vfs_bio.c revision 208745
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 208745 2010-06-02 15:46:37Z alc $"); 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/bio.h> 47#include <sys/conf.h> 48#include <sys/buf.h> 49#include <sys/devicestat.h> 50#include <sys/eventhandler.h> 51#include <sys/fail.h> 52#include <sys/limits.h> 53#include <sys/lock.h> 54#include <sys/malloc.h> 55#include <sys/mount.h> 56#include <sys/mutex.h> 57#include <sys/kernel.h> 58#include <sys/kthread.h> 59#include <sys/proc.h> 60#include <sys/resourcevar.h> 61#include <sys/sysctl.h> 62#include <sys/vmmeter.h> 63#include <sys/vnode.h> 64#include <geom/geom.h> 65#include <vm/vm.h> 66#include <vm/vm_param.h> 67#include <vm/vm_kern.h> 68#include <vm/vm_pageout.h> 69#include <vm/vm_page.h> 70#include <vm/vm_object.h> 71#include <vm/vm_extern.h> 72#include <vm/vm_map.h> 73#include "opt_compat.h" 74#include "opt_directio.h" 75#include "opt_swap.h" 76 77static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 78 79struct bio_ops bioops; /* I/O operation notification */ 80 81struct buf_ops buf_ops_bio = { 82 .bop_name = "buf_ops_bio", 83 .bop_write = bufwrite, 84 .bop_strategy = bufstrategy, 85 .bop_sync = bufsync, 86 .bop_bdflush = bufbdflush, 87}; 88 89/* 90 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 91 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 92 */ 93struct buf *buf; /* buffer header pool */ 94 95static struct proc *bufdaemonproc; 96 97static int inmem(struct vnode *vp, daddr_t blkno); 98static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 99 vm_offset_t to); 100static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 101 vm_offset_t to); 102static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 103static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 104 vm_page_t m); 105static void vfs_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 * Request for the buf daemon to write more buffers than is indicated by 220 * lodirtybuf. This may be necessary to push out excess dependencies or 221 * defragment the address space where a simple count of the number of dirty 222 * buffers is insufficient to characterize the demand for flushing them. 223 */ 224static int bd_speedupreq; 225 226/* 227 * This lock synchronizes access to bd_request. 228 */ 229static struct mtx bdlock; 230 231/* 232 * bogus page -- for I/O to/from partially complete buffers 233 * this is a temporary solution to the problem, but it is not 234 * really that bad. it would be better to split the buffer 235 * for input in the case of buffers partially already in memory, 236 * but the code is intricate enough already. 237 */ 238vm_page_t bogus_page; 239 240/* 241 * Synchronization (sleep/wakeup) variable for active buffer space requests. 242 * Set when wait starts, cleared prior to wakeup(). 243 * Used in runningbufwakeup() and waitrunningbufspace(). 244 */ 245static int runningbufreq; 246 247/* 248 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 249 * waitrunningbufspace(). 250 */ 251static struct mtx rbreqlock; 252 253/* 254 * Synchronization (sleep/wakeup) variable for buffer requests. 255 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 256 * by and/or. 257 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 258 * getnewbuf(), and getblk(). 259 */ 260static int needsbuffer; 261 262/* 263 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 264 */ 265static struct mtx nblock; 266 267/* 268 * Definitions for the buffer free lists. 269 */ 270#define BUFFER_QUEUES 6 /* number of free buffer queues */ 271 272#define QUEUE_NONE 0 /* on no queue */ 273#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 274#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 275#define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ 276#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 277#define QUEUE_EMPTY 5 /* empty buffer headers */ 278#define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 279 280/* Queues for free buffers with various properties */ 281static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 282 283/* Lock for the bufqueues */ 284static struct mtx bqlock; 285 286/* 287 * Single global constant for BUF_WMESG, to avoid getting multiple references. 288 * buf_wmesg is referred from macros. 289 */ 290const char *buf_wmesg = BUF_WMESG; 291 292#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 293#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 294#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 295#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 296 297#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 298 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 299static int 300sysctl_bufspace(SYSCTL_HANDLER_ARGS) 301{ 302 long lvalue; 303 int ivalue; 304 305 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 306 return (sysctl_handle_long(oidp, arg1, arg2, req)); 307 lvalue = *(long *)arg1; 308 if (lvalue > INT_MAX) 309 /* On overflow, still write out a long to trigger ENOMEM. */ 310 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 311 ivalue = lvalue; 312 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 313} 314#endif 315 316#ifdef DIRECTIO 317extern void ffs_rawread_setup(void); 318#endif /* DIRECTIO */ 319/* 320 * numdirtywakeup: 321 * 322 * If someone is blocked due to there being too many dirty buffers, 323 * and numdirtybuffers is now reasonable, wake them up. 324 */ 325 326static __inline void 327numdirtywakeup(int level) 328{ 329 330 if (numdirtybuffers <= level) { 331 mtx_lock(&nblock); 332 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 333 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 334 wakeup(&needsbuffer); 335 } 336 mtx_unlock(&nblock); 337 } 338} 339 340/* 341 * bufspacewakeup: 342 * 343 * Called when buffer space is potentially available for recovery. 344 * getnewbuf() will block on this flag when it is unable to free 345 * sufficient buffer space. Buffer space becomes recoverable when 346 * bp's get placed back in the queues. 347 */ 348 349static __inline void 350bufspacewakeup(void) 351{ 352 353 /* 354 * If someone is waiting for BUF space, wake them up. Even 355 * though we haven't freed the kva space yet, the waiting 356 * process will be able to now. 357 */ 358 mtx_lock(&nblock); 359 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 360 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 361 wakeup(&needsbuffer); 362 } 363 mtx_unlock(&nblock); 364} 365 366/* 367 * runningbufwakeup() - in-progress I/O accounting. 368 * 369 */ 370void 371runningbufwakeup(struct buf *bp) 372{ 373 374 if (bp->b_runningbufspace) { 375 atomic_subtract_long(&runningbufspace, bp->b_runningbufspace); 376 bp->b_runningbufspace = 0; 377 mtx_lock(&rbreqlock); 378 if (runningbufreq && runningbufspace <= lorunningspace) { 379 runningbufreq = 0; 380 wakeup(&runningbufreq); 381 } 382 mtx_unlock(&rbreqlock); 383 } 384} 385 386/* 387 * bufcountwakeup: 388 * 389 * Called when a buffer has been added to one of the free queues to 390 * account for the buffer and to wakeup anyone waiting for free buffers. 391 * This typically occurs when large amounts of metadata are being handled 392 * by the buffer cache ( else buffer space runs out first, usually ). 393 */ 394 395static __inline void 396bufcountwakeup(void) 397{ 398 399 atomic_add_int(&numfreebuffers, 1); 400 mtx_lock(&nblock); 401 if (needsbuffer) { 402 needsbuffer &= ~VFS_BIO_NEED_ANY; 403 if (numfreebuffers >= hifreebuffers) 404 needsbuffer &= ~VFS_BIO_NEED_FREE; 405 wakeup(&needsbuffer); 406 } 407 mtx_unlock(&nblock); 408} 409 410/* 411 * waitrunningbufspace() 412 * 413 * runningbufspace is a measure of the amount of I/O currently 414 * running. This routine is used in async-write situations to 415 * prevent creating huge backups of pending writes to a device. 416 * Only asynchronous writes are governed by this function. 417 * 418 * Reads will adjust runningbufspace, but will not block based on it. 419 * The read load has a side effect of reducing the allowed write load. 420 * 421 * This does NOT turn an async write into a sync write. It waits 422 * for earlier writes to complete and generally returns before the 423 * caller's write has reached the device. 424 */ 425void 426waitrunningbufspace(void) 427{ 428 429 mtx_lock(&rbreqlock); 430 while (runningbufspace > hirunningspace) { 431 ++runningbufreq; 432 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 433 } 434 mtx_unlock(&rbreqlock); 435} 436 437 438/* 439 * vfs_buf_test_cache: 440 * 441 * Called when a buffer is extended. This function clears the B_CACHE 442 * bit if the newly extended portion of the buffer does not contain 443 * valid data. 444 */ 445static __inline 446void 447vfs_buf_test_cache(struct buf *bp, 448 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 449 vm_page_t m) 450{ 451 452 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 453 if (bp->b_flags & B_CACHE) { 454 int base = (foff + off) & PAGE_MASK; 455 if (vm_page_is_valid(m, base, size) == 0) 456 bp->b_flags &= ~B_CACHE; 457 } 458} 459 460/* Wake up the buffer daemon if necessary */ 461static __inline 462void 463bd_wakeup(int dirtybuflevel) 464{ 465 466 mtx_lock(&bdlock); 467 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 468 bd_request = 1; 469 wakeup(&bd_request); 470 } 471 mtx_unlock(&bdlock); 472} 473 474/* 475 * bd_speedup - speedup the buffer cache flushing code 476 */ 477 478void 479bd_speedup(void) 480{ 481 int needwake; 482 483 mtx_lock(&bdlock); 484 needwake = 0; 485 if (bd_speedupreq == 0 || bd_request == 0) 486 needwake = 1; 487 bd_speedupreq = 1; 488 bd_request = 1; 489 if (needwake) 490 wakeup(&bd_request); 491 mtx_unlock(&bdlock); 492} 493 494/* 495 * Calculating buffer cache scaling values and reserve space for buffer 496 * headers. This is called during low level kernel initialization and 497 * may be called more then once. We CANNOT write to the memory area 498 * being reserved at this time. 499 */ 500caddr_t 501kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 502{ 503 int tuned_nbuf; 504 long maxbuf; 505 506 /* 507 * physmem_est is in pages. Convert it to kilobytes (assumes 508 * PAGE_SIZE is >= 1K) 509 */ 510 physmem_est = physmem_est * (PAGE_SIZE / 1024); 511 512 /* 513 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 514 * For the first 64MB of ram nominally allocate sufficient buffers to 515 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 516 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 517 * the buffer cache we limit the eventual kva reservation to 518 * maxbcache bytes. 519 * 520 * factor represents the 1/4 x ram conversion. 521 */ 522 if (nbuf == 0) { 523 int factor = 4 * BKVASIZE / 1024; 524 525 nbuf = 50; 526 if (physmem_est > 4096) 527 nbuf += min((physmem_est - 4096) / factor, 528 65536 / factor); 529 if (physmem_est > 65536) 530 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 531 532 if (maxbcache && nbuf > maxbcache / BKVASIZE) 533 nbuf = maxbcache / BKVASIZE; 534 tuned_nbuf = 1; 535 } else 536 tuned_nbuf = 0; 537 538 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 539 maxbuf = (LONG_MAX / 3) / BKVASIZE; 540 if (nbuf > maxbuf) { 541 if (!tuned_nbuf) 542 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 543 maxbuf); 544 nbuf = maxbuf; 545 } 546 547 /* 548 * swbufs are used as temporary holders for I/O, such as paging I/O. 549 * We have no less then 16 and no more then 256. 550 */ 551 nswbuf = max(min(nbuf/4, 256), 16); 552#ifdef NSWBUF_MIN 553 if (nswbuf < NSWBUF_MIN) 554 nswbuf = NSWBUF_MIN; 555#endif 556#ifdef DIRECTIO 557 ffs_rawread_setup(); 558#endif 559 560 /* 561 * Reserve space for the buffer cache buffers 562 */ 563 swbuf = (void *)v; 564 v = (caddr_t)(swbuf + nswbuf); 565 buf = (void *)v; 566 v = (caddr_t)(buf + nbuf); 567 568 return(v); 569} 570 571/* Initialize the buffer subsystem. Called before use of any buffers. */ 572void 573bufinit(void) 574{ 575 struct buf *bp; 576 int i; 577 578 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 579 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 580 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 581 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 582 583 /* next, make a null set of free lists */ 584 for (i = 0; i < BUFFER_QUEUES; i++) 585 TAILQ_INIT(&bufqueues[i]); 586 587 /* finally, initialize each buffer header and stick on empty q */ 588 for (i = 0; i < nbuf; i++) { 589 bp = &buf[i]; 590 bzero(bp, sizeof *bp); 591 bp->b_flags = B_INVAL; /* we're just an empty header */ 592 bp->b_rcred = NOCRED; 593 bp->b_wcred = NOCRED; 594 bp->b_qindex = QUEUE_EMPTY; 595 bp->b_vflags = 0; 596 bp->b_xflags = 0; 597 LIST_INIT(&bp->b_dep); 598 BUF_LOCKINIT(bp); 599 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 600 } 601 602 /* 603 * maxbufspace is the absolute maximum amount of buffer space we are 604 * allowed to reserve in KVM and in real terms. The absolute maximum 605 * is nominally used by buf_daemon. hibufspace is the nominal maximum 606 * used by most other processes. The differential is required to 607 * ensure that buf_daemon is able to run when other processes might 608 * be blocked waiting for buffer space. 609 * 610 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 611 * this may result in KVM fragmentation which is not handled optimally 612 * by the system. 613 */ 614 maxbufspace = (long)nbuf * BKVASIZE; 615 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 616 lobufspace = hibufspace - MAXBSIZE; 617 618 lorunningspace = 512 * 1024; 619 hirunningspace = 1024 * 1024; 620 621/* 622 * Limit the amount of malloc memory since it is wired permanently into 623 * the kernel space. Even though this is accounted for in the buffer 624 * allocation, we don't want the malloced region to grow uncontrolled. 625 * The malloc scheme improves memory utilization significantly on average 626 * (small) directories. 