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.
466 * We check for a negative count transition using bit 30 since bit 31
467 * has a different meaning.
470 desired = (active & 0x7FFFFFFF) - 1;
472 desired |= active & 0x80000000;
473 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
474 if (desired & 0x40000000)
475 panic("bio_track_rel: bad count: %p\n", track);
476 if (active & 0x80000000)
480 active = track->bk_active;
485 * Wait for the tracking count to reach 0.
487 * Use atomic ops such that the wait flag is only set atomically when
488 * bk_active is non-zero.
493 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
502 if (track->bk_active == 0)
506 * Full-on. Note that the wait flag may only be atomically set if
507 * the active count is non-zero.
509 crit_enter(); /* for tsleep_interlock */
511 while ((active = track->bk_active) != 0) {
512 desired = active | 0x80000000;
513 tsleep_interlock(track);
514 if (active == desired ||
515 atomic_cmpset_int(&track->bk_active, active, desired)) {
516 error = tsleep(track, slp_flags, "iowait", slp_timo);
528 * Load time initialisation of the buffer cache, called from machine
529 * dependant initialization code.
535 vm_offset_t bogus_offset;
538 spin_init(&needsbuffer_spin);
540 /* next, make a null set of free lists */
541 for (i = 0; i < BUFFER_QUEUES; i++)
542 TAILQ_INIT(&bufqueues[i]);
544 /* finally, initialize each buffer header and stick on empty q */
545 for (i = 0; i < nbuf; i++) {
547 bzero(bp, sizeof *bp);
548 bp->b_flags = B_INVAL; /* we're just an empty header */
549 bp->b_cmd = BUF_CMD_DONE;
550 bp->b_qindex = BQUEUE_EMPTY;
552 xio_init(&bp->b_xio);
555 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
559 * maxbufspace is the absolute maximum amount of buffer space we are
560 * allowed to reserve in KVM and in real terms. The absolute maximum
561 * is nominally used by buf_daemon. hibufspace is the nominal maximum
562 * used by most other processes. The differential is required to
563 * ensure that buf_daemon is able to run when other processes might
564 * be blocked waiting for buffer space.
566 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
567 * this may result in KVM fragmentation which is not handled optimally
570 maxbufspace = nbuf * BKVASIZE;
571 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
572 lobufspace = hibufspace - MAXBSIZE;
574 lorunningspace = 512 * 1024;
575 hirunningspace = 1024 * 1024;
578 * Limit the amount of malloc memory since it is wired permanently
579 * into the kernel space. Even though this is accounted for in
580 * the buffer allocation, we don't want the malloced region to grow
581 * uncontrolled. The malloc scheme improves memory utilization
582 * significantly on average (small) directories.
584 maxbufmallocspace = hibufspace / 20;
587 * Reduce the chance of a deadlock occuring by limiting the number
588 * of delayed-write dirty buffers we allow to stack up.
590 hidirtybufspace = hibufspace / 2;
594 lodirtybufspace = hidirtybufspace / 2;
597 * Maximum number of async ops initiated per buf_daemon loop. This is
598 * somewhat of a hack at the moment, we really need to limit ourselves
599 * based on the number of bytes of I/O in-transit that were initiated
603 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
604 bogus_page = vm_page_alloc(&kernel_object,
605 (bogus_offset >> PAGE_SHIFT),
607 vmstats.v_wire_count++;
612 * Initialize the embedded bio structures
615 initbufbio(struct buf *bp)
617 bp->b_bio1.bio_buf = bp;
618 bp->b_bio1.bio_prev = NULL;
619 bp->b_bio1.bio_offset = NOOFFSET;
620 bp->b_bio1.bio_next = &bp->b_bio2;
621 bp->b_bio1.bio_done = NULL;
623 bp->b_bio2.bio_buf = bp;
624 bp->b_bio2.bio_prev = &bp->b_bio1;
625 bp->b_bio2.bio_offset = NOOFFSET;
626 bp->b_bio2.bio_next = NULL;
627 bp->b_bio2.bio_done = NULL;
631 * Reinitialize the embedded bio structures as well as any additional
632 * translation cache layers.
635 reinitbufbio(struct buf *bp)
639 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
640 bio->bio_done = NULL;
641 bio->bio_offset = NOOFFSET;
646 * Push another BIO layer onto an existing BIO and return it. The new
647 * BIO layer may already exist, holding cached translation data.
650 push_bio(struct bio *bio)
654 if ((nbio = bio->bio_next) == NULL) {
655 int index = bio - &bio->bio_buf->b_bio_array[0];
656 if (index >= NBUF_BIO - 1) {
657 panic("push_bio: too many layers bp %p\n",
660 nbio = &bio->bio_buf->b_bio_array[index + 1];
661 bio->bio_next = nbio;
662 nbio->bio_prev = bio;
663 nbio->bio_buf = bio->bio_buf;
664 nbio->bio_offset = NOOFFSET;
665 nbio->bio_done = NULL;
666 nbio->bio_next = NULL;
668 KKASSERT(nbio->bio_done == NULL);
673 * Pop a BIO translation layer, returning the previous layer. The
674 * must have been previously pushed.
677 pop_bio(struct bio *bio)
679 return(bio->bio_prev);
683 clearbiocache(struct bio *bio)
686 bio->bio_offset = NOOFFSET;
694 * Free the KVA allocation for buffer 'bp'.
696 * Must be called from a critical section as this is the only locking for
699 * Since this call frees up buffer space, we call bufspacewakeup().
702 bfreekva(struct buf *bp)
708 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
709 vm_map_lock(&buffer_map);
710 bufspace -= bp->b_kvasize;
711 vm_map_delete(&buffer_map,
712 (vm_offset_t) bp->b_kvabase,
713 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
716 vm_map_unlock(&buffer_map);
717 vm_map_entry_release(count);
726 * Remove the buffer from the appropriate free list.
729 bremfree(struct buf *bp)
733 if (bp->b_qindex != BQUEUE_NONE) {
734 KASSERT(BUF_REFCNTNB(bp) == 1,
735 ("bremfree: bp %p not locked",bp));
736 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
737 bp->b_qindex = BQUEUE_NONE;
739 if (BUF_REFCNTNB(bp) <= 1)
740 panic("bremfree: removing a buffer not on a queue");
750 * Get a buffer with the specified data. Look in the cache first. We
751 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
752 * is set, the buffer is valid and we do not have to do anything ( see
756 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
760 bp = getblk(vp, loffset, size, 0, 0);
763 /* if not found in cache, do some I/O */
764 if ((bp->b_flags & B_CACHE) == 0) {
765 KASSERT(!(bp->b_flags & B_ASYNC),
766 ("bread: illegal async bp %p", bp));
767 bp->b_flags &= ~(B_ERROR | B_INVAL);
768 bp->b_cmd = BUF_CMD_READ;
769 vfs_busy_pages(vp, bp);
770 vn_strategy(vp, &bp->b_bio1);
771 return (biowait(bp));
779 * Operates like bread, but also starts asynchronous I/O on
780 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
781 * to initiating I/O . If B_CACHE is set, the buffer is valid
782 * and we do not have to do anything.
785 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
786 int *rabsize, int cnt, struct buf **bpp)
788 struct buf *bp, *rabp;
790 int rv = 0, readwait = 0;
792 *bpp = bp = getblk(vp, loffset, size, 0, 0);
794 /* if not found in cache, do some I/O */
795 if ((bp->b_flags & B_CACHE) == 0) {
796 bp->b_flags &= ~(B_ERROR | B_INVAL);
797 bp->b_cmd = BUF_CMD_READ;
798 vfs_busy_pages(vp, bp);
799 vn_strategy(vp, &bp->b_bio1);
803 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
804 if (inmem(vp, *raoffset))
806 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
808 if ((rabp->b_flags & B_CACHE) == 0) {
809 rabp->b_flags |= B_ASYNC;
810 rabp->b_flags &= ~(B_ERROR | B_INVAL);
811 rabp->b_cmd = BUF_CMD_READ;
812 vfs_busy_pages(vp, rabp);
814 vn_strategy(vp, &rabp->b_bio1);
829 * Write, release buffer on completion. (Done by iodone
830 * if async). Do not bother writing anything if the buffer
833 * Note that we set B_CACHE here, indicating that buffer is
834 * fully valid and thus cacheable. This is true even of NFS
835 * now so we set it generally. This could be set either here
836 * or in biodone() since the I/O is synchronous. We put it
840 bwrite(struct buf *bp)
844 if (bp->b_flags & B_INVAL) {
849 oldflags = bp->b_flags;
851 if (BUF_REFCNTNB(bp) == 0)
852 panic("bwrite: buffer is not busy???");
855 /* Mark the buffer clean */
858 bp->b_flags &= ~B_ERROR;
859 bp->b_flags |= B_CACHE;
860 bp->b_cmd = BUF_CMD_WRITE;
861 vfs_busy_pages(bp->b_vp, bp);
864 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
865 * valid for vnode-backed buffers.
867 bp->b_runningbufspace = bp->b_bufsize;
868 if (bp->b_runningbufspace) {
869 runningbufspace += bp->b_runningbufspace;
874 if (oldflags & B_ASYNC)
876 vn_strategy(bp->b_vp, &bp->b_bio1);
878 if ((oldflags & B_ASYNC) == 0) {
879 int rtval = biowait(bp);
889 * Delayed write. (Buffer is marked dirty). Do not bother writing
890 * anything if the buffer is marked invalid.
892 * Note that since the buffer must be completely valid, we can safely
893 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
894 * biodone() in order to prevent getblk from writing the buffer
898 bdwrite(struct buf *bp)
900 if (BUF_REFCNTNB(bp) == 0)
901 panic("bdwrite: buffer is not busy");
903 if (bp->b_flags & B_INVAL) {
910 * Set B_CACHE, indicating that the buffer is fully valid. This is
911 * true even of NFS now.
