2 * Copyright (c) 1994,1997 John S. Dyson
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice immediately at the beginning of the file, without modification,
10 * this list of conditions, and the following disclaimer.
11 * 2. Absolutely no warranty of function or purpose is made by the author
14 * $FreeBSD: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $
15 * $DragonFly: src/sys/kern/vfs_bio.c,v 1.115 2008/08/13 11:02:31 swildner Exp $
19 * this file contains a new buffer I/O scheme implementing a coherent
20 * VM object and buffer cache scheme. Pains have been taken to make
21 * sure that the performance degradation associated with schemes such
22 * as this is not realized.
24 * Author: John S. Dyson
25 * Significant help during the development and debugging phases
26 * had been provided by David Greenman, also of the FreeBSD core team.
28 * see man buf(9) for more info.
31 #include <sys/param.h>
32 #include <sys/systm.h>
35 #include <sys/eventhandler.h>
37 #include <sys/malloc.h>
38 #include <sys/mount.h>
39 #include <sys/kernel.h>
40 #include <sys/kthread.h>
42 #include <sys/reboot.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sysctl.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_kern.h>
51 #include <vm/vm_pageout.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_map.h>
58 #include <sys/thread2.h>
59 #include <sys/spinlock2.h>
60 #include <vm/vm_page2.h>
71 BQUEUE_NONE, /* not on any queue */
72 BQUEUE_LOCKED, /* locked buffers */
73 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
74 BQUEUE_DIRTY, /* B_DELWRI buffers */
75 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
76 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
77 BQUEUE_EMPTY, /* empty buffer headers */
79 BUFFER_QUEUES /* number of buffer queues */
82 typedef enum bufq_type bufq_type_t;
84 #define BD_WAKE_SIZE 128
85 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
87 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
88 struct spinlock bufspin = SPINLOCK_INITIALIZER(&bufspin);
90 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
92 struct buf *buf; /* buffer header pool */
94 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
95 int pageno, vm_page_t m);
96 static void vfs_clean_pages(struct buf *bp);
97 static void vfs_setdirty(struct buf *bp);
98 static void vfs_vmio_release(struct buf *bp);
99 static int flushbufqueues(bufq_type_t q);
100 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);
102 static void bd_signal(int totalspace);
103 static void buf_daemon(void);
104 static void buf_daemon_hw(void);
107 * bogus page -- for I/O to/from partially complete buffers
108 * this is a temporary solution to the problem, but it is not
109 * really that bad. it would be better to split the buffer
110 * for input in the case of buffers partially already in memory,
111 * but the code is intricate enough already.
113 vm_page_t bogus_page;
116 * These are all static, but make the ones we export globals so we do
117 * not need to use compiler magic.
119 int bufspace, maxbufspace,
120 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
121 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
122 static int lorunningspace, hirunningspace, runningbufreq;
123 int dirtybufspace, dirtybufspacehw, lodirtybufspace, hidirtybufspace;
124 int dirtybufcount, dirtybufcounthw;
125 int runningbufspace, runningbufcount;
126 static int getnewbufcalls;
127 static int getnewbufrestarts;
128 static int recoverbufcalls;
129 static int needsbuffer; /* locked by needsbuffer_spin */
130 static int bd_request; /* locked by needsbuffer_spin */
131 static int bd_request_hw; /* locked by needsbuffer_spin */
132 static u_int bd_wake_ary[BD_WAKE_SIZE];
133 static u_int bd_wake_index;
134 static struct spinlock needsbuffer_spin;
136 static struct thread *bufdaemon_td;
137 static struct thread *bufdaemonhw_td;
141 * Sysctls for operational control of the buffer cache.
143 SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
144 "Number of dirty buffers to flush before bufdaemon becomes inactive");
145 SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
146 "High watermark used to trigger explicit flushing of dirty buffers");
147 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
148 "Minimum amount of buffer space required for active I/O");
149 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
150 "Maximum amount of buffer space to usable for active I/O");
152 * Sysctls determining current state of the buffer cache.
154 SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
155 "Total number of buffers in buffer cache");
156 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
157 "Pending bytes of dirty buffers (all)");
158 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
159 "Pending bytes of dirty buffers (heavy weight)");
160 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
161 "Pending number of dirty buffers");
162 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
163 "Pending number of dirty buffers (heavy weight)");
164 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
165 "I/O bytes currently in progress due to asynchronous writes");
166 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
167 "I/O buffers currently in progress due to asynchronous writes");
168 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
169 "Hard limit on maximum amount of memory usable for buffer space");
170 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
171 "Soft limit on maximum amount of memory usable for buffer space");
172 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
173 "Minimum amount of memory to reserve for system buffer space");
174 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
175 "Amount of memory available for buffers");
176 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
177 0, "Maximum amount of memory reserved for buffers using malloc");
178 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
179 "Amount of memory left for buffers using malloc-scheme");
180 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
181 "New buffer header acquisition requests");
182 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
183 0, "New buffer header acquisition restarts");
184 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
185 "Recover VM space in an emergency");
186 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
187 "Buffer acquisition restarts due to fragmented buffer map");
188 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
189 "Amount of time KVA space was deallocated in an arbitrary buffer");
190 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
191 "Amount of time buffer re-use operations were successful");
192 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
193 "sizeof(struct buf)");
195 char *buf_wmesg = BUF_WMESG;
197 extern int vm_swap_size;
199 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
200 #define VFS_BIO_NEED_UNUSED02 0x02
201 #define VFS_BIO_NEED_UNUSED04 0x04
202 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
207 * Called when buffer space is potentially available for recovery.
208 * getnewbuf() will block on this flag when it is unable to free
209 * sufficient buffer space. Buffer space becomes recoverable when
210 * bp's get placed back in the queues.
217 * If someone is waiting for BUF space, wake them up. Even
218 * though we haven't freed the kva space yet, the waiting
219 * process will be able to now.
221 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
222 spin_lock_wr(&needsbuffer_spin);
223 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
224 spin_unlock_wr(&needsbuffer_spin);
225 wakeup(&needsbuffer);
232 * Accounting for I/O in progress.
236 runningbufwakeup(struct buf *bp)
240 if ((totalspace = bp->b_runningbufspace) != 0) {
241 runningbufspace -= totalspace;
243 bp->b_runningbufspace = 0;
244 if (runningbufreq && runningbufspace <= lorunningspace) {
246 wakeup(&runningbufreq);
248 bd_signal(totalspace);
255 * Called when a buffer has been added to one of the free queues to
256 * account for the buffer and to wakeup anyone waiting for free buffers.
257 * This typically occurs when large amounts of metadata are being handled
258 * by the buffer cache ( else buffer space runs out first, usually ).
266 spin_lock_wr(&needsbuffer_spin);
267 needsbuffer &= ~VFS_BIO_NEED_ANY;
268 spin_unlock_wr(&needsbuffer_spin);
269 wakeup(&needsbuffer);
274 * waitrunningbufspace()
276 * Wait for the amount of running I/O to drop to a reasonable level.
278 * The caller may be using this function to block in a tight loop, we
279 * must block of runningbufspace is greater then the passed limit.
280 * And even with that it may not be enough, due to the presence of
281 * B_LOCKED dirty buffers, so also wait for at least one running buffer
285 waitrunningbufspace(int limit)
289 if (lorunningspace < limit)
290 lorun = lorunningspace;
295 if (runningbufspace > lorun) {
296 while (runningbufspace > lorun) {
298 tsleep(&runningbufreq, 0, "wdrain", 0);
300 } else if (runningbufspace) {
302 tsleep(&runningbufreq, 0, "wdrain2", 1);
308 * vfs_buf_test_cache:
310 * Called when a buffer is extended. This function clears the B_CACHE
311 * bit if the newly extended portion of the buffer does not contain
316 vfs_buf_test_cache(struct buf *bp,
317 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
320 if (bp->b_flags & B_CACHE) {
321 int base = (foff + off) & PAGE_MASK;
322 if (vm_page_is_valid(m, base, size) == 0)
323 bp->b_flags &= ~B_CACHE;
330 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
339 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
342 if (bd_request == 0 &&
343 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
344 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
345 spin_lock_wr(&needsbuffer_spin);
347 spin_unlock_wr(&needsbuffer_spin);
350 if (bd_request_hw == 0 &&
351 (dirtybufspacehw > lodirtybufspace / 2 ||
352 dirtybufcounthw >= nbuf / 2)) {
353 spin_lock_wr(&needsbuffer_spin);
355 spin_unlock_wr(&needsbuffer_spin);
356 wakeup(&bd_request_hw);
363 * Get the buf_daemon heated up when the number of running and dirty
364 * buffers exceeds the mid-point.
375 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
377 totalspace = runningbufspace + dirtybufspace;
378 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
380 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
381 if (totalspace >= mid2)
382 return(totalspace - mid2);
390 * Wait for the buffer cache to flush (totalspace) bytes worth of
391 * buffers, then return.
393 * Regardless this function blocks while the number of dirty buffers
394 * exceeds hidirtybufspace.
399 bd_wait(int totalspace)
404 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
407 while (totalspace > 0) {
410 if (totalspace > runningbufspace + dirtybufspace)
411 totalspace = runningbufspace + dirtybufspace;
412 count = totalspace / BKVASIZE;
413 if (count >= BD_WAKE_SIZE)
414 count = BD_WAKE_SIZE - 1;
416 spin_lock_wr(&needsbuffer_spin);
417 i = (bd_wake_index + count) & BD_WAKE_MASK;
419 tsleep_interlock(&bd_wake_ary[i]);
420 spin_unlock_wr(&needsbuffer_spin);
422 tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz);
425 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
432 * This function is called whenever runningbufspace or dirtybufspace
433 * is reduced. Track threads waiting for run+dirty buffer I/O
439 bd_signal(int totalspace)
443 if (totalspace > 0) {
444 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
445 totalspace = BKVASIZE * BD_WAKE_SIZE;
446 spin_lock_wr(&needsbuffer_spin);
447 while (totalspace > 0) {
450 if (bd_wake_ary[i]) {
452 spin_unlock_wr(&needsbuffer_spin);
453 wakeup(&bd_wake_ary[i]);
454 spin_lock_wr(&needsbuffer_spin);
456 totalspace -= BKVASIZE;
458 spin_unlock_wr(&needsbuffer_spin);
463 * BIO tracking support routines.
465 * Release a ref on a bio_track. Wakeup requests are atomically released
466 * along with the last reference so bk_active will never wind up set to
473 bio_track_rel(struct bio_track *track)
481 active = track->bk_active;
482 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
486 * Full-on. Note that the wait flag is only atomically released on
487 * the 1->0 count transition.
489 * We check for a negative count transition using bit 30 since bit 31
490 * has a different meaning.
493 desired = (active & 0x7FFFFFFF) - 1;
495 desired |= active & 0x80000000;
496 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
497 if (desired & 0x40000000)
498 panic("bio_track_rel: bad count: %p\n", track);
499 if (active & 0x80000000)
503 active = track->bk_active;
508 * Wait for the tracking count to reach 0.
510 * Use atomic ops such that the wait flag is only set atomically when
511 * bk_active is non-zero.
516 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
525 if (track->bk_active == 0)
529 * Full-on. Note that the wait flag may only be atomically set if
530 * the active count is non-zero.
532 crit_enter(); /* for tsleep_interlock */
534 while ((active = track->bk_active) != 0) {
535 desired = active | 0x80000000;
536 tsleep_interlock(track);
537 if (active == desired ||
538 atomic_cmpset_int(&track->bk_active, active, desired)) {
539 error = tsleep(track, slp_flags | PINTERLOCKED,
552 * Load time initialisation of the buffer cache, called from machine
553 * dependant initialization code.
