vfs_bio.c revision 136966
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 136966 2004-10-26 07:39:12Z 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 64struct buf_ops buf_ops_bio = { 65 .bop_name = "buf_ops_bio", 66 .bop_write = bufwrite, 67 .bop_strategy = bufstrategy, 68}; 69 70/* 71 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 72 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 73 */ 74struct buf *buf; /* buffer header pool */ 75 76static struct proc *bufdaemonproc; 77 78static int inmem(struct vnode * vp, daddr_t blkno); 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 bstrategy(bp); 727 ++readwait; 728 } 729 730 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 731 if (inmem(vp, *rablkno)) 732 continue; 733 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 734 735 if ((rabp->b_flags & B_CACHE) == 0) { 736 if (curthread != PCPU_GET(idlethread)) 737 curthread->td_proc->p_stats->p_ru.ru_inblock++; 738 rabp->b_flags |= B_ASYNC; 739 rabp->b_flags &= ~B_INVAL; 740 rabp->b_ioflags &= ~BIO_ERROR; 741 rabp->b_iocmd = BIO_READ; 742 if (rabp->b_rcred == NOCRED && cred != NOCRED) 743 rabp->b_rcred = crhold(cred); 744 vfs_busy_pages(rabp, 0); 745 BUF_KERNPROC(rabp); 746 rabp->b_iooffset = dbtob(rabp->b_blkno); 747 bstrategy(rabp); 748 } else { 749 brelse(rabp); 750 } 751 } 752 753 if (readwait) { 754 rv = bufwait(bp); 755 } 756 return (rv); 757} 758 759/* 760 * Write, release buffer on completion. (Done by iodone 761 * if async). Do not bother writing anything if the buffer 762 * is invalid. 763 * 764 * Note that we set B_CACHE here, indicating that buffer is 765 * fully valid and thus cacheable. This is true even of NFS 766 * now so we set it generally. This could be set either here 767 * or in biodone() since the I/O is synchronous. We put it 768 * here. 769 */ 770int 771bufwrite(struct buf *bp) 772{ 773 int oldflags, s; 774 struct buf *newbp; 775 776 if (bp->b_flags & B_INVAL) { 777 brelse(bp); 778 return (0); 779 } 780 781 oldflags = bp->b_flags; 782 783 if (BUF_REFCNT(bp) == 0) 784 panic("bufwrite: buffer is not busy???"); 785 s = splbio(); 786 /* 787 * If a background write is already in progress, delay 788 * writing this block if it is asynchronous. Otherwise 789 * wait for the background write to complete. 790 */ 791 BO_LOCK(bp->b_bufobj); 792 if (bp->b_vflags & BV_BKGRDINPROG) { 793 if (bp->b_flags & B_ASYNC) { 794 BO_UNLOCK(bp->b_bufobj); 795 splx(s); 796 bdwrite(bp); 797 return (0); 798 } 799 bp->b_vflags |= BV_BKGRDWAIT; 800 msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), PRIBIO, "bwrbg", 0); 801 if (bp->b_vflags & BV_BKGRDINPROG) 802 panic("bufwrite: still writing"); 803 } 804 BO_UNLOCK(bp->b_bufobj); 805 806 /* Mark the buffer clean */ 807 bundirty(bp); 808 809 /* 810 * If this buffer is marked for background writing and we 811 * do not have to wait for it, make a copy and write the 812 * copy so as to leave this buffer ready for further use. 813 * 814 * This optimization eats a lot of memory. If we have a page 815 * or buffer shortfall we can't do it. 816 */ 817 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 818 (bp->b_flags & B_ASYNC) && 819 !vm_page_count_severe() && 820 !buf_dirty_count_severe()) { 821 KASSERT(bp->b_iodone == NULL, 822 ("bufwrite: needs chained iodone (%p)", bp->b_iodone)); 823 824 /* get a new block */ 825 newbp = geteblk(bp->b_bufsize); 826 827 /* 828 * set it to be identical to the old block. We have to 829 * set b_lblkno and BKGRDMARKER before calling bgetvp() 830 * to avoid confusing the splay tree and gbincore(). 831 */ 832 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 833 newbp->b_lblkno = bp->b_lblkno; 834 newbp->b_xflags |= BX_BKGRDMARKER; 835 BO_LOCK(bp->b_bufobj); 836 bp->b_vflags |= BV_BKGRDINPROG; 837 bgetvp(bp->b_vp, newbp); 838 BO_UNLOCK(bp->b_bufobj); 839 newbp->b_bufobj = &bp->b_vp->v_bufobj; 840 newbp->b_blkno = bp->b_blkno; 841 newbp->b_offset = bp->b_offset; 842 newbp->b_iodone = vfs_backgroundwritedone; 843 newbp->b_flags |= B_ASYNC; 844 newbp->b_flags &= ~B_INVAL; 845 846 /* move over the dependencies */ 847 if (LIST_FIRST(&bp->b_dep) != NULL) 848 buf_movedeps(bp, newbp); 849 850 /* 851 * Initiate write on the copy, release the original to 852 * the B_LOCKED queue so that it cannot go away until 853 * the background write completes. If not locked it could go 854 * away and then be reconstituted while it was being written. 855 * If the reconstituted buffer were written, we could end up 856 * with two background copies being written at the same time. 857 */ 858 bqrelse(bp); 859 bp = newbp; 860 } 861 862 bp->b_flags &= ~B_DONE; 863 bp->b_ioflags &= ~BIO_ERROR; 864 bp->b_flags |= B_CACHE; 865 bp->b_iocmd = BIO_WRITE; 866 867 bufobj_wref(bp->b_bufobj); 868 vfs_busy_pages(bp, 1); 869 870 /* 871 * Normal bwrites pipeline writes 872 */ 873 bp->b_runningbufspace = bp->b_bufsize; 874 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 875 876 if (curthread != PCPU_GET(idlethread)) 877 curthread->td_proc->p_stats->p_ru.ru_oublock++; 878 splx(s); 879 if (oldflags & B_ASYNC) 880 BUF_KERNPROC(bp); 881 bp->b_iooffset = dbtob(bp->b_blkno); 882 bstrategy(bp); 883 884 if ((oldflags & B_ASYNC) == 0) { 885 int rtval = bufwait(bp); 886 brelse(bp); 887 return (rtval); 888 } else { 889 /* 890 * don't allow the async write to saturate the I/O 891 * system. We will not deadlock here because 892 * we are blocking waiting for I/O that is already in-progress 893 * to complete. We do not block here if it is the update 894 * or syncer daemon trying to clean up as that can lead 895 * to deadlock. 896 */ 897 if (curthread->td_proc != bufdaemonproc && 898 curthread->td_proc != updateproc) 899 waitrunningbufspace(); 900 } 901 902 return (0); 903} 904 905/* 906 * Complete a background write started from bwrite. 907 */ 908static void 909vfs_backgroundwritedone(struct buf *bp) 910{ 911 struct buf *origbp; 912 913 /* 914 * Find the original buffer that we are writing. 915 */ 916 BO_LOCK(bp->b_bufobj); 917 if ((origbp = gbincore(bp->b_bufobj, bp->b_lblkno)) == NULL) 918 panic("backgroundwritedone: lost buffer"); 919 920 /* 921 * Clear the BV_BKGRDINPROG flag in the original buffer 922 * and awaken it if it is waiting for the write to complete. 923 * If BV_BKGRDINPROG is not set in the original buffer it must 924 * have been released and re-instantiated - which is not legal. 925 */ 926 KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 927 ("backgroundwritedone: lost buffer2")); 928 origbp->b_vflags &= ~BV_BKGRDINPROG; 929 if (origbp->b_vflags & BV_BKGRDWAIT) { 930 origbp->b_vflags &= ~BV_BKGRDWAIT; 931 wakeup(&origbp->b_xflags); 932 } 933 BO_UNLOCK(bp->b_bufobj); 934 /* 935 * Process dependencies then return any unfinished ones. 936 */ 937 if (LIST_FIRST(&bp->b_dep) != NULL) 938 buf_complete(bp); 939 if (LIST_FIRST(&bp->b_dep) != NULL) 940 buf_movedeps(bp, origbp); 941 942 /* 943 * This buffer is marked B_NOCACHE, so when it is released 944 * by biodone, it will be tossed. We mark it with BIO_READ 945 * to avoid biodone doing a second bufobj_wdrop. 946 */ 947 bp->b_flags |= B_NOCACHE; 948 bp->b_iocmd = BIO_READ; 949 bp->b_flags &= ~(B_CACHE | B_DONE); 950 bp->b_iodone = 0; 951 bufdone(bp); 952} 953 954/* 955 * Delayed write. (Buffer is marked dirty). Do not bother writing 956 * anything if the buffer is marked invalid. 957 * 958 * Note that since the buffer must be completely valid, we can safely 959 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 960 * biodone() in order to prevent getblk from writing the buffer 961 * out synchronously. 962 */ 963void 964bdwrite(struct buf *bp) 965{ 966 struct thread *td = curthread; 967 struct vnode *vp; 968 struct buf *nbp; 969 struct bufobj *bo; 970 971 GIANT_REQUIRED; 972 973 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 974 KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy")); 975 976 if (bp->b_flags & B_INVAL) { 977 brelse(bp); 978 return; 979 } 980 981 /* 982 * If we have too many dirty buffers, don't create any more. 983 * If we are wildly over our limit, then force a complete 984 * cleanup. Otherwise, just keep the situation from getting 985 * out of control. Note that we have to avoid a recursive 986 * disaster and not try to clean up after our own cleanup! 987 */ 988 vp = bp->b_vp; 989 bo = bp->b_bufobj; 990 BO_LOCK(bo); 991 if (td->td_pflags & TDP_COWINPROGRESS) { 992 recursiveflushes++; 993 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 994 BO_UNLOCK(bo); 995 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 996 BO_LOCK(bo); 997 altbufferflushes++; 998 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 999 /* 1000 * Try to find a buffer to flush. 1001 */ 1002 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1003 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1004 buf_countdeps(nbp, 0) || 1005 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1006 continue; 1007 if (bp == nbp) 1008 panic("bdwrite: found ourselves"); 1009 BO_UNLOCK(bo); 1010 if (nbp->b_flags & B_CLUSTEROK) { 1011 vfs_bio_awrite(nbp); 1012 } else { 1013 bremfree(nbp); 1014 bawrite(nbp); 1015 } 1016 BO_LOCK(bo); 1017 dirtybufferflushes++; 1018 break; 1019 } 1020 } 1021 BO_UNLOCK(bo); 1022 1023 bdirty(bp); 1024 /* 1025 * Set B_CACHE, indicating that the buffer is fully valid. This is 1026 * true even of NFS now. 1027 */ 1028 bp->b_flags |= B_CACHE; 1029 1030 /* 1031 * This bmap keeps the system from needing to do the bmap later, 1032 * perhaps when the system is attempting to do a sync. Since it 1033 * is likely that the indirect block -- or whatever other datastructure 1034 * that the filesystem needs is still in memory now, it is a good 1035 * thing to do this. Note also, that if the pageout daemon is 1036 * requesting a sync -- there might not be enough memory to do 1037 * the bmap then... So, this is important to do. 1038 */ 1039 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1040 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1041 } 1042 1043 /* 1044 * Set the *dirty* buffer range based upon the VM system dirty pages. 1045 */ 1046 vfs_setdirty(bp); 1047 1048 /* 1049 * We need to do this here to satisfy the vnode_pager and the 1050 * pageout daemon, so that it thinks that the pages have been 1051 * "cleaned". Note that since the pages are in a delayed write 1052 * buffer -- the VFS layer "will" see that the pages get written 1053 * out on the next sync, or perhaps the cluster will be completed. 1054 */ 1055 vfs_clean_pages(bp); 1056 bqrelse(bp); 1057 1058 /* 1059 * Wakeup the buffer flushing daemon if we have a lot of dirty 1060 * buffers (midpoint between our recovery point and our stall 1061 * point). 1062 */ 1063 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1064 1065 /* 1066 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1067 * due to the softdep code. 1068 */ 1069} 1070 1071/* 1072 * bdirty: 1073 * 1074 * Turn buffer into delayed write request. We must clear BIO_READ and 1075 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1076 * itself to properly update it in the dirty/clean lists. We mark it 1077 * B_DONE to ensure that any asynchronization of the buffer properly 1078 * clears B_DONE ( else a panic will occur later ). 