vfs_bio.c revision 191218
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 191218 2009-04-17 18:11:11Z kan $"); 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/bio.h> 47#include <sys/conf.h> 48#include <sys/buf.h> 49#include <sys/devicestat.h> 50#include <sys/eventhandler.h> 51#include <sys/limits.h> 52#include <sys/lock.h> 53#include <sys/malloc.h> 54#include <sys/mount.h> 55#include <sys/mutex.h> 56#include <sys/kdb.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, 103 vm_page_t m); 104static void vfs_clean_pages(struct buf *bp); 105static void vfs_setdirty(struct buf *bp); 106static void vfs_setdirty_locked_object(struct buf *bp); 107static void vfs_vmio_release(struct buf *bp); 108static int vfs_bio_clcheck(struct vnode *vp, int size, 109 daddr_t lblkno, daddr_t blkno); 110static int buf_do_flush(struct vnode *vp); 111static int flushbufqueues(struct vnode *, int, int); 112static void buf_daemon(void); 113static void bremfreel(struct buf *bp); 114#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 115 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 116static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 117#endif 118 119int vmiodirenable = TRUE; 120SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 121 "Use the VM system for directory writes"); 122long runningbufspace; 123SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 124 "Amount of presently outstanding async buffer io"); 125static long bufspace; 126#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 127 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 128SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 129 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 130#else 131SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 132 "Virtual memory used for buffers"); 133#endif 134static long maxbufspace; 135SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 136 "Maximum allowed value of bufspace (including buf_daemon)"); 137static long bufmallocspace; 138SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 139 "Amount of malloced memory for buffers"); 140static long maxbufmallocspace; 141SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 142 "Maximum amount of malloced memory for buffers"); 143static long lobufspace; 144SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 145 "Minimum amount of buffers we want to have"); 146long hibufspace; 147SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 148 "Maximum allowed value of bufspace (excluding buf_daemon)"); 149static int bufreusecnt; 150SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 151 "Number of times we have reused a buffer"); 152static int buffreekvacnt; 153SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 154 "Number of times we have freed the KVA space from some buffer"); 155static int bufdefragcnt; 156SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 157 "Number of times we have had to repeat buffer allocation to defragment"); 158static long lorunningspace; 159SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 160 "Minimum preferred space used for in-progress I/O"); 161static long hirunningspace; 162SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 163 "Maximum amount of space to use for in-progress I/O"); 164int dirtybufferflushes; 165SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 166 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 167int bdwriteskip; 168SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 169 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 170int altbufferflushes; 171SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 172 0, "Number of fsync flushes to limit dirty buffers"); 173static int recursiveflushes; 174SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 175 0, "Number of flushes skipped due to being recursive"); 176static int numdirtybuffers; 177SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 178 "Number of buffers that are dirty (has unwritten changes) at the moment"); 179static int lodirtybuffers; 180SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 181 "How many buffers we want to have free before bufdaemon can sleep"); 182static int hidirtybuffers; 183SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 184 "When the number of dirty buffers is considered severe"); 185int dirtybufthresh; 186SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 187 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 188static int numfreebuffers; 189SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 190 "Number of free buffers"); 191static int lofreebuffers; 192SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 193 "XXX Unused"); 194static int hifreebuffers; 195SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 196 "XXX Complicatedly unused"); 197static int getnewbufcalls; 198SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 199 "Number of calls to getnewbuf"); 200static int getnewbufrestarts; 201SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 202 "Number of times getnewbuf has had to restart a buffer aquisition"); 203static int flushbufqtarget = 100; 204SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 205 "Amount of work to do in flushbufqueues when helping bufdaemon"); 206static long notbufdflashes; 207SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0, 208 "Number of dirty buffer flushes done by the bufdaemon helpers"); 209 210/* 211 * Wakeup point for bufdaemon, as well as indicator of whether it is already 212 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 213 * is idling. 214 */ 215static int bd_request; 216 217/* 218 * This lock synchronizes access to bd_request. 219 */ 220static struct mtx bdlock; 221 222/* 223 * bogus page -- for I/O to/from partially complete buffers 224 * this is a temporary solution to the problem, but it is not 225 * really that bad. it would be better to split the buffer 226 * for input in the case of buffers partially already in memory, 227 * but the code is intricate enough already. 228 */ 229vm_page_t bogus_page; 230 231/* 232 * Synchronization (sleep/wakeup) variable for active buffer space requests. 233 * Set when wait starts, cleared prior to wakeup(). 234 * Used in runningbufwakeup() and waitrunningbufspace(). 235 */ 236static int runningbufreq; 237 238/* 239 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 240 * waitrunningbufspace(). 241 */ 242static struct mtx rbreqlock; 243 244/* 245 * Synchronization (sleep/wakeup) variable for buffer requests. 246 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 247 * by and/or. 248 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 249 * getnewbuf(), and getblk(). 250 */ 251static int needsbuffer; 252 253/* 254 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 255 */ 256static struct mtx nblock; 257 258/* 259 * Definitions for the buffer free lists. 260 */ 261#define BUFFER_QUEUES 6 /* number of free buffer queues */ 262 263#define QUEUE_NONE 0 /* on no queue */ 264#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 265#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 266#define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ 267#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 268#define QUEUE_EMPTY 5 /* empty buffer headers */ 269#define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 270 271/* Queues for free buffers with various properties */ 272static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 273 274/* Lock for the bufqueues */ 275static struct mtx bqlock; 276 277/* 278 * Single global constant for BUF_WMESG, to avoid getting multiple references. 279 * buf_wmesg is referred from macros. 280 */ 281const char *buf_wmesg = BUF_WMESG; 282 283#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 284#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 285#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 286#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 287 288#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 289 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 290static int 291sysctl_bufspace(SYSCTL_HANDLER_ARGS) 292{ 293 long lvalue; 294 int ivalue; 295 296 if (sizeof(int) == sizeof(long) || req->oldlen == sizeof(long)) 297 return (sysctl_handle_long(oidp, arg1, arg2, req)); 298 lvalue = *(long *)arg1; 299 if (lvalue > INT_MAX) 300 /* On overflow, still write out a long to trigger ENOMEM. */ 301 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 302 ivalue = lvalue; 303 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 304} 305#endif 306 307#ifdef DIRECTIO 308extern void ffs_rawread_setup(void); 309#endif /* DIRECTIO */ 310/* 311 * numdirtywakeup: 312 * 313 * If someone is blocked due to there being too many dirty buffers, 314 * and numdirtybuffers is now reasonable, wake them up. 315 */ 316 317static __inline void 318numdirtywakeup(int level) 319{ 320 321 if (numdirtybuffers <= level) { 322 mtx_lock(&nblock); 323 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 324 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 325 wakeup(&needsbuffer); 326 } 327 mtx_unlock(&nblock); 328 } 329} 330 331/* 332 * bufspacewakeup: 333 * 334 * Called when buffer space is potentially available for recovery. 335 * getnewbuf() will block on this flag when it is unable to free 336 * sufficient buffer space. Buffer space becomes recoverable when 337 * bp's get placed back in the queues. 338 */ 339 340static __inline void 341bufspacewakeup(void) 342{ 343 344 /* 345 * If someone is waiting for BUF space, wake them up. Even 346 * though we haven't freed the kva space yet, the waiting 347 * process will be able to now. 348 */ 349 mtx_lock(&nblock); 350 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 351 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 352 wakeup(&needsbuffer); 353 } 354 mtx_unlock(&nblock); 355} 356 357/* 358 * runningbufwakeup() - in-progress I/O accounting. 359 * 360 */ 361void 362runningbufwakeup(struct buf *bp) 363{ 364 365 if (bp->b_runningbufspace) { 366 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace); 367 bp->b_runningbufspace = 0; 368 mtx_lock(&rbreqlock); 369 if (runningbufreq && runningbufspace <= lorunningspace) { 370 runningbufreq = 0; 371 wakeup(&runningbufreq); 372 } 373 mtx_unlock(&rbreqlock); 374 } 375} 376 377/* 378 * bufcountwakeup: 379 * 380 * Called when a buffer has been added to one of the free queues to 381 * account for the buffer and to wakeup anyone waiting for free buffers. 382 * This typically occurs when large amounts of metadata are being handled 383 * by the buffer cache ( else buffer space runs out first, usually ). 384 */ 385 386static __inline void 387bufcountwakeup(void) 388{ 389 390 atomic_add_int(&numfreebuffers, 1); 391 mtx_lock(&nblock); 392 if (needsbuffer) { 393 needsbuffer &= ~VFS_BIO_NEED_ANY; 394 if (numfreebuffers >= hifreebuffers) 395 needsbuffer &= ~VFS_BIO_NEED_FREE; 396 wakeup(&needsbuffer); 397 } 398 mtx_unlock(&nblock); 399} 400 401/* 402 * waitrunningbufspace() 403 * 404 * runningbufspace is a measure of the amount of I/O currently 405 * running. This routine is used in async-write situations to 406 * prevent creating huge backups of pending writes to a device. 407 * Only asynchronous writes are governed by this function. 408 * 409 * Reads will adjust runningbufspace, but will not block based on it. 410 * The read load has a side effect of reducing the allowed write load. 411 * 412 * This does NOT turn an async write into a sync write. It waits 413 * for earlier writes to complete and generally returns before the 414 * caller's write has reached the device. 415 */ 416void 417waitrunningbufspace(void) 418{ 419 420 mtx_lock(&rbreqlock); 421 while (runningbufspace > hirunningspace) { 422 ++runningbufreq; 423 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 424 } 425 mtx_unlock(&rbreqlock); 426} 427 428 429/* 430 * vfs_buf_test_cache: 431 * 432 * Called when a buffer is extended. This function clears the B_CACHE 433 * bit if the newly extended portion of the buffer does not contain 434 * valid data. 435 */ 436static __inline 437void 438vfs_buf_test_cache(struct buf *bp, 439 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 440 vm_page_t m) 441{ 442 443 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 444 if (bp->b_flags & B_CACHE) { 445 int base = (foff + off) & PAGE_MASK; 446 if (vm_page_is_valid(m, base, size) == 0) 447 bp->b_flags &= ~B_CACHE; 448 } 449} 450 451/* Wake up the buffer daemon if necessary */ 452static __inline 453void 454bd_wakeup(int dirtybuflevel) 455{ 456 457 mtx_lock(&bdlock); 458 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 459 bd_request = 1; 460 wakeup(&bd_request); 461 } 462 mtx_unlock(&bdlock); 463} 464 465/* 466 * bd_speedup - speedup the buffer cache flushing code 467 */ 468 469static __inline 470void 471bd_speedup(void) 472{ 473 474 bd_wakeup(1); 475} 476 477/* 478 * Calculating buffer cache scaling values and reserve space for buffer 479 * headers. This is called during low level kernel initialization and 480 * may be called more then once. We CANNOT write to the memory area 481 * being reserved at this time. 482 */ 483caddr_t 484kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 485{ 486 int tuned_nbuf; 487 long maxbuf; 488 489 /* 490 * physmem_est is in pages. Convert it to kilobytes (assumes 491 * PAGE_SIZE is >= 1K) 492 */ 493 physmem_est = physmem_est * (PAGE_SIZE / 1024); 494 495 /* 496 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 497 * For the first 64MB of ram nominally allocate sufficient buffers to 498 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 499 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 500 * the buffer cache we limit the eventual kva reservation to 501 * maxbcache bytes. 502 * 503 * factor represents the 1/4 x ram conversion. 504 */ 505 if (nbuf == 0) { 506 int factor = 4 * BKVASIZE / 1024; 507 508 nbuf = 50; 509 if (physmem_est > 4096) 510 nbuf += min((physmem_est - 4096) / factor, 511 65536 / factor); 512 if (physmem_est > 65536) 513 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 514 515 if (maxbcache && nbuf > maxbcache / BKVASIZE) 516 nbuf = maxbcache / BKVASIZE; 517 tuned_nbuf = 1; 518 } else 519 tuned_nbuf = 0; 520 521 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 522 maxbuf = (LONG_MAX / 3) / BKVASIZE; 523 if (nbuf > maxbuf) { 524 if (!tuned_nbuf) 525 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 526 maxbuf); 527 nbuf = maxbuf; 528 } 529 530 /* 531 * swbufs are used as temporary holders for I/O, such as paging I/O. 532 * We have no less then 16 and no more then 256. 