vfs_bio.c revision 140721
1169689Skan/*- 2169689Skan * Copyright (c) 2004 Poul-Henning Kamp 3169689Skan * Copyright (c) 1994,1997 John S. Dyson 4169689Skan * All rights reserved. 5169689Skan * 6169689Skan * Redistribution and use in source and binary forms, with or without 7169689Skan * modification, are permitted provided that the following conditions 8169689Skan * are met: 9169689Skan * 1. Redistributions of source code must retain the above copyright 10169689Skan * notice, this list of conditions and the following disclaimer. 11169689Skan * 2. Redistributions in binary form must reproduce the above copyright 12169689Skan * notice, this list of conditions and the following disclaimer in the 13169689Skan * documentation and/or other materials provided with the distribution. 14169689Skan * 15169689Skan * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16169689Skan * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17169689Skan * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18169689Skan * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19169689Skan * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20169689Skan * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21169689Skan * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22169689Skan * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23169689Skan * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24169689Skan * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25169689Skan * SUCH DAMAGE. 26169689Skan */ 27169689Skan 28132718Skan/* 2990075Sobrien * this file contains a new buffer I/O scheme implementing a coherent 3090075Sobrien * VM object and buffer cache scheme. Pains have been taken to make 3190075Sobrien * sure that the performance degradation associated with schemes such 3290075Sobrien * as this is not realized. 3390075Sobrien * 3490075Sobrien * Author: John S. Dyson 35132718Skan * Significant help during the development and debugging phases 36132718Skan * had been provided by David Greenman, also of the FreeBSD core team. 3790075Sobrien * 3890075Sobrien * see man buf(9) for more info. 3990075Sobrien */ 40132718Skan 4190075Sobrien#include <sys/cdefs.h> 42132718Skan__FBSDID("$FreeBSD: head/sys/kern/vfs_bio.c 140721 2005-01-24 10:47:04Z jeff $"); 43132718Skan 4490075Sobrien#include <sys/param.h> 4590075Sobrien#include <sys/systm.h> 4690075Sobrien#include <sys/bio.h> 4790075Sobrien#include <sys/conf.h> 4890075Sobrien#include <sys/buf.h> 4990075Sobrien#include <sys/devicestat.h> 5090075Sobrien#include <sys/eventhandler.h> 5190075Sobrien#include <sys/lock.h> 5290075Sobrien#include <sys/malloc.h> 5390075Sobrien#include <sys/mount.h> 5490075Sobrien#include <sys/mutex.h> 5590075Sobrien#include <sys/kernel.h> 5690075Sobrien#include <sys/kthread.h> 5790075Sobrien#include <sys/proc.h> 5890075Sobrien#include <sys/resourcevar.h> 5990075Sobrien#include <sys/sysctl.h> 6090075Sobrien#include <sys/vmmeter.h> 6190075Sobrien#include <sys/vnode.h> 6290075Sobrien#include <geom/geom.h> 6390075Sobrien#include <vm/vm.h> 6490075Sobrien#include <vm/vm_param.h> 6590075Sobrien#include <vm/vm_kern.h> 6690075Sobrien#include <vm/vm_pageout.h> 6790075Sobrien#include <vm/vm_page.h> 6890075Sobrien#include <vm/vm_object.h> 6990075Sobrien#include <vm/vm_extern.h> 7090075Sobrien#include <vm/vm_map.h> 71132718Skan#include "opt_directio.h" 7290075Sobrien#include "opt_swap.h" 7390075Sobrien 7490075Sobrienstatic MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 7590075Sobrien 7690075Sobrienstruct bio_ops bioops; /* I/O operation notification */ 7790075Sobrien 7890075Sobrienstruct buf_ops buf_ops_bio = { 7990075Sobrien .bop_name = "buf_ops_bio", 8090075Sobrien .bop_write = bufwrite, 8190075Sobrien .bop_strategy = bufstrategy, 8290075Sobrien .bop_sync = bufsync, 8390075Sobrien}; 8490075Sobrien 8590075Sobrien/* 8690075Sobrien * XXX buf is global because kern_shutdown.c and ffs_checkoverlap has 8790075Sobrien * carnal knowledge of buffers. This knowledge should be moved to vfs_bio.c. 8890075Sobrien */ 8990075Sobrienstruct buf *buf; /* buffer header pool */ 9090075Sobrien 9190075Sobrienstatic struct proc *bufdaemonproc; 9290075Sobrien 9390075Sobrienstatic int inmem(struct vnode *vp, daddr_t blkno); 9490075Sobrienstatic void vm_hold_free_pages(struct buf *bp, vm_offset_t from, 9590075Sobrien vm_offset_t to); 9690075Sobrienstatic void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 9790075Sobrien vm_offset_t to); 9890075Sobrienstatic void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 9990075Sobrien int pageno, vm_page_t m); 10090075Sobrienstatic void vfs_clean_pages(struct buf *bp); 10190075Sobrienstatic void vfs_setdirty(struct buf *bp); 10290075Sobrienstatic void vfs_vmio_release(struct buf *bp); 10390075Sobrienstatic void vfs_backgroundwritedone(struct buf *bp); 10490075Sobrienstatic int vfs_bio_clcheck(struct vnode *vp, int size, 10590075Sobrien daddr_t lblkno, daddr_t blkno); 10690075Sobrienstatic int flushbufqueues(int flushdeps); 10790075Sobrienstatic void buf_daemon(void); 10890075Sobrienvoid bremfreel(struct buf *bp); 10990075Sobrien 11090075Sobrienint vmiodirenable = TRUE; 11190075SobrienSYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 11290075Sobrien "Use the VM system for directory writes"); 11390075Sobrienint runningbufspace; 11490075SobrienSYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 11590075Sobrien "Amount of presently outstanding async buffer io"); 11690075Sobrienstatic int bufspace; 11790075SobrienSYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0, 11890075Sobrien "KVA memory used for bufs"); 11990075Sobrienstatic int maxbufspace; 12090075SobrienSYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0, 12190075Sobrien "Maximum allowed value of bufspace (including buf_daemon)"); 12290075Sobrienstatic int bufmallocspace; 12390075SobrienSYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 12490075Sobrien "Amount of malloced memory for buffers"); 12590075Sobrienstatic int maxbufmallocspace; 12690075SobrienSYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, 12790075Sobrien "Maximum amount of malloced memory for buffers"); 12890075Sobrienstatic int lobufspace; 12990075SobrienSYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0, 13090075Sobrien "Minimum amount of buffers we want to have"); 13190075Sobrienstatic int hibufspace; 13290075SobrienSYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0, 13390075Sobrien "Maximum allowed value of bufspace (excluding buf_daemon)"); 13490075Sobrienstatic int bufreusecnt; 13590075SobrienSYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW, &bufreusecnt, 0, 13690075Sobrien "Number of times we have reused a buffer"); 13790075Sobrienstatic int buffreekvacnt; 13890075SobrienSYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 0, 13990075Sobrien "Number of times we have freed the KVA space from some buffer"); 14090075Sobrienstatic int bufdefragcnt; 14190075SobrienSYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 0, 14290075Sobrien "Number of times we have had to repeat buffer allocation to defragment"); 14390075Sobrienstatic int lorunningspace; 14490075SobrienSYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0, 14590075Sobrien "Minimum preferred space used for in-progress I/O"); 14690075Sobrienstatic int hirunningspace; 14790075SobrienSYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0, 14890075Sobrien "Maximum amount of space to use for in-progress I/O"); 14990075Sobrienstatic int dirtybufferflushes; 15090075SobrienSYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 15190075Sobrien 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 15290075Sobrienstatic int altbufferflushes; 15390075SobrienSYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 15490075Sobrien 0, "Number of fsync flushes to limit dirty buffers"); 15590075Sobrienstatic int recursiveflushes; 15690075SobrienSYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 15790075Sobrien 0, "Number of flushes skipped due to being recursive"); 15890075Sobrienstatic int numdirtybuffers; 15990075SobrienSYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0, 16090075Sobrien "Number of buffers that are dirty (has unwritten changes) at the moment"); 16190075Sobrienstatic int lodirtybuffers; 16290075SobrienSYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0, 16390075Sobrien "How many buffers we want to have free before bufdaemon can sleep"); 16490075Sobrienstatic int hidirtybuffers; 16590075SobrienSYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0, 16690075Sobrien "When the number of dirty buffers is considered severe"); 16790075Sobrienstatic int dirtybufthresh; 16890075SobrienSYSCTL_INT(_vfs, OID_AUTO, dirtybufthresh, CTLFLAG_RW, &dirtybufthresh, 169169689Skan 0, "Number of bdwrite to bawrite conversions to clear dirty buffers"); 170169689Skanstatic int numfreebuffers; 17190075SobrienSYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 17290075Sobrien "Number of free buffers"); 17390075Sobrienstatic int lofreebuffers; 17490075SobrienSYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0, 175169689Skan "XXX Unused"); 17690075Sobrienstatic int hifreebuffers; 17790075SobrienSYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0, 17890075Sobrien "XXX Complicatedly unused"); 17990075Sobrienstatic int getnewbufcalls; 18090075SobrienSYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW, &getnewbufcalls, 0, 18190075Sobrien "Number of calls to getnewbuf"); 18290075Sobrienstatic int getnewbufrestarts; 18390075SobrienSYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW, &getnewbufrestarts, 0, 18490075Sobrien "Number of times getnewbuf has had to restart a buffer aquisition"); 18590075Sobrienstatic int dobkgrdwrite = 1; 18690075SobrienSYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0, 18790075Sobrien "Do background writes (honoring the BV_BKGRDWRITE flag)?"); 18890075Sobrien 18990075Sobrien/* 19090075Sobrien * Wakeup point for bufdaemon, as well as indicator of whether it is already 19190075Sobrien * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 19290075Sobrien * is idling. 19390075Sobrien */ 19490075Sobrienstatic int bd_request; 19590075Sobrien 19690075Sobrien/* 19790075Sobrien * This lock synchronizes access to bd_request. 19890075Sobrien */ 19990075Sobrienstatic struct mtx bdlock; 20090075Sobrien 20190075Sobrien/* 20290075Sobrien * bogus page -- for I/O to/from partially complete buffers 20390075Sobrien * this is a temporary solution to the problem, but it is not 20490075Sobrien * really that bad. it would be better to split the buffer 20590075Sobrien * for input in the case of buffers partially already in memory, 20690075Sobrien * but the code is intricate enough already. 20790075Sobrien */ 20890075Sobrienvm_page_t bogus_page; 20990075Sobrien 21090075Sobrien/* 21190075Sobrien * Synchronization (sleep/wakeup) variable for active buffer space requests. 21290075Sobrien * Set when wait starts, cleared prior to wakeup(). 21390075Sobrien * Used in runningbufwakeup() and waitrunningbufspace(). 21490075Sobrien */ 21590075Sobrienstatic int runningbufreq; 21690075Sobrien 21790075Sobrien/* 21890075Sobrien * This lock protects the runningbufreq and synchronizes runningbufwakeup and 21990075Sobrien * waitrunningbufspace(). 22090075Sobrien */ 22190075Sobrienstatic struct mtx rbreqlock; 222169689Skan 223169689Skan/* 22490075Sobrien * Synchronization (sleep/wakeup) variable for buffer requests. 22590075Sobrien * Can contain the VFS_BIO_NEED flags defined below; setting/clearing is done 22690075Sobrien * by and/or. 22790075Sobrien * Used in numdirtywakeup(), bufspacewakeup(), bufcountwakeup(), bwillwrite(), 228169689Skan * getnewbuf(), and getblk(). 22990075Sobrien */ 23090075Sobrienstatic int needsbuffer; 23190075Sobrien 23290075Sobrien/* 23390075Sobrien * Lock that protects needsbuffer and the sleeps/wakeups surrounding it. 23490075Sobrien */ 23590075Sobrienstatic struct mtx nblock; 23690075Sobrien 23790075Sobrien/* 23890075Sobrien * Lock that protects against bwait()/bdone()/B_DONE races. 23990075Sobrien */ 24090075Sobrien 24190075Sobrienstatic struct mtx bdonelock; 24290075Sobrien 24390075Sobrien/* 24490075Sobrien * Definitions for the buffer free lists. 24590075Sobrien */ 24690075Sobrien#define BUFFER_QUEUES 5 /* number of free buffer queues */ 24790075Sobrien 24890075Sobrien#define QUEUE_NONE 0 /* on no queue */ 24990075Sobrien#define QUEUE_CLEAN 1 /* non-B_DELWRI buffers */ 25090075Sobrien#define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 25190075Sobrien#define QUEUE_EMPTYKVA 3 /* empty buffer headers w/KVA assignment */ 25290075Sobrien#define QUEUE_EMPTY 4 /* empty buffer headers */ 25390075Sobrien 25490075Sobrien/* Queues for free buffers with various properties */ 25590075Sobrienstatic TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES] = { { 0 } }; 25690075Sobrien 25790075Sobrien/* Lock for the bufqueues */ 25890075Sobrienstatic struct mtx bqlock; 25990075Sobrien 26090075Sobrien/* 26190075Sobrien * Single global constant for BUF_WMESG, to avoid getting multiple references. 26290075Sobrien * buf_wmesg is referred from macros. 26390075Sobrien */ 26490075Sobrienconst char *buf_wmesg = BUF_WMESG; 26590075Sobrien 26690075Sobrien#define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */ 26790075Sobrien#define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */ 26890075Sobrien#define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */ 26990075Sobrien#define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */ 27090075Sobrien 27190075Sobrien#ifdef DIRECTIO 27290075Sobrienextern void ffs_rawread_setup(void); 27390075Sobrien#endif /* DIRECTIO */ 27490075Sobrien/* 27590075Sobrien * numdirtywakeup: 27690075Sobrien * 27790075Sobrien * If someone is blocked due to there being too many dirty buffers, 27890075Sobrien * and numdirtybuffers is now reasonable, wake them up. 279169689Skan */ 280169689Skan 28190075Sobrienstatic __inline void 28290075Sobriennumdirtywakeup(int level) 28390075Sobrien{ 28490075Sobrien 285169689Skan if (numdirtybuffers <= level) { 28690075Sobrien mtx_lock(&nblock); 28790075Sobrien if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) { 28890075Sobrien needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH; 28990075Sobrien wakeup(&needsbuffer); 29090075Sobrien } 29190075Sobrien mtx_unlock(&nblock); 29290075Sobrien } 29390075Sobrien} 29490075Sobrien 29590075Sobrien/* 29690075Sobrien * bufspacewakeup: 29790075Sobrien * 29890075Sobrien * Called when buffer space is potentially available for recovery. 29990075Sobrien * getnewbuf() will block on this flag when it is unable to free 30090075Sobrien * sufficient buffer space. Buffer space becomes recoverable when 30190075Sobrien * bp's get placed back in the queues. 30290075Sobrien */ 30390075Sobrien 30490075Sobrienstatic __inline void 30590075Sobrienbufspacewakeup(void) 30690075Sobrien{ 30790075Sobrien 30890075Sobrien /* 30990075Sobrien * If someone is waiting for BUF space, wake them up. Even 31090075Sobrien * though we haven't freed the kva space yet, the waiting 31190075Sobrien * process will be able to now. 31290075Sobrien */ 31390075Sobrien mtx_lock(&nblock); 31490075Sobrien if (needsbuffer & VFS_BIO_NEED_BUFSPACE) { 31590075Sobrien needsbuffer &= ~VFS_BIO_NEED_BUFSPACE; 31690075Sobrien wakeup(&needsbuffer); 31790075Sobrien } 31890075Sobrien mtx_unlock(&nblock); 31990075Sobrien} 32090075Sobrien 32190075Sobrien/* 32290075Sobrien * runningbufwakeup() - in-progress I/O accounting. 