913 bp->b_flags |= B_CACHE;
916 * This bmap keeps the system from needing to do the bmap later,
917 * perhaps when the system is attempting to do a sync. Since it
918 * is likely that the indirect block -- or whatever other datastructure
919 * that the filesystem needs is still in memory now, it is a good
920 * thing to do this. Note also, that if the pageout daemon is
921 * requesting a sync -- there might not be enough memory to do
922 * the bmap then... So, this is important to do.
924 if (bp->b_bio2.bio_offset == NOOFFSET) {
925 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
926 NULL, NULL, BUF_CMD_WRITE);
930 * Set the *dirty* buffer range based upon the VM system dirty pages.
935 * We need to do this here to satisfy the vnode_pager and the
936 * pageout daemon, so that it thinks that the pages have been
937 * "cleaned". Note that since the pages are in a delayed write
938 * buffer -- the VFS layer "will" see that the pages get written
939 * out on the next sync, or perhaps the cluster will be completed.
945 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
946 * due to the softdep code.
953 * Turn buffer into delayed write request by marking it B_DELWRI.
954 * B_RELBUF and B_NOCACHE must be cleared.
956 * We reassign the buffer to itself to properly update it in the
959 * Must be called from a critical section.
960 * The buffer must be on BQUEUE_NONE.
963 bdirty(struct buf *bp)
965 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
966 if (bp->b_flags & B_NOCACHE) {
967 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
968 bp->b_flags &= ~B_NOCACHE;
970 if (bp->b_flags & B_INVAL) {
971 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
973 bp->b_flags &= ~B_RELBUF;
975 if ((bp->b_flags & B_DELWRI) == 0) {
976 bp->b_flags |= B_DELWRI;
979 dirtybufspace += bp->b_bufsize;
980 if (bp->b_flags & B_HEAVY) {
982 dirtybufspacehw += bp->b_bufsize;
989 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
990 * needs to be flushed with a different buf_daemon thread to avoid
991 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
994 bheavy(struct buf *bp)
996 if ((bp->b_flags & B_HEAVY) == 0) {
997 bp->b_flags |= B_HEAVY;
998 if (bp->b_flags & B_DELWRI) {
1000 dirtybufspacehw += bp->b_bufsize;
1008 * Clear B_DELWRI for buffer.
1010 * Must be called from a critical section.
1012 * The buffer is typically on BQUEUE_NONE but there is one case in
1013 * brelse() that calls this function after placing the buffer on
1014 * a different queue.
1018 bundirty(struct buf *bp)
1020 if (bp->b_flags & B_DELWRI) {
1021 bp->b_flags &= ~B_DELWRI;
1024 dirtybufspace -= bp->b_bufsize;
1025 if (bp->b_flags & B_HEAVY) {
1027 dirtybufspacehw -= bp->b_bufsize;
1029 bd_signal(bp->b_bufsize);
1032 * Since it is now being written, we can clear its deferred write flag.
1034 bp->b_flags &= ~B_DEFERRED;
1040 * Asynchronous write. Start output on a buffer, but do not wait for
1041 * it to complete. The buffer is released when the output completes.
1043 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1044 * B_INVAL buffers. Not us.
1047 bawrite(struct buf *bp)
1049 bp->b_flags |= B_ASYNC;
1056 * Ordered write. Start output on a buffer, and flag it so that the
1057 * device will write it in the order it was queued. The buffer is
1058 * released when the output completes. bwrite() ( or the VOP routine
1059 * anyway ) is responsible for handling B_INVAL buffers.
1062 bowrite(struct buf *bp)
1064 bp->b_flags |= B_ORDERED | B_ASYNC;
1065 return (bwrite(bp));
1069 * buf_dirty_count_severe:
1071 * Return true if we have too many dirty buffers.
1074 buf_dirty_count_severe(void)
1076 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
1077 dirtybufcount >= nbuf / 2);
1083 * Release a busy buffer and, if requested, free its resources. The
1084 * buffer will be stashed in the appropriate bufqueue[] allowing it
1085 * to be accessed later as a cache entity or reused for other purposes.
1088 brelse(struct buf *bp)
1091 int saved_flags = bp->b_flags;
1094 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1099 * If B_NOCACHE is set we are being asked to destroy the buffer and
1100 * its backing store. Clear B_DELWRI.
1102 * B_NOCACHE is set in two cases: (1) when the caller really wants
1103 * to destroy the buffer and backing store and (2) when the caller
1104 * wants to destroy the buffer and backing store after a write
1107 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1111 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1113 * A re-dirtied buffer is only subject to destruction
1114 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1116 /* leave buffer intact */
1117 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1118 (bp->b_bufsize <= 0)) {
1120 * Either a failed read or we were asked to free or not
1121 * cache the buffer. This path is reached with B_DELWRI
1122 * set only if B_INVAL is already set. B_NOCACHE governs
1123 * backing store destruction.
1125 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1126 * buffer cannot be immediately freed.
1128 bp->b_flags |= B_INVAL;
1129 if (LIST_FIRST(&bp->b_dep) != NULL)
1131 if (bp->b_flags & B_DELWRI) {
1133 dirtybufspace -= bp->b_bufsize;
1134 if (bp->b_flags & B_HEAVY) {
1136 dirtybufspacehw -= bp->b_bufsize;
1138 bd_signal(bp->b_bufsize);
1140 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1144 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1145 * If vfs_vmio_release() is called with either bit set, the
1146 * underlying pages may wind up getting freed causing a previous
1147 * write (bdwrite()) to get 'lost' because pages associated with
1148 * a B_DELWRI bp are marked clean. Pages associated with a
1149 * B_LOCKED buffer may be mapped by the filesystem.
1151 * If we want to release the buffer ourselves (rather then the
1152 * originator asking us to release it), give the originator a
1153 * chance to countermand the release by setting B_LOCKED.
1155 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1156 * if B_DELWRI is set.
1158 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1159 * on pages to return pages to the VM page queues.
1161 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1162 bp->b_flags &= ~B_RELBUF;
1163 } else if (vm_page_count_severe()) {
1164 if (LIST_FIRST(&bp->b_dep) != NULL)
1165 buf_deallocate(bp); /* can set B_LOCKED */
1166 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1167 bp->b_flags &= ~B_RELBUF;
1169 bp->b_flags |= B_RELBUF;
1173 * Make sure b_cmd is clear. It may have already been cleared by
1176 * At this point destroying the buffer is governed by the B_INVAL
1177 * or B_RELBUF flags.
1179 bp->b_cmd = BUF_CMD_DONE;
1182 * VMIO buffer rundown. Make sure the VM page array is restored
1183 * after an I/O may have replaces some of the pages with bogus pages
1184 * in order to not destroy dirty pages in a fill-in read.
1186 * Note that due to the code above, if a buffer is marked B_DELWRI
1187 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1188 * B_INVAL may still be set, however.
1190 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1191 * but not the backing store. B_NOCACHE will destroy the backing
1194 * Note that dirty NFS buffers contain byte-granular write ranges
1195 * and should not be destroyed w/ B_INVAL even if the backing store
1198 if (bp->b_flags & B_VMIO) {
1200 * Rundown for VMIO buffers which are not dirty NFS buffers.
1212 * Get the base offset and length of the buffer. Note that
1213 * in the VMIO case if the buffer block size is not
1214 * page-aligned then b_data pointer may not be page-aligned.
1215 * But our b_xio.xio_pages array *IS* page aligned.
1217 * block sizes less then DEV_BSIZE (usually 512) are not
1218 * supported due to the page granularity bits (m->valid,
1219 * m->dirty, etc...).
1221 * See man buf(9) for more information
1224 resid = bp->b_bufsize;
1225 foff = bp->b_loffset;
1227 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1228 m = bp->b_xio.xio_pages[i];
1229 vm_page_flag_clear(m, PG_ZERO);
1231 * If we hit a bogus page, fixup *all* of them
1232 * now. Note that we left these pages wired
1233 * when we removed them so they had better exist,
1234 * and they cannot be ripped out from under us so
1235 * no critical section protection is necessary.
1237 if (m == bogus_page) {
1239 poff = OFF_TO_IDX(bp->b_loffset);
1241 for (j = i; j < bp->b_xio.xio_npages; j++) {
1244 mtmp = bp->b_xio.xio_pages[j];
1245 if (mtmp == bogus_page) {
1246 mtmp = vm_page_lookup(obj, poff + j);
1248 panic("brelse: page missing");
1250 bp->b_xio.xio_pages[j] = mtmp;
1254 if ((bp->b_flags & B_INVAL) == 0) {
1255 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1256 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1258 m = bp->b_xio.xio_pages[i];
1262 * Invalidate the backing store if B_NOCACHE is set
1263 * (e.g. used with vinvalbuf()). If this is NFS
1264 * we impose a requirement that the block size be
1265 * a multiple of PAGE_SIZE and create a temporary
1266 * hack to basically invalidate the whole page. The
1267 * problem is that NFS uses really odd buffer sizes
1268 * especially when tracking piecemeal writes and
1269 * it also vinvalbuf()'s a lot, which would result
1270 * in only partial page validation and invalidation
1271 * here. If the file page is mmap()'d, however,
1272 * all the valid bits get set so after we invalidate
1273 * here we would end up with weird m->valid values
1274 * like 0xfc. nfs_getpages() can't handle this so
1275 * we clear all the valid bits for the NFS case
1276 * instead of just some of them.
1278 * The real bug is the VM system having to set m->valid
1279 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1280 * itself is an artifact of the whole 512-byte
1281 * granular mess that exists to support odd block
1282 * sizes and UFS meta-data block sizes (e.g. 6144).
1283 * A complete rewrite is required.
1285 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1286 int poffset = foff & PAGE_MASK;
1289 presid = PAGE_SIZE - poffset;
1290 if (bp->b_vp->v_tag == VT_NFS &&
1291 bp->b_vp->v_type == VREG) {
1293 } else if (presid > resid) {
1296 KASSERT(presid >= 0, ("brelse: extra page"));
1297 vm_page_set_invalid(m, poffset, presid);
1299 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1300 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1302 if (bp->b_flags & (B_INVAL | B_RELBUF))
1303 vfs_vmio_release(bp);
1306 * Rundown for non-VMIO buffers.
