vfs_bio.c revision 119599
1/* 2 * Copyright (c) 1994,1997 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Absolutely no warranty of function or purpose is made by the author 12 * John S. Dyson. 13 */ 14 15/* 16 * this file contains a new buffer I/O scheme implementing a coherent 17 * VM object and buffer cache scheme. Pains have been taken to make 18 * sure that the performance degradation associated with schemes such 19 * as this is not realized. 20 * 21 * Author: John S. Dyson 22 * Significant help during the development and debugging phases 23 * had been provided by David Greenman, also of the FreeBSD core team. 24 * 25 * see man buf(9) for more info. 26 */ 27 28#include <sys/cdefs.h> 29__FBSDID("$FreeBSD: head/sys/kern/vfs_bio.c 119599 2003-08-31 01:07:45Z jeff $"); 30 31#include <sys/param.h> 32#include <sys/systm.h> 33#include <sys/bio.h> 34#include <sys/buf.h> 35#include <sys/devicestat.h> 36#include <sys/eventhandler.h> 37#include <sys/lock.h> 38#include <sys/malloc.h> 39#include <sys/mount.h> 40#include <sys/mutex.h> 41#include <sys/kernel.h> 42#include <sys/kthread.h> 43#include <sys/proc.h> 44#include <sys/resourcevar.h> 45#include <sys/sysctl.h> 46#include <sys/vmmeter.h> 47#include <sys/vnode.h> 48#include <vm/vm.h> 49#include <vm/vm_param.h> 50#include <vm/vm_kern.h> 51#include <vm/vm_pageout.h> 52#include <vm/vm_page.h> 53#include <vm/vm_object.h> 54#include <vm/vm_extern.h> 55#include <vm/vm_map.h> 56#include "opt_directio.h" 57#include "opt_swap.h" 58 59static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 60 61struct bio_ops bioops; /* I/O operation notification */ 62 63struct buf_ops buf_ops_bio = { 64 "buf_ops_bio", 65 bwrite 66}; 67 68/* 69 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 70 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 71 */ 72struct buf *buf; /* buffer header pool */ 73 74static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 75 vm_offset_t to); 76static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 77 vm_offset_t to); 78static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 79 int pageno, vm_page_t m); 80static void vfs_clean_pages(struct buf * bp); 81static void vfs_setdirty(struct buf *bp); 82static void vfs_vmio_release(struct buf *bp); 83static void vfs_backgroundwritedone(struct buf *bp); 84static int vfs_bio_clcheck(struct vnode *vp, int size, 85 daddr_t lblkno, daddr_t blkno); 86static int flushbufqueues(int flushdeps); 87static void buf_daemon(void); 88void bremfreel(struct buf * bp); 89 90int vmiodirenable = TRUE; 91SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 92 "Use the VM system for directory writes"); 93int runningbufspace; 94SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 95 "Amount of presently outstanding async buffer io"); 96static int bufspace; 97SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 98 "KVA memory used for bufs"); 99static int maxbufspace; 100SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 101 "Maximum allowed value of bufspace (including buf_daemon)"); 102static int bufmallocspace; 103SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 104 "Amount of malloced memory for buffers"); 105static int maxbufmallocspace; 106SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 107 "Maximum amount of malloced memory for buffers"); 108static int lobufspace; 109SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 110 "Minimum amount of buffers we want to have"); 111static int hibufspace; 112SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 113 "Maximum allowed value of bufspace (excluding buf_daemon)"); 114static int bufreusecnt; 115SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 116 "Number of times we have reused a buffer"); 117static int buffreekvacnt; 118SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 119 "Number of times we have freed the KVA space from some buffer"); 120static int bufdefragcnt; 121SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 122 "Number of times we have had to repeat buffer allocation to defragment"); 123static int lorunningspace; 124SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 125 "Minimum preferred space used for in-progress I/O"); 126static int hirunningspace; 127SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 128 "Maximum amount of space to use for in-progress I/O"); 129static int dirtybufferflushes; 130SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 131 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 132static int altbufferflushes; 133SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 134 0, "Number of fsync flushes to limit dirty buffers"); 135static int recursiveflushes; 136SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 137 0, "Number of flushes skipped due to being recursive"); 138static int numdirtybuffers; 139SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 140 "Number of buffers that are dirty (has unwritten changes) at the moment"); 141static int lodirtybuffers; 142SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 143 "How many buffers we want to have free before bufdaemon can sleep"); 144static int hidirtybuffers; 145SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 146 "When the number of dirty buffers is considered severe"); 147static int dirtybufthresh; 148SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 149 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 150static int numfreebuffers; 151SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 152 "Number of free buffers"); 153static int lofreebuffers; 154SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 155 "XXX Unused"); 156static int hifreebuffers; 157SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 158 "XXX Complicatedly unused"); 159static int getnewbufcalls; 160SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 161 "Number of calls to getnewbuf"); 162static int getnewbufrestarts; 163SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 164 "Number of times getnewbuf has had to restart a buffer aquisition"); 165static int dobkgrdwrite = 1; 166SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 167 "Do background writes (honoring the BV_BKGRDWRITE flag)?"); 168 169/* 170 * Wakeup point for bufdaemon, as well as indicator of whether it is already 171 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 172 * is idling. 173 */ 174static int bd_request; 175 176/* 177 * This lock synchronizes access to bd_request. 178 */ 179static struct mtx bdlock; 180 181/* 182 * bogus page -- for I/O to/from partially complete buffers 183 * this is a temporary solution to the problem, but it is not 184 * really that bad. it would be better to split the buffer 185 * for input in the case of buffers partially already in memory, 186 * but the code is intricate enough already. 187 */ 188vm_page_t bogus_page; 189 190/* 191 * Synchronization (sleep/wakeup) variable for active buffer space requests. 192 * Set when wait starts, cleared prior to wakeup(). 193 * Used in runningbufwakeup() and waitrunningbufspace(). 194 */ 195static int runningbufreq; 196 197/* 198 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 199 * waitrunningbufspace(). 200 */ 201static struct mtx rbreqlock; 202 203/* 204 * Synchronization (sleep/wakeup) variable for buffer requests. 205 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 206 * by and/or. 207 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 208 * getnewbuf(), and getblk(). 209 */ 210static int needsbuffer; 211 212/* 213 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 214 */ 215static struct mtx nblock; 216 217/* 218 * Lock that protects against bwait()/bdone()/B_DONE races. 219 */ 220 221static struct mtx bdonelock; 222 223/* 224 * Definitions for the buffer free lists. 225 */ 226#define BUFFER_QUEUES 5 /* number of free buffer queues */ 227 228#define QUEUE_NONE 0 /* on no queue */ 229#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 230#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 231#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 232#define QUEUE_EMPTY 4 /* empty buffer headers */ 233 234/* Queues for free buffers with various properties */ 235static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 236 237/* Lock for the bufqueues */ 238static struct mtx bqlock; 239 240/* 241 * Single global constant for BUF_WMESG, to avoid getting multiple references. 242 * buf_wmesg is referred from macros. 243 */ 244const char *buf_wmesg = BUF_WMESG; 245 246#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 247#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 248#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 249#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 250 251#ifdef DIRECTIO 252extern void ffs_rawread_setup(void); 253#endif /* DIRECTIO */ 254/* 255 * numdirtywakeup: 256 * 257 * If someone is blocked due to there being too many dirty buffers, 258 * and numdirtybuffers is now reasonable, wake them up. 259 */ 260 261static __inline void 262numdirtywakeup(int level) 263{ 264 if (numdirtybuffers <= level) { 265 mtx_lock(&nblock); 266 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 267 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 268 wakeup(&needsbuffer); 269 } 270 mtx_unlock(&nblock); 271 } 272} 273 274/* 275 * bufspacewakeup: 276 * 277 * Called when buffer space is potentially available for recovery. 278 * getnewbuf() will block on this flag when it is unable to free 279 * sufficient buffer space. Buffer space becomes recoverable when 280 * bp's get placed back in the queues. 281 */ 282 283static __inline void 284bufspacewakeup(void) 285{ 286 /* 287 * If someone is waiting for BUF space, wake them up. Even 288 * though we haven't freed the kva space yet, the waiting 289 * process will be able to now. 290 */ 291 mtx_lock(&nblock); 292 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 293 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 294 wakeup(&needsbuffer); 295 } 296 mtx_unlock(&nblock); 297} 298 299/* 300 * runningbufwakeup() - in-progress I/O accounting. 301 * 302 */ 303static __inline void 304runningbufwakeup(struct buf *bp) 305{ 306 if (bp->b_runningbufspace) { 307 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 308 bp->b_runningbufspace = 0; 309 mtx_lock(&rbreqlock); 310 if (runningbufreq && runningbufspace <= lorunningspace) { 311 runningbufreq = 0; 312 wakeup(&runningbufreq); 313 } 314 mtx_unlock(&rbreqlock); 315 } 316} 317 318/* 319 * bufcountwakeup: 320 * 321 * Called when a buffer has been added to one of the free queues to 322 * account for the buffer and to wakeup anyone waiting for free buffers. 323 * This typically occurs when large amounts of metadata are being handled 324 * by the buffer cache ( else buffer space runs out first, usually ). 325 */ 326 327static __inline void 328bufcountwakeup(void) 329{ 330 atomic_add_int(&numfreebuffers, 1); 331 mtx_lock(&nblock); 332 if (needsbuffer) { 333 needsbuffer &= ~VFS_BIO_NEED_ANY; 334 if (numfreebuffers >= hifreebuffers) 335 needsbuffer &= ~VFS_BIO_NEED_FREE; 336 wakeup(&needsbuffer); 337 } 338 mtx_unlock(&nblock); 339} 340 341/* 342 * waitrunningbufspace() 343 * 344 * runningbufspace is a measure of the amount of I/O currently 345 * running. This routine is used in async-write situations to 346 * prevent creating huge backups of pending writes to a device. 347 * Only asynchronous writes are governed by this function. 348 * 349 * Reads will adjust runningbufspace, but will not block based on it. 350 * The read load has a side effect of reducing the allowed write load. 351 * 352 * This does NOT turn an async write into a sync write. It waits 353 * for earlier writes to complete and generally returns before the 354 * caller's write has reached the device. 355 */ 356static __inline void 357waitrunningbufspace(void) 358{ 359 mtx_lock(&rbreqlock); 360 while (runningbufspace > hirunningspace) { 361 ++runningbufreq; 362 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 363 } 364 mtx_unlock(&rbreqlock); 365} 366 367 368/* 369 * vfs_buf_test_cache: 370 * 371 * Called when a buffer is extended. This function clears the B_CACHE 372 * bit if the newly extended portion of the buffer does not contain 373 * valid data. 374 */ 375static __inline__ 376void 377vfs_buf_test_cache(struct buf *bp, 378 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 379 vm_page_t m) 380{ 381 GIANT_REQUIRED; 382 383 if (bp->b_flags & B_CACHE) { 384 int base = (foff + off) & PAGE_MASK; 385 if (vm_page_is_valid(m, base, size) == 0) 386 bp->b_flags &= ~B_CACHE; 387 } 388} 389 390/* Wake up the buffer deamon if necessary */ 391static __inline__ 392void 393bd_wakeup(int dirtybuflevel) 394{ 395 mtx_lock(&bdlock); 396 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 397 bd_request = 1; 398 wakeup(&bd_request); 399 } 400 mtx_unlock(&bdlock); 401} 402 403/* 404 * bd_speedup - speedup the buffer cache flushing code 405 */ 406 407static __inline__ 408void 409bd_speedup(void) 410{ 411 bd_wakeup(1); 412} 413 414/* 415 * Calculating buffer cache scaling values and reserve space for buffer 416 * headers. This is called during low level kernel initialization and 417 * may be called more then once. We CANNOT write to the memory area 418 * being reserved at this time. 419 */ 420caddr_t 421kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 422{ 423 /* 424 * physmem_est is in pages. Convert it to kilobytes (assumes 425 * PAGE_SIZE is >= 1K) 426 */ 427 physmem_est = physmem_est * (PAGE_SIZE / 1024); 428 429 /* 430 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 431 * For the first 64MB of ram nominally allocate sufficient buffers to 432 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 433 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 434 * the buffer cache we limit the eventual kva reservation to 435 * maxbcache bytes. 