vfs_bio.c revision 111119
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 * $FreeBSD: head/sys/kern/vfs_bio.c 111119 2003-02-19 05:47:46Z imp $ 15 */ 16 17/* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 * 27 * see man buf(9) for more info. 28 */ 29 30#include <sys/param.h> 31#include <sys/systm.h> 32#include <sys/stdint.h> 33#include <sys/bio.h> 34#include <sys/buf.h> 35#include <sys/devicestat.h> 36#include <sys/eventhandler.h> 37#include <sys/lock.h> 38#include <sys/malloc.h> 39#include <sys/mount.h> 40#include <sys/mutex.h> 41#include <sys/kernel.h> 42#include <sys/kthread.h> 43#include <sys/proc.h> 44#include <sys/resourcevar.h> 45#include <sys/sysctl.h> 46#include <sys/vmmeter.h> 47#include <sys/vnode.h> 48#include <vm/vm.h> 49#include <vm/vm_param.h> 50#include <vm/vm_kern.h> 51#include <vm/vm_pageout.h> 52#include <vm/vm_page.h> 53#include <vm/vm_object.h> 54#include <vm/vm_extern.h> 55#include <vm/vm_map.h> 56 57static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 58 59struct bio_ops bioops; /* I/O operation notification */ 60 61struct buf_ops buf_ops_bio = { 62 "buf_ops_bio", 63 bwrite 64}; 65 66/* 67 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 68 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 69 */ 70struct buf *buf; /* buffer header pool */ 71struct mtx buftimelock; /* Interlock on setting prio and timo */ 72 73static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 74 vm_offset_t to); 75static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 76 vm_offset_t to); 77static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 78 int pageno, vm_page_t m); 79static void vfs_clean_pages(struct buf * bp); 80static void vfs_setdirty(struct buf *bp); 81static void vfs_vmio_release(struct buf *bp); 82static void vfs_backgroundwritedone(struct buf *bp); 83static int vfs_bio_clcheck(struct vnode *vp, int size, 84 daddr_t lblkno, daddr_t blkno); 85static int flushbufqueues(void); 86static void buf_daemon(void); 87void bremfreel(struct buf * bp); 88 89int vmiodirenable = TRUE; 90SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 91 "Use the VM system for directory writes"); 92int runningbufspace; 93SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 94 "Amount of presently outstanding async buffer io"); 95static int bufspace; 96SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 97 "KVA memory used for bufs"); 98static int maxbufspace; 99SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 100 "Maximum allowed value of bufspace (including buf_daemon)"); 101static int bufmallocspace; 102SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 103 "Amount of malloced memory for buffers"); 104static int maxbufmallocspace; 105SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 106 "Maximum amount of malloced memory for buffers"); 107static int lobufspace; 108SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 109 "Minimum amount of buffers we want to have"); 110static int hibufspace; 111SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 112 "Maximum allowed value of bufspace (excluding buf_daemon)"); 113static int bufreusecnt; 114SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 115 "Number of times we have reused a buffer"); 116static int buffreekvacnt; 117SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 118 "Number of times we have freed the KVA space from some buffer"); 119static int bufdefragcnt; 120SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 121 "Number of times we have had to repeat buffer allocation to defragment"); 122static int lorunningspace; 123SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 124 "Minimum preferred space used for in-progress I/O"); 125static int hirunningspace; 126SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 127 "Maximum amount of space to use for in-progress I/O"); 128static int numdirtybuffers; 129SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 130 "Number of buffers that are dirty (has unwritten changes) at the moment"); 131static int lodirtybuffers; 132SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 133 "How many buffers we want to have free before bufdaemon can sleep"); 134static int hidirtybuffers; 135SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 136 "When the number of dirty buffers is considered severe"); 137static int numfreebuffers; 138SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 139 "Number of free buffers"); 140static int lofreebuffers; 141SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 142 "XXX Unused"); 143static int hifreebuffers; 144SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 145 "XXX Complicatedly unused"); 146static int getnewbufcalls; 147SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 148 "Number of calls to getnewbuf"); 149static int getnewbufrestarts; 150SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 151 "Number of times getnewbuf has had to restart a buffer aquisition"); 152static int dobkgrdwrite = 1; 153SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 154 "Do background writes (honoring the BX_BKGRDWRITE flag)?"); 155 156/* 157 * Wakeup point for bufdaemon, as well as indicator of whether it is already 158 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 159 * is idling. 160 */ 161static int bd_request; 162 163/* 164 * This lock synchronizes access to bd_request. 165 */ 166static struct mtx bdlock; 167 168/* 169 * bogus page -- for I/O to/from partially complete buffers 170 * this is a temporary solution to the problem, but it is not 171 * really that bad. it would be better to split the buffer 172 * for input in the case of buffers partially already in memory, 173 * but the code is intricate enough already. 174 */ 175vm_page_t bogus_page; 176 177/* 178 * Synchronization (sleep/wakeup) variable for active buffer space requests. 179 * Set when wait starts, cleared prior to wakeup(). 180 * Used in runningbufwakeup() and waitrunningbufspace(). 181 */ 182static int runningbufreq; 183 184/* 185 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 186 * waitrunningbufspace(). 187 */ 188static struct mtx rbreqlock; 189 190/* 191 * Synchronization (sleep/wakeup) variable for buffer requests. 192 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 193 * by and/or. 194 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 195 * getnewbuf(), and getblk(). 196 */ 197static int needsbuffer; 198 199/* 200 * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 201 */ 202static struct mtx nblock; 203 204#ifdef USE_BUFHASH 205/* 206 * Mask for index into the buffer hash table, which needs to be power of 2 in 207 * size. Set in kern_vfs_bio_buffer_alloc. 208 */ 209static int bufhashmask; 210 211/* 212 * Hash table for all buffers, with a linked list hanging from each table 213 * entry. Set in kern_vfs_bio_buffer_alloc, initialized in buf_init. 214 */ 215static LIST_HEAD(bufhashhdr, buf) *bufhashtbl; 216 217/* 218 * Somewhere to store buffers when they are not in another list, to always 219 * have them in a list (and thus being able to use the same set of operations 220 * on them.) 221 */ 222static struct bufhashhdr invalhash; 223 224#endif 225 226/* 227 * Definitions for the buffer free lists. 228 */ 229#define BUFFER_QUEUES 6 /* number of free buffer queues */ 230 231#define QUEUE_NONE 0 /* on no queue */ 232#define QUEUE_LOCKED 1 /* locked buffers */ 233#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ 234#define QUEUE_DIRTY 3 /* B_DELWRI buffers */ 235#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 236#define QUEUE_EMPTY 5 /* empty buffer headers */ 237 238/* Queues for free buffers with various properties */ 239static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 240 241/* Lock for the bufqueues */ 242static struct mtx bqlock; 243 244/* 245 * Single global constant for BUF_WMESG, to avoid getting multiple references. 246 * buf_wmesg is referred from macros. 247 */ 248const char *buf_wmesg = BUF_WMESG; 249 250#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 251#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 252#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 253#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 254 255#ifdef USE_BUFHASH 256/* 257 * Buffer hash table code. Note that the logical block scans linearly, which 258 * gives us some L1 cache locality. 259 */ 260 261static __inline 262struct bufhashhdr * 263bufhash(struct vnode *vnp, daddr_t bn) 264{ 265 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 266} 267 268#endif 269 270/* 271 * numdirtywakeup: 272 * 273 * If someone is blocked due to there being too many dirty buffers, 274 * and numdirtybuffers is now reasonable, wake them up. 275 */ 276 277static __inline void 278numdirtywakeup(int level) 279{ 280 if (numdirtybuffers <= level) { 281 mtx_lock(&nblock); 282 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 283 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 284 wakeup(&needsbuffer); 285 } 286 mtx_unlock(&nblock); 287 } 288} 289 290/* 291 * bufspacewakeup: 292 * 293 * Called when buffer space is potentially available for recovery. 294 * getnewbuf() will block on this flag when it is unable to free 295 * sufficient buffer space. Buffer space becomes recoverable when 296 * bp's get placed back in the queues. 297 */ 298 299static __inline void 300bufspacewakeup(void) 301{ 302 /* 303 * If someone is waiting for BUF space, wake them up. Even 304 * though we haven't freed the kva space yet, the waiting 305 * process will be able to now. 306 */ 307 mtx_lock(&nblock); 308 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 309 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 310 wakeup(&needsbuffer); 311 } 312 mtx_unlock(&nblock); 313} 314 315/* 316 * runningbufwakeup() - in-progress I/O accounting. 317 * 318 */ 319static __inline void 320runningbufwakeup(struct buf *bp) 321{ 322 if (bp->b_runningbufspace) { 323 atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 324 bp->b_runningbufspace = 0; 325 mtx_lock(&rbreqlock); 326 if (runningbufreq && runningbufspace <= lorunningspace) { 327 runningbufreq = 0; 328 wakeup(&runningbufreq); 329 } 330 mtx_unlock(&rbreqlock); 331 } 332} 333 334/* 335 * bufcountwakeup: 336 * 337 * Called when a buffer has been added to one of the free queues to 338 * account for the buffer and to wakeup anyone waiting for free buffers. 339 * This typically occurs when large amounts of metadata are being handled 340 * by the buffer cache ( else buffer space runs out first, usually ). 341 */ 342 343static __inline void 344bufcountwakeup(void) 345{ 346 atomic_add_int(&numfreebuffers, 1); 347 mtx_lock(&nblock); 348 if (needsbuffer) { 349 needsbuffer &= ~VFS_BIO_NEED_ANY; 350 if (numfreebuffers >= hifreebuffers) 351 needsbuffer &= ~VFS_BIO_NEED_FREE; 352 wakeup(&needsbuffer); 353 } 354 mtx_unlock(&nblock); 355} 356 357/* 358 * waitrunningbufspace() 359 * 360 * runningbufspace is a measure of the amount of I/O currently 361 * running. This routine is used in async-write situations to 362 * prevent creating huge backups of pending writes to a device. 363 * Only asynchronous writes are governed by this function. 364 * 365 * Reads will adjust runningbufspace, but will not block based on it. 366 * The read load has a side effect of reducing the allowed write load. 367 * 368 * This does NOT turn an async write into a sync write. It waits 369 * for earlier writes to complete and generally returns before the 370 * caller's write has reached the device. 371 */ 372static __inline void 373waitrunningbufspace(void) 374{ 375 mtx_lock(&rbreqlock); 376 while (runningbufspace > hirunningspace) { 377 ++runningbufreq; 378 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 379 } 380 mtx_unlock(&rbreqlock); 381} 382 383 384/* 385 * vfs_buf_test_cache: 386 * 387 * Called when a buffer is extended. This function clears the B_CACHE 388 * bit if the newly extended portion of the buffer does not contain 389 * valid data. 390 */ 391static __inline__ 392void 393vfs_buf_test_cache(struct buf *bp, 394 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 395 vm_page_t m) 396{ 397 GIANT_REQUIRED; 398 399 if (bp->b_flags & B_CACHE) { 400 int base = (foff + off) & PAGE_MASK; 401 if (vm_page_is_valid(m, base, size) == 0) 402 bp->b_flags &= ~B_CACHE; 403 } 404} 405 406/* Wake up the buffer deamon if necessary */ 407static __inline__ 408void 409bd_wakeup(int dirtybuflevel) 410{ 411 mtx_lock(&bdlock); 412 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 413 bd_request = 1; 414 wakeup(&bd_request); 415 } 416 mtx_unlock(&bdlock); 417} 418 419/* 420 * bd_speedup - speedup the buffer cache flushing code 421 */ 422 423static __inline__ 424void 425bd_speedup(void) 426{ 427 bd_wakeup(1); 428} 429 430/* 431 * Calculating buffer cache scaling values and reserve space for buffer 432 * headers. This is called during low level kernel initialization and 433 * may be called more then once. We CANNOT write to the memory area 434 * being reserved at this time. 435 */ 436caddr_t 437kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 438{ 439 /* 440 * physmem_est is in pages. Convert it to kilobytes (assumes 441 * PAGE_SIZE is >= 1K) 442 */ 443 physmem_est = physmem_est * (PAGE_SIZE / 1024); 444 445 /* 446 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 447 * For the first 64MB of ram nominally allocate sufficient buffers to 448 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 449 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 450 * the buffer cache we limit the eventual kva reservation to 451 * maxbcache bytes. 