533 */ 534 nswbuf = max(min(nbuf/4, 256), 16); 535#ifdef NSWBUF_MIN 536 if (nswbuf < NSWBUF_MIN) 537 nswbuf = NSWBUF_MIN; 538#endif 539#ifdef DIRECTIO 540 ffs_rawread_setup(); 541#endif 542 543 /* 544 * Reserve space for the buffer cache buffers 545 */ 546 swbuf = (void *)v; 547 v = (caddr_t)(swbuf + nswbuf); 548 buf = (void *)v; 549 v = (caddr_t)(buf + nbuf); 550 551 return(v); 552} 553 554/* Initialize the buffer subsystem. Called before use of any buffers. */ 555void 556bufinit(void) 557{ 558 struct buf *bp; 559 int i; 560 561 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 562 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 563 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 564 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 565 566 /* next, make a null set of free lists */ 567 for (i = 0; i < BUFFER_QUEUES; i++) 568 TAILQ_INIT(&bufqueues[i]); 569 570 /* finally, initialize each buffer header and stick on empty q */ 571 for (i = 0; i < nbuf; i++) { 572 bp = &buf[i]; 573 bzero(bp, sizeof *bp); 574 bp->b_flags = B_INVAL; /* we're just an empty header */ 575 bp->b_rcred = NOCRED; 576 bp->b_wcred = NOCRED; 577 bp->b_qindex = QUEUE_EMPTY; 578 bp->b_vflags = 0; 579 bp->b_xflags = 0; 580 LIST_INIT(&bp->b_dep); 581 BUF_LOCKINIT(bp); 582 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 583 } 584 585 /* 586 * maxbufspace is the absolute maximum amount of buffer space we are 587 * allowed to reserve in KVM and in real terms. The absolute maximum 588 * is nominally used by buf_daemon. hibufspace is the nominal maximum 589 * used by most other processes. The differential is required to 590 * ensure that buf_daemon is able to run when other processes might 591 * be blocked waiting for buffer space. 592 * 593 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 594 * this may result in KVM fragmentation which is not handled optimally 595 * by the system. 596 */ 597 maxbufspace = (long)nbuf * BKVASIZE; 598 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 599 lobufspace = hibufspace - MAXBSIZE; 600 601 lorunningspace = 512 * 1024; 602 hirunningspace = 1024 * 1024; 603 604/* 605 * Limit the amount of malloc memory since it is wired permanently into 606 * the kernel space. Even though this is accounted for in the buffer 607 * allocation, we don't want the malloced region to grow uncontrolled. 608 * The malloc scheme improves memory utilization significantly on average 609 * (small) directories. 610 */ 611 maxbufmallocspace = hibufspace / 20; 612 613/* 614 * Reduce the chance of a deadlock occuring by limiting the number 615 * of delayed-write dirty buffers we allow to stack up. 616 */ 617 hidirtybuffers = nbuf / 4 + 20; 618 dirtybufthresh = hidirtybuffers * 9 / 10; 619 numdirtybuffers = 0; 620/* 621 * To support extreme low-memory systems, make sure hidirtybuffers cannot 622 * eat up all available buffer space. This occurs when our minimum cannot 623 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 624 * BKVASIZE'd (8K) buffers. 625 */ 626 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 627 hidirtybuffers >>= 1; 628 } 629 lodirtybuffers = hidirtybuffers / 2; 630 631/* 632 * Try to keep the number of free buffers in the specified range, 633 * and give special processes (e.g. like buf_daemon) access to an 634 * emergency reserve. 635 */ 636 lofreebuffers = nbuf / 18 + 5; 637 hifreebuffers = 2 * lofreebuffers; 638 numfreebuffers = nbuf; 639 640/* 641 * Maximum number of async ops initiated per buf_daemon loop. This is 642 * somewhat of a hack at the moment, we really need to limit ourselves 643 * based on the number of bytes of I/O in-transit that were initiated 644 * from buf_daemon. 645 */ 646 647 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 648 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 649} 650 651/* 652 * bfreekva() - free the kva allocation for a buffer. 653 * 654 * Since this call frees up buffer space, we call bufspacewakeup(). 655 */ 656static void 657bfreekva(struct buf *bp) 658{ 659 660 if (bp->b_kvasize) { 661 atomic_add_int(&buffreekvacnt, 1); 662 atomic_subtract_long(&bufspace, bp->b_kvasize); 663 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 664 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 665 bp->b_kvasize = 0; 666 bufspacewakeup(); 667 } 668} 669 670/* 671 * bremfree: 672 * 673 * Mark the buffer for removal from the appropriate free list in brelse. 674 * 675 */ 676void 677bremfree(struct buf *bp) 678{ 679 680 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 681 KASSERT((bp->b_flags & B_REMFREE) == 0, 682 ("bremfree: buffer %p already marked for delayed removal.", bp)); 683 KASSERT(bp->b_qindex != QUEUE_NONE, 684 ("bremfree: buffer %p not on a queue.", bp)); 685 BUF_ASSERT_HELD(bp); 686 687 bp->b_flags |= B_REMFREE; 688 /* Fixup numfreebuffers count. */ 689 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 690 atomic_subtract_int(&numfreebuffers, 1); 691} 692 693/* 694 * bremfreef: 695 * 696 * Force an immediate removal from a free list. Used only in nfs when 697 * it abuses the b_freelist pointer. 698 */ 699void 700bremfreef(struct buf *bp) 701{ 702 mtx_lock(&bqlock); 703 bremfreel(bp); 704 mtx_unlock(&bqlock); 705} 706 707/* 708 * bremfreel: 709 * 710 * Removes a buffer from the free list, must be called with the 711 * bqlock held. 712 */ 713static void 714bremfreel(struct buf *bp) 715{ 716 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 717 bp, bp->b_vp, bp->b_flags); 718 KASSERT(bp->b_qindex != QUEUE_NONE, 719 ("bremfreel: buffer %p not on a queue.", bp)); 720 BUF_ASSERT_HELD(bp); 721 mtx_assert(&bqlock, MA_OWNED); 722 723 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 724 bp->b_qindex = QUEUE_NONE; 725 /* 726 * If this was a delayed bremfree() we only need to remove the buffer 727 * from the queue and return the stats are already done. 728 */ 729 if (bp->b_flags & B_REMFREE) { 730 bp->b_flags &= ~B_REMFREE; 731 return; 732 } 733 /* 734 * Fixup numfreebuffers count. If the buffer is invalid or not 735 * delayed-write, the buffer was free and we must decrement 736 * numfreebuffers. 737 */ 738 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 739 atomic_subtract_int(&numfreebuffers, 1); 740} 741 742 743/* 744 * Get a buffer with the specified data. Look in the cache first. We 745 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 746 * is set, the buffer is valid and we do not have to do anything ( see 747 * getblk() ). This is really just a special case of breadn(). 748 */ 749int 750bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 751 struct buf **bpp) 752{ 753 754 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 755} 756 757/* 758 * Same as bread, but also checks that cached indirect block 759 * buffers were not tampered with. 760 */ 761int 762breadi(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 763 struct buf **bpp) 764{ 765 struct buf *bp; 766 int rv = 0; 767 768 CTR3(KTR_BUF, "breadi(%p, %jd, %d)", vp, blkno, size); 769 *bpp = bp = getblk(vp, blkno, size, 0, 0, GB_INDIR); 770 771 /* if not found in cache, do some I/O */ 772 if ((bp->b_flags & B_CACHE) == 0) { 773 if (!TD_IS_IDLETHREAD(curthread)) 774 curthread->td_ru.ru_inblock++; 775 bp->b_iocmd = BIO_READ; 776 bp->b_flags &= ~B_INVAL; 777 bp->b_ioflags &= ~BIO_ERROR; 778 if (bp->b_rcred == NOCRED && cred != NOCRED) 779 bp->b_rcred = crhold(cred); 780 vfs_busy_pages(bp, 0); 781 bp->b_iooffset = dbtob(bp->b_blkno); 782 bstrategy(bp); 783 rv = bufwait(bp); 784 } 785 786 return (rv); 787} 788 789/* 790 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 791 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 792 * the buffer is valid and we do not have to do anything. 793 */ 794void 795breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 796 int cnt, struct ucred * cred) 797{ 798 struct buf *rabp; 799 int i; 800 801 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 802 if (inmem(vp, *rablkno)) 803 continue; 804 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 805 806 if ((rabp->b_flags & B_CACHE) == 0) { 807 if (!TD_IS_IDLETHREAD(curthread)) 808 curthread->td_ru.ru_inblock++; 809 rabp->b_flags |= B_ASYNC; 810 rabp->b_flags &= ~B_INVAL; 811 rabp->b_ioflags &= ~BIO_ERROR; 812 rabp->b_iocmd = BIO_READ; 813 if (rabp->b_rcred == NOCRED && cred != NOCRED) 814 rabp->b_rcred = crhold(cred); 815 vfs_busy_pages(rabp, 0); 816 BUF_KERNPROC(rabp); 817 rabp->b_iooffset = dbtob(rabp->b_blkno); 818 bstrategy(rabp); 819 } else { 820 brelse(rabp); 821 } 822 } 823} 824 825/* 826 * Operates like bread, but also starts asynchronous I/O on 827 * read-ahead blocks. 828 */ 829int 830breadn(struct vnode * vp, daddr_t blkno, int size, 831 daddr_t * rablkno, int *rabsize, 832 int cnt, struct ucred * cred, struct buf **bpp) 833{ 834 struct buf *bp; 835 int rv = 0, readwait = 0; 836 837 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 838 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 839 840 /* if not found in cache, do some I/O */ 841 if ((bp->b_flags & B_CACHE) == 0) { 842 if (!TD_IS_IDLETHREAD(curthread)) 843 curthread->td_ru.ru_inblock++; 844 bp->b_iocmd = BIO_READ; 845 bp->b_flags &= ~B_INVAL; 846 bp->b_ioflags &= ~BIO_ERROR; 847 if (bp->b_rcred == NOCRED && cred != NOCRED) 848 bp->b_rcred = crhold(cred); 849 vfs_busy_pages(bp, 0); 850 bp->b_iooffset = dbtob(bp->b_blkno); 851 bstrategy(bp); 852 ++readwait; 853 } 854 855 breada(vp, rablkno, rabsize, cnt, cred); 856 857 if (readwait) { 858 rv = bufwait(bp); 859 } 860 return (rv); 861} 862 863/* 864 * Write, release buffer on completion. (Done by iodone 865 * if async). Do not bother writing anything if the buffer 866 * is invalid. 867 * 868 * Note that we set B_CACHE here, indicating that buffer is 869 * fully valid and thus cacheable. This is true even of NFS 870 * now so we set it generally. This could be set either here 871 * or in biodone() since the I/O is synchronous. We put it 872 * here. 873 */ 874int 875bufwrite(struct buf *bp) 876{ 877 int oldflags; 878 struct vnode *vp; 879 int vp_md; 880 881 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 882 if (bp->b_flags & B_INVAL) { 883 brelse(bp); 884 return (0); 885 } 886 887 oldflags = bp->b_flags; 888 889 BUF_ASSERT_HELD(bp); 890 891 if (bp->b_pin_count > 0) 892 bunpin_wait(bp); 893 894 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 895 ("FFS background buffer should not get here %p", bp)); 896 897 vp = bp->b_vp; 898 if (vp) 899 vp_md = vp->v_vflag & VV_MD; 900 else 901 vp_md = 0; 902 903 /* Mark the buffer clean */ 904 bundirty(bp); 905 906 bp->b_flags &= ~B_DONE; 907 bp->b_ioflags &= ~BIO_ERROR; 908 bp->b_flags |= B_CACHE; 909 bp->b_iocmd = BIO_WRITE; 910 911 bufobj_wref(bp->b_bufobj); 912 vfs_busy_pages(bp, 1); 913 914 /* 915 * Normal bwrites pipeline writes 916 */ 917 bp->b_runningbufspace = bp->b_bufsize; 918 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 919 920 if (!TD_IS_IDLETHREAD(curthread)) 921 curthread->td_ru.ru_oublock++; 922 if (oldflags & B_ASYNC) 923 BUF_KERNPROC(bp); 924 bp->b_iooffset = dbtob(bp->b_blkno); 925 bstrategy(bp); 926 927 if ((oldflags & B_ASYNC) == 0) { 928 int rtval = bufwait(bp); 929 brelse(bp); 930 return (rtval); 931 } else { 932 /* 933 * don't allow the async write to saturate the I/O 934 * system. We will not deadlock here because 935 * we are blocking waiting for I/O that is already in-progress 936 * to complete. We do not block here if it is the update 937 * or syncer daemon trying to clean up as that can lead 938 * to deadlock. 939 */ 940 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 941 waitrunningbufspace(); 942 } 943 944 return (0); 945} 946 947void 948bufbdflush(struct bufobj *bo, struct buf *bp) 949{ 950 struct buf *nbp; 951 952 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 953 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 954 altbufferflushes++; 955 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 956 BO_LOCK(bo); 957 /* 958 * Try to find a buffer to flush. 959 */ 960 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 961 if ((nbp->b_vflags & BV_BKGRDINPROG) || 962 BUF_LOCK(nbp, 963 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 964 continue; 965 if (bp == nbp) 966 panic("bdwrite: found ourselves"); 967 BO_UNLOCK(bo); 968 /* Don't countdeps with the bo lock held. */ 969 if (buf_countdeps(nbp, 0)) { 970 BO_LOCK(bo); 971 BUF_UNLOCK(nbp); 972 continue; 973 } 974 if (nbp->b_flags & B_CLUSTEROK) { 975 vfs_bio_awrite(nbp); 976 } else { 977 bremfree(nbp); 978 bawrite(nbp); 979 } 980 dirtybufferflushes++; 981 break; 982 } 983 if (nbp == NULL) 984 BO_UNLOCK(bo); 985 } 986} 987 988/* 989 * Delayed write. (Buffer is marked dirty). Do not bother writing 990 * anything if the buffer is marked invalid. 991 * 992 * Note that since the buffer must be completely valid, we can safely 993 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 994 * biodone() in order to prevent getblk from writing the buffer 995 * out synchronously. 996 */ 997void 998bdwrite(struct buf *bp) 999{ 1000 struct thread *td = curthread; 1001 struct vnode *vp; 1002 struct bufobj *bo; 1003 1004 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1005 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1006 BUF_ASSERT_HELD(bp); 1007 1008 if (bp->b_flags & B_INVAL) { 1009 brelse(bp); 1010 return; 1011 } 1012 1013 /* 1014 * If we have too many dirty buffers, don't create any more. 1015 * If we are wildly over our limit, then force a complete 1016 * cleanup. Otherwise, just keep the situation from getting 1017 * out of control. Note that we have to avoid a recursive 1018 * disaster and not try to clean up after our own cleanup! 