436 * 437 * factor represents the 1/4 x ram conversion. 438 */ 439 if (nbuf == 0) { 440 int factor = 4 * BKVASIZE / 1024; 441 442 nbuf = 50; 443 if (physmem_est > 4096) 444 nbuf += min((physmem_est - 4096) / factor, 445 65536 / factor); 446 if (physmem_est > 65536) 447 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 448 449 if (maxbcache && nbuf > maxbcache / BKVASIZE) 450 nbuf = maxbcache / BKVASIZE; 451 } 452 453#if 0 454 /* 455 * Do not allow the buffer_map to be more then 1/2 the size of the 456 * kernel_map. 457 */ 458 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 459 (BKVASIZE * 2)) { 460 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 461 (BKVASIZE * 2); 462 printf("Warning: nbufs capped at %d\n", nbuf); 463 } 464#endif 465 466 /* 467 * swbufs are used as temporary holders for I/O, such as paging I/O. 468 * We have no less then 16 and no more then 256. 469 */ 470 nswbuf = max(min(nbuf/4, 256), 16); 471#ifdef NSWBUF_MIN 472 if (nswbuf < NSWBUF_MIN) 473 nswbuf = NSWBUF_MIN; 474#endif 475#ifdef DIRECTIO 476 ffs_rawread_setup(); 477#endif 478 479 /* 480 * Reserve space for the buffer cache buffers 481 */ 482 swbuf = (void *)v; 483 v = (caddr_t)(swbuf + nswbuf); 484 buf = (void *)v; 485 v = (caddr_t)(buf + nbuf); 486 487 return(v); 488} 489 490/* Initialize the buffer subsystem. Called before use of any buffers. */ 491void 492bufinit(void) 493{ 494 struct buf *bp; 495 int i; 496 497 GIANT_REQUIRED; 498 499 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 500 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 501 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 502 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 503 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 504 505 /* next, make a null set of free lists */ 506 for (i = 0; i < BUFFER_QUEUES; i++) 507 TAILQ_INIT(&bufqueues[i]); 508 509 /* finally, initialize each buffer header and stick on empty q */ 510 for (i = 0; i < nbuf; i++) { 511 bp = &buf[i]; 512 bzero(bp, sizeof *bp); 513 bp->b_flags = B_INVAL; /* we're just an empty header */ 514 bp->b_dev = NODEV; 515 bp->b_rcred = NOCRED; 516 bp->b_wcred = NOCRED; 517 bp->b_qindex = QUEUE_EMPTY; 518 bp->b_vflags = 0; 519 bp->b_xflags = 0; 520 LIST_INIT(&bp->b_dep); 521 BUF_LOCKINIT(bp); 522 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 523 } 524 525 /* 526 * maxbufspace is the absolute maximum amount of buffer space we are 527 * allowed to reserve in KVM and in real terms. The absolute maximum 528 * is nominally used by buf_daemon. hibufspace is the nominal maximum 529 * used by most other processes. The differential is required to 530 * ensure that buf_daemon is able to run when other processes might 531 * be blocked waiting for buffer space. 532 * 533 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 534 * this may result in KVM fragmentation which is not handled optimally 535 * by the system. 536 */ 537 maxbufspace = nbuf * BKVASIZE; 538 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 539 lobufspace = hibufspace - MAXBSIZE; 540 541 lorunningspace = 512 * 1024; 542 hirunningspace = 1024 * 1024; 543 544/* 545 * Limit the amount of malloc memory since it is wired permanently into 546 * the kernel space. Even though this is accounted for in the buffer 547 * allocation, we don't want the malloced region to grow uncontrolled. 548 * The malloc scheme improves memory utilization significantly on average 549 * (small) directories. 550 */ 551 maxbufmallocspace = hibufspace / 20; 552 553/* 554 * Reduce the chance of a deadlock occuring by limiting the number 555 * of delayed-write dirty buffers we allow to stack up. 556 */ 557 hidirtybuffers = nbuf / 4 + 20; 558 dirtybufthresh = hidirtybuffers * 9 / 10; 559 numdirtybuffers = 0; 560/* 561 * To support extreme low-memory systems, make sure hidirtybuffers cannot 562 * eat up all available buffer space. This occurs when our minimum cannot 563 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 564 * BKVASIZE'd (8K) buffers. 565 */ 566 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 567 hidirtybuffers >>= 1; 568 } 569 lodirtybuffers = hidirtybuffers / 2; 570 571/* 572 * Try to keep the number of free buffers in the specified range, 573 * and give special processes (e.g. like buf_daemon) access to an 574 * emergency reserve. 575 */ 576 lofreebuffers = nbuf / 18 + 5; 577 hifreebuffers = 2 * lofreebuffers; 578 numfreebuffers = nbuf; 579 580/* 581 * Maximum number of async ops initiated per buf_daemon loop. This is 582 * somewhat of a hack at the moment, we really need to limit ourselves 583 * based on the number of bytes of I/O in-transit that were initiated 584 * from buf_daemon. 585 */ 586 587 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 588 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 589} 590 591/* 592 * bfreekva() - free the kva allocation for a buffer. 593 * 594 * Must be called at splbio() or higher as this is the only locking for 595 * buffer_map. 596 * 597 * Since this call frees up buffer space, we call bufspacewakeup(). 598 */ 599static void 600bfreekva(struct buf * bp) 601{ 602 GIANT_REQUIRED; 603 604 if (bp->b_kvasize) { 605 atomic_add_int(&buffreekvacnt, 1); 606 atomic_subtract_int(&bufspace, bp->b_kvasize); 607 vm_map_delete(buffer_map, 608 (vm_offset_t) bp->b_kvabase, 609 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 610 ); 611 bp->b_kvasize = 0; 612 bufspacewakeup(); 613 } 614} 615 616/* 617 * bremfree: 618 * 619 * Remove the buffer from the appropriate free list. 620 */ 621void 622bremfree(struct buf * bp) 623{ 624 mtx_lock(&bqlock); 625 bremfreel(bp); 626 mtx_unlock(&bqlock); 627} 628 629void 630bremfreel(struct buf * bp) 631{ 632 int s = splbio(); 633 int old_qindex = bp->b_qindex; 634 635 GIANT_REQUIRED; 636 637 if (bp->b_qindex != QUEUE_NONE) { 638 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 639 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 640 bp->b_qindex = QUEUE_NONE; 641 } else { 642 if (BUF_REFCNT(bp) <= 1) 643 panic("bremfree: removing a buffer not on a queue"); 644 } 645 646 /* 647 * Fixup numfreebuffers count. If the buffer is invalid or not 648 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 649 * the buffer was free and we must decrement numfreebuffers. 650 */ 651 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 652 switch(old_qindex) { 653 case QUEUE_DIRTY: 654 case QUEUE_CLEAN: 655 case QUEUE_EMPTY: 656 case QUEUE_EMPTYKVA: 657 atomic_subtract_int(&numfreebuffers, 1); 658 break; 659 default: 660 break; 661 } 662 } 663 splx(s); 664} 665 666 667/* 668 * Get a buffer with the specified data. Look in the cache first. We 669 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 670 * is set, the buffer is valid and we do not have to do anything ( see 671 * getblk() ). This is really just a special case of breadn(). 672 */ 673int 674bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 675 struct buf ** bpp) 676{ 677 678 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 679} 680 681/* 682 * Operates like bread, but also starts asynchronous I/O on 683 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 684 * to initiating I/O . If B_CACHE is set, the buffer is valid 685 * and we do not have to do anything. 686 */ 687int 688breadn(struct vnode * vp, daddr_t blkno, int size, 689 daddr_t * rablkno, int *rabsize, 690 int cnt, struct ucred * cred, struct buf ** bpp) 691{ 692 struct buf *bp, *rabp; 693 int i; 694 int rv = 0, readwait = 0; 695 696 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 697 698 /* if not found in cache, do some I/O */ 699 if ((bp->b_flags & B_CACHE) == 0) { 700 if (curthread != PCPU_GET(idlethread)) 701 curthread->td_proc->p_stats->p_ru.ru_inblock++; 702 bp->b_iocmd = BIO_READ; 703 bp->b_flags &= ~B_INVAL; 704 bp->b_ioflags &= ~BIO_ERROR; 705 if (bp->b_rcred == NOCRED && cred != NOCRED) 706 bp->b_rcred = crhold(cred); 707 vfs_busy_pages(bp, 0); 708 if (vp->v_type == VCHR) 709 VOP_SPECSTRATEGY(vp, bp); 710 else 711 VOP_STRATEGY(vp, bp); 712 ++readwait; 713 } 714 715 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 716 if (inmem(vp, *rablkno)) 717 continue; 718 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 719 720 if ((rabp->b_flags & B_CACHE) == 0) { 721 if (curthread != PCPU_GET(idlethread)) 722 curthread->td_proc->p_stats->p_ru.ru_inblock++; 723 rabp->b_flags |= B_ASYNC; 724 rabp->b_flags &= ~B_INVAL; 725 rabp->b_ioflags &= ~BIO_ERROR; 726 rabp->b_iocmd = BIO_READ; 727 if (rabp->b_rcred == NOCRED && cred != NOCRED) 728 rabp->b_rcred = crhold(cred); 729 vfs_busy_pages(rabp, 0); 730 BUF_KERNPROC(rabp); 731 if (vp->v_type == VCHR) 732 VOP_SPECSTRATEGY(vp, rabp); 733 else 734 VOP_STRATEGY(vp, rabp); 735 } else { 736 brelse(rabp); 737 } 738 } 739 740 if (readwait) { 741 rv = bufwait(bp); 742 } 743 return (rv); 744} 745 746/* 747 * Write, release buffer on completion. (Done by iodone 748 * if async). Do not bother writing anything if the buffer 749 * is invalid. 750 * 751 * Note that we set B_CACHE here, indicating that buffer is 752 * fully valid and thus cacheable. This is true even of NFS 753 * now so we set it generally. This could be set either here 754 * or in biodone() since the I/O is synchronous. We put it 755 * here. 756 */ 757 758int 759bwrite(struct buf * bp) 760{ 761 int oldflags, s; 762 struct buf *newbp; 763 764 if (bp->b_flags & B_INVAL) { 765 brelse(bp); 766 return (0); 767 } 768 769 oldflags = bp->b_flags; 770 771 if (BUF_REFCNT(bp) == 0) 772 panic("bwrite: buffer is not busy???"); 773 s = splbio(); 774 /* 775 * If a background write is already in progress, delay 776 * writing this block if it is asynchronous. Otherwise 777 * wait for the background write to complete. 778 */ 779 VI_LOCK(bp->b_vp); 780 if (bp->b_vflags & BV_BKGRDINPROG) { 781 if (bp->b_flags & B_ASYNC) { 782 VI_UNLOCK(bp->b_vp); 783 splx(s); 784 bdwrite(bp); 785 return (0); 786 } 787 bp->b_vflags |= BV_BKGRDWAIT; 788 msleep(&bp->b_xflags, VI_MTX(bp->b_vp), PRIBIO, "bwrbg", 0); 789 if (bp->b_vflags & BV_BKGRDINPROG) 790 panic("bwrite: still writing"); 791 } 792 VI_UNLOCK(bp->b_vp); 793 794 /* Mark the buffer clean */ 795 bundirty(bp); 796 797 /* 798 * If this buffer is marked for background writing and we 799 * do not have to wait for it, make a copy and write the 800 * copy so as to leave this buffer ready for further use. 801 * 802 * This optimization eats a lot of memory. If we have a page 803 * or buffer shortfall we can't do it. 804 */ 805 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 806 (bp->b_flags & B_ASYNC) && 807 !vm_page_count_severe() && 808 !buf_dirty_count_severe()) { 809 if (bp->b_iodone != NULL) { 810 printf("bp->b_iodone = %p\n", bp->b_iodone); 811 panic("bwrite: need chained iodone"); 812 } 813 814 /* get a new block */ 815 newbp = geteblk(bp->b_bufsize); 816 817 /* 818 * set it to be identical to the old block. We have to 819 * set b_lblkno and BKGRDMARKER before calling bgetvp() 820 * to avoid confusing the splay tree and gbincore(). 821 */ 822 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 823 newbp->b_lblkno = bp->b_lblkno; 824 newbp->b_xflags |= BX_BKGRDMARKER; 825 VI_LOCK(bp->b_vp); 826 bp->b_vflags |= BV_BKGRDINPROG; 827 bgetvp(bp->b_vp, newbp); 828 VI_UNLOCK(bp->b_vp); 829 newbp->b_blkno = bp->b_blkno; 830 newbp->b_offset = bp->b_offset; 831 newbp->b_iodone = vfs_backgroundwritedone; 832 newbp->b_flags |= B_ASYNC; 833 newbp->b_flags &= ~B_INVAL; 834 835 /* move over the dependencies */ 836 if (LIST_FIRST(&bp->b_dep) != NULL) 837 buf_movedeps(bp, newbp); 838 839 /* 840 * Initiate write on the copy, release the original to 841 * the B_LOCKED queue so that it cannot go away until 842 * the background write completes. If not locked it could go 843 * away and then be reconstituted while it was being written. 844 * If the reconstituted buffer were written, we could end up 845 * with two background copies being written at the same time. 846 */ 847 bqrelse(bp); 848 bp = newbp; 849 } 850 851 bp->b_flags &= ~B_DONE; 852 bp->b_ioflags &= ~BIO_ERROR; 853 bp->b_flags |= B_WRITEINPROG | B_CACHE; 854 bp->b_iocmd = BIO_WRITE; 855 856 VI_LOCK(bp->b_vp); 857 bp->b_vp->v_numoutput++; 858 VI_UNLOCK(bp->b_vp); 859 vfs_busy_pages(bp, 1); 860 861 /* 862 * Normal bwrites pipeline writes 863 */ 864 bp->b_runningbufspace = bp->b_bufsize; 865 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 866 867 if (curthread != PCPU_GET(idlethread)) 868 curthread->td_proc->p_stats->p_ru.ru_oublock++; 869 splx(s); 870 if (oldflags & B_ASYNC) 871 BUF_KERNPROC(bp); 872 if (bp->b_vp->v_type == VCHR) 873 VOP_SPECSTRATEGY(bp->b_vp, bp); 874 else 875 VOP_STRATEGY(bp->b_vp, bp); 876 877 if ((oldflags & B_ASYNC) == 0) { 878 int rtval = bufwait(bp); 879 brelse(bp); 880 return (rtval); 881 } else { 882 /* 883 * don't allow the async write to saturate the I/O 884 * system. We will not deadlock here because 885 * we are blocking waiting for I/O that is already in-progress 886 * to complete. 887 */ 888 waitrunningbufspace(); 889 } 890 891 return (0); 892} 893 894/* 895 * Complete a background write started from bwrite. 896 */ 897static void 898vfs_backgroundwritedone(bp) 899 struct buf *bp; 900{ 901 struct buf *origbp; 902 903 /* 904 * Find the original buffer that we are writing. 