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];
89 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
91 struct buf *buf; /* buffer header pool */
93 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
94 int pageno, vm_page_t m);
95 static void vfs_clean_pages(struct buf *bp);
96 static void vfs_setdirty(struct buf *bp);
97 static void vfs_vmio_release(struct buf *bp);
98 static int flushbufqueues(bufq_type_t q);
99 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);
101 static void bd_signal(int totalspace);
102 static void buf_daemon(void);
103 static void buf_daemon_hw(void);
106 * bogus page -- for I/O to/from partially complete buffers
107 * this is a temporary solution to the problem, but it is not
108 * really that bad. it would be better to split the buffer
109 * for input in the case of buffers partially already in memory,
110 * but the code is intricate enough already.
112 vm_page_t bogus_page;
115 * These are all static, but make the ones we export globals so we do
116 * not need to use compiler magic.
118 int bufspace, maxbufspace,
119 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
120 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
121 static int lorunningspace, hirunningspace, runningbufreq;
122 int dirtybufspace, dirtybufspacehw, lodirtybufspace, hidirtybufspace;
123 int dirtybufcount, dirtybufcounthw;
124 int runningbufspace, runningbufcount;
125 static int getnewbufcalls;
126 static int getnewbufrestarts;
127 static int recoverbufcalls;
128 static int needsbuffer; /* locked by needsbuffer_spin */
129 static int bd_request; /* locked by needsbuffer_spin */
130 static int bd_request_hw; /* locked by needsbuffer_spin */
131 static u_int bd_wake_ary[BD_WAKE_SIZE];
132 static u_int bd_wake_index;
133 static struct spinlock needsbuffer_spin;
135 static struct thread *bufdaemon_td;
136 static struct thread *bufdaemonhw_td;
140 * Sysctls for operational control of the buffer cache.
142 SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
143 "Number of dirty buffers to flush before bufdaemon becomes inactive");
144 SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
145 "High watermark used to trigger explicit flushing of dirty buffers");
146 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
147 "Minimum amount of buffer space required for active I/O");
148 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
149 "Maximum amount of buffer space to usable for active I/O");
151 * Sysctls determining current state of the buffer cache.
153 SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
154 "Total number of buffers in buffer cache");
155 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
156 "Pending bytes of dirty buffers (all)");
157 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
158 "Pending bytes of dirty buffers (heavy weight)");
159 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
160 "Pending number of dirty buffers");
161 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
162 "Pending number of dirty buffers (heavy weight)");
163 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
164 "I/O bytes currently in progress due to asynchronous writes");
165 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
166 "I/O buffers currently in progress due to asynchronous writes");
167 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
168 "Hard limit on maximum amount of memory usable for buffer space");
169 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
170 "Soft limit on maximum amount of memory usable for buffer space");
171 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
172 "Minimum amount of memory to reserve for system buffer space");
173 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
174 "Amount of memory available for buffers");
175 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
176 0, "Maximum amount of memory reserved for buffers using malloc");
177 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
178 "Amount of memory left for buffers using malloc-scheme");
179 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
180 "New buffer header acquisition requests");
181 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
182 0, "New buffer header acquisition restarts");
183 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
184 "Recover VM space in an emergency");
185 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
186 "Buffer acquisition restarts due to fragmented buffer map");
187 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
188 "Amount of time KVA space was deallocated in an arbitrary buffer");
189 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
190 "Amount of time buffer re-use operations were successful");
191 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
192 "sizeof(struct buf)");
194 char *buf_wmesg = BUF_WMESG;
196 extern int vm_swap_size;
198 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
199 #define VFS_BIO_NEED_UNUSED02 0x02
200 #define VFS_BIO_NEED_UNUSED04 0x04
201 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
206 * Called when buffer space is potentially available for recovery.
207 * getnewbuf() will block on this flag when it is unable to free
208 * sufficient buffer space. Buffer space becomes recoverable when
209 * bp's get placed back in the queues.
216 * If someone is waiting for BUF space, wake them up. Even
217 * though we haven't freed the kva space yet, the waiting
218 * process will be able to now.
220 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
221 spin_lock_wr(&needsbuffer_spin);
222 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
223 spin_unlock_wr(&needsbuffer_spin);
224 wakeup(&needsbuffer);
231 * Accounting for I/O in progress.
235 runningbufwakeup(struct buf *bp)
239 if ((totalspace = bp->b_runningbufspace) != 0) {
240 runningbufspace -= totalspace;
242 bp->b_runningbufspace = 0;
243 if (runningbufreq && runningbufspace <= lorunningspace) {
245 wakeup(&runningbufreq);
247 bd_signal(totalspace);
254 * Called when a buffer has been added to one of the free queues to
255 * account for the buffer and to wakeup anyone waiting for free buffers.
256 * This typically occurs when large amounts of metadata are being handled
257 * by the buffer cache ( else buffer space runs out first, usually ).
264 spin_lock_wr(&needsbuffer_spin);
265 needsbuffer &= ~VFS_BIO_NEED_ANY;
266 spin_unlock_wr(&needsbuffer_spin);
267 wakeup(&needsbuffer);
272 * waitrunningbufspace()
274 * Wait for the amount of running I/O to drop to a reasonable level.
276 * The caller may be using this function to block in a tight loop, we
277 * must block of runningbufspace is greater then the passed limit.
278 * And even with that it may not be enough, due to the presence of
279 * B_LOCKED dirty buffers, so also wait for at least one running buffer
283 waitrunningbufspace(int limit)
287 if (lorunningspace < limit)
288 lorun = lorunningspace;
293 if (runningbufspace > lorun) {
294 while (runningbufspace > lorun) {
296 tsleep(&runningbufreq, 0, "wdrain", 0);
298 } else if (runningbufspace) {
300 tsleep(&runningbufreq, 0, "wdrain2", 1);
306 * vfs_buf_test_cache:
308 * Called when a buffer is extended. This function clears the B_CACHE
309 * bit if the newly extended portion of the buffer does not contain
314 vfs_buf_test_cache(struct buf *bp,
315 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
318 if (bp->b_flags & B_CACHE) {
319 int base = (foff + off) & PAGE_MASK;
320 if (vm_page_is_valid(m, base, size) == 0)
321 bp->b_flags &= ~B_CACHE;
328 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
335 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
338 if (bd_request == 0 &&
339 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
340 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
341 spin_lock_wr(&needsbuffer_spin);
343 spin_unlock_wr(&needsbuffer_spin);
346 if (bd_request_hw == 0 &&
347 (dirtybufspacehw > lodirtybufspace / 2 ||
348 dirtybufcounthw >= nbuf / 2)) {
349 spin_lock_wr(&needsbuffer_spin);
351 spin_unlock_wr(&needsbuffer_spin);
352 wakeup(&bd_request_hw);
359 * Get the buf_daemon heated up when the number of running and dirty
360 * buffers exceeds the mid-point.
369 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
371 totalspace = runningbufspace + dirtybufspace;
372 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
374 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
375 if (totalspace >= mid2)
376 return(totalspace - mid2);
384 * Wait for the buffer cache to flush (totalspace) bytes worth of
385 * buffers, then return.
387 * Regardless this function blocks while the number of dirty buffers
388 * exceeds hidirtybufspace.
391 bd_wait(int totalspace)
396 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
399 while (totalspace > 0) {
402 if (totalspace > runningbufspace + dirtybufspace)
403 totalspace = runningbufspace + dirtybufspace;
404 count = totalspace / BKVASIZE;
405 if (count >= BD_WAKE_SIZE)
406 count = BD_WAKE_SIZE - 1;
407 i = (bd_wake_index + count) & BD_WAKE_MASK;
409 tsleep(&bd_wake_ary[i], 0, "flstik", hz);
412 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
419 * This function is called whenever runningbufspace or dirtybufspace
420 * is reduced. Track threads waiting for run+dirty buffer I/O
424 bd_signal(int totalspace)
428 while (totalspace > 0) {
429 i = atomic_fetchadd_int(&bd_wake_index, 1);
431 if (bd_wake_ary[i]) {
433 wakeup(&bd_wake_ary[i]);
435 totalspace -= BKVASIZE;
440 * BIO tracking support routines.
442 * Release a ref on a bio_track. Wakeup requests are atomically released
443 * along with the last reference so bk_active will never wind up set to
450 bio_track_rel(struct bio_track *track)
458 active = track->bk_active;
459 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
463 * Full-on. Note that the wait flag is only atomically released on
464 * the 1->0 count transition.
467 desired = (active & 0x7FFFFFFF) - 1;
469 desired |= active & 0x80000000;
470 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
472 panic("bio_track_rel: bad count: %p\n", track);
473 if (active & 0x80000000)
477 active = track->bk_active;
482 * Wait for the tracking count to reach 0.
484 * Use atomic ops such that the wait flag is only set atomically when
485 * bk_active is non-zero.
490 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
499 if (track->bk_active == 0)
503 * Full-on. Note that the wait flag may only be atomically set if
504 * the active count is non-zero.
506 crit_enter(); /* for tsleep_interlock */
508 while ((active = track->bk_active) != 0) {
509 desired = active | 0x80000000;
510 tsleep_interlock(track);
511 if (active == desired ||
512 atomic_cmpset_int(&track->bk_active, active, desired)) {
513 error = tsleep(track, slp_flags, "iowait", slp_timo);
525 * Load time initialisation of the buffer cache, called from machine
526 * dependant initialization code.
532 vm_offset_t bogus_offset;
535 spin_init(&needsbuffer_spin);
537 /* next, make a null set of free lists */
538 for (i = 0; i < BUFFER_QUEUES; i++)
539 TAILQ_INIT(&bufqueues[i]);
541 /* finally, initialize each buffer header and stick on empty q */
542 for (i = 0; i < nbuf; i++) {
544 bzero(bp, sizeof *bp);
545 bp->b_flags = B_INVAL; /* we're just an empty header */
546 bp->b_cmd = BUF_CMD_DONE;
547 bp->b_qindex = BQUEUE_EMPTY;
549 xio_init(&bp->b_xio);
552 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
556 * maxbufspace is the absolute maximum amount of buffer space we are
557 * allowed to reserve in KVM and in real terms. The absolute maximum
558 * is nominally used by buf_daemon. hibufspace is the nominal maximum
559 * used by most other processes. The differential is required to
560 * ensure that buf_daemon is able to run when other processes might
561 * be blocked waiting for buffer space.
563 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
564 * this may result in KVM fragmentation which is not handled optimally
567 maxbufspace = nbuf * BKVASIZE;
568 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
569 lobufspace = hibufspace - MAXBSIZE;
571 lorunningspace = 512 * 1024;
572 hirunningspace = 1024 * 1024;
575 * Limit the amount of malloc memory since it is wired permanently
576 * into the kernel space. Even though this is accounted for in
577 * the buffer allocation, we don't want the malloced region to grow
578 * uncontrolled. The malloc scheme improves memory utilization
579 * significantly on average (small) directories.
