vfs_bio.c revision 136751
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 136751 2004-10-21 15:53:54Z phk $"); 30 31#include <sys/param.h> 32#include <sys/systm.h> 33#include <sys/bio.h> 34#include <sys/conf.h> 35#include <sys/buf.h> 36#include <sys/devicestat.h> 37#include <sys/eventhandler.h> 38#include <sys/lock.h> 39#include <sys/malloc.h> 40#include <sys/mount.h> 41#include <sys/mutex.h> 42#include <sys/kernel.h> 43#include <sys/kthread.h> 44#include <sys/proc.h> 45#include <sys/resourcevar.h> 46#include <sys/sysctl.h> 47#include <sys/vmmeter.h> 48#include <sys/vnode.h> 49#include <vm/vm.h> 50#include <vm/vm_param.h> 51#include <vm/vm_kern.h> 52#include <vm/vm_pageout.h> 53#include <vm/vm_page.h> 54#include <vm/vm_object.h> 55#include <vm/vm_extern.h> 56#include <vm/vm_map.h> 57#include "opt_directio.h" 58#include "opt_swap.h" 59 60static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 61 62struct bio_ops bioops; /* I/O operation notification */ 63 64static int ibwrite(struct buf *); 65 66struct buf_ops buf_ops_bio = { 67 "buf_ops_bio", 68 ibwrite 69}; 70 71/* 72 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 73 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 74 */ 75struct buf *buf; /* buffer header pool */ 76 77static struct proc *bufdaemonproc; 78 79static void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 80 vm_offset_t to); 81static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 82 vm_offset_t to); 83static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 84 int pageno, vm_page_t m); 85static void vfs_clean_pages(struct buf *bp); 86static void vfs_setdirty(struct buf *bp); 87static void vfs_vmio_release(struct buf *bp); 88static void vfs_backgroundwritedone(struct buf *bp); 89static int vfs_bio_clcheck(struct vnode *vp, int size, 90 daddr_t lblkno, daddr_t blkno); 91static int flushbufqueues(int flushdeps); 92static void buf_daemon(void); 93void bremfreel(struct buf *bp); 94 95int vmiodirenable = TRUE; 96SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 97 "Use the VM system for directory writes"); 98int runningbufspace; 99SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 100 "Amount of presently outstanding async buffer io"); 101static int bufspace; 102SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 103 "KVA memory used for bufs"); 104static int maxbufspace; 105SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 106 "Maximum allowed value of bufspace (including buf_daemon)"); 107static int bufmallocspace; 108SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 109 "Amount of malloced memory for buffers"); 110static int maxbufmallocspace; 111SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 112 "Maximum amount of malloced memory for buffers"); 113static int lobufspace; 114SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 115 "Minimum amount of buffers we want to have"); 116static int hibufspace; 117SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 118 "Maximum allowed value of bufspace (excluding buf_daemon)"); 119static int bufreusecnt; 120SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 121 "Number of times we have reused a buffer"); 122static int buffreekvacnt; 123SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 124 "Number of times we have freed the KVA space from some buffer"); 125static int bufdefragcnt; 126SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 127 "Number of times we have had to repeat buffer allocation to defragment"); 128static int lorunningspace; 129SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 130 "Minimum preferred space used for in-progress I/O"); 131static int hirunningspace; 132SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 133 "Maximum amount of space to use for in-progress I/O"); 134static int dirtybufferflushes; 135SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 136 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 137static int altbufferflushes; 138SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 139 0, "Number of fsync flushes to limit dirty buffers"); 140static int recursiveflushes; 141SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 142 0, "Number of flushes skipped due to being recursive"); 143static int numdirtybuffers; 144SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 145 "Number of buffers that are dirty (has unwritten changes) at the moment"); 146static int lodirtybuffers; 147SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 148 "How many buffers we want to have free before bufdaemon can sleep"); 149static int hidirtybuffers; 150SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 151 "When the number of dirty buffers is considered severe"); 152static int dirtybufthresh; 153SYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 154 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 155static int numfreebuffers; 156SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 157 "Number of free buffers"); 158static int lofreebuffers; 159SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 160 "XXX Unused"); 161static int hifreebuffers; 162SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 163 "XXX Complicatedly unused"); 164static int getnewbufcalls; 165SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 166 "Number of calls to getnewbuf"); 167static int getnewbufrestarts; 168SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 169 "Number of times getnewbuf has had to restart a buffer aquisition"); 170static int dobkgrdwrite = 1; 171SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 172 "Do background writes (honoring the BV_BKGRDWRITE flag)?"); 173 174/* 175 * Wakeup point for bufdaemon, as well as indicator of whether it is already 176 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 177 * is idling. 178 */ 179static int bd_request; 180 181/* 182 * This lock synchronizes access to bd_request. 183 */ 184static struct mtx bdlock; 185 186/* 187 * bogus page -- for I/O to/from partially complete buffers 188 * this is a temporary solution to the problem, but it is not 189 * really that bad. it would be better to split the buffer 190 * for input in the case of buffers partially already in memory, 191 * but the code is intricate enough already. 192 */ 193vm_page_t bogus_page; 194 195/* 196 * Synchronization (sleep/wakeup) variable for active buffer space requests. 197 * Set when wait starts, cleared prior to wakeup(). 198 * Used in runningbufwakeup() and waitrunningbufspace(). 199 */ 200static int runningbufreq; 201 202/* 203 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 204 * waitrunningbufspace(). 205 */ 206static struct mtx rbreqlock; 207 208/* 209 * Synchronization (sleep/wakeup) variable for buffer requests. 210 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 211 * by and/or. 212 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 213 * getnewbuf(), and getblk(). 214 */ 215static int needsbuffer; 216 217/* 218 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 219 */ 220static struct mtx nblock; 221 222/* 223 * Lock that protects against bwait()/bdone()/B_DONE races. 224 */ 225 226static struct mtx bdonelock; 227 228/* 229 * Definitions for the buffer free lists. 230 */ 231#define BUFFER_QUEUES 5 /* number of free buffer queues */ 232 233#define QUEUE_NONE 0 /* on no queue */ 234#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 235#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 236#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 237#define QUEUE_EMPTY 4 /* empty buffer headers */ 238 239/* Queues for free buffers with various properties */ 240static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 241 242/* Lock for the bufqueues */ 243static struct mtx bqlock; 244 245/* 246 * Single global constant for BUF_WMESG, to avoid getting multiple references. 247 * buf_wmesg is referred from macros. 248 */ 249const char *buf_wmesg = BUF_WMESG; 250 251#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 252#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 253#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 254#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 255 256#ifdef DIRECTIO 257extern void ffs_rawread_setup(void); 258#endif /* DIRECTIO */ 259/* 260 * numdirtywakeup: 261 * 262 * If someone is blocked due to there being too many dirty buffers, 263 * and numdirtybuffers is now reasonable, wake them up. 264 */ 265 266static __inline void 267numdirtywakeup(int level) 268{ 269 270 if (numdirtybuffers <= level) { 271 mtx_lock(&nblock); 272 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 273 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 274 wakeup(&needsbuffer); 275 } 276 mtx_unlock(&nblock); 277 } 278} 279 280/* 281 * bufspacewakeup: 282 * 283 * Called when buffer space is potentially available for recovery. 284 * getnewbuf() will block on this flag when it is unable to free 285 * sufficient buffer space. Buffer space becomes recoverable when 286 * bp's get placed back in the queues. 287 */ 288 289static __inline void 290bufspacewakeup(void) 291{ 292 293 /* 294 * If someone is waiting for BUF space, wake them up. Even 295 * though we haven't freed the kva space yet, the waiting 296 * process will be able to now. 297 */ 298 mtx_lock(&nblock); 299 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 300 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 301 wakeup(&needsbuffer); 302 } 303 mtx_unlock(&nblock); 304} 305 306/* 307 * runningbufwakeup() - in-progress I/O accounting. 308 * 309 */ 310static __inline void 311runningbufwakeup(struct buf *bp) 312{ 313 314 if (bp->b_runningbufspace) { 315 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 316 bp->b_runningbufspace = 0; 317 mtx_lock(&rbreqlock); 318 if (runningbufreq && runningbufspace <= lorunningspace) { 319 runningbufreq = 0; 320 wakeup(&runningbufreq); 321 } 322 mtx_unlock(&rbreqlock); 323 } 324} 325 326/* 327 * bufcountwakeup: 328 * 329 * Called when a buffer has been added to one of the free queues to 330 * account for the buffer and to wakeup anyone waiting for free buffers. 331 * This typically occurs when large amounts of metadata are being handled 332 * by the buffer cache ( else buffer space runs out first, usually ). 333 */ 334 335static __inline void 336bufcountwakeup(void) 337{ 338 339 atomic_add_int(&numfreebuffers, 1); 340 mtx_lock(&nblock); 341 if (needsbuffer) { 342 needsbuffer &= ~VFS_BIO_NEED_ANY; 343 if (numfreebuffers >= hifreebuffers) 344 needsbuffer &= ~VFS_BIO_NEED_FREE; 345 wakeup(&needsbuffer); 346 } 347 mtx_unlock(&nblock); 348} 349 350/* 351 * waitrunningbufspace() 352 * 353 * runningbufspace is a measure of the amount of I/O currently 354 * running. This routine is used in async-write situations to 355 * prevent creating huge backups of pending writes to a device. 356 * Only asynchronous writes are governed by this function. 357 * 358 * Reads will adjust runningbufspace, but will not block based on it. 359 * The read load has a side effect of reducing the allowed write load. 360 * 361 * This does NOT turn an async write into a sync write. It waits 362 * for earlier writes to complete and generally returns before the 363 * caller's write has reached the device. 364 */ 365static __inline void 366waitrunningbufspace(void) 367{ 368 369 mtx_lock(&rbreqlock); 370 while (runningbufspace > hirunningspace) { 371 ++runningbufreq; 372 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 373 } 374 mtx_unlock(&rbreqlock); 375} 376 377 378/* 379 * vfs_buf_test_cache: 380 * 381 * Called when a buffer is extended. This function clears the B_CACHE 382 * bit if the newly extended portion of the buffer does not contain 383 * valid data. 384 */ 385static __inline 386void 387vfs_buf_test_cache(struct buf *bp, 388 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 389 vm_page_t m) 390{ 391 392 GIANT_REQUIRED; 393 394 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 395 if (bp->b_flags & B_CACHE) { 396 int base = (foff + off) & PAGE_MASK; 397 if (vm_page_is_valid(m, base, size) == 0) 398 bp->b_flags &= ~B_CACHE; 399 } 400} 401 402/* Wake up the buffer deamon if necessary */ 403static __inline 404void 405bd_wakeup(int dirtybuflevel) 406{ 407 408 mtx_lock(&bdlock); 409 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 410 bd_request = 1; 411 wakeup(&bd_request); 412 } 413 mtx_unlock(&bdlock); 414} 415 416/* 417 * bd_speedup - speedup the buffer cache flushing code 418 */ 419 420static __inline 421void 422bd_speedup(void) 423{ 424 425 bd_wakeup(1); 426} 427 428/* 429 * Calculating buffer cache scaling values and reserve space for buffer 430 * headers. This is called during low level kernel initialization and 431 * may be called more then once. We CANNOT write to the memory area 432 * being reserved at this time. 433 */ 434caddr_t 435kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 436{ 437 438 /* 439 * physmem_est is in pages. Convert it to kilobytes (assumes 440 * PAGE_SIZE is >= 1K) 441 */ 442 physmem_est = physmem_est * (PAGE_SIZE / 1024); 443 444 /* 445 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 446 * For the first 64MB of ram nominally allocate sufficient buffers to 447 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 448 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 449 * the buffer cache we limit the eventual kva reservation to 450 * maxbcache bytes. 451 * 452 * factor represents the 1/4 x ram conversion. 