452 * 453 * factor represents the 1/4 x ram conversion. 454 */ 455 if (nbuf == 0) { 456 int factor = 4 * BKVASIZE / 1024; 457 458 nbuf = 50; 459 if (physmem_est > 4096) 460 nbuf += min((physmem_est - 4096) / factor, 461 65536 / factor); 462 if (physmem_est > 65536) 463 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 464 465 if (maxbcache && nbuf > maxbcache / BKVASIZE) 466 nbuf = maxbcache / BKVASIZE; 467 } 468 469#if 0 470 /* 471 * Do not allow the buffer_map to be more then 1/2 the size of the 472 * kernel_map. 473 */ 474 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 475 (BKVASIZE * 2)) { 476 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 477 (BKVASIZE * 2); 478 printf("Warning: nbufs capped at %d\n", nbuf); 479 } 480#endif 481 482 /* 483 * swbufs are used as temporary holders for I/O, such as paging I/O. 484 * We have no less then 16 and no more then 256. 485 */ 486 nswbuf = max(min(nbuf/4, 256), 16); 487 488 /* 489 * Reserve space for the buffer cache buffers 490 */ 491 swbuf = (void *)v; 492 v = (caddr_t)(swbuf + nswbuf); 493 buf = (void *)v; 494 v = (caddr_t)(buf + nbuf); 495 496#ifdef USE_BUFHASH 497 /* 498 * Calculate the hash table size and reserve space 499 */ 500 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 501 ; 502 bufhashtbl = (void *)v; 503 v = (caddr_t)(bufhashtbl + bufhashmask); 504 --bufhashmask; 505#endif 506 return(v); 507} 508 509/* Initialize the buffer subsystem. Called before use of any buffers. */ 510void 511bufinit(void) 512{ 513 struct buf *bp; 514 vm_offset_t bogus_offset; 515 int i; 516 517 GIANT_REQUIRED; 518 519#ifdef USE_BUFHASH 520 LIST_INIT(&invalhash); 521#endif 522 mtx_init(&buftimelock, "buftime lock", NULL, MTX_DEF); 523 mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 524 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 525 mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 526 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 527 528#ifdef USE_BUFHASH 529 for (i = 0; i <= bufhashmask; i++) 530 LIST_INIT(&bufhashtbl[i]); 531#endif 532 533 /* next, make a null set of free lists */ 534 for (i = 0; i < BUFFER_QUEUES; i++) 535 TAILQ_INIT(&bufqueues[i]); 536 537 /* finally, initialize each buffer header and stick on empty q */ 538 for (i = 0; i < nbuf; i++) { 539 bp = &buf[i]; 540 bzero(bp, sizeof *bp); 541 bp->b_flags = B_INVAL; /* we're just an empty header */ 542 bp->b_dev = NODEV; 543 bp->b_rcred = NOCRED; 544 bp->b_wcred = NOCRED; 545 bp->b_qindex = QUEUE_EMPTY; 546 bp->b_vflags = 0; 547 bp->b_xflags = 0; 548 LIST_INIT(&bp->b_dep); 549 BUF_LOCKINIT(bp); 550 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 551#ifdef USE_BUFHASH 552 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 553#endif 554 } 555 556 /* 557 * maxbufspace is the absolute maximum amount of buffer space we are 558 * allowed to reserve in KVM and in real terms. The absolute maximum 559 * is nominally used by buf_daemon. hibufspace is the nominal maximum 560 * used by most other processes. The differential is required to 561 * ensure that buf_daemon is able to run when other processes might 562 * be blocked waiting for buffer space. 563 * 564 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 565 * this may result in KVM fragmentation which is not handled optimally 566 * by the system. 567 */ 568 maxbufspace = nbuf * BKVASIZE; 569 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 570 lobufspace = hibufspace - MAXBSIZE; 571 572 lorunningspace = 512 * 1024; 573 hirunningspace = 1024 * 1024; 574 575/* 576 * Limit the amount of malloc memory since it is wired permanently into 577 * the kernel space. Even though this is accounted for in the buffer 578 * allocation, we don't want the malloced region to grow uncontrolled. 579 * The malloc scheme improves memory utilization significantly on average 580 * (small) directories. 581 */ 582 maxbufmallocspace = hibufspace / 20; 583 584/* 585 * Reduce the chance of a deadlock occuring by limiting the number 586 * of delayed-write dirty buffers we allow to stack up. 587 */ 588 hidirtybuffers = nbuf / 4 + 20; 589 numdirtybuffers = 0; 590/* 591 * To support extreme low-memory systems, make sure hidirtybuffers cannot 592 * eat up all available buffer space. This occurs when our minimum cannot 593 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 594 * BKVASIZE'd (8K) buffers. 595 */ 596 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 597 hidirtybuffers >>= 1; 598 } 599 lodirtybuffers = hidirtybuffers / 2; 600 601/* 602 * Try to keep the number of free buffers in the specified range, 603 * and give special processes (e.g. like buf_daemon) access to an 604 * emergency reserve. 605 */ 606 lofreebuffers = nbuf / 18 + 5; 607 hifreebuffers = 2 * lofreebuffers; 608 numfreebuffers = nbuf; 609 610/* 611 * Maximum number of async ops initiated per buf_daemon loop. This is 612 * somewhat of a hack at the moment, we really need to limit ourselves 613 * based on the number of bytes of I/O in-transit that were initiated 614 * from buf_daemon. 615 */ 616 617 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 618 vm_object_lock(kernel_object); 619 bogus_page = vm_page_alloc(kernel_object, 620 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 621 VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 622 vm_object_unlock(kernel_object); 623} 624 625/* 626 * bfreekva() - free the kva allocation for a buffer. 627 * 628 * Must be called at splbio() or higher as this is the only locking for 629 * buffer_map. 630 * 631 * Since this call frees up buffer space, we call bufspacewakeup(). 632 */ 633static void 634bfreekva(struct buf * bp) 635{ 636 GIANT_REQUIRED; 637 638 if (bp->b_kvasize) { 639 atomic_add_int(&buffreekvacnt, 1); 640 atomic_subtract_int(&bufspace, bp->b_kvasize); 641 vm_map_delete(buffer_map, 642 (vm_offset_t) bp->b_kvabase, 643 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 644 ); 645 bp->b_kvasize = 0; 646 bufspacewakeup(); 647 } 648} 649 650/* 651 * bremfree: 652 * 653 * Remove the buffer from the appropriate free list. 654 */ 655void 656bremfree(struct buf * bp) 657{ 658 mtx_lock(&bqlock); 659 bremfreel(bp); 660 mtx_unlock(&bqlock); 661} 662 663void 664bremfreel(struct buf * bp) 665{ 666 int s = splbio(); 667 int old_qindex = bp->b_qindex; 668 669 GIANT_REQUIRED; 670 671 if (bp->b_qindex != QUEUE_NONE) { 672 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 673 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 674 bp->b_qindex = QUEUE_NONE; 675 } else { 676 if (BUF_REFCNT(bp) <= 1) 677 panic("bremfree: removing a buffer not on a queue"); 678 } 679 680 /* 681 * Fixup numfreebuffers count. If the buffer is invalid or not 682 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 683 * the buffer was free and we must decrement numfreebuffers. 684 */ 685 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 686 switch(old_qindex) { 687 case QUEUE_DIRTY: 688 case QUEUE_CLEAN: 689 case QUEUE_EMPTY: 690 case QUEUE_EMPTYKVA: 691 atomic_subtract_int(&numfreebuffers, 1); 692 break; 693 default: 694 break; 695 } 696 } 697 splx(s); 698} 699 700 701/* 702 * Get a buffer with the specified data. Look in the cache first. We 703 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 704 * is set, the buffer is valid and we do not have to do anything ( see 705 * getblk() ). This is really just a special case of breadn(). 706 */ 707int 708bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 709 struct buf ** bpp) 710{ 711 712 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 713} 714 715/* 716 * Operates like bread, but also starts asynchronous I/O on 717 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 718 * to initiating I/O . If B_CACHE is set, the buffer is valid 719 * and we do not have to do anything. 720 */ 721int 722breadn(struct vnode * vp, daddr_t blkno, int size, 723 daddr_t * rablkno, int *rabsize, 724 int cnt, struct ucred * cred, struct buf ** bpp) 725{ 726 struct buf *bp, *rabp; 727 int i; 728 int rv = 0, readwait = 0; 729 730 *bpp = bp = getblk(vp, blkno, size, 0, 0); 731 732 /* if not found in cache, do some I/O */ 733 if ((bp->b_flags & B_CACHE) == 0) { 734 if (curthread != PCPU_GET(idlethread)) 735 curthread->td_proc->p_stats->p_ru.ru_inblock++; 736 bp->b_iocmd = BIO_READ; 737 bp->b_flags &= ~B_INVAL; 738 bp->b_ioflags &= ~BIO_ERROR; 739 if (bp->b_rcred == NOCRED && cred != NOCRED) 740 bp->b_rcred = crhold(cred); 741 vfs_busy_pages(bp, 0); 742 if (vp->v_type == VCHR) 743 VOP_SPECSTRATEGY(vp, bp); 744 else 745 VOP_STRATEGY(vp, bp); 746 ++readwait; 747 } 748 749 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 750 if (inmem(vp, *rablkno)) 751 continue; 752 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 753 754 if ((rabp->b_flags & B_CACHE) == 0) { 755 if (curthread != PCPU_GET(idlethread)) 756 curthread->td_proc->p_stats->p_ru.ru_inblock++; 757 rabp->b_flags |= B_ASYNC; 758 rabp->b_flags &= ~B_INVAL; 759 rabp->b_ioflags &= ~BIO_ERROR; 760 rabp->b_iocmd = BIO_READ; 761 if (rabp->b_rcred == NOCRED && cred != NOCRED) 762 rabp->b_rcred = crhold(cred); 763 vfs_busy_pages(rabp, 0); 764 BUF_KERNPROC(rabp); 765 if (vp->v_type == VCHR) 766 VOP_SPECSTRATEGY(vp, rabp); 767 else 768 VOP_STRATEGY(vp, rabp); 769 } else { 770 brelse(rabp); 771 } 772 } 773 774 if (readwait) { 775 rv = bufwait(bp); 776 } 777 return (rv); 778} 779 780/* 781 * Write, release buffer on completion. (Done by iodone 782 * if async). Do not bother writing anything if the buffer 783 * is invalid. 784 * 785 * Note that we set B_CACHE here, indicating that buffer is 786 * fully valid and thus cacheable. This is true even of NFS 787 * now so we set it generally. This could be set either here 788 * or in biodone() since the I/O is synchronous. We put it 789 * here. 790 */ 791 792int 793bwrite(struct buf * bp) 794{ 795 int oldflags, s; 796 struct buf *newbp; 797 798 if (bp->b_flags & B_INVAL) { 799 brelse(bp); 800 return (0); 801 } 802 803 oldflags = bp->b_flags; 804 805 if (BUF_REFCNT(bp) == 0) 806 panic("bwrite: buffer is not busy???"); 807 s = splbio(); 808 /* 809 * If a background write is already in progress, delay 810 * writing this block if it is asynchronous. Otherwise 811 * wait for the background write to complete. 812 */ 813 if (bp->b_xflags & BX_BKGRDINPROG) { 814 if (bp->b_flags & B_ASYNC) { 815 splx(s); 816 bdwrite(bp); 817 return (0); 818 } 819 bp->b_xflags |= BX_BKGRDWAIT; 820 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); 821 if (bp->b_xflags & BX_BKGRDINPROG) 822 panic("bwrite: still writing"); 823 } 824 825 /* Mark the buffer clean */ 826 bundirty(bp); 827 828 /* 829 * If this buffer is marked for background writing and we 830 * do not have to wait for it, make a copy and write the 831 * copy so as to leave this buffer ready for further use. 832 * 833 * This optimization eats a lot of memory. If we have a page 834 * or buffer shortfall we can't do it. 835 */ 836 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 837 (bp->b_flags & B_ASYNC) && 838 !vm_page_count_severe() && 839 !buf_dirty_count_severe()) { 840 if (bp->b_iodone != NULL) { 841 printf("bp->b_iodone = %p\n", bp->b_iodone); 842 panic("bwrite: need chained iodone"); 843 } 844 845 /* get a new block */ 846 newbp = geteblk(bp->b_bufsize); 847 848 /* 849 * set it to be identical to the old block. We have to 850 * set b_lblkno and BKGRDMARKER before calling bgetvp() 851 * to avoid confusing the splay tree and gbincore(). 852 */ 853 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 854 newbp->b_lblkno = bp->b_lblkno; 855 newbp->b_xflags |= BX_BKGRDMARKER; 856 bgetvp(bp->b_vp, newbp); 857 newbp->b_blkno = bp->b_blkno; 858 newbp->b_offset = bp->b_offset; 859 newbp->b_iodone = vfs_backgroundwritedone; 860 newbp->b_flags |= B_ASYNC; 861 newbp->b_flags &= ~B_INVAL; 862 863 /* move over the dependencies */ 864 if (LIST_FIRST(&bp->b_dep) != NULL) 865 buf_movedeps(bp, newbp); 866 867 /* 868 * Initiate write on the copy, release the original to 869 * the B_LOCKED queue so that it cannot go away until 870 * the background write completes. If not locked it could go 871 * away and then be reconstituted while it was being written. 872 * If the reconstituted buffer were written, we could end up 873 * with two background copies being written at the same time. 874 */ 875 bp->b_xflags |= BX_BKGRDINPROG; 876 bp->b_flags |= B_LOCKED; 877 bqrelse(bp); 878 bp = newbp; 879 } 880 881 bp->b_flags &= ~B_DONE; 882 bp->b_ioflags &= ~BIO_ERROR; 883 bp->b_flags |= B_WRITEINPROG | B_CACHE; 884 bp->b_iocmd = BIO_WRITE; 885 886 VI_LOCK(bp->b_vp); 887 bp->b_vp->v_numoutput++; 888 VI_UNLOCK(bp->b_vp); 889 vfs_busy_pages(bp, 1); 890 891 /* 892 * Normal bwrites pipeline writes 893 */ 894 bp->b_runningbufspace = bp->b_bufsize; 895 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 896 897 if (curthread != PCPU_GET(idlethread)) 898 curthread->td_proc->p_stats->p_ru.ru_oublock++; 899 splx(s); 900 if (oldflags & B_ASYNC) 901 BUF_KERNPROC(bp); 902 if (bp->b_vp->v_type == VCHR) 903 VOP_SPECSTRATEGY(bp->b_vp, bp); 904 else 905 VOP_STRATEGY(bp->b_vp, bp); 906 907 if ((oldflags & B_ASYNC) == 0) { 908 int rtval = bufwait(bp); 909 brelse(bp); 910 return (rtval); 911 } else if ((oldflags & B_NOWDRAIN) == 0) { 912 /* 913 * don't allow the async write to saturate the I/O 914 * system. Deadlocks can occur only if a device strategy 915 * routine (like in MD) turns around and issues another 916 * high-level write, in which case B_NOWDRAIN is expected 917 * to be set. Otherwise we will not deadlock here because 918 * we are blocking waiting for I/O that is already in-progress 919 * to complete. 