vfs_bio.c revision 99737
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 99737 2002-07-10 17:02:32Z dillon $ 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/eventhandler.h> 36#include <sys/lock.h> 37#include <sys/malloc.h> 38#include <sys/mount.h> 39#include <sys/mutex.h> 40#include <sys/kernel.h> 41#include <sys/kthread.h> 42#include <sys/ktr.h> 43#include <sys/proc.h> 44#include <sys/reboot.h> 45#include <sys/resourcevar.h> 46#include <sys/sysctl.h> 47#include <sys/vmmeter.h> 48#include <sys/vnode.h> 49#include <vm/vm.h> 50#include <vm/vm_param.h> 51#include <vm/vm_kern.h> 52#include <vm/vm_pageout.h> 53#include <vm/vm_page.h> 54#include <vm/vm_object.h> 55#include <vm/vm_extern.h> 56#include <vm/vm_map.h> 57 58static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 59 60struct bio_ops bioops; /* I/O operation notification */ 61 62struct buf_ops buf_ops_bio = { 63 "buf_ops_bio", 64 bwrite 65}; 66 67/* 68 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 69 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 70 */ 71struct buf *buf; /* buffer header pool */ 72struct mtx buftimelock; /* Interlock on setting prio and timo */ 73 74static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 75 vm_offset_t to); 76static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 77 vm_offset_t to); 78static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 79 int pageno, vm_page_t m); 80static void vfs_clean_pages(struct buf * bp); 81static void vfs_setdirty(struct buf *bp); 82static void vfs_vmio_release(struct buf *bp); 83static void vfs_backgroundwritedone(struct buf *bp); 84static int flushbufqueues(void); 85static void buf_daemon(void); 86 87int vmiodirenable = TRUE; 88SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 89 "Use the VM system for directory writes"); 90int runningbufspace; 91SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 92 "Amount of presently outstanding async buffer io"); 93static int bufspace; 94SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 95 "KVA memory used for bufs"); 96static int maxbufspace; 97SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 98 "Maximum allowed value of bufspace (including buf_daemon)"); 99static int bufmallocspace; 100SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 101 "Amount of malloced memory for buffers"); 102static int maxbufmallocspace; 103SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 104 "Maximum amount of malloced memory for buffers"); 105static int lobufspace; 106SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 107 "Minimum amount of buffers we want to have"); 108static int hibufspace; 109SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 110 "Maximum allowed value of bufspace (excluding buf_daemon)"); 111static int bufreusecnt; 112SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 113 "Number of times we have reused a buffer"); 114static int buffreekvacnt; 115SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 116 "Number of times we have freed the KVA space from some buffer"); 117static int bufdefragcnt; 118SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 119 "Number of times we have had to repeat buffer allocation to defragment"); 120static int lorunningspace; 121SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 122 "Minimum preferred space used for in-progress I/O"); 123static int hirunningspace; 124SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 125 "Maximum amount of space to use for in-progress I/O"); 126static int numdirtybuffers; 127SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 128 "Number of buffers that are dirty (has unwritten changes) at the moment"); 129static int lodirtybuffers; 130SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 131 "How many buffers we want to have free before bufdaemon can sleep"); 132static int hidirtybuffers; 133SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 134 "When the number of dirty buffers is considered severe"); 135static int numfreebuffers; 136SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 137 "Number of free buffers"); 138static int lofreebuffers; 139SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 140 "XXX Unused"); 141static int hifreebuffers; 142SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 143 "XXX Complicatedly unused"); 144static int getnewbufcalls; 145SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 146 "Number of calls to getnewbuf"); 147static int getnewbufrestarts; 148SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 149 "Number of times getnewbuf has had to restart a buffer aquisition"); 150static int dobkgrdwrite = 1; 151SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 152 "Do background writes (honoring the BX_BKGRDWRITE flag)?"); 153 154/* 155 * Wakeup point for bufdaemon, as well as indicator of whether it is already 156 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 157 * is idling. 158 */ 159static int bd_request; 160 161/* 162 * bogus page -- for I/O to/from partially complete buffers 163 * this is a temporary solution to the problem, but it is not 164 * really that bad. it would be better to split the buffer 165 * for input in the case of buffers partially already in memory, 166 * but the code is intricate enough already. 167 */ 168vm_page_t bogus_page; 169 170/* 171 * Offset for bogus_page. 172 * XXX bogus_offset should be local to bufinit 173 */ 174static vm_offset_t bogus_offset; 175 176/* 177 * Synchronization (sleep/wakeup) variable for active buffer space requests. 178 * Set when wait starts, cleared prior to wakeup(). 179 * Used in runningbufwakeup() and waitrunningbufspace(). 180 */ 181static int runningbufreq; 182 183/* 184 * Synchronization (sleep/wakeup) variable for buffer requests. 185 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 186 * by and/or. 187 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 188 * getnewbuf(), and getblk(). 189 */ 190static int needsbuffer; 191 192#ifdef USE_BUFHASH 193/* 194 * Mask for index into the buffer hash table, which needs to be power of 2 in 195 * size. Set in kern_vfs_bio_buffer_alloc. 196 */ 197static int bufhashmask; 198 199/* 200 * Hash table for all buffers, with a linked list hanging from each table 201 * entry. Set in kern_vfs_bio_buffer_alloc, initialized in buf_init. 202 */ 203static LIST_HEAD(bufhashhdr, buf) *bufhashtbl; 204 205/* 206 * Somewhere to store buffers when they are not in another list, to always 207 * have them in a list (and thus being able to use the same set of operations 208 * on them.) 209 */ 210static struct bufhashhdr invalhash; 211 212#endif 213 214/* 215 * Definitions for the buffer free lists. 216 */ 217#define BUFFER_QUEUES 6 /* number of free buffer queues */ 218 219#define QUEUE_NONE 0 /* on no queue */ 220#define QUEUE_LOCKED 1 /* locked buffers */ 221#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ 222#define QUEUE_DIRTY 3 /* B_DELWRI buffers */ 223#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 224#define QUEUE_EMPTY 5 /* empty buffer headers */ 225 226/* Queues for free buffers with various properties */ 227static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 228/* 229 * Single global constant for BUF_WMESG, to avoid getting multiple references. 230 * buf_wmesg is referred from macros. 231 */ 232const char *buf_wmesg = BUF_WMESG; 233 234#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 235#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 236#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 237#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 238 239#ifdef USE_BUFHASH 240/* 241 * Buffer hash table code. Note that the logical block scans linearly, which 242 * gives us some L1 cache locality. 243 */ 244 245static __inline 246struct bufhashhdr * 247bufhash(struct vnode *vnp, daddr_t bn) 248{ 249 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 250} 251 252#endif 253 254/* 255 * numdirtywakeup: 256 * 257 * If someone is blocked due to there being too many dirty buffers, 258 * and numdirtybuffers is now reasonable, wake them up. 259 */ 260 261static __inline void 262numdirtywakeup(int level) 263{ 264 if (numdirtybuffers <= level) { 265 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 266 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 267 wakeup(&needsbuffer); 268 } 269 } 270} 271 272/* 273 * bufspacewakeup: 274 * 275 * Called when buffer space is potentially available for recovery. 276 * getnewbuf() will block on this flag when it is unable to free 277 * sufficient buffer space. Buffer space becomes recoverable when 278 * bp's get placed back in the queues. 279 */ 280 281static __inline void 282bufspacewakeup(void) 283{ 284 /* 285 * If someone is waiting for BUF space, wake them up. Even 286 * though we haven't freed the kva space yet, the waiting 287 * process will be able to now. 288 */ 289 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 290 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 291 wakeup(&needsbuffer); 292 } 293} 294 295/* 296 * runningbufwakeup() - in-progress I/O accounting. 297 * 298 */ 299static __inline void 300runningbufwakeup(struct buf *bp) 301{ 302 if (bp->b_runningbufspace) { 303 runningbufspace -= bp->b_runningbufspace; 304 bp->b_runningbufspace = 0; 305 if (runningbufreq && runningbufspace <= lorunningspace) { 306 runningbufreq = 0; 307 wakeup(&runningbufreq); 308 } 309 } 310} 311 312/* 313 * bufcountwakeup: 314 * 315 * Called when a buffer has been added to one of the free queues to 316 * account for the buffer and to wakeup anyone waiting for free buffers. 317 * This typically occurs when large amounts of metadata are being handled 318 * by the buffer cache ( else buffer space runs out first, usually ). 319 */ 320 321static __inline void 322bufcountwakeup(void) 323{ 324 ++numfreebuffers; 325 if (needsbuffer) { 326 needsbuffer &= ~VFS_BIO_NEED_ANY; 327 if (numfreebuffers >= hifreebuffers) 328 needsbuffer &= ~VFS_BIO_NEED_FREE; 329 wakeup(&needsbuffer); 330 } 331} 332 333/* 334 * waitrunningbufspace() 335 * 336 * runningbufspace is a measure of the amount of I/O currently 337 * running. This routine is used in async-write situations to 338 * prevent creating huge backups of pending writes to a device. 339 * Only asynchronous writes are governed by this function. 340 * 341 * Reads will adjust runningbufspace, but will not block based on it. 342 * The read load has a side effect of reducing the allowed write load. 343 * 344 * This does NOT turn an async write into a sync write. It waits 345 * for earlier writes to complete and generally returns before the 346 * caller's write has reached the device. 347 */ 348static __inline void 349waitrunningbufspace(void) 350{ 351 /* 352 * XXX race against wakeup interrupt, currently 353 * protected by Giant. FIXME! 354 */ 355 while (runningbufspace > hirunningspace) { 356 ++runningbufreq; 357 tsleep(&runningbufreq, PVM, "wdrain", 0); 358 } 359} 360 361 362/* 363 * vfs_buf_test_cache: 364 * 365 * Called when a buffer is extended. This function clears the B_CACHE 366 * bit if the newly extended portion of the buffer does not contain 367 * valid data. 368 */ 369static __inline__ 370void 371vfs_buf_test_cache(struct buf *bp, 372 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 373 vm_page_t m) 374{ 375 GIANT_REQUIRED; 376 377 if (bp->b_flags & B_CACHE) { 378 int base = (foff + off) & PAGE_MASK; 379 if (vm_page_is_valid(m, base, size) == 0) 380 bp->b_flags &= ~B_CACHE; 381 } 382} 383 384/* Wake up the buffer deamon if necessary */ 385static __inline__ 386void 387bd_wakeup(int dirtybuflevel) 388{ 389 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 390 bd_request = 1; 391 wakeup(&bd_request); 392 } 393} 394 395/* 396 * bd_speedup - speedup the buffer cache flushing code 397 */ 398 399static __inline__ 400void 401bd_speedup(void) 402{ 403 bd_wakeup(1); 404} 405 406/* 407 * Calculating buffer cache scaling values and reserve space for buffer 408 * headers. This is called during low level kernel initialization and 409 * may be called more then once. We CANNOT write to the memory area 410 * being reserved at this time. 411 */ 412caddr_t 413kern_vfs_bio_buffer_alloc(caddr_t v, int physmem_est) 414{ 415 /* 416 * physmem_est is in pages. Convert it to kilobytes (assumes 417 * PAGE_SIZE is >= 1K) 418 */ 419 physmem_est = physmem_est * (PAGE_SIZE / 1024); 420 421 /* 422 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 423 * For the first 64MB of ram nominally allocate sufficient buffers to 424 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 425 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 426 * the buffer cache we limit the eventual kva reservation to 427 * maxbcache bytes. 