vfs_bio.c revision 140056
1118611Snjl/*- 2118611Snjl * Copyright (c) 2004 Poul-Henning Kamp 3118611Snjl * Copyright (c) 1994,1997 John S. Dyson 4151937Sjkim * All rights reserved. 5118611Snjl * 6118611Snjl * Redistribution and use in source and binary forms, with or without 7118611Snjl * modification, are permitted provided that the following conditions 8217365Sjkim * are met: 9217365Sjkim * 1. Redistributions of source code must retain the above copyright 10118611Snjl * notice, this list of conditions and the following disclaimer. 11118611Snjl * 2. Redistributions in binary form must reproduce the above copyright 12217365Sjkim * notice, this list of conditions and the following disclaimer in the 13217365Sjkim * documentation and/or other materials provided with the distribution. 14217365Sjkim * 15217365Sjkim * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16217365Sjkim * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17217365Sjkim * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18217365Sjkim * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19217365Sjkim * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20217365Sjkim * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21217365Sjkim * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22217365Sjkim * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23217365Sjkim * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24217365Sjkim * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25217365Sjkim * SUCH DAMAGE. 26118611Snjl */ 27217365Sjkim 28217365Sjkim/* 29217365Sjkim * this file contains a new buffer I/O scheme implementing a coherent 30118611Snjl * VM object and buffer cache scheme. Pains have been taken to make 31217365Sjkim * sure that the performance degradation associated with schemes such 32217365Sjkim * as this is not realized. 33217365Sjkim * 34217365Sjkim * Author: John S. Dyson 35217365Sjkim * Significant help during the development and debugging phases 36217365Sjkim * had been provided by David Greenman, also of the FreeBSD core team. 37217365Sjkim * 38217365Sjkim * see man buf(9) for more info. 39217365Sjkim */ 40217365Sjkim 41217365Sjkim#include <sys/cdefs.h> 42217365Sjkim__FBSDID("$FreeBSD: head/sys/kern/vfs_bio.c 140056 2005-01-11 10:43:08Z phk $"); 43217365Sjkim 44118611Snjl#include <sys/param.h> 45118611Snjl#include <sys/systm.h> 46118611Snjl#include <sys/bio.h> 47118611Snjl#include <sys/conf.h> 48118611Snjl#include <sys/buf.h> 49217365Sjkim#include <sys/devicestat.h> 50217365Sjkim#include <sys/eventhandler.h> 51217365Sjkim#include <sys/lock.h> 52217365Sjkim#include <sys/malloc.h> 53118611Snjl#include <sys/mount.h> 54118611Snjl#include <sys/mutex.h> 55118611Snjl#include <sys/kernel.h> 56118611Snjl#include <sys/kthread.h> 57118611Snjl#include <sys/proc.h> 58118611Snjl#include <sys/resourcevar.h> 59118611Snjl#include <sys/sysctl.h> 60118611Snjl#include <sys/vmmeter.h> 61118611Snjl#include <sys/vnode.h> 62118611Snjl#include <geom/geom.h> 63118611Snjl#include <vm/vm.h> 64118611Snjl#include <vm/vm_param.h> 65118611Snjl#include <vm/vm_kern.h> 66118611Snjl#include <vm/vm_pageout.h> 67118611Snjl#include <vm/vm_page.h> 68118611Snjl#include <vm/vm_object.h> 69118611Snjl#include <vm/vm_extern.h> 70118611Snjl#include <vm/vm_map.h> 71118611Snjl#include "opt_directio.h" 72151937Sjkim#include "opt_swap.h" 73118611Snjl 74151937Sjkimstatic MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 75151937Sjkim 76213806Sjkimstruct bio_ops bioops; /* I/O operation notification */ 77151937Sjkim 78118611Snjlstruct buf_ops buf_ops_bio = { 79118611Snjl .bop_name = "buf_ops_bio", 80118611Snjl .bop_write = bufwrite, 81118611Snjl .bop_strategy = bufstrategy, 82118611Snjl .bop_sync = bufsync, 83118611Snjl}; 84118611Snjl 85118611Snjl/* 86151937Sjkim * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 87151937Sjkim * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 88151937Sjkim */ 89118611Snjlstruct buf *buf; /* buffer header pool */ 90118611Snjl 91118611Snjlstatic struct proc *bufdaemonproc; 92118611Snjl 93118611Snjlstatic int inmem(struct vnode *vp, daddr_t blkno); 94118611Snjlstatic void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 95118611Snjl vm_offset_t to); 96118611Snjlstatic void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 97118611Snjl vm_offset_t to); 98118611Snjlstatic void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 99118611Snjl int pageno, vm_page_t m); 100118611Snjlstatic void vfs_clean_pages(struct buf *bp); 101118611Snjlstatic void vfs_setdirty(struct buf *bp); 102118611Snjlstatic void vfs_vmio_release(struct buf *bp); 103118611Snjlstatic void vfs_backgroundwritedone(struct buf *bp); 104118611Snjlstatic int vfs_bio_clcheck(struct vnode *vp, int size, 105118611Snjl daddr_t lblkno, daddr_t blkno); 106118611Snjlstatic int flushbufqueues(int flushdeps); 107118611Snjlstatic void buf_daemon(void); 108118611Snjlvoid bremfreel(struct buf *bp); 109118611Snjl 110118611Snjlint vmiodirenable = TRUE; 111118611SnjlSYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 112118611Snjl "Use the VM system for directory writes"); 113118611Snjlint runningbufspace; 114118611SnjlSYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 115118611Snjl "Amount of presently outstanding async buffer io"); 116118611Snjlstatic int bufspace; 117118611SnjlSYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 118193529Sjkim "KVA memory used for bufs"); 119209746Sjkimstatic int maxbufspace; 120193529SjkimSYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 121209746Sjkim "Maximum allowed value of bufspace (including buf_daemon)"); 122209746Sjkimstatic int bufmallocspace; 123209746SjkimSYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 124209746Sjkim "Amount of malloced memory for buffers"); 125210976Sjkimstatic int maxbufmallocspace; 126210976SjkimSYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 127210976Sjkim "Maximum amount of malloced memory for buffers"); 128210976Sjkimstatic int lobufspace; 129193529SjkimSYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 130193529Sjkim "Minimum amount of buffers we want to have"); 131210976Sjkimstatic int hibufspace; 132210976SjkimSYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 133118611Snjl "Maximum allowed value of bufspace (excluding buf_daemon)"); 134193529Sjkimstatic int bufreusecnt; 135193529SjkimSYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 136193529Sjkim "Number of times we have reused a buffer"); 137193529Sjkimstatic int buffreekvacnt; 138151937SjkimSYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 139151937Sjkim "Number of times we have freed the KVA space from some buffer"); 140151937Sjkimstatic int bufdefragcnt; 141118611SnjlSYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 142118611Snjl "Number of times we have had to repeat buffer allocation to defragment"); 143118611Snjlstatic int lorunningspace; 144118611SnjlSYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 145118611Snjl "Minimum preferred space used for in-progress I/O"); 146118611Snjlstatic int hirunningspace; 147118611SnjlSYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 148118611Snjl "Maximum amount of space to use for in-progress I/O"); 149151937Sjkimstatic int dirtybufferflushes; 150151937SjkimSYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 151151937Sjkim 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 152151937Sjkimstatic int altbufferflushes; 153118611SnjlSYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 154151937Sjkim 0, "Number of fsync flushes to limit dirty buffers"); 155151937Sjkimstatic int recursiveflushes; 156118611SnjlSYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 157151937Sjkim 0, "Number of flushes skipped due to being recursive"); 158151937Sjkimstatic int numdirtybuffers; 159151937SjkimSYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 160118611Snjl "Number of buffers that are dirty (has unwritten changes) at the moment"); 161209746Sjkimstatic int lodirtybuffers; 162209746SjkimSYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 163209746Sjkim "How many buffers we want to have free before bufdaemon can sleep"); 164118611Snjlstatic int hidirtybuffers; 165209746SjkimSYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 166151937Sjkim "When the number of dirty buffers is considered severe"); 167218590Sjkimstatic int dirtybufthresh; 168151937SjkimSYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 169151937Sjkim 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 170151937Sjkimstatic int numfreebuffers; 171151937SjkimSYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 172151937Sjkim "Number of free buffers"); 173151937Sjkimstatic int lofreebuffers; 174118611SnjlSYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 175151937Sjkim "XXX Unused"); 176151937Sjkimstatic int hifreebuffers; 177151937SjkimSYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 178151937Sjkim "XXX Complicatedly unused"); 179151937Sjkimstatic int getnewbufcalls; 180118611SnjlSYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 181151937Sjkim "Number of calls to getnewbuf"); 182151937Sjkimstatic int getnewbufrestarts; 183151937SjkimSYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 184151937Sjkim "Number of times getnewbuf has had to restart a buffer aquisition"); 185151937Sjkimstatic int dobkgrdwrite = 1; 186118611SnjlSYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 187151937Sjkim "Do background writes (honoring the BV_BKGRDWRITE flag)?"); 188151937Sjkim 189151937Sjkim/* 190151937Sjkim * Wakeup point for bufdaemon, as well as indicator of whether it is already 191151937Sjkim * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 192118611Snjl * is idling. 193151937Sjkim */ 194151937Sjkimstatic int bd_request; 195151937Sjkim 196151937Sjkim/* 197151937Sjkim * This lock synchronizes access to bd_request. 198118611Snjl */ 199151937Sjkimstatic struct mtx bdlock; 200218590Sjkim 201151937Sjkim/* 202151937Sjkim * bogus page -- for I/O to/from partially complete buffers 203151937Sjkim * this is a temporary solution to the problem, but it is not 204118611Snjl * really that bad. it would be better to split the buffer 205218590Sjkim * for input in the case of buffers partially already in memory, 206218590Sjkim * but the code is intricate enough already. 207218590Sjkim */ 208218590Sjkimvm_page_t bogus_page; 209218590Sjkim 210218590Sjkim/* 211218590Sjkim * Synchronization (sleep/wakeup) variable for active buffer space requests. 212218590Sjkim * Set when wait starts, cleared prior to wakeup(). 213218590Sjkim * Used in runningbufwakeup() and waitrunningbufspace(). 214218590Sjkim */ 215218590Sjkimstatic int runningbufreq; 216218590Sjkim 217218590Sjkim/* 218218590Sjkim * This lock protects the runningbufreq and synchronizes runningbufwakeup and 219218590Sjkim * waitrunningbufspace(). 220218590Sjkim */ 221218590Sjkimstatic struct mtx rbreqlock; 222218590Sjkim 223218590Sjkim/* 224218590Sjkim * Synchronization (sleep/wakeup) variable for buffer requests. 225218590Sjkim * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 226218590Sjkim * by and/or. 227218590Sjkim * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 228218590Sjkim * getnewbuf(), and getblk(). 229218590Sjkim */ 230218590Sjkimstatic int needsbuffer; 231218590Sjkim 232151937Sjkim/* 233218590Sjkim * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 234118611Snjl */ 235218590Sjkimstatic struct mtx nblock; 236151937Sjkim 237218590Sjkim/* 238218590Sjkim * Lock that protects against bwait()/bdone()/B_DONE races. 239218590Sjkim */ 240218590Sjkim 241218590Sjkimstatic struct mtx bdonelock; 242218590Sjkim 243218590Sjkim/* 244218590Sjkim * Definitions for the buffer free lists. 245218590Sjkim */ 246218590Sjkim#define BUFFER_QUEUES 5 /* number of free buffer queues */ 247218590Sjkim 248218590Sjkim#define QUEUE_NONE 0 /* on no queue */ 249218590Sjkim#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 250218590Sjkim#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 251218590Sjkim#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 252218590Sjkim#define QUEUE_EMPTY 4 /* empty buffer headers */ 253218590Sjkim 254218590Sjkim/* Queues for free buffers with various properties */ 255218590Sjkimstatic TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 256218590Sjkim 257218590Sjkim/* Lock for the bufqueues */ 258218590Sjkimstatic struct mtx bqlock; 259218590Sjkim 260218590Sjkim/* 261218590Sjkim * Single global constant for BUF_WMESG, to avoid getting multiple references. 262218590Sjkim * buf_wmesg is referred from macros. 263218590Sjkim */ 264218590Sjkimconst char *buf_wmesg = BUF_WMESG; 265218590Sjkim 266218590Sjkim#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 267218590Sjkim#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 268218590Sjkim#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 269151937Sjkim#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 270151937Sjkim 271151937Sjkim#ifdef DIRECTIO 272118611Snjlextern void ffs_rawread_setup(void); 273118611Snjl#endif /* DIRECTIO */ 274118611Snjl/* 275118611Snjl * numdirtywakeup: 276118611Snjl * 277118611Snjl * If someone is blocked due to there being too many dirty buffers, 278118611Snjl * and numdirtybuffers is now reasonable, wake them up. 279118611Snjl */ 280118611Snjl 281118611Snjlstatic __inline void 282118611Snjlnumdirtywakeup(int level) 283118611Snjl{ 284118611Snjl 285118611Snjl if (numdirtybuffers <= level) { 286151937Sjkim mtx_lock(&nblock); 287151937Sjkim if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 288151937Sjkim needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 289151937Sjkim wakeup(&needsbuffer); 290118611Snjl } 291118611Snjl mtx_unlock(&nblock); 292118611Snjl } 293118611Snjl} 294118611Snjl 295118611Snjl/* 296118611Snjl * bufspacewakeup: 297118611Snjl * 298118611Snjl * Called when buffer space is potentially available for recovery. 299118611Snjl * getnewbuf() will block on this flag when it is unable to free 300118611Snjl * sufficient buffer space. Buffer space becomes recoverable when 301118611Snjl * bp's get placed back in the queues. 302118611Snjl */ 303118611Snjl 304118611Snjlstatic __inline void 305118611Snjlbufspacewakeup(void) 306118611Snjl{ 307118611Snjl 308118611Snjl /* 309118611Snjl * If someone is waiting for BUF space, wake them up. Even 310118611Snjl * though we haven't freed the kva space yet, the waiting 311193529Sjkim * process will be able to now. 312193529Sjkim */ 313193529Sjkim mtx_lock(&nblock); 314193529Sjkim if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 315167802Sjkim needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 316118611Snjl wakeup(&needsbuffer); 317167802Sjkim } 318118611Snjl mtx_unlock(&nblock); 319118611Snjl} 320151937Sjkim 321151937Sjkim/* 322151937Sjkim * runningbufwakeup() - in-progress I/O accounting. 323118611Snjl * 324151937Sjkim */ 325151937Sjkimstatic __inline void 326118611Snjlrunningbufwakeup(struct buf *bp) 327118611Snjl{ 328118611Snjl 329118611Snjl if (bp->b_runningbufspace) { 330118611Snjl atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 331118611Snjl bp->b_runningbufspace = 0; 332118611Snjl mtx_lock(&rbreqlock); 333118611Snjl if (runningbufreq && runningbufspace <= lorunningspace) { 334118611Snjl runningbufreq = 0; 335118611Snjl wakeup(&runningbufreq); 336118611Snjl } 337118611Snjl mtx_unlock(&rbreqlock); 338118611Snjl } 339118611Snjl} 340167802Sjkim 341167802Sjkim/* 342167802Sjkim * bufcountwakeup: 343118611Snjl * 344151937Sjkim * Called when a buffer has been added to one of the free queues to 345151937Sjkim * account for the buffer and to wakeup anyone waiting for free buffers. 