1079 * 1080 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1081 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1082 * should only be called if the buffer is known-good. 1083 * 1084 * Since the buffer is not on a queue, we do not update the numfreebuffers 1085 * count. 1086 * 1087 * Must be called at splbio(). 1088 * The buffer must be on QUEUE_NONE. 1089 */ 1090void 1091bdirty(struct buf *bp) 1092{ 1093 1094 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1095 KASSERT(bp->b_qindex == QUEUE_NONE, 1096 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1097 bp->b_flags &= ~(B_RELBUF); 1098 bp->b_iocmd = BIO_WRITE; 1099 1100 if ((bp->b_flags & B_DELWRI) == 0) { 1101 bp->b_flags |= B_DONE | B_DELWRI; 1102 reassignbuf(bp); 1103 atomic_add_int(&numdirtybuffers, 1); 1104 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1105 } 1106} 1107 1108/* 1109 * bundirty: 1110 * 1111 * Clear B_DELWRI for buffer. 1112 * 1113 * Since the buffer is not on a queue, we do not update the numfreebuffers 1114 * count. 1115 * 1116 * Must be called at splbio(). 1117 * The buffer must be on QUEUE_NONE. 1118 */ 1119 1120void 1121bundirty(struct buf *bp) 1122{ 1123 1124 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1125 KASSERT(bp->b_qindex == QUEUE_NONE, 1126 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1127 1128 if (bp->b_flags & B_DELWRI) { 1129 bp->b_flags &= ~B_DELWRI; 1130 reassignbuf(bp); 1131 atomic_subtract_int(&numdirtybuffers, 1); 1132 numdirtywakeup(lodirtybuffers); 1133 } 1134 /* 1135 * Since it is now being written, we can clear its deferred write flag. 1136 */ 1137 bp->b_flags &= ~B_DEFERRED; 1138} 1139 1140/* 1141 * bawrite: 1142 * 1143 * Asynchronous write. Start output on a buffer, but do not wait for 1144 * it to complete. The buffer is released when the output completes. 1145 * 1146 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1147 * B_INVAL buffers. Not us. 1148 */ 1149void 1150bawrite(struct buf *bp) 1151{ 1152 1153 bp->b_flags |= B_ASYNC; 1154 (void) bwrite(bp); 1155} 1156 1157/* 1158 * bwillwrite: 1159 * 1160 * Called prior to the locking of any vnodes when we are expecting to 1161 * write. We do not want to starve the buffer cache with too many 1162 * dirty buffers so we block here. By blocking prior to the locking 1163 * of any vnodes we attempt to avoid the situation where a locked vnode 1164 * prevents the various system daemons from flushing related buffers. 1165 */ 1166 1167void 1168bwillwrite(void) 1169{ 1170 1171 if (numdirtybuffers >= hidirtybuffers) { 1172 int s; 1173 1174 mtx_lock(&Giant); 1175 s = splbio(); 1176 mtx_lock(&nblock); 1177 while (numdirtybuffers >= hidirtybuffers) { 1178 bd_wakeup(1); 1179 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1180 msleep(&needsbuffer, &nblock, 1181 (PRIBIO + 4), "flswai", 0); 1182 } 1183 splx(s); 1184 mtx_unlock(&nblock); 1185 mtx_unlock(&Giant); 1186 } 1187} 1188 1189/* 1190 * Return true if we have too many dirty buffers. 1191 */ 1192int 1193buf_dirty_count_severe(void) 1194{ 1195 1196 return(numdirtybuffers >= hidirtybuffers); 1197} 1198 1199/* 1200 * brelse: 1201 * 1202 * Release a busy buffer and, if requested, free its resources. The 1203 * buffer will be stashed in the appropriate bufqueue[] allowing it 1204 * to be accessed later as a cache entity or reused for other purposes. 1205 */ 1206void 1207brelse(struct buf *bp) 1208{ 1209 int s; 1210 1211 GIANT_REQUIRED; 1212 1213 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1214 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1215 1216 s = splbio(); 1217 1218 if (bp->b_iocmd == BIO_WRITE && 1219 (bp->b_ioflags & BIO_ERROR) && 1220 !(bp->b_flags & B_INVAL)) { 1221 /* 1222 * Failed write, redirty. Must clear BIO_ERROR to prevent 1223 * pages from being scrapped. If B_INVAL is set then 1224 * this case is not run and the next case is run to 1225 * destroy the buffer. B_INVAL can occur if the buffer 1226 * is outside the range supported by the underlying device. 1227 */ 1228 bp->b_ioflags &= ~BIO_ERROR; 1229 bdirty(bp); 1230 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1231 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1232 /* 1233 * Either a failed I/O or we were asked to free or not 1234 * cache the buffer. 1235 */ 1236 bp->b_flags |= B_INVAL; 1237 if (LIST_FIRST(&bp->b_dep) != NULL) 1238 buf_deallocate(bp); 1239 if (bp->b_flags & B_DELWRI) { 1240 atomic_subtract_int(&numdirtybuffers, 1); 1241 numdirtywakeup(lodirtybuffers); 1242 } 1243 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1244 if ((bp->b_flags & B_VMIO) == 0) { 1245 if (bp->b_bufsize) 1246 allocbuf(bp, 0); 1247 if (bp->b_vp) 1248 brelvp(bp); 1249 } 1250 } 1251 1252 /* 1253 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1254 * is called with B_DELWRI set, the underlying pages may wind up 1255 * getting freed causing a previous write (bdwrite()) to get 'lost' 1256 * because pages associated with a B_DELWRI bp are marked clean. 1257 * 1258 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1259 * if B_DELWRI is set. 1260 * 1261 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1262 * on pages to return pages to the VM page queues. 1263 */ 1264 if (bp->b_flags & B_DELWRI) 1265 bp->b_flags &= ~B_RELBUF; 1266 else if (vm_page_count_severe()) { 1267 /* 1268 * XXX This lock may not be necessary since BKGRDINPROG 1269 * cannot be set while we hold the buf lock, it can only be 1270 * cleared if it is already pending. 1271 */ 1272 if (bp->b_vp) { 1273 BO_LOCK(bp->b_bufobj); 1274 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1275 bp->b_flags |= B_RELBUF; 1276 BO_UNLOCK(bp->b_bufobj); 1277 } else 1278 bp->b_flags |= B_RELBUF; 1279 } 1280 1281 /* 1282 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1283 * constituted, not even NFS buffers now. Two flags effect this. If 1284 * B_INVAL, the struct buf is invalidated but the VM object is kept 1285 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1286 * 1287 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1288 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1289 * buffer is also B_INVAL because it hits the re-dirtying code above. 1290 * 1291 * Normally we can do this whether a buffer is B_DELWRI or not. If 1292 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1293 * the commit state and we cannot afford to lose the buffer. If the 1294 * buffer has a background write in progress, we need to keep it 1295 * around to prevent it from being reconstituted and starting a second 1296 * background write. 1297 */ 1298 if ((bp->b_flags & B_VMIO) 1299 && !(bp->b_vp->v_mount != NULL && 1300 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1301 !vn_isdisk(bp->b_vp, NULL) && 1302 (bp->b_flags & B_DELWRI)) 1303 ) { 1304 1305 int i, j, resid; 1306 vm_page_t m; 1307 off_t foff; 1308 vm_pindex_t poff; 1309 vm_object_t obj; 1310 1311 obj = bp->b_object; 1312 1313 /* 1314 * Get the base offset and length of the buffer. Note that 1315 * in the VMIO case if the buffer block size is not 1316 * page-aligned then b_data pointer may not be page-aligned. 1317 * But our b_pages[] array *IS* page aligned. 1318 * 1319 * block sizes less then DEV_BSIZE (usually 512) are not 1320 * supported due to the page granularity bits (m->valid, 1321 * m->dirty, etc...). 1322 * 1323 * See man buf(9) for more information 1324 */ 1325 resid = bp->b_bufsize; 1326 foff = bp->b_offset; 1327 VM_OBJECT_LOCK(obj); 1328 for (i = 0; i < bp->b_npages; i++) { 1329 int had_bogus = 0; 1330 1331 m = bp->b_pages[i]; 1332 1333 /* 1334 * If we hit a bogus page, fixup *all* the bogus pages 1335 * now. 1336 */ 1337 if (m == bogus_page) { 1338 poff = OFF_TO_IDX(bp->b_offset); 1339 had_bogus = 1; 1340 1341 for (j = i; j < bp->b_npages; j++) { 1342 vm_page_t mtmp; 1343 mtmp = bp->b_pages[j]; 1344 if (mtmp == bogus_page) { 1345 mtmp = vm_page_lookup(obj, poff + j); 1346 if (!mtmp) { 1347 panic("brelse: page missing\n"); 1348 } 1349 bp->b_pages[j] = mtmp; 1350 } 1351 } 1352 1353 if ((bp->b_flags & B_INVAL) == 0) { 1354 pmap_qenter( 1355 trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1356 } 1357 m = bp->b_pages[i]; 1358 } 1359 if ((bp->b_flags & B_NOCACHE) || 1360 (bp->b_ioflags & BIO_ERROR)) { 1361 int poffset = foff & PAGE_MASK; 1362 int presid = resid > (PAGE_SIZE - poffset) ? 1363 (PAGE_SIZE - poffset) : resid; 1364 1365 KASSERT(presid >= 0, ("brelse: extra page")); 1366 vm_page_lock_queues(); 1367 vm_page_set_invalid(m, poffset, presid); 1368 vm_page_unlock_queues(); 1369 if (had_bogus) 1370 printf("avoided corruption bug in bogus_page/brelse code\n"); 1371 } 1372 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1373 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1374 } 1375 VM_OBJECT_UNLOCK(obj); 1376 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1377 vfs_vmio_release(bp); 1378 1379 } else if (bp->b_flags & B_VMIO) { 1380 1381 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1382 vfs_vmio_release(bp); 1383 } 1384 1385 } 1386 1387 if (bp->b_qindex != QUEUE_NONE) 1388 panic("brelse: free buffer onto another queue???"); 1389 if (BUF_REFCNT(bp) > 1) { 1390 /* do not release to free list */ 1391 BUF_UNLOCK(bp); 1392 splx(s); 1393 return; 1394 } 1395 1396 /* enqueue */ 1397 mtx_lock(&bqlock); 1398 1399 /* buffers with no memory */ 1400 if (bp->b_bufsize == 0) { 1401 bp->b_flags |= B_INVAL; 1402 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1403 if (bp->b_vflags & BV_BKGRDINPROG) 1404 panic("losing buffer 1"); 1405 if (bp->b_kvasize) { 1406 bp->b_qindex = QUEUE_EMPTYKVA; 1407 } else { 1408 bp->b_qindex = QUEUE_EMPTY; 1409 } 1410 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1411 bp->b_dev = NULL; 1412 /* buffers with junk contents */ 1413 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1414 (bp->b_ioflags & BIO_ERROR)) { 1415 bp->b_flags |= B_INVAL; 1416 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1417 if (bp->b_vflags & BV_BKGRDINPROG) 1418 panic("losing buffer 2"); 1419 bp->b_qindex = QUEUE_CLEAN; 1420 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1421 bp->b_dev = NULL; 1422 /* remaining buffers */ 1423 } else { 1424 if (bp->b_flags & B_DELWRI) 1425 bp->b_qindex = QUEUE_DIRTY; 1426 else 1427 bp->b_qindex = QUEUE_CLEAN; 1428 if (bp->b_flags & B_AGE) 1429 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1430 else 1431 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1432 } 1433 mtx_unlock(&bqlock); 1434 1435 /* 1436 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1437 * placed the buffer on the correct queue. We must also disassociate 1438 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1439 * find it. 1440 */ 1441 if (bp->b_flags & B_INVAL) { 1442 if (bp->b_flags & B_DELWRI) 1443 bundirty(bp); 1444 if (bp->b_vp) 1445 brelvp(bp); 1446 } 1447 1448 /* 1449 * Fixup numfreebuffers count. The bp is on an appropriate queue 1450 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1451 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1452 * if B_INVAL is set ). 1453 */ 1454 1455 if (!