1019 */ 1020 vp = bp->b_vp; 1021 bo = bp->b_bufobj; 1022 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1023 td->td_pflags |= TDP_INBDFLUSH; 1024 BO_BDFLUSH(bo, bp); 1025 td->td_pflags &= ~TDP_INBDFLUSH; 1026 } else 1027 recursiveflushes++; 1028 1029 bdirty(bp); 1030 /* 1031 * Set B_CACHE, indicating that the buffer is fully valid. This is 1032 * true even of NFS now. 1033 */ 1034 bp->b_flags |= B_CACHE; 1035 1036 /* 1037 * This bmap keeps the system from needing to do the bmap later, 1038 * perhaps when the system is attempting to do a sync. Since it 1039 * is likely that the indirect block -- or whatever other datastructure 1040 * that the filesystem needs is still in memory now, it is a good 1041 * thing to do this. Note also, that if the pageout daemon is 1042 * requesting a sync -- there might not be enough memory to do 1043 * the bmap then... So, this is important to do. 1044 */ 1045 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1046 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1047 } 1048 1049 /* 1050 * Set the *dirty* buffer range based upon the VM system dirty pages. 1051 */ 1052 vfs_setdirty(bp); 1053 1054 /* 1055 * We need to do this here to satisfy the vnode_pager and the 1056 * pageout daemon, so that it thinks that the pages have been 1057 * "cleaned". Note that since the pages are in a delayed write 1058 * buffer -- the VFS layer "will" see that the pages get written 1059 * out on the next sync, or perhaps the cluster will be completed. 1060 */ 1061 vfs_clean_pages(bp); 1062 bqrelse(bp); 1063 1064 /* 1065 * Wakeup the buffer flushing daemon if we have a lot of dirty 1066 * buffers (midpoint between our recovery point and our stall 1067 * point). 1068 */ 1069 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1070 1071 /* 1072 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1073 * due to the softdep code. 1074 */ 1075} 1076 1077/* 1078 * bdirty: 1079 * 1080 * Turn buffer into delayed write request. We must clear BIO_READ and 1081 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1082 * itself to properly update it in the dirty/clean lists. We mark it 1083 * B_DONE to ensure that any asynchronization of the buffer properly 1084 * clears B_DONE ( else a panic will occur later ). 1085 * 1086 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1087 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1088 * should only be called if the buffer is known-good. 1089 * 1090 * Since the buffer is not on a queue, we do not update the numfreebuffers 1091 * count. 1092 * 1093 * The buffer must be on QUEUE_NONE. 1094 */ 1095void 1096bdirty(struct buf *bp) 1097{ 1098 1099 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1100 bp, bp->b_vp, bp->b_flags); 1101 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1102 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1103 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1104 BUF_ASSERT_HELD(bp); 1105 bp->b_flags &= ~(B_RELBUF); 1106 bp->b_iocmd = BIO_WRITE; 1107 1108 if ((bp->b_flags & B_DELWRI) == 0) { 1109 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1110 reassignbuf(bp); 1111 atomic_add_int(&numdirtybuffers, 1); 1112 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1113 } 1114} 1115 1116/* 1117 * bundirty: 1118 * 1119 * Clear B_DELWRI for buffer. 1120 * 1121 * Since the buffer is not on a queue, we do not update the numfreebuffers 1122 * count. 1123 * 1124 * The buffer must be on QUEUE_NONE. 1125 */ 1126 1127void 1128bundirty(struct buf *bp) 1129{ 1130 1131 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1132 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1133 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1134 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1135 BUF_ASSERT_HELD(bp); 1136 1137 if (bp->b_flags & B_DELWRI) { 1138 bp->b_flags &= ~B_DELWRI; 1139 reassignbuf(bp); 1140 atomic_subtract_int(&numdirtybuffers, 1); 1141 numdirtywakeup(lodirtybuffers); 1142 } 1143 /* 1144 * Since it is now being written, we can clear its deferred write flag. 1145 */ 1146 bp->b_flags &= ~B_DEFERRED; 1147} 1148 1149/* 1150 * bawrite: 1151 * 1152 * Asynchronous write. Start output on a buffer, but do not wait for 1153 * it to complete. The buffer is released when the output completes. 1154 * 1155 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1156 * B_INVAL buffers. Not us. 1157 */ 1158void 1159bawrite(struct buf *bp) 1160{ 1161 1162 bp->b_flags |= B_ASYNC; 1163 (void) bwrite(bp); 1164} 1165 1166/* 1167 * bwillwrite: 1168 * 1169 * Called prior to the locking of any vnodes when we are expecting to 1170 * write. We do not want to starve the buffer cache with too many 1171 * dirty buffers so we block here. By blocking prior to the locking 1172 * of any vnodes we attempt to avoid the situation where a locked vnode 1173 * prevents the various system daemons from flushing related buffers. 1174 */ 1175 1176void 1177bwillwrite(void) 1178{ 1179 1180 if (numdirtybuffers >= hidirtybuffers) { 1181 mtx_lock(&nblock); 1182 while (numdirtybuffers >= hidirtybuffers) { 1183 bd_wakeup(1); 1184 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1185 msleep(&needsbuffer, &nblock, 1186 (PRIBIO + 4), "flswai", 0); 1187 } 1188 mtx_unlock(&nblock); 1189 } 1190} 1191 1192/* 1193 * Return true if we have too many dirty buffers. 1194 */ 1195int 1196buf_dirty_count_severe(void) 1197{ 1198 1199 return(numdirtybuffers >= hidirtybuffers); 1200} 1201 1202/* 1203 * brelse: 1204 * 1205 * Release a busy buffer and, if requested, free its resources. The 1206 * buffer will be stashed in the appropriate bufqueue[] allowing it 1207 * to be accessed later as a cache entity or reused for other purposes. 1208 */ 1209void 1210brelse(struct buf *bp) 1211{ 1212 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1213 bp, bp->b_vp, bp->b_flags); 1214 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1215 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1216 1217 if (bp->b_flags & B_MANAGED) { 1218 bqrelse(bp); 1219 return; 1220 } 1221 1222 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1223 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1224 /* 1225 * Failed write, redirty. Must clear BIO_ERROR to prevent 1226 * pages from being scrapped. If the error is anything 1227 * other than an I/O error (EIO), assume that retryingi 1228 * is futile. 1229 */ 1230 bp->b_ioflags &= ~BIO_ERROR; 1231 bdirty(bp); 1232 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1233 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1234 /* 1235 * Either a failed I/O or we were asked to free or not 1236 * cache the buffer. 1237 */ 1238 bp->b_flags |= B_INVAL; 1239 if (!LIST_EMPTY(&bp->b_dep)) 1240 buf_deallocate(bp); 1241 if (bp->b_flags & B_DELWRI) { 1242 atomic_subtract_int(&numdirtybuffers, 1); 1243 numdirtywakeup(lodirtybuffers); 1244 } 1245 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1246 if ((bp->b_flags & B_VMIO) == 0) { 1247 if (bp->b_bufsize) 1248 allocbuf(bp, 0); 1249 if (bp->b_vp) 1250 brelvp(bp); 1251 } 1252 } 1253 1254 /* Validate indirect buffer before VMIO buffer rundown. */ 1255 if ((bp->b_flags & B_INDIR) && !(bp->b_flags & B_INVAL)) 1256 BO_INDIR_CHECK(bp->b_bufobj, bp); 1257 1258 /* 1259 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1260 * is called with B_DELWRI set, the underlying pages may wind up 1261 * getting freed causing a previous write (bdwrite()) to get 'lost' 1262 * because pages associated with a B_DELWRI bp are marked clean. 1263 * 1264 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1265 * if B_DELWRI is set. 1266 * 1267 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1268 * on pages to return pages to the VM page queues. 1269 */ 1270 if (bp->b_flags & B_DELWRI) 1271 bp->b_flags &= ~B_RELBUF; 1272 else if (vm_page_count_severe()) { 1273 /* 1274 * The locking of the BO_LOCK is not necessary since 1275 * BKGRDINPROG cannot be set while we hold the buf 1276 * lock, it can only be cleared if it is already 1277 * pending. 1278 */ 1279 if (bp->b_vp) { 1280 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1281 bp->b_flags |= B_RELBUF; 1282 } else 1283 bp->b_flags |= B_RELBUF; 1284 } 1285 1286 /* 1287 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1288 * constituted, not even NFS buffers now. Two flags effect this. If 1289 * B_INVAL, the struct buf is invalidated but the VM object is kept 1290 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1291 * 1292 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1293 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1294 * buffer is also B_INVAL because it hits the re-dirtying code above. 1295 * 1296 * Normally we can do this whether a buffer is B_DELWRI or not. If 1297 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1298 * the commit state and we cannot afford to lose the buffer. If the 1299 * buffer has a background write in progress, we need to keep it 1300 * around to prevent it from being reconstituted and starting a second 1301 * background write. 1302 */ 1303 if ((bp->b_flags & B_VMIO) 1304 && !(bp->b_vp->v_mount != NULL && 1305 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1306 !vn_isdisk(bp->b_vp, NULL) && 1307 (bp->b_flags & B_DELWRI)) 1308 ) { 1309 1310 int i, j, resid; 1311 vm_page_t m; 1312 off_t foff; 1313 vm_pindex_t poff; 1314 vm_object_t obj; 1315 1316 obj = bp->b_bufobj->bo_object; 1317 1318 /* 1319 * Get the base offset and length of the buffer. Note that 1320 * in the VMIO case if the buffer block size is not 1321 * page-aligned then b_data pointer may not be page-aligned. 1322 * But our b_pages[] array *IS* page aligned. 1323 * 1324 * block sizes less then DEV_BSIZE (usually 512) are not 1325 * supported due to the page granularity bits (m->valid, 1326 * m->dirty, etc...). 1327 * 1328 * See man buf(9) for more information 1329 */ 1330 resid = bp->b_bufsize; 1331 foff = bp->b_offset; 1332 VM_OBJECT_LOCK(obj); 1333 for (i = 0; i < bp->b_npages; i++) { 1334 int had_bogus = 0; 1335 1336 m = bp->b_pages[i]; 1337 1338 /* 1339 * If we hit a bogus page, fixup *all* the bogus pages 1340 * now. 1341 */ 1342 if (m == bogus_page) { 1343 poff = OFF_TO_IDX(bp->b_offset); 1344 had_bogus = 1; 1345 1346 for (j = i; j < bp->b_npages; j++) { 1347 vm_page_t mtmp; 1348 mtmp = bp->b_pages[j]; 1349 if (mtmp == bogus_page) { 1350 mtmp = vm_page_lookup(obj, poff + j); 1351 if (!mtmp) { 1352 panic("brelse: page missing\n"); 1353 } 1354 bp->b_pages[j] = mtmp; 1355 } 1356 } 1357 1358 if ((bp->b_flags & B_INVAL) == 0) { 1359 pmap_qenter( 1360 trunc_page((vm_offset_t)bp->b_data), 1361 bp->b_pages, bp->b_npages); 1362 } 1363 m = bp->b_pages[i]; 1364 } 1365 if ((bp->b_flags & B_NOCACHE) || 1366 (bp->b_ioflags & BIO_ERROR)) { 1367 int poffset = foff & PAGE_MASK; 1368 int presid = resid > (PAGE_SIZE - poffset) ? 1369 (PAGE_SIZE - poffset) : resid; 1370 1371 KASSERT(presid >= 0, ("brelse: extra page")); 1372 vm_page_lock_queues(); 1373 vm_page_set_invalid(m, poffset, presid); 1374 vm_page_unlock_queues(); 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(); 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 1482 if ((bp->b_flags & B_INDIR) && !(bp->b_flags & B_INVAL)) 1483 BO_INDIR_CHECK(bp->b_bufobj, bp); 1484 /* unlock */ 1485 BUF_UNLOCK(bp); 1486} 1487 1488/* 1489 * Release a buffer back to the appropriate queue but do not try to free 1490 * it. The buffer is expected to be used again soon. 1491 * 1492 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1493 * biodone() to requeue an async I/O on completion. It is also used when 1494 * known good buffers need to be requeued but we think we may need the data 1495 * again soon. 1496 * 1497 * XXX we should be able to leave the B_RELBUF hint set on completion. 1498 */ 1499void 1500bqrelse(struct buf *bp) 1501{ 1502 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1503 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1504 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1505 1506 if (BUF_LOCKRECURSED(bp)) { 1507 /* do not release to free list */ 1508 BUF_UNLOCK(bp); 1509 return; 1510 } 1511 1512 if (bp->b_flags & B_MANAGED) { 1513 if (bp->b_flags & B_REMFREE) { 1514 mtx_lock(&bqlock); 1515 bremfreel(bp); 1516 mtx_unlock(&bqlock); 1517 } 1518 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1519 BUF_UNLOCK(bp); 1520 return; 1521 } 1522 1523 mtx_lock(&bqlock); 1524 /* Handle delayed bremfree() processing. */ 1525 if (bp->b_flags & B_REMFREE) 1526 bremfreel(bp); 1527 if (bp->b_qindex != QUEUE_NONE) 1528 panic("bqrelse: free buffer onto another queue???"); 1529 /* buffers with stale but valid contents */ 1530 if (bp->b_flags & B_DELWRI) { 1531 if (bp->b_flags & B_NEEDSGIANT) 1532 bp->b_qindex = QUEUE_DIRTY_GIANT; 1533 else 1534 bp->b_qindex = QUEUE_DIRTY; 1535 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1536 } else { 1537 /* 1538 * The locking of the BO_LOCK for checking of the 1539 * BV_BKGRDINPROG is not necessary since the 1540 * BV_BKGRDINPROG cannot be set while we hold the buf 1541 * lock, it can only be cleared if it is already 1542 * pending. 1543 */ 1544 if (!vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1545 bp->b_qindex = QUEUE_CLEAN; 1546 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1547 b_freelist); 1548 } else { 1549 /* 1550 * We are too low on memory, we have to try to free 1551 * the buffer (most importantly: the wired pages 1552 * making up its backing store) *now*. 1553 */ 1554 mtx_unlock(&bqlock); 1555 brelse(bp); 1556 return; 1557 } 1558 } 1559 mtx_unlock(&bqlock); 1560 1561 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1562 bufcountwakeup(); 1563 1564 /* 1565 * Something we can maybe free or reuse. 1566 */ 1567 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1568 bufspacewakeup(); 1569 1570 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1571 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1572 panic("bqrelse: not dirty"); 1573 1574 if ((bp->b_flags & B_INDIR) && !(bp->b_flags & B_INVAL)) 1575 BO_INDIR_CHECK(bp->b_bufobj, bp); 1576 1577 /* unlock */ 1578 BUF_UNLOCK(bp); 1579} 1580 1581/* Give pages used by the bp back to the VM system (where possible) */ 1582static void 1583vfs_vmio_release(struct buf *bp) 1584{ 1585 int i; 1586 vm_page_t m; 1587 1588 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1589 vm_page_lock_queues(); 1590 for (i = 0; i < bp->b_npages; i++) { 1591 m = bp->b_pages[i]; 1592 bp->b_pages[i] = NULL; 1593 /* 1594 * In order to keep page LRU ordering consistent, put 1595 * everything on the inactive queue. 