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/limits.h> 52 #include <sys/lock.h> 53 #include <sys/malloc.h> 54 #include <sys/mount.h> 55 #include <sys/mutex.h> 56 #include <sys/kernel.h> 57 #include <sys/kthread.h> 58 #include <sys/proc.h> 59 #include <sys/resourcevar.h> 60 #include <sys/sysctl.h> 61 #include <sys/vmmeter.h> 62 #include <sys/vnode.h> 63 #include <geom/geom.h> 64 #include <vm/vm.h> 65 #include <vm/vm_param.h> 66 #include <vm/vm_kern.h> 67 #include <vm/vm_pageout.h> 68 #include <vm/vm_page.h> 69 #include <vm/vm_object.h> 70 #include <vm/vm_extern.h> 71 #include <vm/vm_map.h> 72 #include "opt_directio.h" 73 #include "opt_swap.h" 74 75 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 76 77 struct bio_ops bioops; /* I/O operation notification */ 78 79 struct buf_ops buf_ops_bio = { 80 .bop_name = "buf_ops_bio", 81 .bop_write = bufwrite, 82 .bop_strategy = bufstrategy, 83 .bop_sync = bufsync, 84 .bop_bdflush = bufbdflush, 85 }; 86 87 /* 88 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 89 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 90 */ 91 struct buf *buf; /* buffer header pool */ 92 93 static struct proc *bufdaemonproc; 94 95 static int inmem(struct vnode *vp, daddr_t blkno); 96 static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 97 vm_offset_t to); 98 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 99 vm_offset_t to); 100 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 101 vm_page_t m); 102 static void vfs_clean_pages(struct buf *bp); 103 static void vfs_setdirty(struct buf *bp); 104 static void vfs_setdirty_locked_object(struct buf *bp); 105 static void vfs_vmio_release(struct buf *bp); 106 static int vfs_bio_clcheck(struct vnode *vp, int size, 107 daddr_t lblkno, daddr_t blkno); 108 static int flushbufqueues(int, int); 109 static void buf_daemon(void); 110 static void bremfreel(struct buf *bp); 111 112 int vmiodirenable = TRUE; 113 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 114 "Use the VM system for directory writes"); 115 int runningbufspace; 116 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 117 "Amount of presently outstanding async buffer io"); 118 static int bufspace; 119 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 120 "KVA memory used for bufs"); 121 static int maxbufspace; 122 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 123 "Maximum allowed value of bufspace (including buf_daemon)"); 124 static int bufmallocspace; 125 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 126 "Amount of malloced memory for buffers"); 127 static int maxbufmallocspace; 128 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 129 "Maximum amount of malloced memory for buffers"); 130 static int lobufspace; 131 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 132 "Minimum amount of buffers we want to have"); 133 int hibufspace; 134 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 135 "Maximum allowed value of bufspace (excluding buf_daemon)"); 136 static int bufreusecnt; 137 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 138 "Number of times we have reused a buffer"); 139 static int buffreekvacnt; 140 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 141 "Number of times we have freed the KVA space from some buffer"); 142 static int bufdefragcnt; 143 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 144 "Number of times we have had to repeat buffer allocation to defragment"); 145 static int lorunningspace; 146 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 147 "Minimum preferred space used for in-progress I/O"); 148 static int hirunningspace; 149 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 150 "Maximum amount of space to use for in-progress I/O"); 151 int dirtybufferflushes; 152 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 153 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 154 int bdwriteskip; 155 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 156 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 157 int altbufferflushes; 158 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 159 0, "Number of fsync flushes to limit dirty buffers"); 160 static int recursiveflushes; 161 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 162 0, "Number of flushes skipped due to being recursive"); 163 static int numdirtybuffers; 164 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 165 "Number of buffers that are dirty (has unwritten changes) at the moment"); 166 static int lodirtybuffers; 167 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 168 "How many buffers we want to have free before bufdaemon can sleep"); 169 static int hidirtybuffers; 170 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 171 "When the number of dirty buffers is considered severe"); 172 int dirtybufthresh; 173 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 174 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 175 static int numfreebuffers; 176 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 177 "Number of free buffers"); 178 static int lofreebuffers; 179 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 180 "XXX Unused"); 181 static int hifreebuffers; 182 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 183 "XXX Complicatedly unused"); 184 static int getnewbufcalls; 185 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 186 "Number of calls to getnewbuf"); 187 static int getnewbufrestarts; 188 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 189 "Number of times getnewbuf has had to restart a buffer aquisition"); 190 191 /* 192 * Wakeup point for bufdaemon, as well as indicator of whether it is already 193 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 194 * is idling. 195 */ 196 static int bd_request; 197 198 /* 199 * This lock synchronizes access to bd_request. 200 */ 201 static struct mtx bdlock; 202 203 /* 204 * bogus page -- for I/O to/from partially complete buffers 205 * this is a temporary solution to the problem, but it is not 206 * really that bad. it would be better to split the buffer 207 * for input in the case of buffers partially already in memory, 208 * but the code is intricate enough already. 209 */ 210 vm_page_t bogus_page; 211 212 /* 213 * Synchronization (sleep/wakeup) variable for active buffer space requests. 214 * Set when wait starts, cleared prior to wakeup(). 215 * Used in runningbufwakeup() and waitrunningbufspace(). 216 */ 217 static int runningbufreq; 218 219 /* 220 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 221 * waitrunningbufspace(). 222 */ 223 static struct mtx rbreqlock; 224 225 /* 226 * Synchronization (sleep/wakeup) variable for buffer requests. 227 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 228 * by and/or. 229 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 230 * getnewbuf(), and getblk(). 231 */ 232 static int needsbuffer; 233 234 /* 235 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 236 */ 237 static struct mtx nblock; 238 239 /* 240 * Definitions for the buffer free lists. 241 */ 242 #define BUFFER_QUEUES 6 /* number of free buffer queues */ 243 244 #define QUEUE_NONE 0 /* on no queue */ 245 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 246 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 247 #define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ 248 #define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 249 #define QUEUE_EMPTY 5 /* empty buffer headers */ 250 251 /* Queues for free buffers with various properties */ 252 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 253 254 /* Lock for the bufqueues */ 255 static struct mtx bqlock; 256 257 /* 258 * Single global constant for BUF_WMESG, to avoid getting multiple references. 259 * buf_wmesg is referred from macros. 260 */ 261 const char *buf_wmesg = BUF_WMESG; 262 263 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 264 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 265 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 266 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 267 268 #ifdef DIRECTIO 269 extern void ffs_rawread_setup(void); 270 #endif /* DIRECTIO */ 271 /* 272 * numdirtywakeup: 273 * 274 * If someone is blocked due to there being too many dirty buffers, 275 * and numdirtybuffers is now reasonable, wake them up. 276 */ 277 278 static __inline void 279 numdirtywakeup(int level) 280 { 281 282 if (numdirtybuffers <= level) { 283 mtx_lock(&nblock); 284 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 285 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 286 wakeup(&needsbuffer); 287 } 288 mtx_unlock(&nblock); 289 } 290 } 291 292 /* 293 * bufspacewakeup: 294 * 295 * Called when buffer space is potentially available for recovery. 296 * getnewbuf() will block on this flag when it is unable to free 297 * sufficient buffer space. Buffer space becomes recoverable when 298 * bp's get placed back in the queues. 299 */ 300 301 static __inline void 302 bufspacewakeup(void) 303 { 304 305 /* 306 * If someone is waiting for BUF space, wake them up. Even 307 * though we haven't freed the kva space yet, the waiting 308 * process will be able to now. 309 */ 310 mtx_lock(&nblock); 311 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 312 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 313 wakeup(&needsbuffer); 314 } 315 mtx_unlock(&nblock); 316 } 317 318 /* 319 * runningbufwakeup() - in-progress I/O accounting. 320 * 321 */ 322 void 323 runningbufwakeup(struct buf *bp) 324 { 325 326 if (bp->b_runningbufspace) { 327 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 328 bp->b_runningbufspace = 0; 329 mtx_lock(&rbreqlock); 330 if (runningbufreq && runningbufspace <= lorunningspace) { 331 runningbufreq = 0; 332 wakeup(&runningbufreq); 333 } 334 mtx_unlock(&rbreqlock); 335 } 336 } 337 338 /* 339 * bufcountwakeup: 340 * 341 * Called when a buffer has been added to one of the free queues to 342 * account for the buffer and to wakeup anyone waiting for free buffers. 343 * This typically occurs when large amounts of metadata are being handled 344 * by the buffer cache ( else buffer space runs out first, usually ). 345 */ 346 347 static __inline void 348 bufcountwakeup(void) 349 { 350 351 atomic_add_int(&numfreebuffers, 1); 352 mtx_lock(&nblock); 353 if (needsbuffer) { 354 needsbuffer &= ~VFS_BIO_NEED_ANY; 355 if (numfreebuffers >= hifreebuffers) 356 needsbuffer &= ~VFS_BIO_NEED_FREE; 357 wakeup(&needsbuffer); 358 } 359 mtx_unlock(&nblock); 360 } 361 362 /* 363 * waitrunningbufspace() 364 * 365 * runningbufspace is a measure of the amount of I/O currently 366 * running. This routine is used in async-write situations to 367 * prevent creating huge backups of pending writes to a device. 368 * Only asynchronous writes are governed by this function. 369 * 370 * Reads will adjust runningbufspace, but will not block based on it. 371 * The read load has a side effect of reducing the allowed write load. 372 * 373 * This does NOT turn an async write into a sync write. It waits 374 * for earlier writes to complete and generally returns before the 375 * caller's write has reached the device. 376 */ 377 void 378 waitrunningbufspace(void) 379 { 380 381 mtx_lock(&rbreqlock); 382 while (runningbufspace > hirunningspace) { 383 ++runningbufreq; 384 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 385 } 386 mtx_unlock(&rbreqlock); 387 } 388 389 390 /* 391 * vfs_buf_test_cache: 392 * 393 * Called when a buffer is extended. This function clears the B_CACHE 394 * bit if the newly extended portion of the buffer does not contain 395 * valid data. 396 */ 397 static __inline 398 void 399 vfs_buf_test_cache(struct buf *bp, 400 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 401 vm_page_t m) 402 { 403 404 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 405 if (bp->b_flags & B_CACHE) { 406 int base = (foff + off) & PAGE_MASK; 407 if (vm_page_is_valid(m, base, size) == 0) 408 bp->b_flags &= ~B_CACHE; 409 } 410 } 411 412 /* Wake up the buffer daemon if necessary */ 413 static __inline 414 void 415 bd_wakeup(int dirtybuflevel) 416 { 417 418 mtx_lock(&bdlock); 419 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 420 bd_request = 1; 421 wakeup(&bd_request); 422 } 423 mtx_unlock(&bdlock); 424 } 425 426 /* 427 * bd_speedup - speedup the buffer cache flushing code 428 */ 429 430 static __inline 431 void 432 bd_speedup(void) 433 { 434 435 bd_wakeup(1); 436 } 437 438 /* 439 * Calculating buffer cache scaling values and reserve space for buffer 440 * headers. This is called during low level kernel initialization and 441 * may be called more then once. We CANNOT write to the memory area 442 * being reserved at this time. 443 */ 444 caddr_t 445 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 446 { 447 int maxbuf; 448 449 /* 450 * physmem_est is in pages. Convert it to kilobytes (assumes 451 * PAGE_SIZE is >= 1K) 452 */ 453 physmem_est = physmem_est * (PAGE_SIZE / 1024); 454 455 /* 456 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 457 * For the first 64MB of ram nominally allocate sufficient buffers to 458 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 459 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 460 * the buffer cache we limit the eventual kva reservation to 461 * maxbcache bytes. 462 * 463 * factor represents the 1/4 x ram conversion. 464 */ 465 if (nbuf == 0) { 466 int factor = 4 * BKVASIZE / 1024; 467 468 nbuf = 50; 469 if (physmem_est > 4096) 470 nbuf += min((physmem_est - 4096) / factor, 471 65536 / factor); 472 if (physmem_est > 65536) 473 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 474 475 if (maxbcache && nbuf > maxbcache / BKVASIZE) 476 nbuf = maxbcache / BKVASIZE; 477 478 /* XXX Avoid integer overflows later on with maxbufspace. */ 479 maxbuf = (INT_MAX / 3) / BKVASIZE; 480 if (nbuf > maxbuf) 481 nbuf = maxbuf; 482 } 483 484 /* 485 * swbufs are used as temporary holders for I/O, such as paging I/O. 486 * We have no less then 16 and no more then 256. 