642 if (vfs_bio_clcheck(vp, size, lblkno + i, 1643 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1644 break; 1645 1646 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1647 if (vfs_bio_clcheck(vp, size, lblkno - j, 1648 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1649 break; 1650 1651 VI_UNLOCK(vp); 1652 --j; 1653 ncl = i + j; 1654 /* 1655 * this is a possible cluster write 1656 */ 1657 if (ncl != 1) { 1658 BUF_UNLOCK(bp); 1659 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1660 return nwritten; 1661 } 1662 } 1663 bremfree(bp); 1664 bp->b_flags |= B_ASYNC; 1665 /* 1666 * default (old) behavior, writing out only one block 1667 * 1668 * XXX returns b_bufsize instead of b_bcount for nwritten? 1669 */ 1670 nwritten = bp->b_bufsize; 1671 (void) bwrite(bp); 1672 1673 return nwritten; 1674 } 1675 1676 /* 1677 * getnewbuf: 1678 * 1679 * Find and initialize a new buffer header, freeing up existing buffers 1680 * in the bufqueues as necessary. The new buffer is returned locked. 1681 * 1682 * Important: B_INVAL is not set. If the caller wishes to throw the 1683 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1684 * 1685 * We block if: 1686 * We have insufficient buffer headers 1687 * We have insufficient buffer space 1688 * buffer_map is too fragmented ( space reservation fails ) 1689 * If we have to flush dirty buffers ( but we try to avoid this ) 1690 * 1691 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1692 * Instead we ask the buf daemon to do it for us. We attempt to 1693 * avoid piecemeal wakeups of the pageout daemon. 1694 */ 1695 1696 static struct buf * 1697 getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1698 { 1699 struct buf *bp; 1700 struct buf *nbp; 1701 int defrag = 0; 1702 int nqindex; 1703 static int flushingbufs; 1704 1705 /* 1706 * We can't afford to block since we might be holding a vnode lock, 1707 * which may prevent system daemons from running. We deal with 1708 * low-memory situations by proactively returning memory and running 1709 * async I/O rather then sync I/O. 1710 */ 1711 1712 atomic_add_int(&getnewbufcalls, 1); 1713 atomic_subtract_int(&getnewbufrestarts, 1); 1714 restart: 1715 atomic_add_int(&getnewbufrestarts, 1); 1716 1717 /* 1718 * Setup for scan. If we do not have enough free buffers, 1719 * we setup a degenerate case that immediately fails. Note 1720 * that if we are specially marked process, we are allowed to 1721 * dip into our reserves. 1722 * 1723 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1724 * 1725 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1726 * However, there are a number of cases (defragging, reusing, ...) 1727 * where we cannot backup. 1728 */ 1729 mtx_lock(&bqlock); 1730 nqindex = QUEUE_EMPTYKVA; 1731 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1732 1733 if (nbp == NULL) { 1734 /* 1735 * If no EMPTYKVA buffers and we are either 1736 * defragging or reusing, locate a CLEAN buffer 1737 * to free or reuse. If bufspace useage is low 1738 * skip this step so we can allocate a new buffer. 1739 */ 1740 if (defrag || bufspace >= lobufspace) { 1741 nqindex = QUEUE_CLEAN; 1742 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1743 } 1744 1745 /* 1746 * If we could not find or were not allowed to reuse a 1747 * CLEAN buffer, check to see if it is ok to use an EMPTY 1748 * buffer. We can only use an EMPTY buffer if allocating 1749 * its KVA would not otherwise run us out of buffer space. 1750 */ 1751 if (nbp == NULL && defrag == 0 && 1752 bufspace + maxsize < hibufspace) { 1753 nqindex = QUEUE_EMPTY; 1754 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1755 } 1756 } 1757 1758 /* 1759 * Run scan, possibly freeing data and/or kva mappings on the fly 1760 * depending. 1761 */ 1762 1763 while ((bp = nbp) != NULL) { 1764 int qindex = nqindex; 1765 1766 /* 1767 * Calculate next bp ( we can only use it if we do not block 1768 * or do other fancy things ). 1769 */ 1770 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1771 switch(qindex) { 1772 case QUEUE_EMPTY: 1773 nqindex = QUEUE_EMPTYKVA; 1774 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1775 break; 1776 /* FALLTHROUGH */ 1777 case QUEUE_EMPTYKVA: 1778 nqindex = QUEUE_CLEAN; 1779 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1780 break; 1781 /* FALLTHROUGH */ 1782 case QUEUE_CLEAN: 1783 /* 1784 * nbp is NULL. 1785 */ 1786 break; 1787 } 1788 } 1789 /* 1790 * If we are defragging then we need a buffer with 1791 * b_kvasize != 0. XXX this situation should no longer 1792 * occur, if defrag is non-zero the buffer's b_kvasize 1793 * should also be non-zero at this point. XXX 1794 */ 1795 if (defrag && bp->b_kvasize == 0) { 1796 printf("Warning: defrag empty buffer %p\n", bp); 1797 continue; 1798 } 1799 1800 /* 1801 * Start freeing the bp. This is somewhat involved. nbp 1802 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1803 */ 1804 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1805 continue; 1806 if (bp->b_vp) { 1807 BO_LOCK(bp->b_bufobj); 1808 if (bp->b_vflags & BV_BKGRDINPROG) { 1809 BO_UNLOCK(bp->b_bufobj); 1810 BUF_UNLOCK(bp); 1811 continue; 1812 } 1813 BO_UNLOCK(bp->b_bufobj); 1814 } 1815 CTR6(KTR_BUF, 1816 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1817 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1818 bp->b_kvasize, bp->b_bufsize, qindex); 1819 1820 /* 1821 * Sanity Checks 1822 */ 1823 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1824 1825 /* 1826 * Note: we no longer distinguish between VMIO and non-VMIO 1827 * buffers. 1828 */ 1829 1830 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1831 1832 bremfreel(bp); 1833 mtx_unlock(&bqlock); 1834 1835 if (qindex == QUEUE_CLEAN) { 1836 if (bp->b_flags & B_VMIO) { 1837 bp->b_flags &= ~B_ASYNC; 1838 vfs_vmio_release(bp); 1839 } 1840 if (bp->b_vp) 1841 brelvp(bp); 1842 } 1843 1844 /* 1845 * NOTE: nbp is now entirely invalid. We can only restart 1846 * the scan from this point on. 1847 * 1848 * Get the rest of the buffer freed up. b_kva* is still 1849 * valid after this operation. 1850 */ 1851 1852 if (bp->b_rcred != NOCRED) { 1853 crfree(bp->b_rcred); 1854 bp->b_rcred = NOCRED; 1855 } 1856 if (bp->b_wcred != NOCRED) { 1857 crfree(bp->b_wcred); 1858 bp->b_wcred = NOCRED; 1859 } 1860 if (!LIST_EMPTY(&bp->b_dep)) 1861 buf_deallocate(bp); 1862 if (bp->b_vflags & BV_BKGRDINPROG) 1863 panic("losing buffer 3"); 1864 KASSERT(bp->b_vp == NULL, 1865 ("bp: %p still has vnode %p. qindex: %d", 1866 bp, bp->b_vp, qindex)); 1867 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1868 ("bp: %p still on a buffer list. xflags %X", 1869 bp, bp->b_xflags)); 1870 1871 if (bp->b_bufsize) 1872 allocbuf(bp, 0); 1873 1874 bp->b_flags = 0; 1875 bp->b_ioflags = 0; 1876 bp->b_xflags = 0; 1877 bp->b_vflags = 0; 1878 bp->b_vp = NULL; 1879 bp->b_blkno = bp->b_lblkno = 0; 1880 bp->b_offset = NOOFFSET; 1881 bp->b_iodone = 0; 1882 bp->b_error = 0; 1883 bp->b_resid = 0; 1884 bp->b_bcount = 0; 1885 bp->b_npages = 0; 1886 bp->b_dirtyoff = bp->b_dirtyend = 0; 1887 bp->b_bufobj = NULL; 1888 bp->b_pin_count = 0; 1889 bp->b_fsprivate1 = NULL; 1890 bp->b_fsprivate2 = NULL; 1891 bp->b_fsprivate3 = NULL; 1892 1893 LIST_INIT(&bp->b_dep); 1894 1895 /* 1896 * If we are defragging then free the buffer. 1897 */ 1898 if (defrag) { 1899 bp->b_flags |= B_INVAL; 1900 bfreekva(bp); 1901 brelse(bp); 1902 defrag = 0; 1903 goto restart; 1904 } 1905 1906 /* 1907 * Notify any waiters for the buffer lock about 1908 * identity change by freeing the buffer. 1909 */ 1910 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp) > 0) { 1911 bp->b_flags |= B_INVAL; 1912 bfreekva(bp); 1913 brelse(bp); 1914 goto restart; 1915 } 1916 1917 /* 1918 * If we are overcomitted then recover the buffer and its 1919 * KVM space. This occurs in rare situations when multiple 1920 * processes are blocked in getnewbuf() or allocbuf(). 1921 */ 1922 if (bufspace >= hibufspace) 1923 flushingbufs = 1; 1924 if (flushingbufs && bp->b_kvasize != 0) { 1925 bp->b_flags |= B_INVAL; 1926 bfreekva(bp); 1927 brelse(bp); 1928 goto restart; 1929 } 1930 if (bufspace < lobufspace) 1931 flushingbufs = 0; 1932 break; 1933 } 1934 1935 /* 1936 * If we exhausted our list, sleep as appropriate. We may have to 1937 * wakeup various daemons and write out some dirty buffers. 