vfs_bio.c revision 108651
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 108651 2003-01-04 09:37:42Z phk $ 15 */ 16 17/* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 * 27 * see man buf(9) for more info. 28 */ 29 30#include <sys/param.h> 31#include <sys/systm.h> 32#include <sys/stdint.h> 33#include <sys/bio.h> 34#include <sys/buf.h> 35#include <sys/devicestat.h> 36#include <sys/eventhandler.h> 37#include <sys/lock.h> 38#include <sys/malloc.h> 39#include <sys/mount.h> 40#include <sys/mutex.h> 41#include <sys/kernel.h> 42#include <sys/kthread.h> 43#include <sys/proc.h> 44#include <sys/resourcevar.h> 45#include <sys/sysctl.h> 46#include <sys/vmmeter.h> 47#include <sys/vnode.h> 48#include <vm/vm.h> 49#include <vm/vm_param.h> 50#include <vm/vm_kern.h> 51#include <vm/vm_pageout.h> 52#include <vm/vm_page.h> 53#include <vm/vm_object.h> 54#include <vm/vm_extern.h> 55#include <vm/vm_map.h> 56 57static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 58 59struct bio_ops bioops; /* I/O operation notification */ 60 61struct buf_ops buf_ops_bio = { 62 "buf_ops_bio", 63 bwrite 64}; 65 66/* 67 * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 68 * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 69 */ 70struct buf *buf; /* buffer header pool */ 71struct mtx buftimelock; /* Interlock on setting prio and timo */ 72 73static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 74 vm_offset_t to); 75static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 76 vm_offset_t to); 77static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 78 int pageno, vm_page_t m); 79static void vfs_clean_pages(struct buf * bp); 80static void vfs_setdirty(struct buf *bp); 81static void vfs_vmio_release(struct buf *bp); 82static void vfs_backgroundwritedone(struct buf *bp); 83static int flushbufqueues(void); 84static void buf_daemon(void); 85 86int vmiodirenable = TRUE; 87SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 88 "Use the VM system for directory writes"); 89int runningbufspace; 90SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 91 "Amount of presently outstanding async buffer io"); 92static int bufspace; 93SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 94 "KVA memory used for bufs"); 95static int maxbufspace; 96SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 97 "Maximum allowed value of bufspace (including buf_daemon)"); 98static int bufmallocspace; 99SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 100 "Amount of malloced memory for buffers"); 101static int maxbufmallocspace; 102SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 103 "Maximum amount of malloced memory for buffers"); 104static int lobufspace; 105SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 106 "Minimum amount of buffers we want to have"); 107static int hibufspace; 108SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 109 "Maximum allowed value of bufspace (excluding buf_daemon)"); 110static int bufreusecnt; 111SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 112 "Number of times we have reused a buffer"); 113static int buffreekvacnt; 114SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 115 "Number of times we have freed the KVA space from some buffer"); 116static int bufdefragcnt; 117SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 118 "Number of times we have had to repeat buffer allocation to defragment"); 119static int lorunningspace; 120SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 121 "Minimum preferred space used for in-progress I/O"); 122static int hirunningspace; 123SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 124 "Maximum amount of space to use for in-progress I/O"); 125static int numdirtybuffers; 126SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 127 "Number of buffers that are dirty (has unwritten changes) at the moment"); 128static int lodirtybuffers; 129SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 130 "How many buffers we want to have free before bufdaemon can sleep"); 131static int hidirtybuffers; 132SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 133 "When the number of dirty buffers is considered severe"); 134static int numfreebuffers; 135SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 136 "Number of free buffers"); 137static int lofreebuffers; 138SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 139 "XXX Unused"); 140static int hifreebuffers; 141SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 142 "XXX Complicatedly unused"); 143static int getnewbufcalls; 144SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 145 "Number of calls to getnewbuf"); 146static int getnewbufrestarts; 147SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 148 "Number of times getnewbuf has had to restart a buffer aquisition"); 149static int dobkgrdwrite = 1; 150SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 151 "Do background writes (honoring the BX_BKGRDWRITE flag)?"); 152 153/* 154 * Wakeup point for bufdaemon, as well as indicator of whether it is already 155 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 156 * is idling. 157 */ 158static int bd_request; 159 160/* 161 * bogus page -- for I/O to/from partially complete buffers 162 * this is a temporary solution to the problem, but it is not 163 * really that bad. it would be better to split the buffer 164 * for input in the case of buffers partially already in memory, 165 * but the code is intricate enough already. 166 */ 167vm_page_t bogus_page; 168 169/* 170 * Offset for bogus_page. 171 * XXX bogus_offset should be local to bufinit 172 */ 173static vm_offset_t bogus_offset; 174 175/* 176 * Synchronization (sleep/wakeup) variable for active buffer space requests. 177 * Set when wait starts, cleared prior to wakeup(). 178 * Used in runningbufwakeup() and waitrunningbufspace(). 179 */ 180static int runningbufreq; 181 182/* 183 * Synchronization (sleep/wakeup) variable for buffer requests. 184 * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 185 * by and/or. 186 * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 187 * getnewbuf(), and getblk(). 188 */ 189static int needsbuffer; 190 191#ifdef USE_BUFHASH 192/* 193 * Mask for index into the buffer hash table, which needs to be power of 2 in 194 * size. Set in kern_vfs_bio_buffer_alloc. 195 */ 196static int bufhashmask; 197 198/* 199 * Hash table for all buffers, with a linked list hanging from each table 200 * entry. Set in kern_vfs_bio_buffer_alloc, initialized in buf_init. 201 */ 202static LIST_HEAD(bufhashhdr, buf) *bufhashtbl; 203 204/* 205 * Somewhere to store buffers when they are not in another list, to always 206 * have them in a list (and thus being able to use the same set of operations 207 * on them.) 208 */ 209static struct bufhashhdr invalhash; 210 211#endif 212 213/* 214 * Definitions for the buffer free lists. 215 */ 216#define BUFFER_QUEUES 6 /* number of free buffer queues */ 217 218#define QUEUE_NONE 0 /* on no queue */ 219#define QUEUE_LOCKED 1 /* locked buffers */ 220#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ 221#define QUEUE_DIRTY 3 /* B_DELWRI buffers */ 222#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 223#define QUEUE_EMPTY 5 /* empty buffer headers */ 224 225/* Queues for free buffers with various properties */ 226static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 227/* 228 * Single global constant for BUF_WMESG, to avoid getting multiple references. 229 * buf_wmesg is referred from macros. 230 */ 231const char *buf_wmesg = BUF_WMESG; 232 233#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 234#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 235#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 236#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 237 238#ifdef USE_BUFHASH 239/* 240 * Buffer hash table code. Note that the logical block scans linearly, which 241 * gives us some L1 cache locality. 242 */ 243 244static __inline 245struct bufhashhdr * 246bufhash(struct vnode *vnp, daddr_t bn) 247{ 248 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 249} 250 251#endif 252 253/* 254 * numdirtywakeup: 255 * 256 * If someone is blocked due to there being too many dirty buffers, 257 * and numdirtybuffers is now reasonable, wake them up. 258 */ 259 260static __inline void 261numdirtywakeup(int level) 262{ 263 if (numdirtybuffers <= level) { 264 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 265 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 266 wakeup(&needsbuffer); 267 } 268 } 269} 270 271/* 272 * bufspacewakeup: 273 * 274 * Called when buffer space is potentially available for recovery. 275 * getnewbuf() will block on this flag when it is unable to free 276 * sufficient buffer space. Buffer space becomes recoverable when 277 * bp's get placed back in the queues. 278 */ 279 280static __inline void 281bufspacewakeup(void) 282{ 283 /* 284 * If someone is waiting for BUF space, wake them up. Even 285 * though we haven't freed the kva space yet, the waiting 286 * process will be able to now. 287 */ 288 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 289 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 290 wakeup(&needsbuffer); 291 } 292} 293 294/* 295 * runningbufwakeup() - in-progress I/O accounting. 296 * 297 */ 298static __inline void 299runningbufwakeup(struct buf *bp) 300{ 301 if (bp->b_runningbufspace) { 302 runningbufspace -= bp->b_runningbufspace; 303 bp->b_runningbufspace = 0; 304 if (runningbufreq && runningbufspace <= lorunningspace) { 305 runningbufreq = 0; 306 wakeup(&runningbufreq); 307 } 308 } 309} 310 311/* 312 * bufcountwakeup: 313 * 314 * Called when a buffer has been added to one of the free queues to 315 * account for the buffer and to wakeup anyone waiting for free buffers. 316 * This typically occurs when large amounts of metadata are being handled 317 * by the buffer cache ( else buffer space runs out first, usually ). 318 */ 319 320static __inline void 321bufcountwakeup(void) 322{ 323 ++numfreebuffers; 324 if (needsbuffer) { 325 needsbuffer &= ~VFS_BIO_NEED_ANY; 326 if (numfreebuffers >= hifreebuffers) 327 needsbuffer &= ~VFS_BIO_NEED_FREE; 328 wakeup(&needsbuffer); 329 } 330} 331 332/* 333 * waitrunningbufspace() 334 * 335 * runningbufspace is a measure of the amount of I/O currently 336 * running. This routine is used in async-write situations to 337 * prevent creating huge backups of pending writes to a device. 338 * Only asynchronous writes are governed by this function. 339 * 340 * Reads will adjust runningbufspace, but will not block based on it. 341 * The read load has a side effect of reducing the allowed write load. 342 * 343 * This does NOT turn an async write into a sync write. It waits 344 * for earlier writes to complete and generally returns before the 345 * caller's write has reached the device. 346 */ 347static __inline void 348waitrunningbufspace(void) 349{ 350 /* 351 * XXX race against wakeup interrupt, currently 352 * protected by Giant. FIXME! 353 */ 354 while (runningbufspace > hirunningspace) { 355 ++runningbufreq; 356 tsleep(&runningbufreq, PVM, "wdrain", 0); 357 } 358} 359 360 361/* 362 * vfs_buf_test_cache: 363 * 364 * Called when a buffer is extended. This function clears the B_CACHE 365 * bit if the newly extended portion of the buffer does not contain 366 * valid data. 367 */ 368static __inline__ 369void 370vfs_buf_test_cache(struct buf *bp, 371 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 372 vm_page_t m) 373{ 374 GIANT_REQUIRED; 375 376 if (bp->b_flags & B_CACHE) { 377 int base = (foff + off) & PAGE_MASK; 378 if (vm_page_is_valid(m, base, size) == 0) 379 bp->b_flags &= ~B_CACHE; 380 } 381} 382 383/* Wake up the buffer deamon if necessary */ 384static __inline__ 385void 386bd_wakeup(int dirtybuflevel) 387{ 388 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 389 bd_request = 1; 390 wakeup(&bd_request); 391 } 392} 393 394/* 395 * bd_speedup - speedup the buffer cache flushing code 396 */ 397 398static __inline__ 399void 400bd_speedup(void) 401{ 402 bd_wakeup(1); 403} 404 405/* 406 * Calculating buffer cache scaling values and reserve space for buffer 407 * headers. This is called during low level kernel initialization and 408 * may be called more then once. We CANNOT write to the memory area 409 * being reserved at this time. 410 */ 411caddr_t 412kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 413{ 414 /* 415 * physmem_est is in pages. Convert it to kilobytes (assumes 416 * PAGE_SIZE is >= 1K) 417 */ 418 physmem_est = physmem_est * (PAGE_SIZE / 1024); 419 420 /* 421 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 422 * For the first 64MB of ram nominally allocate sufficient buffers to 423 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 424 * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 425 * the buffer cache we limit the eventual kva reservation to 426 * maxbcache bytes. 