487 */ 488 nswbuf = max(min(nbuf/4, 256), 16); 489 #ifdef NSWBUF_MIN 490 if (nswbuf < NSWBUF_MIN) 491 nswbuf = NSWBUF_MIN; 492 #endif 493 #ifdef DIRECTIO 494 ffs_rawread_setup(); 495 #endif 496 497 /* 498 * Reserve space for the buffer cache buffers 499 */ 500 swbuf = (void *)v; 501 v = (caddr_t)(swbuf + nswbuf); 502 buf = (void *)v; 503 v = (caddr_t)(buf + nbuf); 504 505 return(v); 506 } 507 508 /* Initialize the buffer subsystem. Called before use of any buffers. */ 509 void 510 bufinit(void) 511 { 512 struct buf *bp; 513 int i; 514 515 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 516 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 517 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 518 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 519 520 /* next, make a null set of free lists */ 521 for (i = 0; i < BUFFER_QUEUES; i++) 522 TAILQ_INIT(&bufqueues[i]); 523 524 /* finally, initialize each buffer header and stick on empty q */ 525 for (i = 0; i < nbuf; i++) { 526 bp = &buf[i]; 527 bzero(bp, sizeof *bp); 528 bp->b_flags = B_INVAL; /* we're just an empty header */ 529 bp->b_rcred = NOCRED; 530 bp->b_wcred = NOCRED; 531 bp->b_qindex = QUEUE_EMPTY; 532 bp->b_vflags = 0; 533 bp->b_xflags = 0; 534 LIST_INIT(&bp->b_dep); 535 BUF_LOCKINIT(bp); 536 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 537 } 538 539 /* 540 * maxbufspace is the absolute maximum amount of buffer space we are 541 * allowed to reserve in KVM and in real terms. The absolute maximum 542 * is nominally used by buf_daemon. hibufspace is the nominal maximum 543 * used by most other processes. The differential is required to 544 * ensure that buf_daemon is able to run when other processes might 545 * be blocked waiting for buffer space. 546 * 547 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 548 * this may result in KVM fragmentation which is not handled optimally 549 * by the system. 550 */ 551 maxbufspace = nbuf * BKVASIZE; 552 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 553 lobufspace = hibufspace - MAXBSIZE; 554 555 lorunningspace = 512 * 1024; 556 hirunningspace = 1024 * 1024; 557 558 /* 559 * Limit the amount of malloc memory since it is wired permanently into 560 * the kernel space. Even though this is accounted for in the buffer 561 * allocation, we don't want the malloced region to grow uncontrolled. 562 * The malloc scheme improves memory utilization significantly on average 563 * (small) directories. 564 */ 565 maxbufmallocspace = hibufspace / 20; 566 567 /* 568 * Reduce the chance of a deadlock occuring by limiting the number 569 * of delayed-write dirty buffers we allow to stack up. 570 */ 571 hidirtybuffers = nbuf / 4 + 20; 572 dirtybufthresh = hidirtybuffers * 9 / 10; 573 numdirtybuffers = 0; 574 /* 575 * To support extreme low-memory systems, make sure hidirtybuffers cannot 576 * eat up all available buffer space. This occurs when our minimum cannot 577 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 578 * BKVASIZE'd (8K) buffers. 579 */ 580 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 581 hidirtybuffers >>= 1; 582 } 583 lodirtybuffers = hidirtybuffers / 2; 584 585 /* 586 * Try to keep the number of free buffers in the specified range, 587 * and give special processes (e.g. like buf_daemon) access to an 588 * emergency reserve. 589 */ 590 lofreebuffers = nbuf / 18 + 5; 591 hifreebuffers = 2 * lofreebuffers; 592 numfreebuffers = nbuf; 593 594 /* 595 * Maximum number of async ops initiated per buf_daemon loop. This is 596 * somewhat of a hack at the moment, we really need to limit ourselves 597 * based on the number of bytes of I/O in-transit that were initiated 598 * from buf_daemon. 599 */ 600 601 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 602 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 603 } 604 605 /* 606 * bfreekva() - free the kva allocation for a buffer. 607 * 608 * Since this call frees up buffer space, we call bufspacewakeup(). 609 */ 610 static void 611 bfreekva(struct buf *bp) 612 { 613 614 if (bp->b_kvasize) { 615 atomic_add_int(&buffreekvacnt, 1); 616 atomic_subtract_int(&bufspace, bp->b_kvasize); 617 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 618 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 619 bp->b_kvasize = 0; 620 bufspacewakeup(); 621 } 622 } 623 624 /* 625 * bremfree: 626 * 627 * Mark the buffer for removal from the appropriate free list in brelse. 628 * 629 */ 630 void 631 bremfree(struct buf *bp) 632 { 633 634 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 635 KASSERT((bp->b_flags & B_REMFREE) == 0, 636 ("bremfree: buffer %p already marked for delayed removal.", bp)); 637 KASSERT(bp->b_qindex != QUEUE_NONE, 638 ("bremfree: buffer %p not on a queue.", bp)); 639 BUF_ASSERT_HELD(bp); 640 641 bp->b_flags |= B_REMFREE; 642 /* Fixup numfreebuffers count. */ 643 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 644 atomic_subtract_int(&numfreebuffers, 1); 645 } 646 647 /* 648 * bremfreef: 649 * 650 * Force an immediate removal from a free list. Used only in nfs when 651 * it abuses the b_freelist pointer. 652 */ 653 void 654 bremfreef(struct buf *bp) 655 { 656 mtx_lock(&bqlock); 657 bremfreel(bp); 658 mtx_unlock(&bqlock); 659 } 660 661 /* 662 * bremfreel: 663 * 664 * Removes a buffer from the free list, must be called with the 665 * bqlock held. 666 */ 667 static void 668 bremfreel(struct buf *bp) 669 { 670 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 671 bp, bp->b_vp, bp->b_flags); 672 KASSERT(bp->b_qindex != QUEUE_NONE, 673 ("bremfreel: buffer %p not on a queue.", bp)); 674 BUF_ASSERT_HELD(bp); 675 mtx_assert(&bqlock, MA_OWNED); 676 677 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 678 bp->b_qindex = QUEUE_NONE; 679 /* 680 * If this was a delayed bremfree() we only need to remove the buffer 681 * from the queue and return the stats are already done. 682 */ 683 if (bp->b_flags & B_REMFREE) { 684 bp->b_flags &= ~B_REMFREE; 685 return; 686 } 687 /* 688 * Fixup numfreebuffers count. If the buffer is invalid or not 689 * delayed-write, the buffer was free and we must decrement 690 * numfreebuffers. 691 */ 692 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 693 atomic_subtract_int(&numfreebuffers, 1); 694 } 695 696 697 /* 698 * Get a buffer with the specified data. Look in the cache first. We 699 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 700 * is set, the buffer is valid and we do not have to do anything ( see 701 * getblk() ). This is really just a special case of breadn(). 702 */ 703 int 704 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 705 struct buf **bpp) 706 { 707 708 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 709 } 710 711 /* 712 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 713 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 714 * the buffer is valid and we do not have to do anything. 715 */ 716 void 717 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 718 int cnt, struct ucred * cred) 719 { 720 struct buf *rabp; 721 int i; 722 723 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 724 if (inmem(vp, *rablkno)) 725 continue; 726 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 727 728 if ((rabp->b_flags & B_CACHE) == 0) { 729 if (!TD_IS_IDLETHREAD(curthread)) 730 curthread->td_ru.ru_inblock++; 731 rabp->b_flags |= B_ASYNC; 732 rabp->b_flags &= ~B_INVAL; 733 rabp->b_ioflags &= ~BIO_ERROR; 734 rabp->b_iocmd = BIO_READ; 735 if (rabp->b_rcred == NOCRED && cred != NOCRED) 736 rabp->b_rcred = crhold(cred); 737 vfs_busy_pages(rabp, 0); 738 BUF_KERNPROC(rabp); 739 rabp->b_iooffset = dbtob(rabp->b_blkno); 740 bstrategy(rabp); 741 } else { 742 brelse(rabp); 743 } 744 } 745 } 746 747 /* 748 * Operates like bread, but also starts asynchronous I/O on 749 * read-ahead blocks. 750 */ 751 int 752 breadn(struct vnode * vp, daddr_t blkno, int size, 753 daddr_t * rablkno, int *rabsize, 754 int cnt, struct ucred * cred, struct buf **bpp) 755 { 756 struct buf *bp; 757 int rv = 0, readwait = 0; 758 759 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 760 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 761 762 /* if not found in cache, do some I/O */ 763 if ((bp->b_flags & B_CACHE) == 0) { 764 if (!TD_IS_IDLETHREAD(curthread)) 765 curthread->td_ru.ru_inblock++; 766 bp->b_iocmd = BIO_READ; 767 bp->b_flags &= ~B_INVAL; 768 bp->b_ioflags &= ~BIO_ERROR; 769 if (bp->b_rcred == NOCRED && cred != NOCRED) 770 bp->b_rcred = crhold(cred); 771 vfs_busy_pages(bp, 0); 772 bp->b_iooffset = dbtob(bp->b_blkno); 773 bstrategy(bp); 774 ++readwait; 775 } 776 777 breada(vp, rablkno, rabsize, cnt, cred); 778 779 if (readwait) { 780 rv = bufwait(bp); 781 } 782 return (rv); 783 } 784 785 /* 786 * Write, release buffer on completion. (Done by iodone 787 * if async). Do not bother writing anything if the buffer 788 * is invalid. 789 * 790 * Note that we set B_CACHE here, indicating that buffer is 791 * fully valid and thus cacheable. This is true even of NFS 792 * now so we set it generally. This could be set either here 793 * or in biodone() since the I/O is synchronous. We put it 794 * here. 795 */ 796 int 797 bufwrite(struct buf *bp) 798 { 799 int oldflags; 800 struct vnode *vp; 801 int vp_md; 802 803 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 804 if (bp->b_flags & B_INVAL) { 805 brelse(bp); 806 return (0); 807 } 808 809 oldflags = bp->b_flags; 810 811 BUF_ASSERT_HELD(bp); 812 813 if (bp->b_pin_count > 0) 814 bunpin_wait(bp); 815 816 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 817 ("FFS background buffer should not get here %p", bp)); 818 819 vp = bp->b_vp; 820 if (vp) 821 vp_md = vp->v_vflag & VV_MD; 822 else 823 vp_md = 0; 824 825 /* Mark the buffer clean */ 826 bundirty(bp); 827 828 bp->b_flags &= ~B_DONE; 829 bp->b_ioflags &= ~BIO_ERROR; 830 bp->b_flags |= B_CACHE; 831 bp->b_iocmd = BIO_WRITE; 832 833 bufobj_wref(bp->b_bufobj); 834 vfs_busy_pages(bp, 1); 835 836 /* 837 * Normal bwrites pipeline writes 838 */ 839 bp->b_runningbufspace = bp->b_bufsize; 840 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 841 842 if (!TD_IS_IDLETHREAD(curthread)) 843 curthread->td_ru.ru_oublock++; 844 if (oldflags & B_ASYNC) 845 BUF_KERNPROC(bp); 846 bp->b_iooffset = dbtob(bp->b_blkno); 847 bstrategy(bp); 848 849 if ((oldflags & B_ASYNC) == 0) { 850 int rtval = bufwait(bp); 851 brelse(bp); 852 return (rtval); 853 } else { 854 /* 855 * don't allow the async write to saturate the I/O 856 * system. We will not deadlock here because 857 * we are blocking waiting for I/O that is already in-progress 858 * to complete. We do not block here if it is the update 859 * or syncer daemon trying to clean up as that can lead 860 * to deadlock. 861 */ 862 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 863 waitrunningbufspace(); 864 } 865 866 return (0); 867 } 868 869 void 870 bufbdflush(struct bufobj *bo, struct buf *bp) 871 { 872 struct buf *nbp; 873 874 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 875 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 876 altbufferflushes++; 877 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 878 BO_LOCK(bo); 879 /* 880 * Try to find a buffer to flush. 881 */ 882 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 883 if ((nbp->b_vflags & BV_BKGRDINPROG) || 884 BUF_LOCK(nbp, 885 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 886 continue; 887 if (bp == nbp) 888 panic("bdwrite: found ourselves"); 889 BO_UNLOCK(bo); 890 /* Don't countdeps with the bo lock held. */ 891 if (buf_countdeps(nbp, 0)) { 892 BO_LOCK(bo); 893 BUF_UNLOCK(nbp); 894 continue; 895 } 896 if (nbp->b_flags & B_CLUSTEROK) { 897 vfs_bio_awrite(nbp); 898 } else { 899 bremfree(nbp); 900 bawrite(nbp); 901 } 902 dirtybufferflushes++; 903 break; 904 } 905 if (nbp == NULL) 906 BO_UNLOCK(bo); 907 } 908 } 909 910 /* 911 * Delayed write. (Buffer is marked dirty). Do not bother writing 912 * anything if the buffer is marked invalid. 913 * 914 * Note that since the buffer must be completely valid, we can safely 915 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 916 * biodone() in order to prevent getblk from writing the buffer 917 * out synchronously. 918 */ 919 void 920 bdwrite(struct buf *bp) 921 { 922 struct thread *td = curthread; 923 struct vnode *vp; 924 struct bufobj *bo; 925 926 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 927 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 928 BUF_ASSERT_HELD(bp); 929 930 if (bp->b_flags & B_INVAL) { 931 brelse(bp); 932 return; 933 } 934 935 /* 936 * If we have too many dirty buffers, don't create any more. 937 * If we are wildly over our limit, then force a complete 938 * cleanup. Otherwise, just keep the situation from getting 939 * out of control. Note that we have to avoid a recursive 940 * disaster and not try to clean up after our own cleanup! 941 */ 942 vp = bp->b_vp; 943 bo = bp->b_bufobj; 944 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 945 td->td_pflags |= TDP_INBDFLUSH; 946 BO_BDFLUSH(bo, bp); 947 td->td_pflags &= ~TDP_INBDFLUSH; 948 } else 949 recursiveflushes++; 950 951 bdirty(bp); 952 /* 953 * Set B_CACHE, indicating that the buffer is fully valid. This is 954 * true even of NFS now. 955 */ 956 bp->b_flags |= B_CACHE; 957 958 /* 959 * This bmap keeps the system from needing to do the bmap later, 960 * perhaps when the system is attempting to do a sync. Since it 961 * is likely that the indirect block -- or whatever other datastructure 962 * that the filesystem needs is still in memory now, it is a good 963 * thing to do this. Note also, that if the pageout daemon is 964 * requesting a sync -- there might not be enough memory to do 965 * the bmap then... So, this is important to do. 966 */ 967 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 968 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 969 } 970 971 /* 972 * Set the *dirty* buffer range based upon the VM system dirty pages. 973 */ 974 vfs_setdirty(bp); 975 976 /* 977 * We need to do this here to satisfy the vnode_pager and the 978 * pageout daemon, so that it thinks that the pages have been 979 * "cleaned". Note that since the pages are in a delayed write 980 * buffer -- the VFS layer "will" see that the pages get written 981 * out on the next sync, or perhaps the cluster will be completed. 