vfs_bio.c revision 222220
1156321Sdamien/*- 2156321Sdamien * Copyright (c) 2004 Poul-Henning Kamp 3156321Sdamien * Copyright (c) 1994,1997 John S. Dyson 4156321Sdamien * All rights reserved. 5156321Sdamien * 6156321Sdamien * Redistribution and use in source and binary forms, with or without 7156321Sdamien * modification, are permitted provided that the following conditions 8156321Sdamien * are met: 9156321Sdamien * 1. Redistributions of source code must retain the above copyright 10156321Sdamien * notice, this list of conditions and the following disclaimer. 11156321Sdamien * 2. Redistributions in binary form must reproduce the above copyright 12156321Sdamien * notice, this list of conditions and the following disclaimer in the 13156321Sdamien * documentation and/or other materials provided with the distribution. 14156321Sdamien * 15156321Sdamien * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16156321Sdamien * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17156321Sdamien * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18156321Sdamien * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19156321Sdamien * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20156321Sdamien * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21156321Sdamien * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22156321Sdamien * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23156321Sdamien * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24156321Sdamien * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25156321Sdamien * SUCH DAMAGE. 26156321Sdamien */ 27156321Sdamien 28156321Sdamien/* 29156321Sdamien * this file contains a new buffer I/O scheme implementing a coherent 30156321Sdamien * VM object and buffer cache scheme. Pains have been taken to make 31156321Sdamien * sure that the performance degradation associated with schemes such 32156321Sdamien * as this is not realized. 33156321Sdamien * 34156321Sdamien * Author: John S. Dyson 35156321Sdamien * Significant help during the development and debugging phases 36164982Skevlo * had been provided by David Greenman, also of the FreeBSD core team. 37164982Skevlo * 38156321Sdamien * see man buf(9) for more info. 39156321Sdamien */ 40156321Sdamien 41178354Ssam#include <sys/cdefs.h> 42156321Sdamien__FBSDID("$FreeBSD: head/sys/kern/vfs_bio.c 222220 2011-05-23 19:59:01Z ru $"); 43156321Sdamien 44156321Sdamien#include <sys/param.h> 45156321Sdamien#include <sys/systm.h> 46156321Sdamien#include <sys/bio.h> 47156321Sdamien#include <sys/conf.h> 48156321Sdamien#include <sys/buf.h> 49156321Sdamien#include <sys/devicestat.h> 50156321Sdamien#include <sys/eventhandler.h> 51156321Sdamien#include <sys/fail.h> 52156321Sdamien#include <sys/limits.h> 53156321Sdamien#include <sys/lock.h> 54156321Sdamien#include <sys/malloc.h> 55156321Sdamien#include <sys/mount.h> 56156321Sdamien#include <sys/mutex.h> 57170530Ssam#include <sys/kernel.h> 58206358Srpaulo#include <sys/kthread.h> 59156321Sdamien#include <sys/proc.h> 60156321Sdamien#include <sys/resourcevar.h> 61156321Sdamien#include <sys/sysctl.h> 62156321Sdamien#include <sys/vmmeter.h> 63156321Sdamien#include <sys/vnode.h> 64156321Sdamien#include <geom/geom.h> 65156321Sdamien#include <vm/vm.h> 66156327Ssilby#include <vm/vm_param.h> 67156327Ssilby#include <vm/vm_kern.h> 68156321Sdamien#include <vm/vm_pageout.h> 69178354Ssam#include <vm/vm_page.h> 70156321Sdamien#include <vm/vm_object.h> 71178354Ssam#include <vm/vm_extern.h> 72178354Ssam#include <vm/vm_map.h> 73178354Ssam#include "opt_compat.h" 74178354Ssam#include "opt_directio.h" 75178354Ssam#include "opt_swap.h" 76178354Ssam 77178354Ssamstatic MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 78178354Ssam 79156321Sdamienstruct bio_ops bioops; /* I/O operation notification */ 80178354Ssam 81178354Ssamstruct buf_ops buf_ops_bio = { 82156321Sdamien .bop_name = "buf_ops_bio", 83156321Sdamien .bop_write = bufwrite, 84178354Ssam .bop_strategy = bufstrategy, 85228621Sbschmidt .bop_sync = bufsync, 86228621Sbschmidt .bop_bdflush = bufbdflush, 87228621Sbschmidt}; 88178354Ssam 89156321Sdamien/* 90156321Sdamien * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 91156321Sdamien * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 92156321Sdamien */ 93156321Sdamienstruct buf *buf; /* buffer header pool */ 94156321Sdamien 95156321Sdamienstatic struct proc *bufdaemonproc; 96156321Sdamien 97156321Sdamienstatic int inmem(struct vnode *vp, daddr_t blkno); 98156321Sdamienstatic void vm_hold_free_pages(struct buf *bp, int newbsize); 99156321Sdamienstatic void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 100156321Sdamien vm_offset_t to); 101156321Sdamienstatic void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 102156321Sdamienstatic void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 103178354Ssam vm_page_t m); 104156321Sdamienstatic void vfs_drain_busy_pages(struct buf *bp); 105156321Sdamienstatic void vfs_clean_pages_dirty_buf(struct buf *bp); 106156321Sdamienstatic void vfs_setdirty_locked_object(struct buf *bp); 107156321Sdamienstatic void vfs_vmio_release(struct buf *bp); 108156321Sdamienstatic int vfs_bio_clcheck(struct vnode *vp, int size, 109156321Sdamien daddr_t lblkno, daddr_t blkno); 110156321Sdamienstatic int buf_do_flush(struct vnode *vp); 111156321Sdamienstatic int flushbufqueues(struct vnode *, int, int); 112156321Sdamienstatic void buf_daemon(void); 113170530Ssamstatic void bremfreel(struct buf *bp); 114170530Ssam#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 115170530Ssam defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 116156321Sdamienstatic int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 117156321Sdamien#endif 118156321Sdamien 119156321Sdamienint vmiodirenable = TRUE; 120156321SdamienSYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 121156321Sdamien "Use the VM system for directory writes"); 122156321Sdamienlong runningbufspace; 123178354SsamSYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 124156321Sdamien "Amount of presently outstanding async buffer io"); 125178354Ssamstatic long bufspace; 126178354Ssam#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 127165352Sbms defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 128156321SdamienSYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 129156321Sdamien &bufspace, 0, sysctl_bufspace, "L", "Virtual memory used for buffers"); 130156321Sdamien#else 131156321SdamienSYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 132156321Sdamien "Virtual memory used for buffers"); 133156321Sdamien#endif 134156321Sdamienstatic long maxbufspace; 135156321SdamienSYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 136156321Sdamien "Maximum allowed value of bufspace (including buf_daemon)"); 137156321Sdamienstatic long bufmallocspace; 138156321SdamienSYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 139156321Sdamien "Amount of malloced memory for buffers"); 140156321Sdamienstatic long maxbufmallocspace; 141156321SdamienSYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 142156321Sdamien "Maximum amount of malloced memory for buffers"); 143156321Sdamienstatic long lobufspace; 144156321SdamienSYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 145156321Sdamien "Minimum amount of buffers we want to have"); 146156321Sdamienlong hibufspace; 147156321SdamienSYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 148156321Sdamien "Maximum allowed value of bufspace (excluding buf_daemon)"); 149178354Ssamstatic int bufreusecnt; 150156321SdamienSYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 151156321Sdamien "Number of times we have reused a buffer"); 152156321Sdamienstatic int buffreekvacnt; 153178354SsamSYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 154190526Ssam "Number of times we have freed the KVA space from some buffer"); 155156321Sdamienstatic int bufdefragcnt; 156178354SsamSYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 157156321Sdamien "Number of times we have had to repeat buffer allocation to defragment"); 158178354Ssamstatic long lorunningspace; 159156321SdamienSYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 160178354Ssam "Minimum preferred space used for in-progress I/O"); 161156321Sdamienstatic long hirunningspace; 162156321SdamienSYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 163156321Sdamien "Maximum amount of space to use for in-progress I/O"); 164156321Sdamienint dirtybufferflushes; 165156321SdamienSYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 166178354Ssam 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 167178354Ssamint bdwriteskip; 168156321SdamienSYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 169192468Ssam 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 170156321Sdamienint altbufferflushes; 171156321SdamienSYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 172156321Sdamien 0, "Number of fsync flushes to limit dirty buffers"); 173156321Sdamienstatic int recursiveflushes; 174156321SdamienSYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 175156321Sdamien 0, "Number of flushes skipped due to being recursive"); 176156321Sdamienstatic int numdirtybuffers; 177156321SdamienSYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 178156321Sdamien "Number of buffers that are dirty (has unwritten changes) at the moment"); 179156321Sdamienstatic int lodirtybuffers; 180156321SdamienSYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 181156321Sdamien "How many buffers we want to have free before bufdaemon can sleep"); 182156321Sdamienstatic int hidirtybuffers; 183156321SdamienSYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 184156321Sdamien "When the number of dirty buffers is considered severe"); 185156321Sdamienint dirtybufthresh; 186156321SdamienSYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 187156321Sdamien 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 188156321Sdamienstatic int numfreebuffers; 189156321SdamienSYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 190156321Sdamien "Number of free buffers"); 191156321Sdamienstatic int lofreebuffers; 192156321SdamienSYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 193156321Sdamien "XXX Unused"); 194156321Sdamienstatic int hifreebuffers; 195156321SdamienSYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 196156321Sdamien "XXX Complicatedly unused"); 197156321Sdamienstatic int getnewbufcalls; 198156321SdamienSYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 199178354Ssam "Number of calls to getnewbuf"); 200156321Sdamienstatic int getnewbufrestarts; 201156321SdamienSYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 202178354Ssam "Number of times getnewbuf has had to restart a buffer aquisition"); 203178354Ssamstatic int flushbufqtarget = 100; 204190526SsamSYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 205156321Sdamien "Amount of work to do in flushbufqueues when helping bufdaemon"); 206178354Ssamstatic long notbufdflashes; 207156321SdamienSYSCTL_LONG(_vfs, OID_AUTO, notbufdflashes, CTLFLAG_RD, ¬bufdflashes, 0, 208156321Sdamien "Number of dirty buffer flushes done by the bufdaemon helpers"); 209178354Ssam 210178354Ssam/* 211178354Ssam * Wakeup point for bufdaemon, as well as indicator of whether it is already 212178354Ssam * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 213178354Ssam * is idling. 214178354Ssam */ 215178354Ssamstatic int bd_request; 216156321Sdamien 217156321Sdamien/* 218156321Sdamien * Request for the buf daemon to write more buffers than is indicated by 219165352Sbms * lodirtybuf. This may be necessary to push out excess dependencies or 220156321Sdamien * defragment the address space where a simple count of the number of dirty 221156321Sdamien * buffers is insufficient to characterize the demand for flushing them. 222156321Sdamien */ 223156321Sdamienstatic int bd_speedupreq; 224156321Sdamien 225156321Sdamien/* 226156321Sdamien * This lock synchronizes access to bd_request. 227156321Sdamien */ 228156321Sdamienstatic struct mtx bdlock; 229156321Sdamien 230156321Sdamien/* 231156321Sdamien * bogus page -- for I/O to/from partially complete buffers 232156321Sdamien * this is a temporary solution to the problem, but it is not 233156321Sdamien * really that bad. it would be better to split the buffer 234156321Sdamien * for input in the case of buffers partially already in memory, 235190526Ssam * but the code is intricate enough already. 236156321Sdamien */ 237156321Sdamienvm_page_t bogus_page; 238156321Sdamien 239156321Sdamien/* 240156321Sdamien * Synchronization (sleep/wakeup) variable for active buffer space requests. 241156321Sdamien * Set when wait starts, cleared prior to wakeup(). 242156321Sdamien * Used in runningbufwakeup() and waitrunningbufspace(). 243156321Sdamien */ 244156321Sdamienstatic int runningbufreq; 245156321Sdamien 246156321Sdamien/* 247156321Sdamien * This lock protects the runningbufreq and synchronizes runningbufwakeup and 248156321Sdamien * waitrunningbufspace(). 249156321Sdamien */ 250156321Sdamienstatic struct mtx rbreqlock; 251156321Sdamien 252156321Sdamien/* 253156321Sdamien * Synchronization (sleep/wakeup) variable for buffer requests. 254156321Sdamien * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 255156321Sdamien * by and/or. 256156321Sdamien * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 257156321Sdamien * getnewbuf(), and getblk(). 258156321Sdamien */ 259156321Sdamienstatic int needsbuffer; 260156321Sdamien 261156321Sdamien/* 262156321Sdamien * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 263156321Sdamien */ 264156321Sdamienstatic struct mtx nblock; 265156321Sdamien 266156321Sdamien/* 267156321Sdamien * Definitions for the buffer free lists. 268156321Sdamien */ 269156321Sdamien#define BUFFER_QUEUES 6 /* number of free buffer queues */ 270156321Sdamien 271207554Ssobomax#define QUEUE_NONE 0 /* on no queue */ 272207554Ssobomax#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 273156321Sdamien#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 274156321Sdamien#define QUEUE_DIRTY_GIANT 3 /* B_DELWRI buffers that need giant */ 275156321Sdamien#define QUEUE_EMPTYKVA 4 /* empty buffer headers w/KVA assignment */ 276178354Ssam#define QUEUE_EMPTY 5 /* empty buffer headers */ 277156321Sdamien#define QUEUE_SENTINEL 1024 /* not an queue index, but mark for sentinel */ 278156321Sdamien 279156321Sdamien/* Queues for free buffers with various properties */ 280156321Sdamienstatic TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 281178957Ssam 282178957Ssam/* Lock for the bufqueues */ 283178354Ssamstatic struct mtx bqlock; 284178354Ssam 285178354Ssam/* 286178354Ssam * Single global constant for BUF_WMESG, to avoid getting multiple references. 287195618Srpaulo * buf_wmesg is referred from macros. 288178354Ssam */ 289178354Ssamconst char *buf_wmesg = BUF_WMESG; 290178354Ssam 291178354Ssam#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 292156407Sdamien#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 293178354Ssam#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 294178354Ssam#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 295156407Sdamien 296178354Ssam#if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 297156321Sdamien defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 298170530Ssamstatic int 299170530Ssamsysctl_bufspace(SYSCTL_HANDLER_ARGS) 300170530Ssam{ 301170530Ssam long lvalue; 302170530Ssam int ivalue; 303178354Ssam 304156321Sdamien if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 305190526Ssam return (sysctl_handle_long(oidp, arg1, arg2, req)); 306178354Ssam lvalue = *(long *)arg1; 307178354Ssam if (lvalue > INT_MAX) 308178354Ssam /* On overflow, still write out a long to trigger ENOMEM. */ 309170530Ssam return (sysctl_handle_long(oidp, &lvalue, 0, req)); 310170530Ssam ivalue = lvalue; 311170530Ssam return (sysctl_handle_int(oidp, &ivalue, 0, req)); 312156321Sdamien} 313178354Ssam#endif 314178354Ssam 315156321Sdamien#ifdef DIRECTIO 316178354Ssamextern void ffs_rawread_setup(void); 317178354Ssam#endif /* DIRECTIO */ 318156321Sdamien/* 319192468Ssam * numdirtywakeup: 320192468Ssam * 321192468Ssam * If someone is blocked due to there being too many dirty buffers, 322192468Ssam * and numdirtybuffers is now reasonable, wake them up. 323192468Ssam */ 324178354Ssam 325178354Ssamstatic __inline void 326156321Sdamiennumdirtywakeup(int level) 327178354Ssam{ 328178354Ssam 329178354Ssam if (numdirtybuffers <= level) { 330156321Sdamien mtx_lock(&nblock); 331156321Sdamien if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 332156321Sdamien needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 333156321Sdamien wakeup(&needsbuffer); 334156321Sdamien } 335156321Sdamien mtx_unlock(&nblock); 336156321Sdamien } 337156321Sdamien} 338156321Sdamien 339178354Ssam/* 340156321Sdamien * bufspacewakeup: 341156321Sdamien * 342156321Sdamien * Called when buffer space is potentially available for recovery. 343156321Sdamien * getnewbuf() will block on this flag when it is unable to free 344156321Sdamien * sufficient buffer space. Buffer space becomes recoverable when 345156321Sdamien * bp's get placed back in the queues. 346156321Sdamien */ 347178354Ssam 348178354Ssamstatic __inline void 349170530Ssambufspacewakeup(void) 350178038Ssam{ 351178038Ssam 352178038Ssam /* 353156321Sdamien * If someone is waiting for BUF space, wake them up. Even 354156321Sdamien * though we haven't freed the kva space yet, the waiting 355156321Sdamien * process will be able to now. 356156321Sdamien */ 357156321Sdamien mtx_lock(&nblock); 358156321Sdamien if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 359156321Sdamien needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 360156321Sdamien wakeup(&needsbuffer); 361156321Sdamien } 362156321Sdamien mtx_unlock(&nblock); 363156321Sdamien} 364156321Sdamien 365156321Sdamien/* 366156321Sdamien * runningbufwakeup() - in-progress I/O accounting. 367156321Sdamien * 368156321Sdamien */ 369156321Sdamienvoid 370178354Ssamrunningbufwakeup(struct buf *bp) 371228621Sbschmidt{ 372228621Sbschmidt 373228621Sbschmidt if (bp->b_runningbufspace) { 374228621Sbschmidt atomic_subtract_long(&runningbufspace, bp->b_runningbufspace); 375178354Ssam bp->b_runningbufspace = 0; 376178354Ssam mtx_lock(&rbreqlock); 377178354Ssam if (runningbufreq && runningbufspace <= lorunningspace) { 378178354Ssam runningbufreq = 0; 379178354Ssam wakeup(&runningbufreq); 380178354Ssam } 381178354Ssam mtx_unlock(&rbreqlock); 382178354Ssam } 383178354Ssam} 384178354Ssam 385178354Ssam/* 386195618Srpaulo * bufcountwakeup: 387195618Srpaulo * 388178354Ssam * Called when a buffer has been added to one of the free queues to 389178354Ssam * account for the buffer and to wakeup anyone waiting for free buffers. 390178354Ssam * This typically occurs when large amounts of metadata are being handled 391178354Ssam * by the buffer cache ( else buffer space runs out first, usually ). 