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>
47 #include <sys/dsched.h>
50 #include <vm/vm_param.h>
51 #include <vm/vm_kern.h>
52 #include <vm/vm_pageout.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_map.h>
57 #include <vm/vm_pager.h>
58 #include <vm/swap_pager.h>
61 #include <sys/thread2.h>
62 #include <sys/spinlock2.h>
63 #include <sys/mplock2.h>
64 #include <vm/vm_page2.h>
75 BQUEUE_NONE, /* not on any queue */
76 BQUEUE_LOCKED, /* locked buffers */
77 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
78 BQUEUE_DIRTY, /* B_DELWRI buffers */
79 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
80 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
81 BQUEUE_EMPTY, /* empty buffer headers */
83 BUFFER_QUEUES /* number of buffer queues */
86 typedef enum bufq_type bufq_type_t;
88 #define BD_WAKE_SIZE 16384
89 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
91 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
92 struct spinlock bufspin = SPINLOCK_INITIALIZER(&bufspin);
94 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
96 struct buf *buf; /* buffer header pool */
98 static void vfs_clean_pages(struct buf *bp);
99 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
100 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m);
101 static void vfs_vmio_release(struct buf *bp);
102 static int flushbufqueues(bufq_type_t q);
103 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);
105 static void bd_signal(int totalspace);
106 static void buf_daemon(void);
107 static void buf_daemon_hw(void);
110 * bogus page -- for I/O to/from partially complete buffers
111 * this is a temporary solution to the problem, but it is not
112 * really that bad. it would be better to split the buffer
113 * for input in the case of buffers partially already in memory,
114 * but the code is intricate enough already.
116 vm_page_t bogus_page;
119 * These are all static, but make the ones we export globals so we do
120 * not need to use compiler magic.
122 int bufspace, maxbufspace,
123 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
124 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
125 static int lorunningspace, hirunningspace, runningbufreq;
126 int dirtybufspace, dirtybufspacehw, lodirtybufspace, hidirtybufspace;
127 int dirtybufcount, dirtybufcounthw;
128 int runningbufspace, runningbufcount;
129 static int getnewbufcalls;
130 static int getnewbufrestarts;
131 static int recoverbufcalls;
132 static int needsbuffer; /* locked by needsbuffer_spin */
133 static int bd_request; /* locked by needsbuffer_spin */
134 static int bd_request_hw; /* locked by needsbuffer_spin */
135 static u_int bd_wake_ary[BD_WAKE_SIZE];
136 static u_int bd_wake_index;
137 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */
138 static struct spinlock needsbuffer_spin;
139 static int debug_commit;
141 static struct thread *bufdaemon_td;
142 static struct thread *bufdaemonhw_td;
146 * Sysctls for operational control of the buffer cache.
148 SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
149 "Number of dirty buffers to flush before bufdaemon becomes inactive");
150 SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
151 "High watermark used to trigger explicit flushing of dirty buffers");
152 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
153 "Minimum amount of buffer space required for active I/O");
154 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
155 "Maximum amount of buffer space to usable for active I/O");
156 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
157 "Recycle pages to active or inactive queue transition pt 0-64");
159 * Sysctls determining current state of the buffer cache.
161 SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
162 "Total number of buffers in buffer cache");
163 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
164 "Pending bytes of dirty buffers (all)");
165 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
166 "Pending bytes of dirty buffers (heavy weight)");
167 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
168 "Pending number of dirty buffers");
169 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
170 "Pending number of dirty buffers (heavy weight)");
171 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
172 "I/O bytes currently in progress due to asynchronous writes");
173 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
174 "I/O buffers currently in progress due to asynchronous writes");
175 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
176 "Hard limit on maximum amount of memory usable for buffer space");
177 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
178 "Soft limit on maximum amount of memory usable for buffer space");
179 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
180 "Minimum amount of memory to reserve for system buffer space");
181 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
182 "Amount of memory available for buffers");
183 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
184 0, "Maximum amount of memory reserved for buffers using malloc");
185 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
186 "Amount of memory left for buffers using malloc-scheme");
187 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
188 "New buffer header acquisition requests");
189 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
190 0, "New buffer header acquisition restarts");
191 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
192 "Recover VM space in an emergency");
193 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
194 "Buffer acquisition restarts due to fragmented buffer map");
195 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
196 "Amount of time KVA space was deallocated in an arbitrary buffer");
197 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
198 "Amount of time buffer re-use operations were successful");
199 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, "");
200 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
201 "sizeof(struct buf)");
203 char *buf_wmesg = BUF_WMESG;
205 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
206 #define VFS_BIO_NEED_UNUSED02 0x02
207 #define VFS_BIO_NEED_UNUSED04 0x04
208 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
213 * Called when buffer space is potentially available for recovery.
214 * getnewbuf() will block on this flag when it is unable to free
215 * sufficient buffer space. Buffer space becomes recoverable when
216 * bp's get placed back in the queues.
223 * If someone is waiting for BUF space, wake them up. Even
224 * though we haven't freed the kva space yet, the waiting
225 * process will be able to now.
227 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
228 spin_lock_wr(&needsbuffer_spin);
229 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
230 spin_unlock_wr(&needsbuffer_spin);
231 wakeup(&needsbuffer);
238 * Accounting for I/O in progress.
242 runningbufwakeup(struct buf *bp)
247 if ((totalspace = bp->b_runningbufspace) != 0) {
248 atomic_subtract_int(&runningbufspace, totalspace);
249 atomic_subtract_int(&runningbufcount, 1);
250 bp->b_runningbufspace = 0;
253 * see waitrunningbufspace() for limit test.
255 limit = hirunningspace * 2 / 3;
256 if (runningbufreq && runningbufspace <= limit) {
258 wakeup(&runningbufreq);
260 bd_signal(totalspace);
267 * Called when a buffer has been added to one of the free queues to
268 * account for the buffer and to wakeup anyone waiting for free buffers.
269 * This typically occurs when large amounts of metadata are being handled
270 * by the buffer cache ( else buffer space runs out first, usually ).
278 spin_lock_wr(&needsbuffer_spin);
279 needsbuffer &= ~VFS_BIO_NEED_ANY;
280 spin_unlock_wr(&needsbuffer_spin);
281 wakeup(&needsbuffer);
286 * waitrunningbufspace()
288 * Wait for the amount of running I/O to drop to hirunningspace * 2 / 3.
289 * This is the point where write bursting stops so we don't want to wait
290 * for the running amount to drop below it (at least if we still want bioq
293 * The caller may be using this function to block in a tight loop, we
294 * must block while runningbufspace is greater then or equal to
295 * hirunningspace * 2 / 3.
297 * And even with that it may not be enough, due to the presence of
298 * B_LOCKED dirty buffers, so also wait for at least one running buffer
302 waitrunningbufspace(void)
304 int limit = hirunningspace * 2 / 3;
307 if (runningbufspace > limit) {
308 while (runningbufspace > limit) {
310 tsleep(&runningbufreq, 0, "wdrn1", 0);
312 } else if (runningbufspace) {
314 tsleep(&runningbufreq, 0, "wdrn2", 1);
320 * buf_dirty_count_severe:
322 * Return true if we have too many dirty buffers.
325 buf_dirty_count_severe(void)
327 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
328 dirtybufcount >= nbuf / 2);
332 * Return true if the amount of running I/O is severe and BIOQ should
336 buf_runningbufspace_severe(void)
338 return (runningbufspace >= hirunningspace * 2 / 3);
342 * vfs_buf_test_cache:
344 * Called when a buffer is extended. This function clears the B_CACHE
345 * bit if the newly extended portion of the buffer does not contain
348 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
349 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
350 * them while a clean buffer was present.
354 vfs_buf_test_cache(struct buf *bp,
355 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
358 if (bp->b_flags & B_CACHE) {
359 int base = (foff + off) & PAGE_MASK;
360 if (vm_page_is_valid(m, base, size) == 0)
361 bp->b_flags &= ~B_CACHE;
368 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
377 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
380 if (bd_request == 0 &&
381 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
382 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
383 spin_lock_wr(&needsbuffer_spin);
385 spin_unlock_wr(&needsbuffer_spin);
388 if (bd_request_hw == 0 &&
389 (dirtybufspacehw > lodirtybufspace / 2 ||
390 dirtybufcounthw >= nbuf / 2)) {
391 spin_lock_wr(&needsbuffer_spin);
393 spin_unlock_wr(&needsbuffer_spin);
394 wakeup(&bd_request_hw);
401 * Get the buf_daemon heated up when the number of running and dirty
402 * buffers exceeds the mid-point.
404 * Return the total number of dirty bytes past the second mid point
405 * as a measure of how much excess dirty data there is in the system.
416 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
418 totalspace = runningbufspace + dirtybufspace;
419 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
421 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
422 if (totalspace >= mid2)
423 return(totalspace - mid2);
431 * Wait for the buffer cache to flush (totalspace) bytes worth of
432 * buffers, then return.
434 * Regardless this function blocks while the number of dirty buffers
435 * exceeds hidirtybufspace.
440 bd_wait(int totalspace)
445 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
448 while (totalspace > 0) {
450 if (totalspace > runningbufspace + dirtybufspace)
451 totalspace = runningbufspace + dirtybufspace;
452 count = totalspace / BKVASIZE;
453 if (count >= BD_WAKE_SIZE)
454 count = BD_WAKE_SIZE - 1;
456 spin_lock_wr(&needsbuffer_spin);
457 i = (bd_wake_index + count) & BD_WAKE_MASK;
459 tsleep_interlock(&bd_wake_ary[i], 0);
460 spin_unlock_wr(&needsbuffer_spin);
461 tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz);
463 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
470 * This function is called whenever runningbufspace or dirtybufspace
471 * is reduced. Track threads waiting for run+dirty buffer I/O
477 bd_signal(int totalspace)
481 if (totalspace > 0) {
482 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
483 totalspace = BKVASIZE * BD_WAKE_SIZE;
484 spin_lock_wr(&needsbuffer_spin);
485 while (totalspace > 0) {
488 if (bd_wake_ary[i]) {
490 spin_unlock_wr(&needsbuffer_spin);
491 wakeup(&bd_wake_ary[i]);
492 spin_lock_wr(&needsbuffer_spin);
494 totalspace -= BKVASIZE;
496 spin_unlock_wr(&needsbuffer_spin);
501 * BIO tracking support routines.
503 * Release a ref on a bio_track. Wakeup requests are atomically released
504 * along with the last reference so bk_active will never wind up set to
511 bio_track_rel(struct bio_track *track)
519 active = track->bk_active;
520 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
524 * Full-on. Note that the wait flag is only atomically released on
525 * the 1->0 count transition.
527 * We check for a negative count transition using bit 30 since bit 31
528 * has a different meaning.
531 desired = (active & 0x7FFFFFFF) - 1;
533 desired |= active & 0x80000000;
534 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
535 if (desired & 0x40000000)
536 panic("bio_track_rel: bad count: %p\n", track);
537 if (active & 0x80000000)
541 active = track->bk_active;
546 * Wait for the tracking count to reach 0.
548 * Use atomic ops such that the wait flag is only set atomically when
549 * bk_active is non-zero.
554 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
563 if (track->bk_active == 0)
567 * Full-on. Note that the wait flag may only be atomically set if
568 * the active count is non-zero.
571 while ((active = track->bk_active) != 0) {
572 desired = active | 0x80000000;
573 tsleep_interlock(track, slp_flags);
574 if (active == desired ||
575 atomic_cmpset_int(&track->bk_active, active, desired)) {
576 error = tsleep(track, slp_flags | PINTERLOCKED,
588 * Load time initialisation of the buffer cache, called from machine
589 * dependant initialization code.
595 vm_offset_t bogus_offset;
598 spin_init(&needsbuffer_spin);
600 /* next, make a null set of free lists */
601 for (i = 0; i < BUFFER_QUEUES; i++)
602 TAILQ_INIT(&bufqueues[i]);
604 /* finally, initialize each buffer header and stick on empty q */
605 for (i = 0; i < nbuf; i++) {
607 bzero(bp, sizeof *bp);
608 bp->b_flags = B_INVAL; /* we're just an empty header */
609 bp->b_cmd = BUF_CMD_DONE;
610 bp->b_qindex = BQUEUE_EMPTY;
612 xio_init(&bp->b_xio);
615 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
619 * maxbufspace is the absolute maximum amount of buffer space we are
620 * allowed to reserve in KVM and in real terms. The absolute maximum
621 * is nominally used by buf_daemon. hibufspace is the nominal maximum
622 * used by most other processes. The differential is required to
623 * ensure that buf_daemon is able to run when other processes might
624 * be blocked waiting for buffer space.
626 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
627 * this may result in KVM fragmentation which is not handled optimally
630 maxbufspace = nbuf * BKVASIZE;
631 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
632 lobufspace = hibufspace - MAXBSIZE;
634 lorunningspace = 512 * 1024;
635 /* hirunningspace -- see below */
638 * Limit the amount of malloc memory since it is wired permanently
639 * into the kernel space. Even though this is accounted for in
640 * the buffer allocation, we don't want the malloced region to grow
641 * uncontrolled. The malloc scheme improves memory utilization
642 * significantly on average (small) directories.
644 maxbufmallocspace = hibufspace / 20;
647 * Reduce the chance of a deadlock occuring by limiting the number
648 * of delayed-write dirty buffers we allow to stack up.