627 */ 628 maxbufmallocspace = hibufspace / 20; 629 630/* 631 * Reduce the chance of a deadlock occuring by limiting the number 632 * of delayed-write dirty buffers we allow to stack up. 633 */ 634 hidirtybuffers = nbuf / 4 + 20; 635 dirtybufthresh = hidirtybuffers * 9 / 10; 636 numdirtybuffers = 0; 637/* 638 * To support extreme low-memory systems, make sure hidirtybuffers cannot 639 * eat up all available buffer space. This occurs when our minimum cannot 640 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 641 * BKVASIZE'd (8K) buffers. 642 */ 643 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 644 hidirtybuffers >>= 1; 645 } 646 lodirtybuffers = hidirtybuffers / 2; 647 648/* 649 * Try to keep the number of free buffers in the specified range, 650 * and give special processes (e.g. like buf_daemon) access to an 651 * emergency reserve. 652 */ 653 lofreebuffers = nbuf / 18 + 5; 654 hifreebuffers = 2 * lofreebuffers; 655 numfreebuffers = nbuf; 656 657 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 658 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 659} 660 661/* 662 * bfreekva() - free the kva allocation for a buffer. 663 * 664 * Since this call frees up buffer space, we call bufspacewakeup(). 665 */ 666static void 667bfreekva(struct buf *bp) 668{ 669 670 if (bp->b_kvasize) { 671 atomic_add_int(&buffreekvacnt, 1); 672 atomic_subtract_long(&bufspace, bp->b_kvasize); 673 vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 674 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 675 bp->b_kvasize = 0; 676 bufspacewakeup(); 677 } 678} 679 680/* 681 * bremfree: 682 * 683 * Mark the buffer for removal from the appropriate free list in brelse. 684 * 685 */ 686void 687bremfree(struct buf *bp) 688{ 689 690 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 691 KASSERT((bp->b_flags & B_REMFREE) == 0, 692 ("bremfree: buffer %p already marked for delayed removal.", bp)); 693 KASSERT(bp->b_qindex != QUEUE_NONE, 694 ("bremfree: buffer %p not on a queue.", bp)); 695 BUF_ASSERT_HELD(bp); 696 697 bp->b_flags |= B_REMFREE; 698 /* Fixup numfreebuffers count. */ 699 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 700 atomic_subtract_int(&numfreebuffers, 1); 701} 702 703/* 704 * bremfreef: 705 * 706 * Force an immediate removal from a free list. Used only in nfs when 707 * it abuses the b_freelist pointer. 708 */ 709void 710bremfreef(struct buf *bp) 711{ 712 mtx_lock(&bqlock); 713 bremfreel(bp); 714 mtx_unlock(&bqlock); 715} 716 717/* 718 * bremfreel: 719 * 720 * Removes a buffer from the free list, must be called with the 721 * bqlock held. 722 */ 723static void 724bremfreel(struct buf *bp) 725{ 726 CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 727 bp, bp->b_vp, bp->b_flags); 728 KASSERT(bp->b_qindex != QUEUE_NONE, 729 ("bremfreel: buffer %p not on a queue.", bp)); 730 BUF_ASSERT_HELD(bp); 731 mtx_assert(&bqlock, MA_OWNED); 732 733 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 734 bp->b_qindex = QUEUE_NONE; 735 /* 736 * If this was a delayed bremfree() we only need to remove the buffer 737 * from the queue and return the stats are already done. 738 */ 739 if (bp->b_flags & B_REMFREE) { 740 bp->b_flags &= ~B_REMFREE; 741 return; 742 } 743 /* 744 * Fixup numfreebuffers count. If the buffer is invalid or not 745 * delayed-write, the buffer was free and we must decrement 746 * numfreebuffers. 747 */ 748 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 749 atomic_subtract_int(&numfreebuffers, 1); 750} 751 752 753/* 754 * Get a buffer with the specified data. Look in the cache first. We 755 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 756 * is set, the buffer is valid and we do not have to do anything ( see 757 * getblk() ). This is really just a special case of breadn(). 758 */ 759int 760bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 761 struct buf **bpp) 762{ 763 764 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 765} 766 767/* 768 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 769 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 770 * the buffer is valid and we do not have to do anything. 771 */ 772void 773breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 774 int cnt, struct ucred * cred) 775{ 776 struct buf *rabp; 777 int i; 778 779 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 780 if (inmem(vp, *rablkno)) 781 continue; 782 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 783 784 if ((rabp->b_flags & B_CACHE) == 0) { 785 if (!TD_IS_IDLETHREAD(curthread)) 786 curthread->td_ru.ru_inblock++; 787 rabp->b_flags |= B_ASYNC; 788 rabp->b_flags &= ~B_INVAL; 789 rabp->b_ioflags &= ~BIO_ERROR; 790 rabp->b_iocmd = BIO_READ; 791 if (rabp->b_rcred == NOCRED && cred != NOCRED) 792 rabp->b_rcred = crhold(cred); 793 vfs_busy_pages(rabp, 0); 794 BUF_KERNPROC(rabp); 795 rabp->b_iooffset = dbtob(rabp->b_blkno); 796 bstrategy(rabp); 797 } else { 798 brelse(rabp); 799 } 800 } 801} 802 803/* 804 * Operates like bread, but also starts asynchronous I/O on 805 * read-ahead blocks. 806 */ 807int 808breadn(struct vnode * vp, daddr_t blkno, int size, 809 daddr_t * rablkno, int *rabsize, 810 int cnt, struct ucred * cred, struct buf **bpp) 811{ 812 struct buf *bp; 813 int rv = 0, readwait = 0; 814 815 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 816 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 817 818 /* if not found in cache, do some I/O */ 819 if ((bp->b_flags & B_CACHE) == 0) { 820 if (!TD_IS_IDLETHREAD(curthread)) 821 curthread->td_ru.ru_inblock++; 822 bp->b_iocmd = BIO_READ; 823 bp->b_flags &= ~B_INVAL; 824 bp->b_ioflags &= ~BIO_ERROR; 825 if (bp->b_rcred == NOCRED && cred != NOCRED) 826 bp->b_rcred = crhold(cred); 827 vfs_busy_pages(bp, 0); 828 bp->b_iooffset = dbtob(bp->b_blkno); 829 bstrategy(bp); 830 ++readwait; 831 } 832 833 breada(vp, rablkno, rabsize, cnt, cred); 834 835 if (readwait) { 836 rv = bufwait(bp); 837 } 838 return (rv); 839} 840 841/* 842 * Write, release buffer on completion. (Done by iodone 843 * if async). Do not bother writing anything if the buffer 844 * is invalid. 845 * 846 * Note that we set B_CACHE here, indicating that buffer is 847 * fully valid and thus cacheable. This is true even of NFS 848 * now so we set it generally. This could be set either here 849 * or in biodone() since the I/O is synchronous. We put it 850 * here. 851 */ 852int 853bufwrite(struct buf *bp) 854{ 855 int oldflags; 856 struct vnode *vp; 857 int vp_md; 858 859 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 860 if (bp->b_flags & B_INVAL) { 861 brelse(bp); 862 return (0); 863 } 864 865 oldflags = bp->b_flags; 866 867 BUF_ASSERT_HELD(bp); 868 869 if (bp->b_pin_count > 0) 870 bunpin_wait(bp); 871 872 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 873 ("FFS background buffer should not get here %p", bp)); 874 875 vp = bp->b_vp; 876 if (vp) 877 vp_md = vp->v_vflag & VV_MD; 878 else 879 vp_md = 0; 880 881 /* Mark the buffer clean */ 882 bundirty(bp); 883 884 bp->b_flags &= ~B_DONE; 885 bp->b_ioflags &= ~BIO_ERROR; 886 bp->b_flags |= B_CACHE; 887 bp->b_iocmd = BIO_WRITE; 888 889 bufobj_wref(bp->b_bufobj); 890 vfs_busy_pages(bp, 1); 891 892 /* 893 * Normal bwrites pipeline writes 894 */ 895 bp->b_runningbufspace = bp->b_bufsize; 896 atomic_add_long(&runningbufspace, bp->b_runningbufspace); 897 898 if (!TD_IS_IDLETHREAD(curthread)) 899 curthread->td_ru.ru_oublock++; 900 if (oldflags & B_ASYNC) 901 BUF_KERNPROC(bp); 902 bp->b_iooffset = dbtob(bp->b_blkno); 903 bstrategy(bp); 904 905 if ((oldflags & B_ASYNC) == 0) { 906 int rtval = bufwait(bp); 907 brelse(bp); 908 return (rtval); 909 } else { 910 /* 911 * don't allow the async write to saturate the I/O 912 * system. We will not deadlock here because 913 * we are blocking waiting for I/O that is already in-progress 914 * to complete. We do not block here if it is the update 915 * or syncer daemon trying to clean up as that can lead 916 * to deadlock. 917 */ 918 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 919 waitrunningbufspace(); 920 } 921 922 return (0); 923} 924 925void 926bufbdflush(struct bufobj *bo, struct buf *bp) 927{ 928 struct buf *nbp; 929 930 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 931 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 932 altbufferflushes++; 933 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 934 BO_LOCK(bo); 935 /* 936 * Try to find a buffer to flush. 937 */ 938 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 939 if ((nbp->b_vflags & BV_BKGRDINPROG) || 940 BUF_LOCK(nbp, 941 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 942 continue; 943 if (bp == nbp) 944 panic("bdwrite: found ourselves"); 945 BO_UNLOCK(bo); 946 /* Don't countdeps with the bo lock held. */ 947 if (buf_countdeps(nbp, 0)) { 948 BO_LOCK(bo); 949 BUF_UNLOCK(nbp); 950 continue; 951 } 952 if (nbp->b_flags & B_CLUSTEROK) { 953 vfs_bio_awrite(nbp); 954 } else { 955 bremfree(nbp); 956 bawrite(nbp); 957 } 958 dirtybufferflushes++; 959 break; 960 } 961 if (nbp == NULL) 962 BO_UNLOCK(bo); 963 } 964} 965 966/* 967 * Delayed write. (Buffer is marked dirty). Do not bother writing 968 * anything if the buffer is marked invalid. 969 * 970 * Note that since the buffer must be completely valid, we can safely 971 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 972 * biodone() in order to prevent getblk from writing the buffer 973 * out synchronously. 974 */ 975void 976bdwrite(struct buf *bp) 977{ 978 struct thread *td = curthread; 979 struct vnode *vp; 980 struct bufobj *bo; 981 982 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 983 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 984 BUF_ASSERT_HELD(bp); 985 986 if (bp->b_flags & B_INVAL) { 987 brelse(bp); 988 return; 989 } 990 991 /* 992 * If we have too many dirty buffers, don't create any more. 993 * If we are wildly over our limit, then force a complete 994 * cleanup. Otherwise, just keep the situation from getting 995 * out of control. Note that we have to avoid a recursive 996 * disaster and not try to clean up after our own cleanup! 997 */ 998 vp = bp->b_vp; 999 bo = bp->b_bufobj; 1000 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1001 td->td_pflags |= TDP_INBDFLUSH; 1002 BO_BDFLUSH(bo, bp); 1003 td->td_pflags &= ~TDP_INBDFLUSH; 1004 } else 1005 recursiveflushes++; 1006 1007 bdirty(bp); 1008 /* 1009 * Set B_CACHE, indicating that the buffer is fully valid. This is 1010 * true even of NFS now. 1011 */ 1012 bp->b_flags |= B_CACHE; 1013 1014 /* 1015 * This bmap keeps the system from needing to do the bmap later, 1016 * perhaps when the system is attempting to do a sync. Since it 1017 * is likely that the indirect block -- or whatever other datastructure 1018 * that the filesystem needs is still in memory now, it is a good 1019 * thing to do this. Note also, that if the pageout daemon is 1020 * requesting a sync -- there might not be enough memory to do 1021 * the bmap then... So, this is important to do. 1022 */ 1023 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1024 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1025 } 1026 1027 /* 1028 * Set the *dirty* buffer range based upon the VM system dirty pages. 1029 */ 1030 vfs_setdirty(bp); 1031 1032 /* 1033 * We need to do this here to satisfy the vnode_pager and the 1034 * pageout daemon, so that it thinks that the pages have been 1035 * "cleaned". Note that since the pages are in a delayed write 1036 * buffer -- the VFS layer "will" see that the pages get written 1037 * out on the next sync, or perhaps the cluster will be completed. 1038 */ 1039 vfs_clean_pages(bp); 1040 bqrelse(bp); 1041 1042 /* 1043 * Wakeup the buffer flushing daemon if we have a lot of dirty 1044 * buffers (midpoint between our recovery point and our stall 1045 * point). 1046 */ 1047 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1048 1049 /* 1050 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1051 * due to the softdep code. 1052 */ 1053} 1054 1055/* 1056 * bdirty: 1057 * 1058 * Turn buffer into delayed write request. We must clear BIO_READ and 1059 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1060 * itself to properly update it in the dirty/clean lists. We mark it 1061 * B_DONE to ensure that any asynchronization of the buffer properly 1062 * clears B_DONE ( else a panic will occur later ). 1063 * 1064 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1065 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1066 * should only be called if the buffer is known-good. 1067 * 1068 * Since the buffer is not on a queue, we do not update the numfreebuffers 1069 * count. 1070 * 1071 * The buffer must be on QUEUE_NONE. 1072 */ 1073void 1074bdirty(struct buf *bp) 1075{ 1076 1077 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1078 bp, bp->b_vp, bp->b_flags); 1079 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1080 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1081 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1082 BUF_ASSERT_HELD(bp); 1083 bp->b_flags &= ~(B_RELBUF); 1084 bp->b_iocmd = BIO_WRITE; 1085 1086 if ((bp->b_flags & B_DELWRI) == 0) { 1087 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1088 reassignbuf(bp); 1089 atomic_add_int(&numdirtybuffers, 1); 1090 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1091 } 1092} 1093 1094/* 1095 * bundirty: 1096 * 1097 * Clear B_DELWRI for buffer. 1098 * 1099 * Since the buffer is not on a queue, we do not update the numfreebuffers 1100 * count. 1101 * 1102 * The buffer must be on QUEUE_NONE. 1103 */ 1104 1105void 1106bundirty(struct buf *bp) 1107{ 1108 1109 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1110 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1111 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1112 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1113 BUF_ASSERT_HELD(bp); 1114 1115 if (bp->b_flags & B_DELWRI) { 1116 bp->b_flags &= ~B_DELWRI; 1117 reassignbuf(bp); 1118 atomic_subtract_int(&numdirtybuffers, 1); 1119 numdirtywakeup(lodirtybuffers); 1120 } 1121 /* 1122 * Since it is now being written, we can clear its deferred write flag. 1123 */ 1124 bp->b_flags &= ~B_DEFERRED; 1125} 1126 1127/* 1128 * bawrite: 1129 * 1130 * Asynchronous write. Start output on a buffer, but do not wait for 1131 * it to complete. The buffer is released when the output completes. 