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)
241 if ((totalspace = bp->b_runningbufspace) != 0) {
242 atomic_subtract_int(&runningbufspace, totalspace);
243 atomic_subtract_int(&runningbufcount, 1);
244 bp->b_runningbufspace = 0;
247 * see waitrunningbufspace() for limit test.
249 limit = hirunningspace * 2 / 3;
250 if (runningbufreq && runningbufspace <= limit) {
252 wakeup(&runningbufreq);
254 bd_signal(totalspace);
261 * Called when a buffer has been added to one of the free queues to
262 * account for the buffer and to wakeup anyone waiting for free buffers.
263 * This typically occurs when large amounts of metadata are being handled
264 * by the buffer cache ( else buffer space runs out first, usually ).
272 spin_lock_wr(&needsbuffer_spin);
273 needsbuffer &= ~VFS_BIO_NEED_ANY;
274 spin_unlock_wr(&needsbuffer_spin);
275 wakeup(&needsbuffer);
280 * waitrunningbufspace()
282 * Wait for the amount of running I/O to drop to hirunningspace * 2 / 3.
283 * This is the point where write bursting stops so we don't want to wait
284 * for the running amount to drop below it (at least if we still want bioq
287 * The caller may be using this function to block in a tight loop, we
288 * must block while runningbufspace is greater then or equal to
289 * hirunningspace * 2 / 3.
291 * And even with that it may not be enough, due to the presence of
292 * B_LOCKED dirty buffers, so also wait for at least one running buffer
296 waitrunningbufspace(void)
298 int limit = hirunningspace * 2 / 3;
301 if (runningbufspace > limit) {
302 while (runningbufspace > limit) {
304 tsleep(&runningbufreq, 0, "wdrn1", 0);
306 } else if (runningbufspace) {
308 tsleep(&runningbufreq, 0, "wdrn2", 1);
314 * buf_dirty_count_severe:
316 * Return true if we have too many dirty buffers.
319 buf_dirty_count_severe(void)
321 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
322 dirtybufcount >= nbuf / 2);
326 * Return true if the amount of running I/O is severe and BIOQ should
330 buf_runningbufspace_severe(void)
332 return (runningbufspace >= hirunningspace * 2 / 3);
336 * vfs_buf_test_cache:
338 * Called when a buffer is extended. This function clears the B_CACHE
339 * bit if the newly extended portion of the buffer does not contain
344 vfs_buf_test_cache(struct buf *bp,
345 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
348 if (bp->b_flags & B_CACHE) {
349 int base = (foff + off) & PAGE_MASK;
350 if (vm_page_is_valid(m, base, size) == 0)
351 bp->b_flags &= ~B_CACHE;
358 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
367 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
370 if (bd_request == 0 &&
371 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
372 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
373 spin_lock_wr(&needsbuffer_spin);
375 spin_unlock_wr(&needsbuffer_spin);
378 if (bd_request_hw == 0 &&
379 (dirtybufspacehw > lodirtybufspace / 2 ||
380 dirtybufcounthw >= nbuf / 2)) {
381 spin_lock_wr(&needsbuffer_spin);
383 spin_unlock_wr(&needsbuffer_spin);
384 wakeup(&bd_request_hw);
391 * Get the buf_daemon heated up when the number of running and dirty
392 * buffers exceeds the mid-point.
403 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
405 totalspace = runningbufspace + dirtybufspace;
406 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
408 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
409 if (totalspace >= mid2)
410 return(totalspace - mid2);
418 * Wait for the buffer cache to flush (totalspace) bytes worth of
419 * buffers, then return.
421 * Regardless this function blocks while the number of dirty buffers
422 * exceeds hidirtybufspace.
427 bd_wait(int totalspace)
432 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
435 while (totalspace > 0) {
437 if (totalspace > runningbufspace + dirtybufspace)
438 totalspace = runningbufspace + dirtybufspace;
439 count = totalspace / BKVASIZE;
440 if (count >= BD_WAKE_SIZE)
441 count = BD_WAKE_SIZE - 1;
443 spin_lock_wr(&needsbuffer_spin);
444 i = (bd_wake_index + count) & BD_WAKE_MASK;
446 tsleep_interlock(&bd_wake_ary[i], 0);
447 spin_unlock_wr(&needsbuffer_spin);
448 tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz);
450 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
457 * This function is called whenever runningbufspace or dirtybufspace
458 * is reduced. Track threads waiting for run+dirty buffer I/O
464 bd_signal(int totalspace)
468 if (totalspace > 0) {
469 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
470 totalspace = BKVASIZE * BD_WAKE_SIZE;
471 spin_lock_wr(&needsbuffer_spin);
472 while (totalspace > 0) {
475 if (bd_wake_ary[i]) {
477 spin_unlock_wr(&needsbuffer_spin);
478 wakeup(&bd_wake_ary[i]);
479 spin_lock_wr(&needsbuffer_spin);
481 totalspace -= BKVASIZE;
483 spin_unlock_wr(&needsbuffer_spin);
488 * BIO tracking support routines.
490 * Release a ref on a bio_track. Wakeup requests are atomically released
491 * along with the last reference so bk_active will never wind up set to
498 bio_track_rel(struct bio_track *track)
506 active = track->bk_active;
507 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
511 * Full-on. Note that the wait flag is only atomically released on
512 * the 1->0 count transition.
514 * We check for a negative count transition using bit 30 since bit 31
515 * has a different meaning.
518 desired = (active & 0x7FFFFFFF) - 1;
520 desired |= active & 0x80000000;
521 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
522 if (desired & 0x40000000)
523 panic("bio_track_rel: bad count: %p\n", track);
524 if (active & 0x80000000)
528 active = track->bk_active;
533 * Wait for the tracking count to reach 0.
535 * Use atomic ops such that the wait flag is only set atomically when
536 * bk_active is non-zero.
541 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
550 if (track->bk_active == 0)
554 * Full-on. Note that the wait flag may only be atomically set if
555 * the active count is non-zero.
558 while ((active = track->bk_active) != 0) {
559 desired = active | 0x80000000;
560 tsleep_interlock(track, slp_flags);
561 if (active == desired ||
562 atomic_cmpset_int(&track->bk_active, active, desired)) {
563 error = tsleep(track, slp_flags | PINTERLOCKED,
575 * Load time initialisation of the buffer cache, called from machine
576 * dependant initialization code.
582 vm_offset_t bogus_offset;
585 spin_init(&needsbuffer_spin);
587 /* next, make a null set of free lists */
588 for (i = 0; i < BUFFER_QUEUES; i++)
589 TAILQ_INIT(&bufqueues[i]);
591 /* finally, initialize each buffer header and stick on empty q */
592 for (i = 0; i < nbuf; i++) {
594 bzero(bp, sizeof *bp);
595 bp->b_flags = B_INVAL; /* we're just an empty header */
596 bp->b_cmd = BUF_CMD_DONE;
597 bp->b_qindex = BQUEUE_EMPTY;
599 xio_init(&bp->b_xio);
602 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
606 * maxbufspace is the absolute maximum amount of buffer space we are
607 * allowed to reserve in KVM and in real terms. The absolute maximum
608 * is nominally used by buf_daemon. hibufspace is the nominal maximum
609 * used by most other processes. The differential is required to
610 * ensure that buf_daemon is able to run when other processes might
611 * be blocked waiting for buffer space.
613 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
614 * this may result in KVM fragmentation which is not handled optimally
617 maxbufspace = nbuf * BKVASIZE;
618 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
619 lobufspace = hibufspace - MAXBSIZE;
621 lorunningspace = 512 * 1024;
622 /* hirunningspace -- see below */
625 * Limit the amount of malloc memory since it is wired permanently
626 * into the kernel space. Even though this is accounted for in
627 * the buffer allocation, we don't want the malloced region to grow
628 * uncontrolled. The malloc scheme improves memory utilization
629 * significantly on average (small) directories.
631 maxbufmallocspace = hibufspace / 20;
634 * Reduce the chance of a deadlock occuring by limiting the number
635 * of delayed-write dirty buffers we allow to stack up.
637 * We don't want too much actually queued to the device at once
638 * (XXX this needs to be per-mount!), because the buffers will
639 * wind up locked for a very long period of time while the I/O
642 hidirtybufspace = hibufspace / 2; /* dirty + running */
643 hirunningspace = hibufspace / 16; /* locked & queued to device */
644 if (hirunningspace < 1024 * 1024)
645 hirunningspace = 1024 * 1024;
650 lodirtybufspace = hidirtybufspace / 2;
653 * Maximum number of async ops initiated per buf_daemon loop. This is
654 * somewhat of a hack at the moment, we really need to limit ourselves
655 * based on the number of bytes of I/O in-transit that were initiated
659 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
660 bogus_page = vm_page_alloc(&kernel_object,
661 (bogus_offset >> PAGE_SHIFT),
663 vmstats.v_wire_count++;
668 * Initialize the embedded bio structures
671 initbufbio(struct buf *bp)
673 bp->b_bio1.bio_buf = bp;
674 bp->b_bio1.bio_prev = NULL;
675 bp->b_bio1.bio_offset = NOOFFSET;
676 bp->b_bio1.bio_next = &bp->b_bio2;
677 bp->b_bio1.bio_done = NULL;
678 bp->b_bio1.bio_flags = 0;
680 bp->b_bio2.bio_buf = bp;
681 bp->b_bio2.bio_prev = &bp->b_bio1;
682 bp->b_bio2.bio_offset = NOOFFSET;
683 bp->b_bio2.bio_next = NULL;
684 bp->b_bio2.bio_done = NULL;
685 bp->b_bio2.bio_flags = 0;
689 * Reinitialize the embedded bio structures as well as any additional
690 * translation cache layers.
693 reinitbufbio(struct buf *bp)
697 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
698 bio->bio_done = NULL;
699 bio->bio_offset = NOOFFSET;
704 * Push another BIO layer onto an existing BIO and return it. The new
705 * BIO layer may already exist, holding cached translation data.
708 push_bio(struct bio *bio)
712 if ((nbio = bio->bio_next) == NULL) {
713 int index = bio - &bio->bio_buf->b_bio_array[0];
714 if (index >= NBUF_BIO - 1) {
715 panic("push_bio: too many layers bp %p\n",
718 nbio = &bio->bio_buf->b_bio_array[index + 1];
719 bio->bio_next = nbio;
720 nbio->bio_prev = bio;
721 nbio->bio_buf = bio->bio_buf;
722 nbio->bio_offset = NOOFFSET;
723 nbio->bio_done = NULL;
724 nbio->bio_next = NULL;
726 KKASSERT(nbio->bio_done == NULL);
731 * Pop a BIO translation layer, returning the previous layer. The
732 * must have been previously pushed.
735 pop_bio(struct bio *bio)
737 return(bio->bio_prev);
741 clearbiocache(struct bio *bio)
744 bio->bio_offset = NOOFFSET;
752 * Free the KVA allocation for buffer 'bp'.
754 * Must be called from a critical section as this is the only locking for
757 * Since this call frees up buffer space, we call bufspacewakeup().