453 */ 454 if (nbuf == 0) { 455 int factor = 4 * BKVASIZE / 1024; 456 457 nbuf = 50; 458 if (physmem_est > 4096) 459 nbuf += min((physmem_est - 4096) / factor, 460 65536 / factor); 461 if (physmem_est > 65536) 462 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 463 464 if (maxbcache && nbuf > maxbcache / BKVASIZE) 465 nbuf = maxbcache / BKVASIZE; 466 } 467 468#if 0 469 /* 470 * Do not allow the buffer_map to be more then 1/2 the size of the 471 * kernel_map. 472 */ 473 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 474 (BKVASIZE * 2)) { 475 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 476 (BKVASIZE * 2); 477 printf("Warning: nbufs capped at %d\n", nbuf); 478 } 479#endif 480 481 /* 482 * swbufs are used as temporary holders for I/O, such as paging I/O. 483 * We have no less then 16 and no more then 256. 484 */ 485 nswbuf = max(min(nbuf/4, 256), 16); 486#ifdef NSWBUF_MIN 487 if (nswbuf < NSWBUF_MIN) 488 nswbuf = NSWBUF_MIN; 489#endif 490#ifdef DIRECTIO 491 ffs_rawread_setup(); 492#endif 493 494 /* 495 * Reserve space for the buffer cache buffers 496 */ 497 swbuf = (void *)v; 498 v = (caddr_t)(swbuf + nswbuf); 499 buf = (void *)v; 500 v = (caddr_t)(buf + nbuf); 501 502 return(v); 503} 504 505/* Initialize the buffer subsystem. Called before use of any buffers. */ 506void 507bufinit(void) 508{ 509 struct buf *bp; 510 int i; 511 512 GIANT_REQUIRED; 513 514 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 515 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 516 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 517 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 518 mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 519 520 /* next, make a null set of free lists */ 521 for (i = 0; i < BUFFER_QUEUES; i++) 522 TAILQ_INIT(&bufqueues[i]); 523 524 /* finally, initialize each buffer header and stick on empty q */ 525 for (i = 0; i < nbuf; i++) { 526 bp = &buf[i]; 527 bzero(bp, sizeof *bp); 528 bp->b_flags = B_INVAL; /* we're just an empty header */ 529 bp->b_dev = NULL; 530 bp->b_rcred = NOCRED; 531 bp->b_wcred = NOCRED; 532 bp->b_qindex = QUEUE_EMPTY; 533 bp->b_vflags = 0; 534 bp->b_xflags = 0; 535 LIST_INIT(&bp->b_dep); 536 BUF_LOCKINIT(bp); 537 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 538 } 539 540 /* 541 * maxbufspace is the absolute maximum amount of buffer space we are 542 * allowed to reserve in KVM and in real terms. The absolute maximum 543 * is nominally used by buf_daemon. hibufspace is the nominal maximum 544 * used by most other processes. The differential is required to 545 * ensure that buf_daemon is able to run when other processes might 546 * be blocked waiting for buffer space. 547 * 548 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 549 * this may result in KVM fragmentation which is not handled optimally 550 * by the system. 551 */ 552 maxbufspace = nbuf * BKVASIZE; 553 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 554 lobufspace = hibufspace - MAXBSIZE; 555 556 lorunningspace = 512 * 1024; 557 hirunningspace = 1024 * 1024; 558 559/* 560 * Limit the amount of malloc memory since it is wired permanently into 561 * the kernel space. Even though this is accounted for in the buffer 562 * allocation, we don't want the malloced region to grow uncontrolled. 563 * The malloc scheme improves memory utilization significantly on average 564 * (small) directories. 565 */ 566 maxbufmallocspace = hibufspace / 20; 567 568/* 569 * Reduce the chance of a deadlock occuring by limiting the number 570 * of delayed-write dirty buffers we allow to stack up. 571 */ 572 hidirtybuffers = nbuf / 4 + 20; 573 dirtybufthresh = hidirtybuffers * 9 / 10; 574 numdirtybuffers = 0; 575/* 576 * To support extreme low-memory systems, make sure hidirtybuffers cannot 577 * eat up all available buffer space. This occurs when our minimum cannot 578 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 579 * BKVASIZE'd (8K) buffers. 580 */ 581 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 582 hidirtybuffers >>= 1; 583 } 584 lodirtybuffers = hidirtybuffers / 2; 585 586/* 587 * Try to keep the number of free buffers in the specified range, 588 * and give special processes (e.g. like buf_daemon) access to an 589 * emergency reserve. 590 */ 591 lofreebuffers = nbuf / 18 + 5; 592 hifreebuffers = 2 * lofreebuffers; 593 numfreebuffers = nbuf; 594 595/* 596 * Maximum number of async ops initiated per buf_daemon loop. This is 597 * somewhat of a hack at the moment, we really need to limit ourselves 598 * based on the number of bytes of I/O in-transit that were initiated 599 * from buf_daemon. 600 */ 601 602 bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 603 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 604} 605 606/* 607 * bfreekva() - free the kva allocation for a buffer. 608 * 609 * Must be called at splbio() or higher as this is the only locking for 610 * buffer_map. 611 * 612 * Since this call frees up buffer space, we call bufspacewakeup(). 613 */ 614static void 615bfreekva(struct buf *bp) 616{ 617 618 GIANT_REQUIRED; 619 620 if (bp->b_kvasize) { 621 atomic_add_int(&buffreekvacnt, 1); 622 atomic_subtract_int(&bufspace, bp->b_kvasize); 623 vm_map_delete(buffer_map, 624 (vm_offset_t) bp->b_kvabase, 625 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 626 ); 627 bp->b_kvasize = 0; 628 bufspacewakeup(); 629 } 630} 631 632/* 633 * bremfree: 634 * 635 * Remove the buffer from the appropriate free list. 636 */ 637void 638bremfree(struct buf *bp) 639{ 640 641 mtx_lock(&bqlock); 642 bremfreel(bp); 643 mtx_unlock(&bqlock); 644} 645 646void 647bremfreel(struct buf *bp) 648{ 649 int s = splbio(); 650 int old_qindex = bp->b_qindex; 651 652 GIANT_REQUIRED; 653 654 if (bp->b_qindex != QUEUE_NONE) { 655 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 656 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 657 bp->b_qindex = QUEUE_NONE; 658 } else { 659 if (BUF_REFCNT(bp) <= 1) 660 panic("bremfree: removing a buffer not on a queue"); 661 } 662 663 /* 664 * Fixup numfreebuffers count. If the buffer is invalid or not 665 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 666 * the buffer was free and we must decrement numfreebuffers. 667 */ 668 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 669 switch(old_qindex) { 670 case QUEUE_DIRTY: 671 case QUEUE_CLEAN: 672 case QUEUE_EMPTY: 673 case QUEUE_EMPTYKVA: 674 atomic_subtract_int(&numfreebuffers, 1); 675 break; 676 default: 677 break; 678 } 679 } 680 splx(s); 681} 682 683 684/* 685 * Get a buffer with the specified data. Look in the cache first. We 686 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 687 * is set, the buffer is valid and we do not have to do anything ( see 688 * getblk() ). This is really just a special case of breadn(). 689 */ 690int 691bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 692 struct buf **bpp) 693{ 694 695 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 696} 697 698/* 699 * Operates like bread, but also starts asynchronous I/O on 700 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 701 * to initiating I/O . If B_CACHE is set, the buffer is valid 702 * and we do not have to do anything. 703 */ 704int 705breadn(struct vnode * vp, daddr_t blkno, int size, 706 daddr_t * rablkno, int *rabsize, 707 int cnt, struct ucred * cred, struct buf **bpp) 708{ 709 struct buf *bp, *rabp; 710 int i; 711 int rv = 0, readwait = 0; 712 713 *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 714 715 /* if not found in cache, do some I/O */ 716 if ((bp->b_flags & B_CACHE) == 0) { 717 if (curthread != PCPU_GET(idlethread)) 718 curthread->td_proc->p_stats->p_ru.ru_inblock++; 719 bp->b_iocmd = BIO_READ; 720 bp->b_flags &= ~B_INVAL; 721 bp->b_ioflags &= ~BIO_ERROR; 722 if (bp->b_rcred == NOCRED && cred != NOCRED) 723 bp->b_rcred = crhold(cred); 724 vfs_busy_pages(bp, 0); 725 bp->b_iooffset = dbtob(bp->b_blkno); 726 if (vp->v_type == VCHR) 727 VOP_SPECSTRATEGY(vp, bp); 728 else 729 VOP_STRATEGY(vp, bp); 730 ++readwait; 731 } 732 733 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 734 if (inmem(vp, *rablkno)) 735 continue; 736 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 737 738 if ((rabp->b_flags & B_CACHE) == 0) { 739 if (curthread != PCPU_GET(idlethread)) 740 curthread->td_proc->p_stats->p_ru.ru_inblock++; 741 rabp->b_flags |= B_ASYNC; 742 rabp->b_flags &= ~B_INVAL; 743 rabp->b_ioflags &= ~BIO_ERROR; 744 rabp->b_iocmd = BIO_READ; 745 if (rabp->b_rcred == NOCRED && cred != NOCRED) 746 rabp->b_rcred = crhold(cred); 747 vfs_busy_pages(rabp, 0); 748 BUF_KERNPROC(rabp); 749 rabp->b_iooffset = dbtob(rabp->b_blkno); 750 if (vp->v_type == VCHR) 751 VOP_SPECSTRATEGY(vp, rabp); 752 else 753 VOP_STRATEGY(vp, rabp); 754 } else { 755 brelse(rabp); 756 } 757 } 758 759 if (readwait) { 760 rv = bufwait(bp); 761 } 762 return (rv); 763} 764 765/* 766 * Write, release buffer on completion. (Done by iodone 767 * if async). Do not bother writing anything if the buffer 768 * is invalid. 769 * 770 * Note that we set B_CACHE here, indicating that buffer is 771 * fully valid and thus cacheable. This is true even of NFS 772 * now so we set it generally. This could be set either here 773 * or in biodone() since the I/O is synchronous. We put it 774 * here. 775 */ 776int 777bwrite(struct buf *bp) 778{ 779 780 KASSERT(bp->b_op != NULL && bp->b_op->bop_write != NULL, 781 ("Martian buffer %p in bwrite: nobody to write it.", bp)); 782 return (bp->b_op->bop_write(bp)); 783} 784 785static int 786ibwrite(struct buf *bp) 787{ 788 int oldflags, s; 789 struct buf *newbp; 790 791 if (bp->b_flags & B_INVAL) { 792 brelse(bp); 793 return (0); 794 } 795 796 oldflags = bp->b_flags; 797 798 if (BUF_REFCNT(bp) == 0) 799 panic("ibwrite: buffer is not busy???"); 800 s = splbio(); 801 /* 802 * If a background write is already in progress, delay 803 * writing this block if it is asynchronous. Otherwise 804 * wait for the background write to complete. 805 */ 806 VI_LOCK(bp->b_vp); 807 if (bp->b_vflags & BV_BKGRDINPROG) { 808 if (bp->b_flags & B_ASYNC) { 809 VI_UNLOCK(bp->b_vp); 810 splx(s); 811 bdwrite(bp); 812 return (0); 813 } 814 bp->b_vflags |= BV_BKGRDWAIT; 815 msleep(&bp->b_xflags, VI_MTX(bp->b_vp), PRIBIO, "bwrbg", 0); 816 if (bp->b_vflags & BV_BKGRDINPROG) 817 panic("ibwrite: still writing"); 818 } 819 VI_UNLOCK(bp->b_vp); 820 821 /* Mark the buffer clean */ 822 bundirty(bp); 823 824 /* 825 * If this buffer is marked for background writing and we 826 * do not have to wait for it, make a copy and write the 827 * copy so as to leave this buffer ready for further use. 828 * 829 * This optimization eats a lot of memory. If we have a page 830 * or buffer shortfall we can't do it. 831 */ 832 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 833 (bp->b_flags & B_ASYNC) && 834 !vm_page_count_severe() && 835 !buf_dirty_count_severe()) { 836 if (bp->b_iodone != NULL) { 837 printf("bp->b_iodone = %p\n", bp->b_iodone); 838 panic("ibwrite: need chained iodone"); 839 } 840 841 /* get a new block */ 842 newbp = geteblk(bp->b_bufsize); 843 844 /* 845 * set it to be identical to the old block. We have to 846 * set b_lblkno and BKGRDMARKER before calling bgetvp() 847 * to avoid confusing the splay tree and gbincore(). 848 */ 849 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 850 newbp->b_lblkno = bp->b_lblkno; 851 newbp->b_xflags |= BX_BKGRDMARKER; 852 VI_LOCK(bp->b_vp); 853 bp->b_vflags |= BV_BKGRDINPROG; 854 bgetvp(bp->b_vp, newbp); 855 VI_UNLOCK(bp->b_vp); 856 newbp->b_blkno = bp->b_blkno; 857 newbp->b_offset = bp->b_offset; 858 newbp->b_iodone = vfs_backgroundwritedone; 859 newbp->b_flags |= B_ASYNC; 860 newbp->b_flags &= ~B_INVAL; 861 862 /* move over the dependencies */ 863 if (LIST_FIRST(&bp->b_dep) != NULL) 864 buf_movedeps(bp, newbp); 865 866 /* 867 * Initiate write on the copy, release the original to 868 * the B_LOCKED queue so that it cannot go away until 869 * the background write completes. If not locked it could go 870 * away and then be reconstituted while it was being written. 871 * If the reconstituted buffer were written, we could end up 872 * with two background copies being written at the same time. 873 */ 874 bqrelse(bp); 875 bp = newbp; 876 } 877 878 bp->b_flags &= ~B_DONE; 879 bp->b_ioflags &= ~BIO_ERROR; 880 bp->b_flags |= B_CACHE; 881 bp->b_iocmd = BIO_WRITE; 882 883 bufobj_wref(&bp->b_vp->v_bufobj); 884 vfs_busy_pages(bp, 1); 885 886 /* 887 * Normal bwrites pipeline writes 888 */ 889 bp->b_runningbufspace = bp->b_bufsize; 890 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 891 892 if (curthread != PCPU_GET(idlethread)) 893 curthread->td_proc->p_stats->p_ru.ru_oublock++; 894 splx(s); 895 if (oldflags & B_ASYNC) 896 BUF_KERNPROC(bp); 897 bp->b_iooffset = dbtob(bp->b_blkno); 898 if (bp->b_vp->v_type == VCHR) { 899 if (!buf_prewrite(bp->b_vp, bp)) 900 VOP_SPECSTRATEGY(bp->b_vp, bp); 901 } else { 902 VOP_STRATEGY(bp->b_vp, bp); 903 } 904 905 if ((oldflags & B_ASYNC) == 0) { 906 int rtval = bufwait(bp); 907 brelse(bp); 908 return (rtval); 909 } else { 910 /* 911 * don't allow the async write to saturate the I/O 912 * system. We will not deadlock here because 913 * we are blocking waiting for I/O that is already in-progress 914 * to complete. We do not block here if it is the update 915 * or syncer daemon trying to clean up as that can lead 916 * to deadlock. 917 */ 918 if (curthread->td_proc != bufdaemonproc && 919 curthread->td_proc != updateproc) 920 waitrunningbufspace(); 921 } 922 923 return (0); 924} 925 926/* 927 * Complete a background write started from bwrite. 928 */ 929static void 930vfs_backgroundwritedone(struct buf *bp) 931{ 932 struct buf *origbp; 933 934 /* 935 * Find the original buffer that we are writing. 936 */ 937 VI_LOCK(bp->b_vp); 938 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 939 panic("backgroundwritedone: lost buffer"); 940 941 /* 942 * Clear the BV_BKGRDINPROG flag in the original buffer 943 * and awaken it if it is waiting for the write to complete. 944 * If BV_BKGRDINPROG is not set in the original buffer it must 945 * have been released and re-instantiated - which is not legal. 