1596 */ 1597 vm_page_unwire(m, 0); 1598 /* 1599 * We don't mess with busy pages, it is 1600 * the responsibility of the process that 1601 * busied the pages to deal with them. 1602 */ 1603 if ((m->oflags & VPO_BUSY) || (m->busy != 0)) 1604 continue; 1605 1606 if (m->wire_count == 0) { 1607 /* 1608 * Might as well free the page if we can and it has 1609 * no valid data. We also free the page if the 1610 * buffer was used for direct I/O 1611 */ 1612 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1613 m->hold_count == 0) { 1614 vm_page_free(m); 1615 } else if (bp->b_flags & B_DIRECT) { 1616 vm_page_try_to_free(m); 1617 } else if (vm_page_count_severe()) { 1618 vm_page_try_to_cache(m); 1619 } 1620 } 1621 } 1622 vm_page_unlock_queues(); 1623 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1624 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1625 1626 if (bp->b_bufsize) { 1627 bufspacewakeup(); 1628 bp->b_bufsize = 0; 1629 } 1630 bp->b_npages = 0; 1631 bp->b_flags &= ~B_VMIO; 1632 if (bp->b_vp) 1633 brelvp(bp); 1634} 1635 1636/* 1637 * Check to see if a block at a particular lbn is available for a clustered 1638 * write. 1639 */ 1640static int 1641vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1642{ 1643 struct buf *bpa; 1644 int match; 1645 1646 match = 0; 1647 1648 /* If the buf isn't in core skip it */ 1649 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1650 return (0); 1651 1652 /* If the buf is busy we don't want to wait for it */ 1653 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1654 return (0); 1655 1656 /* Only cluster with valid clusterable delayed write buffers */ 1657 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1658 (B_DELWRI | B_CLUSTEROK)) 1659 goto done; 1660 1661 if (bpa->b_bufsize != size) 1662 goto done; 1663 1664 /* 1665 * Check to see if it is in the expected place on disk and that the 1666 * block has been mapped. 1667 */ 1668 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1669 match = 1; 1670done: 1671 BUF_UNLOCK(bpa); 1672 return (match); 1673} 1674 1675/* 1676 * vfs_bio_awrite: 1677 * 1678 * Implement clustered async writes for clearing out B_DELWRI buffers. 1679 * This is much better then the old way of writing only one buffer at 1680 * a time. Note that we may not be presented with the buffers in the 1681 * correct order, so we search for the cluster in both directions. 1682 */ 1683int 1684vfs_bio_awrite(struct buf *bp) 1685{ 1686 struct bufobj *bo; 1687 int i; 1688 int j; 1689 daddr_t lblkno = bp->b_lblkno; 1690 struct vnode *vp = bp->b_vp; 1691 int ncl; 1692 int nwritten; 1693 int size; 1694 int maxcl; 1695 1696 bo = &vp->v_bufobj; 1697 /* 1698 * right now we support clustered writing only to regular files. If 1699 * we find a clusterable block we could be in the middle of a cluster 1700 * rather then at the beginning. 1701 */ 1702 if ((vp->v_type == VREG) && 1703 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1704 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1705 1706 size = vp->v_mount->mnt_stat.f_iosize; 1707 maxcl = MAXPHYS / size; 1708 1709 BO_LOCK(bo); 1710 for (i = 1; i < maxcl; i++) 1711 if (vfs_bio_clcheck(vp, size, lblkno + i, 1712 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1713 break; 1714 1715 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1716 if (vfs_bio_clcheck(vp, size, lblkno - j, 1717 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1718 break; 1719 BO_UNLOCK(bo); 1720 --j; 1721 ncl = i + j; 1722 /* 1723 * this is a possible cluster write 1724 */ 1725 if (ncl != 1) { 1726 BUF_UNLOCK(bp); 1727 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1728 return nwritten; 1729 } 1730 } 1731 bremfree(bp); 1732 bp->b_flags |= B_ASYNC; 1733 /* 1734 * default (old) behavior, writing out only one block 1735 * 1736 * XXX returns b_bufsize instead of b_bcount for nwritten? 1737 */ 1738 nwritten = bp->b_bufsize; 1739 (void) bwrite(bp); 1740 1741 return nwritten; 1742} 1743 1744/* 1745 * getnewbuf: 1746 * 1747 * Find and initialize a new buffer header, freeing up existing buffers 1748 * in the bufqueues as necessary. The new buffer is returned locked. 1749 * 1750 * Important: B_INVAL is not set. If the caller wishes to throw the 1751 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1752 * 1753 * We block if: 1754 * We have insufficient buffer headers 1755 * We have insufficient buffer space 1756 * buffer_map is too fragmented ( space reservation fails ) 1757 * If we have to flush dirty buffers ( but we try to avoid this ) 1758 * 1759 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1760 * Instead we ask the buf daemon to do it for us. We attempt to 1761 * avoid piecemeal wakeups of the pageout daemon. 1762 */ 1763 1764static struct buf * 1765getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 1766 int gbflags) 1767{ 1768 struct thread *td; 1769 struct buf *bp; 1770 struct buf *nbp; 1771 int defrag = 0; 1772 int nqindex; 1773 static int flushingbufs; 1774 1775 td = curthread; 1776 /* 1777 * We can't afford to block since we might be holding a vnode lock, 1778 * which may prevent system daemons from running. We deal with 1779 * low-memory situations by proactively returning memory and running 1780 * async I/O rather then sync I/O. 1781 */ 1782 atomic_add_int(&getnewbufcalls, 1); 1783 atomic_subtract_int(&getnewbufrestarts, 1); 1784restart: 1785 atomic_add_int(&getnewbufrestarts, 1); 1786 1787 /* 1788 * Setup for scan. If we do not have enough free buffers, 1789 * we setup a degenerate case that immediately fails. Note 1790 * that if we are specially marked process, we are allowed to 1791 * dip into our reserves. 1792 * 1793 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1794 * 1795 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1796 * However, there are a number of cases (defragging, reusing, ...) 1797 * where we cannot backup. 1798 */ 1799 mtx_lock(&bqlock); 1800 nqindex = QUEUE_EMPTYKVA; 1801 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1802 1803 if (nbp == NULL) { 1804 /* 1805 * If no EMPTYKVA buffers and we are either 1806 * defragging or reusing, locate a CLEAN buffer 1807 * to free or reuse. If bufspace useage is low 1808 * skip this step so we can allocate a new buffer. 1809 */ 1810 if (defrag || bufspace >= lobufspace) { 1811 nqindex = QUEUE_CLEAN; 1812 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1813 } 1814 1815 /* 1816 * If we could not find or were not allowed to reuse a 1817 * CLEAN buffer, check to see if it is ok to use an EMPTY 1818 * buffer. We can only use an EMPTY buffer if allocating 1819 * its KVA would not otherwise run us out of buffer space. 1820 */ 1821 if (nbp == NULL && defrag == 0 && 1822 bufspace + maxsize < hibufspace) { 1823 nqindex = QUEUE_EMPTY; 1824 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1825 } 1826 } 1827 1828 /* 1829 * Run scan, possibly freeing data and/or kva mappings on the fly 1830 * depending. 1831 */ 1832 1833 while ((bp = nbp) != NULL) { 1834 int qindex = nqindex; 1835 1836 /* 1837 * Calculate next bp ( we can only use it if we do not block 1838 * or do other fancy things ). 1839 */ 1840 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1841 switch(qindex) { 1842 case QUEUE_EMPTY: 1843 nqindex = QUEUE_EMPTYKVA; 1844 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1845 break; 1846 /* FALLTHROUGH */ 1847 case QUEUE_EMPTYKVA: 1848 nqindex = QUEUE_CLEAN; 1849 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1850 break; 1851 /* FALLTHROUGH */ 1852 case QUEUE_CLEAN: 1853 /* 1854 * nbp is NULL. 1855 */ 1856 break; 1857 } 1858 } 1859 /* 1860 * If we are defragging then we need a buffer with 1861 * b_kvasize != 0. XXX this situation should no longer 1862 * occur, if defrag is non-zero the buffer's b_kvasize 1863 * should also be non-zero at this point. XXX 1864 */ 1865 if (defrag && bp->b_kvasize == 0) { 1866 printf("Warning: defrag empty buffer %p\n", bp); 1867 continue; 1868 } 1869 1870 /* 1871 * Start freeing the bp. This is somewhat involved. nbp 1872 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1873 */ 1874 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1875 continue; 1876 if (bp->b_vp) { 1877 BO_LOCK(bp->b_bufobj); 1878 if (bp->b_vflags & BV_BKGRDINPROG) { 1879 BO_UNLOCK(bp->b_bufobj); 1880 BUF_UNLOCK(bp); 1881 continue; 1882 } 1883 BO_UNLOCK(bp->b_bufobj); 1884 } 1885 CTR6(KTR_BUF, 1886 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1887 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1888 bp->b_kvasize, bp->b_bufsize, qindex); 1889 1890 /* 1891 * Sanity Checks 1892 */ 1893 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1894 1895 /* 1896 * Note: we no longer distinguish between VMIO and non-VMIO 1897 * buffers. 1898 */ 1899 1900 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1901 1902 bremfreel(bp); 1903 mtx_unlock(&bqlock); 1904 1905 if (qindex == QUEUE_CLEAN) { 1906 if (bp->b_flags & B_VMIO) { 1907 bp->b_flags &= ~B_ASYNC; 1908 vfs_vmio_release(bp); 1909 } 1910 if (bp->b_vp) 1911 brelvp(bp); 1912 } 1913 1914 /* 1915 * NOTE: nbp is now entirely invalid. We can only restart 1916 * the scan from this point on. 1917 * 1918 * Get the rest of the buffer freed up. b_kva* is still 1919 * valid after this operation. 1920 */ 1921 1922 if (bp->b_rcred != NOCRED) { 1923 crfree(bp->b_rcred); 1924 bp->b_rcred = NOCRED; 1925 } 1926 if (bp->b_wcred != NOCRED) { 1927 crfree(bp->b_wcred); 1928 bp->b_wcred = NOCRED; 1929 } 1930 if (!LIST_EMPTY(&bp->b_dep)) 1931 buf_deallocate(bp); 1932 if (bp->b_vflags & BV_BKGRDINPROG) 1933 panic("losing buffer 3"); 1934 KASSERT(bp->b_vp == NULL, 1935 ("bp: %p still has vnode %p. qindex: %d", 1936 bp, bp->b_vp, qindex)); 1937 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1938 ("bp: %p still on a buffer list. xflags %X", 1939 bp, bp->b_xflags)); 1940 1941 if (bp->b_bufsize) 1942 allocbuf(bp, 0); 1943 1944 bp->b_flags = 0; 1945 bp->b_ioflags = 0; 1946 bp->b_xflags = 0; 1947 bp->b_vflags = 0; 1948 bp->b_vp = NULL; 1949 bp->b_blkno = bp->b_lblkno = 0; 1950 bp->b_offset = NOOFFSET; 1951 bp->b_iodone = 0; 1952 bp->b_error = 0; 1953 bp->b_resid = 0; 1954 bp->b_bcount = 0; 1955 bp->b_npages = 0; 1956 bp->b_dirtyoff = bp->b_dirtyend = 0; 1957 bp->b_bufobj = NULL; 1958 bp->b_pin_count = 0; 1959 bp->b_fsprivate1 = NULL; 1960 bp->b_fsprivate2 = NULL; 1961 bp->b_fsprivate3 = NULL; 1962 1963 LIST_INIT(&bp->b_dep); 1964 1965 /* 1966 * If we are defragging then free the buffer. 1967 */ 1968 if (defrag) { 1969 bp->b_flags |= B_INVAL; 1970 bfreekva(bp); 1971 brelse(bp); 1972 defrag = 0; 1973 goto restart; 1974 } 1975 1976 /* 1977 * Notify any waiters for the buffer lock about 1978 * identity change by freeing the buffer. 1979 */ 1980 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 1981 bp->b_flags |= B_INVAL; 1982 bfreekva(bp); 1983 brelse(bp); 1984 goto restart; 1985 } 1986 1987 /* 1988 * If we are overcomitted then recover the buffer and its 1989 * KVM space. This occurs in rare situations when multiple 1990 * processes are blocked in getnewbuf() or allocbuf(). 1991 */ 1992 if (bufspace >= hibufspace) 1993 flushingbufs = 1; 1994 if (flushingbufs && bp->b_kvasize != 0) { 1995 bp->b_flags |= B_INVAL; 1996 bfreekva(bp); 1997 brelse(bp); 1998 goto restart; 1999 } 2000 if (bufspace < lobufspace) 2001 flushingbufs = 0; 2002 break; 2003 } 2004 2005 /* 2006 * If we exhausted our list, sleep as appropriate. We may have to 2007 * wakeup various daemons and write out some dirty buffers. 2008 * 2009 * Generally we are sleeping due to insufficient buffer space. 2010 */ 2011 2012 if (bp == NULL) { 2013 int flags, norunbuf; 2014 char *waitmsg; 2015 int fl; 2016 2017 if (defrag) { 2018 flags = VFS_BIO_NEED_BUFSPACE; 2019 waitmsg = "nbufkv"; 2020 } else if (bufspace >= hibufspace) { 2021 waitmsg = "nbufbs"; 2022 flags = VFS_BIO_NEED_BUFSPACE; 2023 } else { 2024 waitmsg = "newbuf"; 2025 flags = VFS_BIO_NEED_ANY; 2026 } 2027 mtx_lock(&nblock); 2028 needsbuffer |= flags; 2029 mtx_unlock(&nblock); 2030 mtx_unlock(&bqlock); 2031 2032 bd_speedup(); /* heeeelp */ 2033 if (gbflags & GB_NOWAIT_BD) 2034 return (NULL); 2035 2036 mtx_lock(&nblock); 2037 while (needsbuffer & flags) { 2038 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2039 mtx_unlock(&nblock); 2040 /* 2041 * getblk() is called with a vnode 2042 * locked, and some majority of the 2043 * dirty buffers may as well belong to 2044 * the vnode. Flushing the buffers 2045 * there would make a progress that 2046 * cannot be achieved by the 2047 * buf_daemon, that cannot lock the 2048 * vnode. 2049 */ 2050 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2051 (td->td_pflags & TDP_NORUNNINGBUF); 2052 /* play bufdaemon */ 2053 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2054 fl = buf_do_flush(vp); 2055 td->td_pflags &= norunbuf; 2056 mtx_lock(&nblock); 2057 if (fl != 0) 2058 continue; 2059 if ((needsbuffer & flags) == 0) 2060 break; 2061 } 2062 if (msleep(&needsbuffer, &nblock, 2063 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2064 mtx_unlock(&nblock); 2065 return (NULL); 2066 } 2067 } 2068 mtx_unlock(&nblock); 2069 } else { 2070 /* 2071 * We finally have a valid bp. We aren't quite out of the 2072 * woods, we still have to reserve kva space. In order 2073 * to keep fragmentation sane we only allocate kva in 2074 * BKVASIZE chunks. 2075 */ 2076 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2077 2078 if (maxsize != bp->b_kvasize) { 2079 vm_offset_t addr = 0; 2080 2081 bfreekva(bp); 2082 2083 vm_map_lock(buffer_map); 2084 if (vm_map_findspace(buffer_map, 2085 vm_map_min(buffer_map), maxsize, &addr)) { 2086 /* 2087 * Uh oh. Buffer map is to fragmented. We 2088 * must defragment the map. 2089 */ 2090 atomic_add_int(&bufdefragcnt, 1); 2091 vm_map_unlock(buffer_map); 2092 defrag = 1; 2093 bp->b_flags |= B_INVAL; 2094 brelse(bp); 2095 goto restart; 2096 } 2097 if (addr) { 2098 vm_map_insert(buffer_map, NULL, 0, 2099 addr, addr + maxsize, 2100 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2101 2102 bp->b_kvabase = (caddr_t) addr; 2103 bp->b_kvasize = maxsize; 2104 atomic_add_long(&bufspace, bp->b_kvasize); 2105 atomic_add_int(&bufreusecnt, 1); 2106 } 2107 vm_map_unlock(buffer_map); 2108 } 2109 bp->b_saveaddr = bp->b_kvabase; 2110 bp->b_data = bp->b_saveaddr; 2111 } 2112 return(bp); 2113} 2114 2115/* 2116 * buf_daemon: 2117 * 2118 * buffer flushing daemon. Buffers are normally flushed by the 2119 * update daemon but if it cannot keep up this process starts to 2120 * take the load in an attempt to prevent getnewbuf() from blocking. 