905 */ 906 VI_LOCK(bp->b_vp); 907 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 908 panic("backgroundwritedone: lost buffer"); 909 910 /* 911 * Clear the BV_BKGRDINPROG flag in the original buffer 912 * and awaken it if it is waiting for the write to complete. 913 * If BV_BKGRDINPROG is not set in the original buffer it must 914 * have been released and re-instantiated - which is not legal. 915 */ 916 KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 917 ("backgroundwritedone: lost buffer2")); 918 origbp->b_vflags &= ~BV_BKGRDINPROG; 919 if (origbp->b_vflags & BV_BKGRDWAIT) { 920 origbp->b_vflags &= ~BV_BKGRDWAIT; 921 wakeup(&origbp->b_xflags); 922 } 923 VI_UNLOCK(bp->b_vp); 924 /* 925 * Process dependencies then return any unfinished ones. 926 */ 927 if (LIST_FIRST(&bp->b_dep) != NULL) 928 buf_complete(bp); 929 if (LIST_FIRST(&bp->b_dep) != NULL) 930 buf_movedeps(bp, origbp); 931 932 /* 933 * This buffer is marked B_NOCACHE, so when it is released 934 * by biodone, it will be tossed. We mark it with BIO_READ 935 * to avoid biodone doing a second vwakeup. 936 */ 937 bp->b_flags |= B_NOCACHE; 938 bp->b_iocmd = BIO_READ; 939 bp->b_flags &= ~(B_CACHE | B_DONE); 940 bp->b_iodone = 0; 941 bufdone(bp); 942} 943 944/* 945 * Delayed write. (Buffer is marked dirty). Do not bother writing 946 * anything if the buffer is marked invalid. 947 * 948 * Note that since the buffer must be completely valid, we can safely 949 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 950 * biodone() in order to prevent getblk from writing the buffer 951 * out synchronously. 952 */ 953void 954bdwrite(struct buf * bp) 955{ 956 struct thread *td = curthread; 957 struct vnode *vp; 958 struct buf *nbp; 959 960 GIANT_REQUIRED; 961 962 if (BUF_REFCNT(bp) == 0) 963 panic("bdwrite: buffer is not busy"); 964 965 if (bp->b_flags & B_INVAL) { 966 brelse(bp); 967 return; 968 } 969 970 /* 971 * If we have too many dirty buffers, don't create any more. 972 * If we are wildly over our limit, then force a complete 973 * cleanup. Otherwise, just keep the situation from getting 974 * out of control. Note that we have to avoid a recursive 975 * disaster and not try to clean up after our own cleanup! 976 */ 977 vp = bp->b_vp; 978 VI_LOCK(vp); 979 if (td->td_proc->p_flag & P_COWINPROGRESS) { 980 recursiveflushes++; 981 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 982 VI_UNLOCK(vp); 983 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 984 VI_LOCK(vp); 985 altbufferflushes++; 986 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 987 /* 988 * Try to find a buffer to flush. 989 */ 990 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 991 if ((nbp->b_vflags & BV_BKGRDINPROG) || 992 buf_countdeps(nbp, 0) || 993 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 994 continue; 995 if (bp == nbp) 996 panic("bdwrite: found ourselves"); 997 VI_UNLOCK(vp); 998 if (nbp->b_flags & B_CLUSTEROK) { 999 vfs_bio_awrite(nbp); 1000 } else { 1001 bremfree(nbp); 1002 bawrite(nbp); 1003 } 1004 VI_LOCK(vp); 1005 dirtybufferflushes++; 1006 break; 1007 } 1008 } 1009 VI_UNLOCK(vp); 1010 1011 bdirty(bp); 1012 /* 1013 * Set B_CACHE, indicating that the buffer is fully valid. This is 1014 * true even of NFS now. 1015 */ 1016 bp->b_flags |= B_CACHE; 1017 1018 /* 1019 * This bmap keeps the system from needing to do the bmap later, 1020 * perhaps when the system is attempting to do a sync. Since it 1021 * is likely that the indirect block -- or whatever other datastructure 1022 * that the filesystem needs is still in memory now, it is a good 1023 * thing to do this. Note also, that if the pageout daemon is 1024 * requesting a sync -- there might not be enough memory to do 1025 * the bmap then... So, this is important to do. 1026 */ 1027 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1028 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1029 } 1030 1031 /* 1032 * Set the *dirty* buffer range based upon the VM system dirty pages. 1033 */ 1034 vfs_setdirty(bp); 1035 1036 /* 1037 * We need to do this here to satisfy the vnode_pager and the 1038 * pageout daemon, so that it thinks that the pages have been 1039 * "cleaned". Note that since the pages are in a delayed write 1040 * buffer -- the VFS layer "will" see that the pages get written 1041 * out on the next sync, or perhaps the cluster will be completed. 1042 */ 1043 vfs_clean_pages(bp); 1044 bqrelse(bp); 1045 1046 /* 1047 * Wakeup the buffer flushing daemon if we have a lot of dirty 1048 * buffers (midpoint between our recovery point and our stall 1049 * point). 1050 */ 1051 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1052 1053 /* 1054 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1055 * due to the softdep code. 1056 */ 1057} 1058 1059/* 1060 * bdirty: 1061 * 1062 * Turn buffer into delayed write request. We must clear BIO_READ and 1063 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1064 * itself to properly update it in the dirty/clean lists. We mark it 1065 * B_DONE to ensure that any asynchronization of the buffer properly 1066 * clears B_DONE ( else a panic will occur later ). 1067 * 1068 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1069 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1070 * should only be called if the buffer is known-good. 1071 * 1072 * Since the buffer is not on a queue, we do not update the numfreebuffers 1073 * count. 1074 * 1075 * Must be called at splbio(). 1076 * The buffer must be on QUEUE_NONE. 1077 */ 1078void 1079bdirty(bp) 1080 struct buf *bp; 1081{ 1082 KASSERT(bp->b_qindex == QUEUE_NONE, 1083 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1084 bp->b_flags &= ~(B_RELBUF); 1085 bp->b_iocmd = BIO_WRITE; 1086 1087 if ((bp->b_flags & B_DELWRI) == 0) { 1088 bp->b_flags |= B_DONE | B_DELWRI; 1089 reassignbuf(bp, bp->b_vp); 1090 atomic_add_int(&numdirtybuffers, 1); 1091 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1092 } 1093} 1094 1095/* 1096 * bundirty: 1097 * 1098 * Clear B_DELWRI for buffer. 1099 * 1100 * Since the buffer is not on a queue, we do not update the numfreebuffers 1101 * count. 1102 * 1103 * Must be called at splbio(). 1104 * The buffer must be on QUEUE_NONE. 1105 */ 1106 1107void 1108bundirty(bp) 1109 struct buf *bp; 1110{ 1111 KASSERT(bp->b_qindex == QUEUE_NONE, 1112 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1113 1114 if (bp->b_flags & B_DELWRI) { 1115 bp->b_flags &= ~B_DELWRI; 1116 reassignbuf(bp, bp->b_vp); 1117 atomic_subtract_int(&numdirtybuffers, 1); 1118 numdirtywakeup(lodirtybuffers); 1119 } 1120 /* 1121 * Since it is now being written, we can clear its deferred write flag. 1122 */ 1123 bp->b_flags &= ~B_DEFERRED; 1124} 1125 1126/* 1127 * bawrite: 1128 * 1129 * Asynchronous write. Start output on a buffer, but do not wait for 1130 * it to complete. The buffer is released when the output completes. 1131 * 1132 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1133 * B_INVAL buffers. Not us. 1134 */ 1135void 1136bawrite(struct buf * bp) 1137{ 1138 bp->b_flags |= B_ASYNC; 1139 (void) BUF_WRITE(bp); 1140} 1141 1142/* 1143 * bwillwrite: 1144 * 1145 * Called prior to the locking of any vnodes when we are expecting to 1146 * write. We do not want to starve the buffer cache with too many 1147 * dirty buffers so we block here. By blocking prior to the locking 1148 * of any vnodes we attempt to avoid the situation where a locked vnode 1149 * prevents the various system daemons from flushing related buffers. 1150 */ 1151 1152void 1153bwillwrite(void) 1154{ 1155 if (numdirtybuffers >= hidirtybuffers) { 1156 int s; 1157 1158 mtx_lock(&Giant); 1159 s = splbio(); 1160 mtx_lock(&nblock); 1161 while (numdirtybuffers >= hidirtybuffers) { 1162 bd_wakeup(1); 1163 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1164 msleep(&needsbuffer, &nblock, 1165 (PRIBIO + 4), "flswai", 0); 1166 } 1167 splx(s); 1168 mtx_unlock(&nblock); 1169 mtx_unlock(&Giant); 1170 } 1171} 1172 1173/* 1174 * Return true if we have too many dirty buffers. 1175 */ 1176int 1177buf_dirty_count_severe(void) 1178{ 1179 return(numdirtybuffers >= hidirtybuffers); 1180} 1181 1182/* 1183 * brelse: 1184 * 1185 * Release a busy buffer and, if requested, free its resources. The 1186 * buffer will be stashed in the appropriate bufqueue[] allowing it 1187 * to be accessed later as a cache entity or reused for other purposes. 1188 */ 1189void 1190brelse(struct buf * bp) 1191{ 1192 int s; 1193 1194 GIANT_REQUIRED; 1195 1196 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1197 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1198 1199 s = splbio(); 1200 1201 if (bp->b_iocmd == BIO_WRITE && 1202 (bp->b_ioflags & BIO_ERROR) && 1203 !(bp->b_flags & B_INVAL)) { 1204 /* 1205 * Failed write, redirty. Must clear BIO_ERROR to prevent 1206 * pages from being scrapped. If B_INVAL is set then 1207 * this case is not run and the next case is run to 1208 * destroy the buffer. B_INVAL can occur if the buffer 1209 * is outside the range supported by the underlying device. 1210 */ 1211 bp->b_ioflags &= ~BIO_ERROR; 1212 bdirty(bp); 1213 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1214 (bp->b_ioflags & BIO_ERROR) || 1215 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1216 /* 1217 * Either a failed I/O or we were asked to free or not 1218 * cache the buffer. 1219 */ 1220 bp->b_flags |= B_INVAL; 1221 if (LIST_FIRST(&bp->b_dep) != NULL) 1222 buf_deallocate(bp); 1223 if (bp->b_flags & B_DELWRI) { 1224 atomic_subtract_int(&numdirtybuffers, 1); 1225 numdirtywakeup(lodirtybuffers); 1226 } 1227 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1228 if ((bp->b_flags & B_VMIO) == 0) { 1229 if (bp->b_bufsize) 1230 allocbuf(bp, 0); 1231 if (bp->b_vp) 1232 brelvp(bp); 1233 } 1234 } 1235 1236 /* 1237 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1238 * is called with B_DELWRI set, the underlying pages may wind up 1239 * getting freed causing a previous write (bdwrite()) to get 'lost' 1240 * because pages associated with a B_DELWRI bp are marked clean. 1241 * 1242 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1243 * if B_DELWRI is set. 1244 * 1245 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1246 * on pages to return pages to the VM page queues. 1247 */ 1248 if (bp->b_flags & B_DELWRI) 1249 bp->b_flags &= ~B_RELBUF; 1250 else if (vm_page_count_severe()) { 1251 /* 1252 * XXX This lock may not be necessary since BKGRDINPROG 1253 * cannot be set while we hold the buf lock, it can only be 1254 * cleared if it is already pending. 1255 */ 1256 if (bp->b_vp) { 1257 VI_LOCK(bp->b_vp); 1258 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1259 bp->b_flags |= B_RELBUF; 1260 VI_UNLOCK(bp->b_vp); 1261 } else 1262 bp->b_flags |= B_RELBUF; 1263 } 1264 1265 /* 1266 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1267 * constituted, not even NFS buffers now. Two flags effect this. If 1268 * B_INVAL, the struct buf is invalidated but the VM object is kept 1269 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1270 * 1271 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1272 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1273 * buffer is also B_INVAL because it hits the re-dirtying code above. 1274 * 1275 * Normally we can do this whether a buffer is B_DELWRI or not. If 1276 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1277 * the commit state and we cannot afford to lose the buffer. If the 1278 * buffer has a background write in progress, we need to keep it 1279 * around to prevent it from being reconstituted and starting a second 1280 * background write. 1281 */ 1282 if ((bp->b_flags & B_VMIO) 1283 && !(bp->b_vp->v_mount != NULL && 1284 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1285 !vn_isdisk(bp->b_vp, NULL) && 1286 (bp->b_flags & B_DELWRI)) 1287 ) { 1288 1289 int i, j, resid; 1290 vm_page_t m; 1291 off_t foff; 1292 vm_pindex_t poff; 1293 vm_object_t obj; 1294 struct vnode *vp; 1295 1296 vp = bp->b_vp; 1297 obj = bp->b_object; 1298 1299 /* 1300 * Get the base offset and length of the buffer. Note that 1301 * in the VMIO case if the buffer block size is not 1302 * page-aligned then b_data pointer may not be page-aligned. 1303 * But our b_pages[] array *IS* page aligned. 1304 * 1305 * block sizes less then DEV_BSIZE (usually 512) are not 1306 * supported due to the page granularity bits (m->valid, 1307 * m->dirty, etc...). 1308 * 1309 * See man buf(9) for more information 1310 */ 1311 resid = bp->b_bufsize; 1312 foff = bp->b_offset; 1313 if (obj != NULL) 1314 VM_OBJECT_LOCK(obj); 1315 for (i = 0; i < bp->b_npages; i++) { 1316 int had_bogus = 0; 1317 1318 m = bp->b_pages[i]; 1319 vm_page_lock_queues(); 1320 vm_page_flag_clear(m, PG_ZERO); 1321 vm_page_unlock_queues(); 1322 1323 /* 1324 * If we hit a bogus page, fixup *all* the bogus pages 1325 * now. 1326 */ 1327 if (m == bogus_page) { 1328 poff = OFF_TO_IDX(bp->b_offset); 1329 had_bogus = 1; 1330 1331 for (j = i; j < bp->b_npages; j++) { 1332 vm_page_t mtmp; 1333 mtmp = bp->b_pages[j]; 1334 if (mtmp == bogus_page) { 1335 mtmp = vm_page_lookup(obj, poff + j); 1336 if (!mtmp) { 1337 panic("brelse: page missing\n"); 1338 } 1339 bp->b_pages[j] = mtmp; 1340 } 1341 } 1342 1343 if ((bp->b_flags & B_INVAL) == 0) { 1344 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1345 } 1346 m = bp->b_pages[i]; 1347 } 1348 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1349 int poffset = foff & PAGE_MASK; 1350 int presid = resid > (PAGE_SIZE - poffset) ? 