vfs_bio.c revision 135617
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 135617 2004-09-23 14:45:04Z phk $"); 30 31#include <sys/param.h> 32#include <sys/systm.h> 33#include <sys/bio.h> 34#include <sys/conf.h> 35#include <sys/buf.h> 36#include <sys/devicestat.h> 37#include <sys/eventhandler.h> 38#include <sys/lock.h> 39#include <sys/malloc.h> 40#include <sys/mount.h> 41#include <sys/mutex.h> 42#include <sys/kernel.h> 43#include <sys/kthread.h> 44#include <sys/proc.h> 45#include <sys/resourcevar.h> 46#include <sys/sysctl.h> 47#include <sys/vmmeter.h> 48#include <sys/vnode.h> 49#include <vm/vm.h> 50#include <vm/vm_param.h> 51#include <vm/vm_kern.h> 52#include <vm/vm_pageout.h> 53#include <vm/vm_page.h> 54#include <vm/vm_object.h> 55#include <vm/vm_extern.h> 56#include <vm/vm_map.h> 57#include "opt_directio.h" 58#include "opt_swap.h" 59 60static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 61 62struct bio_ops bioops; /* I/O operation notification */ 63 64static int ibwrite(struct buf *); 65 66struct buf_ops buf_ops_bio = { 67 "buf_ops_bio", 68 ibwrite 69}; 70 71/* 72 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 73 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 74 */ 75struct buf *buf; /* buffer header pool */ 76 77static struct proc *bufdaemonproc; 78 79static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 80 vm_offset_t to); 81static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 82 vm_offset_t to); 83static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 84 int pageno, vm_page_t m); 85static void vfs_clean_pages(struct buf *bp); 86static void vfs_setdirty(struct buf *bp); 87static void vfs_vmio_release(struct buf *bp); 88static void vfs_backgroundwritedone(struct buf *bp); 89static int vfs_bio_clcheck(struct vnode *vp, int size, 90 daddr_t lblkno, daddr_t blkno); 91static int flushbufqueues(int flushdeps); 92static void buf_daemon(void); 93void bremfreel(struct buf *bp); 94 95int vmiodirenable = TRUE; 96SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 97 "Use the VM system for directory writes"); 98int runningbufspace; 99SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 100 "Amount of presently outstanding async buffer io"); 101static int bufspace; 102SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 103 "KVA memory used for bufs"); 104static int maxbufspace; 105SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 106 "Maximum allowed value of bufspace (including buf_daemon)"); 107static int bufmallocspace; 108SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 109 "Amount of malloced memory for buffers"); 110static int maxbufmallocspace; 111SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 112 "Maximum amount of malloced memory for buffers"); 113static int lobufspace; 114SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 115 "Minimum amount of buffers we want to have"); 116static int hibufspace; 117SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 118 "Maximum allowed value of bufspace (excluding buf_daemon)"); 119static int bufreusecnt; 120SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 121 "Number of times we have reused a buffer"); 122static int buffreekvacnt; 123SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 124 "Number of times we have freed the KVA space from some buffer"); 125static int bufdefragcnt; 126SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 127 "Number of times we have had to repeat buffer allocation to defragment"); 128static int lorunningspace; 129SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 130 "Minimum preferred space used for in-progress I/O"); 131static int hirunningspace; 132SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 133 "Maximum amount of space to use for in-progress I/O"); 134static int dirtybufferflushes; 135SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 136 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 137static int altbufferflushes; 138SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 139 0, "Number of fsync flushes to limit dirty buffers"); 140static int recursiveflushes; 141SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 142 0, "Number of flushes skipped due to being recursive"); 143static int numdirtybuffers; 144SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 145 "Number of buffers that are dirty (has unwritten changes) at the moment"); 146static int lodirtybuffers; 147SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 148 "How many buffers we want to have free before bufdaemon can sleep"); 149static int hidirtybuffers; 150SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 151 "When the number of dirty buffers is considered severe"); 152static int dirtybufthresh; 153SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 154 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 155static int numfreebuffers; 156SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 157 "Number of free buffers"); 158static int lofreebuffers; 159SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 160 "XXX Unused"); 161static int hifreebuffers; 162SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 163 "XXX Complicatedly unused"); 164static int getnewbufcalls; 165SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 166 "Number of calls to getnewbuf"); 167static int getnewbufrestarts; 168SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 169 "Number of times getnewbuf has had to restart a buffer aquisition"); 170static int dobkgrdwrite = 1; 171SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 172 "Do background writes (honoring the BV_BKGRDWRITE flag)?"); 173 174/* 175 * Wakeup point for bufdaemon, as well as indicator of whether it is already 176 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 177 * is idling. 178 */ 179static int bd_request; 180 181/* 182 * This lock synchronizes access to bd_request. 183 */ 184static struct mtx bdlock; 185 186/* 187 * bogus page -- for I/O to/from partially complete buffers 188 * this is a temporary solution to the problem, but it is not 189 * really that bad. it would be better to split the buffer 190 * for input in the case of buffers partially already in memory, 191 * but the code is intricate enough already. 192 */ 193vm_page_t bogus_page; 194 195/* 196 * Synchronization (sleep/wakeup) variable for active buffer space requests. 197 * Set when wait starts, cleared prior to wakeup(). 198 * Used in runningbufwakeup() and waitrunningbufspace(). 199 */ 200static int runningbufreq; 201 202/* 203 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 204 * waitrunningbufspace(). 205 */ 206static struct mtx rbreqlock; 207 208/* 209 * Synchronization (sleep/wakeup) variable for buffer requests. 210 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 211 * by and/or. 212 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 213 * getnewbuf(), and getblk(). 214 */ 215static int needsbuffer; 216 217/* 218 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 219 */ 220static struct mtx nblock; 221 222/* 223 * Lock that protects against bwait()/bdone()/B_DONE races. 224 */ 225 226static struct mtx bdonelock; 227 228/* 229 * Definitions for the buffer free lists. 230 */ 231#define BUFFER_QUEUES 5 /* number of free buffer queues */ 232 233#define QUEUE_NONE 0 /* on no queue */ 234#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 235#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 236#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 237#define QUEUE_EMPTY 4 /* empty buffer headers */ 238 239/* Queues for free buffers with various properties */ 240static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 241 242/* Lock for the bufqueues */ 243static struct mtx bqlock; 244 245/* 246 * Single global constant for BUF_WMESG, to avoid getting multiple references. 247 * buf_wmesg is referred from macros. 248 */ 249const char *buf_wmesg = BUF_WMESG; 250 251#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 252#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 253#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 254#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 255 256#ifdef DIRECTIO 257extern void ffs_rawread_setup(void); 258#endif /* DIRECTIO */ 259/* 260 * numdirtywakeup: 261 * 262 * If someone is blocked due to there being too many dirty buffers, 263 * and numdirtybuffers is now reasonable, wake them up. 264 */ 265 266static __inline void 267numdirtywakeup(int level) 268{ 269 270 if (numdirtybuffers <= level) { 271 mtx_lock(&nblock); 272 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 273 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 274 wakeup(&needsbuffer); 275 } 276 mtx_unlock(&nblock); 277 } 278} 279 280/* 281 * bufspacewakeup: 282 * 283 * Called when buffer space is potentially available for recovery. 284 * getnewbuf() will block on this flag when it is unable to free 285 * sufficient buffer space. Buffer space becomes recoverable when 286 * bp's get placed back in the queues. 287 */ 288 289static __inline void 290bufspacewakeup(void) 291{ 292 293 /* 294 * If someone is waiting for BUF space, wake them up. Even 295 * though we haven't freed the kva space yet, the waiting 296 * process will be able to now. 297 */ 298 mtx_lock(&nblock); 299 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 300 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 301 wakeup(&needsbuffer); 302 } 303 mtx_unlock(&nblock); 304} 305 306/* 307 * runningbufwakeup() - in-progress I/O accounting. 308 * 309 */ 310static __inline void 311runningbufwakeup(struct buf *bp) 312{ 313 314 if (bp->b_runningbufspace) { 315 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 316 bp->b_runningbufspace = 0; 317 mtx_lock(&rbreqlock); 318 if (runningbufreq && runningbufspace <= lorunningspace) { 319 runningbufreq = 0; 320 wakeup(&runningbufreq); 321 } 322 mtx_unlock(&rbreqlock); 323 } 324} 325 326/* 327 * bufcountwakeup: 328 * 329 * Called when a buffer has been added to one of the free queues to 330 * account for the buffer and to wakeup anyone waiting for free buffers. 331 * This typically occurs when large amounts of metadata are being handled 332 * by the buffer cache ( else buffer space runs out first, usually ). 333 */ 334 335static __inline void 336bufcountwakeup(void) 337{ 338 339 atomic_add_int(&numfreebuffers, 1); 340 mtx_lock(&nblock); 341 if (needsbuffer) { 342 needsbuffer &= ~VFS_BIO_NEED_ANY; 343 if (numfreebuffers >= hifreebuffers) 344 needsbuffer &= ~VFS_BIO_NEED_FREE; 345 wakeup(&needsbuffer); 346 } 347 mtx_unlock(&nblock); 348} 349 350/* 351 * waitrunningbufspace() 352 * 353 * runningbufspace is a measure of the amount of I/O currently 354 * running. This routine is used in async-write situations to 355 * prevent creating huge backups of pending writes to a device. 356 * Only asynchronous writes are governed by this function. 357 * 358 * Reads will adjust runningbufspace, but will not block based on it. 359 * The read load has a side effect of reducing the allowed write load. 360 * 361 * This does NOT turn an async write into a sync write. It waits 362 * for earlier writes to complete and generally returns before the 363 * caller's write has reached the device. 364 */ 365static __inline void 366waitrunningbufspace(void) 367{ 368 369 mtx_lock(&rbreqlock); 370 while (runningbufspace > hirunningspace) { 371 ++runningbufreq; 372 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 373 } 374 mtx_unlock(&rbreqlock); 375} 376 377 378/* 379 * vfs_buf_test_cache: 380 * 381 * Called when a buffer is extended. This function clears the B_CACHE 382 * bit if the newly extended portion of the buffer does not contain 383 * valid data. 384 */ 385static __inline 386void 387vfs_buf_test_cache(struct buf *bp, 388 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 389 vm_page_t m) 390{ 391 392 GIANT_REQUIRED; 393 394 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 395 if (bp->b_flags & B_CACHE) { 396 int base = (foff + off) & PAGE_MASK; 397 if (vm_page_is_valid(m, base, size) == 0) 398 bp->b_flags &= ~B_CACHE; 399 } 400} 401 402/* Wake up the buffer deamon if necessary */ 403static __inline 404void 405bd_wakeup(int dirtybuflevel) 406{ 407 408 mtx_lock(&bdlock); 409 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 410 bd_request = 1; 411 wakeup(&bd_request); 412 } 413 mtx_unlock(&bdlock); 414} 415 416/* 417 * bd_speedup - speedup the buffer cache flushing code 418 */ 419 420static __inline 421void 422bd_speedup(void) 423{ 424 425 bd_wakeup(1); 426} 427 428/* 429 * Calculating buffer cache scaling values and reserve space for buffer 430 * headers. This is called during low level kernel initialization and 431 * may be called more then once. We CANNOT write to the memory area 432 * being reserved at this time. 433 */ 434caddr_t 435kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 436{ 437 438 /* 439 * physmem_est is in pages. Convert it to kilobytes (assumes 440 * PAGE_SIZE is >= 1K) 441 */ 442 physmem_est = physmem_est * (PAGE_SIZE / 1024); 443 444 /* 445 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 446 * For the first 64MB of ram nominally allocate sufficient buffers to 447 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 448 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 449 * the buffer cache we limit the eventual kva reservation to 450 * maxbcache bytes. 451 * 452 * factor represents the 1/4 x ram conversion. 