1132 * 1133 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1134 * B_INVAL buffers. Not us. 1135 */ 1136void 1137bawrite(struct buf *bp) 1138{ 1139 1140 bp->b_flags |= B_ASYNC; 1141 (void) bwrite(bp); 1142} 1143 1144/* 1145 * bwillwrite: 1146 * 1147 * Called prior to the locking of any vnodes when we are expecting to 1148 * write. We do not want to starve the buffer cache with too many 1149 * dirty buffers so we block here. By blocking prior to the locking 1150 * of any vnodes we attempt to avoid the situation where a locked vnode 1151 * prevents the various system daemons from flushing related buffers. 1152 */ 1153 1154void 1155bwillwrite(void) 1156{ 1157 1158 if (numdirtybuffers >= hidirtybuffers) { 1159 mtx_lock(&nblock); 1160 while (numdirtybuffers >= hidirtybuffers) { 1161 bd_wakeup(1); 1162 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1163 msleep(&needsbuffer, &nblock, 1164 (PRIBIO + 4), "flswai", 0); 1165 } 1166 mtx_unlock(&nblock); 1167 } 1168} 1169 1170/* 1171 * Return true if we have too many dirty buffers. 1172 */ 1173int 1174buf_dirty_count_severe(void) 1175{ 1176 1177 return(numdirtybuffers >= hidirtybuffers); 1178} 1179 1180static __noinline int 1181buf_vm_page_count_severe(void) 1182{ 1183 1184 KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1185 1186 return vm_page_count_severe(); 1187} 1188 1189/* 1190 * brelse: 1191 * 1192 * Release a busy buffer and, if requested, free its resources. The 1193 * buffer will be stashed in the appropriate bufqueue[] allowing it 1194 * to be accessed later as a cache entity or reused for other purposes. 1195 */ 1196void 1197brelse(struct buf *bp) 1198{ 1199 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1200 bp, bp->b_vp, bp->b_flags); 1201 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1202 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1203 1204 if (bp->b_flags & B_MANAGED) { 1205 bqrelse(bp); 1206 return; 1207 } 1208 1209 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1210 bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1211 /* 1212 * Failed write, redirty. Must clear BIO_ERROR to prevent 1213 * pages from being scrapped. If the error is anything 1214 * other than an I/O error (EIO), assume that retrying 1215 * is futile. 1216 */ 1217 bp->b_ioflags &= ~BIO_ERROR; 1218 bdirty(bp); 1219 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1220 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1221 /* 1222 * Either a failed I/O or we were asked to free or not 1223 * cache the buffer. 1224 */ 1225 bp->b_flags |= B_INVAL; 1226 if (!LIST_EMPTY(&bp->b_dep)) 1227 buf_deallocate(bp); 1228 if (bp->b_flags & B_DELWRI) { 1229 atomic_subtract_int(&numdirtybuffers, 1); 1230 numdirtywakeup(lodirtybuffers); 1231 } 1232 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1233 if ((bp->b_flags & B_VMIO) == 0) { 1234 if (bp->b_bufsize) 1235 allocbuf(bp, 0); 1236 if (bp->b_vp) 1237 brelvp(bp); 1238 } 1239 } 1240 1241 /* 1242 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1243 * is called with B_DELWRI set, the underlying pages may wind up 1244 * getting freed causing a previous write (bdwrite()) to get 'lost' 1245 * because pages associated with a B_DELWRI bp are marked clean. 1246 * 1247 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1248 * if B_DELWRI is set. 1249 * 1250 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1251 * on pages to return pages to the VM page queues. 1252 */ 1253 if (bp->b_flags & B_DELWRI) 1254 bp->b_flags &= ~B_RELBUF; 1255 else if (buf_vm_page_count_severe()) { 1256 /* 1257 * The locking of the BO_LOCK is not necessary since 1258 * BKGRDINPROG cannot be set while we hold the buf 1259 * lock, it can only be cleared if it is already 1260 * pending. 1261 */ 1262 if (bp->b_vp) { 1263 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1264 bp->b_flags |= B_RELBUF; 1265 } else 1266 bp->b_flags |= B_RELBUF; 1267 } 1268 1269 /* 1270 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1271 * constituted, not even NFS buffers now. Two flags effect this. If 1272 * B_INVAL, the struct buf is invalidated but the VM object is kept 1273 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1274 * 1275 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1276 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1277 * buffer is also B_INVAL because it hits the re-dirtying code above. 1278 * 1279 * Normally we can do this whether a buffer is B_DELWRI or not. If 1280 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1281 * the commit state and we cannot afford to lose the buffer. If the 1282 * buffer has a background write in progress, we need to keep it 1283 * around to prevent it from being reconstituted and starting a second 1284 * background write. 1285 */ 1286 if ((bp->b_flags & B_VMIO) 1287 && !(bp->b_vp->v_mount != NULL && 1288 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1289 !vn_isdisk(bp->b_vp, NULL) && 1290 (bp->b_flags & B_DELWRI)) 1291 ) { 1292 1293 int i, j, resid; 1294 vm_page_t m; 1295 off_t foff; 1296 vm_pindex_t poff; 1297 vm_object_t obj; 1298 1299 obj = bp->b_bufobj->bo_object; 1300 1301 /* 1302 * Get the base offset and length of the buffer. Note that 1303 * in the VMIO case if the buffer block size is not 1304 * page-aligned then b_data pointer may not be page-aligned. 1305 * But our b_pages[] array *IS* page aligned. 1306 * 1307 * block sizes less then DEV_BSIZE (usually 512) are not 1308 * supported due to the page granularity bits (m->valid, 1309 * m->dirty, etc...). 1310 * 1311 * See man buf(9) for more information 1312 */ 1313 resid = bp->b_bufsize; 1314 foff = bp->b_offset; 1315 VM_OBJECT_LOCK(obj); 1316 for (i = 0; i < bp->b_npages; i++) { 1317 int had_bogus = 0; 1318 1319 m = bp->b_pages[i]; 1320 1321 /* 1322 * If we hit a bogus page, fixup *all* the bogus pages 1323 * now. 1324 */ 1325 if (m == bogus_page) { 1326 poff = OFF_TO_IDX(bp->b_offset); 1327 had_bogus = 1; 1328 1329 for (j = i; j < bp->b_npages; j++) { 1330 vm_page_t mtmp; 1331 mtmp = bp->b_pages[j]; 1332 if (mtmp == bogus_page) { 1333 mtmp = vm_page_lookup(obj, poff + j); 1334 if (!mtmp) { 1335 panic("brelse: page missing\n"); 1336 } 1337 bp->b_pages[j] = mtmp; 1338 } 1339 } 1340 1341 if ((bp->b_flags & B_INVAL) == 0) { 1342 pmap_qenter( 1343 trunc_page((vm_offset_t)bp->b_data), 1344 bp->b_pages, bp->b_npages); 1345 } 1346 m = bp->b_pages[i]; 1347 } 1348 if ((bp->b_flags & B_NOCACHE) || 1349 (bp->b_ioflags & BIO_ERROR && 1350 bp->b_iocmd == BIO_READ)) { 1351 int poffset = foff & PAGE_MASK; 1352 int presid = resid > (PAGE_SIZE - poffset) ? 1353 (PAGE_SIZE - poffset) : resid; 1354 1355 KASSERT(presid >= 0, ("brelse: extra page")); 1356 vm_page_set_invalid(m, poffset, presid); 1357 if (had_bogus) 1358 printf("avoided corruption bug in bogus_page/brelse code\n"); 1359 } 1360 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1361 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1362 } 1363 VM_OBJECT_UNLOCK(obj); 1364 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1365 vfs_vmio_release(bp); 1366 1367 } else if (bp->b_flags & B_VMIO) { 1368 1369 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1370 vfs_vmio_release(bp); 1371 } 1372 1373 } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1374 if (bp->b_bufsize != 0) 1375 allocbuf(bp, 0); 1376 if (bp->b_vp != NULL) 1377 brelvp(bp); 1378 } 1379 1380 if (BUF_LOCKRECURSED(bp)) { 1381 /* do not release to free list */ 1382 BUF_UNLOCK(bp); 1383 return; 1384 } 1385 1386 /* enqueue */ 1387 mtx_lock(&bqlock); 1388 /* Handle delayed bremfree() processing. */ 1389 if (bp->b_flags & B_REMFREE) 1390 bremfreel(bp); 1391 if (bp->b_qindex != QUEUE_NONE) 1392 panic("brelse: free buffer onto another queue???"); 1393 1394 /* 1395 * If the buffer has junk contents signal it and eventually 1396 * clean up B_DELWRI and diassociate the vnode so that gbincore() 1397 * doesn't find it. 1398 */ 1399 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1400 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1401 bp->b_flags |= B_INVAL; 1402 if (bp->b_flags & B_INVAL) { 1403 if (bp->b_flags & B_DELWRI) 1404 bundirty(bp); 1405 if (bp->b_vp) 1406 brelvp(bp); 1407 } 1408 1409 /* buffers with no memory */ 1410 if (bp->b_bufsize == 0) { 1411 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1412 if (bp->b_vflags & BV_BKGRDINPROG) 1413 panic("losing buffer 1"); 1414 if (bp->b_kvasize) { 1415 bp->b_qindex = QUEUE_EMPTYKVA; 1416 } else { 1417 bp->b_qindex = QUEUE_EMPTY; 1418 } 1419 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1420 /* buffers with junk contents */ 1421 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1422 (bp->b_ioflags & BIO_ERROR)) { 1423 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1424 if (bp->b_vflags & BV_BKGRDINPROG) 1425 panic("losing buffer 2"); 1426 bp->b_qindex = QUEUE_CLEAN; 1427 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1428 /* remaining buffers */ 1429 } else { 1430 if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1431 (B_DELWRI|B_NEEDSGIANT)) 1432 bp->b_qindex = QUEUE_DIRTY_GIANT; 1433 else if (bp->b_flags & B_DELWRI) 1434 bp->b_qindex = QUEUE_DIRTY; 1435 else 1436 bp->b_qindex = QUEUE_CLEAN; 1437 if (bp->b_flags & B_AGE) 1438 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1439 else 1440 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1441 } 1442 mtx_unlock(&bqlock); 1443 1444 /* 1445 * Fixup numfreebuffers count. The bp is on an appropriate queue 1446 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1447 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1448 * if B_INVAL is set ). 1449 */ 1450 1451 if (!(bp->b_flags & B_DELWRI)) 1452 bufcountwakeup(); 1453 1454 /* 1455 * Something we can maybe free or reuse 1456 */ 1457 if (bp->b_bufsize || bp->b_kvasize) 1458 bufspacewakeup(); 1459 1460 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1461 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1462 panic("brelse: not dirty"); 1463 /* unlock */ 1464 BUF_UNLOCK(bp); 1465} 1466 1467/* 1468 * Release a buffer back to the appropriate queue but do not try to free 1469 * it. The buffer is expected to be used again soon. 1470 * 1471 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1472 * biodone() to requeue an async I/O on completion. It is also used when 1473 * known good buffers need to be requeued but we think we may need the data 1474 * again soon. 1475 * 1476 * XXX we should be able to leave the B_RELBUF hint set on completion. 1477 */ 1478void 1479bqrelse(struct buf *bp) 1480{ 1481 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1482 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1483 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1484 1485 if (BUF_LOCKRECURSED(bp)) { 1486 /* do not release to free list */ 1487 BUF_UNLOCK(bp); 1488 return; 1489 } 1490 1491 if (bp->b_flags & B_MANAGED) { 1492 if (bp->b_flags & B_REMFREE) { 1493 mtx_lock(&bqlock); 1494 bremfreel(bp); 1495 mtx_unlock(&bqlock); 1496 } 1497 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1498 BUF_UNLOCK(bp); 1499 return; 1500 } 1501 1502 mtx_lock(&bqlock); 1503 /* Handle delayed bremfree() processing. */ 1504 if (bp->b_flags & B_REMFREE) 1505 bremfreel(bp); 1506 if (bp->b_qindex != QUEUE_NONE) 1507 panic("bqrelse: free buffer onto another queue???"); 1508 /* buffers with stale but valid contents */ 1509 if (bp->b_flags & B_DELWRI) { 1510 if (bp->b_flags & B_NEEDSGIANT) 1511 bp->b_qindex = QUEUE_DIRTY_GIANT; 1512 else 1513 bp->b_qindex = QUEUE_DIRTY; 1514 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1515 } else { 1516 /* 1517 * The locking of the BO_LOCK for checking of the 1518 * BV_BKGRDINPROG is not necessary since the 1519 * BV_BKGRDINPROG cannot be set while we hold the buf 1520 * lock, it can only be cleared if it is already 1521 * pending. 1522 */ 1523 if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1524 bp->b_qindex = QUEUE_CLEAN; 1525 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1526 b_freelist); 1527 } else { 1528 /* 1529 * We are too low on memory, we have to try to free 1530 * the buffer (most importantly: the wired pages 1531 * making up its backing store) *now*. 1532 */ 1533 mtx_unlock(&bqlock); 1534 brelse(bp); 1535 return; 1536 } 1537 } 1538 mtx_unlock(&bqlock); 1539 1540 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1541 bufcountwakeup(); 1542 1543 /* 1544 * Something we can maybe free or reuse. 1545 */ 1546 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1547 bufspacewakeup(); 1548 1549 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1550 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1551 panic("bqrelse: not dirty"); 1552 /* unlock */ 1553 BUF_UNLOCK(bp); 1554} 1555 1556/* Give pages used by the bp back to the VM system (where possible) */ 1557static void 1558vfs_vmio_release(struct buf *bp) 1559{ 1560 int i; 1561 vm_page_t m; 1562 1563 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1564 for (i = 0; i < bp->b_npages; i++) { 1565 m = bp->b_pages[i]; 1566 bp->b_pages[i] = NULL; 1567 /* 1568 * In order to keep page LRU ordering consistent, put 1569 * everything on the inactive queue. 1570 */ 1571 vm_page_lock(m); 1572 vm_page_unwire(m, 0); 1573 /* 1574 * We don't mess with busy pages, it is 1575 * the responsibility of the process that 1576 * busied the pages to deal with them. 1577 */ 1578 if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 && 1579 m->wire_count == 0) { 1580 /* 1581 * Might as well free the page if we can and it has 1582 * no valid data. We also free the page if the 1583 * buffer was used for direct I/O 1584 */ 1585 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1586 m->hold_count == 0) { 1587 vm_page_free(m); 1588 } else if (bp->b_flags & B_DIRECT) { 1589 vm_page_try_to_free(m); 1590 } else if (buf_vm_page_count_severe()) { 1591 vm_page_try_to_cache(m); 1592 } 1593 } 1594 vm_page_unlock(m); 1595 } 1596 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1597 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1598 1599 if (bp->b_bufsize) { 1600 bufspacewakeup(); 1601 bp->b_bufsize = 0; 1602 } 1603 bp->b_npages = 0; 1604 bp->b_flags &= ~B_VMIO; 1605 if (bp->b_vp) 1606 brelvp(bp); 1607} 1608 1609/* 1610 * Check to see if a block at a particular lbn is available for a clustered 1611 * write. 