982 */ 983 vfs_clean_pages(bp); 984 bqrelse(bp); 985 986 /* 987 * Wakeup the buffer flushing daemon if we have a lot of dirty 988 * buffers (midpoint between our recovery point and our stall 989 * point). 990 */ 991 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 992 993 /* 994 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 995 * due to the softdep code. 996 */ 997 } 998 999 /* 1000 * bdirty: 1001 * 1002 * Turn buffer into delayed write request. We must clear BIO_READ and 1003 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1004 * itself to properly update it in the dirty/clean lists. We mark it 1005 * B_DONE to ensure that any asynchronization of the buffer properly 1006 * clears B_DONE ( else a panic will occur later ). 1007 * 1008 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1009 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1010 * should only be called if the buffer is known-good. 1011 * 1012 * Since the buffer is not on a queue, we do not update the numfreebuffers 1013 * count. 1014 * 1015 * The buffer must be on QUEUE_NONE. 1016 */ 1017 void 1018 bdirty(struct buf *bp) 1019 { 1020 1021 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1022 bp, bp->b_vp, bp->b_flags); 1023 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1024 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1025 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1026 BUF_ASSERT_HELD(bp); 1027 bp->b_flags &= ~(B_RELBUF); 1028 bp->b_iocmd = BIO_WRITE; 1029 1030 if ((bp->b_flags & B_DELWRI) == 0) { 1031 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1032 reassignbuf(bp); 1033 atomic_add_int(&numdirtybuffers, 1); 1034 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1035 } 1036 } 1037 1038 /* 1039 * bundirty: 1040 * 1041 * Clear B_DELWRI for buffer. 1042 * 1043 * Since the buffer is not on a queue, we do not update the numfreebuffers 1044 * count. 1045 * 1046 * The buffer must be on QUEUE_NONE. 1047 */ 1048 1049 void 1050 bundirty(struct buf *bp) 1051 { 1052 1053 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1054 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1055 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1056 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1057 BUF_ASSERT_HELD(bp); 1058 1059 if (bp->b_flags & B_DELWRI) { 1060 bp->b_flags &= ~B_DELWRI; 1061 reassignbuf(bp); 1062 atomic_subtract_int(&numdirtybuffers, 1); 1063 numdirtywakeup(lodirtybuffers); 1064 } 1065 /* 1066 * Since it is now being written, we can clear its deferred write flag. 1067 */ 1068 bp->b_flags &= ~B_DEFERRED; 1069 } 1070 1071 /* 1072 * bawrite: 1073 * 1074 * Asynchronous write. Start output on a buffer, but do not wait for 1075 * it to complete. The buffer is released when the output completes. 1076 * 1077 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1078 * B_INVAL buffers. Not us. 1079 */ 1080 void 1081 bawrite(struct buf *bp) 1082 { 1083 1084 bp->b_flags |= B_ASYNC; 1085 (void) bwrite(bp); 1086 } 1087 1088 /* 1089 * bwillwrite: 1090 * 1091 * Called prior to the locking of any vnodes when we are expecting to 1092 * write. We do not want to starve the buffer cache with too many 1093 * dirty buffers so we block here. By blocking prior to the locking 1094 * of any vnodes we attempt to avoid the situation where a locked vnode 1095 * prevents the various system daemons from flushing related buffers. 1096 */ 1097 1098 void 1099 bwillwrite(void) 1100 { 1101 1102 if (numdirtybuffers >= hidirtybuffers) { 1103 mtx_lock(&nblock); 1104 while (numdirtybuffers >= hidirtybuffers) { 1105 bd_wakeup(1); 1106 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1107 msleep(&needsbuffer, &nblock, 1108 (PRIBIO + 4), "flswai", 0); 1109 } 1110 mtx_unlock(&nblock); 1111 } 1112 } 1113 1114 /* 1115 * Return true if we have too many dirty buffers. 1116 */ 1117 int 1118 buf_dirty_count_severe(void) 1119 { 1120 1121 return(numdirtybuffers >= hidirtybuffers); 1122 } 1123 1124 /* 1125 * brelse: 1126 * 1127 * Release a busy buffer and, if requested, free its resources. The 1128 * buffer will be stashed in the appropriate bufqueue[] allowing it 1129 * to be accessed later as a cache entity or reused for other purposes. 1130 */ 1131 void 1132 brelse(struct buf *bp) 1133 { 1134 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1135 bp, bp->b_vp, bp->b_flags); 1136 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1137 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1138 1139 if (bp->b_flags & B_MANAGED) { 1140 bqrelse(bp); 1141 return; 1142 } 1143 1144 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1145 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1146 /* 1147 * Failed write, redirty. Must clear BIO_ERROR to prevent 1148 * pages from being scrapped. If the error is anything 1149 * other than an I/O error (EIO), assume that retryingi 1150 * is futile. 1151 */ 1152 bp->b_ioflags &= ~BIO_ERROR; 1153 bdirty(bp); 1154 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1155 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1156 /* 1157 * Either a failed I/O or we were asked to free or not 1158 * cache the buffer. 1159 */ 1160 bp->b_flags |= B_INVAL; 1161 if (!LIST_EMPTY(&bp->b_dep)) 1162 buf_deallocate(bp); 1163 if (bp->b_flags & B_DELWRI) { 1164 atomic_subtract_int(&numdirtybuffers, 1); 1165 numdirtywakeup(lodirtybuffers); 1166 } 1167 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1168 if ((bp->b_flags & B_VMIO) == 0) { 1169 if (bp->b_bufsize) 1170 allocbuf(bp, 0); 1171 if (bp->b_vp) 1172 brelvp(bp); 1173 } 1174 } 1175 1176 /* 1177 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1178 * is called with B_DELWRI set, the underlying pages may wind up 1179 * getting freed causing a previous write (bdwrite()) to get 'lost' 1180 * because pages associated with a B_DELWRI bp are marked clean. 1181 * 1182 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1183 * if B_DELWRI is set. 1184 * 1185 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1186 * on pages to return pages to the VM page queues. 1187 */ 1188 if (bp->b_flags & B_DELWRI) 1189 bp->b_flags &= ~B_RELBUF; 1190 else if (vm_page_count_severe()) { 1191 /* 1192 * The locking of the BO_LOCK is not necessary since 1193 * BKGRDINPROG cannot be set while we hold the buf 1194 * lock, it can only be cleared if it is already 1195 * pending. 1196 */ 1197 if (bp->b_vp) { 1198 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1199 bp->b_flags |= B_RELBUF; 1200 } else 1201 bp->b_flags |= B_RELBUF; 1202 } 1203 1204 /* 1205 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1206 * constituted, not even NFS buffers now. Two flags effect this. If 1207 * B_INVAL, the struct buf is invalidated but the VM object is kept 1208 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1209 * 1210 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1211 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1212 * buffer is also B_INVAL because it hits the re-dirtying code above. 1213 * 1214 * Normally we can do this whether a buffer is B_DELWRI or not. If 1215 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1216 * the commit state and we cannot afford to lose the buffer. If the 1217 * buffer has a background write in progress, we need to keep it 1218 * around to prevent it from being reconstituted and starting a second 1219 * background write. 1220 */ 1221 if ((bp->b_flags & B_VMIO) 1222 && !(bp->b_vp->v_mount != NULL && 1223 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1224 !vn_isdisk(bp->b_vp, NULL) && 1225 (bp->b_flags & B_DELWRI)) 1226 ) { 1227 1228 int i, j, resid; 1229 vm_page_t m; 1230 off_t foff; 1231 vm_pindex_t poff; 1232 vm_object_t obj; 1233 1234 obj = bp->b_bufobj->bo_object; 1235 1236 /* 1237 * Get the base offset and length of the buffer. Note that 1238 * in the VMIO case if the buffer block size is not 1239 * page-aligned then b_data pointer may not be page-aligned. 1240 * But our b_pages[] array *IS* page aligned. 1241 * 1242 * block sizes less then DEV_BSIZE (usually 512) are not 1243 * supported due to the page granularity bits (m->valid, 1244 * m->dirty, etc...). 1245 * 1246 * See man buf(9) for more information 1247 */ 1248 resid = bp->b_bufsize; 1249 foff = bp->b_offset; 1250 VM_OBJECT_LOCK(obj); 1251 for (i = 0; i < bp->b_npages; i++) { 1252 int had_bogus = 0; 1253 1254 m = bp->b_pages[i]; 1255 1256 /* 1257 * If we hit a bogus page, fixup *all* the bogus pages 1258 * now. 1259 */ 1260 if (m == bogus_page) { 1261 poff = OFF_TO_IDX(bp->b_offset); 1262 had_bogus = 1; 1263 1264 for (j = i; j < bp->b_npages; j++) { 1265 vm_page_t mtmp; 1266 mtmp = bp->b_pages[j]; 1267 if (mtmp == bogus_page) { 1268 mtmp = vm_page_lookup(obj, poff + j); 1269 if (!mtmp) { 1270 panic("brelse: page missing\n"); 1271 } 1272 bp->b_pages[j] = mtmp; 1273 } 1274 } 1275 1276 if ((bp->b_flags & B_INVAL) == 0) { 1277 pmap_qenter( 1278 trunc_page((vm_offset_t)bp->b_data), 1279 bp->b_pages, bp->b_npages); 1280 } 1281 m = bp->b_pages[i]; 1282 } 1283 if ((bp->b_flags & B_NOCACHE) || 1284 (bp->b_ioflags & BIO_ERROR)) { 1285 int poffset = foff & PAGE_MASK; 1286 int presid = resid > (PAGE_SIZE - poffset) ? 1287 (PAGE_SIZE - poffset) : resid; 1288 1289 KASSERT(presid >= 0, ("brelse: extra page")); 1290 vm_page_lock_queues(); 1291 vm_page_set_invalid(m, poffset, presid); 1292 vm_page_unlock_queues(); 1293 if (had_bogus) 1294 printf("avoided corruption bug in bogus_page/brelse code\n"); 1295 } 1296 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1297 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1298 } 1299 VM_OBJECT_UNLOCK(obj); 1300 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1301 vfs_vmio_release(bp); 1302 1303 } else if (bp->b_flags & B_VMIO) { 1304 1305 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1306 vfs_vmio_release(bp); 1307 } 1308 1309 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1310 if (bp->b_bufsize != 0) 1311 allocbuf(bp, 0); 1312 if (bp->b_vp != NULL) 1313 brelvp(bp); 1314 } 1315 1316 if (BUF_LOCKRECURSED(bp)) { 1317 /* do not release to free list */ 1318 BUF_UNLOCK(bp); 1319 return; 1320 } 1321 1322 /* enqueue */ 1323 mtx_lock(&bqlock); 1324 /* Handle delayed bremfree() processing. */ 1325 if (bp->b_flags & B_REMFREE) 1326 bremfreel(bp); 1327 if (bp->b_qindex != QUEUE_NONE) 1328 panic("brelse: free buffer onto another queue???"); 1329 1330 /* buffers with no memory */ 1331 if (bp->b_bufsize == 0) { 1332 bp->b_flags |= B_INVAL; 1333 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1334 if (bp->b_vflags & BV_BKGRDINPROG) 1335 panic("losing buffer 1"); 1336 if (bp->b_kvasize) { 1337 bp->b_qindex = QUEUE_EMPTYKVA; 1338 } else { 1339 bp->b_qindex = QUEUE_EMPTY; 1340 } 1341 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1342 /* buffers with junk contents */ 1343 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1344 (bp->b_ioflags & BIO_ERROR)) { 1345 bp->b_flags |= B_INVAL; 1346 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1347 if (bp->b_vflags & BV_BKGRDINPROG) 1348 panic("losing buffer 2"); 1349 bp->b_qindex = QUEUE_CLEAN; 1350 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1351 /* remaining buffers */ 1352 } else { 1353 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1354 (B_DELWRI|B_NEEDSGIANT)) 1355 bp->b_qindex = QUEUE_DIRTY_GIANT; 1356 else if (bp->b_flags & B_DELWRI) 1357 bp->b_qindex = QUEUE_DIRTY; 1358 else 1359 bp->b_qindex = QUEUE_CLEAN; 1360 if (bp->b_flags & B_AGE) 1361 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1362 else 1363 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1364 } 1365 mtx_unlock(&bqlock); 1366 1367 /* 1368 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1369 * placed the buffer on the correct queue. We must also disassociate 1370 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1371 * find it. 1372 */ 1373 if (bp->b_flags & B_INVAL) { 1374 if (bp->b_flags & B_DELWRI) 1375 bundirty(bp); 1376 if (bp->b_vp) 1377 brelvp(bp); 1378 } 1379 1380 /* 1381 * Fixup numfreebuffers count. The bp is on an appropriate queue 1382 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1383 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1384 * if B_INVAL is set ). 1385 */ 1386 1387 if (!(bp->b_flags & B_DELWRI)) 1388 bufcountwakeup(); 1389 1390 /* 1391 * Something we can maybe free or reuse 1392 */ 1393 if (bp->b_bufsize || bp->b_kvasize) 1394 bufspacewakeup(); 1395 1396 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1397 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1398 panic("brelse: not dirty"); 1399 /* unlock */ 1400 BUF_UNLOCK(bp); 1401 } 1402 1403 /* 1404 * Release a buffer back to the appropriate queue but do not try to free 1405 * it. The buffer is expected to be used again soon. 1406 * 1407 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1408 * biodone() to requeue an async I/O on completion. It is also used when 1409 * known good buffers need to be requeued but we think we may need the data 1410 * again soon. 1411 * 1412 * XXX we should be able to leave the B_RELBUF hint set on completion. 1413 */ 1414 void 1415 bqrelse(struct buf *bp) 1416 { 1417 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1418 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1419 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1420 1421 if (BUF_LOCKRECURSED(bp)) { 1422 /* do not release to free list */ 1423 BUF_UNLOCK(bp); 1424 return; 1425 } 1426 1427 if (bp->b_flags & B_MANAGED) { 1428 if (bp->b_flags & B_REMFREE) { 1429 mtx_lock(&bqlock); 1430 bremfreel(bp); 1431 mtx_unlock(&bqlock); 1432 } 1433 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1434 BUF_UNLOCK(bp); 1435 return; 1436 } 1437 1438 mtx_lock(&bqlock); 1439 /* Handle delayed bremfree() processing. */ 1440 if (bp->b_flags & B_REMFREE) 1441 bremfreel(bp); 1442 if (bp->b_qindex != QUEUE_NONE) 1443 panic("bqrelse: free buffer onto another queue???"); 1444 /* buffers with stale but valid contents */ 1445 if (bp->b_flags & B_DELWRI) { 1446 if (bp->b_flags & B_NEEDSGIANT) 1447 bp->b_qindex = QUEUE_DIRTY_GIANT; 1448 else 1449 bp->b_qindex = QUEUE_DIRTY; 1450 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1451 } else { 1452 /* 1453 * The locking of the BO_LOCK for checking of the 1454 * BV_BKGRDINPROG is not necessary since the 1455 * BV_BKGRDINPROG cannot be set while we hold the buf 1456 * lock, it can only be cleared if it is already 1457 * pending. 