2 * Copyright (c) 1994,1997 John S. Dyson
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice immediately at the beginning of the file, without modification,
10 * this list of conditions, and the following disclaimer.
11 * 2. Absolutely no warranty of function or purpose is made by the author
14 * $FreeBSD: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $
15 * $DragonFly: src/sys/kern/vfs_bio.c,v 1.115 2008/08/13 11:02:31 swildner Exp $
19 * this file contains a new buffer I/O scheme implementing a coherent
20 * VM object and buffer cache scheme. Pains have been taken to make
21 * sure that the performance degradation associated with schemes such
22 * as this is not realized.
24 * Author: John S. Dyson
25 * Significant help during the development and debugging phases
26 * had been provided by David Greenman, also of the FreeBSD core team.
28 * see man buf(9) for more info.
31 #include <sys/param.h>
32 #include <sys/systm.h>
35 #include <sys/eventhandler.h>
37 #include <sys/malloc.h>
38 #include <sys/mount.h>
39 #include <sys/kernel.h>
40 #include <sys/kthread.h>
42 #include <sys/reboot.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sysctl.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_kern.h>
51 #include <vm/vm_pageout.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_map.h>
58 #include <sys/thread2.h>
59 #include <sys/spinlock2.h>
60 #include <vm/vm_page2.h>
71 BQUEUE_NONE, /* not on any queue */
72 BQUEUE_LOCKED, /* locked buffers */
73 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
74 BQUEUE_DIRTY, /* B_DELWRI buffers */
75 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
76 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
77 BQUEUE_EMPTY, /* empty buffer headers */
79 BUFFER_QUEUES /* number of buffer queues */
82 typedef enum bufq_type bufq_type_t;
84 #define BD_WAKE_SIZE 16384
85 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
87 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
88 struct spinlock bufspin = SPINLOCK_INITIALIZER(&bufspin);
90 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
92 struct buf *buf; /* buffer header pool */
94 static void vfs_clean_pages(struct buf *bp);
95 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
96 static void vfs_vmio_release(struct buf *bp);
97 static int flushbufqueues(bufq_type_t q);
98 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);
100 static void bd_signal(int totalspace);
101 static void buf_daemon(void);
102 static void buf_daemon_hw(void);
105 * bogus page -- for I/O to/from partially complete buffers
106 * this is a temporary solution to the problem, but it is not
107 * really that bad. it would be better to split the buffer
108 * for input in the case of buffers partially already in memory,
109 * but the code is intricate enough already.
111 vm_page_t bogus_page;
114 * These are all static, but make the ones we export globals so we do
115 * not need to use compiler magic.
117 int bufspace, maxbufspace,
118 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
119 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
120 static int lorunningspace, hirunningspace, runningbufreq;
121 int dirtybufspace, dirtybufspacehw, lodirtybufspace, hidirtybufspace;
122 int dirtybufcount, dirtybufcounthw;
123 int runningbufspace, runningbufcount;
124 static int getnewbufcalls;
125 static int getnewbufrestarts;
126 static int recoverbufcalls;
127 static int needsbuffer; /* locked by needsbuffer_spin */
128 static int bd_request; /* locked by needsbuffer_spin */
129 static int bd_request_hw; /* locked by needsbuffer_spin */
130 static u_int bd_wake_ary[BD_WAKE_SIZE];
131 static u_int bd_wake_index;
132 static u_int vm_cycle_point = ACT_INIT + ACT_ADVANCE * 6;
133 static struct spinlock needsbuffer_spin;
135 static struct thread *bufdaemon_td;
136 static struct thread *bufdaemonhw_td;
140 * Sysctls for operational control of the buffer cache.
142 SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
143 "Number of dirty buffers to flush before bufdaemon becomes inactive");
144 SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
145 "High watermark used to trigger explicit flushing of dirty buffers");
146 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
147 "Minimum amount of buffer space required for active I/O");
148 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
149 "Maximum amount of buffer space to usable for active I/O");
150 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
151 "Recycle pages to active or inactive queue transition pt 0-64");
153 * Sysctls determining current state of the buffer cache.
155 SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
156 "Total number of buffers in buffer cache");
157 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
158 "Pending bytes of dirty buffers (all)");
159 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
160 "Pending bytes of dirty buffers (heavy weight)");
161 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
162 "Pending number of dirty buffers");
163 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
164 "Pending number of dirty buffers (heavy weight)");
165 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
166 "I/O bytes currently in progress due to asynchronous writes");
167 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
168 "I/O buffers currently in progress due to asynchronous writes");
169 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
170 "Hard limit on maximum amount of memory usable for buffer space");
171 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
172 "Soft limit on maximum amount of memory usable for buffer space");
173 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
174 "Minimum amount of memory to reserve for system buffer space");
175 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
176 "Amount of memory available for buffers");
177 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
178 0, "Maximum amount of memory reserved for buffers using malloc");
179 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
180 "Amount of memory left for buffers using malloc-scheme");
181 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
182 "New buffer header acquisition requests");
183 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
184 0, "New buffer header acquisition restarts");
185 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
186 "Recover VM space in an emergency");
187 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
188 "Buffer acquisition restarts due to fragmented buffer map");
189 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
190 "Amount of time KVA space was deallocated in an arbitrary buffer");
191 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
192 "Amount of time buffer re-use operations were successful");
193 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
194 "sizeof(struct buf)");
196 char *buf_wmesg = BUF_WMESG;
198 extern int vm_swap_size;
200 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
201 #define VFS_BIO_NEED_UNUSED02 0x02
202 #define VFS_BIO_NEED_UNUSED04 0x04
203 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
208 * Called when buffer space is potentially available for recovery.
