vfs_bio.c revision 210217
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: head/sys/kern/vfs_bio.c 210217 2010-07-18 10:15:33Z ivoras $"); 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 77static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 78 79struct bio_ops bioops; /* I/O operation notification */ 80 81struct 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 */ 93struct buf *buf; /* buffer header pool */ 94 95static struct proc *bufdaemonproc; 96 97static int inmem(struct vnode *vp, daddr_t blkno); 98static void vm_hold_free_pages(struct buf *bp, int newbsize); 99static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 100 vm_offset_t to); 101static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 102static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 103 vm_page_t m); 104static void vfs_drain_busy_pages(struct buf *bp); 105static void vfs_clean_pages_dirty_buf(struct buf *bp); 106static void vfs_setdirty_locked_object(struct buf *bp); 107static void vfs_vmio_release(struct buf *bp); 108static int vfs_bio_clcheck(struct vnode *vp, int size, 109 daddr_t lblkno, daddr_t blkno); 110static int buf_do_flush(struct vnode *vp); 111static int flushbufqueues(struct vnode *, int, int); 112static void buf_daemon(void); 113static void bremfreel(struct buf *bp); 114#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 115 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 116static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 117#endif 118 119int vmiodirenable = TRUE; 120SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 121 "Use the VM system for directory writes"); 122long runningbufspace; 123SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 124 "Amount of presently outstanding async buffer io"); 125static long bufspace; 126#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 127 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 128SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 129 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 130#else 131SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 132 "Virtual memory used for buffers"); 133#endif 134static long maxbufspace; 135SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 136 "Maximum allowed value of bufspace (including buf_daemon)"); 137static long bufmallocspace; 138SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 139 "Amount of malloced memory for buffers"); 140static long maxbufmallocspace; 141SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 142 "Maximum amount of malloced memory for buffers"); 143static long lobufspace; 144SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 145 "Minimum amount of buffers we want to have"); 146long hibufspace; 147SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 148 "Maximum allowed value of bufspace (excluding buf_daemon)"); 149static int bufreusecnt; 150SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 151 "Number of times we have reused a buffer"); 152static int buffreekvacnt; 153SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 154 "Number of times we have freed the KVA space from some buffer"); 155static int bufdefragcnt; 156SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 157 "Number of times we have had to repeat buffer allocation to defragment"); 158static long lorunningspace; 159SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 160 "Minimum preferred space used for in-progress I/O"); 161static long hirunningspace; 162SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 163 "Maximum amount of space to use for in-progress I/O"); 164int dirtybufferflushes; 165SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 166 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 167int bdwriteskip; 168SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 169 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 170int altbufferflushes; 171SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 172 0, "Number of fsync flushes to limit dirty buffers"); 173static int recursiveflushes; 174SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 175 0, "Number of flushes skipped due to being recursive"); 176static int numdirtybuffers; 177SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 178 "Number of buffers that are dirty (has unwritten changes) at the moment"); 179static int lodirtybuffers; 180SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 181 "How many buffers we want to have free before bufdaemon can sleep"); 182static int hidirtybuffers; 183SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 184 "When the number of dirty buffers is considered severe"); 185int dirtybufthresh; 186SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 187 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 188static int numfreebuffers; 189SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 190 "Number of free buffers"); 191static int lofreebuffers; 192SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 193 "XXX Unused"); 194static int hifreebuffers; 195SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 196 "XXX Complicatedly unused"); 197static int getnewbufcalls; 198SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 199 "Number of calls to getnewbuf"); 200static int getnewbufrestarts; 201SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 202 "Number of times getnewbuf has had to restart a buffer aquisition"); 203static int flushbufqtarget = 100; 204SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 205 "Amount of work to do in flushbufqueues when helping bufdaemon"); 206static long notbufdflashes; 207SYSCTL_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 */ 215static 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 */ 223static int bd_speedupreq; 224 225/* 226 * This lock synchronizes access to bd_request. 227 */ 228static 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 */ 237vm_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 */ 244static int runningbufreq; 245 246/* 247 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 248 * waitrunningbufspace(). 249 */ 250static 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 */ 259static int needsbuffer; 260 261/* 262 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 263 */ 264static 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 */ 280static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 281 282/* Lock for the bufqueues */ 283static 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 */ 289const 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) 298static int 299sysctl_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 316extern 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 325static __inline void 326numdirtywakeup(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 348static __inline void 349bufspacewakeup(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 */ 369void 370runningbufwakeup(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 394static __inline void 395bufcountwakeup(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 */ 430void 431waitrunningbufspace(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 */ 450static __inline 451void 452vfs_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 */ 466static __inline 467void 468bd_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 483void 484bd_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 */ 505caddr_t 506kern_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. */ 577void 578bufinit(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 lorunningspace = 512 * 1024; 624 hirunningspace = lmin(roundup(hibufspace/64, MAXBSIZE), 16*1024*1024); 625 if (hirunningspace < 1024 * 1024) 626 hirunningspace = 1024 * 1024; 627 628/* 629 * Limit the amount of malloc memory since it is wired permanently into 630 * the kernel space. Even though this is accounted for in the buffer 631 * allocation, we don't want the malloced region to grow uncontrolled. 632 * The malloc scheme improves memory utilization significantly on average 633 * (small) directories. 634 */ 635 maxbufmallocspace = hibufspace / 20; 636 637/* 638 * Reduce the chance of a deadlock occuring by limiting the number 639 * of delayed-write dirty buffers we allow to stack up. 640 */ 641 hidirtybuffers = nbuf / 4 + 20; 642 dirtybufthresh = hidirtybuffers * 9 / 10; 643 numdirtybuffers = 0; 644/* 645 * To support extreme low-memory systems, make sure hidirtybuffers cannot 646 * eat up all available buffer space. This occurs when our minimum cannot 647 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 648 * BKVASIZE'd (8K) buffers. 649 */ 650 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 651 hidirtybuffers >>= 1; 652 } 653 lodirtybuffers = hidirtybuffers / 2; 654 655/* 656 * Try to keep the number of free buffers in the specified range, 657 * and give special processes (e.g. like buf_daemon) access to an 658 * emergency reserve. 659 */ 660 lofreebuffers = nbuf / 18 + 5; 661 hifreebuffers = 2 * lofreebuffers; 662 numfreebuffers = nbuf; 663 664 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 665 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 666} 667 668/* 669 * bfreekva() - free the kva allocation for a buffer. 670 * 671 * Since this call frees up buffer space, we call bufspacewakeup(). 672 */ 673static void 674bfreekva(struct buf *bp) 675{ 676 677 if (bp->b_kvasize) { 678 atomic_add_int(&buffreekvacnt, 1); 679 atomic_subtract_long(&bufspace, bp->b_kvasize); 680 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 681 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 682 bp->b_kvasize = 0; 683 bufspacewakeup(); 684 } 685} 686 687/* 688 * bremfree: 689 * 690 * Mark the buffer for removal from the appropriate free list in brelse. 691 * 692 */ 693void 694bremfree(struct buf *bp) 695{ 696 int old; 697 698 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 699 KASSERT((bp->b_flags & B_REMFREE) == 0, 700 ("bremfree: buffer %p already marked for delayed removal.", bp)); 701 KASSERT(bp->b_qindex != QUEUE_NONE, 702 ("bremfree: buffer %p not on a queue.", bp)); 703 BUF_ASSERT_HELD(bp); 704 705 bp->b_flags |= B_REMFREE; 706 /* Fixup numfreebuffers count. */ 707 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 708 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 709 ("buf %p not counted in numfreebuffers", bp)); 710 bp->b_vflags &= ~BV_INFREECNT; 711 old = atomic_fetchadd_int(&numfreebuffers, -1); 712 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 713 } 714} 715 716/* 717 * bremfreef: 718 * 719 * Force an immediate removal from a free list. Used only in nfs when 720 * it abuses the b_freelist pointer. 721 */ 722void 723bremfreef(struct buf *bp) 724{ 725 mtx_lock(&bqlock); 726 bremfreel(bp); 727 mtx_unlock(&bqlock); 728} 729 730/* 731 * bremfreel: 732 * 733 * Removes a buffer from the free list, must be called with the 734 * bqlock held. 735 */ 736static void 737bremfreel(struct buf *bp) 738{ 739 int old; 740 741 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 742 bp, bp->b_vp, bp->b_flags); 743 KASSERT(bp->b_qindex != QUEUE_NONE, 744 ("bremfreel: buffer %p not on a queue.", bp)); 745 BUF_ASSERT_HELD(bp); 746 mtx_assert(&bqlock, MA_OWNED); 747 748 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 749 bp->b_qindex = QUEUE_NONE; 750 /* 751 * If this was a delayed bremfree() we only need to remove the buffer 752 * from the queue and return the stats are already done. 753 */ 754 if (bp->b_flags & B_REMFREE) { 755 bp->b_flags &= ~B_REMFREE; 756 return; 757 } 758 /* 759 * Fixup numfreebuffers count. If the buffer is invalid or not 760 * delayed-write, the buffer was free and we must decrement 761 * numfreebuffers. 762 */ 763 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 764 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 765 ("buf %p not counted in numfreebuffers", bp)); 766 bp->b_vflags &= ~BV_INFREECNT; 767 old = atomic_fetchadd_int(&numfreebuffers, -1); 768 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 769 } 770} 771 772 773/* 774 * Get a buffer with the specified data. Look in the cache first. We 775 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 776 * is set, the buffer is valid and we do not have to do anything ( see 777 * getblk() ). This is really just a special case of breadn(). 778 */ 779int 780bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 781 struct buf **bpp) 782{ 783 784 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 785} 786 787/* 788 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 789 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 790 * the buffer is valid and we do not have to do anything. 791 */ 792void 793breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 794 int cnt, struct ucred * cred) 795{ 796 struct buf *rabp; 797 int i; 798 799 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 800 if (inmem(vp, *rablkno)) 801 continue; 802 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 803 804 if ((rabp->b_flags & B_CACHE) == 0) { 805 if (!TD_IS_IDLETHREAD(curthread)) 806 curthread->td_ru.ru_inblock++; 807 rabp->b_flags |= B_ASYNC; 808 rabp->b_flags &= ~B_INVAL; 809 rabp->b_ioflags &= ~BIO_ERROR; 810 rabp->b_iocmd = BIO_READ; 811 if (rabp->b_rcred == NOCRED && cred != NOCRED) 812 rabp->b_rcred = crhold(cred); 813 vfs_busy_pages(rabp, 0); 814 BUF_KERNPROC(rabp); 815 rabp->b_iooffset = dbtob(rabp->b_blkno); 816 bstrategy(rabp); 817 } else { 818 brelse(rabp); 819 } 820 } 821} 822 823/* 824 * Operates like bread, but also starts asynchronous I/O on 825 * read-ahead blocks. 