1612 */ 1613static int 1614vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1615{ 1616 struct buf *bpa; 1617 int match; 1618 1619 match = 0; 1620 1621 /* If the buf isn't in core skip it */ 1622 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1623 return (0); 1624 1625 /* If the buf is busy we don't want to wait for it */ 1626 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1627 return (0); 1628 1629 /* Only cluster with valid clusterable delayed write buffers */ 1630 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1631 (B_DELWRI | B_CLUSTEROK)) 1632 goto done; 1633 1634 if (bpa->b_bufsize != size) 1635 goto done; 1636 1637 /* 1638 * Check to see if it is in the expected place on disk and that the 1639 * block has been mapped. 1640 */ 1641 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1642 match = 1; 1643done: 1644 BUF_UNLOCK(bpa); 1645 return (match); 1646} 1647 1648/* 1649 * vfs_bio_awrite: 1650 * 1651 * Implement clustered async writes for clearing out B_DELWRI buffers. 1652 * This is much better then the old way of writing only one buffer at 1653 * a time. Note that we may not be presented with the buffers in the 1654 * correct order, so we search for the cluster in both directions. 1655 */ 1656int 1657vfs_bio_awrite(struct buf *bp) 1658{ 1659 struct bufobj *bo; 1660 int i; 1661 int j; 1662 daddr_t lblkno = bp->b_lblkno; 1663 struct vnode *vp = bp->b_vp; 1664 int ncl; 1665 int nwritten; 1666 int size; 1667 int maxcl; 1668 1669 bo = &vp->v_bufobj; 1670 /* 1671 * right now we support clustered writing only to regular files. If 1672 * we find a clusterable block we could be in the middle of a cluster 1673 * rather then at the beginning. 1674 */ 1675 if ((vp->v_type == VREG) && 1676 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1677 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1678 1679 size = vp->v_mount->mnt_stat.f_iosize; 1680 maxcl = MAXPHYS / size; 1681 1682 BO_LOCK(bo); 1683 for (i = 1; i < maxcl; i++) 1684 if (vfs_bio_clcheck(vp, size, lblkno + i, 1685 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1686 break; 1687 1688 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1689 if (vfs_bio_clcheck(vp, size, lblkno - j, 1690 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1691 break; 1692 BO_UNLOCK(bo); 1693 --j; 1694 ncl = i + j; 1695 /* 1696 * this is a possible cluster write 1697 */ 1698 if (ncl != 1) { 1699 BUF_UNLOCK(bp); 1700 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1701 return nwritten; 1702 } 1703 } 1704 bremfree(bp); 1705 bp->b_flags |= B_ASYNC; 1706 /* 1707 * default (old) behavior, writing out only one block 1708 * 1709 * XXX returns b_bufsize instead of b_bcount for nwritten? 1710 */ 1711 nwritten = bp->b_bufsize; 1712 (void) bwrite(bp); 1713 1714 return nwritten; 1715} 1716 1717/* 1718 * getnewbuf: 1719 * 1720 * Find and initialize a new buffer header, freeing up existing buffers 1721 * in the bufqueues as necessary. The new buffer is returned locked. 1722 * 1723 * Important: B_INVAL is not set. If the caller wishes to throw the 1724 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1725 * 1726 * We block if: 1727 * We have insufficient buffer headers 1728 * We have insufficient buffer space 1729 * buffer_map is too fragmented ( space reservation fails ) 1730 * If we have to flush dirty buffers ( but we try to avoid this ) 1731 * 1732 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1733 * Instead we ask the buf daemon to do it for us. We attempt to 1734 * avoid piecemeal wakeups of the pageout daemon. 1735 */ 1736 1737static struct buf * 1738getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 1739 int gbflags) 1740{ 1741 struct thread *td; 1742 struct buf *bp; 1743 struct buf *nbp; 1744 int defrag = 0; 1745 int nqindex; 1746 static int flushingbufs; 1747 1748 td = curthread; 1749 /* 1750 * We can't afford to block since we might be holding a vnode lock, 1751 * which may prevent system daemons from running. We deal with 1752 * low-memory situations by proactively returning memory and running 1753 * async I/O rather then sync I/O. 1754 */ 1755 atomic_add_int(&getnewbufcalls, 1); 1756 atomic_subtract_int(&getnewbufrestarts, 1); 1757restart: 1758 atomic_add_int(&getnewbufrestarts, 1); 1759 1760 /* 1761 * Setup for scan. If we do not have enough free buffers, 1762 * we setup a degenerate case that immediately fails. Note 1763 * that if we are specially marked process, we are allowed to 1764 * dip into our reserves. 1765 * 1766 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1767 * 1768 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1769 * However, there are a number of cases (defragging, reusing, ...) 1770 * where we cannot backup. 1771 */ 1772 mtx_lock(&bqlock); 1773 nqindex = QUEUE_EMPTYKVA; 1774 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1775 1776 if (nbp == NULL) { 1777 /* 1778 * If no EMPTYKVA buffers and we are either 1779 * defragging or reusing, locate a CLEAN buffer 1780 * to free or reuse. If bufspace useage is low 1781 * skip this step so we can allocate a new buffer. 1782 */ 1783 if (defrag || bufspace >= lobufspace) { 1784 nqindex = QUEUE_CLEAN; 1785 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1786 } 1787 1788 /* 1789 * If we could not find or were not allowed to reuse a 1790 * CLEAN buffer, check to see if it is ok to use an EMPTY 1791 * buffer. We can only use an EMPTY buffer if allocating 1792 * its KVA would not otherwise run us out of buffer space. 1793 */ 1794 if (nbp == NULL && defrag == 0 && 1795 bufspace + maxsize < hibufspace) { 1796 nqindex = QUEUE_EMPTY; 1797 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1798 } 1799 } 1800 1801 /* 1802 * Run scan, possibly freeing data and/or kva mappings on the fly 1803 * depending. 1804 */ 1805 1806 while ((bp = nbp) != NULL) { 1807 int qindex = nqindex; 1808 1809 /* 1810 * Calculate next bp ( we can only use it if we do not block 1811 * or do other fancy things ). 1812 */ 1813 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1814 switch(qindex) { 1815 case QUEUE_EMPTY: 1816 nqindex = QUEUE_EMPTYKVA; 1817 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1818 break; 1819 /* FALLTHROUGH */ 1820 case QUEUE_EMPTYKVA: 1821 nqindex = QUEUE_CLEAN; 1822 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1823 break; 1824 /* FALLTHROUGH */ 1825 case QUEUE_CLEAN: 1826 /* 1827 * nbp is NULL. 1828 */ 1829 break; 1830 } 1831 } 1832 /* 1833 * If we are defragging then we need a buffer with 1834 * b_kvasize != 0. XXX this situation should no longer 1835 * occur, if defrag is non-zero the buffer's b_kvasize 1836 * should also be non-zero at this point. XXX 1837 */ 1838 if (defrag && bp->b_kvasize == 0) { 1839 printf("Warning: defrag empty buffer %p\n", bp); 1840 continue; 1841 } 1842 1843 /* 1844 * Start freeing the bp. This is somewhat involved. nbp 1845 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1846 */ 1847 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1848 continue; 1849 if (bp->b_vp) { 1850 BO_LOCK(bp->b_bufobj); 1851 if (bp->b_vflags & BV_BKGRDINPROG) { 1852 BO_UNLOCK(bp->b_bufobj); 1853 BUF_UNLOCK(bp); 1854 continue; 1855 } 1856 BO_UNLOCK(bp->b_bufobj); 1857 } 1858 CTR6(KTR_BUF, 1859 "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1860 "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1861 bp->b_kvasize, bp->b_bufsize, qindex); 1862 1863 /* 1864 * Sanity Checks 1865 */ 1866 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1867 1868 /* 1869 * Note: we no longer distinguish between VMIO and non-VMIO 1870 * buffers. 1871 */ 1872 1873 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1874 1875 bremfreel(bp); 1876 mtx_unlock(&bqlock); 1877 1878 if (qindex == QUEUE_CLEAN) { 1879 if (bp->b_flags & B_VMIO) { 1880 bp->b_flags &= ~B_ASYNC; 1881 vfs_vmio_release(bp); 1882 } 1883 if (bp->b_vp) 1884 brelvp(bp); 1885 } 1886 1887 /* 1888 * NOTE: nbp is now entirely invalid. We can only restart 1889 * the scan from this point on. 1890 * 1891 * Get the rest of the buffer freed up. b_kva* is still 1892 * valid after this operation. 1893 */ 1894 1895 if (bp->b_rcred != NOCRED) { 1896 crfree(bp->b_rcred); 1897 bp->b_rcred = NOCRED; 1898 } 1899 if (bp->b_wcred != NOCRED) { 1900 crfree(bp->b_wcred); 1901 bp->b_wcred = NOCRED; 1902 } 1903 if (!LIST_EMPTY(&bp->b_dep)) 1904 buf_deallocate(bp); 1905 if (bp->b_vflags & BV_BKGRDINPROG) 1906 panic("losing buffer 3"); 1907 KASSERT(bp->b_vp == NULL, 1908 ("bp: %p still has vnode %p. qindex: %d", 1909 bp, bp->b_vp, qindex)); 1910 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1911 ("bp: %p still on a buffer list. xflags %X", 1912 bp, bp->b_xflags)); 1913 1914 if (bp->b_bufsize) 1915 allocbuf(bp, 0); 1916 1917 bp->b_flags = 0; 1918 bp->b_ioflags = 0; 1919 bp->b_xflags = 0; 1920 bp->b_vflags = 0; 1921 bp->b_vp = NULL; 1922 bp->b_blkno = bp->b_lblkno = 0; 1923 bp->b_offset = NOOFFSET; 1924 bp->b_iodone = 0; 1925 bp->b_error = 0; 1926 bp->b_resid = 0; 1927 bp->b_bcount = 0; 1928 bp->b_npages = 0; 1929 bp->b_dirtyoff = bp->b_dirtyend = 0; 1930 bp->b_bufobj = NULL; 1931 bp->b_pin_count = 0; 1932 bp->b_fsprivate1 = NULL; 1933 bp->b_fsprivate2 = NULL; 1934 bp->b_fsprivate3 = NULL; 1935 1936 LIST_INIT(&bp->b_dep); 1937 1938 /* 1939 * If we are defragging then free the buffer. 1940 */ 1941 if (defrag) { 1942 bp->b_flags |= B_INVAL; 1943 bfreekva(bp); 1944 brelse(bp); 1945 defrag = 0; 1946 goto restart; 1947 } 1948 1949 /* 1950 * Notify any waiters for the buffer lock about 1951 * identity change by freeing the buffer. 1952 */ 1953 if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 1954 bp->b_flags |= B_INVAL; 1955 bfreekva(bp); 1956 brelse(bp); 1957 goto restart; 1958 } 1959 1960 /* 1961 * If we are overcomitted then recover the buffer and its 1962 * KVM space. This occurs in rare situations when multiple 1963 * processes are blocked in getnewbuf() or allocbuf(). 1964 */ 1965 if (bufspace >= hibufspace) 1966 flushingbufs = 1; 1967 if (flushingbufs && bp->b_kvasize != 0) { 1968 bp->b_flags |= B_INVAL; 1969 bfreekva(bp); 1970 brelse(bp); 1971 goto restart; 1972 } 1973 if (bufspace < lobufspace) 1974 flushingbufs = 0; 1975 break; 1976 } 1977 1978 /* 1979 * If we exhausted our list, sleep as appropriate. We may have to 1980 * wakeup various daemons and write out some dirty buffers. 1981 * 1982 * Generally we are sleeping due to insufficient buffer space. 1983 */ 1984 1985 if (bp == NULL) { 1986 int flags, norunbuf; 1987 char *waitmsg; 1988 int fl; 1989 1990 if (defrag) { 1991 flags = VFS_BIO_NEED_BUFSPACE; 1992 waitmsg = "nbufkv"; 1993 } else if (bufspace >= hibufspace) { 1994 waitmsg = "nbufbs"; 1995 flags = VFS_BIO_NEED_BUFSPACE; 1996 } else { 1997 waitmsg = "newbuf"; 1998 flags = VFS_BIO_NEED_ANY; 1999 } 2000 mtx_lock(&nblock); 2001 needsbuffer |= flags; 2002 mtx_unlock(&nblock); 2003 mtx_unlock(&bqlock); 2004 2005 bd_speedup(); /* heeeelp */ 2006 if (gbflags & GB_NOWAIT_BD) 2007 return (NULL); 2008 2009 mtx_lock(&nblock); 2010 while (needsbuffer & flags) { 2011 if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2012 mtx_unlock(&nblock); 2013 /* 2014 * getblk() is called with a vnode 2015 * locked, and some majority of the 2016 * dirty buffers may as well belong to 2017 * the vnode. Flushing the buffers 2018 * there would make a progress that 2019 * cannot be achieved by the 2020 * buf_daemon, that cannot lock the 2021 * vnode. 2022 */ 2023 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2024 (td->td_pflags & TDP_NORUNNINGBUF); 2025 /* play bufdaemon */ 2026 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2027 fl = buf_do_flush(vp); 2028 td->td_pflags &= norunbuf; 2029 mtx_lock(&nblock); 2030 if (fl != 0) 2031 continue; 2032 if ((needsbuffer & flags) == 0) 2033 break; 2034 } 2035 if (msleep(&needsbuffer, &nblock, 2036 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2037 mtx_unlock(&nblock); 2038 return (NULL); 2039 } 2040 } 2041 mtx_unlock(&nblock); 2042 } else { 2043 /* 2044 * We finally have a valid bp. We aren't quite out of the 2045 * woods, we still have to reserve kva space. In order 2046 * to keep fragmentation sane we only allocate kva in 2047 * BKVASIZE chunks. 2048 */ 2049 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2050 2051 if (maxsize != bp->b_kvasize) { 2052 vm_offset_t addr = 0; 2053 2054 bfreekva(bp); 2055 2056 vm_map_lock(buffer_map); 2057 if (vm_map_findspace(buffer_map, 2058 vm_map_min(buffer_map), maxsize, &addr)) { 2059 /* 2060 * Uh oh. Buffer map is to fragmented. We 2061 * must defragment the map. 2062 */ 2063 atomic_add_int(&bufdefragcnt, 1); 2064 vm_map_unlock(buffer_map); 2065 defrag = 1; 2066 bp->b_flags |= B_INVAL; 2067 brelse(bp); 2068 goto restart; 2069 } 2070 if (addr) { 2071 vm_map_insert(buffer_map, NULL, 0, 2072 addr, addr + maxsize, 2073 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2074 2075 bp->b_kvabase = (caddr_t) addr; 2076 bp->b_kvasize = maxsize; 2077 atomic_add_long(&bufspace, bp->b_kvasize); 2078 atomic_add_int(&bufreusecnt, 1); 2079 } 2080 vm_map_unlock(buffer_map); 2081 } 2082 bp->b_saveaddr = bp->b_kvabase; 2083 bp->b_data = bp->b_saveaddr; 2084 } 2085 return(bp); 2086} 2087 2088/* 2089 * buf_daemon: 2090 * 2091 * buffer flushing daemon. Buffers are normally flushed by the 2092 * update daemon but if it cannot keep up this process starts to 2093 * take the load in an attempt to prevent getnewbuf() from blocking. 2094 */ 2095 2096static struct kproc_desc buf_kp = { 2097 "bufdaemon", 2098 buf_daemon, 2099 &bufdaemonproc 2100}; 2101SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2102 2103static int 2104buf_do_flush(struct vnode *vp) 2105{ 2106 int flushed; 2107 2108 flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2109 /* The list empty check here is slightly racy */ 2110 if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2111 mtx_lock(&Giant); 2112 flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0); 2113 mtx_unlock(&Giant); 2114 } 2115 if (flushed == 0) { 2116 /* 2117 * Could not find any buffers without rollback 2118 * dependencies, so just write the first one 2119 * in the hopes of eventually making progress. 