209 * getnewbuf() will block on this flag when it is unable to free
210 * sufficient buffer space. Buffer space becomes recoverable when
211 * bp's get placed back in the queues.
218 * If someone is waiting for BUF space, wake them up. Even
219 * though we haven't freed the kva space yet, the waiting
220 * process will be able to now.
222 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
223 spin_lock_wr(&needsbuffer_spin);
224 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
225 spin_unlock_wr(&needsbuffer_spin);
226 wakeup(&needsbuffer);
233 * Accounting for I/O in progress.
237 runningbufwakeup(struct buf *bp)
242 if ((totalspace = bp->b_runningbufspace) != 0) {
243 atomic_subtract_int(&runningbufspace, totalspace);
244 atomic_subtract_int(&runningbufcount, 1);
245 bp->b_runningbufspace = 0;
248 * see waitrunningbufspace() for limit test.
250 limit = hirunningspace * 2 / 3;
251 if (runningbufreq && runningbufspace <= limit) {
253 wakeup(&runningbufreq);
255 bd_signal(totalspace);
262 * Called when a buffer has been added to one of the free queues to
263 * account for the buffer and to wakeup anyone waiting for free buffers.
264 * This typically occurs when large amounts of metadata are being handled
265 * by the buffer cache ( else buffer space runs out first, usually ).
273 spin_lock_wr(&needsbuffer_spin);
274 needsbuffer &= ~VFS_BIO_NEED_ANY;
275 spin_unlock_wr(&needsbuffer_spin);
276 wakeup(&needsbuffer);
281 * waitrunningbufspace()
283 * Wait for the amount of running I/O to drop to hirunningspace * 2 / 3.
284 * This is the point where write bursting stops so we don't want to wait
285 * for the running amount to drop below it (at least if we still want bioq
288 * The caller may be using this function to block in a tight loop, we
289 * must block while runningbufspace is greater then or equal to
290 * hirunningspace * 2 / 3.
292 * And even with that it may not be enough, due to the presence of
293 * B_LOCKED dirty buffers, so also wait for at least one running buffer
297 waitrunningbufspace(void)
299 int limit = hirunningspace * 2 / 3;
302 if (runningbufspace > limit) {
303 while (runningbufspace > limit) {
305 tsleep(&runningbufreq, 0, "wdrn1", 0);
307 } else if (runningbufspace) {
309 tsleep(&runningbufreq, 0, "wdrn2", 1);
315 * buf_dirty_count_severe:
317 * Return true if we have too many dirty buffers.
320 buf_dirty_count_severe(void)
322 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
323 dirtybufcount >= nbuf / 2);
327 * Return true if the amount of running I/O is severe and BIOQ should
331 buf_runningbufspace_severe(void)
333 return (runningbufspace >= hirunningspace * 2 / 3);
337 * vfs_buf_test_cache:
339 * Called when a buffer is extended. This function clears the B_CACHE
340 * bit if the newly extended portion of the buffer does not contain
343 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
344 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
345 * them while a clean buffer was present.
349 vfs_buf_test_cache(struct buf *bp,
350 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
353 if (bp->b_flags & B_CACHE) {
354 int base = (foff + off) & PAGE_MASK;
355 if (vm_page_is_valid(m, base, size) == 0)
356 bp->b_flags &= ~B_CACHE;
363 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
372 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
375 if (bd_request == 0 &&
376 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
377 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
378 spin_lock_wr(&needsbuffer_spin);
380 spin_unlock_wr(&needsbuffer_spin);
383 if (bd_request_hw == 0 &&
384 (dirtybufspacehw > lodirtybufspace / 2 ||
385 dirtybufcounthw >= nbuf / 2)) {
386 spin_lock_wr(&needsbuffer_spin);
388 spin_unlock_wr(&needsbuffer_spin);
389 wakeup(&bd_request_hw);
396 * Get the buf_daemon heated up when the number of running and dirty
397 * buffers exceeds the mid-point.
399 * Return the total number of dirty bytes past the second mid point
400 * as a measure of how much excess dirty data there is in the system.
411 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
413 totalspace = runningbufspace + dirtybufspace;
414 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
416 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
417 if (totalspace >= mid2)
418 return(totalspace - mid2);
426 * Wait for the buffer cache to flush (totalspace) bytes worth of
427 * buffers, then return.
429 * Regardless this function blocks while the number of dirty buffers
430 * exceeds hidirtybufspace.