920 */ 921 waitrunningbufspace(); 922 } 923 924 return (0); 925} 926 927/* 928 * Complete a background write started from bwrite. 929 */ 930static void 931vfs_backgroundwritedone(bp) 932 struct buf *bp; 933{ 934 struct buf *origbp; 935 936 /* 937 * Find the original buffer that we are writing. 938 */ 939 VI_LOCK(bp->b_vp); 940 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 941 panic("backgroundwritedone: lost buffer"); 942 VI_UNLOCK(bp->b_vp); 943 /* 944 * Process dependencies then return any unfinished ones. 945 */ 946 if (LIST_FIRST(&bp->b_dep) != NULL) 947 buf_complete(bp); 948 if (LIST_FIRST(&bp->b_dep) != NULL) 949 buf_movedeps(bp, origbp); 950 951 /* XXX Find out if origbp can disappear or get inconsistent */ 952 /* 953 * Clear the BX_BKGRDINPROG flag in the original buffer 954 * and awaken it if it is waiting for the write to complete. 955 * If BX_BKGRDINPROG is not set in the original buffer it must 956 * have been released and re-instantiated - which is not legal. 957 */ 958 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 959 ("backgroundwritedone: lost buffer2")); 960 origbp->b_xflags &= ~BX_BKGRDINPROG; 961 if (origbp->b_xflags & BX_BKGRDWAIT) { 962 origbp->b_xflags &= ~BX_BKGRDWAIT; 963 wakeup(&origbp->b_xflags); 964 } 965 /* 966 * Clear the B_LOCKED flag and remove it from the locked 967 * queue if it currently resides there. 968 */ 969 origbp->b_flags &= ~B_LOCKED; 970 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 971 bremfree(origbp); 972 bqrelse(origbp); 973 } 974 /* 975 * This buffer is marked B_NOCACHE, so when it is released 976 * by biodone, it will be tossed. We mark it with BIO_READ 977 * to avoid biodone doing a second vwakeup. 978 */ 979 bp->b_flags |= B_NOCACHE; 980 bp->b_iocmd = BIO_READ; 981 bp->b_flags &= ~(B_CACHE | B_DONE); 982 bp->b_iodone = 0; 983 bufdone(bp); 984} 985 986/* 987 * Delayed write. (Buffer is marked dirty). Do not bother writing 988 * anything if the buffer is marked invalid. 989 * 990 * Note that since the buffer must be completely valid, we can safely 991 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 992 * biodone() in order to prevent getblk from writing the buffer 993 * out synchronously. 994 */ 995void 996bdwrite(struct buf * bp) 997{ 998 GIANT_REQUIRED; 999 1000 if (BUF_REFCNT(bp) == 0) 1001 panic("bdwrite: buffer is not busy"); 1002 1003 if (bp->b_flags & B_INVAL) { 1004 brelse(bp); 1005 return; 1006 } 1007 bdirty(bp); 1008 1009 /* 1010 * Set B_CACHE, indicating that the buffer is fully valid. This is 1011 * true even of NFS now. 1012 */ 1013 bp->b_flags |= B_CACHE; 1014 1015 /* 1016 * This bmap keeps the system from needing to do the bmap later, 1017 * perhaps when the system is attempting to do a sync. Since it 1018 * is likely that the indirect block -- or whatever other datastructure 1019 * that the filesystem needs is still in memory now, it is a good 1020 * thing to do this. Note also, that if the pageout daemon is 1021 * requesting a sync -- there might not be enough memory to do 1022 * the bmap then... So, this is important to do. 1023 */ 1024 if (bp->b_vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1025 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1026 } 1027 1028 /* 1029 * Set the *dirty* buffer range based upon the VM system dirty pages. 1030 */ 1031 vfs_setdirty(bp); 1032 1033 /* 1034 * We need to do this here to satisfy the vnode_pager and the 1035 * pageout daemon, so that it thinks that the pages have been 1036 * "cleaned". Note that since the pages are in a delayed write 1037 * buffer -- the VFS layer "will" see that the pages get written 1038 * out on the next sync, or perhaps the cluster will be completed. 1039 */ 1040 vfs_clean_pages(bp); 1041 bqrelse(bp); 1042 1043 /* 1044 * Wakeup the buffer flushing daemon if we have a lot of dirty 1045 * buffers (midpoint between our recovery point and our stall 1046 * point). 1047 */ 1048 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1049 1050 /* 1051 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1052 * due to the softdep code. 1053 */ 1054} 1055 1056/* 1057 * bdirty: 1058 * 1059 * Turn buffer into delayed write request. We must clear BIO_READ and 1060 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1061 * itself to properly update it in the dirty/clean lists. We mark it 1062 * B_DONE to ensure that any asynchronization of the buffer properly 1063 * clears B_DONE ( else a panic will occur later ). 1064 * 1065 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1066 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1067 * should only be called if the buffer is known-good. 1068 * 1069 * Since the buffer is not on a queue, we do not update the numfreebuffers 1070 * count. 1071 * 1072 * Must be called at splbio(). 1073 * The buffer must be on QUEUE_NONE. 1074 */ 1075void 1076bdirty(bp) 1077 struct buf *bp; 1078{ 1079 KASSERT(bp->b_qindex == QUEUE_NONE, 1080 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1081 bp->b_flags &= ~(B_RELBUF); 1082 bp->b_iocmd = BIO_WRITE; 1083 1084 if ((bp->b_flags & B_DELWRI) == 0) { 1085 bp->b_flags |= B_DONE | B_DELWRI; 1086 reassignbuf(bp, bp->b_vp); 1087 atomic_add_int(&numdirtybuffers, 1); 1088 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1089 } 1090} 1091 1092/* 1093 * bundirty: 1094 * 1095 * Clear B_DELWRI for buffer. 1096 * 1097 * Since the buffer is not on a queue, we do not update the numfreebuffers 1098 * count. 1099 * 1100 * Must be called at splbio(). 1101 * The buffer must be on QUEUE_NONE. 1102 */ 1103 1104void 1105bundirty(bp) 1106 struct buf *bp; 1107{ 1108 KASSERT(bp->b_qindex == QUEUE_NONE, 1109 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1110 1111 if (bp->b_flags & B_DELWRI) { 1112 bp->b_flags &= ~B_DELWRI; 1113 reassignbuf(bp, bp->b_vp); 1114 atomic_subtract_int(&numdirtybuffers, 1); 1115 numdirtywakeup(lodirtybuffers); 1116 } 1117 /* 1118 * Since it is now being written, we can clear its deferred write flag. 1119 */ 1120 bp->b_flags &= ~B_DEFERRED; 1121} 1122 1123/* 1124 * bawrite: 1125 * 1126 * Asynchronous write. Start output on a buffer, but do not wait for 1127 * it to complete. The buffer is released when the output completes. 1128 * 1129 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1130 * B_INVAL buffers. Not us. 1131 */ 1132void 1133bawrite(struct buf * bp) 1134{ 1135 bp->b_flags |= B_ASYNC; 1136 (void) BUF_WRITE(bp); 1137} 1138 1139/* 1140 * bwillwrite: 1141 * 1142 * Called prior to the locking of any vnodes when we are expecting to 1143 * write. We do not want to starve the buffer cache with too many 1144 * dirty buffers so we block here. By blocking prior to the locking 1145 * of any vnodes we attempt to avoid the situation where a locked vnode 1146 * prevents the various system daemons from flushing related buffers. 1147 */ 1148 1149void 1150bwillwrite(void) 1151{ 1152 if (numdirtybuffers >= hidirtybuffers) { 1153 int s; 1154 1155 mtx_lock(&Giant); 1156 s = splbio(); 1157 mtx_lock(&nblock); 1158 while (numdirtybuffers >= hidirtybuffers) { 1159 bd_wakeup(1); 1160 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1161 msleep(&needsbuffer, &nblock, 1162 (PRIBIO + 4), "flswai", 0); 1163 } 1164 splx(s); 1165 mtx_unlock(&nblock); 1166 mtx_unlock(&Giant); 1167 } 1168} 1169 1170/* 1171 * Return true if we have too many dirty buffers. 1172 */ 1173int 1174buf_dirty_count_severe(void) 1175{ 1176 return(numdirtybuffers >= hidirtybuffers); 1177} 1178 1179/* 1180 * brelse: 1181 * 1182 * Release a busy buffer and, if requested, free its resources. The 1183 * buffer will be stashed in the appropriate bufqueue[] allowing it 1184 * to be accessed later as a cache entity or reused for other purposes. 1185 */ 1186void 1187brelse(struct buf * bp) 1188{ 1189 int s; 1190 1191 GIANT_REQUIRED; 1192 1193 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1194 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1195 1196 s = splbio(); 1197 1198 if (bp->b_flags & B_LOCKED) 1199 bp->b_ioflags &= ~BIO_ERROR; 1200 1201 if (bp->b_iocmd == BIO_WRITE && 1202 (bp->b_ioflags & BIO_ERROR) && 1203 !(bp->b_flags & B_INVAL)) { 1204 /* 1205 * Failed write, redirty. Must clear BIO_ERROR to prevent 1206 * pages from being scrapped. If B_INVAL is set then 1207 * this case is not run and the next case is run to 1208 * destroy the buffer. B_INVAL can occur if the buffer 1209 * is outside the range supported by the underlying device. 1210 */ 1211 bp->b_ioflags &= ~BIO_ERROR; 1212 bdirty(bp); 1213 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1214 (bp->b_ioflags & BIO_ERROR) || 1215 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1216 /* 1217 * Either a failed I/O or we were asked to free or not 1218 * cache the buffer. 1219 */ 1220 bp->b_flags |= B_INVAL; 1221 if (LIST_FIRST(&bp->b_dep) != NULL) 1222 buf_deallocate(bp); 1223 if (bp->b_flags & B_DELWRI) { 1224 atomic_subtract_int(&numdirtybuffers, 1); 1225 numdirtywakeup(lodirtybuffers); 1226 } 1227 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1228 if ((bp->b_flags & B_VMIO) == 0) { 1229 if (bp->b_bufsize) 1230 allocbuf(bp, 0); 1231 if (bp->b_vp) 1232 brelvp(bp); 1233 } 1234 } 1235 1236 /* 1237 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1238 * is called with B_DELWRI set, the underlying pages may wind up 1239 * getting freed causing a previous write (bdwrite()) to get 'lost' 1240 * because pages associated with a B_DELWRI bp are marked clean. 1241 * 1242 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1243 * if B_DELWRI is set. 1244 * 1245 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1246 * on pages to return pages to the VM page queues. 1247 */ 1248 if (bp->b_flags & B_DELWRI) 1249 bp->b_flags &= ~B_RELBUF; 1250 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1251 bp->b_flags |= B_RELBUF; 1252 1253 /* 1254 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1255 * constituted, not even NFS buffers now. Two flags effect this. If 1256 * B_INVAL, the struct buf is invalidated but the VM object is kept 1257 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1258 * 1259 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1260 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1261 * buffer is also B_INVAL because it hits the re-dirtying code above. 1262 * 1263 * Normally we can do this whether a buffer is B_DELWRI or not. If 1264 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1265 * the commit state and we cannot afford to lose the buffer. If the 1266 * buffer has a background write in progress, we need to keep it 1267 * around to prevent it from being reconstituted and starting a second 1268 * background write. 1269 */ 1270 if ((bp->b_flags & B_VMIO) 1271 && !(bp->b_vp->v_mount != NULL && 1272 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1273 !vn_isdisk(bp->b_vp, NULL) && 1274 (bp->b_flags & B_DELWRI)) 1275 ) { 1276 1277 int i, j, resid; 1278 vm_page_t m; 1279 off_t foff; 1280 vm_pindex_t poff; 1281 vm_object_t obj; 1282 struct vnode *vp; 1283 1284 vp = bp->b_vp; 1285 obj = bp->b_object; 1286 1287 /* 1288 * Get the base offset and length of the buffer. Note that 1289 * in the VMIO case if the buffer block size is not 1290 * page-aligned then b_data pointer may not be page-aligned. 1291 * But our b_pages[] array *IS* page aligned. 1292 * 1293 * block sizes less then DEV_BSIZE (usually 512) are not 1294 * supported due to the page granularity bits (m->valid, 1295 * m->dirty, etc...). 1296 * 1297 * See man buf(9) for more information 1298 */ 1299 resid = bp->b_bufsize; 1300 foff = bp->b_offset; 1301 1302 for (i = 0; i < bp->b_npages; i++) { 1303 int had_bogus = 0; 1304 1305 m = bp->b_pages[i]; 1306 vm_page_lock_queues(); 1307 vm_page_flag_clear(m, PG_ZERO); 1308 vm_page_unlock_queues(); 1309 1310 /* 1311 * If we hit a bogus page, fixup *all* the bogus pages 1312 * now. 1313 */ 1314 if (m == bogus_page) { 1315 poff = OFF_TO_IDX(bp->b_offset); 1316 had_bogus = 1; 1317 1318 for (j = i; j < bp->b_npages; j++) { 1319 vm_page_t mtmp; 1320 mtmp = bp->b_pages[j]; 1321 if (mtmp == bogus_page) { 1322 mtmp = vm_page_lookup(obj, poff + j); 1323 if (!mtmp) { 1324 panic("brelse: page missing\n"); 1325 } 1326 bp->b_pages[j] = mtmp; 1327 } 1328 } 1329 1330 if ((bp->b_flags & B_INVAL) == 0) { 1331 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1332 } 1333 m = bp->b_pages[i]; 1334 } 1335 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1336 int poffset = foff & PAGE_MASK; 1337 int presid = resid > (PAGE_SIZE - poffset) ? 