428 * 429 * factor represents the 1/4 x ram conversion. 430 */ 431 if (nbuf == 0) { 432 int factor = 4 * BKVASIZE / 1024; 433 434 nbuf = 50; 435 if (physmem_est > 4096) 436 nbuf += min((physmem_est - 4096) / factor, 437 65536 / factor); 438 if (physmem_est > 65536) 439 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 440 441 if (maxbcache && nbuf > maxbcache / BKVASIZE) 442 nbuf = maxbcache / BKVASIZE; 443 } 444 445#if 0 446 /* 447 * Do not allow the buffer_map to be more then 1/2 the size of the 448 * kernel_map. 449 */ 450 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 451 (BKVASIZE * 2)) { 452 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 453 (BKVASIZE * 2); 454 printf("Warning: nbufs capped at %d\n", nbuf); 455 } 456#endif 457 458 /* 459 * swbufs are used as temporary holders for I/O, such as paging I/O. 460 * We have no less then 16 and no more then 256. 461 */ 462 nswbuf = max(min(nbuf/4, 256), 16); 463 464 /* 465 * Reserve space for the buffer cache buffers 466 */ 467 swbuf = (void *)v; 468 v = (caddr_t)(swbuf + nswbuf); 469 buf = (void *)v; 470 v = (caddr_t)(buf + nbuf); 471 472#ifdef USE_BUFHASH 473 /* 474 * Calculate the hash table size and reserve space 475 */ 476 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 477 ; 478 bufhashtbl = (void *)v; 479 v = (caddr_t)(bufhashtbl + bufhashmask); 480 --bufhashmask; 481#endif 482 return(v); 483} 484 485/* Initialize the buffer subsystem. Called before use of any buffers. */ 486void 487bufinit(void) 488{ 489 struct buf *bp; 490 int i; 491 492 GIANT_REQUIRED; 493 494#ifdef USE_BUFHASH 495 LIST_INIT(&invalhash); 496#endif 497 mtx_init(&buftimelock, "buftime lock", NULL, MTX_DEF); 498 499#ifdef USE_BUFHASH 500 for (i = 0; i <= bufhashmask; i++) 501 LIST_INIT(&bufhashtbl[i]); 502#endif 503 504 /* next, make a null set of free lists */ 505 for (i = 0; i < BUFFER_QUEUES; i++) 506 TAILQ_INIT(&bufqueues[i]); 507 508 /* finally, initialize each buffer header and stick on empty q */ 509 for (i = 0; i < nbuf; i++) { 510 bp = &buf[i]; 511 bzero(bp, sizeof *bp); 512 bp->b_flags = B_INVAL; /* we're just an empty header */ 513 bp->b_dev = NODEV; 514 bp->b_rcred = NOCRED; 515 bp->b_wcred = NOCRED; 516 bp->b_qindex = QUEUE_EMPTY; 517 bp->b_xflags = 0; 518 LIST_INIT(&bp->b_dep); 519 BUF_LOCKINIT(bp); 520 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 521#ifdef USE_BUFHASH 522 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 523#endif 524 } 525 526 /* 527 * maxbufspace is the absolute maximum amount of buffer space we are 528 * allowed to reserve in KVM and in real terms. The absolute maximum 529 * is nominally used by buf_daemon. hibufspace is the nominal maximum 530 * used by most other processes. The differential is required to 531 * ensure that buf_daemon is able to run when other processes might 532 * be blocked waiting for buffer space. 533 * 534 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 535 * this may result in KVM fragmentation which is not handled optimally 536 * by the system. 537 */ 538 maxbufspace = nbuf * BKVASIZE; 539 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 540 lobufspace = hibufspace - MAXBSIZE; 541 542 lorunningspace = 512 * 1024; 543 hirunningspace = 1024 * 1024; 544 545/* 546 * Limit the amount of malloc memory since it is wired permanently into 547 * the kernel space. Even though this is accounted for in the buffer 548 * allocation, we don't want the malloced region to grow uncontrolled. 549 * The malloc scheme improves memory utilization significantly on average 550 * (small) directories. 551 */ 552 maxbufmallocspace = hibufspace / 20; 553 554/* 555 * Reduce the chance of a deadlock occuring by limiting the number 556 * of delayed-write dirty buffers we allow to stack up. 557 */ 558 hidirtybuffers = nbuf / 4 + 20; 559 numdirtybuffers = 0; 560/* 561 * To support extreme low-memory systems, make sure hidirtybuffers cannot 562 * eat up all available buffer space. This occurs when our minimum cannot 563 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 564 * BKVASIZE'd (8K) buffers. 565 */ 566 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 567 hidirtybuffers >>= 1; 568 } 569 lodirtybuffers = hidirtybuffers / 2; 570 571/* 572 * Try to keep the number of free buffers in the specified range, 573 * and give special processes (e.g. like buf_daemon) access to an 574 * emergency reserve. 575 */ 576 lofreebuffers = nbuf / 18 + 5; 577 hifreebuffers = 2 * lofreebuffers; 578 numfreebuffers = nbuf; 579 580/* 581 * Maximum number of async ops initiated per buf_daemon loop. This is 582 * somewhat of a hack at the moment, we really need to limit ourselves 583 * based on the number of bytes of I/O in-transit that were initiated 584 * from buf_daemon. 585 */ 586 587 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 588 bogus_page = vm_page_alloc(kernel_object, 589 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 590 VM_ALLOC_NORMAL); 591 cnt.v_wire_count++; 592} 593 594/* 595 * bfreekva() - free the kva allocation for a buffer. 596 * 597 * Must be called at splbio() or higher as this is the only locking for 598 * buffer_map. 599 * 600 * Since this call frees up buffer space, we call bufspacewakeup(). 601 */ 602static void 603bfreekva(struct buf * bp) 604{ 605 GIANT_REQUIRED; 606 607 if (bp->b_kvasize) { 608 ++buffreekvacnt; 609 bufspace -= bp->b_kvasize; 610 vm_map_delete(buffer_map, 611 (vm_offset_t) bp->b_kvabase, 612 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 613 ); 614 bp->b_kvasize = 0; 615 bufspacewakeup(); 616 } 617} 618 619/* 620 * bremfree: 621 * 622 * Remove the buffer from the appropriate free list. 623 */ 624void 625bremfree(struct buf * bp) 626{ 627 int s = splbio(); 628 int old_qindex = bp->b_qindex; 629 630 GIANT_REQUIRED; 631 632 if (bp->b_qindex != QUEUE_NONE) { 633 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 634 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 635 bp->b_qindex = QUEUE_NONE; 636 } else { 637 if (BUF_REFCNT(bp) <= 1) 638 panic("bremfree: removing a buffer not on a queue"); 639 } 640 641 /* 642 * Fixup numfreebuffers count. If the buffer is invalid or not 643 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 644 * the buffer was free and we must decrement numfreebuffers. 645 */ 646 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 647 switch(old_qindex) { 648 case QUEUE_DIRTY: 649 case QUEUE_CLEAN: 650 case QUEUE_EMPTY: 651 case QUEUE_EMPTYKVA: 652 --numfreebuffers; 653 break; 654 default: 655 break; 656 } 657 } 658 splx(s); 659} 660 661 662/* 663 * Get a buffer with the specified data. Look in the cache first. We 664 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 665 * is set, the buffer is valid and we do not have to do anything ( see 666 * getblk() ). This is really just a special case of breadn(). 667 */ 668int 669bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 670 struct buf ** bpp) 671{ 672 673 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 674} 675 676/* 677 * Operates like bread, but also starts asynchronous I/O on 678 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 679 * to initiating I/O . If B_CACHE is set, the buffer is valid 680 * and we do not have to do anything. 681 */ 682int 683breadn(struct vnode * vp, daddr_t blkno, int size, 684 daddr_t * rablkno, int *rabsize, 685 int cnt, struct ucred * cred, struct buf ** bpp) 686{ 687 struct buf *bp, *rabp; 688 int i; 689 int rv = 0, readwait = 0; 690 691 *bpp = bp = getblk(vp, blkno, size, 0, 0); 692 693 /* if not found in cache, do some I/O */ 694 if ((bp->b_flags & B_CACHE) == 0) { 695 if (curthread != PCPU_GET(idlethread)) 696 curthread->td_proc->p_stats->p_ru.ru_inblock++; 697 bp->b_iocmd = BIO_READ; 698 bp->b_flags &= ~B_INVAL; 699 bp->b_ioflags &= ~BIO_ERROR; 700 if (bp->b_rcred == NOCRED && cred != NOCRED) 701 bp->b_rcred = crhold(cred); 702 vfs_busy_pages(bp, 0); 703 VOP_STRATEGY(vp, bp); 704 ++readwait; 705 } 706 707 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 708 if (inmem(vp, *rablkno)) 709 continue; 710 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 711 712 if ((rabp->b_flags & B_CACHE) == 0) { 713 if (curthread != PCPU_GET(idlethread)) 714 curthread->td_proc->p_stats->p_ru.ru_inblock++; 715 rabp->b_flags |= B_ASYNC; 716 rabp->b_flags &= ~B_INVAL; 717 rabp->b_ioflags &= ~BIO_ERROR; 718 rabp->b_iocmd = BIO_READ; 719 if (rabp->b_rcred == NOCRED && cred != NOCRED) 720 rabp->b_rcred = crhold(cred); 721 vfs_busy_pages(rabp, 0); 722 BUF_KERNPROC(rabp); 723 VOP_STRATEGY(vp, rabp); 724 } else { 725 brelse(rabp); 726 } 727 } 728 729 if (readwait) { 730 rv = bufwait(bp); 731 } 732 return (rv); 733} 734 735/* 736 * Write, release buffer on completion. (Done by iodone 737 * if async). Do not bother writing anything if the buffer 738 * is invalid. 739 * 740 * Note that we set B_CACHE here, indicating that buffer is 741 * fully valid and thus cacheable. This is true even of NFS 742 * now so we set it generally. This could be set either here 743 * or in biodone() since the I/O is synchronous. We put it 744 * here. 745 */ 746 747int 748bwrite(struct buf * bp) 749{ 750 int oldflags, s; 751 struct buf *newbp; 752 753 if (bp->b_flags & B_INVAL) { 754 brelse(bp); 755 return (0); 756 } 757 758 oldflags = bp->b_flags; 759 760 if (BUF_REFCNT(bp) == 0) 761 panic("bwrite: buffer is not busy???"); 762 s = splbio(); 763 /* 764 * If a background write is already in progress, delay 765 * writing this block if it is asynchronous. Otherwise 766 * wait for the background write to complete. 767 */ 768 if (bp->b_xflags & BX_BKGRDINPROG) { 769 if (bp->b_flags & B_ASYNC) { 770 splx(s); 771 bdwrite(bp); 772 return (0); 773 } 774 bp->b_xflags |= BX_BKGRDWAIT; 775 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); 776 if (bp->b_xflags & BX_BKGRDINPROG) 777 panic("bwrite: still writing"); 778 } 779 780 /* Mark the buffer clean */ 781 bundirty(bp); 782 783 /* 784 * If this buffer is marked for background writing and we 785 * do not have to wait for it, make a copy and write the 786 * copy so as to leave this buffer ready for further use. 787 * 788 * This optimization eats a lot of memory. If we have a page 789 * or buffer shortfall we can't do it. 790 */ 791 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 792 (bp->b_flags & B_ASYNC) && 793 !vm_page_count_severe() && 794 !buf_dirty_count_severe()) { 795 if (bp->b_iodone != NULL) { 796 printf("bp->b_iodone = %p\n", bp->b_iodone); 797 panic("bwrite: need chained iodone"); 798 } 799 800 /* get a new block */ 801 newbp = geteblk(bp->b_bufsize); 802 803 /* 804 * set it to be identical to the old block. We have to 805 * set b_lblkno and BKGRDMARKER before calling bgetvp() 806 * to avoid confusing the splay tree and gbincore(). 807 */ 808 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 809 newbp->b_lblkno = bp->b_lblkno; 810 newbp->b_xflags |= BX_BKGRDMARKER; 811 bgetvp(bp->b_vp, newbp); 812 newbp->b_blkno = bp->b_blkno; 813 newbp->b_offset = bp->b_offset; 814 newbp->b_iodone = vfs_backgroundwritedone; 815 newbp->b_flags |= B_ASYNC; 816 newbp->b_flags &= ~B_INVAL; 817 818 /* move over the dependencies */ 819 if (LIST_FIRST(&bp->b_dep) != NULL) 820 buf_movedeps(bp, newbp); 821 822 /* 823 * Initiate write on the copy, release the original to 824 * the B_LOCKED queue so that it cannot go away until 825 * the background write completes. If not locked it could go 826 * away and then be reconstituted while it was being written. 