vfs_bio.c revision 98631
1219888Sed/* 2219888Sed * Copyright (c) 1994,1997 John S. Dyson 3219888Sed * All rights reserved. 4219888Sed * 5219888Sed * Redistribution and use in source and binary forms, with or without 6219888Sed * modification, are permitted provided that the following conditions 7219888Sed * are met: 8219888Sed * 1. Redistributions of source code must retain the above copyright 9219888Sed * notice immediately at the beginning of the file, without modification, 10219888Sed * this list of conditions, and the following disclaimer. 11219888Sed * 2. Absolutely no warranty of function or purpose is made by the author 12219888Sed * John S. Dyson. 13219888Sed * 14219888Sed * $FreeBSD: head/sys/kern/vfs_bio.c 98631 2002-06-22 19:09:35Z dillon $ 15219888Sed */ 16219888Sed 17219888Sed/* 18219888Sed * this file contains a new buffer I/O scheme implementing a coherent 19219888Sed * VM object and buffer cache scheme. Pains have been taken to make 20219888Sed * sure that the performance degradation associated with schemes such 21219888Sed * as this is not realized. 22219888Sed * 23219888Sed * Author: John S. Dyson 24219888Sed * Significant help during the development and debugging phases 25219888Sed * had been provided by David Greenman, also of the FreeBSD core team. 26219888Sed * 27219888Sed * see man buf(9) for more info. 28219888Sed */ 29219888Sed 30219888Sed#include <sys/param.h> 31219888Sed#include <sys/systm.h> 32219888Sed#include <sys/stdint.h> 33219888Sed#include <sys/bio.h> 34219888Sed#include <sys/buf.h> 35219888Sed#include <sys/eventhandler.h> 36219888Sed#include <sys/lock.h> 37219888Sed#include <sys/malloc.h> 38219888Sed#include <sys/mount.h> 39219888Sed#include <sys/mutex.h> 40219888Sed#include <sys/kernel.h> 41219888Sed#include <sys/kthread.h> 42219888Sed#include <sys/ktr.h> 43219888Sed#include <sys/proc.h> 44219888Sed#include <sys/reboot.h> 45219888Sed#include <sys/resourcevar.h> 46219888Sed#include <sys/sysctl.h> 47219888Sed#include <sys/vmmeter.h> 48219888Sed#include <sys/vnode.h> 49219888Sed#include <vm/vm.h> 50219888Sed#include <vm/vm_param.h> 51219888Sed#include <vm/vm_kern.h> 52219888Sed#include <vm/vm_pageout.h> 53219888Sed#include <vm/vm_page.h> 54219888Sed#include <vm/vm_object.h> 55219888Sed#include <vm/vm_extern.h> 56219888Sed#include <vm/vm_map.h> 57219888Sed 58219888Sedstatic MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 59219888Sed 60219888Sedstruct bio_ops bioops; /* I/O operation notification */ 61219888Sed 62219888Sedstruct buf_ops buf_ops_bio = { 63219888Sed "buf_ops_bio", 64219888Sed bwrite 65219888Sed}; 66219888Sed 67219888Sed/* 68219888Sed * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 69219888Sed * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 70219888Sed */ 71219888Sedstruct buf *buf; /* buffer header pool */ 72219888Sedstruct mtx buftimelock; /* Interlock on setting prio and timo */ 73219888Sed 74219888Sedstatic void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 75219888Sed vm_offset_t to); 76219888Sedstatic void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 77219888Sed vm_offset_t to); 78219888Sedstatic void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 79219888Sed int pageno, vm_page_t m); 80219888Sedstatic void vfs_clean_pages(struct buf * bp); 81219888Sedstatic void vfs_setdirty(struct buf *bp); 82219888Sedstatic void vfs_vmio_release(struct buf *bp); 83219888Sedstatic void vfs_backgroundwritedone(struct buf *bp); 84219888Sedstatic int flushbufqueues(void); 85219888Sedstatic void buf_daemon(void); 86219888Sed 87219888Sedint vmiodirenable = TRUE; 88219888SedSYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 89219888Sed "Use the VM system for directory writes"); 90219888Sedint runningbufspace; 91219888SedSYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 92219888Sed "Amount of presently outstanding async buffer io"); 93219888Sedstatic int bufspace; 94219888SedSYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 95219888Sed "KVA memory used for bufs"); 96219888Sedstatic int maxbufspace; 97219888SedSYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 98219888Sed "Maximum allowed value of bufspace (including buf_daemon)"); 99219888Sedstatic int bufmallocspace; 100219888SedSYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 101219888Sed "Amount of malloced memory for buffers"); 102219888Sedstatic int maxbufmallocspace; 103219888SedSYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 104219888Sed "Maximum amount of malloced memory for buffers"); 105219888Sedstatic int lobufspace; 106219888SedSYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 107219888Sed "Minimum amount of buffers we want to have"); 108219888Sedstatic int hibufspace; 109219888SedSYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 110219888Sed "Maximum allowed value of bufspace (excluding buf_daemon)"); 111219888Sedstatic int bufreusecnt; 112219888SedSYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 113219888Sed "Number of times we have reused a buffer"); 114219888Sedstatic int buffreekvacnt; 115219888SedSYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 116219888Sed "Number of times we have freed the KVA space from some buffer"); 117219888Sedstatic int bufdefragcnt; 118219888SedSYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 119219888Sed "Number of times we have had to repeat buffer allocation to defragment"); 120219888Sedstatic int lorunningspace; 121219888SedSYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 122219888Sed "Minimum preferred space used for in-progress I/O"); 123219888Sedstatic int hirunningspace; 124219888SedSYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 125219888Sed "Maximum amount of space to use for in-progress I/O"); 126219888Sedstatic int numdirtybuffers; 127219888SedSYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 128219888Sed "Number of buffers that are dirty (has unwritten changes) at the moment"); 129219888Sedstatic int lodirtybuffers; 130219888SedSYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 131219888Sed "How many buffers we want to have free before bufdaemon can sleep"); 132219888Sedstatic int hidirtybuffers; 133219888SedSYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 134219888Sed "When the number of dirty buffers is considered severe"); 135219888Sedstatic int numfreebuffers; 136219888SedSYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 137219888Sed "Number of free buffers"); 138219888Sedstatic int lofreebuffers; 139219888SedSYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 140219888Sed "XXX Unused"); 141219888Sedstatic int hifreebuffers; 142219888SedSYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 143219888Sed "XXX Complicatedly unused"); 144219888Sedstatic int getnewbufcalls; 145219888SedSYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 146219888Sed "Number of calls to getnewbuf"); 147219888Sedstatic int getnewbufrestarts; 148219888SedSYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 149219888Sed "Number of times getnewbuf has had to restart a buffer aquisition"); 150219888Sedstatic int dobkgrdwrite = 1; 151219888SedSYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 152219888Sed "Do background writes (honoring the BX_BKGRDWRITE flag)?"); 153219888Sed 154219888Sed/* 155219888Sed * Wakeup point for bufdaemon, as well as indicator of whether it is already 156219888Sed * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 157219888Sed * is idling. 158219888Sed */ 159219888Sedstatic int bd_request; 160219888Sed 161219888Sed/* 162219888Sed * bogus page -- for I/O to/from partially complete buffers 163219888Sed * this is a temporary solution to the problem, but it is not 164219888Sed * really that bad. it would be better to split the buffer 165219888Sed * for input in the case of buffers partially already in memory, 166219888Sed * but the code is intricate enough already. 167219888Sed */ 168219888Sedvm_page_t bogus_page; 169219888Sed 170219888Sed/* 171219888Sed * Offset for bogus_page. 172219888Sed * XXX bogus_offset should be local to bufinit 173219888Sed */ 174219888Sedstatic vm_offset_t bogus_offset; 175219888Sed 176219888Sed/* 177219888Sed * Synchronization (sleep/wakeup) variable for active buffer space requests. 178219888Sed * Set when wait starts, cleared prior to wakeup(). 179219888Sed * Used in runningbufwakeup() and waitrunningbufspace(). 180219888Sed */ 181219888Sedstatic int runningbufreq; 182219888Sed 183219888Sed/* 184219888Sed * Synchronization (sleep/wakeup) variable for buffer requests. 185219888Sed * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 186219888Sed * by and/or. 187219888Sed * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 188219888Sed * getnewbuf(), and getblk(). 189219888Sed */ 190219888Sedstatic int needsbuffer; 191219888Sed 192219888Sed/* 193219888Sed * Mask for index into the buffer hash table, which needs to be power of 2 in 194219888Sed * size. Set in kern_vfs_bio_buffer_alloc. 195219888Sed */ 196219888Sedstatic int bufhashmask; 197219888Sed 198219888Sed/* 199219888Sed * Hash table for all buffers, with a linked list hanging from each table 200219888Sed * entry. Set in kern_vfs_bio_buffer_alloc, initialized in buf_init. 201219888Sed */ 202219888Sedstatic LIST_HEAD(bufhashhdr, buf) *bufhashtbl; 203219888Sed 204219888Sed/* 205219888Sed * Somewhere to store buffers when they are not in another list, to always 206219888Sed * have them in a list (and thus being able to use the same set of operations 207219888Sed * on them.) 208219888Sed */ 209219888Sedstatic struct bufhashhdr invalhash; 210219888Sed 211219888Sed/* 212219888Sed * Definitions for the buffer free lists. 213219888Sed */ 214219888Sed#define BUFFER_QUEUES 6 /* number of free buffer queues */ 215219888Sed 216219888Sed#define QUEUE_NONE 0 /* on no queue */ 217219888Sed#define QUEUE_LOCKED 1 /* locked buffers */ 218219888Sed#define QUEUE_CLEAN 2 /* non-B_DELWRI buffers */ 219219888Sed#define QUEUE_DIRTY 3 /* B_DELWRI buffers */ 220219888Sed#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 221219888Sed#define QUEUE_EMPTY 5 /* empty buffer headers */ 222219888Sed 223219888Sed/* Queues for free buffers with various properties */ 224219888Sedstatic TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 225219888Sed/* 226219888Sed * Single global constant for BUF_WMESG, to avoid getting multiple references. 227219888Sed * buf_wmesg is referred from macros. 228219888Sed */ 229219888Sedconst char *buf_wmesg = BUF_WMESG; 230219888Sed 231219888Sed#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 232219888Sed#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 233219888Sed#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 234219888Sed#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 235219888Sed 236219888Sed/* 237219888Sed * Buffer hash table code. Note that the logical block scans linearly, which 238219888Sed * gives us some L1 cache locality. 239219888Sed */ 240219888Sed 241219888Sedstatic __inline 242219888Sedstruct bufhashhdr * 243219888Sedbufhash(struct vnode *vnp, daddr_t bn) 244219888Sed{ 245219888Sed return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]); 246219888Sed} 247219888Sed 248219888Sed/* 249219888Sed * numdirtywakeup: 250219888Sed * 251219888Sed * If someone is blocked due to there being too many dirty buffers, 252219888Sed * and numdirtybuffers is now reasonable, wake them up. 253219888Sed */ 254219888Sed 255219888Sedstatic __inline void 256219888Sednumdirtywakeup(int level) 257219888Sed{ 258219888Sed if (numdirtybuffers <= level) { 259219888Sed if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 260219888Sed needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 261219888Sed wakeup(&needsbuffer); 262219888Sed } 263219888Sed } 264219888Sed} 265219888Sed 266219888Sed/* 267219888Sed * bufspacewakeup: 268219888Sed * 269219888Sed * Called when buffer space is potentially available for recovery. 270219888Sed * getnewbuf() will block on this flag when it is unable to free 271219888Sed * sufficient buffer space. Buffer space becomes recoverable when 272219888Sed * bp's get placed back in the queues. 273219888Sed */ 274219888Sed 275219888Sedstatic __inline void 276219888Sedbufspacewakeup(void) 277219888Sed{ 278219888Sed /* 279219888Sed * If someone is waiting for BUF space, wake them up. Even 280219888Sed * though we haven't freed the kva space yet, the waiting 281219888Sed * process will be able to now. 282219888Sed */ 283219888Sed if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 284219888Sed needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 285219888Sed wakeup(&needsbuffer); 286219888Sed } 287219888Sed} 288219888Sed 289219888Sed/* 290219888Sed * runningbufwakeup() - in-progress I/O accounting. 