346151937Sjkim * This typically occurs when large amounts of metadata are being handled 347151937Sjkim * by the buffer cache ( else buffer space runs out first, usually ). 348151937Sjkim */ 349151937Sjkim 350151937Sjkimstatic __inline void 351151937Sjkimbufcountwakeup(void) 352118611Snjl{ 353118611Snjl 354118611Snjl atomic_add_int(&numfreebuffers, 1); 355118611Snjl mtx_lock(&nblock); 356118611Snjl if (needsbuffer) { 357118611Snjl needsbuffer &= ~VFS_BIO_NEED_ANY; 358118611Snjl if (numfreebuffers >= hifreebuffers) 359118611Snjl needsbuffer &= ~VFS_BIO_NEED_FREE; 360118611Snjl wakeup(&needsbuffer); 361118611Snjl } 362118611Snjl mtx_unlock(&nblock); 363118611Snjl} 364151937Sjkim 365118611Snjl/* 366118611Snjl * waitrunningbufspace() 367118611Snjl * 368118611Snjl * runningbufspace is a measure of the amount of I/O currently 369118611Snjl * running. This routine is used in async-write situations to 370118611Snjl * prevent creating huge backups of pending writes to a device. 371118611Snjl * Only asynchronous writes are governed by this function. 372118611Snjl * 373118611Snjl * Reads will adjust runningbufspace, but will not block based on it. 374118611Snjl * The read load has a side effect of reducing the allowed write load. 375118611Snjl * 376118611Snjl * This does NOT turn an async write into a sync write. It waits 377118611Snjl * for earlier writes to complete and generally returns before the 378118611Snjl * caller's write has reached the device. 379118611Snjl */ 380118611Snjlstatic __inline void 381151937Sjkimwaitrunningbufspace(void) 382118611Snjl{ 383118611Snjl 384118611Snjl mtx_lock(&rbreqlock); 385151937Sjkim while (runningbufspace > hirunningspace) { 386151937Sjkim ++runningbufreq; 387151937Sjkim msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 388151937Sjkim } 389118611Snjl mtx_unlock(&rbreqlock); 390118611Snjl} 391118611Snjl 392118611Snjl 393118611Snjl/* 394118611Snjl * vfs_buf_test_cache: 395118611Snjl * 396118611Snjl * Called when a buffer is extended. This function clears the B_CACHE 397118611Snjl * bit if the newly extended portion of the buffer does not contain 398118611Snjl * valid data. 399118611Snjl */ 400118611Snjlstatic __inline 401118611Snjlvoid 402118611Snjlvfs_buf_test_cache(struct buf *bp, 403118611Snjl vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 404118611Snjl vm_page_t m) 405118611Snjl{ 406118611Snjl 407118611Snjl GIANT_REQUIRED; 408118611Snjl 409118611Snjl VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 410118611Snjl if (bp->b_flags & B_CACHE) { 411118611Snjl int base = (foff + off) & PAGE_MASK; 412151937Sjkim if (vm_page_is_valid(m, base, size) == 0) 413118611Snjl bp->b_flags &= ~B_CACHE; 414118611Snjl } 415151937Sjkim} 416151937Sjkim 417118611Snjl/* Wake up the buffer deamon if necessary */ 418118611Snjlstatic __inline 419151937Sjkimvoid 420209746Sjkimbd_wakeup(int dirtybuflevel) 421209746Sjkim{ 422209746Sjkim 423209746Sjkim mtx_lock(&bdlock); 424209746Sjkim if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 425209746Sjkim bd_request = 1; 426209746Sjkim wakeup(&bd_request); 427209746Sjkim } 428209746Sjkim mtx_unlock(&bdlock); 429209746Sjkim} 430151937Sjkim 431151937Sjkim/* 432118611Snjl * bd_speedup - speedup the buffer cache flushing code 433118611Snjl */ 434118611Snjl 435118611Snjlstatic __inline 436118611Snjlvoid 437118611Snjlbd_speedup(void) 438118611Snjl{ 439118611Snjl 440118611Snjl bd_wakeup(1); 441118611Snjl} 442118611Snjl 443118611Snjl/* 444118611Snjl * Calculating buffer cache scaling values and reserve space for buffer 445151937Sjkim * headers. This is called during low level kernel initialization and 446151937Sjkim * may be called more then once. We CANNOT write to the memory area 447118611Snjl * being reserved at this time. 448118611Snjl */ 449118611Snjlcaddr_t 450151937Sjkimkern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 451118611Snjl{ 452118611Snjl 453118611Snjl /* 454118611Snjl * physmem_est is in pages. Convert it to kilobytes (assumes 455118611Snjl * PAGE_SIZE is >= 1K) 456204773Sjkim */ 457204773Sjkim physmem_est = physmem_est * (PAGE_SIZE / 1024); 458204773Sjkim 459204773Sjkim /* 460151937Sjkim * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 461204773Sjkim * For the first 64MB of ram nominally allocate sufficient buffers to 462204773Sjkim * cover 1/4 of our ram. Beyond the first 64MB allocate additional 463204773Sjkim * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 464204773Sjkim * the buffer cache we limit the eventual kva reservation to 465206117Sjkim * maxbcache bytes. 466206117Sjkim * 467206117Sjkim * factor represents the 1/4 x ram conversion. 468206117Sjkim */ 469206117Sjkim if (nbuf == 0) { 470204773Sjkim int factor = 4 * BKVASIZE / 1024; 471204773Sjkim 472204773Sjkim nbuf = 50; 473204773Sjkim if (physmem_est > 4096) 474204773Sjkim nbuf += min((physmem_est - 4096) / factor, 475204773Sjkim 65536 / factor); 476204773Sjkim if (physmem_est > 65536) 477204773Sjkim nbuf += (physmem_est - 65536) * 2 / (factor * 5); 478204773Sjkim 479204773Sjkim if (maxbcache && nbuf > maxbcache / BKVASIZE) 480204773Sjkim nbuf = maxbcache / BKVASIZE; 481204773Sjkim } 482151937Sjkim 483151937Sjkim#if 0 484151937Sjkim /* 485118611Snjl * Do not allow the buffer_map to be more then 1/2 the size of the 486118611Snjl * kernel_map. 487118611Snjl */ 488118611Snjl if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 489118611Snjl (BKVASIZE * 2)) { 490118611Snjl nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 491118611Snjl (BKVASIZE * 2); 492118611Snjl printf("Warning: nbufs capped at %d\n", nbuf); 493118611Snjl } 494118611Snjl#endif 495118611Snjl 496118611Snjl /* 497118611Snjl * swbufs are used as temporary holders for I/O, such as paging I/O. 498118611Snjl * We have no less then 16 and no more then 256. 499118611Snjl */ 500118611Snjl nswbuf = max(min(nbuf/4, 256), 16); 501118611Snjl#ifdef NSWBUF_MIN 502118611Snjl if (nswbuf < NSWBUF_MIN) 503118611Snjl nswbuf = NSWBUF_MIN; 504118611Snjl#endif 505118611Snjl#ifdef DIRECTIO 506118611Snjl ffs_rawread_setup(); 507118611Snjl#endif 508118611Snjl 509118611Snjl /* 510118611Snjl * Reserve space for the buffer cache buffers 511118611Snjl */ 512118611Snjl swbuf = (void *)v; 513151937Sjkim v = (caddr_t)(swbuf + nswbuf); 514151937Sjkim buf = (void *)v; 515151937Sjkim v = (caddr_t)(buf + nbuf); 516118611Snjl 517151937Sjkim return(v); 518151937Sjkim} 519151937Sjkim 520151937Sjkim/* Initialize the buffer subsystem. Called before use of any buffers. */ 521118611Snjlvoid 522118611Snjlbufinit(void) 523118611Snjl{ 524118611Snjl struct buf *bp; 525118611Snjl int i; 526118611Snjl 527118611Snjl GIANT_REQUIRED; 528118611Snjl 529118611Snjl mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 530118611Snjl mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 531118611Snjl mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 532118611Snjl mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 533118611Snjl mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 534118611Snjl 535118611Snjl /* next, make a null set of free lists */ 536118611Snjl for (i = 0; i < BUFFER_QUEUES; i++) 537118611Snjl TAILQ_INIT(&bufqueues[i]); 538118611Snjl 539202771Sjkim /* finally, initialize each buffer header and stick on empty q */ 540118611Snjl for (i = 0; i < nbuf; i++) { 541118611Snjl bp = &buf[i]; 542218590Sjkim bzero(bp, sizeof *bp); 543218590Sjkim bp->b_flags = B_INVAL; /* we're just an empty header */ 544218590Sjkim bp->b_rcred = NOCRED; 545218590Sjkim bp->b_wcred = NOCRED; 546118611Snjl bp->b_qindex = QUEUE_EMPTY; 547118611Snjl bp->b_vflags = 0; 548118611Snjl bp->b_xflags = 0; 549118611Snjl LIST_INIT(&bp->b_dep); 550118611Snjl BUF_LOCKINIT(bp); 551118611Snjl TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 552118611Snjl } 553118611Snjl 554118611Snjl /* 555118611Snjl * maxbufspace is the absolute maximum amount of buffer space we are 556151937Sjkim * allowed to reserve in KVM and in real terms. The absolute maximum 557151937Sjkim * is nominally used by buf_daemon. hibufspace is the nominal maximum 558118611Snjl * used by most other processes. The differential is required to 559118611Snjl * ensure that buf_daemon is able to run when other processes might 560118611Snjl * be blocked waiting for buffer space. 561118611Snjl * 562118611Snjl * maxbufspace is based on BKVASIZE. Allocating buffers larger then 563118611Snjl * this may result in KVM fragmentation which is not handled optimally 564118611Snjl * by the system. 565118611Snjl */ 566118611Snjl maxbufspace = nbuf * BKVASIZE; 567118611Snjl hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 568118611Snjl lobufspace = hibufspace - MAXBSIZE; 569118611Snjl 570118611Snjl lorunningspace = 512 * 1024; 571118611Snjl hirunningspace = 1024 * 1024; 572118611Snjl 573118611Snjl/* 574118611Snjl * Limit the amount of malloc memory since it is wired permanently into 575118611Snjl * the kernel space. Even though this is accounted for in the buffer 576118611Snjl * allocation, we don't want the malloced region to grow uncontrolled. 577118611Snjl * The malloc scheme improves memory utilization significantly on average 578118611Snjl * (small) directories. 579118611Snjl */ 580118611Snjl maxbufmallocspace = hibufspace / 20; 581118611Snjl 582118611Snjl/* 583118611Snjl * Reduce the chance of a deadlock occuring by limiting the number 584151937Sjkim * of delayed-write dirty buffers we allow to stack up. 585151937Sjkim */ 586118611Snjl hidirtybuffers = nbuf / 4 + 20; 587118611Snjl dirtybufthresh = hidirtybuffers * 9 / 10; 588197104Sjkim numdirtybuffers = 0; 589197104Sjkim/* 590197104Sjkim * To support extreme low-memory systems, make sure hidirtybuffers cannot 591197104Sjkim * eat up all available buffer space. This occurs when our minimum cannot 592118611Snjl * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 593118611Snjl * BKVASIZE'd (8K) buffers. 594118611Snjl */ 595118611Snjl while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 596118611Snjl hidirtybuffers >>= 1; 597118611Snjl } 598118611Snjl lodirtybuffers = hidirtybuffers / 2; 599118611Snjl 600207344Sjkim/* 601207344Sjkim * Try to keep the number of free buffers in the specified range, 602207344Sjkim * and give special processes (e.g. like buf_daemon) access to an 603207344Sjkim * emergency reserve. 604118611Snjl */ 605118611Snjl lofreebuffers = nbuf / 18 + 5; 606118611Snjl hifreebuffers = 2 * lofreebuffers; 607118611Snjl numfreebuffers = nbuf; 608118611Snjl 609118611Snjl/* 610118611Snjl * Maximum number of async ops initiated per buf_daemon loop. This is 611118611Snjl * somewhat of a hack at the moment, we really need to limit ourselves 612118611Snjl * based on the number of bytes of I/O in-transit that were initiated 613118611Snjl * from buf_daemon. 614118611Snjl */ 615118611Snjl 616118611Snjl bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 617118611Snjl VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 618118611Snjl} 619118611Snjl 620118611Snjl/* 621118611Snjl * bfreekva() - free the kva allocation for a buffer. 622118611Snjl * 623118611Snjl * Must be called at splbio() or higher as this is the only locking for 624118611Snjl * buffer_map. 625118611Snjl * 626118611Snjl * Since this call frees up buffer space, we call bufspacewakeup(). 627118611Snjl */ 628118611Snjlstatic void 629118611Snjlbfreekva(struct buf *bp) 630118611Snjl{ 631118611Snjl 632118611Snjl GIANT_REQUIRED; 633118611Snjl 634118611Snjl if (bp->b_kvasize) { 635118611Snjl atomic_add_int(&buffreekvacnt, 1); 636118611Snjl atomic_subtract_int(&bufspace, bp->b_kvasize); 637118611Snjl vm_map_delete(buffer_map, 638118611Snjl (vm_offset_t) bp->b_kvabase, 639118611Snjl (vm_offset_t) bp->b_kvabase + bp->b_kvasize 640118611Snjl ); 641118611Snjl bp->b_kvasize = 0; 642118611Snjl bufspacewakeup(); 643118611Snjl } 644118611Snjl} 645118611Snjl 646118611Snjl/* 647118611Snjl * bremfree: 648151937Sjkim * 649151937Sjkim * Mark the buffer for removal from the appropriate free list in brelse. 650151937Sjkim * 651118611Snjl */ 652118611Snjlvoid 653118611Snjlbremfree(struct buf *bp) 654118611Snjl{ 655118611Snjl 656151937Sjkim KASSERT(BUF_REFCNT(bp), ("bremfree: buf must be locked.")); 657151937Sjkim KASSERT((bp->b_flags & B_REMFREE) == 0 && bp->b_qindex != QUEUE_NONE, 658151937Sjkim ("bremfree: buffer not on a queue.")); 659118611Snjl 660151937Sjkim bp->b_flags |= B_REMFREE; 661151937Sjkim /* Fixup numfreebuffers count. */ 662118611Snjl if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 663167802Sjkim atomic_subtract_int(&numfreebuffers, 1); 664167802Sjkim} 665167802Sjkim 666167802Sjkim/* 667118611Snjl * bremfreef: 668118611Snjl * 669118611Snjl * Force an immediate removal from a free list. Used only in nfs when 670118611Snjl * it abuses the b_freelist pointer. 671212761Sjkim */ 672212761Sjkimvoid 673212761Sjkimbremfreef(struct buf *bp) 674118611Snjl{ 675212761Sjkim mtx_lock(&bqlock); 676151937Sjkim bremfreel(bp); 677151937Sjkim mtx_unlock(&bqlock); 678151937Sjkim} 679151937Sjkim 680151937Sjkim/* 681151937Sjkim * bremfreel: 682151937Sjkim * 683151937Sjkim * Removes a buffer from the free list, must be called with the 684118611Snjl * bqlock held. 685151937Sjkim */ 686118611Snjlvoid 687151937Sjkimbremfreel(struct buf *bp) 688151937Sjkim{ 689151937Sjkim int s = splbio(); 690151937Sjkim 691118611Snjl mtx_assert(&bqlock, MA_OWNED); 692209746Sjkim 693209746Sjkim if (bp->b_qindex != QUEUE_NONE) { 694209746Sjkim KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp)); 695209746Sjkim TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 696118611Snjl bp->b_qindex = QUEUE_NONE; 697118611Snjl } else { 698118611Snjl if (BUF_REFCNT(bp) <= 1) 699151937Sjkim panic("bremfree: removing a buffer not on a queue"); 700118611Snjl } 701118611Snjl /* 702118611Snjl * If this was a delayed bremfree() we only need to remove the buffer 703118611Snjl * from the queue and return the stats are already done. 704118611Snjl */ 705151937Sjkim if (bp->b_flags & B_REMFREE) { 706118611Snjl bp->b_flags &= ~B_REMFREE; 707118611Snjl splx(s); 708118611Snjl return; 709118611Snjl } 710118611Snjl /* 711118611Snjl * Fixup numfreebuffers count. If the buffer is invalid or not 712118611Snjl * delayed-write, the buffer was free and we must decrement 713118611Snjl * numfreebuffers. 714118611Snjl */ 715118611Snjl if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 716118611Snjl atomic_subtract_int(&numfreebuffers, 1); 717118611Snjl splx(s); 718118611Snjl} 719118611Snjl 720118611Snjl 721118611Snjl/* 722151937Sjkim * Get a buffer with the specified data. Look in the cache first. We 723118611Snjl * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 724118611Snjl * is set, the buffer is valid and we do not have to do anything ( see 725118611Snjl * getblk() ). This is really just a special case of breadn(). 726118611Snjl */ 727118611Snjlint 728118611Snjlbread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 729118611Snjl struct buf **bpp) 730118611Snjl{ 731118611Snjl 732118611Snjl return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 733118611Snjl} 734118611Snjl 735118611Snjl/* 736118611Snjl * Operates like bread, but also starts asynchronous I/O on 737118611Snjl * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 738118611Snjl * to initiating I/O . If B_CACHE is set, the buffer is valid 739118611Snjl * and we do not have to do anything. 