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|TDP_INBDFLUSH)) == 0) { 965 td->td_pflags |= TDP_INBDFLUSH; 966 BO_BDFLUSH(bo, bp); 967 td->td_pflags &= ~TDP_INBDFLUSH; 968 } else 969 recursiveflushes++; 970 971 bdirty(bp); 972 /* 973 * Set B_CACHE, indicating that the buffer is fully valid. This is 974 * true even of NFS now. 975 */ 976 bp->b_flags |= B_CACHE; 977 978 /* 979 * This bmap keeps the system from needing to do the bmap later, 980 * perhaps when the system is attempting to do a sync. Since it 981 * is likely that the indirect block -- or whatever other datastructure 982 * that the filesystem needs is still in memory now, it is a good 983 * thing to do this. Note also, that if the pageout daemon is 984 * requesting a sync -- there might not be enough memory to do 985 * the bmap then... So, this is important to do. 986 */ 987 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 988 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 989 } 990 991 /* 992 * Set the *dirty* buffer range based upon the VM system dirty pages. 993 */ 994 vfs_setdirty(bp); 995 996 /* 997 * We need to do this here to satisfy the vnode_pager and the 998 * pageout daemon, so that it thinks that the pages have been 999 * "cleaned". Note that since the pages are in a delayed write 1000 * buffer -- the VFS layer "will" see that the pages get written 1001 * out on the next sync, or perhaps the cluster will be completed. 1002 */ 1003 vfs_clean_pages(bp); 1004 bqrelse(bp); 1005 1006 /* 1007 * Wakeup the buffer flushing daemon if we have a lot of dirty 1008 * buffers (midpoint between our recovery point and our stall 1009 * point). 1010 */ 1011 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1012 1013 /* 1014 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1015 * due to the softdep code. 1016 */ 1017 } 1018 1019 /* 1020 * bdirty: 1021 * 1022 * Turn buffer into delayed write request. We must clear BIO_READ and 1023 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1024 * itself to properly update it in the dirty/clean lists. We mark it 1025 * B_DONE to ensure that any asynchronization of the buffer properly 1026 * clears B_DONE ( else a panic will occur later ). 1027 * 1028 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1029 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1030 * should only be called if the buffer is known-good. 1031 * 1032 * Since the buffer is not on a queue, we do not update the numfreebuffers 1033 * count. 1034 * 1035 * The buffer must be on QUEUE_NONE. 1036 */ 1037 void 1038 bdirty(struct buf *bp) 1039 { 1040 1041 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1042 bp, bp->b_vp, bp->b_flags); 1043 KASSERT(BUF_REFCNT(bp) == 1, ("bdirty: bp %p not locked",bp)); 1044 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1045 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1046 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1047 bp->b_flags &= ~(B_RELBUF); 1048 bp->b_iocmd = BIO_WRITE; 1049 1050 if ((bp->b_flags & B_DELWRI) == 0) { 1051 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1052 reassignbuf(bp); 1053 atomic_add_int(&numdirtybuffers, 1); 1054 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1055 } 1056 } 1057 1058 /* 1059 * bundirty: 1060 * 1061 * Clear B_DELWRI for buffer. 1062 * 1063 * Since the buffer is not on a queue, we do not update the numfreebuffers 1064 * count. 1065 * 1066 * The buffer must be on QUEUE_NONE. 1067 */ 1068 1069 void 1070 bundirty(struct buf *bp) 1071 { 1072 1073 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1074 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1075 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1076 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1077 KASSERT(BUF_REFCNT(bp) == 1, ("bundirty: bp %p not locked",bp)); 1078 1079 if (bp->b_flags & B_DELWRI) { 1080 bp->b_flags &= ~B_DELWRI; 1081 reassignbuf(bp); 1082 atomic_subtract_int(&numdirtybuffers, 1); 1083 numdirtywakeup(lodirtybuffers); 1084 } 1085 /* 1086 * Since it is now being written, we can clear its deferred write flag. 1087 */ 1088 bp->b_flags &= ~B_DEFERRED; 1089 } 1090 1091 /* 1092 * bawrite: 1093 * 1094 * Asynchronous write. Start output on a buffer, but do not wait for 1095 * it to complete. The buffer is released when the output completes. 1096 * 1097 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1098 * B_INVAL buffers. Not us. 1099 */ 1100 void 1101 bawrite(struct buf *bp) 1102 { 1103 1104 bp->b_flags |= B_ASYNC; 1105 (void) bwrite(bp); 1106 } 1107 1108 /* 1109 * bwillwrite: 1110 * 1111 * Called prior to the locking of any vnodes when we are expecting to 1112 * write. We do not want to starve the buffer cache with too many 1113 * dirty buffers so we block here. By blocking prior to the locking 1114 * of any vnodes we attempt to avoid the situation where a locked vnode 1115 * prevents the various system daemons from flushing related buffers. 1116 */ 1117 1118 void 1119 bwillwrite(void) 1120 { 1121 1122 if (numdirtybuffers >= hidirtybuffers) { 1123 mtx_lock(&nblock); 1124 while (numdirtybuffers >= hidirtybuffers) { 1125 bd_wakeup(1); 1126 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1127 msleep(&needsbuffer, &nblock, 1128 (PRIBIO + 4), "flswai", 0); 1129 } 1130 mtx_unlock(&nblock); 1131 } 1132 } 1133 1134 /* 1135 * Return true if we have too many dirty buffers. 1136 */ 1137 int 1138 buf_dirty_count_severe(void) 1139 { 1140 1141 return(numdirtybuffers >= hidirtybuffers); 1142 } 1143 1144 /* 1145 * brelse: 1146 * 1147 * Release a busy buffer and, if requested, free its resources. The 1148 * buffer will be stashed in the appropriate bufqueue[] allowing it 1149 * to be accessed later as a cache entity or reused for other purposes. 1150 */ 1151 void 1152 brelse(struct buf *bp) 1153 { 1154 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1155 bp, bp->b_vp, bp->b_flags); 1156 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1157 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1158 1159 if (bp->b_flags & B_MANAGED) { 1160 bqrelse(bp); 1161 return; 1162 } 1163 1164 if (bp->b_iocmd == BIO_WRITE && 1165 (bp->b_ioflags & BIO_ERROR) && 1166 !(bp->b_flags & B_INVAL)) { 1167 /* 1168 * Failed write, redirty. Must clear BIO_ERROR to prevent 1169 * pages from being scrapped. If B_INVAL is set then 1170 * this case is not run and the next case is run to 1171 * destroy the buffer. B_INVAL can occur if the buffer 1172 * is outside the range supported by the underlying device. 1173 */ 1174 bp->b_ioflags &= ~BIO_ERROR; 1175 bdirty(bp); 1176 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1177 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1178 /* 1179 * Either a failed I/O or we were asked to free or not 1180 * cache the buffer. 1181 */ 1182 bp->b_flags |= B_INVAL; 1183 if (!LIST_EMPTY(&bp->b_dep)) 1184 buf_deallocate(bp); 1185 if (bp->b_flags & B_DELWRI) { 1186 atomic_subtract_int(&numdirtybuffers, 1); 1187 numdirtywakeup(lodirtybuffers); 1188 } 1189 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1190 if ((bp->b_flags & B_VMIO) == 0) { 1191 if (bp->b_bufsize) 1192 allocbuf(bp, 0); 1193 if (bp->b_vp) 1194 brelvp(bp); 1195 } 1196 } 1197 1198 /* 1199 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1200 * is called with B_DELWRI set, the underlying pages may wind up 1201 * getting freed causing a previous write (bdwrite()) to get 'lost' 1202 * because pages associated with a B_DELWRI bp are marked clean. 1203 * 1204 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1205 * if B_DELWRI is set. 1206 * 1207 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1208 * on pages to return pages to the VM page queues. 1209 */ 1210 if (bp->b_flags & B_DELWRI) 1211 bp->b_flags &= ~B_RELBUF; 1212 else if (vm_page_count_severe()) { 1213 /* 1214 * XXX This lock may not be necessary since BKGRDINPROG 1215 * cannot be set while we hold the buf lock, it can only be 1216 * cleared if it is already pending. 1217 */ 1218 if (bp->b_vp) { 1219 BO_LOCK(bp->b_bufobj); 1220 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1221 bp->b_flags |= B_RELBUF; 1222 BO_UNLOCK(bp->b_bufobj); 1223 } else 1224 bp->b_flags |= B_RELBUF; 1225 } 1226 1227 /* 1228 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1229 * constituted, not even NFS buffers now. Two flags effect this. If 1230 * B_INVAL, the struct buf is invalidated but the VM object is kept 1231 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1232 * 1233 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1234 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1235 * buffer is also B_INVAL because it hits the re-dirtying code above. 1236 * 1237 * Normally we can do this whether a buffer is B_DELWRI or not. If 1238 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1239 * the commit state and we cannot afford to lose the buffer. If the 1240 * buffer has a background write in progress, we need to keep it 1241 * around to prevent it from being reconstituted and starting a second 1242 * background write. 1243 */ 1244 if ((bp->b_flags & B_VMIO) 1245 && !(bp->b_vp->v_mount != NULL && 1246 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1247 !vn_isdisk(bp->b_vp, NULL) && 1248 (bp->b_flags & B_DELWRI)) 1249 ) { 1250 1251 int i, j, resid; 1252 vm_page_t m; 1253 off_t foff; 1254 vm_pindex_t poff; 1255 vm_object_t obj; 1256 1257 obj = bp->b_bufobj->bo_object; 1258 1259 /* 1260 * Get the base offset and length of the buffer. Note that 1261 * in the VMIO case if the buffer block size is not 1262 * page-aligned then b_data pointer may not be page-aligned. 1263 * But our b_pages[] array *IS* page aligned. 1264 * 1265 * block sizes less then DEV_BSIZE (usually 512) are not 1266 * supported due to the page granularity bits (m->valid, 1267 * m->dirty, etc...). 1268 * 1269 * See man buf(9) for more information 1270 */ 1271 resid = bp->b_bufsize; 1272 foff = bp->b_offset; 1273 VM_OBJECT_LOCK(obj); 1274 for (i = 0; i < bp->b_npages; i++) { 1275 int had_bogus = 0; 1276 1277 m = bp->b_pages[i]; 1278 1279 /* 1280 * If we hit a bogus page, fixup *all* the bogus pages 1281 * now. 1282 */ 1283 if (m == bogus_page) { 1284 poff = OFF_TO_IDX(bp->b_offset); 1285 had_bogus = 1; 1286 1287 for (j = i; j < bp->b_npages; j++) { 1288 vm_page_t mtmp; 1289 mtmp = bp->b_pages[j]; 1290 if (mtmp == bogus_page) { 1291 mtmp = vm_page_lookup(obj, poff + j); 1292 if (!mtmp) { 1293 panic("brelse: page missing\n"); 1294 } 1295 bp->b_pages[j] = mtmp; 1296 } 1297 } 1298 1299 if ((bp->b_flags & B_INVAL) == 0) { 1300 pmap_qenter( 1301 trunc_page((vm_offset_t)bp->b_data), 1302 bp->b_pages, bp->b_npages); 1303 } 1304 m = bp->b_pages[i]; 1305 } 1306 if ((bp->b_flags & B_NOCACHE) || 1307 (bp->b_ioflags & BIO_ERROR)) { 1308 int poffset = foff & PAGE_MASK; 1309 int presid = resid > (PAGE_SIZE - poffset) ? 