1308 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1311 kprintf("brelse bp %p %08x/%08x: Warning, caught and fixed brelvp bug\n", bp, saved_flags, bp->b_flags);
1315 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1321 if (bp->b_qindex != BQUEUE_NONE)
1322 panic("brelse: free buffer onto another queue???");
1323 if (BUF_REFCNTNB(bp) > 1) {
1324 /* Temporary panic to verify exclusive locking */
1325 /* This panic goes away when we allow shared refs */
1326 panic("brelse: multiple refs");
1327 /* do not release to free list */
1334 * Figure out the correct queue to place the cleaned up buffer on.
1335 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1336 * disassociated from their vnode.
1338 if (bp->b_flags & B_LOCKED) {
1340 * Buffers that are locked are placed in the locked queue
1341 * immediately, regardless of their state.
1343 bp->b_qindex = BQUEUE_LOCKED;
1344 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1345 } else if (bp->b_bufsize == 0) {
1347 * Buffers with no memory. Due to conditionals near the top
1348 * of brelse() such buffers should probably already be
1349 * marked B_INVAL and disassociated from their vnode.
1351 bp->b_flags |= B_INVAL;
1352 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1353 KKASSERT((bp->b_flags & B_HASHED) == 0);
1354 if (bp->b_kvasize) {
1355 bp->b_qindex = BQUEUE_EMPTYKVA;
1357 bp->b_qindex = BQUEUE_EMPTY;
1359 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1360 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1362 * Buffers with junk contents. Again these buffers had better
1363 * already be disassociated from their vnode.
1365 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1366 KKASSERT((bp->b_flags & B_HASHED) == 0);
1367 bp->b_flags |= B_INVAL;
1368 bp->b_qindex = BQUEUE_CLEAN;
1369 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1372 * Remaining buffers. These buffers are still associated with
1375 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1377 bp->b_qindex = BQUEUE_DIRTY;
1378 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1380 case B_DELWRI | B_HEAVY:
1381 bp->b_qindex = BQUEUE_DIRTY_HW;
1382 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1387 * NOTE: Buffers are always placed at the end of the
1388 * queue. If B_AGE is not set the buffer will cycle
1389 * through the queue twice.
1391 bp->b_qindex = BQUEUE_CLEAN;
1392 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1398 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1399 * on the correct queue.
1401 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1405 * The bp is on an appropriate queue unless locked. If it is not
1406 * locked or dirty we can wakeup threads waiting for buffer space.
1408 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1409 * if B_INVAL is set ).
1411 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1415 * Something we can maybe free or reuse
1417 if (bp->b_bufsize || bp->b_kvasize)
1421 * Clean up temporary flags and unlock the buffer.
1423 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
1431 * Release a buffer back to the appropriate queue but do not try to free
1432 * it. The buffer is expected to be used again soon.
1434 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1435 * biodone() to requeue an async I/O on completion. It is also used when
1436 * known good buffers need to be requeued but we think we may need the data
1439 * XXX we should be able to leave the B_RELBUF hint set on completion.
1442 bqrelse(struct buf *bp)
1446 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1448 if (bp->b_qindex != BQUEUE_NONE)
1449 panic("bqrelse: free buffer onto another queue???");
1450 if (BUF_REFCNTNB(bp) > 1) {
1451 /* do not release to free list */
1452 panic("bqrelse: multiple refs");
1457 if (bp->b_flags & B_LOCKED) {
1459 * Locked buffers are released to the locked queue. However,
1460 * if the buffer is dirty it will first go into the dirty
1461 * queue and later on after the I/O completes successfully it
1462 * will be released to the locked queue.
1464 bp->b_qindex = BQUEUE_LOCKED;
1465 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1466 } else if (bp->b_flags & B_DELWRI) {
1467 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1468 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1469 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1470 } else if (vm_page_count_severe()) {
1472 * We are too low on memory, we have to try to free the
1473 * buffer (most importantly: the wired pages making up its
1474 * backing store) *now*.
1480 bp->b_qindex = BQUEUE_CLEAN;
1481 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1484 if ((bp->b_flags & B_LOCKED) == 0 &&
1485 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1490 * Something we can maybe free or reuse.
1492 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1496 * Final cleanup and unlock. Clear bits that are only used while a
1497 * buffer is actively locked.
1499 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF);
1507 * Return backing pages held by the buffer 'bp' back to the VM system
1508 * if possible. The pages are freed if they are no longer valid or
1509 * attempt to free if it was used for direct I/O otherwise they are
1510 * sent to the page cache.
1512 * Pages that were marked busy are left alone and skipped.
1514 * The KVA mapping (b_data) for the underlying pages is removed by
1518 vfs_vmio_release(struct buf *bp)
1524 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1525 m = bp->b_xio.xio_pages[i];
1526 bp->b_xio.xio_pages[i] = NULL;
1528 * In order to keep page LRU ordering consistent, put
1529 * everything on the inactive queue.
1531 vm_page_unwire(m, 0);
1533 * We don't mess with busy pages, it is
1534 * the responsibility of the process that
1535 * busied the pages to deal with them.
1537 if ((m->flags & PG_BUSY) || (m->busy != 0))
1540 if (m->wire_count == 0) {
1541 vm_page_flag_clear(m, PG_ZERO);
1543 * Might as well free the page if we can and it has
1544 * no valid data. We also free the page if the
1545 * buffer was used for direct I/O.
1547 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1548 m->hold_count == 0) {
1550 vm_page_protect(m, VM_PROT_NONE);
1552 } else if (bp->b_flags & B_DIRECT) {
1553 vm_page_try_to_free(m);
1554 } else if (vm_page_count_severe()) {
1555 vm_page_try_to_cache(m);
1560 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1561 if (bp->b_bufsize) {
1565 bp->b_xio.xio_npages = 0;
1566 bp->b_flags &= ~B_VMIO;
1567 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1575 * Implement clustered async writes for clearing out B_DELWRI buffers.
1576 * This is much better then the old way of writing only one buffer at
1577 * a time. Note that we may not be presented with the buffers in the
1578 * correct order, so we search for the cluster in both directions.
1580 * The buffer is locked on call.
1583 vfs_bio_awrite(struct buf *bp)
1587 off_t loffset = bp->b_loffset;
1588 struct vnode *vp = bp->b_vp;
1596 * right now we support clustered writing only to regular files. If
1597 * we find a clusterable block we could be in the middle of a cluster
1598 * rather then at the beginning.
1600 * NOTE: b_bio1 contains the logical loffset and is aliased
1601 * to b_loffset. b_bio2 contains the translated block number.
1603 if ((vp->v_type == VREG) &&
1604 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1605 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1607 size = vp->v_mount->mnt_stat.f_iosize;
1609 for (i = size; i < MAXPHYS; i += size) {
1610 if ((bpa = findblk(vp, loffset + i)) &&
1611 BUF_REFCNT(bpa) == 0 &&
1612 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1613 (B_DELWRI | B_CLUSTEROK)) &&
1614 (bpa->b_bufsize == size)) {
1615 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1616 (bpa->b_bio2.bio_offset !=
1617 bp->b_bio2.bio_offset + i))
1623 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1624 if ((bpa = findblk(vp, loffset - j)) &&
1625 BUF_REFCNT(bpa) == 0 &&
1626 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1627 (B_DELWRI | B_CLUSTEROK)) &&
1628 (bpa->b_bufsize == size)) {
1629 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1630 (bpa->b_bio2.bio_offset !=
1631 bp->b_bio2.bio_offset - j))
1640 * this is a possible cluster write
1642 if (nbytes != size) {
1644 nwritten = cluster_wbuild(vp, size,
1645 loffset - j, nbytes);
1652 bp->b_flags |= B_ASYNC;
1656 * default (old) behavior, writing out only one block
1658 * XXX returns b_bufsize instead of b_bcount for nwritten?
1660 nwritten = bp->b_bufsize;
1669 * Find and initialize a new buffer header, freeing up existing buffers
1670 * in the bufqueues as necessary. The new buffer is returned locked.
1672 * Important: B_INVAL is not set. If the caller wishes to throw the
1673 * buffer away, the caller must set B_INVAL prior to calling brelse().
1676 * We have insufficient buffer headers
1677 * We have insufficient buffer space
1678 * buffer_map is too fragmented ( space reservation fails )
1679 * If we have to flush dirty buffers ( but we try to avoid this )
1681 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1682 * Instead we ask the buf daemon to do it for us. We attempt to
1683 * avoid piecemeal wakeups of the pageout daemon.
1687 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1693 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1694 static int flushingbufs;
1697 * We can't afford to block since we might be holding a vnode lock,
1698 * which may prevent system daemons from running. We deal with
1699 * low-memory situations by proactively returning memory and running
1700 * async I/O rather then sync I/O.
1704 --getnewbufrestarts;
1706 ++getnewbufrestarts;
1709 * Setup for scan. If we do not have enough free buffers,
1710 * we setup a degenerate case that immediately fails. Note
1711 * that if we are specially marked process, we are allowed to
1712 * dip into our reserves.
1714 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1716 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1717 * However, there are a number of cases (defragging, reusing, ...)
1718 * where we cannot backup.
1720 nqindex = BQUEUE_EMPTYKVA;
1721 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1725 * If no EMPTYKVA buffers and we are either
1726 * defragging or reusing, locate a CLEAN buffer
1727 * to free or reuse. If bufspace useage is low
1728 * skip this step so we can allocate a new buffer.
1730 if (defrag || bufspace >= lobufspace) {
1731 nqindex = BQUEUE_CLEAN;
1732 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1736 * If we could not find or were not allowed to reuse a
1737 * CLEAN buffer, check to see if it is ok to use an EMPTY
1738 * buffer. We can only use an EMPTY buffer if allocating
1739 * its KVA would not otherwise run us out of buffer space.