762 bfreekva(struct buf *bp)
769 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
770 vm_map_lock(&buffer_map);
771 bufspace -= bp->b_kvasize;
772 vm_map_delete(&buffer_map,
773 (vm_offset_t) bp->b_kvabase,
774 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
777 vm_map_unlock(&buffer_map);
778 vm_map_entry_release(count);
788 * Remove the buffer from the appropriate free list.
791 _bremfree(struct buf *bp)
793 if (bp->b_qindex != BQUEUE_NONE) {
794 KASSERT(BUF_REFCNTNB(bp) == 1,
795 ("bremfree: bp %p not locked",bp));
796 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
797 bp->b_qindex = BQUEUE_NONE;
799 if (BUF_REFCNTNB(bp) <= 1)
800 panic("bremfree: removing a buffer not on a queue");
805 bremfree(struct buf *bp)
807 spin_lock_wr(&bufspin);
809 spin_unlock_wr(&bufspin);
813 bremfree_locked(struct buf *bp)
821 * Get a buffer with the specified data. Look in the cache first. We
822 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
823 * is set, the buffer is valid and we do not have to do anything ( see
829 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
833 bp = getblk(vp, loffset, size, 0, 0);
836 /* if not found in cache, do some I/O */
837 if ((bp->b_flags & B_CACHE) == 0) {
839 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
840 bp->b_cmd = BUF_CMD_READ;
841 bp->b_bio1.bio_done = biodone_sync;
842 bp->b_bio1.bio_flags |= BIO_SYNC;
843 vfs_busy_pages(vp, bp);
844 vn_strategy(vp, &bp->b_bio1);
846 return (biowait(&bp->b_bio1, "biord"));
854 * Operates like bread, but also starts asynchronous I/O on
855 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
856 * to initiating I/O . If B_CACHE is set, the buffer is valid
857 * and we do not have to do anything.
862 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
863 int *rabsize, int cnt, struct buf **bpp)
865 struct buf *bp, *rabp;
867 int rv = 0, readwait = 0;
869 *bpp = bp = getblk(vp, loffset, size, 0, 0);
871 /* if not found in cache, do some I/O */
872 if ((bp->b_flags & B_CACHE) == 0) {
874 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
875 bp->b_cmd = BUF_CMD_READ;
876 bp->b_bio1.bio_done = biodone_sync;
877 bp->b_bio1.bio_flags |= BIO_SYNC;
878 vfs_busy_pages(vp, bp);
879 vn_strategy(vp, &bp->b_bio1);
884 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
885 if (inmem(vp, *raoffset))
887 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
889 if ((rabp->b_flags & B_CACHE) == 0) {
891 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
892 rabp->b_cmd = BUF_CMD_READ;
893 vfs_busy_pages(vp, rabp);
895 vn_strategy(vp, &rabp->b_bio1);
902 rv = biowait(&bp->b_bio1, "biord");
909 * Synchronous write, waits for completion.
911 * Write, release buffer on completion. (Done by iodone
912 * if async). Do not bother writing anything if the buffer
915 * Note that we set B_CACHE here, indicating that buffer is
916 * fully valid and thus cacheable. This is true even of NFS
917 * now so we set it generally. This could be set either here
918 * or in biodone() since the I/O is synchronous. We put it
922 bwrite(struct buf *bp)
926 if (bp->b_flags & B_INVAL) {
930 if (BUF_REFCNTNB(bp) == 0)
931 panic("bwrite: buffer is not busy???");
933 /* Mark the buffer clean */
936 bp->b_flags &= ~(B_ERROR | B_EINTR);
937 bp->b_flags |= B_CACHE;
938 bp->b_cmd = BUF_CMD_WRITE;
939 bp->b_bio1.bio_done = biodone_sync;
940 bp->b_bio1.bio_flags |= BIO_SYNC;
941 vfs_busy_pages(bp->b_vp, bp);
944 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
945 * valid for vnode-backed buffers.
947 bp->b_runningbufspace = bp->b_bufsize;
948 if (bp->b_runningbufspace) {
949 runningbufspace += bp->b_runningbufspace;
953 vn_strategy(bp->b_vp, &bp->b_bio1);
954 error = biowait(&bp->b_bio1, "biows");
962 * Asynchronous write. Start output on a buffer, but do not wait for
963 * it to complete. The buffer is released when the output completes.
965 * bwrite() ( or the VOP routine anyway ) is responsible for handling
966 * B_INVAL buffers. Not us.
969 bawrite(struct buf *bp)
971 if (bp->b_flags & B_INVAL) {
975 if (BUF_REFCNTNB(bp) == 0)
976 panic("bwrite: buffer is not busy???");
978 /* Mark the buffer clean */
981 bp->b_flags &= ~(B_ERROR | B_EINTR);
982 bp->b_flags |= B_CACHE;
983 bp->b_cmd = BUF_CMD_WRITE;
984 KKASSERT(bp->b_bio1.bio_done == NULL);
985 vfs_busy_pages(bp->b_vp, bp);
988 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
989 * valid for vnode-backed buffers.
991 bp->b_runningbufspace = bp->b_bufsize;
992 if (bp->b_runningbufspace) {
993 runningbufspace += bp->b_runningbufspace;
998 vn_strategy(bp->b_vp, &bp->b_bio1);
1004 * Ordered write. Start output on a buffer, and flag it so that the
1005 * device will write it in the order it was queued. The buffer is
1006 * released when the output completes. bwrite() ( or the VOP routine
1007 * anyway ) is responsible for handling B_INVAL buffers.
1010 bowrite(struct buf *bp)
1012 bp->b_flags |= B_ORDERED;
1020 * Delayed write. (Buffer is marked dirty). Do not bother writing
1021 * anything if the buffer is marked invalid.
1023 * Note that since the buffer must be completely valid, we can safely
1024 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1025 * biodone() in order to prevent getblk from writing the buffer
1026 * out synchronously.
1029 bdwrite(struct buf *bp)
1031 if (BUF_REFCNTNB(bp) == 0)
1032 panic("bdwrite: buffer is not busy");
1034 if (bp->b_flags & B_INVAL) {
1041 * Set B_CACHE, indicating that the buffer is fully valid. This is
1042 * true even of NFS now.
1044 bp->b_flags |= B_CACHE;
1047 * This bmap keeps the system from needing to do the bmap later,
1048 * perhaps when the system is attempting to do a sync. Since it
1049 * is likely that the indirect block -- or whatever other datastructure
1050 * that the filesystem needs is still in memory now, it is a good
1051 * thing to do this. Note also, that if the pageout daemon is
1052 * requesting a sync -- there might not be enough memory to do
1053 * the bmap then... So, this is important to do.
1055 if (bp->b_bio2.bio_offset == NOOFFSET) {
1056 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1057 NULL, NULL, BUF_CMD_WRITE);
1061 * Set the *dirty* buffer range based upon the VM system dirty pages.
1066 * We need to do this here to satisfy the vnode_pager and the
1067 * pageout daemon, so that it thinks that the pages have been
1068 * "cleaned". Note that since the pages are in a delayed write
1069 * buffer -- the VFS layer "will" see that the pages get written
1070 * out on the next sync, or perhaps the cluster will be completed.
1072 vfs_clean_pages(bp);
1076 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1077 * due to the softdep code.
1084 * Turn buffer into delayed write request by marking it B_DELWRI.
1085 * B_RELBUF and B_NOCACHE must be cleared.
1087 * We reassign the buffer to itself to properly update it in the
1088 * dirty/clean lists.
1090 * Must be called from a critical section.
1091 * The buffer must be on BQUEUE_NONE.
1094 bdirty(struct buf *bp)
1096 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1097 if (bp->b_flags & B_NOCACHE) {
1098 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1099 bp->b_flags &= ~B_NOCACHE;
1101 if (bp->b_flags & B_INVAL) {
1102 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1104 bp->b_flags &= ~B_RELBUF;
1106 if ((bp->b_flags & B_DELWRI) == 0) {
1107 bp->b_flags |= B_DELWRI;
1109 atomic_add_int(&dirtybufcount, 1);
1110 dirtybufspace += bp->b_bufsize;
1111 if (bp->b_flags & B_HEAVY) {
1112 atomic_add_int(&dirtybufcounthw, 1);
1113 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
1120 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1121 * needs to be flushed with a different buf_daemon thread to avoid
1122 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1125 bheavy(struct buf *bp)
1127 if ((bp->b_flags & B_HEAVY) == 0) {
1128 bp->b_flags |= B_HEAVY;
1129 if (bp->b_flags & B_DELWRI) {
1130 atomic_add_int(&dirtybufcounthw, 1);
1131 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
1139 * Clear B_DELWRI for buffer.
1141 * Must be called from a critical section.
1143 * The buffer is typically on BQUEUE_NONE but there is one case in
1144 * brelse() that calls this function after placing the buffer on
1145 * a different queue.
1150 bundirty(struct buf *bp)
1152 if (bp->b_flags & B_DELWRI) {
1153 bp->b_flags &= ~B_DELWRI;
1155 atomic_subtract_int(&dirtybufcount, 1);
1156 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
1157 if (bp->b_flags & B_HEAVY) {
1158 atomic_subtract_int(&dirtybufcounthw, 1);
1159 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
1161 bd_signal(bp->b_bufsize);
1164 * Since it is now being written, we can clear its deferred write flag.
1166 bp->b_flags &= ~B_DEFERRED;
1172 * Release a busy buffer and, if requested, free its resources. The
1173 * buffer will be stashed in the appropriate bufqueue[] allowing it
1174 * to be accessed later as a cache entity or reused for other purposes.
1179 brelse(struct buf *bp)
1182 int saved_flags = bp->b_flags;
1185 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1188 * If B_NOCACHE is set we are being asked to destroy the buffer and
1189 * its backing store. Clear B_DELWRI.
1191 * B_NOCACHE is set in two cases: (1) when the caller really wants
1192 * to destroy the buffer and backing store and (2) when the caller
1193 * wants to destroy the buffer and backing store after a write
1196 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1200 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1202 * A re-dirtied buffer is only subject to destruction
1203 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1205 /* leave buffer intact */
1206 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1207 (bp->b_bufsize <= 0)) {
1209 * Either a failed read or we were asked to free or not
1210 * cache the buffer. This path is reached with B_DELWRI
1211 * set only if B_INVAL is already set. B_NOCACHE governs
1212 * backing store destruction.
1214 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1215 * buffer cannot be immediately freed.
1217 bp->b_flags |= B_INVAL;
1218 if (LIST_FIRST(&bp->b_dep) != NULL) {
1223 if (bp->b_flags & B_DELWRI) {
1224 atomic_subtract_int(&dirtybufcount, 1);
1225 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
1226 if (bp->b_flags & B_HEAVY) {
1227 atomic_subtract_int(&dirtybufcounthw, 1);
1228 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
1230 bd_signal(bp->b_bufsize);
1232 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1236 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1237 * If vfs_vmio_release() is called with either bit set, the
1238 * underlying pages may wind up getting freed causing a previous
1239 * write (bdwrite()) to get 'lost' because pages associated with
1240 * a B_DELWRI bp are marked clean. Pages associated with a
1241 * B_LOCKED buffer may be mapped by the filesystem.
1243 * If we want to release the buffer ourselves (rather then the
1244 * originator asking us to release it), give the originator a
1245 * chance to countermand the release by setting B_LOCKED.