650 * We don't want too much actually queued to the device at once
651 * (XXX this needs to be per-mount!), because the buffers will
652 * wind up locked for a very long period of time while the I/O
655 hidirtybufspace = hibufspace / 2; /* dirty + running */
656 hirunningspace = hibufspace / 16; /* locked & queued to device */
657 if (hirunningspace < 1024 * 1024)
658 hirunningspace = 1024 * 1024;
663 lodirtybufspace = hidirtybufspace / 2;
666 * Maximum number of async ops initiated per buf_daemon loop. This is
667 * somewhat of a hack at the moment, we really need to limit ourselves
668 * based on the number of bytes of I/O in-transit that were initiated
672 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
673 bogus_page = vm_page_alloc(&kernel_object,
674 (bogus_offset >> PAGE_SHIFT),
676 vmstats.v_wire_count++;
681 * Initialize the embedded bio structures
684 initbufbio(struct buf *bp)
686 bp->b_bio1.bio_buf = bp;
687 bp->b_bio1.bio_prev = NULL;
688 bp->b_bio1.bio_offset = NOOFFSET;
689 bp->b_bio1.bio_next = &bp->b_bio2;
690 bp->b_bio1.bio_done = NULL;
691 bp->b_bio1.bio_flags = 0;
693 bp->b_bio2.bio_buf = bp;
694 bp->b_bio2.bio_prev = &bp->b_bio1;
695 bp->b_bio2.bio_offset = NOOFFSET;
696 bp->b_bio2.bio_next = NULL;
697 bp->b_bio2.bio_done = NULL;
698 bp->b_bio2.bio_flags = 0;
702 * Reinitialize the embedded bio structures as well as any additional
703 * translation cache layers.
706 reinitbufbio(struct buf *bp)
710 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
711 bio->bio_done = NULL;
712 bio->bio_offset = NOOFFSET;
717 * Push another BIO layer onto an existing BIO and return it. The new
718 * BIO layer may already exist, holding cached translation data.
721 push_bio(struct bio *bio)
725 if ((nbio = bio->bio_next) == NULL) {
726 int index = bio - &bio->bio_buf->b_bio_array[0];
727 if (index >= NBUF_BIO - 1) {
728 panic("push_bio: too many layers bp %p\n",
731 nbio = &bio->bio_buf->b_bio_array[index + 1];
732 bio->bio_next = nbio;
733 nbio->bio_prev = bio;
734 nbio->bio_buf = bio->bio_buf;
735 nbio->bio_offset = NOOFFSET;
736 nbio->bio_done = NULL;
737 nbio->bio_next = NULL;
739 KKASSERT(nbio->bio_done == NULL);
744 * Pop a BIO translation layer, returning the previous layer. The
745 * must have been previously pushed.
748 pop_bio(struct bio *bio)
750 return(bio->bio_prev);
754 clearbiocache(struct bio *bio)
757 bio->bio_offset = NOOFFSET;
765 * Free the KVA allocation for buffer 'bp'.
767 * Must be called from a critical section as this is the only locking for
770 * Since this call frees up buffer space, we call bufspacewakeup().
775 bfreekva(struct buf *bp)
782 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
783 vm_map_lock(&buffer_map);
784 bufspace -= bp->b_kvasize;
785 vm_map_delete(&buffer_map,
786 (vm_offset_t) bp->b_kvabase,
787 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
790 vm_map_unlock(&buffer_map);
791 vm_map_entry_release(count);
793 bp->b_kvabase = NULL;
802 * Remove the buffer from the appropriate free list.
805 _bremfree(struct buf *bp)
807 if (bp->b_qindex != BQUEUE_NONE) {
808 KASSERT(BUF_REFCNTNB(bp) == 1,
809 ("bremfree: bp %p not locked",bp));
810 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
811 bp->b_qindex = BQUEUE_NONE;
813 if (BUF_REFCNTNB(bp) <= 1)
814 panic("bremfree: removing a buffer not on a queue");
819 bremfree(struct buf *bp)
821 spin_lock_wr(&bufspin);
823 spin_unlock_wr(&bufspin);
827 bremfree_locked(struct buf *bp)
835 * Get a buffer with the specified data. Look in the cache first. We
836 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
837 * is set, the buffer is valid and we do not have to do anything ( see
843 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
847 bp = getblk(vp, loffset, size, 0, 0);
850 /* if not found in cache, do some I/O */
851 if ((bp->b_flags & B_CACHE) == 0) {
853 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
854 bp->b_cmd = BUF_CMD_READ;
855 bp->b_bio1.bio_done = biodone_sync;
856 bp->b_bio1.bio_flags |= BIO_SYNC;
857 vfs_busy_pages(vp, bp);
858 vn_strategy(vp, &bp->b_bio1);
860 return (biowait(&bp->b_bio1, "biord"));
868 * Operates like bread, but also starts asynchronous I/O on
869 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
870 * to initiating I/O . If B_CACHE is set, the buffer is valid
871 * and we do not have to do anything.
876 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
877 int *rabsize, int cnt, struct buf **bpp)
879 struct buf *bp, *rabp;
881 int rv = 0, readwait = 0;
883 *bpp = bp = getblk(vp, loffset, size, 0, 0);
885 /* if not found in cache, do some I/O */
886 if ((bp->b_flags & B_CACHE) == 0) {
888 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
889 bp->b_cmd = BUF_CMD_READ;
890 bp->b_bio1.bio_done = biodone_sync;
891 bp->b_bio1.bio_flags |= BIO_SYNC;
892 vfs_busy_pages(vp, bp);
893 vn_strategy(vp, &bp->b_bio1);
898 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
899 if (inmem(vp, *raoffset))
901 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
903 if ((rabp->b_flags & B_CACHE) == 0) {
905 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
906 rabp->b_cmd = BUF_CMD_READ;
907 vfs_busy_pages(vp, rabp);
909 vn_strategy(vp, &rabp->b_bio1);
916 rv = biowait(&bp->b_bio1, "biord");
923 * Synchronous write, waits for completion.
925 * Write, release buffer on completion. (Done by iodone
926 * if async). Do not bother writing anything if the buffer
929 * Note that we set B_CACHE here, indicating that buffer is
930 * fully valid and thus cacheable. This is true even of NFS
931 * now so we set it generally. This could be set either here
932 * or in biodone() since the I/O is synchronous. We put it
936 bwrite(struct buf *bp)
940 if (bp->b_flags & B_INVAL) {
944 if (BUF_REFCNTNB(bp) == 0)
945 panic("bwrite: buffer is not busy???");
947 /* Mark the buffer clean */
950 bp->b_flags &= ~(B_ERROR | B_EINTR);
951 bp->b_flags |= B_CACHE;
952 bp->b_cmd = BUF_CMD_WRITE;
953 bp->b_bio1.bio_done = biodone_sync;
954 bp->b_bio1.bio_flags |= BIO_SYNC;
955 vfs_busy_pages(bp->b_vp, bp);
958 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
959 * valid for vnode-backed buffers.
961 bp->b_runningbufspace = bp->b_bufsize;
962 if (bp->b_runningbufspace) {
963 runningbufspace += bp->b_runningbufspace;
967 vn_strategy(bp->b_vp, &bp->b_bio1);
968 error = biowait(&bp->b_bio1, "biows");
976 * Asynchronous write. Start output on a buffer, but do not wait for
977 * it to complete. The buffer is released when the output completes.
979 * bwrite() ( or the VOP routine anyway ) is responsible for handling
980 * B_INVAL buffers. Not us.
983 bawrite(struct buf *bp)
985 if (bp->b_flags & B_INVAL) {
989 if (BUF_REFCNTNB(bp) == 0)
990 panic("bwrite: buffer is not busy???");
992 /* Mark the buffer clean */
995 bp->b_flags &= ~(B_ERROR | B_EINTR);
996 bp->b_flags |= B_CACHE;
997 bp->b_cmd = BUF_CMD_WRITE;
998 KKASSERT(bp->b_bio1.bio_done == NULL);
999 vfs_busy_pages(bp->b_vp, bp);
1002 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1003 * valid for vnode-backed buffers.
1005 bp->b_runningbufspace = bp->b_bufsize;
1006 if (bp->b_runningbufspace) {
1007 runningbufspace += bp->b_runningbufspace;
1012 vn_strategy(bp->b_vp, &bp->b_bio1);
1018 * Ordered write. Start output on a buffer, and flag it so that the
1019 * device will write it in the order it was queued. The buffer is
1020 * released when the output completes. bwrite() ( or the VOP routine
1021 * anyway ) is responsible for handling B_INVAL buffers.
1024 bowrite(struct buf *bp)
1026 bp->b_flags |= B_ORDERED;
1034 * Delayed write. (Buffer is marked dirty). Do not bother writing
1035 * anything if the buffer is marked invalid.
1037 * Note that since the buffer must be completely valid, we can safely
1038 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1039 * biodone() in order to prevent getblk from writing the buffer
1040 * out synchronously.
1043 bdwrite(struct buf *bp)
1045 if (BUF_REFCNTNB(bp) == 0)
1046 panic("bdwrite: buffer is not busy");
1048 if (bp->b_flags & B_INVAL) {
1054 if (dsched_is_clear_buf_priv(bp))
1058 * Set B_CACHE, indicating that the buffer is fully valid. This is
1059 * true even of NFS now.
1061 bp->b_flags |= B_CACHE;
1064 * This bmap keeps the system from needing to do the bmap later,
1065 * perhaps when the system is attempting to do a sync. Since it
1066 * is likely that the indirect block -- or whatever other datastructure
1067 * that the filesystem needs is still in memory now, it is a good
1068 * thing to do this. Note also, that if the pageout daemon is
1069 * requesting a sync -- there might not be enough memory to do
1070 * the bmap then... So, this is important to do.
1072 if (bp->b_bio2.bio_offset == NOOFFSET) {
1073 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1074 NULL, NULL, BUF_CMD_WRITE);
1078 * Because the underlying pages may still be mapped and
1079 * writable trying to set the dirty buffer (b_dirtyoff/end)
1080 * range here will be inaccurate.
1082 * However, we must still clean the pages to satisfy the
1083 * vnode_pager and pageout daemon, so theythink the pages
1084 * have been "cleaned". What has really occured is that
1085 * they've been earmarked for later writing by the buffer
1088 * So we get the b_dirtyoff/end update but will not actually
1089 * depend on it (NFS that is) until the pages are busied for
1092 vfs_clean_pages(bp);
1096 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1097 * due to the softdep code.
1102 * Fake write - return pages to VM system as dirty, leave the buffer clean.
1103 * This is used by tmpfs.
1105 * It is important for any VFS using this routine to NOT use it for
1106 * IO_SYNC or IO_ASYNC operations which occur when the system really
1107 * wants to flush VM pages to backing store.
1110 buwrite(struct buf *bp)
1116 * Only works for VMIO buffers. If the buffer is already
1117 * marked for delayed-write we can't avoid the bdwrite().
1119 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) {
1125 * Set valid & dirty.
1127 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1128 m = bp->b_xio.xio_pages[i];
1129 vfs_dirty_one_page(bp, i, m);
1137 * Turn buffer into delayed write request by marking it B_DELWRI.
1138 * B_RELBUF and B_NOCACHE must be cleared.
1140 * We reassign the buffer to itself to properly update it in the
1141 * dirty/clean lists.
1143 * Must be called from a critical section.
1144 * The buffer must be on BQUEUE_NONE.
1147 bdirty(struct buf *bp)
1149 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1150 if (bp->b_flags & B_NOCACHE) {
1151 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1152 bp->b_flags &= ~B_NOCACHE;
1154 if (bp->b_flags & B_INVAL) {
1155 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1157 bp->b_flags &= ~B_RELBUF;
1159 if ((bp->b_flags & B_DELWRI) == 0) {
1160 bp->b_flags |= B_DELWRI;
1162 atomic_add_int(&dirtybufcount, 1);
1163 dirtybufspace += bp->b_bufsize;
1164 if (bp->b_flags & B_HEAVY) {
1165 atomic_add_int(&dirtybufcounthw, 1);
1166 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
1173 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1174 * needs to be flushed with a different buf_daemon thread to avoid
1175 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1178 bheavy(struct buf *bp)
1180 if ((bp->b_flags & B_HEAVY) == 0) {
1181 bp->b_flags |= B_HEAVY;
1182 if (bp->b_flags & B_DELWRI) {
1183 atomic_add_int(&dirtybufcounthw, 1);
1184 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
1192 * Clear B_DELWRI for buffer.
1194 * Must be called from a critical section.
1196 * The buffer is typically on BQUEUE_NONE but there is one case in
1197 * brelse() that calls this function after placing the buffer on
1198 * a different queue.
1203 bundirty(struct buf *bp)
1205 if (bp->b_flags & B_DELWRI) {
1206 bp->b_flags &= ~B_DELWRI;
1208 atomic_subtract_int(&dirtybufcount, 1);
1209 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
1210 if (bp->b_flags & B_HEAVY) {
1211 atomic_subtract_int(&dirtybufcounthw, 1);
1212 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
1214 bd_signal(bp->b_bufsize);
1217 * Since it is now being written, we can clear its deferred write flag.