581 maxbufmallocspace = hibufspace / 20;
584 * Reduce the chance of a deadlock occuring by limiting the number
585 * of delayed-write dirty buffers we allow to stack up.
587 hidirtybufspace = hibufspace / 2;
591 lodirtybufspace = hidirtybufspace / 2;
594 * Maximum number of async ops initiated per buf_daemon loop. This is
595 * somewhat of a hack at the moment, we really need to limit ourselves
596 * based on the number of bytes of I/O in-transit that were initiated
600 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
601 bogus_page = vm_page_alloc(&kernel_object,
602 (bogus_offset >> PAGE_SHIFT),
604 vmstats.v_wire_count++;
609 * Initialize the embedded bio structures
612 initbufbio(struct buf *bp)
614 bp->b_bio1.bio_buf = bp;
615 bp->b_bio1.bio_prev = NULL;
616 bp->b_bio1.bio_offset = NOOFFSET;
617 bp->b_bio1.bio_next = &bp->b_bio2;
618 bp->b_bio1.bio_done = NULL;
620 bp->b_bio2.bio_buf = bp;
621 bp->b_bio2.bio_prev = &bp->b_bio1;
622 bp->b_bio2.bio_offset = NOOFFSET;
623 bp->b_bio2.bio_next = NULL;
624 bp->b_bio2.bio_done = NULL;
628 * Reinitialize the embedded bio structures as well as any additional
629 * translation cache layers.
632 reinitbufbio(struct buf *bp)
636 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
637 bio->bio_done = NULL;
638 bio->bio_offset = NOOFFSET;
643 * Push another BIO layer onto an existing BIO and return it. The new
644 * BIO layer may already exist, holding cached translation data.
647 push_bio(struct bio *bio)
651 if ((nbio = bio->bio_next) == NULL) {
652 int index = bio - &bio->bio_buf->b_bio_array[0];
653 if (index >= NBUF_BIO - 1) {
654 panic("push_bio: too many layers bp %p\n",
657 nbio = &bio->bio_buf->b_bio_array[index + 1];
658 bio->bio_next = nbio;
659 nbio->bio_prev = bio;
660 nbio->bio_buf = bio->bio_buf;
661 nbio->bio_offset = NOOFFSET;
662 nbio->bio_done = NULL;
663 nbio->bio_next = NULL;
665 KKASSERT(nbio->bio_done == NULL);
670 * Pop a BIO translation layer, returning the previous layer. The
671 * must have been previously pushed.
674 pop_bio(struct bio *bio)
676 return(bio->bio_prev);
680 clearbiocache(struct bio *bio)
683 bio->bio_offset = NOOFFSET;
691 * Free the KVA allocation for buffer 'bp'.
693 * Must be called from a critical section as this is the only locking for
696 * Since this call frees up buffer space, we call bufspacewakeup().
699 bfreekva(struct buf *bp)
705 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
706 vm_map_lock(&buffer_map);
707 bufspace -= bp->b_kvasize;
708 vm_map_delete(&buffer_map,
709 (vm_offset_t) bp->b_kvabase,
710 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
713 vm_map_unlock(&buffer_map);
714 vm_map_entry_release(count);
723 * Remove the buffer from the appropriate free list.
726 bremfree(struct buf *bp)
730 if (bp->b_qindex != BQUEUE_NONE) {
731 KASSERT(BUF_REFCNTNB(bp) == 1,
732 ("bremfree: bp %p not locked",bp));
733 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
734 bp->b_qindex = BQUEUE_NONE;
736 if (BUF_REFCNTNB(bp) <= 1)
737 panic("bremfree: removing a buffer not on a queue");
747 * Get a buffer with the specified data. Look in the cache first. We
748 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
749 * is set, the buffer is valid and we do not have to do anything ( see
753 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
757 bp = getblk(vp, loffset, size, 0, 0);
760 /* if not found in cache, do some I/O */
761 if ((bp->b_flags & B_CACHE) == 0) {
762 KASSERT(!(bp->b_flags & B_ASYNC),
763 ("bread: illegal async bp %p", bp));
764 bp->b_flags &= ~(B_ERROR | B_INVAL);
765 bp->b_cmd = BUF_CMD_READ;
766 vfs_busy_pages(vp, bp);
767 vn_strategy(vp, &bp->b_bio1);
768 return (biowait(bp));
776 * Operates like bread, but also starts asynchronous I/O on
777 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
778 * to initiating I/O . If B_CACHE is set, the buffer is valid
779 * and we do not have to do anything.
782 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
783 int *rabsize, int cnt, struct buf **bpp)
785 struct buf *bp, *rabp;
787 int rv = 0, readwait = 0;
789 *bpp = bp = getblk(vp, loffset, size, 0, 0);
791 /* if not found in cache, do some I/O */
792 if ((bp->b_flags & B_CACHE) == 0) {
793 bp->b_flags &= ~(B_ERROR | B_INVAL);
794 bp->b_cmd = BUF_CMD_READ;
795 vfs_busy_pages(vp, bp);
796 vn_strategy(vp, &bp->b_bio1);
800 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
801 if (inmem(vp, *raoffset))
803 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
805 if ((rabp->b_flags & B_CACHE) == 0) {
806 rabp->b_flags |= B_ASYNC;
807 rabp->b_flags &= ~(B_ERROR | B_INVAL);
808 rabp->b_cmd = BUF_CMD_READ;
809 vfs_busy_pages(vp, rabp);
811 vn_strategy(vp, &rabp->b_bio1);
826 * Write, release buffer on completion. (Done by iodone
827 * if async). Do not bother writing anything if the buffer
830 * Note that we set B_CACHE here, indicating that buffer is
831 * fully valid and thus cacheable. This is true even of NFS
832 * now so we set it generally. This could be set either here
833 * or in biodone() since the I/O is synchronous. We put it
837 bwrite(struct buf *bp)
841 if (bp->b_flags & B_INVAL) {
846 oldflags = bp->b_flags;
848 if (BUF_REFCNTNB(bp) == 0)
849 panic("bwrite: buffer is not busy???");
852 /* Mark the buffer clean */
855 bp->b_flags &= ~B_ERROR;
856 bp->b_flags |= B_CACHE;
857 bp->b_cmd = BUF_CMD_WRITE;
858 vfs_busy_pages(bp->b_vp, bp);
861 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
862 * valid for vnode-backed buffers.
864 bp->b_runningbufspace = bp->b_bufsize;
865 if (bp->b_runningbufspace) {
866 runningbufspace += bp->b_runningbufspace;
871 if (oldflags & B_ASYNC)
873 vn_strategy(bp->b_vp, &bp->b_bio1);
875 if ((oldflags & B_ASYNC) == 0) {
876 int rtval = biowait(bp);
886 * Delayed write. (Buffer is marked dirty). Do not bother writing
887 * anything if the buffer is marked invalid.
889 * Note that since the buffer must be completely valid, we can safely
890 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
891 * biodone() in order to prevent getblk from writing the buffer
895 bdwrite(struct buf *bp)
897 if (BUF_REFCNTNB(bp) == 0)
898 panic("bdwrite: buffer is not busy");
900 if (bp->b_flags & B_INVAL) {
907 * Set B_CACHE, indicating that the buffer is fully valid. This is
908 * true even of NFS now.
910 bp->b_flags |= B_CACHE;
913 * This bmap keeps the system from needing to do the bmap later,
914 * perhaps when the system is attempting to do a sync. Since it
915 * is likely that the indirect block -- or whatever other datastructure
916 * that the filesystem needs is still in memory now, it is a good
917 * thing to do this. Note also, that if the pageout daemon is
918 * requesting a sync -- there might not be enough memory to do
919 * the bmap then... So, this is important to do.
921 if (bp->b_bio2.bio_offset == NOOFFSET) {
922 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
923 NULL, NULL, BUF_CMD_WRITE);
927 * Set the *dirty* buffer range based upon the VM system dirty pages.
932 * We need to do this here to satisfy the vnode_pager and the
933 * pageout daemon, so that it thinks that the pages have been
934 * "cleaned". Note that since the pages are in a delayed write
935 * buffer -- the VFS layer "will" see that the pages get written
936 * out on the next sync, or perhaps the cluster will be completed.
942 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
943 * due to the softdep code.
950 * Turn buffer into delayed write request by marking it B_DELWRI.
951 * B_RELBUF and B_NOCACHE must be cleared.
953 * We reassign the buffer to itself to properly update it in the
956 * Must be called from a critical section.
957 * The buffer must be on BQUEUE_NONE.
960 bdirty(struct buf *bp)
962 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
963 if (bp->b_flags & B_NOCACHE) {
964 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
965 bp->b_flags &= ~B_NOCACHE;
967 if (bp->b_flags & B_INVAL) {
968 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
970 bp->b_flags &= ~B_RELBUF;
972 if ((bp->b_flags & B_DELWRI) == 0) {
973 bp->b_flags |= B_DELWRI;
976 dirtybufspace += bp->b_bufsize;
977 if (bp->b_flags & B_HEAVY) {
979 dirtybufspacehw += bp->b_bufsize;
986 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
987 * needs to be flushed with a different buf_daemon thread to avoid
988 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
991 bheavy(struct buf *bp)
993 if ((bp->b_flags & B_HEAVY) == 0) {
994 bp->b_flags |= B_HEAVY;
995 if (bp->b_flags & B_DELWRI) {
997 dirtybufspacehw += bp->b_bufsize;
1005 * Clear B_DELWRI for buffer.
1007 * Must be called from a critical section.
1009 * The buffer is typically on BQUEUE_NONE but there is one case in
1010 * brelse() that calls this function after placing the buffer on
1011 * a different queue.
1015 bundirty(struct buf *bp)
1017 if (bp->b_flags & B_DELWRI) {
1018 bp->b_flags &= ~B_DELWRI;
1021 dirtybufspace -= bp->b_bufsize;
1022 if (bp->b_flags & B_HEAVY) {
1024 dirtybufspacehw -= bp->b_bufsize;
1026 bd_signal(bp->b_bufsize);
1029 * Since it is now being written, we can clear its deferred write flag.
1031 bp->b_flags &= ~B_DEFERRED;
1037 * Asynchronous write. Start output on a buffer, but do not wait for
1038 * it to complete. The buffer is released when the output completes.
1040 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1041 * B_INVAL buffers. Not us.
1044 bawrite(struct buf *bp)
1046 bp->b_flags |= B_ASYNC;
1053 * Ordered write. Start output on a buffer, and flag it so that the
1054 * device will write it in the order it was queued. The buffer is
1055 * released when the output completes. bwrite() ( or the VOP routine
1056 * anyway ) is responsible for handling B_INVAL buffers.
1059 bowrite(struct buf *bp)
1061 bp->b_flags |= B_ORDERED | B_ASYNC;
1062 return (bwrite(bp));
1066 * buf_dirty_count_severe:
1068 * Return true if we have too many dirty buffers.
