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 16384
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_clean_pages(struct buf *bp);
95 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
96 static void vfs_vmio_release(struct buf *bp);
97 static int flushbufqueues(bufq_type_t q);
98 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);
100 static void bd_signal(int totalspace);
101 static void buf_daemon(void);
102 static void buf_daemon_hw(void);
105 * bogus page -- for I/O to/from partially complete buffers
106 * this is a temporary solution to the problem, but it is not
107 * really that bad. it would be better to split the buffer
108 * for input in the case of buffers partially already in memory,
109 * but the code is intricate enough already.
111 vm_page_t bogus_page;
114 * These are all static, but make the ones we export globals so we do
115 * not need to use compiler magic.
117 int bufspace, maxbufspace,
118 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
119 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
120 static int lorunningspace, hirunningspace, runningbufreq;
121 int dirtybufspace, dirtybufspacehw, lodirtybufspace, hidirtybufspace;
122 int dirtybufcount, dirtybufcounthw;
123 int runningbufspace, runningbufcount;
124 static int getnewbufcalls;
125 static int getnewbufrestarts;
126 static int recoverbufcalls;
127 static int needsbuffer; /* locked by needsbuffer_spin */
128 static int bd_request; /* locked by needsbuffer_spin */
129 static int bd_request_hw; /* locked by needsbuffer_spin */
130 static u_int bd_wake_ary[BD_WAKE_SIZE];
131 static u_int bd_wake_index;
132 static struct spinlock needsbuffer_spin;
134 static struct thread *bufdaemon_td;
135 static struct thread *bufdaemonhw_td;
139 * Sysctls for operational control of the buffer cache.
141 SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
142 "Number of dirty buffers to flush before bufdaemon becomes inactive");
143 SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
144 "High watermark used to trigger explicit flushing of dirty buffers");
145 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
146 "Minimum amount of buffer space required for active I/O");
147 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
148 "Maximum amount of buffer space to usable for active I/O");
150 * Sysctls determining current state of the buffer cache.
152 SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
153 "Total number of buffers in buffer cache");
154 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
155 "Pending bytes of dirty buffers (all)");
156 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
157 "Pending bytes of dirty buffers (heavy weight)");
158 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
159 "Pending number of dirty buffers");
160 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
161 "Pending number of dirty buffers (heavy weight)");
162 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
163 "I/O bytes currently in progress due to asynchronous writes");
164 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
165 "I/O buffers currently in progress due to asynchronous writes");
166 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
167 "Hard limit on maximum amount of memory usable for buffer space");
168 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
169 "Soft limit on maximum amount of memory usable for buffer space");
170 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
171 "Minimum amount of memory to reserve for system buffer space");
172 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
173 "Amount of memory available for buffers");
174 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
175 0, "Maximum amount of memory reserved for buffers using malloc");
176 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
177 "Amount of memory left for buffers using malloc-scheme");
178 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
179 "New buffer header acquisition requests");
180 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
181 0, "New buffer header acquisition restarts");
182 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
183 "Recover VM space in an emergency");
184 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
185 "Buffer acquisition restarts due to fragmented buffer map");
186 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
187 "Amount of time KVA space was deallocated in an arbitrary buffer");
188 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
189 "Amount of time buffer re-use operations were successful");
190 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
191 "sizeof(struct buf)");
193 char *buf_wmesg = BUF_WMESG;
195 extern int vm_swap_size;
197 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
198 #define VFS_BIO_NEED_UNUSED02 0x02
199 #define VFS_BIO_NEED_UNUSED04 0x04
200 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
205 * Called when buffer space is potentially available for recovery.
206 * getnewbuf() will block on this flag when it is unable to free
207 * sufficient buffer space. Buffer space becomes recoverable when
208 * bp's get placed back in the queues.
215 * If someone is waiting for BUF space, wake them up. Even
216 * though we haven't freed the kva space yet, the waiting
217 * process will be able to now.
219 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
220 spin_lock_wr(&needsbuffer_spin);
221 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
222 spin_unlock_wr(&needsbuffer_spin);
223 wakeup(&needsbuffer);
230 * Accounting for I/O in progress.
234 runningbufwakeup(struct buf *bp)
239 if ((totalspace = bp->b_runningbufspace) != 0) {
240 atomic_subtract_int(&runningbufspace, totalspace);
241 atomic_subtract_int(&runningbufcount, 1);
242 bp->b_runningbufspace = 0;
245 * see waitrunningbufspace() for limit test.
247 limit = hirunningspace * 2 / 3;
248 if (runningbufreq && runningbufspace <= limit) {
250 wakeup(&runningbufreq);
252 bd_signal(totalspace);
259 * Called when a buffer has been added to one of the free queues to
260 * account for the buffer and to wakeup anyone waiting for free buffers.
261 * This typically occurs when large amounts of metadata are being handled
262 * by the buffer cache ( else buffer space runs out first, usually ).
270 spin_lock_wr(&needsbuffer_spin);
271 needsbuffer &= ~VFS_BIO_NEED_ANY;
272 spin_unlock_wr(&needsbuffer_spin);
273 wakeup(&needsbuffer);
278 * waitrunningbufspace()
280 * Wait for the amount of running I/O to drop to hirunningspace * 2 / 3.
281 * This is the point where write bursting stops so we don't want to wait
282 * for the running amount to drop below it (at least if we still want bioq
285 * The caller may be using this function to block in a tight loop, we
286 * must block while runningbufspace is greater then or equal to
287 * hirunningspace * 2 / 3.
289 * And even with that it may not be enough, due to the presence of
290 * B_LOCKED dirty buffers, so also wait for at least one running buffer
294 waitrunningbufspace(void)
296 int limit = hirunningspace * 2 / 3;
299 if (runningbufspace > limit) {
300 while (runningbufspace > limit) {
302 tsleep(&runningbufreq, 0, "wdrn1", 0);
304 } else if (runningbufspace) {
306 tsleep(&runningbufreq, 0, "wdrn2", 1);
312 * buf_dirty_count_severe:
314 * Return true if we have too many dirty buffers.
317 buf_dirty_count_severe(void)
319 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
320 dirtybufcount >= nbuf / 2);
324 * Return true if the amount of running I/O is severe and BIOQ should
328 buf_runningbufspace_severe(void)
330 return (runningbufspace >= hirunningspace * 2 / 3);
334 * vfs_buf_test_cache:
336 * Called when a buffer is extended. This function clears the B_CACHE
337 * bit if the newly extended portion of the buffer does not contain
340 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
341 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
342 * them while a clean buffer was present.
346 vfs_buf_test_cache(struct buf *bp,
347 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
350 if (bp->b_flags & B_CACHE) {
351 int base = (foff + off) & PAGE_MASK;
352 if (vm_page_is_valid(m, base, size) == 0)
353 bp->b_flags &= ~B_CACHE;
360 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
369 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
372 if (bd_request == 0 &&
373 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
374 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
375 spin_lock_wr(&needsbuffer_spin);
377 spin_unlock_wr(&needsbuffer_spin);
380 if (bd_request_hw == 0 &&
381 (dirtybufspacehw > lodirtybufspace / 2 ||
382 dirtybufcounthw >= nbuf / 2)) {
383 spin_lock_wr(&needsbuffer_spin);
385 spin_unlock_wr(&needsbuffer_spin);
386 wakeup(&bd_request_hw);
393 * Get the buf_daemon heated up when the number of running and dirty
394 * buffers exceeds the mid-point.
396 * Return the total number of dirty bytes past the second mid point
397 * as a measure of how much excess dirty data there is in the system.
408 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
410 totalspace = runningbufspace + dirtybufspace;
411 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
413 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
414 if (totalspace >= mid2)
415 return(totalspace - mid2);
423 * Wait for the buffer cache to flush (totalspace) bytes worth of
424 * buffers, then return.
426 * Regardless this function blocks while the number of dirty buffers
427 * exceeds hidirtybufspace.
432 bd_wait(int totalspace)
437 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
440 while (totalspace > 0) {
442 if (totalspace > runningbufspace + dirtybufspace)
443 totalspace = runningbufspace + dirtybufspace;
444 count = totalspace / BKVASIZE;
445 if (count >= BD_WAKE_SIZE)
446 count = BD_WAKE_SIZE - 1;
448 spin_lock_wr(&needsbuffer_spin);
449 i = (bd_wake_index + count) & BD_WAKE_MASK;
451 tsleep_interlock(&bd_wake_ary[i], 0);
452 spin_unlock_wr(&needsbuffer_spin);
453 tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz);
455 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
462 * This function is called whenever runningbufspace or dirtybufspace
463 * is reduced. Track threads waiting for run+dirty buffer I/O
469 bd_signal(int totalspace)
473 if (totalspace > 0) {
474 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
475 totalspace = BKVASIZE * BD_WAKE_SIZE;
476 spin_lock_wr(&needsbuffer_spin);
477 while (totalspace > 0) {
480 if (bd_wake_ary[i]) {
482 spin_unlock_wr(&needsbuffer_spin);
483 wakeup(&bd_wake_ary[i]);
484 spin_lock_wr(&needsbuffer_spin);
486 totalspace -= BKVASIZE;
488 spin_unlock_wr(&needsbuffer_spin);
493 * BIO tracking support routines.
495 * Release a ref on a bio_track. Wakeup requests are atomically released
496 * along with the last reference so bk_active will never wind up set to
503 bio_track_rel(struct bio_track *track)
511 active = track->bk_active;
512 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
516 * Full-on. Note that the wait flag is only atomically released on
517 * the 1->0 count transition.
519 * We check for a negative count transition using bit 30 since bit 31
520 * has a different meaning.
523 desired = (active & 0x7FFFFFFF) - 1;
525 desired |= active & 0x80000000;
526 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
527 if (desired & 0x40000000)
528 panic("bio_track_rel: bad count: %p\n", track);
529 if (active & 0x80000000)
533 active = track->bk_active;
538 * Wait for the tracking count to reach 0.
540 * Use atomic ops such that the wait flag is only set atomically when
541 * bk_active is non-zero.
546 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
555 if (track->bk_active == 0)
559 * Full-on. Note that the wait flag may only be atomically set if
560 * the active count is non-zero.
563 while ((active = track->bk_active) != 0) {
564 desired = active | 0x80000000;
565 tsleep_interlock(track, slp_flags);
566 if (active == desired ||
567 atomic_cmpset_int(&track->bk_active, active, desired)) {
568 error = tsleep(track, slp_flags | PINTERLOCKED,
580 * Load time initialisation of the buffer cache, called from machine
581 * dependant initialization code.
587 vm_offset_t bogus_offset;
590 spin_init(&needsbuffer_spin);
592 /* next, make a null set of free lists */
593 for (i = 0; i < BUFFER_QUEUES; i++)
594 TAILQ_INIT(&bufqueues[i]);
596 /* finally, initialize each buffer header and stick on empty q */
597 for (i = 0; i < nbuf; i++) {
599 bzero(bp, sizeof *bp);
600 bp->b_flags = B_INVAL; /* we're just an empty header */
601 bp->b_cmd = BUF_CMD_DONE;
602 bp->b_qindex = BQUEUE_EMPTY;
604 xio_init(&bp->b_xio);
607 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
611 * maxbufspace is the absolute maximum amount of buffer space we are
612 * allowed to reserve in KVM and in real terms. The absolute maximum
613 * is nominally used by buf_daemon. hibufspace is the nominal maximum
614 * used by most other processes. The differential is required to
615 * ensure that buf_daemon is able to run when other processes might
616 * be blocked waiting for buffer space.
618 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
619 * this may result in KVM fragmentation which is not handled optimally
622 maxbufspace = nbuf * BKVASIZE;
623 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
624 lobufspace = hibufspace - MAXBSIZE;
626 lorunningspace = 512 * 1024;
627 /* hirunningspace -- see below */
630 * Limit the amount of malloc memory since it is wired permanently
631 * into the kernel space. Even though this is accounted for in
632 * the buffer allocation, we don't want the malloced region to grow
633 * uncontrolled. The malloc scheme improves memory utilization
634 * significantly on average (small) directories.
