2 * Copyright (c) 1994,1997 John S. Dyson
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice immediately at the beginning of the file, without modification,
10 * this list of conditions, and the following disclaimer.
11 * 2. Absolutely no warranty of function or purpose is made by the author
14 * $FreeBSD: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $
15 * $DragonFly: src/sys/kern/vfs_bio.c,v 1.115 2008/08/13 11:02:31 swildner Exp $
19 * this file contains a new buffer I/O scheme implementing a coherent
20 * VM object and buffer cache scheme. Pains have been taken to make
21 * sure that the performance degradation associated with schemes such
22 * as this is not realized.
24 * Author: John S. Dyson
25 * Significant help during the development and debugging phases
26 * had been provided by David Greenman, also of the FreeBSD core team.
28 * see man buf(9) for more info.
31 #include <sys/param.h>
32 #include <sys/systm.h>
35 #include <sys/eventhandler.h>
37 #include <sys/malloc.h>
38 #include <sys/mount.h>
39 #include <sys/kernel.h>
40 #include <sys/kthread.h>
42 #include <sys/reboot.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sysctl.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_kern.h>
51 #include <vm/vm_pageout.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_map.h>
58 #include <sys/thread2.h>
59 #include <sys/spinlock2.h>
60 #include <vm/vm_page2.h>
71 BQUEUE_NONE, /* not on any queue */
72 BQUEUE_LOCKED, /* locked buffers */
73 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
74 BQUEUE_DIRTY, /* B_DELWRI buffers */
75 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
76 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
77 BQUEUE_EMPTY, /* empty buffer headers */
79 BUFFER_QUEUES /* number of buffer queues */
82 typedef enum bufq_type bufq_type_t;
84 #define BD_WAKE_SIZE 128
85 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
87 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
88 struct spinlock bufspin = SPINLOCK_INITIALIZER(&bufspin);
90 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
92 struct buf *buf; /* buffer header pool */
94 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
95 int pageno, vm_page_t m);
96 static void vfs_clean_pages(struct buf *bp);
97 static void vfs_setdirty(struct buf *bp);
98 static void vfs_vmio_release(struct buf *bp);
99 static int flushbufqueues(bufq_type_t q);
100 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);
102 static void bd_signal(int totalspace);
103 static void buf_daemon(void);
104 static void buf_daemon_hw(void);
107 * bogus page -- for I/O to/from partially complete buffers
108 * this is a temporary solution to the problem, but it is not
109 * really that bad. it would be better to split the buffer
110 * for input in the case of buffers partially already in memory,
111 * but the code is intricate enough already.
113 vm_page_t bogus_page;
116 * These are all static, but make the ones we export globals so we do
117 * not need to use compiler magic.
119 int bufspace, maxbufspace,
120 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
121 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
122 static int lorunningspace, hirunningspace, runningbufreq;
123 int dirtybufspace, dirtybufspacehw, lodirtybufspace, hidirtybufspace;
124 int dirtybufcount, dirtybufcounthw;
125 int runningbufspace, runningbufcount;
126 static int getnewbufcalls;
127 static int getnewbufrestarts;
128 static int recoverbufcalls;
129 static int needsbuffer; /* locked by needsbuffer_spin */
130 static int bd_request; /* locked by needsbuffer_spin */
131 static int bd_request_hw; /* locked by needsbuffer_spin */
132 static u_int bd_wake_ary[BD_WAKE_SIZE];
133 static u_int bd_wake_index;
134 static struct spinlock needsbuffer_spin;
136 static struct thread *bufdaemon_td;
137 static struct thread *bufdaemonhw_td;
141 * Sysctls for operational control of the buffer cache.
143 SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
144 "Number of dirty buffers to flush before bufdaemon becomes inactive");
145 SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
146 "High watermark used to trigger explicit flushing of dirty buffers");
147 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
148 "Minimum amount of buffer space required for active I/O");
149 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
150 "Maximum amount of buffer space to usable for active I/O");
152 * Sysctls determining current state of the buffer cache.
154 SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
155 "Total number of buffers in buffer cache");
156 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
157 "Pending bytes of dirty buffers (all)");
158 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
159 "Pending bytes of dirty buffers (heavy weight)");
160 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
161 "Pending number of dirty buffers");
162 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
163 "Pending number of dirty buffers (heavy weight)");
164 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
165 "I/O bytes currently in progress due to asynchronous writes");
166 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
167 "I/O buffers currently in progress due to asynchronous writes");
168 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
169 "Hard limit on maximum amount of memory usable for buffer space");
170 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
171 "Soft limit on maximum amount of memory usable for buffer space");
172 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
173 "Minimum amount of memory to reserve for system buffer space");
174 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
175 "Amount of memory available for buffers");
176 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
177 0, "Maximum amount of memory reserved for buffers using malloc");
178 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
179 "Amount of memory left for buffers using malloc-scheme");
180 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
181 "New buffer header acquisition requests");
182 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
183 0, "New buffer header acquisition restarts");
184 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
185 "Recover VM space in an emergency");
186 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
187 "Buffer acquisition restarts due to fragmented buffer map");
188 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
189 "Amount of time KVA space was deallocated in an arbitrary buffer");
190 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
191 "Amount of time buffer re-use operations were successful");
192 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
193 "sizeof(struct buf)");
195 char *buf_wmesg = BUF_WMESG;
197 extern int vm_swap_size;
199 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
200 #define VFS_BIO_NEED_UNUSED02 0x02
201 #define VFS_BIO_NEED_UNUSED04 0x04
202 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
207 * Called when buffer space is potentially available for recovery.
208 * getnewbuf() will block on this flag when it is unable to free
209 * sufficient buffer space. Buffer space becomes recoverable when
210 * bp's get placed back in the queues.
217 * If someone is waiting for BUF space, wake them up. Even
218 * though we haven't freed the kva space yet, the waiting
219 * process will be able to now.
221 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
222 spin_lock_wr(&needsbuffer_spin);
223 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
224 spin_unlock_wr(&needsbuffer_spin);
225 wakeup(&needsbuffer);
232 * Accounting for I/O in progress.
236 runningbufwakeup(struct buf *bp)
240 if ((totalspace = bp->b_runningbufspace) != 0) {
241 runningbufspace -= totalspace;
243 bp->b_runningbufspace = 0;
244 if (runningbufreq && runningbufspace <= lorunningspace) {
246 wakeup(&runningbufreq);
248 bd_signal(totalspace);
255 * Called when a buffer has been added to one of the free queues to
256 * account for the buffer and to wakeup anyone waiting for free buffers.
257 * This typically occurs when large amounts of metadata are being handled
258 * by the buffer cache ( else buffer space runs out first, usually ).
266 spin_lock_wr(&needsbuffer_spin);
267 needsbuffer &= ~VFS_BIO_NEED_ANY;
268 spin_unlock_wr(&needsbuffer_spin);
269 wakeup(&needsbuffer);
274 * waitrunningbufspace()
276 * Wait for the amount of running I/O to drop to a reasonable level.
278 * The caller may be using this function to block in a tight loop, we
279 * must block of runningbufspace is greater then the passed limit.
280 * And even with that it may not be enough, due to the presence of
281 * B_LOCKED dirty buffers, so also wait for at least one running buffer
285 waitrunningbufspace(int limit)
289 if (lorunningspace < limit)
290 lorun = lorunningspace;
295 if (runningbufspace > lorun) {
296 while (runningbufspace > lorun) {
298 tsleep(&runningbufreq, 0, "wdrain", 0);
300 } else if (runningbufspace) {
302 tsleep(&runningbufreq, 0, "wdrain2", 1);
308 * vfs_buf_test_cache:
310 * Called when a buffer is extended. This function clears the B_CACHE
311 * bit if the newly extended portion of the buffer does not contain
316 vfs_buf_test_cache(struct buf *bp,
317 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
320 if (bp->b_flags & B_CACHE) {
321 int base = (foff + off) & PAGE_MASK;
322 if (vm_page_is_valid(m, base, size) == 0)
323 bp->b_flags &= ~B_CACHE;
330 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
339 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
342 if (bd_request == 0 &&
343 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
344 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
345 spin_lock_wr(&needsbuffer_spin);
347 spin_unlock_wr(&needsbuffer_spin);
350 if (bd_request_hw == 0 &&
351 (dirtybufspacehw > lodirtybufspace / 2 ||
352 dirtybufcounthw >= nbuf / 2)) {
353 spin_lock_wr(&needsbuffer_spin);
355 spin_unlock_wr(&needsbuffer_spin);
356 wakeup(&bd_request_hw);
363 * Get the buf_daemon heated up when the number of running and dirty
364 * buffers exceeds the mid-point.
375 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
377 totalspace = runningbufspace + dirtybufspace;
378 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
380 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
381 if (totalspace >= mid2)
382 return(totalspace - mid2);
390 * Wait for the buffer cache to flush (totalspace) bytes worth of
391 * buffers, then return.
393 * Regardless this function blocks while the number of dirty buffers
394 * exceeds hidirtybufspace.
399 bd_wait(int totalspace)
404 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
407 while (totalspace > 0) {
410 if (totalspace > runningbufspace + dirtybufspace)
411 totalspace = runningbufspace + dirtybufspace;
412 count = totalspace / BKVASIZE;
413 if (count >= BD_WAKE_SIZE)
414 count = BD_WAKE_SIZE - 1;
416 spin_lock_wr(&needsbuffer_spin);
417 i = (bd_wake_index + count) & BD_WAKE_MASK;
419 tsleep_interlock(&bd_wake_ary[i]);
420 spin_unlock_wr(&needsbuffer_spin);
422 tsleep(&bd_wake_ary[i], 0, "flstik", hz);
425 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
432 * This function is called whenever runningbufspace or dirtybufspace
433 * is reduced. Track threads waiting for run+dirty buffer I/O
439 bd_signal(int totalspace)
443 if (totalspace > 0) {
444 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
445 totalspace = BKVASIZE * BD_WAKE_SIZE;
446 spin_lock_wr(&needsbuffer_spin);
447 while (totalspace > 0) {
450 if (bd_wake_ary[i]) {
452 spin_unlock_wr(&needsbuffer_spin);
453 wakeup(&bd_wake_ary[i]);
454 spin_lock_wr(&needsbuffer_spin);
456 totalspace -= BKVASIZE;
458 spin_unlock_wr(&needsbuffer_spin);
463 * BIO tracking support routines.
465 * Release a ref on a bio_track. Wakeup requests are atomically released
466 * along with the last reference so bk_active will never wind up set to
473 bio_track_rel(struct bio_track *track)
481 active = track->bk_active;
482 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
486 * Full-on. Note that the wait flag is only atomically released on
487 * the 1->0 count transition.
489 * We check for a negative count transition using bit 30 since bit 31
490 * has a different meaning.
493 desired = (active & 0x7FFFFFFF) - 1;
495 desired |= active & 0x80000000;
496 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
497 if (desired & 0x40000000)
498 panic("bio_track_rel: bad count: %p\n", track);
499 if (active & 0x80000000)
503 active = track->bk_active;
508 * Wait for the tracking count to reach 0.
510 * Use atomic ops such that the wait flag is only set atomically when
511 * bk_active is non-zero.
516 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
525 if (track->bk_active == 0)
529 * Full-on. Note that the wait flag may only be atomically set if
530 * the active count is non-zero.
