vfs_bio.c revision 209605
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 209605 2010-06-30 04:52:42Z alc $"); 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, vm_offset_t from, 99 vm_offset_t to); 100static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 101 vm_offset_t to); 102static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 103static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 104 vm_page_t m); 105static void vfs_drain_busy_pages(struct buf *bp); 106static void vfs_clean_pages_dirty_buf(struct buf *bp); 107static void vfs_setdirty_locked_object(struct buf *bp); 108static void vfs_vmio_release(struct buf *bp); 109static int vfs_bio_clcheck(struct vnode *vp, int size, 110 daddr_t lblkno, daddr_t blkno); 111static int buf_do_flush(struct vnode *vp); 112static int flushbufqueues(struct vnode *, int, int); 113static void buf_daemon(void); 114static void bremfreel(struct buf *bp); 115#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 116 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 117static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 118#endif 119 120int vmiodirenable = TRUE; 121SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 122 "Use the VM system for directory writes"); 123long runningbufspace; 124SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 125 "Amount of presently outstanding async buffer io"); 126static long bufspace; 127#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 128 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 129SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 130 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 131#else 132SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 133 "Virtual memory used for buffers"); 134#endif 135static long maxbufspace; 136SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 137 "Maximum allowed value of bufspace (including buf_daemon)"); 138static long bufmallocspace; 139SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 140 "Amount of malloced memory for buffers"); 141static long maxbufmallocspace; 142SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 143 "Maximum amount of malloced memory for buffers"); 144static long lobufspace; 145SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 146 "Minimum amount of buffers we want to have"); 147long hibufspace; 148SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 149 "Maximum allowed value of bufspace (excluding buf_daemon)"); 150static int bufreusecnt; 151SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 152 "Number of times we have reused a buffer"); 153static int buffreekvacnt; 154SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 155 "Number of times we have freed the KVA space from some buffer"); 156static int bufdefragcnt; 157SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 158 "Number of times we have had to repeat buffer allocation to defragment"); 159static long lorunningspace; 160SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 161 "Minimum preferred space used for in-progress I/O"); 162static long hirunningspace; 163SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 164 "Maximum amount of space to use for in-progress I/O"); 165int dirtybufferflushes; 166SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 167 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 168int bdwriteskip; 169SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 170 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 171int altbufferflushes; 172SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 173 0, "Number of fsync flushes to limit dirty buffers"); 174static int recursiveflushes; 175SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 176 0, "Number of flushes skipped due to being recursive"); 177static int numdirtybuffers; 178SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 179 "Number of buffers that are dirty (has unwritten changes) at the moment"); 180static int lodirtybuffers; 181SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 182 "How many buffers we want to have free before bufdaemon can sleep"); 183static int hidirtybuffers; 184SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 185 "When the number of dirty buffers is considered severe"); 186int dirtybufthresh; 187SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 188 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 189static int numfreebuffers; 190SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 191 "Number of free buffers"); 192static int lofreebuffers; 193SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 194 "XXX Unused"); 195static int hifreebuffers; 196SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 197 "XXX Complicatedly unused"); 198static int getnewbufcalls; 199SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 200 "Number of calls to getnewbuf"); 201static int getnewbufrestarts; 202SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 203 "Number of times getnewbuf has had to restart a buffer aquisition"); 204static int flushbufqtarget = 100; 205SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 206 "Amount of work to do in flushbufqueues when helping bufdaemon"); 207static long notbufdflashes; 208SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0, 209 "Number of dirty buffer flushes done by the bufdaemon helpers"); 210 211/* 212 * Wakeup point for bufdaemon, as well as indicator of whether it is already 213 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 214 * is idling. 215 */ 216static int bd_request; 217 218/* 219 * Request for the buf daemon to write more buffers than is indicated by 220 * lodirtybuf. This may be necessary to push out excess dependencies or 221 * defragment the address space where a simple count of the number of dirty 222 * buffers is insufficient to characterize the demand for flushing them. 223 */ 224static int bd_speedupreq; 225 226/* 227 * This lock synchronizes access to bd_request. 228 */ 229static struct mtx bdlock; 230 231/* 232 * bogus page -- for I/O to/from partially complete buffers 233 * this is a temporary solution to the problem, but it is not 234 * really that bad. it would be better to split the buffer 235 * for input in the case of buffers partially already in memory, 236 * but the code is intricate enough already. 237 */ 238vm_page_t bogus_page; 239 240/* 241 * Synchronization (sleep/wakeup) variable for active buffer space requests. 242 * Set when wait starts, cleared prior to wakeup(). 243 * Used in runningbufwakeup() and waitrunningbufspace(). 244 */ 245static int runningbufreq; 246 247/* 248 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 249 * waitrunningbufspace(). 250 */ 251static struct mtx rbreqlock; 252 253/* 254 * Synchronization (sleep/wakeup) variable for buffer requests. 255 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 256 * by and/or. 257 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 258 * getnewbuf(), and getblk(). 259 */ 260static int needsbuffer; 261 262/* 263 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 264 */ 265static struct mtx nblock; 266 267/* 268 * Definitions for the buffer free lists. 269 */ 270#define BUFFER_QUEUES 6 /* number of free buffer queues */ 271 272#define QUEUE_NONE 0 /* on no queue */ 273#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 274#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 275#define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ 276#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 277#define QUEUE_EMPTY 5 /* empty buffer headers */ 278#define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 279 280/* Queues for free buffers with various properties */ 281static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 282 283/* Lock for the bufqueues */ 284static struct mtx bqlock; 285 286/* 287 * Single global constant for BUF_WMESG, to avoid getting multiple references. 288 * buf_wmesg is referred from macros. 289 */ 290const char *buf_wmesg = BUF_WMESG; 291 292#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 293#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 294#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 295#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 296 297#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 298 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 299static int 300sysctl_bufspace(SYSCTL_HANDLER_ARGS) 301{ 302 long lvalue; 303 int ivalue; 304 305 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 306 return (sysctl_handle_long(oidp, arg1, arg2, req)); 307 lvalue = *(long *)arg1; 308 if (lvalue > INT_MAX) 309 /* On overflow, still write out a long to trigger ENOMEM. */ 310 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 311 ivalue = lvalue; 312 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 313} 314#endif 315 316#ifdef DIRECTIO 317extern void ffs_rawread_setup(void); 318#endif /* DIRECTIO */ 319/* 320 * numdirtywakeup: 321 * 322 * If someone is blocked due to there being too many dirty buffers, 323 * and numdirtybuffers is now reasonable, wake them up. 324 */ 325 326static __inline void 327numdirtywakeup(int level) 328{ 329 330 if (numdirtybuffers <= level) { 331 mtx_lock(&nblock); 332 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 333 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 334 wakeup(&needsbuffer); 335 } 336 mtx_unlock(&nblock); 337 } 338} 339 340/* 341 * bufspacewakeup: 342 * 343 * Called when buffer space is potentially available for recovery. 344 * getnewbuf() will block on this flag when it is unable to free 345 * sufficient buffer space. Buffer space becomes recoverable when 346 * bp's get placed back in the queues. 347 */ 348 349static __inline void 350bufspacewakeup(void) 351{ 352 353 /* 354 * If someone is waiting for BUF space, wake them up. Even 355 * though we haven't freed the kva space yet, the waiting 356 * process will be able to now. 357 */ 358 mtx_lock(&nblock); 359 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 360 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 361 wakeup(&needsbuffer); 362 } 363 mtx_unlock(&nblock); 364} 365 366/* 367 * runningbufwakeup() - in-progress I/O accounting. 368 * 369 */ 370void 371runningbufwakeup(struct buf *bp) 372{ 373 374 if (bp->b_runningbufspace) { 375 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace); 376 bp->b_runningbufspace = 0; 377 mtx_lock(&rbreqlock); 378 if (runningbufreq && runningbufspace <= lorunningspace) { 379 runningbufreq = 0; 380 wakeup(&runningbufreq); 381 } 382 mtx_unlock(&rbreqlock); 383 } 384} 385 386/* 387 * bufcountwakeup: 388 * 389 * Called when a buffer has been added to one of the free queues to 390 * account for the buffer and to wakeup anyone waiting for free buffers. 391 * This typically occurs when large amounts of metadata are being handled 392 * by the buffer cache ( else buffer space runs out first, usually ). 393 */ 394 395static __inline void 396bufcountwakeup(struct buf *bp) 397{ 398 int old; 399 400 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 401 ("buf %p already counted as free", bp)); 402 bp->b_vflags |= BV_INFREECNT; 403 old = atomic_fetchadd_int(&numfreebuffers, 1); 404 KASSERT(old >= 0 && old < nbuf, 405 ("numfreebuffers climbed to %d", old + 1)); 406 mtx_lock(&nblock); 407 if (needsbuffer) { 408 needsbuffer &= ~VFS_BIO_NEED_ANY; 409 if (numfreebuffers >= hifreebuffers) 410 needsbuffer &= ~VFS_BIO_NEED_FREE; 411 wakeup(&needsbuffer); 412 } 413 mtx_unlock(&nblock); 414} 415 416/* 417 * waitrunningbufspace() 418 * 419 * runningbufspace is a measure of the amount of I/O currently 420 * running. This routine is used in async-write situations to 421 * prevent creating huge backups of pending writes to a device. 422 * Only asynchronous writes are governed by this function. 423 * 424 * Reads will adjust runningbufspace, but will not block based on it. 425 * The read load has a side effect of reducing the allowed write load. 426 * 427 * This does NOT turn an async write into a sync write. It waits 428 * for earlier writes to complete and generally returns before the 429 * caller's write has reached the device. 430 */ 431void 432waitrunningbufspace(void) 433{ 434 435 mtx_lock(&rbreqlock); 436 while (runningbufspace > hirunningspace) { 437 ++runningbufreq; 438 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 439 } 440 mtx_unlock(&rbreqlock); 441} 442 443 444/* 445 * vfs_buf_test_cache: 446 * 447 * Called when a buffer is extended. This function clears the B_CACHE 448 * bit if the newly extended portion of the buffer does not contain 449 * valid data. 450 */ 451static __inline 452void 453vfs_buf_test_cache(struct buf *bp, 454 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 455 vm_page_t m) 456{ 457 458 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 459 if (bp->b_flags & B_CACHE) { 460 int base = (foff + off) & PAGE_MASK; 461 if (vm_page_is_valid(m, base, size) == 0) 462 bp->b_flags &= ~B_CACHE; 463 } 464} 465 466/* Wake up the buffer daemon if necessary */ 467static __inline 468void 469bd_wakeup(int dirtybuflevel) 470{ 471 472 mtx_lock(&bdlock); 473 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 474 bd_request = 1; 475 wakeup(&bd_request); 476 } 477 mtx_unlock(&bdlock); 478} 479 480/* 481 * bd_speedup - speedup the buffer cache flushing code 482 */ 483 484void 485bd_speedup(void) 486{ 487 int needwake; 488 489 mtx_lock(&bdlock); 490 needwake = 0; 491 if (bd_speedupreq == 0 || bd_request == 0) 492 needwake = 1; 493 bd_speedupreq = 1; 494 bd_request = 1; 495 if (needwake) 496 wakeup(&bd_request); 497 mtx_unlock(&bdlock); 498} 499 500/* 501 * Calculating buffer cache scaling values and reserve space for buffer 502 * headers. This is called during low level kernel initialization and 503 * may be called more then once. We CANNOT write to the memory area 504 * being reserved at this time. 505 */ 506caddr_t 507kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 508{ 509 int tuned_nbuf; 510 long maxbuf; 511 512 /* 513 * physmem_est is in pages. Convert it to kilobytes (assumes 514 * PAGE_SIZE is >= 1K) 515 */ 516 physmem_est = physmem_est * (PAGE_SIZE / 1024); 517 518 /* 519 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 520 * For the first 64MB of ram nominally allocate sufficient buffers to 521 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 522 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 523 * the buffer cache we limit the eventual kva reservation to 524 * maxbcache bytes. 525 * 526 * factor represents the 1/4 x ram conversion. 