1071 buf_dirty_count_severe(void)
1073 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
1074 dirtybufcount >= nbuf / 2);
1080 * Release a busy buffer and, if requested, free its resources. The
1081 * buffer will be stashed in the appropriate bufqueue[] allowing it
1082 * to be accessed later as a cache entity or reused for other purposes.
1085 brelse(struct buf *bp)
1088 int saved_flags = bp->b_flags;
1091 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1096 * If B_NOCACHE is set we are being asked to destroy the buffer and
1097 * its backing store. Clear B_DELWRI.
1099 * B_NOCACHE is set in two cases: (1) when the caller really wants
1100 * to destroy the buffer and backing store and (2) when the caller
1101 * wants to destroy the buffer and backing store after a write
1104 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1108 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1110 * A re-dirtied buffer is only subject to destruction
1111 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1113 /* leave buffer intact */
1114 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1115 (bp->b_bufsize <= 0)) {
1117 * Either a failed read or we were asked to free or not
1118 * cache the buffer. This path is reached with B_DELWRI
1119 * set only if B_INVAL is already set. B_NOCACHE governs
1120 * backing store destruction.
1122 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1123 * buffer cannot be immediately freed.
1125 bp->b_flags |= B_INVAL;
1126 if (LIST_FIRST(&bp->b_dep) != NULL)
1128 if (bp->b_flags & B_DELWRI) {
1130 dirtybufspace -= bp->b_bufsize;
1131 if (bp->b_flags & B_HEAVY) {
1133 dirtybufspacehw -= bp->b_bufsize;
1135 bd_signal(bp->b_bufsize);
1137 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1141 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1142 * If vfs_vmio_release() is called with either bit set, the
1143 * underlying pages may wind up getting freed causing a previous
1144 * write (bdwrite()) to get 'lost' because pages associated with
1145 * a B_DELWRI bp are marked clean. Pages associated with a
1146 * B_LOCKED buffer may be mapped by the filesystem.
1148 * If we want to release the buffer ourselves (rather then the
1149 * originator asking us to release it), give the originator a
1150 * chance to countermand the release by setting B_LOCKED.
1152 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1153 * if B_DELWRI is set.
1155 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1156 * on pages to return pages to the VM page queues.
1158 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1159 bp->b_flags &= ~B_RELBUF;
1160 } else if (vm_page_count_severe()) {
1161 if (LIST_FIRST(&bp->b_dep) != NULL)
1162 buf_deallocate(bp); /* can set B_LOCKED */
1163 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1164 bp->b_flags &= ~B_RELBUF;
1166 bp->b_flags |= B_RELBUF;
1170 * Make sure b_cmd is clear. It may have already been cleared by
1173 * At this point destroying the buffer is governed by the B_INVAL
1174 * or B_RELBUF flags.
1176 bp->b_cmd = BUF_CMD_DONE;
1179 * VMIO buffer rundown. Make sure the VM page array is restored
1180 * after an I/O may have replaces some of the pages with bogus pages
1181 * in order to not destroy dirty pages in a fill-in read.
1183 * Note that due to the code above, if a buffer is marked B_DELWRI
1184 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1185 * B_INVAL may still be set, however.
1187 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1188 * but not the backing store. B_NOCACHE will destroy the backing
1191 * Note that dirty NFS buffers contain byte-granular write ranges
1192 * and should not be destroyed w/ B_INVAL even if the backing store
1195 if (bp->b_flags & B_VMIO) {
1197 * Rundown for VMIO buffers which are not dirty NFS buffers.
1209 * Get the base offset and length of the buffer. Note that
1210 * in the VMIO case if the buffer block size is not
1211 * page-aligned then b_data pointer may not be page-aligned.
1212 * But our b_xio.xio_pages array *IS* page aligned.
1214 * block sizes less then DEV_BSIZE (usually 512) are not
1215 * supported due to the page granularity bits (m->valid,
1216 * m->dirty, etc...).
1218 * See man buf(9) for more information
1221 resid = bp->b_bufsize;
1222 foff = bp->b_loffset;
1224 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1225 m = bp->b_xio.xio_pages[i];
1226 vm_page_flag_clear(m, PG_ZERO);
1228 * If we hit a bogus page, fixup *all* of them
1229 * now. Note that we left these pages wired
1230 * when we removed them so they had better exist,
1231 * and they cannot be ripped out from under us so
1232 * no critical section protection is necessary.
1234 if (m == bogus_page) {
1236 poff = OFF_TO_IDX(bp->b_loffset);
1238 for (j = i; j < bp->b_xio.xio_npages; j++) {
1241 mtmp = bp->b_xio.xio_pages[j];
1242 if (mtmp == bogus_page) {
1243 mtmp = vm_page_lookup(obj, poff + j);
1245 panic("brelse: page missing");
1247 bp->b_xio.xio_pages[j] = mtmp;
1251 if ((bp->b_flags & B_INVAL) == 0) {
1252 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1253 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1255 m = bp->b_xio.xio_pages[i];
1259 * Invalidate the backing store if B_NOCACHE is set
1260 * (e.g. used with vinvalbuf()). If this is NFS
1261 * we impose a requirement that the block size be
1262 * a multiple of PAGE_SIZE and create a temporary
1263 * hack to basically invalidate the whole page. The
1264 * problem is that NFS uses really odd buffer sizes
1265 * especially when tracking piecemeal writes and
1266 * it also vinvalbuf()'s a lot, which would result
1267 * in only partial page validation and invalidation
1268 * here. If the file page is mmap()'d, however,
1269 * all the valid bits get set so after we invalidate
1270 * here we would end up with weird m->valid values
1271 * like 0xfc. nfs_getpages() can't handle this so
1272 * we clear all the valid bits for the NFS case
1273 * instead of just some of them.
1275 * The real bug is the VM system having to set m->valid
1276 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1277 * itself is an artifact of the whole 512-byte
1278 * granular mess that exists to support odd block
1279 * sizes and UFS meta-data block sizes (e.g. 6144).
1280 * A complete rewrite is required.
1282 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1283 int poffset = foff & PAGE_MASK;
1286 presid = PAGE_SIZE - poffset;
1287 if (bp->b_vp->v_tag == VT_NFS &&
1288 bp->b_vp->v_type == VREG) {
1290 } else if (presid > resid) {
1293 KASSERT(presid >= 0, ("brelse: extra page"));
1294 vm_page_set_invalid(m, poffset, presid);
1296 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1297 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1299 if (bp->b_flags & (B_INVAL | B_RELBUF))
1300 vfs_vmio_release(bp);
1303 * Rundown for non-VMIO buffers.
1305 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1308 kprintf("brelse bp %p %08x/%08x: Warning, caught and fixed brelvp bug\n", bp, saved_flags, bp->b_flags);
1312 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1318 if (bp->b_qindex != BQUEUE_NONE)
1319 panic("brelse: free buffer onto another queue???");
1320 if (BUF_REFCNTNB(bp) > 1) {
1321 /* Temporary panic to verify exclusive locking */
1322 /* This panic goes away when we allow shared refs */
1323 panic("brelse: multiple refs");
1324 /* do not release to free list */
1331 * Figure out the correct queue to place the cleaned up buffer on.
1332 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1333 * disassociated from their vnode.
1335 if (bp->b_flags & B_LOCKED) {
1337 * Buffers that are locked are placed in the locked queue
1338 * immediately, regardless of their state.
1340 bp->b_qindex = BQUEUE_LOCKED;
1341 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1342 } else if (bp->b_bufsize == 0) {
1344 * Buffers with no memory. Due to conditionals near the top
1345 * of brelse() such buffers should probably already be
1346 * marked B_INVAL and disassociated from their vnode.
1348 bp->b_flags |= B_INVAL;
1349 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1350 KKASSERT((bp->b_flags & B_HASHED) == 0);
1351 if (bp->b_kvasize) {
1352 bp->b_qindex = BQUEUE_EMPTYKVA;
1354 bp->b_qindex = BQUEUE_EMPTY;
1356 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1357 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1359 * Buffers with junk contents. Again these buffers had better
1360 * already be disassociated from their vnode.
1362 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1363 KKASSERT((bp->b_flags & B_HASHED) == 0);
1364 bp->b_flags |= B_INVAL;
1365 bp->b_qindex = BQUEUE_CLEAN;
1366 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1369 * Remaining buffers. These buffers are still associated with
1372 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1374 bp->b_qindex = BQUEUE_DIRTY;
1375 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1377 case B_DELWRI | B_HEAVY:
1378 bp->b_qindex = BQUEUE_DIRTY_HW;
1379 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1384 * NOTE: Buffers are always placed at the end of the
1385 * queue. If B_AGE is not set the buffer will cycle
1386 * through the queue twice.
1388 bp->b_qindex = BQUEUE_CLEAN;
1389 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1395 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1396 * on the correct queue.
1398 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1402 * The bp is on an appropriate queue unless locked. If it is not
1403 * locked or dirty we can wakeup threads waiting for buffer space.
1405 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1406 * if B_INVAL is set ).
1408 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1412 * Something we can maybe free or reuse
1414 if (bp->b_bufsize || bp->b_kvasize)
1418 * Clean up temporary flags and unlock the buffer.
1420 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
1428 * Release a buffer back to the appropriate queue but do not try to free
1429 * it. The buffer is expected to be used again soon.
1431 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1432 * biodone() to requeue an async I/O on completion. It is also used when
1433 * known good buffers need to be requeued but we think we may need the data
1436 * XXX we should be able to leave the B_RELBUF hint set on completion.
1439 bqrelse(struct buf *bp)
1443 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1445 if (bp->b_qindex != BQUEUE_NONE)
1446 panic("bqrelse: free buffer onto another queue???");
1447 if (BUF_REFCNTNB(bp) > 1) {
1448 /* do not release to free list */
1449 panic("bqrelse: multiple refs");
1454 if (bp->b_flags & B_LOCKED) {
1456 * Locked buffers are released to the locked queue. However,
1457 * if the buffer is dirty it will first go into the dirty
1458 * queue and later on after the I/O completes successfully it
1459 * will be released to the locked queue.
1461 bp->b_qindex = BQUEUE_LOCKED;
1462 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1463 } else if (bp->b_flags & B_DELWRI) {
1464 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1465 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1466 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1467 } else if (vm_page_count_severe()) {
1469 * We are too low on memory, we have to try to free the
1470 * buffer (most importantly: the wired pages making up its
1471 * backing store) *now*.
1477 bp->b_qindex = BQUEUE_CLEAN;
1478 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1481 if ((bp->b_flags & B_LOCKED) == 0 &&
1482 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1487 * Something we can maybe free or reuse.
1489 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1493 * Final cleanup and unlock. Clear bits that are only used while a
1494 * buffer is actively locked.
1496 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF);
1504 * Return backing pages held by the buffer 'bp' back to the VM system
1505 * if possible. The pages are freed if they are no longer valid or
1506 * attempt to free if it was used for direct I/O otherwise they are
1507 * sent to the page cache.