559 vm_offset_t bogus_offset;
562 spin_init(&needsbuffer_spin);
564 /* next, make a null set of free lists */
565 for (i = 0; i < BUFFER_QUEUES; i++)
566 TAILQ_INIT(&bufqueues[i]);
568 /* finally, initialize each buffer header and stick on empty q */
569 for (i = 0; i < nbuf; i++) {
571 bzero(bp, sizeof *bp);
572 bp->b_flags = B_INVAL; /* we're just an empty header */
573 bp->b_cmd = BUF_CMD_DONE;
574 bp->b_qindex = BQUEUE_EMPTY;
576 xio_init(&bp->b_xio);
579 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
583 * maxbufspace is the absolute maximum amount of buffer space we are
584 * allowed to reserve in KVM and in real terms. The absolute maximum
585 * is nominally used by buf_daemon. hibufspace is the nominal maximum
586 * used by most other processes. The differential is required to
587 * ensure that buf_daemon is able to run when other processes might
588 * be blocked waiting for buffer space.
590 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
591 * this may result in KVM fragmentation which is not handled optimally
594 maxbufspace = nbuf * BKVASIZE;
595 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
596 lobufspace = hibufspace - MAXBSIZE;
598 lorunningspace = 512 * 1024;
599 hirunningspace = 1024 * 1024;
602 * Limit the amount of malloc memory since it is wired permanently
603 * into the kernel space. Even though this is accounted for in
604 * the buffer allocation, we don't want the malloced region to grow
605 * uncontrolled. The malloc scheme improves memory utilization
606 * significantly on average (small) directories.
608 maxbufmallocspace = hibufspace / 20;
611 * Reduce the chance of a deadlock occuring by limiting the number
612 * of delayed-write dirty buffers we allow to stack up.
614 hidirtybufspace = hibufspace / 2;
618 lodirtybufspace = hidirtybufspace / 2;
621 * Maximum number of async ops initiated per buf_daemon loop. This is
622 * somewhat of a hack at the moment, we really need to limit ourselves
623 * based on the number of bytes of I/O in-transit that were initiated
627 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
628 bogus_page = vm_page_alloc(&kernel_object,
629 (bogus_offset >> PAGE_SHIFT),
631 vmstats.v_wire_count++;
636 * Initialize the embedded bio structures
639 initbufbio(struct buf *bp)
641 bp->b_bio1.bio_buf = bp;
642 bp->b_bio1.bio_prev = NULL;
643 bp->b_bio1.bio_offset = NOOFFSET;
644 bp->b_bio1.bio_next = &bp->b_bio2;
645 bp->b_bio1.bio_done = NULL;
647 bp->b_bio2.bio_buf = bp;
648 bp->b_bio2.bio_prev = &bp->b_bio1;
649 bp->b_bio2.bio_offset = NOOFFSET;
650 bp->b_bio2.bio_next = NULL;
651 bp->b_bio2.bio_done = NULL;
655 * Reinitialize the embedded bio structures as well as any additional
656 * translation cache layers.
659 reinitbufbio(struct buf *bp)
663 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
664 bio->bio_done = NULL;
665 bio->bio_offset = NOOFFSET;
670 * Push another BIO layer onto an existing BIO and return it. The new
671 * BIO layer may already exist, holding cached translation data.
674 push_bio(struct bio *bio)
678 if ((nbio = bio->bio_next) == NULL) {
679 int index = bio - &bio->bio_buf->b_bio_array[0];
680 if (index >= NBUF_BIO - 1) {
681 panic("push_bio: too many layers bp %p\n",
684 nbio = &bio->bio_buf->b_bio_array[index + 1];
685 bio->bio_next = nbio;
686 nbio->bio_prev = bio;
687 nbio->bio_buf = bio->bio_buf;
688 nbio->bio_offset = NOOFFSET;
689 nbio->bio_done = NULL;
690 nbio->bio_next = NULL;
692 KKASSERT(nbio->bio_done == NULL);
697 * Pop a BIO translation layer, returning the previous layer. The
698 * must have been previously pushed.
701 pop_bio(struct bio *bio)
703 return(bio->bio_prev);
707 clearbiocache(struct bio *bio)
710 bio->bio_offset = NOOFFSET;
718 * Free the KVA allocation for buffer 'bp'.
720 * Must be called from a critical section as this is the only locking for
723 * Since this call frees up buffer space, we call bufspacewakeup().
728 bfreekva(struct buf *bp)
735 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
736 vm_map_lock(&buffer_map);
737 bufspace -= bp->b_kvasize;
738 vm_map_delete(&buffer_map,
739 (vm_offset_t) bp->b_kvabase,
740 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
743 vm_map_unlock(&buffer_map);
744 vm_map_entry_release(count);
754 * Remove the buffer from the appropriate free list.
757 _bremfree(struct buf *bp)
759 if (bp->b_qindex != BQUEUE_NONE) {
760 KASSERT(BUF_REFCNTNB(bp) == 1,
761 ("bremfree: bp %p not locked",bp));
762 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
763 bp->b_qindex = BQUEUE_NONE;
765 if (BUF_REFCNTNB(bp) <= 1)
766 panic("bremfree: removing a buffer not on a queue");
771 bremfree(struct buf *bp)
773 spin_lock_wr(&bufspin);
775 spin_unlock_wr(&bufspin);
779 bremfree_locked(struct buf *bp)
787 * Get a buffer with the specified data. Look in the cache first. We
788 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
789 * is set, the buffer is valid and we do not have to do anything ( see
795 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
799 bp = getblk(vp, loffset, size, 0, 0);
802 /* if not found in cache, do some I/O */
803 if ((bp->b_flags & B_CACHE) == 0) {
805 KASSERT(!(bp->b_flags & B_ASYNC),
806 ("bread: illegal async bp %p", bp));
807 bp->b_flags &= ~(B_ERROR | B_INVAL);
808 bp->b_cmd = BUF_CMD_READ;
809 vfs_busy_pages(vp, bp);
810 vn_strategy(vp, &bp->b_bio1);
812 return (biowait(bp));
820 * Operates like bread, but also starts asynchronous I/O on
821 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
822 * to initiating I/O . If B_CACHE is set, the buffer is valid
823 * and we do not have to do anything.
828 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
829 int *rabsize, int cnt, struct buf **bpp)
831 struct buf *bp, *rabp;
833 int rv = 0, readwait = 0;
835 *bpp = bp = getblk(vp, loffset, size, 0, 0);
837 /* if not found in cache, do some I/O */
838 if ((bp->b_flags & B_CACHE) == 0) {
840 bp->b_flags &= ~(B_ERROR | B_INVAL);
841 bp->b_cmd = BUF_CMD_READ;
842 vfs_busy_pages(vp, bp);
843 vn_strategy(vp, &bp->b_bio1);
848 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
849 if (inmem(vp, *raoffset))
851 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
853 if ((rabp->b_flags & B_CACHE) == 0) {
855 rabp->b_flags |= B_ASYNC;
856 rabp->b_flags &= ~(B_ERROR | B_INVAL);
857 rabp->b_cmd = BUF_CMD_READ;
858 vfs_busy_pages(vp, rabp);
860 vn_strategy(vp, &rabp->b_bio1);
874 * Write, release buffer on completion. (Done by iodone
875 * if async). Do not bother writing anything if the buffer
878 * Note that we set B_CACHE here, indicating that buffer is
879 * fully valid and thus cacheable. This is true even of NFS
880 * now so we set it generally. This could be set either here
881 * or in biodone() since the I/O is synchronous. We put it
885 bwrite(struct buf *bp)
889 if (bp->b_flags & B_INVAL) {
894 oldflags = bp->b_flags;
896 if (BUF_REFCNTNB(bp) == 0)
897 panic("bwrite: buffer is not busy???");
900 /* Mark the buffer clean */
903 bp->b_flags &= ~B_ERROR;
904 bp->b_flags |= B_CACHE;
905 bp->b_cmd = BUF_CMD_WRITE;
906 vfs_busy_pages(bp->b_vp, bp);
909 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
910 * valid for vnode-backed buffers.
912 bp->b_runningbufspace = bp->b_bufsize;
913 if (bp->b_runningbufspace) {
914 runningbufspace += bp->b_runningbufspace;
919 if (oldflags & B_ASYNC)
921 vn_strategy(bp->b_vp, &bp->b_bio1);
923 if ((oldflags & B_ASYNC) == 0) {
924 int rtval = biowait(bp);
934 * Delayed write. (Buffer is marked dirty). Do not bother writing
935 * anything if the buffer is marked invalid.
937 * Note that since the buffer must be completely valid, we can safely
938 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
939 * biodone() in order to prevent getblk from writing the buffer
943 bdwrite(struct buf *bp)
945 if (BUF_REFCNTNB(bp) == 0)
946 panic("bdwrite: buffer is not busy");
948 if (bp->b_flags & B_INVAL) {
955 * Set B_CACHE, indicating that the buffer is fully valid. This is
956 * true even of NFS now.
958 bp->b_flags |= B_CACHE;
961 * This bmap keeps the system from needing to do the bmap later,
962 * perhaps when the system is attempting to do a sync. Since it
963 * is likely that the indirect block -- or whatever other datastructure
964 * that the filesystem needs is still in memory now, it is a good
965 * thing to do this. Note also, that if the pageout daemon is
966 * requesting a sync -- there might not be enough memory to do
967 * the bmap then... So, this is important to do.
969 if (bp->b_bio2.bio_offset == NOOFFSET) {
970 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
971 NULL, NULL, BUF_CMD_WRITE);
975 * Set the *dirty* buffer range based upon the VM system dirty pages.
980 * We need to do this here to satisfy the vnode_pager and the
981 * pageout daemon, so that it thinks that the pages have been
982 * "cleaned". Note that since the pages are in a delayed write
983 * buffer -- the VFS layer "will" see that the pages get written
984 * out on the next sync, or perhaps the cluster will be completed.
990 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
991 * due to the softdep code.
998 * Turn buffer into delayed write request by marking it B_DELWRI.
999 * B_RELBUF and B_NOCACHE must be cleared.
1001 * We reassign the buffer to itself to properly update it in the
1002 * dirty/clean lists.
1004 * Must be called from a critical section.
1005 * The buffer must be on BQUEUE_NONE.
1008 bdirty(struct buf *bp)
1010 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1011 if (bp->b_flags & B_NOCACHE) {
1012 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1013 bp->b_flags &= ~B_NOCACHE;
1015 if (bp->b_flags & B_INVAL) {
1016 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1018 bp->b_flags &= ~B_RELBUF;
1020 if ((bp->b_flags & B_DELWRI) == 0) {
1021 bp->b_flags |= B_DELWRI;
1023 atomic_add_int(&dirtybufcount, 1);
1024 dirtybufspace += bp->b_bufsize;
1025 if (bp->b_flags & B_HEAVY) {
1026 atomic_add_int(&dirtybufcounthw, 1);
1027 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
1034 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1035 * needs to be flushed with a different buf_daemon thread to avoid
1036 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1039 bheavy(struct buf *bp)
1041 if ((bp->b_flags & B_HEAVY) == 0) {
1042 bp->b_flags |= B_HEAVY;
1043 if (bp->b_flags & B_DELWRI) {
1044 atomic_add_int(&dirtybufcounthw, 1);
1045 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
1053 * Clear B_DELWRI for buffer.