740118611Snjl */ 741118611Snjlint 742118611Snjlbreadn(struct vnode * vp, daddr_t blkno, int size, 743118611Snjl daddr_t * rablkno, int *rabsize, 744118611Snjl int cnt, struct ucred * cred, struct buf **bpp) 745118611Snjl{ 746118611Snjl struct buf *bp, *rabp; 747118611Snjl int i; 748118611Snjl int rv = 0, readwait = 0; 749118611Snjl 750118611Snjl *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 751118611Snjl 752118611Snjl /* if not found in cache, do some I/O */ 753118611Snjl if ((bp->b_flags & B_CACHE) == 0) { 754118611Snjl if (curthread != PCPU_GET(idlethread)) 755118611Snjl curthread->td_proc->p_stats->p_ru.ru_inblock++; 756118611Snjl bp->b_iocmd = BIO_READ; 757118611Snjl bp->b_flags &= ~B_INVAL; 758118611Snjl bp->b_ioflags &= ~BIO_ERROR; 759118611Snjl if (bp->b_rcred == NOCRED && cred != NOCRED) 760118611Snjl bp->b_rcred = crhold(cred); 761118611Snjl vfs_busy_pages(bp, 0); 762202771Sjkim bp->b_iooffset = dbtob(bp->b_blkno); 763118611Snjl bstrategy(bp); 764118611Snjl ++readwait; 765118611Snjl } 766217365Sjkim 767217365Sjkim for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 768217365Sjkim if (inmem(vp, *rablkno)) 769217365Sjkim continue; 770217365Sjkim rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 771217365Sjkim 772118611Snjl if ((rabp->b_flags & B_CACHE) == 0) { 773217365Sjkim if (curthread != PCPU_GET(idlethread)) 774217365Sjkim curthread->td_proc->p_stats->p_ru.ru_inblock++; 775217365Sjkim rabp->b_flags |= B_ASYNC; 776217365Sjkim rabp->b_flags &= ~B_INVAL; 777217365Sjkim rabp->b_ioflags &= ~BIO_ERROR; 778217365Sjkim rabp->b_iocmd = BIO_READ; 779217365Sjkim if (rabp->b_rcred == NOCRED && cred != NOCRED) 780217365Sjkim rabp->b_rcred = crhold(cred); 781217365Sjkim vfs_busy_pages(rabp, 0); 782217365Sjkim BUF_KERNPROC(rabp); 783118611Snjl rabp->b_iooffset = dbtob(rabp->b_blkno); 784151937Sjkim bstrategy(rabp); 785118611Snjl } else { 786207344Sjkim brelse(rabp); 787207344Sjkim } 788207344Sjkim } 789207344Sjkim 790207344Sjkim if (readwait) { 791207344Sjkim rv = bufwait(bp); 792207344Sjkim } 793207344Sjkim return (rv); 794207344Sjkim} 795207344Sjkim 796213806Sjkim/* 797213806Sjkim * Write, release buffer on completion. (Done by iodone 798207344Sjkim * if async). Do not bother writing anything if the buffer 799207344Sjkim * is invalid. 800207344Sjkim * 801207344Sjkim * Note that we set B_CACHE here, indicating that buffer is 802207344Sjkim * fully valid and thus cacheable. This is true even of NFS 803207344Sjkim * now so we set it generally. This could be set either here 804207344Sjkim * or in biodone() since the I/O is synchronous. We put it 805207344Sjkim * here. 806207344Sjkim */ 807207344Sjkimint 808207344Sjkimbufwrite(struct buf *bp) 809213806Sjkim{ 810213806Sjkim int oldflags, s; 811207344Sjkim struct buf *newbp; 812207344Sjkim 813207344Sjkim if (bp->b_flags & B_INVAL) { 814207344Sjkim brelse(bp); 815207344Sjkim return (0); 816118611Snjl } 817118611Snjl 818118611Snjl oldflags = bp->b_flags; 819118611Snjl 820151937Sjkim if (BUF_REFCNT(bp) == 0) 821118611Snjl panic("bufwrite: buffer is not busy???"); 822118611Snjl s = splbio(); 823118611Snjl /* 824118611Snjl * If a background write is already in progress, delay 825118611Snjl * writing this block if it is asynchronous. Otherwise 826118611Snjl * wait for the background write to complete. 827118611Snjl */ 828118611Snjl BO_LOCK(bp->b_bufobj); 829118611Snjl if (bp->b_vflags & BV_BKGRDINPROG) { 830118611Snjl if (bp->b_flags & B_ASYNC) { 831118611Snjl BO_UNLOCK(bp->b_bufobj); 832118611Snjl splx(s); 833118611Snjl bdwrite(bp); 834118611Snjl return (0); 835118611Snjl } 836118611Snjl bp->b_vflags |= BV_BKGRDWAIT; 837118611Snjl msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), PRIBIO, "bwrbg", 0); 838118611Snjl if (bp->b_vflags & BV_BKGRDINPROG) 839118611Snjl panic("bufwrite: still writing"); 840118611Snjl } 841118611Snjl BO_UNLOCK(bp->b_bufobj); 842118611Snjl 843118611Snjl /* Mark the buffer clean */ 844151937Sjkim bundirty(bp); 845151937Sjkim 846151937Sjkim /* 847151937Sjkim * If this buffer is marked for background writing and we 848118611Snjl * do not have to wait for it, make a copy and write the 849118611Snjl * copy so as to leave this buffer ready for further use. 850118611Snjl * 851118611Snjl * This optimization eats a lot of memory. If we have a page 852118611Snjl * or buffer shortfall we can't do it. 853118611Snjl */ 854151937Sjkim if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 855151937Sjkim (bp->b_flags & B_ASYNC) && 856118611Snjl !vm_page_count_severe() && 857118611Snjl !buf_dirty_count_severe()) { 858118611Snjl KASSERT(bp->b_iodone == NULL, 859118611Snjl ("bufwrite: needs chained iodone (%p)", bp->b_iodone)); 860118611Snjl 861118611Snjl /* get a new block */ 862118611Snjl newbp = geteblk(bp->b_bufsize); 863118611Snjl 864118611Snjl /* 865151937Sjkim * set it to be identical to the old block. We have to 866151937Sjkim * set b_lblkno and BKGRDMARKER before calling bgetvp() 867151937Sjkim * to avoid confusing the splay tree and gbincore(). 868118611Snjl */ 869151937Sjkim memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 870118611Snjl newbp->b_lblkno = bp->b_lblkno; 871207344Sjkim newbp->b_xflags |= BX_BKGRDMARKER; 872118611Snjl BO_LOCK(bp->b_bufobj); 873118611Snjl bp->b_vflags |= BV_BKGRDINPROG; 874167802Sjkim bgetvp(bp->b_vp, newbp); 875167802Sjkim BO_UNLOCK(bp->b_bufobj); 876167802Sjkim newbp->b_bufobj = &bp->b_vp->v_bufobj; 877167802Sjkim newbp->b_blkno = bp->b_blkno; 878167802Sjkim newbp->b_offset = bp->b_offset; 879207344Sjkim newbp->b_iodone = vfs_backgroundwritedone; 880118611Snjl newbp->b_flags |= B_ASYNC; 881118611Snjl newbp->b_flags &= ~B_INVAL; 882118611Snjl 883118611Snjl /* move over the dependencies */ 884207344Sjkim if (LIST_FIRST(&bp->b_dep) != NULL) 885118611Snjl buf_movedeps(bp, newbp); 886118611Snjl 887118611Snjl /* 888118611Snjl * Initiate write on the copy, release the original to 889207344Sjkim * the B_LOCKED queue so that it cannot go away until 890118611Snjl * the background write completes. If not locked it could go 891118611Snjl * away and then be reconstituted while it was being written. 892118611Snjl * If the reconstituted buffer were written, we could end up 893118611Snjl * with two background copies being written at the same time. 894207344Sjkim */ 895118611Snjl bqrelse(bp); 896118611Snjl bp = newbp; 897118611Snjl } 898118611Snjl 899207344Sjkim bp->b_flags &= ~B_DONE; 900118611Snjl bp->b_ioflags &= ~BIO_ERROR; 901118611Snjl bp->b_flags |= B_CACHE; 902118611Snjl bp->b_iocmd = BIO_WRITE; 903118611Snjl 904207344Sjkim bufobj_wref(bp->b_bufobj); 905118611Snjl vfs_busy_pages(bp, 1); 906118611Snjl 907118611Snjl /* 908118611Snjl * Normal bwrites pipeline writes 909207344Sjkim */ 910118611Snjl bp->b_runningbufspace = bp->b_bufsize; 911118611Snjl atomic_add_int(&runningbufspace, bp->b_runningbufspace); 912118611Snjl 913207344Sjkim if (curthread != PCPU_GET(idlethread)) 914207344Sjkim curthread->td_proc->p_stats->p_ru.ru_oublock++; 915207344Sjkim splx(s); 916207344Sjkim if (oldflags & B_ASYNC) 917118611Snjl BUF_KERNPROC(bp); 918207344Sjkim bp->b_iooffset = dbtob(bp->b_blkno); 919118611Snjl bstrategy(bp); 920118611Snjl 921118611Snjl if ((oldflags & B_ASYNC) == 0) { 922118611Snjl int rtval = bufwait(bp); 923207344Sjkim brelse(bp); 924118611Snjl return (rtval); 925118611Snjl } else { 926118611Snjl /* 927118611Snjl * don't allow the async write to saturate the I/O 928207344Sjkim * system. We will not deadlock here because 929118611Snjl * we are blocking waiting for I/O that is already in-progress 930118611Snjl * to complete. We do not block here if it is the update 931118611Snjl * or syncer daemon trying to clean up as that can lead 932118611Snjl * to deadlock. 933207344Sjkim */ 934118611Snjl if (curthread->td_proc != bufdaemonproc && 935118611Snjl curthread->td_proc != updateproc) 936118611Snjl waitrunningbufspace(); 937118611Snjl } 938207344Sjkim 939118611Snjl return (0); 940118611Snjl} 941118611Snjl 942118611Snjl/* 943118611Snjl * Complete a background write started from bwrite. 944207344Sjkim */ 945118611Snjlstatic void 946151937Sjkimvfs_backgroundwritedone(struct buf *bp) 947151937Sjkim{ 948151937Sjkim struct buf *origbp; 949151937Sjkim 950118611Snjl /* 951118611Snjl * Find the original buffer that we are writing. 952207344Sjkim */ 953207344Sjkim BO_LOCK(bp->b_bufobj); 954207344Sjkim if ((origbp = gbincore(bp->b_bufobj, bp->b_lblkno)) == NULL) 955207344Sjkim panic("backgroundwritedone: lost buffer"); 956207344Sjkim 957207344Sjkim /* 958207344Sjkim * Clear the BV_BKGRDINPROG flag in the original buffer 959207344Sjkim * and awaken it if it is waiting for the write to complete. 960207344Sjkim * If BV_BKGRDINPROG is not set in the original buffer it must 961207344Sjkim * have been released and re-instantiated - which is not legal. 962207344Sjkim */ 963207344Sjkim KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 964207344Sjkim ("backgroundwritedone: lost buffer2")); 965207344Sjkim origbp->b_vflags &= ~BV_BKGRDINPROG; 966118611Snjl if (origbp->b_vflags & BV_BKGRDWAIT) { 967118611Snjl origbp->b_vflags &= ~BV_BKGRDWAIT; 968118611Snjl wakeup(&origbp->b_xflags); 969118611Snjl } 970118611Snjl BO_UNLOCK(bp->b_bufobj); 971118611Snjl /* 972118611Snjl * Process dependencies then return any unfinished ones. 973118611Snjl */ 974118611Snjl if (LIST_FIRST(&bp->b_dep) != NULL) 975118611Snjl buf_complete(bp); 976151937Sjkim if (LIST_FIRST(&bp->b_dep) != NULL) 977151937Sjkim buf_movedeps(bp, origbp); 978151937Sjkim 979151937Sjkim /* 980207344Sjkim * This buffer is marked B_NOCACHE, so when it is released 981207344Sjkim * by biodone, it will be tossed. We mark it with BIO_READ 982207344Sjkim * to avoid biodone doing a second bufobj_wdrop. 983207344Sjkim */ 984151937Sjkim bp->b_flags |= B_NOCACHE; 985151937Sjkim bp->b_iocmd = BIO_READ; 986151937Sjkim bp->b_flags &= ~(B_CACHE | B_DONE); 987151937Sjkim bp->b_iodone = 0; 988151937Sjkim bufdone(bp); 989151937Sjkim} 990151937Sjkim 991151937Sjkim/* 992151937Sjkim * Delayed write. (Buffer is marked dirty). Do not bother writing 993151937Sjkim * anything if the buffer is marked invalid. 994151937Sjkim * 995151937Sjkim * Note that since the buffer must be completely valid, we can safely 996151937Sjkim * set B_CACHE. In fact, we have to set B_CACHE here rather then in 997151937Sjkim * biodone() in order to prevent getblk from writing the buffer 998151937Sjkim * out synchronously. 999207344Sjkim */ 1000207344Sjkimvoid 1001207344Sjkimbdwrite(struct buf *bp) 1002207344Sjkim{ 1003151937Sjkim struct thread *td = curthread; 1004118611Snjl struct vnode *vp; 1005118611Snjl struct buf *nbp; 1006118611Snjl struct bufobj *bo; 1007118611Snjl 1008118611Snjl GIANT_REQUIRED; 1009118611Snjl 1010118611Snjl KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1011118611Snjl KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy")); 1012118611Snjl 1013118611Snjl if (bp->b_flags & B_INVAL) { 1014151937Sjkim brelse(bp); 1015151937Sjkim return; 1016151937Sjkim } 1017151937Sjkim 1018207344Sjkim /* 1019207344Sjkim * If we have too many dirty buffers, don't create any more. 1020207344Sjkim * If we are wildly over our limit, then force a complete 1021207344Sjkim * cleanup. Otherwise, just keep the situation from getting 1022151937Sjkim * out of control. Note that we have to avoid a recursive 1023118611Snjl * disaster and not try to clean up after our own cleanup! 1024118611Snjl */ 1025118611Snjl vp = bp->b_vp; 1026118611Snjl bo = bp->b_bufobj; 1027118611Snjl BO_LOCK(bo); 1028151937Sjkim if (td->td_pflags & TDP_COWINPROGRESS) { 1029151937Sjkim recursiveflushes++; 1030151937Sjkim } else if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1031151937Sjkim BO_UNLOCK(bo); 1032151937Sjkim (void) VOP_FSYNC(vp, MNT_NOWAIT, td); 1033118611Snjl BO_LOCK(bo); 1034118611Snjl altbufferflushes++; 1035118611Snjl } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1036118611Snjl /* 1037209746Sjkim * Try to find a buffer to flush. 1038209746Sjkim */ 1039209746Sjkim TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1040209746Sjkim if ((nbp->b_vflags & BV_BKGRDINPROG) || 1041209746Sjkim buf_countdeps(nbp, 0) || 1042209746Sjkim BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1043209746Sjkim continue; 1044209746Sjkim if (bp == nbp) 1045209746Sjkim panic("bdwrite: found ourselves"); 1046209746Sjkim BO_UNLOCK(bo); 1047209746Sjkim if (nbp->b_flags & B_CLUSTEROK) { 1048118611Snjl vfs_bio_awrite(nbp); 1049118611Snjl } else { 1050 bremfree(nbp); 1051 bawrite(nbp); 1052 } 1053 BO_LOCK(bo); 1054 dirtybufferflushes++; 1055 break; 1056 } 1057 } 1058 BO_UNLOCK(bo); 1059 1060 bdirty(bp); 1061 /* 1062 * Set B_CACHE, indicating that the buffer is fully valid. This is 1063 * true even of NFS now. 1064 */ 1065 bp->b_flags |= B_CACHE; 1066 1067 /* 1068 * This bmap keeps the system from needing to do the bmap later, 1069 * perhaps when the system is attempting to do a sync. Since it 1070 * is likely that the indirect block -- or whatever other datastructure 1071 * that the filesystem needs is still in memory now, it is a good 1072 * thing to do this. Note also, that if the pageout daemon is 1073 * requesting a sync -- there might not be enough memory to do 1074 * the bmap then... So, this is important to do. 1075 */ 1076 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1077 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1078 } 1079 1080 /* 1081 * Set the *dirty* buffer range based upon the VM system dirty pages. 1082 */ 1083 vfs_setdirty(bp); 1084 1085 /* 1086 * We need to do this here to satisfy the vnode_pager and the 1087 * pageout daemon, so that it thinks that the pages have been 1088 * "cleaned". Note that since the pages are in a delayed write 1089 * buffer -- the VFS layer "will" see that the pages get written 1090 * out on the next sync, or perhaps the cluster will be completed. 1091 */ 1092 vfs_clean_pages(bp); 1093 bqrelse(bp); 1094 1095 /* 1096 * Wakeup the buffer flushing daemon if we have a lot of dirty 1097 * buffers (midpoint between our recovery point and our stall 1098 * point). 1099 */ 1100 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1101 1102 /* 1103 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1104 * due to the softdep code. 1105 */ 1106} 1107 1108/* 1109 * bdirty: 1110 * 1111 * Turn buffer into delayed write request. We must clear BIO_READ and 1112 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1113 * itself to properly update it in the dirty/clean lists. We mark it 1114 * B_DONE to ensure that any asynchronization of the buffer properly 1115 * clears B_DONE ( else a panic will occur later ). 1116 * 1117 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1118 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1119 * should only be called if the buffer is known-good. 1120 * 1121 * Since the buffer is not on a queue, we do not update the numfreebuffers 1122 * count. 1123 * 1124 * Must be called at splbio(). 1125 * The buffer must be on QUEUE_NONE. 1126 */ 1127void 1128bdirty(struct buf *bp) 1129{ 1130 1131 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1132 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1133 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1134 bp->b_flags &= ~(B_RELBUF); 1135 bp->b_iocmd = BIO_WRITE; 1136 1137 if ((bp->b_flags & B_DELWRI) == 0) { 1138 bp->b_flags |= B_DONE | B_DELWRI; 1139 reassignbuf(bp); 1140 atomic_add_int(&numdirtybuffers, 1); 1141 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1142 } 1143} 1144 1145/* 1146 * bundirty: 1147 * 1148 * Clear B_DELWRI for buffer. 1149 * 1150 * Since the buffer is not on a queue, we do not update the numfreebuffers 1151 * count. 1152 * 1153 * Must be called at splbio(). 