291219888Sed * 292219888Sed */ 293219888Sedstatic __inline void 294219888Sedrunningbufwakeup(struct buf *bp) 295219888Sed{ 296219888Sed if (bp->b_runningbufspace) { 297219888Sed runningbufspace -= bp->b_runningbufspace; 298219888Sed bp->b_runningbufspace = 0; 299219888Sed if (runningbufreq && runningbufspace <= lorunningspace) { 300219888Sed runningbufreq = 0; 301219888Sed wakeup(&runningbufreq); 302219888Sed } 303219888Sed } 304219888Sed} 305219888Sed 306219888Sed/* 307219888Sed * bufcountwakeup: 308219888Sed * 309219888Sed * Called when a buffer has been added to one of the free queues to 310219888Sed * account for the buffer and to wakeup anyone waiting for free buffers. 311219888Sed * This typically occurs when large amounts of metadata are being handled 312219888Sed * by the buffer cache ( else buffer space runs out first, usually ). 313219888Sed */ 314219888Sed 315219888Sedstatic __inline void 316219888Sedbufcountwakeup(void) 317219888Sed{ 318219888Sed ++numfreebuffers; 319219888Sed if (needsbuffer) { 320219888Sed needsbuffer &= ~VFS_BIO_NEED_ANY; 321219888Sed if (numfreebuffers >= hifreebuffers) 322219888Sed needsbuffer &= ~VFS_BIO_NEED_FREE; 323219888Sed wakeup(&needsbuffer); 324219888Sed } 325219888Sed} 326219888Sed 327219888Sed/* 328219888Sed * waitrunningbufspace() 329219888Sed * 330219888Sed * runningbufspace is a measure of the amount of I/O currently 331219888Sed * running. This routine is used in async-write situations to 332219888Sed * prevent creating huge backups of pending writes to a device. 333219888Sed * Only asynchronous writes are governed by this function. 334219888Sed * 335219888Sed * Reads will adjust runningbufspace, but will not block based on it. 336219888Sed * The read load has a side effect of reducing the allowed write load. 337219888Sed * 338219888Sed * This does NOT turn an async write into a sync write. It waits 339219888Sed * for earlier writes to complete and generally returns before the 340219888Sed * caller's write has reached the device. 341219888Sed */ 342219888Sedstatic __inline void 343219888Sedwaitrunningbufspace(void) 344219888Sed{ 345219888Sed /* 346219888Sed * XXX race against wakeup interrupt, currently 347219888Sed * protected by Giant. FIXME! 348219888Sed */ 349219888Sed while (runningbufspace > hirunningspace) { 350219888Sed ++runningbufreq; 351219888Sed tsleep(&runningbufreq, PVM, "wdrain", 0); 352219888Sed } 353219888Sed} 354219888Sed 355219888Sed 356219888Sed/* 357219888Sed * vfs_buf_test_cache: 358219888Sed * 359219888Sed * Called when a buffer is extended. This function clears the B_CACHE 360219888Sed * bit if the newly extended portion of the buffer does not contain 361219888Sed * valid data. 362219888Sed */ 363219888Sedstatic __inline__ 364219888Sedvoid 365219888Sedvfs_buf_test_cache(struct buf *bp, 366219888Sed vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 367219888Sed vm_page_t m) 368219888Sed{ 369219888Sed GIANT_REQUIRED; 370219888Sed 371219888Sed if (bp->b_flags & B_CACHE) { 372219888Sed int base = (foff + off) & PAGE_MASK; 373219888Sed if (vm_page_is_valid(m, base, size) == 0) 374219888Sed bp->b_flags &= ~B_CACHE; 375219888Sed } 376219888Sed} 377219888Sed 378219888Sed/* Wake up the buffer deamon if necessary */ 379219888Sedstatic __inline__ 380219888Sedvoid 381219888Sedbd_wakeup(int dirtybuflevel) 382219888Sed{ 383219888Sed if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 384219888Sed bd_request = 1; 385219888Sed wakeup(&bd_request); 386219888Sed } 387219888Sed} 388219888Sed 389219888Sed/* 390219888Sed * bd_speedup - speedup the buffer cache flushing code 391219888Sed */ 392219888Sed 393219888Sedstatic __inline__ 394219888Sedvoid 395219888Sedbd_speedup(void) 396219888Sed{ 397219888Sed bd_wakeup(1); 398219888Sed} 399219888Sed 400219888Sed/* 401219888Sed * Calculating buffer cache scaling values and reserve space for buffer 402219888Sed * headers. This is called during low level kernel initialization and 403219888Sed * may be called more then once. We CANNOT write to the memory area 404219888Sed * being reserved at this time. 405219888Sed */ 406219888Sedcaddr_t 407219888Sedkern_vfs_bio_buffer_alloc(caddr_t v, int physmem_est) 408219888Sed{ 409219888Sed /* 410219888Sed * physmem_est is in pages. Convert it to kilobytes (assumes 411219888Sed * PAGE_SIZE is >= 1K) 412219888Sed */ 413219888Sed physmem_est = physmem_est * (PAGE_SIZE / 1024); 414219888Sed 415219888Sed /* 416219888Sed * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 417219888Sed * For the first 64MB of ram nominally allocate sufficient buffers to 418219888Sed * cover 1/4 of our ram. Beyond the first 64MB allocate additional 419219888Sed * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 420219888Sed * the buffer cache we limit the eventual kva reservation to 421219888Sed * maxbcache bytes. 422219888Sed * 423219888Sed * factor represents the 1/4 x ram conversion. 424219888Sed */ 425219888Sed if (nbuf == 0) { 426219888Sed int factor = 4 * BKVASIZE / 1024; 427219888Sed 428219888Sed nbuf = 50; 429219888Sed if (physmem_est > 4096) 430219888Sed nbuf += min((physmem_est - 4096) / factor, 431219888Sed 65536 / factor); 432219888Sed if (physmem_est > 65536) 433219888Sed nbuf += (physmem_est - 65536) * 2 / (factor * 5); 434219888Sed 435219888Sed if (maxbcache && nbuf > maxbcache / BKVASIZE) 436219888Sed nbuf = maxbcache / BKVASIZE; 437219888Sed } 438219888Sed 439219888Sed#if 0 440219888Sed /* 441219888Sed * Do not allow the buffer_map to be more then 1/2 the size of the 442219888Sed * kernel_map. 443219888Sed */ 444219888Sed if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 445219888Sed (BKVASIZE * 2)) { 446219888Sed nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 447219888Sed (BKVASIZE * 2); 448219888Sed printf("Warning: nbufs capped at %d\n", nbuf); 449219888Sed } 450219888Sed#endif 451219888Sed 452219888Sed /* 453219888Sed * swbufs are used as temporary holders for I/O, such as paging I/O. 454219888Sed * We have no less then 16 and no more then 256. 455219888Sed */ 456219888Sed nswbuf = max(min(nbuf/4, 256), 16); 457219888Sed 458219888Sed /* 459219888Sed * Reserve space for the buffer cache buffers 460219888Sed */ 461219888Sed swbuf = (void *)v; 462219888Sed v = (caddr_t)(swbuf + nswbuf); 463219888Sed buf = (void *)v; 464219888Sed v = (caddr_t)(buf + nbuf); 465219888Sed 466219888Sed /* 467219888Sed * Calculate the hash table size and reserve space 468219888Sed */ 469219888Sed for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1) 470219888Sed ; 471219888Sed bufhashtbl = (void *)v; 472219888Sed v = (caddr_t)(bufhashtbl + bufhashmask); 473219888Sed --bufhashmask; 474219888Sed 475219888Sed return(v); 476219888Sed} 477219888Sed 478219888Sed/* Initialize the buffer subsystem. Called before use of any buffers. */ 479219888Sedvoid 480219888Sedbufinit(void) 481219888Sed{ 482219888Sed struct buf *bp; 483219888Sed int i; 484219888Sed 485219888Sed GIANT_REQUIRED; 486219888Sed 487219888Sed LIST_INIT(&invalhash); 488219888Sed mtx_init(&buftimelock, "buftime lock", NULL, MTX_DEF); 489219888Sed 490219888Sed for (i = 0; i <= bufhashmask; i++) 491219888Sed LIST_INIT(&bufhashtbl[i]); 492219888Sed 493219888Sed /* next, make a null set of free lists */ 494219888Sed for (i = 0; i < BUFFER_QUEUES; i++) 495219888Sed TAILQ_INIT(&bufqueues[i]); 496219888Sed 497219888Sed /* finally, initialize each buffer header and stick on empty q */ 498219888Sed for (i = 0; i < nbuf; i++) { 499219888Sed bp = &buf[i]; 500219888Sed bzero(bp, sizeof *bp); 501219888Sed bp->b_flags = B_INVAL; /* we're just an empty header */ 502219888Sed bp->b_dev = NODEV; 503219888Sed bp->b_rcred = NOCRED; 504219888Sed bp->b_wcred = NOCRED; 505219888Sed bp->b_qindex = QUEUE_EMPTY; 506219888Sed bp->b_xflags = 0; 507219888Sed LIST_INIT(&bp->b_dep); 508219888Sed BUF_LOCKINIT(bp); 509219888Sed TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 510219888Sed LIST_INSERT_HEAD(&invalhash, bp, b_hash); 511219888Sed } 512219888Sed 513219888Sed /* 514219888Sed * maxbufspace is the absolute maximum amount of buffer space we are 515219888Sed * allowed to reserve in KVM and in real terms. The absolute maximum 516219888Sed * is nominally used by buf_daemon. hibufspace is the nominal maximum 517219888Sed * used by most other processes. The differential is required to 518219888Sed * ensure that buf_daemon is able to run when other processes might 519219888Sed * be blocked waiting for buffer space. 520219888Sed * 521219888Sed * maxbufspace is based on BKVASIZE. Allocating buffers larger then 522219888Sed * this may result in KVM fragmentation which is not handled optimally 523219888Sed * by the system. 524219888Sed */ 525219888Sed maxbufspace = nbuf * BKVASIZE; 526219888Sed hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 527219888Sed lobufspace = hibufspace - MAXBSIZE; 528219888Sed 529219888Sed lorunningspace = 512 * 1024; 530219888Sed hirunningspace = 1024 * 1024; 531219888Sed 532219888Sed/* 533219888Sed * Limit the amount of malloc memory since it is wired permanently into 534219888Sed * the kernel space. Even though this is accounted for in the buffer 535219888Sed * allocation, we don't want the malloced region to grow uncontrolled. 536219888Sed * The malloc scheme improves memory utilization significantly on average 537219888Sed * (small) directories. 538219888Sed */ 539219888Sed maxbufmallocspace = hibufspace / 20; 540219888Sed 541219888Sed/* 542219888Sed * Reduce the chance of a deadlock occuring by limiting the number 543219888Sed * of delayed-write dirty buffers we allow to stack up. 544219888Sed */ 545219888Sed hidirtybuffers = nbuf / 4 + 20; 546219888Sed numdirtybuffers = 0; 547219888Sed/* 548219888Sed * To support extreme low-memory systems, make sure hidirtybuffers cannot 549219888Sed * eat up all available buffer space. This occurs when our minimum cannot 550219888Sed * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 551219888Sed * BKVASIZE'd (8K) buffers. 552219888Sed */ 553219888Sed while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 554219888Sed hidirtybuffers >>= 1; 555219888Sed } 556219888Sed lodirtybuffers = hidirtybuffers / 2; 557219888Sed 558219888Sed/* 559219888Sed * Try to keep the number of free buffers in the specified range, 560219888Sed * and give special processes (e.g. like buf_daemon) access to an 561219888Sed * emergency reserve. 562219888Sed */ 563219888Sed lofreebuffers = nbuf / 18 + 5; 564219888Sed hifreebuffers = 2 * lofreebuffers; 565219888Sed numfreebuffers = nbuf; 566219888Sed 567219888Sed/* 568219888Sed * Maximum number of async ops initiated per buf_daemon loop. This is 569219888Sed * somewhat of a hack at the moment, we really need to limit ourselves 570219888Sed * based on the number of bytes of I/O in-transit that were initiated 571219888Sed * from buf_daemon. 572219888Sed */ 573219888Sed 574219888Sed bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 575219888Sed bogus_page = vm_page_alloc(kernel_object, 576219888Sed ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 577219888Sed VM_ALLOC_NORMAL); 578219888Sed cnt.v_wire_count++; 579219888Sed} 580219888Sed 581219888Sed/* 582219888Sed * bfreekva() - free the kva allocation for a buffer. 583219888Sed * 584219888Sed * Must be called at splbio() or higher as this is the only locking for 585219888Sed * buffer_map. 586219888Sed * 587219888Sed * Since this call frees up buffer space, we call bufspacewakeup(). 588219888Sed */ 589219888Sedstatic void 590219888Sedbfreekva(struct buf * bp) 591219888Sed{ 592219888Sed GIANT_REQUIRED; 593219888Sed 594219888Sed if (bp->b_kvasize) { 595219888Sed ++buffreekvacnt; 596219888Sed bufspace -= bp->b_kvasize; 597219888Sed vm_map_delete(buffer_map, 598219888Sed (vm_offset_t) bp->b_kvabase, 599219888Sed (vm_offset_t) bp->b_kvabase + bp->b_kvasize 600219888Sed ); 601219888Sed bp->b_kvasize = 0; 602219888Sed bufspacewakeup(); 603219888Sed } 604219888Sed} 605219888Sed 606219888Sed/* 607219888Sed * bremfree: 608219888Sed * 609219888Sed * Remove the buffer from the appropriate free list. 610219888Sed */ 611219888Sedvoid 612219888Sedbremfree(struct buf * bp) 613219888Sed{ 614219888Sed int s = splbio(); 615219888Sed int old_qindex = bp->b_qindex; 616219888Sed 617219888Sed GIANT_REQUIRED; 618219888Sed 619219888Sed if (bp->b_qindex != QUEUE_NONE) { 620219888Sed KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 621219888Sed TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 622219888Sed bp->b_qindex = QUEUE_NONE; 623219888Sed } else { 624219888Sed if (BUF_REFCNT(bp) <= 1) 625219888Sed panic("bremfree: removing a buffer not on a queue"); 626219888Sed } 627219888Sed 628219888Sed /* 629219888Sed * Fixup numfreebuffers count. If the buffer is invalid or not 630219888Sed * delayed-write, and it was on the EMPTY, LRU, or AGE queues, 631219888Sed * the buffer was free and we must decrement numfreebuffers. 