636 maxbufmallocspace = hibufspace / 20;
639 * Reduce the chance of a deadlock occuring by limiting the number
640 * of delayed-write dirty buffers we allow to stack up.
642 * We don't want too much actually queued to the device at once
643 * (XXX this needs to be per-mount!), because the buffers will
644 * wind up locked for a very long period of time while the I/O
647 hidirtybufspace = hibufspace / 2; /* dirty + running */
648 hirunningspace = hibufspace / 16; /* locked & queued to device */
649 if (hirunningspace < 1024 * 1024)
650 hirunningspace = 1024 * 1024;
655 lodirtybufspace = hidirtybufspace / 2;
658 * Maximum number of async ops initiated per buf_daemon loop. This is
659 * somewhat of a hack at the moment, we really need to limit ourselves
660 * based on the number of bytes of I/O in-transit that were initiated
664 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
665 bogus_page = vm_page_alloc(&kernel_object,
666 (bogus_offset >> PAGE_SHIFT),
668 vmstats.v_wire_count++;
673 * Initialize the embedded bio structures
676 initbufbio(struct buf *bp)
678 bp->b_bio1.bio_buf = bp;
679 bp->b_bio1.bio_prev = NULL;
680 bp->b_bio1.bio_offset = NOOFFSET;
681 bp->b_bio1.bio_next = &bp->b_bio2;
682 bp->b_bio1.bio_done = NULL;
683 bp->b_bio1.bio_flags = 0;
685 bp->b_bio2.bio_buf = bp;
686 bp->b_bio2.bio_prev = &bp->b_bio1;
687 bp->b_bio2.bio_offset = NOOFFSET;
688 bp->b_bio2.bio_next = NULL;
689 bp->b_bio2.bio_done = NULL;
690 bp->b_bio2.bio_flags = 0;
694 * Reinitialize the embedded bio structures as well as any additional
695 * translation cache layers.
698 reinitbufbio(struct buf *bp)
702 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
703 bio->bio_done = NULL;
704 bio->bio_offset = NOOFFSET;
709 * Push another BIO layer onto an existing BIO and return it. The new
710 * BIO layer may already exist, holding cached translation data.
713 push_bio(struct bio *bio)
717 if ((nbio = bio->bio_next) == NULL) {
718 int index = bio - &bio->bio_buf->b_bio_array[0];
719 if (index >= NBUF_BIO - 1) {
720 panic("push_bio: too many layers bp %p\n",
723 nbio = &bio->bio_buf->b_bio_array[index + 1];
724 bio->bio_next = nbio;
725 nbio->bio_prev = bio;
726 nbio->bio_buf = bio->bio_buf;
727 nbio->bio_offset = NOOFFSET;
728 nbio->bio_done = NULL;
729 nbio->bio_next = NULL;
731 KKASSERT(nbio->bio_done == NULL);
736 * Pop a BIO translation layer, returning the previous layer. The
737 * must have been previously pushed.
740 pop_bio(struct bio *bio)
742 return(bio->bio_prev);
746 clearbiocache(struct bio *bio)
749 bio->bio_offset = NOOFFSET;
757 * Free the KVA allocation for buffer 'bp'.
759 * Must be called from a critical section as this is the only locking for
762 * Since this call frees up buffer space, we call bufspacewakeup().
767 bfreekva(struct buf *bp)
774 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
775 vm_map_lock(&buffer_map);
776 bufspace -= bp->b_kvasize;
777 vm_map_delete(&buffer_map,
778 (vm_offset_t) bp->b_kvabase,
779 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
782 vm_map_unlock(&buffer_map);
783 vm_map_entry_release(count);
793 * Remove the buffer from the appropriate free list.
796 _bremfree(struct buf *bp)
798 if (bp->b_qindex != BQUEUE_NONE) {
799 KASSERT(BUF_REFCNTNB(bp) == 1,
800 ("bremfree: bp %p not locked",bp));
801 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
802 bp->b_qindex = BQUEUE_NONE;
804 if (BUF_REFCNTNB(bp) <= 1)
805 panic("bremfree: removing a buffer not on a queue");
810 bremfree(struct buf *bp)
812 spin_lock_wr(&bufspin);
814 spin_unlock_wr(&bufspin);
818 bremfree_locked(struct buf *bp)
826 * Get a buffer with the specified data. Look in the cache first. We
827 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
828 * is set, the buffer is valid and we do not have to do anything ( see
834 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
838 bp = getblk(vp, loffset, size, 0, 0);
841 /* if not found in cache, do some I/O */
842 if ((bp->b_flags & B_CACHE) == 0) {
844 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
845 bp->b_cmd = BUF_CMD_READ;
846 bp->b_bio1.bio_done = biodone_sync;
847 bp->b_bio1.bio_flags |= BIO_SYNC;
848 vfs_busy_pages(vp, bp);
849 vn_strategy(vp, &bp->b_bio1);
851 return (biowait(&bp->b_bio1, "biord"));
859 * Operates like bread, but also starts asynchronous I/O on
860 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
861 * to initiating I/O . If B_CACHE is set, the buffer is valid
862 * and we do not have to do anything.
867 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
868 int *rabsize, int cnt, struct buf **bpp)
870 struct buf *bp, *rabp;
872 int rv = 0, readwait = 0;
874 *bpp = bp = getblk(vp, loffset, size, 0, 0);
876 /* if not found in cache, do some I/O */
877 if ((bp->b_flags & B_CACHE) == 0) {
879 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
880 bp->b_cmd = BUF_CMD_READ;
881 bp->b_bio1.bio_done = biodone_sync;
882 bp->b_bio1.bio_flags |= BIO_SYNC;
883 vfs_busy_pages(vp, bp);
884 vn_strategy(vp, &bp->b_bio1);
889 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
890 if (inmem(vp, *raoffset))
892 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
894 if ((rabp->b_flags & B_CACHE) == 0) {
896 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
897 rabp->b_cmd = BUF_CMD_READ;
898 vfs_busy_pages(vp, rabp);
900 vn_strategy(vp, &rabp->b_bio1);
907 rv = biowait(&bp->b_bio1, "biord");
914 * Synchronous write, waits for completion.
916 * Write, release buffer on completion. (Done by iodone
917 * if async). Do not bother writing anything if the buffer
920 * Note that we set B_CACHE here, indicating that buffer is
921 * fully valid and thus cacheable. This is true even of NFS
922 * now so we set it generally. This could be set either here
923 * or in biodone() since the I/O is synchronous. We put it
927 bwrite(struct buf *bp)
931 if (bp->b_flags & B_INVAL) {
935 if (BUF_REFCNTNB(bp) == 0)
936 panic("bwrite: buffer is not busy???");
938 /* Mark the buffer clean */
941 bp->b_flags &= ~(B_ERROR | B_EINTR);
942 bp->b_flags |= B_CACHE;
943 bp->b_cmd = BUF_CMD_WRITE;
944 bp->b_bio1.bio_done = biodone_sync;
945 bp->b_bio1.bio_flags |= BIO_SYNC;
946 vfs_busy_pages(bp->b_vp, bp);
949 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
950 * valid for vnode-backed buffers.
952 bp->b_runningbufspace = bp->b_bufsize;
953 if (bp->b_runningbufspace) {
954 runningbufspace += bp->b_runningbufspace;
958 vn_strategy(bp->b_vp, &bp->b_bio1);
959 error = biowait(&bp->b_bio1, "biows");
967 * Asynchronous write. Start output on a buffer, but do not wait for
968 * it to complete. The buffer is released when the output completes.
970 * bwrite() ( or the VOP routine anyway ) is responsible for handling
971 * B_INVAL buffers. Not us.
974 bawrite(struct buf *bp)
976 if (bp->b_flags & B_INVAL) {
980 if (BUF_REFCNTNB(bp) == 0)
981 panic("bwrite: buffer is not busy???");
983 /* Mark the buffer clean */
986 bp->b_flags &= ~(B_ERROR | B_EINTR);
987 bp->b_flags |= B_CACHE;
988 bp->b_cmd = BUF_CMD_WRITE;
989 KKASSERT(bp->b_bio1.bio_done == NULL);
990 vfs_busy_pages(bp->b_vp, bp);
993 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
994 * valid for vnode-backed buffers.
996 bp->b_runningbufspace = bp->b_bufsize;
997 if (bp->b_runningbufspace) {
998 runningbufspace += bp->b_runningbufspace;
1003 vn_strategy(bp->b_vp, &bp->b_bio1);
1009 * Ordered write. Start output on a buffer, and flag it so that the
1010 * device will write it in the order it was queued. The buffer is
1011 * released when the output completes. bwrite() ( or the VOP routine
1012 * anyway ) is responsible for handling B_INVAL buffers.
1015 bowrite(struct buf *bp)
1017 bp->b_flags |= B_ORDERED;
1025 * Delayed write. (Buffer is marked dirty). Do not bother writing
1026 * anything if the buffer is marked invalid.
1028 * Note that since the buffer must be completely valid, we can safely
1029 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1030 * biodone() in order to prevent getblk from writing the buffer
1031 * out synchronously.
1034 bdwrite(struct buf *bp)
1036 if (BUF_REFCNTNB(bp) == 0)
1037 panic("bdwrite: buffer is not busy");
1039 if (bp->b_flags & B_INVAL) {
1046 * Set B_CACHE, indicating that the buffer is fully valid. This is
1047 * true even of NFS now.
1049 bp->b_flags |= B_CACHE;
1052 * This bmap keeps the system from needing to do the bmap later,
1053 * perhaps when the system is attempting to do a sync. Since it
1054 * is likely that the indirect block -- or whatever other datastructure
1055 * that the filesystem needs is still in memory now, it is a good
1056 * thing to do this. Note also, that if the pageout daemon is
1057 * requesting a sync -- there might not be enough memory to do
1058 * the bmap then... So, this is important to do.
1060 if (bp->b_bio2.bio_offset == NOOFFSET) {
1061 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1062 NULL, NULL, BUF_CMD_WRITE);
1066 * Because the underlying pages may still be mapped and
1067 * writable trying to set the dirty buffer (b_dirtyoff/end)
1068 * range here will be inaccurate.
1070 * However, we must still clean the pages to satisfy the
1071 * vnode_pager and pageout daemon, so theythink the pages
1072 * have been "cleaned". What has really occured is that
1073 * they've been earmarked for later writing by the buffer
1076 * So we get the b_dirtyoff/end update but will not actually
1077 * depend on it (NFS that is) until the pages are busied for
1080 vfs_clean_pages(bp);
1084 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1085 * due to the softdep code.
1092 * Turn buffer into delayed write request by marking it B_DELWRI.
1093 * B_RELBUF and B_NOCACHE must be cleared.
1095 * We reassign the buffer to itself to properly update it in the
1096 * dirty/clean lists.
1098 * Must be called from a critical section.
1099 * The buffer must be on BQUEUE_NONE.
1102 bdirty(struct buf *bp)
1104 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1105 if (bp->b_flags & B_NOCACHE) {
1106 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1107 bp->b_flags &= ~B_NOCACHE;
1109 if (bp->b_flags & B_INVAL) {
1110 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1112 bp->b_flags &= ~B_RELBUF;
1114 if ((bp->b_flags & B_DELWRI) == 0) {
1115 bp->b_flags |= B_DELWRI;
1117 atomic_add_int(&dirtybufcount, 1);
1118 dirtybufspace += bp->b_bufsize;
1119 if (bp->b_flags & B_HEAVY) {
1120 atomic_add_int(&dirtybufcounthw, 1);
1121 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
1128 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1129 * needs to be flushed with a different buf_daemon thread to avoid
1130 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1133 bheavy(struct buf *bp)
1135 if ((bp->b_flags & B_HEAVY) == 0) {
1136 bp->b_flags |= B_HEAVY;
1137 if (bp->b_flags & B_DELWRI) {
1138 atomic_add_int(&dirtybufcounthw, 1);
1139 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
1147 * Clear B_DELWRI for buffer.