1055 * Must be called from a critical section.
1057 * The buffer is typically on BQUEUE_NONE but there is one case in
1058 * brelse() that calls this function after placing the buffer on
1059 * a different queue.
1064 bundirty(struct buf *bp)
1066 if (bp->b_flags & B_DELWRI) {
1067 bp->b_flags &= ~B_DELWRI;
1069 atomic_subtract_int(&dirtybufcount, 1);
1070 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
1071 if (bp->b_flags & B_HEAVY) {
1072 atomic_subtract_int(&dirtybufcounthw, 1);
1073 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
1075 bd_signal(bp->b_bufsize);
1078 * Since it is now being written, we can clear its deferred write flag.
1080 bp->b_flags &= ~B_DEFERRED;
1086 * Asynchronous write. Start output on a buffer, but do not wait for
1087 * it to complete. The buffer is released when the output completes.
1089 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1090 * B_INVAL buffers. Not us.
1093 bawrite(struct buf *bp)
1095 bp->b_flags |= B_ASYNC;
1102 * Ordered write. Start output on a buffer, and flag it so that the
1103 * device will write it in the order it was queued. The buffer is
1104 * released when the output completes. bwrite() ( or the VOP routine
1105 * anyway ) is responsible for handling B_INVAL buffers.
1108 bowrite(struct buf *bp)
1110 bp->b_flags |= B_ORDERED | B_ASYNC;
1111 return (bwrite(bp));
1115 * buf_dirty_count_severe:
1117 * Return true if we have too many dirty buffers.
1120 buf_dirty_count_severe(void)
1122 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
1123 dirtybufcount >= nbuf / 2);
1129 * Release a busy buffer and, if requested, free its resources. The
1130 * buffer will be stashed in the appropriate bufqueue[] allowing it
1131 * to be accessed later as a cache entity or reused for other purposes.
1136 brelse(struct buf *bp)
1139 int saved_flags = bp->b_flags;
1142 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1145 * If B_NOCACHE is set we are being asked to destroy the buffer and
1146 * its backing store. Clear B_DELWRI.
1148 * B_NOCACHE is set in two cases: (1) when the caller really wants
1149 * to destroy the buffer and backing store and (2) when the caller
1150 * wants to destroy the buffer and backing store after a write
1153 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1157 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1159 * A re-dirtied buffer is only subject to destruction
1160 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1162 /* leave buffer intact */
1163 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1164 (bp->b_bufsize <= 0)) {
1166 * Either a failed read or we were asked to free or not
1167 * cache the buffer. This path is reached with B_DELWRI
1168 * set only if B_INVAL is already set. B_NOCACHE governs
1169 * backing store destruction.
1171 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1172 * buffer cannot be immediately freed.
1174 bp->b_flags |= B_INVAL;
1175 if (LIST_FIRST(&bp->b_dep) != NULL) {
1180 if (bp->b_flags & B_DELWRI) {
1181 atomic_subtract_int(&dirtybufcount, 1);
1182 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
1183 if (bp->b_flags & B_HEAVY) {
1184 atomic_subtract_int(&dirtybufcounthw, 1);
1185 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
1187 bd_signal(bp->b_bufsize);
1189 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1193 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1194 * If vfs_vmio_release() is called with either bit set, the
1195 * underlying pages may wind up getting freed causing a previous
1196 * write (bdwrite()) to get 'lost' because pages associated with
1197 * a B_DELWRI bp are marked clean. Pages associated with a
1198 * B_LOCKED buffer may be mapped by the filesystem.
1200 * If we want to release the buffer ourselves (rather then the
1201 * originator asking us to release it), give the originator a
1202 * chance to countermand the release by setting B_LOCKED.
1204 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1205 * if B_DELWRI is set.
1207 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1208 * on pages to return pages to the VM page queues.
1210 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1211 bp->b_flags &= ~B_RELBUF;
1212 } else if (vm_page_count_severe()) {
1213 if (LIST_FIRST(&bp->b_dep) != NULL) {
1215 buf_deallocate(bp); /* can set B_LOCKED */
1218 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1219 bp->b_flags &= ~B_RELBUF;
1221 bp->b_flags |= B_RELBUF;
1225 * Make sure b_cmd is clear. It may have already been cleared by
1228 * At this point destroying the buffer is governed by the B_INVAL
1229 * or B_RELBUF flags.
1231 bp->b_cmd = BUF_CMD_DONE;
1234 * VMIO buffer rundown. Make sure the VM page array is restored
1235 * after an I/O may have replaces some of the pages with bogus pages
1236 * in order to not destroy dirty pages in a fill-in read.
1238 * Note that due to the code above, if a buffer is marked B_DELWRI
1239 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1240 * B_INVAL may still be set, however.
1242 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1243 * but not the backing store. B_NOCACHE will destroy the backing
1246 * Note that dirty NFS buffers contain byte-granular write ranges
1247 * and should not be destroyed w/ B_INVAL even if the backing store
1250 if (bp->b_flags & B_VMIO) {
1252 * Rundown for VMIO buffers which are not dirty NFS buffers.
1264 * Get the base offset and length of the buffer. Note that
1265 * in the VMIO case if the buffer block size is not
1266 * page-aligned then b_data pointer may not be page-aligned.
1267 * But our b_xio.xio_pages array *IS* page aligned.
1269 * block sizes less then DEV_BSIZE (usually 512) are not
1270 * supported due to the page granularity bits (m->valid,
1271 * m->dirty, etc...).
1273 * See man buf(9) for more information
1276 resid = bp->b_bufsize;
1277 foff = bp->b_loffset;
1280 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1281 m = bp->b_xio.xio_pages[i];
1282 vm_page_flag_clear(m, PG_ZERO);
1284 * If we hit a bogus page, fixup *all* of them
1285 * now. Note that we left these pages wired
1286 * when we removed them so they had better exist,
1287 * and they cannot be ripped out from under us so
1288 * no critical section protection is necessary.
1290 if (m == bogus_page) {
1292 poff = OFF_TO_IDX(bp->b_loffset);
1294 for (j = i; j < bp->b_xio.xio_npages; j++) {
1297 mtmp = bp->b_xio.xio_pages[j];
1298 if (mtmp == bogus_page) {
1299 mtmp = vm_page_lookup(obj, poff + j);
1301 panic("brelse: page missing");
1303 bp->b_xio.xio_pages[j] = mtmp;
1307 if ((bp->b_flags & B_INVAL) == 0) {
1308 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1309 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1311 m = bp->b_xio.xio_pages[i];
1315 * Invalidate the backing store if B_NOCACHE is set
1316 * (e.g. used with vinvalbuf()). If this is NFS
1317 * we impose a requirement that the block size be
1318 * a multiple of PAGE_SIZE and create a temporary
1319 * hack to basically invalidate the whole page. The
1320 * problem is that NFS uses really odd buffer sizes
1321 * especially when tracking piecemeal writes and
1322 * it also vinvalbuf()'s a lot, which would result
1323 * in only partial page validation and invalidation
1324 * here. If the file page is mmap()'d, however,
1325 * all the valid bits get set so after we invalidate
1326 * here we would end up with weird m->valid values
1327 * like 0xfc. nfs_getpages() can't handle this so
1328 * we clear all the valid bits for the NFS case
1329 * instead of just some of them.
1331 * The real bug is the VM system having to set m->valid
1332 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1333 * itself is an artifact of the whole 512-byte
1334 * granular mess that exists to support odd block
1335 * sizes and UFS meta-data block sizes (e.g. 6144).
1336 * A complete rewrite is required.
1338 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1339 int poffset = foff & PAGE_MASK;
1342 presid = PAGE_SIZE - poffset;
1343 if (bp->b_vp->v_tag == VT_NFS &&
1344 bp->b_vp->v_type == VREG) {
1346 } else if (presid > resid) {
1349 KASSERT(presid >= 0, ("brelse: extra page"));
1350 vm_page_set_invalid(m, poffset, presid);
1352 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1353 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1355 if (bp->b_flags & (B_INVAL | B_RELBUF))
1356 vfs_vmio_release(bp);
1360 * Rundown for non-VMIO buffers.
1362 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1366 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1373 if (bp->b_qindex != BQUEUE_NONE)
1374 panic("brelse: free buffer onto another queue???");
1375 if (BUF_REFCNTNB(bp) > 1) {
1376 /* Temporary panic to verify exclusive locking */
1377 /* This panic goes away when we allow shared refs */
1378 panic("brelse: multiple refs");
1384 * Figure out the correct queue to place the cleaned up buffer on.
1385 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1386 * disassociated from their vnode.
1388 spin_lock_wr(&bufspin);
1389 if (bp->b_flags & B_LOCKED) {
1391 * Buffers that are locked are placed in the locked queue
1392 * immediately, regardless of their state.
1394 bp->b_qindex = BQUEUE_LOCKED;
1395 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1396 } else if (bp->b_bufsize == 0) {
1398 * Buffers with no memory. Due to conditionals near the top
1399 * of brelse() such buffers should probably already be
1400 * marked B_INVAL and disassociated from their vnode.
1402 bp->b_flags |= B_INVAL;
1403 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1404 KKASSERT((bp->b_flags & B_HASHED) == 0);
1405 if (bp->b_kvasize) {
1406 bp->b_qindex = BQUEUE_EMPTYKVA;
1408 bp->b_qindex = BQUEUE_EMPTY;
1410 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1411 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1413 * Buffers with junk contents. Again these buffers had better
1414 * already be disassociated from their vnode.
1416 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1417 KKASSERT((bp->b_flags & B_HASHED) == 0);
1418 bp->b_flags |= B_INVAL;
1419 bp->b_qindex = BQUEUE_CLEAN;
1420 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1423 * Remaining buffers. These buffers are still associated with
1426 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1428 bp->b_qindex = BQUEUE_DIRTY;
1429 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1431 case B_DELWRI | B_HEAVY:
1432 bp->b_qindex = BQUEUE_DIRTY_HW;
1433 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1438 * NOTE: Buffers are always placed at the end of the
1439 * queue. If B_AGE is not set the buffer will cycle
1440 * through the queue twice.
1442 bp->b_qindex = BQUEUE_CLEAN;
1443 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1447 spin_unlock_wr(&bufspin);
1450 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1451 * on the correct queue.
1453 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1457 * The bp is on an appropriate queue unless locked. If it is not
1458 * locked or dirty we can wakeup threads waiting for buffer space.
1460 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1461 * if B_INVAL is set ).
1463 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1467 * Something we can maybe free or reuse
1469 if (bp->b_bufsize || bp->b_kvasize)
1473 * Clean up temporary flags and unlock the buffer.
1475 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
1482 * Release a buffer back to the appropriate queue but do not try to free
1483 * it. The buffer is expected to be used again soon.
1485 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1486 * biodone() to requeue an async I/O on completion. It is also used when
1487 * known good buffers need to be requeued but we think we may need the data
1490 * XXX we should be able to leave the B_RELBUF hint set on completion.
1495 bqrelse(struct buf *bp)
1497 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1499 if (bp->b_qindex != BQUEUE_NONE)
1500 panic("bqrelse: free buffer onto another queue???");
1501 if (BUF_REFCNTNB(bp) > 1) {
1502 /* do not release to free list */
1503 panic("bqrelse: multiple refs");
1507 spin_lock_wr(&bufspin);
1508 if (bp->b_flags & B_LOCKED) {
1510 * Locked buffers are released to the locked queue. However,
1511 * if the buffer is dirty it will first go into the dirty
1512 * queue and later on after the I/O completes successfully it
1513 * will be released to the locked queue.