32390075Sobrien * 32490075Sobrien */ 32590075Sobrienstatic __inline void 32690075Sobrienrunningbufwakeup(struct buf *bp) 32790075Sobrien{ 32890075Sobrien 32990075Sobrien if (bp->b_runningbufspace) { 33090075Sobrien atomic_subtract_int(&runningbufspace, bp->b_runningbufspace); 33190075Sobrien bp->b_runningbufspace = 0; 332169689Skan mtx_lock(&rbreqlock); 333169689Skan if (runningbufreq && runningbufspace <= lorunningspace) { 33490075Sobrien runningbufreq = 0; 33590075Sobrien wakeup(&runningbufreq); 33690075Sobrien } 33790075Sobrien mtx_unlock(&rbreqlock); 338169689Skan } 33990075Sobrien} 34090075Sobrien 34190075Sobrien/* 34290075Sobrien * bufcountwakeup: 34390075Sobrien * 34490075Sobrien * Called when a buffer has been added to one of the free queues to 34590075Sobrien * account for the buffer and to wakeup anyone waiting for free buffers. 34690075Sobrien * This typically occurs when large amounts of metadata are being handled 34790075Sobrien * by the buffer cache ( else buffer space runs out first, usually ). 34890075Sobrien */ 34990075Sobrien 35090075Sobrienstatic __inline void 35190075Sobrienbufcountwakeup(void) 35290075Sobrien{ 35390075Sobrien 35490075Sobrien atomic_add_int(&numfreebuffers, 1); 35590075Sobrien mtx_lock(&nblock); 35690075Sobrien if (needsbuffer) { 35790075Sobrien needsbuffer &= ~VFS_BIO_NEED_ANY; 35890075Sobrien if (numfreebuffers >= hifreebuffers) 35990075Sobrien needsbuffer &= ~VFS_BIO_NEED_FREE; 36090075Sobrien wakeup(&needsbuffer); 36190075Sobrien } 36290075Sobrien mtx_unlock(&nblock); 36390075Sobrien} 36490075Sobrien 36590075Sobrien/* 36690075Sobrien * waitrunningbufspace() 36790075Sobrien * 36890075Sobrien * runningbufspace is a measure of the amount of I/O currently 36990075Sobrien * running. This routine is used in async-write situations to 37090075Sobrien * prevent creating huge backups of pending writes to a device. 37190075Sobrien * Only asynchronous writes are governed by this function. 37290075Sobrien * 37390075Sobrien * Reads will adjust runningbufspace, but will not block based on it. 37490075Sobrien * The read load has a side effect of reducing the allowed write load. 37590075Sobrien * 37690075Sobrien * This does NOT turn an async write into a sync write. It waits 37790075Sobrien * for earlier writes to complete and generally returns before the 37890075Sobrien * caller's write has reached the device. 37990075Sobrien */ 38090075Sobrienstatic __inline void 38190075Sobrienwaitrunningbufspace(void) 38290075Sobrien{ 38390075Sobrien 38490075Sobrien mtx_lock(&rbreqlock); 38590075Sobrien while (runningbufspace > hirunningspace) { 38690075Sobrien ++runningbufreq; 387169689Skan msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 388169689Skan } 38990075Sobrien mtx_unlock(&rbreqlock); 39090075Sobrien} 39190075Sobrien 39290075Sobrien 39390075Sobrien/* 394169689Skan * vfs_buf_test_cache: 39590075Sobrien * 39690075Sobrien * Called when a buffer is extended. This function clears the B_CACHE 39790075Sobrien * bit if the newly extended portion of the buffer does not contain 39890075Sobrien * valid data. 39990075Sobrien */ 40090075Sobrienstatic __inline 40190075Sobrienvoid 40290075Sobrienvfs_buf_test_cache(struct buf *bp, 40390075Sobrien vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 40490075Sobrien vm_page_t m) 40590075Sobrien{ 40690075Sobrien 40790075Sobrien VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 40890075Sobrien if (bp->b_flags & B_CACHE) { 40990075Sobrien int base = (foff + off) & PAGE_MASK; 41090075Sobrien if (vm_page_is_valid(m, base, size) == 0) 41190075Sobrien bp->b_flags &= ~B_CACHE; 41290075Sobrien } 41390075Sobrien} 41490075Sobrien 41590075Sobrien/* Wake up the buffer deamon if necessary */ 41690075Sobrienstatic __inline 41790075Sobrienvoid 41890075Sobrienbd_wakeup(int dirtybuflevel) 41990075Sobrien{ 42090075Sobrien 42190075Sobrien mtx_lock(&bdlock); 42290075Sobrien if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) { 42390075Sobrien bd_request = 1; 42490075Sobrien wakeup(&bd_request); 42590075Sobrien } 42690075Sobrien mtx_unlock(&bdlock); 42790075Sobrien} 42890075Sobrien 42990075Sobrien/* 43090075Sobrien * bd_speedup - speedup the buffer cache flushing code 43190075Sobrien */ 43290075Sobrien 43390075Sobrienstatic __inline 43490075Sobrienvoid 43590075Sobrienbd_speedup(void) 43690075Sobrien{ 43790075Sobrien 43890075Sobrien bd_wakeup(1); 439169689Skan} 440169689Skan 44190075Sobrien/* 44290075Sobrien * Calculating buffer cache scaling values and reserve space for buffer 44390075Sobrien * headers. This is called during low level kernel initialization and 44490075Sobrien * may be called more then once. We CANNOT write to the memory area 44590075Sobrien * being reserved at this time. 44690075Sobrien */ 44790075Sobriencaddr_t 448169689Skankern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 44990075Sobrien{ 45090075Sobrien 45190075Sobrien /* 45290075Sobrien * physmem_est is in pages. Convert it to kilobytes (assumes 45390075Sobrien * PAGE_SIZE is >= 1K) 45490075Sobrien */ 45590075Sobrien physmem_est = physmem_est * (PAGE_SIZE / 1024); 45690075Sobrien 45790075Sobrien /* 45890075Sobrien * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 45990075Sobrien * For the first 64MB of ram nominally allocate sufficient buffers to 46090075Sobrien * cover 1/4 of our ram. Beyond the first 64MB allocate additional 46190075Sobrien * buffers to cover 1/20 of our ram over 64MB. When auto-sizing 46290075Sobrien * the buffer cache we limit the eventual kva reservation to 46390075Sobrien * maxbcache bytes. 46490075Sobrien * 46590075Sobrien * factor represents the 1/4 x ram conversion. 46690075Sobrien */ 46790075Sobrien if (nbuf == 0) { 46890075Sobrien int factor = 4 * BKVASIZE / 1024; 46990075Sobrien 47090075Sobrien nbuf = 50; 47190075Sobrien if (physmem_est > 4096) 47290075Sobrien nbuf += min((physmem_est - 4096) / factor, 47390075Sobrien 65536 / factor); 47490075Sobrien if (physmem_est > 65536) 47590075Sobrien nbuf += (physmem_est - 65536) * 2 / (factor * 5); 47690075Sobrien 47790075Sobrien if (maxbcache && nbuf > maxbcache / BKVASIZE) 47890075Sobrien nbuf = maxbcache / BKVASIZE; 47990075Sobrien } 48090075Sobrien 48190075Sobrien#if 0 48290075Sobrien /* 48390075Sobrien * Do not allow the buffer_map to be more then 1/2 the size of the 48490075Sobrien * kernel_map. 48590075Sobrien */ 48690075Sobrien if (nbuf > (kernel_map->max_offset - kernel_map->min_offset) / 48790075Sobrien (BKVASIZE * 2)) { 48890075Sobrien nbuf = (kernel_map->max_offset - kernel_map->min_offset) / 489169689Skan (BKVASIZE * 2); 490169689Skan printf("Warning: nbufs capped at %d\n", nbuf); 49190075Sobrien } 49290075Sobrien#endif 49390075Sobrien 494169689Skan /* 49590075Sobrien * swbufs are used as temporary holders for I/O, such as paging I/O. 49690075Sobrien * We have no less then 16 and no more then 256. 49790075Sobrien */ 49890075Sobrien nswbuf = max(min(nbuf/4, 256), 16); 49990075Sobrien#ifdef NSWBUF_MIN 50090075Sobrien if (nswbuf < NSWBUF_MIN) 50190075Sobrien nswbuf = NSWBUF_MIN; 50290075Sobrien#endif 50390075Sobrien#ifdef DIRECTIO 50490075Sobrien ffs_rawread_setup(); 50590075Sobrien#endif 50690075Sobrien 50790075Sobrien /* 50890075Sobrien * Reserve space for the buffer cache buffers 50990075Sobrien */ 51090075Sobrien swbuf = (void *)v; 51190075Sobrien v = (caddr_t)(swbuf + nswbuf); 51290075Sobrien buf = (void *)v; 51390075Sobrien v = (caddr_t)(buf + nbuf); 51490075Sobrien 51590075Sobrien return(v); 51690075Sobrien} 51790075Sobrien 51890075Sobrien/* Initialize the buffer subsystem. Called before use of any buffers. */ 51990075Sobrienvoid 52090075Sobrienbufinit(void) 52190075Sobrien{ 52290075Sobrien struct buf *bp; 52390075Sobrien int i; 52490075Sobrien 52590075Sobrien mtx_init(&bqlock, "buf queue lock", NULL, MTX_DEF); 52690075Sobrien mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 52790075Sobrien mtx_init(&nblock, "needsbuffer lock", NULL, MTX_DEF); 52890075Sobrien mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 52990075Sobrien mtx_init(&bdonelock, "bdone lock", NULL, MTX_DEF); 53090075Sobrien 53190075Sobrien /* next, make a null set of free lists */ 53290075Sobrien for (i = 0; i < BUFFER_QUEUES; i++) 53390075Sobrien TAILQ_INIT(&bufqueues[i]); 53490075Sobrien 535169689Skan /* finally, initialize each buffer header and stick on empty q */ 536169689Skan for (i = 0; i < nbuf; i++) { 53790075Sobrien bp = &buf[i]; 53890075Sobrien bzero(bp, sizeof *bp); 53990075Sobrien bp->b_flags = B_INVAL; /* we're just an empty header */ 54090075Sobrien bp->b_rcred = NOCRED; 54190075Sobrien bp->b_wcred = NOCRED; 54290075Sobrien bp->b_qindex = QUEUE_EMPTY; 54390075Sobrien bp->b_vflags = 0; 544169689Skan bp->b_xflags = 0; 54590075Sobrien LIST_INIT(&bp->b_dep); 54690075Sobrien BUF_LOCKINIT(bp); 54790075Sobrien TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 54890075Sobrien } 54990075Sobrien 55090075Sobrien /* 55190075Sobrien * maxbufspace is the absolute maximum amount of buffer space we are 55290075Sobrien * allowed to reserve in KVM and in real terms. The absolute maximum 55390075Sobrien * is nominally used by buf_daemon. hibufspace is the nominal maximum 55490075Sobrien * used by most other processes. The differential is required to 55590075Sobrien * ensure that buf_daemon is able to run when other processes might 55690075Sobrien * be blocked waiting for buffer space. 55790075Sobrien * 55890075Sobrien * maxbufspace is based on BKVASIZE. Allocating buffers larger then 55990075Sobrien * this may result in KVM fragmentation which is not handled optimally 56090075Sobrien * by the system. 56190075Sobrien */ 56290075Sobrien maxbufspace = nbuf * BKVASIZE; 56390075Sobrien hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10); 56490075Sobrien lobufspace = hibufspace - MAXBSIZE; 56590075Sobrien 56690075Sobrien lorunningspace = 512 * 1024; 56790075Sobrien hirunningspace = 1024 * 1024; 56890075Sobrien 56990075Sobrien/* 57090075Sobrien * Limit the amount of malloc memory since it is wired permanently into 57190075Sobrien * the kernel space. Even though this is accounted for in the buffer 57290075Sobrien * allocation, we don't want the malloced region to grow uncontrolled. 57390075Sobrien * The malloc scheme improves memory utilization significantly on average 57490075Sobrien * (small) directories. 57590075Sobrien */ 57690075Sobrien maxbufmallocspace = hibufspace / 20; 57790075Sobrien 57890075Sobrien/* 57990075Sobrien * Reduce the chance of a deadlock occuring by limiting the number 58090075Sobrien * of delayed-write dirty buffers we allow to stack up. 58190075Sobrien */ 58290075Sobrien hidirtybuffers = nbuf / 4 + 20; 58390075Sobrien dirtybufthresh = hidirtybuffers * 9 / 10; 58490075Sobrien numdirtybuffers = 0; 58590075Sobrien/* 58690075Sobrien * To support extreme low-memory systems, make sure hidirtybuffers cannot 58790075Sobrien * eat up all available buffer space. This occurs when our minimum cannot 58890075Sobrien * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming 58990075Sobrien * BKVASIZE'd (8K) buffers. 59090075Sobrien */ 59190075Sobrien while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 59290075Sobrien hidirtybuffers >>= 1; 59390075Sobrien } 59490075Sobrien lodirtybuffers = hidirtybuffers / 2; 59590075Sobrien 59690075Sobrien/* 59790075Sobrien * Try to keep the number of free buffers in the specified range, 59890075Sobrien * and give special processes (e.g. like buf_daemon) access to an 59990075Sobrien * emergency reserve. 60090075Sobrien */ 60190075Sobrien lofreebuffers = nbuf / 18 + 5; 60290075Sobrien hifreebuffers = 2 * lofreebuffers; 60390075Sobrien numfreebuffers = nbuf; 60490075Sobrien 60590075Sobrien/* 60690075Sobrien * Maximum number of async ops initiated per buf_daemon loop. This is 60790075Sobrien * somewhat of a hack at the moment, we really need to limit ourselves 60890075Sobrien * based on the number of bytes of I/O in-transit that were initiated 60990075Sobrien * from buf_daemon. 61090075Sobrien */ 61190075Sobrien 61290075Sobrien bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | 61390075Sobrien VM_ALLOC_NORMAL | VM_ALLOC_WIRED); 61490075Sobrien} 61590075Sobrien 61690075Sobrien/* 61790075Sobrien * bfreekva() - free the kva allocation for a buffer. 61890075Sobrien * 61990075Sobrien * Must be called at splbio() or higher as this is the only locking for 62090075Sobrien * buffer_map. 62190075Sobrien * 62290075Sobrien * Since this call frees up buffer space, we call bufspacewakeup(). 62390075Sobrien */ 62490075Sobrienstatic void 62590075Sobrienbfreekva(struct buf *bp) 62690075Sobrien{ 62790075Sobrien 62890075Sobrien if (bp->b_kvasize) { 62990075Sobrien atomic_add_int(&buffreekvacnt, 1); 63090075Sobrien atomic_subtract_int(&bufspace, bp->b_kvasize); 63190075Sobrien vm_map_delete(buffer_map, 63290075Sobrien (vm_offset_t) bp->b_kvabase, 63390075Sobrien (vm_offset_t) bp->b_kvabase + bp->b_kvasize 63490075Sobrien ); 63590075Sobrien bp->b_kvasize = 0; 63690075Sobrien bufspacewakeup(); 63790075Sobrien } 63890075Sobrien} 63990075Sobrien 64090075Sobrien/* 64190075Sobrien * bremfree: 64290075Sobrien * 64390075Sobrien * Mark the buffer for removal from the appropriate free list in brelse. 64490075Sobrien * 64590075Sobrien */ 64690075Sobrienvoid 64790075Sobrienbremfree(struct buf *bp) 64890075Sobrien{ 64990075Sobrien 65090075Sobrien CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 65190075Sobrien KASSERT(BUF_REFCNT(bp), ("bremfree: buf must be locked.")); 65290075Sobrien KASSERT((bp->b_flags & B_REMFREE) == 0 && bp->b_qindex != QUEUE_NONE, 65390075Sobrien ("bremfree: buffer %p not on a queue.", bp)); 65490075Sobrien 65590075Sobrien bp->b_flags |= B_REMFREE; 65690075Sobrien /* Fixup numfreebuffers count. */ 65790075Sobrien if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 65890075Sobrien atomic_subtract_int(&numfreebuffers, 1); 65990075Sobrien} 66090075Sobrien 66190075Sobrien/* 66290075Sobrien * bremfreef: 66390075Sobrien * 66490075Sobrien * Force an immediate removal from a free list. Used only in nfs when 66590075Sobrien * it abuses the b_freelist pointer. 66690075Sobrien */ 66790075Sobrienvoid 66890075Sobrienbremfreef(struct buf *bp) 66990075Sobrien{ 67090075Sobrien mtx_lock(&bqlock); 67190075Sobrien bremfreel(bp); 67290075Sobrien mtx_unlock(&bqlock); 67390075Sobrien} 67490075Sobrien 67590075Sobrien/* 67690075Sobrien * bremfreel: 67790075Sobrien * 67890075Sobrien * Removes a buffer from the free list, must be called with the 67990075Sobrien * bqlock held. 68090075Sobrien */ 68190075Sobrienvoid 68290075Sobrienbremfreel(struct buf *bp) 68390075Sobrien{ 68490075Sobrien int s = splbio(); 68590075Sobrien 68690075Sobrien CTR3(KTR_BUF, "bremfreel(%p) vp %p flags %X", 68790075Sobrien bp, bp->b_vp, bp->b_flags); 68890075Sobrien KASSERT(BUF_REFCNT(bp), ("bremfreel: buffer %p not locked.", bp)); 68990075Sobrien KASSERT(bp->b_qindex != QUEUE_NONE, 69090075Sobrien ("bremfreel: buffer %p not on a queue.", bp)); 69190075Sobrien mtx_assert(&bqlock, MA_OWNED); 69290075Sobrien 69390075Sobrien TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 69490075Sobrien bp->b_qindex = QUEUE_NONE; 69590075Sobrien /* 69690075Sobrien * If this was a delayed bremfree() we only need to remove the buffer 69790075Sobrien * from the queue and return the stats are already done. 69890075Sobrien */ 69990075Sobrien if (bp->b_flags & B_REMFREE) { 70090075Sobrien bp->b_flags &= ~B_REMFREE; 70190075Sobrien splx(s); 70290075Sobrien return; 70390075Sobrien } 70490075Sobrien /* 70590075Sobrien * Fixup numfreebuffers count. If the buffer is invalid or not 70690075Sobrien * delayed-write, the buffer was free and we must decrement 70790075Sobrien * numfreebuffers. 