453 */ 454 if (nbuf == 0) { 455 int factor = 4 * BKVASIZE / 1024; 456 457 nbuf = 50; 458 if (physmem_est > 4096) 459 nbuf += min((physmem_est - 4096) / factor, 460 65536 / factor); 461 if (physmem_est > 65536) 462 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 463 464 if (maxbcache && nbuf > maxbcache / BKVASIZE) 465 nbuf = maxbcache / BKVASIZE; 466 } 467 468#if 0 469 /* 470 * Do not allow the buffer_map to be more then 1/2 the size of the 471 * kernel_map. 472 */ 473 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 474 (BKVASIZE * 2)) { 475 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 476 (BKVASIZE * 2); 477 printf("Warning: nbufs capped at %d\n", nbuf); 478 } 479#endif 480 481 /* 482 * swbufs are used as temporary holders for I/O, such as paging I/O. 483 * We have no less then 16 and no more then 256. 484 */ 485 nswbuf = max(min(nbuf/4, 256), 16); 486#ifdef NSWBUF_MIN 487 if (nswbuf < NSWBUF_MIN) 488 nswbuf = NSWBUF_MIN; 489#endif 490#ifdef DIRECTIO 491 ffs_rawread_setup(); 492#endif 493 494 /* 495 * Reserve space for the buffer cache buffers 496 */ 497 swbuf = (void *)v; 498 v = (caddr_t)(swbuf + nswbuf); 499 buf = (void *)v; 500 v = (caddr_t)(buf + nbuf); 501 502 return(v); 503} 504 505/* Initialize the buffer subsystem. Called before use of any buffers. */ 506void 507bufinit(void) 508{ 509 struct buf *bp; 510 int i; 511 512 GIANT_REQUIRED; 513 514 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 515 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 516 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 517 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 518 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 519 520 /* next, make a null set of free lists */ 521 for (i = 0; i < BUFFER_QUEUES; i++) 522 TAILQ_INIT(&bufqueues[i]); 523 524 /* finally, initialize each buffer header and stick on empty q */ 525 for (i = 0; i < nbuf; i++) { 526 bp = &buf[i]; 527 bzero(bp, sizeof *bp); 528 bp->b_flags = B_INVAL; /* we're just an empty header */ 529 bp->b_dev = NULL; 530 bp->b_rcred = NOCRED; 531 bp->b_wcred = NOCRED; 532 bp->b_qindex = QUEUE_EMPTY; 533 bp->b_vflags = 0; 534 bp->b_xflags = 0; 535 LIST_INIT(&bp->b_dep); 536 BUF_LOCKINIT(bp); 537 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 538 } 539 540 /* 541 * maxbufspace is the absolute maximum amount of buffer space we are 542 * allowed to reserve in KVM and in real terms. The absolute maximum 543 * is nominally used by buf_daemon. hibufspace is the nominal maximum 544 * used by most other processes. The differential is required to 545 * ensure that buf_daemon is able to run when other processes might 546 * be blocked waiting for buffer space. 547 * 548 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 549 * this may result in KVM fragmentation which is not handled optimally 550 * by the system. 551 */ 552 maxbufspace = nbuf * BKVASIZE; 553 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 554 lobufspace = hibufspace - MAXBSIZE; 555 556 lorunningspace = 512 * 1024; 557 hirunningspace = 1024 * 1024; 558 559/* 560 * Limit the amount of malloc memory since it is wired permanently into 561 * the kernel space. Even though this is accounted for in the buffer 562 * allocation, we don't want the malloced region to grow uncontrolled. 563 * The malloc scheme improves memory utilization significantly on average 564 * (small) directories. 565 */ 566 maxbufmallocspace = hibufspace / 20; 567 568/* 569 * Reduce the chance of a deadlock occuring by limiting the number 570 * of delayed-write dirty buffers we allow to stack up. 571 */ 572 hidirtybuffers = nbuf / 4 + 20; 573 dirtybufthresh = hidirtybuffers * 9 / 10; 574 numdirtybuffers = 0; 575/* 576 * To support extreme low-memory systems, make sure hidirtybuffers cannot 577 * eat up all available buffer space. This occurs when our minimum cannot 578 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 579 * BKVASIZE'd (8K) buffers. 580 */ 581 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 582 hidirtybuffers >>= 1; 583 } 584 lodirtybuffers = hidirtybuffers / 2; 585 586/* 587 * Try to keep the number of free buffers in the specified range, 588 * and give special processes (e.g. like buf_daemon) access to an 589 * emergency reserve. 590 */ 591 lofreebuffers = nbuf / 18 + 5; 592 hifreebuffers = 2 * lofreebuffers; 593 numfreebuffers = nbuf; 594 595/* 596 * Maximum number of async ops initiated per buf_daemon loop. This is 597 * somewhat of a hack at the moment, we really need to limit ourselves 598 * based on the number of bytes of I/O in-transit that were initiated 599 * from buf_daemon. 600 */ 601 602 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 603 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 604} 605 606/* 607 * bfreekva() - free the kva allocation for a buffer. 608 * 609 * Must be called at splbio() or higher as this is the only locking for 610 * buffer_map. 611 * 612 * Since this call frees up buffer space, we call bufspacewakeup(). 613 */ 614static void 615bfreekva(struct buf *bp) 616{ 617 618 GIANT_REQUIRED; 619 620 if (bp->b_kvasize) { 621 atomic_add_int(&buffreekvacnt, 1); 622 atomic_subtract_int(&bufspace, bp->b_kvasize); 623 vm_map_delete(buffer_map, 624 (vm_offset_t) bp->b_kvabase, 625 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 626 ); 627 bp->b_kvasize = 0; 628 bufspacewakeup(); 629 } 630} 631 632/* 633 * bremfree: 634 * 635 * Remove the buffer from the appropriate free list. 636 */ 637void 638bremfree(struct buf *bp) 639{ 640 641 mtx_lock(&bqlock); 642 bremfreel(bp); 643 mtx_unlock(&bqlock); 644} 645 646void 647bremfreel(struct buf *bp) 648{ 649 int s = splbio(); 650 int old_qindex = bp->b_qindex; 651 652 GIANT_REQUIRED; 653 654 if (bp->b_qindex != QUEUE_NONE) { 655 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 656 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 657 bp->b_qindex = QUEUE_NONE; 658 } else { 659 if (BUF_REFCNT(bp) <= 1) 660 panic("bremfree: removing a buffer not on a queue"); 661 } 662 663 /* 664 * Fixup numfreebuffers count. If the buffer is invalid or not 665 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 666 * the buffer was free and we must decrement numfreebuffers. 667 */ 668 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 669 switch(old_qindex) { 670 case QUEUE_DIRTY: 671 case QUEUE_CLEAN: 672 case QUEUE_EMPTY: 673 case QUEUE_EMPTYKVA: 674 atomic_subtract_int(&numfreebuffers, 1); 675 break; 676 default: 677 break; 678 } 679 } 680 splx(s); 681} 682 683 684/* 685 * Get a buffer with the specified data. Look in the cache first. We 686 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 687 * is set, the buffer is valid and we do not have to do anything ( see 688 * getblk() ). This is really just a special case of breadn(). 689 */ 690int 691bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 692 struct buf **bpp) 693{ 694 695 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 696} 697 698/* 699 * Operates like bread, but also starts asynchronous I/O on 700 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 701 * to initiating I/O . If B_CACHE is set, the buffer is valid 702 * and we do not have to do anything. 703 */ 704int 705breadn(struct vnode * vp, daddr_t blkno, int size, 706 daddr_t * rablkno, int *rabsize, 707 int cnt, struct ucred * cred, struct buf **bpp) 708{ 709 struct buf *bp, *rabp; 710 int i; 711 int rv = 0, readwait = 0; 712 713 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 714 715 /* if not found in cache, do some I/O */ 716 if ((bp->b_flags & B_CACHE) == 0) { 717 if (curthread != PCPU_GET(idlethread)) 718 curthread->td_proc->p_stats->p_ru.ru_inblock++; 719 bp->b_iocmd = BIO_READ; 720 bp->b_flags &= ~B_INVAL; 721 bp->b_ioflags &= ~BIO_ERROR; 722 if (bp->b_rcred == NOCRED && cred != NOCRED) 723 bp->b_rcred = crhold(cred); 724 vfs_busy_pages(bp, 0); 725 bp->b_iooffset = dbtob(bp->b_blkno); 726 if (vp->v_type == VCHR) 727 VOP_SPECSTRATEGY(vp, bp); 728 else 729 VOP_STRATEGY(vp, bp); 730 ++readwait; 731 } 732 733 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 734 if (inmem(vp, *rablkno)) 735 continue; 736 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 737 738 if ((rabp->b_flags & B_CACHE) == 0) { 739 if (curthread != PCPU_GET(idlethread)) 740 curthread->td_proc->p_stats->p_ru.ru_inblock++; 741 rabp->b_flags |= B_ASYNC; 742 rabp->b_flags &= ~B_INVAL; 743 rabp->b_ioflags &= ~BIO_ERROR; 744 rabp->b_iocmd = BIO_READ; 745 if (rabp->b_rcred == NOCRED && cred != NOCRED) 746 rabp->b_rcred = crhold(cred); 747 vfs_busy_pages(rabp, 0); 748 BUF_KERNPROC(rabp); 749 rabp->b_iooffset = dbtob(rabp->b_blkno); 750 if (vp->v_type == VCHR) 751 VOP_SPECSTRATEGY(vp, rabp); 752 else 753 VOP_STRATEGY(vp, rabp); 754 } else { 755 brelse(rabp); 756 } 757 } 758 759 if (readwait) { 760 rv = bufwait(bp); 761 } 762 return (rv); 763} 764 765/* 766 * Write, release buffer on completion. (Done by iodone 767 * if async). Do not bother writing anything if the buffer 768 * is invalid. 769 * 770 * Note that we set B_CACHE here, indicating that buffer is 771 * fully valid and thus cacheable. This is true even of NFS 772 * now so we set it generally. This could be set either here 773 * or in biodone() since the I/O is synchronous. We put it 774 * here. 775 */ 776int 777bwrite(struct buf *bp) 778{ 779 780 KASSERT(bp->b_op != NULL && bp->b_op->bop_write != NULL, 781 ("Martian buffer %p in bwrite: nobody to write it.", bp)); 782 return (bp->b_op->bop_write(bp)); 783} 784 785static int 786ibwrite(struct buf *bp) 787{ 788 int oldflags, s; 789 struct buf *newbp; 790 791 if (bp->b_flags & B_INVAL) { 792 brelse(bp); 793 return (0); 794 } 795 796 oldflags = bp->b_flags; 797 798 if (BUF_REFCNT(bp) == 0) 799 panic("ibwrite: buffer is not busy???"); 800 s = splbio(); 801 /* 802 * If a background write is already in progress, delay 803 * writing this block if it is asynchronous. Otherwise 804 * wait for the background write to complete. 805 */ 806 VI_LOCK(bp->b_vp); 807 if (bp->b_vflags & BV_BKGRDINPROG) { 808 if (bp->b_flags & B_ASYNC) { 809 VI_UNLOCK(bp->b_vp); 810 splx(s); 811 bdwrite(bp); 812 return (0); 813 } 814 bp->b_vflags |= BV_BKGRDWAIT; 815 msleep(&bp->b_xflags, VI_MTX(bp->b_vp), PRIBIO, "bwrbg", 0); 816 if (bp->b_vflags & BV_BKGRDINPROG) 817 panic("ibwrite: still writing"); 818 } 819 VI_UNLOCK(bp->b_vp); 820 821 /* Mark the buffer clean */ 822 bundirty(bp); 823 824 /* 825 * If this buffer is marked for background writing and we 826 * do not have to wait for it, make a copy and write the 827 * copy so as to leave this buffer ready for further use. 828 * 829 * This optimization eats a lot of memory. If we have a page 830 * or buffer shortfall we can't do it. 831 */ 832 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 833 (bp->b_flags & B_ASYNC) && 834 !vm_page_count_severe() && 835 !buf_dirty_count_severe()) { 836 if (bp->b_iodone != NULL) { 837 printf("bp->b_iodone = %p\n", bp->b_iodone); 838 panic("ibwrite: need chained iodone"); 839 } 840 841 /* get a new block */ 842 newbp = geteblk(bp->b_bufsize); 843 844 /* 845 * set it to be identical to the old block. We have to 846 * set b_lblkno and BKGRDMARKER before calling bgetvp() 847 * to avoid confusing the splay tree and gbincore(). 848 */ 849 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 850 newbp->b_lblkno = bp->b_lblkno; 851 newbp->b_xflags |= BX_BKGRDMARKER; 852 VI_LOCK(bp->b_vp); 853 bp->b_vflags |= BV_BKGRDINPROG; 854 bgetvp(bp->b_vp, newbp); 855 VI_UNLOCK(bp->b_vp); 856 newbp->b_blkno = bp->b_blkno; 857 newbp->b_offset = bp->b_offset; 858 newbp->b_iodone = vfs_backgroundwritedone; 859 newbp->b_flags |= B_ASYNC; 860 newbp->b_flags &= ~B_INVAL; 861 862 /* move over the dependencies */ 863 if (LIST_FIRST(&bp->b_dep) != NULL) 864 buf_movedeps(bp, newbp); 865 866 /* 867 * Initiate write on the copy, release the original to 868 * the B_LOCKED queue so that it cannot go away until 869 * the background write completes. If not locked it could go 870 * away and then be reconstituted while it was being written. 871 * If the reconstituted buffer were written, we could end up 872 * with two background copies being written at the same time. 873 */ 874 bqrelse(bp); 875 bp = newbp; 876 } 877 878 bp->b_flags &= ~B_DONE; 879 bp->b_ioflags &= ~BIO_ERROR; 880 bp->b_flags |= B_CACHE; 881 bp->b_iocmd = BIO_WRITE; 882 883 VI_LOCK(bp->b_vp); 884 bp->b_vp->v_numoutput++; 885 VI_UNLOCK(bp->b_vp); 886 vfs_busy_pages(bp, 1); 887 888 /* 889 * Normal bwrites pipeline writes 890 */ 891 bp->b_runningbufspace = bp->b_bufsize; 892 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 893 894 if (curthread != PCPU_GET(idlethread)) 895 curthread->td_proc->p_stats->p_ru.ru_oublock++; 896 splx(s); 897 if (oldflags & B_ASYNC) 898 BUF_KERNPROC(bp); 899 bp->b_iooffset = dbtob(bp->b_blkno); 900 if (bp->b_vp->v_type == VCHR) { 901 if (!buf_prewrite(bp->b_vp, bp)) 902 VOP_SPECSTRATEGY(bp->b_vp, bp); 903 } else { 904 VOP_STRATEGY(bp->b_vp, bp); 905 } 906 907 if ((oldflags & B_ASYNC) == 0) { 908 int rtval = bufwait(bp); 909 brelse(bp); 910 return (rtval); 911 } else { 912 /* 913 * don't allow the async write to saturate the I/O 914 * system. We will not deadlock here because 915 * we are blocking waiting for I/O that is already in-progress 916 * to complete. We do not block here if it is the update 917 * or syncer daemon trying to clean up as that can lead 918 * to deadlock. 919 */ 920 if (curthread->td_proc != bufdaemonproc && 921 curthread->td_proc != updateproc) 922 waitrunningbufspace(); 923 } 924 925 return (0); 926} 927 928/* 929 * Complete a background write started from bwrite. 930 */ 931static void 932vfs_backgroundwritedone(struct buf *bp) 933{ 934 struct buf *origbp; 935 936 /* 937 * Find the original buffer that we are writing. 