1458 */ 1459 if (!vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1460 bp->b_qindex = QUEUE_CLEAN; 1461 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1462 b_freelist); 1463 } else { 1464 /* 1465 * We are too low on memory, we have to try to free 1466 * the buffer (most importantly: the wired pages 1467 * making up its backing store) *now*. 1468 */ 1469 mtx_unlock(&bqlock); 1470 brelse(bp); 1471 return; 1472 } 1473 } 1474 mtx_unlock(&bqlock); 1475 1476 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1477 bufcountwakeup(); 1478 1479 /* 1480 * Something we can maybe free or reuse. 1481 */ 1482 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1483 bufspacewakeup(); 1484 1485 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1486 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1487 panic("bqrelse: not dirty"); 1488 /* unlock */ 1489 BUF_UNLOCK(bp); 1490 } 1491 1492 /* Give pages used by the bp back to the VM system (where possible) */ 1493 static void 1494 vfs_vmio_release(struct buf *bp) 1495 { 1496 int i; 1497 vm_page_t m; 1498 1499 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1500 vm_page_lock_queues(); 1501 for (i = 0; i < bp->b_npages; i++) { 1502 m = bp->b_pages[i]; 1503 bp->b_pages[i] = NULL; 1504 /* 1505 * In order to keep page LRU ordering consistent, put 1506 * everything on the inactive queue. 1507 */ 1508 vm_page_unwire(m, 0); 1509 /* 1510 * We don't mess with busy pages, it is 1511 * the responsibility of the process that 1512 * busied the pages to deal with them. 1513 */ 1514 if ((m->oflags & VPO_BUSY) || (m->busy != 0)) 1515 continue; 1516 1517 if (m->wire_count == 0) { 1518 /* 1519 * Might as well free the page if we can and it has 1520 * no valid data. We also free the page if the 1521 * buffer was used for direct I/O 1522 */ 1523 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1524 m->hold_count == 0) { 1525 vm_page_free(m); 1526 } else if (bp->b_flags & B_DIRECT) { 1527 vm_page_try_to_free(m); 1528 } else if (vm_page_count_severe()) { 1529 vm_page_try_to_cache(m); 1530 } 1531 } 1532 } 1533 vm_page_unlock_queues(); 1534 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1535 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1536 1537 if (bp->b_bufsize) { 1538 bufspacewakeup(); 1539 bp->b_bufsize = 0; 1540 } 1541 bp->b_npages = 0; 1542 bp->b_flags &= ~B_VMIO; 1543 if (bp->b_vp) 1544 brelvp(bp); 1545 } 1546 1547 /* 1548 * Check to see if a block at a particular lbn is available for a clustered 1549 * write. 1550 */ 1551 static int 1552 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1553 { 1554 struct buf *bpa; 1555 int match; 1556 1557 match = 0; 1558 1559 /* If the buf isn't in core skip it */ 1560 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1561 return (0); 1562 1563 /* If the buf is busy we don't want to wait for it */ 1564 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1565 return (0); 1566 1567 /* Only cluster with valid clusterable delayed write buffers */ 1568 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1569 (B_DELWRI | B_CLUSTEROK)) 1570 goto done; 1571 1572 if (bpa->b_bufsize != size) 1573 goto done; 1574 1575 /* 1576 * Check to see if it is in the expected place on disk and that the 1577 * block has been mapped. 1578 */ 1579 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1580 match = 1; 1581 done: 1582 BUF_UNLOCK(bpa); 1583 return (match); 1584 } 1585 1586 /* 1587 * vfs_bio_awrite: 1588 * 1589 * Implement clustered async writes for clearing out B_DELWRI buffers. 1590 * This is much better then the old way of writing only one buffer at 1591 * a time. Note that we may not be presented with the buffers in the 1592 * correct order, so we search for the cluster in both directions. 1593 */ 1594 int 1595 vfs_bio_awrite(struct buf *bp) 1596 { 1597 struct bufobj *bo; 1598 int i; 1599 int j; 1600 daddr_t lblkno = bp->b_lblkno; 1601 struct vnode *vp = bp->b_vp; 1602 int ncl; 1603 int nwritten; 1604 int size; 1605 int maxcl; 1606 1607 bo = &vp->v_bufobj; 1608 /* 1609 * right now we support clustered writing only to regular files. If 1610 * we find a clusterable block we could be in the middle of a cluster 1611 * rather then at the beginning. 1612 */ 1613 if ((vp->v_type == VREG) && 1614 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1615 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1616 1617 size = vp->v_mount->mnt_stat.f_iosize; 1618 maxcl = MAXPHYS / size; 1619 1620 BO_LOCK(bo); 1621 for (i = 1; i < maxcl; i++) 1622 if (vfs_bio_clcheck(vp, size, lblkno + i, 1623 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1624 break; 1625 1626 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1627 if (vfs_bio_clcheck(vp, size, lblkno - j, 1628 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1629 break; 1630 BO_UNLOCK(bo); 1631 --j; 1632 ncl = i + j; 1633 /* 1634 * this is a possible cluster write 1635 */ 1636 if (ncl != 1) { 1637 BUF_UNLOCK(bp); 1638 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1639 return nwritten; 1640 } 1641 } 1642 bremfree(bp); 1643 bp->b_flags |= B_ASYNC; 1644 /* 1645 * default (old) behavior, writing out only one block 1646 * 1647 * XXX returns b_bufsize instead of b_bcount for nwritten? 1648 */ 1649 nwritten = bp->b_bufsize; 1650 (void) bwrite(bp); 1651 1652 return nwritten; 1653 } 1654 1655 /* 1656 * getnewbuf: 1657 * 1658 * Find and initialize a new buffer header, freeing up existing buffers 1659 * in the bufqueues as necessary. The new buffer is returned locked. 1660 * 1661 * Important: B_INVAL is not set. If the caller wishes to throw the 1662 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1663 * 1664 * We block if: 1665 * We have insufficient buffer headers 1666 * We have insufficient buffer space 1667 * buffer_map is too fragmented ( space reservation fails ) 1668 * If we have to flush dirty buffers ( but we try to avoid this ) 1669 * 1670 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1671 * Instead we ask the buf daemon to do it for us. We attempt to 1672 * avoid piecemeal wakeups of the pageout daemon. 1673 */ 1674 1675 static struct buf * 1676 getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1677 { 1678 struct buf *bp; 1679 struct buf *nbp; 1680 int defrag = 0; 1681 int nqindex; 1682 static int flushingbufs; 1683 1684 /* 1685 * We can't afford to block since we might be holding a vnode lock, 1686 * which may prevent system daemons from running. We deal with 1687 * low-memory situations by proactively returning memory and running 1688 * async I/O rather then sync I/O. 1689 */ 1690 1691 atomic_add_int(&getnewbufcalls, 1); 1692 atomic_subtract_int(&getnewbufrestarts, 1); 1693 restart: 1694 atomic_add_int(&getnewbufrestarts, 1); 1695 1696 /* 1697 * Setup for scan. If we do not have enough free buffers, 1698 * we setup a degenerate case that immediately fails. Note 1699 * that if we are specially marked process, we are allowed to 1700 * dip into our reserves. 1701 * 1702 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1703 * 1704 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1705 * However, there are a number of cases (defragging, reusing, ...) 1706 * where we cannot backup. 1707 */ 1708 mtx_lock(&bqlock); 1709 nqindex = QUEUE_EMPTYKVA; 1710 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1711 1712 if (nbp == NULL) { 1713 /* 1714 * If no EMPTYKVA buffers and we are either 1715 * defragging or reusing, locate a CLEAN buffer 1716 * to free or reuse. If bufspace useage is low 1717 * skip this step so we can allocate a new buffer. 1718 */ 1719 if (defrag || bufspace >= lobufspace) { 1720 nqindex = QUEUE_CLEAN; 1721 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1722 } 1723 1724 /* 1725 * If we could not find or were not allowed to reuse a 1726 * CLEAN buffer, check to see if it is ok to use an EMPTY 1727 * buffer. We can only use an EMPTY buffer if allocating 1728 * its KVA would not otherwise run us out of buffer space. 1729 */ 1730 if (nbp == NULL && defrag == 0 && 1731 bufspace + maxsize < hibufspace) { 1732 nqindex = QUEUE_EMPTY; 1733 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1734 } 1735 } 1736 1737 /* 1738 * Run scan, possibly freeing data and/or kva mappings on the fly 1739 * depending. 1740 */ 1741 1742 while ((bp = nbp) != NULL) { 1743 int qindex = nqindex; 1744 1745 /* 1746 * Calculate next bp ( we can only use it if we do not block 1747 * or do other fancy things ). 1748 */ 1749 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1750 switch(qindex) { 1751 case QUEUE_EMPTY: 1752 nqindex = QUEUE_EMPTYKVA; 1753 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1754 break; 1755 /* FALLTHROUGH */ 1756 case QUEUE_EMPTYKVA: 1757 nqindex = QUEUE_CLEAN; 1758 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1759 break; 1760 /* FALLTHROUGH */ 1761 case QUEUE_CLEAN: 1762 /* 1763 * nbp is NULL. 1764 */ 1765 break; 1766 } 1767 } 1768 /* 1769 * If we are defragging then we need a buffer with 1770 * b_kvasize != 0. XXX this situation should no longer 1771 * occur, if defrag is non-zero the buffer's b_kvasize 1772 * should also be non-zero at this point. XXX 1773 */ 1774 if (defrag && bp->b_kvasize == 0) { 1775 printf("Warning: defrag empty buffer %p\n", bp); 1776 continue; 1777 } 1778 1779 /* 1780 * Start freeing the bp. This is somewhat involved. nbp 1781 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1782 */ 1783 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1784 continue; 1785 if (bp->b_vp) { 1786 BO_LOCK(bp->b_bufobj); 1787 if (bp->b_vflags & BV_BKGRDINPROG) { 1788 BO_UNLOCK(bp->b_bufobj); 1789 BUF_UNLOCK(bp); 1790 continue; 1791 } 1792 BO_UNLOCK(bp->b_bufobj); 1793 } 1794 CTR6(KTR_BUF, 1795 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1796 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1797 bp->b_kvasize, bp->b_bufsize, qindex); 1798 1799 /* 1800 * Sanity Checks 1801 */ 1802 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1803 1804 /* 1805 * Note: we no longer distinguish between VMIO and non-VMIO 1806 * buffers. 1807 */ 1808 1809 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1810 1811 bremfreel(bp); 1812 mtx_unlock(&bqlock); 1813 1814 if (qindex == QUEUE_CLEAN) { 1815 if (bp->b_flags & B_VMIO) { 1816 bp->b_flags &= ~B_ASYNC; 1817 vfs_vmio_release(bp); 1818 } 1819 if (bp->b_vp) 1820 brelvp(bp); 1821 } 1822 1823 /* 1824 * NOTE: nbp is now entirely invalid. We can only restart 1825 * the scan from this point on. 1826 * 1827 * Get the rest of the buffer freed up. b_kva* is still 1828 * valid after this operation. 1829 */ 1830 1831 if (bp->b_rcred != NOCRED) { 1832 crfree(bp->b_rcred); 1833 bp->b_rcred = NOCRED; 1834 } 1835 if (bp->b_wcred != NOCRED) { 1836 crfree(bp->b_wcred); 1837 bp->b_wcred = NOCRED; 1838 } 1839 if (!LIST_EMPTY(&bp->b_dep)) 1840 buf_deallocate(bp); 1841 if (bp->b_vflags & BV_BKGRDINPROG) 1842 panic("losing buffer 3"); 1843 KASSERT(bp->b_vp == NULL, 1844 ("bp: %p still has vnode %p. qindex: %d", 1845 bp, bp->b_vp, qindex)); 1846 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1847 ("bp: %p still on a buffer list. xflags %X", 1848 bp, bp->b_xflags)); 1849 1850 if (bp->b_bufsize) 1851 allocbuf(bp, 0); 1852 1853 bp->b_flags = 0; 1854 bp->b_ioflags = 0; 1855 bp->b_xflags = 0; 1856 bp->b_vflags = 0; 1857 bp->b_vp = NULL; 1858 bp->b_blkno = bp->b_lblkno = 0; 1859 bp->b_offset = NOOFFSET; 1860 bp->b_iodone = 0; 1861 bp->b_error = 0; 1862 bp->b_resid = 0; 1863 bp->b_bcount = 0; 1864 bp->b_npages = 0; 1865 bp->b_dirtyoff = bp->b_dirtyend = 0; 1866 bp->b_bufobj = NULL; 1867 bp->b_pin_count = 0; 1868 bp->b_fsprivate1 = NULL; 1869 bp->b_fsprivate2 = NULL; 1870 bp->b_fsprivate3 = NULL; 1871 1872 LIST_INIT(&bp->b_dep); 1873 1874 /* 1875 * If we are defragging then free the buffer. 1876 */ 1877 if (defrag) { 1878 bp->b_flags |= B_INVAL; 1879 bfreekva(bp); 1880 brelse(bp); 1881 defrag = 0; 1882 goto restart; 1883 } 1884 1885 /* 1886 * Notify any waiters for the buffer lock about 1887 * identity change by freeing the buffer. 1888 */ 1889 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 1890 bp->b_flags |= B_INVAL; 1891 bfreekva(bp); 1892 brelse(bp); 1893 goto restart; 1894 } 1895 1896 /* 1897 * If we are overcomitted then recover the buffer and its 1898 * KVM space. This occurs in rare situations when multiple 1899 * processes are blocked in getnewbuf() or allocbuf(). 1900 */ 1901 if (bufspace >= hibufspace) 1902 flushingbufs = 1; 1903 if (flushingbufs && bp->b_kvasize != 0) { 1904 bp->b_flags |= B_INVAL; 1905 bfreekva(bp); 1906 brelse(bp); 1907 goto restart; 1908 } 1909 if (bufspace < lobufspace) 1910 flushingbufs = 0; 1911 break; 1912 } 1913 1914 /* 1915 * If we exhausted our list, sleep as appropriate. We may have to 1916 * wakeup various daemons and write out some dirty buffers. 1917 * 1918 * Generally we are sleeping due to insufficient buffer space. 1919 */ 1920 1921 if (bp == NULL) { 1922 int flags; 1923 char *waitmsg; 1924 1925 if (defrag) { 1926 flags = VFS_BIO_NEED_BUFSPACE; 1927 waitmsg = "nbufkv"; 1928 } else if (bufspace >= hibufspace) { 1929 waitmsg = "nbufbs"; 1930 flags = VFS_BIO_NEED_BUFSPACE; 1931 } else { 1932 waitmsg = "newbuf"; 1933 flags = VFS_BIO_NEED_ANY; 1934 } 1935 mtx_lock(&nblock); 1936 needsbuffer |= flags; 1937 mtx_unlock(&nblock); 1938 mtx_unlock(&bqlock); 1939 1940 bd_speedup(); /* heeeelp */ 1941 1942 mtx_lock(&nblock); 1943 while (needsbuffer & flags) { 1944 if (msleep(&needsbuffer, &nblock, 1945 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1946 mtx_unlock(&nblock); 1947 return (NULL); 1948 } 1949 } 1950 mtx_unlock(&nblock); 1951 } else { 1952 /* 1953 * We finally have a valid bp. We aren't quite out of the 1954 * woods, we still have to reserve kva space. In order 1955 * to keep fragmentation sane we only allocate kva in 1956 * BKVASIZE chunks. 