527 */ 528 if (nbuf == 0) { 529 int factor = 4 * BKVASIZE / 1024; 530 531 nbuf = 50; 532 if (physmem_est > 4096) 533 nbuf += min((physmem_est - 4096) / factor, 534 65536 / factor); 535 if (physmem_est > 65536) 536 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 537 538 if (maxbcache && nbuf > maxbcache / BKVASIZE) 539 nbuf = maxbcache / BKVASIZE; 540 tuned_nbuf = 1; 541 } else 542 tuned_nbuf = 0; 543 544 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 545 maxbuf = (LONG_MAX / 3) / BKVASIZE; 546 if (nbuf > maxbuf) { 547 if (!tuned_nbuf) 548 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 549 maxbuf); 550 nbuf = maxbuf; 551 } 552 553 /* 554 * swbufs are used as temporary holders for I/O, such as paging I/O. 555 * We have no less then 16 and no more then 256. 556 */ 557 nswbuf = max(min(nbuf/4, 256), 16); 558#ifdef NSWBUF_MIN 559 if (nswbuf < NSWBUF_MIN) 560 nswbuf = NSWBUF_MIN; 561#endif 562#ifdef DIRECTIO 563 ffs_rawread_setup(); 564#endif 565 566 /* 567 * Reserve space for the buffer cache buffers 568 */ 569 swbuf = (void *)v; 570 v = (caddr_t)(swbuf + nswbuf); 571 buf = (void *)v; 572 v = (caddr_t)(buf + nbuf); 573 574 return(v); 575} 576 577/* Initialize the buffer subsystem. Called before use of any buffers. */ 578void 579bufinit(void) 580{ 581 struct buf *bp; 582 int i; 583 584 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 585 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 586 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 587 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 588 589 /* next, make a null set of free lists */ 590 for (i = 0; i < BUFFER_QUEUES; i++) 591 TAILQ_INIT(&bufqueues[i]); 592 593 /* finally, initialize each buffer header and stick on empty q */ 594 for (i = 0; i < nbuf; i++) { 595 bp = &buf[i]; 596 bzero(bp, sizeof *bp); 597 bp->b_flags = B_INVAL; /* we're just an empty header */ 598 bp->b_rcred = NOCRED; 599 bp->b_wcred = NOCRED; 600 bp->b_qindex = QUEUE_EMPTY; 601 bp->b_vflags = BV_INFREECNT; /* buf is counted as free */ 602 bp->b_xflags = 0; 603 LIST_INIT(&bp->b_dep); 604 BUF_LOCKINIT(bp); 605 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 606 } 607 608 /* 609 * maxbufspace is the absolute maximum amount of buffer space we are 610 * allowed to reserve in KVM and in real terms. The absolute maximum 611 * is nominally used by buf_daemon. hibufspace is the nominal maximum 612 * used by most other processes. The differential is required to 613 * ensure that buf_daemon is able to run when other processes might 614 * be blocked waiting for buffer space. 615 * 616 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 617 * this may result in KVM fragmentation which is not handled optimally 618 * by the system. 619 */ 620 maxbufspace = (long)nbuf * BKVASIZE; 621 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 622 lobufspace = hibufspace - MAXBSIZE; 623 624 lorunningspace = 512 * 1024; 625 hirunningspace = 1024 * 1024; 626 627/* 628 * Limit the amount of malloc memory since it is wired permanently into 629 * the kernel space. Even though this is accounted for in the buffer 630 * allocation, we don't want the malloced region to grow uncontrolled. 631 * The malloc scheme improves memory utilization significantly on average 632 * (small) directories. 633 */ 634 maxbufmallocspace = hibufspace / 20; 635 636/* 637 * Reduce the chance of a deadlock occuring by limiting the number 638 * of delayed-write dirty buffers we allow to stack up. 639 */ 640 hidirtybuffers = nbuf / 4 + 20; 641 dirtybufthresh = hidirtybuffers * 9 / 10; 642 numdirtybuffers = 0; 643/* 644 * To support extreme low-memory systems, make sure hidirtybuffers cannot 645 * eat up all available buffer space. This occurs when our minimum cannot 646 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 647 * BKVASIZE'd (8K) buffers. 648 */ 649 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 650 hidirtybuffers >>= 1; 651 } 652 lodirtybuffers = hidirtybuffers / 2; 653 654/* 655 * Try to keep the number of free buffers in the specified range, 656 * and give special processes (e.g. like buf_daemon) access to an 657 * emergency reserve. 658 */ 659 lofreebuffers = nbuf / 18 + 5; 660 hifreebuffers = 2 * lofreebuffers; 661 numfreebuffers = nbuf; 662 663 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 664 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 665} 666 667/* 668 * bfreekva() - free the kva allocation for a buffer. 669 * 670 * Since this call frees up buffer space, we call bufspacewakeup(). 671 */ 672static void 673bfreekva(struct buf *bp) 674{ 675 676 if (bp->b_kvasize) { 677 atomic_add_int(&buffreekvacnt, 1); 678 atomic_subtract_long(&bufspace, bp->b_kvasize); 679 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 680 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 681 bp->b_kvasize = 0; 682 bufspacewakeup(); 683 } 684} 685 686/* 687 * bremfree: 688 * 689 * Mark the buffer for removal from the appropriate free list in brelse. 690 * 691 */ 692void 693bremfree(struct buf *bp) 694{ 695 int old; 696 697 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 698 KASSERT((bp->b_flags & B_REMFREE) == 0, 699 ("bremfree: buffer %p already marked for delayed removal.", bp)); 700 KASSERT(bp->b_qindex != QUEUE_NONE, 701 ("bremfree: buffer %p not on a queue.", bp)); 702 BUF_ASSERT_HELD(bp); 703 704 bp->b_flags |= B_REMFREE; 705 /* Fixup numfreebuffers count. */ 706 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 707 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 708 ("buf %p not counted in numfreebuffers", bp)); 709 bp->b_vflags &= ~BV_INFREECNT; 710 old = atomic_fetchadd_int(&numfreebuffers, -1); 711 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 712 } 713} 714 715/* 716 * bremfreef: 717 * 718 * Force an immediate removal from a free list. Used only in nfs when 719 * it abuses the b_freelist pointer. 720 */ 721void 722bremfreef(struct buf *bp) 723{ 724 mtx_lock(&bqlock); 725 bremfreel(bp); 726 mtx_unlock(&bqlock); 727} 728 729/* 730 * bremfreel: 731 * 732 * Removes a buffer from the free list, must be called with the 733 * bqlock held. 734 */ 735static void 736bremfreel(struct buf *bp) 737{ 738 int old; 739 740 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 741 bp, bp->b_vp, bp->b_flags); 742 KASSERT(bp->b_qindex != QUEUE_NONE, 743 ("bremfreel: buffer %p not on a queue.", bp)); 744 BUF_ASSERT_HELD(bp); 745 mtx_assert(&bqlock, MA_OWNED); 746 747 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 748 bp->b_qindex = QUEUE_NONE; 749 /* 750 * If this was a delayed bremfree() we only need to remove the buffer 751 * from the queue and return the stats are already done. 752 */ 753 if (bp->b_flags & B_REMFREE) { 754 bp->b_flags &= ~B_REMFREE; 755 return; 756 } 757 /* 758 * Fixup numfreebuffers count. If the buffer is invalid or not 759 * delayed-write, the buffer was free and we must decrement 760 * numfreebuffers. 761 */ 762 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 763 KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 764 ("buf %p not counted in numfreebuffers", bp)); 765 bp->b_vflags &= ~BV_INFREECNT; 766 old = atomic_fetchadd_int(&numfreebuffers, -1); 767 KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 768 } 769} 770 771 772/* 773 * Get a buffer with the specified data. Look in the cache first. We 774 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 775 * is set, the buffer is valid and we do not have to do anything ( see 776 * getblk() ). This is really just a special case of breadn(). 777 */ 778int 779bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 780 struct buf **bpp) 781{ 782 783 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 784} 785 786/* 787 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 788 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 789 * the buffer is valid and we do not have to do anything. 790 */ 791void 792breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 793 int cnt, struct ucred * cred) 794{ 795 struct buf *rabp; 796 int i; 797 798 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 799 if (inmem(vp, *rablkno)) 800 continue; 801 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 802 803 if ((rabp->b_flags & B_CACHE) == 0) { 804 if (!TD_IS_IDLETHREAD(curthread)) 805 curthread->td_ru.ru_inblock++; 806 rabp->b_flags |= B_ASYNC; 807 rabp->b_flags &= ~B_INVAL; 808 rabp->b_ioflags &= ~BIO_ERROR; 809 rabp->b_iocmd = BIO_READ; 810 if (rabp->b_rcred == NOCRED && cred != NOCRED) 811 rabp->b_rcred = crhold(cred); 812 vfs_busy_pages(rabp, 0); 813 BUF_KERNPROC(rabp); 814 rabp->b_iooffset = dbtob(rabp->b_blkno); 815 bstrategy(rabp); 816 } else { 817 brelse(rabp); 818 } 819 } 820} 821 822/* 823 * Operates like bread, but also starts asynchronous I/O on 824 * read-ahead blocks. 825 */ 826int 827breadn(struct vnode * vp, daddr_t blkno, int size, 828 daddr_t * rablkno, int *rabsize, 829 int cnt, struct ucred * cred, struct buf **bpp) 830{ 831 struct buf *bp; 832 int rv = 0, readwait = 0; 833 834 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 835 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 836 837 /* if not found in cache, do some I/O */ 838 if ((bp->b_flags & B_CACHE) == 0) { 839 if (!TD_IS_IDLETHREAD(curthread)) 840 curthread->td_ru.ru_inblock++; 841 bp->b_iocmd = BIO_READ; 842 bp->b_flags &= ~B_INVAL; 843 bp->b_ioflags &= ~BIO_ERROR; 844 if (bp->b_rcred == NOCRED && cred != NOCRED) 845 bp->b_rcred = crhold(cred); 846 vfs_busy_pages(bp, 0); 847 bp->b_iooffset = dbtob(bp->b_blkno); 848 bstrategy(bp); 849 ++readwait; 850 } 851 852 breada(vp, rablkno, rabsize, cnt, cred); 853 854 if (readwait) { 855 rv = bufwait(bp); 856 } 857 return (rv); 858} 859 860/* 861 * Write, release buffer on completion. (Done by iodone 862 * if async). Do not bother writing anything if the buffer 863 * is invalid. 864 * 865 * Note that we set B_CACHE here, indicating that buffer is 866 * fully valid and thus cacheable. This is true even of NFS 867 * now so we set it generally. This could be set either here 868 * or in biodone() since the I/O is synchronous. We put it 869 * here. 870 */ 871int 872bufwrite(struct buf *bp) 873{ 874 int oldflags; 875 struct vnode *vp; 876 int vp_md; 877 878 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 879 if (bp->b_flags & B_INVAL) { 880 brelse(bp); 881 return (0); 882 } 883 884 oldflags = bp->b_flags; 885 886 BUF_ASSERT_HELD(bp); 887 888 if (bp->b_pin_count > 0) 889 bunpin_wait(bp); 890 891 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 892 ("FFS background buffer should not get here %p", bp)); 893 894 vp = bp->b_vp; 895 if (vp) 896 vp_md = vp->v_vflag & VV_MD; 897 else 898 vp_md = 0; 899 900 /* Mark the buffer clean */ 901 bundirty(bp); 902 903 bp->b_flags &= ~B_DONE; 904 bp->b_ioflags &= ~BIO_ERROR; 905 bp->b_flags |= B_CACHE; 906 bp->b_iocmd = BIO_WRITE; 907 908 bufobj_wref(bp->b_bufobj); 909 vfs_busy_pages(bp, 1); 910 911 /* 912 * Normal bwrites pipeline writes 913 */ 914 bp->b_runningbufspace = bp->b_bufsize; 915 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 916 917 if (!TD_IS_IDLETHREAD(curthread)) 918 curthread->td_ru.ru_oublock++; 919 if (oldflags & B_ASYNC) 920 BUF_KERNPROC(bp); 921 bp->b_iooffset = dbtob(bp->b_blkno); 922 bstrategy(bp); 923 924 if ((oldflags & B_ASYNC) == 0) { 925 int rtval = bufwait(bp); 926 brelse(bp); 927 return (rtval); 928 } else { 929 /* 930 * don't allow the async write to saturate the I/O 931 * system. We will not deadlock here because 932 * we are blocking waiting for I/O that is already in-progress 933 * to complete. We do not block here if it is the update 934 * or syncer daemon trying to clean up as that can lead 935 * to deadlock. 936 */ 937 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 938 waitrunningbufspace(); 939 } 940 941 return (0); 942} 943 944void 945bufbdflush(struct bufobj *bo, struct buf *bp) 946{ 947 struct buf *nbp; 948 949 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 950 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 951 altbufferflushes++; 952 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 953 BO_LOCK(bo); 954 /* 955 * Try to find a buffer to flush. 956 */ 957 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 958 if ((nbp->b_vflags & BV_BKGRDINPROG) || 959 BUF_LOCK(nbp, 960 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 961 continue; 962 if (bp == nbp) 963 panic("bdwrite: found ourselves"); 964 BO_UNLOCK(bo); 965 /* Don't countdeps with the bo lock held. */ 966 if (buf_countdeps(nbp, 0)) { 967 BO_LOCK(bo); 968 BUF_UNLOCK(nbp); 969 continue; 970 } 971 if (nbp->b_flags & B_CLUSTEROK) { 972 vfs_bio_awrite(nbp); 973 } else { 974 bremfree(nbp); 975 bawrite(nbp); 976 } 977 dirtybufferflushes++; 978 break; 979 } 980 if (nbp == NULL) 981 BO_UNLOCK(bo); 982 } 983} 984 985/* 986 * Delayed write. (Buffer is marked dirty). Do not bother writing 987 * anything if the buffer is marked invalid. 988 * 989 * Note that since the buffer must be completely valid, we can safely 990 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 991 * biodone() in order to prevent getblk from writing the buffer 992 * out synchronously. 993 */ 994void 995bdwrite(struct buf *bp) 996{ 997 struct thread *td = curthread; 998 struct vnode *vp; 999 struct bufobj *bo; 1000 1001 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1002 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1003 BUF_ASSERT_HELD(bp); 1004 1005 if (bp->b_flags & B_INVAL) { 1006 brelse(bp); 1007 return; 1008 } 1009 1010 /* 1011 * If we have too many dirty buffers, don't create any more. 1012 * If we are wildly over our limit, then force a complete 1013 * cleanup. Otherwise, just keep the situation from getting 1014 * out of control. Note that we have to avoid a recursive 1015 * disaster and not try to clean up after our own cleanup! 1016 */ 1017 vp = bp->b_vp; 1018 bo = bp->b_bufobj; 1019 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1020 td->td_pflags |= TDP_INBDFLUSH; 1021 BO_BDFLUSH(bo, bp); 1022 td->td_pflags &= ~TDP_INBDFLUSH; 1023 } else 1024 recursiveflushes++; 1025 1026 bdirty(bp); 1027 /* 1028 * Set B_CACHE, indicating that the buffer is fully valid. This is 1029 * true even of NFS now. 1030 */ 1031 bp->b_flags |= B_CACHE; 1032 1033 /* 1034 * This bmap keeps the system from needing to do the bmap later, 1035 * perhaps when the system is attempting to do a sync. Since it 1036 * is likely that the indirect block -- or whatever other datastructure 1037 * that the filesystem needs is still in memory now, it is a good 1038 * thing to do this. Note also, that if the pageout daemon is 1039 * requesting a sync -- there might not be enough memory to do 1040 * the bmap then... So, this is important to do. 1041 */ 1042 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1043 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1044 } 1045 1046 /* 1047 * Set the *dirty* buffer range based upon the VM system dirty 1048 * pages. 