1219 bp->b_flags &= ~B_DEFERRED;
1225 * Release a busy buffer and, if requested, free its resources. The
1226 * buffer will be stashed in the appropriate bufqueue[] allowing it
1227 * to be accessed later as a cache entity or reused for other purposes.
1232 brelse(struct buf *bp)
1235 int saved_flags = bp->b_flags;
1238 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1241 * If B_NOCACHE is set we are being asked to destroy the buffer and
1242 * its backing store. Clear B_DELWRI.
1244 * B_NOCACHE is set in two cases: (1) when the caller really wants
1245 * to destroy the buffer and backing store and (2) when the caller
1246 * wants to destroy the buffer and backing store after a write
1249 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1253 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1255 * A re-dirtied buffer is only subject to destruction
1256 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1258 /* leave buffer intact */
1259 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1260 (bp->b_bufsize <= 0)) {
1262 * Either a failed read or we were asked to free or not
1263 * cache the buffer. This path is reached with B_DELWRI
1264 * set only if B_INVAL is already set. B_NOCACHE governs
1265 * backing store destruction.
1267 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1268 * buffer cannot be immediately freed.
1270 bp->b_flags |= B_INVAL;
1271 if (LIST_FIRST(&bp->b_dep) != NULL) {
1276 if (bp->b_flags & B_DELWRI) {
1277 atomic_subtract_int(&dirtybufcount, 1);
1278 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
1279 if (bp->b_flags & B_HEAVY) {
1280 atomic_subtract_int(&dirtybufcounthw, 1);
1281 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
1283 bd_signal(bp->b_bufsize);
1285 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1289 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1290 * If vfs_vmio_release() is called with either bit set, the
1291 * underlying pages may wind up getting freed causing a previous
1292 * write (bdwrite()) to get 'lost' because pages associated with
1293 * a B_DELWRI bp are marked clean. Pages associated with a
1294 * B_LOCKED buffer may be mapped by the filesystem.
1296 * If we want to release the buffer ourselves (rather then the
1297 * originator asking us to release it), give the originator a
1298 * chance to countermand the release by setting B_LOCKED.
1300 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1301 * if B_DELWRI is set.
1303 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1304 * on pages to return pages to the VM page queues.
1306 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1307 bp->b_flags &= ~B_RELBUF;
1308 } else if (vm_page_count_severe()) {
1309 if (LIST_FIRST(&bp->b_dep) != NULL) {
1311 buf_deallocate(bp); /* can set B_LOCKED */
1314 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1315 bp->b_flags &= ~B_RELBUF;
1317 bp->b_flags |= B_RELBUF;
1321 * Make sure b_cmd is clear. It may have already been cleared by
1324 * At this point destroying the buffer is governed by the B_INVAL
1325 * or B_RELBUF flags.
1327 bp->b_cmd = BUF_CMD_DONE;
1328 dsched_exit_buf(bp);
1331 * VMIO buffer rundown. Make sure the VM page array is restored
1332 * after an I/O may have replaces some of the pages with bogus pages
1333 * in order to not destroy dirty pages in a fill-in read.
1335 * Note that due to the code above, if a buffer is marked B_DELWRI
1336 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1337 * B_INVAL may still be set, however.
1339 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1340 * but not the backing store. B_NOCACHE will destroy the backing
1343 * Note that dirty NFS buffers contain byte-granular write ranges
1344 * and should not be destroyed w/ B_INVAL even if the backing store
1347 if (bp->b_flags & B_VMIO) {
1349 * Rundown for VMIO buffers which are not dirty NFS buffers.
1361 * Get the base offset and length of the buffer. Note that
1362 * in the VMIO case if the buffer block size is not
1363 * page-aligned then b_data pointer may not be page-aligned.
1364 * But our b_xio.xio_pages array *IS* page aligned.
1366 * block sizes less then DEV_BSIZE (usually 512) are not
1367 * supported due to the page granularity bits (m->valid,
1368 * m->dirty, etc...).
1370 * See man buf(9) for more information
1373 resid = bp->b_bufsize;
1374 foff = bp->b_loffset;
1377 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1378 m = bp->b_xio.xio_pages[i];
1379 vm_page_flag_clear(m, PG_ZERO);
1381 * If we hit a bogus page, fixup *all* of them
1382 * now. Note that we left these pages wired
1383 * when we removed them so they had better exist,
1384 * and they cannot be ripped out from under us so
1385 * no critical section protection is necessary.
1387 if (m == bogus_page) {
1389 poff = OFF_TO_IDX(bp->b_loffset);
1391 for (j = i; j < bp->b_xio.xio_npages; j++) {
1394 mtmp = bp->b_xio.xio_pages[j];
1395 if (mtmp == bogus_page) {
1396 mtmp = vm_page_lookup(obj, poff + j);
1398 panic("brelse: page missing");
1400 bp->b_xio.xio_pages[j] = mtmp;
1403 bp->b_flags &= ~B_HASBOGUS;
1405 if ((bp->b_flags & B_INVAL) == 0) {
1406 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1407 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1409 m = bp->b_xio.xio_pages[i];
1413 * Invalidate the backing store if B_NOCACHE is set
1414 * (e.g. used with vinvalbuf()). If this is NFS
1415 * we impose a requirement that the block size be
1416 * a multiple of PAGE_SIZE and create a temporary
1417 * hack to basically invalidate the whole page. The
1418 * problem is that NFS uses really odd buffer sizes
1419 * especially when tracking piecemeal writes and
1420 * it also vinvalbuf()'s a lot, which would result
1421 * in only partial page validation and invalidation
1422 * here. If the file page is mmap()'d, however,
1423 * all the valid bits get set so after we invalidate
1424 * here we would end up with weird m->valid values
1425 * like 0xfc. nfs_getpages() can't handle this so
1426 * we clear all the valid bits for the NFS case
1427 * instead of just some of them.
1429 * The real bug is the VM system having to set m->valid
1430 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1431 * itself is an artifact of the whole 512-byte
1432 * granular mess that exists to support odd block
1433 * sizes and UFS meta-data block sizes (e.g. 6144).
1434 * A complete rewrite is required.
1438 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1439 int poffset = foff & PAGE_MASK;
1442 presid = PAGE_SIZE - poffset;
1443 if (bp->b_vp->v_tag == VT_NFS &&
1444 bp->b_vp->v_type == VREG) {
1446 } else if (presid > resid) {
1449 KASSERT(presid >= 0, ("brelse: extra page"));
1450 vm_page_set_invalid(m, poffset, presid);
1453 * Also make sure any swap cache is removed
1454 * as it is now stale (HAMMER in particular
1455 * uses B_NOCACHE to deal with buffer
1458 swap_pager_unswapped(m);
1460 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1461 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1463 if (bp->b_flags & (B_INVAL | B_RELBUF))
1464 vfs_vmio_release(bp);
1468 * Rundown for non-VMIO buffers.
1470 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1474 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1481 if (bp->b_qindex != BQUEUE_NONE)
1482 panic("brelse: free buffer onto another queue???");
1483 if (BUF_REFCNTNB(bp) > 1) {
1484 /* Temporary panic to verify exclusive locking */
1485 /* This panic goes away when we allow shared refs */
1486 panic("brelse: multiple refs");
1492 * Figure out the correct queue to place the cleaned up buffer on.
1493 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1494 * disassociated from their vnode.
1496 spin_lock_wr(&bufspin);
1497 if (bp->b_flags & B_LOCKED) {
1499 * Buffers that are locked are placed in the locked queue
1500 * immediately, regardless of their state.
1502 bp->b_qindex = BQUEUE_LOCKED;
1503 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1504 } else if (bp->b_bufsize == 0) {
1506 * Buffers with no memory. Due to conditionals near the top
1507 * of brelse() such buffers should probably already be
1508 * marked B_INVAL and disassociated from their vnode.
1510 bp->b_flags |= B_INVAL;
1511 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1512 KKASSERT((bp->b_flags & B_HASHED) == 0);
1513 if (bp->b_kvasize) {
1514 bp->b_qindex = BQUEUE_EMPTYKVA;
1516 bp->b_qindex = BQUEUE_EMPTY;
1518 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1519 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1521 * Buffers with junk contents. Again these buffers had better
1522 * already be disassociated from their vnode.
1524 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1525 KKASSERT((bp->b_flags & B_HASHED) == 0);
1526 bp->b_flags |= B_INVAL;
1527 bp->b_qindex = BQUEUE_CLEAN;
1528 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1531 * Remaining buffers. These buffers are still associated with
1534 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1536 bp->b_qindex = BQUEUE_DIRTY;
1537 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1539 case B_DELWRI | B_HEAVY:
1540 bp->b_qindex = BQUEUE_DIRTY_HW;
1541 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1546 * NOTE: Buffers are always placed at the end of the
1547 * queue. If B_AGE is not set the buffer will cycle
1548 * through the queue twice.
1550 bp->b_qindex = BQUEUE_CLEAN;
1551 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1555 spin_unlock_wr(&bufspin);
1558 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1559 * on the correct queue.
1561 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1565 * The bp is on an appropriate queue unless locked. If it is not
1566 * locked or dirty we can wakeup threads waiting for buffer space.
1568 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1569 * if B_INVAL is set ).
1571 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1575 * Something we can maybe free or reuse
1577 if (bp->b_bufsize || bp->b_kvasize)
1581 * Clean up temporary flags and unlock the buffer.
1583 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
1590 * Release a buffer back to the appropriate queue but do not try to free
1591 * it. The buffer is expected to be used again soon.
1593 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1594 * biodone() to requeue an async I/O on completion. It is also used when
1595 * known good buffers need to be requeued but we think we may need the data
1598 * XXX we should be able to leave the B_RELBUF hint set on completion.
1603 bqrelse(struct buf *bp)
1605 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1607 if (bp->b_qindex != BQUEUE_NONE)
1608 panic("bqrelse: free buffer onto another queue???");
1609 if (BUF_REFCNTNB(bp) > 1) {
1610 /* do not release to free list */
1611 panic("bqrelse: multiple refs");
1615 buf_act_advance(bp);
1617 spin_lock_wr(&bufspin);
1618 if (bp->b_flags & B_LOCKED) {
1620 * Locked buffers are released to the locked queue. However,
1621 * if the buffer is dirty it will first go into the dirty
1622 * queue and later on after the I/O completes successfully it
1623 * will be released to the locked queue.
1625 bp->b_qindex = BQUEUE_LOCKED;
1626 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1627 } else if (bp->b_flags & B_DELWRI) {
1628 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1629 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1630 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1631 } else if (vm_page_count_severe()) {
1633 * We are too low on memory, we have to try to free the
1634 * buffer (most importantly: the wired pages making up its
1635 * backing store) *now*.
1637 spin_unlock_wr(&bufspin);
1641 bp->b_qindex = BQUEUE_CLEAN;
1642 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1644 spin_unlock_wr(&bufspin);
1646 if ((bp->b_flags & B_LOCKED) == 0 &&
1647 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1652 * Something we can maybe free or reuse.
1654 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1658 * Final cleanup and unlock. Clear bits that are only used while a
1659 * buffer is actively locked.
1661 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
1662 dsched_exit_buf(bp);
1669 * Return backing pages held by the buffer 'bp' back to the VM system
1670 * if possible. The pages are freed if they are no longer valid or
1671 * attempt to free if it was used for direct I/O otherwise they are
1672 * sent to the page cache.
1674 * Pages that were marked busy are left alone and skipped.
1676 * The KVA mapping (b_data) for the underlying pages is removed by
1680 vfs_vmio_release(struct buf *bp)
1685 lwkt_gettoken(&vm_token);
1687 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1688 m = bp->b_xio.xio_pages[i];
1689 bp->b_xio.xio_pages[i] = NULL;
1692 * The VFS is telling us this is not a meta-data buffer
1693 * even if it is backed by a block device.
1695 if (bp->b_flags & B_NOTMETA)
1696 vm_page_flag_set(m, PG_NOTMETA);
1699 * This is a very important bit of code. We try to track
1700 * VM page use whether the pages are wired into the buffer
1701 * cache or not. While wired into the buffer cache the
1702 * bp tracks the act_count.
1704 * We can choose to place unwired pages on the inactive
1705 * queue (0) or active queue (1). If we place too many
1706 * on the active queue the queue will cycle the act_count
1707 * on pages we'd like to keep, just from single-use pages
1708 * (such as when doing a tar-up or file scan).
1710 if (bp->b_act_count < vm_cycle_point)
1711 vm_page_unwire(m, 0);
1713 vm_page_unwire(m, 1);
1716 * We don't mess with busy pages, it is
1717 * the responsibility of the process that
1718 * busied the pages to deal with them.
1720 if ((m->flags & PG_BUSY) || (m->busy != 0))
1723 if (m->wire_count == 0) {
1724 vm_page_flag_clear(m, PG_ZERO);
1726 * Might as well free the page if we can and it has
1727 * no valid data. We also free the page if the
1728 * buffer was used for direct I/O.
1731 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1732 m->hold_count == 0) {
1734 vm_page_protect(m, VM_PROT_NONE);
1738 if (bp->b_flags & B_DIRECT) {
1739 vm_page_try_to_free(m);
1740 } else if (vm_page_count_severe()) {
1741 m->act_count = bp->b_act_count;
1742 vm_page_try_to_cache(m);
1744 m->act_count = bp->b_act_count;
1749 lwkt_reltoken(&vm_token);
1750 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1751 if (bp->b_bufsize) {
1755 bp->b_xio.xio_npages = 0;
1756 bp->b_flags &= ~B_VMIO;
1757 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1768 * Implement clustered async writes for clearing out B_DELWRI buffers.