632219888Sed */ 633219888Sed if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 634219888Sed switch(old_qindex) { 635219888Sed case QUEUE_DIRTY: 636219888Sed case QUEUE_CLEAN: 637219888Sed case QUEUE_EMPTY: 638219888Sed case QUEUE_EMPTYKVA: 639219888Sed --numfreebuffers; 640219888Sed break; 641219888Sed default: 642 break; 643 } 644 } 645 splx(s); 646} 647 648 649/* 650 * Get a buffer with the specified data. Look in the cache first. We 651 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 652 * is set, the buffer is valid and we do not have to do anything ( see 653 * getblk() ). This is really just a special case of breadn(). 654 */ 655int 656bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 657 struct buf ** bpp) 658{ 659 660 return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 661} 662 663/* 664 * Operates like bread, but also starts asynchronous I/O on 665 * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 666 * to initiating I/O . If B_CACHE is set, the buffer is valid 667 * and we do not have to do anything. 668 */ 669int 670breadn(struct vnode * vp, daddr_t blkno, int size, 671 daddr_t * rablkno, int *rabsize, 672 int cnt, struct ucred * cred, struct buf ** bpp) 673{ 674 struct buf *bp, *rabp; 675 int i; 676 int rv = 0, readwait = 0; 677 678 *bpp = bp = getblk(vp, blkno, size, 0, 0); 679 680 /* if not found in cache, do some I/O */ 681 if ((bp->b_flags & B_CACHE) == 0) { 682 if (curthread != PCPU_GET(idlethread)) 683 curthread->td_proc->p_stats->p_ru.ru_inblock++; 684 bp->b_iocmd = BIO_READ; 685 bp->b_flags &= ~B_INVAL; 686 bp->b_ioflags &= ~BIO_ERROR; 687 if (bp->b_rcred == NOCRED && cred != NOCRED) 688 bp->b_rcred = crhold(cred); 689 vfs_busy_pages(bp, 0); 690 VOP_STRATEGY(vp, bp); 691 ++readwait; 692 } 693 694 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 695 if (inmem(vp, *rablkno)) 696 continue; 697 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 698 699 if ((rabp->b_flags & B_CACHE) == 0) { 700 if (curthread != PCPU_GET(idlethread)) 701 curthread->td_proc->p_stats->p_ru.ru_inblock++; 702 rabp->b_flags |= B_ASYNC; 703 rabp->b_flags &= ~B_INVAL; 704 rabp->b_ioflags &= ~BIO_ERROR; 705 rabp->b_iocmd = BIO_READ; 706 if (rabp->b_rcred == NOCRED && cred != NOCRED) 707 rabp->b_rcred = crhold(cred); 708 vfs_busy_pages(rabp, 0); 709 BUF_KERNPROC(rabp); 710 VOP_STRATEGY(vp, rabp); 711 } else { 712 brelse(rabp); 713 } 714 } 715 716 if (readwait) { 717 rv = bufwait(bp); 718 } 719 return (rv); 720} 721 722/* 723 * Write, release buffer on completion. (Done by iodone 724 * if async). Do not bother writing anything if the buffer 725 * is invalid. 726 * 727 * Note that we set B_CACHE here, indicating that buffer is 728 * fully valid and thus cacheable. This is true even of NFS 729 * now so we set it generally. This could be set either here 730 * or in biodone() since the I/O is synchronous. We put it 731 * here. 732 */ 733 734int 735bwrite(struct buf * bp) 736{ 737 int oldflags, s; 738 struct buf *newbp; 739 740 if (bp->b_flags & B_INVAL) { 741 brelse(bp); 742 return (0); 743 } 744 745 oldflags = bp->b_flags; 746 747 if (BUF_REFCNT(bp) == 0) 748 panic("bwrite: buffer is not busy???"); 749 s = splbio(); 750 /* 751 * If a background write is already in progress, delay 752 * writing this block if it is asynchronous. Otherwise 753 * wait for the background write to complete. 754 */ 755 if (bp->b_xflags & BX_BKGRDINPROG) { 756 if (bp->b_flags & B_ASYNC) { 757 splx(s); 758 bdwrite(bp); 759 return (0); 760 } 761 bp->b_xflags |= BX_BKGRDWAIT; 762 tsleep(&bp->b_xflags, PRIBIO, "bwrbg", 0); 763 if (bp->b_xflags & BX_BKGRDINPROG) 764 panic("bwrite: still writing"); 765 } 766 767 /* Mark the buffer clean */ 768 bundirty(bp); 769 770 /* 771 * If this buffer is marked for background writing and we 772 * do not have to wait for it, make a copy and write the 773 * copy so as to leave this buffer ready for further use. 774 * 775 * This optimization eats a lot of memory. If we have a page 776 * or buffer shortfall we can't do it. 777 */ 778 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 779 (bp->b_flags & B_ASYNC) && 780 !vm_page_count_severe() && 781 !buf_dirty_count_severe()) { 782 if (bp->b_iodone != NULL) { 783 printf("bp->b_iodone = %p\n", bp->b_iodone); 784 panic("bwrite: need chained iodone"); 785 } 786 787 /* get a new block */ 788 newbp = geteblk(bp->b_bufsize); 789 790 /* set it to be identical to the old block */ 791 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 792 bgetvp(bp->b_vp, newbp); 793 newbp->b_lblkno = bp->b_lblkno; 794 newbp->b_blkno = bp->b_blkno; 795 newbp->b_offset = bp->b_offset; 796 newbp->b_iodone = vfs_backgroundwritedone; 797 newbp->b_flags |= B_ASYNC; 798 newbp->b_flags &= ~B_INVAL; 799 800 /* move over the dependencies */ 801 if (LIST_FIRST(&bp->b_dep) != NULL) 802 buf_movedeps(bp, newbp); 803 804 /* 805 * Initiate write on the copy, release the original to 806 * the B_LOCKED queue so that it cannot go away until 807 * the background write completes. If not locked it could go 808 * away and then be reconstituted while it was being written. 809 * If the reconstituted buffer were written, we could end up 810 * with two background copies being written at the same time. 811 */ 812 bp->b_xflags |= BX_BKGRDINPROG; 813 bp->b_flags |= B_LOCKED; 814 bqrelse(bp); 815 bp = newbp; 816 } 817 818 bp->b_flags &= ~B_DONE; 819 bp->b_ioflags &= ~BIO_ERROR; 820 bp->b_flags |= B_WRITEINPROG | B_CACHE; 821 bp->b_iocmd = BIO_WRITE; 822 823 bp->b_vp->v_numoutput++; 824 vfs_busy_pages(bp, 1); 825 826 /* 827 * Normal bwrites pipeline writes 828 */ 829 bp->b_runningbufspace = bp->b_bufsize; 830 runningbufspace += bp->b_runningbufspace; 831 832 if (curthread != PCPU_GET(idlethread)) 833 curthread->td_proc->p_stats->p_ru.ru_oublock++; 834 splx(s); 835 if (oldflags & B_ASYNC) 836 BUF_KERNPROC(bp); 837 BUF_STRATEGY(bp); 838 839 if ((oldflags & B_ASYNC) == 0) { 840 int rtval = bufwait(bp); 841 brelse(bp); 842 return (rtval); 843 } else if ((oldflags & B_NOWDRAIN) == 0) { 844 /* 845 * don't allow the async write to saturate the I/O 846 * system. Deadlocks can occur only if a device strategy 847 * routine (like in MD) turns around and issues another 848 * high-level write, in which case B_NOWDRAIN is expected 849 * to be set. Otherwise we will not deadlock here because 850 * we are blocking waiting for I/O that is already in-progress 851 * to complete. 852 */ 853 waitrunningbufspace(); 854 } 855 856 return (0); 857} 858 859/* 860 * Complete a background write started from bwrite. 861 */ 862static void 863vfs_backgroundwritedone(bp) 864 struct buf *bp; 865{ 866 struct buf *origbp; 867 868 /* 869 * Find the original buffer that we are writing. 870 */ 871 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL) 872 panic("backgroundwritedone: lost buffer"); 873 /* 874 * Process dependencies then return any unfinished ones. 875 */ 876 if (LIST_FIRST(&bp->b_dep) != NULL) 877 buf_complete(bp); 878 if (LIST_FIRST(&bp->b_dep) != NULL) 879 buf_movedeps(bp, origbp); 880 /* 881 * Clear the BX_BKGRDINPROG flag in the original buffer 882 * and awaken it if it is waiting for the write to complete. 883 * If BX_BKGRDINPROG is not set in the original buffer it must 884 * have been released and re-instantiated - which is not legal. 885 */ 886 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), 887 ("backgroundwritedone: lost buffer2")); 888 origbp->b_xflags &= ~BX_BKGRDINPROG; 889 if (origbp->b_xflags & BX_BKGRDWAIT) { 890 origbp->b_xflags &= ~BX_BKGRDWAIT; 891 wakeup(&origbp->b_xflags); 892 } 893 /* 894 * Clear the B_LOCKED flag and remove it from the locked 895 * queue if it currently resides there. 896 */ 897 origbp->b_flags &= ~B_LOCKED; 898 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 899 bremfree(origbp); 900 bqrelse(origbp); 901 } 902 /* 903 * This buffer is marked B_NOCACHE, so when it is released 904 * by biodone, it will be tossed. We mark it with BIO_READ 905 * to avoid biodone doing a second vwakeup. 906 */ 907 bp->b_flags |= B_NOCACHE; 908 bp->b_iocmd = BIO_READ; 909 bp->b_flags &= ~(B_CACHE | B_DONE); 910 bp->b_iodone = 0; 911 bufdone(bp); 912} 913 914/* 915 * Delayed write. (Buffer is marked dirty). Do not bother writing 916 * anything if the buffer is marked invalid. 917 * 918 * Note that since the buffer must be completely valid, we can safely 919 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 920 * biodone() in order to prevent getblk from writing the buffer 921 * out synchronously. 922 */ 923void 924bdwrite(struct buf * bp) 925{ 926 GIANT_REQUIRED; 927 928 if (BUF_REFCNT(bp) == 0) 929 panic("bdwrite: buffer is not busy"); 930 931 if (bp->b_flags & B_INVAL) { 932 brelse(bp); 933 return; 934 } 935 bdirty(bp); 936 937 /* 938 * Set B_CACHE, indicating that the buffer is fully valid. This is 939 * true even of NFS now. 940 */ 941 bp->b_flags |= B_CACHE; 942 943 /* 944 * This bmap keeps the system from needing to do the bmap later, 945 * perhaps when the system is attempting to do a sync. Since it 946 * is likely that the indirect block -- or whatever other datastructure 947 * that the filesystem needs is still in memory now, it is a good 948 * thing to do this. Note also, that if the pageout daemon is 949 * requesting a sync -- there might not be enough memory to do 950 * the bmap then... So, this is important to do. 951 */ 952 if (bp->b_lblkno == bp->b_blkno) { 953 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 954 } 955 956 /* 957 * Set the *dirty* buffer range based upon the VM system dirty pages. 958 */ 959 vfs_setdirty(bp); 960 961 /* 962 * We need to do this here to satisfy the vnode_pager and the 963 * pageout daemon, so that it thinks that the pages have been 964 * "cleaned". Note that since the pages are in a delayed write 965 * buffer -- the VFS layer "will" see that the pages get written 966 * out on the next sync, or perhaps the cluster will be completed. 967 */ 968 vfs_clean_pages(bp); 969 bqrelse(bp); 970 971 /* 972 * Wakeup the buffer flushing daemon if we have a lot of dirty 973 * buffers (midpoint between our recovery point and our stall 974 * point). 975 */ 976 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 977 978 /* 979 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 980 * due to the softdep code. 981 */ 982} 983 984/* 985 * bdirty: 986 * 987 * Turn buffer into delayed write request. We must clear BIO_READ and 988 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 989 * itself to properly update it in the dirty/clean lists. We mark it 990 * B_DONE to ensure that any asynchronization of the buffer properly 991 * clears B_DONE ( else a panic will occur later ). 992 * 993 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 994 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 995 * should only be called if the buffer is known-good. 996 * 997 * Since the buffer is not on a queue, we do not update the numfreebuffers 998 * count. 999 * 1000 * Must be called at splbio(). 1001 * The buffer must be on QUEUE_NONE. 1002 */ 1003void 1004bdirty(bp) 1005 struct buf *bp; 1006{ 1007 KASSERT(bp->b_qindex == QUEUE_NONE, 1008 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1009 bp->b_flags &= ~(B_RELBUF); 1010 bp->b_iocmd = BIO_WRITE; 1011 1012 if ((bp->b_flags & B_DELWRI) == 0) { 1013 bp->b_flags |= B_DONE | B_DELWRI; 1014 reassignbuf(bp, bp->b_vp); 1015 ++numdirtybuffers; 1016 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1017 } 1018} 1019 1020/* 1021 * bundirty: 1022 * 1023 * Clear B_DELWRI for buffer. 1024 * 1025 * Since the buffer is not on a queue, we do not update the numfreebuffers 1026 * count. 1027 * 1028 * Must be called at splbio(). 1029 * The buffer must be on QUEUE_NONE. 1030 */ 1031 1032void 1033bundirty(bp) 1034 struct buf *bp; 1035{ 1036 KASSERT(bp->b_qindex == QUEUE_NONE, 1037 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1038 1039 if (bp->b_flags & B_DELWRI) { 1040 bp->b_flags &= ~B_DELWRI; 1041 reassignbuf(bp, bp->b_vp); 1042 --numdirtybuffers; 1043 numdirtywakeup(lodirtybuffers); 1044 } 1045 /* 1046 * Since it is now being written, we can clear its deferred write flag. 1047 */ 1048 bp->b_flags &= ~B_DEFERRED; 1049} 1050 1051/* 1052 * bawrite: 1053 * 1054 * Asynchronous write. Start output on a buffer, but do not wait for 1055 * it to complete. The buffer is released when the output completes. 1056 * 1057 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1058 * B_INVAL buffers. Not us. 1059 */ 1060void 1061bawrite(struct buf * bp) 1062{ 1063 bp->b_flags |= B_ASYNC; 1064 (void) BUF_WRITE(bp); 1065} 1066 1067/* 1068 * bwillwrite: 1069 * 1070 * Called prior to the locking of any vnodes when we are expecting to 1071 * write. We do not want to starve the buffer cache with too many 1072 * dirty buffers so we block here. By blocking prior to the locking 1073 * of any vnodes we attempt to avoid the situation where a locked vnode 1074 * prevents the various system daemons from flushing related buffers. 1075 */ 1076 1077void 1078bwillwrite(void) 1079{ 1080 if (numdirtybuffers >= hidirtybuffers) { 1081 int s; 1082 1083 mtx_lock(&Giant); 1084 s = splbio(); 1085 while (numdirtybuffers >= hidirtybuffers) { 1086 bd_wakeup(1); 1087 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1088 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0); 1089 } 1090 splx(s); 1091 mtx_unlock(&Giant); 1092 } 1093} 1094 1095/* 1096 * Return true if we have too many dirty buffers. 