946 */ 947 KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 948 ("backgroundwritedone: lost buffer2")); 949 origbp->b_vflags &= ~BV_BKGRDINPROG; 950 if (origbp->b_vflags & BV_BKGRDWAIT) { 951 origbp->b_vflags &= ~BV_BKGRDWAIT; 952 wakeup(&origbp->b_xflags); 953 } 954 VI_UNLOCK(bp->b_vp); 955 /* 956 * Process dependencies then return any unfinished ones. 957 */ 958 if (LIST_FIRST(&bp->b_dep) != NULL) 959 buf_complete(bp); 960 if (LIST_FIRST(&bp->b_dep) != NULL) 961 buf_movedeps(bp, origbp); 962 963 /* 964 * This buffer is marked B_NOCACHE, so when it is released 965 * by biodone, it will be tossed. We mark it with BIO_READ 966 * to avoid biodone doing a second bufobj_wakeup. 967 */ 968 bp->b_flags |= B_NOCACHE; 969 bp->b_iocmd = BIO_READ; 970 bp->b_flags &= ~(B_CACHE | B_DONE); 971 bp->b_iodone = 0; 972 bufdone(bp); 973} 974 975/* 976 * Delayed write. (Buffer is marked dirty). Do not bother writing 977 * anything if the buffer is marked invalid. 978 * 979 * Note that since the buffer must be completely valid, we can safely 980 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 981 * biodone() in order to prevent getblk from writing the buffer 982 * out synchronously. 983 */ 984void 985bdwrite(struct buf *bp) 986{ 987 struct thread *td = curthread; 988 struct vnode *vp; 989 struct buf *nbp; 990 991 GIANT_REQUIRED; 992 993 if (BUF_REFCNT(bp) == 0) 994 panic("bdwrite: buffer is not busy"); 995 996 if (bp->b_flags & B_INVAL) { 997 brelse(bp); 998 return; 999 } 1000 1001 /* 1002 * If we have too many dirty buffers, don't create any more. 1003 * If we are wildly over our limit, then force a complete 1004 * cleanup. Otherwise, just keep the situation from getting 1005 * out of control. Note that we have to avoid a recursive 1006 * disaster and not try to clean up after our own cleanup! 1007 */ 1008 vp = bp->b_vp; 1009 VI_LOCK(vp); 1010 if (td->td_pflags & TDP_COWINPROGRESS) { 1011 recursiveflushes++; 1012 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh + 10) { 1013 VI_UNLOCK(vp); 1014 (void) VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td); 1015 VI_LOCK(vp); 1016 altbufferflushes++; 1017 } else if (vp != NULL && vp->v_dirtybufcnt > dirtybufthresh) { 1018 /* 1019 * Try to find a buffer to flush. 1020 */ 1021 TAILQ_FOREACH(nbp, &vp->v_dirtyblkhd, b_bobufs) { 1022 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1023 buf_countdeps(nbp, 0) || 1024 BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1025 continue; 1026 if (bp == nbp) 1027 panic("bdwrite: found ourselves"); 1028 VI_UNLOCK(vp); 1029 if (nbp->b_flags & B_CLUSTEROK) { 1030 vfs_bio_awrite(nbp); 1031 } else { 1032 bremfree(nbp); 1033 bawrite(nbp); 1034 } 1035 VI_LOCK(vp); 1036 dirtybufferflushes++; 1037 break; 1038 } 1039 } 1040 VI_UNLOCK(vp); 1041 1042 bdirty(bp); 1043 /* 1044 * Set B_CACHE, indicating that the buffer is fully valid. This is 1045 * true even of NFS now. 1046 */ 1047 bp->b_flags |= B_CACHE; 1048 1049 /* 1050 * This bmap keeps the system from needing to do the bmap later, 1051 * perhaps when the system is attempting to do a sync. Since it 1052 * is likely that the indirect block -- or whatever other datastructure 1053 * that the filesystem needs is still in memory now, it is a good 1054 * thing to do this. Note also, that if the pageout daemon is 1055 * requesting a sync -- there might not be enough memory to do 1056 * the bmap then... So, this is important to do. 1057 */ 1058 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1059 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1060 } 1061 1062 /* 1063 * Set the *dirty* buffer range based upon the VM system dirty pages. 1064 */ 1065 vfs_setdirty(bp); 1066 1067 /* 1068 * We need to do this here to satisfy the vnode_pager and the 1069 * pageout daemon, so that it thinks that the pages have been 1070 * "cleaned". Note that since the pages are in a delayed write 1071 * buffer -- the VFS layer "will" see that the pages get written 1072 * out on the next sync, or perhaps the cluster will be completed. 1073 */ 1074 vfs_clean_pages(bp); 1075 bqrelse(bp); 1076 1077 /* 1078 * Wakeup the buffer flushing daemon if we have a lot of dirty 1079 * buffers (midpoint between our recovery point and our stall 1080 * point). 1081 */ 1082 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1083 1084 /* 1085 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1086 * due to the softdep code. 1087 */ 1088} 1089 1090/* 1091 * bdirty: 1092 * 1093 * Turn buffer into delayed write request. We must clear BIO_READ and 1094 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1095 * itself to properly update it in the dirty/clean lists. We mark it 1096 * B_DONE to ensure that any asynchronization of the buffer properly 1097 * clears B_DONE ( else a panic will occur later ). 1098 * 1099 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1100 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1101 * should only be called if the buffer is known-good. 1102 * 1103 * Since the buffer is not on a queue, we do not update the numfreebuffers 1104 * count. 1105 * 1106 * Must be called at splbio(). 1107 * The buffer must be on QUEUE_NONE. 1108 */ 1109void 1110bdirty(struct buf *bp) 1111{ 1112 1113 KASSERT(bp->b_qindex == QUEUE_NONE, 1114 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1115 bp->b_flags &= ~(B_RELBUF); 1116 bp->b_iocmd = BIO_WRITE; 1117 1118 if ((bp->b_flags & B_DELWRI) == 0) { 1119 bp->b_flags |= B_DONE | B_DELWRI; 1120 reassignbuf(bp); 1121 atomic_add_int(&numdirtybuffers, 1); 1122 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1123 } 1124} 1125 1126/* 1127 * bundirty: 1128 * 1129 * Clear B_DELWRI for buffer. 1130 * 1131 * Since the buffer is not on a queue, we do not update the numfreebuffers 1132 * count. 1133 * 1134 * Must be called at splbio(). 1135 * The buffer must be on QUEUE_NONE. 1136 */ 1137 1138void 1139bundirty(struct buf *bp) 1140{ 1141 1142 KASSERT(bp->b_qindex == QUEUE_NONE, 1143 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1144 1145 if (bp->b_flags & B_DELWRI) { 1146 bp->b_flags &= ~B_DELWRI; 1147 reassignbuf(bp); 1148 atomic_subtract_int(&numdirtybuffers, 1); 1149 numdirtywakeup(lodirtybuffers); 1150 } 1151 /* 1152 * Since it is now being written, we can clear its deferred write flag. 1153 */ 1154 bp->b_flags &= ~B_DEFERRED; 1155} 1156 1157/* 1158 * bawrite: 1159 * 1160 * Asynchronous write. Start output on a buffer, but do not wait for 1161 * it to complete. The buffer is released when the output completes. 1162 * 1163 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1164 * B_INVAL buffers. Not us. 1165 */ 1166void 1167bawrite(struct buf *bp) 1168{ 1169 1170 bp->b_flags |= B_ASYNC; 1171 (void) bwrite(bp); 1172} 1173 1174/* 1175 * bwillwrite: 1176 * 1177 * Called prior to the locking of any vnodes when we are expecting to 1178 * write. We do not want to starve the buffer cache with too many 1179 * dirty buffers so we block here. By blocking prior to the locking 1180 * of any vnodes we attempt to avoid the situation where a locked vnode 1181 * prevents the various system daemons from flushing related buffers. 1182 */ 1183 1184void 1185bwillwrite(void) 1186{ 1187 1188 if (numdirtybuffers >= hidirtybuffers) { 1189 int s; 1190 1191 mtx_lock(&Giant); 1192 s = splbio(); 1193 mtx_lock(&nblock); 1194 while (numdirtybuffers >= hidirtybuffers) { 1195 bd_wakeup(1); 1196 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1197 msleep(&needsbuffer, &nblock, 1198 (PRIBIO + 4), "flswai", 0); 1199 } 1200 splx(s); 1201 mtx_unlock(&nblock); 1202 mtx_unlock(&Giant); 1203 } 1204} 1205 1206/* 1207 * Return true if we have too many dirty buffers. 1208 */ 1209int 1210buf_dirty_count_severe(void) 1211{ 1212 1213 return(numdirtybuffers >= hidirtybuffers); 1214} 1215 1216/* 1217 * brelse: 1218 * 1219 * Release a busy buffer and, if requested, free its resources. The 1220 * buffer will be stashed in the appropriate bufqueue[] allowing it 1221 * to be accessed later as a cache entity or reused for other purposes. 1222 */ 1223void 1224brelse(struct buf *bp) 1225{ 1226 int s; 1227 1228 GIANT_REQUIRED; 1229 1230 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1231 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1232 1233 s = splbio(); 1234 1235 if (bp->b_iocmd == BIO_WRITE && 1236 (bp->b_ioflags & BIO_ERROR) && 1237 !(bp->b_flags & B_INVAL)) { 1238 /* 1239 * Failed write, redirty. Must clear BIO_ERROR to prevent 1240 * pages from being scrapped. If B_INVAL is set then 1241 * this case is not run and the next case is run to 1242 * destroy the buffer. B_INVAL can occur if the buffer 1243 * is outside the range supported by the underlying device. 1244 */ 1245 bp->b_ioflags &= ~BIO_ERROR; 1246 bdirty(bp); 1247 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1248 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1249 /* 1250 * Either a failed I/O or we were asked to free or not 1251 * cache the buffer. 1252 */ 1253 bp->b_flags |= B_INVAL; 1254 if (LIST_FIRST(&bp->b_dep) != NULL) 1255 buf_deallocate(bp); 1256 if (bp->b_flags & B_DELWRI) { 1257 atomic_subtract_int(&numdirtybuffers, 1); 1258 numdirtywakeup(lodirtybuffers); 1259 } 1260 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1261 if ((bp->b_flags & B_VMIO) == 0) { 1262 if (bp->b_bufsize) 1263 allocbuf(bp, 0); 1264 if (bp->b_vp) 1265 brelvp(bp); 1266 } 1267 } 1268 1269 /* 1270 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1271 * is called with B_DELWRI set, the underlying pages may wind up 1272 * getting freed causing a previous write (bdwrite()) to get 'lost' 1273 * because pages associated with a B_DELWRI bp are marked clean. 1274 * 1275 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1276 * if B_DELWRI is set. 1277 * 1278 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1279 * on pages to return pages to the VM page queues. 1280 */ 1281 if (bp->b_flags & B_DELWRI) 1282 bp->b_flags &= ~B_RELBUF; 1283 else if (vm_page_count_severe()) { 1284 /* 1285 * XXX This lock may not be necessary since BKGRDINPROG 1286 * cannot be set while we hold the buf lock, it can only be 1287 * cleared if it is already pending. 1288 */ 1289 if (bp->b_vp) { 1290 VI_LOCK(bp->b_vp); 1291 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1292 bp->b_flags |= B_RELBUF; 1293 VI_UNLOCK(bp->b_vp); 1294 } else 1295 bp->b_flags |= B_RELBUF; 1296 } 1297 1298 /* 1299 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1300 * constituted, not even NFS buffers now. Two flags effect this. If 1301 * B_INVAL, the struct buf is invalidated but the VM object is kept 1302 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1303 * 1304 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1305 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1306 * buffer is also B_INVAL because it hits the re-dirtying code above. 1307 * 1308 * Normally we can do this whether a buffer is B_DELWRI or not. If 1309 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1310 * the commit state and we cannot afford to lose the buffer. If the 1311 * buffer has a background write in progress, we need to keep it 1312 * around to prevent it from being reconstituted and starting a second 1313 * background write. 1314 */ 1315 if ((bp->b_flags & B_VMIO) 1316 && !(bp->b_vp->v_mount != NULL && 1317 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1318 !vn_isdisk(bp->b_vp, NULL) && 1319 (bp->b_flags & B_DELWRI)) 1320 ) { 1321 1322 int i, j, resid; 1323 vm_page_t m; 1324 off_t foff; 1325 vm_pindex_t poff; 1326 vm_object_t obj; 1327 1328 obj = bp->b_object; 1329 1330 /* 1331 * Get the base offset and length of the buffer. Note that 1332 * in the VMIO case if the buffer block size is not 1333 * page-aligned then b_data pointer may not be page-aligned. 1334 * But our b_pages[] array *IS* page aligned. 1335 * 1336 * block sizes less then DEV_BSIZE (usually 512) are not 1337 * supported due to the page granularity bits (m->valid, 1338 * m->dirty, etc...). 1339 * 1340 * See man buf(9) for more information 1341 */ 1342 resid = bp->b_bufsize; 1343 foff = bp->b_offset; 1344 VM_OBJECT_LOCK(obj); 1345 for (i = 0; i < bp->b_npages; i++) { 1346 int had_bogus = 0; 1347 1348 m = bp->b_pages[i]; 1349 1350 /* 1351 * If we hit a bogus page, fixup *all* the bogus pages 1352 * now. 1353 */ 1354 if (m == bogus_page) { 1355 poff = OFF_TO_IDX(bp->b_offset); 1356 had_bogus = 1; 1357 1358 for (j = i; j < bp->b_npages; j++) { 1359 vm_page_t mtmp; 1360 mtmp = bp->b_pages[j]; 1361 if (mtmp == bogus_page) { 1362 mtmp = vm_page_lookup(obj, poff + j); 1363 if (!mtmp) { 1364 panic("brelse: page missing\n"); 1365 } 1366 bp->b_pages[j] = mtmp; 1367 } 1368 } 1369 1370 if ((bp->b_flags & B_INVAL) == 0) { 1371 pmap_qenter( 1372 trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1373 } 1374 m = bp->b_pages[i]; 1375 } 1376 if ((bp->b_flags & B_NOCACHE) || 1377 (bp->b_ioflags & BIO_ERROR)) { 1378 int poffset = foff & PAGE_MASK; 1379 int presid = resid > (PAGE_SIZE - poffset) ? 