827 * If the reconstituted buffer were written, we could end up 828 * with two background copies being written at the same time. 829 */ 830 bp->b_xflags |= BX_BKGRDINPROG; 831 bp->b_flags |= B_LOCKED; 832 bqrelse(bp); 833 bp = newbp; 834 } 835 836 bp->b_flags &= ~B_DONE; 837 bp->b_ioflags &= ~BIO_ERROR; 838 bp->b_flags |= B_WRITEINPROG | B_CACHE; 839 bp->b_iocmd = BIO_WRITE; 840 841 bp->b_vp->v_numoutput++; 842 vfs_busy_pages(bp, 1); 843 844 /* 845 * Normal bwrites pipeline writes 846 */ 847 bp->b_runningbufspace = bp->b_bufsize; 848 runningbufspace += bp->b_runningbufspace; 849 850 if (curthread != PCPU_GET(idlethread)) 851 curthread->td_proc->p_stats->p_ru.ru_oublock++; 852 splx(s); 853 if (oldflags & B_ASYNC) 854 BUF_KERNPROC(bp); 855 BUF_STRATEGY(bp); 856 857 if ((oldflags & B_ASYNC) == 0) { 858 int rtval = bufwait(bp); 859 brelse(bp); 860 return (rtval); 861 } else if ((oldflags & B_NOWDRAIN) == 0) { 862 /* 863 * don't allow the async write to saturate the I/O 864 * system. Deadlocks can occur only if a device strategy 865 * routine (like in MD) turns around and issues another 866 * high-level write, in which case B_NOWDRAIN is expected 867 * to be set. Otherwise we will not deadlock here because 868 * we are blocking waiting for I/O that is already in-progress 869 * to complete. 870 */ 871 waitrunningbufspace(); 872 } 873 874 return (0); 875} 876 877/* 878 * Complete a background write started from bwrite. 879 */ 880static void 881vfs_backgroundwritedone(bp) 882 struct buf *bp; 883{ 884 struct buf *origbp; 885 886 /* 887 * Find the original buffer that we are writing. 888 */ 889 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 890 panic("backgroundwritedone: lost buffer"); 891 /* 892 * Process dependencies then return any unfinished ones. 893 */ 894 if (LIST_FIRST(&bp->b_dep) != NULL) 895 buf_complete(bp); 896 if (LIST_FIRST(&bp->b_dep) != NULL) 897 buf_movedeps(bp, origbp); 898 /* 899 * Clear the BX_BKGRDINPROG flag in the original buffer 900 * and awaken it if it is waiting for the write to complete. 901 * If BX_BKGRDINPROG is not set in the original buffer it must 902 * have been released and re-instantiated - which is not legal. 903 */ 904 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 905 ("backgroundwritedone: lost buffer2")); 906 origbp->b_xflags &= ~BX_BKGRDINPROG; 907 if (origbp->b_xflags & BX_BKGRDWAIT) { 908 origbp->b_xflags &= ~BX_BKGRDWAIT; 909 wakeup(&origbp->b_xflags); 910 } 911 /* 912 * Clear the B_LOCKED flag and remove it from the locked 913 * queue if it currently resides there. 914 */ 915 origbp->b_flags &= ~B_LOCKED; 916 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 917 bremfree(origbp); 918 bqrelse(origbp); 919 } 920 /* 921 * This buffer is marked B_NOCACHE, so when it is released 922 * by biodone, it will be tossed. We mark it with BIO_READ 923 * to avoid biodone doing a second vwakeup. 924 */ 925 bp->b_flags |= B_NOCACHE; 926 bp->b_iocmd = BIO_READ; 927 bp->b_flags &= ~(B_CACHE | B_DONE); 928 bp->b_iodone = 0; 929 bufdone(bp); 930} 931 932/* 933 * Delayed write. (Buffer is marked dirty). Do not bother writing 934 * anything if the buffer is marked invalid. 935 * 936 * Note that since the buffer must be completely valid, we can safely 937 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 938 * biodone() in order to prevent getblk from writing the buffer 939 * out synchronously. 940 */ 941void 942bdwrite(struct buf * bp) 943{ 944 GIANT_REQUIRED; 945 946 if (BUF_REFCNT(bp) == 0) 947 panic("bdwrite: buffer is not busy"); 948 949 if (bp->b_flags & B_INVAL) { 950 brelse(bp); 951 return; 952 } 953 bdirty(bp); 954 955 /* 956 * Set B_CACHE, indicating that the buffer is fully valid. This is 957 * true even of NFS now. 958 */ 959 bp->b_flags |= B_CACHE; 960 961 /* 962 * This bmap keeps the system from needing to do the bmap later, 963 * perhaps when the system is attempting to do a sync. Since it 964 * is likely that the indirect block -- or whatever other datastructure 965 * that the filesystem needs is still in memory now, it is a good 966 * thing to do this. Note also, that if the pageout daemon is 967 * requesting a sync -- there might not be enough memory to do 968 * the bmap then... So, this is important to do. 969 */ 970 if (bp->b_lblkno == bp->b_blkno) { 971 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 972 } 973 974 /* 975 * Set the *dirty* buffer range based upon the VM system dirty pages. 976 */ 977 vfs_setdirty(bp); 978 979 /* 980 * We need to do this here to satisfy the vnode_pager and the 981 * pageout daemon, so that it thinks that the pages have been 982 * "cleaned". Note that since the pages are in a delayed write 983 * buffer -- the VFS layer "will" see that the pages get written 984 * out on the next sync, or perhaps the cluster will be completed. 985 */ 986 vfs_clean_pages(bp); 987 bqrelse(bp); 988 989 /* 990 * Wakeup the buffer flushing daemon if we have a lot of dirty 991 * buffers (midpoint between our recovery point and our stall 992 * point). 993 */ 994 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 995 996 /* 997 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 998 * due to the softdep code. 999 */ 1000} 1001 1002/* 1003 * bdirty: 1004 * 1005 * Turn buffer into delayed write request. We must clear BIO_READ and 1006 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1007 * itself to properly update it in the dirty/clean lists. We mark it 1008 * B_DONE to ensure that any asynchronization of the buffer properly 1009 * clears B_DONE ( else a panic will occur later ). 1010 * 1011 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1012 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1013 * should only be called if the buffer is known-good. 1014 * 1015 * Since the buffer is not on a queue, we do not update the numfreebuffers 1016 * count. 1017 * 1018 * Must be called at splbio(). 1019 * The buffer must be on QUEUE_NONE. 1020 */ 1021void 1022bdirty(bp) 1023 struct buf *bp; 1024{ 1025 KASSERT(bp->b_qindex == QUEUE_NONE, 1026 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1027 bp->b_flags &= ~(B_RELBUF); 1028 bp->b_iocmd = BIO_WRITE; 1029 1030 if ((bp->b_flags & B_DELWRI) == 0) { 1031 bp->b_flags |= B_DONE | B_DELWRI; 1032 reassignbuf(bp, bp->b_vp); 1033 ++numdirtybuffers; 1034 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1035 } 1036} 1037 1038/* 1039 * bundirty: 1040 * 1041 * Clear B_DELWRI for buffer. 1042 * 1043 * Since the buffer is not on a queue, we do not update the numfreebuffers 1044 * count. 1045 * 1046 * Must be called at splbio(). 1047 * The buffer must be on QUEUE_NONE. 1048 */ 1049 1050void 1051bundirty(bp) 1052 struct buf *bp; 1053{ 1054 KASSERT(bp->b_qindex == QUEUE_NONE, 1055 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1056 1057 if (bp->b_flags & B_DELWRI) { 1058 bp->b_flags &= ~B_DELWRI; 1059 reassignbuf(bp, bp->b_vp); 1060 --numdirtybuffers; 1061 numdirtywakeup(lodirtybuffers); 1062 } 1063 /* 1064 * Since it is now being written, we can clear its deferred write flag. 1065 */ 1066 bp->b_flags &= ~B_DEFERRED; 1067} 1068 1069/* 1070 * bawrite: 1071 * 1072 * Asynchronous write. Start output on a buffer, but do not wait for 1073 * it to complete. The buffer is released when the output completes. 1074 * 1075 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1076 * B_INVAL buffers. Not us. 1077 */ 1078void 1079bawrite(struct buf * bp) 1080{ 1081 bp->b_flags |= B_ASYNC; 1082 (void) BUF_WRITE(bp); 1083} 1084 1085/* 1086 * bwillwrite: 1087 * 1088 * Called prior to the locking of any vnodes when we are expecting to 1089 * write. We do not want to starve the buffer cache with too many 1090 * dirty buffers so we block here. By blocking prior to the locking 1091 * of any vnodes we attempt to avoid the situation where a locked vnode 1092 * prevents the various system daemons from flushing related buffers. 1093 */ 1094 1095void 1096bwillwrite(void) 1097{ 1098 if (numdirtybuffers >= hidirtybuffers) { 1099 int s; 1100 1101 mtx_lock(&Giant); 1102 s = splbio(); 1103 while (numdirtybuffers >= hidirtybuffers) { 1104 bd_wakeup(1); 1105 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1106 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 1107 } 1108 splx(s); 1109 mtx_unlock(&Giant); 1110 } 1111} 1112 1113/* 1114 * Return true if we have too many dirty buffers. 1115 */ 1116int 1117buf_dirty_count_severe(void) 1118{ 1119 return(numdirtybuffers >= hidirtybuffers); 1120} 1121 1122/* 1123 * brelse: 1124 * 1125 * Release a busy buffer and, if requested, free its resources. The 1126 * buffer will be stashed in the appropriate bufqueue[] allowing it 1127 * to be accessed later as a cache entity or reused for other purposes. 1128 */ 1129void 1130brelse(struct buf * bp) 1131{ 1132 int s; 1133 1134 GIANT_REQUIRED; 1135 1136 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1137 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1138 1139 s = splbio(); 1140 1141 if (bp->b_flags & B_LOCKED) 1142 bp->b_ioflags &= ~BIO_ERROR; 1143 1144 if (bp->b_iocmd == BIO_WRITE && 1145 (bp->b_ioflags & BIO_ERROR) && 1146 !(bp->b_flags & B_INVAL)) { 1147 /* 1148 * Failed write, redirty. Must clear BIO_ERROR to prevent 1149 * pages from being scrapped. If B_INVAL is set then 1150 * this case is not run and the next case is run to 1151 * destroy the buffer. B_INVAL can occur if the buffer 1152 * is outside the range supported by the underlying device. 1153 */ 1154 bp->b_ioflags &= ~BIO_ERROR; 1155 bdirty(bp); 1156 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1157 (bp->b_ioflags & BIO_ERROR) || 1158 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1159 /* 1160 * Either a failed I/O or we were asked to free or not 1161 * cache the buffer. 1162 */ 1163 bp->b_flags |= B_INVAL; 1164 if (LIST_FIRST(&bp->b_dep) != NULL) 1165 buf_deallocate(bp); 1166 if (bp->b_flags & B_DELWRI) { 1167 --numdirtybuffers; 1168 numdirtywakeup(lodirtybuffers); 1169 } 1170 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1171 if ((bp->b_flags & B_VMIO) == 0) { 1172 if (bp->b_bufsize) 1173 allocbuf(bp, 0); 1174 if (bp->b_vp) 1175 brelvp(bp); 1176 } 1177 } 1178 1179 /* 1180 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1181 * is called with B_DELWRI set, the underlying pages may wind up 1182 * getting freed causing a previous write (bdwrite()) to get 'lost' 1183 * because pages associated with a B_DELWRI bp are marked clean. 1184 * 1185 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1186 * if B_DELWRI is set. 1187 * 1188 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1189 * on pages to return pages to the VM page queues. 1190 */ 1191 if (bp->b_flags & B_DELWRI) 1192 bp->b_flags &= ~B_RELBUF; 1193 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1194 bp->b_flags |= B_RELBUF; 1195 1196 /* 1197 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1198 * constituted, not even NFS buffers now. Two flags effect this. If 1199 * B_INVAL, the struct buf is invalidated but the VM object is kept 1200 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1201 * 1202 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1203 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1204 * buffer is also B_INVAL because it hits the re-dirtying code above. 1205 * 1206 * Normally we can do this whether a buffer is B_DELWRI or not. If 1207 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1208 * the commit state and we cannot afford to lose the buffer. If the 1209 * buffer has a background write in progress, we need to keep it 1210 * around to prevent it from being reconstituted and starting a second 1211 * background write. 1212 */ 1213 if ((bp->b_flags & B_VMIO) 1214 && !(bp->b_vp->v_tag == VT_NFS && 1215 !vn_isdisk(bp->b_vp, NULL) && 1216 (bp->b_flags & B_DELWRI)) 1217 ) { 1218 1219 int i, j, resid; 1220 vm_page_t m; 1221 off_t foff; 1222 vm_pindex_t poff; 1223 vm_object_t obj; 1224 struct vnode *vp; 1225 1226 vp = bp->b_vp; 1227 obj = bp->b_object; 1228 1229 /* 1230 * Get the base offset and length of the buffer. Note that 1231 * in the VMIO case if the buffer block size is not 1232 * page-aligned then b_data pointer may not be page-aligned. 1233 * But our b_pages[] array *IS* page aligned. 1234 * 1235 * block sizes less then DEV_BSIZE (usually 512) are not 1236 * supported due to the page granularity bits (m->valid, 1237 * m->dirty, etc...). 