1957 */ 1958 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1959 1960 if (maxsize != bp->b_kvasize) { 1961 vm_offset_t addr = 0; 1962 1963 bfreekva(bp); 1964 1965 vm_map_lock(buffer_map); 1966 if (vm_map_findspace(buffer_map, 1967 vm_map_min(buffer_map), maxsize, &addr)) { 1968 /* 1969 * Uh oh. Buffer map is to fragmented. We 1970 * must defragment the map. 1971 */ 1972 atomic_add_int(&bufdefragcnt, 1); 1973 vm_map_unlock(buffer_map); 1974 defrag = 1; 1975 bp->b_flags |= B_INVAL; 1976 brelse(bp); 1977 goto restart; 1978 } 1979 if (addr) { 1980 vm_map_insert(buffer_map, NULL, 0, 1981 addr, addr + maxsize, 1982 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1983 1984 bp->b_kvabase = (caddr_t) addr; 1985 bp->b_kvasize = maxsize; 1986 atomic_add_int(&bufspace, bp->b_kvasize); 1987 atomic_add_int(&bufreusecnt, 1); 1988 } 1989 vm_map_unlock(buffer_map); 1990 } 1991 bp->b_saveaddr = bp->b_kvabase; 1992 bp->b_data = bp->b_saveaddr; 1993 } 1994 return(bp); 1995 } 1996 1997 /* 1998 * buf_daemon: 1999 * 2000 * buffer flushing daemon. Buffers are normally flushed by the 2001 * update daemon but if it cannot keep up this process starts to 2002 * take the load in an attempt to prevent getnewbuf() from blocking. 2003 */ 2004 2005 static struct kproc_desc buf_kp = { 2006 "bufdaemon", 2007 buf_daemon, 2008 &bufdaemonproc 2009 }; 2010 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2011 2012 static void 2013 buf_daemon() 2014 { 2015 2016 /* 2017 * This process needs to be suspended prior to shutdown sync. 2018 */ 2019 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2020 SHUTDOWN_PRI_LAST); 2021 2022 /* 2023 * This process is allowed to take the buffer cache to the limit 2024 */ 2025 curthread->td_pflags |= TDP_NORUNNINGBUF; 2026 mtx_lock(&bdlock); 2027 for (;;) { 2028 bd_request = 0; 2029 mtx_unlock(&bdlock); 2030 2031 kproc_suspend_check(bufdaemonproc); 2032 2033 /* 2034 * Do the flush. Limit the amount of in-transit I/O we 2035 * allow to build up, otherwise we would completely saturate 2036 * the I/O system. Wakeup any waiting processes before we 2037 * normally would so they can run in parallel with our drain. 2038 */ 2039 while (numdirtybuffers > lodirtybuffers) { 2040 int flushed; 2041 2042 flushed = flushbufqueues(QUEUE_DIRTY, 0); 2043 /* The list empty check here is slightly racy */ 2044 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2045 mtx_lock(&Giant); 2046 flushed += flushbufqueues(QUEUE_DIRTY_GIANT, 0); 2047 mtx_unlock(&Giant); 2048 } 2049 if (flushed == 0) { 2050 /* 2051 * Could not find any buffers without rollback 2052 * dependencies, so just write the first one 2053 * in the hopes of eventually making progress. 2054 */ 2055 flushbufqueues(QUEUE_DIRTY, 1); 2056 if (!TAILQ_EMPTY( 2057 &bufqueues[QUEUE_DIRTY_GIANT])) { 2058 mtx_lock(&Giant); 2059 flushbufqueues(QUEUE_DIRTY_GIANT, 1); 2060 mtx_unlock(&Giant); 2061 } 2062 break; 2063 } 2064 uio_yield(); 2065 } 2066 2067 /* 2068 * Only clear bd_request if we have reached our low water 2069 * mark. The buf_daemon normally waits 1 second and 2070 * then incrementally flushes any dirty buffers that have 2071 * built up, within reason. 2072 * 2073 * If we were unable to hit our low water mark and couldn't 2074 * find any flushable buffers, we sleep half a second. 2075 * Otherwise we loop immediately. 2076 */ 2077 mtx_lock(&bdlock); 2078 if (numdirtybuffers <= lodirtybuffers) { 2079 /* 2080 * We reached our low water mark, reset the 2081 * request and sleep until we are needed again. 2082 * The sleep is just so the suspend code works. 2083 */ 2084 bd_request = 0; 2085 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2086 } else { 2087 /* 2088 * We couldn't find any flushable dirty buffers but 2089 * still have too many dirty buffers, we 2090 * have to sleep and try again. (rare) 2091 */ 2092 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2093 } 2094 } 2095 } 2096 2097 /* 2098 * flushbufqueues: 2099 * 2100 * Try to flush a buffer in the dirty queue. We must be careful to 2101 * free up B_INVAL buffers instead of write them, which NFS is 2102 * particularly sensitive to. 2103 */ 2104 static int flushwithdeps = 0; 2105 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2106 0, "Number of buffers flushed with dependecies that require rollbacks"); 2107 2108 static int 2109 flushbufqueues(int queue, int flushdeps) 2110 { 2111 struct buf sentinel; 2112 struct vnode *vp; 2113 struct mount *mp; 2114 struct buf *bp; 2115 int hasdeps; 2116 int flushed; 2117 int target; 2118 2119 target = numdirtybuffers - lodirtybuffers; 2120 if (flushdeps && target > 2) 2121 target /= 2; 2122 flushed = 0; 2123 bp = NULL; 2124 mtx_lock(&bqlock); 2125 TAILQ_INSERT_TAIL(&bufqueues[queue], &sentinel, b_freelist); 2126 while (flushed != target) { 2127 bp = TAILQ_FIRST(&bufqueues[queue]); 2128 if (bp == &sentinel) 2129 break; 2130 TAILQ_REMOVE(&bufqueues[queue], bp, b_freelist); 2131 TAILQ_INSERT_TAIL(&bufqueues[queue], bp, b_freelist); 2132 2133 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2134 continue; 2135 if (bp->b_pin_count > 0) { 2136 BUF_UNLOCK(bp); 2137 continue; 2138 } 2139 BO_LOCK(bp->b_bufobj); 2140 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2141 (bp->b_flags & B_DELWRI) == 0) { 2142 BO_UNLOCK(bp->b_bufobj); 2143 BUF_UNLOCK(bp); 2144 continue; 2145 } 2146 BO_UNLOCK(bp->b_bufobj); 2147 if (bp->b_flags & B_INVAL) { 2148 bremfreel(bp); 2149 mtx_unlock(&bqlock); 2150 brelse(bp); 2151 flushed++; 2152 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2153 mtx_lock(&bqlock); 2154 continue; 2155 } 2156 2157 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2158 if (flushdeps == 0) { 2159 BUF_UNLOCK(bp); 2160 continue; 2161 } 2162 hasdeps = 1; 2163 } else 2164 hasdeps = 0; 2165 /* 2166 * We must hold the lock on a vnode before writing 2167 * one of its buffers. Otherwise we may confuse, or 2168 * in the case of a snapshot vnode, deadlock the 2169 * system. 2170 * 2171 * The lock order here is the reverse of the normal 2172 * of vnode followed by buf lock. This is ok because 2173 * the NOWAIT will prevent deadlock. 2174 */ 2175 vp = bp->b_vp; 2176 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2177 BUF_UNLOCK(bp); 2178 continue; 2179 } 2180 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 2181 mtx_unlock(&bqlock); 2182 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2183 bp, bp->b_vp, bp->b_flags); 2184 vfs_bio_awrite(bp); 2185 vn_finished_write(mp); 2186 VOP_UNLOCK(vp, 0); 2187 flushwithdeps += hasdeps; 2188 flushed++; 2189 waitrunningbufspace(); 2190 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2191 mtx_lock(&bqlock); 2192 continue; 2193 } 2194 vn_finished_write(mp); 2195 BUF_UNLOCK(bp); 2196 } 2197 TAILQ_REMOVE(&bufqueues[queue], &sentinel, b_freelist); 2198 mtx_unlock(&bqlock); 2199 return (flushed); 2200 } 2201 2202 /* 2203 * Check to see if a block is currently memory resident. 2204 */ 2205 struct buf * 2206 incore(struct bufobj *bo, daddr_t blkno) 2207 { 2208 struct buf *bp; 2209 2210 BO_LOCK(bo); 2211 bp = gbincore(bo, blkno); 2212 BO_UNLOCK(bo); 2213 return (bp); 2214 } 2215 2216 /* 2217 * Returns true if no I/O is needed to access the 2218 * associated VM object. This is like incore except 2219 * it also hunts around in the VM system for the data. 2220 */ 2221 2222 static int 2223 inmem(struct vnode * vp, daddr_t blkno) 2224 { 2225 vm_object_t obj; 2226 vm_offset_t toff, tinc, size; 2227 vm_page_t m; 2228 vm_ooffset_t off; 2229 2230 ASSERT_VOP_LOCKED(vp, "inmem"); 2231 2232 if (incore(&vp->v_bufobj, blkno)) 2233 return 1; 2234 if (vp->v_mount == NULL) 2235 return 0; 2236 obj = vp->v_object; 2237 if (obj == NULL) 2238 return (0); 2239 2240 size = PAGE_SIZE; 2241 if (size > vp->v_mount->mnt_stat.f_iosize) 2242 size = vp->v_mount->mnt_stat.f_iosize; 2243 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2244 2245 VM_OBJECT_LOCK(obj); 2246 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2247 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2248 if (!m) 2249 goto notinmem; 2250 tinc = size; 2251 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2252 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2253 if (vm_page_is_valid(m, 2254 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2255 goto notinmem; 2256 } 2257 VM_OBJECT_UNLOCK(obj); 2258 return 1; 2259 2260 notinmem: 2261 VM_OBJECT_UNLOCK(obj); 2262 return (0); 2263 } 2264 2265 /* 2266 * vfs_setdirty: 2267 * 2268 * Sets the dirty range for a buffer based on the status of the dirty 2269 * bits in the pages comprising the buffer. 2270 * 2271 * The range is limited to the size of the buffer. 2272 * 2273 * This routine is primarily used by NFS, but is generalized for the 2274 * B_VMIO case. 2275 */ 2276 static void 2277 vfs_setdirty(struct buf *bp) 2278 { 2279 2280 /* 2281 * Degenerate case - empty buffer 2282 */ 2283 2284 if (bp->b_bufsize == 0) 2285 return; 2286 2287 /* 2288 * We qualify the scan for modified pages on whether the 2289 * object has been flushed yet. 2290 */ 2291 2292 if ((bp->b_flags & B_VMIO) == 0) 2293 return; 2294 2295 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2296 vfs_setdirty_locked_object(bp); 2297 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2298 } 2299 2300 static void 2301 vfs_setdirty_locked_object(struct buf *bp) 2302 { 2303 vm_object_t object; 2304 int i; 2305 2306 object = bp->b_bufobj->bo_object; 2307 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2308 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2309 vm_offset_t boffset; 2310 vm_offset_t eoffset; 2311 2312 vm_page_lock_queues(); 2313 /* 2314 * test the pages to see if they have been modified directly 2315 * by users through the VM system. 2316 */ 2317 for (i = 0; i < bp->b_npages; i++) 2318 vm_page_test_dirty(bp->b_pages[i]); 2319 2320 /* 2321 * Calculate the encompassing dirty range, boffset and eoffset, 2322 * (eoffset - boffset) bytes. 2323 */ 2324 2325 for (i = 0; i < bp->b_npages; i++) { 2326 if (bp->b_pages[i]->dirty) 2327 break; 2328 } 2329 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2330 2331 for (i = bp->b_npages - 1; i >= 0; --i) { 2332 if (bp->b_pages[i]->dirty) { 2333 break; 2334 } 2335 } 2336 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2337 2338 vm_page_unlock_queues(); 2339 /* 2340 * Fit it to the buffer. 2341 */ 2342 2343 if (eoffset > bp->b_bcount) 2344 eoffset = bp->b_bcount; 2345 2346 /* 2347 * If we have a good dirty range, merge with the existing 2348 * dirty range. 2349 */ 2350 2351 if (boffset < eoffset) { 2352 if (bp->b_dirtyoff > boffset) 2353 bp->b_dirtyoff = boffset; 2354 if (bp->b_dirtyend < eoffset) 2355 bp->b_dirtyend = eoffset; 2356 } 2357 } 2358 } 2359 2360 /* 2361 * getblk: 2362 * 2363 * Get a block given a specified block and offset into a file/device. 2364 * The buffers B_DONE bit will be cleared on return, making it almost 2365 * ready for an I/O initiation. B_INVAL may or may not be set on 2366 * return. The caller should clear B_INVAL prior to initiating a 2367 * READ. 2368 * 2369 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2370 * an existing buffer. 2371 * 2372 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2373 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2374 * and then cleared based on the backing VM. If the previous buffer is 2375 * non-0-sized but invalid, B_CACHE will be cleared. 2376 * 2377 * If getblk() must create a new buffer, the new buffer is returned with 2378 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2379 * case it is returned with B_INVAL clear and B_CACHE set based on the 2380 * backing VM. 2381 * 2382 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2383 * B_CACHE bit is clear. 2384 * 2385 * What this means, basically, is that the caller should use B_CACHE to 2386 * determine whether the buffer is fully valid or not and should clear 2387 * B_INVAL prior to issuing a read. If the caller intends to validate 2388 * the buffer by loading its data area with something, the caller needs 2389 * to clear B_INVAL. If the caller does this without issuing an I/O, 2390 * the caller should set B_CACHE ( as an optimization ), else the caller 2391 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2392 * a write attempt or if it was a successfull read. If the caller 2393 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2394 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2395 */ 2396 struct buf * 2397 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2398 int flags) 2399 { 2400 struct buf *bp; 2401 struct bufobj *bo; 2402 int error; 2403 2404 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2405 ASSERT_VOP_LOCKED(vp, "getblk"); 2406 if (size > MAXBSIZE) 2407 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2408 2409 bo = &vp->v_bufobj; 2410 loop: 2411 /* 2412 * Block if we are low on buffers. Certain processes are allowed 2413 * to completely exhaust the buffer cache. 2414 * 2415 * If this check ever becomes a bottleneck it may be better to 2416 * move it into the else, when gbincore() fails. At the moment 2417 * it isn't a problem. 2418 * 2419 * XXX remove if 0 sections (clean this up after its proven) 2420 */ 2421 if (numfreebuffers == 0) { 2422 if (TD_IS_IDLETHREAD(curthread)) 2423 return NULL; 2424 mtx_lock(&nblock); 2425 needsbuffer |= VFS_BIO_NEED_ANY; 2426 mtx_unlock(&nblock); 2427 } 2428 2429 BO_LOCK(bo); 2430 bp = gbincore(bo, blkno); 2431 if (bp != NULL) { 2432 int lockflags; 2433 /* 2434 * Buffer is in-core. If the buffer is not busy, it must 2435 * be on a queue. 2436 */ 2437 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2438 2439 if (flags & GB_LOCK_NOWAIT) 2440 lockflags |= LK_NOWAIT; 2441 2442 error = BUF_TIMELOCK(bp, lockflags, 2443 BO_MTX(bo), "getblk", slpflag, slptimeo); 2444 2445 /* 2446 * If we slept and got the lock we have to restart in case 2447 * the buffer changed identities. 