392178354Ssam */ 393178354Ssam 394178354Ssamstatic __inline void 395178354Ssambufcountwakeup(struct buf *bp) 396178354Ssam{ 397178354Ssam int old; 398178354Ssam 399178354Ssam KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 400178354Ssam ("buf %p already counted as free", bp)); 401178354Ssam if (bp->b_bufobj != NULL) 402178354Ssam mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED); 403178354Ssam bp->b_vflags |= BV_INFREECNT; 404178354Ssam old = atomic_fetchadd_int(&numfreebuffers, 1); 405178354Ssam KASSERT(old >= 0 && old < nbuf, 406178354Ssam ("numfreebuffers climbed to %d", old + 1)); 407178354Ssam mtx_lock(&nblock); 408178354Ssam if (needsbuffer) { 409178354Ssam needsbuffer &= ~VFS_BIO_NEED_ANY; 410178354Ssam if (numfreebuffers >= hifreebuffers) 411178354Ssam needsbuffer &= ~VFS_BIO_NEED_FREE; 412178354Ssam wakeup(&needsbuffer); 413178354Ssam } 414178354Ssam mtx_unlock(&nblock); 415178354Ssam} 416178354Ssam 417178354Ssam/* 418178354Ssam * waitrunningbufspace() 419178354Ssam * 420178354Ssam * runningbufspace is a measure of the amount of I/O currently 421178354Ssam * running. This routine is used in async-write situations to 422178354Ssam * prevent creating huge backups of pending writes to a device. 423178354Ssam * Only asynchronous writes are governed by this function. 424178354Ssam * 425178354Ssam * Reads will adjust runningbufspace, but will not block based on it. 426206358Srpaulo * The read load has a side effect of reducing the allowed write load. 427178354Ssam * 428178354Ssam * This does NOT turn an async write into a sync write. It waits 429178354Ssam * for earlier writes to complete and generally returns before the 430178354Ssam * caller's write has reached the device. 431178354Ssam */ 432178354Ssamvoid 433178354Ssamwaitrunningbufspace(void) 434178354Ssam{ 435178354Ssam 436178354Ssam mtx_lock(&rbreqlock); 437178354Ssam while (runningbufspace > hirunningspace) { 438178354Ssam ++runningbufreq; 439206358Srpaulo msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 440178354Ssam } 441178354Ssam mtx_unlock(&rbreqlock); 442178354Ssam} 443178354Ssam 444156321Sdamien 445156321Sdamien/* 446156321Sdamien * vfs_buf_test_cache: 447156321Sdamien * 448156321Sdamien * Called when a buffer is extended. This function clears the B_CACHE 449156321Sdamien * bit if the newly extended portion of the buffer does not contain 450156321Sdamien * valid data. 451156321Sdamien */ 452156321Sdamienstatic __inline 453156321Sdamienvoid 454156321Sdamienvfs_buf_test_cache(struct buf *bp, 455156321Sdamien vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 456156321Sdamien vm_page_t m) 457156321Sdamien{ 458156321Sdamien 459156321Sdamien VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 460156321Sdamien if (bp->b_flags & B_CACHE) { 461156321Sdamien int base = (foff + off) & PAGE_MASK; 462156321Sdamien if (vm_page_is_valid(m, base, size) == 0) 463156321Sdamien bp->b_flags &= ~B_CACHE; 464178354Ssam } 465156321Sdamien} 466178354Ssam 467178354Ssam/* Wake up the buffer daemon if necessary */ 468156321Sdamienstatic __inline 469156321Sdamienvoid 470156321Sdamienbd_wakeup(int dirtybuflevel) 471156321Sdamien{ 472156321Sdamien 473156321Sdamien mtx_lock(&bdlock); 474156321Sdamien if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 475156321Sdamien bd_request = 1; 476156321Sdamien wakeup(&bd_request); 477156321Sdamien } 478156321Sdamien mtx_unlock(&bdlock); 479156321Sdamien} 480156321Sdamien 481156321Sdamien/* 482156321Sdamien * bd_speedup - speedup the buffer cache flushing code 483156321Sdamien */ 484156321Sdamien 485156321Sdamienvoid 486156321Sdamienbd_speedup(void) 487156321Sdamien{ 488156321Sdamien int needwake; 489156321Sdamien 490156321Sdamien mtx_lock(&bdlock); 491171535Skevlo needwake = 0; 492171535Skevlo if (bd_speedupreq == 0 || bd_request == 0) 493171535Skevlo needwake = 1; 494171535Skevlo bd_speedupreq = 1; 495156321Sdamien bd_request = 1; 496156321Sdamien if (needwake) 497156321Sdamien wakeup(&bd_request); 498156321Sdamien mtx_unlock(&bdlock); 499156321Sdamien} 500156321Sdamien 501156321Sdamien/* 502156321Sdamien * Calculating buffer cache scaling values and reserve space for buffer 503156321Sdamien * headers. This is called during low level kernel initialization and 504156321Sdamien * may be called more then once. We CANNOT write to the memory area 505156321Sdamien * being reserved at this time. 506156321Sdamien */ 507156321Sdamiencaddr_t 508156321Sdamienkern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 509156321Sdamien{ 510156321Sdamien int tuned_nbuf; 511156321Sdamien long maxbuf; 512156321Sdamien 513156321Sdamien /* 514156321Sdamien * physmem_est is in pages. Convert it to kilobytes (assumes 515156321Sdamien * PAGE_SIZE is >= 1K) 516156321Sdamien */ 517156321Sdamien physmem_est = physmem_est * (PAGE_SIZE / 1024); 518156321Sdamien 519156321Sdamien /* 520156321Sdamien * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 521156321Sdamien * For the first 64MB of ram nominally allocate sufficient buffers to 522156321Sdamien * cover 1/4 of our ram. Beyond the first 64MB allocate additional 523171535Skevlo * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 524171535Skevlo * the buffer cache we limit the eventual kva reservation to 525171535Skevlo * maxbcache bytes. 526156321Sdamien * 527156321Sdamien * factor represents the 1/4 x ram conversion. 528156321Sdamien */ 529156321Sdamien if (nbuf == 0) { 530156321Sdamien int factor = 4 * BKVASIZE / 1024; 531156321Sdamien 532156321Sdamien nbuf = 50; 533156321Sdamien if (physmem_est > 4096) 534156321Sdamien nbuf += min((physmem_est - 4096) / factor, 535156321Sdamien 65536 / factor); 536156321Sdamien if (physmem_est > 65536) 537156321Sdamien nbuf += (physmem_est - 65536) * 2 / (factor * 5); 538156321Sdamien 539156321Sdamien if (maxbcache && nbuf > maxbcache / BKVASIZE) 540156321Sdamien nbuf = maxbcache / BKVASIZE; 541156321Sdamien tuned_nbuf = 1; 542156321Sdamien } else 543156321Sdamien tuned_nbuf = 0; 544156321Sdamien 545156321Sdamien /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 546156321Sdamien maxbuf = (LONG_MAX / 3) / BKVASIZE; 547156321Sdamien if (nbuf > maxbuf) { 548156321Sdamien if (!tuned_nbuf) 549156321Sdamien printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 550156321Sdamien maxbuf); 551156321Sdamien nbuf = maxbuf; 552156321Sdamien } 553156321Sdamien 554156321Sdamien /* 555156321Sdamien * swbufs are used as temporary holders for I/O, such as paging I/O. 556156321Sdamien * We have no less then 16 and no more then 256. 557156321Sdamien */ 558156321Sdamien nswbuf = max(min(nbuf/4, 256), 16); 559156321Sdamien#ifdef NSWBUF_MIN 560156321Sdamien if (nswbuf < NSWBUF_MIN) 561156321Sdamien nswbuf = NSWBUF_MIN; 562156321Sdamien#endif 563156321Sdamien#ifdef DIRECTIO 564156321Sdamien ffs_rawread_setup(); 565156321Sdamien#endif 566156321Sdamien 567156321Sdamien /* 568156321Sdamien * Reserve space for the buffer cache buffers 569156321Sdamien */ 570156321Sdamien swbuf = (void *)v; 571156321Sdamien v = (caddr_t)(swbuf + nswbuf); 572156321Sdamien buf = (void *)v; 573156321Sdamien v = (caddr_t)(buf + nbuf); 574156321Sdamien 575156321Sdamien return(v); 576156321Sdamien} 577156321Sdamien 578156321Sdamien/* Initialize the buffer subsystem. Called before use of any buffers. */ 579156321Sdamienvoid 580156321Sdamienbufinit(void) 581156321Sdamien{ 582156321Sdamien struct buf *bp; 583156321Sdamien int i; 584156321Sdamien 585156321Sdamien mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 586156321Sdamien mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 587156321Sdamien mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 588156321Sdamien mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 589156321Sdamien 590156321Sdamien /* next, make a null set of free lists */ 591156321Sdamien for (i = 0; i < BUFFER_QUEUES; i++) 592156321Sdamien TAILQ_INIT(&bufqueues[i]); 593156321Sdamien 594156321Sdamien /* finally, initialize each buffer header and stick on empty q */ 595156321Sdamien for (i = 0; i < nbuf; i++) { 596156321Sdamien bp = &buf[i]; 597156321Sdamien bzero(bp, sizeof *bp); 598156321Sdamien bp->b_flags = B_INVAL; /* we're just an empty header */ 599156321Sdamien bp->b_rcred = NOCRED; 600156321Sdamien bp->b_wcred = NOCRED; 601156321Sdamien bp->b_qindex = QUEUE_EMPTY; 602156321Sdamien bp->b_vflags = BV_INFREECNT; /* buf is counted as free */ 603156321Sdamien bp->b_xflags = 0; 604156321Sdamien LIST_INIT(&bp->b_dep); 605156321Sdamien BUF_LOCKINIT(bp); 606156321Sdamien TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 607156321Sdamien } 608156321Sdamien 609156321Sdamien /* 610156321Sdamien * maxbufspace is the absolute maximum amount of buffer space we are 611156321Sdamien * allowed to reserve in KVM and in real terms. The absolute maximum 612156321Sdamien * is nominally used by buf_daemon. hibufspace is the nominal maximum 613156321Sdamien * used by most other processes. The differential is required to 614156321Sdamien * ensure that buf_daemon is able to run when other processes might 615156321Sdamien * be blocked waiting for buffer space. 616156321Sdamien * 617156321Sdamien * maxbufspace is based on BKVASIZE. Allocating buffers larger then 618156321Sdamien * this may result in KVM fragmentation which is not handled optimally 619156321Sdamien * by the system. 620156321Sdamien */ 621156321Sdamien maxbufspace = (long)nbuf * BKVASIZE; 622156321Sdamien hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 623156321Sdamien lobufspace = hibufspace - MAXBSIZE; 624156321Sdamien 625156321Sdamien /* 626156321Sdamien * Note: The 16 MiB upper limit for hirunningspace was chosen 627156321Sdamien * arbitrarily and may need further tuning. It corresponds to 628156321Sdamien * 128 outstanding write IO requests (if IO size is 128 KiB), 629156321Sdamien * which fits with many RAID controllers' tagged queuing limits. 630156321Sdamien * The lower 1 MiB limit is the historical upper limit for 631156321Sdamien * hirunningspace. 632171535Skevlo */ 633171535Skevlo hirunningspace = lmax(lmin(roundup(hibufspace / 64, MAXBSIZE), 634171535Skevlo 16 * 1024 * 1024), 1024 * 1024); 635171535Skevlo lorunningspace = roundup((hirunningspace * 2) / 3, MAXBSIZE); 636156321Sdamien 637156321Sdamien/* 638156321Sdamien * Limit the amount of malloc memory since it is wired permanently into 639156321Sdamien * the kernel space. Even though this is accounted for in the buffer 640156321Sdamien * allocation, we don't want the malloced region to grow uncontrolled. 641156321Sdamien * The malloc scheme improves memory utilization significantly on average 642156321Sdamien * (small) directories. 643156321Sdamien */ 644156321Sdamien maxbufmallocspace = hibufspace / 20; 645156321Sdamien 646156321Sdamien/* 647156321Sdamien * Reduce the chance of a deadlock occuring by limiting the number 648156321Sdamien * of delayed-write dirty buffers we allow to stack up. 649156321Sdamien */ 650156321Sdamien hidirtybuffers = nbuf / 4 + 20; 651156321Sdamien dirtybufthresh = hidirtybuffers * 9 / 10; 652156321Sdamien numdirtybuffers = 0; 653156321Sdamien/* 654156321Sdamien * To support extreme low-memory systems, make sure hidirtybuffers cannot 655156321Sdamien * eat up all available buffer space. This occurs when our minimum cannot 656156321Sdamien * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 657156321Sdamien * BKVASIZE'd buffers. 658156321Sdamien */ 659156321Sdamien while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 660156321Sdamien hidirtybuffers >>= 1; 661156321Sdamien } 662156321Sdamien lodirtybuffers = hidirtybuffers / 2; 663156321Sdamien 664156321Sdamien/* 665156321Sdamien * Try to keep the number of free buffers in the specified range, 666156321Sdamien * and give special processes (e.g. like buf_daemon) access to an 667171535Skevlo * emergency reserve. 668171535Skevlo */ 669171535Skevlo lofreebuffers = nbuf / 18 + 5; 670156321Sdamien hifreebuffers = 2 * lofreebuffers; 671156321Sdamien numfreebuffers = nbuf; 672156321Sdamien 673156321Sdamien bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 674156321Sdamien VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 675156321Sdamien} 676156321Sdamien 677156321Sdamien/* 678156321Sdamien * bfreekva() - free the kva allocation for a buffer. 679156321Sdamien * 680156321Sdamien * Since this call frees up buffer space, we call bufspacewakeup(). 681156321Sdamien */ 682156321Sdamienstatic void 683156321Sdamienbfreekva(struct buf *bp) 684156321Sdamien{ 685243857Sglebius 686156321Sdamien if (bp->b_kvasize) { 687156321Sdamien atomic_add_int(&buffreekvacnt, 1); 688156321Sdamien atomic_subtract_long(&bufspace, bp->b_kvasize); 689156321Sdamien vm_map_remove(buffer_map, (vm_offset_t) bp->b_kvabase, 690156321Sdamien (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 691156321Sdamien bp->b_kvasize = 0; 692156321Sdamien bufspacewakeup(); 693156321Sdamien } 694156321Sdamien} 695156321Sdamien 696156321Sdamien/* 697156321Sdamien * bremfree: 698156321Sdamien * 699156321Sdamien * Mark the buffer for removal from the appropriate free list in brelse. 700156321Sdamien * 701156321Sdamien */ 702156321Sdamienvoid 703156321Sdamienbremfree(struct buf *bp) 704156321Sdamien{ 705156321Sdamien int old; 706156321Sdamien 707156321Sdamien CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 708156321Sdamien KASSERT((bp->b_flags & B_REMFREE) == 0, 709156321Sdamien ("bremfree: buffer %p already marked for delayed removal.", bp)); 710156321Sdamien KASSERT(bp->b_qindex != QUEUE_NONE, 711156321Sdamien ("bremfree: buffer %p not on a queue.", bp)); 712156321Sdamien BUF_ASSERT_HELD(bp); 713156321Sdamien 714156321Sdamien bp->b_flags |= B_REMFREE; 715156321Sdamien /* Fixup numfreebuffers count. */ 716156321Sdamien if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 717156321Sdamien KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 718156321Sdamien ("buf %p not counted in numfreebuffers", bp)); 719156321Sdamien if (bp->b_bufobj != NULL) 720156321Sdamien mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED); 721156321Sdamien bp->b_vflags &= ~BV_INFREECNT; 722156321Sdamien old = atomic_fetchadd_int(&numfreebuffers, -1); 723156321Sdamien KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 724156321Sdamien } 725156321Sdamien} 726156321Sdamien 727156321Sdamien/* 728156321Sdamien * bremfreef: 729156321Sdamien * 730156321Sdamien * Force an immediate removal from a free list. Used only in nfs when 731156321Sdamien * it abuses the b_freelist pointer. 732156321Sdamien */ 733156321Sdamienvoid 734156321Sdamienbremfreef(struct buf *bp) 735156321Sdamien{ 736156321Sdamien mtx_lock(&bqlock); 737156321Sdamien bremfreel(bp); 738156321Sdamien mtx_unlock(&bqlock); 739156321Sdamien} 740156321Sdamien 741156321Sdamien/* 742156321Sdamien * bremfreel: 743156321Sdamien * 744156321Sdamien * Removes a buffer from the free list, must be called with the 745156321Sdamien * bqlock held. 746156321Sdamien */ 747156321Sdamienstatic void 748156321Sdamienbremfreel(struct buf *bp) 749156321Sdamien{ 750156321Sdamien int old; 751156321Sdamien 752156321Sdamien CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 753156321Sdamien bp, bp->b_vp, bp->b_flags); 754156321Sdamien KASSERT(bp->b_qindex != QUEUE_NONE, 755156321Sdamien ("bremfreel: buffer %p not on a queue.", bp)); 756156321Sdamien BUF_ASSERT_HELD(bp); 757156321Sdamien mtx_assert(&bqlock, MA_OWNED); 758156321Sdamien 759156321Sdamien TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 760156321Sdamien bp->b_qindex = QUEUE_NONE; 761156321Sdamien /* 762156321Sdamien * If this was a delayed bremfree() we only need to remove the buffer 763156321Sdamien * from the queue and return the stats are already done. 764156321Sdamien */ 765156321Sdamien if (bp->b_flags & B_REMFREE) { 766178354Ssam bp->b_flags &= ~B_REMFREE; 767156321Sdamien return; 768178354Ssam } 769178354Ssam /* 770156321Sdamien * Fixup numfreebuffers count. If the buffer is invalid or not 771178354Ssam * delayed-write, the buffer was free and we must decrement 772156321Sdamien * numfreebuffers. 773178354Ssam */ 774178354Ssam if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) { 775156321Sdamien KASSERT((bp->b_vflags & BV_INFREECNT) != 0, 776178354Ssam ("buf %p not counted in numfreebuffers", bp)); 777178354Ssam if (bp->b_bufobj != NULL) 778178354Ssam mtx_assert(BO_MTX(bp->b_bufobj), MA_OWNED); 779178354Ssam bp->b_vflags &= ~BV_INFREECNT; 780156321Sdamien old = atomic_fetchadd_int(&numfreebuffers, -1); 781178354Ssam KASSERT(old > 0, ("numfreebuffers dropped to %d", old - 1)); 782156321Sdamien } 783178354Ssam} 784178354Ssam 785178354Ssam 786178354Ssam/* 787156321Sdamien * Get a buffer with the specified data. Look in the cache first. We 788156321Sdamien * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 789156321Sdamien * is set, the buffer is valid and we do not have to do anything ( see 790156321Sdamien * getblk() ). This is really just a special case of breadn(). 791156321Sdamien */ 792156321Sdamienint 793178354Ssambread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 794195618Srpaulo struct buf **bpp) 795195618Srpaulo{ 796178354Ssam 797178354Ssam return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 798178354Ssam} 799156321Sdamien 800184345Ssam/* 801156321Sdamien * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 802192468Ssam * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 803192468Ssam * the buffer is valid and we do not have to do anything. 804178354Ssam */ 805178354Ssamvoid 806156321Sdamienbreada(struct vnode * vp, daddr_t * rablkno, int * rabsize, 807156321Sdamien int cnt, struct ucred * cred) 808156321Sdamien{ 809156321Sdamien struct buf *rabp; 810156321Sdamien int i; 811156321Sdamien 812156321Sdamien for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 813156321Sdamien if (inmem(vp, *rablkno)) 814156321Sdamien continue; 815156321Sdamien rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 816156321Sdamien 817156321Sdamien if ((rabp->b_flags & B_CACHE) == 0) { 818156321Sdamien if (!TD_IS_IDLETHREAD(curthread)) 819156321Sdamien curthread->td_ru.ru_inblock++; 820156321Sdamien rabp->b_flags |= B_ASYNC; 821156321Sdamien rabp->b_flags &= ~B_INVAL; 822156321Sdamien rabp->b_ioflags &= ~BIO_ERROR; 823156321Sdamien rabp->b_iocmd = BIO_READ; 824156321Sdamien if (rabp->b_rcred == NOCRED && cred != NOCRED) 825156321Sdamien rabp->b_rcred = crhold(cred); 826156321Sdamien vfs_busy_pages(rabp, 0); 827156321Sdamien BUF_KERNPROC(rabp); 828156321Sdamien rabp->b_iooffset = dbtob(rabp->b_blkno); 829156321Sdamien bstrategy(rabp); 830156321Sdamien } else { 831156321Sdamien brelse(rabp); 832156321Sdamien } 833156321Sdamien } 834156321Sdamien} 835156321Sdamien 836156321Sdamien/* 837156321Sdamien * Operates like bread, but also starts asynchronous I/O on 838156321Sdamien * read-ahead blocks. 839156321Sdamien */ 840156321Sdamienint 841156321Sdamienbreadn(struct vnode * vp, daddr_t blkno, int size, 842156321Sdamien daddr_t * rablkno, int *rabsize, 843156321Sdamien int cnt, struct ucred * cred, struct buf **bpp) 844156321Sdamien{ 845156321Sdamien struct buf *bp; 846156321Sdamien int rv = 0, readwait = 0; 847156321Sdamien 848156321Sdamien CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 849156321Sdamien *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 850156321Sdamien 851156321Sdamien /* if not found in cache, do some I/O */ 852156321Sdamien if ((bp->b_flags & B_CACHE) == 0) { 853156321Sdamien if (!TD_IS_IDLETHREAD(curthread)) 854156321Sdamien curthread->td_ru.ru_inblock++; 855156321Sdamien bp->b_iocmd = BIO_READ; 856156321Sdamien bp->b_flags &= ~B_INVAL; 857156321Sdamien bp->b_ioflags &= ~BIO_ERROR; 858156321Sdamien if (bp->b_rcred == NOCRED && cred != NOCRED) 859156321Sdamien bp->b_rcred = crhold(cred); 860156321Sdamien vfs_busy_pages(bp, 0); 861156321Sdamien bp->b_iooffset = dbtob(bp->b_blkno); 862156321Sdamien bstrategy(bp); 863156321Sdamien ++readwait; 864156321Sdamien } 865156321Sdamien 866156321Sdamien breada(vp, rablkno, rabsize, cnt, cred); 867156321Sdamien 868156321Sdamien if (readwait) { 869178354Ssam rv = bufwait(bp); 870156321Sdamien } 871156321Sdamien return (rv); 872156321Sdamien} 873156321Sdamien 874206358Srpaulo/* 875156321Sdamien * Write, release buffer on completion. (Done by iodone 876156321Sdamien * if async). Do not bother writing anything if the buffer 877170530Ssam * is invalid. 