1380 (PAGE_SIZE - poffset) : resid; 1381 1382 KASSERT(presid >= 0, ("brelse: extra page")); 1383 vm_page_lock_queues(); 1384 vm_page_set_invalid(m, poffset, presid); 1385 vm_page_unlock_queues(); 1386 if (had_bogus) 1387 printf("avoided corruption bug in bogus_page/brelse code\n"); 1388 } 1389 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1390 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1391 } 1392 VM_OBJECT_UNLOCK(obj); 1393 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1394 vfs_vmio_release(bp); 1395 1396 } else if (bp->b_flags & B_VMIO) { 1397 1398 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1399 vfs_vmio_release(bp); 1400 } 1401 1402 } 1403 1404 if (bp->b_qindex != QUEUE_NONE) 1405 panic("brelse: free buffer onto another queue???"); 1406 if (BUF_REFCNT(bp) > 1) { 1407 /* do not release to free list */ 1408 BUF_UNLOCK(bp); 1409 splx(s); 1410 return; 1411 } 1412 1413 /* enqueue */ 1414 mtx_lock(&bqlock); 1415 1416 /* buffers with no memory */ 1417 if (bp->b_bufsize == 0) { 1418 bp->b_flags |= B_INVAL; 1419 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1420 if (bp->b_vflags & BV_BKGRDINPROG) 1421 panic("losing buffer 1"); 1422 if (bp->b_kvasize) { 1423 bp->b_qindex = QUEUE_EMPTYKVA; 1424 } else { 1425 bp->b_qindex = QUEUE_EMPTY; 1426 } 1427 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1428 bp->b_dev = NULL; 1429 /* buffers with junk contents */ 1430 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1431 (bp->b_ioflags & BIO_ERROR)) { 1432 bp->b_flags |= B_INVAL; 1433 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1434 if (bp->b_vflags & BV_BKGRDINPROG) 1435 panic("losing buffer 2"); 1436 bp->b_qindex = QUEUE_CLEAN; 1437 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1438 bp->b_dev = NULL; 1439 /* remaining buffers */ 1440 } else { 1441 if (bp->b_flags & B_DELWRI) 1442 bp->b_qindex = QUEUE_DIRTY; 1443 else 1444 bp->b_qindex = QUEUE_CLEAN; 1445 if (bp->b_flags & B_AGE) 1446 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1447 else 1448 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1449 } 1450 mtx_unlock(&bqlock); 1451 1452 /* 1453 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1454 * placed the buffer on the correct queue. We must also disassociate 1455 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1456 * find it. 1457 */ 1458 if (bp->b_flags & B_INVAL) { 1459 if (bp->b_flags & B_DELWRI) 1460 bundirty(bp); 1461 if (bp->b_vp) 1462 brelvp(bp); 1463 } 1464 1465 /* 1466 * Fixup numfreebuffers count. The bp is on an appropriate queue 1467 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1468 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1469 * if B_INVAL is set ). 1470 */ 1471 1472 if (!(bp->b_flags & B_DELWRI)) 1473 bufcountwakeup(); 1474 1475 /* 1476 * Something we can maybe free or reuse 1477 */ 1478 if (bp->b_bufsize || bp->b_kvasize) 1479 bufspacewakeup(); 1480 1481 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1482 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1483 panic("brelse: not dirty"); 1484 /* unlock */ 1485 BUF_UNLOCK(bp); 1486 splx(s); 1487} 1488 1489/* 1490 * Release a buffer back to the appropriate queue but do not try to free 1491 * it. The buffer is expected to be used again soon. 1492 * 1493 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1494 * biodone() to requeue an async I/O on completion. It is also used when 1495 * known good buffers need to be requeued but we think we may need the data 1496 * again soon. 1497 * 1498 * XXX we should be able to leave the B_RELBUF hint set on completion. 1499 */ 1500void 1501bqrelse(struct buf *bp) 1502{ 1503 int s; 1504 1505 s = splbio(); 1506 1507 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1508 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1509 1510 if (bp->b_qindex != QUEUE_NONE) 1511 panic("bqrelse: free buffer onto another queue???"); 1512 if (BUF_REFCNT(bp) > 1) { 1513 /* do not release to free list */ 1514 BUF_UNLOCK(bp); 1515 splx(s); 1516 return; 1517 } 1518 mtx_lock(&bqlock); 1519 /* buffers with stale but valid contents */ 1520 if (bp->b_flags & B_DELWRI) { 1521 bp->b_qindex = QUEUE_DIRTY; 1522 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1523 } else { 1524 /* 1525 * XXX This lock may not be necessary since BKGRDINPROG 1526 * cannot be set while we hold the buf lock, it can only be 1527 * cleared if it is already pending. 1528 */ 1529 VI_LOCK(bp->b_vp); 1530 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1531 VI_UNLOCK(bp->b_vp); 1532 bp->b_qindex = QUEUE_CLEAN; 1533 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1534 b_freelist); 1535 } else { 1536 /* 1537 * We are too low on memory, we have to try to free 1538 * the buffer (most importantly: the wired pages 1539 * making up its backing store) *now*. 1540 */ 1541 VI_UNLOCK(bp->b_vp); 1542 mtx_unlock(&bqlock); 1543 splx(s); 1544 brelse(bp); 1545 return; 1546 } 1547 } 1548 mtx_unlock(&bqlock); 1549 1550 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1551 bufcountwakeup(); 1552 1553 /* 1554 * Something we can maybe free or reuse. 1555 */ 1556 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1557 bufspacewakeup(); 1558 1559 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1560 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1561 panic("bqrelse: not dirty"); 1562 /* unlock */ 1563 BUF_UNLOCK(bp); 1564 splx(s); 1565} 1566 1567/* Give pages used by the bp back to the VM system (where possible) */ 1568static void 1569vfs_vmio_release(struct buf *bp) 1570{ 1571 int i; 1572 vm_page_t m; 1573 1574 GIANT_REQUIRED; 1575 VM_OBJECT_LOCK(bp->b_object); 1576 vm_page_lock_queues(); 1577 for (i = 0; i < bp->b_npages; i++) { 1578 m = bp->b_pages[i]; 1579 bp->b_pages[i] = NULL; 1580 /* 1581 * In order to keep page LRU ordering consistent, put 1582 * everything on the inactive queue. 1583 */ 1584 vm_page_unwire(m, 0); 1585 /* 1586 * We don't mess with busy pages, it is 1587 * the responsibility of the process that 1588 * busied the pages to deal with them. 1589 */ 1590 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1591 continue; 1592 1593 if (m->wire_count == 0) { 1594 /* 1595 * Might as well free the page if we can and it has 1596 * no valid data. We also free the page if the 1597 * buffer was used for direct I/O 1598 */ 1599 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1600 m->hold_count == 0) { 1601 vm_page_busy(m); 1602 pmap_remove_all(m); 1603 vm_page_free(m); 1604 } else if (bp->b_flags & B_DIRECT) { 1605 vm_page_try_to_free(m); 1606 } else if (vm_page_count_severe()) { 1607 vm_page_try_to_cache(m); 1608 } 1609 } 1610 } 1611 vm_page_unlock_queues(); 1612 VM_OBJECT_UNLOCK(bp->b_object); 1613 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1614 1615 if (bp->b_bufsize) { 1616 bufspacewakeup(); 1617 bp->b_bufsize = 0; 1618 } 1619 bp->b_npages = 0; 1620 bp->b_flags &= ~B_VMIO; 1621 if (bp->b_vp) 1622 brelvp(bp); 1623} 1624 1625/* 1626 * Check to see if a block at a particular lbn is available for a clustered 1627 * write. 1628 */ 1629static int 1630vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1631{ 1632 struct buf *bpa; 1633 int match; 1634 1635 match = 0; 1636 1637 /* If the buf isn't in core skip it */ 1638 if ((bpa = gbincore(vp, lblkno)) == NULL) 1639 return (0); 1640 1641 /* If the buf is busy we don't want to wait for it */ 1642 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1643 return (0); 1644 1645 /* Only cluster with valid clusterable delayed write buffers */ 1646 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1647 (B_DELWRI | B_CLUSTEROK)) 1648 goto done; 1649 1650 if (bpa->b_bufsize != size) 1651 goto done; 1652 1653 /* 1654 * Check to see if it is in the expected place on disk and that the 1655 * block has been mapped. 1656 */ 1657 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1658 match = 1; 1659done: 1660 BUF_UNLOCK(bpa); 1661 return (match); 1662} 1663 1664/* 1665 * vfs_bio_awrite: 1666 * 1667 * Implement clustered async writes for clearing out B_DELWRI buffers. 1668 * This is much better then the old way of writing only one buffer at 1669 * a time. Note that we may not be presented with the buffers in the 1670 * correct order, so we search for the cluster in both directions. 1671 */ 1672int 1673vfs_bio_awrite(struct buf *bp) 1674{ 1675 int i; 1676 int j; 1677 daddr_t lblkno = bp->b_lblkno; 1678 struct vnode *vp = bp->b_vp; 1679 int s; 1680 int ncl; 1681 int nwritten; 1682 int size; 1683 int maxcl; 1684 1685 s = splbio(); 1686 /* 1687 * right now we support clustered writing only to regular files. If 1688 * we find a clusterable block we could be in the middle of a cluster 1689 * rather then at the beginning. 1690 */ 1691 if ((vp->v_type == VREG) && 1692 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1693 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1694 1695 size = vp->v_mount->mnt_stat.f_iosize; 1696 maxcl = MAXPHYS / size; 1697 1698 VI_LOCK(vp); 1699 for (i = 1; i < maxcl; i++) 1700 if (vfs_bio_clcheck(vp, size, lblkno + i, 1701 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1702 break; 1703 1704 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1705 if (vfs_bio_clcheck(vp, size, lblkno - j, 1706 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1707 break; 1708 1709 VI_UNLOCK(vp); 1710 --j; 1711 ncl = i + j; 1712 /* 1713 * this is a possible cluster write 1714 */ 1715 if (ncl != 1) { 1716 BUF_UNLOCK(bp); 1717 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1718 splx(s); 1719 return nwritten; 1720 } 1721 } 1722 1723 bremfree(bp); 1724 bp->b_flags |= B_ASYNC; 1725 1726 splx(s); 1727 /* 1728 * default (old) behavior, writing out only one block 1729 * 1730 * XXX returns b_bufsize instead of b_bcount for nwritten? 1731 */ 1732 nwritten = bp->b_bufsize; 1733 (void) bwrite(bp); 1734 1735 return nwritten; 1736} 1737 1738/* 1739 * getnewbuf: 1740 * 1741 * Find and initialize a new buffer header, freeing up existing buffers 1742 * in the bufqueues as necessary. The new buffer is returned locked. 1743 * 1744 * Important: B_INVAL is not set. If the caller wishes to throw the 1745 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1746 * 1747 * We block if: 1748 * We have insufficient buffer headers 1749 * We have insufficient buffer space 1750 * buffer_map is too fragmented ( space reservation fails ) 1751 * If we have to flush dirty buffers ( but we try to avoid this ) 1752 * 1753 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1754 * Instead we ask the buf daemon to do it for us. We attempt to 1755 * avoid piecemeal wakeups of the pageout daemon. 1756 */ 1757 1758static struct buf * 1759getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1760{ 1761 struct buf *bp; 1762 struct buf *nbp; 1763 int defrag = 0; 1764 int nqindex; 1765 static int flushingbufs; 1766 1767 GIANT_REQUIRED; 1768 1769 /* 1770 * We can't afford to block since we might be holding a vnode lock, 1771 * which may prevent system daemons from running. We deal with 1772 * low-memory situations by proactively returning memory and running 1773 * async I/O rather then sync I/O. 1774 */ 1775 1776 atomic_add_int(&getnewbufcalls, 1); 1777 atomic_subtract_int(&getnewbufrestarts, 1); 1778restart: 1779 atomic_add_int(&getnewbufrestarts, 1); 1780 1781 /* 1782 * Setup for scan. If we do not have enough free buffers, 1783 * we setup a degenerate case that immediately fails. Note 1784 * that if we are specially marked process, we are allowed to 1785 * dip into our reserves. 1786 * 1787 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1788 * 1789 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1790 * However, there are a number of cases (defragging, reusing, ...) 1791 * where we cannot backup. 1792 */ 1793 mtx_lock(&bqlock); 1794 nqindex = QUEUE_EMPTYKVA; 1795 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1796 1797 if (nbp == NULL) { 1798 /* 1799 * If no EMPTYKVA buffers and we are either 1800 * defragging or reusing, locate a CLEAN buffer 1801 * to free or reuse. If bufspace useage is low 1802 * skip this step so we can allocate a new buffer. 1803 */ 1804 if (defrag || bufspace >= lobufspace) { 1805 nqindex = QUEUE_CLEAN; 1806 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1807 } 1808 1809 /* 1810 * If we could not find or were not allowed to reuse a 1811 * CLEAN buffer, check to see if it is ok to use an EMPTY 1812 * buffer. We can only use an EMPTY buffer if allocating 1813 * its KVA would not otherwise run us out of buffer space. 1814 */ 1815 if (nbp == NULL && defrag == 0 && 1816 bufspace + maxsize < hibufspace) { 1817 nqindex = QUEUE_EMPTY; 1818 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1819 } 1820 } 1821 1822 /* 1823 * Run scan, possibly freeing data and/or kva mappings on the fly 1824 * depending. 1825 */ 1826 1827 while ((bp = nbp) != NULL) { 1828 int qindex = nqindex; 1829 1830 /* 1831 * Calculate next bp ( we can only use it if we do not block 1832 * or do other fancy things ). 1833 */ 1834 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1835 switch(qindex) { 1836 case QUEUE_EMPTY: 1837 nqindex = QUEUE_EMPTYKVA; 1838 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1839 break; 1840 /* FALLTHROUGH */ 1841 case QUEUE_EMPTYKVA: 1842 nqindex = QUEUE_CLEAN; 1843 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1844 break; 1845 /* FALLTHROUGH */ 1846 case QUEUE_CLEAN: 1847 /* 1848 * nbp is NULL. 1849 */ 1850 break; 1851 } 1852 } 1853 if (bp->b_vp) { 1854 VI_LOCK(bp->b_vp); 1855 if (bp->b_vflags & BV_BKGRDINPROG) { 1856 VI_UNLOCK(bp->b_vp); 1857 continue; 1858 } 1859 VI_UNLOCK(bp->b_vp); 1860 } 1861 1862 /* 1863 * Sanity Checks 1864 */ 1865 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1866 1867 /* 1868 * Note: we no longer distinguish between VMIO and non-VMIO 1869 * buffers. 1870 */ 1871 1872 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1873 1874 /* 1875 * If we are defragging then we need a buffer with 1876 * b_kvasize != 0. XXX this situation should no longer 1877 * occur, if defrag is non-zero the buffer's b_kvasize 1878 * should also be non-zero at this point. XXX 1879 */ 1880 if (defrag && bp->b_kvasize == 0) { 1881 printf("Warning: defrag empty buffer %p\n", bp); 1882 continue; 1883 } 1884 1885 /* 1886 * Start freeing the bp. This is somewhat involved. nbp 1887 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1888 */ 1889 1890 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1891 panic("getnewbuf: locked buf"); 1892 bremfreel(bp); 1893 mtx_unlock(&bqlock); 1894 1895 if (qindex == QUEUE_CLEAN) { 1896 if (bp->b_flags & B_VMIO) { 1897 bp->b_flags &= ~B_ASYNC; 1898 vfs_vmio_release(bp); 1899 } 1900 if (bp->b_vp) 1901 brelvp(bp); 1902 } 1903 1904 /* 1905 * NOTE: nbp is now entirely invalid. We can only restart 1906 * the scan from this point on. 1907 * 1908 * Get the rest of the buffer freed up. b_kva* is still 1909 * valid after this operation. 