(bp->b_flags & B_DELWRI)) 1456 bufcountwakeup(); 1457 1458 /* 1459 * Something we can maybe free or reuse 1460 */ 1461 if (bp->b_bufsize || bp->b_kvasize) 1462 bufspacewakeup(); 1463 1464 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1465 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1466 panic("brelse: not dirty"); 1467 /* unlock */ 1468 BUF_UNLOCK(bp); 1469 splx(s); 1470} 1471 1472/* 1473 * Release a buffer back to the appropriate queue but do not try to free 1474 * it. The buffer is expected to be used again soon. 1475 * 1476 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1477 * biodone() to requeue an async I/O on completion. It is also used when 1478 * known good buffers need to be requeued but we think we may need the data 1479 * again soon. 1480 * 1481 * XXX we should be able to leave the B_RELBUF hint set on completion. 1482 */ 1483void 1484bqrelse(struct buf *bp) 1485{ 1486 int s; 1487 1488 s = splbio(); 1489 1490 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1491 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1492 1493 if (bp->b_qindex != QUEUE_NONE) 1494 panic("bqrelse: free buffer onto another queue???"); 1495 if (BUF_REFCNT(bp) > 1) { 1496 /* do not release to free list */ 1497 BUF_UNLOCK(bp); 1498 splx(s); 1499 return; 1500 } 1501 mtx_lock(&bqlock); 1502 /* buffers with stale but valid contents */ 1503 if (bp->b_flags & B_DELWRI) { 1504 bp->b_qindex = QUEUE_DIRTY; 1505 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1506 } else { 1507 /* 1508 * XXX This lock may not be necessary since BKGRDINPROG 1509 * cannot be set while we hold the buf lock, it can only be 1510 * cleared if it is already pending. 1511 */ 1512 BO_LOCK(bp->b_bufobj); 1513 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1514 BO_UNLOCK(bp->b_bufobj); 1515 bp->b_qindex = QUEUE_CLEAN; 1516 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1517 b_freelist); 1518 } else { 1519 /* 1520 * We are too low on memory, we have to try to free 1521 * the buffer (most importantly: the wired pages 1522 * making up its backing store) *now*. 1523 */ 1524 BO_UNLOCK(bp->b_bufobj); 1525 mtx_unlock(&bqlock); 1526 splx(s); 1527 brelse(bp); 1528 return; 1529 } 1530 } 1531 mtx_unlock(&bqlock); 1532 1533 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1534 bufcountwakeup(); 1535 1536 /* 1537 * Something we can maybe free or reuse. 1538 */ 1539 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1540 bufspacewakeup(); 1541 1542 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1543 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1544 panic("bqrelse: not dirty"); 1545 /* unlock */ 1546 BUF_UNLOCK(bp); 1547 splx(s); 1548} 1549 1550/* Give pages used by the bp back to the VM system (where possible) */ 1551static void 1552vfs_vmio_release(struct buf *bp) 1553{ 1554 int i; 1555 vm_page_t m; 1556 1557 GIANT_REQUIRED; 1558 VM_OBJECT_LOCK(bp->b_object); 1559 vm_page_lock_queues(); 1560 for (i = 0; i < bp->b_npages; i++) { 1561 m = bp->b_pages[i]; 1562 bp->b_pages[i] = NULL; 1563 /* 1564 * In order to keep page LRU ordering consistent, put 1565 * everything on the inactive queue. 1566 */ 1567 vm_page_unwire(m, 0); 1568 /* 1569 * We don't mess with busy pages, it is 1570 * the responsibility of the process that 1571 * busied the pages to deal with them. 1572 */ 1573 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1574 continue; 1575 1576 if (m->wire_count == 0) { 1577 /* 1578 * Might as well free the page if we can and it has 1579 * no valid data. We also free the page if the 1580 * buffer was used for direct I/O 1581 */ 1582 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1583 m->hold_count == 0) { 1584 vm_page_busy(m); 1585 pmap_remove_all(m); 1586 vm_page_free(m); 1587 } else if (bp->b_flags & B_DIRECT) { 1588 vm_page_try_to_free(m); 1589 } else if (vm_page_count_severe()) { 1590 vm_page_try_to_cache(m); 1591 } 1592 } 1593 } 1594 vm_page_unlock_queues(); 1595 VM_OBJECT_UNLOCK(bp->b_object); 1596 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1597 1598 if (bp->b_bufsize) { 1599 bufspacewakeup(); 1600 bp->b_bufsize = 0; 1601 } 1602 bp->b_npages = 0; 1603 bp->b_flags &= ~B_VMIO; 1604 if (bp->b_vp) 1605 brelvp(bp); 1606} 1607 1608/* 1609 * Check to see if a block at a particular lbn is available for a clustered 1610 * write. 1611 */ 1612static int 1613vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1614{ 1615 struct buf *bpa; 1616 int match; 1617 1618 match = 0; 1619 1620 /* If the buf isn't in core skip it */ 1621 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1622 return (0); 1623 1624 /* If the buf is busy we don't want to wait for it */ 1625 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1626 return (0); 1627 1628 /* Only cluster with valid clusterable delayed write buffers */ 1629 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1630 (B_DELWRI | B_CLUSTEROK)) 1631 goto done; 1632 1633 if (bpa->b_bufsize != size) 1634 goto done; 1635 1636 /* 1637 * Check to see if it is in the expected place on disk and that the 1638 * block has been mapped. 1639 */ 1640 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1641 match = 1; 1642done: 1643 BUF_UNLOCK(bpa); 1644 return (match); 1645} 1646 1647/* 1648 * vfs_bio_awrite: 1649 * 1650 * Implement clustered async writes for clearing out B_DELWRI buffers. 1651 * This is much better then the old way of writing only one buffer at 1652 * a time. Note that we may not be presented with the buffers in the 1653 * correct order, so we search for the cluster in both directions. 1654 */ 1655int 1656vfs_bio_awrite(struct buf *bp) 1657{ 1658 int i; 1659 int j; 1660 daddr_t lblkno = bp->b_lblkno; 1661 struct vnode *vp = bp->b_vp; 1662 int s; 1663 int ncl; 1664 int nwritten; 1665 int size; 1666 int maxcl; 1667 1668 s = splbio(); 1669 /* 1670 * right now we support clustered writing only to regular files. If 1671 * we find a clusterable block we could be in the middle of a cluster 1672 * rather then at the beginning. 1673 */ 1674 if ((vp->v_type == VREG) && 1675 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1676 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1677 1678 size = vp->v_mount->mnt_stat.f_iosize; 1679 maxcl = MAXPHYS / size; 1680 1681 VI_LOCK(vp); 1682 for (i = 1; i < maxcl; i++) 1683 if (vfs_bio_clcheck(vp, size, lblkno + i, 1684 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1685 break; 1686 1687 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1688 if (vfs_bio_clcheck(vp, size, lblkno - j, 1689 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1690 break; 1691 1692 VI_UNLOCK(vp); 1693 --j; 1694 ncl = i + j; 1695 /* 1696 * this is a possible cluster write 1697 */ 1698 if (ncl != 1) { 1699 BUF_UNLOCK(bp); 1700 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1701 splx(s); 1702 return nwritten; 1703 } 1704 } 1705 1706 bremfree(bp); 1707 bp->b_flags |= B_ASYNC; 1708 1709 splx(s); 1710 /* 1711 * default (old) behavior, writing out only one block 1712 * 1713 * XXX returns b_bufsize instead of b_bcount for nwritten? 1714 */ 1715 nwritten = bp->b_bufsize; 1716 (void) bwrite(bp); 1717 1718 return nwritten; 1719} 1720 1721/* 1722 * getnewbuf: 1723 * 1724 * Find and initialize a new buffer header, freeing up existing buffers 1725 * in the bufqueues as necessary. The new buffer is returned locked. 1726 * 1727 * Important: B_INVAL is not set. If the caller wishes to throw the 1728 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1729 * 1730 * We block if: 1731 * We have insufficient buffer headers 1732 * We have insufficient buffer space 1733 * buffer_map is too fragmented ( space reservation fails ) 1734 * If we have to flush dirty buffers ( but we try to avoid this ) 1735 * 1736 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1737 * Instead we ask the buf daemon to do it for us. We attempt to 1738 * avoid piecemeal wakeups of the pageout daemon. 1739 */ 1740 1741static struct buf * 1742getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1743{ 1744 struct buf *bp; 1745 struct buf *nbp; 1746 int defrag = 0; 1747 int nqindex; 1748 static int flushingbufs; 1749 1750 GIANT_REQUIRED; 1751 1752 /* 1753 * We can't afford to block since we might be holding a vnode lock, 1754 * which may prevent system daemons from running. We deal with 1755 * low-memory situations by proactively returning memory and running 1756 * async I/O rather then sync I/O. 1757 */ 1758 1759 atomic_add_int(&getnewbufcalls, 1); 1760 atomic_subtract_int(&getnewbufrestarts, 1); 1761restart: 1762 atomic_add_int(&getnewbufrestarts, 1); 1763 1764 /* 1765 * Setup for scan. If we do not have enough free buffers, 1766 * we setup a degenerate case that immediately fails. Note 1767 * that if we are specially marked process, we are allowed to 1768 * dip into our reserves. 1769 * 1770 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1771 * 1772 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1773 * However, there are a number of cases (defragging, reusing, ...) 1774 * where we cannot backup. 1775 */ 1776 mtx_lock(&bqlock); 1777 nqindex = QUEUE_EMPTYKVA; 1778 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1779 1780 if (nbp == NULL) { 1781 /* 1782 * If no EMPTYKVA buffers and we are either 1783 * defragging or reusing, locate a CLEAN buffer 1784 * to free or reuse. If bufspace useage is low 1785 * skip this step so we can allocate a new buffer. 1786 */ 1787 if (defrag || bufspace >= lobufspace) { 1788 nqindex = QUEUE_CLEAN; 1789 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1790 } 1791 1792 /* 1793 * If we could not find or were not allowed to reuse a 1794 * CLEAN buffer, check to see if it is ok to use an EMPTY 1795 * buffer. We can only use an EMPTY buffer if allocating 1796 * its KVA would not otherwise run us out of buffer space. 1797 */ 1798 if (nbp == NULL && defrag == 0 && 1799 bufspace + maxsize < hibufspace) { 1800 nqindex = QUEUE_EMPTY; 1801 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1802 } 1803 } 1804 1805 /* 1806 * Run scan, possibly freeing data and/or kva mappings on the fly 1807 * depending. 1808 */ 1809 1810 while ((bp = nbp) != NULL) { 1811 int qindex = nqindex; 1812 1813 /* 1814 * Calculate next bp ( we can only use it if we do not block 1815 * or do other fancy things ). 1816 */ 1817 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1818 switch(qindex) { 1819 case QUEUE_EMPTY: 1820 nqindex = QUEUE_EMPTYKVA; 1821 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1822 break; 1823 /* FALLTHROUGH */ 1824 case QUEUE_EMPTYKVA: 1825 nqindex = QUEUE_CLEAN; 1826 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1827 break; 1828 /* FALLTHROUGH */ 1829 case QUEUE_CLEAN: 1830 /* 1831 * nbp is NULL. 1832 */ 1833 break; 1834 } 1835 } 1836 if (bp->b_vp) { 1837 BO_LOCK(bp->b_bufobj); 1838 if (bp->b_vflags & BV_BKGRDINPROG) { 1839 BO_UNLOCK(bp->b_bufobj); 1840 continue; 1841 } 1842 BO_UNLOCK(bp->b_bufobj); 1843 } 1844 1845 /* 1846 * Sanity Checks 1847 */ 1848 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1849 1850 /* 1851 * Note: we no longer distinguish between VMIO and non-VMIO 1852 * buffers. 