878170530Ssam * 879170530Ssam * Note that we set B_CACHE here, indicating that buffer is 880156321Sdamien * fully valid and thus cacheable. This is true even of NFS 881156321Sdamien * now so we set it generally. This could be set either here 882156321Sdamien * or in biodone() since the I/O is synchronous. We put it 883156321Sdamien * here. 884156321Sdamien */ 885156321Sdamienint 886156321Sdamienbufwrite(struct buf *bp) 887156321Sdamien{ 888156321Sdamien int oldflags; 889156321Sdamien struct vnode *vp; 890170530Ssam int vp_md; 891170530Ssam 892170530Ssam CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 893170530Ssam if (bp->b_flags & B_INVAL) { 894156321Sdamien brelse(bp); 895159301Sfjoe return (0); 896170530Ssam } 897159301Sfjoe 898206371Srpaulo oldflags = bp->b_flags; 899206371Srpaulo 900159301Sfjoe BUF_ASSERT_HELD(bp); 901156321Sdamien 902156321Sdamien if (bp->b_pin_count > 0) 903156321Sdamien bunpin_wait(bp); 904156321Sdamien 905178354Ssam KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 906178354Ssam ("FFS background buffer should not get here %p", bp)); 907178354Ssam 908206358Srpaulo vp = bp->b_vp; 909206358Srpaulo if (vp) 910206358Srpaulo vp_md = vp->v_vflag & VV_MD; 911156321Sdamien else 912156321Sdamien vp_md = 0; 913156321Sdamien 914156321Sdamien /* Mark the buffer clean */ 915178354Ssam bundirty(bp); 916178354Ssam 917178354Ssam bp->b_flags &= ~B_DONE; 918178354Ssam bp->b_ioflags &= ~BIO_ERROR; 919178354Ssam bp->b_flags |= B_CACHE; 920206358Srpaulo bp->b_iocmd = BIO_WRITE; 921206358Srpaulo 922206358Srpaulo bufobj_wref(bp->b_bufobj); 923156321Sdamien vfs_busy_pages(bp, 1); 924156321Sdamien 925156321Sdamien /* 926156321Sdamien * Normal bwrites pipeline writes 927156321Sdamien */ 928156321Sdamien bp->b_runningbufspace = bp->b_bufsize; 929156321Sdamien atomic_add_long(&runningbufspace, bp->b_runningbufspace); 930156321Sdamien 931156321Sdamien if (!TD_IS_IDLETHREAD(curthread)) 932156321Sdamien curthread->td_ru.ru_oublock++; 933178354Ssam if (oldflags & B_ASYNC) 934156321Sdamien BUF_KERNPROC(bp); 935156321Sdamien bp->b_iooffset = dbtob(bp->b_blkno); 936156321Sdamien bstrategy(bp); 937156321Sdamien 938170530Ssam if ((oldflags & B_ASYNC) == 0) { 939170530Ssam int rtval = bufwait(bp); 940170530Ssam brelse(bp); 941170530Ssam return (rtval); 942170530Ssam } else { 943170530Ssam /* 944156321Sdamien * don't allow the async write to saturate the I/O 945156321Sdamien * system. We will not deadlock here because 946156321Sdamien * we are blocking waiting for I/O that is already in-progress 947156321Sdamien * to complete. We do not block here if it is the update 948178354Ssam * or syncer daemon trying to clean up as that can lead 949178354Ssam * to deadlock. 950156321Sdamien */ 951156321Sdamien if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 952156321Sdamien waitrunningbufspace(); 953156321Sdamien } 954156321Sdamien 955156321Sdamien return (0); 956156321Sdamien} 957156321Sdamien 958156321Sdamienvoid 959156321Sdamienbufbdflush(struct bufobj *bo, struct buf *bp) 960156321Sdamien{ 961156321Sdamien struct buf *nbp; 962156321Sdamien 963156321Sdamien if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 964156321Sdamien (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 965156321Sdamien altbufferflushes++; 966156321Sdamien } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 967156321Sdamien BO_LOCK(bo); 968156321Sdamien /* 969156321Sdamien * Try to find a buffer to flush. 970156321Sdamien */ 971156321Sdamien TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 972156321Sdamien if ((nbp->b_vflags & BV_BKGRDINPROG) || 973156321Sdamien BUF_LOCK(nbp, 974156321Sdamien LK_EXCLUSIVE | LK_NOWAIT, NULL)) 975178354Ssam continue; 976156321Sdamien if (bp == nbp) 977156321Sdamien panic("bdwrite: found ourselves"); 978156321Sdamien BO_UNLOCK(bo); 979156321Sdamien /* Don't countdeps with the bo lock held. */ 980156321Sdamien if (buf_countdeps(nbp, 0)) { 981156321Sdamien BO_LOCK(bo); 982156321Sdamien BUF_UNLOCK(nbp); 983156321Sdamien continue; 984156321Sdamien } 985156321Sdamien if (nbp->b_flags & B_CLUSTEROK) { 986156321Sdamien vfs_bio_awrite(nbp); 987178354Ssam } else { 988178354Ssam bremfree(nbp); 989156321Sdamien bawrite(nbp); 990156321Sdamien } 991156321Sdamien dirtybufferflushes++; 992156321Sdamien break; 993156321Sdamien } 994156321Sdamien if (nbp == NULL) 995156321Sdamien BO_UNLOCK(bo); 996156321Sdamien } 997156321Sdamien} 998156321Sdamien 999156321Sdamien/* 1000156321Sdamien * Delayed write. (Buffer is marked dirty). Do not bother writing 1001192468Ssam * anything if the buffer is marked invalid. 1002170530Ssam * 1003156321Sdamien * Note that since the buffer must be completely valid, we can safely 1004156321Sdamien * set B_CACHE. In fact, we have to set B_CACHE here rather then in 1005156321Sdamien * biodone() in order to prevent getblk from writing the buffer 1006156321Sdamien * out synchronously. 1007156321Sdamien */ 1008156321Sdamienvoid 1009156321Sdamienbdwrite(struct buf *bp) 1010156321Sdamien{ 1011156321Sdamien struct thread *td = curthread; 1012156321Sdamien struct vnode *vp; 1013156321Sdamien struct bufobj *bo; 1014156321Sdamien 1015178354Ssam CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1016178354Ssam KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1017156321Sdamien BUF_ASSERT_HELD(bp); 1018156321Sdamien 1019156321Sdamien if (bp->b_flags & B_INVAL) { 1020156321Sdamien brelse(bp); 1021156321Sdamien return; 1022156321Sdamien } 1023156321Sdamien 1024156321Sdamien /* 1025156321Sdamien * If we have too many dirty buffers, don't create any more. 1026156321Sdamien * If we are wildly over our limit, then force a complete 1027156321Sdamien * cleanup. Otherwise, just keep the situation from getting 1028156321Sdamien * out of control. Note that we have to avoid a recursive 1029156321Sdamien * disaster and not try to clean up after our own cleanup! 1030156321Sdamien */ 1031156321Sdamien vp = bp->b_vp; 1032156321Sdamien bo = bp->b_bufobj; 1033243857Sglebius if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 1034156321Sdamien td->td_pflags |= TDP_INBDFLUSH; 1035156321Sdamien BO_BDFLUSH(bo, bp); 1036156321Sdamien td->td_pflags &= ~TDP_INBDFLUSH; 1037156321Sdamien } else 1038156321Sdamien recursiveflushes++; 1039156321Sdamien 1040156321Sdamien bdirty(bp); 1041156321Sdamien /* 1042156321Sdamien * Set B_CACHE, indicating that the buffer is fully valid. This is 1043156321Sdamien * true even of NFS now. 1044156321Sdamien */ 1045156321Sdamien bp->b_flags |= B_CACHE; 1046156321Sdamien 1047156321Sdamien /* 1048156321Sdamien * This bmap keeps the system from needing to do the bmap later, 1049156321Sdamien * perhaps when the system is attempting to do a sync. Since it 1050156321Sdamien * is likely that the indirect block -- or whatever other datastructure 1051156321Sdamien * that the filesystem needs is still in memory now, it is a good 1052156321Sdamien * thing to do this. Note also, that if the pageout daemon is 1053156321Sdamien * requesting a sync -- there might not be enough memory to do 1054156321Sdamien * the bmap then... So, this is important to do. 1055156321Sdamien */ 1056156321Sdamien if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1057156321Sdamien VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1058156321Sdamien } 1059156321Sdamien 1060156321Sdamien /* 1061156321Sdamien * Set the *dirty* buffer range based upon the VM system dirty 1062156321Sdamien * pages. 1063156321Sdamien * 1064156321Sdamien * Mark the buffer pages as clean. We need to do this here to 1065156321Sdamien * satisfy the vnode_pager and the pageout daemon, so that it 1066156321Sdamien * thinks that the pages have been "cleaned". Note that since 1067156321Sdamien * the pages are in a delayed write buffer -- the VFS layer 1068156321Sdamien * "will" see that the pages get written out on the next sync, 1069156321Sdamien * or perhaps the cluster will be completed. 1070156321Sdamien */ 1071156321Sdamien vfs_clean_pages_dirty_buf(bp); 1072156321Sdamien bqrelse(bp); 1073156321Sdamien 1074156321Sdamien /* 1075170530Ssam * Wakeup the buffer flushing daemon if we have a lot of dirty 1076192468Ssam * buffers (midpoint between our recovery point and our stall 1077192468Ssam * point). 1078192468Ssam */ 1079192468Ssam bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1080170530Ssam 1081192468Ssam /* 1082156321Sdamien * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1083156321Sdamien * due to the softdep code. 1084156321Sdamien */ 1085156321Sdamien} 1086156321Sdamien 1087156321Sdamien/* 1088156321Sdamien * bdirty: 1089156321Sdamien * 1090156321Sdamien * Turn buffer into delayed write request. We must clear BIO_READ and 1091156321Sdamien * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1092178354Ssam * itself to properly update it in the dirty/clean lists. We mark it 1093178958Ssam * B_DONE to ensure that any asynchronization of the buffer properly 1094178958Ssam * clears B_DONE ( else a panic will occur later ). 1095192468Ssam * 1096192468Ssam * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1097156321Sdamien * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1098170530Ssam * should only be called if the buffer is known-good. 1099170530Ssam * 1100156321Sdamien * Since the buffer is not on a queue, we do not update the numfreebuffers 1101178354Ssam * count. 1102178354Ssam * 1103156321Sdamien * The buffer must be on QUEUE_NONE. 1104156321Sdamien */ 1105178354Ssamvoid 1106192468Ssambdirty(struct buf *bp) 1107178354Ssam{ 1108178354Ssam 1109192468Ssam CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1110170530Ssam bp, bp->b_vp, bp->b_flags); 1111170530Ssam KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1112170530Ssam KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1113156321Sdamien ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1114156321Sdamien BUF_ASSERT_HELD(bp); 1115156321Sdamien bp->b_flags &= ~(B_RELBUF); 1116178354Ssam bp->b_iocmd = BIO_WRITE; 1117156321Sdamien 1118156321Sdamien if ((bp->b_flags & B_DELWRI) == 0) { 1119156321Sdamien bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1120156321Sdamien reassignbuf(bp); 1121156321Sdamien atomic_add_int(&numdirtybuffers, 1); 1122156321Sdamien bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1123156321Sdamien } 1124156321Sdamien} 1125156321Sdamien 1126156321Sdamien/* 1127156321Sdamien * bundirty: 1128156321Sdamien * 1129156321Sdamien * Clear B_DELWRI for buffer. 1130156321Sdamien * 1131156321Sdamien * Since the buffer is not on a queue, we do not update the numfreebuffers 1132156321Sdamien * count. 1133156321Sdamien * 1134156321Sdamien * The buffer must be on QUEUE_NONE. 1135156321Sdamien */ 1136156321Sdamien 1137156321Sdamienvoid 1138156321Sdamienbundirty(struct buf *bp) 1139156321Sdamien{ 1140156321Sdamien 1141156321Sdamien CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1142156321Sdamien KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1143156321Sdamien KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1144156321Sdamien ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1145156321Sdamien BUF_ASSERT_HELD(bp); 1146156321Sdamien 1147156321Sdamien if (bp->b_flags & B_DELWRI) { 1148156321Sdamien bp->b_flags &= ~B_DELWRI; 1149156321Sdamien reassignbuf(bp); 1150156321Sdamien atomic_subtract_int(&numdirtybuffers, 1); 1151156321Sdamien numdirtywakeup(lodirtybuffers); 1152156321Sdamien } 1153156321Sdamien /* 1154156321Sdamien * Since it is now being written, we can clear its deferred write flag. 1155156321Sdamien */ 1156156975Sdamien bp->b_flags &= ~B_DEFERRED; 1157156321Sdamien} 1158156321Sdamien 1159156321Sdamien/* 1160156321Sdamien * bawrite: 1161156321Sdamien * 1162156321Sdamien * Asynchronous write. Start output on a buffer, but do not wait for 1163156321Sdamien * it to complete. The buffer is released when the output completes. 1164156321Sdamien * 1165156975Sdamien * bwrite() ( or the VOP routine anyway ) is responsible for handling 1166156975Sdamien * B_INVAL buffers. Not us. 1167156975Sdamien */ 1168156975Sdamienvoid 1169156975Sdamienbawrite(struct buf *bp) 1170156975Sdamien{ 1171156321Sdamien 1172156321Sdamien bp->b_flags |= B_ASYNC; 1173156321Sdamien (void) bwrite(bp); 1174156321Sdamien} 1175156321Sdamien 1176156321Sdamien/* 1177156321Sdamien * bwillwrite: 1178156321Sdamien * 1179156321Sdamien * Called prior to the locking of any vnodes when we are expecting to 1180156321Sdamien * write. We do not want to starve the buffer cache with too many 1181156321Sdamien * dirty buffers so we block here. By blocking prior to the locking 1182156321Sdamien * of any vnodes we attempt to avoid the situation where a locked vnode 1183156321Sdamien * prevents the various system daemons from flushing related buffers. 1184156321Sdamien */ 1185156321Sdamien 1186156321Sdamienvoid 1187156321Sdamienbwillwrite(void) 1188156321Sdamien{ 1189156321Sdamien 1190156321Sdamien if (numdirtybuffers >= hidirtybuffers) { 1191156321Sdamien mtx_lock(&nblock); 1192156321Sdamien while (numdirtybuffers >= hidirtybuffers) { 1193156321Sdamien bd_wakeup(1); 1194156321Sdamien needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1195156321Sdamien msleep(&needsbuffer, &nblock, 1196156321Sdamien (PRIBIO + 4), "flswai", 0); 1197156321Sdamien } 1198156321Sdamien mtx_unlock(&nblock); 1199156321Sdamien } 1200156321Sdamien} 1201156321Sdamien 1202156321Sdamien/* 1203156321Sdamien * Return true if we have too many dirty buffers. 1204156321Sdamien */ 1205156321Sdamienint 1206156321Sdamienbuf_dirty_count_severe(void) 1207156321Sdamien{ 1208156321Sdamien 1209156321Sdamien return(numdirtybuffers >= hidirtybuffers); 1210156321Sdamien} 1211156321Sdamien 1212156321Sdamienstatic __noinline int 1213156321Sdamienbuf_vm_page_count_severe(void) 1214178958Ssam{ 1215178958Ssam 1216178958Ssam KFAIL_POINT_CODE(DEBUG_FP, buf_pressure, return 1); 1217178958Ssam 1218178958Ssam return vm_page_count_severe(); 1219178958Ssam} 1220178958Ssam 1221178958Ssam/* 1222178958Ssam * brelse: 1223178958Ssam * 1224178958Ssam * Release a busy buffer and, if requested, free its resources. The 1225178958Ssam * buffer will be stashed in the appropriate bufqueue[] allowing it 1226178958Ssam * to be accessed later as a cache entity or reused for other purposes. 1227178958Ssam */ 1228178958Ssamvoid 1229178958Ssambrelse(struct buf *bp) 1230178958Ssam{ 1231178958Ssam CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1232178958Ssam bp, bp->b_vp, bp->b_flags); 1233178958Ssam KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1234178958Ssam ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1235178958Ssam 1236178958Ssam if (bp->b_flags & B_MANAGED) { 1237156321Sdamien bqrelse(bp); 1238156321Sdamien return; 1239156321Sdamien } 1240156321Sdamien 1241156321Sdamien if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 1242178354Ssam bp->b_error == EIO && !(bp->b_flags & B_INVAL)) { 1243178354Ssam /* 1244156321Sdamien * Failed write, redirty. Must clear BIO_ERROR to prevent 1245156321Sdamien * pages from being scrapped. If the error is anything 1246156321Sdamien * other than an I/O error (EIO), assume that retrying 1247156321Sdamien * is futile. 1248156321Sdamien */ 1249156321Sdamien bp->b_ioflags &= ~BIO_ERROR; 1250156321Sdamien bdirty(bp); 1251156321Sdamien } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1252156321Sdamien (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1253156321Sdamien /* 1254156321Sdamien * Either a failed I/O or we were asked to free or not 1255156321Sdamien * cache the buffer. 1256156321Sdamien */ 1257156321Sdamien bp->b_flags |= B_INVAL; 1258156321Sdamien if (!LIST_EMPTY(&bp->b_dep)) 1259156321Sdamien buf_deallocate(bp); 1260156321Sdamien if (bp->b_flags & B_DELWRI) { 1261156321Sdamien atomic_subtract_int(&numdirtybuffers, 1); 1262156321Sdamien numdirtywakeup(lodirtybuffers); 1263156321Sdamien } 1264156321Sdamien bp->b_flags &= ~(B_DELWRI | B_CACHE); 1265156321Sdamien if ((bp->b_flags & B_VMIO) == 0) { 1266156321Sdamien if (bp->b_bufsize) 1267156321Sdamien allocbuf(bp, 0); 1268178958Ssam if (bp->b_vp) 1269156321Sdamien brelvp(bp); 1270156321Sdamien } 1271156321Sdamien } 1272190532Ssam 1273156321Sdamien /* 1274156321Sdamien * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1275156321Sdamien * is called with B_DELWRI set, the underlying pages may wind up 1276156321Sdamien * getting freed causing a previous write (bdwrite()) to get 'lost' 1277156321Sdamien * because pages associated with a B_DELWRI bp are marked clean. 1278156321Sdamien * 1279156321Sdamien * We still allow the B_INVAL case to call vfs_vmio_release(), even 1280156321Sdamien * if B_DELWRI is set. 1281156321Sdamien * 1282156321Sdamien * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1283156321Sdamien * on pages to return pages to the VM page queues. 1284156321Sdamien */ 1285156321Sdamien if (bp->b_flags & B_DELWRI) 1286156321Sdamien bp->b_flags &= ~B_RELBUF; 1287156321Sdamien else if (buf_vm_page_count_severe()) { 1288156321Sdamien /* 1289156321Sdamien * The locking of the BO_LOCK is not necessary since 1290156321Sdamien * BKGRDINPROG cannot be set while we hold the buf 1291156321Sdamien * lock, it can only be cleared if it is already 1292156321Sdamien * pending. 1293156321Sdamien */ 1294156321Sdamien if (bp->b_vp) { 1295156321Sdamien if (!(bp->b_vflags & BV_BKGRDINPROG)) 1296156321Sdamien bp->b_flags |= B_RELBUF; 1297156321Sdamien } else 1298156321Sdamien bp->b_flags |= B_RELBUF; 1299156321Sdamien } 1300156321Sdamien 1301156321Sdamien /* 1302156321Sdamien * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1303178354Ssam * constituted, not even NFS buffers now. Two flags effect this. If 1304178354Ssam * B_INVAL, the struct buf is invalidated but the VM object is kept 1305156321Sdamien * around ( i.e. so it is trivial to reconstitute the buffer later ). 1306156321Sdamien * 1307156321Sdamien * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1308173386Skevlo * invalidated. BIO_ERROR cannot be set for a failed write unless the 1309156321Sdamien * buffer is also B_INVAL because it hits the re-dirtying code above. 1310156321Sdamien * 1311156321Sdamien * Normally we can do this whether a buffer is B_DELWRI or not. If 1312156321Sdamien * the buffer is an NFS buffer, it is tracking piecemeal writes or 1313156321Sdamien * the commit state and we cannot afford to lose the buffer. If the 1314156321Sdamien * buffer has a background write in progress, we need to keep it 1315156321Sdamien * around to prevent it from being reconstituted and starting a second 1316156321Sdamien * background write. 1317178354Ssam */ 1318156321Sdamien if ((bp->b_flags & B_VMIO) 1319173386Skevlo && !