1049 * 1050 * Mark the buffer pages as clean. We need to do this here to 1051 * satisfy the vnode_pager and the pageout daemon, so that it 1052 * thinks that the pages have been "cleaned". Note that since 1053 * the pages are in a delayed write buffer -- the VFS layer 1054 * "will" see that the pages get written out on the next sync, 1055 * or perhaps the cluster will be completed. 1056 */ 1057 vfs_clean_pages_dirty_buf(bp); 1058 bqrelse(bp); 1059 1060 /* 1061 * Wakeup the buffer flushing daemon if we have a lot of dirty 1062 * buffers (midpoint between our recovery point and our stall 1063 * point). 1064 */ 1065 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1066 1067 /* 1068 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1069 * due to the softdep code. 1070 */ 1071} 1072 1073/* 1074 * bdirty: 1075 * 1076 * Turn buffer into delayed write request. We must clear BIO_READ and 1077 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1078 * itself to properly update it in the dirty/clean lists. We mark it 1079 * B_DONE to ensure that any asynchronization of the buffer properly 1080 * clears B_DONE ( else a panic will occur later ). 1081 * 1082 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1083 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1084 * should only be called if the buffer is known-good. 1085 * 1086 * Since the buffer is not on a queue, we do not update the numfreebuffers 1087 * count. 1088 * 1089 * The buffer must be on QUEUE_NONE. 1090 */ 1091void 1092bdirty(struct buf *bp) 1093{ 1094 1095 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1096 bp, bp->b_vp, bp->b_flags); 1097 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1098 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1099 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1100 BUF_ASSERT_HELD(bp); 1101 bp->b_flags &= ~(B_RELBUF); 1102 bp->b_iocmd = BIO_WRITE; 1103 1104 if ((bp->b_flags & B_DELWRI) == 0) { 1105 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1106 reassignbuf(bp); 1107 atomic_add_int(&numdirtybuffers, 1); 1108 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1109 } 1110} 1111 1112/* 1113 * bundirty: 1114 * 1115 * Clear B_DELWRI for buffer. 1116 * 1117 * Since the buffer is not on a queue, we do not update the numfreebuffers 1118 * count. 1119 * 1120 * The buffer must be on QUEUE_NONE. 1121 */ 1122 1123void 1124bundirty(struct buf *bp) 1125{ 1126 1127 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1128 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1129 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1130 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1131 BUF_ASSERT_HELD(bp); 1132 1133 if (bp->b_flags & B_DELWRI) { 1134 bp->b_flags &= ~B_DELWRI; 1135 reassignbuf(bp); 1136 atomic_subtract_int(&numdirtybuffers, 1); 1137 numdirtywakeup(lodirtybuffers); 1138 } 1139 /* 1140 * Since it is now being written, we can clear its deferred write flag. 1141 */ 1142 bp->b_flags &= ~B_DEFERRED; 1143} 1144 1145/* 1146 * bawrite: 1147 * 1148 * Asynchronous write. Start output on a buffer, but do not wait for 1149 * it to complete. The buffer is released when the output completes. 1150 * 1151 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1152 * B_INVAL buffers. Not us. 1153 */ 1154void 1155bawrite(struct buf *bp) 1156{ 1157 1158 bp->b_flags |= B_ASYNC; 1159 (void) bwrite(bp); 1160} 1161 1162/* 1163 * bwillwrite: 1164 * 1165 * Called prior to the locking of any vnodes when we are expecting to 1166 * write. We do not want to starve the buffer cache with too many 1167 * dirty buffers so we block here. By blocking prior to the locking 1168 * of any vnodes we attempt to avoid the situation where a locked vnode 1169 * prevents the various system daemons from flushing related buffers. 1170 */ 1171 1172void 1173bwillwrite(void) 1174{ 1175 1176 if (numdirtybuffers >= hidirtybuffers) { 1177 mtx_lock(&nblock); 1178 while (numdirtybuffers >= hidirtybuffers) { 1179 bd_wakeup(1); 1180 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1181 msleep(&needsbuffer, &nblock, 1182 (PRIBIO + 4), "flswai", 0); 1183 } 1184 mtx_unlock(&nblock); 1185 } 1186} 1187 1188/* 1189 * Return true if we have too many dirty buffers. 1190 */ 1191int 1192buf_dirty_count_severe(void) 1193{ 1194 1195 return(numdirtybuffers >= hidirtybuffers); 1196} 1197 1198static __noinline int 1199buf_vm_page_count_severe(void) 1200{ 1201 1202 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1203 1204 return vm_page_count_severe(); 1205} 1206 1207/* 1208 * brelse: 1209 * 1210 * Release a busy buffer and, if requested, free its resources. The 1211 * buffer will be stashed in the appropriate bufqueue[] allowing it 1212 * to be accessed later as a cache entity or reused for other purposes. 1213 */ 1214void 1215brelse(struct buf *bp) 1216{ 1217 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1218 bp, bp->b_vp, bp->b_flags); 1219 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1220 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1221 1222 if (bp->b_flags & B_MANAGED) { 1223 bqrelse(bp); 1224 return; 1225 } 1226 1227 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1228 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1229 /* 1230 * Failed write, redirty. Must clear BIO_ERROR to prevent 1231 * pages from being scrapped. If the error is anything 1232 * other than an I/O error (EIO), assume that retrying 1233 * is futile. 1234 */ 1235 bp->b_ioflags &= ~BIO_ERROR; 1236 bdirty(bp); 1237 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1238 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1239 /* 1240 * Either a failed I/O or we were asked to free or not 1241 * cache the buffer. 1242 */ 1243 bp->b_flags |= B_INVAL; 1244 if (!LIST_EMPTY(&bp->b_dep)) 1245 buf_deallocate(bp); 1246 if (bp->b_flags & B_DELWRI) { 1247 atomic_subtract_int(&numdirtybuffers, 1); 1248 numdirtywakeup(lodirtybuffers); 1249 } 1250 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1251 if ((bp->b_flags & B_VMIO) == 0) { 1252 if (bp->b_bufsize) 1253 allocbuf(bp, 0); 1254 if (bp->b_vp) 1255 brelvp(bp); 1256 } 1257 } 1258 1259 /* 1260 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1261 * is called with B_DELWRI set, the underlying pages may wind up 1262 * getting freed causing a previous write (bdwrite()) to get 'lost' 1263 * because pages associated with a B_DELWRI bp are marked clean. 1264 * 1265 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1266 * if B_DELWRI is set. 1267 * 1268 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1269 * on pages to return pages to the VM page queues. 1270 */ 1271 if (bp->b_flags & B_DELWRI) 1272 bp->b_flags &= ~B_RELBUF; 1273 else if (buf_vm_page_count_severe()) { 1274 /* 1275 * The locking of the BO_LOCK is not necessary since 1276 * BKGRDINPROG cannot be set while we hold the buf 1277 * lock, it can only be cleared if it is already 1278 * pending. 1279 */ 1280 if (bp->b_vp) { 1281 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1282 bp->b_flags |= B_RELBUF; 1283 } else 1284 bp->b_flags |= B_RELBUF; 1285 } 1286 1287 /* 1288 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1289 * constituted, not even NFS buffers now. Two flags effect this. If 1290 * B_INVAL, the struct buf is invalidated but the VM object is kept 1291 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1292 * 1293 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1294 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1295 * buffer is also B_INVAL because it hits the re-dirtying code above. 1296 * 1297 * Normally we can do this whether a buffer is B_DELWRI or not. If 1298 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1299 * the commit state and we cannot afford to lose the buffer. If the 1300 * buffer has a background write in progress, we need to keep it 1301 * around to prevent it from being reconstituted and starting a second 1302 * background write. 1303 */ 1304 if ((bp->b_flags & B_VMIO) 1305 && !(bp->b_vp->v_mount != NULL && 1306 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1307 !vn_isdisk(bp->b_vp, NULL) && 1308 (bp->b_flags & B_DELWRI)) 1309 ) { 1310 1311 int i, j, resid; 1312 vm_page_t m; 1313 off_t foff; 1314 vm_pindex_t poff; 1315 vm_object_t obj; 1316 1317 obj = bp->b_bufobj->bo_object; 1318 1319 /* 1320 * Get the base offset and length of the buffer. Note that 1321 * in the VMIO case if the buffer block size is not 1322 * page-aligned then b_data pointer may not be page-aligned. 1323 * But our b_pages[] array *IS* page aligned. 1324 * 1325 * block sizes less then DEV_BSIZE (usually 512) are not 1326 * supported due to the page granularity bits (m->valid, 1327 * m->dirty, etc...). 1328 * 1329 * See man buf(9) for more information 1330 */ 1331 resid = bp->b_bufsize; 1332 foff = bp->b_offset; 1333 VM_OBJECT_LOCK(obj); 1334 for (i = 0; i < bp->b_npages; i++) { 1335 int had_bogus = 0; 1336 1337 m = bp->b_pages[i]; 1338 1339 /* 1340 * If we hit a bogus page, fixup *all* the bogus pages 1341 * now. 1342 */ 1343 if (m == bogus_page) { 1344 poff = OFF_TO_IDX(bp->b_offset); 1345 had_bogus = 1; 1346 1347 for (j = i; j < bp->b_npages; j++) { 1348 vm_page_t mtmp; 1349 mtmp = bp->b_pages[j]; 1350 if (mtmp == bogus_page) { 1351 mtmp = vm_page_lookup(obj, poff + j); 1352 if (!mtmp) { 1353 panic("brelse: page missing\n"); 1354 } 1355 bp->b_pages[j] = mtmp; 1356 } 1357 } 1358 1359 if ((bp->b_flags & B_INVAL) == 0) { 1360 pmap_qenter( 1361 trunc_page((vm_offset_t)bp->b_data), 1362 bp->b_pages, bp->b_npages); 1363 } 1364 m = bp->b_pages[i]; 1365 } 1366 if ((bp->b_flags & B_NOCACHE) || 1367 (bp->b_ioflags & BIO_ERROR && 1368 bp->b_iocmd == BIO_READ)) { 1369 int poffset = foff & PAGE_MASK; 1370 int presid = resid > (PAGE_SIZE - poffset) ? 1371 (PAGE_SIZE - poffset) : resid; 1372 1373 KASSERT(presid >= 0, ("brelse: extra page")); 1374 vm_page_set_invalid(m, poffset, presid); 1375 if (had_bogus) 1376 printf("avoided corruption bug in bogus_page/brelse code\n"); 1377 } 1378 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1379 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1380 } 1381 VM_OBJECT_UNLOCK(obj); 1382 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1383 vfs_vmio_release(bp); 1384 1385 } else if (bp->b_flags & B_VMIO) { 1386 1387 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1388 vfs_vmio_release(bp); 1389 } 1390 1391 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1392 if (bp->b_bufsize != 0) 1393 allocbuf(bp, 0); 1394 if (bp->b_vp != NULL) 1395 brelvp(bp); 1396 } 1397 1398 if (BUF_LOCKRECURSED(bp)) { 1399 /* do not release to free list */ 1400 BUF_UNLOCK(bp); 1401 return; 1402 } 1403 1404 /* enqueue */ 1405 mtx_lock(&bqlock); 1406 /* Handle delayed bremfree() processing. */ 1407 if (bp->b_flags & B_REMFREE) 1408 bremfreel(bp); 1409 if (bp->b_qindex != QUEUE_NONE) 1410 panic("brelse: free buffer onto another queue???"); 1411 1412 /* 1413 * If the buffer has junk contents signal it and eventually 1414 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1415 * doesn't find it. 1416 */ 1417 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1418 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1419 bp->b_flags |= B_INVAL; 1420 if (bp->b_flags & B_INVAL) { 1421 if (bp->b_flags & B_DELWRI) 1422 bundirty(bp); 1423 if (bp->b_vp) 1424 brelvp(bp); 1425 } 1426 1427 /* buffers with no memory */ 1428 if (bp->b_bufsize == 0) { 1429 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1430 if (bp->b_vflags & BV_BKGRDINPROG) 1431 panic("losing buffer 1"); 1432 if (bp->b_kvasize) { 1433 bp->b_qindex = QUEUE_EMPTYKVA; 1434 } else { 1435 bp->b_qindex = QUEUE_EMPTY; 1436 } 1437 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1438 /* buffers with junk contents */ 1439 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1440 (bp->b_ioflags & BIO_ERROR)) { 1441 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1442 if (bp->b_vflags & BV_BKGRDINPROG) 1443 panic("losing buffer 2"); 1444 bp->b_qindex = QUEUE_CLEAN; 1445 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1446 /* remaining buffers */ 1447 } else { 1448 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1449 (B_DELWRI|B_NEEDSGIANT)) 1450 bp->b_qindex = QUEUE_DIRTY_GIANT; 1451 else if (bp->b_flags & B_DELWRI) 1452 bp->b_qindex = QUEUE_DIRTY; 1453 else 1454 bp->b_qindex = QUEUE_CLEAN; 1455 if (bp->b_flags & B_AGE) 1456 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1457 else 1458 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1459 } 1460 mtx_unlock(&bqlock); 1461 1462 /* 1463 * Fixup numfreebuffers count. The bp is on an appropriate queue 1464 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1465 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1466 * if B_INVAL is set ). 1467 */ 1468 1469 if (!(bp->b_flags & B_DELWRI)) 1470 bufcountwakeup(bp); 1471 1472 /* 1473 * Something we can maybe free or reuse 1474 */ 1475 if (bp->b_bufsize || bp->b_kvasize) 1476 bufspacewakeup(); 1477 1478 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1479 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1480 panic("brelse: not dirty"); 1481 /* unlock */ 1482 BUF_UNLOCK(bp); 1483} 1484 1485/* 1486 * Release a buffer back to the appropriate queue but do not try to free 1487 * it. The buffer is expected to be used again soon. 1488 * 1489 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1490 * biodone() to requeue an async I/O on completion. It is also used when 1491 * known good buffers need to be requeued but we think we may need the data 1492 * again soon. 1493 * 1494 * XXX we should be able to leave the B_RELBUF hint set on completion. 1495 */ 1496void 1497bqrelse(struct buf *bp) 1498{ 1499 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1500 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1501 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1502 1503 if (BUF_LOCKRECURSED(bp)) { 1504 /* do not release to free list */ 1505 BUF_UNLOCK(bp); 1506 return; 1507 } 1508 1509 if (bp->b_flags & B_MANAGED) { 1510 if (bp->b_flags & B_REMFREE) { 1511 mtx_lock(&bqlock); 1512 bremfreel(bp); 1513 mtx_unlock(&bqlock); 1514 } 1515 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1516 BUF_UNLOCK(bp); 1517 return; 1518 } 1519 1520 mtx_lock(&bqlock); 1521 /* Handle delayed bremfree() processing. */ 1522 if (bp->b_flags & B_REMFREE) 1523 bremfreel(bp); 1524 if (bp->b_qindex != QUEUE_NONE) 1525 panic("bqrelse: free buffer onto another queue???"); 1526 /* buffers with stale but valid contents */ 1527 if (bp->b_flags & B_DELWRI) { 1528 if (bp->b_flags & B_NEEDSGIANT) 1529 bp->b_qindex = QUEUE_DIRTY_GIANT; 1530 else 1531 bp->b_qindex = QUEUE_DIRTY; 1532 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1533 } else { 1534 /* 1535 * The locking of the BO_LOCK for checking of the 1536 * BV_BKGRDINPROG is not necessary since the 1537 * BV_BKGRDINPROG cannot be set while we hold the buf 1538 * lock, it can only be cleared if it is already 1539 * pending. 