1769 * This is much better then the old way of writing only one buffer at
1770 * a time. Note that we may not be presented with the buffers in the
1771 * correct order, so we search for the cluster in both directions.
1773 * The buffer is locked on call.
1776 vfs_bio_awrite(struct buf *bp)
1780 off_t loffset = bp->b_loffset;
1781 struct vnode *vp = bp->b_vp;
1788 * right now we support clustered writing only to regular files. If
1789 * we find a clusterable block we could be in the middle of a cluster
1790 * rather then at the beginning.
1792 * NOTE: b_bio1 contains the logical loffset and is aliased
1793 * to b_loffset. b_bio2 contains the translated block number.
1795 if ((vp->v_type == VREG) &&
1796 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1797 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1799 size = vp->v_mount->mnt_stat.f_iosize;
1801 for (i = size; i < MAXPHYS; i += size) {
1802 if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) &&
1803 BUF_REFCNT(bpa) == 0 &&
1804 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1805 (B_DELWRI | B_CLUSTEROK)) &&
1806 (bpa->b_bufsize == size)) {
1807 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1808 (bpa->b_bio2.bio_offset !=
1809 bp->b_bio2.bio_offset + i))
1815 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1816 if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) &&
1817 BUF_REFCNT(bpa) == 0 &&
1818 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1819 (B_DELWRI | B_CLUSTEROK)) &&
1820 (bpa->b_bufsize == size)) {
1821 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1822 (bpa->b_bio2.bio_offset !=
1823 bp->b_bio2.bio_offset - j))
1833 * this is a possible cluster write
1835 if (nbytes != size) {
1837 nwritten = cluster_wbuild(vp, size,
1838 loffset - j, nbytes);
1844 * default (old) behavior, writing out only one block
1846 * XXX returns b_bufsize instead of b_bcount for nwritten?
1848 nwritten = bp->b_bufsize;
1858 * Find and initialize a new buffer header, freeing up existing buffers
1859 * in the bufqueues as necessary. The new buffer is returned locked.
1861 * Important: B_INVAL is not set. If the caller wishes to throw the
1862 * buffer away, the caller must set B_INVAL prior to calling brelse().
1865 * We have insufficient buffer headers
1866 * We have insufficient buffer space
1867 * buffer_map is too fragmented ( space reservation fails )
1868 * If we have to flush dirty buffers ( but we try to avoid this )
1870 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1871 * Instead we ask the buf daemon to do it for us. We attempt to
1872 * avoid piecemeal wakeups of the pageout daemon.
1877 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1883 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1884 static int flushingbufs;
1887 * We can't afford to block since we might be holding a vnode lock,
1888 * which may prevent system daemons from running. We deal with
1889 * low-memory situations by proactively returning memory and running
1890 * async I/O rather then sync I/O.
1894 --getnewbufrestarts;
1896 ++getnewbufrestarts;
1899 * Setup for scan. If we do not have enough free buffers,
1900 * we setup a degenerate case that immediately fails. Note
1901 * that if we are specially marked process, we are allowed to
1902 * dip into our reserves.
1904 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1906 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1907 * However, there are a number of cases (defragging, reusing, ...)
1908 * where we cannot backup.
1910 nqindex = BQUEUE_EMPTYKVA;
1911 spin_lock_wr(&bufspin);
1912 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1916 * If no EMPTYKVA buffers and we are either
1917 * defragging or reusing, locate a CLEAN buffer
1918 * to free or reuse. If bufspace useage is low
1919 * skip this step so we can allocate a new buffer.
1921 if (defrag || bufspace >= lobufspace) {
1922 nqindex = BQUEUE_CLEAN;
1923 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1927 * If we could not find or were not allowed to reuse a
1928 * CLEAN buffer, check to see if it is ok to use an EMPTY
1929 * buffer. We can only use an EMPTY buffer if allocating
1930 * its KVA would not otherwise run us out of buffer space.
1932 if (nbp == NULL && defrag == 0 &&
1933 bufspace + maxsize < hibufspace) {
1934 nqindex = BQUEUE_EMPTY;
1935 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1940 * Run scan, possibly freeing data and/or kva mappings on the fly
1943 * WARNING! bufspin is held!
1945 while ((bp = nbp) != NULL) {
1946 int qindex = nqindex;
1948 nbp = TAILQ_NEXT(bp, b_freelist);
1951 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1952 * cycles through the queue twice before being selected.
1954 if (qindex == BQUEUE_CLEAN &&
1955 (bp->b_flags & B_AGE) == 0 && nbp) {
1956 bp->b_flags |= B_AGE;
1957 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1958 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1963 * Calculate next bp ( we can only use it if we do not block
1964 * or do other fancy things ).
1969 nqindex = BQUEUE_EMPTYKVA;
1970 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1973 case BQUEUE_EMPTYKVA:
1974 nqindex = BQUEUE_CLEAN;
1975 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1989 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1992 * Note: we no longer distinguish between VMIO and non-VMIO
1996 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1999 * If we are defragging then we need a buffer with
2000 * b_kvasize != 0. XXX this situation should no longer
2001 * occur, if defrag is non-zero the buffer's b_kvasize
2002 * should also be non-zero at this point. XXX
2004 if (defrag && bp->b_kvasize == 0) {
2005 kprintf("Warning: defrag empty buffer %p\n", bp);
2010 * Start freeing the bp. This is somewhat involved. nbp
2011 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
2012 * on the clean list must be disassociated from their
2013 * current vnode. Buffers on the empty[kva] lists have
2014 * already been disassociated.
2017 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2018 spin_unlock_wr(&bufspin);
2019 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
2022 if (bp->b_qindex != qindex) {
2023 spin_unlock_wr(&bufspin);
2024 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
2028 bremfree_locked(bp);
2029 spin_unlock_wr(&bufspin);
2032 * Dependancies must be handled before we disassociate the
2035 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2036 * be immediately disassociated. HAMMER then becomes
2037 * responsible for releasing the buffer.
2039 * NOTE: bufspin is UNLOCKED now.
2041 if (LIST_FIRST(&bp->b_dep) != NULL) {
2045 if (bp->b_flags & B_LOCKED) {
2049 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2052 if (qindex == BQUEUE_CLEAN) {
2054 if (bp->b_flags & B_VMIO) {
2056 vfs_vmio_release(bp);
2065 * NOTE: nbp is now entirely invalid. We can only restart
2066 * the scan from this point on.
2068 * Get the rest of the buffer freed up. b_kva* is still
2069 * valid after this operation.
2072 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
2073 KKASSERT((bp->b_flags & B_HASHED) == 0);
2076 * critical section protection is not required when
2077 * scrapping a buffer's contents because it is already
2080 if (bp->b_bufsize) {
2086 bp->b_flags = B_BNOCLIP;
2087 bp->b_cmd = BUF_CMD_DONE;
2092 bp->b_xio.xio_npages = 0;
2093 bp->b_dirtyoff = bp->b_dirtyend = 0;
2094 bp->b_act_count = ACT_INIT;
2096 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2098 if (blkflags & GETBLK_BHEAVY)
2099 bp->b_flags |= B_HEAVY;
2102 * If we are defragging then free the buffer.
2105 bp->b_flags |= B_INVAL;
2113 * If we are overcomitted then recover the buffer and its
2114 * KVM space. This occurs in rare situations when multiple
2115 * processes are blocked in getnewbuf() or allocbuf().
2117 if (bufspace >= hibufspace)
2119 if (flushingbufs && bp->b_kvasize != 0) {
2120 bp->b_flags |= B_INVAL;
2125 if (bufspace < lobufspace)
2128 /* NOT REACHED, bufspin not held */
2132 * If we exhausted our list, sleep as appropriate. We may have to
2133 * wakeup various daemons and write out some dirty buffers.
2135 * Generally we are sleeping due to insufficient buffer space.
2137 * NOTE: bufspin is held if bp is NULL, else it is not held.
2143 spin_unlock_wr(&bufspin);
2145 flags = VFS_BIO_NEED_BUFSPACE;
2147 } else if (bufspace >= hibufspace) {
2149 flags = VFS_BIO_NEED_BUFSPACE;
2152 flags = VFS_BIO_NEED_ANY;
2155 needsbuffer |= flags;
2156 bd_speedup(); /* heeeelp */
2157 while (needsbuffer & flags) {
2158 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
2163 * We finally have a valid bp. We aren't quite out of the
2164 * woods, we still have to reserve kva space. In order
2165 * to keep fragmentation sane we only allocate kva in
2168 * (bufspin is not held)
2170 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2172 if (maxsize != bp->b_kvasize) {
2173 vm_offset_t addr = 0;
2179 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2180 vm_map_lock(&buffer_map);
2182 if (vm_map_findspace(&buffer_map,
2183 vm_map_min(&buffer_map), maxsize,
2184 maxsize, 0, &addr)) {
2186 * Uh oh. Buffer map is too fragmented. We
2187 * must defragment the map.
2189 vm_map_unlock(&buffer_map);
2190 vm_map_entry_release(count);
2193 bp->b_flags |= B_INVAL;
2199 vm_map_insert(&buffer_map, &count,
2201 addr, addr + maxsize,
2203 VM_PROT_ALL, VM_PROT_ALL,
2206 bp->b_kvabase = (caddr_t) addr;
2207 bp->b_kvasize = maxsize;
2208 bufspace += bp->b_kvasize;
2211 vm_map_unlock(&buffer_map);
2212 vm_map_entry_release(count);
2215 bp->b_data = bp->b_kvabase;
2221 * This routine is called in an emergency to recover VM pages from the
2222 * buffer cache by cashing in clean buffers. The idea is to recover
2223 * enough pages to be able to satisfy a stuck bio_page_alloc().
2226 recoverbufpages(void)
2233 spin_lock_wr(&bufspin);
2234 while (bytes < MAXBSIZE) {
2235 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2240 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2241 * cycles through the queue twice before being selected.
2243 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
2244 bp->b_flags |= B_AGE;
2245 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
2246 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
2254 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
2255 KKASSERT((bp->b_flags & B_DELWRI) == 0);
2258 * Start freeing the bp. This is somewhat involved.
2260 * Buffers on the clean list must be disassociated from
2261 * their current vnode
2264 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2265 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp);
2266 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
2269 if (bp->b_qindex != BQUEUE_CLEAN) {
2270 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp, bp->b_qindex);
2274 bremfree_locked(bp);
2275 spin_unlock_wr(&bufspin);
2278 * Dependancies must be handled before we disassociate the
2281 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2282 * be immediately disassociated. HAMMER then becomes
2283 * responsible for releasing the buffer.
2285 if (LIST_FIRST(&bp->b_dep) != NULL) {
2287 if (bp->b_flags & B_LOCKED) {
2289 spin_lock_wr(&bufspin);
2292 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2295 bytes += bp->b_bufsize;
2298 if (bp->b_flags & B_VMIO) {
2299 bp->b_flags |= B_DIRECT; /* try to free pages */
2300 vfs_vmio_release(bp);
2305 KKASSERT(bp->b_vp == NULL);
2306 KKASSERT((bp->b_flags & B_HASHED) == 0);
2309 * critical section protection is not required when
2310 * scrapping a buffer's contents because it is already
2317 bp->b_flags = B_BNOCLIP;
2318 bp->b_cmd = BUF_CMD_DONE;
2323 bp->b_xio.xio_npages = 0;
2324 bp->b_dirtyoff = bp->b_dirtyend = 0;
2326 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2328 bp->b_flags |= B_INVAL;
2331 spin_lock_wr(&bufspin);
2333 spin_unlock_wr(&bufspin);
2340 * Buffer flushing daemon. Buffers are normally flushed by the
2341 * update daemon but if it cannot keep up this process starts to
2342 * take the load in an attempt to prevent getnewbuf() from blocking.
2344 * Once a flush is initiated it does not stop until the number
2345 * of buffers falls below lodirtybuffers, but we will wake up anyone
2346 * waiting at the mid-point.
2349 static struct kproc_desc buf_kp = {
2354 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2355 kproc_start, &buf_kp)
2357 static struct kproc_desc bufhw_kp = {
2362 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2363 kproc_start, &bufhw_kp)
2371 * This process needs to be suspended prior to shutdown sync.
2373 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2374 bufdaemon_td, SHUTDOWN_PRI_LAST);
2375 curthread->td_flags |= TDF_SYSTHREAD;
2378 * This process is allowed to take the buffer cache to the limit
2383 kproc_suspend_loop();
2386 * Do the flush as long as the number of dirty buffers
2387 * (including those running) exceeds lodirtybufspace.
2389 * When flushing limit running I/O to hirunningspace
2390 * Do the flush. Limit the amount of in-transit I/O we
2391 * allow to build up, otherwise we would completely saturate
2392 * the I/O system. Wakeup any waiting processes before we
2393 * normally would so they can run in parallel with our drain.
2395 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2396 * but because we split the operation into two threads we
2397 * have to cut it in half for each thread.