1910 */ 1911 1912 if (bp->b_rcred != NOCRED) { 1913 crfree(bp->b_rcred); 1914 bp->b_rcred = NOCRED; 1915 } 1916 if (bp->b_wcred != NOCRED) { 1917 crfree(bp->b_wcred); 1918 bp->b_wcred = NOCRED; 1919 } 1920 if (LIST_FIRST(&bp->b_dep) != NULL) 1921 buf_deallocate(bp); 1922 if (bp->b_vflags & BV_BKGRDINPROG) 1923 panic("losing buffer 3"); 1924 1925 if (bp->b_bufsize) 1926 allocbuf(bp, 0); 1927 1928 bp->b_flags = 0; 1929 bp->b_ioflags = 0; 1930 bp->b_xflags = 0; 1931 bp->b_vflags = 0; 1932 bp->b_dev = NULL; 1933 bp->b_vp = NULL; 1934 bp->b_blkno = bp->b_lblkno = 0; 1935 bp->b_offset = NOOFFSET; 1936 bp->b_iodone = 0; 1937 bp->b_error = 0; 1938 bp->b_resid = 0; 1939 bp->b_bcount = 0; 1940 bp->b_npages = 0; 1941 bp->b_dirtyoff = bp->b_dirtyend = 0; 1942 bp->b_magic = B_MAGIC_BIO; 1943 bp->b_op = &buf_ops_bio; 1944 bp->b_object = NULL; 1945 1946 LIST_INIT(&bp->b_dep); 1947 1948 /* 1949 * If we are defragging then free the buffer. 1950 */ 1951 if (defrag) { 1952 bp->b_flags |= B_INVAL; 1953 bfreekva(bp); 1954 brelse(bp); 1955 defrag = 0; 1956 goto restart; 1957 } 1958 1959 /* 1960 * If we are overcomitted then recover the buffer and its 1961 * KVM space. This occurs in rare situations when multiple 1962 * processes are blocked in getnewbuf() or allocbuf(). 1963 */ 1964 if (bufspace >= hibufspace) 1965 flushingbufs = 1; 1966 if (flushingbufs && bp->b_kvasize != 0) { 1967 bp->b_flags |= B_INVAL; 1968 bfreekva(bp); 1969 brelse(bp); 1970 goto restart; 1971 } 1972 if (bufspace < lobufspace) 1973 flushingbufs = 0; 1974 break; 1975 } 1976 1977 /* 1978 * If we exhausted our list, sleep as appropriate. We may have to 1979 * wakeup various daemons and write out some dirty buffers. 1980 * 1981 * Generally we are sleeping due to insufficient buffer space. 1982 */ 1983 1984 if (bp == NULL) { 1985 int flags; 1986 char *waitmsg; 1987 1988 mtx_unlock(&bqlock); 1989 if (defrag) { 1990 flags = VFS_BIO_NEED_BUFSPACE; 1991 waitmsg = "nbufkv"; 1992 } else if (bufspace >= hibufspace) { 1993 waitmsg = "nbufbs"; 1994 flags = VFS_BIO_NEED_BUFSPACE; 1995 } else { 1996 waitmsg = "newbuf"; 1997 flags = VFS_BIO_NEED_ANY; 1998 } 1999 2000 bd_speedup(); /* heeeelp */ 2001 2002 mtx_lock(&nblock); 2003 needsbuffer |= flags; 2004 while (needsbuffer & flags) { 2005 if (msleep(&needsbuffer, &nblock, 2006 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2007 mtx_unlock(&nblock); 2008 return (NULL); 2009 } 2010 } 2011 mtx_unlock(&nblock); 2012 } else { 2013 /* 2014 * We finally have a valid bp. We aren't quite out of the 2015 * woods, we still have to reserve kva space. In order 2016 * to keep fragmentation sane we only allocate kva in 2017 * BKVASIZE chunks. 2018 */ 2019 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2020 2021 if (maxsize != bp->b_kvasize) { 2022 vm_offset_t addr = 0; 2023 2024 bfreekva(bp); 2025 2026 if (vm_map_findspace(buffer_map, 2027 vm_map_min(buffer_map), maxsize, &addr)) { 2028 /* 2029 * Uh oh. Buffer map is to fragmented. We 2030 * must defragment the map. 2031 */ 2032 atomic_add_int(&bufdefragcnt, 1); 2033 defrag = 1; 2034 bp->b_flags |= B_INVAL; 2035 brelse(bp); 2036 goto restart; 2037 } 2038 if (addr) { 2039 vm_map_insert(buffer_map, NULL, 0, 2040 addr, addr + maxsize, 2041 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2042 2043 bp->b_kvabase = (caddr_t) addr; 2044 bp->b_kvasize = maxsize; 2045 atomic_add_int(&bufspace, bp->b_kvasize); 2046 atomic_add_int(&bufreusecnt, 1); 2047 } 2048 } 2049 bp->b_saveaddr = bp->b_kvabase; 2050 bp->b_data = bp->b_saveaddr; 2051 } 2052 return(bp); 2053} 2054 2055/* 2056 * buf_daemon: 2057 * 2058 * buffer flushing daemon. Buffers are normally flushed by the 2059 * update daemon but if it cannot keep up this process starts to 2060 * take the load in an attempt to prevent getnewbuf() from blocking. 2061 */ 2062 2063static struct kproc_desc buf_kp = { 2064 "bufdaemon", 2065 buf_daemon, 2066 &bufdaemonproc 2067}; 2068SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2069 2070static void 2071buf_daemon() 2072{ 2073 int s; 2074 2075 mtx_lock(&Giant); 2076 2077 /* 2078 * This process needs to be suspended prior to shutdown sync. 2079 */ 2080 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2081 SHUTDOWN_PRI_LAST); 2082 2083 /* 2084 * This process is allowed to take the buffer cache to the limit 2085 */ 2086 s = splbio(); 2087 mtx_lock(&bdlock); 2088 2089 for (;;) { 2090 bd_request = 0; 2091 mtx_unlock(&bdlock); 2092 2093 kthread_suspend_check(bufdaemonproc); 2094 2095 /* 2096 * Do the flush. Limit the amount of in-transit I/O we 2097 * allow to build up, otherwise we would completely saturate 2098 * the I/O system. Wakeup any waiting processes before we 2099 * normally would so they can run in parallel with our drain. 2100 */ 2101 while (numdirtybuffers > lodirtybuffers) { 2102 if (flushbufqueues(0) == 0) { 2103 /* 2104 * Could not find any buffers without rollback 2105 * dependencies, so just write the first one 2106 * in the hopes of eventually making progress. 2107 */ 2108 flushbufqueues(1); 2109 break; 2110 } 2111 waitrunningbufspace(); 2112 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2113 } 2114 2115 /* 2116 * Only clear bd_request if we have reached our low water 2117 * mark. The buf_daemon normally waits 1 second and 2118 * then incrementally flushes any dirty buffers that have 2119 * built up, within reason. 2120 * 2121 * If we were unable to hit our low water mark and couldn't 2122 * find any flushable buffers, we sleep half a second. 2123 * Otherwise we loop immediately. 2124 */ 2125 mtx_lock(&bdlock); 2126 if (numdirtybuffers <= lodirtybuffers) { 2127 /* 2128 * We reached our low water mark, reset the 2129 * request and sleep until we are needed again. 2130 * The sleep is just so the suspend code works. 2131 */ 2132 bd_request = 0; 2133 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2134 } else { 2135 /* 2136 * We couldn't find any flushable dirty buffers but 2137 * still have too many dirty buffers, we 2138 * have to sleep and try again. (rare) 2139 */ 2140 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2141 } 2142 } 2143} 2144 2145/* 2146 * flushbufqueues: 2147 * 2148 * Try to flush a buffer in the dirty queue. We must be careful to 2149 * free up B_INVAL buffers instead of write them, which NFS is 2150 * particularly sensitive to. 2151 */ 2152int flushwithdeps = 0; 2153SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2154 0, "Number of buffers flushed with dependecies that require rollbacks"); 2155 2156static int 2157flushbufqueues(int flushdeps) 2158{ 2159 struct thread *td = curthread; 2160 struct vnode *vp; 2161 struct mount *mp; 2162 struct buf *bp; 2163 int hasdeps; 2164 2165 mtx_lock(&bqlock); 2166 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2167 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2168 continue; 2169 KASSERT((bp->b_flags & B_DELWRI), 2170 ("unexpected clean buffer %p", bp)); 2171 VI_LOCK(bp->b_vp); 2172 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2173 VI_UNLOCK(bp->b_vp); 2174 BUF_UNLOCK(bp); 2175 continue; 2176 } 2177 VI_UNLOCK(bp->b_vp); 2178 if (bp->b_flags & B_INVAL) { 2179 bremfreel(bp); 2180 mtx_unlock(&bqlock); 2181 brelse(bp); 2182 return (1); 2183 } 2184 2185 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2186 if (flushdeps == 0) { 2187 BUF_UNLOCK(bp); 2188 continue; 2189 } 2190 hasdeps = 1; 2191 } else 2192 hasdeps = 0; 2193 /* 2194 * We must hold the lock on a vnode before writing 2195 * one of its buffers. Otherwise we may confuse, or 2196 * in the case of a snapshot vnode, deadlock the 2197 * system. 2198 * 2199 * The lock order here is the reverse of the normal 2200 * of vnode followed by buf lock. This is ok because 2201 * the NOWAIT will prevent deadlock. 2202 */ 2203 vp = bp->b_vp; 2204 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2205 BUF_UNLOCK(bp); 2206 continue; 2207 } 2208 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2209 mtx_unlock(&bqlock); 2210 vfs_bio_awrite(bp); 2211 vn_finished_write(mp); 2212 VOP_UNLOCK(vp, 0, td); 2213 flushwithdeps += hasdeps; 2214 return (1); 2215 } 2216 vn_finished_write(mp); 2217 BUF_UNLOCK(bp); 2218 } 2219 mtx_unlock(&bqlock); 2220 return (0); 2221} 2222 2223/* 2224 * Check to see if a block is currently memory resident. 2225 */ 2226struct buf * 2227incore(struct vnode * vp, daddr_t blkno) 2228{ 2229 struct buf *bp; 2230 2231 int s = splbio(); 2232 VI_LOCK(vp); 2233 bp = gbincore(vp, blkno); 2234 VI_UNLOCK(vp); 2235 splx(s); 2236 return (bp); 2237} 2238 2239/* 2240 * Returns true if no I/O is needed to access the 2241 * associated VM object. This is like incore except 2242 * it also hunts around in the VM system for the data. 2243 */ 2244 2245int 2246inmem(struct vnode * vp, daddr_t blkno) 2247{ 2248 vm_object_t obj; 2249 vm_offset_t toff, tinc, size; 2250 vm_page_t m; 2251 vm_ooffset_t off; 2252 2253 GIANT_REQUIRED; 2254 ASSERT_VOP_LOCKED(vp, "inmem"); 2255 2256 if (incore(vp, blkno)) 2257 return 1; 2258 if (vp->v_mount == NULL) 2259 return 0; 2260 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2261 return 0; 2262 2263 size = PAGE_SIZE; 2264 if (size > vp->v_mount->mnt_stat.f_iosize) 2265 size = vp->v_mount->mnt_stat.f_iosize; 2266 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2267 2268 VM_OBJECT_LOCK(obj); 2269 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2270 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2271 if (!m) 2272 goto notinmem; 2273 tinc = size; 2274 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2275 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2276 if (vm_page_is_valid(m, 2277 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2278 goto notinmem; 2279 } 2280 VM_OBJECT_UNLOCK(obj); 2281 return 1; 2282 2283notinmem: 2284 VM_OBJECT_UNLOCK(obj); 2285 return (0); 2286} 2287 2288/* 2289 * vfs_setdirty: 2290 * 2291 * Sets the dirty range for a buffer based on the status of the dirty 2292 * bits in the pages comprising the buffer. 2293 * 2294 * The range is limited to the size of the buffer. 2295 * 2296 * This routine is primarily used by NFS, but is generalized for the 2297 * B_VMIO case. 2298 */ 2299static void 2300vfs_setdirty(struct buf *bp) 2301{ 2302 int i; 2303 vm_object_t object; 2304 2305 GIANT_REQUIRED; 2306 /* 2307 * Degenerate case - empty buffer 2308 */ 2309 2310 if (bp->b_bufsize == 0) 2311 return; 2312 2313 /* 2314 * We qualify the scan for modified pages on whether the 2315 * object has been flushed yet. The OBJ_WRITEABLE flag 2316 * is not cleared simply by protecting pages off. 2317 */ 2318 2319 if ((bp->b_flags & B_VMIO) == 0) 2320 return; 2321 2322 object = bp->b_pages[0]->object; 2323 VM_OBJECT_LOCK(object); 2324 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2325 printf("Warning: object %p writeable but not mightbedirty\n", object); 2326 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2327 printf("Warning: object %p mightbedirty but not writeable\n", object); 2328 2329 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2330 vm_offset_t boffset; 2331 vm_offset_t eoffset; 2332 2333 vm_page_lock_queues(); 2334 /* 2335 * test the pages to see if they have been modified directly 2336 * by users through the VM system. 2337 */ 2338 for (i = 0; i < bp->b_npages; i++) 2339 vm_page_test_dirty(bp->b_pages[i]); 2340 2341 /* 2342 * Calculate the encompassing dirty range, boffset and eoffset, 2343 * (eoffset - boffset) bytes. 2344 */ 2345 2346 for (i = 0; i < bp->b_npages; i++) { 2347 if (bp->b_pages[i]->dirty) 2348 break; 2349 } 2350 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2351 2352 for (i = bp->b_npages - 1; i >= 0; --i) { 2353 if (bp->b_pages[i]->dirty) { 2354 break; 2355 } 2356 } 2357 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2358 2359 vm_page_unlock_queues(); 2360 /* 2361 * Fit it to the buffer. 2362 */ 2363 2364 if (eoffset > bp->b_bcount) 2365 eoffset = bp->b_bcount; 2366 2367 /* 2368 * If we have a good dirty range, merge with the existing 2369 * dirty range. 2370 */ 2371 2372 if (boffset < eoffset) { 2373 if (bp->b_dirtyoff > boffset) 2374 bp->b_dirtyoff = boffset; 2375 if (bp->b_dirtyend < eoffset) 2376 bp->b_dirtyend = eoffset; 2377 } 2378 } 2379 VM_OBJECT_UNLOCK(object); 2380} 2381 2382/* 2383 * getblk: 2384 * 2385 * Get a block given a specified block and offset into a file/device. 2386 * The buffers B_DONE bit will be cleared on return, making it almost 2387 * ready for an I/O initiation. B_INVAL may or may not be set on 2388 * return. The caller should clear B_INVAL prior to initiating a 2389 * READ. 2390 * 2391 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2392 * an existing buffer. 2393 * 2394 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2395 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2396 * and then cleared based on the backing VM. If the previous buffer is 2397 * non-0-sized but invalid, B_CACHE will be cleared. 2398 * 2399 * If getblk() must create a new buffer, the new buffer is returned with 2400 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2401 * case it is returned with B_INVAL clear and B_CACHE set based on the 2402 * backing VM. 2403 * 2404 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2405 * B_CACHE bit is clear. 