1247 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1248 * if B_DELWRI is set.
1250 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1251 * on pages to return pages to the VM page queues.
1253 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1254 bp->b_flags &= ~B_RELBUF;
1255 } else if (vm_page_count_severe()) {
1256 if (LIST_FIRST(&bp->b_dep) != NULL) {
1258 buf_deallocate(bp); /* can set B_LOCKED */
1261 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1262 bp->b_flags &= ~B_RELBUF;
1264 bp->b_flags |= B_RELBUF;
1268 * Make sure b_cmd is clear. It may have already been cleared by
1271 * At this point destroying the buffer is governed by the B_INVAL
1272 * or B_RELBUF flags.
1274 bp->b_cmd = BUF_CMD_DONE;
1277 * VMIO buffer rundown. Make sure the VM page array is restored
1278 * after an I/O may have replaces some of the pages with bogus pages
1279 * in order to not destroy dirty pages in a fill-in read.
1281 * Note that due to the code above, if a buffer is marked B_DELWRI
1282 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1283 * B_INVAL may still be set, however.
1285 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1286 * but not the backing store. B_NOCACHE will destroy the backing
1289 * Note that dirty NFS buffers contain byte-granular write ranges
1290 * and should not be destroyed w/ B_INVAL even if the backing store
1293 if (bp->b_flags & B_VMIO) {
1295 * Rundown for VMIO buffers which are not dirty NFS buffers.
1307 * Get the base offset and length of the buffer. Note that
1308 * in the VMIO case if the buffer block size is not
1309 * page-aligned then b_data pointer may not be page-aligned.
1310 * But our b_xio.xio_pages array *IS* page aligned.
1312 * block sizes less then DEV_BSIZE (usually 512) are not
1313 * supported due to the page granularity bits (m->valid,
1314 * m->dirty, etc...).
1316 * See man buf(9) for more information
1319 resid = bp->b_bufsize;
1320 foff = bp->b_loffset;
1323 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1324 m = bp->b_xio.xio_pages[i];
1325 vm_page_flag_clear(m, PG_ZERO);
1327 * If we hit a bogus page, fixup *all* of them
1328 * now. Note that we left these pages wired
1329 * when we removed them so they had better exist,
1330 * and they cannot be ripped out from under us so
1331 * no critical section protection is necessary.
1333 if (m == bogus_page) {
1335 poff = OFF_TO_IDX(bp->b_loffset);
1337 for (j = i; j < bp->b_xio.xio_npages; j++) {
1340 mtmp = bp->b_xio.xio_pages[j];
1341 if (mtmp == bogus_page) {
1342 mtmp = vm_page_lookup(obj, poff + j);
1344 panic("brelse: page missing");
1346 bp->b_xio.xio_pages[j] = mtmp;
1350 if ((bp->b_flags & B_INVAL) == 0) {
1351 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1352 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1354 m = bp->b_xio.xio_pages[i];
1358 * Invalidate the backing store if B_NOCACHE is set
1359 * (e.g. used with vinvalbuf()). If this is NFS
1360 * we impose a requirement that the block size be
1361 * a multiple of PAGE_SIZE and create a temporary
1362 * hack to basically invalidate the whole page. The
1363 * problem is that NFS uses really odd buffer sizes
1364 * especially when tracking piecemeal writes and
1365 * it also vinvalbuf()'s a lot, which would result
1366 * in only partial page validation and invalidation
1367 * here. If the file page is mmap()'d, however,
1368 * all the valid bits get set so after we invalidate
1369 * here we would end up with weird m->valid values
1370 * like 0xfc. nfs_getpages() can't handle this so
1371 * we clear all the valid bits for the NFS case
1372 * instead of just some of them.
1374 * The real bug is the VM system having to set m->valid
1375 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1376 * itself is an artifact of the whole 512-byte
1377 * granular mess that exists to support odd block
1378 * sizes and UFS meta-data block sizes (e.g. 6144).
1379 * A complete rewrite is required.
1381 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1382 int poffset = foff & PAGE_MASK;
1385 presid = PAGE_SIZE - poffset;
1386 if (bp->b_vp->v_tag == VT_NFS &&
1387 bp->b_vp->v_type == VREG) {
1389 } else if (presid > resid) {
1392 KASSERT(presid >= 0, ("brelse: extra page"));
1393 vm_page_set_invalid(m, poffset, presid);
1395 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1396 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1398 if (bp->b_flags & (B_INVAL | B_RELBUF))
1399 vfs_vmio_release(bp);
1403 * Rundown for non-VMIO buffers.
1405 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1409 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1416 if (bp->b_qindex != BQUEUE_NONE)
1417 panic("brelse: free buffer onto another queue???");
1418 if (BUF_REFCNTNB(bp) > 1) {
1419 /* Temporary panic to verify exclusive locking */
1420 /* This panic goes away when we allow shared refs */
1421 panic("brelse: multiple refs");
1427 * Figure out the correct queue to place the cleaned up buffer on.
1428 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1429 * disassociated from their vnode.
1431 spin_lock_wr(&bufspin);
1432 if (bp->b_flags & B_LOCKED) {
1434 * Buffers that are locked are placed in the locked queue
1435 * immediately, regardless of their state.
1437 bp->b_qindex = BQUEUE_LOCKED;
1438 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1439 } else if (bp->b_bufsize == 0) {
1441 * Buffers with no memory. Due to conditionals near the top
1442 * of brelse() such buffers should probably already be
1443 * marked B_INVAL and disassociated from their vnode.
1445 bp->b_flags |= B_INVAL;
1446 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1447 KKASSERT((bp->b_flags & B_HASHED) == 0);
1448 if (bp->b_kvasize) {
1449 bp->b_qindex = BQUEUE_EMPTYKVA;
1451 bp->b_qindex = BQUEUE_EMPTY;
1453 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1454 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1456 * Buffers with junk contents. Again these buffers had better
1457 * already be disassociated from their vnode.
1459 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1460 KKASSERT((bp->b_flags & B_HASHED) == 0);
1461 bp->b_flags |= B_INVAL;
1462 bp->b_qindex = BQUEUE_CLEAN;
1463 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1466 * Remaining buffers. These buffers are still associated with
1469 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1471 bp->b_qindex = BQUEUE_DIRTY;
1472 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1474 case B_DELWRI | B_HEAVY:
1475 bp->b_qindex = BQUEUE_DIRTY_HW;
1476 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1481 * NOTE: Buffers are always placed at the end of the
1482 * queue. If B_AGE is not set the buffer will cycle
1483 * through the queue twice.
1485 bp->b_qindex = BQUEUE_CLEAN;
1486 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1490 spin_unlock_wr(&bufspin);
1493 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1494 * on the correct queue.
1496 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1500 * The bp is on an appropriate queue unless locked. If it is not
1501 * locked or dirty we can wakeup threads waiting for buffer space.
1503 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1504 * if B_INVAL is set ).
1506 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1510 * Something we can maybe free or reuse
1512 if (bp->b_bufsize || bp->b_kvasize)
1516 * Clean up temporary flags and unlock the buffer.
1518 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
1525 * Release a buffer back to the appropriate queue but do not try to free
1526 * it. The buffer is expected to be used again soon.
1528 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1529 * biodone() to requeue an async I/O on completion. It is also used when
1530 * known good buffers need to be requeued but we think we may need the data
1533 * XXX we should be able to leave the B_RELBUF hint set on completion.
1538 bqrelse(struct buf *bp)
1540 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1542 if (bp->b_qindex != BQUEUE_NONE)
1543 panic("bqrelse: free buffer onto another queue???");
1544 if (BUF_REFCNTNB(bp) > 1) {
1545 /* do not release to free list */
1546 panic("bqrelse: multiple refs");
1550 spin_lock_wr(&bufspin);
1551 if (bp->b_flags & B_LOCKED) {
1553 * Locked buffers are released to the locked queue. However,
1554 * if the buffer is dirty it will first go into the dirty
1555 * queue and later on after the I/O completes successfully it
1556 * will be released to the locked queue.
1558 bp->b_qindex = BQUEUE_LOCKED;
1559 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1560 } else if (bp->b_flags & B_DELWRI) {
1561 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1562 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1563 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1564 } else if (vm_page_count_severe()) {
1566 * We are too low on memory, we have to try to free the
1567 * buffer (most importantly: the wired pages making up its
1568 * backing store) *now*.
1570 spin_unlock_wr(&bufspin);
1574 bp->b_qindex = BQUEUE_CLEAN;
1575 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1577 spin_unlock_wr(&bufspin);
1579 if ((bp->b_flags & B_LOCKED) == 0 &&
1580 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1585 * Something we can maybe free or reuse.
1587 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1591 * Final cleanup and unlock. Clear bits that are only used while a
1592 * buffer is actively locked.
1594 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
1601 * Return backing pages held by the buffer 'bp' back to the VM system
1602 * if possible. The pages are freed if they are no longer valid or
1603 * attempt to free if it was used for direct I/O otherwise they are
1604 * sent to the page cache.
1606 * Pages that were marked busy are left alone and skipped.
1608 * The KVA mapping (b_data) for the underlying pages is removed by
1612 vfs_vmio_release(struct buf *bp)
1618 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1619 m = bp->b_xio.xio_pages[i];
1620 bp->b_xio.xio_pages[i] = NULL;
1622 * In order to keep page LRU ordering consistent, put
1623 * everything on the inactive queue.
1625 vm_page_unwire(m, 0);
1627 * We don't mess with busy pages, it is
1628 * the responsibility of the process that
1629 * busied the pages to deal with them.
1631 if ((m->flags & PG_BUSY) || (m->busy != 0))
1634 if (m->wire_count == 0) {
1635 vm_page_flag_clear(m, PG_ZERO);
1637 * Might as well free the page if we can and it has
1638 * no valid data. We also free the page if the
1639 * buffer was used for direct I/O.
1642 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1643 m->hold_count == 0) {
1645 vm_page_protect(m, VM_PROT_NONE);
1649 if (bp->b_flags & B_DIRECT) {
1650 vm_page_try_to_free(m);
1651 } else if (vm_page_count_severe()) {
1652 vm_page_try_to_cache(m);
1657 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1658 if (bp->b_bufsize) {
1662 bp->b_xio.xio_npages = 0;
1663 bp->b_flags &= ~B_VMIO;
1664 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1675 * Implement clustered async writes for clearing out B_DELWRI buffers.
1676 * This is much better then the old way of writing only one buffer at
1677 * a time. Note that we may not be presented with the buffers in the
1678 * correct order, so we search for the cluster in both directions.
1680 * The buffer is locked on call.
1683 vfs_bio_awrite(struct buf *bp)
1687 off_t loffset = bp->b_loffset;
1688 struct vnode *vp = bp->b_vp;
1695 * right now we support clustered writing only to regular files. If
1696 * we find a clusterable block we could be in the middle of a cluster
1697 * rather then at the beginning.
1699 * NOTE: b_bio1 contains the logical loffset and is aliased
1700 * to b_loffset. b_bio2 contains the translated block number.
