vfs_bio.c revision 192908
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 192908 2009-05-27 16:36:54Z zml $"); 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/bio.h> 47#include <sys/conf.h> 48#include <sys/buf.h> 49#include <sys/devicestat.h> 50#include <sys/eventhandler.h> 51#include <sys/fail.h> 52#include <sys/limits.h> 53#include <sys/lock.h> 54#include <sys/malloc.h> 55#include <sys/mount.h> 56#include <sys/mutex.h> 57#include <sys/kernel.h> 58#include <sys/kthread.h> 59#include <sys/proc.h> 60#include <sys/resourcevar.h> 61#include <sys/sysctl.h> 62#include <sys/vmmeter.h> 63#include <sys/vnode.h> 64#include <geom/geom.h> 65#include <vm/vm.h> 66#include <vm/vm_param.h> 67#include <vm/vm_kern.h> 68#include <vm/vm_pageout.h> 69#include <vm/vm_page.h> 70#include <vm/vm_object.h> 71#include <vm/vm_extern.h> 72#include <vm/vm_map.h> 73#include "opt_compat.h" 74#include "opt_directio.h" 75#include "opt_swap.h" 76 77static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 78 79struct bio_ops bioops; /* I/O operation notification */ 80 81struct buf_ops buf_ops_bio = { 82 .bop_name = "buf_ops_bio", 83 .bop_write = bufwrite, 84 .bop_strategy = bufstrategy, 85 .bop_sync = bufsync, 86 .bop_bdflush = bufbdflush, 87}; 88 89/* 90 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 91 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 92 */ 93struct buf *buf; /* buffer header pool */ 94 95static struct proc *bufdaemonproc; 96 97static int inmem(struct vnode *vp, daddr_t blkno); 98static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 99 vm_offset_t to); 100static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 101 vm_offset_t to); 102static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 103static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 104 vm_page_t m); 105static void vfs_clean_pages(struct buf *bp); 106static void vfs_setdirty(struct buf *bp); 107static void vfs_setdirty_locked_object(struct buf *bp); 108static void vfs_vmio_release(struct buf *bp); 109static int vfs_bio_clcheck(struct vnode *vp, int size, 110 daddr_t lblkno, daddr_t blkno); 111static int buf_do_flush(struct vnode *vp); 112static int flushbufqueues(struct vnode *, int, int); 113static void buf_daemon(void); 114static void bremfreel(struct buf *bp); 115#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 116 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 117static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 118#endif 119 120int vmiodirenable = TRUE; 121SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 122 "Use the VM system for directory writes"); 123long runningbufspace; 124SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 125 "Amount of presently outstanding async buffer io"); 126static long bufspace; 127#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 128 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 129SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 130 &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 131#else 132SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 133 "Virtual memory used for buffers"); 134#endif 135static long maxbufspace; 136SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 137 "Maximum allowed value of bufspace (including buf_daemon)"); 138static long bufmallocspace; 139SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 140 "Amount of malloced memory for buffers"); 141static long maxbufmallocspace; 142SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 143 "Maximum amount of malloced memory for buffers"); 144static long lobufspace; 145SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 146 "Minimum amount of buffers we want to have"); 147long hibufspace; 148SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 149 "Maximum allowed value of bufspace (excluding buf_daemon)"); 150static int bufreusecnt; 151SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 152 "Number of times we have reused a buffer"); 153static int buffreekvacnt; 154SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 155 "Number of times we have freed the KVA space from some buffer"); 156static int bufdefragcnt; 157SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 158 "Number of times we have had to repeat buffer allocation to defragment"); 159static long lorunningspace; 160SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 161 "Minimum preferred space used for in-progress I/O"); 162static long hirunningspace; 163SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 164 "Maximum amount of space to use for in-progress I/O"); 165int dirtybufferflushes; 166SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 167 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 168int bdwriteskip; 169SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 170 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 171int altbufferflushes; 172SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 173 0, "Number of fsync flushes to limit dirty buffers"); 174static int recursiveflushes; 175SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 176 0, "Number of flushes skipped due to being recursive"); 177static int numdirtybuffers; 178SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 179 "Number of buffers that are dirty (has unwritten changes) at the moment"); 180static int lodirtybuffers; 181SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 182 "How many buffers we want to have free before bufdaemon can sleep"); 183static int hidirtybuffers; 184SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 185 "When the number of dirty buffers is considered severe"); 186int dirtybufthresh; 187SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 188 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 189static int numfreebuffers; 190SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 191 "Number of free buffers"); 192static int lofreebuffers; 193SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 194 "XXX Unused"); 195static int hifreebuffers; 196SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 197 "XXX Complicatedly unused"); 198static int getnewbufcalls; 199SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 200 "Number of calls to getnewbuf"); 201static int getnewbufrestarts; 202SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 203 "Number of times getnewbuf has had to restart a buffer aquisition"); 204static int flushbufqtarget = 100; 205SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 206 "Amount of work to do in flushbufqueues when helping bufdaemon"); 207static long notbufdflashes; 208SYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0, 209 "Number of dirty buffer flushes done by the bufdaemon helpers"); 210 211/* 212 * Wakeup point for bufdaemon, as well as indicator of whether it is already 213 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 214 * is idling. 215 */ 216static int bd_request; 217 218/* 219 * This lock synchronizes access to bd_request. 220 */ 221static struct mtx bdlock; 222 223/* 224 * bogus page -- for I/O to/from partially complete buffers 225 * this is a temporary solution to the problem, but it is not 226 * really that bad. it would be better to split the buffer 227 * for input in the case of buffers partially already in memory, 228 * but the code is intricate enough already. 229 */ 230vm_page_t bogus_page; 231 232/* 233 * Synchronization (sleep/wakeup) variable for active buffer space requests. 234 * Set when wait starts, cleared prior to wakeup(). 235 * Used in runningbufwakeup() and waitrunningbufspace(). 236 */ 237static int runningbufreq; 238 239/* 240 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 241 * waitrunningbufspace(). 242 */ 243static struct mtx rbreqlock; 244 245/* 246 * Synchronization (sleep/wakeup) variable for buffer requests. 247 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 248 * by and/or. 249 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 250 * getnewbuf(), and getblk(). 251 */ 252static int needsbuffer; 253 254/* 255 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 256 */ 257static struct mtx nblock; 258 259/* 260 * Definitions for the buffer free lists. 261 */ 262#define BUFFER_QUEUES 6 /* number of free buffer queues */ 263 264#define QUEUE_NONE 0 /* on no queue */ 265#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 266#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 267#define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ 268#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 269#define QUEUE_EMPTY 5 /* empty buffer headers */ 270#define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 271 272/* Queues for free buffers with various properties */ 273static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 274 275/* Lock for the bufqueues */ 276static struct mtx bqlock; 277 278/* 279 * Single global constant for BUF_WMESG, to avoid getting multiple references. 280 * buf_wmesg is referred from macros. 281 */ 282const char *buf_wmesg = BUF_WMESG; 283 284#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 285#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 286#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 287#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 288 289#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 290 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 291static int 292sysctl_bufspace(SYSCTL_HANDLER_ARGS) 293{ 294 long lvalue; 295 int ivalue; 296 297 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 298 return (sysctl_handle_long(oidp, arg1, arg2, req)); 299 lvalue = *(long *)arg1; 300 if (lvalue > INT_MAX) 301 /* On overflow, still write out a long to trigger ENOMEM. */ 302 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 303 ivalue = lvalue; 304 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 305} 306#endif 307 308#ifdef DIRECTIO 309extern void ffs_rawread_setup(void); 310#endif /* DIRECTIO */ 311/* 312 * numdirtywakeup: 313 * 314 * If someone is blocked due to there being too many dirty buffers, 315 * and numdirtybuffers is now reasonable, wake them up. 316 */ 317 318static __inline void 319numdirtywakeup(int level) 320{ 321 322 if (numdirtybuffers <= level) { 323 mtx_lock(&nblock); 324 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 325 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 326 wakeup(&needsbuffer); 327 } 328 mtx_unlock(&nblock); 329 } 330} 331 332/* 333 * bufspacewakeup: 334 * 335 * Called when buffer space is potentially available for recovery. 336 * getnewbuf() will block on this flag when it is unable to free 337 * sufficient buffer space. Buffer space becomes recoverable when 338 * bp's get placed back in the queues. 339 */ 340 341static __inline void 342bufspacewakeup(void) 343{ 344 345 /* 346 * If someone is waiting for BUF space, wake them up. Even 347 * though we haven't freed the kva space yet, the waiting 348 * process will be able to now. 349 */ 350 mtx_lock(&nblock); 351 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 352 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 353 wakeup(&needsbuffer); 354 } 355 mtx_unlock(&nblock); 356} 357 358/* 359 * runningbufwakeup() - in-progress I/O accounting. 360 * 361 */ 362void 363runningbufwakeup(struct buf *bp) 364{ 365 366 if (bp->b_runningbufspace) { 367 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace); 368 bp->b_runningbufspace = 0; 369 mtx_lock(&rbreqlock); 370 if (runningbufreq && runningbufspace <= lorunningspace) { 371 runningbufreq = 0; 372 wakeup(&runningbufreq); 373 } 374 mtx_unlock(&rbreqlock); 375 } 376} 377 378/* 379 * bufcountwakeup: 380 * 381 * Called when a buffer has been added to one of the free queues to 382 * account for the buffer and to wakeup anyone waiting for free buffers. 383 * This typically occurs when large amounts of metadata are being handled 384 * by the buffer cache ( else buffer space runs out first, usually ). 385 */ 386 387static __inline void 388bufcountwakeup(void) 389{ 390 391 atomic_add_int(&numfreebuffers, 1); 392 mtx_lock(&nblock); 393 if (needsbuffer) { 394 needsbuffer &= ~VFS_BIO_NEED_ANY; 395 if (numfreebuffers >= hifreebuffers) 396 needsbuffer &= ~VFS_BIO_NEED_FREE; 397 wakeup(&needsbuffer); 398 } 399 mtx_unlock(&nblock); 400} 401 402/* 403 * waitrunningbufspace() 404 * 405 * runningbufspace is a measure of the amount of I/O currently 406 * running. This routine is used in async-write situations to 407 * prevent creating huge backups of pending writes to a device. 408 * Only asynchronous writes are governed by this function. 409 * 410 * Reads will adjust runningbufspace, but will not block based on it. 411 * The read load has a side effect of reducing the allowed write load. 412 * 413 * This does NOT turn an async write into a sync write. It waits 414 * for earlier writes to complete and generally returns before the 415 * caller's write has reached the device. 416 */ 417void 418waitrunningbufspace(void) 419{ 420 421 mtx_lock(&rbreqlock); 422 while (runningbufspace > hirunningspace) { 423 ++runningbufreq; 424 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 425 } 426 mtx_unlock(&rbreqlock); 427} 428 429 430/* 431 * vfs_buf_test_cache: 432 * 433 * Called when a buffer is extended. This function clears the B_CACHE 434 * bit if the newly extended portion of the buffer does not contain 435 * valid data. 436 */ 437static __inline 438void 439vfs_buf_test_cache(struct buf *bp, 440 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 441 vm_page_t m) 442{ 443 444 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 445 if (bp->b_flags & B_CACHE) { 446 int base = (foff + off) & PAGE_MASK; 447 if (vm_page_is_valid(m, base, size) == 0) 448 bp->b_flags &= ~B_CACHE; 449 } 450} 451 452/* Wake up the buffer daemon if necessary */ 453static __inline 454void 455bd_wakeup(int dirtybuflevel) 456{ 457 458 mtx_lock(&bdlock); 459 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 460 bd_request = 1; 461 wakeup(&bd_request); 462 } 463 mtx_unlock(&bdlock); 464} 465 466/* 467 * bd_speedup - speedup the buffer cache flushing code 468 */ 469 470static __inline 471void 472bd_speedup(void) 473{ 474 475 bd_wakeup(1); 476} 477 478/* 479 * Calculating buffer cache scaling values and reserve space for buffer 480 * headers. This is called during low level kernel initialization and 481 * may be called more then once. We CANNOT write to the memory area 482 * being reserved at this time. 483 */ 484caddr_t 485kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 486{ 487 int tuned_nbuf; 488 long maxbuf; 489 490 /* 491 * physmem_est is in pages. Convert it to kilobytes (assumes 492 * PAGE_SIZE is >= 1K) 493 */ 494 physmem_est = physmem_est * (PAGE_SIZE / 1024); 495 496 /* 497 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 498 * For the first 64MB of ram nominally allocate sufficient buffers to 499 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 500 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 501 * the buffer cache we limit the eventual kva reservation to 502 * maxbcache bytes. 