vfs_bio.c revision 119536
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 119536 2003-08-28 19:58:11Z marcel $"); 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 VI_LOCK(bp->b_vp); 1257 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1258 bp->b_flags |= B_RELBUF; 1259 VI_UNLOCK(bp->b_vp); 1260 } 1261 1262 /* 1263 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1264 * constituted, not even NFS buffers now. Two flags effect this. If 1265 * B_INVAL, the struct buf is invalidated but the VM object is kept 1266 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1267 * 1268 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1269 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1270 * buffer is also B_INVAL because it hits the re-dirtying code above. 1271 * 1272 * Normally we can do this whether a buffer is B_DELWRI or not. If 1273 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1274 * the commit state and we cannot afford to lose the buffer. If the 1275 * buffer has a background write in progress, we need to keep it 1276 * around to prevent it from being reconstituted and starting a second 1277 * background write. 1278 */ 1279 if ((bp->b_flags & B_VMIO) 1280 && !(bp->b_vp->v_mount != NULL && 1281 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1282 !vn_isdisk(bp->b_vp, NULL) && 1283 (bp->b_flags & B_DELWRI)) 1284 ) { 1285 1286 int i, j, resid; 1287 vm_page_t m; 1288 off_t foff; 1289 vm_pindex_t poff; 1290 vm_object_t obj; 1291 struct vnode *vp; 1292 1293 vp = bp->b_vp; 1294 obj = bp->b_object; 1295 1296 /* 1297 * Get the base offset and length of the buffer. Note that 1298 * in the VMIO case if the buffer block size is not 1299 * page-aligned then b_data pointer may not be page-aligned. 1300 * But our b_pages[] array *IS* page aligned. 1301 * 1302 * block sizes less then DEV_BSIZE (usually 512) are not 1303 * supported due to the page granularity bits (m->valid, 1304 * m->dirty, etc...). 1305 * 1306 * See man buf(9) for more information 1307 */ 1308 resid = bp->b_bufsize; 1309 foff = bp->b_offset; 1310 if (obj != NULL) 1311 VM_OBJECT_LOCK(obj); 1312 for (i = 0; i < bp->b_npages; i++) { 1313 int had_bogus = 0; 1314 1315 m = bp->b_pages[i]; 1316 vm_page_lock_queues(); 1317 vm_page_flag_clear(m, PG_ZERO); 1318 vm_page_unlock_queues(); 1319 1320 /* 1321 * If we hit a bogus page, fixup *all* the bogus pages 1322 * now. 1323 */ 1324 if (m == bogus_page) { 1325 poff = OFF_TO_IDX(bp->b_offset); 1326 had_bogus = 1; 1327 1328 for (j = i; j < bp->b_npages; j++) { 1329 vm_page_t mtmp; 1330 mtmp = bp->b_pages[j]; 1331 if (mtmp == bogus_page) { 1332 mtmp = vm_page_lookup(obj, poff + j); 1333 if (!mtmp) { 1334 panic("brelse: page missing\n"); 1335 } 1336 bp->b_pages[j] = mtmp; 1337 } 1338 } 1339 1340 if ((bp->b_flags & B_INVAL) == 0) { 1341 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1342 } 1343 m = bp->b_pages[i]; 1344 } 1345 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1346 int poffset = foff & PAGE_MASK; 1347 int presid = resid > (PAGE_SIZE - poffset) ? 1348 (PAGE_SIZE - poffset) : resid; 1349 1350 KASSERT(presid >= 0, ("brelse: extra page")); 1351 vm_page_lock_queues(); 1352 vm_page_set_invalid(m, poffset, presid); 1353 vm_page_unlock_queues(); 1354 if (had_bogus) 1355 printf("avoided corruption bug in bogus_page/brelse code\n"); 1356 } 1357 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1358 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1359 } 1360 if (obj != NULL) 1361 VM_OBJECT_UNLOCK(obj); 1362 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1363 vfs_vmio_release(bp); 1364 1365 } else if (bp->b_flags & B_VMIO) { 1366 1367 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1368 vfs_vmio_release(bp); 1369 } 1370 1371 } 1372 1373 if (bp->b_qindex != QUEUE_NONE) 1374 panic("brelse: free buffer onto another queue???"); 1375 if (BUF_REFCNT(bp) > 1) { 1376 /* do not release to free list */ 1377 BUF_UNLOCK(bp); 1378 splx(s); 1379 return; 1380 } 1381 1382 /* enqueue */ 1383 mtx_lock(&bqlock); 1384 1385 /* buffers with no memory */ 1386 if (bp->b_bufsize == 0) { 1387 bp->b_flags |= B_INVAL; 1388 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1389 if (bp->b_vflags & BV_BKGRDINPROG) 1390 panic("losing buffer 1"); 1391 if (bp->b_kvasize) { 1392 bp->b_qindex = QUEUE_EMPTYKVA; 1393 } else { 1394 bp->b_qindex = QUEUE_EMPTY; 1395 } 1396 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1397 bp->b_dev = NODEV; 1398 /* buffers with junk contents */ 1399 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1400 (bp->b_ioflags & BIO_ERROR)) { 1401 bp->b_flags |= B_INVAL; 1402 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1403 if (bp->b_vflags & BV_BKGRDINPROG) 1404 panic("losing buffer 2"); 1405 bp->b_qindex = QUEUE_CLEAN; 1406 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1407 bp->b_dev = NODEV; 1408 /* remaining buffers */ 1409 } else { 1410 if (bp->b_flags & B_DELWRI) 1411 bp->b_qindex = QUEUE_DIRTY; 1412 else 1413 bp->b_qindex = QUEUE_CLEAN; 1414 if (bp->b_flags & B_AGE) 1415 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1416 else 1417 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1418 } 1419 mtx_unlock(&bqlock); 1420 1421 /* 1422 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1423 * placed the buffer on the correct queue. We must also disassociate 1424 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1425 * find it. 1426 */ 1427 if (bp->b_flags & B_INVAL) { 1428 if (bp->b_flags & B_DELWRI) 1429 bundirty(bp); 1430 if (bp->b_vp) 1431 brelvp(bp); 1432 } 1433 1434 /* 1435 * Fixup numfreebuffers count. The bp is on an appropriate queue 1436 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1437 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1438 * if B_INVAL is set ). 1439 */ 1440 1441 if (!(bp->b_flags & B_DELWRI)) 1442 bufcountwakeup(); 1443 1444 /* 1445 * Something we can maybe free or reuse 1446 */ 1447 if (bp->b_bufsize || bp->b_kvasize) 1448 bufspacewakeup(); 1449 1450 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1451 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1452 panic("brelse: not dirty"); 1453 /* unlock */ 1454 BUF_UNLOCK(bp); 1455 splx(s); 1456} 1457 1458/* 1459 * Release a buffer back to the appropriate queue but do not try to free 1460 * it. The buffer is expected to be used again soon. 1461 * 1462 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1463 * biodone() to requeue an async I/O on completion. It is also used when 1464 * known good buffers need to be requeued but we think we may need the data 1465 * again soon. 1466 * 1467 * XXX we should be able to leave the B_RELBUF hint set on completion. 1468 */ 1469void 1470bqrelse(struct buf * bp) 1471{ 1472 int s; 1473 1474 s = splbio(); 1475 1476 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1477 1478 if (bp->b_qindex != QUEUE_NONE) 1479 panic("bqrelse: free buffer onto another queue???"); 1480 if (BUF_REFCNT(bp) > 1) { 1481 /* do not release to free list */ 1482 BUF_UNLOCK(bp); 1483 splx(s); 1484 return; 1485 } 1486 mtx_lock(&bqlock); 1487 /* buffers with stale but valid contents */ 1488 if (bp->b_flags & B_DELWRI) { 1489 bp->b_qindex = QUEUE_DIRTY; 1490 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1491 } else { 1492 /* 1493 * XXX This lock may not be necessary since BKGRDINPROG 1494 * cannot be set while we hold the buf lock, it can only be 1495 * cleared if it is already pending. 1496 */ 1497 VI_LOCK(bp->b_vp); 1498 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1499 VI_UNLOCK(bp->b_vp); 1500 bp->b_qindex = QUEUE_CLEAN; 1501 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1502 b_freelist); 1503 } else { 1504 /* 1505 * We are too low on memory, we have to try to free 1506 * the buffer (most importantly: the wired pages 1507 * making up its backing store) *now*. 1508 */ 1509 VI_UNLOCK(bp->b_vp); 1510 mtx_unlock(&bqlock); 1511 splx(s); 1512 brelse(bp); 1513 return; 1514 } 1515 } 1516 mtx_unlock(&bqlock); 1517 1518 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1519 bufcountwakeup(); 1520 1521 /* 1522 * Something we can maybe free or reuse. 1523 */ 1524 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1525 bufspacewakeup(); 1526 1527 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1528 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1529 panic("bqrelse: not dirty"); 1530 /* unlock */ 1531 BUF_UNLOCK(bp); 1532 splx(s); 1533} 1534 1535/* Give pages used by the bp back to the VM system (where possible) */ 1536static void 1537vfs_vmio_release(bp) 1538 struct buf *bp; 1539{ 1540 int i; 1541 vm_page_t m; 1542 1543 GIANT_REQUIRED; 1544 if (bp->b_object != NULL) 1545 VM_OBJECT_LOCK(bp->b_object); 1546 vm_page_lock_queues(); 1547 for (i = 0; i < bp->b_npages; i++) { 1548 m = bp->b_pages[i]; 1549 bp->b_pages[i] = NULL; 1550 /* 1551 * In order to keep page LRU ordering consistent, put 1552 * everything on the inactive queue. 