1702 if ((vp->v_type == VREG) &&
1703 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1704 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1706 size = vp->v_mount->mnt_stat.f_iosize;
1708 for (i = size; i < MAXPHYS; i += size) {
1709 if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) &&
1710 BUF_REFCNT(bpa) == 0 &&
1711 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1712 (B_DELWRI | B_CLUSTEROK)) &&
1713 (bpa->b_bufsize == size)) {
1714 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1715 (bpa->b_bio2.bio_offset !=
1716 bp->b_bio2.bio_offset + i))
1722 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1723 if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) &&
1724 BUF_REFCNT(bpa) == 0 &&
1725 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1726 (B_DELWRI | B_CLUSTEROK)) &&
1727 (bpa->b_bufsize == size)) {
1728 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1729 (bpa->b_bio2.bio_offset !=
1730 bp->b_bio2.bio_offset - j))
1740 * this is a possible cluster write
1742 if (nbytes != size) {
1744 nwritten = cluster_wbuild(vp, size,
1745 loffset - j, nbytes);
1751 * default (old) behavior, writing out only one block
1753 * XXX returns b_bufsize instead of b_bcount for nwritten?
1755 nwritten = bp->b_bufsize;
1765 * Find and initialize a new buffer header, freeing up existing buffers
1766 * in the bufqueues as necessary. The new buffer is returned locked.
1768 * Important: B_INVAL is not set. If the caller wishes to throw the
1769 * buffer away, the caller must set B_INVAL prior to calling brelse().
1772 * We have insufficient buffer headers
1773 * We have insufficient buffer space
1774 * buffer_map is too fragmented ( space reservation fails )
1775 * If we have to flush dirty buffers ( but we try to avoid this )
1777 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1778 * Instead we ask the buf daemon to do it for us. We attempt to
1779 * avoid piecemeal wakeups of the pageout daemon.
1784 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1790 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1791 static int flushingbufs;
1794 * We can't afford to block since we might be holding a vnode lock,
1795 * which may prevent system daemons from running. We deal with
1796 * low-memory situations by proactively returning memory and running
1797 * async I/O rather then sync I/O.
1801 --getnewbufrestarts;
1803 ++getnewbufrestarts;
1806 * Setup for scan. If we do not have enough free buffers,
1807 * we setup a degenerate case that immediately fails. Note
1808 * that if we are specially marked process, we are allowed to
1809 * dip into our reserves.
1811 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1813 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1814 * However, there are a number of cases (defragging, reusing, ...)
1815 * where we cannot backup.
1817 nqindex = BQUEUE_EMPTYKVA;
1818 spin_lock_wr(&bufspin);
1819 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1823 * If no EMPTYKVA buffers and we are either
1824 * defragging or reusing, locate a CLEAN buffer
1825 * to free or reuse. If bufspace useage is low
1826 * skip this step so we can allocate a new buffer.
1828 if (defrag || bufspace >= lobufspace) {
1829 nqindex = BQUEUE_CLEAN;
1830 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1834 * If we could not find or were not allowed to reuse a
1835 * CLEAN buffer, check to see if it is ok to use an EMPTY
1836 * buffer. We can only use an EMPTY buffer if allocating
1837 * its KVA would not otherwise run us out of buffer space.
1839 if (nbp == NULL && defrag == 0 &&
1840 bufspace + maxsize < hibufspace) {
1841 nqindex = BQUEUE_EMPTY;
1842 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1847 * Run scan, possibly freeing data and/or kva mappings on the fly
1850 * WARNING! bufspin is held!
1852 while ((bp = nbp) != NULL) {
1853 int qindex = nqindex;
1855 nbp = TAILQ_NEXT(bp, b_freelist);
1858 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1859 * cycles through the queue twice before being selected.
1861 if (qindex == BQUEUE_CLEAN &&
1862 (bp->b_flags & B_AGE) == 0 && nbp) {
1863 bp->b_flags |= B_AGE;
1864 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1865 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1870 * Calculate next bp ( we can only use it if we do not block
1871 * or do other fancy things ).
1876 nqindex = BQUEUE_EMPTYKVA;
1877 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1880 case BQUEUE_EMPTYKVA:
1881 nqindex = BQUEUE_CLEAN;
1882 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1896 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1899 * Note: we no longer distinguish between VMIO and non-VMIO
1903 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1906 * If we are defragging then we need a buffer with
1907 * b_kvasize != 0. XXX this situation should no longer
1908 * occur, if defrag is non-zero the buffer's b_kvasize
1909 * should also be non-zero at this point. XXX
1911 if (defrag && bp->b_kvasize == 0) {
1912 kprintf("Warning: defrag empty buffer %p\n", bp);
1917 * Start freeing the bp. This is somewhat involved. nbp
1918 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1919 * on the clean list must be disassociated from their
1920 * current vnode. Buffers on the empty[kva] lists have
1921 * already been disassociated.
1924 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1925 spin_unlock_wr(&bufspin);
1926 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1927 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1930 if (bp->b_qindex != qindex) {
1931 spin_unlock_wr(&bufspin);
1932 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1936 bremfree_locked(bp);
1937 spin_unlock_wr(&bufspin);
1940 * Dependancies must be handled before we disassociate the
1943 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1944 * be immediately disassociated. HAMMER then becomes
1945 * responsible for releasing the buffer.
1947 * NOTE: bufspin is UNLOCKED now.
1949 if (LIST_FIRST(&bp->b_dep) != NULL) {
1953 if (bp->b_flags & B_LOCKED) {
1957 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1960 if (qindex == BQUEUE_CLEAN) {
1962 if (bp->b_flags & B_VMIO) {
1964 vfs_vmio_release(bp);
1973 * NOTE: nbp is now entirely invalid. We can only restart
1974 * the scan from this point on.
1976 * Get the rest of the buffer freed up. b_kva* is still
1977 * valid after this operation.
1980 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1981 KKASSERT((bp->b_flags & B_HASHED) == 0);
1984 * critical section protection is not required when
1985 * scrapping a buffer's contents because it is already
1988 if (bp->b_bufsize) {
1994 bp->b_flags = B_BNOCLIP;
1995 bp->b_cmd = BUF_CMD_DONE;
2000 bp->b_xio.xio_npages = 0;
2001 bp->b_dirtyoff = bp->b_dirtyend = 0;
2003 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2005 if (blkflags & GETBLK_BHEAVY)
2006 bp->b_flags |= B_HEAVY;
2009 * If we are defragging then free the buffer.
2012 bp->b_flags |= B_INVAL;
2020 * If we are overcomitted then recover the buffer and its
2021 * KVM space. This occurs in rare situations when multiple
2022 * processes are blocked in getnewbuf() or allocbuf().
2024 if (bufspace >= hibufspace)
2026 if (flushingbufs && bp->b_kvasize != 0) {
2027 bp->b_flags |= B_INVAL;
2032 if (bufspace < lobufspace)
2035 /* NOT REACHED, bufspin not held */
2039 * If we exhausted our list, sleep as appropriate. We may have to
2040 * wakeup various daemons and write out some dirty buffers.
2042 * Generally we are sleeping due to insufficient buffer space.
2044 * NOTE: bufspin is held if bp is NULL, else it is not held.
2050 spin_unlock_wr(&bufspin);
2052 flags = VFS_BIO_NEED_BUFSPACE;
2054 } else if (bufspace >= hibufspace) {
2056 flags = VFS_BIO_NEED_BUFSPACE;
2059 flags = VFS_BIO_NEED_ANY;
2062 needsbuffer |= flags;
2063 bd_speedup(); /* heeeelp */
2064 while (needsbuffer & flags) {
2065 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
2070 * We finally have a valid bp. We aren't quite out of the
2071 * woods, we still have to reserve kva space. In order
2072 * to keep fragmentation sane we only allocate kva in
2075 * (bufspin is not held)
2077 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2079 if (maxsize != bp->b_kvasize) {
2080 vm_offset_t addr = 0;
2086 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2087 vm_map_lock(&buffer_map);
2089 if (vm_map_findspace(&buffer_map,
2090 vm_map_min(&buffer_map), maxsize,
2091 maxsize, 0, &addr)) {
2093 * Uh oh. Buffer map is too fragmented. We
2094 * must defragment the map.
2096 vm_map_unlock(&buffer_map);
2097 vm_map_entry_release(count);
2100 bp->b_flags |= B_INVAL;
2106 vm_map_insert(&buffer_map, &count,
2108 addr, addr + maxsize,
2110 VM_PROT_ALL, VM_PROT_ALL,
2113 bp->b_kvabase = (caddr_t) addr;
2114 bp->b_kvasize = maxsize;
2115 bufspace += bp->b_kvasize;
2118 vm_map_unlock(&buffer_map);
2119 vm_map_entry_release(count);
2122 bp->b_data = bp->b_kvabase;
2128 * This routine is called in an emergency to recover VM pages from the
2129 * buffer cache by cashing in clean buffers. The idea is to recover
2130 * enough pages to be able to satisfy a stuck bio_page_alloc().
2133 recoverbufpages(void)
2140 spin_lock_wr(&bufspin);
2141 while (bytes < MAXBSIZE) {
2142 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2147 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2148 * cycles through the queue twice before being selected.
2150 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
2151 bp->b_flags |= B_AGE;
2152 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
2153 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
2161 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
2162 KKASSERT((bp->b_flags & B_DELWRI) == 0);
2165 * Start freeing the bp. This is somewhat involved.
2167 * Buffers on the clean list must be disassociated from
2168 * their current vnode
2171 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2172 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp);
2173 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
2176 if (bp->b_qindex != BQUEUE_CLEAN) {
2177 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp, bp->b_qindex);
2181 bremfree_locked(bp);
2182 spin_unlock_wr(&bufspin);
2185 * Dependancies must be handled before we disassociate the
2188 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2189 * be immediately disassociated. HAMMER then becomes
2190 * responsible for releasing the buffer.
2192 if (LIST_FIRST(&bp->b_dep) != NULL) {
2194 if (bp->b_flags & B_LOCKED) {
2196 spin_lock_wr(&bufspin);
2199 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2202 bytes += bp->b_bufsize;
2205 if (bp->b_flags & B_VMIO) {
2206 bp->b_flags |= B_DIRECT; /* try to free pages */
2207 vfs_vmio_release(bp);
2212 KKASSERT(bp->b_vp == NULL);
2213 KKASSERT((bp->b_flags & B_HASHED) == 0);
2216 * critical section protection is not required when
2217 * scrapping a buffer's contents because it is already
2224 bp->b_flags = B_BNOCLIP;
2225 bp->b_cmd = BUF_CMD_DONE;
2230 bp->b_xio.xio_npages = 0;
2231 bp->b_dirtyoff = bp->b_dirtyend = 0;
2233 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2235 bp->b_flags |= B_INVAL;
2238 spin_lock_wr(&bufspin);
2240 spin_unlock_wr(&bufspin);
2247 * Buffer flushing daemon. Buffers are normally flushed by the
2248 * update daemon but if it cannot keep up this process starts to
2249 * take the load in an attempt to prevent getnewbuf() from blocking.
2251 * Once a flush is initiated it does not stop until the number
2252 * of buffers falls below lodirtybuffers, but we will wake up anyone
2253 * waiting at the mid-point.
