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