1853 */ 1854 1855 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1856 1857 /* 1858 * If we are defragging then we need a buffer with 1859 * b_kvasize != 0. XXX this situation should no longer 1860 * occur, if defrag is non-zero the buffer's b_kvasize 1861 * should also be non-zero at this point. XXX 1862 */ 1863 if (defrag && bp->b_kvasize == 0) { 1864 printf("Warning: defrag empty buffer %p\n", bp); 1865 continue; 1866 } 1867 1868 /* 1869 * Start freeing the bp. This is somewhat involved. nbp 1870 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1871 */ 1872 1873 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1874 panic("getnewbuf: locked buf"); 1875 bremfreel(bp); 1876 mtx_unlock(&bqlock); 1877 1878 if (qindex == QUEUE_CLEAN) { 1879 if (bp->b_flags & B_VMIO) { 1880 bp->b_flags &= ~B_ASYNC; 1881 vfs_vmio_release(bp); 1882 } 1883 if (bp->b_vp) 1884 brelvp(bp); 1885 } 1886 1887 /* 1888 * NOTE: nbp is now entirely invalid. We can only restart 1889 * the scan from this point on. 1890 * 1891 * Get the rest of the buffer freed up. b_kva* is still 1892 * valid after this operation. 1893 */ 1894 1895 if (bp->b_rcred != NOCRED) { 1896 crfree(bp->b_rcred); 1897 bp->b_rcred = NOCRED; 1898 } 1899 if (bp->b_wcred != NOCRED) { 1900 crfree(bp->b_wcred); 1901 bp->b_wcred = NOCRED; 1902 } 1903 if (LIST_FIRST(&bp->b_dep) != NULL) 1904 buf_deallocate(bp); 1905 if (bp->b_vflags & BV_BKGRDINPROG) 1906 panic("losing buffer 3"); 1907 1908 if (bp->b_bufsize) 1909 allocbuf(bp, 0); 1910 1911 bp->b_flags = 0; 1912 bp->b_ioflags = 0; 1913 bp->b_xflags = 0; 1914 bp->b_vflags = 0; 1915 bp->b_dev = NULL; 1916 bp->b_vp = NULL; 1917 bp->b_blkno = bp->b_lblkno = 0; 1918 bp->b_offset = NOOFFSET; 1919 bp->b_iodone = 0; 1920 bp->b_error = 0; 1921 bp->b_resid = 0; 1922 bp->b_bcount = 0; 1923 bp->b_npages = 0; 1924 bp->b_dirtyoff = bp->b_dirtyend = 0; 1925 bp->b_magic = B_MAGIC_BIO; 1926 bp->b_bufobj = NULL; 1927 1928 LIST_INIT(&bp->b_dep); 1929 1930 /* 1931 * If we are defragging then free the buffer. 1932 */ 1933 if (defrag) { 1934 bp->b_flags |= B_INVAL; 1935 bfreekva(bp); 1936 brelse(bp); 1937 defrag = 0; 1938 goto restart; 1939 } 1940 1941 /* 1942 * If we are overcomitted then recover the buffer and its 1943 * KVM space. This occurs in rare situations when multiple 1944 * processes are blocked in getnewbuf() or allocbuf(). 1945 */ 1946 if (bufspace >= hibufspace) 1947 flushingbufs = 1; 1948 if (flushingbufs && bp->b_kvasize != 0) { 1949 bp->b_flags |= B_INVAL; 1950 bfreekva(bp); 1951 brelse(bp); 1952 goto restart; 1953 } 1954 if (bufspace < lobufspace) 1955 flushingbufs = 0; 1956 break; 1957 } 1958 1959 /* 1960 * If we exhausted our list, sleep as appropriate. We may have to 1961 * wakeup various daemons and write out some dirty buffers. 1962 * 1963 * Generally we are sleeping due to insufficient buffer space. 1964 */ 1965 1966 if (bp == NULL) { 1967 int flags; 1968 char *waitmsg; 1969 1970 mtx_unlock(&bqlock); 1971 if (defrag) { 1972 flags = VFS_BIO_NEED_BUFSPACE; 1973 waitmsg = "nbufkv"; 1974 } else if (bufspace >= hibufspace) { 1975 waitmsg = "nbufbs"; 1976 flags = VFS_BIO_NEED_BUFSPACE; 1977 } else { 1978 waitmsg = "newbuf"; 1979 flags = VFS_BIO_NEED_ANY; 1980 } 1981 1982 bd_speedup(); /* heeeelp */ 1983 1984 mtx_lock(&nblock); 1985 needsbuffer |= flags; 1986 while (needsbuffer & flags) { 1987 if (msleep(&needsbuffer, &nblock, 1988 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1989 mtx_unlock(&nblock); 1990 return (NULL); 1991 } 1992 } 1993 mtx_unlock(&nblock); 1994 } else { 1995 /* 1996 * We finally have a valid bp. We aren't quite out of the 1997 * woods, we still have to reserve kva space. In order 1998 * to keep fragmentation sane we only allocate kva in 1999 * BKVASIZE chunks. 2000 */ 2001 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2002 2003 if (maxsize != bp->b_kvasize) { 2004 vm_offset_t addr = 0; 2005 2006 bfreekva(bp); 2007 2008 if (vm_map_findspace(buffer_map, 2009 vm_map_min(buffer_map), maxsize, &addr)) { 2010 /* 2011 * Uh oh. Buffer map is to fragmented. We 2012 * must defragment the map. 2013 */ 2014 atomic_add_int(&bufdefragcnt, 1); 2015 defrag = 1; 2016 bp->b_flags |= B_INVAL; 2017 brelse(bp); 2018 goto restart; 2019 } 2020 if (addr) { 2021 vm_map_insert(buffer_map, NULL, 0, 2022 addr, addr + maxsize, 2023 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2024 2025 bp->b_kvabase = (caddr_t) addr; 2026 bp->b_kvasize = maxsize; 2027 atomic_add_int(&bufspace, bp->b_kvasize); 2028 atomic_add_int(&bufreusecnt, 1); 2029 } 2030 } 2031 bp->b_saveaddr = bp->b_kvabase; 2032 bp->b_data = bp->b_saveaddr; 2033 } 2034 return(bp); 2035} 2036 2037/* 2038 * buf_daemon: 2039 * 2040 * buffer flushing daemon. Buffers are normally flushed by the 2041 * update daemon but if it cannot keep up this process starts to 2042 * take the load in an attempt to prevent getnewbuf() from blocking. 2043 */ 2044 2045static struct kproc_desc buf_kp = { 2046 "bufdaemon", 2047 buf_daemon, 2048 &bufdaemonproc 2049}; 2050SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2051 2052static void 2053buf_daemon() 2054{ 2055 int s; 2056 2057 mtx_lock(&Giant); 2058 2059 /* 2060 * This process needs to be suspended prior to shutdown sync. 2061 */ 2062 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2063 SHUTDOWN_PRI_LAST); 2064 2065 /* 2066 * This process is allowed to take the buffer cache to the limit 2067 */ 2068 s = splbio(); 2069 mtx_lock(&bdlock); 2070 2071 for (;;) { 2072 bd_request = 0; 2073 mtx_unlock(&bdlock); 2074 2075 kthread_suspend_check(bufdaemonproc); 2076 2077 /* 2078 * Do the flush. Limit the amount of in-transit I/O we 2079 * allow to build up, otherwise we would completely saturate 2080 * the I/O system. Wakeup any waiting processes before we 2081 * normally would so they can run in parallel with our drain. 2082 */ 2083 while (numdirtybuffers > lodirtybuffers) { 2084 if (flushbufqueues(0) == 0) { 2085 /* 2086 * Could not find any buffers without rollback 2087 * dependencies, so just write the first one 2088 * in the hopes of eventually making progress. 2089 */ 2090 flushbufqueues(1); 2091 break; 2092 } 2093 waitrunningbufspace(); 2094 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2095 } 2096 2097 /* 2098 * Only clear bd_request if we have reached our low water 2099 * mark. The buf_daemon normally waits 1 second and 2100 * then incrementally flushes any dirty buffers that have 2101 * built up, within reason. 2102 * 2103 * If we were unable to hit our low water mark and couldn't 2104 * find any flushable buffers, we sleep half a second. 2105 * Otherwise we loop immediately. 2106 */ 2107 mtx_lock(&bdlock); 2108 if (numdirtybuffers <= lodirtybuffers) { 2109 /* 2110 * We reached our low water mark, reset the 2111 * request and sleep until we are needed again. 2112 * The sleep is just so the suspend code works. 2113 */ 2114 bd_request = 0; 2115 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2116 } else { 2117 /* 2118 * We couldn't find any flushable dirty buffers but 2119 * still have too many dirty buffers, we 2120 * have to sleep and try again. (rare) 2121 */ 2122 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2123 } 2124 } 2125} 2126 2127/* 2128 * flushbufqueues: 2129 * 2130 * Try to flush a buffer in the dirty queue. We must be careful to 2131 * free up B_INVAL buffers instead of write them, which NFS is 2132 * particularly sensitive to. 2133 */ 2134int flushwithdeps = 0; 2135SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2136 0, "Number of buffers flushed with dependecies that require rollbacks"); 2137 2138static int 2139flushbufqueues(int flushdeps) 2140{ 2141 struct thread *td = curthread; 2142 struct vnode *vp; 2143 struct mount *mp; 2144 struct buf *bp; 2145 int hasdeps; 2146 2147 mtx_lock(&bqlock); 2148 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2149 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2150 continue; 2151 KASSERT((bp->b_flags & B_DELWRI), 2152 ("unexpected clean buffer %p", bp)); 2153 BO_LOCK(bp->b_bufobj); 2154 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2155 BO_UNLOCK(bp->b_bufobj); 2156 BUF_UNLOCK(bp); 2157 continue; 2158 } 2159 BO_UNLOCK(bp->b_bufobj); 2160 if (bp->b_flags & B_INVAL) { 2161 bremfreel(bp); 2162 mtx_unlock(&bqlock); 2163 brelse(bp); 2164 return (1); 2165 } 2166 2167 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2168 if (flushdeps == 0) { 2169 BUF_UNLOCK(bp); 2170 continue; 2171 } 2172 hasdeps = 1; 2173 } else 2174 hasdeps = 0; 2175 /* 2176 * We must hold the lock on a vnode before writing 2177 * one of its buffers. Otherwise we may confuse, or 2178 * in the case of a snapshot vnode, deadlock the 2179 * system. 2180 * 2181 * The lock order here is the reverse of the normal 2182 * of vnode followed by buf lock. This is ok because 2183 * the NOWAIT will prevent deadlock. 2184 */ 2185 vp = bp->b_vp; 2186 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2187 BUF_UNLOCK(bp); 2188 continue; 2189 } 2190 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2191 mtx_unlock(&bqlock); 2192 vfs_bio_awrite(bp); 2193 vn_finished_write(mp); 2194 VOP_UNLOCK(vp, 0, td); 2195 flushwithdeps += hasdeps; 2196 return (1); 2197 } 2198 vn_finished_write(mp); 2199 BUF_UNLOCK(bp); 2200 } 2201 mtx_unlock(&bqlock); 2202 return (0); 2203} 2204 2205/* 2206 * Check to see if a block is currently memory resident. 2207 */ 2208struct buf * 2209incore(struct bufobj *bo, daddr_t blkno) 2210{ 2211 struct buf *bp; 2212 2213 int s = splbio(); 2214 BO_LOCK(bo); 2215 bp = gbincore(bo, blkno); 2216 BO_UNLOCK(bo); 2217 splx(s); 2218 return (bp); 2219} 2220 2221/* 2222 * Returns true if no I/O is needed to access the 2223 * associated VM object. This is like incore except 2224 * it also hunts around in the VM system for the data. 2225 */ 2226 2227static int 2228inmem(struct vnode * vp, daddr_t blkno) 2229{ 2230 vm_object_t obj; 2231 vm_offset_t toff, tinc, size; 2232 vm_page_t m; 2233 vm_ooffset_t off; 2234 2235 GIANT_REQUIRED; 2236 ASSERT_VOP_LOCKED(vp, "inmem"); 2237 2238 if (incore(&vp->v_bufobj, blkno)) 2239 return 1; 2240 if (vp->v_mount == NULL) 2241 return 0; 2242 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2243 return 0; 2244 2245 size = PAGE_SIZE; 2246 if (size > vp->v_mount->mnt_stat.f_iosize) 2247 size = vp->v_mount->mnt_stat.f_iosize; 2248 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2249 2250 VM_OBJECT_LOCK(obj); 2251 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2252 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2253 if (!m) 2254 goto notinmem; 2255 tinc = size; 2256 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2257 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2258 if (vm_page_is_valid(m, 2259 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2260 goto notinmem; 2261 } 2262 VM_OBJECT_UNLOCK(obj); 2263 return 1; 2264 2265notinmem: 2266 VM_OBJECT_UNLOCK(obj); 2267 return (0); 2268} 2269 2270/* 2271 * vfs_setdirty: 2272 * 2273 * Sets the dirty range for a buffer based on the status of the dirty 2274 * bits in the pages comprising the buffer. 2275 * 2276 * The range is limited to the size of the buffer. 