1351 (PAGE_SIZE - poffset) : resid; 1352 1353 KASSERT(presid >= 0, ("brelse: extra page")); 1354 vm_page_lock_queues(); 1355 vm_page_set_invalid(m, poffset, presid); 1356 vm_page_unlock_queues(); 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 if (obj != NULL) 1364 VM_OBJECT_UNLOCK(obj); 1365 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1366 vfs_vmio_release(bp); 1367 1368 } else if (bp->b_flags & B_VMIO) { 1369 1370 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1371 vfs_vmio_release(bp); 1372 } 1373 1374 } 1375 1376 if (bp->b_qindex != QUEUE_NONE) 1377 panic("brelse: free buffer onto another queue???"); 1378 if (BUF_REFCNT(bp) > 1) { 1379 /* do not release to free list */ 1380 BUF_UNLOCK(bp); 1381 splx(s); 1382 return; 1383 } 1384 1385 /* enqueue */ 1386 mtx_lock(&bqlock); 1387 1388 /* buffers with no memory */ 1389 if (bp->b_bufsize == 0) { 1390 bp->b_flags |= B_INVAL; 1391 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1392 if (bp->b_vflags & BV_BKGRDINPROG) 1393 panic("losing buffer 1"); 1394 if (bp->b_kvasize) { 1395 bp->b_qindex = QUEUE_EMPTYKVA; 1396 } else { 1397 bp->b_qindex = QUEUE_EMPTY; 1398 } 1399 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1400 bp->b_dev = NODEV; 1401 /* buffers with junk contents */ 1402 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1403 (bp->b_ioflags & BIO_ERROR)) { 1404 bp->b_flags |= B_INVAL; 1405 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1406 if (bp->b_vflags & BV_BKGRDINPROG) 1407 panic("losing buffer 2"); 1408 bp->b_qindex = QUEUE_CLEAN; 1409 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1410 bp->b_dev = NODEV; 1411 /* remaining buffers */ 1412 } else { 1413 if (bp->b_flags & B_DELWRI) 1414 bp->b_qindex = QUEUE_DIRTY; 1415 else 1416 bp->b_qindex = QUEUE_CLEAN; 1417 if (bp->b_flags & B_AGE) 1418 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1419 else 1420 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1421 } 1422 mtx_unlock(&bqlock); 1423 1424 /* 1425 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1426 * placed the buffer on the correct queue. We must also disassociate 1427 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1428 * find it. 1429 */ 1430 if (bp->b_flags & B_INVAL) { 1431 if (bp->b_flags & B_DELWRI) 1432 bundirty(bp); 1433 if (bp->b_vp) 1434 brelvp(bp); 1435 } 1436 1437 /* 1438 * Fixup numfreebuffers count. The bp is on an appropriate queue 1439 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1440 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1441 * if B_INVAL is set ). 1442 */ 1443 1444 if (!(bp->b_flags & B_DELWRI)) 1445 bufcountwakeup(); 1446 1447 /* 1448 * Something we can maybe free or reuse 1449 */ 1450 if (bp->b_bufsize || bp->b_kvasize) 1451 bufspacewakeup(); 1452 1453 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1454 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1455 panic("brelse: not dirty"); 1456 /* unlock */ 1457 BUF_UNLOCK(bp); 1458 splx(s); 1459} 1460 1461/* 1462 * Release a buffer back to the appropriate queue but do not try to free 1463 * it. The buffer is expected to be used again soon. 1464 * 1465 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1466 * biodone() to requeue an async I/O on completion. It is also used when 1467 * known good buffers need to be requeued but we think we may need the data 1468 * again soon. 1469 * 1470 * XXX we should be able to leave the B_RELBUF hint set on completion. 1471 */ 1472void 1473bqrelse(struct buf * bp) 1474{ 1475 int s; 1476 1477 s = splbio(); 1478 1479 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1480 1481 if (bp->b_qindex != QUEUE_NONE) 1482 panic("bqrelse: free buffer onto another queue???"); 1483 if (BUF_REFCNT(bp) > 1) { 1484 /* do not release to free list */ 1485 BUF_UNLOCK(bp); 1486 splx(s); 1487 return; 1488 } 1489 mtx_lock(&bqlock); 1490 /* buffers with stale but valid contents */ 1491 if (bp->b_flags & B_DELWRI) { 1492 bp->b_qindex = QUEUE_DIRTY; 1493 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1494 } else { 1495 /* 1496 * XXX This lock may not be necessary since BKGRDINPROG 1497 * cannot be set while we hold the buf lock, it can only be 1498 * cleared if it is already pending. 1499 */ 1500 VI_LOCK(bp->b_vp); 1501 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1502 VI_UNLOCK(bp->b_vp); 1503 bp->b_qindex = QUEUE_CLEAN; 1504 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1505 b_freelist); 1506 } else { 1507 /* 1508 * We are too low on memory, we have to try to free 1509 * the buffer (most importantly: the wired pages 1510 * making up its backing store) *now*. 1511 */ 1512 VI_UNLOCK(bp->b_vp); 1513 mtx_unlock(&bqlock); 1514 splx(s); 1515 brelse(bp); 1516 return; 1517 } 1518 } 1519 mtx_unlock(&bqlock); 1520 1521 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1522 bufcountwakeup(); 1523 1524 /* 1525 * Something we can maybe free or reuse. 1526 */ 1527 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1528 bufspacewakeup(); 1529 1530 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1531 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1532 panic("bqrelse: not dirty"); 1533 /* unlock */ 1534 BUF_UNLOCK(bp); 1535 splx(s); 1536} 1537 1538/* Give pages used by the bp back to the VM system (where possible) */ 1539static void 1540vfs_vmio_release(bp) 1541 struct buf *bp; 1542{ 1543 int i; 1544 vm_page_t m; 1545 1546 GIANT_REQUIRED; 1547 if (bp->b_object != NULL) 1548 VM_OBJECT_LOCK(bp->b_object); 1549 vm_page_lock_queues(); 1550 for (i = 0; i < bp->b_npages; i++) { 1551 m = bp->b_pages[i]; 1552 bp->b_pages[i] = NULL; 1553 /* 1554 * In order to keep page LRU ordering consistent, put 1555 * everything on the inactive queue. 1556 */ 1557 vm_page_unwire(m, 0); 1558 /* 1559 * We don't mess with busy pages, it is 1560 * the responsibility of the process that 1561 * busied the pages to deal with them. 1562 */ 1563 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1564 continue; 1565 1566 if (m->wire_count == 0) { 1567 vm_page_flag_clear(m, PG_ZERO); 1568 /* 1569 * Might as well free the page if we can and it has 1570 * no valid data. We also free the page if the 1571 * buffer was used for direct I/O 1572 */ 1573 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1574 m->hold_count == 0) { 1575 vm_page_busy(m); 1576 pmap_remove_all(m); 1577 vm_page_free(m); 1578 } else if (bp->b_flags & B_DIRECT) { 1579 vm_page_try_to_free(m); 1580 } else if (vm_page_count_severe()) { 1581 vm_page_try_to_cache(m); 1582 } 1583 } 1584 } 1585 vm_page_unlock_queues(); 1586 if (bp->b_object != NULL) 1587 VM_OBJECT_UNLOCK(bp->b_object); 1588 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1589 1590 if (bp->b_bufsize) { 1591 bufspacewakeup(); 1592 bp->b_bufsize = 0; 1593 } 1594 bp->b_npages = 0; 1595 bp->b_flags &= ~B_VMIO; 1596 if (bp->b_vp) 1597 brelvp(bp); 1598} 1599 1600/* 1601 * Check to see if a block at a particular lbn is available for a clustered 1602 * write. 1603 */ 1604static int 1605vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1606{ 1607 struct buf *bpa; 1608 int match; 1609 1610 match = 0; 1611 1612 /* If the buf isn't in core skip it */ 1613 if ((bpa = gbincore(vp, lblkno)) == NULL) 1614 return (0); 1615 1616 /* If the buf is busy we don't want to wait for it */ 1617 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1618 return (0); 1619 1620 /* Only cluster with valid clusterable delayed write buffers */ 1621 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1622 (B_DELWRI | B_CLUSTEROK)) 1623 goto done; 1624 1625 if (bpa->b_bufsize != size) 1626 goto done; 1627 1628 /* 1629 * Check to see if it is in the expected place on disk and that the 1630 * block has been mapped. 1631 */ 1632 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1633 match = 1; 1634done: 1635 BUF_UNLOCK(bpa); 1636 return (match); 1637} 1638 1639/* 1640 * vfs_bio_awrite: 1641 * 1642 * Implement clustered async writes for clearing out B_DELWRI buffers. 1643 * This is much better then the old way of writing only one buffer at 1644 * a time. Note that we may not be presented with the buffers in the 1645 * correct order, so we search for the cluster in both directions. 1646 */ 1647int 1648vfs_bio_awrite(struct buf * bp) 1649{ 1650 int i; 1651 int j; 1652 daddr_t lblkno = bp->b_lblkno; 1653 struct vnode *vp = bp->b_vp; 1654 int s; 1655 int ncl; 1656 int nwritten; 1657 int size; 1658 int maxcl; 1659 1660 s = splbio(); 1661 /* 1662 * right now we support clustered writing only to regular files. If 1663 * we find a clusterable block we could be in the middle of a cluster 1664 * rather then at the beginning. 1665 */ 1666 if ((vp->v_type == VREG) && 1667 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1668 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1669 1670 size = vp->v_mount->mnt_stat.f_iosize; 1671 maxcl = MAXPHYS / size; 1672 1673 VI_LOCK(vp); 1674 for (i = 1; i < maxcl; i++) 1675 if (vfs_bio_clcheck(vp, size, lblkno + i, 1676 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1677 break; 1678 1679 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1680 if (vfs_bio_clcheck(vp, size, lblkno - j, 1681 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1682 break; 1683 1684 VI_UNLOCK(vp); 1685 --j; 1686 ncl = i + j; 1687 /* 1688 * this is a possible cluster write 1689 */ 1690 if (ncl != 1) { 1691 BUF_UNLOCK(bp); 1692 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1693 splx(s); 1694 return nwritten; 1695 } 1696 } 1697 1698 bremfree(bp); 1699 bp->b_flags |= B_ASYNC; 1700 1701 splx(s); 1702 /* 1703 * default (old) behavior, writing out only one block 1704 * 1705 * XXX returns b_bufsize instead of b_bcount for nwritten? 1706 */ 1707 nwritten = bp->b_bufsize; 1708 (void) BUF_WRITE(bp); 1709 1710 return nwritten; 1711} 1712 1713/* 1714 * getnewbuf: 1715 * 1716 * Find and initialize a new buffer header, freeing up existing buffers 1717 * in the bufqueues as necessary. The new buffer is returned locked. 1718 * 1719 * Important: B_INVAL is not set. If the caller wishes to throw the 1720 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1721 * 1722 * We block if: 1723 * We have insufficient buffer headers 1724 * We have insufficient buffer space 1725 * buffer_map is too fragmented ( space reservation fails ) 1726 * If we have to flush dirty buffers ( but we try to avoid this ) 1727 * 1728 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1729 * Instead we ask the buf daemon to do it for us. We attempt to 1730 * avoid piecemeal wakeups of the pageout daemon. 1731 */ 1732 1733static struct buf * 1734getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1735{ 1736 struct buf *bp; 1737 struct buf *nbp; 1738 int defrag = 0; 1739 int nqindex; 1740 static int flushingbufs; 1741 1742 GIANT_REQUIRED; 1743 1744 /* 1745 * We can't afford to block since we might be holding a vnode lock, 1746 * which may prevent system daemons from running. We deal with 1747 * low-memory situations by proactively returning memory and running 1748 * async I/O rather then sync I/O. 1749 */ 1750 1751 atomic_add_int(&getnewbufcalls, 1); 1752 atomic_subtract_int(&getnewbufrestarts, 1); 1753restart: 1754 atomic_add_int(&getnewbufrestarts, 1); 1755 1756 /* 1757 * Setup for scan. If we do not have enough free buffers, 1758 * we setup a degenerate case that immediately fails. Note 1759 * that if we are specially marked process, we are allowed to 1760 * dip into our reserves. 1761 * 1762 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1763 * 1764 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1765 * However, there are a number of cases (defragging, reusing, ...) 1766 * where we cannot backup. 1767 */ 1768 mtx_lock(&bqlock); 1769 nqindex = QUEUE_EMPTYKVA; 1770 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1771 1772 if (nbp == NULL) { 1773 /* 1774 * If no EMPTYKVA buffers and we are either 1775 * defragging or reusing, locate a CLEAN buffer 1776 * to free or reuse. If bufspace useage is low 1777 * skip this step so we can allocate a new buffer. 1778 */ 1779 if (defrag || bufspace >= lobufspace) { 1780 nqindex = QUEUE_CLEAN; 1781 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1782 } 1783 1784 /* 1785 * If we could not find or were not allowed to reuse a 1786 * CLEAN buffer, check to see if it is ok to use an EMPTY 1787 * buffer. We can only use an EMPTY buffer if allocating 1788 * its KVA would not otherwise run us out of buffer space. 1789 */ 1790 if (nbp == NULL && defrag == 0 && 1791 bufspace + maxsize < hibufspace) { 1792 nqindex = QUEUE_EMPTY; 1793 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1794 } 1795 } 1796 1797 /* 1798 * Run scan, possibly freeing data and/or kva mappings on the fly 1799 * depending. 1800 */ 1801 1802 while ((bp = nbp) != NULL) { 1803 int qindex = nqindex; 1804 1805 /* 1806 * Calculate next bp ( we can only use it if we do not block 1807 * or do other fancy things ). 1808 */ 1809 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1810 switch(qindex) { 1811 case QUEUE_EMPTY: 1812 nqindex = QUEUE_EMPTYKVA; 1813 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1814 break; 1815 /* FALLTHROUGH */ 1816 case QUEUE_EMPTYKVA: 1817 nqindex = QUEUE_CLEAN; 1818 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1819 break; 1820 /* FALLTHROUGH */ 1821 case QUEUE_CLEAN: 1822 /* 1823 * nbp is NULL. 1824 */ 1825 break; 1826 } 1827 } 1828 if (bp->b_vp) { 1829 VI_LOCK(bp->b_vp); 1830 if (bp->b_vflags & BV_BKGRDINPROG) { 1831 VI_UNLOCK(bp->b_vp); 1832 continue; 1833 } 1834 VI_UNLOCK(bp->b_vp); 1835 } 1836 1837 /* 1838 * Sanity Checks 1839 */ 1840 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1841 1842 /* 1843 * Note: we no longer distinguish between VMIO and non-VMIO 1844 * buffers. 