532 crit_enter(); /* for tsleep_interlock */
534 while ((active = track->bk_active) != 0) {
535 desired = active | 0x80000000;
536 tsleep_interlock(track);
537 if (active == desired ||
538 atomic_cmpset_int(&track->bk_active, active, desired)) {
539 error = tsleep(track, slp_flags, "iowait", slp_timo);
551 * Load time initialisation of the buffer cache, called from machine
552 * dependant initialization code.
558 vm_offset_t bogus_offset;
561 spin_init(&needsbuffer_spin);
563 /* next, make a null set of free lists */
564 for (i = 0; i < BUFFER_QUEUES; i++)
565 TAILQ_INIT(&bufqueues[i]);
567 /* finally, initialize each buffer header and stick on empty q */
568 for (i = 0; i < nbuf; i++) {
570 bzero(bp, sizeof *bp);
571 bp->b_flags = B_INVAL; /* we're just an empty header */
572 bp->b_cmd = BUF_CMD_DONE;
573 bp->b_qindex = BQUEUE_EMPTY;
575 xio_init(&bp->b_xio);
578 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
582 * maxbufspace is the absolute maximum amount of buffer space we are
583 * allowed to reserve in KVM and in real terms. The absolute maximum
584 * is nominally used by buf_daemon. hibufspace is the nominal maximum
585 * used by most other processes. The differential is required to
586 * ensure that buf_daemon is able to run when other processes might
587 * be blocked waiting for buffer space.
589 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
590 * this may result in KVM fragmentation which is not handled optimally
593 maxbufspace = nbuf * BKVASIZE;
594 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
595 lobufspace = hibufspace - MAXBSIZE;
597 lorunningspace = 512 * 1024;
598 hirunningspace = 1024 * 1024;
601 * Limit the amount of malloc memory since it is wired permanently
602 * into the kernel space. Even though this is accounted for in
603 * the buffer allocation, we don't want the malloced region to grow
604 * uncontrolled. The malloc scheme improves memory utilization
605 * significantly on average (small) directories.
607 maxbufmallocspace = hibufspace / 20;
610 * Reduce the chance of a deadlock occuring by limiting the number
611 * of delayed-write dirty buffers we allow to stack up.
613 hidirtybufspace = hibufspace / 2;
617 lodirtybufspace = hidirtybufspace / 2;
620 * Maximum number of async ops initiated per buf_daemon loop. This is
621 * somewhat of a hack at the moment, we really need to limit ourselves
622 * based on the number of bytes of I/O in-transit that were initiated
626 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
627 bogus_page = vm_page_alloc(&kernel_object,
628 (bogus_offset >> PAGE_SHIFT),
630 vmstats.v_wire_count++;
635 * Initialize the embedded bio structures
638 initbufbio(struct buf *bp)
640 bp->b_bio1.bio_buf = bp;
641 bp->b_bio1.bio_prev = NULL;
642 bp->b_bio1.bio_offset = NOOFFSET;
643 bp->b_bio1.bio_next = &bp->b_bio2;
644 bp->b_bio1.bio_done = NULL;
646 bp->b_bio2.bio_buf = bp;
647 bp->b_bio2.bio_prev = &bp->b_bio1;
648 bp->b_bio2.bio_offset = NOOFFSET;
649 bp->b_bio2.bio_next = NULL;
650 bp->b_bio2.bio_done = NULL;
654 * Reinitialize the embedded bio structures as well as any additional
655 * translation cache layers.
658 reinitbufbio(struct buf *bp)
662 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
663 bio->bio_done = NULL;
664 bio->bio_offset = NOOFFSET;
669 * Push another BIO layer onto an existing BIO and return it. The new
670 * BIO layer may already exist, holding cached translation data.
673 push_bio(struct bio *bio)
677 if ((nbio = bio->bio_next) == NULL) {
678 int index = bio - &bio->bio_buf->b_bio_array[0];
679 if (index >= NBUF_BIO - 1) {
680 panic("push_bio: too many layers bp %p\n",
683 nbio = &bio->bio_buf->b_bio_array[index + 1];
684 bio->bio_next = nbio;
685 nbio->bio_prev = bio;
686 nbio->bio_buf = bio->bio_buf;
687 nbio->bio_offset = NOOFFSET;
688 nbio->bio_done = NULL;
689 nbio->bio_next = NULL;
691 KKASSERT(nbio->bio_done == NULL);
696 * Pop a BIO translation layer, returning the previous layer. The
697 * must have been previously pushed.
700 pop_bio(struct bio *bio)
702 return(bio->bio_prev);
706 clearbiocache(struct bio *bio)
709 bio->bio_offset = NOOFFSET;
717 * Free the KVA allocation for buffer 'bp'.
719 * Must be called from a critical section as this is the only locking for
722 * Since this call frees up buffer space, we call bufspacewakeup().
727 bfreekva(struct buf *bp)
734 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
735 vm_map_lock(&buffer_map);
736 bufspace -= bp->b_kvasize;
737 vm_map_delete(&buffer_map,
738 (vm_offset_t) bp->b_kvabase,
739 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
742 vm_map_unlock(&buffer_map);
743 vm_map_entry_release(count);
753 * Remove the buffer from the appropriate free list.
756 _bremfree(struct buf *bp)
758 if (bp->b_qindex != BQUEUE_NONE) {
759 KASSERT(BUF_REFCNTNB(bp) == 1,
760 ("bremfree: bp %p not locked",bp));
761 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
762 bp->b_qindex = BQUEUE_NONE;
764 if (BUF_REFCNTNB(bp) <= 1)
765 panic("bremfree: removing a buffer not on a queue");
770 bremfree(struct buf *bp)
772 spin_lock_wr(&bufspin);
774 spin_unlock_wr(&bufspin);
778 bremfree_locked(struct buf *bp)
786 * Get a buffer with the specified data. Look in the cache first. We
787 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
788 * is set, the buffer is valid and we do not have to do anything ( see
794 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
798 bp = getblk(vp, loffset, size, 0, 0);
801 /* if not found in cache, do some I/O */
802 if ((bp->b_flags & B_CACHE) == 0) {
804 KASSERT(!(bp->b_flags & B_ASYNC),
805 ("bread: illegal async bp %p", bp));
806 bp->b_flags &= ~(B_ERROR | B_INVAL);
807 bp->b_cmd = BUF_CMD_READ;
808 vfs_busy_pages(vp, bp);
809 vn_strategy(vp, &bp->b_bio1);
811 return (biowait(bp));
819 * Operates like bread, but also starts asynchronous I/O on
820 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
821 * to initiating I/O . If B_CACHE is set, the buffer is valid
822 * and we do not have to do anything.
827 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
828 int *rabsize, int cnt, struct buf **bpp)
830 struct buf *bp, *rabp;
832 int rv = 0, readwait = 0;
834 *bpp = bp = getblk(vp, loffset, size, 0, 0);
836 /* if not found in cache, do some I/O */
837 if ((bp->b_flags & B_CACHE) == 0) {
839 bp->b_flags &= ~(B_ERROR | B_INVAL);
840 bp->b_cmd = BUF_CMD_READ;
841 vfs_busy_pages(vp, bp);
842 vn_strategy(vp, &bp->b_bio1);
847 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
848 if (inmem(vp, *raoffset))
850 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
852 if ((rabp->b_flags & B_CACHE) == 0) {
854 rabp->b_flags |= B_ASYNC;
855 rabp->b_flags &= ~(B_ERROR | B_INVAL);
856 rabp->b_cmd = BUF_CMD_READ;
857 vfs_busy_pages(vp, rabp);
859 vn_strategy(vp, &rabp->b_bio1);
873 * Write, release buffer on completion. (Done by iodone
874 * if async). Do not bother writing anything if the buffer
877 * Note that we set B_CACHE here, indicating that buffer is
878 * fully valid and thus cacheable. This is true even of NFS
879 * now so we set it generally. This could be set either here
880 * or in biodone() since the I/O is synchronous. We put it
884 bwrite(struct buf *bp)
888 if (bp->b_flags & B_INVAL) {
893 oldflags = bp->b_flags;
895 if (BUF_REFCNTNB(bp) == 0)
896 panic("bwrite: buffer is not busy???");
899 /* Mark the buffer clean */
902 bp->b_flags &= ~B_ERROR;
903 bp->b_flags |= B_CACHE;
904 bp->b_cmd = BUF_CMD_WRITE;
905 vfs_busy_pages(bp->b_vp, bp);
908 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
909 * valid for vnode-backed buffers.
911 bp->b_runningbufspace = bp->b_bufsize;
912 if (bp->b_runningbufspace) {
913 runningbufspace += bp->b_runningbufspace;
918 if (oldflags & B_ASYNC)
920 vn_strategy(bp->b_vp, &bp->b_bio1);
922 if ((oldflags & B_ASYNC) == 0) {
923 int rtval = biowait(bp);
933 * Delayed write. (Buffer is marked dirty). Do not bother writing
934 * anything if the buffer is marked invalid.
936 * Note that since the buffer must be completely valid, we can safely
937 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
938 * biodone() in order to prevent getblk from writing the buffer
942 bdwrite(struct buf *bp)
944 if (BUF_REFCNTNB(bp) == 0)
945 panic("bdwrite: buffer is not busy");
947 if (bp->b_flags & B_INVAL) {
954 * Set B_CACHE, indicating that the buffer is fully valid. This is
955 * true even of NFS now.
957 bp->b_flags |= B_CACHE;
960 * This bmap keeps the system from needing to do the bmap later,
961 * perhaps when the system is attempting to do a sync. Since it
962 * is likely that the indirect block -- or whatever other datastructure
963 * that the filesystem needs is still in memory now, it is a good
964 * thing to do this. Note also, that if the pageout daemon is
965 * requesting a sync -- there might not be enough memory to do
966 * the bmap then... So, this is important to do.
968 if (bp->b_bio2.bio_offset == NOOFFSET) {
969 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
970 NULL, NULL, BUF_CMD_WRITE);
974 * Set the *dirty* buffer range based upon the VM system dirty pages.
979 * We need to do this here to satisfy the vnode_pager and the
980 * pageout daemon, so that it thinks that the pages have been
981 * "cleaned". Note that since the pages are in a delayed write
982 * buffer -- the VFS layer "will" see that the pages get written
983 * out on the next sync, or perhaps the cluster will be completed.
989 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
990 * due to the softdep code.
997 * Turn buffer into delayed write request by marking it B_DELWRI.
998 * B_RELBUF and B_NOCACHE must be cleared.
1000 * We reassign the buffer to itself to properly update it in the
1001 * dirty/clean lists.
1003 * Must be called from a critical section.
1004 * The buffer must be on BQUEUE_NONE.
1007 bdirty(struct buf *bp)
1009 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1010 if (bp->b_flags & B_NOCACHE) {
1011 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1012 bp->b_flags &= ~B_NOCACHE;
1014 if (bp->b_flags & B_INVAL) {
1015 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1017 bp->b_flags &= ~B_RELBUF;
1019 if ((bp->b_flags & B_DELWRI) == 0) {
1020 bp->b_flags |= B_DELWRI;
1022 atomic_add_int(&dirtybufcount, 1);
1023 dirtybufspace += bp->b_bufsize;
1024 if (bp->b_flags & B_HEAVY) {
1025 atomic_add_int(&dirtybufcounthw, 1);
1026 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
1033 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1034 * needs to be flushed with a different buf_daemon thread to avoid
1035 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1038 bheavy(struct buf *bp)
1040 if ((bp->b_flags & B_HEAVY) == 0) {
1041 bp->b_flags |= B_HEAVY;
1042 if (bp->b_flags & B_DELWRI) {
1043 atomic_add_int(&dirtybufcounthw, 1);
1044 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
1052 * Clear B_DELWRI for buffer.