826 */ 827int 828breadn(struct vnode * vp, daddr_t blkno, int size, 829 daddr_t * rablkno, int *rabsize, 830 int cnt, struct ucred * cred, struct buf **bpp) 831{ 832 struct buf *bp; 833 int rv = 0, readwait = 0; 834 835 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 836 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 837 838 /* if not found in cache, do some I/O */ 839 if ((bp->b_flags & B_CACHE) == 0) { 840 if (!TD_IS_IDLETHREAD(curthread)) 841 curthread->td_ru.ru_inblock++; 842 bp->b_iocmd = BIO_READ; 843 bp->b_flags &= ~B_INVAL; 844 bp->b_ioflags &= ~BIO_ERROR; 845 if (bp->b_rcred == NOCRED && cred != NOCRED) 846 bp->b_rcred = crhold(cred); 847 vfs_busy_pages(bp, 0); 848 bp->b_iooffset = dbtob(bp->b_blkno); 849 bstrategy(bp); 850 ++readwait; 851 } 852 853 breada(vp, rablkno, rabsize, cnt, cred); 854 855 if (readwait) { 856 rv = bufwait(bp); 857 } 858 return (rv); 859} 860 861/* 862 * Write, release buffer on completion. (Done by iodone 863 * if async). Do not bother writing anything if the buffer 864 * is invalid. 865 * 866 * Note that we set B_CACHE here, indicating that buffer is 867 * fully valid and thus cacheable. This is true even of NFS 868 * now so we set it generally. This could be set either here 869 * or in biodone() since the I/O is synchronous. We put it 870 * here. 871 */ 872int 873bufwrite(struct buf *bp) 874{ 875 int oldflags; 876 struct vnode *vp; 877 int vp_md; 878 879 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 880 if (bp->b_flags & B_INVAL) { 881 brelse(bp); 882 return (0); 883 } 884 885 oldflags = bp->b_flags; 886 887 BUF_ASSERT_HELD(bp); 888 889 if (bp->b_pin_count > 0) 890 bunpin_wait(bp); 891 892 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 893 ("FFS background buffer should not get here %p", bp)); 894 895 vp = bp->b_vp; 896 if (vp) 897 vp_md = vp->v_vflag & VV_MD; 898 else 899 vp_md = 0; 900 901 /* Mark the buffer clean */ 902 bundirty(bp); 903 904 bp->b_flags &= ~B_DONE; 905 bp->b_ioflags &= ~BIO_ERROR; 906 bp->b_flags |= B_CACHE; 907 bp->b_iocmd = BIO_WRITE; 908 909 bufobj_wref(bp->b_bufobj); 910 vfs_busy_pages(bp, 1); 911 912 /* 913 * Normal bwrites pipeline writes 914 */ 915 bp->b_runningbufspace = bp->b_bufsize; 916 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 917 918 if (!TD_IS_IDLETHREAD(curthread)) 919 curthread->td_ru.ru_oublock++; 920 if (oldflags & B_ASYNC) 921 BUF_KERNPROC(bp); 922 bp->b_iooffset = dbtob(bp->b_blkno); 923 bstrategy(bp); 924 925 if ((oldflags & B_ASYNC) == 0) { 926 int rtval = bufwait(bp); 927 brelse(bp); 928 return (rtval); 929 } else { 930 /* 931 * don't allow the async write to saturate the I/O 932 * system. We will not deadlock here because 933 * we are blocking waiting for I/O that is already in-progress 934 * to complete. We do not block here if it is the update 935 * or syncer daemon trying to clean up as that can lead 936 * to deadlock. 937 */ 938 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 939 waitrunningbufspace(); 940 } 941 942 return (0); 943} 944 945void 946bufbdflush(struct bufobj *bo, struct buf *bp) 947{ 948 struct buf *nbp; 949 950 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 951 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 952 altbufferflushes++; 953 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 954 BO_LOCK(bo); 955 /* 956 * Try to find a buffer to flush. 957 */ 958 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 959 if ((nbp->b_vflags & BV_BKGRDINPROG) || 960 BUF_LOCK(nbp, 961 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 962 continue; 963 if (bp == nbp) 964 panic("bdwrite: found ourselves"); 965 BO_UNLOCK(bo); 966 /* Don't countdeps with the bo lock held. */ 967 if (buf_countdeps(nbp, 0)) { 968 BO_LOCK(bo); 969 BUF_UNLOCK(nbp); 970 continue; 971 } 972 if (nbp->b_flags & B_CLUSTEROK) { 973 vfs_bio_awrite(nbp); 974 } else { 975 bremfree(nbp); 976 bawrite(nbp); 977 } 978 dirtybufferflushes++; 979 break; 980 } 981 if (nbp == NULL) 982 BO_UNLOCK(bo); 983 } 984} 985 986/* 987 * Delayed write. (Buffer is marked dirty). Do not bother writing 988 * anything if the buffer is marked invalid. 989 * 990 * Note that since the buffer must be completely valid, we can safely 991 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 992 * biodone() in order to prevent getblk from writing the buffer 993 * out synchronously. 994 */ 995void 996bdwrite(struct buf *bp) 997{ 998 struct thread *td = curthread; 999 struct vnode *vp; 1000 struct bufobj *bo; 1001 1002 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1003 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1004 BUF_ASSERT_HELD(bp); 1005 1006 if (bp->b_flags & B_INVAL) { 1007 brelse(bp); 1008 return; 1009 } 1010 1011 /* 1012 * If we have too many dirty buffers, don't create any more. 1013 * If we are wildly over our limit, then force a complete 1014 * cleanup. Otherwise, just keep the situation from getting 1015 * out of control. Note that we have to avoid a recursive 1016 * disaster and not try to clean up after our own cleanup! 1017 */ 1018 vp = bp->b_vp; 1019 bo = bp->b_bufobj; 1020 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1021 td->td_pflags |= TDP_INBDFLUSH; 1022 BO_BDFLUSH(bo, bp); 1023 td->td_pflags &= ~TDP_INBDFLUSH; 1024 } else 1025 recursiveflushes++; 1026 1027 bdirty(bp); 1028 /* 1029 * Set B_CACHE, indicating that the buffer is fully valid. This is 1030 * true even of NFS now. 1031 */ 1032 bp->b_flags |= B_CACHE; 1033 1034 /* 1035 * This bmap keeps the system from needing to do the bmap later, 1036 * perhaps when the system is attempting to do a sync. Since it 1037 * is likely that the indirect block -- or whatever other datastructure 1038 * that the filesystem needs is still in memory now, it is a good 1039 * thing to do this. Note also, that if the pageout daemon is 1040 * requesting a sync -- there might not be enough memory to do 1041 * the bmap then... So, this is important to do. 1042 */ 1043 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1044 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1045 } 1046 1047 /* 1048 * Set the *dirty* buffer range based upon the VM system dirty 1049 * pages. 1050 * 1051 * Mark the buffer pages as clean. We need to do this here to 1052 * satisfy the vnode_pager and the pageout daemon, so that it 1053 * thinks that the pages have been "cleaned". Note that since 1054 * the pages are in a delayed write buffer -- the VFS layer 1055 * "will" see that the pages get written out on the next sync, 1056 * or perhaps the cluster will be completed. 1057 */ 1058 vfs_clean_pages_dirty_buf(bp); 1059 bqrelse(bp); 1060 1061 /* 1062 * Wakeup the buffer flushing daemon if we have a lot of dirty 1063 * buffers (midpoint between our recovery point and our stall 1064 * point). 1065 */ 1066 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1067 1068 /* 1069 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1070 * due to the softdep code. 1071 */ 1072} 1073 1074/* 1075 * bdirty: 1076 * 1077 * Turn buffer into delayed write request. We must clear BIO_READ and 1078 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1079 * itself to properly update it in the dirty/clean lists. We mark it 1080 * B_DONE to ensure that any asynchronization of the buffer properly 1081 * clears B_DONE ( else a panic will occur later ). 1082 * 1083 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1084 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1085 * should only be called if the buffer is known-good. 1086 * 1087 * Since the buffer is not on a queue, we do not update the numfreebuffers 1088 * count. 1089 * 1090 * The buffer must be on QUEUE_NONE. 1091 */ 1092void 1093bdirty(struct buf *bp) 1094{ 1095 1096 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1097 bp, bp->b_vp, bp->b_flags); 1098 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1099 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1100 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1101 BUF_ASSERT_HELD(bp); 1102 bp->b_flags &= ~(B_RELBUF); 1103 bp->b_iocmd = BIO_WRITE; 1104 1105 if ((bp->b_flags & B_DELWRI) == 0) { 1106 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1107 reassignbuf(bp); 1108 atomic_add_int(&numdirtybuffers, 1); 1109 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1110 } 1111} 1112 1113/* 1114 * bundirty: 1115 * 1116 * Clear B_DELWRI for buffer. 1117 * 1118 * Since the buffer is not on a queue, we do not update the numfreebuffers 1119 * count. 1120 * 1121 * The buffer must be on QUEUE_NONE. 1122 */ 1123 1124void 1125bundirty(struct buf *bp) 1126{ 1127 1128 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1129 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1130 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1131 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1132 BUF_ASSERT_HELD(bp); 1133 1134 if (bp->b_flags & B_DELWRI) { 1135 bp->b_flags &= ~B_DELWRI; 1136 reassignbuf(bp); 1137 atomic_subtract_int(&numdirtybuffers, 1); 1138 numdirtywakeup(lodirtybuffers); 1139 } 1140 /* 1141 * Since it is now being written, we can clear its deferred write flag. 1142 */ 1143 bp->b_flags &= ~B_DEFERRED; 1144} 1145 1146/* 1147 * bawrite: 1148 * 1149 * Asynchronous write. Start output on a buffer, but do not wait for 1150 * it to complete. The buffer is released when the output completes. 1151 * 1152 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1153 * B_INVAL buffers. Not us. 1154 */ 1155void 1156bawrite(struct buf *bp) 1157{ 1158 1159 bp->b_flags |= B_ASYNC; 1160 (void) bwrite(bp); 1161} 1162 1163/* 1164 * bwillwrite: 1165 * 1166 * Called prior to the locking of any vnodes when we are expecting to 1167 * write. We do not want to starve the buffer cache with too many 1168 * dirty buffers so we block here. By blocking prior to the locking 1169 * of any vnodes we attempt to avoid the situation where a locked vnode 1170 * prevents the various system daemons from flushing related buffers. 1171 */ 1172 1173void 1174bwillwrite(void) 1175{ 1176 1177 if (numdirtybuffers >= hidirtybuffers) { 1178 mtx_lock(&nblock); 1179 while (numdirtybuffers >= hidirtybuffers) { 1180 bd_wakeup(1); 1181 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1182 msleep(&needsbuffer, &nblock, 1183 (PRIBIO + 4), "flswai", 0); 1184 } 1185 mtx_unlock(&nblock); 1186 } 1187} 1188 1189/* 1190 * Return true if we have too many dirty buffers. 1191 */ 1192int 1193buf_dirty_count_severe(void) 1194{ 1195 1196 return(numdirtybuffers >= hidirtybuffers); 1197} 1198 1199static __noinline int 1200buf_vm_page_count_severe(void) 1201{ 1202 1203 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1204 1205 return vm_page_count_severe(); 1206} 1207 1208/* 1209 * brelse: 1210 * 1211 * Release a busy buffer and, if requested, free its resources. The 1212 * buffer will be stashed in the appropriate bufqueue[] allowing it 1213 * to be accessed later as a cache entity or reused for other purposes. 1214 */ 1215void 1216brelse(struct buf *bp) 1217{ 1218 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1219 bp, bp->b_vp, bp->b_flags); 1220 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1221 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1222 1223 if (bp->b_flags & B_MANAGED) { 1224 bqrelse(bp); 1225 return; 1226 } 1227 1228 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1229 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1230 /* 1231 * Failed write, redirty. Must clear BIO_ERROR to prevent 1232 * pages from being scrapped. If the error is anything 1233 * other than an I/O error (EIO), assume that retrying 1234 * is futile. 1235 */ 1236 bp->b_ioflags &= ~BIO_ERROR; 1237 bdirty(bp); 1238 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1239 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1240 /* 1241 * Either a failed I/O or we were asked to free or not 1242 * cache the buffer. 1243 */ 1244 bp->b_flags |= B_INVAL; 1245 if (!LIST_EMPTY(&bp->b_dep)) 1246 buf_deallocate(bp); 1247 if (bp->b_flags & B_DELWRI) { 1248 atomic_subtract_int(&numdirtybuffers, 1); 1249 numdirtywakeup(lodirtybuffers); 1250 } 1251 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1252 if ((bp->b_flags & B_VMIO) == 0) { 1253 if (bp->b_bufsize) 1254 allocbuf(bp, 0); 1255 if (bp->b_vp) 1256 brelvp(bp); 1257 } 1258 } 1259 1260 /* 1261 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1262 * is called with B_DELWRI set, the underlying pages may wind up 1263 * getting freed causing a previous write (bdwrite()) to get 'lost' 1264 * because pages associated with a B_DELWRI bp are marked clean. 1265 * 1266 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1267 * if B_DELWRI is set. 1268 * 1269 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1270 * on pages to return pages to the VM page queues. 1271 */ 1272 if (bp->b_flags & B_DELWRI) 1273 bp->b_flags &= ~B_RELBUF; 1274 else if (buf_vm_page_count_severe()) { 1275 /* 1276 * The locking of the BO_LOCK is not necessary since 1277 * BKGRDINPROG cannot be set while we hold the buf 1278 * lock, it can only be cleared if it is already 1279 * pending. 