1845 */ 1846 1847 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1848 1849 /* 1850 * If we are defragging then we need a buffer with 1851 * b_kvasize != 0. XXX this situation should no longer 1852 * occur, if defrag is non-zero the buffer's b_kvasize 1853 * should also be non-zero at this point. XXX 1854 */ 1855 if (defrag && bp->b_kvasize == 0) { 1856 printf("Warning: defrag empty buffer %p\n", bp); 1857 continue; 1858 } 1859 1860 /* 1861 * Start freeing the bp. This is somewhat involved. nbp 1862 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1863 */ 1864 1865 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1866 panic("getnewbuf: locked buf"); 1867 bremfreel(bp); 1868 mtx_unlock(&bqlock); 1869 1870 if (qindex == QUEUE_CLEAN) { 1871 if (bp->b_flags & B_VMIO) { 1872 bp->b_flags &= ~B_ASYNC; 1873 vfs_vmio_release(bp); 1874 } 1875 if (bp->b_vp) 1876 brelvp(bp); 1877 } 1878 1879 /* 1880 * NOTE: nbp is now entirely invalid. We can only restart 1881 * the scan from this point on. 1882 * 1883 * Get the rest of the buffer freed up. b_kva* is still 1884 * valid after this operation. 1885 */ 1886 1887 if (bp->b_rcred != NOCRED) { 1888 crfree(bp->b_rcred); 1889 bp->b_rcred = NOCRED; 1890 } 1891 if (bp->b_wcred != NOCRED) { 1892 crfree(bp->b_wcred); 1893 bp->b_wcred = NOCRED; 1894 } 1895 if (LIST_FIRST(&bp->b_dep) != NULL) 1896 buf_deallocate(bp); 1897 if (bp->b_vflags & BV_BKGRDINPROG) 1898 panic("losing buffer 3"); 1899 1900 if (bp->b_bufsize) 1901 allocbuf(bp, 0); 1902 1903 bp->b_flags = 0; 1904 bp->b_ioflags = 0; 1905 bp->b_xflags = 0; 1906 bp->b_vflags = 0; 1907 bp->b_dev = NODEV; 1908 bp->b_vp = NULL; 1909 bp->b_blkno = bp->b_lblkno = 0; 1910 bp->b_offset = NOOFFSET; 1911 bp->b_iodone = 0; 1912 bp->b_error = 0; 1913 bp->b_resid = 0; 1914 bp->b_bcount = 0; 1915 bp->b_npages = 0; 1916 bp->b_dirtyoff = bp->b_dirtyend = 0; 1917 bp->b_magic = B_MAGIC_BIO; 1918 bp->b_op = &buf_ops_bio; 1919 bp->b_object = NULL; 1920 1921 LIST_INIT(&bp->b_dep); 1922 1923 /* 1924 * If we are defragging then free the buffer. 1925 */ 1926 if (defrag) { 1927 bp->b_flags |= B_INVAL; 1928 bfreekva(bp); 1929 brelse(bp); 1930 defrag = 0; 1931 goto restart; 1932 } 1933 1934 /* 1935 * If we are overcomitted then recover the buffer and its 1936 * KVM space. This occurs in rare situations when multiple 1937 * processes are blocked in getnewbuf() or allocbuf(). 1938 */ 1939 if (bufspace >= hibufspace) 1940 flushingbufs = 1; 1941 if (flushingbufs && bp->b_kvasize != 0) { 1942 bp->b_flags |= B_INVAL; 1943 bfreekva(bp); 1944 brelse(bp); 1945 goto restart; 1946 } 1947 if (bufspace < lobufspace) 1948 flushingbufs = 0; 1949 break; 1950 } 1951 1952 /* 1953 * If we exhausted our list, sleep as appropriate. We may have to 1954 * wakeup various daemons and write out some dirty buffers. 1955 * 1956 * Generally we are sleeping due to insufficient buffer space. 1957 */ 1958 1959 if (bp == NULL) { 1960 int flags; 1961 char *waitmsg; 1962 1963 mtx_unlock(&bqlock); 1964 if (defrag) { 1965 flags = VFS_BIO_NEED_BUFSPACE; 1966 waitmsg = "nbufkv"; 1967 } else if (bufspace >= hibufspace) { 1968 waitmsg = "nbufbs"; 1969 flags = VFS_BIO_NEED_BUFSPACE; 1970 } else { 1971 waitmsg = "newbuf"; 1972 flags = VFS_BIO_NEED_ANY; 1973 } 1974 1975 bd_speedup(); /* heeeelp */ 1976 1977 mtx_lock(&nblock); 1978 needsbuffer |= flags; 1979 while (needsbuffer & flags) { 1980 if (msleep(&needsbuffer, &nblock, 1981 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1982 mtx_unlock(&nblock); 1983 return (NULL); 1984 } 1985 } 1986 mtx_unlock(&nblock); 1987 } else { 1988 /* 1989 * We finally have a valid bp. We aren't quite out of the 1990 * woods, we still have to reserve kva space. In order 1991 * to keep fragmentation sane we only allocate kva in 1992 * BKVASIZE chunks. 1993 */ 1994 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1995 1996 if (maxsize != bp->b_kvasize) { 1997 vm_offset_t addr = 0; 1998 1999 bfreekva(bp); 2000 2001 if (vm_map_findspace(buffer_map, 2002 vm_map_min(buffer_map), maxsize, &addr)) { 2003 /* 2004 * Uh oh. Buffer map is to fragmented. We 2005 * must defragment the map. 2006 */ 2007 atomic_add_int(&bufdefragcnt, 1); 2008 defrag = 1; 2009 bp->b_flags |= B_INVAL; 2010 brelse(bp); 2011 goto restart; 2012 } 2013 if (addr) { 2014 vm_map_insert(buffer_map, NULL, 0, 2015 addr, addr + maxsize, 2016 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2017 2018 bp->b_kvabase = (caddr_t) addr; 2019 bp->b_kvasize = maxsize; 2020 atomic_add_int(&bufspace, bp->b_kvasize); 2021 atomic_add_int(&bufreusecnt, 1); 2022 } 2023 } 2024 bp->b_saveaddr = bp->b_kvabase; 2025 bp->b_data = bp->b_saveaddr; 2026 } 2027 return(bp); 2028} 2029 2030/* 2031 * buf_daemon: 2032 * 2033 * buffer flushing daemon. Buffers are normally flushed by the 2034 * update daemon but if it cannot keep up this process starts to 2035 * take the load in an attempt to prevent getnewbuf() from blocking. 2036 */ 2037 2038static struct proc *bufdaemonproc; 2039 2040static struct kproc_desc buf_kp = { 2041 "bufdaemon", 2042 buf_daemon, 2043 &bufdaemonproc 2044}; 2045SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2046 2047static void 2048buf_daemon() 2049{ 2050 int s; 2051 2052 mtx_lock(&Giant); 2053 2054 /* 2055 * This process needs to be suspended prior to shutdown sync. 2056 */ 2057 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2058 SHUTDOWN_PRI_LAST); 2059 2060 /* 2061 * This process is allowed to take the buffer cache to the limit 2062 */ 2063 s = splbio(); 2064 mtx_lock(&bdlock); 2065 2066 for (;;) { 2067 bd_request = 0; 2068 mtx_unlock(&bdlock); 2069 2070 kthread_suspend_check(bufdaemonproc); 2071 2072 /* 2073 * Do the flush. Limit the amount of in-transit I/O we 2074 * allow to build up, otherwise we would completely saturate 2075 * the I/O system. Wakeup any waiting processes before we 2076 * normally would so they can run in parallel with our drain. 2077 */ 2078 while (numdirtybuffers > lodirtybuffers) { 2079 if (flushbufqueues(0) == 0) { 2080 /* 2081 * Could not find any buffers without rollback 2082 * dependencies, so just write the first one 2083 * in the hopes of eventually making progress. 2084 */ 2085 flushbufqueues(1); 2086 break; 2087 } 2088 waitrunningbufspace(); 2089 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2090 } 2091 2092 /* 2093 * Only clear bd_request if we have reached our low water 2094 * mark. The buf_daemon normally waits 1 second and 2095 * then incrementally flushes any dirty buffers that have 2096 * built up, within reason. 2097 * 2098 * If we were unable to hit our low water mark and couldn't 2099 * find any flushable buffers, we sleep half a second. 2100 * Otherwise we loop immediately. 2101 */ 2102 mtx_lock(&bdlock); 2103 if (numdirtybuffers <= lodirtybuffers) { 2104 /* 2105 * We reached our low water mark, reset the 2106 * request and sleep until we are needed again. 2107 * The sleep is just so the suspend code works. 2108 */ 2109 bd_request = 0; 2110 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2111 } else { 2112 /* 2113 * We couldn't find any flushable dirty buffers but 2114 * still have too many dirty buffers, we 2115 * have to sleep and try again. (rare) 2116 */ 2117 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2118 } 2119 } 2120} 2121 2122/* 2123 * flushbufqueues: 2124 * 2125 * Try to flush a buffer in the dirty queue. We must be careful to 2126 * free up B_INVAL buffers instead of write them, which NFS is 2127 * particularly sensitive to. 2128 */ 2129int flushwithdeps = 0; 2130SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2131 0, "Number of buffers flushed with dependecies that require rollbacks"); 2132static int 2133flushbufqueues(int flushdeps) 2134{ 2135 struct thread *td = curthread; 2136 struct vnode *vp; 2137 struct buf *bp; 2138 int hasdeps; 2139 2140 mtx_lock(&bqlock); 2141 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2142 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2143 continue; 2144 KASSERT((bp->b_flags & B_DELWRI), 2145 ("unexpected clean buffer %p", bp)); 2146 VI_LOCK(bp->b_vp); 2147 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2148 VI_UNLOCK(bp->b_vp); 2149 BUF_UNLOCK(bp); 2150 continue; 2151 } 2152 VI_UNLOCK(bp->b_vp); 2153 if (bp->b_flags & B_INVAL) { 2154 bremfreel(bp); 2155 mtx_unlock(&bqlock); 2156 brelse(bp); 2157 return (1); 2158 } 2159 2160 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2161 if (flushdeps == 0) { 2162 BUF_UNLOCK(bp); 2163 continue; 2164 } 2165 hasdeps = 1; 2166 } else 2167 hasdeps = 0; 2168 /* 2169 * We must hold the lock on a vnode before writing 2170 * one of its buffers. Otherwise we may confuse, or 2171 * in the case of a snapshot vnode, deadlock the 2172 * system. 2173 * 2174 * The lock order here is the reverse of the normal 2175 * of vnode followed by buf lock. This is ok because 2176 * the NOWAIT will prevent deadlock. 2177 */ 2178 if ((vp = bp->b_vp) == NULL || 2179 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2180 mtx_unlock(&bqlock); 2181 vfs_bio_awrite(bp); 2182 if (vp != NULL) 2183 VOP_UNLOCK(vp, 0, td); 2184 flushwithdeps += hasdeps; 2185 return (1); 2186 } 2187 BUF_UNLOCK(bp); 2188 } 2189 mtx_unlock(&bqlock); 2190 return (0); 2191} 2192 2193/* 2194 * Check to see if a block is currently memory resident. 2195 */ 2196struct buf * 2197incore(struct vnode * vp, daddr_t blkno) 2198{ 2199 struct buf *bp; 2200 2201 int s = splbio(); 2202 VI_LOCK(vp); 2203 bp = gbincore(vp, blkno); 2204 VI_UNLOCK(vp); 2205 splx(s); 2206 return (bp); 2207} 2208 2209/* 2210 * Returns true if no I/O is needed to access the 2211 * associated VM object. This is like incore except 2212 * it also hunts around in the VM system for the data. 2213 */ 2214 2215int 2216inmem(struct vnode * vp, daddr_t blkno) 2217{ 2218 vm_object_t obj; 2219 vm_offset_t toff, tinc, size; 2220 vm_page_t m; 2221 vm_ooffset_t off; 2222 2223 GIANT_REQUIRED; 2224 ASSERT_VOP_LOCKED(vp, "inmem"); 2225 2226 if (incore(vp, blkno)) 2227 return 1; 2228 if (vp->v_mount == NULL) 2229 return 0; 2230 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2231 return 0; 2232 2233 size = PAGE_SIZE; 2234 if (size > vp->v_mount->mnt_stat.f_iosize) 2235 size = vp->v_mount->mnt_stat.f_iosize; 2236 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2237 2238 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2239 VM_OBJECT_LOCK(obj); 2240 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2241 VM_OBJECT_UNLOCK(obj); 2242 if (!m) 2243 goto notinmem; 2244 tinc = size; 2245 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2246 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2247 if (vm_page_is_valid(m, 2248 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2249 goto notinmem; 2250 } 2251 return 1; 2252 2253notinmem: 2254 return (0); 2255} 2256 2257/* 2258 * vfs_setdirty: 2259 * 2260 * Sets the dirty range for a buffer based on the status of the dirty 2261 * bits in the pages comprising the buffer. 2262 * 2263 * The range is limited to the size of the buffer. 2264 * 2265 * This routine is primarily used by NFS, but is generalized for the 2266 * B_VMIO case. 2267 */ 2268static void 2269vfs_setdirty(struct buf *bp) 2270{ 2271 int i; 2272 vm_object_t object; 2273 2274 GIANT_REQUIRED; 2275 /* 2276 * Degenerate case - empty buffer 2277 */ 2278 2279 if (bp->b_bufsize == 0) 2280 return; 2281 2282 /* 2283 * We qualify the scan for modified pages on whether the 2284 * object has been flushed yet. The OBJ_WRITEABLE flag 2285 * is not cleared simply by protecting pages off. 2286 */ 2287 2288 if ((bp->b_flags & B_VMIO) == 0) 2289 return; 2290 2291 object = bp->b_pages[0]->object; 2292 VM_OBJECT_LOCK(object); 2293 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2294 printf("Warning: object %p writeable but not mightbedirty\n", object); 2295 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2296 printf("Warning: object %p mightbedirty but not writeable\n", object); 2297 2298 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2299 vm_offset_t boffset; 2300 vm_offset_t eoffset; 2301 2302 vm_page_lock_queues(); 2303 /* 2304 * test the pages to see if they have been modified directly 2305 * by users through the VM system. 2306 */ 2307 for (i = 0; i < bp->b_npages; i++) { 2308 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2309 vm_page_test_dirty(bp->b_pages[i]); 2310 } 2311 2312 /* 2313 * Calculate the encompassing dirty range, boffset and eoffset, 2314 * (eoffset - boffset) bytes. 2315 */ 2316 2317 for (i = 0; i < bp->b_npages; i++) { 2318 if (bp->b_pages[i]->dirty) 2319 break; 2320 } 2321 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2322 2323 for (i = bp->b_npages - 1; i >= 0; --i) { 2324 if (bp->b_pages[i]->dirty) { 2325 break; 2326 } 2327 } 2328 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2329 2330 vm_page_unlock_queues(); 2331 /* 2332 * Fit it to the buffer. 2333 */ 2334 2335 if (eoffset > bp->b_bcount) 2336 eoffset = bp->b_bcount; 2337 2338 /* 2339 * If we have a good dirty range, merge with the existing 2340 * dirty range. 2341 */ 2342 2343 if (boffset < eoffset) { 2344 if (bp->b_dirtyoff > boffset) 2345 bp->b_dirtyoff = boffset; 2346 if (bp->b_dirtyend < eoffset) 2347 bp->b_dirtyend = eoffset; 2348 } 2349 } 2350 VM_OBJECT_UNLOCK(object); 2351} 2352 2353/* 2354 * getblk: 2355 * 2356 * Get a block given a specified block and offset into a file/device. 