938 */ 939 VI_LOCK(bp->b_vp); 940 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 941 panic("backgroundwritedone: lost buffer"); 942 943 /* 944 * Clear the BV_BKGRDINPROG flag in the original buffer 945 * and awaken it if it is waiting for the write to complete. 946 * If BV_BKGRDINPROG is not set in the original buffer it must 947 * have been released and re-instantiated - which is not legal. 948 */ 949 KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 950 ("backgroundwritedone: lost buffer2")); 951 origbp->b_vflags &= ~BV_BKGRDINPROG; 952 if (origbp->b_vflags & BV_BKGRDWAIT) { 953 origbp->b_vflags &= ~BV_BKGRDWAIT; 954 wakeup(&origbp->b_xflags); 955 } 956 VI_UNLOCK(bp->b_vp); 957 /* 958 * Process dependencies then return any unfinished ones. 959 */ 960 if (LIST_FIRST(&bp->b_dep) != NULL) 961 buf_complete(bp); 962 if (LIST_FIRST(&bp->b_dep) != NULL) 963 buf_movedeps(bp, origbp); 964 965 /* 966 * This buffer is marked B_NOCACHE, so when it is released 967 * by biodone, it will be tossed. We mark it with BIO_READ 968 * to avoid biodone doing a second vwakeup. 969 */ 970 bp->b_flags |= B_NOCACHE; 971 bp->b_iocmd = BIO_READ; 972 bp->b_flags &= ~(B_CACHE | B_DONE); 973 bp->b_iodone = 0; 974 bufdone(bp); 975} 976 977/* 978 * Delayed write. (Buffer is marked dirty). Do not bother writing 979 * anything if the buffer is marked invalid. 980 * 981 * Note that since the buffer must be completely valid, we can safely 982 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 983 * biodone() in order to prevent getblk from writing the buffer 984 * out synchronously. 985 */ 986void 987bdwrite(struct buf *bp) 988{ 989 struct thread *td = curthread; 990 struct vnode *vp; 991 struct buf *nbp; 992 993 GIANT_REQUIRED; 994 995 if (BUF_REFCNT(bp) == 0) 996 panic("bdwrite: buffer is not busy"); 997 998 if (bp->b_flags & B_INVAL) { 999 brelse(bp); 1000 return; 1001 } 1002 1003 /* 1004 * If we have too many dirty buffers, don't create any more. 1005 * If we are wildly over our limit, then force a complete 1006 * cleanup. Otherwise, just keep the situation from getting 1007 * out of control. Note that we have to avoid a recursive 1008 * disaster and not try to clean up after our own cleanup! 1009 */ 1010 vp = bp->b_vp; 1011 VI_LOCK(vp); 1012 if (td->td_pflags & TDP_COWINPROGRESS) { 1013 recursiveflushes++; 1014 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 1015 VI_UNLOCK(vp); 1016 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 1017 VI_LOCK(vp); 1018 altbufferflushes++; 1019 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 1020 /* 1021 * Try to find a buffer to flush. 1022 */ 1023 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_vnbufs) { 1024 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1025 buf_countdeps(nbp, 0) || 1026 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1027 continue; 1028 if (bp == nbp) 1029 panic("bdwrite: found ourselves"); 1030 VI_UNLOCK(vp); 1031 if (nbp->b_flags & B_CLUSTEROK) { 1032 vfs_bio_awrite(nbp); 1033 } else { 1034 bremfree(nbp); 1035 bawrite(nbp); 1036 } 1037 VI_LOCK(vp); 1038 dirtybufferflushes++; 1039 break; 1040 } 1041 } 1042 VI_UNLOCK(vp); 1043 1044 bdirty(bp); 1045 /* 1046 * Set B_CACHE, indicating that the buffer is fully valid. This is 1047 * true even of NFS now. 1048 */ 1049 bp->b_flags |= B_CACHE; 1050 1051 /* 1052 * This bmap keeps the system from needing to do the bmap later, 1053 * perhaps when the system is attempting to do a sync. Since it 1054 * is likely that the indirect block -- or whatever other datastructure 1055 * that the filesystem needs is still in memory now, it is a good 1056 * thing to do this. Note also, that if the pageout daemon is 1057 * requesting a sync -- there might not be enough memory to do 1058 * the bmap then... So, this is important to do. 1059 */ 1060 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1061 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1062 } 1063 1064 /* 1065 * Set the *dirty* buffer range based upon the VM system dirty pages. 1066 */ 1067 vfs_setdirty(bp); 1068 1069 /* 1070 * We need to do this here to satisfy the vnode_pager and the 1071 * pageout daemon, so that it thinks that the pages have been 1072 * "cleaned". Note that since the pages are in a delayed write 1073 * buffer -- the VFS layer "will" see that the pages get written 1074 * out on the next sync, or perhaps the cluster will be completed. 1075 */ 1076 vfs_clean_pages(bp); 1077 bqrelse(bp); 1078 1079 /* 1080 * Wakeup the buffer flushing daemon if we have a lot of dirty 1081 * buffers (midpoint between our recovery point and our stall 1082 * point). 1083 */ 1084 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1085 1086 /* 1087 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1088 * due to the softdep code. 1089 */ 1090} 1091 1092/* 1093 * bdirty: 1094 * 1095 * Turn buffer into delayed write request. We must clear BIO_READ and 1096 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1097 * itself to properly update it in the dirty/clean lists. We mark it 1098 * B_DONE to ensure that any asynchronization of the buffer properly 1099 * clears B_DONE ( else a panic will occur later ). 1100 * 1101 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1102 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1103 * should only be called if the buffer is known-good. 1104 * 1105 * Since the buffer is not on a queue, we do not update the numfreebuffers 1106 * count. 1107 * 1108 * Must be called at splbio(). 1109 * The buffer must be on QUEUE_NONE. 1110 */ 1111void 1112bdirty(struct buf *bp) 1113{ 1114 1115 KASSERT(bp->b_qindex == QUEUE_NONE, 1116 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1117 bp->b_flags &= ~(B_RELBUF); 1118 bp->b_iocmd = BIO_WRITE; 1119 1120 if ((bp->b_flags & B_DELWRI) == 0) { 1121 bp->b_flags |= B_DONE | B_DELWRI; 1122 reassignbuf(bp); 1123 atomic_add_int(&numdirtybuffers, 1); 1124 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1125 } 1126} 1127 1128/* 1129 * bundirty: 1130 * 1131 * Clear B_DELWRI for buffer. 1132 * 1133 * Since the buffer is not on a queue, we do not update the numfreebuffers 1134 * count. 1135 * 1136 * Must be called at splbio(). 1137 * The buffer must be on QUEUE_NONE. 1138 */ 1139 1140void 1141bundirty(struct buf *bp) 1142{ 1143 1144 KASSERT(bp->b_qindex == QUEUE_NONE, 1145 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1146 1147 if (bp->b_flags & B_DELWRI) { 1148 bp->b_flags &= ~B_DELWRI; 1149 reassignbuf(bp); 1150 atomic_subtract_int(&numdirtybuffers, 1); 1151 numdirtywakeup(lodirtybuffers); 1152 } 1153 /* 1154 * Since it is now being written, we can clear its deferred write flag. 1155 */ 1156 bp->b_flags &= ~B_DEFERRED; 1157} 1158 1159/* 1160 * bawrite: 1161 * 1162 * Asynchronous write. Start output on a buffer, but do not wait for 1163 * it to complete. The buffer is released when the output completes. 1164 * 1165 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1166 * B_INVAL buffers. Not us. 1167 */ 1168void 1169bawrite(struct buf *bp) 1170{ 1171 1172 bp->b_flags |= B_ASYNC; 1173 (void) bwrite(bp); 1174} 1175 1176/* 1177 * bwillwrite: 1178 * 1179 * Called prior to the locking of any vnodes when we are expecting to 1180 * write. We do not want to starve the buffer cache with too many 1181 * dirty buffers so we block here. By blocking prior to the locking 1182 * of any vnodes we attempt to avoid the situation where a locked vnode 1183 * prevents the various system daemons from flushing related buffers. 1184 */ 1185 1186void 1187bwillwrite(void) 1188{ 1189 1190 if (numdirtybuffers >= hidirtybuffers) { 1191 int s; 1192 1193 mtx_lock(&Giant); 1194 s = splbio(); 1195 mtx_lock(&nblock); 1196 while (numdirtybuffers >= hidirtybuffers) { 1197 bd_wakeup(1); 1198 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1199 msleep(&needsbuffer, &nblock, 1200 (PRIBIO + 4), "flswai", 0); 1201 } 1202 splx(s); 1203 mtx_unlock(&nblock); 1204 mtx_unlock(&Giant); 1205 } 1206} 1207 1208/* 1209 * Return true if we have too many dirty buffers. 1210 */ 1211int 1212buf_dirty_count_severe(void) 1213{ 1214 1215 return(numdirtybuffers >= hidirtybuffers); 1216} 1217 1218/* 1219 * brelse: 1220 * 1221 * Release a busy buffer and, if requested, free its resources. The 1222 * buffer will be stashed in the appropriate bufqueue[] allowing it 1223 * to be accessed later as a cache entity or reused for other purposes. 1224 */ 1225void 1226brelse(struct buf *bp) 1227{ 1228 int s; 1229 1230 GIANT_REQUIRED; 1231 1232 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1233 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1234 1235 s = splbio(); 1236 1237 if (bp->b_iocmd == BIO_WRITE && 1238 (bp->b_ioflags & BIO_ERROR) && 1239 !(bp->b_flags & B_INVAL)) { 1240 /* 1241 * Failed write, redirty. Must clear BIO_ERROR to prevent 1242 * pages from being scrapped. If B_INVAL is set then 1243 * this case is not run and the next case is run to 1244 * destroy the buffer. B_INVAL can occur if the buffer 1245 * is outside the range supported by the underlying device. 1246 */ 1247 bp->b_ioflags &= ~BIO_ERROR; 1248 bdirty(bp); 1249 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1250 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1251 /* 1252 * Either a failed I/O or we were asked to free or not 1253 * cache the buffer. 1254 */ 1255 bp->b_flags |= B_INVAL; 1256 if (LIST_FIRST(&bp->b_dep) != NULL) 1257 buf_deallocate(bp); 1258 if (bp->b_flags & B_DELWRI) { 1259 atomic_subtract_int(&numdirtybuffers, 1); 1260 numdirtywakeup(lodirtybuffers); 1261 } 1262 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1263 if ((bp->b_flags & B_VMIO) == 0) { 1264 if (bp->b_bufsize) 1265 allocbuf(bp, 0); 1266 if (bp->b_vp) 1267 brelvp(bp); 1268 } 1269 } 1270 1271 /* 1272 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1273 * is called with B_DELWRI set, the underlying pages may wind up 1274 * getting freed causing a previous write (bdwrite()) to get 'lost' 1275 * because pages associated with a B_DELWRI bp are marked clean. 1276 * 1277 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1278 * if B_DELWRI is set. 1279 * 1280 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1281 * on pages to return pages to the VM page queues. 1282 */ 1283 if (bp->b_flags & B_DELWRI) 1284 bp->b_flags &= ~B_RELBUF; 1285 else if (vm_page_count_severe()) { 1286 /* 1287 * XXX This lock may not be necessary since BKGRDINPROG 1288 * cannot be set while we hold the buf lock, it can only be 1289 * cleared if it is already pending. 1290 */ 1291 if (bp->b_vp) { 1292 VI_LOCK(bp->b_vp); 1293 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1294 bp->b_flags |= B_RELBUF; 1295 VI_UNLOCK(bp->b_vp); 1296 } else 1297 bp->b_flags |= B_RELBUF; 1298 } 1299 1300 /* 1301 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1302 * constituted, not even NFS buffers now. Two flags effect this. If 1303 * B_INVAL, the struct buf is invalidated but the VM object is kept 1304 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1305 * 1306 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1307 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1308 * buffer is also B_INVAL because it hits the re-dirtying code above. 1309 * 1310 * Normally we can do this whether a buffer is B_DELWRI or not. If 1311 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1312 * the commit state and we cannot afford to lose the buffer. If the 1313 * buffer has a background write in progress, we need to keep it 1314 * around to prevent it from being reconstituted and starting a second 1315 * background write. 1316 */ 1317 if ((bp->b_flags & B_VMIO) 1318 && !(bp->b_vp->v_mount != NULL && 1319 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1320 !vn_isdisk(bp->b_vp, NULL) && 1321 (bp->b_flags & B_DELWRI)) 1322 ) { 1323 1324 int i, j, resid; 1325 vm_page_t m; 1326 off_t foff; 1327 vm_pindex_t poff; 1328 vm_object_t obj; 1329 1330 obj = bp->b_object; 1331 1332 /* 1333 * Get the base offset and length of the buffer. Note that 1334 * in the VMIO case if the buffer block size is not 1335 * page-aligned then b_data pointer may not be page-aligned. 1336 * But our b_pages[] array *IS* page aligned. 1337 * 1338 * block sizes less then DEV_BSIZE (usually 512) are not 1339 * supported due to the page granularity bits (m->valid, 1340 * m->dirty, etc...). 1341 * 1342 * See man buf(9) for more information 1343 */ 1344 resid = bp->b_bufsize; 1345 foff = bp->b_offset; 1346 VM_OBJECT_LOCK(obj); 1347 for (i = 0; i < bp->b_npages; i++) { 1348 int had_bogus = 0; 1349 1350 m = bp->b_pages[i]; 1351 1352 /* 1353 * If we hit a bogus page, fixup *all* the bogus pages 1354 * now. 1355 */ 1356 if (m == bogus_page) { 1357 poff = OFF_TO_IDX(bp->b_offset); 1358 had_bogus = 1; 1359 1360 for (j = i; j < bp->b_npages; j++) { 1361 vm_page_t mtmp; 1362 mtmp = bp->b_pages[j]; 1363 if (mtmp == bogus_page) { 1364 mtmp = vm_page_lookup(obj, poff + j); 1365 if (!mtmp) { 1366 panic("brelse: page missing\n"); 1367 } 1368 bp->b_pages[j] = mtmp; 1369 } 1370 } 1371 1372 if ((bp->b_flags & B_INVAL) == 0) { 1373 pmap_qenter( 1374 trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1375 } 1376 m = bp->b_pages[i]; 1377 } 1378 if ((bp->b_flags & B_NOCACHE) || 1379 (bp->b_ioflags & BIO_ERROR)) { 1380 int poffset = foff & PAGE_MASK; 1381 int presid = resid > (PAGE_SIZE - poffset) ? 