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/fail.h> 52 #include <sys/limits.h> 53 #include <sys/lock.h> 54 #include <sys/malloc.h> 55 #include <sys/mount.h> 56 #include <sys/mutex.h> 57 #include <sys/kernel.h> 58 #include <sys/kthread.h> 59 #include <sys/proc.h> 60 #include <sys/resourcevar.h> 61 #include <sys/sysctl.h> 62 #include <sys/vmmeter.h> 63 #include <sys/vnode.h> 64 #include <geom/geom.h> 65 #include <vm/vm.h> 66 #include <vm/vm_param.h> 67 #include <vm/vm_kern.h> 68 #include <vm/vm_pageout.h> 69 #include <vm/vm_page.h> 70 #include <vm/vm_object.h> 71 #include <vm/vm_extern.h> 72 #include <vm/vm_map.h> 73 #include "opt_compat.h" 74 #include "opt_directio.h" 75 #include "opt_swap.h" 76 77 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 78 79 struct bio_ops bioops; /* I/O operation notification */ 80 81 struct buf_ops buf_ops_bio = { 82 .bop_name = "buf_ops_bio", 83 .bop_write = bufwrite, 84 .bop_strategy = bufstrategy, 85 .bop_sync = bufsync, 86 .bop_bdflush = bufbdflush, 87 }; 88 89 /* 90 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 91 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 92 */ 93 struct buf *buf; /* buffer header pool */ 94 95 static struct proc *bufdaemonproc; 96 97 static int inmem(struct vnode *vp, daddr_t blkno); 98 static void vm_hold_free_pages(struct buf *bp, int newbsize); 99 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 100 vm_offset_t to); 101 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 102 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 103 vm_page_t m); 104 static void vfs_drain_busy_pages(struct buf *bp); 105 static void vfs_clean_pages_dirty_buf(struct buf *bp); 106 static void vfs_setdirty_locked_object(struct buf *bp); 107 static void vfs_vmio_release(struct buf *bp); 108 static int vfs_bio_clcheck(struct vnode *vp, int size, 109 daddr_t lblkno, daddr_t blkno); 110 static int buf_do_flush(struct vnode *vp); 111 static int flushbufqueues(struct vnode *, int, int); 112 static void buf_daemon(void); 113 static void bremfreel(struct buf *bp); 114 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 115 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 116 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 117 #endif 118 119 int vmiodirenable = TRUE; 120 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 121 "Use the VM system for directory writes"); 122 long runningbufspace; 123 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 124 "Amount of presently outstanding async buffer io"); 125 static long bufspace; 126 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 127 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 128 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 129 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 130 #else 131 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 132 "Virtual memory used for buffers"); 133 #endif 134 static long maxbufspace; 135 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 136 "Maximum allowed value of bufspace (including buf_daemon)"); 137 static long bufmallocspace; 138 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 139 "Amount of malloced memory for buffers"); 140 static long maxbufmallocspace; 141 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 142 "Maximum amount of malloced memory for buffers"); 143 static long lobufspace; 144 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 145 "Minimum amount of buffers we want to have"); 146 long hibufspace; 147 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 148 "Maximum allowed value of bufspace (excluding buf_daemon)"); 149 static int bufreusecnt; 150 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 151 "Number of times we have reused a buffer"); 152 static int buffreekvacnt; 153 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 154 "Number of times we have freed the KVA space from some buffer"); 155 static int bufdefragcnt; 156 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 157 "Number of times we have had to repeat buffer allocation to defragment"); 158 static long lorunningspace; 159 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 160 "Minimum preferred space used for in-progress I/O"); 161 static long hirunningspace; 162 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 163 "Maximum amount of space to use for in-progress I/O"); 164 int dirtybufferflushes; 165 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 166 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 167 int bdwriteskip; 168 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 169 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 170 int altbufferflushes; 171 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 172 0, "Number of fsync flushes to limit dirty buffers"); 173 static int recursiveflushes; 174 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 175 0, "Number of flushes skipped due to being recursive"); 176 static int numdirtybuffers; 177 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 178 "Number of buffers that are dirty (has unwritten changes) at the moment"); 179 static int lodirtybuffers; 180 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 181 "How many buffers we want to have free before bufdaemon can sleep"); 182 static int hidirtybuffers; 183 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 184 "When the number of dirty buffers is considered severe"); 185 int dirtybufthresh; 186 SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 187 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 188 static int numfreebuffers; 189 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 190 "Number of free buffers"); 191 static int lofreebuffers; 192 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 193 "XXX Unused"); 194 static int hifreebuffers; 195 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 196 "XXX Complicatedly unused"); 197 static int getnewbufcalls; 198 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 199 "Number of calls to getnewbuf"); 200 static int getnewbufrestarts; 201 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 202 "Number of times getnewbuf has had to restart a buffer aquisition"); 203 static int flushbufqtarget = 100; 204 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 205 "Amount of work to do in flushbufqueues when helping bufdaemon"); 206 static long notbufdflashes; 207 SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0, 208 "Number of dirty buffer flushes done by the bufdaemon helpers"); 209 210 /* 211 * Wakeup point for bufdaemon, as well as indicator of whether it is already 212 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 213 * is idling. 214 */ 215 static int bd_request; 216 217 /* 218 * Request for the buf daemon to write more buffers than is indicated by 219 * lodirtybuf. This may be necessary to push out excess dependencies or 220 * defragment the address space where a simple count of the number of dirty 221 * buffers is insufficient to characterize the demand for flushing them. 222 */ 223 static int bd_speedupreq; 224 225 /* 226 * This lock synchronizes access to bd_request. 227 */ 228 static struct mtx bdlock; 229 230 /* 231 * bogus page -- for I/O to/from partially complete buffers 232 * this is a temporary solution to the problem, but it is not 233 * really that bad. it would be better to split the buffer 234 * for input in the case of buffers partially already in memory, 235 * but the code is intricate enough already. 236 */ 237 vm_page_t bogus_page; 238 239 /* 240 * Synchronization (sleep/wakeup) variable for active buffer space requests. 241 * Set when wait starts, cleared prior to wakeup(). 242 * Used in runningbufwakeup() and waitrunningbufspace(). 243 */ 244 static int runningbufreq; 245 246 /* 247 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 248 * waitrunningbufspace(). 249 */ 250 static struct mtx rbreqlock; 251 252 /* 253 * Synchronization (sleep/wakeup) variable for buffer requests. 254 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 255 * by and/or. 256 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 257 * getnewbuf(), and getblk(). 258 */ 259 static int needsbuffer; 260 261 /* 262 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 263 */ 264 static struct mtx nblock; 265 266 /* 267 * Definitions for the buffer free lists. 268 */ 269 #define BUFFER_QUEUES 6 /* number of free buffer queues */ 270 271 #define QUEUE_NONE 0 /* on no queue */ 272 #define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 273 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 274 #define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ 275 #define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 276 #define QUEUE_EMPTY 5 /* empty buffer headers */ 277 #define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 278 279 /* Queues for free buffers with various properties */ 280 static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 281 282 /* Lock for the bufqueues */ 283 static struct mtx bqlock; 284 285 /* 286 * Single global constant for BUF_WMESG, to avoid getting multiple references. 287 * buf_wmesg is referred from macros. 288 */ 289 const char *buf_wmesg = BUF_WMESG; 290 291 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 292 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 293 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 294 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 295 296 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 297 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 298 static int 299 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 300 { 301 long lvalue; 302 int ivalue; 303 304 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 305 return (sysctl_handle_long(oidp, arg1, arg2, req)); 306 lvalue = *(long *)arg1; 307 if (lvalue > INT_MAX) 308 /* On overflow, still write out a long to trigger ENOMEM. */ 309 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 310 ivalue = lvalue; 311 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 312 } 313 #endif 314 315 #ifdef DIRECTIO 316 extern void ffs_rawread_setup(void); 317 #endif /* DIRECTIO */ 318 /* 319 * numdirtywakeup: 320 * 321 * If someone is blocked due to there being too many dirty buffers, 322 * and numdirtybuffers is now reasonable, wake them up. 323 */ 324 325 static __inline void 326 numdirtywakeup(int level) 327 { 328 329 if (numdirtybuffers <= level) { 330 mtx_lock(&nblock); 331 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 332 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 333 wakeup(&needsbuffer); 334 } 335 mtx_unlock(&nblock); 336 } 337 } 338 339 /* 340 * bufspacewakeup: 341 * 342 * Called when buffer space is potentially available for recovery. 343 * getnewbuf() will block on this flag when it is unable to free 344 * sufficient buffer space. Buffer space becomes recoverable when 345 * bp's get placed back in the queues. 346 */ 347 348 static __inline void 349 bufspacewakeup(void) 350 { 351 352 /* 353 * If someone is waiting for BUF space, wake them up. Even 354 * though we haven't freed the kva space yet, the waiting 355 * process will be able to now. 356 */ 357 mtx_lock(&nblock); 358 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 359 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 360 wakeup(&needsbuffer); 361 } 362 mtx_unlock(&nblock); 363 } 364 365 /* 366 * runningbufwakeup() - in-progress I/O accounting. 367 * 368 */ 369 void 370 runningbufwakeup(struct buf *bp) 371 { 372 373 if (bp->b_runningbufspace) { 374 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace); 375 bp->b_runningbufspace = 0; 376 mtx_lock(&rbreqlock); 377 if (runningbufreq && runningbufspace <= lorunningspace) { 378 runningbufreq = 0; 379 wakeup(&runningbufreq); 380 } 381 mtx_unlock(&rbreqlock); 382 } 383 } 384 385 /* 386 * bufcountwakeup: 387 * 388 * Called when a buffer has been added to one of the free queues to 389 * account for the buffer and to wakeup anyone waiting for free buffers. 390 * This typically occurs when large amounts of metadata are being handled 391 * by the buffer cache ( else buffer space runs out first, usually ). 392 */ 393 394 static __inline void 395 bufcountwakeup(struct buf *bp) 396 { 397 int old; 398 399 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 400 ("buf %p already counted as free", bp)); 401 bp->b_vflags |= BV_INFREECNT; 402 old = atomic_fetchadd_int(&numfreebuffers, 1); 403 KASSERT(old >= 0 && old < nbuf, 404 ("numfreebuffers climbed to %d", old + 1)); 405 mtx_lock(&nblock); 406 if (needsbuffer) { 407 needsbuffer &= ~VFS_BIO_NEED_ANY; 408 if (numfreebuffers >= hifreebuffers) 409 needsbuffer &= ~VFS_BIO_NEED_FREE; 410 wakeup(&needsbuffer); 411 } 412 mtx_unlock(&nblock); 413 } 414 415 /* 416 * waitrunningbufspace() 417 * 418 * runningbufspace is a measure of the amount of I/O currently 419 * running. This routine is used in async-write situations to 420 * prevent creating huge backups of pending writes to a device. 421 * Only asynchronous writes are governed by this function. 422 * 423 * Reads will adjust runningbufspace, but will not block based on it. 424 * The read load has a side effect of reducing the allowed write load. 425 * 426 * This does NOT turn an async write into a sync write. It waits 427 * for earlier writes to complete and generally returns before the 428 * caller's write has reached the device. 429 */ 430 void 431 waitrunningbufspace(void) 432 { 433 434 mtx_lock(&rbreqlock); 435 while (runningbufspace > hirunningspace) { 436 ++runningbufreq; 437 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 438 } 439 mtx_unlock(&rbreqlock); 440 } 441 442 443 /* 444 * vfs_buf_test_cache: 445 * 446 * Called when a buffer is extended. This function clears the B_CACHE 447 * bit if the newly extended portion of the buffer does not contain 448 * valid data. 449 */ 450 static __inline 451 void 452 vfs_buf_test_cache(struct buf *bp, 453 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 454 vm_page_t m) 455 { 456 457 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 458 if (bp->b_flags & B_CACHE) { 459 int base = (foff + off) & PAGE_MASK; 460 if (vm_page_is_valid(m, base, size) == 0) 461 bp->b_flags &= ~B_CACHE; 462 } 463 } 464 465 /* Wake up the buffer daemon if necessary */ 466 static __inline 467 void 468 bd_wakeup(int dirtybuflevel) 469 { 470 471 mtx_lock(&bdlock); 472 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 473 bd_request = 1; 474 wakeup(&bd_request); 475 } 476 mtx_unlock(&bdlock); 477 } 478 479 /* 480 * bd_speedup - speedup the buffer cache flushing code 481 */ 482 483 void 484 bd_speedup(void) 485 { 486 int needwake; 487 488 mtx_lock(&bdlock); 489 needwake = 0; 490 if (bd_speedupreq == 0 || bd_request == 0) 491 needwake = 1; 492 bd_speedupreq = 1; 493 bd_request = 1; 494 if (needwake) 495 wakeup(&bd_request); 496 mtx_unlock(&bdlock); 497 } 498 499 /* 500 * Calculating buffer cache scaling values and reserve space for buffer 501 * headers. This is called during low level kernel initialization and 502 * may be called more then once. We CANNOT write to the memory area 503 * being reserved at this time. 504 */ 505 caddr_t 506 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 507 { 508 int tuned_nbuf; 509 long maxbuf; 510 511 /* 512 * physmem_est is in pages. Convert it to kilobytes (assumes 513 * PAGE_SIZE is >= 1K) 514 */ 515 physmem_est = physmem_est * (PAGE_SIZE / 1024); 516 517 /* 518 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 519 * For the first 64MB of ram nominally allocate sufficient buffers to 520 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 521 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 522 * the buffer cache we limit the eventual kva reservation to 523 * maxbcache bytes. 