2357 * The buffers B_DONE bit will be cleared on return, making it almost 2358 * ready for an I/O initiation. B_INVAL may or may not be set on 2359 * return. The caller should clear B_INVAL prior to initiating a 2360 * READ. 2361 * 2362 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2363 * an existing buffer. 2364 * 2365 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2366 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2367 * and then cleared based on the backing VM. If the previous buffer is 2368 * non-0-sized but invalid, B_CACHE will be cleared. 2369 * 2370 * If getblk() must create a new buffer, the new buffer is returned with 2371 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2372 * case it is returned with B_INVAL clear and B_CACHE set based on the 2373 * backing VM. 2374 * 2375 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2376 * B_CACHE bit is clear. 2377 * 2378 * What this means, basically, is that the caller should use B_CACHE to 2379 * determine whether the buffer is fully valid or not and should clear 2380 * B_INVAL prior to issuing a read. If the caller intends to validate 2381 * the buffer by loading its data area with something, the caller needs 2382 * to clear B_INVAL. If the caller does this without issuing an I/O, 2383 * the caller should set B_CACHE ( as an optimization ), else the caller 2384 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2385 * a write attempt or if it was a successfull read. If the caller 2386 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2387 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2388 */ 2389struct buf * 2390getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2391 int flags) 2392{ 2393 struct buf *bp; 2394 int s; 2395 int error; 2396 ASSERT_VOP_LOCKED(vp, "getblk"); 2397 2398 if (size > MAXBSIZE) 2399 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2400 2401 s = splbio(); 2402loop: 2403 /* 2404 * Block if we are low on buffers. Certain processes are allowed 2405 * to completely exhaust the buffer cache. 2406 * 2407 * If this check ever becomes a bottleneck it may be better to 2408 * move it into the else, when gbincore() fails. At the moment 2409 * it isn't a problem. 2410 * 2411 * XXX remove if 0 sections (clean this up after its proven) 2412 */ 2413 if (numfreebuffers == 0) { 2414 if (curthread == PCPU_GET(idlethread)) 2415 return NULL; 2416 mtx_lock(&nblock); 2417 needsbuffer |= VFS_BIO_NEED_ANY; 2418 mtx_unlock(&nblock); 2419 } 2420 2421 VI_LOCK(vp); 2422 if ((bp = gbincore(vp, blkno))) { 2423 int lockflags; 2424 /* 2425 * Buffer is in-core. If the buffer is not busy, it must 2426 * be on a queue. 2427 */ 2428 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2429 2430 if (flags & GB_LOCK_NOWAIT) 2431 lockflags |= LK_NOWAIT; 2432 2433 error = BUF_TIMELOCK(bp, lockflags, 2434 VI_MTX(vp), "getblk", slpflag, slptimeo); 2435 2436 /* 2437 * If we slept and got the lock we have to restart in case 2438 * the buffer changed identities. 2439 */ 2440 if (error == ENOLCK) 2441 goto loop; 2442 /* We timed out or were interrupted. */ 2443 else if (error) 2444 return (NULL); 2445 2446 /* 2447 * The buffer is locked. B_CACHE is cleared if the buffer is 2448 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2449 * and for a VMIO buffer B_CACHE is adjusted according to the 2450 * backing VM cache. 2451 */ 2452 if (bp->b_flags & B_INVAL) 2453 bp->b_flags &= ~B_CACHE; 2454 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2455 bp->b_flags |= B_CACHE; 2456 bremfree(bp); 2457 2458 /* 2459 * check for size inconsistancies for non-VMIO case. 2460 */ 2461 2462 if (bp->b_bcount != size) { 2463 if ((bp->b_flags & B_VMIO) == 0 || 2464 (size > bp->b_kvasize)) { 2465 if (bp->b_flags & B_DELWRI) { 2466 bp->b_flags |= B_NOCACHE; 2467 BUF_WRITE(bp); 2468 } else { 2469 if ((bp->b_flags & B_VMIO) && 2470 (LIST_FIRST(&bp->b_dep) == NULL)) { 2471 bp->b_flags |= B_RELBUF; 2472 brelse(bp); 2473 } else { 2474 bp->b_flags |= B_NOCACHE; 2475 BUF_WRITE(bp); 2476 } 2477 } 2478 goto loop; 2479 } 2480 } 2481 2482 /* 2483 * If the size is inconsistant in the VMIO case, we can resize 2484 * the buffer. This might lead to B_CACHE getting set or 2485 * cleared. If the size has not changed, B_CACHE remains 2486 * unchanged from its previous state. 2487 */ 2488 2489 if (bp->b_bcount != size) 2490 allocbuf(bp, size); 2491 2492 KASSERT(bp->b_offset != NOOFFSET, 2493 ("getblk: no buffer offset")); 2494 2495 /* 2496 * A buffer with B_DELWRI set and B_CACHE clear must 2497 * be committed before we can return the buffer in 2498 * order to prevent the caller from issuing a read 2499 * ( due to B_CACHE not being set ) and overwriting 2500 * it. 2501 * 2502 * Most callers, including NFS and FFS, need this to 2503 * operate properly either because they assume they 2504 * can issue a read if B_CACHE is not set, or because 2505 * ( for example ) an uncached B_DELWRI might loop due 2506 * to softupdates re-dirtying the buffer. In the latter 2507 * case, B_CACHE is set after the first write completes, 2508 * preventing further loops. 2509 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2510 * above while extending the buffer, we cannot allow the 2511 * buffer to remain with B_CACHE set after the write 2512 * completes or it will represent a corrupt state. To 2513 * deal with this we set B_NOCACHE to scrap the buffer 2514 * after the write. 2515 * 2516 * We might be able to do something fancy, like setting 2517 * B_CACHE in bwrite() except if B_DELWRI is already set, 2518 * so the below call doesn't set B_CACHE, but that gets real 2519 * confusing. This is much easier. 2520 */ 2521 2522 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2523 bp->b_flags |= B_NOCACHE; 2524 BUF_WRITE(bp); 2525 goto loop; 2526 } 2527 2528 splx(s); 2529 bp->b_flags &= ~B_DONE; 2530 } else { 2531 int bsize, maxsize, vmio; 2532 off_t offset; 2533 2534 /* 2535 * Buffer is not in-core, create new buffer. The buffer 2536 * returned by getnewbuf() is locked. Note that the returned 2537 * buffer is also considered valid (not marked B_INVAL). 2538 */ 2539 VI_UNLOCK(vp); 2540 if (vn_isdisk(vp, NULL)) 2541 bsize = DEV_BSIZE; 2542 else if (vp->v_mountedhere) 2543 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2544 else if (vp->v_mount) 2545 bsize = vp->v_mount->mnt_stat.f_iosize; 2546 else 2547 bsize = size; 2548 2549 offset = blkno * bsize; 2550 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2551 (vp->v_vflag & VV_OBJBUF); 2552 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2553 maxsize = imax(maxsize, bsize); 2554 2555 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2556 if (slpflag || slptimeo) { 2557 splx(s); 2558 return NULL; 2559 } 2560 goto loop; 2561 } 2562 2563 /* 2564 * This code is used to make sure that a buffer is not 2565 * created while the getnewbuf routine is blocked. 2566 * This can be a problem whether the vnode is locked or not. 2567 * If the buffer is created out from under us, we have to 2568 * throw away the one we just created. There is now window 2569 * race because we are safely running at splbio() from the 2570 * point of the duplicate buffer creation through to here, 2571 * and we've locked the buffer. 2572 * 2573 * Note: this must occur before we associate the buffer 2574 * with the vp especially considering limitations in 2575 * the splay tree implementation when dealing with duplicate 2576 * lblkno's. 2577 */ 2578 VI_LOCK(vp); 2579 if (gbincore(vp, blkno)) { 2580 VI_UNLOCK(vp); 2581 bp->b_flags |= B_INVAL; 2582 brelse(bp); 2583 goto loop; 2584 } 2585 2586 /* 2587 * Insert the buffer into the hash, so that it can 2588 * be found by incore. 2589 */ 2590 bp->b_blkno = bp->b_lblkno = blkno; 2591 bp->b_offset = offset; 2592 2593 bgetvp(vp, bp); 2594 VI_UNLOCK(vp); 2595 2596 /* 2597 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2598 * buffer size starts out as 0, B_CACHE will be set by 2599 * allocbuf() for the VMIO case prior to it testing the 2600 * backing store for validity. 2601 */ 2602 2603 if (vmio) { 2604 bp->b_flags |= B_VMIO; 2605#if defined(VFS_BIO_DEBUG) 2606 if (vp->v_type != VREG) 2607 printf("getblk: vmioing file type %d???\n", vp->v_type); 2608#endif 2609 VOP_GETVOBJECT(vp, &bp->b_object); 2610 } else { 2611 bp->b_flags &= ~B_VMIO; 2612 bp->b_object = NULL; 2613 } 2614 2615 allocbuf(bp, size); 2616 2617 splx(s); 2618 bp->b_flags &= ~B_DONE; 2619 } 2620 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2621 return (bp); 2622} 2623 2624/* 2625 * Get an empty, disassociated buffer of given size. The buffer is initially 2626 * set to B_INVAL. 2627 */ 2628struct buf * 2629geteblk(int size) 2630{ 2631 struct buf *bp; 2632 int s; 2633 int maxsize; 2634 2635 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2636 2637 s = splbio(); 2638 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2639 continue; 2640 splx(s); 2641 allocbuf(bp, size); 2642 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2643 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2644 return (bp); 2645} 2646 2647 2648/* 2649 * This code constitutes the buffer memory from either anonymous system 2650 * memory (in the case of non-VMIO operations) or from an associated 2651 * VM object (in the case of VMIO operations). This code is able to 2652 * resize a buffer up or down. 2653 * 2654 * Note that this code is tricky, and has many complications to resolve 2655 * deadlock or inconsistant data situations. Tread lightly!!! 2656 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2657 * the caller. Calling this code willy nilly can result in the loss of data. 2658 * 2659 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2660 * B_CACHE for the non-VMIO case. 2661 */ 2662 2663int 2664allocbuf(struct buf *bp, int size) 2665{ 2666 int newbsize, mbsize; 2667 int i; 2668 2669 GIANT_REQUIRED; 2670 2671 if (BUF_REFCNT(bp) == 0) 2672 panic("allocbuf: buffer not busy"); 2673 2674 if (bp->b_kvasize < size) 2675 panic("allocbuf: buffer too small"); 2676 2677 if ((bp->b_flags & B_VMIO) == 0) { 2678 caddr_t origbuf; 2679 int origbufsize; 2680 /* 2681 * Just get anonymous memory from the kernel. Don't 2682 * mess with B_CACHE. 2683 */ 2684 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2685 if (bp->b_flags & B_MALLOC) 2686 newbsize = mbsize; 2687 else 2688 newbsize = round_page(size); 2689 2690 if (newbsize < bp->b_bufsize) { 2691 /* 2692 * malloced buffers are not shrunk 2693 */ 2694 if (bp->b_flags & B_MALLOC) { 2695 if (newbsize) { 2696 bp->b_bcount = size; 2697 } else { 2698 free(bp->b_data, M_BIOBUF); 2699 if (bp->b_bufsize) { 2700 atomic_subtract_int( 2701 &bufmallocspace, 2702 bp->b_bufsize); 2703 bufspacewakeup(); 2704 bp->b_bufsize = 0; 2705 } 2706 bp->b_saveaddr = bp->b_kvabase; 2707 bp->b_data = bp->b_saveaddr; 2708 bp->b_bcount = 0; 2709 bp->b_flags &= ~B_MALLOC; 2710 } 2711 return 1; 2712 } 2713 vm_hold_free_pages( 2714 bp, 2715 (vm_offset_t) bp->b_data + newbsize, 2716 (vm_offset_t) bp->b_data + bp->b_bufsize); 2717 } else if (newbsize > bp->b_bufsize) { 2718 /* 2719 * We only use malloced memory on the first allocation. 2720 * and revert to page-allocated memory when the buffer 2721 * grows. 2722 */ 2723 /* 2724 * There is a potential smp race here that could lead 2725 * to bufmallocspace slightly passing the max. It 2726 * is probably extremely rare and not worth worrying 2727 * over. 2728 */ 2729 if ( (bufmallocspace < maxbufmallocspace) && 2730 (bp->b_bufsize == 0) && 2731 (mbsize <= PAGE_SIZE/2)) { 2732 2733 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2734 bp->b_bufsize = mbsize; 2735 bp->b_bcount = size; 2736 bp->b_flags |= B_MALLOC; 2737 atomic_add_int(&bufmallocspace, mbsize); 2738 return 1; 2739 } 2740 origbuf = NULL; 2741 origbufsize = 0; 2742 /* 2743 * If the buffer is growing on its other-than-first allocation, 2744 * then we revert to the page-allocation scheme. 2745 */ 2746 if (bp->b_flags & B_MALLOC) { 2747 origbuf = bp->b_data; 2748 origbufsize = bp->b_bufsize; 2749 bp->b_data = bp->b_kvabase; 2750 if (bp->b_bufsize) { 2751 atomic_subtract_int(&bufmallocspace, 2752 bp->b_bufsize); 2753 bufspacewakeup(); 2754 bp->b_bufsize = 0; 2755 } 2756 bp->b_flags &= ~B_MALLOC; 2757 newbsize = round_page(newbsize); 2758 } 2759 vm_hold_load_pages( 2760 bp, 2761 (vm_offset_t) bp->b_data + bp->b_bufsize, 2762 (vm_offset_t) bp->b_data + newbsize); 2763 if (origbuf) { 2764 bcopy(origbuf, bp->b_data, origbufsize); 2765 free(origbuf, M_BIOBUF); 2766 } 2767 } 2768 } else { 2769 int desiredpages; 2770 2771 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2772 desiredpages = (size == 0) ? 0 : 2773 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2774 2775 if (bp->b_flags & B_MALLOC) 2776 panic("allocbuf: VMIO buffer can't be malloced"); 2777 /* 2778 * Set B_CACHE initially if buffer is 0 length or will become 2779 * 0-length. 2780 */ 2781 if (size == 0 || bp->b_bufsize == 0) 2782 bp->b_flags |= B_CACHE; 2783 2784 if (newbsize < bp->b_bufsize) { 2785 /* 2786 * DEV_BSIZE aligned new buffer size is less then the 2787 * DEV_BSIZE aligned existing buffer size. Figure out 2788 * if we have to remove any pages. 2789 */ 2790 if (desiredpages < bp->b_npages) { 2791 vm_page_t m; 2792 2793 vm_page_lock_queues(); 2794 for (i = desiredpages; i < bp->b_npages; i++) { 2795 /* 2796 * the page is not freed here -- it 2797 * is the responsibility of 2798 * vnode_pager_setsize 2799 */ 2800 m = bp->b_pages[i]; 2801 KASSERT(m != bogus_page, 2802 ("allocbuf: bogus page found")); 2803 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2804 vm_page_lock_queues(); 2805 2806 bp->b_pages[i] = NULL; 2807 vm_page_unwire(m, 0); 2808 } 2809 vm_page_unlock_queues(); 2810 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2811 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2812 bp->b_npages = desiredpages; 2813 } 2814 } else if (size > bp->b_bcount) { 2815 /* 2816 * We are growing the buffer, possibly in a 2817 * byte-granular fashion. 