435 bd_wait(int totalspace)
440 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
443 while (totalspace > 0) {
445 if (totalspace > runningbufspace + dirtybufspace)
446 totalspace = runningbufspace + dirtybufspace;
447 count = totalspace / BKVASIZE;
448 if (count >= BD_WAKE_SIZE)
449 count = BD_WAKE_SIZE - 1;
451 spin_lock_wr(&needsbuffer_spin);
452 i = (bd_wake_index + count) & BD_WAKE_MASK;
454 tsleep_interlock(&bd_wake_ary[i], 0);
455 spin_unlock_wr(&needsbuffer_spin);
456 tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz);
458 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
465 * This function is called whenever runningbufspace or dirtybufspace
466 * is reduced. Track threads waiting for run+dirty buffer I/O
472 bd_signal(int totalspace)
476 if (totalspace > 0) {
477 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
478 totalspace = BKVASIZE * BD_WAKE_SIZE;
479 spin_lock_wr(&needsbuffer_spin);
480 while (totalspace > 0) {
483 if (bd_wake_ary[i]) {
485 spin_unlock_wr(&needsbuffer_spin);
486 wakeup(&bd_wake_ary[i]);
487 spin_lock_wr(&needsbuffer_spin);
489 totalspace -= BKVASIZE;
491 spin_unlock_wr(&needsbuffer_spin);
496 * BIO tracking support routines.
498 * Release a ref on a bio_track. Wakeup requests are atomically released
499 * along with the last reference so bk_active will never wind up set to
506 bio_track_rel(struct bio_track *track)
514 active = track->bk_active;
515 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
519 * Full-on. Note that the wait flag is only atomically released on
520 * the 1->0 count transition.
522 * We check for a negative count transition using bit 30 since bit 31
523 * has a different meaning.
526 desired = (active & 0x7FFFFFFF) - 1;
528 desired |= active & 0x80000000;
529 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
530 if (desired & 0x40000000)
531 panic("bio_track_rel: bad count: %p\n", track);
532 if (active & 0x80000000)
536 active = track->bk_active;
541 * Wait for the tracking count to reach 0.
543 * Use atomic ops such that the wait flag is only set atomically when
544 * bk_active is non-zero.
549 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
558 if (track->bk_active == 0)
562 * Full-on. Note that the wait flag may only be atomically set if
563 * the active count is non-zero.
566 while ((active = track->bk_active) != 0) {
567 desired = active | 0x80000000;
568 tsleep_interlock(track, slp_flags);
569 if (active == desired ||
570 atomic_cmpset_int(&track->bk_active, active, desired)) {
571 error = tsleep(track, slp_flags | PINTERLOCKED,
583 * Load time initialisation of the buffer cache, called from machine
584 * dependant initialization code.
590 vm_offset_t bogus_offset;
593 spin_init(&needsbuffer_spin);
595 /* next, make a null set of free lists */
596 for (i = 0; i < BUFFER_QUEUES; i++)
597 TAILQ_INIT(&bufqueues[i]);
599 /* finally, initialize each buffer header and stick on empty q */
600 for (i = 0; i < nbuf; i++) {
602 bzero(bp, sizeof *bp);
603 bp->b_flags = B_INVAL; /* we're just an empty header */
604 bp->b_cmd = BUF_CMD_DONE;
605 bp->b_qindex = BQUEUE_EMPTY;
607 xio_init(&bp->b_xio);
610 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
614 * maxbufspace is the absolute maximum amount of buffer space we are
615 * allowed to reserve in KVM and in real terms. The absolute maximum
616 * is nominally used by buf_daemon. hibufspace is the nominal maximum
617 * used by most other processes. The differential is required to
618 * ensure that buf_daemon is able to run when other processes might
619 * be blocked waiting for buffer space.
621 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
622 * this may result in KVM fragmentation which is not handled optimally
625 maxbufspace = nbuf * BKVASIZE;
626 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
627 lobufspace = hibufspace - MAXBSIZE;
629 lorunningspace = 512 * 1024;
630 /* hirunningspace -- see below */
633 * Limit the amount of malloc memory since it is wired permanently
634 * into the kernel space. Even though this is accounted for in
635 * the buffer allocation, we don't want the malloced region to grow
636 * uncontrolled. The malloc scheme improves memory utilization
637 * significantly on average (small) directories.
639 maxbufmallocspace = hibufspace / 20;
642 * Reduce the chance of a deadlock occuring by limiting the number
643 * of delayed-write dirty buffers we allow to stack up.
645 * We don't want too much actually queued to the device at once
646 * (XXX this needs to be per-mount!), because the buffers will
647 * wind up locked for a very long period of time while the I/O
650 hidirtybufspace = hibufspace / 2; /* dirty + running */
651 hirunningspace = hibufspace / 16; /* locked & queued to device */
652 if (hirunningspace < 1024 * 1024)
653 hirunningspace = 1024 * 1024;
658 lodirtybufspace = hidirtybufspace / 2;
661 * Maximum number of async ops initiated per buf_daemon loop. This is
662 * somewhat of a hack at the moment, we really need to limit ourselves
663 * based on the number of bytes of I/O in-transit that were initiated
667 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
668 bogus_page = vm_page_alloc(&kernel_object,
669 (bogus_offset >> PAGE_SHIFT),
671 vmstats.v_wire_count++;
676 * Initialize the embedded bio structures
679 initbufbio(struct buf *bp)
681 bp->b_bio1.bio_buf = bp;
682 bp->b_bio1.bio_prev = NULL;
683 bp->b_bio1.bio_offset = NOOFFSET;
684 bp->b_bio1.bio_next = &bp->b_bio2;
685 bp->b_bio1.bio_done = NULL;
686 bp->b_bio1.bio_flags = 0;
688 bp->b_bio2.bio_buf = bp;
689 bp->b_bio2.bio_prev = &bp->b_bio1;
690 bp->b_bio2.bio_offset = NOOFFSET;
691 bp->b_bio2.bio_next = NULL;
692 bp->b_bio2.bio_done = NULL;
693 bp->b_bio2.bio_flags = 0;
697 * Reinitialize the embedded bio structures as well as any additional
698 * translation cache layers.