1509 * Pages that were marked busy are left alone and skipped.
1511 * The KVA mapping (b_data) for the underlying pages is removed by
1515 vfs_vmio_release(struct buf *bp)
1521 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1522 m = bp->b_xio.xio_pages[i];
1523 bp->b_xio.xio_pages[i] = NULL;
1525 * In order to keep page LRU ordering consistent, put
1526 * everything on the inactive queue.
1528 vm_page_unwire(m, 0);
1530 * We don't mess with busy pages, it is
1531 * the responsibility of the process that
1532 * busied the pages to deal with them.
1534 if ((m->flags & PG_BUSY) || (m->busy != 0))
1537 if (m->wire_count == 0) {
1538 vm_page_flag_clear(m, PG_ZERO);
1540 * Might as well free the page if we can and it has
1541 * no valid data. We also free the page if the
1542 * buffer was used for direct I/O.
1544 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1545 m->hold_count == 0) {
1547 vm_page_protect(m, VM_PROT_NONE);
1549 } else if (bp->b_flags & B_DIRECT) {
1550 vm_page_try_to_free(m);
1551 } else if (vm_page_count_severe()) {
1552 vm_page_try_to_cache(m);
1557 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1558 if (bp->b_bufsize) {
1562 bp->b_xio.xio_npages = 0;
1563 bp->b_flags &= ~B_VMIO;
1564 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1572 * Implement clustered async writes for clearing out B_DELWRI buffers.
1573 * This is much better then the old way of writing only one buffer at
1574 * a time. Note that we may not be presented with the buffers in the
1575 * correct order, so we search for the cluster in both directions.
1577 * The buffer is locked on call.
1580 vfs_bio_awrite(struct buf *bp)
1584 off_t loffset = bp->b_loffset;
1585 struct vnode *vp = bp->b_vp;
1593 * right now we support clustered writing only to regular files. If
1594 * we find a clusterable block we could be in the middle of a cluster
1595 * rather then at the beginning.
1597 * NOTE: b_bio1 contains the logical loffset and is aliased
1598 * to b_loffset. b_bio2 contains the translated block number.
1600 if ((vp->v_type == VREG) &&
1601 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1602 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1604 size = vp->v_mount->mnt_stat.f_iosize;
1606 for (i = size; i < MAXPHYS; i += size) {
1607 if ((bpa = findblk(vp, loffset + i)) &&
1608 BUF_REFCNT(bpa) == 0 &&
1609 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1610 (B_DELWRI | B_CLUSTEROK)) &&
1611 (bpa->b_bufsize == size)) {
1612 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1613 (bpa->b_bio2.bio_offset !=
1614 bp->b_bio2.bio_offset + i))
1620 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1621 if ((bpa = findblk(vp, loffset - j)) &&
1622 BUF_REFCNT(bpa) == 0 &&
1623 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1624 (B_DELWRI | B_CLUSTEROK)) &&
1625 (bpa->b_bufsize == size)) {
1626 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1627 (bpa->b_bio2.bio_offset !=
1628 bp->b_bio2.bio_offset - j))
1637 * this is a possible cluster write
1639 if (nbytes != size) {
1641 nwritten = cluster_wbuild(vp, size,
1642 loffset - j, nbytes);
1649 bp->b_flags |= B_ASYNC;
1653 * default (old) behavior, writing out only one block
1655 * XXX returns b_bufsize instead of b_bcount for nwritten?
1657 nwritten = bp->b_bufsize;
1666 * Find and initialize a new buffer header, freeing up existing buffers
1667 * in the bufqueues as necessary. The new buffer is returned locked.
1669 * Important: B_INVAL is not set. If the caller wishes to throw the
1670 * buffer away, the caller must set B_INVAL prior to calling brelse().
1673 * We have insufficient buffer headers
1674 * We have insufficient buffer space
1675 * buffer_map is too fragmented ( space reservation fails )
1676 * If we have to flush dirty buffers ( but we try to avoid this )
1678 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1679 * Instead we ask the buf daemon to do it for us. We attempt to
1680 * avoid piecemeal wakeups of the pageout daemon.
1684 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1690 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1691 static int flushingbufs;
1694 * We can't afford to block since we might be holding a vnode lock,
1695 * which may prevent system daemons from running. We deal with
1696 * low-memory situations by proactively returning memory and running
1697 * async I/O rather then sync I/O.
1701 --getnewbufrestarts;
1703 ++getnewbufrestarts;
1706 * Setup for scan. If we do not have enough free buffers,
1707 * we setup a degenerate case that immediately fails. Note
1708 * that if we are specially marked process, we are allowed to
1709 * dip into our reserves.
1711 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1713 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1714 * However, there are a number of cases (defragging, reusing, ...)
1715 * where we cannot backup.
1717 nqindex = BQUEUE_EMPTYKVA;
1718 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1722 * If no EMPTYKVA buffers and we are either
1723 * defragging or reusing, locate a CLEAN buffer
1724 * to free or reuse. If bufspace useage is low
1725 * skip this step so we can allocate a new buffer.
1727 if (defrag || bufspace >= lobufspace) {
1728 nqindex = BQUEUE_CLEAN;
1729 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1733 * If we could not find or were not allowed to reuse a
1734 * CLEAN buffer, check to see if it is ok to use an EMPTY
1735 * buffer. We can only use an EMPTY buffer if allocating
1736 * its KVA would not otherwise run us out of buffer space.
1738 if (nbp == NULL && defrag == 0 &&
1739 bufspace + maxsize < hibufspace) {
1740 nqindex = BQUEUE_EMPTY;
1741 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1746 * Run scan, possibly freeing data and/or kva mappings on the fly
1750 while ((bp = nbp) != NULL) {
1751 int qindex = nqindex;
1753 nbp = TAILQ_NEXT(bp, b_freelist);
1756 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1757 * cycles through the queue twice before being selected.
1759 if (qindex == BQUEUE_CLEAN &&
1760 (bp->b_flags & B_AGE) == 0 && nbp) {
1761 bp->b_flags |= B_AGE;
1762 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1763 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1768 * Calculate next bp ( we can only use it if we do not block
1769 * or do other fancy things ).
1774 nqindex = BQUEUE_EMPTYKVA;
1775 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1778 case BQUEUE_EMPTYKVA:
1779 nqindex = BQUEUE_CLEAN;
1780 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1794 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1797 * Note: we no longer distinguish between VMIO and non-VMIO
1801 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1804 * If we are defragging then we need a buffer with
1805 * b_kvasize != 0. XXX this situation should no longer
1806 * occur, if defrag is non-zero the buffer's b_kvasize
1807 * should also be non-zero at this point. XXX
1809 if (defrag && bp->b_kvasize == 0) {
1810 kprintf("Warning: defrag empty buffer %p\n", bp);
1815 * Start freeing the bp. This is somewhat involved. nbp
1816 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1817 * on the clean list must be disassociated from their
1818 * current vnode. Buffers on the empty[kva] lists have
1819 * already been disassociated.
1822 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1823 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1824 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1827 if (bp->b_qindex != qindex) {
1828 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1835 * Dependancies must be handled before we disassociate the
1838 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1839 * be immediately disassociated. HAMMER then becomes
1840 * responsible for releasing the buffer.
1842 if (LIST_FIRST(&bp->b_dep) != NULL) {
1844 if (bp->b_flags & B_LOCKED) {
1848 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1851 if (qindex == BQUEUE_CLEAN) {
1852 if (bp->b_flags & B_VMIO) {
1853 bp->b_flags &= ~B_ASYNC;
1854 vfs_vmio_release(bp);
1861 * NOTE: nbp is now entirely invalid. We can only restart
1862 * the scan from this point on.
1864 * Get the rest of the buffer freed up. b_kva* is still
1865 * valid after this operation.
1868 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1869 KKASSERT((bp->b_flags & B_HASHED) == 0);
1872 * critical section protection is not required when
1873 * scrapping a buffer's contents because it is already
1879 bp->b_flags = B_BNOCLIP;
1880 bp->b_cmd = BUF_CMD_DONE;
1885 bp->b_xio.xio_npages = 0;
1886 bp->b_dirtyoff = bp->b_dirtyend = 0;
1888 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1890 if (blkflags & GETBLK_BHEAVY)
1891 bp->b_flags |= B_HEAVY;
1894 * If we are defragging then free the buffer.
1897 bp->b_flags |= B_INVAL;
1905 * If we are overcomitted then recover the buffer and its
1906 * KVM space. This occurs in rare situations when multiple
1907 * processes are blocked in getnewbuf() or allocbuf().
1909 if (bufspace >= hibufspace)
1911 if (flushingbufs && bp->b_kvasize != 0) {
1912 bp->b_flags |= B_INVAL;
1917 if (bufspace < lobufspace)
1923 * If we exhausted our list, sleep as appropriate. We may have to
1924 * wakeup various daemons and write out some dirty buffers.
1926 * Generally we are sleeping due to insufficient buffer space.
1934 flags = VFS_BIO_NEED_BUFSPACE;
1936 } else if (bufspace >= hibufspace) {
1938 flags = VFS_BIO_NEED_BUFSPACE;
1941 flags = VFS_BIO_NEED_ANY;
1944 needsbuffer |= flags;
1945 bd_speedup(); /* heeeelp */
1946 while (needsbuffer & flags) {
1947 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
1952 * We finally have a valid bp. We aren't quite out of the
1953 * woods, we still have to reserve kva space. In order
1954 * to keep fragmentation sane we only allocate kva in
1957 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1959 if (maxsize != bp->b_kvasize) {
1960 vm_offset_t addr = 0;
1965 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1966 vm_map_lock(&buffer_map);
1968 if (vm_map_findspace(&buffer_map,
1969 vm_map_min(&buffer_map), maxsize,
1970 maxsize, 0, &addr)) {
1972 * Uh oh. Buffer map is too fragmented. We
1973 * must defragment the map.
1975 vm_map_unlock(&buffer_map);
1976 vm_map_entry_release(count);
1979 bp->b_flags |= B_INVAL;
1984 vm_map_insert(&buffer_map, &count,
1986 addr, addr + maxsize,
1988 VM_PROT_ALL, VM_PROT_ALL,
1991 bp->b_kvabase = (caddr_t) addr;
1992 bp->b_kvasize = maxsize;
1993 bufspace += bp->b_kvasize;
1996 vm_map_unlock(&buffer_map);
1997 vm_map_entry_release(count);
1999 bp->b_data = bp->b_kvabase;
2005 * This routine is called in an emergency to recover VM pages from the
2006 * buffer cache by cashing in clean buffers. The idea is to recover
2007 * enough pages to be able to satisfy a stuck bio_page_alloc().
2010 recoverbufpages(void)
2017 while (bytes < MAXBSIZE) {
2018 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2023 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2024 * cycles through the queue twice before being selected.