2256 static struct kproc_desc buf_kp = {
2261 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2262 kproc_start, &buf_kp)
2264 static struct kproc_desc bufhw_kp = {
2269 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2270 kproc_start, &bufhw_kp)
2278 * This process needs to be suspended prior to shutdown sync.
2280 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2281 bufdaemon_td, SHUTDOWN_PRI_LAST);
2282 curthread->td_flags |= TDF_SYSTHREAD;
2285 * This process is allowed to take the buffer cache to the limit
2290 kproc_suspend_loop();
2293 * Do the flush as long as the number of dirty buffers
2294 * (including those running) exceeds lodirtybufspace.
2296 * When flushing limit running I/O to hirunningspace
2297 * Do the flush. Limit the amount of in-transit I/O we
2298 * allow to build up, otherwise we would completely saturate
2299 * the I/O system. Wakeup any waiting processes before we
2300 * normally would so they can run in parallel with our drain.
2302 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2303 * but because we split the operation into two threads we
2304 * have to cut it in half for each thread.
2306 waitrunningbufspace();
2307 limit = lodirtybufspace / 2;
2308 while (runningbufspace + dirtybufspace > limit ||
2309 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
2310 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2312 if (runningbufspace < hirunningspace)
2314 waitrunningbufspace();
2318 * We reached our low water mark, reset the
2319 * request and sleep until we are needed again.
2320 * The sleep is just so the suspend code works.
2322 spin_lock_wr(&needsbuffer_spin);
2323 if (bd_request == 0) {
2324 msleep(&bd_request, &needsbuffer_spin, 0,
2328 spin_unlock_wr(&needsbuffer_spin);
2338 * This process needs to be suspended prior to shutdown sync.
2340 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2341 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2342 curthread->td_flags |= TDF_SYSTHREAD;
2345 * This process is allowed to take the buffer cache to the limit
2350 kproc_suspend_loop();
2353 * Do the flush. Limit the amount of in-transit I/O we
2354 * allow to build up, otherwise we would completely saturate
2355 * the I/O system. Wakeup any waiting processes before we
2356 * normally would so they can run in parallel with our drain.
2358 * Once we decide to flush push the queued I/O up to
2359 * hirunningspace in order to trigger bursting by the bioq
2362 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2363 * but because we split the operation into two threads we
2364 * have to cut it in half for each thread.
2366 waitrunningbufspace();
2367 limit = lodirtybufspace / 2;
2368 while (runningbufspace + dirtybufspacehw > limit ||
2369 dirtybufcounthw >= nbuf / 2) {
2370 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2372 if (runningbufspace < hirunningspace)
2374 waitrunningbufspace();
2378 * We reached our low water mark, reset the
2379 * request and sleep until we are needed again.
2380 * The sleep is just so the suspend code works.
2382 spin_lock_wr(&needsbuffer_spin);
2383 if (bd_request_hw == 0) {
2384 msleep(&bd_request_hw, &needsbuffer_spin, 0,
2388 spin_unlock_wr(&needsbuffer_spin);
2395 * Try to flush a buffer in the dirty queue. We must be careful to
2396 * free up B_INVAL buffers instead of write them, which NFS is
2397 * particularly sensitive to.
2399 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2400 * that we really want to try to get the buffer out and reuse it
2401 * due to the write load on the machine.
2404 flushbufqueues(bufq_type_t q)
2410 spin_lock_wr(&bufspin);
2413 bp = TAILQ_FIRST(&bufqueues[q]);
2415 KASSERT((bp->b_flags & B_DELWRI),
2416 ("unexpected clean buffer %p", bp));
2418 if (bp->b_flags & B_DELWRI) {
2419 if (bp->b_flags & B_INVAL) {
2420 spin_unlock_wr(&bufspin);
2422 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2423 panic("flushbufqueues: locked buf");
2429 if (LIST_FIRST(&bp->b_dep) != NULL &&
2430 (bp->b_flags & B_DEFERRED) == 0 &&
2431 buf_countdeps(bp, 0)) {
2432 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2433 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2435 bp->b_flags |= B_DEFERRED;
2436 bp = TAILQ_FIRST(&bufqueues[q]);
2441 * Only write it out if we can successfully lock
2442 * it. If the buffer has a dependancy,
2443 * buf_checkwrite must also return 0 for us to
2444 * be able to initate the write.
2446 * If the buffer is flagged B_ERROR it may be
2447 * requeued over and over again, we try to
2448 * avoid a live lock.
2450 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
2451 spin_unlock_wr(&bufspin);
2453 if (LIST_FIRST(&bp->b_dep) != NULL &&
2454 buf_checkwrite(bp)) {
2457 } else if (bp->b_flags & B_ERROR) {
2458 tsleep(bp, 0, "bioer", 1);
2459 bp->b_flags &= ~B_AGE;
2462 bp->b_flags |= B_AGE;
2469 bp = TAILQ_NEXT(bp, b_freelist);
2472 spin_unlock_wr(&bufspin);
2479 * Returns true if no I/O is needed to access the associated VM object.
2480 * This is like findblk except it also hunts around in the VM system for
2483 * Note that we ignore vm_page_free() races from interrupts against our
2484 * lookup, since if the caller is not protected our return value will not
2485 * be any more valid then otherwise once we exit the critical section.
2488 inmem(struct vnode *vp, off_t loffset)
2491 vm_offset_t toff, tinc, size;
2494 if (findblk(vp, loffset, FINDBLK_TEST))
2496 if (vp->v_mount == NULL)
2498 if ((obj = vp->v_object) == NULL)
2502 if (size > vp->v_mount->mnt_stat.f_iosize)
2503 size = vp->v_mount->mnt_stat.f_iosize;
2505 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2506 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2510 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2511 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2512 if (vm_page_is_valid(m,
2513 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2522 * Sets the dirty range for a buffer based on the status of the dirty
2523 * bits in the pages comprising the buffer.
2525 * The range is limited to the size of the buffer.
2527 * This routine is primarily used by NFS, but is generalized for the
2531 vfs_setdirty(struct buf *bp)
2537 * Degenerate case - empty buffer
2540 if (bp->b_bufsize == 0)
2544 * We qualify the scan for modified pages on whether the
2545 * object has been flushed yet. The OBJ_WRITEABLE flag
2546 * is not cleared simply by protecting pages off.
2549 if ((bp->b_flags & B_VMIO) == 0)
2552 object = bp->b_xio.xio_pages[0]->object;
2554 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2555 kprintf("Warning: object %p writeable but not mightbedirty\n", object);
2556 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2557 kprintf("Warning: object %p mightbedirty but not writeable\n", object);
2559 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2560 vm_offset_t boffset;
2561 vm_offset_t eoffset;
2564 * test the pages to see if they have been modified directly
2565 * by users through the VM system.
2567 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2568 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2569 vm_page_test_dirty(bp->b_xio.xio_pages[i]);
2573 * Calculate the encompassing dirty range, boffset and eoffset,
2574 * (eoffset - boffset) bytes.
2577 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2578 if (bp->b_xio.xio_pages[i]->dirty)
2581 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2583 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2584 if (bp->b_xio.xio_pages[i]->dirty) {
2588 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2591 * Fit it to the buffer.
2594 if (eoffset > bp->b_bcount)
2595 eoffset = bp->b_bcount;
2598 * If we have a good dirty range, merge with the existing
2602 if (boffset < eoffset) {
2603 if (bp->b_dirtyoff > boffset)
2604 bp->b_dirtyoff = boffset;
2605 if (bp->b_dirtyend < eoffset)
2606 bp->b_dirtyend = eoffset;
2614 * Locate and return the specified buffer. Unless flagged otherwise,
2615 * a locked buffer will be returned if it exists or NULL if it does not.
2617 * findblk()'d buffers are still on the bufqueues and if you intend
2618 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2619 * and possibly do other stuff to it.
2621 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2622 * for locking the buffer and ensuring that it remains
2623 * the desired buffer after locking.
2625 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2626 * to acquire the lock we return NULL, even if the
2629 * (0) - Lock the buffer blocking.
2634 findblk(struct vnode *vp, off_t loffset, int flags)
2640 lkflags = LK_EXCLUSIVE;
2641 if (flags & FINDBLK_NBLOCK)
2642 lkflags |= LK_NOWAIT;
2645 lwkt_gettoken(&vlock, &vp->v_token);
2646 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2647 lwkt_reltoken(&vlock);
2648 if (bp == NULL || (flags & FINDBLK_TEST))
2650 if (BUF_LOCK(bp, lkflags)) {
2654 if (bp->b_vp == vp && bp->b_loffset == loffset)
2664 * Similar to getblk() except only returns the buffer if it is
2665 * B_CACHE and requires no other manipulation. Otherwise NULL
2668 * If B_RAM is set the buffer might be just fine, but we return
2669 * NULL anyway because we want the code to fall through to the
2670 * cluster read. Otherwise read-ahead breaks.
2673 getcacheblk(struct vnode *vp, off_t loffset)
2677 bp = findblk(vp, loffset, 0);
2679 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
2680 bp->b_flags &= ~B_AGE;
2693 * Get a block given a specified block and offset into a file/device.
2694 * B_INVAL may or may not be set on return. The caller should clear
2695 * B_INVAL prior to initiating a READ.
2697 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2698 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2699 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2700 * without doing any of those things the system will likely believe
2701 * the buffer to be valid (especially if it is not B_VMIO), and the
2702 * next getblk() will return the buffer with B_CACHE set.
2704 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2705 * an existing buffer.
2707 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2708 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2709 * and then cleared based on the backing VM. If the previous buffer is
2710 * non-0-sized but invalid, B_CACHE will be cleared.
2712 * If getblk() must create a new buffer, the new buffer is returned with
2713 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2714 * case it is returned with B_INVAL clear and B_CACHE set based on the
2717 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2718 * B_CACHE bit is clear.
2720 * What this means, basically, is that the caller should use B_CACHE to
2721 * determine whether the buffer is fully valid or not and should clear
2722 * B_INVAL prior to issuing a read. If the caller intends to validate
2723 * the buffer by loading its data area with something, the caller needs
2724 * to clear B_INVAL. If the caller does this without issuing an I/O,
2725 * the caller should set B_CACHE ( as an optimization ), else the caller
2726 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2727 * a write attempt or if it was a successfull read. If the caller
2728 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2729 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2733 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2734 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2739 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2742 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2746 if (size > MAXBSIZE)
2747 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2748 if (vp->v_object == NULL)
2749 panic("getblk: vnode %p has no object!", vp);
2752 if ((bp = findblk(vp, loffset, FINDBLK_TEST)) != NULL) {
2754 * The buffer was found in the cache, but we need to lock it.
2755 * Even with LK_NOWAIT the lockmgr may break our critical
2756 * section, so double-check the validity of the buffer
2757 * once the lock has been obtained.
2759 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2760 if (blkflags & GETBLK_NOWAIT)
2762 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2763 if (blkflags & GETBLK_PCATCH)
2764 lkflags |= LK_PCATCH;
2765 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2767 if (error == ENOLCK)
2771 /* buffer may have changed on us */
2775 * Once the buffer has been locked, make sure we didn't race
2776 * a buffer recyclement. Buffers that are no longer hashed
2777 * will have b_vp == NULL, so this takes care of that check
2780 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2781 kprintf("Warning buffer %p (vp %p loffset %lld) "
2783 bp, vp, (long long)loffset);
2789 * If SZMATCH any pre-existing buffer must be of the requested
2790 * size or NULL is returned. The caller absolutely does not
2791 * want getblk() to bwrite() the buffer on a size mismatch.