1054 * Must be called from a critical section.
1056 * The buffer is typically on BQUEUE_NONE but there is one case in
1057 * brelse() that calls this function after placing the buffer on
1058 * a different queue.
1063 bundirty(struct buf *bp)
1065 if (bp->b_flags & B_DELWRI) {
1066 bp->b_flags &= ~B_DELWRI;
1068 atomic_subtract_int(&dirtybufcount, 1);
1069 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
1070 if (bp->b_flags & B_HEAVY) {
1071 atomic_subtract_int(&dirtybufcounthw, 1);
1072 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
1074 bd_signal(bp->b_bufsize);
1077 * Since it is now being written, we can clear its deferred write flag.
1079 bp->b_flags &= ~B_DEFERRED;
1085 * Asynchronous write. Start output on a buffer, but do not wait for
1086 * it to complete. The buffer is released when the output completes.
1088 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1089 * B_INVAL buffers. Not us.
1092 bawrite(struct buf *bp)
1094 bp->b_flags |= B_ASYNC;
1101 * Ordered write. Start output on a buffer, and flag it so that the
1102 * device will write it in the order it was queued. The buffer is
1103 * released when the output completes. bwrite() ( or the VOP routine
1104 * anyway ) is responsible for handling B_INVAL buffers.
1107 bowrite(struct buf *bp)
1109 bp->b_flags |= B_ORDERED | B_ASYNC;
1110 return (bwrite(bp));
1114 * buf_dirty_count_severe:
1116 * Return true if we have too many dirty buffers.
1119 buf_dirty_count_severe(void)
1121 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
1122 dirtybufcount >= nbuf / 2);
1128 * Release a busy buffer and, if requested, free its resources. The
1129 * buffer will be stashed in the appropriate bufqueue[] allowing it
1130 * to be accessed later as a cache entity or reused for other purposes.
1135 brelse(struct buf *bp)
1138 int saved_flags = bp->b_flags;
1141 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1144 * If B_NOCACHE is set we are being asked to destroy the buffer and
1145 * its backing store. Clear B_DELWRI.
1147 * B_NOCACHE is set in two cases: (1) when the caller really wants
1148 * to destroy the buffer and backing store and (2) when the caller
1149 * wants to destroy the buffer and backing store after a write
1152 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1156 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1158 * A re-dirtied buffer is only subject to destruction
1159 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1161 /* leave buffer intact */
1162 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1163 (bp->b_bufsize <= 0)) {
1165 * Either a failed read or we were asked to free or not
1166 * cache the buffer. This path is reached with B_DELWRI
1167 * set only if B_INVAL is already set. B_NOCACHE governs
1168 * backing store destruction.
1170 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1171 * buffer cannot be immediately freed.
1173 bp->b_flags |= B_INVAL;
1174 if (LIST_FIRST(&bp->b_dep) != NULL) {
1179 if (bp->b_flags & B_DELWRI) {
1180 atomic_subtract_int(&dirtybufcount, 1);
1181 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
1182 if (bp->b_flags & B_HEAVY) {
1183 atomic_subtract_int(&dirtybufcounthw, 1);
1184 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
1186 bd_signal(bp->b_bufsize);
1188 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1192 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1193 * If vfs_vmio_release() is called with either bit set, the
1194 * underlying pages may wind up getting freed causing a previous
1195 * write (bdwrite()) to get 'lost' because pages associated with
1196 * a B_DELWRI bp are marked clean. Pages associated with a
1197 * B_LOCKED buffer may be mapped by the filesystem.
1199 * If we want to release the buffer ourselves (rather then the
1200 * originator asking us to release it), give the originator a
1201 * chance to countermand the release by setting B_LOCKED.
1203 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1204 * if B_DELWRI is set.
1206 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1207 * on pages to return pages to the VM page queues.
1209 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1210 bp->b_flags &= ~B_RELBUF;
1211 } else if (vm_page_count_severe()) {
1212 if (LIST_FIRST(&bp->b_dep) != NULL) {
1214 buf_deallocate(bp); /* can set B_LOCKED */
1217 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1218 bp->b_flags &= ~B_RELBUF;
1220 bp->b_flags |= B_RELBUF;
1224 * Make sure b_cmd is clear. It may have already been cleared by
1227 * At this point destroying the buffer is governed by the B_INVAL
1228 * or B_RELBUF flags.
1230 bp->b_cmd = BUF_CMD_DONE;
1233 * VMIO buffer rundown. Make sure the VM page array is restored
1234 * after an I/O may have replaces some of the pages with bogus pages
1235 * in order to not destroy dirty pages in a fill-in read.
1237 * Note that due to the code above, if a buffer is marked B_DELWRI
1238 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1239 * B_INVAL may still be set, however.
1241 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1242 * but not the backing store. B_NOCACHE will destroy the backing
1245 * Note that dirty NFS buffers contain byte-granular write ranges
1246 * and should not be destroyed w/ B_INVAL even if the backing store
1249 if (bp->b_flags & B_VMIO) {
1251 * Rundown for VMIO buffers which are not dirty NFS buffers.
1263 * Get the base offset and length of the buffer. Note that
1264 * in the VMIO case if the buffer block size is not
1265 * page-aligned then b_data pointer may not be page-aligned.
1266 * But our b_xio.xio_pages array *IS* page aligned.
1268 * block sizes less then DEV_BSIZE (usually 512) are not
1269 * supported due to the page granularity bits (m->valid,
1270 * m->dirty, etc...).
1272 * See man buf(9) for more information
1275 resid = bp->b_bufsize;
1276 foff = bp->b_loffset;
1279 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1280 m = bp->b_xio.xio_pages[i];
1281 vm_page_flag_clear(m, PG_ZERO);
1283 * If we hit a bogus page, fixup *all* of them
1284 * now. Note that we left these pages wired
1285 * when we removed them so they had better exist,
1286 * and they cannot be ripped out from under us so
1287 * no critical section protection is necessary.
1289 if (m == bogus_page) {
1291 poff = OFF_TO_IDX(bp->b_loffset);
1293 for (j = i; j < bp->b_xio.xio_npages; j++) {
1296 mtmp = bp->b_xio.xio_pages[j];
1297 if (mtmp == bogus_page) {
1298 mtmp = vm_page_lookup(obj, poff + j);
1300 panic("brelse: page missing");
1302 bp->b_xio.xio_pages[j] = mtmp;
1306 if ((bp->b_flags & B_INVAL) == 0) {
1307 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1308 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1310 m = bp->b_xio.xio_pages[i];
1314 * Invalidate the backing store if B_NOCACHE is set
1315 * (e.g. used with vinvalbuf()). If this is NFS
1316 * we impose a requirement that the block size be
1317 * a multiple of PAGE_SIZE and create a temporary
1318 * hack to basically invalidate the whole page. The
1319 * problem is that NFS uses really odd buffer sizes
1320 * especially when tracking piecemeal writes and
1321 * it also vinvalbuf()'s a lot, which would result
1322 * in only partial page validation and invalidation
1323 * here. If the file page is mmap()'d, however,
1324 * all the valid bits get set so after we invalidate
1325 * here we would end up with weird m->valid values
1326 * like 0xfc. nfs_getpages() can't handle this so
1327 * we clear all the valid bits for the NFS case
1328 * instead of just some of them.
1330 * The real bug is the VM system having to set m->valid
1331 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1332 * itself is an artifact of the whole 512-byte
1333 * granular mess that exists to support odd block
1334 * sizes and UFS meta-data block sizes (e.g. 6144).
1335 * A complete rewrite is required.
1337 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1338 int poffset = foff & PAGE_MASK;
1341 presid = PAGE_SIZE - poffset;
1342 if (bp->b_vp->v_tag == VT_NFS &&
1343 bp->b_vp->v_type == VREG) {
1345 } else if (presid > resid) {
1348 KASSERT(presid >= 0, ("brelse: extra page"));
1349 vm_page_set_invalid(m, poffset, presid);
1351 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1352 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1354 if (bp->b_flags & (B_INVAL | B_RELBUF))
1355 vfs_vmio_release(bp);
1359 * Rundown for non-VMIO buffers.
1361 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1365 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1372 if (bp->b_qindex != BQUEUE_NONE)
1373 panic("brelse: free buffer onto another queue???");
1374 if (BUF_REFCNTNB(bp) > 1) {
1375 /* Temporary panic to verify exclusive locking */
1376 /* This panic goes away when we allow shared refs */
1377 panic("brelse: multiple refs");
1383 * Figure out the correct queue to place the cleaned up buffer on.
1384 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1385 * disassociated from their vnode.
1387 spin_lock_wr(&bufspin);
1388 if (bp->b_flags & B_LOCKED) {
1390 * Buffers that are locked are placed in the locked queue
1391 * immediately, regardless of their state.
1393 bp->b_qindex = BQUEUE_LOCKED;
1394 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1395 } else if (bp->b_bufsize == 0) {
1397 * Buffers with no memory. Due to conditionals near the top
1398 * of brelse() such buffers should probably already be
1399 * marked B_INVAL and disassociated from their vnode.
1401 bp->b_flags |= B_INVAL;
1402 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1403 KKASSERT((bp->b_flags & B_HASHED) == 0);
1404 if (bp->b_kvasize) {
1405 bp->b_qindex = BQUEUE_EMPTYKVA;
1407 bp->b_qindex = BQUEUE_EMPTY;
1409 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1410 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1412 * Buffers with junk contents. Again these buffers had better
1413 * already be disassociated from their vnode.
1415 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1416 KKASSERT((bp->b_flags & B_HASHED) == 0);
1417 bp->b_flags |= B_INVAL;
1418 bp->b_qindex = BQUEUE_CLEAN;
1419 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1422 * Remaining buffers. These buffers are still associated with
1425 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1427 bp->b_qindex = BQUEUE_DIRTY;
1428 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1430 case B_DELWRI | B_HEAVY:
1431 bp->b_qindex = BQUEUE_DIRTY_HW;
1432 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1437 * NOTE: Buffers are always placed at the end of the
1438 * queue. If B_AGE is not set the buffer will cycle
1439 * through the queue twice.
1441 bp->b_qindex = BQUEUE_CLEAN;
1442 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1446 spin_unlock_wr(&bufspin);
1449 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1450 * on the correct queue.
1452 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1456 * The bp is on an appropriate queue unless locked. If it is not
1457 * locked or dirty we can wakeup threads waiting for buffer space.
1459 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1460 * if B_INVAL is set ).
1462 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1466 * Something we can maybe free or reuse
1468 if (bp->b_bufsize || bp->b_kvasize)
1472 * Clean up temporary flags and unlock the buffer.
1474 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
1481 * Release a buffer back to the appropriate queue but do not try to free
1482 * it. The buffer is expected to be used again soon.
1484 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1485 * biodone() to requeue an async I/O on completion. It is also used when
1486 * known good buffers need to be requeued but we think we may need the data
1489 * XXX we should be able to leave the B_RELBUF hint set on completion.
1494 bqrelse(struct buf *bp)
1496 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1498 if (bp->b_qindex != BQUEUE_NONE)
1499 panic("bqrelse: free buffer onto another queue???");
1500 if (BUF_REFCNTNB(bp) > 1) {
1501 /* do not release to free list */
1502 panic("bqrelse: multiple refs");
1506 spin_lock_wr(&bufspin);
1507 if (bp->b_flags & B_LOCKED) {
1509 * Locked buffers are released to the locked queue. However,
1510 * if the buffer is dirty it will first go into the dirty
1511 * queue and later on after the I/O completes successfully it
1512 * will be released to the locked queue.