1338 (PAGE_SIZE - poffset) : resid; 1339 1340 KASSERT(presid >= 0, ("brelse: extra page")); 1341 vm_page_set_invalid(m, poffset, presid); 1342 if (had_bogus) 1343 printf("avoided corruption bug in bogus_page/brelse code\n"); 1344 } 1345 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1346 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1347 } 1348 1349 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1350 vfs_vmio_release(bp); 1351 1352 } else if (bp->b_flags & B_VMIO) { 1353 1354 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1355 vfs_vmio_release(bp); 1356 } 1357 1358 } 1359 1360 if (bp->b_qindex != QUEUE_NONE) 1361 panic("brelse: free buffer onto another queue???"); 1362 if (BUF_REFCNT(bp) > 1) { 1363 /* do not release to free list */ 1364 BUF_UNLOCK(bp); 1365 splx(s); 1366 return; 1367 } 1368 1369 /* enqueue */ 1370 mtx_lock(&bqlock); 1371 1372 /* buffers with no memory */ 1373 if (bp->b_bufsize == 0) { 1374 bp->b_flags |= B_INVAL; 1375 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1376 if (bp->b_xflags & BX_BKGRDINPROG) 1377 panic("losing buffer 1"); 1378 if (bp->b_kvasize) { 1379 bp->b_qindex = QUEUE_EMPTYKVA; 1380 } else { 1381 bp->b_qindex = QUEUE_EMPTY; 1382 } 1383 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1384#ifdef USE_BUFHASH 1385 LIST_REMOVE(bp, b_hash); 1386 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1387#endif 1388 bp->b_dev = NODEV; 1389 /* buffers with junk contents */ 1390 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1391 (bp->b_ioflags & BIO_ERROR)) { 1392 bp->b_flags |= B_INVAL; 1393 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1394 if (bp->b_xflags & BX_BKGRDINPROG) 1395 panic("losing buffer 2"); 1396 bp->b_qindex = QUEUE_CLEAN; 1397 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1398#ifdef USE_BUFHASH 1399 LIST_REMOVE(bp, b_hash); 1400 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1401#endif 1402 bp->b_dev = NODEV; 1403 1404 /* buffers that are locked */ 1405 } else if (bp->b_flags & B_LOCKED) { 1406 bp->b_qindex = QUEUE_LOCKED; 1407 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1408 1409 /* remaining buffers */ 1410 } else { 1411 if (bp->b_flags & B_DELWRI) 1412 bp->b_qindex = QUEUE_DIRTY; 1413 else 1414 bp->b_qindex = QUEUE_CLEAN; 1415 if (bp->b_flags & B_AGE) 1416 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1417 else 1418 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1419 } 1420 mtx_unlock(&bqlock); 1421 1422 /* 1423 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1424 * placed the buffer on the correct queue. We must also disassociate 1425 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1426 * find it. 1427 */ 1428 if (bp->b_flags & B_INVAL) { 1429 if (bp->b_flags & B_DELWRI) 1430 bundirty(bp); 1431 if (bp->b_vp) 1432 brelvp(bp); 1433 } 1434 1435 /* 1436 * Fixup numfreebuffers count. The bp is on an appropriate queue 1437 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1438 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1439 * if B_INVAL is set ). 1440 */ 1441 1442 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1443 bufcountwakeup(); 1444 1445 /* 1446 * Something we can maybe free or reuse 1447 */ 1448 if (bp->b_bufsize || bp->b_kvasize) 1449 bufspacewakeup(); 1450 1451 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1452 B_DIRECT | B_NOWDRAIN); 1453 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1454 panic("brelse: not dirty"); 1455 /* unlock */ 1456 BUF_UNLOCK(bp); 1457 splx(s); 1458} 1459 1460/* 1461 * Release a buffer back to the appropriate queue but do not try to free 1462 * it. The buffer is expected to be used again soon. 1463 * 1464 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1465 * biodone() to requeue an async I/O on completion. It is also used when 1466 * known good buffers need to be requeued but we think we may need the data 1467 * again soon. 1468 * 1469 * XXX we should be able to leave the B_RELBUF hint set on completion. 1470 */ 1471void 1472bqrelse(struct buf * bp) 1473{ 1474 int s; 1475 1476 s = splbio(); 1477 1478 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1479 1480 if (bp->b_qindex != QUEUE_NONE) 1481 panic("bqrelse: free buffer onto another queue???"); 1482 if (BUF_REFCNT(bp) > 1) { 1483 /* do not release to free list */ 1484 BUF_UNLOCK(bp); 1485 splx(s); 1486 return; 1487 } 1488 mtx_lock(&bqlock); 1489 if (bp->b_flags & B_LOCKED) { 1490 bp->b_ioflags &= ~BIO_ERROR; 1491 bp->b_qindex = QUEUE_LOCKED; 1492 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1493 /* buffers with stale but valid contents */ 1494 } else if (bp->b_flags & B_DELWRI) { 1495 bp->b_qindex = QUEUE_DIRTY; 1496 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1497 } else if (vm_page_count_severe()) { 1498 /* 1499 * We are too low on memory, we have to try to free the 1500 * buffer (most importantly: the wired pages making up its 1501 * backing store) *now*. 1502 */ 1503 mtx_unlock(&bqlock); 1504 splx(s); 1505 brelse(bp); 1506 return; 1507 } else { 1508 bp->b_qindex = QUEUE_CLEAN; 1509 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1510 } 1511 mtx_unlock(&bqlock); 1512 1513 if ((bp->b_flags & B_LOCKED) == 0 && 1514 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1515 bufcountwakeup(); 1516 } 1517 1518 /* 1519 * Something we can maybe free or reuse. 1520 */ 1521 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1522 bufspacewakeup(); 1523 1524 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1525 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1526 panic("bqrelse: not dirty"); 1527 /* unlock */ 1528 BUF_UNLOCK(bp); 1529 splx(s); 1530} 1531 1532/* Give pages used by the bp back to the VM system (where possible) */ 1533static void 1534vfs_vmio_release(bp) 1535 struct buf *bp; 1536{ 1537 int i; 1538 vm_page_t m; 1539 1540 GIANT_REQUIRED; 1541 vm_page_lock_queues(); 1542 for (i = 0; i < bp->b_npages; i++) { 1543 m = bp->b_pages[i]; 1544 bp->b_pages[i] = NULL; 1545 /* 1546 * In order to keep page LRU ordering consistent, put 1547 * everything on the inactive queue. 1548 */ 1549 vm_page_unwire(m, 0); 1550 /* 1551 * We don't mess with busy pages, it is 1552 * the responsibility of the process that 1553 * busied the pages to deal with them. 1554 */ 1555 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1556 continue; 1557 1558 if (m->wire_count == 0) { 1559 vm_page_flag_clear(m, PG_ZERO); 1560 /* 1561 * Might as well free the page if we can and it has 1562 * no valid data. We also free the page if the 1563 * buffer was used for direct I/O 1564 */ 1565 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1566 m->hold_count == 0) { 1567 vm_page_busy(m); 1568 pmap_remove_all(m); 1569 vm_page_free(m); 1570 } else if (bp->b_flags & B_DIRECT) { 1571 vm_page_try_to_free(m); 1572 } else if (vm_page_count_severe()) { 1573 vm_page_try_to_cache(m); 1574 } 1575 } 1576 } 1577 vm_page_unlock_queues(); 1578 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1579 1580 if (bp->b_bufsize) { 1581 bufspacewakeup(); 1582 bp->b_bufsize = 0; 1583 } 1584 bp->b_npages = 0; 1585 bp->b_flags &= ~B_VMIO; 1586 if (bp->b_vp) 1587 brelvp(bp); 1588} 1589 1590#ifdef USE_BUFHASH 1591/* 1592 * XXX MOVED TO VFS_SUBR.C 1593 * 1594 * Check to see if a block is currently memory resident. 1595 */ 1596struct buf * 1597gbincore(struct vnode * vp, daddr_t blkno) 1598{ 1599 struct buf *bp; 1600 struct bufhashhdr *bh; 1601 1602 bh = bufhash(vp, blkno); 1603 1604 /* Search hash chain */ 1605 LIST_FOREACH(bp, bh, b_hash) { 1606 /* hit */ 1607 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1608 (bp->b_flags & B_INVAL) == 0) { 1609 break; 1610 } 1611 } 1612 return (bp); 1613} 1614#endif 1615 1616/* 1617 * Check to see if a block at a particular lbn is available for a clustered 1618 * write. 1619 */ 1620static int 1621vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1622{ 1623 struct buf *bpa; 1624 int match; 1625 1626 match = 0; 1627 1628 /* If the buf isn't in core skip it */ 1629 if ((bpa = gbincore(vp, lblkno)) == NULL) 1630 return (0); 1631 1632 /* If the buf is busy we don't want to wait for it */ 1633 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1634 return (0); 1635 1636 /* Only cluster with valid clusterable delayed write buffers */ 1637 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1638 (B_DELWRI | B_CLUSTEROK)) 1639 goto done; 1640 1641 if (bpa->b_bufsize != size) 1642 goto done; 1643 1644 /* 1645 * Check to see if it is in the expected place on disk and that the 1646 * block has been mapped. 1647 */ 1648 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1649 match = 1; 1650done: 1651 BUF_UNLOCK(bpa); 1652 return (match); 1653} 1654 1655/* 1656 * vfs_bio_awrite: 1657 * 1658 * Implement clustered async writes for clearing out B_DELWRI buffers. 1659 * This is much better then the old way of writing only one buffer at 1660 * a time. Note that we may not be presented with the buffers in the 1661 * correct order, so we search for the cluster in both directions. 1662 */ 1663int 1664vfs_bio_awrite(struct buf * bp) 1665{ 1666 int i; 1667 int j; 1668 daddr_t lblkno = bp->b_lblkno; 1669 struct vnode *vp = bp->b_vp; 1670 int s; 1671 int ncl; 1672 int nwritten; 1673 int size; 1674 int maxcl; 1675 1676 s = splbio(); 1677 /* 1678 * right now we support clustered writing only to regular files. If 1679 * we find a clusterable block we could be in the middle of a cluster 1680 * rather then at the beginning. 1681 */ 1682 if ((vp->v_type == VREG) && 1683 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1684 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1685 1686 size = vp->v_mount->mnt_stat.f_iosize; 1687 maxcl = MAXPHYS / size; 1688 1689 VI_LOCK(vp); 1690 for (i = 1; i < maxcl; i++) 1691 if (vfs_bio_clcheck(vp, size, lblkno + i, 1692 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1693 break; 1694 1695 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1696 if (vfs_bio_clcheck(vp, size, lblkno - j, 1697 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1698 break; 1699 1700 VI_UNLOCK(vp); 1701 --j; 1702 ncl = i + j; 1703 /* 1704 * this is a possible cluster write 1705 */ 1706 if (ncl != 1) { 1707 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1708 splx(s); 1709 return nwritten; 1710 } 1711 } 1712 1713 BUF_LOCK(bp, LK_EXCLUSIVE); 1714 bremfree(bp); 1715 bp->b_flags |= B_ASYNC; 1716 1717 splx(s); 1718 /* 1719 * default (old) behavior, writing out only one block 1720 * 1721 * XXX returns b_bufsize instead of b_bcount for nwritten? 1722 */ 1723 nwritten = bp->b_bufsize; 1724 (void) BUF_WRITE(bp); 1725 1726 return nwritten; 1727} 1728 1729/* 1730 * getnewbuf: 1731 * 1732 * Find and initialize a new buffer header, freeing up existing buffers 1733 * in the bufqueues as necessary. The new buffer is returned locked. 1734 * 1735 * Important: B_INVAL is not set. If the caller wishes to throw the 1736 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1737 * 1738 * We block if: 1739 * We have insufficient buffer headers 1740 * We have insufficient buffer space 1741 * buffer_map is too fragmented ( space reservation fails ) 1742 * If we have to flush dirty buffers ( but we try to avoid this ) 1743 * 1744 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1745 * Instead we ask the buf daemon to do it for us. We attempt to 1746 * avoid piecemeal wakeups of the pageout daemon. 1747 */ 1748 1749static struct buf * 1750getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1751{ 1752 struct buf *bp; 1753 struct buf *nbp; 1754 int defrag = 0; 1755 int nqindex; 1756 static int flushingbufs; 1757 1758 GIANT_REQUIRED; 1759 1760 /* 1761 * We can't afford to block since we might be holding a vnode lock, 1762 * which may prevent system daemons from running. We deal with 1763 * low-memory situations by proactively returning memory and running 1764 * async I/O rather then sync I/O. 1765 */ 1766 1767 atomic_add_int(&getnewbufcalls, 1); 1768 atomic_subtract_int(&getnewbufrestarts, 1); 1769restart: 1770 atomic_add_int(&getnewbufrestarts, 1); 1771 1772 /* 1773 * Setup for scan. If we do not have enough free buffers, 1774 * we setup a degenerate case that immediately fails. Note 1775 * that if we are specially marked process, we are allowed to 1776 * dip into our reserves. 1777 * 1778 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1779 * 1780 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1781 * However, there are a number of cases (defragging, reusing, ...) 1782 * where we cannot backup. 1783 */ 1784 mtx_lock(&bqlock); 1785 nqindex = QUEUE_EMPTYKVA; 1786 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1787 1788 if (nbp == NULL) { 1789 /* 1790 * If no EMPTYKVA buffers and we are either 1791 * defragging or reusing, locate a CLEAN buffer 1792 * to free or reuse. If bufspace useage is low 1793 * skip this step so we can allocate a new buffer. 1794 */ 1795 if (defrag || bufspace >= lobufspace) { 1796 nqindex = QUEUE_CLEAN; 1797 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1798 } 1799 1800 /* 1801 * If we could not find or were not allowed to reuse a 1802 * CLEAN buffer, check to see if it is ok to use an EMPTY 1803 * buffer. We can only use an EMPTY buffer if allocating 1804 * its KVA would not otherwise run us out of buffer space. 1805 */ 1806 if (nbp == NULL && defrag == 0 && 1807 bufspace + maxsize < hibufspace) { 1808 nqindex = QUEUE_EMPTY; 1809 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1810 } 1811 } 1812 1813 /* 1814 * Run scan, possibly freeing data and/or kva mappings on the fly 1815 * depending. 1816 */ 1817 1818 while ((bp = nbp) != NULL) { 1819 int qindex = nqindex; 1820 1821 /* 1822 * Calculate next bp ( we can only use it if we do not block 1823 * or do other fancy things ). 1824 */ 1825 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1826 switch(qindex) { 1827 case QUEUE_EMPTY: 1828 nqindex = QUEUE_EMPTYKVA; 1829 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1830 break; 1831 /* FALLTHROUGH */ 1832 case QUEUE_EMPTYKVA: 1833 nqindex = QUEUE_CLEAN; 1834 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1835 break; 1836 /* FALLTHROUGH */ 1837 case QUEUE_CLEAN: 1838 /* 1839 * nbp is NULL. 1840 */ 1841 break; 1842 } 1843 } 1844 1845 /* 1846 * Sanity Checks 1847 */ 1848 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1849 1850 /* 1851 * Note: we no longer distinguish between VMIO and non-VMIO 1852 * buffers. 