1097 */ 1098int 1099buf_dirty_count_severe(void) 1100{ 1101 return(numdirtybuffers >= hidirtybuffers); 1102} 1103 1104/* 1105 * brelse: 1106 * 1107 * Release a busy buffer and, if requested, free its resources. The 1108 * buffer will be stashed in the appropriate bufqueue[] allowing it 1109 * to be accessed later as a cache entity or reused for other purposes. 1110 */ 1111void 1112brelse(struct buf * bp) 1113{ 1114 int s; 1115 1116 GIANT_REQUIRED; 1117 1118 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1119 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1120 1121 s = splbio(); 1122 1123 if (bp->b_flags & B_LOCKED) 1124 bp->b_ioflags &= ~BIO_ERROR; 1125 1126 if (bp->b_iocmd == BIO_WRITE && 1127 (bp->b_ioflags & BIO_ERROR) && 1128 !(bp->b_flags & B_INVAL)) { 1129 /* 1130 * Failed write, redirty. Must clear BIO_ERROR to prevent 1131 * pages from being scrapped. If B_INVAL is set then 1132 * this case is not run and the next case is run to 1133 * destroy the buffer. B_INVAL can occur if the buffer 1134 * is outside the range supported by the underlying device. 1135 */ 1136 bp->b_ioflags &= ~BIO_ERROR; 1137 bdirty(bp); 1138 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1139 (bp->b_ioflags & BIO_ERROR) || 1140 bp->b_iocmd == BIO_DELETE || (bp->b_bufsize <= 0)) { 1141 /* 1142 * Either a failed I/O or we were asked to free or not 1143 * cache the buffer. 1144 */ 1145 bp->b_flags |= B_INVAL; 1146 if (LIST_FIRST(&bp->b_dep) != NULL) 1147 buf_deallocate(bp); 1148 if (bp->b_flags & B_DELWRI) { 1149 --numdirtybuffers; 1150 numdirtywakeup(lodirtybuffers); 1151 } 1152 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1153 if ((bp->b_flags & B_VMIO) == 0) { 1154 if (bp->b_bufsize) 1155 allocbuf(bp, 0); 1156 if (bp->b_vp) 1157 brelvp(bp); 1158 } 1159 } 1160 1161 /* 1162 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1163 * is called with B_DELWRI set, the underlying pages may wind up 1164 * getting freed causing a previous write (bdwrite()) to get 'lost' 1165 * because pages associated with a B_DELWRI bp are marked clean. 1166 * 1167 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1168 * if B_DELWRI is set. 1169 * 1170 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1171 * on pages to return pages to the VM page queues. 1172 */ 1173 if (bp->b_flags & B_DELWRI) 1174 bp->b_flags &= ~B_RELBUF; 1175 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG)) 1176 bp->b_flags |= B_RELBUF; 1177 1178 /* 1179 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1180 * constituted, not even NFS buffers now. Two flags effect this. If 1181 * B_INVAL, the struct buf is invalidated but the VM object is kept 1182 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1183 * 1184 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1185 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1186 * buffer is also B_INVAL because it hits the re-dirtying code above. 1187 * 1188 * Normally we can do this whether a buffer is B_DELWRI or not. If 1189 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1190 * the commit state and we cannot afford to lose the buffer. If the 1191 * buffer has a background write in progress, we need to keep it 1192 * around to prevent it from being reconstituted and starting a second 1193 * background write. 1194 */ 1195 if ((bp->b_flags & B_VMIO) 1196 && !(bp->b_vp->v_tag == VT_NFS && 1197 !vn_isdisk(bp->b_vp, NULL) && 1198 (bp->b_flags & B_DELWRI)) 1199 ) { 1200 1201 int i, j, resid; 1202 vm_page_t m; 1203 off_t foff; 1204 vm_pindex_t poff; 1205 vm_object_t obj; 1206 struct vnode *vp; 1207 1208 vp = bp->b_vp; 1209 1210 /* 1211 * Get the base offset and length of the buffer. Note that 1212 * in the VMIO case if the buffer block size is not 1213 * page-aligned then b_data pointer may not be page-aligned. 1214 * But our b_pages[] array *IS* page aligned. 1215 * 1216 * block sizes less then DEV_BSIZE (usually 512) are not 1217 * supported due to the page granularity bits (m->valid, 1218 * m->dirty, etc...). 1219 * 1220 * See man buf(9) for more information 1221 */ 1222 resid = bp->b_bufsize; 1223 foff = bp->b_offset; 1224 1225 for (i = 0; i < bp->b_npages; i++) { 1226 int had_bogus = 0; 1227 1228 m = bp->b_pages[i]; 1229 vm_page_flag_clear(m, PG_ZERO); 1230 1231 /* 1232 * If we hit a bogus page, fixup *all* the bogus pages 1233 * now. 1234 */ 1235 if (m == bogus_page) { 1236 VOP_GETVOBJECT(vp, &obj); 1237 poff = OFF_TO_IDX(bp->b_offset); 1238 had_bogus = 1; 1239 1240 for (j = i; j < bp->b_npages; j++) { 1241 vm_page_t mtmp; 1242 mtmp = bp->b_pages[j]; 1243 if (mtmp == bogus_page) { 1244 mtmp = vm_page_lookup(obj, poff + j); 1245 if (!mtmp) { 1246 panic("brelse: page missing\n"); 1247 } 1248 bp->b_pages[j] = mtmp; 1249 } 1250 } 1251 1252 if ((bp->b_flags & B_INVAL) == 0) { 1253 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1254 } 1255 m = bp->b_pages[i]; 1256 } 1257 if ((bp->b_flags & B_NOCACHE) || (bp->b_ioflags & BIO_ERROR)) { 1258 int poffset = foff & PAGE_MASK; 1259 int presid = resid > (PAGE_SIZE - poffset) ? 1260 (PAGE_SIZE - poffset) : resid; 1261 1262 KASSERT(presid >= 0, ("brelse: extra page")); 1263 vm_page_set_invalid(m, poffset, presid); 1264 if (had_bogus) 1265 printf("avoided corruption bug in bogus_page/brelse code\n"); 1266 } 1267 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1268 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1269 } 1270 1271 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1272 vfs_vmio_release(bp); 1273 1274 } else if (bp->b_flags & B_VMIO) { 1275 1276 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1277 vfs_vmio_release(bp); 1278 } 1279 1280 } 1281 1282 if (bp->b_qindex != QUEUE_NONE) 1283 panic("brelse: free buffer onto another queue???"); 1284 if (BUF_REFCNT(bp) > 1) { 1285 /* do not release to free list */ 1286 BUF_UNLOCK(bp); 1287 splx(s); 1288 return; 1289 } 1290 1291 /* enqueue */ 1292 1293 /* buffers with no memory */ 1294 if (bp->b_bufsize == 0) { 1295 bp->b_flags |= B_INVAL; 1296 bp->b_xflags &= ~BX_BKGRDWRITE; 1297 if (bp->b_xflags & BX_BKGRDINPROG) 1298 panic("losing buffer 1"); 1299 if (bp->b_kvasize) { 1300 bp->b_qindex = QUEUE_EMPTYKVA; 1301 } else { 1302 bp->b_qindex = QUEUE_EMPTY; 1303 } 1304 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1305 LIST_REMOVE(bp, b_hash); 1306 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1307 bp->b_dev = NODEV; 1308 /* buffers with junk contents */ 1309 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1310 (bp->b_ioflags & BIO_ERROR)) { 1311 bp->b_flags |= B_INVAL; 1312 bp->b_xflags &= ~BX_BKGRDWRITE; 1313 if (bp->b_xflags & BX_BKGRDINPROG) 1314 panic("losing buffer 2"); 1315 bp->b_qindex = QUEUE_CLEAN; 1316 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1317 LIST_REMOVE(bp, b_hash); 1318 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1319 bp->b_dev = NODEV; 1320 1321 /* buffers that are locked */ 1322 } else if (bp->b_flags & B_LOCKED) { 1323 bp->b_qindex = QUEUE_LOCKED; 1324 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1325 1326 /* remaining buffers */ 1327 } else { 1328 if (bp->b_flags & B_DELWRI) 1329 bp->b_qindex = QUEUE_DIRTY; 1330 else 1331 bp->b_qindex = QUEUE_CLEAN; 1332 if (bp->b_flags & B_AGE) 1333 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1334 else 1335 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1336 } 1337 1338 /* 1339 * If B_INVAL, clear B_DELWRI. We've already placed the buffer 1340 * on the correct queue. 1341 */ 1342 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI)) 1343 bundirty(bp); 1344 1345 /* 1346 * Fixup numfreebuffers count. The bp is on an appropriate queue 1347 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1348 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1349 * if B_INVAL is set ). 1350 */ 1351 1352 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI)) 1353 bufcountwakeup(); 1354 1355 /* 1356 * Something we can maybe free or reuse 1357 */ 1358 if (bp->b_bufsize || bp->b_kvasize) 1359 bufspacewakeup(); 1360 1361 /* unlock */ 1362 BUF_UNLOCK(bp); 1363 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | 1364 B_DIRECT | B_NOWDRAIN); 1365 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1366 panic("brelse: not dirty"); 1367 splx(s); 1368} 1369 1370/* 1371 * Release a buffer back to the appropriate queue but do not try to free 1372 * it. The buffer is expected to be used again soon. 1373 * 1374 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1375 * biodone() to requeue an async I/O on completion. It is also used when 1376 * known good buffers need to be requeued but we think we may need the data 1377 * again soon. 1378 * 1379 * XXX we should be able to leave the B_RELBUF hint set on completion. 1380 */ 1381void 1382bqrelse(struct buf * bp) 1383{ 1384 int s; 1385 1386 s = splbio(); 1387 1388 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1389 1390 if (bp->b_qindex != QUEUE_NONE) 1391 panic("bqrelse: free buffer onto another queue???"); 1392 if (BUF_REFCNT(bp) > 1) { 1393 /* do not release to free list */ 1394 BUF_UNLOCK(bp); 1395 splx(s); 1396 return; 1397 } 1398 if (bp->b_flags & B_LOCKED) { 1399 bp->b_ioflags &= ~BIO_ERROR; 1400 bp->b_qindex = QUEUE_LOCKED; 1401 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 1402 /* buffers with stale but valid contents */ 1403 } else if (bp->b_flags & B_DELWRI) { 1404 bp->b_qindex = QUEUE_DIRTY; 1405 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1406 } else if (vm_page_count_severe()) { 1407 /* 1408 * We are too low on memory, we have to try to free the 1409 * buffer (most importantly: the wired pages making up its 1410 * backing store) *now*. 1411 */ 1412 splx(s); 1413 brelse(bp); 1414 return; 1415 } else { 1416 bp->b_qindex = QUEUE_CLEAN; 1417 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1418 } 1419 1420 if ((bp->b_flags & B_LOCKED) == 0 && 1421 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) { 1422 bufcountwakeup(); 1423 } 1424 1425 /* 1426 * Something we can maybe free or reuse. 1427 */ 1428 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1429 bufspacewakeup(); 1430 1431 /* unlock */ 1432 BUF_UNLOCK(bp); 1433 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1434 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1435 panic("bqrelse: not dirty"); 1436 splx(s); 1437} 1438 1439/* Give pages used by the bp back to the VM system (where possible) */ 1440static void 1441vfs_vmio_release(bp) 1442 struct buf *bp; 1443{ 1444 int i; 1445 vm_page_t m; 1446 1447 GIANT_REQUIRED; 1448 1449 for (i = 0; i < bp->b_npages; i++) { 1450 m = bp->b_pages[i]; 1451 bp->b_pages[i] = NULL; 1452 /* 1453 * In order to keep page LRU ordering consistent, put 1454 * everything on the inactive queue. 1455 */ 1456 vm_page_unwire(m, 0); 1457 /* 1458 * We don't mess with busy pages, it is 1459 * the responsibility of the process that 1460 * busied the pages to deal with them. 1461 */ 1462 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1463 continue; 1464 1465 if (m->wire_count == 0) { 1466 vm_page_flag_clear(m, PG_ZERO); 1467 /* 1468 * Might as well free the page if we can and it has 1469 * no valid data. We also free the page if the 1470 * buffer was used for direct I/O 1471 */ 1472 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1473 m->hold_count == 0) { 1474 vm_page_busy(m); 1475 vm_page_protect(m, VM_PROT_NONE); 1476 vm_page_free(m); 1477 } else if (bp->b_flags & B_DIRECT) { 1478 vm_page_try_to_free(m); 1479 } else if (vm_page_count_severe()) { 1480 vm_page_try_to_cache(m); 1481 } 1482 } 1483 } 1484 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1485 1486 if (bp->b_bufsize) { 1487 bufspacewakeup(); 1488 bp->b_bufsize = 0; 1489 } 1490 bp->b_npages = 0; 1491 bp->b_flags &= ~B_VMIO; 1492 if (bp->b_vp) 1493 brelvp(bp); 1494} 1495 1496/* 1497 * Check to see if a block is currently memory resident. 1498 */ 1499struct buf * 1500gbincore(struct vnode * vp, daddr_t blkno) 1501{ 1502 struct buf *bp; 1503 struct bufhashhdr *bh; 1504 1505 bh = bufhash(vp, blkno); 1506 1507 /* Search hash chain */ 1508 LIST_FOREACH(bp, bh, b_hash) { 1509 /* hit */ 1510 if (bp->b_vp == vp && bp->b_lblkno == blkno && 1511 (bp->b_flags & B_INVAL) == 0) { 1512 break; 1513 } 1514 } 1515 return (bp); 1516} 1517 1518/* 1519 * vfs_bio_awrite: 1520 * 1521 * Implement clustered async writes for clearing out B_DELWRI buffers. 1522 * This is much better then the old way of writing only one buffer at 1523 * a time. Note that we may not be presented with the buffers in the 1524 * correct order, so we search for the cluster in both directions. 