2121 */ 2122 2123static struct kproc_desc buf_kp = { 2124 "bufdaemon", 2125 buf_daemon, 2126 &bufdaemonproc 2127}; 2128SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2129 2130static int 2131buf_do_flush(struct vnode *vp) 2132{ 2133 int flushed; 2134 2135 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2136 /* The list empty check here is slightly racy */ 2137 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2138 mtx_lock(&Giant); 2139 flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0); 2140 mtx_unlock(&Giant); 2141 } 2142 if (flushed == 0) { 2143 /* 2144 * Could not find any buffers without rollback 2145 * dependencies, so just write the first one 2146 * in the hopes of eventually making progress. 2147 */ 2148 flushbufqueues(vp, QUEUE_DIRTY, 1); 2149 if (!TAILQ_EMPTY( 2150 &bufqueues[QUEUE_DIRTY_GIANT])) { 2151 mtx_lock(&Giant); 2152 flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1); 2153 mtx_unlock(&Giant); 2154 } 2155 } 2156 return (flushed); 2157} 2158 2159static void 2160buf_daemon() 2161{ 2162 2163 /* 2164 * This process needs to be suspended prior to shutdown sync. 2165 */ 2166 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2167 SHUTDOWN_PRI_LAST); 2168 2169 /* 2170 * This process is allowed to take the buffer cache to the limit 2171 */ 2172 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2173 mtx_lock(&bdlock); 2174 for (;;) { 2175 bd_request = 0; 2176 mtx_unlock(&bdlock); 2177 2178 kproc_suspend_check(bufdaemonproc); 2179 2180 /* 2181 * Do the flush. Limit the amount of in-transit I/O we 2182 * allow to build up, otherwise we would completely saturate 2183 * the I/O system. Wakeup any waiting processes before we 2184 * normally would so they can run in parallel with our drain. 2185 */ 2186 while (numdirtybuffers > lodirtybuffers) { 2187 if (buf_do_flush(NULL) == 0) 2188 break; 2189 uio_yield(); 2190 } 2191 2192 /* 2193 * Only clear bd_request if we have reached our low water 2194 * mark. The buf_daemon normally waits 1 second and 2195 * then incrementally flushes any dirty buffers that have 2196 * built up, within reason. 2197 * 2198 * If we were unable to hit our low water mark and couldn't 2199 * find any flushable buffers, we sleep half a second. 2200 * Otherwise we loop immediately. 2201 */ 2202 mtx_lock(&bdlock); 2203 if (numdirtybuffers <= lodirtybuffers) { 2204 /* 2205 * We reached our low water mark, reset the 2206 * request and sleep until we are needed again. 2207 * The sleep is just so the suspend code works. 2208 */ 2209 bd_request = 0; 2210 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2211 } else { 2212 /* 2213 * We couldn't find any flushable dirty buffers but 2214 * still have too many dirty buffers, we 2215 * have to sleep and try again. (rare) 2216 */ 2217 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2218 } 2219 } 2220} 2221 2222/* 2223 * flushbufqueues: 2224 * 2225 * Try to flush a buffer in the dirty queue. We must be careful to 2226 * free up B_INVAL buffers instead of write them, which NFS is 2227 * particularly sensitive to. 2228 */ 2229static int flushwithdeps = 0; 2230SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2231 0, "Number of buffers flushed with dependecies that require rollbacks"); 2232 2233static int 2234flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2235{ 2236 struct buf *sentinel; 2237 struct vnode *vp; 2238 struct mount *mp; 2239 struct buf *bp; 2240 int hasdeps; 2241 int flushed; 2242 int target; 2243 2244 if (lvp == NULL) { 2245 target = numdirtybuffers - lodirtybuffers; 2246 if (flushdeps && target > 2) 2247 target /= 2; 2248 } else 2249 target = flushbufqtarget; 2250 flushed = 0; 2251 bp = NULL; 2252 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2253 sentinel->b_qindex = QUEUE_SENTINEL; 2254 mtx_lock(&bqlock); 2255 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2256 while (flushed != target) { 2257 bp = TAILQ_NEXT(sentinel, b_freelist); 2258 if (bp != NULL) { 2259 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2260 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2261 b_freelist); 2262 } else 2263 break; 2264 /* 2265 * Skip sentinels inserted by other invocations of the 2266 * flushbufqueues(), taking care to not reorder them. 2267 */ 2268 if (bp->b_qindex == QUEUE_SENTINEL) 2269 continue; 2270 /* 2271 * Only flush the buffers that belong to the 2272 * vnode locked by the curthread. 2273 */ 2274 if (lvp != NULL && bp->b_vp != lvp) 2275 continue; 2276 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2277 continue; 2278 if (bp->b_pin_count > 0) { 2279 BUF_UNLOCK(bp); 2280 continue; 2281 } 2282 BO_LOCK(bp->b_bufobj); 2283 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2284 (bp->b_flags & B_DELWRI) == 0) { 2285 BO_UNLOCK(bp->b_bufobj); 2286 BUF_UNLOCK(bp); 2287 continue; 2288 } 2289 BO_UNLOCK(bp->b_bufobj); 2290 if (bp->b_flags & B_INVAL) { 2291 bremfreel(bp); 2292 mtx_unlock(&bqlock); 2293 brelse(bp); 2294 flushed++; 2295 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2296 mtx_lock(&bqlock); 2297 continue; 2298 } 2299 2300 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2301 if (flushdeps == 0) { 2302 BUF_UNLOCK(bp); 2303 continue; 2304 } 2305 hasdeps = 1; 2306 } else 2307 hasdeps = 0; 2308 /* 2309 * We must hold the lock on a vnode before writing 2310 * one of its buffers. Otherwise we may confuse, or 2311 * in the case of a snapshot vnode, deadlock the 2312 * system. 2313 * 2314 * The lock order here is the reverse of the normal 2315 * of vnode followed by buf lock. This is ok because 2316 * the NOWAIT will prevent deadlock. 2317 */ 2318 vp = bp->b_vp; 2319 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2320 BUF_UNLOCK(bp); 2321 continue; 2322 } 2323 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2324 mtx_unlock(&bqlock); 2325 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2326 bp, bp->b_vp, bp->b_flags); 2327 if (curproc == bufdaemonproc) 2328 vfs_bio_awrite(bp); 2329 else { 2330 bremfree(bp); 2331 bwrite(bp); 2332 notbufdflashes++; 2333 } 2334 vn_finished_write(mp); 2335 VOP_UNLOCK(vp, 0); 2336 flushwithdeps += hasdeps; 2337 flushed++; 2338 2339 /* 2340 * Sleeping on runningbufspace while holding 2341 * vnode lock leads to deadlock. 2342 */ 2343 if (curproc == bufdaemonproc) 2344 waitrunningbufspace(); 2345 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2346 mtx_lock(&bqlock); 2347 continue; 2348 } 2349 vn_finished_write(mp); 2350 BUF_UNLOCK(bp); 2351 } 2352 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2353 mtx_unlock(&bqlock); 2354 free(sentinel, M_TEMP); 2355 return (flushed); 2356} 2357 2358/* 2359 * Check to see if a block is currently memory resident. 2360 */ 2361struct buf * 2362incore(struct bufobj *bo, daddr_t blkno) 2363{ 2364 struct buf *bp; 2365 2366 BO_LOCK(bo); 2367 bp = gbincore(bo, blkno); 2368 BO_UNLOCK(bo); 2369 return (bp); 2370} 2371 2372/* 2373 * Returns true if no I/O is needed to access the 2374 * associated VM object. This is like incore except 2375 * it also hunts around in the VM system for the data. 2376 */ 2377 2378static int 2379inmem(struct vnode * vp, daddr_t blkno) 2380{ 2381 vm_object_t obj; 2382 vm_offset_t toff, tinc, size; 2383 vm_page_t m; 2384 vm_ooffset_t off; 2385 2386 ASSERT_VOP_LOCKED(vp, "inmem"); 2387 2388 if (incore(&vp->v_bufobj, blkno)) 2389 return 1; 2390 if (vp->v_mount == NULL) 2391 return 0; 2392 obj = vp->v_object; 2393 if (obj == NULL) 2394 return (0); 2395 2396 size = PAGE_SIZE; 2397 if (size > vp->v_mount->mnt_stat.f_iosize) 2398 size = vp->v_mount->mnt_stat.f_iosize; 2399 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2400 2401 VM_OBJECT_LOCK(obj); 2402 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2403 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2404 if (!m) 2405 goto notinmem; 2406 tinc = size; 2407 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2408 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2409 if (vm_page_is_valid(m, 2410 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2411 goto notinmem; 2412 } 2413 VM_OBJECT_UNLOCK(obj); 2414 return 1; 2415 2416notinmem: 2417 VM_OBJECT_UNLOCK(obj); 2418 return (0); 2419} 2420 2421/* 2422 * vfs_setdirty: 2423 * 2424 * Sets the dirty range for a buffer based on the status of the dirty 2425 * bits in the pages comprising the buffer. 2426 * 2427 * The range is limited to the size of the buffer. 2428 * 2429 * This routine is primarily used by NFS, but is generalized for the 2430 * B_VMIO case. 2431 */ 2432static void 2433vfs_setdirty(struct buf *bp) 2434{ 2435 2436 /* 2437 * Degenerate case - empty buffer 2438 */ 2439 2440 if (bp->b_bufsize == 0) 2441 return; 2442 2443 /* 2444 * We qualify the scan for modified pages on whether the 2445 * object has been flushed yet. 2446 */ 2447 2448 if ((bp->b_flags & B_VMIO) == 0) 2449 return; 2450 2451 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2452 vfs_setdirty_locked_object(bp); 2453 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2454} 2455 2456static void 2457vfs_setdirty_locked_object(struct buf *bp) 2458{ 2459 vm_object_t object; 2460 int i; 2461 2462 object = bp->b_bufobj->bo_object; 2463 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2464 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2465 vm_offset_t boffset; 2466 vm_offset_t eoffset; 2467 2468 vm_page_lock_queues(); 2469 /* 2470 * test the pages to see if they have been modified directly 2471 * by users through the VM system. 2472 */ 2473 for (i = 0; i < bp->b_npages; i++) 2474 vm_page_test_dirty(bp->b_pages[i]); 2475 2476 /* 2477 * Calculate the encompassing dirty range, boffset and eoffset, 2478 * (eoffset - boffset) bytes. 2479 */ 2480 2481 for (i = 0; i < bp->b_npages; i++) { 2482 if (bp->b_pages[i]->dirty) 2483 break; 2484 } 2485 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2486 2487 for (i = bp->b_npages - 1; i >= 0; --i) { 2488 if (bp->b_pages[i]->dirty) { 2489 break; 2490 } 2491 } 2492 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2493 2494 vm_page_unlock_queues(); 2495 /* 2496 * Fit it to the buffer. 2497 */ 2498 2499 if (eoffset > bp->b_bcount) 2500 eoffset = bp->b_bcount; 2501 2502 /* 2503 * If we have a good dirty range, merge with the existing 2504 * dirty range. 2505 */ 2506 2507 if (boffset < eoffset) { 2508 if (bp->b_dirtyoff > boffset) 2509 bp->b_dirtyoff = boffset; 2510 if (bp->b_dirtyend < eoffset) 2511 bp->b_dirtyend = eoffset; 2512 } 2513 } 2514} 2515 2516/* 2517 * getblk: 2518 * 2519 * Get a block given a specified block and offset into a file/device. 2520 * The buffers B_DONE bit will be cleared on return, making it almost 2521 * ready for an I/O initiation. B_INVAL may or may not be set on 2522 * return. The caller should clear B_INVAL prior to initiating a 2523 * READ. 2524 * 2525 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2526 * an existing buffer. 2527 * 2528 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2529 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2530 * and then cleared based on the backing VM. If the previous buffer is 2531 * non-0-sized but invalid, B_CACHE will be cleared. 2532 * 2533 * If getblk() must create a new buffer, the new buffer is returned with 2534 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2535 * case it is returned with B_INVAL clear and B_CACHE set based on the 2536 * backing VM. 2537 * 2538 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2539 * B_CACHE bit is clear. 2540 * 2541 * What this means, basically, is that the caller should use B_CACHE to 2542 * determine whether the buffer is fully valid or not and should clear 2543 * B_INVAL prior to issuing a read. If the caller intends to validate 2544 * the buffer by loading its data area with something, the caller needs 2545 * to clear B_INVAL. If the caller does this without issuing an I/O, 2546 * the caller should set B_CACHE ( as an optimization ), else the caller 2547 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2548 * a write attempt or if it was a successfull read. If the caller 2549 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2550 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2551 */ 2552struct buf * 2553getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2554 int flags) 2555{ 2556 struct buf *bp; 2557 struct bufobj *bo; 2558 int error; 2559 2560 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2561 ASSERT_VOP_LOCKED(vp, "getblk"); 2562 if (size > MAXBSIZE) 2563 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2564 2565 bo = &vp->v_bufobj; 2566loop: 2567 /* 2568 * Block if we are low on buffers. Certain processes are allowed 2569 * to completely exhaust the buffer cache. 2570 * 2571 * If this check ever becomes a bottleneck it may be better to 2572 * move it into the else, when gbincore() fails. At the moment 2573 * it isn't a problem. 2574 * 2575 * XXX remove if 0 sections (clean this up after its proven) 2576 */ 2577 if (numfreebuffers == 0) { 2578 if (TD_IS_IDLETHREAD(curthread)) 2579 return NULL; 2580 mtx_lock(&nblock); 2581 needsbuffer |= VFS_BIO_NEED_ANY; 2582 mtx_unlock(&nblock); 2583 } 2584 2585 BO_LOCK(bo); 2586 bp = gbincore(bo, blkno); 2587 if (bp != NULL) { 2588 int lockflags; 2589 /* 2590 * Buffer is in-core. If the buffer is not busy, it must 2591 * be on a queue. 2592 */ 2593 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2594 2595 if (flags & GB_LOCK_NOWAIT) 2596 lockflags |= LK_NOWAIT; 2597 2598 error = BUF_TIMELOCK(bp, lockflags, 2599 BO_MTX(bo), "getblk", slpflag, slptimeo); 2600 2601 /* 2602 * If we slept and got the lock we have to restart in case 2603 * the buffer changed identities. 2604 */ 2605 if (error == ENOLCK) 2606 goto loop; 2607 /* We timed out or were interrupted. */ 2608 else if (error) 2609 return (NULL); 2610 2611 /* 2612 * The buffer is locked. B_CACHE is cleared if the buffer is 2613 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2614 * and for a VMIO buffer B_CACHE is adjusted according to the 2615 * backing VM cache. 