427 * 428 * factor represents the 1/4 x ram conversion. 429 */ 430 if (nbuf == 0) { 431 int factor = 4 * BKVASIZE / 1024; 432 433 nbuf = 50; 434 if (physmem_est > 4096) 435 nbuf += min((physmem_est - 4096) / factor, 436 65536 / factor); 437 if (physmem_est > 65536) 438 nbuf += (physmem_est - 65536) * 2 / (factor * 5); 439 440 if (maxbcache && nbuf > maxbcache / BKVASIZE) 441 nbuf = maxbcache / BKVASIZE; 442 } 443 444#if 0 445 /* 446 * Do not allow the buffer_map to be more then 1/2 the size of the 447 * kernel_map. 448 */ 449 if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 450 (BKVASIZE * 2)) { 451 nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 452 (BKVASIZE * 2); 453 printf("Warning: nbufs capped at %d\n", nbuf); 454 } 455#endif 456 457 /* 458 * swbufs are used as temporary holders for I/O, such as paging I/O. 459 * We have no less then 16 and no more then 256. 460 */ 461 nswbuf = max(min(nbuf/4, 256), 16); 462 463 /* 464 * Reserve space for the buffer cache buffers 465 */ 466 swbuf = (void *)v; 467 v = (caddr_t)(swbuf + nswbuf); 468 buf = (void *)v; 469 v = (caddr_t)(buf + nbuf); 470 471#ifdef USE_BUFHASH 472 /* 473 * Calculate the hash table size and reserve space 474 */ 475 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 476 ; 477 bufhashtbl = (void *)v; 478 v = (caddr_t)(bufhashtbl + bufhashmask); 479 --bufhashmask; 480#endif 481 return(v); 482} 483 484/* Initialize the buffer subsystem. Called before use of any buffers. */ 485void 486bufinit(void) 487{ 488 struct buf *bp; 489 int i; 490 491 GIANT_REQUIRED; 492 493#ifdef USE_BUFHASH 494 LIST_INIT(&invalhash); 495#endif 496 mtx_init(&buftimelock, "buftime lock", NULL, MTX_DEF); 497 498#ifdef USE_BUFHASH 499 for (i = 0; i <= bufhashmask; i++) 500 LIST_INIT(&bufhashtbl[i]); 501#endif 502 503 /* next, make a null set of free lists */ 504 for (i = 0; i < BUFFER_QUEUES; i++) 505 TAILQ_INIT(&bufqueues[i]); 506 507 /* finally, initialize each buffer header and stick on empty q */ 508 for (i = 0; i < nbuf; i++) { 509 bp = &buf[i]; 510 bzero(bp, sizeof *bp); 511 bp->b_flags = B_INVAL; /* we're just an empty header */ 512 bp->b_dev = NODEV; 513 bp->b_rcred = NOCRED; 514 bp->b_wcred = NOCRED; 515 bp->b_qindex = QUEUE_EMPTY; 516 bp->b_xflags = 0; 517 LIST_INIT(&bp->b_dep); 518 BUF_LOCKINIT(bp); 519 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 520#ifdef USE_BUFHASH 521 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 522#endif 523 } 524 525 /* 526 * maxbufspace is the absolute maximum amount of buffer space we are 527 * allowed to reserve in KVM and in real terms. The absolute maximum 528 * is nominally used by buf_daemon. hibufspace is the nominal maximum 529 * used by most other processes. The differential is required to 530 * ensure that buf_daemon is able to run when other processes might 531 * be blocked waiting for buffer space. 532 * 533 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 534 * this may result in KVM fragmentation which is not handled optimally 535 * by the system. 536 */ 537 maxbufspace = nbuf * BKVASIZE; 538 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 539 lobufspace = hibufspace - MAXBSIZE; 540 541 lorunningspace = 512 * 1024; 542 hirunningspace = 1024 * 1024; 543 544/* 545 * Limit the amount of malloc memory since it is wired permanently into 546 * the kernel space. Even though this is accounted for in the buffer 547 * allocation, we don't want the malloced region to grow uncontrolled. 548 * The malloc scheme improves memory utilization significantly on average 549 * (small) directories. 550 */ 551 maxbufmallocspace = hibufspace / 20; 552 553/* 554 * Reduce the chance of a deadlock occuring by limiting the number 555 * of delayed-write dirty buffers we allow to stack up. 556 */ 557 hidirtybuffers = nbuf / 4 + 20; 558 numdirtybuffers = 0; 559/* 560 * To support extreme low-memory systems, make sure hidirtybuffers cannot 561 * eat up all available buffer space. This occurs when our minimum cannot 562 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 563 * BKVASIZE'd (8K) buffers. 564 */ 565 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 566 hidirtybuffers >>= 1; 567 } 568 lodirtybuffers = hidirtybuffers / 2; 569 570/* 571 * Try to keep the number of free buffers in the specified range, 572 * and give special processes (e.g. like buf_daemon) access to an 573 * emergency reserve. 574 */ 575 lofreebuffers = nbuf / 18 + 5; 576 hifreebuffers = 2 * lofreebuffers; 577 numfreebuffers = nbuf; 578 579/* 580 * Maximum number of async ops initiated per buf_daemon loop. This is 581 * somewhat of a hack at the moment, we really need to limit ourselves 582 * based on the number of bytes of I/O in-transit that were initiated 583 * from buf_daemon. 584 */ 585 586 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 587 vm_object_lock(kernel_object); 588 bogus_page = vm_page_alloc(kernel_object, 589 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 590 VM_ALLOC_NORMAL); 591 vm_object_unlock(kernel_object); 592 cnt.v_wire_count++; 593} 594 595/* 596 * bfreekva() - free the kva allocation for a buffer. 597 * 598 * Must be called at splbio() or higher as this is the only locking for 599 * buffer_map. 600 * 601 * Since this call frees up buffer space, we call bufspacewakeup(). 602 */ 603static void 604bfreekva(struct buf * bp) 605{ 606 GIANT_REQUIRED; 607 608 if (bp->b_kvasize) { 609 ++buffreekvacnt; 610 bufspace -= bp->b_kvasize; 611 vm_map_delete(buffer_map, 612 (vm_offset_t) bp->b_kvabase, 613 (vm_offset_t) bp->b_kvabase + bp->b_kvasize 614 ); 615 bp->b_kvasize = 0; 616 bufspacewakeup(); 617 } 618} 619 620/* 621 * bremfree: 622 * 623 * Remove the buffer from the appropriate free list. 624 */ 625void 626bremfree(struct buf * bp) 627{ 628 int s = splbio(); 629 int old_qindex = bp->b_qindex; 630 631 GIANT_REQUIRED; 632 633 if (bp->b_qindex != QUEUE_NONE) { 634 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 635 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 636 bp->b_qindex = QUEUE_NONE; 637 } else { 638 if (BUF_REFCNT(bp) <= 1) 639 panic("bremfree: removing a buffer not on a queue"); 640 } 641 642 /* 643 * Fixup numfreebuffers count. If the buffer is invalid or not 644 * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 645 * the buffer was free and we must decrement numfreebuffers. 646 */ 647 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 648 switch(old_qindex) { 649 case QUEUE_DIRTY: 650 case QUEUE_CLEAN: 651 case QUEUE_EMPTY: 652 case QUEUE_EMPTYKVA: 653 --numfreebuffers; 654 break; 655 default: 656 break; 657 } 658 } 659 splx(s); 660} 661 662 663/* 664 * Get a buffer with the specified data. Look in the cache first. We 665 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 666 * is set, the buffer is valid and we do not have to do anything ( see 667 * getblk() ). This is really just a special case of breadn(). 668 */ 669int 670bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 671 struct buf ** bpp) 672{ 673 674 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 675} 676 677/* 678 * Operates like bread, but also starts asynchronous I/O on 679 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 680 * to initiating I/O . If B_CACHE is set, the buffer is valid 681 * and we do not have to do anything. 682 */ 683int 684breadn(struct vnode * vp, daddr_t blkno, int size, 685 daddr_t * rablkno, int *rabsize, 686 int cnt, struct ucred * cred, struct buf ** bpp) 687{ 688 struct buf *bp, *rabp; 689 int i; 690 int rv = 0, readwait = 0; 691 692 *bpp = bp = getblk(vp, blkno, size, 0, 0); 693 694 /* if not found in cache, do some I/O */ 695 if ((bp->b_flags & B_CACHE) == 0) { 696 if (curthread != PCPU_GET(idlethread)) 697 curthread->td_proc->p_stats->p_ru.ru_inblock++; 698 bp->b_iocmd = BIO_READ; 699 bp->b_flags &= ~B_INVAL; 700 bp->b_ioflags &= ~BIO_ERROR; 701 if (bp->b_rcred == NOCRED && cred != NOCRED) 702 bp->b_rcred = crhold(cred); 703 vfs_busy_pages(bp, 0); 704 VOP_STRATEGY(vp, bp); 705 ++readwait; 706 } 707 708 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 709 if (inmem(vp, *rablkno)) 710 continue; 711 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 712 713 if ((rabp->b_flags & B_CACHE) == 0) { 714 if (curthread != PCPU_GET(idlethread)) 715 curthread->td_proc->p_stats->p_ru.ru_inblock++; 716 rabp->b_flags |= B_ASYNC; 717 rabp->b_flags &= ~B_INVAL; 718 rabp->b_ioflags &= ~BIO_ERROR; 719 rabp->b_iocmd = BIO_READ; 720 if (rabp->b_rcred == NOCRED && cred != NOCRED) 721 rabp->b_rcred = crhold(cred); 722 vfs_busy_pages(rabp, 0); 723 BUF_KERNPROC(rabp); 724 VOP_STRATEGY(vp, rabp); 725 } else { 726 brelse(rabp); 727 } 728 } 729 730 if (readwait) { 731 rv = bufwait(bp); 732 } 733 return (rv); 734} 735 736/* 737 * Write, release buffer on completion. (Done by iodone 738 * if async). Do not bother writing anything if the buffer 739 * is invalid. 740 * 741 * Note that we set B_CACHE here, indicating that buffer is 742 * fully valid and thus cacheable. This is true even of NFS 743 * now so we set it generally. This could be set either here 744 * or in biodone() since the I/O is synchronous. We put it 745 * here. 746 */ 747 748int 749bwrite(struct buf * bp) 750{ 751 int oldflags, s; 752 struct buf *newbp; 753 754 if (bp->b_flags & B_INVAL) { 755 brelse(bp); 756 return (0); 757 } 758 759 oldflags = bp->b_flags; 760 761 if (BUF_REFCNT(bp) == 0) 762 panic("bwrite: buffer is not busy???"); 763 s = splbio(); 764 /* 765 * If a background write is already in progress, delay 766 * writing this block if it is asynchronous. Otherwise 767 * wait for the background write to complete. 768 */ 769 if (bp->b_xflags & BX_BKGRDINPROG) { 770 if (bp->b_flags & B_ASYNC) { 771 splx(s); 772 bdwrite(bp); 773 return (0); 774 } 775 bp->b_xflags |= BX_BKGRDWAIT; 776 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); 777 if (bp->b_xflags & BX_BKGRDINPROG) 778 panic("bwrite: still writing"); 779 } 780 781 /* Mark the buffer clean */ 782 bundirty(bp); 783 784 /* 785 * If this buffer is marked for background writing and we 786 * do not have to wait for it, make a copy and write the 787 * copy so as to leave this buffer ready for further use. 788 * 789 * This optimization eats a lot of memory. If we have a page 790 * or buffer shortfall we can't do it. 791 */ 792 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 793 (bp->b_flags & B_ASYNC) && 794 !vm_page_count_severe() && 795 !buf_dirty_count_severe()) { 796 if (bp->b_iodone != NULL) { 797 printf("bp->b_iodone = %p\n", bp->b_iodone); 798 panic("bwrite: need chained iodone"); 799 } 800 801 /* get a new block */ 802 newbp = geteblk(bp->b_bufsize); 803 804 /* 805 * set it to be identical to the old block. We have to 806 * set b_lblkno and BKGRDMARKER before calling bgetvp() 807 * to avoid confusing the splay tree and gbincore(). 808 */ 809 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 810 newbp->b_lblkno = bp->b_lblkno; 811 newbp->b_xflags |= BX_BKGRDMARKER; 812 bgetvp(bp->b_vp, newbp); 813 newbp->b_blkno = bp->b_blkno; 814 newbp->b_offset = bp->b_offset; 815 newbp->b_iodone = vfs_backgroundwritedone; 816 newbp->b_flags |= B_ASYNC; 817 newbp->b_flags &= ~B_INVAL; 818 819 /* move over the dependencies */ 820 if (LIST_FIRST(&bp->b_dep) != NULL) 821 buf_movedeps(bp, newbp); 822 823 /* 824 * Initiate write on the copy, release the original to 825 * the B_LOCKED queue so that it cannot go away until 826 * the background write completes. If not locked it could go 827 * away and then be reconstituted while it was being written. 828 * If the reconstituted buffer were written, we could end up 829 * with two background copies being written at the same time. 830 */ 831 bp->b_xflags |= BX_BKGRDINPROG; 832 bp->b_flags |= B_LOCKED; 833 bqrelse(bp); 834 bp = newbp; 835 } 836 837 bp->b_flags &= ~B_DONE; 838 bp->b_ioflags &= ~BIO_ERROR; 839 bp->b_flags |= B_WRITEINPROG | B_CACHE; 840 bp->b_iocmd = BIO_WRITE; 841 842 VI_LOCK(bp->b_vp); 843 bp->b_vp->v_numoutput++; 844 VI_UNLOCK(bp->b_vp); 845 vfs_busy_pages(bp, 1); 846 847 /* 848 * Normal bwrites pipeline writes 849 */ 850 bp->b_runningbufspace = bp->b_bufsize; 851 runningbufspace += bp->b_runningbufspace; 852 853 if (curthread != PCPU_GET(idlethread)) 854 curthread->td_proc->p_stats->p_ru.ru_oublock++; 855 splx(s); 856 if (oldflags & B_ASYNC) 857 BUF_KERNPROC(bp); 858 VOP_STRATEGY(bp->b_vp, bp); 859 860 if ((oldflags & B_ASYNC) == 0) { 861 int rtval = bufwait(bp); 862 brelse(bp); 863 return (rtval); 864 } else if ((oldflags & B_NOWDRAIN) == 0) { 865 /* 866 * don't allow the async write to saturate the I/O 867 * system. Deadlocks can occur only if a device strategy 868 * routine (like in MD) turns around and issues another 869 * high-level write, in which case B_NOWDRAIN is expected 870 * to be set. Otherwise we will not deadlock here because 871 * we are blocking waiting for I/O that is already in-progress 872 * to complete. 