1149 * Must be called from a critical section.
1151 * The buffer is typically on BQUEUE_NONE but there is one case in
1152 * brelse() that calls this function after placing the buffer on
1153 * a different queue.
1158 bundirty(struct buf *bp)
1160 if (bp->b_flags & B_DELWRI) {
1161 bp->b_flags &= ~B_DELWRI;
1163 atomic_subtract_int(&dirtybufcount, 1);
1164 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
1165 if (bp->b_flags & B_HEAVY) {
1166 atomic_subtract_int(&dirtybufcounthw, 1);
1167 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
1169 bd_signal(bp->b_bufsize);
1172 * Since it is now being written, we can clear its deferred write flag.
1174 bp->b_flags &= ~B_DEFERRED;
1180 * Release a busy buffer and, if requested, free its resources. The
1181 * buffer will be stashed in the appropriate bufqueue[] allowing it
1182 * to be accessed later as a cache entity or reused for other purposes.
1187 brelse(struct buf *bp)
1190 int saved_flags = bp->b_flags;
1193 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1196 * If B_NOCACHE is set we are being asked to destroy the buffer and
1197 * its backing store. Clear B_DELWRI.
1199 * B_NOCACHE is set in two cases: (1) when the caller really wants
1200 * to destroy the buffer and backing store and (2) when the caller
1201 * wants to destroy the buffer and backing store after a write
1204 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1208 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1210 * A re-dirtied buffer is only subject to destruction
1211 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1213 /* leave buffer intact */
1214 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1215 (bp->b_bufsize <= 0)) {
1217 * Either a failed read or we were asked to free or not
1218 * cache the buffer. This path is reached with B_DELWRI
1219 * set only if B_INVAL is already set. B_NOCACHE governs
1220 * backing store destruction.
1222 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1223 * buffer cannot be immediately freed.
1225 bp->b_flags |= B_INVAL;
1226 if (LIST_FIRST(&bp->b_dep) != NULL) {
1231 if (bp->b_flags & B_DELWRI) {
1232 atomic_subtract_int(&dirtybufcount, 1);
1233 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
1234 if (bp->b_flags & B_HEAVY) {
1235 atomic_subtract_int(&dirtybufcounthw, 1);
1236 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
1238 bd_signal(bp->b_bufsize);
1240 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1244 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1245 * If vfs_vmio_release() is called with either bit set, the
1246 * underlying pages may wind up getting freed causing a previous
1247 * write (bdwrite()) to get 'lost' because pages associated with
1248 * a B_DELWRI bp are marked clean. Pages associated with a
1249 * B_LOCKED buffer may be mapped by the filesystem.
1251 * If we want to release the buffer ourselves (rather then the
1252 * originator asking us to release it), give the originator a
1253 * chance to countermand the release by setting B_LOCKED.
1255 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1256 * if B_DELWRI is set.
1258 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1259 * on pages to return pages to the VM page queues.
1261 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1262 bp->b_flags &= ~B_RELBUF;
1263 } else if (vm_page_count_severe()) {
1264 if (LIST_FIRST(&bp->b_dep) != NULL) {
1266 buf_deallocate(bp); /* can set B_LOCKED */
1269 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1270 bp->b_flags &= ~B_RELBUF;
1272 bp->b_flags |= B_RELBUF;
1276 * Make sure b_cmd is clear. It may have already been cleared by
1279 * At this point destroying the buffer is governed by the B_INVAL
1280 * or B_RELBUF flags.
1282 bp->b_cmd = BUF_CMD_DONE;
1285 * VMIO buffer rundown. Make sure the VM page array is restored
1286 * after an I/O may have replaces some of the pages with bogus pages
1287 * in order to not destroy dirty pages in a fill-in read.
1289 * Note that due to the code above, if a buffer is marked B_DELWRI
1290 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1291 * B_INVAL may still be set, however.
1293 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1294 * but not the backing store. B_NOCACHE will destroy the backing
1297 * Note that dirty NFS buffers contain byte-granular write ranges
1298 * and should not be destroyed w/ B_INVAL even if the backing store
1301 if (bp->b_flags & B_VMIO) {
1303 * Rundown for VMIO buffers which are not dirty NFS buffers.
1315 * Get the base offset and length of the buffer. Note that
1316 * in the VMIO case if the buffer block size is not
1317 * page-aligned then b_data pointer may not be page-aligned.
1318 * But our b_xio.xio_pages array *IS* page aligned.
1320 * block sizes less then DEV_BSIZE (usually 512) are not
1321 * supported due to the page granularity bits (m->valid,
1322 * m->dirty, etc...).
1324 * See man buf(9) for more information
1327 resid = bp->b_bufsize;
1328 foff = bp->b_loffset;
1331 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1332 m = bp->b_xio.xio_pages[i];
1333 vm_page_flag_clear(m, PG_ZERO);
1335 * If we hit a bogus page, fixup *all* of them
1336 * now. Note that we left these pages wired
1337 * when we removed them so they had better exist,
1338 * and they cannot be ripped out from under us so
1339 * no critical section protection is necessary.
1341 if (m == bogus_page) {
1343 poff = OFF_TO_IDX(bp->b_loffset);
1345 for (j = i; j < bp->b_xio.xio_npages; j++) {
1348 mtmp = bp->b_xio.xio_pages[j];
1349 if (mtmp == bogus_page) {
1350 mtmp = vm_page_lookup(obj, poff + j);
1352 panic("brelse: page missing");
1354 bp->b_xio.xio_pages[j] = mtmp;
1358 if ((bp->b_flags & B_INVAL) == 0) {
1359 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1360 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1362 m = bp->b_xio.xio_pages[i];
1366 * Invalidate the backing store if B_NOCACHE is set
1367 * (e.g. used with vinvalbuf()). If this is NFS
1368 * we impose a requirement that the block size be
1369 * a multiple of PAGE_SIZE and create a temporary
1370 * hack to basically invalidate the whole page. The
1371 * problem is that NFS uses really odd buffer sizes
1372 * especially when tracking piecemeal writes and
1373 * it also vinvalbuf()'s a lot, which would result
1374 * in only partial page validation and invalidation
1375 * here. If the file page is mmap()'d, however,
1376 * all the valid bits get set so after we invalidate
1377 * here we would end up with weird m->valid values
1378 * like 0xfc. nfs_getpages() can't handle this so
1379 * we clear all the valid bits for the NFS case
1380 * instead of just some of them.
1382 * The real bug is the VM system having to set m->valid
1383 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1384 * itself is an artifact of the whole 512-byte
1385 * granular mess that exists to support odd block
1386 * sizes and UFS meta-data block sizes (e.g. 6144).
1387 * A complete rewrite is required.
1391 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1392 int poffset = foff & PAGE_MASK;
1395 presid = PAGE_SIZE - poffset;
1396 if (bp->b_vp->v_tag == VT_NFS &&
1397 bp->b_vp->v_type == VREG) {
1399 } else if (presid > resid) {
1402 KASSERT(presid >= 0, ("brelse: extra page"));
1403 vm_page_set_invalid(m, poffset, presid);
1405 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1406 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1408 if (bp->b_flags & (B_INVAL | B_RELBUF))
1409 vfs_vmio_release(bp);
1413 * Rundown for non-VMIO buffers.
1415 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1419 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1426 if (bp->b_qindex != BQUEUE_NONE)
1427 panic("brelse: free buffer onto another queue???");
1428 if (BUF_REFCNTNB(bp) > 1) {
1429 /* Temporary panic to verify exclusive locking */
1430 /* This panic goes away when we allow shared refs */
1431 panic("brelse: multiple refs");
1437 * Figure out the correct queue to place the cleaned up buffer on.
1438 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1439 * disassociated from their vnode.
1441 spin_lock_wr(&bufspin);
1442 if (bp->b_flags & B_LOCKED) {
1444 * Buffers that are locked are placed in the locked queue
1445 * immediately, regardless of their state.
1447 bp->b_qindex = BQUEUE_LOCKED;
1448 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1449 } else if (bp->b_bufsize == 0) {
1451 * Buffers with no memory. Due to conditionals near the top
1452 * of brelse() such buffers should probably already be
1453 * marked B_INVAL and disassociated from their vnode.
1455 bp->b_flags |= B_INVAL;
1456 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1457 KKASSERT((bp->b_flags & B_HASHED) == 0);
1458 if (bp->b_kvasize) {
1459 bp->b_qindex = BQUEUE_EMPTYKVA;
1461 bp->b_qindex = BQUEUE_EMPTY;
1463 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1464 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1466 * Buffers with junk contents. Again these buffers had better
1467 * already be disassociated from their vnode.
1469 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1470 KKASSERT((bp->b_flags & B_HASHED) == 0);
1471 bp->b_flags |= B_INVAL;
1472 bp->b_qindex = BQUEUE_CLEAN;
1473 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1476 * Remaining buffers. These buffers are still associated with
1479 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1481 bp->b_qindex = BQUEUE_DIRTY;
1482 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1484 case B_DELWRI | B_HEAVY:
1485 bp->b_qindex = BQUEUE_DIRTY_HW;
1486 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1491 * NOTE: Buffers are always placed at the end of the
1492 * queue. If B_AGE is not set the buffer will cycle
1493 * through the queue twice.
1495 bp->b_qindex = BQUEUE_CLEAN;
1496 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1500 spin_unlock_wr(&bufspin);
1503 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1504 * on the correct queue.
1506 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1510 * The bp is on an appropriate queue unless locked. If it is not
1511 * locked or dirty we can wakeup threads waiting for buffer space.
1513 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1514 * if B_INVAL is set ).
1516 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1520 * Something we can maybe free or reuse
1522 if (bp->b_bufsize || bp->b_kvasize)
1526 * Clean up temporary flags and unlock the buffer.
1528 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
1535 * Release a buffer back to the appropriate queue but do not try to free
1536 * it. The buffer is expected to be used again soon.
1538 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1539 * biodone() to requeue an async I/O on completion. It is also used when
1540 * known good buffers need to be requeued but we think we may need the data
1543 * XXX we should be able to leave the B_RELBUF hint set on completion.
1548 bqrelse(struct buf *bp)
1550 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1552 if (bp->b_qindex != BQUEUE_NONE)
1553 panic("bqrelse: free buffer onto another queue???");
1554 if (BUF_REFCNTNB(bp) > 1) {
1555 /* do not release to free list */
1556 panic("bqrelse: multiple refs");
1560 spin_lock_wr(&bufspin);
1561 if (bp->b_flags & B_LOCKED) {
1563 * Locked buffers are released to the locked queue. However,
1564 * if the buffer is dirty it will first go into the dirty
1565 * queue and later on after the I/O completes successfully it
1566 * will be released to the locked queue.
1568 bp->b_qindex = BQUEUE_LOCKED;
1569 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1570 } else if (bp->b_flags & B_DELWRI) {
1571 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1572 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1573 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1574 } else if (vm_page_count_severe()) {
1576 * We are too low on memory, we have to try to free the
1577 * buffer (most importantly: the wired pages making up its
1578 * backing store) *now*.
1580 spin_unlock_wr(&bufspin);
1584 bp->b_qindex = BQUEUE_CLEAN;
1585 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1587 spin_unlock_wr(&bufspin);
1589 if ((bp->b_flags & B_LOCKED) == 0 &&
1590 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1595 * Something we can maybe free or reuse.
1597 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1601 * Final cleanup and unlock. Clear bits that are only used while a
1602 * buffer is actively locked.
1604 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
1611 * Return backing pages held by the buffer 'bp' back to the VM system
1612 * if possible. The pages are freed if they are no longer valid or
1613 * attempt to free if it was used for direct I/O otherwise they are
1614 * sent to the page cache.
1616 * Pages that were marked busy are left alone and skipped.