1515 bp->b_qindex = BQUEUE_LOCKED;
1516 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1517 } else if (bp->b_flags & B_DELWRI) {
1518 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1519 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1520 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1521 } else if (vm_page_count_severe()) {
1523 * We are too low on memory, we have to try to free the
1524 * buffer (most importantly: the wired pages making up its
1525 * backing store) *now*.
1527 spin_unlock_wr(&bufspin);
1531 bp->b_qindex = BQUEUE_CLEAN;
1532 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1534 spin_unlock_wr(&bufspin);
1536 if ((bp->b_flags & B_LOCKED) == 0 &&
1537 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1542 * Something we can maybe free or reuse.
1544 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1548 * Final cleanup and unlock. Clear bits that are only used while a
1549 * buffer is actively locked.
1551 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF);
1558 * Return backing pages held by the buffer 'bp' back to the VM system
1559 * if possible. The pages are freed if they are no longer valid or
1560 * attempt to free if it was used for direct I/O otherwise they are
1561 * sent to the page cache.
1563 * Pages that were marked busy are left alone and skipped.
1565 * The KVA mapping (b_data) for the underlying pages is removed by
1569 vfs_vmio_release(struct buf *bp)
1575 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1576 m = bp->b_xio.xio_pages[i];
1577 bp->b_xio.xio_pages[i] = NULL;
1579 * In order to keep page LRU ordering consistent, put
1580 * everything on the inactive queue.
1582 vm_page_unwire(m, 0);
1584 * We don't mess with busy pages, it is
1585 * the responsibility of the process that
1586 * busied the pages to deal with them.
1588 if ((m->flags & PG_BUSY) || (m->busy != 0))
1591 if (m->wire_count == 0) {
1592 vm_page_flag_clear(m, PG_ZERO);
1594 * Might as well free the page if we can and it has
1595 * no valid data. We also free the page if the
1596 * buffer was used for direct I/O.
1598 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1599 m->hold_count == 0) {
1601 vm_page_protect(m, VM_PROT_NONE);
1603 } else if (bp->b_flags & B_DIRECT) {
1604 vm_page_try_to_free(m);
1605 } else if (vm_page_count_severe()) {
1606 vm_page_try_to_cache(m);
1611 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1612 if (bp->b_bufsize) {
1616 bp->b_xio.xio_npages = 0;
1617 bp->b_flags &= ~B_VMIO;
1618 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1629 * Implement clustered async writes for clearing out B_DELWRI buffers.
1630 * This is much better then the old way of writing only one buffer at
1631 * a time. Note that we may not be presented with the buffers in the
1632 * correct order, so we search for the cluster in both directions.
1634 * The buffer is locked on call.
1637 vfs_bio_awrite(struct buf *bp)
1641 off_t loffset = bp->b_loffset;
1642 struct vnode *vp = bp->b_vp;
1649 * right now we support clustered writing only to regular files. If
1650 * we find a clusterable block we could be in the middle of a cluster
1651 * rather then at the beginning.
1653 * NOTE: b_bio1 contains the logical loffset and is aliased
1654 * to b_loffset. b_bio2 contains the translated block number.
1656 if ((vp->v_type == VREG) &&
1657 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1658 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1660 size = vp->v_mount->mnt_stat.f_iosize;
1662 for (i = size; i < MAXPHYS; i += size) {
1663 if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) &&
1664 BUF_REFCNT(bpa) == 0 &&
1665 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1666 (B_DELWRI | B_CLUSTEROK)) &&
1667 (bpa->b_bufsize == size)) {
1668 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1669 (bpa->b_bio2.bio_offset !=
1670 bp->b_bio2.bio_offset + i))
1676 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1677 if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) &&
1678 BUF_REFCNT(bpa) == 0 &&
1679 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1680 (B_DELWRI | B_CLUSTEROK)) &&
1681 (bpa->b_bufsize == size)) {
1682 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1683 (bpa->b_bio2.bio_offset !=
1684 bp->b_bio2.bio_offset - j))
1694 * this is a possible cluster write
1696 if (nbytes != size) {
1698 nwritten = cluster_wbuild(vp, size,
1699 loffset - j, nbytes);
1705 bp->b_flags |= B_ASYNC;
1708 * default (old) behavior, writing out only one block
1710 * XXX returns b_bufsize instead of b_bcount for nwritten?
1712 nwritten = bp->b_bufsize;
1721 * Find and initialize a new buffer header, freeing up existing buffers
1722 * in the bufqueues as necessary. The new buffer is returned locked.
1724 * Important: B_INVAL is not set. If the caller wishes to throw the
1725 * buffer away, the caller must set B_INVAL prior to calling brelse().
1728 * We have insufficient buffer headers
1729 * We have insufficient buffer space
1730 * buffer_map is too fragmented ( space reservation fails )
1731 * If we have to flush dirty buffers ( but we try to avoid this )
1733 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1734 * Instead we ask the buf daemon to do it for us. We attempt to
1735 * avoid piecemeal wakeups of the pageout daemon.
1740 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1746 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1747 static int flushingbufs;
1750 * We can't afford to block since we might be holding a vnode lock,
1751 * which may prevent system daemons from running. We deal with
1752 * low-memory situations by proactively returning memory and running
1753 * async I/O rather then sync I/O.
1757 --getnewbufrestarts;
1759 ++getnewbufrestarts;
1762 * Setup for scan. If we do not have enough free buffers,
1763 * we setup a degenerate case that immediately fails. Note
1764 * that if we are specially marked process, we are allowed to
1765 * dip into our reserves.
1767 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1769 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1770 * However, there are a number of cases (defragging, reusing, ...)
1771 * where we cannot backup.
1773 nqindex = BQUEUE_EMPTYKVA;
1774 spin_lock_wr(&bufspin);
1775 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1779 * If no EMPTYKVA buffers and we are either
1780 * defragging or reusing, locate a CLEAN buffer
1781 * to free or reuse. If bufspace useage is low
1782 * skip this step so we can allocate a new buffer.
1784 if (defrag || bufspace >= lobufspace) {
1785 nqindex = BQUEUE_CLEAN;
1786 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1790 * If we could not find or were not allowed to reuse a
1791 * CLEAN buffer, check to see if it is ok to use an EMPTY
1792 * buffer. We can only use an EMPTY buffer if allocating
1793 * its KVA would not otherwise run us out of buffer space.
1795 if (nbp == NULL && defrag == 0 &&
1796 bufspace + maxsize < hibufspace) {
1797 nqindex = BQUEUE_EMPTY;
1798 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1803 * Run scan, possibly freeing data and/or kva mappings on the fly
1806 * WARNING! bufspin is held!
1808 while ((bp = nbp) != NULL) {
1809 int qindex = nqindex;
1811 nbp = TAILQ_NEXT(bp, b_freelist);
1814 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1815 * cycles through the queue twice before being selected.
1817 if (qindex == BQUEUE_CLEAN &&
1818 (bp->b_flags & B_AGE) == 0 && nbp) {
1819 bp->b_flags |= B_AGE;
1820 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1821 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1826 * Calculate next bp ( we can only use it if we do not block
1827 * or do other fancy things ).
1832 nqindex = BQUEUE_EMPTYKVA;
1833 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1836 case BQUEUE_EMPTYKVA:
1837 nqindex = BQUEUE_CLEAN;
1838 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1852 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1855 * Note: we no longer distinguish between VMIO and non-VMIO
1859 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1862 * If we are defragging then we need a buffer with
1863 * b_kvasize != 0. XXX this situation should no longer
1864 * occur, if defrag is non-zero the buffer's b_kvasize
1865 * should also be non-zero at this point. XXX
1867 if (defrag && bp->b_kvasize == 0) {
1868 kprintf("Warning: defrag empty buffer %p\n", bp);
1873 * Start freeing the bp. This is somewhat involved. nbp
1874 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1875 * on the clean list must be disassociated from their
1876 * current vnode. Buffers on the empty[kva] lists have
1877 * already been disassociated.
1880 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1881 spin_unlock_wr(&bufspin);
1882 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1883 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1886 if (bp->b_qindex != qindex) {
1887 spin_unlock_wr(&bufspin);
1888 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1892 bremfree_locked(bp);
1893 spin_unlock_wr(&bufspin);
1896 * Dependancies must be handled before we disassociate the
1899 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1900 * be immediately disassociated. HAMMER then becomes
1901 * responsible for releasing the buffer.
1903 * NOTE: bufspin is UNLOCKED now.
1905 if (LIST_FIRST(&bp->b_dep) != NULL) {
1909 if (bp->b_flags & B_LOCKED) {
1913 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1916 if (qindex == BQUEUE_CLEAN) {
1918 if (bp->b_flags & B_VMIO) {
1919 bp->b_flags &= ~B_ASYNC;
1921 vfs_vmio_release(bp);
1930 * NOTE: nbp is now entirely invalid. We can only restart
1931 * the scan from this point on.
1933 * Get the rest of the buffer freed up. b_kva* is still
1934 * valid after this operation.
1937 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1938 KKASSERT((bp->b_flags & B_HASHED) == 0);
1941 * critical section protection is not required when
1942 * scrapping a buffer's contents because it is already
1945 if (bp->b_bufsize) {
1951 bp->b_flags = B_BNOCLIP;
1952 bp->b_cmd = BUF_CMD_DONE;
1957 bp->b_xio.xio_npages = 0;
1958 bp->b_dirtyoff = bp->b_dirtyend = 0;
1960 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1962 if (blkflags & GETBLK_BHEAVY)
1963 bp->b_flags |= B_HEAVY;
1966 * If we are defragging then free the buffer.
1969 bp->b_flags |= B_INVAL;
1977 * If we are overcomitted then recover the buffer and its
1978 * KVM space. This occurs in rare situations when multiple
1979 * processes are blocked in getnewbuf() or allocbuf().
1981 if (bufspace >= hibufspace)
1983 if (flushingbufs && bp->b_kvasize != 0) {
1984 bp->b_flags |= B_INVAL;
1989 if (bufspace < lobufspace)
1992 /* NOT REACHED, bufspin not held */
1996 * If we exhausted our list, sleep as appropriate. We may have to
1997 * wakeup various daemons and write out some dirty buffers.
1999 * Generally we are sleeping due to insufficient buffer space.
2001 * NOTE: bufspin is held if bp is NULL, else it is not held.
2007 spin_unlock_wr(&bufspin);
2009 flags = VFS_BIO_NEED_BUFSPACE;
2011 } else if (bufspace >= hibufspace) {
2013 flags = VFS_BIO_NEED_BUFSPACE;
2016 flags = VFS_BIO_NEED_ANY;
2019 needsbuffer |= flags;
2020 bd_speedup(); /* heeeelp */
2021 while (needsbuffer & flags) {
2022 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
2027 * We finally have a valid bp. We aren't quite out of the
2028 * woods, we still have to reserve kva space. In order
2029 * to keep fragmentation sane we only allocate kva in
2032 * (bufspin is not held)
2034 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2036 if (maxsize != bp->b_kvasize) {
2037 vm_offset_t addr = 0;
2043 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2044 vm_map_lock(&buffer_map);
2046 if (vm_map_findspace(&buffer_map,
2047 vm_map_min(&buffer_map), maxsize,
2048 maxsize, 0, &addr)) {
2050 * Uh oh. Buffer map is too fragmented. We
2051 * must defragment the map.