1280 */ 1281 if (bp->b_vp) { 1282 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1283 bp->b_flags |= B_RELBUF; 1284 } else 1285 bp->b_flags |= B_RELBUF; 1286 } 1287 1288 /* 1289 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1290 * constituted, not even NFS buffers now. Two flags effect this. If 1291 * B_INVAL, the struct buf is invalidated but the VM object is kept 1292 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1293 * 1294 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1295 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1296 * buffer is also B_INVAL because it hits the re-dirtying code above. 1297 * 1298 * Normally we can do this whether a buffer is B_DELWRI or not. If 1299 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1300 * the commit state and we cannot afford to lose the buffer. If the 1301 * buffer has a background write in progress, we need to keep it 1302 * around to prevent it from being reconstituted and starting a second 1303 * background write. 1304 */ 1305 if ((bp->b_flags & B_VMIO) 1306 && !(bp->b_vp->v_mount != NULL && 1307 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1308 !vn_isdisk(bp->b_vp, NULL) && 1309 (bp->b_flags & B_DELWRI)) 1310 ) { 1311 1312 int i, j, resid; 1313 vm_page_t m; 1314 off_t foff; 1315 vm_pindex_t poff; 1316 vm_object_t obj; 1317 1318 obj = bp->b_bufobj->bo_object; 1319 1320 /* 1321 * Get the base offset and length of the buffer. Note that 1322 * in the VMIO case if the buffer block size is not 1323 * page-aligned then b_data pointer may not be page-aligned. 1324 * But our b_pages[] array *IS* page aligned. 1325 * 1326 * block sizes less then DEV_BSIZE (usually 512) are not 1327 * supported due to the page granularity bits (m->valid, 1328 * m->dirty, etc...). 1329 * 1330 * See man buf(9) for more information 1331 */ 1332 resid = bp->b_bufsize; 1333 foff = bp->b_offset; 1334 VM_OBJECT_LOCK(obj); 1335 for (i = 0; i < bp->b_npages; i++) { 1336 int had_bogus = 0; 1337 1338 m = bp->b_pages[i]; 1339 1340 /* 1341 * If we hit a bogus page, fixup *all* the bogus pages 1342 * now. 1343 */ 1344 if (m == bogus_page) { 1345 poff = OFF_TO_IDX(bp->b_offset); 1346 had_bogus = 1; 1347 1348 for (j = i; j < bp->b_npages; j++) { 1349 vm_page_t mtmp; 1350 mtmp = bp->b_pages[j]; 1351 if (mtmp == bogus_page) { 1352 mtmp = vm_page_lookup(obj, poff + j); 1353 if (!mtmp) { 1354 panic("brelse: page missing\n"); 1355 } 1356 bp->b_pages[j] = mtmp; 1357 } 1358 } 1359 1360 if ((bp->b_flags & B_INVAL) == 0) { 1361 pmap_qenter( 1362 trunc_page((vm_offset_t)bp->b_data), 1363 bp->b_pages, bp->b_npages); 1364 } 1365 m = bp->b_pages[i]; 1366 } 1367 if ((bp->b_flags & B_NOCACHE) || 1368 (bp->b_ioflags & BIO_ERROR && 1369 bp->b_iocmd == BIO_READ)) { 1370 int poffset = foff & PAGE_MASK; 1371 int presid = resid > (PAGE_SIZE - poffset) ? 1372 (PAGE_SIZE - poffset) : resid; 1373 1374 KASSERT(presid >= 0, ("brelse: extra page")); 1375 vm_page_set_invalid(m, poffset, presid); 1376 if (had_bogus) 1377 printf("avoided corruption bug in bogus_page/brelse code\n"); 1378 } 1379 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1380 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1381 } 1382 VM_OBJECT_UNLOCK(obj); 1383 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1384 vfs_vmio_release(bp); 1385 1386 } else if (bp->b_flags & B_VMIO) { 1387 1388 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1389 vfs_vmio_release(bp); 1390 } 1391 1392 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1393 if (bp->b_bufsize != 0) 1394 allocbuf(bp, 0); 1395 if (bp->b_vp != NULL) 1396 brelvp(bp); 1397 } 1398 1399 if (BUF_LOCKRECURSED(bp)) { 1400 /* do not release to free list */ 1401 BUF_UNLOCK(bp); 1402 return; 1403 } 1404 1405 /* enqueue */ 1406 mtx_lock(&bqlock); 1407 /* Handle delayed bremfree() processing. */ 1408 if (bp->b_flags & B_REMFREE) 1409 bremfreel(bp); 1410 if (bp->b_qindex != QUEUE_NONE) 1411 panic("brelse: free buffer onto another queue???"); 1412 1413 /* 1414 * If the buffer has junk contents signal it and eventually 1415 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1416 * doesn't find it. 1417 */ 1418 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1419 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1420 bp->b_flags |= B_INVAL; 1421 if (bp->b_flags & B_INVAL) { 1422 if (bp->b_flags & B_DELWRI) 1423 bundirty(bp); 1424 if (bp->b_vp) 1425 brelvp(bp); 1426 } 1427 1428 /* buffers with no memory */ 1429 if (bp->b_bufsize == 0) { 1430 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1431 if (bp->b_vflags & BV_BKGRDINPROG) 1432 panic("losing buffer 1"); 1433 if (bp->b_kvasize) { 1434 bp->b_qindex = QUEUE_EMPTYKVA; 1435 } else { 1436 bp->b_qindex = QUEUE_EMPTY; 1437 } 1438 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1439 /* buffers with junk contents */ 1440 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1441 (bp->b_ioflags & BIO_ERROR)) { 1442 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1443 if (bp->b_vflags & BV_BKGRDINPROG) 1444 panic("losing buffer 2"); 1445 bp->b_qindex = QUEUE_CLEAN; 1446 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1447 /* remaining buffers */ 1448 } else { 1449 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1450 (B_DELWRI|B_NEEDSGIANT)) 1451 bp->b_qindex = QUEUE_DIRTY_GIANT; 1452 else if (bp->b_flags & B_DELWRI) 1453 bp->b_qindex = QUEUE_DIRTY; 1454 else 1455 bp->b_qindex = QUEUE_CLEAN; 1456 if (bp->b_flags & B_AGE) 1457 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1458 else 1459 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1460 } 1461 mtx_unlock(&bqlock); 1462 1463 /* 1464 * Fixup numfreebuffers count. The bp is on an appropriate queue 1465 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1466 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1467 * if B_INVAL is set ). 1468 */ 1469 1470 if (!(bp->b_flags & B_DELWRI)) 1471 bufcountwakeup(bp); 1472 1473 /* 1474 * Something we can maybe free or reuse 1475 */ 1476 if (bp->b_bufsize || bp->b_kvasize) 1477 bufspacewakeup(); 1478 1479 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1480 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1481 panic("brelse: not dirty"); 1482 /* unlock */ 1483 BUF_UNLOCK(bp); 1484} 1485 1486/* 1487 * Release a buffer back to the appropriate queue but do not try to free 1488 * it. The buffer is expected to be used again soon. 1489 * 1490 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1491 * biodone() to requeue an async I/O on completion. It is also used when 1492 * known good buffers need to be requeued but we think we may need the data 1493 * again soon. 1494 * 1495 * XXX we should be able to leave the B_RELBUF hint set on completion. 1496 */ 1497void 1498bqrelse(struct buf *bp) 1499{ 1500 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1501 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1502 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1503 1504 if (BUF_LOCKRECURSED(bp)) { 1505 /* do not release to free list */ 1506 BUF_UNLOCK(bp); 1507 return; 1508 } 1509 1510 if (bp->b_flags & B_MANAGED) { 1511 if (bp->b_flags & B_REMFREE) { 1512 mtx_lock(&bqlock); 1513 bremfreel(bp); 1514 mtx_unlock(&bqlock); 1515 } 1516 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1517 BUF_UNLOCK(bp); 1518 return; 1519 } 1520 1521 mtx_lock(&bqlock); 1522 /* Handle delayed bremfree() processing. */ 1523 if (bp->b_flags & B_REMFREE) 1524 bremfreel(bp); 1525 if (bp->b_qindex != QUEUE_NONE) 1526 panic("bqrelse: free buffer onto another queue???"); 1527 /* buffers with stale but valid contents */ 1528 if (bp->b_flags & B_DELWRI) { 1529 if (bp->b_flags & B_NEEDSGIANT) 1530 bp->b_qindex = QUEUE_DIRTY_GIANT; 1531 else 1532 bp->b_qindex = QUEUE_DIRTY; 1533 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1534 } else { 1535 /* 1536 * The locking of the BO_LOCK for checking of the 1537 * BV_BKGRDINPROG is not necessary since the 1538 * BV_BKGRDINPROG cannot be set while we hold the buf 1539 * lock, it can only be cleared if it is already 1540 * pending. 1541 */ 1542 if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1543 bp->b_qindex = QUEUE_CLEAN; 1544 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1545 b_freelist); 1546 } else { 1547 /* 1548 * We are too low on memory, we have to try to free 1549 * the buffer (most importantly: the wired pages 1550 * making up its backing store) *now*. 1551 */ 1552 mtx_unlock(&bqlock); 1553 brelse(bp); 1554 return; 1555 } 1556 } 1557 mtx_unlock(&bqlock); 1558 1559 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1560 bufcountwakeup(bp); 1561 1562 /* 1563 * Something we can maybe free or reuse. 1564 */ 1565 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1566 bufspacewakeup(); 1567 1568 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1569 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1570 panic("bqrelse: not dirty"); 1571 /* unlock */ 1572 BUF_UNLOCK(bp); 1573} 1574 1575/* Give pages used by the bp back to the VM system (where possible) */ 1576static void 1577vfs_vmio_release(struct buf *bp) 1578{ 1579 int i; 1580 vm_page_t m; 1581 1582 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1583 for (i = 0; i < bp->b_npages; i++) { 1584 m = bp->b_pages[i]; 1585 bp->b_pages[i] = NULL; 1586 /* 1587 * In order to keep page LRU ordering consistent, put 1588 * everything on the inactive queue. 1589 */ 1590 vm_page_lock(m); 1591 vm_page_unwire(m, 0); 1592 /* 1593 * We don't mess with busy pages, it is 1594 * the responsibility of the process that 1595 * busied the pages to deal with them. 1596 */ 1597 if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 && 1598 m->wire_count == 0) { 1599 /* 1600 * Might as well free the page if we can and it has 1601 * no valid data. We also free the page if the 1602 * buffer was used for direct I/O 1603 */ 1604 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1605 m->hold_count == 0) { 1606 vm_page_free(m); 1607 } else if (bp->b_flags & B_DIRECT) { 1608 vm_page_try_to_free(m); 1609 } else if (buf_vm_page_count_severe()) { 1610 vm_page_try_to_cache(m); 1611 } 1612 } 1613 vm_page_unlock(m); 1614 } 1615 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1616 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1617 1618 if (bp->b_bufsize) { 1619 bufspacewakeup(); 1620 bp->b_bufsize = 0; 1621 } 1622 bp->b_npages = 0; 1623 bp->b_flags &= ~B_VMIO; 1624 if (bp->b_vp) 1625 brelvp(bp); 1626} 1627 1628/* 1629 * Check to see if a block at a particular lbn is available for a clustered 1630 * write. 1631 */ 1632static int 1633vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1634{ 1635 struct buf *bpa; 1636 int match; 1637 1638 match = 0; 1639 1640 /* If the buf isn't in core skip it */ 1641 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1642 return (0); 1643 1644 /* If the buf is busy we don't want to wait for it */ 1645 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1646 return (0); 1647 1648 /* Only cluster with valid clusterable delayed write buffers */ 1649 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1650 (B_DELWRI | B_CLUSTEROK)) 1651 goto done; 1652 1653 if (bpa->b_bufsize != size) 1654 goto done; 1655 1656 /* 1657 * Check to see if it is in the expected place on disk and that the 1658 * block has been mapped. 1659 */ 1660 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1661 match = 1; 1662done: 1663 BUF_UNLOCK(bpa); 1664 return (match); 1665} 1666 1667/* 1668 * vfs_bio_awrite: 1669 * 1670 * Implement clustered async writes for clearing out B_DELWRI buffers. 1671 * This is much better then the old way of writing only one buffer at 1672 * a time. Note that we may not be presented with the buffers in the 1673 * correct order, so we search for the cluster in both directions. 1674 */ 1675int 1676vfs_bio_awrite(struct buf *bp) 1677{ 1678 struct bufobj *bo; 1679 int i; 1680 int j; 1681 daddr_t lblkno = bp->b_lblkno; 1682 struct vnode *vp = bp->b_vp; 1683 int ncl; 1684 int nwritten; 1685 int size; 1686 int maxcl; 1687 1688 bo = &vp->v_bufobj; 1689 /* 1690 * right now we support clustered writing only to regular files. If 1691 * we find a clusterable block we could be in the middle of a cluster 1692 * rather then at the beginning. 1693 */ 1694 if ((vp->v_type == VREG) && 1695 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1696 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1697 1698 size = vp->v_mount->mnt_stat.f_iosize; 1699 maxcl = MAXPHYS / size; 1700 1701 BO_LOCK(bo); 1702 for (i = 1; i < maxcl; i++) 1703 if (vfs_bio_clcheck(vp, size, lblkno + i, 1704 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1705 break; 1706 1707 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1708 if (vfs_bio_clcheck(vp, size, lblkno - j, 1709 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1710 break; 1711 BO_UNLOCK(bo); 1712 --j; 1713 ncl = i + j; 1714 /* 1715 * this is a possible cluster write 1716 */ 1717 if (ncl != 1) { 1718 BUF_UNLOCK(bp); 1719 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1720 return nwritten; 1721 } 1722 } 1723 bremfree(bp); 1724 bp->b_flags |= B_ASYNC; 1725 /* 1726 * default (old) behavior, writing out only one block 1727 * 1728 * XXX returns b_bufsize instead of b_bcount for nwritten? 1729 */ 1730 nwritten = bp->b_bufsize; 1731 (void) bwrite(bp); 1732 1733 return nwritten; 1734} 1735 1736/* 1737 * getnewbuf: 1738 * 1739 * Find and initialize a new buffer header, freeing up existing buffers 1740 * in the bufqueues as necessary. The new buffer is returned locked. 