(bp->b_vp->v_mount != NULL && 1320173386Skevlo (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1321262007Skevlo !vn_isdisk(bp->b_vp, NULL) && 1322178354Ssam (bp->b_flags & B_DELWRI)) 1323173386Skevlo ) { 1324173386Skevlo 1325173386Skevlo int i, j, resid; 1326173386Skevlo vm_page_t m; 1327173386Skevlo off_t foff; 1328173386Skevlo vm_pindex_t poff; 1329156321Sdamien vm_object_t obj; 1330156321Sdamien 1331156321Sdamien obj = bp->b_bufobj->bo_object; 1332156321Sdamien 1333156321Sdamien /* 1334156321Sdamien * Get the base offset and length of the buffer. Note that 1335156321Sdamien * in the VMIO case if the buffer block size is not 1336156321Sdamien * page-aligned then b_data pointer may not be page-aligned. 1337156321Sdamien * But our b_pages[] array *IS* page aligned. 1338192468Ssam * 1339156321Sdamien * block sizes less then DEV_BSIZE (usually 512) are not 1340156321Sdamien * supported due to the page granularity bits (m->valid, 1341156321Sdamien * m->dirty, etc...). 1342156321Sdamien * 1343156321Sdamien * See man buf(9) for more information 1344192468Ssam */ 1345156321Sdamien resid = bp->b_bufsize; 1346156321Sdamien foff = bp->b_offset; 1347156321Sdamien VM_OBJECT_LOCK(obj); 1348156321Sdamien for (i = 0; i < bp->b_npages; i++) { 1349178354Ssam int had_bogus = 0; 1350178354Ssam 1351156321Sdamien m = bp->b_pages[i]; 1352156321Sdamien 1353156321Sdamien /* 1354156321Sdamien * If we hit a bogus page, fixup *all* the bogus pages 1355156321Sdamien * now. 1356156321Sdamien */ 1357190532Ssam if (m == bogus_page) { 1358178354Ssam poff = OFF_TO_IDX(bp->b_offset); 1359156321Sdamien had_bogus = 1; 1360156321Sdamien 1361156321Sdamien for (j = i; j < bp->b_npages; j++) { 1362156321Sdamien vm_page_t mtmp; 1363156321Sdamien mtmp = bp->b_pages[j]; 1364156321Sdamien if (mtmp == bogus_page) { 1365156321Sdamien mtmp = vm_page_lookup(obj, poff + j); 1366156321Sdamien if (!mtmp) { 1367156321Sdamien panic("brelse: page missing\n"); 1368156321Sdamien } 1369156321Sdamien bp->b_pages[j] = mtmp; 1370156321Sdamien } 1371156321Sdamien } 1372156321Sdamien 1373156321Sdamien if ((bp->b_flags & B_INVAL) == 0) { 1374156321Sdamien pmap_qenter( 1375178354Ssam trunc_page((vm_offset_t)bp->b_data), 1376178354Ssam bp->b_pages, bp->b_npages); 1377156321Sdamien } 1378156321Sdamien m = bp->b_pages[i]; 1379156321Sdamien } 1380156321Sdamien if ((bp->b_flags & B_NOCACHE) || 1381156321Sdamien (bp->b_ioflags & BIO_ERROR && 1382156321Sdamien bp->b_iocmd == BIO_READ)) { 1383156321Sdamien int poffset = foff & PAGE_MASK; 1384156321Sdamien int presid = resid > (PAGE_SIZE - poffset) ? 1385156321Sdamien (PAGE_SIZE - poffset) : resid; 1386178354Ssam 1387178354Ssam KASSERT(presid >= 0, ("brelse: extra page")); 1388178354Ssam vm_page_set_invalid(m, poffset, presid); 1389156321Sdamien if (had_bogus) 1390178354Ssam printf("avoided corruption bug in bogus_page/brelse code\n"); 1391178354Ssam } 1392178354Ssam resid -= PAGE_SIZE - (foff & PAGE_MASK); 1393178354Ssam foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1394178354Ssam } 1395178354Ssam VM_OBJECT_UNLOCK(obj); 1396178354Ssam if (bp->b_flags & (B_INVAL | B_RELBUF)) 1397178354Ssam vfs_vmio_release(bp); 1398178354Ssam 1399178354Ssam } else if (bp->b_flags & B_VMIO) { 1400156321Sdamien 1401178354Ssam if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1402178354Ssam vfs_vmio_release(bp); 1403178354Ssam } 1404178354Ssam 1405178354Ssam } else if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0) { 1406178354Ssam if (bp->b_bufsize != 0) 1407190532Ssam allocbuf(bp, 0); 1408190532Ssam if (bp->b_vp != NULL) 1409178354Ssam brelvp(bp); 1410178354Ssam } 1411190532Ssam 1412190532Ssam if (BUF_LOCKRECURSED(bp)) { 1413178354Ssam /* do not release to free list */ 1414178354Ssam BUF_UNLOCK(bp); 1415178354Ssam return; 1416190532Ssam } 1417178354Ssam 1418178354Ssam /* enqueue */ 1419178354Ssam mtx_lock(&bqlock); 1420178354Ssam /* Handle delayed bremfree() processing. */ 1421156321Sdamien if (bp->b_flags & B_REMFREE) { 1422178354Ssam struct bufobj *bo; 1423178354Ssam 1424178354Ssam bo = bp->b_bufobj; 1425178354Ssam if (bo != NULL) 1426156321Sdamien BO_LOCK(bo); 1427178354Ssam bremfreel(bp); 1428178354Ssam if (bo != NULL) 1429156321Sdamien BO_UNLOCK(bo); 1430178354Ssam } 1431178354Ssam if (bp->b_qindex != QUEUE_NONE) 1432178354Ssam panic("brelse: free buffer onto another queue???"); 1433178354Ssam 1434178354Ssam /* 1435178354Ssam * If the buffer has junk contents signal it and eventually 1436178354Ssam * clean up B_DELWRI and diassociate the vnode so that gbincore() 1437178354Ssam * doesn't find it. 1438156321Sdamien */ 1439178354Ssam if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 1440178354Ssam (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 1441178354Ssam bp->b_flags |= B_INVAL; 1442178354Ssam if (bp->b_flags & B_INVAL) { 1443156321Sdamien if (bp->b_flags & B_DELWRI) 1444178354Ssam bundirty(bp); 1445178354Ssam if (bp->b_vp) 1446178354Ssam brelvp(bp); 1447178354Ssam } 1448178354Ssam 1449178354Ssam /* buffers with no memory */ 1450178354Ssam if (bp->b_bufsize == 0) { 1451178354Ssam bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1452178354Ssam if (bp->b_vflags & BV_BKGRDINPROG) 1453178354Ssam panic("losing buffer 1"); 1454156321Sdamien if (bp->b_kvasize) { 1455156321Sdamien bp->b_qindex = QUEUE_EMPTYKVA; 1456156321Sdamien } else { 1457156321Sdamien bp->b_qindex = QUEUE_EMPTY; 1458156321Sdamien } 1459156321Sdamien TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1460178354Ssam /* buffers with junk contents */ 1461178354Ssam } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1462178354Ssam (bp->b_ioflags & BIO_ERROR)) { 1463156321Sdamien bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1464156321Sdamien if (bp->b_vflags & BV_BKGRDINPROG) 1465156321Sdamien panic("losing buffer 2"); 1466156321Sdamien bp->b_qindex = QUEUE_CLEAN; 1467178354Ssam TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1468156321Sdamien /* remaining buffers */ 1469156321Sdamien } else { 1470156321Sdamien if ((bp->b_flags & (B_DELWRI|B_NEEDSGIANT)) == 1471156321Sdamien (B_DELWRI|B_NEEDSGIANT)) 1472156321Sdamien bp->b_qindex = QUEUE_DIRTY_GIANT; 1473178354Ssam else if (bp->b_flags & B_DELWRI) 1474156321Sdamien bp->b_qindex = QUEUE_DIRTY; 1475156321Sdamien else 1476156321Sdamien bp->b_qindex = QUEUE_CLEAN; 1477156321Sdamien if (bp->b_flags & B_AGE) 1478178354Ssam TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1479178354Ssam else 1480178354Ssam TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1481178354Ssam } 1482178354Ssam mtx_unlock(&bqlock); 1483178354Ssam 1484178354Ssam /* 1485156321Sdamien * Fixup numfreebuffers count. The bp is on an appropriate queue 1486206358Srpaulo * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1487178354Ssam * We've already handled the B_INVAL case ( B_DELWRI will be clear 1488156321Sdamien * if B_INVAL is set ). 1489156321Sdamien */ 1490156321Sdamien 1491156321Sdamien if (!(bp->b_flags & B_DELWRI)) { 1492156321Sdamien struct bufobj *bo; 1493156321Sdamien 1494156321Sdamien bo = bp->b_bufobj; 1495156321Sdamien if (bo != NULL) 1496262007Skevlo BO_LOCK(bo); 1497178354Ssam bufcountwakeup(bp); 1498156321Sdamien if (bo != NULL) 1499156321Sdamien BO_UNLOCK(bo); 1500156321Sdamien } 1501156321Sdamien 1502156321Sdamien /* 1503156321Sdamien * Something we can maybe free or reuse 1504156321Sdamien */ 1505156321Sdamien if (bp->b_bufsize || bp->b_kvasize) 1506156321Sdamien bufspacewakeup(); 1507178354Ssam 1508178354Ssam bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1509178354Ssam if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1510178354Ssam panic("brelse: not dirty"); 1511178354Ssam /* unlock */ 1512178354Ssam BUF_UNLOCK(bp); 1513190532Ssam} 1514178354Ssam 1515178354Ssam/* 1516178354Ssam * Release a buffer back to the appropriate queue but do not try to free 1517178354Ssam * it. The buffer is expected to be used again soon. 1518178354Ssam * 1519178354Ssam * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1520178354Ssam * biodone() to requeue an async I/O on completion. It is also used when 1521178354Ssam * known good buffers need to be requeued but we think we may need the data 1522156321Sdamien * again soon. 1523156321Sdamien * 1524156321Sdamien * XXX we should be able to leave the B_RELBUF hint set on completion. 1525156321Sdamien */ 1526156321Sdamienvoid 1527156321Sdamienbqrelse(struct buf *bp) 1528156321Sdamien{ 1529156321Sdamien struct bufobj *bo; 1530156321Sdamien 1531156321Sdamien CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1532156321Sdamien KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1533156321Sdamien ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1534156321Sdamien 1535156321Sdamien if (BUF_LOCKRECURSED(bp)) { 1536156321Sdamien /* do not release to free list */ 1537243857Sglebius BUF_UNLOCK(bp); 1538156321Sdamien return; 1539156321Sdamien } 1540156321Sdamien 1541156321Sdamien bo = bp->b_bufobj; 1542156321Sdamien if (bp->b_flags & B_MANAGED) { 1543156321Sdamien if (bp->b_flags & B_REMFREE) { 1544156321Sdamien mtx_lock(&bqlock); 1545156321Sdamien if (bo != NULL) 1546156321Sdamien BO_LOCK(bo); 1547156321Sdamien bremfreel(bp); 1548156321Sdamien if (bo != NULL) 1549156321Sdamien BO_UNLOCK(bo); 1550156321Sdamien mtx_unlock(&bqlock); 1551156321Sdamien } 1552156321Sdamien bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1553156321Sdamien BUF_UNLOCK(bp); 1554156321Sdamien return; 1555156321Sdamien } 1556156321Sdamien 1557156321Sdamien mtx_lock(&bqlock); 1558156321Sdamien /* Handle delayed bremfree() processing. */ 1559192468Ssam if (bp->b_flags & B_REMFREE) { 1560156321Sdamien if (bo != NULL) 1561156321Sdamien BO_LOCK(bo); 1562156321Sdamien bremfreel(bp); 1563156321Sdamien if (bo != NULL) 1564156321Sdamien BO_UNLOCK(bo); 1565192468Ssam } 1566156321Sdamien if (bp->b_qindex != QUEUE_NONE) 1567156321Sdamien panic("bqrelse: free buffer onto another queue???"); 1568156321Sdamien /* buffers with stale but valid contents */ 1569156321Sdamien if (bp->b_flags & B_DELWRI) { 1570156321Sdamien if (bp->b_flags & B_NEEDSGIANT) 1571156321Sdamien bp->b_qindex = QUEUE_DIRTY_GIANT; 1572178354Ssam else 1573178354Ssam bp->b_qindex = QUEUE_DIRTY; 1574178354Ssam TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1575178354Ssam } else { 1576156321Sdamien /* 1577178354Ssam * The locking of the BO_LOCK for checking of the 1578156321Sdamien * BV_BKGRDINPROG is not necessary since the 1579156321Sdamien * BV_BKGRDINPROG cannot be set while we hold the buf 1580156321Sdamien * lock, it can only be cleared if it is already 1581156321Sdamien * pending. 1582190532Ssam */ 1583178354Ssam if (!buf_vm_page_count_severe() || (bp->b_vflags & BV_BKGRDINPROG)) { 1584156321Sdamien bp->b_qindex = QUEUE_CLEAN; 1585156321Sdamien TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1586156321Sdamien b_freelist); 1587156321Sdamien } else { 1588156321Sdamien /* 1589156321Sdamien * We are too low on memory, we have to try to free 1590156321Sdamien * the buffer (most importantly: the wired pages 1591156321Sdamien * making up its backing store) *now*. 1592156321Sdamien */ 1593178354Ssam mtx_unlock(&bqlock); 1594178354Ssam brelse(bp); 1595156321Sdamien return; 1596156321Sdamien } 1597156321Sdamien } 1598156321Sdamien mtx_unlock(&bqlock); 1599156321Sdamien 1600156321Sdamien if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) { 1601156321Sdamien if (bo != NULL) 1602156321Sdamien BO_LOCK(bo); 1603156321Sdamien bufcountwakeup(bp); 1604156321Sdamien if (bo != NULL) 1605178354Ssam BO_UNLOCK(bo); 1606156321Sdamien } 1607156321Sdamien 1608178354Ssam /* 1609156321Sdamien * Something we can maybe free or reuse. 1610156321Sdamien */ 1611156321Sdamien if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1612178354Ssam bufspacewakeup(); 1613156321Sdamien 1614156975Sdamien bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1615178354Ssam if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1616156975Sdamien panic("bqrelse: not dirty"); 1617156975Sdamien /* unlock */ 1618156321Sdamien BUF_UNLOCK(bp); 1619178354Ssam} 1620178354Ssam 1621178354Ssam/* Give pages used by the bp back to the VM system (where possible) */ 1622156321Sdamienstatic void 1623178354Ssamvfs_vmio_release(struct buf *bp) 1624178354Ssam{ 1625178354Ssam int i; 1626178354Ssam vm_page_t m; 1627178354Ssam 1628178354Ssam VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1629178354Ssam for (i = 0; i < bp->b_npages; i++) { 1630178354Ssam m = bp->b_pages[i]; 1631178354Ssam bp->b_pages[i] = NULL; 1632178354Ssam /* 1633178354Ssam * In order to keep page LRU ordering consistent, put 1634178354Ssam * everything on the inactive queue. 1635178354Ssam */ 1636156321Sdamien vm_page_lock(m); 1637178354Ssam vm_page_unwire(m, 0); 1638178354Ssam /* 1639178354Ssam * We don't mess with busy pages, it is 1640156321Sdamien * the responsibility of the process that 1641178354Ssam * busied the pages to deal with them. 1642178354Ssam */ 1643178354Ssam if ((m->oflags & VPO_BUSY) == 0 && m->busy == 0 && 1644178354Ssam m->wire_count == 0) { 1645156321Sdamien /* 1646178354Ssam * Might as well free the page if we can and it has 1647178354Ssam * no valid data. We also free the page if the 1648178354Ssam * buffer was used for direct I/O 1649178354Ssam */ 1650156321Sdamien if ((bp->b_flags & B_ASYNC) == 0 && !m->valid) { 1651178354Ssam vm_page_free(m); 1652178354Ssam } else if (bp->b_flags & B_DIRECT) { 1653178354Ssam vm_page_try_to_free(m); 1654178354Ssam } else if (buf_vm_page_count_severe()) { 1655178354Ssam vm_page_try_to_cache(m); 1656178354Ssam } 1657178354Ssam } 1658156321Sdamien vm_page_unlock(m); 1659178354Ssam } 1660156321Sdamien VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1661178354Ssam pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1662178354Ssam 1663178354Ssam if (bp->b_bufsize) { 1664178354Ssam bufspacewakeup(); 1665178354Ssam bp->b_bufsize = 0; 1666178354Ssam } 1667178354Ssam bp->b_npages = 0; 1668178354Ssam bp->b_flags &= ~B_VMIO; 1669178354Ssam if (bp->b_vp) 1670178354Ssam brelvp(bp); 1671178354Ssam} 1672178354Ssam 1673178354Ssam/* 1674178354Ssam * Check to see if a block at a particular lbn is available for a clustered 1675156321Sdamien * write. 1676178354Ssam */ 1677156321Sdamienstatic int 1678178354Ssamvfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1679178354Ssam{ 1680178354Ssam struct buf *bpa; 1681178354Ssam int match; 1682178354Ssam 1683178354Ssam match = 0; 1684178354Ssam 1685178354Ssam /* If the buf isn't in core skip it */ 1686156321Sdamien if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1687178354Ssam return (0); 1688156321Sdamien 1689178354Ssam /* If the buf is busy we don't want to wait for it */ 1690178354Ssam if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1691178354Ssam return (0); 1692178354Ssam 1693156321Sdamien /* Only cluster with valid clusterable delayed write buffers */ 1694178354Ssam if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1695156321Sdamien (B_DELWRI | B_CLUSTEROK)) 1696156321Sdamien goto done; 1697156321Sdamien 1698165352Sbms if (bpa->b_bufsize != size) 1699156321Sdamien goto done; 1700165352Sbms 1701178354Ssam /* 1702156321Sdamien * Check to see if it is in the expected place on disk and that the 1703178354Ssam * block has been mapped. 1704156321Sdamien */ 1705178354Ssam if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1706156321Sdamien match = 1; 1707178354Ssamdone: 1708178354Ssam BUF_UNLOCK(bpa); 1709156321Sdamien return (match); 1710178354Ssam} 1711178354Ssam 1712178354Ssam/* 1713178354Ssam * vfs_bio_awrite: 1714178354Ssam * 1715178354Ssam * Implement clustered async writes for clearing out B_DELWRI buffers. 1716178354Ssam * This is much better then the old way of writing only one buffer at 1717178354Ssam * a time. Note that we may not be presented with the buffers in the 1718156321Sdamien * correct order, so we search for the cluster in both directions. 1719156321Sdamien */ 1720156321Sdamienint 1721156321Sdamienvfs_bio_awrite(struct buf *bp) 1722156321Sdamien{ 1723156321Sdamien struct bufobj *bo; 1724178354Ssam int i; 1725178354Ssam int j; 1726178354Ssam daddr_t lblkno = bp->b_lblkno; 1727156321Sdamien struct vnode *vp = bp->b_vp; 1728156321Sdamien int ncl; 1729156321Sdamien int nwritten; 1730178704Sthompsa int size; 1731156321Sdamien int maxcl; 1732178354Ssam 1733178354Ssam bo = &vp->v_bufobj; 1734178354Ssam /* 1735178354Ssam * right now we support clustered writing only to regular files. If 1736178354Ssam * we find a clusterable block we could be in the middle of a cluster 1737156321Sdamien * rather then at the beginning. 1738156321Sdamien */ 1739178354Ssam if ((vp->v_type == VREG) && 1740156321Sdamien (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1741178704Sthompsa (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1742178704Sthompsa 1743178704Sthompsa size = vp->v_mount->mnt_stat.f_iosize; 1744156321Sdamien maxcl = MAXPHYS / size; 1745178354Ssam 1746178354Ssam BO_LOCK(bo); 1747178354Ssam for (i = 1; i < maxcl; i++) 1748178704Sthompsa if (vfs_bio_clcheck(vp, size, lblkno + i, 1749178354Ssam bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1750178354Ssam break; 1751178704Sthompsa 1752178704Sthompsa for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1753178704Sthompsa if (vfs_bio_clcheck(vp, size, lblkno - j, 1754156321Sdamien bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1755156321Sdamien break; 1756156321Sdamien BO_UNLOCK(bo); 1757156321Sdamien --j; 1758156321Sdamien ncl = i + j; 1759156321Sdamien /* 1760156321Sdamien * this is a possible cluster write 1761156321Sdamien */ 1762156321Sdamien if (ncl != 1) { 1763156321Sdamien BUF_UNLOCK(bp); 1764156321Sdamien nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1765156321Sdamien return nwritten; 1766156321Sdamien } 1767156321Sdamien } 1768156321Sdamien bremfree(bp); 1769156321Sdamien bp->b_flags |= B_ASYNC; 1770156321Sdamien /* 1771156321Sdamien * default (old) behavior, writing out only one block 1772156321Sdamien * 1773156321Sdamien * XXX returns b_bufsize instead of b_bcount for nwritten? 1774156321Sdamien */ 1775156321Sdamien nwritten = bp->b_bufsize; 1776156321Sdamien (void) bwrite(bp); 1777178354Ssam 1778156321Sdamien return nwritten; 1779156321Sdamien} 1780156321Sdamien 1781156321Sdamien/* 1782156321Sdamien * getnewbuf: 1783156321Sdamien * 1784156321Sdamien * Find and initialize a new buffer header, freeing up existing buffers 1785156321Sdamien * in the bufqueues as necessary. The new buffer is returned locked. 1786156321Sdamien * 1787156321Sdamien * Important: B_INVAL is not set. If the caller wishes to throw the 1788156321Sdamien * buffer away, the caller must set B_INVAL prior to calling brelse(). 1789156321Sdamien * 1790156321Sdamien * We block if: 1791156321Sdamien * We have insufficient buffer headers 1792156321Sdamien * We have insufficient buffer space 1793156321Sdamien * buffer_map is too fragmented ( space reservation fails ) 1794156321Sdamien * If we have to flush dirty buffers ( but we try to avoid this ) 1795156321Sdamien * 1796156321Sdamien * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1797156321Sdamien * Instead we ask the buf daemon to do it for us. We attempt to 1798156321Sdamien * avoid piecemeal wakeups of the pageout daemon. 