2399 waitrunningbufspace();
2400 limit = lodirtybufspace / 2;
2401 while (runningbufspace + dirtybufspace > limit ||
2402 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
2403 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2405 if (runningbufspace < hirunningspace)
2407 waitrunningbufspace();
2411 * We reached our low water mark, reset the
2412 * request and sleep until we are needed again.
2413 * The sleep is just so the suspend code works.
2415 spin_lock_wr(&needsbuffer_spin);
2416 if (bd_request == 0) {
2417 ssleep(&bd_request, &needsbuffer_spin, 0,
2421 spin_unlock_wr(&needsbuffer_spin);
2431 * This process needs to be suspended prior to shutdown sync.
2433 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2434 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2435 curthread->td_flags |= TDF_SYSTHREAD;
2438 * This process is allowed to take the buffer cache to the limit
2443 kproc_suspend_loop();
2446 * Do the flush. Limit the amount of in-transit I/O we
2447 * allow to build up, otherwise we would completely saturate
2448 * the I/O system. Wakeup any waiting processes before we
2449 * normally would so they can run in parallel with our drain.
2451 * Once we decide to flush push the queued I/O up to
2452 * hirunningspace in order to trigger bursting by the bioq
2455 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2456 * but because we split the operation into two threads we
2457 * have to cut it in half for each thread.
2459 waitrunningbufspace();
2460 limit = lodirtybufspace / 2;
2461 while (runningbufspace + dirtybufspacehw > limit ||
2462 dirtybufcounthw >= nbuf / 2) {
2463 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2465 if (runningbufspace < hirunningspace)
2467 waitrunningbufspace();
2471 * We reached our low water mark, reset the
2472 * request and sleep until we are needed again.
2473 * The sleep is just so the suspend code works.
2475 spin_lock_wr(&needsbuffer_spin);
2476 if (bd_request_hw == 0) {
2477 ssleep(&bd_request_hw, &needsbuffer_spin, 0,
2481 spin_unlock_wr(&needsbuffer_spin);
2488 * Try to flush a buffer in the dirty queue. We must be careful to
2489 * free up B_INVAL buffers instead of write them, which NFS is
2490 * particularly sensitive to.
2492 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2493 * that we really want to try to get the buffer out and reuse it
2494 * due to the write load on the machine.
2497 flushbufqueues(bufq_type_t q)
2503 spin_lock_wr(&bufspin);
2506 bp = TAILQ_FIRST(&bufqueues[q]);
2508 KASSERT((bp->b_flags & B_DELWRI),
2509 ("unexpected clean buffer %p", bp));
2511 if (bp->b_flags & B_DELWRI) {
2512 if (bp->b_flags & B_INVAL) {
2513 spin_unlock_wr(&bufspin);
2515 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2516 panic("flushbufqueues: locked buf");
2522 if (LIST_FIRST(&bp->b_dep) != NULL &&
2523 (bp->b_flags & B_DEFERRED) == 0 &&
2524 buf_countdeps(bp, 0)) {
2525 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2526 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2528 bp->b_flags |= B_DEFERRED;
2529 bp = TAILQ_FIRST(&bufqueues[q]);
2534 * Only write it out if we can successfully lock
2535 * it. If the buffer has a dependancy,
2536 * buf_checkwrite must also return 0 for us to
2537 * be able to initate the write.
2539 * If the buffer is flagged B_ERROR it may be
2540 * requeued over and over again, we try to
2541 * avoid a live lock.
2543 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
2544 spin_unlock_wr(&bufspin);
2546 if (LIST_FIRST(&bp->b_dep) != NULL &&
2547 buf_checkwrite(bp)) {
2550 } else if (bp->b_flags & B_ERROR) {
2551 tsleep(bp, 0, "bioer", 1);
2552 bp->b_flags &= ~B_AGE;
2555 bp->b_flags |= B_AGE;
2562 bp = TAILQ_NEXT(bp, b_freelist);
2565 spin_unlock_wr(&bufspin);
2572 * Returns true if no I/O is needed to access the associated VM object.
2573 * This is like findblk except it also hunts around in the VM system for
2576 * Note that we ignore vm_page_free() races from interrupts against our
2577 * lookup, since if the caller is not protected our return value will not
2578 * be any more valid then otherwise once we exit the critical section.
2581 inmem(struct vnode *vp, off_t loffset)
2584 vm_offset_t toff, tinc, size;
2587 if (findblk(vp, loffset, FINDBLK_TEST))
2589 if (vp->v_mount == NULL)
2591 if ((obj = vp->v_object) == NULL)
2595 if (size > vp->v_mount->mnt_stat.f_iosize)
2596 size = vp->v_mount->mnt_stat.f_iosize;
2598 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2599 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2603 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2604 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2605 if (vm_page_is_valid(m,
2606 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2615 * Locate and return the specified buffer. Unless flagged otherwise,
2616 * a locked buffer will be returned if it exists or NULL if it does not.
2618 * findblk()'d buffers are still on the bufqueues and if you intend
2619 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2620 * and possibly do other stuff to it.
2622 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2623 * for locking the buffer and ensuring that it remains
2624 * the desired buffer after locking.
2626 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2627 * to acquire the lock we return NULL, even if the
2630 * (0) - Lock the buffer blocking.
2635 findblk(struct vnode *vp, off_t loffset, int flags)
2640 lkflags = LK_EXCLUSIVE;
2641 if (flags & FINDBLK_NBLOCK)
2642 lkflags |= LK_NOWAIT;
2645 lwkt_gettoken(&vp->v_token);
2646 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2647 lwkt_reltoken(&vp->v_token);
2648 if (bp == NULL || (flags & FINDBLK_TEST))
2650 if (BUF_LOCK(bp, lkflags)) {
2654 if (bp->b_vp == vp && bp->b_loffset == loffset)
2664 * Similar to getblk() except only returns the buffer if it is
2665 * B_CACHE and requires no other manipulation. Otherwise NULL
2668 * If B_RAM is set the buffer might be just fine, but we return
2669 * NULL anyway because we want the code to fall through to the
2670 * cluster read. Otherwise read-ahead breaks.
2673 getcacheblk(struct vnode *vp, off_t loffset)
2677 bp = findblk(vp, loffset, 0);
2679 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
2680 bp->b_flags &= ~B_AGE;
2693 * Get a block given a specified block and offset into a file/device.
2694 * B_INVAL may or may not be set on return. The caller should clear
2695 * B_INVAL prior to initiating a READ.
2697 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2698 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2699 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2700 * without doing any of those things the system will likely believe
2701 * the buffer to be valid (especially if it is not B_VMIO), and the
2702 * next getblk() will return the buffer with B_CACHE set.
2704 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2705 * an existing buffer.
2707 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2708 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2709 * and then cleared based on the backing VM. If the previous buffer is
2710 * non-0-sized but invalid, B_CACHE will be cleared.
2712 * If getblk() must create a new buffer, the new buffer is returned with
2713 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2714 * case it is returned with B_INVAL clear and B_CACHE set based on the
2717 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2718 * B_CACHE bit is clear.
2720 * What this means, basically, is that the caller should use B_CACHE to
2721 * determine whether the buffer is fully valid or not and should clear
2722 * B_INVAL prior to issuing a read. If the caller intends to validate
2723 * the buffer by loading its data area with something, the caller needs
2724 * to clear B_INVAL. If the caller does this without issuing an I/O,
2725 * the caller should set B_CACHE ( as an optimization ), else the caller
2726 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2727 * a write attempt or if it was a successfull read. If the caller
2728 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2729 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2733 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2734 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2739 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2742 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2746 if (size > MAXBSIZE)
2747 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2748 if (vp->v_object == NULL)
2749 panic("getblk: vnode %p has no object!", vp);
2752 if ((bp = findblk(vp, loffset, FINDBLK_TEST)) != NULL) {
2754 * The buffer was found in the cache, but we need to lock it.
2755 * Even with LK_NOWAIT the lockmgr may break our critical
2756 * section, so double-check the validity of the buffer
2757 * once the lock has been obtained.
2759 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2760 if (blkflags & GETBLK_NOWAIT)
2762 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2763 if (blkflags & GETBLK_PCATCH)
2764 lkflags |= LK_PCATCH;
2765 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2767 if (error == ENOLCK)
2771 /* buffer may have changed on us */
2775 * Once the buffer has been locked, make sure we didn't race
2776 * a buffer recyclement. Buffers that are no longer hashed
2777 * will have b_vp == NULL, so this takes care of that check
2780 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2781 kprintf("Warning buffer %p (vp %p loffset %lld) "
2783 bp, vp, (long long)loffset);
2789 * If SZMATCH any pre-existing buffer must be of the requested
2790 * size or NULL is returned. The caller absolutely does not
2791 * want getblk() to bwrite() the buffer on a size mismatch.
2793 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2799 * All vnode-based buffers must be backed by a VM object.
2801 KKASSERT(bp->b_flags & B_VMIO);
2802 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2803 bp->b_flags &= ~B_AGE;
2806 * Make sure that B_INVAL buffers do not have a cached
2807 * block number translation.
2809 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2810 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2811 " did not have cleared bio_offset cache\n",
2812 bp, vp, (long long)loffset);
2813 clearbiocache(&bp->b_bio2);
2817 * The buffer is locked. B_CACHE is cleared if the buffer is
2820 if (bp->b_flags & B_INVAL)
2821 bp->b_flags &= ~B_CACHE;
2825 * Any size inconsistancy with a dirty buffer or a buffer
2826 * with a softupdates dependancy must be resolved. Resizing
2827 * the buffer in such circumstances can lead to problems.
2829 * Dirty or dependant buffers are written synchronously.
2830 * Other types of buffers are simply released and
2831 * reconstituted as they may be backed by valid, dirty VM
2832 * pages (but not marked B_DELWRI).
2834 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2835 * and may be left over from a prior truncation (and thus
2836 * no longer represent the actual EOF point), so we
2837 * definitely do not want to B_NOCACHE the backing store.
2839 if (size != bp->b_bcount) {
2841 if (bp->b_flags & B_DELWRI) {
2842 bp->b_flags |= B_RELBUF;
2844 } else if (LIST_FIRST(&bp->b_dep)) {
2845 bp->b_flags |= B_RELBUF;
2848 bp->b_flags |= B_RELBUF;
2854 KKASSERT(size <= bp->b_kvasize);
2855 KASSERT(bp->b_loffset != NOOFFSET,
2856 ("getblk: no buffer offset"));
2859 * A buffer with B_DELWRI set and B_CACHE clear must
2860 * be committed before we can return the buffer in
2861 * order to prevent the caller from issuing a read
2862 * ( due to B_CACHE not being set ) and overwriting
2865 * Most callers, including NFS and FFS, need this to
2866 * operate properly either because they assume they
2867 * can issue a read if B_CACHE is not set, or because
2868 * ( for example ) an uncached B_DELWRI might loop due
2869 * to softupdates re-dirtying the buffer. In the latter
2870 * case, B_CACHE is set after the first write completes,
2871 * preventing further loops.
2873 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2874 * above while extending the buffer, we cannot allow the
2875 * buffer to remain with B_CACHE set after the write
2876 * completes or it will represent a corrupt state. To
2877 * deal with this we set B_NOCACHE to scrap the buffer
2880 * XXX Should this be B_RELBUF instead of B_NOCACHE?
2881 * I'm not even sure this state is still possible
2882 * now that getblk() writes out any dirty buffers
2885 * We might be able to do something fancy, like setting
2886 * B_CACHE in bwrite() except if B_DELWRI is already set,
2887 * so the below call doesn't set B_CACHE, but that gets real
2888 * confusing. This is much easier.
2891 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2893 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
2894 "and CACHE clear, b_flags %08x\n",
2895 bp, (intmax_t)bp->b_loffset, bp->b_flags);
2896 bp->b_flags |= B_NOCACHE;
2903 * Buffer is not in-core, create new buffer. The buffer
2904 * returned by getnewbuf() is locked. Note that the returned
2905 * buffer is also considered valid (not marked B_INVAL).
2907 * Calculating the offset for the I/O requires figuring out
2908 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2909 * the mount's f_iosize otherwise. If the vnode does not
2910 * have an associated mount we assume that the passed size is
2913 * Note that vn_isdisk() cannot be used here since it may
2914 * return a failure for numerous reasons. Note that the
2915 * buffer size may be larger then the block size (the caller
2916 * will use block numbers with the proper multiple). Beware
2917 * of using any v_* fields which are part of unions. In
2918 * particular, in DragonFly the mount point overloading
2919 * mechanism uses the namecache only and the underlying
2920 * directory vnode is not a special case.
2924 if (vp->v_type == VBLK || vp->v_type == VCHR)
2926 else if (vp->v_mount)
2927 bsize = vp->v_mount->mnt_stat.f_iosize;
2931 maxsize = size + (loffset & PAGE_MASK);
2932 maxsize = imax(maxsize, bsize);
2934 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
2936 if (slpflags || slptimeo)
2942 * Atomically insert the buffer into the hash, so that it can
2943 * be found by findblk().
2945 * If bgetvp() returns non-zero a collision occured, and the
2946 * bp will not be associated with the vnode.
2948 * Make sure the translation layer has been cleared.
2950 bp->b_loffset = loffset;
2951 bp->b_bio2.bio_offset = NOOFFSET;
2952 /* bp->b_bio2.bio_next = NULL; */
2954 if (bgetvp(vp, bp, size)) {
2955 bp->b_flags |= B_INVAL;
2961 * All vnode-based buffers must be backed by a VM object.