1741 * 1742 * Important: B_INVAL is not set. If the caller wishes to throw the 1743 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1744 * 1745 * We block if: 1746 * We have insufficient buffer headers 1747 * We have insufficient buffer space 1748 * buffer_map is too fragmented ( space reservation fails ) 1749 * If we have to flush dirty buffers ( but we try to avoid this ) 1750 * 1751 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1752 * Instead we ask the buf daemon to do it for us. We attempt to 1753 * avoid piecemeal wakeups of the pageout daemon. 1754 */ 1755 1756static struct buf * 1757getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 1758 int gbflags) 1759{ 1760 struct thread *td; 1761 struct buf *bp; 1762 struct buf *nbp; 1763 int defrag = 0; 1764 int nqindex; 1765 static int flushingbufs; 1766 1767 td = curthread; 1768 /* 1769 * We can't afford to block since we might be holding a vnode lock, 1770 * which may prevent system daemons from running. We deal with 1771 * low-memory situations by proactively returning memory and running 1772 * async I/O rather then sync I/O. 1773 */ 1774 atomic_add_int(&getnewbufcalls, 1); 1775 atomic_subtract_int(&getnewbufrestarts, 1); 1776restart: 1777 atomic_add_int(&getnewbufrestarts, 1); 1778 1779 /* 1780 * Setup for scan. If we do not have enough free buffers, 1781 * we setup a degenerate case that immediately fails. Note 1782 * that if we are specially marked process, we are allowed to 1783 * dip into our reserves. 1784 * 1785 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1786 * 1787 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1788 * However, there are a number of cases (defragging, reusing, ...) 1789 * where we cannot backup. 1790 */ 1791 mtx_lock(&bqlock); 1792 nqindex = QUEUE_EMPTYKVA; 1793 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1794 1795 if (nbp == NULL) { 1796 /* 1797 * If no EMPTYKVA buffers and we are either 1798 * defragging or reusing, locate a CLEAN buffer 1799 * to free or reuse. If bufspace useage is low 1800 * skip this step so we can allocate a new buffer. 1801 */ 1802 if (defrag || bufspace >= lobufspace) { 1803 nqindex = QUEUE_CLEAN; 1804 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1805 } 1806 1807 /* 1808 * If we could not find or were not allowed to reuse a 1809 * CLEAN buffer, check to see if it is ok to use an EMPTY 1810 * buffer. We can only use an EMPTY buffer if allocating 1811 * its KVA would not otherwise run us out of buffer space. 1812 */ 1813 if (nbp == NULL && defrag == 0 && 1814 bufspace + maxsize < hibufspace) { 1815 nqindex = QUEUE_EMPTY; 1816 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1817 } 1818 } 1819 1820 /* 1821 * Run scan, possibly freeing data and/or kva mappings on the fly 1822 * depending. 1823 */ 1824 1825 while ((bp = nbp) != NULL) { 1826 int qindex = nqindex; 1827 1828 /* 1829 * Calculate next bp ( we can only use it if we do not block 1830 * or do other fancy things ). 1831 */ 1832 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1833 switch(qindex) { 1834 case QUEUE_EMPTY: 1835 nqindex = QUEUE_EMPTYKVA; 1836 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1837 break; 1838 /* FALLTHROUGH */ 1839 case QUEUE_EMPTYKVA: 1840 nqindex = QUEUE_CLEAN; 1841 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1842 break; 1843 /* FALLTHROUGH */ 1844 case QUEUE_CLEAN: 1845 /* 1846 * nbp is NULL. 1847 */ 1848 break; 1849 } 1850 } 1851 /* 1852 * If we are defragging then we need a buffer with 1853 * b_kvasize != 0. XXX this situation should no longer 1854 * occur, if defrag is non-zero the buffer's b_kvasize 1855 * should also be non-zero at this point. XXX 1856 */ 1857 if (defrag && bp->b_kvasize == 0) { 1858 printf("Warning: defrag empty buffer %p\n", bp); 1859 continue; 1860 } 1861 1862 /* 1863 * Start freeing the bp. This is somewhat involved. nbp 1864 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1865 */ 1866 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1867 continue; 1868 if (bp->b_vp) { 1869 BO_LOCK(bp->b_bufobj); 1870 if (bp->b_vflags & BV_BKGRDINPROG) { 1871 BO_UNLOCK(bp->b_bufobj); 1872 BUF_UNLOCK(bp); 1873 continue; 1874 } 1875 BO_UNLOCK(bp->b_bufobj); 1876 } 1877 CTR6(KTR_BUF, 1878 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1879 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1880 bp->b_kvasize, bp->b_bufsize, qindex); 1881 1882 /* 1883 * Sanity Checks 1884 */ 1885 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1886 1887 /* 1888 * Note: we no longer distinguish between VMIO and non-VMIO 1889 * buffers. 1890 */ 1891 1892 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1893 1894 bremfreel(bp); 1895 mtx_unlock(&bqlock); 1896 1897 if (qindex == QUEUE_CLEAN) { 1898 if (bp->b_flags & B_VMIO) { 1899 bp->b_flags &= ~B_ASYNC; 1900 vfs_vmio_release(bp); 1901 } 1902 if (bp->b_vp) 1903 brelvp(bp); 1904 } 1905 1906 /* 1907 * NOTE: nbp is now entirely invalid. We can only restart 1908 * the scan from this point on. 1909 * 1910 * Get the rest of the buffer freed up. b_kva* is still 1911 * valid after this operation. 1912 */ 1913 1914 if (bp->b_rcred != NOCRED) { 1915 crfree(bp->b_rcred); 1916 bp->b_rcred = NOCRED; 1917 } 1918 if (bp->b_wcred != NOCRED) { 1919 crfree(bp->b_wcred); 1920 bp->b_wcred = NOCRED; 1921 } 1922 if (!LIST_EMPTY(&bp->b_dep)) 1923 buf_deallocate(bp); 1924 if (bp->b_vflags & BV_BKGRDINPROG) 1925 panic("losing buffer 3"); 1926 KASSERT(bp->b_vp == NULL, 1927 ("bp: %p still has vnode %p. qindex: %d", 1928 bp, bp->b_vp, qindex)); 1929 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1930 ("bp: %p still on a buffer list. xflags %X", 1931 bp, bp->b_xflags)); 1932 1933 if (bp->b_bufsize) 1934 allocbuf(bp, 0); 1935 1936 bp->b_flags = 0; 1937 bp->b_ioflags = 0; 1938 bp->b_xflags = 0; 1939 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 1940 ("buf %p still counted as free?", bp)); 1941 bp->b_vflags = 0; 1942 bp->b_vp = NULL; 1943 bp->b_blkno = bp->b_lblkno = 0; 1944 bp->b_offset = NOOFFSET; 1945 bp->b_iodone = 0; 1946 bp->b_error = 0; 1947 bp->b_resid = 0; 1948 bp->b_bcount = 0; 1949 bp->b_npages = 0; 1950 bp->b_dirtyoff = bp->b_dirtyend = 0; 1951 bp->b_bufobj = NULL; 1952 bp->b_pin_count = 0; 1953 bp->b_fsprivate1 = NULL; 1954 bp->b_fsprivate2 = NULL; 1955 bp->b_fsprivate3 = NULL; 1956 1957 LIST_INIT(&bp->b_dep); 1958 1959 /* 1960 * If we are defragging then free the buffer. 1961 */ 1962 if (defrag) { 1963 bp->b_flags |= B_INVAL; 1964 bfreekva(bp); 1965 brelse(bp); 1966 defrag = 0; 1967 goto restart; 1968 } 1969 1970 /* 1971 * Notify any waiters for the buffer lock about 1972 * identity change by freeing the buffer. 1973 */ 1974 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 1975 bp->b_flags |= B_INVAL; 1976 bfreekva(bp); 1977 brelse(bp); 1978 goto restart; 1979 } 1980 1981 /* 1982 * If we are overcomitted then recover the buffer and its 1983 * KVM space. This occurs in rare situations when multiple 1984 * processes are blocked in getnewbuf() or allocbuf(). 1985 */ 1986 if (bufspace >= hibufspace) 1987 flushingbufs = 1; 1988 if (flushingbufs && bp->b_kvasize != 0) { 1989 bp->b_flags |= B_INVAL; 1990 bfreekva(bp); 1991 brelse(bp); 1992 goto restart; 1993 } 1994 if (bufspace < lobufspace) 1995 flushingbufs = 0; 1996 break; 1997 } 1998 1999 /* 2000 * If we exhausted our list, sleep as appropriate. We may have to 2001 * wakeup various daemons and write out some dirty buffers. 2002 * 2003 * Generally we are sleeping due to insufficient buffer space. 2004 */ 2005 2006 if (bp == NULL) { 2007 int flags, norunbuf; 2008 char *waitmsg; 2009 int fl; 2010 2011 if (defrag) { 2012 flags = VFS_BIO_NEED_BUFSPACE; 2013 waitmsg = "nbufkv"; 2014 } else if (bufspace >= hibufspace) { 2015 waitmsg = "nbufbs"; 2016 flags = VFS_BIO_NEED_BUFSPACE; 2017 } else { 2018 waitmsg = "newbuf"; 2019 flags = VFS_BIO_NEED_ANY; 2020 } 2021 mtx_lock(&nblock); 2022 needsbuffer |= flags; 2023 mtx_unlock(&nblock); 2024 mtx_unlock(&bqlock); 2025 2026 bd_speedup(); /* heeeelp */ 2027 if (gbflags & GB_NOWAIT_BD) 2028 return (NULL); 2029 2030 mtx_lock(&nblock); 2031 while (needsbuffer & flags) { 2032 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2033 mtx_unlock(&nblock); 2034 /* 2035 * getblk() is called with a vnode 2036 * locked, and some majority of the 2037 * dirty buffers may as well belong to 2038 * the vnode. Flushing the buffers 2039 * there would make a progress that 2040 * cannot be achieved by the 2041 * buf_daemon, that cannot lock the 2042 * vnode. 2043 */ 2044 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2045 (td->td_pflags & TDP_NORUNNINGBUF); 2046 /* play bufdaemon */ 2047 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2048 fl = buf_do_flush(vp); 2049 td->td_pflags &= norunbuf; 2050 mtx_lock(&nblock); 2051 if (fl != 0) 2052 continue; 2053 if ((needsbuffer & flags) == 0) 2054 break; 2055 } 2056 if (msleep(&needsbuffer, &nblock, 2057 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2058 mtx_unlock(&nblock); 2059 return (NULL); 2060 } 2061 } 2062 mtx_unlock(&nblock); 2063 } else { 2064 /* 2065 * We finally have a valid bp. We aren't quite out of the 2066 * woods, we still have to reserve kva space. In order 2067 * to keep fragmentation sane we only allocate kva in 2068 * BKVASIZE chunks. 2069 */ 2070 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2071 2072 if (maxsize != bp->b_kvasize) { 2073 vm_offset_t addr = 0; 2074 2075 bfreekva(bp); 2076 2077 vm_map_lock(buffer_map); 2078 if (vm_map_findspace(buffer_map, 2079 vm_map_min(buffer_map), maxsize, &addr)) { 2080 /* 2081 * Uh oh. Buffer map is to fragmented. We 2082 * must defragment the map. 2083 */ 2084 atomic_add_int(&bufdefragcnt, 1); 2085 vm_map_unlock(buffer_map); 2086 defrag = 1; 2087 bp->b_flags |= B_INVAL; 2088 brelse(bp); 2089 goto restart; 2090 } 2091 if (addr) { 2092 vm_map_insert(buffer_map, NULL, 0, 2093 addr, addr + maxsize, 2094 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2095 2096 bp->b_kvabase = (caddr_t) addr; 2097 bp->b_kvasize = maxsize; 2098 atomic_add_long(&bufspace, bp->b_kvasize); 2099 atomic_add_int(&bufreusecnt, 1); 2100 } 2101 vm_map_unlock(buffer_map); 2102 } 2103 bp->b_saveaddr = bp->b_kvabase; 2104 bp->b_data = bp->b_saveaddr; 2105 } 2106 return(bp); 2107} 2108 2109/* 2110 * buf_daemon: 2111 * 2112 * buffer flushing daemon. Buffers are normally flushed by the 2113 * update daemon but if it cannot keep up this process starts to 2114 * take the load in an attempt to prevent getnewbuf() from blocking. 2115 */ 2116 2117static struct kproc_desc buf_kp = { 2118 "bufdaemon", 2119 buf_daemon, 2120 &bufdaemonproc 2121}; 2122SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2123 2124static int 2125buf_do_flush(struct vnode *vp) 2126{ 2127 int flushed; 2128 2129 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2130 /* The list empty check here is slightly racy */ 2131 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2132 mtx_lock(&Giant); 2133 flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0); 2134 mtx_unlock(&Giant); 2135 } 2136 if (flushed == 0) { 2137 /* 2138 * Could not find any buffers without rollback 2139 * dependencies, so just write the first one 2140 * in the hopes of eventually making progress. 2141 */ 2142 flushbufqueues(vp, QUEUE_DIRTY, 1); 2143 if (!TAILQ_EMPTY( 2144 &bufqueues[QUEUE_DIRTY_GIANT])) { 2145 mtx_lock(&Giant); 2146 flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1); 2147 mtx_unlock(&Giant); 2148 } 2149 } 2150 return (flushed); 2151} 2152 2153static void 2154buf_daemon() 2155{ 2156 int lodirtysave; 2157 2158 /* 2159 * This process needs to be suspended prior to shutdown sync. 2160 */ 2161 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2162 SHUTDOWN_PRI_LAST); 2163 2164 /* 2165 * This process is allowed to take the buffer cache to the limit 2166 */ 2167 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2168 mtx_lock(&bdlock); 2169 for (;;) { 2170 bd_request = 0; 2171 mtx_unlock(&bdlock); 2172 2173 kproc_suspend_check(bufdaemonproc); 2174 lodirtysave = lodirtybuffers; 2175 if (bd_speedupreq) { 2176 lodirtybuffers = numdirtybuffers / 2; 2177 bd_speedupreq = 0; 2178 } 2179 /* 2180 * Do the flush. Limit the amount of in-transit I/O we 2181 * allow to build up, otherwise we would completely saturate 2182 * the I/O system. Wakeup any waiting processes before we 2183 * normally would so they can run in parallel with our drain. 2184 */ 2185 while (numdirtybuffers > lodirtybuffers) { 2186 if (buf_do_flush(NULL) == 0) 2187 break; 2188 uio_yield(); 2189 } 2190 lodirtybuffers = lodirtysave; 2191 2192 /* 2193 * Only clear bd_request if we have reached our low water 2194 * mark. The buf_daemon normally waits 1 second and 2195 * then incrementally flushes any dirty buffers that have 2196 * built up, within reason. 2197 * 2198 * If we were unable to hit our low water mark and couldn't 2199 * find any flushable buffers, we sleep half a second. 2200 * Otherwise we loop immediately. 2201 */ 2202 mtx_lock(&bdlock); 2203 if (numdirtybuffers <= lodirtybuffers) { 2204 /* 2205 * We reached our low water mark, reset the 2206 * request and sleep until we are needed again. 2207 * The sleep is just so the suspend code works. 