1514 bp->b_qindex = BQUEUE_LOCKED;
1515 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1516 } else if (bp->b_flags & B_DELWRI) {
1517 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1518 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1519 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1520 } else if (vm_page_count_severe()) {
1522 * We are too low on memory, we have to try to free the
1523 * buffer (most importantly: the wired pages making up its
1524 * backing store) *now*.
1526 spin_unlock_wr(&bufspin);
1530 bp->b_qindex = BQUEUE_CLEAN;
1531 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1533 spin_unlock_wr(&bufspin);
1535 if ((bp->b_flags & B_LOCKED) == 0 &&
1536 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1541 * Something we can maybe free or reuse.
1543 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1547 * Final cleanup and unlock. Clear bits that are only used while a
1548 * buffer is actively locked.
1550 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF);
1557 * Return backing pages held by the buffer 'bp' back to the VM system
1558 * if possible. The pages are freed if they are no longer valid or
1559 * attempt to free if it was used for direct I/O otherwise they are
1560 * sent to the page cache.
1562 * Pages that were marked busy are left alone and skipped.
1564 * The KVA mapping (b_data) for the underlying pages is removed by
1568 vfs_vmio_release(struct buf *bp)
1574 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1575 m = bp->b_xio.xio_pages[i];
1576 bp->b_xio.xio_pages[i] = NULL;
1578 * In order to keep page LRU ordering consistent, put
1579 * everything on the inactive queue.
1581 vm_page_unwire(m, 0);
1583 * We don't mess with busy pages, it is
1584 * the responsibility of the process that
1585 * busied the pages to deal with them.
1587 if ((m->flags & PG_BUSY) || (m->busy != 0))
1590 if (m->wire_count == 0) {
1591 vm_page_flag_clear(m, PG_ZERO);
1593 * Might as well free the page if we can and it has
1594 * no valid data. We also free the page if the
1595 * buffer was used for direct I/O.
1597 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1598 m->hold_count == 0) {
1600 vm_page_protect(m, VM_PROT_NONE);
1602 } else if (bp->b_flags & B_DIRECT) {
1603 vm_page_try_to_free(m);
1604 } else if (vm_page_count_severe()) {
1605 vm_page_try_to_cache(m);
1610 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1611 if (bp->b_bufsize) {
1615 bp->b_xio.xio_npages = 0;
1616 bp->b_flags &= ~B_VMIO;
1617 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1628 * Implement clustered async writes for clearing out B_DELWRI buffers.
1629 * This is much better then the old way of writing only one buffer at
1630 * a time. Note that we may not be presented with the buffers in the
1631 * correct order, so we search for the cluster in both directions.
1633 * The buffer is locked on call.
1636 vfs_bio_awrite(struct buf *bp)
1640 off_t loffset = bp->b_loffset;
1641 struct vnode *vp = bp->b_vp;
1648 * right now we support clustered writing only to regular files. If
1649 * we find a clusterable block we could be in the middle of a cluster
1650 * rather then at the beginning.
1652 * NOTE: b_bio1 contains the logical loffset and is aliased
1653 * to b_loffset. b_bio2 contains the translated block number.
1655 if ((vp->v_type == VREG) &&
1656 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1657 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1659 size = vp->v_mount->mnt_stat.f_iosize;
1661 for (i = size; i < MAXPHYS; i += size) {
1662 if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) &&
1663 BUF_REFCNT(bpa) == 0 &&
1664 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1665 (B_DELWRI | B_CLUSTEROK)) &&
1666 (bpa->b_bufsize == size)) {
1667 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1668 (bpa->b_bio2.bio_offset !=
1669 bp->b_bio2.bio_offset + i))
1675 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1676 if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) &&
1677 BUF_REFCNT(bpa) == 0 &&
1678 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1679 (B_DELWRI | B_CLUSTEROK)) &&
1680 (bpa->b_bufsize == size)) {
1681 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1682 (bpa->b_bio2.bio_offset !=
1683 bp->b_bio2.bio_offset - j))
1693 * this is a possible cluster write
1695 if (nbytes != size) {
1697 nwritten = cluster_wbuild(vp, size,
1698 loffset - j, nbytes);
1704 bp->b_flags |= B_ASYNC;
1707 * default (old) behavior, writing out only one block
1709 * XXX returns b_bufsize instead of b_bcount for nwritten?
1711 nwritten = bp->b_bufsize;
1720 * Find and initialize a new buffer header, freeing up existing buffers
1721 * in the bufqueues as necessary. The new buffer is returned locked.
1723 * Important: B_INVAL is not set. If the caller wishes to throw the
1724 * buffer away, the caller must set B_INVAL prior to calling brelse().
1727 * We have insufficient buffer headers
1728 * We have insufficient buffer space
1729 * buffer_map is too fragmented ( space reservation fails )
1730 * If we have to flush dirty buffers ( but we try to avoid this )
1732 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1733 * Instead we ask the buf daemon to do it for us. We attempt to
1734 * avoid piecemeal wakeups of the pageout daemon.
1739 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1745 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1746 static int flushingbufs;
1749 * We can't afford to block since we might be holding a vnode lock,
1750 * which may prevent system daemons from running. We deal with
1751 * low-memory situations by proactively returning memory and running
1752 * async I/O rather then sync I/O.
1756 --getnewbufrestarts;
1758 ++getnewbufrestarts;
1761 * Setup for scan. If we do not have enough free buffers,
1762 * we setup a degenerate case that immediately fails. Note
1763 * that if we are specially marked process, we are allowed to
1764 * dip into our reserves.
1766 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1768 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1769 * However, there are a number of cases (defragging, reusing, ...)
1770 * where we cannot backup.
1772 nqindex = BQUEUE_EMPTYKVA;
1773 spin_lock_wr(&bufspin);
1774 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1778 * If no EMPTYKVA buffers and we are either
1779 * defragging or reusing, locate a CLEAN buffer
1780 * to free or reuse. If bufspace useage is low
1781 * skip this step so we can allocate a new buffer.
1783 if (defrag || bufspace >= lobufspace) {
1784 nqindex = BQUEUE_CLEAN;
1785 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1789 * If we could not find or were not allowed to reuse a
1790 * CLEAN buffer, check to see if it is ok to use an EMPTY
1791 * buffer. We can only use an EMPTY buffer if allocating
1792 * its KVA would not otherwise run us out of buffer space.
1794 if (nbp == NULL && defrag == 0 &&
1795 bufspace + maxsize < hibufspace) {
1796 nqindex = BQUEUE_EMPTY;
1797 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1802 * Run scan, possibly freeing data and/or kva mappings on the fly
1805 * WARNING! bufspin is held!
1807 while ((bp = nbp) != NULL) {
1808 int qindex = nqindex;
1810 nbp = TAILQ_NEXT(bp, b_freelist);
1813 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1814 * cycles through the queue twice before being selected.
1816 if (qindex == BQUEUE_CLEAN &&
1817 (bp->b_flags & B_AGE) == 0 && nbp) {
1818 bp->b_flags |= B_AGE;
1819 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1820 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1825 * Calculate next bp ( we can only use it if we do not block
1826 * or do other fancy things ).
1831 nqindex = BQUEUE_EMPTYKVA;
1832 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1835 case BQUEUE_EMPTYKVA:
1836 nqindex = BQUEUE_CLEAN;
1837 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1851 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1854 * Note: we no longer distinguish between VMIO and non-VMIO
1858 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1861 * If we are defragging then we need a buffer with
1862 * b_kvasize != 0. XXX this situation should no longer
1863 * occur, if defrag is non-zero the buffer's b_kvasize
1864 * should also be non-zero at this point. XXX
1866 if (defrag && bp->b_kvasize == 0) {
1867 kprintf("Warning: defrag empty buffer %p\n", bp);
1872 * Start freeing the bp. This is somewhat involved. nbp
1873 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1874 * on the clean list must be disassociated from their
1875 * current vnode. Buffers on the empty[kva] lists have
1876 * already been disassociated.
1879 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1880 spin_unlock_wr(&bufspin);
1881 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1882 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1885 if (bp->b_qindex != qindex) {
1886 spin_unlock_wr(&bufspin);
1887 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1891 bremfree_locked(bp);
1892 spin_unlock_wr(&bufspin);
1895 * Dependancies must be handled before we disassociate the
1898 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1899 * be immediately disassociated. HAMMER then becomes
1900 * responsible for releasing the buffer.
1902 * NOTE: bufspin is UNLOCKED now.
1904 if (LIST_FIRST(&bp->b_dep) != NULL) {
1908 if (bp->b_flags & B_LOCKED) {
1912 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1915 if (qindex == BQUEUE_CLEAN) {
1917 if (bp->b_flags & B_VMIO) {
1918 bp->b_flags &= ~B_ASYNC;
1920 vfs_vmio_release(bp);
1929 * NOTE: nbp is now entirely invalid. We can only restart
1930 * the scan from this point on.
1932 * Get the rest of the buffer freed up. b_kva* is still
1933 * valid after this operation.
1936 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1937 KKASSERT((bp->b_flags & B_HASHED) == 0);
1940 * critical section protection is not required when
1941 * scrapping a buffer's contents because it is already
1944 if (bp->b_bufsize) {
1950 bp->b_flags = B_BNOCLIP;
1951 bp->b_cmd = BUF_CMD_DONE;
1956 bp->b_xio.xio_npages = 0;
1957 bp->b_dirtyoff = bp->b_dirtyend = 0;
1959 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1961 if (blkflags & GETBLK_BHEAVY)
1962 bp->b_flags |= B_HEAVY;
1965 * If we are defragging then free the buffer.
1968 bp->b_flags |= B_INVAL;
1976 * If we are overcomitted then recover the buffer and its
1977 * KVM space. This occurs in rare situations when multiple
1978 * processes are blocked in getnewbuf() or allocbuf().
1980 if (bufspace >= hibufspace)
1982 if (flushingbufs && bp->b_kvasize != 0) {
1983 bp->b_flags |= B_INVAL;
1988 if (bufspace < lobufspace)
1991 /* NOT REACHED, bufspin not held */
1995 * If we exhausted our list, sleep as appropriate. We may have to
1996 * wakeup various daemons and write out some dirty buffers.
1998 * Generally we are sleeping due to insufficient buffer space.
2000 * NOTE: bufspin is held if bp is NULL, else it is not held.
2006 spin_unlock_wr(&bufspin);
2008 flags = VFS_BIO_NEED_BUFSPACE;
2010 } else if (bufspace >= hibufspace) {
2012 flags = VFS_BIO_NEED_BUFSPACE;
2015 flags = VFS_BIO_NEED_ANY;
2018 needsbuffer |= flags;
2019 bd_speedup(); /* heeeelp */
2020 while (needsbuffer & flags) {
2021 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
2026 * We finally have a valid bp. We aren't quite out of the
2027 * woods, we still have to reserve kva space. In order
2028 * to keep fragmentation sane we only allocate kva in
2031 * (bufspin is not held)
2033 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2035 if (maxsize != bp->b_kvasize) {
2036 vm_offset_t addr = 0;
2042 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2043 vm_map_lock(&buffer_map);
2045 if (vm_map_findspace(&buffer_map,
2046 vm_map_min(&buffer_map), maxsize,
2047 maxsize, 0, &addr)) {
2049 * Uh oh. Buffer map is too fragmented. We
2050 * must defragment the map.