1238 * 1239 * See man buf(9) for more information 1240 */ 1241 resid = bp->b_bufsize; 1242 foff = bp->b_offset; 1243 1244 for (i = 0; i < bp->b_npages; i++) { 1245 int had_bogus = 0; 1246 1247 m = bp->b_pages[i]; 1248 vm_page_flag_clear(m, PG_ZERO); 1249 1250 /* 1251 * If we hit a bogus page, fixup *all* the bogus pages 1252 * now. 1253 */ 1254 if (m == bogus_page) { 1255 poff = OFF_TO_IDX(bp->b_offset); 1256 had_bogus = 1; 1257 1258 for (j = i; j < bp->b_npages; j++) { 1259 vm_page_t mtmp; 1260 mtmp = bp->b_pages[j]; 1261 if (mtmp == bogus_page) { 1262 mtmp = vm_page_lookup(obj, poff + j); 1263 if (!mtmp) { 1264 panic("brelse: page missing\n"); 1265 } 1266 bp->b_pages[j] = mtmp; 1267 } 1268 } 1269 1270 if ((bp->b_flags & B_INVAL) == 0) { 1271 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1272 } 1273 m = bp->b_pages[i]; 1274 } 1275 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1276 int poffset = foff & PAGE_MASK; 1277 int presid = resid > (PAGE_SIZE - poffset) ? 1278 (PAGE_SIZE - poffset) : resid; 1279 1280 KASSERT(presid >= 0, ("brelse: extra page")); 1281 vm_page_set_invalid(m, poffset, presid); 1282 if (had_bogus) 1283 printf("avoided corruption bug in bogus_page/brelse code\n"); 1284 } 1285 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1286 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1287 } 1288 1289 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1290 vfs_vmio_release(bp); 1291 1292 } else if (bp->b_flags & B_VMIO) { 1293 1294 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1295 vfs_vmio_release(bp); 1296 } 1297 1298 } 1299 1300 if (bp->b_qindex != QUEUE_NONE) 1301 panic("brelse: free buffer onto another queue???"); 1302 if (BUF_REFCNT(bp) > 1) { 1303 /* do not release to free list */ 1304 BUF_UNLOCK(bp); 1305 splx(s); 1306 return; 1307 } 1308 1309 /* enqueue */ 1310 1311 /* buffers with no memory */ 1312 if (bp->b_bufsize == 0) { 1313 bp->b_flags |= B_INVAL; 1314 bp->b_xflags &= ~BX_BKGRDWRITE; 1315 if (bp->b_xflags & BX_BKGRDINPROG) 1316 panic("losing buffer 1"); 1317 if (bp->b_kvasize) { 1318 bp->b_qindex = QUEUE_EMPTYKVA; 1319 } else { 1320 bp->b_qindex = QUEUE_EMPTY; 1321 } 1322 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1323#ifdef USE_BUFHASH 1324 LIST_REMOVE(bp, b_hash); 1325 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1326#endif 1327 bp->b_dev = NODEV; 1328 /* buffers with junk contents */ 1329 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1330 (bp->b_ioflags & BIO_ERROR)) { 1331 bp->b_flags |= B_INVAL; 1332 bp->b_xflags &= ~BX_BKGRDWRITE; 1333 if (bp->b_xflags & BX_BKGRDINPROG) 1334 panic("losing buffer 2"); 1335 bp->b_qindex = QUEUE_CLEAN; 1336 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1337#ifdef USE_BUFHASH 1338 LIST_REMOVE(bp, b_hash); 1339 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1340#endif 1341 bp->b_dev = NODEV; 1342 1343 /* buffers that are locked */ 1344 } else if (bp->b_flags & B_LOCKED) { 1345 bp->b_qindex = QUEUE_LOCKED; 1346 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1347 1348 /* remaining buffers */ 1349 } else { 1350 if (bp->b_flags & B_DELWRI) 1351 bp->b_qindex = QUEUE_DIRTY; 1352 else 1353 bp->b_qindex = QUEUE_CLEAN; 1354 if (bp->b_flags & B_AGE) 1355 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1356 else 1357 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1358 } 1359 1360 /* 1361 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1362 * placed the buffer on the correct queue. We must also disassociate 1363 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1364 * find it. 1365 */ 1366 if (bp->b_flags & B_INVAL) { 1367 if (bp->b_flags & B_DELWRI) 1368 bundirty(bp); 1369 if (bp->b_vp) 1370 brelvp(bp); 1371 } 1372 1373 /* 1374 * Fixup numfreebuffers count. The bp is on an appropriate queue 1375 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1376 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1377 * if B_INVAL is set ). 1378 */ 1379 1380 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1381 bufcountwakeup(); 1382 1383 /* 1384 * Something we can maybe free or reuse 1385 */ 1386 if (bp->b_bufsize || bp->b_kvasize) 1387 bufspacewakeup(); 1388 1389 /* unlock */ 1390 BUF_UNLOCK(bp); 1391 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1392 B_DIRECT | B_NOWDRAIN); 1393 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1394 panic("brelse: not dirty"); 1395 splx(s); 1396} 1397 1398/* 1399 * Release a buffer back to the appropriate queue but do not try to free 1400 * it. The buffer is expected to be used again soon. 1401 * 1402 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1403 * biodone() to requeue an async I/O on completion. It is also used when 1404 * known good buffers need to be requeued but we think we may need the data 1405 * again soon. 1406 * 1407 * XXX we should be able to leave the B_RELBUF hint set on completion. 1408 */ 1409void 1410bqrelse(struct buf * bp) 1411{ 1412 int s; 1413 1414 s = splbio(); 1415 1416 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1417 1418 if (bp->b_qindex != QUEUE_NONE) 1419 panic("bqrelse: free buffer onto another queue???"); 1420 if (BUF_REFCNT(bp) > 1) { 1421 /* do not release to free list */ 1422 BUF_UNLOCK(bp); 1423 splx(s); 1424 return; 1425 } 1426 if (bp->b_flags & B_LOCKED) { 1427 bp->b_ioflags &= ~BIO_ERROR; 1428 bp->b_qindex = QUEUE_LOCKED; 1429 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1430 /* buffers with stale but valid contents */ 1431 } else if (bp->b_flags & B_DELWRI) { 1432 bp->b_qindex = QUEUE_DIRTY; 1433 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1434 } else if (vm_page_count_severe()) { 1435 /* 1436 * We are too low on memory, we have to try to free the 1437 * buffer (most importantly: the wired pages making up its 1438 * backing store) *now*. 1439 */ 1440 splx(s); 1441 brelse(bp); 1442 return; 1443 } else { 1444 bp->b_qindex = QUEUE_CLEAN; 1445 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1446 } 1447 1448 if ((bp->b_flags & B_LOCKED) == 0 && 1449 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1450 bufcountwakeup(); 1451 } 1452 1453 /* 1454 * Something we can maybe free or reuse. 1455 */ 1456 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1457 bufspacewakeup(); 1458 1459 /* unlock */ 1460 BUF_UNLOCK(bp); 1461 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1462 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1463 panic("bqrelse: not dirty"); 1464 splx(s); 1465} 1466 1467/* Give pages used by the bp back to the VM system (where possible) */ 1468static void 1469vfs_vmio_release(bp) 1470 struct buf *bp; 1471{ 1472 int i; 1473 vm_page_t m; 1474 1475 GIANT_REQUIRED; 1476 1477 for (i = 0; i < bp->b_npages; i++) { 1478 m = bp->b_pages[i]; 1479 bp->b_pages[i] = NULL; 1480 /* 1481 * In order to keep page LRU ordering consistent, put 1482 * everything on the inactive queue. 1483 */ 1484 vm_page_unwire(m, 0); 1485 /* 1486 * We don't mess with busy pages, it is 1487 * the responsibility of the process that 1488 * busied the pages to deal with them. 1489 */ 1490 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1491 continue; 1492 1493 if (m->wire_count == 0) { 1494 vm_page_flag_clear(m, PG_ZERO); 1495 /* 1496 * Might as well free the page if we can and it has 1497 * no valid data. We also free the page if the 1498 * buffer was used for direct I/O 1499 */ 1500 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1501 m->hold_count == 0) { 1502 vm_page_busy(m); 1503 vm_page_protect(m, VM_PROT_NONE); 1504 vm_page_free(m); 1505 } else if (bp->b_flags & B_DIRECT) { 1506 vm_page_try_to_free(m); 1507 } else if (vm_page_count_severe()) { 1508 vm_page_try_to_cache(m); 1509 } 1510 } 1511 } 1512 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1513 1514 if (bp->b_bufsize) { 1515 bufspacewakeup(); 1516 bp->b_bufsize = 0; 1517 } 1518 bp->b_npages = 0; 1519 bp->b_flags &= ~B_VMIO; 1520 if (bp->b_vp) 1521 brelvp(bp); 1522} 1523 1524#ifdef USE_BUFHASH 1525/* 1526 * XXX MOVED TO VFS_SUBR.C 1527 * 1528 * Check to see if a block is currently memory resident. 1529 */ 1530struct buf * 1531gbincore(struct vnode * vp, daddr_t blkno) 1532{ 1533 struct buf *bp; 1534 struct bufhashhdr *bh; 1535 1536 bh = bufhash(vp, blkno); 1537 1538 /* Search hash chain */ 1539 LIST_FOREACH(bp, bh, b_hash) { 1540 /* hit */ 1541 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1542 (bp->b_flags & B_INVAL) == 0) { 1543 break; 1544 } 1545 } 1546 return (bp); 1547} 1548#endif 1549 1550/* 1551 * vfs_bio_awrite: 1552 * 1553 * Implement clustered async writes for clearing out B_DELWRI buffers. 1554 * This is much better then the old way of writing only one buffer at 1555 * a time. Note that we may not be presented with the buffers in the 1556 * correct order, so we search for the cluster in both directions. 1557 */ 1558int 1559vfs_bio_awrite(struct buf * bp) 1560{ 1561 int i; 1562 int j; 1563 daddr_t lblkno = bp->b_lblkno; 1564 struct vnode *vp = bp->b_vp; 1565 int s; 1566 int ncl; 1567 struct buf *bpa; 1568 int nwritten; 1569 int size; 1570 int maxcl; 1571 1572 s = splbio(); 1573 /* 1574 * right now we support clustered writing only to regular files. If 1575 * we find a clusterable block we could be in the middle of a cluster 1576 * rather then at the beginning. 1577 */ 1578 if ((vp->v_type == VREG) && 1579 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1580 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1581 1582 size = vp->v_mount->mnt_stat.f_iosize; 1583 maxcl = MAXPHYS / size; 1584 1585 for (i = 1; i < maxcl; i++) { 1586 if ((bpa = gbincore(vp, lblkno + i)) && 1587 BUF_REFCNT(bpa) == 0 && 1588 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1589 (B_DELWRI | B_CLUSTEROK)) && 1590 (bpa->b_bufsize == size)) { 1591 if ((bpa->b_blkno == bpa->b_lblkno) || 1592 (bpa->b_blkno != 1593 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1594 break; 1595 } else { 1596 break; 1597 } 1598 } 1599 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1600 if ((bpa = gbincore(vp, lblkno - j)) && 1601 BUF_REFCNT(bpa) == 0 && 1602 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1603 (B_DELWRI | B_CLUSTEROK)) && 1604 (bpa->b_bufsize == size)) { 1605 if ((bpa->b_blkno == bpa->b_lblkno) || 1606 (bpa->b_blkno != 1607 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1608 break; 1609 } else { 1610 break; 1611 } 1612 } 1613 --j; 1614 ncl = i + j; 1615 /* 1616 * this is a possible cluster write 1617 */ 1618 if (ncl != 1) { 1619 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1620 splx(s); 1621 return nwritten; 1622 } 1623 } 1624 1625 BUF_LOCK(bp, LK_EXCLUSIVE); 1626 bremfree(bp); 1627 bp->b_flags |= B_ASYNC; 1628 1629 splx(s); 1630 /* 1631 * default (old) behavior, writing out only one block 1632 * 1633 * XXX returns b_bufsize instead of b_bcount for nwritten? 1634 */ 1635 nwritten = bp->b_bufsize; 1636 (void) BUF_WRITE(bp); 1637 1638 return nwritten; 1639} 1640 1641/* 1642 * getnewbuf: 1643 * 1644 * Find and initialize a new buffer header, freeing up existing buffers 1645 * in the bufqueues as necessary. The new buffer is returned locked. 1646 * 1647 * Important: B_INVAL is not set. If the caller wishes to throw the 1648 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1649 * 1650 * We block if: 1651 * We have insufficient buffer headers 1652 * We have insufficient buffer space 1653 * buffer_map is too fragmented ( space reservation fails ) 1654 * If we have to flush dirty buffers ( but we try to avoid this ) 1655 * 1656 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1657 * Instead we ask the buf daemon to do it for us. We attempt to 1658 * avoid piecemeal wakeups of the pageout daemon. 1659 */ 1660 1661static struct buf * 1662getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1663{ 1664 struct buf *bp; 1665 struct buf *nbp; 1666 int defrag = 0; 1667 int nqindex; 1668 static int flushingbufs; 1669 1670 GIANT_REQUIRED; 1671 1672 /* 1673 * We can't afford to block since we might be holding a vnode lock, 1674 * which may prevent system daemons from running. We deal with 1675 * low-memory situations by proactively returning memory and running 1676 * async I/O rather then sync I/O. 1677 */ 1678 1679 ++getnewbufcalls; 1680 --getnewbufrestarts; 1681restart: 1682 ++getnewbufrestarts; 1683 1684 /* 1685 * Setup for scan. If we do not have enough free buffers, 1686 * we setup a degenerate case that immediately fails. Note 1687 * that if we are specially marked process, we are allowed to 1688 * dip into our reserves. 