2026 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
2027 bp->b_flags |= B_AGE;
2028 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
2029 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
2037 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
2038 KKASSERT((bp->b_flags & B_DELWRI) == 0);
2041 * Start freeing the bp. This is somewhat involved.
2043 * Buffers on the clean list must be disassociated from
2044 * their current vnode
2047 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2048 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp);
2049 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
2052 if (bp->b_qindex != BQUEUE_CLEAN) {
2053 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp, bp->b_qindex);
2060 * Dependancies must be handled before we disassociate the
2063 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2064 * be immediately disassociated. HAMMER then becomes
2065 * responsible for releasing the buffer.
2067 if (LIST_FIRST(&bp->b_dep) != NULL) {
2069 if (bp->b_flags & B_LOCKED) {
2073 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2076 bytes += bp->b_bufsize;
2078 if (bp->b_flags & B_VMIO) {
2079 bp->b_flags &= ~B_ASYNC;
2080 bp->b_flags |= B_DIRECT; /* try to free pages */
2081 vfs_vmio_release(bp);
2086 KKASSERT(bp->b_vp == NULL);
2087 KKASSERT((bp->b_flags & B_HASHED) == 0);
2090 * critical section protection is not required when
2091 * scrapping a buffer's contents because it is already
2097 bp->b_flags = B_BNOCLIP;
2098 bp->b_cmd = BUF_CMD_DONE;
2103 bp->b_xio.xio_npages = 0;
2104 bp->b_dirtyoff = bp->b_dirtyend = 0;
2106 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2108 bp->b_flags |= B_INVAL;
2118 * Buffer flushing daemon. Buffers are normally flushed by the
2119 * update daemon but if it cannot keep up this process starts to
2120 * take the load in an attempt to prevent getnewbuf() from blocking.
2122 * Once a flush is initiated it does not stop until the number
2123 * of buffers falls below lodirtybuffers, but we will wake up anyone
2124 * waiting at the mid-point.
2127 static struct kproc_desc buf_kp = {
2132 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2133 kproc_start, &buf_kp)
2135 static struct kproc_desc bufhw_kp = {
2140 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2141 kproc_start, &bufhw_kp)
2149 * This process needs to be suspended prior to shutdown sync.
2151 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2152 bufdaemon_td, SHUTDOWN_PRI_LAST);
2153 curthread->td_flags |= TDF_SYSTHREAD;
2156 * This process is allowed to take the buffer cache to the limit
2161 kproc_suspend_loop();
2164 * Do the flush. Limit the amount of in-transit I/O we
2165 * allow to build up, otherwise we would completely saturate
2166 * the I/O system. Wakeup any waiting processes before we
2167 * normally would so they can run in parallel with our drain.
2169 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2170 * but because we split the operation into two threads we
2171 * have to cut it in half for each thread.
2173 limit = lodirtybufspace / 2;
2174 waitrunningbufspace(limit);
2175 while (runningbufspace + dirtybufspace > limit ||
2176 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
2177 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2179 waitrunningbufspace(limit);
2183 * We reached our low water mark, reset the
2184 * request and sleep until we are needed again.
2185 * The sleep is just so the suspend code works.
2187 spin_lock_wr(&needsbuffer_spin);
2188 if (bd_request == 0) {
2189 msleep(&bd_request, &needsbuffer_spin, 0,
2193 spin_unlock_wr(&needsbuffer_spin);
2203 * This process needs to be suspended prior to shutdown sync.
2205 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2206 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2207 curthread->td_flags |= TDF_SYSTHREAD;
2210 * This process is allowed to take the buffer cache to the limit
2215 kproc_suspend_loop();
2218 * Do the flush. Limit the amount of in-transit I/O we
2219 * allow to build up, otherwise we would completely saturate
2220 * the I/O system. Wakeup any waiting processes before we
2221 * normally would so they can run in parallel with our drain.
2223 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2224 * but because we split the operation into two threads we
2225 * have to cut it in half for each thread.
2227 limit = lodirtybufspace / 2;
2228 waitrunningbufspace(limit);
2229 while (runningbufspace + dirtybufspacehw > limit ||
2230 dirtybufcounthw >= nbuf / 2) {
2231 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2233 waitrunningbufspace(limit);
2237 * We reached our low water mark, reset the
2238 * request and sleep until we are needed again.
2239 * The sleep is just so the suspend code works.
2241 spin_lock_wr(&needsbuffer_spin);
2242 if (bd_request_hw == 0) {
2243 msleep(&bd_request_hw, &needsbuffer_spin, 0,
2247 spin_unlock_wr(&needsbuffer_spin);
2254 * Try to flush a buffer in the dirty queue. We must be careful to
2255 * free up B_INVAL buffers instead of write them, which NFS is
2256 * particularly sensitive to.
2258 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2259 * that we really want to try to get the buffer out and reuse it
2260 * due to the write load on the machine.
2264 flushbufqueues(bufq_type_t q)
2269 bp = TAILQ_FIRST(&bufqueues[q]);
2271 KASSERT((bp->b_flags & B_DELWRI),
2272 ("unexpected clean buffer %p", bp));
2274 if (bp->b_flags & B_DELWRI) {
2275 if (bp->b_flags & B_INVAL) {
2276 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2277 panic("flushbufqueues: locked buf");
2283 if (LIST_FIRST(&bp->b_dep) != NULL &&
2284 (bp->b_flags & B_DEFERRED) == 0 &&
2285 buf_countdeps(bp, 0)) {
2286 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2287 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2289 bp->b_flags |= B_DEFERRED;
2290 bp = TAILQ_FIRST(&bufqueues[q]);
2295 * Only write it out if we can successfully lock
2296 * it. If the buffer has a dependancy,
2297 * buf_checkwrite must also return 0 for us to
2298 * be able to initate the write.
2300 * If the buffer is flagged B_ERROR it may be
2301 * requeued over and over again, we try to
2302 * avoid a live lock.
2304 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
2305 if (LIST_FIRST(&bp->b_dep) != NULL &&
2306 buf_checkwrite(bp)) {
2309 } else if (bp->b_flags & B_ERROR) {
2310 tsleep(bp, 0, "bioer", 1);
2311 bp->b_flags &= ~B_AGE;
2314 bp->b_flags |= B_AGE;
2321 bp = TAILQ_NEXT(bp, b_freelist);
2329 * Returns true if no I/O is needed to access the associated VM object.
2330 * This is like findblk except it also hunts around in the VM system for
2333 * Note that we ignore vm_page_free() races from interrupts against our
2334 * lookup, since if the caller is not protected our return value will not
2335 * be any more valid then otherwise once we exit the critical section.
2338 inmem(struct vnode *vp, off_t loffset)
2341 vm_offset_t toff, tinc, size;
2344 if (findblk(vp, loffset))
2346 if (vp->v_mount == NULL)
2348 if ((obj = vp->v_object) == NULL)
2352 if (size > vp->v_mount->mnt_stat.f_iosize)
2353 size = vp->v_mount->mnt_stat.f_iosize;
2355 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2356 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2360 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2361 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2362 if (vm_page_is_valid(m,
2363 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2372 * Sets the dirty range for a buffer based on the status of the dirty
2373 * bits in the pages comprising the buffer.
2375 * The range is limited to the size of the buffer.
2377 * This routine is primarily used by NFS, but is generalized for the
2381 vfs_setdirty(struct buf *bp)
2387 * Degenerate case - empty buffer
2390 if (bp->b_bufsize == 0)
2394 * We qualify the scan for modified pages on whether the
2395 * object has been flushed yet. The OBJ_WRITEABLE flag
2396 * is not cleared simply by protecting pages off.
2399 if ((bp->b_flags & B_VMIO) == 0)
2402 object = bp->b_xio.xio_pages[0]->object;
2404 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2405 kprintf("Warning: object %p writeable but not mightbedirty\n", object);
2406 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2407 kprintf("Warning: object %p mightbedirty but not writeable\n", object);
2409 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2410 vm_offset_t boffset;
2411 vm_offset_t eoffset;
2414 * test the pages to see if they have been modified directly
2415 * by users through the VM system.
2417 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2418 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2419 vm_page_test_dirty(bp->b_xio.xio_pages[i]);
2423 * Calculate the encompassing dirty range, boffset and eoffset,
2424 * (eoffset - boffset) bytes.
2427 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2428 if (bp->b_xio.xio_pages[i]->dirty)
2431 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2433 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2434 if (bp->b_xio.xio_pages[i]->dirty) {
2438 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2441 * Fit it to the buffer.
2444 if (eoffset > bp->b_bcount)
2445 eoffset = bp->b_bcount;
2448 * If we have a good dirty range, merge with the existing
2452 if (boffset < eoffset) {
2453 if (bp->b_dirtyoff > boffset)
2454 bp->b_dirtyoff = boffset;
2455 if (bp->b_dirtyend < eoffset)
2456 bp->b_dirtyend = eoffset;
2464 * Locate and return the specified buffer, or NULL if the buffer does
2465 * not exist. Do not attempt to lock the buffer or manipulate it in
2466 * any way. The caller must validate that the correct buffer has been
2467 * obtain after locking it.
2470 findblk(struct vnode *vp, off_t loffset)
2475 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2483 * Get a block given a specified block and offset into a file/device.
2484 * B_INVAL may or may not be set on return. The caller should clear
2485 * B_INVAL prior to initiating a READ.
2487 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2488 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2489 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2490 * without doing any of those things the system will likely believe
2491 * the buffer to be valid (especially if it is not B_VMIO), and the
2492 * next getblk() will return the buffer with B_CACHE set.
2494 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2495 * an existing buffer.
2497 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2498 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2499 * and then cleared based on the backing VM. If the previous buffer is
2500 * non-0-sized but invalid, B_CACHE will be cleared.
2502 * If getblk() must create a new buffer, the new buffer is returned with
2503 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2504 * case it is returned with B_INVAL clear and B_CACHE set based on the
2507 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2508 * B_CACHE bit is clear.
2510 * What this means, basically, is that the caller should use B_CACHE to
2511 * determine whether the buffer is fully valid or not and should clear
2512 * B_INVAL prior to issuing a read. If the caller intends to validate
2513 * the buffer by loading its data area with something, the caller needs
2514 * to clear B_INVAL. If the caller does this without issuing an I/O,
2515 * the caller should set B_CACHE ( as an optimization ), else the caller
2516 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2517 * a write attempt or if it was a successfull read. If the caller
2518 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2519 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2523 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2524 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2527 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2530 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2533 if (size > MAXBSIZE)
2534 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2535 if (vp->v_object == NULL)
2536 panic("getblk: vnode %p has no object!", vp);
2540 if ((bp = findblk(vp, loffset))) {
2542 * The buffer was found in the cache, but we need to lock it.
2543 * Even with LK_NOWAIT the lockmgr may break our critical
2544 * section, so double-check the validity of the buffer
2545 * once the lock has been obtained.