2053 vm_map_unlock(&buffer_map);
2054 vm_map_entry_release(count);
2057 bp->b_flags |= B_INVAL;
2063 vm_map_insert(&buffer_map, &count,
2065 addr, addr + maxsize,
2067 VM_PROT_ALL, VM_PROT_ALL,
2070 bp->b_kvabase = (caddr_t) addr;
2071 bp->b_kvasize = maxsize;
2072 bufspace += bp->b_kvasize;
2075 vm_map_unlock(&buffer_map);
2076 vm_map_entry_release(count);
2079 bp->b_data = bp->b_kvabase;
2085 * This routine is called in an emergency to recover VM pages from the
2086 * buffer cache by cashing in clean buffers. The idea is to recover
2087 * enough pages to be able to satisfy a stuck bio_page_alloc().
2090 recoverbufpages(void)
2097 spin_lock_wr(&bufspin);
2098 while (bytes < MAXBSIZE) {
2099 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2104 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2105 * cycles through the queue twice before being selected.
2107 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
2108 bp->b_flags |= B_AGE;
2109 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
2110 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
2118 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
2119 KKASSERT((bp->b_flags & B_DELWRI) == 0);
2122 * Start freeing the bp. This is somewhat involved.
2124 * Buffers on the clean list must be disassociated from
2125 * their current vnode
2128 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2129 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp);
2130 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
2133 if (bp->b_qindex != BQUEUE_CLEAN) {
2134 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp, bp->b_qindex);
2138 bremfree_locked(bp);
2139 spin_unlock_wr(&bufspin);
2142 * Dependancies must be handled before we disassociate the
2145 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2146 * be immediately disassociated. HAMMER then becomes
2147 * responsible for releasing the buffer.
2149 if (LIST_FIRST(&bp->b_dep) != NULL) {
2151 if (bp->b_flags & B_LOCKED) {
2153 spin_lock_wr(&bufspin);
2156 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2159 bytes += bp->b_bufsize;
2162 if (bp->b_flags & B_VMIO) {
2163 bp->b_flags &= ~B_ASYNC;
2164 bp->b_flags |= B_DIRECT; /* try to free pages */
2165 vfs_vmio_release(bp);
2170 KKASSERT(bp->b_vp == NULL);
2171 KKASSERT((bp->b_flags & B_HASHED) == 0);
2174 * critical section protection is not required when
2175 * scrapping a buffer's contents because it is already
2182 bp->b_flags = B_BNOCLIP;
2183 bp->b_cmd = BUF_CMD_DONE;
2188 bp->b_xio.xio_npages = 0;
2189 bp->b_dirtyoff = bp->b_dirtyend = 0;
2191 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2193 bp->b_flags |= B_INVAL;
2196 spin_lock_wr(&bufspin);
2198 spin_unlock_wr(&bufspin);
2205 * Buffer flushing daemon. Buffers are normally flushed by the
2206 * update daemon but if it cannot keep up this process starts to
2207 * take the load in an attempt to prevent getnewbuf() from blocking.
2209 * Once a flush is initiated it does not stop until the number
2210 * of buffers falls below lodirtybuffers, but we will wake up anyone
2211 * waiting at the mid-point.
2214 static struct kproc_desc buf_kp = {
2219 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2220 kproc_start, &buf_kp)
2222 static struct kproc_desc bufhw_kp = {
2227 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2228 kproc_start, &bufhw_kp)
2236 * This process needs to be suspended prior to shutdown sync.
2238 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2239 bufdaemon_td, SHUTDOWN_PRI_LAST);
2240 curthread->td_flags |= TDF_SYSTHREAD;
2243 * This process is allowed to take the buffer cache to the limit
2248 kproc_suspend_loop();
2251 * Do the flush. Limit the amount of in-transit I/O we
2252 * allow to build up, otherwise we would completely saturate
2253 * the I/O system. Wakeup any waiting processes before we
2254 * normally would so they can run in parallel with our drain.
2256 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2257 * but because we split the operation into two threads we
2258 * have to cut it in half for each thread.
2260 limit = lodirtybufspace / 2;
2261 waitrunningbufspace(limit);
2262 while (runningbufspace + dirtybufspace > limit ||
2263 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
2264 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2266 waitrunningbufspace(limit);
2270 * We reached our low water mark, reset the
2271 * request and sleep until we are needed again.
2272 * The sleep is just so the suspend code works.
2274 spin_lock_wr(&needsbuffer_spin);
2275 if (bd_request == 0) {
2276 msleep(&bd_request, &needsbuffer_spin, 0,
2280 spin_unlock_wr(&needsbuffer_spin);
2290 * This process needs to be suspended prior to shutdown sync.
2292 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2293 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2294 curthread->td_flags |= TDF_SYSTHREAD;
2297 * This process is allowed to take the buffer cache to the limit
2302 kproc_suspend_loop();
2305 * Do the flush. Limit the amount of in-transit I/O we
2306 * allow to build up, otherwise we would completely saturate
2307 * the I/O system. Wakeup any waiting processes before we
2308 * normally would so they can run in parallel with our drain.
2310 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2311 * but because we split the operation into two threads we
2312 * have to cut it in half for each thread.
2314 limit = lodirtybufspace / 2;
2315 waitrunningbufspace(limit);
2316 while (runningbufspace + dirtybufspacehw > limit ||
2317 dirtybufcounthw >= nbuf / 2) {
2318 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2320 waitrunningbufspace(limit);
2324 * We reached our low water mark, reset the
2325 * request and sleep until we are needed again.
2326 * The sleep is just so the suspend code works.
2328 spin_lock_wr(&needsbuffer_spin);
2329 if (bd_request_hw == 0) {
2330 msleep(&bd_request_hw, &needsbuffer_spin, 0,
2334 spin_unlock_wr(&needsbuffer_spin);
2341 * Try to flush a buffer in the dirty queue. We must be careful to
2342 * free up B_INVAL buffers instead of write them, which NFS is
2343 * particularly sensitive to.
2345 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2346 * that we really want to try to get the buffer out and reuse it
2347 * due to the write load on the machine.
2350 flushbufqueues(bufq_type_t q)
2356 spin_lock_wr(&bufspin);
2359 bp = TAILQ_FIRST(&bufqueues[q]);
2361 KASSERT((bp->b_flags & B_DELWRI),
2362 ("unexpected clean buffer %p", bp));
2364 if (bp->b_flags & B_DELWRI) {
2365 if (bp->b_flags & B_INVAL) {
2366 spin_unlock_wr(&bufspin);
2368 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2369 panic("flushbufqueues: locked buf");
2375 if (LIST_FIRST(&bp->b_dep) != NULL &&
2376 (bp->b_flags & B_DEFERRED) == 0 &&
2377 buf_countdeps(bp, 0)) {
2378 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2379 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2381 bp->b_flags |= B_DEFERRED;
2382 bp = TAILQ_FIRST(&bufqueues[q]);
2387 * Only write it out if we can successfully lock
2388 * it. If the buffer has a dependancy,
2389 * buf_checkwrite must also return 0 for us to
2390 * be able to initate the write.
2392 * If the buffer is flagged B_ERROR it may be
2393 * requeued over and over again, we try to
2394 * avoid a live lock.
2396 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
2397 spin_unlock_wr(&bufspin);
2399 if (LIST_FIRST(&bp->b_dep) != NULL &&
2400 buf_checkwrite(bp)) {
2403 } else if (bp->b_flags & B_ERROR) {
2404 tsleep(bp, 0, "bioer", 1);
2405 bp->b_flags &= ~B_AGE;
2408 bp->b_flags |= B_AGE;
2415 bp = TAILQ_NEXT(bp, b_freelist);
2418 spin_unlock_wr(&bufspin);
2425 * Returns true if no I/O is needed to access the associated VM object.
2426 * This is like findblk except it also hunts around in the VM system for
2429 * Note that we ignore vm_page_free() races from interrupts against our
2430 * lookup, since if the caller is not protected our return value will not
2431 * be any more valid then otherwise once we exit the critical section.
2434 inmem(struct vnode *vp, off_t loffset)
2437 vm_offset_t toff, tinc, size;
2440 if (findblk(vp, loffset, FINDBLK_TEST))
2442 if (vp->v_mount == NULL)
2444 if ((obj = vp->v_object) == NULL)
2448 if (size > vp->v_mount->mnt_stat.f_iosize)
2449 size = vp->v_mount->mnt_stat.f_iosize;
2451 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2452 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2456 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2457 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2458 if (vm_page_is_valid(m,
2459 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2468 * Sets the dirty range for a buffer based on the status of the dirty
2469 * bits in the pages comprising the buffer.
2471 * The range is limited to the size of the buffer.
2473 * This routine is primarily used by NFS, but is generalized for the
2477 vfs_setdirty(struct buf *bp)
2483 * Degenerate case - empty buffer
2486 if (bp->b_bufsize == 0)
2490 * We qualify the scan for modified pages on whether the
2491 * object has been flushed yet. The OBJ_WRITEABLE flag
2492 * is not cleared simply by protecting pages off.
2495 if ((bp->b_flags & B_VMIO) == 0)
2498 object = bp->b_xio.xio_pages[0]->object;
2500 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2501 kprintf("Warning: object %p writeable but not mightbedirty\n", object);
2502 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2503 kprintf("Warning: object %p mightbedirty but not writeable\n", object);
2505 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2506 vm_offset_t boffset;
2507 vm_offset_t eoffset;
2510 * test the pages to see if they have been modified directly
2511 * by users through the VM system.
2513 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2514 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2515 vm_page_test_dirty(bp->b_xio.xio_pages[i]);
2519 * Calculate the encompassing dirty range, boffset and eoffset,
2520 * (eoffset - boffset) bytes.
2523 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2524 if (bp->b_xio.xio_pages[i]->dirty)
2527 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2529 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2530 if (bp->b_xio.xio_pages[i]->dirty) {
2534 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2537 * Fit it to the buffer.
2540 if (eoffset > bp->b_bcount)
2541 eoffset = bp->b_bcount;
2544 * If we have a good dirty range, merge with the existing
2548 if (boffset < eoffset) {
2549 if (bp->b_dirtyoff > boffset)
2550 bp->b_dirtyoff = boffset;
2551 if (bp->b_dirtyend < eoffset)
2552 bp->b_dirtyend = eoffset;
2560 * Locate and return the specified buffer. Unless flagged otherwise,
2561 * a locked buffer will be returned if it exists or NULL if it does not.
2563 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2564 * for locking the buffer and ensuring that it remains
2565 * the desired buffer after locking.
2567 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2568 * to acquire the lock we return NULL, even if the
2571 * (0) - Lock the buffer blocking.
2576 findblk(struct vnode *vp, off_t loffset, int flags)
2582 lkflags = LK_EXCLUSIVE;
2583 if (flags & FINDBLK_NBLOCK)
2584 lkflags |= LK_NOWAIT;
2587 lwkt_gettoken(&vlock, &vp->v_token);
2588 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2589 lwkt_reltoken(&vlock);
2590 if (bp == NULL || (flags & FINDBLK_TEST))
2592 if (BUF_LOCK(bp, lkflags)) {
2596 if (bp->b_vp == vp && bp->b_loffset == loffset)
2606 * Similar to getblk() except only returns the buffer if it is
2607 * B_CACHE and requires no other manipulation. Otherwise NULL
2610 * If B_RAM is set the buffer might be just fine, but we return
2611 * NULL anyway because we want the code to fall through to the
2612 * cluster read. Otherwise read-ahead breaks.