1310 (PAGE_SIZE - poffset) : resid; 1311 1312 KASSERT(presid >= 0, ("brelse: extra page")); 1313 vm_page_lock_queues(); 1314 vm_page_set_invalid(m, poffset, presid); 1315 vm_page_unlock_queues(); 1316 if (had_bogus) 1317 printf("avoided corruption bug in bogus_page/brelse code\n"); 1318 } 1319 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1320 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1321 } 1322 VM_OBJECT_UNLOCK(obj); 1323 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1324 vfs_vmio_release(bp); 1325 1326 } else if (bp->b_flags & B_VMIO) { 1327 1328 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1329 vfs_vmio_release(bp); 1330 } 1331 1332 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1333 if (bp->b_bufsize != 0) 1334 allocbuf(bp, 0); 1335 if (bp->b_vp != NULL) 1336 brelvp(bp); 1337 } 1338 1339 if (BUF_REFCNT(bp) > 1) { 1340 /* do not release to free list */ 1341 BUF_UNLOCK(bp); 1342 return; 1343 } 1344 1345 /* enqueue */ 1346 mtx_lock(&bqlock); 1347 /* Handle delayed bremfree() processing. */ 1348 if (bp->b_flags & B_REMFREE) 1349 bremfreel(bp); 1350 if (bp->b_qindex != QUEUE_NONE) 1351 panic("brelse: free buffer onto another queue???"); 1352 1353 /* buffers with no memory */ 1354 if (bp->b_bufsize == 0) { 1355 bp->b_flags |= B_INVAL; 1356 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1357 if (bp->b_vflags & BV_BKGRDINPROG) 1358 panic("losing buffer 1"); 1359 if (bp->b_kvasize) { 1360 bp->b_qindex = QUEUE_EMPTYKVA; 1361 } else { 1362 bp->b_qindex = QUEUE_EMPTY; 1363 } 1364 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1365 /* buffers with junk contents */ 1366 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1367 (bp->b_ioflags & BIO_ERROR)) { 1368 bp->b_flags |= B_INVAL; 1369 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1370 if (bp->b_vflags & BV_BKGRDINPROG) 1371 panic("losing buffer 2"); 1372 bp->b_qindex = QUEUE_CLEAN; 1373 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1374 /* remaining buffers */ 1375 } else { 1376 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1377 (B_DELWRI|B_NEEDSGIANT)) 1378 bp->b_qindex = QUEUE_DIRTY_GIANT; 1379 if (bp->b_flags & B_DELWRI) 1380 bp->b_qindex = QUEUE_DIRTY; 1381 else 1382 bp->b_qindex = QUEUE_CLEAN; 1383 if (bp->b_flags & B_AGE) 1384 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1385 else 1386 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1387 } 1388 mtx_unlock(&bqlock); 1389 1390 /* 1391 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1392 * placed the buffer on the correct queue. We must also disassociate 1393 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1394 * find it. 1395 */ 1396 if (bp->b_flags & B_INVAL) { 1397 if (bp->b_flags & B_DELWRI) 1398 bundirty(bp); 1399 if (bp->b_vp) 1400 brelvp(bp); 1401 } 1402 1403 /* 1404 * Fixup numfreebuffers count. The bp is on an appropriate queue 1405 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1406 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1407 * if B_INVAL is set ). 1408 */ 1409 1410 if (!(bp->b_flags & B_DELWRI)) 1411 bufcountwakeup(); 1412 1413 /* 1414 * Something we can maybe free or reuse 1415 */ 1416 if (bp->b_bufsize || bp->b_kvasize) 1417 bufspacewakeup(); 1418 1419 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1420 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1421 panic("brelse: not dirty"); 1422 /* unlock */ 1423 BUF_UNLOCK(bp); 1424 } 1425 1426 /* 1427 * Release a buffer back to the appropriate queue but do not try to free 1428 * it. The buffer is expected to be used again soon. 1429 * 1430 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1431 * biodone() to requeue an async I/O on completion. It is also used when 1432 * known good buffers need to be requeued but we think we may need the data 1433 * again soon. 1434 * 1435 * XXX we should be able to leave the B_RELBUF hint set on completion. 1436 */ 1437 void 1438 bqrelse(struct buf *bp) 1439 { 1440 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1441 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1442 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1443 1444 if (BUF_REFCNT(bp) > 1) { 1445 /* do not release to free list */ 1446 BUF_UNLOCK(bp); 1447 return; 1448 } 1449 1450 if (bp->b_flags & B_MANAGED) { 1451 if (bp->b_flags & B_REMFREE) { 1452 mtx_lock(&bqlock); 1453 bremfreel(bp); 1454 mtx_unlock(&bqlock); 1455 } 1456 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1457 BUF_UNLOCK(bp); 1458 return; 1459 } 1460 1461 mtx_lock(&bqlock); 1462 /* Handle delayed bremfree() processing. */ 1463 if (bp->b_flags & B_REMFREE) 1464 bremfreel(bp); 1465 if (bp->b_qindex != QUEUE_NONE) 1466 panic("bqrelse: free buffer onto another queue???"); 1467 /* buffers with stale but valid contents */ 1468 if (bp->b_flags & B_DELWRI) { 1469 if (bp->b_flags & B_NEEDSGIANT) 1470 bp->b_qindex = QUEUE_DIRTY_GIANT; 1471 else 1472 bp->b_qindex = QUEUE_DIRTY; 1473 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1474 } else { 1475 /* 1476 * XXX This lock may not be necessary since BKGRDINPROG 1477 * cannot be set while we hold the buf lock, it can only be 1478 * cleared if it is already pending. 1479 */ 1480 BO_LOCK(bp->b_bufobj); 1481 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1482 BO_UNLOCK(bp->b_bufobj); 1483 bp->b_qindex = QUEUE_CLEAN; 1484 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1485 b_freelist); 1486 } else { 1487 /* 1488 * We are too low on memory, we have to try to free 1489 * the buffer (most importantly: the wired pages 1490 * making up its backing store) *now*. 1491 */ 1492 BO_UNLOCK(bp->b_bufobj); 1493 mtx_unlock(&bqlock); 1494 brelse(bp); 1495 return; 1496 } 1497 } 1498 mtx_unlock(&bqlock); 1499 1500 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1501 bufcountwakeup(); 1502 1503 /* 1504 * Something we can maybe free or reuse. 1505 */ 1506 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1507 bufspacewakeup(); 1508 1509 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1510 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1511 panic("bqrelse: not dirty"); 1512 /* unlock */ 1513 BUF_UNLOCK(bp); 1514 } 1515 1516 /* Give pages used by the bp back to the VM system (where possible) */ 1517 static void 1518 vfs_vmio_release(struct buf *bp) 1519 { 1520 int i; 1521 vm_page_t m; 1522 1523 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1524 vm_page_lock_queues(); 1525 for (i = 0; i < bp->b_npages; i++) { 1526 m = bp->b_pages[i]; 1527 bp->b_pages[i] = NULL; 1528 /* 1529 * In order to keep page LRU ordering consistent, put 1530 * everything on the inactive queue. 1531 */ 1532 vm_page_unwire(m, 0); 1533 /* 1534 * We don't mess with busy pages, it is 1535 * the responsibility of the process that 1536 * busied the pages to deal with them. 1537 */ 1538 if ((m->oflags & VPO_BUSY) || (m->busy != 0)) 1539 continue; 1540 1541 if (m->wire_count == 0) { 1542 /* 1543 * Might as well free the page if we can and it has 1544 * no valid data. We also free the page if the 1545 * buffer was used for direct I/O 1546 */ 1547 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1548 m->hold_count == 0) { 1549 vm_page_free(m); 1550 } else if (bp->b_flags & B_DIRECT) { 1551 vm_page_try_to_free(m); 1552 } else if (vm_page_count_severe()) { 1553 vm_page_try_to_cache(m); 1554 } 1555 } 1556 } 1557 vm_page_unlock_queues(); 1558 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1559 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1560 1561 if (bp->b_bufsize) { 1562 bufspacewakeup(); 1563 bp->b_bufsize = 0; 1564 } 1565 bp->b_npages = 0; 1566 bp->b_flags &= ~B_VMIO; 1567 if (bp->b_vp) 1568 brelvp(bp); 1569 } 1570 1571 /* 1572 * Check to see if a block at a particular lbn is available for a clustered 1573 * write. 1574 */ 1575 static int 1576 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1577 { 1578 struct buf *bpa; 1579 int match; 1580 1581 match = 0; 1582 1583 /* If the buf isn't in core skip it */ 1584 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1585 return (0); 1586 1587 /* If the buf is busy we don't want to wait for it */ 1588 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1589 return (0); 1590 1591 /* Only cluster with valid clusterable delayed write buffers */ 1592 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1593 (B_DELWRI | B_CLUSTEROK)) 1594 goto done; 1595 1596 if (bpa->b_bufsize != size) 1597 goto done; 1598 1599 /* 1600 * Check to see if it is in the expected place on disk and that the 1601 * block has been mapped. 1602 */ 1603 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1604 match = 1; 1605 done: 1606 BUF_UNLOCK(bpa); 1607 return (match); 1608 } 1609 1610 /* 1611 * vfs_bio_awrite: 1612 * 1613 * Implement clustered async writes for clearing out B_DELWRI buffers. 1614 * This is much better then the old way of writing only one buffer at 1615 * a time. Note that we may not be presented with the buffers in the 1616 * correct order, so we search for the cluster in both directions. 1617 */ 1618 int 1619 vfs_bio_awrite(struct buf *bp) 1620 { 1621 int i; 1622 int j; 1623 daddr_t lblkno = bp->b_lblkno; 1624 struct vnode *vp = bp->b_vp; 1625 int ncl; 1626 int nwritten; 1627 int size; 1628 int maxcl; 1629 1630 /* 1631 * right now we support clustered writing only to regular files. If 1632 * we find a clusterable block we could be in the middle of a cluster 1633 * rather then at the beginning. 1634 */ 1635 if ((vp->v_type == VREG) && 1636 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1637 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1638 1639 size = vp->v_mount->mnt_stat.f_iosize; 1640 maxcl = MAXPHYS / size; 1641 1642 VI_LOCK(vp); 1643 for (i = 1; i < maxcl; i++) 1644 if (vfs_bio_clcheck(vp, size, lblkno + i, 1645 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1646 break; 1647 1648 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1649 if (vfs_bio_clcheck(vp, size, lblkno - j, 1650 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1651 break; 1652 1653 VI_UNLOCK(vp); 1654 --j; 1655 ncl = i + j; 1656 /* 1657 * this is a possible cluster write 1658 */ 1659 if (ncl != 1) { 1660 BUF_UNLOCK(bp); 1661 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1662 return nwritten; 1663 } 1664 } 1665 bremfree(bp); 1666 bp->b_flags |= B_ASYNC; 1667 /* 1668 * default (old) behavior, writing out only one block 1669 * 1670 * XXX returns b_bufsize instead of b_bcount for nwritten? 1671 */ 1672 nwritten = bp->b_bufsize; 1673 (void) bwrite(bp); 1674 1675 return nwritten; 1676 } 1677 1678 /* 1679 * getnewbuf: 1680 * 1681 * Find and initialize a new buffer header, freeing up existing buffers 1682 * in the bufqueues as necessary. The new buffer is returned locked. 