2277 * 2278 * This routine is primarily used by NFS, but is generalized for the 2279 * B_VMIO case. 2280 */ 2281static void 2282vfs_setdirty(struct buf *bp) 2283{ 2284 int i; 2285 vm_object_t object; 2286 2287 GIANT_REQUIRED; 2288 /* 2289 * Degenerate case - empty buffer 2290 */ 2291 2292 if (bp->b_bufsize == 0) 2293 return; 2294 2295 /* 2296 * We qualify the scan for modified pages on whether the 2297 * object has been flushed yet. The OBJ_WRITEABLE flag 2298 * is not cleared simply by protecting pages off. 2299 */ 2300 2301 if ((bp->b_flags & B_VMIO) == 0) 2302 return; 2303 2304 object = bp->b_pages[0]->object; 2305 VM_OBJECT_LOCK(object); 2306 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2307 printf("Warning: object %p writeable but not mightbedirty\n", object); 2308 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2309 printf("Warning: object %p mightbedirty but not writeable\n", object); 2310 2311 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2312 vm_offset_t boffset; 2313 vm_offset_t eoffset; 2314 2315 vm_page_lock_queues(); 2316 /* 2317 * test the pages to see if they have been modified directly 2318 * by users through the VM system. 2319 */ 2320 for (i = 0; i < bp->b_npages; i++) 2321 vm_page_test_dirty(bp->b_pages[i]); 2322 2323 /* 2324 * Calculate the encompassing dirty range, boffset and eoffset, 2325 * (eoffset - boffset) bytes. 2326 */ 2327 2328 for (i = 0; i < bp->b_npages; i++) { 2329 if (bp->b_pages[i]->dirty) 2330 break; 2331 } 2332 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2333 2334 for (i = bp->b_npages - 1; i >= 0; --i) { 2335 if (bp->b_pages[i]->dirty) { 2336 break; 2337 } 2338 } 2339 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2340 2341 vm_page_unlock_queues(); 2342 /* 2343 * Fit it to the buffer. 2344 */ 2345 2346 if (eoffset > bp->b_bcount) 2347 eoffset = bp->b_bcount; 2348 2349 /* 2350 * If we have a good dirty range, merge with the existing 2351 * dirty range. 2352 */ 2353 2354 if (boffset < eoffset) { 2355 if (bp->b_dirtyoff > boffset) 2356 bp->b_dirtyoff = boffset; 2357 if (bp->b_dirtyend < eoffset) 2358 bp->b_dirtyend = eoffset; 2359 } 2360 } 2361 VM_OBJECT_UNLOCK(object); 2362} 2363 2364/* 2365 * getblk: 2366 * 2367 * Get a block given a specified block and offset into a file/device. 2368 * The buffers B_DONE bit will be cleared on return, making it almost 2369 * ready for an I/O initiation. B_INVAL may or may not be set on 2370 * return. The caller should clear B_INVAL prior to initiating a 2371 * READ. 2372 * 2373 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2374 * an existing buffer. 2375 * 2376 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2377 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2378 * and then cleared based on the backing VM. If the previous buffer is 2379 * non-0-sized but invalid, B_CACHE will be cleared. 2380 * 2381 * If getblk() must create a new buffer, the new buffer is returned with 2382 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2383 * case it is returned with B_INVAL clear and B_CACHE set based on the 2384 * backing VM. 2385 * 2386 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2387 * B_CACHE bit is clear. 2388 * 2389 * What this means, basically, is that the caller should use B_CACHE to 2390 * determine whether the buffer is fully valid or not and should clear 2391 * B_INVAL prior to issuing a read. If the caller intends to validate 2392 * the buffer by loading its data area with something, the caller needs 2393 * to clear B_INVAL. If the caller does this without issuing an I/O, 2394 * the caller should set B_CACHE ( as an optimization ), else the caller 2395 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2396 * a write attempt or if it was a successfull read. If the caller 2397 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2398 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2399 */ 2400struct buf * 2401getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2402 int flags) 2403{ 2404 struct buf *bp; 2405 struct bufobj *bo; 2406 int s; 2407 int error; 2408 ASSERT_VOP_LOCKED(vp, "getblk"); 2409 struct vm_object *vmo; 2410 2411 if (size > MAXBSIZE) 2412 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2413 2414 bo = &vp->v_bufobj; 2415 s = splbio(); 2416loop: 2417 /* 2418 * Block if we are low on buffers. Certain processes are allowed 2419 * to completely exhaust the buffer cache. 2420 * 2421 * If this check ever becomes a bottleneck it may be better to 2422 * move it into the else, when gbincore() fails. At the moment 2423 * it isn't a problem. 2424 * 2425 * XXX remove if 0 sections (clean this up after its proven) 2426 */ 2427 if (numfreebuffers == 0) { 2428 if (curthread == PCPU_GET(idlethread)) 2429 return NULL; 2430 mtx_lock(&nblock); 2431 needsbuffer |= VFS_BIO_NEED_ANY; 2432 mtx_unlock(&nblock); 2433 } 2434 2435 VI_LOCK(vp); 2436 bp = gbincore(bo, blkno); 2437 if (bp != NULL) { 2438 int lockflags; 2439 /* 2440 * Buffer is in-core. If the buffer is not busy, it must 2441 * be on a queue. 2442 */ 2443 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2444 2445 if (flags & GB_LOCK_NOWAIT) 2446 lockflags |= LK_NOWAIT; 2447 2448 error = BUF_TIMELOCK(bp, lockflags, 2449 VI_MTX(vp), "getblk", slpflag, slptimeo); 2450 2451 /* 2452 * If we slept and got the lock we have to restart in case 2453 * the buffer changed identities. 2454 */ 2455 if (error == ENOLCK) 2456 goto loop; 2457 /* We timed out or were interrupted. */ 2458 else if (error) 2459 return (NULL); 2460 2461 /* 2462 * The buffer is locked. B_CACHE is cleared if the buffer is 2463 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2464 * and for a VMIO buffer B_CACHE is adjusted according to the 2465 * backing VM cache. 2466 */ 2467 if (bp->b_flags & B_INVAL) 2468 bp->b_flags &= ~B_CACHE; 2469 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2470 bp->b_flags |= B_CACHE; 2471 bremfree(bp); 2472 2473 /* 2474 * check for size inconsistancies for non-VMIO case. 2475 */ 2476 2477 if (bp->b_bcount != size) { 2478 if ((bp->b_flags & B_VMIO) == 0 || 2479 (size > bp->b_kvasize)) { 2480 if (bp->b_flags & B_DELWRI) { 2481 bp->b_flags |= B_NOCACHE; 2482 bwrite(bp); 2483 } else { 2484 if ((bp->b_flags & B_VMIO) && 2485 (LIST_FIRST(&bp->b_dep) == NULL)) { 2486 bp->b_flags |= B_RELBUF; 2487 brelse(bp); 2488 } else { 2489 bp->b_flags |= B_NOCACHE; 2490 bwrite(bp); 2491 } 2492 } 2493 goto loop; 2494 } 2495 } 2496 2497 /* 2498 * If the size is inconsistant in the VMIO case, we can resize 2499 * the buffer. This might lead to B_CACHE getting set or 2500 * cleared. If the size has not changed, B_CACHE remains 2501 * unchanged from its previous state. 2502 */ 2503 2504 if (bp->b_bcount != size) 2505 allocbuf(bp, size); 2506 2507 KASSERT(bp->b_offset != NOOFFSET, 2508 ("getblk: no buffer offset")); 2509 2510 /* 2511 * A buffer with B_DELWRI set and B_CACHE clear must 2512 * be committed before we can return the buffer in 2513 * order to prevent the caller from issuing a read 2514 * ( due to B_CACHE not being set ) and overwriting 2515 * it. 2516 * 2517 * Most callers, including NFS and FFS, need this to 2518 * operate properly either because they assume they 2519 * can issue a read if B_CACHE is not set, or because 2520 * ( for example ) an uncached B_DELWRI might loop due 2521 * to softupdates re-dirtying the buffer. In the latter 2522 * case, B_CACHE is set after the first write completes, 2523 * preventing further loops. 2524 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2525 * above while extending the buffer, we cannot allow the 2526 * buffer to remain with B_CACHE set after the write 2527 * completes or it will represent a corrupt state. To 2528 * deal with this we set B_NOCACHE to scrap the buffer 2529 * after the write. 2530 * 2531 * We might be able to do something fancy, like setting 2532 * B_CACHE in bwrite() except if B_DELWRI is already set, 2533 * so the below call doesn't set B_CACHE, but that gets real 2534 * confusing. This is much easier. 2535 */ 2536 2537 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2538 bp->b_flags |= B_NOCACHE; 2539 bwrite(bp); 2540 goto loop; 2541 } 2542 2543 splx(s); 2544 bp->b_flags &= ~B_DONE; 2545 } else { 2546 int bsize, maxsize, vmio; 2547 off_t offset; 2548 2549 /* 2550 * Buffer is not in-core, create new buffer. The buffer 2551 * returned by getnewbuf() is locked. Note that the returned 2552 * buffer is also considered valid (not marked B_INVAL). 2553 */ 2554 VI_UNLOCK(vp); 2555 /* 2556 * If the user does not want us to create the buffer, bail out 2557 * here. 2558 */ 2559 if (flags & GB_NOCREAT) { 2560 splx(s); 2561 return NULL; 2562 } 2563 2564 bsize = bo->bo_bsize; 2565 offset = blkno * bsize; 2566 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2567 (vp->v_vflag & VV_OBJBUF); 2568 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2569 maxsize = imax(maxsize, bsize); 2570 2571 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2572 if (bp == NULL) { 2573 if (slpflag || slptimeo) { 2574 splx(s); 2575 return NULL; 2576 } 2577 goto loop; 2578 } 2579 2580 /* 2581 * This code is used to make sure that a buffer is not 2582 * created while the getnewbuf routine is blocked. 2583 * This can be a problem whether the vnode is locked or not. 2584 * If the buffer is created out from under us, we have to 2585 * throw away the one we just created. There is now window 2586 * race because we are safely running at splbio() from the 2587 * point of the duplicate buffer creation through to here, 2588 * and we've locked the buffer. 2589 * 2590 * Note: this must occur before we associate the buffer 2591 * with the vp especially considering limitations in 2592 * the splay tree implementation when dealing with duplicate 2593 * lblkno's. 2594 */ 2595 BO_LOCK(bo); 2596 if (gbincore(bo, blkno)) { 2597 BO_UNLOCK(bo); 2598 bp->b_flags |= B_INVAL; 2599 brelse(bp); 2600 goto loop; 2601 } 2602 2603 /* 2604 * Insert the buffer into the hash, so that it can 2605 * be found by incore. 2606 */ 2607 bp->b_blkno = bp->b_lblkno = blkno; 2608 bp->b_offset = offset; 2609 2610 bgetvp(vp, bp); 2611 BO_UNLOCK(bo); 2612 2613 /* 2614 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2615 * buffer size starts out as 0, B_CACHE will be set by 2616 * allocbuf() for the VMIO case prior to it testing the 2617 * backing store for validity. 2618 */ 2619 2620 if (vmio) { 2621 bp->b_flags |= B_VMIO; 2622#if defined(VFS_BIO_DEBUG) 2623 if (vn_canvmio(vp) != TRUE) 2624 printf("getblk: VMIO on vnode type %d\n", 2625 vp->v_type); 2626#endif 2627 VOP_GETVOBJECT(vp, &vmo); 2628 KASSERT(vmo == bp->b_object, 2629 ("ARGH! different b_object %p %p %p\n", bp, vmo, bp->b_object)); 2630 } else { 2631 bp->b_flags &= ~B_VMIO; 2632 KASSERT(bp->b_object == NULL, 2633 ("ARGH! has b_object %p %p\n", bp, bp->b_object)); 2634 } 2635 2636 allocbuf(bp, size); 2637 2638 splx(s); 2639 bp->b_flags &= ~B_DONE; 2640 } 2641 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2642 KASSERT(bp->b_bufobj == bo, 2643 ("wrong b_bufobj %p should be %p", bp->b_bufobj, bo)); 2644 return (bp); 2645} 2646 2647/* 2648 * Get an empty, disassociated buffer of given size. The buffer is initially 2649 * set to B_INVAL. 