2963 KKASSERT(vp->v_object != NULL);
2964 bp->b_flags |= B_VMIO;
2965 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2971 KKASSERT(dsched_is_clear_buf_priv(bp));
2978 * Reacquire a buffer that was previously released to the locked queue,
2979 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2980 * set B_LOCKED (which handles the acquisition race).
2982 * To this end, either B_LOCKED must be set or the dependancy list must be
2988 regetblk(struct buf *bp)
2990 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
2991 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2998 * Get an empty, disassociated buffer of given size. The buffer is
2999 * initially set to B_INVAL.
3001 * critical section protection is not required for the allocbuf()
3002 * call because races are impossible here.
3012 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3014 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
3019 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3020 KKASSERT(dsched_is_clear_buf_priv(bp));
3028 * This code constitutes the buffer memory from either anonymous system
3029 * memory (in the case of non-VMIO operations) or from an associated
3030 * VM object (in the case of VMIO operations). This code is able to
3031 * resize a buffer up or down.
3033 * Note that this code is tricky, and has many complications to resolve
3034 * deadlock or inconsistant data situations. Tread lightly!!!
3035 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3036 * the caller. Calling this code willy nilly can result in the loss of data.
3038 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3039 * B_CACHE for the non-VMIO case.
3041 * This routine does not need to be called from a critical section but you
3042 * must own the buffer.
3047 allocbuf(struct buf *bp, int size)
3049 int newbsize, mbsize;
3052 if (BUF_REFCNT(bp) == 0)
3053 panic("allocbuf: buffer not busy");
3055 if (bp->b_kvasize < size)
3056 panic("allocbuf: buffer too small");
3058 if ((bp->b_flags & B_VMIO) == 0) {
3062 * Just get anonymous memory from the kernel. Don't
3063 * mess with B_CACHE.
3065 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3066 if (bp->b_flags & B_MALLOC)
3069 newbsize = round_page(size);
3071 if (newbsize < bp->b_bufsize) {
3073 * Malloced buffers are not shrunk
3075 if (bp->b_flags & B_MALLOC) {
3077 bp->b_bcount = size;
3079 kfree(bp->b_data, M_BIOBUF);
3080 if (bp->b_bufsize) {
3081 bufmallocspace -= bp->b_bufsize;
3085 bp->b_data = bp->b_kvabase;
3087 bp->b_flags &= ~B_MALLOC;
3093 (vm_offset_t) bp->b_data + newbsize,
3094 (vm_offset_t) bp->b_data + bp->b_bufsize);
3095 } else if (newbsize > bp->b_bufsize) {
3097 * We only use malloced memory on the first allocation.
3098 * and revert to page-allocated memory when the buffer
3101 if ((bufmallocspace < maxbufmallocspace) &&
3102 (bp->b_bufsize == 0) &&
3103 (mbsize <= PAGE_SIZE/2)) {
3105 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3106 bp->b_bufsize = mbsize;
3107 bp->b_bcount = size;
3108 bp->b_flags |= B_MALLOC;
3109 bufmallocspace += mbsize;
3115 * If the buffer is growing on its other-than-first
3116 * allocation, then we revert to the page-allocation
3119 if (bp->b_flags & B_MALLOC) {
3120 origbuf = bp->b_data;
3121 origbufsize = bp->b_bufsize;
3122 bp->b_data = bp->b_kvabase;
3123 if (bp->b_bufsize) {
3124 bufmallocspace -= bp->b_bufsize;
3128 bp->b_flags &= ~B_MALLOC;
3129 newbsize = round_page(newbsize);
3133 (vm_offset_t) bp->b_data + bp->b_bufsize,
3134 (vm_offset_t) bp->b_data + newbsize);
3136 bcopy(origbuf, bp->b_data, origbufsize);
3137 kfree(origbuf, M_BIOBUF);
3144 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3145 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3146 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3147 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3149 if (bp->b_flags & B_MALLOC)
3150 panic("allocbuf: VMIO buffer can't be malloced");
3152 * Set B_CACHE initially if buffer is 0 length or will become
3155 if (size == 0 || bp->b_bufsize == 0)
3156 bp->b_flags |= B_CACHE;
3158 if (newbsize < bp->b_bufsize) {
3160 * DEV_BSIZE aligned new buffer size is less then the
3161 * DEV_BSIZE aligned existing buffer size. Figure out
3162 * if we have to remove any pages.
3164 if (desiredpages < bp->b_xio.xio_npages) {
3165 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3167 * the page is not freed here -- it
3168 * is the responsibility of
3169 * vnode_pager_setsize
3171 m = bp->b_xio.xio_pages[i];
3172 KASSERT(m != bogus_page,
3173 ("allocbuf: bogus page found"));
3174 while (vm_page_sleep_busy(m, TRUE, "biodep"))
3177 bp->b_xio.xio_pages[i] = NULL;
3178 vm_page_unwire(m, 0);
3180 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3181 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3182 bp->b_xio.xio_npages = desiredpages;
3184 } else if (size > bp->b_bcount) {
3186 * We are growing the buffer, possibly in a
3187 * byte-granular fashion.
3195 * Step 1, bring in the VM pages from the object,
3196 * allocating them if necessary. We must clear
3197 * B_CACHE if these pages are not valid for the
3198 * range covered by the buffer.
3200 * critical section protection is required to protect
3201 * against interrupts unbusying and freeing pages
3202 * between our vm_page_lookup() and our
3203 * busycheck/wiring call.
3209 while (bp->b_xio.xio_npages < desiredpages) {
3213 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
3214 if ((m = vm_page_lookup(obj, pi)) == NULL) {
3216 * note: must allocate system pages
3217 * since blocking here could intefere
3218 * with paging I/O, no matter which
3221 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
3225 vm_page_flag_clear(m, PG_ZERO);
3226 bp->b_flags &= ~B_CACHE;
3227 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3228 ++bp->b_xio.xio_npages;
3234 * We found a page. If we have to sleep on it,
3235 * retry because it might have gotten freed out
3238 * We can only test PG_BUSY here. Blocking on
3239 * m->busy might lead to a deadlock:
3241 * vm_fault->getpages->cluster_read->allocbuf
3245 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
3247 vm_page_flag_clear(m, PG_ZERO);
3249 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3250 ++bp->b_xio.xio_npages;
3251 if (bp->b_act_count < m->act_count)
3252 bp->b_act_count = m->act_count;
3257 * Step 2. We've loaded the pages into the buffer,
3258 * we have to figure out if we can still have B_CACHE
3259 * set. Note that B_CACHE is set according to the
3260 * byte-granular range ( bcount and size ), not the
3261 * aligned range ( newbsize ).
3263 * The VM test is against m->valid, which is DEV_BSIZE
3264 * aligned. Needless to say, the validity of the data
3265 * needs to also be DEV_BSIZE aligned. Note that this
3266 * fails with NFS if the server or some other client
3267 * extends the file's EOF. If our buffer is resized,
3268 * B_CACHE may remain set! XXX
3271 toff = bp->b_bcount;
3272 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3274 while ((bp->b_flags & B_CACHE) && toff < size) {
3277 if (tinc > (size - toff))
3280 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3288 bp->b_xio.xio_pages[pi]
3295 * Step 3, fixup the KVM pmap. Remember that
3296 * bp->b_data is relative to bp->b_loffset, but
3297 * bp->b_loffset may be offset into the first page.
3300 bp->b_data = (caddr_t)
3301 trunc_page((vm_offset_t)bp->b_data);
3303 (vm_offset_t)bp->b_data,
3304 bp->b_xio.xio_pages,
3305 bp->b_xio.xio_npages
3307 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3308 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3312 /* adjust space use on already-dirty buffer */
3313 if (bp->b_flags & B_DELWRI) {
3314 dirtybufspace += newbsize - bp->b_bufsize;
3315 if (bp->b_flags & B_HEAVY)
3316 dirtybufspacehw += newbsize - bp->b_bufsize;
3318 if (newbsize < bp->b_bufsize)
3320 bp->b_bufsize = newbsize; /* actual buffer allocation */
3321 bp->b_bcount = size; /* requested buffer size */
3328 * Wait for buffer I/O completion, returning error status. B_EINTR
3329 * is converted into an EINTR error but not cleared (since a chain
3330 * of biowait() calls may occur).
3332 * On return bpdone() will have been called but the buffer will remain
3333 * locked and will not have been brelse()'d.
3335 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3336 * likely still in progress on return.
3338 * NOTE! This operation is on a BIO, not a BUF.
3340 * NOTE! BIO_DONE is cleared by vn_strategy()
3345 _biowait(struct bio *bio, const char *wmesg, int to)
3347 struct buf *bp = bio->bio_buf;
3352 KKASSERT(bio == &bp->b_bio1);
3354 flags = bio->bio_flags;
3355 if (flags & BIO_DONE)
3357 tsleep_interlock(bio, 0);
3358 nflags = flags | BIO_WANT;
3359 tsleep_interlock(bio, 0);
3360 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3362 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3363 else if (bp->b_cmd == BUF_CMD_READ)
3364 error = tsleep(bio, PINTERLOCKED, "biord", to);
3366 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3368 kprintf("tsleep error biowait %d\n", error);
3377 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3378 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3379 if (bp->b_flags & B_EINTR)
3381 if (bp->b_flags & B_ERROR)
3382 return (bp->b_error ? bp->b_error : EIO);
3387 biowait(struct bio *bio, const char *wmesg)
3389 return(_biowait(bio, wmesg, 0));
3393 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3395 return(_biowait(bio, wmesg, to));
3399 * This associates a tracking count with an I/O. vn_strategy() and
3400 * dev_dstrategy() do this automatically but there are a few cases
3401 * where a vnode or device layer is bypassed when a block translation
3402 * is cached. In such cases bio_start_transaction() may be called on
3403 * the bypassed layers so the system gets an I/O in progress indication
3404 * for those higher layers.
3407 bio_start_transaction(struct bio *bio, struct bio_track *track)
3409 bio->bio_track = track;
3410 if (dsched_is_clear_buf_priv(bio->bio_buf))
3411 dsched_new_buf(bio->bio_buf);
3412 bio_track_ref(track);
3416 * Initiate I/O on a vnode.
3418 * SWAPCACHE OPERATION:
3420 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately
3421 * devfs also uses b_vp for fake buffers so we also have to check
3422 * that B_PAGING is 0. In this case the passed 'vp' is probably the
3423 * underlying block device. The swap assignments are related to the
3424 * buffer cache buffer's b_vp, not the passed vp.
3426 * The passed vp == bp->b_vp only in the case where the strategy call
3427 * is made on the vp itself for its own buffers (a regular file or
3428 * block device vp). The filesystem usually then re-calls vn_strategy()
3429 * after translating the request to an underlying device.
3431 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the
3432 * underlying buffer cache buffers.
3434 * We can only deal with page-aligned buffers at the moment, because
3435 * we can't tell what the real dirty state for pages straddling a buffer
3438 * In order to call swap_pager_strategy() we must provide the VM object
3439 * and base offset for the underlying buffer cache pages so it can find
3443 vn_strategy(struct vnode *vp, struct bio *bio)
3445 struct bio_track *track;
3446 struct buf *bp = bio->bio_buf;
3448 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3451 * Handle the swap cache intercept.
3453 if (vn_cache_strategy(vp, bio))
3457 * Otherwise do the operation through the filesystem
3459 if (bp->b_cmd == BUF_CMD_READ)
3460 track = &vp->v_track_read;
3462 track = &vp->v_track_write;
3463 KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3464 bio->bio_track = track;
3465 if (dsched_is_clear_buf_priv(bio->bio_buf))
3466 dsched_new_buf(bio->bio_buf);
3467 bio_track_ref(track);
3468 vop_strategy(*vp->v_ops, vp, bio);
3472 vn_cache_strategy(struct vnode *vp, struct bio *bio)
3474 struct buf *bp = bio->bio_buf;
3481 * Is this buffer cache buffer suitable for reading from
3484 if (vm_swapcache_read_enable == 0 ||
3485 bp->b_cmd != BUF_CMD_READ ||
3486 ((bp->b_flags & B_CLUSTER) == 0 &&
3487 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) ||
3488 ((int)bp->b_loffset & PAGE_MASK) != 0 ||
3489 (bp->b_bcount & PAGE_MASK) != 0) {
3494 * Figure out the original VM object (it will match the underlying
3495 * VM pages). Note that swap cached data uses page indices relative
3496 * to that object, not relative to bio->bio_offset.
3498 if (bp->b_flags & B_CLUSTER)
3499 object = vp->v_object;
3501 object = bp->b_vp->v_object;
3504 * In order to be able to use the swap cache all underlying VM
3505 * pages must be marked as such, and we can't have any bogus pages.
3507 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
3508 m = bp->b_xio.xio_pages[i];
3509 if ((m->flags & PG_SWAPPED) == 0)
3511 if (m == bogus_page)
3516 * If we are good then issue the I/O using swap_pager_strategy()
3518 if (i == bp->b_xio.xio_npages) {
3519 m = bp->b_xio.xio_pages[0];
3520 nbio = push_bio(bio);
3521 nbio->bio_offset = ptoa(m->pindex);
3522 KKASSERT(m->object == object);
3523 swap_pager_strategy(object, nbio);
3532 * Finish I/O on a buffer after all BIOs have been processed.