2793 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2799 * All vnode-based buffers must be backed by a VM object.
2801 KKASSERT(bp->b_flags & B_VMIO);
2802 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2803 bp->b_flags &= ~B_AGE;
2806 * Make sure that B_INVAL buffers do not have a cached
2807 * block number translation.
2809 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2810 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2811 " did not have cleared bio_offset cache\n",
2812 bp, vp, (long long)loffset);
2813 clearbiocache(&bp->b_bio2);
2817 * The buffer is locked. B_CACHE is cleared if the buffer is
2820 if (bp->b_flags & B_INVAL)
2821 bp->b_flags &= ~B_CACHE;
2825 * Any size inconsistancy with a dirty buffer or a buffer
2826 * with a softupdates dependancy must be resolved. Resizing
2827 * the buffer in such circumstances can lead to problems.
2829 if (size != bp->b_bcount) {
2831 if (bp->b_flags & B_DELWRI) {
2832 bp->b_flags |= B_NOCACHE;
2834 } else if (LIST_FIRST(&bp->b_dep)) {
2835 bp->b_flags |= B_NOCACHE;
2838 bp->b_flags |= B_RELBUF;
2844 KKASSERT(size <= bp->b_kvasize);
2845 KASSERT(bp->b_loffset != NOOFFSET,
2846 ("getblk: no buffer offset"));
2849 * A buffer with B_DELWRI set and B_CACHE clear must
2850 * be committed before we can return the buffer in
2851 * order to prevent the caller from issuing a read
2852 * ( due to B_CACHE not being set ) and overwriting
2855 * Most callers, including NFS and FFS, need this to
2856 * operate properly either because they assume they
2857 * can issue a read if B_CACHE is not set, or because
2858 * ( for example ) an uncached B_DELWRI might loop due
2859 * to softupdates re-dirtying the buffer. In the latter
2860 * case, B_CACHE is set after the first write completes,
2861 * preventing further loops.
2863 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2864 * above while extending the buffer, we cannot allow the
2865 * buffer to remain with B_CACHE set after the write
2866 * completes or it will represent a corrupt state. To
2867 * deal with this we set B_NOCACHE to scrap the buffer
2870 * We might be able to do something fancy, like setting
2871 * B_CACHE in bwrite() except if B_DELWRI is already set,
2872 * so the below call doesn't set B_CACHE, but that gets real
2873 * confusing. This is much easier.
2876 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2878 bp->b_flags |= B_NOCACHE;
2885 * Buffer is not in-core, create new buffer. The buffer
2886 * returned by getnewbuf() is locked. Note that the returned
2887 * buffer is also considered valid (not marked B_INVAL).
2889 * Calculating the offset for the I/O requires figuring out
2890 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2891 * the mount's f_iosize otherwise. If the vnode does not
2892 * have an associated mount we assume that the passed size is
2895 * Note that vn_isdisk() cannot be used here since it may
2896 * return a failure for numerous reasons. Note that the
2897 * buffer size may be larger then the block size (the caller
2898 * will use block numbers with the proper multiple). Beware
2899 * of using any v_* fields which are part of unions. In
2900 * particular, in DragonFly the mount point overloading
2901 * mechanism uses the namecache only and the underlying
2902 * directory vnode is not a special case.
2906 if (vp->v_type == VBLK || vp->v_type == VCHR)
2908 else if (vp->v_mount)
2909 bsize = vp->v_mount->mnt_stat.f_iosize;
2913 maxsize = size + (loffset & PAGE_MASK);
2914 maxsize = imax(maxsize, bsize);
2916 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
2918 if (slpflags || slptimeo)
2924 * Atomically insert the buffer into the hash, so that it can
2925 * be found by findblk().
2927 * If bgetvp() returns non-zero a collision occured, and the
2928 * bp will not be associated with the vnode.
2930 * Make sure the translation layer has been cleared.
2932 bp->b_loffset = loffset;
2933 bp->b_bio2.bio_offset = NOOFFSET;
2934 /* bp->b_bio2.bio_next = NULL; */
2936 if (bgetvp(vp, bp)) {
2937 bp->b_flags |= B_INVAL;
2943 * All vnode-based buffers must be backed by a VM object.
2945 KKASSERT(vp->v_object != NULL);
2946 bp->b_flags |= B_VMIO;
2947 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2959 * Reacquire a buffer that was previously released to the locked queue,
2960 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2961 * set B_LOCKED (which handles the acquisition race).
2963 * To this end, either B_LOCKED must be set or the dependancy list must be
2969 regetblk(struct buf *bp)
2971 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
2972 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2979 * Get an empty, disassociated buffer of given size. The buffer is
2980 * initially set to B_INVAL.
2982 * critical section protection is not required for the allocbuf()
2983 * call because races are impossible here.
2993 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2995 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
3000 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3008 * This code constitutes the buffer memory from either anonymous system
3009 * memory (in the case of non-VMIO operations) or from an associated
3010 * VM object (in the case of VMIO operations). This code is able to
3011 * resize a buffer up or down.
3013 * Note that this code is tricky, and has many complications to resolve
3014 * deadlock or inconsistant data situations. Tread lightly!!!
3015 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3016 * the caller. Calling this code willy nilly can result in the loss of data.
3018 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3019 * B_CACHE for the non-VMIO case.
3021 * This routine does not need to be called from a critical section but you
3022 * must own the buffer.
3027 allocbuf(struct buf *bp, int size)
3029 int newbsize, mbsize;
3032 if (BUF_REFCNT(bp) == 0)
3033 panic("allocbuf: buffer not busy");
3035 if (bp->b_kvasize < size)
3036 panic("allocbuf: buffer too small");
3038 if ((bp->b_flags & B_VMIO) == 0) {
3042 * Just get anonymous memory from the kernel. Don't
3043 * mess with B_CACHE.
3045 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3046 if (bp->b_flags & B_MALLOC)
3049 newbsize = round_page(size);
3051 if (newbsize < bp->b_bufsize) {
3053 * Malloced buffers are not shrunk
3055 if (bp->b_flags & B_MALLOC) {
3057 bp->b_bcount = size;
3059 kfree(bp->b_data, M_BIOBUF);
3060 if (bp->b_bufsize) {
3061 bufmallocspace -= bp->b_bufsize;
3065 bp->b_data = bp->b_kvabase;
3067 bp->b_flags &= ~B_MALLOC;
3073 (vm_offset_t) bp->b_data + newbsize,
3074 (vm_offset_t) bp->b_data + bp->b_bufsize);
3075 } else if (newbsize > bp->b_bufsize) {
3077 * We only use malloced memory on the first allocation.
3078 * and revert to page-allocated memory when the buffer
3081 if ((bufmallocspace < maxbufmallocspace) &&
3082 (bp->b_bufsize == 0) &&
3083 (mbsize <= PAGE_SIZE/2)) {
3085 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3086 bp->b_bufsize = mbsize;
3087 bp->b_bcount = size;
3088 bp->b_flags |= B_MALLOC;
3089 bufmallocspace += mbsize;
3095 * If the buffer is growing on its other-than-first
3096 * allocation, then we revert to the page-allocation
3099 if (bp->b_flags & B_MALLOC) {
3100 origbuf = bp->b_data;
3101 origbufsize = bp->b_bufsize;
3102 bp->b_data = bp->b_kvabase;
3103 if (bp->b_bufsize) {
3104 bufmallocspace -= bp->b_bufsize;
3108 bp->b_flags &= ~B_MALLOC;
3109 newbsize = round_page(newbsize);
3113 (vm_offset_t) bp->b_data + bp->b_bufsize,
3114 (vm_offset_t) bp->b_data + newbsize);
3116 bcopy(origbuf, bp->b_data, origbufsize);
3117 kfree(origbuf, M_BIOBUF);
3124 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3125 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3126 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3127 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3129 if (bp->b_flags & B_MALLOC)
3130 panic("allocbuf: VMIO buffer can't be malloced");
3132 * Set B_CACHE initially if buffer is 0 length or will become
3135 if (size == 0 || bp->b_bufsize == 0)
3136 bp->b_flags |= B_CACHE;
3138 if (newbsize < bp->b_bufsize) {
3140 * DEV_BSIZE aligned new buffer size is less then the
3141 * DEV_BSIZE aligned existing buffer size. Figure out
3142 * if we have to remove any pages.
3144 if (desiredpages < bp->b_xio.xio_npages) {
3145 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3147 * the page is not freed here -- it
3148 * is the responsibility of
3149 * vnode_pager_setsize
3151 m = bp->b_xio.xio_pages[i];
3152 KASSERT(m != bogus_page,
3153 ("allocbuf: bogus page found"));
3154 while (vm_page_sleep_busy(m, TRUE, "biodep"))
3157 bp->b_xio.xio_pages[i] = NULL;
3158 vm_page_unwire(m, 0);
3160 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3161 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3162 bp->b_xio.xio_npages = desiredpages;
3164 } else if (size > bp->b_bcount) {
3166 * We are growing the buffer, possibly in a
3167 * byte-granular fashion.
3175 * Step 1, bring in the VM pages from the object,
3176 * allocating them if necessary. We must clear
3177 * B_CACHE if these pages are not valid for the
3178 * range covered by the buffer.
3180 * critical section protection is required to protect
3181 * against interrupts unbusying and freeing pages
3182 * between our vm_page_lookup() and our
3183 * busycheck/wiring call.
3189 while (bp->b_xio.xio_npages < desiredpages) {
3193 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
3194 if ((m = vm_page_lookup(obj, pi)) == NULL) {
3196 * note: must allocate system pages
3197 * since blocking here could intefere
3198 * with paging I/O, no matter which
3201 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
3205 bp->b_flags &= ~B_CACHE;
3206 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3207 ++bp->b_xio.xio_npages;
3213 * We found a page. If we have to sleep on it,
3214 * retry because it might have gotten freed out
3217 * We can only test PG_BUSY here. Blocking on
3218 * m->busy might lead to a deadlock:
3220 * vm_fault->getpages->cluster_read->allocbuf
3224 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
3226 vm_page_flag_clear(m, PG_ZERO);
3228 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3229 ++bp->b_xio.xio_npages;
3234 * Step 2. We've loaded the pages into the buffer,
3235 * we have to figure out if we can still have B_CACHE
3236 * set. Note that B_CACHE is set according to the
3237 * byte-granular range ( bcount and size ), not the
3238 * aligned range ( newbsize ).
3240 * The VM test is against m->valid, which is DEV_BSIZE
3241 * aligned. Needless to say, the validity of the data
3242 * needs to also be DEV_BSIZE aligned. Note that this
3243 * fails with NFS if the server or some other client
3244 * extends the file's EOF. If our buffer is resized,
3245 * B_CACHE may remain set! XXX
3248 toff = bp->b_bcount;
3249 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3251 while ((bp->b_flags & B_CACHE) && toff < size) {
3254 if (tinc > (size - toff))
3257 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3265 bp->b_xio.xio_pages[pi]
3272 * Step 3, fixup the KVM pmap. Remember that
3273 * bp->b_data is relative to bp->b_loffset, but
3274 * bp->b_loffset may be offset into the first page.