2406 * 2407 * What this means, basically, is that the caller should use B_CACHE to 2408 * determine whether the buffer is fully valid or not and should clear 2409 * B_INVAL prior to issuing a read. If the caller intends to validate 2410 * the buffer by loading its data area with something, the caller needs 2411 * to clear B_INVAL. If the caller does this without issuing an I/O, 2412 * the caller should set B_CACHE ( as an optimization ), else the caller 2413 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2414 * a write attempt or if it was a successfull read. If the caller 2415 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2416 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2417 */ 2418struct buf * 2419getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2420 int flags) 2421{ 2422 struct buf *bp; 2423 int s; 2424 int error; 2425 ASSERT_VOP_LOCKED(vp, "getblk"); 2426 2427 if (size > MAXBSIZE) 2428 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2429 2430 s = splbio(); 2431loop: 2432 /* 2433 * Block if we are low on buffers. Certain processes are allowed 2434 * to completely exhaust the buffer cache. 2435 * 2436 * If this check ever becomes a bottleneck it may be better to 2437 * move it into the else, when gbincore() fails. At the moment 2438 * it isn't a problem. 2439 * 2440 * XXX remove if 0 sections (clean this up after its proven) 2441 */ 2442 if (numfreebuffers == 0) { 2443 if (curthread == PCPU_GET(idlethread)) 2444 return NULL; 2445 mtx_lock(&nblock); 2446 needsbuffer |= VFS_BIO_NEED_ANY; 2447 mtx_unlock(&nblock); 2448 } 2449 2450 VI_LOCK(vp); 2451 if ((bp = gbincore(vp, blkno))) { 2452 int lockflags; 2453 /* 2454 * Buffer is in-core. If the buffer is not busy, it must 2455 * be on a queue. 2456 */ 2457 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2458 2459 if (flags & GB_LOCK_NOWAIT) 2460 lockflags |= LK_NOWAIT; 2461 2462 error = BUF_TIMELOCK(bp, lockflags, 2463 VI_MTX(vp), "getblk", slpflag, slptimeo); 2464 2465 /* 2466 * If we slept and got the lock we have to restart in case 2467 * the buffer changed identities. 2468 */ 2469 if (error == ENOLCK) 2470 goto loop; 2471 /* We timed out or were interrupted. */ 2472 else if (error) 2473 return (NULL); 2474 2475 /* 2476 * The buffer is locked. B_CACHE is cleared if the buffer is 2477 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2478 * and for a VMIO buffer B_CACHE is adjusted according to the 2479 * backing VM cache. 2480 */ 2481 if (bp->b_flags & B_INVAL) 2482 bp->b_flags &= ~B_CACHE; 2483 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2484 bp->b_flags |= B_CACHE; 2485 bremfree(bp); 2486 2487 /* 2488 * check for size inconsistancies for non-VMIO case. 2489 */ 2490 2491 if (bp->b_bcount != size) { 2492 if ((bp->b_flags & B_VMIO) == 0 || 2493 (size > bp->b_kvasize)) { 2494 if (bp->b_flags & B_DELWRI) { 2495 bp->b_flags |= B_NOCACHE; 2496 bwrite(bp); 2497 } else { 2498 if ((bp->b_flags & B_VMIO) && 2499 (LIST_FIRST(&bp->b_dep) == NULL)) { 2500 bp->b_flags |= B_RELBUF; 2501 brelse(bp); 2502 } else { 2503 bp->b_flags |= B_NOCACHE; 2504 bwrite(bp); 2505 } 2506 } 2507 goto loop; 2508 } 2509 } 2510 2511 /* 2512 * If the size is inconsistant in the VMIO case, we can resize 2513 * the buffer. This might lead to B_CACHE getting set or 2514 * cleared. If the size has not changed, B_CACHE remains 2515 * unchanged from its previous state. 2516 */ 2517 2518 if (bp->b_bcount != size) 2519 allocbuf(bp, size); 2520 2521 KASSERT(bp->b_offset != NOOFFSET, 2522 ("getblk: no buffer offset")); 2523 2524 /* 2525 * A buffer with B_DELWRI set and B_CACHE clear must 2526 * be committed before we can return the buffer in 2527 * order to prevent the caller from issuing a read 2528 * ( due to B_CACHE not being set ) and overwriting 2529 * it. 2530 * 2531 * Most callers, including NFS and FFS, need this to 2532 * operate properly either because they assume they 2533 * can issue a read if B_CACHE is not set, or because 2534 * ( for example ) an uncached B_DELWRI might loop due 2535 * to softupdates re-dirtying the buffer. In the latter 2536 * case, B_CACHE is set after the first write completes, 2537 * preventing further loops. 2538 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2539 * above while extending the buffer, we cannot allow the 2540 * buffer to remain with B_CACHE set after the write 2541 * completes or it will represent a corrupt state. To 2542 * deal with this we set B_NOCACHE to scrap the buffer 2543 * after the write. 2544 * 2545 * We might be able to do something fancy, like setting 2546 * B_CACHE in bwrite() except if B_DELWRI is already set, 2547 * so the below call doesn't set B_CACHE, but that gets real 2548 * confusing. This is much easier. 2549 */ 2550 2551 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2552 bp->b_flags |= B_NOCACHE; 2553 bwrite(bp); 2554 goto loop; 2555 } 2556 2557 splx(s); 2558 bp->b_flags &= ~B_DONE; 2559 } else { 2560 int bsize, maxsize, vmio; 2561 off_t offset; 2562 2563 /* 2564 * Buffer is not in-core, create new buffer. The buffer 2565 * returned by getnewbuf() is locked. Note that the returned 2566 * buffer is also considered valid (not marked B_INVAL). 2567 */ 2568 VI_UNLOCK(vp); 2569 /* 2570 * If the user does not want us to create the buffer, bail out 2571 * here. 2572 */ 2573 if (flags & GB_NOCREAT) { 2574 splx(s); 2575 return NULL; 2576 } 2577 if (vn_isdisk(vp, NULL)) 2578 bsize = DEV_BSIZE; 2579 else if (vp->v_mountedhere) 2580 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2581 else if (vp->v_mount) 2582 bsize = vp->v_mount->mnt_stat.f_iosize; 2583 else 2584 bsize = size; 2585 2586 if (vp->v_bsize != bsize) { 2587#if 0 2588 printf("WARNING: Wrong block size on vnode: %d should be %d\n", vp->v_bsize, bsize); 2589#endif 2590 vp->v_bsize = bsize; 2591 } 2592 2593 offset = blkno * bsize; 2594 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2595 (vp->v_vflag & VV_OBJBUF); 2596 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2597 maxsize = imax(maxsize, bsize); 2598 2599 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2600 if (slpflag || slptimeo) { 2601 splx(s); 2602 return NULL; 2603 } 2604 goto loop; 2605 } 2606 2607 /* 2608 * This code is used to make sure that a buffer is not 2609 * created while the getnewbuf routine is blocked. 2610 * This can be a problem whether the vnode is locked or not. 2611 * If the buffer is created out from under us, we have to 2612 * throw away the one we just created. There is now window 2613 * race because we are safely running at splbio() from the 2614 * point of the duplicate buffer creation through to here, 2615 * and we've locked the buffer. 2616 * 2617 * Note: this must occur before we associate the buffer 2618 * with the vp especially considering limitations in 2619 * the splay tree implementation when dealing with duplicate 2620 * lblkno's. 2621 */ 2622 VI_LOCK(vp); 2623 if (gbincore(vp, blkno)) { 2624 VI_UNLOCK(vp); 2625 bp->b_flags |= B_INVAL; 2626 brelse(bp); 2627 goto loop; 2628 } 2629 2630 /* 2631 * Insert the buffer into the hash, so that it can 2632 * be found by incore. 2633 */ 2634 bp->b_blkno = bp->b_lblkno = blkno; 2635 bp->b_offset = offset; 2636 2637 bgetvp(vp, bp); 2638 VI_UNLOCK(vp); 2639 2640 /* 2641 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2642 * buffer size starts out as 0, B_CACHE will be set by 2643 * allocbuf() for the VMIO case prior to it testing the 2644 * backing store for validity. 2645 */ 2646 2647 if (vmio) { 2648 bp->b_flags |= B_VMIO; 2649#if defined(VFS_BIO_DEBUG) 2650 if (vn_canvmio(vp) != TRUE) 2651 printf("getblk: VMIO on vnode type %d\n", 2652 vp->v_type); 2653#endif 2654 VOP_GETVOBJECT(vp, &bp->b_object); 2655 } else { 2656 bp->b_flags &= ~B_VMIO; 2657 bp->b_object = NULL; 2658 } 2659 2660 allocbuf(bp, size); 2661 2662 splx(s); 2663 bp->b_flags &= ~B_DONE; 2664 } 2665 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2666 return (bp); 2667} 2668 2669/* 2670 * Get an empty, disassociated buffer of given size. The buffer is initially 2671 * set to B_INVAL. 2672 */ 2673struct buf * 2674geteblk(int size) 2675{ 2676 struct buf *bp; 2677 int s; 2678 int maxsize; 2679 2680 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2681 2682 s = splbio(); 2683 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2684 continue; 2685 splx(s); 2686 allocbuf(bp, size); 2687 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2688 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2689 return (bp); 2690} 2691 2692 2693/* 2694 * This code constitutes the buffer memory from either anonymous system 2695 * memory (in the case of non-VMIO operations) or from an associated 2696 * VM object (in the case of VMIO operations). This code is able to 2697 * resize a buffer up or down. 2698 * 2699 * Note that this code is tricky, and has many complications to resolve 2700 * deadlock or inconsistant data situations. Tread lightly!!! 2701 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2702 * the caller. Calling this code willy nilly can result in the loss of data. 2703 * 2704 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2705 * B_CACHE for the non-VMIO case. 2706 */ 2707 2708int 2709allocbuf(struct buf *bp, int size) 2710{ 2711 int newbsize, mbsize; 2712 int i; 2713 2714 GIANT_REQUIRED; 2715 2716 if (BUF_REFCNT(bp) == 0) 2717 panic("allocbuf: buffer not busy"); 2718 2719 if (bp->b_kvasize < size) 2720 panic("allocbuf: buffer too small"); 2721 2722 if ((bp->b_flags & B_VMIO) == 0) { 2723 caddr_t origbuf; 2724 int origbufsize; 2725 /* 2726 * Just get anonymous memory from the kernel. Don't 2727 * mess with B_CACHE. 2728 */ 2729 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2730 if (bp->b_flags & B_MALLOC) 2731 newbsize = mbsize; 2732 else 2733 newbsize = round_page(size); 2734 2735 if (newbsize < bp->b_bufsize) { 2736 /* 2737 * malloced buffers are not shrunk 2738 */ 2739 if (bp->b_flags & B_MALLOC) { 2740 if (newbsize) { 2741 bp->b_bcount = size; 2742 } else { 2743 free(bp->b_data, M_BIOBUF); 2744 if (bp->b_bufsize) { 2745 atomic_subtract_int( 2746 &bufmallocspace, 2747 bp->b_bufsize); 2748 bufspacewakeup(); 2749 bp->b_bufsize = 0; 2750 } 2751 bp->b_saveaddr = bp->b_kvabase; 2752 bp->b_data = bp->b_saveaddr; 2753 bp->b_bcount = 0; 2754 bp->b_flags &= ~B_MALLOC; 2755 } 2756 return 1; 2757 } 2758 vm_hold_free_pages( 2759 bp, 2760 (vm_offset_t) bp->b_data + newbsize, 2761 (vm_offset_t) bp->b_data + bp->b_bufsize); 2762 } else if (newbsize > bp->b_bufsize) { 2763 /* 2764 * We only use malloced memory on the first allocation. 2765 * and revert to page-allocated memory when the buffer 2766 * grows. 2767 */ 2768 /* 2769 * There is a potential smp race here that could lead 2770 * to bufmallocspace slightly passing the max. It 2771 * is probably extremely rare and not worth worrying 2772 * over. 2773 */ 2774 if ( (bufmallocspace < maxbufmallocspace) && 2775 (bp->b_bufsize == 0) && 2776 (mbsize <= PAGE_SIZE/2)) { 2777 2778 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2779 bp->b_bufsize = mbsize; 2780 bp->b_bcount = size; 2781 bp->b_flags |= B_MALLOC; 2782 atomic_add_int(&bufmallocspace, mbsize); 2783 return 1; 2784 } 2785 origbuf = NULL; 2786 origbufsize = 0; 2787 /* 2788 * If the buffer is growing on its other-than-first allocation, 2789 * then we revert to the page-allocation scheme. 2790 */ 2791 if (bp->b_flags & B_MALLOC) { 2792 origbuf = bp->b_data; 2793 origbufsize = bp->b_bufsize; 2794 bp->b_data = bp->b_kvabase; 2795 if (bp->b_bufsize) { 2796 atomic_subtract_int(&bufmallocspace, 2797 bp->b_bufsize); 2798 bufspacewakeup(); 2799 bp->b_bufsize = 0; 2800 } 2801 bp->b_flags &= ~B_MALLOC; 2802 newbsize = round_page(newbsize); 2803 } 2804 vm_hold_load_pages( 2805 bp, 2806 (vm_offset_t) bp->b_data + bp->b_bufsize, 2807 (vm_offset_t) bp->b_data + newbsize); 2808 if (origbuf) { 2809 bcopy(origbuf, bp->b_data, origbufsize); 2810 free(origbuf, M_BIOBUF); 2811 } 2812 } 2813 } else { 2814 int desiredpages; 2815 2816 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2817 desiredpages = (size == 0) ? 0 : 2818 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2819 2820 if (bp->b_flags & B_MALLOC) 2821 panic("allocbuf: VMIO buffer can't be malloced"); 2822 /* 2823 * Set B_CACHE initially if buffer is 0 length or will become 2824 * 0-length. 2825 */ 2826 if (size == 0 || bp->b_bufsize == 0) 2827 bp->b_flags |= B_CACHE; 2828 2829 if (newbsize < bp->b_bufsize) { 2830 /* 2831 * DEV_BSIZE aligned new buffer size is less then the 2832 * DEV_BSIZE aligned existing buffer size. Figure out 2833 * if we have to remove any pages. 2834 */ 2835 if (desiredpages < bp->b_npages) { 2836 vm_page_t m; 2837 2838 vm_page_lock_queues(); 2839 for (i = desiredpages; i < bp->b_npages; i++) { 2840 /* 2841 * the page is not freed here -- it 2842 * is the responsibility of 2843 * vnode_pager_setsize 2844 */ 2845 m = bp->b_pages[i]; 2846 KASSERT(m != bogus_page, 2847 ("allocbuf: bogus page found")); 2848 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2849 vm_page_lock_queues(); 2850 2851 bp->b_pages[i] = NULL; 2852 vm_page_unwire(m, 0); 2853 } 2854 vm_page_unlock_queues(); 2855 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2856 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2857 bp->b_npages = desiredpages; 2858 } 2859 } else if (size > bp->b_bcount) { 2860 /* 2861 * We are growing the buffer, possibly in a 2862 * byte-granular fashion. 2863 */ 2864 struct vnode *vp; 2865 vm_object_t obj; 2866 vm_offset_t toff; 2867 vm_offset_t tinc; 2868 2869 /* 2870 * Step 1, bring in the VM pages from the object, 2871 * allocating them if necessary. We must clear 2872 * B_CACHE if these pages are not valid for the 2873 * range covered by the buffer. 