1741 if (nbp == NULL && defrag == 0 &&
1742 bufspace + maxsize < hibufspace) {
1743 nqindex = BQUEUE_EMPTY;
1744 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1749 * Run scan, possibly freeing data and/or kva mappings on the fly
1753 while ((bp = nbp) != NULL) {
1754 int qindex = nqindex;
1756 nbp = TAILQ_NEXT(bp, b_freelist);
1759 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1760 * cycles through the queue twice before being selected.
1762 if (qindex == BQUEUE_CLEAN &&
1763 (bp->b_flags & B_AGE) == 0 && nbp) {
1764 bp->b_flags |= B_AGE;
1765 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1766 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1771 * Calculate next bp ( we can only use it if we do not block
1772 * or do other fancy things ).
1777 nqindex = BQUEUE_EMPTYKVA;
1778 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1781 case BQUEUE_EMPTYKVA:
1782 nqindex = BQUEUE_CLEAN;
1783 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1797 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1800 * Note: we no longer distinguish between VMIO and non-VMIO
1804 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1807 * If we are defragging then we need a buffer with
1808 * b_kvasize != 0. XXX this situation should no longer
1809 * occur, if defrag is non-zero the buffer's b_kvasize
1810 * should also be non-zero at this point. XXX
1812 if (defrag && bp->b_kvasize == 0) {
1813 kprintf("Warning: defrag empty buffer %p\n", bp);
1818 * Start freeing the bp. This is somewhat involved. nbp
1819 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1820 * on the clean list must be disassociated from their
1821 * current vnode. Buffers on the empty[kva] lists have
1822 * already been disassociated.
1825 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1826 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1827 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1830 if (bp->b_qindex != qindex) {
1831 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1838 * Dependancies must be handled before we disassociate the
1841 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1842 * be immediately disassociated. HAMMER then becomes
1843 * responsible for releasing the buffer.
1845 if (LIST_FIRST(&bp->b_dep) != NULL) {
1847 if (bp->b_flags & B_LOCKED) {
1851 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1854 if (qindex == BQUEUE_CLEAN) {
1855 if (bp->b_flags & B_VMIO) {
1856 bp->b_flags &= ~B_ASYNC;
1857 vfs_vmio_release(bp);
1864 * NOTE: nbp is now entirely invalid. We can only restart
1865 * the scan from this point on.
1867 * Get the rest of the buffer freed up. b_kva* is still
1868 * valid after this operation.
1871 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1872 KKASSERT((bp->b_flags & B_HASHED) == 0);
1875 * critical section protection is not required when
1876 * scrapping a buffer's contents because it is already
1882 bp->b_flags = B_BNOCLIP;
1883 bp->b_cmd = BUF_CMD_DONE;
1888 bp->b_xio.xio_npages = 0;
1889 bp->b_dirtyoff = bp->b_dirtyend = 0;
1891 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1893 if (blkflags & GETBLK_BHEAVY)
1894 bp->b_flags |= B_HEAVY;
1897 * If we are defragging then free the buffer.
1900 bp->b_flags |= B_INVAL;
1908 * If we are overcomitted then recover the buffer and its
1909 * KVM space. This occurs in rare situations when multiple
1910 * processes are blocked in getnewbuf() or allocbuf().
1912 if (bufspace >= hibufspace)
1914 if (flushingbufs && bp->b_kvasize != 0) {
1915 bp->b_flags |= B_INVAL;
1920 if (bufspace < lobufspace)
1926 * If we exhausted our list, sleep as appropriate. We may have to
1927 * wakeup various daemons and write out some dirty buffers.
1929 * Generally we are sleeping due to insufficient buffer space.
1937 flags = VFS_BIO_NEED_BUFSPACE;
1939 } else if (bufspace >= hibufspace) {
1941 flags = VFS_BIO_NEED_BUFSPACE;
1944 flags = VFS_BIO_NEED_ANY;
1947 needsbuffer |= flags;
1948 bd_speedup(); /* heeeelp */
1949 while (needsbuffer & flags) {
1950 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
1955 * We finally have a valid bp. We aren't quite out of the
1956 * woods, we still have to reserve kva space. In order
1957 * to keep fragmentation sane we only allocate kva in
1960 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1962 if (maxsize != bp->b_kvasize) {
1963 vm_offset_t addr = 0;
1968 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1969 vm_map_lock(&buffer_map);
1971 if (vm_map_findspace(&buffer_map,
1972 vm_map_min(&buffer_map), maxsize,
1973 maxsize, 0, &addr)) {
1975 * Uh oh. Buffer map is too fragmented. We
1976 * must defragment the map.
1978 vm_map_unlock(&buffer_map);
1979 vm_map_entry_release(count);
1982 bp->b_flags |= B_INVAL;
1987 vm_map_insert(&buffer_map, &count,
1989 addr, addr + maxsize,
1991 VM_PROT_ALL, VM_PROT_ALL,
1994 bp->b_kvabase = (caddr_t) addr;
1995 bp->b_kvasize = maxsize;
1996 bufspace += bp->b_kvasize;
1999 vm_map_unlock(&buffer_map);
2000 vm_map_entry_release(count);
2002 bp->b_data = bp->b_kvabase;
2008 * This routine is called in an emergency to recover VM pages from the
2009 * buffer cache by cashing in clean buffers. The idea is to recover
2010 * enough pages to be able to satisfy a stuck bio_page_alloc().
2013 recoverbufpages(void)
2020 while (bytes < MAXBSIZE) {
2021 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2026 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2027 * cycles through the queue twice before being selected.
2029 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
2030 bp->b_flags |= B_AGE;
2031 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
2032 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
2040 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
2041 KKASSERT((bp->b_flags & B_DELWRI) == 0);
2044 * Start freeing the bp. This is somewhat involved.
2046 * Buffers on the clean list must be disassociated from
2047 * their current vnode
2050 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2051 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp);
2052 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
2055 if (bp->b_qindex != BQUEUE_CLEAN) {
2056 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp, bp->b_qindex);
2063 * Dependancies must be handled before we disassociate the
2066 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2067 * be immediately disassociated. HAMMER then becomes
2068 * responsible for releasing the buffer.
2070 if (LIST_FIRST(&bp->b_dep) != NULL) {
2072 if (bp->b_flags & B_LOCKED) {
2076 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2079 bytes += bp->b_bufsize;
2081 if (bp->b_flags & B_VMIO) {
2082 bp->b_flags &= ~B_ASYNC;
2083 bp->b_flags |= B_DIRECT; /* try to free pages */
2084 vfs_vmio_release(bp);
2089 KKASSERT(bp->b_vp == NULL);
2090 KKASSERT((bp->b_flags & B_HASHED) == 0);
2093 * critical section protection is not required when
2094 * scrapping a buffer's contents because it is already
2100 bp->b_flags = B_BNOCLIP;
2101 bp->b_cmd = BUF_CMD_DONE;
2106 bp->b_xio.xio_npages = 0;
2107 bp->b_dirtyoff = bp->b_dirtyend = 0;
2109 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2111 bp->b_flags |= B_INVAL;
2121 * Buffer flushing daemon. Buffers are normally flushed by the
2122 * update daemon but if it cannot keep up this process starts to
2123 * take the load in an attempt to prevent getnewbuf() from blocking.
2125 * Once a flush is initiated it does not stop until the number
2126 * of buffers falls below lodirtybuffers, but we will wake up anyone
2127 * waiting at the mid-point.
2130 static struct kproc_desc buf_kp = {
2135 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2136 kproc_start, &buf_kp)
2138 static struct kproc_desc bufhw_kp = {
2143 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2144 kproc_start, &bufhw_kp)
2152 * This process needs to be suspended prior to shutdown sync.
2154 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2155 bufdaemon_td, SHUTDOWN_PRI_LAST);
2156 curthread->td_flags |= TDF_SYSTHREAD;
2159 * This process is allowed to take the buffer cache to the limit
2164 kproc_suspend_loop();
2167 * Do the flush. Limit the amount of in-transit I/O we
2168 * allow to build up, otherwise we would completely saturate
2169 * the I/O system. Wakeup any waiting processes before we
2170 * normally would so they can run in parallel with our drain.
2172 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2173 * but because we split the operation into two threads we
2174 * have to cut it in half for each thread.
2176 limit = lodirtybufspace / 2;
2177 waitrunningbufspace(limit);
2178 while (runningbufspace + dirtybufspace > limit ||
2179 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
2180 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2182 waitrunningbufspace(limit);
2186 * We reached our low water mark, reset the
2187 * request and sleep until we are needed again.
2188 * The sleep is just so the suspend code works.
2190 spin_lock_wr(&needsbuffer_spin);
2191 if (bd_request == 0) {
2192 msleep(&bd_request, &needsbuffer_spin, 0,
2196 spin_unlock_wr(&needsbuffer_spin);
2206 * This process needs to be suspended prior to shutdown sync.
2208 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2209 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2210 curthread->td_flags |= TDF_SYSTHREAD;
2213 * This process is allowed to take the buffer cache to the limit
2218 kproc_suspend_loop();
2221 * Do the flush. Limit the amount of in-transit I/O we
2222 * allow to build up, otherwise we would completely saturate
2223 * the I/O system. Wakeup any waiting processes before we
2224 * normally would so they can run in parallel with our drain.
2226 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2227 * but because we split the operation into two threads we
2228 * have to cut it in half for each thread.
2230 limit = lodirtybufspace / 2;
2231 waitrunningbufspace(limit);
2232 while (runningbufspace + dirtybufspacehw > limit ||
2233 dirtybufcounthw >= nbuf / 2) {
2234 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2236 waitrunningbufspace(limit);
2240 * We reached our low water mark, reset the
2241 * request and sleep until we are needed again.
2242 * The sleep is just so the suspend code works.