1553 */ 1554 vm_page_unwire(m, 0); 1555 /* 1556 * We don't mess with busy pages, it is 1557 * the responsibility of the process that 1558 * busied the pages to deal with them. 1559 */ 1560 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1561 continue; 1562 1563 if (m->wire_count == 0) { 1564 vm_page_flag_clear(m, PG_ZERO); 1565 /* 1566 * Might as well free the page if we can and it has 1567 * no valid data. We also free the page if the 1568 * buffer was used for direct I/O 1569 */ 1570 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1571 m->hold_count == 0) { 1572 vm_page_busy(m); 1573 pmap_remove_all(m); 1574 vm_page_free(m); 1575 } else if (bp->b_flags & B_DIRECT) { 1576 vm_page_try_to_free(m); 1577 } else if (vm_page_count_severe()) { 1578 vm_page_try_to_cache(m); 1579 } 1580 } 1581 } 1582 vm_page_unlock_queues(); 1583 if (bp->b_object != NULL) 1584 VM_OBJECT_UNLOCK(bp->b_object); 1585 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1586 1587 if (bp->b_bufsize) { 1588 bufspacewakeup(); 1589 bp->b_bufsize = 0; 1590 } 1591 bp->b_npages = 0; 1592 bp->b_flags &= ~B_VMIO; 1593 if (bp->b_vp) 1594 brelvp(bp); 1595} 1596 1597/* 1598 * Check to see if a block at a particular lbn is available for a clustered 1599 * write. 1600 */ 1601static int 1602vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1603{ 1604 struct buf *bpa; 1605 int match; 1606 1607 match = 0; 1608 1609 /* If the buf isn't in core skip it */ 1610 if ((bpa = gbincore(vp, lblkno)) == NULL) 1611 return (0); 1612 1613 /* If the buf is busy we don't want to wait for it */ 1614 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1615 return (0); 1616 1617 /* Only cluster with valid clusterable delayed write buffers */ 1618 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1619 (B_DELWRI | B_CLUSTEROK)) 1620 goto done; 1621 1622 if (bpa->b_bufsize != size) 1623 goto done; 1624 1625 /* 1626 * Check to see if it is in the expected place on disk and that the 1627 * block has been mapped. 1628 */ 1629 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1630 match = 1; 1631done: 1632 BUF_UNLOCK(bpa); 1633 return (match); 1634} 1635 1636/* 1637 * vfs_bio_awrite: 1638 * 1639 * Implement clustered async writes for clearing out B_DELWRI buffers. 1640 * This is much better then the old way of writing only one buffer at 1641 * a time. Note that we may not be presented with the buffers in the 1642 * correct order, so we search for the cluster in both directions. 1643 */ 1644int 1645vfs_bio_awrite(struct buf * bp) 1646{ 1647 int i; 1648 int j; 1649 daddr_t lblkno = bp->b_lblkno; 1650 struct vnode *vp = bp->b_vp; 1651 int s; 1652 int ncl; 1653 int nwritten; 1654 int size; 1655 int maxcl; 1656 1657 s = splbio(); 1658 /* 1659 * right now we support clustered writing only to regular files. If 1660 * we find a clusterable block we could be in the middle of a cluster 1661 * rather then at the beginning. 1662 */ 1663 if ((vp->v_type == VREG) && 1664 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1665 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1666 1667 size = vp->v_mount->mnt_stat.f_iosize; 1668 maxcl = MAXPHYS / size; 1669 1670 VI_LOCK(vp); 1671 for (i = 1; i < maxcl; i++) 1672 if (vfs_bio_clcheck(vp, size, lblkno + i, 1673 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1674 break; 1675 1676 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1677 if (vfs_bio_clcheck(vp, size, lblkno - j, 1678 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1679 break; 1680 1681 VI_UNLOCK(vp); 1682 --j; 1683 ncl = i + j; 1684 /* 1685 * this is a possible cluster write 1686 */ 1687 if (ncl != 1) { 1688 BUF_UNLOCK(bp); 1689 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1690 splx(s); 1691 return nwritten; 1692 } 1693 } 1694 1695 bremfree(bp); 1696 bp->b_flags |= B_ASYNC; 1697 1698 splx(s); 1699 /* 1700 * default (old) behavior, writing out only one block 1701 * 1702 * XXX returns b_bufsize instead of b_bcount for nwritten? 1703 */ 1704 nwritten = bp->b_bufsize; 1705 (void) BUF_WRITE(bp); 1706 1707 return nwritten; 1708} 1709 1710/* 1711 * getnewbuf: 1712 * 1713 * Find and initialize a new buffer header, freeing up existing buffers 1714 * in the bufqueues as necessary. The new buffer is returned locked. 1715 * 1716 * Important: B_INVAL is not set. If the caller wishes to throw the 1717 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1718 * 1719 * We block if: 1720 * We have insufficient buffer headers 1721 * We have insufficient buffer space 1722 * buffer_map is too fragmented ( space reservation fails ) 1723 * If we have to flush dirty buffers ( but we try to avoid this ) 1724 * 1725 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1726 * Instead we ask the buf daemon to do it for us. We attempt to 1727 * avoid piecemeal wakeups of the pageout daemon. 1728 */ 1729 1730static struct buf * 1731getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1732{ 1733 struct buf *bp; 1734 struct buf *nbp; 1735 int defrag = 0; 1736 int nqindex; 1737 static int flushingbufs; 1738 1739 GIANT_REQUIRED; 1740 1741 /* 1742 * We can't afford to block since we might be holding a vnode lock, 1743 * which may prevent system daemons from running. We deal with 1744 * low-memory situations by proactively returning memory and running 1745 * async I/O rather then sync I/O. 1746 */ 1747 1748 atomic_add_int(&getnewbufcalls, 1); 1749 atomic_subtract_int(&getnewbufrestarts, 1); 1750restart: 1751 atomic_add_int(&getnewbufrestarts, 1); 1752 1753 /* 1754 * Setup for scan. If we do not have enough free buffers, 1755 * we setup a degenerate case that immediately fails. Note 1756 * that if we are specially marked process, we are allowed to 1757 * dip into our reserves. 1758 * 1759 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1760 * 1761 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1762 * However, there are a number of cases (defragging, reusing, ...) 1763 * where we cannot backup. 1764 */ 1765 mtx_lock(&bqlock); 1766 nqindex = QUEUE_EMPTYKVA; 1767 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1768 1769 if (nbp == NULL) { 1770 /* 1771 * If no EMPTYKVA buffers and we are either 1772 * defragging or reusing, locate a CLEAN buffer 1773 * to free or reuse. If bufspace useage is low 1774 * skip this step so we can allocate a new buffer. 1775 */ 1776 if (defrag || bufspace >= lobufspace) { 1777 nqindex = QUEUE_CLEAN; 1778 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1779 } 1780 1781 /* 1782 * If we could not find or were not allowed to reuse a 1783 * CLEAN buffer, check to see if it is ok to use an EMPTY 1784 * buffer. We can only use an EMPTY buffer if allocating 1785 * its KVA would not otherwise run us out of buffer space. 1786 */ 1787 if (nbp == NULL && defrag == 0 && 1788 bufspace + maxsize < hibufspace) { 1789 nqindex = QUEUE_EMPTY; 1790 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1791 } 1792 } 1793 1794 /* 1795 * Run scan, possibly freeing data and/or kva mappings on the fly 1796 * depending. 1797 */ 1798 1799 while ((bp = nbp) != NULL) { 1800 int qindex = nqindex; 1801 1802 /* 1803 * Calculate next bp ( we can only use it if we do not block 1804 * or do other fancy things ). 1805 */ 1806 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1807 switch(qindex) { 1808 case QUEUE_EMPTY: 1809 nqindex = QUEUE_EMPTYKVA; 1810 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1811 break; 1812 /* FALLTHROUGH */ 1813 case QUEUE_EMPTYKVA: 1814 nqindex = QUEUE_CLEAN; 1815 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1816 break; 1817 /* FALLTHROUGH */ 1818 case QUEUE_CLEAN: 1819 /* 1820 * nbp is NULL. 1821 */ 1822 break; 1823 } 1824 } 1825 if (bp->b_vp) { 1826 VI_LOCK(bp->b_vp); 1827 if (bp->b_vflags & BV_BKGRDINPROG) { 1828 VI_UNLOCK(bp->b_vp); 1829 continue; 1830 } 1831 VI_UNLOCK(bp->b_vp); 1832 } 1833 1834 /* 1835 * Sanity Checks 1836 */ 1837 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1838 1839 /* 1840 * Note: we no longer distinguish between VMIO and non-VMIO 1841 * buffers. 1842 */ 1843 1844 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1845 1846 /* 1847 * If we are defragging then we need a buffer with 1848 * b_kvasize != 0. XXX this situation should no longer 1849 * occur, if defrag is non-zero the buffer's b_kvasize 1850 * should also be non-zero at this point. XXX 1851 */ 1852 if (defrag && bp->b_kvasize == 0) { 1853 printf("Warning: defrag empty buffer %p\n", bp); 1854 continue; 1855 } 1856 1857 /* 1858 * Start freeing the bp. This is somewhat involved. nbp 1859 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1860 */ 1861 1862 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1863 panic("getnewbuf: locked buf"); 1864 bremfreel(bp); 1865 mtx_unlock(&bqlock); 1866 1867 if (qindex == QUEUE_CLEAN) { 1868 if (bp->b_flags & B_VMIO) { 1869 bp->b_flags &= ~B_ASYNC; 1870 vfs_vmio_release(bp); 1871 } 1872 if (bp->b_vp) 1873 brelvp(bp); 1874 } 1875 1876 /* 1877 * NOTE: nbp is now entirely invalid. We can only restart 1878 * the scan from this point on. 1879 * 1880 * Get the rest of the buffer freed up. b_kva* is still 1881 * valid after this operation. 