503 * 504 * factor represents the 1/4 x ram conversion. 505 */ 506 if (nbuf == 0) { 507 int factor = 4 * BKVASIZE / 1024; 508 509 nbuf = 50; 510 if (physmem_est > 4096) 511 nbuf += min((physmem_est - 4096) / factor, 512 65536 / factor); 513 if (physmem_est > 65536) 514 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 515 516 if (maxbcache && nbuf > maxbcache / BKVASIZE) 517 nbuf = maxbcache / BKVASIZE; 518 tuned_nbuf = 1; 519 } else 520 tuned_nbuf = 0; 521 522 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 523 maxbuf = (LONG_MAX / 3) / BKVASIZE; 524 if (nbuf > maxbuf) { 525 if (!tuned_nbuf) 526 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 527 maxbuf); 528 nbuf = maxbuf; 529 } 530 531 /* 532 * swbufs are used as temporary holders for I/O, such as paging I/O. 533 * We have no less then 16 and no more then 256. 534 */ 535 nswbuf = max(min(nbuf/4, 256), 16); 536#ifdef NSWBUF_MIN 537 if (nswbuf < NSWBUF_MIN) 538 nswbuf = NSWBUF_MIN; 539#endif 540#ifdef DIRECTIO 541 ffs_rawread_setup(); 542#endif 543 544 /* 545 * Reserve space for the buffer cache buffers 546 */ 547 swbuf = (void *)v; 548 v = (caddr_t)(swbuf + nswbuf); 549 buf = (void *)v; 550 v = (caddr_t)(buf + nbuf); 551 552 return(v); 553} 554 555/* Initialize the buffer subsystem. Called before use of any buffers. */ 556void 557bufinit(void) 558{ 559 struct buf *bp; 560 int i; 561 562 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 563 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 564 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 565 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 566 567 /* next, make a null set of free lists */ 568 for (i = 0; i < BUFFER_QUEUES; i++) 569 TAILQ_INIT(&bufqueues[i]); 570 571 /* finally, initialize each buffer header and stick on empty q */ 572 for (i = 0; i < nbuf; i++) { 573 bp = &buf[i]; 574 bzero(bp, sizeof *bp); 575 bp->b_flags = B_INVAL; /* we're just an empty header */ 576 bp->b_rcred = NOCRED; 577 bp->b_wcred = NOCRED; 578 bp->b_qindex = QUEUE_EMPTY; 579 bp->b_vflags = 0; 580 bp->b_xflags = 0; 581 LIST_INIT(&bp->b_dep); 582 BUF_LOCKINIT(bp); 583 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 584 } 585 586 /* 587 * maxbufspace is the absolute maximum amount of buffer space we are 588 * allowed to reserve in KVM and in real terms. The absolute maximum 589 * is nominally used by buf_daemon. hibufspace is the nominal maximum 590 * used by most other processes. The differential is required to 591 * ensure that buf_daemon is able to run when other processes might 592 * be blocked waiting for buffer space. 593 * 594 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 595 * this may result in KVM fragmentation which is not handled optimally 596 * by the system. 597 */ 598 maxbufspace = (long)nbuf * BKVASIZE; 599 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 600 lobufspace = hibufspace - MAXBSIZE; 601 602 lorunningspace = 512 * 1024; 603 hirunningspace = 1024 * 1024; 604 605/* 606 * Limit the amount of malloc memory since it is wired permanently into 607 * the kernel space. Even though this is accounted for in the buffer 608 * allocation, we don't want the malloced region to grow uncontrolled. 609 * The malloc scheme improves memory utilization significantly on average 610 * (small) directories. 611 */ 612 maxbufmallocspace = hibufspace / 20; 613 614/* 615 * Reduce the chance of a deadlock occuring by limiting the number 616 * of delayed-write dirty buffers we allow to stack up. 617 */ 618 hidirtybuffers = nbuf / 4 + 20; 619 dirtybufthresh = hidirtybuffers * 9 / 10; 620 numdirtybuffers = 0; 621/* 622 * To support extreme low-memory systems, make sure hidirtybuffers cannot 623 * eat up all available buffer space. This occurs when our minimum cannot 624 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 625 * BKVASIZE'd (8K) buffers. 626 */ 627 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 628 hidirtybuffers >>= 1; 629 } 630 lodirtybuffers = hidirtybuffers / 2; 631 632/* 633 * Try to keep the number of free buffers in the specified range, 634 * and give special processes (e.g. like buf_daemon) access to an 635 * emergency reserve. 636 */ 637 lofreebuffers = nbuf / 18 + 5; 638 hifreebuffers = 2 * lofreebuffers; 639 numfreebuffers = nbuf; 640 641/* 642 * Maximum number of async ops initiated per buf_daemon loop. This is 643 * somewhat of a hack at the moment, we really need to limit ourselves 644 * based on the number of bytes of I/O in-transit that were initiated 645 * from buf_daemon. 646 */ 647 648 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 649 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 650} 651 652/* 653 * bfreekva() - free the kva allocation for a buffer. 654 * 655 * Since this call frees up buffer space, we call bufspacewakeup(). 656 */ 657static void 658bfreekva(struct buf *bp) 659{ 660 661 if (bp->b_kvasize) { 662 atomic_add_int(&buffreekvacnt, 1); 663 atomic_subtract_long(&bufspace, bp->b_kvasize); 664 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 665 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 666 bp->b_kvasize = 0; 667 bufspacewakeup(); 668 } 669} 670 671/* 672 * bremfree: 673 * 674 * Mark the buffer for removal from the appropriate free list in brelse. 675 * 676 */ 677void 678bremfree(struct buf *bp) 679{ 680 681 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 682 KASSERT((bp->b_flags & B_REMFREE) == 0, 683 ("bremfree: buffer %p already marked for delayed removal.", bp)); 684 KASSERT(bp->b_qindex != QUEUE_NONE, 685 ("bremfree: buffer %p not on a queue.", bp)); 686 BUF_ASSERT_HELD(bp); 687 688 bp->b_flags |= B_REMFREE; 689 /* Fixup numfreebuffers count. */ 690 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 691 atomic_subtract_int(&numfreebuffers, 1); 692} 693 694/* 695 * bremfreef: 696 * 697 * Force an immediate removal from a free list. Used only in nfs when 698 * it abuses the b_freelist pointer. 699 */ 700void 701bremfreef(struct buf *bp) 702{ 703 mtx_lock(&bqlock); 704 bremfreel(bp); 705 mtx_unlock(&bqlock); 706} 707 708/* 709 * bremfreel: 710 * 711 * Removes a buffer from the free list, must be called with the 712 * bqlock held. 713 */ 714static void 715bremfreel(struct buf *bp) 716{ 717 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 718 bp, bp->b_vp, bp->b_flags); 719 KASSERT(bp->b_qindex != QUEUE_NONE, 720 ("bremfreel: buffer %p not on a queue.", bp)); 721 BUF_ASSERT_HELD(bp); 722 mtx_assert(&bqlock, MA_OWNED); 723 724 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 725 bp->b_qindex = QUEUE_NONE; 726 /* 727 * If this was a delayed bremfree() we only need to remove the buffer 728 * from the queue and return the stats are already done. 729 */ 730 if (bp->b_flags & B_REMFREE) { 731 bp->b_flags &= ~B_REMFREE; 732 return; 733 } 734 /* 735 * Fixup numfreebuffers count. If the buffer is invalid or not 736 * delayed-write, the buffer was free and we must decrement 737 * numfreebuffers. 738 */ 739 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 740 atomic_subtract_int(&numfreebuffers, 1); 741} 742 743 744/* 745 * Get a buffer with the specified data. Look in the cache first. We 746 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 747 * is set, the buffer is valid and we do not have to do anything ( see 748 * getblk() ). This is really just a special case of breadn(). 749 */ 750int 751bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 752 struct buf **bpp) 753{ 754 755 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 756} 757 758/* 759 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 760 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 761 * the buffer is valid and we do not have to do anything. 762 */ 763void 764breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 765 int cnt, struct ucred * cred) 766{ 767 struct buf *rabp; 768 int i; 769 770 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 771 if (inmem(vp, *rablkno)) 772 continue; 773 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 774 775 if ((rabp->b_flags & B_CACHE) == 0) { 776 if (!TD_IS_IDLETHREAD(curthread)) 777 curthread->td_ru.ru_inblock++; 778 rabp->b_flags |= B_ASYNC; 779 rabp->b_flags &= ~B_INVAL; 780 rabp->b_ioflags &= ~BIO_ERROR; 781 rabp->b_iocmd = BIO_READ; 782 if (rabp->b_rcred == NOCRED && cred != NOCRED) 783 rabp->b_rcred = crhold(cred); 784 vfs_busy_pages(rabp, 0); 785 BUF_KERNPROC(rabp); 786 rabp->b_iooffset = dbtob(rabp->b_blkno); 787 bstrategy(rabp); 788 } else { 789 brelse(rabp); 790 } 791 } 792} 793 794/* 795 * Operates like bread, but also starts asynchronous I/O on 796 * read-ahead blocks. 797 */ 798int 799breadn(struct vnode * vp, daddr_t blkno, int size, 800 daddr_t * rablkno, int *rabsize, 801 int cnt, struct ucred * cred, struct buf **bpp) 802{ 803 struct buf *bp; 804 int rv = 0, readwait = 0; 805 806 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 807 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 808 809 /* if not found in cache, do some I/O */ 810 if ((bp->b_flags & B_CACHE) == 0) { 811 if (!TD_IS_IDLETHREAD(curthread)) 812 curthread->td_ru.ru_inblock++; 813 bp->b_iocmd = BIO_READ; 814 bp->b_flags &= ~B_INVAL; 815 bp->b_ioflags &= ~BIO_ERROR; 816 if (bp->b_rcred == NOCRED && cred != NOCRED) 817 bp->b_rcred = crhold(cred); 818 vfs_busy_pages(bp, 0); 819 bp->b_iooffset = dbtob(bp->b_blkno); 820 bstrategy(bp); 821 ++readwait; 822 } 823 824 breada(vp, rablkno, rabsize, cnt, cred); 825 826 if (readwait) { 827 rv = bufwait(bp); 828 } 829 return (rv); 830} 831 832/* 833 * Write, release buffer on completion. (Done by iodone 834 * if async). Do not bother writing anything if the buffer 835 * is invalid. 836 * 837 * Note that we set B_CACHE here, indicating that buffer is 838 * fully valid and thus cacheable. This is true even of NFS 839 * now so we set it generally. This could be set either here 840 * or in biodone() since the I/O is synchronous. We put it 841 * here. 842 */ 843int 844bufwrite(struct buf *bp) 845{ 846 int oldflags; 847 struct vnode *vp; 848 int vp_md; 849 850 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 851 if (bp->b_flags & B_INVAL) { 852 brelse(bp); 853 return (0); 854 } 855 856 oldflags = bp->b_flags; 857 858 BUF_ASSERT_HELD(bp); 859 860 if (bp->b_pin_count > 0) 861 bunpin_wait(bp); 862 863 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 864 ("FFS background buffer should not get here %p", bp)); 865 866 vp = bp->b_vp; 867 if (vp) 868 vp_md = vp->v_vflag & VV_MD; 869 else 870 vp_md = 0; 871 872 /* Mark the buffer clean */ 873 bundirty(bp); 874 875 bp->b_flags &= ~B_DONE; 876 bp->b_ioflags &= ~BIO_ERROR; 877 bp->b_flags |= B_CACHE; 878 bp->b_iocmd = BIO_WRITE; 879 880 bufobj_wref(bp->b_bufobj); 881 vfs_busy_pages(bp, 1); 882 883 /* 884 * Normal bwrites pipeline writes 885 */ 886 bp->b_runningbufspace = bp->b_bufsize; 887 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 888 889 if (!TD_IS_IDLETHREAD(curthread)) 890 curthread->td_ru.ru_oublock++; 891 if (oldflags & B_ASYNC) 892 BUF_KERNPROC(bp); 893 bp->b_iooffset = dbtob(bp->b_blkno); 894 bstrategy(bp); 895 896 if ((oldflags & B_ASYNC) == 0) { 897 int rtval = bufwait(bp); 898 brelse(bp); 899 return (rtval); 900 } else { 901 /* 902 * don't allow the async write to saturate the I/O 903 * system. We will not deadlock here because 904 * we are blocking waiting for I/O that is already in-progress 905 * to complete. We do not block here if it is the update 906 * or syncer daemon trying to clean up as that can lead 907 * to deadlock. 908 */ 909 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 910 waitrunningbufspace(); 911 } 912 913 return (0); 914} 915 916void 917bufbdflush(struct bufobj *bo, struct buf *bp) 918{ 919 struct buf *nbp; 920 921 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 922 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 923 altbufferflushes++; 924 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 925 BO_LOCK(bo); 926 /* 927 * Try to find a buffer to flush. 928 */ 929 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 930 if ((nbp->b_vflags & BV_BKGRDINPROG) || 931 BUF_LOCK(nbp, 932 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 933 continue; 934 if (bp == nbp) 935 panic("bdwrite: found ourselves"); 936 BO_UNLOCK(bo); 937 /* Don't countdeps with the bo lock held. */ 938 if (buf_countdeps(nbp, 0)) { 939 BO_LOCK(bo); 940 BUF_UNLOCK(nbp); 941 continue; 942 } 943 if (nbp->b_flags & B_CLUSTEROK) { 944 vfs_bio_awrite(nbp); 945 } else { 946 bremfree(nbp); 947 bawrite(nbp); 948 } 949 dirtybufferflushes++; 950 break; 951 } 952 if (nbp == NULL) 953 BO_UNLOCK(bo); 954 } 955} 956 957/* 958 * Delayed write. (Buffer is marked dirty). Do not bother writing 959 * anything if the buffer is marked invalid. 960 * 961 * Note that since the buffer must be completely valid, we can safely 962 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 963 * biodone() in order to prevent getblk from writing the buffer 964 * out synchronously. 965 */ 966void 967bdwrite(struct buf *bp) 968{ 969 struct thread *td = curthread; 970 struct vnode *vp; 971 struct bufobj *bo; 972 973 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 974 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 975 BUF_ASSERT_HELD(bp); 976 977 if (bp->b_flags & B_INVAL) { 978 brelse(bp); 979 return; 980 } 981 982 /* 983 * If we have too many dirty buffers, don't create any more. 984 * If we are wildly over our limit, then force a complete 985 * cleanup. Otherwise, just keep the situation from getting 986 * out of control. Note that we have to avoid a recursive 987 * disaster and not try to clean up after our own cleanup! 988 */ 989 vp = bp->b_vp; 990 bo = bp->b_bufobj; 991 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 992 td->td_pflags |= TDP_INBDFLUSH; 993 BO_BDFLUSH(bo, bp); 994 td->td_pflags &= ~TDP_INBDFLUSH; 995 } else 996 recursiveflushes++; 997 998 bdirty(bp); 999 /* 1000 * Set B_CACHE, indicating that the buffer is fully valid. This is 1001 * true even of NFS now. 1002 */ 1003 bp->b_flags |= B_CACHE; 1004 1005 /* 1006 * This bmap keeps the system from needing to do the bmap later, 1007 * perhaps when the system is attempting to do a sync. Since it 1008 * is likely that the indirect block -- or whatever other datastructure 1009 * that the filesystem needs is still in memory now, it is a good 1010 * thing to do this. Note also, that if the pageout daemon is 1011 * requesting a sync -- there might not be enough memory to do 1012 * the bmap then... So, this is important to do. 1013 */ 1014 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1015 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1016 } 1017 1018 /* 1019 * Set the *dirty* buffer range based upon the VM system dirty pages. 