2244 spin_lock_wr(&needsbuffer_spin);
2245 if (bd_request_hw == 0) {
2246 msleep(&bd_request_hw, &needsbuffer_spin, 0,
2250 spin_unlock_wr(&needsbuffer_spin);
2257 * Try to flush a buffer in the dirty queue. We must be careful to
2258 * free up B_INVAL buffers instead of write them, which NFS is
2259 * particularly sensitive to.
2261 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2262 * that we really want to try to get the buffer out and reuse it
2263 * due to the write load on the machine.
2267 flushbufqueues(bufq_type_t q)
2272 bp = TAILQ_FIRST(&bufqueues[q]);
2274 KASSERT((bp->b_flags & B_DELWRI),
2275 ("unexpected clean buffer %p", bp));
2277 if (bp->b_flags & B_DELWRI) {
2278 if (bp->b_flags & B_INVAL) {
2279 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2280 panic("flushbufqueues: locked buf");
2286 if (LIST_FIRST(&bp->b_dep) != NULL &&
2287 (bp->b_flags & B_DEFERRED) == 0 &&
2288 buf_countdeps(bp, 0)) {
2289 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2290 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2292 bp->b_flags |= B_DEFERRED;
2293 bp = TAILQ_FIRST(&bufqueues[q]);
2298 * Only write it out if we can successfully lock
2299 * it. If the buffer has a dependancy,
2300 * buf_checkwrite must also return 0 for us to
2301 * be able to initate the write.
2303 * If the buffer is flagged B_ERROR it may be
2304 * requeued over and over again, we try to
2305 * avoid a live lock.
2307 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
2308 if (LIST_FIRST(&bp->b_dep) != NULL &&
2309 buf_checkwrite(bp)) {
2312 } else if (bp->b_flags & B_ERROR) {
2313 tsleep(bp, 0, "bioer", 1);
2314 bp->b_flags &= ~B_AGE;
2317 bp->b_flags |= B_AGE;
2324 bp = TAILQ_NEXT(bp, b_freelist);
2332 * Returns true if no I/O is needed to access the associated VM object.
2333 * This is like findblk except it also hunts around in the VM system for
2336 * Note that we ignore vm_page_free() races from interrupts against our
2337 * lookup, since if the caller is not protected our return value will not
2338 * be any more valid then otherwise once we exit the critical section.
2341 inmem(struct vnode *vp, off_t loffset)
2344 vm_offset_t toff, tinc, size;
2347 if (findblk(vp, loffset))
2349 if (vp->v_mount == NULL)
2351 if ((obj = vp->v_object) == NULL)
2355 if (size > vp->v_mount->mnt_stat.f_iosize)
2356 size = vp->v_mount->mnt_stat.f_iosize;
2358 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2359 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2363 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2364 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2365 if (vm_page_is_valid(m,
2366 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2375 * Sets the dirty range for a buffer based on the status of the dirty
2376 * bits in the pages comprising the buffer.
2378 * The range is limited to the size of the buffer.
2380 * This routine is primarily used by NFS, but is generalized for the
2384 vfs_setdirty(struct buf *bp)
2390 * Degenerate case - empty buffer
2393 if (bp->b_bufsize == 0)
2397 * We qualify the scan for modified pages on whether the
2398 * object has been flushed yet. The OBJ_WRITEABLE flag
2399 * is not cleared simply by protecting pages off.
2402 if ((bp->b_flags & B_VMIO) == 0)
2405 object = bp->b_xio.xio_pages[0]->object;
2407 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2408 kprintf("Warning: object %p writeable but not mightbedirty\n", object);
2409 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2410 kprintf("Warning: object %p mightbedirty but not writeable\n", object);
2412 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2413 vm_offset_t boffset;
2414 vm_offset_t eoffset;
2417 * test the pages to see if they have been modified directly
2418 * by users through the VM system.
2420 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2421 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2422 vm_page_test_dirty(bp->b_xio.xio_pages[i]);
2426 * Calculate the encompassing dirty range, boffset and eoffset,
2427 * (eoffset - boffset) bytes.
2430 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2431 if (bp->b_xio.xio_pages[i]->dirty)
2434 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2436 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2437 if (bp->b_xio.xio_pages[i]->dirty) {
2441 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2444 * Fit it to the buffer.
2447 if (eoffset > bp->b_bcount)
2448 eoffset = bp->b_bcount;
2451 * If we have a good dirty range, merge with the existing
2455 if (boffset < eoffset) {
2456 if (bp->b_dirtyoff > boffset)
2457 bp->b_dirtyoff = boffset;
2458 if (bp->b_dirtyend < eoffset)
2459 bp->b_dirtyend = eoffset;
2467 * Locate and return the specified buffer, or NULL if the buffer does
2468 * not exist. Do not attempt to lock the buffer or manipulate it in
2469 * any way. The caller must validate that the correct buffer has been
2470 * obtain after locking it.
2475 findblk(struct vnode *vp, off_t loffset)
2480 lwkt_gettoken(&vlock, &vp->v_token);
2481 /* ASSERT_LWKT_TOKEN_HELD(&vp->v_token);*/
2482 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2483 lwkt_reltoken(&vlock);
2490 * Get a block given a specified block and offset into a file/device.
2491 * B_INVAL may or may not be set on return. The caller should clear
2492 * B_INVAL prior to initiating a READ.
2494 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2495 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2496 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2497 * without doing any of those things the system will likely believe
2498 * the buffer to be valid (especially if it is not B_VMIO), and the
2499 * next getblk() will return the buffer with B_CACHE set.
2501 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2502 * an existing buffer.
2504 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2505 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2506 * and then cleared based on the backing VM. If the previous buffer is
2507 * non-0-sized but invalid, B_CACHE will be cleared.
2509 * If getblk() must create a new buffer, the new buffer is returned with
2510 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2511 * case it is returned with B_INVAL clear and B_CACHE set based on the
2514 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2515 * B_CACHE bit is clear.
2517 * What this means, basically, is that the caller should use B_CACHE to
2518 * determine whether the buffer is fully valid or not and should clear
2519 * B_INVAL prior to issuing a read. If the caller intends to validate
2520 * the buffer by loading its data area with something, the caller needs
2521 * to clear B_INVAL. If the caller does this without issuing an I/O,
2522 * the caller should set B_CACHE ( as an optimization ), else the caller
2523 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2524 * a write attempt or if it was a successfull read. If the caller
2525 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2526 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2530 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2531 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2534 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2537 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2540 if (size > MAXBSIZE)
2541 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2542 if (vp->v_object == NULL)
2543 panic("getblk: vnode %p has no object!", vp);
2547 if ((bp = findblk(vp, loffset))) {
2549 * The buffer was found in the cache, but we need to lock it.
2550 * Even with LK_NOWAIT the lockmgr may break our critical
2551 * section, so double-check the validity of the buffer
2552 * once the lock has been obtained.
2554 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2555 if (blkflags & GETBLK_NOWAIT) {
2559 int lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2560 if (blkflags & GETBLK_PCATCH)
2561 lkflags |= LK_PCATCH;
2562 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2564 if (error == ENOLCK)
2572 * Once the buffer has been locked, make sure we didn't race
2573 * a buffer recyclement. Buffers that are no longer hashed
2574 * will have b_vp == NULL, so this takes care of that check
2577 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2578 kprintf("Warning buffer %p (vp %p loffset %lld) "
2580 bp, vp, (long long)loffset);
2586 * If SZMATCH any pre-existing buffer must be of the requested
2587 * size or NULL is returned. The caller absolutely does not
2588 * want getblk() to bwrite() the buffer on a size mismatch.
2590 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2597 * All vnode-based buffers must be backed by a VM object.
2599 KKASSERT(bp->b_flags & B_VMIO);
2600 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2601 bp->b_flags &= ~B_AGE;
2604 * Make sure that B_INVAL buffers do not have a cached
2605 * block number translation.
2607 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2608 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2609 " did not have cleared bio_offset cache\n",
2610 bp, vp, (long long)loffset);
2611 clearbiocache(&bp->b_bio2);
2615 * The buffer is locked. B_CACHE is cleared if the buffer is
2618 if (bp->b_flags & B_INVAL)
2619 bp->b_flags &= ~B_CACHE;
2623 * Any size inconsistancy with a dirty buffer or a buffer
2624 * with a softupdates dependancy must be resolved. Resizing
2625 * the buffer in such circumstances can lead to problems.
2627 if (size != bp->b_bcount) {
2628 if (bp->b_flags & B_DELWRI) {
2629 bp->b_flags |= B_NOCACHE;
2631 } else if (LIST_FIRST(&bp->b_dep)) {
2632 bp->b_flags |= B_NOCACHE;
2635 bp->b_flags |= B_RELBUF;
2640 KKASSERT(size <= bp->b_kvasize);
2641 KASSERT(bp->b_loffset != NOOFFSET,
2642 ("getblk: no buffer offset"));
2645 * A buffer with B_DELWRI set and B_CACHE clear must
2646 * be committed before we can return the buffer in
2647 * order to prevent the caller from issuing a read
2648 * ( due to B_CACHE not being set ) and overwriting
2651 * Most callers, including NFS and FFS, need this to
2652 * operate properly either because they assume they
2653 * can issue a read if B_CACHE is not set, or because
2654 * ( for example ) an uncached B_DELWRI might loop due
2655 * to softupdates re-dirtying the buffer. In the latter
2656 * case, B_CACHE is set after the first write completes,
2657 * preventing further loops.
2659 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2660 * above while extending the buffer, we cannot allow the
2661 * buffer to remain with B_CACHE set after the write
2662 * completes or it will represent a corrupt state. To
2663 * deal with this we set B_NOCACHE to scrap the buffer
2666 * We might be able to do something fancy, like setting
2667 * B_CACHE in bwrite() except if B_DELWRI is already set,
2668 * so the below call doesn't set B_CACHE, but that gets real
2669 * confusing. This is much easier.