1882 */ 1883 1884 if (bp->b_rcred != NOCRED) { 1885 crfree(bp->b_rcred); 1886 bp->b_rcred = NOCRED; 1887 } 1888 if (bp->b_wcred != NOCRED) { 1889 crfree(bp->b_wcred); 1890 bp->b_wcred = NOCRED; 1891 } 1892 if (LIST_FIRST(&bp->b_dep) != NULL) 1893 buf_deallocate(bp); 1894 if (bp->b_vflags & BV_BKGRDINPROG) 1895 panic("losing buffer 3"); 1896 1897 if (bp->b_bufsize) 1898 allocbuf(bp, 0); 1899 1900 bp->b_flags = 0; 1901 bp->b_ioflags = 0; 1902 bp->b_xflags = 0; 1903 bp->b_vflags = 0; 1904 bp->b_dev = NODEV; 1905 bp->b_vp = NULL; 1906 bp->b_blkno = bp->b_lblkno = 0; 1907 bp->b_offset = NOOFFSET; 1908 bp->b_iodone = 0; 1909 bp->b_error = 0; 1910 bp->b_resid = 0; 1911 bp->b_bcount = 0; 1912 bp->b_npages = 0; 1913 bp->b_dirtyoff = bp->b_dirtyend = 0; 1914 bp->b_magic = B_MAGIC_BIO; 1915 bp->b_op = &buf_ops_bio; 1916 bp->b_object = NULL; 1917 1918 LIST_INIT(&bp->b_dep); 1919 1920 /* 1921 * If we are defragging then free the buffer. 1922 */ 1923 if (defrag) { 1924 bp->b_flags |= B_INVAL; 1925 bfreekva(bp); 1926 brelse(bp); 1927 defrag = 0; 1928 goto restart; 1929 } 1930 1931 /* 1932 * If we are overcomitted then recover the buffer and its 1933 * KVM space. This occurs in rare situations when multiple 1934 * processes are blocked in getnewbuf() or allocbuf(). 1935 */ 1936 if (bufspace >= hibufspace) 1937 flushingbufs = 1; 1938 if (flushingbufs && bp->b_kvasize != 0) { 1939 bp->b_flags |= B_INVAL; 1940 bfreekva(bp); 1941 brelse(bp); 1942 goto restart; 1943 } 1944 if (bufspace < lobufspace) 1945 flushingbufs = 0; 1946 break; 1947 } 1948 1949 /* 1950 * If we exhausted our list, sleep as appropriate. We may have to 1951 * wakeup various daemons and write out some dirty buffers. 1952 * 1953 * Generally we are sleeping due to insufficient buffer space. 1954 */ 1955 1956 if (bp == NULL) { 1957 int flags; 1958 char *waitmsg; 1959 1960 mtx_unlock(&bqlock); 1961 if (defrag) { 1962 flags = VFS_BIO_NEED_BUFSPACE; 1963 waitmsg = "nbufkv"; 1964 } else if (bufspace >= hibufspace) { 1965 waitmsg = "nbufbs"; 1966 flags = VFS_BIO_NEED_BUFSPACE; 1967 } else { 1968 waitmsg = "newbuf"; 1969 flags = VFS_BIO_NEED_ANY; 1970 } 1971 1972 bd_speedup(); /* heeeelp */ 1973 1974 mtx_lock(&nblock); 1975 needsbuffer |= flags; 1976 while (needsbuffer & flags) { 1977 if (msleep(&needsbuffer, &nblock, 1978 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1979 mtx_unlock(&nblock); 1980 return (NULL); 1981 } 1982 } 1983 mtx_unlock(&nblock); 1984 } else { 1985 /* 1986 * We finally have a valid bp. We aren't quite out of the 1987 * woods, we still have to reserve kva space. In order 1988 * to keep fragmentation sane we only allocate kva in 1989 * BKVASIZE chunks. 1990 */ 1991 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1992 1993 if (maxsize != bp->b_kvasize) { 1994 vm_offset_t addr = 0; 1995 1996 bfreekva(bp); 1997 1998 if (vm_map_findspace(buffer_map, 1999 vm_map_min(buffer_map), maxsize, &addr)) { 2000 /* 2001 * Uh oh. Buffer map is to fragmented. We 2002 * must defragment the map. 2003 */ 2004 atomic_add_int(&bufdefragcnt, 1); 2005 defrag = 1; 2006 bp->b_flags |= B_INVAL; 2007 brelse(bp); 2008 goto restart; 2009 } 2010 if (addr) { 2011 vm_map_insert(buffer_map, NULL, 0, 2012 addr, addr + maxsize, 2013 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2014 2015 bp->b_kvabase = (caddr_t) addr; 2016 bp->b_kvasize = maxsize; 2017 atomic_add_int(&bufspace, bp->b_kvasize); 2018 atomic_add_int(&bufreusecnt, 1); 2019 } 2020 } 2021 bp->b_saveaddr = bp->b_kvabase; 2022 bp->b_data = bp->b_saveaddr; 2023 } 2024 return(bp); 2025} 2026 2027/* 2028 * buf_daemon: 2029 * 2030 * buffer flushing daemon. Buffers are normally flushed by the 2031 * update daemon but if it cannot keep up this process starts to 2032 * take the load in an attempt to prevent getnewbuf() from blocking. 2033 */ 2034 2035static struct proc *bufdaemonproc; 2036 2037static struct kproc_desc buf_kp = { 2038 "bufdaemon", 2039 buf_daemon, 2040 &bufdaemonproc 2041}; 2042SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2043 2044static void 2045buf_daemon() 2046{ 2047 int s; 2048 2049 mtx_lock(&Giant); 2050 2051 /* 2052 * This process needs to be suspended prior to shutdown sync. 2053 */ 2054 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2055 SHUTDOWN_PRI_LAST); 2056 2057 /* 2058 * This process is allowed to take the buffer cache to the limit 2059 */ 2060 s = splbio(); 2061 mtx_lock(&bdlock); 2062 2063 for (;;) { 2064 bd_request = 0; 2065 mtx_unlock(&bdlock); 2066 2067 kthread_suspend_check(bufdaemonproc); 2068 2069 /* 2070 * Do the flush. Limit the amount of in-transit I/O we 2071 * allow to build up, otherwise we would completely saturate 2072 * the I/O system. Wakeup any waiting processes before we 2073 * normally would so they can run in parallel with our drain. 2074 */ 2075 while (numdirtybuffers > lodirtybuffers) { 2076 if (flushbufqueues(0) == 0) { 2077 /* 2078 * Could not find any buffers without rollback 2079 * dependencies, so just write the first one 2080 * in the hopes of eventually making progress. 2081 */ 2082 flushbufqueues(1); 2083 break; 2084 } 2085 waitrunningbufspace(); 2086 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2087 } 2088 2089 /* 2090 * Only clear bd_request if we have reached our low water 2091 * mark. The buf_daemon normally waits 1 second and 2092 * then incrementally flushes any dirty buffers that have 2093 * built up, within reason. 2094 * 2095 * If we were unable to hit our low water mark and couldn't 2096 * find any flushable buffers, we sleep half a second. 2097 * Otherwise we loop immediately. 2098 */ 2099 mtx_lock(&bdlock); 2100 if (numdirtybuffers <= lodirtybuffers) { 2101 /* 2102 * We reached our low water mark, reset the 2103 * request and sleep until we are needed again. 2104 * The sleep is just so the suspend code works. 2105 */ 2106 bd_request = 0; 2107 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2108 } else { 2109 /* 2110 * We couldn't find any flushable dirty buffers but 2111 * still have too many dirty buffers, we 2112 * have to sleep and try again. (rare) 2113 */ 2114 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2115 } 2116 } 2117} 2118 2119/* 2120 * flushbufqueues: 2121 * 2122 * Try to flush a buffer in the dirty queue. We must be careful to 2123 * free up B_INVAL buffers instead of write them, which NFS is 2124 * particularly sensitive to. 2125 */ 2126int flushwithdeps = 0; 2127SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2128 0, "Number of buffers flushed with dependecies that require rollbacks"); 2129static int 2130flushbufqueues(int flushdeps) 2131{ 2132 struct thread *td = curthread; 2133 struct vnode *vp; 2134 struct buf *bp; 2135 int hasdeps; 2136 2137 mtx_lock(&bqlock); 2138 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2139 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2140 continue; 2141 KASSERT((bp->b_flags & B_DELWRI), 2142 ("unexpected clean buffer %p", bp)); 2143 VI_LOCK(bp->b_vp); 2144 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2145 VI_UNLOCK(bp->b_vp); 2146 BUF_UNLOCK(bp); 2147 continue; 2148 } 2149 VI_UNLOCK(bp->b_vp); 2150 if (bp->b_flags & B_INVAL) { 2151 bremfreel(bp); 2152 mtx_unlock(&bqlock); 2153 brelse(bp); 2154 return (1); 2155 } 2156 2157 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2158 if (flushdeps == 0) { 2159 BUF_UNLOCK(bp); 2160 continue; 2161 } 2162 hasdeps = 1; 2163 } else 2164 hasdeps = 0; 2165 /* 2166 * We must hold the lock on a vnode before writing 2167 * one of its buffers. Otherwise we may confuse, or 2168 * in the case of a snapshot vnode, deadlock the 2169 * system. 2170 * 2171 * The lock order here is the reverse of the normal 2172 * of vnode followed by buf lock. This is ok because 2173 * the NOWAIT will prevent deadlock. 2174 */ 2175 if ((vp = bp->b_vp) == NULL || 2176 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2177 mtx_unlock(&bqlock); 2178 vfs_bio_awrite(bp); 2179 if (vp != NULL) 2180 VOP_UNLOCK(vp, 0, td); 2181 flushwithdeps += hasdeps; 2182 return (1); 2183 } 2184 BUF_UNLOCK(bp); 2185 } 2186 mtx_unlock(&bqlock); 2187 return (0); 2188} 2189 2190/* 2191 * Check to see if a block is currently memory resident. 