701 reinitbufbio(struct buf *bp)
705 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
706 bio->bio_done = NULL;
707 bio->bio_offset = NOOFFSET;
712 * Push another BIO layer onto an existing BIO and return it. The new
713 * BIO layer may already exist, holding cached translation data.
716 push_bio(struct bio *bio)
720 if ((nbio = bio->bio_next) == NULL) {
721 int index = bio - &bio->bio_buf->b_bio_array[0];
722 if (index >= NBUF_BIO - 1) {
723 panic("push_bio: too many layers bp %p\n",
726 nbio = &bio->bio_buf->b_bio_array[index + 1];
727 bio->bio_next = nbio;
728 nbio->bio_prev = bio;
729 nbio->bio_buf = bio->bio_buf;
730 nbio->bio_offset = NOOFFSET;
731 nbio->bio_done = NULL;
732 nbio->bio_next = NULL;
734 KKASSERT(nbio->bio_done == NULL);
739 * Pop a BIO translation layer, returning the previous layer. The
740 * must have been previously pushed.
743 pop_bio(struct bio *bio)
745 return(bio->bio_prev);
749 clearbiocache(struct bio *bio)
752 bio->bio_offset = NOOFFSET;
760 * Free the KVA allocation for buffer 'bp'.
762 * Must be called from a critical section as this is the only locking for
765 * Since this call frees up buffer space, we call bufspacewakeup().
770 bfreekva(struct buf *bp)
777 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
778 vm_map_lock(&buffer_map);
779 bufspace -= bp->b_kvasize;
780 vm_map_delete(&buffer_map,
781 (vm_offset_t) bp->b_kvabase,
782 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
785 vm_map_unlock(&buffer_map);
786 vm_map_entry_release(count);
796 * Remove the buffer from the appropriate free list.
799 _bremfree(struct buf *bp)
801 if (bp->b_qindex != BQUEUE_NONE) {
802 KASSERT(BUF_REFCNTNB(bp) == 1,
803 ("bremfree: bp %p not locked",bp));
804 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
805 bp->b_qindex = BQUEUE_NONE;
807 if (BUF_REFCNTNB(bp) <= 1)
808 panic("bremfree: removing a buffer not on a queue");
813 bremfree(struct buf *bp)
815 spin_lock_wr(&bufspin);
817 spin_unlock_wr(&bufspin);
821 bremfree_locked(struct buf *bp)
829 * Get a buffer with the specified data. Look in the cache first. We
830 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
831 * is set, the buffer is valid and we do not have to do anything ( see
837 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
841 bp = getblk(vp, loffset, size, 0, 0);
844 /* if not found in cache, do some I/O */
845 if ((bp->b_flags & B_CACHE) == 0) {
847 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
848 bp->b_cmd = BUF_CMD_READ;
849 bp->b_bio1.bio_done = biodone_sync;
850 bp->b_bio1.bio_flags |= BIO_SYNC;
851 vfs_busy_pages(vp, bp);
852 vn_strategy(vp, &bp->b_bio1);
854 return (biowait(&bp->b_bio1, "biord"));
862 * Operates like bread, but also starts asynchronous I/O on
863 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
864 * to initiating I/O . If B_CACHE is set, the buffer is valid
865 * and we do not have to do anything.
870 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
871 int *rabsize, int cnt, struct buf **bpp)
873 struct buf *bp, *rabp;
875 int rv = 0, readwait = 0;
877 *bpp = bp = getblk(vp, loffset, size, 0, 0);
879 /* if not found in cache, do some I/O */
880 if ((bp->b_flags & B_CACHE) == 0) {
882 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
883 bp->b_cmd = BUF_CMD_READ;
884 bp->b_bio1.bio_done = biodone_sync;
885 bp->b_bio1.bio_flags |= BIO_SYNC;
886 vfs_busy_pages(vp, bp);
887 vn_strategy(vp, &bp->b_bio1);
892 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
893 if (inmem(vp, *raoffset))
895 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
897 if ((rabp->b_flags & B_CACHE) == 0) {
899 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
900 rabp->b_cmd = BUF_CMD_READ;
901 vfs_busy_pages(vp, rabp);
903 vn_strategy(vp, &rabp->b_bio1);
910 rv = biowait(&bp->b_bio1, "biord");
917 * Synchronous write, waits for completion.
919 * Write, release buffer on completion. (Done by iodone
920 * if async). Do not bother writing anything if the buffer
923 * Note that we set B_CACHE here, indicating that buffer is
924 * fully valid and thus cacheable. This is true even of NFS
925 * now so we set it generally. This could be set either here
926 * or in biodone() since the I/O is synchronous. We put it
930 bwrite(struct buf *bp)
934 if (bp->b_flags & B_INVAL) {
938 if (BUF_REFCNTNB(bp) == 0)
939 panic("bwrite: buffer is not busy???");
941 /* Mark the buffer clean */
944 bp->b_flags &= ~(B_ERROR | B_EINTR);
945 bp->b_flags |= B_CACHE;
946 bp->b_cmd = BUF_CMD_WRITE;
947 bp->b_bio1.bio_done = biodone_sync;
948 bp->b_bio1.bio_flags |= BIO_SYNC;
949 vfs_busy_pages(bp->b_vp, bp);
952 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
953 * valid for vnode-backed buffers.