1938 * 1939 * Generally we are sleeping due to insufficient buffer space. 1940 */ 1941 1942 if (bp == NULL) { 1943 int flags; 1944 char *waitmsg; 1945 1946 if (defrag) { 1947 flags = VFS_BIO_NEED_BUFSPACE; 1948 waitmsg = "nbufkv"; 1949 } else if (bufspace >= hibufspace) { 1950 waitmsg = "nbufbs"; 1951 flags = VFS_BIO_NEED_BUFSPACE; 1952 } else { 1953 waitmsg = "newbuf"; 1954 flags = VFS_BIO_NEED_ANY; 1955 } 1956 mtx_lock(&nblock); 1957 needsbuffer |= flags; 1958 mtx_unlock(&nblock); 1959 mtx_unlock(&bqlock); 1960 1961 bd_speedup(); /* heeeelp */ 1962 1963 mtx_lock(&nblock); 1964 while (needsbuffer & flags) { 1965 if (msleep(&needsbuffer, &nblock, 1966 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1967 mtx_unlock(&nblock); 1968 return (NULL); 1969 } 1970 } 1971 mtx_unlock(&nblock); 1972 } else { 1973 /* 1974 * We finally have a valid bp. We aren't quite out of the 1975 * woods, we still have to reserve kva space. In order 1976 * to keep fragmentation sane we only allocate kva in 1977 * BKVASIZE chunks. 1978 */ 1979 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1980 1981 if (maxsize != bp->b_kvasize) { 1982 vm_offset_t addr = 0; 1983 1984 bfreekva(bp); 1985 1986 vm_map_lock(buffer_map); 1987 if (vm_map_findspace(buffer_map, 1988 vm_map_min(buffer_map), maxsize, &addr)) { 1989 /* 1990 * Uh oh. Buffer map is to fragmented. We 1991 * must defragment the map. 1992 */ 1993 atomic_add_int(&bufdefragcnt, 1); 1994 vm_map_unlock(buffer_map); 1995 defrag = 1; 1996 bp->b_flags |= B_INVAL; 1997 brelse(bp); 1998 goto restart; 1999 } 2000 if (addr) { 2001 vm_map_insert(buffer_map, NULL, 0, 2002 addr, addr + maxsize, 2003 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2004 2005 bp->b_kvabase = (caddr_t) addr; 2006 bp->b_kvasize = maxsize; 2007 atomic_add_int(&bufspace, bp->b_kvasize); 2008 atomic_add_int(&bufreusecnt, 1); 2009 } 2010 vm_map_unlock(buffer_map); 2011 } 2012 bp->b_saveaddr = bp->b_kvabase; 2013 bp->b_data = bp->b_saveaddr; 2014 } 2015 return(bp); 2016 } 2017 2018 /* 2019 * buf_daemon: 2020 * 2021 * buffer flushing daemon. Buffers are normally flushed by the 2022 * update daemon but if it cannot keep up this process starts to 2023 * take the load in an attempt to prevent getnewbuf() from blocking. 2024 */ 2025 2026 static struct kproc_desc buf_kp = { 2027 "bufdaemon", 2028 buf_daemon, 2029 &bufdaemonproc 2030 }; 2031 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2032 2033 static void 2034 buf_daemon() 2035 { 2036 2037 /* 2038 * This process needs to be suspended prior to shutdown sync. 2039 */ 2040 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2041 SHUTDOWN_PRI_LAST); 2042 2043 /* 2044 * This process is allowed to take the buffer cache to the limit 2045 */ 2046 curthread->td_pflags |= TDP_NORUNNINGBUF; 2047 mtx_lock(&bdlock); 2048 for (;;) { 2049 bd_request = 0; 2050 mtx_unlock(&bdlock); 2051 2052 kthread_suspend_check(bufdaemonproc); 2053 2054 /* 2055 * Do the flush. Limit the amount of in-transit I/O we 2056 * allow to build up, otherwise we would completely saturate 2057 * the I/O system. Wakeup any waiting processes before we 2058 * normally would so they can run in parallel with our drain. 2059 */ 2060 while (numdirtybuffers > lodirtybuffers) { 2061 int flushed; 2062 2063 flushed = flushbufqueues(QUEUE_DIRTY, 0); 2064 /* The list empty check here is slightly racy */ 2065 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2066 mtx_lock(&Giant); 2067 flushed += flushbufqueues(QUEUE_DIRTY_GIANT, 0); 2068 mtx_unlock(&Giant); 2069 } 2070 if (flushed == 0) { 2071 /* 2072 * Could not find any buffers without rollback 2073 * dependencies, so just write the first one 2074 * in the hopes of eventually making progress. 2075 */ 2076 flushbufqueues(QUEUE_DIRTY, 1); 2077 if (!TAILQ_EMPTY( 2078 &bufqueues[QUEUE_DIRTY_GIANT])) { 2079 mtx_lock(&Giant); 2080 flushbufqueues(QUEUE_DIRTY_GIANT, 1); 2081 mtx_unlock(&Giant); 2082 } 2083 break; 2084 } 2085 uio_yield(); 2086 } 2087 2088 /* 2089 * Only clear bd_request if we have reached our low water 2090 * mark. The buf_daemon normally waits 1 second and 2091 * then incrementally flushes any dirty buffers that have 2092 * built up, within reason. 2093 * 2094 * If we were unable to hit our low water mark and couldn't 2095 * find any flushable buffers, we sleep half a second. 2096 * Otherwise we loop immediately. 2097 */ 2098 mtx_lock(&bdlock); 2099 if (numdirtybuffers <= lodirtybuffers) { 2100 /* 2101 * We reached our low water mark, reset the 2102 * request and sleep until we are needed again. 2103 * The sleep is just so the suspend code works. 2104 */ 2105 bd_request = 0; 2106 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2107 } else { 2108 /* 2109 * We couldn't find any flushable dirty buffers but 2110 * still have too many dirty buffers, we 2111 * have to sleep and try again. (rare) 2112 */ 2113 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2114 } 2115 } 2116 } 2117 2118 /* 2119 * flushbufqueues: 2120 * 2121 * Try to flush a buffer in the dirty queue. We must be careful to 2122 * free up B_INVAL buffers instead of write them, which NFS is 2123 * particularly sensitive to. 2124 */ 2125 static int flushwithdeps = 0; 2126 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2127 0, "Number of buffers flushed with dependecies that require rollbacks"); 2128 2129 static int 2130 flushbufqueues(int queue, int flushdeps) 2131 { 2132 struct thread *td = curthread; 2133 struct buf sentinel; 2134 struct vnode *vp; 2135 struct mount *mp; 2136 struct buf *bp; 2137 int hasdeps; 2138 int flushed; 2139 int target; 2140 2141 target = numdirtybuffers - lodirtybuffers; 2142 if (flushdeps && target > 2) 2143 target /= 2; 2144 flushed = 0; 2145 bp = NULL; 2146 mtx_lock(&bqlock); 2147 TAILQ_INSERT_TAIL(&bufqueues[queue], &sentinel, b_freelist); 2148 while (flushed != target) { 2149 bp = TAILQ_FIRST(&bufqueues[queue]); 2150 if (bp == &sentinel) 2151 break; 2152 TAILQ_REMOVE(&bufqueues[queue], bp, b_freelist); 2153 TAILQ_INSERT_TAIL(&bufqueues[queue], bp, b_freelist); 2154 2155 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2156 continue; 2157 if (bp->b_pin_count > 0) { 2158 BUF_UNLOCK(bp); 2159 continue; 2160 } 2161 BO_LOCK(bp->b_bufobj); 2162 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2163 (bp->b_flags & B_DELWRI) == 0) { 2164 BO_UNLOCK(bp->b_bufobj); 2165 BUF_UNLOCK(bp); 2166 continue; 2167 } 2168 BO_UNLOCK(bp->b_bufobj); 2169 if (bp->b_flags & B_INVAL) { 2170 bremfreel(bp); 2171 mtx_unlock(&bqlock); 2172 brelse(bp); 2173 flushed++; 2174 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2175 mtx_lock(&bqlock); 2176 continue; 2177 } 2178 2179 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2180 if (flushdeps == 0) { 2181 BUF_UNLOCK(bp); 2182 continue; 2183 } 2184 hasdeps = 1; 2185 } else 2186 hasdeps = 0; 2187 /* 2188 * We must hold the lock on a vnode before writing 2189 * one of its buffers. Otherwise we may confuse, or 2190 * in the case of a snapshot vnode, deadlock the 2191 * system. 2192 * 2193 * The lock order here is the reverse of the normal 2194 * of vnode followed by buf lock. This is ok because 2195 * the NOWAIT will prevent deadlock. 2196 */ 2197 vp = bp->b_vp; 2198 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2199 BUF_UNLOCK(bp); 2200 continue; 2201 } 2202 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2203 mtx_unlock(&bqlock); 2204 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2205 bp, bp->b_vp, bp->b_flags); 2206 vfs_bio_awrite(bp); 2207 vn_finished_write(mp); 2208 VOP_UNLOCK(vp, 0, td); 2209 flushwithdeps += hasdeps; 2210 flushed++; 2211 waitrunningbufspace(); 2212 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2213 mtx_lock(&bqlock); 2214 continue; 2215 } 2216 vn_finished_write(mp); 2217 BUF_UNLOCK(bp); 2218 } 2219 TAILQ_REMOVE(&bufqueues[queue], &sentinel, b_freelist); 2220 mtx_unlock(&bqlock); 2221 return (flushed); 2222 } 2223 2224 /* 2225 * Check to see if a block is currently memory resident. 