1689 * 1690 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1691 * 1692 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1693 * However, there are a number of cases (defragging, reusing, ...) 1694 * where we cannot backup. 1695 */ 1696 nqindex = QUEUE_EMPTYKVA; 1697 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1698 1699 if (nbp == NULL) { 1700 /* 1701 * If no EMPTYKVA buffers and we are either 1702 * defragging or reusing, locate a CLEAN buffer 1703 * to free or reuse. If bufspace useage is low 1704 * skip this step so we can allocate a new buffer. 1705 */ 1706 if (defrag || bufspace >= lobufspace) { 1707 nqindex = QUEUE_CLEAN; 1708 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1709 } 1710 1711 /* 1712 * If we could not find or were not allowed to reuse a 1713 * CLEAN buffer, check to see if it is ok to use an EMPTY 1714 * buffer. We can only use an EMPTY buffer if allocating 1715 * its KVA would not otherwise run us out of buffer space. 1716 */ 1717 if (nbp == NULL && defrag == 0 && 1718 bufspace + maxsize < hibufspace) { 1719 nqindex = QUEUE_EMPTY; 1720 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1721 } 1722 } 1723 1724 /* 1725 * Run scan, possibly freeing data and/or kva mappings on the fly 1726 * depending. 1727 */ 1728 1729 while ((bp = nbp) != NULL) { 1730 int qindex = nqindex; 1731 1732 /* 1733 * Calculate next bp ( we can only use it if we do not block 1734 * or do other fancy things ). 1735 */ 1736 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1737 switch(qindex) { 1738 case QUEUE_EMPTY: 1739 nqindex = QUEUE_EMPTYKVA; 1740 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1741 break; 1742 /* fall through */ 1743 case QUEUE_EMPTYKVA: 1744 nqindex = QUEUE_CLEAN; 1745 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1746 break; 1747 /* fall through */ 1748 case QUEUE_CLEAN: 1749 /* 1750 * nbp is NULL. 1751 */ 1752 break; 1753 } 1754 } 1755 1756 /* 1757 * Sanity Checks 1758 */ 1759 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1760 1761 /* 1762 * Note: we no longer distinguish between VMIO and non-VMIO 1763 * buffers. 1764 */ 1765 1766 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1767 1768 /* 1769 * If we are defragging then we need a buffer with 1770 * b_kvasize != 0. XXX this situation should no longer 1771 * occur, if defrag is non-zero the buffer's b_kvasize 1772 * should also be non-zero at this point. XXX 1773 */ 1774 if (defrag && bp->b_kvasize == 0) { 1775 printf("Warning: defrag empty buffer %p\n", bp); 1776 continue; 1777 } 1778 1779 /* 1780 * Start freeing the bp. This is somewhat involved. nbp 1781 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1782 */ 1783 1784 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1785 panic("getnewbuf: locked buf"); 1786 bremfree(bp); 1787 1788 if (qindex == QUEUE_CLEAN) { 1789 if (bp->b_flags & B_VMIO) { 1790 bp->b_flags &= ~B_ASYNC; 1791 vfs_vmio_release(bp); 1792 } 1793 if (bp->b_vp) 1794 brelvp(bp); 1795 } 1796 1797 /* 1798 * NOTE: nbp is now entirely invalid. We can only restart 1799 * the scan from this point on. 1800 * 1801 * Get the rest of the buffer freed up. b_kva* is still 1802 * valid after this operation. 1803 */ 1804 1805 if (bp->b_rcred != NOCRED) { 1806 crfree(bp->b_rcred); 1807 bp->b_rcred = NOCRED; 1808 } 1809 if (bp->b_wcred != NOCRED) { 1810 crfree(bp->b_wcred); 1811 bp->b_wcred = NOCRED; 1812 } 1813 if (LIST_FIRST(&bp->b_dep) != NULL) 1814 buf_deallocate(bp); 1815 if (bp->b_xflags & BX_BKGRDINPROG) 1816 panic("losing buffer 3"); 1817#ifdef USE_BUFHASH 1818 LIST_REMOVE(bp, b_hash); 1819 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1820#endif 1821 1822 if (bp->b_bufsize) 1823 allocbuf(bp, 0); 1824 1825 bp->b_flags = 0; 1826 bp->b_ioflags = 0; 1827 bp->b_xflags = 0; 1828 bp->b_dev = NODEV; 1829 bp->b_vp = NULL; 1830 bp->b_blkno = bp->b_lblkno = 0; 1831 bp->b_offset = NOOFFSET; 1832 bp->b_iodone = 0; 1833 bp->b_error = 0; 1834 bp->b_resid = 0; 1835 bp->b_bcount = 0; 1836 bp->b_npages = 0; 1837 bp->b_dirtyoff = bp->b_dirtyend = 0; 1838 bp->b_magic = B_MAGIC_BIO; 1839 bp->b_op = &buf_ops_bio; 1840 bp->b_object = NULL; 1841 1842 LIST_INIT(&bp->b_dep); 1843 1844 /* 1845 * If we are defragging then free the buffer. 1846 */ 1847 if (defrag) { 1848 bp->b_flags |= B_INVAL; 1849 bfreekva(bp); 1850 brelse(bp); 1851 defrag = 0; 1852 goto restart; 1853 } 1854 1855 /* 1856 * If we are overcomitted then recover the buffer and its 1857 * KVM space. This occurs in rare situations when multiple 1858 * processes are blocked in getnewbuf() or allocbuf(). 1859 */ 1860 if (bufspace >= hibufspace) 1861 flushingbufs = 1; 1862 if (flushingbufs && bp->b_kvasize != 0) { 1863 bp->b_flags |= B_INVAL; 1864 bfreekva(bp); 1865 brelse(bp); 1866 goto restart; 1867 } 1868 if (bufspace < lobufspace) 1869 flushingbufs = 0; 1870 break; 1871 } 1872 1873 /* 1874 * If we exhausted our list, sleep as appropriate. We may have to 1875 * wakeup various daemons and write out some dirty buffers. 1876 * 1877 * Generally we are sleeping due to insufficient buffer space. 1878 */ 1879 1880 if (bp == NULL) { 1881 int flags; 1882 char *waitmsg; 1883 1884 if (defrag) { 1885 flags = VFS_BIO_NEED_BUFSPACE; 1886 waitmsg = "nbufkv"; 1887 } else if (bufspace >= hibufspace) { 1888 waitmsg = "nbufbs"; 1889 flags = VFS_BIO_NEED_BUFSPACE; 1890 } else { 1891 waitmsg = "newbuf"; 1892 flags = VFS_BIO_NEED_ANY; 1893 } 1894 1895 bd_speedup(); /* heeeelp */ 1896 1897 needsbuffer |= flags; 1898 while (needsbuffer & flags) { 1899 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1900 waitmsg, slptimeo)) 1901 return (NULL); 1902 } 1903 } else { 1904 /* 1905 * We finally have a valid bp. We aren't quite out of the 1906 * woods, we still have to reserve kva space. In order 1907 * to keep fragmentation sane we only allocate kva in 1908 * BKVASIZE chunks. 1909 */ 1910 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1911 1912 if (maxsize != bp->b_kvasize) { 1913 vm_offset_t addr = 0; 1914 1915 bfreekva(bp); 1916 1917 if (vm_map_findspace(buffer_map, 1918 vm_map_min(buffer_map), maxsize, &addr)) { 1919 /* 1920 * Uh oh. Buffer map is to fragmented. We 1921 * must defragment the map. 1922 */ 1923 ++bufdefragcnt; 1924 defrag = 1; 1925 bp->b_flags |= B_INVAL; 1926 brelse(bp); 1927 goto restart; 1928 } 1929 if (addr) { 1930 vm_map_insert(buffer_map, NULL, 0, 1931 addr, addr + maxsize, 1932 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1933 1934 bp->b_kvabase = (caddr_t) addr; 1935 bp->b_kvasize = maxsize; 1936 bufspace += bp->b_kvasize; 1937 ++bufreusecnt; 1938 } 1939 } 1940 bp->b_data = bp->b_kvabase; 1941 } 1942 return(bp); 1943} 1944 1945/* 1946 * buf_daemon: 1947 * 1948 * buffer flushing daemon. Buffers are normally flushed by the 1949 * update daemon but if it cannot keep up this process starts to 1950 * take the load in an attempt to prevent getnewbuf() from blocking. 1951 */ 1952 1953static struct proc *bufdaemonproc; 1954 1955static struct kproc_desc buf_kp = { 1956 "bufdaemon", 1957 buf_daemon, 1958 &bufdaemonproc 1959}; 1960SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1961 1962static void 1963buf_daemon() 1964{ 1965 int s; 1966 1967 mtx_lock(&Giant); 1968 1969 /* 1970 * This process needs to be suspended prior to shutdown sync. 1971 */ 1972 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 1973 SHUTDOWN_PRI_LAST); 1974 1975 /* 1976 * This process is allowed to take the buffer cache to the limit 1977 */ 1978 s = splbio(); 1979 1980 for (;;) { 1981 kthread_suspend_check(bufdaemonproc); 1982 1983 bd_request = 0; 1984 1985 /* 1986 * Do the flush. Limit the amount of in-transit I/O we 1987 * allow to build up, otherwise we would completely saturate 1988 * the I/O system. Wakeup any waiting processes before we 1989 * normally would so they can run in parallel with our drain. 1990 */ 1991 while (numdirtybuffers > lodirtybuffers) { 1992 if (flushbufqueues() == 0) 1993 break; 1994 waitrunningbufspace(); 1995 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 1996 } 1997 1998 /* 1999 * Only clear bd_request if we have reached our low water 2000 * mark. The buf_daemon normally waits 1 second and 2001 * then incrementally flushes any dirty buffers that have 2002 * built up, within reason. 2003 * 2004 * If we were unable to hit our low water mark and couldn't 2005 * find any flushable buffers, we sleep half a second. 2006 * Otherwise we loop immediately. 2007 */ 2008 if (numdirtybuffers <= lodirtybuffers) { 2009 /* 2010 * We reached our low water mark, reset the 2011 * request and sleep until we are needed again. 2012 * The sleep is just so the suspend code works. 2013 */ 2014 bd_request = 0; 2015 tsleep(&bd_request, PVM, "psleep", hz); 2016 } else { 2017 /* 2018 * We couldn't find any flushable dirty buffers but 2019 * still have too many dirty buffers, we 2020 * have to sleep and try again. (rare) 2021 */ 2022 tsleep(&bd_request, PVM, "qsleep", hz / 2); 2023 } 2024 } 2025} 2026 2027/* 2028 * flushbufqueues: 2029 * 2030 * Try to flush a buffer in the dirty queue. We must be careful to 2031 * free up B_INVAL buffers instead of write them, which NFS is 2032 * particularly sensitive to. 2033 */ 2034 2035static int 2036flushbufqueues(void) 2037{ 2038 struct buf *bp; 2039 int r = 0; 2040 2041 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 2042 2043 while (bp) { 2044 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 2045 if ((bp->b_flags & B_DELWRI) != 0 && 2046 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 2047 if (bp->b_flags & B_INVAL) { 2048 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 2049 panic("flushbufqueues: locked buf"); 2050 bremfree(bp); 2051 brelse(bp); 2052 ++r; 2053 break; 2054 } 2055 if (LIST_FIRST(&bp->b_dep) != NULL && 2056 (bp->b_flags & B_DEFERRED) == 0 && 2057 buf_countdeps(bp, 0)) { 2058 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 2059 bp, b_freelist); 2060 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 2061 bp, b_freelist); 2062 bp->b_flags |= B_DEFERRED; 2063 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 2064 continue; 2065 } 2066 vfs_bio_awrite(bp); 2067 ++r; 2068 break; 2069 } 2070 bp = TAILQ_NEXT(bp, b_freelist); 2071 } 2072 return (r); 2073} 2074 2075/* 2076 * Check to see if a block is currently memory resident. 2077 */ 2078struct buf * 2079incore(struct vnode * vp, daddr_t blkno) 2080{ 2081 struct buf *bp; 2082 2083 int s = splbio(); 2084 bp = gbincore(vp, blkno); 2085 splx(s); 2086 return (bp); 2087} 2088 2089/* 2090 * Returns true if no I/O is needed to access the 2091 * associated VM object. This is like incore except 2092 * it also hunts around in the VM system for the data. 2093 */ 2094 2095int 2096inmem(struct vnode * vp, daddr_t blkno) 2097{ 2098 vm_object_t obj; 2099 vm_offset_t toff, tinc, size; 2100 vm_page_t m; 2101 vm_ooffset_t off; 2102 2103 GIANT_REQUIRED; 2104 2105 if (incore(vp, blkno)) 2106 return 1; 2107 if (vp->v_mount == NULL) 2108 return 0; 2109 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) 2110 return 0; 2111 2112 size = PAGE_SIZE; 2113 if (size > vp->v_mount->mnt_stat.f_iosize) 2114 size = vp->v_mount->mnt_stat.f_iosize; 2115 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2116 2117 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2118 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2119 if (!m) 2120 goto notinmem; 2121 tinc = size; 2122 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2123 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2124 if (vm_page_is_valid(m, 2125 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2126 goto notinmem; 2127 } 2128 return 1; 2129 2130notinmem: 2131 return (0); 2132} 2133 2134/* 2135 * vfs_setdirty: 2136 * 2137 * Sets the dirty range for a buffer based on the status of the dirty 2138 * bits in the pages comprising the buffer. 2139 * 2140 * The range is limited to the size of the buffer. 