2547 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2548 if (blkflags & GETBLK_NOWAIT) {
2552 int lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2553 if (blkflags & GETBLK_PCATCH)
2554 lkflags |= LK_PCATCH;
2555 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2557 if (error == ENOLCK)
2565 * Once the buffer has been locked, make sure we didn't race
2566 * a buffer recyclement. Buffers that are no longer hashed
2567 * will have b_vp == NULL, so this takes care of that check
2570 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2571 kprintf("Warning buffer %p (vp %p loffset %lld) "
2573 bp, vp, (long long)loffset);
2579 * If SZMATCH any pre-existing buffer must be of the requested
2580 * size or NULL is returned. The caller absolutely does not
2581 * want getblk() to bwrite() the buffer on a size mismatch.
2583 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2590 * All vnode-based buffers must be backed by a VM object.
2592 KKASSERT(bp->b_flags & B_VMIO);
2593 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2594 bp->b_flags &= ~B_AGE;
2597 * Make sure that B_INVAL buffers do not have a cached
2598 * block number translation.
2600 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2601 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2602 " did not have cleared bio_offset cache\n",
2603 bp, vp, (long long)loffset);
2604 clearbiocache(&bp->b_bio2);
2608 * The buffer is locked. B_CACHE is cleared if the buffer is
2611 if (bp->b_flags & B_INVAL)
2612 bp->b_flags &= ~B_CACHE;
2616 * Any size inconsistancy with a dirty buffer or a buffer
2617 * with a softupdates dependancy must be resolved. Resizing
2618 * the buffer in such circumstances can lead to problems.
2620 if (size != bp->b_bcount) {
2621 if (bp->b_flags & B_DELWRI) {
2622 bp->b_flags |= B_NOCACHE;
2624 } else if (LIST_FIRST(&bp->b_dep)) {
2625 bp->b_flags |= B_NOCACHE;
2628 bp->b_flags |= B_RELBUF;
2633 KKASSERT(size <= bp->b_kvasize);
2634 KASSERT(bp->b_loffset != NOOFFSET,
2635 ("getblk: no buffer offset"));
2638 * A buffer with B_DELWRI set and B_CACHE clear must
2639 * be committed before we can return the buffer in
2640 * order to prevent the caller from issuing a read
2641 * ( due to B_CACHE not being set ) and overwriting
2644 * Most callers, including NFS and FFS, need this to
2645 * operate properly either because they assume they
2646 * can issue a read if B_CACHE is not set, or because
2647 * ( for example ) an uncached B_DELWRI might loop due
2648 * to softupdates re-dirtying the buffer. In the latter
2649 * case, B_CACHE is set after the first write completes,
2650 * preventing further loops.
2652 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2653 * above while extending the buffer, we cannot allow the
2654 * buffer to remain with B_CACHE set after the write
2655 * completes or it will represent a corrupt state. To
2656 * deal with this we set B_NOCACHE to scrap the buffer
2659 * We might be able to do something fancy, like setting
2660 * B_CACHE in bwrite() except if B_DELWRI is already set,
2661 * so the below call doesn't set B_CACHE, but that gets real
2662 * confusing. This is much easier.
2665 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2666 bp->b_flags |= B_NOCACHE;
2673 * Buffer is not in-core, create new buffer. The buffer
2674 * returned by getnewbuf() is locked. Note that the returned
2675 * buffer is also considered valid (not marked B_INVAL).
2677 * Calculating the offset for the I/O requires figuring out
2678 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2679 * the mount's f_iosize otherwise. If the vnode does not
2680 * have an associated mount we assume that the passed size is
2683 * Note that vn_isdisk() cannot be used here since it may
2684 * return a failure for numerous reasons. Note that the
2685 * buffer size may be larger then the block size (the caller
2686 * will use block numbers with the proper multiple). Beware
2687 * of using any v_* fields which are part of unions. In
2688 * particular, in DragonFly the mount point overloading
2689 * mechanism uses the namecache only and the underlying
2690 * directory vnode is not a special case.
2694 if (vp->v_type == VBLK || vp->v_type == VCHR)
2696 else if (vp->v_mount)
2697 bsize = vp->v_mount->mnt_stat.f_iosize;
2701 maxsize = size + (loffset & PAGE_MASK);
2702 maxsize = imax(maxsize, bsize);
2704 if ((bp = getnewbuf(blkflags, slptimeo, size, maxsize)) == NULL) {
2705 if (slpflags || slptimeo) {
2713 * This code is used to make sure that a buffer is not
2714 * created while the getnewbuf routine is blocked.
2715 * This can be a problem whether the vnode is locked or not.
2716 * If the buffer is created out from under us, we have to
2717 * throw away the one we just created. There is no window
2718 * race because we are safely running in a critical section
2719 * from the point of the duplicate buffer creation through
2720 * to here, and we've locked the buffer.
2722 if (findblk(vp, loffset)) {
2723 bp->b_flags |= B_INVAL;
2729 * Insert the buffer into the hash, so that it can
2730 * be found by findblk().
2732 * Make sure the translation layer has been cleared.
2734 bp->b_loffset = loffset;
2735 bp->b_bio2.bio_offset = NOOFFSET;
2736 /* bp->b_bio2.bio_next = NULL; */
2741 * All vnode-based buffers must be backed by a VM object.
2743 KKASSERT(vp->v_object != NULL);
2744 bp->b_flags |= B_VMIO;
2745 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2757 * Reacquire a buffer that was previously released to the locked queue,
2758 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2759 * set B_LOCKED (which handles the acquisition race).
2761 * To this end, either B_LOCKED must be set or the dependancy list must be
2765 regetblk(struct buf *bp)
2767 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
2768 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2777 * Get an empty, disassociated buffer of given size. The buffer is
2778 * initially set to B_INVAL.
2780 * critical section protection is not required for the allocbuf()
2781 * call because races are impossible here.
2789 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2792 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2796 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2804 * This code constitutes the buffer memory from either anonymous system
2805 * memory (in the case of non-VMIO operations) or from an associated
2806 * VM object (in the case of VMIO operations). This code is able to
2807 * resize a buffer up or down.
2809 * Note that this code is tricky, and has many complications to resolve
2810 * deadlock or inconsistant data situations. Tread lightly!!!
2811 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2812 * the caller. Calling this code willy nilly can result in the loss of data.
2814 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2815 * B_CACHE for the non-VMIO case.
2817 * This routine does not need to be called from a critical section but you
2818 * must own the buffer.
2821 allocbuf(struct buf *bp, int size)
2823 int newbsize, mbsize;
2826 if (BUF_REFCNT(bp) == 0)
2827 panic("allocbuf: buffer not busy");
2829 if (bp->b_kvasize < size)
2830 panic("allocbuf: buffer too small");
2832 if ((bp->b_flags & B_VMIO) == 0) {
2836 * Just get anonymous memory from the kernel. Don't
2837 * mess with B_CACHE.
2839 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2840 if (bp->b_flags & B_MALLOC)
2843 newbsize = round_page(size);
2845 if (newbsize < bp->b_bufsize) {
2847 * Malloced buffers are not shrunk
2849 if (bp->b_flags & B_MALLOC) {
2851 bp->b_bcount = size;
2853 kfree(bp->b_data, M_BIOBUF);
2854 if (bp->b_bufsize) {
2855 bufmallocspace -= bp->b_bufsize;
2859 bp->b_data = bp->b_kvabase;
2861 bp->b_flags &= ~B_MALLOC;
2867 (vm_offset_t) bp->b_data + newbsize,
2868 (vm_offset_t) bp->b_data + bp->b_bufsize);
2869 } else if (newbsize > bp->b_bufsize) {
2871 * We only use malloced memory on the first allocation.
2872 * and revert to page-allocated memory when the buffer
2875 if ((bufmallocspace < maxbufmallocspace) &&
2876 (bp->b_bufsize == 0) &&
2877 (mbsize <= PAGE_SIZE/2)) {
2879 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
2880 bp->b_bufsize = mbsize;
2881 bp->b_bcount = size;
2882 bp->b_flags |= B_MALLOC;
2883 bufmallocspace += mbsize;
2889 * If the buffer is growing on its other-than-first
2890 * allocation, then we revert to the page-allocation
2893 if (bp->b_flags & B_MALLOC) {
2894 origbuf = bp->b_data;
2895 origbufsize = bp->b_bufsize;
2896 bp->b_data = bp->b_kvabase;
2897 if (bp->b_bufsize) {
2898 bufmallocspace -= bp->b_bufsize;
2902 bp->b_flags &= ~B_MALLOC;
2903 newbsize = round_page(newbsize);
2907 (vm_offset_t) bp->b_data + bp->b_bufsize,
2908 (vm_offset_t) bp->b_data + newbsize);
2910 bcopy(origbuf, bp->b_data, origbufsize);
2911 kfree(origbuf, M_BIOBUF);
2918 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2919 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
2920 newbsize + PAGE_MASK) >> PAGE_SHIFT;
2921 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
2923 if (bp->b_flags & B_MALLOC)
2924 panic("allocbuf: VMIO buffer can't be malloced");
2926 * Set B_CACHE initially if buffer is 0 length or will become
2929 if (size == 0 || bp->b_bufsize == 0)
2930 bp->b_flags |= B_CACHE;
2932 if (newbsize < bp->b_bufsize) {
2934 * DEV_BSIZE aligned new buffer size is less then the
2935 * DEV_BSIZE aligned existing buffer size. Figure out
2936 * if we have to remove any pages.
2938 if (desiredpages < bp->b_xio.xio_npages) {
2939 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
2941 * the page is not freed here -- it
2942 * is the responsibility of
2943 * vnode_pager_setsize
2945 m = bp->b_xio.xio_pages[i];
2946 KASSERT(m != bogus_page,
2947 ("allocbuf: bogus page found"));
2948 while (vm_page_sleep_busy(m, TRUE, "biodep"))
2951 bp->b_xio.xio_pages[i] = NULL;
2952 vm_page_unwire(m, 0);
2954 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2955 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
2956 bp->b_xio.xio_npages = desiredpages;
2958 } else if (size > bp->b_bcount) {
2960 * We are growing the buffer, possibly in a
2961 * byte-granular fashion.
2969 * Step 1, bring in the VM pages from the object,
2970 * allocating them if necessary. We must clear
2971 * B_CACHE if these pages are not valid for the
2972 * range covered by the buffer.
2974 * critical section protection is required to protect
2975 * against interrupts unbusying and freeing pages
2976 * between our vm_page_lookup() and our
2977 * busycheck/wiring call.
2983 while (bp->b_xio.xio_npages < desiredpages) {
2987 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
2988 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2990 * note: must allocate system pages
2991 * since blocking here could intefere
2992 * with paging I/O, no matter which
2995 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
2999 bp->b_flags &= ~B_CACHE;
3000 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3001 ++bp->b_xio.xio_npages;
3007 * We found a page. If we have to sleep on it,
3008 * retry because it might have gotten freed out
3011 * We can only test PG_BUSY here. Blocking on
3012 * m->busy might lead to a deadlock:
3014 * vm_fault->getpages->cluster_read->allocbuf
3018 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
3020 vm_page_flag_clear(m, PG_ZERO);
3022 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3023 ++bp->b_xio.xio_npages;
3028 * Step 2. We've loaded the pages into the buffer,
3029 * we have to figure out if we can still have B_CACHE
3030 * set. Note that B_CACHE is set according to the
3031 * byte-granular range ( bcount and size ), not the
3032 * aligned range ( newbsize ).