2120 */ 2121 flushbufqueues(vp, QUEUE_DIRTY, 1); 2122 if (!TAILQ_EMPTY( 2123 &bufqueues[QUEUE_DIRTY_GIANT])) { 2124 mtx_lock(&Giant); 2125 flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1); 2126 mtx_unlock(&Giant); 2127 } 2128 } 2129 return (flushed); 2130} 2131 2132static void 2133buf_daemon() 2134{ 2135 int lodirtysave; 2136 2137 /* 2138 * This process needs to be suspended prior to shutdown sync. 2139 */ 2140 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2141 SHUTDOWN_PRI_LAST); 2142 2143 /* 2144 * This process is allowed to take the buffer cache to the limit 2145 */ 2146 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2147 mtx_lock(&bdlock); 2148 for (;;) { 2149 bd_request = 0; 2150 mtx_unlock(&bdlock); 2151 2152 kproc_suspend_check(bufdaemonproc); 2153 lodirtysave = lodirtybuffers; 2154 if (bd_speedupreq) { 2155 lodirtybuffers = numdirtybuffers / 2; 2156 bd_speedupreq = 0; 2157 } 2158 /* 2159 * Do the flush. Limit the amount of in-transit I/O we 2160 * allow to build up, otherwise we would completely saturate 2161 * the I/O system. Wakeup any waiting processes before we 2162 * normally would so they can run in parallel with our drain. 2163 */ 2164 while (numdirtybuffers > lodirtybuffers) { 2165 if (buf_do_flush(NULL) == 0) 2166 break; 2167 uio_yield(); 2168 } 2169 lodirtybuffers = lodirtysave; 2170 2171 /* 2172 * Only clear bd_request if we have reached our low water 2173 * mark. The buf_daemon normally waits 1 second and 2174 * then incrementally flushes any dirty buffers that have 2175 * built up, within reason. 2176 * 2177 * If we were unable to hit our low water mark and couldn't 2178 * find any flushable buffers, we sleep half a second. 2179 * Otherwise we loop immediately. 2180 */ 2181 mtx_lock(&bdlock); 2182 if (numdirtybuffers <= lodirtybuffers) { 2183 /* 2184 * We reached our low water mark, reset the 2185 * request and sleep until we are needed again. 2186 * The sleep is just so the suspend code works. 2187 */ 2188 bd_request = 0; 2189 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2190 } else { 2191 /* 2192 * We couldn't find any flushable dirty buffers but 2193 * still have too many dirty buffers, we 2194 * have to sleep and try again. (rare) 2195 */ 2196 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2197 } 2198 } 2199} 2200 2201/* 2202 * flushbufqueues: 2203 * 2204 * Try to flush a buffer in the dirty queue. We must be careful to 2205 * free up B_INVAL buffers instead of write them, which NFS is 2206 * particularly sensitive to. 2207 */ 2208static int flushwithdeps = 0; 2209SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2210 0, "Number of buffers flushed with dependecies that require rollbacks"); 2211 2212static int 2213flushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2214{ 2215 struct buf *sentinel; 2216 struct vnode *vp; 2217 struct mount *mp; 2218 struct buf *bp; 2219 int hasdeps; 2220 int flushed; 2221 int target; 2222 2223 if (lvp == NULL) { 2224 target = numdirtybuffers - lodirtybuffers; 2225 if (flushdeps && target > 2) 2226 target /= 2; 2227 } else 2228 target = flushbufqtarget; 2229 flushed = 0; 2230 bp = NULL; 2231 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2232 sentinel->b_qindex = QUEUE_SENTINEL; 2233 mtx_lock(&bqlock); 2234 TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2235 while (flushed != target) { 2236 bp = TAILQ_NEXT(sentinel, b_freelist); 2237 if (bp != NULL) { 2238 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2239 TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2240 b_freelist); 2241 } else 2242 break; 2243 /* 2244 * Skip sentinels inserted by other invocations of the 2245 * flushbufqueues(), taking care to not reorder them. 2246 */ 2247 if (bp->b_qindex == QUEUE_SENTINEL) 2248 continue; 2249 /* 2250 * Only flush the buffers that belong to the 2251 * vnode locked by the curthread. 2252 */ 2253 if (lvp != NULL && bp->b_vp != lvp) 2254 continue; 2255 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2256 continue; 2257 if (bp->b_pin_count > 0) { 2258 BUF_UNLOCK(bp); 2259 continue; 2260 } 2261 BO_LOCK(bp->b_bufobj); 2262 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2263 (bp->b_flags & B_DELWRI) == 0) { 2264 BO_UNLOCK(bp->b_bufobj); 2265 BUF_UNLOCK(bp); 2266 continue; 2267 } 2268 BO_UNLOCK(bp->b_bufobj); 2269 if (bp->b_flags & B_INVAL) { 2270 bremfreel(bp); 2271 mtx_unlock(&bqlock); 2272 brelse(bp); 2273 flushed++; 2274 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2275 mtx_lock(&bqlock); 2276 continue; 2277 } 2278 2279 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2280 if (flushdeps == 0) { 2281 BUF_UNLOCK(bp); 2282 continue; 2283 } 2284 hasdeps = 1; 2285 } else 2286 hasdeps = 0; 2287 /* 2288 * We must hold the lock on a vnode before writing 2289 * one of its buffers. Otherwise we may confuse, or 2290 * in the case of a snapshot vnode, deadlock the 2291 * system. 2292 * 2293 * The lock order here is the reverse of the normal 2294 * of vnode followed by buf lock. This is ok because 2295 * the NOWAIT will prevent deadlock. 2296 */ 2297 vp = bp->b_vp; 2298 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2299 BUF_UNLOCK(bp); 2300 continue; 2301 } 2302 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2303 mtx_unlock(&bqlock); 2304 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2305 bp, bp->b_vp, bp->b_flags); 2306 if (curproc == bufdaemonproc) 2307 vfs_bio_awrite(bp); 2308 else { 2309 bremfree(bp); 2310 bwrite(bp); 2311 notbufdflashes++; 2312 } 2313 vn_finished_write(mp); 2314 VOP_UNLOCK(vp, 0); 2315 flushwithdeps += hasdeps; 2316 flushed++; 2317 2318 /* 2319 * Sleeping on runningbufspace while holding 2320 * vnode lock leads to deadlock. 2321 */ 2322 if (curproc == bufdaemonproc) 2323 waitrunningbufspace(); 2324 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2325 mtx_lock(&bqlock); 2326 continue; 2327 } 2328 vn_finished_write(mp); 2329 BUF_UNLOCK(bp); 2330 } 2331 TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2332 mtx_unlock(&bqlock); 2333 free(sentinel, M_TEMP); 2334 return (flushed); 2335} 2336 2337/* 2338 * Check to see if a block is currently memory resident. 2339 */ 2340struct buf * 2341incore(struct bufobj *bo, daddr_t blkno) 2342{ 2343 struct buf *bp; 2344 2345 BO_LOCK(bo); 2346 bp = gbincore(bo, blkno); 2347 BO_UNLOCK(bo); 2348 return (bp); 2349} 2350 2351/* 2352 * Returns true if no I/O is needed to access the 2353 * associated VM object. This is like incore except 2354 * it also hunts around in the VM system for the data. 2355 */ 2356 2357static int 2358inmem(struct vnode * vp, daddr_t blkno) 2359{ 2360 vm_object_t obj; 2361 vm_offset_t toff, tinc, size; 2362 vm_page_t m; 2363 vm_ooffset_t off; 2364 2365 ASSERT_VOP_LOCKED(vp, "inmem"); 2366 2367 if (incore(&vp->v_bufobj, blkno)) 2368 return 1; 2369 if (vp->v_mount == NULL) 2370 return 0; 2371 obj = vp->v_object; 2372 if (obj == NULL) 2373 return (0); 2374 2375 size = PAGE_SIZE; 2376 if (size > vp->v_mount->mnt_stat.f_iosize) 2377 size = vp->v_mount->mnt_stat.f_iosize; 2378 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2379 2380 VM_OBJECT_LOCK(obj); 2381 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2382 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2383 if (!m) 2384 goto notinmem; 2385 tinc = size; 2386 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2387 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2388 if (vm_page_is_valid(m, 2389 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2390 goto notinmem; 2391 } 2392 VM_OBJECT_UNLOCK(obj); 2393 return 1; 2394 2395notinmem: 2396 VM_OBJECT_UNLOCK(obj); 2397 return (0); 2398} 2399 2400/* 2401 * vfs_setdirty: 2402 * 2403 * Sets the dirty range for a buffer based on the status of the dirty 2404 * bits in the pages comprising the buffer. 2405 * 2406 * The range is limited to the size of the buffer. 2407 * 2408 * This routine is primarily used by NFS, but is generalized for the 2409 * B_VMIO case. 2410 */ 2411static void 2412vfs_setdirty(struct buf *bp) 2413{ 2414 2415 /* 2416 * Degenerate case - empty buffer 2417 */ 2418 if (bp->b_bufsize == 0) 2419 return; 2420 2421 if ((bp->b_flags & B_VMIO) == 0) 2422 return; 2423 2424 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2425 vfs_setdirty_locked_object(bp); 2426 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2427} 2428 2429static void 2430vfs_setdirty_locked_object(struct buf *bp) 2431{ 2432 vm_object_t object; 2433 int i; 2434 2435 object = bp->b_bufobj->bo_object; 2436 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2437 2438 /* 2439 * We qualify the scan for modified pages on whether the 2440 * object has been flushed yet. 2441 */ 2442 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2443 vm_offset_t boffset; 2444 vm_offset_t eoffset; 2445 2446 /* 2447 * test the pages to see if they have been modified directly 2448 * by users through the VM system. 2449 */ 2450 for (i = 0; i < bp->b_npages; i++) 2451 vm_page_test_dirty(bp->b_pages[i]); 2452 2453 /* 2454 * Calculate the encompassing dirty range, boffset and eoffset, 2455 * (eoffset - boffset) bytes. 2456 */ 2457 2458 for (i = 0; i < bp->b_npages; i++) { 2459 if (bp->b_pages[i]->dirty) 2460 break; 2461 } 2462 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2463 2464 for (i = bp->b_npages - 1; i >= 0; --i) { 2465 if (bp->b_pages[i]->dirty) { 2466 break; 2467 } 2468 } 2469 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2470 2471 /* 2472 * Fit it to the buffer. 2473 */ 2474 2475 if (eoffset > bp->b_bcount) 2476 eoffset = bp->b_bcount; 2477 2478 /* 2479 * If we have a good dirty range, merge with the existing 2480 * dirty range. 2481 */ 2482 2483 if (boffset < eoffset) { 2484 if (bp->b_dirtyoff > boffset) 2485 bp->b_dirtyoff = boffset; 2486 if (bp->b_dirtyend < eoffset) 2487 bp->b_dirtyend = eoffset; 2488 } 2489 } 2490} 2491 2492/* 2493 * getblk: 2494 * 2495 * Get a block given a specified block and offset into a file/device. 2496 * The buffers B_DONE bit will be cleared on return, making it almost 2497 * ready for an I/O initiation. B_INVAL may or may not be set on 2498 * return. The caller should clear B_INVAL prior to initiating a 2499 * READ. 2500 * 2501 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2502 * an existing buffer. 2503 * 2504 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2505 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2506 * and then cleared based on the backing VM. If the previous buffer is 2507 * non-0-sized but invalid, B_CACHE will be cleared. 2508 * 2509 * If getblk() must create a new buffer, the new buffer is returned with 2510 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2511 * case it is returned with B_INVAL clear and B_CACHE set based on the 2512 * backing VM. 2513 * 2514 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2515 * B_CACHE bit is clear. 2516 * 2517 * What this means, basically, is that the caller should use B_CACHE to 2518 * determine whether the buffer is fully valid or not and should clear 2519 * B_INVAL prior to issuing a read. If the caller intends to validate 2520 * the buffer by loading its data area with something, the caller needs 2521 * to clear B_INVAL. If the caller does this without issuing an I/O, 2522 * the caller should set B_CACHE ( as an optimization ), else the caller 2523 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2524 * a write attempt or if it was a successfull read. If the caller 2525 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2526 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2527 */ 2528struct buf * 2529getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2530 int flags) 2531{ 2532 struct buf *bp; 2533 struct bufobj *bo; 2534 int error; 2535 2536 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2537 ASSERT_VOP_LOCKED(vp, "getblk"); 2538 if (size > MAXBSIZE) 2539 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2540 2541 bo = &vp->v_bufobj; 2542loop: 2543 /* 2544 * Block if we are low on buffers. Certain processes are allowed 2545 * to completely exhaust the buffer cache. 2546 * 2547 * If this check ever becomes a bottleneck it may be better to 2548 * move it into the else, when gbincore() fails. At the moment 2549 * it isn't a problem. 2550 * 2551 * XXX remove if 0 sections (clean this up after its proven) 2552 */ 2553 if (numfreebuffers == 0) { 2554 if (TD_IS_IDLETHREAD(curthread)) 2555 return NULL; 2556 mtx_lock(&nblock); 2557 needsbuffer |= VFS_BIO_NEED_ANY; 2558 mtx_unlock(&nblock); 2559 } 2560 2561 BO_LOCK(bo); 2562 bp = gbincore(bo, blkno); 2563 if (bp != NULL) { 2564 int lockflags; 2565 /* 2566 * Buffer is in-core. If the buffer is not busy, it must 2567 * be on a queue. 2568 */ 2569 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2570 2571 if (flags & GB_LOCK_NOWAIT) 2572 lockflags |= LK_NOWAIT; 2573 2574 error = BUF_TIMELOCK(bp, lockflags, 2575 BO_MTX(bo), "getblk", slpflag, slptimeo); 2576 2577 /* 2578 * If we slept and got the lock we have to restart in case 2579 * the buffer changed identities. 2580 */ 2581 if (error == ENOLCK) 2582 goto loop; 2583 /* We timed out or were interrupted. */ 2584 else if (error) 2585 return (NULL); 2586 2587 /* 2588 * The buffer is locked. B_CACHE is cleared if the buffer is 2589 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2590 * and for a VMIO buffer B_CACHE is adjusted according to the 2591 * backing VM cache. 