2615 getcacheblk(struct vnode *vp, off_t loffset)
2619 bp = findblk(vp, loffset, 0);
2621 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
2622 bp->b_flags &= ~B_AGE;
2635 * Get a block given a specified block and offset into a file/device.
2636 * B_INVAL may or may not be set on return. The caller should clear
2637 * B_INVAL prior to initiating a READ.
2639 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2640 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2641 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2642 * without doing any of those things the system will likely believe
2643 * the buffer to be valid (especially if it is not B_VMIO), and the
2644 * next getblk() will return the buffer with B_CACHE set.
2646 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2647 * an existing buffer.
2649 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2650 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2651 * and then cleared based on the backing VM. If the previous buffer is
2652 * non-0-sized but invalid, B_CACHE will be cleared.
2654 * If getblk() must create a new buffer, the new buffer is returned with
2655 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2656 * case it is returned with B_INVAL clear and B_CACHE set based on the
2659 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2660 * B_CACHE bit is clear.
2662 * What this means, basically, is that the caller should use B_CACHE to
2663 * determine whether the buffer is fully valid or not and should clear
2664 * B_INVAL prior to issuing a read. If the caller intends to validate
2665 * the buffer by loading its data area with something, the caller needs
2666 * to clear B_INVAL. If the caller does this without issuing an I/O,
2667 * the caller should set B_CACHE ( as an optimization ), else the caller
2668 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2669 * a write attempt or if it was a successfull read. If the caller
2670 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2671 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2675 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2676 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2681 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2684 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2688 if (size > MAXBSIZE)
2689 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2690 if (vp->v_object == NULL)
2691 panic("getblk: vnode %p has no object!", vp);
2694 if ((bp = findblk(vp, loffset, FINDBLK_TEST)) != NULL) {
2696 * The buffer was found in the cache, but we need to lock it.
2697 * Even with LK_NOWAIT the lockmgr may break our critical
2698 * section, so double-check the validity of the buffer
2699 * once the lock has been obtained.
2701 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2702 if (blkflags & GETBLK_NOWAIT)
2704 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2705 if (blkflags & GETBLK_PCATCH)
2706 lkflags |= LK_PCATCH;
2707 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2709 if (error == ENOLCK)
2713 /* buffer may have changed on us */
2717 * Once the buffer has been locked, make sure we didn't race
2718 * a buffer recyclement. Buffers that are no longer hashed
2719 * will have b_vp == NULL, so this takes care of that check
2722 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2723 kprintf("Warning buffer %p (vp %p loffset %lld) "
2725 bp, vp, (long long)loffset);
2731 * If SZMATCH any pre-existing buffer must be of the requested
2732 * size or NULL is returned. The caller absolutely does not
2733 * want getblk() to bwrite() the buffer on a size mismatch.
2735 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2741 * All vnode-based buffers must be backed by a VM object.
2743 KKASSERT(bp->b_flags & B_VMIO);
2744 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2745 bp->b_flags &= ~B_AGE;
2748 * Make sure that B_INVAL buffers do not have a cached
2749 * block number translation.
2751 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2752 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2753 " did not have cleared bio_offset cache\n",
2754 bp, vp, (long long)loffset);
2755 clearbiocache(&bp->b_bio2);
2759 * The buffer is locked. B_CACHE is cleared if the buffer is
2762 if (bp->b_flags & B_INVAL)
2763 bp->b_flags &= ~B_CACHE;
2767 * Any size inconsistancy with a dirty buffer or a buffer
2768 * with a softupdates dependancy must be resolved. Resizing
2769 * the buffer in such circumstances can lead to problems.
2771 if (size != bp->b_bcount) {
2773 if (bp->b_flags & B_DELWRI) {
2774 bp->b_flags |= B_NOCACHE;
2776 } else if (LIST_FIRST(&bp->b_dep)) {
2777 bp->b_flags |= B_NOCACHE;
2780 bp->b_flags |= B_RELBUF;
2786 KKASSERT(size <= bp->b_kvasize);
2787 KASSERT(bp->b_loffset != NOOFFSET,
2788 ("getblk: no buffer offset"));
2791 * A buffer with B_DELWRI set and B_CACHE clear must
2792 * be committed before we can return the buffer in
2793 * order to prevent the caller from issuing a read
2794 * ( due to B_CACHE not being set ) and overwriting
2797 * Most callers, including NFS and FFS, need this to
2798 * operate properly either because they assume they
2799 * can issue a read if B_CACHE is not set, or because
2800 * ( for example ) an uncached B_DELWRI might loop due
2801 * to softupdates re-dirtying the buffer. In the latter
2802 * case, B_CACHE is set after the first write completes,
2803 * preventing further loops.
2805 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2806 * above while extending the buffer, we cannot allow the
2807 * buffer to remain with B_CACHE set after the write
2808 * completes or it will represent a corrupt state. To
2809 * deal with this we set B_NOCACHE to scrap the buffer
2812 * We might be able to do something fancy, like setting
2813 * B_CACHE in bwrite() except if B_DELWRI is already set,
2814 * so the below call doesn't set B_CACHE, but that gets real
2815 * confusing. This is much easier.
2818 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2820 bp->b_flags |= B_NOCACHE;
2827 * Buffer is not in-core, create new buffer. The buffer
2828 * returned by getnewbuf() is locked. Note that the returned
2829 * buffer is also considered valid (not marked B_INVAL).
2831 * Calculating the offset for the I/O requires figuring out
2832 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2833 * the mount's f_iosize otherwise. If the vnode does not
2834 * have an associated mount we assume that the passed size is
2837 * Note that vn_isdisk() cannot be used here since it may
2838 * return a failure for numerous reasons. Note that the
2839 * buffer size may be larger then the block size (the caller
2840 * will use block numbers with the proper multiple). Beware
2841 * of using any v_* fields which are part of unions. In
2842 * particular, in DragonFly the mount point overloading
2843 * mechanism uses the namecache only and the underlying
2844 * directory vnode is not a special case.
2848 if (vp->v_type == VBLK || vp->v_type == VCHR)
2850 else if (vp->v_mount)
2851 bsize = vp->v_mount->mnt_stat.f_iosize;
2855 maxsize = size + (loffset & PAGE_MASK);
2856 maxsize = imax(maxsize, bsize);
2858 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
2860 if (slpflags || slptimeo)
2866 * Atomically insert the buffer into the hash, so that it can
2867 * be found by findblk().
2869 * If bgetvp() returns non-zero a collision occured, and the
2870 * bp will not be associated with the vnode.
2872 * Make sure the translation layer has been cleared.
2874 bp->b_loffset = loffset;
2875 bp->b_bio2.bio_offset = NOOFFSET;
2876 /* bp->b_bio2.bio_next = NULL; */
2878 if (bgetvp(vp, bp)) {
2879 bp->b_flags |= B_INVAL;
2885 * All vnode-based buffers must be backed by a VM object.
2887 KKASSERT(vp->v_object != NULL);
2888 bp->b_flags |= B_VMIO;
2889 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2901 * Reacquire a buffer that was previously released to the locked queue,
2902 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2903 * set B_LOCKED (which handles the acquisition race).
2905 * To this end, either B_LOCKED must be set or the dependancy list must be
2911 regetblk(struct buf *bp)
2913 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
2914 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2921 * Get an empty, disassociated buffer of given size. The buffer is
2922 * initially set to B_INVAL.
2924 * critical section protection is not required for the allocbuf()
2925 * call because races are impossible here.
2935 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2937 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2942 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2950 * This code constitutes the buffer memory from either anonymous system
2951 * memory (in the case of non-VMIO operations) or from an associated
2952 * VM object (in the case of VMIO operations). This code is able to
2953 * resize a buffer up or down.
2955 * Note that this code is tricky, and has many complications to resolve
2956 * deadlock or inconsistant data situations. Tread lightly!!!
2957 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2958 * the caller. Calling this code willy nilly can result in the loss of data.
2960 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2961 * B_CACHE for the non-VMIO case.
2963 * This routine does not need to be called from a critical section but you
2964 * must own the buffer.
2969 allocbuf(struct buf *bp, int size)
2971 int newbsize, mbsize;
2974 if (BUF_REFCNT(bp) == 0)
2975 panic("allocbuf: buffer not busy");
2977 if (bp->b_kvasize < size)
2978 panic("allocbuf: buffer too small");
2980 if ((bp->b_flags & B_VMIO) == 0) {
2984 * Just get anonymous memory from the kernel. Don't
2985 * mess with B_CACHE.
2987 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2988 if (bp->b_flags & B_MALLOC)
2991 newbsize = round_page(size);
2993 if (newbsize < bp->b_bufsize) {
2995 * Malloced buffers are not shrunk
2997 if (bp->b_flags & B_MALLOC) {
2999 bp->b_bcount = size;
3001 kfree(bp->b_data, M_BIOBUF);
3002 if (bp->b_bufsize) {
3003 bufmallocspace -= bp->b_bufsize;
3007 bp->b_data = bp->b_kvabase;
3009 bp->b_flags &= ~B_MALLOC;
3015 (vm_offset_t) bp->b_data + newbsize,
3016 (vm_offset_t) bp->b_data + bp->b_bufsize);
3017 } else if (newbsize > bp->b_bufsize) {
3019 * We only use malloced memory on the first allocation.
3020 * and revert to page-allocated memory when the buffer
3023 if ((bufmallocspace < maxbufmallocspace) &&
3024 (bp->b_bufsize == 0) &&
3025 (mbsize <= PAGE_SIZE/2)) {
3027 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3028 bp->b_bufsize = mbsize;
3029 bp->b_bcount = size;
3030 bp->b_flags |= B_MALLOC;
3031 bufmallocspace += mbsize;
3037 * If the buffer is growing on its other-than-first
3038 * allocation, then we revert to the page-allocation
3041 if (bp->b_flags & B_MALLOC) {
3042 origbuf = bp->b_data;
3043 origbufsize = bp->b_bufsize;
3044 bp->b_data = bp->b_kvabase;
3045 if (bp->b_bufsize) {
3046 bufmallocspace -= bp->b_bufsize;
3050 bp->b_flags &= ~B_MALLOC;
3051 newbsize = round_page(newbsize);
3055 (vm_offset_t) bp->b_data + bp->b_bufsize,
3056 (vm_offset_t) bp->b_data + newbsize);
3058 bcopy(origbuf, bp->b_data, origbufsize);
3059 kfree(origbuf, M_BIOBUF);
3066 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3067 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3068 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3069 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3071 if (bp->b_flags & B_MALLOC)
3072 panic("allocbuf: VMIO buffer can't be malloced");
3074 * Set B_CACHE initially if buffer is 0 length or will become
3077 if (size == 0 || bp->b_bufsize == 0)
3078 bp->b_flags |= B_CACHE;
3080 if (newbsize < bp->b_bufsize) {
3082 * DEV_BSIZE aligned new buffer size is less then the
3083 * DEV_BSIZE aligned existing buffer size. Figure out
3084 * if we have to remove any pages.
3086 if (desiredpages < bp->b_xio.xio_npages) {
3087 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3089 * the page is not freed here -- it
3090 * is the responsibility of
3091 * vnode_pager_setsize
3093 m = bp->b_xio.xio_pages[i];
3094 KASSERT(m != bogus_page,
3095 ("allocbuf: bogus page found"));
3096 while (vm_page_sleep_busy(m, TRUE, "biodep"))
3099 bp->b_xio.xio_pages[i] = NULL;
3100 vm_page_unwire(m, 0);
3102 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3103 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3104 bp->b_xio.xio_npages = desiredpages;
3106 } else if (size > bp->b_bcount) {
3108 * We are growing the buffer, possibly in a
3109 * byte-granular fashion.