1618 * The KVA mapping (b_data) for the underlying pages is removed by
1622 vfs_vmio_release(struct buf *bp)
1628 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1629 m = bp->b_xio.xio_pages[i];
1630 bp->b_xio.xio_pages[i] = NULL;
1632 * In order to keep page LRU ordering consistent, put
1633 * everything on the inactive queue.
1635 vm_page_unwire(m, 0);
1637 * We don't mess with busy pages, it is
1638 * the responsibility of the process that
1639 * busied the pages to deal with them.
1641 if ((m->flags & PG_BUSY) || (m->busy != 0))
1644 if (m->wire_count == 0) {
1645 vm_page_flag_clear(m, PG_ZERO);
1647 * Might as well free the page if we can and it has
1648 * no valid data. We also free the page if the
1649 * buffer was used for direct I/O.
1652 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1653 m->hold_count == 0) {
1655 vm_page_protect(m, VM_PROT_NONE);
1659 if (bp->b_flags & B_DIRECT) {
1660 vm_page_try_to_free(m);
1661 } else if (vm_page_count_severe()) {
1662 vm_page_try_to_cache(m);
1667 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1668 if (bp->b_bufsize) {
1672 bp->b_xio.xio_npages = 0;
1673 bp->b_flags &= ~B_VMIO;
1674 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1685 * Implement clustered async writes for clearing out B_DELWRI buffers.
1686 * This is much better then the old way of writing only one buffer at
1687 * a time. Note that we may not be presented with the buffers in the
1688 * correct order, so we search for the cluster in both directions.
1690 * The buffer is locked on call.
1693 vfs_bio_awrite(struct buf *bp)
1697 off_t loffset = bp->b_loffset;
1698 struct vnode *vp = bp->b_vp;
1705 * right now we support clustered writing only to regular files. If
1706 * we find a clusterable block we could be in the middle of a cluster
1707 * rather then at the beginning.
1709 * NOTE: b_bio1 contains the logical loffset and is aliased
1710 * to b_loffset. b_bio2 contains the translated block number.
1712 if ((vp->v_type == VREG) &&
1713 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1714 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1716 size = vp->v_mount->mnt_stat.f_iosize;
1718 for (i = size; i < MAXPHYS; i += size) {
1719 if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) &&
1720 BUF_REFCNT(bpa) == 0 &&
1721 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1722 (B_DELWRI | B_CLUSTEROK)) &&
1723 (bpa->b_bufsize == size)) {
1724 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1725 (bpa->b_bio2.bio_offset !=
1726 bp->b_bio2.bio_offset + i))
1732 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1733 if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) &&
1734 BUF_REFCNT(bpa) == 0 &&
1735 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1736 (B_DELWRI | B_CLUSTEROK)) &&
1737 (bpa->b_bufsize == size)) {
1738 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1739 (bpa->b_bio2.bio_offset !=
1740 bp->b_bio2.bio_offset - j))
1750 * this is a possible cluster write
1752 if (nbytes != size) {
1754 nwritten = cluster_wbuild(vp, size,
1755 loffset - j, nbytes);
1761 * default (old) behavior, writing out only one block
1763 * XXX returns b_bufsize instead of b_bcount for nwritten?
1765 nwritten = bp->b_bufsize;
1775 * Find and initialize a new buffer header, freeing up existing buffers
1776 * in the bufqueues as necessary. The new buffer is returned locked.
1778 * Important: B_INVAL is not set. If the caller wishes to throw the
1779 * buffer away, the caller must set B_INVAL prior to calling brelse().
1782 * We have insufficient buffer headers
1783 * We have insufficient buffer space
1784 * buffer_map is too fragmented ( space reservation fails )
1785 * If we have to flush dirty buffers ( but we try to avoid this )
1787 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1788 * Instead we ask the buf daemon to do it for us. We attempt to
1789 * avoid piecemeal wakeups of the pageout daemon.
1794 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1800 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1801 static int flushingbufs;
1804 * We can't afford to block since we might be holding a vnode lock,
1805 * which may prevent system daemons from running. We deal with
1806 * low-memory situations by proactively returning memory and running
1807 * async I/O rather then sync I/O.
1811 --getnewbufrestarts;
1813 ++getnewbufrestarts;
1816 * Setup for scan. If we do not have enough free buffers,
1817 * we setup a degenerate case that immediately fails. Note
1818 * that if we are specially marked process, we are allowed to
1819 * dip into our reserves.
1821 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1823 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1824 * However, there are a number of cases (defragging, reusing, ...)
1825 * where we cannot backup.
1827 nqindex = BQUEUE_EMPTYKVA;
1828 spin_lock_wr(&bufspin);
1829 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1833 * If no EMPTYKVA buffers and we are either
1834 * defragging or reusing, locate a CLEAN buffer
1835 * to free or reuse. If bufspace useage is low
1836 * skip this step so we can allocate a new buffer.
1838 if (defrag || bufspace >= lobufspace) {
1839 nqindex = BQUEUE_CLEAN;
1840 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1844 * If we could not find or were not allowed to reuse a
1845 * CLEAN buffer, check to see if it is ok to use an EMPTY
1846 * buffer. We can only use an EMPTY buffer if allocating
1847 * its KVA would not otherwise run us out of buffer space.
1849 if (nbp == NULL && defrag == 0 &&
1850 bufspace + maxsize < hibufspace) {
1851 nqindex = BQUEUE_EMPTY;
1852 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1857 * Run scan, possibly freeing data and/or kva mappings on the fly
1860 * WARNING! bufspin is held!
1862 while ((bp = nbp) != NULL) {
1863 int qindex = nqindex;
1865 nbp = TAILQ_NEXT(bp, b_freelist);
1868 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1869 * cycles through the queue twice before being selected.
1871 if (qindex == BQUEUE_CLEAN &&
1872 (bp->b_flags & B_AGE) == 0 && nbp) {
1873 bp->b_flags |= B_AGE;
1874 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1875 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1880 * Calculate next bp ( we can only use it if we do not block
1881 * or do other fancy things ).
1886 nqindex = BQUEUE_EMPTYKVA;
1887 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1890 case BQUEUE_EMPTYKVA:
1891 nqindex = BQUEUE_CLEAN;
1892 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1906 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1909 * Note: we no longer distinguish between VMIO and non-VMIO
1913 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1916 * If we are defragging then we need a buffer with
1917 * b_kvasize != 0. XXX this situation should no longer
1918 * occur, if defrag is non-zero the buffer's b_kvasize
1919 * should also be non-zero at this point. XXX
1921 if (defrag && bp->b_kvasize == 0) {
1922 kprintf("Warning: defrag empty buffer %p\n", bp);
1927 * Start freeing the bp. This is somewhat involved. nbp
1928 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1929 * on the clean list must be disassociated from their
1930 * current vnode. Buffers on the empty[kva] lists have
1931 * already been disassociated.
1934 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1935 spin_unlock_wr(&bufspin);
1936 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1937 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1940 if (bp->b_qindex != qindex) {
1941 spin_unlock_wr(&bufspin);
1942 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1946 bremfree_locked(bp);
1947 spin_unlock_wr(&bufspin);
1950 * Dependancies must be handled before we disassociate the
1953 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1954 * be immediately disassociated. HAMMER then becomes
1955 * responsible for releasing the buffer.
1957 * NOTE: bufspin is UNLOCKED now.
1959 if (LIST_FIRST(&bp->b_dep) != NULL) {
1963 if (bp->b_flags & B_LOCKED) {
1967 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1970 if (qindex == BQUEUE_CLEAN) {
1972 if (bp->b_flags & B_VMIO) {
1974 vfs_vmio_release(bp);
1983 * NOTE: nbp is now entirely invalid. We can only restart
1984 * the scan from this point on.
1986 * Get the rest of the buffer freed up. b_kva* is still
1987 * valid after this operation.
1990 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1991 KKASSERT((bp->b_flags & B_HASHED) == 0);
1994 * critical section protection is not required when
1995 * scrapping a buffer's contents because it is already
1998 if (bp->b_bufsize) {
2004 bp->b_flags = B_BNOCLIP;
2005 bp->b_cmd = BUF_CMD_DONE;
2010 bp->b_xio.xio_npages = 0;
2011 bp->b_dirtyoff = bp->b_dirtyend = 0;
2013 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2015 if (blkflags & GETBLK_BHEAVY)
2016 bp->b_flags |= B_HEAVY;
2019 * If we are defragging then free the buffer.
2022 bp->b_flags |= B_INVAL;
2030 * If we are overcomitted then recover the buffer and its
2031 * KVM space. This occurs in rare situations when multiple
2032 * processes are blocked in getnewbuf() or allocbuf().
2034 if (bufspace >= hibufspace)
2036 if (flushingbufs && bp->b_kvasize != 0) {
2037 bp->b_flags |= B_INVAL;
2042 if (bufspace < lobufspace)
2045 /* NOT REACHED, bufspin not held */
2049 * If we exhausted our list, sleep as appropriate. We may have to
2050 * wakeup various daemons and write out some dirty buffers.
2052 * Generally we are sleeping due to insufficient buffer space.
2054 * NOTE: bufspin is held if bp is NULL, else it is not held.
2060 spin_unlock_wr(&bufspin);
2062 flags = VFS_BIO_NEED_BUFSPACE;
2064 } else if (bufspace >= hibufspace) {
2066 flags = VFS_BIO_NEED_BUFSPACE;
2069 flags = VFS_BIO_NEED_ANY;
2072 needsbuffer |= flags;
2073 bd_speedup(); /* heeeelp */
2074 while (needsbuffer & flags) {
2075 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
2080 * We finally have a valid bp. We aren't quite out of the
2081 * woods, we still have to reserve kva space. In order
2082 * to keep fragmentation sane we only allocate kva in
2085 * (bufspin is not held)
2087 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2089 if (maxsize != bp->b_kvasize) {
2090 vm_offset_t addr = 0;
2096 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2097 vm_map_lock(&buffer_map);
2099 if (vm_map_findspace(&buffer_map,
2100 vm_map_min(&buffer_map), maxsize,
2101 maxsize, 0, &addr)) {
2103 * Uh oh. Buffer map is too fragmented. We
2104 * must defragment the map.
2106 vm_map_unlock(&buffer_map);
2107 vm_map_entry_release(count);
2110 bp->b_flags |= B_INVAL;
2116 vm_map_insert(&buffer_map, &count,
2118 addr, addr + maxsize,
2120 VM_PROT_ALL, VM_PROT_ALL,
2123 bp->b_kvabase = (caddr_t) addr;
2124 bp->b_kvasize = maxsize;
2125 bufspace += bp->b_kvasize;
2128 vm_map_unlock(&buffer_map);
2129 vm_map_entry_release(count);
2132 bp->b_data = bp->b_kvabase;
2138 * This routine is called in an emergency to recover VM pages from the
2139 * buffer cache by cashing in clean buffers. The idea is to recover
2140 * enough pages to be able to satisfy a stuck bio_page_alloc().
2143 recoverbufpages(void)
2150 spin_lock_wr(&bufspin);
2151 while (bytes < MAXBSIZE) {
2152 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2157 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2158 * cycles through the queue twice before being selected.
2160 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
2161 bp->b_flags |= B_AGE;
2162 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
2163 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
2171 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
2172 KKASSERT((bp->b_flags & B_DELWRI) == 0);
2175 * Start freeing the bp. This is somewhat involved.
2177 * Buffers on the clean list must be disassociated from
2178 * their current vnode
2181 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2182 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp);
2183 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
2186 if (bp->b_qindex != BQUEUE_CLEAN) {
2187 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp, bp->b_qindex);
2191 bremfree_locked(bp);
2192 spin_unlock_wr(&bufspin);
2195 * Dependancies must be handled before we disassociate the
2198 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2199 * be immediately disassociated. HAMMER then becomes
2200 * responsible for releasing the buffer.