873 */ 874 waitrunningbufspace(); 875 } 876 877 return (0); 878} 879 880/* 881 * Complete a background write started from bwrite. 882 */ 883static void 884vfs_backgroundwritedone(bp) 885 struct buf *bp; 886{ 887 struct buf *origbp; 888 889 /* 890 * Find the original buffer that we are writing. 891 */ 892 VI_LOCK(bp->b_vp); 893 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 894 panic("backgroundwritedone: lost buffer"); 895 VI_UNLOCK(bp->b_vp); 896 /* 897 * Process dependencies then return any unfinished ones. 898 */ 899 if (LIST_FIRST(&bp->b_dep) != NULL) 900 buf_complete(bp); 901 if (LIST_FIRST(&bp->b_dep) != NULL) 902 buf_movedeps(bp, origbp); 903 /* 904 * Clear the BX_BKGRDINPROG flag in the original buffer 905 * and awaken it if it is waiting for the write to complete. 906 * If BX_BKGRDINPROG is not set in the original buffer it must 907 * have been released and re-instantiated - which is not legal. 908 */ 909 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 910 ("backgroundwritedone: lost buffer2")); 911 origbp->b_xflags &= ~BX_BKGRDINPROG; 912 if (origbp->b_xflags & BX_BKGRDWAIT) { 913 origbp->b_xflags &= ~BX_BKGRDWAIT; 914 wakeup(&origbp->b_xflags); 915 } 916 /* 917 * Clear the B_LOCKED flag and remove it from the locked 918 * queue if it currently resides there. 919 */ 920 origbp->b_flags &= ~B_LOCKED; 921 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 922 bremfree(origbp); 923 bqrelse(origbp); 924 } 925 /* 926 * This buffer is marked B_NOCACHE, so when it is released 927 * by biodone, it will be tossed. We mark it with BIO_READ 928 * to avoid biodone doing a second vwakeup. 929 */ 930 bp->b_flags |= B_NOCACHE; 931 bp->b_iocmd = BIO_READ; 932 bp->b_flags &= ~(B_CACHE | B_DONE); 933 bp->b_iodone = 0; 934 bufdone(bp); 935} 936 937/* 938 * Delayed write. (Buffer is marked dirty). Do not bother writing 939 * anything if the buffer is marked invalid. 940 * 941 * Note that since the buffer must be completely valid, we can safely 942 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 943 * biodone() in order to prevent getblk from writing the buffer 944 * out synchronously. 945 */ 946void 947bdwrite(struct buf * bp) 948{ 949 GIANT_REQUIRED; 950 951 if (BUF_REFCNT(bp) == 0) 952 panic("bdwrite: buffer is not busy"); 953 954 if (bp->b_flags & B_INVAL) { 955 brelse(bp); 956 return; 957 } 958 bdirty(bp); 959 960 /* 961 * Set B_CACHE, indicating that the buffer is fully valid. This is 962 * true even of NFS now. 963 */ 964 bp->b_flags |= B_CACHE; 965 966 /* 967 * This bmap keeps the system from needing to do the bmap later, 968 * perhaps when the system is attempting to do a sync. Since it 969 * is likely that the indirect block -- or whatever other datastructure 970 * that the filesystem needs is still in memory now, it is a good 971 * thing to do this. Note also, that if the pageout daemon is 972 * requesting a sync -- there might not be enough memory to do 973 * the bmap then... So, this is important to do. 974 */ 975 if (bp->b_vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 976 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 977 } 978 979 /* 980 * Set the *dirty* buffer range based upon the VM system dirty pages. 981 */ 982 vfs_setdirty(bp); 983 984 /* 985 * We need to do this here to satisfy the vnode_pager and the 986 * pageout daemon, so that it thinks that the pages have been 987 * "cleaned". Note that since the pages are in a delayed write 988 * buffer -- the VFS layer "will" see that the pages get written 989 * out on the next sync, or perhaps the cluster will be completed. 990 */ 991 vfs_clean_pages(bp); 992 bqrelse(bp); 993 994 /* 995 * Wakeup the buffer flushing daemon if we have a lot of dirty 996 * buffers (midpoint between our recovery point and our stall 997 * point). 998 */ 999 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1000 1001 /* 1002 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1003 * due to the softdep code. 1004 */ 1005} 1006 1007/* 1008 * bdirty: 1009 * 1010 * Turn buffer into delayed write request. We must clear BIO_READ and 1011 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1012 * itself to properly update it in the dirty/clean lists. We mark it 1013 * B_DONE to ensure that any asynchronization of the buffer properly 1014 * clears B_DONE ( else a panic will occur later ). 1015 * 1016 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1017 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1018 * should only be called if the buffer is known-good. 1019 * 1020 * Since the buffer is not on a queue, we do not update the numfreebuffers 1021 * count. 1022 * 1023 * Must be called at splbio(). 1024 * The buffer must be on QUEUE_NONE. 1025 */ 1026void 1027bdirty(bp) 1028 struct buf *bp; 1029{ 1030 KASSERT(bp->b_qindex == QUEUE_NONE, 1031 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1032 bp->b_flags &= ~(B_RELBUF); 1033 bp->b_iocmd = BIO_WRITE; 1034 1035 if ((bp->b_flags & B_DELWRI) == 0) { 1036 bp->b_flags |= B_DONE | B_DELWRI; 1037 reassignbuf(bp, bp->b_vp); 1038 ++numdirtybuffers; 1039 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1040 } 1041} 1042 1043/* 1044 * bundirty: 1045 * 1046 * Clear B_DELWRI for buffer. 1047 * 1048 * Since the buffer is not on a queue, we do not update the numfreebuffers 1049 * count. 1050 * 1051 * Must be called at splbio(). 1052 * The buffer must be on QUEUE_NONE. 1053 */ 1054 1055void 1056bundirty(bp) 1057 struct buf *bp; 1058{ 1059 KASSERT(bp->b_qindex == QUEUE_NONE, 1060 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1061 1062 if (bp->b_flags & B_DELWRI) { 1063 bp->b_flags &= ~B_DELWRI; 1064 reassignbuf(bp, bp->b_vp); 1065 --numdirtybuffers; 1066 numdirtywakeup(lodirtybuffers); 1067 } 1068 /* 1069 * Since it is now being written, we can clear its deferred write flag. 1070 */ 1071 bp->b_flags &= ~B_DEFERRED; 1072} 1073 1074/* 1075 * bawrite: 1076 * 1077 * Asynchronous write. Start output on a buffer, but do not wait for 1078 * it to complete. The buffer is released when the output completes. 1079 * 1080 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1081 * B_INVAL buffers. Not us. 1082 */ 1083void 1084bawrite(struct buf * bp) 1085{ 1086 bp->b_flags |= B_ASYNC; 1087 (void) BUF_WRITE(bp); 1088} 1089 1090/* 1091 * bwillwrite: 1092 * 1093 * Called prior to the locking of any vnodes when we are expecting to 1094 * write. We do not want to starve the buffer cache with too many 1095 * dirty buffers so we block here. By blocking prior to the locking 1096 * of any vnodes we attempt to avoid the situation where a locked vnode 1097 * prevents the various system daemons from flushing related buffers. 1098 */ 1099 1100void 1101bwillwrite(void) 1102{ 1103 if (numdirtybuffers >= hidirtybuffers) { 1104 int s; 1105 1106 mtx_lock(&Giant); 1107 s = splbio(); 1108 while (numdirtybuffers >= hidirtybuffers) { 1109 bd_wakeup(1); 1110 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1111 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 1112 } 1113 splx(s); 1114 mtx_unlock(&Giant); 1115 } 1116} 1117 1118/* 1119 * Return true if we have too many dirty buffers. 1120 */ 1121int 1122buf_dirty_count_severe(void) 1123{ 1124 return(numdirtybuffers >= hidirtybuffers); 1125} 1126 1127/* 1128 * brelse: 1129 * 1130 * Release a busy buffer and, if requested, free its resources. The 1131 * buffer will be stashed in the appropriate bufqueue[] allowing it 1132 * to be accessed later as a cache entity or reused for other purposes. 1133 */ 1134void 1135brelse(struct buf * bp) 1136{ 1137 int s; 1138 1139 GIANT_REQUIRED; 1140 1141 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1142 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1143 1144 s = splbio(); 1145 1146 if (bp->b_flags & B_LOCKED) 1147 bp->b_ioflags &= ~BIO_ERROR; 1148 1149 if (bp->b_iocmd == BIO_WRITE && 1150 (bp->b_ioflags & BIO_ERROR) && 1151 !(bp->b_flags & B_INVAL)) { 1152 /* 1153 * Failed write, redirty. Must clear BIO_ERROR to prevent 1154 * pages from being scrapped. If B_INVAL is set then 1155 * this case is not run and the next case is run to 1156 * destroy the buffer. B_INVAL can occur if the buffer 1157 * is outside the range supported by the underlying device. 1158 */ 1159 bp->b_ioflags &= ~BIO_ERROR; 1160 bdirty(bp); 1161 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1162 (bp->b_ioflags & BIO_ERROR) || 1163 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1164 /* 1165 * Either a failed I/O or we were asked to free or not 1166 * cache the buffer. 1167 */ 1168 bp->b_flags |= B_INVAL; 1169 if (LIST_FIRST(&bp->b_dep) != NULL) 1170 buf_deallocate(bp); 1171 if (bp->b_flags & B_DELWRI) { 1172 --numdirtybuffers; 1173 numdirtywakeup(lodirtybuffers); 1174 } 1175 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1176 if ((bp->b_flags & B_VMIO) == 0) { 1177 if (bp->b_bufsize) 1178 allocbuf(bp, 0); 1179 if (bp->b_vp) 1180 brelvp(bp); 1181 } 1182 } 1183 1184 /* 1185 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1186 * is called with B_DELWRI set, the underlying pages may wind up 1187 * getting freed causing a previous write (bdwrite()) to get 'lost' 1188 * because pages associated with a B_DELWRI bp are marked clean. 1189 * 1190 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1191 * if B_DELWRI is set. 1192 * 1193 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1194 * on pages to return pages to the VM page queues. 1195 */ 1196 if (bp->b_flags & B_DELWRI) 1197 bp->b_flags &= ~B_RELBUF; 1198 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1199 bp->b_flags |= B_RELBUF; 1200 1201 /* 1202 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1203 * constituted, not even NFS buffers now. Two flags effect this. If 1204 * B_INVAL, the struct buf is invalidated but the VM object is kept 1205 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1206 * 1207 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1208 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1209 * buffer is also B_INVAL because it hits the re-dirtying code above. 1210 * 1211 * Normally we can do this whether a buffer is B_DELWRI or not. If 1212 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1213 * the commit state and we cannot afford to lose the buffer. If the 1214 * buffer has a background write in progress, we need to keep it 1215 * around to prevent it from being reconstituted and starting a second 1216 * background write. 1217 */ 1218 if ((bp->b_flags & B_VMIO) 1219 && !(bp->b_vp->v_mount != NULL && 1220 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1221 !vn_isdisk(bp->b_vp, NULL) && 1222 (bp->b_flags & B_DELWRI)) 1223 ) { 1224 1225 int i, j, resid; 1226 vm_page_t m; 1227 off_t foff; 1228 vm_pindex_t poff; 1229 vm_object_t obj; 1230 struct vnode *vp; 1231 1232 vp = bp->b_vp; 1233 obj = bp->b_object; 1234 1235 /* 1236 * Get the base offset and length of the buffer. Note that 1237 * in the VMIO case if the buffer block size is not 1238 * page-aligned then b_data pointer may not be page-aligned. 1239 * But our b_pages[] array *IS* page aligned. 1240 * 1241 * block sizes less then DEV_BSIZE (usually 512) are not 1242 * supported due to the page granularity bits (m->valid, 1243 * m->dirty, etc...). 1244 * 1245 * See man buf(9) for more information 1246 */ 1247 resid = bp->b_bufsize; 1248 foff = bp->b_offset; 1249 1250 for (i = 0; i < bp->b_npages; i++) { 1251 int had_bogus = 0; 1252 1253 m = bp->b_pages[i]; 1254 vm_page_lock_queues(); 1255 vm_page_flag_clear(m, PG_ZERO); 1256 vm_page_unlock_queues(); 1257 1258 /* 1259 * If we hit a bogus page, fixup *all* the bogus pages 1260 * now. 1261 */ 1262 if (m == bogus_page) { 1263 poff = OFF_TO_IDX(bp->b_offset); 1264 had_bogus = 1; 1265 1266 for (j = i; j < bp->b_npages; j++) { 1267 vm_page_t mtmp; 1268 mtmp = bp->b_pages[j]; 1269 if (mtmp == bogus_page) { 1270 mtmp = vm_page_lookup(obj, poff + j); 1271 if (!mtmp) { 1272 panic("brelse: page missing\n"); 1273 } 1274 bp->b_pages[j] = mtmp; 1275 } 1276 } 1277 1278 if ((bp->b_flags & B_INVAL) == 0) { 1279 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1280 } 1281 m = bp->b_pages[i]; 1282 } 1283 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1284 int poffset = foff & PAGE_MASK; 1285 int presid = resid > (PAGE_SIZE - poffset) ? 