2650 */ 2651struct buf * 2652geteblk(int size) 2653{ 2654 struct buf *bp; 2655 int s; 2656 int maxsize; 2657 2658 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2659 2660 s = splbio(); 2661 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2662 continue; 2663 splx(s); 2664 allocbuf(bp, size); 2665 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2666 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2667 return (bp); 2668} 2669 2670 2671/* 2672 * This code constitutes the buffer memory from either anonymous system 2673 * memory (in the case of non-VMIO operations) or from an associated 2674 * VM object (in the case of VMIO operations). This code is able to 2675 * resize a buffer up or down. 2676 * 2677 * Note that this code is tricky, and has many complications to resolve 2678 * deadlock or inconsistant data situations. Tread lightly!!! 2679 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2680 * the caller. Calling this code willy nilly can result in the loss of data. 2681 * 2682 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2683 * B_CACHE for the non-VMIO case. 2684 */ 2685 2686int 2687allocbuf(struct buf *bp, int size) 2688{ 2689 int newbsize, mbsize; 2690 int i; 2691 2692 GIANT_REQUIRED; 2693 2694 if (BUF_REFCNT(bp) == 0) 2695 panic("allocbuf: buffer not busy"); 2696 2697 if (bp->b_kvasize < size) 2698 panic("allocbuf: buffer too small"); 2699 2700 if ((bp->b_flags & B_VMIO) == 0) { 2701 caddr_t origbuf; 2702 int origbufsize; 2703 /* 2704 * Just get anonymous memory from the kernel. Don't 2705 * mess with B_CACHE. 2706 */ 2707 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2708 if (bp->b_flags & B_MALLOC) 2709 newbsize = mbsize; 2710 else 2711 newbsize = round_page(size); 2712 2713 if (newbsize < bp->b_bufsize) { 2714 /* 2715 * malloced buffers are not shrunk 2716 */ 2717 if (bp->b_flags & B_MALLOC) { 2718 if (newbsize) { 2719 bp->b_bcount = size; 2720 } else { 2721 free(bp->b_data, M_BIOBUF); 2722 if (bp->b_bufsize) { 2723 atomic_subtract_int( 2724 &bufmallocspace, 2725 bp->b_bufsize); 2726 bufspacewakeup(); 2727 bp->b_bufsize = 0; 2728 } 2729 bp->b_saveaddr = bp->b_kvabase; 2730 bp->b_data = bp->b_saveaddr; 2731 bp->b_bcount = 0; 2732 bp->b_flags &= ~B_MALLOC; 2733 } 2734 return 1; 2735 } 2736 vm_hold_free_pages( 2737 bp, 2738 (vm_offset_t) bp->b_data + newbsize, 2739 (vm_offset_t) bp->b_data + bp->b_bufsize); 2740 } else if (newbsize > bp->b_bufsize) { 2741 /* 2742 * We only use malloced memory on the first allocation. 2743 * and revert to page-allocated memory when the buffer 2744 * grows. 2745 */ 2746 /* 2747 * There is a potential smp race here that could lead 2748 * to bufmallocspace slightly passing the max. It 2749 * is probably extremely rare and not worth worrying 2750 * over. 2751 */ 2752 if ( (bufmallocspace < maxbufmallocspace) && 2753 (bp->b_bufsize == 0) && 2754 (mbsize <= PAGE_SIZE/2)) { 2755 2756 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2757 bp->b_bufsize = mbsize; 2758 bp->b_bcount = size; 2759 bp->b_flags |= B_MALLOC; 2760 atomic_add_int(&bufmallocspace, mbsize); 2761 return 1; 2762 } 2763 origbuf = NULL; 2764 origbufsize = 0; 2765 /* 2766 * If the buffer is growing on its other-than-first allocation, 2767 * then we revert to the page-allocation scheme. 2768 */ 2769 if (bp->b_flags & B_MALLOC) { 2770 origbuf = bp->b_data; 2771 origbufsize = bp->b_bufsize; 2772 bp->b_data = bp->b_kvabase; 2773 if (bp->b_bufsize) { 2774 atomic_subtract_int(&bufmallocspace, 2775 bp->b_bufsize); 2776 bufspacewakeup(); 2777 bp->b_bufsize = 0; 2778 } 2779 bp->b_flags &= ~B_MALLOC; 2780 newbsize = round_page(newbsize); 2781 } 2782 vm_hold_load_pages( 2783 bp, 2784 (vm_offset_t) bp->b_data + bp->b_bufsize, 2785 (vm_offset_t) bp->b_data + newbsize); 2786 if (origbuf) { 2787 bcopy(origbuf, bp->b_data, origbufsize); 2788 free(origbuf, M_BIOBUF); 2789 } 2790 } 2791 } else { 2792 int desiredpages; 2793 2794 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2795 desiredpages = (size == 0) ? 0 : 2796 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2797 2798 if (bp->b_flags & B_MALLOC) 2799 panic("allocbuf: VMIO buffer can't be malloced"); 2800 /* 2801 * Set B_CACHE initially if buffer is 0 length or will become 2802 * 0-length. 2803 */ 2804 if (size == 0 || bp->b_bufsize == 0) 2805 bp->b_flags |= B_CACHE; 2806 2807 if (newbsize < bp->b_bufsize) { 2808 /* 2809 * DEV_BSIZE aligned new buffer size is less then the 2810 * DEV_BSIZE aligned existing buffer size. Figure out 2811 * if we have to remove any pages. 2812 */ 2813 if (desiredpages < bp->b_npages) { 2814 vm_page_t m; 2815 2816 VM_OBJECT_LOCK(bp->b_object); 2817 vm_page_lock_queues(); 2818 for (i = desiredpages; i < bp->b_npages; i++) { 2819 /* 2820 * the page is not freed here -- it 2821 * is the responsibility of 2822 * vnode_pager_setsize 2823 */ 2824 m = bp->b_pages[i]; 2825 KASSERT(m != bogus_page, 2826 ("allocbuf: bogus page found")); 2827 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2828 vm_page_lock_queues(); 2829 2830 bp->b_pages[i] = NULL; 2831 vm_page_unwire(m, 0); 2832 } 2833 vm_page_unlock_queues(); 2834 VM_OBJECT_UNLOCK(bp->b_object); 2835 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2836 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2837 bp->b_npages = desiredpages; 2838 } 2839 } else if (size > bp->b_bcount) { 2840 /* 2841 * We are growing the buffer, possibly in a 2842 * byte-granular fashion. 2843 */ 2844 struct vnode *vp; 2845 vm_object_t obj; 2846 vm_offset_t toff; 2847 vm_offset_t tinc; 2848 2849 /* 2850 * Step 1, bring in the VM pages from the object, 2851 * allocating them if necessary. We must clear 2852 * B_CACHE if these pages are not valid for the 2853 * range covered by the buffer. 2854 */ 2855 2856 vp = bp->b_vp; 2857 obj = bp->b_object; 2858 2859 VM_OBJECT_LOCK(obj); 2860 while (bp->b_npages < desiredpages) { 2861 vm_page_t m; 2862 vm_pindex_t pi; 2863 2864 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2865 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2866 /* 2867 * note: must allocate system pages 2868 * since blocking here could intefere 2869 * with paging I/O, no matter which 2870 * process we are. 2871 */ 2872 m = vm_page_alloc(obj, pi, 2873 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2874 VM_ALLOC_WIRED); 2875 if (m == NULL) { 2876 atomic_add_int(&vm_pageout_deficit, 2877 desiredpages - bp->b_npages); 2878 VM_OBJECT_UNLOCK(obj); 2879 VM_WAIT; 2880 VM_OBJECT_LOCK(obj); 2881 } else { 2882 bp->b_flags &= ~B_CACHE; 2883 bp->b_pages[bp->b_npages] = m; 2884 ++bp->b_npages; 2885 } 2886 continue; 2887 } 2888 2889 /* 2890 * We found a page. If we have to sleep on it, 2891 * retry because it might have gotten freed out 2892 * from under us. 2893 * 2894 * We can only test PG_BUSY here. Blocking on 2895 * m->busy might lead to a deadlock: 2896 * 2897 * vm_fault->getpages->cluster_read->allocbuf 2898 * 2899 */ 2900 vm_page_lock_queues(); 2901 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2902 continue; 2903 2904 /* 2905 * We have a good page. Should we wakeup the 2906 * page daemon? 2907 */ 2908 if ((curproc != pageproc) && 2909 ((m->queue - m->pc) == PQ_CACHE) && 2910 ((cnt.v_free_count + cnt.v_cache_count) < 2911 (cnt.v_free_min + cnt.v_cache_min))) { 2912 pagedaemon_wakeup(); 2913 } 2914 vm_page_wire(m); 2915 vm_page_unlock_queues(); 2916 bp->b_pages[bp->b_npages] = m; 2917 ++bp->b_npages; 2918 } 2919 2920 /* 2921 * Step 2. We've loaded the pages into the buffer, 2922 * we have to figure out if we can still have B_CACHE 2923 * set. Note that B_CACHE is set according to the 2924 * byte-granular range ( bcount and size ), new the 2925 * aligned range ( newbsize ). 2926 * 2927 * The VM test is against m->valid, which is DEV_BSIZE 2928 * aligned. Needless to say, the validity of the data 2929 * needs to also be DEV_BSIZE aligned. Note that this 2930 * fails with NFS if the server or some other client 2931 * extends the file's EOF. If our buffer is resized, 2932 * B_CACHE may remain set! XXX 2933 */ 2934 2935 toff = bp->b_bcount; 2936 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2937 2938 while ((bp->b_flags & B_CACHE) && toff < size) { 2939 vm_pindex_t pi; 2940 2941 if (tinc > (size - toff)) 2942 tinc = size - toff; 2943 2944 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2945 PAGE_SHIFT; 2946 2947 vfs_buf_test_cache( 2948 bp, 2949 bp->b_offset, 2950 toff, 2951 tinc, 2952 bp->b_pages[pi] 2953 ); 2954 toff += tinc; 2955 tinc = PAGE_SIZE; 2956 } 2957 VM_OBJECT_UNLOCK(obj); 2958 2959 /* 2960 * Step 3, fixup the KVM pmap. Remember that 2961 * bp->b_data is relative to bp->b_offset, but 2962 * bp->b_offset may be offset into the first page. 2963 */ 2964 2965 bp->b_data = (caddr_t) 2966 trunc_page((vm_offset_t)bp->b_data); 2967 pmap_qenter( 2968 (vm_offset_t)bp->b_data, 2969 bp->b_pages, 2970 bp->b_npages 2971 ); 2972 2973 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2974 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2975 } 2976 } 2977 if (newbsize < bp->b_bufsize) 2978 bufspacewakeup(); 2979 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2980 bp->b_bcount = size; /* requested buffer size */ 2981 return 1; 2982} 2983 2984void 2985biodone(struct bio *bp) 2986{ 2987 2988 mtx_lock(&bdonelock); 2989 bp->bio_flags |= BIO_DONE; 2990 if (bp->bio_done == NULL) 2991 wakeup(bp); 2992 mtx_unlock(&bdonelock); 2993 if (bp->bio_done != NULL) 2994 bp->bio_done(bp); 2995} 2996 2997/* 2998 * Wait for a BIO to finish. 2999 * 3000 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3001 * case is not yet clear. 3002 */ 3003int 3004biowait(struct bio *bp, const char *wchan) 3005{ 3006 3007 mtx_lock(&bdonelock); 3008 while ((bp->bio_flags & BIO_DONE) == 0) 3009 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3010 mtx_unlock(&bdonelock); 3011 if (bp->bio_error != 0) 3012 return (bp->bio_error); 3013 if (!(bp->bio_flags & BIO_ERROR)) 3014 return (0); 3015 return (EIO); 3016} 3017 3018void 3019biofinish(struct bio *bp, struct devstat *stat, int error) 3020{ 3021 3022 if (error) { 3023 bp->bio_error = error; 3024 bp->bio_flags |= BIO_ERROR; 3025 } 3026 if (stat != NULL) 3027 devstat_end_transaction_bio(stat, bp); 3028 biodone(bp); 3029} 3030 3031/* 3032 * bufwait: 3033 * 3034 * Wait for buffer I/O completion, returning error status. The buffer 3035 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3036 * error and cleared. 3037 */ 3038int 3039bufwait(struct buf *bp) 3040{ 3041 int s; 3042 3043 s = splbio(); 3044 if (bp->b_iocmd == BIO_READ) 3045 bwait(bp, PRIBIO, "biord"); 3046 else 3047 bwait(bp, PRIBIO, "biowr"); 3048 splx(s); 3049 if (bp->b_flags & B_EINTR) { 3050 bp->b_flags &= ~B_EINTR; 3051 return (EINTR); 3052 } 3053 if (bp->b_ioflags & BIO_ERROR) { 3054 return (bp->b_error ? bp->b_error : EIO); 3055 } else { 3056 return (0); 3057 } 3058} 3059 3060 /* 3061 * Call back function from struct bio back up to struct buf. 3062 */ 3063static void 3064bufdonebio(struct bio *bp) 3065{ 3066 3067 /* Device drivers may or may not hold giant, hold it here. */ 3068 mtx_lock(&Giant); 3069 bufdone(bp->bio_caller2); 3070 mtx_unlock(&Giant); 3071} 3072 3073void 3074dev_strategy(struct buf *bp) 3075{ 3076 struct cdevsw *csw; 3077 struct cdev *dev; 3078 3079 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3080 panic("b_iocmd botch"); 3081 bp->b_io.bio_done = bufdonebio; 3082 bp->b_io.bio_caller2 = bp; 3083 dev = bp->b_io.bio_dev; 3084 KASSERT(dev->si_refcount > 0, 3085 ("dev_strategy on un-referenced struct cdev *(%s)", 3086 devtoname(dev))); 3087 csw = dev_refthread(dev); 3088 if (csw == NULL) { 3089 bp->b_error = ENXIO; 3090 bp->b_ioflags = BIO_ERROR; 3091 mtx_lock(&Giant); /* XXX: too defensive ? */ 3092 bufdone(bp); 3093 mtx_unlock(&Giant); /* XXX: too defensive ? */ 3094 return; 3095 } 3096 (*csw->d_strategy)(&bp->b_io); 3097 dev_relthread(dev); 3098} 3099 3100/* 3101 * bufdone: 3102 * 3103 * Finish I/O on a buffer, optionally calling a completion function. 