70890075Sobrien */ 70990075Sobrien if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) 71090075Sobrien atomic_subtract_int(&numfreebuffers, 1); 71190075Sobrien splx(s); 71290075Sobrien} 71390075Sobrien 71490075Sobrien 71590075Sobrien/* 71690075Sobrien * Get a buffer with the specified data. Look in the cache first. We 71790075Sobrien * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 71890075Sobrien * is set, the buffer is valid and we do not have to do anything ( see 71990075Sobrien * getblk() ). This is really just a special case of breadn(). 72090075Sobrien */ 72190075Sobrienint 72290075Sobrienbread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 72390075Sobrien struct buf **bpp) 72490075Sobrien{ 72590075Sobrien 72690075Sobrien return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp)); 72790075Sobrien} 72890075Sobrien 72990075Sobrien/* 73090075Sobrien * Operates like bread, but also starts asynchronous I/O on 73190075Sobrien * read-ahead blocks. We must clear BIO_ERROR and B_INVAL prior 73290075Sobrien * to initiating I/O . If B_CACHE is set, the buffer is valid 73390075Sobrien * and we do not have to do anything. 73490075Sobrien */ 73590075Sobrienint 73690075Sobrienbreadn(struct vnode * vp, daddr_t blkno, int size, 73790075Sobrien daddr_t * rablkno, int *rabsize, 73890075Sobrien int cnt, struct ucred * cred, struct buf **bpp) 73990075Sobrien{ 74090075Sobrien struct buf *bp, *rabp; 74190075Sobrien int i; 74290075Sobrien int rv = 0, readwait = 0; 74390075Sobrien 74490075Sobrien CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 74590075Sobrien *bpp = bp = getblk(vp, blkno, size, 0, 0, 0); 74690075Sobrien 74790075Sobrien /* if not found in cache, do some I/O */ 74890075Sobrien if ((bp->b_flags & B_CACHE) == 0) { 74990075Sobrien if (curthread != PCPU_GET(idlethread)) 75090075Sobrien curthread->td_proc->p_stats->p_ru.ru_inblock++; 75190075Sobrien bp->b_iocmd = BIO_READ; 75290075Sobrien bp->b_flags &= ~B_INVAL; 75390075Sobrien bp->b_ioflags &= ~BIO_ERROR; 75490075Sobrien if (bp->b_rcred == NOCRED && cred != NOCRED) 75590075Sobrien bp->b_rcred = crhold(cred); 75690075Sobrien vfs_busy_pages(bp, 0); 75790075Sobrien bp->b_iooffset = dbtob(bp->b_blkno); 75890075Sobrien bstrategy(bp); 759132718Skan ++readwait; 760132718Skan } 761169689Skan 762132718Skan for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 763132718Skan if (inmem(vp, *rablkno)) 764132718Skan continue; 765132718Skan rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 766132718Skan 767132718Skan if ((rabp->b_flags & B_CACHE) == 0) { 768132718Skan if (curthread != PCPU_GET(idlethread)) 769132718Skan curthread->td_proc->p_stats->p_ru.ru_inblock++; 770132718Skan rabp->b_flags |= B_ASYNC; 771132718Skan rabp->b_flags &= ~B_INVAL; 772169689Skan rabp->b_ioflags &= ~BIO_ERROR; 773169689Skan rabp->b_iocmd = BIO_READ; 774132718Skan if (rabp->b_rcred == NOCRED && cred != NOCRED) 775132718Skan rabp->b_rcred = crhold(cred); 776132718Skan vfs_busy_pages(rabp, 0); 777132718Skan BUF_KERNPROC(rabp); 778132718Skan rabp->b_iooffset = dbtob(rabp->b_blkno); 779132718Skan bstrategy(rabp); 780132718Skan } else { 781132718Skan brelse(rabp); 782132718Skan } 783169689Skan } 784169689Skan 785132718Skan if (readwait) { 786132718Skan rv = bufwait(bp); 787132718Skan } 788132718Skan return (rv); 789132718Skan} 790132718Skan 791132718Skan/* 792132718Skan * Write, release buffer on completion. (Done by iodone 793132718Skan * if async). Do not bother writing anything if the buffer 794132718Skan * is invalid. 795 * 796 * Note that we set B_CACHE here, indicating that buffer is 797 * fully valid and thus cacheable. This is true even of NFS 798 * now so we set it generally. This could be set either here 799 * or in biodone() since the I/O is synchronous. We put it 800 * here. 801 */ 802int 803bufwrite(struct buf *bp) 804{ 805 int oldflags, s; 806 struct buf *newbp; 807 808 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 809 if (bp->b_flags & B_INVAL) { 810 brelse(bp); 811 return (0); 812 } 813 814 oldflags = bp->b_flags; 815 816 if (BUF_REFCNT(bp) == 0) 817 panic("bufwrite: buffer is not busy???"); 818 s = splbio(); 819 /* 820 * If a background write is already in progress, delay 821 * writing this block if it is asynchronous. Otherwise 822 * wait for the background write to complete. 823 */ 824 BO_LOCK(bp->b_bufobj); 825 if (bp->b_vflags & BV_BKGRDINPROG) { 826 if (bp->b_flags & B_ASYNC) { 827 BO_UNLOCK(bp->b_bufobj); 828 splx(s); 829 bdwrite(bp); 830 return (0); 831 } 832 bp->b_vflags |= BV_BKGRDWAIT; 833 msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), PRIBIO, "bwrbg", 0); 834 if (bp->b_vflags & BV_BKGRDINPROG) 835 panic("bufwrite: still writing"); 836 } 837 BO_UNLOCK(bp->b_bufobj); 838 839 /* Mark the buffer clean */ 840 bundirty(bp); 841 842 /* 843 * If this buffer is marked for background writing and we 844 * do not have to wait for it, make a copy and write the 845 * copy so as to leave this buffer ready for further use. 846 * 847 * This optimization eats a lot of memory. If we have a page 848 * or buffer shortfall we can't do it. 849 */ 850 if (dobkgrdwrite && (bp->b_xflags & BX_BKGRDWRITE) && 851 (bp->b_flags & B_ASYNC) && 852 !vm_page_count_severe() && 853 !buf_dirty_count_severe()) { 854 KASSERT(bp->b_iodone == NULL, 855 ("bufwrite: needs chained iodone (%p)", bp->b_iodone)); 856 857 /* get a new block */ 858 newbp = geteblk(bp->b_bufsize); 859 860 /* 861 * set it to be identical to the old block. We have to 862 * set b_lblkno and BKGRDMARKER before calling bgetvp() 863 * to avoid confusing the splay tree and gbincore(). 864 */ 865 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize); 866 newbp->b_lblkno = bp->b_lblkno; 867 newbp->b_xflags |= BX_BKGRDMARKER; 868 BO_LOCK(bp->b_bufobj); 869 bp->b_vflags |= BV_BKGRDINPROG; 870 bgetvp(bp->b_vp, newbp); 871 BO_UNLOCK(bp->b_bufobj); 872 newbp->b_bufobj = &bp->b_vp->v_bufobj; 873 newbp->b_blkno = bp->b_blkno; 874 newbp->b_offset = bp->b_offset; 875 newbp->b_iodone = vfs_backgroundwritedone; 876 newbp->b_flags |= B_ASYNC; 877 newbp->b_flags &= ~B_INVAL; 878 879 /* move over the dependencies */ 880 if (LIST_FIRST(&bp->b_dep) != NULL) 881 buf_movedeps(bp, newbp); 882 883 /* 884 * Initiate write on the copy, release the original to 885 * the B_LOCKED queue so that it cannot go away until 886 * the background write completes. If not locked it could go 887 * away and then be reconstituted while it was being written. 888 * If the reconstituted buffer were written, we could end up 889 * with two background copies being written at the same time. 890 */ 891 bqrelse(bp); 892 bp = newbp; 893 } 894 895 bp->b_flags &= ~B_DONE; 896 bp->b_ioflags &= ~BIO_ERROR; 897 bp->b_flags |= B_CACHE; 898 bp->b_iocmd = BIO_WRITE; 899 900 bufobj_wref(bp->b_bufobj); 901 vfs_busy_pages(bp, 1); 902 903 /* 904 * Normal bwrites pipeline writes 905 */ 906 bp->b_runningbufspace = bp->b_bufsize; 907 atomic_add_int(&runningbufspace, bp->b_runningbufspace); 908 909 if (curthread != PCPU_GET(idlethread)) 910 curthread->td_proc->p_stats->p_ru.ru_oublock++; 911 splx(s); 912 if (oldflags & B_ASYNC) 913 BUF_KERNPROC(bp); 914 bp->b_iooffset = dbtob(bp->b_blkno); 915 bstrategy(bp); 916 917 if ((oldflags & B_ASYNC) == 0) { 918 int rtval = bufwait(bp); 919 brelse(bp); 920 return (rtval); 921 } else { 922 /* 923 * don't allow the async write to saturate the I/O 924 * system. We will not deadlock here because 925 * we are blocking waiting for I/O that is already in-progress 926 * to complete. We do not block here if it is the update 927 * or syncer daemon trying to clean up as that can lead 928 * to deadlock. 929 */ 930 if (curthread->td_proc != bufdaemonproc && 931 curthread->td_proc != updateproc) 932 waitrunningbufspace(); 933 } 934 935 return (0); 936} 937 938/* 939 * Complete a background write started from bwrite. 940 */ 941static void 942vfs_backgroundwritedone(struct buf *bp) 943{ 944 struct buf *origbp; 945 946 /* 947 * Find the original buffer that we are writing. 948 */ 949 BO_LOCK(bp->b_bufobj); 950 if ((origbp = gbincore(bp->b_bufobj, bp->b_lblkno)) == NULL) 951 panic("backgroundwritedone: lost buffer"); 952 953 /* 954 * Clear the BV_BKGRDINPROG flag in the original buffer 955 * and awaken it if it is waiting for the write to complete. 956 * If BV_BKGRDINPROG is not set in the original buffer it must 957 * have been released and re-instantiated - which is not legal. 958 */ 959 KASSERT((origbp->b_vflags & BV_BKGRDINPROG), 960 ("backgroundwritedone: lost buffer2")); 961 origbp->b_vflags &= ~BV_BKGRDINPROG; 962 if (origbp->b_vflags & BV_BKGRDWAIT) { 963 origbp->b_vflags &= ~BV_BKGRDWAIT; 964 wakeup(&origbp->b_xflags); 965 } 966 BO_UNLOCK(bp->b_bufobj); 967 /* 968 * Process dependencies then return any unfinished ones. 969 */ 970 if (LIST_FIRST(&bp->b_dep) != NULL) 971 buf_complete(bp); 972 if (LIST_FIRST(&bp->b_dep) != NULL) 973 buf_movedeps(bp, origbp); 974 975 /* 976 * This buffer is marked B_NOCACHE, so when it is released 977 * by biodone, it will be tossed. We mark it with BIO_READ 978 * to avoid biodone doing a second bufobj_wdrop. 979 */ 980 bp->b_flags |= B_NOCACHE; 981 bp->b_iocmd = BIO_READ; 982 bp->b_flags &= ~(B_CACHE | B_DONE); 983 bp->b_iodone = 0; 984 bufdone(bp); 985} 986 987/* 988 * Delayed write. (Buffer is marked dirty). Do not bother writing 989 * anything if the buffer is marked invalid. 990 * 991 * Note that since the buffer must be completely valid, we can safely 992 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 993 * biodone() in order to prevent getblk from writing the buffer 994 * out synchronously. 995 */ 996void 997bdwrite(struct buf *bp) 998{ 999 struct thread *td = curthread; 1000 struct vnode *vp; 1001 struct buf *nbp; 1002 struct bufobj *bo; 1003 1004 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1005 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1006 KASSERT(BUF_REFCNT(bp) != 0, ("bdwrite: buffer is not busy")); 1007 1008 if (bp->b_flags & B_INVAL) { 1009 brelse(bp); 1010 return; 1011 } 1012 1013 /* 1014 * If we have too many dirty buffers, don't create any more. 1015 * If we are wildly over our limit, then force a complete 1016 * cleanup. Otherwise, just keep the situation from getting 1017 * out of control. Note that we have to avoid a recursive 1018 * disaster and not try to clean up after our own cleanup! 1019 */ 1020 vp = bp->b_vp; 1021 bo = bp->b_bufobj; 1022 if ((td->td_pflags & TDP_COWINPROGRESS) == 0) { 1023 BO_LOCK(bo); 1024 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 1025 BO_UNLOCK(bo); 1026 (void) VOP_FSYNC(vp, MNT_NOWAIT, td); 1027 altbufferflushes++; 1028 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 1029 /* 1030 * Try to find a buffer to flush. 1031 */ 1032 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 1033 if ((nbp->b_vflags & BV_BKGRDINPROG) || 1034 BUF_LOCK(nbp, 1035 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 1036 continue; 1037 if (bp == nbp) 1038 panic("bdwrite: found ourselves"); 1039 BO_UNLOCK(bo); 1040 /* Don't countdeps with the bo lock held. */ 1041 if (buf_countdeps(nbp, 0)) { 1042 BO_LOCK(bo); 1043 BUF_UNLOCK(nbp); 1044 continue; 1045 } 1046 if (nbp->b_flags & B_CLUSTEROK) { 1047 vfs_bio_awrite(nbp); 1048 } else { 1049 bremfree(nbp); 1050 bawrite(nbp); 1051 } 1052 dirtybufferflushes++; 1053 break; 1054 } 1055 if (nbp == NULL) 1056 BO_UNLOCK(bo); 1057 } else 1058 BO_UNLOCK(bo); 1059 } else 1060 recursiveflushes++; 1061 1062 bdirty(bp); 1063 /* 1064 * Set B_CACHE, indicating that the buffer is fully valid. This is 1065 * true even of NFS now. 1066 */ 1067 bp->b_flags |= B_CACHE; 1068 1069 /* 1070 * This bmap keeps the system from needing to do the bmap later, 1071 * perhaps when the system is attempting to do a sync. Since it 1072 * is likely that the indirect block -- or whatever other datastructure 1073 * that the filesystem needs is still in memory now, it is a good 1074 * thing to do this. Note also, that if the pageout daemon is 1075 * requesting a sync -- there might not be enough memory to do 1076 * the bmap then... So, this is important to do. 1077 */ 1078 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 1079 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 1080 } 1081 1082 /* 1083 * Set the *dirty* buffer range based upon the VM system dirty pages. 1084 */ 1085 vfs_setdirty(bp); 1086 1087 /* 1088 * We need to do this here to satisfy the vnode_pager and the 1089 * pageout daemon, so that it thinks that the pages have been 1090 * "cleaned". Note that since the pages are in a delayed write 1091 * buffer -- the VFS layer "will" see that the pages get written 1092 * out on the next sync, or perhaps the cluster will be completed. 1093 */ 1094 vfs_clean_pages(bp); 1095 bqrelse(bp); 1096 1097 /* 1098 * Wakeup the buffer flushing daemon if we have a lot of dirty 1099 * buffers (midpoint between our recovery point and our stall 1100 * point). 1101 */ 1102 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1103 1104 /* 1105 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 1106 * due to the softdep code. 1107 */ 1108} 1109 1110/* 1111 * bdirty: 1112 * 1113 * Turn buffer into delayed write request. We must clear BIO_READ and 1114 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 1115 * itself to properly update it in the dirty/clean lists. We mark it 1116 * B_DONE to ensure that any asynchronization of the buffer properly 1117 * clears B_DONE ( else a panic will occur later ). 1118 * 1119 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 1120 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 1121 * should only be called if the buffer is known-good. 1122 * 1123 * Since the buffer is not on a queue, we do not update the numfreebuffers 1124 * count. 1125 * 1126 * Must be called at splbio(). 1127 * The buffer must be on QUEUE_NONE. 1128 */ 1129void 1130bdirty(struct buf *bp) 1131{ 1132 1133 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 1134 bp, bp->b_vp, bp->b_flags); 1135 KASSERT(BUF_REFCNT(bp) == 1, ("bdirty: bp %p not locked",bp)); 1136 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1137 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1138 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1139 bp->b_flags &= ~(B_RELBUF); 1140 bp->b_iocmd = BIO_WRITE; 1141 1142 if ((bp->b_flags & B_DELWRI) == 0) { 1143 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 1144 reassignbuf(bp); 1145 atomic_add_int(&numdirtybuffers, 1); 1146 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2); 1147 } 1148} 1149 1150/* 1151 * bundirty: 1152 * 1153 * Clear B_DELWRI for buffer. 1154 * 1155 * Since the buffer is not on a queue, we do not update the numfreebuffers 1156 * count. 1157 * 1158 * Must be called at splbio(). 1159 * The buffer must be on QUEUE_NONE. 1160 */ 1161 1162void 1163bundirty(struct buf *bp) 1164{ 1165 1166 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1167 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 1168 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 1169 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 1170 KASSERT(BUF_REFCNT(bp) == 1, ("bundirty: bp %p not locked",bp)); 1171 1172 if (bp->b_flags & B_DELWRI) { 1173 bp->b_flags &= ~B_DELWRI; 1174 reassignbuf(bp); 1175 atomic_subtract_int(&numdirtybuffers, 1); 1176 numdirtywakeup(lodirtybuffers); 1177 } 1178 /* 1179 * Since it is now being written, we can clear its deferred write flag. 1180 */ 1181 bp->b_flags &= ~B_DEFERRED; 1182} 1183 1184/* 1185 * bawrite: 1186 * 1187 * Asynchronous write. Start output on a buffer, but do not wait for 1188 * it to complete. The buffer is released when the output completes. 