2202 if (LIST_FIRST(&bp->b_dep) != NULL) {
2204 if (bp->b_flags & B_LOCKED) {
2206 spin_lock_wr(&bufspin);
2209 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2212 bytes += bp->b_bufsize;
2215 if (bp->b_flags & B_VMIO) {
2216 bp->b_flags |= B_DIRECT; /* try to free pages */
2217 vfs_vmio_release(bp);
2222 KKASSERT(bp->b_vp == NULL);
2223 KKASSERT((bp->b_flags & B_HASHED) == 0);
2226 * critical section protection is not required when
2227 * scrapping a buffer's contents because it is already
2234 bp->b_flags = B_BNOCLIP;
2235 bp->b_cmd = BUF_CMD_DONE;
2240 bp->b_xio.xio_npages = 0;
2241 bp->b_dirtyoff = bp->b_dirtyend = 0;
2243 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2245 bp->b_flags |= B_INVAL;
2248 spin_lock_wr(&bufspin);
2250 spin_unlock_wr(&bufspin);
2257 * Buffer flushing daemon. Buffers are normally flushed by the
2258 * update daemon but if it cannot keep up this process starts to
2259 * take the load in an attempt to prevent getnewbuf() from blocking.
2261 * Once a flush is initiated it does not stop until the number
2262 * of buffers falls below lodirtybuffers, but we will wake up anyone
2263 * waiting at the mid-point.
2266 static struct kproc_desc buf_kp = {
2271 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2272 kproc_start, &buf_kp)
2274 static struct kproc_desc bufhw_kp = {
2279 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2280 kproc_start, &bufhw_kp)
2288 * This process needs to be suspended prior to shutdown sync.
2290 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2291 bufdaemon_td, SHUTDOWN_PRI_LAST);
2292 curthread->td_flags |= TDF_SYSTHREAD;
2295 * This process is allowed to take the buffer cache to the limit
2300 kproc_suspend_loop();
2303 * Do the flush as long as the number of dirty buffers
2304 * (including those running) exceeds lodirtybufspace.
2306 * When flushing limit running I/O to hirunningspace
2307 * Do the flush. Limit the amount of in-transit I/O we
2308 * allow to build up, otherwise we would completely saturate
2309 * the I/O system. Wakeup any waiting processes before we
2310 * normally would so they can run in parallel with our drain.
2312 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2313 * but because we split the operation into two threads we
2314 * have to cut it in half for each thread.
2316 waitrunningbufspace();
2317 limit = lodirtybufspace / 2;
2318 while (runningbufspace + dirtybufspace > limit ||
2319 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
2320 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2322 if (runningbufspace < hirunningspace)
2324 waitrunningbufspace();
2328 * We reached our low water mark, reset the
2329 * request and sleep until we are needed again.
2330 * The sleep is just so the suspend code works.
2332 spin_lock_wr(&needsbuffer_spin);
2333 if (bd_request == 0) {
2334 ssleep(&bd_request, &needsbuffer_spin, 0,
2338 spin_unlock_wr(&needsbuffer_spin);
2348 * This process needs to be suspended prior to shutdown sync.
2350 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2351 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2352 curthread->td_flags |= TDF_SYSTHREAD;
2355 * This process is allowed to take the buffer cache to the limit
2360 kproc_suspend_loop();
2363 * Do the flush. Limit the amount of in-transit I/O we
2364 * allow to build up, otherwise we would completely saturate
2365 * the I/O system. Wakeup any waiting processes before we
2366 * normally would so they can run in parallel with our drain.
2368 * Once we decide to flush push the queued I/O up to
2369 * hirunningspace in order to trigger bursting by the bioq
2372 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2373 * but because we split the operation into two threads we
2374 * have to cut it in half for each thread.
2376 waitrunningbufspace();
2377 limit = lodirtybufspace / 2;
2378 while (runningbufspace + dirtybufspacehw > limit ||
2379 dirtybufcounthw >= nbuf / 2) {
2380 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2382 if (runningbufspace < hirunningspace)
2384 waitrunningbufspace();
2388 * We reached our low water mark, reset the
2389 * request and sleep until we are needed again.
2390 * The sleep is just so the suspend code works.
2392 spin_lock_wr(&needsbuffer_spin);
2393 if (bd_request_hw == 0) {
2394 ssleep(&bd_request_hw, &needsbuffer_spin, 0,
2398 spin_unlock_wr(&needsbuffer_spin);
2405 * Try to flush a buffer in the dirty queue. We must be careful to
2406 * free up B_INVAL buffers instead of write them, which NFS is
2407 * particularly sensitive to.
2409 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2410 * that we really want to try to get the buffer out and reuse it
2411 * due to the write load on the machine.
2414 flushbufqueues(bufq_type_t q)
2420 spin_lock_wr(&bufspin);
2423 bp = TAILQ_FIRST(&bufqueues[q]);
2425 KASSERT((bp->b_flags & B_DELWRI),
2426 ("unexpected clean buffer %p", bp));
2428 if (bp->b_flags & B_DELWRI) {
2429 if (bp->b_flags & B_INVAL) {
2430 spin_unlock_wr(&bufspin);
2432 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2433 panic("flushbufqueues: locked buf");
2439 if (LIST_FIRST(&bp->b_dep) != NULL &&
2440 (bp->b_flags & B_DEFERRED) == 0 &&
2441 buf_countdeps(bp, 0)) {
2442 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2443 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2445 bp->b_flags |= B_DEFERRED;
2446 bp = TAILQ_FIRST(&bufqueues[q]);
2451 * Only write it out if we can successfully lock
2452 * it. If the buffer has a dependancy,
2453 * buf_checkwrite must also return 0 for us to
2454 * be able to initate the write.
2456 * If the buffer is flagged B_ERROR it may be
2457 * requeued over and over again, we try to
2458 * avoid a live lock.
2460 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
2461 spin_unlock_wr(&bufspin);
2463 if (LIST_FIRST(&bp->b_dep) != NULL &&
2464 buf_checkwrite(bp)) {
2467 } else if (bp->b_flags & B_ERROR) {
2468 tsleep(bp, 0, "bioer", 1);
2469 bp->b_flags &= ~B_AGE;
2472 bp->b_flags |= B_AGE;
2479 bp = TAILQ_NEXT(bp, b_freelist);
2482 spin_unlock_wr(&bufspin);
2489 * Returns true if no I/O is needed to access the associated VM object.
2490 * This is like findblk except it also hunts around in the VM system for
2493 * Note that we ignore vm_page_free() races from interrupts against our
2494 * lookup, since if the caller is not protected our return value will not
2495 * be any more valid then otherwise once we exit the critical section.
2498 inmem(struct vnode *vp, off_t loffset)
2501 vm_offset_t toff, tinc, size;
2504 if (findblk(vp, loffset, FINDBLK_TEST))
2506 if (vp->v_mount == NULL)
2508 if ((obj = vp->v_object) == NULL)
2512 if (size > vp->v_mount->mnt_stat.f_iosize)
2513 size = vp->v_mount->mnt_stat.f_iosize;
2515 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2516 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2520 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2521 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2522 if (vm_page_is_valid(m,
2523 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2532 * Locate and return the specified buffer. Unless flagged otherwise,
2533 * a locked buffer will be returned if it exists or NULL if it does not.
2535 * findblk()'d buffers are still on the bufqueues and if you intend
2536 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2537 * and possibly do other stuff to it.
2539 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2540 * for locking the buffer and ensuring that it remains
2541 * the desired buffer after locking.
2543 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2544 * to acquire the lock we return NULL, even if the
2547 * (0) - Lock the buffer blocking.
2552 findblk(struct vnode *vp, off_t loffset, int flags)
2558 lkflags = LK_EXCLUSIVE;
2559 if (flags & FINDBLK_NBLOCK)
2560 lkflags |= LK_NOWAIT;
2563 lwkt_gettoken(&vlock, &vp->v_token);
2564 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2565 lwkt_reltoken(&vlock);
2566 if (bp == NULL || (flags & FINDBLK_TEST))
2568 if (BUF_LOCK(bp, lkflags)) {
2572 if (bp->b_vp == vp && bp->b_loffset == loffset)
2582 * Similar to getblk() except only returns the buffer if it is
2583 * B_CACHE and requires no other manipulation. Otherwise NULL
2586 * If B_RAM is set the buffer might be just fine, but we return
2587 * NULL anyway because we want the code to fall through to the
2588 * cluster read. Otherwise read-ahead breaks.
2591 getcacheblk(struct vnode *vp, off_t loffset)
2595 bp = findblk(vp, loffset, 0);
2597 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
2598 bp->b_flags &= ~B_AGE;
2611 * Get a block given a specified block and offset into a file/device.
2612 * B_INVAL may or may not be set on return. The caller should clear
2613 * B_INVAL prior to initiating a READ.
2615 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2616 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2617 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2618 * without doing any of those things the system will likely believe
2619 * the buffer to be valid (especially if it is not B_VMIO), and the
2620 * next getblk() will return the buffer with B_CACHE set.
2622 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2623 * an existing buffer.
2625 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2626 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2627 * and then cleared based on the backing VM. If the previous buffer is
2628 * non-0-sized but invalid, B_CACHE will be cleared.
2630 * If getblk() must create a new buffer, the new buffer is returned with
2631 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2632 * case it is returned with B_INVAL clear and B_CACHE set based on the
2635 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2636 * B_CACHE bit is clear.
2638 * What this means, basically, is that the caller should use B_CACHE to
2639 * determine whether the buffer is fully valid or not and should clear
2640 * B_INVAL prior to issuing a read. If the caller intends to validate
2641 * the buffer by loading its data area with something, the caller needs
2642 * to clear B_INVAL. If the caller does this without issuing an I/O,
2643 * the caller should set B_CACHE ( as an optimization ), else the caller
2644 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2645 * a write attempt or if it was a successfull read. If the caller
2646 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2647 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2651 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2652 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2657 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2660 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2664 if (size > MAXBSIZE)
2665 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2666 if (vp->v_object == NULL)
2667 panic("getblk: vnode %p has no object!", vp);
2670 if ((bp = findblk(vp, loffset, FINDBLK_TEST)) != NULL) {
2672 * The buffer was found in the cache, but we need to lock it.
2673 * Even with LK_NOWAIT the lockmgr may break our critical
2674 * section, so double-check the validity of the buffer
2675 * once the lock has been obtained.
2677 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2678 if (blkflags & GETBLK_NOWAIT)
2680 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2681 if (blkflags & GETBLK_PCATCH)
2682 lkflags |= LK_PCATCH;
2683 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2685 if (error == ENOLCK)
2689 /* buffer may have changed on us */
2693 * Once the buffer has been locked, make sure we didn't race
2694 * a buffer recyclement. Buffers that are no longer hashed
2695 * will have b_vp == NULL, so this takes care of that check
2698 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2699 kprintf("Warning buffer %p (vp %p loffset %lld) "
2701 bp, vp, (long long)loffset);
2707 * If SZMATCH any pre-existing buffer must be of the requested
2708 * size or NULL is returned. The caller absolutely does not
2709 * want getblk() to bwrite() the buffer on a size mismatch.
2711 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2717 * All vnode-based buffers must be backed by a VM object.
2719 KKASSERT(bp->b_flags & B_VMIO);
2720 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2721 bp->b_flags &= ~B_AGE;
2724 * Make sure that B_INVAL buffers do not have a cached
2725 * block number translation.
2727 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2728 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2729 " did not have cleared bio_offset cache\n",
2730 bp, vp, (long long)loffset);
2731 clearbiocache(&bp->b_bio2);
2735 * The buffer is locked. B_CACHE is cleared if the buffer is
2738 if (bp->b_flags & B_INVAL)
2739 bp->b_flags &= ~B_CACHE;
2743 * Any size inconsistancy with a dirty buffer or a buffer
2744 * with a softupdates dependancy must be resolved. Resizing
2745 * the buffer in such circumstances can lead to problems.