2208 */ 2209 bd_request = 0; 2210 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2211 } else { 2212 /* 2213 * We couldn't find any flushable dirty buffers but 2214 * still have too many dirty buffers, we 2215 * have to sleep and try again. (rare) 2216 */ 2217 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2218 } 2219 } 2220} 2221 2222/* 2223 * flushbufqueues: 2224 * 2225 * Try to flush a buffer in the dirty queue. We must be careful to 2226 * free up B_INVAL buffers instead of write them, which NFS is 2227 * particularly sensitive to. 2228 */ 2229static int flushwithdeps = 0; 2230SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2231 0, "Number of buffers flushed with dependecies that require rollbacks"); 2232 2233static int 2234flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2235{ 2236 struct buf *sentinel; 2237 struct vnode *vp; 2238 struct mount *mp; 2239 struct buf *bp; 2240 int hasdeps; 2241 int flushed; 2242 int target; 2243 2244 if (lvp == NULL) { 2245 target = numdirtybuffers - lodirtybuffers; 2246 if (flushdeps && target > 2) 2247 target /= 2; 2248 } else 2249 target = flushbufqtarget; 2250 flushed = 0; 2251 bp = NULL; 2252 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2253 sentinel->b_qindex = QUEUE_SENTINEL; 2254 mtx_lock(&bqlock); 2255 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2256 while (flushed != target) { 2257 bp = TAILQ_NEXT(sentinel, b_freelist); 2258 if (bp != NULL) { 2259 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2260 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2261 b_freelist); 2262 } else 2263 break; 2264 /* 2265 * Skip sentinels inserted by other invocations of the 2266 * flushbufqueues(), taking care to not reorder them. 2267 */ 2268 if (bp->b_qindex == QUEUE_SENTINEL) 2269 continue; 2270 /* 2271 * Only flush the buffers that belong to the 2272 * vnode locked by the curthread. 2273 */ 2274 if (lvp != NULL && bp->b_vp != lvp) 2275 continue; 2276 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2277 continue; 2278 if (bp->b_pin_count > 0) { 2279 BUF_UNLOCK(bp); 2280 continue; 2281 } 2282 BO_LOCK(bp->b_bufobj); 2283 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2284 (bp->b_flags & B_DELWRI) == 0) { 2285 BO_UNLOCK(bp->b_bufobj); 2286 BUF_UNLOCK(bp); 2287 continue; 2288 } 2289 BO_UNLOCK(bp->b_bufobj); 2290 if (bp->b_flags & B_INVAL) { 2291 bremfreel(bp); 2292 mtx_unlock(&bqlock); 2293 brelse(bp); 2294 flushed++; 2295 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2296 mtx_lock(&bqlock); 2297 continue; 2298 } 2299 2300 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2301 if (flushdeps == 0) { 2302 BUF_UNLOCK(bp); 2303 continue; 2304 } 2305 hasdeps = 1; 2306 } else 2307 hasdeps = 0; 2308 /* 2309 * We must hold the lock on a vnode before writing 2310 * one of its buffers. Otherwise we may confuse, or 2311 * in the case of a snapshot vnode, deadlock the 2312 * system. 2313 * 2314 * The lock order here is the reverse of the normal 2315 * of vnode followed by buf lock. This is ok because 2316 * the NOWAIT will prevent deadlock. 2317 */ 2318 vp = bp->b_vp; 2319 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2320 BUF_UNLOCK(bp); 2321 continue; 2322 } 2323 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2324 mtx_unlock(&bqlock); 2325 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2326 bp, bp->b_vp, bp->b_flags); 2327 if (curproc == bufdaemonproc) 2328 vfs_bio_awrite(bp); 2329 else { 2330 bremfree(bp); 2331 bwrite(bp); 2332 notbufdflashes++; 2333 } 2334 vn_finished_write(mp); 2335 VOP_UNLOCK(vp, 0); 2336 flushwithdeps += hasdeps; 2337 flushed++; 2338 2339 /* 2340 * Sleeping on runningbufspace while holding 2341 * vnode lock leads to deadlock. 2342 */ 2343 if (curproc == bufdaemonproc) 2344 waitrunningbufspace(); 2345 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2346 mtx_lock(&bqlock); 2347 continue; 2348 } 2349 vn_finished_write(mp); 2350 BUF_UNLOCK(bp); 2351 } 2352 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2353 mtx_unlock(&bqlock); 2354 free(sentinel, M_TEMP); 2355 return (flushed); 2356} 2357 2358/* 2359 * Check to see if a block is currently memory resident. 2360 */ 2361struct buf * 2362incore(struct bufobj *bo, daddr_t blkno) 2363{ 2364 struct buf *bp; 2365 2366 BO_LOCK(bo); 2367 bp = gbincore(bo, blkno); 2368 BO_UNLOCK(bo); 2369 return (bp); 2370} 2371 2372/* 2373 * Returns true if no I/O is needed to access the 2374 * associated VM object. This is like incore except 2375 * it also hunts around in the VM system for the data. 2376 */ 2377 2378static int 2379inmem(struct vnode * vp, daddr_t blkno) 2380{ 2381 vm_object_t obj; 2382 vm_offset_t toff, tinc, size; 2383 vm_page_t m; 2384 vm_ooffset_t off; 2385 2386 ASSERT_VOP_LOCKED(vp, "inmem"); 2387 2388 if (incore(&vp->v_bufobj, blkno)) 2389 return 1; 2390 if (vp->v_mount == NULL) 2391 return 0; 2392 obj = vp->v_object; 2393 if (obj == NULL) 2394 return (0); 2395 2396 size = PAGE_SIZE; 2397 if (size > vp->v_mount->mnt_stat.f_iosize) 2398 size = vp->v_mount->mnt_stat.f_iosize; 2399 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2400 2401 VM_OBJECT_LOCK(obj); 2402 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2403 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2404 if (!m) 2405 goto notinmem; 2406 tinc = size; 2407 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2408 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2409 if (vm_page_is_valid(m, 2410 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2411 goto notinmem; 2412 } 2413 VM_OBJECT_UNLOCK(obj); 2414 return 1; 2415 2416notinmem: 2417 VM_OBJECT_UNLOCK(obj); 2418 return (0); 2419} 2420 2421/* 2422 * Set the dirty range for a buffer based on the status of the dirty 2423 * bits in the pages comprising the buffer. The range is limited 2424 * to the size of the buffer. 2425 * 2426 * Tell the VM system that the pages associated with this buffer 2427 * are clean. This is used for delayed writes where the data is 2428 * going to go to disk eventually without additional VM intevention. 2429 * 2430 * Note that while we only really need to clean through to b_bcount, we 2431 * just go ahead and clean through to b_bufsize. 2432 */ 2433static void 2434vfs_clean_pages_dirty_buf(struct buf *bp) 2435{ 2436 vm_ooffset_t foff, noff, eoff; 2437 vm_page_t m; 2438 int i; 2439 2440 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2441 return; 2442 2443 foff = bp->b_offset; 2444 KASSERT(bp->b_offset != NOOFFSET, 2445 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2446 2447 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2448 vfs_drain_busy_pages(bp); 2449 vfs_setdirty_locked_object(bp); 2450 for (i = 0; i < bp->b_npages; i++) { 2451 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2452 eoff = noff; 2453 if (eoff > bp->b_offset + bp->b_bufsize) 2454 eoff = bp->b_offset + bp->b_bufsize; 2455 m = bp->b_pages[i]; 2456 vfs_page_set_validclean(bp, foff, m); 2457 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2458 foff = noff; 2459 } 2460 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2461} 2462 2463static void 2464vfs_setdirty_locked_object(struct buf *bp) 2465{ 2466 vm_object_t object; 2467 int i; 2468 2469 object = bp->b_bufobj->bo_object; 2470 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2471 2472 /* 2473 * We qualify the scan for modified pages on whether the 2474 * object has been flushed yet. 2475 */ 2476 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2477 vm_offset_t boffset; 2478 vm_offset_t eoffset; 2479 2480 /* 2481 * test the pages to see if they have been modified directly 2482 * by users through the VM system. 2483 */ 2484 for (i = 0; i < bp->b_npages; i++) 2485 vm_page_test_dirty(bp->b_pages[i]); 2486 2487 /* 2488 * Calculate the encompassing dirty range, boffset and eoffset, 2489 * (eoffset - boffset) bytes. 2490 */ 2491 2492 for (i = 0; i < bp->b_npages; i++) { 2493 if (bp->b_pages[i]->dirty) 2494 break; 2495 } 2496 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2497 2498 for (i = bp->b_npages - 1; i >= 0; --i) { 2499 if (bp->b_pages[i]->dirty) { 2500 break; 2501 } 2502 } 2503 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2504 2505 /* 2506 * Fit it to the buffer. 2507 */ 2508 2509 if (eoffset > bp->b_bcount) 2510 eoffset = bp->b_bcount; 2511 2512 /* 2513 * If we have a good dirty range, merge with the existing 2514 * dirty range. 2515 */ 2516 2517 if (boffset < eoffset) { 2518 if (bp->b_dirtyoff > boffset) 2519 bp->b_dirtyoff = boffset; 2520 if (bp->b_dirtyend < eoffset) 2521 bp->b_dirtyend = eoffset; 2522 } 2523 } 2524} 2525 2526/* 2527 * getblk: 2528 * 2529 * Get a block given a specified block and offset into a file/device. 2530 * The buffers B_DONE bit will be cleared on return, making it almost 2531 * ready for an I/O initiation. B_INVAL may or may not be set on 2532 * return. The caller should clear B_INVAL prior to initiating a 2533 * READ. 2534 * 2535 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2536 * an existing buffer. 2537 * 2538 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2539 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2540 * and then cleared based on the backing VM. If the previous buffer is 2541 * non-0-sized but invalid, B_CACHE will be cleared. 2542 * 2543 * If getblk() must create a new buffer, the new buffer is returned with 2544 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2545 * case it is returned with B_INVAL clear and B_CACHE set based on the 2546 * backing VM. 2547 * 2548 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2549 * B_CACHE bit is clear. 2550 * 2551 * What this means, basically, is that the caller should use B_CACHE to 2552 * determine whether the buffer is fully valid or not and should clear 2553 * B_INVAL prior to issuing a read. If the caller intends to validate 2554 * the buffer by loading its data area with something, the caller needs 2555 * to clear B_INVAL. If the caller does this without issuing an I/O, 2556 * the caller should set B_CACHE ( as an optimization ), else the caller 2557 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2558 * a write attempt or if it was a successfull read. If the caller 2559 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2560 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2561 */ 2562struct buf * 2563getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2564 int flags) 2565{ 2566 struct buf *bp; 2567 struct bufobj *bo; 2568 int error; 2569 2570 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2571 ASSERT_VOP_LOCKED(vp, "getblk"); 2572 if (size > MAXBSIZE) 2573 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2574 2575 bo = &vp->v_bufobj; 2576loop: 2577 /* 2578 * Block if we are low on buffers. Certain processes are allowed 2579 * to completely exhaust the buffer cache. 2580 * 2581 * If this check ever becomes a bottleneck it may be better to 2582 * move it into the else, when gbincore() fails. At the moment 2583 * it isn't a problem. 2584 * 2585 * XXX remove if 0 sections (clean this up after its proven) 2586 */ 2587 if (numfreebuffers == 0) { 2588 if (TD_IS_IDLETHREAD(curthread)) 2589 return NULL; 2590 mtx_lock(&nblock); 2591 needsbuffer |= VFS_BIO_NEED_ANY; 2592 mtx_unlock(&nblock); 2593 } 2594 2595 BO_LOCK(bo); 2596 bp = gbincore(bo, blkno); 2597 if (bp != NULL) { 2598 int lockflags; 2599 /* 2600 * Buffer is in-core. If the buffer is not busy, it must 2601 * be on a queue. 2602 */ 2603 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2604 2605 if (flags & GB_LOCK_NOWAIT) 2606 lockflags |= LK_NOWAIT; 2607 2608 error = BUF_TIMELOCK(bp, lockflags, 2609 BO_MTX(bo), "getblk", slpflag, slptimeo); 2610 2611 /* 2612 * If we slept and got the lock we have to restart in case 2613 * the buffer changed identities. 2614 */ 2615 if (error == ENOLCK) 2616 goto loop; 2617 /* We timed out or were interrupted. */ 2618 else if (error) 2619 return (NULL); 2620 2621 /* 2622 * The buffer is locked. B_CACHE is cleared if the buffer is 2623 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2624 * and for a VMIO buffer B_CACHE is adjusted according to the 2625 * backing VM cache. 2626 */ 2627 if (bp->b_flags & B_INVAL) 2628 bp->b_flags &= ~B_CACHE; 2629 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2630 bp->b_flags |= B_CACHE; 2631 bremfree(bp); 2632 2633 /* 2634 * check for size inconsistancies for non-VMIO case. 2635 */ 2636 2637 if (bp->b_bcount != size) { 2638 if ((bp->b_flags & B_VMIO) == 0 || 2639 (size > bp->b_kvasize)) { 2640 if (bp->b_flags & B_DELWRI) { 2641 /* 2642 * If buffer is pinned and caller does 2643 * not want sleep waiting for it to be 2644 * unpinned, bail out 2645 * */ 2646 if (bp->b_pin_count > 0) { 2647 if (flags & GB_LOCK_NOWAIT) { 2648 bqrelse(bp); 2649 return (NULL); 2650 } else { 2651 bunpin_wait(bp); 2652 } 2653 } 2654 bp->b_flags |= B_NOCACHE; 2655 bwrite(bp); 2656 } else { 2657 if (LIST_EMPTY(&bp->b_dep)) { 2658 bp->b_flags |= B_RELBUF; 2659 brelse(bp); 2660 } else { 2661 bp->b_flags |= B_NOCACHE; 2662 bwrite(bp); 2663 } 2664 } 2665 goto loop; 2666 } 2667 } 2668 2669 /* 2670 * If the size is inconsistant in the VMIO case, we can resize 2671 * the buffer. This might lead to B_CACHE getting set or 2672 * cleared. If the size has not changed, B_CACHE remains 2673 * unchanged from its previous state. 