2616 */ 2617 if (bp->b_flags & B_INVAL) 2618 bp->b_flags &= ~B_CACHE; 2619 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2620 bp->b_flags |= B_CACHE; 2621 bremfree(bp); 2622 2623 /* 2624 * check for size inconsistancies for non-VMIO case. 2625 */ 2626 2627 if (bp->b_bcount != size) { 2628 if ((bp->b_flags & B_VMIO) == 0 || 2629 (size > bp->b_kvasize)) { 2630 if (bp->b_flags & B_DELWRI) { 2631 /* 2632 * If buffer is pinned and caller does 2633 * not want sleep waiting for it to be 2634 * unpinned, bail out 2635 * */ 2636 if (bp->b_pin_count > 0) { 2637 if (flags & GB_LOCK_NOWAIT) { 2638 bqrelse(bp); 2639 return (NULL); 2640 } else { 2641 bunpin_wait(bp); 2642 } 2643 } 2644 bp->b_flags |= B_NOCACHE; 2645 bwrite(bp); 2646 } else { 2647 if (LIST_EMPTY(&bp->b_dep)) { 2648 bp->b_flags |= B_RELBUF; 2649 brelse(bp); 2650 } else { 2651 bp->b_flags |= B_NOCACHE; 2652 bwrite(bp); 2653 } 2654 } 2655 goto loop; 2656 } 2657 } 2658 2659 /* 2660 * If the size is inconsistant in the VMIO case, we can resize 2661 * the buffer. This might lead to B_CACHE getting set or 2662 * cleared. If the size has not changed, B_CACHE remains 2663 * unchanged from its previous state. 2664 */ 2665 2666 if (bp->b_bcount != size) 2667 allocbuf(bp, size); 2668 2669 KASSERT(bp->b_offset != NOOFFSET, 2670 ("getblk: no buffer offset")); 2671 2672 /* 2673 * A buffer with B_DELWRI set and B_CACHE clear must 2674 * be committed before we can return the buffer in 2675 * order to prevent the caller from issuing a read 2676 * ( due to B_CACHE not being set ) and overwriting 2677 * it. 2678 * 2679 * Most callers, including NFS and FFS, need this to 2680 * operate properly either because they assume they 2681 * can issue a read if B_CACHE is not set, or because 2682 * ( for example ) an uncached B_DELWRI might loop due 2683 * to softupdates re-dirtying the buffer. In the latter 2684 * case, B_CACHE is set after the first write completes, 2685 * preventing further loops. 2686 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2687 * above while extending the buffer, we cannot allow the 2688 * buffer to remain with B_CACHE set after the write 2689 * completes or it will represent a corrupt state. To 2690 * deal with this we set B_NOCACHE to scrap the buffer 2691 * after the write. 2692 * 2693 * We might be able to do something fancy, like setting 2694 * B_CACHE in bwrite() except if B_DELWRI is already set, 2695 * so the below call doesn't set B_CACHE, but that gets real 2696 * confusing. This is much easier. 2697 */ 2698 2699 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2700 bp->b_flags |= B_NOCACHE; 2701 bwrite(bp); 2702 goto loop; 2703 } 2704 bp->b_flags &= ~B_DONE; 2705 2706 if (flags & GB_INDIR) { 2707 if (bp->b_flags & B_INDIR) { 2708 if (bp->b_flags & B_CACHE) 2709 BO_INDIR_CHECK(bo, bp); 2710 } else { 2711 printf("XXKAN: Looking incore indirect buffer " 2712 "found non-indirect one"); 2713 kdb_backtrace(); 2714 bp->b_flags |= B_INDIR; 2715 } 2716 } 2717 } else { 2718 int bsize, maxsize, vmio; 2719 off_t offset; 2720 2721 /* 2722 * Buffer is not in-core, create new buffer. The buffer 2723 * returned by getnewbuf() is locked. Note that the returned 2724 * buffer is also considered valid (not marked B_INVAL). 2725 */ 2726 BO_UNLOCK(bo); 2727 /* 2728 * If the user does not want us to create the buffer, bail out 2729 * here. 2730 */ 2731 if (flags & GB_NOCREAT) 2732 return NULL; 2733 bsize = bo->bo_bsize; 2734 offset = blkno * bsize; 2735 vmio = vp->v_object != NULL; 2736 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2737 maxsize = imax(maxsize, bsize); 2738 2739 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 2740 if (bp == NULL) { 2741 if (slpflag || slptimeo) 2742 return NULL; 2743 goto loop; 2744 } 2745 2746 /* 2747 * This code is used to make sure that a buffer is not 2748 * created while the getnewbuf routine is blocked. 2749 * This can be a problem whether the vnode is locked or not. 2750 * If the buffer is created out from under us, we have to 2751 * throw away the one we just created. 2752 * 2753 * Note: this must occur before we associate the buffer 2754 * with the vp especially considering limitations in 2755 * the splay tree implementation when dealing with duplicate 2756 * lblkno's. 2757 */ 2758 BO_LOCK(bo); 2759 if (gbincore(bo, blkno)) { 2760 BO_UNLOCK(bo); 2761 bp->b_flags |= B_INVAL; 2762 brelse(bp); 2763 goto loop; 2764 } 2765 2766 /* 2767 * Insert the buffer into the hash, so that it can 2768 * be found by incore. 2769 */ 2770 bp->b_blkno = bp->b_lblkno = blkno; 2771 bp->b_offset = offset; 2772 bgetvp(vp, bp); 2773 BO_UNLOCK(bo); 2774 2775 /* 2776 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2777 * buffer size starts out as 0, B_CACHE will be set by 2778 * allocbuf() for the VMIO case prior to it testing the 2779 * backing store for validity. 2780 */ 2781 2782 if (vmio) { 2783 bp->b_flags |= B_VMIO; 2784#if defined(VFS_BIO_DEBUG) 2785 if (vn_canvmio(vp) != TRUE) 2786 printf("getblk: VMIO on vnode type %d\n", 2787 vp->v_type); 2788#endif 2789 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2790 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2791 bp, vp->v_object, bp->b_bufobj->bo_object)); 2792 } else { 2793 bp->b_flags &= ~B_VMIO; 2794 KASSERT(bp->b_bufobj->bo_object == NULL, 2795 ("ARGH! has b_bufobj->bo_object %p %p\n", 2796 bp, bp->b_bufobj->bo_object)); 2797 } 2798 2799 allocbuf(bp, size); 2800 bp->b_flags &= ~B_DONE; 2801 if (flags & GB_INDIR) 2802 bp->b_flags |= B_INDIR; 2803 } 2804 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2805 BUF_ASSERT_HELD(bp); 2806 KASSERT(bp->b_bufobj == bo, 2807 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2808 return (bp); 2809} 2810 2811/* 2812 * Get an empty, disassociated buffer of given size. The buffer is initially 2813 * set to B_INVAL. 2814 */ 2815struct buf * 2816geteblk(int size, int flags) 2817{ 2818 struct buf *bp; 2819 int maxsize; 2820 2821 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2822 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 2823 if ((flags & GB_NOWAIT_BD) && 2824 (curthread->td_pflags & TDP_BUFNEED) != 0) 2825 return (NULL); 2826 } 2827 allocbuf(bp, size); 2828 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2829 BUF_ASSERT_HELD(bp); 2830 return (bp); 2831} 2832 2833 2834/* 2835 * This code constitutes the buffer memory from either anonymous system 2836 * memory (in the case of non-VMIO operations) or from an associated 2837 * VM object (in the case of VMIO operations). This code is able to 2838 * resize a buffer up or down. 2839 * 2840 * Note that this code is tricky, and has many complications to resolve 2841 * deadlock or inconsistant data situations. Tread lightly!!! 2842 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2843 * the caller. Calling this code willy nilly can result in the loss of data. 2844 * 2845 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2846 * B_CACHE for the non-VMIO case. 2847 */ 2848 2849int 2850allocbuf(struct buf *bp, int size) 2851{ 2852 int newbsize, mbsize; 2853 int i; 2854 2855 BUF_ASSERT_HELD(bp); 2856 2857 if (bp->b_kvasize < size) 2858 panic("allocbuf: buffer too small"); 2859 2860 if ((bp->b_flags & B_VMIO) == 0) { 2861 caddr_t origbuf; 2862 int origbufsize; 2863 /* 2864 * Just get anonymous memory from the kernel. Don't 2865 * mess with B_CACHE. 2866 */ 2867 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2868 if (bp->b_flags & B_MALLOC) 2869 newbsize = mbsize; 2870 else 2871 newbsize = round_page(size); 2872 2873 if (newbsize < bp->b_bufsize) { 2874 /* 2875 * malloced buffers are not shrunk 2876 */ 2877 if (bp->b_flags & B_MALLOC) { 2878 if (newbsize) { 2879 bp->b_bcount = size; 2880 } else { 2881 free(bp->b_data, M_BIOBUF); 2882 if (bp->b_bufsize) { 2883 atomic_subtract_long( 2884 &bufmallocspace, 2885 bp->b_bufsize); 2886 bufspacewakeup(); 2887 bp->b_bufsize = 0; 2888 } 2889 bp->b_saveaddr = bp->b_kvabase; 2890 bp->b_data = bp->b_saveaddr; 2891 bp->b_bcount = 0; 2892 bp->b_flags &= ~B_MALLOC; 2893 } 2894 return 1; 2895 } 2896 vm_hold_free_pages( 2897 bp, 2898 (vm_offset_t) bp->b_data + newbsize, 2899 (vm_offset_t) bp->b_data + bp->b_bufsize); 2900 } else if (newbsize > bp->b_bufsize) { 2901 /* 2902 * We only use malloced memory on the first allocation. 2903 * and revert to page-allocated memory when the buffer 2904 * grows. 2905 */ 2906 /* 2907 * There is a potential smp race here that could lead 2908 * to bufmallocspace slightly passing the max. It 2909 * is probably extremely rare and not worth worrying 2910 * over. 2911 */ 2912 if ( (bufmallocspace < maxbufmallocspace) && 2913 (bp->b_bufsize == 0) && 2914 (mbsize <= PAGE_SIZE/2)) { 2915 2916 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2917 bp->b_bufsize = mbsize; 2918 bp->b_bcount = size; 2919 bp->b_flags |= B_MALLOC; 2920 atomic_add_long(&bufmallocspace, mbsize); 2921 return 1; 2922 } 2923 origbuf = NULL; 2924 origbufsize = 0; 2925 /* 2926 * If the buffer is growing on its other-than-first allocation, 2927 * then we revert to the page-allocation scheme. 2928 */ 2929 if (bp->b_flags & B_MALLOC) { 2930 origbuf = bp->b_data; 2931 origbufsize = bp->b_bufsize; 2932 bp->b_data = bp->b_kvabase; 2933 if (bp->b_bufsize) { 2934 atomic_subtract_long(&bufmallocspace, 2935 bp->b_bufsize); 2936 bufspacewakeup(); 2937 bp->b_bufsize = 0; 2938 } 2939 bp->b_flags &= ~B_MALLOC; 2940 newbsize = round_page(newbsize); 2941 } 2942 vm_hold_load_pages( 2943 bp, 2944 (vm_offset_t) bp->b_data + bp->b_bufsize, 2945 (vm_offset_t) bp->b_data + newbsize); 2946 if (origbuf) { 2947 bcopy(origbuf, bp->b_data, origbufsize); 2948 free(origbuf, M_BIOBUF); 2949 } 2950 } 2951 } else { 2952 int desiredpages; 2953 2954 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2955 desiredpages = (size == 0) ? 0 : 2956 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2957 2958 if (bp->b_flags & B_MALLOC) 2959 panic("allocbuf: VMIO buffer can't be malloced"); 2960 /* 2961 * Set B_CACHE initially if buffer is 0 length or will become 2962 * 0-length. 2963 */ 2964 if (size == 0 || bp->b_bufsize == 0) 2965 bp->b_flags |= B_CACHE; 2966 2967 if (newbsize < bp->b_bufsize) { 2968 /* 2969 * DEV_BSIZE aligned new buffer size is less then the 2970 * DEV_BSIZE aligned existing buffer size. Figure out 2971 * if we have to remove any pages. 2972 */ 2973 if (desiredpages < bp->b_npages) { 2974 vm_page_t m; 2975 2976 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2977 vm_page_lock_queues(); 2978 for (i = desiredpages; i < bp->b_npages; i++) { 2979 /* 2980 * the page is not freed here -- it 2981 * is the responsibility of 2982 * vnode_pager_setsize 2983 */ 2984 m = bp->b_pages[i]; 2985 KASSERT(m != bogus_page, 2986 ("allocbuf: bogus page found")); 2987 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2988 vm_page_lock_queues(); 2989 2990 bp->b_pages[i] = NULL; 2991 vm_page_unwire(m, 0); 2992 } 2993 vm_page_unlock_queues(); 2994 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2995 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2996 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2997 bp->b_npages = desiredpages; 2998 } 2999 } else if (size > bp->b_bcount) { 3000 /* 3001 * We are growing the buffer, possibly in a 3002 * byte-granular fashion. 3003 */ 3004 struct vnode *vp; 3005 vm_object_t obj; 3006 vm_offset_t toff; 3007 vm_offset_t tinc; 3008 3009 /* 3010 * Step 1, bring in the VM pages from the object, 3011 * allocating them if necessary. We must clear 3012 * B_CACHE if these pages are not valid for the 3013 * range covered by the buffer. 3014 */ 3015 3016 vp = bp->b_vp; 3017 obj = bp->b_bufobj->bo_object; 3018 3019 VM_OBJECT_LOCK(obj); 3020 while (bp->b_npages < desiredpages) { 3021 vm_page_t m; 3022 vm_pindex_t pi; 3023 3024 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 3025 if ((m = vm_page_lookup(obj, pi)) == NULL) { 3026 /* 3027 * note: must allocate system pages 3028 * since blocking here could intefere 3029 * with paging I/O, no matter which 3030 * process we are. 3031 */ 3032 m = vm_page_alloc(obj, pi, 3033 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 3034 VM_ALLOC_WIRED); 3035 if (m == NULL) { 3036 atomic_add_int(&vm_pageout_deficit, 3037 desiredpages - bp->b_npages); 3038 VM_OBJECT_UNLOCK(obj); 3039 VM_WAIT; 3040 VM_OBJECT_LOCK(obj); 3041 } else { 3042 if (m->valid == 0) 3043 bp->b_flags &= ~B_CACHE; 3044 bp->b_pages[bp->b_npages] = m; 3045 ++bp->b_npages; 3046 } 3047 continue; 3048 } 3049 3050 /* 3051 * We found a page. If we have to sleep on it, 3052 * retry because it might have gotten freed out 3053 * from under us. 3054 * 3055 * We can only test VPO_BUSY here. Blocking on 3056 * m->busy might lead to a deadlock: 3057 * 3058 * vm_fault->getpages->cluster_read->allocbuf 3059 * 3060 */ 3061 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 3062 continue; 3063 3064 /* 3065 * We have a good page. 3066 */ 3067 vm_page_lock_queues(); 3068 vm_page_wire(m); 3069 vm_page_unlock_queues(); 3070 bp->b_pages[bp->b_npages] = m; 3071 ++bp->b_npages; 3072 } 3073 3074 /* 3075 * Step 2. We've loaded the pages into the buffer, 3076 * we have to figure out if we can still have B_CACHE 3077 * set. Note that B_CACHE is set according to the 3078 * byte-granular range ( bcount and size ), new the 3079 * aligned range ( newbsize ). 3080 * 3081 * The VM test is against m->valid, which is DEV_BSIZE 3082 * aligned. Needless to say, the validity of the data 3083 * needs to also be DEV_BSIZE aligned. Note that this 3084 * fails with NFS if the server or some other client 3085 * extends the file's EOF. If our buffer is resized, 3086 * B_CACHE may remain set! XXX 3087 */ 3088 3089 toff = bp->b_bcount; 3090 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3091 3092 while ((bp->b_flags & B_CACHE) && toff < size) { 3093 vm_pindex_t pi; 3094 3095 if (tinc > (size - toff)) 3096 tinc = size - toff; 3097 3098 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3099 PAGE_SHIFT; 3100 3101 vfs_buf_test_cache( 3102 bp, 3103 bp->b_offset, 3104 toff, 3105 tinc, 3106 bp->b_pages[pi] 3107 ); 3108 toff += tinc; 3109 tinc = PAGE_SIZE; 3110 } 3111 VM_OBJECT_UNLOCK(obj); 3112 3113 /* 3114 * Step 3, fixup the KVM pmap. Remember that 3115 * bp->b_data is relative to bp->b_offset, but 3116 * bp->b_offset may be offset into the first page. 3117 */ 3118 3119 bp->b_data = (caddr_t) 3120 trunc_page((vm_offset_t)bp->b_data); 3121 pmap_qenter( 3122 (vm_offset_t)bp->b_data, 3123 bp->b_pages, 3124 bp->b_npages 3125 ); 3126 3127 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3128 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3129 } 3130 } 3131 if (newbsize < bp->b_bufsize) 3132 bufspacewakeup(); 3133 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3134 bp->b_bcount = size; /* requested buffer size */ 3135 return 1; 3136} 3137 3138void 3139biodone(struct bio *bp) 3140{ 3141 struct mtx *mtxp; 3142 void (*done)(struct bio *); 3143 3144 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3145 mtx_lock(mtxp); 3146 bp->bio_flags |= BIO_DONE; 3147 done = bp->bio_done; 3148 if (done == NULL) 3149 wakeup(bp); 3150 mtx_unlock(mtxp); 3151 if (done != NULL) 3152 done(bp); 3153} 3154 3155/* 3156 * Wait for a BIO to finish. 