1540 */ 1541 if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1542 bp->b_qindex = QUEUE_CLEAN; 1543 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1544 b_freelist); 1545 } else { 1546 /* 1547 * We are too low on memory, we have to try to free 1548 * the buffer (most importantly: the wired pages 1549 * making up its backing store) *now*. 1550 */ 1551 mtx_unlock(&bqlock); 1552 brelse(bp); 1553 return; 1554 } 1555 } 1556 mtx_unlock(&bqlock); 1557 1558 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1559 bufcountwakeup(bp); 1560 1561 /* 1562 * Something we can maybe free or reuse. 1563 */ 1564 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1565 bufspacewakeup(); 1566 1567 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1568 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1569 panic("bqrelse: not dirty"); 1570 /* unlock */ 1571 BUF_UNLOCK(bp); 1572} 1573 1574/* Give pages used by the bp back to the VM system (where possible) */ 1575static void 1576vfs_vmio_release(struct buf *bp) 1577{ 1578 int i; 1579 vm_page_t m; 1580 1581 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1582 for (i = 0; i < bp->b_npages; i++) { 1583 m = bp->b_pages[i]; 1584 bp->b_pages[i] = NULL; 1585 /* 1586 * In order to keep page LRU ordering consistent, put 1587 * everything on the inactive queue. 1588 */ 1589 vm_page_lock(m); 1590 vm_page_unwire(m, 0); 1591 /* 1592 * We don't mess with busy pages, it is 1593 * the responsibility of the process that 1594 * busied the pages to deal with them. 1595 */ 1596 if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 && 1597 m->wire_count == 0) { 1598 /* 1599 * Might as well free the page if we can and it has 1600 * no valid data. We also free the page if the 1601 * buffer was used for direct I/O 1602 */ 1603 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1604 m->hold_count == 0) { 1605 vm_page_free(m); 1606 } else if (bp->b_flags & B_DIRECT) { 1607 vm_page_try_to_free(m); 1608 } else if (buf_vm_page_count_severe()) { 1609 vm_page_try_to_cache(m); 1610 } 1611 } 1612 vm_page_unlock(m); 1613 } 1614 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1615 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1616 1617 if (bp->b_bufsize) { 1618 bufspacewakeup(); 1619 bp->b_bufsize = 0; 1620 } 1621 bp->b_npages = 0; 1622 bp->b_flags &= ~B_VMIO; 1623 if (bp->b_vp) 1624 brelvp(bp); 1625} 1626 1627/* 1628 * Check to see if a block at a particular lbn is available for a clustered 1629 * write. 1630 */ 1631static int 1632vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1633{ 1634 struct buf *bpa; 1635 int match; 1636 1637 match = 0; 1638 1639 /* If the buf isn't in core skip it */ 1640 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1641 return (0); 1642 1643 /* If the buf is busy we don't want to wait for it */ 1644 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1645 return (0); 1646 1647 /* Only cluster with valid clusterable delayed write buffers */ 1648 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1649 (B_DELWRI | B_CLUSTEROK)) 1650 goto done; 1651 1652 if (bpa->b_bufsize != size) 1653 goto done; 1654 1655 /* 1656 * Check to see if it is in the expected place on disk and that the 1657 * block has been mapped. 1658 */ 1659 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1660 match = 1; 1661done: 1662 BUF_UNLOCK(bpa); 1663 return (match); 1664} 1665 1666/* 1667 * vfs_bio_awrite: 1668 * 1669 * Implement clustered async writes for clearing out B_DELWRI buffers. 1670 * This is much better then the old way of writing only one buffer at 1671 * a time. Note that we may not be presented with the buffers in the 1672 * correct order, so we search for the cluster in both directions. 1673 */ 1674int 1675vfs_bio_awrite(struct buf *bp) 1676{ 1677 struct bufobj *bo; 1678 int i; 1679 int j; 1680 daddr_t lblkno = bp->b_lblkno; 1681 struct vnode *vp = bp->b_vp; 1682 int ncl; 1683 int nwritten; 1684 int size; 1685 int maxcl; 1686 1687 bo = &vp->v_bufobj; 1688 /* 1689 * right now we support clustered writing only to regular files. If 1690 * we find a clusterable block we could be in the middle of a cluster 1691 * rather then at the beginning. 1692 */ 1693 if ((vp->v_type == VREG) && 1694 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1695 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1696 1697 size = vp->v_mount->mnt_stat.f_iosize; 1698 maxcl = MAXPHYS / size; 1699 1700 BO_LOCK(bo); 1701 for (i = 1; i < maxcl; i++) 1702 if (vfs_bio_clcheck(vp, size, lblkno + i, 1703 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1704 break; 1705 1706 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1707 if (vfs_bio_clcheck(vp, size, lblkno - j, 1708 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1709 break; 1710 BO_UNLOCK(bo); 1711 --j; 1712 ncl = i + j; 1713 /* 1714 * this is a possible cluster write 1715 */ 1716 if (ncl != 1) { 1717 BUF_UNLOCK(bp); 1718 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1719 return nwritten; 1720 } 1721 } 1722 bremfree(bp); 1723 bp->b_flags |= B_ASYNC; 1724 /* 1725 * default (old) behavior, writing out only one block 1726 * 1727 * XXX returns b_bufsize instead of b_bcount for nwritten? 1728 */ 1729 nwritten = bp->b_bufsize; 1730 (void) bwrite(bp); 1731 1732 return nwritten; 1733} 1734 1735/* 1736 * getnewbuf: 1737 * 1738 * Find and initialize a new buffer header, freeing up existing buffers 1739 * in the bufqueues as necessary. The new buffer is returned locked. 1740 * 1741 * Important: B_INVAL is not set. If the caller wishes to throw the 1742 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1743 * 1744 * We block if: 1745 * We have insufficient buffer headers 1746 * We have insufficient buffer space 1747 * buffer_map is too fragmented ( space reservation fails ) 1748 * If we have to flush dirty buffers ( but we try to avoid this ) 1749 * 1750 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1751 * Instead we ask the buf daemon to do it for us. We attempt to 1752 * avoid piecemeal wakeups of the pageout daemon. 1753 */ 1754 1755static struct buf * 1756getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 1757 int gbflags) 1758{ 1759 struct thread *td; 1760 struct buf *bp; 1761 struct buf *nbp; 1762 int defrag = 0; 1763 int nqindex; 1764 static int flushingbufs; 1765 1766 td = curthread; 1767 /* 1768 * We can't afford to block since we might be holding a vnode lock, 1769 * which may prevent system daemons from running. We deal with 1770 * low-memory situations by proactively returning memory and running 1771 * async I/O rather then sync I/O. 1772 */ 1773 atomic_add_int(&getnewbufcalls, 1); 1774 atomic_subtract_int(&getnewbufrestarts, 1); 1775restart: 1776 atomic_add_int(&getnewbufrestarts, 1); 1777 1778 /* 1779 * Setup for scan. If we do not have enough free buffers, 1780 * we setup a degenerate case that immediately fails. Note 1781 * that if we are specially marked process, we are allowed to 1782 * dip into our reserves. 1783 * 1784 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1785 * 1786 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1787 * However, there are a number of cases (defragging, reusing, ...) 1788 * where we cannot backup. 1789 */ 1790 mtx_lock(&bqlock); 1791 nqindex = QUEUE_EMPTYKVA; 1792 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1793 1794 if (nbp == NULL) { 1795 /* 1796 * If no EMPTYKVA buffers and we are either 1797 * defragging or reusing, locate a CLEAN buffer 1798 * to free or reuse. If bufspace useage is low 1799 * skip this step so we can allocate a new buffer. 1800 */ 1801 if (defrag || bufspace >= lobufspace) { 1802 nqindex = QUEUE_CLEAN; 1803 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1804 } 1805 1806 /* 1807 * If we could not find or were not allowed to reuse a 1808 * CLEAN buffer, check to see if it is ok to use an EMPTY 1809 * buffer. We can only use an EMPTY buffer if allocating 1810 * its KVA would not otherwise run us out of buffer space. 1811 */ 1812 if (nbp == NULL && defrag == 0 && 1813 bufspace + maxsize < hibufspace) { 1814 nqindex = QUEUE_EMPTY; 1815 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1816 } 1817 } 1818 1819 /* 1820 * Run scan, possibly freeing data and/or kva mappings on the fly 1821 * depending. 1822 */ 1823 1824 while ((bp = nbp) != NULL) { 1825 int qindex = nqindex; 1826 1827 /* 1828 * Calculate next bp ( we can only use it if we do not block 1829 * or do other fancy things ). 1830 */ 1831 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1832 switch(qindex) { 1833 case QUEUE_EMPTY: 1834 nqindex = QUEUE_EMPTYKVA; 1835 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1836 break; 1837 /* FALLTHROUGH */ 1838 case QUEUE_EMPTYKVA: 1839 nqindex = QUEUE_CLEAN; 1840 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1841 break; 1842 /* FALLTHROUGH */ 1843 case QUEUE_CLEAN: 1844 /* 1845 * nbp is NULL. 1846 */ 1847 break; 1848 } 1849 } 1850 /* 1851 * If we are defragging then we need a buffer with 1852 * b_kvasize != 0. XXX this situation should no longer 1853 * occur, if defrag is non-zero the buffer's b_kvasize 1854 * should also be non-zero at this point. XXX 1855 */ 1856 if (defrag && bp->b_kvasize == 0) { 1857 printf("Warning: defrag empty buffer %p\n", bp); 1858 continue; 1859 } 1860 1861 /* 1862 * Start freeing the bp. This is somewhat involved. nbp 1863 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1864 */ 1865 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1866 continue; 1867 if (bp->b_vp) { 1868 BO_LOCK(bp->b_bufobj); 1869 if (bp->b_vflags & BV_BKGRDINPROG) { 1870 BO_UNLOCK(bp->b_bufobj); 1871 BUF_UNLOCK(bp); 1872 continue; 1873 } 1874 BO_UNLOCK(bp->b_bufobj); 1875 } 1876 CTR6(KTR_BUF, 1877 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1878 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1879 bp->b_kvasize, bp->b_bufsize, qindex); 1880 1881 /* 1882 * Sanity Checks 1883 */ 1884 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1885 1886 /* 1887 * Note: we no longer distinguish between VMIO and non-VMIO 1888 * buffers. 1889 */ 1890 1891 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1892 1893 bremfreel(bp); 1894 mtx_unlock(&bqlock); 1895 1896 if (qindex == QUEUE_CLEAN) { 1897 if (bp->b_flags & B_VMIO) { 1898 bp->b_flags &= ~B_ASYNC; 1899 vfs_vmio_release(bp); 1900 } 1901 if (bp->b_vp) 1902 brelvp(bp); 1903 } 1904 1905 /* 1906 * NOTE: nbp is now entirely invalid. We can only restart 1907 * the scan from this point on. 1908 * 1909 * Get the rest of the buffer freed up. b_kva* is still 1910 * valid after this operation. 1911 */ 1912 1913 if (bp->b_rcred != NOCRED) { 1914 crfree(bp->b_rcred); 1915 bp->b_rcred = NOCRED; 1916 } 1917 if (bp->b_wcred != NOCRED) { 1918 crfree(bp->b_wcred); 1919 bp->b_wcred = NOCRED; 1920 } 1921 if (!LIST_EMPTY(&bp->b_dep)) 1922 buf_deallocate(bp); 1923 if (bp->b_vflags & BV_BKGRDINPROG) 1924 panic("losing buffer 3"); 1925 KASSERT(bp->b_vp == NULL, 1926 ("bp: %p still has vnode %p. qindex: %d", 1927 bp, bp->b_vp, qindex)); 1928 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1929 ("bp: %p still on a buffer list. xflags %X", 1930 bp, bp->b_xflags)); 1931 1932 if (bp->b_bufsize) 1933 allocbuf(bp, 0); 1934 1935 bp->b_flags = 0; 1936 bp->b_ioflags = 0; 1937 bp->b_xflags = 0; 1938 KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 1939 ("buf %p still counted as free?", bp)); 1940 bp->b_vflags = 0; 1941 bp->b_vp = NULL; 1942 bp->b_blkno = bp->b_lblkno = 0; 1943 bp->b_offset = NOOFFSET; 1944 bp->b_iodone = 0; 1945 bp->b_error = 0; 1946 bp->b_resid = 0; 1947 bp->b_bcount = 0; 1948 bp->b_npages = 0; 1949 bp->b_dirtyoff = bp->b_dirtyend = 0; 1950 bp->b_bufobj = NULL; 1951 bp->b_pin_count = 0; 1952 bp->b_fsprivate1 = NULL; 1953 bp->b_fsprivate2 = NULL; 1954 bp->b_fsprivate3 = NULL; 1955 1956 LIST_INIT(&bp->b_dep); 1957 1958 /* 1959 * If we are defragging then free the buffer. 1960 */ 1961 if (defrag) { 1962 bp->b_flags |= B_INVAL; 1963 bfreekva(bp); 1964 brelse(bp); 1965 defrag = 0; 1966 goto restart; 1967 } 1968 1969 /* 1970 * Notify any waiters for the buffer lock about 1971 * identity change by freeing the buffer. 1972 */ 1973 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 1974 bp->b_flags |= B_INVAL; 1975 bfreekva(bp); 1976 brelse(bp); 1977 goto restart; 1978 } 1979 1980 /* 1981 * If we are overcomitted then recover the buffer and its 1982 * KVM space. This occurs in rare situations when multiple 1983 * processes are blocked in getnewbuf() or allocbuf(). 1984 */ 1985 if (bufspace >= hibufspace) 1986 flushingbufs = 1; 1987 if (flushingbufs && bp->b_kvasize != 0) { 1988 bp->b_flags |= B_INVAL; 1989 bfreekva(bp); 1990 brelse(bp); 1991 goto restart; 1992 } 1993 if (bufspace < lobufspace) 1994 flushingbufs = 0; 1995 break; 1996 } 1997 1998 /* 1999 * If we exhausted our list, sleep as appropriate. We may have to 2000 * wakeup various daemons and write out some dirty buffers. 2001 * 2002 * Generally we are sleeping due to insufficient buffer space. 2003 */ 2004 2005 if (bp == NULL) { 2006 int flags, norunbuf; 2007 char *waitmsg; 2008 int fl; 2009 2010 if (defrag) { 2011 flags = VFS_BIO_NEED_BUFSPACE; 2012 waitmsg = "nbufkv"; 2013 } else if (bufspace >= hibufspace) { 2014 waitmsg = "nbufbs"; 2015 flags = VFS_BIO_NEED_BUFSPACE; 2016 } else { 2017 waitmsg = "newbuf"; 2018 flags = VFS_BIO_NEED_ANY; 2019 } 2020 mtx_lock(&nblock); 2021 needsbuffer |= flags; 2022 mtx_unlock(&nblock); 2023 mtx_unlock(&bqlock); 2024 2025 bd_speedup(); /* heeeelp */ 2026 if (gbflags & GB_NOWAIT_BD) 2027 return (NULL); 2028 2029 mtx_lock(&nblock); 2030 while (needsbuffer & flags) { 2031 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2032 mtx_unlock(&nblock); 2033 /* 2034 * getblk() is called with a vnode 2035 * locked, and some majority of the 2036 * dirty buffers may as well belong to 2037 * the vnode. Flushing the buffers 2038 * there would make a progress that 2039 * cannot be achieved by the 2040 * buf_daemon, that cannot lock the 2041 * vnode. 2042 */ 2043 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2044 (td->td_pflags & TDP_NORUNNINGBUF); 2045 /* play bufdaemon */ 2046 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2047 fl = buf_do_flush(vp); 2048 td->td_pflags &= norunbuf; 2049 mtx_lock(&nblock); 2050 if (fl != 0) 2051 continue; 2052 if ((needsbuffer & flags) == 0) 2053 break; 2054 } 2055 if (msleep(&needsbuffer, &nblock, 2056 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2057 mtx_unlock(&nblock); 2058 return (NULL); 2059 } 2060 } 2061 mtx_unlock(&nblock); 2062 } else { 2063 /* 2064 * We finally have a valid bp. We aren't quite out of the 2065 * woods, we still have to reserve kva space. In order 2066 * to keep fragmentation sane we only allocate kva in 2067 * BKVASIZE chunks. 