1286 (PAGE_SIZE - poffset) : resid; 1287 1288 KASSERT(presid >= 0, ("brelse: extra page")); 1289 vm_page_set_invalid(m, poffset, presid); 1290 if (had_bogus) 1291 printf("avoided corruption bug in bogus_page/brelse code\n"); 1292 } 1293 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1294 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1295 } 1296 1297 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1298 vfs_vmio_release(bp); 1299 1300 } else if (bp->b_flags & B_VMIO) { 1301 1302 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1303 vfs_vmio_release(bp); 1304 } 1305 1306 } 1307 1308 if (bp->b_qindex != QUEUE_NONE) 1309 panic("brelse: free buffer onto another queue???"); 1310 if (BUF_REFCNT(bp) > 1) { 1311 /* do not release to free list */ 1312 BUF_UNLOCK(bp); 1313 splx(s); 1314 return; 1315 } 1316 1317 /* enqueue */ 1318 1319 /* buffers with no memory */ 1320 if (bp->b_bufsize == 0) { 1321 bp->b_flags |= B_INVAL; 1322 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1323 if (bp->b_xflags & BX_BKGRDINPROG) 1324 panic("losing buffer 1"); 1325 if (bp->b_kvasize) { 1326 bp->b_qindex = QUEUE_EMPTYKVA; 1327 } else { 1328 bp->b_qindex = QUEUE_EMPTY; 1329 } 1330 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1331#ifdef USE_BUFHASH 1332 LIST_REMOVE(bp, b_hash); 1333 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1334#endif 1335 bp->b_dev = NODEV; 1336 /* buffers with junk contents */ 1337 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1338 (bp->b_ioflags & BIO_ERROR)) { 1339 bp->b_flags |= B_INVAL; 1340 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1341 if (bp->b_xflags & BX_BKGRDINPROG) 1342 panic("losing buffer 2"); 1343 bp->b_qindex = QUEUE_CLEAN; 1344 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1345#ifdef USE_BUFHASH 1346 LIST_REMOVE(bp, b_hash); 1347 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1348#endif 1349 bp->b_dev = NODEV; 1350 1351 /* buffers that are locked */ 1352 } else if (bp->b_flags & B_LOCKED) { 1353 bp->b_qindex = QUEUE_LOCKED; 1354 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1355 1356 /* remaining buffers */ 1357 } else { 1358 if (bp->b_flags & B_DELWRI) 1359 bp->b_qindex = QUEUE_DIRTY; 1360 else 1361 bp->b_qindex = QUEUE_CLEAN; 1362 if (bp->b_flags & B_AGE) 1363 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1364 else 1365 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1366 } 1367 1368 /* 1369 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1370 * placed the buffer on the correct queue. We must also disassociate 1371 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1372 * find it. 1373 */ 1374 if (bp->b_flags & B_INVAL) { 1375 if (bp->b_flags & B_DELWRI) 1376 bundirty(bp); 1377 if (bp->b_vp) 1378 brelvp(bp); 1379 } 1380 1381 /* 1382 * Fixup numfreebuffers count. The bp is on an appropriate queue 1383 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1384 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1385 * if B_INVAL is set ). 1386 */ 1387 1388 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1389 bufcountwakeup(); 1390 1391 /* 1392 * Something we can maybe free or reuse 1393 */ 1394 if (bp->b_bufsize || bp->b_kvasize) 1395 bufspacewakeup(); 1396 1397 /* unlock */ 1398 BUF_UNLOCK(bp); 1399 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1400 B_DIRECT | B_NOWDRAIN); 1401 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1402 panic("brelse: not dirty"); 1403 splx(s); 1404} 1405 1406/* 1407 * Release a buffer back to the appropriate queue but do not try to free 1408 * it. The buffer is expected to be used again soon. 1409 * 1410 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1411 * biodone() to requeue an async I/O on completion. It is also used when 1412 * known good buffers need to be requeued but we think we may need the data 1413 * again soon. 1414 * 1415 * XXX we should be able to leave the B_RELBUF hint set on completion. 1416 */ 1417void 1418bqrelse(struct buf * bp) 1419{ 1420 int s; 1421 1422 s = splbio(); 1423 1424 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1425 1426 if (bp->b_qindex != QUEUE_NONE) 1427 panic("bqrelse: free buffer onto another queue???"); 1428 if (BUF_REFCNT(bp) > 1) { 1429 /* do not release to free list */ 1430 BUF_UNLOCK(bp); 1431 splx(s); 1432 return; 1433 } 1434 if (bp->b_flags & B_LOCKED) { 1435 bp->b_ioflags &= ~BIO_ERROR; 1436 bp->b_qindex = QUEUE_LOCKED; 1437 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1438 /* buffers with stale but valid contents */ 1439 } else if (bp->b_flags & B_DELWRI) { 1440 bp->b_qindex = QUEUE_DIRTY; 1441 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1442 } else if (vm_page_count_severe()) { 1443 /* 1444 * We are too low on memory, we have to try to free the 1445 * buffer (most importantly: the wired pages making up its 1446 * backing store) *now*. 1447 */ 1448 splx(s); 1449 brelse(bp); 1450 return; 1451 } else { 1452 bp->b_qindex = QUEUE_CLEAN; 1453 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1454 } 1455 1456 if ((bp->b_flags & B_LOCKED) == 0 && 1457 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1458 bufcountwakeup(); 1459 } 1460 1461 /* 1462 * Something we can maybe free or reuse. 1463 */ 1464 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1465 bufspacewakeup(); 1466 1467 /* unlock */ 1468 BUF_UNLOCK(bp); 1469 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1470 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1471 panic("bqrelse: not dirty"); 1472 splx(s); 1473} 1474 1475/* Give pages used by the bp back to the VM system (where possible) */ 1476static void 1477vfs_vmio_release(bp) 1478 struct buf *bp; 1479{ 1480 int i; 1481 vm_page_t m; 1482 1483 GIANT_REQUIRED; 1484 vm_page_lock_queues(); 1485 for (i = 0; i < bp->b_npages; i++) { 1486 m = bp->b_pages[i]; 1487 bp->b_pages[i] = NULL; 1488 /* 1489 * In order to keep page LRU ordering consistent, put 1490 * everything on the inactive queue. 1491 */ 1492 vm_page_unwire(m, 0); 1493 /* 1494 * We don't mess with busy pages, it is 1495 * the responsibility of the process that 1496 * busied the pages to deal with them. 1497 */ 1498 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1499 continue; 1500 1501 if (m->wire_count == 0) { 1502 vm_page_flag_clear(m, PG_ZERO); 1503 /* 1504 * Might as well free the page if we can and it has 1505 * no valid data. We also free the page if the 1506 * buffer was used for direct I/O 1507 */ 1508 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1509 m->hold_count == 0) { 1510 vm_page_busy(m); 1511 pmap_remove_all(m); 1512 vm_page_free(m); 1513 } else if (bp->b_flags & B_DIRECT) { 1514 vm_page_try_to_free(m); 1515 } else if (vm_page_count_severe()) { 1516 vm_page_try_to_cache(m); 1517 } 1518 } 1519 } 1520 vm_page_unlock_queues(); 1521 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1522 1523 if (bp->b_bufsize) { 1524 bufspacewakeup(); 1525 bp->b_bufsize = 0; 1526 } 1527 bp->b_npages = 0; 1528 bp->b_flags &= ~B_VMIO; 1529 if (bp->b_vp) 1530 brelvp(bp); 1531} 1532 1533#ifdef USE_BUFHASH 1534/* 1535 * XXX MOVED TO VFS_SUBR.C 1536 * 1537 * Check to see if a block is currently memory resident. 1538 */ 1539struct buf * 1540gbincore(struct vnode * vp, daddr_t blkno) 1541{ 1542 struct buf *bp; 1543 struct bufhashhdr *bh; 1544 1545 bh = bufhash(vp, blkno); 1546 1547 /* Search hash chain */ 1548 LIST_FOREACH(bp, bh, b_hash) { 1549 /* hit */ 1550 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1551 (bp->b_flags & B_INVAL) == 0) { 1552 break; 1553 } 1554 } 1555 return (bp); 1556} 1557#endif 1558 1559/* 1560 * vfs_bio_awrite: 1561 * 1562 * Implement clustered async writes for clearing out B_DELWRI buffers. 1563 * This is much better then the old way of writing only one buffer at 1564 * a time. Note that we may not be presented with the buffers in the 1565 * correct order, so we search for the cluster in both directions. 1566 */ 1567int 1568vfs_bio_awrite(struct buf * bp) 1569{ 1570 int i; 1571 int j; 1572 daddr_t lblkno = bp->b_lblkno; 1573 struct vnode *vp = bp->b_vp; 1574 int s; 1575 int ncl; 1576 struct buf *bpa; 1577 int nwritten; 1578 int size; 1579 int maxcl; 1580 1581 s = splbio(); 1582 /* 1583 * right now we support clustered writing only to regular files. If 1584 * we find a clusterable block we could be in the middle of a cluster 1585 * rather then at the beginning. 1586 */ 1587 if ((vp->v_type == VREG) && 1588 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1589 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1590 1591 size = vp->v_mount->mnt_stat.f_iosize; 1592 maxcl = MAXPHYS / size; 1593 1594 VI_LOCK(vp); 1595 for (i = 1; i < maxcl; i++) { 1596 if ((bpa = gbincore(vp, lblkno + i)) && 1597 BUF_REFCNT(bpa) == 0 && 1598 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1599 (B_DELWRI | B_CLUSTEROK)) && 1600 (bpa->b_bufsize == size)) { 1601 if ((bpa->b_blkno == bpa->b_lblkno) || 1602 (bpa->b_blkno != 1603 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1604 break; 1605 } else { 1606 break; 1607 } 1608 } 1609 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1610 if ((bpa = gbincore(vp, lblkno - j)) && 1611 BUF_REFCNT(bpa) == 0 && 1612 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1613 (B_DELWRI | B_CLUSTEROK)) && 1614 (bpa->b_bufsize == size)) { 1615 if ((bpa->b_blkno == bpa->b_lblkno) || 1616 (bpa->b_blkno != 1617 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1618 break; 1619 } else { 1620 break; 1621 } 1622 } 1623 VI_UNLOCK(vp); 1624 --j; 1625 ncl = i + j; 1626 /* 1627 * this is a possible cluster write 1628 */ 1629 if (ncl != 1) { 1630 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1631 splx(s); 1632 return nwritten; 1633 } 1634 } 1635 1636 BUF_LOCK(bp, LK_EXCLUSIVE); 1637 bremfree(bp); 1638 bp->b_flags |= B_ASYNC; 1639 1640 splx(s); 1641 /* 1642 * default (old) behavior, writing out only one block 1643 * 1644 * XXX returns b_bufsize instead of b_bcount for nwritten? 1645 */ 1646 nwritten = bp->b_bufsize; 1647 (void) BUF_WRITE(bp); 1648 1649 return nwritten; 1650} 1651 1652/* 1653 * getnewbuf: 1654 * 1655 * Find and initialize a new buffer header, freeing up existing buffers 1656 * in the bufqueues as necessary. The new buffer is returned locked. 1657 * 1658 * Important: B_INVAL is not set. If the caller wishes to throw the 1659 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1660 * 1661 * We block if: 1662 * We have insufficient buffer headers 1663 * We have insufficient buffer space 1664 * buffer_map is too fragmented ( space reservation fails ) 1665 * If we have to flush dirty buffers ( but we try to avoid this ) 1666 * 1667 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1668 * Instead we ask the buf daemon to do it for us. We attempt to 1669 * avoid piecemeal wakeups of the pageout daemon. 1670 */ 1671 1672static struct buf * 1673getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1674{ 1675 struct buf *bp; 1676 struct buf *nbp; 1677 int defrag = 0; 1678 int nqindex; 1679 static int flushingbufs; 1680 1681 GIANT_REQUIRED; 1682 1683 /* 1684 * We can't afford to block since we might be holding a vnode lock, 1685 * which may prevent system daemons from running. We deal with 1686 * low-memory situations by proactively returning memory and running 1687 * async I/O rather then sync I/O. 1688 */ 1689 1690 ++getnewbufcalls; 1691 --getnewbufrestarts; 1692restart: 1693 ++getnewbufrestarts; 1694 1695 /* 1696 * Setup for scan. If we do not have enough free buffers, 1697 * we setup a degenerate case that immediately fails. Note 1698 * that if we are specially marked process, we are allowed to 1699 * dip into our reserves. 1700 * 1701 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1702 * 1703 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1704 * However, there are a number of cases (defragging, reusing, ...) 1705 * where we cannot backup. 1706 */ 1707 nqindex = QUEUE_EMPTYKVA; 1708 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1709 1710 if (nbp == NULL) { 1711 /* 1712 * If no EMPTYKVA buffers and we are either 1713 * defragging or reusing, locate a CLEAN buffer 1714 * to free or reuse. If bufspace useage is low 1715 * skip this step so we can allocate a new buffer. 1716 */ 1717 if (defrag || bufspace >= lobufspace) { 1718 nqindex = QUEUE_CLEAN; 1719 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1720 } 1721 1722 /* 1723 * If we could not find or were not allowed to reuse a 1724 * CLEAN buffer, check to see if it is ok to use an EMPTY 1725 * buffer. We can only use an EMPTY buffer if allocating 1726 * its KVA would not otherwise run us out of buffer space. 1727 */ 1728 if (nbp == NULL && defrag == 0 && 1729 bufspace + maxsize < hibufspace) { 1730 nqindex = QUEUE_EMPTY; 1731 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1732 } 1733 } 1734 1735 /* 1736 * Run scan, possibly freeing data and/or kva mappings on the fly 1737 * depending. 1738 */ 1739 1740 while ((bp = nbp) != NULL) { 1741 int qindex = nqindex; 1742 1743 /* 1744 * Calculate next bp ( we can only use it if we do not block 1745 * or do other fancy things ). 