1683 * 1684 * Important: B_INVAL is not set. If the caller wishes to throw the 1685 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1686 * 1687 * We block if: 1688 * We have insufficient buffer headers 1689 * We have insufficient buffer space 1690 * buffer_map is too fragmented ( space reservation fails ) 1691 * If we have to flush dirty buffers ( but we try to avoid this ) 1692 * 1693 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1694 * Instead we ask the buf daemon to do it for us. We attempt to 1695 * avoid piecemeal wakeups of the pageout daemon. 1696 */ 1697 1698 static struct buf * 1699 getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1700 { 1701 struct buf *bp; 1702 struct buf *nbp; 1703 int defrag = 0; 1704 int nqindex; 1705 static int flushingbufs; 1706 1707 /* 1708 * We can't afford to block since we might be holding a vnode lock, 1709 * which may prevent system daemons from running. We deal with 1710 * low-memory situations by proactively returning memory and running 1711 * async I/O rather then sync I/O. 1712 */ 1713 1714 atomic_add_int(&getnewbufcalls, 1); 1715 atomic_subtract_int(&getnewbufrestarts, 1); 1716 restart: 1717 atomic_add_int(&getnewbufrestarts, 1); 1718 1719 /* 1720 * Setup for scan. If we do not have enough free buffers, 1721 * we setup a degenerate case that immediately fails. Note 1722 * that if we are specially marked process, we are allowed to 1723 * dip into our reserves. 1724 * 1725 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1726 * 1727 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1728 * However, there are a number of cases (defragging, reusing, ...) 1729 * where we cannot backup. 1730 */ 1731 mtx_lock(&bqlock); 1732 nqindex = QUEUE_EMPTYKVA; 1733 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1734 1735 if (nbp == NULL) { 1736 /* 1737 * If no EMPTYKVA buffers and we are either 1738 * defragging or reusing, locate a CLEAN buffer 1739 * to free or reuse. If bufspace useage is low 1740 * skip this step so we can allocate a new buffer. 1741 */ 1742 if (defrag || bufspace >= lobufspace) { 1743 nqindex = QUEUE_CLEAN; 1744 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1745 } 1746 1747 /* 1748 * If we could not find or were not allowed to reuse a 1749 * CLEAN buffer, check to see if it is ok to use an EMPTY 1750 * buffer. We can only use an EMPTY buffer if allocating 1751 * its KVA would not otherwise run us out of buffer space. 1752 */ 1753 if (nbp == NULL && defrag == 0 && 1754 bufspace + maxsize < hibufspace) { 1755 nqindex = QUEUE_EMPTY; 1756 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1757 } 1758 } 1759 1760 /* 1761 * Run scan, possibly freeing data and/or kva mappings on the fly 1762 * depending. 1763 */ 1764 1765 while ((bp = nbp) != NULL) { 1766 int qindex = nqindex; 1767 1768 /* 1769 * Calculate next bp ( we can only use it if we do not block 1770 * or do other fancy things ). 1771 */ 1772 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1773 switch(qindex) { 1774 case QUEUE_EMPTY: 1775 nqindex = QUEUE_EMPTYKVA; 1776 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1777 break; 1778 /* FALLTHROUGH */ 1779 case QUEUE_EMPTYKVA: 1780 nqindex = QUEUE_CLEAN; 1781 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1782 break; 1783 /* FALLTHROUGH */ 1784 case QUEUE_CLEAN: 1785 /* 1786 * nbp is NULL. 1787 */ 1788 break; 1789 } 1790 } 1791 /* 1792 * If we are defragging then we need a buffer with 1793 * b_kvasize != 0. XXX this situation should no longer 1794 * occur, if defrag is non-zero the buffer's b_kvasize 1795 * should also be non-zero at this point. XXX 1796 */ 1797 if (defrag && bp->b_kvasize == 0) { 1798 printf("Warning: defrag empty buffer %p\n", bp); 1799 continue; 1800 } 1801 1802 /* 1803 * Start freeing the bp. This is somewhat involved. nbp 1804 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1805 */ 1806 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1807 continue; 1808 if (bp->b_vp) { 1809 BO_LOCK(bp->b_bufobj); 1810 if (bp->b_vflags & BV_BKGRDINPROG) { 1811 BO_UNLOCK(bp->b_bufobj); 1812 BUF_UNLOCK(bp); 1813 continue; 1814 } 1815 BO_UNLOCK(bp->b_bufobj); 1816 } 1817 CTR6(KTR_BUF, 1818 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1819 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1820 bp->b_kvasize, bp->b_bufsize, qindex); 1821 1822 /* 1823 * Sanity Checks 1824 */ 1825 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1826 1827 /* 1828 * Note: we no longer distinguish between VMIO and non-VMIO 1829 * buffers. 1830 */ 1831 1832 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1833 1834 bremfreel(bp); 1835 mtx_unlock(&bqlock); 1836 1837 if (qindex == QUEUE_CLEAN) { 1838 if (bp->b_flags & B_VMIO) { 1839 bp->b_flags &= ~B_ASYNC; 1840 vfs_vmio_release(bp); 1841 } 1842 if (bp->b_vp) 1843 brelvp(bp); 1844 } 1845 1846 /* 1847 * NOTE: nbp is now entirely invalid. We can only restart 1848 * the scan from this point on. 1849 * 1850 * Get the rest of the buffer freed up. b_kva* is still 1851 * valid after this operation. 1852 */ 1853 1854 if (bp->b_rcred != NOCRED) { 1855 crfree(bp->b_rcred); 1856 bp->b_rcred = NOCRED; 1857 } 1858 if (bp->b_wcred != NOCRED) { 1859 crfree(bp->b_wcred); 1860 bp->b_wcred = NOCRED; 1861 } 1862 if (!LIST_EMPTY(&bp->b_dep)) 1863 buf_deallocate(bp); 1864 if (bp->b_vflags & BV_BKGRDINPROG) 1865 panic("losing buffer 3"); 1866 KASSERT(bp->b_vp == NULL, 1867 ("bp: %p still has vnode %p. qindex: %d", 1868 bp, bp->b_vp, qindex)); 1869 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1870 ("bp: %p still on a buffer list. xflags %X", 1871 bp, bp->b_xflags)); 1872 1873 if (bp->b_bufsize) 1874 allocbuf(bp, 0); 1875 1876 bp->b_flags = 0; 1877 bp->b_ioflags = 0; 1878 bp->b_xflags = 0; 1879 bp->b_vflags = 0; 1880 bp->b_vp = NULL; 1881 bp->b_blkno = bp->b_lblkno = 0; 1882 bp->b_offset = NOOFFSET; 1883 bp->b_iodone = 0; 1884 bp->b_error = 0; 1885 bp->b_resid = 0; 1886 bp->b_bcount = 0; 1887 bp->b_npages = 0; 1888 bp->b_dirtyoff = bp->b_dirtyend = 0; 1889 bp->b_bufobj = NULL; 1890 bp->b_pin_count = 0; 1891 bp->b_fsprivate1 = NULL; 1892 bp->b_fsprivate2 = NULL; 1893 bp->b_fsprivate3 = NULL; 1894 1895 LIST_INIT(&bp->b_dep); 1896 1897 /* 1898 * If we are defragging then free the buffer. 1899 */ 1900 if (defrag) { 1901 bp->b_flags |= B_INVAL; 1902 bfreekva(bp); 1903 brelse(bp); 1904 defrag = 0; 1905 goto restart; 1906 } 1907 1908 /* 1909 * Notify any waiters for the buffer lock about 1910 * identity change by freeing the buffer. 1911 */ 1912 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp) > 0) { 1913 bp->b_flags |= B_INVAL; 1914 bfreekva(bp); 1915 brelse(bp); 1916 goto restart; 1917 } 1918 1919 /* 1920 * If we are overcomitted then recover the buffer and its 1921 * KVM space. This occurs in rare situations when multiple 1922 * processes are blocked in getnewbuf() or allocbuf(). 1923 */ 1924 if (bufspace >= hibufspace) 1925 flushingbufs = 1; 1926 if (flushingbufs && bp->b_kvasize != 0) { 1927 bp->b_flags |= B_INVAL; 1928 bfreekva(bp); 1929 brelse(bp); 1930 goto restart; 1931 } 1932 if (bufspace < lobufspace) 1933 flushingbufs = 0; 1934 break; 1935 } 1936 1937 /* 1938 * If we exhausted our list, sleep as appropriate. We may have to 1939 * wakeup various daemons and write out some dirty buffers. 1940 * 1941 * Generally we are sleeping due to insufficient buffer space. 1942 */ 1943 1944 if (bp == NULL) { 1945 int flags; 1946 char *waitmsg; 1947 1948 if (defrag) { 1949 flags = VFS_BIO_NEED_BUFSPACE; 1950 waitmsg = "nbufkv"; 1951 } else if (bufspace >= hibufspace) { 1952 waitmsg = "nbufbs"; 1953 flags = VFS_BIO_NEED_BUFSPACE; 1954 } else { 1955 waitmsg = "newbuf"; 1956 flags = VFS_BIO_NEED_ANY; 1957 } 1958 mtx_lock(&nblock); 1959 needsbuffer |= flags; 1960 mtx_unlock(&nblock); 1961 mtx_unlock(&bqlock); 1962 1963 bd_speedup(); /* heeeelp */ 1964 1965 mtx_lock(&nblock); 1966 while (needsbuffer & flags) { 1967 if (msleep(&needsbuffer, &nblock, 1968 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1969 mtx_unlock(&nblock); 1970 return (NULL); 1971 } 1972 } 1973 mtx_unlock(&nblock); 1974 } else { 1975 /* 1976 * We finally have a valid bp. We aren't quite out of the 1977 * woods, we still have to reserve kva space. In order 1978 * to keep fragmentation sane we only allocate kva in 1979 * BKVASIZE chunks. 1980 */ 1981 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1982 1983 if (maxsize != bp->b_kvasize) { 1984 vm_offset_t addr = 0; 1985 1986 bfreekva(bp); 1987 1988 vm_map_lock(buffer_map); 1989 if (vm_map_findspace(buffer_map, 1990 vm_map_min(buffer_map), maxsize, &addr)) { 1991 /* 1992 * Uh oh. Buffer map is to fragmented. We 1993 * must defragment the map. 1994 */ 1995 atomic_add_int(&bufdefragcnt, 1); 1996 vm_map_unlock(buffer_map); 1997 defrag = 1; 1998 bp->b_flags |= B_INVAL; 1999 brelse(bp); 2000 goto restart; 2001 } 2002 if (addr) { 2003 vm_map_insert(buffer_map, NULL, 0, 2004 addr, addr + maxsize, 2005 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2006 2007 bp->b_kvabase = (caddr_t) addr; 2008 bp->b_kvasize = maxsize; 2009 atomic_add_int(&bufspace, bp->b_kvasize); 2010 atomic_add_int(&bufreusecnt, 1); 2011 } 2012 vm_map_unlock(buffer_map); 2013 } 2014 bp->b_saveaddr = bp->b_kvabase; 2015 bp->b_data = bp->b_saveaddr; 2016 } 2017 return(bp); 2018 } 2019 2020 /* 2021 * buf_daemon: 2022 * 2023 * buffer flushing daemon. Buffers are normally flushed by the 2024 * update daemon but if it cannot keep up this process starts to 2025 * take the load in an attempt to prevent getnewbuf() from blocking. 2026 */ 2027 2028 static struct kproc_desc buf_kp = { 2029 "bufdaemon", 2030 buf_daemon, 2031 &bufdaemonproc 2032 }; 2033 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2034 2035 static void 2036 buf_daemon() 2037 { 2038 2039 /* 2040 * This process needs to be suspended prior to shutdown sync. 2041 */ 2042 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2043 SHUTDOWN_PRI_LAST); 2044 2045 /* 2046 * This process is allowed to take the buffer cache to the limit 2047 */ 2048 curthread->td_pflags |= TDP_NORUNNINGBUF; 2049 mtx_lock(&bdlock); 2050 for (;;) { 2051 bd_request = 0; 2052 mtx_unlock(&bdlock); 2053 2054 kthread_suspend_check(bufdaemonproc); 2055 2056 /* 2057 * Do the flush. Limit the amount of in-transit I/O we 2058 * allow to build up, otherwise we would completely saturate 2059 * the I/O system. Wakeup any waiting processes before we 2060 * normally would so they can run in parallel with our drain. 