2674 */ 2675 2676 if (bp->b_bcount != size) 2677 allocbuf(bp, size); 2678 2679 KASSERT(bp->b_offset != NOOFFSET, 2680 ("getblk: no buffer offset")); 2681 2682 /* 2683 * A buffer with B_DELWRI set and B_CACHE clear must 2684 * be committed before we can return the buffer in 2685 * order to prevent the caller from issuing a read 2686 * ( due to B_CACHE not being set ) and overwriting 2687 * it. 2688 * 2689 * Most callers, including NFS and FFS, need this to 2690 * operate properly either because they assume they 2691 * can issue a read if B_CACHE is not set, or because 2692 * ( for example ) an uncached B_DELWRI might loop due 2693 * to softupdates re-dirtying the buffer. In the latter 2694 * case, B_CACHE is set after the first write completes, 2695 * preventing further loops. 2696 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2697 * above while extending the buffer, we cannot allow the 2698 * buffer to remain with B_CACHE set after the write 2699 * completes or it will represent a corrupt state. To 2700 * deal with this we set B_NOCACHE to scrap the buffer 2701 * after the write. 2702 * 2703 * We might be able to do something fancy, like setting 2704 * B_CACHE in bwrite() except if B_DELWRI is already set, 2705 * so the below call doesn't set B_CACHE, but that gets real 2706 * confusing. This is much easier. 2707 */ 2708 2709 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2710 bp->b_flags |= B_NOCACHE; 2711 bwrite(bp); 2712 goto loop; 2713 } 2714 bp->b_flags &= ~B_DONE; 2715 } else { 2716 int bsize, maxsize, vmio; 2717 off_t offset; 2718 2719 /* 2720 * Buffer is not in-core, create new buffer. The buffer 2721 * returned by getnewbuf() is locked. Note that the returned 2722 * buffer is also considered valid (not marked B_INVAL). 2723 */ 2724 BO_UNLOCK(bo); 2725 /* 2726 * If the user does not want us to create the buffer, bail out 2727 * here. 2728 */ 2729 if (flags & GB_NOCREAT) 2730 return NULL; 2731 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 2732 offset = blkno * bsize; 2733 vmio = vp->v_object != NULL; 2734 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2735 maxsize = imax(maxsize, bsize); 2736 2737 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 2738 if (bp == NULL) { 2739 if (slpflag || slptimeo) 2740 return NULL; 2741 goto loop; 2742 } 2743 2744 /* 2745 * This code is used to make sure that a buffer is not 2746 * created while the getnewbuf routine is blocked. 2747 * This can be a problem whether the vnode is locked or not. 2748 * If the buffer is created out from under us, we have to 2749 * throw away the one we just created. 2750 * 2751 * Note: this must occur before we associate the buffer 2752 * with the vp especially considering limitations in 2753 * the splay tree implementation when dealing with duplicate 2754 * lblkno's. 2755 */ 2756 BO_LOCK(bo); 2757 if (gbincore(bo, blkno)) { 2758 BO_UNLOCK(bo); 2759 bp->b_flags |= B_INVAL; 2760 brelse(bp); 2761 goto loop; 2762 } 2763 2764 /* 2765 * Insert the buffer into the hash, so that it can 2766 * be found by incore. 2767 */ 2768 bp->b_blkno = bp->b_lblkno = blkno; 2769 bp->b_offset = offset; 2770 bgetvp(vp, bp); 2771 BO_UNLOCK(bo); 2772 2773 /* 2774 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2775 * buffer size starts out as 0, B_CACHE will be set by 2776 * allocbuf() for the VMIO case prior to it testing the 2777 * backing store for validity. 2778 */ 2779 2780 if (vmio) { 2781 bp->b_flags |= B_VMIO; 2782#if defined(VFS_BIO_DEBUG) 2783 if (vn_canvmio(vp) != TRUE) 2784 printf("getblk: VMIO on vnode type %d\n", 2785 vp->v_type); 2786#endif 2787 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2788 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2789 bp, vp->v_object, bp->b_bufobj->bo_object)); 2790 } else { 2791 bp->b_flags &= ~B_VMIO; 2792 KASSERT(bp->b_bufobj->bo_object == NULL, 2793 ("ARGH! has b_bufobj->bo_object %p %p\n", 2794 bp, bp->b_bufobj->bo_object)); 2795 } 2796 2797 allocbuf(bp, size); 2798 bp->b_flags &= ~B_DONE; 2799 } 2800 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2801 BUF_ASSERT_HELD(bp); 2802 KASSERT(bp->b_bufobj == bo, 2803 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2804 return (bp); 2805} 2806 2807/* 2808 * Get an empty, disassociated buffer of given size. The buffer is initially 2809 * set to B_INVAL. 2810 */ 2811struct buf * 2812geteblk(int size, int flags) 2813{ 2814 struct buf *bp; 2815 int maxsize; 2816 2817 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2818 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 2819 if ((flags & GB_NOWAIT_BD) && 2820 (curthread->td_pflags & TDP_BUFNEED) != 0) 2821 return (NULL); 2822 } 2823 allocbuf(bp, size); 2824 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2825 BUF_ASSERT_HELD(bp); 2826 return (bp); 2827} 2828 2829 2830/* 2831 * This code constitutes the buffer memory from either anonymous system 2832 * memory (in the case of non-VMIO operations) or from an associated 2833 * VM object (in the case of VMIO operations). This code is able to 2834 * resize a buffer up or down. 2835 * 2836 * Note that this code is tricky, and has many complications to resolve 2837 * deadlock or inconsistant data situations. Tread lightly!!! 2838 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2839 * the caller. Calling this code willy nilly can result in the loss of data. 2840 * 2841 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2842 * B_CACHE for the non-VMIO case. 2843 */ 2844 2845int 2846allocbuf(struct buf *bp, int size) 2847{ 2848 int newbsize, mbsize; 2849 int i; 2850 2851 BUF_ASSERT_HELD(bp); 2852 2853 if (bp->b_kvasize < size) 2854 panic("allocbuf: buffer too small"); 2855 2856 if ((bp->b_flags & B_VMIO) == 0) { 2857 caddr_t origbuf; 2858 int origbufsize; 2859 /* 2860 * Just get anonymous memory from the kernel. Don't 2861 * mess with B_CACHE. 2862 */ 2863 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2864 if (bp->b_flags & B_MALLOC) 2865 newbsize = mbsize; 2866 else 2867 newbsize = round_page(size); 2868 2869 if (newbsize < bp->b_bufsize) { 2870 /* 2871 * malloced buffers are not shrunk 2872 */ 2873 if (bp->b_flags & B_MALLOC) { 2874 if (newbsize) { 2875 bp->b_bcount = size; 2876 } else { 2877 free(bp->b_data, M_BIOBUF); 2878 if (bp->b_bufsize) { 2879 atomic_subtract_long( 2880 &bufmallocspace, 2881 bp->b_bufsize); 2882 bufspacewakeup(); 2883 bp->b_bufsize = 0; 2884 } 2885 bp->b_saveaddr = bp->b_kvabase; 2886 bp->b_data = bp->b_saveaddr; 2887 bp->b_bcount = 0; 2888 bp->b_flags &= ~B_MALLOC; 2889 } 2890 return 1; 2891 } 2892 vm_hold_free_pages(bp, newbsize); 2893 } else if (newbsize > bp->b_bufsize) { 2894 /* 2895 * We only use malloced memory on the first allocation. 2896 * and revert to page-allocated memory when the buffer 2897 * grows. 2898 */ 2899 /* 2900 * There is a potential smp race here that could lead 2901 * to bufmallocspace slightly passing the max. It 2902 * is probably extremely rare and not worth worrying 2903 * over. 2904 */ 2905 if ( (bufmallocspace < maxbufmallocspace) && 2906 (bp->b_bufsize == 0) && 2907 (mbsize <= PAGE_SIZE/2)) { 2908 2909 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2910 bp->b_bufsize = mbsize; 2911 bp->b_bcount = size; 2912 bp->b_flags |= B_MALLOC; 2913 atomic_add_long(&bufmallocspace, mbsize); 2914 return 1; 2915 } 2916 origbuf = NULL; 2917 origbufsize = 0; 2918 /* 2919 * If the buffer is growing on its other-than-first allocation, 2920 * then we revert to the page-allocation scheme. 2921 */ 2922 if (bp->b_flags & B_MALLOC) { 2923 origbuf = bp->b_data; 2924 origbufsize = bp->b_bufsize; 2925 bp->b_data = bp->b_kvabase; 2926 if (bp->b_bufsize) { 2927 atomic_subtract_long(&bufmallocspace, 2928 bp->b_bufsize); 2929 bufspacewakeup(); 2930 bp->b_bufsize = 0; 2931 } 2932 bp->b_flags &= ~B_MALLOC; 2933 newbsize = round_page(newbsize); 2934 } 2935 vm_hold_load_pages( 2936 bp, 2937 (vm_offset_t) bp->b_data + bp->b_bufsize, 2938 (vm_offset_t) bp->b_data + newbsize); 2939 if (origbuf) { 2940 bcopy(origbuf, bp->b_data, origbufsize); 2941 free(origbuf, M_BIOBUF); 2942 } 2943 } 2944 } else { 2945 int desiredpages; 2946 2947 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2948 desiredpages = (size == 0) ? 0 : 2949 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2950 2951 if (bp->b_flags & B_MALLOC) 2952 panic("allocbuf: VMIO buffer can't be malloced"); 2953 /* 2954 * Set B_CACHE initially if buffer is 0 length or will become 2955 * 0-length. 2956 */ 2957 if (size == 0 || bp->b_bufsize == 0) 2958 bp->b_flags |= B_CACHE; 2959 2960 if (newbsize < bp->b_bufsize) { 2961 /* 2962 * DEV_BSIZE aligned new buffer size is less then the 2963 * DEV_BSIZE aligned existing buffer size. Figure out 2964 * if we have to remove any pages. 2965 */ 2966 if (desiredpages < bp->b_npages) { 2967 vm_page_t m; 2968 2969 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2970 for (i = desiredpages; i < bp->b_npages; i++) { 2971 /* 2972 * the page is not freed here -- it 2973 * is the responsibility of 2974 * vnode_pager_setsize 2975 */ 2976 m = bp->b_pages[i]; 2977 KASSERT(m != bogus_page, 2978 ("allocbuf: bogus page found")); 2979 while (vm_page_sleep_if_busy(m, TRUE, 2980 "biodep")) 2981 continue; 2982 2983 bp->b_pages[i] = NULL; 2984 vm_page_lock(m); 2985 vm_page_unwire(m, 0); 2986 vm_page_unlock(m); 2987 } 2988 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2989 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2990 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2991 bp->b_npages = desiredpages; 2992 } 2993 } else if (size > bp->b_bcount) { 2994 /* 2995 * We are growing the buffer, possibly in a 2996 * byte-granular fashion. 2997 */ 2998 vm_object_t obj; 2999 vm_offset_t toff; 3000 vm_offset_t tinc; 3001 3002 /* 3003 * Step 1, bring in the VM pages from the object, 3004 * allocating them if necessary. We must clear 3005 * B_CACHE if these pages are not valid for the 3006 * range covered by the buffer. 3007 */ 3008 3009 obj = bp->b_bufobj->bo_object; 3010 3011 VM_OBJECT_LOCK(obj); 3012 while (bp->b_npages < desiredpages) { 3013 vm_page_t m; 3014 3015 /* 3016 * We must allocate system pages since blocking 3017 * here could intefere with paging I/O, no 3018 * matter which process we are. 3019 * 3020 * We can only test VPO_BUSY here. Blocking on 3021 * m->busy might lead to a deadlock: 3022 * vm_fault->getpages->cluster_read->allocbuf 3023 * Thus, we specify VM_ALLOC_IGN_SBUSY. 3024 */ 3025 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3026 bp->b_npages, VM_ALLOC_NOBUSY | 3027 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3028 VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY | 3029 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3030 if (m->valid == 0) 3031 bp->b_flags &= ~B_CACHE; 3032 bp->b_pages[bp->b_npages] = m; 3033 ++bp->b_npages; 3034 } 3035 3036 /* 3037 * Step 2. We've loaded the pages into the buffer, 3038 * we have to figure out if we can still have B_CACHE 3039 * set. Note that B_CACHE is set according to the 3040 * byte-granular range ( bcount and size ), new the 3041 * aligned range ( newbsize ). 3042 * 3043 * The VM test is against m->valid, which is DEV_BSIZE 3044 * aligned. Needless to say, the validity of the data 3045 * needs to also be DEV_BSIZE aligned. Note that this 3046 * fails with NFS if the server or some other client 3047 * extends the file's EOF. If our buffer is resized, 3048 * B_CACHE may remain set! XXX 3049 */ 3050 3051 toff = bp->b_bcount; 3052 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3053 3054 while ((bp->b_flags & B_CACHE) && toff < size) { 3055 vm_pindex_t pi; 3056 3057 if (tinc > (size - toff)) 3058 tinc = size - toff; 3059 3060 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3061 PAGE_SHIFT; 3062 3063 vfs_buf_test_cache( 3064 bp, 3065 bp->b_offset, 3066 toff, 3067 tinc, 3068 bp->b_pages[pi] 3069 ); 3070 toff += tinc; 3071 tinc = PAGE_SIZE; 3072 } 3073 VM_OBJECT_UNLOCK(obj); 3074 3075 /* 3076 * Step 3, fixup the KVM pmap. Remember that 3077 * bp->b_data is relative to bp->b_offset, but 3078 * bp->b_offset may be offset into the first page. 3079 */ 3080 3081 bp->b_data = (caddr_t) 3082 trunc_page((vm_offset_t)bp->b_data); 3083 pmap_qenter( 3084 (vm_offset_t)bp->b_data, 3085 bp->b_pages, 3086 bp->b_npages 3087 ); 3088 3089 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3090 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3091 } 3092 } 3093 if (newbsize < bp->b_bufsize) 3094 bufspacewakeup(); 3095 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3096 bp->b_bcount = size; /* requested buffer size */ 3097 return 1; 3098} 3099 3100void 3101biodone(struct bio *bp) 3102{ 3103 struct mtx *mtxp; 3104 void (*done)(struct bio *); 3105 3106 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3107 mtx_lock(mtxp); 3108 bp->bio_flags |= BIO_DONE; 3109 done = bp->bio_done; 3110 if (done == NULL) 3111 wakeup(bp); 3112 mtx_unlock(mtxp); 3113 if (done != NULL) 3114 done(bp); 3115} 3116 3117/* 3118 * Wait for a BIO to finish. 3119 * 3120 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3121 * case is not yet clear. 3122 */ 3123int 3124biowait(struct bio *bp, const char *wchan) 3125{ 3126 struct mtx *mtxp; 3127 3128 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3129 mtx_lock(mtxp); 3130 while ((bp->bio_flags & BIO_DONE) == 0) 3131 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3132 mtx_unlock(mtxp); 3133 if (bp->bio_error != 0) 3134 return (bp->bio_error); 3135 if (!