1020 */ 1021 vfs_setdirty(bp); 1022 1023 /* 1024 * We need to do this here to satisfy the vnode_pager and the 1025 * pageout daemon, so that it thinks that the pages have been 1026 * "cleaned". Note that since the pages are in a delayed write 1027 * buffer -- the VFS layer "will" see that the pages get written 1028 * out on the next sync, or perhaps the cluster will be completed. 1029 */ 1030 vfs_clean_pages(bp); 1031 bqrelse(bp); 1032 1033 /* 1034 * Wakeup the buffer flushing daemon if we have a lot of dirty 1035 * buffers (midpoint between our recovery point and our stall 1036 * point). 1037 */ 1038 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1039 1040 /* 1041 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1042 * due to the softdep code. 1043 */ 1044} 1045 1046/* 1047 * bdirty: 1048 * 1049 * Turn buffer into delayed write request. We must clear BIO_READ and 1050 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1051 * itself to properly update it in the dirty/clean lists. We mark it 1052 * B_DONE to ensure that any asynchronization of the buffer properly 1053 * clears B_DONE ( else a panic will occur later ). 1054 * 1055 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1056 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1057 * should only be called if the buffer is known-good. 1058 * 1059 * Since the buffer is not on a queue, we do not update the numfreebuffers 1060 * count. 1061 * 1062 * The buffer must be on QUEUE_NONE. 1063 */ 1064void 1065bdirty(struct buf *bp) 1066{ 1067 1068 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1069 bp, bp->b_vp, bp->b_flags); 1070 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1071 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1072 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1073 BUF_ASSERT_HELD(bp); 1074 bp->b_flags &= ~(B_RELBUF); 1075 bp->b_iocmd = BIO_WRITE; 1076 1077 if ((bp->b_flags & B_DELWRI) == 0) { 1078 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1079 reassignbuf(bp); 1080 atomic_add_int(&numdirtybuffers, 1); 1081 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1082 } 1083} 1084 1085/* 1086 * bundirty: 1087 * 1088 * Clear B_DELWRI for buffer. 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 */ 1095 1096void 1097bundirty(struct buf *bp) 1098{ 1099 1100 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", 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 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1104 BUF_ASSERT_HELD(bp); 1105 1106 if (bp->b_flags & B_DELWRI) { 1107 bp->b_flags &= ~B_DELWRI; 1108 reassignbuf(bp); 1109 atomic_subtract_int(&numdirtybuffers, 1); 1110 numdirtywakeup(lodirtybuffers); 1111 } 1112 /* 1113 * Since it is now being written, we can clear its deferred write flag. 1114 */ 1115 bp->b_flags &= ~B_DEFERRED; 1116} 1117 1118/* 1119 * bawrite: 1120 * 1121 * Asynchronous write. Start output on a buffer, but do not wait for 1122 * it to complete. The buffer is released when the output completes. 1123 * 1124 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1125 * B_INVAL buffers. Not us. 1126 */ 1127void 1128bawrite(struct buf *bp) 1129{ 1130 1131 bp->b_flags |= B_ASYNC; 1132 (void) bwrite(bp); 1133} 1134 1135/* 1136 * bwillwrite: 1137 * 1138 * Called prior to the locking of any vnodes when we are expecting to 1139 * write. We do not want to starve the buffer cache with too many 1140 * dirty buffers so we block here. By blocking prior to the locking 1141 * of any vnodes we attempt to avoid the situation where a locked vnode 1142 * prevents the various system daemons from flushing related buffers. 1143 */ 1144 1145void 1146bwillwrite(void) 1147{ 1148 1149 if (numdirtybuffers >= hidirtybuffers) { 1150 mtx_lock(&nblock); 1151 while (numdirtybuffers >= hidirtybuffers) { 1152 bd_wakeup(1); 1153 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1154 msleep(&needsbuffer, &nblock, 1155 (PRIBIO + 4), "flswai", 0); 1156 } 1157 mtx_unlock(&nblock); 1158 } 1159} 1160 1161/* 1162 * Return true if we have too many dirty buffers. 1163 */ 1164int 1165buf_dirty_count_severe(void) 1166{ 1167 1168 return(numdirtybuffers >= hidirtybuffers); 1169} 1170 1171static __noinline int 1172buf_vm_page_count_severe(void) 1173{ 1174 1175 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1176 1177 return vm_page_count_severe(); 1178} 1179 1180/* 1181 * brelse: 1182 * 1183 * Release a busy buffer and, if requested, free its resources. The 1184 * buffer will be stashed in the appropriate bufqueue[] allowing it 1185 * to be accessed later as a cache entity or reused for other purposes. 1186 */ 1187void 1188brelse(struct buf *bp) 1189{ 1190 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1191 bp, bp->b_vp, bp->b_flags); 1192 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1193 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1194 1195 if (bp->b_flags & B_MANAGED) { 1196 bqrelse(bp); 1197 return; 1198 } 1199 1200 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1201 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1202 /* 1203 * Failed write, redirty. Must clear BIO_ERROR to prevent 1204 * pages from being scrapped. If the error is anything 1205 * other than an I/O error (EIO), assume that retryingi 1206 * is futile. 1207 */ 1208 bp->b_ioflags &= ~BIO_ERROR; 1209 bdirty(bp); 1210 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1211 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1212 /* 1213 * Either a failed I/O or we were asked to free or not 1214 * cache the buffer. 1215 */ 1216 bp->b_flags |= B_INVAL; 1217 if (!LIST_EMPTY(&bp->b_dep)) 1218 buf_deallocate(bp); 1219 if (bp->b_flags & B_DELWRI) { 1220 atomic_subtract_int(&numdirtybuffers, 1); 1221 numdirtywakeup(lodirtybuffers); 1222 } 1223 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1224 if ((bp->b_flags & B_VMIO) == 0) { 1225 if (bp->b_bufsize) 1226 allocbuf(bp, 0); 1227 if (bp->b_vp) 1228 brelvp(bp); 1229 } 1230 } 1231 1232 /* 1233 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1234 * is called with B_DELWRI set, the underlying pages may wind up 1235 * getting freed causing a previous write (bdwrite()) to get 'lost' 1236 * because pages associated with a B_DELWRI bp are marked clean. 1237 * 1238 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1239 * if B_DELWRI is set. 1240 * 1241 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1242 * on pages to return pages to the VM page queues. 1243 */ 1244 if (bp->b_flags & B_DELWRI) 1245 bp->b_flags &= ~B_RELBUF; 1246 else if (buf_vm_page_count_severe()) { 1247 /* 1248 * The locking of the BO_LOCK is not necessary since 1249 * BKGRDINPROG cannot be set while we hold the buf 1250 * lock, it can only be cleared if it is already 1251 * pending. 1252 */ 1253 if (bp->b_vp) { 1254 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1255 bp->b_flags |= B_RELBUF; 1256 } else 1257 bp->b_flags |= B_RELBUF; 1258 } 1259 1260 /* 1261 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1262 * constituted, not even NFS buffers now. Two flags effect this. If 1263 * B_INVAL, the struct buf is invalidated but the VM object is kept 1264 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1265 * 1266 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1267 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1268 * buffer is also B_INVAL because it hits the re-dirtying code above. 1269 * 1270 * Normally we can do this whether a buffer is B_DELWRI or not. If 1271 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1272 * the commit state and we cannot afford to lose the buffer. If the 1273 * buffer has a background write in progress, we need to keep it 1274 * around to prevent it from being reconstituted and starting a second 1275 * background write. 1276 */ 1277 if ((bp->b_flags & B_VMIO) 1278 && !(bp->b_vp->v_mount != NULL && 1279 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1280 !vn_isdisk(bp->b_vp, NULL) && 1281 (bp->b_flags & B_DELWRI)) 1282 ) { 1283 1284 int i, j, resid; 1285 vm_page_t m; 1286 off_t foff; 1287 vm_pindex_t poff; 1288 vm_object_t obj; 1289 1290 obj = bp->b_bufobj->bo_object; 1291 1292 /* 1293 * Get the base offset and length of the buffer. Note that 1294 * in the VMIO case if the buffer block size is not 1295 * page-aligned then b_data pointer may not be page-aligned. 1296 * But our b_pages[] array *IS* page aligned. 1297 * 1298 * block sizes less then DEV_BSIZE (usually 512) are not 1299 * supported due to the page granularity bits (m->valid, 1300 * m->dirty, etc...). 1301 * 1302 * See man buf(9) for more information 1303 */ 1304 resid = bp->b_bufsize; 1305 foff = bp->b_offset; 1306 VM_OBJECT_LOCK(obj); 1307 for (i = 0; i < bp->b_npages; i++) { 1308 int had_bogus = 0; 1309 1310 m = bp->b_pages[i]; 1311 1312 /* 1313 * If we hit a bogus page, fixup *all* the bogus pages 1314 * now. 1315 */ 1316 if (m == bogus_page) { 1317 poff = OFF_TO_IDX(bp->b_offset); 1318 had_bogus = 1; 1319 1320 for (j = i; j < bp->b_npages; j++) { 1321 vm_page_t mtmp; 1322 mtmp = bp->b_pages[j]; 1323 if (mtmp == bogus_page) { 1324 mtmp = vm_page_lookup(obj, poff + j); 1325 if (!mtmp) { 1326 panic("brelse: page missing\n"); 1327 } 1328 bp->b_pages[j] = mtmp; 1329 } 1330 } 1331 1332 if ((bp->b_flags & B_INVAL) == 0) { 1333 pmap_qenter( 1334 trunc_page((vm_offset_t)bp->b_data), 1335 bp->b_pages, bp->b_npages); 1336 } 1337 m = bp->b_pages[i]; 1338 } 1339 if ((bp->b_flags & B_NOCACHE) || 1340 (bp->b_ioflags & BIO_ERROR)) { 1341 int poffset = foff & PAGE_MASK; 1342 int presid = resid > (PAGE_SIZE - poffset) ? 1343 (PAGE_SIZE - poffset) : resid; 1344 1345 KASSERT(presid >= 0, ("brelse: extra page")); 1346 vm_page_lock_queues(); 1347 vm_page_set_invalid(m, poffset, presid); 1348 vm_page_unlock_queues(); 1349 if (had_bogus) 1350 printf("avoided corruption bug in bogus_page/brelse code\n"); 1351 } 1352 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1353 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1354 } 1355 VM_OBJECT_UNLOCK(obj); 1356 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1357 vfs_vmio_release(bp); 1358 1359 } else if (bp->b_flags & B_VMIO) { 1360 1361 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1362 vfs_vmio_release(bp); 1363 } 1364 1365 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1366 if (bp->b_bufsize != 0) 1367 allocbuf(bp, 0); 1368 if (bp->b_vp != NULL) 1369 brelvp(bp); 1370 } 1371 1372 if (BUF_LOCKRECURSED(bp)) { 1373 /* do not release to free list */ 1374 BUF_UNLOCK(bp); 1375 return; 1376 } 1377 1378 /* enqueue */ 1379 mtx_lock(&bqlock); 1380 /* Handle delayed bremfree() processing. */ 1381 if (bp->b_flags & B_REMFREE) 1382 bremfreel(bp); 1383 if (bp->b_qindex != QUEUE_NONE) 1384 panic("brelse: free buffer onto another queue???"); 1385 1386 /* 1387 * If the buffer has junk contents signal it and eventually 1388 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1389 * doesn't find it. 1390 */ 1391 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1392 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1393 bp->b_flags |= B_INVAL; 1394 if (bp->b_flags & B_INVAL) { 1395 if (bp->b_flags & B_DELWRI) 1396 bundirty(bp); 1397 if (bp->b_vp) 1398 brelvp(bp); 1399 } 1400 1401 /* buffers with no memory */ 1402 if (bp->b_bufsize == 0) { 1403 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1404 if (bp->b_vflags & BV_BKGRDINPROG) 1405 panic("losing buffer 1"); 1406 if (bp->b_kvasize) { 1407 bp->b_qindex = QUEUE_EMPTYKVA; 1408 } else { 1409 bp->b_qindex = QUEUE_EMPTY; 1410 } 1411 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1412 /* buffers with junk contents */ 1413 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1414 (bp->b_ioflags & BIO_ERROR)) { 1415 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1416 if (bp->b_vflags & BV_BKGRDINPROG) 1417 panic("losing buffer 2"); 1418 bp->b_qindex = QUEUE_CLEAN; 1419 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1420 /* remaining buffers */ 1421 } else { 1422 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1423 (B_DELWRI|B_NEEDSGIANT)) 1424 bp->b_qindex = QUEUE_DIRTY_GIANT; 1425 else if (bp->b_flags & B_DELWRI) 1426 bp->b_qindex = QUEUE_DIRTY; 1427 else 1428 bp->b_qindex = QUEUE_CLEAN; 1429 if (bp->b_flags & B_AGE) 1430 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1431 else 1432 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1433 } 1434 mtx_unlock(&bqlock); 1435 1436 /* 1437 * Fixup numfreebuffers count. The bp is on an appropriate queue 1438 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1439 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1440 * if B_INVAL is set ). 1441 */ 1442 1443 if (!(bp->b_flags & B_DELWRI)) 1444 bufcountwakeup(); 1445 1446 /* 1447 * Something we can maybe free or reuse 1448 */ 1449 if (bp->b_bufsize || bp->b_kvasize) 1450 bufspacewakeup(); 1451 1452 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1453 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1454 panic("brelse: not dirty"); 1455 /* unlock */ 1456 BUF_UNLOCK(bp); 1457} 1458 1459/* 1460 * Release a buffer back to the appropriate queue but do not try to free 1461 * it. The buffer is expected to be used again soon. 1462 * 1463 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1464 * biodone() to requeue an async I/O on completion. It is also used when 1465 * known good buffers need to be requeued but we think we may need the data 1466 * again soon. 1467 * 1468 * XXX we should be able to leave the B_RELBUF hint set on completion. 1469 */ 1470void 1471bqrelse(struct buf *bp) 1472{ 1473 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1474 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1475 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1476 1477 if (BUF_LOCKRECURSED(bp)) { 1478 /* do not release to free list */ 1479 BUF_UNLOCK(bp); 1480 return; 1481 } 1482 1483 if (bp->b_flags & B_MANAGED) { 1484 if (bp->b_flags & B_REMFREE) { 1485 mtx_lock(&bqlock); 1486 bremfreel(bp); 1487 mtx_unlock(&bqlock); 1488 } 1489 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1490 BUF_UNLOCK(bp); 1491 return; 1492 } 1493 1494 mtx_lock(&bqlock); 1495 /* Handle delayed bremfree() processing. */ 1496 if (bp->b_flags & B_REMFREE) 1497 bremfreel(bp); 1498 if (bp->b_qindex != QUEUE_NONE) 1499 panic("bqrelse: free buffer onto another queue???"); 1500 /* buffers with stale but valid contents */ 1501 if (bp->b_flags & B_DELWRI) { 1502 if (bp->b_flags & B_NEEDSGIANT) 1503 bp->b_qindex = QUEUE_DIRTY_GIANT; 1504 else 1505 bp->b_qindex = QUEUE_DIRTY; 1506 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1507 } else { 1508 /* 1509 * The locking of the BO_LOCK for checking of the 1510 * BV_BKGRDINPROG is not necessary since the 1511 * BV_BKGRDINPROG cannot be set while we hold the buf 1512 * lock, it can only be cleared if it is already 1513 * pending. 