1525 */ 1526int 1527vfs_bio_awrite(struct buf * bp) 1528{ 1529 int i; 1530 int j; 1531 daddr_t lblkno = bp->b_lblkno; 1532 struct vnode *vp = bp->b_vp; 1533 int s; 1534 int ncl; 1535 struct buf *bpa; 1536 int nwritten; 1537 int size; 1538 int maxcl; 1539 1540 s = splbio(); 1541 /* 1542 * right now we support clustered writing only to regular files. If 1543 * we find a clusterable block we could be in the middle of a cluster 1544 * rather then at the beginning. 1545 */ 1546 if ((vp->v_type == VREG) && 1547 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1548 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1549 1550 size = vp->v_mount->mnt_stat.f_iosize; 1551 maxcl = MAXPHYS / size; 1552 1553 for (i = 1; i < maxcl; i++) { 1554 if ((bpa = gbincore(vp, lblkno + i)) && 1555 BUF_REFCNT(bpa) == 0 && 1556 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1557 (B_DELWRI | B_CLUSTEROK)) && 1558 (bpa->b_bufsize == size)) { 1559 if ((bpa->b_blkno == bpa->b_lblkno) || 1560 (bpa->b_blkno != 1561 bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 1562 break; 1563 } else { 1564 break; 1565 } 1566 } 1567 for (j = 1; i + j <= maxcl && j <= lblkno; j++) { 1568 if ((bpa = gbincore(vp, lblkno - j)) && 1569 BUF_REFCNT(bpa) == 0 && 1570 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) == 1571 (B_DELWRI | B_CLUSTEROK)) && 1572 (bpa->b_bufsize == size)) { 1573 if ((bpa->b_blkno == bpa->b_lblkno) || 1574 (bpa->b_blkno != 1575 bp->b_blkno - ((j * size) >> DEV_BSHIFT))) 1576 break; 1577 } else { 1578 break; 1579 } 1580 } 1581 --j; 1582 ncl = i + j; 1583 /* 1584 * this is a possible cluster write 1585 */ 1586 if (ncl != 1) { 1587 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1588 splx(s); 1589 return nwritten; 1590 } 1591 } 1592 1593 BUF_LOCK(bp, LK_EXCLUSIVE); 1594 bremfree(bp); 1595 bp->b_flags |= B_ASYNC; 1596 1597 splx(s); 1598 /* 1599 * default (old) behavior, writing out only one block 1600 * 1601 * XXX returns b_bufsize instead of b_bcount for nwritten? 1602 */ 1603 nwritten = bp->b_bufsize; 1604 (void) BUF_WRITE(bp); 1605 1606 return nwritten; 1607} 1608 1609/* 1610 * getnewbuf: 1611 * 1612 * Find and initialize a new buffer header, freeing up existing buffers 1613 * in the bufqueues as necessary. The new buffer is returned locked. 1614 * 1615 * Important: B_INVAL is not set. If the caller wishes to throw the 1616 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1617 * 1618 * We block if: 1619 * We have insufficient buffer headers 1620 * We have insufficient buffer space 1621 * buffer_map is too fragmented ( space reservation fails ) 1622 * If we have to flush dirty buffers ( but we try to avoid this ) 1623 * 1624 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1625 * Instead we ask the buf daemon to do it for us. We attempt to 1626 * avoid piecemeal wakeups of the pageout daemon. 1627 */ 1628 1629static struct buf * 1630getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1631{ 1632 struct buf *bp; 1633 struct buf *nbp; 1634 int defrag = 0; 1635 int nqindex; 1636 static int flushingbufs; 1637 1638 GIANT_REQUIRED; 1639 1640 /* 1641 * We can't afford to block since we might be holding a vnode lock, 1642 * which may prevent system daemons from running. We deal with 1643 * low-memory situations by proactively returning memory and running 1644 * async I/O rather then sync I/O. 1645 */ 1646 1647 ++getnewbufcalls; 1648 --getnewbufrestarts; 1649restart: 1650 ++getnewbufrestarts; 1651 1652 /* 1653 * Setup for scan. If we do not have enough free buffers, 1654 * we setup a degenerate case that immediately fails. Note 1655 * that if we are specially marked process, we are allowed to 1656 * dip into our reserves. 1657 * 1658 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1659 * 1660 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1661 * However, there are a number of cases (defragging, reusing, ...) 1662 * where we cannot backup. 1663 */ 1664 nqindex = QUEUE_EMPTYKVA; 1665 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1666 1667 if (nbp == NULL) { 1668 /* 1669 * If no EMPTYKVA buffers and we are either 1670 * defragging or reusing, locate a CLEAN buffer 1671 * to free or reuse. If bufspace useage is low 1672 * skip this step so we can allocate a new buffer. 1673 */ 1674 if (defrag || bufspace >= lobufspace) { 1675 nqindex = QUEUE_CLEAN; 1676 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1677 } 1678 1679 /* 1680 * If we could not find or were not allowed to reuse a 1681 * CLEAN buffer, check to see if it is ok to use an EMPTY 1682 * buffer. We can only use an EMPTY buffer if allocating 1683 * its KVA would not otherwise run us out of buffer space. 1684 */ 1685 if (nbp == NULL && defrag == 0 && 1686 bufspace + maxsize < hibufspace) { 1687 nqindex = QUEUE_EMPTY; 1688 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1689 } 1690 } 1691 1692 /* 1693 * Run scan, possibly freeing data and/or kva mappings on the fly 1694 * depending. 1695 */ 1696 1697 while ((bp = nbp) != NULL) { 1698 int qindex = nqindex; 1699 1700 /* 1701 * Calculate next bp ( we can only use it if we do not block 1702 * or do other fancy things ). 1703 */ 1704 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1705 switch(qindex) { 1706 case QUEUE_EMPTY: 1707 nqindex = QUEUE_EMPTYKVA; 1708 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1709 break; 1710 /* fall through */ 1711 case QUEUE_EMPTYKVA: 1712 nqindex = QUEUE_CLEAN; 1713 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1714 break; 1715 /* fall through */ 1716 case QUEUE_CLEAN: 1717 /* 1718 * nbp is NULL. 1719 */ 1720 break; 1721 } 1722 } 1723 1724 /* 1725 * Sanity Checks 1726 */ 1727 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1728 1729 /* 1730 * Note: we no longer distinguish between VMIO and non-VMIO 1731 * buffers. 1732 */ 1733 1734 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1735 1736 /* 1737 * If we are defragging then we need a buffer with 1738 * b_kvasize != 0. XXX this situation should no longer 1739 * occur, if defrag is non-zero the buffer's b_kvasize 1740 * should also be non-zero at this point. XXX 1741 */ 1742 if (defrag && bp->b_kvasize == 0) { 1743 printf("Warning: defrag empty buffer %p\n", bp); 1744 continue; 1745 } 1746 1747 /* 1748 * Start freeing the bp. This is somewhat involved. nbp 1749 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1750 */ 1751 1752 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 1753 panic("getnewbuf: locked buf"); 1754 bremfree(bp); 1755 1756 if (qindex == QUEUE_CLEAN) { 1757 if (bp->b_flags & B_VMIO) { 1758 bp->b_flags &= ~B_ASYNC; 1759 vfs_vmio_release(bp); 1760 } 1761 if (bp->b_vp) 1762 brelvp(bp); 1763 } 1764 1765 /* 1766 * NOTE: nbp is now entirely invalid. We can only restart 1767 * the scan from this point on. 1768 * 1769 * Get the rest of the buffer freed up. b_kva* is still 1770 * valid after this operation. 1771 */ 1772 1773 if (bp->b_rcred != NOCRED) { 1774 crfree(bp->b_rcred); 1775 bp->b_rcred = NOCRED; 1776 } 1777 if (bp->b_wcred != NOCRED) { 1778 crfree(bp->b_wcred); 1779 bp->b_wcred = NOCRED; 1780 } 1781 if (LIST_FIRST(&bp->b_dep) != NULL) 1782 buf_deallocate(bp); 1783 if (bp->b_xflags & BX_BKGRDINPROG) 1784 panic("losing buffer 3"); 1785 LIST_REMOVE(bp, b_hash); 1786 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1787 1788 if (bp->b_bufsize) 1789 allocbuf(bp, 0); 1790 1791 bp->b_flags = 0; 1792 bp->b_ioflags = 0; 1793 bp->b_xflags = 0; 1794 bp->b_dev = NODEV; 1795 bp->b_vp = NULL; 1796 bp->b_blkno = bp->b_lblkno = 0; 1797 bp->b_offset = NOOFFSET; 1798 bp->b_iodone = 0; 1799 bp->b_error = 0; 1800 bp->b_resid = 0; 1801 bp->b_bcount = 0; 1802 bp->b_npages = 0; 1803 bp->b_dirtyoff = bp->b_dirtyend = 0; 1804 bp->b_magic = B_MAGIC_BIO; 1805 bp->b_op = &buf_ops_bio; 1806 1807 LIST_INIT(&bp->b_dep); 1808 1809 /* 1810 * If we are defragging then free the buffer. 1811 */ 1812 if (defrag) { 1813 bp->b_flags |= B_INVAL; 1814 bfreekva(bp); 1815 brelse(bp); 1816 defrag = 0; 1817 goto restart; 1818 } 1819 1820 /* 1821 * If we are overcomitted then recover the buffer and its 1822 * KVM space. This occurs in rare situations when multiple 1823 * processes are blocked in getnewbuf() or allocbuf(). 1824 */ 1825 if (bufspace >= hibufspace) 1826 flushingbufs = 1; 1827 if (flushingbufs && bp->b_kvasize != 0) { 1828 bp->b_flags |= B_INVAL; 1829 bfreekva(bp); 1830 brelse(bp); 1831 goto restart; 1832 } 1833 if (bufspace < lobufspace) 1834 flushingbufs = 0; 1835 break; 1836 } 1837 1838 /* 1839 * If we exhausted our list, sleep as appropriate. We may have to 1840 * wakeup various daemons and write out some dirty buffers. 1841 * 1842 * Generally we are sleeping due to insufficient buffer space. 1843 */ 1844 1845 if (bp == NULL) { 1846 int flags; 1847 char *waitmsg; 1848 1849 if (defrag) { 1850 flags = VFS_BIO_NEED_BUFSPACE; 1851 waitmsg = "nbufkv"; 1852 } else if (bufspace >= hibufspace) { 1853 waitmsg = "nbufbs"; 1854 flags = VFS_BIO_NEED_BUFSPACE; 1855 } else { 1856 waitmsg = "newbuf"; 1857 flags = VFS_BIO_NEED_ANY; 1858 } 1859 1860 bd_speedup(); /* heeeelp */ 1861 1862 needsbuffer |= flags; 1863 while (needsbuffer & flags) { 1864 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, 1865 waitmsg, slptimeo)) 1866 return (NULL); 1867 } 1868 } else { 1869 /* 1870 * We finally have a valid bp. We aren't quite out of the 1871 * woods, we still have to reserve kva space. In order 1872 * to keep fragmentation sane we only allocate kva in 1873 * BKVASIZE chunks. 1874 */ 1875 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 1876 1877 if (maxsize != bp->b_kvasize) { 1878 vm_offset_t addr = 0; 1879 1880 bfreekva(bp); 1881 1882 if (vm_map_findspace(buffer_map, 1883 vm_map_min(buffer_map), maxsize, &addr)) { 1884 /* 1885 * Uh oh. Buffer map is to fragmented. We 1886 * must defragment the map. 1887 */ 1888 ++bufdefragcnt; 1889 defrag = 1; 1890 bp->b_flags |= B_INVAL; 1891 brelse(bp); 1892 goto restart; 1893 } 1894 if (addr) { 1895 vm_map_insert(buffer_map, NULL, 0, 1896 addr, addr + maxsize, 1897 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1898 1899 bp->b_kvabase = (caddr_t) addr; 1900 bp->b_kvasize = maxsize; 1901 bufspace += bp->b_kvasize; 1902 ++bufreusecnt; 1903 } 1904 } 1905 bp->b_data = bp->b_kvabase; 1906 } 1907 return(bp); 1908} 1909 1910/* 1911 * buf_daemon: 1912 * 1913 * buffer flushing daemon. Buffers are normally flushed by the 1914 * update daemon but if it cannot keep up this process starts to 1915 * take the load in an attempt to prevent getnewbuf() from blocking. 1916 */ 1917 1918static struct proc *bufdaemonproc; 1919 1920static struct kproc_desc buf_kp = { 1921 "bufdaemon", 1922 buf_daemon, 1923 &bufdaemonproc 1924}; 1925SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 1926 1927static void 1928buf_daemon() 1929{ 1930 int s; 1931 1932 mtx_lock(&Giant); 1933 1934 /* 1935 * This process needs to be suspended prior to shutdown sync. 1936 */ 1937 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 1938 SHUTDOWN_PRI_LAST); 1939 1940 /* 1941 * This process is allowed to take the buffer cache to the limit 1942 */ 1943 s = splbio(); 1944 1945 for (;;) { 1946 kthread_suspend_check(bufdaemonproc); 1947 1948 bd_request = 0; 1949 1950 /* 1951 * Do the flush. Limit the amount of in-transit I/O we 1952 * allow to build up, otherwise we would completely saturate 1953 * the I/O system. Wakeup any waiting processes before we 1954 * normally would so they can run in parallel with our drain. 1955 */ 1956 while (numdirtybuffers > lodirtybuffers) { 1957 if (flushbufqueues() == 0) 1958 break; 1959 waitrunningbufspace(); 1960 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 1961 } 1962 1963 /* 1964 * Only clear bd_request if we have reached our low water 1965 * mark. The buf_daemon normally waits 1 second and 1966 * then incrementally flushes any dirty buffers that have 1967 * built up, within reason. 1968 * 1969 * If we were unable to hit our low water mark and couldn't 1970 * find any flushable buffers, we sleep half a second. 1971 * Otherwise we loop immediately. 1972 */ 1973 if (numdirtybuffers <= lodirtybuffers) { 1974 /* 1975 * We reached our low water mark, reset the 1976 * request and sleep until we are needed again. 1977 * The sleep is just so the suspend code works. 1978 */ 1979 bd_request = 0; 1980 tsleep(&bd_request, PVM, "psleep", hz); 1981 } else { 1982 /* 1983 * We couldn't find any flushable dirty buffers but 1984 * still have too many dirty buffers, we 1985 * have to sleep and try again. (rare) 1986 */ 1987 tsleep(&bd_request, PVM, "qsleep", hz / 2); 1988 } 1989 } 1990} 1991 1992/* 1993 * flushbufqueues: 1994 * 1995 * Try to flush a buffer in the dirty queue. We must be careful to 1996 * free up B_INVAL buffers instead of write them, which NFS is 1997 * particularly sensitive to. 1998 */ 1999 2000static int 2001flushbufqueues(void) 2002{ 2003 struct buf *bp; 2004 int r = 0; 2005 2006 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 2007 2008 while (bp) { 2009 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp)); 2010 if ((bp->b_flags & B_DELWRI) != 0 && 2011 (bp->b_xflags & BX_BKGRDINPROG) == 0) { 2012 if (bp->b_flags & B_INVAL) { 2013 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) 2014 panic("flushbufqueues: locked buf"); 2015 bremfree(bp); 2016 brelse(bp); 2017 ++r; 2018 break; 2019 } 2020 if (LIST_FIRST(&bp->b_dep) != NULL && 2021 (bp->b_flags & B_DEFERRED) == 0 && 2022 buf_countdeps(bp, 0)) { 2023 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY], 2024 bp, b_freelist); 2025 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], 2026 bp, b_freelist); 2027 bp->b_flags |= B_DEFERRED; 2028 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]); 2029 continue; 2030 } 2031 vfs_bio_awrite(bp); 2032 ++r; 2033 break; 2034 } 2035 bp = TAILQ_NEXT(bp, b_freelist); 2036 } 2037 return (r); 2038} 2039 2040/* 2041 * Check to see if a block is currently memory resident. 