2874 */ 2875 2876 vp = bp->b_vp; 2877 obj = bp->b_object; 2878 2879 VM_OBJECT_LOCK(obj); 2880 while (bp->b_npages < desiredpages) { 2881 vm_page_t m; 2882 vm_pindex_t pi; 2883 2884 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2885 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2886 /* 2887 * note: must allocate system pages 2888 * since blocking here could intefere 2889 * with paging I/O, no matter which 2890 * process we are. 2891 */ 2892 m = vm_page_alloc(obj, pi, 2893 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2894 if (m == NULL) { 2895 atomic_add_int(&vm_pageout_deficit, 2896 desiredpages - bp->b_npages); 2897 VM_OBJECT_UNLOCK(obj); 2898 VM_WAIT; 2899 VM_OBJECT_LOCK(obj); 2900 } else { 2901 vm_page_lock_queues(); 2902 vm_page_wakeup(m); 2903 vm_page_unlock_queues(); 2904 bp->b_flags &= ~B_CACHE; 2905 bp->b_pages[bp->b_npages] = m; 2906 ++bp->b_npages; 2907 } 2908 continue; 2909 } 2910 2911 /* 2912 * We found a page. If we have to sleep on it, 2913 * retry because it might have gotten freed out 2914 * from under us. 2915 * 2916 * We can only test PG_BUSY here. Blocking on 2917 * m->busy might lead to a deadlock: 2918 * 2919 * vm_fault->getpages->cluster_read->allocbuf 2920 * 2921 */ 2922 vm_page_lock_queues(); 2923 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2924 continue; 2925 2926 /* 2927 * We have a good page. Should we wakeup the 2928 * page daemon? 2929 */ 2930 if ((curproc != pageproc) && 2931 ((m->queue - m->pc) == PQ_CACHE) && 2932 ((cnt.v_free_count + cnt.v_cache_count) < 2933 (cnt.v_free_min + cnt.v_cache_min))) { 2934 pagedaemon_wakeup(); 2935 } 2936 vm_page_wire(m); 2937 vm_page_unlock_queues(); 2938 bp->b_pages[bp->b_npages] = m; 2939 ++bp->b_npages; 2940 } 2941 2942 /* 2943 * Step 2. We've loaded the pages into the buffer, 2944 * we have to figure out if we can still have B_CACHE 2945 * set. Note that B_CACHE is set according to the 2946 * byte-granular range ( bcount and size ), new the 2947 * aligned range ( newbsize ). 2948 * 2949 * The VM test is against m->valid, which is DEV_BSIZE 2950 * aligned. Needless to say, the validity of the data 2951 * needs to also be DEV_BSIZE aligned. Note that this 2952 * fails with NFS if the server or some other client 2953 * extends the file's EOF. If our buffer is resized, 2954 * B_CACHE may remain set! XXX 2955 */ 2956 2957 toff = bp->b_bcount; 2958 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2959 2960 while ((bp->b_flags & B_CACHE) && toff < size) { 2961 vm_pindex_t pi; 2962 2963 if (tinc > (size - toff)) 2964 tinc = size - toff; 2965 2966 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2967 PAGE_SHIFT; 2968 2969 vfs_buf_test_cache( 2970 bp, 2971 bp->b_offset, 2972 toff, 2973 tinc, 2974 bp->b_pages[pi] 2975 ); 2976 toff += tinc; 2977 tinc = PAGE_SIZE; 2978 } 2979 VM_OBJECT_UNLOCK(obj); 2980 2981 /* 2982 * Step 3, fixup the KVM pmap. Remember that 2983 * bp->b_data is relative to bp->b_offset, but 2984 * bp->b_offset may be offset into the first page. 2985 */ 2986 2987 bp->b_data = (caddr_t) 2988 trunc_page((vm_offset_t)bp->b_data); 2989 pmap_qenter( 2990 (vm_offset_t)bp->b_data, 2991 bp->b_pages, 2992 bp->b_npages 2993 ); 2994 2995 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2996 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2997 } 2998 } 2999 if (newbsize < bp->b_bufsize) 3000 bufspacewakeup(); 3001 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3002 bp->b_bcount = size; /* requested buffer size */ 3003 return 1; 3004} 3005 3006void 3007biodone(struct bio *bp) 3008{ 3009 3010 mtx_lock(&bdonelock); 3011 bp->bio_flags |= BIO_DONE; 3012 if (bp->bio_done == NULL) 3013 wakeup(bp); 3014 mtx_unlock(&bdonelock); 3015 if (bp->bio_done != NULL) 3016 bp->bio_done(bp); 3017} 3018 3019/* 3020 * Wait for a BIO to finish. 3021 * 3022 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3023 * case is not yet clear. 3024 */ 3025int 3026biowait(struct bio *bp, const char *wchan) 3027{ 3028 3029 mtx_lock(&bdonelock); 3030 while ((bp->bio_flags & BIO_DONE) == 0) 3031 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3032 mtx_unlock(&bdonelock); 3033 if (bp->bio_error != 0) 3034 return (bp->bio_error); 3035 if (!(bp->bio_flags & BIO_ERROR)) 3036 return (0); 3037 return (EIO); 3038} 3039 3040void 3041biofinish(struct bio *bp, struct devstat *stat, int error) 3042{ 3043 3044 if (error) { 3045 bp->bio_error = error; 3046 bp->bio_flags |= BIO_ERROR; 3047 } 3048 if (stat != NULL) 3049 devstat_end_transaction_bio(stat, bp); 3050 biodone(bp); 3051} 3052 3053/* 3054 * bufwait: 3055 * 3056 * Wait for buffer I/O completion, returning error status. The buffer 3057 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3058 * error and cleared. 3059 */ 3060int 3061bufwait(struct buf *bp) 3062{ 3063 int s; 3064 3065 s = splbio(); 3066 if (bp->b_iocmd == BIO_READ) 3067 bwait(bp, PRIBIO, "biord"); 3068 else 3069 bwait(bp, PRIBIO, "biowr"); 3070 splx(s); 3071 if (bp->b_flags & B_EINTR) { 3072 bp->b_flags &= ~B_EINTR; 3073 return (EINTR); 3074 } 3075 if (bp->b_ioflags & BIO_ERROR) { 3076 return (bp->b_error ? bp->b_error : EIO); 3077 } else { 3078 return (0); 3079 } 3080} 3081 3082 /* 3083 * Call back function from struct bio back up to struct buf. 3084 */ 3085static void 3086bufdonebio(struct bio *bp) 3087{ 3088 3089 /* Device drivers may or may not hold giant, hold it here. */ 3090 mtx_lock(&Giant); 3091 bufdone(bp->bio_caller2); 3092 mtx_unlock(&Giant); 3093} 3094 3095void 3096dev_strategy(struct buf *bp) 3097{ 3098 struct cdevsw *csw; 3099 struct cdev *dev; 3100 3101 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3102 panic("b_iocmd botch"); 3103 bp->b_io.bio_done = bufdonebio; 3104 bp->b_io.bio_caller2 = bp; 3105 dev = bp->b_io.bio_dev; 3106 KASSERT(dev->si_refcount > 0, 3107 ("dev_strategy on un-referenced struct cdev *(%s)", 3108 devtoname(dev))); 3109 csw = dev_refthread(dev); 3110 if (csw == NULL) { 3111 bp->b_error = ENXIO; 3112 bp->b_ioflags = BIO_ERROR; 3113 mtx_lock(&Giant); /* XXX: too defensive ? */ 3114 bufdone(bp); 3115 mtx_unlock(&Giant); /* XXX: too defensive ? */ 3116 return; 3117 } 3118 (*csw->d_strategy)(&bp->b_io); 3119 dev_relthread(dev); 3120} 3121 3122/* 3123 * bufdone: 3124 * 3125 * Finish I/O on a buffer, optionally calling a completion function. 3126 * This is usually called from an interrupt so process blocking is 3127 * not allowed. 3128 * 3129 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3130 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3131 * assuming B_INVAL is clear. 3132 * 3133 * For the VMIO case, we set B_CACHE if the op was a read and no 3134 * read error occured, or if the op was a write. B_CACHE is never 3135 * set if the buffer is invalid or otherwise uncacheable. 3136 * 3137 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3138 * initiator to leave B_INVAL set to brelse the buffer out of existance 3139 * in the biodone routine. 3140 */ 3141void 3142bufdone(struct buf *bp) 3143{ 3144 int s; 3145 void (*biodone)(struct buf *); 3146 3147 3148 s = splbio(); 3149 3150 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3151 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3152 3153 bp->b_flags |= B_DONE; 3154 runningbufwakeup(bp); 3155 3156 if (bp->b_iocmd == BIO_WRITE && bp->b_vp != NULL) 3157 bufobj_wdrop(&bp->b_vp->v_bufobj); 3158 3159 /* call optional completion function if requested */ 3160 if (bp->b_iodone != NULL) { 3161 biodone = bp->b_iodone; 3162 bp->b_iodone = NULL; 3163 (*biodone) (bp); 3164 splx(s); 3165 return; 3166 } 3167 if (LIST_FIRST(&bp->b_dep) != NULL) 3168 buf_complete(bp); 3169 3170 if (bp->b_flags & B_VMIO) { 3171 int i; 3172 vm_ooffset_t foff; 3173 vm_page_t m; 3174 vm_object_t obj; 3175 int iosize; 3176 struct vnode *vp = bp->b_vp; 3177 3178 obj = bp->b_object; 3179 3180#if defined(VFS_BIO_DEBUG) 3181 mp_fixme("usecount and vflag accessed without locks."); 3182 if (vp->v_usecount == 0) { 3183 panic("biodone: zero vnode ref count"); 3184 } 3185 3186 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3187 panic("biodone: vnode is not setup for merged cache"); 3188 } 3189#endif 3190 3191 foff = bp->b_offset; 3192 KASSERT(bp->b_offset != NOOFFSET, 3193 ("biodone: no buffer offset")); 3194 3195 VM_OBJECT_LOCK(obj); 3196#if defined(VFS_BIO_DEBUG) 3197 if (obj->paging_in_progress < bp->b_npages) { 3198 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3199 obj->paging_in_progress, bp->b_npages); 3200 } 3201#endif 3202 3203 /* 3204 * Set B_CACHE if the op was a normal read and no error 3205 * occured. B_CACHE is set for writes in the b*write() 3206 * routines. 3207 */ 3208 iosize = bp->b_bcount - bp->b_resid; 3209 if (bp->b_iocmd == BIO_READ && 3210 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3211 !(bp->b_ioflags & BIO_ERROR)) { 3212 bp->b_flags |= B_CACHE; 3213 } 3214 vm_page_lock_queues(); 3215 for (i = 0; i < bp->b_npages; i++) { 3216 int bogusflag = 0; 3217 int resid; 3218 3219 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3220 if (resid > iosize) 3221 resid = iosize; 3222 3223 /* 3224 * cleanup bogus pages, restoring the originals 3225 */ 3226 m = bp->b_pages[i]; 3227 if (m == bogus_page) { 3228 bogusflag = 1; 3229 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3230 if (m == NULL) 3231 panic("biodone: page disappeared!"); 3232 bp->b_pages[i] = m; 3233 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3234 } 3235#if defined(VFS_BIO_DEBUG) 3236 if (OFF_TO_IDX(foff) != m->pindex) { 3237 printf( 3238"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3239 (intmax_t)foff, (uintmax_t)m->pindex); 3240 } 3241#endif 3242 3243 /* 3244 * In the write case, the valid and clean bits are 3245 * already changed correctly ( see bdwrite() ), so we 3246 * only need to do this here in the read case. 3247 */ 3248 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3249 vfs_page_set_valid(bp, foff, i, m); 3250 } 3251 3252 /* 3253 * when debugging new filesystems or buffer I/O methods, this 3254 * is the most common error that pops up. if you see this, you 3255 * have not set the page busy flag correctly!!! 3256 */ 3257 if (m->busy == 0) { 3258 printf("biodone: page busy < 0, " 3259 "pindex: %d, foff: 0x(%x,%x), " 3260 "resid: %d, index: %d\n", 3261 (int) m->pindex, (int)(foff >> 32), 3262 (int) foff & 0xffffffff, resid, i); 3263 if (!vn_isdisk(vp, NULL)) 3264 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3265 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3266 (intmax_t) bp->b_lblkno, 3267 bp->b_flags, bp->b_npages); 3268 else 3269 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3270 (intmax_t) bp->b_lblkno, 3271 bp->b_flags, bp->b_npages); 3272 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3273 (u_long)m->valid, (u_long)m->dirty, 3274 m->wire_count); 3275 panic("biodone: page busy < 0\n"); 3276 } 3277 vm_page_io_finish(m); 3278 vm_object_pip_subtract(obj, 1); 3279 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3280 iosize -= resid; 3281 } 3282 vm_page_unlock_queues(); 3283 vm_object_pip_wakeupn(obj, 0); 3284 VM_OBJECT_UNLOCK(obj); 3285 } 3286 3287 /* 3288 * For asynchronous completions, release the buffer now. The brelse 3289 * will do a wakeup there if necessary - so no need to do a wakeup 3290 * here in the async case. The sync case always needs to do a wakeup. 3291 */ 3292 3293 if (bp->b_flags & B_ASYNC) { 3294 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3295 brelse(bp); 3296 else 3297 bqrelse(bp); 3298 } else { 3299 bdone(bp); 3300 } 3301 splx(s); 3302} 3303 3304/* 3305 * This routine is called in lieu of iodone in the case of 3306 * incomplete I/O. This keeps the busy status for pages 3307 * consistant. 3308 */ 3309void 3310vfs_unbusy_pages(struct buf *bp) 3311{ 3312 int i; 3313 vm_object_t obj; 3314 vm_page_t m; 3315 3316 runningbufwakeup(bp); 3317 if (!(bp->b_flags & B_VMIO)) 3318 return; 3319 3320 obj = bp->b_object; 3321 VM_OBJECT_LOCK(obj); 3322 vm_page_lock_queues(); 3323 for (i = 0; i < bp->b_npages; i++) { 3324 m = bp->b_pages[i]; 3325 if (m == bogus_page) { 3326 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3327 if (!m) { 3328 panic("vfs_unbusy_pages: page missing\n"); 3329 } 3330 bp->b_pages[i] = m; 3331 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3332 } 3333 vm_object_pip_subtract(obj, 1); 3334 vm_page_io_finish(m); 3335 } 3336 vm_page_unlock_queues(); 3337 vm_object_pip_wakeupn(obj, 0); 3338 VM_OBJECT_UNLOCK(obj); 3339} 3340 3341/* 3342 * vfs_page_set_valid: 3343 * 3344 * Set the valid bits in a page based on the supplied offset. The 3345 * range is restricted to the buffer's size. 3346 * 3347 * This routine is typically called after a read completes. 3348 */ 3349static void 3350vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3351{ 3352 vm_ooffset_t soff, eoff; 3353 3354 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3355 /* 3356 * Start and end offsets in buffer. eoff - soff may not cross a 3357 * page boundry or cross the end of the buffer. The end of the 3358 * buffer, in this case, is our file EOF, not the allocation size 3359 * of the buffer. 3360 */ 3361 soff = off; 3362 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3363 if (eoff > bp->b_offset + bp->b_bcount) 3364 eoff = bp->b_offset + bp->b_bcount; 3365 3366 /* 3367 * Set valid range. This is typically the entire buffer and thus the 3368 * entire page. 3369 */ 3370 if (eoff > soff) { 3371 vm_page_set_validclean( 3372 m, 3373 (vm_offset_t) (soff & PAGE_MASK), 3374 (vm_offset_t) (eoff - soff) 3375 ); 3376 } 3377} 3378 3379/* 3380 * This routine is called before a device strategy routine. 3381 * It is used to tell the VM system that paging I/O is in 3382 * progress, and treat the pages associated with the buffer 3383 * almost as being PG_BUSY. Also the object paging_in_progress 3384 * flag is handled to make sure that the object doesn't become 3385 * inconsistant. 3386 * 3387 * Since I/O has not been initiated yet, certain buffer flags 3388 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3389 * and should be ignored. 3390 */ 3391void 3392vfs_busy_pages(struct buf *bp, int clear_modify) 3393{ 3394 int i, bogus; 3395 vm_object_t obj; 3396 vm_ooffset_t foff; 3397 vm_page_t m; 3398 3399 if (!