2747 * Dirty or dependant buffers are written synchronously.
2748 * Other types of buffers are simply released and
2749 * reconstituted as they may be backed by valid, dirty VM
2750 * pages (but not marked B_DELWRI).
2752 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2753 * and may be left over from a prior truncation (and thus
2754 * no longer represent the actual EOF point), so we
2755 * definitely do not want to B_NOCACHE the backing store.
2757 if (size != bp->b_bcount) {
2759 if (bp->b_flags & B_DELWRI) {
2760 bp->b_flags |= B_RELBUF;
2762 } else if (LIST_FIRST(&bp->b_dep)) {
2763 bp->b_flags |= B_RELBUF;
2766 bp->b_flags |= B_RELBUF;
2772 KKASSERT(size <= bp->b_kvasize);
2773 KASSERT(bp->b_loffset != NOOFFSET,
2774 ("getblk: no buffer offset"));
2777 * A buffer with B_DELWRI set and B_CACHE clear must
2778 * be committed before we can return the buffer in
2779 * order to prevent the caller from issuing a read
2780 * ( due to B_CACHE not being set ) and overwriting
2783 * Most callers, including NFS and FFS, need this to
2784 * operate properly either because they assume they
2785 * can issue a read if B_CACHE is not set, or because
2786 * ( for example ) an uncached B_DELWRI might loop due
2787 * to softupdates re-dirtying the buffer. In the latter
2788 * case, B_CACHE is set after the first write completes,
2789 * preventing further loops.
2791 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2792 * above while extending the buffer, we cannot allow the
2793 * buffer to remain with B_CACHE set after the write
2794 * completes or it will represent a corrupt state. To
2795 * deal with this we set B_NOCACHE to scrap the buffer
2798 * XXX Should this be B_RELBUF instead of B_NOCACHE?
2799 * I'm not even sure this state is still possible
2800 * now that getblk() writes out any dirty buffers
2803 * We might be able to do something fancy, like setting
2804 * B_CACHE in bwrite() except if B_DELWRI is already set,
2805 * so the below call doesn't set B_CACHE, but that gets real
2806 * confusing. This is much easier.
2809 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2811 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
2812 "and CACHE clear, b_flags %08x\n",
2813 bp, (intmax_t)bp->b_loffset, bp->b_flags);
2814 bp->b_flags |= B_NOCACHE;
2821 * Buffer is not in-core, create new buffer. The buffer
2822 * returned by getnewbuf() is locked. Note that the returned
2823 * buffer is also considered valid (not marked B_INVAL).
2825 * Calculating the offset for the I/O requires figuring out
2826 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2827 * the mount's f_iosize otherwise. If the vnode does not
2828 * have an associated mount we assume that the passed size is
2831 * Note that vn_isdisk() cannot be used here since it may
2832 * return a failure for numerous reasons. Note that the
2833 * buffer size may be larger then the block size (the caller
2834 * will use block numbers with the proper multiple). Beware
2835 * of using any v_* fields which are part of unions. In
2836 * particular, in DragonFly the mount point overloading
2837 * mechanism uses the namecache only and the underlying
2838 * directory vnode is not a special case.
2842 if (vp->v_type == VBLK || vp->v_type == VCHR)
2844 else if (vp->v_mount)
2845 bsize = vp->v_mount->mnt_stat.f_iosize;
2849 maxsize = size + (loffset & PAGE_MASK);
2850 maxsize = imax(maxsize, bsize);
2852 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
2854 if (slpflags || slptimeo)
2860 * Atomically insert the buffer into the hash, so that it can
2861 * be found by findblk().
2863 * If bgetvp() returns non-zero a collision occured, and the
2864 * bp will not be associated with the vnode.
2866 * Make sure the translation layer has been cleared.
2868 bp->b_loffset = loffset;
2869 bp->b_bio2.bio_offset = NOOFFSET;
2870 /* bp->b_bio2.bio_next = NULL; */
2872 if (bgetvp(vp, bp)) {
2873 bp->b_flags |= B_INVAL;
2879 * All vnode-based buffers must be backed by a VM object.
2881 KKASSERT(vp->v_object != NULL);
2882 bp->b_flags |= B_VMIO;
2883 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2895 * Reacquire a buffer that was previously released to the locked queue,
2896 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2897 * set B_LOCKED (which handles the acquisition race).
2899 * To this end, either B_LOCKED must be set or the dependancy list must be
2905 regetblk(struct buf *bp)
2907 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
2908 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2915 * Get an empty, disassociated buffer of given size. The buffer is
2916 * initially set to B_INVAL.
2918 * critical section protection is not required for the allocbuf()
2919 * call because races are impossible here.
2929 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2931 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2936 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2944 * This code constitutes the buffer memory from either anonymous system
2945 * memory (in the case of non-VMIO operations) or from an associated
2946 * VM object (in the case of VMIO operations). This code is able to
2947 * resize a buffer up or down.
2949 * Note that this code is tricky, and has many complications to resolve
2950 * deadlock or inconsistant data situations. Tread lightly!!!
2951 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2952 * the caller. Calling this code willy nilly can result in the loss of data.
2954 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2955 * B_CACHE for the non-VMIO case.
2957 * This routine does not need to be called from a critical section but you
2958 * must own the buffer.
2963 allocbuf(struct buf *bp, int size)
2965 int newbsize, mbsize;
2968 if (BUF_REFCNT(bp) == 0)
2969 panic("allocbuf: buffer not busy");
2971 if (bp->b_kvasize < size)
2972 panic("allocbuf: buffer too small");
2974 if ((bp->b_flags & B_VMIO) == 0) {
2978 * Just get anonymous memory from the kernel. Don't
2979 * mess with B_CACHE.
2981 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2982 if (bp->b_flags & B_MALLOC)
2985 newbsize = round_page(size);
2987 if (newbsize < bp->b_bufsize) {
2989 * Malloced buffers are not shrunk
2991 if (bp->b_flags & B_MALLOC) {
2993 bp->b_bcount = size;
2995 kfree(bp->b_data, M_BIOBUF);
2996 if (bp->b_bufsize) {
2997 bufmallocspace -= bp->b_bufsize;
3001 bp->b_data = bp->b_kvabase;
3003 bp->b_flags &= ~B_MALLOC;
3009 (vm_offset_t) bp->b_data + newbsize,
3010 (vm_offset_t) bp->b_data + bp->b_bufsize);
3011 } else if (newbsize > bp->b_bufsize) {
3013 * We only use malloced memory on the first allocation.
3014 * and revert to page-allocated memory when the buffer
3017 if ((bufmallocspace < maxbufmallocspace) &&
3018 (bp->b_bufsize == 0) &&
3019 (mbsize <= PAGE_SIZE/2)) {
3021 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3022 bp->b_bufsize = mbsize;
3023 bp->b_bcount = size;
3024 bp->b_flags |= B_MALLOC;
3025 bufmallocspace += mbsize;
3031 * If the buffer is growing on its other-than-first
3032 * allocation, then we revert to the page-allocation
3035 if (bp->b_flags & B_MALLOC) {
3036 origbuf = bp->b_data;
3037 origbufsize = bp->b_bufsize;
3038 bp->b_data = bp->b_kvabase;
3039 if (bp->b_bufsize) {
3040 bufmallocspace -= bp->b_bufsize;
3044 bp->b_flags &= ~B_MALLOC;
3045 newbsize = round_page(newbsize);
3049 (vm_offset_t) bp->b_data + bp->b_bufsize,
3050 (vm_offset_t) bp->b_data + newbsize);
3052 bcopy(origbuf, bp->b_data, origbufsize);
3053 kfree(origbuf, M_BIOBUF);
3060 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3061 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3062 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3063 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3065 if (bp->b_flags & B_MALLOC)
3066 panic("allocbuf: VMIO buffer can't be malloced");
3068 * Set B_CACHE initially if buffer is 0 length or will become
3071 if (size == 0 || bp->b_bufsize == 0)
3072 bp->b_flags |= B_CACHE;
3074 if (newbsize < bp->b_bufsize) {
3076 * DEV_BSIZE aligned new buffer size is less then the
3077 * DEV_BSIZE aligned existing buffer size. Figure out
3078 * if we have to remove any pages.
3080 if (desiredpages < bp->b_xio.xio_npages) {
3081 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3083 * the page is not freed here -- it
3084 * is the responsibility of
3085 * vnode_pager_setsize
3087 m = bp->b_xio.xio_pages[i];
3088 KASSERT(m != bogus_page,
3089 ("allocbuf: bogus page found"));
3090 while (vm_page_sleep_busy(m, TRUE, "biodep"))
3093 bp->b_xio.xio_pages[i] = NULL;
3094 vm_page_unwire(m, 0);
3096 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3097 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3098 bp->b_xio.xio_npages = desiredpages;
3100 } else if (size > bp->b_bcount) {
3102 * We are growing the buffer, possibly in a
3103 * byte-granular fashion.
3111 * Step 1, bring in the VM pages from the object,
3112 * allocating them if necessary. We must clear
3113 * B_CACHE if these pages are not valid for the
3114 * range covered by the buffer.
3116 * critical section protection is required to protect
3117 * against interrupts unbusying and freeing pages
3118 * between our vm_page_lookup() and our
3119 * busycheck/wiring call.
3125 while (bp->b_xio.xio_npages < desiredpages) {
3129 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
3130 if ((m = vm_page_lookup(obj, pi)) == NULL) {
3132 * note: must allocate system pages
3133 * since blocking here could intefere
3134 * with paging I/O, no matter which
3137 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
3141 bp->b_flags &= ~B_CACHE;
3142 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3143 ++bp->b_xio.xio_npages;
3149 * We found a page. If we have to sleep on it,
3150 * retry because it might have gotten freed out
3153 * We can only test PG_BUSY here. Blocking on
3154 * m->busy might lead to a deadlock:
3156 * vm_fault->getpages->cluster_read->allocbuf
3160 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
3162 vm_page_flag_clear(m, PG_ZERO);
3164 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3165 ++bp->b_xio.xio_npages;
3170 * Step 2. We've loaded the pages into the buffer,
3171 * we have to figure out if we can still have B_CACHE
3172 * set. Note that B_CACHE is set according to the
3173 * byte-granular range ( bcount and size ), not the
3174 * aligned range ( newbsize ).
3176 * The VM test is against m->valid, which is DEV_BSIZE
3177 * aligned. Needless to say, the validity of the data
3178 * needs to also be DEV_BSIZE aligned. Note that this
3179 * fails with NFS if the server or some other client
3180 * extends the file's EOF. If our buffer is resized,
3181 * B_CACHE may remain set! XXX
3184 toff = bp->b_bcount;
3185 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3187 while ((bp->b_flags & B_CACHE) && toff < size) {
3190 if (tinc > (size - toff))
3193 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3201 bp->b_xio.xio_pages[pi]
3208 * Step 3, fixup the KVM pmap. Remember that
3209 * bp->b_data is relative to bp->b_loffset, but
3210 * bp->b_loffset may be offset into the first page.
3213 bp->b_data = (caddr_t)
3214 trunc_page((vm_offset_t)bp->b_data);
3216 (vm_offset_t)bp->b_data,
3217 bp->b_xio.xio_pages,
3218 bp->b_xio.xio_npages
3220 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3221 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3225 /* adjust space use on already-dirty buffer */
3226 if (bp->b_flags & B_DELWRI) {
3227 dirtybufspace += newbsize - bp->b_bufsize;
3228 if (bp->b_flags & B_HEAVY)
3229 dirtybufspacehw += newbsize - bp->b_bufsize;
3231 if (newbsize < bp->b_bufsize)
3233 bp->b_bufsize = newbsize; /* actual buffer allocation */
3234 bp->b_bcount = size; /* requested buffer size */
3241 * Wait for buffer I/O completion, returning error status. B_EINTR
3242 * is converted into an EINTR error but not cleared (since a chain
3243 * of biowait() calls may occur).