2192 */ 2193struct buf * 2194incore(struct vnode * vp, daddr_t blkno) 2195{ 2196 struct buf *bp; 2197 2198 int s = splbio(); 2199 VI_LOCK(vp); 2200 bp = gbincore(vp, blkno); 2201 VI_UNLOCK(vp); 2202 splx(s); 2203 return (bp); 2204} 2205 2206/* 2207 * Returns true if no I/O is needed to access the 2208 * associated VM object. This is like incore except 2209 * it also hunts around in the VM system for the data. 2210 */ 2211 2212int 2213inmem(struct vnode * vp, daddr_t blkno) 2214{ 2215 vm_object_t obj; 2216 vm_offset_t toff, tinc, size; 2217 vm_page_t m; 2218 vm_ooffset_t off; 2219 2220 GIANT_REQUIRED; 2221 ASSERT_VOP_LOCKED(vp, "inmem"); 2222 2223 if (incore(vp, blkno)) 2224 return 1; 2225 if (vp->v_mount == NULL) 2226 return 0; 2227 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2228 return 0; 2229 2230 size = PAGE_SIZE; 2231 if (size > vp->v_mount->mnt_stat.f_iosize) 2232 size = vp->v_mount->mnt_stat.f_iosize; 2233 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2234 2235 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2236 VM_OBJECT_LOCK(obj); 2237 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2238 VM_OBJECT_UNLOCK(obj); 2239 if (!m) 2240 goto notinmem; 2241 tinc = size; 2242 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2243 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2244 if (vm_page_is_valid(m, 2245 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2246 goto notinmem; 2247 } 2248 return 1; 2249 2250notinmem: 2251 return (0); 2252} 2253 2254/* 2255 * vfs_setdirty: 2256 * 2257 * Sets the dirty range for a buffer based on the status of the dirty 2258 * bits in the pages comprising the buffer. 2259 * 2260 * The range is limited to the size of the buffer. 2261 * 2262 * This routine is primarily used by NFS, but is generalized for the 2263 * B_VMIO case. 2264 */ 2265static void 2266vfs_setdirty(struct buf *bp) 2267{ 2268 int i; 2269 vm_object_t object; 2270 2271 GIANT_REQUIRED; 2272 /* 2273 * Degenerate case - empty buffer 2274 */ 2275 2276 if (bp->b_bufsize == 0) 2277 return; 2278 2279 /* 2280 * We qualify the scan for modified pages on whether the 2281 * object has been flushed yet. The OBJ_WRITEABLE flag 2282 * is not cleared simply by protecting pages off. 2283 */ 2284 2285 if ((bp->b_flags & B_VMIO) == 0) 2286 return; 2287 2288 object = bp->b_pages[0]->object; 2289 VM_OBJECT_LOCK(object); 2290 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2291 printf("Warning: object %p writeable but not mightbedirty\n", object); 2292 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2293 printf("Warning: object %p mightbedirty but not writeable\n", object); 2294 2295 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2296 vm_offset_t boffset; 2297 vm_offset_t eoffset; 2298 2299 vm_page_lock_queues(); 2300 /* 2301 * test the pages to see if they have been modified directly 2302 * by users through the VM system. 2303 */ 2304 for (i = 0; i < bp->b_npages; i++) { 2305 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2306 vm_page_test_dirty(bp->b_pages[i]); 2307 } 2308 2309 /* 2310 * Calculate the encompassing dirty range, boffset and eoffset, 2311 * (eoffset - boffset) bytes. 2312 */ 2313 2314 for (i = 0; i < bp->b_npages; i++) { 2315 if (bp->b_pages[i]->dirty) 2316 break; 2317 } 2318 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2319 2320 for (i = bp->b_npages - 1; i >= 0; --i) { 2321 if (bp->b_pages[i]->dirty) { 2322 break; 2323 } 2324 } 2325 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2326 2327 vm_page_unlock_queues(); 2328 /* 2329 * Fit it to the buffer. 2330 */ 2331 2332 if (eoffset > bp->b_bcount) 2333 eoffset = bp->b_bcount; 2334 2335 /* 2336 * If we have a good dirty range, merge with the existing 2337 * dirty range. 2338 */ 2339 2340 if (boffset < eoffset) { 2341 if (bp->b_dirtyoff > boffset) 2342 bp->b_dirtyoff = boffset; 2343 if (bp->b_dirtyend < eoffset) 2344 bp->b_dirtyend = eoffset; 2345 } 2346 } 2347 VM_OBJECT_UNLOCK(object); 2348} 2349 2350/* 2351 * getblk: 2352 * 2353 * Get a block given a specified block and offset into a file/device. 2354 * The buffers B_DONE bit will be cleared on return, making it almost 2355 * ready for an I/O initiation. B_INVAL may or may not be set on 2356 * return. The caller should clear B_INVAL prior to initiating a 2357 * READ. 2358 * 2359 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2360 * an existing buffer. 2361 * 2362 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2363 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2364 * and then cleared based on the backing VM. If the previous buffer is 2365 * non-0-sized but invalid, B_CACHE will be cleared. 2366 * 2367 * If getblk() must create a new buffer, the new buffer is returned with 2368 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2369 * case it is returned with B_INVAL clear and B_CACHE set based on the 2370 * backing VM. 2371 * 2372 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2373 * B_CACHE bit is clear. 2374 * 2375 * What this means, basically, is that the caller should use B_CACHE to 2376 * determine whether the buffer is fully valid or not and should clear 2377 * B_INVAL prior to issuing a read. If the caller intends to validate 2378 * the buffer by loading its data area with something, the caller needs 2379 * to clear B_INVAL. If the caller does this without issuing an I/O, 2380 * the caller should set B_CACHE ( as an optimization ), else the caller 2381 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2382 * a write attempt or if it was a successfull read. If the caller 2383 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2384 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2385 */ 2386struct buf * 2387getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2388 int flags) 2389{ 2390 struct buf *bp; 2391 int s; 2392 int error; 2393 ASSERT_VOP_LOCKED(vp, "getblk"); 2394 2395 if (size > MAXBSIZE) 2396 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2397 2398 s = splbio(); 2399loop: 2400 /* 2401 * Block if we are low on buffers. Certain processes are allowed 2402 * to completely exhaust the buffer cache. 2403 * 2404 * If this check ever becomes a bottleneck it may be better to 2405 * move it into the else, when gbincore() fails. At the moment 2406 * it isn't a problem. 2407 * 2408 * XXX remove if 0 sections (clean this up after its proven) 2409 */ 2410 if (numfreebuffers == 0) { 2411 if (curthread == PCPU_GET(idlethread)) 2412 return NULL; 2413 mtx_lock(&nblock); 2414 needsbuffer |= VFS_BIO_NEED_ANY; 2415 mtx_unlock(&nblock); 2416 } 2417 2418 VI_LOCK(vp); 2419 if ((bp = gbincore(vp, blkno))) { 2420 int lockflags; 2421 /* 2422 * Buffer is in-core. If the buffer is not busy, it must 2423 * be on a queue. 2424 */ 2425 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2426 2427 if (flags & GB_LOCK_NOWAIT) 2428 lockflags |= LK_NOWAIT; 2429 2430 error = BUF_TIMELOCK(bp, lockflags, 2431 VI_MTX(vp), "getblk", slpflag, slptimeo); 2432 2433 /* 2434 * If we slept and got the lock we have to restart in case 2435 * the buffer changed identities. 2436 */ 2437 if (error == ENOLCK) 2438 goto loop; 2439 /* We timed out or were interrupted. */ 2440 else if (error) 2441 return (NULL); 2442 2443 /* 2444 * The buffer is locked. B_CACHE is cleared if the buffer is 2445 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2446 * and for a VMIO buffer B_CACHE is adjusted according to the 2447 * backing VM cache. 2448 */ 2449 if (bp->b_flags & B_INVAL) 2450 bp->b_flags &= ~B_CACHE; 2451 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2452 bp->b_flags |= B_CACHE; 2453 bremfree(bp); 2454 2455 /* 2456 * check for size inconsistancies for non-VMIO case. 2457 */ 2458 2459 if (bp->b_bcount != size) { 2460 if ((bp->b_flags & B_VMIO) == 0 || 2461 (size > bp->b_kvasize)) { 2462 if (bp->b_flags & B_DELWRI) { 2463 bp->b_flags |= B_NOCACHE; 2464 BUF_WRITE(bp); 2465 } else { 2466 if ((bp->b_flags & B_VMIO) && 2467 (LIST_FIRST(&bp->b_dep) == NULL)) { 2468 bp->b_flags |= B_RELBUF; 2469 brelse(bp); 2470 } else { 2471 bp->b_flags |= B_NOCACHE; 2472 BUF_WRITE(bp); 2473 } 2474 } 2475 goto loop; 2476 } 2477 } 2478 2479 /* 2480 * If the size is inconsistant in the VMIO case, we can resize 2481 * the buffer. This might lead to B_CACHE getting set or 2482 * cleared. If the size has not changed, B_CACHE remains 2483 * unchanged from its previous state. 2484 */ 2485 2486 if (bp->b_bcount != size) 2487 allocbuf(bp, size); 2488 2489 KASSERT(bp->b_offset != NOOFFSET, 2490 ("getblk: no buffer offset")); 2491 2492 /* 2493 * A buffer with B_DELWRI set and B_CACHE clear must 2494 * be committed before we can return the buffer in 2495 * order to prevent the caller from issuing a read 2496 * ( due to B_CACHE not being set ) and overwriting 2497 * it. 2498 * 2499 * Most callers, including NFS and FFS, need this to 2500 * operate properly either because they assume they 2501 * can issue a read if B_CACHE is not set, or because 2502 * ( for example ) an uncached B_DELWRI might loop due 2503 * to softupdates re-dirtying the buffer. In the latter 2504 * case, B_CACHE is set after the first write completes, 2505 * preventing further loops. 