1746 */ 1747 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1748 switch(qindex) { 1749 case QUEUE_EMPTY: 1750 nqindex = QUEUE_EMPTYKVA; 1751 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1752 break; 1753 /* FALLTHROUGH */ 1754 case QUEUE_EMPTYKVA: 1755 nqindex = QUEUE_CLEAN; 1756 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1757 break; 1758 /* FALLTHROUGH */ 1759 case QUEUE_CLEAN: 1760 /* 1761 * nbp is NULL. 1762 */ 1763 break; 1764 } 1765 } 1766 1767 /* 1768 * Sanity Checks 1769 */ 1770 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1771 1772 /* 1773 * Note: we no longer distinguish between VMIO and non-VMIO 1774 * buffers. 1775 */ 1776 1777 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1778 1779 /* 1780 * If we are defragging then we need a buffer with 1781 * b_kvasize != 0. XXX this situation should no longer 1782 * occur, if defrag is non-zero the buffer's b_kvasize 1783 * should also be non-zero at this point. XXX 1784 */ 1785 if (defrag && bp->b_kvasize == 0) { 1786 printf("Warning: defrag empty buffer %p\n", bp); 1787 continue; 1788 } 1789 1790 /* 1791 * Start freeing the bp. This is somewhat involved. nbp 1792 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1793 */ 1794 1795 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1796 panic("getnewbuf: locked buf"); 1797 bremfree(bp); 1798 1799 if (qindex == QUEUE_CLEAN) { 1800 if (bp->b_flags & B_VMIO) { 1801 bp->b_flags &= ~B_ASYNC; 1802 vfs_vmio_release(bp); 1803 } 1804 if (bp->b_vp) 1805 brelvp(bp); 1806 } 1807 1808 /* 1809 * NOTE: nbp is now entirely invalid. We can only restart 1810 * the scan from this point on. 1811 * 1812 * Get the rest of the buffer freed up. b_kva* is still 1813 * valid after this operation. 1814 */ 1815 1816 if (bp->b_rcred != NOCRED) { 1817 crfree(bp->b_rcred); 1818 bp->b_rcred = NOCRED; 1819 } 1820 if (bp->b_wcred != NOCRED) { 1821 crfree(bp->b_wcred); 1822 bp->b_wcred = NOCRED; 1823 } 1824 if (LIST_FIRST(&bp->b_dep) != NULL) 1825 buf_deallocate(bp); 1826 if (bp->b_xflags & BX_BKGRDINPROG) 1827 panic("losing buffer 3"); 1828#ifdef USE_BUFHASH 1829 LIST_REMOVE(bp, b_hash); 1830 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1831#endif 1832 1833 if (bp->b_bufsize) 1834 allocbuf(bp, 0); 1835 1836 bp->b_flags = 0; 1837 bp->b_ioflags = 0; 1838 bp->b_xflags = 0; 1839 bp->b_dev = NODEV; 1840 bp->b_vp = NULL; 1841 bp->b_blkno = bp->b_lblkno = 0; 1842 bp->b_offset = NOOFFSET; 1843 bp->b_iodone = 0; 1844 bp->b_error = 0; 1845 bp->b_resid = 0; 1846 bp->b_bcount = 0; 1847 bp->b_npages = 0; 1848 bp->b_dirtyoff = bp->b_dirtyend = 0; 1849 bp->b_magic = B_MAGIC_BIO; 1850 bp->b_op = &buf_ops_bio; 1851 bp->b_object = NULL; 1852 1853 LIST_INIT(&bp->b_dep); 1854 1855 /* 1856 * If we are defragging then free the buffer. 1857 */ 1858 if (defrag) { 1859 bp->b_flags |= B_INVAL; 1860 bfreekva(bp); 1861 brelse(bp); 1862 defrag = 0; 1863 goto restart; 1864 } 1865 1866 /* 1867 * If we are overcomitted then recover the buffer and its 1868 * KVM space. This occurs in rare situations when multiple 1869 * processes are blocked in getnewbuf() or allocbuf(). 1870 */ 1871 if (bufspace >= hibufspace) 1872 flushingbufs = 1; 1873 if (flushingbufs && bp->b_kvasize != 0) { 1874 bp->b_flags |= B_INVAL; 1875 bfreekva(bp); 1876 brelse(bp); 1877 goto restart; 1878 } 1879 if (bufspace < lobufspace) 1880 flushingbufs = 0; 1881 break; 1882 } 1883 1884 /* 1885 * If we exhausted our list, sleep as appropriate. We may have to 1886 * wakeup various daemons and write out some dirty buffers. 1887 * 1888 * Generally we are sleeping due to insufficient buffer space. 1889 */ 1890 1891 if (bp == NULL) { 1892 int flags; 1893 char *waitmsg; 1894 1895 if (defrag) { 1896 flags = VFS_BIO_NEED_BUFSPACE; 1897 waitmsg = "nbufkv"; 1898 } else if (bufspace >= hibufspace) { 1899 waitmsg = "nbufbs"; 1900 flags = VFS_BIO_NEED_BUFSPACE; 1901 } else { 1902 waitmsg = "newbuf"; 1903 flags = VFS_BIO_NEED_ANY; 1904 } 1905 1906 bd_speedup(); /* heeeelp */ 1907 1908 needsbuffer |= flags; 1909 while (needsbuffer & flags) { 1910 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1911 waitmsg, slptimeo)) 1912 return (NULL); 1913 } 1914 } else { 1915 /* 1916 * We finally have a valid bp. We aren't quite out of the 1917 * woods, we still have to reserve kva space. In order 1918 * to keep fragmentation sane we only allocate kva in 1919 * BKVASIZE chunks. 1920 */ 1921 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1922 1923 if (maxsize != bp->b_kvasize) { 1924 vm_offset_t addr = 0; 1925 1926 bfreekva(bp); 1927 1928 if (vm_map_findspace(buffer_map, 1929 vm_map_min(buffer_map), maxsize, &addr)) { 1930 /* 1931 * Uh oh. Buffer map is to fragmented. We 1932 * must defragment the map. 1933 */ 1934 ++bufdefragcnt; 1935 defrag = 1; 1936 bp->b_flags |= B_INVAL; 1937 brelse(bp); 1938 goto restart; 1939 } 1940 if (addr) { 1941 vm_map_insert(buffer_map, NULL, 0, 1942 addr, addr + maxsize, 1943 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1944 1945 bp->b_kvabase = (caddr_t) addr; 1946 bp->b_kvasize = maxsize; 1947 bufspace += bp->b_kvasize; 1948 ++bufreusecnt; 1949 } 1950 } 1951 bp->b_data = bp->b_kvabase; 1952 } 1953 return(bp); 1954} 1955 1956/* 1957 * buf_daemon: 1958 * 1959 * buffer flushing daemon. Buffers are normally flushed by the 1960 * update daemon but if it cannot keep up this process starts to 1961 * take the load in an attempt to prevent getnewbuf() from blocking. 1962 */ 1963 1964static struct proc *bufdaemonproc; 1965 1966static struct kproc_desc buf_kp = { 1967 "bufdaemon", 1968 buf_daemon, 1969 &bufdaemonproc 1970}; 1971SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1972 1973static void 1974buf_daemon() 1975{ 1976 int s; 1977 1978 mtx_lock(&Giant); 1979 1980 /* 1981 * This process needs to be suspended prior to shutdown sync. 1982 */ 1983 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 1984 SHUTDOWN_PRI_LAST); 1985 1986 /* 1987 * This process is allowed to take the buffer cache to the limit 1988 */ 1989 s = splbio(); 1990 1991 for (;;) { 1992 kthread_suspend_check(bufdaemonproc); 1993 1994 bd_request = 0; 1995 1996 /* 1997 * Do the flush. Limit the amount of in-transit I/O we 1998 * allow to build up, otherwise we would completely saturate 1999 * the I/O system. Wakeup any waiting processes before we 2000 * normally would so they can run in parallel with our drain. 2001 */ 2002 while (numdirtybuffers > lodirtybuffers) { 2003 if (flushbufqueues() == 0) 2004 break; 2005 waitrunningbufspace(); 2006 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2007 } 2008 2009 /* 2010 * Only clear bd_request if we have reached our low water 2011 * mark. The buf_daemon normally waits 1 second and 2012 * then incrementally flushes any dirty buffers that have 2013 * built up, within reason. 2014 * 2015 * If we were unable to hit our low water mark and couldn't 2016 * find any flushable buffers, we sleep half a second. 2017 * Otherwise we loop immediately. 2018 */ 2019 if (numdirtybuffers <= lodirtybuffers) { 2020 /* 2021 * We reached our low water mark, reset the 2022 * request and sleep until we are needed again. 2023 * The sleep is just so the suspend code works. 2024 */ 2025 bd_request = 0; 2026 tsleep(&bd_request, PVM, "psleep", hz); 2027 } else { 2028 /* 2029 * We couldn't find any flushable dirty buffers but 2030 * still have too many dirty buffers, we 2031 * have to sleep and try again. (rare) 2032 */ 2033 tsleep(&bd_request, PVM, "qsleep", hz / 10); 2034 } 2035 } 2036} 2037 2038/* 2039 * flushbufqueues: 2040 * 2041 * Try to flush a buffer in the dirty queue. We must be careful to 2042 * free up B_INVAL buffers instead of write them, which NFS is 2043 * particularly sensitive to. 2044 */ 2045int flushwithdeps = 0; 2046SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2047 0, "Number of buffers flushed with dependecies that require rollbacks"); 2048static int 2049flushbufqueues(void) 2050{ 2051 struct thread *td = curthread; 2052 struct vnode *vp; 2053 struct buf *bp; 2054 2055 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2056 KASSERT((bp->b_flags & B_DELWRI), 2057 ("unexpected clean buffer %p", bp)); 2058 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) 2059 continue; 2060 if (bp->b_flags & B_INVAL) { 2061 if (BUF_LOCK(bp, LK_EXCLUSIVE) != 0) 2062 panic("flushbufqueues: locked buf"); 2063 bremfree(bp); 2064 brelse(bp); 2065 return (1); 2066 } 2067 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) 2068 continue; 2069 /* 2070 * We must hold the lock on a vnode before writing 2071 * one of its buffers. Otherwise we may confuse, or 2072 * in the case of a snapshot vnode, deadlock the 2073 * system. 2074 */ 2075 if ((vp = bp->b_vp) == NULL || 2076 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2077 vfs_bio_awrite(bp); 2078 if (vp != NULL) 2079 VOP_UNLOCK(vp, 0, td); 2080 return (1); 2081 } 2082 } 2083 /* 2084 * Could not find any buffers without rollback dependencies, 2085 * so just write the first one in the hopes of eventually 2086 * making progress. 2087 */ 2088 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2089 KASSERT((bp->b_flags & B_DELWRI), 2090 ("unexpected clean buffer %p", bp)); 2091 if ((bp->b_xflags & BX_BKGRDINPROG) != 0) 2092 continue; 2093 if (bp->b_flags & B_INVAL) { 2094 if (BUF_LOCK(bp, LK_EXCLUSIVE) != 0) 2095 panic("flushbufqueues: locked buf"); 2096 bremfree(bp); 2097 brelse(bp); 2098 return (1); 2099 } 2100 /* 2101 * We must hold the lock on a vnode before writing 2102 * one of its buffers. Otherwise we may confuse, or 2103 * in the case of a snapshot vnode, deadlock the 2104 * system. 2105 */ 2106 if ((vp = bp->b_vp) == NULL || 2107 vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2108 vfs_bio_awrite(bp); 2109 if (vp != NULL) 2110 VOP_UNLOCK(vp, 0, td); 2111 flushwithdeps += 1; 2112 return (0); 2113 } 2114 } 2115 return (0); 2116} 2117 2118/* 2119 * Check to see if a block is currently memory resident. 2120 */ 2121struct buf * 2122incore(struct vnode * vp, daddr_t blkno) 2123{ 2124 struct buf *bp; 2125 2126 int s = splbio(); 2127 VI_LOCK(vp); 2128 bp = gbincore(vp, blkno); 2129 VI_UNLOCK(vp); 2130 splx(s); 2131 return (bp); 2132} 2133 2134/* 2135 * Returns true if no I/O is needed to access the 2136 * associated VM object. This is like incore except 2137 * it also hunts around in the VM system for the data. 2138 */ 2139 2140int 2141inmem(struct vnode * vp, daddr_t blkno) 2142{ 2143 vm_object_t obj; 2144 vm_offset_t toff, tinc, size; 2145 vm_page_t m; 2146 vm_ooffset_t off; 2147 2148 GIANT_REQUIRED; 2149 ASSERT_VOP_LOCKED(vp, "inmem"); 2150 2151 if (incore(vp, blkno)) 2152 return 1; 2153 if (vp->v_mount == NULL) 2154 return 0; 2155 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2156 return 0; 2157 2158 size = PAGE_SIZE; 2159 if (size > vp->v_mount->mnt_stat.f_iosize) 2160 size = vp->v_mount->mnt_stat.f_iosize; 2161 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2162 2163 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2164 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2165 if (!m) 2166 goto notinmem; 2167 tinc = size; 2168 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2169 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2170 if (vm_page_is_valid(m, 2171 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2172 goto notinmem; 2173 } 2174 return 1; 2175 2176notinmem: 2177 return (0); 2178} 2179 2180/* 2181 * vfs_setdirty: 2182 * 2183 * Sets the dirty range for a buffer based on the status of the dirty 2184 * bits in the pages comprising the buffer. 2185 * 2186 * The range is limited to the size of the buffer. 2187 * 2188 * This routine is primarily used by NFS, but is generalized for the 2189 * B_VMIO case. 2190 */ 2191static void 2192vfs_setdirty(struct buf *bp) 2193{ 2194 int i; 2195 vm_object_t object; 2196 2197 GIANT_REQUIRED; 2198 /* 2199 * Degenerate case - empty buffer 2200 */ 2201 2202 if (bp->b_bufsize == 0) 2203 return; 2204 2205 /* 2206 * We qualify the scan for modified pages on whether the 2207 * object has been flushed yet. The OBJ_WRITEABLE flag 2208 * is not cleared simply by protecting pages off. 2209 */ 2210 2211 if ((bp->b_flags & B_VMIO) == 0) 2212 return; 2213 2214 object = bp->b_pages[0]->object; 2215 2216 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2217 printf("Warning: object %p writeable but not mightbedirty\n", object); 2218 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2219 printf("Warning: object %p mightbedirty but not writeable\n", object); 2220 2221 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2222 vm_offset_t boffset; 2223 vm_offset_t eoffset; 2224 2225 vm_page_lock_queues(); 2226 /* 2227 * test the pages to see if they have been modified directly 2228 * by users through the VM system. 