3117 * Step 1, bring in the VM pages from the object,
3118 * allocating them if necessary. We must clear
3119 * B_CACHE if these pages are not valid for the
3120 * range covered by the buffer.
3122 * critical section protection is required to protect
3123 * against interrupts unbusying and freeing pages
3124 * between our vm_page_lookup() and our
3125 * busycheck/wiring call.
3131 while (bp->b_xio.xio_npages < desiredpages) {
3135 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
3136 if ((m = vm_page_lookup(obj, pi)) == NULL) {
3138 * note: must allocate system pages
3139 * since blocking here could intefere
3140 * with paging I/O, no matter which
3143 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
3147 bp->b_flags &= ~B_CACHE;
3148 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3149 ++bp->b_xio.xio_npages;
3155 * We found a page. If we have to sleep on it,
3156 * retry because it might have gotten freed out
3159 * We can only test PG_BUSY here. Blocking on
3160 * m->busy might lead to a deadlock:
3162 * vm_fault->getpages->cluster_read->allocbuf
3166 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
3168 vm_page_flag_clear(m, PG_ZERO);
3170 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3171 ++bp->b_xio.xio_npages;
3176 * Step 2. We've loaded the pages into the buffer,
3177 * we have to figure out if we can still have B_CACHE
3178 * set. Note that B_CACHE is set according to the
3179 * byte-granular range ( bcount and size ), not the
3180 * aligned range ( newbsize ).
3182 * The VM test is against m->valid, which is DEV_BSIZE
3183 * aligned. Needless to say, the validity of the data
3184 * needs to also be DEV_BSIZE aligned. Note that this
3185 * fails with NFS if the server or some other client
3186 * extends the file's EOF. If our buffer is resized,
3187 * B_CACHE may remain set! XXX
3190 toff = bp->b_bcount;
3191 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3193 while ((bp->b_flags & B_CACHE) && toff < size) {
3196 if (tinc > (size - toff))
3199 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3207 bp->b_xio.xio_pages[pi]
3214 * Step 3, fixup the KVM pmap. Remember that
3215 * bp->b_data is relative to bp->b_loffset, but
3216 * bp->b_loffset may be offset into the first page.
3219 bp->b_data = (caddr_t)
3220 trunc_page((vm_offset_t)bp->b_data);
3222 (vm_offset_t)bp->b_data,
3223 bp->b_xio.xio_pages,
3224 bp->b_xio.xio_npages
3226 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3227 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3231 /* adjust space use on already-dirty buffer */
3232 if (bp->b_flags & B_DELWRI) {
3233 dirtybufspace += newbsize - bp->b_bufsize;
3234 if (bp->b_flags & B_HEAVY)
3235 dirtybufspacehw += newbsize - bp->b_bufsize;
3237 if (newbsize < bp->b_bufsize)
3239 bp->b_bufsize = newbsize; /* actual buffer allocation */
3240 bp->b_bcount = size; /* requested buffer size */
3247 * Wait for buffer I/O completion, returning error status. The buffer
3248 * is left locked on return. B_EINTR is converted into an EINTR error
3251 * NOTE! The original b_cmd is lost on return, since b_cmd will be
3252 * set to BUF_CMD_DONE.
3257 biowait(struct buf *bp)
3259 if (bp->b_cmd != BUF_CMD_DONE) {
3262 tsleep_interlock(bp);
3263 if (bp->b_cmd == BUF_CMD_DONE)
3265 if (bp->b_cmd == BUF_CMD_READ)
3266 tsleep(bp, PINTERLOCKED, "biord", 0);
3268 tsleep(bp, PINTERLOCKED, "biowr", 0);
3272 if (bp->b_flags & B_EINTR) {
3273 bp->b_flags &= ~B_EINTR;
3276 if (bp->b_flags & B_ERROR) {
3277 return (bp->b_error ? bp->b_error : EIO);
3284 * This associates a tracking count with an I/O. vn_strategy() and
3285 * dev_dstrategy() do this automatically but there are a few cases
3286 * where a vnode or device layer is bypassed when a block translation
3287 * is cached. In such cases bio_start_transaction() may be called on
3288 * the bypassed layers so the system gets an I/O in progress indication
3289 * for those higher layers.
3292 bio_start_transaction(struct bio *bio, struct bio_track *track)
3294 bio->bio_track = track;
3295 bio_track_ref(track);
3299 * Initiate I/O on a vnode.
3302 vn_strategy(struct vnode *vp, struct bio *bio)
3304 struct bio_track *track;
3306 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
3307 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
3308 track = &vp->v_track_read;
3310 track = &vp->v_track_write;
3311 bio->bio_track = track;
3312 bio_track_ref(track);
3313 vop_strategy(*vp->v_ops, vp, bio);
3319 * Finish I/O on a buffer, optionally calling a completion function.
3320 * This is usually called from an interrupt so process blocking is
3323 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3324 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3325 * assuming B_INVAL is clear.
3327 * For the VMIO case, we set B_CACHE if the op was a read and no
3328 * read error occured, or if the op was a write. B_CACHE is never
3329 * set if the buffer is invalid or otherwise uncacheable.
3331 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3332 * initiator to leave B_INVAL set to brelse the buffer out of existance
3333 * in the biodone routine.
3336 biodone(struct bio *bio)
3338 struct buf *bp = bio->bio_buf;
3343 KASSERT(BUF_REFCNTNB(bp) > 0,
3344 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3345 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3346 ("biodone: bp %p already done!", bp));
3348 runningbufwakeup(bp);
3351 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3354 biodone_t *done_func;
3355 struct bio_track *track;
3358 * BIO tracking. Most but not all BIOs are tracked.
3360 if ((track = bio->bio_track) != NULL) {
3361 bio_track_rel(track);
3362 bio->bio_track = NULL;
3366 * A bio_done function terminates the loop. The function
3367 * will be responsible for any further chaining and/or
3368 * buffer management.
3370 * WARNING! The done function can deallocate the buffer!
3372 if ((done_func = bio->bio_done) != NULL) {
3373 bio->bio_done = NULL;
3378 bio = bio->bio_prev;
3382 bp->b_cmd = BUF_CMD_DONE;
3385 * Only reads and writes are processed past this point.
3387 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3388 if (cmd == BUF_CMD_FREEBLKS)
3389 bp->b_flags |= B_NOCACHE;
3396 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3397 * a lot worse. XXX - move this above the clearing of b_cmd
3399 if (LIST_FIRST(&bp->b_dep) != NULL)
3403 * A failed write must re-dirty the buffer unless B_INVAL
3406 if (cmd == BUF_CMD_WRITE &&
3407 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3408 bp->b_flags &= ~B_NOCACHE;
3413 if (bp->b_flags & B_VMIO) {
3419 struct vnode *vp = bp->b_vp;
3423 #if defined(VFS_BIO_DEBUG)
3424 if (vp->v_auxrefs == 0)
3425 panic("biodone: zero vnode hold count");
3426 if ((vp->v_flag & VOBJBUF) == 0)
3427 panic("biodone: vnode is not setup for merged cache");
3430 foff = bp->b_loffset;
3431 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3432 KASSERT(obj != NULL, ("biodone: missing VM object"));
3434 #if defined(VFS_BIO_DEBUG)
3435 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3436 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3437 obj->paging_in_progress, bp->b_xio.xio_npages);
3442 * Set B_CACHE if the op was a normal read and no error
3443 * occured. B_CACHE is set for writes in the b*write()
3446 iosize = bp->b_bcount - bp->b_resid;
3447 if (cmd == BUF_CMD_READ && (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3448 bp->b_flags |= B_CACHE;
3451 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3455 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3460 * cleanup bogus pages, restoring the originals. Since
3461 * the originals should still be wired, we don't have
3462 * to worry about interrupt/freeing races destroying
3463 * the VM object association.
3465 m = bp->b_xio.xio_pages[i];
3466 if (m == bogus_page) {
3468 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3470 panic("biodone: page disappeared");
3471 bp->b_xio.xio_pages[i] = m;
3472 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3473 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3475 #if defined(VFS_BIO_DEBUG)
3476 if (OFF_TO_IDX(foff) != m->pindex) {
3478 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
3479 (unsigned long)foff, m->pindex);
3484 * In the write case, the valid and clean bits are
3485 * already changed correctly ( see bdwrite() ), so we
3486 * only need to do this here in the read case.
3488 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3489 vfs_page_set_valid(bp, foff, i, m);
3491 vm_page_flag_clear(m, PG_ZERO);
3494 * when debugging new filesystems or buffer I/O methods, this
3495 * is the most common error that pops up. if you see this, you
3496 * have not set the page busy flag correctly!!!
3499 kprintf("biodone: page busy < 0, "
3500 "pindex: %d, foff: 0x(%x,%x), "
3501 "resid: %d, index: %d\n",
3502 (int) m->pindex, (int)(foff >> 32),
3503 (int) foff & 0xffffffff, resid, i);
3504 if (!vn_isdisk(vp, NULL))
3505 kprintf(" iosize: %ld, loffset: %lld, "
3506 "flags: 0x%08x, npages: %d\n",
3507 bp->b_vp->v_mount->mnt_stat.f_iosize,
3508 (long long)bp->b_loffset,
3509 bp->b_flags, bp->b_xio.xio_npages);
3511 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3512 (long long)bp->b_loffset,
3513 bp->b_flags, bp->b_xio.xio_npages);
3514 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3515 m->valid, m->dirty, m->wire_count);
3516 panic("biodone: page busy < 0");
3518 vm_page_io_finish(m);
3519 vm_object_pip_subtract(obj, 1);
3520 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3524 vm_object_pip_wakeupn(obj, 0);
3528 * For asynchronous completions, release the buffer now. The brelse
3529 * will do a wakeup there if necessary - so no need to do a wakeup
3530 * here in the async case. The sync case always needs to do a wakeup.
3533 if (bp->b_flags & B_ASYNC) {
3534 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
3547 * This routine is called in lieu of iodone in the case of
3548 * incomplete I/O. This keeps the busy status for pages
3552 vfs_unbusy_pages(struct buf *bp)
3556 runningbufwakeup(bp);
3557 if (bp->b_flags & B_VMIO) {
3558 struct vnode *vp = bp->b_vp;
3563 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3564 vm_page_t m = bp->b_xio.xio_pages[i];
3567 * When restoring bogus changes the original pages
3568 * should still be wired, so we are in no danger of
3569 * losing the object association and do not need
3570 * critical section protection particularly.
3572 if (m == bogus_page) {
3573 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3575 panic("vfs_unbusy_pages: page missing");
3577 bp->b_xio.xio_pages[i] = m;
3578 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3579 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3581 vm_object_pip_subtract(obj, 1);
3582 vm_page_flag_clear(m, PG_ZERO);
3583 vm_page_io_finish(m);
3585 vm_object_pip_wakeupn(obj, 0);
3590 * vfs_page_set_valid:
3592 * Set the valid bits in a page based on the supplied offset. The
3593 * range is restricted to the buffer's size.
3595 * This routine is typically called after a read completes.