2506 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2507 * above while extending the buffer, we cannot allow the 2508 * buffer to remain with B_CACHE set after the write 2509 * completes or it will represent a corrupt state. To 2510 * deal with this we set B_NOCACHE to scrap the buffer 2511 * after the write. 2512 * 2513 * We might be able to do something fancy, like setting 2514 * B_CACHE in bwrite() except if B_DELWRI is already set, 2515 * so the below call doesn't set B_CACHE, but that gets real 2516 * confusing. This is much easier. 2517 */ 2518 2519 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2520 bp->b_flags |= B_NOCACHE; 2521 BUF_WRITE(bp); 2522 goto loop; 2523 } 2524 2525 splx(s); 2526 bp->b_flags &= ~B_DONE; 2527 } else { 2528 int bsize, maxsize, vmio; 2529 off_t offset; 2530 2531 /* 2532 * Buffer is not in-core, create new buffer. The buffer 2533 * returned by getnewbuf() is locked. Note that the returned 2534 * buffer is also considered valid (not marked B_INVAL). 2535 */ 2536 VI_UNLOCK(vp); 2537 if (vn_isdisk(vp, NULL)) 2538 bsize = DEV_BSIZE; 2539 else if (vp->v_mountedhere) 2540 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2541 else if (vp->v_mount) 2542 bsize = vp->v_mount->mnt_stat.f_iosize; 2543 else 2544 bsize = size; 2545 2546 offset = blkno * bsize; 2547 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2548 (vp->v_vflag & VV_OBJBUF); 2549 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2550 maxsize = imax(maxsize, bsize); 2551 2552 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2553 if (slpflag || slptimeo) { 2554 splx(s); 2555 return NULL; 2556 } 2557 goto loop; 2558 } 2559 2560 /* 2561 * This code is used to make sure that a buffer is not 2562 * created while the getnewbuf routine is blocked. 2563 * This can be a problem whether the vnode is locked or not. 2564 * If the buffer is created out from under us, we have to 2565 * throw away the one we just created. There is now window 2566 * race because we are safely running at splbio() from the 2567 * point of the duplicate buffer creation through to here, 2568 * and we've locked the buffer. 2569 * 2570 * Note: this must occur before we associate the buffer 2571 * with the vp especially considering limitations in 2572 * the splay tree implementation when dealing with duplicate 2573 * lblkno's. 2574 */ 2575 VI_LOCK(vp); 2576 if (gbincore(vp, blkno)) { 2577 VI_UNLOCK(vp); 2578 bp->b_flags |= B_INVAL; 2579 brelse(bp); 2580 goto loop; 2581 } 2582 2583 /* 2584 * Insert the buffer into the hash, so that it can 2585 * be found by incore. 2586 */ 2587 bp->b_blkno = bp->b_lblkno = blkno; 2588 bp->b_offset = offset; 2589 2590 bgetvp(vp, bp); 2591 VI_UNLOCK(vp); 2592 2593 /* 2594 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2595 * buffer size starts out as 0, B_CACHE will be set by 2596 * allocbuf() for the VMIO case prior to it testing the 2597 * backing store for validity. 2598 */ 2599 2600 if (vmio) { 2601 bp->b_flags |= B_VMIO; 2602#if defined(VFS_BIO_DEBUG) 2603 if (vp->v_type != VREG) 2604 printf("getblk: vmioing file type %d???\n", vp->v_type); 2605#endif 2606 VOP_GETVOBJECT(vp, &bp->b_object); 2607 } else { 2608 bp->b_flags &= ~B_VMIO; 2609 bp->b_object = NULL; 2610 } 2611 2612 allocbuf(bp, size); 2613 2614 splx(s); 2615 bp->b_flags &= ~B_DONE; 2616 } 2617 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2618 return (bp); 2619} 2620 2621/* 2622 * Get an empty, disassociated buffer of given size. The buffer is initially 2623 * set to B_INVAL. 2624 */ 2625struct buf * 2626geteblk(int size) 2627{ 2628 struct buf *bp; 2629 int s; 2630 int maxsize; 2631 2632 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2633 2634 s = splbio(); 2635 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2636 continue; 2637 splx(s); 2638 allocbuf(bp, size); 2639 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2640 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2641 return (bp); 2642} 2643 2644 2645/* 2646 * This code constitutes the buffer memory from either anonymous system 2647 * memory (in the case of non-VMIO operations) or from an associated 2648 * VM object (in the case of VMIO operations). This code is able to 2649 * resize a buffer up or down. 2650 * 2651 * Note that this code is tricky, and has many complications to resolve 2652 * deadlock or inconsistant data situations. Tread lightly!!! 2653 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2654 * the caller. Calling this code willy nilly can result in the loss of data. 2655 * 2656 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2657 * B_CACHE for the non-VMIO case. 2658 */ 2659 2660int 2661allocbuf(struct buf *bp, int size) 2662{ 2663 int newbsize, mbsize; 2664 int i; 2665 2666 GIANT_REQUIRED; 2667 2668 if (BUF_REFCNT(bp) == 0) 2669 panic("allocbuf: buffer not busy"); 2670 2671 if (bp->b_kvasize < size) 2672 panic("allocbuf: buffer too small"); 2673 2674 if ((bp->b_flags & B_VMIO) == 0) { 2675 caddr_t origbuf; 2676 int origbufsize; 2677 /* 2678 * Just get anonymous memory from the kernel. Don't 2679 * mess with B_CACHE. 2680 */ 2681 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2682 if (bp->b_flags & B_MALLOC) 2683 newbsize = mbsize; 2684 else 2685 newbsize = round_page(size); 2686 2687 if (newbsize < bp->b_bufsize) { 2688 /* 2689 * malloced buffers are not shrunk 2690 */ 2691 if (bp->b_flags & B_MALLOC) { 2692 if (newbsize) { 2693 bp->b_bcount = size; 2694 } else { 2695 free(bp->b_data, M_BIOBUF); 2696 if (bp->b_bufsize) { 2697 atomic_subtract_int( 2698 &bufmallocspace, 2699 bp->b_bufsize); 2700 bufspacewakeup(); 2701 bp->b_bufsize = 0; 2702 } 2703 bp->b_saveaddr = bp->b_kvabase; 2704 bp->b_data = bp->b_saveaddr; 2705 bp->b_bcount = 0; 2706 bp->b_flags &= ~B_MALLOC; 2707 } 2708 return 1; 2709 } 2710 vm_hold_free_pages( 2711 bp, 2712 (vm_offset_t) bp->b_data + newbsize, 2713 (vm_offset_t) bp->b_data + bp->b_bufsize); 2714 } else if (newbsize > bp->b_bufsize) { 2715 /* 2716 * We only use malloced memory on the first allocation. 2717 * and revert to page-allocated memory when the buffer 2718 * grows. 2719 */ 2720 /* 2721 * There is a potential smp race here that could lead 2722 * to bufmallocspace slightly passing the max. It 2723 * is probably extremely rare and not worth worrying 2724 * over. 2725 */ 2726 if ( (bufmallocspace < maxbufmallocspace) && 2727 (bp->b_bufsize == 0) && 2728 (mbsize <= PAGE_SIZE/2)) { 2729 2730 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2731 bp->b_bufsize = mbsize; 2732 bp->b_bcount = size; 2733 bp->b_flags |= B_MALLOC; 2734 atomic_add_int(&bufmallocspace, mbsize); 2735 return 1; 2736 } 2737 origbuf = NULL; 2738 origbufsize = 0; 2739 /* 2740 * If the buffer is growing on its other-than-first allocation, 2741 * then we revert to the page-allocation scheme. 2742 */ 2743 if (bp->b_flags & B_MALLOC) { 2744 origbuf = bp->b_data; 2745 origbufsize = bp->b_bufsize; 2746 bp->b_data = bp->b_kvabase; 2747 if (bp->b_bufsize) { 2748 atomic_subtract_int(&bufmallocspace, 2749 bp->b_bufsize); 2750 bufspacewakeup(); 2751 bp->b_bufsize = 0; 2752 } 2753 bp->b_flags &= ~B_MALLOC; 2754 newbsize = round_page(newbsize); 2755 } 2756 vm_hold_load_pages( 2757 bp, 2758 (vm_offset_t) bp->b_data + bp->b_bufsize, 2759 (vm_offset_t) bp->b_data + newbsize); 2760 if (origbuf) { 2761 bcopy(origbuf, bp->b_data, origbufsize); 2762 free(origbuf, M_BIOBUF); 2763 } 2764 } 2765 } else { 2766 int desiredpages; 2767 2768 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2769 desiredpages = (size == 0) ? 0 : 2770 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2771 2772 if (bp->b_flags & B_MALLOC) 2773 panic("allocbuf: VMIO buffer can't be malloced"); 2774 /* 2775 * Set B_CACHE initially if buffer is 0 length or will become 2776 * 0-length. 2777 */ 2778 if (size == 0 || bp->b_bufsize == 0) 2779 bp->b_flags |= B_CACHE; 2780 2781 if (newbsize < bp->b_bufsize) { 2782 /* 2783 * DEV_BSIZE aligned new buffer size is less then the 2784 * DEV_BSIZE aligned existing buffer size. Figure out 2785 * if we have to remove any pages. 2786 */ 2787 if (desiredpages < bp->b_npages) { 2788 vm_page_t m; 2789 2790 vm_page_lock_queues(); 2791 for (i = desiredpages; i < bp->b_npages; i++) { 2792 /* 2793 * the page is not freed here -- it 2794 * is the responsibility of 2795 * vnode_pager_setsize 2796 */ 2797 m = bp->b_pages[i]; 2798 KASSERT(m != bogus_page, 2799 ("allocbuf: bogus page found")); 2800 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2801 vm_page_lock_queues(); 2802 2803 bp->b_pages[i] = NULL; 2804 vm_page_unwire(m, 0); 2805 } 2806 vm_page_unlock_queues(); 2807 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2808 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2809 bp->b_npages = desiredpages; 2810 } 2811 } else if (size > bp->b_bcount) { 2812 /* 2813 * We are growing the buffer, possibly in a 2814 * byte-granular fashion. 