3245 * On return bpdone() will have been called but the buffer will remain
3246 * locked and will not have been brelse()'d.
3248 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3249 * likely still in progress on return.
3251 * NOTE! This operation is on a BIO, not a BUF.
3253 * NOTE! BIO_DONE is cleared by vn_strategy()
3258 _biowait(struct bio *bio, const char *wmesg, int to)
3260 struct buf *bp = bio->bio_buf;
3265 KKASSERT(bio == &bp->b_bio1);
3267 flags = bio->bio_flags;
3268 if (flags & BIO_DONE)
3270 tsleep_interlock(bio, 0);
3271 nflags = flags | BIO_WANT;
3272 tsleep_interlock(bio, 0);
3273 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3275 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3276 else if (bp->b_cmd == BUF_CMD_READ)
3277 error = tsleep(bio, PINTERLOCKED, "biord", to);
3279 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3281 kprintf("tsleep error biowait %d\n", error);
3291 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3292 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3293 if (bp->b_flags & B_EINTR)
3295 if (bp->b_flags & B_ERROR)
3296 return (bp->b_error ? bp->b_error : EIO);
3301 biowait(struct bio *bio, const char *wmesg)
3303 return(_biowait(bio, wmesg, 0));
3307 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3309 return(_biowait(bio, wmesg, to));
3313 * This associates a tracking count with an I/O. vn_strategy() and
3314 * dev_dstrategy() do this automatically but there are a few cases
3315 * where a vnode or device layer is bypassed when a block translation
3316 * is cached. In such cases bio_start_transaction() may be called on
3317 * the bypassed layers so the system gets an I/O in progress indication
3318 * for those higher layers.
3321 bio_start_transaction(struct bio *bio, struct bio_track *track)
3323 bio->bio_track = track;
3324 bio_track_ref(track);
3328 * Initiate I/O on a vnode.
3331 vn_strategy(struct vnode *vp, struct bio *bio)
3333 struct bio_track *track;
3335 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
3336 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
3337 track = &vp->v_track_read;
3339 track = &vp->v_track_write;
3340 KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3341 bio->bio_track = track;
3342 bio_track_ref(track);
3343 vop_strategy(*vp->v_ops, vp, bio);
3349 * Finish I/O on a buffer after all BIOs have been processed.
3350 * Called when the bio chain is exhausted or by biowait. If called
3351 * by biowait, elseit is typically 0.
3353 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3354 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3355 * assuming B_INVAL is clear.
3357 * For the VMIO case, we set B_CACHE if the op was a read and no
3358 * read error occured, or if the op was a write. B_CACHE is never
3359 * set if the buffer is invalid or otherwise uncacheable.
3361 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3362 * initiator to leave B_INVAL set to brelse the buffer out of existance
3363 * in the biodone routine.
3366 bpdone(struct buf *bp, int elseit)
3370 KASSERT(BUF_REFCNTNB(bp) > 0,
3371 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3372 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3373 ("biodone: bp %p already done!", bp));
3376 * No more BIOs are left. All completion functions have been dealt
3377 * with, now we clean up the buffer.
3380 bp->b_cmd = BUF_CMD_DONE;
3383 * Only reads and writes are processed past this point.
3385 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3386 if (cmd == BUF_CMD_FREEBLKS)
3387 bp->b_flags |= B_NOCACHE;
3394 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3395 * a lot worse. XXX - move this above the clearing of b_cmd
3397 if (LIST_FIRST(&bp->b_dep) != NULL)
3401 * A failed write must re-dirty the buffer unless B_INVAL
3402 * was set. Only applicable to normal buffers (with VPs).
3403 * vinum buffers may not have a vp.
3405 if (cmd == BUF_CMD_WRITE &&
3406 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3407 bp->b_flags &= ~B_NOCACHE;
3412 if (bp->b_flags & B_VMIO) {
3418 struct vnode *vp = bp->b_vp;
3422 #if defined(VFS_BIO_DEBUG)
3423 if (vp->v_auxrefs == 0)
3424 panic("biodone: zero vnode hold count");
3425 if ((vp->v_flag & VOBJBUF) == 0)
3426 panic("biodone: vnode is not setup for merged cache");
3429 foff = bp->b_loffset;
3430 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3431 KASSERT(obj != NULL, ("biodone: missing VM object"));
3433 #if defined(VFS_BIO_DEBUG)
3434 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3435 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3436 obj->paging_in_progress, bp->b_xio.xio_npages);
3441 * Set B_CACHE if the op was a normal read and no error
3442 * occured. B_CACHE is set for writes in the b*write()
3445 iosize = bp->b_bcount - bp->b_resid;
3446 if (cmd == BUF_CMD_READ &&
3447 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3448 bp->b_flags |= B_CACHE;
3453 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3457 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3462 * cleanup bogus pages, restoring the originals. Since
3463 * the originals should still be wired, we don't have
3464 * to worry about interrupt/freeing races destroying
3465 * the VM object association.
3467 m = bp->b_xio.xio_pages[i];
3468 if (m == bogus_page) {
3470 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3472 panic("biodone: page disappeared");
3473 bp->b_xio.xio_pages[i] = m;
3474 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3475 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3477 #if defined(VFS_BIO_DEBUG)
3478 if (OFF_TO_IDX(foff) != m->pindex) {
3479 kprintf("biodone: foff(%lu)/m->pindex(%ld) "
3481 (unsigned long)foff, (long)m->pindex);
3486 * In the write case, the valid and clean bits are
3487 * already changed correctly (see bdwrite()), so we
3488 * only need to do this here in the read case.
3490 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3491 vfs_clean_one_page(bp, i, m);
3493 vm_page_flag_clear(m, PG_ZERO);
3496 * when debugging new filesystems or buffer I/O
3497 * methods, this is the most common error that pops
3498 * up. if you see this, you have not set the page
3499 * busy flag correctly!!!
3502 kprintf("biodone: page busy < 0, "
3503 "pindex: %d, foff: 0x(%x,%x), "
3504 "resid: %d, index: %d\n",
3505 (int) m->pindex, (int)(foff >> 32),
3506 (int) foff & 0xffffffff, resid, i);
3507 if (!vn_isdisk(vp, NULL))
3508 kprintf(" iosize: %ld, loffset: %lld, "
3509 "flags: 0x%08x, npages: %d\n",
3510 bp->b_vp->v_mount->mnt_stat.f_iosize,
3511 (long long)bp->b_loffset,
3512 bp->b_flags, bp->b_xio.xio_npages);
3514 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3515 (long long)bp->b_loffset,
3516 bp->b_flags, bp->b_xio.xio_npages);
3517 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3518 m->valid, m->dirty, m->wire_count);
3519 panic("biodone: page busy < 0");
3521 vm_page_io_finish(m);
3522 vm_object_pip_subtract(obj, 1);
3523 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3527 vm_object_pip_wakeupn(obj, 0);
3533 * Finish up by releasing the buffer. There are no more synchronous
3534 * or asynchronous completions, those were handled by bio_done
3538 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
3549 biodone(struct bio *bio)
3551 struct buf *bp = bio->bio_buf;
3553 runningbufwakeup(bp);
3556 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3559 biodone_t *done_func;
3560 struct bio_track *track;
3563 * BIO tracking. Most but not all BIOs are tracked.
3565 if ((track = bio->bio_track) != NULL) {
3566 bio_track_rel(track);
3567 bio->bio_track = NULL;
3571 * A bio_done function terminates the loop. The function
3572 * will be responsible for any further chaining and/or
3573 * buffer management.
3575 * WARNING! The done function can deallocate the buffer!
3577 if ((done_func = bio->bio_done) != NULL) {
3578 bio->bio_done = NULL;
3582 bio = bio->bio_prev;
3586 * If we've run out of bio's do normal [a]synchronous completion.
3592 * Synchronous biodone - this terminates a synchronous BIO.
3594 * bpdone() is called with elseit=FALSE, leaving the buffer completed
3595 * but still locked. The caller must brelse() the buffer after waiting
3599 biodone_sync(struct bio *bio)
3601 struct buf *bp = bio->bio_buf;
3605 KKASSERT(bio == &bp->b_bio1);
3609 flags = bio->bio_flags;
3610 nflags = (flags | BIO_DONE) & ~BIO_WANT;
3612 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3613 if (flags & BIO_WANT)
3623 * This routine is called in lieu of iodone in the case of
3624 * incomplete I/O. This keeps the busy status for pages
3628 vfs_unbusy_pages(struct buf *bp)
3632 runningbufwakeup(bp);
3633 if (bp->b_flags & B_VMIO) {
3634 struct vnode *vp = bp->b_vp;
3639 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3640 vm_page_t m = bp->b_xio.xio_pages[i];
3643 * When restoring bogus changes the original pages
3644 * should still be wired, so we are in no danger of
3645 * losing the object association and do not need
3646 * critical section protection particularly.
3648 if (m == bogus_page) {
3649 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3651 panic("vfs_unbusy_pages: page missing");
3653 bp->b_xio.xio_pages[i] = m;
3654 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3655 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3657 vm_object_pip_subtract(obj, 1);
3658 vm_page_flag_clear(m, PG_ZERO);
3659 vm_page_io_finish(m);
3661 vm_object_pip_wakeupn(obj, 0);
3668 * This routine is called before a device strategy routine.
3669 * It is used to tell the VM system that paging I/O is in
3670 * progress, and treat the pages associated with the buffer
3671 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3672 * flag is handled to make sure that the object doesn't become
3675 * Since I/O has not been initiated yet, certain buffer flags
3676 * such as B_ERROR or B_INVAL may be in an inconsistant state
3677 * and should be ignored.
3680 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3683 struct lwp *lp = curthread->td_lwp;
3686 * The buffer's I/O command must already be set. If reading,
3687 * B_CACHE must be 0 (double check against callers only doing
3688 * I/O when B_CACHE is 0).
3690 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3691 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3693 if (bp->b_flags & B_VMIO) {
3697 KASSERT(bp->b_loffset != NOOFFSET,
3698 ("vfs_busy_pages: no buffer offset"));
3701 * Loop until none of the pages are busy.
3704 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3705 vm_page_t m = bp->b_xio.xio_pages[i];
3707 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3712 * Setup for I/O, soft-busy the page right now because
3713 * the next loop may block.
3715 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3716 vm_page_t m = bp->b_xio.xio_pages[i];
3718 vm_page_flag_clear(m, PG_ZERO);
3719 if ((bp->b_flags & B_CLUSTER) == 0) {
3720 vm_object_pip_add(obj, 1);
3721 vm_page_io_start(m);
3726 * Adjust protections for I/O and do bogus-page mapping.
3727 * Assume that vm_page_protect() can block (it can block
3728 * if VM_PROT_NONE, don't take any chances regardless).
3730 * In particularly note that for writes we must incorporate
3731 * page dirtyness from the VM system into the buffer's
3734 * For reads we theoretically must incorporate page dirtyness
3735 * from the VM system to determine if the page needs bogus
3736 * replacement, but we shortcut the test by simply checking
3737 * that all m->valid bits are set, indicating that the page
3738 * is fully valid and does not need to be re-read. For any
3739 * VM system dirtyness the page will also be fully valid
3740 * since it was mapped at one point.
3743 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3744 vm_page_t m = bp->b_xio.xio_pages[i];
3746 vm_page_flag_clear(m, PG_ZERO); /* XXX */
3747 if (bp->b_cmd == BUF_CMD_WRITE) {
3749 * When readying a vnode-backed buffer for
3750 * a write we must zero-fill any invalid
3751 * portions of the backing VM pages, mark
3752 * it valid and clear related dirty bits.
3754 * vfs_clean_one_page() incorporates any
3755 * VM dirtyness and updates the b_dirtyoff
3756 * range (after we've made the page RO).
3758 * It is also expected that the pmap modified
3759 * bit has already been cleared by the
3760 * vm_page_protect(). We may not be able
3761 * to clear all dirty bits for a page if it
3762 * was also memory mapped (NFS).