2042 */ 2043struct buf * 2044incore(struct vnode * vp, daddr_t blkno) 2045{ 2046 struct buf *bp; 2047 2048 int s = splbio(); 2049 bp = gbincore(vp, blkno); 2050 splx(s); 2051 return (bp); 2052} 2053 2054/* 2055 * Returns true if no I/O is needed to access the 2056 * associated VM object. This is like incore except 2057 * it also hunts around in the VM system for the data. 2058 */ 2059 2060int 2061inmem(struct vnode * vp, daddr_t blkno) 2062{ 2063 vm_object_t obj; 2064 vm_offset_t toff, tinc, size; 2065 vm_page_t m; 2066 vm_ooffset_t off; 2067 2068 GIANT_REQUIRED; 2069 2070 if (incore(vp, blkno)) 2071 return 1; 2072 if (vp->v_mount == NULL) 2073 return 0; 2074 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0) 2075 return 0; 2076 2077 size = PAGE_SIZE; 2078 if (size > vp->v_mount->mnt_stat.f_iosize) 2079 size = vp->v_mount->mnt_stat.f_iosize; 2080 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2081 2082 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2083 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2084 if (!m) 2085 goto notinmem; 2086 tinc = size; 2087 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2088 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2089 if (vm_page_is_valid(m, 2090 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2091 goto notinmem; 2092 } 2093 return 1; 2094 2095notinmem: 2096 return (0); 2097} 2098 2099/* 2100 * vfs_setdirty: 2101 * 2102 * Sets the dirty range for a buffer based on the status of the dirty 2103 * bits in the pages comprising the buffer. 2104 * 2105 * The range is limited to the size of the buffer. 2106 * 2107 * This routine is primarily used by NFS, but is generalized for the 2108 * B_VMIO case. 2109 */ 2110static void 2111vfs_setdirty(struct buf *bp) 2112{ 2113 int i; 2114 vm_object_t object; 2115 2116 GIANT_REQUIRED; 2117 /* 2118 * Degenerate case - empty buffer 2119 */ 2120 2121 if (bp->b_bufsize == 0) 2122 return; 2123 2124 /* 2125 * We qualify the scan for modified pages on whether the 2126 * object has been flushed yet. The OBJ_WRITEABLE flag 2127 * is not cleared simply by protecting pages off. 2128 */ 2129 2130 if ((bp->b_flags & B_VMIO) == 0) 2131 return; 2132 2133 object = bp->b_pages[0]->object; 2134 2135 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2136 printf("Warning: object %p writeable but not mightbedirty\n", object); 2137 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2138 printf("Warning: object %p mightbedirty but not writeable\n", object); 2139 2140 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2141 vm_offset_t boffset; 2142 vm_offset_t eoffset; 2143 2144 /* 2145 * test the pages to see if they have been modified directly 2146 * by users through the VM system. 2147 */ 2148 for (i = 0; i < bp->b_npages; i++) { 2149 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2150 vm_page_test_dirty(bp->b_pages[i]); 2151 } 2152 2153 /* 2154 * Calculate the encompassing dirty range, boffset and eoffset, 2155 * (eoffset - boffset) bytes. 2156 */ 2157 2158 for (i = 0; i < bp->b_npages; i++) { 2159 if (bp->b_pages[i]->dirty) 2160 break; 2161 } 2162 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2163 2164 for (i = bp->b_npages - 1; i >= 0; --i) { 2165 if (bp->b_pages[i]->dirty) { 2166 break; 2167 } 2168 } 2169 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2170 2171 /* 2172 * Fit it to the buffer. 2173 */ 2174 2175 if (eoffset > bp->b_bcount) 2176 eoffset = bp->b_bcount; 2177 2178 /* 2179 * If we have a good dirty range, merge with the existing 2180 * dirty range. 2181 */ 2182 2183 if (boffset < eoffset) { 2184 if (bp->b_dirtyoff > boffset) 2185 bp->b_dirtyoff = boffset; 2186 if (bp->b_dirtyend < eoffset) 2187 bp->b_dirtyend = eoffset; 2188 } 2189 } 2190} 2191 2192/* 2193 * getblk: 2194 * 2195 * Get a block given a specified block and offset into a file/device. 2196 * The buffers B_DONE bit will be cleared on return, making it almost 2197 * ready for an I/O initiation. B_INVAL may or may not be set on 2198 * return. The caller should clear B_INVAL prior to initiating a 2199 * READ. 2200 * 2201 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2202 * an existing buffer. 2203 * 2204 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2205 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2206 * and then cleared based on the backing VM. If the previous buffer is 2207 * non-0-sized but invalid, B_CACHE will be cleared. 2208 * 2209 * If getblk() must create a new buffer, the new buffer is returned with 2210 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2211 * case it is returned with B_INVAL clear and B_CACHE set based on the 2212 * backing VM. 2213 * 2214 * getblk() also forces a BUF_WRITE() for any B_DELWRI buffer whos 2215 * B_CACHE bit is clear. 2216 * 2217 * What this means, basically, is that the caller should use B_CACHE to 2218 * determine whether the buffer is fully valid or not and should clear 2219 * B_INVAL prior to issuing a read. If the caller intends to validate 2220 * the buffer by loading its data area with something, the caller needs 2221 * to clear B_INVAL. If the caller does this without issuing an I/O, 2222 * the caller should set B_CACHE ( as an optimization ), else the caller 2223 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2224 * a write attempt or if it was a successfull read. If the caller 2225 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2226 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2227 */ 2228struct buf * 2229getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 2230{ 2231 struct buf *bp; 2232 int s; 2233 struct bufhashhdr *bh; 2234 2235 if (size > MAXBSIZE) 2236 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2237 2238 s = splbio(); 2239loop: 2240 /* 2241 * Block if we are low on buffers. Certain processes are allowed 2242 * to completely exhaust the buffer cache. 2243 * 2244 * If this check ever becomes a bottleneck it may be better to 2245 * move it into the else, when gbincore() fails. At the moment 2246 * it isn't a problem. 2247 * 2248 * XXX remove if 0 sections (clean this up after its proven) 2249 */ 2250 if (numfreebuffers == 0) { 2251 if (curthread == PCPU_GET(idlethread)) 2252 return NULL; 2253 needsbuffer |= VFS_BIO_NEED_ANY; 2254 } 2255 2256 if ((bp = gbincore(vp, blkno))) { 2257 /* 2258 * Buffer is in-core. If the buffer is not busy, it must 2259 * be on a queue. 2260 */ 2261 2262 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) { 2263 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL, 2264 "getblk", slpflag, slptimeo) == ENOLCK) 2265 goto loop; 2266 splx(s); 2267 return (struct buf *) NULL; 2268 } 2269 2270 /* 2271 * The buffer is locked. B_CACHE is cleared if the buffer is 2272 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2273 * and for a VMIO buffer B_CACHE is adjusted according to the 2274 * backing VM cache. 2275 */ 2276 if (bp->b_flags & B_INVAL) 2277 bp->b_flags &= ~B_CACHE; 2278 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2279 bp->b_flags |= B_CACHE; 2280 bremfree(bp); 2281 2282 /* 2283 * check for size inconsistancies for non-VMIO case. 2284 */ 2285 2286 if (bp->b_bcount != size) { 2287 if ((bp->b_flags & B_VMIO) == 0 || 2288 (size > bp->b_kvasize)) { 2289 if (bp->b_flags & B_DELWRI) { 2290 bp->b_flags |= B_NOCACHE; 2291 BUF_WRITE(bp); 2292 } else { 2293 if ((bp->b_flags & B_VMIO) && 2294 (LIST_FIRST(&bp->b_dep) == NULL)) { 2295 bp->b_flags |= B_RELBUF; 2296 brelse(bp); 2297 } else { 2298 bp->b_flags |= B_NOCACHE; 2299 BUF_WRITE(bp); 2300 } 2301 } 2302 goto loop; 2303 } 2304 } 2305 2306 /* 2307 * If the size is inconsistant in the VMIO case, we can resize 2308 * the buffer. This might lead to B_CACHE getting set or 2309 * cleared. If the size has not changed, B_CACHE remains 2310 * unchanged from its previous state. 2311 */ 2312 2313 if (bp->b_bcount != size) 2314 allocbuf(bp, size); 2315 2316 KASSERT(bp->b_offset != NOOFFSET, 2317 ("getblk: no buffer offset")); 2318 2319 /* 2320 * A buffer with B_DELWRI set and B_CACHE clear must 2321 * be committed before we can return the buffer in 2322 * order to prevent the caller from issuing a read 2323 * ( due to B_CACHE not being set ) and overwriting 2324 * it. 2325 * 2326 * Most callers, including NFS and FFS, need this to 2327 * operate properly either because they assume they 2328 * can issue a read if B_CACHE is not set, or because 2329 * ( for example ) an uncached B_DELWRI might loop due 2330 * to softupdates re-dirtying the buffer. In the latter 2331 * case, B_CACHE is set after the first write completes, 2332 * preventing further loops. 2333 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2334 * above while extending the buffer, we cannot allow the 2335 * buffer to remain with B_CACHE set after the write 2336 * completes or it will represent a corrupt state. To 2337 * deal with this we set B_NOCACHE to scrap the buffer 2338 * after the write. 2339 * 2340 * We might be able to do something fancy, like setting 2341 * B_CACHE in bwrite() except if B_DELWRI is already set, 2342 * so the below call doesn't set B_CACHE, but that gets real 2343 * confusing. This is much easier. 2344 */ 2345 2346 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2347 bp->b_flags |= B_NOCACHE; 2348 BUF_WRITE(bp); 2349 goto loop; 2350 } 2351 2352 splx(s); 2353 bp->b_flags &= ~B_DONE; 2354 } else { 2355 /* 2356 * Buffer is not in-core, create new buffer. The buffer 2357 * returned by getnewbuf() is locked. Note that the returned 2358 * buffer is also considered valid (not marked B_INVAL). 2359 */ 2360 int bsize, maxsize, vmio; 2361 off_t offset; 2362 2363 if (vn_isdisk(vp, NULL)) 2364 bsize = DEV_BSIZE; 2365 else if (vp->v_mountedhere) 2366 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 2367 else if (vp->v_mount) 2368 bsize = vp->v_mount->mnt_stat.f_iosize; 2369 else 2370 bsize = size; 2371 2372 offset = blkno * bsize; 2373 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF); 2374 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2375 maxsize = imax(maxsize, bsize); 2376 2377 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) { 2378 if (slpflag || slptimeo) { 2379 splx(s); 2380 return NULL; 2381 } 2382 goto loop; 2383 } 2384 2385 /* 2386 * This code is used to make sure that a buffer is not 2387 * created while the getnewbuf routine is blocked. 2388 * This can be a problem whether the vnode is locked or not. 2389 * If the buffer is created out from under us, we have to 2390 * throw away the one we just created. There is now window 2391 * race because we are safely running at splbio() from the 2392 * point of the duplicate buffer creation through to here, 2393 * and we've locked the buffer. 2394 */ 2395 if (gbincore(vp, blkno)) { 2396 bp->b_flags |= B_INVAL; 2397 brelse(bp); 2398 goto loop; 2399 } 2400 2401 /* 2402 * Insert the buffer into the hash, so that it can 2403 * be found by incore. 2404 */ 2405 bp->b_blkno = bp->b_lblkno = blkno; 2406 bp->b_offset = offset; 2407 2408 bgetvp(vp, bp); 2409 LIST_REMOVE(bp, b_hash); 2410 bh = bufhash(vp, blkno); 2411 LIST_INSERT_HEAD(bh, bp, b_hash); 2412 2413 /* 2414 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2415 * buffer size starts out as 0, B_CACHE will be set by 2416 * allocbuf() for the VMIO case prior to it testing the 2417 * backing store for validity. 2418 */ 2419 2420 if (vmio) { 2421 bp->b_flags |= B_VMIO; 2422#if defined(VFS_BIO_DEBUG) 2423 if (vp->v_type != VREG) 2424 printf("getblk: vmioing file type %d???\n", vp->v_type); 2425#endif 2426 } else { 2427 bp->b_flags &= ~B_VMIO; 2428 } 2429 2430 allocbuf(bp, size); 2431 2432 splx(s); 2433 bp->b_flags &= ~B_DONE; 2434 } 2435 return (bp); 2436} 2437 2438/* 2439 * Get an empty, disassociated buffer of given size. The buffer is initially 2440 * set to B_INVAL. 