1799156321Sdamien */ 1800156321Sdamien 1801156321Sdamienstatic struct buf * 1802156321Sdamiengetnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize, 1803156321Sdamien int gbflags) 1804156321Sdamien{ 1805156321Sdamien struct thread *td; 1806156321Sdamien struct buf *bp; 1807156321Sdamien struct buf *nbp; 1808156321Sdamien int defrag = 0; 1809156321Sdamien int nqindex; 1810156321Sdamien static int flushingbufs; 1811156321Sdamien 1812156321Sdamien td = curthread; 1813156321Sdamien /* 1814156321Sdamien * We can't afford to block since we might be holding a vnode lock, 1815156321Sdamien * which may prevent system daemons from running. We deal with 1816156321Sdamien * low-memory situations by proactively returning memory and running 1817156321Sdamien * async I/O rather then sync I/O. 1818156321Sdamien */ 1819156321Sdamien atomic_add_int(&getnewbufcalls, 1); 1820156321Sdamien atomic_subtract_int(&getnewbufrestarts, 1); 1821156321Sdamienrestart: 1822156321Sdamien atomic_add_int(&getnewbufrestarts, 1); 1823156321Sdamien 1824156321Sdamien /* 1825156321Sdamien * Setup for scan. If we do not have enough free buffers, 1826156321Sdamien * we setup a degenerate case that immediately fails. Note 1827156321Sdamien * that if we are specially marked process, we are allowed to 1828156321Sdamien * dip into our reserves. 1829156321Sdamien * 1830156321Sdamien * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1831156321Sdamien * 1832156321Sdamien * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1833178354Ssam * However, there are a number of cases (defragging, reusing, ...) 1834156321Sdamien * where we cannot backup. 1835156321Sdamien */ 1836156321Sdamien mtx_lock(&bqlock); 1837156321Sdamien nqindex = QUEUE_EMPTYKVA; 1838156321Sdamien nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1839156321Sdamien 1840156321Sdamien if (nbp == NULL) { 1841156321Sdamien /* 1842156321Sdamien * If no EMPTYKVA buffers and we are either 1843156321Sdamien * defragging or reusing, locate a CLEAN buffer 1844156321Sdamien * to free or reuse. If bufspace useage is low 1845156321Sdamien * skip this step so we can allocate a new buffer. 1846156321Sdamien */ 1847156321Sdamien if (defrag || bufspace >= lobufspace) { 1848156321Sdamien nqindex = QUEUE_CLEAN; 1849156321Sdamien nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1850156321Sdamien } 1851156321Sdamien 1852156321Sdamien /* 1853156321Sdamien * If we could not find or were not allowed to reuse a 1854156321Sdamien * CLEAN buffer, check to see if it is ok to use an EMPTY 1855156321Sdamien * buffer. We can only use an EMPTY buffer if allocating 1856156321Sdamien * its KVA would not otherwise run us out of buffer space. 1857156321Sdamien */ 1858156321Sdamien if (nbp == NULL && defrag == 0 && 1859156321Sdamien bufspace + maxsize < hibufspace) { 1860156321Sdamien nqindex = QUEUE_EMPTY; 1861156321Sdamien nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1862156321Sdamien } 1863156321Sdamien } 1864156321Sdamien 1865156321Sdamien /* 1866156321Sdamien * Run scan, possibly freeing data and/or kva mappings on the fly 1867156321Sdamien * depending. 1868156321Sdamien */ 1869156321Sdamien 1870156321Sdamien while ((bp = nbp) != NULL) { 1871156321Sdamien int qindex = nqindex; 1872156321Sdamien 1873156321Sdamien /* 1874156321Sdamien * Calculate next bp ( we can only use it if we do not block 1875156321Sdamien * or do other fancy things ). 1876156321Sdamien */ 1877156321Sdamien if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1878156321Sdamien switch(qindex) { 1879178354Ssam case QUEUE_EMPTY: 1880178354Ssam nqindex = QUEUE_EMPTYKVA; 1881156321Sdamien if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1882156321Sdamien break; 1883156321Sdamien /* FALLTHROUGH */ 1884156321Sdamien case QUEUE_EMPTYKVA: 1885156321Sdamien nqindex = QUEUE_CLEAN; 1886178354Ssam if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1887156321Sdamien break; 1888156321Sdamien /* FALLTHROUGH */ 1889156321Sdamien case QUEUE_CLEAN: 1890156321Sdamien /* 1891156321Sdamien * nbp is NULL. 1892156321Sdamien */ 1893156321Sdamien break; 1894156321Sdamien } 1895156321Sdamien } 1896178354Ssam /* 1897178354Ssam * If we are defragging then we need a buffer with 1898156321Sdamien * b_kvasize != 0. XXX this situation should no longer 1899156321Sdamien * occur, if defrag is non-zero the buffer's b_kvasize 1900156321Sdamien * should also be non-zero at this point. XXX 1901156321Sdamien */ 1902156321Sdamien if (defrag && bp->b_kvasize == 0) { 1903178354Ssam printf("Warning: defrag empty buffer %p\n", bp); 1904156321Sdamien continue; 1905156321Sdamien } 1906156321Sdamien 1907156321Sdamien /* 1908156321Sdamien * Start freeing the bp. This is somewhat involved. nbp 1909156321Sdamien * remains valid only for QUEUE_EMPTY[KVA] bp's. 1910156321Sdamien */ 1911156321Sdamien if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1912156321Sdamien continue; 1913156321Sdamien if (bp->b_vp) { 1914178354Ssam BO_LOCK(bp->b_bufobj); 1915178354Ssam if (bp->b_vflags & BV_BKGRDINPROG) { 1916156321Sdamien BO_UNLOCK(bp->b_bufobj); 1917156321Sdamien BUF_UNLOCK(bp); 1918220502Sbschmidt continue; 1919156321Sdamien } 1920156321Sdamien BO_UNLOCK(bp->b_bufobj); 1921156321Sdamien } 1922156321Sdamien CTR6(KTR_BUF, 1923156321Sdamien "getnewbuf(%p) vp %p flags %X kvasize %d bufsize %d " 1924156321Sdamien "queue %d (recycling)", bp, bp->b_vp, bp->b_flags, 1925156321Sdamien bp->b_kvasize, bp->b_bufsize, qindex); 1926252727Sadrian 1927156321Sdamien /* 1928156321Sdamien * Sanity Checks 1929156321Sdamien */ 1930156321Sdamien KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1931178354Ssam 1932156321Sdamien /* 1933156321Sdamien * Note: we no longer distinguish between VMIO and non-VMIO 1934156321Sdamien * buffers. 1935156321Sdamien */ 1936156321Sdamien 1937156321Sdamien KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1938156321Sdamien 1939156321Sdamien if (bp->b_bufobj != NULL) 1940156321Sdamien BO_LOCK(bp->b_bufobj); 1941156321Sdamien bremfreel(bp); 1942156321Sdamien if (bp->b_bufobj != NULL) 1943156321Sdamien BO_UNLOCK(bp->b_bufobj); 1944156321Sdamien mtx_unlock(&bqlock); 1945156321Sdamien 1946156321Sdamien if (qindex == QUEUE_CLEAN) { 1947156321Sdamien if (bp->b_flags & B_VMIO) { 1948156321Sdamien bp->b_flags &= ~B_ASYNC; 1949156321Sdamien vfs_vmio_release(bp); 1950156321Sdamien } 1951156321Sdamien if (bp->b_vp) 1952156321Sdamien brelvp(bp); 1953156321Sdamien } 1954156321Sdamien 1955156321Sdamien /* 1956156321Sdamien * NOTE: nbp is now entirely invalid. We can only restart 1957156321Sdamien * the scan from this point on. 1958156321Sdamien * 1959156321Sdamien * Get the rest of the buffer freed up. b_kva* is still 1960156321Sdamien * valid after this operation. 1961156321Sdamien */ 1962156321Sdamien 1963156321Sdamien if (bp->b_rcred != NOCRED) { 1964156321Sdamien crfree(bp->b_rcred); 1965156321Sdamien bp->b_rcred = NOCRED; 1966156321Sdamien } 1967156321Sdamien if (bp->b_wcred != NOCRED) { 1968156321Sdamien crfree(bp->b_wcred); 1969156321Sdamien bp->b_wcred = NOCRED; 1970156321Sdamien } 1971156321Sdamien if (!LIST_EMPTY(&bp->b_dep)) 1972156321Sdamien buf_deallocate(bp); 1973156321Sdamien if (bp->b_vflags & BV_BKGRDINPROG) 1974156321Sdamien panic("losing buffer 3"); 1975156321Sdamien KASSERT(bp->b_vp == NULL, 1976156321Sdamien ("bp: %p still has vnode %p. qindex: %d", 1977156321Sdamien bp, bp->b_vp, qindex)); 1978156321Sdamien KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1979156321Sdamien ("bp: %p still on a buffer list. xflags %X", 1980156321Sdamien bp, bp->b_xflags)); 1981156321Sdamien 1982156321Sdamien if (bp->b_bufsize) 1983156321Sdamien allocbuf(bp, 0); 1984178354Ssam 1985178354Ssam bp->b_flags = 0; 1986156321Sdamien bp->b_ioflags = 0; 1987156321Sdamien bp->b_xflags = 0; 1988156321Sdamien KASSERT((bp->b_vflags & BV_INFREECNT) == 0, 1989156321Sdamien ("buf %p still counted as free?", bp)); 1990156321Sdamien bp->b_vflags = 0; 1991156321Sdamien bp->b_vp = NULL; 1992178354Ssam bp->b_blkno = bp->b_lblkno = 0; 1993156321Sdamien bp->b_offset = NOOFFSET; 1994156321Sdamien bp->b_iodone = 0; 1995156321Sdamien bp->b_error = 0; 1996156321Sdamien bp->b_resid = 0; 1997156321Sdamien bp->b_bcount = 0; 1998156321Sdamien bp->b_npages = 0; 1999156321Sdamien bp->b_dirtyoff = bp->b_dirtyend = 0; 2000156321Sdamien bp->b_bufobj = NULL; 2001156321Sdamien bp->b_pin_count = 0; 2002156321Sdamien bp->b_fsprivate1 = NULL; 2003156321Sdamien bp->b_fsprivate2 = NULL; 2004156321Sdamien bp->b_fsprivate3 = NULL; 2005156321Sdamien 2006156321Sdamien LIST_INIT(&bp->b_dep); 2007156321Sdamien 2008156321Sdamien /* 2009156321Sdamien * If we are defragging then free the buffer. 2010156321Sdamien */ 2011156321Sdamien if (defrag) { 2012156321Sdamien bp->b_flags |= B_INVAL; 2013156321Sdamien bfreekva(bp); 2014156321Sdamien brelse(bp); 2015156321Sdamien defrag = 0; 2016156321Sdamien goto restart; 2017156321Sdamien } 2018156321Sdamien 2019156321Sdamien /* 2020156321Sdamien * Notify any waiters for the buffer lock about 2021156321Sdamien * identity change by freeing the buffer. 2022156321Sdamien */ 2023156321Sdamien if (qindex == QUEUE_CLEAN && BUF_LOCKWAITERS(bp)) { 2024156321Sdamien bp->b_flags |= B_INVAL; 2025156321Sdamien bfreekva(bp); 2026156321Sdamien brelse(bp); 2027156321Sdamien goto restart; 2028156321Sdamien } 2029156321Sdamien 2030156321Sdamien /* 2031156321Sdamien * If we are overcomitted then recover the buffer and its 2032156321Sdamien * KVM space. This occurs in rare situations when multiple 2033156321Sdamien * processes are blocked in getnewbuf() or allocbuf(). 2034156321Sdamien */ 2035156321Sdamien if (bufspace >= hibufspace) 2036156321Sdamien flushingbufs = 1; 2037156321Sdamien if (flushingbufs && bp->b_kvasize != 0) { 2038156321Sdamien bp->b_flags |= B_INVAL; 2039156321Sdamien bfreekva(bp); 2040156321Sdamien brelse(bp); 2041156321Sdamien goto restart; 2042156321Sdamien } 2043156321Sdamien if (bufspace < lobufspace) 2044156321Sdamien flushingbufs = 0; 2045156321Sdamien break; 2046156321Sdamien } 2047156321Sdamien 2048156321Sdamien /* 2049156321Sdamien * If we exhausted our list, sleep as appropriate. We may have to 2050156321Sdamien * wakeup various daemons and write out some dirty buffers. 2051156321Sdamien * 2052156321Sdamien * Generally we are sleeping due to insufficient buffer space. 2053156321Sdamien */ 2054156321Sdamien 2055156321Sdamien if (bp == NULL) { 2056156321Sdamien int flags, norunbuf; 2057156321Sdamien char *waitmsg; 2058156321Sdamien int fl; 2059156321Sdamien 2060156321Sdamien if (defrag) { 2061156321Sdamien flags = VFS_BIO_NEED_BUFSPACE; 2062156321Sdamien waitmsg = "nbufkv"; 2063156321Sdamien } else if (bufspace >= hibufspace) { 2064156321Sdamien waitmsg = "nbufbs"; 2065156321Sdamien flags = VFS_BIO_NEED_BUFSPACE; 2066156321Sdamien } else { 2067156321Sdamien waitmsg = "newbuf"; 2068156321Sdamien flags = VFS_BIO_NEED_ANY; 2069156321Sdamien } 2070156321Sdamien mtx_lock(&nblock); 2071156321Sdamien needsbuffer |= flags; 2072156321Sdamien mtx_unlock(&nblock); 2073156321Sdamien mtx_unlock(&bqlock); 2074156321Sdamien 2075156321Sdamien bd_speedup(); /* heeeelp */ 2076156321Sdamien if (gbflags & GB_NOWAIT_BD) 2077156321Sdamien return (NULL); 2078156321Sdamien 2079156321Sdamien mtx_lock(&nblock); 2080178354Ssam while (needsbuffer & flags) { 2081156321Sdamien if (vp != NULL && (td->td_pflags & TDP_BUFNEED) == 0) { 2082178354Ssam mtx_unlock(&nblock); 2083156321Sdamien /* 2084156321Sdamien * getblk() is called with a vnode 2085156321Sdamien * locked, and some majority of the 2086156321Sdamien * dirty buffers may as well belong to 2087156321Sdamien * the vnode. Flushing the buffers 2088156321Sdamien * there would make a progress that 2089156321Sdamien * cannot be achieved by the 2090156321Sdamien * buf_daemon, that cannot lock the 2091156321Sdamien * vnode. 2092156321Sdamien */ 2093178354Ssam norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 2094178354Ssam (td->td_pflags & TDP_NORUNNINGBUF); 2095156321Sdamien /* play bufdaemon */ 2096156321Sdamien td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 2097156321Sdamien fl = buf_do_flush(vp); 2098156321Sdamien td->td_pflags &= norunbuf; 2099156321Sdamien mtx_lock(&nblock); 2100156321Sdamien if (fl != 0) 2101156321Sdamien continue; 2102156321Sdamien if ((needsbuffer & flags) == 0) 2103156321Sdamien break; 2104156321Sdamien } 2105156321Sdamien if (msleep(&needsbuffer, &nblock, 2106156321Sdamien (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2107156321Sdamien mtx_unlock(&nblock); 2108156321Sdamien return (NULL); 2109156321Sdamien } 2110156321Sdamien } 2111156321Sdamien mtx_unlock(&nblock); 2112156321Sdamien } else { 2113156321Sdamien /* 2114156321Sdamien * We finally have a valid bp. We aren't quite out of the 2115156321Sdamien * woods, we still have to reserve kva space. In order 2116156321Sdamien * to keep fragmentation sane we only allocate kva in 2117156321Sdamien * BKVASIZE chunks. 2118156321Sdamien */ 2119156321Sdamien maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2120156321Sdamien 2121156321Sdamien if (maxsize != bp->b_kvasize) { 2122156321Sdamien vm_offset_t addr = 0; 2123156321Sdamien 2124156321Sdamien bfreekva(bp); 2125156321Sdamien 2126156321Sdamien vm_map_lock(buffer_map); 2127156321Sdamien if (vm_map_findspace(buffer_map, 2128156321Sdamien vm_map_min(buffer_map), maxsize, &addr)) { 2129156321Sdamien /* 2130156321Sdamien * Uh oh. Buffer map is to fragmented. We 2131156321Sdamien * must defragment the map. 2132156321Sdamien */ 2133156321Sdamien atomic_add_int(&bufdefragcnt, 1); 2134156321Sdamien vm_map_unlock(buffer_map); 2135156321Sdamien defrag = 1; 2136156321Sdamien bp->b_flags |= B_INVAL; 2137156321Sdamien brelse(bp); 2138156321Sdamien goto restart; 2139156321Sdamien } 2140156321Sdamien if (addr) { 2141156321Sdamien vm_map_insert(buffer_map, NULL, 0, 2142156321Sdamien addr, addr + maxsize, 2143156321Sdamien VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2144156321Sdamien 2145156321Sdamien bp->b_kvabase = (caddr_t) addr; 2146156321Sdamien bp->b_kvasize = maxsize; 2147178354Ssam atomic_add_long(&bufspace, bp->b_kvasize); 2148156321Sdamien atomic_add_int(&bufreusecnt, 1); 2149156321Sdamien } 2150156321Sdamien vm_map_unlock(buffer_map); 2151156321Sdamien } 2152156321Sdamien bp->b_saveaddr = bp->b_kvabase; 2153156321Sdamien bp->b_data = bp->b_saveaddr; 2154156321Sdamien } 2155156321Sdamien return(bp); 2156156321Sdamien} 2157156321Sdamien 2158156321Sdamien/* 2159156321Sdamien * buf_daemon: 2160156321Sdamien * 2161156321Sdamien * buffer flushing daemon. Buffers are normally flushed by the 2162156321Sdamien * update daemon but if it cannot keep up this process starts to 2163156321Sdamien * take the load in an attempt to prevent getnewbuf() from blocking. 2164156321Sdamien */ 2165156321Sdamien 2166156321Sdamienstatic struct kproc_desc buf_kp = { 2167156321Sdamien "bufdaemon", 2168156321Sdamien buf_daemon, 2169156321Sdamien &bufdaemonproc 2170156321Sdamien}; 2171190526SsamSYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 2172156321Sdamien 2173156321Sdamienstatic int 2174156321Sdamienbuf_do_flush(struct vnode *vp) 2175156321Sdamien{ 2176156321Sdamien int flushed; 2177156321Sdamien 2178190526Ssam flushed = flushbufqueues(vp, QUEUE_DIRTY, 0); 2179190526Ssam /* The list empty check here is slightly racy */ 2180156321Sdamien if (!TAILQ_EMPTY(&bufqueues[QUEUE_DIRTY_GIANT])) { 2181156321Sdamien mtx_lock(&Giant); 2182190526Ssam flushed += flushbufqueues(vp, QUEUE_DIRTY_GIANT, 0); 2183190526Ssam mtx_unlock(&Giant); 2184156321Sdamien } 2185156321Sdamien if (flushed == 0) { 2186190526Ssam /* 2187190526Ssam * Could not find any buffers without rollback 2188156321Sdamien * dependencies, so just write the first one 2189156321Sdamien * in the hopes of eventually making progress. 2190156321Sdamien */ 2191156321Sdamien flushbufqueues(vp, QUEUE_DIRTY, 1); 2192156321Sdamien if (!TAILQ_EMPTY( 2193156321Sdamien &bufqueues[QUEUE_DIRTY_GIANT])) { 2194156321Sdamien mtx_lock(&Giant); 2195156321Sdamien flushbufqueues(vp, QUEUE_DIRTY_GIANT, 1); 2196156321Sdamien mtx_unlock(&Giant); 2197178354Ssam } 2198156321Sdamien } 2199156321Sdamien return (flushed); 2200156321Sdamien} 2201156321Sdamien 2202156321Sdamienstatic void 2203178354Ssambuf_daemon() 2204178354Ssam{ 2205156321Sdamien int lodirtysave; 2206156321Sdamien 2207156321Sdamien /* 2208156321Sdamien * This process needs to be suspended prior to shutdown sync. 2209156321Sdamien */ 2210170530Ssam EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2211170530Ssam SHUTDOWN_PRI_LAST); 2212170530Ssam 2213170530Ssam /* 2214156321Sdamien * This process is allowed to take the buffer cache to the limit 2215156321Sdamien */ 2216156321Sdamien curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 2217156321Sdamien mtx_lock(&bdlock); 2218170530Ssam for (;;) { 2219170530Ssam bd_request = 0; 2220170530Ssam mtx_unlock(&bdlock); 2221170530Ssam 2222156321Sdamien kproc_suspend_check(bufdaemonproc); 2223156321Sdamien lodirtysave = lodirtybuffers; 2224156321Sdamien if (bd_speedupreq) { 2225156321Sdamien lodirtybuffers = numdirtybuffers / 2; 2226156321Sdamien bd_speedupreq = 0; 2227156321Sdamien } 2228178354Ssam /* 2229178354Ssam * Do the flush. Limit the amount of in-transit I/O we 2230156321Sdamien * allow to build up, otherwise we would completely saturate 2231156321Sdamien * the I/O system. Wakeup any waiting processes before we 2232156321Sdamien * normally would so they can run in parallel with our drain. 2233156321Sdamien */ 2234156321Sdamien while (numdirtybuffers > lodirtybuffers) { 2235156321Sdamien if (buf_do_flush(NULL) == 0) 2236156321Sdamien break; 2237178354Ssam kern_yield(PRI_UNCHANGED); 2238156321Sdamien } 2239156321Sdamien lodirtybuffers = lodirtysave; 2240156321Sdamien 2241156321Sdamien /* 2242156321Sdamien * Only clear bd_request if we have reached our low water 2243178354Ssam * mark. The buf_daemon normally waits 1 second and 2244178354Ssam * then incrementally flushes any dirty buffers that have 2245156321Sdamien * built up, within reason. 2246178354Ssam * 2247178354Ssam * If we were unable to hit our low water mark and couldn't 2248156321Sdamien * find any flushable buffers, we sleep half a second. 2249156321Sdamien * Otherwise we loop immediately. 2250156321Sdamien */ 2251156321Sdamien mtx_lock(&bdlock); 2252156321Sdamien if (numdirtybuffers <= lodirtybuffers) { 2253156321Sdamien /* 2254156321Sdamien * We reached our low water mark, reset the 2255156321Sdamien * request and sleep until we are needed again. 2256156321Sdamien * The sleep is just so the suspend code works. 