955 bp->b_runningbufspace = bp->b_bufsize;
956 if (bp->b_runningbufspace) {
957 runningbufspace += bp->b_runningbufspace;
961 vn_strategy(bp->b_vp, &bp->b_bio1);
962 error = biowait(&bp->b_bio1, "biows");
970 * Asynchronous write. Start output on a buffer, but do not wait for
971 * it to complete. The buffer is released when the output completes.
973 * bwrite() ( or the VOP routine anyway ) is responsible for handling
974 * B_INVAL buffers. Not us.
977 bawrite(struct buf *bp)
979 if (bp->b_flags & B_INVAL) {
983 if (BUF_REFCNTNB(bp) == 0)
984 panic("bwrite: buffer is not busy???");
986 /* Mark the buffer clean */
989 bp->b_flags &= ~(B_ERROR | B_EINTR);
990 bp->b_flags |= B_CACHE;
991 bp->b_cmd = BUF_CMD_WRITE;
992 KKASSERT(bp->b_bio1.bio_done == NULL);
993 vfs_busy_pages(bp->b_vp, bp);
996 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
997 * valid for vnode-backed buffers.
999 bp->b_runningbufspace = bp->b_bufsize;
1000 if (bp->b_runningbufspace) {
1001 runningbufspace += bp->b_runningbufspace;
1006 vn_strategy(bp->b_vp, &bp->b_bio1);
1012 * Ordered write. Start output on a buffer, and flag it so that the
1013 * device will write it in the order it was queued. The buffer is
1014 * released when the output completes. bwrite() ( or the VOP routine
1015 * anyway ) is responsible for handling B_INVAL buffers.
1018 bowrite(struct buf *bp)
1020 bp->b_flags |= B_ORDERED;
1028 * Delayed write. (Buffer is marked dirty). Do not bother writing
1029 * anything if the buffer is marked invalid.
1031 * Note that since the buffer must be completely valid, we can safely
1032 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1033 * biodone() in order to prevent getblk from writing the buffer
1034 * out synchronously.
1037 bdwrite(struct buf *bp)
1039 if (BUF_REFCNTNB(bp) == 0)
1040 panic("bdwrite: buffer is not busy");
1042 if (bp->b_flags & B_INVAL) {
1049 * Set B_CACHE, indicating that the buffer is fully valid. This is
1050 * true even of NFS now.
1052 bp->b_flags |= B_CACHE;
1055 * This bmap keeps the system from needing to do the bmap later,
1056 * perhaps when the system is attempting to do a sync. Since it
1057 * is likely that the indirect block -- or whatever other datastructure
1058 * that the filesystem needs is still in memory now, it is a good
1059 * thing to do this. Note also, that if the pageout daemon is
1060 * requesting a sync -- there might not be enough memory to do
1061 * the bmap then... So, this is important to do.
1063 if (bp->b_bio2.bio_offset == NOOFFSET) {
1064 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1065 NULL, NULL, BUF_CMD_WRITE);
1069 * Because the underlying pages may still be mapped and
1070 * writable trying to set the dirty buffer (b_dirtyoff/end)
1071 * range here will be inaccurate.
1073 * However, we must still clean the pages to satisfy the
1074 * vnode_pager and pageout daemon, so theythink the pages
1075 * have been "cleaned". What has really occured is that
1076 * they've been earmarked for later writing by the buffer
1079 * So we get the b_dirtyoff/end update but will not actually
1080 * depend on it (NFS that is) until the pages are busied for
1083 vfs_clean_pages(bp);
1087 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1088 * due to the softdep code.
1095 * Turn buffer into delayed write request by marking it B_DELWRI.
1096 * B_RELBUF and B_NOCACHE must be cleared.
1098 * We reassign the buffer to itself to properly update it in the
1099 * dirty/clean lists.
1101 * Must be called from a critical section.
1102 * The buffer must be on BQUEUE_NONE.
1105 bdirty(struct buf *bp)
1107 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1108 if (bp->b_flags & B_NOCACHE) {
1109 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1110 bp->b_flags &= ~B_NOCACHE;
1112 if (bp->b_flags & B_INVAL) {
1113 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1115 bp->b_flags &= ~B_RELBUF;
1117 if ((bp->b_flags & B_DELWRI) == 0) {
1118 bp->b_flags |= B_DELWRI;
1120 atomic_add_int(&dirtybufcount, 1);
1121 dirtybufspace += bp->b_bufsize;
1122 if (bp->b_flags & B_HEAVY) {
1123 atomic_add_int(&dirtybufcounthw, 1);
1124 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
1131 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1132 * needs to be flushed with a different buf_daemon thread to avoid
1133 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1136 bheavy(struct buf *bp)
1138 if ((bp->b_flags & B_HEAVY) == 0) {
1139 bp->b_flags |= B_HEAVY;
1140 if (bp->b_flags & B_DELWRI) {
1141 atomic_add_int(&dirtybufcounthw, 1);
1142 atomic_add_int(&dirtybufspacehw, bp->b_bufsize);
1150 * Clear B_DELWRI for buffer.
1152 * Must be called from a critical section.
1154 * The buffer is typically on BQUEUE_NONE but there is one case in
1155 * brelse() that calls this function after placing the buffer on
1156 * a different queue.