3104 * This is usually called from an interrupt so process blocking is 3105 * not allowed. 3106 * 3107 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3108 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3109 * assuming B_INVAL is clear. 3110 * 3111 * For the VMIO case, we set B_CACHE if the op was a read and no 3112 * read error occured, or if the op was a write. B_CACHE is never 3113 * set if the buffer is invalid or otherwise uncacheable. 3114 * 3115 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3116 * initiator to leave B_INVAL set to brelse the buffer out of existance 3117 * in the biodone routine. 3118 */ 3119void 3120bufdone(struct buf *bp) 3121{ 3122 int s; 3123 void (*biodone)(struct buf *); 3124 3125 3126 s = splbio(); 3127 3128 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3129 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3130 3131 bp->b_flags |= B_DONE; 3132 runningbufwakeup(bp); 3133 3134 if (bp->b_iocmd == BIO_WRITE && bp->b_bufobj != NULL) 3135 bufobj_wdrop(bp->b_bufobj); 3136 3137 /* call optional completion function if requested */ 3138 if (bp->b_iodone != NULL) { 3139 biodone = bp->b_iodone; 3140 bp->b_iodone = NULL; 3141 (*biodone) (bp); 3142 splx(s); 3143 return; 3144 } 3145 if (LIST_FIRST(&bp->b_dep) != NULL) 3146 buf_complete(bp); 3147 3148 if (bp->b_flags & B_VMIO) { 3149 int i; 3150 vm_ooffset_t foff; 3151 vm_page_t m; 3152 vm_object_t obj; 3153 int iosize; 3154 struct vnode *vp = bp->b_vp; 3155 3156 obj = bp->b_object; 3157 3158#if defined(VFS_BIO_DEBUG) 3159 mp_fixme("usecount and vflag accessed without locks."); 3160 if (vp->v_usecount == 0) { 3161 panic("biodone: zero vnode ref count"); 3162 } 3163 3164 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3165 panic("biodone: vnode is not setup for merged cache"); 3166 } 3167#endif 3168 3169 foff = bp->b_offset; 3170 KASSERT(bp->b_offset != NOOFFSET, 3171 ("biodone: no buffer offset")); 3172 3173 VM_OBJECT_LOCK(obj); 3174#if defined(VFS_BIO_DEBUG) 3175 if (obj->paging_in_progress < bp->b_npages) { 3176 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3177 obj->paging_in_progress, bp->b_npages); 3178 } 3179#endif 3180 3181 /* 3182 * Set B_CACHE if the op was a normal read and no error 3183 * occured. B_CACHE is set for writes in the b*write() 3184 * routines. 3185 */ 3186 iosize = bp->b_bcount - bp->b_resid; 3187 if (bp->b_iocmd == BIO_READ && 3188 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3189 !(bp->b_ioflags & BIO_ERROR)) { 3190 bp->b_flags |= B_CACHE; 3191 } 3192 vm_page_lock_queues(); 3193 for (i = 0; i < bp->b_npages; i++) { 3194 int bogusflag = 0; 3195 int resid; 3196 3197 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3198 if (resid > iosize) 3199 resid = iosize; 3200 3201 /* 3202 * cleanup bogus pages, restoring the originals 3203 */ 3204 m = bp->b_pages[i]; 3205 if (m == bogus_page) { 3206 bogusflag = 1; 3207 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3208 if (m == NULL) 3209 panic("biodone: page disappeared!"); 3210 bp->b_pages[i] = m; 3211 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3212 } 3213#if defined(VFS_BIO_DEBUG) 3214 if (OFF_TO_IDX(foff) != m->pindex) { 3215 printf( 3216"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3217 (intmax_t)foff, (uintmax_t)m->pindex); 3218 } 3219#endif 3220 3221 /* 3222 * In the write case, the valid and clean bits are 3223 * already changed correctly ( see bdwrite() ), so we 3224 * only need to do this here in the read case. 3225 */ 3226 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3227 vfs_page_set_valid(bp, foff, i, m); 3228 } 3229 3230 /* 3231 * when debugging new filesystems or buffer I/O methods, this 3232 * is the most common error that pops up. if you see this, you 3233 * have not set the page busy flag correctly!!! 3234 */ 3235 if (m->busy == 0) { 3236 printf("biodone: page busy < 0, " 3237 "pindex: %d, foff: 0x(%x,%x), " 3238 "resid: %d, index: %d\n", 3239 (int) m->pindex, (int)(foff >> 32), 3240 (int) foff & 0xffffffff, resid, i); 3241 if (!vn_isdisk(vp, NULL)) 3242 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3243 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3244 (intmax_t) bp->b_lblkno, 3245 bp->b_flags, bp->b_npages); 3246 else 3247 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3248 (intmax_t) bp->b_lblkno, 3249 bp->b_flags, bp->b_npages); 3250 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3251 (u_long)m->valid, (u_long)m->dirty, 3252 m->wire_count); 3253 panic("biodone: page busy < 0\n"); 3254 } 3255 vm_page_io_finish(m); 3256 vm_object_pip_subtract(obj, 1); 3257 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3258 iosize -= resid; 3259 } 3260 vm_page_unlock_queues(); 3261 vm_object_pip_wakeupn(obj, 0); 3262 VM_OBJECT_UNLOCK(obj); 3263 } 3264 3265 /* 3266 * For asynchronous completions, release the buffer now. The brelse 3267 * will do a wakeup there if necessary - so no need to do a wakeup 3268 * here in the async case. The sync case always needs to do a wakeup. 3269 */ 3270 3271 if (bp->b_flags & B_ASYNC) { 3272 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3273 brelse(bp); 3274 else 3275 bqrelse(bp); 3276 } else { 3277 bdone(bp); 3278 } 3279 splx(s); 3280} 3281 3282/* 3283 * This routine is called in lieu of iodone in the case of 3284 * incomplete I/O. This keeps the busy status for pages 3285 * consistant. 3286 */ 3287void 3288vfs_unbusy_pages(struct buf *bp) 3289{ 3290 int i; 3291 vm_object_t obj; 3292 vm_page_t m; 3293 3294 runningbufwakeup(bp); 3295 if (!(bp->b_flags & B_VMIO)) 3296 return; 3297 3298 obj = bp->b_object; 3299 VM_OBJECT_LOCK(obj); 3300 vm_page_lock_queues(); 3301 for (i = 0; i < bp->b_npages; i++) { 3302 m = bp->b_pages[i]; 3303 if (m == bogus_page) { 3304 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3305 if (!m) 3306 panic("vfs_unbusy_pages: page missing\n"); 3307 bp->b_pages[i] = m; 3308 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3309 bp->b_pages, bp->b_npages); 3310 } 3311 vm_object_pip_subtract(obj, 1); 3312 vm_page_io_finish(m); 3313 } 3314 vm_page_unlock_queues(); 3315 vm_object_pip_wakeupn(obj, 0); 3316 VM_OBJECT_UNLOCK(obj); 3317} 3318 3319/* 3320 * vfs_page_set_valid: 3321 * 3322 * Set the valid bits in a page based on the supplied offset. The 3323 * range is restricted to the buffer's size. 3324 * 3325 * This routine is typically called after a read completes. 3326 */ 3327static void 3328vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3329{ 3330 vm_ooffset_t soff, eoff; 3331 3332 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3333 /* 3334 * Start and end offsets in buffer. eoff - soff may not cross a 3335 * page boundry or cross the end of the buffer. The end of the 3336 * buffer, in this case, is our file EOF, not the allocation size 3337 * of the buffer. 3338 */ 3339 soff = off; 3340 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3341 if (eoff > bp->b_offset + bp->b_bcount) 3342 eoff = bp->b_offset + bp->b_bcount; 3343 3344 /* 3345 * Set valid range. This is typically the entire buffer and thus the 3346 * entire page. 3347 */ 3348 if (eoff > soff) { 3349 vm_page_set_validclean( 3350 m, 3351 (vm_offset_t) (soff & PAGE_MASK), 3352 (vm_offset_t) (eoff - soff) 3353 ); 3354 } 3355} 3356 3357/* 3358 * This routine is called before a device strategy routine. 3359 * It is used to tell the VM system that paging I/O is in 3360 * progress, and treat the pages associated with the buffer 3361 * almost as being PG_BUSY. Also the object paging_in_progress 3362 * flag is handled to make sure that the object doesn't become 3363 * inconsistant. 3364 * 3365 * Since I/O has not been initiated yet, certain buffer flags 3366 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3367 * and should be ignored. 3368 */ 3369void 3370vfs_busy_pages(struct buf *bp, int clear_modify) 3371{ 3372 int i, bogus; 3373 vm_object_t obj; 3374 vm_ooffset_t foff; 3375 vm_page_t m; 3376 3377 if (!(bp->b_flags & B_VMIO)) 3378 return; 3379 3380 obj = bp->b_object; 3381 foff = bp->b_offset; 3382 KASSERT(bp->b_offset != NOOFFSET, 3383 ("vfs_busy_pages: no buffer offset")); 3384 vfs_setdirty(bp); 3385 VM_OBJECT_LOCK(obj); 3386retry: 3387 vm_page_lock_queues(); 3388 for (i = 0; i < bp->b_npages; i++) { 3389 m = bp->b_pages[i]; 3390 3391 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3392 goto retry; 3393 } 3394 bogus = 0; 3395 for (i = 0; i < bp->b_npages; i++) { 3396 m = bp->b_pages[i]; 3397 3398 if ((bp->b_flags & B_CLUSTER) == 0) { 3399 vm_object_pip_add(obj, 1); 3400 vm_page_io_start(m); 3401 } 3402 /* 3403 * When readying a buffer for a read ( i.e 3404 * clear_modify == 0 ), it is important to do 3405 * bogus_page replacement for valid pages in 3406 * partially instantiated buffers. Partially 3407 * instantiated buffers can, in turn, occur when 3408 * reconstituting a buffer from its VM backing store 3409 * base. We only have to do this if B_CACHE is 3410 * clear ( which causes the I/O to occur in the 3411 * first place ). The replacement prevents the read 3412 * I/O from overwriting potentially dirty VM-backed 3413 * pages. XXX bogus page replacement is, uh, bogus. 3414 * It may not work properly with small-block devices. 3415 * We need to find a better way. 3416 */ 3417 pmap_remove_all(m); 3418 if (clear_modify) 3419 vfs_page_set_valid(bp, foff, i, m); 3420 else if (m->valid == VM_PAGE_BITS_ALL && 3421 (bp->b_flags & B_CACHE) == 0) { 3422 bp->b_pages[i] = bogus_page; 3423 bogus++; 3424 } 3425 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3426 } 3427 vm_page_unlock_queues(); 3428 VM_OBJECT_UNLOCK(obj); 3429 if (bogus) 3430 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3431 bp->b_pages, bp->b_npages); 3432} 3433 3434/* 3435 * Tell the VM system that the pages associated with this buffer 3436 * are clean. This is used for delayed writes where the data is 3437 * going to go to disk eventually without additional VM intevention. 3438 * 3439 * Note that while we only really need to clean through to b_bcount, we 3440 * just go ahead and clean through to b_bufsize. 3441 */ 3442static void 3443vfs_clean_pages(struct buf *bp) 3444{ 3445 int i; 3446 vm_ooffset_t foff, noff, eoff; 3447 vm_page_t m; 3448 3449 if (!