2257178354Ssam */ 2258178354Ssam bd_request = 0; 2259156321Sdamien msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2260156321Sdamien } else { 2261156321Sdamien /* 2262156321Sdamien * We couldn't find any flushable dirty buffers but 2263156321Sdamien * still have too many dirty buffers, we 2264156321Sdamien * have to sleep and try again. (rare) 2265156321Sdamien */ 2266156321Sdamien msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2267156321Sdamien } 2268156321Sdamien } 2269156321Sdamien} 2270156321Sdamien 2271156321Sdamien/* 2272156321Sdamien * flushbufqueues: 2273156321Sdamien * 2274156321Sdamien * Try to flush a buffer in the dirty queue. We must be careful to 2275156321Sdamien * free up B_INVAL buffers instead of write them, which NFS is 2276156321Sdamien * particularly sensitive to. 2277156321Sdamien */ 2278156321Sdamienstatic int flushwithdeps = 0; 2279156321SdamienSYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2280156321Sdamien 0, "Number of buffers flushed with dependecies that require rollbacks"); 2281156321Sdamien 2282156321Sdamienstatic int 2283156321Sdamienflushbufqueues(struct vnode *lvp, int queue, int flushdeps) 2284156321Sdamien{ 2285156321Sdamien struct buf *sentinel; 2286156321Sdamien struct vnode *vp; 2287156321Sdamien struct mount *mp; 2288156321Sdamien struct buf *bp; 2289156321Sdamien int hasdeps; 2290156321Sdamien int flushed; 2291156321Sdamien int target; 2292156321Sdamien 2293156321Sdamien if (lvp == NULL) { 2294156321Sdamien target = numdirtybuffers - lodirtybuffers; 2295156321Sdamien if (flushdeps && target > 2) 2296156321Sdamien target /= 2; 2297156321Sdamien } else 2298156321Sdamien target = flushbufqtarget; 2299178354Ssam flushed = 0; 2300156321Sdamien bp = NULL; 2301156321Sdamien sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 2302178354Ssam sentinel->b_qindex = QUEUE_SENTINEL; 2303178354Ssam mtx_lock(&bqlock); 2304156321Sdamien TAILQ_INSERT_HEAD(&bufqueues[queue], sentinel, b_freelist); 2305178354Ssam while (flushed != target) { 2306156321Sdamien bp = TAILQ_NEXT(sentinel, b_freelist); 2307178354Ssam if (bp != NULL) { 2308156975Sdamien TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2309178354Ssam TAILQ_INSERT_AFTER(&bufqueues[queue], bp, sentinel, 2310178354Ssam b_freelist); 2311178354Ssam } else 2312178354Ssam break; 2313178354Ssam /* 2314178354Ssam * Skip sentinels inserted by other invocations of the 2315178354Ssam * flushbufqueues(), taking care to not reorder them. 2316178354Ssam */ 2317178354Ssam if (bp->b_qindex == QUEUE_SENTINEL) 2318178354Ssam continue; 2319178354Ssam /* 2320170530Ssam * Only flush the buffers that belong to the 2321156321Sdamien * vnode locked by the curthread. 2322156321Sdamien */ 2323156321Sdamien if (lvp != NULL && bp->b_vp != lvp) 2324156321Sdamien continue; 2325156321Sdamien if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2326156321Sdamien continue; 2327156321Sdamien if (bp->b_pin_count > 0) { 2328156321Sdamien BUF_UNLOCK(bp); 2329156321Sdamien continue; 2330156321Sdamien } 2331156321Sdamien BO_LOCK(bp->b_bufobj); 2332156321Sdamien if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2333156321Sdamien (bp->b_flags & B_DELWRI) == 0) { 2334156321Sdamien BO_UNLOCK(bp->b_bufobj); 2335156321Sdamien BUF_UNLOCK(bp); 2336156321Sdamien continue; 2337156321Sdamien } 2338156321Sdamien BO_UNLOCK(bp->b_bufobj); 2339156321Sdamien if (bp->b_flags & B_INVAL) { 2340156321Sdamien bremfreel(bp); 2341156321Sdamien mtx_unlock(&bqlock); 2342156321Sdamien brelse(bp); 2343156321Sdamien flushed++; 2344156321Sdamien numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2345156321Sdamien mtx_lock(&bqlock); 2346156321Sdamien continue; 2347156321Sdamien } 2348156321Sdamien 2349156321Sdamien if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 2350156321Sdamien if (flushdeps == 0) { 2351156321Sdamien BUF_UNLOCK(bp); 2352156321Sdamien continue; 2353156321Sdamien } 2354156321Sdamien hasdeps = 1; 2355156321Sdamien } else 2356156321Sdamien hasdeps = 0; 2357156321Sdamien /* 2358156321Sdamien * We must hold the lock on a vnode before writing 2359156321Sdamien * one of its buffers. Otherwise we may confuse, or 2360156321Sdamien * in the case of a snapshot vnode, deadlock the 2361156321Sdamien * system. 2362156321Sdamien * 2363156321Sdamien * The lock order here is the reverse of the normal 2364156321Sdamien * of vnode followed by buf lock. This is ok because 2365190526Ssam * the NOWAIT will prevent deadlock. 2366156321Sdamien */ 2367156321Sdamien vp = bp->b_vp; 2368156321Sdamien if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2369156321Sdamien BUF_UNLOCK(bp); 2370156321Sdamien continue; 2371156321Sdamien } 2372156321Sdamien if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_CANRECURSE) == 0) { 2373156321Sdamien mtx_unlock(&bqlock); 2374156321Sdamien CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2375156321Sdamien bp, bp->b_vp, bp->b_flags); 2376156321Sdamien if (curproc == bufdaemonproc) 2377156321Sdamien vfs_bio_awrite(bp); 2378156321Sdamien else { 2379170530Ssam bremfree(bp); 2380156321Sdamien bwrite(bp); 2381156321Sdamien notbufdflashes++; 2382156321Sdamien } 2383156321Sdamien vn_finished_write(mp); 2384170530Ssam VOP_UNLOCK(vp, 0); 2385156321Sdamien flushwithdeps += hasdeps; 2386156321Sdamien flushed++; 2387156321Sdamien 2388156321Sdamien /* 2389156321Sdamien * Sleeping on runningbufspace while holding 2390156321Sdamien * vnode lock leads to deadlock. 2391156321Sdamien */ 2392156321Sdamien if (curproc == bufdaemonproc) 2393156321Sdamien waitrunningbufspace(); 2394156321Sdamien numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2395156321Sdamien mtx_lock(&bqlock); 2396156321Sdamien continue; 2397156321Sdamien } 2398156321Sdamien vn_finished_write(mp); 2399156321Sdamien BUF_UNLOCK(bp); 2400156321Sdamien } 2401195618Srpaulo TAILQ_REMOVE(&bufqueues[queue], sentinel, b_freelist); 2402195618Srpaulo mtx_unlock(&bqlock); 2403156321Sdamien free(sentinel, M_TEMP); 2404156321Sdamien return (flushed); 2405156321Sdamien} 2406156321Sdamien 2407156321Sdamien/* 2408156321Sdamien * Check to see if a block is currently memory resident. 2409156321Sdamien */ 2410156321Sdamienstruct buf * 2411156321Sdamienincore(struct bufobj *bo, daddr_t blkno) 2412156321Sdamien{ 2413156321Sdamien struct buf *bp; 2414156321Sdamien 2415156321Sdamien BO_LOCK(bo); 2416156321Sdamien bp = gbincore(bo, blkno); 2417156321Sdamien BO_UNLOCK(bo); 2418156321Sdamien return (bp); 2419156321Sdamien} 2420156321Sdamien 2421156321Sdamien/* 2422156321Sdamien * Returns true if no I/O is needed to access the 2423156321Sdamien * associated VM object. This is like incore except 2424156321Sdamien * it also hunts around in the VM system for the data. 2425156321Sdamien */ 2426156321Sdamien 2427156321Sdamienstatic int 2428156321Sdamieninmem(struct vnode * vp, daddr_t blkno) 2429178354Ssam{ 2430156975Sdamien vm_object_t obj; 2431156321Sdamien vm_offset_t toff, tinc, size; 2432156321Sdamien vm_page_t m; 2433178354Ssam vm_ooffset_t off; 2434178354Ssam 2435156321Sdamien ASSERT_VOP_LOCKED(vp, "inmem"); 2436156321Sdamien 2437178354Ssam if (incore(&vp->v_bufobj, blkno)) 2438178354Ssam return 1; 2439170530Ssam if (vp->v_mount == NULL) 2440170530Ssam return 0; 2441178354Ssam obj = vp->v_object; 2442170530Ssam if (obj == NULL) 2443178354Ssam return (0); 2444178931Sthompsa 2445178931Sthompsa size = PAGE_SIZE; 2446170530Ssam if (size > vp->v_mount->mnt_stat.f_iosize) 2447170530Ssam size = vp->v_mount->mnt_stat.f_iosize; 2448170530Ssam off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2449170530Ssam 2450170530Ssam VM_OBJECT_LOCK(obj); 2451178354Ssam for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2452156321Sdamien m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2453170530Ssam if (!m) 2454156321Sdamien goto notinmem; 2455178354Ssam tinc = size; 2456170530Ssam if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2457156321Sdamien tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2458178354Ssam if (vm_page_is_valid(m, 2459178354Ssam (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2460178354Ssam goto notinmem; 2461170530Ssam } 2462170530Ssam VM_OBJECT_UNLOCK(obj); 2463178354Ssam return 1; 2464170530Ssam 2465170530Ssamnotinmem: 2466170530Ssam VM_OBJECT_UNLOCK(obj); 2467170530Ssam return (0); 2468170530Ssam} 2469170530Ssam 2470170530Ssam/* 2471170530Ssam * Set the dirty range for a buffer based on the status of the dirty 2472170530Ssam * bits in the pages comprising the buffer. The range is limited 2473170530Ssam * to the size of the buffer. 2474170530Ssam * 2475170530Ssam * Tell the VM system that the pages associated with this buffer 2476170530Ssam * are clean. This is used for delayed writes where the data is 2477170530Ssam * going to go to disk eventually without additional VM intevention. 2478170530Ssam * 2479170530Ssam * Note that while we only really need to clean through to b_bcount, we 2480170530Ssam * just go ahead and clean through to b_bufsize. 2481170530Ssam */ 2482170530Ssamstatic void 2483170530Ssamvfs_clean_pages_dirty_buf(struct buf *bp) 2484170530Ssam{ 2485170530Ssam vm_ooffset_t foff, noff, eoff; 2486170530Ssam vm_page_t m; 2487170530Ssam int i; 2488170530Ssam 2489170530Ssam if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 2490170530Ssam return; 2491156321Sdamien 2492156321Sdamien foff = bp->b_offset; 2493178354Ssam KASSERT(bp->b_offset != NOOFFSET, 2494178354Ssam ("vfs_clean_pages_dirty_buf: no buffer offset")); 2495178354Ssam 2496178354Ssam VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2497178354Ssam vfs_drain_busy_pages(bp); 2498178354Ssam vfs_setdirty_locked_object(bp); 2499178354Ssam for (i = 0; i < bp->b_npages; i++) { 2500178354Ssam noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2501178354Ssam eoff = noff; 2502178354Ssam if (eoff > bp->b_offset + bp->b_bufsize) 2503156321Sdamien eoff = bp->b_offset + bp->b_bufsize; 2504178354Ssam m = bp->b_pages[i]; 2505156321Sdamien vfs_page_set_validclean(bp, foff, m); 2506178354Ssam /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 2507178354Ssam foff = noff; 2508178354Ssam } 2509178354Ssam VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2510156321Sdamien} 2511178354Ssam 2512178354Ssamstatic void 2513178354Ssamvfs_setdirty_locked_object(struct buf *bp) 2514178354Ssam{ 2515178354Ssam vm_object_t object; 2516178354Ssam int i; 2517178354Ssam 2518178354Ssam object = bp->b_bufobj->bo_object; 2519178354Ssam VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2520178354Ssam 2521178354Ssam /* 2522178354Ssam * We qualify the scan for modified pages on whether the 2523178354Ssam * object has been flushed yet. 2524178354Ssam */ 2525178354Ssam if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 2526178354Ssam vm_offset_t boffset; 2527178354Ssam vm_offset_t eoffset; 2528178354Ssam 2529178354Ssam /* 2530178354Ssam * test the pages to see if they have been modified directly 2531178354Ssam * by users through the VM system. 2532178354Ssam */ 2533178354Ssam for (i = 0; i < bp->b_npages; i++) 2534178354Ssam vm_page_test_dirty(bp->b_pages[i]); 2535156321Sdamien 2536156321Sdamien /* 2537156321Sdamien * Calculate the encompassing dirty range, boffset and eoffset, 2538156321Sdamien * (eoffset - boffset) bytes. 2539156321Sdamien */ 2540156321Sdamien 2541156321Sdamien for (i = 0; i < bp->b_npages; i++) { 2542156321Sdamien if (bp->b_pages[i]->dirty) 2543156321Sdamien break; 2544156321Sdamien } 2545178354Ssam boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2546156321Sdamien 2547156321Sdamien for (i = bp->b_npages - 1; i >= 0; --i) { 2548156321Sdamien if (bp->b_pages[i]->dirty) { 2549156321Sdamien break; 2550156321Sdamien } 2551156321Sdamien } 2552156321Sdamien eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2553156321Sdamien 2554156321Sdamien /* 2555156321Sdamien * Fit it to the buffer. 2556156321Sdamien */ 2557156321Sdamien 2558178354Ssam if (eoffset > bp->b_bcount) 2559178354Ssam eoffset = bp->b_bcount; 2560178354Ssam 2561178354Ssam /* 2562178354Ssam * If we have a good dirty range, merge with the existing 2563178354Ssam * dirty range. 2564178354Ssam */ 2565178354Ssam 2566156321Sdamien if (boffset < eoffset) { 2567156321Sdamien if (bp->b_dirtyoff > boffset) 2568156321Sdamien bp->b_dirtyoff = boffset; 2569156321Sdamien if (bp->b_dirtyend < eoffset) 2570156321Sdamien bp->b_dirtyend = eoffset; 2571156321Sdamien } 2572156321Sdamien } 2573156321Sdamien} 2574156321Sdamien 2575156321Sdamien/* 2576156321Sdamien * getblk: 2577156321Sdamien * 2578156321Sdamien * Get a block given a specified block and offset into a file/device. 2579156321Sdamien * The buffers B_DONE bit will be cleared on return, making it almost 2580156321Sdamien * ready for an I/O initiation. B_INVAL may or may not be set on 2581156321Sdamien * return. The caller should clear B_INVAL prior to initiating a 2582156321Sdamien * READ. 2583156321Sdamien * 2584156321Sdamien * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2585156321Sdamien * an existing buffer. 2586156321Sdamien * 2587156321Sdamien * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2588156321Sdamien * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2589156321Sdamien * and then cleared based on the backing VM. If the previous buffer is 2590156321Sdamien * non-0-sized but invalid, B_CACHE will be cleared. 2591156321Sdamien * 2592156321Sdamien * If getblk() must create a new buffer, the new buffer is returned with 2593156321Sdamien * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2594156321Sdamien * case it is returned with B_INVAL clear and B_CACHE set based on the 2595156321Sdamien * backing VM. 2596156321Sdamien * 2597156321Sdamien * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2598156321Sdamien * B_CACHE bit is clear. 2599156321Sdamien * 2600156321Sdamien * What this means, basically, is that the caller should use B_CACHE to 2601156321Sdamien * determine whether the buffer is fully valid or not and should clear 2602156321Sdamien * B_INVAL prior to issuing a read. If the caller intends to validate 2603156321Sdamien * the buffer by loading its data area with something, the caller needs 2604156321Sdamien * to clear B_INVAL. If the caller does this without issuing an I/O, 2605156321Sdamien * the caller should set B_CACHE ( as an optimization ), else the caller 2606156321Sdamien * should issue the I/O and biodone() will set B_CACHE if the I/O was 2607156321Sdamien * a write attempt or if it was a successfull read. If the caller 2608156321Sdamien * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2609156321Sdamien * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2610156321Sdamien */ 2611156321Sdamienstruct buf * 2612156321Sdamiengetblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2613156321Sdamien int flags) 2614156321Sdamien{ 2615156321Sdamien struct buf *bp; 2616156321Sdamien struct bufobj *bo; 2617156321Sdamien int error; 2618156321Sdamien 2619156321Sdamien CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2620156321Sdamien ASSERT_VOP_LOCKED(vp, "getblk"); 2621156321Sdamien if (size > MAXBSIZE) 2622156321Sdamien panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2623156321Sdamien 2624156321Sdamien bo = &vp->v_bufobj; 2625156321Sdamienloop: 2626156321Sdamien /* 2627156321Sdamien * Block if we are low on buffers. Certain processes are allowed 2628156321Sdamien * to completely exhaust the buffer cache. 2629156321Sdamien * 2630156321Sdamien * If this check ever becomes a bottleneck it may be better to 2631156321Sdamien * move it into the else, when gbincore() fails. At the moment 2632156321Sdamien * it isn't a problem. 2633156321Sdamien * 2634156321Sdamien * XXX remove if 0 sections (clean this up after its proven) 2635156321Sdamien */ 2636156321Sdamien if (numfreebuffers == 0) { 2637156321Sdamien if (TD_IS_IDLETHREAD(curthread)) 2638156321Sdamien return NULL; 2639156321Sdamien mtx_lock(&nblock); 2640156321Sdamien needsbuffer |= VFS_BIO_NEED_ANY; 2641156321Sdamien mtx_unlock(&nblock); 2642156321Sdamien } 2643156321Sdamien 2644156321Sdamien BO_LOCK(bo); 2645156321Sdamien bp = gbincore(bo, blkno); 2646156321Sdamien if (bp != NULL) { 2647156321Sdamien int lockflags; 2648156321Sdamien /* 2649156321Sdamien * Buffer is in-core. If the buffer is not busy, it must 2650156321Sdamien * be on a queue. 2651156321Sdamien */ 2652156321Sdamien lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2653156321Sdamien 2654156321Sdamien if (flags & GB_LOCK_NOWAIT) 2655156321Sdamien lockflags |= LK_NOWAIT; 2656156321Sdamien 2657156321Sdamien error = BUF_TIMELOCK(bp, lockflags, 2658156321Sdamien BO_MTX(bo), "getblk", slpflag, slptimeo); 2659156321Sdamien 2660156321Sdamien /* 2661156321Sdamien * If we slept and got the lock we have to restart in case 2662156321Sdamien * the buffer changed identities. 2663156321Sdamien */ 2664156321Sdamien if (error == ENOLCK) 2665156321Sdamien goto loop; 2666156321Sdamien /* We timed out or were interrupted. */ 2667156321Sdamien else if (error) 2668156321Sdamien return (NULL); 2669156321Sdamien 2670156321Sdamien /* 2671156321Sdamien * The buffer is locked. B_CACHE is cleared if the buffer is 2672156321Sdamien * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2673156321Sdamien * and for a VMIO buffer B_CACHE is adjusted according to the 2674156321Sdamien * backing VM cache. 2675156321Sdamien */ 2676156321Sdamien if (bp->b_flags & B_INVAL) 2677156321Sdamien bp->b_flags &= ~B_CACHE; 2678156321Sdamien else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2679156321Sdamien bp->b_flags |= B_CACHE; 2680156321Sdamien BO_LOCK(bo); 2681156321Sdamien bremfree(bp); 2682156321Sdamien BO_UNLOCK(bo); 2683156321Sdamien 2684156321Sdamien /* 2685156321Sdamien * check for size inconsistancies for non-VMIO case. 2686156321Sdamien */ 2687156321Sdamien 2688178354Ssam if (bp->b_bcount != size) { 2689156321Sdamien if ((bp->b_flags & B_VMIO) == 0 || 2690178354Ssam (size > bp->b_kvasize)) { 2691156321Sdamien if (bp->b_flags & B_DELWRI) { 2692156321Sdamien /* 2693156321Sdamien * If buffer is pinned and caller does 2694156321Sdamien * not want sleep waiting for it to be 2695156321Sdamien * unpinned, bail out 2696178354Ssam * */ 2697156321Sdamien if (bp->b_pin_count > 0) { 2698156321Sdamien if (flags & GB_LOCK_NOWAIT) { 2699156321Sdamien bqrelse(bp); 2700156321Sdamien return (NULL); 2701156321Sdamien } else { 2702156321Sdamien bunpin_wait(bp); 2703178354Ssam } 2704156321Sdamien } 2705156321Sdamien bp->b_flags |= B_NOCACHE; 2706156321Sdamien bwrite(bp); 2707156321Sdamien } else { 2708156321Sdamien if (LIST_EMPTY(&bp->b_dep)) { 2709156321Sdamien bp->b_flags |= B_RELBUF; 2710156321Sdamien brelse(bp); 2711156321Sdamien } else { 2712156321Sdamien bp->b_flags |= B_NOCACHE; 2713156321Sdamien bwrite(bp); 2714156321Sdamien } 2715156321Sdamien } 2716156321Sdamien goto loop; 2717156321Sdamien } 2718156321Sdamien } 2719156321Sdamien 2720156321Sdamien /* 2721156321Sdamien * If the size is inconsistant in the VMIO case, we can resize 2722156321Sdamien * the buffer. This might lead to B_CACHE getting set or 2723156321Sdamien * cleared. If the size has not changed, B_CACHE remains 2724156321Sdamien * unchanged from its previous state. 2725156321Sdamien */ 2726156321Sdamien 2727178354Ssam if (bp->b_bcount != size) 2728178354Ssam allocbuf(bp, size); 2729178354Ssam 2730156321Sdamien KASSERT(bp->b_offset != NOOFFSET, 2731156321Sdamien ("getblk: no buffer offset")); 2732178354Ssam 2733156321Sdamien /* 2734156321Sdamien * A buffer with B_DELWRI set and B_CACHE clear must 2735156321Sdamien * be committed before we can return the buffer in 2736156321Sdamien * order to prevent the caller from issuing a read 2737156321Sdamien * ( due to B_CACHE not being set ) and overwriting 2738156321Sdamien * it. 2739156321Sdamien * 2740156321Sdamien * Most callers, including NFS and FFS, need this to 2741156321Sdamien * operate properly either because they assume they 2742156321Sdamien * can issue a read if B_CACHE is not set, or because 2743178354Ssam * ( for example ) an uncached B_DELWRI might loop due 2744156321Sdamien * to softupdates re-dirtying the buffer. In the latter 2745156321Sdamien * case, B_CACHE is set after the first write completes, 2746178354Ssam * preventing further loops. 