2441 */ 2442struct buf * 2443geteblk(int size) 2444{ 2445 struct buf *bp; 2446 int s; 2447 int maxsize; 2448 2449 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2450 2451 s = splbio(); 2452 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0); 2453 splx(s); 2454 allocbuf(bp, size); 2455 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2456 return (bp); 2457} 2458 2459 2460/* 2461 * This code constitutes the buffer memory from either anonymous system 2462 * memory (in the case of non-VMIO operations) or from an associated 2463 * VM object (in the case of VMIO operations). This code is able to 2464 * resize a buffer up or down. 2465 * 2466 * Note that this code is tricky, and has many complications to resolve 2467 * deadlock or inconsistant data situations. Tread lightly!!! 2468 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2469 * the caller. Calling this code willy nilly can result in the loss of data. 2470 * 2471 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2472 * B_CACHE for the non-VMIO case. 2473 */ 2474 2475int 2476allocbuf(struct buf *bp, int size) 2477{ 2478 int newbsize, mbsize; 2479 int i; 2480 2481 GIANT_REQUIRED; 2482 2483 if (BUF_REFCNT(bp) == 0) 2484 panic("allocbuf: buffer not busy"); 2485 2486 if (bp->b_kvasize < size) 2487 panic("allocbuf: buffer too small"); 2488 2489 if ((bp->b_flags & B_VMIO) == 0) { 2490 caddr_t origbuf; 2491 int origbufsize; 2492 /* 2493 * Just get anonymous memory from the kernel. Don't 2494 * mess with B_CACHE. 2495 */ 2496 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2497 if (bp->b_flags & B_MALLOC) 2498 newbsize = mbsize; 2499 else 2500 newbsize = round_page(size); 2501 2502 if (newbsize < bp->b_bufsize) { 2503 /* 2504 * malloced buffers are not shrunk 2505 */ 2506 if (bp->b_flags & B_MALLOC) { 2507 if (newbsize) { 2508 bp->b_bcount = size; 2509 } else { 2510 free(bp->b_data, M_BIOBUF); 2511 if (bp->b_bufsize) { 2512 bufmallocspace -= bp->b_bufsize; 2513 bufspacewakeup(); 2514 bp->b_bufsize = 0; 2515 } 2516 bp->b_data = bp->b_kvabase; 2517 bp->b_bcount = 0; 2518 bp->b_flags &= ~B_MALLOC; 2519 } 2520 return 1; 2521 } 2522 vm_hold_free_pages( 2523 bp, 2524 (vm_offset_t) bp->b_data + newbsize, 2525 (vm_offset_t) bp->b_data + bp->b_bufsize); 2526 } else if (newbsize > bp->b_bufsize) { 2527 /* 2528 * We only use malloced memory on the first allocation. 2529 * and revert to page-allocated memory when the buffer 2530 * grows. 2531 */ 2532 if ( (bufmallocspace < maxbufmallocspace) && 2533 (bp->b_bufsize == 0) && 2534 (mbsize <= PAGE_SIZE/2)) { 2535 2536 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2537 bp->b_bufsize = mbsize; 2538 bp->b_bcount = size; 2539 bp->b_flags |= B_MALLOC; 2540 bufmallocspace += mbsize; 2541 return 1; 2542 } 2543 origbuf = NULL; 2544 origbufsize = 0; 2545 /* 2546 * If the buffer is growing on its other-than-first allocation, 2547 * then we revert to the page-allocation scheme. 2548 */ 2549 if (bp->b_flags & B_MALLOC) { 2550 origbuf = bp->b_data; 2551 origbufsize = bp->b_bufsize; 2552 bp->b_data = bp->b_kvabase; 2553 if (bp->b_bufsize) { 2554 bufmallocspace -= bp->b_bufsize; 2555 bufspacewakeup(); 2556 bp->b_bufsize = 0; 2557 } 2558 bp->b_flags &= ~B_MALLOC; 2559 newbsize = round_page(newbsize); 2560 } 2561 vm_hold_load_pages( 2562 bp, 2563 (vm_offset_t) bp->b_data + bp->b_bufsize, 2564 (vm_offset_t) bp->b_data + newbsize); 2565 if (origbuf) { 2566 bcopy(origbuf, bp->b_data, origbufsize); 2567 free(origbuf, M_BIOBUF); 2568 } 2569 } 2570 } else { 2571 vm_page_t m; 2572 int desiredpages; 2573 2574 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2575 desiredpages = (size == 0) ? 0 : 2576 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2577 2578 if (bp->b_flags & B_MALLOC) 2579 panic("allocbuf: VMIO buffer can't be malloced"); 2580 /* 2581 * Set B_CACHE initially if buffer is 0 length or will become 2582 * 0-length. 2583 */ 2584 if (size == 0 || bp->b_bufsize == 0) 2585 bp->b_flags |= B_CACHE; 2586 2587 if (newbsize < bp->b_bufsize) { 2588 /* 2589 * DEV_BSIZE aligned new buffer size is less then the 2590 * DEV_BSIZE aligned existing buffer size. Figure out 2591 * if we have to remove any pages. 2592 */ 2593 if (desiredpages < bp->b_npages) { 2594 for (i = desiredpages; i < bp->b_npages; i++) { 2595 /* 2596 * the page is not freed here -- it 2597 * is the responsibility of 2598 * vnode_pager_setsize 2599 */ 2600 m = bp->b_pages[i]; 2601 KASSERT(m != bogus_page, 2602 ("allocbuf: bogus page found")); 2603 while (vm_page_sleep_busy(m, TRUE, "biodep")) 2604 ; 2605 2606 bp->b_pages[i] = NULL; 2607 vm_page_unwire(m, 0); 2608 } 2609 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2610 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2611 bp->b_npages = desiredpages; 2612 } 2613 } else if (size > bp->b_bcount) { 2614 /* 2615 * We are growing the buffer, possibly in a 2616 * byte-granular fashion. 2617 */ 2618 struct vnode *vp; 2619 vm_object_t obj; 2620 vm_offset_t toff; 2621 vm_offset_t tinc; 2622 2623 /* 2624 * Step 1, bring in the VM pages from the object, 2625 * allocating them if necessary. We must clear 2626 * B_CACHE if these pages are not valid for the 2627 * range covered by the buffer. 2628 */ 2629 2630 vp = bp->b_vp; 2631 VOP_GETVOBJECT(vp, &obj); 2632 2633 while (bp->b_npages < desiredpages) { 2634 vm_page_t m; 2635 vm_pindex_t pi; 2636 2637 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2638 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2639 /* 2640 * note: must allocate system pages 2641 * since blocking here could intefere 2642 * with paging I/O, no matter which 2643 * process we are. 2644 */ 2645 m = vm_page_alloc(obj, pi, VM_ALLOC_SYSTEM); 2646 if (m == NULL) { 2647 VM_WAIT; 2648 vm_pageout_deficit += desiredpages - bp->b_npages; 2649 } else { 2650 vm_page_wire(m); 2651 vm_page_wakeup(m); 2652 bp->b_flags &= ~B_CACHE; 2653 bp->b_pages[bp->b_npages] = m; 2654 ++bp->b_npages; 2655 } 2656 continue; 2657 } 2658 2659 /* 2660 * We found a page. If we have to sleep on it, 2661 * retry because it might have gotten freed out 2662 * from under us. 2663 * 2664 * We can only test PG_BUSY here. Blocking on 2665 * m->busy might lead to a deadlock: 2666 * 2667 * vm_fault->getpages->cluster_read->allocbuf 2668 * 2669 */ 2670 2671 if (vm_page_sleep_busy(m, FALSE, "pgtblk")) 2672 continue; 2673 2674 /* 2675 * We have a good page. Should we wakeup the 2676 * page daemon? 2677 */ 2678 if ((curproc != pageproc) && 2679 ((m->queue - m->pc) == PQ_CACHE) && 2680 ((cnt.v_free_count + cnt.v_cache_count) < 2681 (cnt.v_free_min + cnt.v_cache_min))) { 2682 pagedaemon_wakeup(); 2683 } 2684 vm_page_flag_clear(m, PG_ZERO); 2685 vm_page_wire(m); 2686 bp->b_pages[bp->b_npages] = m; 2687 ++bp->b_npages; 2688 } 2689 2690 /* 2691 * Step 2. We've loaded the pages into the buffer, 2692 * we have to figure out if we can still have B_CACHE 2693 * set. Note that B_CACHE is set according to the 2694 * byte-granular range ( bcount and size ), new the 2695 * aligned range ( newbsize ). 2696 * 2697 * The VM test is against m->valid, which is DEV_BSIZE 2698 * aligned. Needless to say, the validity of the data 2699 * needs to also be DEV_BSIZE aligned. Note that this 2700 * fails with NFS if the server or some other client 2701 * extends the file's EOF. If our buffer is resized, 2702 * B_CACHE may remain set! XXX 2703 */ 2704 2705 toff = bp->b_bcount; 2706 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2707 2708 while ((bp->b_flags & B_CACHE) && toff < size) { 2709 vm_pindex_t pi; 2710 2711 if (tinc > (size - toff)) 2712 tinc = size - toff; 2713 2714 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2715 PAGE_SHIFT; 2716 2717 vfs_buf_test_cache( 2718 bp, 2719 bp->b_offset, 2720 toff, 2721 tinc, 2722 bp->b_pages[pi] 2723 ); 2724 toff += tinc; 2725 tinc = PAGE_SIZE; 2726 } 2727 2728 /* 2729 * Step 3, fixup the KVM pmap. Remember that 2730 * bp->b_data is relative to bp->b_offset, but 2731 * bp->b_offset may be offset into the first page. 2732 */ 2733 2734 bp->b_data = (caddr_t) 2735 trunc_page((vm_offset_t)bp->b_data); 2736 pmap_qenter( 2737 (vm_offset_t)bp->b_data, 2738 bp->b_pages, 2739 bp->b_npages 2740 ); 2741 2742 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 2743 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 2744 } 2745 } 2746 if (newbsize < bp->b_bufsize) 2747 bufspacewakeup(); 2748 bp->b_bufsize = newbsize; /* actual buffer allocation */ 2749 bp->b_bcount = size; /* requested buffer size */ 2750 return 1; 2751} 2752 2753/* 2754 * bufwait: 2755 * 2756 * Wait for buffer I/O completion, returning error status. The buffer 2757 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR 2758 * error and cleared. 2759 */ 2760int 2761bufwait(register struct buf * bp) 2762{ 2763 int s; 2764 2765 s = splbio(); 2766 while ((bp->b_flags & B_DONE) == 0) { 2767 if (bp->b_iocmd == BIO_READ) 2768 tsleep(bp, PRIBIO, "biord", 0); 2769 else 2770 tsleep(bp, PRIBIO, "biowr", 0); 2771 } 2772 splx(s); 2773 if (bp->b_flags & B_EINTR) { 2774 bp->b_flags &= ~B_EINTR; 2775 return (EINTR); 2776 } 2777 if (bp->b_ioflags & BIO_ERROR) { 2778 return (bp->b_error ? bp->b_error : EIO); 2779 } else { 2780 return (0); 2781 } 2782} 2783 2784 /* 2785 * Call back function from struct bio back up to struct buf. 2786 * The corresponding initialization lives in sys/conf.h:DEV_STRATEGY(). 2787 */ 2788void 2789bufdonebio(struct bio *bp) 2790{ 2791 bufdone(bp->bio_caller2); 2792} 2793 2794/* 2795 * bufdone: 2796 * 2797 * Finish I/O on a buffer, optionally calling a completion function. 2798 * This is usually called from an interrupt so process blocking is 2799 * not allowed. 2800 * 2801 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 2802 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 2803 * assuming B_INVAL is clear. 2804 * 2805 * For the VMIO case, we set B_CACHE if the op was a read and no 2806 * read error occured, or if the op was a write. B_CACHE is never 2807 * set if the buffer is invalid or otherwise uncacheable. 2808 * 2809 * biodone does not mess with B_INVAL, allowing the I/O routine or the 2810 * initiator to leave B_INVAL set to brelse the buffer out of existance 2811 * in the biodone routine. 2812 */ 2813void 2814bufdone(struct buf *bp) 2815{ 2816 int s, error; 2817 void (*biodone)(struct buf *); 2818 2819 GIANT_REQUIRED; 2820 2821 s = splbio(); 2822 2823 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 2824 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 2825 2826 bp->b_flags |= B_DONE; 2827 runningbufwakeup(bp); 2828 2829 if (bp->b_iocmd == BIO_DELETE) { 2830 brelse(bp); 2831 splx(s); 2832 return; 2833 } 2834 2835 if (bp->b_iocmd == BIO_WRITE) { 2836 vwakeup(bp); 2837 } 2838 2839 /* call optional completion function if requested */ 2840 if (bp->b_iodone != NULL) { 2841 biodone = bp->b_iodone; 2842 bp->b_iodone = NULL; 2843 (*biodone) (bp); 2844 splx(s); 2845 return; 2846 } 2847 if (LIST_FIRST(&bp->b_dep) != NULL) 2848 buf_complete(bp); 2849 2850 if (bp->b_flags & B_VMIO) { 2851 int i; 2852 vm_ooffset_t foff; 2853 vm_page_t m; 2854 vm_object_t obj; 2855 int iosize; 2856 struct vnode *vp = bp->b_vp; 2857 2858 error = VOP_GETVOBJECT(vp, &obj); 2859 2860#if defined(VFS_BIO_DEBUG) 2861 if (vp->v_usecount == 0) { 2862 panic("biodone: zero vnode ref count"); 2863 } 2864 2865 if (error) { 2866 panic("biodone: missing VM object"); 2867 } 2868 2869 if ((vp->v_flag & VOBJBUF) == 0) { 2870 panic("biodone: vnode is not setup for merged cache"); 2871 } 2872#endif 2873 2874 foff = bp->b_offset; 2875 KASSERT(bp->b_offset != NOOFFSET, 2876 ("biodone: no buffer offset")); 2877 2878 if (error) { 2879 panic("biodone: no object"); 2880 } 2881#if defined(VFS_BIO_DEBUG) 2882 if (obj->paging_in_progress < bp->b_npages) { 2883 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 2884 obj->paging_in_progress, bp->b_npages); 2885 } 2886#endif 2887 2888 /* 2889 * Set B_CACHE if the op was a normal read and no error 2890 * occured. B_CACHE is set for writes in the b*write() 2891 * routines. 2892 */ 2893 iosize = bp->b_bcount - bp->b_resid; 2894 if (bp->b_iocmd == BIO_READ && 2895 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 2896 !(bp->b_ioflags & BIO_ERROR)) { 2897 bp->b_flags |= B_CACHE; 2898 } 2899 2900 for (i = 0; i < bp->b_npages; i++) { 2901 int bogusflag = 0; 2902 int resid; 2903 2904 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2905 if (resid > iosize) 2906 resid = iosize; 2907 2908 /* 2909 * cleanup bogus pages, restoring the originals 2910 */ 2911 m = bp->b_pages[i]; 2912 if (m == bogus_page) { 2913 bogusflag = 1; 2914 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2915 if (m == NULL) 2916 panic("biodone: page disappeared!"); 2917 bp->b_pages[i] = m; 2918 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2919 } 2920#if defined(VFS_BIO_DEBUG) 2921 if (OFF_TO_IDX(foff) != m->pindex) { 2922 printf( 2923"biodone: foff(%lu)/m->pindex(%d) mismatch\n", 2924 (unsigned long)foff, m->pindex); 2925 } 2926#endif 2927 2928 /* 2929 * In the write case, the valid and clean bits are 2930 * already changed correctly ( see bdwrite() ), so we 2931 * only need to do this here in the read case. 2932 */ 2933 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 2934 vfs_page_set_valid(bp, foff, i, m); 2935 } 2936 vm_page_flag_clear(m, PG_ZERO); 2937 2938 /* 2939 * when debugging new filesystems or buffer I/O methods, this 2940 * is the most common error that pops up. if you see this, you 2941 * have not set the page busy flag correctly!!! 