1382 (PAGE_SIZE - poffset) : resid; 1383 1384 KASSERT(presid >= 0, ("brelse: extra page")); 1385 vm_page_lock_queues(); 1386 vm_page_set_invalid(m, poffset, presid); 1387 vm_page_unlock_queues(); 1388 if (had_bogus) 1389 printf("avoided corruption bug in bogus_page/brelse code\n"); 1390 } 1391 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1392 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1393 } 1394 VM_OBJECT_UNLOCK(obj); 1395 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1396 vfs_vmio_release(bp); 1397 1398 } else if (bp->b_flags & B_VMIO) { 1399 1400 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1401 vfs_vmio_release(bp); 1402 } 1403 1404 } 1405 1406 if (bp->b_qindex != QUEUE_NONE) 1407 panic("brelse: free buffer onto another queue???"); 1408 if (BUF_REFCNT(bp) > 1) { 1409 /* do not release to free list */ 1410 BUF_UNLOCK(bp); 1411 splx(s); 1412 return; 1413 } 1414 1415 /* enqueue */ 1416 mtx_lock(&bqlock); 1417 1418 /* buffers with no memory */ 1419 if (bp->b_bufsize == 0) { 1420 bp->b_flags |= B_INVAL; 1421 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1422 if (bp->b_vflags & BV_BKGRDINPROG) 1423 panic("losing buffer 1"); 1424 if (bp->b_kvasize) { 1425 bp->b_qindex = QUEUE_EMPTYKVA; 1426 } else { 1427 bp->b_qindex = QUEUE_EMPTY; 1428 } 1429 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1430 bp->b_dev = NULL; 1431 /* buffers with junk contents */ 1432 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1433 (bp->b_ioflags & BIO_ERROR)) { 1434 bp->b_flags |= B_INVAL; 1435 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1436 if (bp->b_vflags & BV_BKGRDINPROG) 1437 panic("losing buffer 2"); 1438 bp->b_qindex = QUEUE_CLEAN; 1439 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1440 bp->b_dev = NULL; 1441 /* remaining buffers */ 1442 } else { 1443 if (bp->b_flags & B_DELWRI) 1444 bp->b_qindex = QUEUE_DIRTY; 1445 else 1446 bp->b_qindex = QUEUE_CLEAN; 1447 if (bp->b_flags & B_AGE) 1448 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1449 else 1450 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1451 } 1452 mtx_unlock(&bqlock); 1453 1454 /* 1455 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1456 * placed the buffer on the correct queue. We must also disassociate 1457 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1458 * find it. 1459 */ 1460 if (bp->b_flags & B_INVAL) { 1461 if (bp->b_flags & B_DELWRI) 1462 bundirty(bp); 1463 if (bp->b_vp) 1464 brelvp(bp); 1465 } 1466 1467 /* 1468 * Fixup numfreebuffers count. The bp is on an appropriate queue 1469 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1470 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1471 * if B_INVAL is set ). 1472 */ 1473 1474 if (!(bp->b_flags & B_DELWRI)) 1475 bufcountwakeup(); 1476 1477 /* 1478 * Something we can maybe free or reuse 1479 */ 1480 if (bp->b_bufsize || bp->b_kvasize) 1481 bufspacewakeup(); 1482 1483 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1484 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1485 panic("brelse: not dirty"); 1486 /* unlock */ 1487 BUF_UNLOCK(bp); 1488 splx(s); 1489} 1490 1491/* 1492 * Release a buffer back to the appropriate queue but do not try to free 1493 * it. The buffer is expected to be used again soon. 1494 * 1495 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1496 * biodone() to requeue an async I/O on completion. It is also used when 1497 * known good buffers need to be requeued but we think we may need the data 1498 * again soon. 1499 * 1500 * XXX we should be able to leave the B_RELBUF hint set on completion. 1501 */ 1502void 1503bqrelse(struct buf *bp) 1504{ 1505 int s; 1506 1507 s = splbio(); 1508 1509 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1510 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1511 1512 if (bp->b_qindex != QUEUE_NONE) 1513 panic("bqrelse: free buffer onto another queue???"); 1514 if (BUF_REFCNT(bp) > 1) { 1515 /* do not release to free list */ 1516 BUF_UNLOCK(bp); 1517 splx(s); 1518 return; 1519 } 1520 mtx_lock(&bqlock); 1521 /* buffers with stale but valid contents */ 1522 if (bp->b_flags & B_DELWRI) { 1523 bp->b_qindex = QUEUE_DIRTY; 1524 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1525 } else { 1526 /* 1527 * XXX This lock may not be necessary since BKGRDINPROG 1528 * cannot be set while we hold the buf lock, it can only be 1529 * cleared if it is already pending. 1530 */ 1531 VI_LOCK(bp->b_vp); 1532 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1533 VI_UNLOCK(bp->b_vp); 1534 bp->b_qindex = QUEUE_CLEAN; 1535 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1536 b_freelist); 1537 } else { 1538 /* 1539 * We are too low on memory, we have to try to free 1540 * the buffer (most importantly: the wired pages 1541 * making up its backing store) *now*. 1542 */ 1543 VI_UNLOCK(bp->b_vp); 1544 mtx_unlock(&bqlock); 1545 splx(s); 1546 brelse(bp); 1547 return; 1548 } 1549 } 1550 mtx_unlock(&bqlock); 1551 1552 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1553 bufcountwakeup(); 1554 1555 /* 1556 * Something we can maybe free or reuse. 1557 */ 1558 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1559 bufspacewakeup(); 1560 1561 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1562 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1563 panic("bqrelse: not dirty"); 1564 /* unlock */ 1565 BUF_UNLOCK(bp); 1566 splx(s); 1567} 1568 1569/* Give pages used by the bp back to the VM system (where possible) */ 1570static void 1571vfs_vmio_release(struct buf *bp) 1572{ 1573 int i; 1574 vm_page_t m; 1575 1576 GIANT_REQUIRED; 1577 VM_OBJECT_LOCK(bp->b_object); 1578 vm_page_lock_queues(); 1579 for (i = 0; i < bp->b_npages; i++) { 1580 m = bp->b_pages[i]; 1581 bp->b_pages[i] = NULL; 1582 /* 1583 * In order to keep page LRU ordering consistent, put 1584 * everything on the inactive queue. 1585 */ 1586 vm_page_unwire(m, 0); 1587 /* 1588 * We don't mess with busy pages, it is 1589 * the responsibility of the process that 1590 * busied the pages to deal with them. 1591 */ 1592 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1593 continue; 1594 1595 if (m->wire_count == 0) { 1596 /* 1597 * Might as well free the page if we can and it has 1598 * no valid data. We also free the page if the 1599 * buffer was used for direct I/O 1600 */ 1601 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1602 m->hold_count == 0) { 1603 vm_page_busy(m); 1604 pmap_remove_all(m); 1605 vm_page_free(m); 1606 } else if (bp->b_flags & B_DIRECT) { 1607 vm_page_try_to_free(m); 1608 } else if (vm_page_count_severe()) { 1609 vm_page_try_to_cache(m); 1610 } 1611 } 1612 } 1613 vm_page_unlock_queues(); 1614 VM_OBJECT_UNLOCK(bp->b_object); 1615 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1616 1617 if (bp->b_bufsize) { 1618 bufspacewakeup(); 1619 bp->b_bufsize = 0; 1620 } 1621 bp->b_npages = 0; 1622 bp->b_flags &= ~B_VMIO; 1623 if (bp->b_vp) 1624 brelvp(bp); 1625} 1626 1627/* 1628 * Check to see if a block at a particular lbn is available for a clustered 1629 * write. 1630 */ 1631static int 1632vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1633{ 1634 struct buf *bpa; 1635 int match; 1636 1637 match = 0; 1638 1639 /* If the buf isn't in core skip it */ 1640 if ((bpa = gbincore(vp, lblkno)) == NULL) 1641 return (0); 1642 1643 /* If the buf is busy we don't want to wait for it */ 1644 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1645 return (0); 1646 1647 /* Only cluster with valid clusterable delayed write buffers */ 1648 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1649 (B_DELWRI | B_CLUSTEROK)) 1650 goto done; 1651 1652 if (bpa->b_bufsize != size) 1653 goto done; 1654 1655 /* 1656 * Check to see if it is in the expected place on disk and that the 1657 * block has been mapped. 1658 */ 1659 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1660 match = 1; 1661done: 1662 BUF_UNLOCK(bpa); 1663 return (match); 1664} 1665 1666/* 1667 * vfs_bio_awrite: 1668 * 1669 * Implement clustered async writes for clearing out B_DELWRI buffers. 1670 * This is much better then the old way of writing only one buffer at 1671 * a time. Note that we may not be presented with the buffers in the 1672 * correct order, so we search for the cluster in both directions. 1673 */ 1674int 1675vfs_bio_awrite(struct buf *bp) 1676{ 1677 int i; 1678 int j; 1679 daddr_t lblkno = bp->b_lblkno; 1680 struct vnode *vp = bp->b_vp; 1681 int s; 1682 int ncl; 1683 int nwritten; 1684 int size; 1685 int maxcl; 1686 1687 s = splbio(); 1688 /* 1689 * right now we support clustered writing only to regular files. If 1690 * we find a clusterable block we could be in the middle of a cluster 1691 * rather then at the beginning. 1692 */ 1693 if ((vp->v_type == VREG) && 1694 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1695 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1696 1697 size = vp->v_mount->mnt_stat.f_iosize; 1698 maxcl = MAXPHYS / size; 1699 1700 VI_LOCK(vp); 1701 for (i = 1; i < maxcl; i++) 1702 if (vfs_bio_clcheck(vp, size, lblkno + i, 1703 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1704 break; 1705 1706 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1707 if (vfs_bio_clcheck(vp, size, lblkno - j, 1708 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1709 break; 1710 1711 VI_UNLOCK(vp); 1712 --j; 1713 ncl = i + j; 1714 /* 1715 * this is a possible cluster write 1716 */ 1717 if (ncl != 1) { 1718 BUF_UNLOCK(bp); 1719 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1720 splx(s); 1721 return nwritten; 1722 } 1723 } 1724 1725 bremfree(bp); 1726 bp->b_flags |= B_ASYNC; 1727 1728 splx(s); 1729 /* 1730 * default (old) behavior, writing out only one block 1731 * 1732 * XXX returns b_bufsize instead of b_bcount for nwritten? 1733 */ 1734 nwritten = bp->b_bufsize; 1735 (void) bwrite(bp); 1736 1737 return nwritten; 1738} 1739 1740/* 1741 * getnewbuf: 1742 * 1743 * Find and initialize a new buffer header, freeing up existing buffers 1744 * in the bufqueues as necessary. The new buffer is returned locked. 1745 * 1746 * Important: B_INVAL is not set. If the caller wishes to throw the 1747 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1748 * 1749 * We block if: 1750 * We have insufficient buffer headers 1751 * We have insufficient buffer space 1752 * buffer_map is too fragmented ( space reservation fails ) 1753 * If we have to flush dirty buffers ( but we try to avoid this ) 1754 * 1755 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1756 * Instead we ask the buf daemon to do it for us. We attempt to 1757 * avoid piecemeal wakeups of the pageout daemon. 1758 */ 1759 1760static struct buf * 1761getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1762{ 1763 struct buf *bp; 1764 struct buf *nbp; 1765 int defrag = 0; 1766 int nqindex; 1767 static int flushingbufs; 1768 1769 GIANT_REQUIRED; 1770 1771 /* 1772 * We can't afford to block since we might be holding a vnode lock, 1773 * which may prevent system daemons from running. We deal with 1774 * low-memory situations by proactively returning memory and running 1775 * async I/O rather then sync I/O. 1776 */ 1777 1778 atomic_add_int(&getnewbufcalls, 1); 1779 atomic_subtract_int(&getnewbufrestarts, 1); 1780restart: 1781 atomic_add_int(&getnewbufrestarts, 1); 1782 1783 /* 1784 * Setup for scan. If we do not have enough free buffers, 1785 * we setup a degenerate case that immediately fails. Note 1786 * that if we are specially marked process, we are allowed to 1787 * dip into our reserves. 1788 * 1789 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1790 * 1791 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1792 * However, there are a number of cases (defragging, reusing, ...) 1793 * where we cannot backup. 1794 */ 1795 mtx_lock(&bqlock); 1796 nqindex = QUEUE_EMPTYKVA; 1797 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1798 1799 if (nbp == NULL) { 1800 /* 1801 * If no EMPTYKVA buffers and we are either 1802 * defragging or reusing, locate a CLEAN buffer 1803 * to free or reuse. If bufspace useage is low 1804 * skip this step so we can allocate a new buffer. 1805 */ 1806 if (defrag || bufspace >= lobufspace) { 1807 nqindex = QUEUE_CLEAN; 1808 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1809 } 1810 1811 /* 1812 * If we could not find or were not allowed to reuse a 1813 * CLEAN buffer, check to see if it is ok to use an EMPTY 1814 * buffer. We can only use an EMPTY buffer if allocating 1815 * its KVA would not otherwise run us out of buffer space. 1816 */ 1817 if (nbp == NULL && defrag == 0 && 1818 bufspace + maxsize < hibufspace) { 1819 nqindex = QUEUE_EMPTY; 1820 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1821 } 1822 } 1823 1824 /* 1825 * Run scan, possibly freeing data and/or kva mappings on the fly 1826 * depending. 1827 */ 1828 1829 while ((bp = nbp) != NULL) { 1830 int qindex = nqindex; 1831 1832 /* 1833 * Calculate next bp ( we can only use it if we do not block 1834 * or do other fancy things ). 1835 */ 1836 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1837 switch(qindex) { 1838 case QUEUE_EMPTY: 1839 nqindex = QUEUE_EMPTYKVA; 1840 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1841 break; 1842 /* FALLTHROUGH */ 1843 case QUEUE_EMPTYKVA: 1844 nqindex = QUEUE_CLEAN; 1845 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1846 break; 1847 /* FALLTHROUGH */ 1848 case QUEUE_CLEAN: 1849 /* 1850 * nbp is NULL. 1851 */ 1852 break; 1853 } 1854 } 1855 if (bp->b_vp) { 1856 VI_LOCK(bp->b_vp); 1857 if (bp->b_vflags & BV_BKGRDINPROG) { 1858 VI_UNLOCK(bp->b_vp); 1859 continue; 1860 } 1861 VI_UNLOCK(bp->b_vp); 1862 } 1863 1864 /* 1865 * Sanity Checks 1866 */ 1867 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1868 1869 /* 1870 * Note: we no longer distinguish between VMIO and non-VMIO 1871 * buffers. 