1189 * 1190 * bwrite() ( or the VOP routine anyway ) is responsible for handling 1191 * B_INVAL buffers. Not us. 1192 */ 1193void 1194bawrite(struct buf *bp) 1195{ 1196 1197 bp->b_flags |= B_ASYNC; 1198 (void) bwrite(bp); 1199} 1200 1201/* 1202 * bwillwrite: 1203 * 1204 * Called prior to the locking of any vnodes when we are expecting to 1205 * write. We do not want to starve the buffer cache with too many 1206 * dirty buffers so we block here. By blocking prior to the locking 1207 * of any vnodes we attempt to avoid the situation where a locked vnode 1208 * prevents the various system daemons from flushing related buffers. 1209 */ 1210 1211void 1212bwillwrite(void) 1213{ 1214 1215 if (numdirtybuffers >= hidirtybuffers) { 1216 int s; 1217 1218 s = splbio(); 1219 mtx_lock(&nblock); 1220 while (numdirtybuffers >= hidirtybuffers) { 1221 bd_wakeup(1); 1222 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH; 1223 msleep(&needsbuffer, &nblock, 1224 (PRIBIO + 4), "flswai", 0); 1225 } 1226 splx(s); 1227 mtx_unlock(&nblock); 1228 } 1229} 1230 1231/* 1232 * Return true if we have too many dirty buffers. 1233 */ 1234int 1235buf_dirty_count_severe(void) 1236{ 1237 1238 return(numdirtybuffers >= hidirtybuffers); 1239} 1240 1241/* 1242 * brelse: 1243 * 1244 * Release a busy buffer and, if requested, free its resources. The 1245 * buffer will be stashed in the appropriate bufqueue[] allowing it 1246 * to be accessed later as a cache entity or reused for other purposes. 1247 */ 1248void 1249brelse(struct buf *bp) 1250{ 1251 int s; 1252 1253 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 1254 bp, bp->b_vp, bp->b_flags); 1255 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1256 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1257 1258 s = splbio(); 1259 1260 if (bp->b_iocmd == BIO_WRITE && 1261 (bp->b_ioflags & BIO_ERROR) && 1262 !(bp->b_flags & B_INVAL)) { 1263 /* 1264 * Failed write, redirty. Must clear BIO_ERROR to prevent 1265 * pages from being scrapped. If B_INVAL is set then 1266 * this case is not run and the next case is run to 1267 * destroy the buffer. B_INVAL can occur if the buffer 1268 * is outside the range supported by the underlying device. 1269 */ 1270 bp->b_ioflags &= ~BIO_ERROR; 1271 bdirty(bp); 1272 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 1273 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 1274 /* 1275 * Either a failed I/O or we were asked to free or not 1276 * cache the buffer. 1277 */ 1278 bp->b_flags |= B_INVAL; 1279 if (LIST_FIRST(&bp->b_dep) != NULL) 1280 buf_deallocate(bp); 1281 if (bp->b_flags & B_DELWRI) { 1282 atomic_subtract_int(&numdirtybuffers, 1); 1283 numdirtywakeup(lodirtybuffers); 1284 } 1285 bp->b_flags &= ~(B_DELWRI | B_CACHE); 1286 if ((bp->b_flags & B_VMIO) == 0) { 1287 if (bp->b_bufsize) 1288 allocbuf(bp, 0); 1289 if (bp->b_vp) 1290 brelvp(bp); 1291 } 1292 } 1293 1294 /* 1295 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 1296 * is called with B_DELWRI set, the underlying pages may wind up 1297 * getting freed causing a previous write (bdwrite()) to get 'lost' 1298 * because pages associated with a B_DELWRI bp are marked clean. 1299 * 1300 * We still allow the B_INVAL case to call vfs_vmio_release(), even 1301 * if B_DELWRI is set. 1302 * 1303 * If B_DELWRI is not set we may have to set B_RELBUF if we are low 1304 * on pages to return pages to the VM page queues. 1305 */ 1306 if (bp->b_flags & B_DELWRI) 1307 bp->b_flags &= ~B_RELBUF; 1308 else if (vm_page_count_severe()) { 1309 /* 1310 * XXX This lock may not be necessary since BKGRDINPROG 1311 * cannot be set while we hold the buf lock, it can only be 1312 * cleared if it is already pending. 1313 */ 1314 if (bp->b_vp) { 1315 BO_LOCK(bp->b_bufobj); 1316 if (!(bp->b_vflags & BV_BKGRDINPROG)) 1317 bp->b_flags |= B_RELBUF; 1318 BO_UNLOCK(bp->b_bufobj); 1319 } else 1320 bp->b_flags |= B_RELBUF; 1321 } 1322 1323 /* 1324 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 1325 * constituted, not even NFS buffers now. Two flags effect this. If 1326 * B_INVAL, the struct buf is invalidated but the VM object is kept 1327 * around ( i.e. so it is trivial to reconstitute the buffer later ). 1328 * 1329 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 1330 * invalidated. BIO_ERROR cannot be set for a failed write unless the 1331 * buffer is also B_INVAL because it hits the re-dirtying code above. 1332 * 1333 * Normally we can do this whether a buffer is B_DELWRI or not. If 1334 * the buffer is an NFS buffer, it is tracking piecemeal writes or 1335 * the commit state and we cannot afford to lose the buffer. If the 1336 * buffer has a background write in progress, we need to keep it 1337 * around to prevent it from being reconstituted and starting a second 1338 * background write. 1339 */ 1340 if ((bp->b_flags & B_VMIO) 1341 && !(bp->b_vp->v_mount != NULL && 1342 (bp->b_vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 1343 !vn_isdisk(bp->b_vp, NULL) && 1344 (bp->b_flags & B_DELWRI)) 1345 ) { 1346 1347 int i, j, resid; 1348 vm_page_t m; 1349 off_t foff; 1350 vm_pindex_t poff; 1351 vm_object_t obj; 1352 1353 obj = bp->b_bufobj->bo_object; 1354 1355 /* 1356 * Get the base offset and length of the buffer. Note that 1357 * in the VMIO case if the buffer block size is not 1358 * page-aligned then b_data pointer may not be page-aligned. 1359 * But our b_pages[] array *IS* page aligned. 1360 * 1361 * block sizes less then DEV_BSIZE (usually 512) are not 1362 * supported due to the page granularity bits (m->valid, 1363 * m->dirty, etc...). 1364 * 1365 * See man buf(9) for more information 1366 */ 1367 resid = bp->b_bufsize; 1368 foff = bp->b_offset; 1369 VM_OBJECT_LOCK(obj); 1370 for (i = 0; i < bp->b_npages; i++) { 1371 int had_bogus = 0; 1372 1373 m = bp->b_pages[i]; 1374 1375 /* 1376 * If we hit a bogus page, fixup *all* the bogus pages 1377 * now. 1378 */ 1379 if (m == bogus_page) { 1380 poff = OFF_TO_IDX(bp->b_offset); 1381 had_bogus = 1; 1382 1383 for (j = i; j < bp->b_npages; j++) { 1384 vm_page_t mtmp; 1385 mtmp = bp->b_pages[j]; 1386 if (mtmp == bogus_page) { 1387 mtmp = vm_page_lookup(obj, poff + j); 1388 if (!mtmp) { 1389 panic("brelse: page missing\n"); 1390 } 1391 bp->b_pages[j] = mtmp; 1392 } 1393 } 1394 1395 if ((bp->b_flags & B_INVAL) == 0) { 1396 pmap_qenter( 1397 trunc_page((vm_offset_t)bp->b_data), 1398 bp->b_pages, bp->b_npages); 1399 } 1400 m = bp->b_pages[i]; 1401 } 1402 if ((bp->b_flags & B_NOCACHE) || 1403 (bp->b_ioflags & BIO_ERROR)) { 1404 int poffset = foff & PAGE_MASK; 1405 int presid = resid > (PAGE_SIZE - poffset) ? 1406 (PAGE_SIZE - poffset) : resid; 1407 1408 KASSERT(presid >= 0, ("brelse: extra page")); 1409 vm_page_lock_queues(); 1410 vm_page_set_invalid(m, poffset, presid); 1411 vm_page_unlock_queues(); 1412 if (had_bogus) 1413 printf("avoided corruption bug in bogus_page/brelse code\n"); 1414 } 1415 resid -= PAGE_SIZE - (foff & PAGE_MASK); 1416 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 1417 } 1418 VM_OBJECT_UNLOCK(obj); 1419 if (bp->b_flags & (B_INVAL | B_RELBUF)) 1420 vfs_vmio_release(bp); 1421 1422 } else if (bp->b_flags & B_VMIO) { 1423 1424 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 1425 vfs_vmio_release(bp); 1426 } 1427 1428 } 1429 1430 if (BUF_REFCNT(bp) > 1) { 1431 /* do not release to free list */ 1432 BUF_UNLOCK(bp); 1433 splx(s); 1434 return; 1435 } 1436 1437 /* enqueue */ 1438 mtx_lock(&bqlock); 1439 /* Handle delayed bremfree() processing. */ 1440 if (bp->b_flags & B_REMFREE) 1441 bremfreel(bp); 1442 if (bp->b_qindex != QUEUE_NONE) 1443 panic("brelse: free buffer onto another queue???"); 1444 1445 /* buffers with no memory */ 1446 if (bp->b_bufsize == 0) { 1447 bp->b_flags |= B_INVAL; 1448 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1449 if (bp->b_vflags & BV_BKGRDINPROG) 1450 panic("losing buffer 1"); 1451 if (bp->b_kvasize) { 1452 bp->b_qindex = QUEUE_EMPTYKVA; 1453 } else { 1454 bp->b_qindex = QUEUE_EMPTY; 1455 } 1456 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1457 /* buffers with junk contents */ 1458 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 1459 (bp->b_ioflags & BIO_ERROR)) { 1460 bp->b_flags |= B_INVAL; 1461 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 1462 if (bp->b_vflags & BV_BKGRDINPROG) 1463 panic("losing buffer 2"); 1464 bp->b_qindex = QUEUE_CLEAN; 1465 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist); 1466 /* remaining buffers */ 1467 } else { 1468 if (bp->b_flags & B_DELWRI) 1469 bp->b_qindex = QUEUE_DIRTY; 1470 else 1471 bp->b_qindex = QUEUE_CLEAN; 1472 if (bp->b_flags & B_AGE) 1473 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist); 1474 else 1475 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist); 1476 } 1477 mtx_unlock(&bqlock); 1478 1479 /* 1480 * If B_INVAL and B_DELWRI is set, clear B_DELWRI. We have already 1481 * placed the buffer on the correct queue. We must also disassociate 1482 * the device and vnode for a B_INVAL buffer so gbincore() doesn't 1483 * find it. 1484 */ 1485 if (bp->b_flags & B_INVAL) { 1486 if (bp->b_flags & B_DELWRI) 1487 bundirty(bp); 1488 if (bp->b_vp) 1489 brelvp(bp); 1490 } 1491 1492 /* 1493 * Fixup numfreebuffers count. The bp is on an appropriate queue 1494 * unless locked. We then bump numfreebuffers if it is not B_DELWRI. 1495 * We've already handled the B_INVAL case ( B_DELWRI will be clear 1496 * if B_INVAL is set ). 1497 */ 1498 1499 if (!(bp->b_flags & B_DELWRI)) 1500 bufcountwakeup(); 1501 1502 /* 1503 * Something we can maybe free or reuse 1504 */ 1505 if (bp->b_bufsize || bp->b_kvasize) 1506 bufspacewakeup(); 1507 1508 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF | B_DIRECT); 1509 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1510 panic("brelse: not dirty"); 1511 /* unlock */ 1512 BUF_UNLOCK(bp); 1513 splx(s); 1514} 1515 1516/* 1517 * Release a buffer back to the appropriate queue but do not try to free 1518 * it. The buffer is expected to be used again soon. 1519 * 1520 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 1521 * biodone() to requeue an async I/O on completion. It is also used when 1522 * known good buffers need to be requeued but we think we may need the data 1523 * again soon. 1524 * 1525 * XXX we should be able to leave the B_RELBUF hint set on completion. 1526 */ 1527void 1528bqrelse(struct buf *bp) 1529{ 1530 int s; 1531 1532 s = splbio(); 1533 1534 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1535 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 1536 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 1537 1538 if (BUF_REFCNT(bp) > 1) { 1539 /* do not release to free list */ 1540 BUF_UNLOCK(bp); 1541 splx(s); 1542 return; 1543 } 1544 mtx_lock(&bqlock); 1545 /* Handle delayed bremfree() processing. */ 1546 if (bp->b_flags & B_REMFREE) 1547 bremfreel(bp); 1548 if (bp->b_qindex != QUEUE_NONE) 1549 panic("bqrelse: free buffer onto another queue???"); 1550 /* buffers with stale but valid contents */ 1551 if (bp->b_flags & B_DELWRI) { 1552 bp->b_qindex = QUEUE_DIRTY; 1553 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist); 1554 } else { 1555 /* 1556 * XXX This lock may not be necessary since BKGRDINPROG 1557 * cannot be set while we hold the buf lock, it can only be 1558 * cleared if it is already pending. 1559 */ 1560 BO_LOCK(bp->b_bufobj); 1561 if (!vm_page_count_severe() || bp->b_vflags & BV_BKGRDINPROG) { 1562 BO_UNLOCK(bp->b_bufobj); 1563 bp->b_qindex = QUEUE_CLEAN; 1564 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, 1565 b_freelist); 1566 } else { 1567 /* 1568 * We are too low on memory, we have to try to free 1569 * the buffer (most importantly: the wired pages 1570 * making up its backing store) *now*. 1571 */ 1572 BO_UNLOCK(bp->b_bufobj); 1573 mtx_unlock(&bqlock); 1574 splx(s); 1575 brelse(bp); 1576 return; 1577 } 1578 } 1579 mtx_unlock(&bqlock); 1580 1581 if ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI)) 1582 bufcountwakeup(); 1583 1584 /* 1585 * Something we can maybe free or reuse. 1586 */ 1587 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI)) 1588 bufspacewakeup(); 1589 1590 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 1591 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 1592 panic("bqrelse: not dirty"); 1593 /* unlock */ 1594 BUF_UNLOCK(bp); 1595 splx(s); 1596} 1597 1598/* Give pages used by the bp back to the VM system (where possible) */ 1599static void 1600vfs_vmio_release(struct buf *bp) 1601{ 1602 int i; 1603 vm_page_t m; 1604 1605 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 1606 vm_page_lock_queues(); 1607 for (i = 0; i < bp->b_npages; i++) { 1608 m = bp->b_pages[i]; 1609 bp->b_pages[i] = NULL; 1610 /* 1611 * In order to keep page LRU ordering consistent, put 1612 * everything on the inactive queue. 1613 */ 1614 vm_page_unwire(m, 0); 1615 /* 1616 * We don't mess with busy pages, it is 1617 * the responsibility of the process that 1618 * busied the pages to deal with them. 1619 */ 1620 if ((m->flags & PG_BUSY) || (m->busy != 0)) 1621 continue; 1622 1623 if (m->wire_count == 0) { 1624 /* 1625 * Might as well free the page if we can and it has 1626 * no valid data. We also free the page if the 1627 * buffer was used for direct I/O 1628 */ 1629 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && 1630 m->hold_count == 0) { 1631 pmap_remove_all(m); 1632 vm_page_free(m); 1633 } else if (bp->b_flags & B_DIRECT) { 1634 vm_page_try_to_free(m); 1635 } else if (vm_page_count_severe()) { 1636 vm_page_try_to_cache(m); 1637 } 1638 } 1639 } 1640 vm_page_unlock_queues(); 1641 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 1642 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 1643 1644 if (bp->b_bufsize) { 1645 bufspacewakeup(); 1646 bp->b_bufsize = 0; 1647 } 1648 bp->b_npages = 0; 1649 bp->b_flags &= ~B_VMIO; 1650 if (bp->b_vp) 1651 brelvp(bp); 1652} 1653 1654/* 1655 * Check to see if a block at a particular lbn is available for a clustered 1656 * write. 1657 */ 1658static int 1659vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 1660{ 1661 struct buf *bpa; 1662 int match; 1663 1664 match = 0; 1665 1666 /* If the buf isn't in core skip it */ 1667 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 1668 return (0); 1669 1670 /* If the buf is busy we don't want to wait for it */ 1671 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1672 return (0); 1673 1674 /* Only cluster with valid clusterable delayed write buffers */ 1675 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 1676 (B_DELWRI | B_CLUSTEROK)) 1677 goto done; 1678 1679 if (bpa->b_bufsize != size) 1680 goto done; 1681 1682 /* 1683 * Check to see if it is in the expected place on disk and that the 1684 * block has been mapped. 1685 */ 1686 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 1687 match = 1; 1688done: 1689 BUF_UNLOCK(bpa); 1690 return (match); 1691} 1692 1693/* 1694 * vfs_bio_awrite: 1695 * 1696 * Implement clustered async writes for clearing out B_DELWRI buffers. 1697 * This is much better then the old way of writing only one buffer at 1698 * a time. Note that we may not be presented with the buffers in the 1699 * correct order, so we search for the cluster in both directions. 1700 */ 1701int 1702vfs_bio_awrite(struct buf *bp) 1703{ 1704 int i; 1705 int j; 1706 daddr_t lblkno = bp->b_lblkno; 1707 struct vnode *vp = bp->b_vp; 1708 int s; 1709 int ncl; 1710 int nwritten; 1711 int size; 1712 int maxcl; 1713 1714 s = splbio(); 1715 /* 1716 * right now we support clustered writing only to regular files. If 1717 * we find a clusterable block we could be in the middle of a cluster 1718 * rather then at the beginning. 