(bp->bio_flags & BIO_ERROR)) 3136 return (0); 3137 return (EIO); 3138} 3139 3140void 3141biofinish(struct bio *bp, struct devstat *stat, int error) 3142{ 3143 3144 if (error) { 3145 bp->bio_error = error; 3146 bp->bio_flags |= BIO_ERROR; 3147 } 3148 if (stat != NULL) 3149 devstat_end_transaction_bio(stat, bp); 3150 biodone(bp); 3151} 3152 3153/* 3154 * bufwait: 3155 * 3156 * Wait for buffer I/O completion, returning error status. The buffer 3157 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3158 * error and cleared. 3159 */ 3160int 3161bufwait(struct buf *bp) 3162{ 3163 if (bp->b_iocmd == BIO_READ) 3164 bwait(bp, PRIBIO, "biord"); 3165 else 3166 bwait(bp, PRIBIO, "biowr"); 3167 if (bp->b_flags & B_EINTR) { 3168 bp->b_flags &= ~B_EINTR; 3169 return (EINTR); 3170 } 3171 if (bp->b_ioflags & BIO_ERROR) { 3172 return (bp->b_error ? bp->b_error : EIO); 3173 } else { 3174 return (0); 3175 } 3176} 3177 3178 /* 3179 * Call back function from struct bio back up to struct buf. 3180 */ 3181static void 3182bufdonebio(struct bio *bip) 3183{ 3184 struct buf *bp; 3185 3186 bp = bip->bio_caller2; 3187 bp->b_resid = bp->b_bcount - bip->bio_completed; 3188 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3189 bp->b_ioflags = bip->bio_flags; 3190 bp->b_error = bip->bio_error; 3191 if (bp->b_error) 3192 bp->b_ioflags |= BIO_ERROR; 3193 bufdone(bp); 3194 g_destroy_bio(bip); 3195} 3196 3197void 3198dev_strategy(struct cdev *dev, struct buf *bp) 3199{ 3200 struct cdevsw *csw; 3201 struct bio *bip; 3202 3203 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3204 panic("b_iocmd botch"); 3205 for (;;) { 3206 bip = g_new_bio(); 3207 if (bip != NULL) 3208 break; 3209 /* Try again later */ 3210 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3211 } 3212 bip->bio_cmd = bp->b_iocmd; 3213 bip->bio_offset = bp->b_iooffset; 3214 bip->bio_length = bp->b_bcount; 3215 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3216 bip->bio_data = bp->b_data; 3217 bip->bio_done = bufdonebio; 3218 bip->bio_caller2 = bp; 3219 bip->bio_dev = dev; 3220 KASSERT(dev->si_refcount > 0, 3221 ("dev_strategy on un-referenced struct cdev *(%s)", 3222 devtoname(dev))); 3223 csw = dev_refthread(dev); 3224 if (csw == NULL) { 3225 g_destroy_bio(bip); 3226 bp->b_error = ENXIO; 3227 bp->b_ioflags = BIO_ERROR; 3228 bufdone(bp); 3229 return; 3230 } 3231 (*csw->d_strategy)(bip); 3232 dev_relthread(dev); 3233} 3234 3235/* 3236 * bufdone: 3237 * 3238 * Finish I/O on a buffer, optionally calling a completion function. 3239 * This is usually called from an interrupt so process blocking is 3240 * not allowed. 3241 * 3242 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3243 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3244 * assuming B_INVAL is clear. 3245 * 3246 * For the VMIO case, we set B_CACHE if the op was a read and no 3247 * read error occured, or if the op was a write. B_CACHE is never 3248 * set if the buffer is invalid or otherwise uncacheable. 3249 * 3250 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3251 * initiator to leave B_INVAL set to brelse the buffer out of existance 3252 * in the biodone routine. 3253 */ 3254void 3255bufdone(struct buf *bp) 3256{ 3257 struct bufobj *dropobj; 3258 void (*biodone)(struct buf *); 3259 3260 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3261 dropobj = NULL; 3262 3263 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3264 BUF_ASSERT_HELD(bp); 3265 3266 runningbufwakeup(bp); 3267 if (bp->b_iocmd == BIO_WRITE) 3268 dropobj = bp->b_bufobj; 3269 /* call optional completion function if requested */ 3270 if (bp->b_iodone != NULL) { 3271 biodone = bp->b_iodone; 3272 bp->b_iodone = NULL; 3273 (*biodone) (bp); 3274 if (dropobj) 3275 bufobj_wdrop(dropobj); 3276 return; 3277 } 3278 3279 bufdone_finish(bp); 3280 3281 if (dropobj) 3282 bufobj_wdrop(dropobj); 3283} 3284 3285void 3286bufdone_finish(struct buf *bp) 3287{ 3288 BUF_ASSERT_HELD(bp); 3289 3290 if (!LIST_EMPTY(&bp->b_dep)) 3291 buf_complete(bp); 3292 3293 if (bp->b_flags & B_VMIO) { 3294 int i; 3295 vm_ooffset_t foff; 3296 vm_page_t m; 3297 vm_object_t obj; 3298 int bogus, iosize; 3299 struct vnode *vp = bp->b_vp; 3300 3301 obj = bp->b_bufobj->bo_object; 3302 3303#if defined(VFS_BIO_DEBUG) 3304 mp_fixme("usecount and vflag accessed without locks."); 3305 if (vp->v_usecount == 0) { 3306 panic("biodone: zero vnode ref count"); 3307 } 3308 3309 KASSERT(vp->v_object != NULL, 3310 ("biodone: vnode %p has no vm_object", vp)); 3311#endif 3312 3313 foff = bp->b_offset; 3314 KASSERT(bp->b_offset != NOOFFSET, 3315 ("biodone: no buffer offset")); 3316 3317 VM_OBJECT_LOCK(obj); 3318#if defined(VFS_BIO_DEBUG) 3319 if (obj->paging_in_progress < bp->b_npages) { 3320 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3321 obj->paging_in_progress, bp->b_npages); 3322 } 3323#endif 3324 3325 /* 3326 * Set B_CACHE if the op was a normal read and no error 3327 * occured. B_CACHE is set for writes in the b*write() 3328 * routines. 3329 */ 3330 iosize = bp->b_bcount - bp->b_resid; 3331 if (bp->b_iocmd == BIO_READ && 3332 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3333 !(bp->b_ioflags & BIO_ERROR)) { 3334 bp->b_flags |= B_CACHE; 3335 } 3336 bogus = 0; 3337 for (i = 0; i < bp->b_npages; i++) { 3338 int bogusflag = 0; 3339 int resid; 3340 3341 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3342 if (resid > iosize) 3343 resid = iosize; 3344 3345 /* 3346 * cleanup bogus pages, restoring the originals 3347 */ 3348 m = bp->b_pages[i]; 3349 if (m == bogus_page) { 3350 bogus = bogusflag = 1; 3351 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3352 if (m == NULL) 3353 panic("biodone: page disappeared!"); 3354 bp->b_pages[i] = m; 3355 } 3356#if defined(VFS_BIO_DEBUG) 3357 if (OFF_TO_IDX(foff) != m->pindex) { 3358 printf( 3359"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3360 (intmax_t)foff, (uintmax_t)m->pindex); 3361 } 3362#endif 3363 3364 /* 3365 * In the write case, the valid and clean bits are 3366 * already changed correctly ( see bdwrite() ), so we 3367 * only need to do this here in the read case. 3368 */ 3369 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3370 KASSERT((m->dirty & vm_page_bits(foff & 3371 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3372 " page %p has unexpected dirty bits", m)); 3373 vfs_page_set_valid(bp, foff, m); 3374 } 3375 3376 /* 3377 * when debugging new filesystems or buffer I/O methods, this 3378 * is the most common error that pops up. if you see this, you 3379 * have not set the page busy flag correctly!!! 3380 */ 3381 if (m->busy == 0) { 3382 printf("biodone: page busy < 0, " 3383 "pindex: %d, foff: 0x(%x,%x), " 3384 "resid: %d, index: %d\n", 3385 (int) m->pindex, (int)(foff >> 32), 3386 (int) foff & 0xffffffff, resid, i); 3387 if (!vn_isdisk(vp, NULL)) 3388 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3389 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3390 (intmax_t) bp->b_lblkno, 3391 bp->b_flags, bp->b_npages); 3392 else 3393 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3394 (intmax_t) bp->b_lblkno, 3395 bp->b_flags, bp->b_npages); 3396 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3397 (u_long)m->valid, (u_long)m->dirty, 3398 m->wire_count); 3399 panic("biodone: page busy < 0\n"); 3400 } 3401 vm_page_io_finish(m); 3402 vm_object_pip_subtract(obj, 1); 3403 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3404 iosize -= resid; 3405 } 3406 vm_object_pip_wakeupn(obj, 0); 3407 VM_OBJECT_UNLOCK(obj); 3408 if (bogus) 3409 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3410 bp->b_pages, bp->b_npages); 3411 } 3412 3413 /* 3414 * For asynchronous completions, release the buffer now. The brelse 3415 * will do a wakeup there if necessary - so no need to do a wakeup 3416 * here in the async case. The sync case always needs to do a wakeup. 3417 */ 3418 3419 if (bp->b_flags & B_ASYNC) { 3420 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3421 brelse(bp); 3422 else 3423 bqrelse(bp); 3424 } else 3425 bdone(bp); 3426} 3427 3428/* 3429 * This routine is called in lieu of iodone in the case of 3430 * incomplete I/O. This keeps the busy status for pages 3431 * consistant. 3432 */ 3433void 3434vfs_unbusy_pages(struct buf *bp) 3435{ 3436 int i; 3437 vm_object_t obj; 3438 vm_page_t m; 3439 3440 runningbufwakeup(bp); 3441 if (!(bp->b_flags & B_VMIO)) 3442 return; 3443 3444 obj = bp->b_bufobj->bo_object; 3445 VM_OBJECT_LOCK(obj); 3446 for (i = 0; i < bp->b_npages; i++) { 3447 m = bp->b_pages[i]; 3448 if (m == bogus_page) { 3449 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3450 if (!m) 3451 panic("vfs_unbusy_pages: page missing\n"); 3452 bp->b_pages[i] = m; 3453 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3454 bp->b_pages, bp->b_npages); 3455 } 3456 vm_object_pip_subtract(obj, 1); 3457 vm_page_io_finish(m); 3458 } 3459 vm_object_pip_wakeupn(obj, 0); 3460 VM_OBJECT_UNLOCK(obj); 3461} 3462 3463/* 3464 * vfs_page_set_valid: 3465 * 3466 * Set the valid bits in a page based on the supplied offset. The 3467 * range is restricted to the buffer's size. 3468 * 3469 * This routine is typically called after a read completes. 3470 */ 3471static void 3472vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3473{ 3474 vm_ooffset_t eoff; 3475 3476 /* 3477 * Compute the end offset, eoff, such that [off, eoff) does not span a 3478 * page boundary and eoff is not greater than the end of the buffer. 3479 * The end of the buffer, in this case, is our file EOF, not the 3480 * allocation size of the buffer. 3481 */ 3482 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3483 if (eoff > bp->b_offset + bp->b_bcount) 3484 eoff = bp->b_offset + bp->b_bcount; 3485 3486 /* 3487 * Set valid range. This is typically the entire buffer and thus the 3488 * entire page. 3489 */ 3490 if (eoff > off) 3491 vm_page_set_valid(m, off & PAGE_MASK, eoff - off); 3492} 3493 3494/* 3495 * vfs_page_set_validclean: 3496 * 3497 * Set the valid bits and clear the dirty bits in a page based on the 3498 * supplied offset. The range is restricted to the buffer's size. 3499 */ 3500static void 3501vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3502{ 3503 vm_ooffset_t soff, eoff; 3504 3505 /* 3506 * Start and end offsets in buffer. eoff - soff may not cross a 3507 * page boundry or cross the end of the buffer. The end of the 3508 * buffer, in this case, is our file EOF, not the allocation size 3509 * of the buffer. 3510 */ 3511 soff = off; 3512 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3513 if (eoff > bp->b_offset + bp->b_bcount) 3514 eoff = bp->b_offset + bp->b_bcount; 3515 3516 /* 3517 * Set valid range. This is typically the entire buffer and thus the 3518 * entire page. 3519 */ 3520 if (eoff > soff) { 3521 vm_page_set_validclean( 3522 m, 3523 (vm_offset_t) (soff & PAGE_MASK), 3524 (vm_offset_t) (eoff - soff) 3525 ); 3526 } 3527} 3528 3529/* 3530 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If 3531 * any page is busy, drain the flag. 3532 */ 3533static void 3534vfs_drain_busy_pages(struct buf *bp) 3535{ 3536 vm_page_t m; 3537 int i, last_busied; 3538 3539 VM_OBJECT_LOCK_ASSERT(bp->b_bufobj->bo_object, MA_OWNED); 3540 last_busied = 0; 3541 for (i = 0; i < bp->b_npages; i++) { 3542 m = bp->b_pages[i]; 3543 if ((m->oflags & VPO_BUSY) != 0) { 3544 for (; last_busied < i; last_busied++) 3545 vm_page_busy(bp->b_pages[last_busied]); 3546 while ((m->oflags & VPO_BUSY) != 0) 3547 vm_page_sleep(m, "vbpage"); 3548 } 3549 } 3550 for (i = 0; i < last_busied; i++) 3551 vm_page_wakeup(bp->b_pages[i]); 3552} 3553 3554/* 3555 * This routine is called before a device strategy routine. 3556 * It is used to tell the VM system that paging I/O is in 3557 * progress, and treat the pages associated with the buffer 3558 * almost as being VPO_BUSY. Also the object paging_in_progress 3559 * flag is handled to make sure that the object doesn't become 3560 * inconsistant. 3561 * 3562 * Since I/O has not been initiated yet, certain buffer flags 3563 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3564 * and should be ignored. 3565 */ 3566void 3567vfs_busy_pages(struct buf *bp, int clear_modify) 3568{ 3569 int i, bogus; 3570 vm_object_t obj; 3571 vm_ooffset_t foff; 3572 vm_page_t m; 3573 3574 if (!(bp->b_flags & B_VMIO)) 3575 return; 3576 3577 obj = bp->b_bufobj->bo_object; 3578 foff = bp->b_offset; 3579 KASSERT(bp->b_offset != NOOFFSET, 3580 ("vfs_busy_pages: no buffer offset")); 3581 VM_OBJECT_LOCK(obj); 3582 vfs_drain_busy_pages(bp); 3583 if (bp->b_bufsize != 0) 3584 vfs_setdirty_locked_object(bp); 3585 bogus = 0; 3586 for (i = 0; i < bp->b_npages; i++) { 3587 m = bp->b_pages[i]; 3588 3589 if ((bp->b_flags & B_CLUSTER) == 0) { 3590 vm_object_pip_add(obj, 1); 3591 vm_page_io_start(m); 3592 } 3593 /* 3594 * When readying a buffer for a read ( i.e 3595 * clear_modify == 0 ), it is important to do 3596 * bogus_page replacement for valid pages in 3597 * partially instantiated buffers. Partially 3598 * instantiated buffers can, in turn, occur when 3599 * reconstituting a buffer from its VM backing store 3600 * base. We only have to do this if B_CACHE is 3601 * clear ( which causes the I/O to occur in the 3602 * first place ). The replacement prevents the read 3603 * I/O from overwriting potentially dirty VM-backed 3604 * pages. XXX bogus page replacement is, uh, bogus. 3605 * It may not work properly with small-block devices. 