3034 * The VM test is against m->valid, which is DEV_BSIZE
3035 * aligned. Needless to say, the validity of the data
3036 * needs to also be DEV_BSIZE aligned. Note that this
3037 * fails with NFS if the server or some other client
3038 * extends the file's EOF. If our buffer is resized,
3039 * B_CACHE may remain set! XXX
3042 toff = bp->b_bcount;
3043 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3045 while ((bp->b_flags & B_CACHE) && toff < size) {
3048 if (tinc > (size - toff))
3051 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3059 bp->b_xio.xio_pages[pi]
3066 * Step 3, fixup the KVM pmap. Remember that
3067 * bp->b_data is relative to bp->b_loffset, but
3068 * bp->b_loffset may be offset into the first page.
3071 bp->b_data = (caddr_t)
3072 trunc_page((vm_offset_t)bp->b_data);
3074 (vm_offset_t)bp->b_data,
3075 bp->b_xio.xio_pages,
3076 bp->b_xio.xio_npages
3078 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3079 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3083 /* adjust space use on already-dirty buffer */
3084 if (bp->b_flags & B_DELWRI) {
3085 dirtybufspace += newbsize - bp->b_bufsize;
3086 if (bp->b_flags & B_HEAVY)
3087 dirtybufspacehw += newbsize - bp->b_bufsize;
3089 if (newbsize < bp->b_bufsize)
3091 bp->b_bufsize = newbsize; /* actual buffer allocation */
3092 bp->b_bcount = size; /* requested buffer size */
3099 * Wait for buffer I/O completion, returning error status. The buffer
3100 * is left locked on return. B_EINTR is converted into an EINTR error
3103 * NOTE! The original b_cmd is lost on return, since b_cmd will be
3104 * set to BUF_CMD_DONE.
3107 biowait(struct buf *bp)
3110 while (bp->b_cmd != BUF_CMD_DONE) {
3111 if (bp->b_cmd == BUF_CMD_READ)
3112 tsleep(bp, 0, "biord", 0);
3114 tsleep(bp, 0, "biowr", 0);
3117 if (bp->b_flags & B_EINTR) {
3118 bp->b_flags &= ~B_EINTR;
3121 if (bp->b_flags & B_ERROR) {
3122 return (bp->b_error ? bp->b_error : EIO);
3129 * This associates a tracking count with an I/O. vn_strategy() and
3130 * dev_dstrategy() do this automatically but there are a few cases
3131 * where a vnode or device layer is bypassed when a block translation
3132 * is cached. In such cases bio_start_transaction() may be called on
3133 * the bypassed layers so the system gets an I/O in progress indication
3134 * for those higher layers.
3137 bio_start_transaction(struct bio *bio, struct bio_track *track)
3139 bio->bio_track = track;
3140 bio_track_ref(track);
3144 * Initiate I/O on a vnode.
3147 vn_strategy(struct vnode *vp, struct bio *bio)
3149 struct bio_track *track;
3151 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
3152 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
3153 track = &vp->v_track_read;
3155 track = &vp->v_track_write;
3156 bio->bio_track = track;
3157 bio_track_ref(track);
3158 vop_strategy(*vp->v_ops, vp, bio);
3165 * Finish I/O on a buffer, optionally calling a completion function.
3166 * This is usually called from an interrupt so process blocking is
3169 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3170 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3171 * assuming B_INVAL is clear.
3173 * For the VMIO case, we set B_CACHE if the op was a read and no
3174 * read error occured, or if the op was a write. B_CACHE is never
3175 * set if the buffer is invalid or otherwise uncacheable.
3177 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3178 * initiator to leave B_INVAL set to brelse the buffer out of existance
3179 * in the biodone routine.
3182 biodone(struct bio *bio)
3184 struct buf *bp = bio->bio_buf;
3189 KASSERT(BUF_REFCNTNB(bp) > 0,
3190 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3191 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3192 ("biodone: bp %p already done!", bp));
3194 runningbufwakeup(bp);
3197 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3200 biodone_t *done_func;
3201 struct bio_track *track;
3204 * BIO tracking. Most but not all BIOs are tracked.
3206 if ((track = bio->bio_track) != NULL) {
3207 bio_track_rel(track);
3208 bio->bio_track = NULL;
3212 * A bio_done function terminates the loop. The function
3213 * will be responsible for any further chaining and/or
3214 * buffer management.
3216 * WARNING! The done function can deallocate the buffer!
3218 if ((done_func = bio->bio_done) != NULL) {
3219 bio->bio_done = NULL;
3224 bio = bio->bio_prev;
3228 bp->b_cmd = BUF_CMD_DONE;
3231 * Only reads and writes are processed past this point.
3233 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3234 if (cmd == BUF_CMD_FREEBLKS)
3235 bp->b_flags |= B_NOCACHE;
3242 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3243 * a lot worse. XXX - move this above the clearing of b_cmd
3245 if (LIST_FIRST(&bp->b_dep) != NULL)
3249 * A failed write must re-dirty the buffer unless B_INVAL
3252 if (cmd == BUF_CMD_WRITE &&
3253 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3254 bp->b_flags &= ~B_NOCACHE;
3259 if (bp->b_flags & B_VMIO) {
3265 struct vnode *vp = bp->b_vp;
3269 #if defined(VFS_BIO_DEBUG)
3270 if (vp->v_auxrefs == 0)
3271 panic("biodone: zero vnode hold count");
3272 if ((vp->v_flag & VOBJBUF) == 0)
3273 panic("biodone: vnode is not setup for merged cache");
3276 foff = bp->b_loffset;
3277 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3278 KASSERT(obj != NULL, ("biodone: missing VM object"));
3280 #if defined(VFS_BIO_DEBUG)
3281 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3282 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3283 obj->paging_in_progress, bp->b_xio.xio_npages);
3288 * Set B_CACHE if the op was a normal read and no error
3289 * occured. B_CACHE is set for writes in the b*write()
3292 iosize = bp->b_bcount - bp->b_resid;
3293 if (cmd == BUF_CMD_READ && (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3294 bp->b_flags |= B_CACHE;
3297 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3301 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3306 * cleanup bogus pages, restoring the originals. Since
3307 * the originals should still be wired, we don't have
3308 * to worry about interrupt/freeing races destroying
3309 * the VM object association.
3311 m = bp->b_xio.xio_pages[i];
3312 if (m == bogus_page) {
3314 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3316 panic("biodone: page disappeared");
3317 bp->b_xio.xio_pages[i] = m;
3318 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3319 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3321 #if defined(VFS_BIO_DEBUG)
3322 if (OFF_TO_IDX(foff) != m->pindex) {
3324 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
3325 (unsigned long)foff, m->pindex);
3330 * In the write case, the valid and clean bits are
3331 * already changed correctly ( see bdwrite() ), so we
3332 * only need to do this here in the read case.
3334 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3335 vfs_page_set_valid(bp, foff, i, m);
3337 vm_page_flag_clear(m, PG_ZERO);
3340 * when debugging new filesystems or buffer I/O methods, this
3341 * is the most common error that pops up. if you see this, you
3342 * have not set the page busy flag correctly!!!
3345 kprintf("biodone: page busy < 0, "
3346 "pindex: %d, foff: 0x(%x,%x), "
3347 "resid: %d, index: %d\n",
3348 (int) m->pindex, (int)(foff >> 32),
3349 (int) foff & 0xffffffff, resid, i);
3350 if (!vn_isdisk(vp, NULL))
3351 kprintf(" iosize: %ld, loffset: %lld, "
3352 "flags: 0x%08x, npages: %d\n",
3353 bp->b_vp->v_mount->mnt_stat.f_iosize,
3354 (long long)bp->b_loffset,
3355 bp->b_flags, bp->b_xio.xio_npages);
3357 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3358 (long long)bp->b_loffset,
3359 bp->b_flags, bp->b_xio.xio_npages);
3360 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3361 m->valid, m->dirty, m->wire_count);
3362 panic("biodone: page busy < 0");
3364 vm_page_io_finish(m);
3365 vm_object_pip_subtract(obj, 1);
3366 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3370 vm_object_pip_wakeupn(obj, 0);
3374 * For asynchronous completions, release the buffer now. The brelse
3375 * will do a wakeup there if necessary - so no need to do a wakeup
3376 * here in the async case. The sync case always needs to do a wakeup.
3379 if (bp->b_flags & B_ASYNC) {
3380 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
3393 * This routine is called in lieu of iodone in the case of
3394 * incomplete I/O. This keeps the busy status for pages
3398 vfs_unbusy_pages(struct buf *bp)
3402 runningbufwakeup(bp);
3403 if (bp->b_flags & B_VMIO) {
3404 struct vnode *vp = bp->b_vp;
3409 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3410 vm_page_t m = bp->b_xio.xio_pages[i];
3413 * When restoring bogus changes the original pages
3414 * should still be wired, so we are in no danger of
3415 * losing the object association and do not need
3416 * critical section protection particularly.
3418 if (m == bogus_page) {
3419 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3421 panic("vfs_unbusy_pages: page missing");
3423 bp->b_xio.xio_pages[i] = m;
3424 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3425 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3427 vm_object_pip_subtract(obj, 1);
3428 vm_page_flag_clear(m, PG_ZERO);
3429 vm_page_io_finish(m);
3431 vm_object_pip_wakeupn(obj, 0);
3436 * vfs_page_set_valid:
3438 * Set the valid bits in a page based on the supplied offset. The
3439 * range is restricted to the buffer's size.
3441 * This routine is typically called after a read completes.
3444 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3446 vm_ooffset_t soff, eoff;
3449 * Start and end offsets in buffer. eoff - soff may not cross a
3450 * page boundry or cross the end of the buffer. The end of the
3451 * buffer, in this case, is our file EOF, not the allocation size
3455 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3456 if (eoff > bp->b_loffset + bp->b_bcount)
3457 eoff = bp->b_loffset + bp->b_bcount;
3460 * Set valid range. This is typically the entire buffer and thus the
3464 vm_page_set_validclean(
3466 (vm_offset_t) (soff & PAGE_MASK),
3467 (vm_offset_t) (eoff - soff)
3475 * This routine is called before a device strategy routine.
3476 * It is used to tell the VM system that paging I/O is in
3477 * progress, and treat the pages associated with the buffer
3478 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3479 * flag is handled to make sure that the object doesn't become
3482 * Since I/O has not been initiated yet, certain buffer flags
3483 * such as B_ERROR or B_INVAL may be in an inconsistant state
3484 * and should be ignored.