524 * 525 * factor represents the 1/4 x ram conversion. 526 */ 527 if (nbuf == 0) { 528 int factor = 4 * BKVASIZE / 1024; 529 530 nbuf = 50; 531 if (physmem_est > 4096) 532 nbuf += min((physmem_est - 4096) / factor, 533 65536 / factor); 534 if (physmem_est > 65536) 535 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 536 537 if (maxbcache && nbuf > maxbcache / BKVASIZE) 538 nbuf = maxbcache / BKVASIZE; 539 tuned_nbuf = 1; 540 } else 541 tuned_nbuf = 0; 542 543 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 544 maxbuf = (LONG_MAX / 3) / BKVASIZE; 545 if (nbuf > maxbuf) { 546 if (!tuned_nbuf) 547 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 548 maxbuf); 549 nbuf = maxbuf; 550 } 551 552 /* 553 * swbufs are used as temporary holders for I/O, such as paging I/O. 554 * We have no less then 16 and no more then 256. 555 */ 556 nswbuf = max(min(nbuf/4, 256), 16); 557 #ifdef NSWBUF_MIN 558 if (nswbuf < NSWBUF_MIN) 559 nswbuf = NSWBUF_MIN; 560 #endif 561 #ifdef DIRECTIO 562 ffs_rawread_setup(); 563 #endif 564 565 /* 566 * Reserve space for the buffer cache buffers 567 */ 568 swbuf = (void *)v; 569 v = (caddr_t)(swbuf + nswbuf); 570 buf = (void *)v; 571 v = (caddr_t)(buf + nbuf); 572 573 return(v); 574 } 575 576 /* Initialize the buffer subsystem. Called before use of any buffers. */ 577 void 578 bufinit(void) 579 { 580 struct buf *bp; 581 int i; 582 583 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 584 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 585 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 586 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 587 588 /* next, make a null set of free lists */ 589 for (i = 0; i < BUFFER_QUEUES; i++) 590 TAILQ_INIT(&bufqueues[i]); 591 592 /* finally, initialize each buffer header and stick on empty q */ 593 for (i = 0; i < nbuf; i++) { 594 bp = &buf[i]; 595 bzero(bp, sizeof *bp); 596 bp->b_flags = B_INVAL; /* we're just an empty header */ 597 bp->b_rcred = NOCRED; 598 bp->b_wcred = NOCRED; 599 bp->b_qindex = QUEUE_EMPTY; 600 bp->b_vflags = BV_INFREECNT; /* buf is counted as free */ 601 bp->b_xflags = 0; 602 LIST_INIT(&bp->b_dep); 603 BUF_LOCKINIT(bp); 604 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 605 } 606 607 /* 608 * maxbufspace is the absolute maximum amount of buffer space we are 609 * allowed to reserve in KVM and in real terms. The absolute maximum 610 * is nominally used by buf_daemon. hibufspace is the nominal maximum 611 * used by most other processes. The differential is required to 612 * ensure that buf_daemon is able to run when other processes might 613 * be blocked waiting for buffer space. 614 * 615 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 616 * this may result in KVM fragmentation which is not handled optimally 617 * by the system. 618 */ 619 maxbufspace = (long)nbuf * BKVASIZE; 620 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 621 lobufspace = hibufspace - MAXBSIZE; 622 623 /* 624 * Note: The 16 MB upper limit for hirunningspace was chosen 625 * arbitrarily and may need further tuning. The lower 1 MB 626 * limit is the historical upper limit for hirunningspace. 627 */ 628 hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE), 629 16 * 1024 * 1024), 1024 * 1024); 630 lorunningspace = roundup(hirunningspace / 2, MAXBSIZE); 631 632 /* 633 * Limit the amount of malloc memory since it is wired permanently into 634 * the kernel space. Even though this is accounted for in the buffer 635 * allocation, we don't want the malloced region to grow uncontrolled. 636 * The malloc scheme improves memory utilization significantly on average 637 * (small) directories. 638 */ 639 maxbufmallocspace = hibufspace / 20; 640 641 /* 642 * Reduce the chance of a deadlock occuring by limiting the number 643 * of delayed-write dirty buffers we allow to stack up. 644 */ 645 hidirtybuffers = nbuf / 4 + 20; 646 dirtybufthresh = hidirtybuffers * 9 / 10; 647 numdirtybuffers = 0; 648 /* 649 * To support extreme low-memory systems, make sure hidirtybuffers cannot 650 * eat up all available buffer space. This occurs when our minimum cannot 651 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 652 * BKVASIZE'd (8K) buffers. 653 */ 654 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 655 hidirtybuffers >>= 1; 656 } 657 lodirtybuffers = hidirtybuffers / 2; 658 659 /* 660 * Try to keep the number of free buffers in the specified range, 661 * and give special processes (e.g. like buf_daemon) access to an 662 * emergency reserve. 663 */ 664 lofreebuffers = nbuf / 18 + 5; 665 hifreebuffers = 2 * lofreebuffers; 666 numfreebuffers = nbuf; 667 668 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 669 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 670 } 671 672 /* 673 * bfreekva() - free the kva allocation for a buffer. 674 * 675 * Since this call frees up buffer space, we call bufspacewakeup(). 676 */ 677 static void 678 bfreekva(struct buf *bp) 679 { 680 681 if (bp->b_kvasize) { 682 atomic_add_int(&buffreekvacnt, 1); 683 atomic_subtract_long(&bufspace, bp->b_kvasize); 684 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 685 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 686 bp->b_kvasize = 0; 687 bufspacewakeup(); 688 } 689 } 690 691 /* 692 * bremfree: 693 * 694 * Mark the buffer for removal from the appropriate free list in brelse. 695 * 696 */ 697 void 698 bremfree(struct buf *bp) 699 { 700 int old; 701 702 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 703 KASSERT((bp->b_flags & B_REMFREE) == 0, 704 ("bremfree: buffer %p already marked for delayed removal.", bp)); 705 KASSERT(bp->b_qindex != QUEUE_NONE, 706 ("bremfree: buffer %p not on a queue.", bp)); 707 BUF_ASSERT_HELD(bp); 708 709 bp->b_flags |= B_REMFREE; 710 /* Fixup numfreebuffers count. */ 711 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 712 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 713 ("buf %p not counted in numfreebuffers", bp)); 714 bp->b_vflags &= ~BV_INFREECNT; 715 old = atomic_fetchadd_int(&numfreebuffers, -1); 716 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 717 } 718 } 719 720 /* 721 * bremfreef: 722 * 723 * Force an immediate removal from a free list. Used only in nfs when 724 * it abuses the b_freelist pointer. 725 */ 726 void 727 bremfreef(struct buf *bp) 728 { 729 mtx_lock(&bqlock); 730 bremfreel(bp); 731 mtx_unlock(&bqlock); 732 } 733 734 /* 735 * bremfreel: 736 * 737 * Removes a buffer from the free list, must be called with the 738 * bqlock held. 739 */ 740 static void 741 bremfreel(struct buf *bp) 742 { 743 int old; 744 745 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 746 bp, bp->b_vp, bp->b_flags); 747 KASSERT(bp->b_qindex != QUEUE_NONE, 748 ("bremfreel: buffer %p not on a queue.", bp)); 749 BUF_ASSERT_HELD(bp); 750 mtx_assert(&bqlock, MA_OWNED); 751 752 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 753 bp->b_qindex = QUEUE_NONE; 754 /* 755 * If this was a delayed bremfree() we only need to remove the buffer 756 * from the queue and return the stats are already done. 757 */ 758 if (bp->b_flags & B_REMFREE) { 759 bp->b_flags &= ~B_REMFREE; 760 return; 761 } 762 /* 763 * Fixup numfreebuffers count. If the buffer is invalid or not 764 * delayed-write, the buffer was free and we must decrement 765 * numfreebuffers. 766 */ 767 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 768 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 769 ("buf %p not counted in numfreebuffers", bp)); 770 bp->b_vflags &= ~BV_INFREECNT; 771 old = atomic_fetchadd_int(&numfreebuffers, -1); 772 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 773 } 774 } 775 776 777 /* 778 * Get a buffer with the specified data. Look in the cache first. We 779 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 780 * is set, the buffer is valid and we do not have to do anything ( see 781 * getblk() ). This is really just a special case of breadn(). 782 */ 783 int 784 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 785 struct buf **bpp) 786 { 787 788 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 789 } 790 791 /* 792 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 793 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 794 * the buffer is valid and we do not have to do anything. 795 */ 796 void 797 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 798 int cnt, struct ucred * cred) 799 { 800 struct buf *rabp; 801 int i; 802 803 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 804 if (inmem(vp, *rablkno)) 805 continue; 806 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 807 808 if ((rabp->b_flags & B_CACHE) == 0) { 809 if (!TD_IS_IDLETHREAD(curthread)) 810 curthread->td_ru.ru_inblock++; 811 rabp->b_flags |= B_ASYNC; 812 rabp->b_flags &= ~B_INVAL; 813 rabp->b_ioflags &= ~BIO_ERROR; 814 rabp->b_iocmd = BIO_READ; 815 if (rabp->b_rcred == NOCRED && cred != NOCRED) 816 rabp->b_rcred = crhold(cred); 817 vfs_busy_pages(rabp, 0); 818 BUF_KERNPROC(rabp); 819 rabp->b_iooffset = dbtob(rabp->b_blkno); 820 bstrategy(rabp); 821 } else { 822 brelse(rabp); 823 } 824 } 825 } 826 827 /* 828 * Operates like bread, but also starts asynchronous I/O on 829 * read-ahead blocks. 830 */ 831 int 832 breadn(struct vnode * vp, daddr_t blkno, int size, 833 daddr_t * rablkno, int *rabsize, 834 int cnt, struct ucred * cred, struct buf **bpp) 835 { 836 struct buf *bp; 837 int rv = 0, readwait = 0; 838 839 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 840 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 841 842 /* if not found in cache, do some I/O */ 843 if ((bp->b_flags & B_CACHE) == 0) { 844 if (!TD_IS_IDLETHREAD(curthread)) 845 curthread->td_ru.ru_inblock++; 846 bp->b_iocmd = BIO_READ; 847 bp->b_flags &= ~B_INVAL; 848 bp->b_ioflags &= ~BIO_ERROR; 849 if (bp->b_rcred == NOCRED && cred != NOCRED) 850 bp->b_rcred = crhold(cred); 851 vfs_busy_pages(bp, 0); 852 bp->b_iooffset = dbtob(bp->b_blkno); 853 bstrategy(bp); 854 ++readwait; 855 } 856 857 breada(vp, rablkno, rabsize, cnt, cred); 858 859 if (readwait) { 860 rv = bufwait(bp); 861 } 862 return (rv); 863 } 864 865 /* 866 * Write, release buffer on completion. (Done by iodone 867 * if async). Do not bother writing anything if the buffer 868 * is invalid. 869 * 870 * Note that we set B_CACHE here, indicating that buffer is 871 * fully valid and thus cacheable. This is true even of NFS 872 * now so we set it generally. This could be set either here 873 * or in biodone() since the I/O is synchronous. We put it 874 * here. 875 */ 876 int 877 bufwrite(struct buf *bp) 878 { 879 int oldflags; 880 struct vnode *vp; 881 int vp_md; 882 883 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 884 if (bp->b_flags & B_INVAL) { 885 brelse(bp); 886 return (0); 887 } 888 889 oldflags = bp->b_flags; 890 891 BUF_ASSERT_HELD(bp); 892 893 if (bp->b_pin_count > 0) 894 bunpin_wait(bp); 895 896 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 897 ("FFS background buffer should not get here %p", bp)); 898 899 vp = bp->b_vp; 900 if (vp) 901 vp_md = vp->v_vflag & VV_MD; 902 else 903 vp_md = 0; 904 905 /* Mark the buffer clean */ 906 bundirty(bp); 907 908 bp->b_flags &= ~B_DONE; 909 bp->b_ioflags &= ~BIO_ERROR; 910 bp->b_flags |= B_CACHE; 911 bp->b_iocmd = BIO_WRITE; 912 913 bufobj_wref(bp->b_bufobj); 914 vfs_busy_pages(bp, 1); 915 916 /* 917 * Normal bwrites pipeline writes 918 */ 919 bp->b_runningbufspace = bp->b_bufsize; 920 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 921 922 if (!TD_IS_IDLETHREAD(curthread)) 923 curthread->td_ru.ru_oublock++; 924 if (oldflags & B_ASYNC) 925 BUF_KERNPROC(bp); 926 bp->b_iooffset = dbtob(bp->b_blkno); 927 bstrategy(bp); 928 929 if ((oldflags & B_ASYNC) == 0) { 930 int rtval = bufwait(bp); 931 brelse(bp); 932 return (rtval); 933 } else { 934 /* 935 * don't allow the async write to saturate the I/O 936 * system. We will not deadlock here because 937 * we are blocking waiting for I/O that is already in-progress 938 * to complete. We do not block here if it is the update 939 * or syncer daemon trying to clean up as that can lead 940 * to deadlock. 941 */ 942 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 943 waitrunningbufspace(); 944 } 945 946 return (0); 947 } 948 949 void 950 bufbdflush(struct bufobj *bo, struct buf *bp) 951 { 952 struct buf *nbp; 953 954 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 955 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 956 altbufferflushes++; 957 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 958 BO_LOCK(bo); 959 /* 960 * Try to find a buffer to flush. 961 */ 962 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 963 if ((nbp->b_vflags & BV_BKGRDINPROG) || 964 BUF_LOCK(nbp, 965 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 966 continue; 967 if (bp == nbp) 968 panic("bdwrite: found ourselves"); 969 BO_UNLOCK(bo); 970 /* Don't countdeps with the bo lock held. */ 971 if (buf_countdeps(nbp, 0)) { 972 BO_LOCK(bo); 973 BUF_UNLOCK(nbp); 974 continue; 975 } 976 if (nbp->b_flags & B_CLUSTEROK) { 977 vfs_bio_awrite(nbp); 978 } else { 979 bremfree(nbp); 980 bawrite(nbp); 981 } 982 dirtybufferflushes++; 983 break; 984 } 985 if (nbp == NULL) 986 BO_UNLOCK(bo); 987 } 988 } 989 990 /* 991 * Delayed write. (Buffer is marked dirty). Do not bother writing 992 * anything if the buffer is marked invalid. 993 * 994 * Note that since the buffer must be completely valid, we can safely 995 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 996 * biodone() in order to prevent getblk from writing the buffer 997 * out synchronously. 998 */ 999 void 1000 bdwrite(struct buf *bp) 1001 { 1002 struct thread *td = curthread; 1003 struct vnode *vp; 1004 struct bufobj *bo; 1005 1006 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1007 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1008 BUF_ASSERT_HELD(bp); 1009 1010 if (bp->b_flags & B_INVAL) { 1011 brelse(bp); 1012 return; 1013 } 1014 1015 /* 1016 * If we have too many dirty buffers, don't create any more. 1017 * If we are wildly over our limit, then force a complete 1018 * cleanup. Otherwise, just keep the situation from getting 1019 * out of control. Note that we have to avoid a recursive 1020 * disaster and not try to clean up after our own cleanup! 