2592 */ 2593 if (bp->b_flags & B_INVAL) 2594 bp->b_flags &= ~B_CACHE; 2595 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2596 bp->b_flags |= B_CACHE; 2597 bremfree(bp); 2598 2599 /* 2600 * check for size inconsistancies for non-VMIO case. 2601 */ 2602 2603 if (bp->b_bcount != size) { 2604 if ((bp->b_flags & B_VMIO) == 0 || 2605 (size > bp->b_kvasize)) { 2606 if (bp->b_flags & B_DELWRI) { 2607 /* 2608 * If buffer is pinned and caller does 2609 * not want sleep waiting for it to be 2610 * unpinned, bail out 2611 * */ 2612 if (bp->b_pin_count > 0) { 2613 if (flags & GB_LOCK_NOWAIT) { 2614 bqrelse(bp); 2615 return (NULL); 2616 } else { 2617 bunpin_wait(bp); 2618 } 2619 } 2620 bp->b_flags |= B_NOCACHE; 2621 bwrite(bp); 2622 } else { 2623 if (LIST_EMPTY(&bp->b_dep)) { 2624 bp->b_flags |= B_RELBUF; 2625 brelse(bp); 2626 } else { 2627 bp->b_flags |= B_NOCACHE; 2628 bwrite(bp); 2629 } 2630 } 2631 goto loop; 2632 } 2633 } 2634 2635 /* 2636 * If the size is inconsistant in the VMIO case, we can resize 2637 * the buffer. This might lead to B_CACHE getting set or 2638 * cleared. If the size has not changed, B_CACHE remains 2639 * unchanged from its previous state. 2640 */ 2641 2642 if (bp->b_bcount != size) 2643 allocbuf(bp, size); 2644 2645 KASSERT(bp->b_offset != NOOFFSET, 2646 ("getblk: no buffer offset")); 2647 2648 /* 2649 * A buffer with B_DELWRI set and B_CACHE clear must 2650 * be committed before we can return the buffer in 2651 * order to prevent the caller from issuing a read 2652 * ( due to B_CACHE not being set ) and overwriting 2653 * it. 2654 * 2655 * Most callers, including NFS and FFS, need this to 2656 * operate properly either because they assume they 2657 * can issue a read if B_CACHE is not set, or because 2658 * ( for example ) an uncached B_DELWRI might loop due 2659 * to softupdates re-dirtying the buffer. In the latter 2660 * case, B_CACHE is set after the first write completes, 2661 * preventing further loops. 2662 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2663 * above while extending the buffer, we cannot allow the 2664 * buffer to remain with B_CACHE set after the write 2665 * completes or it will represent a corrupt state. To 2666 * deal with this we set B_NOCACHE to scrap the buffer 2667 * after the write. 2668 * 2669 * We might be able to do something fancy, like setting 2670 * B_CACHE in bwrite() except if B_DELWRI is already set, 2671 * so the below call doesn't set B_CACHE, but that gets real 2672 * confusing. This is much easier. 2673 */ 2674 2675 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2676 bp->b_flags |= B_NOCACHE; 2677 bwrite(bp); 2678 goto loop; 2679 } 2680 bp->b_flags &= ~B_DONE; 2681 } else { 2682 int bsize, maxsize, vmio; 2683 off_t offset; 2684 2685 /* 2686 * Buffer is not in-core, create new buffer. The buffer 2687 * returned by getnewbuf() is locked. Note that the returned 2688 * buffer is also considered valid (not marked B_INVAL). 2689 */ 2690 BO_UNLOCK(bo); 2691 /* 2692 * If the user does not want us to create the buffer, bail out 2693 * here. 2694 */ 2695 if (flags & GB_NOCREAT) 2696 return NULL; 2697 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 2698 offset = blkno * bsize; 2699 vmio = vp->v_object != NULL; 2700 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2701 maxsize = imax(maxsize, bsize); 2702 2703 bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 2704 if (bp == NULL) { 2705 if (slpflag || slptimeo) 2706 return NULL; 2707 goto loop; 2708 } 2709 2710 /* 2711 * This code is used to make sure that a buffer is not 2712 * created while the getnewbuf routine is blocked. 2713 * This can be a problem whether the vnode is locked or not. 2714 * If the buffer is created out from under us, we have to 2715 * throw away the one we just created. 2716 * 2717 * Note: this must occur before we associate the buffer 2718 * with the vp especially considering limitations in 2719 * the splay tree implementation when dealing with duplicate 2720 * lblkno's. 2721 */ 2722 BO_LOCK(bo); 2723 if (gbincore(bo, blkno)) { 2724 BO_UNLOCK(bo); 2725 bp->b_flags |= B_INVAL; 2726 brelse(bp); 2727 goto loop; 2728 } 2729 2730 /* 2731 * Insert the buffer into the hash, so that it can 2732 * be found by incore. 2733 */ 2734 bp->b_blkno = bp->b_lblkno = blkno; 2735 bp->b_offset = offset; 2736 bgetvp(vp, bp); 2737 BO_UNLOCK(bo); 2738 2739 /* 2740 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2741 * buffer size starts out as 0, B_CACHE will be set by 2742 * allocbuf() for the VMIO case prior to it testing the 2743 * backing store for validity. 2744 */ 2745 2746 if (vmio) { 2747 bp->b_flags |= B_VMIO; 2748#if defined(VFS_BIO_DEBUG) 2749 if (vn_canvmio(vp) != TRUE) 2750 printf("getblk: VMIO on vnode type %d\n", 2751 vp->v_type); 2752#endif 2753 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2754 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2755 bp, vp->v_object, bp->b_bufobj->bo_object)); 2756 } else { 2757 bp->b_flags &= ~B_VMIO; 2758 KASSERT(bp->b_bufobj->bo_object == NULL, 2759 ("ARGH! has b_bufobj->bo_object %p %p\n", 2760 bp, bp->b_bufobj->bo_object)); 2761 } 2762 2763 allocbuf(bp, size); 2764 bp->b_flags &= ~B_DONE; 2765 } 2766 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2767 BUF_ASSERT_HELD(bp); 2768 KASSERT(bp->b_bufobj == bo, 2769 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2770 return (bp); 2771} 2772 2773/* 2774 * Get an empty, disassociated buffer of given size. The buffer is initially 2775 * set to B_INVAL. 2776 */ 2777struct buf * 2778geteblk(int size, int flags) 2779{ 2780 struct buf *bp; 2781 int maxsize; 2782 2783 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2784 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 2785 if ((flags & GB_NOWAIT_BD) && 2786 (curthread->td_pflags & TDP_BUFNEED) != 0) 2787 return (NULL); 2788 } 2789 allocbuf(bp, size); 2790 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2791 BUF_ASSERT_HELD(bp); 2792 return (bp); 2793} 2794 2795 2796/* 2797 * This code constitutes the buffer memory from either anonymous system 2798 * memory (in the case of non-VMIO operations) or from an associated 2799 * VM object (in the case of VMIO operations). This code is able to 2800 * resize a buffer up or down. 2801 * 2802 * Note that this code is tricky, and has many complications to resolve 2803 * deadlock or inconsistant data situations. Tread lightly!!! 2804 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2805 * the caller. Calling this code willy nilly can result in the loss of data. 2806 * 2807 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2808 * B_CACHE for the non-VMIO case. 2809 */ 2810 2811int 2812allocbuf(struct buf *bp, int size) 2813{ 2814 int newbsize, mbsize; 2815 int i; 2816 2817 BUF_ASSERT_HELD(bp); 2818 2819 if (bp->b_kvasize < size) 2820 panic("allocbuf: buffer too small"); 2821 2822 if ((bp->b_flags & B_VMIO) == 0) { 2823 caddr_t origbuf; 2824 int origbufsize; 2825 /* 2826 * Just get anonymous memory from the kernel. Don't 2827 * mess with B_CACHE. 2828 */ 2829 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2830 if (bp->b_flags & B_MALLOC) 2831 newbsize = mbsize; 2832 else 2833 newbsize = round_page(size); 2834 2835 if (newbsize < bp->b_bufsize) { 2836 /* 2837 * malloced buffers are not shrunk 2838 */ 2839 if (bp->b_flags & B_MALLOC) { 2840 if (newbsize) { 2841 bp->b_bcount = size; 2842 } else { 2843 free(bp->b_data, M_BIOBUF); 2844 if (bp->b_bufsize) { 2845 atomic_subtract_long( 2846 &bufmallocspace, 2847 bp->b_bufsize); 2848 bufspacewakeup(); 2849 bp->b_bufsize = 0; 2850 } 2851 bp->b_saveaddr = bp->b_kvabase; 2852 bp->b_data = bp->b_saveaddr; 2853 bp->b_bcount = 0; 2854 bp->b_flags &= ~B_MALLOC; 2855 } 2856 return 1; 2857 } 2858 vm_hold_free_pages( 2859 bp, 2860 (vm_offset_t) bp->b_data + newbsize, 2861 (vm_offset_t) bp->b_data + bp->b_bufsize); 2862 } else if (newbsize > bp->b_bufsize) { 2863 /* 2864 * We only use malloced memory on the first allocation. 2865 * and revert to page-allocated memory when the buffer 2866 * grows. 2867 */ 2868 /* 2869 * There is a potential smp race here that could lead 2870 * to bufmallocspace slightly passing the max. It 2871 * is probably extremely rare and not worth worrying 2872 * over. 2873 */ 2874 if ( (bufmallocspace < maxbufmallocspace) && 2875 (bp->b_bufsize == 0) && 2876 (mbsize <= PAGE_SIZE/2)) { 2877 2878 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2879 bp->b_bufsize = mbsize; 2880 bp->b_bcount = size; 2881 bp->b_flags |= B_MALLOC; 2882 atomic_add_long(&bufmallocspace, mbsize); 2883 return 1; 2884 } 2885 origbuf = NULL; 2886 origbufsize = 0; 2887 /* 2888 * If the buffer is growing on its other-than-first allocation, 2889 * then we revert to the page-allocation scheme. 2890 */ 2891 if (bp->b_flags & B_MALLOC) { 2892 origbuf = bp->b_data; 2893 origbufsize = bp->b_bufsize; 2894 bp->b_data = bp->b_kvabase; 2895 if (bp->b_bufsize) { 2896 atomic_subtract_long(&bufmallocspace, 2897 bp->b_bufsize); 2898 bufspacewakeup(); 2899 bp->b_bufsize = 0; 2900 } 2901 bp->b_flags &= ~B_MALLOC; 2902 newbsize = round_page(newbsize); 2903 } 2904 vm_hold_load_pages( 2905 bp, 2906 (vm_offset_t) bp->b_data + bp->b_bufsize, 2907 (vm_offset_t) bp->b_data + newbsize); 2908 if (origbuf) { 2909 bcopy(origbuf, bp->b_data, origbufsize); 2910 free(origbuf, M_BIOBUF); 2911 } 2912 } 2913 } else { 2914 int desiredpages; 2915 2916 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2917 desiredpages = (size == 0) ? 0 : 2918 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2919 2920 if (bp->b_flags & B_MALLOC) 2921 panic("allocbuf: VMIO buffer can't be malloced"); 2922 /* 2923 * Set B_CACHE initially if buffer is 0 length or will become 2924 * 0-length. 2925 */ 2926 if (size == 0 || bp->b_bufsize == 0) 2927 bp->b_flags |= B_CACHE; 2928 2929 if (newbsize < bp->b_bufsize) { 2930 /* 2931 * DEV_BSIZE aligned new buffer size is less then the 2932 * DEV_BSIZE aligned existing buffer size. Figure out 2933 * if we have to remove any pages. 2934 */ 2935 if (desiredpages < bp->b_npages) { 2936 vm_page_t m; 2937 2938 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2939 for (i = desiredpages; i < bp->b_npages; i++) { 2940 /* 2941 * the page is not freed here -- it 2942 * is the responsibility of 2943 * vnode_pager_setsize 2944 */ 2945 m = bp->b_pages[i]; 2946 KASSERT(m != bogus_page, 2947 ("allocbuf: bogus page found")); 2948 while (vm_page_sleep_if_busy(m, TRUE, 2949 "biodep")) 2950 continue; 2951 2952 bp->b_pages[i] = NULL; 2953 vm_page_lock(m); 2954 vm_page_unwire(m, 0); 2955 vm_page_unlock(m); 2956 } 2957 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2958 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2959 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2960 bp->b_npages = desiredpages; 2961 } 2962 } else if (size > bp->b_bcount) { 2963 /* 2964 * We are growing the buffer, possibly in a 2965 * byte-granular fashion. 2966 */ 2967 vm_object_t obj; 2968 vm_offset_t toff; 2969 vm_offset_t tinc; 2970 2971 /* 2972 * Step 1, bring in the VM pages from the object, 2973 * allocating them if necessary. We must clear 2974 * B_CACHE if these pages are not valid for the 2975 * range covered by the buffer. 2976 */ 2977 2978 obj = bp->b_bufobj->bo_object; 2979 2980 VM_OBJECT_LOCK(obj); 2981 while (bp->b_npages < desiredpages) { 2982 vm_page_t m; 2983 vm_pindex_t pi; 2984 2985 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2986 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2987 /* 2988 * note: must allocate system pages 2989 * since blocking here could intefere 2990 * with paging I/O, no matter which 2991 * process we are. 2992 */ 2993 m = vm_page_alloc(obj, pi, 2994 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2995 VM_ALLOC_WIRED); 2996 if (m == NULL) { 2997 atomic_add_int(&vm_pageout_deficit, 2998 desiredpages - bp->b_npages); 2999 VM_OBJECT_UNLOCK(obj); 3000 VM_WAIT; 3001 VM_OBJECT_LOCK(obj); 3002 } else { 3003 if (m->valid == 0) 3004 bp->b_flags &= ~B_CACHE; 3005 bp->b_pages[bp->b_npages] = m; 3006 ++bp->b_npages; 3007 } 3008 continue; 3009 } 3010 3011 /* 3012 * We found a page. If we have to sleep on it, 3013 * retry because it might have gotten freed out 3014 * from under us. 3015 * 3016 * We can only test VPO_BUSY here. Blocking on 3017 * m->busy might lead to a deadlock: 3018 * 3019 * vm_fault->getpages->cluster_read->allocbuf 3020 * 3021 */ 3022 if ((m->oflags & VPO_BUSY) != 0) { 3023 /* 3024 * Reference the page before unlocking 3025 * and sleeping so that the page daemon 3026 * is less likely to reclaim it. 3027 */ 3028 vm_page_lock_queues(); 3029 vm_page_flag_set(m, PG_REFERENCED); 3030 vm_page_sleep(m, "pgtblk"); 3031 continue; 3032 } 3033 3034 /* 3035 * We have a good page. 3036 */ 3037 vm_page_lock(m); 3038 vm_page_wire(m); 3039 vm_page_unlock(m); 3040 bp->b_pages[bp->b_npages] = m; 3041 ++bp->b_npages; 3042 } 3043 3044 /* 3045 * Step 2. We've loaded the pages into the buffer, 3046 * we have to figure out if we can still have B_CACHE 3047 * set. Note that B_CACHE is set according to the 3048 * byte-granular range ( bcount and size ), new the 3049 * aligned range ( newbsize ). 3050 * 3051 * The VM test is against m->valid, which is DEV_BSIZE 3052 * aligned. Needless to say, the validity of the data 3053 * needs to also be DEV_BSIZE aligned. Note that this 3054 * fails with NFS if the server or some other client 3055 * extends the file's EOF. If our buffer is resized, 3056 * B_CACHE may remain set! XXX 3057 */ 3058 3059 toff = bp->b_bcount; 3060 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3061 3062 while ((bp->b_flags & B_CACHE) && toff < size) { 3063 vm_pindex_t pi; 3064 3065 if (tinc > (size - toff)) 3066 tinc = size - toff; 3067 3068 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3069 PAGE_SHIFT; 3070 3071 vfs_buf_test_cache( 3072 bp, 3073 bp->b_offset, 3074 toff, 3075 tinc, 3076 bp->b_pages[pi] 3077 ); 3078 toff += tinc; 3079 tinc = PAGE_SIZE; 3080 } 3081 VM_OBJECT_UNLOCK(obj); 3082 3083 /* 3084 * Step 3, fixup the KVM pmap. Remember that 3085 * bp->b_data is relative to bp->b_offset, but 3086 * bp->b_offset may be offset into the first page. 3087 */ 3088 3089 bp->b_data = (caddr_t) 3090 trunc_page((vm_offset_t)bp->b_data); 3091 pmap_qenter( 3092 (vm_offset_t)bp->b_data, 3093 bp->b_pages, 3094 bp->b_npages 3095 ); 3096 3097 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3098 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3099 } 3100 } 3101 if (newbsize < bp->b_bufsize) 3102 bufspacewakeup(); 3103 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3104 bp->b_bcount = size; /* requested buffer size */ 3105 return 1; 3106} 3107 3108void 3109biodone(struct bio *bp) 3110{ 3111 struct mtx *mtxp; 3112 void (*done)(struct bio *); 3113 3114 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3115 mtx_lock(mtxp); 3116 bp->bio_flags |= BIO_DONE; 3117 done = bp->bio_done; 3118 if (done == NULL) 3119 wakeup(bp); 3120 mtx_unlock(mtxp); 3121 if (done != NULL) 3122 done(bp); 3123} 3124 3125/* 3126 * Wait for a BIO to finish. 