3487 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3490 struct lwp *lp = curthread->td_lwp;
3493 * The buffer's I/O command must already be set. If reading,
3494 * B_CACHE must be 0 (double check against callers only doing
3495 * I/O when B_CACHE is 0).
3497 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3498 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3500 if (bp->b_flags & B_VMIO) {
3505 foff = bp->b_loffset;
3506 KASSERT(bp->b_loffset != NOOFFSET,
3507 ("vfs_busy_pages: no buffer offset"));
3511 * Loop until none of the pages are busy.
3514 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3515 vm_page_t m = bp->b_xio.xio_pages[i];
3517 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3522 * Setup for I/O, soft-busy the page right now because
3523 * the next loop may block.
3525 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3526 vm_page_t m = bp->b_xio.xio_pages[i];
3528 vm_page_flag_clear(m, PG_ZERO);
3529 if ((bp->b_flags & B_CLUSTER) == 0) {
3530 vm_object_pip_add(obj, 1);
3531 vm_page_io_start(m);
3536 * Adjust protections for I/O and do bogus-page mapping.
3537 * Assume that vm_page_protect() can block (it can block
3538 * if VM_PROT_NONE, don't take any chances regardless).
3541 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3542 vm_page_t m = bp->b_xio.xio_pages[i];
3545 * When readying a vnode-backed buffer for a write
3546 * we must zero-fill any invalid portions of the
3549 * When readying a vnode-backed buffer for a read
3550 * we must replace any dirty pages with a bogus
3551 * page so we do not destroy dirty data when
3552 * filling in gaps. Dirty pages might not
3553 * necessarily be marked dirty yet, so use m->valid
3554 * as a reasonable test.
3556 * Bogus page replacement is, uh, bogus. We need
3557 * to find a better way.
3559 if (bp->b_cmd == BUF_CMD_WRITE) {
3560 vm_page_protect(m, VM_PROT_READ);
3561 vfs_page_set_valid(bp, foff, i, m);
3562 } else if (m->valid == VM_PAGE_BITS_ALL) {
3563 bp->b_xio.xio_pages[i] = bogus_page;
3566 vm_page_protect(m, VM_PROT_NONE);
3568 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3571 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3572 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3576 * This is the easiest place to put the process accounting for the I/O
3580 if (bp->b_cmd == BUF_CMD_READ)
3581 lp->lwp_ru.ru_inblock++;
3583 lp->lwp_ru.ru_oublock++;
3590 * Tell the VM system that the pages associated with this buffer
3591 * are clean. This is used for delayed writes where the data is
3592 * going to go to disk eventually without additional VM intevention.
3594 * Note that while we only really need to clean through to b_bcount, we
3595 * just go ahead and clean through to b_bufsize.
3598 vfs_clean_pages(struct buf *bp)
3602 if (bp->b_flags & B_VMIO) {
3605 foff = bp->b_loffset;
3606 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset"));
3607 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3608 vm_page_t m = bp->b_xio.xio_pages[i];
3609 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3611 vfs_page_set_valid(bp, foff, i, m);
3618 * vfs_bio_set_validclean:
3620 * Set the range within the buffer to valid and clean. The range is
3621 * relative to the beginning of the buffer, b_loffset. Note that
3622 * b_loffset itself may be offset from the beginning of the first page.
3626 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3628 if (bp->b_flags & B_VMIO) {
3633 * Fixup base to be relative to beginning of first page.
3634 * Set initial n to be the maximum number of bytes in the
3635 * first page that can be validated.
3638 base += (bp->b_loffset & PAGE_MASK);
3639 n = PAGE_SIZE - (base & PAGE_MASK);
3641 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) {
3642 vm_page_t m = bp->b_xio.xio_pages[i];
3647 vm_page_set_validclean(m, base & PAGE_MASK, n);
3658 * Clear a buffer. This routine essentially fakes an I/O, so we need
3659 * to clear B_ERROR and B_INVAL.
3661 * Note that while we only theoretically need to clear through b_bcount,
3662 * we go ahead and clear through b_bufsize.
3666 vfs_bio_clrbuf(struct buf *bp)
3670 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3671 bp->b_flags &= ~(B_INVAL|B_ERROR);
3672 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3673 (bp->b_loffset & PAGE_MASK) == 0) {
3674 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3675 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3679 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3680 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3681 bzero(bp->b_data, bp->b_bufsize);
3682 bp->b_xio.xio_pages[0]->valid |= mask;
3688 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3689 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3690 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3691 ea = (caddr_t)(vm_offset_t)ulmin(
3692 (u_long)(vm_offset_t)ea,
3693 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3694 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3695 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3697 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3698 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3702 for (; sa < ea; sa += DEV_BSIZE, j++) {
3703 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3704 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3705 bzero(sa, DEV_BSIZE);
3708 bp->b_xio.xio_pages[i]->valid |= mask;
3709 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
3718 * vm_hold_load_pages:
3720 * Load pages into the buffer's address space. The pages are
3721 * allocated from the kernel object in order to reduce interference
3722 * with the any VM paging I/O activity. The range of loaded
3723 * pages will be wired.
3725 * If a page cannot be allocated, the 'pagedaemon' is woken up to
3726 * retrieve the full range (to - from) of pages.
3730 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3736 to = round_page(to);
3737 from = round_page(from);
3738 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3743 * Note: must allocate system pages since blocking here
3744 * could intefere with paging I/O, no matter which
3747 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT,
3748 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
3751 p->valid = VM_PAGE_BITS_ALL;
3752 vm_page_flag_clear(p, PG_ZERO);
3753 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3754 bp->b_xio.xio_pages[index] = p;
3761 bp->b_xio.xio_npages = index;
3765 * Allocate pages for a buffer cache buffer.
3767 * Under extremely severe memory conditions even allocating out of the
3768 * system reserve can fail. If this occurs we must allocate out of the
3769 * interrupt reserve to avoid a deadlock with the pageout daemon.
3771 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
3772 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
3773 * against the pageout daemon if pages are not freed from other sources.
3777 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit)
3782 * Try a normal allocation, allow use of system reserve.
3784 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
3789 * The normal allocation failed and we clearly have a page
3790 * deficit. Try to reclaim some clean VM pages directly
3791 * from the buffer cache.
3793 vm_pageout_deficit += deficit;
3797 * We may have blocked, the caller will know what to do if the
3800 if (vm_page_lookup(obj, pg))
3804 * Allocate and allow use of the interrupt reserve.
3806 * If after all that we still can't allocate a VM page we are
3807 * in real trouble, but we slog on anyway hoping that the system
3810 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
3811 VM_ALLOC_INTERRUPT);
3813 if (vm_page_count_severe()) {
3814 kprintf("bio_page_alloc: WARNING emergency page "
3819 kprintf("bio_page_alloc: WARNING emergency page "
3820 "allocation failed\n");
3827 * vm_hold_free_pages:
3829 * Return pages associated with the buffer back to the VM system.
3831 * The range of pages underlying the buffer's address space will
3832 * be unmapped and un-wired.
3835 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3839 int index, newnpages;
3841 from = round_page(from);
3842 to = round_page(to);
3843 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3845 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3846 p = bp->b_xio.xio_pages[index];
3847 if (p && (index < bp->b_xio.xio_npages)) {
3849 kprintf("vm_hold_free_pages: doffset: %lld, "
3851 (long long)bp->b_bio2.bio_offset,
3852 (long long)bp->b_loffset);
3854 bp->b_xio.xio_pages[index] = NULL;
3857 vm_page_unwire(p, 0);
3861 bp->b_xio.xio_npages = newnpages;
3867 * Map a user buffer into KVM via a pbuf. On return the buffer's
3868 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
3872 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
3883 * bp had better have a command and it better be a pbuf.
3885 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3886 KKASSERT(bp->b_flags & B_PAGING);
3892 * Map the user data into KVM. Mappings have to be page-aligned.
3894 addr = (caddr_t)trunc_page((vm_offset_t)udata);
3897 vmprot = VM_PROT_READ;
3898 if (bp->b_cmd == BUF_CMD_READ)
3899 vmprot |= VM_PROT_WRITE;
3901 while (addr < udata + bytes) {
3903 * Do the vm_fault if needed; do the copy-on-write thing
3904 * when reading stuff off device into memory.
3906 * vm_fault_page*() returns a held VM page.
3908 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
3909 va = trunc_page(va);
3911 m = vm_fault_page_quick(va, vmprot, &error);
3913 for (i = 0; i < pidx; ++i) {
3914 vm_page_unhold(bp->b_xio.xio_pages[i]);
3915 bp->b_xio.xio_pages[i] = NULL;
3919 bp->b_xio.xio_pages[pidx] = m;
3925 * Map the page array and set the buffer fields to point to
3926 * the mapped data buffer.
3928 if (pidx > btoc(MAXPHYS))
3929 panic("vmapbuf: mapped more than MAXPHYS");
3930 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
3932 bp->b_xio.xio_npages = pidx;
3933 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
3934 bp->b_bcount = bytes;
3935 bp->b_bufsize = bytes;
3942 * Free the io map PTEs associated with this IO operation.
3943 * We also invalidate the TLB entries and restore the original b_addr.
3946 vunmapbuf(struct buf *bp)
3951 KKASSERT(bp->b_flags & B_PAGING);
3953 npages = bp->b_xio.xio_npages;
3954 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3955 for (pidx = 0; pidx < npages; ++pidx) {
3956 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
3957 bp->b_xio.xio_pages[pidx] = NULL;
3959 bp->b_xio.xio_npages = 0;
3960 bp->b_data = bp->b_kvabase;
3964 * Scan all buffers in the system and issue the callback.
3967 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
3973 for (n = 0; n < nbuf; ++n) {
3974 if ((error = callback(&buf[n], info)) < 0) {
3984 * print out statistics from the current status of the buffer pool
3985 * this can be toggeled by the system control option debug.syncprt
3994 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
3995 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
3997 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
3999 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4002 TAILQ_FOREACH(bp, dp, b_freelist) {
4003 counts[bp->b_bufsize/PAGE_SIZE]++;
4007 kprintf("%s: total-%d", bname[i], count);
4008 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4010 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
4018 DB_SHOW_COMMAND(buffer, db_show_buffer)
4021 struct buf *bp = (struct buf *)addr;
4024 db_printf("usage: show buffer <addr>\n");
4028 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4029 db_printf("b_cmd = %d\n", bp->b_cmd);
4030 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4031 "b_resid = %d\n, b_data = %p, "
4032 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4033 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4035 (long long)bp->b_bio2.bio_offset,
4036 (long long)(bp->b_bio2.bio_next ?
4037 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4038 if (bp->b_xio.xio_npages) {
4040 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4041 bp->b_xio.xio_npages);
4042 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4044 m = bp->b_xio.xio_pages[i];
4045 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4046 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4047 if ((i + 1) < bp->b_xio.xio_npages)