3598 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3600 vm_ooffset_t soff, eoff;
3603 * Start and end offsets in buffer. eoff - soff may not cross a
3604 * page boundry or cross the end of the buffer. The end of the
3605 * buffer, in this case, is our file EOF, not the allocation size
3609 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3610 if (eoff > bp->b_loffset + bp->b_bcount)
3611 eoff = bp->b_loffset + bp->b_bcount;
3614 * Set valid range. This is typically the entire buffer and thus the
3618 vm_page_set_validclean(
3620 (vm_offset_t) (soff & PAGE_MASK),
3621 (vm_offset_t) (eoff - soff)
3629 * This routine is called before a device strategy routine.
3630 * It is used to tell the VM system that paging I/O is in
3631 * progress, and treat the pages associated with the buffer
3632 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3633 * flag is handled to make sure that the object doesn't become
3636 * Since I/O has not been initiated yet, certain buffer flags
3637 * such as B_ERROR or B_INVAL may be in an inconsistant state
3638 * and should be ignored.
3641 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3644 struct lwp *lp = curthread->td_lwp;
3647 * The buffer's I/O command must already be set. If reading,
3648 * B_CACHE must be 0 (double check against callers only doing
3649 * I/O when B_CACHE is 0).
3651 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3652 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3654 if (bp->b_flags & B_VMIO) {
3659 foff = bp->b_loffset;
3660 KASSERT(bp->b_loffset != NOOFFSET,
3661 ("vfs_busy_pages: no buffer offset"));
3665 * Loop until none of the pages are busy.
3668 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3669 vm_page_t m = bp->b_xio.xio_pages[i];
3671 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3676 * Setup for I/O, soft-busy the page right now because
3677 * the next loop may block.
3679 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3680 vm_page_t m = bp->b_xio.xio_pages[i];
3682 vm_page_flag_clear(m, PG_ZERO);
3683 if ((bp->b_flags & B_CLUSTER) == 0) {
3684 vm_object_pip_add(obj, 1);
3685 vm_page_io_start(m);
3690 * Adjust protections for I/O and do bogus-page mapping.
3691 * Assume that vm_page_protect() can block (it can block
3692 * if VM_PROT_NONE, don't take any chances regardless).
3695 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3696 vm_page_t m = bp->b_xio.xio_pages[i];
3699 * When readying a vnode-backed buffer for a write
3700 * we must zero-fill any invalid portions of the
3703 * When readying a vnode-backed buffer for a read
3704 * we must replace any dirty pages with a bogus
3705 * page so we do not destroy dirty data when
3706 * filling in gaps. Dirty pages might not
3707 * necessarily be marked dirty yet, so use m->valid
3708 * as a reasonable test.
3710 * Bogus page replacement is, uh, bogus. We need
3711 * to find a better way.
3713 if (bp->b_cmd == BUF_CMD_WRITE) {
3714 vm_page_protect(m, VM_PROT_READ);
3715 vfs_page_set_valid(bp, foff, i, m);
3716 } else if (m->valid == VM_PAGE_BITS_ALL) {
3717 bp->b_xio.xio_pages[i] = bogus_page;
3720 vm_page_protect(m, VM_PROT_NONE);
3722 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3725 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3726 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3730 * This is the easiest place to put the process accounting for the I/O
3734 if (bp->b_cmd == BUF_CMD_READ)
3735 lp->lwp_ru.ru_inblock++;
3737 lp->lwp_ru.ru_oublock++;
3744 * Tell the VM system that the pages associated with this buffer
3745 * are clean. This is used for delayed writes where the data is
3746 * going to go to disk eventually without additional VM intevention.
3748 * Note that while we only really need to clean through to b_bcount, we
3749 * just go ahead and clean through to b_bufsize.
3752 vfs_clean_pages(struct buf *bp)
3756 if (bp->b_flags & B_VMIO) {
3759 foff = bp->b_loffset;
3760 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset"));
3761 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3762 vm_page_t m = bp->b_xio.xio_pages[i];
3763 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3765 vfs_page_set_valid(bp, foff, i, m);
3772 * vfs_bio_set_validclean:
3774 * Set the range within the buffer to valid and clean. The range is
3775 * relative to the beginning of the buffer, b_loffset. Note that
3776 * b_loffset itself may be offset from the beginning of the first page.
3780 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3782 if (bp->b_flags & B_VMIO) {
3787 * Fixup base to be relative to beginning of first page.
3788 * Set initial n to be the maximum number of bytes in the
3789 * first page that can be validated.
3792 base += (bp->b_loffset & PAGE_MASK);
3793 n = PAGE_SIZE - (base & PAGE_MASK);
3795 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) {
3796 vm_page_t m = bp->b_xio.xio_pages[i];
3801 vm_page_set_validclean(m, base & PAGE_MASK, n);
3812 * Clear a buffer. This routine essentially fakes an I/O, so we need
3813 * to clear B_ERROR and B_INVAL.
3815 * Note that while we only theoretically need to clear through b_bcount,
3816 * we go ahead and clear through b_bufsize.
3820 vfs_bio_clrbuf(struct buf *bp)
3824 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3825 bp->b_flags &= ~(B_INVAL|B_ERROR);
3826 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3827 (bp->b_loffset & PAGE_MASK) == 0) {
3828 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3829 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3833 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3834 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3835 bzero(bp->b_data, bp->b_bufsize);
3836 bp->b_xio.xio_pages[0]->valid |= mask;
3842 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3843 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3844 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3845 ea = (caddr_t)(vm_offset_t)ulmin(
3846 (u_long)(vm_offset_t)ea,
3847 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3848 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3849 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3851 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3852 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3856 for (; sa < ea; sa += DEV_BSIZE, j++) {
3857 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3858 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3859 bzero(sa, DEV_BSIZE);
3862 bp->b_xio.xio_pages[i]->valid |= mask;
3863 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
3872 * vm_hold_load_pages:
3874 * Load pages into the buffer's address space. The pages are
3875 * allocated from the kernel object in order to reduce interference
3876 * with the any VM paging I/O activity. The range of loaded
3877 * pages will be wired.
3879 * If a page cannot be allocated, the 'pagedaemon' is woken up to
3880 * retrieve the full range (to - from) of pages.
3884 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3890 to = round_page(to);
3891 from = round_page(from);
3892 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3897 * Note: must allocate system pages since blocking here
3898 * could intefere with paging I/O, no matter which
3901 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT,
3902 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
3905 p->valid = VM_PAGE_BITS_ALL;
3906 vm_page_flag_clear(p, PG_ZERO);
3907 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3908 bp->b_xio.xio_pages[index] = p;
3915 bp->b_xio.xio_npages = index;
3919 * Allocate pages for a buffer cache buffer.
3921 * Under extremely severe memory conditions even allocating out of the
3922 * system reserve can fail. If this occurs we must allocate out of the
3923 * interrupt reserve to avoid a deadlock with the pageout daemon.
3925 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
3926 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
3927 * against the pageout daemon if pages are not freed from other sources.
3931 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit)
3936 * Try a normal allocation, allow use of system reserve.
3938 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
3943 * The normal allocation failed and we clearly have a page
3944 * deficit. Try to reclaim some clean VM pages directly
3945 * from the buffer cache.
3947 vm_pageout_deficit += deficit;
3951 * We may have blocked, the caller will know what to do if the
3954 if (vm_page_lookup(obj, pg))
3958 * Allocate and allow use of the interrupt reserve.
3960 * If after all that we still can't allocate a VM page we are
3961 * in real trouble, but we slog on anyway hoping that the system
3964 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
3965 VM_ALLOC_INTERRUPT);
3967 if (vm_page_count_severe()) {
3968 kprintf("bio_page_alloc: WARNING emergency page "
3973 kprintf("bio_page_alloc: WARNING emergency page "
3974 "allocation failed\n");
3981 * vm_hold_free_pages:
3983 * Return pages associated with the buffer back to the VM system.
3985 * The range of pages underlying the buffer's address space will
3986 * be unmapped and un-wired.
3989 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3993 int index, newnpages;
3995 from = round_page(from);
3996 to = round_page(to);
3997 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3999 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4000 p = bp->b_xio.xio_pages[index];
4001 if (p && (index < bp->b_xio.xio_npages)) {
4003 kprintf("vm_hold_free_pages: doffset: %lld, "
4005 (long long)bp->b_bio2.bio_offset,
4006 (long long)bp->b_loffset);
4008 bp->b_xio.xio_pages[index] = NULL;
4011 vm_page_unwire(p, 0);
4015 bp->b_xio.xio_npages = newnpages;
4021 * Map a user buffer into KVM via a pbuf. On return the buffer's
4022 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4026 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4037 * bp had better have a command and it better be a pbuf.
4039 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4040 KKASSERT(bp->b_flags & B_PAGING);
4046 * Map the user data into KVM. Mappings have to be page-aligned.
4048 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4051 vmprot = VM_PROT_READ;
4052 if (bp->b_cmd == BUF_CMD_READ)
4053 vmprot |= VM_PROT_WRITE;
4055 while (addr < udata + bytes) {
4057 * Do the vm_fault if needed; do the copy-on-write thing
4058 * when reading stuff off device into memory.
4060 * vm_fault_page*() returns a held VM page.
4062 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4063 va = trunc_page(va);
4065 m = vm_fault_page_quick(va, vmprot, &error);
4067 for (i = 0; i < pidx; ++i) {
4068 vm_page_unhold(bp->b_xio.xio_pages[i]);
4069 bp->b_xio.xio_pages[i] = NULL;
4073 bp->b_xio.xio_pages[pidx] = m;
4079 * Map the page array and set the buffer fields to point to
4080 * the mapped data buffer.
4082 if (pidx > btoc(MAXPHYS))
4083 panic("vmapbuf: mapped more than MAXPHYS");
4084 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4086 bp->b_xio.xio_npages = pidx;
4087 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4088 bp->b_bcount = bytes;
4089 bp->b_bufsize = bytes;
4096 * Free the io map PTEs associated with this IO operation.
4097 * We also invalidate the TLB entries and restore the original b_addr.
4100 vunmapbuf(struct buf *bp)
4105 KKASSERT(bp->b_flags & B_PAGING);
4107 npages = bp->b_xio.xio_npages;
4108 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4109 for (pidx = 0; pidx < npages; ++pidx) {
4110 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4111 bp->b_xio.xio_pages[pidx] = NULL;
4113 bp->b_xio.xio_npages = 0;
4114 bp->b_data = bp->b_kvabase;
4118 * Scan all buffers in the system and issue the callback.
4121 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4127 for (n = 0; n < nbuf; ++n) {
4128 if ((error = callback(&buf[n], info)) < 0) {
4138 * print out statistics from the current status of the buffer pool
4139 * this can be toggeled by the system control option debug.syncprt
4148 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
4149 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
4151 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
4153 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4156 TAILQ_FOREACH(bp, dp, b_freelist) {
4157 counts[bp->b_bufsize/PAGE_SIZE]++;
4161 kprintf("%s: total-%d", bname[i], count);
4162 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4164 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
4172 DB_SHOW_COMMAND(buffer, db_show_buffer)
4175 struct buf *bp = (struct buf *)addr;
4178 db_printf("usage: show buffer <addr>\n");
4182 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4183 db_printf("b_cmd = %d\n", bp->b_cmd);
4184 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4185 "b_resid = %d\n, b_data = %p, "
4186 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4187 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4189 (long long)bp->b_bio2.bio_offset,
4190 (long long)(bp->b_bio2.bio_next ?
4191 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4192 if (bp->b_xio.xio_npages) {
4194 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4195 bp->b_xio.xio_npages);
4196 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4198 m = bp->b_xio.xio_pages[i];
4199 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4200 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4201 if ((i + 1) < bp->b_xio.xio_npages)