3157 * 3158 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3159 * case is not yet clear. 3160 */ 3161int 3162biowait(struct bio *bp, const char *wchan) 3163{ 3164 struct mtx *mtxp; 3165 3166 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3167 mtx_lock(mtxp); 3168 while ((bp->bio_flags & BIO_DONE) == 0) 3169 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3170 mtx_unlock(mtxp); 3171 if (bp->bio_error != 0) 3172 return (bp->bio_error); 3173 if (!(bp->bio_flags & BIO_ERROR)) 3174 return (0); 3175 return (EIO); 3176} 3177 3178void 3179biofinish(struct bio *bp, struct devstat *stat, int error) 3180{ 3181 3182 if (error) { 3183 bp->bio_error = error; 3184 bp->bio_flags |= BIO_ERROR; 3185 } 3186 if (stat != NULL) 3187 devstat_end_transaction_bio(stat, bp); 3188 biodone(bp); 3189} 3190 3191/* 3192 * bufwait: 3193 * 3194 * Wait for buffer I/O completion, returning error status. The buffer 3195 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3196 * error and cleared. 3197 */ 3198int 3199bufwait(struct buf *bp) 3200{ 3201 if (bp->b_iocmd == BIO_READ) 3202 bwait(bp, PRIBIO, "biord"); 3203 else 3204 bwait(bp, PRIBIO, "biowr"); 3205 if (bp->b_flags & B_EINTR) { 3206 bp->b_flags &= ~B_EINTR; 3207 return (EINTR); 3208 } 3209 if (bp->b_ioflags & BIO_ERROR) { 3210 return (bp->b_error ? bp->b_error : EIO); 3211 } else { 3212 return (0); 3213 } 3214} 3215 3216 /* 3217 * Call back function from struct bio back up to struct buf. 3218 */ 3219static void 3220bufdonebio(struct bio *bip) 3221{ 3222 struct buf *bp; 3223 3224 bp = bip->bio_caller2; 3225 bp->b_resid = bp->b_bcount - bip->bio_completed; 3226 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3227 bp->b_ioflags = bip->bio_flags; 3228 bp->b_error = bip->bio_error; 3229 if (bp->b_error) 3230 bp->b_ioflags |= BIO_ERROR; 3231 bufdone(bp); 3232 g_destroy_bio(bip); 3233} 3234 3235void 3236dev_strategy(struct cdev *dev, struct buf *bp) 3237{ 3238 struct cdevsw *csw; 3239 struct bio *bip; 3240 3241 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3242 panic("b_iocmd botch"); 3243 for (;;) { 3244 bip = g_new_bio(); 3245 if (bip != NULL) 3246 break; 3247 /* Try again later */ 3248 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3249 } 3250 bip->bio_cmd = bp->b_iocmd; 3251 bip->bio_offset = bp->b_iooffset; 3252 bip->bio_length = bp->b_bcount; 3253 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3254 bip->bio_data = bp->b_data; 3255 bip->bio_done = bufdonebio; 3256 bip->bio_caller2 = bp; 3257 bip->bio_dev = dev; 3258 KASSERT(dev->si_refcount > 0, 3259 ("dev_strategy on un-referenced struct cdev *(%s)", 3260 devtoname(dev))); 3261 csw = dev_refthread(dev); 3262 if (csw == NULL) { 3263 g_destroy_bio(bip); 3264 bp->b_error = ENXIO; 3265 bp->b_ioflags = BIO_ERROR; 3266 bufdone(bp); 3267 return; 3268 } 3269 (*csw->d_strategy)(bip); 3270 dev_relthread(dev); 3271} 3272 3273/* 3274 * bufdone: 3275 * 3276 * Finish I/O on a buffer, optionally calling a completion function. 3277 * This is usually called from an interrupt so process blocking is 3278 * not allowed. 3279 * 3280 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3281 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3282 * assuming B_INVAL is clear. 3283 * 3284 * For the VMIO case, we set B_CACHE if the op was a read and no 3285 * read error occured, or if the op was a write. B_CACHE is never 3286 * set if the buffer is invalid or otherwise uncacheable. 3287 * 3288 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3289 * initiator to leave B_INVAL set to brelse the buffer out of existance 3290 * in the biodone routine. 3291 */ 3292void 3293bufdone(struct buf *bp) 3294{ 3295 struct bufobj *dropobj; 3296 void (*biodone)(struct buf *); 3297 3298 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3299 dropobj = NULL; 3300 3301 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3302 BUF_ASSERT_HELD(bp); 3303 3304 runningbufwakeup(bp); 3305 if (bp->b_iocmd == BIO_WRITE) 3306 dropobj = bp->b_bufobj; 3307 /* call optional completion function if requested */ 3308 if (bp->b_iodone != NULL) { 3309 biodone = bp->b_iodone; 3310 bp->b_iodone = NULL; 3311 if (bp->b_flags & B_INDIR) { 3312 if ((bp->b_flags & B_INVAL) == 0 && 3313 (bp->b_ioflags & BIO_ERROR) == 0) { 3314 BO_INDIR_CHECK(bp->b_bufobj, bp); 3315 } 3316 } 3317 (*biodone) (bp); 3318 if (dropobj) 3319 bufobj_wdrop(dropobj); 3320 return; 3321 } 3322 3323 bufdone_finish(bp); 3324 3325 if (dropobj) 3326 bufobj_wdrop(dropobj); 3327} 3328 3329void 3330bufdone_finish(struct buf *bp) 3331{ 3332 BUF_ASSERT_HELD(bp); 3333 3334 if (!LIST_EMPTY(&bp->b_dep)) 3335 buf_complete(bp); 3336 3337 if (bp->b_flags & B_VMIO) { 3338 int i; 3339 vm_ooffset_t foff; 3340 vm_page_t m; 3341 vm_object_t obj; 3342 int iosize; 3343 struct vnode *vp = bp->b_vp; 3344 boolean_t are_queues_locked; 3345 3346 obj = bp->b_bufobj->bo_object; 3347 3348#if defined(VFS_BIO_DEBUG) 3349 mp_fixme("usecount and vflag accessed without locks."); 3350 if (vp->v_usecount == 0) { 3351 panic("biodone: zero vnode ref count"); 3352 } 3353 3354 KASSERT(vp->v_object != NULL, 3355 ("biodone: vnode %p has no vm_object", vp)); 3356#endif 3357 3358 foff = bp->b_offset; 3359 KASSERT(bp->b_offset != NOOFFSET, 3360 ("biodone: no buffer offset")); 3361 3362 VM_OBJECT_LOCK(obj); 3363#if defined(VFS_BIO_DEBUG) 3364 if (obj->paging_in_progress < bp->b_npages) { 3365 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3366 obj->paging_in_progress, bp->b_npages); 3367 } 3368#endif 3369 3370 /* 3371 * Set B_CACHE if the op was a normal read and no error 3372 * occured. B_CACHE is set for writes in the b*write() 3373 * routines. 3374 */ 3375 iosize = bp->b_bcount - bp->b_resid; 3376 if (bp->b_iocmd == BIO_READ && 3377 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3378 !(bp->b_ioflags & BIO_ERROR)) { 3379 bp->b_flags |= B_CACHE; 3380 } 3381 if (bp->b_iocmd == BIO_READ) { 3382 vm_page_lock_queues(); 3383 are_queues_locked = TRUE; 3384 } else 3385 are_queues_locked = FALSE; 3386 for (i = 0; i < bp->b_npages; i++) { 3387 int bogusflag = 0; 3388 int resid; 3389 3390 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3391 if (resid > iosize) 3392 resid = iosize; 3393 3394 /* 3395 * cleanup bogus pages, restoring the originals 3396 */ 3397 m = bp->b_pages[i]; 3398 if (m == bogus_page) { 3399 bogusflag = 1; 3400 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3401 if (m == NULL) 3402 panic("biodone: page disappeared!"); 3403 bp->b_pages[i] = m; 3404 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3405 bp->b_pages, bp->b_npages); 3406 } 3407#if defined(VFS_BIO_DEBUG) 3408 if (OFF_TO_IDX(foff) != m->pindex) { 3409 printf( 3410"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3411 (intmax_t)foff, (uintmax_t)m->pindex); 3412 } 3413#endif 3414 3415 /* 3416 * In the write case, the valid and clean bits are 3417 * already changed correctly ( see bdwrite() ), so we 3418 * only need to do this here in the read case. 3419 */ 3420 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3421 vfs_page_set_valid(bp, foff, m); 3422 } 3423 3424 /* 3425 * when debugging new filesystems or buffer I/O methods, this 3426 * is the most common error that pops up. if you see this, you 3427 * have not set the page busy flag correctly!!! 3428 */ 3429 if (m->busy == 0) { 3430 printf("biodone: page busy < 0, " 3431 "pindex: %d, foff: 0x(%x,%x), " 3432 "resid: %d, index: %d\n", 3433 (int) m->pindex, (int)(foff >> 32), 3434 (int) foff & 0xffffffff, resid, i); 3435 if (!vn_isdisk(vp, NULL)) 3436 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3437 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3438 (intmax_t) bp->b_lblkno, 3439 bp->b_flags, bp->b_npages); 3440 else 3441 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3442 (intmax_t) bp->b_lblkno, 3443 bp->b_flags, bp->b_npages); 3444 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3445 (u_long)m->valid, (u_long)m->dirty, 3446 m->wire_count); 3447 panic("biodone: page busy < 0\n"); 3448 } 3449 vm_page_io_finish(m); 3450 vm_object_pip_subtract(obj, 1); 3451 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3452 iosize -= resid; 3453 } 3454 if (are_queues_locked) 3455 vm_page_unlock_queues(); 3456 vm_object_pip_wakeupn(obj, 0); 3457 VM_OBJECT_UNLOCK(obj); 3458 } 3459 3460 if (bp->b_flags & B_INDIR) { 3461 if ((bp->b_flags & B_INVAL) == 0 && 3462 (bp->b_ioflags & BIO_ERROR) == 0) { 3463 BO_INDIR_CHECK(bp->b_bufobj, bp); 3464 } 3465 } 3466 /* 3467 * For asynchronous completions, release the buffer now. The brelse 3468 * will do a wakeup there if necessary - so no need to do a wakeup 3469 * here in the async case. The sync case always needs to do a wakeup. 3470 */ 3471 3472 if (bp->b_flags & B_ASYNC) { 3473 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3474 brelse(bp); 3475 else 3476 bqrelse(bp); 3477 } else 3478 bdone(bp); 3479} 3480 3481/* 3482 * This routine is called in lieu of iodone in the case of 3483 * incomplete I/O. This keeps the busy status for pages 3484 * consistant. 3485 */ 3486void 3487vfs_unbusy_pages(struct buf *bp) 3488{ 3489 int i; 3490 vm_object_t obj; 3491 vm_page_t m; 3492 3493 runningbufwakeup(bp); 3494 if (!(bp->b_flags & B_VMIO)) 3495 return; 3496 3497 obj = bp->b_bufobj->bo_object; 3498 VM_OBJECT_LOCK(obj); 3499 for (i = 0; i < bp->b_npages; i++) { 3500 m = bp->b_pages[i]; 3501 if (m == bogus_page) { 3502 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3503 if (!m) 3504 panic("vfs_unbusy_pages: page missing\n"); 3505 bp->b_pages[i] = m; 3506 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3507 bp->b_pages, bp->b_npages); 3508 } 3509 vm_object_pip_subtract(obj, 1); 3510 vm_page_io_finish(m); 3511 } 3512 vm_object_pip_wakeupn(obj, 0); 3513 VM_OBJECT_UNLOCK(obj); 3514} 3515 3516/* 3517 * vfs_page_set_valid: 3518 * 3519 * Set the valid bits in a page based on the supplied offset. The 3520 * range is restricted to the buffer's size. 3521 * 3522 * This routine is typically called after a read completes. 3523 */ 3524static void 3525vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3526{ 3527 vm_ooffset_t soff, eoff; 3528 3529 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3530 /* 3531 * Start and end offsets in buffer. eoff - soff may not cross a 3532 * page boundry or cross the end of the buffer. The end of the 3533 * buffer, in this case, is our file EOF, not the allocation size 3534 * of the buffer. 3535 */ 3536 soff = off; 3537 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3538 if (eoff > bp->b_offset + bp->b_bcount) 3539 eoff = bp->b_offset + bp->b_bcount; 3540 3541 /* 3542 * Set valid range. This is typically the entire buffer and thus the 3543 * entire page. 3544 */ 3545 if (eoff > soff) { 3546 vm_page_set_validclean( 3547 m, 3548 (vm_offset_t) (soff & PAGE_MASK), 3549 (vm_offset_t) (eoff - soff) 3550 ); 3551 } 3552} 3553 3554/* 3555 * This routine is called before a device strategy routine. 3556 * It is used to tell the VM system that paging I/O is in 3557 * progress, and treat the pages associated with the buffer 3558 * almost as being VPO_BUSY. Also the object paging_in_progress 3559 * flag is handled to make sure that the object doesn't become 3560 * inconsistant. 3561 * 3562 * Since I/O has not been initiated yet, certain buffer flags 3563 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3564 * and should be ignored. 3565 */ 3566void 3567vfs_busy_pages(struct buf *bp, int clear_modify) 3568{ 3569 int i, bogus; 3570 vm_object_t obj; 3571 vm_ooffset_t foff; 3572 vm_page_t m; 3573 3574 if (!(bp->b_flags & B_VMIO)) 3575 return; 3576 3577 obj = bp->b_bufobj->bo_object; 3578 foff = bp->b_offset; 3579 KASSERT(bp->b_offset != NOOFFSET, 3580 ("vfs_busy_pages: no buffer offset")); 3581 VM_OBJECT_LOCK(obj); 3582 if (bp->b_bufsize != 0) 3583 vfs_setdirty_locked_object(bp); 3584retry: 3585 for (i = 0; i < bp->b_npages; i++) { 3586 m = bp->b_pages[i]; 3587 3588 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3589 goto retry; 3590 } 3591 bogus = 0; 3592 vm_page_lock_queues(); 3593 for (i = 0; i < bp->b_npages; i++) { 3594 m = bp->b_pages[i]; 3595 3596 if ((bp->b_flags & B_CLUSTER) == 0) { 3597 vm_object_pip_add(obj, 1); 3598 vm_page_io_start(m); 3599 } 3600 /* 3601 * When readying a buffer for a read ( i.e 3602 * clear_modify == 0 ), it is important to do 3603 * bogus_page replacement for valid pages in 3604 * partially instantiated buffers. Partially 3605 * instantiated buffers can, in turn, occur when 3606 * reconstituting a buffer from its VM backing store 3607 * base. We only have to do this if B_CACHE is 3608 * clear ( which causes the I/O to occur in the 3609 * first place ). The replacement prevents the read 3610 * I/O from overwriting potentially dirty VM-backed 3611 * pages. XXX bogus page replacement is, uh, bogus. 3612 * It may not work properly with small-block devices. 3613 * We need to find a better way. 3614 */ 3615 pmap_remove_all(m); 3616 if (clear_modify) 3617 vfs_page_set_valid(bp, foff, m); 3618 else if (m->valid == VM_PAGE_BITS_ALL && 3619 (bp->b_flags & B_CACHE) == 0) { 3620 bp->b_pages[i] = bogus_page; 3621 bogus++; 3622 } 3623 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3624 } 3625 vm_page_unlock_queues(); 3626 VM_OBJECT_UNLOCK(obj); 3627 if (bogus) 3628 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3629 bp->b_pages, bp->b_npages); 3630} 3631 3632/* 3633 * Tell the VM system that the pages associated with this buffer 3634 * are clean. This is used for delayed writes where the data is 3635 * going to go to disk eventually without additional VM intevention. 