2226 */ 2227 struct buf * 2228 incore(struct bufobj *bo, daddr_t blkno) 2229 { 2230 struct buf *bp; 2231 2232 BO_LOCK(bo); 2233 bp = gbincore(bo, blkno); 2234 BO_UNLOCK(bo); 2235 return (bp); 2236 } 2237 2238 /* 2239 * Returns true if no I/O is needed to access the 2240 * associated VM object. This is like incore except 2241 * it also hunts around in the VM system for the data. 2242 */ 2243 2244 static int 2245 inmem(struct vnode * vp, daddr_t blkno) 2246 { 2247 vm_object_t obj; 2248 vm_offset_t toff, tinc, size; 2249 vm_page_t m; 2250 vm_ooffset_t off; 2251 2252 ASSERT_VOP_LOCKED(vp, "inmem"); 2253 2254 if (incore(&vp->v_bufobj, blkno)) 2255 return 1; 2256 if (vp->v_mount == NULL) 2257 return 0; 2258 obj = vp->v_object; 2259 if (obj == NULL) 2260 return (0); 2261 2262 size = PAGE_SIZE; 2263 if (size > vp->v_mount->mnt_stat.f_iosize) 2264 size = vp->v_mount->mnt_stat.f_iosize; 2265 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2266 2267 VM_OBJECT_LOCK(obj); 2268 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2269 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2270 if (!m) 2271 goto notinmem; 2272 tinc = size; 2273 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2274 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2275 if (vm_page_is_valid(m, 2276 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2277 goto notinmem; 2278 } 2279 VM_OBJECT_UNLOCK(obj); 2280 return 1; 2281 2282 notinmem: 2283 VM_OBJECT_UNLOCK(obj); 2284 return (0); 2285 } 2286 2287 /* 2288 * vfs_setdirty: 2289 * 2290 * Sets the dirty range for a buffer based on the status of the dirty 2291 * bits in the pages comprising the buffer. 2292 * 2293 * The range is limited to the size of the buffer. 2294 * 2295 * This routine is primarily used by NFS, but is generalized for the 2296 * B_VMIO case. 2297 */ 2298 static void 2299 vfs_setdirty(struct buf *bp) 2300 { 2301 2302 /* 2303 * Degenerate case - empty buffer 2304 */ 2305 2306 if (bp->b_bufsize == 0) 2307 return; 2308 2309 /* 2310 * We qualify the scan for modified pages on whether the 2311 * object has been flushed yet. 2312 */ 2313 2314 if ((bp->b_flags & B_VMIO) == 0) 2315 return; 2316 2317 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2318 vfs_setdirty_locked_object(bp); 2319 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2320 } 2321 2322 static void 2323 vfs_setdirty_locked_object(struct buf *bp) 2324 { 2325 vm_object_t object; 2326 int i; 2327 2328 object = bp->b_bufobj->bo_object; 2329 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2330 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2331 vm_offset_t boffset; 2332 vm_offset_t eoffset; 2333 2334 vm_page_lock_queues(); 2335 /* 2336 * test the pages to see if they have been modified directly 2337 * by users through the VM system. 2338 */ 2339 for (i = 0; i < bp->b_npages; i++) 2340 vm_page_test_dirty(bp->b_pages[i]); 2341 2342 /* 2343 * Calculate the encompassing dirty range, boffset and eoffset, 2344 * (eoffset - boffset) bytes. 2345 */ 2346 2347 for (i = 0; i < bp->b_npages; i++) { 2348 if (bp->b_pages[i]->dirty) 2349 break; 2350 } 2351 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2352 2353 for (i = bp->b_npages - 1; i >= 0; --i) { 2354 if (bp->b_pages[i]->dirty) { 2355 break; 2356 } 2357 } 2358 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2359 2360 vm_page_unlock_queues(); 2361 /* 2362 * Fit it to the buffer. 2363 */ 2364 2365 if (eoffset > bp->b_bcount) 2366 eoffset = bp->b_bcount; 2367 2368 /* 2369 * If we have a good dirty range, merge with the existing 2370 * dirty range. 2371 */ 2372 2373 if (boffset < eoffset) { 2374 if (bp->b_dirtyoff > boffset) 2375 bp->b_dirtyoff = boffset; 2376 if (bp->b_dirtyend < eoffset) 2377 bp->b_dirtyend = eoffset; 2378 } 2379 } 2380 } 2381 2382 /* 2383 * getblk: 2384 * 2385 * Get a block given a specified block and offset into a file/device. 2386 * The buffers B_DONE bit will be cleared on return, making it almost 2387 * ready for an I/O initiation. B_INVAL may or may not be set on 2388 * return. The caller should clear B_INVAL prior to initiating a 2389 * READ. 2390 * 2391 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2392 * an existing buffer. 2393 * 2394 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2395 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2396 * and then cleared based on the backing VM. If the previous buffer is 2397 * non-0-sized but invalid, B_CACHE will be cleared. 2398 * 2399 * If getblk() must create a new buffer, the new buffer is returned with 2400 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2401 * case it is returned with B_INVAL clear and B_CACHE set based on the 2402 * backing VM. 2403 * 2404 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2405 * B_CACHE bit is clear. 2406 * 2407 * What this means, basically, is that the caller should use B_CACHE to 2408 * determine whether the buffer is fully valid or not and should clear 2409 * B_INVAL prior to issuing a read. If the caller intends to validate 2410 * the buffer by loading its data area with something, the caller needs 2411 * to clear B_INVAL. If the caller does this without issuing an I/O, 2412 * the caller should set B_CACHE ( as an optimization ), else the caller 2413 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2414 * a write attempt or if it was a successfull read. If the caller 2415 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2416 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2417 */ 2418 struct buf * 2419 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2420 int flags) 2421 { 2422 struct buf *bp; 2423 struct bufobj *bo; 2424 int error; 2425 2426 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2427 ASSERT_VOP_LOCKED(vp, "getblk"); 2428 if (size > MAXBSIZE) 2429 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2430 2431 bo = &vp->v_bufobj; 2432 loop: 2433 /* 2434 * Block if we are low on buffers. Certain processes are allowed 2435 * to completely exhaust the buffer cache. 2436 * 2437 * If this check ever becomes a bottleneck it may be better to 2438 * move it into the else, when gbincore() fails. At the moment 2439 * it isn't a problem. 2440 * 2441 * XXX remove if 0 sections (clean this up after its proven) 2442 */ 2443 if (numfreebuffers == 0) { 2444 if (TD_IS_IDLETHREAD(curthread)) 2445 return NULL; 2446 mtx_lock(&nblock); 2447 needsbuffer |= VFS_BIO_NEED_ANY; 2448 mtx_unlock(&nblock); 2449 } 2450 2451 VI_LOCK(vp); 2452 bp = gbincore(bo, blkno); 2453 if (bp != NULL) { 2454 int lockflags; 2455 /* 2456 * Buffer is in-core. If the buffer is not busy, it must 2457 * be on a queue. 2458 */ 2459 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2460 2461 if (flags & GB_LOCK_NOWAIT) 2462 lockflags |= LK_NOWAIT; 2463 2464 error = BUF_TIMELOCK(bp, lockflags, 2465 VI_MTX(vp), "getblk", slpflag, slptimeo); 2466 2467 /* 2468 * If we slept and got the lock we have to restart in case 2469 * the buffer changed identities. 2470 */ 2471 if (error == ENOLCK) 2472 goto loop; 2473 /* We timed out or were interrupted. */ 2474 else if (error) 2475 return (NULL); 2476 2477 /* 2478 * The buffer is locked. B_CACHE is cleared if the buffer is 2479 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2480 * and for a VMIO buffer B_CACHE is adjusted according to the 2481 * backing VM cache. 2482 */ 2483 if (bp->b_flags & B_INVAL) 2484 bp->b_flags &= ~B_CACHE; 2485 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2486 bp->b_flags |= B_CACHE; 2487 bremfree(bp); 2488 2489 /* 2490 * check for size inconsistancies for non-VMIO case. 2491 */ 2492 2493 if (bp->b_bcount != size) { 2494 if ((bp->b_flags & B_VMIO) == 0 || 2495 (size > bp->b_kvasize)) { 2496 if (bp->b_flags & B_DELWRI) { 2497 /* 2498 * If buffer is pinned and caller does 2499 * not want sleep waiting for it to be 2500 * unpinned, bail out 2501 * */ 2502 if (bp->b_pin_count > 0) { 2503 if (flags & GB_LOCK_NOWAIT) { 2504 bqrelse(bp); 2505 return (NULL); 2506 } else { 2507 bunpin_wait(bp); 2508 } 2509 } 2510 bp->b_flags |= B_NOCACHE; 2511 bwrite(bp); 2512 } else { 2513 if (LIST_EMPTY(&bp->b_dep)) { 2514 bp->b_flags |= B_RELBUF; 2515 brelse(bp); 2516 } else { 2517 bp->b_flags |= B_NOCACHE; 2518 bwrite(bp); 2519 } 2520 } 2521 goto loop; 2522 } 2523 } 2524 2525 /* 2526 * If the size is inconsistant in the VMIO case, we can resize 2527 * the buffer. This might lead to B_CACHE getting set or 2528 * cleared. If the size has not changed, B_CACHE remains 2529 * unchanged from its previous state. 