1719 */ 1720 if ((vp->v_type == VREG) && 1721 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 1722 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 1723 1724 size = vp->v_mount->mnt_stat.f_iosize; 1725 maxcl = MAXPHYS / size; 1726 1727 VI_LOCK(vp); 1728 for (i = 1; i < maxcl; i++) 1729 if (vfs_bio_clcheck(vp, size, lblkno + i, 1730 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 1731 break; 1732 1733 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 1734 if (vfs_bio_clcheck(vp, size, lblkno - j, 1735 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 1736 break; 1737 1738 VI_UNLOCK(vp); 1739 --j; 1740 ncl = i + j; 1741 /* 1742 * this is a possible cluster write 1743 */ 1744 if (ncl != 1) { 1745 BUF_UNLOCK(bp); 1746 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl); 1747 splx(s); 1748 return nwritten; 1749 } 1750 } 1751 1752 bremfree(bp); 1753 bp->b_flags |= B_ASYNC; 1754 1755 splx(s); 1756 /* 1757 * default (old) behavior, writing out only one block 1758 * 1759 * XXX returns b_bufsize instead of b_bcount for nwritten? 1760 */ 1761 nwritten = bp->b_bufsize; 1762 (void) bwrite(bp); 1763 1764 return nwritten; 1765} 1766 1767/* 1768 * getnewbuf: 1769 * 1770 * Find and initialize a new buffer header, freeing up existing buffers 1771 * in the bufqueues as necessary. The new buffer is returned locked. 1772 * 1773 * Important: B_INVAL is not set. If the caller wishes to throw the 1774 * buffer away, the caller must set B_INVAL prior to calling brelse(). 1775 * 1776 * We block if: 1777 * We have insufficient buffer headers 1778 * We have insufficient buffer space 1779 * buffer_map is too fragmented ( space reservation fails ) 1780 * If we have to flush dirty buffers ( but we try to avoid this ) 1781 * 1782 * To avoid VFS layer recursion we do not flush dirty buffers ourselves. 1783 * Instead we ask the buf daemon to do it for us. We attempt to 1784 * avoid piecemeal wakeups of the pageout daemon. 1785 */ 1786 1787static struct buf * 1788getnewbuf(int slpflag, int slptimeo, int size, int maxsize) 1789{ 1790 struct buf *bp; 1791 struct buf *nbp; 1792 int defrag = 0; 1793 int nqindex; 1794 static int flushingbufs; 1795 1796 /* 1797 * We can't afford to block since we might be holding a vnode lock, 1798 * which may prevent system daemons from running. We deal with 1799 * low-memory situations by proactively returning memory and running 1800 * async I/O rather then sync I/O. 1801 */ 1802 1803 atomic_add_int(&getnewbufcalls, 1); 1804 atomic_subtract_int(&getnewbufrestarts, 1); 1805restart: 1806 atomic_add_int(&getnewbufrestarts, 1); 1807 1808 /* 1809 * Setup for scan. If we do not have enough free buffers, 1810 * we setup a degenerate case that immediately fails. Note 1811 * that if we are specially marked process, we are allowed to 1812 * dip into our reserves. 1813 * 1814 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN 1815 * 1816 * We start with EMPTYKVA. If the list is empty we backup to EMPTY. 1817 * However, there are a number of cases (defragging, reusing, ...) 1818 * where we cannot backup. 1819 */ 1820 mtx_lock(&bqlock); 1821 nqindex = QUEUE_EMPTYKVA; 1822 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]); 1823 1824 if (nbp == NULL) { 1825 /* 1826 * If no EMPTYKVA buffers and we are either 1827 * defragging or reusing, locate a CLEAN buffer 1828 * to free or reuse. If bufspace useage is low 1829 * skip this step so we can allocate a new buffer. 1830 */ 1831 if (defrag || bufspace >= lobufspace) { 1832 nqindex = QUEUE_CLEAN; 1833 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]); 1834 } 1835 1836 /* 1837 * If we could not find or were not allowed to reuse a 1838 * CLEAN buffer, check to see if it is ok to use an EMPTY 1839 * buffer. We can only use an EMPTY buffer if allocating 1840 * its KVA would not otherwise run us out of buffer space. 1841 */ 1842 if (nbp == NULL && defrag == 0 && 1843 bufspace + maxsize < hibufspace) { 1844 nqindex = QUEUE_EMPTY; 1845 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1846 } 1847 } 1848 1849 /* 1850 * Run scan, possibly freeing data and/or kva mappings on the fly 1851 * depending. 1852 */ 1853 1854 while ((bp = nbp) != NULL) { 1855 int qindex = nqindex; 1856 1857 /* 1858 * Calculate next bp ( we can only use it if we do not block 1859 * or do other fancy things ). 1860 */ 1861 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) { 1862 switch(qindex) { 1863 case QUEUE_EMPTY: 1864 nqindex = QUEUE_EMPTYKVA; 1865 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]))) 1866 break; 1867 /* FALLTHROUGH */ 1868 case QUEUE_EMPTYKVA: 1869 nqindex = QUEUE_CLEAN; 1870 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]))) 1871 break; 1872 /* FALLTHROUGH */ 1873 case QUEUE_CLEAN: 1874 /* 1875 * nbp is NULL. 1876 */ 1877 break; 1878 } 1879 } 1880 /* 1881 * If we are defragging then we need a buffer with 1882 * b_kvasize != 0. XXX this situation should no longer 1883 * occur, if defrag is non-zero the buffer's b_kvasize 1884 * should also be non-zero at this point. XXX 1885 */ 1886 if (defrag && bp->b_kvasize == 0) { 1887 printf("Warning: defrag empty buffer %p\n", bp); 1888 continue; 1889 } 1890 1891 /* 1892 * Start freeing the bp. This is somewhat involved. nbp 1893 * remains valid only for QUEUE_EMPTY[KVA] bp's. 1894 */ 1895 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1896 continue; 1897 if (bp->b_vp) { 1898 BO_LOCK(bp->b_bufobj); 1899 if (bp->b_vflags & BV_BKGRDINPROG) { 1900 BO_UNLOCK(bp->b_bufobj); 1901 BUF_UNLOCK(bp); 1902 continue; 1903 } 1904 BO_UNLOCK(bp->b_bufobj); 1905 } 1906 CTR3(KTR_BUF, "getnewbuf(%p) vp %p flags %X (recycling)", 1907 bp, bp->b_vp, bp->b_flags); 1908 1909 /* 1910 * Sanity Checks 1911 */ 1912 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp)); 1913 1914 /* 1915 * Note: we no longer distinguish between VMIO and non-VMIO 1916 * buffers. 1917 */ 1918 1919 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex)); 1920 1921 bremfreel(bp); 1922 mtx_unlock(&bqlock); 1923 1924 if (qindex == QUEUE_CLEAN) { 1925 if (bp->b_flags & B_VMIO) { 1926 bp->b_flags &= ~B_ASYNC; 1927 vfs_vmio_release(bp); 1928 } 1929 if (bp->b_vp) 1930 brelvp(bp); 1931 } 1932 1933 /* 1934 * NOTE: nbp is now entirely invalid. We can only restart 1935 * the scan from this point on. 1936 * 1937 * Get the rest of the buffer freed up. b_kva* is still 1938 * valid after this operation. 1939 */ 1940 1941 if (bp->b_rcred != NOCRED) { 1942 crfree(bp->b_rcred); 1943 bp->b_rcred = NOCRED; 1944 } 1945 if (bp->b_wcred != NOCRED) { 1946 crfree(bp->b_wcred); 1947 bp->b_wcred = NOCRED; 1948 } 1949 if (LIST_FIRST(&bp->b_dep) != NULL) 1950 buf_deallocate(bp); 1951 if (bp->b_vflags & BV_BKGRDINPROG) 1952 panic("losing buffer 3"); 1953 1954 if (bp->b_bufsize) 1955 allocbuf(bp, 0); 1956 1957 bp->b_flags = 0; 1958 bp->b_ioflags = 0; 1959 bp->b_xflags = 0; 1960 bp->b_vflags = 0; 1961 bp->b_vp = NULL; 1962 bp->b_blkno = bp->b_lblkno = 0; 1963 bp->b_offset = NOOFFSET; 1964 bp->b_iodone = 0; 1965 bp->b_error = 0; 1966 bp->b_resid = 0; 1967 bp->b_bcount = 0; 1968 bp->b_npages = 0; 1969 bp->b_dirtyoff = bp->b_dirtyend = 0; 1970 bp->b_bufobj = NULL; 1971 1972 LIST_INIT(&bp->b_dep); 1973 1974 /* 1975 * If we are defragging then free the buffer. 1976 */ 1977 if (defrag) { 1978 bp->b_flags |= B_INVAL; 1979 bfreekva(bp); 1980 brelse(bp); 1981 defrag = 0; 1982 goto restart; 1983 } 1984 1985 /* 1986 * If we are overcomitted then recover the buffer and its 1987 * KVM space. This occurs in rare situations when multiple 1988 * processes are blocked in getnewbuf() or allocbuf(). 1989 */ 1990 if (bufspace >= hibufspace) 1991 flushingbufs = 1; 1992 if (flushingbufs && bp->b_kvasize != 0) { 1993 bp->b_flags |= B_INVAL; 1994 bfreekva(bp); 1995 brelse(bp); 1996 goto restart; 1997 } 1998 if (bufspace < lobufspace) 1999 flushingbufs = 0; 2000 break; 2001 } 2002 2003 /* 2004 * If we exhausted our list, sleep as appropriate. We may have to 2005 * wakeup various daemons and write out some dirty buffers. 2006 * 2007 * Generally we are sleeping due to insufficient buffer space. 2008 */ 2009 2010 if (bp == NULL) { 2011 int flags; 2012 char *waitmsg; 2013 2014 mtx_unlock(&bqlock); 2015 if (defrag) { 2016 flags = VFS_BIO_NEED_BUFSPACE; 2017 waitmsg = "nbufkv"; 2018 } else if (bufspace >= hibufspace) { 2019 waitmsg = "nbufbs"; 2020 flags = VFS_BIO_NEED_BUFSPACE; 2021 } else { 2022 waitmsg = "newbuf"; 2023 flags = VFS_BIO_NEED_ANY; 2024 } 2025 2026 bd_speedup(); /* heeeelp */ 2027 2028 mtx_lock(&nblock); 2029 needsbuffer |= flags; 2030 while (needsbuffer & flags) { 2031 if (msleep(&needsbuffer, &nblock, 2032 (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) { 2033 mtx_unlock(&nblock); 2034 return (NULL); 2035 } 2036 } 2037 mtx_unlock(&nblock); 2038 } else { 2039 /* 2040 * We finally have a valid bp. We aren't quite out of the 2041 * woods, we still have to reserve kva space. In order 2042 * to keep fragmentation sane we only allocate kva in 2043 * BKVASIZE chunks. 2044 */ 2045 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 2046 2047 if (maxsize != bp->b_kvasize) { 2048 vm_offset_t addr = 0; 2049 2050 bfreekva(bp); 2051 2052 if (vm_map_findspace(buffer_map, 2053 vm_map_min(buffer_map), maxsize, &addr)) { 2054 /* 2055 * Uh oh. Buffer map is to fragmented. We 2056 * must defragment the map. 2057 */ 2058 atomic_add_int(&bufdefragcnt, 1); 2059 defrag = 1; 2060 bp->b_flags |= B_INVAL; 2061 brelse(bp); 2062 goto restart; 2063 } 2064 if (addr) { 2065 vm_map_insert(buffer_map, NULL, 0, 2066 addr, addr + maxsize, 2067 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 2068 2069 bp->b_kvabase = (caddr_t) addr; 2070 bp->b_kvasize = maxsize; 2071 atomic_add_int(&bufspace, bp->b_kvasize); 2072 atomic_add_int(&bufreusecnt, 1); 2073 } 2074 } 2075 bp->b_saveaddr = bp->b_kvabase; 2076 bp->b_data = bp->b_saveaddr; 2077 } 2078 return(bp); 2079} 2080 2081/* 2082 * buf_daemon: 2083 * 2084 * buffer flushing daemon. Buffers are normally flushed by the 2085 * update daemon but if it cannot keep up this process starts to 2086 * take the load in an attempt to prevent getnewbuf() from blocking. 2087 */ 2088 2089static struct kproc_desc buf_kp = { 2090 "bufdaemon", 2091 buf_daemon, 2092 &bufdaemonproc 2093}; 2094SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp) 2095 2096static void 2097buf_daemon() 2098{ 2099 int s; 2100 2101 mtx_lock(&Giant); 2102 2103 /* 2104 * This process needs to be suspended prior to shutdown sync. 2105 */ 2106 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, bufdaemonproc, 2107 SHUTDOWN_PRI_LAST); 2108 2109 /* 2110 * This process is allowed to take the buffer cache to the limit 2111 */ 2112 s = splbio(); 2113 mtx_lock(&bdlock); 2114 2115 for (;;) { 2116 bd_request = 0; 2117 mtx_unlock(&bdlock); 2118 2119 kthread_suspend_check(bufdaemonproc); 2120 2121 /* 2122 * Do the flush. Limit the amount of in-transit I/O we 2123 * allow to build up, otherwise we would completely saturate 2124 * the I/O system. Wakeup any waiting processes before we 2125 * normally would so they can run in parallel with our drain. 2126 */ 2127 while (numdirtybuffers > lodirtybuffers) { 2128 if (flushbufqueues(0) == 0) { 2129 /* 2130 * Could not find any buffers without rollback 2131 * dependencies, so just write the first one 2132 * in the hopes of eventually making progress. 2133 */ 2134 flushbufqueues(1); 2135 break; 2136 } 2137 waitrunningbufspace(); 2138 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2); 2139 } 2140 2141 /* 2142 * Only clear bd_request if we have reached our low water 2143 * mark. The buf_daemon normally waits 1 second and 2144 * then incrementally flushes any dirty buffers that have 2145 * built up, within reason. 2146 * 2147 * If we were unable to hit our low water mark and couldn't 2148 * find any flushable buffers, we sleep half a second. 2149 * Otherwise we loop immediately. 2150 */ 2151 mtx_lock(&bdlock); 2152 if (numdirtybuffers <= lodirtybuffers) { 2153 /* 2154 * We reached our low water mark, reset the 2155 * request and sleep until we are needed again. 2156 * The sleep is just so the suspend code works. 2157 */ 2158 bd_request = 0; 2159 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 2160 } else { 2161 /* 2162 * We couldn't find any flushable dirty buffers but 2163 * still have too many dirty buffers, we 2164 * have to sleep and try again. (rare) 2165 */ 2166 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 2167 } 2168 } 2169} 2170 2171/* 2172 * flushbufqueues: 2173 * 2174 * Try to flush a buffer in the dirty queue. We must be careful to 2175 * free up B_INVAL buffers instead of write them, which NFS is 2176 * particularly sensitive to. 2177 */ 2178int flushwithdeps = 0; 2179SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 2180 0, "Number of buffers flushed with dependecies that require rollbacks"); 2181 2182static int 2183flushbufqueues(int flushdeps) 2184{ 2185 struct thread *td = curthread; 2186 struct vnode *vp; 2187 struct mount *mp; 2188 struct buf *bp; 2189 int hasdeps; 2190 2191 mtx_lock(&bqlock); 2192 TAILQ_FOREACH(bp, &bufqueues[QUEUE_DIRTY], b_freelist) { 2193 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 2194 continue; 2195 BO_LOCK(bp->b_bufobj); 2196 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 2197 (bp->b_flags & B_DELWRI) == 0) { 2198 BO_UNLOCK(bp->b_bufobj); 2199 BUF_UNLOCK(bp); 2200 continue; 2201 } 2202 BO_UNLOCK(bp->b_bufobj); 2203 if (bp->b_flags & B_INVAL) { 2204 bremfreel(bp); 2205 mtx_unlock(&bqlock); 2206 brelse(bp); 2207 return (1); 2208 } 2209 2210 if (LIST_FIRST(&bp->b_dep) != NULL && buf_countdeps(bp, 0)) { 2211 if (flushdeps == 0) { 2212 BUF_UNLOCK(bp); 2213 continue; 2214 } 2215 hasdeps = 1; 2216 } else 2217 hasdeps = 0; 2218 /* 2219 * We must hold the lock on a vnode before writing 2220 * one of its buffers. Otherwise we may confuse, or 2221 * in the case of a snapshot vnode, deadlock the 2222 * system. 2223 * 2224 * The lock order here is the reverse of the normal 2225 * of vnode followed by buf lock. This is ok because 2226 * the NOWAIT will prevent deadlock. 2227 */ 2228 vp = bp->b_vp; 2229 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 2230 BUF_UNLOCK(bp); 2231 continue; 2232 } 2233 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) == 0) { 2234 mtx_unlock(&bqlock); 2235 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 2236 bp, bp->b_vp, bp->b_flags); 2237 vfs_bio_awrite(bp); 2238 vn_finished_write(mp); 2239 VOP_UNLOCK(vp, 0, td); 2240 flushwithdeps += hasdeps; 2241 return (1); 2242 } 2243 vn_finished_write(mp); 2244 BUF_UNLOCK(bp); 2245 } 2246 mtx_unlock(&bqlock); 2247 return (0); 2248} 2249 2250/* 2251 * Check to see if a block is currently memory resident. 2252 */ 2253struct buf * 2254incore(struct bufobj *bo, daddr_t blkno) 2255{ 2256 struct buf *bp; 2257 2258 int s = splbio(); 2259 BO_LOCK(bo); 2260 bp = gbincore(bo, blkno); 2261 BO_UNLOCK(bo); 2262 splx(s); 2263 return (bp); 2264} 2265 2266/* 2267 * Returns true if no I/O is needed to access the 2268 * associated VM object. This is like incore except 2269 * it also hunts around in the VM system for the data. 2270 */ 2271 2272static int 2273inmem(struct vnode * vp, daddr_t blkno) 2274{ 2275 vm_object_t obj; 2276 vm_offset_t toff, tinc, size; 2277 vm_page_t m; 2278 vm_ooffset_t off; 2279 2280 ASSERT_VOP_LOCKED(vp, "inmem"); 2281 2282 if (incore(&vp->v_bufobj, blkno)) 2283 return 1; 2284 if (vp->v_mount == NULL) 2285 return 0; 2286 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_vflag & VV_OBJBUF) == 0) 2287 return 0; 2288 2289 size = PAGE_SIZE; 2290 if (size > vp->v_mount->mnt_stat.f_iosize) 2291 size = vp->v_mount->mnt_stat.f_iosize; 2292 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 2293 2294 VM_OBJECT_LOCK(obj); 2295 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 2296 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 2297 if (!m) 2298 goto notinmem; 2299 tinc = size; 2300 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 2301 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 2302 if (vm_page_is_valid(m, 2303 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 2304 goto notinmem; 2305 } 2306 VM_OBJECT_UNLOCK(obj); 2307 return 1; 2308 2309notinmem: 2310 VM_OBJECT_UNLOCK(obj); 2311 return (0); 2312} 2313 2314/* 2315 * vfs_setdirty: 2316 * 2317 * Sets the dirty range for a buffer based on the status of the dirty 2318 * bits in the pages comprising the buffer. 