2052 vm_map_unlock(&buffer_map);
2053 vm_map_entry_release(count);
2056 bp->b_flags |= B_INVAL;
2062 vm_map_insert(&buffer_map, &count,
2064 addr, addr + maxsize,
2066 VM_PROT_ALL, VM_PROT_ALL,
2069 bp->b_kvabase = (caddr_t) addr;
2070 bp->b_kvasize = maxsize;
2071 bufspace += bp->b_kvasize;
2074 vm_map_unlock(&buffer_map);
2075 vm_map_entry_release(count);
2078 bp->b_data = bp->b_kvabase;
2084 * This routine is called in an emergency to recover VM pages from the
2085 * buffer cache by cashing in clean buffers. The idea is to recover
2086 * enough pages to be able to satisfy a stuck bio_page_alloc().
2089 recoverbufpages(void)
2096 spin_lock_wr(&bufspin);
2097 while (bytes < MAXBSIZE) {
2098 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2103 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2104 * cycles through the queue twice before being selected.
2106 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
2107 bp->b_flags |= B_AGE;
2108 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
2109 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
2117 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
2118 KKASSERT((bp->b_flags & B_DELWRI) == 0);
2121 * Start freeing the bp. This is somewhat involved.
2123 * Buffers on the clean list must be disassociated from
2124 * their current vnode
2127 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2128 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp);
2129 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
2132 if (bp->b_qindex != BQUEUE_CLEAN) {
2133 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp, bp->b_qindex);
2137 bremfree_locked(bp);
2138 spin_unlock_wr(&bufspin);
2141 * Dependancies must be handled before we disassociate the
2144 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2145 * be immediately disassociated. HAMMER then becomes
2146 * responsible for releasing the buffer.
2148 if (LIST_FIRST(&bp->b_dep) != NULL) {
2150 if (bp->b_flags & B_LOCKED) {
2152 spin_lock_wr(&bufspin);
2155 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2158 bytes += bp->b_bufsize;
2161 if (bp->b_flags & B_VMIO) {
2162 bp->b_flags &= ~B_ASYNC;
2163 bp->b_flags |= B_DIRECT; /* try to free pages */
2164 vfs_vmio_release(bp);
2169 KKASSERT(bp->b_vp == NULL);
2170 KKASSERT((bp->b_flags & B_HASHED) == 0);
2173 * critical section protection is not required when
2174 * scrapping a buffer's contents because it is already
2181 bp->b_flags = B_BNOCLIP;
2182 bp->b_cmd = BUF_CMD_DONE;
2187 bp->b_xio.xio_npages = 0;
2188 bp->b_dirtyoff = bp->b_dirtyend = 0;
2190 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2192 bp->b_flags |= B_INVAL;
2195 spin_lock_wr(&bufspin);
2197 spin_unlock_wr(&bufspin);
2204 * Buffer flushing daemon. Buffers are normally flushed by the
2205 * update daemon but if it cannot keep up this process starts to
2206 * take the load in an attempt to prevent getnewbuf() from blocking.
2208 * Once a flush is initiated it does not stop until the number
2209 * of buffers falls below lodirtybuffers, but we will wake up anyone
2210 * waiting at the mid-point.
2213 static struct kproc_desc buf_kp = {
2218 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2219 kproc_start, &buf_kp)
2221 static struct kproc_desc bufhw_kp = {
2226 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2227 kproc_start, &bufhw_kp)
2235 * This process needs to be suspended prior to shutdown sync.
2237 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2238 bufdaemon_td, SHUTDOWN_PRI_LAST);
2239 curthread->td_flags |= TDF_SYSTHREAD;
2242 * This process is allowed to take the buffer cache to the limit
2247 kproc_suspend_loop();
2250 * Do the flush. Limit the amount of in-transit I/O we
2251 * allow to build up, otherwise we would completely saturate
2252 * the I/O system. Wakeup any waiting processes before we
2253 * normally would so they can run in parallel with our drain.
2255 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2256 * but because we split the operation into two threads we
2257 * have to cut it in half for each thread.
2259 limit = lodirtybufspace / 2;
2260 waitrunningbufspace(limit);
2261 while (runningbufspace + dirtybufspace > limit ||
2262 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
2263 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2265 waitrunningbufspace(limit);
2269 * We reached our low water mark, reset the
2270 * request and sleep until we are needed again.
2271 * The sleep is just so the suspend code works.
2273 spin_lock_wr(&needsbuffer_spin);
2274 if (bd_request == 0) {
2275 msleep(&bd_request, &needsbuffer_spin, 0,
2279 spin_unlock_wr(&needsbuffer_spin);
2289 * This process needs to be suspended prior to shutdown sync.
2291 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2292 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2293 curthread->td_flags |= TDF_SYSTHREAD;
2296 * This process is allowed to take the buffer cache to the limit
2301 kproc_suspend_loop();
2304 * Do the flush. Limit the amount of in-transit I/O we
2305 * allow to build up, otherwise we would completely saturate
2306 * the I/O system. Wakeup any waiting processes before we
2307 * normally would so they can run in parallel with our drain.
2309 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2310 * but because we split the operation into two threads we
2311 * have to cut it in half for each thread.
2313 limit = lodirtybufspace / 2;
2314 waitrunningbufspace(limit);
2315 while (runningbufspace + dirtybufspacehw > limit ||
2316 dirtybufcounthw >= nbuf / 2) {
2317 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2319 waitrunningbufspace(limit);
2323 * We reached our low water mark, reset the
2324 * request and sleep until we are needed again.
2325 * The sleep is just so the suspend code works.
2327 spin_lock_wr(&needsbuffer_spin);
2328 if (bd_request_hw == 0) {
2329 msleep(&bd_request_hw, &needsbuffer_spin, 0,
2333 spin_unlock_wr(&needsbuffer_spin);
2340 * Try to flush a buffer in the dirty queue. We must be careful to
2341 * free up B_INVAL buffers instead of write them, which NFS is
2342 * particularly sensitive to.
2344 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2345 * that we really want to try to get the buffer out and reuse it
2346 * due to the write load on the machine.
2349 flushbufqueues(bufq_type_t q)
2355 spin_lock_wr(&bufspin);
2358 bp = TAILQ_FIRST(&bufqueues[q]);
2360 KASSERT((bp->b_flags & B_DELWRI),
2361 ("unexpected clean buffer %p", bp));
2363 if (bp->b_flags & B_DELWRI) {
2364 if (bp->b_flags & B_INVAL) {
2365 spin_unlock_wr(&bufspin);
2367 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2368 panic("flushbufqueues: locked buf");
2374 if (LIST_FIRST(&bp->b_dep) != NULL &&
2375 (bp->b_flags & B_DEFERRED) == 0 &&
2376 buf_countdeps(bp, 0)) {
2377 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2378 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2380 bp->b_flags |= B_DEFERRED;
2381 bp = TAILQ_FIRST(&bufqueues[q]);
2386 * Only write it out if we can successfully lock
2387 * it. If the buffer has a dependancy,
2388 * buf_checkwrite must also return 0 for us to
2389 * be able to initate the write.
2391 * If the buffer is flagged B_ERROR it may be
2392 * requeued over and over again, we try to
2393 * avoid a live lock.
2395 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
2396 spin_unlock_wr(&bufspin);
2398 if (LIST_FIRST(&bp->b_dep) != NULL &&
2399 buf_checkwrite(bp)) {
2402 } else if (bp->b_flags & B_ERROR) {
2403 tsleep(bp, 0, "bioer", 1);
2404 bp->b_flags &= ~B_AGE;
2407 bp->b_flags |= B_AGE;
2414 bp = TAILQ_NEXT(bp, b_freelist);
2417 spin_unlock_wr(&bufspin);
2424 * Returns true if no I/O is needed to access the associated VM object.
2425 * This is like findblk except it also hunts around in the VM system for
2428 * Note that we ignore vm_page_free() races from interrupts against our
2429 * lookup, since if the caller is not protected our return value will not
2430 * be any more valid then otherwise once we exit the critical section.
2433 inmem(struct vnode *vp, off_t loffset)
2436 vm_offset_t toff, tinc, size;
2439 if (findblk(vp, loffset, FINDBLK_TEST))
2441 if (vp->v_mount == NULL)
2443 if ((obj = vp->v_object) == NULL)
2447 if (size > vp->v_mount->mnt_stat.f_iosize)
2448 size = vp->v_mount->mnt_stat.f_iosize;
2450 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2451 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2455 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2456 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2457 if (vm_page_is_valid(m,
2458 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2467 * Sets the dirty range for a buffer based on the status of the dirty
2468 * bits in the pages comprising the buffer.
2470 * The range is limited to the size of the buffer.
2472 * This routine is primarily used by NFS, but is generalized for the
2476 vfs_setdirty(struct buf *bp)
2482 * Degenerate case - empty buffer
2485 if (bp->b_bufsize == 0)
2489 * We qualify the scan for modified pages on whether the
2490 * object has been flushed yet. The OBJ_WRITEABLE flag
2491 * is not cleared simply by protecting pages off.
2494 if ((bp->b_flags & B_VMIO) == 0)
2497 object = bp->b_xio.xio_pages[0]->object;
2499 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2500 kprintf("Warning: object %p writeable but not mightbedirty\n", object);
2501 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2502 kprintf("Warning: object %p mightbedirty but not writeable\n", object);
2504 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2505 vm_offset_t boffset;
2506 vm_offset_t eoffset;
2509 * test the pages to see if they have been modified directly
2510 * by users through the VM system.
2512 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2513 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2514 vm_page_test_dirty(bp->b_xio.xio_pages[i]);
2518 * Calculate the encompassing dirty range, boffset and eoffset,
2519 * (eoffset - boffset) bytes.
2522 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2523 if (bp->b_xio.xio_pages[i]->dirty)
2526 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2528 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2529 if (bp->b_xio.xio_pages[i]->dirty) {
2533 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2536 * Fit it to the buffer.
2539 if (eoffset > bp->b_bcount)
2540 eoffset = bp->b_bcount;
2543 * If we have a good dirty range, merge with the existing
2547 if (boffset < eoffset) {
2548 if (bp->b_dirtyoff > boffset)
2549 bp->b_dirtyoff = boffset;
2550 if (bp->b_dirtyend < eoffset)
2551 bp->b_dirtyend = eoffset;
2559 * Locate and return the specified buffer. Unless flagged otherwise,
2560 * a locked buffer will be returned if it exists or NULL if it does not.
2562 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2563 * for locking the buffer and ensuring that it remains
2564 * the desired buffer after locking.
2566 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2567 * to acquire the lock we return NULL, even if the
2570 * (0) - Lock the buffer blocking.
2575 findblk(struct vnode *vp, off_t loffset, int flags)
2581 lkflags = LK_EXCLUSIVE;
2582 if (flags & FINDBLK_NBLOCK)
2583 lkflags |= LK_NOWAIT;
2586 lwkt_gettoken(&vlock, &vp->v_token);
2587 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2588 lwkt_reltoken(&vlock);
2589 if (bp == NULL || (flags & FINDBLK_TEST))
2591 if (BUF_LOCK(bp, lkflags)) {
2595 if (bp->b_vp == vp && bp->b_loffset == loffset)
2605 * Similar to getblk() except only returns the buffer if it is
2606 * B_CACHE and requires no other manipulation. Otherwise NULL
2609 * If B_RAM is set the buffer might be just fine, but we return
2610 * NULL anyway because we want the code to fall through to the
2611 * cluster read. Otherwise read-ahead breaks.