2061 */ 2062 while (numdirtybuffers > lodirtybuffers) { 2063 int flushed; 2064 2065 flushed = flushbufqueues(QUEUE_DIRTY, 0); 2066 /* The list empty check here is slightly racy */ 2067 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2068 mtx_lock(&Giant); 2069 flushed += flushbufqueues(QUEUE_DIRTY_GIANT, 0); 2070 mtx_unlock(&Giant); 2071 } 2072 if (flushed == 0) { 2073 /* 2074 * Could not find any buffers without rollback 2075 * dependencies, so just write the first one 2076 * in the hopes of eventually making progress. 2077 */ 2078 flushbufqueues(QUEUE_DIRTY, 1); 2079 if (!TAILQ_EMPTY( 2080 &bufqueues[QUEUE_DIRTY_GIANT])) { 2081 mtx_lock(&Giant); 2082 flushbufqueues(QUEUE_DIRTY_GIANT, 1); 2083 mtx_unlock(&Giant); 2084 } 2085 break; 2086 } 2087 uio_yield(); 2088 } 2089 2090 /* 2091 * Only clear bd_request if we have reached our low water 2092 * mark. The buf_daemon normally waits 1 second and 2093 * then incrementally flushes any dirty buffers that have 2094 * built up, within reason. 2095 * 2096 * If we were unable to hit our low water mark and couldn't 2097 * find any flushable buffers, we sleep half a second. 2098 * Otherwise we loop immediately. 2099 */ 2100 mtx_lock(&bdlock); 2101 if (numdirtybuffers <= lodirtybuffers) { 2102 /* 2103 * We reached our low water mark, reset the 2104 * request and sleep until we are needed again. 2105 * The sleep is just so the suspend code works. 2106 */ 2107 bd_request = 0; 2108 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2109 } else { 2110 /* 2111 * We couldn't find any flushable dirty buffers but 2112 * still have too many dirty buffers, we 2113 * have to sleep and try again. (rare) 2114 */ 2115 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2116 } 2117 } 2118 } 2119 2120 /* 2121 * flushbufqueues: 2122 * 2123 * Try to flush a buffer in the dirty queue. We must be careful to 2124 * free up B_INVAL buffers instead of write them, which NFS is 2125 * particularly sensitive to. 2126 */ 2127 static int flushwithdeps = 0; 2128 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2129 0, "Number of buffers flushed with dependecies that require rollbacks"); 2130 2131 static int 2132 flushbufqueues(int queue, int flushdeps) 2133 { 2134 struct thread *td = curthread; 2135 struct buf sentinel; 2136 struct vnode *vp; 2137 struct mount *mp; 2138 struct buf *bp; 2139 int hasdeps; 2140 int flushed; 2141 int target; 2142 2143 target = numdirtybuffers - lodirtybuffers; 2144 if (flushdeps && target > 2) 2145 target /= 2; 2146 flushed = 0; 2147 bp = NULL; 2148 mtx_lock(&bqlock); 2149 TAILQ_INSERT_TAIL(&bufqueues[queue], &sentinel, b_freelist); 2150 while (flushed != target) { 2151 bp = TAILQ_FIRST(&bufqueues[queue]); 2152 if (bp == &sentinel) 2153 break; 2154 TAILQ_REMOVE(&bufqueues[queue], bp, b_freelist); 2155 TAILQ_INSERT_TAIL(&bufqueues[queue], bp, b_freelist); 2156 2157 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2158 continue; 2159 if (bp->b_pin_count > 0) { 2160 BUF_UNLOCK(bp); 2161 continue; 2162 } 2163 BO_LOCK(bp->b_bufobj); 2164 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2165 (bp->b_flags & B_DELWRI) == 0) { 2166 BO_UNLOCK(bp->b_bufobj); 2167 BUF_UNLOCK(bp); 2168 continue; 2169 } 2170 BO_UNLOCK(bp->b_bufobj); 2171 if (bp->b_flags & B_INVAL) { 2172 bremfreel(bp); 2173 mtx_unlock(&bqlock); 2174 brelse(bp); 2175 flushed++; 2176 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2177 mtx_lock(&bqlock); 2178 continue; 2179 } 2180 2181 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2182 if (flushdeps == 0) { 2183 BUF_UNLOCK(bp); 2184 continue; 2185 } 2186 hasdeps = 1; 2187 } else 2188 hasdeps = 0; 2189 /* 2190 * We must hold the lock on a vnode before writing 2191 * one of its buffers. Otherwise we may confuse, or 2192 * in the case of a snapshot vnode, deadlock the 2193 * system. 2194 * 2195 * The lock order here is the reverse of the normal 2196 * of vnode followed by buf lock. This is ok because 2197 * the NOWAIT will prevent deadlock. 2198 */ 2199 vp = bp->b_vp; 2200 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2201 BUF_UNLOCK(bp); 2202 continue; 2203 } 2204 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2205 mtx_unlock(&bqlock); 2206 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2207 bp, bp->b_vp, bp->b_flags); 2208 vfs_bio_awrite(bp); 2209 vn_finished_write(mp); 2210 VOP_UNLOCK(vp, 0, td); 2211 flushwithdeps += hasdeps; 2212 flushed++; 2213 waitrunningbufspace(); 2214 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2215 mtx_lock(&bqlock); 2216 continue; 2217 } 2218 vn_finished_write(mp); 2219 BUF_UNLOCK(bp); 2220 } 2221 TAILQ_REMOVE(&bufqueues[queue], &sentinel, b_freelist); 2222 mtx_unlock(&bqlock); 2223 return (flushed); 2224 } 2225 2226 /* 2227 * Check to see if a block is currently memory resident. 2228 */ 2229 struct buf * 2230 incore(struct bufobj *bo, daddr_t blkno) 2231 { 2232 struct buf *bp; 2233 2234 BO_LOCK(bo); 2235 bp = gbincore(bo, blkno); 2236 BO_UNLOCK(bo); 2237 return (bp); 2238 } 2239 2240 /* 2241 * Returns true if no I/O is needed to access the 2242 * associated VM object. This is like incore except 2243 * it also hunts around in the VM system for the data. 2244 */ 2245 2246 static int 2247 inmem(struct vnode * vp, daddr_t blkno) 2248 { 2249 vm_object_t obj; 2250 vm_offset_t toff, tinc, size; 2251 vm_page_t m; 2252 vm_ooffset_t off; 2253 2254 ASSERT_VOP_LOCKED(vp, "inmem"); 2255 2256 if (incore(&vp->v_bufobj, blkno)) 2257 return 1; 2258 if (vp->v_mount == NULL) 2259 return 0; 2260 obj = vp->v_object; 2261 if (obj == NULL) 2262 return (0); 2263 2264 size = PAGE_SIZE; 2265 if (size > vp->v_mount->mnt_stat.f_iosize) 2266 size = vp->v_mount->mnt_stat.f_iosize; 2267 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2268 2269 VM_OBJECT_LOCK(obj); 2270 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2271 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2272 if (!m) 2273 goto notinmem; 2274 tinc = size; 2275 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2276 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2277 if (vm_page_is_valid(m, 2278 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2279 goto notinmem; 2280 } 2281 VM_OBJECT_UNLOCK(obj); 2282 return 1; 2283 2284 notinmem: 2285 VM_OBJECT_UNLOCK(obj); 2286 return (0); 2287 } 2288 2289 /* 2290 * vfs_setdirty: 2291 * 2292 * Sets the dirty range for a buffer based on the status of the dirty 2293 * bits in the pages comprising the buffer. 2294 * 2295 * The range is limited to the size of the buffer. 2296 * 2297 * This routine is primarily used by NFS, but is generalized for the 2298 * B_VMIO case. 2299 */ 2300 static void 2301 vfs_setdirty(struct buf *bp) 2302 { 2303 2304 /* 2305 * Degenerate case - empty buffer 2306 */ 2307 2308 if (bp->b_bufsize == 0) 2309 return; 2310 2311 /* 2312 * We qualify the scan for modified pages on whether the 2313 * object has been flushed yet. 2314 */ 2315 2316 if ((bp->b_flags & B_VMIO) == 0) 2317 return; 2318 2319 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2320 vfs_setdirty_locked_object(bp); 2321 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2322 } 2323 2324 static void 2325 vfs_setdirty_locked_object(struct buf *bp) 2326 { 2327 vm_object_t object; 2328 int i; 2329 2330 object = bp->b_bufobj->bo_object; 2331 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2332 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2333 vm_offset_t boffset; 2334 vm_offset_t eoffset; 2335 2336 vm_page_lock_queues(); 2337 /* 2338 * test the pages to see if they have been modified directly 2339 * by users through the VM system. 2340 */ 2341 for (i = 0; i < bp->b_npages; i++) 2342 vm_page_test_dirty(bp->b_pages[i]); 2343 2344 /* 2345 * Calculate the encompassing dirty range, boffset and eoffset, 2346 * (eoffset - boffset) bytes. 2347 */ 2348 2349 for (i = 0; i < bp->b_npages; i++) { 2350 if (bp->b_pages[i]->dirty) 2351 break; 2352 } 2353 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2354 2355 for (i = bp->b_npages - 1; i >= 0; --i) { 2356 if (bp->b_pages[i]->dirty) { 2357 break; 2358 } 2359 } 2360 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2361 2362 vm_page_unlock_queues(); 2363 /* 2364 * Fit it to the buffer. 2365 */ 2366 2367 if (eoffset > bp->b_bcount) 2368 eoffset = bp->b_bcount; 2369 2370 /* 2371 * If we have a good dirty range, merge with the existing 2372 * dirty range. 2373 */ 2374 2375 if (boffset < eoffset) { 2376 if (bp->b_dirtyoff > boffset) 2377 bp->b_dirtyoff = boffset; 2378 if (bp->b_dirtyend < eoffset) 2379 bp->b_dirtyend = eoffset; 2380 } 2381 } 2382 } 2383 2384 /* 2385 * getblk: 2386 * 2387 * Get a block given a specified block and offset into a file/device. 2388 * The buffers B_DONE bit will be cleared on return, making it almost 2389 * ready for an I/O initiation. B_INVAL may or may not be set on 2390 * return. The caller should clear B_INVAL prior to initiating a 2391 * READ. 2392 * 2393 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2394 * an existing buffer. 2395 * 2396 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2397 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2398 * and then cleared based on the backing VM. If the previous buffer is 2399 * non-0-sized but invalid, B_CACHE will be cleared. 2400 * 2401 * If getblk() must create a new buffer, the new buffer is returned with 2402 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2403 * case it is returned with B_INVAL clear and B_CACHE set based on the 2404 * backing VM. 2405 * 2406 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2407 * B_CACHE bit is clear. 2408 * 2409 * What this means, basically, is that the caller should use B_CACHE to 2410 * determine whether the buffer is fully valid or not and should clear 2411 * B_INVAL prior to issuing a read. If the caller intends to validate 2412 * the buffer by loading its data area with something, the caller needs 2413 * to clear B_INVAL. If the caller does this without issuing an I/O, 2414 * the caller should set B_CACHE ( as an optimization ), else the caller 2415 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2416 * a write attempt or if it was a successfull read. If the caller 2417 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2418 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2419 */ 2420 struct buf * 2421 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2422 int flags) 2423 { 2424 struct buf *bp; 2425 struct bufobj *bo; 2426 int error; 2427 2428 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2429 ASSERT_VOP_LOCKED(vp, "getblk"); 2430 if (size > MAXBSIZE) 2431 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2432 2433 bo = &vp->v_bufobj; 2434 loop: 2435 /* 2436 * Block if we are low on buffers. Certain processes are allowed 2437 * to completely exhaust the buffer cache. 