2530 */ 2531 2532 if (bp->b_bcount != size) 2533 allocbuf(bp, size); 2534 2535 KASSERT(bp->b_offset != NOOFFSET, 2536 ("getblk: no buffer offset")); 2537 2538 /* 2539 * A buffer with B_DELWRI set and B_CACHE clear must 2540 * be committed before we can return the buffer in 2541 * order to prevent the caller from issuing a read 2542 * ( due to B_CACHE not being set ) and overwriting 2543 * it. 2544 * 2545 * Most callers, including NFS and FFS, need this to 2546 * operate properly either because they assume they 2547 * can issue a read if B_CACHE is not set, or because 2548 * ( for example ) an uncached B_DELWRI might loop due 2549 * to softupdates re-dirtying the buffer. In the latter 2550 * case, B_CACHE is set after the first write completes, 2551 * preventing further loops. 2552 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2553 * above while extending the buffer, we cannot allow the 2554 * buffer to remain with B_CACHE set after the write 2555 * completes or it will represent a corrupt state. To 2556 * deal with this we set B_NOCACHE to scrap the buffer 2557 * after the write. 2558 * 2559 * We might be able to do something fancy, like setting 2560 * B_CACHE in bwrite() except if B_DELWRI is already set, 2561 * so the below call doesn't set B_CACHE, but that gets real 2562 * confusing. This is much easier. 2563 */ 2564 2565 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2566 bp->b_flags |= B_NOCACHE; 2567 bwrite(bp); 2568 goto loop; 2569 } 2570 bp->b_flags &= ~B_DONE; 2571 } else { 2572 int bsize, maxsize, vmio; 2573 off_t offset; 2574 2575 /* 2576 * Buffer is not in-core, create new buffer. The buffer 2577 * returned by getnewbuf() is locked. Note that the returned 2578 * buffer is also considered valid (not marked B_INVAL). 2579 */ 2580 VI_UNLOCK(vp); 2581 /* 2582 * If the user does not want us to create the buffer, bail out 2583 * here. 2584 */ 2585 if (flags & GB_NOCREAT) 2586 return NULL; 2587 bsize = bo->bo_bsize; 2588 offset = blkno * bsize; 2589 vmio = vp->v_object != NULL; 2590 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2591 maxsize = imax(maxsize, bsize); 2592 2593 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2594 if (bp == NULL) { 2595 if (slpflag || slptimeo) 2596 return NULL; 2597 goto loop; 2598 } 2599 2600 /* 2601 * This code is used to make sure that a buffer is not 2602 * created while the getnewbuf routine is blocked. 2603 * This can be a problem whether the vnode is locked or not. 2604 * If the buffer is created out from under us, we have to 2605 * throw away the one we just created. 2606 * 2607 * Note: this must occur before we associate the buffer 2608 * with the vp especially considering limitations in 2609 * the splay tree implementation when dealing with duplicate 2610 * lblkno's. 2611 */ 2612 BO_LOCK(bo); 2613 if (gbincore(bo, blkno)) { 2614 BO_UNLOCK(bo); 2615 bp->b_flags |= B_INVAL; 2616 brelse(bp); 2617 goto loop; 2618 } 2619 2620 /* 2621 * Insert the buffer into the hash, so that it can 2622 * be found by incore. 2623 */ 2624 bp->b_blkno = bp->b_lblkno = blkno; 2625 bp->b_offset = offset; 2626 bgetvp(vp, bp); 2627 BO_UNLOCK(bo); 2628 2629 /* 2630 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2631 * buffer size starts out as 0, B_CACHE will be set by 2632 * allocbuf() for the VMIO case prior to it testing the 2633 * backing store for validity. 2634 */ 2635 2636 if (vmio) { 2637 bp->b_flags |= B_VMIO; 2638 #if defined(VFS_BIO_DEBUG) 2639 if (vn_canvmio(vp) != TRUE) 2640 printf("getblk: VMIO on vnode type %d\n", 2641 vp->v_type); 2642 #endif 2643 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2644 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2645 bp, vp->v_object, bp->b_bufobj->bo_object)); 2646 } else { 2647 bp->b_flags &= ~B_VMIO; 2648 KASSERT(bp->b_bufobj->bo_object == NULL, 2649 ("ARGH! has b_bufobj->bo_object %p %p\n", 2650 bp, bp->b_bufobj->bo_object)); 2651 } 2652 2653 allocbuf(bp, size); 2654 bp->b_flags &= ~B_DONE; 2655 } 2656 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2657 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2658 KASSERT(bp->b_bufobj == bo, 2659 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2660 return (bp); 2661 } 2662 2663 /* 2664 * Get an empty, disassociated buffer of given size. The buffer is initially 2665 * set to B_INVAL. 2666 */ 2667 struct buf * 2668 geteblk(int size) 2669 { 2670 struct buf *bp; 2671 int maxsize; 2672 2673 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2674 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2675 continue; 2676 allocbuf(bp, size); 2677 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2678 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2679 return (bp); 2680 } 2681 2682 2683 /* 2684 * This code constitutes the buffer memory from either anonymous system 2685 * memory (in the case of non-VMIO operations) or from an associated 2686 * VM object (in the case of VMIO operations). This code is able to 2687 * resize a buffer up or down. 2688 * 2689 * Note that this code is tricky, and has many complications to resolve 2690 * deadlock or inconsistant data situations. Tread lightly!!! 2691 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2692 * the caller. Calling this code willy nilly can result in the loss of data. 2693 * 2694 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2695 * B_CACHE for the non-VMIO case. 2696 */ 2697 2698 int 2699 allocbuf(struct buf *bp, int size) 2700 { 2701 int newbsize, mbsize; 2702 int i; 2703 2704 if (BUF_REFCNT(bp) == 0) 2705 panic("allocbuf: buffer not busy"); 2706 2707 if (bp->b_kvasize < size) 2708 panic("allocbuf: buffer too small"); 2709 2710 if ((bp->b_flags & B_VMIO) == 0) { 2711 caddr_t origbuf; 2712 int origbufsize; 2713 /* 2714 * Just get anonymous memory from the kernel. Don't 2715 * mess with B_CACHE. 2716 */ 2717 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2718 if (bp->b_flags & B_MALLOC) 2719 newbsize = mbsize; 2720 else 2721 newbsize = round_page(size); 2722 2723 if (newbsize < bp->b_bufsize) { 2724 /* 2725 * malloced buffers are not shrunk 2726 */ 2727 if (bp->b_flags & B_MALLOC) { 2728 if (newbsize) { 2729 bp->b_bcount = size; 2730 } else { 2731 free(bp->b_data, M_BIOBUF); 2732 if (bp->b_bufsize) { 2733 atomic_subtract_int( 2734 &bufmallocspace, 2735 bp->b_bufsize); 2736 bufspacewakeup(); 2737 bp->b_bufsize = 0; 2738 } 2739 bp->b_saveaddr = bp->b_kvabase; 2740 bp->b_data = bp->b_saveaddr; 2741 bp->b_bcount = 0; 2742 bp->b_flags &= ~B_MALLOC; 2743 } 2744 return 1; 2745 } 2746 vm_hold_free_pages( 2747 bp, 2748 (vm_offset_t) bp->b_data + newbsize, 2749 (vm_offset_t) bp->b_data + bp->b_bufsize); 2750 } else if (newbsize > bp->b_bufsize) { 2751 /* 2752 * We only use malloced memory on the first allocation. 2753 * and revert to page-allocated memory when the buffer 2754 * grows. 2755 */ 2756 /* 2757 * There is a potential smp race here that could lead 2758 * to bufmallocspace slightly passing the max. It 2759 * is probably extremely rare and not worth worrying 2760 * over. 2761 */ 2762 if ( (bufmallocspace < maxbufmallocspace) && 2763 (bp->b_bufsize == 0) && 2764 (mbsize <= PAGE_SIZE/2)) { 2765 2766 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2767 bp->b_bufsize = mbsize; 2768 bp->b_bcount = size; 2769 bp->b_flags |= B_MALLOC; 2770 atomic_add_int(&bufmallocspace, mbsize); 2771 return 1; 2772 } 2773 origbuf = NULL; 2774 origbufsize = 0; 2775 /* 2776 * If the buffer is growing on its other-than-first allocation, 2777 * then we revert to the page-allocation scheme. 