1514 */ 1515 if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1516 bp->b_qindex = QUEUE_CLEAN; 1517 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1518 b_freelist); 1519 } else { 1520 /* 1521 * We are too low on memory, we have to try to free 1522 * the buffer (most importantly: the wired pages 1523 * making up its backing store) *now*. 1524 */ 1525 mtx_unlock(&bqlock); 1526 brelse(bp); 1527 return; 1528 } 1529 } 1530 mtx_unlock(&bqlock); 1531 1532 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1533 bufcountwakeup(); 1534 1535 /* 1536 * Something we can maybe free or reuse. 1537 */ 1538 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1539 bufspacewakeup(); 1540 1541 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1542 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1543 panic("bqrelse: not dirty"); 1544 /* unlock */ 1545 BUF_UNLOCK(bp); 1546} 1547 1548/* Give pages used by the bp back to the VM system (where possible) */ 1549static void 1550vfs_vmio_release(struct buf *bp) 1551{ 1552 int i; 1553 vm_page_t m; 1554 1555 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1556 vm_page_lock_queues(); 1557 for (i = 0; i < bp->b_npages; i++) { 1558 m = bp->b_pages[i]; 1559 bp->b_pages[i] = NULL; 1560 /* 1561 * In order to keep page LRU ordering consistent, put 1562 * everything on the inactive queue. 1563 */ 1564 vm_page_unwire(m, 0); 1565 /* 1566 * We don't mess with busy pages, it is 1567 * the responsibility of the process that 1568 * busied the pages to deal with them. 1569 */ 1570 if ((m->oflags & VPO_BUSY) || (m->busy != 0)) 1571 continue; 1572 1573 if (m->wire_count == 0) { 1574 /* 1575 * Might as well free the page if we can and it has 1576 * no valid data. We also free the page if the 1577 * buffer was used for direct I/O 1578 */ 1579 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1580 m->hold_count == 0) { 1581 vm_page_free(m); 1582 } else if (bp->b_flags & B_DIRECT) { 1583 vm_page_try_to_free(m); 1584 } else if (buf_vm_page_count_severe()) { 1585 vm_page_try_to_cache(m); 1586 } 1587 } 1588 } 1589 vm_page_unlock_queues(); 1590 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1591 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1592 1593 if (bp->b_bufsize) { 1594 bufspacewakeup(); 1595 bp->b_bufsize = 0; 1596 } 1597 bp->b_npages = 0; 1598 bp->b_flags &= ~B_VMIO; 1599 if (bp->b_vp) 1600 brelvp(bp); 1601} 1602 1603/* 1604 * Check to see if a block at a particular lbn is available for a clustered 1605 * write. 1606 */ 1607static int 1608vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1609{ 1610 struct buf *bpa; 1611 int match; 1612 1613 match = 0; 1614 1615 /* If the buf isn't in core skip it */ 1616 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1617 return (0); 1618 1619 /* If the buf is busy we don't want to wait for it */ 1620 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1621 return (0); 1622 1623 /* Only cluster with valid clusterable delayed write buffers */ 1624 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1625 (B_DELWRI | B_CLUSTEROK)) 1626 goto done; 1627 1628 if (bpa->b_bufsize != size) 1629 goto done; 1630 1631 /* 1632 * Check to see if it is in the expected place on disk and that the 1633 * block has been mapped. 1634 */ 1635 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1636 match = 1; 1637done: 1638 BUF_UNLOCK(bpa); 1639 return (match); 1640} 1641 1642/* 1643 * vfs_bio_awrite: 1644 * 1645 * Implement clustered async writes for clearing out B_DELWRI buffers. 1646 * This is much better then the old way of writing only one buffer at 1647 * a time. Note that we may not be presented with the buffers in the 1648 * correct order, so we search for the cluster in both directions. 1649 */ 1650int 1651vfs_bio_awrite(struct buf *bp) 1652{ 1653 struct bufobj *bo; 1654 int i; 1655 int j; 1656 daddr_t lblkno = bp->b_lblkno; 1657 struct vnode *vp = bp->b_vp; 1658 int ncl; 1659 int nwritten; 1660 int size; 1661 int maxcl; 1662 1663 bo = &vp->v_bufobj; 1664 /* 1665 * right now we support clustered writing only to regular files. If 1666 * we find a clusterable block we could be in the middle of a cluster 1667 * rather then at the beginning. 1668 */ 1669 if ((vp->v_type == VREG) && 1670 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1671 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1672 1673 size = vp->v_mount->mnt_stat.f_iosize; 1674 maxcl = MAXPHYS / size; 1675 1676 BO_LOCK(bo); 1677 for (i = 1; i < maxcl; i++) 1678 if (vfs_bio_clcheck(vp, size, lblkno + i, 1679 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1680 break; 1681 1682 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1683 if (vfs_bio_clcheck(vp, size, lblkno - j, 1684 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1685 break; 1686 BO_UNLOCK(bo); 1687 --j; 1688 ncl = i + j; 1689 /* 1690 * this is a possible cluster write 1691 */ 1692 if (ncl != 1) { 1693 BUF_UNLOCK(bp); 1694 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1695 return nwritten; 1696 } 1697 } 1698 bremfree(bp); 1699 bp->b_flags |= B_ASYNC; 1700 /* 1701 * default (old) behavior, writing out only one block 1702 * 1703 * XXX returns b_bufsize instead of b_bcount for nwritten? 1704 */ 1705 nwritten = bp->b_bufsize; 1706 (void) bwrite(bp); 1707 1708 return nwritten; 1709} 1710 1711/* 1712 * getnewbuf: 1713 * 1714 * Find and initialize a new buffer header, freeing up existing buffers 1715 * in the bufqueues as necessary. The new buffer is returned locked. 1716 * 1717 * Important: B_INVAL is not set. If the caller wishes to throw the 1718 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1719 * 1720 * We block if: 1721 * We have insufficient buffer headers 1722 * We have insufficient buffer space 1723 * buffer_map is too fragmented ( space reservation fails ) 1724 * If we have to flush dirty buffers ( but we try to avoid this ) 1725 * 1726 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1727 * Instead we ask the buf daemon to do it for us. We attempt to 1728 * avoid piecemeal wakeups of the pageout daemon. 1729 */ 1730 1731static struct buf * 1732getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 1733 int gbflags) 1734{ 1735 struct thread *td; 1736 struct buf *bp; 1737 struct buf *nbp; 1738 int defrag = 0; 1739 int nqindex; 1740 static int flushingbufs; 1741 1742 td = curthread; 1743 /* 1744 * We can't afford to block since we might be holding a vnode lock, 1745 * which may prevent system daemons from running. We deal with 1746 * low-memory situations by proactively returning memory and running 1747 * async I/O rather then sync I/O. 1748 */ 1749 atomic_add_int(&getnewbufcalls, 1); 1750 atomic_subtract_int(&getnewbufrestarts, 1); 1751restart: 1752 atomic_add_int(&getnewbufrestarts, 1); 1753 1754 /* 1755 * Setup for scan. If we do not have enough free buffers, 1756 * we setup a degenerate case that immediately fails. Note 1757 * that if we are specially marked process, we are allowed to 1758 * dip into our reserves. 1759 * 1760 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1761 * 1762 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1763 * However, there are a number of cases (defragging, reusing, ...) 1764 * where we cannot backup. 1765 */ 1766 mtx_lock(&bqlock); 1767 nqindex = QUEUE_EMPTYKVA; 1768 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1769 1770 if (nbp == NULL) { 1771 /* 1772 * If no EMPTYKVA buffers and we are either 1773 * defragging or reusing, locate a CLEAN buffer 1774 * to free or reuse. If bufspace useage is low 1775 * skip this step so we can allocate a new buffer. 1776 */ 1777 if (defrag || bufspace >= lobufspace) { 1778 nqindex = QUEUE_CLEAN; 1779 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1780 } 1781 1782 /* 1783 * If we could not find or were not allowed to reuse a 1784 * CLEAN buffer, check to see if it is ok to use an EMPTY 1785 * buffer. We can only use an EMPTY buffer if allocating 1786 * its KVA would not otherwise run us out of buffer space. 1787 */ 1788 if (nbp == NULL && defrag == 0 && 1789 bufspace + maxsize < hibufspace) { 1790 nqindex = QUEUE_EMPTY; 1791 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1792 } 1793 } 1794 1795 /* 1796 * Run scan, possibly freeing data and/or kva mappings on the fly 1797 * depending. 1798 */ 1799 1800 while ((bp = nbp) != NULL) { 1801 int qindex = nqindex; 1802 1803 /* 1804 * Calculate next bp ( we can only use it if we do not block 1805 * or do other fancy things ). 1806 */ 1807 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1808 switch(qindex) { 1809 case QUEUE_EMPTY: 1810 nqindex = QUEUE_EMPTYKVA; 1811 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1812 break; 1813 /* FALLTHROUGH */ 1814 case QUEUE_EMPTYKVA: 1815 nqindex = QUEUE_CLEAN; 1816 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1817 break; 1818 /* FALLTHROUGH */ 1819 case QUEUE_CLEAN: 1820 /* 1821 * nbp is NULL. 1822 */ 1823 break; 1824 } 1825 } 1826 /* 1827 * If we are defragging then we need a buffer with 1828 * b_kvasize != 0. XXX this situation should no longer 1829 * occur, if defrag is non-zero the buffer's b_kvasize 1830 * should also be non-zero at this point. XXX 1831 */ 1832 if (defrag && bp->b_kvasize == 0) { 1833 printf("Warning: defrag empty buffer %p\n", bp); 1834 continue; 1835 } 1836 1837 /* 1838 * Start freeing the bp. This is somewhat involved. nbp 1839 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1840 */ 1841 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1842 continue; 1843 if (bp->b_vp) { 1844 BO_LOCK(bp->b_bufobj); 1845 if (bp->b_vflags & BV_BKGRDINPROG) { 1846 BO_UNLOCK(bp->b_bufobj); 1847 BUF_UNLOCK(bp); 1848 continue; 1849 } 1850 BO_UNLOCK(bp->b_bufobj); 1851 } 1852 CTR6(KTR_BUF, 1853 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1854 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1855 bp->b_kvasize, bp->b_bufsize, qindex); 1856 1857 /* 1858 * Sanity Checks 1859 */ 1860 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1861 1862 /* 1863 * Note: we no longer distinguish between VMIO and non-VMIO 1864 * buffers. 1865 */ 1866 1867 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1868 1869 bremfreel(bp); 1870 mtx_unlock(&bqlock); 1871 1872 if (qindex == QUEUE_CLEAN) { 1873 if (bp->b_flags & B_VMIO) { 1874 bp->b_flags &= ~B_ASYNC; 1875 vfs_vmio_release(bp); 1876 } 1877 if (bp->b_vp) 1878 brelvp(bp); 1879 } 1880 1881 /* 1882 * NOTE: nbp is now entirely invalid. We can only restart 1883 * the scan from this point on. 1884 * 1885 * Get the rest of the buffer freed up. b_kva* is still 1886 * valid after this operation. 1887 */ 1888 1889 if (bp->b_rcred != NOCRED) { 1890 crfree(bp->b_rcred); 1891 bp->b_rcred = NOCRED; 1892 } 1893 if (bp->b_wcred != NOCRED) { 1894 crfree(bp->b_wcred); 1895 bp->b_wcred = NOCRED; 1896 } 1897 if (!LIST_EMPTY(&bp->b_dep)) 1898 buf_deallocate(bp); 1899 if (bp->b_vflags & BV_BKGRDINPROG) 1900 panic("losing buffer 3"); 1901 KASSERT(bp->b_vp == NULL, 1902 ("bp: %p still has vnode %p. qindex: %d", 1903 bp, bp->b_vp, qindex)); 1904 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1905 ("bp: %p still on a buffer list. xflags %X", 1906 bp, bp->b_xflags)); 1907 1908 if (bp->b_bufsize) 1909 allocbuf(bp, 0); 1910 1911 bp->b_flags = 0; 1912 bp->b_ioflags = 0; 1913 bp->b_xflags = 0; 1914 bp->b_vflags = 0; 1915 bp->b_vp = NULL; 1916 bp->b_blkno = bp->b_lblkno = 0; 1917 bp->b_offset = NOOFFSET; 1918 bp->b_iodone = 0; 1919 bp->b_error = 0; 1920 bp->b_resid = 0; 1921 bp->b_bcount = 0; 1922 bp->b_npages = 0; 1923 bp->b_dirtyoff = bp->b_dirtyend = 0; 1924 bp->b_bufobj = NULL; 1925 bp->b_pin_count = 0; 1926 bp->b_fsprivate1 = NULL; 1927 bp->b_fsprivate2 = NULL; 1928 bp->b_fsprivate3 = NULL; 1929 1930 LIST_INIT(&bp->b_dep); 1931 1932 /* 1933 * If we are defragging then free the buffer. 1934 */ 1935 if (defrag) { 1936 bp->b_flags |= B_INVAL; 1937 bfreekva(bp); 1938 brelse(bp); 1939 defrag = 0; 1940 goto restart; 1941 } 1942 1943 /* 1944 * Notify any waiters for the buffer lock about 1945 * identity change by freeing the buffer. 1946 */ 1947 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 1948 bp->b_flags |= B_INVAL; 1949 bfreekva(bp); 1950 brelse(bp); 1951 goto restart; 1952 } 1953 1954 /* 1955 * If we are overcomitted then recover the buffer and its 1956 * KVM space. This occurs in rare situations when multiple 1957 * processes are blocked in getnewbuf() or allocbuf(). 1958 */ 1959 if (bufspace >= hibufspace) 1960 flushingbufs = 1; 1961 if (flushingbufs && bp->b_kvasize != 0) { 1962 bp->b_flags |= B_INVAL; 1963 bfreekva(bp); 1964 brelse(bp); 1965 goto restart; 1966 } 1967 if (bufspace < lobufspace) 1968 flushingbufs = 0; 1969 break; 1970 } 1971 1972 /* 1973 * If we exhausted our list, sleep as appropriate. We may have to 1974 * wakeup various daemons and write out some dirty buffers. 1975 * 1976 * Generally we are sleeping due to insufficient buffer space. 1977 */ 1978 1979 if (bp == NULL) { 1980 int flags, norunbuf; 1981 char *waitmsg; 1982 int fl; 1983 1984 if (defrag) { 1985 flags = VFS_BIO_NEED_BUFSPACE; 1986 waitmsg = "nbufkv"; 1987 } else if (bufspace >= hibufspace) { 1988 waitmsg = "nbufbs"; 1989 flags = VFS_BIO_NEED_BUFSPACE; 1990 } else { 1991 waitmsg = "newbuf"; 1992 flags = VFS_BIO_NEED_ANY; 1993 } 1994 mtx_lock(&nblock); 1995 needsbuffer |= flags; 1996 mtx_unlock(&nblock); 1997 mtx_unlock(&bqlock); 1998 1999 bd_speedup(); /* heeeelp */ 2000 if (gbflags & GB_NOWAIT_BD) 2001 return (NULL); 2002 2003 mtx_lock(&nblock); 2004 while (needsbuffer & flags) { 2005 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2006 mtx_unlock(&nblock); 2007 /* 2008 * getblk() is called with a vnode 2009 * locked, and some majority of the 2010 * dirty buffers may as well belong to 2011 * the vnode. Flushing the buffers 2012 * there would make a progress that 2013 * cannot be achieved by the 2014 * buf_daemon, that cannot lock the 2015 * vnode. 2016 */ 2017 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2018 (td->td_pflags & TDP_NORUNNINGBUF); 2019 /* play bufdaemon */ 2020 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2021 fl = buf_do_flush(vp); 2022 td->td_pflags &= norunbuf; 2023 mtx_lock(&nblock); 2024 if (fl != 0) 2025 continue; 2026 if ((needsbuffer & flags) == 0) 2027 break; 2028 } 2029 if (msleep(&needsbuffer, &nblock, 2030 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2031 mtx_unlock(&nblock); 2032 return (NULL); 2033 } 2034 } 2035 mtx_unlock(&nblock); 2036 } else { 2037 /* 2038 * We finally have a valid bp. We aren't quite out of the 2039 * woods, we still have to reserve kva space. In order 2040 * to keep fragmentation sane we only allocate kva in 2041 * BKVASIZE chunks. 