2438 * 2439 * If this check ever becomes a bottleneck it may be better to 2440 * move it into the else, when gbincore() fails. At the moment 2441 * it isn't a problem. 2442 * 2443 * XXX remove if 0 sections (clean this up after its proven) 2444 */ 2445 if (numfreebuffers == 0) { 2446 if (TD_IS_IDLETHREAD(curthread)) 2447 return NULL; 2448 mtx_lock(&nblock); 2449 needsbuffer |= VFS_BIO_NEED_ANY; 2450 mtx_unlock(&nblock); 2451 } 2452 2453 VI_LOCK(vp); 2454 bp = gbincore(bo, blkno); 2455 if (bp != NULL) { 2456 int lockflags; 2457 /* 2458 * Buffer is in-core. If the buffer is not busy, it must 2459 * be on a queue. 2460 */ 2461 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2462 2463 if (flags & GB_LOCK_NOWAIT) 2464 lockflags |= LK_NOWAIT; 2465 2466 error = BUF_TIMELOCK(bp, lockflags, 2467 VI_MTX(vp), "getblk", slpflag, slptimeo); 2468 2469 /* 2470 * If we slept and got the lock we have to restart in case 2471 * the buffer changed identities. 2472 */ 2473 if (error == ENOLCK) 2474 goto loop; 2475 /* We timed out or were interrupted. */ 2476 else if (error) 2477 return (NULL); 2478 2479 /* 2480 * The buffer is locked. B_CACHE is cleared if the buffer is 2481 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2482 * and for a VMIO buffer B_CACHE is adjusted according to the 2483 * backing VM cache. 2484 */ 2485 if (bp->b_flags & B_INVAL) 2486 bp->b_flags &= ~B_CACHE; 2487 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2488 bp->b_flags |= B_CACHE; 2489 bremfree(bp); 2490 2491 /* 2492 * check for size inconsistancies for non-VMIO case. 2493 */ 2494 2495 if (bp->b_bcount != size) { 2496 if ((bp->b_flags & B_VMIO) == 0 || 2497 (size > bp->b_kvasize)) { 2498 if (bp->b_flags & B_DELWRI) { 2499 /* 2500 * If buffer is pinned and caller does 2501 * not want sleep waiting for it to be 2502 * unpinned, bail out 2503 * */ 2504 if (bp->b_pin_count > 0) { 2505 if (flags & GB_LOCK_NOWAIT) { 2506 bqrelse(bp); 2507 return (NULL); 2508 } else { 2509 bunpin_wait(bp); 2510 } 2511 } 2512 bp->b_flags |= B_NOCACHE; 2513 bwrite(bp); 2514 } else { 2515 if (LIST_EMPTY(&bp->b_dep)) { 2516 bp->b_flags |= B_RELBUF; 2517 brelse(bp); 2518 } else { 2519 bp->b_flags |= B_NOCACHE; 2520 bwrite(bp); 2521 } 2522 } 2523 goto loop; 2524 } 2525 } 2526 2527 /* 2528 * If the size is inconsistant in the VMIO case, we can resize 2529 * the buffer. This might lead to B_CACHE getting set or 2530 * cleared. If the size has not changed, B_CACHE remains 2531 * unchanged from its previous state. 2532 */ 2533 2534 if (bp->b_bcount != size) 2535 allocbuf(bp, size); 2536 2537 KASSERT(bp->b_offset != NOOFFSET, 2538 ("getblk: no buffer offset")); 2539 2540 /* 2541 * A buffer with B_DELWRI set and B_CACHE clear must 2542 * be committed before we can return the buffer in 2543 * order to prevent the caller from issuing a read 2544 * ( due to B_CACHE not being set ) and overwriting 2545 * it. 2546 * 2547 * Most callers, including NFS and FFS, need this to 2548 * operate properly either because they assume they 2549 * can issue a read if B_CACHE is not set, or because 2550 * ( for example ) an uncached B_DELWRI might loop due 2551 * to softupdates re-dirtying the buffer. In the latter 2552 * case, B_CACHE is set after the first write completes, 2553 * preventing further loops. 2554 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2555 * above while extending the buffer, we cannot allow the 2556 * buffer to remain with B_CACHE set after the write 2557 * completes or it will represent a corrupt state. To 2558 * deal with this we set B_NOCACHE to scrap the buffer 2559 * after the write. 2560 * 2561 * We might be able to do something fancy, like setting 2562 * B_CACHE in bwrite() except if B_DELWRI is already set, 2563 * so the below call doesn't set B_CACHE, but that gets real 2564 * confusing. This is much easier. 2565 */ 2566 2567 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2568 bp->b_flags |= B_NOCACHE; 2569 bwrite(bp); 2570 goto loop; 2571 } 2572 bp->b_flags &= ~B_DONE; 2573 } else { 2574 int bsize, maxsize, vmio; 2575 off_t offset; 2576 2577 /* 2578 * Buffer is not in-core, create new buffer. The buffer 2579 * returned by getnewbuf() is locked. Note that the returned 2580 * buffer is also considered valid (not marked B_INVAL). 2581 */ 2582 VI_UNLOCK(vp); 2583 /* 2584 * If the user does not want us to create the buffer, bail out 2585 * here. 2586 */ 2587 if (flags & GB_NOCREAT) 2588 return NULL; 2589 bsize = bo->bo_bsize; 2590 offset = blkno * bsize; 2591 vmio = vp->v_object != NULL; 2592 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2593 maxsize = imax(maxsize, bsize); 2594 2595 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2596 if (bp == NULL) { 2597 if (slpflag || slptimeo) 2598 return NULL; 2599 goto loop; 2600 } 2601 2602 /* 2603 * This code is used to make sure that a buffer is not 2604 * created while the getnewbuf routine is blocked. 2605 * This can be a problem whether the vnode is locked or not. 2606 * If the buffer is created out from under us, we have to 2607 * throw away the one we just created. 2608 * 2609 * Note: this must occur before we associate the buffer 2610 * with the vp especially considering limitations in 2611 * the splay tree implementation when dealing with duplicate 2612 * lblkno's. 2613 */ 2614 BO_LOCK(bo); 2615 if (gbincore(bo, blkno)) { 2616 BO_UNLOCK(bo); 2617 bp->b_flags |= B_INVAL; 2618 brelse(bp); 2619 goto loop; 2620 } 2621 2622 /* 2623 * Insert the buffer into the hash, so that it can 2624 * be found by incore. 2625 */ 2626 bp->b_blkno = bp->b_lblkno = blkno; 2627 bp->b_offset = offset; 2628 bgetvp(vp, bp); 2629 BO_UNLOCK(bo); 2630 2631 /* 2632 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2633 * buffer size starts out as 0, B_CACHE will be set by 2634 * allocbuf() for the VMIO case prior to it testing the 2635 * backing store for validity. 2636 */ 2637 2638 if (vmio) { 2639 bp->b_flags |= B_VMIO; 2640 #if defined(VFS_BIO_DEBUG) 2641 if (vn_canvmio(vp) != TRUE) 2642 printf("getblk: VMIO on vnode type %d\n", 2643 vp->v_type); 2644 #endif 2645 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2646 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2647 bp, vp->v_object, bp->b_bufobj->bo_object)); 2648 } else { 2649 bp->b_flags &= ~B_VMIO; 2650 KASSERT(bp->b_bufobj->bo_object == NULL, 2651 ("ARGH! has b_bufobj->bo_object %p %p\n", 2652 bp, bp->b_bufobj->bo_object)); 2653 } 2654 2655 allocbuf(bp, size); 2656 bp->b_flags &= ~B_DONE; 2657 } 2658 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2659 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2660 KASSERT(bp->b_bufobj == bo, 2661 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2662 return (bp); 2663 } 2664 2665 /* 2666 * Get an empty, disassociated buffer of given size. The buffer is initially 2667 * set to B_INVAL. 2668 */ 2669 struct buf * 2670 geteblk(int size) 2671 { 2672 struct buf *bp; 2673 int maxsize; 2674 2675 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2676 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2677 continue; 2678 allocbuf(bp, size); 2679 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2680 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2681 return (bp); 2682 } 2683 2684 2685 /* 2686 * This code constitutes the buffer memory from either anonymous system 2687 * memory (in the case of non-VMIO operations) or from an associated 2688 * VM object (in the case of VMIO operations). This code is able to 2689 * resize a buffer up or down. 2690 * 2691 * Note that this code is tricky, and has many complications to resolve 2692 * deadlock or inconsistant data situations. Tread lightly!!! 2693 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2694 * the caller. Calling this code willy nilly can result in the loss of data. 2695 * 2696 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2697 * B_CACHE for the non-VMIO case. 2698 */ 2699 2700 int 2701 allocbuf(struct buf *bp, int size) 2702 { 2703 int newbsize, mbsize; 2704 int i; 2705 2706 if (BUF_REFCNT(bp) == 0) 2707 panic("allocbuf: buffer not busy"); 2708 2709 if (bp->b_kvasize < size) 2710 panic("allocbuf: buffer too small"); 2711 2712 if ((bp->b_flags & B_VMIO) == 0) { 2713 caddr_t origbuf; 2714 int origbufsize; 2715 /* 2716 * Just get anonymous memory from the kernel. Don't 2717 * mess with B_CACHE. 2718 */ 2719 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2720 if (bp->b_flags & B_MALLOC) 2721 newbsize = mbsize; 2722 else 2723 newbsize = round_page(size); 2724 2725 if (newbsize < bp->b_bufsize) { 2726 /* 2727 * malloced buffers are not shrunk 2728 */ 2729 if (bp->b_flags & B_MALLOC) { 2730 if (newbsize) { 2731 bp->b_bcount = size; 2732 } else { 2733 free(bp->b_data, M_BIOBUF); 2734 if (bp->b_bufsize) { 2735 atomic_subtract_int( 2736 &bufmallocspace, 2737 bp->b_bufsize); 2738 bufspacewakeup(); 2739 bp->b_bufsize = 0; 2740 } 2741 bp->b_saveaddr = bp->b_kvabase; 2742 bp->b_data = bp->b_saveaddr; 2743 bp->b_bcount = 0; 2744 bp->b_flags &= ~B_MALLOC; 2745 } 2746 return 1; 2747 } 2748 vm_hold_free_pages( 2749 bp, 2750 (vm_offset_t) bp->b_data + newbsize, 2751 (vm_offset_t) bp->b_data + bp->b_bufsize); 2752 } else if (newbsize > bp->b_bufsize) { 2753 /* 2754 * We only use malloced memory on the first allocation. 2755 * and revert to page-allocated memory when the buffer 2756 * grows. 2757 */ 2758 /* 2759 * There is a potential smp race here that could lead 2760 * to bufmallocspace slightly passing the max. It 2761 * is probably extremely rare and not worth worrying 2762 * over. 2763 */ 2764 if ( (bufmallocspace < maxbufmallocspace) && 2765 (bp->b_bufsize == 0) && 2766 (mbsize <= PAGE_SIZE/2)) { 2767 2768 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2769 bp->b_bufsize = mbsize; 2770 bp->b_bcount = size; 2771 bp->b_flags |= B_MALLOC; 2772 atomic_add_int(&bufmallocspace, mbsize); 2773 return 1; 2774 } 2775 origbuf = NULL; 2776 origbufsize = 0; 2777 /* 2778 * If the buffer is growing on its other-than-first allocation, 2779 * then we revert to the page-allocation scheme. 2780 */ 2781 if (bp->b_flags & B_MALLOC) { 2782 origbuf = bp->b_data; 2783 origbufsize = bp->b_bufsize; 2784 bp->b_data = bp->b_kvabase; 2785 if (bp->b_bufsize) { 2786 atomic_subtract_int(&bufmallocspace, 2787 bp->b_bufsize); 2788 bufspacewakeup(); 2789 bp->b_bufsize = 0; 2790 } 2791 bp->b_flags &= ~B_MALLOC; 2792 newbsize = round_page(newbsize); 2793 } 2794 vm_hold_load_pages( 2795 bp, 2796 (vm_offset_t) bp->b_data + bp->b_bufsize, 2797 (vm_offset_t) bp->b_data + newbsize); 2798 if (origbuf) { 2799 bcopy(origbuf, bp->b_data, origbufsize); 2800 free(origbuf, M_BIOBUF); 2801 } 2802 } 2803 } else { 2804 int desiredpages; 2805 2806 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2807 desiredpages = (size == 0) ? 