2942 */ 2943 if (m->busy == 0) { 2944 printf("biodone: page busy < 0, " 2945 "pindex: %d, foff: 0x(%x,%x), " 2946 "resid: %d, index: %d\n", 2947 (int) m->pindex, (int)(foff >> 32), 2948 (int) foff & 0xffffffff, resid, i); 2949 if (!vn_isdisk(vp, NULL)) 2950 printf(" iosize: %ld, lblkno: %lld, flags: 0x%lx, npages: %d\n", 2951 bp->b_vp->v_mount->mnt_stat.f_iosize, 2952 (intmax_t) bp->b_lblkno, 2953 bp->b_flags, bp->b_npages); 2954 else 2955 printf(" VDEV, lblkno: %lld, flags: 0x%lx, npages: %d\n", 2956 (intmax_t) bp->b_lblkno, 2957 bp->b_flags, bp->b_npages); 2958 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2959 m->valid, m->dirty, m->wire_count); 2960 panic("biodone: page busy < 0\n"); 2961 } 2962 vm_page_io_finish(m); 2963 vm_object_pip_subtract(obj, 1); 2964 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2965 iosize -= resid; 2966 } 2967 if (obj) 2968 vm_object_pip_wakeupn(obj, 0); 2969 } 2970 2971 /* 2972 * For asynchronous completions, release the buffer now. The brelse 2973 * will do a wakeup there if necessary - so no need to do a wakeup 2974 * here in the async case. The sync case always needs to do a wakeup. 2975 */ 2976 2977 if (bp->b_flags & B_ASYNC) { 2978 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 2979 brelse(bp); 2980 else 2981 bqrelse(bp); 2982 } else { 2983 wakeup(bp); 2984 } 2985 splx(s); 2986} 2987 2988/* 2989 * This routine is called in lieu of iodone in the case of 2990 * incomplete I/O. This keeps the busy status for pages 2991 * consistant. 2992 */ 2993void 2994vfs_unbusy_pages(struct buf * bp) 2995{ 2996 int i; 2997 2998 GIANT_REQUIRED; 2999 3000 runningbufwakeup(bp); 3001 if (bp->b_flags & B_VMIO) { 3002 struct vnode *vp = bp->b_vp; 3003 vm_object_t obj; 3004 3005 VOP_GETVOBJECT(vp, &obj); 3006 3007 for (i = 0; i < bp->b_npages; i++) { 3008 vm_page_t m = bp->b_pages[i]; 3009 3010 if (m == bogus_page) { 3011 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3012 if (!m) { 3013 panic("vfs_unbusy_pages: page missing\n"); 3014 } 3015 bp->b_pages[i] = m; 3016 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3017 } 3018 vm_object_pip_subtract(obj, 1); 3019 vm_page_flag_clear(m, PG_ZERO); 3020 vm_page_io_finish(m); 3021 } 3022 vm_object_pip_wakeupn(obj, 0); 3023 } 3024} 3025 3026/* 3027 * vfs_page_set_valid: 3028 * 3029 * Set the valid bits in a page based on the supplied offset. The 3030 * range is restricted to the buffer's size. 3031 * 3032 * This routine is typically called after a read completes. 3033 */ 3034static void 3035vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3036{ 3037 vm_ooffset_t soff, eoff; 3038 3039 GIANT_REQUIRED; 3040 /* 3041 * Start and end offsets in buffer. eoff - soff may not cross a 3042 * page boundry or cross the end of the buffer. The end of the 3043 * buffer, in this case, is our file EOF, not the allocation size 3044 * of the buffer. 3045 */ 3046 soff = off; 3047 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3048 if (eoff > bp->b_offset + bp->b_bcount) 3049 eoff = bp->b_offset + bp->b_bcount; 3050 3051 /* 3052 * Set valid range. This is typically the entire buffer and thus the 3053 * entire page. 3054 */ 3055 if (eoff > soff) { 3056 vm_page_set_validclean( 3057 m, 3058 (vm_offset_t) (soff & PAGE_MASK), 3059 (vm_offset_t) (eoff - soff) 3060 ); 3061 } 3062} 3063 3064/* 3065 * This routine is called before a device strategy routine. 3066 * It is used to tell the VM system that paging I/O is in 3067 * progress, and treat the pages associated with the buffer 3068 * almost as being PG_BUSY. Also the object paging_in_progress 3069 * flag is handled to make sure that the object doesn't become 3070 * inconsistant. 3071 * 3072 * Since I/O has not been initiated yet, certain buffer flags 3073 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3074 * and should be ignored. 3075 */ 3076void 3077vfs_busy_pages(struct buf * bp, int clear_modify) 3078{ 3079 int i, bogus; 3080 3081 GIANT_REQUIRED; 3082 3083 if (bp->b_flags & B_VMIO) { 3084 struct vnode *vp = bp->b_vp; 3085 vm_object_t obj; 3086 vm_ooffset_t foff; 3087 3088 VOP_GETVOBJECT(vp, &obj); 3089 foff = bp->b_offset; 3090 KASSERT(bp->b_offset != NOOFFSET, 3091 ("vfs_busy_pages: no buffer offset")); 3092 vfs_setdirty(bp); 3093 3094retry: 3095 for (i = 0; i < bp->b_npages; i++) { 3096 vm_page_t m = bp->b_pages[i]; 3097 if (vm_page_sleep_busy(m, FALSE, "vbpage")) 3098 goto retry; 3099 } 3100 3101 bogus = 0; 3102 for (i = 0; i < bp->b_npages; i++) { 3103 vm_page_t m = bp->b_pages[i]; 3104 3105 vm_page_flag_clear(m, PG_ZERO); 3106 if ((bp->b_flags & B_CLUSTER) == 0) { 3107 vm_object_pip_add(obj, 1); 3108 vm_page_io_start(m); 3109 } 3110 3111 /* 3112 * When readying a buffer for a read ( i.e 3113 * clear_modify == 0 ), it is important to do 3114 * bogus_page replacement for valid pages in 3115 * partially instantiated buffers. Partially 3116 * instantiated buffers can, in turn, occur when 3117 * reconstituting a buffer from its VM backing store 3118 * base. We only have to do this if B_CACHE is 3119 * clear ( which causes the I/O to occur in the 3120 * first place ). The replacement prevents the read 3121 * I/O from overwriting potentially dirty VM-backed 3122 * pages. XXX bogus page replacement is, uh, bogus. 3123 * It may not work properly with small-block devices. 3124 * We need to find a better way. 3125 */ 3126 3127 vm_page_protect(m, VM_PROT_NONE); 3128 if (clear_modify) 3129 vfs_page_set_valid(bp, foff, i, m); 3130 else if (m->valid == VM_PAGE_BITS_ALL && 3131 (bp->b_flags & B_CACHE) == 0) { 3132 bp->b_pages[i] = bogus_page; 3133 bogus++; 3134 } 3135 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3136 } 3137 if (bogus) 3138 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3139 } 3140} 3141 3142/* 3143 * Tell the VM system that the pages associated with this buffer 3144 * are clean. This is used for delayed writes where the data is 3145 * going to go to disk eventually without additional VM intevention. 3146 * 3147 * Note that while we only really need to clean through to b_bcount, we 3148 * just go ahead and clean through to b_bufsize. 3149 */ 3150static void 3151vfs_clean_pages(struct buf * bp) 3152{ 3153 int i; 3154 3155 GIANT_REQUIRED; 3156 3157 if (bp->b_flags & B_VMIO) { 3158 vm_ooffset_t foff; 3159 3160 foff = bp->b_offset; 3161 KASSERT(bp->b_offset != NOOFFSET, 3162 ("vfs_clean_pages: no buffer offset")); 3163 for (i = 0; i < bp->b_npages; i++) { 3164 vm_page_t m = bp->b_pages[i]; 3165 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3166 vm_ooffset_t eoff = noff; 3167 3168 if (eoff > bp->b_offset + bp->b_bufsize) 3169 eoff = bp->b_offset + bp->b_bufsize; 3170 vfs_page_set_valid(bp, foff, i, m); 3171 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3172 foff = noff; 3173 } 3174 } 3175} 3176 3177/* 3178 * vfs_bio_set_validclean: 3179 * 3180 * Set the range within the buffer to valid and clean. The range is 3181 * relative to the beginning of the buffer, b_offset. Note that b_offset 3182 * itself may be offset from the beginning of the first page. 3183 * 3184 */ 3185 3186void 3187vfs_bio_set_validclean(struct buf *bp, int base, int size) 3188{ 3189 if (bp->b_flags & B_VMIO) { 3190 int i; 3191 int n; 3192 3193 /* 3194 * Fixup base to be relative to beginning of first page. 3195 * Set initial n to be the maximum number of bytes in the 3196 * first page that can be validated. 3197 */ 3198 3199 base += (bp->b_offset & PAGE_MASK); 3200 n = PAGE_SIZE - (base & PAGE_MASK); 3201 3202 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3203 vm_page_t m = bp->b_pages[i]; 3204 3205 if (n > size) 3206 n = size; 3207 3208 vm_page_set_validclean(m, base & PAGE_MASK, n); 3209 base += n; 3210 size -= n; 3211 n = PAGE_SIZE; 3212 } 3213 } 3214} 3215 3216/* 3217 * vfs_bio_clrbuf: 3218 * 3219 * clear a buffer. This routine essentially fakes an I/O, so we need 3220 * to clear BIO_ERROR and B_INVAL. 3221 * 3222 * Note that while we only theoretically need to clear through b_bcount, 3223 * we go ahead and clear through b_bufsize. 3224 */ 3225 3226void 3227vfs_bio_clrbuf(struct buf *bp) 3228{ 3229 int i, mask = 0; 3230 caddr_t sa, ea; 3231 3232 GIANT_REQUIRED; 3233 3234 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 3235 bp->b_flags &= ~B_INVAL; 3236 bp->b_ioflags &= ~BIO_ERROR; 3237 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3238 (bp->b_offset & PAGE_MASK) == 0) { 3239 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3240 if ((bp->b_pages[0]->valid & mask) == mask) { 3241 bp->b_resid = 0; 3242 return; 3243 } 3244 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3245 ((bp->b_pages[0]->valid & mask) == 0)) { 3246 bzero(bp->b_data, bp->b_bufsize); 3247 bp->b_pages[0]->valid |= mask; 3248 bp->b_resid = 0; 3249 return; 3250 } 3251 } 3252 ea = sa = bp->b_data; 3253 for(i=0;i<bp->b_npages;i++,sa=ea) { 3254 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3255 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3256 ea = (caddr_t)(vm_offset_t)ulmin( 3257 (u_long)(vm_offset_t)ea, 3258 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3259 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3260 if ((bp->b_pages[i]->valid & mask) == mask) 3261 continue; 3262 if ((bp->b_pages[i]->valid & mask) == 0) { 3263 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 3264 bzero(sa, ea - sa); 3265 } 3266 } else { 3267 for (; sa < ea; sa += DEV_BSIZE, j++) { 3268 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3269 (bp->b_pages[i]->valid & (1<<j)) == 0) 3270 bzero(sa, DEV_BSIZE); 3271 } 3272 } 3273 bp->b_pages[i]->valid |= mask; 3274 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 3275 } 3276 bp->b_resid = 0; 3277 } else { 3278 clrbuf(bp); 3279 } 3280} 3281 3282/* 3283 * vm_hold_load_pages and vm_hold_free_pages get pages into 3284 * a buffers address space. The pages are anonymous and are 3285 * not associated with a file object. 3286 */ 3287static void 3288vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3289{ 3290 vm_offset_t pg; 3291 vm_page_t p; 3292 int index; 3293 3294 GIANT_REQUIRED; 3295 3296 to = round_page(to); 3297 from = round_page(from); 3298 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3299 3300 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3301tryagain: 3302 /* 3303 * note: must allocate system pages since blocking here 3304 * could intefere with paging I/O, no matter which 3305 * process we are. 3306 */ 3307 p = vm_page_alloc(kernel_object, 3308 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3309 VM_ALLOC_SYSTEM); 3310 if (!p) { 3311 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 3312 VM_WAIT; 3313 goto tryagain; 3314 } 3315 vm_page_wire(p); 3316 p->valid = VM_PAGE_BITS_ALL; 3317 vm_page_flag_clear(p, PG_ZERO); 3318 pmap_qenter(pg, &p, 1); 3319 bp->b_pages[index] = p; 3320 vm_page_wakeup(p); 3321 } 3322 bp->b_npages = index; 3323} 3324 3325/* Return pages associated with this buf to the vm system */ 3326void 3327vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 3328{ 3329 vm_offset_t pg; 3330 vm_page_t p; 3331 int index, newnpages; 3332 3333 GIANT_REQUIRED; 3334 3335 from = round_page(from); 3336 to = round_page(to); 3337 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3338 3339 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3340 p = bp->b_pages[index]; 3341 if (p && (index < bp->b_npages)) { 3342 if (p->busy) { 3343 printf( 3344 "vm_hold_free_pages: blkno: %lld, lblkno: %lld\n", 3345 (intmax_t)bp->b_blkno, 3346 (intmax_t)bp->b_lblkno); 3347 } 3348 bp->b_pages[index] = NULL; 3349 pmap_qremove(pg, 1); 3350 vm_page_busy(p); 3351 vm_page_unwire(p, 0); 3352 vm_page_free(p); 3353 } 3354 } 3355 bp->b_npages = newnpages; 3356} 3357 3358 3359#include "opt_ddb.h" 3360#ifdef DDB 3361#include <ddb/ddb.h> 3362 3363/* DDB command to show buffer data */ 3364DB_SHOW_COMMAND(buffer, db_show_buffer) 3365{ 3366 /* get args */ 3367 struct buf *bp = (struct buf *)addr; 3368 3369 if (!have_addr) { 3370 db_printf("usage: show buffer <addr>\n"); 3371 return; 3372 } 3373 3374 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3375 db_printf( 3376 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3377 "b_dev = (%d,%d), b_data = %p, b_blkno = %lld, b_pblkno = %lld\n", 3378 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3379 major(bp->b_dev), minor(bp->b_dev), bp->b_data, 3380 (intmax_t)bp->b_blkno, (intmax_t)bp->b_pblkno); 3381 if (bp->b_npages) { 3382 int i; 3383 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3384 for (i = 0; i < bp->b_npages; i++) { 3385 vm_page_t m; 3386 m = bp->b_pages[i]; 3387 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3388 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3389 if ((i + 1) < bp->b_npages) 3390 db_printf(","); 3391 } 3392 db_printf("\n"); 3393 } 3394} 3395#endif /* DDB */ 3396