2815 */ 2816 struct vnode *vp; 2817 vm_object_t obj; 2818 vm_offset_t toff; 2819 vm_offset_t tinc; 2820 2821 /* 2822 * Step 1, bring in the VM pages from the object, 2823 * allocating them if necessary. We must clear 2824 * B_CACHE if these pages are not valid for the 2825 * range covered by the buffer. 2826 */ 2827 2828 vp = bp->b_vp; 2829 obj = bp->b_object; 2830 2831 VM_OBJECT_LOCK(obj); 2832 while (bp->b_npages < desiredpages) { 2833 vm_page_t m; 2834 vm_pindex_t pi; 2835 2836 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2837 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2838 /* 2839 * note: must allocate system pages 2840 * since blocking here could intefere 2841 * with paging I/O, no matter which 2842 * process we are. 2843 */ 2844 m = vm_page_alloc(obj, pi, 2845 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2846 if (m == NULL) { 2847 atomic_add_int(&vm_pageout_deficit, 2848 desiredpages - bp->b_npages); 2849 VM_OBJECT_UNLOCK(obj); 2850 VM_WAIT; 2851 VM_OBJECT_LOCK(obj); 2852 } else { 2853 vm_page_lock_queues(); 2854 vm_page_wakeup(m); 2855 vm_page_unlock_queues(); 2856 bp->b_flags &= ~B_CACHE; 2857 bp->b_pages[bp->b_npages] = m; 2858 ++bp->b_npages; 2859 } 2860 continue; 2861 } 2862 2863 /* 2864 * We found a page. If we have to sleep on it, 2865 * retry because it might have gotten freed out 2866 * from under us. 2867 * 2868 * We can only test PG_BUSY here. Blocking on 2869 * m->busy might lead to a deadlock: 2870 * 2871 * vm_fault->getpages->cluster_read->allocbuf 2872 * 2873 */ 2874 vm_page_lock_queues(); 2875 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2876 continue; 2877 2878 /* 2879 * We have a good page. Should we wakeup the 2880 * page daemon? 2881 */ 2882 if ((curproc != pageproc) && 2883 ((m->queue - m->pc) == PQ_CACHE) && 2884 ((cnt.v_free_count + cnt.v_cache_count) < 2885 (cnt.v_free_min + cnt.v_cache_min))) { 2886 pagedaemon_wakeup(); 2887 } 2888 vm_page_flag_clear(m, PG_ZERO); 2889 vm_page_wire(m); 2890 vm_page_unlock_queues(); 2891 bp->b_pages[bp->b_npages] = m; 2892 ++bp->b_npages; 2893 } 2894 VM_OBJECT_UNLOCK(obj); 2895 2896 /* 2897 * Step 2. We've loaded the pages into the buffer, 2898 * we have to figure out if we can still have B_CACHE 2899 * set. Note that B_CACHE is set according to the 2900 * byte-granular range ( bcount and size ), new the 2901 * aligned range ( newbsize ). 2902 * 2903 * The VM test is against m->valid, which is DEV_BSIZE 2904 * aligned. Needless to say, the validity of the data 2905 * needs to also be DEV_BSIZE aligned. Note that this 2906 * fails with NFS if the server or some other client 2907 * extends the file's EOF. If our buffer is resized, 2908 * B_CACHE may remain set! XXX 2909 */ 2910 2911 toff = bp->b_bcount; 2912 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2913 2914 while ((bp->b_flags & B_CACHE) && toff < size) { 2915 vm_pindex_t pi; 2916 2917 if (tinc > (size - toff)) 2918 tinc = size - toff; 2919 2920 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2921 PAGE_SHIFT; 2922 2923 vfs_buf_test_cache( 2924 bp, 2925 bp->b_offset, 2926 toff, 2927 tinc, 2928 bp->b_pages[pi] 2929 ); 2930 toff += tinc; 2931 tinc = PAGE_SIZE; 2932 } 2933 2934 /* 2935 * Step 3, fixup the KVM pmap. Remember that 2936 * bp->b_data is relative to bp->b_offset, but 2937 * bp->b_offset may be offset into the first page. 2938 */ 2939 2940 bp->b_data = (caddr_t) 2941 trunc_page((vm_offset_t)bp->b_data); 2942 pmap_qenter( 2943 (vm_offset_t)bp->b_data, 2944 bp->b_pages, 2945 bp->b_npages 2946 ); 2947 2948 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2949 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2950 } 2951 } 2952 if (newbsize < bp->b_bufsize) 2953 bufspacewakeup(); 2954 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2955 bp->b_bcount = size; /* requested buffer size */ 2956 return 1; 2957} 2958 2959void 2960biodone(struct bio *bp) 2961{ 2962 mtx_lock(&bdonelock); 2963 bp->bio_flags |= BIO_DONE; 2964 if (bp->bio_done == NULL) 2965 wakeup(bp); 2966 mtx_unlock(&bdonelock); 2967 if (bp->bio_done != NULL) 2968 bp->bio_done(bp); 2969} 2970 2971/* 2972 * Wait for a BIO to finish. 2973 * 2974 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2975 * case is not yet clear. 2976 */ 2977int 2978biowait(struct bio *bp, const char *wchan) 2979{ 2980 2981 mtx_lock(&bdonelock); 2982 while ((bp->bio_flags & BIO_DONE) == 0) 2983 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 2984 mtx_unlock(&bdonelock); 2985 if (bp->bio_error != 0) 2986 return (bp->bio_error); 2987 if (!(bp->bio_flags & BIO_ERROR)) 2988 return (0); 2989 return (EIO); 2990} 2991 2992void 2993biofinish(struct bio *bp, struct devstat *stat, int error) 2994{ 2995 2996 if (error) { 2997 bp->bio_error = error; 2998 bp->bio_flags |= BIO_ERROR; 2999 } 3000 if (stat != NULL) 3001 devstat_end_transaction_bio(stat, bp); 3002 biodone(bp); 3003} 3004 3005/* 3006 * bufwait: 3007 * 3008 * Wait for buffer I/O completion, returning error status. The buffer 3009 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3010 * error and cleared. 3011 */ 3012int 3013bufwait(register struct buf * bp) 3014{ 3015 int s; 3016 3017 s = splbio(); 3018 if (bp->b_iocmd == BIO_READ) 3019 bwait(bp, PRIBIO, "biord"); 3020 else 3021 bwait(bp, PRIBIO, "biowr"); 3022 splx(s); 3023 if (bp->b_flags & B_EINTR) { 3024 bp->b_flags &= ~B_EINTR; 3025 return (EINTR); 3026 } 3027 if (bp->b_ioflags & BIO_ERROR) { 3028 return (bp->b_error ? bp->b_error : EIO); 3029 } else { 3030 return (0); 3031 } 3032} 3033 3034 /* 3035 * Call back function from struct bio back up to struct buf. 3036 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3037 */ 3038void 3039bufdonebio(struct bio *bp) 3040{ 3041 3042 /* Device drivers may or may not hold giant, hold it here. */ 3043 mtx_lock(&Giant); 3044 bufdone(bp->bio_caller2); 3045 mtx_unlock(&Giant); 3046} 3047 3048/* 3049 * bufdone: 3050 * 3051 * Finish I/O on a buffer, optionally calling a completion function. 3052 * This is usually called from an interrupt so process blocking is 3053 * not allowed. 3054 * 3055 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3056 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3057 * assuming B_INVAL is clear. 3058 * 3059 * For the VMIO case, we set B_CACHE if the op was a read and no 3060 * read error occured, or if the op was a write. B_CACHE is never 3061 * set if the buffer is invalid or otherwise uncacheable. 3062 * 3063 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3064 * initiator to leave B_INVAL set to brelse the buffer out of existance 3065 * in the biodone routine. 3066 */ 3067void 3068bufdone(struct buf *bp) 3069{ 3070 int s; 3071 void (*biodone)(struct buf *); 3072 3073 GIANT_REQUIRED; 3074 3075 s = splbio(); 3076 3077 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3078 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3079 3080 bp->b_flags |= B_DONE; 3081 runningbufwakeup(bp); 3082 3083 if (bp->b_iocmd == BIO_DELETE) { 3084 brelse(bp); 3085 splx(s); 3086 return; 3087 } 3088 3089 if (bp->b_iocmd == BIO_WRITE) { 3090 vwakeup(bp); 3091 } 3092 3093 /* call optional completion function if requested */ 3094 if (bp->b_iodone != NULL) { 3095 biodone = bp->b_iodone; 3096 bp->b_iodone = NULL; 3097 (*biodone) (bp); 3098 splx(s); 3099 return; 3100 } 3101 if (LIST_FIRST(&bp->b_dep) != NULL) 3102 buf_complete(bp); 3103 3104 if (bp->b_flags & B_VMIO) { 3105 int i; 3106 vm_ooffset_t foff; 3107 vm_page_t m; 3108 vm_object_t obj; 3109 int iosize; 3110 struct vnode *vp = bp->b_vp; 3111 3112 obj = bp->b_object; 3113 3114#if defined(VFS_BIO_DEBUG) 3115 mp_fixme("usecount and vflag accessed without locks."); 3116 if (vp->v_usecount == 0) { 3117 panic("biodone: zero vnode ref count"); 3118 } 3119 3120 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3121 panic("biodone: vnode is not setup for merged cache"); 3122 } 3123#endif 3124 3125 foff = bp->b_offset; 3126 KASSERT(bp->b_offset != NOOFFSET, 3127 ("biodone: no buffer offset")); 3128 3129 if (obj != NULL) 3130 VM_OBJECT_LOCK(obj); 3131#if defined(VFS_BIO_DEBUG) 3132 if (obj->paging_in_progress < bp->b_npages) { 3133 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3134 obj->paging_in_progress, bp->b_npages); 3135 } 3136#endif 3137 3138 /* 3139 * Set B_CACHE if the op was a normal read and no error 3140 * occured. B_CACHE is set for writes in the b*write() 3141 * routines. 3142 */ 3143 iosize = bp->b_bcount - bp->b_resid; 3144 if (bp->b_iocmd == BIO_READ && 3145 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3146 !