2747156321Sdamien * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2748156321Sdamien * above while extending the buffer, we cannot allow the 2749156321Sdamien * buffer to remain with B_CACHE set after the write 2750156321Sdamien * completes or it will represent a corrupt state. To 2751156321Sdamien * deal with this we set B_NOCACHE to scrap the buffer 2752156321Sdamien * after the write. 2753156321Sdamien * 2754192468Ssam * We might be able to do something fancy, like setting 2755192468Ssam * B_CACHE in bwrite() except if B_DELWRI is already set, 2756192468Ssam * so the below call doesn't set B_CACHE, but that gets real 2757192468Ssam * confusing. This is much easier. 2758192468Ssam */ 2759192468Ssam 2760192468Ssam if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2761192468Ssam bp->b_flags |= B_NOCACHE; 2762156321Sdamien bwrite(bp); 2763156321Sdamien goto loop; 2764156321Sdamien } 2765156321Sdamien bp->b_flags &= ~B_DONE; 2766156321Sdamien } else { 2767156321Sdamien int bsize, maxsize, vmio; 2768156321Sdamien off_t offset; 2769156321Sdamien 2770156321Sdamien /* 2771156321Sdamien * Buffer is not in-core, create new buffer. The buffer 2772156321Sdamien * returned by getnewbuf() is locked. Note that the returned 2773156321Sdamien * buffer is also considered valid (not marked B_INVAL). 2774156321Sdamien */ 2775170530Ssam BO_UNLOCK(bo); 2776170530Ssam /* 2777170530Ssam * If the user does not want us to create the buffer, bail out 2778170530Ssam * here. 2779170530Ssam */ 2780170530Ssam if (flags & GB_NOCREAT) 2781170530Ssam return NULL; 2782170530Ssam bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 2783170530Ssam offset = blkno * bsize; 2784156321Sdamien vmio = vp->v_object != NULL; 2785170530Ssam maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2786170530Ssam maxsize = imax(maxsize, bsize); 2787156321Sdamien 2788156321Sdamien bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize, flags); 2789156321Sdamien if (bp == NULL) { 2790156321Sdamien if (slpflag || slptimeo) 2791156321Sdamien return NULL; 2792156321Sdamien goto loop; 2793156321Sdamien } 2794156321Sdamien 2795156321Sdamien /* 2796156321Sdamien * This code is used to make sure that a buffer is not 2797156321Sdamien * created while the getnewbuf routine is blocked. 2798156321Sdamien * This can be a problem whether the vnode is locked or not. 2799156321Sdamien * If the buffer is created out from under us, we have to 2800156321Sdamien * throw away the one we just created. 2801156321Sdamien * 2802156321Sdamien * Note: this must occur before we associate the buffer 2803156321Sdamien * with the vp especially considering limitations in 2804156321Sdamien * the splay tree implementation when dealing with duplicate 2805156321Sdamien * lblkno's. 2806156321Sdamien */ 2807156321Sdamien BO_LOCK(bo); 2808170530Ssam if (gbincore(bo, blkno)) { 2809170530Ssam BO_UNLOCK(bo); 2810170530Ssam bp->b_flags |= B_INVAL; 2811170530Ssam brelse(bp); 2812170530Ssam goto loop; 2813170530Ssam } 2814170530Ssam 2815170530Ssam /* 2816170530Ssam * Insert the buffer into the hash, so that it can 2817170530Ssam * be found by incore. 2818170530Ssam */ 2819170530Ssam bp->b_blkno = bp->b_lblkno = blkno; 2820170530Ssam bp->b_offset = offset; 2821170530Ssam bgetvp(vp, bp); 2822170530Ssam BO_UNLOCK(bo); 2823170530Ssam 2824170530Ssam /* 2825170530Ssam * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2826170530Ssam * buffer size starts out as 0, B_CACHE will be set by 2827178354Ssam * allocbuf() for the VMIO case prior to it testing the 2828170530Ssam * backing store for validity. 2829170530Ssam */ 2830170530Ssam 2831178354Ssam if (vmio) { 2832170530Ssam bp->b_flags |= B_VMIO; 2833170530Ssam KASSERT(vp->v_object == bp->b_bufobj->bo_object, 2834170530Ssam ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2835170530Ssam bp, vp->v_object, bp->b_bufobj->bo_object)); 2836170530Ssam } else { 2837170530Ssam bp->b_flags &= ~B_VMIO; 2838170530Ssam KASSERT(bp->b_bufobj->bo_object == NULL, 2839170530Ssam ("ARGH! has b_bufobj->bo_object %p %p\n", 2840170530Ssam bp, bp->b_bufobj->bo_object)); 2841170530Ssam } 2842170530Ssam 2843170530Ssam allocbuf(bp, size); 2844170530Ssam bp->b_flags &= ~B_DONE; 2845 } 2846 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2847 BUF_ASSERT_HELD(bp); 2848 KASSERT(bp->b_bufobj == bo, 2849 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 2850 return (bp); 2851} 2852 2853/* 2854 * Get an empty, disassociated buffer of given size. The buffer is initially 2855 * set to B_INVAL. 2856 */ 2857struct buf * 2858geteblk(int size, int flags) 2859{ 2860 struct buf *bp; 2861 int maxsize; 2862 2863 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2864 while ((bp = getnewbuf(NULL, 0, 0, size, maxsize, flags)) == NULL) { 2865 if ((flags & GB_NOWAIT_BD) && 2866 (curthread->td_pflags & TDP_BUFNEED) != 0) 2867 return (NULL); 2868 } 2869 allocbuf(bp, size); 2870 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2871 BUF_ASSERT_HELD(bp); 2872 return (bp); 2873} 2874 2875 2876/* 2877 * This code constitutes the buffer memory from either anonymous system 2878 * memory (in the case of non-VMIO operations) or from an associated 2879 * VM object (in the case of VMIO operations). This code is able to 2880 * resize a buffer up or down. 2881 * 2882 * Note that this code is tricky, and has many complications to resolve 2883 * deadlock or inconsistant data situations. Tread lightly!!! 2884 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2885 * the caller. Calling this code willy nilly can result in the loss of data. 2886 * 2887 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2888 * B_CACHE for the non-VMIO case. 2889 */ 2890 2891int 2892allocbuf(struct buf *bp, int size) 2893{ 2894 int newbsize, mbsize; 2895 int i; 2896 2897 BUF_ASSERT_HELD(bp); 2898 2899 if (bp->b_kvasize < size) 2900 panic("allocbuf: buffer too small"); 2901 2902 if ((bp->b_flags & B_VMIO) == 0) { 2903 caddr_t origbuf; 2904 int origbufsize; 2905 /* 2906 * Just get anonymous memory from the kernel. Don't 2907 * mess with B_CACHE. 2908 */ 2909 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2910 if (bp->b_flags & B_MALLOC) 2911 newbsize = mbsize; 2912 else 2913 newbsize = round_page(size); 2914 2915 if (newbsize < bp->b_bufsize) { 2916 /* 2917 * malloced buffers are not shrunk 2918 */ 2919 if (bp->b_flags & B_MALLOC) { 2920 if (newbsize) { 2921 bp->b_bcount = size; 2922 } else { 2923 free(bp->b_data, M_BIOBUF); 2924 if (bp->b_bufsize) { 2925 atomic_subtract_long( 2926 &bufmallocspace, 2927 bp->b_bufsize); 2928 bufspacewakeup(); 2929 bp->b_bufsize = 0; 2930 } 2931 bp->b_saveaddr = bp->b_kvabase; 2932 bp->b_data = bp->b_saveaddr; 2933 bp->b_bcount = 0; 2934 bp->b_flags &= ~B_MALLOC; 2935 } 2936 return 1; 2937 } 2938 vm_hold_free_pages(bp, newbsize); 2939 } else if (newbsize > bp->b_bufsize) { 2940 /* 2941 * We only use malloced memory on the first allocation. 2942 * and revert to page-allocated memory when the buffer 2943 * grows. 2944 */ 2945 /* 2946 * There is a potential smp race here that could lead 2947 * to bufmallocspace slightly passing the max. It 2948 * is probably extremely rare and not worth worrying 2949 * over. 2950 */ 2951 if ( (bufmallocspace < maxbufmallocspace) && 2952 (bp->b_bufsize == 0) && 2953 (mbsize <= PAGE_SIZE/2)) { 2954 2955 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2956 bp->b_bufsize = mbsize; 2957 bp->b_bcount = size; 2958 bp->b_flags |= B_MALLOC; 2959 atomic_add_long(&bufmallocspace, mbsize); 2960 return 1; 2961 } 2962 origbuf = NULL; 2963 origbufsize = 0; 2964 /* 2965 * If the buffer is growing on its other-than-first allocation, 2966 * then we revert to the page-allocation scheme. 2967 */ 2968 if (bp->b_flags & B_MALLOC) { 2969 origbuf = bp->b_data; 2970 origbufsize = bp->b_bufsize; 2971 bp->b_data = bp->b_kvabase; 2972 if (bp->b_bufsize) { 2973 atomic_subtract_long(&bufmallocspace, 2974 bp->b_bufsize); 2975 bufspacewakeup(); 2976 bp->b_bufsize = 0; 2977 } 2978 bp->b_flags &= ~B_MALLOC; 2979 newbsize = round_page(newbsize); 2980 } 2981 vm_hold_load_pages( 2982 bp, 2983 (vm_offset_t) bp->b_data + bp->b_bufsize, 2984 (vm_offset_t) bp->b_data + newbsize); 2985 if (origbuf) { 2986 bcopy(origbuf, bp->b_data, origbufsize); 2987 free(origbuf, M_BIOBUF); 2988 } 2989 } 2990 } else { 2991 int desiredpages; 2992 2993 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2994 desiredpages = (size == 0) ? 0 : 2995 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2996 2997 if (bp->b_flags & B_MALLOC) 2998 panic("allocbuf: VMIO buffer can't be malloced"); 2999 /* 3000 * Set B_CACHE initially if buffer is 0 length or will become 3001 * 0-length. 3002 */ 3003 if (size == 0 || bp->b_bufsize == 0) 3004 bp->b_flags |= B_CACHE; 3005 3006 if (newbsize < bp->b_bufsize) { 3007 /* 3008 * DEV_BSIZE aligned new buffer size is less then the 3009 * DEV_BSIZE aligned existing buffer size. Figure out 3010 * if we have to remove any pages. 3011 */ 3012 if (desiredpages < bp->b_npages) { 3013 vm_page_t m; 3014 3015 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3016 for (i = desiredpages; i < bp->b_npages; i++) { 3017 /* 3018 * the page is not freed here -- it 3019 * is the responsibility of 3020 * vnode_pager_setsize 3021 */ 3022 m = bp->b_pages[i]; 3023 KASSERT(m != bogus_page, 3024 ("allocbuf: bogus page found")); 3025 while (vm_page_sleep_if_busy(m, TRUE, 3026 "biodep")) 3027 continue; 3028 3029 bp->b_pages[i] = NULL; 3030 vm_page_lock(m); 3031 vm_page_unwire(m, 0); 3032 vm_page_unlock(m); 3033 } 3034 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3035 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 3036 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 3037 bp->b_npages = desiredpages; 3038 } 3039 } else if (size > bp->b_bcount) { 3040 /* 3041 * We are growing the buffer, possibly in a 3042 * byte-granular fashion. 3043 */ 3044 vm_object_t obj; 3045 vm_offset_t toff; 3046 vm_offset_t tinc; 3047 3048 /* 3049 * Step 1, bring in the VM pages from the object, 3050 * allocating them if necessary. We must clear 3051 * B_CACHE if these pages are not valid for the 3052 * range covered by the buffer. 3053 */ 3054 3055 obj = bp->b_bufobj->bo_object; 3056 3057 VM_OBJECT_LOCK(obj); 3058 while (bp->b_npages < desiredpages) { 3059 vm_page_t m; 3060 3061 /* 3062 * We must allocate system pages since blocking 3063 * here could intefere with paging I/O, no 3064 * matter which process we are. 3065 * 3066 * We can only test VPO_BUSY here. Blocking on 3067 * m->busy might lead to a deadlock: 3068 * vm_fault->getpages->cluster_read->allocbuf 3069 * Thus, we specify VM_ALLOC_IGN_SBUSY. 3070 */ 3071 m = vm_page_grab(obj, OFF_TO_IDX(bp->b_offset) + 3072 bp->b_npages, VM_ALLOC_NOBUSY | 3073 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3074 VM_ALLOC_RETRY | VM_ALLOC_IGN_SBUSY | 3075 VM_ALLOC_COUNT(desiredpages - bp->b_npages)); 3076 if (m->valid == 0) 3077 bp->b_flags &= ~B_CACHE; 3078 bp->b_pages[bp->b_npages] = m; 3079 ++bp->b_npages; 3080 } 3081 3082 /* 3083 * Step 2. We've loaded the pages into the buffer, 3084 * we have to figure out if we can still have B_CACHE 3085 * set. Note that B_CACHE is set according to the 3086 * byte-granular range ( bcount and size ), new the 3087 * aligned range ( newbsize ). 3088 * 3089 * The VM test is against m->valid, which is DEV_BSIZE 3090 * aligned. Needless to say, the validity of the data 3091 * needs to also be DEV_BSIZE aligned. Note that this 3092 * fails with NFS if the server or some other client 3093 * extends the file's EOF. If our buffer is resized, 3094 * B_CACHE may remain set! XXX 3095 */ 3096 3097 toff = bp->b_bcount; 3098 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3099 3100 while ((bp->b_flags & B_CACHE) && toff < size) { 3101 vm_pindex_t pi; 3102 3103 if (tinc > (size - toff)) 3104 tinc = size - toff; 3105 3106 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 3107 PAGE_SHIFT; 3108 3109 vfs_buf_test_cache( 3110 bp, 3111 bp->b_offset, 3112 toff, 3113 tinc, 3114 bp->b_pages[pi] 3115 ); 3116 toff += tinc; 3117 tinc = PAGE_SIZE; 3118 } 3119 VM_OBJECT_UNLOCK(obj); 3120 3121 /* 3122 * Step 3, fixup the KVM pmap. Remember that 3123 * bp->b_data is relative to bp->b_offset, but 3124 * bp->b_offset may be offset into the first page. 3125 */ 3126 3127 bp->b_data = (caddr_t) 3128 trunc_page((vm_offset_t)bp->b_data); 3129 pmap_qenter( 3130 (vm_offset_t)bp->b_data, 3131 bp->b_pages, 3132 bp->b_npages 3133 ); 3134 3135 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3136 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3137 } 3138 } 3139 if (newbsize < bp->b_bufsize) 3140 bufspacewakeup(); 3141 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3142 bp->b_bcount = size; /* requested buffer size */ 3143 return 1; 3144} 3145 3146void 3147biodone(struct bio *bp) 3148{ 3149 struct mtx *mtxp; 3150 void (*done)(struct bio *); 3151 3152 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3153 mtx_lock(mtxp); 3154 bp->bio_flags |= BIO_DONE; 3155 done = bp->bio_done; 3156 if (done == NULL) 3157 wakeup(bp); 3158 mtx_unlock(mtxp); 3159 if (done != NULL) 3160 done(bp); 3161} 3162 3163/* 3164 * Wait for a BIO to finish. 3165 * 3166 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3167 * case is not yet clear. 3168 */ 3169int 3170biowait(struct bio *bp, const char *wchan) 3171{ 3172 struct mtx *mtxp; 3173 3174 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3175 mtx_lock(mtxp); 3176 while ((bp->bio_flags & BIO_DONE) == 0) 3177 msleep(bp, mtxp, PRIBIO, wchan, hz / 10); 3178 mtx_unlock(mtxp); 3179 if (bp->bio_error != 0) 3180 return (bp->bio_error); 3181 if (!(bp->bio_flags & BIO_ERROR)) 3182 return (0); 3183 return (EIO); 3184} 3185 3186void 3187biofinish(struct bio *bp, struct devstat *stat, int error) 3188{ 3189 3190 if (error) { 3191 bp->bio_error = error; 3192 bp->bio_flags |= BIO_ERROR; 3193 } 3194 if (stat != NULL) 3195 devstat_end_transaction_bio(stat, bp); 3196 biodone(bp); 3197} 3198 3199/* 3200 * bufwait: 3201 * 3202 * Wait for buffer I/O completion, returning error status. The buffer 3203 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3204 * error and cleared. 3205 */ 3206int 3207bufwait(struct buf *bp) 3208{ 3209 if (bp->b_iocmd == BIO_READ) 3210 bwait(bp, PRIBIO, "biord"); 3211 else 3212 bwait(bp, PRIBIO, "biowr"); 3213 if (bp->b_flags & B_EINTR) { 3214 bp->b_flags &= ~B_EINTR; 3215 return (EINTR); 3216 } 3217 if (bp->b_ioflags & BIO_ERROR) { 3218 return (bp->b_error ? bp->b_error : EIO); 3219 } else { 3220 return (0); 3221 } 3222} 3223 3224 /* 3225 * Call back function from struct bio back up to struct buf. 3226 */ 3227static void 3228bufdonebio(struct bio *bip) 3229{ 3230 struct buf *bp; 3231 3232 bp = bip->bio_caller2; 3233 bp->b_resid = bp->b_bcount - bip->bio_completed; 3234 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3235 bp->b_ioflags = bip->bio_flags; 3236 bp->b_error = bip->bio_error; 3237 if (bp->b_error) 3238 bp->b_ioflags |= BIO_ERROR; 3239 bufdone(bp); 3240 g_destroy_bio(bip); 3241} 3242 3243void 3244dev_strategy(struct cdev *dev, struct buf *bp) 3245{ 3246 struct cdevsw *csw; 3247 struct bio *bip; 3248 int ref; 3249 3250 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3251 panic("b_iocmd botch"); 3252 for (;;) { 3253 bip = g_new_bio(); 3254 if (bip != NULL) 3255 break; 3256 /* Try again later */ 3257 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3258 } 3259 bip->bio_cmd = bp->b_iocmd; 3260 bip->bio_offset = bp->b_iooffset; 3261 bip->bio_length = bp->b_bcount; 3262 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3263 bip->bio_data = bp->b_data; 3264 bip->bio_done = bufdonebio; 3265 bip->bio_caller2 = bp; 3266 bip->bio_dev = dev; 3267 KASSERT(dev->si_refcount > 0, 3268 ("dev_strategy on un-referenced struct cdev *(%s)", 3269 devtoname(dev))); 3270 csw = dev_refthread(dev, &ref); 3271 if (csw == NULL) { 3272 g_destroy_bio(bip); 3273 bp->b_error = ENXIO; 3274 bp->b_ioflags = BIO_ERROR; 3275 bufdone(bp); 3276 return; 3277 } 3278 (*csw->d_strategy)(bip); 3279 dev_relthread(dev, ref); 3280} 3281 3282/* 3283 * bufdone: 3284 * 3285 * Finish I/O on a buffer, optionally calling a completion function. 3286 * This is usually called from an interrupt so process blocking is 3287 * not allowed. 3288 * 3289 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3290 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3291 * assuming B_INVAL is clear. 3292 * 3293 * For the VMIO case, we set B_CACHE if the op was a read and no 3294 * read error occured, or if the op was a write. B_CACHE is never 3295 * set if the buffer is invalid or otherwise uncacheable. 3296 * 3297 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3298 * initiator to leave B_INVAL set to brelse the buffer out of existance 3299 * in the biodone routine. 3300 */ 3301void 3302bufdone(struct buf *bp) 3303{ 3304 struct bufobj *dropobj; 3305 void (*biodone)(struct buf *); 3306 3307 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3308 dropobj = NULL; 3309 3310 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3311 BUF_ASSERT_HELD(bp); 3312 3313 runningbufwakeup(bp); 3314 if (bp->b_iocmd == BIO_WRITE) 3315 dropobj = bp->b_bufobj; 3316 /* call optional completion function if requested */ 3317 if (bp->b_iodone != NULL) { 3318 biodone = bp->b_iodone; 3319 bp->b_iodone = NULL; 3320 (*biodone) (bp); 3321 if (dropobj) 3322 bufobj_wdrop(dropobj); 3323 return; 3324 } 3325 3326 bufdone_finish(bp); 3327 3328 if (dropobj) 3329 bufobj_wdrop(dropobj); 3330} 3331 3332void 3333bufdone_finish(struct buf *bp) 3334{ 3335 BUF_ASSERT_HELD(bp); 3336 3337 if (!LIST_EMPTY(&bp->b_dep)) 3338 buf_complete(bp); 3339 3340 if (bp->b_flags & B_VMIO) { 3341 vm_ooffset_t foff; 3342 vm_page_t m; 3343 vm_object_t obj; 3344 struct vnode *vp; 3345 int bogus, i, iosize; 3346 3347 obj = bp->b_bufobj->bo_object; 3348 KASSERT(obj->paging_in_progress >= bp->b_npages, 3349 ("biodone_finish: paging in progress(%d) < b_npages(%d)", 3350 obj->paging_in_progress, bp->b_npages)); 3351 3352 vp = bp->b_vp; 3353 KASSERT(vp->v_holdcnt > 0, 3354 ("biodone_finish: vnode %p has zero hold count", vp)); 3355 KASSERT(vp->v_object != NULL, 3356 ("biodone_finish: vnode %p has no vm_object", vp)); 3357 3358 foff = bp->b_offset; 3359 KASSERT(bp->b_offset != NOOFFSET, 3360 ("biodone_finish: bp %p has no buffer offset", bp)); 3361 3362 /* 3363 * Set B_CACHE if the op was a normal read and no error 3364 * occured. B_CACHE is set for writes in the b*write() 3365 * routines. 3366 */ 3367 iosize = bp->b_bcount - bp->b_resid; 3368 if (bp->b_iocmd == BIO_READ && 3369 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3370 !