2068 */ 2069 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2070 2071 if (maxsize != bp->b_kvasize) { 2072 vm_offset_t addr = 0; 2073 2074 bfreekva(bp); 2075 2076 vm_map_lock(buffer_map); 2077 if (vm_map_findspace(buffer_map, 2078 vm_map_min(buffer_map), maxsize, &addr)) { 2079 /* 2080 * Uh oh. Buffer map is to fragmented. We 2081 * must defragment the map. 2082 */ 2083 atomic_add_int(&bufdefragcnt, 1); 2084 vm_map_unlock(buffer_map); 2085 defrag = 1; 2086 bp->b_flags |= B_INVAL; 2087 brelse(bp); 2088 goto restart; 2089 } 2090 if (addr) { 2091 vm_map_insert(buffer_map, NULL, 0, 2092 addr, addr + maxsize, 2093 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2094 2095 bp->b_kvabase = (caddr_t) addr; 2096 bp->b_kvasize = maxsize; 2097 atomic_add_long(&bufspace, bp->b_kvasize); 2098 atomic_add_int(&bufreusecnt, 1); 2099 } 2100 vm_map_unlock(buffer_map); 2101 } 2102 bp->b_saveaddr = bp->b_kvabase; 2103 bp->b_data = bp->b_saveaddr; 2104 } 2105 return(bp); 2106} 2107 2108/* 2109 * buf_daemon: 2110 * 2111 * buffer flushing daemon. Buffers are normally flushed by the 2112 * update daemon but if it cannot keep up this process starts to 2113 * take the load in an attempt to prevent getnewbuf() from blocking. 2114 */ 2115 2116static struct kproc_desc buf_kp = { 2117 "bufdaemon", 2118 buf_daemon, 2119 &bufdaemonproc 2120}; 2121SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2122 2123static int 2124buf_do_flush(struct vnode *vp) 2125{ 2126 int flushed; 2127 2128 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2129 /* The list empty check here is slightly racy */ 2130 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2131 mtx_lock(&Giant); 2132 flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0); 2133 mtx_unlock(&Giant); 2134 } 2135 if (flushed == 0) { 2136 /* 2137 * Could not find any buffers without rollback 2138 * dependencies, so just write the first one 2139 * in the hopes of eventually making progress. 2140 */ 2141 flushbufqueues(vp, QUEUE_DIRTY, 1); 2142 if (!TAILQ_EMPTY( 2143 &bufqueues[QUEUE_DIRTY_GIANT])) { 2144 mtx_lock(&Giant); 2145 flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1); 2146 mtx_unlock(&Giant); 2147 } 2148 } 2149 return (flushed); 2150} 2151 2152static void 2153buf_daemon() 2154{ 2155 int lodirtysave; 2156 2157 /* 2158 * This process needs to be suspended prior to shutdown sync. 2159 */ 2160 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2161 SHUTDOWN_PRI_LAST); 2162 2163 /* 2164 * This process is allowed to take the buffer cache to the limit 2165 */ 2166 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2167 mtx_lock(&bdlock); 2168 for (;;) { 2169 bd_request = 0; 2170 mtx_unlock(&bdlock); 2171 2172 kproc_suspend_check(bufdaemonproc); 2173 lodirtysave = lodirtybuffers; 2174 if (bd_speedupreq) { 2175 lodirtybuffers = numdirtybuffers / 2; 2176 bd_speedupreq = 0; 2177 } 2178 /* 2179 * Do the flush. Limit the amount of in-transit I/O we 2180 * allow to build up, otherwise we would completely saturate 2181 * the I/O system. Wakeup any waiting processes before we 2182 * normally would so they can run in parallel with our drain. 2183 */ 2184 while (numdirtybuffers > lodirtybuffers) { 2185 if (buf_do_flush(NULL) == 0) 2186 break; 2187 uio_yield(); 2188 } 2189 lodirtybuffers = lodirtysave; 2190 2191 /* 2192 * Only clear bd_request if we have reached our low water 2193 * mark. The buf_daemon normally waits 1 second and 2194 * then incrementally flushes any dirty buffers that have 2195 * built up, within reason. 2196 * 2197 * If we were unable to hit our low water mark and couldn't 2198 * find any flushable buffers, we sleep half a second. 2199 * Otherwise we loop immediately. 2200 */ 2201 mtx_lock(&bdlock); 2202 if (numdirtybuffers <= lodirtybuffers) { 2203 /* 2204 * We reached our low water mark, reset the 2205 * request and sleep until we are needed again. 2206 * The sleep is just so the suspend code works. 2207 */ 2208 bd_request = 0; 2209 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2210 } else { 2211 /* 2212 * We couldn't find any flushable dirty buffers but 2213 * still have too many dirty buffers, we 2214 * have to sleep and try again. (rare) 2215 */ 2216 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2217 } 2218 } 2219} 2220 2221/* 2222 * flushbufqueues: 2223 * 2224 * Try to flush a buffer in the dirty queue. We must be careful to 2225 * free up B_INVAL buffers instead of write them, which NFS is 2226 * particularly sensitive to. 2227 */ 2228static int flushwithdeps = 0; 2229SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2230 0, "Number of buffers flushed with dependecies that require rollbacks"); 2231 2232static int 2233flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2234{ 2235 struct buf *sentinel; 2236 struct vnode *vp; 2237 struct mount *mp; 2238 struct buf *bp; 2239 int hasdeps; 2240 int flushed; 2241 int target; 2242 2243 if (lvp == NULL) { 2244 target = numdirtybuffers - lodirtybuffers; 2245 if (flushdeps && target > 2) 2246 target /= 2; 2247 } else 2248 target = flushbufqtarget; 2249 flushed = 0; 2250 bp = NULL; 2251 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2252 sentinel->b_qindex = QUEUE_SENTINEL; 2253 mtx_lock(&bqlock); 2254 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2255 while (flushed != target) { 2256 bp = TAILQ_NEXT(sentinel, b_freelist); 2257 if (bp != NULL) { 2258 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2259 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2260 b_freelist); 2261 } else 2262 break; 2263 /* 2264 * Skip sentinels inserted by other invocations of the 2265 * flushbufqueues(), taking care to not reorder them. 2266 */ 2267 if (bp->b_qindex == QUEUE_SENTINEL) 2268 continue; 2269 /* 2270 * Only flush the buffers that belong to the 2271 * vnode locked by the curthread. 2272 */ 2273 if (lvp != NULL && bp->b_vp != lvp) 2274 continue; 2275 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2276 continue; 2277 if (bp->b_pin_count > 0) { 2278 BUF_UNLOCK(bp); 2279 continue; 2280 } 2281 BO_LOCK(bp->b_bufobj); 2282 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2283 (bp->b_flags & B_DELWRI) == 0) { 2284 BO_UNLOCK(bp->b_bufobj); 2285 BUF_UNLOCK(bp); 2286 continue; 2287 } 2288 BO_UNLOCK(bp->b_bufobj); 2289 if (bp->b_flags & B_INVAL) { 2290 bremfreel(bp); 2291 mtx_unlock(&bqlock); 2292 brelse(bp); 2293 flushed++; 2294 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2295 mtx_lock(&bqlock); 2296 continue; 2297 } 2298 2299 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2300 if (flushdeps == 0) { 2301 BUF_UNLOCK(bp); 2302 continue; 2303 } 2304 hasdeps = 1; 2305 } else 2306 hasdeps = 0; 2307 /* 2308 * We must hold the lock on a vnode before writing 2309 * one of its buffers. Otherwise we may confuse, or 2310 * in the case of a snapshot vnode, deadlock the 2311 * system. 2312 * 2313 * The lock order here is the reverse of the normal 2314 * of vnode followed by buf lock. This is ok because 2315 * the NOWAIT will prevent deadlock. 2316 */ 2317 vp = bp->b_vp; 2318 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2319 BUF_UNLOCK(bp); 2320 continue; 2321 } 2322 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2323 mtx_unlock(&bqlock); 2324 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2325 bp, bp->b_vp, bp->b_flags); 2326 if (curproc == bufdaemonproc) 2327 vfs_bio_awrite(bp); 2328 else { 2329 bremfree(bp); 2330 bwrite(bp); 2331 notbufdflashes++; 2332 } 2333 vn_finished_write(mp); 2334 VOP_UNLOCK(vp, 0); 2335 flushwithdeps += hasdeps; 2336 flushed++; 2337 2338 /* 2339 * Sleeping on runningbufspace while holding 2340 * vnode lock leads to deadlock. 2341 */ 2342 if (curproc == bufdaemonproc) 2343 waitrunningbufspace(); 2344 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2345 mtx_lock(&bqlock); 2346 continue; 2347 } 2348 vn_finished_write(mp); 2349 BUF_UNLOCK(bp); 2350 } 2351 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2352 mtx_unlock(&bqlock); 2353 free(sentinel, M_TEMP); 2354 return (flushed); 2355} 2356 2357/* 2358 * Check to see if a block is currently memory resident. 2359 */ 2360struct buf * 2361incore(struct bufobj *bo, daddr_t blkno) 2362{ 2363 struct buf *bp; 2364 2365 BO_LOCK(bo); 2366 bp = gbincore(bo, blkno); 2367 BO_UNLOCK(bo); 2368 return (bp); 2369} 2370 2371/* 2372 * Returns true if no I/O is needed to access the 2373 * associated VM object. This is like incore except 2374 * it also hunts around in the VM system for the data. 2375 */ 2376 2377static int 2378inmem(struct vnode * vp, daddr_t blkno) 2379{ 2380 vm_object_t obj; 2381 vm_offset_t toff, tinc, size; 2382 vm_page_t m; 2383 vm_ooffset_t off; 2384 2385 ASSERT_VOP_LOCKED(vp, "inmem"); 2386 2387 if (incore(&vp->v_bufobj, blkno)) 2388 return 1; 2389 if (vp->v_mount == NULL) 2390 return 0; 2391 obj = vp->v_object; 2392 if (obj == NULL) 2393 return (0); 2394 2395 size = PAGE_SIZE; 2396 if (size > vp->v_mount->mnt_stat.f_iosize) 2397 size = vp->v_mount->mnt_stat.f_iosize; 2398 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2399 2400 VM_OBJECT_LOCK(obj); 2401 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2402 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2403 if (!m) 2404 goto notinmem; 2405 tinc = size; 2406 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2407 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2408 if (vm_page_is_valid(m, 2409 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2410 goto notinmem; 2411 } 2412 VM_OBJECT_UNLOCK(obj); 2413 return 1; 2414 2415notinmem: 2416 VM_OBJECT_UNLOCK(obj); 2417 return (0); 2418} 2419 2420/* 2421 * Set the dirty range for a buffer based on the status of the dirty 2422 * bits in the pages comprising the buffer. The range is limited 2423 * to the size of the buffer. 2424 * 2425 * Tell the VM system that the pages associated with this buffer 2426 * are clean. This is used for delayed writes where the data is 2427 * going to go to disk eventually without additional VM intevention. 2428 * 2429 * Note that while we only really need to clean through to b_bcount, we 2430 * just go ahead and clean through to b_bufsize. 2431 */ 2432static void 2433vfs_clean_pages_dirty_buf(struct buf *bp) 2434{ 2435 vm_ooffset_t foff, noff, eoff; 2436 vm_page_t m; 2437 int i; 2438 2439 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2440 return; 2441 2442 foff = bp->b_offset; 2443 KASSERT(bp->b_offset != NOOFFSET, 2444 ("vfs_clean_pages_dirty_buf: no buffer offset")); 2445 2446 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2447 vfs_drain_busy_pages(bp); 2448 vfs_setdirty_locked_object(bp); 2449 for (i = 0; i < bp->b_npages; i++) { 2450 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2451 eoff = noff; 2452 if (eoff > bp->b_offset + bp->b_bufsize) 2453 eoff = bp->b_offset + bp->b_bufsize; 2454 m = bp->b_pages[i]; 2455 vfs_page_set_validclean(bp, foff, m); 2456 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2457 foff = noff; 2458 } 2459 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2460} 2461 2462static void 2463vfs_setdirty_locked_object(struct buf *bp) 2464{ 2465 vm_object_t object; 2466 int i; 2467 2468 object = bp->b_bufobj->bo_object; 2469 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2470 2471 /* 2472 * We qualify the scan for modified pages on whether the 2473 * object has been flushed yet. 2474 */ 2475 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2476 vm_offset_t boffset; 2477 vm_offset_t eoffset; 2478 2479 /* 2480 * test the pages to see if they have been modified directly 2481 * by users through the VM system. 2482 */ 2483 for (i = 0; i < bp->b_npages; i++) 2484 vm_page_test_dirty(bp->b_pages[i]); 2485 2486 /* 2487 * Calculate the encompassing dirty range, boffset and eoffset, 2488 * (eoffset - boffset) bytes. 2489 */ 2490 2491 for (i = 0; i < bp->b_npages; i++) { 2492 if (bp->b_pages[i]->dirty) 2493 break; 2494 } 2495 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2496 2497 for (i = bp->b_npages - 1; i >= 0; --i) { 2498 if (bp->b_pages[i]->dirty) { 2499 break; 2500 } 2501 } 2502 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2503 2504 /* 2505 * Fit it to the buffer. 2506 */ 2507 2508 if (eoffset > bp->b_bcount) 2509 eoffset = bp->b_bcount; 2510 2511 /* 2512 * If we have a good dirty range, merge with the existing 2513 * dirty range. 2514 */ 2515 2516 if (boffset < eoffset) { 2517 if (bp->b_dirtyoff > boffset) 2518 bp->b_dirtyoff = boffset; 2519 if (bp->b_dirtyend < eoffset) 2520 bp->b_dirtyend = eoffset; 2521 } 2522 } 2523} 2524 2525/* 2526 * getblk: 2527 * 2528 * Get a block given a specified block and offset into a file/device. 2529 * The buffers B_DONE bit will be cleared on return, making it almost 2530 * ready for an I/O initiation. B_INVAL may or may not be set on 2531 * return. The caller should clear B_INVAL prior to initiating a 2532 * READ. 2533 * 2534 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2535 * an existing buffer. 2536 * 2537 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2538 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2539 * and then cleared based on the backing VM. If the previous buffer is 2540 * non-0-sized but invalid, B_CACHE will be cleared. 2541 * 2542 * If getblk() must create a new buffer, the new buffer is returned with 2543 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2544 * case it is returned with B_INVAL clear and B_CACHE set based on the 2545 * backing VM. 2546 * 2547 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2548 * B_CACHE bit is clear. 2549 * 2550 * What this means, basically, is that the caller should use B_CACHE to 2551 * determine whether the buffer is fully valid or not and should clear 2552 * B_INVAL prior to issuing a read. If the caller intends to validate 2553 * the buffer by loading its data area with something, the caller needs 2554 * to clear B_INVAL. If the caller does this without issuing an I/O, 2555 * the caller should set B_CACHE ( as an optimization ), else the caller 2556 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2557 * a write attempt or if it was a successfull read. If the caller 2558 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2559 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2560 */ 2561struct buf * 2562getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2563 int flags) 2564{ 2565 struct buf *bp; 2566 struct bufobj *bo; 2567 int error; 2568 2569 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2570 ASSERT_VOP_LOCKED(vp, "getblk"); 2571 if (size > MAXBSIZE) 2572 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2573 2574 bo = &vp->v_bufobj; 2575loop: 2576 /* 2577 * Block if we are low on buffers. Certain processes are allowed 2578 * to completely exhaust the buffer cache. 