1154 * The buffer must be on QUEUE_NONE. 1155 */ 1156 1157void 1158bundirty(struct buf *bp) 1159{ 1160 1161 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1162 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1163 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1164 1165 if (bp->b_flags & B_DELWRI) { 1166 bp->b_flags &= ~B_DELWRI; 1167 reassignbuf(bp); 1168 atomic_subtract_int(&numdirtybuffers, 1); 1169 numdirtywakeup(lodirtybuffers); 1170 } 1171 /* 1172 * Since it is now being written, we can clear its deferred write flag. 1173 */ 1174 bp->b_flags &= ~B_DEFERRED; 1175} 1176 1177/* 1178 * bawrite: 1179 * 1180 * Asynchronous write. Start output on a buffer, but do not wait for 1181 * it to complete. The buffer is released when the output completes. 1182 * 1183 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1184 * B_INVAL buffers. Not us. 1185 */ 1186void 1187bawrite(struct buf *bp) 1188{ 1189 1190 bp->b_flags |= B_ASYNC; 1191 (void) bwrite(bp); 1192} 1193 1194/* 1195 * bwillwrite: 1196 * 1197 * Called prior to the locking of any vnodes when we are expecting to 1198 * write. We do not want to starve the buffer cache with too many 1199 * dirty buffers so we block here. By blocking prior to the locking 1200 * of any vnodes we attempt to avoid the situation where a locked vnode 1201 * prevents the various system daemons from flushing related buffers. 1202 */ 1203 1204void 1205bwillwrite(void) 1206{ 1207 1208 if (numdirtybuffers >= hidirtybuffers) { 1209 int s; 1210 1211 mtx_lock(&Giant); 1212 s = splbio(); 1213 mtx_lock(&nblock); 1214 while (numdirtybuffers >= hidirtybuffers) { 1215 bd_wakeup(1); 1216 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1217 msleep(&needsbuffer, &nblock, 1218 (PRIBIO + 4), "flswai", 0); 1219 } 1220 splx(s); 1221 mtx_unlock(&nblock); 1222 mtx_unlock(&Giant); 1223 } 1224} 1225 1226/* 1227 * Return true if we have too many dirty buffers. 1228 */ 1229int 1230buf_dirty_count_severe(void) 1231{ 1232 1233 return(numdirtybuffers >= hidirtybuffers); 1234} 1235 1236/* 1237 * brelse: 1238 * 1239 * Release a busy buffer and, if requested, free its resources. The 1240 * buffer will be stashed in the appropriate bufqueue[] allowing it 1241 * to be accessed later as a cache entity or reused for other purposes. 1242 */ 1243void 1244brelse(struct buf *bp) 1245{ 1246 int s; 1247 1248 GIANT_REQUIRED; 1249 1250 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1251 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1252 1253 s = splbio(); 1254 1255 if (bp->b_iocmd == BIO_WRITE && 1256 (bp->b_ioflags & BIO_ERROR) && 1257 !(bp->b_flags & B_INVAL)) { 1258 /* 1259 * Failed write, redirty. Must clear BIO_ERROR to prevent 1260 * pages from being scrapped. If B_INVAL is set then 1261 * this case is not run and the next case is run to 1262 * destroy the buffer. B_INVAL can occur if the buffer 1263 * is outside the range supported by the underlying device. 1264 */ 1265 bp->b_ioflags &= ~BIO_ERROR; 1266 bdirty(bp); 1267 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1268 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1269 /* 1270 * Either a failed I/O or we were asked to free or not 1271 * cache the buffer. 1272 */ 1273 bp->b_flags |= B_INVAL; 1274 if (LIST_FIRST(&bp->b_dep) != NULL) 1275 buf_deallocate(bp); 1276 if (bp->b_flags & B_DELWRI) { 1277 atomic_subtract_int(&numdirtybuffers, 1); 1278 numdirtywakeup(lodirtybuffers); 1279 } 1280 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1281 if ((bp->b_flags & B_VMIO) == 0) { 1282 if (bp->b_bufsize) 1283 allocbuf(bp, 0); 1284 if (bp->b_vp) 1285 brelvp(bp); 1286 } 1287 } 1288 1289 /* 1290 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1291 * is called with B_DELWRI set, the underlying pages may wind up 1292 * getting freed causing a previous write (bdwrite()) to get 'lost' 1293 * because pages associated with a B_DELWRI bp are marked clean. 1294 * 1295 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1296 * if B_DELWRI is set. 1297 * 1298 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1299 * on pages to return pages to the VM page queues. 1300 */ 1301 if (bp->b_flags & B_DELWRI) 1302 bp->b_flags &= ~B_RELBUF; 1303 else if (vm_page_count_severe()) { 1304 /* 1305 * XXX This lock may not be necessary since BKGRDINPROG 1306 * cannot be set while we hold the buf lock, it can only be 1307 * cleared if it is already pending. 1308 */ 1309 if (bp->b_vp) { 1310 BO_LOCK(bp->b_bufobj); 1311 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1312 bp->b_flags |= B_RELBUF; 1313 BO_UNLOCK(bp->b_bufobj); 1314 } else 1315 bp->b_flags |= B_RELBUF; 1316 } 1317 1318 /* 1319 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1320 * constituted, not even NFS buffers now. Two flags effect this. If 1321 * B_INVAL, the struct buf is invalidated but the VM object is kept 1322 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1323 * 1324 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1325 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1326 * buffer is also B_INVAL because it hits the re-dirtying code above. 1327 * 1328 * Normally we can do this whether a buffer is B_DELWRI or not. If 1329 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1330 * the commit state and we cannot afford to lose the buffer. If the 1331 * buffer has a background write in progress, we need to keep it 1332 * around to prevent it from being reconstituted and starting a second 1333 * background write. 1334 */ 1335 if ((bp->b_flags & B_VMIO) 1336 && !(bp->b_vp->v_mount != NULL && 1337 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1338 !vn_isdisk(bp->b_vp, NULL) && 1339 (bp->b_flags & B_DELWRI)) 1340 ) { 1341 1342 int i, j, resid; 1343 vm_page_t m; 1344 off_t foff; 1345 vm_pindex_t poff; 1346 vm_object_t obj; 1347 1348 obj = bp->b_bufobj->bo_object; 1349 1350 /* 1351 * Get the base offset and length of the buffer. Note that 1352 * in the VMIO case if the buffer block size is not 1353 * page-aligned then b_data pointer may not be page-aligned. 1354 * But our b_pages[] array *IS* page aligned. 1355 * 1356 * block sizes less then DEV_BSIZE (usually 512) are not 1357 * supported due to the page granularity bits (m->valid, 1358 * m->dirty, etc...). 1359 * 1360 * See man buf(9) for more information 1361 */ 1362 resid = bp->b_bufsize; 1363 foff = bp->b_offset; 1364 VM_OBJECT_LOCK(obj); 1365 for (i = 0; i < bp->b_npages; i++) { 1366 int had_bogus = 0; 1367 1368 m = bp->b_pages[i]; 1369 1370 /* 1371 * If we hit a bogus page, fixup *all* the bogus pages 1372 * now. 1373 */ 1374 if (m == bogus_page) { 1375 poff = OFF_TO_IDX(bp->b_offset); 1376 had_bogus = 1; 1377 1378 for (j = i; j < bp->b_npages; j++) { 1379 vm_page_t mtmp; 1380 mtmp = bp->b_pages[j]; 1381 if (mtmp == bogus_page) { 1382 mtmp = vm_page_lookup(obj, poff + j); 1383 if (!mtmp) { 1384 panic("brelse: page missing\n"); 1385 } 1386 bp->b_pages[j] = mtmp; 1387 } 1388 } 1389 1390 if ((bp->b_flags & B_INVAL) == 0) { 1391 pmap_qenter( 1392 trunc_page((vm_offset_t)bp->b_data), 1393 bp->b_pages, bp->b_npages); 1394 } 1395 m = bp->b_pages[i]; 1396 } 1397 if ((bp->b_flags & B_NOCACHE) || 1398 (bp->b_ioflags & BIO_ERROR)) { 1399 int poffset = foff & PAGE_MASK; 1400 int presid = resid > (PAGE_SIZE - poffset) ? 1401 (PAGE_SIZE - poffset) : resid; 1402 1403 KASSERT(presid >= 0, ("brelse: extra page")); 1404 vm_page_lock_queues(); 1405 vm_page_set_invalid(m, poffset, presid); 1406 vm_page_unlock_queues(); 1407 if (had_bogus) 1408 printf("avoided corruption bug in bogus_page/brelse code\n"); 1409 } 1410 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1411 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1412 } 1413 VM_OBJECT_UNLOCK(obj); 1414 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1415 vfs_vmio_release(bp); 1416 1417 } else if (bp->b_flags & B_VMIO) { 1418 1419 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1420 vfs_vmio_release(bp); 1421 } 1422 1423 } 1424 1425 if (BUF_REFCNT(bp) > 1) { 1426 /* do not release to free list */ 1427 BUF_UNLOCK(bp); 1428 splx(s); 1429 return; 1430 } 1431 1432 /* enqueue */ 1433 mtx_lock(&bqlock); 1434 /* Handle delayed bremfree() processing. */ 1435 if (bp->b_flags & B_REMFREE) 1436 bremfreel(bp); 1437 if (bp->b_qindex != QUEUE_NONE) 1438 panic("brelse: free buffer onto another queue???"); 1439 1440 /* buffers with no memory */ 1441 if (bp->b_bufsize == 0) { 1442 bp->b_flags |= B_INVAL; 1443 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1444 if (bp->b_vflags & BV_BKGRDINPROG) 1445 panic("losing buffer 1"); 1446 if (bp->b_kvasize) { 1447 bp->b_qindex = QUEUE_EMPTYKVA; 1448 } else { 1449 bp->b_qindex = QUEUE_EMPTY; 1450 } 1451 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1452 /* buffers with junk contents */ 1453 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1454 (bp->b_ioflags & BIO_ERROR)) { 1455 bp->b_flags |= B_INVAL; 1456 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1457 if (bp->b_vflags & BV_BKGRDINPROG) 1458 panic("losing buffer 2"); 1459 bp->b_qindex = QUEUE_CLEAN; 1460 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1461 /* remaining buffers */ 1462 } else { 1463 if (bp->b_flags & B_DELWRI) 1464 bp->b_qindex = QUEUE_DIRTY; 1465 else 1466 bp->b_qindex = QUEUE_CLEAN; 1467 if (bp->b_flags & B_AGE) 1468 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1469 else 1470 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1471 } 1472 mtx_unlock(&bqlock); 1473 1474 /* 1475 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1476 * placed the buffer on the correct queue. We must also disassociate 1477 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1478 * find it. 1479 */ 1480 if (bp->b_flags & B_INVAL) { 1481 if (bp->b_flags & B_DELWRI) 1482 bundirty(bp); 1483 if (bp->b_vp) 1484 brelvp(bp); 1485 } 1486 1487 /* 1488 * Fixup numfreebuffers count. The bp is on an appropriate queue 1489 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1490 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1491 * if B_INVAL is set ). 1492 */ 1493 1494 if (!(bp->b_flags & B_DELWRI)) 1495 bufcountwakeup(); 1496 1497 /* 1498 * Something we can maybe free or reuse 1499 */ 1500 if (bp->b_bufsize || bp->b_kvasize) 1501 bufspacewakeup(); 1502 1503 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1504 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1505 panic("brelse: not dirty"); 1506 /* unlock */ 1507 BUF_UNLOCK(bp); 1508 splx(s); 1509} 1510 1511/* 1512 * Release a buffer back to the appropriate queue but do not try to free 1513 * it. The buffer is expected to be used again soon. 1514 * 1515 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1516 * biodone() to requeue an async I/O on completion. It is also used when 1517 * known good buffers need to be requeued but we think we may need the data 1518 * again soon. 1519 * 1520 * XXX we should be able to leave the B_RELBUF hint set on completion. 1521 */ 1522void 1523bqrelse(struct buf *bp) 1524{ 1525 int s; 1526 1527 s = splbio(); 1528 1529 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1530 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1531 1532 if (BUF_REFCNT(bp) > 1) { 1533 /* do not release to free list */ 1534 BUF_UNLOCK(bp); 1535 splx(s); 1536 return; 1537 } 1538 mtx_lock(&bqlock); 1539 /* Handle delayed bremfree() processing. */ 1540 if (bp->b_flags & B_REMFREE) 1541 bremfreel(bp); 1542 if (bp->b_qindex != QUEUE_NONE) 1543 panic("bqrelse: free buffer onto another queue???"); 1544 /* buffers with stale but valid contents */ 1545 if (bp->b_flags & B_DELWRI) { 1546 bp->b_qindex = QUEUE_DIRTY; 1547 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1548 } else { 1549 /* 1550 * XXX This lock may not be necessary since BKGRDINPROG 1551 * cannot be set while we hold the buf lock, it can only be 1552 * cleared if it is already pending. 1553 */ 1554 BO_LOCK(bp->b_bufobj); 1555 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1556 BO_UNLOCK(bp->b_bufobj); 1557 bp->b_qindex = QUEUE_CLEAN; 1558 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1559 b_freelist); 1560 } else { 1561 /* 1562 * We are too low on memory, we have to try to free 1563 * the buffer (most importantly: the wired pages 1564 * making up its backing store) *now*. 1565 */ 1566 BO_UNLOCK(bp->b_bufobj); 1567 mtx_unlock(&bqlock); 1568 splx(s); 1569 brelse(bp); 1570 return; 1571 } 1572 } 1573 mtx_unlock(&bqlock); 1574 1575 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1576 bufcountwakeup(); 1577 1578 /* 1579 * Something we can maybe free or reuse. 1580 */ 1581 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1582 bufspacewakeup(); 1583 1584 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1585 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1586 panic("bqrelse: not dirty"); 1587 /* unlock */ 1588 BUF_UNLOCK(bp); 1589 splx(s); 1590} 1591 1592/* Give pages used by the bp back to the VM system (where possible) */ 1593static void 1594vfs_vmio_release(struct buf *bp) 1595{ 1596 int i; 1597 vm_page_t m; 1598 1599 GIANT_REQUIRED; 1600 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1601 vm_page_lock_queues(); 1602 for (i = 0; i < bp->b_npages; i++) { 1603 m = bp->b_pages[i]; 1604 bp->b_pages[i] = NULL; 1605 /* 1606 * In order to keep page LRU ordering consistent, put 1607 * everything on the inactive queue. 1608 */ 1609 vm_page_unwire(m, 0); 1610 /* 1611 * We don't mess with busy pages, it is 1612 * the responsibility of the process that 1613 * busied the pages to deal with them. 1614 */ 1615 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1616 continue; 1617 1618 if (m->wire_count == 0) { 1619 /* 1620 * Might as well free the page if we can and it has 1621 * no valid data. We also free the page if the 1622 * buffer was used for direct I/O 1623 */ 1624 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1625 m->hold_count == 0) { 1626 pmap_remove_all(m); 1627 vm_page_free(m); 1628 } else if (bp->b_flags & B_DIRECT) { 1629 vm_page_try_to_free(m); 1630 } else if (vm_page_count_severe()) { 1631 vm_page_try_to_cache(m); 1632 } 1633 } 1634 } 1635 vm_page_unlock_queues(); 1636 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1637 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1638 1639 if (bp->b_bufsize) { 1640 bufspacewakeup(); 1641 bp->b_bufsize = 0; 1642 } 1643 bp->b_npages = 0; 1644 bp->b_flags &= ~B_VMIO; 1645 if (bp->b_vp) 1646 brelvp(bp); 1647} 1648 1649/* 1650 * Check to see if a block at a particular lbn is available for a clustered 1651 * write. 1652 */ 1653static int 1654vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1655{ 1656 struct buf *bpa; 1657 int match; 1658 1659 match = 0; 1660 1661 /* If the buf isn't in core skip it */ 1662 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1663 return (0); 1664 1665 /* If the buf is busy we don't want to wait for it */ 1666 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1667 return (0); 1668 1669 /* Only cluster with valid clusterable delayed write buffers */ 1670 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1671 (B_DELWRI | B_CLUSTEROK)) 1672 goto done; 1673 1674 if (bpa->b_bufsize != size) 1675 goto done; 1676 1677 /* 1678 * Check to see if it is in the expected place on disk and that the 1679 * block has been mapped. 1680 */ 1681 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1682 match = 1; 1683done: 1684 BUF_UNLOCK(bpa); 1685 return (match); 1686} 1687 1688/* 1689 * vfs_bio_awrite: 1690 * 1691 * Implement clustered async writes for clearing out B_DELWRI buffers. 1692 * This is much better then the old way of writing only one buffer at 1693 * a time. Note that we may not be presented with the buffers in the 1694 * correct order, so we search for the cluster in both directions. 