1161 bundirty(struct buf *bp)
1163 if (bp->b_flags & B_DELWRI) {
1164 bp->b_flags &= ~B_DELWRI;
1166 atomic_subtract_int(&dirtybufcount, 1);
1167 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
1168 if (bp->b_flags & B_HEAVY) {
1169 atomic_subtract_int(&dirtybufcounthw, 1);
1170 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
1172 bd_signal(bp->b_bufsize);
1175 * Since it is now being written, we can clear its deferred write flag.
1177 bp->b_flags &= ~B_DEFERRED;
1183 * Release a busy buffer and, if requested, free its resources. The
1184 * buffer will be stashed in the appropriate bufqueue[] allowing it
1185 * to be accessed later as a cache entity or reused for other purposes.
1190 brelse(struct buf *bp)
1193 int saved_flags = bp->b_flags;
1196 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1199 * If B_NOCACHE is set we are being asked to destroy the buffer and
1200 * its backing store. Clear B_DELWRI.
1202 * B_NOCACHE is set in two cases: (1) when the caller really wants
1203 * to destroy the buffer and backing store and (2) when the caller
1204 * wants to destroy the buffer and backing store after a write
1207 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1211 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1213 * A re-dirtied buffer is only subject to destruction
1214 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1216 /* leave buffer intact */
1217 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1218 (bp->b_bufsize <= 0)) {
1220 * Either a failed read or we were asked to free or not
1221 * cache the buffer. This path is reached with B_DELWRI
1222 * set only if B_INVAL is already set. B_NOCACHE governs
1223 * backing store destruction.
1225 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1226 * buffer cannot be immediately freed.
1228 bp->b_flags |= B_INVAL;
1229 if (LIST_FIRST(&bp->b_dep) != NULL) {
1234 if (bp->b_flags & B_DELWRI) {
1235 atomic_subtract_int(&dirtybufcount, 1);
1236 atomic_subtract_int(&dirtybufspace, bp->b_bufsize);
1237 if (bp->b_flags & B_HEAVY) {
1238 atomic_subtract_int(&dirtybufcounthw, 1);
1239 atomic_subtract_int(&dirtybufspacehw, bp->b_bufsize);
1241 bd_signal(bp->b_bufsize);
1243 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1247 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1248 * If vfs_vmio_release() is called with either bit set, the
1249 * underlying pages may wind up getting freed causing a previous
1250 * write (bdwrite()) to get 'lost' because pages associated with
1251 * a B_DELWRI bp are marked clean. Pages associated with a
1252 * B_LOCKED buffer may be mapped by the filesystem.
1254 * If we want to release the buffer ourselves (rather then the
1255 * originator asking us to release it), give the originator a
1256 * chance to countermand the release by setting B_LOCKED.
1258 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1259 * if B_DELWRI is set.
1261 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1262 * on pages to return pages to the VM page queues.
1264 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1265 bp->b_flags &= ~B_RELBUF;
1266 } else if (vm_page_count_severe()) {
1267 if (LIST_FIRST(&bp->b_dep) != NULL) {
1269 buf_deallocate(bp); /* can set B_LOCKED */
1272 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1273 bp->b_flags &= ~B_RELBUF;
1275 bp->b_flags |= B_RELBUF;
1279 * Make sure b_cmd is clear. It may have already been cleared by
1282 * At this point destroying the buffer is governed by the B_INVAL
1283 * or B_RELBUF flags.
1285 bp->b_cmd = BUF_CMD_DONE;
1288 * VMIO buffer rundown. Make sure the VM page array is restored
1289 * after an I/O may have replaces some of the pages with bogus pages
1290 * in order to not destroy dirty pages in a fill-in read.
1292 * Note that due to the code above, if a buffer is marked B_DELWRI
1293 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1294 * B_INVAL may still be set, however.
1296 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1297 * but not the backing store. B_NOCACHE will destroy the backing
1300 * Note that dirty NFS buffers contain byte-granular write ranges
1301 * and should not be destroyed w/ B_INVAL even if the backing store
1304 if (bp->b_flags & B_VMIO) {
1306 * Rundown for VMIO buffers which are not dirty NFS buffers.
1318 * Get the base offset and length of the buffer. Note that
1319 * in the VMIO case if the buffer block size is not
1320 * page-aligned then b_data pointer may not be page-aligned.
1321 * But our b_xio.xio_pages array *IS* page aligned.
1323 * block sizes less then DEV_BSIZE (usually 512) are not
1324 * supported due to the page granularity bits (m->valid,
1325 * m->dirty, etc...).
1327 * See man buf(9) for more information
1330 resid = bp->b_bufsize;
1331 foff = bp->b_loffset;
1334 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1335 m = bp->b_xio.xio_pages[i];
1336 vm_page_flag_clear(m, PG_ZERO);
1338 * If we hit a bogus page, fixup *all* of them
1339 * now. Note that we left these pages wired
1340 * when we removed them so they had better exist,
1341 * and they cannot be ripped out from under us so
1342 * no critical section protection is necessary.