2319 * 2320 * The range is limited to the size of the buffer. 2321 * 2322 * This routine is primarily used by NFS, but is generalized for the 2323 * B_VMIO case. 2324 */ 2325static void 2326vfs_setdirty(struct buf *bp) 2327{ 2328 int i; 2329 vm_object_t object; 2330 2331 /* 2332 * Degenerate case - empty buffer 2333 */ 2334 2335 if (bp->b_bufsize == 0) 2336 return; 2337 2338 /* 2339 * We qualify the scan for modified pages on whether the 2340 * object has been flushed yet. The OBJ_WRITEABLE flag 2341 * is not cleared simply by protecting pages off. 2342 */ 2343 2344 if ((bp->b_flags & B_VMIO) == 0) 2345 return; 2346 2347 object = bp->b_pages[0]->object; 2348 VM_OBJECT_LOCK(object); 2349 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY)) 2350 printf("Warning: object %p writeable but not mightbedirty\n", object); 2351 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY)) 2352 printf("Warning: object %p mightbedirty but not writeable\n", object); 2353 2354 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) { 2355 vm_offset_t boffset; 2356 vm_offset_t eoffset; 2357 2358 vm_page_lock_queues(); 2359 /* 2360 * test the pages to see if they have been modified directly 2361 * by users through the VM system. 2362 */ 2363 for (i = 0; i < bp->b_npages; i++) 2364 vm_page_test_dirty(bp->b_pages[i]); 2365 2366 /* 2367 * Calculate the encompassing dirty range, boffset and eoffset, 2368 * (eoffset - boffset) bytes. 2369 */ 2370 2371 for (i = 0; i < bp->b_npages; i++) { 2372 if (bp->b_pages[i]->dirty) 2373 break; 2374 } 2375 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2376 2377 for (i = bp->b_npages - 1; i >= 0; --i) { 2378 if (bp->b_pages[i]->dirty) { 2379 break; 2380 } 2381 } 2382 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 2383 2384 vm_page_unlock_queues(); 2385 /* 2386 * Fit it to the buffer. 2387 */ 2388 2389 if (eoffset > bp->b_bcount) 2390 eoffset = bp->b_bcount; 2391 2392 /* 2393 * If we have a good dirty range, merge with the existing 2394 * dirty range. 2395 */ 2396 2397 if (boffset < eoffset) { 2398 if (bp->b_dirtyoff > boffset) 2399 bp->b_dirtyoff = boffset; 2400 if (bp->b_dirtyend < eoffset) 2401 bp->b_dirtyend = eoffset; 2402 } 2403 } 2404 VM_OBJECT_UNLOCK(object); 2405} 2406 2407/* 2408 * getblk: 2409 * 2410 * Get a block given a specified block and offset into a file/device. 2411 * The buffers B_DONE bit will be cleared on return, making it almost 2412 * ready for an I/O initiation. B_INVAL may or may not be set on 2413 * return. The caller should clear B_INVAL prior to initiating a 2414 * READ. 2415 * 2416 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 2417 * an existing buffer. 2418 * 2419 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 2420 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 2421 * and then cleared based on the backing VM. If the previous buffer is 2422 * non-0-sized but invalid, B_CACHE will be cleared. 2423 * 2424 * If getblk() must create a new buffer, the new buffer is returned with 2425 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 2426 * case it is returned with B_INVAL clear and B_CACHE set based on the 2427 * backing VM. 2428 * 2429 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 2430 * B_CACHE bit is clear. 2431 * 2432 * What this means, basically, is that the caller should use B_CACHE to 2433 * determine whether the buffer is fully valid or not and should clear 2434 * B_INVAL prior to issuing a read. If the caller intends to validate 2435 * the buffer by loading its data area with something, the caller needs 2436 * to clear B_INVAL. If the caller does this without issuing an I/O, 2437 * the caller should set B_CACHE ( as an optimization ), else the caller 2438 * should issue the I/O and biodone() will set B_CACHE if the I/O was 2439 * a write attempt or if it was a successfull read. If the caller 2440 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 2441 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 2442 */ 2443struct buf * 2444getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo, 2445 int flags) 2446{ 2447 struct buf *bp; 2448 struct bufobj *bo; 2449 int s; 2450 int error; 2451 struct vm_object *vmo; 2452 2453 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 2454 ASSERT_VOP_LOCKED(vp, "getblk"); 2455 if (size > MAXBSIZE) 2456 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 2457 2458 bo = &vp->v_bufobj; 2459 s = splbio(); 2460loop: 2461 /* 2462 * Block if we are low on buffers. Certain processes are allowed 2463 * to completely exhaust the buffer cache. 2464 * 2465 * If this check ever becomes a bottleneck it may be better to 2466 * move it into the else, when gbincore() fails. At the moment 2467 * it isn't a problem. 2468 * 2469 * XXX remove if 0 sections (clean this up after its proven) 2470 */ 2471 if (numfreebuffers == 0) { 2472 if (curthread == PCPU_GET(idlethread)) 2473 return NULL; 2474 mtx_lock(&nblock); 2475 needsbuffer |= VFS_BIO_NEED_ANY; 2476 mtx_unlock(&nblock); 2477 } 2478 2479 VI_LOCK(vp); 2480 bp = gbincore(bo, blkno); 2481 if (bp != NULL) { 2482 int lockflags; 2483 /* 2484 * Buffer is in-core. If the buffer is not busy, it must 2485 * be on a queue. 2486 */ 2487 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 2488 2489 if (flags & GB_LOCK_NOWAIT) 2490 lockflags |= LK_NOWAIT; 2491 2492 error = BUF_TIMELOCK(bp, lockflags, 2493 VI_MTX(vp), "getblk", slpflag, slptimeo); 2494 2495 /* 2496 * If we slept and got the lock we have to restart in case 2497 * the buffer changed identities. 2498 */ 2499 if (error == ENOLCK) 2500 goto loop; 2501 /* We timed out or were interrupted. */ 2502 else if (error) 2503 return (NULL); 2504 2505 /* 2506 * The buffer is locked. B_CACHE is cleared if the buffer is 2507 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 2508 * and for a VMIO buffer B_CACHE is adjusted according to the 2509 * backing VM cache. 2510 */ 2511 if (bp->b_flags & B_INVAL) 2512 bp->b_flags &= ~B_CACHE; 2513 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 2514 bp->b_flags |= B_CACHE; 2515 bremfree(bp); 2516 2517 /* 2518 * check for size inconsistancies for non-VMIO case. 2519 */ 2520 2521 if (bp->b_bcount != size) { 2522 if ((bp->b_flags & B_VMIO) == 0 || 2523 (size > bp->b_kvasize)) { 2524 if (bp->b_flags & B_DELWRI) { 2525 bp->b_flags |= B_NOCACHE; 2526 bwrite(bp); 2527 } else { 2528 if ((bp->b_flags & B_VMIO) && 2529 (LIST_FIRST(&bp->b_dep) == NULL)) { 2530 bp->b_flags |= B_RELBUF; 2531 brelse(bp); 2532 } else { 2533 bp->b_flags |= B_NOCACHE; 2534 bwrite(bp); 2535 } 2536 } 2537 goto loop; 2538 } 2539 } 2540 2541 /* 2542 * If the size is inconsistant in the VMIO case, we can resize 2543 * the buffer. This might lead to B_CACHE getting set or 2544 * cleared. If the size has not changed, B_CACHE remains 2545 * unchanged from its previous state. 2546 */ 2547 2548 if (bp->b_bcount != size) 2549 allocbuf(bp, size); 2550 2551 KASSERT(bp->b_offset != NOOFFSET, 2552 ("getblk: no buffer offset")); 2553 2554 /* 2555 * A buffer with B_DELWRI set and B_CACHE clear must 2556 * be committed before we can return the buffer in 2557 * order to prevent the caller from issuing a read 2558 * ( due to B_CACHE not being set ) and overwriting 2559 * it. 2560 * 2561 * Most callers, including NFS and FFS, need this to 2562 * operate properly either because they assume they 2563 * can issue a read if B_CACHE is not set, or because 2564 * ( for example ) an uncached B_DELWRI might loop due 2565 * to softupdates re-dirtying the buffer. In the latter 2566 * case, B_CACHE is set after the first write completes, 2567 * preventing further loops. 2568 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 2569 * above while extending the buffer, we cannot allow the 2570 * buffer to remain with B_CACHE set after the write 2571 * completes or it will represent a corrupt state. To 2572 * deal with this we set B_NOCACHE to scrap the buffer 2573 * after the write. 2574 * 2575 * We might be able to do something fancy, like setting 2576 * B_CACHE in bwrite() except if B_DELWRI is already set, 2577 * so the below call doesn't set B_CACHE, but that gets real 2578 * confusing. This is much easier. 2579 */ 2580 2581 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 2582 bp->b_flags |= B_NOCACHE; 2583 bwrite(bp); 2584 goto loop; 2585 } 2586 2587 splx(s); 2588 bp->b_flags &= ~B_DONE; 2589 } else { 2590 int bsize, maxsize, vmio; 2591 off_t offset; 2592 2593 /* 2594 * Buffer is not in-core, create new buffer. The buffer 2595 * returned by getnewbuf() is locked. Note that the returned 2596 * buffer is also considered valid (not marked B_INVAL). 2597 */ 2598 VI_UNLOCK(vp); 2599 /* 2600 * If the user does not want us to create the buffer, bail out 2601 * here. 2602 */ 2603 if (flags & GB_NOCREAT) { 2604 splx(s); 2605 return NULL; 2606 } 2607 2608 bsize = bo->bo_bsize; 2609 offset = blkno * bsize; 2610 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && 2611 (vp->v_vflag & VV_OBJBUF); 2612 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 2613 maxsize = imax(maxsize, bsize); 2614 2615 bp = getnewbuf(slpflag, slptimeo, size, maxsize); 2616 if (bp == NULL) { 2617 if (slpflag || slptimeo) { 2618 splx(s); 2619 return NULL; 2620 } 2621 goto loop; 2622 } 2623 2624 /* 2625 * This code is used to make sure that a buffer is not 2626 * created while the getnewbuf routine is blocked. 2627 * This can be a problem whether the vnode is locked or not. 2628 * If the buffer is created out from under us, we have to 2629 * throw away the one we just created. There is now window 2630 * race because we are safely running at splbio() from the 2631 * point of the duplicate buffer creation through to here, 2632 * and we've locked the buffer. 2633 * 2634 * Note: this must occur before we associate the buffer 2635 * with the vp especially considering limitations in 2636 * the splay tree implementation when dealing with duplicate 2637 * lblkno's. 2638 */ 2639 BO_LOCK(bo); 2640 if (gbincore(bo, blkno)) { 2641 BO_UNLOCK(bo); 2642 bp->b_flags |= B_INVAL; 2643 brelse(bp); 2644 goto loop; 2645 } 2646 2647 /* 2648 * Insert the buffer into the hash, so that it can 2649 * be found by incore. 2650 */ 2651 bp->b_blkno = bp->b_lblkno = blkno; 2652 bp->b_offset = offset; 2653 2654 bgetvp(vp, bp); 2655 BO_UNLOCK(bo); 2656 2657 /* 2658 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 2659 * buffer size starts out as 0, B_CACHE will be set by 2660 * allocbuf() for the VMIO case prior to it testing the 2661 * backing store for validity. 2662 */ 2663 2664 if (vmio) { 2665 bp->b_flags |= B_VMIO; 2666#if defined(VFS_BIO_DEBUG) 2667 if (vn_canvmio(vp) != TRUE) 2668 printf("getblk: VMIO on vnode type %d\n", 2669 vp->v_type); 2670#endif 2671 VOP_GETVOBJECT(vp, &vmo); 2672 KASSERT(vmo == bp->b_bufobj->bo_object, 2673 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 2674 bp, vmo, bp->b_bufobj->bo_object)); 2675 } else { 2676 bp->b_flags &= ~B_VMIO; 2677 KASSERT(bp->b_bufobj->bo_object == NULL, 2678 ("ARGH! has b_bufobj->bo_object %p %p\n", 2679 bp, bp->b_bufobj->bo_object)); 2680 } 2681 2682 allocbuf(bp, size); 2683 2684 splx(s); 2685 bp->b_flags &= ~B_DONE; 2686 } 2687 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 2688 KASSERT(BUF_REFCNT(bp) == 1, ("getblk: bp %p not locked",bp)); 2689 KASSERT(bp->b_bufobj == bo, 2690 ("wrong b_bufobj %p should be %p", bp->b_bufobj, bo)); 2691 return (bp); 2692} 2693 2694/* 2695 * Get an empty, disassociated buffer of given size. The buffer is initially 2696 * set to B_INVAL. 2697 */ 2698struct buf * 2699geteblk(int size) 2700{ 2701 struct buf *bp; 2702 int s; 2703 int maxsize; 2704 2705 maxsize = (size + BKVAMASK) & ~BKVAMASK; 2706 2707 s = splbio(); 2708 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0) 2709 continue; 2710 splx(s); 2711 allocbuf(bp, size); 2712 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 2713 KASSERT(BUF_REFCNT(bp) == 1, ("geteblk: bp %p not locked",bp)); 2714 return (bp); 2715} 2716 2717 2718/* 2719 * This code constitutes the buffer memory from either anonymous system 2720 * memory (in the case of non-VMIO operations) or from an associated 2721 * VM object (in the case of VMIO operations). This code is able to 2722 * resize a buffer up or down. 2723 * 2724 * Note that this code is tricky, and has many complications to resolve 2725 * deadlock or inconsistant data situations. Tread lightly!!! 2726 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 2727 * the caller. Calling this code willy nilly can result in the loss of data. 2728 * 2729 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 2730 * B_CACHE for the non-VMIO case. 2731 */ 2732 2733int 2734allocbuf(struct buf *bp, int size) 2735{ 2736 int newbsize, mbsize; 2737 int i; 2738 2739 if (BUF_REFCNT(bp) == 0) 2740 panic("allocbuf: buffer not busy"); 2741 2742 if (bp->b_kvasize < size) 2743 panic("allocbuf: buffer too small"); 2744 2745 if ((bp->b_flags & B_VMIO) == 0) { 2746 caddr_t origbuf; 2747 int origbufsize; 2748 /* 2749 * Just get anonymous memory from the kernel. Don't 2750 * mess with B_CACHE. 2751 */ 2752 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2753 if (bp->b_flags & B_MALLOC) 2754 newbsize = mbsize; 2755 else 2756 newbsize = round_page(size); 2757 2758 if (newbsize < bp->b_bufsize) { 2759 /* 2760 * malloced buffers are not shrunk 2761 */ 2762 if (bp->b_flags & B_MALLOC) { 2763 if (newbsize) { 2764 bp->b_bcount = size; 2765 } else { 2766 free(bp->b_data, M_BIOBUF); 2767 if (bp->b_bufsize) { 2768 atomic_subtract_int( 2769 &bufmallocspace, 2770 bp->b_bufsize); 2771 bufspacewakeup(); 2772 bp->b_bufsize = 0; 2773 } 2774 bp->b_saveaddr = bp->b_kvabase; 2775 bp->b_data = bp->b_saveaddr; 2776 bp->b_bcount = 0; 2777 bp->b_flags &= ~B_MALLOC; 2778 } 2779 return 1; 2780 } 2781 vm_hold_free_pages( 2782 bp, 2783 (vm_offset_t) bp->b_data + newbsize, 2784 (vm_offset_t) bp->b_data + bp->b_bufsize); 2785 } else if (newbsize > bp->b_bufsize) { 2786 /* 2787 * We only use malloced memory on the first allocation. 2788 * and revert to page-allocated memory when the buffer 2789 * grows. 2790 */ 2791 /* 2792 * There is a potential smp race here that could lead 2793 * to bufmallocspace slightly passing the max. It 2794 * is probably extremely rare and not worth worrying 2795 * over. 2796 */ 2797 if ( (bufmallocspace < maxbufmallocspace) && 2798 (bp->b_bufsize == 0) && 2799 (mbsize <= PAGE_SIZE/2)) { 2800 2801 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 2802 bp->b_bufsize = mbsize; 2803 bp->b_bcount = size; 2804 bp->b_flags |= B_MALLOC; 2805 atomic_add_int(&bufmallocspace, mbsize); 2806 return 1; 2807 } 2808 origbuf = NULL; 2809 origbufsize = 0; 2810 /* 2811 * If the buffer is growing on its other-than-first allocation, 2812 * then we revert to the page-allocation scheme. 