2778 */ 2779 if (bp->b_flags & B_MALLOC) { 2780 origbuf = bp->b_data; 2781 origbufsize = bp->b_bufsize; 2782 bp->b_data = bp->b_kvabase; 2783 if (bp->b_bufsize) { 2784 atomic_subtract_int(&bufmallocspace, 2785 bp->b_bufsize); 2786 bufspacewakeup(); 2787 bp->b_bufsize = 0; 2788 } 2789 bp->b_flags &= ~B_MALLOC; 2790 newbsize = round_page(newbsize); 2791 } 2792 vm_hold_load_pages( 2793 bp, 2794 (vm_offset_t) bp->b_data + bp->b_bufsize, 2795 (vm_offset_t) bp->b_data + newbsize); 2796 if (origbuf) { 2797 bcopy(origbuf, bp->b_data, origbufsize); 2798 free(origbuf, M_BIOBUF); 2799 } 2800 } 2801 } else { 2802 int desiredpages; 2803 2804 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2805 desiredpages = (size == 0) ? 0 : 2806 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2807 2808 if (bp->b_flags & B_MALLOC) 2809 panic("allocbuf: VMIO buffer can't be malloced"); 2810 /* 2811 * Set B_CACHE initially if buffer is 0 length or will become 2812 * 0-length. 2813 */ 2814 if (size == 0 || bp->b_bufsize == 0) 2815 bp->b_flags |= B_CACHE; 2816 2817 if (newbsize < bp->b_bufsize) { 2818 /* 2819 * DEV_BSIZE aligned new buffer size is less then the 2820 * DEV_BSIZE aligned existing buffer size. Figure out 2821 * if we have to remove any pages. 2822 */ 2823 if (desiredpages < bp->b_npages) { 2824 vm_page_t m; 2825 2826 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2827 vm_page_lock_queues(); 2828 for (i = desiredpages; i < bp->b_npages; i++) { 2829 /* 2830 * the page is not freed here -- it 2831 * is the responsibility of 2832 * vnode_pager_setsize 2833 */ 2834 m = bp->b_pages[i]; 2835 KASSERT(m != bogus_page, 2836 ("allocbuf: bogus page found")); 2837 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2838 vm_page_lock_queues(); 2839 2840 bp->b_pages[i] = NULL; 2841 vm_page_unwire(m, 0); 2842 } 2843 vm_page_unlock_queues(); 2844 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2845 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2846 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2847 bp->b_npages = desiredpages; 2848 } 2849 } else if (size > bp->b_bcount) { 2850 /* 2851 * We are growing the buffer, possibly in a 2852 * byte-granular fashion. 2853 */ 2854 struct vnode *vp; 2855 vm_object_t obj; 2856 vm_offset_t toff; 2857 vm_offset_t tinc; 2858 2859 /* 2860 * Step 1, bring in the VM pages from the object, 2861 * allocating them if necessary. We must clear 2862 * B_CACHE if these pages are not valid for the 2863 * range covered by the buffer. 2864 */ 2865 2866 vp = bp->b_vp; 2867 obj = bp->b_bufobj->bo_object; 2868 2869 VM_OBJECT_LOCK(obj); 2870 while (bp->b_npages < desiredpages) { 2871 vm_page_t m; 2872 vm_pindex_t pi; 2873 2874 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2875 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2876 /* 2877 * note: must allocate system pages 2878 * since blocking here could intefere 2879 * with paging I/O, no matter which 2880 * process we are. 2881 */ 2882 m = vm_page_alloc(obj, pi, 2883 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2884 VM_ALLOC_WIRED); 2885 if (m == NULL) { 2886 atomic_add_int(&vm_pageout_deficit, 2887 desiredpages - bp->b_npages); 2888 VM_OBJECT_UNLOCK(obj); 2889 VM_WAIT; 2890 VM_OBJECT_LOCK(obj); 2891 } else { 2892 bp->b_flags &= ~B_CACHE; 2893 bp->b_pages[bp->b_npages] = m; 2894 ++bp->b_npages; 2895 } 2896 continue; 2897 } 2898 2899 /* 2900 * We found a page. If we have to sleep on it, 2901 * retry because it might have gotten freed out 2902 * from under us. 2903 * 2904 * We can only test VPO_BUSY here. Blocking on 2905 * m->busy might lead to a deadlock: 2906 * 2907 * vm_fault->getpages->cluster_read->allocbuf 2908 * 2909 */ 2910 vm_page_lock_queues(); 2911 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2912 continue; 2913 2914 /* 2915 * We have a good page. Should we wakeup the 2916 * page daemon? 2917 */ 2918 if ((curproc != pageproc) && 2919 (VM_PAGE_INQUEUE1(m, PQ_CACHE)) && 2920 ((cnt.v_free_count + cnt.v_cache_count) < 2921 (cnt.v_free_min + cnt.v_cache_min))) { 2922 pagedaemon_wakeup(); 2923 } 2924 vm_page_wire(m); 2925 vm_page_unlock_queues(); 2926 bp->b_pages[bp->b_npages] = m; 2927 ++bp->b_npages; 2928 } 2929 2930 /* 2931 * Step 2. We've loaded the pages into the buffer, 2932 * we have to figure out if we can still have B_CACHE 2933 * set. Note that B_CACHE is set according to the 2934 * byte-granular range ( bcount and size ), new the 2935 * aligned range ( newbsize ). 2936 * 2937 * The VM test is against m->valid, which is DEV_BSIZE 2938 * aligned. Needless to say, the validity of the data 2939 * needs to also be DEV_BSIZE aligned. Note that this 2940 * fails with NFS if the server or some other client 2941 * extends the file's EOF. If our buffer is resized, 2942 * B_CACHE may remain set! XXX 2943 */ 2944 2945 toff = bp->b_bcount; 2946 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2947 2948 while ((bp->b_flags & B_CACHE) && toff < size) { 2949 vm_pindex_t pi; 2950 2951 if (tinc > (size - toff)) 2952 tinc = size - toff; 2953 2954 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2955 PAGE_SHIFT; 2956 2957 vfs_buf_test_cache( 2958 bp, 2959 bp->b_offset, 2960 toff, 2961 tinc, 2962 bp->b_pages[pi] 2963 ); 2964 toff += tinc; 2965 tinc = PAGE_SIZE; 2966 } 2967 VM_OBJECT_UNLOCK(obj); 2968 2969 /* 2970 * Step 3, fixup the KVM pmap. Remember that 2971 * bp->b_data is relative to bp->b_offset, but 2972 * bp->b_offset may be offset into the first page. 2973 */ 2974 2975 bp->b_data = (caddr_t) 2976 trunc_page((vm_offset_t)bp->b_data); 2977 pmap_qenter( 2978 (vm_offset_t)bp->b_data, 2979 bp->b_pages, 2980 bp->b_npages 2981 ); 2982 2983 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2984 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2985 } 2986 } 2987 if (newbsize < bp->b_bufsize) 2988 bufspacewakeup(); 2989 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2990 bp->b_bcount = size; /* requested buffer size */ 2991 return 1; 2992 } 2993 2994 void 2995 biodone(struct bio *bp) 2996 { 2997 void (*done)(struct bio *); 2998 2999 mtx_lock(&bdonelock); 3000 bp->bio_flags |= BIO_DONE; 3001 done = bp->bio_done; 3002 if (done == NULL) 3003 wakeup(bp); 3004 mtx_unlock(&bdonelock); 3005 if (done != NULL) 3006 done(bp); 3007 } 3008 3009 /* 3010 * Wait for a BIO to finish. 3011 * 3012 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3013 * case is not yet clear. 3014 */ 3015 int 3016 biowait(struct bio *bp, const char *wchan) 3017 { 3018 3019 mtx_lock(&bdonelock); 3020 while ((bp->bio_flags & BIO_DONE) == 0) 3021 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3022 mtx_unlock(&bdonelock); 3023 if (bp->bio_error != 0) 3024 return (bp->bio_error); 3025 if (!(bp->bio_flags & BIO_ERROR)) 3026 return (0); 3027 return (EIO); 3028 } 3029 3030 void 3031 biofinish(struct bio *bp, struct devstat *stat, int error) 3032 { 3033 3034 if (error) { 3035 bp->bio_error = error; 3036 bp->bio_flags |= BIO_ERROR; 3037 } 3038 if (stat != NULL) 3039 devstat_end_transaction_bio(stat, bp); 3040 biodone(bp); 3041 } 3042 3043 /* 3044 * bufwait: 3045 * 3046 * Wait for buffer I/O completion, returning error status. The buffer 3047 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3048 * error and cleared. 3049 */ 3050 int 3051 bufwait(struct buf *bp) 3052 { 3053 if (bp->b_iocmd == BIO_READ) 3054 bwait(bp, PRIBIO, "biord"); 3055 else 3056 bwait(bp, PRIBIO, "biowr"); 3057 if (bp->b_flags & B_EINTR) { 3058 bp->b_flags &= ~B_EINTR; 3059 return (EINTR); 3060 } 3061 if (bp->b_ioflags & BIO_ERROR) { 3062 return (bp->b_error ? bp->b_error : EIO); 3063 } else { 3064 return (0); 3065 } 3066 } 3067 3068 /* 3069 * Call back function from struct bio back up to struct buf. 3070 */ 3071 static void 3072 bufdonebio(struct bio *bip) 3073 { 3074 struct buf *bp; 3075 3076 bp = bip->bio_caller2; 3077 bp->b_resid = bp->b_bcount - bip->bio_completed; 3078 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3079 bp->b_ioflags = bip->bio_flags; 3080 bp->b_error = bip->bio_error; 3081 if (bp->b_error) 3082 bp->b_ioflags |= BIO_ERROR; 3083 bufdone(bp); 3084 g_destroy_bio(bip); 3085 } 3086 3087 void 3088 dev_strategy(struct cdev *dev, struct buf *bp) 3089 { 3090 struct cdevsw *csw; 3091 struct bio *bip; 3092 3093 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3094 panic("b_iocmd botch"); 3095 for (;;) { 3096 bip = g_new_bio(); 3097 if (bip != NULL) 3098 break; 3099 /* Try again later */ 3100 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3101 } 3102 bip->bio_cmd = bp->b_iocmd; 3103 bip->bio_offset = bp->b_iooffset; 3104 bip->bio_length = bp->b_bcount; 3105 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3106 bip->bio_data = bp->b_data; 3107 bip->bio_done = bufdonebio; 3108 bip->bio_caller2 = bp; 3109 bip->bio_dev = dev; 3110 KASSERT(dev->si_refcount > 0, 3111 ("dev_strategy on un-referenced struct cdev *(%s)", 3112 devtoname(dev))); 3113 csw = dev_refthread(dev); 3114 if (csw == NULL) { 3115 g_destroy_bio(bip); 3116 bp->b_error = ENXIO; 3117 bp->b_ioflags = BIO_ERROR; 3118 bufdone(bp); 3119 return; 3120 } 3121 (*csw->d_strategy)(bip); 3122 dev_relthread(dev); 3123 } 3124 3125 /* 3126 * bufdone: 3127 * 3128 * Finish I/O on a buffer, optionally calling a completion function. 