3764 vm_page_protect(m, VM_PROT_READ);
3765 vfs_clean_one_page(bp, i, m);
3766 } else if (m->valid == VM_PAGE_BITS_ALL) {
3768 * When readying a vnode-backed buffer for
3769 * read we must replace any dirty pages with
3770 * a bogus page so dirty data is not destroyed
3771 * when filling gaps.
3773 * To avoid testing whether the page is
3774 * dirty we instead test that the page was
3775 * at some point mapped (m->valid fully
3776 * valid) with the understanding that
3777 * this also covers the dirty case.
3779 bp->b_xio.xio_pages[i] = bogus_page;
3781 } else if (m->valid & m->dirty) {
3783 * This case should not occur as partial
3784 * dirtyment can only happen if the buffer
3785 * is B_CACHE, and this code is not entered
3786 * if the buffer is B_CACHE.
3788 kprintf("Warning: vfs_busy_pages - page not "
3789 "fully valid! loff=%jx bpf=%08x "
3790 "idx=%d val=%02x dir=%02x\n",
3791 (intmax_t)bp->b_loffset, bp->b_flags,
3792 i, m->valid, m->dirty);
3793 vm_page_protect(m, VM_PROT_NONE);
3796 * The page is not valid and can be made
3799 vm_page_protect(m, VM_PROT_NONE);
3803 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3804 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3809 * This is the easiest place to put the process accounting for the I/O
3813 if (bp->b_cmd == BUF_CMD_READ)
3814 lp->lwp_ru.ru_inblock++;
3816 lp->lwp_ru.ru_oublock++;
3823 * Tell the VM system that the pages associated with this buffer
3824 * are clean. This is used for delayed writes where the data is
3825 * going to go to disk eventually without additional VM intevention.
3827 * Note that while we only really need to clean through to b_bcount, we
3828 * just go ahead and clean through to b_bufsize.
3831 vfs_clean_pages(struct buf *bp)
3836 if ((bp->b_flags & B_VMIO) == 0)
3839 KASSERT(bp->b_loffset != NOOFFSET,
3840 ("vfs_clean_pages: no buffer offset"));
3842 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3843 m = bp->b_xio.xio_pages[i];
3844 vfs_clean_one_page(bp, i, m);
3849 * vfs_clean_one_page:
3851 * Set the valid bits and clear the dirty bits in a page within a
3852 * buffer. The range is restricted to the buffer's size and the
3853 * buffer's logical offset might index into the first page.
3855 * The caller has busied or soft-busied the page and it is not mapped,
3856 * test and incorporate the dirty bits into b_dirtyoff/end before
3857 * clearing them. Note that we need to clear the pmap modified bits
3858 * after determining the the page was dirty, vm_page_set_validclean()
3859 * does not do it for us.
3861 * This routine is typically called after a read completes (dirty should
3862 * be zero in that case as we are not called on bogus-replace pages),
3863 * or before a write is initiated.
3866 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m)
3874 * Calculate offset range within the page but relative to buffer's
3875 * loffset. loffset might be offset into the first page.
3877 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
3878 bcount = bp->b_bcount + xoff; /* offset adjusted */
3884 soff = (pageno << PAGE_SHIFT);
3885 eoff = soff + PAGE_SIZE;
3893 * Test dirty bits and adjust b_dirtyoff/end.
3895 * If dirty pages are incorporated into the bp any prior
3896 * B_NEEDCOMMIT state (NFS) must be cleared because the
3897 * caller has not taken into account the new dirty data.
3899 * If the page was memory mapped the dirty bits might go beyond the
3900 * end of the buffer, but we can't really make the assumption that
3901 * a file EOF straddles the buffer (even though this is the case for
3902 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
3903 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
3904 * This also saves some console spam.
3906 vm_page_test_dirty(m);
3908 pmap_clear_modify(m);
3909 if ((bp->b_flags & B_NEEDCOMMIT) &&
3910 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) {
3911 kprintf("Warning: vfs_clean_one_page: bp %p "
3912 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT\n",
3913 bp, (intmax_t)bp->b_loffset, bp->b_bcount,
3915 bp->b_flags &= ~B_NEEDCOMMIT;
3917 if (bp->b_dirtyoff > soff - xoff)
3918 bp->b_dirtyoff = soff - xoff;
3919 if (bp->b_dirtyend < eoff - xoff)
3920 bp->b_dirtyend = eoff - xoff;
3924 * Set related valid bits, clear related dirty bits.
3925 * Does not mess with the pmap modified bit.
3927 * WARNING! We cannot just clear all of m->dirty here as the
3928 * buffer cache buffers may use a DEV_BSIZE'd aligned
3929 * block size, or have an odd size (e.g. NFS at file EOF).
3930 * The putpages code can clear m->dirty to 0.
3932 * If a VOP_WRITE generates a buffer cache buffer which
3933 * covers the same space as mapped writable pages the
3934 * buffer flush might not be able to clear all the dirty
3935 * bits and still require a putpages from the VM system
3938 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff);
3944 * Clear a buffer. This routine essentially fakes an I/O, so we need
3945 * to clear B_ERROR and B_INVAL.
3947 * Note that while we only theoretically need to clear through b_bcount,
3948 * we go ahead and clear through b_bufsize.
3952 vfs_bio_clrbuf(struct buf *bp)
3956 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3957 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
3958 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3959 (bp->b_loffset & PAGE_MASK) == 0) {
3960 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3961 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3965 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3966 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3967 bzero(bp->b_data, bp->b_bufsize);
3968 bp->b_xio.xio_pages[0]->valid |= mask;
3974 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3975 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3976 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3977 ea = (caddr_t)(vm_offset_t)ulmin(
3978 (u_long)(vm_offset_t)ea,
3979 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3980 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3981 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3983 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3984 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3988 for (; sa < ea; sa += DEV_BSIZE, j++) {
3989 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3990 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3991 bzero(sa, DEV_BSIZE);
3994 bp->b_xio.xio_pages[i]->valid |= mask;
3995 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
4004 * vm_hold_load_pages:
4006 * Load pages into the buffer's address space. The pages are
4007 * allocated from the kernel object in order to reduce interference
4008 * with the any VM paging I/O activity. The range of loaded
4009 * pages will be wired.
4011 * If a page cannot be allocated, the 'pagedaemon' is woken up to
4012 * retrieve the full range (to - from) of pages.
4016 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4022 to = round_page(to);
4023 from = round_page(from);
4024 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4029 * Note: must allocate system pages since blocking here
4030 * could intefere with paging I/O, no matter which
4033 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT,
4034 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
4037 p->valid = VM_PAGE_BITS_ALL;
4038 vm_page_flag_clear(p, PG_ZERO);
4039 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
4040 bp->b_xio.xio_pages[index] = p;
4047 bp->b_xio.xio_npages = index;
4051 * Allocate pages for a buffer cache buffer.
4053 * Under extremely severe memory conditions even allocating out of the
4054 * system reserve can fail. If this occurs we must allocate out of the
4055 * interrupt reserve to avoid a deadlock with the pageout daemon.
4057 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
4058 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
4059 * against the pageout daemon if pages are not freed from other sources.
4063 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit)
4068 * Try a normal allocation, allow use of system reserve.
4070 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
4075 * The normal allocation failed and we clearly have a page
4076 * deficit. Try to reclaim some clean VM pages directly
4077 * from the buffer cache.
4079 vm_pageout_deficit += deficit;
4083 * We may have blocked, the caller will know what to do if the
4086 if (vm_page_lookup(obj, pg))
4090 * Allocate and allow use of the interrupt reserve.
4092 * If after all that we still can't allocate a VM page we are
4093 * in real trouble, but we slog on anyway hoping that the system
4096 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
4097 VM_ALLOC_INTERRUPT);
4099 if (vm_page_count_severe()) {
4100 kprintf("bio_page_alloc: WARNING emergency page "
4105 kprintf("bio_page_alloc: WARNING emergency page "
4106 "allocation failed\n");
4113 * vm_hold_free_pages:
4115 * Return pages associated with the buffer back to the VM system.
4117 * The range of pages underlying the buffer's address space will
4118 * be unmapped and un-wired.
4121 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4125 int index, newnpages;
4127 from = round_page(from);
4128 to = round_page(to);
4129 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4131 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4132 p = bp->b_xio.xio_pages[index];
4133 if (p && (index < bp->b_xio.xio_npages)) {
4135 kprintf("vm_hold_free_pages: doffset: %lld, "
4137 (long long)bp->b_bio2.bio_offset,
4138 (long long)bp->b_loffset);
4140 bp->b_xio.xio_pages[index] = NULL;
4143 vm_page_unwire(p, 0);
4147 bp->b_xio.xio_npages = newnpages;
4153 * Map a user buffer into KVM via a pbuf. On return the buffer's
4154 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4158 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4169 * bp had better have a command and it better be a pbuf.
4171 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4172 KKASSERT(bp->b_flags & B_PAGING);
4178 * Map the user data into KVM. Mappings have to be page-aligned.
4180 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4183 vmprot = VM_PROT_READ;
4184 if (bp->b_cmd == BUF_CMD_READ)
4185 vmprot |= VM_PROT_WRITE;
4187 while (addr < udata + bytes) {
4189 * Do the vm_fault if needed; do the copy-on-write thing
4190 * when reading stuff off device into memory.
4192 * vm_fault_page*() returns a held VM page.
4194 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4195 va = trunc_page(va);
4197 m = vm_fault_page_quick(va, vmprot, &error);
4199 for (i = 0; i < pidx; ++i) {
4200 vm_page_unhold(bp->b_xio.xio_pages[i]);
4201 bp->b_xio.xio_pages[i] = NULL;
4205 bp->b_xio.xio_pages[pidx] = m;
4211 * Map the page array and set the buffer fields to point to
4212 * the mapped data buffer.
4214 if (pidx > btoc(MAXPHYS))
4215 panic("vmapbuf: mapped more than MAXPHYS");
4216 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4218 bp->b_xio.xio_npages = pidx;
4219 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4220 bp->b_bcount = bytes;
4221 bp->b_bufsize = bytes;
4228 * Free the io map PTEs associated with this IO operation.
4229 * We also invalidate the TLB entries and restore the original b_addr.
4232 vunmapbuf(struct buf *bp)
4237 KKASSERT(bp->b_flags & B_PAGING);
4239 npages = bp->b_xio.xio_npages;
4240 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4241 for (pidx = 0; pidx < npages; ++pidx) {
4242 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4243 bp->b_xio.xio_pages[pidx] = NULL;
4245 bp->b_xio.xio_npages = 0;
4246 bp->b_data = bp->b_kvabase;
4250 * Scan all buffers in the system and issue the callback.
4253 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4259 for (n = 0; n < nbuf; ++n) {
4260 if ((error = callback(&buf[n], info)) < 0) {
4270 * print out statistics from the current status of the buffer pool
4271 * this can be toggeled by the system control option debug.syncprt
4280 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
4281 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
4283 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
4285 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4288 TAILQ_FOREACH(bp, dp, b_freelist) {
4289 counts[bp->b_bufsize/PAGE_SIZE]++;
4293 kprintf("%s: total-%d", bname[i], count);
4294 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4296 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
4304 DB_SHOW_COMMAND(buffer, db_show_buffer)
4307 struct buf *bp = (struct buf *)addr;
4310 db_printf("usage: show buffer <addr>\n");
4314 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4315 db_printf("b_cmd = %d\n", bp->b_cmd);
4316 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4317 "b_resid = %d\n, b_data = %p, "
4318 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4319 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4321 (long long)bp->b_bio2.bio_offset,
4322 (long long)(bp->b_bio2.bio_next ?
4323 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4324 if (bp->b_xio.xio_npages) {
4326 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4327 bp->b_xio.xio_npages);
4328 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4330 m = bp->b_xio.xio_pages[i];
4331 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4332 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4333 if ((i + 1) < bp->b_xio.xio_npages)