1695 */ 1696int 1697vfs_bio_awrite(struct buf *bp) 1698{ 1699 int i; 1700 int j; 1701 daddr_t lblkno = bp->b_lblkno; 1702 struct vnode *vp = bp->b_vp; 1703 int s; 1704 int ncl; 1705 int nwritten; 1706 int size; 1707 int maxcl; 1708 1709 s = splbio(); 1710 /* 1711 * right now we support clustered writing only to regular files. If 1712 * we find a clusterable block we could be in the middle of a cluster 1713 * rather then at the beginning. 1714 */ 1715 if ((vp->v_type == VREG) && 1716 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1717 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1718 1719 size = vp->v_mount->mnt_stat.f_iosize; 1720 maxcl = MAXPHYS / size; 1721 1722 VI_LOCK(vp); 1723 for (i = 1; i < maxcl; i++) 1724 if (vfs_bio_clcheck(vp, size, lblkno + i, 1725 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1726 break; 1727 1728 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1729 if (vfs_bio_clcheck(vp, size, lblkno - j, 1730 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1731 break; 1732 1733 VI_UNLOCK(vp); 1734 --j; 1735 ncl = i + j; 1736 /* 1737 * this is a possible cluster write 1738 */ 1739 if (ncl != 1) { 1740 BUF_UNLOCK(bp); 1741 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1742 splx(s); 1743 return nwritten; 1744 } 1745 } 1746 1747 bremfree(bp); 1748 bp->b_flags |= B_ASYNC; 1749 1750 splx(s); 1751 /* 1752 * default (old) behavior, writing out only one block 1753 * 1754 * XXX returns b_bufsize instead of b_bcount for nwritten? 1755 */ 1756 nwritten = bp->b_bufsize; 1757 (void) bwrite(bp); 1758 1759 return nwritten; 1760} 1761 1762/* 1763 * getnewbuf: 1764 * 1765 * Find and initialize a new buffer header, freeing up existing buffers 1766 * in the bufqueues as necessary. The new buffer is returned locked. 1767 * 1768 * Important: B_INVAL is not set. If the caller wishes to throw the 1769 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1770 * 1771 * We block if: 1772 * We have insufficient buffer headers 1773 * We have insufficient buffer space 1774 * buffer_map is too fragmented ( space reservation fails ) 1775 * If we have to flush dirty buffers ( but we try to avoid this ) 1776 * 1777 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1778 * Instead we ask the buf daemon to do it for us. We attempt to 1779 * avoid piecemeal wakeups of the pageout daemon. 1780 */ 1781 1782static struct buf * 1783getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1784{ 1785 struct buf *bp; 1786 struct buf *nbp; 1787 int defrag = 0; 1788 int nqindex; 1789 static int flushingbufs; 1790 1791 GIANT_REQUIRED; 1792 1793 /* 1794 * We can't afford to block since we might be holding a vnode lock, 1795 * which may prevent system daemons from running. We deal with 1796 * low-memory situations by proactively returning memory and running 1797 * async I/O rather then sync I/O. 1798 */ 1799 1800 atomic_add_int(&getnewbufcalls, 1); 1801 atomic_subtract_int(&getnewbufrestarts, 1); 1802restart: 1803 atomic_add_int(&getnewbufrestarts, 1); 1804 1805 /* 1806 * Setup for scan. If we do not have enough free buffers, 1807 * we setup a degenerate case that immediately fails. Note 1808 * that if we are specially marked process, we are allowed to 1809 * dip into our reserves. 1810 * 1811 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1812 * 1813 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1814 * However, there are a number of cases (defragging, reusing, ...) 1815 * where we cannot backup. 1816 */ 1817 mtx_lock(&bqlock); 1818 nqindex = QUEUE_EMPTYKVA; 1819 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1820 1821 if (nbp == NULL) { 1822 /* 1823 * If no EMPTYKVA buffers and we are either 1824 * defragging or reusing, locate a CLEAN buffer 1825 * to free or reuse. If bufspace useage is low 1826 * skip this step so we can allocate a new buffer. 1827 */ 1828 if (defrag || bufspace >= lobufspace) { 1829 nqindex = QUEUE_CLEAN; 1830 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1831 } 1832 1833 /* 1834 * If we could not find or were not allowed to reuse a 1835 * CLEAN buffer, check to see if it is ok to use an EMPTY 1836 * buffer. We can only use an EMPTY buffer if allocating 1837 * its KVA would not otherwise run us out of buffer space. 1838 */ 1839 if (nbp == NULL && defrag == 0 && 1840 bufspace + maxsize < hibufspace) { 1841 nqindex = QUEUE_EMPTY; 1842 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1843 } 1844 } 1845 1846 /* 1847 * Run scan, possibly freeing data and/or kva mappings on the fly 1848 * depending. 1849 */ 1850 1851 while ((bp = nbp) != NULL) { 1852 int qindex = nqindex; 1853 1854 /* 1855 * Calculate next bp ( we can only use it if we do not block 1856 * or do other fancy things ). 1857 */ 1858 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1859 switch(qindex) { 1860 case QUEUE_EMPTY: 1861 nqindex = QUEUE_EMPTYKVA; 1862 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1863 break; 1864 /* FALLTHROUGH */ 1865 case QUEUE_EMPTYKVA: 1866 nqindex = QUEUE_CLEAN; 1867 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1868 break; 1869 /* FALLTHROUGH */ 1870 case QUEUE_CLEAN: 1871 /* 1872 * nbp is NULL. 1873 */ 1874 break; 1875 } 1876 } 1877 if (bp->b_vp) { 1878 BO_LOCK(bp->b_bufobj); 1879 if (bp->b_vflags & BV_BKGRDINPROG) { 1880 BO_UNLOCK(bp->b_bufobj); 1881 continue; 1882 } 1883 BO_UNLOCK(bp->b_bufobj); 1884 } 1885 1886 /* 1887 * If we are defragging then we need a buffer with 1888 * b_kvasize != 0. XXX this situation should no longer 1889 * occur, if defrag is non-zero the buffer's b_kvasize 1890 * should also be non-zero at this point. XXX 1891 */ 1892 if (defrag && bp->b_kvasize == 0) { 1893 printf("Warning: defrag empty buffer %p\n", bp); 1894 continue; 1895 } 1896 1897 /* 1898 * Start freeing the bp. This is somewhat involved. nbp 1899 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1900 */ 1901 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1902 continue; 1903 /* 1904 * Sanity Checks 1905 */ 1906 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1907 1908 /* 1909 * Note: we no longer distinguish between VMIO and non-VMIO 1910 * buffers. 1911 */ 1912 1913 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1914 1915 bremfreel(bp); 1916 mtx_unlock(&bqlock); 1917 1918 if (qindex == QUEUE_CLEAN) { 1919 if (bp->b_flags & B_VMIO) { 1920 bp->b_flags &= ~B_ASYNC; 1921 vfs_vmio_release(bp); 1922 } 1923 if (bp->b_vp) 1924 brelvp(bp); 1925 } 1926 1927 /* 1928 * NOTE: nbp is now entirely invalid. We can only restart 1929 * the scan from this point on. 1930 * 1931 * Get the rest of the buffer freed up. b_kva* is still 1932 * valid after this operation. 1933 */ 1934 1935 if (bp->b_rcred != NOCRED) { 1936 crfree(bp->b_rcred); 1937 bp->b_rcred = NOCRED; 1938 } 1939 if (bp->b_wcred != NOCRED) { 1940 crfree(bp->b_wcred); 1941 bp->b_wcred = NOCRED; 1942 } 1943 if (LIST_FIRST(&bp->b_dep) != NULL) 1944 buf_deallocate(bp); 1945 if (bp->b_vflags & BV_BKGRDINPROG) 1946 panic("losing buffer 3"); 1947 1948 if (bp->b_bufsize) 1949 allocbuf(bp, 0); 1950 1951 bp->b_flags = 0; 1952 bp->b_ioflags = 0; 1953 bp->b_xflags = 0; 1954 bp->b_vflags = 0; 1955 bp->b_vp = NULL; 1956 bp->b_blkno = bp->b_lblkno = 0; 1957 bp->b_offset = NOOFFSET; 1958 bp->b_iodone = 0; 1959 bp->b_error = 0; 1960 bp->b_resid = 0; 1961 bp->b_bcount = 0; 1962 bp->b_npages = 0; 1963 bp->b_dirtyoff = bp->b_dirtyend = 0; 1964 bp->b_bufobj = NULL; 1965 1966 LIST_INIT(&bp->b_dep); 1967 1968 /* 1969 * If we are defragging then free the buffer. 1970 */ 1971 if (defrag) { 1972 bp->b_flags |= B_INVAL; 1973 bfreekva(bp); 1974 brelse(bp); 1975 defrag = 0; 1976 goto restart; 1977 } 1978 1979 /* 1980 * If we are overcomitted then recover the buffer and its 1981 * KVM space. This occurs in rare situations when multiple 1982 * processes are blocked in getnewbuf() or allocbuf(). 1983 */ 1984 if (bufspace >= hibufspace) 1985 flushingbufs = 1; 1986 if (flushingbufs && bp->b_kvasize != 0) { 1987 bp->b_flags |= B_INVAL; 1988 bfreekva(bp); 1989 brelse(bp); 1990 goto restart; 1991 } 1992 if (bufspace < lobufspace) 1993 flushingbufs = 0; 1994 break; 1995 } 1996 1997 /* 1998 * If we exhausted our list, sleep as appropriate. We may have to 1999 * wakeup various daemons and write out some dirty buffers. 2000 * 2001 * Generally we are sleeping due to insufficient buffer space. 2002 */ 2003 2004 if (bp == NULL) { 2005 int flags; 2006 char *waitmsg; 2007 2008 mtx_unlock(&bqlock); 2009 if (defrag) { 2010 flags = VFS_BIO_NEED_BUFSPACE; 2011 waitmsg = "nbufkv"; 2012 } else if (bufspace >= hibufspace) { 2013 waitmsg = "nbufbs"; 2014 flags = VFS_BIO_NEED_BUFSPACE; 2015 } else { 2016 waitmsg = "newbuf"; 2017 flags = VFS_BIO_NEED_ANY; 2018 } 2019 2020 bd_speedup(); /* heeeelp */ 2021 2022 mtx_lock(&nblock); 2023 needsbuffer |= flags; 2024 while (needsbuffer & flags) { 2025 if (msleep(&needsbuffer, &nblock, 2026 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2027 mtx_unlock(&nblock); 2028 return (NULL); 2029 } 2030 } 2031 mtx_unlock(&nblock); 2032 } else { 2033 /* 2034 * We finally have a valid bp. We aren't quite out of the 2035 * woods, we still have to reserve kva space. In order 2036 * to keep fragmentation sane we only allocate kva in 2037 * BKVASIZE chunks. 2038 */ 2039 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2040 2041 if (maxsize != bp->b_kvasize) { 2042 vm_offset_t addr = 0; 2043 2044 bfreekva(bp); 2045 2046 if (vm_map_findspace(buffer_map, 2047 vm_map_min(buffer_map), maxsize, &addr)) { 2048 /* 2049 * Uh oh. Buffer map is to fragmented. We 2050 * must defragment the map. 2051 */ 2052 atomic_add_int(&bufdefragcnt, 1); 2053 defrag = 1; 2054 bp->b_flags |= B_INVAL; 2055 brelse(bp); 2056 goto restart; 2057 } 2058 if (addr) { 2059 vm_map_insert(buffer_map, NULL, 0, 2060 addr, addr + maxsize, 2061 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2062 2063 bp->b_kvabase = (caddr_t) addr; 2064 bp->b_kvasize = maxsize; 2065 atomic_add_int(&bufspace, bp->b_kvasize); 2066 atomic_add_int(&bufreusecnt, 1); 2067 } 2068 } 2069 bp->b_saveaddr = bp->b_kvabase; 2070 bp->b_data = bp->b_saveaddr; 2071 } 2072 return(bp); 2073} 2074 2075/* 2076 * buf_daemon: 2077 * 2078 * buffer flushing daemon. Buffers are normally flushed by the 2079 * update daemon but if it cannot keep up this process starts to 2080 * take the load in an attempt to prevent getnewbuf() from blocking. 2081 */ 2082 2083static struct kproc_desc buf_kp = { 2084 "bufdaemon", 2085 buf_daemon, 2086 &bufdaemonproc 2087}; 2088SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2089 2090static void 2091buf_daemon() 2092{ 2093 int s; 2094 2095 mtx_lock(&Giant); 2096 2097 /* 2098 * This process needs to be suspended prior to shutdown sync. 2099 */ 2100 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2101 SHUTDOWN_PRI_LAST); 2102 2103 /* 2104 * This process is allowed to take the buffer cache to the limit 2105 */ 2106 s = splbio(); 2107 mtx_lock(&bdlock); 2108 2109 for (;;) { 2110 bd_request = 0; 2111 mtx_unlock(&bdlock); 2112 2113 kthread_suspend_check(bufdaemonproc); 2114 2115 /* 2116 * Do the flush. Limit the amount of in-transit I/O we 2117 * allow to build up, otherwise we would completely saturate 2118 * the I/O system. Wakeup any waiting processes before we 2119 * normally would so they can run in parallel with our drain. 2120 */ 2121 while (numdirtybuffers > lodirtybuffers) { 2122 if (flushbufqueues(0) == 0) { 2123 /* 2124 * Could not find any buffers without rollback 2125 * dependencies, so just write the first one 2126 * in the hopes of eventually making progress. 2127 */ 2128 flushbufqueues(1); 2129 break; 2130 } 2131 waitrunningbufspace(); 2132 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2133 } 2134 2135 /* 2136 * Only clear bd_request if we have reached our low water 2137 * mark. The buf_daemon normally waits 1 second and 2138 * then incrementally flushes any dirty buffers that have 2139 * built up, within reason. 2140 * 2141 * If we were unable to hit our low water mark and couldn't 2142 * find any flushable buffers, we sleep half a second. 2143 * Otherwise we loop immediately. 2144 */ 2145 mtx_lock(&bdlock); 2146 if (numdirtybuffers <= lodirtybuffers) { 2147 /* 2148 * We reached our low water mark, reset the 2149 * request and sleep until we are needed again. 2150 * The sleep is just so the suspend code works. 2151 */ 2152 bd_request = 0; 2153 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2154 } else { 2155 /* 2156 * We couldn't find any flushable dirty buffers but 2157 * still have too many dirty buffers, we 2158 * have to sleep and try again. (rare) 2159 */ 2160 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2161 } 2162 } 2163} 2164 2165/* 2166 * flushbufqueues: 2167 * 2168 * Try to flush a buffer in the dirty queue. We must be careful to 2169 * free up B_INVAL buffers instead of write them, which NFS is 2170 * particularly sensitive to. 2171 */ 2172int flushwithdeps = 0; 2173SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2174 0, "Number of buffers flushed with dependecies that require rollbacks"); 2175 2176static int 2177flushbufqueues(int flushdeps) 2178{ 2179 struct thread *td = curthread; 2180 struct vnode *vp; 2181 struct mount *mp; 2182 struct buf *bp; 2183 int hasdeps; 2184 2185 mtx_lock(&bqlock); 2186 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2187 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2188 continue; 2189 KASSERT((bp->b_flags & B_DELWRI), 2190 ("unexpected clean buffer %p", bp)); 2191 BO_LOCK(bp->b_bufobj); 2192 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 2193 BO_UNLOCK(bp->b_bufobj); 2194 BUF_UNLOCK(bp); 2195 continue; 2196 } 2197 BO_UNLOCK(bp->b_bufobj); 2198 if (bp->b_flags & B_INVAL) { 2199 bremfreel(bp); 2200 mtx_unlock(&bqlock); 2201 brelse(bp); 2202 return (1); 2203 } 2204 2205 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2206 if (flushdeps == 0) { 2207 BUF_UNLOCK(bp); 2208 continue; 2209 } 2210 hasdeps = 1; 2211 } else 2212 hasdeps = 0; 2213 /* 2214 * We must hold the lock on a vnode before writing 2215 * one of its buffers. Otherwise we may confuse, or 2216 * in the case of a snapshot vnode, deadlock the 2217 * system. 2218 * 2219 * The lock order here is the reverse of the normal 2220 * of vnode followed by buf lock. This is ok because 2221 * the NOWAIT will prevent deadlock. 2222 */ 2223 vp = bp->b_vp; 2224 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2225 BUF_UNLOCK(bp); 2226 continue; 2227 } 2228 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2229 mtx_unlock(&bqlock); 2230 vfs_bio_awrite(bp); 2231 vn_finished_write(mp); 2232 VOP_UNLOCK(vp, 0, td); 2233 flushwithdeps += hasdeps; 2234 return (1); 2235 } 2236 vn_finished_write(mp); 2237 BUF_UNLOCK(bp); 2238 } 2239 mtx_unlock(&bqlock); 2240 return (0); 2241} 2242 2243/* 2244 * Check to see if a block is currently memory resident. 