2229 */ 2230 for (i = 0; i < bp->b_npages; i++) { 2231 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2232 vm_page_test_dirty(bp->b_pages[i]); 2233 } 2234 2235 /* 2236 * Calculate the encompassing dirty range, boffset and eoffset, 2237 * (eoffset - boffset) bytes. 2238 */ 2239 2240 for (i = 0; i < bp->b_npages; i++) { 2241 if (bp->b_pages[i]->dirty) 2242 break; 2243 } 2244 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2245 2246 for (i = bp->b_npages - 1; i >= 0; --i) { 2247 if (bp->b_pages[i]->dirty) { 2248 break; 2249 } 2250 } 2251 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2252 2253 vm_page_unlock_queues(); 2254 /* 2255 * Fit it to the buffer. 2256 */ 2257 2258 if (eoffset > bp->b_bcount) 2259 eoffset = bp->b_bcount; 2260 2261 /* 2262 * If we have a good dirty range, merge with the existing 2263 * dirty range. 2264 */ 2265 2266 if (boffset < eoffset) { 2267 if (bp->b_dirtyoff > boffset) 2268 bp->b_dirtyoff = boffset; 2269 if (bp->b_dirtyend < eoffset) 2270 bp->b_dirtyend = eoffset; 2271 } 2272 } 2273} 2274 2275/* 2276 * getblk: 2277 * 2278 * Get a block given a specified block and offset into a file/device. 2279 * The buffers B_DONE bit will be cleared on return, making it almost 2280 * ready for an I/O initiation. B_INVAL may or may not be set on 2281 * return. The caller should clear B_INVAL prior to initiating a 2282 * READ. 2283 * 2284 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2285 * an existing buffer. 2286 * 2287 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2288 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2289 * and then cleared based on the backing VM. If the previous buffer is 2290 * non-0-sized but invalid, B_CACHE will be cleared. 2291 * 2292 * If getblk() must create a new buffer, the new buffer is returned with 2293 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2294 * case it is returned with B_INVAL clear and B_CACHE set based on the 2295 * backing VM. 2296 * 2297 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2298 * B_CACHE bit is clear. 2299 * 2300 * What this means, basically, is that the caller should use B_CACHE to 2301 * determine whether the buffer is fully valid or not and should clear 2302 * B_INVAL prior to issuing a read. If the caller intends to validate 2303 * the buffer by loading its data area with something, the caller needs 2304 * to clear B_INVAL. If the caller does this without issuing an I/O, 2305 * the caller should set B_CACHE ( as an optimization ), else the caller 2306 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2307 * a write attempt or if it was a successfull read. If the caller 2308 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2309 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2310 */ 2311struct buf * 2312getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2313{ 2314 struct buf *bp; 2315 int s; 2316#ifdef USE_BUFHASH 2317 struct bufhashhdr *bh; 2318#endif 2319 ASSERT_VOP_LOCKED(vp, "getblk"); 2320 2321 if (size > MAXBSIZE) 2322 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2323 2324 s = splbio(); 2325loop: 2326 /* 2327 * Block if we are low on buffers. Certain processes are allowed 2328 * to completely exhaust the buffer cache. 2329 * 2330 * If this check ever becomes a bottleneck it may be better to 2331 * move it into the else, when gbincore() fails. At the moment 2332 * it isn't a problem. 2333 * 2334 * XXX remove if 0 sections (clean this up after its proven) 2335 */ 2336 if (numfreebuffers == 0) { 2337 if (curthread == PCPU_GET(idlethread)) 2338 return NULL; 2339 needsbuffer |= VFS_BIO_NEED_ANY; 2340 } 2341 2342 VI_LOCK(vp); 2343 if ((bp = gbincore(vp, blkno))) { 2344 VI_UNLOCK(vp); 2345 /* 2346 * Buffer is in-core. If the buffer is not busy, it must 2347 * be on a queue. 2348 */ 2349 2350 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2351 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2352 "getblk", slpflag, slptimeo) == ENOLCK) 2353 goto loop; 2354 splx(s); 2355 return (struct buf *) NULL; 2356 } 2357 2358 /* 2359 * The buffer is locked. B_CACHE is cleared if the buffer is 2360 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2361 * and for a VMIO buffer B_CACHE is adjusted according to the 2362 * backing VM cache. 2363 */ 2364 if (bp->b_flags & B_INVAL) 2365 bp->b_flags &= ~B_CACHE; 2366 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2367 bp->b_flags |= B_CACHE; 2368 bremfree(bp); 2369 2370 /* 2371 * check for size inconsistancies for non-VMIO case. 2372 */ 2373 2374 if (bp->b_bcount != size) { 2375 if ((bp->b_flags & B_VMIO) == 0 || 2376 (size > bp->b_kvasize)) { 2377 if (bp->b_flags & B_DELWRI) { 2378 bp->b_flags |= B_NOCACHE; 2379 BUF_WRITE(bp); 2380 } else { 2381 if ((bp->b_flags & B_VMIO) && 2382 (LIST_FIRST(&bp->b_dep) == NULL)) { 2383 bp->b_flags |= B_RELBUF; 2384 brelse(bp); 2385 } else { 2386 bp->b_flags |= B_NOCACHE; 2387 BUF_WRITE(bp); 2388 } 2389 } 2390 goto loop; 2391 } 2392 } 2393 2394 /* 2395 * If the size is inconsistant in the VMIO case, we can resize 2396 * the buffer. This might lead to B_CACHE getting set or 2397 * cleared. If the size has not changed, B_CACHE remains 2398 * unchanged from its previous state. 2399 */ 2400 2401 if (bp->b_bcount != size) 2402 allocbuf(bp, size); 2403 2404 KASSERT(bp->b_offset != NOOFFSET, 2405 ("getblk: no buffer offset")); 2406 2407 /* 2408 * A buffer with B_DELWRI set and B_CACHE clear must 2409 * be committed before we can return the buffer in 2410 * order to prevent the caller from issuing a read 2411 * ( due to B_CACHE not being set ) and overwriting 2412 * it. 2413 * 2414 * Most callers, including NFS and FFS, need this to 2415 * operate properly either because they assume they 2416 * can issue a read if B_CACHE is not set, or because 2417 * ( for example ) an uncached B_DELWRI might loop due 2418 * to softupdates re-dirtying the buffer. In the latter 2419 * case, B_CACHE is set after the first write completes, 2420 * preventing further loops. 2421 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2422 * above while extending the buffer, we cannot allow the 2423 * buffer to remain with B_CACHE set after the write 2424 * completes or it will represent a corrupt state. To 2425 * deal with this we set B_NOCACHE to scrap the buffer 2426 * after the write. 2427 * 2428 * We might be able to do something fancy, like setting 2429 * B_CACHE in bwrite() except if B_DELWRI is already set, 2430 * so the below call doesn't set B_CACHE, but that gets real 2431 * confusing. This is much easier. 2432 */ 2433 2434 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2435 bp->b_flags |= B_NOCACHE; 2436 BUF_WRITE(bp); 2437 goto loop; 2438 } 2439 2440 splx(s); 2441 bp->b_flags &= ~B_DONE; 2442 } else { 2443 int bsize, maxsize, vmio; 2444 off_t offset; 2445 2446 /* 2447 * Buffer is not in-core, create new buffer. The buffer 2448 * returned by getnewbuf() is locked. Note that the returned 2449 * buffer is also considered valid (not marked B_INVAL). 2450 */ 2451 VI_UNLOCK(vp); 2452 if (vn_isdisk(vp, NULL)) 2453 bsize = DEV_BSIZE; 2454 else if (vp->v_mountedhere) 2455 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2456 else if (vp->v_mount) 2457 bsize = vp->v_mount->mnt_stat.f_iosize; 2458 else 2459 bsize = size; 2460 2461 offset = blkno * bsize; 2462 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2463 (vp->v_vflag & VV_OBJBUF); 2464 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2465 maxsize = imax(maxsize, bsize); 2466 2467 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2468 if (slpflag || slptimeo) { 2469 splx(s); 2470 return NULL; 2471 } 2472 goto loop; 2473 } 2474 2475 /* 2476 * This code is used to make sure that a buffer is not 2477 * created while the getnewbuf routine is blocked. 2478 * This can be a problem whether the vnode is locked or not. 2479 * If the buffer is created out from under us, we have to 2480 * throw away the one we just created. There is now window 2481 * race because we are safely running at splbio() from the 2482 * point of the duplicate buffer creation through to here, 2483 * and we've locked the buffer. 2484 * 2485 * Note: this must occur before we associate the buffer 2486 * with the vp especially considering limitations in 2487 * the splay tree implementation when dealing with duplicate 2488 * lblkno's. 2489 */ 2490 VI_LOCK(vp); 2491 if (gbincore(vp, blkno)) { 2492 VI_UNLOCK(vp); 2493 bp->b_flags |= B_INVAL; 2494 brelse(bp); 2495 goto loop; 2496 } 2497 VI_UNLOCK(vp); 2498 2499 /* 2500 * Insert the buffer into the hash, so that it can 2501 * be found by incore. 2502 */ 2503 bp->b_blkno = bp->b_lblkno = blkno; 2504 bp->b_offset = offset; 2505 2506 bgetvp(vp, bp); 2507#ifdef USE_BUFHASH 2508 LIST_REMOVE(bp, b_hash); 2509 bh = bufhash(vp, blkno); 2510 LIST_INSERT_HEAD(bh, bp, b_hash); 2511#endif 2512 2513 /* 2514 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2515 * buffer size starts out as 0, B_CACHE will be set by 2516 * allocbuf() for the VMIO case prior to it testing the 2517 * backing store for validity. 2518 */ 2519 2520 if (vmio) { 2521 bp->b_flags |= B_VMIO; 2522#if defined(VFS_BIO_DEBUG) 2523 if (vp->v_type != VREG) 2524 printf("getblk: vmioing file type %d???\n", vp->v_type); 2525#endif 2526 VOP_GETVOBJECT(vp, &bp->b_object); 2527 } else { 2528 bp->b_flags &= ~B_VMIO; 2529 bp->b_object = NULL; 2530 } 2531 2532 allocbuf(bp, size); 2533 2534 splx(s); 2535 bp->b_flags &= ~B_DONE; 2536 } 2537 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2538 return (bp); 2539} 2540 2541/* 2542 * Get an empty, disassociated buffer of given size. The buffer is initially 2543 * set to B_INVAL. 2544 */ 2545struct buf * 2546geteblk(int size) 2547{ 2548 struct buf *bp; 2549 int s; 2550 int maxsize; 2551 2552 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2553 2554 s = splbio(); 2555 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2556 continue; 2557 splx(s); 2558 allocbuf(bp, size); 2559 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2560 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2561 return (bp); 2562} 2563 2564 2565/* 2566 * This code constitutes the buffer memory from either anonymous system 2567 * memory (in the case of non-VMIO operations) or from an associated 2568 * VM object (in the case of VMIO operations). This code is able to 2569 * resize a buffer up or down. 2570 * 2571 * Note that this code is tricky, and has many complications to resolve 2572 * deadlock or inconsistant data situations. Tread lightly!!! 2573 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2574 * the caller. Calling this code willy nilly can result in the loss of data. 2575 * 2576 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2577 * B_CACHE for the non-VMIO case. 2578 */ 2579 2580int 2581allocbuf(struct buf *bp, int size) 2582{ 2583 int newbsize, mbsize; 2584 int i; 2585 2586 GIANT_REQUIRED; 2587 2588 if (BUF_REFCNT(bp) == 0) 2589 panic("allocbuf: buffer not busy"); 2590 2591 if (bp->b_kvasize < size) 2592 panic("allocbuf: buffer too small"); 2593 2594 if ((bp->b_flags & B_VMIO) == 0) { 2595 caddr_t origbuf; 2596 int origbufsize; 2597 /* 2598 * Just get anonymous memory from the kernel. Don't 2599 * mess with B_CACHE. 2600 */ 2601 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2602 if (bp->b_flags & B_MALLOC) 2603 newbsize = mbsize; 2604 else 2605 newbsize = round_page(size); 2606 2607 if (newbsize < bp->b_bufsize) { 2608 /* 2609 * malloced buffers are not shrunk 2610 */ 2611 if (bp->b_flags & B_MALLOC) { 2612 if (newbsize) { 2613 bp->b_bcount = size; 2614 } else { 2615 free(bp->b_data, M_BIOBUF); 2616 if (bp->b_bufsize) { 2617 bufmallocspace -= bp->b_bufsize; 2618 bufspacewakeup(); 2619 bp->b_bufsize = 0; 2620 } 2621 bp->b_data = bp->b_kvabase; 2622 bp->b_bcount = 0; 2623 bp->b_flags &= ~B_MALLOC; 2624 } 2625 return 1; 2626 } 2627 vm_hold_free_pages( 2628 bp, 2629 (vm_offset_t) bp->b_data + newbsize, 2630 (vm_offset_t) bp->b_data + bp->b_bufsize); 2631 } else if (newbsize > bp->b_bufsize) { 2632 /* 2633 * We only use malloced memory on the first allocation. 2634 * and revert to page-allocated memory when the buffer 2635 * grows. 2636 */ 2637 if ( (bufmallocspace < maxbufmallocspace) && 2638 (bp->b_bufsize == 0) && 2639 (mbsize <= PAGE_SIZE/2)) { 2640 2641 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2642 bp->b_bufsize = mbsize; 2643 bp->b_bcount = size; 2644 bp->b_flags |= B_MALLOC; 2645 bufmallocspace += mbsize; 2646 return 1; 2647 } 2648 origbuf = NULL; 2649 origbufsize = 0; 2650 /* 2651 * If the buffer is growing on its other-than-first allocation, 2652 * then we revert to the page-allocation scheme. 