(bp->b_flags & B_VMIO)) 3450 return; 3451 3452 foff = bp->b_offset; 3453 KASSERT(bp->b_offset != NOOFFSET, 3454 ("vfs_clean_pages: no buffer offset")); 3455 VM_OBJECT_LOCK(bp->b_object); 3456 vm_page_lock_queues(); 3457 for (i = 0; i < bp->b_npages; i++) { 3458 m = bp->b_pages[i]; 3459 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3460 eoff = noff; 3461 3462 if (eoff > bp->b_offset + bp->b_bufsize) 3463 eoff = bp->b_offset + bp->b_bufsize; 3464 vfs_page_set_valid(bp, foff, i, m); 3465 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3466 foff = noff; 3467 } 3468 vm_page_unlock_queues(); 3469 VM_OBJECT_UNLOCK(bp->b_object); 3470} 3471 3472/* 3473 * vfs_bio_set_validclean: 3474 * 3475 * Set the range within the buffer to valid and clean. The range is 3476 * relative to the beginning of the buffer, b_offset. Note that b_offset 3477 * itself may be offset from the beginning of the first page. 3478 * 3479 */ 3480 3481void 3482vfs_bio_set_validclean(struct buf *bp, int base, int size) 3483{ 3484 int i, n; 3485 vm_page_t m; 3486 3487 if (!(bp->b_flags & B_VMIO)) 3488 return; 3489 3490 /* 3491 * Fixup base to be relative to beginning of first page. 3492 * Set initial n to be the maximum number of bytes in the 3493 * first page that can be validated. 3494 */ 3495 3496 base += (bp->b_offset & PAGE_MASK); 3497 n = PAGE_SIZE - (base & PAGE_MASK); 3498 3499 VM_OBJECT_LOCK(bp->b_object); 3500 vm_page_lock_queues(); 3501 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3502 m = bp->b_pages[i]; 3503 3504 if (n > size) 3505 n = size; 3506 3507 vm_page_set_validclean(m, base & PAGE_MASK, n); 3508 base += n; 3509 size -= n; 3510 n = PAGE_SIZE; 3511 } 3512 vm_page_unlock_queues(); 3513 VM_OBJECT_UNLOCK(bp->b_object); 3514} 3515 3516/* 3517 * vfs_bio_clrbuf: 3518 * 3519 * clear a buffer. This routine essentially fakes an I/O, so we need 3520 * to clear BIO_ERROR and B_INVAL. 3521 * 3522 * Note that while we only theoretically need to clear through b_bcount, 3523 * we go ahead and clear through b_bufsize. 3524 */ 3525 3526void 3527vfs_bio_clrbuf(struct buf *bp) 3528{ 3529 int i, j, mask = 0; 3530 caddr_t sa, ea; 3531 3532 GIANT_REQUIRED; 3533 3534 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3535 clrbuf(bp); 3536 return; 3537 } 3538 bp->b_flags &= ~B_INVAL; 3539 bp->b_ioflags &= ~BIO_ERROR; 3540 VM_OBJECT_LOCK(bp->b_object); 3541 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3542 (bp->b_offset & PAGE_MASK) == 0) { 3543 if (bp->b_pages[0] == bogus_page) 3544 goto unlock; 3545 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3546 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3547 if ((bp->b_pages[0]->valid & mask) == mask) 3548 goto unlock; 3549 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3550 ((bp->b_pages[0]->valid & mask) == 0)) { 3551 bzero(bp->b_data, bp->b_bufsize); 3552 bp->b_pages[0]->valid |= mask; 3553 goto unlock; 3554 } 3555 } 3556 ea = sa = bp->b_data; 3557 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3558 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3559 ea = (caddr_t)(vm_offset_t)ulmin( 3560 (u_long)(vm_offset_t)ea, 3561 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3562 if (bp->b_pages[i] == bogus_page) 3563 continue; 3564 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3565 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3566 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3567 if ((bp->b_pages[i]->valid & mask) == mask) 3568 continue; 3569 if ((bp->b_pages[i]->valid & mask) == 0) { 3570 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3571 bzero(sa, ea - sa); 3572 } else { 3573 for (; sa < ea; sa += DEV_BSIZE, j++) { 3574 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3575 (bp->b_pages[i]->valid & (1<<j)) == 0) 3576 bzero(sa, DEV_BSIZE); 3577 } 3578 } 3579 bp->b_pages[i]->valid |= mask; 3580 } 3581unlock: 3582 VM_OBJECT_UNLOCK(bp->b_object); 3583 bp->b_resid = 0; 3584} 3585 3586/* 3587 * vm_hold_load_pages and vm_hold_free_pages get pages into 3588 * a buffers address space. The pages are anonymous and are 3589 * not associated with a file object. 3590 */ 3591static void 3592vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3593{ 3594 vm_offset_t pg; 3595 vm_page_t p; 3596 int index; 3597 3598 to = round_page(to); 3599 from = round_page(from); 3600 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3601 3602 VM_OBJECT_LOCK(kernel_object); 3603 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3604tryagain: 3605 /* 3606 * note: must allocate system pages since blocking here 3607 * could intefere with paging I/O, no matter which 3608 * process we are. 3609 */ 3610 p = vm_page_alloc(kernel_object, 3611 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3612 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3613 if (!p) { 3614 atomic_add_int(&vm_pageout_deficit, 3615 (to - pg) >> PAGE_SHIFT); 3616 VM_OBJECT_UNLOCK(kernel_object); 3617 VM_WAIT; 3618 VM_OBJECT_LOCK(kernel_object); 3619 goto tryagain; 3620 } 3621 p->valid = VM_PAGE_BITS_ALL; 3622 pmap_qenter(pg, &p, 1); 3623 bp->b_pages[index] = p; 3624 } 3625 VM_OBJECT_UNLOCK(kernel_object); 3626 bp->b_npages = index; 3627} 3628 3629/* Return pages associated with this buf to the vm system */ 3630static void 3631vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3632{ 3633 vm_offset_t pg; 3634 vm_page_t p; 3635 int index, newnpages; 3636 3637 from = round_page(from); 3638 to = round_page(to); 3639 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3640 3641 VM_OBJECT_LOCK(kernel_object); 3642 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3643 p = bp->b_pages[index]; 3644 if (p && (index < bp->b_npages)) { 3645 if (p->busy) { 3646 printf( 3647 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3648 (intmax_t)bp->b_blkno, 3649 (intmax_t)bp->b_lblkno); 3650 } 3651 bp->b_pages[index] = NULL; 3652 pmap_qremove(pg, 1); 3653 vm_page_lock_queues(); 3654 vm_page_busy(p); 3655 vm_page_unwire(p, 0); 3656 vm_page_free(p); 3657 vm_page_unlock_queues(); 3658 } 3659 } 3660 VM_OBJECT_UNLOCK(kernel_object); 3661 bp->b_npages = newnpages; 3662} 3663 3664/* 3665 * Map an IO request into kernel virtual address space. 3666 * 3667 * All requests are (re)mapped into kernel VA space. 3668 * Notice that we use b_bufsize for the size of the buffer 3669 * to be mapped. b_bcount might be modified by the driver. 3670 * 3671 * Note that even if the caller determines that the address space should 3672 * be valid, a race or a smaller-file mapped into a larger space may 3673 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3674 * check the return value. 3675 */ 3676int 3677vmapbuf(struct buf *bp) 3678{ 3679 caddr_t addr, kva; 3680 vm_prot_t prot; 3681 int pidx, i; 3682 struct vm_page *m; 3683 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3684 3685 if (bp->b_bufsize < 0) 3686 return (-1); 3687 prot = VM_PROT_READ; 3688 if (bp->b_iocmd == BIO_READ) 3689 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3690 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3691 addr < bp->b_data + bp->b_bufsize; 3692 addr += PAGE_SIZE, pidx++) { 3693 /* 3694 * Do the vm_fault if needed; do the copy-on-write thing 3695 * when reading stuff off device into memory. 3696 * 3697 * NOTE! Must use pmap_extract() because addr may be in 3698 * the userland address space, and kextract is only guarenteed 3699 * to work for the kernland address space (see: sparc64 port). 3700 */ 3701retry: 3702 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3703 prot) < 0) { 3704 vm_page_lock_queues(); 3705 for (i = 0; i < pidx; ++i) { 3706 vm_page_unhold(bp->b_pages[i]); 3707 bp->b_pages[i] = NULL; 3708 } 3709 vm_page_unlock_queues(); 3710 return(-1); 3711 } 3712 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3713 if (m == NULL) 3714 goto retry; 3715 bp->b_pages[pidx] = m; 3716 } 3717 if (pidx > btoc(MAXPHYS)) 3718 panic("vmapbuf: mapped more than MAXPHYS"); 3719 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3720 3721 kva = bp->b_saveaddr; 3722 bp->b_npages = pidx; 3723 bp->b_saveaddr = bp->b_data; 3724 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3725 return(0); 3726} 3727 3728/* 3729 * Free the io map PTEs associated with this IO operation. 3730 * We also invalidate the TLB entries and restore the original b_addr. 3731 */ 3732void 3733vunmapbuf(struct buf *bp) 3734{ 3735 int pidx; 3736 int npages; 3737 3738 npages = bp->b_npages; 3739 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3740 vm_page_lock_queues(); 3741 for (pidx = 0; pidx < npages; pidx++) 3742 vm_page_unhold(bp->b_pages[pidx]); 3743 vm_page_unlock_queues(); 3744 3745 bp->b_data = bp->b_saveaddr; 3746} 3747 3748void 3749bdone(struct buf *bp) 3750{ 3751 3752 mtx_lock(&bdonelock); 3753 bp->b_flags |= B_DONE; 3754 wakeup(bp); 3755 mtx_unlock(&bdonelock); 3756} 3757 3758void 3759bwait(struct buf *bp, u_char pri, const char *wchan) 3760{ 3761 3762 mtx_lock(&bdonelock); 3763 while ((bp->b_flags & B_DONE) == 0) 3764 msleep(bp, &bdonelock, pri, wchan, 0); 3765 mtx_unlock(&bdonelock); 3766} 3767 3768void 3769bufstrategy(struct bufobj *bo, struct buf *bp) 3770{ 3771 int i = 0; 3772 struct vnode *vp; 3773 3774 vp = bp->b_vp; 3775#if 0 3776 KASSERT(vp == bo->bo_vnode, ("Inconsistent vnode bufstrategy")); 3777 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3778 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3779#endif 3780 if (vp->v_type == VCHR) { 3781 if (!buf_prewrite(bp->b_vp, bp)) 3782 i = VOP_SPECSTRATEGY(vp, bp); 3783 } else { 3784 i = VOP_STRATEGY(vp, bp); 3785 } 3786 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3787} 3788 3789void 3790bufobj_wref(struct bufobj *bo) 3791{ 3792 3793 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3794 BO_LOCK(bo); 3795 bo->bo_numoutput++; 3796 BO_UNLOCK(bo); 3797} 3798 3799void 3800bufobj_wdrop(struct bufobj *bo) 3801{ 3802 3803 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3804 BO_LOCK(bo); 3805 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3806 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3807 bo->bo_flag &= ~BO_WWAIT; 3808 wakeup(&bo->bo_numoutput); 3809 } 3810 BO_UNLOCK(bo); 3811} 3812 3813int 3814bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3815{ 3816 int error; 3817 3818 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3819 ASSERT_BO_LOCKED(bo); 3820 error = 0; 3821 while (bo->bo_numoutput) { 3822 bo->bo_flag |= BO_WWAIT; 3823 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3824 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3825 if (error) 3826 break; 3827 } 3828 return (error); 3829} 3830 3831#include "opt_ddb.h" 3832#ifdef DDB 3833#include <ddb/ddb.h> 3834 3835/* DDB command to show buffer data */ 3836DB_SHOW_COMMAND(buffer, db_show_buffer) 3837{ 3838 /* get args */ 3839 struct buf *bp = (struct buf *)addr; 3840 3841 if (!have_addr) { 3842 db_printf("usage: show buffer <addr>\n"); 3843 return; 3844 } 3845 3846 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3847 db_printf( 3848 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3849 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd\n", 3850 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3851 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3852 (intmax_t)bp->b_blkno); 3853 if (bp->b_npages) { 3854 int i; 3855 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3856 for (i = 0; i < bp->b_npages; i++) { 3857 vm_page_t m; 3858 m = bp->b_pages[i]; 3859 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3860 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3861 if ((i + 1) < bp->b_npages) 3862 db_printf(","); 3863 } 3864 db_printf("\n"); 3865 } 3866} 3867#endif /* DDB */ 3868