1853 */ 1854 1855 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1856 1857 /* 1858 * If we are defragging then we need a buffer with 1859 * b_kvasize != 0. XXX this situation should no longer 1860 * occur, if defrag is non-zero the buffer's b_kvasize 1861 * should also be non-zero at this point. XXX 1862 */ 1863 if (defrag && bp->b_kvasize == 0) { 1864 printf("Warning: defrag empty buffer %p\n", bp); 1865 continue; 1866 } 1867 1868 /* 1869 * Start freeing the bp. This is somewhat involved. nbp 1870 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1871 */ 1872 1873 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1874 panic("getnewbuf: locked buf"); 1875 bremfreel(bp); 1876 mtx_unlock(&bqlock); 1877 1878 if (qindex == QUEUE_CLEAN) { 1879 if (bp->b_flags & B_VMIO) { 1880 bp->b_flags &= ~B_ASYNC; 1881 vfs_vmio_release(bp); 1882 } 1883 if (bp->b_vp) 1884 brelvp(bp); 1885 } 1886 1887 /* 1888 * NOTE: nbp is now entirely invalid. We can only restart 1889 * the scan from this point on. 1890 * 1891 * Get the rest of the buffer freed up. b_kva* is still 1892 * valid after this operation. 1893 */ 1894 1895 if (bp->b_rcred != NOCRED) { 1896 crfree(bp->b_rcred); 1897 bp->b_rcred = NOCRED; 1898 } 1899 if (bp->b_wcred != NOCRED) { 1900 crfree(bp->b_wcred); 1901 bp->b_wcred = NOCRED; 1902 } 1903 if (LIST_FIRST(&bp->b_dep) != NULL) 1904 buf_deallocate(bp); 1905 if (bp->b_xflags & BX_BKGRDINPROG) 1906 panic("losing buffer 3"); 1907#ifdef USE_BUFHASH 1908 LIST_REMOVE(bp, b_hash); 1909 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1910#endif 1911 1912 if (bp->b_bufsize) 1913 allocbuf(bp, 0); 1914 1915 bp->b_flags = 0; 1916 bp->b_ioflags = 0; 1917 bp->b_xflags = 0; 1918 bp->b_vflags = 0; 1919 bp->b_dev = NODEV; 1920 bp->b_vp = NULL; 1921 bp->b_blkno = bp->b_lblkno = 0; 1922 bp->b_offset = NOOFFSET; 1923 bp->b_iodone = 0; 1924 bp->b_error = 0; 1925 bp->b_resid = 0; 1926 bp->b_bcount = 0; 1927 bp->b_npages = 0; 1928 bp->b_dirtyoff = bp->b_dirtyend = 0; 1929 bp->b_magic = B_MAGIC_BIO; 1930 bp->b_op = &buf_ops_bio; 1931 bp->b_object = NULL; 1932 1933 LIST_INIT(&bp->b_dep); 1934 1935 /* 1936 * If we are defragging then free the buffer. 1937 */ 1938 if (defrag) { 1939 bp->b_flags |= B_INVAL; 1940 bfreekva(bp); 1941 brelse(bp); 1942 defrag = 0; 1943 goto restart; 1944 } 1945 1946 /* 1947 * If we are overcomitted then recover the buffer and its 1948 * KVM space. This occurs in rare situations when multiple 1949 * processes are blocked in getnewbuf() or allocbuf(). 1950 */ 1951 if (bufspace >= hibufspace) 1952 flushingbufs = 1; 1953 if (flushingbufs && bp->b_kvasize != 0) { 1954 bp->b_flags |= B_INVAL; 1955 bfreekva(bp); 1956 brelse(bp); 1957 goto restart; 1958 } 1959 if (bufspace < lobufspace) 1960 flushingbufs = 0; 1961 break; 1962 } 1963 1964 /* 1965 * If we exhausted our list, sleep as appropriate. We may have to 1966 * wakeup various daemons and write out some dirty buffers. 1967 * 1968 * Generally we are sleeping due to insufficient buffer space. 1969 */ 1970 1971 if (bp == NULL) { 1972 int flags; 1973 char *waitmsg; 1974 1975 mtx_unlock(&bqlock); 1976 if (defrag) { 1977 flags = VFS_BIO_NEED_BUFSPACE; 1978 waitmsg = "nbufkv"; 1979 } else if (bufspace >= hibufspace) { 1980 waitmsg = "nbufbs"; 1981 flags = VFS_BIO_NEED_BUFSPACE; 1982 } else { 1983 waitmsg = "newbuf"; 1984 flags = VFS_BIO_NEED_ANY; 1985 } 1986 1987 bd_speedup(); /* heeeelp */ 1988 1989 mtx_lock(&nblock); 1990 needsbuffer |= flags; 1991 while (needsbuffer & flags) { 1992 if (msleep(&needsbuffer, &nblock, 1993 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 1994 mtx_unlock(&nblock); 1995 return (NULL); 1996 } 1997 } 1998 mtx_unlock(&nblock); 1999 } else { 2000 /* 2001 * We finally have a valid bp. We aren't quite out of the 2002 * woods, we still have to reserve kva space. In order 2003 * to keep fragmentation sane we only allocate kva in 2004 * BKVASIZE chunks. 2005 */ 2006 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2007 2008 if (maxsize != bp->b_kvasize) { 2009 vm_offset_t addr = 0; 2010 2011 bfreekva(bp); 2012 2013 if (vm_map_findspace(buffer_map, 2014 vm_map_min(buffer_map), maxsize, &addr)) { 2015 /* 2016 * Uh oh. Buffer map is to fragmented. We 2017 * must defragment the map. 2018 */ 2019 atomic_add_int(&bufdefragcnt, 1); 2020 defrag = 1; 2021 bp->b_flags |= B_INVAL; 2022 brelse(bp); 2023 goto restart; 2024 } 2025 if (addr) { 2026 vm_map_insert(buffer_map, NULL, 0, 2027 addr, addr + maxsize, 2028 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2029 2030 bp->b_kvabase = (caddr_t) addr; 2031 bp->b_kvasize = maxsize; 2032 atomic_add_int(&bufspace, bp->b_kvasize); 2033 atomic_add_int(&bufreusecnt, 1); 2034 } 2035 } 2036 bp->b_data = bp->b_kvabase; 2037 } 2038 return(bp); 2039} 2040 2041/* 2042 * buf_daemon: 2043 * 2044 * buffer flushing daemon. Buffers are normally flushed by the 2045 * update daemon but if it cannot keep up this process starts to 2046 * take the load in an attempt to prevent getnewbuf() from blocking. 2047 */ 2048 2049static struct proc *bufdaemonproc; 2050 2051static struct kproc_desc buf_kp = { 2052 "bufdaemon", 2053 buf_daemon, 2054 &bufdaemonproc 2055}; 2056SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2057 2058static void 2059buf_daemon() 2060{ 2061 int s; 2062 2063 mtx_lock(&Giant); 2064 2065 /* 2066 * This process needs to be suspended prior to shutdown sync. 2067 */ 2068 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2069 SHUTDOWN_PRI_LAST); 2070 2071 /* 2072 * This process is allowed to take the buffer cache to the limit 2073 */ 2074 s = splbio(); 2075 mtx_lock(&bdlock); 2076 2077 for (;;) { 2078 bd_request = 0; 2079 mtx_unlock(&bdlock); 2080 2081 kthread_suspend_check(bufdaemonproc); 2082 2083 /* 2084 * Do the flush. Limit the amount of in-transit I/O we 2085 * allow to build up, otherwise we would completely saturate 2086 * the I/O system. Wakeup any waiting processes before we 2087 * normally would so they can run in parallel with our drain. 2088 */ 2089 while (numdirtybuffers > lodirtybuffers) { 2090 if (flushbufqueues() == 0) 2091 break; 2092 waitrunningbufspace(); 2093 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2094 } 2095 2096 /* 2097 * Only clear bd_request if we have reached our low water 2098 * mark. The buf_daemon normally waits 1 second and 2099 * then incrementally flushes any dirty buffers that have 2100 * built up, within reason. 2101 * 2102 * If we were unable to hit our low water mark and couldn't 2103 * find any flushable buffers, we sleep half a second. 2104 * Otherwise we loop immediately. 2105 */ 2106 mtx_lock(&bdlock); 2107 if (numdirtybuffers <= lodirtybuffers) { 2108 /* 2109 * We reached our low water mark, reset the 2110 * request and sleep until we are needed again. 2111 * The sleep is just so the suspend code works. 2112 */ 2113 bd_request = 0; 2114 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2115 } else { 2116 /* 2117 * We couldn't find any flushable dirty buffers but 2118 * still have too many dirty buffers, we 2119 * have to sleep and try again. (rare) 2120 */ 2121 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2122 } 2123 } 2124} 2125 2126/* 2127 * flushbufqueues: 2128 * 2129 * Try to flush a buffer in the dirty queue. We must be careful to 2130 * free up B_INVAL buffers instead of write them, which NFS is 2131 * particularly sensitive to. 2132 */ 2133int flushwithdeps = 0; 2134SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2135 0, "Number of buffers flushed with dependecies that require rollbacks"); 2136static int 2137flushbufqueues(void) 2138{ 2139 struct thread *td = curthread; 2140 struct vnode *vp; 2141 struct buf *bp; 2142 2143 mtx_lock(&bqlock); 2144 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2145 KASSERT((bp->b_flags & B_DELWRI), 2146 ("unexpected clean buffer %p", bp)); 2147 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) 2148 continue; 2149 if (bp->b_flags & B_INVAL) { 2150 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 2151 panic("flushbufqueues: locked buf"); 2152 bremfreel(bp); 2153 mtx_unlock(&bqlock); 2154 brelse(bp); 2155 return (1); 2156 } 2157 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) 2158 continue; 2159 /* 2160 * We must hold the lock on a vnode before writing 2161 * one of its buffers. Otherwise we may confuse, or 2162 * in the case of a snapshot vnode, deadlock the 2163 * system. 2164 */ 2165 if ((vp = bp->b_vp) == NULL || 2166 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2167 mtx_unlock(&bqlock); 2168 vfs_bio_awrite(bp); 2169 if (vp != NULL) 2170 VOP_UNLOCK(vp, 0, td); 2171 return (1); 2172 } 2173 } 2174 /* 2175 * Could not find any buffers without rollback dependencies, 2176 * so just write the first one in the hopes of eventually 2177 * making progress. 2178 */ 2179 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2180 KASSERT((bp->b_flags & B_DELWRI), 2181 ("unexpected clean buffer %p", bp)); 2182 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) 2183 continue; 2184 if (bp->b_flags & B_INVAL) { 2185 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 2186 panic("flushbufqueues: locked buf"); 2187 bremfreel(bp); 2188 mtx_unlock(&bqlock); 2189 brelse(bp); 2190 return (1); 2191 } 2192 /* 2193 * We must hold the lock on a vnode before writing 2194 * one of its buffers. Otherwise we may confuse, or 2195 * in the case of a snapshot vnode, deadlock the 2196 * system. 2197 */ 2198 if ((vp = bp->b_vp) == NULL || 2199 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2200 mtx_unlock(&bqlock); 2201 vfs_bio_awrite(bp); 2202 if (vp != NULL) 2203 VOP_UNLOCK(vp, 0, td); 2204 flushwithdeps += 1; 2205 return (0); 2206 } 2207 } 2208 mtx_unlock(&bqlock); 2209 return (0); 2210} 2211 2212/* 2213 * Check to see if a block is currently memory resident. 2214 */ 2215struct buf * 2216incore(struct vnode * vp, daddr_t blkno) 2217{ 2218 struct buf *bp; 2219 2220 int s = splbio(); 2221 VI_LOCK(vp); 2222 bp = gbincore(vp, blkno); 2223 VI_UNLOCK(vp); 2224 splx(s); 2225 return (bp); 2226} 2227 2228/* 2229 * Returns true if no I/O is needed to access the 2230 * associated VM object. This is like incore except 2231 * it also hunts around in the VM system for the data. 2232 */ 2233 2234int 2235inmem(struct vnode * vp, daddr_t blkno) 2236{ 2237 vm_object_t obj; 2238 vm_offset_t toff, tinc, size; 2239 vm_page_t m; 2240 vm_ooffset_t off; 2241 2242 GIANT_REQUIRED; 2243 ASSERT_VOP_LOCKED(vp, "inmem"); 2244 2245 if (incore(vp, blkno)) 2246 return 1; 2247 if (vp->v_mount == NULL) 2248 return 0; 2249 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2250 return 0; 2251 2252 size = PAGE_SIZE; 2253 if (size > vp->v_mount->mnt_stat.f_iosize) 2254 size = vp->v_mount->mnt_stat.f_iosize; 2255 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2256 2257 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2258 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2259 if (!m) 2260 goto notinmem; 2261 tinc = size; 2262 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2263 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2264 if (vm_page_is_valid(m, 2265 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2266 goto notinmem; 2267 } 2268 return 1; 2269 2270notinmem: 2271 return (0); 2272} 2273 2274/* 2275 * vfs_setdirty: 2276 * 2277 * Sets the dirty range for a buffer based on the status of the dirty 2278 * bits in the pages comprising the buffer. 2279 * 2280 * The range is limited to the size of the buffer. 2281 * 2282 * This routine is primarily used by NFS, but is generalized for the 2283 * B_VMIO case. 2284 */ 2285static void 2286vfs_setdirty(struct buf *bp) 2287{ 2288 int i; 2289 vm_object_t object; 2290 2291 GIANT_REQUIRED; 2292 /* 2293 * Degenerate case - empty buffer 2294 */ 2295 2296 if (bp->b_bufsize == 0) 2297 return; 2298 2299 /* 2300 * We qualify the scan for modified pages on whether the 2301 * object has been flushed yet. The OBJ_WRITEABLE flag 2302 * is not cleared simply by protecting pages off. 2303 */ 2304 2305 if ((bp->b_flags & B_VMIO) == 0) 2306 return; 2307 2308 object = bp->b_pages[0]->object; 2309 2310 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2311 printf("Warning: object %p writeable but not mightbedirty\n", object); 2312 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2313 printf("Warning: object %p mightbedirty but not writeable\n", object); 2314 2315 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2316 vm_offset_t boffset; 2317 vm_offset_t eoffset; 2318 2319 vm_page_lock_queues(); 2320 /* 2321 * test the pages to see if they have been modified directly 2322 * by users through the VM system. 2323 */ 2324 for (i = 0; i < bp->b_npages; i++) { 2325 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2326 vm_page_test_dirty(bp->b_pages[i]); 2327 } 2328 2329 /* 2330 * Calculate the encompassing dirty range, boffset and eoffset, 2331 * (eoffset - boffset) bytes. 2332 */ 2333 2334 for (i = 0; i < bp->b_npages; i++) { 2335 if (bp->b_pages[i]->dirty) 2336 break; 2337 } 2338 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2339 2340 for (i = bp->b_npages - 1; i >= 0; --i) { 2341 if (bp->b_pages[i]->dirty) { 2342 break; 2343 } 2344 } 2345 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2346 2347 vm_page_unlock_queues(); 2348 /* 2349 * Fit it to the buffer. 2350 */ 2351 2352 if (eoffset > bp->b_bcount) 2353 eoffset = bp->b_bcount; 2354 2355 /* 2356 * If we have a good dirty range, merge with the existing 2357 * dirty range. 