1021 */ 1022 vp = bp->b_vp; 1023 bo = bp->b_bufobj; 1024 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1025 td->td_pflags |= TDP_INBDFLUSH; 1026 BO_BDFLUSH(bo, bp); 1027 td->td_pflags &= ~TDP_INBDFLUSH; 1028 } else 1029 recursiveflushes++; 1030 1031 bdirty(bp); 1032 /* 1033 * Set B_CACHE, indicating that the buffer is fully valid. This is 1034 * true even of NFS now. 1035 */ 1036 bp->b_flags |= B_CACHE; 1037 1038 /* 1039 * This bmap keeps the system from needing to do the bmap later, 1040 * perhaps when the system is attempting to do a sync. Since it 1041 * is likely that the indirect block -- or whatever other datastructure 1042 * that the filesystem needs is still in memory now, it is a good 1043 * thing to do this. Note also, that if the pageout daemon is 1044 * requesting a sync -- there might not be enough memory to do 1045 * the bmap then... So, this is important to do. 1046 */ 1047 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1048 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1049 } 1050 1051 /* 1052 * Set the *dirty* buffer range based upon the VM system dirty 1053 * pages. 1054 * 1055 * Mark the buffer pages as clean. We need to do this here to 1056 * satisfy the vnode_pager and the pageout daemon, so that it 1057 * thinks that the pages have been "cleaned". Note that since 1058 * the pages are in a delayed write buffer -- the VFS layer 1059 * "will" see that the pages get written out on the next sync, 1060 * or perhaps the cluster will be completed. 1061 */ 1062 vfs_clean_pages_dirty_buf(bp); 1063 bqrelse(bp); 1064 1065 /* 1066 * Wakeup the buffer flushing daemon if we have a lot of dirty 1067 * buffers (midpoint between our recovery point and our stall 1068 * point). 1069 */ 1070 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1071 1072 /* 1073 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1074 * due to the softdep code. 1075 */ 1076 } 1077 1078 /* 1079 * bdirty: 1080 * 1081 * Turn buffer into delayed write request. We must clear BIO_READ and 1082 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1083 * itself to properly update it in the dirty/clean lists. We mark it 1084 * B_DONE to ensure that any asynchronization of the buffer properly 1085 * clears B_DONE ( else a panic will occur later ). 1086 * 1087 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1088 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1089 * should only be called if the buffer is known-good. 1090 * 1091 * Since the buffer is not on a queue, we do not update the numfreebuffers 1092 * count. 1093 * 1094 * The buffer must be on QUEUE_NONE. 1095 */ 1096 void 1097 bdirty(struct buf *bp) 1098 { 1099 1100 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1101 bp, bp->b_vp, bp->b_flags); 1102 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1103 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1104 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1105 BUF_ASSERT_HELD(bp); 1106 bp->b_flags &= ~(B_RELBUF); 1107 bp->b_iocmd = BIO_WRITE; 1108 1109 if ((bp->b_flags & B_DELWRI) == 0) { 1110 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1111 reassignbuf(bp); 1112 atomic_add_int(&numdirtybuffers, 1); 1113 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1114 } 1115 } 1116 1117 /* 1118 * bundirty: 1119 * 1120 * Clear B_DELWRI for buffer. 1121 * 1122 * Since the buffer is not on a queue, we do not update the numfreebuffers 1123 * count. 1124 * 1125 * The buffer must be on QUEUE_NONE. 1126 */ 1127 1128 void 1129 bundirty(struct buf *bp) 1130 { 1131 1132 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1133 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1134 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1135 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1136 BUF_ASSERT_HELD(bp); 1137 1138 if (bp->b_flags & B_DELWRI) { 1139 bp->b_flags &= ~B_DELWRI; 1140 reassignbuf(bp); 1141 atomic_subtract_int(&numdirtybuffers, 1); 1142 numdirtywakeup(lodirtybuffers); 1143 } 1144 /* 1145 * Since it is now being written, we can clear its deferred write flag. 1146 */ 1147 bp->b_flags &= ~B_DEFERRED; 1148 } 1149 1150 /* 1151 * bawrite: 1152 * 1153 * Asynchronous write. Start output on a buffer, but do not wait for 1154 * it to complete. The buffer is released when the output completes. 1155 * 1156 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1157 * B_INVAL buffers. Not us. 1158 */ 1159 void 1160 bawrite(struct buf *bp) 1161 { 1162 1163 bp->b_flags |= B_ASYNC; 1164 (void) bwrite(bp); 1165 } 1166 1167 /* 1168 * bwillwrite: 1169 * 1170 * Called prior to the locking of any vnodes when we are expecting to 1171 * write. We do not want to starve the buffer cache with too many 1172 * dirty buffers so we block here. By blocking prior to the locking 1173 * of any vnodes we attempt to avoid the situation where a locked vnode 1174 * prevents the various system daemons from flushing related buffers. 1175 */ 1176 1177 void 1178 bwillwrite(void) 1179 { 1180 1181 if (numdirtybuffers >= hidirtybuffers) { 1182 mtx_lock(&nblock); 1183 while (numdirtybuffers >= hidirtybuffers) { 1184 bd_wakeup(1); 1185 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1186 msleep(&needsbuffer, &nblock, 1187 (PRIBIO + 4), "flswai", 0); 1188 } 1189 mtx_unlock(&nblock); 1190 } 1191 } 1192 1193 /* 1194 * Return true if we have too many dirty buffers. 1195 */ 1196 int 1197 buf_dirty_count_severe(void) 1198 { 1199 1200 return(numdirtybuffers >= hidirtybuffers); 1201 } 1202 1203 static __noinline int 1204 buf_vm_page_count_severe(void) 1205 { 1206 1207 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1208 1209 return vm_page_count_severe(); 1210 } 1211 1212 /* 1213 * brelse: 1214 * 1215 * Release a busy buffer and, if requested, free its resources. The 1216 * buffer will be stashed in the appropriate bufqueue[] allowing it 1217 * to be accessed later as a cache entity or reused for other purposes. 1218 */ 1219 void 1220 brelse(struct buf *bp) 1221 { 1222 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1223 bp, bp->b_vp, bp->b_flags); 1224 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1225 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1226 1227 if (bp->b_flags & B_MANAGED) { 1228 bqrelse(bp); 1229 return; 1230 } 1231 1232 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1233 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1234 /* 1235 * Failed write, redirty. Must clear BIO_ERROR to prevent 1236 * pages from being scrapped. If the error is anything 1237 * other than an I/O error (EIO), assume that retrying 1238 * is futile. 1239 */ 1240 bp->b_ioflags &= ~BIO_ERROR; 1241 bdirty(bp); 1242 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1243 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1244 /* 1245 * Either a failed I/O or we were asked to free or not 1246 * cache the buffer. 1247 */ 1248 bp->b_flags |= B_INVAL; 1249 if (!LIST_EMPTY(&bp->b_dep)) 1250 buf_deallocate(bp); 1251 if (bp->b_flags & B_DELWRI) { 1252 atomic_subtract_int(&numdirtybuffers, 1); 1253 numdirtywakeup(lodirtybuffers); 1254 } 1255 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1256 if ((bp->b_flags & B_VMIO) == 0) { 1257 if (bp->b_bufsize) 1258 allocbuf(bp, 0); 1259 if (bp->b_vp) 1260 brelvp(bp); 1261 } 1262 } 1263 1264 /* 1265 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1266 * is called with B_DELWRI set, the underlying pages may wind up 1267 * getting freed causing a previous write (bdwrite()) to get 'lost' 1268 * because pages associated with a B_DELWRI bp are marked clean. 1269 * 1270 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1271 * if B_DELWRI is set. 1272 * 1273 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1274 * on pages to return pages to the VM page queues. 1275 */ 1276 if (bp->b_flags & B_DELWRI) 1277 bp->b_flags &= ~B_RELBUF; 1278 else if (buf_vm_page_count_severe()) { 1279 /* 1280 * The locking of the BO_LOCK is not necessary since 1281 * BKGRDINPROG cannot be set while we hold the buf 1282 * lock, it can only be cleared if it is already 1283 * pending. 1284 */ 1285 if (bp->b_vp) { 1286 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1287 bp->b_flags |= B_RELBUF; 1288 } else 1289 bp->b_flags |= B_RELBUF; 1290 } 1291 1292 /* 1293 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1294 * constituted, not even NFS buffers now. Two flags effect this. If 1295 * B_INVAL, the struct buf is invalidated but the VM object is kept 1296 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1297 * 1298 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1299 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1300 * buffer is also B_INVAL because it hits the re-dirtying code above. 1301 * 1302 * Normally we can do this whether a buffer is B_DELWRI or not. If 1303 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1304 * the commit state and we cannot afford to lose the buffer. If the 1305 * buffer has a background write in progress, we need to keep it 1306 * around to prevent it from being reconstituted and starting a second 1307 * background write. 1308 */ 1309 if ((bp->b_flags & B_VMIO) 1310 && !(bp->b_vp->v_mount != NULL && 1311 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1312 !vn_isdisk(bp->b_vp, NULL) && 1313 (bp->b_flags & B_DELWRI)) 1314 ) { 1315 1316 int i, j, resid; 1317 vm_page_t m; 1318 off_t foff; 1319 vm_pindex_t poff; 1320 vm_object_t obj; 1321 1322 obj = bp->b_bufobj->bo_object; 1323 1324 /* 1325 * Get the base offset and length of the buffer. Note that 1326 * in the VMIO case if the buffer block size is not 1327 * page-aligned then b_data pointer may not be page-aligned. 1328 * But our b_pages[] array *IS* page aligned. 1329 * 1330 * block sizes less then DEV_BSIZE (usually 512) are not 1331 * supported due to the page granularity bits (m->valid, 1332 * m->dirty, etc...). 1333 * 1334 * See man buf(9) for more information 1335 */ 1336 resid = bp->b_bufsize; 1337 foff = bp->b_offset; 1338 VM_OBJECT_LOCK(obj); 1339 for (i = 0; i < bp->b_npages; i++) { 1340 int had_bogus = 0; 1341 1342 m = bp->b_pages[i]; 1343 1344 /* 1345 * If we hit a bogus page, fixup *all* the bogus pages 1346 * now. 1347 */ 1348 if (m == bogus_page) { 1349 poff = OFF_TO_IDX(bp->b_offset); 1350 had_bogus = 1; 1351 1352 for (j = i; j < bp->b_npages; j++) { 1353 vm_page_t mtmp; 1354 mtmp = bp->b_pages[j]; 1355 if (mtmp == bogus_page) { 1356 mtmp = vm_page_lookup(obj, poff + j); 1357 if (!mtmp) { 1358 panic("brelse: page missing\n"); 1359 } 1360 bp->b_pages[j] = mtmp; 1361 } 1362 } 1363 1364 if ((bp->b_flags & B_INVAL) == 0) { 1365 pmap_qenter( 1366 trunc_page((vm_offset_t)bp->b_data), 1367 bp->b_pages, bp->b_npages); 1368 } 1369 m = bp->b_pages[i]; 1370 } 1371 if ((bp->b_flags & B_NOCACHE) || 1372 (bp->b_ioflags & BIO_ERROR && 1373 bp->b_iocmd == BIO_READ)) { 1374 int poffset = foff & PAGE_MASK; 1375 int presid = resid > (PAGE_SIZE - poffset) ? 1376 (PAGE_SIZE - poffset) : resid; 1377 1378 KASSERT(presid >= 0, ("brelse: extra page")); 1379 vm_page_set_invalid(m, poffset, presid); 1380 if (had_bogus) 1381 printf("avoided corruption bug in bogus_page/brelse code\n"); 1382 } 1383 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1384 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1385 } 1386 VM_OBJECT_UNLOCK(obj); 1387 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1388 vfs_vmio_release(bp); 1389 1390 } else if (bp->b_flags & B_VMIO) { 1391 1392 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1393 vfs_vmio_release(bp); 1394 } 1395 1396 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1397 if (bp->b_bufsize != 0) 1398 allocbuf(bp, 0); 1399 if (bp->b_vp != NULL) 1400 brelvp(bp); 1401 } 1402 1403 if (BUF_LOCKRECURSED(bp)) { 1404 /* do not release to free list */ 1405 BUF_UNLOCK(bp); 1406 return; 1407 } 1408 1409 /* enqueue */ 1410 mtx_lock(&bqlock); 1411 /* Handle delayed bremfree() processing. */ 1412 if (bp->b_flags & B_REMFREE) 1413 bremfreel(bp); 1414 if (bp->b_qindex != QUEUE_NONE) 1415 panic("brelse: free buffer onto another queue???"); 1416 1417 /* 1418 * If the buffer has junk contents signal it and eventually 1419 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1420 * doesn't find it. 1421 */ 1422 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1423 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1424 bp->b_flags |= B_INVAL; 1425 if (bp->b_flags & B_INVAL) { 1426 if (bp->b_flags & B_DELWRI) 1427 bundirty(bp); 1428 if (bp->b_vp) 1429 brelvp(bp); 1430 } 1431 1432 /* buffers with no memory */ 1433 if (bp->b_bufsize == 0) { 1434 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1435 if (bp->b_vflags & BV_BKGRDINPROG) 1436 panic("losing buffer 1"); 1437 if (bp->b_kvasize) { 1438 bp->b_qindex = QUEUE_EMPTYKVA; 1439 } else { 1440 bp->b_qindex = QUEUE_EMPTY; 1441 } 1442 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1443 /* buffers with junk contents */ 1444 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1445 (bp->b_ioflags & BIO_ERROR)) { 1446 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1447 if (bp->b_vflags & BV_BKGRDINPROG) 1448 panic("losing buffer 2"); 1449 bp->b_qindex = QUEUE_CLEAN; 1450 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1451 /* remaining buffers */ 1452 } else { 1453 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1454 (B_DELWRI|B_NEEDSGIANT)) 1455 bp->b_qindex = QUEUE_DIRTY_GIANT; 1456 else if (bp->b_flags & B_DELWRI) 1457 bp->b_qindex = QUEUE_DIRTY; 1458 else 1459 bp->b_qindex = QUEUE_CLEAN; 1460 if (bp->b_flags & B_AGE) 1461 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1462 else 1463 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1464 } 1465 mtx_unlock(&bqlock); 1466 1467 /* 1468 * Fixup numfreebuffers count. The bp is on an appropriate queue 1469 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1470 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1471 * if B_INVAL is set ). 1472 */ 1473 1474 if (!(bp->b_flags & B_DELWRI)) 1475 bufcountwakeup(bp); 1476 1477 /* 1478 * Something we can maybe free or reuse 1479 */ 1480 if (bp->b_bufsize || bp->b_kvasize) 1481 bufspacewakeup(); 1482 1483 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1484 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1485 panic("brelse: not dirty"); 1486 /* unlock */ 1487 BUF_UNLOCK(bp); 1488 } 1489 1490 /* 1491 * Release a buffer back to the appropriate queue but do not try to free 1492 * it. The buffer is expected to be used again soon. 1493 * 1494 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1495 * biodone() to requeue an async I/O on completion. It is also used when 1496 * known good buffers need to be requeued but we think we may need the data 1497 * again soon. 1498 * 1499 * XXX we should be able to leave the B_RELBUF hint set on completion. 