2245 */ 2246struct buf * 2247incore(struct bufobj *bo, daddr_t blkno) 2248{ 2249 struct buf *bp; 2250 2251 int s = splbio(); 2252 BO_LOCK(bo); 2253 bp = gbincore(bo, blkno); 2254 BO_UNLOCK(bo); 2255 splx(s); 2256 return (bp); 2257} 2258 2259/* 2260 * Returns true if no I/O is needed to access the 2261 * associated VM object. This is like incore except 2262 * it also hunts around in the VM system for the data. 2263 */ 2264 2265static int 2266inmem(struct vnode * vp, daddr_t blkno) 2267{ 2268 vm_object_t obj; 2269 vm_offset_t toff, tinc, size; 2270 vm_page_t m; 2271 vm_ooffset_t off; 2272 2273 GIANT_REQUIRED; 2274 ASSERT_VOP_LOCKED(vp, "inmem"); 2275 2276 if (incore(&vp->v_bufobj, blkno)) 2277 return 1; 2278 if (vp->v_mount == NULL) 2279 return 0; 2280 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2281 return 0; 2282 2283 size = PAGE_SIZE; 2284 if (size > vp->v_mount->mnt_stat.f_iosize) 2285 size = vp->v_mount->mnt_stat.f_iosize; 2286 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2287 2288 VM_OBJECT_LOCK(obj); 2289 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2290 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2291 if (!m) 2292 goto notinmem; 2293 tinc = size; 2294 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2295 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2296 if (vm_page_is_valid(m, 2297 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2298 goto notinmem; 2299 } 2300 VM_OBJECT_UNLOCK(obj); 2301 return 1; 2302 2303notinmem: 2304 VM_OBJECT_UNLOCK(obj); 2305 return (0); 2306} 2307 2308/* 2309 * vfs_setdirty: 2310 * 2311 * Sets the dirty range for a buffer based on the status of the dirty 2312 * bits in the pages comprising the buffer. 2313 * 2314 * The range is limited to the size of the buffer. 2315 * 2316 * This routine is primarily used by NFS, but is generalized for the 2317 * B_VMIO case. 2318 */ 2319static void 2320vfs_setdirty(struct buf *bp) 2321{ 2322 int i; 2323 vm_object_t object; 2324 2325 GIANT_REQUIRED; 2326 /* 2327 * Degenerate case - empty buffer 2328 */ 2329 2330 if (bp->b_bufsize == 0) 2331 return; 2332 2333 /* 2334 * We qualify the scan for modified pages on whether the 2335 * object has been flushed yet. The OBJ_WRITEABLE flag 2336 * is not cleared simply by protecting pages off. 2337 */ 2338 2339 if ((bp->b_flags & B_VMIO) == 0) 2340 return; 2341 2342 object = bp->b_pages[0]->object; 2343 VM_OBJECT_LOCK(object); 2344 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2345 printf("Warning: object %p writeable but not mightbedirty\n", object); 2346 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2347 printf("Warning: object %p mightbedirty but not writeable\n", object); 2348 2349 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2350 vm_offset_t boffset; 2351 vm_offset_t eoffset; 2352 2353 vm_page_lock_queues(); 2354 /* 2355 * test the pages to see if they have been modified directly 2356 * by users through the VM system. 2357 */ 2358 for (i = 0; i < bp->b_npages; i++) 2359 vm_page_test_dirty(bp->b_pages[i]); 2360 2361 /* 2362 * Calculate the encompassing dirty range, boffset and eoffset, 2363 * (eoffset - boffset) bytes. 2364 */ 2365 2366 for (i = 0; i < bp->b_npages; i++) { 2367 if (bp->b_pages[i]->dirty) 2368 break; 2369 } 2370 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2371 2372 for (i = bp->b_npages - 1; i >= 0; --i) { 2373 if (bp->b_pages[i]->dirty) { 2374 break; 2375 } 2376 } 2377 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2378 2379 vm_page_unlock_queues(); 2380 /* 2381 * Fit it to the buffer. 2382 */ 2383 2384 if (eoffset > bp->b_bcount) 2385 eoffset = bp->b_bcount; 2386 2387 /* 2388 * If we have a good dirty range, merge with the existing 2389 * dirty range. 2390 */ 2391 2392 if (boffset < eoffset) { 2393 if (bp->b_dirtyoff > boffset) 2394 bp->b_dirtyoff = boffset; 2395 if (bp->b_dirtyend < eoffset) 2396 bp->b_dirtyend = eoffset; 2397 } 2398 } 2399 VM_OBJECT_UNLOCK(object); 2400} 2401 2402/* 2403 * getblk: 2404 * 2405 * Get a block given a specified block and offset into a file/device. 2406 * The buffers B_DONE bit will be cleared on return, making it almost 2407 * ready for an I/O initiation. B_INVAL may or may not be set on 2408 * return. The caller should clear B_INVAL prior to initiating a 2409 * READ. 2410 * 2411 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2412 * an existing buffer. 2413 * 2414 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2415 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2416 * and then cleared based on the backing VM. If the previous buffer is 2417 * non-0-sized but invalid, B_CACHE will be cleared. 2418 * 2419 * If getblk() must create a new buffer, the new buffer is returned with 2420 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2421 * case it is returned with B_INVAL clear and B_CACHE set based on the 2422 * backing VM. 2423 * 2424 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2425 * B_CACHE bit is clear. 2426 * 2427 * What this means, basically, is that the caller should use B_CACHE to 2428 * determine whether the buffer is fully valid or not and should clear 2429 * B_INVAL prior to issuing a read. If the caller intends to validate 2430 * the buffer by loading its data area with something, the caller needs 2431 * to clear B_INVAL. If the caller does this without issuing an I/O, 2432 * the caller should set B_CACHE ( as an optimization ), else the caller 2433 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2434 * a write attempt or if it was a successfull read. If the caller 2435 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2436 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2437 */ 2438struct buf * 2439getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2440 int flags) 2441{ 2442 struct buf *bp; 2443 struct bufobj *bo; 2444 int s; 2445 int error; 2446 ASSERT_VOP_LOCKED(vp, "getblk"); 2447 struct vm_object *vmo; 2448 2449 if (size > MAXBSIZE) 2450 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2451 2452 bo = &vp->v_bufobj; 2453 s = splbio(); 2454loop: 2455 /* 2456 * Block if we are low on buffers. Certain processes are allowed 2457 * to completely exhaust the buffer cache. 2458 * 2459 * If this check ever becomes a bottleneck it may be better to 2460 * move it into the else, when gbincore() fails. At the moment 2461 * it isn't a problem. 2462 * 2463 * XXX remove if 0 sections (clean this up after its proven) 2464 */ 2465 if (numfreebuffers == 0) { 2466 if (curthread == PCPU_GET(idlethread)) 2467 return NULL; 2468 mtx_lock(&nblock); 2469 needsbuffer |= VFS_BIO_NEED_ANY; 2470 mtx_unlock(&nblock); 2471 } 2472 2473 VI_LOCK(vp); 2474 bp = gbincore(bo, blkno); 2475 if (bp != NULL) { 2476 int lockflags; 2477 /* 2478 * Buffer is in-core. If the buffer is not busy, it must 2479 * be on a queue. 2480 */ 2481 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2482 2483 if (flags & GB_LOCK_NOWAIT) 2484 lockflags |= LK_NOWAIT; 2485 2486 error = BUF_TIMELOCK(bp, lockflags, 2487 VI_MTX(vp), "getblk", slpflag, slptimeo); 2488 2489 /* 2490 * If we slept and got the lock we have to restart in case 2491 * the buffer changed identities. 2492 */ 2493 if (error == ENOLCK) 2494 goto loop; 2495 /* We timed out or were interrupted. */ 2496 else if (error) 2497 return (NULL); 2498 2499 /* 2500 * The buffer is locked. B_CACHE is cleared if the buffer is 2501 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2502 * and for a VMIO buffer B_CACHE is adjusted according to the 2503 * backing VM cache. 2504 */ 2505 if (bp->b_flags & B_INVAL) 2506 bp->b_flags &= ~B_CACHE; 2507 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2508 bp->b_flags |= B_CACHE; 2509 bremfree(bp); 2510 2511 /* 2512 * check for size inconsistancies for non-VMIO case. 2513 */ 2514 2515 if (bp->b_bcount != size) { 2516 if ((bp->b_flags & B_VMIO) == 0 || 2517 (size > bp->b_kvasize)) { 2518 if (bp->b_flags & B_DELWRI) { 2519 bp->b_flags |= B_NOCACHE; 2520 bwrite(bp); 2521 } else { 2522 if ((bp->b_flags & B_VMIO) && 2523 (LIST_FIRST(&bp->b_dep) == NULL)) { 2524 bp->b_flags |= B_RELBUF; 2525 brelse(bp); 2526 } else { 2527 bp->b_flags |= B_NOCACHE; 2528 bwrite(bp); 2529 } 2530 } 2531 goto loop; 2532 } 2533 } 2534 2535 /* 2536 * If the size is inconsistant in the VMIO case, we can resize 2537 * the buffer. This might lead to B_CACHE getting set or 2538 * cleared. If the size has not changed, B_CACHE remains 2539 * unchanged from its previous state. 2540 */ 2541 2542 if (bp->b_bcount != size) 2543 allocbuf(bp, size); 2544 2545 KASSERT(bp->b_offset != NOOFFSET, 2546 ("getblk: no buffer offset")); 2547 2548 /* 2549 * A buffer with B_DELWRI set and B_CACHE clear must 2550 * be committed before we can return the buffer in 2551 * order to prevent the caller from issuing a read 2552 * ( due to B_CACHE not being set ) and overwriting 2553 * it. 2554 * 2555 * Most callers, including NFS and FFS, need this to 2556 * operate properly either because they assume they 2557 * can issue a read if B_CACHE is not set, or because 2558 * ( for example ) an uncached B_DELWRI might loop due 2559 * to softupdates re-dirtying the buffer. In the latter 2560 * case, B_CACHE is set after the first write completes, 2561 * preventing further loops. 2562 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2563 * above while extending the buffer, we cannot allow the 2564 * buffer to remain with B_CACHE set after the write 2565 * completes or it will represent a corrupt state. To 2566 * deal with this we set B_NOCACHE to scrap the buffer 2567 * after the write. 2568 * 2569 * We might be able to do something fancy, like setting 2570 * B_CACHE in bwrite() except if B_DELWRI is already set, 2571 * so the below call doesn't set B_CACHE, but that gets real 2572 * confusing. This is much easier. 2573 */ 2574 2575 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2576 bp->b_flags |= B_NOCACHE; 2577 bwrite(bp); 2578 goto loop; 2579 } 2580 2581 splx(s); 2582 bp->b_flags &= ~B_DONE; 2583 } else { 2584 int bsize, maxsize, vmio; 2585 off_t offset; 2586 2587 /* 2588 * Buffer is not in-core, create new buffer. The buffer 2589 * returned by getnewbuf() is locked. Note that the returned 2590 * buffer is also considered valid (not marked B_INVAL). 2591 */ 2592 VI_UNLOCK(vp); 2593 /* 2594 * If the user does not want us to create the buffer, bail out 2595 * here. 2596 */ 2597 if (flags & GB_NOCREAT) { 2598 splx(s); 2599 return NULL; 2600 } 2601 2602 bsize = bo->bo_bsize; 2603 offset = blkno * bsize; 2604 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2605 (vp->v_vflag & VV_OBJBUF); 2606 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2607 maxsize = imax(maxsize, bsize); 2608 2609 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2610 if (bp == NULL) { 2611 if (slpflag || slptimeo) { 2612 splx(s); 2613 return NULL; 2614 } 2615 goto loop; 2616 } 2617 2618 /* 2619 * This code is used to make sure that a buffer is not 2620 * created while the getnewbuf routine is blocked. 2621 * This can be a problem whether the vnode is locked or not. 2622 * If the buffer is created out from under us, we have to 2623 * throw away the one we just created. There is now window 2624 * race because we are safely running at splbio() from the 2625 * point of the duplicate buffer creation through to here, 2626 * and we've locked the buffer. 2627 * 2628 * Note: this must occur before we associate the buffer 2629 * with the vp especially considering limitations in 2630 * the splay tree implementation when dealing with duplicate 2631 * lblkno's. 2632 */ 2633 BO_LOCK(bo); 2634 if (gbincore(bo, blkno)) { 2635 BO_UNLOCK(bo); 2636 bp->b_flags |= B_INVAL; 2637 brelse(bp); 2638 goto loop; 2639 } 2640 2641 /* 2642 * Insert the buffer into the hash, so that it can 2643 * be found by incore. 2644 */ 2645 bp->b_blkno = bp->b_lblkno = blkno; 2646 bp->b_offset = offset; 2647 2648 bgetvp(vp, bp); 2649 BO_UNLOCK(bo); 2650 2651 /* 2652 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2653 * buffer size starts out as 0, B_CACHE will be set by 2654 * allocbuf() for the VMIO case prior to it testing the 2655 * backing store for validity. 2656 */ 2657 2658 if (vmio) { 2659 bp->b_flags |= B_VMIO; 2660#if defined(VFS_BIO_DEBUG) 2661 if (vn_canvmio(vp) != TRUE) 2662 printf("getblk: VMIO on vnode type %d\n", 2663 vp->v_type); 2664#endif 2665 VOP_GETVOBJECT(vp, &vmo); 2666 KASSERT(vmo == bp->b_bufobj->bo_object, 2667 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2668 bp, vmo, bp->b_bufobj->bo_object)); 2669 } else { 2670 bp->b_flags &= ~B_VMIO; 2671 KASSERT(bp->b_bufobj->bo_object == NULL, 2672 ("ARGH! has b_bufobj->bo_object %p %p\n", 2673 bp, bp->b_bufobj->bo_object)); 2674 } 2675 2676 allocbuf(bp, size); 2677 2678 splx(s); 2679 bp->b_flags &= ~B_DONE; 2680 } 2681 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2682 KASSERT(bp->b_bufobj == bo, 2683 ("wrong b_bufobj %p should be %p", bp->b_bufobj, bo)); 2684 return (bp); 2685} 2686 2687/* 2688 * Get an empty, disassociated buffer of given size. The buffer is initially 2689 * set to B_INVAL. 2690 */ 2691struct buf * 2692geteblk(int size) 2693{ 2694 struct buf *bp; 2695 int s; 2696 int maxsize; 2697 2698 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2699 2700 s = splbio(); 2701 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2702 continue; 2703 splx(s); 2704 allocbuf(bp, size); 2705 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2706 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2707 return (bp); 2708} 2709 2710 2711/* 2712 * This code constitutes the buffer memory from either anonymous system 2713 * memory (in the case of non-VMIO operations) or from an associated 2714 * VM object (in the case of VMIO operations). This code is able to 2715 * resize a buffer up or down. 2716 * 2717 * Note that this code is tricky, and has many complications to resolve 2718 * deadlock or inconsistant data situations. Tread lightly!!! 2719 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2720 * the caller. Calling this code willy nilly can result in the loss of data. 2721 * 2722 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2723 * B_CACHE for the non-VMIO case. 2724 */ 2725 2726int 2727allocbuf(struct buf *bp, int size) 2728{ 2729 int newbsize, mbsize; 2730 int i; 2731 2732 GIANT_REQUIRED; 2733 2734 if (BUF_REFCNT(bp) == 0) 2735 panic("allocbuf: buffer not busy"); 2736 2737 if (bp->b_kvasize < size) 2738 panic("allocbuf: buffer too small"); 2739 2740 if ((bp->b_flags & B_VMIO) == 0) { 2741 caddr_t origbuf; 2742 int origbufsize; 2743 /* 2744 * Just get anonymous memory from the kernel. Don't 2745 * mess with B_CACHE. 2746 */ 2747 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2748 if (bp->b_flags & B_MALLOC) 2749 newbsize = mbsize; 2750 else 2751 newbsize = round_page(size); 2752 2753 if (newbsize < bp->b_bufsize) { 2754 /* 2755 * malloced buffers are not shrunk 2756 */ 2757 if (bp->b_flags & B_MALLOC) { 2758 if (newbsize) { 2759 bp->b_bcount = size; 2760 } else { 2761 free(bp->b_data, M_BIOBUF); 2762 if (bp->b_bufsize) { 2763 atomic_subtract_int( 2764 &bufmallocspace, 2765 bp->b_bufsize); 2766 bufspacewakeup(); 2767 bp->b_bufsize = 0; 2768 } 2769 bp->b_saveaddr = bp->b_kvabase; 2770 bp->b_data = bp->b_saveaddr; 2771 bp->b_bcount = 0; 2772 bp->b_flags &= ~B_MALLOC; 2773 } 2774 return 1; 2775 } 2776 vm_hold_free_pages( 2777 bp, 2778 (vm_offset_t) bp->b_data + newbsize, 2779 (vm_offset_t) bp->b_data + bp->b_bufsize); 2780 } else if (newbsize > bp->b_bufsize) { 2781 /* 2782 * We only use malloced memory on the first allocation. 2783 * and revert to page-allocated memory when the buffer 2784 * grows. 2785 */ 2786 /* 2787 * There is a potential smp race here that could lead 2788 * to bufmallocspace slightly passing the max. It 2789 * is probably extremely rare and not worth worrying 2790 * over. 2791 */ 2792 if ( (bufmallocspace < maxbufmallocspace) && 2793 (bp->b_bufsize == 0) && 2794 (mbsize <= PAGE_SIZE/2)) { 2795 2796 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2797 bp->b_bufsize = mbsize; 2798 bp->b_bcount = size; 2799 bp->b_flags |= B_MALLOC; 2800 atomic_add_int(&bufmallocspace, mbsize); 2801 return 1; 2802 } 2803 origbuf = NULL; 2804 origbufsize = 0; 2805 /* 2806 * If the buffer is growing on its other-than-first allocation, 2807 * then we revert to the page-allocation scheme. 