2358 */ 2359 2360 if (boffset < eoffset) { 2361 if (bp->b_dirtyoff > boffset) 2362 bp->b_dirtyoff = boffset; 2363 if (bp->b_dirtyend < eoffset) 2364 bp->b_dirtyend = eoffset; 2365 } 2366 } 2367} 2368 2369/* 2370 * getblk: 2371 * 2372 * Get a block given a specified block and offset into a file/device. 2373 * The buffers B_DONE bit will be cleared on return, making it almost 2374 * ready for an I/O initiation. B_INVAL may or may not be set on 2375 * return. The caller should clear B_INVAL prior to initiating a 2376 * READ. 2377 * 2378 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2379 * an existing buffer. 2380 * 2381 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2382 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2383 * and then cleared based on the backing VM. If the previous buffer is 2384 * non-0-sized but invalid, B_CACHE will be cleared. 2385 * 2386 * If getblk() must create a new buffer, the new buffer is returned with 2387 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2388 * case it is returned with B_INVAL clear and B_CACHE set based on the 2389 * backing VM. 2390 * 2391 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2392 * B_CACHE bit is clear. 2393 * 2394 * What this means, basically, is that the caller should use B_CACHE to 2395 * determine whether the buffer is fully valid or not and should clear 2396 * B_INVAL prior to issuing a read. If the caller intends to validate 2397 * the buffer by loading its data area with something, the caller needs 2398 * to clear B_INVAL. If the caller does this without issuing an I/O, 2399 * the caller should set B_CACHE ( as an optimization ), else the caller 2400 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2401 * a write attempt or if it was a successfull read. If the caller 2402 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2403 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2404 */ 2405struct buf * 2406getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2407{ 2408 struct buf *bp; 2409 int s; 2410#ifdef USE_BUFHASH 2411 struct bufhashhdr *bh; 2412#endif 2413 ASSERT_VOP_LOCKED(vp, "getblk"); 2414 2415 if (size > MAXBSIZE) 2416 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2417 2418 s = splbio(); 2419loop: 2420 /* 2421 * Block if we are low on buffers. Certain processes are allowed 2422 * to completely exhaust the buffer cache. 2423 * 2424 * If this check ever becomes a bottleneck it may be better to 2425 * move it into the else, when gbincore() fails. At the moment 2426 * it isn't a problem. 2427 * 2428 * XXX remove if 0 sections (clean this up after its proven) 2429 */ 2430 if (numfreebuffers == 0) { 2431 if (curthread == PCPU_GET(idlethread)) 2432 return NULL; 2433 mtx_lock(&nblock); 2434 needsbuffer |= VFS_BIO_NEED_ANY; 2435 mtx_unlock(&nblock); 2436 } 2437 2438 VI_LOCK(vp); 2439 if ((bp = gbincore(vp, blkno))) { 2440 VI_UNLOCK(vp); 2441 /* 2442 * Buffer is in-core. If the buffer is not busy, it must 2443 * be on a queue. 2444 */ 2445 2446 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2447 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2448 "getblk", slpflag, slptimeo) == ENOLCK) 2449 goto loop; 2450 splx(s); 2451 return (struct buf *) NULL; 2452 } 2453 2454 /* 2455 * The buffer is locked. B_CACHE is cleared if the buffer is 2456 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2457 * and for a VMIO buffer B_CACHE is adjusted according to the 2458 * backing VM cache. 2459 */ 2460 if (bp->b_flags & B_INVAL) 2461 bp->b_flags &= ~B_CACHE; 2462 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2463 bp->b_flags |= B_CACHE; 2464 bremfree(bp); 2465 2466 /* 2467 * check for size inconsistancies for non-VMIO case. 2468 */ 2469 2470 if (bp->b_bcount != size) { 2471 if ((bp->b_flags & B_VMIO) == 0 || 2472 (size > bp->b_kvasize)) { 2473 if (bp->b_flags & B_DELWRI) { 2474 bp->b_flags |= B_NOCACHE; 2475 BUF_WRITE(bp); 2476 } else { 2477 if ((bp->b_flags & B_VMIO) && 2478 (LIST_FIRST(&bp->b_dep) == NULL)) { 2479 bp->b_flags |= B_RELBUF; 2480 brelse(bp); 2481 } else { 2482 bp->b_flags |= B_NOCACHE; 2483 BUF_WRITE(bp); 2484 } 2485 } 2486 goto loop; 2487 } 2488 } 2489 2490 /* 2491 * If the size is inconsistant in the VMIO case, we can resize 2492 * the buffer. This might lead to B_CACHE getting set or 2493 * cleared. If the size has not changed, B_CACHE remains 2494 * unchanged from its previous state. 2495 */ 2496 2497 if (bp->b_bcount != size) 2498 allocbuf(bp, size); 2499 2500 KASSERT(bp->b_offset != NOOFFSET, 2501 ("getblk: no buffer offset")); 2502 2503 /* 2504 * A buffer with B_DELWRI set and B_CACHE clear must 2505 * be committed before we can return the buffer in 2506 * order to prevent the caller from issuing a read 2507 * ( due to B_CACHE not being set ) and overwriting 2508 * it. 2509 * 2510 * Most callers, including NFS and FFS, need this to 2511 * operate properly either because they assume they 2512 * can issue a read if B_CACHE is not set, or because 2513 * ( for example ) an uncached B_DELWRI might loop due 2514 * to softupdates re-dirtying the buffer. In the latter 2515 * case, B_CACHE is set after the first write completes, 2516 * preventing further loops. 2517 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2518 * above while extending the buffer, we cannot allow the 2519 * buffer to remain with B_CACHE set after the write 2520 * completes or it will represent a corrupt state. To 2521 * deal with this we set B_NOCACHE to scrap the buffer 2522 * after the write. 2523 * 2524 * We might be able to do something fancy, like setting 2525 * B_CACHE in bwrite() except if B_DELWRI is already set, 2526 * so the below call doesn't set B_CACHE, but that gets real 2527 * confusing. This is much easier. 2528 */ 2529 2530 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2531 bp->b_flags |= B_NOCACHE; 2532 BUF_WRITE(bp); 2533 goto loop; 2534 } 2535 2536 splx(s); 2537 bp->b_flags &= ~B_DONE; 2538 } else { 2539 int bsize, maxsize, vmio; 2540 off_t offset; 2541 2542 /* 2543 * Buffer is not in-core, create new buffer. The buffer 2544 * returned by getnewbuf() is locked. Note that the returned 2545 * buffer is also considered valid (not marked B_INVAL). 2546 */ 2547 VI_UNLOCK(vp); 2548 if (vn_isdisk(vp, NULL)) 2549 bsize = DEV_BSIZE; 2550 else if (vp->v_mountedhere) 2551 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2552 else if (vp->v_mount) 2553 bsize = vp->v_mount->mnt_stat.f_iosize; 2554 else 2555 bsize = size; 2556 2557 offset = blkno * bsize; 2558 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2559 (vp->v_vflag & VV_OBJBUF); 2560 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2561 maxsize = imax(maxsize, bsize); 2562 2563 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2564 if (slpflag || slptimeo) { 2565 splx(s); 2566 return NULL; 2567 } 2568 goto loop; 2569 } 2570 2571 /* 2572 * This code is used to make sure that a buffer is not 2573 * created while the getnewbuf routine is blocked. 2574 * This can be a problem whether the vnode is locked or not. 2575 * If the buffer is created out from under us, we have to 2576 * throw away the one we just created. There is now window 2577 * race because we are safely running at splbio() from the 2578 * point of the duplicate buffer creation through to here, 2579 * and we've locked the buffer. 2580 * 2581 * Note: this must occur before we associate the buffer 2582 * with the vp especially considering limitations in 2583 * the splay tree implementation when dealing with duplicate 2584 * lblkno's. 2585 */ 2586 VI_LOCK(vp); 2587 if (gbincore(vp, blkno)) { 2588 VI_UNLOCK(vp); 2589 bp->b_flags |= B_INVAL; 2590 brelse(bp); 2591 goto loop; 2592 } 2593 VI_UNLOCK(vp); 2594 2595 /* 2596 * Insert the buffer into the hash, so that it can 2597 * be found by incore. 2598 */ 2599 bp->b_blkno = bp->b_lblkno = blkno; 2600 bp->b_offset = offset; 2601 2602 bgetvp(vp, bp); 2603#ifdef USE_BUFHASH 2604 LIST_REMOVE(bp, b_hash); 2605 bh = bufhash(vp, blkno); 2606 LIST_INSERT_HEAD(bh, bp, b_hash); 2607#endif 2608 2609 /* 2610 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2611 * buffer size starts out as 0, B_CACHE will be set by 2612 * allocbuf() for the VMIO case prior to it testing the 2613 * backing store for validity. 2614 */ 2615 2616 if (vmio) { 2617 bp->b_flags |= B_VMIO; 2618#if defined(VFS_BIO_DEBUG) 2619 if (vp->v_type != VREG) 2620 printf("getblk: vmioing file type %d???\n", vp->v_type); 2621#endif 2622 VOP_GETVOBJECT(vp, &bp->b_object); 2623 } else { 2624 bp->b_flags &= ~B_VMIO; 2625 bp->b_object = NULL; 2626 } 2627 2628 allocbuf(bp, size); 2629 2630 splx(s); 2631 bp->b_flags &= ~B_DONE; 2632 } 2633 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2634 return (bp); 2635} 2636 2637/* 2638 * Get an empty, disassociated buffer of given size. The buffer is initially 2639 * set to B_INVAL. 2640 */ 2641struct buf * 2642geteblk(int size) 2643{ 2644 struct buf *bp; 2645 int s; 2646 int maxsize; 2647 2648 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2649 2650 s = splbio(); 2651 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2652 continue; 2653 splx(s); 2654 allocbuf(bp, size); 2655 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2656 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2657 return (bp); 2658} 2659 2660 2661/* 2662 * This code constitutes the buffer memory from either anonymous system 2663 * memory (in the case of non-VMIO operations) or from an associated 2664 * VM object (in the case of VMIO operations). This code is able to 2665 * resize a buffer up or down. 2666 * 2667 * Note that this code is tricky, and has many complications to resolve 2668 * deadlock or inconsistant data situations. Tread lightly!!! 2669 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2670 * the caller. Calling this code willy nilly can result in the loss of data. 2671 * 2672 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2673 * B_CACHE for the non-VMIO case. 2674 */ 2675 2676int 2677allocbuf(struct buf *bp, int size) 2678{ 2679 int newbsize, mbsize; 2680 int i; 2681 2682 GIANT_REQUIRED; 2683 2684 if (BUF_REFCNT(bp) == 0) 2685 panic("allocbuf: buffer not busy"); 2686 2687 if (bp->b_kvasize < size) 2688 panic("allocbuf: buffer too small"); 2689 2690 if ((bp->b_flags & B_VMIO) == 0) { 2691 caddr_t origbuf; 2692 int origbufsize; 2693 /* 2694 * Just get anonymous memory from the kernel. Don't 2695 * mess with B_CACHE. 2696 */ 2697 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2698 if (bp->b_flags & B_MALLOC) 2699 newbsize = mbsize; 2700 else 2701 newbsize = round_page(size); 2702 2703 if (newbsize < bp->b_bufsize) { 2704 /* 2705 * malloced buffers are not shrunk 2706 */ 2707 if (bp->b_flags & B_MALLOC) { 2708 if (newbsize) { 2709 bp->b_bcount = size; 2710 } else { 2711 free(bp->b_data, M_BIOBUF); 2712 if (bp->b_bufsize) { 2713 atomic_subtract_int( 2714 &bufmallocspace, 2715 bp->b_bufsize); 2716 bufspacewakeup(); 2717 bp->b_bufsize = 0; 2718 } 2719 bp->b_data = bp->b_kvabase; 2720 bp->b_bcount = 0; 2721 bp->b_flags &= ~B_MALLOC; 2722 } 2723 return 1; 2724 } 2725 vm_hold_free_pages( 2726 bp, 2727 (vm_offset_t) bp->b_data + newbsize, 2728 (vm_offset_t) bp->b_data + bp->b_bufsize); 2729 } else if (newbsize > bp->b_bufsize) { 2730 /* 2731 * We only use malloced memory on the first allocation. 2732 * and revert to page-allocated memory when the buffer 2733 * grows. 2734 */ 2735 /* 2736 * There is a potential smp race here that could lead 2737 * to bufmallocspace slightly passing the max. It 2738 * is probably extremely rare and not worth worrying 2739 * over. 2740 */ 2741 if ( (bufmallocspace < maxbufmallocspace) && 2742 (bp->b_bufsize == 0) && 2743 (mbsize <= PAGE_SIZE/2)) { 2744 2745 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2746 bp->b_bufsize = mbsize; 2747 bp->b_bcount = size; 2748 bp->b_flags |= B_MALLOC; 2749 atomic_add_int(&bufmallocspace, mbsize); 2750 return 1; 2751 } 2752 origbuf = NULL; 2753 origbufsize = 0; 2754 /* 2755 * If the buffer is growing on its other-than-first allocation, 2756 * then we revert to the page-allocation scheme. 2757 */ 2758 if (bp->b_flags & B_MALLOC) { 2759 origbuf = bp->b_data; 2760 origbufsize = bp->b_bufsize; 2761 bp->b_data = bp->b_kvabase; 2762 if (bp->b_bufsize) { 2763 atomic_subtract_int(&bufmallocspace, 2764 bp->b_bufsize); 2765 bufspacewakeup(); 2766 bp->b_bufsize = 0; 2767 } 2768 bp->b_flags &= ~B_MALLOC; 2769 newbsize = round_page(newbsize); 2770 } 2771 vm_hold_load_pages( 2772 bp, 2773 (vm_offset_t) bp->b_data + bp->b_bufsize, 2774 (vm_offset_t) bp->b_data + newbsize); 2775 if (origbuf) { 2776 bcopy(origbuf, bp->b_data, origbufsize); 2777 free(origbuf, M_BIOBUF); 2778 } 2779 } 2780 } else { 2781 int desiredpages; 2782 2783 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2784 desiredpages = (size == 0) ? 0 : 2785 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2786 2787 if (bp->b_flags & B_MALLOC) 2788 panic("allocbuf: VMIO buffer can't be malloced"); 2789 /* 2790 * Set B_CACHE initially if buffer is 0 length or will become 2791 * 0-length. 2792 */ 2793 if (size == 0 || bp->b_bufsize == 0) 2794 bp->b_flags |= B_CACHE; 2795 2796 if (newbsize < bp->b_bufsize) { 2797 /* 2798 * DEV_BSIZE aligned new buffer size is less then the 2799 * DEV_BSIZE aligned existing buffer size. Figure out 2800 * if we have to remove any pages. 