3129 * This is usually called from an interrupt so process blocking is 3130 * not allowed. 3131 * 3132 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3133 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3134 * assuming B_INVAL is clear. 3135 * 3136 * For the VMIO case, we set B_CACHE if the op was a read and no 3137 * read error occured, or if the op was a write. B_CACHE is never 3138 * set if the buffer is invalid or otherwise uncacheable. 3139 * 3140 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3141 * initiator to leave B_INVAL set to brelse the buffer out of existance 3142 * in the biodone routine. 3143 */ 3144 void 3145 bufdone(struct buf *bp) 3146 { 3147 struct bufobj *dropobj; 3148 void (*biodone)(struct buf *); 3149 3150 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3151 dropobj = NULL; 3152 3153 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, 3154 BUF_REFCNT(bp))); 3155 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3156 3157 runningbufwakeup(bp); 3158 if (bp->b_iocmd == BIO_WRITE) 3159 dropobj = bp->b_bufobj; 3160 /* call optional completion function if requested */ 3161 if (bp->b_iodone != NULL) { 3162 biodone = bp->b_iodone; 3163 bp->b_iodone = NULL; 3164 (*biodone) (bp); 3165 if (dropobj) 3166 bufobj_wdrop(dropobj); 3167 return; 3168 } 3169 3170 bufdone_finish(bp); 3171 3172 if (dropobj) 3173 bufobj_wdrop(dropobj); 3174 } 3175 3176 void 3177 bufdone_finish(struct buf *bp) 3178 { 3179 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, 3180 BUF_REFCNT(bp))); 3181 3182 if (!LIST_EMPTY(&bp->b_dep)) 3183 buf_complete(bp); 3184 3185 if (bp->b_flags & B_VMIO) { 3186 int i; 3187 vm_ooffset_t foff; 3188 vm_page_t m; 3189 vm_object_t obj; 3190 int iosize; 3191 struct vnode *vp = bp->b_vp; 3192 boolean_t are_queues_locked; 3193 3194 obj = bp->b_bufobj->bo_object; 3195 3196 #if defined(VFS_BIO_DEBUG) 3197 mp_fixme("usecount and vflag accessed without locks."); 3198 if (vp->v_usecount == 0) { 3199 panic("biodone: zero vnode ref count"); 3200 } 3201 3202 KASSERT(vp->v_object != NULL, 3203 ("biodone: vnode %p has no vm_object", vp)); 3204 #endif 3205 3206 foff = bp->b_offset; 3207 KASSERT(bp->b_offset != NOOFFSET, 3208 ("biodone: no buffer offset")); 3209 3210 VM_OBJECT_LOCK(obj); 3211 #if defined(VFS_BIO_DEBUG) 3212 if (obj->paging_in_progress < bp->b_npages) { 3213 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3214 obj->paging_in_progress, bp->b_npages); 3215 } 3216 #endif 3217 3218 /* 3219 * Set B_CACHE if the op was a normal read and no error 3220 * occured. B_CACHE is set for writes in the b*write() 3221 * routines. 3222 */ 3223 iosize = bp->b_bcount - bp->b_resid; 3224 if (bp->b_iocmd == BIO_READ && 3225 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3226 !(bp->b_ioflags & BIO_ERROR)) { 3227 bp->b_flags |= B_CACHE; 3228 } 3229 if (bp->b_iocmd == BIO_READ) { 3230 vm_page_lock_queues(); 3231 are_queues_locked = TRUE; 3232 } else 3233 are_queues_locked = FALSE; 3234 for (i = 0; i < bp->b_npages; i++) { 3235 int bogusflag = 0; 3236 int resid; 3237 3238 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3239 if (resid > iosize) 3240 resid = iosize; 3241 3242 /* 3243 * cleanup bogus pages, restoring the originals 3244 */ 3245 m = bp->b_pages[i]; 3246 if (m == bogus_page) { 3247 bogusflag = 1; 3248 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3249 if (m == NULL) 3250 panic("biodone: page disappeared!"); 3251 bp->b_pages[i] = m; 3252 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3253 bp->b_pages, bp->b_npages); 3254 } 3255 #if defined(VFS_BIO_DEBUG) 3256 if (OFF_TO_IDX(foff) != m->pindex) { 3257 printf( 3258 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3259 (intmax_t)foff, (uintmax_t)m->pindex); 3260 } 3261 #endif 3262 3263 /* 3264 * In the write case, the valid and clean bits are 3265 * already changed correctly ( see bdwrite() ), so we 3266 * only need to do this here in the read case. 3267 */ 3268 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3269 vfs_page_set_valid(bp, foff, i, m); 3270 } 3271 3272 /* 3273 * when debugging new filesystems or buffer I/O methods, this 3274 * is the most common error that pops up. if you see this, you 3275 * have not set the page busy flag correctly!!! 3276 */ 3277 if (m->busy == 0) { 3278 printf("biodone: page busy < 0, " 3279 "pindex: %d, foff: 0x(%x,%x), " 3280 "resid: %d, index: %d\n", 3281 (int) m->pindex, (int)(foff >> 32), 3282 (int) foff & 0xffffffff, resid, i); 3283 if (!vn_isdisk(vp, NULL)) 3284 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3285 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3286 (intmax_t) bp->b_lblkno, 3287 bp->b_flags, bp->b_npages); 3288 else 3289 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3290 (intmax_t) bp->b_lblkno, 3291 bp->b_flags, bp->b_npages); 3292 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3293 (u_long)m->valid, (u_long)m->dirty, 3294 m->wire_count); 3295 panic("biodone: page busy < 0\n"); 3296 } 3297 vm_page_io_finish(m); 3298 vm_object_pip_subtract(obj, 1); 3299 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3300 iosize -= resid; 3301 } 3302 if (are_queues_locked) 3303 vm_page_unlock_queues(); 3304 vm_object_pip_wakeupn(obj, 0); 3305 VM_OBJECT_UNLOCK(obj); 3306 } 3307 3308 /* 3309 * For asynchronous completions, release the buffer now. The brelse 3310 * will do a wakeup there if necessary - so no need to do a wakeup 3311 * here in the async case. The sync case always needs to do a wakeup. 3312 */ 3313 3314 if (bp->b_flags & B_ASYNC) { 3315 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3316 brelse(bp); 3317 else 3318 bqrelse(bp); 3319 } else 3320 bdone(bp); 3321 } 3322 3323 /* 3324 * This routine is called in lieu of iodone in the case of 3325 * incomplete I/O. This keeps the busy status for pages 3326 * consistant. 3327 */ 3328 void 3329 vfs_unbusy_pages(struct buf *bp) 3330 { 3331 int i; 3332 vm_object_t obj; 3333 vm_page_t m; 3334 3335 runningbufwakeup(bp); 3336 if (!(bp->b_flags & B_VMIO)) 3337 return; 3338 3339 obj = bp->b_bufobj->bo_object; 3340 VM_OBJECT_LOCK(obj); 3341 for (i = 0; i < bp->b_npages; i++) { 3342 m = bp->b_pages[i]; 3343 if (m == bogus_page) { 3344 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3345 if (!m) 3346 panic("vfs_unbusy_pages: page missing\n"); 3347 bp->b_pages[i] = m; 3348 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3349 bp->b_pages, bp->b_npages); 3350 } 3351 vm_object_pip_subtract(obj, 1); 3352 vm_page_io_finish(m); 3353 } 3354 vm_object_pip_wakeupn(obj, 0); 3355 VM_OBJECT_UNLOCK(obj); 3356 } 3357 3358 /* 3359 * vfs_page_set_valid: 3360 * 3361 * Set the valid bits in a page based on the supplied offset. The 3362 * range is restricted to the buffer's size. 3363 * 3364 * This routine is typically called after a read completes. 3365 */ 3366 static void 3367 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3368 { 3369 vm_ooffset_t soff, eoff; 3370 3371 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3372 /* 3373 * Start and end offsets in buffer. eoff - soff may not cross a 3374 * page boundry or cross the end of the buffer. The end of the 3375 * buffer, in this case, is our file EOF, not the allocation size 3376 * of the buffer. 