2672 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2673 bp->b_flags |= B_NOCACHE;
2680 * Buffer is not in-core, create new buffer. The buffer
2681 * returned by getnewbuf() is locked. Note that the returned
2682 * buffer is also considered valid (not marked B_INVAL).
2684 * Calculating the offset for the I/O requires figuring out
2685 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2686 * the mount's f_iosize otherwise. If the vnode does not
2687 * have an associated mount we assume that the passed size is
2690 * Note that vn_isdisk() cannot be used here since it may
2691 * return a failure for numerous reasons. Note that the
2692 * buffer size may be larger then the block size (the caller
2693 * will use block numbers with the proper multiple). Beware
2694 * of using any v_* fields which are part of unions. In
2695 * particular, in DragonFly the mount point overloading
2696 * mechanism uses the namecache only and the underlying
2697 * directory vnode is not a special case.
2701 if (vp->v_type == VBLK || vp->v_type == VCHR)
2703 else if (vp->v_mount)
2704 bsize = vp->v_mount->mnt_stat.f_iosize;
2708 maxsize = size + (loffset & PAGE_MASK);
2709 maxsize = imax(maxsize, bsize);
2711 if ((bp = getnewbuf(blkflags, slptimeo, size, maxsize)) == NULL) {
2712 if (slpflags || slptimeo) {
2720 * This code is used to make sure that a buffer is not
2721 * created while the getnewbuf routine is blocked.
2722 * This can be a problem whether the vnode is locked or not.
2723 * If the buffer is created out from under us, we have to
2724 * throw away the one we just created. There is no window
2725 * race because we are safely running in a critical section
2726 * from the point of the duplicate buffer creation through
2727 * to here, and we've locked the buffer.
2729 if (findblk(vp, loffset)) {
2730 bp->b_flags |= B_INVAL;
2736 * Insert the buffer into the hash, so that it can
2737 * be found by findblk().
2739 * Make sure the translation layer has been cleared.
2741 bp->b_loffset = loffset;
2742 bp->b_bio2.bio_offset = NOOFFSET;
2743 /* bp->b_bio2.bio_next = NULL; */
2748 * All vnode-based buffers must be backed by a VM object.
2750 KKASSERT(vp->v_object != NULL);
2751 bp->b_flags |= B_VMIO;
2752 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2764 * Reacquire a buffer that was previously released to the locked queue,
2765 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2766 * set B_LOCKED (which handles the acquisition race).
2768 * To this end, either B_LOCKED must be set or the dependancy list must be
2772 regetblk(struct buf *bp)
2774 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
2775 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2784 * Get an empty, disassociated buffer of given size. The buffer is
2785 * initially set to B_INVAL.
2787 * critical section protection is not required for the allocbuf()
2788 * call because races are impossible here.
2796 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2799 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2803 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2811 * This code constitutes the buffer memory from either anonymous system
2812 * memory (in the case of non-VMIO operations) or from an associated
2813 * VM object (in the case of VMIO operations). This code is able to
2814 * resize a buffer up or down.
2816 * Note that this code is tricky, and has many complications to resolve
2817 * deadlock or inconsistant data situations. Tread lightly!!!
2818 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2819 * the caller. Calling this code willy nilly can result in the loss of data.
2821 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2822 * B_CACHE for the non-VMIO case.
2824 * This routine does not need to be called from a critical section but you
2825 * must own the buffer.
2828 allocbuf(struct buf *bp, int size)
2830 int newbsize, mbsize;
2833 if (BUF_REFCNT(bp) == 0)
2834 panic("allocbuf: buffer not busy");
2836 if (bp->b_kvasize < size)
2837 panic("allocbuf: buffer too small");
2839 if ((bp->b_flags & B_VMIO) == 0) {
2843 * Just get anonymous memory from the kernel. Don't
2844 * mess with B_CACHE.
2846 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2847 if (bp->b_flags & B_MALLOC)
2850 newbsize = round_page(size);
2852 if (newbsize < bp->b_bufsize) {
2854 * Malloced buffers are not shrunk
2856 if (bp->b_flags & B_MALLOC) {
2858 bp->b_bcount = size;
2860 kfree(bp->b_data, M_BIOBUF);
2861 if (bp->b_bufsize) {
2862 bufmallocspace -= bp->b_bufsize;
2866 bp->b_data = bp->b_kvabase;
2868 bp->b_flags &= ~B_MALLOC;
2874 (vm_offset_t) bp->b_data + newbsize,
2875 (vm_offset_t) bp->b_data + bp->b_bufsize);
2876 } else if (newbsize > bp->b_bufsize) {
2878 * We only use malloced memory on the first allocation.
2879 * and revert to page-allocated memory when the buffer
2882 if ((bufmallocspace < maxbufmallocspace) &&
2883 (bp->b_bufsize == 0) &&
2884 (mbsize <= PAGE_SIZE/2)) {
2886 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
2887 bp->b_bufsize = mbsize;
2888 bp->b_bcount = size;
2889 bp->b_flags |= B_MALLOC;
2890 bufmallocspace += mbsize;
2896 * If the buffer is growing on its other-than-first
2897 * allocation, then we revert to the page-allocation
2900 if (bp->b_flags & B_MALLOC) {
2901 origbuf = bp->b_data;
2902 origbufsize = bp->b_bufsize;
2903 bp->b_data = bp->b_kvabase;
2904 if (bp->b_bufsize) {
2905 bufmallocspace -= bp->b_bufsize;
2909 bp->b_flags &= ~B_MALLOC;
2910 newbsize = round_page(newbsize);
2914 (vm_offset_t) bp->b_data + bp->b_bufsize,
2915 (vm_offset_t) bp->b_data + newbsize);
2917 bcopy(origbuf, bp->b_data, origbufsize);
2918 kfree(origbuf, M_BIOBUF);
2925 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2926 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
2927 newbsize + PAGE_MASK) >> PAGE_SHIFT;
2928 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
2930 if (bp->b_flags & B_MALLOC)
2931 panic("allocbuf: VMIO buffer can't be malloced");
2933 * Set B_CACHE initially if buffer is 0 length or will become
2936 if (size == 0 || bp->b_bufsize == 0)
2937 bp->b_flags |= B_CACHE;
2939 if (newbsize < bp->b_bufsize) {
2941 * DEV_BSIZE aligned new buffer size is less then the
2942 * DEV_BSIZE aligned existing buffer size. Figure out
2943 * if we have to remove any pages.
2945 if (desiredpages < bp->b_xio.xio_npages) {
2946 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
2948 * the page is not freed here -- it
2949 * is the responsibility of
2950 * vnode_pager_setsize
2952 m = bp->b_xio.xio_pages[i];
2953 KASSERT(m != bogus_page,
2954 ("allocbuf: bogus page found"));
2955 while (vm_page_sleep_busy(m, TRUE, "biodep"))
2958 bp->b_xio.xio_pages[i] = NULL;
2959 vm_page_unwire(m, 0);
2961 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2962 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
2963 bp->b_xio.xio_npages = desiredpages;
2965 } else if (size > bp->b_bcount) {
2967 * We are growing the buffer, possibly in a
2968 * byte-granular fashion.
2976 * Step 1, bring in the VM pages from the object,
2977 * allocating them if necessary. We must clear
2978 * B_CACHE if these pages are not valid for the
2979 * range covered by the buffer.
2981 * critical section protection is required to protect
2982 * against interrupts unbusying and freeing pages
2983 * between our vm_page_lookup() and our
2984 * busycheck/wiring call.
2990 while (bp->b_xio.xio_npages < desiredpages) {
2994 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
2995 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2997 * note: must allocate system pages
2998 * since blocking here could intefere
2999 * with paging I/O, no matter which
3002 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
3006 bp->b_flags &= ~B_CACHE;
3007 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3008 ++bp->b_xio.xio_npages;
3014 * We found a page. If we have to sleep on it,
3015 * retry because it might have gotten freed out
3018 * We can only test PG_BUSY here. Blocking on
3019 * m->busy might lead to a deadlock:
3021 * vm_fault->getpages->cluster_read->allocbuf
3025 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
3027 vm_page_flag_clear(m, PG_ZERO);
3029 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3030 ++bp->b_xio.xio_npages;
3035 * Step 2. We've loaded the pages into the buffer,
3036 * we have to figure out if we can still have B_CACHE
3037 * set. Note that B_CACHE is set according to the
3038 * byte-granular range ( bcount and size ), not the
3039 * aligned range ( newbsize ).
3041 * The VM test is against m->valid, which is DEV_BSIZE
3042 * aligned. Needless to say, the validity of the data
3043 * needs to also be DEV_BSIZE aligned. Note that this
3044 * fails with NFS if the server or some other client
3045 * extends the file's EOF. If our buffer is resized,
3046 * B_CACHE may remain set! XXX
3049 toff = bp->b_bcount;
3050 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3052 while ((bp->b_flags & B_CACHE) && toff < size) {
3055 if (tinc > (size - toff))
3058 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3066 bp->b_xio.xio_pages[pi]
3073 * Step 3, fixup the KVM pmap. Remember that
3074 * bp->b_data is relative to bp->b_loffset, but
3075 * bp->b_loffset may be offset into the first page.
3078 bp->b_data = (caddr_t)
3079 trunc_page((vm_offset_t)bp->b_data);
3081 (vm_offset_t)bp->b_data,
3082 bp->b_xio.xio_pages,
3083 bp->b_xio.xio_npages
3085 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3086 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3090 /* adjust space use on already-dirty buffer */
3091 if (bp->b_flags & B_DELWRI) {
3092 dirtybufspace += newbsize - bp->b_bufsize;
3093 if (bp->b_flags & B_HEAVY)
3094 dirtybufspacehw += newbsize - bp->b_bufsize;
3096 if (newbsize < bp->b_bufsize)
3098 bp->b_bufsize = newbsize; /* actual buffer allocation */
3099 bp->b_bcount = size; /* requested buffer size */
3106 * Wait for buffer I/O completion, returning error status. The buffer
3107 * is left locked on return. B_EINTR is converted into an EINTR error
3110 * NOTE! The original b_cmd is lost on return, since b_cmd will be
3111 * set to BUF_CMD_DONE.