2042 */ 2043 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2044 2045 if (maxsize != bp->b_kvasize) { 2046 vm_offset_t addr = 0; 2047 2048 bfreekva(bp); 2049 2050 vm_map_lock(buffer_map); 2051 if (vm_map_findspace(buffer_map, 2052 vm_map_min(buffer_map), maxsize, &addr)) { 2053 /* 2054 * Uh oh. Buffer map is to fragmented. We 2055 * must defragment the map. 2056 */ 2057 atomic_add_int(&bufdefragcnt, 1); 2058 vm_map_unlock(buffer_map); 2059 defrag = 1; 2060 bp->b_flags |= B_INVAL; 2061 brelse(bp); 2062 goto restart; 2063 } 2064 if (addr) { 2065 vm_map_insert(buffer_map, NULL, 0, 2066 addr, addr + maxsize, 2067 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2068 2069 bp->b_kvabase = (caddr_t) addr; 2070 bp->b_kvasize = maxsize; 2071 atomic_add_long(&bufspace, bp->b_kvasize); 2072 atomic_add_int(&bufreusecnt, 1); 2073 } 2074 vm_map_unlock(buffer_map); 2075 } 2076 bp->b_saveaddr = bp->b_kvabase; 2077 bp->b_data = bp->b_saveaddr; 2078 } 2079 return(bp); 2080} 2081 2082/* 2083 * buf_daemon: 2084 * 2085 * buffer flushing daemon. Buffers are normally flushed by the 2086 * update daemon but if it cannot keep up this process starts to 2087 * take the load in an attempt to prevent getnewbuf() from blocking. 2088 */ 2089 2090static struct kproc_desc buf_kp = { 2091 "bufdaemon", 2092 buf_daemon, 2093 &bufdaemonproc 2094}; 2095SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2096 2097static int 2098buf_do_flush(struct vnode *vp) 2099{ 2100 int flushed; 2101 2102 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2103 /* The list empty check here is slightly racy */ 2104 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2105 mtx_lock(&Giant); 2106 flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0); 2107 mtx_unlock(&Giant); 2108 } 2109 if (flushed == 0) { 2110 /* 2111 * Could not find any buffers without rollback 2112 * dependencies, so just write the first one 2113 * in the hopes of eventually making progress. 2114 */ 2115 flushbufqueues(vp, QUEUE_DIRTY, 1); 2116 if (!TAILQ_EMPTY( 2117 &bufqueues[QUEUE_DIRTY_GIANT])) { 2118 mtx_lock(&Giant); 2119 flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1); 2120 mtx_unlock(&Giant); 2121 } 2122 } 2123 return (flushed); 2124} 2125 2126static void 2127buf_daemon() 2128{ 2129 2130 /* 2131 * This process needs to be suspended prior to shutdown sync. 2132 */ 2133 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2134 SHUTDOWN_PRI_LAST); 2135 2136 /* 2137 * This process is allowed to take the buffer cache to the limit 2138 */ 2139 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2140 mtx_lock(&bdlock); 2141 for (;;) { 2142 bd_request = 0; 2143 mtx_unlock(&bdlock); 2144 2145 kproc_suspend_check(bufdaemonproc); 2146 2147 /* 2148 * Do the flush. Limit the amount of in-transit I/O we 2149 * allow to build up, otherwise we would completely saturate 2150 * the I/O system. Wakeup any waiting processes before we 2151 * normally would so they can run in parallel with our drain. 2152 */ 2153 while (numdirtybuffers > lodirtybuffers) { 2154 if (buf_do_flush(NULL) == 0) 2155 break; 2156 uio_yield(); 2157 } 2158 2159 /* 2160 * Only clear bd_request if we have reached our low water 2161 * mark. The buf_daemon normally waits 1 second and 2162 * then incrementally flushes any dirty buffers that have 2163 * built up, within reason. 2164 * 2165 * If we were unable to hit our low water mark and couldn't 2166 * find any flushable buffers, we sleep half a second. 2167 * Otherwise we loop immediately. 2168 */ 2169 mtx_lock(&bdlock); 2170 if (numdirtybuffers <= lodirtybuffers) { 2171 /* 2172 * We reached our low water mark, reset the 2173 * request and sleep until we are needed again. 2174 * The sleep is just so the suspend code works. 2175 */ 2176 bd_request = 0; 2177 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2178 } else { 2179 /* 2180 * We couldn't find any flushable dirty buffers but 2181 * still have too many dirty buffers, we 2182 * have to sleep and try again. (rare) 2183 */ 2184 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2185 } 2186 } 2187} 2188 2189/* 2190 * flushbufqueues: 2191 * 2192 * Try to flush a buffer in the dirty queue. We must be careful to 2193 * free up B_INVAL buffers instead of write them, which NFS is 2194 * particularly sensitive to. 2195 */ 2196static int flushwithdeps = 0; 2197SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2198 0, "Number of buffers flushed with dependecies that require rollbacks"); 2199 2200static int 2201flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2202{ 2203 struct buf *sentinel; 2204 struct vnode *vp; 2205 struct mount *mp; 2206 struct buf *bp; 2207 int hasdeps; 2208 int flushed; 2209 int target; 2210 2211 if (lvp == NULL) { 2212 target = numdirtybuffers - lodirtybuffers; 2213 if (flushdeps && target > 2) 2214 target /= 2; 2215 } else 2216 target = flushbufqtarget; 2217 flushed = 0; 2218 bp = NULL; 2219 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2220 sentinel->b_qindex = QUEUE_SENTINEL; 2221 mtx_lock(&bqlock); 2222 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2223 while (flushed != target) { 2224 bp = TAILQ_NEXT(sentinel, b_freelist); 2225 if (bp != NULL) { 2226 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2227 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2228 b_freelist); 2229 } else 2230 break; 2231 /* 2232 * Skip sentinels inserted by other invocations of the 2233 * flushbufqueues(), taking care to not reorder them. 2234 */ 2235 if (bp->b_qindex == QUEUE_SENTINEL) 2236 continue; 2237 /* 2238 * Only flush the buffers that belong to the 2239 * vnode locked by the curthread. 2240 */ 2241 if (lvp != NULL && bp->b_vp != lvp) 2242 continue; 2243 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2244 continue; 2245 if (bp->b_pin_count > 0) { 2246 BUF_UNLOCK(bp); 2247 continue; 2248 } 2249 BO_LOCK(bp->b_bufobj); 2250 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2251 (bp->b_flags & B_DELWRI) == 0) { 2252 BO_UNLOCK(bp->b_bufobj); 2253 BUF_UNLOCK(bp); 2254 continue; 2255 } 2256 BO_UNLOCK(bp->b_bufobj); 2257 if (bp->b_flags & B_INVAL) { 2258 bremfreel(bp); 2259 mtx_unlock(&bqlock); 2260 brelse(bp); 2261 flushed++; 2262 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2263 mtx_lock(&bqlock); 2264 continue; 2265 } 2266 2267 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2268 if (flushdeps == 0) { 2269 BUF_UNLOCK(bp); 2270 continue; 2271 } 2272 hasdeps = 1; 2273 } else 2274 hasdeps = 0; 2275 /* 2276 * We must hold the lock on a vnode before writing 2277 * one of its buffers. Otherwise we may confuse, or 2278 * in the case of a snapshot vnode, deadlock the 2279 * system. 2280 * 2281 * The lock order here is the reverse of the normal 2282 * of vnode followed by buf lock. This is ok because 2283 * the NOWAIT will prevent deadlock. 2284 */ 2285 vp = bp->b_vp; 2286 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2287 BUF_UNLOCK(bp); 2288 continue; 2289 } 2290 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2291 mtx_unlock(&bqlock); 2292 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2293 bp, bp->b_vp, bp->b_flags); 2294 if (curproc == bufdaemonproc) 2295 vfs_bio_awrite(bp); 2296 else { 2297 bremfree(bp); 2298 bwrite(bp); 2299 notbufdflashes++; 2300 } 2301 vn_finished_write(mp); 2302 VOP_UNLOCK(vp, 0); 2303 flushwithdeps += hasdeps; 2304 flushed++; 2305 2306 /* 2307 * Sleeping on runningbufspace while holding 2308 * vnode lock leads to deadlock. 2309 */ 2310 if (curproc == bufdaemonproc) 2311 waitrunningbufspace(); 2312 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2313 mtx_lock(&bqlock); 2314 continue; 2315 } 2316 vn_finished_write(mp); 2317 BUF_UNLOCK(bp); 2318 } 2319 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2320 mtx_unlock(&bqlock); 2321 free(sentinel, M_TEMP); 2322 return (flushed); 2323} 2324 2325/* 2326 * Check to see if a block is currently memory resident. 2327 */ 2328struct buf * 2329incore(struct bufobj *bo, daddr_t blkno) 2330{ 2331 struct buf *bp; 2332 2333 BO_LOCK(bo); 2334 bp = gbincore(bo, blkno); 2335 BO_UNLOCK(bo); 2336 return (bp); 2337} 2338 2339/* 2340 * Returns true if no I/O is needed to access the 2341 * associated VM object. This is like incore except 2342 * it also hunts around in the VM system for the data. 2343 */ 2344 2345static int 2346inmem(struct vnode * vp, daddr_t blkno) 2347{ 2348 vm_object_t obj; 2349 vm_offset_t toff, tinc, size; 2350 vm_page_t m; 2351 vm_ooffset_t off; 2352 2353 ASSERT_VOP_LOCKED(vp, "inmem"); 2354 2355 if (incore(&vp->v_bufobj, blkno)) 2356 return 1; 2357 if (vp->v_mount == NULL) 2358 return 0; 2359 obj = vp->v_object; 2360 if (obj == NULL) 2361 return (0); 2362 2363 size = PAGE_SIZE; 2364 if (size > vp->v_mount->mnt_stat.f_iosize) 2365 size = vp->v_mount->mnt_stat.f_iosize; 2366 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2367 2368 VM_OBJECT_LOCK(obj); 2369 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2370 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2371 if (!m) 2372 goto notinmem; 2373 tinc = size; 2374 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2375 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2376 if (vm_page_is_valid(m, 2377 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2378 goto notinmem; 2379 } 2380 VM_OBJECT_UNLOCK(obj); 2381 return 1; 2382 2383notinmem: 2384 VM_OBJECT_UNLOCK(obj); 2385 return (0); 2386} 2387 2388/* 2389 * vfs_setdirty: 2390 * 2391 * Sets the dirty range for a buffer based on the status of the dirty 2392 * bits in the pages comprising the buffer. 2393 * 2394 * The range is limited to the size of the buffer. 2395 * 2396 * This routine is primarily used by NFS, but is generalized for the 2397 * B_VMIO case. 2398 */ 2399static void 2400vfs_setdirty(struct buf *bp) 2401{ 2402 2403 /* 2404 * Degenerate case - empty buffer 2405 */ 2406 2407 if (bp->b_bufsize == 0) 2408 return; 2409 2410 /* 2411 * We qualify the scan for modified pages on whether the 2412 * object has been flushed yet. 2413 */ 2414 2415 if ((bp->b_flags & B_VMIO) == 0) 2416 return; 2417 2418 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2419 vfs_setdirty_locked_object(bp); 2420 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2421} 2422 2423static void 2424vfs_setdirty_locked_object(struct buf *bp) 2425{ 2426 vm_object_t object; 2427 int i; 2428 2429 object = bp->b_bufobj->bo_object; 2430 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2431 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2432 vm_offset_t boffset; 2433 vm_offset_t eoffset; 2434 2435 vm_page_lock_queues(); 2436 /* 2437 * test the pages to see if they have been modified directly 2438 * by users through the VM system. 2439 */ 2440 for (i = 0; i < bp->b_npages; i++) 2441 vm_page_test_dirty(bp->b_pages[i]); 2442 2443 /* 2444 * Calculate the encompassing dirty range, boffset and eoffset, 2445 * (eoffset - boffset) bytes. 2446 */ 2447 2448 for (i = 0; i < bp->b_npages; i++) { 2449 if (bp->b_pages[i]->dirty) 2450 break; 2451 } 2452 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2453 2454 for (i = bp->b_npages - 1; i >= 0; --i) { 2455 if (bp->b_pages[i]->dirty) { 2456 break; 2457 } 2458 } 2459 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2460 2461 vm_page_unlock_queues(); 2462 /* 2463 * Fit it to the buffer. 2464 */ 2465 2466 if (eoffset > bp->b_bcount) 2467 eoffset = bp->b_bcount; 2468 2469 /* 2470 * If we have a good dirty range, merge with the existing 2471 * dirty range. 2472 */ 2473 2474 if (boffset < eoffset) { 2475 if (bp->b_dirtyoff > boffset) 2476 bp->b_dirtyoff = boffset; 2477 if (bp->b_dirtyend < eoffset) 2478 bp->b_dirtyend = eoffset; 2479 } 2480 } 2481} 2482 2483/* 2484 * getblk: 2485 * 2486 * Get a block given a specified block and offset into a file/device. 2487 * The buffers B_DONE bit will be cleared on return, making it almost 2488 * ready for an I/O initiation. B_INVAL may or may not be set on 2489 * return. The caller should clear B_INVAL prior to initiating a 2490 * READ. 2491 * 2492 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2493 * an existing buffer. 2494 * 2495 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2496 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2497 * and then cleared based on the backing VM. If the previous buffer is 2498 * non-0-sized but invalid, B_CACHE will be cleared. 2499 * 2500 * If getblk() must create a new buffer, the new buffer is returned with 2501 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2502 * case it is returned with B_INVAL clear and B_CACHE set based on the 2503 * backing VM. 2504 * 2505 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2506 * B_CACHE bit is clear. 2507 * 2508 * What this means, basically, is that the caller should use B_CACHE to 2509 * determine whether the buffer is fully valid or not and should clear 2510 * B_INVAL prior to issuing a read. If the caller intends to validate 2511 * the buffer by loading its data area with something, the caller needs 2512 * to clear B_INVAL. If the caller does this without issuing an I/O, 2513 * the caller should set B_CACHE ( as an optimization ), else the caller 2514 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2515 * a write attempt or if it was a successfull read. If the caller 2516 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2517 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2518 */ 2519struct buf * 2520getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2521 int flags) 2522{ 2523 struct buf *bp; 2524 struct bufobj *bo; 2525 int error; 2526 2527 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2528 ASSERT_VOP_LOCKED(vp, "getblk"); 2529 if (size > MAXBSIZE) 2530 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2531 2532 bo = &vp->v_bufobj; 2533loop: 2534 /* 2535 * Block if we are low on buffers. Certain processes are allowed 2536 * to completely exhaust the buffer cache. 2537 * 2538 * If this check ever becomes a bottleneck it may be better to 2539 * move it into the else, when gbincore() fails. At the moment 2540 * it isn't a problem. 2541 * 2542 * XXX remove if 0 sections (clean this up after its proven) 2543 */ 2544 if (numfreebuffers == 0) { 2545 if (TD_IS_IDLETHREAD(curthread)) 2546 return NULL; 2547 mtx_lock(&nblock); 2548 needsbuffer |= VFS_BIO_NEED_ANY; 2549 mtx_unlock(&nblock); 2550 } 2551 2552 BO_LOCK(bo); 2553 bp = gbincore(bo, blkno); 2554 if (bp != NULL) { 2555 int lockflags; 2556 /* 2557 * Buffer is in-core. If the buffer is not busy, it must 2558 * be on a queue. 2559 */ 2560 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2561 2562 if (flags & GB_LOCK_NOWAIT) 2563 lockflags |= LK_NOWAIT; 2564 2565 error = BUF_TIMELOCK(bp, lockflags, 2566 BO_MTX(bo), "getblk", slpflag, slptimeo); 2567 2568 /* 2569 * If we slept and got the lock we have to restart in case 2570 * the buffer changed identities. 