2448 */ 2449 if (error == ENOLCK) 2450 goto loop; 2451 /* We timed out or were interrupted. */ 2452 else if (error) 2453 return (NULL); 2454 2455 /* 2456 * The buffer is locked. B_CACHE is cleared if the buffer is 2457 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2458 * and for a VMIO buffer B_CACHE is adjusted according to the 2459 * backing VM cache. 2460 */ 2461 if (bp->b_flags & B_INVAL) 2462 bp->b_flags &= ~B_CACHE; 2463 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2464 bp->b_flags |= B_CACHE; 2465 bremfree(bp); 2466 2467 /* 2468 * check for size inconsistancies for non-VMIO case. 2469 */ 2470 2471 if (bp->b_bcount != size) { 2472 if ((bp->b_flags & B_VMIO) == 0 || 2473 (size > bp->b_kvasize)) { 2474 if (bp->b_flags & B_DELWRI) { 2475 /* 2476 * If buffer is pinned and caller does 2477 * not want sleep waiting for it to be 2478 * unpinned, bail out 2479 * */ 2480 if (bp->b_pin_count > 0) { 2481 if (flags & GB_LOCK_NOWAIT) { 2482 bqrelse(bp); 2483 return (NULL); 2484 } else { 2485 bunpin_wait(bp); 2486 } 2487 } 2488 bp->b_flags |= B_NOCACHE; 2489 bwrite(bp); 2490 } else { 2491 if (LIST_EMPTY(&bp->b_dep)) { 2492 bp->b_flags |= B_RELBUF; 2493 brelse(bp); 2494 } else { 2495 bp->b_flags |= B_NOCACHE; 2496 bwrite(bp); 2497 } 2498 } 2499 goto loop; 2500 } 2501 } 2502 2503 /* 2504 * If the size is inconsistant in the VMIO case, we can resize 2505 * the buffer. This might lead to B_CACHE getting set or 2506 * cleared. If the size has not changed, B_CACHE remains 2507 * unchanged from its previous state. 2508 */ 2509 2510 if (bp->b_bcount != size) 2511 allocbuf(bp, size); 2512 2513 KASSERT(bp->b_offset != NOOFFSET, 2514 ("getblk: no buffer offset")); 2515 2516 /* 2517 * A buffer with B_DELWRI set and B_CACHE clear must 2518 * be committed before we can return the buffer in 2519 * order to prevent the caller from issuing a read 2520 * ( due to B_CACHE not being set ) and overwriting 2521 * it. 2522 * 2523 * Most callers, including NFS and FFS, need this to 2524 * operate properly either because they assume they 2525 * can issue a read if B_CACHE is not set, or because 2526 * ( for example ) an uncached B_DELWRI might loop due 2527 * to softupdates re-dirtying the buffer. In the latter 2528 * case, B_CACHE is set after the first write completes, 2529 * preventing further loops. 2530 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2531 * above while extending the buffer, we cannot allow the 2532 * buffer to remain with B_CACHE set after the write 2533 * completes or it will represent a corrupt state. To 2534 * deal with this we set B_NOCACHE to scrap the buffer 2535 * after the write. 2536 * 2537 * We might be able to do something fancy, like setting 2538 * B_CACHE in bwrite() except if B_DELWRI is already set, 2539 * so the below call doesn't set B_CACHE, but that gets real 2540 * confusing. This is much easier. 2541 */ 2542 2543 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2544 bp->b_flags |= B_NOCACHE; 2545 bwrite(bp); 2546 goto loop; 2547 } 2548 bp->b_flags &= ~B_DONE; 2549 } else { 2550 int bsize, maxsize, vmio; 2551 off_t offset; 2552 2553 /* 2554 * Buffer is not in-core, create new buffer. The buffer 2555 * returned by getnewbuf() is locked. Note that the returned 2556 * buffer is also considered valid (not marked B_INVAL). 2557 */ 2558 BO_UNLOCK(bo); 2559 /* 2560 * If the user does not want us to create the buffer, bail out 2561 * here. 2562 */ 2563 if (flags & GB_NOCREAT) 2564 return NULL; 2565 bsize = bo->bo_bsize; 2566 offset = blkno * bsize; 2567 vmio = vp->v_object != NULL; 2568 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2569 maxsize = imax(maxsize, bsize); 2570 2571 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2572 if (bp == NULL) { 2573 if (slpflag || slptimeo) 2574 return NULL; 2575 goto loop; 2576 } 2577 2578 /* 2579 * This code is used to make sure that a buffer is not 2580 * created while the getnewbuf routine is blocked. 2581 * This can be a problem whether the vnode is locked or not. 2582 * If the buffer is created out from under us, we have to 2583 * throw away the one we just created. 2584 * 2585 * Note: this must occur before we associate the buffer 2586 * with the vp especially considering limitations in 2587 * the splay tree implementation when dealing with duplicate 2588 * lblkno's. 2589 */ 2590 BO_LOCK(bo); 2591 if (gbincore(bo, blkno)) { 2592 BO_UNLOCK(bo); 2593 bp->b_flags |= B_INVAL; 2594 brelse(bp); 2595 goto loop; 2596 } 2597 2598 /* 2599 * Insert the buffer into the hash, so that it can 2600 * be found by incore. 2601 */ 2602 bp->b_blkno = bp->b_lblkno = blkno; 2603 bp->b_offset = offset; 2604 bgetvp(vp, bp); 2605 BO_UNLOCK(bo); 2606 2607 /* 2608 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2609 * buffer size starts out as 0, B_CACHE will be set by 2610 * allocbuf() for the VMIO case prior to it testing the 2611 * backing store for validity. 2612 */ 2613 2614 if (vmio) { 2615 bp->b_flags |= B_VMIO; 2616 #if defined(VFS_BIO_DEBUG) 2617 if (vn_canvmio(vp) != TRUE) 2618 printf("getblk: VMIO on vnode type %d\n", 2619 vp->v_type); 2620 #endif 2621 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2622 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2623 bp, vp->v_object, bp->b_bufobj->bo_object)); 2624 } else { 2625 bp->b_flags &= ~B_VMIO; 2626 KASSERT(bp->b_bufobj->bo_object == NULL, 2627 ("ARGH! has b_bufobj->bo_object %p %p\n", 2628 bp, bp->b_bufobj->bo_object)); 2629 } 2630 2631 allocbuf(bp, size); 2632 bp->b_flags &= ~B_DONE; 2633 } 2634 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2635 BUF_ASSERT_HELD(bp); 2636 KASSERT(bp->b_bufobj == bo, 2637 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2638 return (bp); 2639 } 2640 2641 /* 2642 * Get an empty, disassociated buffer of given size. The buffer is initially 2643 * set to B_INVAL. 2644 */ 2645 struct buf * 2646 geteblk(int size) 2647 { 2648 struct buf *bp; 2649 int maxsize; 2650 2651 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2652 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2653 continue; 2654 allocbuf(bp, size); 2655 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2656 BUF_ASSERT_HELD(bp); 2657 return (bp); 2658 } 2659 2660 2661 /* 2662 * This code constitutes the buffer memory from either anonymous system 2663 * memory (in the case of non-VMIO operations) or from an associated 2664 * VM object (in the case of VMIO operations). This code is able to 2665 * resize a buffer up or down. 2666 * 2667 * Note that this code is tricky, and has many complications to resolve 2668 * deadlock or inconsistant data situations. Tread lightly!!! 2669 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2670 * the caller. Calling this code willy nilly can result in the loss of data. 2671 * 2672 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2673 * B_CACHE for the non-VMIO case. 2674 */ 2675 2676 int 2677 allocbuf(struct buf *bp, int size) 2678 { 2679 int newbsize, mbsize; 2680 int i; 2681 2682 BUF_ASSERT_HELD(bp); 2683 2684 if (bp->b_kvasize < size) 2685 panic("allocbuf: buffer too small"); 2686 2687 if ((bp->b_flags & B_VMIO) == 0) { 2688 caddr_t origbuf; 2689 int origbufsize; 2690 /* 2691 * Just get anonymous memory from the kernel. Don't 2692 * mess with B_CACHE. 2693 */ 2694 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2695 if (bp->b_flags & B_MALLOC) 2696 newbsize = mbsize; 2697 else 2698 newbsize = round_page(size); 2699 2700 if (newbsize < bp->b_bufsize) { 2701 /* 2702 * malloced buffers are not shrunk 2703 */ 2704 if (bp->b_flags & B_MALLOC) { 2705 if (newbsize) { 2706 bp->b_bcount = size; 2707 } else { 2708 free(bp->b_data, M_BIOBUF); 2709 if (bp->b_bufsize) { 2710 atomic_subtract_int( 2711 &bufmallocspace, 2712 bp->b_bufsize); 2713 bufspacewakeup(); 2714 bp->b_bufsize = 0; 2715 } 2716 bp->b_saveaddr = bp->b_kvabase; 2717 bp->b_data = bp->b_saveaddr; 2718 bp->b_bcount = 0; 2719 bp->b_flags &= ~B_MALLOC; 2720 } 2721 return 1; 2722 } 2723 vm_hold_free_pages( 2724 bp, 2725 (vm_offset_t) bp->b_data + newbsize, 2726 (vm_offset_t) bp->b_data + bp->b_bufsize); 2727 } else if (newbsize > bp->b_bufsize) { 2728 /* 2729 * We only use malloced memory on the first allocation. 2730 * and revert to page-allocated memory when the buffer 2731 * grows. 2732 */ 2733 /* 2734 * There is a potential smp race here that could lead 2735 * to bufmallocspace slightly passing the max. It 2736 * is probably extremely rare and not worth worrying 2737 * over. 2738 */ 2739 if ( (bufmallocspace < maxbufmallocspace) && 2740 (bp->b_bufsize == 0) && 2741 (mbsize <= PAGE_SIZE/2)) { 2742 2743 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2744 bp->b_bufsize = mbsize; 2745 bp->b_bcount = size; 2746 bp->b_flags |= B_MALLOC; 2747 atomic_add_int(&bufmallocspace, mbsize); 2748 return 1; 2749 } 2750 origbuf = NULL; 2751 origbufsize = 0; 2752 /* 2753 * If the buffer is growing on its other-than-first allocation, 2754 * then we revert to the page-allocation scheme. 2755 */ 2756 if (bp->b_flags & B_MALLOC) { 2757 origbuf = bp->b_data; 2758 origbufsize = bp->b_bufsize; 2759 bp->b_data = bp->b_kvabase; 2760 if (bp->b_bufsize) { 2761 atomic_subtract_int(&bufmallocspace, 2762 bp->b_bufsize); 2763 bufspacewakeup(); 2764 bp->b_bufsize = 0; 2765 } 2766 bp->b_flags &= ~B_MALLOC; 2767 newbsize = round_page(newbsize); 2768 } 2769 vm_hold_load_pages( 2770 bp, 2771 (vm_offset_t) bp->b_data + bp->b_bufsize, 2772 (vm_offset_t) bp->b_data + newbsize); 2773 if (origbuf) { 2774 bcopy(origbuf, bp->b_data, origbufsize); 2775 free(origbuf, M_BIOBUF); 2776 } 2777 } 2778 } else { 2779 int desiredpages; 2780 2781 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2782 desiredpages = (size == 0) ? 0 : 2783 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2784 2785 if (bp->b_flags & B_MALLOC) 2786 panic("allocbuf: VMIO buffer can't be malloced"); 2787 /* 2788 * Set B_CACHE initially if buffer is 0 length or will become 2789 * 0-length. 2790 */ 2791 if (size == 0 || bp->b_bufsize == 0) 2792 bp->b_flags |= B_CACHE; 2793 2794 if (newbsize < bp->b_bufsize) { 2795 /* 2796 * DEV_BSIZE aligned new buffer size is less then the 2797 * DEV_BSIZE aligned existing buffer size. Figure out 2798 * if we have to remove any pages. 2799 */ 2800 if (desiredpages < bp->b_npages) { 2801 vm_page_t m; 2802 2803 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2804 vm_page_lock_queues(); 2805 for (i = desiredpages; i < bp->b_npages; i++) { 2806 /* 2807 * the page is not freed here -- it 2808 * is the responsibility of 2809 * vnode_pager_setsize 2810 */ 2811 m = bp->b_pages[i]; 2812 KASSERT(m != bogus_page, 2813 ("allocbuf: bogus page found")); 2814 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2815 vm_page_lock_queues(); 2816 2817 bp->b_pages[i] = NULL; 2818 vm_page_unwire(m, 0); 2819 } 2820 vm_page_unlock_queues(); 2821 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2822 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2823 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2824 bp->b_npages = desiredpages; 2825 } 2826 } else if (size > bp->b_bcount) { 2827 /* 2828 * We are growing the buffer, possibly in a 2829 * byte-granular fashion. 2830 */ 2831 struct vnode *vp; 2832 vm_object_t obj; 2833 vm_offset_t toff; 2834 vm_offset_t tinc; 2835 2836 /* 2837 * Step 1, bring in the VM pages from the object, 2838 * allocating them if necessary. We must clear 2839 * B_CACHE if these pages are not valid for the 2840 * range covered by the buffer. 2841 */ 2842 2843 vp = bp->b_vp; 2844 obj = bp->b_bufobj->bo_object; 2845 2846 VM_OBJECT_LOCK(obj); 2847 while (bp->b_npages < desiredpages) { 2848 vm_page_t m; 2849 vm_pindex_t pi; 2850 2851 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2852 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2853 /* 2854 * note: must allocate system pages 2855 * since blocking here could intefere 2856 * with paging I/O, no matter which 2857 * process we are. 2858 */ 2859 m = vm_page_alloc(obj, pi, 2860 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2861 VM_ALLOC_WIRED); 2862 if (m == NULL) { 2863 atomic_add_int(&vm_pageout_deficit, 2864 desiredpages - bp->b_npages); 2865 VM_OBJECT_UNLOCK(obj); 2866 VM_WAIT; 2867 VM_OBJECT_LOCK(obj); 2868 } else { 2869 if (m->valid == 0) 2870 bp->b_flags &= ~B_CACHE; 2871 bp->b_pages[bp->b_npages] = m; 2872 ++bp->b_npages; 2873 } 2874 continue; 2875 } 2876 2877 /* 2878 * We found a page. If we have to sleep on it, 2879 * retry because it might have gotten freed out 2880 * from under us. 2881 * 2882 * We can only test VPO_BUSY here. Blocking on 2883 * m->busy might lead to a deadlock: 2884 * 2885 * vm_fault->getpages->cluster_read->allocbuf 2886 * 2887 */ 2888 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2889 continue; 2890 2891 /* 2892 * We have a good page. 2893 */ 2894 vm_page_lock_queues(); 2895 vm_page_wire(m); 2896 vm_page_unlock_queues(); 2897 bp->b_pages[bp->b_npages] = m; 2898 ++bp->b_npages; 2899 } 2900 2901 /* 2902 * Step 2. We've loaded the pages into the buffer, 2903 * we have to figure out if we can still have B_CACHE 2904 * set. Note that B_CACHE is set according to the 2905 * byte-granular range ( bcount and size ), new the 2906 * aligned range ( newbsize ). 2907 * 2908 * The VM test is against m->valid, which is DEV_BSIZE 2909 * aligned. Needless to say, the validity of the data 2910 * needs to also be DEV_BSIZE aligned. Note that this 2911 * fails with NFS if the server or some other client 2912 * extends the file's EOF. If our buffer is resized, 2913 * B_CACHE may remain set! XXX 2914 */ 2915 2916 toff = bp->b_bcount; 2917 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2918 2919 while ((bp->b_flags & B_CACHE) && toff < size) { 2920 vm_pindex_t pi; 2921 2922 if (tinc > (size - toff)) 2923 tinc = size - toff; 2924 2925 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2926 PAGE_SHIFT; 2927 2928 vfs_buf_test_cache( 2929 bp, 2930 bp->b_offset, 2931 toff, 2932 tinc, 2933 bp->b_pages[pi] 2934 ); 2935 toff += tinc; 2936 tinc = PAGE_SIZE; 2937 } 2938 VM_OBJECT_UNLOCK(obj); 2939 2940 /* 2941 * Step 3, fixup the KVM pmap. Remember that 2942 * bp->b_data is relative to bp->b_offset, but 2943 * bp->b_offset may be offset into the first page. 2944 */ 2945 2946 bp->b_data = (caddr_t) 2947 trunc_page((vm_offset_t)bp->b_data); 2948 pmap_qenter( 2949 (vm_offset_t)bp->b_data, 2950 bp->b_pages, 2951 bp->b_npages 2952 ); 2953 2954 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2955 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2956 } 2957 } 2958 if (newbsize < bp->b_bufsize) 2959 bufspacewakeup(); 2960 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2961 bp->b_bcount = size; /* requested buffer size */ 2962 return 1; 2963 } 2964 2965 void 2966 biodone(struct bio *bp) 2967 { 2968 struct mtx *mtxp; 2969 void (*done)(struct bio *); 2970 2971 mtxp = mtx_pool_find(mtxpool_sleep, bp); 2972 mtx_lock(mtxp); 2973 bp->bio_flags |= BIO_DONE; 2974 done = bp->bio_done; 2975 if (done == NULL) 2976 wakeup(bp); 2977 mtx_unlock(mtxp); 2978 if (done != NULL) 2979 done(bp); 2980 } 2981 2982 /* 2983 * Wait for a BIO to finish. 