3116 biowait(struct buf *bp)
3118 if (bp->b_cmd != BUF_CMD_DONE) {
3121 tsleep_interlock(bp);
3122 if (bp->b_cmd == BUF_CMD_DONE)
3124 if (bp->b_cmd == BUF_CMD_READ)
3125 tsleep(bp, 0, "biord", 0);
3127 tsleep(bp, 0, "biowr", 0);
3131 if (bp->b_flags & B_EINTR) {
3132 bp->b_flags &= ~B_EINTR;
3135 if (bp->b_flags & B_ERROR) {
3136 return (bp->b_error ? bp->b_error : EIO);
3143 * This associates a tracking count with an I/O. vn_strategy() and
3144 * dev_dstrategy() do this automatically but there are a few cases
3145 * where a vnode or device layer is bypassed when a block translation
3146 * is cached. In such cases bio_start_transaction() may be called on
3147 * the bypassed layers so the system gets an I/O in progress indication
3148 * for those higher layers.
3151 bio_start_transaction(struct bio *bio, struct bio_track *track)
3153 bio->bio_track = track;
3154 bio_track_ref(track);
3158 * Initiate I/O on a vnode.
3161 vn_strategy(struct vnode *vp, struct bio *bio)
3163 struct bio_track *track;
3165 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
3166 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
3167 track = &vp->v_track_read;
3169 track = &vp->v_track_write;
3170 bio->bio_track = track;
3171 bio_track_ref(track);
3172 vop_strategy(*vp->v_ops, vp, bio);
3179 * Finish I/O on a buffer, optionally calling a completion function.
3180 * This is usually called from an interrupt so process blocking is
3183 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3184 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3185 * assuming B_INVAL is clear.
3187 * For the VMIO case, we set B_CACHE if the op was a read and no
3188 * read error occured, or if the op was a write. B_CACHE is never
3189 * set if the buffer is invalid or otherwise uncacheable.
3191 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3192 * initiator to leave B_INVAL set to brelse the buffer out of existance
3193 * in the biodone routine.
3196 biodone(struct bio *bio)
3198 struct buf *bp = bio->bio_buf;
3203 KASSERT(BUF_REFCNTNB(bp) > 0,
3204 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3205 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3206 ("biodone: bp %p already done!", bp));
3208 runningbufwakeup(bp);
3211 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3214 biodone_t *done_func;
3215 struct bio_track *track;
3218 * BIO tracking. Most but not all BIOs are tracked.
3220 if ((track = bio->bio_track) != NULL) {
3221 bio_track_rel(track);
3222 bio->bio_track = NULL;
3226 * A bio_done function terminates the loop. The function
3227 * will be responsible for any further chaining and/or
3228 * buffer management.
3230 * WARNING! The done function can deallocate the buffer!
3232 if ((done_func = bio->bio_done) != NULL) {
3233 bio->bio_done = NULL;
3238 bio = bio->bio_prev;
3242 bp->b_cmd = BUF_CMD_DONE;
3245 * Only reads and writes are processed past this point.
3247 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3248 if (cmd == BUF_CMD_FREEBLKS)
3249 bp->b_flags |= B_NOCACHE;
3256 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3257 * a lot worse. XXX - move this above the clearing of b_cmd
3259 if (LIST_FIRST(&bp->b_dep) != NULL)
3263 * A failed write must re-dirty the buffer unless B_INVAL
3266 if (cmd == BUF_CMD_WRITE &&
3267 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3268 bp->b_flags &= ~B_NOCACHE;
3273 if (bp->b_flags & B_VMIO) {
3279 struct vnode *vp = bp->b_vp;
3283 #if defined(VFS_BIO_DEBUG)
3284 if (vp->v_auxrefs == 0)
3285 panic("biodone: zero vnode hold count");
3286 if ((vp->v_flag & VOBJBUF) == 0)
3287 panic("biodone: vnode is not setup for merged cache");
3290 foff = bp->b_loffset;
3291 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3292 KASSERT(obj != NULL, ("biodone: missing VM object"));
3294 #if defined(VFS_BIO_DEBUG)
3295 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3296 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3297 obj->paging_in_progress, bp->b_xio.xio_npages);
3302 * Set B_CACHE if the op was a normal read and no error
3303 * occured. B_CACHE is set for writes in the b*write()
3306 iosize = bp->b_bcount - bp->b_resid;
3307 if (cmd == BUF_CMD_READ && (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3308 bp->b_flags |= B_CACHE;
3311 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3315 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3320 * cleanup bogus pages, restoring the originals. Since
3321 * the originals should still be wired, we don't have
3322 * to worry about interrupt/freeing races destroying
3323 * the VM object association.
3325 m = bp->b_xio.xio_pages[i];
3326 if (m == bogus_page) {
3328 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3330 panic("biodone: page disappeared");
3331 bp->b_xio.xio_pages[i] = m;
3332 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3333 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3335 #if defined(VFS_BIO_DEBUG)
3336 if (OFF_TO_IDX(foff) != m->pindex) {
3338 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
3339 (unsigned long)foff, m->pindex);
3344 * In the write case, the valid and clean bits are
3345 * already changed correctly ( see bdwrite() ), so we
3346 * only need to do this here in the read case.
3348 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3349 vfs_page_set_valid(bp, foff, i, m);
3351 vm_page_flag_clear(m, PG_ZERO);
3354 * when debugging new filesystems or buffer I/O methods, this
3355 * is the most common error that pops up. if you see this, you
3356 * have not set the page busy flag correctly!!!
3359 kprintf("biodone: page busy < 0, "
3360 "pindex: %d, foff: 0x(%x,%x), "
3361 "resid: %d, index: %d\n",
3362 (int) m->pindex, (int)(foff >> 32),
3363 (int) foff & 0xffffffff, resid, i);
3364 if (!vn_isdisk(vp, NULL))
3365 kprintf(" iosize: %ld, loffset: %lld, "
3366 "flags: 0x%08x, npages: %d\n",
3367 bp->b_vp->v_mount->mnt_stat.f_iosize,
3368 (long long)bp->b_loffset,
3369 bp->b_flags, bp->b_xio.xio_npages);
3371 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3372 (long long)bp->b_loffset,
3373 bp->b_flags, bp->b_xio.xio_npages);
3374 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3375 m->valid, m->dirty, m->wire_count);
3376 panic("biodone: page busy < 0");
3378 vm_page_io_finish(m);
3379 vm_object_pip_subtract(obj, 1);
3380 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3384 vm_object_pip_wakeupn(obj, 0);
3388 * For asynchronous completions, release the buffer now. The brelse
3389 * will do a wakeup there if necessary - so no need to do a wakeup
3390 * here in the async case. The sync case always needs to do a wakeup.
3393 if (bp->b_flags & B_ASYNC) {
3394 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
3407 * This routine is called in lieu of iodone in the case of
3408 * incomplete I/O. This keeps the busy status for pages
3412 vfs_unbusy_pages(struct buf *bp)
3416 runningbufwakeup(bp);
3417 if (bp->b_flags & B_VMIO) {
3418 struct vnode *vp = bp->b_vp;
3423 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3424 vm_page_t m = bp->b_xio.xio_pages[i];
3427 * When restoring bogus changes the original pages
3428 * should still be wired, so we are in no danger of
3429 * losing the object association and do not need
3430 * critical section protection particularly.
3432 if (m == bogus_page) {
3433 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3435 panic("vfs_unbusy_pages: page missing");
3437 bp->b_xio.xio_pages[i] = m;
3438 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3439 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3441 vm_object_pip_subtract(obj, 1);
3442 vm_page_flag_clear(m, PG_ZERO);
3443 vm_page_io_finish(m);
3445 vm_object_pip_wakeupn(obj, 0);
3450 * vfs_page_set_valid:
3452 * Set the valid bits in a page based on the supplied offset. The
3453 * range is restricted to the buffer's size.
3455 * This routine is typically called after a read completes.
3458 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3460 vm_ooffset_t soff, eoff;
3463 * Start and end offsets in buffer. eoff - soff may not cross a
3464 * page boundry or cross the end of the buffer. The end of the
3465 * buffer, in this case, is our file EOF, not the allocation size
3469 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3470 if (eoff > bp->b_loffset + bp->b_bcount)
3471 eoff = bp->b_loffset + bp->b_bcount;
3474 * Set valid range. This is typically the entire buffer and thus the
3478 vm_page_set_validclean(
3480 (vm_offset_t) (soff & PAGE_MASK),
3481 (vm_offset_t) (eoff - soff)
3489 * This routine is called before a device strategy routine.
3490 * It is used to tell the VM system that paging I/O is in
3491 * progress, and treat the pages associated with the buffer
3492 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3493 * flag is handled to make sure that the object doesn't become
3496 * Since I/O has not been initiated yet, certain buffer flags
3497 * such as B_ERROR or B_INVAL may be in an inconsistant state
3498 * and should be ignored.
3501 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3504 struct lwp *lp = curthread->td_lwp;
3507 * The buffer's I/O command must already be set. If reading,
3508 * B_CACHE must be 0 (double check against callers only doing
3509 * I/O when B_CACHE is 0).
3511 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3512 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3514 if (bp->b_flags & B_VMIO) {
3519 foff = bp->b_loffset;
3520 KASSERT(bp->b_loffset != NOOFFSET,
3521 ("vfs_busy_pages: no buffer offset"));
3525 * Loop until none of the pages are busy.
3528 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3529 vm_page_t m = bp->b_xio.xio_pages[i];
3531 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3536 * Setup for I/O, soft-busy the page right now because
3537 * the next loop may block.
3539 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3540 vm_page_t m = bp->b_xio.xio_pages[i];
3542 vm_page_flag_clear(m, PG_ZERO);
3543 if ((bp->b_flags & B_CLUSTER) == 0) {
3544 vm_object_pip_add(obj, 1);
3545 vm_page_io_start(m);
3550 * Adjust protections for I/O and do bogus-page mapping.