3636 * 3637 * Note that while we only really need to clean through to b_bcount, we 3638 * just go ahead and clean through to b_bufsize. 3639 */ 3640static void 3641vfs_clean_pages(struct buf *bp) 3642{ 3643 int i; 3644 vm_ooffset_t foff, noff, eoff; 3645 vm_page_t m; 3646 3647 if (!(bp->b_flags & B_VMIO)) 3648 return; 3649 3650 foff = bp->b_offset; 3651 KASSERT(bp->b_offset != NOOFFSET, 3652 ("vfs_clean_pages: no buffer offset")); 3653 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3654 vm_page_lock_queues(); 3655 for (i = 0; i < bp->b_npages; i++) { 3656 m = bp->b_pages[i]; 3657 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3658 eoff = noff; 3659 3660 if (eoff > bp->b_offset + bp->b_bufsize) 3661 eoff = bp->b_offset + bp->b_bufsize; 3662 vfs_page_set_valid(bp, foff, m); 3663 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3664 foff = noff; 3665 } 3666 vm_page_unlock_queues(); 3667 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3668} 3669 3670/* 3671 * vfs_bio_set_validclean: 3672 * 3673 * Set the range within the buffer to valid and clean. The range is 3674 * relative to the beginning of the buffer, b_offset. Note that b_offset 3675 * itself may be offset from the beginning of the first page. 3676 * 3677 */ 3678 3679void 3680vfs_bio_set_validclean(struct buf *bp, int base, int size) 3681{ 3682 int i, n; 3683 vm_page_t m; 3684 3685 if (!(bp->b_flags & B_VMIO)) 3686 return; 3687 /* 3688 * Fixup base to be relative to beginning of first page. 3689 * Set initial n to be the maximum number of bytes in the 3690 * first page that can be validated. 3691 */ 3692 3693 base += (bp->b_offset & PAGE_MASK); 3694 n = PAGE_SIZE - (base & PAGE_MASK); 3695 3696 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3697 vm_page_lock_queues(); 3698 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3699 m = bp->b_pages[i]; 3700 if (n > size) 3701 n = size; 3702 vm_page_set_validclean(m, base & PAGE_MASK, n); 3703 base += n; 3704 size -= n; 3705 n = PAGE_SIZE; 3706 } 3707 vm_page_unlock_queues(); 3708 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3709} 3710 3711/* 3712 * vfs_bio_clrbuf: 3713 * 3714 * clear a buffer. This routine essentially fakes an I/O, so we need 3715 * to clear BIO_ERROR and B_INVAL. 3716 * 3717 * Note that while we only theoretically need to clear through b_bcount, 3718 * we go ahead and clear through b_bufsize. 3719 */ 3720 3721void 3722vfs_bio_clrbuf(struct buf *bp) 3723{ 3724 int i, j, mask = 0; 3725 caddr_t sa, ea; 3726 3727 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3728 clrbuf(bp); 3729 return; 3730 } 3731 3732 bp->b_flags &= ~B_INVAL; 3733 bp->b_ioflags &= ~BIO_ERROR; 3734 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3735 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3736 (bp->b_offset & PAGE_MASK) == 0) { 3737 if (bp->b_pages[0] == bogus_page) 3738 goto unlock; 3739 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3740 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3741 if ((bp->b_pages[0]->valid & mask) == mask) 3742 goto unlock; 3743 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3744 ((bp->b_pages[0]->valid & mask) == 0)) { 3745 bzero(bp->b_data, bp->b_bufsize); 3746 bp->b_pages[0]->valid |= mask; 3747 goto unlock; 3748 } 3749 } 3750 ea = sa = bp->b_data; 3751 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3752 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3753 ea = (caddr_t)(vm_offset_t)ulmin( 3754 (u_long)(vm_offset_t)ea, 3755 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3756 if (bp->b_pages[i] == bogus_page) 3757 continue; 3758 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3759 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3760 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3761 if ((bp->b_pages[i]->valid & mask) == mask) 3762 continue; 3763 if ((bp->b_pages[i]->valid & mask) == 0) { 3764 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3765 bzero(sa, ea - sa); 3766 } else { 3767 for (; sa < ea; sa += DEV_BSIZE, j++) { 3768 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3769 (bp->b_pages[i]->valid & (1 << j)) == 0) 3770 bzero(sa, DEV_BSIZE); 3771 } 3772 } 3773 bp->b_pages[i]->valid |= mask; 3774 } 3775unlock: 3776 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3777 bp->b_resid = 0; 3778} 3779 3780/* 3781 * vm_hold_load_pages and vm_hold_free_pages get pages into 3782 * a buffers address space. The pages are anonymous and are 3783 * not associated with a file object. 3784 */ 3785static void 3786vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3787{ 3788 vm_offset_t pg; 3789 vm_page_t p; 3790 int index; 3791 3792 to = round_page(to); 3793 from = round_page(from); 3794 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3795 3796 VM_OBJECT_LOCK(kernel_object); 3797 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3798tryagain: 3799 /* 3800 * note: must allocate system pages since blocking here 3801 * could intefere with paging I/O, no matter which 3802 * process we are. 3803 */ 3804 p = vm_page_alloc(kernel_object, 3805 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3806 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3807 if (!p) { 3808 atomic_add_int(&vm_pageout_deficit, 3809 (to - pg) >> PAGE_SHIFT); 3810 VM_OBJECT_UNLOCK(kernel_object); 3811 VM_WAIT; 3812 VM_OBJECT_LOCK(kernel_object); 3813 goto tryagain; 3814 } 3815 p->valid = VM_PAGE_BITS_ALL; 3816 pmap_qenter(pg, &p, 1); 3817 bp->b_pages[index] = p; 3818 } 3819 VM_OBJECT_UNLOCK(kernel_object); 3820 bp->b_npages = index; 3821} 3822 3823/* Return pages associated with this buf to the vm system */ 3824static void 3825vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3826{ 3827 vm_offset_t pg; 3828 vm_page_t p; 3829 int index, newnpages; 3830 3831 from = round_page(from); 3832 to = round_page(to); 3833 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3834 3835 VM_OBJECT_LOCK(kernel_object); 3836 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3837 p = bp->b_pages[index]; 3838 if (p && (index < bp->b_npages)) { 3839 if (p->busy) { 3840 printf( 3841 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3842 (intmax_t)bp->b_blkno, 3843 (intmax_t)bp->b_lblkno); 3844 } 3845 bp->b_pages[index] = NULL; 3846 pmap_qremove(pg, 1); 3847 vm_page_lock_queues(); 3848 vm_page_unwire(p, 0); 3849 vm_page_free(p); 3850 vm_page_unlock_queues(); 3851 } 3852 } 3853 VM_OBJECT_UNLOCK(kernel_object); 3854 bp->b_npages = newnpages; 3855} 3856 3857/* 3858 * Map an IO request into kernel virtual address space. 3859 * 3860 * All requests are (re)mapped into kernel VA space. 3861 * Notice that we use b_bufsize for the size of the buffer 3862 * to be mapped. b_bcount might be modified by the driver. 3863 * 3864 * Note that even if the caller determines that the address space should 3865 * be valid, a race or a smaller-file mapped into a larger space may 3866 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3867 * check the return value. 3868 */ 3869int 3870vmapbuf(struct buf *bp) 3871{ 3872 caddr_t addr, kva; 3873 vm_prot_t prot; 3874 int pidx, i; 3875 struct vm_page *m; 3876 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3877 3878 if (bp->b_bufsize < 0) 3879 return (-1); 3880 prot = VM_PROT_READ; 3881 if (bp->b_iocmd == BIO_READ) 3882 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3883 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3884 addr < bp->b_data + bp->b_bufsize; 3885 addr += PAGE_SIZE, pidx++) { 3886 /* 3887 * Do the vm_fault if needed; do the copy-on-write thing 3888 * when reading stuff off device into memory. 3889 * 3890 * NOTE! Must use pmap_extract() because addr may be in 3891 * the userland address space, and kextract is only guarenteed 3892 * to work for the kernland address space (see: sparc64 port). 3893 */ 3894retry: 3895 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3896 prot) < 0) { 3897 vm_page_lock_queues(); 3898 for (i = 0; i < pidx; ++i) { 3899 vm_page_unhold(bp->b_pages[i]); 3900 bp->b_pages[i] = NULL; 3901 } 3902 vm_page_unlock_queues(); 3903 return(-1); 3904 } 3905 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3906 if (m == NULL) 3907 goto retry; 3908 bp->b_pages[pidx] = m; 3909 } 3910 if (pidx > btoc(MAXPHYS)) 3911 panic("vmapbuf: mapped more than MAXPHYS"); 3912 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3913 3914 kva = bp->b_saveaddr; 3915 bp->b_npages = pidx; 3916 bp->b_saveaddr = bp->b_data; 3917 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3918 return(0); 3919} 3920 3921/* 3922 * Free the io map PTEs associated with this IO operation. 3923 * We also invalidate the TLB entries and restore the original b_addr. 3924 */ 3925void 3926vunmapbuf(struct buf *bp) 3927{ 3928 int pidx; 3929 int npages; 3930 3931 npages = bp->b_npages; 3932 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3933 vm_page_lock_queues(); 3934 for (pidx = 0; pidx < npages; pidx++) 3935 vm_page_unhold(bp->b_pages[pidx]); 3936 vm_page_unlock_queues(); 3937 3938 bp->b_data = bp->b_saveaddr; 3939} 3940 3941void 3942bdone(struct buf *bp) 3943{ 3944 struct mtx *mtxp; 3945 3946 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3947 mtx_lock(mtxp); 3948 bp->b_flags |= B_DONE; 3949 wakeup(bp); 3950 mtx_unlock(mtxp); 3951} 3952 3953void 3954bwait(struct buf *bp, u_char pri, const char *wchan) 3955{ 3956 struct mtx *mtxp; 3957 3958 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3959 mtx_lock(mtxp); 3960 while ((bp->b_flags & B_DONE) == 0) 3961 msleep(bp, mtxp, pri, wchan, 0); 3962 mtx_unlock(mtxp); 3963} 3964 3965int 3966bufsync(struct bufobj *bo, int waitfor) 3967{ 3968 3969 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 3970} 3971 3972void 3973bufstrategy(struct bufobj *bo, struct buf *bp) 3974{ 3975 int i = 0; 3976 struct vnode *vp; 3977 3978 vp = bp->b_vp; 3979 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3980 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3981 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3982 i = VOP_STRATEGY(vp, bp); 3983 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3984} 3985 3986void 3987bufobj_wrefl(struct bufobj *bo) 3988{ 3989 3990 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3991 ASSERT_BO_LOCKED(bo); 3992 bo->bo_numoutput++; 3993} 3994 3995void 3996bufobj_wref(struct bufobj *bo) 3997{ 3998 3999 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4000 BO_LOCK(bo); 4001 bo->bo_numoutput++; 4002 BO_UNLOCK(bo); 4003} 4004 4005void 4006bufobj_wdrop(struct bufobj *bo) 4007{ 4008 4009 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4010 BO_LOCK(bo); 4011 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4012 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4013 bo->bo_flag &= ~BO_WWAIT; 4014 wakeup(&bo->bo_numoutput); 4015 } 4016 BO_UNLOCK(bo); 4017} 4018 4019int 4020bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4021{ 4022 int error; 4023 4024 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4025 ASSERT_BO_LOCKED(bo); 4026 error = 0; 4027 while (bo->bo_numoutput) { 4028 bo->bo_flag |= BO_WWAIT; 4029 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 4030 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4031 if (error) 4032 break; 4033 } 4034 return (error); 4035} 4036 4037void 4038bpin(struct buf *bp) 4039{ 4040 struct mtx *mtxp; 4041 4042 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4043 mtx_lock(mtxp); 4044 bp->b_pin_count++; 4045 mtx_unlock(mtxp); 4046} 4047 4048void 4049bunpin(struct buf *bp) 4050{ 4051 struct mtx *mtxp; 4052 4053 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4054 mtx_lock(mtxp); 4055 if (--bp->b_pin_count == 0) 4056 wakeup(bp); 4057 mtx_unlock(mtxp); 4058} 4059 4060void 4061bunpin_wait(struct buf *bp) 4062{ 4063 struct mtx *mtxp; 4064 4065 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4066 mtx_lock(mtxp); 4067 while (bp->b_pin_count > 0) 4068 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4069 mtx_unlock(mtxp); 4070} 4071 4072#include "opt_ddb.h" 4073#ifdef DDB 4074#include <ddb/ddb.h> 4075 4076/* DDB command to show buffer data */ 4077DB_SHOW_COMMAND(buffer, db_show_buffer) 4078{ 4079 /* get args */ 4080 struct buf *bp = (struct buf *)addr; 4081 4082 if (!have_addr) { 4083 db_printf("usage: show buffer <addr>\n"); 4084 return; 4085 } 4086 4087 db_printf("buf at %p\n", bp); 4088 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4089 db_printf( 4090 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4091 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n", 4092 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4093 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4094 bp->b_dep.lh_first); 4095 if (bp->b_npages) { 4096 int i; 4097 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4098 for (i = 0; i < bp->b_npages; i++) { 4099 vm_page_t m; 4100 m = bp->b_pages[i]; 4101 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4102 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4103 if ((i + 1) < bp->b_npages) 4104 db_printf(","); 4105 } 4106 db_printf("\n"); 4107 } 4108 db_printf(" "); 4109 lockmgr_printinfo(&bp->b_lock); 4110} 4111 4112DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4113{ 4114 struct buf *bp; 4115 int i; 4116 4117 for (i = 0; i < nbuf; i++) { 4118 bp = &buf[i]; 4119 if (BUF_ISLOCKED(bp)) { 4120 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4121 db_printf("\n"); 4122 } 4123 } 4124} 4125 4126DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4127{ 4128 struct vnode *vp; 4129 struct buf *bp; 4130 4131 if (!have_addr) { 4132 db_printf("usage: show vnodebufs <addr>\n"); 4133 return; 4134 } 4135 vp = (struct vnode *)addr; 4136 db_printf("Clean buffers:\n"); 4137 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4138 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4139 db_printf("\n"); 4140 } 4141 db_printf("Dirty buffers:\n"); 4142 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4143 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4144 db_printf("\n"); 4145 } 4146} 4147#endif /* DDB */ 4148