2818 */ 2819 struct vnode *vp; 2820 vm_object_t obj; 2821 vm_offset_t toff; 2822 vm_offset_t tinc; 2823 2824 /* 2825 * Step 1, bring in the VM pages from the object, 2826 * allocating them if necessary. We must clear 2827 * B_CACHE if these pages are not valid for the 2828 * range covered by the buffer. 2829 */ 2830 2831 vp = bp->b_vp; 2832 obj = bp->b_object; 2833 2834 VM_OBJECT_LOCK(obj); 2835 while (bp->b_npages < desiredpages) { 2836 vm_page_t m; 2837 vm_pindex_t pi; 2838 2839 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2840 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2841 /* 2842 * note: must allocate system pages 2843 * since blocking here could intefere 2844 * with paging I/O, no matter which 2845 * process we are. 2846 */ 2847 m = vm_page_alloc(obj, pi, 2848 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2849 if (m == NULL) { 2850 atomic_add_int(&vm_pageout_deficit, 2851 desiredpages - bp->b_npages); 2852 VM_OBJECT_UNLOCK(obj); 2853 VM_WAIT; 2854 VM_OBJECT_LOCK(obj); 2855 } else { 2856 vm_page_lock_queues(); 2857 vm_page_wakeup(m); 2858 vm_page_unlock_queues(); 2859 bp->b_flags &= ~B_CACHE; 2860 bp->b_pages[bp->b_npages] = m; 2861 ++bp->b_npages; 2862 } 2863 continue; 2864 } 2865 2866 /* 2867 * We found a page. If we have to sleep on it, 2868 * retry because it might have gotten freed out 2869 * from under us. 2870 * 2871 * We can only test PG_BUSY here. Blocking on 2872 * m->busy might lead to a deadlock: 2873 * 2874 * vm_fault->getpages->cluster_read->allocbuf 2875 * 2876 */ 2877 vm_page_lock_queues(); 2878 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2879 continue; 2880 2881 /* 2882 * We have a good page. Should we wakeup the 2883 * page daemon? 2884 */ 2885 if ((curproc != pageproc) && 2886 ((m->queue - m->pc) == PQ_CACHE) && 2887 ((cnt.v_free_count + cnt.v_cache_count) < 2888 (cnt.v_free_min + cnt.v_cache_min))) { 2889 pagedaemon_wakeup(); 2890 } 2891 vm_page_flag_clear(m, PG_ZERO); 2892 vm_page_wire(m); 2893 vm_page_unlock_queues(); 2894 bp->b_pages[bp->b_npages] = m; 2895 ++bp->b_npages; 2896 } 2897 VM_OBJECT_UNLOCK(obj); 2898 2899 /* 2900 * Step 2. We've loaded the pages into the buffer, 2901 * we have to figure out if we can still have B_CACHE 2902 * set. Note that B_CACHE is set according to the 2903 * byte-granular range ( bcount and size ), new the 2904 * aligned range ( newbsize ). 2905 * 2906 * The VM test is against m->valid, which is DEV_BSIZE 2907 * aligned. Needless to say, the validity of the data 2908 * needs to also be DEV_BSIZE aligned. Note that this 2909 * fails with NFS if the server or some other client 2910 * extends the file's EOF. If our buffer is resized, 2911 * B_CACHE may remain set! XXX 2912 */ 2913 2914 toff = bp->b_bcount; 2915 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2916 2917 while ((bp->b_flags & B_CACHE) && toff < size) { 2918 vm_pindex_t pi; 2919 2920 if (tinc > (size - toff)) 2921 tinc = size - toff; 2922 2923 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2924 PAGE_SHIFT; 2925 2926 vfs_buf_test_cache( 2927 bp, 2928 bp->b_offset, 2929 toff, 2930 tinc, 2931 bp->b_pages[pi] 2932 ); 2933 toff += tinc; 2934 tinc = PAGE_SIZE; 2935 } 2936 2937 /* 2938 * Step 3, fixup the KVM pmap. Remember that 2939 * bp->b_data is relative to bp->b_offset, but 2940 * bp->b_offset may be offset into the first page. 2941 */ 2942 2943 bp->b_data = (caddr_t) 2944 trunc_page((vm_offset_t)bp->b_data); 2945 pmap_qenter( 2946 (vm_offset_t)bp->b_data, 2947 bp->b_pages, 2948 bp->b_npages 2949 ); 2950 2951 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2952 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2953 } 2954 } 2955 if (newbsize < bp->b_bufsize) 2956 bufspacewakeup(); 2957 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2958 bp->b_bcount = size; /* requested buffer size */ 2959 return 1; 2960} 2961 2962void 2963biodone(struct bio *bp) 2964{ 2965 mtx_lock(&bdonelock); 2966 bp->bio_flags |= BIO_DONE; 2967 if (bp->bio_done == NULL) 2968 wakeup(bp); 2969 mtx_unlock(&bdonelock); 2970 if (bp->bio_done != NULL) 2971 bp->bio_done(bp); 2972} 2973 2974/* 2975 * Wait for a BIO to finish. 2976 * 2977 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2978 * case is not yet clear. 2979 */ 2980int 2981biowait(struct bio *bp, const char *wchan) 2982{ 2983 2984 mtx_lock(&bdonelock); 2985 while ((bp->bio_flags & BIO_DONE) == 0) 2986 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 2987 mtx_unlock(&bdonelock); 2988 if (bp->bio_error != 0) 2989 return (bp->bio_error); 2990 if (!(bp->bio_flags & BIO_ERROR)) 2991 return (0); 2992 return (EIO); 2993} 2994 2995void 2996biofinish(struct bio *bp, struct devstat *stat, int error) 2997{ 2998 2999 if (error) { 3000 bp->bio_error = error; 3001 bp->bio_flags |= BIO_ERROR; 3002 } 3003 if (stat != NULL) 3004 devstat_end_transaction_bio(stat, bp); 3005 biodone(bp); 3006} 3007 3008/* 3009 * bufwait: 3010 * 3011 * Wait for buffer I/O completion, returning error status. The buffer 3012 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3013 * error and cleared. 3014 */ 3015int 3016bufwait(register struct buf * bp) 3017{ 3018 int s; 3019 3020 s = splbio(); 3021 if (bp->b_iocmd == BIO_READ) 3022 bwait(bp, PRIBIO, "biord"); 3023 else 3024 bwait(bp, PRIBIO, "biowr"); 3025 splx(s); 3026 if (bp->b_flags & B_EINTR) { 3027 bp->b_flags &= ~B_EINTR; 3028 return (EINTR); 3029 } 3030 if (bp->b_ioflags & BIO_ERROR) { 3031 return (bp->b_error ? bp->b_error : EIO); 3032 } else { 3033 return (0); 3034 } 3035} 3036 3037 /* 3038 * Call back function from struct bio back up to struct buf. 3039 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3040 */ 3041void 3042bufdonebio(struct bio *bp) 3043{ 3044 3045 /* Device drivers may or may not hold giant, hold it here. */ 3046 mtx_lock(&Giant); 3047 bufdone(bp->bio_caller2); 3048 mtx_unlock(&Giant); 3049} 3050 3051/* 3052 * bufdone: 3053 * 3054 * Finish I/O on a buffer, optionally calling a completion function. 3055 * This is usually called from an interrupt so process blocking is 3056 * not allowed. 3057 * 3058 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3059 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3060 * assuming B_INVAL is clear. 3061 * 3062 * For the VMIO case, we set B_CACHE if the op was a read and no 3063 * read error occured, or if the op was a write. B_CACHE is never 3064 * set if the buffer is invalid or otherwise uncacheable. 3065 * 3066 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3067 * initiator to leave B_INVAL set to brelse the buffer out of existance 3068 * in the biodone routine. 3069 */ 3070void 3071bufdone(struct buf *bp) 3072{ 3073 int s; 3074 void (*biodone)(struct buf *); 3075 3076 GIANT_REQUIRED; 3077 3078 s = splbio(); 3079 3080 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3081 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3082 3083 bp->b_flags |= B_DONE; 3084 runningbufwakeup(bp); 3085 3086 if (bp->b_iocmd == BIO_DELETE) { 3087 brelse(bp); 3088 splx(s); 3089 return; 3090 } 3091 3092 if (bp->b_iocmd == BIO_WRITE) { 3093 vwakeup(bp); 3094 } 3095 3096 /* call optional completion function if requested */ 3097 if (bp->b_iodone != NULL) { 3098 biodone = bp->b_iodone; 3099 bp->b_iodone = NULL; 3100 (*biodone) (bp); 3101 splx(s); 3102 return; 3103 } 3104 if (LIST_FIRST(&bp->b_dep) != NULL) 3105 buf_complete(bp); 3106 3107 if (bp->b_flags & B_VMIO) { 3108 int i; 3109 vm_ooffset_t foff; 3110 vm_page_t m; 3111 vm_object_t obj; 3112 int iosize; 3113 struct vnode *vp = bp->b_vp; 3114 3115 obj = bp->b_object; 3116 3117#if defined(VFS_BIO_DEBUG) 3118 mp_fixme("usecount and vflag accessed without locks."); 3119 if (vp->v_usecount == 0) { 3120 panic("biodone: zero vnode ref count"); 3121 } 3122 3123 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3124 panic("biodone: vnode is not setup for merged cache"); 3125 } 3126#endif 3127 3128 foff = bp->b_offset; 3129 KASSERT(bp->b_offset != NOOFFSET, 3130 ("biodone: no buffer offset")); 3131 3132 if (obj != NULL) 3133 VM_OBJECT_LOCK(obj); 3134#if defined(VFS_BIO_DEBUG) 3135 if (obj->paging_in_progress < bp->b_npages) { 3136 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3137 obj->paging_in_progress, bp->b_npages); 3138 } 3139#endif 3140 3141 /* 3142 * Set B_CACHE if the op was a normal read and no error 3143 * occured. B_CACHE is set for writes in the b*write() 3144 * routines. 3145 */ 3146 iosize = bp->b_bcount - bp->b_resid; 3147 if (bp->b_iocmd == BIO_READ && 3148 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3149 !(bp->b_ioflags & BIO_ERROR)) { 3150 bp->b_flags |= B_CACHE; 3151 } 3152 vm_page_lock_queues(); 3153 for (i = 0; i < bp->b_npages; i++) { 3154 int bogusflag = 0; 3155 int resid; 3156 3157 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3158 if (resid > iosize) 3159 resid = iosize; 3160 3161 /* 3162 * cleanup bogus pages, restoring the originals 3163 */ 3164 m = bp->b_pages[i]; 3165 if (m == bogus_page) { 3166 bogusflag = 1; 3167 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3168 if (m == NULL) 3169 panic("biodone: page disappeared!"); 3170 bp->b_pages[i] = m; 3171 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3172 } 3173#if defined(VFS_BIO_DEBUG) 3174 if (OFF_TO_IDX(foff) != m->pindex) { 3175 printf( 3176"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3177 (intmax_t)foff, (uintmax_t)m->pindex); 3178 } 3179#endif 3180 3181 /* 3182 * In the write case, the valid and clean bits are 3183 * already changed correctly ( see bdwrite() ), so we 3184 * only need to do this here in the read case. 3185 */ 3186 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3187 vfs_page_set_valid(bp, foff, i, m); 3188 } 3189 vm_page_flag_clear(m, PG_ZERO); 3190 3191 /* 3192 * when debugging new filesystems or buffer I/O methods, this 3193 * is the most common error that pops up. if you see this, you 3194 * have not set the page busy flag correctly!!! 3195 */ 3196 if (m->busy == 0) { 3197 printf("biodone: page busy < 0, " 3198 "pindex: %d, foff: 0x(%x,%x), " 3199 "resid: %d, index: %d\n", 3200 (int) m->pindex, (int)(foff >> 32), 3201 (int) foff & 0xffffffff, resid, i); 3202 if (!vn_isdisk(vp, NULL)) 3203 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3204 bp->b_vp->v_mount->mnt_stat.f_iosize, 3205 (intmax_t) bp->b_lblkno, 3206 bp->b_flags, bp->b_npages); 3207 else 3208 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3209 (intmax_t) bp->b_lblkno, 3210 bp->b_flags, bp->b_npages); 3211 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3212 (u_long)m->valid, (u_long)m->dirty, 3213 m->wire_count); 3214 panic("biodone: page busy < 0\n"); 3215 } 3216 vm_page_io_finish(m); 3217 vm_object_pip_subtract(obj, 1); 3218 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3219 iosize -= resid; 3220 } 3221 vm_page_unlock_queues(); 3222 if (obj != NULL) { 3223 vm_object_pip_wakeupn(obj, 0); 3224 VM_OBJECT_UNLOCK(obj); 3225 } 3226 } 3227 3228 /* 3229 * For asynchronous completions, release the buffer now. The brelse 3230 * will do a wakeup there if necessary - so no need to do a wakeup 3231 * here in the async case. The sync case always needs to do a wakeup. 3232 */ 3233 3234 if (bp->b_flags & B_ASYNC) { 3235 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3236 brelse(bp); 3237 else 3238 bqrelse(bp); 3239 } else { 3240 bdone(bp); 3241 } 3242 splx(s); 3243} 3244 3245/* 3246 * This routine is called in lieu of iodone in the case of 3247 * incomplete I/O. This keeps the busy status for pages 3248 * consistant. 3249 */ 3250void 3251vfs_unbusy_pages(struct buf * bp) 3252{ 3253 int i; 3254 3255 GIANT_REQUIRED; 3256 3257 runningbufwakeup(bp); 3258 if (bp->b_flags & B_VMIO) { 3259 vm_object_t obj; 3260 3261 obj = bp->b_object; 3262 VM_OBJECT_LOCK(obj); 3263 vm_page_lock_queues(); 3264 for (i = 0; i < bp->b_npages; i++) { 3265 vm_page_t m = bp->b_pages[i]; 3266 3267 if (m == bogus_page) { 3268 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3269 if (!m) { 3270 panic("vfs_unbusy_pages: page missing\n"); 3271 } 3272 bp->b_pages[i] = m; 3273 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3274 } 3275 vm_object_pip_subtract(obj, 1); 3276 vm_page_flag_clear(m, PG_ZERO); 3277 vm_page_io_finish(m); 3278 } 3279 vm_page_unlock_queues(); 3280 vm_object_pip_wakeupn(obj, 0); 3281 VM_OBJECT_UNLOCK(obj); 3282 } 3283} 3284 3285/* 3286 * vfs_page_set_valid: 3287 * 3288 * Set the valid bits in a page based on the supplied offset. The 3289 * range is restricted to the buffer's size. 3290 * 3291 * This routine is typically called after a read completes. 3292 */ 3293static void 3294vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3295{ 3296 vm_ooffset_t soff, eoff; 3297 3298 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3299 /* 3300 * Start and end offsets in buffer. eoff - soff may not cross a 3301 * page boundry or cross the end of the buffer. The end of the 3302 * buffer, in this case, is our file EOF, not the allocation size 3303 * of the buffer. 3304 */ 3305 soff = off; 3306 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3307 if (eoff > bp->b_offset + bp->b_bcount) 3308 eoff = bp->b_offset + bp->b_bcount; 3309 3310 /* 3311 * Set valid range. This is typically the entire buffer and thus the 3312 * entire page. 3313 */ 3314 if (eoff > soff) { 3315 vm_page_set_validclean( 3316 m, 3317 (vm_offset_t) (soff & PAGE_MASK), 3318 (vm_offset_t) (eoff - soff) 3319 ); 3320 } 3321} 3322 3323/* 3324 * This routine is called before a device strategy routine. 