2579 * 2580 * If this check ever becomes a bottleneck it may be better to 2581 * move it into the else, when gbincore() fails. At the moment 2582 * it isn't a problem. 2583 * 2584 * XXX remove if 0 sections (clean this up after its proven) 2585 */ 2586 if (numfreebuffers == 0) { 2587 if (TD_IS_IDLETHREAD(curthread)) 2588 return NULL; 2589 mtx_lock(&nblock); 2590 needsbuffer |= VFS_BIO_NEED_ANY; 2591 mtx_unlock(&nblock); 2592 } 2593 2594 BO_LOCK(bo); 2595 bp = gbincore(bo, blkno); 2596 if (bp != NULL) { 2597 int lockflags; 2598 /* 2599 * Buffer is in-core. If the buffer is not busy, it must 2600 * be on a queue. 2601 */ 2602 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2603 2604 if (flags & GB_LOCK_NOWAIT) 2605 lockflags |= LK_NOWAIT; 2606 2607 error = BUF_TIMELOCK(bp, lockflags, 2608 BO_MTX(bo), "getblk", slpflag, slptimeo); 2609 2610 /* 2611 * If we slept and got the lock we have to restart in case 2612 * the buffer changed identities. 2613 */ 2614 if (error == ENOLCK) 2615 goto loop; 2616 /* We timed out or were interrupted. */ 2617 else if (error) 2618 return (NULL); 2619 2620 /* 2621 * The buffer is locked. B_CACHE is cleared if the buffer is 2622 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2623 * and for a VMIO buffer B_CACHE is adjusted according to the 2624 * backing VM cache. 2625 */ 2626 if (bp->b_flags & B_INVAL) 2627 bp->b_flags &= ~B_CACHE; 2628 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2629 bp->b_flags |= B_CACHE; 2630 bremfree(bp); 2631 2632 /* 2633 * check for size inconsistancies for non-VMIO case. 2634 */ 2635 2636 if (bp->b_bcount != size) { 2637 if ((bp->b_flags & B_VMIO) == 0 || 2638 (size > bp->b_kvasize)) { 2639 if (bp->b_flags & B_DELWRI) { 2640 /* 2641 * If buffer is pinned and caller does 2642 * not want sleep waiting for it to be 2643 * unpinned, bail out 2644 * */ 2645 if (bp->b_pin_count > 0) { 2646 if (flags & GB_LOCK_NOWAIT) { 2647 bqrelse(bp); 2648 return (NULL); 2649 } else { 2650 bunpin_wait(bp); 2651 } 2652 } 2653 bp->b_flags |= B_NOCACHE; 2654 bwrite(bp); 2655 } else { 2656 if (LIST_EMPTY(&bp->b_dep)) { 2657 bp->b_flags |= B_RELBUF; 2658 brelse(bp); 2659 } else { 2660 bp->b_flags |= B_NOCACHE; 2661 bwrite(bp); 2662 } 2663 } 2664 goto loop; 2665 } 2666 } 2667 2668 /* 2669 * If the size is inconsistant in the VMIO case, we can resize 2670 * the buffer. This might lead to B_CACHE getting set or 2671 * cleared. If the size has not changed, B_CACHE remains 2672 * unchanged from its previous state. 2673 */ 2674 2675 if (bp->b_bcount != size) 2676 allocbuf(bp, size); 2677 2678 KASSERT(bp->b_offset != NOOFFSET, 2679 ("getblk: no buffer offset")); 2680 2681 /* 2682 * A buffer with B_DELWRI set and B_CACHE clear must 2683 * be committed before we can return the buffer in 2684 * order to prevent the caller from issuing a read 2685 * ( due to B_CACHE not being set ) and overwriting 2686 * it. 2687 * 2688 * Most callers, including NFS and FFS, need this to 2689 * operate properly either because they assume they 2690 * can issue a read if B_CACHE is not set, or because 2691 * ( for example ) an uncached B_DELWRI might loop due 2692 * to softupdates re-dirtying the buffer. In the latter 2693 * case, B_CACHE is set after the first write completes, 2694 * preventing further loops. 2695 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2696 * above while extending the buffer, we cannot allow the 2697 * buffer to remain with B_CACHE set after the write 2698 * completes or it will represent a corrupt state. To 2699 * deal with this we set B_NOCACHE to scrap the buffer 2700 * after the write. 2701 * 2702 * We might be able to do something fancy, like setting 2703 * B_CACHE in bwrite() except if B_DELWRI is already set, 2704 * so the below call doesn't set B_CACHE, but that gets real 2705 * confusing. This is much easier. 2706 */ 2707 2708 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2709 bp->b_flags |= B_NOCACHE; 2710 bwrite(bp); 2711 goto loop; 2712 } 2713 bp->b_flags &= ~B_DONE; 2714 } else { 2715 int bsize, maxsize, vmio; 2716 off_t offset; 2717 2718 /* 2719 * Buffer is not in-core, create new buffer. The buffer 2720 * returned by getnewbuf() is locked. Note that the returned 2721 * buffer is also considered valid (not marked B_INVAL). 2722 */ 2723 BO_UNLOCK(bo); 2724 /* 2725 * If the user does not want us to create the buffer, bail out 2726 * here. 2727 */ 2728 if (flags & GB_NOCREAT) 2729 return NULL; 2730 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 2731 offset = blkno * bsize; 2732 vmio = vp->v_object != NULL; 2733 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2734 maxsize = imax(maxsize, bsize); 2735 2736 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 2737 if (bp == NULL) { 2738 if (slpflag || slptimeo) 2739 return NULL; 2740 goto loop; 2741 } 2742 2743 /* 2744 * This code is used to make sure that a buffer is not 2745 * created while the getnewbuf routine is blocked. 2746 * This can be a problem whether the vnode is locked or not. 2747 * If the buffer is created out from under us, we have to 2748 * throw away the one we just created. 2749 * 2750 * Note: this must occur before we associate the buffer 2751 * with the vp especially considering limitations in 2752 * the splay tree implementation when dealing with duplicate 2753 * lblkno's. 2754 */ 2755 BO_LOCK(bo); 2756 if (gbincore(bo, blkno)) { 2757 BO_UNLOCK(bo); 2758 bp->b_flags |= B_INVAL; 2759 brelse(bp); 2760 goto loop; 2761 } 2762 2763 /* 2764 * Insert the buffer into the hash, so that it can 2765 * be found by incore. 2766 */ 2767 bp->b_blkno = bp->b_lblkno = blkno; 2768 bp->b_offset = offset; 2769 bgetvp(vp, bp); 2770 BO_UNLOCK(bo); 2771 2772 /* 2773 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2774 * buffer size starts out as 0, B_CACHE will be set by 2775 * allocbuf() for the VMIO case prior to it testing the 2776 * backing store for validity. 2777 */ 2778 2779 if (vmio) { 2780 bp->b_flags |= B_VMIO; 2781#if defined(VFS_BIO_DEBUG) 2782 if (vn_canvmio(vp) != TRUE) 2783 printf("getblk: VMIO on vnode type %d\n", 2784 vp->v_type); 2785#endif 2786 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2787 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2788 bp, vp->v_object, bp->b_bufobj->bo_object)); 2789 } else { 2790 bp->b_flags &= ~B_VMIO; 2791 KASSERT(bp->b_bufobj->bo_object == NULL, 2792 ("ARGH! has b_bufobj->bo_object %p %p\n", 2793 bp, bp->b_bufobj->bo_object)); 2794 } 2795 2796 allocbuf(bp, size); 2797 bp->b_flags &= ~B_DONE; 2798 } 2799 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2800 BUF_ASSERT_HELD(bp); 2801 KASSERT(bp->b_bufobj == bo, 2802 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2803 return (bp); 2804} 2805 2806/* 2807 * Get an empty, disassociated buffer of given size. The buffer is initially 2808 * set to B_INVAL. 2809 */ 2810struct buf * 2811geteblk(int size, int flags) 2812{ 2813 struct buf *bp; 2814 int maxsize; 2815 2816 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2817 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 2818 if ((flags & GB_NOWAIT_BD) && 2819 (curthread->td_pflags & TDP_BUFNEED) != 0) 2820 return (NULL); 2821 } 2822 allocbuf(bp, size); 2823 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2824 BUF_ASSERT_HELD(bp); 2825 return (bp); 2826} 2827 2828 2829/* 2830 * This code constitutes the buffer memory from either anonymous system 2831 * memory (in the case of non-VMIO operations) or from an associated 2832 * VM object (in the case of VMIO operations). This code is able to 2833 * resize a buffer up or down. 2834 * 2835 * Note that this code is tricky, and has many complications to resolve 2836 * deadlock or inconsistant data situations. Tread lightly!!! 2837 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2838 * the caller. Calling this code willy nilly can result in the loss of data. 2839 * 2840 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2841 * B_CACHE for the non-VMIO case. 2842 */ 2843 2844int 2845allocbuf(struct buf *bp, int size) 2846{ 2847 int newbsize, mbsize; 2848 int i; 2849 2850 BUF_ASSERT_HELD(bp); 2851 2852 if (bp->b_kvasize < size) 2853 panic("allocbuf: buffer too small"); 2854 2855 if ((bp->b_flags & B_VMIO) == 0) { 2856 caddr_t origbuf; 2857 int origbufsize; 2858 /* 2859 * Just get anonymous memory from the kernel. Don't 2860 * mess with B_CACHE. 2861 */ 2862 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2863 if (bp->b_flags & B_MALLOC) 2864 newbsize = mbsize; 2865 else 2866 newbsize = round_page(size); 2867 2868 if (newbsize < bp->b_bufsize) { 2869 /* 2870 * malloced buffers are not shrunk 2871 */ 2872 if (bp->b_flags & B_MALLOC) { 2873 if (newbsize) { 2874 bp->b_bcount = size; 2875 } else { 2876 free(bp->b_data, M_BIOBUF); 2877 if (bp->b_bufsize) { 2878 atomic_subtract_long( 2879 &bufmallocspace, 2880 bp->b_bufsize); 2881 bufspacewakeup(); 2882 bp->b_bufsize = 0; 2883 } 2884 bp->b_saveaddr = bp->b_kvabase; 2885 bp->b_data = bp->b_saveaddr; 2886 bp->b_bcount = 0; 2887 bp->b_flags &= ~B_MALLOC; 2888 } 2889 return 1; 2890 } 2891 vm_hold_free_pages( 2892 bp, 2893 (vm_offset_t) bp->b_data + newbsize, 2894 (vm_offset_t) bp->b_data + bp->b_bufsize); 2895 } else if (newbsize > bp->b_bufsize) { 2896 /* 2897 * We only use malloced memory on the first allocation. 2898 * and revert to page-allocated memory when the buffer 2899 * grows. 2900 */ 2901 /* 2902 * There is a potential smp race here that could lead 2903 * to bufmallocspace slightly passing the max. It 2904 * is probably extremely rare and not worth worrying 2905 * over. 2906 */ 2907 if ( (bufmallocspace < maxbufmallocspace) && 2908 (bp->b_bufsize == 0) && 2909 (mbsize <= PAGE_SIZE/2)) { 2910 2911 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2912 bp->b_bufsize = mbsize; 2913 bp->b_bcount = size; 2914 bp->b_flags |= B_MALLOC; 2915 atomic_add_long(&bufmallocspace, mbsize); 2916 return 1; 2917 } 2918 origbuf = NULL; 2919 origbufsize = 0; 2920 /* 2921 * If the buffer is growing on its other-than-first allocation, 2922 * then we revert to the page-allocation scheme. 2923 */ 2924 if (bp->b_flags & B_MALLOC) { 2925 origbuf = bp->b_data; 2926 origbufsize = bp->b_bufsize; 2927 bp->b_data = bp->b_kvabase; 2928 if (bp->b_bufsize) { 2929 atomic_subtract_long(&bufmallocspace, 2930 bp->b_bufsize); 2931 bufspacewakeup(); 2932 bp->b_bufsize = 0; 2933 } 2934 bp->b_flags &= ~B_MALLOC; 2935 newbsize = round_page(newbsize); 2936 } 2937 vm_hold_load_pages( 2938 bp, 2939 (vm_offset_t) bp->b_data + bp->b_bufsize, 2940 (vm_offset_t) bp->b_data + newbsize); 2941 if (origbuf) { 2942 bcopy(origbuf, bp->b_data, origbufsize); 2943 free(origbuf, M_BIOBUF); 2944 } 2945 } 2946 } else { 2947 int desiredpages; 2948 2949 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2950 desiredpages = (size == 0) ? 0 : 2951 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2952 2953 if (bp->b_flags & B_MALLOC) 2954 panic("allocbuf: VMIO buffer can't be malloced"); 2955 /* 2956 * Set B_CACHE initially if buffer is 0 length or will become 2957 * 0-length. 2958 */ 2959 if (size == 0 || bp->b_bufsize == 0) 2960 bp->b_flags |= B_CACHE; 2961 2962 if (newbsize < bp->b_bufsize) { 2963 /* 2964 * DEV_BSIZE aligned new buffer size is less then the 2965 * DEV_BSIZE aligned existing buffer size. Figure out 2966 * if we have to remove any pages. 2967 */ 2968 if (desiredpages < bp->b_npages) { 2969 vm_page_t m; 2970 2971 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2972 for (i = desiredpages; i < bp->b_npages; i++) { 2973 /* 2974 * the page is not freed here -- it 2975 * is the responsibility of 2976 * vnode_pager_setsize 2977 */ 2978 m = bp->b_pages[i]; 2979 KASSERT(m != bogus_page, 2980 ("allocbuf: bogus page found")); 2981 while (vm_page_sleep_if_busy(m, TRUE, 2982 "biodep")) 2983 continue; 2984 2985 bp->b_pages[i] = NULL; 2986 vm_page_lock(m); 2987 vm_page_unwire(m, 0); 2988 vm_page_unlock(m); 2989 } 2990 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2991 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2992 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2993 bp->b_npages = desiredpages; 2994 } 2995 } else if (size > bp->b_bcount) { 2996 /* 2997 * We are growing the buffer, possibly in a 2998 * byte-granular fashion. 2999 */ 3000 vm_object_t obj; 3001 vm_offset_t toff; 3002 vm_offset_t tinc; 3003 3004 /* 3005 * Step 1, bring in the VM pages from the object, 3006 * allocating them if necessary. We must clear 3007 * B_CACHE if these pages are not valid for the 3008 * range covered by the buffer. 3009 */ 3010 3011 obj = bp->b_bufobj->bo_object; 3012 3013 VM_OBJECT_LOCK(obj); 3014 while (bp->b_npages < desiredpages) { 3015 vm_page_t m; 3016 vm_pindex_t pi; 3017 3018 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 3019 if ((m = vm_page_lookup(obj, pi)) == NULL) { 3020 /* 3021 * note: must allocate system pages 3022 * since blocking here could intefere 3023 * with paging I/O, no matter which 3024 * process we are. 3025 */ 3026 m = vm_page_alloc(obj, pi, 3027 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 3028 VM_ALLOC_WIRED); 3029 if (m == NULL) { 3030 atomic_add_int(&vm_pageout_deficit, 3031 desiredpages - bp->b_npages); 3032 VM_OBJECT_UNLOCK(obj); 3033 VM_WAIT; 3034 VM_OBJECT_LOCK(obj); 3035 } else { 3036 if (m->valid == 0) 3037 bp->b_flags &= ~B_CACHE; 3038 bp->b_pages[bp->b_npages] = m; 3039 ++bp->b_npages; 3040 } 3041 continue; 3042 } 3043 3044 /* 3045 * We found a page. If we have to sleep on it, 3046 * retry because it might have gotten freed out 3047 * from under us. 3048 * 3049 * We can only test VPO_BUSY here. Blocking on 3050 * m->busy might lead to a deadlock: 3051 * 3052 * vm_fault->getpages->cluster_read->allocbuf 3053 * 3054 */ 3055 if ((m->oflags & VPO_BUSY) != 0) { 3056 /* 3057 * Reference the page before unlocking 3058 * and sleeping so that the page daemon 3059 * is less likely to reclaim it. 3060 */ 3061 vm_page_lock_queues(); 3062 vm_page_flag_set(m, PG_REFERENCED); 3063 vm_page_sleep(m, "pgtblk"); 3064 continue; 3065 } 3066 3067 /* 3068 * We have a good page. 3069 */ 3070 vm_page_lock(m); 3071 vm_page_wire(m); 3072 vm_page_unlock(m); 3073 bp->b_pages[bp->b_npages] = m; 3074 ++bp->b_npages; 3075 } 3076 3077 /* 3078 * Step 2. We've loaded the pages into the buffer, 3079 * we have to figure out if we can still have B_CACHE 3080 * set. Note that B_CACHE is set according to the 3081 * byte-granular range ( bcount and size ), new the 3082 * aligned range ( newbsize ). 