3606 * We need to find a better way. 3607 */ 3608 if (clear_modify) { 3609 pmap_remove_write(m); 3610 vfs_page_set_validclean(bp, foff, m); 3611 } else if (m->valid == VM_PAGE_BITS_ALL && 3612 (bp->b_flags & B_CACHE) == 0) { 3613 bp->b_pages[i] = bogus_page; 3614 bogus++; 3615 } 3616 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3617 } 3618 VM_OBJECT_UNLOCK(obj); 3619 if (bogus) 3620 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3621 bp->b_pages, bp->b_npages); 3622} 3623 3624/* 3625 * vfs_bio_set_valid: 3626 * 3627 * Set the range within the buffer to valid. The range is 3628 * relative to the beginning of the buffer, b_offset. Note that 3629 * b_offset itself may be offset from the beginning of the first 3630 * page. 3631 */ 3632void 3633vfs_bio_set_valid(struct buf *bp, int base, int size) 3634{ 3635 int i, n; 3636 vm_page_t m; 3637 3638 if (!(bp->b_flags & B_VMIO)) 3639 return; 3640 3641 /* 3642 * Fixup base to be relative to beginning of first page. 3643 * Set initial n to be the maximum number of bytes in the 3644 * first page that can be validated. 3645 */ 3646 base += (bp->b_offset & PAGE_MASK); 3647 n = PAGE_SIZE - (base & PAGE_MASK); 3648 3649 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3650 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3651 m = bp->b_pages[i]; 3652 if (n > size) 3653 n = size; 3654 vm_page_set_valid(m, base & PAGE_MASK, n); 3655 base += n; 3656 size -= n; 3657 n = PAGE_SIZE; 3658 } 3659 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3660} 3661 3662/* 3663 * vfs_bio_clrbuf: 3664 * 3665 * If the specified buffer is a non-VMIO buffer, clear the entire 3666 * buffer. If the specified buffer is a VMIO buffer, clear and 3667 * validate only the previously invalid portions of the buffer. 3668 * This routine essentially fakes an I/O, so we need to clear 3669 * BIO_ERROR and B_INVAL. 3670 * 3671 * Note that while we only theoretically need to clear through b_bcount, 3672 * we go ahead and clear through b_bufsize. 3673 */ 3674void 3675vfs_bio_clrbuf(struct buf *bp) 3676{ 3677 int i, j, mask; 3678 caddr_t sa, ea; 3679 3680 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3681 clrbuf(bp); 3682 return; 3683 } 3684 bp->b_flags &= ~B_INVAL; 3685 bp->b_ioflags &= ~BIO_ERROR; 3686 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3687 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3688 (bp->b_offset & PAGE_MASK) == 0) { 3689 if (bp->b_pages[0] == bogus_page) 3690 goto unlock; 3691 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3692 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3693 if ((bp->b_pages[0]->valid & mask) == mask) 3694 goto unlock; 3695 if ((bp->b_pages[0]->valid & mask) == 0) { 3696 bzero(bp->b_data, bp->b_bufsize); 3697 bp->b_pages[0]->valid |= mask; 3698 goto unlock; 3699 } 3700 } 3701 ea = sa = bp->b_data; 3702 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3703 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3704 ea = (caddr_t)(vm_offset_t)ulmin( 3705 (u_long)(vm_offset_t)ea, 3706 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3707 if (bp->b_pages[i] == bogus_page) 3708 continue; 3709 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3710 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3711 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3712 if ((bp->b_pages[i]->valid & mask) == mask) 3713 continue; 3714 if ((bp->b_pages[i]->valid & mask) == 0) 3715 bzero(sa, ea - sa); 3716 else { 3717 for (; sa < ea; sa += DEV_BSIZE, j++) { 3718 if ((bp->b_pages[i]->valid & (1 << j)) == 0) 3719 bzero(sa, DEV_BSIZE); 3720 } 3721 } 3722 bp->b_pages[i]->valid |= mask; 3723 } 3724unlock: 3725 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3726 bp->b_resid = 0; 3727} 3728 3729/* 3730 * vm_hold_load_pages and vm_hold_free_pages get pages into 3731 * a buffers address space. The pages are anonymous and are 3732 * not associated with a file object. 3733 */ 3734static void 3735vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3736{ 3737 vm_offset_t pg; 3738 vm_page_t p; 3739 int index; 3740 3741 to = round_page(to); 3742 from = round_page(from); 3743 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3744 3745 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3746tryagain: 3747 /* 3748 * note: must allocate system pages since blocking here 3749 * could interfere with paging I/O, no matter which 3750 * process we are. 3751 */ 3752 p = vm_page_alloc(NULL, pg >> PAGE_SHIFT, VM_ALLOC_NOOBJ | 3753 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3754 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 3755 if (!p) { 3756 VM_WAIT; 3757 goto tryagain; 3758 } 3759 pmap_qenter(pg, &p, 1); 3760 bp->b_pages[index] = p; 3761 } 3762 bp->b_npages = index; 3763} 3764 3765/* Return pages associated with this buf to the vm system */ 3766static void 3767vm_hold_free_pages(struct buf *bp, int newbsize) 3768{ 3769 vm_offset_t from; 3770 vm_page_t p; 3771 int index, newnpages; 3772 3773 from = round_page((vm_offset_t)bp->b_data + newbsize); 3774 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3775 if (bp->b_npages > newnpages) 3776 pmap_qremove(from, bp->b_npages - newnpages); 3777 for (index = newnpages; index < bp->b_npages; index++) { 3778 p = bp->b_pages[index]; 3779 bp->b_pages[index] = NULL; 3780 if (p->busy != 0) 3781 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3782 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 3783 p->wire_count--; 3784 vm_page_free(p); 3785 atomic_subtract_int(&cnt.v_wire_count, 1); 3786 } 3787 bp->b_npages = newnpages; 3788} 3789 3790/* 3791 * Map an IO request into kernel virtual address space. 3792 * 3793 * All requests are (re)mapped into kernel VA space. 3794 * Notice that we use b_bufsize for the size of the buffer 3795 * to be mapped. b_bcount might be modified by the driver. 3796 * 3797 * Note that even if the caller determines that the address space should 3798 * be valid, a race or a smaller-file mapped into a larger space may 3799 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3800 * check the return value. 3801 */ 3802int 3803vmapbuf(struct buf *bp) 3804{ 3805 caddr_t addr, kva; 3806 vm_prot_t prot; 3807 int pidx, i; 3808 struct vm_page *m; 3809 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3810 3811 if (bp->b_bufsize < 0) 3812 return (-1); 3813 prot = VM_PROT_READ; 3814 if (bp->b_iocmd == BIO_READ) 3815 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3816 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3817 addr < bp->b_data + bp->b_bufsize; 3818 addr += PAGE_SIZE, pidx++) { 3819 /* 3820 * Do the vm_fault if needed; do the copy-on-write thing 3821 * when reading stuff off device into memory. 3822 * 3823 * NOTE! Must use pmap_extract() because addr may be in 3824 * the userland address space, and kextract is only guarenteed 3825 * to work for the kernland address space (see: sparc64 port). 3826 */ 3827retry: 3828 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3829 prot) < 0) { 3830 for (i = 0; i < pidx; ++i) { 3831 vm_page_lock(bp->b_pages[i]); 3832 vm_page_unhold(bp->b_pages[i]); 3833 vm_page_unlock(bp->b_pages[i]); 3834 bp->b_pages[i] = NULL; 3835 } 3836 return(-1); 3837 } 3838 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3839 if (m == NULL) 3840 goto retry; 3841 bp->b_pages[pidx] = m; 3842 } 3843 if (pidx > btoc(MAXPHYS)) 3844 panic("vmapbuf: mapped more than MAXPHYS"); 3845 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3846 3847 kva = bp->b_saveaddr; 3848 bp->b_npages = pidx; 3849 bp->b_saveaddr = bp->b_data; 3850 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3851 return(0); 3852} 3853 3854/* 3855 * Free the io map PTEs associated with this IO operation. 3856 * We also invalidate the TLB entries and restore the original b_addr. 3857 */ 3858void 3859vunmapbuf(struct buf *bp) 3860{ 3861 int pidx; 3862 int npages; 3863 3864 npages = bp->b_npages; 3865 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3866 for (pidx = 0; pidx < npages; pidx++) { 3867 vm_page_lock(bp->b_pages[pidx]); 3868 vm_page_unhold(bp->b_pages[pidx]); 3869 vm_page_unlock(bp->b_pages[pidx]); 3870 } 3871 3872 bp->b_data = bp->b_saveaddr; 3873} 3874 3875void 3876bdone(struct buf *bp) 3877{ 3878 struct mtx *mtxp; 3879 3880 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3881 mtx_lock(mtxp); 3882 bp->b_flags |= B_DONE; 3883 wakeup(bp); 3884 mtx_unlock(mtxp); 3885} 3886 3887void 3888bwait(struct buf *bp, u_char pri, const char *wchan) 3889{ 3890 struct mtx *mtxp; 3891 3892 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3893 mtx_lock(mtxp); 3894 while ((bp->b_flags & B_DONE) == 0) 3895 msleep(bp, mtxp, pri, wchan, 0); 3896 mtx_unlock(mtxp); 3897} 3898 3899int 3900bufsync(struct bufobj *bo, int waitfor) 3901{ 3902 3903 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 3904} 3905 3906void 3907bufstrategy(struct bufobj *bo, struct buf *bp) 3908{ 3909 int i = 0; 3910 struct vnode *vp; 3911 3912 vp = bp->b_vp; 3913 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3914 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3915 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3916 i = VOP_STRATEGY(vp, bp); 3917 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3918} 3919 3920void 3921bufobj_wrefl(struct bufobj *bo) 3922{ 3923 3924 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3925 ASSERT_BO_LOCKED(bo); 3926 bo->bo_numoutput++; 3927} 3928 3929void 3930bufobj_wref(struct bufobj *bo) 3931{ 3932 3933 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3934 BO_LOCK(bo); 3935 bo->bo_numoutput++; 3936 BO_UNLOCK(bo); 3937} 3938 3939void 3940bufobj_wdrop(struct bufobj *bo) 3941{ 3942 3943 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3944 BO_LOCK(bo); 3945 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3946 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3947 bo->bo_flag &= ~BO_WWAIT; 3948 wakeup(&bo->bo_numoutput); 3949 } 3950 BO_UNLOCK(bo); 3951} 3952 3953int 3954bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3955{ 3956 int error; 3957 3958 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3959 ASSERT_BO_LOCKED(bo); 3960 error = 0; 3961 while (bo->bo_numoutput) { 3962 bo->bo_flag |= BO_WWAIT; 3963 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3964 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3965 if (error) 3966 break; 3967 } 3968 return (error); 3969} 3970 3971void 3972bpin(struct buf *bp) 3973{ 3974 struct mtx *mtxp; 3975 3976 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3977 mtx_lock(mtxp); 3978 bp->b_pin_count++; 3979 mtx_unlock(mtxp); 3980} 3981 3982void 3983bunpin(struct buf *bp) 3984{ 3985 struct mtx *mtxp; 3986 3987 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3988 mtx_lock(mtxp); 3989 if (--bp->b_pin_count == 0) 3990 wakeup(bp); 3991 mtx_unlock(mtxp); 3992} 3993 3994void 3995bunpin_wait(struct buf *bp) 3996{ 3997 struct mtx *mtxp; 3998 3999 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4000 mtx_lock(mtxp); 4001 while (bp->b_pin_count > 0) 4002 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4003 mtx_unlock(mtxp); 4004} 4005 4006#include "opt_ddb.h" 4007#ifdef DDB 4008#include <ddb/ddb.h> 4009 4010/* DDB command to show buffer data */ 4011DB_SHOW_COMMAND(buffer, db_show_buffer) 4012{ 4013 /* get args */ 4014 struct buf *bp = (struct buf *)addr; 4015 4016 if (!have_addr) { 4017 db_printf("usage: show buffer <addr>\n"); 4018 return; 4019 } 4020 4021 db_printf("buf at %p\n", bp); 4022 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4023 db_printf( 4024 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4025 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n", 4026 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4027 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4028 bp->b_dep.lh_first); 4029 if (bp->b_npages) { 4030 int i; 4031 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4032 for (i = 0; i < bp->b_npages; i++) { 4033 vm_page_t m; 4034 m = bp->b_pages[i]; 4035 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4036 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4037 if ((i + 1) < bp->b_npages) 4038 db_printf(","); 4039 } 4040 db_printf("\n"); 4041 } 4042 db_printf(" "); 4043 lockmgr_printinfo(&bp->b_lock); 4044} 4045 4046DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4047{ 4048 struct buf *bp; 4049 int i; 4050 4051 for (i = 0; i < nbuf; i++) { 4052 bp = &buf[i]; 4053 if (BUF_ISLOCKED(bp)) { 4054 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4055 db_printf("\n"); 4056 } 4057 } 4058} 4059 4060DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4061{ 4062 struct vnode *vp; 4063 struct buf *bp; 4064 4065 if (!have_addr) { 4066 db_printf("usage: show vnodebufs <addr>\n"); 4067 return; 4068 } 4069 vp = (struct vnode *)addr; 4070 db_printf("Clean buffers:\n"); 4071 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4072 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4073 db_printf("\n"); 4074 } 4075 db_printf("Dirty buffers:\n"); 4076 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4077 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4078 db_printf("\n"); 4079 } 4080} 4081 4082DB_COMMAND(countfreebufs, db_coundfreebufs) 4083{ 4084 struct buf *bp; 4085 int i, used = 0, nfree = 0; 4086 4087 if (have_addr) { 4088 db_printf("usage: countfreebufs\n"); 4089 return; 4090 } 4091 4092 for (i = 0; i < nbuf; i++) { 4093 bp = &buf[i]; 4094 if ((bp->b_vflags & BV_INFREECNT) != 0) 4095 nfree++; 4096 else 4097 used++; 4098 } 4099 4100 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4101 nfree + used); 4102 db_printf("numfreebuffers is %d\n", numfreebuffers); 4103} 4104#endif /* DDB */ 4105