3325 * It is used to tell the VM system that paging I/O is in 3326 * progress, and treat the pages associated with the buffer 3327 * almost as being PG_BUSY. Also the object paging_in_progress 3328 * flag is handled to make sure that the object doesn't become 3329 * inconsistant. 3330 * 3331 * Since I/O has not been initiated yet, certain buffer flags 3332 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3333 * and should be ignored. 3334 */ 3335void 3336vfs_busy_pages(struct buf * bp, int clear_modify) 3337{ 3338 int i, bogus; 3339 3340 if (bp->b_flags & B_VMIO) { 3341 vm_object_t obj; 3342 vm_ooffset_t foff; 3343 3344 obj = bp->b_object; 3345 foff = bp->b_offset; 3346 KASSERT(bp->b_offset != NOOFFSET, 3347 ("vfs_busy_pages: no buffer offset")); 3348 vfs_setdirty(bp); 3349 if (obj != NULL) 3350 VM_OBJECT_LOCK(obj); 3351retry: 3352 vm_page_lock_queues(); 3353 for (i = 0; i < bp->b_npages; i++) { 3354 vm_page_t m = bp->b_pages[i]; 3355 3356 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3357 goto retry; 3358 } 3359 bogus = 0; 3360 for (i = 0; i < bp->b_npages; i++) { 3361 vm_page_t m = bp->b_pages[i]; 3362 3363 vm_page_flag_clear(m, PG_ZERO); 3364 if ((bp->b_flags & B_CLUSTER) == 0) { 3365 vm_object_pip_add(obj, 1); 3366 vm_page_io_start(m); 3367 } 3368 /* 3369 * When readying a buffer for a read ( i.e 3370 * clear_modify == 0 ), it is important to do 3371 * bogus_page replacement for valid pages in 3372 * partially instantiated buffers. Partially 3373 * instantiated buffers can, in turn, occur when 3374 * reconstituting a buffer from its VM backing store 3375 * base. We only have to do this if B_CACHE is 3376 * clear ( which causes the I/O to occur in the 3377 * first place ). The replacement prevents the read 3378 * I/O from overwriting potentially dirty VM-backed 3379 * pages. XXX bogus page replacement is, uh, bogus. 3380 * It may not work properly with small-block devices. 3381 * We need to find a better way. 3382 */ 3383 pmap_remove_all(m); 3384 if (clear_modify) 3385 vfs_page_set_valid(bp, foff, i, m); 3386 else if (m->valid == VM_PAGE_BITS_ALL && 3387 (bp->b_flags & B_CACHE) == 0) { 3388 bp->b_pages[i] = bogus_page; 3389 bogus++; 3390 } 3391 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3392 } 3393 vm_page_unlock_queues(); 3394 if (obj != NULL) 3395 VM_OBJECT_UNLOCK(obj); 3396 if (bogus) 3397 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3398 } 3399} 3400 3401/* 3402 * Tell the VM system that the pages associated with this buffer 3403 * are clean. This is used for delayed writes where the data is 3404 * going to go to disk eventually without additional VM intevention. 3405 * 3406 * Note that while we only really need to clean through to b_bcount, we 3407 * just go ahead and clean through to b_bufsize. 3408 */ 3409static void 3410vfs_clean_pages(struct buf * bp) 3411{ 3412 int i; 3413 3414 GIANT_REQUIRED; 3415 3416 if (bp->b_flags & B_VMIO) { 3417 vm_ooffset_t foff; 3418 3419 foff = bp->b_offset; 3420 KASSERT(bp->b_offset != NOOFFSET, 3421 ("vfs_clean_pages: no buffer offset")); 3422 vm_page_lock_queues(); 3423 for (i = 0; i < bp->b_npages; i++) { 3424 vm_page_t m = bp->b_pages[i]; 3425 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3426 vm_ooffset_t eoff = noff; 3427 3428 if (eoff > bp->b_offset + bp->b_bufsize) 3429 eoff = bp->b_offset + bp->b_bufsize; 3430 vfs_page_set_valid(bp, foff, i, m); 3431 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3432 foff = noff; 3433 } 3434 vm_page_unlock_queues(); 3435 } 3436} 3437 3438/* 3439 * vfs_bio_set_validclean: 3440 * 3441 * Set the range within the buffer to valid and clean. The range is 3442 * relative to the beginning of the buffer, b_offset. Note that b_offset 3443 * itself may be offset from the beginning of the first page. 3444 * 3445 */ 3446 3447void 3448vfs_bio_set_validclean(struct buf *bp, int base, int size) 3449{ 3450 if (bp->b_flags & B_VMIO) { 3451 int i; 3452 int n; 3453 3454 /* 3455 * Fixup base to be relative to beginning of first page. 3456 * Set initial n to be the maximum number of bytes in the 3457 * first page that can be validated. 3458 */ 3459 3460 base += (bp->b_offset & PAGE_MASK); 3461 n = PAGE_SIZE - (base & PAGE_MASK); 3462 3463 vm_page_lock_queues(); 3464 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3465 vm_page_t m = bp->b_pages[i]; 3466 3467 if (n > size) 3468 n = size; 3469 3470 vm_page_set_validclean(m, base & PAGE_MASK, n); 3471 base += n; 3472 size -= n; 3473 n = PAGE_SIZE; 3474 } 3475 vm_page_unlock_queues(); 3476 } 3477} 3478 3479/* 3480 * vfs_bio_clrbuf: 3481 * 3482 * clear a buffer. This routine essentially fakes an I/O, so we need 3483 * to clear BIO_ERROR and B_INVAL. 3484 * 3485 * Note that while we only theoretically need to clear through b_bcount, 3486 * we go ahead and clear through b_bufsize. 3487 */ 3488 3489void 3490vfs_bio_clrbuf(struct buf *bp) 3491{ 3492 int i, mask = 0; 3493 caddr_t sa, ea; 3494 3495 GIANT_REQUIRED; 3496 3497 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3498 bp->b_flags &= ~B_INVAL; 3499 bp->b_ioflags &= ~BIO_ERROR; 3500 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3501 (bp->b_offset & PAGE_MASK) == 0) { 3502 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3503 if ((bp->b_pages[0]->valid & mask) == mask) { 3504 bp->b_resid = 0; 3505 return; 3506 } 3507 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3508 ((bp->b_pages[0]->valid & mask) == 0)) { 3509 bzero(bp->b_data, bp->b_bufsize); 3510 bp->b_pages[0]->valid |= mask; 3511 bp->b_resid = 0; 3512 return; 3513 } 3514 } 3515 ea = sa = bp->b_data; 3516 for(i=0;i<bp->b_npages;i++,sa=ea) { 3517 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3518 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3519 ea = (caddr_t)(vm_offset_t)ulmin( 3520 (u_long)(vm_offset_t)ea, 3521 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3522 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3523 if ((bp->b_pages[i]->valid & mask) == mask) 3524 continue; 3525 if ((bp->b_pages[i]->valid & mask) == 0) { 3526 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3527 bzero(sa, ea - sa); 3528 } 3529 } else { 3530 for (; sa < ea; sa += DEV_BSIZE, j++) { 3531 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3532 (bp->b_pages[i]->valid & (1<<j)) == 0) 3533 bzero(sa, DEV_BSIZE); 3534 } 3535 } 3536 bp->b_pages[i]->valid |= mask; 3537 vm_page_lock_queues(); 3538 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3539 vm_page_unlock_queues(); 3540 } 3541 bp->b_resid = 0; 3542 } else { 3543 clrbuf(bp); 3544 } 3545} 3546 3547/* 3548 * vm_hold_load_pages and vm_hold_free_pages get pages into 3549 * a buffers address space. The pages are anonymous and are 3550 * not associated with a file object. 3551 */ 3552static void 3553vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3554{ 3555 vm_offset_t pg; 3556 vm_page_t p; 3557 int index; 3558 3559 GIANT_REQUIRED; 3560 3561 to = round_page(to); 3562 from = round_page(from); 3563 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3564 3565 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3566tryagain: 3567 /* 3568 * note: must allocate system pages since blocking here 3569 * could intefere with paging I/O, no matter which 3570 * process we are. 3571 */ 3572 VM_OBJECT_LOCK(kernel_object); 3573 p = vm_page_alloc(kernel_object, 3574 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3575 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3576 VM_OBJECT_UNLOCK(kernel_object); 3577 if (!p) { 3578 atomic_add_int(&vm_pageout_deficit, 3579 (to - pg) >> PAGE_SHIFT); 3580 VM_WAIT; 3581 goto tryagain; 3582 } 3583 vm_page_lock_queues(); 3584 p->valid = VM_PAGE_BITS_ALL; 3585 vm_page_unlock_queues(); 3586 pmap_qenter(pg, &p, 1); 3587 bp->b_pages[index] = p; 3588 vm_page_lock_queues(); 3589 vm_page_wakeup(p); 3590 vm_page_unlock_queues(); 3591 } 3592 bp->b_npages = index; 3593} 3594 3595/* Return pages associated with this buf to the vm system */ 3596static void 3597vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3598{ 3599 vm_offset_t pg; 3600 vm_page_t p; 3601 int index, newnpages; 3602 3603 GIANT_REQUIRED; 3604 3605 from = round_page(from); 3606 to = round_page(to); 3607 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3608 3609 if (bp->b_object != NULL) 3610 VM_OBJECT_LOCK(bp->b_object); 3611 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3612 p = bp->b_pages[index]; 3613 if (p && (index < bp->b_npages)) { 3614 if (p->busy) { 3615 printf( 3616 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3617 (intmax_t)bp->b_blkno, 3618 (intmax_t)bp->b_lblkno); 3619 } 3620 bp->b_pages[index] = NULL; 3621 pmap_qremove(pg, 1); 3622 vm_page_lock_queues(); 3623 vm_page_busy(p); 3624 vm_page_unwire(p, 0); 3625 vm_page_free(p); 3626 vm_page_unlock_queues(); 3627 } 3628 } 3629 if (bp->b_object != NULL) 3630 VM_OBJECT_UNLOCK(bp->b_object); 3631 bp->b_npages = newnpages; 3632} 3633 3634/* 3635 * Map an IO request into kernel virtual address space. 3636 * 3637 * All requests are (re)mapped into kernel VA space. 3638 * Notice that we use b_bufsize for the size of the buffer 3639 * to be mapped. b_bcount might be modified by the driver. 3640 * 3641 * Note that even if the caller determines that the address space should 3642 * be valid, a race or a smaller-file mapped into a larger space may 3643 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3644 * check the return value. 3645 */ 3646int 3647vmapbuf(struct buf *bp) 3648{ 3649 caddr_t addr, kva; 3650 vm_paddr_t pa; 3651 int pidx, i; 3652 struct vm_page *m; 3653 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3654 3655 GIANT_REQUIRED; 3656 3657 if ((bp->b_flags & B_PHYS) == 0) 3658 panic("vmapbuf"); 3659 if (bp->b_bufsize < 0) 3660 return (-1); 3661 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3662 addr < bp->b_data + bp->b_bufsize; 3663 addr += PAGE_SIZE, pidx++) { 3664 /* 3665 * Do the vm_fault if needed; do the copy-on-write thing 3666 * when reading stuff off device into memory. 3667 * 3668 * NOTE! Must use pmap_extract() because addr may be in 3669 * the userland address space, and kextract is only guarenteed 3670 * to work for the kernland address space (see: sparc64 port). 3671 */ 3672retry: 3673 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3674 (bp->b_iocmd == BIO_READ) ? 3675 (VM_PROT_READ|VM_PROT_WRITE) : VM_PROT_READ); 3676 if (i < 0) { 3677 vm_page_lock_queues(); 3678 for (i = 0; i < pidx; ++i) { 3679 vm_page_unhold(bp->b_pages[i]); 3680 bp->b_pages[i] = NULL; 3681 } 3682 vm_page_unlock_queues(); 3683 return(-1); 3684 } 3685 pa = pmap_extract(pmap, (vm_offset_t)addr); 3686 if (pa == 0) { 3687 printf("vmapbuf: warning, race against user address during I/O"); 3688 goto retry; 3689 } 3690 m = PHYS_TO_VM_PAGE(pa); 3691 vm_page_lock_queues(); 3692 vm_page_hold(m); 3693 vm_page_unlock_queues(); 3694 bp->b_pages[pidx] = m; 3695 } 3696 if (pidx > btoc(MAXPHYS)) 3697 panic("vmapbuf: mapped more than MAXPHYS"); 3698 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3699 3700 kva = bp->b_saveaddr; 3701 bp->b_npages = pidx; 3702 bp->b_saveaddr = bp->b_data; 3703 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3704 return(0); 3705} 3706 3707/* 3708 * Free the io map PTEs associated with this IO operation. 3709 * We also invalidate the TLB entries and restore the original b_addr. 3710 */ 3711void 3712vunmapbuf(struct buf *bp) 3713{ 3714 int pidx; 3715 int npages; 3716 3717 GIANT_REQUIRED; 3718 3719 if ((bp->b_flags & B_PHYS) == 0) 3720 panic("vunmapbuf"); 3721 3722 npages = bp->b_npages; 3723 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3724 npages); 3725 vm_page_lock_queues(); 3726 for (pidx = 0; pidx < npages; pidx++) 3727 vm_page_unhold(bp->b_pages[pidx]); 3728 vm_page_unlock_queues(); 3729 3730 bp->b_data = bp->b_saveaddr; 3731} 3732 3733void 3734bdone(struct buf *bp) 3735{ 3736 mtx_lock(&bdonelock); 3737 bp->b_flags |= B_DONE; 3738 wakeup(bp); 3739 mtx_unlock(&bdonelock); 3740} 3741 3742void 3743bwait(struct buf *bp, u_char pri, const char *wchan) 3744{ 3745 mtx_lock(&bdonelock); 3746 while ((bp->b_flags & B_DONE) == 0) 3747 msleep(bp, &bdonelock, pri, wchan, 0); 3748 mtx_unlock(&bdonelock); 3749} 3750 3751#include "opt_ddb.h" 3752#ifdef DDB 3753#include <ddb/ddb.h> 3754 3755/* DDB command to show buffer data */ 3756DB_SHOW_COMMAND(buffer, db_show_buffer) 3757{ 3758 /* get args */ 3759 struct buf *bp = (struct buf *)addr; 3760 3761 if (!have_addr) { 3762 db_printf("usage: show buffer <addr>\n"); 3763 return; 3764 } 3765 3766 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3767 db_printf( 3768 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3769 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3770 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3771 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3772 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3773 if (bp->b_npages) { 3774 int i; 3775 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3776 for (i = 0; i < bp->b_npages; i++) { 3777 vm_page_t m; 3778 m = bp->b_pages[i]; 3779 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3780 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3781 if ((i + 1) < bp->b_npages) 3782 db_printf(","); 3783 } 3784 db_printf("\n"); 3785 } 3786} 3787#endif /* DDB */ 3788