3083 * 3084 * The VM test is against m->valid, which is DEV_BSIZE 3085 * aligned. Needless to say, the validity of the data 3086 * needs to also be DEV_BSIZE aligned. Note that this 3087 * fails with NFS if the server or some other client 3088 * extends the file's EOF. If our buffer is resized, 3089 * B_CACHE may remain set! XXX 3090 */ 3091 3092 toff = bp->b_bcount; 3093 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3094 3095 while ((bp->b_flags & B_CACHE) && toff < size) { 3096 vm_pindex_t pi; 3097 3098 if (tinc > (size - toff)) 3099 tinc = size - toff; 3100 3101 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3102 PAGE_SHIFT; 3103 3104 vfs_buf_test_cache( 3105 bp, 3106 bp->b_offset, 3107 toff, 3108 tinc, 3109 bp->b_pages[pi] 3110 ); 3111 toff += tinc; 3112 tinc = PAGE_SIZE; 3113 } 3114 VM_OBJECT_UNLOCK(obj); 3115 3116 /* 3117 * Step 3, fixup the KVM pmap. Remember that 3118 * bp->b_data is relative to bp->b_offset, but 3119 * bp->b_offset may be offset into the first page. 3120 */ 3121 3122 bp->b_data = (caddr_t) 3123 trunc_page((vm_offset_t)bp->b_data); 3124 pmap_qenter( 3125 (vm_offset_t)bp->b_data, 3126 bp->b_pages, 3127 bp->b_npages 3128 ); 3129 3130 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3131 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3132 } 3133 } 3134 if (newbsize < bp->b_bufsize) 3135 bufspacewakeup(); 3136 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3137 bp->b_bcount = size; /* requested buffer size */ 3138 return 1; 3139} 3140 3141void 3142biodone(struct bio *bp) 3143{ 3144 struct mtx *mtxp; 3145 void (*done)(struct bio *); 3146 3147 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3148 mtx_lock(mtxp); 3149 bp->bio_flags |= BIO_DONE; 3150 done = bp->bio_done; 3151 if (done == NULL) 3152 wakeup(bp); 3153 mtx_unlock(mtxp); 3154 if (done != NULL) 3155 done(bp); 3156} 3157 3158/* 3159 * Wait for a BIO to finish. 3160 * 3161 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3162 * case is not yet clear. 3163 */ 3164int 3165biowait(struct bio *bp, const char *wchan) 3166{ 3167 struct mtx *mtxp; 3168 3169 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3170 mtx_lock(mtxp); 3171 while ((bp->bio_flags & BIO_DONE) == 0) 3172 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3173 mtx_unlock(mtxp); 3174 if (bp->bio_error != 0) 3175 return (bp->bio_error); 3176 if (!(bp->bio_flags & BIO_ERROR)) 3177 return (0); 3178 return (EIO); 3179} 3180 3181void 3182biofinish(struct bio *bp, struct devstat *stat, int error) 3183{ 3184 3185 if (error) { 3186 bp->bio_error = error; 3187 bp->bio_flags |= BIO_ERROR; 3188 } 3189 if (stat != NULL) 3190 devstat_end_transaction_bio(stat, bp); 3191 biodone(bp); 3192} 3193 3194/* 3195 * bufwait: 3196 * 3197 * Wait for buffer I/O completion, returning error status. The buffer 3198 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3199 * error and cleared. 3200 */ 3201int 3202bufwait(struct buf *bp) 3203{ 3204 if (bp->b_iocmd == BIO_READ) 3205 bwait(bp, PRIBIO, "biord"); 3206 else 3207 bwait(bp, PRIBIO, "biowr"); 3208 if (bp->b_flags & B_EINTR) { 3209 bp->b_flags &= ~B_EINTR; 3210 return (EINTR); 3211 } 3212 if (bp->b_ioflags & BIO_ERROR) { 3213 return (bp->b_error ? bp->b_error : EIO); 3214 } else { 3215 return (0); 3216 } 3217} 3218 3219 /* 3220 * Call back function from struct bio back up to struct buf. 3221 */ 3222static void 3223bufdonebio(struct bio *bip) 3224{ 3225 struct buf *bp; 3226 3227 bp = bip->bio_caller2; 3228 bp->b_resid = bp->b_bcount - bip->bio_completed; 3229 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3230 bp->b_ioflags = bip->bio_flags; 3231 bp->b_error = bip->bio_error; 3232 if (bp->b_error) 3233 bp->b_ioflags |= BIO_ERROR; 3234 bufdone(bp); 3235 g_destroy_bio(bip); 3236} 3237 3238void 3239dev_strategy(struct cdev *dev, struct buf *bp) 3240{ 3241 struct cdevsw *csw; 3242 struct bio *bip; 3243 3244 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3245 panic("b_iocmd botch"); 3246 for (;;) { 3247 bip = g_new_bio(); 3248 if (bip != NULL) 3249 break; 3250 /* Try again later */ 3251 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3252 } 3253 bip->bio_cmd = bp->b_iocmd; 3254 bip->bio_offset = bp->b_iooffset; 3255 bip->bio_length = bp->b_bcount; 3256 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3257 bip->bio_data = bp->b_data; 3258 bip->bio_done = bufdonebio; 3259 bip->bio_caller2 = bp; 3260 bip->bio_dev = dev; 3261 KASSERT(dev->si_refcount > 0, 3262 ("dev_strategy on un-referenced struct cdev *(%s)", 3263 devtoname(dev))); 3264 csw = dev_refthread(dev); 3265 if (csw == NULL) { 3266 g_destroy_bio(bip); 3267 bp->b_error = ENXIO; 3268 bp->b_ioflags = BIO_ERROR; 3269 bufdone(bp); 3270 return; 3271 } 3272 (*csw->d_strategy)(bip); 3273 dev_relthread(dev); 3274} 3275 3276/* 3277 * bufdone: 3278 * 3279 * Finish I/O on a buffer, optionally calling a completion function. 3280 * This is usually called from an interrupt so process blocking is 3281 * not allowed. 3282 * 3283 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3284 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3285 * assuming B_INVAL is clear. 3286 * 3287 * For the VMIO case, we set B_CACHE if the op was a read and no 3288 * read error occured, or if the op was a write. B_CACHE is never 3289 * set if the buffer is invalid or otherwise uncacheable. 3290 * 3291 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3292 * initiator to leave B_INVAL set to brelse the buffer out of existance 3293 * in the biodone routine. 3294 */ 3295void 3296bufdone(struct buf *bp) 3297{ 3298 struct bufobj *dropobj; 3299 void (*biodone)(struct buf *); 3300 3301 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3302 dropobj = NULL; 3303 3304 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3305 BUF_ASSERT_HELD(bp); 3306 3307 runningbufwakeup(bp); 3308 if (bp->b_iocmd == BIO_WRITE) 3309 dropobj = bp->b_bufobj; 3310 /* call optional completion function if requested */ 3311 if (bp->b_iodone != NULL) { 3312 biodone = bp->b_iodone; 3313 bp->b_iodone = NULL; 3314 (*biodone) (bp); 3315 if (dropobj) 3316 bufobj_wdrop(dropobj); 3317 return; 3318 } 3319 3320 bufdone_finish(bp); 3321 3322 if (dropobj) 3323 bufobj_wdrop(dropobj); 3324} 3325 3326void 3327bufdone_finish(struct buf *bp) 3328{ 3329 BUF_ASSERT_HELD(bp); 3330 3331 if (!LIST_EMPTY(&bp->b_dep)) 3332 buf_complete(bp); 3333 3334 if (bp->b_flags & B_VMIO) { 3335 int i; 3336 vm_ooffset_t foff; 3337 vm_page_t m; 3338 vm_object_t obj; 3339 int bogus, iosize; 3340 struct vnode *vp = bp->b_vp; 3341 3342 obj = bp->b_bufobj->bo_object; 3343 3344#if defined(VFS_BIO_DEBUG) 3345 mp_fixme("usecount and vflag accessed without locks."); 3346 if (vp->v_usecount == 0) { 3347 panic("biodone: zero vnode ref count"); 3348 } 3349 3350 KASSERT(vp->v_object != NULL, 3351 ("biodone: vnode %p has no vm_object", vp)); 3352#endif 3353 3354 foff = bp->b_offset; 3355 KASSERT(bp->b_offset != NOOFFSET, 3356 ("biodone: no buffer offset")); 3357 3358 VM_OBJECT_LOCK(obj); 3359#if defined(VFS_BIO_DEBUG) 3360 if (obj->paging_in_progress < bp->b_npages) { 3361 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3362 obj->paging_in_progress, bp->b_npages); 3363 } 3364#endif 3365 3366 /* 3367 * Set B_CACHE if the op was a normal read and no error 3368 * occured. B_CACHE is set for writes in the b*write() 3369 * routines. 3370 */ 3371 iosize = bp->b_bcount - bp->b_resid; 3372 if (bp->b_iocmd == BIO_READ && 3373 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3374 !(bp->b_ioflags & BIO_ERROR)) { 3375 bp->b_flags |= B_CACHE; 3376 } 3377 bogus = 0; 3378 for (i = 0; i < bp->b_npages; i++) { 3379 int bogusflag = 0; 3380 int resid; 3381 3382 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3383 if (resid > iosize) 3384 resid = iosize; 3385 3386 /* 3387 * cleanup bogus pages, restoring the originals 3388 */ 3389 m = bp->b_pages[i]; 3390 if (m == bogus_page) { 3391 bogus = bogusflag = 1; 3392 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3393 if (m == NULL) 3394 panic("biodone: page disappeared!"); 3395 bp->b_pages[i] = m; 3396 } 3397#if defined(VFS_BIO_DEBUG) 3398 if (OFF_TO_IDX(foff) != m->pindex) { 3399 printf( 3400"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3401 (intmax_t)foff, (uintmax_t)m->pindex); 3402 } 3403#endif 3404 3405 /* 3406 * In the write case, the valid and clean bits are 3407 * already changed correctly ( see bdwrite() ), so we 3408 * only need to do this here in the read case. 3409 */ 3410 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3411 KASSERT((m->dirty & vm_page_bits(foff & 3412 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3413 " page %p has unexpected dirty bits", m)); 3414 vfs_page_set_valid(bp, foff, m); 3415 } 3416 3417 /* 3418 * when debugging new filesystems or buffer I/O methods, this 3419 * is the most common error that pops up. if you see this, you 3420 * have not set the page busy flag correctly!!! 3421 */ 3422 if (m->busy == 0) { 3423 printf("biodone: page busy < 0, " 3424 "pindex: %d, foff: 0x(%x,%x), " 3425 "resid: %d, index: %d\n", 3426 (int) m->pindex, (int)(foff >> 32), 3427 (int) foff & 0xffffffff, resid, i); 3428 if (!vn_isdisk(vp, NULL)) 3429 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3430 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3431 (intmax_t) bp->b_lblkno, 3432 bp->b_flags, bp->b_npages); 3433 else 3434 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3435 (intmax_t) bp->b_lblkno, 3436 bp->b_flags, bp->b_npages); 3437 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3438 (u_long)m->valid, (u_long)m->dirty, 3439 m->wire_count); 3440 panic("biodone: page busy < 0\n"); 3441 } 3442 vm_page_io_finish(m); 3443 vm_object_pip_subtract(obj, 1); 3444 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3445 iosize -= resid; 3446 } 3447 vm_object_pip_wakeupn(obj, 0); 3448 VM_OBJECT_UNLOCK(obj); 3449 if (bogus) 3450 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3451 bp->b_pages, bp->b_npages); 3452 } 3453 3454 /* 3455 * For asynchronous completions, release the buffer now. The brelse 3456 * will do a wakeup there if necessary - so no need to do a wakeup 3457 * here in the async case. The sync case always needs to do a wakeup. 3458 */ 3459 3460 if (bp->b_flags & B_ASYNC) { 3461 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3462 brelse(bp); 3463 else 3464 bqrelse(bp); 3465 } else 3466 bdone(bp); 3467} 3468 3469/* 3470 * This routine is called in lieu of iodone in the case of 3471 * incomplete I/O. This keeps the busy status for pages 3472 * consistant. 3473 */ 3474void 3475vfs_unbusy_pages(struct buf *bp) 3476{ 3477 int i; 3478 vm_object_t obj; 3479 vm_page_t m; 3480 3481 runningbufwakeup(bp); 3482 if (!(bp->b_flags & B_VMIO)) 3483 return; 3484 3485 obj = bp->b_bufobj->bo_object; 3486 VM_OBJECT_LOCK(obj); 3487 for (i = 0; i < bp->b_npages; i++) { 3488 m = bp->b_pages[i]; 3489 if (m == bogus_page) { 3490 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3491 if (!m) 3492 panic("vfs_unbusy_pages: page missing\n"); 3493 bp->b_pages[i] = m; 3494 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3495 bp->b_pages, bp->b_npages); 3496 } 3497 vm_object_pip_subtract(obj, 1); 3498 vm_page_io_finish(m); 3499 } 3500 vm_object_pip_wakeupn(obj, 0); 3501 VM_OBJECT_UNLOCK(obj); 3502} 3503 3504/* 3505 * vfs_page_set_valid: 3506 * 3507 * Set the valid bits in a page based on the supplied offset. The 3508 * range is restricted to the buffer's size. 3509 * 3510 * This routine is typically called after a read completes. 3511 */ 3512static void 3513vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3514{ 3515 vm_ooffset_t eoff; 3516 3517 /* 3518 * Compute the end offset, eoff, such that [off, eoff) does not span a 3519 * page boundary and eoff is not greater than the end of the buffer. 3520 * The end of the buffer, in this case, is our file EOF, not the 3521 * allocation size of the buffer. 3522 */ 3523 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3524 if (eoff > bp->b_offset + bp->b_bcount) 3525 eoff = bp->b_offset + bp->b_bcount; 3526 3527 /* 3528 * Set valid range. This is typically the entire buffer and thus the 3529 * entire page. 3530 */ 3531 if (eoff > off) 3532 vm_page_set_valid(m, off & PAGE_MASK, eoff - off); 3533} 3534 3535/* 3536 * vfs_page_set_validclean: 3537 * 3538 * Set the valid bits and clear the dirty bits in a page based on the 3539 * supplied offset. The range is restricted to the buffer's size. 3540 */ 3541static void 3542vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3543{ 3544 vm_ooffset_t soff, eoff; 3545 3546 /* 3547 * Start and end offsets in buffer. eoff - soff may not cross a 3548 * page boundry or cross the end of the buffer. The end of the 3549 * buffer, in this case, is our file EOF, not the allocation size 3550 * of the buffer. 3551 */ 3552 soff = off; 3553 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3554 if (eoff > bp->b_offset + bp->b_bcount) 3555 eoff = bp->b_offset + bp->b_bcount; 3556 3557 /* 3558 * Set valid range. This is typically the entire buffer and thus the 3559 * entire page. 3560 */ 3561 if (eoff > soff) { 3562 vm_page_set_validclean( 3563 m, 3564 (vm_offset_t) (soff & PAGE_MASK), 3565 (vm_offset_t) (eoff - soff) 3566 ); 3567 } 3568} 3569 3570/* 3571 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If 3572 * any page is busy, drain the flag. 3573 */ 3574static void 3575vfs_drain_busy_pages(struct buf *bp) 3576{ 3577 vm_page_t m; 3578 int i, last_busied; 3579 3580 VM_OBJECT_LOCK_ASSERT(bp->b_bufobj->bo_object, MA_OWNED); 3581 last_busied = 0; 3582 for (i = 0; i < bp->b_npages; i++) { 3583 m = bp->b_pages[i]; 3584 if ((m->oflags & VPO_BUSY) != 0) { 3585 for (; last_busied < i; last_busied++) 3586 vm_page_busy(bp->b_pages[last_busied]); 3587 while ((m->oflags & VPO_BUSY) != 0) 3588 vm_page_sleep(m, "vbpage"); 3589 } 3590 } 3591 for (i = 0; i < last_busied; i++) 3592 vm_page_wakeup(bp->b_pages[i]); 3593} 3594 3595/* 3596 * This routine is called before a device strategy routine. 3597 * It is used to tell the VM system that paging I/O is in 3598 * progress, and treat the pages associated with the buffer 3599 * almost as being VPO_BUSY. Also the object paging_in_progress 3600 * flag is handled to make sure that the object doesn't become 3601 * inconsistant. 3602 * 3603 * Since I/O has not been initiated yet, certain buffer flags 3604 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3605 * and should be ignored. 3606 */ 3607void 3608vfs_busy_pages(struct buf *bp, int clear_modify) 3609{ 3610 int i, bogus; 3611 vm_object_t obj; 3612 vm_ooffset_t foff; 3613 vm_page_t m; 3614 3615 if (!