3277 bp->b_data = (caddr_t)
3278 trunc_page((vm_offset_t)bp->b_data);
3280 (vm_offset_t)bp->b_data,
3281 bp->b_xio.xio_pages,
3282 bp->b_xio.xio_npages
3284 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3285 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3289 /* adjust space use on already-dirty buffer */
3290 if (bp->b_flags & B_DELWRI) {
3291 dirtybufspace += newbsize - bp->b_bufsize;
3292 if (bp->b_flags & B_HEAVY)
3293 dirtybufspacehw += newbsize - bp->b_bufsize;
3295 if (newbsize < bp->b_bufsize)
3297 bp->b_bufsize = newbsize; /* actual buffer allocation */
3298 bp->b_bcount = size; /* requested buffer size */
3305 * Wait for buffer I/O completion, returning error status. B_EINTR
3306 * is converted into an EINTR error but not cleared (since a chain
3307 * of biowait() calls may occur).
3309 * On return bpdone() will have been called but the buffer will remain
3310 * locked and will not have been brelse()'d.
3312 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3313 * likely still in progress on return.
3315 * NOTE! This operation is on a BIO, not a BUF.
3317 * NOTE! BIO_DONE is cleared by vn_strategy()
3322 _biowait(struct bio *bio, const char *wmesg, int to)
3324 struct buf *bp = bio->bio_buf;
3329 KKASSERT(bio == &bp->b_bio1);
3331 flags = bio->bio_flags;
3332 if (flags & BIO_DONE)
3334 tsleep_interlock(bio, 0);
3335 nflags = flags | BIO_WANT;
3336 tsleep_interlock(bio, 0);
3337 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3339 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3340 else if (bp->b_cmd == BUF_CMD_READ)
3341 error = tsleep(bio, PINTERLOCKED, "biord", to);
3343 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3345 kprintf("tsleep error biowait %d\n", error);
3355 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3356 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3357 if (bp->b_flags & B_EINTR)
3359 if (bp->b_flags & B_ERROR)
3360 return (bp->b_error ? bp->b_error : EIO);
3365 biowait(struct bio *bio, const char *wmesg)
3367 return(_biowait(bio, wmesg, 0));
3371 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3373 return(_biowait(bio, wmesg, to));
3377 * This associates a tracking count with an I/O. vn_strategy() and
3378 * dev_dstrategy() do this automatically but there are a few cases
3379 * where a vnode or device layer is bypassed when a block translation
3380 * is cached. In such cases bio_start_transaction() may be called on
3381 * the bypassed layers so the system gets an I/O in progress indication
3382 * for those higher layers.
3385 bio_start_transaction(struct bio *bio, struct bio_track *track)
3387 bio->bio_track = track;
3388 bio_track_ref(track);
3392 * Initiate I/O on a vnode.
3395 vn_strategy(struct vnode *vp, struct bio *bio)
3397 struct bio_track *track;
3399 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
3400 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
3401 track = &vp->v_track_read;
3403 track = &vp->v_track_write;
3404 KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3405 bio->bio_track = track;
3406 bio_track_ref(track);
3407 vop_strategy(*vp->v_ops, vp, bio);
3413 * Finish I/O on a buffer after all BIOs have been processed.
3414 * Called when the bio chain is exhausted or by biowait. If called
3415 * by biowait, elseit is typically 0.
3417 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3418 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3419 * assuming B_INVAL is clear.
3421 * For the VMIO case, we set B_CACHE if the op was a read and no
3422 * read error occured, or if the op was a write. B_CACHE is never
3423 * set if the buffer is invalid or otherwise uncacheable.
3425 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3426 * initiator to leave B_INVAL set to brelse the buffer out of existance
3427 * in the biodone routine.
3430 bpdone(struct buf *bp, int elseit)
3434 KASSERT(BUF_REFCNTNB(bp) > 0,
3435 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3436 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3437 ("biodone: bp %p already done!", bp));
3440 * No more BIOs are left. All completion functions have been dealt
3441 * with, now we clean up the buffer.
3444 bp->b_cmd = BUF_CMD_DONE;
3447 * Only reads and writes are processed past this point.
3449 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3450 if (cmd == BUF_CMD_FREEBLKS)
3451 bp->b_flags |= B_NOCACHE;
3458 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3459 * a lot worse. XXX - move this above the clearing of b_cmd
3461 if (LIST_FIRST(&bp->b_dep) != NULL)
3465 * A failed write must re-dirty the buffer unless B_INVAL
3466 * was set. Only applicable to normal buffers (with VPs).
3467 * vinum buffers may not have a vp.
3469 if (cmd == BUF_CMD_WRITE &&
3470 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3471 bp->b_flags &= ~B_NOCACHE;
3476 if (bp->b_flags & B_VMIO) {
3482 struct vnode *vp = bp->b_vp;
3486 #if defined(VFS_BIO_DEBUG)
3487 if (vp->v_auxrefs == 0)
3488 panic("biodone: zero vnode hold count");
3489 if ((vp->v_flag & VOBJBUF) == 0)
3490 panic("biodone: vnode is not setup for merged cache");
3493 foff = bp->b_loffset;
3494 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3495 KASSERT(obj != NULL, ("biodone: missing VM object"));
3497 #if defined(VFS_BIO_DEBUG)
3498 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3499 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3500 obj->paging_in_progress, bp->b_xio.xio_npages);
3505 * Set B_CACHE if the op was a normal read and no error
3506 * occured. B_CACHE is set for writes in the b*write()
3509 iosize = bp->b_bcount - bp->b_resid;
3510 if (cmd == BUF_CMD_READ &&
3511 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3512 bp->b_flags |= B_CACHE;
3517 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3521 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3526 * cleanup bogus pages, restoring the originals. Since
3527 * the originals should still be wired, we don't have
3528 * to worry about interrupt/freeing races destroying
3529 * the VM object association.
3531 m = bp->b_xio.xio_pages[i];
3532 if (m == bogus_page) {
3534 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3536 panic("biodone: page disappeared");
3537 bp->b_xio.xio_pages[i] = m;
3538 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3539 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3541 #if defined(VFS_BIO_DEBUG)
3542 if (OFF_TO_IDX(foff) != m->pindex) {
3543 kprintf("biodone: foff(%lu)/m->pindex(%ld) "
3545 (unsigned long)foff, (long)m->pindex);
3550 * In the write case, the valid and clean bits are
3551 * already changed correctly ( see bdwrite() ), so we
3552 * only need to do this here in the read case.
3554 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3555 vfs_page_set_valid(bp, foff, i, m);
3557 vm_page_flag_clear(m, PG_ZERO);
3560 * when debugging new filesystems or buffer I/O methods, this
3561 * is the most common error that pops up. if you see this, you
3562 * have not set the page busy flag correctly!!!
3565 kprintf("biodone: page busy < 0, "
3566 "pindex: %d, foff: 0x(%x,%x), "
3567 "resid: %d, index: %d\n",
3568 (int) m->pindex, (int)(foff >> 32),
3569 (int) foff & 0xffffffff, resid, i);
3570 if (!vn_isdisk(vp, NULL))
3571 kprintf(" iosize: %ld, loffset: %lld, "
3572 "flags: 0x%08x, npages: %d\n",
3573 bp->b_vp->v_mount->mnt_stat.f_iosize,
3574 (long long)bp->b_loffset,
3575 bp->b_flags, bp->b_xio.xio_npages);
3577 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3578 (long long)bp->b_loffset,
3579 bp->b_flags, bp->b_xio.xio_npages);
3580 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3581 m->valid, m->dirty, m->wire_count);
3582 panic("biodone: page busy < 0");
3584 vm_page_io_finish(m);
3585 vm_object_pip_subtract(obj, 1);
3586 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3590 vm_object_pip_wakeupn(obj, 0);
3596 * Finish up by releasing the buffer. There are no more synchronous
3597 * or asynchronous completions, those were handled by bio_done
3601 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
3612 biodone(struct bio *bio)
3614 struct buf *bp = bio->bio_buf;
3616 runningbufwakeup(bp);
3619 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3622 biodone_t *done_func;
3623 struct bio_track *track;
3626 * BIO tracking. Most but not all BIOs are tracked.
3628 if ((track = bio->bio_track) != NULL) {
3629 bio_track_rel(track);
3630 bio->bio_track = NULL;
3634 * A bio_done function terminates the loop. The function
3635 * will be responsible for any further chaining and/or
3636 * buffer management.
3638 * WARNING! The done function can deallocate the buffer!
3640 if ((done_func = bio->bio_done) != NULL) {
3641 bio->bio_done = NULL;
3645 bio = bio->bio_prev;
3649 * If we've run out of bio's do normal [a]synchronous completion.
3655 * Synchronous biodone - this terminates a synchronous BIO.
3657 * bpdone() is called with elseit=FALSE, leaving the buffer completed
3658 * but still locked. The caller must brelse() the buffer after waiting
3662 biodone_sync(struct bio *bio)
3664 struct buf *bp = bio->bio_buf;
3668 KKASSERT(bio == &bp->b_bio1);
3672 flags = bio->bio_flags;
3673 nflags = (flags | BIO_DONE) & ~BIO_WANT;
3675 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3676 if (flags & BIO_WANT)
3686 * This routine is called in lieu of iodone in the case of
3687 * incomplete I/O. This keeps the busy status for pages
3691 vfs_unbusy_pages(struct buf *bp)
3695 runningbufwakeup(bp);
3696 if (bp->b_flags & B_VMIO) {
3697 struct vnode *vp = bp->b_vp;
3702 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3703 vm_page_t m = bp->b_xio.xio_pages[i];
3706 * When restoring bogus changes the original pages
3707 * should still be wired, so we are in no danger of
3708 * losing the object association and do not need
3709 * critical section protection particularly.
3711 if (m == bogus_page) {
3712 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3714 panic("vfs_unbusy_pages: page missing");
3716 bp->b_xio.xio_pages[i] = m;
3717 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3718 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3720 vm_object_pip_subtract(obj, 1);
3721 vm_page_flag_clear(m, PG_ZERO);
3722 vm_page_io_finish(m);
3724 vm_object_pip_wakeupn(obj, 0);
3729 * vfs_page_set_valid:
3731 * Set the valid bits in a page based on the supplied offset. The
3732 * range is restricted to the buffer's size.
3734 * This routine is typically called after a read completes.
3737 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3739 vm_ooffset_t soff, eoff;
3742 * Start and end offsets in buffer. eoff - soff may not cross a
3743 * page boundry or cross the end of the buffer. The end of the
3744 * buffer, in this case, is our file EOF, not the allocation size
3748 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3749 if (eoff > bp->b_loffset + bp->b_bcount)
3750 eoff = bp->b_loffset + bp->b_bcount;
3753 * Set valid range. This is typically the entire buffer and thus the
3757 vm_page_set_validclean(
3759 (vm_offset_t) (soff & PAGE_MASK),
3760 (vm_offset_t) (eoff - soff)
3768 * This routine is called before a device strategy routine.