1344 if (m == bogus_page) {
1346 poff = OFF_TO_IDX(bp->b_loffset);
1348 for (j = i; j < bp->b_xio.xio_npages; j++) {
1351 mtmp = bp->b_xio.xio_pages[j];
1352 if (mtmp == bogus_page) {
1353 mtmp = vm_page_lookup(obj, poff + j);
1355 panic("brelse: page missing");
1357 bp->b_xio.xio_pages[j] = mtmp;
1361 if ((bp->b_flags & B_INVAL) == 0) {
1362 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1363 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1365 m = bp->b_xio.xio_pages[i];
1369 * Invalidate the backing store if B_NOCACHE is set
1370 * (e.g. used with vinvalbuf()). If this is NFS
1371 * we impose a requirement that the block size be
1372 * a multiple of PAGE_SIZE and create a temporary
1373 * hack to basically invalidate the whole page. The
1374 * problem is that NFS uses really odd buffer sizes
1375 * especially when tracking piecemeal writes and
1376 * it also vinvalbuf()'s a lot, which would result
1377 * in only partial page validation and invalidation
1378 * here. If the file page is mmap()'d, however,
1379 * all the valid bits get set so after we invalidate
1380 * here we would end up with weird m->valid values
1381 * like 0xfc. nfs_getpages() can't handle this so
1382 * we clear all the valid bits for the NFS case
1383 * instead of just some of them.
1385 * The real bug is the VM system having to set m->valid
1386 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1387 * itself is an artifact of the whole 512-byte
1388 * granular mess that exists to support odd block
1389 * sizes and UFS meta-data block sizes (e.g. 6144).
1390 * A complete rewrite is required.
1394 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1395 int poffset = foff & PAGE_MASK;
1398 presid = PAGE_SIZE - poffset;
1399 if (bp->b_vp->v_tag == VT_NFS &&
1400 bp->b_vp->v_type == VREG) {
1402 } else if (presid > resid) {
1405 KASSERT(presid >= 0, ("brelse: extra page"));
1406 vm_page_set_invalid(m, poffset, presid);
1408 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1409 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1411 if (bp->b_flags & (B_INVAL | B_RELBUF))
1412 vfs_vmio_release(bp);
1416 * Rundown for non-VMIO buffers.
1418 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1422 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1429 if (bp->b_qindex != BQUEUE_NONE)
1430 panic("brelse: free buffer onto another queue???");
1431 if (BUF_REFCNTNB(bp) > 1) {
1432 /* Temporary panic to verify exclusive locking */
1433 /* This panic goes away when we allow shared refs */
1434 panic("brelse: multiple refs");
1440 * Figure out the correct queue to place the cleaned up buffer on.
1441 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1442 * disassociated from their vnode.
1444 spin_lock_wr(&bufspin);
1445 if (bp->b_flags & B_LOCKED) {
1447 * Buffers that are locked are placed in the locked queue
1448 * immediately, regardless of their state.
1450 bp->b_qindex = BQUEUE_LOCKED;
1451 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1452 } else if (bp->b_bufsize == 0) {
1454 * Buffers with no memory. Due to conditionals near the top
1455 * of brelse() such buffers should probably already be
1456 * marked B_INVAL and disassociated from their vnode.
1458 bp->b_flags |= B_INVAL;
1459 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1460 KKASSERT((bp->b_flags & B_HASHED) == 0);
1461 if (bp->b_kvasize) {
1462 bp->b_qindex = BQUEUE_EMPTYKVA;
1464 bp->b_qindex = BQUEUE_EMPTY;
1466 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1467 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1469 * Buffers with junk contents. Again these buffers had better
1470 * already be disassociated from their vnode.
1472 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1473 KKASSERT((bp->b_flags & B_HASHED) == 0);
1474 bp->b_flags |= B_INVAL;
1475 bp->b_qindex = BQUEUE_CLEAN;
1476 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1479 * Remaining buffers. These buffers are still associated with
1482 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1484 bp->b_qindex = BQUEUE_DIRTY;
1485 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1487 case B_DELWRI | B_HEAVY:
1488 bp->b_qindex = BQUEUE_DIRTY_HW;
1489 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1494 * NOTE: Buffers are always placed at the end of the
1495 * queue. If B_AGE is not set the buffer will cycle
1496 * through the queue twice.
1498 bp->b_qindex = BQUEUE_CLEAN;
1499 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1503 spin_unlock_wr(&bufspin);
1506 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1507 * on the correct queue.
1509 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1513 * The bp is on an appropriate queue unless locked. If it is not
1514 * locked or dirty we can wakeup threads waiting for buffer space.
1516 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1517 * if B_INVAL is set ).
1519 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1523 * Something we can maybe free or reuse
1525 if (bp->b_bufsize || bp->b_kvasize)
1529 * Clean up temporary flags and unlock the buffer.
1531 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
1538 * Release a buffer back to the appropriate queue but do not try to free
1539 * it. The buffer is expected to be used again soon.
1541 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1542 * biodone() to requeue an async I/O on completion. It is also used when
1543 * known good buffers need to be requeued but we think we may need the data
1546 * XXX we should be able to leave the B_RELBUF hint set on completion.
1551 bqrelse(struct buf *bp)
1553 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1555 if (bp->b_qindex != BQUEUE_NONE)
1556 panic("bqrelse: free buffer onto another queue???");
1557 if (BUF_REFCNTNB(bp) > 1) {
1558 /* do not release to free list */
1559 panic("bqrelse: multiple refs");
1563 buf_act_advance(bp);
1565 spin_lock_wr(&bufspin);
1566 if (bp->b_flags & B_LOCKED) {
1568 * Locked buffers are released to the locked queue. However,
1569 * if the buffer is dirty it will first go into the dirty
1570 * queue and later on after the I/O completes successfully it
1571 * will be released to the locked queue.