3533 * Called when the bio chain is exhausted or by biowait. If called
3534 * by biowait, elseit is typically 0.
3536 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3537 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3538 * assuming B_INVAL is clear.
3540 * For the VMIO case, we set B_CACHE if the op was a read and no
3541 * read error occured, or if the op was a write. B_CACHE is never
3542 * set if the buffer is invalid or otherwise uncacheable.
3544 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3545 * initiator to leave B_INVAL set to brelse the buffer out of existance
3546 * in the biodone routine.
3549 bpdone(struct buf *bp, int elseit)
3553 KASSERT(BUF_REFCNTNB(bp) > 0,
3554 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3555 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3556 ("biodone: bp %p already done!", bp));
3559 * No more BIOs are left. All completion functions have been dealt
3560 * with, now we clean up the buffer.
3563 bp->b_cmd = BUF_CMD_DONE;
3566 * Only reads and writes are processed past this point.
3568 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3569 if (cmd == BUF_CMD_FREEBLKS)
3570 bp->b_flags |= B_NOCACHE;
3577 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3578 * a lot worse. XXX - move this above the clearing of b_cmd
3580 if (LIST_FIRST(&bp->b_dep) != NULL)
3584 * A failed write must re-dirty the buffer unless B_INVAL
3585 * was set. Only applicable to normal buffers (with VPs).
3586 * vinum buffers may not have a vp.
3588 if (cmd == BUF_CMD_WRITE &&
3589 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3590 bp->b_flags &= ~B_NOCACHE;
3595 if (bp->b_flags & B_VMIO) {
3601 struct vnode *vp = bp->b_vp;
3605 #if defined(VFS_BIO_DEBUG)
3606 if (vp->v_auxrefs == 0)
3607 panic("biodone: zero vnode hold count");
3608 if ((vp->v_flag & VOBJBUF) == 0)
3609 panic("biodone: vnode is not setup for merged cache");
3612 foff = bp->b_loffset;
3613 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3614 KASSERT(obj != NULL, ("biodone: missing VM object"));
3616 #if defined(VFS_BIO_DEBUG)
3617 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3618 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3619 obj->paging_in_progress, bp->b_xio.xio_npages);
3624 * Set B_CACHE if the op was a normal read and no error
3625 * occured. B_CACHE is set for writes in the b*write()
3628 iosize = bp->b_bcount - bp->b_resid;
3629 if (cmd == BUF_CMD_READ &&
3630 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3631 bp->b_flags |= B_CACHE;
3636 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3640 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3645 * cleanup bogus pages, restoring the originals. Since
3646 * the originals should still be wired, we don't have
3647 * to worry about interrupt/freeing races destroying
3648 * the VM object association.
3650 m = bp->b_xio.xio_pages[i];
3651 if (m == bogus_page) {
3653 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3655 panic("biodone: page disappeared");
3656 bp->b_xio.xio_pages[i] = m;
3657 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3658 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3660 #if defined(VFS_BIO_DEBUG)
3661 if (OFF_TO_IDX(foff) != m->pindex) {
3662 kprintf("biodone: foff(%lu)/m->pindex(%ld) "
3664 (unsigned long)foff, (long)m->pindex);
3669 * In the write case, the valid and clean bits are
3670 * already changed correctly (see bdwrite()), so we
3671 * only need to do this here in the read case.
3673 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3674 vfs_clean_one_page(bp, i, m);
3676 vm_page_flag_clear(m, PG_ZERO);
3679 * when debugging new filesystems or buffer I/O
3680 * methods, this is the most common error that pops
3681 * up. if you see this, you have not set the page
3682 * busy flag correctly!!!
3685 kprintf("biodone: page busy < 0, "
3686 "pindex: %d, foff: 0x(%x,%x), "
3687 "resid: %d, index: %d\n",
3688 (int) m->pindex, (int)(foff >> 32),
3689 (int) foff & 0xffffffff, resid, i);
3690 if (!vn_isdisk(vp, NULL))
3691 kprintf(" iosize: %ld, loffset: %lld, "
3692 "flags: 0x%08x, npages: %d\n",
3693 bp->b_vp->v_mount->mnt_stat.f_iosize,
3694 (long long)bp->b_loffset,
3695 bp->b_flags, bp->b_xio.xio_npages);
3697 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3698 (long long)bp->b_loffset,
3699 bp->b_flags, bp->b_xio.xio_npages);
3700 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3701 m->valid, m->dirty, m->wire_count);
3702 panic("biodone: page busy < 0");
3704 vm_page_io_finish(m);
3705 vm_object_pip_subtract(obj, 1);
3706 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3709 bp->b_flags &= ~B_HASBOGUS;
3711 vm_object_pip_wakeupn(obj, 0);
3717 * Finish up by releasing the buffer. There are no more synchronous
3718 * or asynchronous completions, those were handled by bio_done
3722 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
3733 biodone(struct bio *bio)
3735 struct buf *bp = bio->bio_buf;
3737 runningbufwakeup(bp);
3740 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3743 biodone_t *done_func;
3744 struct bio_track *track;
3747 * BIO tracking. Most but not all BIOs are tracked.
3749 if ((track = bio->bio_track) != NULL) {
3750 bio_track_rel(track);
3751 bio->bio_track = NULL;
3755 * A bio_done function terminates the loop. The function
3756 * will be responsible for any further chaining and/or
3757 * buffer management.
3759 * WARNING! The done function can deallocate the buffer!
3761 if ((done_func = bio->bio_done) != NULL) {
3762 bio->bio_done = NULL;
3766 bio = bio->bio_prev;
3770 * If we've run out of bio's do normal [a]synchronous completion.
3776 * Synchronous biodone - this terminates a synchronous BIO.
3778 * bpdone() is called with elseit=FALSE, leaving the buffer completed
3779 * but still locked. The caller must brelse() the buffer after waiting
3783 biodone_sync(struct bio *bio)
3785 struct buf *bp = bio->bio_buf;
3789 KKASSERT(bio == &bp->b_bio1);
3793 flags = bio->bio_flags;
3794 nflags = (flags | BIO_DONE) & ~BIO_WANT;
3796 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3797 if (flags & BIO_WANT)
3807 * This routine is called in lieu of iodone in the case of
3808 * incomplete I/O. This keeps the busy status for pages
3812 vfs_unbusy_pages(struct buf *bp)
3816 runningbufwakeup(bp);
3817 if (bp->b_flags & B_VMIO) {
3818 struct vnode *vp = bp->b_vp;
3823 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3824 vm_page_t m = bp->b_xio.xio_pages[i];
3827 * When restoring bogus changes the original pages
3828 * should still be wired, so we are in no danger of
3829 * losing the object association and do not need
3830 * critical section protection particularly.
3832 if (m == bogus_page) {
3833 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3835 panic("vfs_unbusy_pages: page missing");
3837 bp->b_xio.xio_pages[i] = m;
3838 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3839 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3841 vm_object_pip_subtract(obj, 1);
3842 vm_page_flag_clear(m, PG_ZERO);
3843 vm_page_io_finish(m);
3845 bp->b_flags &= ~B_HASBOGUS;
3846 vm_object_pip_wakeupn(obj, 0);
3853 * This routine is called before a device strategy routine.
3854 * It is used to tell the VM system that paging I/O is in
3855 * progress, and treat the pages associated with the buffer
3856 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3857 * flag is handled to make sure that the object doesn't become
3860 * Since I/O has not been initiated yet, certain buffer flags
3861 * such as B_ERROR or B_INVAL may be in an inconsistant state
3862 * and should be ignored.
3865 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3868 struct lwp *lp = curthread->td_lwp;
3871 * The buffer's I/O command must already be set. If reading,
3872 * B_CACHE must be 0 (double check against callers only doing
3873 * I/O when B_CACHE is 0).
3875 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3876 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3878 if (bp->b_flags & B_VMIO) {
3882 KASSERT(bp->b_loffset != NOOFFSET,
3883 ("vfs_busy_pages: no buffer offset"));
3886 * Loop until none of the pages are busy.
3889 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3890 vm_page_t m = bp->b_xio.xio_pages[i];
3892 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3897 * Setup for I/O, soft-busy the page right now because
3898 * the next loop may block.
3900 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3901 vm_page_t m = bp->b_xio.xio_pages[i];
3903 vm_page_flag_clear(m, PG_ZERO);
3904 if ((bp->b_flags & B_CLUSTER) == 0) {
3905 vm_object_pip_add(obj, 1);
3906 vm_page_io_start(m);
3911 * Adjust protections for I/O and do bogus-page mapping.
3912 * Assume that vm_page_protect() can block (it can block
3913 * if VM_PROT_NONE, don't take any chances regardless).
3915 * In particular note that for writes we must incorporate
3916 * page dirtyness from the VM system into the buffer's
3919 * For reads we theoretically must incorporate page dirtyness
3920 * from the VM system to determine if the page needs bogus
3921 * replacement, but we shortcut the test by simply checking
3922 * that all m->valid bits are set, indicating that the page
3923 * is fully valid and does not need to be re-read. For any
3924 * VM system dirtyness the page will also be fully valid
3925 * since it was mapped at one point.
3928 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3929 vm_page_t m = bp->b_xio.xio_pages[i];
3931 vm_page_flag_clear(m, PG_ZERO); /* XXX */
3932 if (bp->b_cmd == BUF_CMD_WRITE) {
3934 * When readying a vnode-backed buffer for
3935 * a write we must zero-fill any invalid
3936 * portions of the backing VM pages, mark
3937 * it valid and clear related dirty bits.
3939 * vfs_clean_one_page() incorporates any
3940 * VM dirtyness and updates the b_dirtyoff
3941 * range (after we've made the page RO).
3943 * It is also expected that the pmap modified
3944 * bit has already been cleared by the
3945 * vm_page_protect(). We may not be able
3946 * to clear all dirty bits for a page if it
3947 * was also memory mapped (NFS).
3949 * Finally be sure to unassign any swap-cache
3950 * backing store as it is now stale.
3952 vm_page_protect(m, VM_PROT_READ);
3953 vfs_clean_one_page(bp, i, m);
3954 swap_pager_unswapped(m);
3955 } else if (m->valid == VM_PAGE_BITS_ALL) {
3957 * When readying a vnode-backed buffer for
3958 * read we must replace any dirty pages with
3959 * a bogus page so dirty data is not destroyed
3960 * when filling gaps.
3962 * To avoid testing whether the page is
3963 * dirty we instead test that the page was
3964 * at some point mapped (m->valid fully
3965 * valid) with the understanding that
3966 * this also covers the dirty case.
3968 bp->b_xio.xio_pages[i] = bogus_page;
3969 bp->b_flags |= B_HASBOGUS;
3971 } else if (m->valid & m->dirty) {
3973 * This case should not occur as partial
3974 * dirtyment can only happen if the buffer
3975 * is B_CACHE, and this code is not entered
3976 * if the buffer is B_CACHE.
3978 kprintf("Warning: vfs_busy_pages - page not "
3979 "fully valid! loff=%jx bpf=%08x "
3980 "idx=%d val=%02x dir=%02x\n",
3981 (intmax_t)bp->b_loffset, bp->b_flags,
3982 i, m->valid, m->dirty);
3983 vm_page_protect(m, VM_PROT_NONE);
3986 * The page is not valid and can be made
3989 vm_page_protect(m, VM_PROT_NONE);
3993 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3994 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3999 * This is the easiest place to put the process accounting for the I/O
4003 if (bp->b_cmd == BUF_CMD_READ)
4004 lp->lwp_ru.ru_inblock++;
4006 lp->lwp_ru.ru_oublock++;
4013 * Tell the VM system that the pages associated with this buffer
4014 * are clean. This is used for delayed writes where the data is
4015 * going to go to disk eventually without additional VM intevention.
4017 * Note that while we only really need to clean through to b_bcount, we
4018 * just go ahead and clean through to b_bufsize.
4021 vfs_clean_pages(struct buf *bp)
4026 if ((bp->b_flags & B_VMIO) == 0)
4029 KASSERT(bp->b_loffset != NOOFFSET,
4030 ("vfs_clean_pages: no buffer offset"));
4032 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4033 m = bp->b_xio.xio_pages[i];
4034 vfs_clean_one_page(bp, i, m);
4039 * vfs_clean_one_page:
4041 * Set the valid bits and clear the dirty bits in a page within a
4042 * buffer. The range is restricted to the buffer's size and the
4043 * buffer's logical offset might index into the first page.
4045 * The caller has busied or soft-busied the page and it is not mapped,
4046 * test and incorporate the dirty bits into b_dirtyoff/end before
4047 * clearing them. Note that we need to clear the pmap modified bits
4048 * after determining the the page was dirty, vm_page_set_validclean()
4049 * does not do it for us.
4051 * This routine is typically called after a read completes (dirty should
4052 * be zero in that case as we are not called on bogus-replace pages),
4053 * or before a write is initiated.
4056 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m)
4064 * Calculate offset range within the page but relative to buffer's
4065 * loffset. loffset might be offset into the first page.
4067 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4068 bcount = bp->b_bcount + xoff; /* offset adjusted */
4074 soff = (pageno << PAGE_SHIFT);
4075 eoff = soff + PAGE_SIZE;
4083 * Test dirty bits and adjust b_dirtyoff/end.