2808 */ 2809 if (bp->b_flags & B_MALLOC) { 2810 origbuf = bp->b_data; 2811 origbufsize = bp->b_bufsize; 2812 bp->b_data = bp->b_kvabase; 2813 if (bp->b_bufsize) { 2814 atomic_subtract_int(&bufmallocspace, 2815 bp->b_bufsize); 2816 bufspacewakeup(); 2817 bp->b_bufsize = 0; 2818 } 2819 bp->b_flags &= ~B_MALLOC; 2820 newbsize = round_page(newbsize); 2821 } 2822 vm_hold_load_pages( 2823 bp, 2824 (vm_offset_t) bp->b_data + bp->b_bufsize, 2825 (vm_offset_t) bp->b_data + newbsize); 2826 if (origbuf) { 2827 bcopy(origbuf, bp->b_data, origbufsize); 2828 free(origbuf, M_BIOBUF); 2829 } 2830 } 2831 } else { 2832 int desiredpages; 2833 2834 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2835 desiredpages = (size == 0) ? 0 : 2836 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2837 2838 if (bp->b_flags & B_MALLOC) 2839 panic("allocbuf: VMIO buffer can't be malloced"); 2840 /* 2841 * Set B_CACHE initially if buffer is 0 length or will become 2842 * 0-length. 2843 */ 2844 if (size == 0 || bp->b_bufsize == 0) 2845 bp->b_flags |= B_CACHE; 2846 2847 if (newbsize < bp->b_bufsize) { 2848 /* 2849 * DEV_BSIZE aligned new buffer size is less then the 2850 * DEV_BSIZE aligned existing buffer size. Figure out 2851 * if we have to remove any pages. 2852 */ 2853 if (desiredpages < bp->b_npages) { 2854 vm_page_t m; 2855 2856 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2857 vm_page_lock_queues(); 2858 for (i = desiredpages; i < bp->b_npages; i++) { 2859 /* 2860 * the page is not freed here -- it 2861 * is the responsibility of 2862 * vnode_pager_setsize 2863 */ 2864 m = bp->b_pages[i]; 2865 KASSERT(m != bogus_page, 2866 ("allocbuf: bogus page found")); 2867 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2868 vm_page_lock_queues(); 2869 2870 bp->b_pages[i] = NULL; 2871 vm_page_unwire(m, 0); 2872 } 2873 vm_page_unlock_queues(); 2874 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2875 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2876 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2877 bp->b_npages = desiredpages; 2878 } 2879 } else if (size > bp->b_bcount) { 2880 /* 2881 * We are growing the buffer, possibly in a 2882 * byte-granular fashion. 2883 */ 2884 struct vnode *vp; 2885 vm_object_t obj; 2886 vm_offset_t toff; 2887 vm_offset_t tinc; 2888 2889 /* 2890 * Step 1, bring in the VM pages from the object, 2891 * allocating them if necessary. We must clear 2892 * B_CACHE if these pages are not valid for the 2893 * range covered by the buffer. 2894 */ 2895 2896 vp = bp->b_vp; 2897 obj = bp->b_bufobj->bo_object; 2898 2899 VM_OBJECT_LOCK(obj); 2900 while (bp->b_npages < desiredpages) { 2901 vm_page_t m; 2902 vm_pindex_t pi; 2903 2904 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2905 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2906 /* 2907 * note: must allocate system pages 2908 * since blocking here could intefere 2909 * with paging I/O, no matter which 2910 * process we are. 2911 */ 2912 m = vm_page_alloc(obj, pi, 2913 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2914 VM_ALLOC_WIRED); 2915 if (m == NULL) { 2916 atomic_add_int(&vm_pageout_deficit, 2917 desiredpages - bp->b_npages); 2918 VM_OBJECT_UNLOCK(obj); 2919 VM_WAIT; 2920 VM_OBJECT_LOCK(obj); 2921 } else { 2922 bp->b_flags &= ~B_CACHE; 2923 bp->b_pages[bp->b_npages] = m; 2924 ++bp->b_npages; 2925 } 2926 continue; 2927 } 2928 2929 /* 2930 * We found a page. If we have to sleep on it, 2931 * retry because it might have gotten freed out 2932 * from under us. 2933 * 2934 * We can only test PG_BUSY here. Blocking on 2935 * m->busy might lead to a deadlock: 2936 * 2937 * vm_fault->getpages->cluster_read->allocbuf 2938 * 2939 */ 2940 vm_page_lock_queues(); 2941 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2942 continue; 2943 2944 /* 2945 * We have a good page. Should we wakeup the 2946 * page daemon? 2947 */ 2948 if ((curproc != pageproc) && 2949 ((m->queue - m->pc) == PQ_CACHE) && 2950 ((cnt.v_free_count + cnt.v_cache_count) < 2951 (cnt.v_free_min + cnt.v_cache_min))) { 2952 pagedaemon_wakeup(); 2953 } 2954 vm_page_wire(m); 2955 vm_page_unlock_queues(); 2956 bp->b_pages[bp->b_npages] = m; 2957 ++bp->b_npages; 2958 } 2959 2960 /* 2961 * Step 2. We've loaded the pages into the buffer, 2962 * we have to figure out if we can still have B_CACHE 2963 * set. Note that B_CACHE is set according to the 2964 * byte-granular range ( bcount and size ), new the 2965 * aligned range ( newbsize ). 2966 * 2967 * The VM test is against m->valid, which is DEV_BSIZE 2968 * aligned. Needless to say, the validity of the data 2969 * needs to also be DEV_BSIZE aligned. Note that this 2970 * fails with NFS if the server or some other client 2971 * extends the file's EOF. If our buffer is resized, 2972 * B_CACHE may remain set! XXX 2973 */ 2974 2975 toff = bp->b_bcount; 2976 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2977 2978 while ((bp->b_flags & B_CACHE) && toff < size) { 2979 vm_pindex_t pi; 2980 2981 if (tinc > (size - toff)) 2982 tinc = size - toff; 2983 2984 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2985 PAGE_SHIFT; 2986 2987 vfs_buf_test_cache( 2988 bp, 2989 bp->b_offset, 2990 toff, 2991 tinc, 2992 bp->b_pages[pi] 2993 ); 2994 toff += tinc; 2995 tinc = PAGE_SIZE; 2996 } 2997 VM_OBJECT_UNLOCK(obj); 2998 2999 /* 3000 * Step 3, fixup the KVM pmap. Remember that 3001 * bp->b_data is relative to bp->b_offset, but 3002 * bp->b_offset may be offset into the first page. 3003 */ 3004 3005 bp->b_data = (caddr_t) 3006 trunc_page((vm_offset_t)bp->b_data); 3007 pmap_qenter( 3008 (vm_offset_t)bp->b_data, 3009 bp->b_pages, 3010 bp->b_npages 3011 ); 3012 3013 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3014 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3015 } 3016 } 3017 if (newbsize < bp->b_bufsize) 3018 bufspacewakeup(); 3019 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3020 bp->b_bcount = size; /* requested buffer size */ 3021 return 1; 3022} 3023 3024void 3025biodone(struct bio *bp) 3026{ 3027 3028 mtx_lock(&bdonelock); 3029 bp->bio_flags |= BIO_DONE; 3030 if (bp->bio_done == NULL) 3031 wakeup(bp); 3032 mtx_unlock(&bdonelock); 3033 if (bp->bio_done != NULL) 3034 bp->bio_done(bp); 3035} 3036 3037/* 3038 * Wait for a BIO to finish. 3039 * 3040 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3041 * case is not yet clear. 3042 */ 3043int 3044biowait(struct bio *bp, const char *wchan) 3045{ 3046 3047 mtx_lock(&bdonelock); 3048 while ((bp->bio_flags & BIO_DONE) == 0) 3049 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3050 mtx_unlock(&bdonelock); 3051 if (bp->bio_error != 0) 3052 return (bp->bio_error); 3053 if (!(bp->bio_flags & BIO_ERROR)) 3054 return (0); 3055 return (EIO); 3056} 3057 3058void 3059biofinish(struct bio *bp, struct devstat *stat, int error) 3060{ 3061 3062 if (error) { 3063 bp->bio_error = error; 3064 bp->bio_flags |= BIO_ERROR; 3065 } 3066 if (stat != NULL) 3067 devstat_end_transaction_bio(stat, bp); 3068 biodone(bp); 3069} 3070 3071/* 3072 * bufwait: 3073 * 3074 * Wait for buffer I/O completion, returning error status. The buffer 3075 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3076 * error and cleared. 3077 */ 3078int 3079bufwait(struct buf *bp) 3080{ 3081 int s; 3082 3083 s = splbio(); 3084 if (bp->b_iocmd == BIO_READ) 3085 bwait(bp, PRIBIO, "biord"); 3086 else 3087 bwait(bp, PRIBIO, "biowr"); 3088 splx(s); 3089 if (bp->b_flags & B_EINTR) { 3090 bp->b_flags &= ~B_EINTR; 3091 return (EINTR); 3092 } 3093 if (bp->b_ioflags & BIO_ERROR) { 3094 return (bp->b_error ? bp->b_error : EIO); 3095 } else { 3096 return (0); 3097 } 3098} 3099 3100 /* 3101 * Call back function from struct bio back up to struct buf. 3102 */ 3103static void 3104bufdonebio(struct bio *bip) 3105{ 3106 struct buf *bp; 3107 3108 /* Device drivers may or may not hold giant, hold it here. */ 3109 mtx_lock(&Giant); 3110 bp = bip->bio_caller2; 3111 bp->b_resid = bp->b_bcount - bip->bio_completed; 3112 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3113 bp->b_ioflags = bip->bio_flags; 3114 bp->b_error = bip->bio_error; 3115 if (bp->b_error) 3116 bp->b_ioflags |= BIO_ERROR; 3117 bufdone(bp); 3118 mtx_unlock(&Giant); 3119 g_destroy_bio(bip); 3120} 3121 3122void 3123dev_strategy(struct cdev *dev, struct buf *bp) 3124{ 3125 struct cdevsw *csw; 3126 struct bio *bip; 3127 3128 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3129 panic("b_iocmd botch"); 3130 for (;;) { 3131 bip = g_new_bio(); 3132 if (bip != NULL) 3133 break; 3134 /* Try again later */ 3135 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3136 } 3137 bip->bio_cmd = bp->b_iocmd; 3138 bip->bio_offset = bp->b_iooffset; 3139 bip->bio_length = bp->b_bcount; 3140 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3141 bip->bio_data = bp->b_data; 3142 bip->bio_done = bufdonebio; 3143 bip->bio_caller2 = bp; 3144 bip->bio_dev = dev; 3145 KASSERT(dev->si_refcount > 0, 3146 ("dev_strategy on un-referenced struct cdev *(%s)", 3147 devtoname(dev))); 3148 csw = dev_refthread(dev); 3149 if (csw == NULL) { 3150 bp->b_error = ENXIO; 3151 bp->b_ioflags = BIO_ERROR; 3152 mtx_lock(&Giant); /* XXX: too defensive ? */ 3153 bufdone(bp); 3154 mtx_unlock(&Giant); /* XXX: too defensive ? */ 3155 return; 3156 } 3157 (*csw->d_strategy)(bip); 3158 dev_relthread(dev); 3159} 3160 3161/* 3162 * bufdone: 3163 * 3164 * Finish I/O on a buffer, optionally calling a completion function. 3165 * This is usually called from an interrupt so process blocking is 3166 * not allowed. 3167 * 3168 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3169 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3170 * assuming B_INVAL is clear. 3171 * 3172 * For the VMIO case, we set B_CACHE if the op was a read and no 3173 * read error occured, or if the op was a write. B_CACHE is never 3174 * set if the buffer is invalid or otherwise uncacheable. 3175 * 3176 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3177 * initiator to leave B_INVAL set to brelse the buffer out of existance 3178 * in the biodone routine. 3179 */ 3180void 3181bufdone(struct buf *bp) 3182{ 3183 int s; 3184 void (*biodone)(struct buf *); 3185 3186 3187 s = splbio(); 3188 3189 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3190 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3191 3192 bp->b_flags |= B_DONE; 3193 runningbufwakeup(bp); 3194 3195 if (bp->b_iocmd == BIO_WRITE && bp->b_bufobj != NULL) 3196 bufobj_wdrop(bp->b_bufobj); 3197 3198 /* call optional completion function if requested */ 3199 if (bp->b_iodone != NULL) { 3200 biodone = bp->b_iodone; 3201 bp->b_iodone = NULL; 3202 (*biodone) (bp); 3203 splx(s); 3204 return; 3205 } 3206 if (LIST_FIRST(&bp->b_dep) != NULL) 3207 buf_complete(bp); 3208 3209 if (bp->b_flags & B_VMIO) { 3210 int i; 3211 vm_ooffset_t foff; 3212 vm_page_t m; 3213 vm_object_t obj; 3214 int iosize; 3215 struct vnode *vp = bp->b_vp; 3216 3217 obj = bp->b_bufobj->bo_object; 3218 3219#if defined(VFS_BIO_DEBUG) 3220 mp_fixme("usecount and vflag accessed without locks."); 3221 if (vp->v_usecount == 0) { 3222 panic("biodone: zero vnode ref count"); 3223 } 3224 3225 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3226 panic("biodone: vnode is not setup for merged cache"); 3227 } 3228#endif 3229 3230 foff = bp->b_offset; 3231 KASSERT(bp->b_offset != NOOFFSET, 3232 ("biodone: no buffer offset")); 3233 3234 VM_OBJECT_LOCK(obj); 3235#if defined(VFS_BIO_DEBUG) 3236 if (obj->paging_in_progress < bp->b_npages) { 3237 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3238 obj->paging_in_progress, bp->b_npages); 3239 } 3240#endif 3241 3242 /* 3243 * Set B_CACHE if the op was a normal read and no error 3244 * occured. B_CACHE is set for writes in the b*write() 3245 * routines. 3246 */ 3247 iosize = bp->b_bcount - bp->b_resid; 3248 if (bp->b_iocmd == BIO_READ && 3249 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3250 !(bp->b_ioflags & BIO_ERROR)) { 3251 bp->b_flags |= B_CACHE; 3252 } 3253 vm_page_lock_queues(); 3254 for (i = 0; i < bp->b_npages; i++) { 3255 int bogusflag = 0; 3256 int resid; 3257 3258 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3259 if (resid > iosize) 3260 resid = iosize; 3261 3262 /* 3263 * cleanup bogus pages, restoring the originals 3264 */ 3265 m = bp->b_pages[i]; 3266 if (m == bogus_page) { 3267 bogusflag = 1; 3268 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3269 if (m == NULL) 3270 panic("biodone: page disappeared!"); 3271 bp->b_pages[i] = m; 3272 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3273 } 3274#if defined(VFS_BIO_DEBUG) 3275 if (OFF_TO_IDX(foff) != m->pindex) { 3276 printf( 3277"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3278 (intmax_t)foff, (uintmax_t)m->pindex); 3279 } 3280#endif 3281 3282 /* 3283 * In the write case, the valid and clean bits are 3284 * already changed correctly ( see bdwrite() ), so we 3285 * only need to do this here in the read case. 3286 */ 3287 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3288 vfs_page_set_valid(bp, foff, i, m); 3289 } 3290 3291 /* 3292 * when debugging new filesystems or buffer I/O methods, this 3293 * is the most common error that pops up. if you see this, you 3294 * have not set the page busy flag correctly!!! 3295 */ 3296 if (m->busy == 0) { 3297 printf("biodone: page busy < 0, " 3298 "pindex: %d, foff: 0x(%x,%x), " 3299 "resid: %d, index: %d\n", 3300 (int) m->pindex, (int)(foff >> 32), 3301 (int) foff & 0xffffffff, resid, i); 3302 if (!vn_isdisk(vp, NULL)) 3303 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3304 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3305 (intmax_t) bp->b_lblkno, 3306 bp->b_flags, bp->b_npages); 3307 else 3308 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3309 (intmax_t) bp->b_lblkno, 3310 bp->b_flags, bp->b_npages); 3311 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3312 (u_long)m->valid, (u_long)m->dirty, 3313 m->wire_count); 3314 panic("biodone: page busy < 0\n"); 3315 } 3316 vm_page_io_finish(m); 3317 vm_object_pip_subtract(obj, 1); 3318 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3319 iosize -= resid; 3320 } 3321 vm_page_unlock_queues(); 3322 vm_object_pip_wakeupn(obj, 0); 3323 VM_OBJECT_UNLOCK(obj); 3324 } 3325 3326 /* 3327 * For asynchronous completions, release the buffer now. The brelse 3328 * will do a wakeup there if necessary - so no need to do a wakeup 3329 * here in the async case. The sync case always needs to do a wakeup. 3330 */ 3331 3332 if (bp->b_flags & B_ASYNC) { 3333 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3334 brelse(bp); 3335 else 3336 bqrelse(bp); 3337 } else { 3338 bdone(bp); 3339 } 3340 splx(s); 3341} 3342 3343/* 3344 * This routine is called in lieu of iodone in the case of 3345 * incomplete I/O. This keeps the busy status for pages 3346 * consistant. 3347 */ 3348void 3349vfs_unbusy_pages(struct buf *bp) 3350{ 3351 int i; 3352 vm_object_t obj; 3353 vm_page_t m; 3354 3355 runningbufwakeup(bp); 3356 if (!(bp->b_flags & B_VMIO)) 3357 return; 3358 3359 obj = bp->b_bufobj->bo_object; 3360 VM_OBJECT_LOCK(obj); 3361 vm_page_lock_queues(); 3362 for (i = 0; i < bp->b_npages; i++) { 3363 m = bp->b_pages[i]; 3364 if (m == bogus_page) { 3365 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3366 if (!m) 3367 panic("vfs_unbusy_pages: page missing\n"); 3368 bp->b_pages[i] = m; 3369 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3370 bp->b_pages, bp->b_npages); 3371 } 3372 vm_object_pip_subtract(obj, 1); 3373 vm_page_io_finish(m); 3374 } 3375 vm_page_unlock_queues(); 3376 vm_object_pip_wakeupn(obj, 0); 3377 VM_OBJECT_UNLOCK(obj); 3378} 3379 3380/* 3381 * vfs_page_set_valid: 3382 * 3383 * Set the valid bits in a page based on the supplied offset. The 3384 * range is restricted to the buffer's size. 3385 * 3386 * This routine is typically called after a read completes. 