3377 */ 3378 soff = off; 3379 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3380 if (eoff > bp->b_offset + bp->b_bcount) 3381 eoff = bp->b_offset + bp->b_bcount; 3382 3383 /* 3384 * Set valid range. This is typically the entire buffer and thus the 3385 * entire page. 3386 */ 3387 if (eoff > soff) { 3388 vm_page_set_validclean( 3389 m, 3390 (vm_offset_t) (soff & PAGE_MASK), 3391 (vm_offset_t) (eoff - soff) 3392 ); 3393 } 3394 } 3395 3396 /* 3397 * This routine is called before a device strategy routine. 3398 * It is used to tell the VM system that paging I/O is in 3399 * progress, and treat the pages associated with the buffer 3400 * almost as being VPO_BUSY. Also the object paging_in_progress 3401 * flag is handled to make sure that the object doesn't become 3402 * inconsistant. 3403 * 3404 * Since I/O has not been initiated yet, certain buffer flags 3405 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3406 * and should be ignored. 3407 */ 3408 void 3409 vfs_busy_pages(struct buf *bp, int clear_modify) 3410 { 3411 int i, bogus; 3412 vm_object_t obj; 3413 vm_ooffset_t foff; 3414 vm_page_t m; 3415 3416 if (!(bp->b_flags & B_VMIO)) 3417 return; 3418 3419 obj = bp->b_bufobj->bo_object; 3420 foff = bp->b_offset; 3421 KASSERT(bp->b_offset != NOOFFSET, 3422 ("vfs_busy_pages: no buffer offset")); 3423 VM_OBJECT_LOCK(obj); 3424 if (bp->b_bufsize != 0) 3425 vfs_setdirty_locked_object(bp); 3426 retry: 3427 for (i = 0; i < bp->b_npages; i++) { 3428 m = bp->b_pages[i]; 3429 3430 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3431 goto retry; 3432 } 3433 bogus = 0; 3434 vm_page_lock_queues(); 3435 for (i = 0; i < bp->b_npages; i++) { 3436 m = bp->b_pages[i]; 3437 3438 if ((bp->b_flags & B_CLUSTER) == 0) { 3439 vm_object_pip_add(obj, 1); 3440 vm_page_io_start(m); 3441 } 3442 /* 3443 * When readying a buffer for a read ( i.e 3444 * clear_modify == 0 ), it is important to do 3445 * bogus_page replacement for valid pages in 3446 * partially instantiated buffers. Partially 3447 * instantiated buffers can, in turn, occur when 3448 * reconstituting a buffer from its VM backing store 3449 * base. We only have to do this if B_CACHE is 3450 * clear ( which causes the I/O to occur in the 3451 * first place ). The replacement prevents the read 3452 * I/O from overwriting potentially dirty VM-backed 3453 * pages. XXX bogus page replacement is, uh, bogus. 3454 * It may not work properly with small-block devices. 3455 * We need to find a better way. 3456 */ 3457 pmap_remove_all(m); 3458 if (clear_modify) 3459 vfs_page_set_valid(bp, foff, i, m); 3460 else if (m->valid == VM_PAGE_BITS_ALL && 3461 (bp->b_flags & B_CACHE) == 0) { 3462 bp->b_pages[i] = bogus_page; 3463 bogus++; 3464 } 3465 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3466 } 3467 vm_page_unlock_queues(); 3468 VM_OBJECT_UNLOCK(obj); 3469 if (bogus) 3470 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3471 bp->b_pages, bp->b_npages); 3472 } 3473 3474 /* 3475 * Tell the VM system that the pages associated with this buffer 3476 * are clean. This is used for delayed writes where the data is 3477 * going to go to disk eventually without additional VM intevention. 3478 * 3479 * Note that while we only really need to clean through to b_bcount, we 3480 * just go ahead and clean through to b_bufsize. 3481 */ 3482 static void 3483 vfs_clean_pages(struct buf *bp) 3484 { 3485 int i; 3486 vm_ooffset_t foff, noff, eoff; 3487 vm_page_t m; 3488 3489 if (!(bp->b_flags & B_VMIO)) 3490 return; 3491 3492 foff = bp->b_offset; 3493 KASSERT(bp->b_offset != NOOFFSET, 3494 ("vfs_clean_pages: no buffer offset")); 3495 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3496 vm_page_lock_queues(); 3497 for (i = 0; i < bp->b_npages; i++) { 3498 m = bp->b_pages[i]; 3499 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3500 eoff = noff; 3501 3502 if (eoff > bp->b_offset + bp->b_bufsize) 3503 eoff = bp->b_offset + bp->b_bufsize; 3504 vfs_page_set_valid(bp, foff, i, m); 3505 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3506 foff = noff; 3507 } 3508 vm_page_unlock_queues(); 3509 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3510 } 3511 3512 /* 3513 * vfs_bio_set_validclean: 3514 * 3515 * Set the range within the buffer to valid and clean. The range is 3516 * relative to the beginning of the buffer, b_offset. Note that b_offset 3517 * itself may be offset from the beginning of the first page. 3518 * 3519 */ 3520 3521 void 3522 vfs_bio_set_validclean(struct buf *bp, int base, int size) 3523 { 3524 int i, n; 3525 vm_page_t m; 3526 3527 if (!(bp->b_flags & B_VMIO)) 3528 return; 3529 /* 3530 * Fixup base to be relative to beginning of first page. 3531 * Set initial n to be the maximum number of bytes in the 3532 * first page that can be validated. 3533 */ 3534 3535 base += (bp->b_offset & PAGE_MASK); 3536 n = PAGE_SIZE - (base & PAGE_MASK); 3537 3538 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3539 vm_page_lock_queues(); 3540 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3541 m = bp->b_pages[i]; 3542 if (n > size) 3543 n = size; 3544 vm_page_set_validclean(m, base & PAGE_MASK, n); 3545 base += n; 3546 size -= n; 3547 n = PAGE_SIZE; 3548 } 3549 vm_page_unlock_queues(); 3550 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3551 } 3552 3553 /* 3554 * vfs_bio_clrbuf: 3555 * 3556 * clear a buffer. This routine essentially fakes an I/O, so we need 3557 * to clear BIO_ERROR and B_INVAL. 3558 * 3559 * Note that while we only theoretically need to clear through b_bcount, 3560 * we go ahead and clear through b_bufsize. 3561 */ 3562 3563 void 3564 vfs_bio_clrbuf(struct buf *bp) 3565 { 3566 int i, j, mask = 0; 3567 caddr_t sa, ea; 3568 3569 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3570 clrbuf(bp); 3571 return; 3572 } 3573 3574 bp->b_flags &= ~B_INVAL; 3575 bp->b_ioflags &= ~BIO_ERROR; 3576 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3577 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3578 (bp->b_offset & PAGE_MASK) == 0) { 3579 if (bp->b_pages[0] == bogus_page) 3580 goto unlock; 3581 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3582 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3583 if ((bp->b_pages[0]->valid & mask) == mask) 3584 goto unlock; 3585 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3586 ((bp->b_pages[0]->valid & mask) == 0)) { 3587 bzero(bp->b_data, bp->b_bufsize); 3588 bp->b_pages[0]->valid |= mask; 3589 goto unlock; 3590 } 3591 } 3592 ea = sa = bp->b_data; 3593 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3594 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3595 ea = (caddr_t)(vm_offset_t)ulmin( 3596 (u_long)(vm_offset_t)ea, 3597 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3598 if (bp->b_pages[i] == bogus_page) 3599 continue; 3600 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3601 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3602 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3603 if ((bp->b_pages[i]->valid & mask) == mask) 3604 continue; 3605 if ((bp->b_pages[i]->valid & mask) == 0) { 3606 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3607 bzero(sa, ea - sa); 3608 } else { 3609 for (; sa < ea; sa += DEV_BSIZE, j++) { 3610 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3611 (bp->b_pages[i]->valid & (1 << j)) == 0) 3612 bzero(sa, DEV_BSIZE); 3613 } 3614 } 3615 bp->b_pages[i]->valid |= mask; 3616 } 3617 unlock: 3618 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3619 bp->b_resid = 0; 3620 } 3621 3622 /* 3623 * vm_hold_load_pages and vm_hold_free_pages get pages into 3624 * a buffers address space. The pages are anonymous and are 3625 * not associated with a file object. 