2141 * 2142 * This routine is primarily used by NFS, but is generalized for the 2143 * B_VMIO case. 2144 */ 2145static void 2146vfs_setdirty(struct buf *bp) 2147{ 2148 int i; 2149 vm_object_t object; 2150 2151 GIANT_REQUIRED; 2152 /* 2153 * Degenerate case - empty buffer 2154 */ 2155 2156 if (bp->b_bufsize == 0) 2157 return; 2158 2159 /* 2160 * We qualify the scan for modified pages on whether the 2161 * object has been flushed yet. The OBJ_WRITEABLE flag 2162 * is not cleared simply by protecting pages off. 2163 */ 2164 2165 if ((bp->b_flags & B_VMIO) == 0) 2166 return; 2167 2168 object = bp->b_pages[0]->object; 2169 2170 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2171 printf("Warning: object %p writeable but not mightbedirty\n", object); 2172 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2173 printf("Warning: object %p mightbedirty but not writeable\n", object); 2174 2175 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2176 vm_offset_t boffset; 2177 vm_offset_t eoffset; 2178 2179 /* 2180 * test the pages to see if they have been modified directly 2181 * by users through the VM system. 2182 */ 2183 for (i = 0; i < bp->b_npages; i++) { 2184 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2185 vm_page_test_dirty(bp->b_pages[i]); 2186 } 2187 2188 /* 2189 * Calculate the encompassing dirty range, boffset and eoffset, 2190 * (eoffset - boffset) bytes. 2191 */ 2192 2193 for (i = 0; i < bp->b_npages; i++) { 2194 if (bp->b_pages[i]->dirty) 2195 break; 2196 } 2197 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2198 2199 for (i = bp->b_npages - 1; i >= 0; --i) { 2200 if (bp->b_pages[i]->dirty) { 2201 break; 2202 } 2203 } 2204 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2205 2206 /* 2207 * Fit it to the buffer. 2208 */ 2209 2210 if (eoffset > bp->b_bcount) 2211 eoffset = bp->b_bcount; 2212 2213 /* 2214 * If we have a good dirty range, merge with the existing 2215 * dirty range. 2216 */ 2217 2218 if (boffset < eoffset) { 2219 if (bp->b_dirtyoff > boffset) 2220 bp->b_dirtyoff = boffset; 2221 if (bp->b_dirtyend < eoffset) 2222 bp->b_dirtyend = eoffset; 2223 } 2224 } 2225} 2226 2227/* 2228 * getblk: 2229 * 2230 * Get a block given a specified block and offset into a file/device. 2231 * The buffers B_DONE bit will be cleared on return, making it almost 2232 * ready for an I/O initiation. B_INVAL may or may not be set on 2233 * return. The caller should clear B_INVAL prior to initiating a 2234 * READ. 2235 * 2236 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2237 * an existing buffer. 2238 * 2239 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2240 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2241 * and then cleared based on the backing VM. If the previous buffer is 2242 * non-0-sized but invalid, B_CACHE will be cleared. 2243 * 2244 * If getblk() must create a new buffer, the new buffer is returned with 2245 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2246 * case it is returned with B_INVAL clear and B_CACHE set based on the 2247 * backing VM. 2248 * 2249 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2250 * B_CACHE bit is clear. 2251 * 2252 * What this means, basically, is that the caller should use B_CACHE to 2253 * determine whether the buffer is fully valid or not and should clear 2254 * B_INVAL prior to issuing a read. If the caller intends to validate 2255 * the buffer by loading its data area with something, the caller needs 2256 * to clear B_INVAL. If the caller does this without issuing an I/O, 2257 * the caller should set B_CACHE ( as an optimization ), else the caller 2258 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2259 * a write attempt or if it was a successfull read. If the caller 2260 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2261 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2262 */ 2263struct buf * 2264getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2265{ 2266 struct buf *bp; 2267 int s; 2268#ifdef USE_BUFHASH 2269 struct bufhashhdr *bh; 2270#endif 2271 2272 if (size > MAXBSIZE) 2273 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2274 2275 s = splbio(); 2276loop: 2277 /* 2278 * Block if we are low on buffers. Certain processes are allowed 2279 * to completely exhaust the buffer cache. 2280 * 2281 * If this check ever becomes a bottleneck it may be better to 2282 * move it into the else, when gbincore() fails. At the moment 2283 * it isn't a problem. 2284 * 2285 * XXX remove if 0 sections (clean this up after its proven) 2286 */ 2287 if (numfreebuffers == 0) { 2288 if (curthread == PCPU_GET(idlethread)) 2289 return NULL; 2290 needsbuffer |= VFS_BIO_NEED_ANY; 2291 } 2292 2293 if ((bp = gbincore(vp, blkno))) { 2294 /* 2295 * Buffer is in-core. If the buffer is not busy, it must 2296 * be on a queue. 2297 */ 2298 2299 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2300 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2301 "getblk", slpflag, slptimeo) == ENOLCK) 2302 goto loop; 2303 splx(s); 2304 return (struct buf *) NULL; 2305 } 2306 2307 /* 2308 * The buffer is locked. B_CACHE is cleared if the buffer is 2309 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2310 * and for a VMIO buffer B_CACHE is adjusted according to the 2311 * backing VM cache. 2312 */ 2313 if (bp->b_flags & B_INVAL) 2314 bp->b_flags &= ~B_CACHE; 2315 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2316 bp->b_flags |= B_CACHE; 2317 bremfree(bp); 2318 2319 /* 2320 * check for size inconsistancies for non-VMIO case. 2321 */ 2322 2323 if (bp->b_bcount != size) { 2324 if ((bp->b_flags & B_VMIO) == 0 || 2325 (size > bp->b_kvasize)) { 2326 if (bp->b_flags & B_DELWRI) { 2327 bp->b_flags |= B_NOCACHE; 2328 BUF_WRITE(bp); 2329 } else { 2330 if ((bp->b_flags & B_VMIO) && 2331 (LIST_FIRST(&bp->b_dep) == NULL)) { 2332 bp->b_flags |= B_RELBUF; 2333 brelse(bp); 2334 } else { 2335 bp->b_flags |= B_NOCACHE; 2336 BUF_WRITE(bp); 2337 } 2338 } 2339 goto loop; 2340 } 2341 } 2342 2343 /* 2344 * If the size is inconsistant in the VMIO case, we can resize 2345 * the buffer. This might lead to B_CACHE getting set or 2346 * cleared. If the size has not changed, B_CACHE remains 2347 * unchanged from its previous state. 2348 */ 2349 2350 if (bp->b_bcount != size) 2351 allocbuf(bp, size); 2352 2353 KASSERT(bp->b_offset != NOOFFSET, 2354 ("getblk: no buffer offset")); 2355 2356 /* 2357 * A buffer with B_DELWRI set and B_CACHE clear must 2358 * be committed before we can return the buffer in 2359 * order to prevent the caller from issuing a read 2360 * ( due to B_CACHE not being set ) and overwriting 2361 * it. 2362 * 2363 * Most callers, including NFS and FFS, need this to 2364 * operate properly either because they assume they 2365 * can issue a read if B_CACHE is not set, or because 2366 * ( for example ) an uncached B_DELWRI might loop due 2367 * to softupdates re-dirtying the buffer. In the latter 2368 * case, B_CACHE is set after the first write completes, 2369 * preventing further loops. 2370 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2371 * above while extending the buffer, we cannot allow the 2372 * buffer to remain with B_CACHE set after the write 2373 * completes or it will represent a corrupt state. To 2374 * deal with this we set B_NOCACHE to scrap the buffer 2375 * after the write. 2376 * 2377 * We might be able to do something fancy, like setting 2378 * B_CACHE in bwrite() except if B_DELWRI is already set, 2379 * so the below call doesn't set B_CACHE, but that gets real 2380 * confusing. This is much easier. 2381 */ 2382 2383 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2384 bp->b_flags |= B_NOCACHE; 2385 BUF_WRITE(bp); 2386 goto loop; 2387 } 2388 2389 splx(s); 2390 bp->b_flags &= ~B_DONE; 2391 } else { 2392 /* 2393 * Buffer is not in-core, create new buffer. The buffer 2394 * returned by getnewbuf() is locked. Note that the returned 2395 * buffer is also considered valid (not marked B_INVAL). 2396 */ 2397 int bsize, maxsize, vmio; 2398 off_t offset; 2399 2400 if (vn_isdisk(vp, NULL)) 2401 bsize = DEV_BSIZE; 2402 else if (vp->v_mountedhere) 2403 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2404 else if (vp->v_mount) 2405 bsize = vp->v_mount->mnt_stat.f_iosize; 2406 else 2407 bsize = size; 2408 2409 offset = blkno * bsize; 2410 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); 2411 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2412 maxsize = imax(maxsize, bsize); 2413 2414 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2415 if (slpflag || slptimeo) { 2416 splx(s); 2417 return NULL; 2418 } 2419 goto loop; 2420 } 2421 2422 /* 2423 * This code is used to make sure that a buffer is not 2424 * created while the getnewbuf routine is blocked. 2425 * This can be a problem whether the vnode is locked or not. 2426 * If the buffer is created out from under us, we have to 2427 * throw away the one we just created. There is now window 2428 * race because we are safely running at splbio() from the 2429 * point of the duplicate buffer creation through to here, 2430 * and we've locked the buffer. 2431 * 2432 * Note: this must occur before we associate the buffer 2433 * with the vp especially considering limitations in 2434 * the splay tree implementation when dealing with duplicate 2435 * lblkno's. 2436 */ 2437 if (gbincore(vp, blkno)) { 2438 bp->b_flags |= B_INVAL; 2439 brelse(bp); 2440 goto loop; 2441 } 2442 2443 /* 2444 * Insert the buffer into the hash, so that it can 2445 * be found by incore. 2446 */ 2447 bp->b_blkno = bp->b_lblkno = blkno; 2448 bp->b_offset = offset; 2449 2450 bgetvp(vp, bp); 2451#ifdef USE_BUFHASH 2452 LIST_REMOVE(bp, b_hash); 2453 bh = bufhash(vp, blkno); 2454 LIST_INSERT_HEAD(bh, bp, b_hash); 2455#endif 2456 2457 /* 2458 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2459 * buffer size starts out as 0, B_CACHE will be set by 2460 * allocbuf() for the VMIO case prior to it testing the 2461 * backing store for validity. 2462 */ 2463 2464 if (vmio) { 2465 bp->b_flags |= B_VMIO; 2466#if defined(VFS_BIO_DEBUG) 2467 if (vp->v_type != VREG) 2468 printf("getblk: vmioing file type %d???\n", vp->v_type); 2469#endif 2470 VOP_GETVOBJECT(vp, &bp->b_object); 2471 } else { 2472 bp->b_flags &= ~B_VMIO; 2473 bp->b_object = NULL; 2474 } 2475 2476 allocbuf(bp, size); 2477 2478 splx(s); 2479 bp->b_flags &= ~B_DONE; 2480 } 2481 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2482 return (bp); 2483} 2484 2485/* 2486 * Get an empty, disassociated buffer of given size. The buffer is initially 2487 * set to B_INVAL. 2488 */ 2489struct buf * 2490geteblk(int size) 2491{ 2492 struct buf *bp; 2493 int s; 2494 int maxsize; 2495 2496 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2497 2498 s = splbio(); 2499 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); 2500 splx(s); 2501 allocbuf(bp, size); 2502 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2503 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2504 return (bp); 2505} 2506 2507 2508/* 2509 * This code constitutes the buffer memory from either anonymous system 2510 * memory (in the case of non-VMIO operations) or from an associated 2511 * VM object (in the case of VMIO operations). This code is able to 2512 * resize a buffer up or down. 2513 * 2514 * Note that this code is tricky, and has many complications to resolve 2515 * deadlock or inconsistant data situations. Tread lightly!!! 2516 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2517 * the caller. Calling this code willy nilly can result in the loss of data. 2518 * 2519 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2520 * B_CACHE for the non-VMIO case. 2521 */ 2522 2523int 2524allocbuf(struct buf *bp, int size) 2525{ 2526 int newbsize, mbsize; 2527 int i; 2528 2529 GIANT_REQUIRED; 2530 2531 if (BUF_REFCNT(bp) == 0) 2532 panic("allocbuf: buffer not busy"); 2533 2534 if (bp->b_kvasize < size) 2535 panic("allocbuf: buffer too small"); 2536 2537 if ((bp->b_flags & B_VMIO) == 0) { 2538 caddr_t origbuf; 2539 int origbufsize; 2540 /* 2541 * Just get anonymous memory from the kernel. Don't 2542 * mess with B_CACHE. 