3769 * It is used to tell the VM system that paging I/O is in
3770 * progress, and treat the pages associated with the buffer
3771 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3772 * flag is handled to make sure that the object doesn't become
3775 * Since I/O has not been initiated yet, certain buffer flags
3776 * such as B_ERROR or B_INVAL may be in an inconsistant state
3777 * and should be ignored.
3780 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3783 struct lwp *lp = curthread->td_lwp;
3786 * The buffer's I/O command must already be set. If reading,
3787 * B_CACHE must be 0 (double check against callers only doing
3788 * I/O when B_CACHE is 0).
3790 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3791 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3793 if (bp->b_flags & B_VMIO) {
3798 foff = bp->b_loffset;
3799 KASSERT(bp->b_loffset != NOOFFSET,
3800 ("vfs_busy_pages: no buffer offset"));
3804 * Loop until none of the pages are busy.
3807 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3808 vm_page_t m = bp->b_xio.xio_pages[i];
3810 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3815 * Setup for I/O, soft-busy the page right now because
3816 * the next loop may block.
3818 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3819 vm_page_t m = bp->b_xio.xio_pages[i];
3821 vm_page_flag_clear(m, PG_ZERO);
3822 if ((bp->b_flags & B_CLUSTER) == 0) {
3823 vm_object_pip_add(obj, 1);
3824 vm_page_io_start(m);
3829 * Adjust protections for I/O and do bogus-page mapping.
3830 * Assume that vm_page_protect() can block (it can block
3831 * if VM_PROT_NONE, don't take any chances regardless).
3834 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3835 vm_page_t m = bp->b_xio.xio_pages[i];
3838 * When readying a vnode-backed buffer for a write
3839 * we must zero-fill any invalid portions of the
3842 * When readying a vnode-backed buffer for a read
3843 * we must replace any dirty pages with a bogus
3844 * page so we do not destroy dirty data when
3845 * filling in gaps. Dirty pages might not
3846 * necessarily be marked dirty yet, so use m->valid
3847 * as a reasonable test.
3849 * Bogus page replacement is, uh, bogus. We need
3850 * to find a better way.
3852 if (bp->b_cmd == BUF_CMD_WRITE) {
3853 vm_page_protect(m, VM_PROT_READ);
3854 vfs_page_set_valid(bp, foff, i, m);
3855 } else if (m->valid == VM_PAGE_BITS_ALL) {
3856 bp->b_xio.xio_pages[i] = bogus_page;
3859 vm_page_protect(m, VM_PROT_NONE);
3861 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3864 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3865 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3869 * This is the easiest place to put the process accounting for the I/O
3873 if (bp->b_cmd == BUF_CMD_READ)
3874 lp->lwp_ru.ru_inblock++;
3876 lp->lwp_ru.ru_oublock++;
3883 * Tell the VM system that the pages associated with this buffer
3884 * are clean. This is used for delayed writes where the data is
3885 * going to go to disk eventually without additional VM intevention.
3887 * Note that while we only really need to clean through to b_bcount, we
3888 * just go ahead and clean through to b_bufsize.
3891 vfs_clean_pages(struct buf *bp)
3895 if (bp->b_flags & B_VMIO) {
3898 foff = bp->b_loffset;
3899 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset"));
3900 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3901 vm_page_t m = bp->b_xio.xio_pages[i];
3902 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3904 vfs_page_set_valid(bp, foff, i, m);
3911 * vfs_bio_set_validclean:
3913 * Set the range within the buffer to valid and clean. The range is
3914 * relative to the beginning of the buffer, b_loffset. Note that
3915 * b_loffset itself may be offset from the beginning of the first page.
3919 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3921 if (bp->b_flags & B_VMIO) {
3926 * Fixup base to be relative to beginning of first page.
3927 * Set initial n to be the maximum number of bytes in the
3928 * first page that can be validated.
3931 base += (bp->b_loffset & PAGE_MASK);
3932 n = PAGE_SIZE - (base & PAGE_MASK);
3934 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) {
3935 vm_page_t m = bp->b_xio.xio_pages[i];
3940 vm_page_set_validclean(m, base & PAGE_MASK, n);
3951 * Clear a buffer. This routine essentially fakes an I/O, so we need
3952 * to clear B_ERROR and B_INVAL.
3954 * Note that while we only theoretically need to clear through b_bcount,
3955 * we go ahead and clear through b_bufsize.
3959 vfs_bio_clrbuf(struct buf *bp)
3963 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3964 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
3965 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3966 (bp->b_loffset & PAGE_MASK) == 0) {
3967 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3968 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3972 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3973 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3974 bzero(bp->b_data, bp->b_bufsize);
3975 bp->b_xio.xio_pages[0]->valid |= mask;
3981 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3982 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3983 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3984 ea = (caddr_t)(vm_offset_t)ulmin(
3985 (u_long)(vm_offset_t)ea,
3986 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3987 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3988 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3990 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3991 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3995 for (; sa < ea; sa += DEV_BSIZE, j++) {
3996 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3997 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3998 bzero(sa, DEV_BSIZE);
4001 bp->b_xio.xio_pages[i]->valid |= mask;
4002 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
4011 * vm_hold_load_pages:
4013 * Load pages into the buffer's address space. The pages are
4014 * allocated from the kernel object in order to reduce interference
4015 * with the any VM paging I/O activity. The range of loaded
4016 * pages will be wired.
4018 * If a page cannot be allocated, the 'pagedaemon' is woken up to
4019 * retrieve the full range (to - from) of pages.
4023 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4029 to = round_page(to);
4030 from = round_page(from);
4031 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4036 * Note: must allocate system pages since blocking here
4037 * could intefere with paging I/O, no matter which
4040 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT,
4041 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
4044 p->valid = VM_PAGE_BITS_ALL;
4045 vm_page_flag_clear(p, PG_ZERO);
4046 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
4047 bp->b_xio.xio_pages[index] = p;
4054 bp->b_xio.xio_npages = index;
4058 * Allocate pages for a buffer cache buffer.
4060 * Under extremely severe memory conditions even allocating out of the
4061 * system reserve can fail. If this occurs we must allocate out of the
4062 * interrupt reserve to avoid a deadlock with the pageout daemon.
4064 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
4065 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
4066 * against the pageout daemon if pages are not freed from other sources.
4070 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit)
4075 * Try a normal allocation, allow use of system reserve.
4077 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
4082 * The normal allocation failed and we clearly have a page
4083 * deficit. Try to reclaim some clean VM pages directly
4084 * from the buffer cache.
4086 vm_pageout_deficit += deficit;
4090 * We may have blocked, the caller will know what to do if the
4093 if (vm_page_lookup(obj, pg))
4097 * Allocate and allow use of the interrupt reserve.
4099 * If after all that we still can't allocate a VM page we are
4100 * in real trouble, but we slog on anyway hoping that the system
4103 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
4104 VM_ALLOC_INTERRUPT);
4106 if (vm_page_count_severe()) {
4107 kprintf("bio_page_alloc: WARNING emergency page "
4112 kprintf("bio_page_alloc: WARNING emergency page "
4113 "allocation failed\n");
4120 * vm_hold_free_pages:
4122 * Return pages associated with the buffer back to the VM system.
4124 * The range of pages underlying the buffer's address space will
4125 * be unmapped and un-wired.
4128 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4132 int index, newnpages;
4134 from = round_page(from);
4135 to = round_page(to);
4136 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4138 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4139 p = bp->b_xio.xio_pages[index];
4140 if (p && (index < bp->b_xio.xio_npages)) {
4142 kprintf("vm_hold_free_pages: doffset: %lld, "
4144 (long long)bp->b_bio2.bio_offset,
4145 (long long)bp->b_loffset);
4147 bp->b_xio.xio_pages[index] = NULL;
4150 vm_page_unwire(p, 0);
4154 bp->b_xio.xio_npages = newnpages;
4160 * Map a user buffer into KVM via a pbuf. On return the buffer's
4161 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4165 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4176 * bp had better have a command and it better be a pbuf.
4178 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4179 KKASSERT(bp->b_flags & B_PAGING);
4185 * Map the user data into KVM. Mappings have to be page-aligned.
4187 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4190 vmprot = VM_PROT_READ;
4191 if (bp->b_cmd == BUF_CMD_READ)
4192 vmprot |= VM_PROT_WRITE;
4194 while (addr < udata + bytes) {
4196 * Do the vm_fault if needed; do the copy-on-write thing
4197 * when reading stuff off device into memory.
4199 * vm_fault_page*() returns a held VM page.
4201 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4202 va = trunc_page(va);
4204 m = vm_fault_page_quick(va, vmprot, &error);
4206 for (i = 0; i < pidx; ++i) {
4207 vm_page_unhold(bp->b_xio.xio_pages[i]);
4208 bp->b_xio.xio_pages[i] = NULL;
4212 bp->b_xio.xio_pages[pidx] = m;
4218 * Map the page array and set the buffer fields to point to
4219 * the mapped data buffer.
4221 if (pidx > btoc(MAXPHYS))
4222 panic("vmapbuf: mapped more than MAXPHYS");
4223 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4225 bp->b_xio.xio_npages = pidx;
4226 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4227 bp->b_bcount = bytes;
4228 bp->b_bufsize = bytes;
4235 * Free the io map PTEs associated with this IO operation.
4236 * We also invalidate the TLB entries and restore the original b_addr.
4239 vunmapbuf(struct buf *bp)
4244 KKASSERT(bp->b_flags & B_PAGING);
4246 npages = bp->b_xio.xio_npages;
4247 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4248 for (pidx = 0; pidx < npages; ++pidx) {
4249 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4250 bp->b_xio.xio_pages[pidx] = NULL;
4252 bp->b_xio.xio_npages = 0;
4253 bp->b_data = bp->b_kvabase;
4257 * Scan all buffers in the system and issue the callback.
4260 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4266 for (n = 0; n < nbuf; ++n) {
4267 if ((error = callback(&buf[n], info)) < 0) {
4277 * print out statistics from the current status of the buffer pool
4278 * this can be toggeled by the system control option debug.syncprt
4287 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
4288 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
4290 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
4292 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4295 TAILQ_FOREACH(bp, dp, b_freelist) {
4296 counts[bp->b_bufsize/PAGE_SIZE]++;
4300 kprintf("%s: total-%d", bname[i], count);
4301 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4303 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
4311 DB_SHOW_COMMAND(buffer, db_show_buffer)
4314 struct buf *bp = (struct buf *)addr;
4317 db_printf("usage: show buffer <addr>\n");
4321 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4322 db_printf("b_cmd = %d\n", bp->b_cmd);
4323 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4324 "b_resid = %d\n, b_data = %p, "
4325 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4326 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4328 (long long)bp->b_bio2.bio_offset,
4329 (long long)(bp->b_bio2.bio_next ?
4330 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4331 if (bp->b_xio.xio_npages) {
4333 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4334 bp->b_xio.xio_npages);
4335 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4337 m = bp->b_xio.xio_pages[i];
4338 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4339 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4340 if ((i + 1) < bp->b_xio.xio_npages)