3551 * Assume that vm_page_protect() can block (it can block
3552 * if VM_PROT_NONE, don't take any chances regardless).
3555 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3556 vm_page_t m = bp->b_xio.xio_pages[i];
3559 * When readying a vnode-backed buffer for a write
3560 * we must zero-fill any invalid portions of the
3563 * When readying a vnode-backed buffer for a read
3564 * we must replace any dirty pages with a bogus
3565 * page so we do not destroy dirty data when
3566 * filling in gaps. Dirty pages might not
3567 * necessarily be marked dirty yet, so use m->valid
3568 * as a reasonable test.
3570 * Bogus page replacement is, uh, bogus. We need
3571 * to find a better way.
3573 if (bp->b_cmd == BUF_CMD_WRITE) {
3574 vm_page_protect(m, VM_PROT_READ);
3575 vfs_page_set_valid(bp, foff, i, m);
3576 } else if (m->valid == VM_PAGE_BITS_ALL) {
3577 bp->b_xio.xio_pages[i] = bogus_page;
3580 vm_page_protect(m, VM_PROT_NONE);
3582 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3585 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3586 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3590 * This is the easiest place to put the process accounting for the I/O
3594 if (bp->b_cmd == BUF_CMD_READ)
3595 lp->lwp_ru.ru_inblock++;
3597 lp->lwp_ru.ru_oublock++;
3604 * Tell the VM system that the pages associated with this buffer
3605 * are clean. This is used for delayed writes where the data is
3606 * going to go to disk eventually without additional VM intevention.
3608 * Note that while we only really need to clean through to b_bcount, we
3609 * just go ahead and clean through to b_bufsize.
3612 vfs_clean_pages(struct buf *bp)
3616 if (bp->b_flags & B_VMIO) {
3619 foff = bp->b_loffset;
3620 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset"));
3621 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3622 vm_page_t m = bp->b_xio.xio_pages[i];
3623 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3625 vfs_page_set_valid(bp, foff, i, m);
3632 * vfs_bio_set_validclean:
3634 * Set the range within the buffer to valid and clean. The range is
3635 * relative to the beginning of the buffer, b_loffset. Note that
3636 * b_loffset itself may be offset from the beginning of the first page.
3640 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3642 if (bp->b_flags & B_VMIO) {
3647 * Fixup base to be relative to beginning of first page.
3648 * Set initial n to be the maximum number of bytes in the
3649 * first page that can be validated.
3652 base += (bp->b_loffset & PAGE_MASK);
3653 n = PAGE_SIZE - (base & PAGE_MASK);
3655 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) {
3656 vm_page_t m = bp->b_xio.xio_pages[i];
3661 vm_page_set_validclean(m, base & PAGE_MASK, n);
3672 * Clear a buffer. This routine essentially fakes an I/O, so we need
3673 * to clear B_ERROR and B_INVAL.
3675 * Note that while we only theoretically need to clear through b_bcount,
3676 * we go ahead and clear through b_bufsize.
3680 vfs_bio_clrbuf(struct buf *bp)
3684 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3685 bp->b_flags &= ~(B_INVAL|B_ERROR);
3686 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3687 (bp->b_loffset & PAGE_MASK) == 0) {
3688 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3689 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3693 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3694 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3695 bzero(bp->b_data, bp->b_bufsize);
3696 bp->b_xio.xio_pages[0]->valid |= mask;
3702 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3703 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3704 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3705 ea = (caddr_t)(vm_offset_t)ulmin(
3706 (u_long)(vm_offset_t)ea,
3707 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3708 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3709 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3711 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3712 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3716 for (; sa < ea; sa += DEV_BSIZE, j++) {
3717 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3718 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3719 bzero(sa, DEV_BSIZE);
3722 bp->b_xio.xio_pages[i]->valid |= mask;
3723 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
3732 * vm_hold_load_pages:
3734 * Load pages into the buffer's address space. The pages are
3735 * allocated from the kernel object in order to reduce interference
3736 * with the any VM paging I/O activity. The range of loaded
3737 * pages will be wired.
3739 * If a page cannot be allocated, the 'pagedaemon' is woken up to
3740 * retrieve the full range (to - from) of pages.
3744 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3750 to = round_page(to);
3751 from = round_page(from);
3752 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3757 * Note: must allocate system pages since blocking here
3758 * could intefere with paging I/O, no matter which
3761 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT,
3762 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
3765 p->valid = VM_PAGE_BITS_ALL;
3766 vm_page_flag_clear(p, PG_ZERO);
3767 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3768 bp->b_xio.xio_pages[index] = p;
3775 bp->b_xio.xio_npages = index;
3779 * Allocate pages for a buffer cache buffer.
3781 * Under extremely severe memory conditions even allocating out of the
3782 * system reserve can fail. If this occurs we must allocate out of the
3783 * interrupt reserve to avoid a deadlock with the pageout daemon.
3785 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
3786 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
3787 * against the pageout daemon if pages are not freed from other sources.
3791 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit)
3796 * Try a normal allocation, allow use of system reserve.
3798 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
3803 * The normal allocation failed and we clearly have a page
3804 * deficit. Try to reclaim some clean VM pages directly
3805 * from the buffer cache.
3807 vm_pageout_deficit += deficit;
3811 * We may have blocked, the caller will know what to do if the
3814 if (vm_page_lookup(obj, pg))
3818 * Allocate and allow use of the interrupt reserve.
3820 * If after all that we still can't allocate a VM page we are
3821 * in real trouble, but we slog on anyway hoping that the system
3824 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
3825 VM_ALLOC_INTERRUPT);
3827 if (vm_page_count_severe()) {
3828 kprintf("bio_page_alloc: WARNING emergency page "
3833 kprintf("bio_page_alloc: WARNING emergency page "
3834 "allocation failed\n");
3841 * vm_hold_free_pages:
3843 * Return pages associated with the buffer back to the VM system.
3845 * The range of pages underlying the buffer's address space will
3846 * be unmapped and un-wired.
3849 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3853 int index, newnpages;
3855 from = round_page(from);
3856 to = round_page(to);
3857 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3859 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3860 p = bp->b_xio.xio_pages[index];
3861 if (p && (index < bp->b_xio.xio_npages)) {
3863 kprintf("vm_hold_free_pages: doffset: %lld, "
3865 (long long)bp->b_bio2.bio_offset,
3866 (long long)bp->b_loffset);
3868 bp->b_xio.xio_pages[index] = NULL;
3871 vm_page_unwire(p, 0);
3875 bp->b_xio.xio_npages = newnpages;
3881 * Map a user buffer into KVM via a pbuf. On return the buffer's
3882 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
3886 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
3897 * bp had better have a command and it better be a pbuf.
3899 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3900 KKASSERT(bp->b_flags & B_PAGING);
3906 * Map the user data into KVM. Mappings have to be page-aligned.
3908 addr = (caddr_t)trunc_page((vm_offset_t)udata);
3911 vmprot = VM_PROT_READ;
3912 if (bp->b_cmd == BUF_CMD_READ)
3913 vmprot |= VM_PROT_WRITE;
3915 while (addr < udata + bytes) {
3917 * Do the vm_fault if needed; do the copy-on-write thing
3918 * when reading stuff off device into memory.
3920 * vm_fault_page*() returns a held VM page.
3922 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
3923 va = trunc_page(va);
3925 m = vm_fault_page_quick(va, vmprot, &error);
3927 for (i = 0; i < pidx; ++i) {
3928 vm_page_unhold(bp->b_xio.xio_pages[i]);
3929 bp->b_xio.xio_pages[i] = NULL;
3933 bp->b_xio.xio_pages[pidx] = m;
3939 * Map the page array and set the buffer fields to point to
3940 * the mapped data buffer.
3942 if (pidx > btoc(MAXPHYS))
3943 panic("vmapbuf: mapped more than MAXPHYS");
3944 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
3946 bp->b_xio.xio_npages = pidx;
3947 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
3948 bp->b_bcount = bytes;
3949 bp->b_bufsize = bytes;
3956 * Free the io map PTEs associated with this IO operation.
3957 * We also invalidate the TLB entries and restore the original b_addr.
3960 vunmapbuf(struct buf *bp)
3965 KKASSERT(bp->b_flags & B_PAGING);
3967 npages = bp->b_xio.xio_npages;
3968 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3969 for (pidx = 0; pidx < npages; ++pidx) {
3970 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
3971 bp->b_xio.xio_pages[pidx] = NULL;
3973 bp->b_xio.xio_npages = 0;
3974 bp->b_data = bp->b_kvabase;
3978 * Scan all buffers in the system and issue the callback.
3981 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
3987 for (n = 0; n < nbuf; ++n) {
3988 if ((error = callback(&buf[n], info)) < 0) {
3998 * print out statistics from the current status of the buffer pool
3999 * this can be toggeled by the system control option debug.syncprt
4008 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
4009 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
4011 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
4013 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4016 TAILQ_FOREACH(bp, dp, b_freelist) {
4017 counts[bp->b_bufsize/PAGE_SIZE]++;
4021 kprintf("%s: total-%d", bname[i], count);
4022 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4024 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
4032 DB_SHOW_COMMAND(buffer, db_show_buffer)
4035 struct buf *bp = (struct buf *)addr;
4038 db_printf("usage: show buffer <addr>\n");
4042 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4043 db_printf("b_cmd = %d\n", bp->b_cmd);
4044 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4045 "b_resid = %d\n, b_data = %p, "
4046 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4047 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4049 (long long)bp->b_bio2.bio_offset,
4050 (long long)(bp->b_bio2.bio_next ?
4051 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4052 if (bp->b_xio.xio_npages) {
4054 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4055 bp->b_xio.xio_npages);
4056 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4058 m = bp->b_xio.xio_pages[i];
4059 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4060 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4061 if ((i + 1) < bp->b_xio.xio_npages)