2801 */ 2802 if (desiredpages < bp->b_npages) { 2803 vm_page_t m; 2804 2805 vm_page_lock_queues(); 2806 for (i = desiredpages; i < bp->b_npages; i++) { 2807 /* 2808 * the page is not freed here -- it 2809 * is the responsibility of 2810 * vnode_pager_setsize 2811 */ 2812 m = bp->b_pages[i]; 2813 KASSERT(m != bogus_page, 2814 ("allocbuf: bogus page found")); 2815 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2816 vm_page_lock_queues(); 2817 2818 bp->b_pages[i] = NULL; 2819 vm_page_unwire(m, 0); 2820 } 2821 vm_page_unlock_queues(); 2822 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2823 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2824 bp->b_npages = desiredpages; 2825 } 2826 } else if (size > bp->b_bcount) { 2827 /* 2828 * We are growing the buffer, possibly in a 2829 * byte-granular fashion. 2830 */ 2831 struct vnode *vp; 2832 vm_object_t obj; 2833 vm_offset_t toff; 2834 vm_offset_t tinc; 2835 2836 /* 2837 * Step 1, bring in the VM pages from the object, 2838 * allocating them if necessary. We must clear 2839 * B_CACHE if these pages are not valid for the 2840 * range covered by the buffer. 2841 */ 2842 2843 vp = bp->b_vp; 2844 obj = bp->b_object; 2845 2846 while (bp->b_npages < desiredpages) { 2847 vm_page_t m; 2848 vm_pindex_t pi; 2849 2850 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2851 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2852 /* 2853 * note: must allocate system pages 2854 * since blocking here could intefere 2855 * with paging I/O, no matter which 2856 * process we are. 2857 */ 2858 m = vm_page_alloc(obj, pi, 2859 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2860 if (m == NULL) { 2861 atomic_add_int(&vm_pageout_deficit, 2862 desiredpages - bp->b_npages); 2863 VM_WAIT; 2864 } else { 2865 vm_page_lock_queues(); 2866 vm_page_wakeup(m); 2867 vm_page_unlock_queues(); 2868 bp->b_flags &= ~B_CACHE; 2869 bp->b_pages[bp->b_npages] = m; 2870 ++bp->b_npages; 2871 } 2872 continue; 2873 } 2874 2875 /* 2876 * We found a page. If we have to sleep on it, 2877 * retry because it might have gotten freed out 2878 * from under us. 2879 * 2880 * We can only test PG_BUSY here. Blocking on 2881 * m->busy might lead to a deadlock: 2882 * 2883 * vm_fault->getpages->cluster_read->allocbuf 2884 * 2885 */ 2886 vm_page_lock_queues(); 2887 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2888 continue; 2889 2890 /* 2891 * We have a good page. Should we wakeup the 2892 * page daemon? 2893 */ 2894 if ((curproc != pageproc) && 2895 ((m->queue - m->pc) == PQ_CACHE) && 2896 ((cnt.v_free_count + cnt.v_cache_count) < 2897 (cnt.v_free_min + cnt.v_cache_min))) { 2898 pagedaemon_wakeup(); 2899 } 2900 vm_page_flag_clear(m, PG_ZERO); 2901 vm_page_wire(m); 2902 vm_page_unlock_queues(); 2903 bp->b_pages[bp->b_npages] = m; 2904 ++bp->b_npages; 2905 } 2906 2907 /* 2908 * Step 2. We've loaded the pages into the buffer, 2909 * we have to figure out if we can still have B_CACHE 2910 * set. Note that B_CACHE is set according to the 2911 * byte-granular range ( bcount and size ), new the 2912 * aligned range ( newbsize ). 2913 * 2914 * The VM test is against m->valid, which is DEV_BSIZE 2915 * aligned. Needless to say, the validity of the data 2916 * needs to also be DEV_BSIZE aligned. Note that this 2917 * fails with NFS if the server or some other client 2918 * extends the file's EOF. If our buffer is resized, 2919 * B_CACHE may remain set! XXX 2920 */ 2921 2922 toff = bp->b_bcount; 2923 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2924 2925 while ((bp->b_flags & B_CACHE) && toff < size) { 2926 vm_pindex_t pi; 2927 2928 if (tinc > (size - toff)) 2929 tinc = size - toff; 2930 2931 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2932 PAGE_SHIFT; 2933 2934 vfs_buf_test_cache( 2935 bp, 2936 bp->b_offset, 2937 toff, 2938 tinc, 2939 bp->b_pages[pi] 2940 ); 2941 toff += tinc; 2942 tinc = PAGE_SIZE; 2943 } 2944 2945 /* 2946 * Step 3, fixup the KVM pmap. Remember that 2947 * bp->b_data is relative to bp->b_offset, but 2948 * bp->b_offset may be offset into the first page. 2949 */ 2950 2951 bp->b_data = (caddr_t) 2952 trunc_page((vm_offset_t)bp->b_data); 2953 pmap_qenter( 2954 (vm_offset_t)bp->b_data, 2955 bp->b_pages, 2956 bp->b_npages 2957 ); 2958 2959 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2960 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2961 } 2962 } 2963 if (newbsize < bp->b_bufsize) 2964 bufspacewakeup(); 2965 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2966 bp->b_bcount = size; /* requested buffer size */ 2967 return 1; 2968} 2969 2970void 2971biodone(struct bio *bp) 2972{ 2973 bp->bio_flags |= BIO_DONE; 2974 if (bp->bio_done != NULL) 2975 bp->bio_done(bp); 2976 else 2977 wakeup(bp); 2978} 2979 2980/* 2981 * Wait for a BIO to finish. 2982 * 2983 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2984 * case is not yet clear. 2985 */ 2986int 2987biowait(struct bio *bp, const char *wchan) 2988{ 2989 2990 while ((bp->bio_flags & BIO_DONE) == 0) 2991 msleep(bp, NULL, PRIBIO, wchan, hz / 10); 2992 if (bp->bio_error != 0) 2993 return (bp->bio_error); 2994 if (!(bp->bio_flags & BIO_ERROR)) 2995 return (0); 2996 return (EIO); 2997} 2998 2999void 3000biofinish(struct bio *bp, struct devstat *stat, int error) 3001{ 3002 3003 if (error) { 3004 bp->bio_error = error; 3005 bp->bio_flags |= BIO_ERROR; 3006 } 3007 if (stat != NULL) 3008 devstat_end_transaction_bio(stat, bp); 3009 biodone(bp); 3010} 3011 3012void 3013bioq_init(struct bio_queue_head *head) 3014{ 3015 TAILQ_INIT(&head->queue); 3016 head->last_pblkno = 0; 3017 head->insert_point = NULL; 3018 head->switch_point = NULL; 3019} 3020 3021void 3022bioq_remove(struct bio_queue_head *head, struct bio *bp) 3023{ 3024 if (bp == head->switch_point) 3025 head->switch_point = TAILQ_NEXT(bp, bio_queue); 3026 if (bp == head->insert_point) { 3027 head->insert_point = TAILQ_PREV(bp, bio_queue, bio_queue); 3028 if (head->insert_point == NULL) 3029 head->last_pblkno = 0; 3030 } else if (bp == TAILQ_FIRST(&head->queue)) 3031 head->last_pblkno = bp->bio_pblkno; 3032 TAILQ_REMOVE(&head->queue, bp, bio_queue); 3033 if (TAILQ_FIRST(&head->queue) == head->switch_point) 3034 head->switch_point = NULL; 3035} 3036 3037/* 3038 * bufwait: 3039 * 3040 * Wait for buffer I/O completion, returning error status. The buffer 3041 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3042 * error and cleared. 3043 */ 3044int 3045bufwait(register struct buf * bp) 3046{ 3047 int s; 3048 3049 s = splbio(); 3050 while ((bp->b_flags & B_DONE) == 0) { 3051 if (bp->b_iocmd == BIO_READ) 3052 tsleep(bp, PRIBIO, "biord", 0); 3053 else 3054 tsleep(bp, PRIBIO, "biowr", 0); 3055 } 3056 splx(s); 3057 if (bp->b_flags & B_EINTR) { 3058 bp->b_flags &= ~B_EINTR; 3059 return (EINTR); 3060 } 3061 if (bp->b_ioflags & BIO_ERROR) { 3062 return (bp->b_error ? bp->b_error : EIO); 3063 } else { 3064 return (0); 3065 } 3066} 3067 3068 /* 3069 * Call back function from struct bio back up to struct buf. 3070 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 3071 */ 3072void 3073bufdonebio(struct bio *bp) 3074{ 3075 bufdone(bp->bio_caller2); 3076} 3077 3078/* 3079 * bufdone: 3080 * 3081 * Finish I/O on a buffer, optionally calling a completion function. 3082 * This is usually called from an interrupt so process blocking is 3083 * not allowed. 3084 * 3085 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3086 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3087 * assuming B_INVAL is clear. 3088 * 3089 * For the VMIO case, we set B_CACHE if the op was a read and no 3090 * read error occured, or if the op was a write. B_CACHE is never 3091 * set if the buffer is invalid or otherwise uncacheable. 3092 * 3093 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3094 * initiator to leave B_INVAL set to brelse the buffer out of existance 3095 * in the biodone routine. 3096 */ 3097void 3098bufdone(struct buf *bp) 3099{ 3100 int s; 3101 void (*biodone)(struct buf *); 3102 3103 GIANT_REQUIRED; 3104 3105 s = splbio(); 3106 3107 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3108 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3109 3110 bp->b_flags |= B_DONE; 3111 runningbufwakeup(bp); 3112 3113 if (bp->b_iocmd == BIO_DELETE) { 3114 brelse(bp); 3115 splx(s); 3116 return; 3117 } 3118 3119 if (bp->b_iocmd == BIO_WRITE) { 3120 vwakeup(bp); 3121 } 3122 3123 /* call optional completion function if requested */ 3124 if (bp->b_iodone != NULL) { 3125 biodone = bp->b_iodone; 3126 bp->b_iodone = NULL; 3127 (*biodone) (bp); 3128 splx(s); 3129 return; 3130 } 3131 if (LIST_FIRST(&bp->b_dep) != NULL) 3132 buf_complete(bp); 3133 3134 if (bp->b_flags & B_VMIO) { 3135 int i; 3136 vm_ooffset_t foff; 3137 vm_page_t m; 3138 vm_object_t obj; 3139 int iosize; 3140 struct vnode *vp = bp->b_vp; 3141 3142 obj = bp->b_object; 3143 3144#if defined(VFS_BIO_DEBUG) 3145 mp_fixme("usecount and vflag accessed without locks."); 3146 if (vp->v_usecount == 0) { 3147 panic("biodone: zero vnode ref count"); 3148 } 3149 3150 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3151 panic("biodone: vnode is not setup for merged cache"); 3152 } 3153#endif 3154 3155 foff = bp->b_offset; 3156 KASSERT(bp->b_offset != NOOFFSET, 3157 ("biodone: no buffer offset")); 3158 3159#if defined(VFS_BIO_DEBUG) 3160 if (obj->paging_in_progress < bp->b_npages) { 3161 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3162 obj->paging_in_progress, bp->b_npages); 3163 } 3164#endif 3165 3166 /* 3167 * Set B_CACHE if the op was a normal read and no error 3168 * occured. B_CACHE is set for writes in the b*write() 3169 * routines. 3170 */ 3171 iosize = bp->b_bcount - bp->b_resid; 3172 if (bp->b_iocmd == BIO_READ && 3173 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3174 !(bp->b_ioflags & BIO_ERROR)) { 3175 bp->b_flags |= B_CACHE; 3176 } 3177 vm_page_lock_queues(); 3178 for (i = 0; i < bp->b_npages; i++) { 3179 int bogusflag = 0; 3180 int resid; 3181 3182 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3183 if (resid > iosize) 3184 resid = iosize; 3185 3186 /* 3187 * cleanup bogus pages, restoring the originals 3188 */ 3189 m = bp->b_pages[i]; 3190 if (m == bogus_page) { 3191 bogusflag = 1; 3192 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3193 if (m == NULL) 3194 panic("biodone: page disappeared!"); 3195 bp->b_pages[i] = m; 3196 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3197 } 3198#if defined(VFS_BIO_DEBUG) 3199 if (OFF_TO_IDX(foff) != m->pindex) { 3200 printf( 3201"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3202 (intmax_t)foff, (uintmax_t)m->pindex); 3203 } 3204#endif 3205 3206 /* 3207 * In the write case, the valid and clean bits are 3208 * already changed correctly ( see bdwrite() ), so we 3209 * only need to do this here in the read case. 3210 */ 3211 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3212 vfs_page_set_valid(bp, foff, i, m); 3213 } 3214 vm_page_flag_clear(m, PG_ZERO); 3215 3216 /* 3217 * when debugging new filesystems or buffer I/O methods, this 3218 * is the most common error that pops up. if you see this, you 3219 * have not set the page busy flag correctly!!! 3220 */ 3221 if (m->busy == 0) { 3222 printf("biodone: page busy < 0, " 3223 "pindex: %d, foff: 0x(%x,%x), " 3224 "resid: %d, index: %d\n", 3225 (int) m->pindex, (int)(foff >> 32), 3226 (int) foff & 0xffffffff, resid, i); 3227 if (!vn_isdisk(vp, NULL)) 3228 printf(" iosize: %ld, lblkno: %jd, flags: 0x%x, npages: %d\n", 3229 bp->b_vp->v_mount->mnt_stat.f_iosize, 3230 (intmax_t) bp->b_lblkno, 3231 bp->b_flags, bp->b_npages); 3232 else 3233 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3234 (intmax_t) bp->b_lblkno, 3235 bp->b_flags, bp->b_npages); 3236 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3237 m->valid, m->dirty, m->wire_count); 3238 panic("biodone: page busy < 0\n"); 3239 } 3240 vm_page_io_finish(m); 3241 vm_object_pip_subtract(obj, 1); 3242 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3243 iosize -= resid; 3244 } 3245 vm_page_unlock_queues(); 3246 if (obj) 3247 vm_object_pip_wakeupn(obj, 0); 3248 } 3249 3250 /* 3251 * For asynchronous completions, release the buffer now. The brelse 3252 * will do a wakeup there if necessary - so no need to do a wakeup 3253 * here in the async case. The sync case always needs to do a wakeup. 3254 */ 3255 3256 if (bp->b_flags & B_ASYNC) { 3257 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3258 brelse(bp); 3259 else 3260 bqrelse(bp); 3261 } else { 3262 wakeup(bp); 3263 } 3264 splx(s); 3265} 3266 3267/* 3268 * This routine is called in lieu of iodone in the case of 3269 * incomplete I/O. This keeps the busy status for pages 3270 * consistant. 3271 */ 3272void 3273vfs_unbusy_pages(struct buf * bp) 3274{ 3275 int i; 3276 3277 GIANT_REQUIRED; 3278 3279 runningbufwakeup(bp); 3280 if (bp->b_flags & B_VMIO) { 3281 vm_object_t obj; 3282 3283 obj = bp->b_object; 3284 vm_page_lock_queues(); 3285 for (i = 0; i < bp->b_npages; i++) { 3286 vm_page_t m = bp->b_pages[i]; 3287 3288 if (m == bogus_page) { 3289 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3290 if (!m) { 3291 panic("vfs_unbusy_pages: page missing\n"); 3292 } 3293 bp->b_pages[i] = m; 3294 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3295 } 3296 vm_object_pip_subtract(obj, 1); 3297 vm_page_flag_clear(m, PG_ZERO); 3298 vm_page_io_finish(m); 3299 } 3300 vm_page_unlock_queues(); 3301 vm_object_pip_wakeupn(obj, 0); 3302 } 3303} 3304 3305/* 3306 * vfs_page_set_valid: 3307 * 3308 * Set the valid bits in a page based on the supplied offset. The 3309 * range is restricted to the buffer's size. 3310 * 3311 * This routine is typically called after a read completes. 