1500 */ 1501 void 1502 bqrelse(struct buf *bp) 1503 { 1504 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1505 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1506 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1507 1508 if (BUF_LOCKRECURSED(bp)) { 1509 /* do not release to free list */ 1510 BUF_UNLOCK(bp); 1511 return; 1512 } 1513 1514 if (bp->b_flags & B_MANAGED) { 1515 if (bp->b_flags & B_REMFREE) { 1516 mtx_lock(&bqlock); 1517 bremfreel(bp); 1518 mtx_unlock(&bqlock); 1519 } 1520 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1521 BUF_UNLOCK(bp); 1522 return; 1523 } 1524 1525 mtx_lock(&bqlock); 1526 /* Handle delayed bremfree() processing. */ 1527 if (bp->b_flags & B_REMFREE) 1528 bremfreel(bp); 1529 if (bp->b_qindex != QUEUE_NONE) 1530 panic("bqrelse: free buffer onto another queue???"); 1531 /* buffers with stale but valid contents */ 1532 if (bp->b_flags & B_DELWRI) { 1533 if (bp->b_flags & B_NEEDSGIANT) 1534 bp->b_qindex = QUEUE_DIRTY_GIANT; 1535 else 1536 bp->b_qindex = QUEUE_DIRTY; 1537 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1538 } else { 1539 /* 1540 * The locking of the BO_LOCK for checking of the 1541 * BV_BKGRDINPROG is not necessary since the 1542 * BV_BKGRDINPROG cannot be set while we hold the buf 1543 * lock, it can only be cleared if it is already 1544 * pending. 1545 */ 1546 if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1547 bp->b_qindex = QUEUE_CLEAN; 1548 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1549 b_freelist); 1550 } else { 1551 /* 1552 * We are too low on memory, we have to try to free 1553 * the buffer (most importantly: the wired pages 1554 * making up its backing store) *now*. 1555 */ 1556 mtx_unlock(&bqlock); 1557 brelse(bp); 1558 return; 1559 } 1560 } 1561 mtx_unlock(&bqlock); 1562 1563 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1564 bufcountwakeup(bp); 1565 1566 /* 1567 * Something we can maybe free or reuse. 1568 */ 1569 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1570 bufspacewakeup(); 1571 1572 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1573 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1574 panic("bqrelse: not dirty"); 1575 /* unlock */ 1576 BUF_UNLOCK(bp); 1577 } 1578 1579 /* Give pages used by the bp back to the VM system (where possible) */ 1580 static void 1581 vfs_vmio_release(struct buf *bp) 1582 { 1583 int i; 1584 vm_page_t m; 1585 1586 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1587 for (i = 0; i < bp->b_npages; i++) { 1588 m = bp->b_pages[i]; 1589 bp->b_pages[i] = NULL; 1590 /* 1591 * In order to keep page LRU ordering consistent, put 1592 * everything on the inactive queue. 1593 */ 1594 vm_page_lock(m); 1595 vm_page_unwire(m, 0); 1596 /* 1597 * We don't mess with busy pages, it is 1598 * the responsibility of the process that 1599 * busied the pages to deal with them. 1600 */ 1601 if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 && 1602 m->wire_count == 0) { 1603 /* 1604 * Might as well free the page if we can and it has 1605 * no valid data. We also free the page if the 1606 * buffer was used for direct I/O 1607 */ 1608 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1609 m->hold_count == 0) { 1610 vm_page_free(m); 1611 } else if (bp->b_flags & B_DIRECT) { 1612 vm_page_try_to_free(m); 1613 } else if (buf_vm_page_count_severe()) { 1614 vm_page_try_to_cache(m); 1615 } 1616 } 1617 vm_page_unlock(m); 1618 } 1619 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1620 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1621 1622 if (bp->b_bufsize) { 1623 bufspacewakeup(); 1624 bp->b_bufsize = 0; 1625 } 1626 bp->b_npages = 0; 1627 bp->b_flags &= ~B_VMIO; 1628 if (bp->b_vp) 1629 brelvp(bp); 1630 } 1631 1632 /* 1633 * Check to see if a block at a particular lbn is available for a clustered 1634 * write. 1635 */ 1636 static int 1637 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1638 { 1639 struct buf *bpa; 1640 int match; 1641 1642 match = 0; 1643 1644 /* If the buf isn't in core skip it */ 1645 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1646 return (0); 1647 1648 /* If the buf is busy we don't want to wait for it */ 1649 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1650 return (0); 1651 1652 /* Only cluster with valid clusterable delayed write buffers */ 1653 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1654 (B_DELWRI | B_CLUSTEROK)) 1655 goto done; 1656 1657 if (bpa->b_bufsize != size) 1658 goto done; 1659 1660 /* 1661 * Check to see if it is in the expected place on disk and that the 1662 * block has been mapped. 1663 */ 1664 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1665 match = 1; 1666 done: 1667 BUF_UNLOCK(bpa); 1668 return (match); 1669 } 1670 1671 /* 1672 * vfs_bio_awrite: 1673 * 1674 * Implement clustered async writes for clearing out B_DELWRI buffers. 1675 * This is much better then the old way of writing only one buffer at 1676 * a time. Note that we may not be presented with the buffers in the 1677 * correct order, so we search for the cluster in both directions. 1678 */ 1679 int 1680 vfs_bio_awrite(struct buf *bp) 1681 { 1682 struct bufobj *bo; 1683 int i; 1684 int j; 1685 daddr_t lblkno = bp->b_lblkno; 1686 struct vnode *vp = bp->b_vp; 1687 int ncl; 1688 int nwritten; 1689 int size; 1690 int maxcl; 1691 1692 bo = &vp->v_bufobj; 1693 /* 1694 * right now we support clustered writing only to regular files. If 1695 * we find a clusterable block we could be in the middle of a cluster 1696 * rather then at the beginning. 1697 */ 1698 if ((vp->v_type == VREG) && 1699 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1700 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1701 1702 size = vp->v_mount->mnt_stat.f_iosize; 1703 maxcl = MAXPHYS / size; 1704 1705 BO_LOCK(bo); 1706 for (i = 1; i < maxcl; i++) 1707 if (vfs_bio_clcheck(vp, size, lblkno + i, 1708 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1709 break; 1710 1711 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1712 if (vfs_bio_clcheck(vp, size, lblkno - j, 1713 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1714 break; 1715 BO_UNLOCK(bo); 1716 --j; 1717 ncl = i + j; 1718 /* 1719 * this is a possible cluster write 1720 */ 1721 if (ncl != 1) { 1722 BUF_UNLOCK(bp); 1723 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1724 return nwritten; 1725 } 1726 } 1727 bremfree(bp); 1728 bp->b_flags |= B_ASYNC; 1729 /* 1730 * default (old) behavior, writing out only one block 1731 * 1732 * XXX returns b_bufsize instead of b_bcount for nwritten? 1733 */ 1734 nwritten = bp->b_bufsize; 1735 (void) bwrite(bp); 1736 1737 return nwritten; 1738 } 1739 1740 /* 1741 * getnewbuf: 1742 * 1743 * Find and initialize a new buffer header, freeing up existing buffers 1744 * in the bufqueues as necessary. The new buffer is returned locked. 1745 * 1746 * Important: B_INVAL is not set. If the caller wishes to throw the 1747 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1748 * 1749 * We block if: 1750 * We have insufficient buffer headers 1751 * We have insufficient buffer space 1752 * buffer_map is too fragmented ( space reservation fails ) 1753 * If we have to flush dirty buffers ( but we try to avoid this ) 1754 * 1755 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1756 * Instead we ask the buf daemon to do it for us. We attempt to 1757 * avoid piecemeal wakeups of the pageout daemon. 1758 */ 1759 1760 static struct buf * 1761 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 1762 int gbflags) 1763 { 1764 struct thread *td; 1765 struct buf *bp; 1766 struct buf *nbp; 1767 int defrag = 0; 1768 int nqindex; 1769 static int flushingbufs; 1770 1771 td = curthread; 1772 /* 1773 * We can't afford to block since we might be holding a vnode lock, 1774 * which may prevent system daemons from running. We deal with 1775 * low-memory situations by proactively returning memory and running 1776 * async I/O rather then sync I/O. 1777 */ 1778 atomic_add_int(&getnewbufcalls, 1); 1779 atomic_subtract_int(&getnewbufrestarts, 1); 1780 restart: 1781 atomic_add_int(&getnewbufrestarts, 1); 1782 1783 /* 1784 * Setup for scan. If we do not have enough free buffers, 1785 * we setup a degenerate case that immediately fails. Note 1786 * that if we are specially marked process, we are allowed to 1787 * dip into our reserves. 1788 * 1789 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1790 * 1791 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1792 * However, there are a number of cases (defragging, reusing, ...) 1793 * where we cannot backup. 1794 */ 1795 mtx_lock(&bqlock); 1796 nqindex = QUEUE_EMPTYKVA; 1797 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1798 1799 if (nbp == NULL) { 1800 /* 1801 * If no EMPTYKVA buffers and we are either 1802 * defragging or reusing, locate a CLEAN buffer 1803 * to free or reuse. If bufspace useage is low 1804 * skip this step so we can allocate a new buffer. 1805 */ 1806 if (defrag || bufspace >= lobufspace) { 1807 nqindex = QUEUE_CLEAN; 1808 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1809 } 1810 1811 /* 1812 * If we could not find or were not allowed to reuse a 1813 * CLEAN buffer, check to see if it is ok to use an EMPTY 1814 * buffer. We can only use an EMPTY buffer if allocating 1815 * its KVA would not otherwise run us out of buffer space. 1816 */ 1817 if (nbp == NULL && defrag == 0 && 1818 bufspace + maxsize < hibufspace) { 1819 nqindex = QUEUE_EMPTY; 1820 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1821 } 1822 } 1823 1824 /* 1825 * Run scan, possibly freeing data and/or kva mappings on the fly 1826 * depending. 1827 */ 1828 1829 while ((bp = nbp) != NULL) { 1830 int qindex = nqindex; 1831 1832 /* 1833 * Calculate next bp ( we can only use it if we do not block 1834 * or do other fancy things ). 1835 */ 1836 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1837 switch(qindex) { 1838 case QUEUE_EMPTY: 1839 nqindex = QUEUE_EMPTYKVA; 1840 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1841 break; 1842 /* FALLTHROUGH */ 1843 case QUEUE_EMPTYKVA: 1844 nqindex = QUEUE_CLEAN; 1845 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1846 break; 1847 /* FALLTHROUGH */ 1848 case QUEUE_CLEAN: 1849 /* 1850 * nbp is NULL. 1851 */ 1852 break; 1853 } 1854 } 1855 /* 1856 * If we are defragging then we need a buffer with 1857 * b_kvasize != 0. XXX this situation should no longer 1858 * occur, if defrag is non-zero the buffer's b_kvasize 1859 * should also be non-zero at this point. XXX 1860 */ 1861 if (defrag && bp->b_kvasize == 0) { 1862 printf("Warning: defrag empty buffer %p\n", bp); 1863 continue; 1864 } 1865 1866 /* 1867 * Start freeing the bp. This is somewhat involved. nbp 1868 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1869 */ 1870 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1871 continue; 1872 if (bp->b_vp) { 1873 BO_LOCK(bp->b_bufobj); 1874 if (bp->b_vflags & BV_BKGRDINPROG) { 1875 BO_UNLOCK(bp->b_bufobj); 1876 BUF_UNLOCK(bp); 1877 continue; 1878 } 1879 BO_UNLOCK(bp->b_bufobj); 1880 } 1881 CTR6(KTR_BUF, 1882 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1883 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1884 bp->b_kvasize, bp->b_bufsize, qindex); 1885 1886 /* 1887 * Sanity Checks 1888 */ 1889 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1890 1891 /* 1892 * Note: we no longer distinguish between VMIO and non-VMIO 1893 * buffers. 1894 */ 1895 1896 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1897 1898 bremfreel(bp); 1899 mtx_unlock(&bqlock); 1900 1901 if (qindex == QUEUE_CLEAN) { 1902 if (bp->b_flags & B_VMIO) { 1903 bp->b_flags &= ~B_ASYNC; 1904 vfs_vmio_release(bp); 1905 } 1906 if (bp->b_vp) 1907 brelvp(bp); 1908 } 1909 1910 /* 1911 * NOTE: nbp is now entirely invalid. We can only restart 1912 * the scan from this point on. 1913 * 1914 * Get the rest of the buffer freed up. b_kva* is still 1915 * valid after this operation. 1916 */ 1917 1918 if (bp->b_rcred != NOCRED) { 1919 crfree(bp->b_rcred); 1920 bp->b_rcred = NOCRED; 1921 } 1922 if (bp->b_wcred != NOCRED) { 1923 crfree(bp->b_wcred); 1924 bp->b_wcred = NOCRED; 1925 } 1926 if (!LIST_EMPTY(&bp->b_dep)) 1927 buf_deallocate(bp); 1928 if (bp->b_vflags & BV_BKGRDINPROG) 1929 panic("losing buffer 3"); 1930 KASSERT(bp->b_vp == NULL, 1931 ("bp: %p still has vnode %p. qindex: %d", 1932 bp, bp->b_vp, qindex)); 1933 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1934 ("bp: %p still on a buffer list. xflags %X", 1935 bp, bp->b_xflags)); 1936 1937 if (bp->b_bufsize) 1938 allocbuf(bp, 0); 1939 1940 bp->b_flags = 0; 1941 bp->b_ioflags = 0; 1942 bp->b_xflags = 0; 1943 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 1944 ("buf %p still counted as free?", bp)); 1945 bp->b_vflags = 0; 1946 bp->b_vp = NULL; 1947 bp->b_blkno = bp->b_lblkno = 0; 1948 bp->b_offset = NOOFFSET; 1949 bp->b_iodone = 0; 1950 bp->b_error = 0; 1951 bp->b_resid = 0; 1952 bp->b_bcount = 0; 1953 bp->b_npages = 0; 1954 bp->b_dirtyoff = bp->b_dirtyend = 0; 1955 bp->b_bufobj = NULL; 1956 bp->b_pin_count = 0; 1957 bp->b_fsprivate1 = NULL; 1958 bp->b_fsprivate2 = NULL; 1959 bp->b_fsprivate3 = NULL; 1960 1961 LIST_INIT(&bp->b_dep); 1962 1963 /* 1964 * If we are defragging then free the buffer. 1965 */ 1966 if (defrag) { 1967 bp->b_flags |= B_INVAL; 1968 bfreekva(bp); 1969 brelse(bp); 1970 defrag = 0; 1971 goto restart; 1972 } 1973 1974 /* 1975 * Notify any waiters for the buffer lock about 1976 * identity change by freeing the buffer. 1977 */ 1978 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 1979 bp->b_flags |= B_INVAL; 1980 bfreekva(bp); 1981 brelse(bp); 1982 goto restart; 1983 } 1984 1985 /* 1986 * If we are overcomitted then recover the buffer and its 1987 * KVM space. This occurs in rare situations when multiple 1988 * processes are blocked in getnewbuf() or allocbuf(). 1989 */ 1990 if (bufspace >= hibufspace) 1991 flushingbufs = 1; 1992 if (flushingbufs && bp->b_kvasize != 0) { 1993 bp->b_flags |= B_INVAL; 1994 bfreekva(bp); 1995 brelse(bp); 1996 goto restart; 1997 } 1998 if (bufspace < lobufspace) 1999 flushingbufs = 0; 2000 break; 2001 } 2002 2003 /* 2004 * If we exhausted our list, sleep as appropriate. We may have to 2005 * wakeup various daemons and write out some dirty buffers. 2006 * 2007 * Generally we are sleeping due to insufficient buffer space. 