2571 */ 2572 if (error == ENOLCK) 2573 goto loop; 2574 /* We timed out or were interrupted. */ 2575 else if (error) 2576 return (NULL); 2577 2578 /* 2579 * The buffer is locked. B_CACHE is cleared if the buffer is 2580 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2581 * and for a VMIO buffer B_CACHE is adjusted according to the 2582 * backing VM cache. 2583 */ 2584 if (bp->b_flags & B_INVAL) 2585 bp->b_flags &= ~B_CACHE; 2586 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2587 bp->b_flags |= B_CACHE; 2588 bremfree(bp); 2589 2590 /* 2591 * check for size inconsistancies for non-VMIO case. 2592 */ 2593 2594 if (bp->b_bcount != size) { 2595 if ((bp->b_flags & B_VMIO) == 0 || 2596 (size > bp->b_kvasize)) { 2597 if (bp->b_flags & B_DELWRI) { 2598 /* 2599 * If buffer is pinned and caller does 2600 * not want sleep waiting for it to be 2601 * unpinned, bail out 2602 * */ 2603 if (bp->b_pin_count > 0) { 2604 if (flags & GB_LOCK_NOWAIT) { 2605 bqrelse(bp); 2606 return (NULL); 2607 } else { 2608 bunpin_wait(bp); 2609 } 2610 } 2611 bp->b_flags |= B_NOCACHE; 2612 bwrite(bp); 2613 } else { 2614 if (LIST_EMPTY(&bp->b_dep)) { 2615 bp->b_flags |= B_RELBUF; 2616 brelse(bp); 2617 } else { 2618 bp->b_flags |= B_NOCACHE; 2619 bwrite(bp); 2620 } 2621 } 2622 goto loop; 2623 } 2624 } 2625 2626 /* 2627 * If the size is inconsistant in the VMIO case, we can resize 2628 * the buffer. This might lead to B_CACHE getting set or 2629 * cleared. If the size has not changed, B_CACHE remains 2630 * unchanged from its previous state. 2631 */ 2632 2633 if (bp->b_bcount != size) 2634 allocbuf(bp, size); 2635 2636 KASSERT(bp->b_offset != NOOFFSET, 2637 ("getblk: no buffer offset")); 2638 2639 /* 2640 * A buffer with B_DELWRI set and B_CACHE clear must 2641 * be committed before we can return the buffer in 2642 * order to prevent the caller from issuing a read 2643 * ( due to B_CACHE not being set ) and overwriting 2644 * it. 2645 * 2646 * Most callers, including NFS and FFS, need this to 2647 * operate properly either because they assume they 2648 * can issue a read if B_CACHE is not set, or because 2649 * ( for example ) an uncached B_DELWRI might loop due 2650 * to softupdates re-dirtying the buffer. In the latter 2651 * case, B_CACHE is set after the first write completes, 2652 * preventing further loops. 2653 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2654 * above while extending the buffer, we cannot allow the 2655 * buffer to remain with B_CACHE set after the write 2656 * completes or it will represent a corrupt state. To 2657 * deal with this we set B_NOCACHE to scrap the buffer 2658 * after the write. 2659 * 2660 * We might be able to do something fancy, like setting 2661 * B_CACHE in bwrite() except if B_DELWRI is already set, 2662 * so the below call doesn't set B_CACHE, but that gets real 2663 * confusing. This is much easier. 2664 */ 2665 2666 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2667 bp->b_flags |= B_NOCACHE; 2668 bwrite(bp); 2669 goto loop; 2670 } 2671 bp->b_flags &= ~B_DONE; 2672 } else { 2673 int bsize, maxsize, vmio; 2674 off_t offset; 2675 2676 /* 2677 * Buffer is not in-core, create new buffer. The buffer 2678 * returned by getnewbuf() is locked. Note that the returned 2679 * buffer is also considered valid (not marked B_INVAL). 2680 */ 2681 BO_UNLOCK(bo); 2682 /* 2683 * If the user does not want us to create the buffer, bail out 2684 * here. 2685 */ 2686 if (flags & GB_NOCREAT) 2687 return NULL; 2688 bsize = bo->bo_bsize; 2689 offset = blkno * bsize; 2690 vmio = vp->v_object != NULL; 2691 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2692 maxsize = imax(maxsize, bsize); 2693 2694 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 2695 if (bp == NULL) { 2696 if (slpflag || slptimeo) 2697 return NULL; 2698 goto loop; 2699 } 2700 2701 /* 2702 * This code is used to make sure that a buffer is not 2703 * created while the getnewbuf routine is blocked. 2704 * This can be a problem whether the vnode is locked or not. 2705 * If the buffer is created out from under us, we have to 2706 * throw away the one we just created. 2707 * 2708 * Note: this must occur before we associate the buffer 2709 * with the vp especially considering limitations in 2710 * the splay tree implementation when dealing with duplicate 2711 * lblkno's. 2712 */ 2713 BO_LOCK(bo); 2714 if (gbincore(bo, blkno)) { 2715 BO_UNLOCK(bo); 2716 bp->b_flags |= B_INVAL; 2717 brelse(bp); 2718 goto loop; 2719 } 2720 2721 /* 2722 * Insert the buffer into the hash, so that it can 2723 * be found by incore. 2724 */ 2725 bp->b_blkno = bp->b_lblkno = blkno; 2726 bp->b_offset = offset; 2727 bgetvp(vp, bp); 2728 BO_UNLOCK(bo); 2729 2730 /* 2731 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2732 * buffer size starts out as 0, B_CACHE will be set by 2733 * allocbuf() for the VMIO case prior to it testing the 2734 * backing store for validity. 2735 */ 2736 2737 if (vmio) { 2738 bp->b_flags |= B_VMIO; 2739#if defined(VFS_BIO_DEBUG) 2740 if (vn_canvmio(vp) != TRUE) 2741 printf("getblk: VMIO on vnode type %d\n", 2742 vp->v_type); 2743#endif 2744 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2745 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2746 bp, vp->v_object, bp->b_bufobj->bo_object)); 2747 } else { 2748 bp->b_flags &= ~B_VMIO; 2749 KASSERT(bp->b_bufobj->bo_object == NULL, 2750 ("ARGH! has b_bufobj->bo_object %p %p\n", 2751 bp, bp->b_bufobj->bo_object)); 2752 } 2753 2754 allocbuf(bp, size); 2755 bp->b_flags &= ~B_DONE; 2756 } 2757 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2758 BUF_ASSERT_HELD(bp); 2759 KASSERT(bp->b_bufobj == bo, 2760 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2761 return (bp); 2762} 2763 2764/* 2765 * Get an empty, disassociated buffer of given size. The buffer is initially 2766 * set to B_INVAL. 2767 */ 2768struct buf * 2769geteblk(int size, int flags) 2770{ 2771 struct buf *bp; 2772 int maxsize; 2773 2774 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2775 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 2776 if ((flags & GB_NOWAIT_BD) && 2777 (curthread->td_pflags & TDP_BUFNEED) != 0) 2778 return (NULL); 2779 } 2780 allocbuf(bp, size); 2781 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2782 BUF_ASSERT_HELD(bp); 2783 return (bp); 2784} 2785 2786 2787/* 2788 * This code constitutes the buffer memory from either anonymous system 2789 * memory (in the case of non-VMIO operations) or from an associated 2790 * VM object (in the case of VMIO operations). This code is able to 2791 * resize a buffer up or down. 2792 * 2793 * Note that this code is tricky, and has many complications to resolve 2794 * deadlock or inconsistant data situations. Tread lightly!!! 2795 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2796 * the caller. Calling this code willy nilly can result in the loss of data. 2797 * 2798 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2799 * B_CACHE for the non-VMIO case. 2800 */ 2801 2802int 2803allocbuf(struct buf *bp, int size) 2804{ 2805 int newbsize, mbsize; 2806 int i; 2807 2808 BUF_ASSERT_HELD(bp); 2809 2810 if (bp->b_kvasize < size) 2811 panic("allocbuf: buffer too small"); 2812 2813 if ((bp->b_flags & B_VMIO) == 0) { 2814 caddr_t origbuf; 2815 int origbufsize; 2816 /* 2817 * Just get anonymous memory from the kernel. Don't 2818 * mess with B_CACHE. 2819 */ 2820 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2821 if (bp->b_flags & B_MALLOC) 2822 newbsize = mbsize; 2823 else 2824 newbsize = round_page(size); 2825 2826 if (newbsize < bp->b_bufsize) { 2827 /* 2828 * malloced buffers are not shrunk 2829 */ 2830 if (bp->b_flags & B_MALLOC) { 2831 if (newbsize) { 2832 bp->b_bcount = size; 2833 } else { 2834 free(bp->b_data, M_BIOBUF); 2835 if (bp->b_bufsize) { 2836 atomic_subtract_long( 2837 &bufmallocspace, 2838 bp->b_bufsize); 2839 bufspacewakeup(); 2840 bp->b_bufsize = 0; 2841 } 2842 bp->b_saveaddr = bp->b_kvabase; 2843 bp->b_data = bp->b_saveaddr; 2844 bp->b_bcount = 0; 2845 bp->b_flags &= ~B_MALLOC; 2846 } 2847 return 1; 2848 } 2849 vm_hold_free_pages( 2850 bp, 2851 (vm_offset_t) bp->b_data + newbsize, 2852 (vm_offset_t) bp->b_data + bp->b_bufsize); 2853 } else if (newbsize > bp->b_bufsize) { 2854 /* 2855 * We only use malloced memory on the first allocation. 2856 * and revert to page-allocated memory when the buffer 2857 * grows. 2858 */ 2859 /* 2860 * There is a potential smp race here that could lead 2861 * to bufmallocspace slightly passing the max. It 2862 * is probably extremely rare and not worth worrying 2863 * over. 2864 */ 2865 if ( (bufmallocspace < maxbufmallocspace) && 2866 (bp->b_bufsize == 0) && 2867 (mbsize <= PAGE_SIZE/2)) { 2868 2869 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2870 bp->b_bufsize = mbsize; 2871 bp->b_bcount = size; 2872 bp->b_flags |= B_MALLOC; 2873 atomic_add_long(&bufmallocspace, mbsize); 2874 return 1; 2875 } 2876 origbuf = NULL; 2877 origbufsize = 0; 2878 /* 2879 * If the buffer is growing on its other-than-first allocation, 2880 * then we revert to the page-allocation scheme. 2881 */ 2882 if (bp->b_flags & B_MALLOC) { 2883 origbuf = bp->b_data; 2884 origbufsize = bp->b_bufsize; 2885 bp->b_data = bp->b_kvabase; 2886 if (bp->b_bufsize) { 2887 atomic_subtract_long(&bufmallocspace, 2888 bp->b_bufsize); 2889 bufspacewakeup(); 2890 bp->b_bufsize = 0; 2891 } 2892 bp->b_flags &= ~B_MALLOC; 2893 newbsize = round_page(newbsize); 2894 } 2895 vm_hold_load_pages( 2896 bp, 2897 (vm_offset_t) bp->b_data + bp->b_bufsize, 2898 (vm_offset_t) bp->b_data + newbsize); 2899 if (origbuf) { 2900 bcopy(origbuf, bp->b_data, origbufsize); 2901 free(origbuf, M_BIOBUF); 2902 } 2903 } 2904 } else { 2905 int desiredpages; 2906 2907 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2908 desiredpages = (size == 0) ? 0 : 2909 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2910 2911 if (bp->b_flags & B_MALLOC) 2912 panic("allocbuf: VMIO buffer can't be malloced"); 2913 /* 2914 * Set B_CACHE initially if buffer is 0 length or will become 2915 * 0-length. 2916 */ 2917 if (size == 0 || bp->b_bufsize == 0) 2918 bp->b_flags |= B_CACHE; 2919 2920 if (newbsize < bp->b_bufsize) { 2921 /* 2922 * DEV_BSIZE aligned new buffer size is less then the 2923 * DEV_BSIZE aligned existing buffer size. Figure out 2924 * if we have to remove any pages. 2925 */ 2926 if (desiredpages < bp->b_npages) { 2927 vm_page_t m; 2928 2929 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2930 vm_page_lock_queues(); 2931 for (i = desiredpages; i < bp->b_npages; i++) { 2932 /* 2933 * the page is not freed here -- it 2934 * is the responsibility of 2935 * vnode_pager_setsize 2936 */ 2937 m = bp->b_pages[i]; 2938 KASSERT(m != bogus_page, 2939 ("allocbuf: bogus page found")); 2940 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2941 vm_page_lock_queues(); 2942 2943 bp->b_pages[i] = NULL; 2944 vm_page_unwire(m, 0); 2945 } 2946 vm_page_unlock_queues(); 2947 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2948 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2949 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2950 bp->b_npages = desiredpages; 2951 } 2952 } else if (size > bp->b_bcount) { 2953 /* 2954 * We are growing the buffer, possibly in a 2955 * byte-granular fashion. 2956 */ 2957 struct vnode *vp; 2958 vm_object_t obj; 2959 vm_offset_t toff; 2960 vm_offset_t tinc; 2961 2962 /* 2963 * Step 1, bring in the VM pages from the object, 2964 * allocating them if necessary. We must clear 2965 * B_CACHE if these pages are not valid for the 2966 * range covered by the buffer. 2967 */ 2968 2969 vp = bp->b_vp; 2970 obj = bp->b_bufobj->bo_object; 2971 2972 VM_OBJECT_LOCK(obj); 2973 while (bp->b_npages < desiredpages) { 2974 vm_page_t m; 2975 vm_pindex_t pi; 2976 2977 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2978 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2979 /* 2980 * note: must allocate system pages 2981 * since blocking here could intefere 2982 * with paging I/O, no matter which 2983 * process we are. 2984 */ 2985 m = vm_page_alloc(obj, pi, 2986 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2987 VM_ALLOC_WIRED); 2988 if (m == NULL) { 2989 atomic_add_int(&vm_pageout_deficit, 2990 desiredpages - bp->b_npages); 2991 VM_OBJECT_UNLOCK(obj); 2992 VM_WAIT; 2993 VM_OBJECT_LOCK(obj); 2994 } else { 2995 if (m->valid == 0) 2996 bp->b_flags &= ~B_CACHE; 2997 bp->b_pages[bp->b_npages] = m; 2998 ++bp->b_npages; 2999 } 3000 continue; 3001 } 3002 3003 /* 3004 * We found a page. If we have to sleep on it, 3005 * retry because it might have gotten freed out 3006 * from under us. 3007 * 3008 * We can only test VPO_BUSY here. Blocking on 3009 * m->busy might lead to a deadlock: 3010 * 3011 * vm_fault->getpages->cluster_read->allocbuf 3012 * 3013 */ 3014 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 3015 continue; 3016 3017 /* 3018 * We have a good page. 3019 */ 3020 vm_page_lock_queues(); 3021 vm_page_wire(m); 3022 vm_page_unlock_queues(); 3023 bp->b_pages[bp->b_npages] = m; 3024 ++bp->b_npages; 3025 } 3026 3027 /* 3028 * Step 2. We've loaded the pages into the buffer, 3029 * we have to figure out if we can still have B_CACHE 3030 * set. Note that B_CACHE is set according to the 3031 * byte-granular range ( bcount and size ), new the 3032 * aligned range ( newbsize ). 3033 * 3034 * The VM test is against m->valid, which is DEV_BSIZE 3035 * aligned. Needless to say, the validity of the data 3036 * needs to also be DEV_BSIZE aligned. Note that this 3037 * fails with NFS if the server or some other client 3038 * extends the file's EOF. If our buffer is resized, 3039 * B_CACHE may remain set! XXX 3040 */ 3041 3042 toff = bp->b_bcount; 3043 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3044 3045 while ((bp->b_flags & B_CACHE) && toff < size) { 3046 vm_pindex_t pi; 3047 3048 if (tinc > (size - toff)) 3049 tinc = size - toff; 3050 3051 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3052 PAGE_SHIFT; 3053 3054 vfs_buf_test_cache( 3055 bp, 3056 bp->b_offset, 3057 toff, 3058 tinc, 3059 bp->b_pages[pi] 3060 ); 3061 toff += tinc; 3062 tinc = PAGE_SIZE; 3063 } 3064 VM_OBJECT_UNLOCK(obj); 3065 3066 /* 3067 * Step 3, fixup the KVM pmap. Remember that 3068 * bp->b_data is relative to bp->b_offset, but 3069 * bp->b_offset may be offset into the first page. 3070 */ 3071 3072 bp->b_data = (caddr_t) 3073 trunc_page((vm_offset_t)bp->b_data); 3074 pmap_qenter( 3075 (vm_offset_t)bp->b_data, 3076 bp->b_pages, 3077 bp->b_npages 3078 ); 3079 3080 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3081 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3082 } 3083 } 3084 if (newbsize < bp->b_bufsize) 3085 bufspacewakeup(); 3086 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3087 bp->b_bcount = size; /* requested buffer size */ 3088 return 1; 3089} 3090 3091void 3092biodone(struct bio *bp) 3093{ 3094 struct mtx *mtxp; 3095 void (*done)(struct bio *); 3096 3097 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3098 mtx_lock(mtxp); 3099 bp->bio_flags |= BIO_DONE; 3100 done = bp->bio_done; 3101 if (done == NULL) 3102 wakeup(bp); 3103 mtx_unlock(mtxp); 3104 if (done != NULL) 3105 done(bp); 3106} 3107 3108/* 3109 * Wait for a BIO to finish. 