1872 */ 1873 1874 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1875 1876 /* 1877 * If we are defragging then we need a buffer with 1878 * b_kvasize != 0. XXX this situation should no longer 1879 * occur, if defrag is non-zero the buffer's b_kvasize 1880 * should also be non-zero at this point. XXX 1881 */ 1882 if (defrag && bp->b_kvasize == 0) { 1883 printf("Warning: defrag empty buffer %p\n", bp); 1884 continue; 1885 } 1886 1887 /* 1888 * Start freeing the bp. This is somewhat involved. nbp 1889 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1890 */ 1891 1892 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1893 panic("getnewbuf: locked buf"); 1894 bremfreel(bp); 1895 mtx_unlock(&bqlock); 1896 1897 if (qindex == QUEUE_CLEAN) { 1898 if (bp->b_flags & B_VMIO) { 1899 bp->b_flags &= ~B_ASYNC; 1900 vfs_vmio_release(bp); 1901 } 1902 if (bp->b_vp) 1903 brelvp(bp); 1904 } 1905 1906 /* 1907 * NOTE: nbp is now entirely invalid. We can only restart 1908 * the scan from this point on. 1909 * 1910 * Get the rest of the buffer freed up. b_kva* is still 1911 * valid after this operation. 1912 */ 1913 1914 if (bp->b_rcred != NOCRED) { 1915 crfree(bp->b_rcred); 1916 bp->b_rcred = NOCRED; 1917 } 1918 if (bp->b_wcred != NOCRED) { 1919 crfree(bp->b_wcred); 1920 bp->b_wcred = NOCRED; 1921 } 1922 if (LIST_FIRST(&bp->b_dep) != NULL) 1923 buf_deallocate(bp); 1924 if (bp->b_vflags & BV_BKGRDINPROG) 1925 panic("losing buffer 3"); 1926 1927 if (bp->b_bufsize) 1928 allocbuf(bp, 0); 1929 1930 bp->b_flags = 0; 1931 bp->b_ioflags = 0; 1932 bp->b_xflags = 0; 1933 bp->b_vflags = 0; 1934 bp->b_dev = NULL; 1935 bp->b_vp = NULL; 1936 bp->b_blkno = bp->b_lblkno = 0; 1937 bp->b_offset = NOOFFSET; 1938 bp->b_iodone = 0; 1939 bp->b_error = 0; 1940 bp->b_resid = 0; 1941 bp->b_bcount = 0; 1942 bp->b_npages = 0; 1943 bp->b_dirtyoff = bp->b_dirtyend = 0; 1944 bp->b_magic = B_MAGIC_BIO; 1945 bp->b_op = &buf_ops_bio; 1946 bp->b_object = NULL; 1947 1948 LIST_INIT(&bp->b_dep); 1949 1950 /* 1951 * If we are defragging then free the buffer. 1952 */ 1953 if (defrag) { 1954 bp->b_flags |= B_INVAL; 1955 bfreekva(bp); 1956 brelse(bp); 1957 defrag = 0; 1958 goto restart; 1959 } 1960 1961 /* 1962 * If we are overcomitted then recover the buffer and its 1963 * KVM space. This occurs in rare situations when multiple 1964 * processes are blocked in getnewbuf() or allocbuf(). 1965 */ 1966 if (bufspace >= hibufspace) 1967 flushingbufs = 1; 1968 if (flushingbufs && bp->b_kvasize != 0) { 1969 bp->b_flags |= B_INVAL; 1970 bfreekva(bp); 1971 brelse(bp); 1972 goto restart; 1973 } 1974 if (bufspace < lobufspace) 1975 flushingbufs = 0; 1976 break; 1977 } 1978 1979 /* 1980 * If we exhausted our list, sleep as appropriate. We may have to 1981 * wakeup various daemons and write out some dirty buffers. 1982 * 1983 * Generally we are sleeping due to insufficient buffer space. 1984 */ 1985 1986 if (bp == NULL) { 1987 int flags; 1988 char *waitmsg; 1989 1990 mtx_unlock(&bqlock); 1991 if (defrag) { 1992 flags = VFS_BIO_NEED_BUFSPACE; 1993 waitmsg = "nbufkv"; 1994 } else if (bufspace >= hibufspace) { 1995 waitmsg = "nbufbs"; 1996 flags = VFS_BIO_NEED_BUFSPACE; 1997 } else { 1998 waitmsg = "newbuf"; 1999 flags = VFS_BIO_NEED_ANY; 2000 } 2001 2002 bd_speedup(); /* heeeelp */ 2003 2004 mtx_lock(&nblock); 2005 needsbuffer |= flags; 2006 while (needsbuffer & flags) { 2007 if (msleep(&needsbuffer, &nblock, 2008 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2009 mtx_unlock(&nblock); 2010 return (NULL); 2011 } 2012 } 2013 mtx_unlock(&nblock); 2014 } else { 2015 /* 2016 * We finally have a valid bp. We aren't quite out of the 2017 * woods, we still have to reserve kva space. In order 2018 * to keep fragmentation sane we only allocate kva in 2019 * BKVASIZE chunks. 2020 */ 2021 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2022 2023 if (maxsize != bp->b_kvasize) { 2024 vm_offset_t addr = 0; 2025 2026 bfreekva(bp); 2027 2028 if (vm_map_findspace(buffer_map, 2029 vm_map_min(buffer_map), maxsize, &addr)) { 2030 /* 2031 * Uh oh. Buffer map is to fragmented. We 2032 * must defragment the map. 2033 */ 2034 atomic_add_int(&bufdefragcnt, 1); 2035 defrag = 1; 2036 bp->b_flags |= B_INVAL; 2037 brelse(bp); 2038 goto restart; 2039 } 2040 if (addr) { 2041 vm_map_insert(buffer_map, NULL, 0, 2042 addr, addr + maxsize, 2043 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2044 2045 bp->b_kvabase = (caddr_t) addr; 2046 bp->b_kvasize = maxsize; 2047 atomic_add_int(&bufspace, bp->b_kvasize); 2048 atomic_add_int(&bufreusecnt, 1); 2049 } 2050 } 2051 bp->b_saveaddr = bp->b_kvabase; 2052 bp->b_data = bp->b_saveaddr; 2053 } 2054 return(bp); 2055} 2056 2057/* 2058 * buf_daemon: 2059 * 2060 * buffer flushing daemon. Buffers are normally flushed by the 2061 * update daemon but if it cannot keep up this process starts to 2062 * take the load in an attempt to prevent getnewbuf() from blocking. 2063 */ 2064 2065static struct kproc_desc buf_kp = { 2066 "bufdaemon", 2067 buf_daemon, 2068 &bufdaemonproc 2069}; 2070SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2071 2072static void 2073buf_daemon() 2074{ 2075 int s; 2076 2077 mtx_lock(&Giant); 2078 2079 /* 2080 * This process needs to be suspended prior to shutdown sync. 2081 */ 2082 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2083 SHUTDOWN_PRI_LAST); 2084 2085 /* 2086 * This process is allowed to take the buffer cache to the limit 2087 */ 2088 s = splbio(); 2089 mtx_lock(&bdlock); 2090 2091 for (;;) { 2092 bd_request = 0; 2093 mtx_unlock(&bdlock); 2094 2095 kthread_suspend_check(bufdaemonproc); 2096 2097 /* 2098 * Do the flush. Limit the amount of in-transit I/O we 2099 * allow to build up, otherwise we would completely saturate 2100 * the I/O system. Wakeup any waiting processes before we 2101 * normally would so they can run in parallel with our drain. 2102 */ 2103 while (numdirtybuffers > lodirtybuffers) { 2104 if (flushbufqueues(0) == 0) { 2105 /* 2106 * Could not find any buffers without rollback 2107 * dependencies, so just write the first one 2108 * in the hopes of eventually making progress. 2109 */ 2110 flushbufqueues(1); 2111 break; 2112 } 2113 waitrunningbufspace(); 2114 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2115 } 2116 2117 /* 2118 * Only clear bd_request if we have reached our low water 2119 * mark. The buf_daemon normally waits 1 second and 2120 * then incrementally flushes any dirty buffers that have 2121 * built up, within reason. 2122 * 2123 * If we were unable to hit our low water mark and couldn't 2124 * find any flushable buffers, we sleep half a second. 2125 * Otherwise we loop immediately. 2126 */ 2127 mtx_lock(&bdlock); 2128 if (numdirtybuffers <= lodirtybuffers) { 2129 /* 2130 * We reached our low water mark, reset the 2131 * request and sleep until we are needed again. 2132 * The sleep is just so the suspend code works. 2133 */ 2134 bd_request = 0; 2135 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2136 } else { 2137 /* 2138 * We couldn't find any flushable dirty buffers but 2139 * still have too many dirty buffers, we 2140 * have to sleep and try again. (rare) 2141 */ 2142 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2143 } 2144 } 2145} 2146 2147/* 2148 * flushbufqueues: 2149 * 2150 * Try to flush a buffer in the dirty queue. We must be careful to 2151 * free up B_INVAL buffers instead of write them, which NFS is 2152 * particularly sensitive to. 2153 */ 2154int flushwithdeps = 0; 2155SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2156 0, "Number of buffers flushed with dependecies that require rollbacks"); 2157 2158static int 2159flushbufqueues(int flushdeps) 2160{ 2161 struct thread *td = curthread; 2162 struct vnode *vp; 2163 struct mount *mp; 2164 struct buf *bp; 2165 int hasdeps; 2166 2167 mtx_lock(&bqlock); 2168 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2169 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2170 continue; 2171 KASSERT((bp->b_flags & B_DELWRI), 2172 ("unexpected clean buffer %p", bp)); 2173 VI_LOCK(bp->b_vp); 2174 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2175 VI_UNLOCK(bp->b_vp); 2176 BUF_UNLOCK(bp); 2177 continue; 2178 } 2179 VI_UNLOCK(bp->b_vp); 2180 if (bp->b_flags & B_INVAL) { 2181 bremfreel(bp); 2182 mtx_unlock(&bqlock); 2183 brelse(bp); 2184 return (1); 2185 } 2186 2187 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2188 if (flushdeps == 0) { 2189 BUF_UNLOCK(bp); 2190 continue; 2191 } 2192 hasdeps = 1; 2193 } else 2194 hasdeps = 0; 2195 /* 2196 * We must hold the lock on a vnode before writing 2197 * one of its buffers. Otherwise we may confuse, or 2198 * in the case of a snapshot vnode, deadlock the 2199 * system. 2200 * 2201 * The lock order here is the reverse of the normal 2202 * of vnode followed by buf lock. This is ok because 2203 * the NOWAIT will prevent deadlock. 2204 */ 2205 vp = bp->b_vp; 2206 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2207 BUF_UNLOCK(bp); 2208 continue; 2209 } 2210 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2211 mtx_unlock(&bqlock); 2212 vfs_bio_awrite(bp); 2213 vn_finished_write(mp); 2214 VOP_UNLOCK(vp, 0, td); 2215 flushwithdeps += hasdeps; 2216 return (1); 2217 } 2218 vn_finished_write(mp); 2219 BUF_UNLOCK(bp); 2220 } 2221 mtx_unlock(&bqlock); 2222 return (0); 2223} 2224 2225/* 2226 * Check to see if a block is currently memory resident. 2227 */ 2228struct buf * 2229incore(struct vnode * vp, daddr_t blkno) 2230{ 2231 struct buf *bp; 2232 2233 int s = splbio(); 2234 VI_LOCK(vp); 2235 bp = gbincore(vp, blkno); 2236 VI_UNLOCK(vp); 2237 splx(s); 2238 return (bp); 2239} 2240 2241/* 2242 * Returns true if no I/O is needed to access the 2243 * associated VM object. This is like incore except 2244 * it also hunts around in the VM system for the data. 2245 */ 2246 2247int 2248inmem(struct vnode * vp, daddr_t blkno) 2249{ 2250 vm_object_t obj; 2251 vm_offset_t toff, tinc, size; 2252 vm_page_t m; 2253 vm_ooffset_t off; 2254 2255 GIANT_REQUIRED; 2256 ASSERT_VOP_LOCKED(vp, "inmem"); 2257 2258 if (incore(vp, blkno)) 2259 return 1; 2260 if (vp->v_mount == NULL) 2261 return 0; 2262 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2263 return 0; 2264 2265 size = PAGE_SIZE; 2266 if (size > vp->v_mount->mnt_stat.f_iosize) 2267 size = vp->v_mount->mnt_stat.f_iosize; 2268 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2269 2270 VM_OBJECT_LOCK(obj); 2271 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2272 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2273 if (!m) 2274 goto notinmem; 2275 tinc = size; 2276 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2277 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2278 if (vm_page_is_valid(m, 2279 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2280 goto notinmem; 2281 } 2282 VM_OBJECT_UNLOCK(obj); 2283 return 1; 2284 2285notinmem: 2286 VM_OBJECT_UNLOCK(obj); 2287 return (0); 2288} 2289 2290/* 2291 * vfs_setdirty: 2292 * 2293 * Sets the dirty range for a buffer based on the status of the dirty 2294 * bits in the pages comprising the buffer. 2295 * 2296 * The range is limited to the size of the buffer. 2297 * 2298 * This routine is primarily used by NFS, but is generalized for the 2299 * B_VMIO case. 2300 */ 2301static void 2302vfs_setdirty(struct buf *bp) 2303{ 2304 int i; 2305 vm_object_t object; 2306 2307 GIANT_REQUIRED; 2308 /* 2309 * Degenerate case - empty buffer 2310 */ 2311 2312 if (bp->b_bufsize == 0) 2313 return; 2314 2315 /* 2316 * We qualify the scan for modified pages on whether the 2317 * object has been flushed yet. The OBJ_WRITEABLE flag 2318 * is not cleared simply by protecting pages off. 2319 */ 2320 2321 if ((bp->b_flags & B_VMIO) == 0) 2322 return; 2323 2324 object = bp->b_pages[0]->object; 2325 VM_OBJECT_LOCK(object); 2326 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2327 printf("Warning: object %p writeable but not mightbedirty\n", object); 2328 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2329 printf("Warning: object %p mightbedirty but not writeable\n", object); 2330 2331 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2332 vm_offset_t boffset; 2333 vm_offset_t eoffset; 2334 2335 vm_page_lock_queues(); 2336 /* 2337 * test the pages to see if they have been modified directly 2338 * by users through the VM system. 2339 */ 2340 for (i = 0; i < bp->b_npages; i++) 2341 vm_page_test_dirty(bp->b_pages[i]); 2342 2343 /* 2344 * Calculate the encompassing dirty range, boffset and eoffset, 2345 * (eoffset - boffset) bytes. 2346 */ 2347 2348 for (i = 0; i < bp->b_npages; i++) { 2349 if (bp->b_pages[i]->dirty) 2350 break; 2351 } 2352 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2353 2354 for (i = bp->b_npages - 1; i >= 0; --i) { 2355 if (bp->b_pages[i]->dirty) { 2356 break; 2357 } 2358 } 2359 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2360 2361 vm_page_unlock_queues(); 2362 /* 2363 * Fit it to the buffer. 2364 */ 2365 2366 if (eoffset > bp->b_bcount) 2367 eoffset = bp->b_bcount; 2368 2369 /* 2370 * If we have a good dirty range, merge with the existing 2371 * dirty range. 2372 */ 2373 2374 if (boffset < eoffset) { 2375 if (bp->b_dirtyoff > boffset) 2376 bp->b_dirtyoff = boffset; 2377 if (bp->b_dirtyend < eoffset) 2378 bp->b_dirtyend = eoffset; 2379 } 2380 } 2381 VM_OBJECT_UNLOCK(object); 2382} 2383 2384/* 2385 * getblk: 2386 * 2387 * Get a block given a specified block and offset into a file/device. 2388 * The buffers B_DONE bit will be cleared on return, making it almost 2389 * ready for an I/O initiation. B_INVAL may or may not be set on 2390 * return. The caller should clear B_INVAL prior to initiating a 2391 * READ. 2392 * 2393 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2394 * an existing buffer. 2395 * 2396 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2397 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2398 * and then cleared based on the backing VM. If the previous buffer is 2399 * non-0-sized but invalid, B_CACHE will be cleared. 