2008 */ 2009 2010 if (bp == NULL) { 2011 int flags, norunbuf; 2012 char *waitmsg; 2013 int fl; 2014 2015 if (defrag) { 2016 flags = VFS_BIO_NEED_BUFSPACE; 2017 waitmsg = "nbufkv"; 2018 } else if (bufspace >= hibufspace) { 2019 waitmsg = "nbufbs"; 2020 flags = VFS_BIO_NEED_BUFSPACE; 2021 } else { 2022 waitmsg = "newbuf"; 2023 flags = VFS_BIO_NEED_ANY; 2024 } 2025 mtx_lock(&nblock); 2026 needsbuffer |= flags; 2027 mtx_unlock(&nblock); 2028 mtx_unlock(&bqlock); 2029 2030 bd_speedup(); /* heeeelp */ 2031 if (gbflags & GB_NOWAIT_BD) 2032 return (NULL); 2033 2034 mtx_lock(&nblock); 2035 while (needsbuffer & flags) { 2036 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2037 mtx_unlock(&nblock); 2038 /* 2039 * getblk() is called with a vnode 2040 * locked, and some majority of the 2041 * dirty buffers may as well belong to 2042 * the vnode. Flushing the buffers 2043 * there would make a progress that 2044 * cannot be achieved by the 2045 * buf_daemon, that cannot lock the 2046 * vnode. 2047 */ 2048 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2049 (td->td_pflags & TDP_NORUNNINGBUF); 2050 /* play bufdaemon */ 2051 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2052 fl = buf_do_flush(vp); 2053 td->td_pflags &= norunbuf; 2054 mtx_lock(&nblock); 2055 if (fl != 0) 2056 continue; 2057 if ((needsbuffer & flags) == 0) 2058 break; 2059 } 2060 if (msleep(&needsbuffer, &nblock, 2061 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2062 mtx_unlock(&nblock); 2063 return (NULL); 2064 } 2065 } 2066 mtx_unlock(&nblock); 2067 } else { 2068 /* 2069 * We finally have a valid bp. We aren't quite out of the 2070 * woods, we still have to reserve kva space. In order 2071 * to keep fragmentation sane we only allocate kva in 2072 * BKVASIZE chunks. 2073 */ 2074 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2075 2076 if (maxsize != bp->b_kvasize) { 2077 vm_offset_t addr = 0; 2078 2079 bfreekva(bp); 2080 2081 vm_map_lock(buffer_map); 2082 if (vm_map_findspace(buffer_map, 2083 vm_map_min(buffer_map), maxsize, &addr)) { 2084 /* 2085 * Uh oh. Buffer map is to fragmented. We 2086 * must defragment the map. 2087 */ 2088 atomic_add_int(&bufdefragcnt, 1); 2089 vm_map_unlock(buffer_map); 2090 defrag = 1; 2091 bp->b_flags |= B_INVAL; 2092 brelse(bp); 2093 goto restart; 2094 } 2095 if (addr) { 2096 vm_map_insert(buffer_map, NULL, 0, 2097 addr, addr + maxsize, 2098 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2099 2100 bp->b_kvabase = (caddr_t) addr; 2101 bp->b_kvasize = maxsize; 2102 atomic_add_long(&bufspace, bp->b_kvasize); 2103 atomic_add_int(&bufreusecnt, 1); 2104 } 2105 vm_map_unlock(buffer_map); 2106 } 2107 bp->b_saveaddr = bp->b_kvabase; 2108 bp->b_data = bp->b_saveaddr; 2109 } 2110 return(bp); 2111 } 2112 2113 /* 2114 * buf_daemon: 2115 * 2116 * buffer flushing daemon. Buffers are normally flushed by the 2117 * update daemon but if it cannot keep up this process starts to 2118 * take the load in an attempt to prevent getnewbuf() from blocking. 2119 */ 2120 2121 static struct kproc_desc buf_kp = { 2122 "bufdaemon", 2123 buf_daemon, 2124 &bufdaemonproc 2125 }; 2126 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2127 2128 static int 2129 buf_do_flush(struct vnode *vp) 2130 { 2131 int flushed; 2132 2133 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2134 /* The list empty check here is slightly racy */ 2135 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2136 mtx_lock(&Giant); 2137 flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0); 2138 mtx_unlock(&Giant); 2139 } 2140 if (flushed == 0) { 2141 /* 2142 * Could not find any buffers without rollback 2143 * dependencies, so just write the first one 2144 * in the hopes of eventually making progress. 2145 */ 2146 flushbufqueues(vp, QUEUE_DIRTY, 1); 2147 if (!TAILQ_EMPTY( 2148 &bufqueues[QUEUE_DIRTY_GIANT])) { 2149 mtx_lock(&Giant); 2150 flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1); 2151 mtx_unlock(&Giant); 2152 } 2153 } 2154 return (flushed); 2155 } 2156 2157 static void 2158 buf_daemon() 2159 { 2160 int lodirtysave; 2161 2162 /* 2163 * This process needs to be suspended prior to shutdown sync. 2164 */ 2165 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2166 SHUTDOWN_PRI_LAST); 2167 2168 /* 2169 * This process is allowed to take the buffer cache to the limit 2170 */ 2171 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2172 mtx_lock(&bdlock); 2173 for (;;) { 2174 bd_request = 0; 2175 mtx_unlock(&bdlock); 2176 2177 kproc_suspend_check(bufdaemonproc); 2178 lodirtysave = lodirtybuffers; 2179 if (bd_speedupreq) { 2180 lodirtybuffers = numdirtybuffers / 2; 2181 bd_speedupreq = 0; 2182 } 2183 /* 2184 * Do the flush. Limit the amount of in-transit I/O we 2185 * allow to build up, otherwise we would completely saturate 2186 * the I/O system. Wakeup any waiting processes before we 2187 * normally would so they can run in parallel with our drain. 2188 */ 2189 while (numdirtybuffers > lodirtybuffers) { 2190 if (buf_do_flush(NULL) == 0) 2191 break; 2192 uio_yield(); 2193 } 2194 lodirtybuffers = lodirtysave; 2195 2196 /* 2197 * Only clear bd_request if we have reached our low water 2198 * mark. The buf_daemon normally waits 1 second and 2199 * then incrementally flushes any dirty buffers that have 2200 * built up, within reason. 2201 * 2202 * If we were unable to hit our low water mark and couldn't 2203 * find any flushable buffers, we sleep half a second. 2204 * Otherwise we loop immediately. 2205 */ 2206 mtx_lock(&bdlock); 2207 if (numdirtybuffers <= lodirtybuffers) { 2208 /* 2209 * We reached our low water mark, reset the 2210 * request and sleep until we are needed again. 2211 * The sleep is just so the suspend code works. 2212 */ 2213 bd_request = 0; 2214 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2215 } else { 2216 /* 2217 * We couldn't find any flushable dirty buffers but 2218 * still have too many dirty buffers, we 2219 * have to sleep and try again. (rare) 2220 */ 2221 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2222 } 2223 } 2224 } 2225 2226 /* 2227 * flushbufqueues: 2228 * 2229 * Try to flush a buffer in the dirty queue. We must be careful to 2230 * free up B_INVAL buffers instead of write them, which NFS is 2231 * particularly sensitive to. 2232 */ 2233 static int flushwithdeps = 0; 2234 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2235 0, "Number of buffers flushed with dependecies that require rollbacks"); 2236 2237 static int 2238 flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2239 { 2240 struct buf *sentinel; 2241 struct vnode *vp; 2242 struct mount *mp; 2243 struct buf *bp; 2244 int hasdeps; 2245 int flushed; 2246 int target; 2247 2248 if (lvp == NULL) { 2249 target = numdirtybuffers - lodirtybuffers; 2250 if (flushdeps && target > 2) 2251 target /= 2; 2252 } else 2253 target = flushbufqtarget; 2254 flushed = 0; 2255 bp = NULL; 2256 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2257 sentinel->b_qindex = QUEUE_SENTINEL; 2258 mtx_lock(&bqlock); 2259 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2260 while (flushed != target) { 2261 bp = TAILQ_NEXT(sentinel, b_freelist); 2262 if (bp != NULL) { 2263 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2264 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2265 b_freelist); 2266 } else 2267 break; 2268 /* 2269 * Skip sentinels inserted by other invocations of the 2270 * flushbufqueues(), taking care to not reorder them. 2271 */ 2272 if (bp->b_qindex == QUEUE_SENTINEL) 2273 continue; 2274 /* 2275 * Only flush the buffers that belong to the 2276 * vnode locked by the curthread. 2277 */ 2278 if (lvp != NULL && bp->b_vp != lvp) 2279 continue; 2280 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2281 continue; 2282 if (bp->b_pin_count > 0) { 2283 BUF_UNLOCK(bp); 2284 continue; 2285 } 2286 BO_LOCK(bp->b_bufobj); 2287 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2288 (bp->b_flags & B_DELWRI) == 0) { 2289 BO_UNLOCK(bp->b_bufobj); 2290 BUF_UNLOCK(bp); 2291 continue; 2292 } 2293 BO_UNLOCK(bp->b_bufobj); 2294 if (bp->b_flags & B_INVAL) { 2295 bremfreel(bp); 2296 mtx_unlock(&bqlock); 2297 brelse(bp); 2298 flushed++; 2299 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2300 mtx_lock(&bqlock); 2301 continue; 2302 } 2303 2304 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2305 if (flushdeps == 0) { 2306 BUF_UNLOCK(bp); 2307 continue; 2308 } 2309 hasdeps = 1; 2310 } else 2311 hasdeps = 0; 2312 /* 2313 * We must hold the lock on a vnode before writing 2314 * one of its buffers. Otherwise we may confuse, or 2315 * in the case of a snapshot vnode, deadlock the 2316 * system. 2317 * 2318 * The lock order here is the reverse of the normal 2319 * of vnode followed by buf lock. This is ok because 2320 * the NOWAIT will prevent deadlock. 2321 */ 2322 vp = bp->b_vp; 2323 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2324 BUF_UNLOCK(bp); 2325 continue; 2326 } 2327 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2328 mtx_unlock(&bqlock); 2329 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2330 bp, bp->b_vp, bp->b_flags); 2331 if (curproc == bufdaemonproc) 2332 vfs_bio_awrite(bp); 2333 else { 2334 bremfree(bp); 2335 bwrite(bp); 2336 notbufdflashes++; 2337 } 2338 vn_finished_write(mp); 2339 VOP_UNLOCK(vp, 0); 2340 flushwithdeps += hasdeps; 2341 flushed++; 2342 2343 /* 2344 * Sleeping on runningbufspace while holding 2345 * vnode lock leads to deadlock. 2346 */ 2347 if (curproc == bufdaemonproc) 2348 waitrunningbufspace(); 2349 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2350 mtx_lock(&bqlock); 2351 continue; 2352 } 2353 vn_finished_write(mp); 2354 BUF_UNLOCK(bp); 2355 } 2356 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2357 mtx_unlock(&bqlock); 2358 free(sentinel, M_TEMP); 2359 return (flushed); 2360 } 2361 2362 /* 2363 * Check to see if a block is currently memory resident. 2364 */ 2365 struct buf * 2366 incore(struct bufobj *bo, daddr_t blkno) 2367 { 2368 struct buf *bp; 2369 2370 BO_LOCK(bo); 2371 bp = gbincore(bo, blkno); 2372 BO_UNLOCK(bo); 2373 return (bp); 2374 } 2375 2376 /* 2377 * Returns true if no I/O is needed to access the 2378 * associated VM object. This is like incore except 2379 * it also hunts around in the VM system for the data. 2380 */ 2381 2382 static int 2383 inmem(struct vnode * vp, daddr_t blkno) 2384 { 2385 vm_object_t obj; 2386 vm_offset_t toff, tinc, size; 2387 vm_page_t m; 2388 vm_ooffset_t off; 2389 2390 ASSERT_VOP_LOCKED(vp, "inmem"); 2391 2392 if (incore(&vp->v_bufobj, blkno)) 2393 return 1; 2394 if (vp->v_mount == NULL) 2395 return 0; 2396 obj = vp->v_object; 2397 if (obj == NULL) 2398 return (0); 2399 2400 size = PAGE_SIZE; 2401 if (size > vp->v_mount->mnt_stat.f_iosize) 2402 size = vp->v_mount->mnt_stat.f_iosize; 2403 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2404 2405 VM_OBJECT_LOCK(obj); 2406 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2407 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2408 if (!m) 2409 goto notinmem; 2410 tinc = size; 2411 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2412 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2413 if (vm_page_is_valid(m, 2414 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2415 goto notinmem; 2416 } 2417 VM_OBJECT_UNLOCK(obj); 2418 return 1; 2419 2420 notinmem: 2421 VM_OBJECT_UNLOCK(obj); 2422 return (0); 2423 } 2424 2425 /* 2426 * Set the dirty range for a buffer based on the status of the dirty 2427 * bits in the pages comprising the buffer. The range is limited 2428 * to the size of the buffer. 2429 * 2430 * Tell the VM system that the pages associated with this buffer 2431 * are clean. This is used for delayed writes where the data is 2432 * going to go to disk eventually without additional VM intevention. 2433 * 2434 * Note that while we only really need to clean through to b_bcount, we 2435 * just go ahead and clean through to b_bufsize. 2436 */ 2437 static void 2438 vfs_clean_pages_dirty_buf(struct buf *bp) 2439 { 2440 vm_ooffset_t foff, noff, eoff; 2441 vm_page_t m; 2442 int i; 2443 2444 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2445 return; 2446 2447 foff = bp->b_offset; 2448 KASSERT(bp->b_offset != NOOFFSET, 2449 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2450 2451 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2452 vfs_drain_busy_pages(bp); 2453 vfs_setdirty_locked_object(bp); 2454 for (i = 0; i < bp->b_npages; i++) { 2455 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2456 eoff = noff; 2457 if (eoff > bp->b_offset + bp->b_bufsize) 2458 eoff = bp->b_offset + bp->b_bufsize; 2459 m = bp->b_pages[i]; 2460 vfs_page_set_validclean(bp, foff, m); 2461 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2462 foff = noff; 2463 } 2464 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2465 } 2466 2467 static void 2468 vfs_setdirty_locked_object(struct buf *bp) 2469 { 2470 vm_object_t object; 2471 int i; 2472 2473 object = bp->b_bufobj->bo_object; 2474 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2475 2476 /* 2477 * We qualify the scan for modified pages on whether the 2478 * object has been flushed yet. 2479 */ 2480 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2481 vm_offset_t boffset; 2482 vm_offset_t eoffset; 2483 2484 /* 2485 * test the pages to see if they have been modified directly 2486 * by users through the VM system. 2487 */ 2488 for (i = 0; i < bp->b_npages; i++) 2489 vm_page_test_dirty(bp->b_pages[i]); 2490 2491 /* 2492 * Calculate the encompassing dirty range, boffset and eoffset, 2493 * (eoffset - boffset) bytes. 2494 */ 2495 2496 for (i = 0; i < bp->b_npages; i++) { 2497 if (bp->b_pages[i]->dirty) 2498 break; 2499 } 2500 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2501 2502 for (i = bp->b_npages - 1; i >= 0; --i) { 2503 if (bp->b_pages[i]->dirty) { 2504 break; 2505 } 2506 } 2507 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2508 2509 /* 2510 * Fit it to the buffer. 2511 */ 2512 2513 if (eoffset > bp->b_bcount) 2514 eoffset = bp->b_bcount; 2515 2516 /* 2517 * If we have a good dirty range, merge with the existing 2518 * dirty range. 2519 */ 2520 2521 if (boffset < eoffset) { 2522 if (bp->b_dirtyoff > boffset) 2523 bp->b_dirtyoff = boffset; 2524 if (bp->b_dirtyend < eoffset) 2525 bp->b_dirtyend = eoffset; 2526 } 2527 } 2528 } 2529 2530 /* 2531 * getblk: 2532 * 2533 * Get a block given a specified block and offset into a file/device. 2534 * The buffers B_DONE bit will be cleared on return, making it almost 2535 * ready for an I/O initiation. B_INVAL may or may not be set on 2536 * return. The caller should clear B_INVAL prior to initiating a 2537 * READ. 2538 * 2539 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2540 * an existing buffer. 2541 * 2542 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2543 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2544 * and then cleared based on the backing VM. If the previous buffer is 2545 * non-0-sized but invalid, B_CACHE will be cleared. 