3110 * 3111 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3112 * case is not yet clear. 3113 */ 3114int 3115biowait(struct bio *bp, const char *wchan) 3116{ 3117 struct mtx *mtxp; 3118 3119 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3120 mtx_lock(mtxp); 3121 while ((bp->bio_flags & BIO_DONE) == 0) 3122 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3123 mtx_unlock(mtxp); 3124 if (bp->bio_error != 0) 3125 return (bp->bio_error); 3126 if (!(bp->bio_flags & BIO_ERROR)) 3127 return (0); 3128 return (EIO); 3129} 3130 3131void 3132biofinish(struct bio *bp, struct devstat *stat, int error) 3133{ 3134 3135 if (error) { 3136 bp->bio_error = error; 3137 bp->bio_flags |= BIO_ERROR; 3138 } 3139 if (stat != NULL) 3140 devstat_end_transaction_bio(stat, bp); 3141 biodone(bp); 3142} 3143 3144/* 3145 * bufwait: 3146 * 3147 * Wait for buffer I/O completion, returning error status. The buffer 3148 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3149 * error and cleared. 3150 */ 3151int 3152bufwait(struct buf *bp) 3153{ 3154 if (bp->b_iocmd == BIO_READ) 3155 bwait(bp, PRIBIO, "biord"); 3156 else 3157 bwait(bp, PRIBIO, "biowr"); 3158 if (bp->b_flags & B_EINTR) { 3159 bp->b_flags &= ~B_EINTR; 3160 return (EINTR); 3161 } 3162 if (bp->b_ioflags & BIO_ERROR) { 3163 return (bp->b_error ? bp->b_error : EIO); 3164 } else { 3165 return (0); 3166 } 3167} 3168 3169 /* 3170 * Call back function from struct bio back up to struct buf. 3171 */ 3172static void 3173bufdonebio(struct bio *bip) 3174{ 3175 struct buf *bp; 3176 3177 bp = bip->bio_caller2; 3178 bp->b_resid = bp->b_bcount - bip->bio_completed; 3179 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3180 bp->b_ioflags = bip->bio_flags; 3181 bp->b_error = bip->bio_error; 3182 if (bp->b_error) 3183 bp->b_ioflags |= BIO_ERROR; 3184 bufdone(bp); 3185 g_destroy_bio(bip); 3186} 3187 3188void 3189dev_strategy(struct cdev *dev, struct buf *bp) 3190{ 3191 struct cdevsw *csw; 3192 struct bio *bip; 3193 3194 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3195 panic("b_iocmd botch"); 3196 for (;;) { 3197 bip = g_new_bio(); 3198 if (bip != NULL) 3199 break; 3200 /* Try again later */ 3201 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3202 } 3203 bip->bio_cmd = bp->b_iocmd; 3204 bip->bio_offset = bp->b_iooffset; 3205 bip->bio_length = bp->b_bcount; 3206 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3207 bip->bio_data = bp->b_data; 3208 bip->bio_done = bufdonebio; 3209 bip->bio_caller2 = bp; 3210 bip->bio_dev = dev; 3211 KASSERT(dev->si_refcount > 0, 3212 ("dev_strategy on un-referenced struct cdev *(%s)", 3213 devtoname(dev))); 3214 csw = dev_refthread(dev); 3215 if (csw == NULL) { 3216 g_destroy_bio(bip); 3217 bp->b_error = ENXIO; 3218 bp->b_ioflags = BIO_ERROR; 3219 bufdone(bp); 3220 return; 3221 } 3222 (*csw->d_strategy)(bip); 3223 dev_relthread(dev); 3224} 3225 3226/* 3227 * bufdone: 3228 * 3229 * Finish I/O on a buffer, optionally calling a completion function. 3230 * This is usually called from an interrupt so process blocking is 3231 * not allowed. 3232 * 3233 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3234 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3235 * assuming B_INVAL is clear. 3236 * 3237 * For the VMIO case, we set B_CACHE if the op was a read and no 3238 * read error occured, or if the op was a write. B_CACHE is never 3239 * set if the buffer is invalid or otherwise uncacheable. 3240 * 3241 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3242 * initiator to leave B_INVAL set to brelse the buffer out of existance 3243 * in the biodone routine. 3244 */ 3245void 3246bufdone(struct buf *bp) 3247{ 3248 struct bufobj *dropobj; 3249 void (*biodone)(struct buf *); 3250 3251 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3252 dropobj = NULL; 3253 3254 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3255 BUF_ASSERT_HELD(bp); 3256 3257 runningbufwakeup(bp); 3258 if (bp->b_iocmd == BIO_WRITE) 3259 dropobj = bp->b_bufobj; 3260 /* call optional completion function if requested */ 3261 if (bp->b_iodone != NULL) { 3262 biodone = bp->b_iodone; 3263 bp->b_iodone = NULL; 3264 (*biodone) (bp); 3265 if (dropobj) 3266 bufobj_wdrop(dropobj); 3267 return; 3268 } 3269 3270 bufdone_finish(bp); 3271 3272 if (dropobj) 3273 bufobj_wdrop(dropobj); 3274} 3275 3276void 3277bufdone_finish(struct buf *bp) 3278{ 3279 BUF_ASSERT_HELD(bp); 3280 3281 if (!LIST_EMPTY(&bp->b_dep)) 3282 buf_complete(bp); 3283 3284 if (bp->b_flags & B_VMIO) { 3285 int i; 3286 vm_ooffset_t foff; 3287 vm_page_t m; 3288 vm_object_t obj; 3289 int iosize; 3290 struct vnode *vp = bp->b_vp; 3291 3292 obj = bp->b_bufobj->bo_object; 3293 3294#if defined(VFS_BIO_DEBUG) 3295 mp_fixme("usecount and vflag accessed without locks."); 3296 if (vp->v_usecount == 0) { 3297 panic("biodone: zero vnode ref count"); 3298 } 3299 3300 KASSERT(vp->v_object != NULL, 3301 ("biodone: vnode %p has no vm_object", vp)); 3302#endif 3303 3304 foff = bp->b_offset; 3305 KASSERT(bp->b_offset != NOOFFSET, 3306 ("biodone: no buffer offset")); 3307 3308 VM_OBJECT_LOCK(obj); 3309#if defined(VFS_BIO_DEBUG) 3310 if (obj->paging_in_progress < bp->b_npages) { 3311 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3312 obj->paging_in_progress, bp->b_npages); 3313 } 3314#endif 3315 3316 /* 3317 * Set B_CACHE if the op was a normal read and no error 3318 * occured. B_CACHE is set for writes in the b*write() 3319 * routines. 3320 */ 3321 iosize = bp->b_bcount - bp->b_resid; 3322 if (bp->b_iocmd == BIO_READ && 3323 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3324 !(bp->b_ioflags & BIO_ERROR)) { 3325 bp->b_flags |= B_CACHE; 3326 } 3327 for (i = 0; i < bp->b_npages; i++) { 3328 int bogusflag = 0; 3329 int resid; 3330 3331 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3332 if (resid > iosize) 3333 resid = iosize; 3334 3335 /* 3336 * cleanup bogus pages, restoring the originals 3337 */ 3338 m = bp->b_pages[i]; 3339 if (m == bogus_page) { 3340 bogusflag = 1; 3341 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3342 if (m == NULL) 3343 panic("biodone: page disappeared!"); 3344 bp->b_pages[i] = m; 3345 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3346 bp->b_pages, bp->b_npages); 3347 } 3348#if defined(VFS_BIO_DEBUG) 3349 if (OFF_TO_IDX(foff) != m->pindex) { 3350 printf( 3351"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3352 (intmax_t)foff, (uintmax_t)m->pindex); 3353 } 3354#endif 3355 3356 /* 3357 * In the write case, the valid and clean bits are 3358 * already changed correctly ( see bdwrite() ), so we 3359 * only need to do this here in the read case. 3360 */ 3361 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3362 KASSERT((m->dirty & vm_page_bits(foff & 3363 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3364 " page %p has unexpected dirty bits", m)); 3365 vfs_page_set_valid(bp, foff, m); 3366 } 3367 3368 /* 3369 * when debugging new filesystems or buffer I/O methods, this 3370 * is the most common error that pops up. if you see this, you 3371 * have not set the page busy flag correctly!!! 3372 */ 3373 if (m->busy == 0) { 3374 printf("biodone: page busy < 0, " 3375 "pindex: %d, foff: 0x(%x,%x), " 3376 "resid: %d, index: %d\n", 3377 (int) m->pindex, (int)(foff >> 32), 3378 (int) foff & 0xffffffff, resid, i); 3379 if (!vn_isdisk(vp, NULL)) 3380 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3381 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3382 (intmax_t) bp->b_lblkno, 3383 bp->b_flags, bp->b_npages); 3384 else 3385 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3386 (intmax_t) bp->b_lblkno, 3387 bp->b_flags, bp->b_npages); 3388 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3389 (u_long)m->valid, (u_long)m->dirty, 3390 m->wire_count); 3391 panic("biodone: page busy < 0\n"); 3392 } 3393 vm_page_io_finish(m); 3394 vm_object_pip_subtract(obj, 1); 3395 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3396 iosize -= resid; 3397 } 3398 vm_object_pip_wakeupn(obj, 0); 3399 VM_OBJECT_UNLOCK(obj); 3400 } 3401 3402 /* 3403 * For asynchronous completions, release the buffer now. The brelse 3404 * will do a wakeup there if necessary - so no need to do a wakeup 3405 * here in the async case. The sync case always needs to do a wakeup. 3406 */ 3407 3408 if (bp->b_flags & B_ASYNC) { 3409 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3410 brelse(bp); 3411 else 3412 bqrelse(bp); 3413 } else 3414 bdone(bp); 3415} 3416 3417/* 3418 * This routine is called in lieu of iodone in the case of 3419 * incomplete I/O. This keeps the busy status for pages 3420 * consistant. 3421 */ 3422void 3423vfs_unbusy_pages(struct buf *bp) 3424{ 3425 int i; 3426 vm_object_t obj; 3427 vm_page_t m; 3428 3429 runningbufwakeup(bp); 3430 if (!(bp->b_flags & B_VMIO)) 3431 return; 3432 3433 obj = bp->b_bufobj->bo_object; 3434 VM_OBJECT_LOCK(obj); 3435 for (i = 0; i < bp->b_npages; i++) { 3436 m = bp->b_pages[i]; 3437 if (m == bogus_page) { 3438 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3439 if (!m) 3440 panic("vfs_unbusy_pages: page missing\n"); 3441 bp->b_pages[i] = m; 3442 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3443 bp->b_pages, bp->b_npages); 3444 } 3445 vm_object_pip_subtract(obj, 1); 3446 vm_page_io_finish(m); 3447 } 3448 vm_object_pip_wakeupn(obj, 0); 3449 VM_OBJECT_UNLOCK(obj); 3450} 3451 3452/* 3453 * vfs_page_set_valid: 3454 * 3455 * Set the valid bits in a page based on the supplied offset. The 3456 * range is restricted to the buffer's size. 3457 * 3458 * This routine is typically called after a read completes. 3459 */ 3460static void 3461vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3462{ 3463 vm_ooffset_t eoff; 3464 3465 /* 3466 * Compute the end offset, eoff, such that [off, eoff) does not span a 3467 * page boundary and eoff is not greater than the end of the buffer. 3468 * The end of the buffer, in this case, is our file EOF, not the 3469 * allocation size of the buffer. 3470 */ 3471 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3472 if (eoff > bp->b_offset + bp->b_bcount) 3473 eoff = bp->b_offset + bp->b_bcount; 3474 3475 /* 3476 * Set valid range. This is typically the entire buffer and thus the 3477 * entire page. 3478 */ 3479 if (eoff > off) 3480 vm_page_set_valid(m, off & PAGE_MASK, eoff - off); 3481} 3482 3483/* 3484 * vfs_page_set_validclean: 3485 * 3486 * Set the valid bits and clear the dirty bits in a page based on the 3487 * supplied offset. The range is restricted to the buffer's size. 3488 */ 3489static void 3490vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3491{ 3492 vm_ooffset_t soff, eoff; 3493 3494 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3495 /* 3496 * Start and end offsets in buffer. eoff - soff may not cross a 3497 * page boundry or cross the end of the buffer. The end of the 3498 * buffer, in this case, is our file EOF, not the allocation size 3499 * of the buffer. 3500 */ 3501 soff = off; 3502 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3503 if (eoff > bp->b_offset + bp->b_bcount) 3504 eoff = bp->b_offset + bp->b_bcount; 3505 3506 /* 3507 * Set valid range. This is typically the entire buffer and thus the 3508 * entire page. 3509 */ 3510 if (eoff > soff) { 3511 vm_page_set_validclean( 3512 m, 3513 (vm_offset_t) (soff & PAGE_MASK), 3514 (vm_offset_t) (eoff - soff) 3515 ); 3516 } 3517} 3518 3519/* 3520 * This routine is called before a device strategy routine. 3521 * It is used to tell the VM system that paging I/O is in 3522 * progress, and treat the pages associated with the buffer 3523 * almost as being VPO_BUSY. Also the object paging_in_progress 3524 * flag is handled to make sure that the object doesn't become 3525 * inconsistant. 3526 * 3527 * Since I/O has not been initiated yet, certain buffer flags 3528 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3529 * and should be ignored. 3530 */ 3531void 3532vfs_busy_pages(struct buf *bp, int clear_modify) 3533{ 3534 int i, bogus; 3535 vm_object_t obj; 3536 vm_ooffset_t foff; 3537 vm_page_t m; 3538 3539 if (!(bp->b_flags & B_VMIO)) 3540 return; 3541 3542 obj = bp->b_bufobj->bo_object; 3543 foff = bp->b_offset; 3544 KASSERT(bp->b_offset != NOOFFSET, 3545 ("vfs_busy_pages: no buffer offset")); 3546 VM_OBJECT_LOCK(obj); 3547 if (bp->b_bufsize != 0) 3548 vfs_setdirty_locked_object(bp); 3549retry: 3550 for (i = 0; i < bp->b_npages; i++) { 3551 m = bp->b_pages[i]; 3552 3553 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3554 goto retry; 3555 } 3556 bogus = 0; 3557 if (clear_modify) 3558 vm_page_lock_queues(); 3559 for (i = 0; i < bp->b_npages; i++) { 3560 m = bp->b_pages[i]; 3561 3562 if ((bp->b_flags & B_CLUSTER) == 0) { 3563 vm_object_pip_add(obj, 1); 3564 vm_page_io_start(m); 3565 } 3566 /* 3567 * When readying a buffer for a read ( i.e 3568 * clear_modify == 0 ), it is important to do 3569 * bogus_page replacement for valid pages in 3570 * partially instantiated buffers. Partially 3571 * instantiated buffers can, in turn, occur when 3572 * reconstituting a buffer from its VM backing store 3573 * base. We only have to do this if B_CACHE is 3574 * clear ( which causes the I/O to occur in the 3575 * first place ). The replacement prevents the read 3576 * I/O from overwriting potentially dirty VM-backed 3577 * pages. XXX bogus page replacement is, uh, bogus. 3578 * It may not work properly with small-block devices. 3579 * We need to find a better way. 3580 */ 3581 if (clear_modify) { 3582 pmap_remove_write(m); 3583 vfs_page_set_validclean(bp, foff, m); 3584 } else if (m->valid == VM_PAGE_BITS_ALL && 3585 (bp->b_flags & B_CACHE) == 0) { 3586 bp->b_pages[i] = bogus_page; 3587 bogus++; 3588 } 3589 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3590 } 3591 if (clear_modify) 3592 vm_page_unlock_queues(); 3593 VM_OBJECT_UNLOCK(obj); 3594 if (bogus) 3595 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3596 bp->b_pages, bp->b_npages); 3597} 3598 3599/* 3600 * Tell the VM system that the pages associated with this buffer 3601 * are clean. This is used for delayed writes where the data is 3602 * going to go to disk eventually without additional VM intevention. 3603 * 3604 * Note that while we only really need to clean through to b_bcount, we 3605 * just go ahead and clean through to b_bufsize. 3606 */ 3607static void 3608vfs_clean_pages(struct buf *bp) 3609{ 3610 int i; 3611 vm_ooffset_t foff, noff, eoff; 3612 vm_page_t m; 3613 3614 if (!