2400 * 2401 * If getblk() must create a new buffer, the new buffer is returned with 2402 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2403 * case it is returned with B_INVAL clear and B_CACHE set based on the 2404 * backing VM. 2405 * 2406 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2407 * B_CACHE bit is clear. 2408 * 2409 * What this means, basically, is that the caller should use B_CACHE to 2410 * determine whether the buffer is fully valid or not and should clear 2411 * B_INVAL prior to issuing a read. If the caller intends to validate 2412 * the buffer by loading its data area with something, the caller needs 2413 * to clear B_INVAL. If the caller does this without issuing an I/O, 2414 * the caller should set B_CACHE ( as an optimization ), else the caller 2415 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2416 * a write attempt or if it was a successfull read. If the caller 2417 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2418 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2419 */ 2420struct buf * 2421getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2422 int flags) 2423{ 2424 struct buf *bp; 2425 int s; 2426 int error; 2427 ASSERT_VOP_LOCKED(vp, "getblk"); 2428 2429 if (size > MAXBSIZE) 2430 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2431 2432 s = splbio(); 2433loop: 2434 /* 2435 * Block if we are low on buffers. Certain processes are allowed 2436 * to completely exhaust the buffer cache. 2437 * 2438 * If this check ever becomes a bottleneck it may be better to 2439 * move it into the else, when gbincore() fails. At the moment 2440 * it isn't a problem. 2441 * 2442 * XXX remove if 0 sections (clean this up after its proven) 2443 */ 2444 if (numfreebuffers == 0) { 2445 if (curthread == PCPU_GET(idlethread)) 2446 return NULL; 2447 mtx_lock(&nblock); 2448 needsbuffer |= VFS_BIO_NEED_ANY; 2449 mtx_unlock(&nblock); 2450 } 2451 2452 VI_LOCK(vp); 2453 if ((bp = gbincore(vp, blkno))) { 2454 int lockflags; 2455 /* 2456 * Buffer is in-core. If the buffer is not busy, it must 2457 * be on a queue. 2458 */ 2459 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2460 2461 if (flags & GB_LOCK_NOWAIT) 2462 lockflags |= LK_NOWAIT; 2463 2464 error = BUF_TIMELOCK(bp, lockflags, 2465 VI_MTX(vp), "getblk", slpflag, slptimeo); 2466 2467 /* 2468 * If we slept and got the lock we have to restart in case 2469 * the buffer changed identities. 2470 */ 2471 if (error == ENOLCK) 2472 goto loop; 2473 /* We timed out or were interrupted. */ 2474 else if (error) 2475 return (NULL); 2476 2477 /* 2478 * The buffer is locked. B_CACHE is cleared if the buffer is 2479 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2480 * and for a VMIO buffer B_CACHE is adjusted according to the 2481 * backing VM cache. 2482 */ 2483 if (bp->b_flags & B_INVAL) 2484 bp->b_flags &= ~B_CACHE; 2485 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2486 bp->b_flags |= B_CACHE; 2487 bremfree(bp); 2488 2489 /* 2490 * check for size inconsistancies for non-VMIO case. 2491 */ 2492 2493 if (bp->b_bcount != size) { 2494 if ((bp->b_flags & B_VMIO) == 0 || 2495 (size > bp->b_kvasize)) { 2496 if (bp->b_flags & B_DELWRI) { 2497 bp->b_flags |= B_NOCACHE; 2498 bwrite(bp); 2499 } else { 2500 if ((bp->b_flags & B_VMIO) && 2501 (LIST_FIRST(&bp->b_dep) == NULL)) { 2502 bp->b_flags |= B_RELBUF; 2503 brelse(bp); 2504 } else { 2505 bp->b_flags |= B_NOCACHE; 2506 bwrite(bp); 2507 } 2508 } 2509 goto loop; 2510 } 2511 } 2512 2513 /* 2514 * If the size is inconsistant in the VMIO case, we can resize 2515 * the buffer. This might lead to B_CACHE getting set or 2516 * cleared. If the size has not changed, B_CACHE remains 2517 * unchanged from its previous state. 2518 */ 2519 2520 if (bp->b_bcount != size) 2521 allocbuf(bp, size); 2522 2523 KASSERT(bp->b_offset != NOOFFSET, 2524 ("getblk: no buffer offset")); 2525 2526 /* 2527 * A buffer with B_DELWRI set and B_CACHE clear must 2528 * be committed before we can return the buffer in 2529 * order to prevent the caller from issuing a read 2530 * ( due to B_CACHE not being set ) and overwriting 2531 * it. 2532 * 2533 * Most callers, including NFS and FFS, need this to 2534 * operate properly either because they assume they 2535 * can issue a read if B_CACHE is not set, or because 2536 * ( for example ) an uncached B_DELWRI might loop due 2537 * to softupdates re-dirtying the buffer. In the latter 2538 * case, B_CACHE is set after the first write completes, 2539 * preventing further loops. 2540 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2541 * above while extending the buffer, we cannot allow the 2542 * buffer to remain with B_CACHE set after the write 2543 * completes or it will represent a corrupt state. To 2544 * deal with this we set B_NOCACHE to scrap the buffer 2545 * after the write. 2546 * 2547 * We might be able to do something fancy, like setting 2548 * B_CACHE in bwrite() except if B_DELWRI is already set, 2549 * so the below call doesn't set B_CACHE, but that gets real 2550 * confusing. This is much easier. 2551 */ 2552 2553 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2554 bp->b_flags |= B_NOCACHE; 2555 bwrite(bp); 2556 goto loop; 2557 } 2558 2559 splx(s); 2560 bp->b_flags &= ~B_DONE; 2561 } else { 2562 int bsize, maxsize, vmio; 2563 off_t offset; 2564 2565 /* 2566 * Buffer is not in-core, create new buffer. The buffer 2567 * returned by getnewbuf() is locked. Note that the returned 2568 * buffer is also considered valid (not marked B_INVAL). 2569 */ 2570 VI_UNLOCK(vp); 2571 /* 2572 * If the user does not want us to create the buffer, bail out 2573 * here. 2574 */ 2575 if (flags & GB_NOCREAT) { 2576 splx(s); 2577 return NULL; 2578 } 2579 if (vn_isdisk(vp, NULL)) 2580 bsize = DEV_BSIZE; 2581 else if (vp->v_mountedhere) 2582 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2583 else if (vp->v_mount) 2584 bsize = vp->v_mount->mnt_stat.f_iosize; 2585 else 2586 bsize = size; 2587 2588 if (vp->v_bsize != bsize) { 2589#if 0 2590 printf("WARNING: Wrong block size on vnode: %d should be %d\n", vp->v_bsize, bsize); 2591#endif 2592 vp->v_bsize = bsize; 2593 } 2594 2595 offset = blkno * bsize; 2596 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2597 (vp->v_vflag & VV_OBJBUF); 2598 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2599 maxsize = imax(maxsize, bsize); 2600 2601 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2602 if (slpflag || slptimeo) { 2603 splx(s); 2604 return NULL; 2605 } 2606 goto loop; 2607 } 2608 2609 /* 2610 * This code is used to make sure that a buffer is not 2611 * created while the getnewbuf routine is blocked. 2612 * This can be a problem whether the vnode is locked or not. 2613 * If the buffer is created out from under us, we have to 2614 * throw away the one we just created. There is now window 2615 * race because we are safely running at splbio() from the 2616 * point of the duplicate buffer creation through to here, 2617 * and we've locked the buffer. 2618 * 2619 * Note: this must occur before we associate the buffer 2620 * with the vp especially considering limitations in 2621 * the splay tree implementation when dealing with duplicate 2622 * lblkno's. 2623 */ 2624 VI_LOCK(vp); 2625 if (gbincore(vp, blkno)) { 2626 VI_UNLOCK(vp); 2627 bp->b_flags |= B_INVAL; 2628 brelse(bp); 2629 goto loop; 2630 } 2631 2632 /* 2633 * Insert the buffer into the hash, so that it can 2634 * be found by incore. 2635 */ 2636 bp->b_blkno = bp->b_lblkno = blkno; 2637 bp->b_offset = offset; 2638 2639 bgetvp(vp, bp); 2640 VI_UNLOCK(vp); 2641 2642 /* 2643 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2644 * buffer size starts out as 0, B_CACHE will be set by 2645 * allocbuf() for the VMIO case prior to it testing the 2646 * backing store for validity. 2647 */ 2648 2649 if (vmio) { 2650 bp->b_flags |= B_VMIO; 2651#if defined(VFS_BIO_DEBUG) 2652 if (vn_canvmio(vp) != TRUE) 2653 printf("getblk: VMIO on vnode type %d\n", 2654 vp->v_type); 2655#endif 2656 VOP_GETVOBJECT(vp, &bp->b_object); 2657 } else { 2658 bp->b_flags &= ~B_VMIO; 2659 bp->b_object = NULL; 2660 } 2661 2662 allocbuf(bp, size); 2663 2664 splx(s); 2665 bp->b_flags &= ~B_DONE; 2666 } 2667 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2668 return (bp); 2669} 2670 2671/* 2672 * Get an empty, disassociated buffer of given size. The buffer is initially 2673 * set to B_INVAL. 2674 */ 2675struct buf * 2676geteblk(int size) 2677{ 2678 struct buf *bp; 2679 int s; 2680 int maxsize; 2681 2682 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2683 2684 s = splbio(); 2685 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2686 continue; 2687 splx(s); 2688 allocbuf(bp, size); 2689 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2690 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2691 return (bp); 2692} 2693 2694 2695/* 2696 * This code constitutes the buffer memory from either anonymous system 2697 * memory (in the case of non-VMIO operations) or from an associated 2698 * VM object (in the case of VMIO operations). This code is able to 2699 * resize a buffer up or down. 2700 * 2701 * Note that this code is tricky, and has many complications to resolve 2702 * deadlock or inconsistant data situations. Tread lightly!!! 2703 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2704 * the caller. Calling this code willy nilly can result in the loss of data. 2705 * 2706 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2707 * B_CACHE for the non-VMIO case. 2708 */ 2709 2710int 2711allocbuf(struct buf *bp, int size) 2712{ 2713 int newbsize, mbsize; 2714 int i; 2715 2716 GIANT_REQUIRED; 2717 2718 if (BUF_REFCNT(bp) == 0) 2719 panic("allocbuf: buffer not busy"); 2720 2721 if (bp->b_kvasize < size) 2722 panic("allocbuf: buffer too small"); 2723 2724 if ((bp->b_flags & B_VMIO) == 0) { 2725 caddr_t origbuf; 2726 int origbufsize; 2727 /* 2728 * Just get anonymous memory from the kernel. Don't 2729 * mess with B_CACHE. 2730 */ 2731 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2732 if (bp->b_flags & B_MALLOC) 2733 newbsize = mbsize; 2734 else 2735 newbsize = round_page(size); 2736 2737 if (newbsize < bp->b_bufsize) { 2738 /* 2739 * malloced buffers are not shrunk 2740 */ 2741 if (bp->b_flags & B_MALLOC) { 2742 if (newbsize) { 2743 bp->b_bcount = size; 2744 } else { 2745 free(bp->b_data, M_BIOBUF); 2746 if (bp->b_bufsize) { 2747 atomic_subtract_int( 2748 &bufmallocspace, 2749 bp->b_bufsize); 2750 bufspacewakeup(); 2751 bp->b_bufsize = 0; 2752 } 2753 bp->b_saveaddr = bp->b_kvabase; 2754 bp->b_data = bp->b_saveaddr; 2755 bp->b_bcount = 0; 2756 bp->b_flags &= ~B_MALLOC; 2757 } 2758 return 1; 2759 } 2760 vm_hold_free_pages( 2761 bp, 2762 (vm_offset_t) bp->b_data + newbsize, 2763 (vm_offset_t) bp->b_data + bp->b_bufsize); 2764 } else if (newbsize > bp->b_bufsize) { 2765 /* 2766 * We only use malloced memory on the first allocation. 2767 * and revert to page-allocated memory when the buffer 2768 * grows. 2769 */ 2770 /* 2771 * There is a potential smp race here that could lead 2772 * to bufmallocspace slightly passing the max. It 2773 * is probably extremely rare and not worth worrying 2774 * over. 2775 */ 2776 if ( (bufmallocspace < maxbufmallocspace) && 2777 (bp->b_bufsize == 0) && 2778 (mbsize <= PAGE_SIZE/2)) { 2779 2780 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2781 bp->b_bufsize = mbsize; 2782 bp->b_bcount = size; 2783 bp->b_flags |= B_MALLOC; 2784 atomic_add_int(&bufmallocspace, mbsize); 2785 return 1; 2786 } 2787 origbuf = NULL; 2788 origbufsize = 0; 2789 /* 2790 * If the buffer is growing on its other-than-first allocation, 2791 * then we revert to the page-allocation scheme. 2792 */ 2793 if (bp->b_flags & B_MALLOC) { 2794 origbuf = bp->b_data; 2795 origbufsize = bp->b_bufsize; 2796 bp->b_data = bp->b_kvabase; 2797 if (bp->b_bufsize) { 2798 atomic_subtract_int(&bufmallocspace, 2799 bp->b_bufsize); 2800 bufspacewakeup(); 2801 bp->b_bufsize = 0; 2802 } 2803 bp->b_flags &= ~B_MALLOC; 2804 newbsize = round_page(newbsize); 2805 } 2806 vm_hold_load_pages( 2807 bp, 2808 (vm_offset_t) bp->b_data + bp->b_bufsize, 2809 (vm_offset_t) bp->b_data + newbsize); 2810 if (origbuf) { 2811 bcopy(origbuf, bp->b_data, origbufsize); 2812 free(origbuf, M_BIOBUF); 2813 } 2814 } 2815 } else { 2816 int desiredpages; 2817 2818 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2819 desiredpages = (size == 0) ? 0 : 2820 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2821 2822 if (bp->b_flags & B_MALLOC) 2823 panic("allocbuf: VMIO buffer can't be malloced"); 2824 /* 2825 * Set B_CACHE initially if buffer is 0 length or will become 2826 * 0-length. 2827 */ 2828 if (size == 0 || bp->b_bufsize == 0) 2829 bp->b_flags |= B_CACHE; 2830 2831 if (newbsize < bp->b_bufsize) { 2832 /* 2833 * DEV_BSIZE aligned new buffer size is less then the 2834 * DEV_BSIZE aligned existing buffer size. Figure out 2835 * if we have to remove any pages. 2836 */ 2837 if (desiredpages < bp->b_npages) { 2838 vm_page_t m; 2839 2840 vm_page_lock_queues(); 2841 for (i = desiredpages; i < bp->b_npages; i++) { 2842 /* 2843 * the page is not freed here -- it 2844 * is the responsibility of 2845 * vnode_pager_setsize 2846 */ 2847 m = bp->b_pages[i]; 2848 KASSERT(m != bogus_page, 2849 ("allocbuf: bogus page found")); 2850 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2851 vm_page_lock_queues(); 2852 2853 bp->b_pages[i] = NULL; 2854 vm_page_unwire(m, 0); 2855 } 2856 vm_page_unlock_queues(); 2857 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2858 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2859 bp->b_npages = desiredpages; 2860 } 2861 } else if (size > bp->b_bcount) { 2862 /* 2863 * We are growing the buffer, possibly in a 2864 * byte-granular fashion. 