2984 * 2985 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2986 * case is not yet clear. 2987 */ 2988 int 2989 biowait(struct bio *bp, const char *wchan) 2990 { 2991 struct mtx *mtxp; 2992 2993 mtxp = mtx_pool_find(mtxpool_sleep, bp); 2994 mtx_lock(mtxp); 2995 while ((bp->bio_flags & BIO_DONE) == 0) 2996 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 2997 mtx_unlock(mtxp); 2998 if (bp->bio_error != 0) 2999 return (bp->bio_error); 3000 if (!(bp->bio_flags & BIO_ERROR)) 3001 return (0); 3002 return (EIO); 3003 } 3004 3005 void 3006 biofinish(struct bio *bp, struct devstat *stat, int error) 3007 { 3008 3009 if (error) { 3010 bp->bio_error = error; 3011 bp->bio_flags |= BIO_ERROR; 3012 } 3013 if (stat != NULL) 3014 devstat_end_transaction_bio(stat, bp); 3015 biodone(bp); 3016 } 3017 3018 /* 3019 * bufwait: 3020 * 3021 * Wait for buffer I/O completion, returning error status. The buffer 3022 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3023 * error and cleared. 3024 */ 3025 int 3026 bufwait(struct buf *bp) 3027 { 3028 if (bp->b_iocmd == BIO_READ) 3029 bwait(bp, PRIBIO, "biord"); 3030 else 3031 bwait(bp, PRIBIO, "biowr"); 3032 if (bp->b_flags & B_EINTR) { 3033 bp->b_flags &= ~B_EINTR; 3034 return (EINTR); 3035 } 3036 if (bp->b_ioflags & BIO_ERROR) { 3037 return (bp->b_error ? bp->b_error : EIO); 3038 } else { 3039 return (0); 3040 } 3041 } 3042 3043 /* 3044 * Call back function from struct bio back up to struct buf. 3045 */ 3046 static void 3047 bufdonebio(struct bio *bip) 3048 { 3049 struct buf *bp; 3050 3051 bp = bip->bio_caller2; 3052 bp->b_resid = bp->b_bcount - bip->bio_completed; 3053 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3054 bp->b_ioflags = bip->bio_flags; 3055 bp->b_error = bip->bio_error; 3056 if (bp->b_error) 3057 bp->b_ioflags |= BIO_ERROR; 3058 bufdone(bp); 3059 g_destroy_bio(bip); 3060 } 3061 3062 void 3063 dev_strategy(struct cdev *dev, struct buf *bp) 3064 { 3065 struct cdevsw *csw; 3066 struct bio *bip; 3067 3068 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3069 panic("b_iocmd botch"); 3070 for (;;) { 3071 bip = g_new_bio(); 3072 if (bip != NULL) 3073 break; 3074 /* Try again later */ 3075 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3076 } 3077 bip->bio_cmd = bp->b_iocmd; 3078 bip->bio_offset = bp->b_iooffset; 3079 bip->bio_length = bp->b_bcount; 3080 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3081 bip->bio_data = bp->b_data; 3082 bip->bio_done = bufdonebio; 3083 bip->bio_caller2 = bp; 3084 bip->bio_dev = dev; 3085 KASSERT(dev->si_refcount > 0, 3086 ("dev_strategy on un-referenced struct cdev *(%s)", 3087 devtoname(dev))); 3088 csw = dev_refthread(dev); 3089 if (csw == NULL) { 3090 g_destroy_bio(bip); 3091 bp->b_error = ENXIO; 3092 bp->b_ioflags = BIO_ERROR; 3093 bufdone(bp); 3094 return; 3095 } 3096 (*csw->d_strategy)(bip); 3097 dev_relthread(dev); 3098 } 3099 3100 /* 3101 * bufdone: 3102 * 3103 * Finish I/O on a buffer, optionally calling a completion function. 3104 * This is usually called from an interrupt so process blocking is 3105 * not allowed. 3106 * 3107 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3108 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3109 * assuming B_INVAL is clear. 3110 * 3111 * For the VMIO case, we set B_CACHE if the op was a read and no 3112 * read error occured, or if the op was a write. B_CACHE is never 3113 * set if the buffer is invalid or otherwise uncacheable. 3114 * 3115 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3116 * initiator to leave B_INVAL set to brelse the buffer out of existance 3117 * in the biodone routine. 3118 */ 3119 void 3120 bufdone(struct buf *bp) 3121 { 3122 struct bufobj *dropobj; 3123 void (*biodone)(struct buf *); 3124 3125 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3126 dropobj = NULL; 3127 3128 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3129 BUF_ASSERT_HELD(bp); 3130 3131 runningbufwakeup(bp); 3132 if (bp->b_iocmd == BIO_WRITE) 3133 dropobj = bp->b_bufobj; 3134 /* call optional completion function if requested */ 3135 if (bp->b_iodone != NULL) { 3136 biodone = bp->b_iodone; 3137 bp->b_iodone = NULL; 3138 (*biodone) (bp); 3139 if (dropobj) 3140 bufobj_wdrop(dropobj); 3141 return; 3142 } 3143 3144 bufdone_finish(bp); 3145 3146 if (dropobj) 3147 bufobj_wdrop(dropobj); 3148 } 3149 3150 void 3151 bufdone_finish(struct buf *bp) 3152 { 3153 BUF_ASSERT_HELD(bp); 3154 3155 if (!LIST_EMPTY(&bp->b_dep)) 3156 buf_complete(bp); 3157 3158 if (bp->b_flags & B_VMIO) { 3159 int i; 3160 vm_ooffset_t foff; 3161 vm_page_t m; 3162 vm_object_t obj; 3163 int iosize; 3164 struct vnode *vp = bp->b_vp; 3165 boolean_t are_queues_locked; 3166 3167 obj = bp->b_bufobj->bo_object; 3168 3169 #if defined(VFS_BIO_DEBUG) 3170 mp_fixme("usecount and vflag accessed without locks."); 3171 if (vp->v_usecount == 0) { 3172 panic("biodone: zero vnode ref count"); 3173 } 3174 3175 KASSERT(vp->v_object != NULL, 3176 ("biodone: vnode %p has no vm_object", vp)); 3177 #endif 3178 3179 foff = bp->b_offset; 3180 KASSERT(bp->b_offset != NOOFFSET, 3181 ("biodone: no buffer offset")); 3182 3183 VM_OBJECT_LOCK(obj); 3184 #if defined(VFS_BIO_DEBUG) 3185 if (obj->paging_in_progress < bp->b_npages) { 3186 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3187 obj->paging_in_progress, bp->b_npages); 3188 } 3189 #endif 3190 3191 /* 3192 * Set B_CACHE if the op was a normal read and no error 3193 * occured. B_CACHE is set for writes in the b*write() 3194 * routines. 3195 */ 3196 iosize = bp->b_bcount - bp->b_resid; 3197 if (bp->b_iocmd == BIO_READ && 3198 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3199 !(bp->b_ioflags & BIO_ERROR)) { 3200 bp->b_flags |= B_CACHE; 3201 } 3202 if (bp->b_iocmd == BIO_READ) { 3203 vm_page_lock_queues(); 3204 are_queues_locked = TRUE; 3205 } else 3206 are_queues_locked = FALSE; 3207 for (i = 0; i < bp->b_npages; i++) { 3208 int bogusflag = 0; 3209 int resid; 3210 3211 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3212 if (resid > iosize) 3213 resid = iosize; 3214 3215 /* 3216 * cleanup bogus pages, restoring the originals 3217 */ 3218 m = bp->b_pages[i]; 3219 if (m == bogus_page) { 3220 bogusflag = 1; 3221 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3222 if (m == NULL) 3223 panic("biodone: page disappeared!"); 3224 bp->b_pages[i] = m; 3225 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3226 bp->b_pages, bp->b_npages); 3227 } 3228 #if defined(VFS_BIO_DEBUG) 3229 if (OFF_TO_IDX(foff) != m->pindex) { 3230 printf( 3231 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3232 (intmax_t)foff, (uintmax_t)m->pindex); 3233 } 3234 #endif 3235 3236 /* 3237 * In the write case, the valid and clean bits are 3238 * already changed correctly ( see bdwrite() ), so we 3239 * only need to do this here in the read case. 3240 */ 3241 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3242 vfs_page_set_valid(bp, foff, m); 3243 } 3244 3245 /* 3246 * when debugging new filesystems or buffer I/O methods, this 3247 * is the most common error that pops up. if you see this, you 3248 * have not set the page busy flag correctly!!! 3249 */ 3250 if (m->busy == 0) { 3251 printf("biodone: page busy < 0, " 3252 "pindex: %d, foff: 0x(%x,%x), " 3253 "resid: %d, index: %d\n", 3254 (int) m->pindex, (int)(foff >> 32), 3255 (int) foff & 0xffffffff, resid, i); 3256 if (!vn_isdisk(vp, NULL)) 3257 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3258 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3259 (intmax_t) bp->b_lblkno, 3260 bp->b_flags, bp->b_npages); 3261 else 3262 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3263 (intmax_t) bp->b_lblkno, 3264 bp->b_flags, bp->b_npages); 3265 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3266 (u_long)m->valid, (u_long)m->dirty, 3267 m->wire_count); 3268 panic("biodone: page busy < 0\n"); 3269 } 3270 vm_page_io_finish(m); 3271 vm_object_pip_subtract(obj, 1); 3272 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3273 iosize -= resid; 3274 } 3275 if (are_queues_locked) 3276 vm_page_unlock_queues(); 3277 vm_object_pip_wakeupn(obj, 0); 3278 VM_OBJECT_UNLOCK(obj); 3279 } 3280 3281 /* 3282 * For asynchronous completions, release the buffer now. The brelse 3283 * will do a wakeup there if necessary - so no need to do a wakeup 3284 * here in the async case. The sync case always needs to do a wakeup. 3285 */ 3286 3287 if (bp->b_flags & B_ASYNC) { 3288 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3289 brelse(bp); 3290 else 3291 bqrelse(bp); 3292 } else 3293 bdone(bp); 3294 } 3295 3296 /* 3297 * This routine is called in lieu of iodone in the case of 3298 * incomplete I/O. This keeps the busy status for pages 3299 * consistant. 3300 */ 3301 void 3302 vfs_unbusy_pages(struct buf *bp) 3303 { 3304 int i; 3305 vm_object_t obj; 3306 vm_page_t m; 3307 3308 runningbufwakeup(bp); 3309 if (!(bp->b_flags & B_VMIO)) 3310 return; 3311 3312 obj = bp->b_bufobj->bo_object; 3313 VM_OBJECT_LOCK(obj); 3314 for (i = 0; i < bp->b_npages; i++) { 3315 m = bp->b_pages[i]; 3316 if (m == bogus_page) { 3317 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3318 if (!m) 3319 panic("vfs_unbusy_pages: page missing\n"); 3320 bp->b_pages[i] = m; 3321 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3322 bp->b_pages, bp->b_npages); 3323 } 3324 vm_object_pip_subtract(obj, 1); 3325 vm_page_io_finish(m); 3326 } 3327 vm_object_pip_wakeupn(obj, 0); 3328 VM_OBJECT_UNLOCK(obj); 3329 } 3330 3331 /* 3332 * vfs_page_set_valid: 3333 * 3334 * Set the valid bits in a page based on the supplied offset. The 3335 * range is restricted to the buffer's size. 3336 * 3337 * This routine is typically called after a read completes. 3338 */ 3339 static void 3340 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3341 { 3342 vm_ooffset_t soff, eoff; 3343 3344 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3345 /* 3346 * Start and end offsets in buffer. eoff - soff may not cross a 3347 * page boundry or cross the end of the buffer. The end of the 3348 * buffer, in this case, is our file EOF, not the allocation size 3349 * of the buffer. 3350 */ 3351 soff = off; 3352 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3353 if (eoff > bp->b_offset + bp->b_bcount) 3354 eoff = bp->b_offset + bp->b_bcount; 3355 3356 /* 3357 * Set valid range. This is typically the entire buffer and thus the 3358 * entire page. 3359 */ 3360 if (eoff > soff) { 3361 vm_page_set_validclean( 3362 m, 3363 (vm_offset_t) (soff & PAGE_MASK), 3364 (vm_offset_t) (eoff - soff) 3365 ); 3366 } 3367 } 3368 3369 /* 3370 * This routine is called before a device strategy routine. 3371 * It is used to tell the VM system that paging I/O is in 3372 * progress, and treat the pages associated with the buffer 3373 * almost as being VPO_BUSY. Also the object paging_in_progress 3374 * flag is handled to make sure that the object doesn't become 3375 * inconsistant. 3376 * 3377 * Since I/O has not been initiated yet, certain buffer flags 3378 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3379 * and should be ignored. 3380 */ 3381 void 3382 vfs_busy_pages(struct buf *bp, int clear_modify) 3383 { 3384 int i, bogus; 3385 vm_object_t obj; 3386 vm_ooffset_t foff; 3387 vm_page_t m; 3388 3389 if (!(bp->b_flags & B_VMIO)) 3390 return; 3391 3392 obj = bp->b_bufobj->bo_object; 3393 foff = bp->b_offset; 3394 KASSERT(bp->b_offset != NOOFFSET, 3395 ("vfs_busy_pages: no buffer offset")); 3396 VM_OBJECT_LOCK(obj); 3397 if (bp->b_bufsize != 0) 3398 vfs_setdirty_locked_object(bp); 3399 retry: 3400 for (i = 0; i < bp->b_npages; i++) { 3401 m = bp->b_pages[i]; 3402 3403 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3404 goto retry; 3405 } 3406 bogus = 0; 3407 vm_page_lock_queues(); 3408 for (i = 0; i < bp->b_npages; i++) { 3409 m = bp->b_pages[i]; 3410 3411 if ((bp->b_flags & B_CLUSTER) == 0) { 3412 vm_object_pip_add(obj, 1); 3413 vm_page_io_start(m); 3414 } 3415 /* 3416 * When readying a buffer for a read ( i.e 3417 * clear_modify == 0 ), it is important to do 3418 * bogus_page replacement for valid pages in 3419 * partially instantiated buffers. Partially 3420 * instantiated buffers can, in turn, occur when 3421 * reconstituting a buffer from its VM backing store 3422 * base. We only have to do this if B_CACHE is 3423 * clear ( which causes the I/O to occur in the 3424 * first place ). The replacement prevents the read 3425 * I/O from overwriting potentially dirty VM-backed 3426 * pages. XXX bogus page replacement is, uh, bogus. 3427 * It may not work properly with small-block devices. 3428 * We need to find a better way. 3429 */ 3430 pmap_remove_all(m); 3431 if (clear_modify) 3432 vfs_page_set_valid(bp, foff, m); 3433 else if (m->valid == VM_PAGE_BITS_ALL && 3434 (bp->b_flags & B_CACHE) == 0) { 3435 bp->b_pages[i] = bogus_page; 3436 bogus++; 3437 } 3438 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3439 } 3440 vm_page_unlock_queues(); 3441 VM_OBJECT_UNLOCK(obj); 3442 if (bogus) 3443 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3444 bp->b_pages, bp->b_npages); 3445 } 3446 3447 /* 3448 * Tell the VM system that the pages associated with this buffer 3449 * are clean. This is used for delayed writes where the data is 3450 * going to go to disk eventually without additional VM intevention. 