3387 */ 3388static void 3389vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3390{ 3391 vm_ooffset_t soff, eoff; 3392 3393 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3394 /* 3395 * Start and end offsets in buffer. eoff - soff may not cross a 3396 * page boundry or cross the end of the buffer. The end of the 3397 * buffer, in this case, is our file EOF, not the allocation size 3398 * of the buffer. 3399 */ 3400 soff = off; 3401 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3402 if (eoff > bp->b_offset + bp->b_bcount) 3403 eoff = bp->b_offset + bp->b_bcount; 3404 3405 /* 3406 * Set valid range. This is typically the entire buffer and thus the 3407 * entire page. 3408 */ 3409 if (eoff > soff) { 3410 vm_page_set_validclean( 3411 m, 3412 (vm_offset_t) (soff & PAGE_MASK), 3413 (vm_offset_t) (eoff - soff) 3414 ); 3415 } 3416} 3417 3418/* 3419 * This routine is called before a device strategy routine. 3420 * It is used to tell the VM system that paging I/O is in 3421 * progress, and treat the pages associated with the buffer 3422 * almost as being PG_BUSY. Also the object paging_in_progress 3423 * flag is handled to make sure that the object doesn't become 3424 * inconsistant. 3425 * 3426 * Since I/O has not been initiated yet, certain buffer flags 3427 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3428 * and should be ignored. 3429 */ 3430void 3431vfs_busy_pages(struct buf *bp, int clear_modify) 3432{ 3433 int i, bogus; 3434 vm_object_t obj; 3435 vm_ooffset_t foff; 3436 vm_page_t m; 3437 3438 if (!(bp->b_flags & B_VMIO)) 3439 return; 3440 3441 obj = bp->b_bufobj->bo_object; 3442 foff = bp->b_offset; 3443 KASSERT(bp->b_offset != NOOFFSET, 3444 ("vfs_busy_pages: no buffer offset")); 3445 vfs_setdirty(bp); 3446 VM_OBJECT_LOCK(obj); 3447retry: 3448 vm_page_lock_queues(); 3449 for (i = 0; i < bp->b_npages; i++) { 3450 m = bp->b_pages[i]; 3451 3452 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3453 goto retry; 3454 } 3455 bogus = 0; 3456 for (i = 0; i < bp->b_npages; i++) { 3457 m = bp->b_pages[i]; 3458 3459 if ((bp->b_flags & B_CLUSTER) == 0) { 3460 vm_object_pip_add(obj, 1); 3461 vm_page_io_start(m); 3462 } 3463 /* 3464 * When readying a buffer for a read ( i.e 3465 * clear_modify == 0 ), it is important to do 3466 * bogus_page replacement for valid pages in 3467 * partially instantiated buffers. Partially 3468 * instantiated buffers can, in turn, occur when 3469 * reconstituting a buffer from its VM backing store 3470 * base. We only have to do this if B_CACHE is 3471 * clear ( which causes the I/O to occur in the 3472 * first place ). The replacement prevents the read 3473 * I/O from overwriting potentially dirty VM-backed 3474 * pages. XXX bogus page replacement is, uh, bogus. 3475 * It may not work properly with small-block devices. 3476 * We need to find a better way. 3477 */ 3478 pmap_remove_all(m); 3479 if (clear_modify) 3480 vfs_page_set_valid(bp, foff, i, m); 3481 else if (m->valid == VM_PAGE_BITS_ALL && 3482 (bp->b_flags & B_CACHE) == 0) { 3483 bp->b_pages[i] = bogus_page; 3484 bogus++; 3485 } 3486 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3487 } 3488 vm_page_unlock_queues(); 3489 VM_OBJECT_UNLOCK(obj); 3490 if (bogus) 3491 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3492 bp->b_pages, bp->b_npages); 3493} 3494 3495/* 3496 * Tell the VM system that the pages associated with this buffer 3497 * are clean. This is used for delayed writes where the data is 3498 * going to go to disk eventually without additional VM intevention. 3499 * 3500 * Note that while we only really need to clean through to b_bcount, we 3501 * just go ahead and clean through to b_bufsize. 3502 */ 3503static void 3504vfs_clean_pages(struct buf *bp) 3505{ 3506 int i; 3507 vm_ooffset_t foff, noff, eoff; 3508 vm_page_t m; 3509 3510 if (!(bp->b_flags & B_VMIO)) 3511 return; 3512 3513 foff = bp->b_offset; 3514 KASSERT(bp->b_offset != NOOFFSET, 3515 ("vfs_clean_pages: no buffer offset")); 3516 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3517 vm_page_lock_queues(); 3518 for (i = 0; i < bp->b_npages; i++) { 3519 m = bp->b_pages[i]; 3520 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3521 eoff = noff; 3522 3523 if (eoff > bp->b_offset + bp->b_bufsize) 3524 eoff = bp->b_offset + bp->b_bufsize; 3525 vfs_page_set_valid(bp, foff, i, m); 3526 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3527 foff = noff; 3528 } 3529 vm_page_unlock_queues(); 3530 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3531} 3532 3533/* 3534 * vfs_bio_set_validclean: 3535 * 3536 * Set the range within the buffer to valid and clean. The range is 3537 * relative to the beginning of the buffer, b_offset. Note that b_offset 3538 * itself may be offset from the beginning of the first page. 3539 * 3540 */ 3541 3542void 3543vfs_bio_set_validclean(struct buf *bp, int base, int size) 3544{ 3545 int i, n; 3546 vm_page_t m; 3547 3548 if (!(bp->b_flags & B_VMIO)) 3549 return; 3550 /* 3551 * Fixup base to be relative to beginning of first page. 3552 * Set initial n to be the maximum number of bytes in the 3553 * first page that can be validated. 3554 */ 3555 3556 base += (bp->b_offset & PAGE_MASK); 3557 n = PAGE_SIZE - (base & PAGE_MASK); 3558 3559 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3560 vm_page_lock_queues(); 3561 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3562 m = bp->b_pages[i]; 3563 if (n > size) 3564 n = size; 3565 vm_page_set_validclean(m, base & PAGE_MASK, n); 3566 base += n; 3567 size -= n; 3568 n = PAGE_SIZE; 3569 } 3570 vm_page_unlock_queues(); 3571 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3572} 3573 3574/* 3575 * vfs_bio_clrbuf: 3576 * 3577 * clear a buffer. This routine essentially fakes an I/O, so we need 3578 * to clear BIO_ERROR and B_INVAL. 3579 * 3580 * Note that while we only theoretically need to clear through b_bcount, 3581 * we go ahead and clear through b_bufsize. 3582 */ 3583 3584void 3585vfs_bio_clrbuf(struct buf *bp) 3586{ 3587 int i, j, mask = 0; 3588 caddr_t sa, ea; 3589 3590 GIANT_REQUIRED; 3591 3592 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3593 clrbuf(bp); 3594 return; 3595 } 3596 3597 bp->b_flags &= ~B_INVAL; 3598 bp->b_ioflags &= ~BIO_ERROR; 3599 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3600 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3601 (bp->b_offset & PAGE_MASK) == 0) { 3602 if (bp->b_pages[0] == bogus_page) 3603 goto unlock; 3604 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3605 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3606 if ((bp->b_pages[0]->valid & mask) == mask) 3607 goto unlock; 3608 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3609 ((bp->b_pages[0]->valid & mask) == 0)) { 3610 bzero(bp->b_data, bp->b_bufsize); 3611 bp->b_pages[0]->valid |= mask; 3612 goto unlock; 3613 } 3614 } 3615 ea = sa = bp->b_data; 3616 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3617 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3618 ea = (caddr_t)(vm_offset_t)ulmin( 3619 (u_long)(vm_offset_t)ea, 3620 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3621 if (bp->b_pages[i] == bogus_page) 3622 continue; 3623 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3624 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3625 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3626 if ((bp->b_pages[i]->valid & mask) == mask) 3627 continue; 3628 if ((bp->b_pages[i]->valid & mask) == 0) { 3629 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3630 bzero(sa, ea - sa); 3631 } else { 3632 for (; sa < ea; sa += DEV_BSIZE, j++) { 3633 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3634 (bp->b_pages[i]->valid & (1 << j)) == 0) 3635 bzero(sa, DEV_BSIZE); 3636 } 3637 } 3638 bp->b_pages[i]->valid |= mask; 3639 } 3640unlock: 3641 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3642 bp->b_resid = 0; 3643} 3644 3645/* 3646 * vm_hold_load_pages and vm_hold_free_pages get pages into 3647 * a buffers address space. The pages are anonymous and are 3648 * not associated with a file object. 3649 */ 3650static void 3651vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3652{ 3653 vm_offset_t pg; 3654 vm_page_t p; 3655 int index; 3656 3657 to = round_page(to); 3658 from = round_page(from); 3659 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3660 3661 VM_OBJECT_LOCK(kernel_object); 3662 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3663tryagain: 3664 /* 3665 * note: must allocate system pages since blocking here 3666 * could intefere with paging I/O, no matter which 3667 * process we are. 3668 */ 3669 p = vm_page_alloc(kernel_object, 3670 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3671 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3672 if (!p) { 3673 atomic_add_int(&vm_pageout_deficit, 3674 (to - pg) >> PAGE_SHIFT); 3675 VM_OBJECT_UNLOCK(kernel_object); 3676 VM_WAIT; 3677 VM_OBJECT_LOCK(kernel_object); 3678 goto tryagain; 3679 } 3680 p->valid = VM_PAGE_BITS_ALL; 3681 pmap_qenter(pg, &p, 1); 3682 bp->b_pages[index] = p; 3683 } 3684 VM_OBJECT_UNLOCK(kernel_object); 3685 bp->b_npages = index; 3686} 3687 3688/* Return pages associated with this buf to the vm system */ 3689static void 3690vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3691{ 3692 vm_offset_t pg; 3693 vm_page_t p; 3694 int index, newnpages; 3695 3696 from = round_page(from); 3697 to = round_page(to); 3698 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3699 3700 VM_OBJECT_LOCK(kernel_object); 3701 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3702 p = bp->b_pages[index]; 3703 if (p && (index < bp->b_npages)) { 3704 if (p->busy) { 3705 printf( 3706 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3707 (intmax_t)bp->b_blkno, 3708 (intmax_t)bp->b_lblkno); 3709 } 3710 bp->b_pages[index] = NULL; 3711 pmap_qremove(pg, 1); 3712 vm_page_lock_queues(); 3713 vm_page_unwire(p, 0); 3714 vm_page_free(p); 3715 vm_page_unlock_queues(); 3716 } 3717 } 3718 VM_OBJECT_UNLOCK(kernel_object); 3719 bp->b_npages = newnpages; 3720} 3721 3722/* 3723 * Map an IO request into kernel virtual address space. 3724 * 3725 * All requests are (re)mapped into kernel VA space. 3726 * Notice that we use b_bufsize for the size of the buffer 3727 * to be mapped. b_bcount might be modified by the driver. 3728 * 3729 * Note that even if the caller determines that the address space should 3730 * be valid, a race or a smaller-file mapped into a larger space may 3731 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3732 * check the return value. 3733 */ 3734int 3735vmapbuf(struct buf *bp) 3736{ 3737 caddr_t addr, kva; 3738 vm_prot_t prot; 3739 int pidx, i; 3740 struct vm_page *m; 3741 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3742 3743 if (bp->b_bufsize < 0) 3744 return (-1); 3745 prot = VM_PROT_READ; 3746 if (bp->b_iocmd == BIO_READ) 3747 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3748 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3749 addr < bp->b_data + bp->b_bufsize; 3750 addr += PAGE_SIZE, pidx++) { 3751 /* 3752 * Do the vm_fault if needed; do the copy-on-write thing 3753 * when reading stuff off device into memory. 3754 * 3755 * NOTE! Must use pmap_extract() because addr may be in 3756 * the userland address space, and kextract is only guarenteed 3757 * to work for the kernland address space (see: sparc64 port). 3758 */ 3759retry: 3760 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3761 prot) < 0) { 3762 vm_page_lock_queues(); 3763 for (i = 0; i < pidx; ++i) { 3764 vm_page_unhold(bp->b_pages[i]); 3765 bp->b_pages[i] = NULL; 3766 } 3767 vm_page_unlock_queues(); 3768 return(-1); 3769 } 3770 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3771 if (m == NULL) 3772 goto retry; 3773 bp->b_pages[pidx] = m; 3774 } 3775 if (pidx > btoc(MAXPHYS)) 3776 panic("vmapbuf: mapped more than MAXPHYS"); 3777 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3778 3779 kva = bp->b_saveaddr; 3780 bp->b_npages = pidx; 3781 bp->b_saveaddr = bp->b_data; 3782 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3783 return(0); 3784} 3785 3786/* 3787 * Free the io map PTEs associated with this IO operation. 3788 * We also invalidate the TLB entries and restore the original b_addr. 3789 */ 3790void 3791vunmapbuf(struct buf *bp) 3792{ 3793 int pidx; 3794 int npages; 3795 3796 npages = bp->b_npages; 3797 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3798 vm_page_lock_queues(); 3799 for (pidx = 0; pidx < npages; pidx++) 3800 vm_page_unhold(bp->b_pages[pidx]); 3801 vm_page_unlock_queues(); 3802 3803 bp->b_data = bp->b_saveaddr; 3804} 3805 3806void 3807bdone(struct buf *bp) 3808{ 3809 3810 mtx_lock(&bdonelock); 3811 bp->b_flags |= B_DONE; 3812 wakeup(bp); 3813 mtx_unlock(&bdonelock); 3814} 3815 3816void 3817bwait(struct buf *bp, u_char pri, const char *wchan) 3818{ 3819 3820 mtx_lock(&bdonelock); 3821 while ((bp->b_flags & B_DONE) == 0) 3822 msleep(bp, &bdonelock, pri, wchan, 0); 3823 mtx_unlock(&bdonelock); 3824} 3825 3826int 3827bufsync(struct bufobj *bo, int waitfor, struct thread *td) 3828{ 3829 3830 return (VOP_FSYNC(bo->__bo_vnode, waitfor, td)); 3831} 3832 3833void 3834bufstrategy(struct bufobj *bo, struct buf *bp) 3835{ 3836 int i = 0; 3837 struct vnode *vp; 3838 3839 vp = bp->b_vp; 3840 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3841 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3842 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3843 i = VOP_STRATEGY(vp, bp); 3844 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3845} 3846 3847void 3848bufobj_wref(struct bufobj *bo) 3849{ 3850 3851 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3852 BO_LOCK(bo); 3853 bo->bo_numoutput++; 3854 BO_UNLOCK(bo); 3855} 3856 3857void 3858bufobj_wdrop(struct bufobj *bo) 3859{ 3860 3861 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3862 BO_LOCK(bo); 3863 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3864 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3865 bo->bo_flag &= ~BO_WWAIT; 3866 wakeup(&bo->bo_numoutput); 3867 } 3868 BO_UNLOCK(bo); 3869} 3870 3871int 3872bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3873{ 3874 int error; 3875 3876 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3877 ASSERT_BO_LOCKED(bo); 3878 error = 0; 3879 while (bo->bo_numoutput) { 3880 bo->bo_flag |= BO_WWAIT; 3881 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3882 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3883 if (error) 3884 break; 3885 } 3886 return (error); 3887} 3888 3889#include "opt_ddb.h" 3890#ifdef DDB 3891#include <ddb/ddb.h> 3892 3893/* DDB command to show buffer data */ 3894DB_SHOW_COMMAND(buffer, db_show_buffer) 3895{ 3896 /* get args */ 3897 struct buf *bp = (struct buf *)addr; 3898 3899 if (!have_addr) { 3900 db_printf("usage: show buffer <addr>\n"); 3901 return; 3902 } 3903 3904 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3905 db_printf( 3906 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3907 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", 3908 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3909 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); 3910 if (bp->b_npages) { 3911 int i; 3912 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3913 for (i = 0; i < bp->b_npages; i++) { 3914 vm_page_t m; 3915 m = bp->b_pages[i]; 3916 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3917 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3918 if ((i + 1) < bp->b_npages) 3919 db_printf(","); 3920 } 3921 db_printf("\n"); 3922 } 3923} 3924#endif /* DDB */ 3925