2865 */ 2866 struct vnode *vp; 2867 vm_object_t obj; 2868 vm_offset_t toff; 2869 vm_offset_t tinc; 2870 2871 /* 2872 * Step 1, bring in the VM pages from the object, 2873 * allocating them if necessary. We must clear 2874 * B_CACHE if these pages are not valid for the 2875 * range covered by the buffer. 2876 */ 2877 2878 vp = bp->b_vp; 2879 obj = bp->b_object; 2880 2881 VM_OBJECT_LOCK(obj); 2882 while (bp->b_npages < desiredpages) { 2883 vm_page_t m; 2884 vm_pindex_t pi; 2885 2886 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2887 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2888 /* 2889 * note: must allocate system pages 2890 * since blocking here could intefere 2891 * with paging I/O, no matter which 2892 * process we are. 2893 */ 2894 m = vm_page_alloc(obj, pi, 2895 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2896 if (m == NULL) { 2897 atomic_add_int(&vm_pageout_deficit, 2898 desiredpages - bp->b_npages); 2899 VM_OBJECT_UNLOCK(obj); 2900 VM_WAIT; 2901 VM_OBJECT_LOCK(obj); 2902 } else { 2903 vm_page_lock_queues(); 2904 vm_page_wakeup(m); 2905 vm_page_unlock_queues(); 2906 bp->b_flags &= ~B_CACHE; 2907 bp->b_pages[bp->b_npages] = m; 2908 ++bp->b_npages; 2909 } 2910 continue; 2911 } 2912 2913 /* 2914 * We found a page. If we have to sleep on it, 2915 * retry because it might have gotten freed out 2916 * from under us. 2917 * 2918 * We can only test PG_BUSY here. Blocking on 2919 * m->busy might lead to a deadlock: 2920 * 2921 * vm_fault->getpages->cluster_read->allocbuf 2922 * 2923 */ 2924 vm_page_lock_queues(); 2925 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2926 continue; 2927 2928 /* 2929 * We have a good page. Should we wakeup the 2930 * page daemon? 2931 */ 2932 if ((curproc != pageproc) && 2933 ((m->queue - m->pc) == PQ_CACHE) && 2934 ((cnt.v_free_count + cnt.v_cache_count) < 2935 (cnt.v_free_min + cnt.v_cache_min))) { 2936 pagedaemon_wakeup(); 2937 } 2938 vm_page_wire(m); 2939 vm_page_unlock_queues(); 2940 bp->b_pages[bp->b_npages] = m; 2941 ++bp->b_npages; 2942 } 2943 2944 /* 2945 * Step 2. We've loaded the pages into the buffer, 2946 * we have to figure out if we can still have B_CACHE 2947 * set. Note that B_CACHE is set according to the 2948 * byte-granular range ( bcount and size ), new the 2949 * aligned range ( newbsize ). 2950 * 2951 * The VM test is against m->valid, which is DEV_BSIZE 2952 * aligned. Needless to say, the validity of the data 2953 * needs to also be DEV_BSIZE aligned. Note that this 2954 * fails with NFS if the server or some other client 2955 * extends the file's EOF. If our buffer is resized, 2956 * B_CACHE may remain set! XXX 2957 */ 2958 2959 toff = bp->b_bcount; 2960 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2961 2962 while ((bp->b_flags & B_CACHE) && toff < size) { 2963 vm_pindex_t pi; 2964 2965 if (tinc > (size - toff)) 2966 tinc = size - toff; 2967 2968 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2969 PAGE_SHIFT; 2970 2971 vfs_buf_test_cache( 2972 bp, 2973 bp->b_offset, 2974 toff, 2975 tinc, 2976 bp->b_pages[pi] 2977 ); 2978 toff += tinc; 2979 tinc = PAGE_SIZE; 2980 } 2981 VM_OBJECT_UNLOCK(obj); 2982 2983 /* 2984 * Step 3, fixup the KVM pmap. Remember that 2985 * bp->b_data is relative to bp->b_offset, but 2986 * bp->b_offset may be offset into the first page. 2987 */ 2988 2989 bp->b_data = (caddr_t) 2990 trunc_page((vm_offset_t)bp->b_data); 2991 pmap_qenter( 2992 (vm_offset_t)bp->b_data, 2993 bp->b_pages, 2994 bp->b_npages 2995 ); 2996 2997 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2998 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2999 } 3000 } 3001 if (newbsize < bp->b_bufsize) 3002 bufspacewakeup(); 3003 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3004 bp->b_bcount = size; /* requested buffer size */ 3005 return 1; 3006} 3007 3008void 3009biodone(struct bio *bp) 3010{ 3011 3012 mtx_lock(&bdonelock); 3013 bp->bio_flags |= BIO_DONE; 3014 if (bp->bio_done == NULL) 3015 wakeup(bp); 3016 mtx_unlock(&bdonelock); 3017 if (bp->bio_done != NULL) 3018 bp->bio_done(bp); 3019} 3020 3021/* 3022 * Wait for a BIO to finish. 3023 * 3024 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3025 * case is not yet clear. 3026 */ 3027int 3028biowait(struct bio *bp, const char *wchan) 3029{ 3030 3031 mtx_lock(&bdonelock); 3032 while ((bp->bio_flags & BIO_DONE) == 0) 3033 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3034 mtx_unlock(&bdonelock); 3035 if (bp->bio_error != 0) 3036 return (bp->bio_error); 3037 if (!(bp->bio_flags & BIO_ERROR)) 3038 return (0); 3039 return (EIO); 3040} 3041 3042void 3043biofinish(struct bio *bp, struct devstat *stat, int error) 3044{ 3045 3046 if (error) { 3047 bp->bio_error = error; 3048 bp->bio_flags |= BIO_ERROR; 3049 } 3050 if (stat != NULL) 3051 devstat_end_transaction_bio(stat, bp); 3052 biodone(bp); 3053} 3054 3055/* 3056 * bufwait: 3057 * 3058 * Wait for buffer I/O completion, returning error status. The buffer 3059 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3060 * error and cleared. 3061 */ 3062int 3063bufwait(struct buf *bp) 3064{ 3065 int s; 3066 3067 s = splbio(); 3068 if (bp->b_iocmd == BIO_READ) 3069 bwait(bp, PRIBIO, "biord"); 3070 else 3071 bwait(bp, PRIBIO, "biowr"); 3072 splx(s); 3073 if (bp->b_flags & B_EINTR) { 3074 bp->b_flags &= ~B_EINTR; 3075 return (EINTR); 3076 } 3077 if (bp->b_ioflags & BIO_ERROR) { 3078 return (bp->b_error ? bp->b_error : EIO); 3079 } else { 3080 return (0); 3081 } 3082} 3083 3084 /* 3085 * Call back function from struct bio back up to struct buf. 3086 */ 3087static void 3088bufdonebio(struct bio *bp) 3089{ 3090 3091 /* Device drivers may or may not hold giant, hold it here. */ 3092 mtx_lock(&Giant); 3093 bufdone(bp->bio_caller2); 3094 mtx_unlock(&Giant); 3095} 3096 3097void 3098dev_strategy(struct buf *bp) 3099{ 3100 struct cdevsw *csw; 3101 struct cdev *dev; 3102 3103 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3104 panic("b_iocmd botch"); 3105 bp->b_io.bio_done = bufdonebio; 3106 bp->b_io.bio_caller2 = bp; 3107 dev = bp->b_io.bio_dev; 3108 KASSERT(dev->si_refcount > 0, 3109 ("dev_strategy on un-referenced struct cdev *(%s)", 3110 devtoname(dev))); 3111 dev_lock(); 3112 csw = devsw(dev); 3113 if (csw != NULL) 3114 dev->si_threadcount++; 3115 dev_unlock(); 3116 if (csw == NULL) { 3117 bp->b_error = ENXIO; 3118 bp->b_ioflags = BIO_ERROR; 3119 mtx_lock(&Giant); /* XXX: too defensive ? */ 3120 bufdone(bp); 3121 mtx_unlock(&Giant); /* XXX: too defensive ? */ 3122 return; 3123 } 3124 (*csw->d_strategy)(&bp->b_io); 3125 dev_lock(); 3126 dev->si_threadcount--; 3127 dev_unlock(); 3128} 3129 3130/* 3131 * bufdone: 3132 * 3133 * Finish I/O on a buffer, optionally calling a completion function. 3134 * This is usually called from an interrupt so process blocking is 3135 * not allowed. 3136 * 3137 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3138 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3139 * assuming B_INVAL is clear. 3140 * 3141 * For the VMIO case, we set B_CACHE if the op was a read and no 3142 * read error occured, or if the op was a write. B_CACHE is never 3143 * set if the buffer is invalid or otherwise uncacheable. 3144 * 3145 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3146 * initiator to leave B_INVAL set to brelse the buffer out of existance 3147 * in the biodone routine. 3148 */ 3149void 3150bufdone(struct buf *bp) 3151{ 3152 int s; 3153 void (*biodone)(struct buf *); 3154 3155 3156 s = splbio(); 3157 3158 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3159 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3160 3161 bp->b_flags |= B_DONE; 3162 runningbufwakeup(bp); 3163 3164 if (bp->b_iocmd == BIO_WRITE) { 3165 vwakeup(bp); 3166 } 3167 3168 /* call optional completion function if requested */ 3169 if (bp->b_iodone != NULL) { 3170 biodone = bp->b_iodone; 3171 bp->b_iodone = NULL; 3172 (*biodone) (bp); 3173 splx(s); 3174 return; 3175 } 3176 if (LIST_FIRST(&bp->b_dep) != NULL) 3177 buf_complete(bp); 3178 3179 if (bp->b_flags & B_VMIO) { 3180 int i; 3181 vm_ooffset_t foff; 3182 vm_page_t m; 3183 vm_object_t obj; 3184 int iosize; 3185 struct vnode *vp = bp->b_vp; 3186 3187 obj = bp->b_object; 3188 3189#if defined(VFS_BIO_DEBUG) 3190 mp_fixme("usecount and vflag accessed without locks."); 3191 if (vp->v_usecount == 0) { 3192 panic("biodone: zero vnode ref count"); 3193 } 3194 3195 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3196 panic("biodone: vnode is not setup for merged cache"); 3197 } 3198#endif 3199 3200 foff = bp->b_offset; 3201 KASSERT(bp->b_offset != NOOFFSET, 3202 ("biodone: no buffer offset")); 3203 3204 VM_OBJECT_LOCK(obj); 3205#if defined(VFS_BIO_DEBUG) 3206 if (obj->paging_in_progress < bp->b_npages) { 3207 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3208 obj->paging_in_progress, bp->b_npages); 3209 } 3210#endif 3211 3212 /* 3213 * Set B_CACHE if the op was a normal read and no error 3214 * occured. B_CACHE is set for writes in the b*write() 3215 * routines. 3216 */ 3217 iosize = bp->b_bcount - bp->b_resid; 3218 if (bp->b_iocmd == BIO_READ && 3219 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3220 !(bp->b_ioflags & BIO_ERROR)) { 3221 bp->b_flags |= B_CACHE; 3222 } 3223 vm_page_lock_queues(); 3224 for (i = 0; i < bp->b_npages; i++) { 3225 int bogusflag = 0; 3226 int resid; 3227 3228 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3229 if (resid > iosize) 3230 resid = iosize; 3231 3232 /* 3233 * cleanup bogus pages, restoring the originals 3234 */ 3235 m = bp->b_pages[i]; 3236 if (m == bogus_page) { 3237 bogusflag = 1; 3238 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3239 if (m == NULL) 3240 panic("biodone: page disappeared!"); 3241 bp->b_pages[i] = m; 3242 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3243 } 3244#if defined(VFS_BIO_DEBUG) 3245 if (OFF_TO_IDX(foff) != m->pindex) { 3246 printf( 3247"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3248 (intmax_t)foff, (uintmax_t)m->pindex); 3249 } 3250#endif 3251 3252 /* 3253 * In the write case, the valid and clean bits are 3254 * already changed correctly ( see bdwrite() ), so we 3255 * only need to do this here in the read case. 3256 */ 3257 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3258 vfs_page_set_valid(bp, foff, i, m); 3259 } 3260 3261 /* 3262 * when debugging new filesystems or buffer I/O methods, this 3263 * is the most common error that pops up. if you see this, you 3264 * have not set the page busy flag correctly!!! 3265 */ 3266 if (m->busy == 0) { 3267 printf("biodone: page busy < 0, " 3268 "pindex: %d, foff: 0x(%x,%x), " 3269 "resid: %d, index: %d\n", 3270 (int) m->pindex, (int)(foff >> 32), 3271 (int) foff & 0xffffffff, resid, i); 3272 if (!vn_isdisk(vp, NULL)) 3273 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3274 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3275 (intmax_t) bp->b_lblkno, 3276 bp->b_flags, bp->b_npages); 3277 else 3278 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3279 (intmax_t) bp->b_lblkno, 3280 bp->b_flags, bp->b_npages); 3281 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3282 (u_long)m->valid, (u_long)m->dirty, 3283 m->wire_count); 3284 panic("biodone: page busy < 0\n"); 3285 } 3286 vm_page_io_finish(m); 3287 vm_object_pip_subtract(obj, 1); 3288 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3289 iosize -= resid; 3290 } 3291 vm_page_unlock_queues(); 3292 vm_object_pip_wakeupn(obj, 0); 3293 VM_OBJECT_UNLOCK(obj); 3294 } 3295 3296 /* 3297 * For asynchronous completions, release the buffer now. The brelse 3298 * will do a wakeup there if necessary - so no need to do a wakeup 3299 * here in the async case. The sync case always needs to do a wakeup. 3300 */ 3301 3302 if (bp->b_flags & B_ASYNC) { 3303 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3304 brelse(bp); 3305 else 3306 bqrelse(bp); 3307 } else { 3308 bdone(bp); 3309 } 3310 splx(s); 3311} 3312 3313/* 3314 * This routine is called in lieu of iodone in the case of 3315 * incomplete I/O. This keeps the busy status for pages 3316 * consistant. 3317 */ 3318void 3319vfs_unbusy_pages(struct buf *bp) 3320{ 3321 int i; 3322 vm_object_t obj; 3323 vm_page_t m; 3324 3325 runningbufwakeup(bp); 3326 if (!(bp->b_flags & B_VMIO)) 3327 return; 3328 3329 obj = bp->b_object; 3330 VM_OBJECT_LOCK(obj); 3331 vm_page_lock_queues(); 3332 for (i = 0; i < bp->b_npages; i++) { 3333 m = bp->b_pages[i]; 3334 if (m == bogus_page) { 3335 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3336 if (!m) { 3337 panic("vfs_unbusy_pages: page missing\n"); 3338 } 3339 bp->b_pages[i] = m; 3340 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3341 } 3342 vm_object_pip_subtract(obj, 1); 3343 vm_page_io_finish(m); 3344 } 3345 vm_page_unlock_queues(); 3346 vm_object_pip_wakeupn(obj, 0); 3347 VM_OBJECT_UNLOCK(obj); 3348} 3349 3350/* 3351 * vfs_page_set_valid: 3352 * 3353 * Set the valid bits in a page based on the supplied offset. The 3354 * range is restricted to the buffer's size. 3355 * 3356 * This routine is typically called after a read completes. 3357 */ 3358static void 3359vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3360{ 3361 vm_ooffset_t soff, eoff; 3362 3363 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3364 /* 3365 * Start and end offsets in buffer. eoff - soff may not cross a 3366 * page boundry or cross the end of the buffer. The end of the 3367 * buffer, in this case, is our file EOF, not the allocation size 3368 * of the buffer. 