2614 getcacheblk(struct vnode *vp, off_t loffset)
2618 bp = findblk(vp, loffset, 0);
2620 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
2621 bp->b_flags &= ~B_AGE;
2634 * Get a block given a specified block and offset into a file/device.
2635 * B_INVAL may or may not be set on return. The caller should clear
2636 * B_INVAL prior to initiating a READ.
2638 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2639 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2640 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2641 * without doing any of those things the system will likely believe
2642 * the buffer to be valid (especially if it is not B_VMIO), and the
2643 * next getblk() will return the buffer with B_CACHE set.
2645 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2646 * an existing buffer.
2648 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2649 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2650 * and then cleared based on the backing VM. If the previous buffer is
2651 * non-0-sized but invalid, B_CACHE will be cleared.
2653 * If getblk() must create a new buffer, the new buffer is returned with
2654 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2655 * case it is returned with B_INVAL clear and B_CACHE set based on the
2658 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2659 * B_CACHE bit is clear.
2661 * What this means, basically, is that the caller should use B_CACHE to
2662 * determine whether the buffer is fully valid or not and should clear
2663 * B_INVAL prior to issuing a read. If the caller intends to validate
2664 * the buffer by loading its data area with something, the caller needs
2665 * to clear B_INVAL. If the caller does this without issuing an I/O,
2666 * the caller should set B_CACHE ( as an optimization ), else the caller
2667 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2668 * a write attempt or if it was a successfull read. If the caller
2669 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2670 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2674 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2675 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2680 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2683 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2687 if (size > MAXBSIZE)
2688 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2689 if (vp->v_object == NULL)
2690 panic("getblk: vnode %p has no object!", vp);
2693 if ((bp = findblk(vp, loffset, FINDBLK_TEST)) != NULL) {
2695 * The buffer was found in the cache, but we need to lock it.
2696 * Even with LK_NOWAIT the lockmgr may break our critical
2697 * section, so double-check the validity of the buffer
2698 * once the lock has been obtained.
2700 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2701 if (blkflags & GETBLK_NOWAIT)
2703 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2704 if (blkflags & GETBLK_PCATCH)
2705 lkflags |= LK_PCATCH;
2706 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2708 if (error == ENOLCK)
2712 /* buffer may have changed on us */
2716 * Once the buffer has been locked, make sure we didn't race
2717 * a buffer recyclement. Buffers that are no longer hashed
2718 * will have b_vp == NULL, so this takes care of that check
2721 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2722 kprintf("Warning buffer %p (vp %p loffset %lld) "
2724 bp, vp, (long long)loffset);
2730 * If SZMATCH any pre-existing buffer must be of the requested
2731 * size or NULL is returned. The caller absolutely does not
2732 * want getblk() to bwrite() the buffer on a size mismatch.
2734 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2740 * All vnode-based buffers must be backed by a VM object.
2742 KKASSERT(bp->b_flags & B_VMIO);
2743 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2744 bp->b_flags &= ~B_AGE;
2747 * Make sure that B_INVAL buffers do not have a cached
2748 * block number translation.
2750 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2751 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2752 " did not have cleared bio_offset cache\n",
2753 bp, vp, (long long)loffset);
2754 clearbiocache(&bp->b_bio2);
2758 * The buffer is locked. B_CACHE is cleared if the buffer is
2761 if (bp->b_flags & B_INVAL)
2762 bp->b_flags &= ~B_CACHE;
2766 * Any size inconsistancy with a dirty buffer or a buffer
2767 * with a softupdates dependancy must be resolved. Resizing
2768 * the buffer in such circumstances can lead to problems.
2770 if (size != bp->b_bcount) {
2772 if (bp->b_flags & B_DELWRI) {
2773 bp->b_flags |= B_NOCACHE;
2775 } else if (LIST_FIRST(&bp->b_dep)) {
2776 bp->b_flags |= B_NOCACHE;
2779 bp->b_flags |= B_RELBUF;
2785 KKASSERT(size <= bp->b_kvasize);
2786 KASSERT(bp->b_loffset != NOOFFSET,
2787 ("getblk: no buffer offset"));
2790 * A buffer with B_DELWRI set and B_CACHE clear must
2791 * be committed before we can return the buffer in
2792 * order to prevent the caller from issuing a read
2793 * ( due to B_CACHE not being set ) and overwriting
2796 * Most callers, including NFS and FFS, need this to
2797 * operate properly either because they assume they
2798 * can issue a read if B_CACHE is not set, or because
2799 * ( for example ) an uncached B_DELWRI might loop due
2800 * to softupdates re-dirtying the buffer. In the latter
2801 * case, B_CACHE is set after the first write completes,
2802 * preventing further loops.
2804 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2805 * above while extending the buffer, we cannot allow the
2806 * buffer to remain with B_CACHE set after the write
2807 * completes or it will represent a corrupt state. To
2808 * deal with this we set B_NOCACHE to scrap the buffer
2811 * We might be able to do something fancy, like setting
2812 * B_CACHE in bwrite() except if B_DELWRI is already set,
2813 * so the below call doesn't set B_CACHE, but that gets real
2814 * confusing. This is much easier.
2817 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2819 bp->b_flags |= B_NOCACHE;
2826 * Buffer is not in-core, create new buffer. The buffer
2827 * returned by getnewbuf() is locked. Note that the returned
2828 * buffer is also considered valid (not marked B_INVAL).
2830 * Calculating the offset for the I/O requires figuring out
2831 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2832 * the mount's f_iosize otherwise. If the vnode does not
2833 * have an associated mount we assume that the passed size is
2836 * Note that vn_isdisk() cannot be used here since it may
2837 * return a failure for numerous reasons. Note that the
2838 * buffer size may be larger then the block size (the caller
2839 * will use block numbers with the proper multiple). Beware
2840 * of using any v_* fields which are part of unions. In
2841 * particular, in DragonFly the mount point overloading
2842 * mechanism uses the namecache only and the underlying
2843 * directory vnode is not a special case.
2847 if (vp->v_type == VBLK || vp->v_type == VCHR)
2849 else if (vp->v_mount)
2850 bsize = vp->v_mount->mnt_stat.f_iosize;
2854 maxsize = size + (loffset & PAGE_MASK);
2855 maxsize = imax(maxsize, bsize);
2857 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
2859 if (slpflags || slptimeo)
2865 * Atomically insert the buffer into the hash, so that it can
2866 * be found by findblk().
2868 * If bgetvp() returns non-zero a collision occured, and the
2869 * bp will not be associated with the vnode.
2871 * Make sure the translation layer has been cleared.
2873 bp->b_loffset = loffset;
2874 bp->b_bio2.bio_offset = NOOFFSET;
2875 /* bp->b_bio2.bio_next = NULL; */
2877 if (bgetvp(vp, bp)) {
2878 bp->b_flags |= B_INVAL;
2884 * All vnode-based buffers must be backed by a VM object.
2886 KKASSERT(vp->v_object != NULL);
2887 bp->b_flags |= B_VMIO;
2888 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2900 * Reacquire a buffer that was previously released to the locked queue,
2901 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2902 * set B_LOCKED (which handles the acquisition race).
2904 * To this end, either B_LOCKED must be set or the dependancy list must be
2910 regetblk(struct buf *bp)
2912 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
2913 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2920 * Get an empty, disassociated buffer of given size. The buffer is
2921 * initially set to B_INVAL.
2923 * critical section protection is not required for the allocbuf()
2924 * call because races are impossible here.
2934 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2936 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2941 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2949 * This code constitutes the buffer memory from either anonymous system
2950 * memory (in the case of non-VMIO operations) or from an associated
2951 * VM object (in the case of VMIO operations). This code is able to
2952 * resize a buffer up or down.
2954 * Note that this code is tricky, and has many complications to resolve
2955 * deadlock or inconsistant data situations. Tread lightly!!!
2956 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2957 * the caller. Calling this code willy nilly can result in the loss of data.
2959 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2960 * B_CACHE for the non-VMIO case.
2962 * This routine does not need to be called from a critical section but you
2963 * must own the buffer.
2968 allocbuf(struct buf *bp, int size)
2970 int newbsize, mbsize;
2973 if (BUF_REFCNT(bp) == 0)
2974 panic("allocbuf: buffer not busy");
2976 if (bp->b_kvasize < size)
2977 panic("allocbuf: buffer too small");
2979 if ((bp->b_flags & B_VMIO) == 0) {
2983 * Just get anonymous memory from the kernel. Don't
2984 * mess with B_CACHE.
2986 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2987 if (bp->b_flags & B_MALLOC)
2990 newbsize = round_page(size);
2992 if (newbsize < bp->b_bufsize) {
2994 * Malloced buffers are not shrunk
2996 if (bp->b_flags & B_MALLOC) {
2998 bp->b_bcount = size;
3000 kfree(bp->b_data, M_BIOBUF);
3001 if (bp->b_bufsize) {
3002 bufmallocspace -= bp->b_bufsize;
3006 bp->b_data = bp->b_kvabase;
3008 bp->b_flags &= ~B_MALLOC;
3014 (vm_offset_t) bp->b_data + newbsize,
3015 (vm_offset_t) bp->b_data + bp->b_bufsize);
3016 } else if (newbsize > bp->b_bufsize) {
3018 * We only use malloced memory on the first allocation.
3019 * and revert to page-allocated memory when the buffer
3022 if ((bufmallocspace < maxbufmallocspace) &&
3023 (bp->b_bufsize == 0) &&
3024 (mbsize <= PAGE_SIZE/2)) {
3026 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3027 bp->b_bufsize = mbsize;
3028 bp->b_bcount = size;
3029 bp->b_flags |= B_MALLOC;
3030 bufmallocspace += mbsize;
3036 * If the buffer is growing on its other-than-first
3037 * allocation, then we revert to the page-allocation
3040 if (bp->b_flags & B_MALLOC) {
3041 origbuf = bp->b_data;
3042 origbufsize = bp->b_bufsize;
3043 bp->b_data = bp->b_kvabase;
3044 if (bp->b_bufsize) {
3045 bufmallocspace -= bp->b_bufsize;
3049 bp->b_flags &= ~B_MALLOC;
3050 newbsize = round_page(newbsize);
3054 (vm_offset_t) bp->b_data + bp->b_bufsize,
3055 (vm_offset_t) bp->b_data + newbsize);
3057 bcopy(origbuf, bp->b_data, origbufsize);
3058 kfree(origbuf, M_BIOBUF);
3065 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3066 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3067 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3068 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3070 if (bp->b_flags & B_MALLOC)
3071 panic("allocbuf: VMIO buffer can't be malloced");
3073 * Set B_CACHE initially if buffer is 0 length or will become
3076 if (size == 0 || bp->b_bufsize == 0)
3077 bp->b_flags |= B_CACHE;
3079 if (newbsize < bp->b_bufsize) {
3081 * DEV_BSIZE aligned new buffer size is less then the
3082 * DEV_BSIZE aligned existing buffer size. Figure out
3083 * if we have to remove any pages.
3085 if (desiredpages < bp->b_xio.xio_npages) {
3086 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3088 * the page is not freed here -- it
3089 * is the responsibility of
3090 * vnode_pager_setsize
3092 m = bp->b_xio.xio_pages[i];
3093 KASSERT(m != bogus_page,
3094 ("allocbuf: bogus page found"));
3095 while (vm_page_sleep_busy(m, TRUE, "biodep"))
3098 bp->b_xio.xio_pages[i] = NULL;
3099 vm_page_unwire(m, 0);
3101 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3102 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3103 bp->b_xio.xio_npages = desiredpages;
3105 } else if (size > bp->b_bcount) {
3107 * We are growing the buffer, possibly in a
3108 * byte-granular fashion.