3626 */ 3627 static void 3628 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3629 { 3630 vm_offset_t pg; 3631 vm_page_t p; 3632 int index; 3633 3634 to = round_page(to); 3635 from = round_page(from); 3636 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3637 3638 VM_OBJECT_LOCK(kernel_object); 3639 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3640 tryagain: 3641 /* 3642 * note: must allocate system pages since blocking here 3643 * could intefere with paging I/O, no matter which 3644 * process we are. 3645 */ 3646 p = vm_page_alloc(kernel_object, 3647 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3648 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3649 if (!p) { 3650 atomic_add_int(&vm_pageout_deficit, 3651 (to - pg) >> PAGE_SHIFT); 3652 VM_OBJECT_UNLOCK(kernel_object); 3653 VM_WAIT; 3654 VM_OBJECT_LOCK(kernel_object); 3655 goto tryagain; 3656 } 3657 p->valid = VM_PAGE_BITS_ALL; 3658 pmap_qenter(pg, &p, 1); 3659 bp->b_pages[index] = p; 3660 } 3661 VM_OBJECT_UNLOCK(kernel_object); 3662 bp->b_npages = index; 3663 } 3664 3665 /* Return pages associated with this buf to the vm system */ 3666 static void 3667 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3668 { 3669 vm_offset_t pg; 3670 vm_page_t p; 3671 int index, newnpages; 3672 3673 from = round_page(from); 3674 to = round_page(to); 3675 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3676 3677 VM_OBJECT_LOCK(kernel_object); 3678 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3679 p = bp->b_pages[index]; 3680 if (p && (index < bp->b_npages)) { 3681 if (p->busy) { 3682 printf( 3683 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3684 (intmax_t)bp->b_blkno, 3685 (intmax_t)bp->b_lblkno); 3686 } 3687 bp->b_pages[index] = NULL; 3688 pmap_qremove(pg, 1); 3689 vm_page_lock_queues(); 3690 vm_page_unwire(p, 0); 3691 vm_page_free(p); 3692 vm_page_unlock_queues(); 3693 } 3694 } 3695 VM_OBJECT_UNLOCK(kernel_object); 3696 bp->b_npages = newnpages; 3697 } 3698 3699 /* 3700 * Map an IO request into kernel virtual address space. 3701 * 3702 * All requests are (re)mapped into kernel VA space. 3703 * Notice that we use b_bufsize for the size of the buffer 3704 * to be mapped. b_bcount might be modified by the driver. 3705 * 3706 * Note that even if the caller determines that the address space should 3707 * be valid, a race or a smaller-file mapped into a larger space may 3708 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3709 * check the return value. 3710 */ 3711 int 3712 vmapbuf(struct buf *bp) 3713 { 3714 caddr_t addr, kva; 3715 vm_prot_t prot; 3716 int pidx, i; 3717 struct vm_page *m; 3718 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3719 3720 if (bp->b_bufsize < 0) 3721 return (-1); 3722 prot = VM_PROT_READ; 3723 if (bp->b_iocmd == BIO_READ) 3724 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3725 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3726 addr < bp->b_data + bp->b_bufsize; 3727 addr += PAGE_SIZE, pidx++) { 3728 /* 3729 * Do the vm_fault if needed; do the copy-on-write thing 3730 * when reading stuff off device into memory. 3731 * 3732 * NOTE! Must use pmap_extract() because addr may be in 3733 * the userland address space, and kextract is only guarenteed 3734 * to work for the kernland address space (see: sparc64 port). 3735 */ 3736 retry: 3737 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3738 prot) < 0) { 3739 vm_page_lock_queues(); 3740 for (i = 0; i < pidx; ++i) { 3741 vm_page_unhold(bp->b_pages[i]); 3742 bp->b_pages[i] = NULL; 3743 } 3744 vm_page_unlock_queues(); 3745 return(-1); 3746 } 3747 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3748 if (m == NULL) 3749 goto retry; 3750 bp->b_pages[pidx] = m; 3751 } 3752 if (pidx > btoc(MAXPHYS)) 3753 panic("vmapbuf: mapped more than MAXPHYS"); 3754 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3755 3756 kva = bp->b_saveaddr; 3757 bp->b_npages = pidx; 3758 bp->b_saveaddr = bp->b_data; 3759 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3760 return(0); 3761 } 3762 3763 /* 3764 * Free the io map PTEs associated with this IO operation. 3765 * We also invalidate the TLB entries and restore the original b_addr. 3766 */ 3767 void 3768 vunmapbuf(struct buf *bp) 3769 { 3770 int pidx; 3771 int npages; 3772 3773 npages = bp->b_npages; 3774 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3775 vm_page_lock_queues(); 3776 for (pidx = 0; pidx < npages; pidx++) 3777 vm_page_unhold(bp->b_pages[pidx]); 3778 vm_page_unlock_queues(); 3779 3780 bp->b_data = bp->b_saveaddr; 3781 } 3782 3783 void 3784 bdone(struct buf *bp) 3785 { 3786 3787 mtx_lock(&bdonelock); 3788 bp->b_flags |= B_DONE; 3789 wakeup(bp); 3790 mtx_unlock(&bdonelock); 3791 } 3792 3793 void 3794 bwait(struct buf *bp, u_char pri, const char *wchan) 3795 { 3796 3797 mtx_lock(&bdonelock); 3798 while ((bp->b_flags & B_DONE) == 0) 3799 msleep(bp, &bdonelock, pri, wchan, 0); 3800 mtx_unlock(&bdonelock); 3801 } 3802 3803 int 3804 bufsync(struct bufobj *bo, int waitfor, struct thread *td) 3805 { 3806 3807 return (VOP_FSYNC(bo->__bo_vnode, waitfor, td)); 3808 } 3809 3810 void 3811 bufstrategy(struct bufobj *bo, struct buf *bp) 3812 { 3813 int i = 0; 3814 struct vnode *vp; 3815 3816 vp = bp->b_vp; 3817 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3818 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3819 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3820 i = VOP_STRATEGY(vp, bp); 3821 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3822 } 3823 3824 void 3825 bufobj_wrefl(struct bufobj *bo) 3826 { 3827 3828 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3829 ASSERT_BO_LOCKED(bo); 3830 bo->bo_numoutput++; 3831 } 3832 3833 void 3834 bufobj_wref(struct bufobj *bo) 3835 { 3836 3837 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3838 BO_LOCK(bo); 3839 bo->bo_numoutput++; 3840 BO_UNLOCK(bo); 3841 } 3842 3843 void 3844 bufobj_wdrop(struct bufobj *bo) 3845 { 3846 3847 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3848 BO_LOCK(bo); 3849 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3850 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3851 bo->bo_flag &= ~BO_WWAIT; 3852 wakeup(&bo->bo_numoutput); 3853 } 3854 BO_UNLOCK(bo); 3855 } 3856 3857 int 3858 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3859 { 3860 int error; 3861 3862 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3863 ASSERT_BO_LOCKED(bo); 3864 error = 0; 3865 while (bo->bo_numoutput) { 3866 bo->bo_flag |= BO_WWAIT; 3867 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3868 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3869 if (error) 3870 break; 3871 } 3872 return (error); 3873 } 3874 3875 void 3876 bpin(struct buf *bp) 3877 { 3878 mtx_lock(&bpinlock); 3879 bp->b_pin_count++; 3880 mtx_unlock(&bpinlock); 3881 } 3882 3883 void 3884 bunpin(struct buf *bp) 3885 { 3886 mtx_lock(&bpinlock); 3887 if (--bp->b_pin_count == 0) 3888 wakeup(bp); 3889 mtx_unlock(&bpinlock); 3890 } 3891 3892 void 3893 bunpin_wait(struct buf *bp) 3894 { 3895 mtx_lock(&bpinlock); 3896 while (bp->b_pin_count > 0) 3897 msleep(bp, &bpinlock, PRIBIO, "bwunpin", 0); 3898 mtx_unlock(&bpinlock); 3899 } 3900 3901 #include "opt_ddb.h" 3902 #ifdef DDB 3903 #include <ddb/ddb.h> 3904 3905 /* DDB command to show buffer data */ 3906 DB_SHOW_COMMAND(buffer, db_show_buffer) 3907 { 3908 /* get args */ 3909 struct buf *bp = (struct buf *)addr; 3910 3911 if (!have_addr) { 3912 db_printf("usage: show buffer <addr>\n"); 3913 return; 3914 } 3915 3916 db_printf("buf at %p\n", bp); 3917 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3918 db_printf( 3919 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3920 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", 3921 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3922 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); 3923 if (bp->b_npages) { 3924 int i; 3925 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3926 for (i = 0; i < bp->b_npages; i++) { 3927 vm_page_t m; 3928 m = bp->b_pages[i]; 3929 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3930 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3931 if ((i + 1) < bp->b_npages) 3932 db_printf(","); 3933 } 3934 db_printf("\n"); 3935 } 3936 lockmgr_printinfo(&bp->b_lock); 3937 } 3938 3939 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 3940 { 3941 struct buf *bp; 3942 int i; 3943 3944 for (i = 0; i < nbuf; i++) { 3945 bp = &buf[i]; 3946 if (lockcount(&bp->b_lock)) { 3947 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 3948 db_printf("\n"); 3949 } 3950 } 3951 } 3952 #endif /* DDB */ 3953