3369 */ 3370 soff = off; 3371 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3372 if (eoff > bp->b_offset + bp->b_bcount) 3373 eoff = bp->b_offset + bp->b_bcount; 3374 3375 /* 3376 * Set valid range. This is typically the entire buffer and thus the 3377 * entire page. 3378 */ 3379 if (eoff > soff) { 3380 vm_page_set_validclean( 3381 m, 3382 (vm_offset_t) (soff & PAGE_MASK), 3383 (vm_offset_t) (eoff - soff) 3384 ); 3385 } 3386} 3387 3388/* 3389 * This routine is called before a device strategy routine. 3390 * It is used to tell the VM system that paging I/O is in 3391 * progress, and treat the pages associated with the buffer 3392 * almost as being PG_BUSY. Also the object paging_in_progress 3393 * flag is handled to make sure that the object doesn't become 3394 * inconsistant. 3395 * 3396 * Since I/O has not been initiated yet, certain buffer flags 3397 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3398 * and should be ignored. 3399 */ 3400void 3401vfs_busy_pages(struct buf *bp, int clear_modify) 3402{ 3403 int i, bogus; 3404 vm_object_t obj; 3405 vm_ooffset_t foff; 3406 vm_page_t m; 3407 3408 if (!(bp->b_flags & B_VMIO)) 3409 return; 3410 3411 obj = bp->b_object; 3412 foff = bp->b_offset; 3413 KASSERT(bp->b_offset != NOOFFSET, 3414 ("vfs_busy_pages: no buffer offset")); 3415 vfs_setdirty(bp); 3416 VM_OBJECT_LOCK(obj); 3417retry: 3418 vm_page_lock_queues(); 3419 for (i = 0; i < bp->b_npages; i++) { 3420 m = bp->b_pages[i]; 3421 3422 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3423 goto retry; 3424 } 3425 bogus = 0; 3426 for (i = 0; i < bp->b_npages; i++) { 3427 m = bp->b_pages[i]; 3428 3429 if ((bp->b_flags & B_CLUSTER) == 0) { 3430 vm_object_pip_add(obj, 1); 3431 vm_page_io_start(m); 3432 } 3433 /* 3434 * When readying a buffer for a read ( i.e 3435 * clear_modify == 0 ), it is important to do 3436 * bogus_page replacement for valid pages in 3437 * partially instantiated buffers. Partially 3438 * instantiated buffers can, in turn, occur when 3439 * reconstituting a buffer from its VM backing store 3440 * base. We only have to do this if B_CACHE is 3441 * clear ( which causes the I/O to occur in the 3442 * first place ). The replacement prevents the read 3443 * I/O from overwriting potentially dirty VM-backed 3444 * pages. XXX bogus page replacement is, uh, bogus. 3445 * It may not work properly with small-block devices. 3446 * We need to find a better way. 3447 */ 3448 pmap_remove_all(m); 3449 if (clear_modify) 3450 vfs_page_set_valid(bp, foff, i, m); 3451 else if (m->valid == VM_PAGE_BITS_ALL && 3452 (bp->b_flags & B_CACHE) == 0) { 3453 bp->b_pages[i] = bogus_page; 3454 bogus++; 3455 } 3456 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3457 } 3458 vm_page_unlock_queues(); 3459 VM_OBJECT_UNLOCK(obj); 3460 if (bogus) 3461 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3462 bp->b_pages, bp->b_npages); 3463} 3464 3465/* 3466 * Tell the VM system that the pages associated with this buffer 3467 * are clean. This is used for delayed writes where the data is 3468 * going to go to disk eventually without additional VM intevention. 3469 * 3470 * Note that while we only really need to clean through to b_bcount, we 3471 * just go ahead and clean through to b_bufsize. 3472 */ 3473static void 3474vfs_clean_pages(struct buf *bp) 3475{ 3476 int i; 3477 vm_ooffset_t foff, noff, eoff; 3478 vm_page_t m; 3479 3480 if (!(bp->b_flags & B_VMIO)) 3481 return; 3482 3483 foff = bp->b_offset; 3484 KASSERT(bp->b_offset != NOOFFSET, 3485 ("vfs_clean_pages: no buffer offset")); 3486 VM_OBJECT_LOCK(bp->b_object); 3487 vm_page_lock_queues(); 3488 for (i = 0; i < bp->b_npages; i++) { 3489 m = bp->b_pages[i]; 3490 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3491 eoff = noff; 3492 3493 if (eoff > bp->b_offset + bp->b_bufsize) 3494 eoff = bp->b_offset + bp->b_bufsize; 3495 vfs_page_set_valid(bp, foff, i, m); 3496 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3497 foff = noff; 3498 } 3499 vm_page_unlock_queues(); 3500 VM_OBJECT_UNLOCK(bp->b_object); 3501} 3502 3503/* 3504 * vfs_bio_set_validclean: 3505 * 3506 * Set the range within the buffer to valid and clean. The range is 3507 * relative to the beginning of the buffer, b_offset. Note that b_offset 3508 * itself may be offset from the beginning of the first page. 3509 * 3510 */ 3511 3512void 3513vfs_bio_set_validclean(struct buf *bp, int base, int size) 3514{ 3515 int i, n; 3516 vm_page_t m; 3517 3518 if (!(bp->b_flags & B_VMIO)) 3519 return; 3520 3521 /* 3522 * Fixup base to be relative to beginning of first page. 3523 * Set initial n to be the maximum number of bytes in the 3524 * first page that can be validated. 3525 */ 3526 3527 base += (bp->b_offset & PAGE_MASK); 3528 n = PAGE_SIZE - (base & PAGE_MASK); 3529 3530 VM_OBJECT_LOCK(bp->b_object); 3531 vm_page_lock_queues(); 3532 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3533 m = bp->b_pages[i]; 3534 3535 if (n > size) 3536 n = size; 3537 3538 vm_page_set_validclean(m, base & PAGE_MASK, n); 3539 base += n; 3540 size -= n; 3541 n = PAGE_SIZE; 3542 } 3543 vm_page_unlock_queues(); 3544 VM_OBJECT_UNLOCK(bp->b_object); 3545} 3546 3547/* 3548 * vfs_bio_clrbuf: 3549 * 3550 * clear a buffer. This routine essentially fakes an I/O, so we need 3551 * to clear BIO_ERROR and B_INVAL. 3552 * 3553 * Note that while we only theoretically need to clear through b_bcount, 3554 * we go ahead and clear through b_bufsize. 3555 */ 3556 3557void 3558vfs_bio_clrbuf(struct buf *bp) 3559{ 3560 int i, j, mask = 0; 3561 caddr_t sa, ea; 3562 3563 GIANT_REQUIRED; 3564 3565 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3566 clrbuf(bp); 3567 return; 3568 } 3569 bp->b_flags &= ~B_INVAL; 3570 bp->b_ioflags &= ~BIO_ERROR; 3571 VM_OBJECT_LOCK(bp->b_object); 3572 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3573 (bp->b_offset & PAGE_MASK) == 0) { 3574 if (bp->b_pages[0] == bogus_page) 3575 goto unlock; 3576 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3577 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3578 if ((bp->b_pages[0]->valid & mask) == mask) 3579 goto unlock; 3580 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3581 ((bp->b_pages[0]->valid & mask) == 0)) { 3582 bzero(bp->b_data, bp->b_bufsize); 3583 bp->b_pages[0]->valid |= mask; 3584 goto unlock; 3585 } 3586 } 3587 ea = sa = bp->b_data; 3588 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3589 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3590 ea = (caddr_t)(vm_offset_t)ulmin( 3591 (u_long)(vm_offset_t)ea, 3592 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3593 if (bp->b_pages[i] == bogus_page) 3594 continue; 3595 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3596 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3597 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3598 if ((bp->b_pages[i]->valid & mask) == mask) 3599 continue; 3600 if ((bp->b_pages[i]->valid & mask) == 0) { 3601 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3602 bzero(sa, ea - sa); 3603 } else { 3604 for (; sa < ea; sa += DEV_BSIZE, j++) { 3605 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3606 (bp->b_pages[i]->valid & (1<<j)) == 0) 3607 bzero(sa, DEV_BSIZE); 3608 } 3609 } 3610 bp->b_pages[i]->valid |= mask; 3611 } 3612unlock: 3613 VM_OBJECT_UNLOCK(bp->b_object); 3614 bp->b_resid = 0; 3615} 3616 3617/* 3618 * vm_hold_load_pages and vm_hold_free_pages get pages into 3619 * a buffers address space. The pages are anonymous and are 3620 * not associated with a file object. 3621 */ 3622static void 3623vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3624{ 3625 vm_offset_t pg; 3626 vm_page_t p; 3627 int index; 3628 3629 to = round_page(to); 3630 from = round_page(from); 3631 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3632 3633 VM_OBJECT_LOCK(kernel_object); 3634 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3635tryagain: 3636 /* 3637 * note: must allocate system pages since blocking here 3638 * could intefere with paging I/O, no matter which 3639 * process we are. 3640 */ 3641 p = vm_page_alloc(kernel_object, 3642 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3643 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3644 if (!p) { 3645 atomic_add_int(&vm_pageout_deficit, 3646 (to - pg) >> PAGE_SHIFT); 3647 VM_OBJECT_UNLOCK(kernel_object); 3648 VM_WAIT; 3649 VM_OBJECT_LOCK(kernel_object); 3650 goto tryagain; 3651 } 3652 p->valid = VM_PAGE_BITS_ALL; 3653 pmap_qenter(pg, &p, 1); 3654 bp->b_pages[index] = p; 3655 vm_page_lock_queues(); 3656 vm_page_wakeup(p); 3657 vm_page_unlock_queues(); 3658 } 3659 VM_OBJECT_UNLOCK(kernel_object); 3660 bp->b_npages = index; 3661} 3662 3663/* Return pages associated with this buf to the vm system */ 3664static void 3665vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3666{ 3667 vm_offset_t pg; 3668 vm_page_t p; 3669 int index, newnpages; 3670 3671 GIANT_REQUIRED; 3672 3673 from = round_page(from); 3674 to = round_page(to); 3675 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3676 3677 VM_OBJECT_LOCK(kernel_object); 3678 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3679 p = bp->b_pages[index]; 3680 if (p && (index < bp->b_npages)) { 3681 if (p->busy) { 3682 printf( 3683 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3684 (intmax_t)bp->b_blkno, 3685 (intmax_t)bp->b_lblkno); 3686 } 3687 bp->b_pages[index] = NULL; 3688 pmap_qremove(pg, 1); 3689 vm_page_lock_queues(); 3690 vm_page_busy(p); 3691 vm_page_unwire(p, 0); 3692 vm_page_free(p); 3693 vm_page_unlock_queues(); 3694 } 3695 } 3696 VM_OBJECT_UNLOCK(kernel_object); 3697 bp->b_npages = newnpages; 3698} 3699 3700/* 3701 * Map an IO request into kernel virtual address space. 3702 * 3703 * All requests are (re)mapped into kernel VA space. 3704 * Notice that we use b_bufsize for the size of the buffer 3705 * to be mapped. b_bcount might be modified by the driver. 3706 * 3707 * Note that even if the caller determines that the address space should 3708 * be valid, a race or a smaller-file mapped into a larger space may 3709 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3710 * check the return value. 3711 */ 3712int 3713vmapbuf(struct buf *bp) 3714{ 3715 caddr_t addr, kva; 3716 vm_prot_t prot; 3717 int pidx, i; 3718 struct vm_page *m; 3719 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3720 3721 if (bp->b_bufsize < 0) 3722 return (-1); 3723 prot = VM_PROT_READ; 3724 if (bp->b_iocmd == BIO_READ) 3725 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3726 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3727 addr < bp->b_data + bp->b_bufsize; 3728 addr += PAGE_SIZE, pidx++) { 3729 /* 3730 * Do the vm_fault if needed; do the copy-on-write thing 3731 * when reading stuff off device into memory. 3732 * 3733 * NOTE! Must use pmap_extract() because addr may be in 3734 * the userland address space, and kextract is only guarenteed 3735 * to work for the kernland address space (see: sparc64 port). 3736 */ 3737retry: 3738 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3739 prot) < 0) { 3740 vm_page_lock_queues(); 3741 for (i = 0; i < pidx; ++i) { 3742 vm_page_unhold(bp->b_pages[i]); 3743 bp->b_pages[i] = NULL; 3744 } 3745 vm_page_unlock_queues(); 3746 return(-1); 3747 } 3748 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3749 if (m == NULL) 3750 goto retry; 3751 bp->b_pages[pidx] = m; 3752 } 3753 if (pidx > btoc(MAXPHYS)) 3754 panic("vmapbuf: mapped more than MAXPHYS"); 3755 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3756 3757 kva = bp->b_saveaddr; 3758 bp->b_npages = pidx; 3759 bp->b_saveaddr = bp->b_data; 3760 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3761 return(0); 3762} 3763 3764/* 3765 * Free the io map PTEs associated with this IO operation. 3766 * We also invalidate the TLB entries and restore the original b_addr. 3767 */ 3768void 3769vunmapbuf(struct buf *bp) 3770{ 3771 int pidx; 3772 int npages; 3773 3774 npages = bp->b_npages; 3775 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3776 vm_page_lock_queues(); 3777 for (pidx = 0; pidx < npages; pidx++) 3778 vm_page_unhold(bp->b_pages[pidx]); 3779 vm_page_unlock_queues(); 3780 3781 bp->b_data = bp->b_saveaddr; 3782} 3783 3784void 3785bdone(struct buf *bp) 3786{ 3787 3788 mtx_lock(&bdonelock); 3789 bp->b_flags |= B_DONE; 3790 wakeup(bp); 3791 mtx_unlock(&bdonelock); 3792} 3793 3794void 3795bwait(struct buf *bp, u_char pri, const char *wchan) 3796{ 3797 3798 mtx_lock(&bdonelock); 3799 while ((bp->b_flags & B_DONE) == 0) 3800 msleep(bp, &bdonelock, pri, wchan, 0); 3801 mtx_unlock(&bdonelock); 3802} 3803 3804#include "opt_ddb.h" 3805#ifdef DDB 3806#include <ddb/ddb.h> 3807 3808/* DDB command to show buffer data */ 3809DB_SHOW_COMMAND(buffer, db_show_buffer) 3810{ 3811 /* get args */ 3812 struct buf *bp = (struct buf *)addr; 3813 3814 if (!have_addr) { 3815 db_printf("usage: show buffer <addr>\n"); 3816 return; 3817 } 3818 3819 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3820 db_printf( 3821 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3822 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd\n", 3823 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3824 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3825 (intmax_t)bp->b_blkno); 3826 if (bp->b_npages) { 3827 int i; 3828 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3829 for (i = 0; i < bp->b_npages; i++) { 3830 vm_page_t m; 3831 m = bp->b_pages[i]; 3832 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3833 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3834 if ((i + 1) < bp->b_npages) 3835 db_printf(","); 3836 } 3837 db_printf("\n"); 3838 } 3839} 3840#endif /* DDB */ 3841