3127 * 3128 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3129 * case is not yet clear. 3130 */ 3131int 3132biowait(struct bio *bp, const char *wchan) 3133{ 3134 struct mtx *mtxp; 3135 3136 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3137 mtx_lock(mtxp); 3138 while ((bp->bio_flags & BIO_DONE) == 0) 3139 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3140 mtx_unlock(mtxp); 3141 if (bp->bio_error != 0) 3142 return (bp->bio_error); 3143 if (!(bp->bio_flags & BIO_ERROR)) 3144 return (0); 3145 return (EIO); 3146} 3147 3148void 3149biofinish(struct bio *bp, struct devstat *stat, int error) 3150{ 3151 3152 if (error) { 3153 bp->bio_error = error; 3154 bp->bio_flags |= BIO_ERROR; 3155 } 3156 if (stat != NULL) 3157 devstat_end_transaction_bio(stat, bp); 3158 biodone(bp); 3159} 3160 3161/* 3162 * bufwait: 3163 * 3164 * Wait for buffer I/O completion, returning error status. The buffer 3165 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3166 * error and cleared. 3167 */ 3168int 3169bufwait(struct buf *bp) 3170{ 3171 if (bp->b_iocmd == BIO_READ) 3172 bwait(bp, PRIBIO, "biord"); 3173 else 3174 bwait(bp, PRIBIO, "biowr"); 3175 if (bp->b_flags & B_EINTR) { 3176 bp->b_flags &= ~B_EINTR; 3177 return (EINTR); 3178 } 3179 if (bp->b_ioflags & BIO_ERROR) { 3180 return (bp->b_error ? bp->b_error : EIO); 3181 } else { 3182 return (0); 3183 } 3184} 3185 3186 /* 3187 * Call back function from struct bio back up to struct buf. 3188 */ 3189static void 3190bufdonebio(struct bio *bip) 3191{ 3192 struct buf *bp; 3193 3194 bp = bip->bio_caller2; 3195 bp->b_resid = bp->b_bcount - bip->bio_completed; 3196 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3197 bp->b_ioflags = bip->bio_flags; 3198 bp->b_error = bip->bio_error; 3199 if (bp->b_error) 3200 bp->b_ioflags |= BIO_ERROR; 3201 bufdone(bp); 3202 g_destroy_bio(bip); 3203} 3204 3205void 3206dev_strategy(struct cdev *dev, struct buf *bp) 3207{ 3208 struct cdevsw *csw; 3209 struct bio *bip; 3210 3211 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3212 panic("b_iocmd botch"); 3213 for (;;) { 3214 bip = g_new_bio(); 3215 if (bip != NULL) 3216 break; 3217 /* Try again later */ 3218 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3219 } 3220 bip->bio_cmd = bp->b_iocmd; 3221 bip->bio_offset = bp->b_iooffset; 3222 bip->bio_length = bp->b_bcount; 3223 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3224 bip->bio_data = bp->b_data; 3225 bip->bio_done = bufdonebio; 3226 bip->bio_caller2 = bp; 3227 bip->bio_dev = dev; 3228 KASSERT(dev->si_refcount > 0, 3229 ("dev_strategy on un-referenced struct cdev *(%s)", 3230 devtoname(dev))); 3231 csw = dev_refthread(dev); 3232 if (csw == NULL) { 3233 g_destroy_bio(bip); 3234 bp->b_error = ENXIO; 3235 bp->b_ioflags = BIO_ERROR; 3236 bufdone(bp); 3237 return; 3238 } 3239 (*csw->d_strategy)(bip); 3240 dev_relthread(dev); 3241} 3242 3243/* 3244 * bufdone: 3245 * 3246 * Finish I/O on a buffer, optionally calling a completion function. 3247 * This is usually called from an interrupt so process blocking is 3248 * not allowed. 3249 * 3250 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3251 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3252 * assuming B_INVAL is clear. 3253 * 3254 * For the VMIO case, we set B_CACHE if the op was a read and no 3255 * read error occured, or if the op was a write. B_CACHE is never 3256 * set if the buffer is invalid or otherwise uncacheable. 3257 * 3258 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3259 * initiator to leave B_INVAL set to brelse the buffer out of existance 3260 * in the biodone routine. 3261 */ 3262void 3263bufdone(struct buf *bp) 3264{ 3265 struct bufobj *dropobj; 3266 void (*biodone)(struct buf *); 3267 3268 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3269 dropobj = NULL; 3270 3271 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3272 BUF_ASSERT_HELD(bp); 3273 3274 runningbufwakeup(bp); 3275 if (bp->b_iocmd == BIO_WRITE) 3276 dropobj = bp->b_bufobj; 3277 /* call optional completion function if requested */ 3278 if (bp->b_iodone != NULL) { 3279 biodone = bp->b_iodone; 3280 bp->b_iodone = NULL; 3281 (*biodone) (bp); 3282 if (dropobj) 3283 bufobj_wdrop(dropobj); 3284 return; 3285 } 3286 3287 bufdone_finish(bp); 3288 3289 if (dropobj) 3290 bufobj_wdrop(dropobj); 3291} 3292 3293void 3294bufdone_finish(struct buf *bp) 3295{ 3296 BUF_ASSERT_HELD(bp); 3297 3298 if (!LIST_EMPTY(&bp->b_dep)) 3299 buf_complete(bp); 3300 3301 if (bp->b_flags & B_VMIO) { 3302 int i; 3303 vm_ooffset_t foff; 3304 vm_page_t m; 3305 vm_object_t obj; 3306 int iosize; 3307 struct vnode *vp = bp->b_vp; 3308 3309 obj = bp->b_bufobj->bo_object; 3310 3311#if defined(VFS_BIO_DEBUG) 3312 mp_fixme("usecount and vflag accessed without locks."); 3313 if (vp->v_usecount == 0) { 3314 panic("biodone: zero vnode ref count"); 3315 } 3316 3317 KASSERT(vp->v_object != NULL, 3318 ("biodone: vnode %p has no vm_object", vp)); 3319#endif 3320 3321 foff = bp->b_offset; 3322 KASSERT(bp->b_offset != NOOFFSET, 3323 ("biodone: no buffer offset")); 3324 3325 VM_OBJECT_LOCK(obj); 3326#if defined(VFS_BIO_DEBUG) 3327 if (obj->paging_in_progress < bp->b_npages) { 3328 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3329 obj->paging_in_progress, bp->b_npages); 3330 } 3331#endif 3332 3333 /* 3334 * Set B_CACHE if the op was a normal read and no error 3335 * occured. B_CACHE is set for writes in the b*write() 3336 * routines. 3337 */ 3338 iosize = bp->b_bcount - bp->b_resid; 3339 if (bp->b_iocmd == BIO_READ && 3340 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3341 !(bp->b_ioflags & BIO_ERROR)) { 3342 bp->b_flags |= B_CACHE; 3343 } 3344 for (i = 0; i < bp->b_npages; i++) { 3345 int bogusflag = 0; 3346 int resid; 3347 3348 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3349 if (resid > iosize) 3350 resid = iosize; 3351 3352 /* 3353 * cleanup bogus pages, restoring the originals 3354 */ 3355 m = bp->b_pages[i]; 3356 if (m == bogus_page) { 3357 bogusflag = 1; 3358 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3359 if (m == NULL) 3360 panic("biodone: page disappeared!"); 3361 bp->b_pages[i] = m; 3362 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3363 bp->b_pages, bp->b_npages); 3364 } 3365#if defined(VFS_BIO_DEBUG) 3366 if (OFF_TO_IDX(foff) != m->pindex) { 3367 printf( 3368"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3369 (intmax_t)foff, (uintmax_t)m->pindex); 3370 } 3371#endif 3372 3373 /* 3374 * In the write case, the valid and clean bits are 3375 * already changed correctly ( see bdwrite() ), so we 3376 * only need to do this here in the read case. 3377 */ 3378 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3379 KASSERT((m->dirty & vm_page_bits(foff & 3380 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3381 " page %p has unexpected dirty bits", m)); 3382 vfs_page_set_valid(bp, foff, m); 3383 } 3384 3385 /* 3386 * when debugging new filesystems or buffer I/O methods, this 3387 * is the most common error that pops up. if you see this, you 3388 * have not set the page busy flag correctly!!! 3389 */ 3390 if (m->busy == 0) { 3391 printf("biodone: page busy < 0, " 3392 "pindex: %d, foff: 0x(%x,%x), " 3393 "resid: %d, index: %d\n", 3394 (int) m->pindex, (int)(foff >> 32), 3395 (int) foff & 0xffffffff, resid, i); 3396 if (!vn_isdisk(vp, NULL)) 3397 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3398 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3399 (intmax_t) bp->b_lblkno, 3400 bp->b_flags, bp->b_npages); 3401 else 3402 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3403 (intmax_t) bp->b_lblkno, 3404 bp->b_flags, bp->b_npages); 3405 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3406 (u_long)m->valid, (u_long)m->dirty, 3407 m->wire_count); 3408 panic("biodone: page busy < 0\n"); 3409 } 3410 vm_page_io_finish(m); 3411 vm_object_pip_subtract(obj, 1); 3412 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3413 iosize -= resid; 3414 } 3415 vm_object_pip_wakeupn(obj, 0); 3416 VM_OBJECT_UNLOCK(obj); 3417 } 3418 3419 /* 3420 * For asynchronous completions, release the buffer now. The brelse 3421 * will do a wakeup there if necessary - so no need to do a wakeup 3422 * here in the async case. The sync case always needs to do a wakeup. 3423 */ 3424 3425 if (bp->b_flags & B_ASYNC) { 3426 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3427 brelse(bp); 3428 else 3429 bqrelse(bp); 3430 } else 3431 bdone(bp); 3432} 3433 3434/* 3435 * This routine is called in lieu of iodone in the case of 3436 * incomplete I/O. This keeps the busy status for pages 3437 * consistant. 3438 */ 3439void 3440vfs_unbusy_pages(struct buf *bp) 3441{ 3442 int i; 3443 vm_object_t obj; 3444 vm_page_t m; 3445 3446 runningbufwakeup(bp); 3447 if (!(bp->b_flags & B_VMIO)) 3448 return; 3449 3450 obj = bp->b_bufobj->bo_object; 3451 VM_OBJECT_LOCK(obj); 3452 for (i = 0; i < bp->b_npages; i++) { 3453 m = bp->b_pages[i]; 3454 if (m == bogus_page) { 3455 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3456 if (!m) 3457 panic("vfs_unbusy_pages: page missing\n"); 3458 bp->b_pages[i] = m; 3459 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3460 bp->b_pages, bp->b_npages); 3461 } 3462 vm_object_pip_subtract(obj, 1); 3463 vm_page_io_finish(m); 3464 } 3465 vm_object_pip_wakeupn(obj, 0); 3466 VM_OBJECT_UNLOCK(obj); 3467} 3468 3469/* 3470 * vfs_page_set_valid: 3471 * 3472 * Set the valid bits in a page based on the supplied offset. The 3473 * range is restricted to the buffer's size. 3474 * 3475 * This routine is typically called after a read completes. 3476 */ 3477static void 3478vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3479{ 3480 vm_ooffset_t eoff; 3481 3482 /* 3483 * Compute the end offset, eoff, such that [off, eoff) does not span a 3484 * page boundary and eoff is not greater than the end of the buffer. 3485 * The end of the buffer, in this case, is our file EOF, not the 3486 * allocation size of the buffer. 3487 */ 3488 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3489 if (eoff > bp->b_offset + bp->b_bcount) 3490 eoff = bp->b_offset + bp->b_bcount; 3491 3492 /* 3493 * Set valid range. This is typically the entire buffer and thus the 3494 * entire page. 3495 */ 3496 if (eoff > off) 3497 vm_page_set_valid(m, off & PAGE_MASK, eoff - off); 3498} 3499 3500/* 3501 * vfs_page_set_validclean: 3502 * 3503 * Set the valid bits and clear the dirty bits in a page based on the 3504 * supplied offset. The range is restricted to the buffer's size. 3505 */ 3506static void 3507vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3508{ 3509 vm_ooffset_t soff, eoff; 3510 3511 /* 3512 * Start and end offsets in buffer. eoff - soff may not cross a 3513 * page boundry or cross the end of the buffer. The end of the 3514 * buffer, in this case, is our file EOF, not the allocation size 3515 * of the buffer. 3516 */ 3517 soff = off; 3518 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3519 if (eoff > bp->b_offset + bp->b_bcount) 3520 eoff = bp->b_offset + bp->b_bcount; 3521 3522 /* 3523 * Set valid range. This is typically the entire buffer and thus the 3524 * entire page. 3525 */ 3526 if (eoff > soff) { 3527 vm_page_set_validclean( 3528 m, 3529 (vm_offset_t) (soff & PAGE_MASK), 3530 (vm_offset_t) (eoff - soff) 3531 ); 3532 } 3533} 3534 3535/* 3536 * This routine is called before a device strategy routine. 3537 * It is used to tell the VM system that paging I/O is in 3538 * progress, and treat the pages associated with the buffer 3539 * almost as being VPO_BUSY. Also the object paging_in_progress 3540 * flag is handled to make sure that the object doesn't become 3541 * inconsistant. 3542 * 3543 * Since I/O has not been initiated yet, certain buffer flags 3544 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3545 * and should be ignored. 3546 */ 3547void 3548vfs_busy_pages(struct buf *bp, int clear_modify) 3549{ 3550 int i, bogus; 3551 vm_object_t obj; 3552 vm_ooffset_t foff; 3553 vm_page_t m; 3554 3555 if (!(bp->b_flags & B_VMIO)) 3556 return; 3557 3558 obj = bp->b_bufobj->bo_object; 3559 foff = bp->b_offset; 3560 KASSERT(bp->b_offset != NOOFFSET, 3561 ("vfs_busy_pages: no buffer offset")); 3562 VM_OBJECT_LOCK(obj); 3563 if (bp->b_bufsize != 0) 3564 vfs_setdirty_locked_object(bp); 3565retry: 3566 for (i = 0; i < bp->b_npages; i++) { 3567 m = bp->b_pages[i]; 3568 3569 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3570 goto retry; 3571 } 3572 bogus = 0; 3573 for (i = 0; i < bp->b_npages; i++) { 3574 m = bp->b_pages[i]; 3575 3576 if ((bp->b_flags & B_CLUSTER) == 0) { 3577 vm_object_pip_add(obj, 1); 3578 vm_page_io_start(m); 3579 } 3580 /* 3581 * When readying a buffer for a read ( i.e 3582 * clear_modify == 0 ), it is important to do 3583 * bogus_page replacement for valid pages in 3584 * partially instantiated buffers. Partially 3585 * instantiated buffers can, in turn, occur when 3586 * reconstituting a buffer from its VM backing store 3587 * base. We only have to do this if B_CACHE is 3588 * clear ( which causes the I/O to occur in the 3589 * first place ). The replacement prevents the read 3590 * I/O from overwriting potentially dirty VM-backed 3591 * pages. XXX bogus page replacement is, uh, bogus. 3592 * It may not work properly with small-block devices. 