(bp->b_flags & B_VMIO)) 3616 return; 3617 3618 obj = bp->b_bufobj->bo_object; 3619 foff = bp->b_offset; 3620 KASSERT(bp->b_offset != NOOFFSET, 3621 ("vfs_busy_pages: no buffer offset")); 3622 VM_OBJECT_LOCK(obj); 3623 vfs_drain_busy_pages(bp); 3624 if (bp->b_bufsize != 0) 3625 vfs_setdirty_locked_object(bp); 3626 bogus = 0; 3627 for (i = 0; i < bp->b_npages; i++) { 3628 m = bp->b_pages[i]; 3629 3630 if ((bp->b_flags & B_CLUSTER) == 0) { 3631 vm_object_pip_add(obj, 1); 3632 vm_page_io_start(m); 3633 } 3634 /* 3635 * When readying a buffer for a read ( i.e 3636 * clear_modify == 0 ), it is important to do 3637 * bogus_page replacement for valid pages in 3638 * partially instantiated buffers. Partially 3639 * instantiated buffers can, in turn, occur when 3640 * reconstituting a buffer from its VM backing store 3641 * base. We only have to do this if B_CACHE is 3642 * clear ( which causes the I/O to occur in the 3643 * first place ). The replacement prevents the read 3644 * I/O from overwriting potentially dirty VM-backed 3645 * pages. XXX bogus page replacement is, uh, bogus. 3646 * It may not work properly with small-block devices. 3647 * We need to find a better way. 3648 */ 3649 if (clear_modify) { 3650 pmap_remove_write(m); 3651 vfs_page_set_validclean(bp, foff, m); 3652 } else if (m->valid == VM_PAGE_BITS_ALL && 3653 (bp->b_flags & B_CACHE) == 0) { 3654 bp->b_pages[i] = bogus_page; 3655 bogus++; 3656 } 3657 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3658 } 3659 VM_OBJECT_UNLOCK(obj); 3660 if (bogus) 3661 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3662 bp->b_pages, bp->b_npages); 3663} 3664 3665/* 3666 * vfs_bio_set_valid: 3667 * 3668 * Set the range within the buffer to valid. The range is 3669 * relative to the beginning of the buffer, b_offset. Note that 3670 * b_offset itself may be offset from the beginning of the first 3671 * page. 3672 */ 3673void 3674vfs_bio_set_valid(struct buf *bp, int base, int size) 3675{ 3676 int i, n; 3677 vm_page_t m; 3678 3679 if (!(bp->b_flags & B_VMIO)) 3680 return; 3681 3682 /* 3683 * Fixup base to be relative to beginning of first page. 3684 * Set initial n to be the maximum number of bytes in the 3685 * first page that can be validated. 3686 */ 3687 base += (bp->b_offset & PAGE_MASK); 3688 n = PAGE_SIZE - (base & PAGE_MASK); 3689 3690 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3691 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3692 m = bp->b_pages[i]; 3693 if (n > size) 3694 n = size; 3695 vm_page_set_valid(m, base & PAGE_MASK, n); 3696 base += n; 3697 size -= n; 3698 n = PAGE_SIZE; 3699 } 3700 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3701} 3702 3703/* 3704 * vfs_bio_clrbuf: 3705 * 3706 * If the specified buffer is a non-VMIO buffer, clear the entire 3707 * buffer. If the specified buffer is a VMIO buffer, clear and 3708 * validate only the previously invalid portions of the buffer. 3709 * This routine essentially fakes an I/O, so we need to clear 3710 * BIO_ERROR and B_INVAL. 3711 * 3712 * Note that while we only theoretically need to clear through b_bcount, 3713 * we go ahead and clear through b_bufsize. 3714 */ 3715void 3716vfs_bio_clrbuf(struct buf *bp) 3717{ 3718 int i, j, mask; 3719 caddr_t sa, ea; 3720 3721 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3722 clrbuf(bp); 3723 return; 3724 } 3725 bp->b_flags &= ~B_INVAL; 3726 bp->b_ioflags &= ~BIO_ERROR; 3727 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3728 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3729 (bp->b_offset & PAGE_MASK) == 0) { 3730 if (bp->b_pages[0] == bogus_page) 3731 goto unlock; 3732 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3733 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3734 if ((bp->b_pages[0]->valid & mask) == mask) 3735 goto unlock; 3736 if ((bp->b_pages[0]->valid & mask) == 0) { 3737 bzero(bp->b_data, bp->b_bufsize); 3738 bp->b_pages[0]->valid |= mask; 3739 goto unlock; 3740 } 3741 } 3742 ea = sa = bp->b_data; 3743 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3744 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3745 ea = (caddr_t)(vm_offset_t)ulmin( 3746 (u_long)(vm_offset_t)ea, 3747 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3748 if (bp->b_pages[i] == bogus_page) 3749 continue; 3750 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3751 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3752 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3753 if ((bp->b_pages[i]->valid & mask) == mask) 3754 continue; 3755 if ((bp->b_pages[i]->valid & mask) == 0) 3756 bzero(sa, ea - sa); 3757 else { 3758 for (; sa < ea; sa += DEV_BSIZE, j++) { 3759 if ((bp->b_pages[i]->valid & (1 << j)) == 0) 3760 bzero(sa, DEV_BSIZE); 3761 } 3762 } 3763 bp->b_pages[i]->valid |= mask; 3764 } 3765unlock: 3766 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3767 bp->b_resid = 0; 3768} 3769 3770/* 3771 * vm_hold_load_pages and vm_hold_free_pages get pages into 3772 * a buffers address space. The pages are anonymous and are 3773 * not associated with a file object. 3774 */ 3775static void 3776vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3777{ 3778 vm_offset_t pg; 3779 vm_page_t p; 3780 int index; 3781 3782 to = round_page(to); 3783 from = round_page(from); 3784 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3785 3786 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3787tryagain: 3788 /* 3789 * note: must allocate system pages since blocking here 3790 * could interfere with paging I/O, no matter which 3791 * process we are. 3792 */ 3793 p = vm_page_alloc(NULL, pg >> PAGE_SHIFT, VM_ALLOC_NOOBJ | 3794 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3795 if (!p) { 3796 atomic_add_int(&vm_pageout_deficit, 3797 (to - pg) >> PAGE_SHIFT); 3798 VM_WAIT; 3799 goto tryagain; 3800 } 3801 pmap_qenter(pg, &p, 1); 3802 bp->b_pages[index] = p; 3803 } 3804 bp->b_npages = index; 3805} 3806 3807/* Return pages associated with this buf to the vm system */ 3808static void 3809vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3810{ 3811 vm_offset_t pg; 3812 vm_page_t p; 3813 int index, newnpages; 3814 3815 from = round_page(from); 3816 to = round_page(to); 3817 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3818 3819 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3820 p = bp->b_pages[index]; 3821 if (p && (index < bp->b_npages)) { 3822 if (p->busy) { 3823 printf( 3824 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3825 (intmax_t)bp->b_blkno, 3826 (intmax_t)bp->b_lblkno); 3827 } 3828 bp->b_pages[index] = NULL; 3829 pmap_qremove(pg, 1); 3830 p->wire_count--; 3831 vm_page_free(p); 3832 atomic_subtract_int(&cnt.v_wire_count, 1); 3833 } 3834 } 3835 bp->b_npages = newnpages; 3836} 3837 3838/* 3839 * Map an IO request into kernel virtual address space. 3840 * 3841 * All requests are (re)mapped into kernel VA space. 3842 * Notice that we use b_bufsize for the size of the buffer 3843 * to be mapped. b_bcount might be modified by the driver. 3844 * 3845 * Note that even if the caller determines that the address space should 3846 * be valid, a race or a smaller-file mapped into a larger space may 3847 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3848 * check the return value. 3849 */ 3850int 3851vmapbuf(struct buf *bp) 3852{ 3853 caddr_t addr, kva; 3854 vm_prot_t prot; 3855 int pidx, i; 3856 struct vm_page *m; 3857 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3858 3859 if (bp->b_bufsize < 0) 3860 return (-1); 3861 prot = VM_PROT_READ; 3862 if (bp->b_iocmd == BIO_READ) 3863 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3864 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3865 addr < bp->b_data + bp->b_bufsize; 3866 addr += PAGE_SIZE, pidx++) { 3867 /* 3868 * Do the vm_fault if needed; do the copy-on-write thing 3869 * when reading stuff off device into memory. 3870 * 3871 * NOTE! Must use pmap_extract() because addr may be in 3872 * the userland address space, and kextract is only guarenteed 3873 * to work for the kernland address space (see: sparc64 port). 3874 */ 3875retry: 3876 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3877 prot) < 0) { 3878 for (i = 0; i < pidx; ++i) { 3879 vm_page_lock(bp->b_pages[i]); 3880 vm_page_unhold(bp->b_pages[i]); 3881 vm_page_unlock(bp->b_pages[i]); 3882 bp->b_pages[i] = NULL; 3883 } 3884 return(-1); 3885 } 3886 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3887 if (m == NULL) 3888 goto retry; 3889 bp->b_pages[pidx] = m; 3890 } 3891 if (pidx > btoc(MAXPHYS)) 3892 panic("vmapbuf: mapped more than MAXPHYS"); 3893 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3894 3895 kva = bp->b_saveaddr; 3896 bp->b_npages = pidx; 3897 bp->b_saveaddr = bp->b_data; 3898 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3899 return(0); 3900} 3901 3902/* 3903 * Free the io map PTEs associated with this IO operation. 3904 * We also invalidate the TLB entries and restore the original b_addr. 3905 */ 3906void 3907vunmapbuf(struct buf *bp) 3908{ 3909 int pidx; 3910 int npages; 3911 3912 npages = bp->b_npages; 3913 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3914 for (pidx = 0; pidx < npages; pidx++) { 3915 vm_page_lock(bp->b_pages[pidx]); 3916 vm_page_unhold(bp->b_pages[pidx]); 3917 vm_page_unlock(bp->b_pages[pidx]); 3918 } 3919 3920 bp->b_data = bp->b_saveaddr; 3921} 3922 3923void 3924bdone(struct buf *bp) 3925{ 3926 struct mtx *mtxp; 3927 3928 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3929 mtx_lock(mtxp); 3930 bp->b_flags |= B_DONE; 3931 wakeup(bp); 3932 mtx_unlock(mtxp); 3933} 3934 3935void 3936bwait(struct buf *bp, u_char pri, const char *wchan) 3937{ 3938 struct mtx *mtxp; 3939 3940 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3941 mtx_lock(mtxp); 3942 while ((bp->b_flags & B_DONE) == 0) 3943 msleep(bp, mtxp, pri, wchan, 0); 3944 mtx_unlock(mtxp); 3945} 3946 3947int 3948bufsync(struct bufobj *bo, int waitfor) 3949{ 3950 3951 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 3952} 3953 3954void 3955bufstrategy(struct bufobj *bo, struct buf *bp) 3956{ 3957 int i = 0; 3958 struct vnode *vp; 3959 3960 vp = bp->b_vp; 3961 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3962 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3963 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3964 i = VOP_STRATEGY(vp, bp); 3965 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3966} 3967 3968void 3969bufobj_wrefl(struct bufobj *bo) 3970{ 3971 3972 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3973 ASSERT_BO_LOCKED(bo); 3974 bo->bo_numoutput++; 3975} 3976 3977void 3978bufobj_wref(struct bufobj *bo) 3979{ 3980 3981 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3982 BO_LOCK(bo); 3983 bo->bo_numoutput++; 3984 BO_UNLOCK(bo); 3985} 3986 3987void 3988bufobj_wdrop(struct bufobj *bo) 3989{ 3990 3991 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3992 BO_LOCK(bo); 3993 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3994 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3995 bo->bo_flag &= ~BO_WWAIT; 3996 wakeup(&bo->bo_numoutput); 3997 } 3998 BO_UNLOCK(bo); 3999} 4000 4001int 4002bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4003{ 4004 int error; 4005 4006 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4007 ASSERT_BO_LOCKED(bo); 4008 error = 0; 4009 while (bo->bo_numoutput) { 4010 bo->bo_flag |= BO_WWAIT; 4011 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 4012 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4013 if (error) 4014 break; 4015 } 4016 return (error); 4017} 4018 4019void 4020bpin(struct buf *bp) 4021{ 4022 struct mtx *mtxp; 4023 4024 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4025 mtx_lock(mtxp); 4026 bp->b_pin_count++; 4027 mtx_unlock(mtxp); 4028} 4029 4030void 4031bunpin(struct buf *bp) 4032{ 4033 struct mtx *mtxp; 4034 4035 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4036 mtx_lock(mtxp); 4037 if (--bp->b_pin_count == 0) 4038 wakeup(bp); 4039 mtx_unlock(mtxp); 4040} 4041 4042void 4043bunpin_wait(struct buf *bp) 4044{ 4045 struct mtx *mtxp; 4046 4047 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4048 mtx_lock(mtxp); 4049 while (bp->b_pin_count > 0) 4050 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4051 mtx_unlock(mtxp); 4052} 4053 4054#include "opt_ddb.h" 4055#ifdef DDB 4056#include <ddb/ddb.h> 4057 4058/* DDB command to show buffer data */ 4059DB_SHOW_COMMAND(buffer, db_show_buffer) 4060{ 4061 /* get args */ 4062 struct buf *bp = (struct buf *)addr; 4063 4064 if (!have_addr) { 4065 db_printf("usage: show buffer <addr>\n"); 4066 return; 4067 } 4068 4069 db_printf("buf at %p\n", bp); 4070 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4071 db_printf( 4072 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4073 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n", 4074 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4075 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4076 bp->b_dep.lh_first); 4077 if (bp->b_npages) { 4078 int i; 4079 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4080 for (i = 0; i < bp->b_npages; i++) { 4081 vm_page_t m; 4082 m = bp->b_pages[i]; 4083 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4084 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4085 if ((i + 1) < bp->b_npages) 4086 db_printf(","); 4087 } 4088 db_printf("\n"); 4089 } 4090 db_printf(" "); 4091 lockmgr_printinfo(&bp->b_lock); 4092} 4093 4094DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4095{ 4096 struct buf *bp; 4097 int i; 4098 4099 for (i = 0; i < nbuf; i++) { 4100 bp = &buf[i]; 4101 if (BUF_ISLOCKED(bp)) { 4102 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4103 db_printf("\n"); 4104 } 4105 } 4106} 4107 4108DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4109{ 4110 struct vnode *vp; 4111 struct buf *bp; 4112 4113 if (!have_addr) { 4114 db_printf("usage: show vnodebufs <addr>\n"); 4115 return; 4116 } 4117 vp = (struct vnode *)addr; 4118 db_printf("Clean buffers:\n"); 4119 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4120 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4121 db_printf("\n"); 4122 } 4123 db_printf("Dirty buffers:\n"); 4124 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4125 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4126 db_printf("\n"); 4127 } 4128} 4129 4130DB_COMMAND(countfreebufs, db_coundfreebufs) 4131{ 4132 struct buf *bp; 4133 int i, used = 0, nfree = 0; 4134 4135 if (have_addr) { 4136 db_printf("usage: countfreebufs\n"); 4137 return; 4138 } 4139 4140 for (i = 0; i < nbuf; i++) { 4141 bp = &buf[i]; 4142 if ((bp->b_vflags & BV_INFREECNT) != 0) 4143 nfree++; 4144 else 4145 used++; 4146 } 4147 4148 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4149 nfree + used); 4150 db_printf("numfreebuffers is %d\n", numfreebuffers); 4151} 4152#endif /* DDB */ 4153