3116 * Step 1, bring in the VM pages from the object,
3117 * allocating them if necessary. We must clear
3118 * B_CACHE if these pages are not valid for the
3119 * range covered by the buffer.
3121 * critical section protection is required to protect
3122 * against interrupts unbusying and freeing pages
3123 * between our vm_page_lookup() and our
3124 * busycheck/wiring call.
3130 while (bp->b_xio.xio_npages < desiredpages) {
3134 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
3135 if ((m = vm_page_lookup(obj, pi)) == NULL) {
3137 * note: must allocate system pages
3138 * since blocking here could intefere
3139 * with paging I/O, no matter which
3142 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
3146 bp->b_flags &= ~B_CACHE;
3147 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3148 ++bp->b_xio.xio_npages;
3154 * We found a page. If we have to sleep on it,
3155 * retry because it might have gotten freed out
3158 * We can only test PG_BUSY here. Blocking on
3159 * m->busy might lead to a deadlock:
3161 * vm_fault->getpages->cluster_read->allocbuf
3165 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
3167 vm_page_flag_clear(m, PG_ZERO);
3169 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3170 ++bp->b_xio.xio_npages;
3175 * Step 2. We've loaded the pages into the buffer,
3176 * we have to figure out if we can still have B_CACHE
3177 * set. Note that B_CACHE is set according to the
3178 * byte-granular range ( bcount and size ), not the
3179 * aligned range ( newbsize ).
3181 * The VM test is against m->valid, which is DEV_BSIZE
3182 * aligned. Needless to say, the validity of the data
3183 * needs to also be DEV_BSIZE aligned. Note that this
3184 * fails with NFS if the server or some other client
3185 * extends the file's EOF. If our buffer is resized,
3186 * B_CACHE may remain set! XXX
3189 toff = bp->b_bcount;
3190 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3192 while ((bp->b_flags & B_CACHE) && toff < size) {
3195 if (tinc > (size - toff))
3198 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3206 bp->b_xio.xio_pages[pi]
3213 * Step 3, fixup the KVM pmap. Remember that
3214 * bp->b_data is relative to bp->b_loffset, but
3215 * bp->b_loffset may be offset into the first page.
3218 bp->b_data = (caddr_t)
3219 trunc_page((vm_offset_t)bp->b_data);
3221 (vm_offset_t)bp->b_data,
3222 bp->b_xio.xio_pages,
3223 bp->b_xio.xio_npages
3225 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3226 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3230 /* adjust space use on already-dirty buffer */
3231 if (bp->b_flags & B_DELWRI) {
3232 dirtybufspace += newbsize - bp->b_bufsize;
3233 if (bp->b_flags & B_HEAVY)
3234 dirtybufspacehw += newbsize - bp->b_bufsize;
3236 if (newbsize < bp->b_bufsize)
3238 bp->b_bufsize = newbsize; /* actual buffer allocation */
3239 bp->b_bcount = size; /* requested buffer size */
3246 * Wait for buffer I/O completion, returning error status. The buffer
3247 * is left locked on return. B_EINTR is converted into an EINTR error
3250 * NOTE! The original b_cmd is lost on return, since b_cmd will be
3251 * set to BUF_CMD_DONE.
3256 biowait(struct buf *bp)
3258 if (bp->b_cmd != BUF_CMD_DONE) {
3261 tsleep_interlock(bp);
3262 if (bp->b_cmd == BUF_CMD_DONE)
3264 if (bp->b_cmd == BUF_CMD_READ)
3265 tsleep(bp, 0, "biord", 0);
3267 tsleep(bp, 0, "biowr", 0);
3271 if (bp->b_flags & B_EINTR) {
3272 bp->b_flags &= ~B_EINTR;
3275 if (bp->b_flags & B_ERROR) {
3276 return (bp->b_error ? bp->b_error : EIO);
3283 * This associates a tracking count with an I/O. vn_strategy() and
3284 * dev_dstrategy() do this automatically but there are a few cases
3285 * where a vnode or device layer is bypassed when a block translation
3286 * is cached. In such cases bio_start_transaction() may be called on
3287 * the bypassed layers so the system gets an I/O in progress indication
3288 * for those higher layers.
3291 bio_start_transaction(struct bio *bio, struct bio_track *track)
3293 bio->bio_track = track;
3294 bio_track_ref(track);
3298 * Initiate I/O on a vnode.
3301 vn_strategy(struct vnode *vp, struct bio *bio)
3303 struct bio_track *track;
3305 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
3306 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
3307 track = &vp->v_track_read;
3309 track = &vp->v_track_write;
3310 bio->bio_track = track;
3311 bio_track_ref(track);
3312 vop_strategy(*vp->v_ops, vp, bio);
3318 * Finish I/O on a buffer, optionally calling a completion function.
3319 * This is usually called from an interrupt so process blocking is
3322 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3323 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3324 * assuming B_INVAL is clear.
3326 * For the VMIO case, we set B_CACHE if the op was a read and no
3327 * read error occured, or if the op was a write. B_CACHE is never
3328 * set if the buffer is invalid or otherwise uncacheable.
3330 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3331 * initiator to leave B_INVAL set to brelse the buffer out of existance
3332 * in the biodone routine.
3335 biodone(struct bio *bio)
3337 struct buf *bp = bio->bio_buf;
3342 KASSERT(BUF_REFCNTNB(bp) > 0,
3343 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3344 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3345 ("biodone: bp %p already done!", bp));
3347 runningbufwakeup(bp);
3350 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3353 biodone_t *done_func;
3354 struct bio_track *track;
3357 * BIO tracking. Most but not all BIOs are tracked.
3359 if ((track = bio->bio_track) != NULL) {
3360 bio_track_rel(track);
3361 bio->bio_track = NULL;
3365 * A bio_done function terminates the loop. The function
3366 * will be responsible for any further chaining and/or
3367 * buffer management.
3369 * WARNING! The done function can deallocate the buffer!
3371 if ((done_func = bio->bio_done) != NULL) {
3372 bio->bio_done = NULL;
3377 bio = bio->bio_prev;
3381 bp->b_cmd = BUF_CMD_DONE;
3384 * Only reads and writes are processed past this point.
3386 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3387 if (cmd == BUF_CMD_FREEBLKS)
3388 bp->b_flags |= B_NOCACHE;
3395 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3396 * a lot worse. XXX - move this above the clearing of b_cmd
3398 if (LIST_FIRST(&bp->b_dep) != NULL)
3402 * A failed write must re-dirty the buffer unless B_INVAL
3405 if (cmd == BUF_CMD_WRITE &&
3406 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3407 bp->b_flags &= ~B_NOCACHE;
3412 if (bp->b_flags & B_VMIO) {
3418 struct vnode *vp = bp->b_vp;
3422 #if defined(VFS_BIO_DEBUG)
3423 if (vp->v_auxrefs == 0)
3424 panic("biodone: zero vnode hold count");
3425 if ((vp->v_flag & VOBJBUF) == 0)
3426 panic("biodone: vnode is not setup for merged cache");
3429 foff = bp->b_loffset;
3430 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3431 KASSERT(obj != NULL, ("biodone: missing VM object"));
3433 #if defined(VFS_BIO_DEBUG)
3434 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3435 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3436 obj->paging_in_progress, bp->b_xio.xio_npages);
3441 * Set B_CACHE if the op was a normal read and no error
3442 * occured. B_CACHE is set for writes in the b*write()
3445 iosize = bp->b_bcount - bp->b_resid;
3446 if (cmd == BUF_CMD_READ && (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3447 bp->b_flags |= B_CACHE;
3450 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3454 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3459 * cleanup bogus pages, restoring the originals. Since
3460 * the originals should still be wired, we don't have
3461 * to worry about interrupt/freeing races destroying
3462 * the VM object association.
3464 m = bp->b_xio.xio_pages[i];
3465 if (m == bogus_page) {
3467 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3469 panic("biodone: page disappeared");
3470 bp->b_xio.xio_pages[i] = m;
3471 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3472 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3474 #if defined(VFS_BIO_DEBUG)
3475 if (OFF_TO_IDX(foff) != m->pindex) {
3477 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
3478 (unsigned long)foff, m->pindex);
3483 * In the write case, the valid and clean bits are
3484 * already changed correctly ( see bdwrite() ), so we
3485 * only need to do this here in the read case.
3487 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3488 vfs_page_set_valid(bp, foff, i, m);
3490 vm_page_flag_clear(m, PG_ZERO);
3493 * when debugging new filesystems or buffer I/O methods, this
3494 * is the most common error that pops up. if you see this, you
3495 * have not set the page busy flag correctly!!!
3498 kprintf("biodone: page busy < 0, "
3499 "pindex: %d, foff: 0x(%x,%x), "
3500 "resid: %d, index: %d\n",
3501 (int) m->pindex, (int)(foff >> 32),
3502 (int) foff & 0xffffffff, resid, i);
3503 if (!vn_isdisk(vp, NULL))
3504 kprintf(" iosize: %ld, loffset: %lld, "
3505 "flags: 0x%08x, npages: %d\n",
3506 bp->b_vp->v_mount->mnt_stat.f_iosize,
3507 (long long)bp->b_loffset,
3508 bp->b_flags, bp->b_xio.xio_npages);
3510 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3511 (long long)bp->b_loffset,
3512 bp->b_flags, bp->b_xio.xio_npages);
3513 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3514 m->valid, m->dirty, m->wire_count);
3515 panic("biodone: page busy < 0");
3517 vm_page_io_finish(m);
3518 vm_object_pip_subtract(obj, 1);
3519 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3523 vm_object_pip_wakeupn(obj, 0);
3527 * For asynchronous completions, release the buffer now. The brelse
3528 * will do a wakeup there if necessary - so no need to do a wakeup
3529 * here in the async case. The sync case always needs to do a wakeup.
3532 if (bp->b_flags & B_ASYNC) {
3533 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
3546 * This routine is called in lieu of iodone in the case of
3547 * incomplete I/O. This keeps the busy status for pages
3551 vfs_unbusy_pages(struct buf *bp)
3555 runningbufwakeup(bp);
3556 if (bp->b_flags & B_VMIO) {
3557 struct vnode *vp = bp->b_vp;
3562 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3563 vm_page_t m = bp->b_xio.xio_pages[i];
3566 * When restoring bogus changes the original pages
3567 * should still be wired, so we are in no danger of
3568 * losing the object association and do not need
3569 * critical section protection particularly.
3571 if (m == bogus_page) {
3572 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3574 panic("vfs_unbusy_pages: page missing");
3576 bp->b_xio.xio_pages[i] = m;
3577 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3578 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3580 vm_object_pip_subtract(obj, 1);
3581 vm_page_flag_clear(m, PG_ZERO);
3582 vm_page_io_finish(m);
3584 vm_object_pip_wakeupn(obj, 0);
3589 * vfs_page_set_valid:
3591 * Set the valid bits in a page based on the supplied offset. The
3592 * range is restricted to the buffer's size.
3594 * This routine is typically called after a read completes.