(bp->b_flags & B_VMIO)) 3400 return; 3401 3402 obj = bp->b_object; 3403 foff = bp->b_offset; 3404 KASSERT(bp->b_offset != NOOFFSET, 3405 ("vfs_busy_pages: no buffer offset")); 3406 vfs_setdirty(bp); 3407 VM_OBJECT_LOCK(obj); 3408retry: 3409 vm_page_lock_queues(); 3410 for (i = 0; i < bp->b_npages; i++) { 3411 m = bp->b_pages[i]; 3412 3413 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3414 goto retry; 3415 } 3416 bogus = 0; 3417 for (i = 0; i < bp->b_npages; i++) { 3418 m = bp->b_pages[i]; 3419 3420 if ((bp->b_flags & B_CLUSTER) == 0) { 3421 vm_object_pip_add(obj, 1); 3422 vm_page_io_start(m); 3423 } 3424 /* 3425 * When readying a buffer for a read ( i.e 3426 * clear_modify == 0 ), it is important to do 3427 * bogus_page replacement for valid pages in 3428 * partially instantiated buffers. Partially 3429 * instantiated buffers can, in turn, occur when 3430 * reconstituting a buffer from its VM backing store 3431 * base. We only have to do this if B_CACHE is 3432 * clear ( which causes the I/O to occur in the 3433 * first place ). The replacement prevents the read 3434 * I/O from overwriting potentially dirty VM-backed 3435 * pages. XXX bogus page replacement is, uh, bogus. 3436 * It may not work properly with small-block devices. 3437 * We need to find a better way. 3438 */ 3439 pmap_remove_all(m); 3440 if (clear_modify) 3441 vfs_page_set_valid(bp, foff, i, m); 3442 else if (m->valid == VM_PAGE_BITS_ALL && 3443 (bp->b_flags & B_CACHE) == 0) { 3444 bp->b_pages[i] = bogus_page; 3445 bogus++; 3446 } 3447 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3448 } 3449 vm_page_unlock_queues(); 3450 VM_OBJECT_UNLOCK(obj); 3451 if (bogus) 3452 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3453 bp->b_pages, bp->b_npages); 3454} 3455 3456/* 3457 * Tell the VM system that the pages associated with this buffer 3458 * are clean. This is used for delayed writes where the data is 3459 * going to go to disk eventually without additional VM intevention. 3460 * 3461 * Note that while we only really need to clean through to b_bcount, we 3462 * just go ahead and clean through to b_bufsize. 3463 */ 3464static void 3465vfs_clean_pages(struct buf *bp) 3466{ 3467 int i; 3468 vm_ooffset_t foff, noff, eoff; 3469 vm_page_t m; 3470 3471 if (!(bp->b_flags & B_VMIO)) 3472 return; 3473 3474 foff = bp->b_offset; 3475 KASSERT(bp->b_offset != NOOFFSET, 3476 ("vfs_clean_pages: no buffer offset")); 3477 VM_OBJECT_LOCK(bp->b_object); 3478 vm_page_lock_queues(); 3479 for (i = 0; i < bp->b_npages; i++) { 3480 m = bp->b_pages[i]; 3481 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3482 eoff = noff; 3483 3484 if (eoff > bp->b_offset + bp->b_bufsize) 3485 eoff = bp->b_offset + bp->b_bufsize; 3486 vfs_page_set_valid(bp, foff, i, m); 3487 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3488 foff = noff; 3489 } 3490 vm_page_unlock_queues(); 3491 VM_OBJECT_UNLOCK(bp->b_object); 3492} 3493 3494/* 3495 * vfs_bio_set_validclean: 3496 * 3497 * Set the range within the buffer to valid and clean. The range is 3498 * relative to the beginning of the buffer, b_offset. Note that b_offset 3499 * itself may be offset from the beginning of the first page. 3500 * 3501 */ 3502 3503void 3504vfs_bio_set_validclean(struct buf *bp, int base, int size) 3505{ 3506 int i, n; 3507 vm_page_t m; 3508 3509 if (!(bp->b_flags & B_VMIO)) 3510 return; 3511 3512 /* 3513 * Fixup base to be relative to beginning of first page. 3514 * Set initial n to be the maximum number of bytes in the 3515 * first page that can be validated. 3516 */ 3517 3518 base += (bp->b_offset & PAGE_MASK); 3519 n = PAGE_SIZE - (base & PAGE_MASK); 3520 3521 VM_OBJECT_LOCK(bp->b_object); 3522 vm_page_lock_queues(); 3523 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3524 m = bp->b_pages[i]; 3525 3526 if (n > size) 3527 n = size; 3528 3529 vm_page_set_validclean(m, base & PAGE_MASK, n); 3530 base += n; 3531 size -= n; 3532 n = PAGE_SIZE; 3533 } 3534 vm_page_unlock_queues(); 3535 VM_OBJECT_UNLOCK(bp->b_object); 3536} 3537 3538/* 3539 * vfs_bio_clrbuf: 3540 * 3541 * clear a buffer. This routine essentially fakes an I/O, so we need 3542 * to clear BIO_ERROR and B_INVAL. 3543 * 3544 * Note that while we only theoretically need to clear through b_bcount, 3545 * we go ahead and clear through b_bufsize. 3546 */ 3547 3548void 3549vfs_bio_clrbuf(struct buf *bp) 3550{ 3551 int i, j, mask = 0; 3552 caddr_t sa, ea; 3553 3554 GIANT_REQUIRED; 3555 3556 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3557 clrbuf(bp); 3558 return; 3559 } 3560 bp->b_flags &= ~B_INVAL; 3561 bp->b_ioflags &= ~BIO_ERROR; 3562 VM_OBJECT_LOCK(bp->b_object); 3563 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3564 (bp->b_offset & PAGE_MASK) == 0) { 3565 if (bp->b_pages[0] == bogus_page) 3566 goto unlock; 3567 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3568 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3569 if ((bp->b_pages[0]->valid & mask) == mask) 3570 goto unlock; 3571 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3572 ((bp->b_pages[0]->valid & mask) == 0)) { 3573 bzero(bp->b_data, bp->b_bufsize); 3574 bp->b_pages[0]->valid |= mask; 3575 goto unlock; 3576 } 3577 } 3578 ea = sa = bp->b_data; 3579 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3580 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3581 ea = (caddr_t)(vm_offset_t)ulmin( 3582 (u_long)(vm_offset_t)ea, 3583 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3584 if (bp->b_pages[i] == bogus_page) 3585 continue; 3586 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3587 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3588 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3589 if ((bp->b_pages[i]->valid & mask) == mask) 3590 continue; 3591 if ((bp->b_pages[i]->valid & mask) == 0) { 3592 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3593 bzero(sa, ea - sa); 3594 } else { 3595 for (; sa < ea; sa += DEV_BSIZE, j++) { 3596 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3597 (bp->b_pages[i]->valid & (1<<j)) == 0) 3598 bzero(sa, DEV_BSIZE); 3599 } 3600 } 3601 bp->b_pages[i]->valid |= mask; 3602 } 3603unlock: 3604 VM_OBJECT_UNLOCK(bp->b_object); 3605 bp->b_resid = 0; 3606} 3607 3608/* 3609 * vm_hold_load_pages and vm_hold_free_pages get pages into 3610 * a buffers address space. The pages are anonymous and are 3611 * not associated with a file object. 3612 */ 3613static void 3614vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3615{ 3616 vm_offset_t pg; 3617 vm_page_t p; 3618 int index; 3619 3620 to = round_page(to); 3621 from = round_page(from); 3622 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3623 3624 VM_OBJECT_LOCK(kernel_object); 3625 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3626tryagain: 3627 /* 3628 * note: must allocate system pages since blocking here 3629 * could intefere with paging I/O, no matter which 3630 * process we are. 3631 */ 3632 p = vm_page_alloc(kernel_object, 3633 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3634 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3635 if (!p) { 3636 atomic_add_int(&vm_pageout_deficit, 3637 (to - pg) >> PAGE_SHIFT); 3638 VM_OBJECT_UNLOCK(kernel_object); 3639 VM_WAIT; 3640 VM_OBJECT_LOCK(kernel_object); 3641 goto tryagain; 3642 } 3643 p->valid = VM_PAGE_BITS_ALL; 3644 pmap_qenter(pg, &p, 1); 3645 bp->b_pages[index] = p; 3646 vm_page_lock_queues(); 3647 vm_page_wakeup(p); 3648 vm_page_unlock_queues(); 3649 } 3650 VM_OBJECT_UNLOCK(kernel_object); 3651 bp->b_npages = index; 3652} 3653 3654/* Return pages associated with this buf to the vm system */ 3655static void 3656vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3657{ 3658 vm_offset_t pg; 3659 vm_page_t p; 3660 int index, newnpages; 3661 3662 GIANT_REQUIRED; 3663 3664 from = round_page(from); 3665 to = round_page(to); 3666 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3667 3668 VM_OBJECT_LOCK(kernel_object); 3669 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3670 p = bp->b_pages[index]; 3671 if (p && (index < bp->b_npages)) { 3672 if (p->busy) { 3673 printf( 3674 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3675 (intmax_t)bp->b_blkno, 3676 (intmax_t)bp->b_lblkno); 3677 } 3678 bp->b_pages[index] = NULL; 3679 pmap_qremove(pg, 1); 3680 vm_page_lock_queues(); 3681 vm_page_busy(p); 3682 vm_page_unwire(p, 0); 3683 vm_page_free(p); 3684 vm_page_unlock_queues(); 3685 } 3686 } 3687 VM_OBJECT_UNLOCK(kernel_object); 3688 bp->b_npages = newnpages; 3689} 3690 3691/* 3692 * Map an IO request into kernel virtual address space. 3693 * 3694 * All requests are (re)mapped into kernel VA space. 3695 * Notice that we use b_bufsize for the size of the buffer 3696 * to be mapped. b_bcount might be modified by the driver. 3697 * 3698 * Note that even if the caller determines that the address space should 3699 * be valid, a race or a smaller-file mapped into a larger space may 3700 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3701 * check the return value. 3702 */ 3703int 3704vmapbuf(struct buf *bp) 3705{ 3706 caddr_t addr, kva; 3707 vm_prot_t prot; 3708 int pidx, i; 3709 struct vm_page *m; 3710 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3711 3712 if (bp->b_bufsize < 0) 3713 return (-1); 3714 prot = VM_PROT_READ; 3715 if (bp->b_iocmd == BIO_READ) 3716 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3717 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3718 addr < bp->b_data + bp->b_bufsize; 3719 addr += PAGE_SIZE, pidx++) { 3720 /* 3721 * Do the vm_fault if needed; do the copy-on-write thing 3722 * when reading stuff off device into memory. 3723 * 3724 * NOTE! Must use pmap_extract() because addr may be in 3725 * the userland address space, and kextract is only guarenteed 3726 * to work for the kernland address space (see: sparc64 port). 3727 */ 3728retry: 3729 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3730 prot) < 0) { 3731 vm_page_lock_queues(); 3732 for (i = 0; i < pidx; ++i) { 3733 vm_page_unhold(bp->b_pages[i]); 3734 bp->b_pages[i] = NULL; 3735 } 3736 vm_page_unlock_queues(); 3737 return(-1); 3738 } 3739 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3740 if (m == NULL) 3741 goto retry; 3742 bp->b_pages[pidx] = m; 3743 } 3744 if (pidx > btoc(MAXPHYS)) 3745 panic("vmapbuf: mapped more than MAXPHYS"); 3746 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3747 3748 kva = bp->b_saveaddr; 3749 bp->b_npages = pidx; 3750 bp->b_saveaddr = bp->b_data; 3751 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3752 return(0); 3753} 3754 3755/* 3756 * Free the io map PTEs associated with this IO operation. 3757 * We also invalidate the TLB entries and restore the original b_addr. 3758 */ 3759void 3760vunmapbuf(struct buf *bp) 3761{ 3762 int pidx; 3763 int npages; 3764 3765 npages = bp->b_npages; 3766 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3767 vm_page_lock_queues(); 3768 for (pidx = 0; pidx < npages; pidx++) 3769 vm_page_unhold(bp->b_pages[pidx]); 3770 vm_page_unlock_queues(); 3771 3772 bp->b_data = bp->b_saveaddr; 3773} 3774 3775void 3776bdone(struct buf *bp) 3777{ 3778 3779 mtx_lock(&bdonelock); 3780 bp->b_flags |= B_DONE; 3781 wakeup(bp); 3782 mtx_unlock(&bdonelock); 3783} 3784 3785void 3786bwait(struct buf *bp, u_char pri, const char *wchan) 3787{ 3788 3789 mtx_lock(&bdonelock); 3790 while ((bp->b_flags & B_DONE) == 0) 3791 msleep(bp, &bdonelock, pri, wchan, 0); 3792 mtx_unlock(&bdonelock); 3793} 3794 3795#if 0 /* this is here to unconfuse p4 diff */ 3796 3797void 3798bufstrategy(struct bufobj *bo, struct buf *bp) 3799{ 3800 int i = 0; 3801 struct vnode *vp; 3802 3803 vp = bp->b_vp; 3804 KASSERT(vp == bo->bo_vnode, ("Inconsistent vnode bufstrategy")); 3805 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3806 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3807 i = VOP_STRATEGY(vp, bp); 3808 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3809} 3810 3811#endif 3812 3813void 3814bufobj_wref(struct bufobj *bo) 3815{ 3816 3817 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3818 BO_LOCK(bo); 3819 bo->bo_numoutput++; 3820 BO_UNLOCK(bo); 3821} 3822 3823void 3824bufobj_wdrop(struct bufobj *bo) 3825{ 3826 3827 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3828 BO_LOCK(bo); 3829 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3830 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3831 bo->bo_flag &= ~BO_WWAIT; 3832 wakeup(&bo->bo_numoutput); 3833 } 3834 BO_UNLOCK(bo); 3835} 3836 3837int 3838bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3839{ 3840 int error; 3841 3842 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3843 ASSERT_BO_LOCKED(bo); 3844 error = 0; 3845 while (bo->bo_numoutput) { 3846 bo->bo_flag |= BO_WWAIT; 3847 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3848 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3849 if (error) 3850 break; 3851 } 3852 return (error); 3853} 3854 3855 3856#include "opt_ddb.h" 3857#ifdef DDB 3858#include <ddb/ddb.h> 3859 3860/* DDB command to show buffer data */ 3861DB_SHOW_COMMAND(buffer, db_show_buffer) 3862{ 3863 /* get args */ 3864 struct buf *bp = (struct buf *)addr; 3865 3866 if (!have_addr) { 3867 db_printf("usage: show buffer <addr>\n"); 3868 return; 3869 } 3870 3871 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3872 db_printf( 3873 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3874 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd\n", 3875 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3876 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3877 (intmax_t)bp->b_blkno); 3878 if (bp->b_npages) { 3879 int i; 3880 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3881 for (i = 0; i < bp->b_npages; i++) { 3882 vm_page_t m; 3883 m = bp->b_pages[i]; 3884 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3885 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3886 if ((i + 1) < bp->b_npages) 3887 db_printf(","); 3888 } 3889 db_printf("\n"); 3890 } 3891} 3892#endif /* DDB */ 3893