(bp->b_ioflags & BIO_ERROR)) { 3371 bp->b_flags |= B_CACHE; 3372 } 3373 bogus = 0; 3374 VM_OBJECT_LOCK(obj); 3375 for (i = 0; i < bp->b_npages; i++) { 3376 int bogusflag = 0; 3377 int resid; 3378 3379 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3380 if (resid > iosize) 3381 resid = iosize; 3382 3383 /* 3384 * cleanup bogus pages, restoring the originals 3385 */ 3386 m = bp->b_pages[i]; 3387 if (m == bogus_page) { 3388 bogus = bogusflag = 1; 3389 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3390 if (m == NULL) 3391 panic("biodone: page disappeared!"); 3392 bp->b_pages[i] = m; 3393 } 3394 KASSERT(OFF_TO_IDX(foff) == m->pindex, 3395 ("biodone_finish: foff(%jd)/pindex(%ju) mismatch", 3396 (intmax_t)foff, (uintmax_t)m->pindex)); 3397 3398 /* 3399 * In the write case, the valid and clean bits are 3400 * already changed correctly ( see bdwrite() ), so we 3401 * only need to do this here in the read case. 3402 */ 3403 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3404 KASSERT((m->dirty & vm_page_bits(foff & 3405 PAGE_MASK, resid)) == 0, ("bufdone_finish:" 3406 " page %p has unexpected dirty bits", m)); 3407 vfs_page_set_valid(bp, foff, m); 3408 } 3409 3410 vm_page_io_finish(m); 3411 vm_object_pip_subtract(obj, 1); 3412 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3413 iosize -= resid; 3414 } 3415 vm_object_pip_wakeupn(obj, 0); 3416 VM_OBJECT_UNLOCK(obj); 3417 if (bogus) 3418 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3419 bp->b_pages, bp->b_npages); 3420 } 3421 3422 /* 3423 * For asynchronous completions, release the buffer now. The brelse 3424 * will do a wakeup there if necessary - so no need to do a wakeup 3425 * here in the async case. The sync case always needs to do a wakeup. 3426 */ 3427 3428 if (bp->b_flags & B_ASYNC) { 3429 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3430 brelse(bp); 3431 else 3432 bqrelse(bp); 3433 } else 3434 bdone(bp); 3435} 3436 3437/* 3438 * This routine is called in lieu of iodone in the case of 3439 * incomplete I/O. This keeps the busy status for pages 3440 * consistant. 3441 */ 3442void 3443vfs_unbusy_pages(struct buf *bp) 3444{ 3445 int i; 3446 vm_object_t obj; 3447 vm_page_t m; 3448 3449 runningbufwakeup(bp); 3450 if (!(bp->b_flags & B_VMIO)) 3451 return; 3452 3453 obj = bp->b_bufobj->bo_object; 3454 VM_OBJECT_LOCK(obj); 3455 for (i = 0; i < bp->b_npages; i++) { 3456 m = bp->b_pages[i]; 3457 if (m == bogus_page) { 3458 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3459 if (!m) 3460 panic("vfs_unbusy_pages: page missing\n"); 3461 bp->b_pages[i] = m; 3462 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3463 bp->b_pages, bp->b_npages); 3464 } 3465 vm_object_pip_subtract(obj, 1); 3466 vm_page_io_finish(m); 3467 } 3468 vm_object_pip_wakeupn(obj, 0); 3469 VM_OBJECT_UNLOCK(obj); 3470} 3471 3472/* 3473 * vfs_page_set_valid: 3474 * 3475 * Set the valid bits in a page based on the supplied offset. The 3476 * range is restricted to the buffer's size. 3477 * 3478 * This routine is typically called after a read completes. 3479 */ 3480static void 3481vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3482{ 3483 vm_ooffset_t eoff; 3484 3485 /* 3486 * Compute the end offset, eoff, such that [off, eoff) does not span a 3487 * page boundary and eoff is not greater than the end of the buffer. 3488 * The end of the buffer, in this case, is our file EOF, not the 3489 * allocation size of the buffer. 3490 */ 3491 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 3492 if (eoff > bp->b_offset + bp->b_bcount) 3493 eoff = bp->b_offset + bp->b_bcount; 3494 3495 /* 3496 * Set valid range. This is typically the entire buffer and thus the 3497 * entire page. 3498 */ 3499 if (eoff > off) 3500 vm_page_set_valid(m, off & PAGE_MASK, eoff - off); 3501} 3502 3503/* 3504 * vfs_page_set_validclean: 3505 * 3506 * Set the valid bits and clear the dirty bits in a page based on the 3507 * supplied offset. The range is restricted to the buffer's size. 3508 */ 3509static void 3510vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 3511{ 3512 vm_ooffset_t soff, eoff; 3513 3514 /* 3515 * Start and end offsets in buffer. eoff - soff may not cross a 3516 * page boundry or cross the end of the buffer. The end of the 3517 * buffer, in this case, is our file EOF, not the allocation size 3518 * of the buffer. 3519 */ 3520 soff = off; 3521 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3522 if (eoff > bp->b_offset + bp->b_bcount) 3523 eoff = bp->b_offset + bp->b_bcount; 3524 3525 /* 3526 * Set valid range. This is typically the entire buffer and thus the 3527 * entire page. 3528 */ 3529 if (eoff > soff) { 3530 vm_page_set_validclean( 3531 m, 3532 (vm_offset_t) (soff & PAGE_MASK), 3533 (vm_offset_t) (eoff - soff) 3534 ); 3535 } 3536} 3537 3538/* 3539 * Ensure that all buffer pages are not busied by VPO_BUSY flag. If 3540 * any page is busy, drain the flag. 3541 */ 3542static void 3543vfs_drain_busy_pages(struct buf *bp) 3544{ 3545 vm_page_t m; 3546 int i, last_busied; 3547 3548 VM_OBJECT_LOCK_ASSERT(bp->b_bufobj->bo_object, MA_OWNED); 3549 last_busied = 0; 3550 for (i = 0; i < bp->b_npages; i++) { 3551 m = bp->b_pages[i]; 3552 if ((m->oflags & VPO_BUSY) != 0) { 3553 for (; last_busied < i; last_busied++) 3554 vm_page_busy(bp->b_pages[last_busied]); 3555 while ((m->oflags & VPO_BUSY) != 0) 3556 vm_page_sleep(m, "vbpage"); 3557 } 3558 } 3559 for (i = 0; i < last_busied; i++) 3560 vm_page_wakeup(bp->b_pages[i]); 3561} 3562 3563/* 3564 * This routine is called before a device strategy routine. 3565 * It is used to tell the VM system that paging I/O is in 3566 * progress, and treat the pages associated with the buffer 3567 * almost as being VPO_BUSY. Also the object paging_in_progress 3568 * flag is handled to make sure that the object doesn't become 3569 * inconsistant. 3570 * 3571 * Since I/O has not been initiated yet, certain buffer flags 3572 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3573 * and should be ignored. 3574 */ 3575void 3576vfs_busy_pages(struct buf *bp, int clear_modify) 3577{ 3578 int i, bogus; 3579 vm_object_t obj; 3580 vm_ooffset_t foff; 3581 vm_page_t m; 3582 3583 if (!(bp->b_flags & B_VMIO)) 3584 return; 3585 3586 obj = bp->b_bufobj->bo_object; 3587 foff = bp->b_offset; 3588 KASSERT(bp->b_offset != NOOFFSET, 3589 ("vfs_busy_pages: no buffer offset")); 3590 VM_OBJECT_LOCK(obj); 3591 vfs_drain_busy_pages(bp); 3592 if (bp->b_bufsize != 0) 3593 vfs_setdirty_locked_object(bp); 3594 bogus = 0; 3595 for (i = 0; i < bp->b_npages; i++) { 3596 m = bp->b_pages[i]; 3597 3598 if ((bp->b_flags & B_CLUSTER) == 0) { 3599 vm_object_pip_add(obj, 1); 3600 vm_page_io_start(m); 3601 } 3602 /* 3603 * When readying a buffer for a read ( i.e 3604 * clear_modify == 0 ), it is important to do 3605 * bogus_page replacement for valid pages in 3606 * partially instantiated buffers. Partially 3607 * instantiated buffers can, in turn, occur when 3608 * reconstituting a buffer from its VM backing store 3609 * base. We only have to do this if B_CACHE is 3610 * clear ( which causes the I/O to occur in the 3611 * first place ). The replacement prevents the read 3612 * I/O from overwriting potentially dirty VM-backed 3613 * pages. XXX bogus page replacement is, uh, bogus. 3614 * It may not work properly with small-block devices. 3615 * We need to find a better way. 3616 */ 3617 if (clear_modify) { 3618 pmap_remove_write(m); 3619 vfs_page_set_validclean(bp, foff, m); 3620 } else if (m->valid == VM_PAGE_BITS_ALL && 3621 (bp->b_flags & B_CACHE) == 0) { 3622 bp->b_pages[i] = bogus_page; 3623 bogus++; 3624 } 3625 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3626 } 3627 VM_OBJECT_UNLOCK(obj); 3628 if (bogus) 3629 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3630 bp->b_pages, bp->b_npages); 3631} 3632 3633/* 3634 * vfs_bio_set_valid: 3635 * 3636 * Set the range within the buffer to valid. The range is 3637 * relative to the beginning of the buffer, b_offset. Note that 3638 * b_offset itself may be offset from the beginning of the first 3639 * page. 3640 */ 3641void 3642vfs_bio_set_valid(struct buf *bp, int base, int size) 3643{ 3644 int i, n; 3645 vm_page_t m; 3646 3647 if (!(bp->b_flags & B_VMIO)) 3648 return; 3649 3650 /* 3651 * Fixup base to be relative to beginning of first page. 3652 * Set initial n to be the maximum number of bytes in the 3653 * first page that can be validated. 3654 */ 3655 base += (bp->b_offset & PAGE_MASK); 3656 n = PAGE_SIZE - (base & PAGE_MASK); 3657 3658 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3659 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3660 m = bp->b_pages[i]; 3661 if (n > size) 3662 n = size; 3663 vm_page_set_valid(m, base & PAGE_MASK, n); 3664 base += n; 3665 size -= n; 3666 n = PAGE_SIZE; 3667 } 3668 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3669} 3670 3671/* 3672 * vfs_bio_clrbuf: 3673 * 3674 * If the specified buffer is a non-VMIO buffer, clear the entire 3675 * buffer. If the specified buffer is a VMIO buffer, clear and 3676 * validate only the previously invalid portions of the buffer. 3677 * This routine essentially fakes an I/O, so we need to clear 3678 * BIO_ERROR and B_INVAL. 3679 * 3680 * Note that while we only theoretically need to clear through b_bcount, 3681 * we go ahead and clear through b_bufsize. 3682 */ 3683void 3684vfs_bio_clrbuf(struct buf *bp) 3685{ 3686 int i, j, mask; 3687 caddr_t sa, ea; 3688 3689 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3690 clrbuf(bp); 3691 return; 3692 } 3693 bp->b_flags &= ~B_INVAL; 3694 bp->b_ioflags &= ~BIO_ERROR; 3695 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3696 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3697 (bp->b_offset & PAGE_MASK) == 0) { 3698 if (bp->b_pages[0] == bogus_page) 3699 goto unlock; 3700 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3701 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3702 if ((bp->b_pages[0]->valid & mask) == mask) 3703 goto unlock; 3704 if ((bp->b_pages[0]->valid & mask) == 0) { 3705 bzero(bp->b_data, bp->b_bufsize); 3706 bp->b_pages[0]->valid |= mask; 3707 goto unlock; 3708 } 3709 } 3710 ea = sa = bp->b_data; 3711 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3712 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3713 ea = (caddr_t)(vm_offset_t)ulmin( 3714 (u_long)(vm_offset_t)ea, 3715 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3716 if (bp->b_pages[i] == bogus_page) 3717 continue; 3718 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3719 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3720 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3721 if ((bp->b_pages[i]->valid & mask) == mask) 3722 continue; 3723 if ((bp->b_pages[i]->valid & mask) == 0) 3724 bzero(sa, ea - sa); 3725 else { 3726 for (; sa < ea; sa += DEV_BSIZE, j++) { 3727 if ((bp->b_pages[i]->valid & (1 << j)) == 0) 3728 bzero(sa, DEV_BSIZE); 3729 } 3730 } 3731 bp->b_pages[i]->valid |= mask; 3732 } 3733unlock: 3734 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3735 bp->b_resid = 0; 3736} 3737 3738/* 3739 * vm_hold_load_pages and vm_hold_free_pages get pages into 3740 * a buffers address space. The pages are anonymous and are 3741 * not associated with a file object. 3742 */ 3743static void 3744vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3745{ 3746 vm_offset_t pg; 3747 vm_page_t p; 3748 int index; 3749 3750 to = round_page(to); 3751 from = round_page(from); 3752 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3753 3754 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3755tryagain: 3756 /* 3757 * note: must allocate system pages since blocking here 3758 * could interfere with paging I/O, no matter which 3759 * process we are. 3760 */ 3761 p = vm_page_alloc(NULL, pg >> PAGE_SHIFT, VM_ALLOC_NOOBJ | 3762 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | 3763 VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT)); 3764 if (!p) { 3765 VM_WAIT; 3766 goto tryagain; 3767 } 3768 pmap_qenter(pg, &p, 1); 3769 bp->b_pages[index] = p; 3770 } 3771 bp->b_npages = index; 3772} 3773 3774/* Return pages associated with this buf to the vm system */ 3775static void 3776vm_hold_free_pages(struct buf *bp, int newbsize) 3777{ 3778 vm_offset_t from; 3779 vm_page_t p; 3780 int index, newnpages; 3781 3782 from = round_page((vm_offset_t)bp->b_data + newbsize); 3783 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3784 if (bp->b_npages > newnpages) 3785 pmap_qremove(from, bp->b_npages - newnpages); 3786 for (index = newnpages; index < bp->b_npages; index++) { 3787 p = bp->b_pages[index]; 3788 bp->b_pages[index] = NULL; 3789 if (p->busy != 0) 3790 printf("vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3791 (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); 3792 p->wire_count--; 3793 vm_page_free(p); 3794 atomic_subtract_int(&cnt.v_wire_count, 1); 3795 } 3796 bp->b_npages = newnpages; 3797} 3798 3799/* 3800 * Map an IO request into kernel virtual address space. 3801 * 3802 * All requests are (re)mapped into kernel VA space. 3803 * Notice that we use b_bufsize for the size of the buffer 3804 * to be mapped. b_bcount might be modified by the driver. 3805 * 3806 * Note that even if the caller determines that the address space should 3807 * be valid, a race or a smaller-file mapped into a larger space may 3808 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3809 * check the return value. 3810 */ 3811int 3812vmapbuf(struct buf *bp) 3813{ 3814 caddr_t kva; 3815 vm_prot_t prot; 3816 int pidx; 3817 3818 if (bp->b_bufsize < 0) 3819 return (-1); 3820 prot = VM_PROT_READ; 3821 if (bp->b_iocmd == BIO_READ) 3822 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3823 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 3824 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 3825 btoc(MAXPHYS))) < 0) 3826 return (-1); 3827 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3828 3829 kva = bp->b_saveaddr; 3830 bp->b_npages = pidx; 3831 bp->b_saveaddr = bp->b_data; 3832 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3833 return(0); 3834} 3835 3836/* 3837 * Free the io map PTEs associated with this IO operation. 3838 * We also invalidate the TLB entries and restore the original b_addr. 3839 */ 3840void 3841vunmapbuf(struct buf *bp) 3842{ 3843 int npages; 3844 3845 npages = bp->b_npages; 3846 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3847 vm_page_unhold_pages(bp->b_pages, npages); 3848 3849 bp->b_data = bp->b_saveaddr; 3850} 3851 3852void 3853bdone(struct buf *bp) 3854{ 3855 struct mtx *mtxp; 3856 3857 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3858 mtx_lock(mtxp); 3859 bp->b_flags |= B_DONE; 3860 wakeup(bp); 3861 mtx_unlock(mtxp); 3862} 3863 3864void 3865bwait(struct buf *bp, u_char pri, const char *wchan) 3866{ 3867 struct mtx *mtxp; 3868 3869 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3870 mtx_lock(mtxp); 3871 while ((bp->b_flags & B_DONE) == 0) 3872 msleep(bp, mtxp, pri, wchan, 0); 3873 mtx_unlock(mtxp); 3874} 3875 3876int 3877bufsync(struct bufobj *bo, int waitfor) 3878{ 3879 3880 return (VOP_FSYNC(bo->__bo_vnode, waitfor, curthread)); 3881} 3882 3883void 3884bufstrategy(struct bufobj *bo, struct buf *bp) 3885{ 3886 int i = 0; 3887 struct vnode *vp; 3888 3889 vp = bp->b_vp; 3890 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3891 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3892 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3893 i = VOP_STRATEGY(vp, bp); 3894 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3895} 3896 3897void 3898bufobj_wrefl(struct bufobj *bo) 3899{ 3900 3901 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3902 ASSERT_BO_LOCKED(bo); 3903 bo->bo_numoutput++; 3904} 3905 3906void 3907bufobj_wref(struct bufobj *bo) 3908{ 3909 3910 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3911 BO_LOCK(bo); 3912 bo->bo_numoutput++; 3913 BO_UNLOCK(bo); 3914} 3915 3916void 3917bufobj_wdrop(struct bufobj *bo) 3918{ 3919 3920 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3921 BO_LOCK(bo); 3922 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3923 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3924 bo->bo_flag &= ~BO_WWAIT; 3925 wakeup(&bo->bo_numoutput); 3926 } 3927 BO_UNLOCK(bo); 3928} 3929 3930int 3931bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3932{ 3933 int error; 3934 3935 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3936 ASSERT_BO_LOCKED(bo); 3937 error = 0; 3938 while (bo->bo_numoutput) { 3939 bo->bo_flag |= BO_WWAIT; 3940 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3941 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3942 if (error) 3943 break; 3944 } 3945 return (error); 3946} 3947 3948void 3949bpin(struct buf *bp) 3950{ 3951 struct mtx *mtxp; 3952 3953 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3954 mtx_lock(mtxp); 3955 bp->b_pin_count++; 3956 mtx_unlock(mtxp); 3957} 3958 3959void 3960bunpin(struct buf *bp) 3961{ 3962 struct mtx *mtxp; 3963 3964 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3965 mtx_lock(mtxp); 3966 if (--bp->b_pin_count == 0) 3967 wakeup(bp); 3968 mtx_unlock(mtxp); 3969} 3970 3971void 3972bunpin_wait(struct buf *bp) 3973{ 3974 struct mtx *mtxp; 3975 3976 mtxp = mtx_pool_find(mtxpool_sleep, bp); 3977 mtx_lock(mtxp); 3978 while (bp->b_pin_count > 0) 3979 msleep(bp, mtxp, PRIBIO, "bwunpin", 0); 3980 mtx_unlock(mtxp); 3981} 3982 3983#include "opt_ddb.h" 3984#ifdef DDB 3985#include <ddb/ddb.h> 3986 3987/* DDB command to show buffer data */ 3988DB_SHOW_COMMAND(buffer, db_show_buffer) 3989{ 3990 /* get args */ 3991 struct buf *bp = (struct buf *)addr; 3992 3993 if (!have_addr) { 3994 db_printf("usage: show buffer <addr>\n"); 3995 return; 3996 } 3997 3998 db_printf("buf at %p\n", bp); 3999 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 4000 db_printf( 4001 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 4002 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_dep = %p\n", 4003 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 4004 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 4005 bp->b_dep.lh_first); 4006 if (bp->b_npages) { 4007 int i; 4008 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 4009 for (i = 0; i < bp->b_npages; i++) { 4010 vm_page_t m; 4011 m = bp->b_pages[i]; 4012 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 4013 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 4014 if ((i + 1) < bp->b_npages) 4015 db_printf(","); 4016 } 4017 db_printf("\n"); 4018 } 4019 db_printf(" "); 4020 lockmgr_printinfo(&bp->b_lock); 4021} 4022 4023DB_SHOW_COMMAND(lockedbufs, lockedbufs) 4024{ 4025 struct buf *bp; 4026 int i; 4027 4028 for (i = 0; i < nbuf; i++) { 4029 bp = &buf[i]; 4030 if (BUF_ISLOCKED(bp)) { 4031 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4032 db_printf("\n"); 4033 } 4034 } 4035} 4036 4037DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 4038{ 4039 struct vnode *vp; 4040 struct buf *bp; 4041 4042 if (!have_addr) { 4043 db_printf("usage: show vnodebufs <addr>\n"); 4044 return; 4045 } 4046 vp = (struct vnode *)addr; 4047 db_printf("Clean buffers:\n"); 4048 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 4049 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4050 db_printf("\n"); 4051 } 4052 db_printf("Dirty buffers:\n"); 4053 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 4054 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 4055 db_printf("\n"); 4056 } 4057} 4058 4059DB_COMMAND(countfreebufs, db_coundfreebufs) 4060{ 4061 struct buf *bp; 4062 int i, used = 0, nfree = 0; 4063 4064 if (have_addr) { 4065 db_printf("usage: countfreebufs\n"); 4066 return; 4067 } 4068 4069 for (i = 0; i < nbuf; i++) { 4070 bp = &buf[i]; 4071 if ((bp->b_vflags & BV_INFREECNT) != 0) 4072 nfree++; 4073 else 4074 used++; 4075 } 4076 4077 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 4078 nfree + used); 4079 db_printf("numfreebuffers is %d\n", numfreebuffers); 4080} 4081#endif /* DDB */ 4082