(bp->b_ioflags & BIO_ERROR)) { 3147 bp->b_flags |= B_CACHE; 3148 } 3149 vm_page_lock_queues(); 3150 for (i = 0; i < bp->b_npages; i++) { 3151 int bogusflag = 0; 3152 int resid; 3153 3154 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3155 if (resid > iosize) 3156 resid = iosize; 3157 3158 /* 3159 * cleanup bogus pages, restoring the originals 3160 */ 3161 m = bp->b_pages[i]; 3162 if (m == bogus_page) { 3163 bogusflag = 1; 3164 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3165 if (m == NULL) 3166 panic("biodone: page disappeared!"); 3167 bp->b_pages[i] = m; 3168 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3169 } 3170#if defined(VFS_BIO_DEBUG) 3171 if (OFF_TO_IDX(foff) != m->pindex) { 3172 printf( 3173"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3174 (intmax_t)foff, (uintmax_t)m->pindex); 3175 } 3176#endif 3177 3178 /* 3179 * In the write case, the valid and clean bits are 3180 * already changed correctly ( see bdwrite() ), so we 3181 * only need to do this here in the read case. 3182 */ 3183 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3184 vfs_page_set_valid(bp, foff, i, m); 3185 } 3186 vm_page_flag_clear(m, PG_ZERO); 3187 3188 /* 3189 * when debugging new filesystems or buffer I/O methods, this 3190 * is the most common error that pops up. if you see this, you 3191 * have not set the page busy flag correctly!!! 3192 */ 3193 if (m->busy == 0) { 3194 printf("biodone: page busy < 0, " 3195 "pindex: %d, foff: 0x(%x,%x), " 3196 "resid: %d, index: %d\n", 3197 (int) m->pindex, (int)(foff >> 32), 3198 (int) foff & 0xffffffff, resid, i); 3199 if (!vn_isdisk(vp, NULL)) 3200 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3201 bp->b_vp->v_mount->mnt_stat.f_iosize, 3202 (intmax_t) bp->b_lblkno, 3203 bp->b_flags, bp->b_npages); 3204 else 3205 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3206 (intmax_t) bp->b_lblkno, 3207 bp->b_flags, bp->b_npages); 3208 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3209 (u_long)m->valid, (u_long)m->dirty, 3210 m->wire_count); 3211 panic("biodone: page busy < 0\n"); 3212 } 3213 vm_page_io_finish(m); 3214 vm_object_pip_subtract(obj, 1); 3215 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3216 iosize -= resid; 3217 } 3218 vm_page_unlock_queues(); 3219 if (obj != NULL) { 3220 vm_object_pip_wakeupn(obj, 0); 3221 VM_OBJECT_UNLOCK(obj); 3222 } 3223 } 3224 3225 /* 3226 * For asynchronous completions, release the buffer now. The brelse 3227 * will do a wakeup there if necessary - so no need to do a wakeup 3228 * here in the async case. The sync case always needs to do a wakeup. 3229 */ 3230 3231 if (bp->b_flags & B_ASYNC) { 3232 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3233 brelse(bp); 3234 else 3235 bqrelse(bp); 3236 } else { 3237 bdone(bp); 3238 } 3239 splx(s); 3240} 3241 3242/* 3243 * This routine is called in lieu of iodone in the case of 3244 * incomplete I/O. This keeps the busy status for pages 3245 * consistant. 3246 */ 3247void 3248vfs_unbusy_pages(struct buf * bp) 3249{ 3250 int i; 3251 3252 GIANT_REQUIRED; 3253 3254 runningbufwakeup(bp); 3255 if (bp->b_flags & B_VMIO) { 3256 vm_object_t obj; 3257 3258 obj = bp->b_object; 3259 VM_OBJECT_LOCK(obj); 3260 vm_page_lock_queues(); 3261 for (i = 0; i < bp->b_npages; i++) { 3262 vm_page_t m = bp->b_pages[i]; 3263 3264 if (m == bogus_page) { 3265 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3266 if (!m) { 3267 panic("vfs_unbusy_pages: page missing\n"); 3268 } 3269 bp->b_pages[i] = m; 3270 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3271 } 3272 vm_object_pip_subtract(obj, 1); 3273 vm_page_flag_clear(m, PG_ZERO); 3274 vm_page_io_finish(m); 3275 } 3276 vm_page_unlock_queues(); 3277 vm_object_pip_wakeupn(obj, 0); 3278 VM_OBJECT_UNLOCK(obj); 3279 } 3280} 3281 3282/* 3283 * vfs_page_set_valid: 3284 * 3285 * Set the valid bits in a page based on the supplied offset. The 3286 * range is restricted to the buffer's size. 3287 * 3288 * This routine is typically called after a read completes. 3289 */ 3290static void 3291vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3292{ 3293 vm_ooffset_t soff, eoff; 3294 3295 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3296 /* 3297 * Start and end offsets in buffer. eoff - soff may not cross a 3298 * page boundry or cross the end of the buffer. The end of the 3299 * buffer, in this case, is our file EOF, not the allocation size 3300 * of the buffer. 3301 */ 3302 soff = off; 3303 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3304 if (eoff > bp->b_offset + bp->b_bcount) 3305 eoff = bp->b_offset + bp->b_bcount; 3306 3307 /* 3308 * Set valid range. This is typically the entire buffer and thus the 3309 * entire page. 3310 */ 3311 if (eoff > soff) { 3312 vm_page_set_validclean( 3313 m, 3314 (vm_offset_t) (soff & PAGE_MASK), 3315 (vm_offset_t) (eoff - soff) 3316 ); 3317 } 3318} 3319 3320/* 3321 * This routine is called before a device strategy routine. 3322 * It is used to tell the VM system that paging I/O is in 3323 * progress, and treat the pages associated with the buffer 3324 * almost as being PG_BUSY. Also the object paging_in_progress 3325 * flag is handled to make sure that the object doesn't become 3326 * inconsistant. 3327 * 3328 * Since I/O has not been initiated yet, certain buffer flags 3329 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3330 * and should be ignored. 3331 */ 3332void 3333vfs_busy_pages(struct buf * bp, int clear_modify) 3334{ 3335 int i, bogus; 3336 3337 if (bp->b_flags & B_VMIO) { 3338 vm_object_t obj; 3339 vm_ooffset_t foff; 3340 3341 obj = bp->b_object; 3342 foff = bp->b_offset; 3343 KASSERT(bp->b_offset != NOOFFSET, 3344 ("vfs_busy_pages: no buffer offset")); 3345 vfs_setdirty(bp); 3346 if (obj != NULL) 3347 VM_OBJECT_LOCK(obj); 3348retry: 3349 vm_page_lock_queues(); 3350 for (i = 0; i < bp->b_npages; i++) { 3351 vm_page_t m = bp->b_pages[i]; 3352 3353 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3354 goto retry; 3355 } 3356 bogus = 0; 3357 for (i = 0; i < bp->b_npages; i++) { 3358 vm_page_t m = bp->b_pages[i]; 3359 3360 vm_page_flag_clear(m, PG_ZERO); 3361 if ((bp->b_flags & B_CLUSTER) == 0) { 3362 vm_object_pip_add(obj, 1); 3363 vm_page_io_start(m); 3364 } 3365 /* 3366 * When readying a buffer for a read ( i.e 3367 * clear_modify == 0 ), it is important to do 3368 * bogus_page replacement for valid pages in 3369 * partially instantiated buffers. Partially 3370 * instantiated buffers can, in turn, occur when 3371 * reconstituting a buffer from its VM backing store 3372 * base. We only have to do this if B_CACHE is 3373 * clear ( which causes the I/O to occur in the 3374 * first place ). The replacement prevents the read 3375 * I/O from overwriting potentially dirty VM-backed 3376 * pages. XXX bogus page replacement is, uh, bogus. 3377 * It may not work properly with small-block devices. 3378 * We need to find a better way. 3379 */ 3380 pmap_remove_all(m); 3381 if (clear_modify) 3382 vfs_page_set_valid(bp, foff, i, m); 3383 else if (m->valid == VM_PAGE_BITS_ALL && 3384 (bp->b_flags & B_CACHE) == 0) { 3385 bp->b_pages[i] = bogus_page; 3386 bogus++; 3387 } 3388 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3389 } 3390 vm_page_unlock_queues(); 3391 if (obj != NULL) 3392 VM_OBJECT_UNLOCK(obj); 3393 if (bogus) 3394 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3395 } 3396} 3397 3398/* 3399 * Tell the VM system that the pages associated with this buffer 3400 * are clean. This is used for delayed writes where the data is 3401 * going to go to disk eventually without additional VM intevention. 3402 * 3403 * Note that while we only really need to clean through to b_bcount, we 3404 * just go ahead and clean through to b_bufsize. 3405 */ 3406static void 3407vfs_clean_pages(struct buf * bp) 3408{ 3409 int i; 3410 3411 GIANT_REQUIRED; 3412 3413 if (bp->b_flags & B_VMIO) { 3414 vm_ooffset_t foff; 3415 3416 foff = bp->b_offset; 3417 KASSERT(bp->b_offset != NOOFFSET, 3418 ("vfs_clean_pages: no buffer offset")); 3419 vm_page_lock_queues(); 3420 for (i = 0; i < bp->b_npages; i++) { 3421 vm_page_t m = bp->b_pages[i]; 3422 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3423 vm_ooffset_t eoff = noff; 3424 3425 if (eoff > bp->b_offset + bp->b_bufsize) 3426 eoff = bp->b_offset + bp->b_bufsize; 3427 vfs_page_set_valid(bp, foff, i, m); 3428 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3429 foff = noff; 3430 } 3431 vm_page_unlock_queues(); 3432 } 3433} 3434 3435/* 3436 * vfs_bio_set_validclean: 3437 * 3438 * Set the range within the buffer to valid and clean. The range is 3439 * relative to the beginning of the buffer, b_offset. Note that b_offset 3440 * itself may be offset from the beginning of the first page. 3441 * 3442 */ 3443 3444void 3445vfs_bio_set_validclean(struct buf *bp, int base, int size) 3446{ 3447 if (bp->b_flags & B_VMIO) { 3448 int i; 3449 int n; 3450 3451 /* 3452 * Fixup base to be relative to beginning of first page. 3453 * Set initial n to be the maximum number of bytes in the 3454 * first page that can be validated. 3455 */ 3456 3457 base += (bp->b_offset & PAGE_MASK); 3458 n = PAGE_SIZE - (base & PAGE_MASK); 3459 3460 vm_page_lock_queues(); 3461 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3462 vm_page_t m = bp->b_pages[i]; 3463 3464 if (n > size) 3465 n = size; 3466 3467 vm_page_set_validclean(m, base & PAGE_MASK, n); 3468 base += n; 3469 size -= n; 3470 n = PAGE_SIZE; 3471 } 3472 vm_page_unlock_queues(); 3473 } 3474} 3475 3476/* 3477 * vfs_bio_clrbuf: 3478 * 3479 * clear a buffer. This routine essentially fakes an I/O, so we need 3480 * to clear BIO_ERROR and B_INVAL. 