4085 * If dirty pages are incorporated into the bp any prior
4086 * B_NEEDCOMMIT state (NFS) must be cleared because the
4087 * caller has not taken into account the new dirty data.
4089 * If the page was memory mapped the dirty bits might go beyond the
4090 * end of the buffer, but we can't really make the assumption that
4091 * a file EOF straddles the buffer (even though this is the case for
4092 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
4093 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
4094 * This also saves some console spam.
4096 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK,
4097 * NFS can handle huge commits but not huge writes.
4099 vm_page_test_dirty(m);
4101 if ((bp->b_flags & B_NEEDCOMMIT) &&
4102 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) {
4104 kprintf("Warning: vfs_clean_one_page: bp %p "
4105 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT"
4106 " cmd %d vd %02x/%02x x/s/e %d %d %d "
4108 bp, (intmax_t)bp->b_loffset, bp->b_bcount,
4109 bp->b_flags, bp->b_cmd,
4110 m->valid, m->dirty, xoff, soff, eoff,
4111 bp->b_dirtyoff, bp->b_dirtyend);
4112 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
4114 print_backtrace(-1);
4117 * Only clear the pmap modified bits if ALL the dirty bits
4118 * are set, otherwise the system might mis-clear portions
4121 if (m->dirty == VM_PAGE_BITS_ALL &&
4122 (bp->b_flags & B_NEEDCOMMIT) == 0) {
4123 pmap_clear_modify(m);
4125 if (bp->b_dirtyoff > soff - xoff)
4126 bp->b_dirtyoff = soff - xoff;
4127 if (bp->b_dirtyend < eoff - xoff)
4128 bp->b_dirtyend = eoff - xoff;
4132 * Set related valid bits, clear related dirty bits.
4133 * Does not mess with the pmap modified bit.
4135 * WARNING! We cannot just clear all of m->dirty here as the
4136 * buffer cache buffers may use a DEV_BSIZE'd aligned
4137 * block size, or have an odd size (e.g. NFS at file EOF).
4138 * The putpages code can clear m->dirty to 0.
4140 * If a VOP_WRITE generates a buffer cache buffer which
4141 * covers the same space as mapped writable pages the
4142 * buffer flush might not be able to clear all the dirty
4143 * bits and still require a putpages from the VM system
4146 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff);
4150 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty.
4151 * The page data is assumed to be valid (there is no zeroing here).
4154 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m)
4162 * Calculate offset range within the page but relative to buffer's
4163 * loffset. loffset might be offset into the first page.
4165 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4166 bcount = bp->b_bcount + xoff; /* offset adjusted */
4172 soff = (pageno << PAGE_SHIFT);
4173 eoff = soff + PAGE_SIZE;
4179 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff);
4185 * Clear a buffer. This routine essentially fakes an I/O, so we need
4186 * to clear B_ERROR and B_INVAL.
4188 * Note that while we only theoretically need to clear through b_bcount,
4189 * we go ahead and clear through b_bufsize.
4193 vfs_bio_clrbuf(struct buf *bp)
4197 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
4198 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
4199 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4200 (bp->b_loffset & PAGE_MASK) == 0) {
4201 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4202 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
4206 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
4207 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
4208 bzero(bp->b_data, bp->b_bufsize);
4209 bp->b_xio.xio_pages[0]->valid |= mask;
4215 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
4216 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
4217 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
4218 ea = (caddr_t)(vm_offset_t)ulmin(
4219 (u_long)(vm_offset_t)ea,
4220 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
4221 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4222 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
4224 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
4225 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
4229 for (; sa < ea; sa += DEV_BSIZE, j++) {
4230 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
4231 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
4232 bzero(sa, DEV_BSIZE);
4235 bp->b_xio.xio_pages[i]->valid |= mask;
4236 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
4245 * vm_hold_load_pages:
4247 * Load pages into the buffer's address space. The pages are
4248 * allocated from the kernel object in order to reduce interference
4249 * with the any VM paging I/O activity. The range of loaded
4250 * pages will be wired.
4252 * If a page cannot be allocated, the 'pagedaemon' is woken up to
4253 * retrieve the full range (to - from) of pages.
4257 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4263 to = round_page(to);
4264 from = round_page(from);
4265 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4270 * Note: must allocate system pages since blocking here
4271 * could intefere with paging I/O, no matter which
4274 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT,
4275 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
4278 p->valid = VM_PAGE_BITS_ALL;
4279 vm_page_flag_clear(p, PG_ZERO);
4280 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
4281 bp->b_xio.xio_pages[index] = p;
4288 bp->b_xio.xio_npages = index;
4292 * Allocate pages for a buffer cache buffer.
4294 * Under extremely severe memory conditions even allocating out of the
4295 * system reserve can fail. If this occurs we must allocate out of the
4296 * interrupt reserve to avoid a deadlock with the pageout daemon.
4298 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
4299 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
4300 * against the pageout daemon if pages are not freed from other sources.
4304 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit)
4309 * Try a normal allocation, allow use of system reserve.
4311 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
4316 * The normal allocation failed and we clearly have a page
4317 * deficit. Try to reclaim some clean VM pages directly
4318 * from the buffer cache.
4320 vm_pageout_deficit += deficit;
4324 * We may have blocked, the caller will know what to do if the
4327 if (vm_page_lookup(obj, pg))
4331 * Allocate and allow use of the interrupt reserve.
4333 * If after all that we still can't allocate a VM page we are
4334 * in real trouble, but we slog on anyway hoping that the system
4337 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
4338 VM_ALLOC_INTERRUPT);
4340 if (vm_page_count_severe()) {
4341 kprintf("bio_page_alloc: WARNING emergency page "
4346 kprintf("bio_page_alloc: WARNING emergency page "
4347 "allocation failed\n");
4354 * vm_hold_free_pages:
4356 * Return pages associated with the buffer back to the VM system.
4358 * The range of pages underlying the buffer's address space will
4359 * be unmapped and un-wired.
4362 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4366 int index, newnpages;
4368 from = round_page(from);
4369 to = round_page(to);
4370 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4373 lwkt_gettoken(&vm_token);
4374 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4375 p = bp->b_xio.xio_pages[index];
4376 if (p && (index < bp->b_xio.xio_npages)) {
4378 kprintf("vm_hold_free_pages: doffset: %lld, "
4380 (long long)bp->b_bio2.bio_offset,
4381 (long long)bp->b_loffset);
4383 bp->b_xio.xio_pages[index] = NULL;
4386 vm_page_unwire(p, 0);
4390 bp->b_xio.xio_npages = newnpages;
4391 lwkt_reltoken(&vm_token);
4397 * Map a user buffer into KVM via a pbuf. On return the buffer's
4398 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4402 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4413 * bp had better have a command and it better be a pbuf.
4415 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4416 KKASSERT(bp->b_flags & B_PAGING);
4417 KKASSERT(bp->b_kvabase);
4423 * Map the user data into KVM. Mappings have to be page-aligned.
4425 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4428 vmprot = VM_PROT_READ;
4429 if (bp->b_cmd == BUF_CMD_READ)
4430 vmprot |= VM_PROT_WRITE;
4432 while (addr < udata + bytes) {
4434 * Do the vm_fault if needed; do the copy-on-write thing
4435 * when reading stuff off device into memory.
4437 * vm_fault_page*() returns a held VM page.
4439 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4440 va = trunc_page(va);
4442 m = vm_fault_page_quick(va, vmprot, &error);
4444 for (i = 0; i < pidx; ++i) {
4445 vm_page_unhold(bp->b_xio.xio_pages[i]);
4446 bp->b_xio.xio_pages[i] = NULL;
4450 bp->b_xio.xio_pages[pidx] = m;
4456 * Map the page array and set the buffer fields to point to
4457 * the mapped data buffer.
4459 if (pidx > btoc(MAXPHYS))
4460 panic("vmapbuf: mapped more than MAXPHYS");
4461 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4463 bp->b_xio.xio_npages = pidx;
4464 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4465 bp->b_bcount = bytes;
4466 bp->b_bufsize = bytes;
4473 * Free the io map PTEs associated with this IO operation.
4474 * We also invalidate the TLB entries and restore the original b_addr.
4477 vunmapbuf(struct buf *bp)
4482 KKASSERT(bp->b_flags & B_PAGING);
4484 npages = bp->b_xio.xio_npages;
4485 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4486 for (pidx = 0; pidx < npages; ++pidx) {
4487 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4488 bp->b_xio.xio_pages[pidx] = NULL;
4490 bp->b_xio.xio_npages = 0;
4491 bp->b_data = bp->b_kvabase;
4495 * Scan all buffers in the system and issue the callback.
4498 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4504 for (n = 0; n < nbuf; ++n) {
4505 if ((error = callback(&buf[n], info)) < 0) {
4515 * nestiobuf_iodone: biodone callback for nested buffers and propagate
4516 * completion to the master buffer.
4519 nestiobuf_iodone(struct bio *bio)
4522 struct buf *mbp, *bp;
4527 mbio = bio->bio_caller_info1.ptr;
4528 mbp = mbio->bio_buf;
4530 KKASSERT(bp->b_bcount <= bp->b_bufsize);
4531 KKASSERT(mbp != bp);
4533 error = bp->b_error;
4534 if (bp->b_error == 0 &&
4535 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
4537 * Not all got transfered, raise an error. We have no way to
4538 * propagate these conditions to mbp.
4543 donebytes = bp->b_bufsize;
4546 nestiobuf_done(mbio, donebytes, error);
4550 nestiobuf_done(struct bio *mbio, int donebytes, int error)
4554 mbp = mbio->bio_buf;
4556 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0);
4559 * If an error occured, propagate it to the master buffer.
4561 * Several biodone()s may wind up running concurrently so
4562 * use an atomic op to adjust b_flags.
4565 mbp->b_error = error;
4566 atomic_set_int(&mbp->b_flags, B_ERROR);
4570 * Decrement the master buf b_resid according to our donebytes, and
4571 * also check if this is the last missing bit for the whole nestio
4572 * mess to complete. If so, call biodone() on the master buf mbp.
4574 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4581 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent
4582 * the mbio from being biodone()'d while we are still adding sub-bios to
4586 nestiobuf_init(struct bio *bio)
4588 bio->bio_driver_info = (void *)1;
4592 * The BIOs added to the nestedio have already been started, remove the
4593 * count that placeheld our mbio and biodone() it if the count would
4597 nestiobuf_start(struct bio *mbio)
4599 struct buf *mbp = mbio->bio_buf;
4602 * Decrement the master buf b_resid according to our donebytes, and
4603 * also check if this is the last missing bit for the whole nestio
4604 * mess to complete. If so, call biodone() on the master buf mbp.
4606 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4607 if (mbp->b_flags & B_ERROR)
4608 mbp->b_resid = mbp->b_bcount;
4616 * Set an intermediate error prior to calling nestiobuf_start()
4619 nestiobuf_error(struct bio *mbio, int error)
4621 struct buf *mbp = mbio->bio_buf;
4624 mbp->b_error = error;
4625 atomic_set_int(&mbp->b_flags, B_ERROR);
4630 * nestiobuf_add: setup a "nested" buffer.
4632 * => 'mbp' is a "master" buffer which is being divided into sub pieces.
4633 * => 'bp' should be a buffer allocated by getiobuf.
4634 * => 'offset' is a byte offset in the master buffer.
4635 * => 'size' is a size in bytes of this nested buffer.
4638 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size)
4640 struct buf *mbp = mbio->bio_buf;
4641 struct vnode *vp = mbp->b_vp;
4643 KKASSERT(mbp->b_bcount >= offset + size);
4645 atomic_add_int((int *)&mbio->bio_driver_info, 1);
4647 /* kernel needs to own the lock for it to be released in biodone */
4650 bp->b_cmd = mbp->b_cmd;
4651 bp->b_bio1.bio_done = nestiobuf_iodone;
4652 bp->b_data = (char *)mbp->b_data + offset;
4653 bp->b_resid = bp->b_bcount = size;
4654 bp->b_bufsize = bp->b_bcount;
4656 bp->b_bio1.bio_track = NULL;
4657 bp->b_bio1.bio_caller_info1.ptr = mbio;
4661 * print out statistics from the current status of the buffer pool
4662 * this can be toggeled by the system control option debug.syncprt
4671 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
4672 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
4674 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
4676 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4679 TAILQ_FOREACH(bp, dp, b_freelist) {
4680 counts[bp->b_bufsize/PAGE_SIZE]++;
4684 kprintf("%s: total-%d", bname[i], count);
4685 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4687 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
4695 DB_SHOW_COMMAND(buffer, db_show_buffer)
4698 struct buf *bp = (struct buf *)addr;
4701 db_printf("usage: show buffer <addr>\n");
4705 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4706 db_printf("b_cmd = %d\n", bp->b_cmd);
4707 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4708 "b_resid = %d\n, b_data = %p, "
4709 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4710 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4712 (long long)bp->b_bio2.bio_offset,
4713 (long long)(bp->b_bio2.bio_next ?
4714 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4715 if (bp->b_xio.xio_npages) {
4717 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4718 bp->b_xio.xio_npages);
4719 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4721 m = bp->b_xio.xio_pages[i];
4722 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4723 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4724 if ((i + 1) < bp->b_xio.xio_npages)