2813 */ 2814 if (bp->b_flags & B_MALLOC) { 2815 origbuf = bp->b_data; 2816 origbufsize = bp->b_bufsize; 2817 bp->b_data = bp->b_kvabase; 2818 if (bp->b_bufsize) { 2819 atomic_subtract_int(&bufmallocspace, 2820 bp->b_bufsize); 2821 bufspacewakeup(); 2822 bp->b_bufsize = 0; 2823 } 2824 bp->b_flags &= ~B_MALLOC; 2825 newbsize = round_page(newbsize); 2826 } 2827 vm_hold_load_pages( 2828 bp, 2829 (vm_offset_t) bp->b_data + bp->b_bufsize, 2830 (vm_offset_t) bp->b_data + newbsize); 2831 if (origbuf) { 2832 bcopy(origbuf, bp->b_data, origbufsize); 2833 free(origbuf, M_BIOBUF); 2834 } 2835 } 2836 } else { 2837 int desiredpages; 2838 2839 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2840 desiredpages = (size == 0) ? 0 : 2841 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 2842 2843 if (bp->b_flags & B_MALLOC) 2844 panic("allocbuf: VMIO buffer can't be malloced"); 2845 /* 2846 * Set B_CACHE initially if buffer is 0 length or will become 2847 * 0-length. 2848 */ 2849 if (size == 0 || bp->b_bufsize == 0) 2850 bp->b_flags |= B_CACHE; 2851 2852 if (newbsize < bp->b_bufsize) { 2853 /* 2854 * DEV_BSIZE aligned new buffer size is less then the 2855 * DEV_BSIZE aligned existing buffer size. Figure out 2856 * if we have to remove any pages. 2857 */ 2858 if (desiredpages < bp->b_npages) { 2859 vm_page_t m; 2860 2861 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 2862 vm_page_lock_queues(); 2863 for (i = desiredpages; i < bp->b_npages; i++) { 2864 /* 2865 * the page is not freed here -- it 2866 * is the responsibility of 2867 * vnode_pager_setsize 2868 */ 2869 m = bp->b_pages[i]; 2870 KASSERT(m != bogus_page, 2871 ("allocbuf: bogus page found")); 2872 while (vm_page_sleep_if_busy(m, TRUE, "biodep")) 2873 vm_page_lock_queues(); 2874 2875 bp->b_pages[i] = NULL; 2876 vm_page_unwire(m, 0); 2877 } 2878 vm_page_unlock_queues(); 2879 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 2880 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 2881 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 2882 bp->b_npages = desiredpages; 2883 } 2884 } else if (size > bp->b_bcount) { 2885 /* 2886 * We are growing the buffer, possibly in a 2887 * byte-granular fashion. 2888 */ 2889 struct vnode *vp; 2890 vm_object_t obj; 2891 vm_offset_t toff; 2892 vm_offset_t tinc; 2893 2894 /* 2895 * Step 1, bring in the VM pages from the object, 2896 * allocating them if necessary. We must clear 2897 * B_CACHE if these pages are not valid for the 2898 * range covered by the buffer. 2899 */ 2900 2901 vp = bp->b_vp; 2902 obj = bp->b_bufobj->bo_object; 2903 2904 VM_OBJECT_LOCK(obj); 2905 while (bp->b_npages < desiredpages) { 2906 vm_page_t m; 2907 vm_pindex_t pi; 2908 2909 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages; 2910 if ((m = vm_page_lookup(obj, pi)) == NULL) { 2911 /* 2912 * note: must allocate system pages 2913 * since blocking here could intefere 2914 * with paging I/O, no matter which 2915 * process we are. 2916 */ 2917 m = vm_page_alloc(obj, pi, 2918 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | 2919 VM_ALLOC_WIRED); 2920 if (m == NULL) { 2921 atomic_add_int(&vm_pageout_deficit, 2922 desiredpages - bp->b_npages); 2923 VM_OBJECT_UNLOCK(obj); 2924 VM_WAIT; 2925 VM_OBJECT_LOCK(obj); 2926 } else { 2927 bp->b_flags &= ~B_CACHE; 2928 bp->b_pages[bp->b_npages] = m; 2929 ++bp->b_npages; 2930 } 2931 continue; 2932 } 2933 2934 /* 2935 * We found a page. If we have to sleep on it, 2936 * retry because it might have gotten freed out 2937 * from under us. 2938 * 2939 * We can only test PG_BUSY here. Blocking on 2940 * m->busy might lead to a deadlock: 2941 * 2942 * vm_fault->getpages->cluster_read->allocbuf 2943 * 2944 */ 2945 vm_page_lock_queues(); 2946 if (vm_page_sleep_if_busy(m, FALSE, "pgtblk")) 2947 continue; 2948 2949 /* 2950 * We have a good page. Should we wakeup the 2951 * page daemon? 2952 */ 2953 if ((curproc != pageproc) && 2954 ((m->queue - m->pc) == PQ_CACHE) && 2955 ((cnt.v_free_count + cnt.v_cache_count) < 2956 (cnt.v_free_min + cnt.v_cache_min))) { 2957 pagedaemon_wakeup(); 2958 } 2959 vm_page_wire(m); 2960 vm_page_unlock_queues(); 2961 bp->b_pages[bp->b_npages] = m; 2962 ++bp->b_npages; 2963 } 2964 2965 /* 2966 * Step 2. We've loaded the pages into the buffer, 2967 * we have to figure out if we can still have B_CACHE 2968 * set. Note that B_CACHE is set according to the 2969 * byte-granular range ( bcount and size ), new the 2970 * aligned range ( newbsize ). 2971 * 2972 * The VM test is against m->valid, which is DEV_BSIZE 2973 * aligned. Needless to say, the validity of the data 2974 * needs to also be DEV_BSIZE aligned. Note that this 2975 * fails with NFS if the server or some other client 2976 * extends the file's EOF. If our buffer is resized, 2977 * B_CACHE may remain set! XXX 2978 */ 2979 2980 toff = bp->b_bcount; 2981 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 2982 2983 while ((bp->b_flags & B_CACHE) && toff < size) { 2984 vm_pindex_t pi; 2985 2986 if (tinc > (size - toff)) 2987 tinc = size - toff; 2988 2989 pi = ((bp->b_offset & PAGE_MASK) + toff) >> 2990 PAGE_SHIFT; 2991 2992 vfs_buf_test_cache( 2993 bp, 2994 bp->b_offset, 2995 toff, 2996 tinc, 2997 bp->b_pages[pi] 2998 ); 2999 toff += tinc; 3000 tinc = PAGE_SIZE; 3001 } 3002 VM_OBJECT_UNLOCK(obj); 3003 3004 /* 3005 * Step 3, fixup the KVM pmap. Remember that 3006 * bp->b_data is relative to bp->b_offset, but 3007 * bp->b_offset may be offset into the first page. 3008 */ 3009 3010 bp->b_data = (caddr_t) 3011 trunc_page((vm_offset_t)bp->b_data); 3012 pmap_qenter( 3013 (vm_offset_t)bp->b_data, 3014 bp->b_pages, 3015 bp->b_npages 3016 ); 3017 3018 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 3019 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 3020 } 3021 } 3022 if (newbsize < bp->b_bufsize) 3023 bufspacewakeup(); 3024 bp->b_bufsize = newbsize; /* actual buffer allocation */ 3025 bp->b_bcount = size; /* requested buffer size */ 3026 return 1; 3027} 3028 3029void 3030biodone(struct bio *bp) 3031{ 3032 3033 mtx_lock(&bdonelock); 3034 bp->bio_flags |= BIO_DONE; 3035 if (bp->bio_done == NULL) 3036 wakeup(bp); 3037 mtx_unlock(&bdonelock); 3038 if (bp->bio_done != NULL) 3039 bp->bio_done(bp); 3040} 3041 3042/* 3043 * Wait for a BIO to finish. 3044 * 3045 * XXX: resort to a timeout for now. The optimal locking (if any) for this 3046 * case is not yet clear. 3047 */ 3048int 3049biowait(struct bio *bp, const char *wchan) 3050{ 3051 3052 mtx_lock(&bdonelock); 3053 while ((bp->bio_flags & BIO_DONE) == 0) 3054 msleep(bp, &bdonelock, PRIBIO, wchan, hz / 10); 3055 mtx_unlock(&bdonelock); 3056 if (bp->bio_error != 0) 3057 return (bp->bio_error); 3058 if (!(bp->bio_flags & BIO_ERROR)) 3059 return (0); 3060 return (EIO); 3061} 3062 3063void 3064biofinish(struct bio *bp, struct devstat *stat, int error) 3065{ 3066 3067 if (error) { 3068 bp->bio_error = error; 3069 bp->bio_flags |= BIO_ERROR; 3070 } 3071 if (stat != NULL) 3072 devstat_end_transaction_bio(stat, bp); 3073 biodone(bp); 3074} 3075 3076/* 3077 * bufwait: 3078 * 3079 * Wait for buffer I/O completion, returning error status. The buffer 3080 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 3081 * error and cleared. 3082 */ 3083int 3084bufwait(struct buf *bp) 3085{ 3086 int s; 3087 3088 s = splbio(); 3089 if (bp->b_iocmd == BIO_READ) 3090 bwait(bp, PRIBIO, "biord"); 3091 else 3092 bwait(bp, PRIBIO, "biowr"); 3093 splx(s); 3094 if (bp->b_flags & B_EINTR) { 3095 bp->b_flags &= ~B_EINTR; 3096 return (EINTR); 3097 } 3098 if (bp->b_ioflags & BIO_ERROR) { 3099 return (bp->b_error ? bp->b_error : EIO); 3100 } else { 3101 return (0); 3102 } 3103} 3104 3105 /* 3106 * Call back function from struct bio back up to struct buf. 3107 */ 3108static void 3109bufdonebio(struct bio *bip) 3110{ 3111 struct buf *bp; 3112 3113 bp = bip->bio_caller2; 3114 bp->b_resid = bp->b_bcount - bip->bio_completed; 3115 bp->b_resid = bip->bio_resid; /* XXX: remove */ 3116 bp->b_ioflags = bip->bio_flags; 3117 bp->b_error = bip->bio_error; 3118 if (bp->b_error) 3119 bp->b_ioflags |= BIO_ERROR; 3120 bufdone(bp); 3121 g_destroy_bio(bip); 3122} 3123 3124void 3125dev_strategy(struct cdev *dev, struct buf *bp) 3126{ 3127 struct cdevsw *csw; 3128 struct bio *bip; 3129 3130 if ((!bp->b_iocmd) || (bp->b_iocmd & (bp->b_iocmd - 1))) 3131 panic("b_iocmd botch"); 3132 for (;;) { 3133 bip = g_new_bio(); 3134 if (bip != NULL) 3135 break; 3136 /* Try again later */ 3137 tsleep(&bp, PRIBIO, "dev_strat", hz/10); 3138 } 3139 bip->bio_cmd = bp->b_iocmd; 3140 bip->bio_offset = bp->b_iooffset; 3141 bip->bio_length = bp->b_bcount; 3142 bip->bio_bcount = bp->b_bcount; /* XXX: remove */ 3143 bip->bio_data = bp->b_data; 3144 bip->bio_done = bufdonebio; 3145 bip->bio_caller2 = bp; 3146 bip->bio_dev = dev; 3147 KASSERT(dev->si_refcount > 0, 3148 ("dev_strategy on un-referenced struct cdev *(%s)", 3149 devtoname(dev))); 3150 csw = dev_refthread(dev); 3151 if (csw == NULL) { 3152 bp->b_error = ENXIO; 3153 bp->b_ioflags = BIO_ERROR; 3154 bufdone(bp); 3155 return; 3156 } 3157 (*csw->d_strategy)(bip); 3158 dev_relthread(dev); 3159} 3160 3161/* 3162 * bufdone: 3163 * 3164 * Finish I/O on a buffer, optionally calling a completion function. 3165 * This is usually called from an interrupt so process blocking is 3166 * not allowed. 3167 * 3168 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 3169 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 3170 * assuming B_INVAL is clear. 3171 * 3172 * For the VMIO case, we set B_CACHE if the op was a read and no 3173 * read error occured, or if the op was a write. B_CACHE is never 3174 * set if the buffer is invalid or otherwise uncacheable. 3175 * 3176 * biodone does not mess with B_INVAL, allowing the I/O routine or the 3177 * initiator to leave B_INVAL set to brelse the buffer out of existance 3178 * in the biodone routine. 3179 */ 3180void 3181bufdone(struct buf *bp) 3182{ 3183 int s; 3184 void (*biodone)(struct buf *); 3185 3186 3187 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 3188 s = splbio(); 3189 3190 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp))); 3191 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 3192 3193 runningbufwakeup(bp); 3194 3195 if (bp->b_iocmd == BIO_WRITE && bp->b_bufobj != NULL) 3196 bufobj_wdrop(bp->b_bufobj); 3197 3198 /* call optional completion function if requested */ 3199 if (bp->b_iodone != NULL) { 3200 biodone = bp->b_iodone; 3201 bp->b_iodone = NULL; 3202 /* 3203 * Device drivers may or may not hold giant, hold it here 3204 * if we're calling into unknown code. 3205 */ 3206 mtx_lock(&Giant); 3207 bp->b_flags |= B_DONE; /* XXX Should happen after biodone? */ 3208 (*biodone) (bp); 3209 mtx_unlock(&Giant); 3210 splx(s); 3211 return; 3212 } 3213 if (LIST_FIRST(&bp->b_dep) != NULL) 3214 buf_complete(bp); 3215 3216 if (bp->b_flags & B_VMIO) { 3217 int i; 3218 vm_ooffset_t foff; 3219 vm_page_t m; 3220 vm_object_t obj; 3221 int iosize; 3222 struct vnode *vp = bp->b_vp; 3223 3224 obj = bp->b_bufobj->bo_object; 3225 3226#if defined(VFS_BIO_DEBUG) 3227 mp_fixme("usecount and vflag accessed without locks."); 3228 if (vp->v_usecount == 0) { 3229 panic("biodone: zero vnode ref count"); 3230 } 3231 3232 if ((vp->v_vflag & VV_OBJBUF) == 0) { 3233 panic("biodone: vnode is not setup for merged cache"); 3234 } 3235#endif 3236 3237 foff = bp->b_offset; 3238 KASSERT(bp->b_offset != NOOFFSET, 3239 ("biodone: no buffer offset")); 3240 3241 VM_OBJECT_LOCK(obj); 3242#if defined(VFS_BIO_DEBUG) 3243 if (obj->paging_in_progress < bp->b_npages) { 3244 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 3245 obj->paging_in_progress, bp->b_npages); 3246 } 3247#endif 3248 3249 /* 3250 * Set B_CACHE if the op was a normal read and no error 3251 * occured. B_CACHE is set for writes in the b*write() 3252 * routines. 3253 */ 3254 iosize = bp->b_bcount - bp->b_resid; 3255 if (bp->b_iocmd == BIO_READ && 3256 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 3257 !(bp->b_ioflags & BIO_ERROR)) { 3258 bp->b_flags |= B_CACHE; 3259 } 3260 vm_page_lock_queues(); 3261 for (i = 0; i < bp->b_npages; i++) { 3262 int bogusflag = 0; 3263 int resid; 3264 3265 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 3266 if (resid > iosize) 3267 resid = iosize; 3268 3269 /* 3270 * cleanup bogus pages, restoring the originals 3271 */ 3272 m = bp->b_pages[i]; 3273 if (m == bogus_page) { 3274 bogusflag = 1; 3275 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 3276 if (m == NULL) 3277 panic("biodone: page disappeared!"); 3278 bp->b_pages[i] = m; 3279 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 3280 } 3281#if defined(VFS_BIO_DEBUG) 3282 if (OFF_TO_IDX(foff) != m->pindex) { 3283 printf( 3284"biodone: foff(%jd)/m->pindex(%ju) mismatch\n", 3285 (intmax_t)foff, (uintmax_t)m->pindex); 3286 } 3287#endif 3288 3289 /* 3290 * In the write case, the valid and clean bits are 3291 * already changed correctly ( see bdwrite() ), so we 3292 * only need to do this here in the read case. 3293 */ 3294 if ((bp->b_iocmd == BIO_READ) && !bogusflag && resid > 0) { 3295 vfs_page_set_valid(bp, foff, i, m); 3296 } 3297 3298 /* 3299 * when debugging new filesystems or buffer I/O methods, this 3300 * is the most common error that pops up. if you see this, you 3301 * have not set the page busy flag correctly!!! 3302 */ 3303 if (m->busy == 0) { 3304 printf("biodone: page busy < 0, " 3305 "pindex: %d, foff: 0x(%x,%x), " 3306 "resid: %d, index: %d\n", 3307 (int) m->pindex, (int)(foff >> 32), 3308 (int) foff & 0xffffffff, resid, i); 3309 if (!vn_isdisk(vp, NULL)) 3310 printf(" iosize: %jd, lblkno: %jd, flags: 0x%x, npages: %d\n", 3311 (intmax_t)bp->b_vp->v_mount->mnt_stat.f_iosize, 3312 (intmax_t) bp->b_lblkno, 3313 bp->b_flags, bp->b_npages); 3314 else 3315 printf(" VDEV, lblkno: %jd, flags: 0x%x, npages: %d\n", 3316 (intmax_t) bp->b_lblkno, 3317 bp->b_flags, bp->b_npages); 3318 printf(" valid: 0x%lx, dirty: 0x%lx, wired: %d\n", 3319 (u_long)m->valid, (u_long)m->dirty, 3320 m->wire_count); 3321 panic("biodone: page busy < 0\n"); 3322 } 3323 vm_page_io_finish(m); 3324 vm_object_pip_subtract(obj, 1); 3325 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3326 iosize -= resid; 3327 } 3328 vm_page_unlock_queues(); 3329 vm_object_pip_wakeupn(obj, 0); 3330 VM_OBJECT_UNLOCK(obj); 3331 } 3332 3333 /* 3334 * For asynchronous completions, release the buffer now. The brelse 3335 * will do a wakeup there if necessary - so no need to do a wakeup 3336 * here in the async case. The sync case always needs to do a wakeup. 3337 */ 3338 3339 if (bp->b_flags & B_ASYNC) { 3340 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) 3341 brelse(bp); 3342 else 3343 bqrelse(bp); 3344 } else { 3345 bdone(bp); 3346 } 3347 splx(s); 3348} 3349 3350/* 3351 * This routine is called in lieu of iodone in the case of 3352 * incomplete I/O. This keeps the busy status for pages 3353 * consistant. 3354 */ 3355void 3356vfs_unbusy_pages(struct buf *bp) 3357{ 3358 int i; 3359 vm_object_t obj; 3360 vm_page_t m; 3361 3362 runningbufwakeup(bp); 3363 if (!(bp->b_flags & B_VMIO)) 3364 return; 3365 3366 obj = bp->b_bufobj->bo_object; 3367 VM_OBJECT_LOCK(obj); 3368 vm_page_lock_queues(); 3369 for (i = 0; i < bp->b_npages; i++) { 3370 m = bp->b_pages[i]; 3371 if (m == bogus_page) { 3372 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 3373 if (!m) 3374 panic("vfs_unbusy_pages: page missing\n"); 3375 bp->b_pages[i] = m; 3376 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3377 bp->b_pages, bp->b_npages); 3378 } 3379 vm_object_pip_subtract(obj, 1); 3380 vm_page_io_finish(m); 3381 } 3382 vm_page_unlock_queues(); 3383 vm_object_pip_wakeupn(obj, 0); 3384 VM_OBJECT_UNLOCK(obj); 3385} 3386 3387/* 3388 * vfs_page_set_valid: 3389 * 3390 * Set the valid bits in a page based on the supplied offset. The 3391 * range is restricted to the buffer's size. 3392 * 3393 * This routine is typically called after a read completes. 