0 : 2808 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2809 2810 if (bp->b_flags & B_MALLOC) 2811 panic("allocbuf: VMIO buffer can't be malloced"); 2812 /* 2813 * Set B_CACHE initially if buffer is 0 length or will become 2814 * 0-length. 2815 */ 2816 if (size == 0 || bp->b_bufsize == 0) 2817 bp->b_flags |= B_CACHE; 2818 2819 if (newbsize < bp->b_bufsize) { 2820 /* 2821 * DEV_BSIZE aligned new buffer size is less then the 2822 * DEV_BSIZE aligned existing buffer size. Figure out 2823 * if we have to remove any pages. 2824 */ 2825 if (desiredpages < bp->b_npages) { 2826 vm_page_t m; 2827 2828 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2829 vm_page_lock_queues(); 2830 for (i = desiredpages; i < bp->b_npages; i++) { 2831 /* 2832 * the page is not freed here -- it 2833 * is the responsibility of 2834 * vnode_pager_setsize 2835 */ 2836 m = bp->b_pages[i]; 2837 KASSERT(m != bogus_page, 2838 ("allocbuf: bogus page found")); 2839 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2840 vm_page_lock_queues(); 2841 2842 bp->b_pages[i] = NULL; 2843 vm_page_unwire(m, 0); 2844 } 2845 vm_page_unlock_queues(); 2846 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2847 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2848 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2849 bp->b_npages = desiredpages; 2850 } 2851 } else if (size > bp->b_bcount) { 2852 /* 2853 * We are growing the buffer, possibly in a 2854 * byte-granular fashion. 2855 */ 2856 struct vnode *vp; 2857 vm_object_t obj; 2858 vm_offset_t toff; 2859 vm_offset_t tinc; 2860 2861 /* 2862 * Step 1, bring in the VM pages from the object, 2863 * allocating them if necessary. We must clear 2864 * B_CACHE if these pages are not valid for the 2865 * range covered by the buffer. 2866 */ 2867 2868 vp = bp->b_vp; 2869 obj = bp->b_bufobj->bo_object; 2870 2871 VM_OBJECT_LOCK(obj); 2872 while (bp->b_npages < desiredpages) { 2873 vm_page_t m; 2874 vm_pindex_t pi; 2875 2876 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2877 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2878 /* 2879 * note: must allocate system pages 2880 * since blocking here could intefere 2881 * with paging I/O, no matter which 2882 * process we are. 2883 */ 2884 m = vm_page_alloc(obj, pi, 2885 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2886 VM_ALLOC_WIRED); 2887 if (m == NULL) { 2888 atomic_add_int(&vm_pageout_deficit, 2889 desiredpages - bp->b_npages); 2890 VM_OBJECT_UNLOCK(obj); 2891 VM_WAIT; 2892 VM_OBJECT_LOCK(obj); 2893 } else { 2894 bp->b_flags &= ~B_CACHE; 2895 bp->b_pages[bp->b_npages] = m; 2896 ++bp->b_npages; 2897 } 2898 continue; 2899 } 2900 2901 /* 2902 * We found a page. If we have to sleep on it, 2903 * retry because it might have gotten freed out 2904 * from under us. 2905 * 2906 * We can only test VPO_BUSY here. Blocking on 2907 * m->busy might lead to a deadlock: 2908 * 2909 * vm_fault->getpages->cluster_read->allocbuf 2910 * 2911 */ 2912 vm_page_lock_queues(); 2913 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2914 continue; 2915 2916 /* 2917 * We have a good page. Should we wakeup the 2918 * page daemon? 2919 */ 2920 if ((curproc != pageproc) && 2921 (VM_PAGE_INQUEUE1(m, PQ_CACHE)) && 2922 ((cnt.v_free_count + cnt.v_cache_count) < 2923 (cnt.v_free_min + cnt.v_cache_min))) { 2924 pagedaemon_wakeup(); 2925 } 2926 vm_page_wire(m); 2927 vm_page_unlock_queues(); 2928 bp->b_pages[bp->b_npages] = m; 2929 ++bp->b_npages; 2930 } 2931 2932 /* 2933 * Step 2. We've loaded the pages into the buffer, 2934 * we have to figure out if we can still have B_CACHE 2935 * set. Note that B_CACHE is set according to the 2936 * byte-granular range ( bcount and size ), new the 2937 * aligned range ( newbsize ). 2938 * 2939 * The VM test is against m->valid, which is DEV_BSIZE 2940 * aligned. Needless to say, the validity of the data 2941 * needs to also be DEV_BSIZE aligned. Note that this 2942 * fails with NFS if the server or some other client 2943 * extends the file's EOF. If our buffer is resized, 2944 * B_CACHE may remain set! XXX 2945 */ 2946 2947 toff = bp->b_bcount; 2948 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2949 2950 while ((bp->b_flags & B_CACHE) && toff < size) { 2951 vm_pindex_t pi; 2952 2953 if (tinc > (size - toff)) 2954 tinc = size - toff; 2955 2956 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2957 PAGE_SHIFT; 2958 2959 vfs_buf_test_cache( 2960 bp, 2961 bp->b_offset, 2962 toff, 2963 tinc, 2964 bp->b_pages[pi] 2965 ); 2966 toff += tinc; 2967 tinc = PAGE_SIZE; 2968 } 2969 VM_OBJECT_UNLOCK(obj); 2970 2971 /* 2972 * Step 3, fixup the KVM pmap. Remember that 2973 * bp->b_data is relative to bp->b_offset, but 2974 * bp->b_offset may be offset into the first page. 2975 */ 2976 2977 bp->b_data = (caddr_t) 2978 trunc_page((vm_offset_t)bp->b_data); 2979 pmap_qenter( 2980 (vm_offset_t)bp->b_data, 2981 bp->b_pages, 2982 bp->b_npages 2983 ); 2984 2985 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2986 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2987 } 2988 } 2989 if (newbsize < bp->b_bufsize) 2990 bufspacewakeup(); 2991 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2992 bp->b_bcount = size; /* requested buffer size */ 2993 return 1; 2994 } 2995 2996 void 2997 biodone(struct bio *bp) 2998 { 2999 void (*done)(struct bio *); 3000 3001 mtx_lock(&bdonelock); 3002 bp->bio_flags |= BIO_DONE; 3003 done = bp->bio_done; 3004 if (done == NULL) 3005 wakeup(bp); 3006 mtx_unlock(&bdonelock); 3007 if (done != NULL) 3008 done(bp); 3009 } 3010 3011 /* 3012 * Wait for a BIO to finish. 3013 * 3014 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3015 * case is not yet clear. 3016 */ 3017 int 3018 biowait(struct bio *bp, const char *wchan) 3019 { 3020 3021 mtx_lock(&bdonelock); 3022 while ((bp->bio_flags & BIO_DONE) == 0) 3023 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3024 mtx_unlock(&bdonelock); 3025 if (bp->bio_error != 0) 3026 return (bp->bio_error); 3027 if (!(bp->bio_flags & BIO_ERROR)) 3028 return (0); 3029 return (EIO); 3030 } 3031 3032 void 3033 biofinish(struct bio *bp, struct devstat *stat, int error) 3034 { 3035 3036 if (error) { 3037 bp->bio_error = error; 3038 bp->bio_flags |= BIO_ERROR; 3039 } 3040 if (stat != NULL) 3041 devstat_end_transaction_bio(stat, bp); 3042 biodone(bp); 3043 } 3044 3045 /* 3046 * bufwait: 3047 * 3048 * Wait for buffer I/O completion, returning error status. The buffer 3049 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3050 * error and cleared. 3051 */ 3052 int 3053 bufwait(struct buf *bp) 3054 { 3055 if (bp->b_iocmd == BIO_READ) 3056 bwait(bp, PRIBIO, "biord"); 3057 else 3058 bwait(bp, PRIBIO, "biowr"); 3059 if (bp->b_flags & B_EINTR) { 3060 bp->b_flags &= ~B_EINTR; 3061 return (EINTR); 3062 } 3063 if (bp->b_ioflags & BIO_ERROR) { 3064 return (bp->b_error ? bp->b_error : EIO); 3065 } else { 3066 return (0); 3067 } 3068 } 3069 3070 /* 3071 * Call back function from struct bio back up to struct buf. 3072 */ 3073 static void 3074 bufdonebio(struct bio *bip) 3075 { 3076 struct buf *bp; 3077 3078 bp = bip->bio_caller2; 3079 bp->b_resid = bp->b_bcount - bip->bio_completed; 3080 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3081 bp->b_ioflags = bip->bio_flags; 3082 bp->b_error = bip->bio_error; 3083 if (bp->b_error) 3084 bp->b_ioflags |= BIO_ERROR; 3085 bufdone(bp); 3086 g_destroy_bio(bip); 3087 } 3088 3089 void 3090 dev_strategy(struct cdev *dev, struct buf *bp) 3091 { 3092 struct cdevsw *csw; 3093 struct bio *bip; 3094 3095 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3096 panic("b_iocmd botch"); 3097 for (;;) { 3098 bip = g_new_bio(); 3099 if (bip != NULL) 3100 break; 3101 /* Try again later */ 3102 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3103 } 3104 bip->bio_cmd = bp->b_iocmd; 3105 bip->bio_offset = bp->b_iooffset; 3106 bip->bio_length = bp->b_bcount; 3107 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3108 bip->bio_data = bp->b_data; 3109 bip->bio_done = bufdonebio; 3110 bip->bio_caller2 = bp; 3111 bip->bio_dev = dev; 3112 KASSERT(dev->si_refcount > 0, 3113 ("dev_strategy on un-referenced struct cdev *(%s)", 3114 devtoname(dev))); 3115 csw = dev_refthread(dev); 3116 if (csw == NULL) { 3117 g_destroy_bio(bip); 3118 bp->b_error = ENXIO; 3119 bp->b_ioflags = BIO_ERROR; 3120 bufdone(bp); 3121 return; 3122 } 3123 (*csw->d_strategy)(bip); 3124 dev_relthread(dev); 3125 } 3126 3127 /* 3128 * bufdone: 3129 * 3130 * Finish I/O on a buffer, optionally calling a completion function. 3131 * This is usually called from an interrupt so process blocking is 3132 * not allowed. 3133 * 3134 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3135 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3136 * assuming B_INVAL is clear. 3137 * 3138 * For the VMIO case, we set B_CACHE if the op was a read and no 3139 * read error occured, or if the op was a write. B_CACHE is never 3140 * set if the buffer is invalid or otherwise uncacheable. 3141 * 3142 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3143 * initiator to leave B_INVAL set to brelse the buffer out of existance 3144 * in the biodone routine. 3145 */ 3146 void 3147 bufdone(struct buf *bp) 3148 { 3149 struct bufobj *dropobj; 3150 void (*biodone)(struct buf *); 3151 3152 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3153 dropobj = NULL; 3154 3155 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, 3156 BUF_REFCNT(bp))); 3157 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3158 3159 runningbufwakeup(bp); 3160 if (bp->b_iocmd == BIO_WRITE) 3161 dropobj = bp->b_bufobj; 3162 /* call optional completion function if requested */ 3163 if (bp->b_iodone != NULL) { 3164 biodone = bp->b_iodone; 3165 bp->b_iodone = NULL; 3166 (*biodone) (bp); 3167 if (dropobj) 3168 bufobj_wdrop(dropobj); 3169 return; 3170 } 3171 3172 bufdone_finish(bp); 3173 3174 if (dropobj) 3175 bufobj_wdrop(dropobj); 3176 } 3177 3178 void 3179 bufdone_finish(struct buf *bp) 3180 { 3181 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, 3182 BUF_REFCNT(bp))); 3183 3184 if (!LIST_EMPTY(&bp->b_dep)) 3185 buf_complete(bp); 3186 3187 if (bp->b_flags & B_VMIO) { 3188 int i; 3189 vm_ooffset_t foff; 3190 vm_page_t m; 3191 vm_object_t obj; 3192 int iosize; 3193 struct vnode *vp = bp->b_vp; 3194 boolean_t are_queues_locked; 3195 3196 obj = bp->b_bufobj->bo_object; 3197 3198 #if defined(VFS_BIO_DEBUG) 3199 mp_fixme("usecount and vflag accessed without locks."); 3200 if (vp->v_usecount == 0) { 3201 panic("biodone: zero vnode ref count"); 3202 } 3203 3204 KASSERT(vp->v_object != NULL, 3205 ("biodone: vnode %p has no vm_object", vp)); 3206 #endif 3207 3208 foff = bp->b_offset; 3209 KASSERT(bp->b_offset != NOOFFSET, 3210 ("biodone: no buffer offset")); 3211 3212 VM_OBJECT_LOCK(obj); 3213 #if defined(VFS_BIO_DEBUG) 3214 if (obj->paging_in_progress < bp->b_npages) { 3215 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3216 obj->paging_in_progress, bp->b_npages); 3217 } 3218 #endif 3219 3220 /* 3221 * Set B_CACHE if the op was a normal read and no error 3222 * occured. B_CACHE is set for writes in the b*write() 3223 * routines. 