3593 * We need to find a better way. 3594 */ 3595 if (clear_modify) { 3596 pmap_remove_write(m); 3597 vfs_page_set_validclean(bp, foff, m); 3598 } else if (m->valid == VM_PAGE_BITS_ALL && 3599 (bp->b_flags & B_CACHE) == 0) { 3600 bp->b_pages[i] = bogus_page; 3601 bogus++; 3602 } 3603 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3604 } 3605 VM_OBJECT_UNLOCK(obj); 3606 if (bogus) 3607 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3608 bp->b_pages, bp->b_npages); 3609} 3610 3611/* 3612 * Tell the VM system that the pages associated with this buffer 3613 * are clean. This is used for delayed writes where the data is 3614 * going to go to disk eventually without additional VM intevention. 3615 * 3616 * Note that while we only really need to clean through to b_bcount, we 3617 * just go ahead and clean through to b_bufsize. 3618 */ 3619static void 3620vfs_clean_pages(struct buf *bp) 3621{ 3622 int i; 3623 vm_ooffset_t foff, noff, eoff; 3624 vm_page_t m; 3625 3626 if (!(bp->b_flags & B_VMIO)) 3627 return; 3628 3629 foff = bp->b_offset; 3630 KASSERT(bp->b_offset != NOOFFSET, 3631 ("vfs_clean_pages: no buffer offset")); 3632 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3633 for (i = 0; i < bp->b_npages; i++) { 3634 m = bp->b_pages[i]; 3635 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3636 eoff = noff; 3637 3638 if (eoff > bp->b_offset + bp->b_bufsize) 3639 eoff = bp->b_offset + bp->b_bufsize; 3640 vfs_page_set_validclean(bp, foff, m); 3641 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3642 foff = noff; 3643 } 3644 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3645} 3646 3647/* 3648 * vfs_bio_set_valid: 3649 * 3650 * Set the range within the buffer to valid. The range is 3651 * relative to the beginning of the buffer, b_offset. Note that 3652 * b_offset itself may be offset from the beginning of the first 3653 * page. 3654 */ 3655void 3656vfs_bio_set_valid(struct buf *bp, int base, int size) 3657{ 3658 int i, n; 3659 vm_page_t m; 3660 3661 if (!(bp->b_flags & B_VMIO)) 3662 return; 3663 3664 /* 3665 * Fixup base to be relative to beginning of first page. 3666 * Set initial n to be the maximum number of bytes in the 3667 * first page that can be validated. 3668 */ 3669 base += (bp->b_offset & PAGE_MASK); 3670 n = PAGE_SIZE - (base & PAGE_MASK); 3671 3672 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3673 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3674 m = bp->b_pages[i]; 3675 if (n > size) 3676 n = size; 3677 vm_page_set_valid(m, base & PAGE_MASK, n); 3678 base += n; 3679 size -= n; 3680 n = PAGE_SIZE; 3681 } 3682 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3683} 3684 3685/* 3686 * vfs_bio_clrbuf: 3687 * 3688 * If the specified buffer is a non-VMIO buffer, clear the entire 3689 * buffer. If the specified buffer is a VMIO buffer, clear and 3690 * validate only the previously invalid portions of the buffer. 3691 * This routine essentially fakes an I/O, so we need to clear 3692 * BIO_ERROR and B_INVAL. 3693 * 3694 * Note that while we only theoretically need to clear through b_bcount, 3695 * we go ahead and clear through b_bufsize. 3696 */ 3697void 3698vfs_bio_clrbuf(struct buf *bp) 3699{ 3700 int i, j, mask; 3701 caddr_t sa, ea; 3702 3703 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3704 clrbuf(bp); 3705 return; 3706 } 3707 bp->b_flags &= ~B_INVAL; 3708 bp->b_ioflags &= ~BIO_ERROR; 3709 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3710 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3711 (bp->b_offset & PAGE_MASK) == 0) { 3712 if (bp->b_pages[0] == bogus_page) 3713 goto unlock; 3714 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3715 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3716 if ((bp->b_pages[0]->valid & mask) == mask) 3717 goto unlock; 3718 if ((bp->b_pages[0]->valid & mask) == 0) { 3719 bzero(bp->b_data, bp->b_bufsize); 3720 bp->b_pages[0]->valid |= mask; 3721 goto unlock; 3722 } 3723 } 3724 ea = sa = bp->b_data; 3725 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3726 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3727 ea = (caddr_t)(vm_offset_t)ulmin( 3728 (u_long)(vm_offset_t)ea, 3729 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3730 if (bp->b_pages[i] == bogus_page) 3731 continue; 3732 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3733 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3734 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3735 if ((bp->b_pages[i]->valid & mask) == mask) 3736 continue; 3737 if ((bp->b_pages[i]->valid & mask) == 0) 3738 bzero(sa, ea - sa); 3739 else { 3740 for (; sa < ea; sa += DEV_BSIZE, j++) { 3741 if ((bp->b_pages[i]->valid & (1 << j)) == 0) 3742 bzero(sa, DEV_BSIZE); 3743 } 3744 } 3745 bp->b_pages[i]->valid |= mask; 3746 } 3747unlock: 3748 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3749 bp->b_resid = 0; 3750} 3751 3752/* 3753 * vm_hold_load_pages and vm_hold_free_pages get pages into 3754 * a buffers address space. The pages are anonymous and are 3755 * not associated with a file object. 3756 */ 3757static void 3758vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3759{ 3760 vm_offset_t pg; 3761 vm_page_t p; 3762 int index; 3763 3764 to = round_page(to); 3765 from = round_page(from); 3766 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3767 3768 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3769tryagain: 3770 /* 3771 * note: must allocate system pages since blocking here 3772 * could interfere with paging I/O, no matter which 3773 * process we are. 3774 */ 3775 p = vm_page_alloc(NULL, pg >> PAGE_SHIFT, VM_ALLOC_NOOBJ | 3776 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3777 if (!p) { 3778 atomic_add_int(&vm_pageout_deficit, 3779 (to - pg) >> PAGE_SHIFT); 3780 VM_WAIT; 3781 goto tryagain; 3782 } 3783 pmap_qenter(pg, &p, 1); 3784 bp->b_pages[index] = p; 3785 } 3786 bp->b_npages = index; 3787} 3788 3789/* Return pages associated with this buf to the vm system */ 3790static void 3791vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3792{ 3793 vm_offset_t pg; 3794 vm_page_t p; 3795 int index, newnpages; 3796 3797 from = round_page(from); 3798 to = round_page(to); 3799 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3800 3801 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3802 p = bp->b_pages[index]; 3803 if (p && (index < bp->b_npages)) { 3804 if (p->busy) { 3805 printf( 3806 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3807 (intmax_t)bp->b_blkno, 3808 (intmax_t)bp->b_lblkno); 3809 } 3810 bp->b_pages[index] = NULL; 3811 pmap_qremove(pg, 1); 3812 p->wire_count--; 3813 vm_page_free(p); 3814 atomic_subtract_int(&cnt.v_wire_count, 1); 3815 } 3816 } 3817 bp->b_npages = newnpages; 3818} 3819 3820/* 3821 * Map an IO request into kernel virtual address space. 3822 * 3823 * All requests are (re)mapped into kernel VA space. 3824 * Notice that we use b_bufsize for the size of the buffer 3825 * to be mapped. b_bcount might be modified by the driver. 3826 * 3827 * Note that even if the caller determines that the address space should 3828 * be valid, a race or a smaller-file mapped into a larger space may 3829 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3830 * check the return value. 3831 */ 3832int 3833vmapbuf(struct buf *bp) 3834{ 3835 caddr_t addr, kva; 3836 vm_prot_t prot; 3837 int pidx, i; 3838 struct vm_page *m; 3839 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3840 3841 if (bp->b_bufsize < 0) 3842 return (-1); 3843 prot = VM_PROT_READ; 3844 if (bp->b_iocmd == BIO_READ) 3845 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3846 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3847 addr < bp->b_data + bp->b_bufsize; 3848 addr += PAGE_SIZE, pidx++) { 3849 /* 3850 * Do the vm_fault if needed; do the copy-on-write thing 3851 * when reading stuff off device into memory. 3852 * 3853 * NOTE! Must use pmap_extract() because addr may be in 3854 * the userland address space, and kextract is only guarenteed 3855 * to work for the kernland address space (see: sparc64 port). 3856 */ 3857retry: 3858 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3859 prot) < 0) { 3860 for (i = 0; i < pidx; ++i) { 3861 vm_page_lock(bp->b_pages[i]); 3862 vm_page_unhold(bp->b_pages[i]); 3863 vm_page_unlock(bp->b_pages[i]); 3864 bp->b_pages[i] = NULL; 3865 } 3866 return(-1); 3867 } 3868 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3869 if (m == NULL) 3870 goto retry; 3871 bp->b_pages[pidx] = m; 3872 } 3873 if (pidx > btoc(MAXPHYS)) 3874 panic("vmapbuf: mapped more than MAXPHYS"); 3875 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3876 3877 kva = bp->b_saveaddr; 3878 bp->b_npages = pidx; 3879 bp->b_saveaddr = bp->b_data; 3880 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3881 return(0); 3882} 3883 3884/* 3885 * Free the io map PTEs associated with this IO operation. 3886 * We also invalidate the TLB entries and restore the original b_addr. 3887 */ 3888void 3889vunmapbuf(struct buf *bp) 3890{ 3891 int pidx; 3892 int npages; 3893 3894 npages = bp->b_npages; 3895 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3896 for (pidx = 0; pidx < npages; pidx++) { 3897 vm_page_lock(bp->b_pages[pidx]); 3898 vm_page_unhold(bp->b_pages[pidx]); 3899 vm_page_unlock(bp->b_pages[pidx]); 3900 } 3901 3902 bp->b_data = bp->b_saveaddr; 3903} 3904 3905void 3906bdone(struct buf *bp) 3907{ 3908 struct mtx *mtxp; 3909 3910 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3911 mtx_lock(mtxp); 3912 bp->b_flags |= B_DONE; 3913 wakeup(bp); 3914 mtx_unlock(mtxp); 3915} 3916 3917void 3918bwait(struct buf *bp, u_char pri, const char *wchan) 3919{ 3920 struct mtx *mtxp; 3921 3922 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3923 mtx_lock(mtxp); 3924 while ((bp->b_flags & B_DONE) == 0) 3925 msleep(bp, mtxp, pri, wchan, 0); 3926 mtx_unlock(mtxp); 3927} 3928 3929int 3930bufsync(struct bufobj *bo, int waitfor) 3931{ 3932 3933 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 3934} 3935 3936void 3937bufstrategy(struct bufobj *bo, struct buf *bp) 3938{ 3939 int i = 0; 3940 struct vnode *vp; 3941 3942 vp = bp->b_vp; 3943 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3944 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3945 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3946 i = VOP_STRATEGY(vp, bp); 3947 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3948} 3949 3950void 3951bufobj_wrefl(struct bufobj *bo) 3952{ 3953 3954 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3955 ASSERT_BO_LOCKED(bo); 3956 bo->bo_numoutput++; 3957} 3958 3959void 3960bufobj_wref(struct bufobj *bo) 3961{ 3962 3963 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3964 BO_LOCK(bo); 3965 bo->bo_numoutput++; 3966 BO_UNLOCK(bo); 3967} 3968 3969void 3970bufobj_wdrop(struct bufobj *bo) 3971{ 3972 3973 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3974 BO_LOCK(bo); 3975 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3976 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3977 bo->bo_flag &= ~BO_WWAIT; 3978 wakeup(&bo->bo_numoutput); 3979 } 3980 BO_UNLOCK(bo); 3981} 3982 3983int 3984bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3985{ 3986 int error; 3987 3988 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3989 ASSERT_BO_LOCKED(bo); 3990 error = 0; 3991 while (bo->bo_numoutput) { 3992 bo->bo_flag |= BO_WWAIT; 3993 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3994 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3995 if (error) 3996 break; 3997 } 3998 return (error); 3999} 4000 4001void 4002bpin(struct buf *bp) 4003{ 4004 struct mtx *mtxp; 4005 4006 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4007 mtx_lock(mtxp); 4008 bp->b_pin_count++; 4009 mtx_unlock(mtxp); 4010} 4011 4012void 4013bunpin(struct buf *bp) 4014{ 4015 struct mtx *mtxp; 4016 4017 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4018 mtx_lock(mtxp); 4019 if (--bp->b_pin_count == 0) 4020 wakeup(bp); 4021 mtx_unlock(mtxp); 4022} 4023 4024void 4025bunpin_wait(struct buf *bp) 4026{ 4027 struct mtx *mtxp; 4028 4029 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4030 mtx_lock(mtxp); 4031 while (bp->b_pin_count > 0) 4032 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 4033 mtx_unlock(mtxp); 4034} 4035 4036#include "opt_ddb.h" 4037#ifdef DDB 4038#include <ddb/ddb.h> 4039 4040/* DDB command to show buffer data */ 4041DB_SHOW_COMMAND(buffer, db_show_buffer) 4042{ 4043 /* get args */ 4044 struct buf *bp = (struct buf *)addr; 4045 4046 if (!have_addr) { 4047 db_printf("usage: show buffer <addr>\n"); 4048 return; 4049 } 4050 4051 db_printf("buf at %p\n", bp); 4052 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4053 db_printf( 4054 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4055 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n", 4056 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4057 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4058 bp->b_dep.lh_first); 4059 if (bp->b_npages) { 4060 int i; 4061 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4062 for (i = 0; i < bp->b_npages; i++) { 4063 vm_page_t m; 4064 m = bp->b_pages[i]; 4065 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4066 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4067 if ((i + 1) < bp->b_npages) 4068 db_printf(","); 4069 } 4070 db_printf("\n"); 4071 } 4072 db_printf(" "); 4073 lockmgr_printinfo(&bp->b_lock); 4074} 4075 4076DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4077{ 4078 struct buf *bp; 4079 int i; 4080 4081 for (i = 0; i < nbuf; i++) { 4082 bp = &buf[i]; 4083 if (BUF_ISLOCKED(bp)) { 4084 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4085 db_printf("\n"); 4086 } 4087 } 4088} 4089 4090DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4091{ 4092 struct vnode *vp; 4093 struct buf *bp; 4094 4095 if (!have_addr) { 4096 db_printf("usage: show vnodebufs <addr>\n"); 4097 return; 4098 } 4099 vp = (struct vnode *)addr; 4100 db_printf("Clean buffers:\n"); 4101 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4102 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4103 db_printf("\n"); 4104 } 4105 db_printf("Dirty buffers:\n"); 4106 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4107 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4108 db_printf("\n"); 4109 } 4110} 4111#endif /* DDB */ 4112