2543 */ 2544 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2545 if (bp->b_flags & B_MALLOC) 2546 newbsize = mbsize; 2547 else 2548 newbsize = round_page(size); 2549 2550 if (newbsize < bp->b_bufsize) { 2551 /* 2552 * malloced buffers are not shrunk 2553 */ 2554 if (bp->b_flags & B_MALLOC) { 2555 if (newbsize) { 2556 bp->b_bcount = size; 2557 } else { 2558 free(bp->b_data, M_BIOBUF); 2559 if (bp->b_bufsize) { 2560 bufmallocspace -= bp->b_bufsize; 2561 bufspacewakeup(); 2562 bp->b_bufsize = 0; 2563 } 2564 bp->b_data = bp->b_kvabase; 2565 bp->b_bcount = 0; 2566 bp->b_flags &= ~B_MALLOC; 2567 } 2568 return 1; 2569 } 2570 vm_hold_free_pages( 2571 bp, 2572 (vm_offset_t) bp->b_data + newbsize, 2573 (vm_offset_t) bp->b_data + bp->b_bufsize); 2574 } else if (newbsize > bp->b_bufsize) { 2575 /* 2576 * We only use malloced memory on the first allocation. 2577 * and revert to page-allocated memory when the buffer 2578 * grows. 2579 */ 2580 if ( (bufmallocspace < maxbufmallocspace) && 2581 (bp->b_bufsize == 0) && 2582 (mbsize <= PAGE_SIZE/2)) { 2583 2584 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2585 bp->b_bufsize = mbsize; 2586 bp->b_bcount = size; 2587 bp->b_flags |= B_MALLOC; 2588 bufmallocspace += mbsize; 2589 return 1; 2590 } 2591 origbuf = NULL; 2592 origbufsize = 0; 2593 /* 2594 * If the buffer is growing on its other-than-first allocation, 2595 * then we revert to the page-allocation scheme. 2596 */ 2597 if (bp->b_flags & B_MALLOC) { 2598 origbuf = bp->b_data; 2599 origbufsize = bp->b_bufsize; 2600 bp->b_data = bp->b_kvabase; 2601 if (bp->b_bufsize) { 2602 bufmallocspace -= bp->b_bufsize; 2603 bufspacewakeup(); 2604 bp->b_bufsize = 0; 2605 } 2606 bp->b_flags &= ~B_MALLOC; 2607 newbsize = round_page(newbsize); 2608 } 2609 vm_hold_load_pages( 2610 bp, 2611 (vm_offset_t) bp->b_data + bp->b_bufsize, 2612 (vm_offset_t) bp->b_data + newbsize); 2613 if (origbuf) { 2614 bcopy(origbuf, bp->b_data, origbufsize); 2615 free(origbuf, M_BIOBUF); 2616 } 2617 } 2618 } else { 2619 vm_page_t m; 2620 int desiredpages; 2621 2622 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2623 desiredpages = (size == 0) ? 0 : 2624 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2625 2626 if (bp->b_flags & B_MALLOC) 2627 panic("allocbuf: VMIO buffer can't be malloced"); 2628 /* 2629 * Set B_CACHE initially if buffer is 0 length or will become 2630 * 0-length. 2631 */ 2632 if (size == 0 || bp->b_bufsize == 0) 2633 bp->b_flags |= B_CACHE; 2634 2635 if (newbsize < bp->b_bufsize) { 2636 /* 2637 * DEV_BSIZE aligned new buffer size is less then the 2638 * DEV_BSIZE aligned existing buffer size. Figure out 2639 * if we have to remove any pages. 2640 */ 2641 if (desiredpages < bp->b_npages) { 2642 for (i = desiredpages; i < bp->b_npages; i++) { 2643 /* 2644 * the page is not freed here -- it 2645 * is the responsibility of 2646 * vnode_pager_setsize 2647 */ 2648 m = bp->b_pages[i]; 2649 KASSERT(m != bogus_page, 2650 ("allocbuf: bogus page found")); 2651 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2652 ; 2653 2654 bp->b_pages[i] = NULL; 2655 vm_page_unwire(m, 0); 2656 } 2657 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2658 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2659 bp->b_npages = desiredpages; 2660 } 2661 } else if (size > bp->b_bcount) { 2662 /* 2663 * We are growing the buffer, possibly in a 2664 * byte-granular fashion. 2665 */ 2666 struct vnode *vp; 2667 vm_object_t obj; 2668 vm_offset_t toff; 2669 vm_offset_t tinc; 2670 2671 /* 2672 * Step 1, bring in the VM pages from the object, 2673 * allocating them if necessary. We must clear 2674 * B_CACHE if these pages are not valid for the 2675 * range covered by the buffer. 2676 */ 2677 2678 vp = bp->b_vp; 2679 obj = bp->b_object; 2680 2681 while (bp->b_npages < desiredpages) { 2682 vm_page_t m; 2683 vm_pindex_t pi; 2684 2685 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2686 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2687 /* 2688 * note: must allocate system pages 2689 * since blocking here could intefere 2690 * with paging I/O, no matter which 2691 * process we are. 2692 */ 2693 m = vm_page_alloc(obj, pi, VM_ALLOC_SYSTEM); 2694 if (m == NULL) { 2695 VM_WAIT; 2696 vm_pageout_deficit += desiredpages - bp->b_npages; 2697 } else { 2698 vm_page_wire(m); 2699 vm_page_wakeup(m); 2700 bp->b_flags &= ~B_CACHE; 2701 bp->b_pages[bp->b_npages] = m; 2702 ++bp->b_npages; 2703 } 2704 continue; 2705 } 2706 2707 /* 2708 * We found a page. If we have to sleep on it, 2709 * retry because it might have gotten freed out 2710 * from under us. 2711 * 2712 * We can only test PG_BUSY here. Blocking on 2713 * m->busy might lead to a deadlock: 2714 * 2715 * vm_fault->getpages->cluster_read->allocbuf 2716 * 2717 */ 2718 2719 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2720 continue; 2721 2722 /* 2723 * We have a good page. Should we wakeup the 2724 * page daemon? 2725 */ 2726 if ((curproc != pageproc) && 2727 ((m->queue - m->pc) == PQ_CACHE) && 2728 ((cnt.v_free_count + cnt.v_cache_count) < 2729 (cnt.v_free_min + cnt.v_cache_min))) { 2730 pagedaemon_wakeup(); 2731 } 2732 vm_page_flag_clear(m, PG_ZERO); 2733 vm_page_wire(m); 2734 bp->b_pages[bp->b_npages] = m; 2735 ++bp->b_npages; 2736 } 2737 2738 /* 2739 * Step 2. We've loaded the pages into the buffer, 2740 * we have to figure out if we can still have B_CACHE 2741 * set. Note that B_CACHE is set according to the 2742 * byte-granular range ( bcount and size ), new the 2743 * aligned range ( newbsize ). 2744 * 2745 * The VM test is against m->valid, which is DEV_BSIZE 2746 * aligned. Needless to say, the validity of the data 2747 * needs to also be DEV_BSIZE aligned. Note that this 2748 * fails with NFS if the server or some other client 2749 * extends the file's EOF. If our buffer is resized, 2750 * B_CACHE may remain set! XXX 2751 */ 2752 2753 toff = bp->b_bcount; 2754 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2755 2756 while ((bp->b_flags & B_CACHE) && toff < size) { 2757 vm_pindex_t pi; 2758 2759 if (tinc > (size - toff)) 2760 tinc = size - toff; 2761 2762 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2763 PAGE_SHIFT; 2764 2765 vfs_buf_test_cache( 2766 bp, 2767 bp->b_offset, 2768 toff, 2769 tinc, 2770 bp->b_pages[pi] 2771 ); 2772 toff += tinc; 2773 tinc = PAGE_SIZE; 2774 } 2775 2776 /* 2777 * Step 3, fixup the KVM pmap. Remember that 2778 * bp->b_data is relative to bp->b_offset, but 2779 * bp->b_offset may be offset into the first page. 2780 */ 2781 2782 bp->b_data = (caddr_t) 2783 trunc_page((vm_offset_t)bp->b_data); 2784 pmap_qenter( 2785 (vm_offset_t)bp->b_data, 2786 bp->b_pages, 2787 bp->b_npages 2788 ); 2789 2790 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2791 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2792 } 2793 } 2794 if (newbsize < bp->b_bufsize) 2795 bufspacewakeup(); 2796 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2797 bp->b_bcount = size; /* requested buffer size */ 2798 return 1; 2799} 2800 2801/* 2802 * bufwait: 2803 * 2804 * Wait for buffer I/O completion, returning error status. The buffer 2805 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2806 * error and cleared. 2807 */ 2808int 2809bufwait(register struct buf * bp) 2810{ 2811 int s; 2812 2813 s = splbio(); 2814 while ((bp->b_flags & B_DONE) == 0) { 2815 if (bp->b_iocmd == BIO_READ) 2816 tsleep(bp, PRIBIO, "biord", 0); 2817 else 2818 tsleep(bp, PRIBIO, "biowr", 0); 2819 } 2820 splx(s); 2821 if (bp->b_flags & B_EINTR) { 2822 bp->b_flags &= ~B_EINTR; 2823 return (EINTR); 2824 } 2825 if (bp->b_ioflags & BIO_ERROR) { 2826 return (bp->b_error ? bp->b_error : EIO); 2827 } else { 2828 return (0); 2829 } 2830} 2831 2832 /* 2833 * Call back function from struct bio back up to struct buf. 2834 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 2835 */ 2836void 2837bufdonebio(struct bio *bp) 2838{ 2839 bufdone(bp->bio_caller2); 2840} 2841 2842/* 2843 * bufdone: 2844 * 2845 * Finish I/O on a buffer, optionally calling a completion function. 2846 * This is usually called from an interrupt so process blocking is 2847 * not allowed. 2848 * 2849 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2850 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2851 * assuming B_INVAL is clear. 2852 * 2853 * For the VMIO case, we set B_CACHE if the op was a read and no 2854 * read error occured, or if the op was a write. B_CACHE is never 2855 * set if the buffer is invalid or otherwise uncacheable. 2856 * 2857 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2858 * initiator to leave B_INVAL set to brelse the buffer out of existance 2859 * in the biodone routine. 2860 */ 2861void 2862bufdone(struct buf *bp) 2863{ 2864 int s; 2865 void (*biodone)(struct buf *); 2866 2867 GIANT_REQUIRED; 2868 2869 s = splbio(); 2870 2871 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2872 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2873 2874 bp->b_flags |= B_DONE; 2875 runningbufwakeup(bp); 2876 2877 if (bp->b_iocmd == BIO_DELETE) { 2878 brelse(bp); 2879 splx(s); 2880 return; 2881 } 2882 2883 if (bp->b_iocmd == BIO_WRITE) { 2884 vwakeup(bp); 2885 } 2886 2887 /* call optional completion function if requested */ 2888 if (bp->b_iodone != NULL) { 2889 biodone = bp->b_iodone; 2890 bp->b_iodone = NULL; 2891 (*biodone) (bp); 2892 splx(s); 2893 return; 2894 } 2895 if (LIST_FIRST(&bp->b_dep) != NULL) 2896 buf_complete(bp); 2897 2898 if (bp->b_flags & B_VMIO) { 2899 int i; 2900 vm_ooffset_t foff; 2901 vm_page_t m; 2902 vm_object_t obj; 2903 int iosize; 2904 struct vnode *vp = bp->b_vp; 2905 2906 obj = bp->b_object; 2907 2908#if defined(VFS_BIO_DEBUG) 2909 if (vp->v_usecount == 0) { 2910 panic("biodone: zero vnode ref count"); 2911 } 2912 2913 if ((vp->v_flag & VOBJBUF) == 0) { 2914 panic("biodone: vnode is not setup for merged cache"); 2915 } 2916#endif 2917 2918 foff = bp->b_offset; 2919 KASSERT(bp->b_offset != NOOFFSET, 2920 ("biodone: no buffer offset")); 2921 2922#if defined(VFS_BIO_DEBUG) 2923 if (obj->paging_in_progress < bp->b_npages) { 2924 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2925 obj->paging_in_progress, bp->b_npages); 2926 } 2927#endif 2928 2929 /* 2930 * Set B_CACHE if the op was a normal read and no error 2931 * occured. B_CACHE is set for writes in the b*write() 2932 * routines. 2933 */ 2934 iosize = bp->b_bcount - bp->b_resid; 2935 if (bp->b_iocmd == BIO_READ && 2936 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 2937 !(bp->b_ioflags & BIO_ERROR)) { 2938 bp->b_flags |= B_CACHE; 2939 } 2940 2941 for (i = 0; i < bp->b_npages; i++) { 2942 int bogusflag = 0; 2943 int resid; 2944 2945 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2946 if (resid > iosize) 2947 resid = iosize; 2948 2949 /* 2950 * cleanup bogus pages, restoring the originals 2951 */ 2952 m = bp->b_pages[i]; 2953 if (m == bogus_page) { 2954 bogusflag = 1; 2955 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2956 if (m == NULL) 2957 panic("biodone: page disappeared!"); 2958 bp->b_pages[i] = m; 2959 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2960 } 2961#if defined(VFS_BIO_DEBUG) 2962 if (OFF_TO_IDX(foff) != m->pindex) { 2963 printf( 2964"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 2965 (intmax_t)foff, (uintmax_t)m->pindex); 2966 } 2967#endif 2968 2969 /* 2970 * In the write case, the valid and clean bits are 2971 * already changed correctly ( see bdwrite() ), so we 2972 * only need to do this here in the read case. 2973 */ 2974 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 2975 vfs_page_set_valid(bp, foff, i, m); 2976 } 2977 vm_page_flag_clear(m, PG_ZERO); 2978 2979 /* 2980 * when debugging new filesystems or buffer I/O methods, this 2981 * is the most common error that pops up. if you see this, you 2982 * have not set the page busy flag correctly!!! 