3224 */ 3225 iosize = bp->b_bcount - bp->b_resid; 3226 if (bp->b_iocmd == BIO_READ && 3227 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3228 !(bp->b_ioflags & BIO_ERROR)) { 3229 bp->b_flags |= B_CACHE; 3230 } 3231 if (bp->b_iocmd == BIO_READ) { 3232 vm_page_lock_queues(); 3233 are_queues_locked = TRUE; 3234 } else 3235 are_queues_locked = FALSE; 3236 for (i = 0; i < bp->b_npages; i++) { 3237 int bogusflag = 0; 3238 int resid; 3239 3240 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3241 if (resid > iosize) 3242 resid = iosize; 3243 3244 /* 3245 * cleanup bogus pages, restoring the originals 3246 */ 3247 m = bp->b_pages[i]; 3248 if (m == bogus_page) { 3249 bogusflag = 1; 3250 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3251 if (m == NULL) 3252 panic("biodone: page disappeared!"); 3253 bp->b_pages[i] = m; 3254 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3255 bp->b_pages, bp->b_npages); 3256 } 3257 #if defined(VFS_BIO_DEBUG) 3258 if (OFF_TO_IDX(foff) != m->pindex) { 3259 printf( 3260 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3261 (intmax_t)foff, (uintmax_t)m->pindex); 3262 } 3263 #endif 3264 3265 /* 3266 * In the write case, the valid and clean bits are 3267 * already changed correctly ( see bdwrite() ), so we 3268 * only need to do this here in the read case. 3269 */ 3270 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3271 vfs_page_set_valid(bp, foff, i, m); 3272 } 3273 3274 /* 3275 * when debugging new filesystems or buffer I/O methods, this 3276 * is the most common error that pops up. if you see this, you 3277 * have not set the page busy flag correctly!!! 3278 */ 3279 if (m->busy == 0) { 3280 printf("biodone: page busy < 0, " 3281 "pindex: %d, foff: 0x(%x,%x), " 3282 "resid: %d, index: %d\n", 3283 (int) m->pindex, (int)(foff >> 32), 3284 (int) foff & 0xffffffff, resid, i); 3285 if (!vn_isdisk(vp, NULL)) 3286 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3287 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3288 (intmax_t) bp->b_lblkno, 3289 bp->b_flags, bp->b_npages); 3290 else 3291 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3292 (intmax_t) bp->b_lblkno, 3293 bp->b_flags, bp->b_npages); 3294 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3295 (u_long)m->valid, (u_long)m->dirty, 3296 m->wire_count); 3297 panic("biodone: page busy < 0\n"); 3298 } 3299 vm_page_io_finish(m); 3300 vm_object_pip_subtract(obj, 1); 3301 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3302 iosize -= resid; 3303 } 3304 if (are_queues_locked) 3305 vm_page_unlock_queues(); 3306 vm_object_pip_wakeupn(obj, 0); 3307 VM_OBJECT_UNLOCK(obj); 3308 } 3309 3310 /* 3311 * For asynchronous completions, release the buffer now. The brelse 3312 * will do a wakeup there if necessary - so no need to do a wakeup 3313 * here in the async case. The sync case always needs to do a wakeup. 3314 */ 3315 3316 if (bp->b_flags & B_ASYNC) { 3317 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3318 brelse(bp); 3319 else 3320 bqrelse(bp); 3321 } else 3322 bdone(bp); 3323 } 3324 3325 /* 3326 * This routine is called in lieu of iodone in the case of 3327 * incomplete I/O. This keeps the busy status for pages 3328 * consistant. 3329 */ 3330 void 3331 vfs_unbusy_pages(struct buf *bp) 3332 { 3333 int i; 3334 vm_object_t obj; 3335 vm_page_t m; 3336 3337 runningbufwakeup(bp); 3338 if (!(bp->b_flags & B_VMIO)) 3339 return; 3340 3341 obj = bp->b_bufobj->bo_object; 3342 VM_OBJECT_LOCK(obj); 3343 for (i = 0; i < bp->b_npages; i++) { 3344 m = bp->b_pages[i]; 3345 if (m == bogus_page) { 3346 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3347 if (!m) 3348 panic("vfs_unbusy_pages: page missing\n"); 3349 bp->b_pages[i] = m; 3350 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3351 bp->b_pages, bp->b_npages); 3352 } 3353 vm_object_pip_subtract(obj, 1); 3354 vm_page_io_finish(m); 3355 } 3356 vm_object_pip_wakeupn(obj, 0); 3357 VM_OBJECT_UNLOCK(obj); 3358 } 3359 3360 /* 3361 * vfs_page_set_valid: 3362 * 3363 * Set the valid bits in a page based on the supplied offset. The 3364 * range is restricted to the buffer's size. 3365 * 3366 * This routine is typically called after a read completes. 3367 */ 3368 static void 3369 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3370 { 3371 vm_ooffset_t soff, eoff; 3372 3373 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3374 /* 3375 * Start and end offsets in buffer. eoff - soff may not cross a 3376 * page boundry or cross the end of the buffer. The end of the 3377 * buffer, in this case, is our file EOF, not the allocation size 3378 * of the buffer. 3379 */ 3380 soff = off; 3381 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3382 if (eoff > bp->b_offset + bp->b_bcount) 3383 eoff = bp->b_offset + bp->b_bcount; 3384 3385 /* 3386 * Set valid range. This is typically the entire buffer and thus the 3387 * entire page. 3388 */ 3389 if (eoff > soff) { 3390 vm_page_set_validclean( 3391 m, 3392 (vm_offset_t) (soff & PAGE_MASK), 3393 (vm_offset_t) (eoff - soff) 3394 ); 3395 } 3396 } 3397 3398 /* 3399 * This routine is called before a device strategy routine. 3400 * It is used to tell the VM system that paging I/O is in 3401 * progress, and treat the pages associated with the buffer 3402 * almost as being VPO_BUSY. Also the object paging_in_progress 3403 * flag is handled to make sure that the object doesn't become 3404 * inconsistant. 3405 * 3406 * Since I/O has not been initiated yet, certain buffer flags 3407 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3408 * and should be ignored. 3409 */ 3410 void 3411 vfs_busy_pages(struct buf *bp, int clear_modify) 3412 { 3413 int i, bogus; 3414 vm_object_t obj; 3415 vm_ooffset_t foff; 3416 vm_page_t m; 3417 3418 if (!(bp->b_flags & B_VMIO)) 3419 return; 3420 3421 obj = bp->b_bufobj->bo_object; 3422 foff = bp->b_offset; 3423 KASSERT(bp->b_offset != NOOFFSET, 3424 ("vfs_busy_pages: no buffer offset")); 3425 VM_OBJECT_LOCK(obj); 3426 if (bp->b_bufsize != 0) 3427 vfs_setdirty_locked_object(bp); 3428 retry: 3429 for (i = 0; i < bp->b_npages; i++) { 3430 m = bp->b_pages[i]; 3431 3432 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3433 goto retry; 3434 } 3435 bogus = 0; 3436 vm_page_lock_queues(); 3437 for (i = 0; i < bp->b_npages; i++) { 3438 m = bp->b_pages[i]; 3439 3440 if ((bp->b_flags & B_CLUSTER) == 0) { 3441 vm_object_pip_add(obj, 1); 3442 vm_page_io_start(m); 3443 } 3444 /* 3445 * When readying a buffer for a read ( i.e 3446 * clear_modify == 0 ), it is important to do 3447 * bogus_page replacement for valid pages in 3448 * partially instantiated buffers. Partially 3449 * instantiated buffers can, in turn, occur when 3450 * reconstituting a buffer from its VM backing store 3451 * base. We only have to do this if B_CACHE is 3452 * clear ( which causes the I/O to occur in the 3453 * first place ). The replacement prevents the read 3454 * I/O from overwriting potentially dirty VM-backed 3455 * pages. XXX bogus page replacement is, uh, bogus. 3456 * It may not work properly with small-block devices. 3457 * We need to find a better way. 3458 */ 3459 pmap_remove_all(m); 3460 if (clear_modify) 3461 vfs_page_set_valid(bp, foff, i, m); 3462 else if (m->valid == VM_PAGE_BITS_ALL && 3463 (bp->b_flags & B_CACHE) == 0) { 3464 bp->b_pages[i] = bogus_page; 3465 bogus++; 3466 } 3467 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3468 } 3469 vm_page_unlock_queues(); 3470 VM_OBJECT_UNLOCK(obj); 3471 if (bogus) 3472 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3473 bp->b_pages, bp->b_npages); 3474 } 3475 3476 /* 3477 * Tell the VM system that the pages associated with this buffer 3478 * are clean. This is used for delayed writes where the data is 3479 * going to go to disk eventually without additional VM intevention. 3480 * 3481 * Note that while we only really need to clean through to b_bcount, we 3482 * just go ahead and clean through to b_bufsize. 3483 */ 3484 static void 3485 vfs_clean_pages(struct buf *bp) 3486 { 3487 int i; 3488 vm_ooffset_t foff, noff, eoff; 3489 vm_page_t m; 3490 3491 if (!(bp->b_flags & B_VMIO)) 3492 return; 3493 3494 foff = bp->b_offset; 3495 KASSERT(bp->b_offset != NOOFFSET, 3496 ("vfs_clean_pages: no buffer offset")); 3497 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3498 vm_page_lock_queues(); 3499 for (i = 0; i < bp->b_npages; i++) { 3500 m = bp->b_pages[i]; 3501 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3502 eoff = noff; 3503 3504 if (eoff > bp->b_offset + bp->b_bufsize) 3505 eoff = bp->b_offset + bp->b_bufsize; 3506 vfs_page_set_valid(bp, foff, i, m); 3507 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3508 foff = noff; 3509 } 3510 vm_page_unlock_queues(); 3511 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3512 } 3513 3514 /* 3515 * vfs_bio_set_validclean: 3516 * 3517 * Set the range within the buffer to valid and clean. The range is 3518 * relative to the beginning of the buffer, b_offset. Note that b_offset 3519 * itself may be offset from the beginning of the first page. 3520 * 3521 */ 3522 3523 void 3524 vfs_bio_set_validclean(struct buf *bp, int base, int size) 3525 { 3526 int i, n; 3527 vm_page_t m; 3528 3529 if (!(bp->b_flags & B_VMIO)) 3530 return; 3531 /* 3532 * Fixup base to be relative to beginning of first page. 3533 * Set initial n to be the maximum number of bytes in the 3534 * first page that can be validated. 3535 */ 3536 3537 base += (bp->b_offset & PAGE_MASK); 3538 n = PAGE_SIZE - (base & PAGE_MASK); 3539 3540 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3541 vm_page_lock_queues(); 3542 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3543 m = bp->b_pages[i]; 3544 if (n > size) 3545 n = size; 3546 vm_page_set_validclean(m, base & PAGE_MASK, n); 3547 base += n; 3548 size -= n; 3549 n = PAGE_SIZE; 3550 } 3551 vm_page_unlock_queues(); 3552 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3553 } 3554 3555 /* 3556 * vfs_bio_clrbuf: 3557 * 3558 * clear a buffer. This routine essentially fakes an I/O, so we need 3559 * to clear BIO_ERROR and B_INVAL. 3560 * 3561 * Note that while we only theoretically need to clear through b_bcount, 3562 * we go ahead and clear through b_bufsize. 