2653 */ 2654 if (bp->b_flags & B_MALLOC) { 2655 origbuf = bp->b_data; 2656 origbufsize = bp->b_bufsize; 2657 bp->b_data = bp->b_kvabase; 2658 if (bp->b_bufsize) { 2659 bufmallocspace -= bp->b_bufsize; 2660 bufspacewakeup(); 2661 bp->b_bufsize = 0; 2662 } 2663 bp->b_flags &= ~B_MALLOC; 2664 newbsize = round_page(newbsize); 2665 } 2666 vm_hold_load_pages( 2667 bp, 2668 (vm_offset_t) bp->b_data + bp->b_bufsize, 2669 (vm_offset_t) bp->b_data + newbsize); 2670 if (origbuf) { 2671 bcopy(origbuf, bp->b_data, origbufsize); 2672 free(origbuf, M_BIOBUF); 2673 } 2674 } 2675 } else { 2676 int desiredpages; 2677 2678 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2679 desiredpages = (size == 0) ? 0 : 2680 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2681 2682 if (bp->b_flags & B_MALLOC) 2683 panic("allocbuf: VMIO buffer can't be malloced"); 2684 /* 2685 * Set B_CACHE initially if buffer is 0 length or will become 2686 * 0-length. 2687 */ 2688 if (size == 0 || bp->b_bufsize == 0) 2689 bp->b_flags |= B_CACHE; 2690 2691 if (newbsize < bp->b_bufsize) { 2692 /* 2693 * DEV_BSIZE aligned new buffer size is less then the 2694 * DEV_BSIZE aligned existing buffer size. Figure out 2695 * if we have to remove any pages. 2696 */ 2697 if (desiredpages < bp->b_npages) { 2698 vm_page_t m; 2699 2700 vm_page_lock_queues(); 2701 for (i = desiredpages; i < bp->b_npages; i++) { 2702 /* 2703 * the page is not freed here -- it 2704 * is the responsibility of 2705 * vnode_pager_setsize 2706 */ 2707 m = bp->b_pages[i]; 2708 KASSERT(m != bogus_page, 2709 ("allocbuf: bogus page found")); 2710 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2711 vm_page_lock_queues(); 2712 2713 bp->b_pages[i] = NULL; 2714 vm_page_unwire(m, 0); 2715 } 2716 vm_page_unlock_queues(); 2717 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2718 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2719 bp->b_npages = desiredpages; 2720 } 2721 } else if (size > bp->b_bcount) { 2722 /* 2723 * We are growing the buffer, possibly in a 2724 * byte-granular fashion. 2725 */ 2726 struct vnode *vp; 2727 vm_object_t obj; 2728 vm_offset_t toff; 2729 vm_offset_t tinc; 2730 2731 /* 2732 * Step 1, bring in the VM pages from the object, 2733 * allocating them if necessary. We must clear 2734 * B_CACHE if these pages are not valid for the 2735 * range covered by the buffer. 2736 */ 2737 2738 vp = bp->b_vp; 2739 obj = bp->b_object; 2740 2741 while (bp->b_npages < desiredpages) { 2742 vm_page_t m; 2743 vm_pindex_t pi; 2744 2745 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2746 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2747 /* 2748 * note: must allocate system pages 2749 * since blocking here could intefere 2750 * with paging I/O, no matter which 2751 * process we are. 2752 */ 2753 m = vm_page_alloc(obj, pi, 2754 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 2755 if (m == NULL) { 2756 VM_WAIT; 2757 vm_pageout_deficit += desiredpages - bp->b_npages; 2758 } else { 2759 vm_page_lock_queues(); 2760 vm_page_wakeup(m); 2761 vm_page_unlock_queues(); 2762 bp->b_flags &= ~B_CACHE; 2763 bp->b_pages[bp->b_npages] = m; 2764 ++bp->b_npages; 2765 } 2766 continue; 2767 } 2768 2769 /* 2770 * We found a page. If we have to sleep on it, 2771 * retry because it might have gotten freed out 2772 * from under us. 2773 * 2774 * We can only test PG_BUSY here. Blocking on 2775 * m->busy might lead to a deadlock: 2776 * 2777 * vm_fault->getpages->cluster_read->allocbuf 2778 * 2779 */ 2780 vm_page_lock_queues(); 2781 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2782 continue; 2783 2784 /* 2785 * We have a good page. Should we wakeup the 2786 * page daemon? 2787 */ 2788 if ((curproc != pageproc) && 2789 ((m->queue - m->pc) == PQ_CACHE) && 2790 ((cnt.v_free_count + cnt.v_cache_count) < 2791 (cnt.v_free_min + cnt.v_cache_min))) { 2792 pagedaemon_wakeup(); 2793 } 2794 vm_page_flag_clear(m, PG_ZERO); 2795 vm_page_wire(m); 2796 vm_page_unlock_queues(); 2797 bp->b_pages[bp->b_npages] = m; 2798 ++bp->b_npages; 2799 } 2800 2801 /* 2802 * Step 2. We've loaded the pages into the buffer, 2803 * we have to figure out if we can still have B_CACHE 2804 * set. Note that B_CACHE is set according to the 2805 * byte-granular range ( bcount and size ), new the 2806 * aligned range ( newbsize ). 2807 * 2808 * The VM test is against m->valid, which is DEV_BSIZE 2809 * aligned. Needless to say, the validity of the data 2810 * needs to also be DEV_BSIZE aligned. Note that this 2811 * fails with NFS if the server or some other client 2812 * extends the file's EOF. If our buffer is resized, 2813 * B_CACHE may remain set! XXX 2814 */ 2815 2816 toff = bp->b_bcount; 2817 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2818 2819 while ((bp->b_flags & B_CACHE) && toff < size) { 2820 vm_pindex_t pi; 2821 2822 if (tinc > (size - toff)) 2823 tinc = size - toff; 2824 2825 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2826 PAGE_SHIFT; 2827 2828 vfs_buf_test_cache( 2829 bp, 2830 bp->b_offset, 2831 toff, 2832 tinc, 2833 bp->b_pages[pi] 2834 ); 2835 toff += tinc; 2836 tinc = PAGE_SIZE; 2837 } 2838 2839 /* 2840 * Step 3, fixup the KVM pmap. Remember that 2841 * bp->b_data is relative to bp->b_offset, but 2842 * bp->b_offset may be offset into the first page. 2843 */ 2844 2845 bp->b_data = (caddr_t) 2846 trunc_page((vm_offset_t)bp->b_data); 2847 pmap_qenter( 2848 (vm_offset_t)bp->b_data, 2849 bp->b_pages, 2850 bp->b_npages 2851 ); 2852 2853 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2854 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2855 } 2856 } 2857 if (newbsize < bp->b_bufsize) 2858 bufspacewakeup(); 2859 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2860 bp->b_bcount = size; /* requested buffer size */ 2861 return 1; 2862} 2863 2864void 2865biodone(struct bio *bp) 2866{ 2867 bp->bio_flags |= BIO_DONE; 2868 if (bp->bio_done != NULL) 2869 bp->bio_done(bp); 2870 else 2871 wakeup(bp); 2872} 2873 2874/* 2875 * Wait for a BIO to finish. 2876 * 2877 * XXX: resort to a timeout for now. The optimal locking (if any) for this 2878 * case is not yet clear. 2879 */ 2880int 2881biowait(struct bio *bp, const char *wchan) 2882{ 2883 2884 while ((bp->bio_flags & BIO_DONE) == 0) 2885 msleep(bp, NULL, PRIBIO, wchan, hz / 10); 2886 if (bp->bio_error != 0) 2887 return (bp->bio_error); 2888 if (!(bp->bio_flags & BIO_ERROR)) 2889 return (0); 2890 return (EIO); 2891} 2892 2893void 2894biofinish(struct bio *bp, struct devstat *stat, int error) 2895{ 2896 2897 if (error) { 2898 bp->bio_error = error; 2899 bp->bio_flags |= BIO_ERROR; 2900 } 2901 if (stat != NULL) 2902 devstat_end_transaction_bio(stat, bp); 2903 biodone(bp); 2904} 2905 2906void 2907bioq_init(struct bio_queue_head *head) 2908{ 2909 TAILQ_INIT(&head->queue); 2910 head->last_pblkno = 0; 2911 head->insert_point = NULL; 2912 head->switch_point = NULL; 2913} 2914 2915void 2916bioq_remove(struct bio_queue_head *head, struct bio *bp) 2917{ 2918 if (bp == head->switch_point) 2919 head->switch_point = TAILQ_NEXT(bp, bio_queue); 2920 if (bp == head->insert_point) { 2921 head->insert_point = TAILQ_PREV(bp, bio_queue, bio_queue); 2922 if (head->insert_point == NULL) 2923 head->last_pblkno = 0; 2924 } else if (bp == TAILQ_FIRST(&head->queue)) 2925 head->last_pblkno = bp->bio_pblkno; 2926 TAILQ_REMOVE(&head->queue, bp, bio_queue); 2927 if (TAILQ_FIRST(&head->queue) == head->switch_point) 2928 head->switch_point = NULL; 2929} 2930 2931/* 2932 * bufwait: 2933 * 2934 * Wait for buffer I/O completion, returning error status. The buffer 2935 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 2936 * error and cleared. 2937 */ 2938int 2939bufwait(register struct buf * bp) 2940{ 2941 int s; 2942 2943 s = splbio(); 2944 while ((bp->b_flags & B_DONE) == 0) { 2945 if (bp->b_iocmd == BIO_READ) 2946 tsleep(bp, PRIBIO, "biord", 0); 2947 else 2948 tsleep(bp, PRIBIO, "biowr", 0); 2949 } 2950 splx(s); 2951 if (bp->b_flags & B_EINTR) { 2952 bp->b_flags &= ~B_EINTR; 2953 return (EINTR); 2954 } 2955 if (bp->b_ioflags & BIO_ERROR) { 2956 return (bp->b_error ? bp->b_error : EIO); 2957 } else { 2958 return (0); 2959 } 2960} 2961 2962 /* 2963 * Call back function from struct bio back up to struct buf. 2964 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 2965 */ 2966void 2967bufdonebio(struct bio *bp) 2968{ 2969 bufdone(bp->bio_caller2); 2970} 2971 2972/* 2973 * bufdone: 2974 * 2975 * Finish I/O on a buffer, optionally calling a completion function. 2976 * This is usually called from an interrupt so process blocking is 2977 * not allowed. 2978 * 2979 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2980 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2981 * assuming B_INVAL is clear. 2982 * 2983 * For the VMIO case, we set B_CACHE if the op was a read and no 2984 * read error occured, or if the op was a write. B_CACHE is never 2985 * set if the buffer is invalid or otherwise uncacheable. 2986 * 2987 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2988 * initiator to leave B_INVAL set to brelse the buffer out of existance 2989 * in the biodone routine. 2990 */ 2991void 2992bufdone(struct buf *bp) 2993{ 2994 int s; 2995 void (*biodone)(struct buf *); 2996 2997 GIANT_REQUIRED; 2998 2999 s = splbio(); 3000 3001 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3002 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3003 3004 bp->b_flags |= B_DONE; 3005 runningbufwakeup(bp); 3006 3007 if (bp->b_iocmd == BIO_DELETE) { 3008 brelse(bp); 3009 splx(s); 3010 return; 3011 } 3012 3013 if (bp->b_iocmd == BIO_WRITE) { 3014 vwakeup(bp); 3015 } 3016 3017 /* call optional completion function if requested */ 3018 if (bp->b_iodone != NULL) { 3019 biodone = bp->b_iodone; 3020 bp->b_iodone = NULL; 3021 (*biodone) (bp); 3022 splx(s); 3023 return; 3024 } 3025 if (LIST_FIRST(&bp->b_dep) != NULL) 3026 buf_complete(bp); 3027 3028 if (bp->b_flags & B_VMIO) { 3029 int i; 3030 vm_ooffset_t foff; 3031 vm_page_t m; 3032 vm_object_t obj; 3033 int iosize; 3034 struct vnode *vp = bp->b_vp; 3035 3036 obj = bp->b_object; 3037 3038#if defined(VFS_BIO_DEBUG) 3039 mp_fixme("usecount and vflag accessed without locks."); 3040 if (vp->v_usecount == 0) { 3041 panic("biodone: zero vnode ref count"); 3042 } 3043 3044 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3045 panic("biodone: vnode is not setup for merged cache"); 3046 } 3047#endif 3048 3049 foff = bp->b_offset; 3050 KASSERT(bp->b_offset != NOOFFSET, 3051 ("biodone: no buffer offset")); 3052 3053#if defined(VFS_BIO_DEBUG) 3054 if (obj->paging_in_progress < bp->b_npages) { 3055 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3056 obj->paging_in_progress, bp->b_npages); 3057 } 3058#endif 3059 3060 /* 3061 * Set B_CACHE if the op was a normal read and no error 3062 * occured. B_CACHE is set for writes in the b*write() 3063 * routines. 3064 */ 3065 iosize = bp->b_bcount - bp->b_resid; 3066 if (bp->b_iocmd == BIO_READ && 3067 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3068 !(bp->b_ioflags & BIO_ERROR)) { 3069 bp->b_flags |= B_CACHE; 3070 } 3071 vm_page_lock_queues(); 3072 for (i = 0; i < bp->b_npages; i++) { 3073 int bogusflag = 0; 3074 int resid; 3075 3076 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3077 if (resid > iosize) 3078 resid = iosize; 3079 3080 /* 3081 * cleanup bogus pages, restoring the originals 3082 */ 3083 m = bp->b_pages[i]; 3084 if (m == bogus_page) { 3085 bogusflag = 1; 3086 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3087 if (m == NULL) 3088 panic("biodone: page disappeared!"); 3089 bp->b_pages[i] = m; 3090 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3091 } 3092#if defined(VFS_BIO_DEBUG) 3093 if (OFF_TO_IDX(foff) != m->pindex) { 3094 printf( 3095"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3096 (intmax_t)foff, (uintmax_t)m->pindex); 3097 } 3098#endif 3099 3100 /* 3101 * In the write case, the valid and clean bits are 3102 * already changed correctly ( see bdwrite() ), so we 3103 * only need to do this here in the read case. 3104 */ 3105 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3106 vfs_page_set_valid(bp, foff, i, m); 3107 } 3108 vm_page_flag_clear(m, PG_ZERO); 3109 3110 /* 3111 * when debugging new filesystems or buffer I/O methods, this 3112 * is the most common error that pops up. if you see this, you 3113 * have not set the page busy flag correctly!!! 