2546 * 2547 * If getblk() must create a new buffer, the new buffer is returned with 2548 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2549 * case it is returned with B_INVAL clear and B_CACHE set based on the 2550 * backing VM. 2551 * 2552 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2553 * B_CACHE bit is clear. 2554 * 2555 * What this means, basically, is that the caller should use B_CACHE to 2556 * determine whether the buffer is fully valid or not and should clear 2557 * B_INVAL prior to issuing a read. If the caller intends to validate 2558 * the buffer by loading its data area with something, the caller needs 2559 * to clear B_INVAL. If the caller does this without issuing an I/O, 2560 * the caller should set B_CACHE ( as an optimization ), else the caller 2561 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2562 * a write attempt or if it was a successfull read. If the caller 2563 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2564 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2565 */ 2566 struct buf * 2567 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2568 int flags) 2569 { 2570 struct buf *bp; 2571 struct bufobj *bo; 2572 int error; 2573 2574 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2575 ASSERT_VOP_LOCKED(vp, "getblk"); 2576 if (size > MAXBSIZE) 2577 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2578 2579 bo = &vp->v_bufobj; 2580 loop: 2581 /* 2582 * Block if we are low on buffers. Certain processes are allowed 2583 * to completely exhaust the buffer cache. 2584 * 2585 * If this check ever becomes a bottleneck it may be better to 2586 * move it into the else, when gbincore() fails. At the moment 2587 * it isn't a problem. 2588 * 2589 * XXX remove if 0 sections (clean this up after its proven) 2590 */ 2591 if (numfreebuffers == 0) { 2592 if (TD_IS_IDLETHREAD(curthread)) 2593 return NULL; 2594 mtx_lock(&nblock); 2595 needsbuffer |= VFS_BIO_NEED_ANY; 2596 mtx_unlock(&nblock); 2597 } 2598 2599 BO_LOCK(bo); 2600 bp = gbincore(bo, blkno); 2601 if (bp != NULL) { 2602 int lockflags; 2603 /* 2604 * Buffer is in-core. If the buffer is not busy, it must 2605 * be on a queue. 2606 */ 2607 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2608 2609 if (flags & GB_LOCK_NOWAIT) 2610 lockflags |= LK_NOWAIT; 2611 2612 error = BUF_TIMELOCK(bp, lockflags, 2613 BO_MTX(bo), "getblk", slpflag, slptimeo); 2614 2615 /* 2616 * If we slept and got the lock we have to restart in case 2617 * the buffer changed identities. 2618 */ 2619 if (error == ENOLCK) 2620 goto loop; 2621 /* We timed out or were interrupted. */ 2622 else if (error) 2623 return (NULL); 2624 2625 /* 2626 * The buffer is locked. B_CACHE is cleared if the buffer is 2627 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2628 * and for a VMIO buffer B_CACHE is adjusted according to the 2629 * backing VM cache. 2630 */ 2631 if (bp->b_flags & B_INVAL) 2632 bp->b_flags &= ~B_CACHE; 2633 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2634 bp->b_flags |= B_CACHE; 2635 bremfree(bp); 2636 2637 /* 2638 * check for size inconsistancies for non-VMIO case. 2639 */ 2640 2641 if (bp->b_bcount != size) { 2642 if ((bp->b_flags & B_VMIO) == 0 || 2643 (size > bp->b_kvasize)) { 2644 if (bp->b_flags & B_DELWRI) { 2645 /* 2646 * If buffer is pinned and caller does 2647 * not want sleep waiting for it to be 2648 * unpinned, bail out 2649 * */ 2650 if (bp->b_pin_count > 0) { 2651 if (flags & GB_LOCK_NOWAIT) { 2652 bqrelse(bp); 2653 return (NULL); 2654 } else { 2655 bunpin_wait(bp); 2656 } 2657 } 2658 bp->b_flags |= B_NOCACHE; 2659 bwrite(bp); 2660 } else { 2661 if (LIST_EMPTY(&bp->b_dep)) { 2662 bp->b_flags |= B_RELBUF; 2663 brelse(bp); 2664 } else { 2665 bp->b_flags |= B_NOCACHE; 2666 bwrite(bp); 2667 } 2668 } 2669 goto loop; 2670 } 2671 } 2672 2673 /* 2674 * If the size is inconsistant in the VMIO case, we can resize 2675 * the buffer. This might lead to B_CACHE getting set or 2676 * cleared. If the size has not changed, B_CACHE remains 2677 * unchanged from its previous state. 2678 */ 2679 2680 if (bp->b_bcount != size) 2681 allocbuf(bp, size); 2682 2683 KASSERT(bp->b_offset != NOOFFSET, 2684 ("getblk: no buffer offset")); 2685 2686 /* 2687 * A buffer with B_DELWRI set and B_CACHE clear must 2688 * be committed before we can return the buffer in 2689 * order to prevent the caller from issuing a read 2690 * ( due to B_CACHE not being set ) and overwriting 2691 * it. 2692 * 2693 * Most callers, including NFS and FFS, need this to 2694 * operate properly either because they assume they 2695 * can issue a read if B_CACHE is not set, or because 2696 * ( for example ) an uncached B_DELWRI might loop due 2697 * to softupdates re-dirtying the buffer. In the latter 2698 * case, B_CACHE is set after the first write completes, 2699 * preventing further loops. 2700 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2701 * above while extending the buffer, we cannot allow the 2702 * buffer to remain with B_CACHE set after the write 2703 * completes or it will represent a corrupt state. To 2704 * deal with this we set B_NOCACHE to scrap the buffer 2705 * after the write. 2706 * 2707 * We might be able to do something fancy, like setting 2708 * B_CACHE in bwrite() except if B_DELWRI is already set, 2709 * so the below call doesn't set B_CACHE, but that gets real 2710 * confusing. This is much easier. 2711 */ 2712 2713 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2714 bp->b_flags |= B_NOCACHE; 2715 bwrite(bp); 2716 goto loop; 2717 } 2718 bp->b_flags &= ~B_DONE; 2719 } else { 2720 int bsize, maxsize, vmio; 2721 off_t offset; 2722 2723 /* 2724 * Buffer is not in-core, create new buffer. The buffer 2725 * returned by getnewbuf() is locked. Note that the returned 2726 * buffer is also considered valid (not marked B_INVAL). 2727 */ 2728 BO_UNLOCK(bo); 2729 /* 2730 * If the user does not want us to create the buffer, bail out 2731 * here. 2732 */ 2733 if (flags & GB_NOCREAT) 2734 return NULL; 2735 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 2736 offset = blkno * bsize; 2737 vmio = vp->v_object != NULL; 2738 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2739 maxsize = imax(maxsize, bsize); 2740 2741 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 2742 if (bp == NULL) { 2743 if (slpflag || slptimeo) 2744 return NULL; 2745 goto loop; 2746 } 2747 2748 /* 2749 * This code is used to make sure that a buffer is not 2750 * created while the getnewbuf routine is blocked. 2751 * This can be a problem whether the vnode is locked or not. 2752 * If the buffer is created out from under us, we have to 2753 * throw away the one we just created. 2754 * 2755 * Note: this must occur before we associate the buffer 2756 * with the vp especially considering limitations in 2757 * the splay tree implementation when dealing with duplicate 2758 * lblkno's. 2759 */ 2760 BO_LOCK(bo); 2761 if (gbincore(bo, blkno)) { 2762 BO_UNLOCK(bo); 2763 bp->b_flags |= B_INVAL; 2764 brelse(bp); 2765 goto loop; 2766 } 2767 2768 /* 2769 * Insert the buffer into the hash, so that it can 2770 * be found by incore. 2771 */ 2772 bp->b_blkno = bp->b_lblkno = blkno; 2773 bp->b_offset = offset; 2774 bgetvp(vp, bp); 2775 BO_UNLOCK(bo); 2776 2777 /* 2778 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2779 * buffer size starts out as 0, B_CACHE will be set by 2780 * allocbuf() for the VMIO case prior to it testing the 2781 * backing store for validity. 2782 */ 2783 2784 if (vmio) { 2785 bp->b_flags |= B_VMIO; 2786 #if defined(VFS_BIO_DEBUG) 2787 if (vn_canvmio(vp) != TRUE) 2788 printf("getblk: VMIO on vnode type %d\n", 2789 vp->v_type); 2790 #endif 2791 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2792 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2793 bp, vp->v_object, bp->b_bufobj->bo_object)); 2794 } else { 2795 bp->b_flags &= ~B_VMIO; 2796 KASSERT(bp->b_bufobj->bo_object == NULL, 2797 ("ARGH! has b_bufobj->bo_object %p %p\n", 2798 bp, bp->b_bufobj->bo_object)); 2799 } 2800 2801 allocbuf(bp, size); 2802 bp->b_flags &= ~B_DONE; 2803 } 2804 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2805 BUF_ASSERT_HELD(bp); 2806 KASSERT(bp->b_bufobj == bo, 2807 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2808 return (bp); 2809 } 2810 2811 /* 2812 * Get an empty, disassociated buffer of given size. The buffer is initially 2813 * set to B_INVAL. 2814 */ 2815 struct buf * 2816 geteblk(int size, int flags) 2817 { 2818 struct buf *bp; 2819 int maxsize; 2820 2821 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2822 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 2823 if ((flags & GB_NOWAIT_BD) && 2824 (curthread->td_pflags & TDP_BUFNEED) != 0) 2825 return (NULL); 2826 } 2827 allocbuf(bp, size); 2828 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2829 BUF_ASSERT_HELD(bp); 2830 return (bp); 2831 } 2832 2833 2834 /* 2835 * This code constitutes the buffer memory from either anonymous system 2836 * memory (in the case of non-VMIO operations) or from an associated 2837 * VM object (in the case of VMIO operations). This code is able to 2838 * resize a buffer up or down. 2839 * 2840 * Note that this code is tricky, and has many complications to resolve 2841 * deadlock or inconsistant data situations. Tread lightly!!! 2842 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2843 * the caller. Calling this code willy nilly can result in the loss of data. 2844 * 2845 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2846 * B_CACHE for the non-VMIO case. 2847 */ 2848 2849 int 2850 allocbuf(struct buf *bp, int size) 2851 { 2852 int newbsize, mbsize; 2853 int i; 2854 2855 BUF_ASSERT_HELD(bp); 2856 2857 if (bp->b_kvasize < size) 2858 panic("allocbuf: buffer too small"); 2859 2860 if ((bp->b_flags & B_VMIO) == 0) { 2861 caddr_t origbuf; 2862 int origbufsize; 2863 /* 2864 * Just get anonymous memory from the kernel. Don't 2865 * mess with B_CACHE. 2866 */ 2867 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2868 if (bp->b_flags & B_MALLOC) 2869 newbsize = mbsize; 2870 else 2871 newbsize = round_page(size); 2872 2873 if (newbsize < bp->b_bufsize) { 2874 /* 2875 * malloced buffers are not shrunk 2876 */ 2877 if (bp->b_flags & B_MALLOC) { 2878 if (newbsize) { 2879 bp->b_bcount = size; 2880 } else { 2881 free(bp->b_data, M_BIOBUF); 2882 if (bp->b_bufsize) { 2883 atomic_subtract_long( 2884 &bufmallocspace, 2885 bp->b_bufsize); 2886 bufspacewakeup(); 2887 bp->b_bufsize = 0; 2888 } 2889 bp->b_saveaddr = bp->b_kvabase; 2890 bp->b_data = bp->b_saveaddr; 2891 bp->b_bcount = 0; 2892 bp->b_flags &= ~B_MALLOC; 2893 } 2894 return 1; 2895 } 2896 vm_hold_free_pages(bp, newbsize); 2897 } else if (newbsize > bp->b_bufsize) { 2898 /* 2899 * We only use malloced memory on the first allocation. 2900 * and revert to page-allocated memory when the buffer 2901 * grows. 2902 */ 2903 /* 2904 * There is a potential smp race here that could lead 2905 * to bufmallocspace slightly passing the max. It 2906 * is probably extremely rare and not worth worrying 2907 * over. 2908 */ 2909 if ( (bufmallocspace < maxbufmallocspace) && 2910 (bp->b_bufsize == 0) && 2911 (mbsize <= PAGE_SIZE/2)) { 2912 2913 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2914 bp->b_bufsize = mbsize; 2915 bp->b_bcount = size; 2916 bp->b_flags |= B_MALLOC; 2917 atomic_add_long(&bufmallocspace, mbsize); 2918 return 1; 2919 } 2920 origbuf = NULL; 2921 origbufsize = 0; 2922 /* 2923 * If the buffer is growing on its other-than-first allocation, 2924 * then we revert to the page-allocation scheme. 2925 */ 2926 if (bp->b_flags & B_MALLOC) { 2927 origbuf = bp->b_data; 2928 origbufsize = bp->b_bufsize; 2929 bp->b_data = bp->b_kvabase; 2930 if (bp->b_bufsize) { 2931 atomic_subtract_long(&bufmallocspace, 2932 bp->b_bufsize); 2933 bufspacewakeup(); 2934 bp->b_bufsize = 0; 2935 } 2936 bp->b_flags &= ~B_MALLOC; 2937 newbsize = round_page(newbsize); 2938 } 2939 vm_hold_load_pages( 2940 bp, 2941 (vm_offset_t) bp->b_data + bp->b_bufsize, 2942 (vm_offset_t) bp->b_data + newbsize); 2943 if (origbuf) { 2944 bcopy(origbuf, bp->b_data, origbufsize); 2945 free(origbuf, M_BIOBUF); 2946 } 2947 } 2948 } else { 2949 int desiredpages; 2950 2951 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2952 desiredpages = (size == 0) ? 0 : 2953 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2954 2955 if (bp->b_flags & B_MALLOC) 2956 panic("allocbuf: VMIO buffer can't be malloced"); 2957 /* 2958 * Set B_CACHE initially if buffer is 0 length or will become 2959 * 0-length. 2960 */ 2961 if (size == 0 || bp->b_bufsize == 0) 2962 bp->b_flags |= B_CACHE; 2963 2964 if (newbsize < bp->b_bufsize) { 2965 /* 2966 * DEV_BSIZE aligned new buffer size is less then the 2967 * DEV_BSIZE aligned existing buffer size. Figure out 2968 * if we have to remove any pages. 2969 */ 2970 if (desiredpages < bp->b_npages) { 2971 vm_page_t m; 2972 2973 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2974 for (i = desiredpages; i < bp->b_npages; i++) { 2975 /* 2976 * the page is not freed here -- it 2977 * is the responsibility of 2978 * vnode_pager_setsize 2979 */ 2980 m = bp->b_pages[i]; 2981 KASSERT(m != bogus_page, 2982 ("allocbuf: bogus page found")); 2983 while (vm_page_sleep_if_busy(m, TRUE, 2984 "biodep")) 2985 continue; 2986 2987 bp->b_pages[i] = NULL; 2988 vm_page_lock(m); 2989 vm_page_unwire(m, 0); 2990 vm_page_unlock(m); 2991 } 2992 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2993 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2994 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2995 bp->b_npages = desiredpages; 2996 } 2997 } else if (size > bp->b_bcount) { 2998 /* 2999 * We are growing the buffer, possibly in a 3000 * byte-granular fashion. 3001 */ 3002 vm_object_t obj; 3003 vm_offset_t toff; 3004 vm_offset_t tinc; 3005 3006 /* 3007 * Step 1, bring in the VM pages from the object, 3008 * allocating them if necessary. We must clear 3009 * B_CACHE if these pages are not valid for the 3010 * range covered by the buffer. 3011 */ 3012 3013 obj = bp->b_bufobj->bo_object; 3014 3015 VM_OBJECT_LOCK(obj); 3016 while (bp->b_npages < desiredpages) { 3017 vm_page_t m; 3018 3019 /* 3020 * We must allocate system pages since blocking 3021 * here could intefere with paging I/O, no 3022 * matter which process we are. 3023 * 3024 * We can only test VPO_BUSY here. Blocking on 3025 * m->busy might lead to a deadlock: 3026 * vm_fault->getpages->cluster_read->allocbuf 3027 * Thus, we specify VM_ALLOC_IGN_SBUSY. 3028 */ 3029 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3030 bp->b_npages, VM_ALLOC_NOBUSY | 3031 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3032 VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY | 3033 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3034 if (m->valid == 0) 3035 bp->b_flags &= ~B_CACHE; 3036 bp->b_pages[bp->b_npages] = m; 3037 ++bp->b_npages; 3038 } 3039 3040 /* 3041 * Step 2. We've loaded the pages into the buffer, 3042 * we have to figure out if we can still have B_CACHE 3043 * set. Note that B_CACHE is set according to the 3044 * byte-granular range ( bcount and size ), new the 3045 * aligned range ( newbsize ). 