(bp->b_flags & B_VMIO)) 3615 return; 3616 3617 foff = bp->b_offset; 3618 KASSERT(bp->b_offset != NOOFFSET, 3619 ("vfs_clean_pages: no buffer offset")); 3620 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3621 vm_page_lock_queues(); 3622 for (i = 0; i < bp->b_npages; i++) { 3623 m = bp->b_pages[i]; 3624 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3625 eoff = noff; 3626 3627 if (eoff > bp->b_offset + bp->b_bufsize) 3628 eoff = bp->b_offset + bp->b_bufsize; 3629 vfs_page_set_validclean(bp, foff, m); 3630 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3631 foff = noff; 3632 } 3633 vm_page_unlock_queues(); 3634 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3635} 3636 3637/* 3638 * vfs_bio_set_valid: 3639 * 3640 * Set the range within the buffer to valid. The range is 3641 * relative to the beginning of the buffer, b_offset. Note that 3642 * b_offset itself may be offset from the beginning of the first 3643 * page. 3644 */ 3645void 3646vfs_bio_set_valid(struct buf *bp, int base, int size) 3647{ 3648 int i, n; 3649 vm_page_t m; 3650 3651 if (!(bp->b_flags & B_VMIO)) 3652 return; 3653 3654 /* 3655 * Fixup base to be relative to beginning of first page. 3656 * Set initial n to be the maximum number of bytes in the 3657 * first page that can be validated. 3658 */ 3659 base += (bp->b_offset & PAGE_MASK); 3660 n = PAGE_SIZE - (base & PAGE_MASK); 3661 3662 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3663 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3664 m = bp->b_pages[i]; 3665 if (n > size) 3666 n = size; 3667 vm_page_set_valid(m, base & PAGE_MASK, n); 3668 base += n; 3669 size -= n; 3670 n = PAGE_SIZE; 3671 } 3672 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3673} 3674 3675/* 3676 * vfs_bio_set_validclean: 3677 * 3678 * Set the range within the buffer to valid and clean. The range is 3679 * relative to the beginning of the buffer, b_offset. Note that b_offset 3680 * itself may be offset from the beginning of the first page. 3681 * 3682 */ 3683 3684void 3685vfs_bio_set_validclean(struct buf *bp, int base, int size) 3686{ 3687 int i, n; 3688 vm_page_t m; 3689 3690 if (!(bp->b_flags & B_VMIO)) 3691 return; 3692 /* 3693 * Fixup base to be relative to beginning of first page. 3694 * Set initial n to be the maximum number of bytes in the 3695 * first page that can be validated. 3696 */ 3697 3698 base += (bp->b_offset & PAGE_MASK); 3699 n = PAGE_SIZE - (base & PAGE_MASK); 3700 3701 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3702 vm_page_lock_queues(); 3703 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3704 m = bp->b_pages[i]; 3705 if (n > size) 3706 n = size; 3707 vm_page_set_validclean(m, base & PAGE_MASK, n); 3708 base += n; 3709 size -= n; 3710 n = PAGE_SIZE; 3711 } 3712 vm_page_unlock_queues(); 3713 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3714} 3715 3716/* 3717 * vfs_bio_clrbuf: 3718 * 3719 * If the specified buffer is a non-VMIO buffer, clear the entire 3720 * buffer. If the specified buffer is a VMIO buffer, clear and 3721 * validate only the previously invalid portions of the buffer. 3722 * This routine essentially fakes an I/O, so we need to clear 3723 * BIO_ERROR and B_INVAL. 3724 * 3725 * Note that while we only theoretically need to clear through b_bcount, 3726 * we go ahead and clear through b_bufsize. 3727 */ 3728void 3729vfs_bio_clrbuf(struct buf *bp) 3730{ 3731 int i, j, mask; 3732 caddr_t sa, ea; 3733 3734 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3735 clrbuf(bp); 3736 return; 3737 } 3738 bp->b_flags &= ~B_INVAL; 3739 bp->b_ioflags &= ~BIO_ERROR; 3740 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3741 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3742 (bp->b_offset & PAGE_MASK) == 0) { 3743 if (bp->b_pages[0] == bogus_page) 3744 goto unlock; 3745 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3746 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3747 if ((bp->b_pages[0]->valid & mask) == mask) 3748 goto unlock; 3749 if ((bp->b_pages[0]->valid & mask) == 0) { 3750 bzero(bp->b_data, bp->b_bufsize); 3751 bp->b_pages[0]->valid |= mask; 3752 goto unlock; 3753 } 3754 } 3755 ea = sa = bp->b_data; 3756 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3757 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3758 ea = (caddr_t)(vm_offset_t)ulmin( 3759 (u_long)(vm_offset_t)ea, 3760 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3761 if (bp->b_pages[i] == bogus_page) 3762 continue; 3763 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3764 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3765 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3766 if ((bp->b_pages[i]->valid & mask) == mask) 3767 continue; 3768 if ((bp->b_pages[i]->valid & mask) == 0) 3769 bzero(sa, ea - sa); 3770 else { 3771 for (; sa < ea; sa += DEV_BSIZE, j++) { 3772 if ((bp->b_pages[i]->valid & (1 << j)) == 0) 3773 bzero(sa, DEV_BSIZE); 3774 } 3775 } 3776 bp->b_pages[i]->valid |= mask; 3777 } 3778unlock: 3779 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3780 bp->b_resid = 0; 3781} 3782 3783/* 3784 * vm_hold_load_pages and vm_hold_free_pages get pages into 3785 * a buffers address space. The pages are anonymous and are 3786 * not associated with a file object. 3787 */ 3788static void 3789vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3790{ 3791 vm_offset_t pg; 3792 vm_page_t p; 3793 int index; 3794 3795 to = round_page(to); 3796 from = round_page(from); 3797 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3798 3799 VM_OBJECT_LOCK(kernel_object); 3800 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3801tryagain: 3802 /* 3803 * note: must allocate system pages since blocking here 3804 * could intefere with paging I/O, no matter which 3805 * process we are. 3806 */ 3807 p = vm_page_alloc(kernel_object, 3808 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3809 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3810 if (!p) { 3811 atomic_add_int(&vm_pageout_deficit, 3812 (to - pg) >> PAGE_SHIFT); 3813 VM_OBJECT_UNLOCK(kernel_object); 3814 VM_WAIT; 3815 VM_OBJECT_LOCK(kernel_object); 3816 goto tryagain; 3817 } 3818 p->valid = VM_PAGE_BITS_ALL; 3819 pmap_qenter(pg, &p, 1); 3820 bp->b_pages[index] = p; 3821 } 3822 VM_OBJECT_UNLOCK(kernel_object); 3823 bp->b_npages = index; 3824} 3825 3826/* Return pages associated with this buf to the vm system */ 3827static void 3828vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3829{ 3830 vm_offset_t pg; 3831 vm_page_t p; 3832 int index, newnpages; 3833 3834 from = round_page(from); 3835 to = round_page(to); 3836 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3837 3838 VM_OBJECT_LOCK(kernel_object); 3839 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3840 p = bp->b_pages[index]; 3841 if (p && (index < bp->b_npages)) { 3842 if (p->busy) { 3843 printf( 3844 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3845 (intmax_t)bp->b_blkno, 3846 (intmax_t)bp->b_lblkno); 3847 } 3848 bp->b_pages[index] = NULL; 3849 pmap_qremove(pg, 1); 3850 vm_page_lock_queues(); 3851 vm_page_unwire(p, 0); 3852 vm_page_free(p); 3853 vm_page_unlock_queues(); 3854 } 3855 } 3856 VM_OBJECT_UNLOCK(kernel_object); 3857 bp->b_npages = newnpages; 3858} 3859 3860/* 3861 * Map an IO request into kernel virtual address space. 3862 * 3863 * All requests are (re)mapped into kernel VA space. 3864 * Notice that we use b_bufsize for the size of the buffer 3865 * to be mapped. b_bcount might be modified by the driver. 3866 * 3867 * Note that even if the caller determines that the address space should 3868 * be valid, a race or a smaller-file mapped into a larger space may 3869 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3870 * check the return value. 3871 */ 3872int 3873vmapbuf(struct buf *bp) 3874{ 3875 caddr_t addr, kva; 3876 vm_prot_t prot; 3877 int pidx, i; 3878 struct vm_page *m; 3879 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3880 3881 if (bp->b_bufsize < 0) 3882 return (-1); 3883 prot = VM_PROT_READ; 3884 if (bp->b_iocmd == BIO_READ) 3885 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3886 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3887 addr < bp->b_data + bp->b_bufsize; 3888 addr += PAGE_SIZE, pidx++) { 3889 /* 3890 * Do the vm_fault if needed; do the copy-on-write thing 3891 * when reading stuff off device into memory. 3892 * 3893 * NOTE! Must use pmap_extract() because addr may be in 3894 * the userland address space, and kextract is only guarenteed 3895 * to work for the kernland address space (see: sparc64 port). 3896 */ 3897retry: 3898 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3899 prot) < 0) { 3900 vm_page_lock_queues(); 3901 for (i = 0; i < pidx; ++i) { 3902 vm_page_unhold(bp->b_pages[i]); 3903 bp->b_pages[i] = NULL; 3904 } 3905 vm_page_unlock_queues(); 3906 return(-1); 3907 } 3908 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3909 if (m == NULL) 3910 goto retry; 3911 bp->b_pages[pidx] = m; 3912 } 3913 if (pidx > btoc(MAXPHYS)) 3914 panic("vmapbuf: mapped more than MAXPHYS"); 3915 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3916 3917 kva = bp->b_saveaddr; 3918 bp->b_npages = pidx; 3919 bp->b_saveaddr = bp->b_data; 3920 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3921 return(0); 3922} 3923 3924/* 3925 * Free the io map PTEs associated with this IO operation. 3926 * We also invalidate the TLB entries and restore the original b_addr. 3927 */ 3928void 3929vunmapbuf(struct buf *bp) 3930{ 3931 int pidx; 3932 int npages; 3933 3934 npages = bp->b_npages; 3935 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3936 vm_page_lock_queues(); 3937 for (pidx = 0; pidx < npages; pidx++) 3938 vm_page_unhold(bp->b_pages[pidx]); 3939 vm_page_unlock_queues(); 3940 3941 bp->b_data = bp->b_saveaddr; 3942} 3943 3944void 3945bdone(struct buf *bp) 3946{ 3947 struct mtx *mtxp; 3948 3949 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3950 mtx_lock(mtxp); 3951 bp->b_flags |= B_DONE; 3952 wakeup(bp); 3953 mtx_unlock(mtxp); 3954} 3955 3956void 3957bwait(struct buf *bp, u_char pri, const char *wchan) 3958{ 3959 struct mtx *mtxp; 3960 3961 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3962 mtx_lock(mtxp); 3963 while ((bp->b_flags & B_DONE) == 0) 3964 msleep(bp, mtxp, pri, wchan, 0); 3965 mtx_unlock(mtxp); 3966} 3967 3968int 3969bufsync(struct bufobj *bo, int waitfor) 3970{ 3971 3972 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 3973} 3974 3975void 3976bufstrategy(struct bufobj *bo, struct buf *bp) 3977{ 3978 int i = 0; 3979 struct vnode *vp; 3980 3981 vp = bp->b_vp; 3982 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3983 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3984 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3985 i = VOP_STRATEGY(vp, bp); 3986 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3987} 3988 3989void 3990bufobj_wrefl(struct bufobj *bo) 3991{ 3992 3993 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3994 ASSERT_BO_LOCKED(bo); 3995 bo->bo_numoutput++; 3996} 3997 3998void 3999bufobj_wref(struct bufobj *bo) 4000{ 4001 4002 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 4003 BO_LOCK(bo); 4004 bo->bo_numoutput++; 4005 BO_UNLOCK(bo); 4006} 4007 4008void 4009bufobj_wdrop(struct bufobj *bo) 4010{ 4011 4012 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 4013 BO_LOCK(bo); 4014 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 4015 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 4016 bo->bo_flag &= ~BO_WWAIT; 4017 wakeup(&bo->bo_numoutput); 4018 } 4019 BO_UNLOCK(bo); 4020} 4021 4022int 4023bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 4024{ 4025 int error; 4026 4027 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 4028 ASSERT_BO_LOCKED(bo); 4029 error = 0; 4030 while (bo->bo_numoutput) { 4031 bo->bo_flag |= BO_WWAIT; 4032 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 4033 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 4034 if (error) 4035 break; 4036 } 4037 return (error); 4038} 4039 4040void 4041bpin(struct buf *bp) 4042{ 4043 struct mtx *mtxp; 4044 4045 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4046 mtx_lock(mtxp); 4047 bp->b_pin_count++; 4048 mtx_unlock(mtxp); 4049} 4050 4051void 4052bunpin(struct buf *bp) 4053{ 4054 struct mtx *mtxp; 4055 4056 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4057 mtx_lock(mtxp); 4058 if (--bp->b_pin_count == 0) 4059 wakeup(bp); 4060 mtx_unlock(mtxp); 4061} 4062 4063void 4064bunpin_wait(struct buf *bp) 4065{ 4066 struct mtx *mtxp; 4067 4068 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4069 mtx_lock(mtxp); 4070 while (bp->b_pin_count > 0) 4071 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4072 mtx_unlock(mtxp); 4073} 4074 4075#include "opt_ddb.h" 4076#ifdef DDB 4077#include <ddb/ddb.h> 4078 4079/* DDB command to show buffer data */ 4080DB_SHOW_COMMAND(buffer, db_show_buffer) 4081{ 4082 /* get args */ 4083 struct buf *bp = (struct buf *)addr; 4084 4085 if (!have_addr) { 4086 db_printf("usage: show buffer <addr>\n"); 4087 return; 4088 } 4089 4090 db_printf("buf at %p\n", bp); 4091 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4092 db_printf( 4093 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4094 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n", 4095 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4096 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4097 bp->b_dep.lh_first); 4098 if (bp->b_npages) { 4099 int i; 4100 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4101 for (i = 0; i < bp->b_npages; i++) { 4102 vm_page_t m; 4103 m = bp->b_pages[i]; 4104 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4105 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4106 if ((i + 1) < bp->b_npages) 4107 db_printf(","); 4108 } 4109 db_printf("\n"); 4110 } 4111 db_printf(" "); 4112 lockmgr_printinfo(&bp->b_lock); 4113} 4114 4115DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4116{ 4117 struct buf *bp; 4118 int i; 4119 4120 for (i = 0; i < nbuf; i++) { 4121 bp = &buf[i]; 4122 if (BUF_ISLOCKED(bp)) { 4123 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4124 db_printf("\n"); 4125 } 4126 } 4127} 4128 4129DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4130{ 4131 struct vnode *vp; 4132 struct buf *bp; 4133 4134 if (!have_addr) { 4135 db_printf("usage: show vnodebufs <addr>\n"); 4136 return; 4137 } 4138 vp = (struct vnode *)addr; 4139 db_printf("Clean buffers:\n"); 4140 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4141 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4142 db_printf("\n"); 4143 } 4144 db_printf("Dirty buffers:\n"); 4145 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4146 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4147 db_printf("\n"); 4148 } 4149} 4150#endif /* DDB */ 4151