3394 */ 3395static void 3396vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 3397{ 3398 vm_ooffset_t soff, eoff; 3399 3400 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 3401 /* 3402 * Start and end offsets in buffer. eoff - soff may not cross a 3403 * page boundry or cross the end of the buffer. The end of the 3404 * buffer, in this case, is our file EOF, not the allocation size 3405 * of the buffer. 3406 */ 3407 soff = off; 3408 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3409 if (eoff > bp->b_offset + bp->b_bcount) 3410 eoff = bp->b_offset + bp->b_bcount; 3411 3412 /* 3413 * Set valid range. This is typically the entire buffer and thus the 3414 * entire page. 3415 */ 3416 if (eoff > soff) { 3417 vm_page_set_validclean( 3418 m, 3419 (vm_offset_t) (soff & PAGE_MASK), 3420 (vm_offset_t) (eoff - soff) 3421 ); 3422 } 3423} 3424 3425/* 3426 * This routine is called before a device strategy routine. 3427 * It is used to tell the VM system that paging I/O is in 3428 * progress, and treat the pages associated with the buffer 3429 * almost as being PG_BUSY. Also the object paging_in_progress 3430 * flag is handled to make sure that the object doesn't become 3431 * inconsistant. 3432 * 3433 * Since I/O has not been initiated yet, certain buffer flags 3434 * such as BIO_ERROR or B_INVAL may be in an inconsistant state 3435 * and should be ignored. 3436 */ 3437void 3438vfs_busy_pages(struct buf *bp, int clear_modify) 3439{ 3440 int i, bogus; 3441 vm_object_t obj; 3442 vm_ooffset_t foff; 3443 vm_page_t m; 3444 3445 if (!(bp->b_flags & B_VMIO)) 3446 return; 3447 3448 obj = bp->b_bufobj->bo_object; 3449 foff = bp->b_offset; 3450 KASSERT(bp->b_offset != NOOFFSET, 3451 ("vfs_busy_pages: no buffer offset")); 3452 vfs_setdirty(bp); 3453 VM_OBJECT_LOCK(obj); 3454retry: 3455 vm_page_lock_queues(); 3456 for (i = 0; i < bp->b_npages; i++) { 3457 m = bp->b_pages[i]; 3458 3459 if (vm_page_sleep_if_busy(m, FALSE, "vbpage")) 3460 goto retry; 3461 } 3462 bogus = 0; 3463 for (i = 0; i < bp->b_npages; i++) { 3464 m = bp->b_pages[i]; 3465 3466 if ((bp->b_flags & B_CLUSTER) == 0) { 3467 vm_object_pip_add(obj, 1); 3468 vm_page_io_start(m); 3469 } 3470 /* 3471 * When readying a buffer for a read ( i.e 3472 * clear_modify == 0 ), it is important to do 3473 * bogus_page replacement for valid pages in 3474 * partially instantiated buffers. Partially 3475 * instantiated buffers can, in turn, occur when 3476 * reconstituting a buffer from its VM backing store 3477 * base. We only have to do this if B_CACHE is 3478 * clear ( which causes the I/O to occur in the 3479 * first place ). The replacement prevents the read 3480 * I/O from overwriting potentially dirty VM-backed 3481 * pages. XXX bogus page replacement is, uh, bogus. 3482 * It may not work properly with small-block devices. 3483 * We need to find a better way. 3484 */ 3485 pmap_remove_all(m); 3486 if (clear_modify) 3487 vfs_page_set_valid(bp, foff, i, m); 3488 else if (m->valid == VM_PAGE_BITS_ALL && 3489 (bp->b_flags & B_CACHE) == 0) { 3490 bp->b_pages[i] = bogus_page; 3491 bogus++; 3492 } 3493 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3494 } 3495 vm_page_unlock_queues(); 3496 VM_OBJECT_UNLOCK(obj); 3497 if (bogus) 3498 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 3499 bp->b_pages, bp->b_npages); 3500} 3501 3502/* 3503 * Tell the VM system that the pages associated with this buffer 3504 * are clean. This is used for delayed writes where the data is 3505 * going to go to disk eventually without additional VM intevention. 3506 * 3507 * Note that while we only really need to clean through to b_bcount, we 3508 * just go ahead and clean through to b_bufsize. 3509 */ 3510static void 3511vfs_clean_pages(struct buf *bp) 3512{ 3513 int i; 3514 vm_ooffset_t foff, noff, eoff; 3515 vm_page_t m; 3516 3517 if (!(bp->b_flags & B_VMIO)) 3518 return; 3519 3520 foff = bp->b_offset; 3521 KASSERT(bp->b_offset != NOOFFSET, 3522 ("vfs_clean_pages: no buffer offset")); 3523 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3524 vm_page_lock_queues(); 3525 for (i = 0; i < bp->b_npages; i++) { 3526 m = bp->b_pages[i]; 3527 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3528 eoff = noff; 3529 3530 if (eoff > bp->b_offset + bp->b_bufsize) 3531 eoff = bp->b_offset + bp->b_bufsize; 3532 vfs_page_set_valid(bp, foff, i, m); 3533 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3534 foff = noff; 3535 } 3536 vm_page_unlock_queues(); 3537 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3538} 3539 3540/* 3541 * vfs_bio_set_validclean: 3542 * 3543 * Set the range within the buffer to valid and clean. The range is 3544 * relative to the beginning of the buffer, b_offset. Note that b_offset 3545 * itself may be offset from the beginning of the first page. 3546 * 3547 */ 3548 3549void 3550vfs_bio_set_validclean(struct buf *bp, int base, int size) 3551{ 3552 int i, n; 3553 vm_page_t m; 3554 3555 if (!(bp->b_flags & B_VMIO)) 3556 return; 3557 /* 3558 * Fixup base to be relative to beginning of first page. 3559 * Set initial n to be the maximum number of bytes in the 3560 * first page that can be validated. 3561 */ 3562 3563 base += (bp->b_offset & PAGE_MASK); 3564 n = PAGE_SIZE - (base & PAGE_MASK); 3565 3566 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3567 vm_page_lock_queues(); 3568 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 3569 m = bp->b_pages[i]; 3570 if (n > size) 3571 n = size; 3572 vm_page_set_validclean(m, base & PAGE_MASK, n); 3573 base += n; 3574 size -= n; 3575 n = PAGE_SIZE; 3576 } 3577 vm_page_unlock_queues(); 3578 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3579} 3580 3581/* 3582 * vfs_bio_clrbuf: 3583 * 3584 * clear a buffer. This routine essentially fakes an I/O, so we need 3585 * to clear BIO_ERROR and B_INVAL. 3586 * 3587 * Note that while we only theoretically need to clear through b_bcount, 3588 * we go ahead and clear through b_bufsize. 3589 */ 3590 3591void 3592vfs_bio_clrbuf(struct buf *bp) 3593{ 3594 int i, j, mask = 0; 3595 caddr_t sa, ea; 3596 3597 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 3598 clrbuf(bp); 3599 return; 3600 } 3601 3602 bp->b_flags &= ~B_INVAL; 3603 bp->b_ioflags &= ~BIO_ERROR; 3604 VM_OBJECT_LOCK(bp->b_bufobj->bo_object); 3605 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 3606 (bp->b_offset & PAGE_MASK) == 0) { 3607 if (bp->b_pages[0] == bogus_page) 3608 goto unlock; 3609 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 3610 VM_OBJECT_LOCK_ASSERT(bp->b_pages[0]->object, MA_OWNED); 3611 if ((bp->b_pages[0]->valid & mask) == mask) 3612 goto unlock; 3613 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 3614 ((bp->b_pages[0]->valid & mask) == 0)) { 3615 bzero(bp->b_data, bp->b_bufsize); 3616 bp->b_pages[0]->valid |= mask; 3617 goto unlock; 3618 } 3619 } 3620 ea = sa = bp->b_data; 3621 for(i = 0; i < bp->b_npages; i++, sa = ea) { 3622 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 3623 ea = (caddr_t)(vm_offset_t)ulmin( 3624 (u_long)(vm_offset_t)ea, 3625 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize); 3626 if (bp->b_pages[i] == bogus_page) 3627 continue; 3628 j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE; 3629 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 3630 VM_OBJECT_LOCK_ASSERT(bp->b_pages[i]->object, MA_OWNED); 3631 if ((bp->b_pages[i]->valid & mask) == mask) 3632 continue; 3633 if ((bp->b_pages[i]->valid & mask) == 0) { 3634 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) 3635 bzero(sa, ea - sa); 3636 } else { 3637 for (; sa < ea; sa += DEV_BSIZE, j++) { 3638 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 3639 (bp->b_pages[i]->valid & (1 << j)) == 0) 3640 bzero(sa, DEV_BSIZE); 3641 } 3642 } 3643 bp->b_pages[i]->valid |= mask; 3644 } 3645unlock: 3646 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object); 3647 bp->b_resid = 0; 3648} 3649 3650/* 3651 * vm_hold_load_pages and vm_hold_free_pages get pages into 3652 * a buffers address space. The pages are anonymous and are 3653 * not associated with a file object. 3654 */ 3655static void 3656vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3657{ 3658 vm_offset_t pg; 3659 vm_page_t p; 3660 int index; 3661 3662 to = round_page(to); 3663 from = round_page(from); 3664 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3665 3666 VM_OBJECT_LOCK(kernel_object); 3667 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3668tryagain: 3669 /* 3670 * note: must allocate system pages since blocking here 3671 * could intefere with paging I/O, no matter which 3672 * process we are. 3673 */ 3674 p = vm_page_alloc(kernel_object, 3675 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 3676 VM_ALLOC_NOBUSY | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED); 3677 if (!p) { 3678 atomic_add_int(&vm_pageout_deficit, 3679 (to - pg) >> PAGE_SHIFT); 3680 VM_OBJECT_UNLOCK(kernel_object); 3681 VM_WAIT; 3682 VM_OBJECT_LOCK(kernel_object); 3683 goto tryagain; 3684 } 3685 p->valid = VM_PAGE_BITS_ALL; 3686 pmap_qenter(pg, &p, 1); 3687 bp->b_pages[index] = p; 3688 } 3689 VM_OBJECT_UNLOCK(kernel_object); 3690 bp->b_npages = index; 3691} 3692 3693/* Return pages associated with this buf to the vm system */ 3694static void 3695vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 3696{ 3697 vm_offset_t pg; 3698 vm_page_t p; 3699 int index, newnpages; 3700 3701 from = round_page(from); 3702 to = round_page(to); 3703 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 3704 3705 VM_OBJECT_LOCK(kernel_object); 3706 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 3707 p = bp->b_pages[index]; 3708 if (p && (index < bp->b_npages)) { 3709 if (p->busy) { 3710 printf( 3711 "vm_hold_free_pages: blkno: %jd, lblkno: %jd\n", 3712 (intmax_t)bp->b_blkno, 3713 (intmax_t)bp->b_lblkno); 3714 } 3715 bp->b_pages[index] = NULL; 3716 pmap_qremove(pg, 1); 3717 vm_page_lock_queues(); 3718 vm_page_unwire(p, 0); 3719 vm_page_free(p); 3720 vm_page_unlock_queues(); 3721 } 3722 } 3723 VM_OBJECT_UNLOCK(kernel_object); 3724 bp->b_npages = newnpages; 3725} 3726 3727/* 3728 * Map an IO request into kernel virtual address space. 3729 * 3730 * All requests are (re)mapped into kernel VA space. 3731 * Notice that we use b_bufsize for the size of the buffer 3732 * to be mapped. b_bcount might be modified by the driver. 3733 * 3734 * Note that even if the caller determines that the address space should 3735 * be valid, a race or a smaller-file mapped into a larger space may 3736 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 3737 * check the return value. 3738 */ 3739int 3740vmapbuf(struct buf *bp) 3741{ 3742 caddr_t addr, kva; 3743 vm_prot_t prot; 3744 int pidx, i; 3745 struct vm_page *m; 3746 struct pmap *pmap = &curproc->p_vmspace->vm_pmap; 3747 3748 if (bp->b_bufsize < 0) 3749 return (-1); 3750 prot = VM_PROT_READ; 3751 if (bp->b_iocmd == BIO_READ) 3752 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 3753 for (addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data), pidx = 0; 3754 addr < bp->b_data + bp->b_bufsize; 3755 addr += PAGE_SIZE, pidx++) { 3756 /* 3757 * Do the vm_fault if needed; do the copy-on-write thing 3758 * when reading stuff off device into memory. 3759 * 3760 * NOTE! Must use pmap_extract() because addr may be in 3761 * the userland address space, and kextract is only guarenteed 3762 * to work for the kernland address space (see: sparc64 port). 3763 */ 3764retry: 3765 if (vm_fault_quick(addr >= bp->b_data ? addr : bp->b_data, 3766 prot) < 0) { 3767 vm_page_lock_queues(); 3768 for (i = 0; i < pidx; ++i) { 3769 vm_page_unhold(bp->b_pages[i]); 3770 bp->b_pages[i] = NULL; 3771 } 3772 vm_page_unlock_queues(); 3773 return(-1); 3774 } 3775 m = pmap_extract_and_hold(pmap, (vm_offset_t)addr, prot); 3776 if (m == NULL) 3777 goto retry; 3778 bp->b_pages[pidx] = m; 3779 } 3780 if (pidx > btoc(MAXPHYS)) 3781 panic("vmapbuf: mapped more than MAXPHYS"); 3782 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx); 3783 3784 kva = bp->b_saveaddr; 3785 bp->b_npages = pidx; 3786 bp->b_saveaddr = bp->b_data; 3787 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK); 3788 return(0); 3789} 3790 3791/* 3792 * Free the io map PTEs associated with this IO operation. 3793 * We also invalidate the TLB entries and restore the original b_addr. 3794 */ 3795void 3796vunmapbuf(struct buf *bp) 3797{ 3798 int pidx; 3799 int npages; 3800 3801 npages = bp->b_npages; 3802 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 3803 vm_page_lock_queues(); 3804 for (pidx = 0; pidx < npages; pidx++) 3805 vm_page_unhold(bp->b_pages[pidx]); 3806 vm_page_unlock_queues(); 3807 3808 bp->b_data = bp->b_saveaddr; 3809} 3810 3811void 3812bdone(struct buf *bp) 3813{ 3814 3815 mtx_lock(&bdonelock); 3816 bp->b_flags |= B_DONE; 3817 wakeup(bp); 3818 mtx_unlock(&bdonelock); 3819} 3820 3821void 3822bwait(struct buf *bp, u_char pri, const char *wchan) 3823{ 3824 3825 mtx_lock(&bdonelock); 3826 while ((bp->b_flags & B_DONE) == 0) 3827 msleep(bp, &bdonelock, pri, wchan, 0); 3828 mtx_unlock(&bdonelock); 3829} 3830 3831int 3832bufsync(struct bufobj *bo, int waitfor, struct thread *td) 3833{ 3834 3835 return (VOP_FSYNC(bo->__bo_vnode, waitfor, td)); 3836} 3837 3838void 3839bufstrategy(struct bufobj *bo, struct buf *bp) 3840{ 3841 int i = 0; 3842 struct vnode *vp; 3843 3844 vp = bp->b_vp; 3845 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 3846 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 3847 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 3848 i = VOP_STRATEGY(vp, bp); 3849 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 3850} 3851 3852void 3853bufobj_wref(struct bufobj *bo) 3854{ 3855 3856 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 3857 BO_LOCK(bo); 3858 bo->bo_numoutput++; 3859 BO_UNLOCK(bo); 3860} 3861 3862void 3863bufobj_wdrop(struct bufobj *bo) 3864{ 3865 3866 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 3867 BO_LOCK(bo); 3868 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 3869 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 3870 bo->bo_flag &= ~BO_WWAIT; 3871 wakeup(&bo->bo_numoutput); 3872 } 3873 BO_UNLOCK(bo); 3874} 3875 3876int 3877bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 3878{ 3879 int error; 3880 3881 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 3882 ASSERT_BO_LOCKED(bo); 3883 error = 0; 3884 while (bo->bo_numoutput) { 3885 bo->bo_flag |= BO_WWAIT; 3886 error = msleep(&bo->bo_numoutput, BO_MTX(bo), 3887 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 3888 if (error) 3889 break; 3890 } 3891 return (error); 3892} 3893 3894#include "opt_ddb.h" 3895#ifdef DDB 3896#include <ddb/ddb.h> 3897 3898/* DDB command to show buffer data */ 3899DB_SHOW_COMMAND(buffer, db_show_buffer) 3900{ 3901 /* get args */ 3902 struct buf *bp = (struct buf *)addr; 3903 3904 if (!have_addr) { 3905 db_printf("usage: show buffer <addr>\n"); 3906 return; 3907 } 3908 3909 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS); 3910 db_printf( 3911 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 3912 "b_bufobj = (%p), b_data = %p, b_blkno = %jd\n", 3913 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 3914 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno); 3915 if (bp->b_npages) { 3916 int i; 3917 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 3918 for (i = 0; i < bp->b_npages; i++) { 3919 vm_page_t m; 3920 m = bp->b_pages[i]; 3921 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 3922 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 3923 if ((i + 1) < bp->b_npages) 3924 db_printf(","); 3925 } 3926 db_printf("\n"); 3927 } 3928} 3929#endif /* DDB */ 3930