2983 */ 2984 if (m->busy == 0) { 2985 printf("biodone: page busy < 0, " 2986 "pindex: %d, foff: 0x(%x,%x), " 2987 "resid: %d, index: %d\n", 2988 (int) m->pindex, (int)(foff >> 32), 2989 (int) foff & 0xffffffff, resid, i); 2990 if (!vn_isdisk(vp, NULL)) 2991 printf(" iosize: %ld, lblkno: %jd, flags: 0x%lx, npages: %d\n", 2992 bp->b_vp->v_mount->mnt_stat.f_iosize, 2993 (intmax_t) bp->b_lblkno, 2994 bp->b_flags, bp->b_npages); 2995 else 2996 printf(" VDEV, lblkno: %jd, flags: 0x%lx, npages: %d\n", 2997 (intmax_t) bp->b_lblkno, 2998 bp->b_flags, bp->b_npages); 2999 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3000 m->valid, m->dirty, m->wire_count); 3001 panic("biodone: page busy < 0\n"); 3002 } 3003 vm_page_io_finish(m); 3004 vm_object_pip_subtract(obj, 1); 3005 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3006 iosize -= resid; 3007 } 3008 if (obj) 3009 vm_object_pip_wakeupn(obj, 0); 3010 } 3011 3012 /* 3013 * For asynchronous completions, release the buffer now. The brelse 3014 * will do a wakeup there if necessary - so no need to do a wakeup 3015 * here in the async case. The sync case always needs to do a wakeup. 3016 */ 3017 3018 if (bp->b_flags & B_ASYNC) { 3019 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3020 brelse(bp); 3021 else 3022 bqrelse(bp); 3023 } else { 3024 wakeup(bp); 3025 } 3026 splx(s); 3027} 3028 3029/* 3030 * This routine is called in lieu of iodone in the case of 3031 * incomplete I/O. This keeps the busy status for pages 3032 * consistant. 3033 */ 3034void 3035vfs_unbusy_pages(struct buf * bp) 3036{ 3037 int i; 3038 3039 GIANT_REQUIRED; 3040 3041 runningbufwakeup(bp); 3042 if (bp->b_flags & B_VMIO) { 3043 vm_object_t obj; 3044 3045 obj = bp->b_object; 3046 3047 for (i = 0; i < bp->b_npages; i++) { 3048 vm_page_t m = bp->b_pages[i]; 3049 3050 if (m == bogus_page) { 3051 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3052 if (!m) { 3053 panic("vfs_unbusy_pages: page missing\n"); 3054 } 3055 bp->b_pages[i] = m; 3056 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3057 } 3058 vm_object_pip_subtract(obj, 1); 3059 vm_page_flag_clear(m, PG_ZERO); 3060 vm_page_io_finish(m); 3061 } 3062 vm_object_pip_wakeupn(obj, 0); 3063 } 3064} 3065 3066/* 3067 * vfs_page_set_valid: 3068 * 3069 * Set the valid bits in a page based on the supplied offset. The 3070 * range is restricted to the buffer's size. 3071 * 3072 * This routine is typically called after a read completes. 3073 */ 3074static void 3075vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3076{ 3077 vm_ooffset_t soff, eoff; 3078 3079 GIANT_REQUIRED; 3080 /* 3081 * Start and end offsets in buffer. eoff - soff may not cross a 3082 * page boundry or cross the end of the buffer. The end of the 3083 * buffer, in this case, is our file EOF, not the allocation size 3084 * of the buffer. 3085 */ 3086 soff = off; 3087 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3088 if (eoff > bp->b_offset + bp->b_bcount) 3089 eoff = bp->b_offset + bp->b_bcount; 3090 3091 /* 3092 * Set valid range. This is typically the entire buffer and thus the 3093 * entire page. 3094 */ 3095 if (eoff > soff) { 3096 vm_page_set_validclean( 3097 m, 3098 (vm_offset_t) (soff & PAGE_MASK), 3099 (vm_offset_t) (eoff - soff) 3100 ); 3101 } 3102} 3103 3104/* 3105 * This routine is called before a device strategy routine. 3106 * It is used to tell the VM system that paging I/O is in 3107 * progress, and treat the pages associated with the buffer 3108 * almost as being PG_BUSY. Also the object paging_in_progress 3109 * flag is handled to make sure that the object doesn't become 3110 * inconsistant. 3111 * 3112 * Since I/O has not been initiated yet, certain buffer flags 3113 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3114 * and should be ignored. 3115 */ 3116void 3117vfs_busy_pages(struct buf * bp, int clear_modify) 3118{ 3119 int i, bogus; 3120 3121 GIANT_REQUIRED; 3122 3123 if (bp->b_flags & B_VMIO) { 3124 vm_object_t obj; 3125 vm_ooffset_t foff; 3126 3127 obj = bp->b_object; 3128 foff = bp->b_offset; 3129 KASSERT(bp->b_offset != NOOFFSET, 3130 ("vfs_busy_pages: no buffer offset")); 3131 vfs_setdirty(bp); 3132 3133retry: 3134 for (i = 0; i < bp->b_npages; i++) { 3135 vm_page_t m = bp->b_pages[i]; 3136 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 3137 goto retry; 3138 } 3139 3140 bogus = 0; 3141 for (i = 0; i < bp->b_npages; i++) { 3142 vm_page_t m = bp->b_pages[i]; 3143 3144 vm_page_flag_clear(m, PG_ZERO); 3145 if ((bp->b_flags & B_CLUSTER) == 0) { 3146 vm_object_pip_add(obj, 1); 3147 vm_page_io_start(m); 3148 } 3149 3150 /* 3151 * When readying a buffer for a read ( i.e 3152 * clear_modify == 0 ), it is important to do 3153 * bogus_page replacement for valid pages in 3154 * partially instantiated buffers. Partially 3155 * instantiated buffers can, in turn, occur when 3156 * reconstituting a buffer from its VM backing store 3157 * base. We only have to do this if B_CACHE is 3158 * clear ( which causes the I/O to occur in the 3159 * first place ). The replacement prevents the read 3160 * I/O from overwriting potentially dirty VM-backed 3161 * pages. XXX bogus page replacement is, uh, bogus. 3162 * It may not work properly with small-block devices. 3163 * We need to find a better way. 3164 */ 3165 3166 vm_page_protect(m, VM_PROT_NONE); 3167 if (clear_modify) 3168 vfs_page_set_valid(bp, foff, i, m); 3169 else if (m->valid == VM_PAGE_BITS_ALL && 3170 (bp->b_flags & B_CACHE) == 0) { 3171 bp->b_pages[i] = bogus_page; 3172 bogus++; 3173 } 3174 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3175 } 3176 if (bogus) 3177 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3178 } 3179} 3180 3181/* 3182 * Tell the VM system that the pages associated with this buffer 3183 * are clean. This is used for delayed writes where the data is 3184 * going to go to disk eventually without additional VM intevention. 3185 * 3186 * Note that while we only really need to clean through to b_bcount, we 3187 * just go ahead and clean through to b_bufsize. 3188 */ 3189static void 3190vfs_clean_pages(struct buf * bp) 3191{ 3192 int i; 3193 3194 GIANT_REQUIRED; 3195 3196 if (bp->b_flags & B_VMIO) { 3197 vm_ooffset_t foff; 3198 3199 foff = bp->b_offset; 3200 KASSERT(bp->b_offset != NOOFFSET, 3201 ("vfs_clean_pages: no buffer offset")); 3202 for (i = 0; i < bp->b_npages; i++) { 3203 vm_page_t m = bp->b_pages[i]; 3204 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3205 vm_ooffset_t eoff = noff; 3206 3207 if (eoff > bp->b_offset + bp->b_bufsize) 3208 eoff = bp->b_offset + bp->b_bufsize; 3209 vfs_page_set_valid(bp, foff, i, m); 3210 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3211 foff = noff; 3212 } 3213 } 3214} 3215 3216/* 3217 * vfs_bio_set_validclean: 3218 * 3219 * Set the range within the buffer to valid and clean. The range is 3220 * relative to the beginning of the buffer, b_offset. Note that b_offset 3221 * itself may be offset from the beginning of the first page. 3222 * 3223 */ 3224 3225void 3226vfs_bio_set_validclean(struct buf *bp, int base, int size) 3227{ 3228 if (bp->b_flags & B_VMIO) { 3229 int i; 3230 int n; 3231 3232 /* 3233 * Fixup base to be relative to beginning of first page. 3234 * Set initial n to be the maximum number of bytes in the 3235 * first page that can be validated. 3236 */ 3237 3238 base += (bp->b_offset & PAGE_MASK); 3239 n = PAGE_SIZE - (base & PAGE_MASK); 3240 3241 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3242 vm_page_t m = bp->b_pages[i]; 3243 3244 if (n > size) 3245 n = size; 3246 3247 vm_page_set_validclean(m, base & PAGE_MASK, n); 3248 base += n; 3249 size -= n; 3250 n = PAGE_SIZE; 3251 } 3252 } 3253} 3254 3255/* 3256 * vfs_bio_clrbuf: 3257 * 3258 * clear a buffer. This routine essentially fakes an I/O, so we need 3259 * to clear BIO_ERROR and B_INVAL. 3260 * 3261 * Note that while we only theoretically need to clear through b_bcount, 3262 * we go ahead and clear through b_bufsize. 3263 */ 3264 3265void 3266vfs_bio_clrbuf(struct buf *bp) 3267{ 3268 int i, mask = 0; 3269 caddr_t sa, ea; 3270 3271 GIANT_REQUIRED; 3272 3273 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3274 bp->b_flags &= ~B_INVAL; 3275 bp->b_ioflags &= ~BIO_ERROR; 3276 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3277 (bp->b_offset & PAGE_MASK) == 0) { 3278 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3279 if ((bp->b_pages[0]->valid & mask) == mask) { 3280 bp->b_resid = 0; 3281 return; 3282 } 3283 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3284 ((bp->b_pages[0]->valid & mask) == 0)) { 3285 bzero(bp->b_data, bp->b_bufsize); 3286 bp->b_pages[0]->valid |= mask; 3287 bp->b_resid = 0; 3288 return; 3289 } 3290 } 3291 ea = sa = bp->b_data; 3292 for(i=0;i<bp->b_npages;i++,sa=ea) { 3293 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3294 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3295 ea = (caddr_t)(vm_offset_t)ulmin( 3296 (u_long)(vm_offset_t)ea, 3297 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3298 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3299 if ((bp->b_pages[i]->valid & mask) == mask) 3300 continue; 3301 if ((bp->b_pages[i]->valid & mask) == 0) { 3302 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3303 bzero(sa, ea - sa); 3304 } 3305 } else { 3306 for (; sa < ea; sa += DEV_BSIZE, j++) { 3307 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3308 (bp->b_pages[i]->valid & (1<<j)) == 0) 3309 bzero(sa, DEV_BSIZE); 3310 } 3311 } 3312 bp->b_pages[i]->valid |= mask; 3313 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3314 } 3315 bp->b_resid = 0; 3316 } else { 3317 clrbuf(bp); 3318 } 3319} 3320 3321/* 3322 * vm_hold_load_pages and vm_hold_free_pages get pages into 3323 * a buffers address space. The pages are anonymous and are 3324 * not associated with a file object. 3325 */ 3326static void 3327vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3328{ 3329 vm_offset_t pg; 3330 vm_page_t p; 3331 int index; 3332 3333 GIANT_REQUIRED; 3334 3335 to = round_page(to); 3336 from = round_page(from); 3337 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3338 3339 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3340tryagain: 3341 /* 3342 * note: must allocate system pages since blocking here 3343 * could intefere with paging I/O, no matter which 3344 * process we are. 3345 */ 3346 p = vm_page_alloc(kernel_object, 3347 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3348 VM_ALLOC_SYSTEM); 3349 if (!p) { 3350 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3351 VM_WAIT; 3352 goto tryagain; 3353 } 3354 vm_page_wire(p); 3355 p->valid = VM_PAGE_BITS_ALL; 3356 vm_page_flag_clear(p, PG_ZERO); 3357 pmap_qenter(pg, &p, 1); 3358 bp->b_pages[index] = p; 3359 vm_page_wakeup(p); 3360 } 3361 bp->b_npages = index; 3362} 3363 3364/* Return pages associated with this buf to the vm system */ 3365void 3366vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3367{ 3368 vm_offset_t pg; 3369 vm_page_t p; 3370 int index, newnpages; 3371 3372 GIANT_REQUIRED; 3373 3374 from = round_page(from); 3375 to = round_page(to); 3376 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3377 3378 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3379 p = bp->b_pages[index]; 3380 if (p && (index < bp->b_npages)) { 3381 if (p->busy) { 3382 printf( 3383 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3384 (intmax_t)bp->b_blkno, 3385 (intmax_t)bp->b_lblkno); 3386 } 3387 bp->b_pages[index] = NULL; 3388 pmap_qremove(pg, 1); 3389 vm_page_busy(p); 3390 vm_page_unwire(p, 0); 3391 vm_page_free(p); 3392 } 3393 } 3394 bp->b_npages = newnpages; 3395} 3396 3397 3398#include "opt_ddb.h" 3399#ifdef DDB 3400#include <ddb/ddb.h> 3401 3402/* DDB command to show buffer data */ 3403DB_SHOW_COMMAND(buffer, db_show_buffer) 3404{ 3405 /* get args */ 3406 struct buf *bp = (struct buf *)addr; 3407 3408 if (!have_addr) { 3409 db_printf("usage: show buffer <addr>\n"); 3410 return; 3411 } 3412 3413 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3414 db_printf( 3415 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3416 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3417 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3418 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3419 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3420 if (bp->b_npages) { 3421 int i; 3422 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3423 for (i = 0; i < bp->b_npages; i++) { 3424 vm_page_t m; 3425 m = bp->b_pages[i]; 3426 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3427 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3428 if ((i + 1) < bp->b_npages) 3429 db_printf(","); 3430 } 3431 db_printf("\n"); 3432 } 3433} 3434#endif /* DDB */ 3435