3451 * 3452 * Note that while we only really need to clean through to b_bcount, we 3453 * just go ahead and clean through to b_bufsize. 3454 */ 3455 static void 3456 vfs_clean_pages(struct buf *bp) 3457 { 3458 int i; 3459 vm_ooffset_t foff, noff, eoff; 3460 vm_page_t m; 3461 3462 if (!(bp->b_flags & B_VMIO)) 3463 return; 3464 3465 foff = bp->b_offset; 3466 KASSERT(bp->b_offset != NOOFFSET, 3467 ("vfs_clean_pages: no buffer offset")); 3468 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3469 vm_page_lock_queues(); 3470 for (i = 0; i < bp->b_npages; i++) { 3471 m = bp->b_pages[i]; 3472 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3473 eoff = noff; 3474 3475 if (eoff > bp->b_offset + bp->b_bufsize) 3476 eoff = bp->b_offset + bp->b_bufsize; 3477 vfs_page_set_valid(bp, foff, m); 3478 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3479 foff = noff; 3480 } 3481 vm_page_unlock_queues(); 3482 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3483 } 3484 3485 /* 3486 * vfs_bio_set_validclean: 3487 * 3488 * Set the range within the buffer to valid and clean. The range is 3489 * relative to the beginning of the buffer, b_offset. Note that b_offset 3490 * itself may be offset from the beginning of the first page. 3491 * 3492 */ 3493 3494 void 3495 vfs_bio_set_validclean(struct buf *bp, int base, int size) 3496 { 3497 int i, n; 3498 vm_page_t m; 3499 3500 if (!(bp->b_flags & B_VMIO)) 3501 return; 3502 /* 3503 * Fixup base to be relative to beginning of first page. 3504 * Set initial n to be the maximum number of bytes in the 3505 * first page that can be validated. 3506 */ 3507 3508 base += (bp->b_offset & PAGE_MASK); 3509 n = PAGE_SIZE - (base & PAGE_MASK); 3510 3511 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3512 vm_page_lock_queues(); 3513 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3514 m = bp->b_pages[i]; 3515 if (n > size) 3516 n = size; 3517 vm_page_set_validclean(m, base & PAGE_MASK, n); 3518 base += n; 3519 size -= n; 3520 n = PAGE_SIZE; 3521 } 3522 vm_page_unlock_queues(); 3523 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3524 } 3525 3526 /* 3527 * vfs_bio_clrbuf: 3528 * 3529 * clear a buffer. This routine essentially fakes an I/O, so we need 3530 * to clear BIO_ERROR and B_INVAL. 3531 * 3532 * Note that while we only theoretically need to clear through b_bcount, 3533 * we go ahead and clear through b_bufsize. 3534 */ 3535 3536 void 3537 vfs_bio_clrbuf(struct buf *bp) 3538 { 3539 int i, j, mask = 0; 3540 caddr_t sa, ea; 3541 3542 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3543 clrbuf(bp); 3544 return; 3545 } 3546 3547 bp->b_flags &= ~B_INVAL; 3548 bp->b_ioflags &= ~BIO_ERROR; 3549 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3550 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3551 (bp->b_offset & PAGE_MASK) == 0) { 3552 if (bp->b_pages[0] == bogus_page) 3553 goto unlock; 3554 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3555 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3556 if ((bp->b_pages[0]->valid & mask) == mask) 3557 goto unlock; 3558 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3559 ((bp->b_pages[0]->valid & mask) == 0)) { 3560 bzero(bp->b_data, bp->b_bufsize); 3561 bp->b_pages[0]->valid |= mask; 3562 goto unlock; 3563 } 3564 } 3565 ea = sa = bp->b_data; 3566 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3567 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3568 ea = (caddr_t)(vm_offset_t)ulmin( 3569 (u_long)(vm_offset_t)ea, 3570 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3571 if (bp->b_pages[i] == bogus_page) 3572 continue; 3573 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3574 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3575 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3576 if ((bp->b_pages[i]->valid & mask) == mask) 3577 continue; 3578 if ((bp->b_pages[i]->valid & mask) == 0) { 3579 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3580 bzero(sa, ea - sa); 3581 } else { 3582 for (; sa < ea; sa += DEV_BSIZE, j++) { 3583 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3584 (bp->b_pages[i]->valid & (1 << j)) == 0) 3585 bzero(sa, DEV_BSIZE); 3586 } 3587 } 3588 bp->b_pages[i]->valid |= mask; 3589 } 3590 unlock: 3591 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3592 bp->b_resid = 0; 3593 } 3594 3595 /* 3596 * vm_hold_load_pages and vm_hold_free_pages get pages into 3597 * a buffers address space. The pages are anonymous and are 3598 * not associated with a file object. 3599 */ 3600 static void 3601 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3602 { 3603 vm_offset_t pg; 3604 vm_page_t p; 3605 int index; 3606 3607 to = round_page(to); 3608 from = round_page(from); 3609 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3610 3611 VM_OBJECT_LOCK(kernel_object); 3612 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3613 tryagain: 3614 /* 3615 * note: must allocate system pages since blocking here 3616 * could intefere with paging I/O, no matter which 3617 * process we are. 3618 */ 3619 p = vm_page_alloc(kernel_object, 3620 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3621 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3622 if (!p) { 3623 atomic_add_int(&vm_pageout_deficit, 3624 (to - pg) >> PAGE_SHIFT); 3625 VM_OBJECT_UNLOCK(kernel_object); 3626 VM_WAIT; 3627 VM_OBJECT_LOCK(kernel_object); 3628 goto tryagain; 3629 } 3630 p->valid = VM_PAGE_BITS_ALL; 3631 pmap_qenter(pg, &p, 1); 3632 bp->b_pages[index] = p; 3633 } 3634 VM_OBJECT_UNLOCK(kernel_object); 3635 bp->b_npages = index; 3636 } 3637 3638 /* Return pages associated with this buf to the vm system */ 3639 static void 3640 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3641 { 3642 vm_offset_t pg; 3643 vm_page_t p; 3644 int index, newnpages; 3645 3646 from = round_page(from); 3647 to = round_page(to); 3648 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3649 3650 VM_OBJECT_LOCK(kernel_object); 3651 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3652 p = bp->b_pages[index]; 3653 if (p && (index < bp->b_npages)) { 3654 if (p->busy) { 3655 printf( 3656 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3657 (intmax_t)bp->b_blkno, 3658 (intmax_t)bp->b_lblkno); 3659 } 3660 bp->b_pages[index] = NULL; 3661 pmap_qremove(pg, 1); 3662 vm_page_lock_queues(); 3663 vm_page_unwire(p, 0); 3664 vm_page_free(p); 3665 vm_page_unlock_queues(); 3666 } 3667 } 3668 VM_OBJECT_UNLOCK(kernel_object); 3669 bp->b_npages = newnpages; 3670 } 3671 3672 /* 3673 * Map an IO request into kernel virtual address space. 3674 * 3675 * All requests are (re)mapped into kernel VA space. 3676 * Notice that we use b_bufsize for the size of the buffer 3677 * to be mapped. b_bcount might be modified by the driver. 3678 * 3679 * Note that even if the caller determines that the address space should 3680 * be valid, a race or a smaller-file mapped into a larger space may 3681 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3682 * check the return value. 3683 */ 3684 int 3685 vmapbuf(struct buf *bp) 3686 { 3687 caddr_t addr, kva; 3688 vm_prot_t prot; 3689 int pidx, i; 3690 struct vm_page *m; 3691 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3692 3693 if (bp->b_bufsize < 0) 3694 return (-1); 3695 prot = VM_PROT_READ; 3696 if (bp->b_iocmd == BIO_READ) 3697 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3698 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3699 addr < bp->b_data + bp->b_bufsize; 3700 addr += PAGE_SIZE, pidx++) { 3701 /* 3702 * Do the vm_fault if needed; do the copy-on-write thing 3703 * when reading stuff off device into memory. 3704 * 3705 * NOTE! Must use pmap_extract() because addr may be in 3706 * the userland address space, and kextract is only guarenteed 3707 * to work for the kernland address space (see: sparc64 port). 3708 */ 3709 retry: 3710 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3711 prot) < 0) { 3712 vm_page_lock_queues(); 3713 for (i = 0; i < pidx; ++i) { 3714 vm_page_unhold(bp->b_pages[i]); 3715 bp->b_pages[i] = NULL; 3716 } 3717 vm_page_unlock_queues(); 3718 return(-1); 3719 } 3720 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3721 if (m == NULL) 3722 goto retry; 3723 bp->b_pages[pidx] = m; 3724 } 3725 if (pidx > btoc(MAXPHYS)) 3726 panic("vmapbuf: mapped more than MAXPHYS"); 3727 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3728 3729 kva = bp->b_saveaddr; 3730 bp->b_npages = pidx; 3731 bp->b_saveaddr = bp->b_data; 3732 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3733 return(0); 3734 } 3735 3736 /* 3737 * Free the io map PTEs associated with this IO operation. 3738 * We also invalidate the TLB entries and restore the original b_addr. 3739 */ 3740 void 3741 vunmapbuf(struct buf *bp) 3742 { 3743 int pidx; 3744 int npages; 3745 3746 npages = bp->b_npages; 3747 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3748 vm_page_lock_queues(); 3749 for (pidx = 0; pidx < npages; pidx++) 3750 vm_page_unhold(bp->b_pages[pidx]); 3751 vm_page_unlock_queues(); 3752 3753 bp->b_data = bp->b_saveaddr; 3754 } 3755 3756 void 3757 bdone(struct buf *bp) 3758 { 3759 struct mtx *mtxp; 3760 3761 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3762 mtx_lock(mtxp); 3763 bp->b_flags |= B_DONE; 3764 wakeup(bp); 3765 mtx_unlock(mtxp); 3766 } 3767 3768 void 3769 bwait(struct buf *bp, u_char pri, const char *wchan) 3770 { 3771 struct mtx *mtxp; 3772 3773 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3774 mtx_lock(mtxp); 3775 while ((bp->b_flags & B_DONE) == 0) 3776 msleep(bp, mtxp, pri, wchan, 0); 3777 mtx_unlock(mtxp); 3778 } 3779 3780 int 3781 bufsync(struct bufobj *bo, int waitfor) 3782 { 3783 3784 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 3785 } 3786 3787 void 3788 bufstrategy(struct bufobj *bo, struct buf *bp) 3789 { 3790 int i = 0; 3791 struct vnode *vp; 3792 3793 vp = bp->b_vp; 3794 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3795 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3796 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3797 i = VOP_STRATEGY(vp, bp); 3798 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3799 } 3800 3801 void 3802 bufobj_wrefl(struct bufobj *bo) 3803 { 3804 3805 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3806 ASSERT_BO_LOCKED(bo); 3807 bo->bo_numoutput++; 3808 } 3809 3810 void 3811 bufobj_wref(struct bufobj *bo) 3812 { 3813 3814 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3815 BO_LOCK(bo); 3816 bo->bo_numoutput++; 3817 BO_UNLOCK(bo); 3818 } 3819 3820 void 3821 bufobj_wdrop(struct bufobj *bo) 3822 { 3823 3824 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3825 BO_LOCK(bo); 3826 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3827 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3828 bo->bo_flag &= ~BO_WWAIT; 3829 wakeup(&bo->bo_numoutput); 3830 } 3831 BO_UNLOCK(bo); 3832 } 3833 3834 int 3835 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3836 { 3837 int error; 3838 3839 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3840 ASSERT_BO_LOCKED(bo); 3841 error = 0; 3842 while (bo->bo_numoutput) { 3843 bo->bo_flag |= BO_WWAIT; 3844 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3845 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3846 if (error) 3847 break; 3848 } 3849 return (error); 3850 } 3851 3852 void 3853 bpin(struct buf *bp) 3854 { 3855 struct mtx *mtxp; 3856 3857 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3858 mtx_lock(mtxp); 3859 bp->b_pin_count++; 3860 mtx_unlock(mtxp); 3861 } 3862 3863 void 3864 bunpin(struct buf *bp) 3865 { 3866 struct mtx *mtxp; 3867 3868 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3869 mtx_lock(mtxp); 3870 if (--bp->b_pin_count == 0) 3871 wakeup(bp); 3872 mtx_unlock(mtxp); 3873 } 3874 3875 void 3876 bunpin_wait(struct buf *bp) 3877 { 3878 struct mtx *mtxp; 3879 3880 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3881 mtx_lock(mtxp); 3882 while (bp->b_pin_count > 0) 3883 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 3884 mtx_unlock(mtxp); 3885 } 3886 3887 #include "opt_ddb.h" 3888 #ifdef DDB 3889 #include <ddb/ddb.h> 3890 3891 /* DDB command to show buffer data */ 3892 DB_SHOW_COMMAND(buffer, db_show_buffer) 3893 { 3894 /* get args */ 3895 struct buf *bp = (struct buf *)addr; 3896 3897 if (!have_addr) { 3898 db_printf("usage: show buffer <addr>\n"); 3899 return; 3900 } 3901 3902 db_printf("buf at %p\n", bp); 3903 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3904 db_printf( 3905 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3906 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n", 3907 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3908 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 3909 bp->b_dep.lh_first); 3910 if (bp->b_npages) { 3911 int i; 3912 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3913 for (i = 0; i < bp->b_npages; i++) { 3914 vm_page_t m; 3915 m = bp->b_pages[i]; 3916 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3917 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3918 if ((i + 1) < bp->b_npages) 3919 db_printf(","); 3920 } 3921 db_printf("\n"); 3922 } 3923 lockmgr_printinfo(&bp->b_lock); 3924 } 3925 3926 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 3927 { 3928 struct buf *bp; 3929 int i; 3930 3931 for (i = 0; i < nbuf; i++) { 3932 bp = &buf[i]; 3933 if (BUF_ISLOCKED(bp)) { 3934 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 3935 db_printf("\n"); 3936 } 3937 } 3938 } 3939 3940 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 3941 { 3942 struct vnode *vp; 3943 struct buf *bp; 3944 3945 if (!have_addr) { 3946 db_printf("usage: show vnodebufs <addr>\n"); 3947 return; 3948 } 3949 vp = (struct vnode *)addr; 3950 db_printf("Clean buffers:\n"); 3951 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 3952 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 3953 db_printf("\n"); 3954 } 3955 db_printf("Dirty buffers:\n"); 3956 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 3957 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 3958 db_printf("\n"); 3959 } 3960 } 3961 #endif /* DDB */ 3962