3481 * 3482 * Note that while we only theoretically need to clear through b_bcount, 3483 * we go ahead and clear through b_bufsize. 3484 */ 3485 3486void 3487vfs_bio_clrbuf(struct buf *bp) 3488{ 3489 int i, mask = 0; 3490 caddr_t sa, ea; 3491 3492 GIANT_REQUIRED; 3493 3494 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3495 bp->b_flags &= ~B_INVAL; 3496 bp->b_ioflags &= ~BIO_ERROR; 3497 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3498 (bp->b_offset & PAGE_MASK) == 0) { 3499 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3500 if ((bp->b_pages[0]->valid & mask) == mask) { 3501 bp->b_resid = 0; 3502 return; 3503 } 3504 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3505 ((bp->b_pages[0]->valid & mask) == 0)) { 3506 bzero(bp->b_data, bp->b_bufsize); 3507 bp->b_pages[0]->valid |= mask; 3508 bp->b_resid = 0; 3509 return; 3510 } 3511 } 3512 ea = sa = bp->b_data; 3513 for(i=0;i<bp->b_npages;i++,sa=ea) { 3514 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3515 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3516 ea = (caddr_t)(vm_offset_t)ulmin( 3517 (u_long)(vm_offset_t)ea, 3518 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3519 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3520 if ((bp->b_pages[i]->valid & mask) == mask) 3521 continue; 3522 if ((bp->b_pages[i]->valid & mask) == 0) { 3523 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3524 bzero(sa, ea - sa); 3525 } 3526 } else { 3527 for (; sa < ea; sa += DEV_BSIZE, j++) { 3528 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3529 (bp->b_pages[i]->valid & (1<<j)) == 0) 3530 bzero(sa, DEV_BSIZE); 3531 } 3532 } 3533 bp->b_pages[i]->valid |= mask; 3534 vm_page_lock_queues(); 3535 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3536 vm_page_unlock_queues(); 3537 } 3538 bp->b_resid = 0; 3539 } else { 3540 clrbuf(bp); 3541 } 3542} 3543 3544/* 3545 * vm_hold_load_pages and vm_hold_free_pages get pages into 3546 * a buffers address space. The pages are anonymous and are 3547 * not associated with a file object. 3548 */ 3549static void 3550vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3551{ 3552 vm_offset_t pg; 3553 vm_page_t p; 3554 int index; 3555 3556 GIANT_REQUIRED; 3557 3558 to = round_page(to); 3559 from = round_page(from); 3560 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3561 3562 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3563tryagain: 3564 /* 3565 * note: must allocate system pages since blocking here 3566 * could intefere with paging I/O, no matter which 3567 * process we are. 3568 */ 3569 VM_OBJECT_LOCK(kernel_object); 3570 p = vm_page_alloc(kernel_object, 3571 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3572 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3573 VM_OBJECT_UNLOCK(kernel_object); 3574 if (!p) { 3575 atomic_add_int(&vm_pageout_deficit, 3576 (to - pg) >> PAGE_SHIFT); 3577 VM_WAIT; 3578 goto tryagain; 3579 } 3580 vm_page_lock_queues(); 3581 p->valid = VM_PAGE_BITS_ALL; 3582 vm_page_unlock_queues(); 3583 pmap_qenter(pg, &p, 1); 3584 bp->b_pages[index] = p; 3585 vm_page_lock_queues(); 3586 vm_page_wakeup(p); 3587 vm_page_unlock_queues(); 3588 } 3589 bp->b_npages = index; 3590} 3591 3592/* Return pages associated with this buf to the vm system */ 3593static void 3594vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3595{ 3596 vm_offset_t pg; 3597 vm_page_t p; 3598 int index, newnpages; 3599 3600 GIANT_REQUIRED; 3601 3602 from = round_page(from); 3603 to = round_page(to); 3604 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3605 3606 if (bp->b_object != NULL) 3607 VM_OBJECT_LOCK(bp->b_object); 3608 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3609 p = bp->b_pages[index]; 3610 if (p && (index < bp->b_npages)) { 3611 if (p->busy) { 3612 printf( 3613 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3614 (intmax_t)bp->b_blkno, 3615 (intmax_t)bp->b_lblkno); 3616 } 3617 bp->b_pages[index] = NULL; 3618 pmap_qremove(pg, 1); 3619 vm_page_lock_queues(); 3620 vm_page_busy(p); 3621 vm_page_unwire(p, 0); 3622 vm_page_free(p); 3623 vm_page_unlock_queues(); 3624 } 3625 } 3626 if (bp->b_object != NULL) 3627 VM_OBJECT_UNLOCK(bp->b_object); 3628 bp->b_npages = newnpages; 3629} 3630 3631/* 3632 * Map an IO request into kernel virtual address space. 3633 * 3634 * All requests are (re)mapped into kernel VA space. 3635 * Notice that we use b_bufsize for the size of the buffer 3636 * to be mapped. b_bcount might be modified by the driver. 3637 * 3638 * Note that even if the caller determines that the address space should 3639 * be valid, a race or a smaller-file mapped into a larger space may 3640 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3641 * check the return value. 3642 */ 3643int 3644vmapbuf(struct buf *bp) 3645{ 3646 caddr_t addr, kva; 3647 vm_paddr_t pa; 3648 int pidx, i; 3649 struct vm_page *m; 3650 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3651 3652 GIANT_REQUIRED; 3653 3654 if ((bp->b_flags & B_PHYS) == 0) 3655 panic("vmapbuf"); 3656 if (bp->b_bufsize < 0) 3657 return (-1); 3658 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3659 addr < bp->b_data + bp->b_bufsize; 3660 addr += PAGE_SIZE, pidx++) { 3661 /* 3662 * Do the vm_fault if needed; do the copy-on-write thing 3663 * when reading stuff off device into memory. 3664 * 3665 * NOTE! Must use pmap_extract() because addr may be in 3666 * the userland address space, and kextract is only guarenteed 3667 * to work for the kernland address space (see: sparc64 port). 3668 */ 3669retry: 3670 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3671 (bp->b_iocmd == BIO_READ) ? 3672 (VM_PROT_READ|VM_PROT_WRITE) : VM_PROT_READ); 3673 if (i < 0) { 3674 vm_page_lock_queues(); 3675 for (i = 0; i < pidx; ++i) { 3676 vm_page_unhold(bp->b_pages[i]); 3677 bp->b_pages[i] = NULL; 3678 } 3679 vm_page_unlock_queues(); 3680 return(-1); 3681 } 3682 pa = pmap_extract(pmap, (vm_offset_t)addr); 3683 if (pa == 0) { 3684 printf("vmapbuf: warning, race against user address during I/O"); 3685 goto retry; 3686 } 3687 m = PHYS_TO_VM_PAGE(pa); 3688 vm_page_lock_queues(); 3689 vm_page_hold(m); 3690 vm_page_unlock_queues(); 3691 bp->b_pages[pidx] = m; 3692 } 3693 if (pidx > btoc(MAXPHYS)) 3694 panic("vmapbuf: mapped more than MAXPHYS"); 3695 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3696 3697 kva = bp->b_saveaddr; 3698 bp->b_npages = pidx; 3699 bp->b_saveaddr = bp->b_data; 3700 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3701 return(0); 3702} 3703 3704/* 3705 * Free the io map PTEs associated with this IO operation. 3706 * We also invalidate the TLB entries and restore the original b_addr. 3707 */ 3708void 3709vunmapbuf(struct buf *bp) 3710{ 3711 int pidx; 3712 int npages; 3713 3714 GIANT_REQUIRED; 3715 3716 if ((bp->b_flags & B_PHYS) == 0) 3717 panic("vunmapbuf"); 3718 3719 npages = bp->b_npages; 3720 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3721 npages); 3722 vm_page_lock_queues(); 3723 for (pidx = 0; pidx < npages; pidx++) 3724 vm_page_unhold(bp->b_pages[pidx]); 3725 vm_page_unlock_queues(); 3726 3727 bp->b_data = bp->b_saveaddr; 3728} 3729 3730void 3731bdone(struct buf *bp) 3732{ 3733 mtx_lock(&bdonelock); 3734 bp->b_flags |= B_DONE; 3735 wakeup(bp); 3736 mtx_unlock(&bdonelock); 3737} 3738 3739void 3740bwait(struct buf *bp, u_char pri, const char *wchan) 3741{ 3742 mtx_lock(&bdonelock); 3743 while ((bp->b_flags & B_DONE) == 0) 3744 msleep(bp, &bdonelock, pri, wchan, 0); 3745 mtx_unlock(&bdonelock); 3746} 3747 3748#include "opt_ddb.h" 3749#ifdef DDB 3750#include <ddb/ddb.h> 3751 3752/* DDB command to show buffer data */ 3753DB_SHOW_COMMAND(buffer, db_show_buffer) 3754{ 3755 /* get args */ 3756 struct buf *bp = (struct buf *)addr; 3757 3758 if (!have_addr) { 3759 db_printf("usage: show buffer <addr>\n"); 3760 return; 3761 } 3762 3763 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3764 db_printf( 3765 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3766 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3767 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3768 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3769 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3770 if (bp->b_npages) { 3771 int i; 3772 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3773 for (i = 0; i < bp->b_npages; i++) { 3774 vm_page_t m; 3775 m = bp->b_pages[i]; 3776 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3777 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3778 if ((i + 1) < bp->b_npages) 3779 db_printf(","); 3780 } 3781 db_printf("\n"); 3782 } 3783} 3784#endif /* DDB */ 3785