3046 * 3047 * The VM test is against m->valid, which is DEV_BSIZE 3048 * aligned. Needless to say, the validity of the data 3049 * needs to also be DEV_BSIZE aligned. Note that this 3050 * fails with NFS if the server or some other client 3051 * extends the file's EOF. If our buffer is resized, 3052 * B_CACHE may remain set! XXX 3053 */ 3054 3055 toff = bp->b_bcount; 3056 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3057 3058 while ((bp->b_flags & B_CACHE) && toff < size) { 3059 vm_pindex_t pi; 3060 3061 if (tinc > (size - toff)) 3062 tinc = size - toff; 3063 3064 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3065 PAGE_SHIFT; 3066 3067 vfs_buf_test_cache( 3068 bp, 3069 bp->b_offset, 3070 toff, 3071 tinc, 3072 bp->b_pages[pi] 3073 ); 3074 toff += tinc; 3075 tinc = PAGE_SIZE; 3076 } 3077 VM_OBJECT_UNLOCK(obj); 3078 3079 /* 3080 * Step 3, fixup the KVM pmap. Remember that 3081 * bp->b_data is relative to bp->b_offset, but 3082 * bp->b_offset may be offset into the first page. 3083 */ 3084 3085 bp->b_data = (caddr_t) 3086 trunc_page((vm_offset_t)bp->b_data); 3087 pmap_qenter( 3088 (vm_offset_t)bp->b_data, 3089 bp->b_pages, 3090 bp->b_npages 3091 ); 3092 3093 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3094 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3095 } 3096 } 3097 if (newbsize < bp->b_bufsize) 3098 bufspacewakeup(); 3099 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3100 bp->b_bcount = size; /* requested buffer size */ 3101 return 1; 3102 } 3103 3104 void 3105 biodone(struct bio *bp) 3106 { 3107 struct mtx *mtxp; 3108 void (*done)(struct bio *); 3109 3110 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3111 mtx_lock(mtxp); 3112 bp->bio_flags |= BIO_DONE; 3113 done = bp->bio_done; 3114 if (done == NULL) 3115 wakeup(bp); 3116 mtx_unlock(mtxp); 3117 if (done != NULL) 3118 done(bp); 3119 } 3120 3121 /* 3122 * Wait for a BIO to finish. 3123 * 3124 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3125 * case is not yet clear. 3126 */ 3127 int 3128 biowait(struct bio *bp, const char *wchan) 3129 { 3130 struct mtx *mtxp; 3131 3132 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3133 mtx_lock(mtxp); 3134 while ((bp->bio_flags & BIO_DONE) == 0) 3135 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3136 mtx_unlock(mtxp); 3137 if (bp->bio_error != 0) 3138 return (bp->bio_error); 3139 if (!(bp->bio_flags & BIO_ERROR)) 3140 return (0); 3141 return (EIO); 3142 } 3143 3144 void 3145 biofinish(struct bio *bp, struct devstat *stat, int error) 3146 { 3147 3148 if (error) { 3149 bp->bio_error = error; 3150 bp->bio_flags |= BIO_ERROR; 3151 } 3152 if (stat != NULL) 3153 devstat_end_transaction_bio(stat, bp); 3154 biodone(bp); 3155 } 3156 3157 /* 3158 * bufwait: 3159 * 3160 * Wait for buffer I/O completion, returning error status. The buffer 3161 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3162 * error and cleared. 3163 */ 3164 int 3165 bufwait(struct buf *bp) 3166 { 3167 if (bp->b_iocmd == BIO_READ) 3168 bwait(bp, PRIBIO, "biord"); 3169 else 3170 bwait(bp, PRIBIO, "biowr"); 3171 if (bp->b_flags & B_EINTR) { 3172 bp->b_flags &= ~B_EINTR; 3173 return (EINTR); 3174 } 3175 if (bp->b_ioflags & BIO_ERROR) { 3176 return (bp->b_error ? bp->b_error : EIO); 3177 } else { 3178 return (0); 3179 } 3180 } 3181 3182 /* 3183 * Call back function from struct bio back up to struct buf. 3184 */ 3185 static void 3186 bufdonebio(struct bio *bip) 3187 { 3188 struct buf *bp; 3189 3190 bp = bip->bio_caller2; 3191 bp->b_resid = bp->b_bcount - bip->bio_completed; 3192 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3193 bp->b_ioflags = bip->bio_flags; 3194 bp->b_error = bip->bio_error; 3195 if (bp->b_error) 3196 bp->b_ioflags |= BIO_ERROR; 3197 bufdone(bp); 3198 g_destroy_bio(bip); 3199 } 3200 3201 void 3202 dev_strategy(struct cdev *dev, struct buf *bp) 3203 { 3204 struct cdevsw *csw; 3205 struct bio *bip; 3206 3207 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3208 panic("b_iocmd botch"); 3209 for (;;) { 3210 bip = g_new_bio(); 3211 if (bip != NULL) 3212 break; 3213 /* Try again later */ 3214 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3215 } 3216 bip->bio_cmd = bp->b_iocmd; 3217 bip->bio_offset = bp->b_iooffset; 3218 bip->bio_length = bp->b_bcount; 3219 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3220 bip->bio_data = bp->b_data; 3221 bip->bio_done = bufdonebio; 3222 bip->bio_caller2 = bp; 3223 bip->bio_dev = dev; 3224 KASSERT(dev->si_refcount > 0, 3225 ("dev_strategy on un-referenced struct cdev *(%s)", 3226 devtoname(dev))); 3227 csw = dev_refthread(dev); 3228 if (csw == NULL) { 3229 g_destroy_bio(bip); 3230 bp->b_error = ENXIO; 3231 bp->b_ioflags = BIO_ERROR; 3232 bufdone(bp); 3233 return; 3234 } 3235 (*csw->d_strategy)(bip); 3236 dev_relthread(dev); 3237 } 3238 3239 /* 3240 * bufdone: 3241 * 3242 * Finish I/O on a buffer, optionally calling a completion function. 3243 * This is usually called from an interrupt so process blocking is 3244 * not allowed. 3245 * 3246 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3247 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3248 * assuming B_INVAL is clear. 3249 * 3250 * For the VMIO case, we set B_CACHE if the op was a read and no 3251 * read error occured, or if the op was a write. B_CACHE is never 3252 * set if the buffer is invalid or otherwise uncacheable. 3253 * 3254 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3255 * initiator to leave B_INVAL set to brelse the buffer out of existance 3256 * in the biodone routine. 3257 */ 3258 void 3259 bufdone(struct buf *bp) 3260 { 3261 struct bufobj *dropobj; 3262 void (*biodone)(struct buf *); 3263 3264 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3265 dropobj = NULL; 3266 3267 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3268 BUF_ASSERT_HELD(bp); 3269 3270 runningbufwakeup(bp); 3271 if (bp->b_iocmd == BIO_WRITE) 3272 dropobj = bp->b_bufobj; 3273 /* call optional completion function if requested */ 3274 if (bp->b_iodone != NULL) { 3275 biodone = bp->b_iodone; 3276 bp->b_iodone = NULL; 3277 (*biodone) (bp); 3278 if (dropobj) 3279 bufobj_wdrop(dropobj); 3280 return; 3281 } 3282 3283 bufdone_finish(bp); 3284 3285 if (dropobj) 3286 bufobj_wdrop(dropobj); 3287 } 3288 3289 void 3290 bufdone_finish(struct buf *bp) 3291 { 3292 BUF_ASSERT_HELD(bp); 3293 3294 if (!LIST_EMPTY(&bp->b_dep)) 3295 buf_complete(bp); 3296 3297 if (bp->b_flags & B_VMIO) { 3298 int i; 3299 vm_ooffset_t foff; 3300 vm_page_t m; 3301 vm_object_t obj; 3302 int bogus, iosize; 3303 struct vnode *vp = bp->b_vp; 3304 3305 obj = bp->b_bufobj->bo_object; 3306 3307 #if defined(VFS_BIO_DEBUG) 3308 mp_fixme("usecount and vflag accessed without locks."); 3309 if (vp->v_usecount == 0) { 3310 panic("biodone: zero vnode ref count"); 3311 } 3312 3313 KASSERT(vp->v_object != NULL, 3314 ("biodone: vnode %p has no vm_object", vp)); 3315 #endif 3316 3317 foff = bp->b_offset; 3318 KASSERT(bp->b_offset != NOOFFSET, 3319 ("biodone: no buffer offset")); 3320 3321 VM_OBJECT_LOCK(obj); 3322 #if defined(VFS_BIO_DEBUG) 3323 if (obj->paging_in_progress < bp->b_npages) { 3324 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3325 obj->paging_in_progress, bp->b_npages); 3326 } 3327 #endif 3328 3329 /* 3330 * Set B_CACHE if the op was a normal read and no error 3331 * occured. B_CACHE is set for writes in the b*write() 3332 * routines. 3333 */ 3334 iosize = bp->b_bcount - bp->b_resid; 3335 if (bp->b_iocmd == BIO_READ && 3336 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3337 !(bp->b_ioflags & BIO_ERROR)) { 3338 bp->b_flags |= B_CACHE; 3339 } 3340 bogus = 0; 3341 for (i = 0; i < bp->b_npages; i++) { 3342 int bogusflag = 0; 3343 int resid; 3344 3345 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3346 if (resid > iosize) 3347 resid = iosize; 3348 3349 /* 3350 * cleanup bogus pages, restoring the originals 3351 */ 3352 m = bp->b_pages[i]; 3353 if (m == bogus_page) { 3354 bogus = bogusflag = 1; 3355 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3356 if (m == NULL) 3357 panic("biodone: page disappeared!"); 3358 bp->b_pages[i] = m; 3359 } 3360 #if defined(VFS_BIO_DEBUG) 3361 if (OFF_TO_IDX(foff) != m->pindex) { 3362 printf( 3363 "biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3364 (intmax_t)foff, (uintmax_t)m->pindex); 3365 } 3366 #endif 3367 3368 /* 3369 * In the write case, the valid and clean bits are 3370 * already changed correctly ( see bdwrite() ), so we 3371 * only need to do this here in the read case. 3372 */ 3373 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3374 KASSERT((m->dirty & vm_page_bits(foff & 3375 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3376 " page %p has unexpected dirty bits", m)); 3377 vfs_page_set_valid(bp, foff, m); 3378 } 3379 3380 /* 3381 * when debugging new filesystems or buffer I/O methods, this 3382 * is the most common error that pops up. if you see this, you 3383 * have not set the page busy flag correctly!!! 3384 */ 3385 if (m->busy == 0) { 3386 printf("biodone: page busy < 0, " 3387 "pindex: %d, foff: 0x(%x,%x), " 3388 "resid: %d, index: %d\n", 3389 (int) m->pindex, (int)(foff >> 32), 3390 (int) foff & 0xffffffff, resid, i); 3391 if (!vn_isdisk(vp, NULL)) 3392 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3393 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3394 (intmax_t) bp->b_lblkno, 3395 bp->b_flags, bp->b_npages); 3396 else 3397 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3398 (intmax_t) bp->b_lblkno, 3399 bp->b_flags, bp->b_npages); 3400 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3401 (u_long)m->valid, (u_long)m->dirty, 3402 m->wire_count); 3403 panic("biodone: page busy < 0\n"); 3404 } 3405 vm_page_io_finish(m); 3406 vm_object_pip_subtract(obj, 1); 3407 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3408 iosize -= resid; 3409 } 3410 vm_object_pip_wakeupn(obj, 0); 3411 VM_OBJECT_UNLOCK(obj); 3412 if (bogus) 3413 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3414 bp->b_pages, bp->b_npages); 3415 } 3416 3417 /* 3418 * For asynchronous completions, release the buffer now. The brelse 3419 * will do a wakeup there if necessary - so no need to do a wakeup 3420 * here in the async case. The sync case always needs to do a wakeup. 3421 */ 3422 3423 if (bp->b_flags & B_ASYNC) { 3424 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3425 brelse(bp); 3426 else 3427 bqrelse(bp); 3428 } else 3429 bdone(bp); 3430 } 3431 3432 /* 3433 * This routine is called in lieu of iodone in the case of 3434 * incomplete I/O. This keeps the busy status for pages 3435 * consistant. 3436 */ 3437 void 3438 vfs_unbusy_pages(struct buf *bp) 3439 { 3440 int i; 3441 vm_object_t obj; 3442 vm_page_t m; 3443 3444 runningbufwakeup(bp); 3445 if (!(bp->b_flags & B_VMIO)) 3446 return; 3447 3448 obj = bp->b_bufobj->bo_object; 3449 VM_OBJECT_LOCK(obj); 3450 for (i = 0; i < bp->b_npages; i++) { 3451 m = bp->b_pages[i]; 3452 if (m == bogus_page) { 3453 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3454 if (!m) 3455 panic("vfs_unbusy_pages: page missing\n"); 3456 bp->b_pages[i] = m; 3457 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3458 bp->b_pages, bp->b_npages); 3459 } 3460 vm_object_pip_subtract(obj, 1); 3461 vm_page_io_finish(m); 3462 } 3463 vm_object_pip_wakeupn(obj, 0); 3464 VM_OBJECT_UNLOCK(obj); 3465 } 3466 3467 /* 3468 * vfs_page_set_valid: 3469 * 3470 * Set the valid bits in a page based on the supplied offset. The 3471 * range is restricted to the buffer's size. 3472 * 3473 * This routine is typically called after a read completes. 3474 */ 3475 static void 3476 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3477 { 3478 vm_ooffset_t eoff; 3479 3480 /* 3481 * Compute the end offset, eoff, such that [off, eoff) does not span a 3482 * page boundary and eoff is not greater than the end of the buffer. 3483 * The end of the buffer, in this case, is our file EOF, not the 3484 * allocation size of the buffer. 3485 */ 3486 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3487 if (eoff > bp->b_offset + bp->b_bcount) 3488 eoff = bp->b_offset + bp->b_bcount; 3489 3490 /* 3491 * Set valid range. This is typically the entire buffer and thus the 3492 * entire page. 3493 */ 3494 if (eoff > off) 3495 vm_page_set_valid(m, off & PAGE_MASK, eoff - off); 3496 } 3497 3498 /* 3499 * vfs_page_set_validclean: 3500 * 3501 * Set the valid bits and clear the dirty bits in a page based on the 3502 * supplied offset. The range is restricted to the buffer's size. 3503 */ 3504 static void 3505 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3506 { 3507 vm_ooffset_t soff, eoff; 3508 3509 /* 3510 * Start and end offsets in buffer. eoff - soff may not cross a 3511 * page boundry or cross the end of the buffer. The end of the 3512 * buffer, in this case, is our file EOF, not the allocation size 3513 * of the buffer. 3514 */ 3515 soff = off; 3516 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3517 if (eoff > bp->b_offset + bp->b_bcount) 3518 eoff = bp->b_offset + bp->b_bcount; 3519 3520 /* 3521 * Set valid range. This is typically the entire buffer and thus the 3522 * entire page. 3523 */ 3524 if (eoff > soff) { 3525 vm_page_set_validclean( 3526 m, 3527 (vm_offset_t) (soff & PAGE_MASK), 3528 (vm_offset_t) (eoff - soff) 3529 ); 3530 } 3531 } 3532 3533 /* 3534 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If 3535 * any page is busy, drain the flag. 3536 */ 3537 static void 3538 vfs_drain_busy_pages(struct buf *bp) 3539 { 3540 vm_page_t m; 3541 int i, last_busied; 3542 3543 VM_OBJECT_LOCK_ASSERT(bp->b_bufobj->bo_object, MA_OWNED); 3544 last_busied = 0; 3545 for (i = 0; i < bp->b_npages; i++) { 3546 m = bp->b_pages[i]; 3547 if ((m->oflags & VPO_BUSY) != 0) { 3548 for (; last_busied < i; last_busied++) 3549 vm_page_busy(bp->b_pages[last_busied]); 3550 while ((m->oflags & VPO_BUSY) != 0) 3551 vm_page_sleep(m, "vbpage"); 3552 } 3553 } 3554 for (i = 0; i < last_busied; i++) 3555 vm_page_wakeup(bp->b_pages[i]); 3556 } 3557 3558 /* 3559 * This routine is called before a device strategy routine. 3560 * It is used to tell the VM system that paging I/O is in 3561 * progress, and treat the pages associated with the buffer 3562 * almost as being VPO_BUSY. Also the object paging_in_progress 3563 * flag is handled to make sure that the object doesn't become 3564 * inconsistant. 3565 * 3566 * Since I/O has not been initiated yet, certain buffer flags 3567 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3568 * and should be ignored. 3569 */ 3570 void 3571 vfs_busy_pages(struct buf *bp, int clear_modify) 3572 { 3573 int i, bogus; 3574 vm_object_t obj; 3575 vm_ooffset_t foff; 3576 vm_page_t m; 3577 3578 if (!