3597 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3599 vm_ooffset_t soff, eoff;
3602 * Start and end offsets in buffer. eoff - soff may not cross a
3603 * page boundry or cross the end of the buffer. The end of the
3604 * buffer, in this case, is our file EOF, not the allocation size
3608 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3609 if (eoff > bp->b_loffset + bp->b_bcount)
3610 eoff = bp->b_loffset + bp->b_bcount;
3613 * Set valid range. This is typically the entire buffer and thus the
3617 vm_page_set_validclean(
3619 (vm_offset_t) (soff & PAGE_MASK),
3620 (vm_offset_t) (eoff - soff)
3628 * This routine is called before a device strategy routine.
3629 * It is used to tell the VM system that paging I/O is in
3630 * progress, and treat the pages associated with the buffer
3631 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3632 * flag is handled to make sure that the object doesn't become
3635 * Since I/O has not been initiated yet, certain buffer flags
3636 * such as B_ERROR or B_INVAL may be in an inconsistant state
3637 * and should be ignored.
3640 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3643 struct lwp *lp = curthread->td_lwp;
3646 * The buffer's I/O command must already be set. If reading,
3647 * B_CACHE must be 0 (double check against callers only doing
3648 * I/O when B_CACHE is 0).
3650 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3651 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3653 if (bp->b_flags & B_VMIO) {
3658 foff = bp->b_loffset;
3659 KASSERT(bp->b_loffset != NOOFFSET,
3660 ("vfs_busy_pages: no buffer offset"));
3664 * Loop until none of the pages are busy.
3667 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3668 vm_page_t m = bp->b_xio.xio_pages[i];
3670 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3675 * Setup for I/O, soft-busy the page right now because
3676 * the next loop may block.
3678 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3679 vm_page_t m = bp->b_xio.xio_pages[i];
3681 vm_page_flag_clear(m, PG_ZERO);
3682 if ((bp->b_flags & B_CLUSTER) == 0) {
3683 vm_object_pip_add(obj, 1);
3684 vm_page_io_start(m);
3689 * Adjust protections for I/O and do bogus-page mapping.
3690 * Assume that vm_page_protect() can block (it can block
3691 * if VM_PROT_NONE, don't take any chances regardless).
3694 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3695 vm_page_t m = bp->b_xio.xio_pages[i];
3698 * When readying a vnode-backed buffer for a write
3699 * we must zero-fill any invalid portions of the
3702 * When readying a vnode-backed buffer for a read
3703 * we must replace any dirty pages with a bogus
3704 * page so we do not destroy dirty data when
3705 * filling in gaps. Dirty pages might not
3706 * necessarily be marked dirty yet, so use m->valid
3707 * as a reasonable test.
3709 * Bogus page replacement is, uh, bogus. We need
3710 * to find a better way.
3712 if (bp->b_cmd == BUF_CMD_WRITE) {
3713 vm_page_protect(m, VM_PROT_READ);
3714 vfs_page_set_valid(bp, foff, i, m);
3715 } else if (m->valid == VM_PAGE_BITS_ALL) {
3716 bp->b_xio.xio_pages[i] = bogus_page;
3719 vm_page_protect(m, VM_PROT_NONE);
3721 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3724 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3725 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3729 * This is the easiest place to put the process accounting for the I/O
3733 if (bp->b_cmd == BUF_CMD_READ)
3734 lp->lwp_ru.ru_inblock++;
3736 lp->lwp_ru.ru_oublock++;
3743 * Tell the VM system that the pages associated with this buffer
3744 * are clean. This is used for delayed writes where the data is
3745 * going to go to disk eventually without additional VM intevention.
3747 * Note that while we only really need to clean through to b_bcount, we
3748 * just go ahead and clean through to b_bufsize.
3751 vfs_clean_pages(struct buf *bp)
3755 if (bp->b_flags & B_VMIO) {
3758 foff = bp->b_loffset;
3759 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset"));
3760 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3761 vm_page_t m = bp->b_xio.xio_pages[i];
3762 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3764 vfs_page_set_valid(bp, foff, i, m);
3771 * vfs_bio_set_validclean:
3773 * Set the range within the buffer to valid and clean. The range is
3774 * relative to the beginning of the buffer, b_loffset. Note that
3775 * b_loffset itself may be offset from the beginning of the first page.
3779 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3781 if (bp->b_flags & B_VMIO) {
3786 * Fixup base to be relative to beginning of first page.
3787 * Set initial n to be the maximum number of bytes in the
3788 * first page that can be validated.
3791 base += (bp->b_loffset & PAGE_MASK);
3792 n = PAGE_SIZE - (base & PAGE_MASK);
3794 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) {
3795 vm_page_t m = bp->b_xio.xio_pages[i];
3800 vm_page_set_validclean(m, base & PAGE_MASK, n);
3811 * Clear a buffer. This routine essentially fakes an I/O, so we need
3812 * to clear B_ERROR and B_INVAL.
3814 * Note that while we only theoretically need to clear through b_bcount,
3815 * we go ahead and clear through b_bufsize.
3819 vfs_bio_clrbuf(struct buf *bp)
3823 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3824 bp->b_flags &= ~(B_INVAL|B_ERROR);
3825 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3826 (bp->b_loffset & PAGE_MASK) == 0) {
3827 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3828 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3832 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3833 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3834 bzero(bp->b_data, bp->b_bufsize);
3835 bp->b_xio.xio_pages[0]->valid |= mask;
3841 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3842 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3843 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3844 ea = (caddr_t)(vm_offset_t)ulmin(
3845 (u_long)(vm_offset_t)ea,
3846 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3847 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3848 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3850 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3851 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3855 for (; sa < ea; sa += DEV_BSIZE, j++) {
3856 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3857 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3858 bzero(sa, DEV_BSIZE);
3861 bp->b_xio.xio_pages[i]->valid |= mask;
3862 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
3871 * vm_hold_load_pages:
3873 * Load pages into the buffer's address space. The pages are
3874 * allocated from the kernel object in order to reduce interference
3875 * with the any VM paging I/O activity. The range of loaded
3876 * pages will be wired.
3878 * If a page cannot be allocated, the 'pagedaemon' is woken up to
3879 * retrieve the full range (to - from) of pages.
3883 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3889 to = round_page(to);
3890 from = round_page(from);
3891 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3896 * Note: must allocate system pages since blocking here
3897 * could intefere with paging I/O, no matter which
3900 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT,
3901 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
3904 p->valid = VM_PAGE_BITS_ALL;
3905 vm_page_flag_clear(p, PG_ZERO);
3906 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3907 bp->b_xio.xio_pages[index] = p;
3914 bp->b_xio.xio_npages = index;
3918 * Allocate pages for a buffer cache buffer.
3920 * Under extremely severe memory conditions even allocating out of the
3921 * system reserve can fail. If this occurs we must allocate out of the
3922 * interrupt reserve to avoid a deadlock with the pageout daemon.
3924 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
3925 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
3926 * against the pageout daemon if pages are not freed from other sources.
3930 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit)
3935 * Try a normal allocation, allow use of system reserve.
3937 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
3942 * The normal allocation failed and we clearly have a page
3943 * deficit. Try to reclaim some clean VM pages directly
3944 * from the buffer cache.
3946 vm_pageout_deficit += deficit;
3950 * We may have blocked, the caller will know what to do if the
3953 if (vm_page_lookup(obj, pg))
3957 * Allocate and allow use of the interrupt reserve.
3959 * If after all that we still can't allocate a VM page we are
3960 * in real trouble, but we slog on anyway hoping that the system
3963 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
3964 VM_ALLOC_INTERRUPT);
3966 if (vm_page_count_severe()) {
3967 kprintf("bio_page_alloc: WARNING emergency page "
3972 kprintf("bio_page_alloc: WARNING emergency page "
3973 "allocation failed\n");
3980 * vm_hold_free_pages:
3982 * Return pages associated with the buffer back to the VM system.
3984 * The range of pages underlying the buffer's address space will
3985 * be unmapped and un-wired.
3988 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3992 int index, newnpages;
3994 from = round_page(from);
3995 to = round_page(to);
3996 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3998 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3999 p = bp->b_xio.xio_pages[index];
4000 if (p && (index < bp->b_xio.xio_npages)) {
4002 kprintf("vm_hold_free_pages: doffset: %lld, "
4004 (long long)bp->b_bio2.bio_offset,
4005 (long long)bp->b_loffset);
4007 bp->b_xio.xio_pages[index] = NULL;
4010 vm_page_unwire(p, 0);
4014 bp->b_xio.xio_npages = newnpages;
4020 * Map a user buffer into KVM via a pbuf. On return the buffer's
4021 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4025 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4036 * bp had better have a command and it better be a pbuf.
4038 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4039 KKASSERT(bp->b_flags & B_PAGING);
4045 * Map the user data into KVM. Mappings have to be page-aligned.
4047 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4050 vmprot = VM_PROT_READ;
4051 if (bp->b_cmd == BUF_CMD_READ)
4052 vmprot |= VM_PROT_WRITE;
4054 while (addr < udata + bytes) {
4056 * Do the vm_fault if needed; do the copy-on-write thing
4057 * when reading stuff off device into memory.
4059 * vm_fault_page*() returns a held VM page.
4061 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4062 va = trunc_page(va);
4064 m = vm_fault_page_quick(va, vmprot, &error);
4066 for (i = 0; i < pidx; ++i) {
4067 vm_page_unhold(bp->b_xio.xio_pages[i]);
4068 bp->b_xio.xio_pages[i] = NULL;
4072 bp->b_xio.xio_pages[pidx] = m;
4078 * Map the page array and set the buffer fields to point to
4079 * the mapped data buffer.
4081 if (pidx > btoc(MAXPHYS))
4082 panic("vmapbuf: mapped more than MAXPHYS");
4083 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4085 bp->b_xio.xio_npages = pidx;
4086 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4087 bp->b_bcount = bytes;
4088 bp->b_bufsize = bytes;
4095 * Free the io map PTEs associated with this IO operation.
4096 * We also invalidate the TLB entries and restore the original b_addr.
4099 vunmapbuf(struct buf *bp)
4104 KKASSERT(bp->b_flags & B_PAGING);
4106 npages = bp->b_xio.xio_npages;
4107 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4108 for (pidx = 0; pidx < npages; ++pidx) {
4109 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4110 bp->b_xio.xio_pages[pidx] = NULL;
4112 bp->b_xio.xio_npages = 0;
4113 bp->b_data = bp->b_kvabase;
4117 * Scan all buffers in the system and issue the callback.
4120 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4126 for (n = 0; n < nbuf; ++n) {
4127 if ((error = callback(&buf[n], info)) < 0) {
4137 * print out statistics from the current status of the buffer pool
4138 * this can be toggeled by the system control option debug.syncprt
4147 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
4148 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
4150 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
4152 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4155 TAILQ_FOREACH(bp, dp, b_freelist) {
4156 counts[bp->b_bufsize/PAGE_SIZE]++;
4160 kprintf("%s: total-%d", bname[i], count);
4161 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4163 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
4171 DB_SHOW_COMMAND(buffer, db_show_buffer)
4174 struct buf *bp = (struct buf *)addr;
4177 db_printf("usage: show buffer <addr>\n");
4181 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4182 db_printf("b_cmd = %d\n", bp->b_cmd);
4183 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4184 "b_resid = %d\n, b_data = %p, "
4185 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4186 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4188 (long long)bp->b_bio2.bio_offset,
4189 (long long)(bp->b_bio2.bio_next ?
4190 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4191 if (bp->b_xio.xio_npages) {
4193 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4194 bp->b_xio.xio_npages);
4195 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4197 m = bp->b_xio.xio_pages[i];
4198 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4199 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4200 if ((i + 1) < bp->b_xio.xio_npages)