3114 */ 3115 if (m->busy == 0) { 3116 printf("biodone: page busy < 0, " 3117 "pindex: %d, foff: 0x(%x,%x), " 3118 "resid: %d, index: %d\n", 3119 (int) m->pindex, (int)(foff >> 32), 3120 (int) foff & 0xffffffff, resid, i); 3121 if (!vn_isdisk(vp, NULL)) 3122 printf(" iosize: %ld, lblkno: %jd, flags: 0x%lx, npages: %d\n", 3123 bp->b_vp->v_mount->mnt_stat.f_iosize, 3124 (intmax_t) bp->b_lblkno, 3125 bp->b_flags, bp->b_npages); 3126 else 3127 printf(" VDEV, lblkno: %jd, flags: 0x%lx, npages: %d\n", 3128 (intmax_t) bp->b_lblkno, 3129 bp->b_flags, bp->b_npages); 3130 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 3131 m->valid, m->dirty, m->wire_count); 3132 panic("biodone: page busy < 0\n"); 3133 } 3134 vm_page_io_finish(m); 3135 vm_object_pip_subtract(obj, 1); 3136 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3137 iosize -= resid; 3138 } 3139 vm_page_unlock_queues(); 3140 if (obj) 3141 vm_object_pip_wakeupn(obj, 0); 3142 } 3143 3144 /* 3145 * For asynchronous completions, release the buffer now. The brelse 3146 * will do a wakeup there if necessary - so no need to do a wakeup 3147 * here in the async case. The sync case always needs to do a wakeup. 3148 */ 3149 3150 if (bp->b_flags & B_ASYNC) { 3151 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3152 brelse(bp); 3153 else 3154 bqrelse(bp); 3155 } else { 3156 wakeup(bp); 3157 } 3158 splx(s); 3159} 3160 3161/* 3162 * This routine is called in lieu of iodone in the case of 3163 * incomplete I/O. This keeps the busy status for pages 3164 * consistant. 3165 */ 3166void 3167vfs_unbusy_pages(struct buf * bp) 3168{ 3169 int i; 3170 3171 GIANT_REQUIRED; 3172 3173 runningbufwakeup(bp); 3174 if (bp->b_flags & B_VMIO) { 3175 vm_object_t obj; 3176 3177 obj = bp->b_object; 3178 vm_page_lock_queues(); 3179 for (i = 0; i < bp->b_npages; i++) { 3180 vm_page_t m = bp->b_pages[i]; 3181 3182 if (m == bogus_page) { 3183 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3184 if (!m) { 3185 panic("vfs_unbusy_pages: page missing\n"); 3186 } 3187 bp->b_pages[i] = m; 3188 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3189 } 3190 vm_object_pip_subtract(obj, 1); 3191 vm_page_flag_clear(m, PG_ZERO); 3192 vm_page_io_finish(m); 3193 } 3194 vm_page_unlock_queues(); 3195 vm_object_pip_wakeupn(obj, 0); 3196 } 3197} 3198 3199/* 3200 * vfs_page_set_valid: 3201 * 3202 * Set the valid bits in a page based on the supplied offset. The 3203 * range is restricted to the buffer's size. 3204 * 3205 * This routine is typically called after a read completes. 3206 */ 3207static void 3208vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3209{ 3210 vm_ooffset_t soff, eoff; 3211 3212 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3213 /* 3214 * Start and end offsets in buffer. eoff - soff may not cross a 3215 * page boundry or cross the end of the buffer. The end of the 3216 * buffer, in this case, is our file EOF, not the allocation size 3217 * of the buffer. 3218 */ 3219 soff = off; 3220 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3221 if (eoff > bp->b_offset + bp->b_bcount) 3222 eoff = bp->b_offset + bp->b_bcount; 3223 3224 /* 3225 * Set valid range. This is typically the entire buffer and thus the 3226 * entire page. 3227 */ 3228 if (eoff > soff) { 3229 vm_page_set_validclean( 3230 m, 3231 (vm_offset_t) (soff & PAGE_MASK), 3232 (vm_offset_t) (eoff - soff) 3233 ); 3234 } 3235} 3236 3237/* 3238 * This routine is called before a device strategy routine. 3239 * It is used to tell the VM system that paging I/O is in 3240 * progress, and treat the pages associated with the buffer 3241 * almost as being PG_BUSY. Also the object paging_in_progress 3242 * flag is handled to make sure that the object doesn't become 3243 * inconsistant. 3244 * 3245 * Since I/O has not been initiated yet, certain buffer flags 3246 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3247 * and should be ignored. 3248 */ 3249void 3250vfs_busy_pages(struct buf * bp, int clear_modify) 3251{ 3252 int i, bogus; 3253 3254 if (bp->b_flags & B_VMIO) { 3255 vm_object_t obj; 3256 vm_ooffset_t foff; 3257 3258 obj = bp->b_object; 3259 foff = bp->b_offset; 3260 KASSERT(bp->b_offset != NOOFFSET, 3261 ("vfs_busy_pages: no buffer offset")); 3262 vfs_setdirty(bp); 3263retry: 3264 vm_page_lock_queues(); 3265 for (i = 0; i < bp->b_npages; i++) { 3266 vm_page_t m = bp->b_pages[i]; 3267 3268 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3269 goto retry; 3270 } 3271 bogus = 0; 3272 for (i = 0; i < bp->b_npages; i++) { 3273 vm_page_t m = bp->b_pages[i]; 3274 3275 vm_page_flag_clear(m, PG_ZERO); 3276 if ((bp->b_flags & B_CLUSTER) == 0) { 3277 vm_object_pip_add(obj, 1); 3278 vm_page_io_start(m); 3279 } 3280 /* 3281 * When readying a buffer for a read ( i.e 3282 * clear_modify == 0 ), it is important to do 3283 * bogus_page replacement for valid pages in 3284 * partially instantiated buffers. Partially 3285 * instantiated buffers can, in turn, occur when 3286 * reconstituting a buffer from its VM backing store 3287 * base. We only have to do this if B_CACHE is 3288 * clear ( which causes the I/O to occur in the 3289 * first place ). The replacement prevents the read 3290 * I/O from overwriting potentially dirty VM-backed 3291 * pages. XXX bogus page replacement is, uh, bogus. 3292 * It may not work properly with small-block devices. 3293 * We need to find a better way. 3294 */ 3295 pmap_remove_all(m); 3296 if (clear_modify) 3297 vfs_page_set_valid(bp, foff, i, m); 3298 else if (m->valid == VM_PAGE_BITS_ALL && 3299 (bp->b_flags & B_CACHE) == 0) { 3300 bp->b_pages[i] = bogus_page; 3301 bogus++; 3302 } 3303 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3304 } 3305 vm_page_unlock_queues(); 3306 if (bogus) 3307 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3308 } 3309} 3310 3311/* 3312 * Tell the VM system that the pages associated with this buffer 3313 * are clean. This is used for delayed writes where the data is 3314 * going to go to disk eventually without additional VM intevention. 3315 * 3316 * Note that while we only really need to clean through to b_bcount, we 3317 * just go ahead and clean through to b_bufsize. 3318 */ 3319static void 3320vfs_clean_pages(struct buf * bp) 3321{ 3322 int i; 3323 3324 GIANT_REQUIRED; 3325 3326 if (bp->b_flags & B_VMIO) { 3327 vm_ooffset_t foff; 3328 3329 foff = bp->b_offset; 3330 KASSERT(bp->b_offset != NOOFFSET, 3331 ("vfs_clean_pages: no buffer offset")); 3332 vm_page_lock_queues(); 3333 for (i = 0; i < bp->b_npages; i++) { 3334 vm_page_t m = bp->b_pages[i]; 3335 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3336 vm_ooffset_t eoff = noff; 3337 3338 if (eoff > bp->b_offset + bp->b_bufsize) 3339 eoff = bp->b_offset + bp->b_bufsize; 3340 vfs_page_set_valid(bp, foff, i, m); 3341 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3342 foff = noff; 3343 } 3344 vm_page_unlock_queues(); 3345 } 3346} 3347 3348/* 3349 * vfs_bio_set_validclean: 3350 * 3351 * Set the range within the buffer to valid and clean. The range is 3352 * relative to the beginning of the buffer, b_offset. Note that b_offset 3353 * itself may be offset from the beginning of the first page. 3354 * 3355 */ 3356 3357void 3358vfs_bio_set_validclean(struct buf *bp, int base, int size) 3359{ 3360 if (bp->b_flags & B_VMIO) { 3361 int i; 3362 int n; 3363 3364 /* 3365 * Fixup base to be relative to beginning of first page. 3366 * Set initial n to be the maximum number of bytes in the 3367 * first page that can be validated. 3368 */ 3369 3370 base += (bp->b_offset & PAGE_MASK); 3371 n = PAGE_SIZE - (base & PAGE_MASK); 3372 3373 vm_page_lock_queues(); 3374 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3375 vm_page_t m = bp->b_pages[i]; 3376 3377 if (n > size) 3378 n = size; 3379 3380 vm_page_set_validclean(m, base & PAGE_MASK, n); 3381 base += n; 3382 size -= n; 3383 n = PAGE_SIZE; 3384 } 3385 vm_page_unlock_queues(); 3386 } 3387} 3388 3389/* 3390 * vfs_bio_clrbuf: 3391 * 3392 * clear a buffer. This routine essentially fakes an I/O, so we need 3393 * to clear BIO_ERROR and B_INVAL. 3394 * 3395 * Note that while we only theoretically need to clear through b_bcount, 3396 * we go ahead and clear through b_bufsize. 3397 */ 3398 3399void 3400vfs_bio_clrbuf(struct buf *bp) 3401{ 3402 int i, mask = 0; 3403 caddr_t sa, ea; 3404 3405 GIANT_REQUIRED; 3406 3407 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3408 bp->b_flags &= ~B_INVAL; 3409 bp->b_ioflags &= ~BIO_ERROR; 3410 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3411 (bp->b_offset & PAGE_MASK) == 0) { 3412 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3413 if ((bp->b_pages[0]->valid & mask) == mask) { 3414 bp->b_resid = 0; 3415 return; 3416 } 3417 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3418 ((bp->b_pages[0]->valid & mask) == 0)) { 3419 bzero(bp->b_data, bp->b_bufsize); 3420 bp->b_pages[0]->valid |= mask; 3421 bp->b_resid = 0; 3422 return; 3423 } 3424 } 3425 ea = sa = bp->b_data; 3426 for(i=0;i<bp->b_npages;i++,sa=ea) { 3427 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3428 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3429 ea = (caddr_t)(vm_offset_t)ulmin( 3430 (u_long)(vm_offset_t)ea, 3431 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3432 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3433 if ((bp->b_pages[i]->valid & mask) == mask) 3434 continue; 3435 if ((bp->b_pages[i]->valid & mask) == 0) { 3436 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3437 bzero(sa, ea - sa); 3438 } 3439 } else { 3440 for (; sa < ea; sa += DEV_BSIZE, j++) { 3441 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3442 (bp->b_pages[i]->valid & (1<<j)) == 0) 3443 bzero(sa, DEV_BSIZE); 3444 } 3445 } 3446 bp->b_pages[i]->valid |= mask; 3447 vm_page_lock_queues(); 3448 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3449 vm_page_unlock_queues(); 3450 } 3451 bp->b_resid = 0; 3452 } else { 3453 clrbuf(bp); 3454 } 3455} 3456 3457/* 3458 * vm_hold_load_pages and vm_hold_free_pages get pages into 3459 * a buffers address space. The pages are anonymous and are 3460 * not associated with a file object. 3461 */ 3462static void 3463vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3464{ 3465 vm_offset_t pg; 3466 vm_page_t p; 3467 int index; 3468 3469 GIANT_REQUIRED; 3470 3471 to = round_page(to); 3472 from = round_page(from); 3473 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3474 3475 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3476tryagain: 3477 /* 3478 * note: must allocate system pages since blocking here 3479 * could intefere with paging I/O, no matter which 3480 * process we are. 3481 */ 3482 vm_object_lock(kernel_object); 3483 p = vm_page_alloc(kernel_object, 3484 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3485 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3486 vm_object_unlock(kernel_object); 3487 if (!p) { 3488 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3489 VM_WAIT; 3490 goto tryagain; 3491 } 3492 vm_page_lock_queues(); 3493 p->valid = VM_PAGE_BITS_ALL; 3494 vm_page_flag_clear(p, PG_ZERO); 3495 vm_page_unlock_queues(); 3496 pmap_qenter(pg, &p, 1); 3497 bp->b_pages[index] = p; 3498 vm_page_lock_queues(); 3499 vm_page_wakeup(p); 3500 vm_page_unlock_queues(); 3501 } 3502 bp->b_npages = index; 3503} 3504 3505/* Return pages associated with this buf to the vm system */ 3506static void 3507vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3508{ 3509 vm_offset_t pg; 3510 vm_page_t p; 3511 int index, newnpages; 3512 3513 GIANT_REQUIRED; 3514 3515 from = round_page(from); 3516 to = round_page(to); 3517 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3518 3519 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3520 p = bp->b_pages[index]; 3521 if (p && (index < bp->b_npages)) { 3522 if (p->busy) { 3523 printf( 3524 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3525 (intmax_t)bp->b_blkno, 3526 (intmax_t)bp->b_lblkno); 3527 } 3528 bp->b_pages[index] = NULL; 3529 pmap_qremove(pg, 1); 3530 vm_page_lock_queues(); 3531 vm_page_busy(p); 3532 vm_page_unwire(p, 0); 3533 vm_page_free(p); 3534 vm_page_unlock_queues(); 3535 } 3536 } 3537 bp->b_npages = newnpages; 3538} 3539 3540 3541#include "opt_ddb.h" 3542#ifdef DDB 3543#include <ddb/ddb.h> 3544 3545/* DDB command to show buffer data */ 3546DB_SHOW_COMMAND(buffer, db_show_buffer) 3547{ 3548 /* get args */ 3549 struct buf *bp = (struct buf *)addr; 3550 3551 if (!have_addr) { 3552 db_printf("usage: show buffer <addr>\n"); 3553 return; 3554 } 3555 3556 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3557 db_printf( 3558 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3559 "b_dev = (%d,%d), b_data = %p, b_blkno = %jd, b_pblkno = %jd\n", 3560 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3561 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3562 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3563 if (bp->b_npages) { 3564 int i; 3565 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3566 for (i = 0; i < bp->b_npages; i++) { 3567 vm_page_t m; 3568 m = bp->b_pages[i]; 3569 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3570 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3571 if ((i + 1) < bp->b_npages) 3572 db_printf(","); 3573 } 3574 db_printf("\n"); 3575 } 3576} 3577#endif /* DDB */ 3578