(bp->b_flags & B_VMIO)) 3579 return; 3580 3581 obj = bp->b_bufobj->bo_object; 3582 foff = bp->b_offset; 3583 KASSERT(bp->b_offset != NOOFFSET, 3584 ("vfs_busy_pages: no buffer offset")); 3585 VM_OBJECT_LOCK(obj); 3586 vfs_drain_busy_pages(bp); 3587 if (bp->b_bufsize != 0) 3588 vfs_setdirty_locked_object(bp); 3589 bogus = 0; 3590 for (i = 0; i < bp->b_npages; i++) { 3591 m = bp->b_pages[i]; 3592 3593 if ((bp->b_flags & B_CLUSTER) == 0) { 3594 vm_object_pip_add(obj, 1); 3595 vm_page_io_start(m); 3596 } 3597 /* 3598 * When readying a buffer for a read ( i.e 3599 * clear_modify == 0 ), it is important to do 3600 * bogus_page replacement for valid pages in 3601 * partially instantiated buffers. Partially 3602 * instantiated buffers can, in turn, occur when 3603 * reconstituting a buffer from its VM backing store 3604 * base. We only have to do this if B_CACHE is 3605 * clear ( which causes the I/O to occur in the 3606 * first place ). The replacement prevents the read 3607 * I/O from overwriting potentially dirty VM-backed 3608 * pages. XXX bogus page replacement is, uh, bogus. 3609 * It may not work properly with small-block devices. 3610 * We need to find a better way. 3611 */ 3612 if (clear_modify) { 3613 pmap_remove_write(m); 3614 vfs_page_set_validclean(bp, foff, m); 3615 } else if (m->valid == VM_PAGE_BITS_ALL && 3616 (bp->b_flags & B_CACHE) == 0) { 3617 bp->b_pages[i] = bogus_page; 3618 bogus++; 3619 } 3620 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3621 } 3622 VM_OBJECT_UNLOCK(obj); 3623 if (bogus) 3624 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3625 bp->b_pages, bp->b_npages); 3626 } 3627 3628 /* 3629 * vfs_bio_set_valid: 3630 * 3631 * Set the range within the buffer to valid. The range is 3632 * relative to the beginning of the buffer, b_offset. Note that 3633 * b_offset itself may be offset from the beginning of the first 3634 * page. 3635 */ 3636 void 3637 vfs_bio_set_valid(struct buf *bp, int base, int size) 3638 { 3639 int i, n; 3640 vm_page_t m; 3641 3642 if (!(bp->b_flags & B_VMIO)) 3643 return; 3644 3645 /* 3646 * Fixup base to be relative to beginning of first page. 3647 * Set initial n to be the maximum number of bytes in the 3648 * first page that can be validated. 3649 */ 3650 base += (bp->b_offset & PAGE_MASK); 3651 n = PAGE_SIZE - (base & PAGE_MASK); 3652 3653 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3654 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3655 m = bp->b_pages[i]; 3656 if (n > size) 3657 n = size; 3658 vm_page_set_valid(m, base & PAGE_MASK, n); 3659 base += n; 3660 size -= n; 3661 n = PAGE_SIZE; 3662 } 3663 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3664 } 3665 3666 /* 3667 * vfs_bio_clrbuf: 3668 * 3669 * If the specified buffer is a non-VMIO buffer, clear the entire 3670 * buffer. If the specified buffer is a VMIO buffer, clear and 3671 * validate only the previously invalid portions of the buffer. 3672 * This routine essentially fakes an I/O, so we need to clear 3673 * BIO_ERROR and B_INVAL. 3674 * 3675 * Note that while we only theoretically need to clear through b_bcount, 3676 * we go ahead and clear through b_bufsize. 3677 */ 3678 void 3679 vfs_bio_clrbuf(struct buf *bp) 3680 { 3681 int i, j, mask; 3682 caddr_t sa, ea; 3683 3684 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3685 clrbuf(bp); 3686 return; 3687 } 3688 bp->b_flags &= ~B_INVAL; 3689 bp->b_ioflags &= ~BIO_ERROR; 3690 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3691 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3692 (bp->b_offset & PAGE_MASK) == 0) { 3693 if (bp->b_pages[0] == bogus_page) 3694 goto unlock; 3695 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3696 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3697 if ((bp->b_pages[0]->valid & mask) == mask) 3698 goto unlock; 3699 if ((bp->b_pages[0]->valid & mask) == 0) { 3700 bzero(bp->b_data, bp->b_bufsize); 3701 bp->b_pages[0]->valid |= mask; 3702 goto unlock; 3703 } 3704 } 3705 ea = sa = bp->b_data; 3706 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3707 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3708 ea = (caddr_t)(vm_offset_t)ulmin( 3709 (u_long)(vm_offset_t)ea, 3710 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3711 if (bp->b_pages[i] == bogus_page) 3712 continue; 3713 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3714 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3715 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3716 if ((bp->b_pages[i]->valid & mask) == mask) 3717 continue; 3718 if ((bp->b_pages[i]->valid & mask) == 0) 3719 bzero(sa, ea - sa); 3720 else { 3721 for (; sa < ea; sa += DEV_BSIZE, j++) { 3722 if ((bp->b_pages[i]->valid & (1 << j)) == 0) 3723 bzero(sa, DEV_BSIZE); 3724 } 3725 } 3726 bp->b_pages[i]->valid |= mask; 3727 } 3728 unlock: 3729 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3730 bp->b_resid = 0; 3731 } 3732 3733 /* 3734 * vm_hold_load_pages and vm_hold_free_pages get pages into 3735 * a buffers address space. The pages are anonymous and are 3736 * not associated with a file object. 3737 */ 3738 static void 3739 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3740 { 3741 vm_offset_t pg; 3742 vm_page_t p; 3743 int index; 3744 3745 to = round_page(to); 3746 from = round_page(from); 3747 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3748 3749 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3750 tryagain: 3751 /* 3752 * note: must allocate system pages since blocking here 3753 * could interfere with paging I/O, no matter which 3754 * process we are. 3755 */ 3756 p = vm_page_alloc(NULL, pg >> PAGE_SHIFT, VM_ALLOC_NOOBJ | 3757 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3758 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 3759 if (!p) { 3760 VM_WAIT; 3761 goto tryagain; 3762 } 3763 pmap_qenter(pg, &p, 1); 3764 bp->b_pages[index] = p; 3765 } 3766 bp->b_npages = index; 3767 } 3768 3769 /* Return pages associated with this buf to the vm system */ 3770 static void 3771 vm_hold_free_pages(struct buf *bp, int newbsize) 3772 { 3773 vm_offset_t from; 3774 vm_page_t p; 3775 int index, newnpages; 3776 3777 from = round_page((vm_offset_t)bp->b_data + newbsize); 3778 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3779 if (bp->b_npages > newnpages) 3780 pmap_qremove(from, bp->b_npages - newnpages); 3781 for (index = newnpages; index < bp->b_npages; index++) { 3782 p = bp->b_pages[index]; 3783 bp->b_pages[index] = NULL; 3784 if (p->busy != 0) 3785 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3786 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 3787 p->wire_count--; 3788 vm_page_free(p); 3789 atomic_subtract_int(&cnt.v_wire_count, 1); 3790 } 3791 bp->b_npages = newnpages; 3792 } 3793 3794 /* 3795 * Map an IO request into kernel virtual address space. 3796 * 3797 * All requests are (re)mapped into kernel VA space. 3798 * Notice that we use b_bufsize for the size of the buffer 3799 * to be mapped. b_bcount might be modified by the driver. 3800 * 3801 * Note that even if the caller determines that the address space should 3802 * be valid, a race or a smaller-file mapped into a larger space may 3803 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3804 * check the return value. 3805 */ 3806 int 3807 vmapbuf(struct buf *bp) 3808 { 3809 caddr_t addr, kva; 3810 vm_prot_t prot; 3811 int pidx, i; 3812 struct vm_page *m; 3813 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3814 3815 if (bp->b_bufsize < 0) 3816 return (-1); 3817 prot = VM_PROT_READ; 3818 if (bp->b_iocmd == BIO_READ) 3819 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3820 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3821 addr < bp->b_data + bp->b_bufsize; 3822 addr += PAGE_SIZE, pidx++) { 3823 /* 3824 * Do the vm_fault if needed; do the copy-on-write thing 3825 * when reading stuff off device into memory. 3826 * 3827 * NOTE! Must use pmap_extract() because addr may be in 3828 * the userland address space, and kextract is only guarenteed 3829 * to work for the kernland address space (see: sparc64 port). 3830 */ 3831 retry: 3832 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3833 prot) < 0) { 3834 for (i = 0; i < pidx; ++i) { 3835 vm_page_lock(bp->b_pages[i]); 3836 vm_page_unhold(bp->b_pages[i]); 3837 vm_page_unlock(bp->b_pages[i]); 3838 bp->b_pages[i] = NULL; 3839 } 3840 return(-1); 3841 } 3842 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3843 if (m == NULL) 3844 goto retry; 3845 bp->b_pages[pidx] = m; 3846 } 3847 if (pidx > btoc(MAXPHYS)) 3848 panic("vmapbuf: mapped more than MAXPHYS"); 3849 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3850 3851 kva = bp->b_saveaddr; 3852 bp->b_npages = pidx; 3853 bp->b_saveaddr = bp->b_data; 3854 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3855 return(0); 3856 } 3857 3858 /* 3859 * Free the io map PTEs associated with this IO operation. 3860 * We also invalidate the TLB entries and restore the original b_addr. 3861 */ 3862 void 3863 vunmapbuf(struct buf *bp) 3864 { 3865 int pidx; 3866 int npages; 3867 3868 npages = bp->b_npages; 3869 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3870 for (pidx = 0; pidx < npages; pidx++) { 3871 vm_page_lock(bp->b_pages[pidx]); 3872 vm_page_unhold(bp->b_pages[pidx]); 3873 vm_page_unlock(bp->b_pages[pidx]); 3874 } 3875 3876 bp->b_data = bp->b_saveaddr; 3877 } 3878 3879 void 3880 bdone(struct buf *bp) 3881 { 3882 struct mtx *mtxp; 3883 3884 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3885 mtx_lock(mtxp); 3886 bp->b_flags |= B_DONE; 3887 wakeup(bp); 3888 mtx_unlock(mtxp); 3889 } 3890 3891 void 3892 bwait(struct buf *bp, u_char pri, const char *wchan) 3893 { 3894 struct mtx *mtxp; 3895 3896 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3897 mtx_lock(mtxp); 3898 while ((bp->b_flags & B_DONE) == 0) 3899 msleep(bp, mtxp, pri, wchan, 0); 3900 mtx_unlock(mtxp); 3901 } 3902 3903 int 3904 bufsync(struct bufobj *bo, int waitfor) 3905 { 3906 3907 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 3908 } 3909 3910 void 3911 bufstrategy(struct bufobj *bo, struct buf *bp) 3912 { 3913 int i = 0; 3914 struct vnode *vp; 3915 3916 vp = bp->b_vp; 3917 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3918 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3919 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3920 i = VOP_STRATEGY(vp, bp); 3921 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3922 } 3923 3924 void 3925 bufobj_wrefl(struct bufobj *bo) 3926 { 3927 3928 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3929 ASSERT_BO_LOCKED(bo); 3930 bo->bo_numoutput++; 3931 } 3932 3933 void 3934 bufobj_wref(struct bufobj *bo) 3935 { 3936 3937 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3938 BO_LOCK(bo); 3939 bo->bo_numoutput++; 3940 BO_UNLOCK(bo); 3941 } 3942 3943 void 3944 bufobj_wdrop(struct bufobj *bo) 3945 { 3946 3947 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3948 BO_LOCK(bo); 3949 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3950 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3951 bo->bo_flag &= ~BO_WWAIT; 3952 wakeup(&bo->bo_numoutput); 3953 } 3954 BO_UNLOCK(bo); 3955 } 3956 3957 int 3958 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3959 { 3960 int error; 3961 3962 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3963 ASSERT_BO_LOCKED(bo); 3964 error = 0; 3965 while (bo->bo_numoutput) { 3966 bo->bo_flag |= BO_WWAIT; 3967 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3968 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3969 if (error) 3970 break; 3971 } 3972 return (error); 3973 } 3974 3975 void 3976 bpin(struct buf *bp) 3977 { 3978 struct mtx *mtxp; 3979 3980 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3981 mtx_lock(mtxp); 3982 bp->b_pin_count++; 3983 mtx_unlock(mtxp); 3984 } 3985 3986 void 3987 bunpin(struct buf *bp) 3988 { 3989 struct mtx *mtxp; 3990 3991 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3992 mtx_lock(mtxp); 3993 if (--bp->b_pin_count == 0) 3994 wakeup(bp); 3995 mtx_unlock(mtxp); 3996 } 3997 3998 void 3999 bunpin_wait(struct buf *bp) 4000 { 4001 struct mtx *mtxp; 4002 4003 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4004 mtx_lock(mtxp); 4005 while (bp->b_pin_count > 0) 4006 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4007 mtx_unlock(mtxp); 4008 } 4009 4010 #include "opt_ddb.h" 4011 #ifdef DDB 4012 #include <ddb/ddb.h> 4013 4014 /* DDB command to show buffer data */ 4015 DB_SHOW_COMMAND(buffer, db_show_buffer) 4016 { 4017 /* get args */ 4018 struct buf *bp = (struct buf *)addr; 4019 4020 if (!have_addr) { 4021 db_printf("usage: show buffer <addr>\n"); 4022 return; 4023 } 4024 4025 db_printf("buf at %p\n", bp); 4026 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4027 db_printf( 4028 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4029 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n", 4030 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4031 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4032 bp->b_dep.lh_first); 4033 if (bp->b_npages) { 4034 int i; 4035 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4036 for (i = 0; i < bp->b_npages; i++) { 4037 vm_page_t m; 4038 m = bp->b_pages[i]; 4039 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4040 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4041 if ((i + 1) < bp->b_npages) 4042 db_printf(","); 4043 } 4044 db_printf("\n"); 4045 } 4046 db_printf(" "); 4047 lockmgr_printinfo(&bp->b_lock); 4048 } 4049 4050 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4051 { 4052 struct buf *bp; 4053 int i; 4054 4055 for (i = 0; i < nbuf; i++) { 4056 bp = &buf[i]; 4057 if (BUF_ISLOCKED(bp)) { 4058 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4059 db_printf("\n"); 4060 } 4061 } 4062 } 4063 4064 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4065 { 4066 struct vnode *vp; 4067 struct buf *bp; 4068 4069 if (!have_addr) { 4070 db_printf("usage: show vnodebufs <addr>\n"); 4071 return; 4072 } 4073 vp = (struct vnode *)addr; 4074 db_printf("Clean buffers:\n"); 4075 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4076 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4077 db_printf("\n"); 4078 } 4079 db_printf("Dirty buffers:\n"); 4080 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4081 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4082 db_printf("\n"); 4083 } 4084 } 4085 4086 DB_COMMAND(countfreebufs, db_coundfreebufs) 4087 { 4088 struct buf *bp; 4089 int i, used = 0, nfree = 0; 4090 4091 if (have_addr) { 4092 db_printf("usage: countfreebufs\n"); 4093 return; 4094 } 4095 4096 for (i = 0; i < nbuf; i++) { 4097 bp = &buf[i]; 4098 if ((bp->b_vflags & BV_INFREECNT) != 0) 4099 nfree++; 4100 else 4101 used++; 4102 } 4103 4104 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4105 nfree + used); 4106 db_printf("numfreebuffers is %d\n", numfreebuffers); 4107 } 4108 #endif /* DDB */ 4109