3563 */ 3564 3565 void 3566 vfs_bio_clrbuf(struct buf *bp) 3567 { 3568 int i, j, mask = 0; 3569 caddr_t sa, ea; 3570 3571 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3572 clrbuf(bp); 3573 return; 3574 } 3575 3576 bp->b_flags &= ~B_INVAL; 3577 bp->b_ioflags &= ~BIO_ERROR; 3578 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3579 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3580 (bp->b_offset & PAGE_MASK) == 0) { 3581 if (bp->b_pages[0] == bogus_page) 3582 goto unlock; 3583 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3584 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3585 if ((bp->b_pages[0]->valid & mask) == mask) 3586 goto unlock; 3587 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3588 ((bp->b_pages[0]->valid & mask) == 0)) { 3589 bzero(bp->b_data, bp->b_bufsize); 3590 bp->b_pages[0]->valid |= mask; 3591 goto unlock; 3592 } 3593 } 3594 ea = sa = bp->b_data; 3595 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3596 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3597 ea = (caddr_t)(vm_offset_t)ulmin( 3598 (u_long)(vm_offset_t)ea, 3599 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3600 if (bp->b_pages[i] == bogus_page) 3601 continue; 3602 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3603 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3604 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3605 if ((bp->b_pages[i]->valid & mask) == mask) 3606 continue; 3607 if ((bp->b_pages[i]->valid & mask) == 0) { 3608 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3609 bzero(sa, ea - sa); 3610 } else { 3611 for (; sa < ea; sa += DEV_BSIZE, j++) { 3612 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3613 (bp->b_pages[i]->valid & (1 << j)) == 0) 3614 bzero(sa, DEV_BSIZE); 3615 } 3616 } 3617 bp->b_pages[i]->valid |= mask; 3618 } 3619 unlock: 3620 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3621 bp->b_resid = 0; 3622 } 3623 3624 /* 3625 * vm_hold_load_pages and vm_hold_free_pages get pages into 3626 * a buffers address space. The pages are anonymous and are 3627 * not associated with a file object. 3628 */ 3629 static void 3630 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3631 { 3632 vm_offset_t pg; 3633 vm_page_t p; 3634 int index; 3635 3636 to = round_page(to); 3637 from = round_page(from); 3638 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3639 3640 VM_OBJECT_LOCK(kernel_object); 3641 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3642 tryagain: 3643 /* 3644 * note: must allocate system pages since blocking here 3645 * could intefere with paging I/O, no matter which 3646 * process we are. 3647 */ 3648 p = vm_page_alloc(kernel_object, 3649 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3650 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3651 if (!p) { 3652 atomic_add_int(&vm_pageout_deficit, 3653 (to - pg) >> PAGE_SHIFT); 3654 VM_OBJECT_UNLOCK(kernel_object); 3655 VM_WAIT; 3656 VM_OBJECT_LOCK(kernel_object); 3657 goto tryagain; 3658 } 3659 p->valid = VM_PAGE_BITS_ALL; 3660 pmap_qenter(pg, &p, 1); 3661 bp->b_pages[index] = p; 3662 } 3663 VM_OBJECT_UNLOCK(kernel_object); 3664 bp->b_npages = index; 3665 } 3666 3667 /* Return pages associated with this buf to the vm system */ 3668 static void 3669 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3670 { 3671 vm_offset_t pg; 3672 vm_page_t p; 3673 int index, newnpages; 3674 3675 from = round_page(from); 3676 to = round_page(to); 3677 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3678 3679 VM_OBJECT_LOCK(kernel_object); 3680 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3681 p = bp->b_pages[index]; 3682 if (p && (index < bp->b_npages)) { 3683 if (p->busy) { 3684 printf( 3685 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3686 (intmax_t)bp->b_blkno, 3687 (intmax_t)bp->b_lblkno); 3688 } 3689 bp->b_pages[index] = NULL; 3690 pmap_qremove(pg, 1); 3691 vm_page_lock_queues(); 3692 vm_page_unwire(p, 0); 3693 vm_page_free(p); 3694 vm_page_unlock_queues(); 3695 } 3696 } 3697 VM_OBJECT_UNLOCK(kernel_object); 3698 bp->b_npages = newnpages; 3699 } 3700 3701 /* 3702 * Map an IO request into kernel virtual address space. 3703 * 3704 * All requests are (re)mapped into kernel VA space. 3705 * Notice that we use b_bufsize for the size of the buffer 3706 * to be mapped. b_bcount might be modified by the driver. 3707 * 3708 * Note that even if the caller determines that the address space should 3709 * be valid, a race or a smaller-file mapped into a larger space may 3710 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3711 * check the return value. 3712 */ 3713 int 3714 vmapbuf(struct buf *bp) 3715 { 3716 caddr_t addr, kva; 3717 vm_prot_t prot; 3718 int pidx, i; 3719 struct vm_page *m; 3720 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3721 3722 if (bp->b_bufsize < 0) 3723 return (-1); 3724 prot = VM_PROT_READ; 3725 if (bp->b_iocmd == BIO_READ) 3726 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3727 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3728 addr < bp->b_data + bp->b_bufsize; 3729 addr += PAGE_SIZE, pidx++) { 3730 /* 3731 * Do the vm_fault if needed; do the copy-on-write thing 3732 * when reading stuff off device into memory. 3733 * 3734 * NOTE! Must use pmap_extract() because addr may be in 3735 * the userland address space, and kextract is only guarenteed 3736 * to work for the kernland address space (see: sparc64 port). 3737 */ 3738 retry: 3739 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3740 prot) < 0) { 3741 vm_page_lock_queues(); 3742 for (i = 0; i < pidx; ++i) { 3743 vm_page_unhold(bp->b_pages[i]); 3744 bp->b_pages[i] = NULL; 3745 } 3746 vm_page_unlock_queues(); 3747 return(-1); 3748 } 3749 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3750 if (m == NULL) 3751 goto retry; 3752 bp->b_pages[pidx] = m; 3753 } 3754 if (pidx > btoc(MAXPHYS)) 3755 panic("vmapbuf: mapped more than MAXPHYS"); 3756 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3757 3758 kva = bp->b_saveaddr; 3759 bp->b_npages = pidx; 3760 bp->b_saveaddr = bp->b_data; 3761 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3762 return(0); 3763 } 3764 3765 /* 3766 * Free the io map PTEs associated with this IO operation. 3767 * We also invalidate the TLB entries and restore the original b_addr. 3768 */ 3769 void 3770 vunmapbuf(struct buf *bp) 3771 { 3772 int pidx; 3773 int npages; 3774 3775 npages = bp->b_npages; 3776 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3777 vm_page_lock_queues(); 3778 for (pidx = 0; pidx < npages; pidx++) 3779 vm_page_unhold(bp->b_pages[pidx]); 3780 vm_page_unlock_queues(); 3781 3782 bp->b_data = bp->b_saveaddr; 3783 } 3784 3785 void 3786 bdone(struct buf *bp) 3787 { 3788 3789 mtx_lock(&bdonelock); 3790 bp->b_flags |= B_DONE; 3791 wakeup(bp); 3792 mtx_unlock(&bdonelock); 3793 } 3794 3795 void 3796 bwait(struct buf *bp, u_char pri, const char *wchan) 3797 { 3798 3799 mtx_lock(&bdonelock); 3800 while ((bp->b_flags & B_DONE) == 0) 3801 msleep(bp, &bdonelock, pri, wchan, 0); 3802 mtx_unlock(&bdonelock); 3803 } 3804 3805 int 3806 bufsync(struct bufobj *bo, int waitfor, struct thread *td) 3807 { 3808 3809 return (VOP_FSYNC(bo->__bo_vnode, waitfor, td)); 3810 } 3811 3812 void 3813 bufstrategy(struct bufobj *bo, struct buf *bp) 3814 { 3815 int i = 0; 3816 struct vnode *vp; 3817 3818 vp = bp->b_vp; 3819 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3820 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3821 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3822 i = VOP_STRATEGY(vp, bp); 3823 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3824 } 3825 3826 void 3827 bufobj_wrefl(struct bufobj *bo) 3828 { 3829 3830 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3831 ASSERT_BO_LOCKED(bo); 3832 bo->bo_numoutput++; 3833 } 3834 3835 void 3836 bufobj_wref(struct bufobj *bo) 3837 { 3838 3839 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3840 BO_LOCK(bo); 3841 bo->bo_numoutput++; 3842 BO_UNLOCK(bo); 3843 } 3844 3845 void 3846 bufobj_wdrop(struct bufobj *bo) 3847 { 3848 3849 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3850 BO_LOCK(bo); 3851 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3852 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3853 bo->bo_flag &= ~BO_WWAIT; 3854 wakeup(&bo->bo_numoutput); 3855 } 3856 BO_UNLOCK(bo); 3857 } 3858 3859 int 3860 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3861 { 3862 int error; 3863 3864 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3865 ASSERT_BO_LOCKED(bo); 3866 error = 0; 3867 while (bo->bo_numoutput) { 3868 bo->bo_flag |= BO_WWAIT; 3869 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3870 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3871 if (error) 3872 break; 3873 } 3874 return (error); 3875 } 3876 3877 void 3878 bpin(struct buf *bp) 3879 { 3880 mtx_lock(&bpinlock); 3881 bp->b_pin_count++; 3882 mtx_unlock(&bpinlock); 3883 } 3884 3885 void 3886 bunpin(struct buf *bp) 3887 { 3888 mtx_lock(&bpinlock); 3889 if (--bp->b_pin_count == 0) 3890 wakeup(bp); 3891 mtx_unlock(&bpinlock); 3892 } 3893 3894 void 3895 bunpin_wait(struct buf *bp) 3896 { 3897 mtx_lock(&bpinlock); 3898 while (bp->b_pin_count > 0) 3899 msleep(bp, &bpinlock, PRIBIO, "bwunpin", 0); 3900 mtx_unlock(&bpinlock); 3901 } 3902 3903 #include "opt_ddb.h" 3904 #ifdef DDB 3905 #include <ddb/ddb.h> 3906 3907 /* DDB command to show buffer data */ 3908 DB_SHOW_COMMAND(buffer, db_show_buffer) 3909 { 3910 /* get args */ 3911 struct buf *bp = (struct buf *)addr; 3912 3913 if (!have_addr) { 3914 db_printf("usage: show buffer <addr>\n"); 3915 return; 3916 } 3917 3918 db_printf("buf at %p\n", bp); 3919 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3920 db_printf( 3921 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3922 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", 3923 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3924 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); 3925 if (bp->b_npages) { 3926 int i; 3927 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3928 for (i = 0; i < bp->b_npages; i++) { 3929 vm_page_t m; 3930 m = bp->b_pages[i]; 3931 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3932 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3933 if ((i + 1) < bp->b_npages) 3934 db_printf(","); 3935 } 3936 db_printf("\n"); 3937 } 3938 lockmgr_printinfo(&bp->b_lock); 3939 } 3940 3941 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 3942 { 3943 struct buf *bp; 3944 int i; 3945 3946 for (i = 0; i < nbuf; i++) { 3947 bp = &buf[i]; 3948 if (lockcount(&bp->b_lock)) { 3949 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 3950 db_printf("\n"); 3951 } 3952 } 3953 } 3954 #endif /* DDB */ 3955