3312 */ 3313static void 3314vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3315{ 3316 vm_ooffset_t soff, eoff; 3317 3318 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3319 /* 3320 * Start and end offsets in buffer. eoff - soff may not cross a 3321 * page boundry or cross the end of the buffer. The end of the 3322 * buffer, in this case, is our file EOF, not the allocation size 3323 * of the buffer. 3324 */ 3325 soff = off; 3326 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3327 if (eoff > bp->b_offset + bp->b_bcount) 3328 eoff = bp->b_offset + bp->b_bcount; 3329 3330 /* 3331 * Set valid range. This is typically the entire buffer and thus the 3332 * entire page. 3333 */ 3334 if (eoff > soff) { 3335 vm_page_set_validclean( 3336 m, 3337 (vm_offset_t) (soff & PAGE_MASK), 3338 (vm_offset_t) (eoff - soff) 3339 ); 3340 } 3341} 3342 3343/* 3344 * This routine is called before a device strategy routine. 3345 * It is used to tell the VM system that paging I/O is in 3346 * progress, and treat the pages associated with the buffer 3347 * almost as being PG_BUSY. Also the object paging_in_progress 3348 * flag is handled to make sure that the object doesn't become 3349 * inconsistant. 3350 * 3351 * Since I/O has not been initiated yet, certain buffer flags 3352 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3353 * and should be ignored. 3354 */ 3355void 3356vfs_busy_pages(struct buf * bp, int clear_modify) 3357{ 3358 int i, bogus; 3359 3360 if (bp->b_flags & B_VMIO) { 3361 vm_object_t obj; 3362 vm_ooffset_t foff; 3363 3364 obj = bp->b_object; 3365 foff = bp->b_offset; 3366 KASSERT(bp->b_offset != NOOFFSET, 3367 ("vfs_busy_pages: no buffer offset")); 3368 vfs_setdirty(bp); 3369retry: 3370 vm_page_lock_queues(); 3371 for (i = 0; i < bp->b_npages; i++) { 3372 vm_page_t m = bp->b_pages[i]; 3373 3374 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3375 goto retry; 3376 } 3377 bogus = 0; 3378 for (i = 0; i < bp->b_npages; i++) { 3379 vm_page_t m = bp->b_pages[i]; 3380 3381 vm_page_flag_clear(m, PG_ZERO); 3382 if ((bp->b_flags & B_CLUSTER) == 0) { 3383 vm_object_pip_add(obj, 1); 3384 vm_page_io_start(m); 3385 } 3386 /* 3387 * When readying a buffer for a read ( i.e 3388 * clear_modify == 0 ), it is important to do 3389 * bogus_page replacement for valid pages in 3390 * partially instantiated buffers. Partially 3391 * instantiated buffers can, in turn, occur when 3392 * reconstituting a buffer from its VM backing store 3393 * base. We only have to do this if B_CACHE is 3394 * clear ( which causes the I/O to occur in the 3395 * first place ). The replacement prevents the read 3396 * I/O from overwriting potentially dirty VM-backed 3397 * pages. XXX bogus page replacement is, uh, bogus. 3398 * It may not work properly with small-block devices. 3399 * We need to find a better way. 3400 */ 3401 pmap_remove_all(m); 3402 if (clear_modify) 3403 vfs_page_set_valid(bp, foff, i, m); 3404 else if (m->valid == VM_PAGE_BITS_ALL && 3405 (bp->b_flags & B_CACHE) == 0) { 3406 bp->b_pages[i] = bogus_page; 3407 bogus++; 3408 } 3409 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3410 } 3411 vm_page_unlock_queues(); 3412 if (bogus) 3413 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3414 } 3415} 3416 3417/* 3418 * Tell the VM system that the pages associated with this buffer 3419 * are clean. This is used for delayed writes where the data is 3420 * going to go to disk eventually without additional VM intevention. 3421 * 3422 * Note that while we only really need to clean through to b_bcount, we 3423 * just go ahead and clean through to b_bufsize. 3424 */ 3425static void 3426vfs_clean_pages(struct buf * bp) 3427{ 3428 int i; 3429 3430 GIANT_REQUIRED; 3431 3432 if (bp->b_flags & B_VMIO) { 3433 vm_ooffset_t foff; 3434 3435 foff = bp->b_offset; 3436 KASSERT(bp->b_offset != NOOFFSET, 3437 ("vfs_clean_pages: no buffer offset")); 3438 vm_page_lock_queues(); 3439 for (i = 0; i < bp->b_npages; i++) { 3440 vm_page_t m = bp->b_pages[i]; 3441 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3442 vm_ooffset_t eoff = noff; 3443 3444 if (eoff > bp->b_offset + bp->b_bufsize) 3445 eoff = bp->b_offset + bp->b_bufsize; 3446 vfs_page_set_valid(bp, foff, i, m); 3447 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3448 foff = noff; 3449 } 3450 vm_page_unlock_queues(); 3451 } 3452} 3453 3454/* 3455 * vfs_bio_set_validclean: 3456 * 3457 * Set the range within the buffer to valid and clean. The range is 3458 * relative to the beginning of the buffer, b_offset. Note that b_offset 3459 * itself may be offset from the beginning of the first page. 3460 * 3461 */ 3462 3463void 3464vfs_bio_set_validclean(struct buf *bp, int base, int size) 3465{ 3466 if (bp->b_flags & B_VMIO) { 3467 int i; 3468 int n; 3469 3470 /* 3471 * Fixup base to be relative to beginning of first page. 3472 * Set initial n to be the maximum number of bytes in the 3473 * first page that can be validated. 3474 */ 3475 3476 base += (bp->b_offset & PAGE_MASK); 3477 n = PAGE_SIZE - (base & PAGE_MASK); 3478 3479 vm_page_lock_queues(); 3480 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3481 vm_page_t m = bp->b_pages[i]; 3482 3483 if (n > size) 3484 n = size; 3485 3486 vm_page_set_validclean(m, base & PAGE_MASK, n); 3487 base += n; 3488 size -= n; 3489 n = PAGE_SIZE; 3490 } 3491 vm_page_unlock_queues(); 3492 } 3493} 3494 3495/* 3496 * vfs_bio_clrbuf: 3497 * 3498 * clear a buffer. This routine essentially fakes an I/O, so we need 3499 * to clear BIO_ERROR and B_INVAL. 3500 * 3501 * Note that while we only theoretically need to clear through b_bcount, 3502 * we go ahead and clear through b_bufsize. 3503 */ 3504 3505void 3506vfs_bio_clrbuf(struct buf *bp) 3507{ 3508 int i, mask = 0; 3509 caddr_t sa, ea; 3510 3511 GIANT_REQUIRED; 3512 3513 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3514 bp->b_flags &= ~B_INVAL; 3515 bp->b_ioflags &= ~BIO_ERROR; 3516 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3517 (bp->b_offset & PAGE_MASK) == 0) { 3518 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3519 if ((bp->b_pages[0]->valid & mask) == mask) { 3520 bp->b_resid = 0; 3521 return; 3522 } 3523 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3524 ((bp->b_pages[0]->valid & mask) == 0)) { 3525 bzero(bp->b_data, bp->b_bufsize); 3526 bp->b_pages[0]->valid |= mask; 3527 bp->b_resid = 0; 3528 return; 3529 } 3530 } 3531 ea = sa = bp->b_data; 3532 for(i=0;i<bp->b_npages;i++,sa=ea) { 3533 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3534 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3535 ea = (caddr_t)(vm_offset_t)ulmin( 3536 (u_long)(vm_offset_t)ea, 3537 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3538 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3539 if ((bp->b_pages[i]->valid & mask) == mask) 3540 continue; 3541 if ((bp->b_pages[i]->valid & mask) == 0) { 3542 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3543 bzero(sa, ea - sa); 3544 } 3545 } else { 3546 for (; sa < ea; sa += DEV_BSIZE, j++) { 3547 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3548 (bp->b_pages[i]->valid & (1<<j)) == 0) 3549 bzero(sa, DEV_BSIZE); 3550 } 3551 } 3552 bp->b_pages[i]->valid |= mask; 3553 vm_page_lock_queues(); 3554 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3555 vm_page_unlock_queues(); 3556 } 3557 bp->b_resid = 0; 3558 } else { 3559 clrbuf(bp); 3560 } 3561} 3562 3563/* 3564 * vm_hold_load_pages and vm_hold_free_pages get pages into 3565 * a buffers address space. The pages are anonymous and are 3566 * not associated with a file object. 3567 */ 3568static void 3569vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3570{ 3571 vm_offset_t pg; 3572 vm_page_t p; 3573 int index; 3574 3575 GIANT_REQUIRED; 3576 3577 to = round_page(to); 3578 from = round_page(from); 3579 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3580 3581 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3582tryagain: 3583 /* 3584 * note: must allocate system pages since blocking here 3585 * could intefere with paging I/O, no matter which 3586 * process we are. 3587 */ 3588 vm_object_lock(kernel_object); 3589 p = vm_page_alloc(kernel_object, 3590 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3591 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3592 vm_object_unlock(kernel_object); 3593 if (!p) { 3594 atomic_add_int(&vm_pageout_deficit, 3595 (to - pg) >> PAGE_SHIFT); 3596 VM_WAIT; 3597 goto tryagain; 3598 } 3599 vm_page_lock_queues(); 3600 p->valid = VM_PAGE_BITS_ALL; 3601 vm_page_unlock_queues(); 3602 pmap_qenter(pg, &p, 1); 3603 bp->b_pages[index] = p; 3604 vm_page_lock_queues(); 3605 vm_page_wakeup(p); 3606 vm_page_unlock_queues(); 3607 } 3608 bp->b_npages = index; 3609} 3610 3611/* Return pages associated with this buf to the vm system */ 3612static void 3613vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3614{ 3615 vm_offset_t pg; 3616 vm_page_t p; 3617 int index, newnpages; 3618 3619 GIANT_REQUIRED; 3620 3621 from = round_page(from); 3622 to = round_page(to); 3623 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3624 3625 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3626 p = bp->b_pages[index]; 3627 if (p && (index < bp->b_npages)) { 3628 if (p->busy) { 3629 printf( 3630 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3631 (intmax_t)bp->b_blkno, 3632 (intmax_t)bp->b_lblkno); 3633 } 3634 bp->b_pages[index] = NULL; 3635 pmap_qremove(pg, 1); 3636 vm_page_lock_queues(); 3637 vm_page_busy(p); 3638 vm_page_unwire(p, 0); 3639 vm_page_free(p); 3640 vm_page_unlock_queues(); 3641 } 3642 } 3643 bp->b_npages = newnpages; 3644} 3645 3646/* 3647 * Map an IO request into kernel virtual address space. 3648 * 3649 * All requests are (re)mapped into kernel VA space. 3650 * Notice that we use b_bufsize for the size of the buffer 3651 * to be mapped. b_bcount might be modified by the driver. 3652 * 3653 * Note that even if the caller determines that the address space should 3654 * be valid, a race or a smaller-file mapped into a larger space may 3655 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3656 * check the return value. 3657 */ 3658int 3659vmapbuf(struct buf *bp) 3660{ 3661 caddr_t addr, kva; 3662 vm_offset_t pa; 3663 int pidx, i; 3664 struct vm_page *m; 3665 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3666 3667 GIANT_REQUIRED; 3668 3669 if ((bp->b_flags & B_PHYS) == 0) 3670 panic("vmapbuf"); 3671 3672 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3673 addr < bp->b_data + bp->b_bufsize; 3674 addr += PAGE_SIZE, pidx++) { 3675 /* 3676 * Do the vm_fault if needed; do the copy-on-write thing 3677 * when reading stuff off device into memory. 3678 * 3679 * NOTE! Must use pmap_extract() because addr may be in 3680 * the userland address space, and kextract is only guarenteed 3681 * to work for the kernland address space (see: sparc64 port). 3682 */ 3683retry: 3684 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data, 3685 (bp->b_iocmd == BIO_READ) ? 3686 (VM_PROT_READ|VM_PROT_WRITE) : VM_PROT_READ); 3687 if (i < 0) { 3688 printf("vmapbuf: warning, bad user address during I/O\n"); 3689 vm_page_lock_queues(); 3690 for (i = 0; i < pidx; ++i) { 3691 vm_page_unhold(bp->b_pages[i]); 3692 bp->b_pages[i] = NULL; 3693 } 3694 vm_page_unlock_queues(); 3695 return(-1); 3696 } 3697 pa = trunc_page(pmap_extract(pmap, (vm_offset_t) addr)); 3698 if (pa == 0) { 3699 printf("vmapbuf: warning, race against user address during I/O"); 3700 goto retry; 3701 } 3702 m = PHYS_TO_VM_PAGE(pa); 3703 vm_page_lock_queues(); 3704 vm_page_hold(m); 3705 vm_page_unlock_queues(); 3706 bp->b_pages[pidx] = m; 3707 } 3708 if (pidx > btoc(MAXPHYS)) 3709 panic("vmapbuf: mapped more than MAXPHYS"); 3710 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3711 3712 kva = bp->b_saveaddr; 3713 bp->b_npages = pidx; 3714 bp->b_saveaddr = bp->b_data; 3715 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3716 return(0); 3717} 3718 3719/* 3720 * Free the io map PTEs associated with this IO operation. 3721 * We also invalidate the TLB entries and restore the original b_addr. 3722 */ 3723void 3724vunmapbuf(struct buf *bp) 3725{ 3726 int pidx; 3727 int npages; 3728 3729 GIANT_REQUIRED; 3730 3731 if ((bp->b_flags & B_PHYS) == 0) 3732 panic("vunmapbuf"); 3733 3734 npages = bp->b_npages; 3735 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), 3736 npages); 3737 vm_page_lock_queues(); 3738 for (pidx = 0; pidx < npages; pidx++) 3739 vm_page_unhold(bp->b_pages[pidx]); 3740 vm_page_unlock_queues(); 3741 3742 bp->b_data = bp->b_saveaddr; 3743} 3744 3745#include "opt_ddb.h" 3746#ifdef DDB 3747#include <ddb/ddb.h> 3748 3749/* DDB command to show buffer data */ 3750DB_SHOW_COMMAND(buffer, db_show_buffer) 3751{ 3752 /* get args */ 3753 struct buf *bp = (struct buf *)addr; 3754 3755 if (!have_addr) { 3756 db_printf("usage: show buffer <addr>\n"); 3757 return; 3758 } 3759 3760 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3761 db_printf( 3762 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3763 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3764 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3765 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3766 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3767 if (bp->b_npages) { 3768 int i; 3769 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3770 for (i = 0; i < bp->b_npages; i++) { 3771 vm_page_t m; 3772 m = bp->b_pages[i]; 3773 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3774 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3775 if ((i + 1) < bp->b_npages) 3776 db_printf(","); 3777 } 3778 db_printf("\n"); 3779 } 3780} 3781#endif /* DDB */ 3782