1573 bp->b_qindex = BQUEUE_LOCKED;
1574 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1575 } else if (bp->b_flags & B_DELWRI) {
1576 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1577 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1578 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1579 } else if (vm_page_count_severe()) {
1581 * We are too low on memory, we have to try to free the
1582 * buffer (most importantly: the wired pages making up its
1583 * backing store) *now*.
1585 spin_unlock_wr(&bufspin);
1589 bp->b_qindex = BQUEUE_CLEAN;
1590 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1592 spin_unlock_wr(&bufspin);
1594 if ((bp->b_flags & B_LOCKED) == 0 &&
1595 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1600 * Something we can maybe free or reuse.
1602 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1606 * Final cleanup and unlock. Clear bits that are only used while a
1607 * buffer is actively locked.
1609 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
1616 * Return backing pages held by the buffer 'bp' back to the VM system
1617 * if possible. The pages are freed if they are no longer valid or
1618 * attempt to free if it was used for direct I/O otherwise they are
1619 * sent to the page cache.
1621 * Pages that were marked busy are left alone and skipped.
1623 * The KVA mapping (b_data) for the underlying pages is removed by
1627 vfs_vmio_release(struct buf *bp)
1633 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1634 m = bp->b_xio.xio_pages[i];
1635 bp->b_xio.xio_pages[i] = NULL;
1638 * This is a very important bit of code. We try to track
1639 * VM page use whether the pages are wired into the buffer
1640 * cache or not. While wired into the buffer cache the
1641 * bp tracks the act_count.
1643 * We can choose to place unwired pages on the inactive
1644 * queue (0) or active queue (1). If we place too many
1645 * on the active queue the queue will cycle the act_count
1646 * on pages we'd like to keep, just from single-use pages
1647 * (such as when doing a tar-up or file scan).
1649 if (bp->b_act_count < vm_cycle_point)
1650 vm_page_unwire(m, 0);
1652 vm_page_unwire(m, 1);
1655 * We don't mess with busy pages, it is
1656 * the responsibility of the process that
1657 * busied the pages to deal with them.
1659 if ((m->flags & PG_BUSY) || (m->busy != 0))
1662 if (m->wire_count == 0) {
1663 vm_page_flag_clear(m, PG_ZERO);
1665 * Might as well free the page if we can and it has
1666 * no valid data. We also free the page if the
1667 * buffer was used for direct I/O.
1670 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1671 m->hold_count == 0) {
1673 vm_page_protect(m, VM_PROT_NONE);
1677 if (bp->b_flags & B_DIRECT) {
1678 vm_page_try_to_free(m);
1679 } else if (vm_page_count_severe()) {
1680 m->act_count = bp->b_act_count;
1681 vm_page_try_to_cache(m);
1683 m->act_count = bp->b_act_count;
1688 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1689 if (bp->b_bufsize) {
1693 bp->b_xio.xio_npages = 0;
1694 bp->b_flags &= ~B_VMIO;
1695 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1706 * Implement clustered async writes for clearing out B_DELWRI buffers.
1707 * This is much better then the old way of writing only one buffer at
1708 * a time. Note that we may not be presented with the buffers in the
1709 * correct order, so we search for the cluster in both directions.
1711 * The buffer is locked on call.
1714 vfs_bio_awrite(struct buf *bp)
1718 off_t loffset = bp->b_loffset;
1719 struct vnode *vp = bp->b_vp;
1726 * right now we support clustered writing only to regular files. If
1727 * we find a clusterable block we could be in the middle of a cluster
1728 * rather then at the beginning.
1730 * NOTE: b_bio1 contains the logical loffset and is aliased
1731 * to b_loffset. b_bio2 contains the translated block number.
1733 if ((vp->v_type == VREG) &&
1734 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1735 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1737 size = vp->v_mount->mnt_stat.f_iosize;
1739 for (i = size; i < MAXPHYS; i += size) {
1740 if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) &&
1741 BUF_REFCNT(bpa) == 0 &&
1742 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1743 (B_DELWRI | B_CLUSTEROK)) &&
1744 (bpa->b_bufsize == size)) {
1745 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1746 (bpa->b_bio2.bio_offset !=
1747 bp->b_bio2.bio_offset + i))
1753 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1754 if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) &&
1755 BUF_REFCNT(bpa) == 0 &&
1756 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1757 (B_DELWRI | B_CLUSTEROK)) &&
1758 (bpa->b_bufsize == size)) {
1759 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1760 (bpa->b_bio2.bio_offset !=
1761 bp->b_bio2.bio_offset - j))
1771 * this is a possible cluster write
1773 if (nbytes != size) {
1775 nwritten = cluster_wbuild(vp, size,
1776 loffset - j, nbytes);
1782 * default (old) behavior, writing out only one block
1784 * XXX returns b_bufsize instead of b_bcount for nwritten?
1786 nwritten = bp->b_bufsize;
1796 * Find and initialize a new buffer header, freeing up existing buffers
1797 * in the bufqueues as necessary. The new buffer is returned locked.
1799 * Important: B_INVAL is not set. If the caller wishes to throw the
1800 * buffer away, the caller must set B_INVAL prior to calling brelse().
1803 * We have insufficient buffer headers
1804 * We have insufficient buffer space
1805 * buffer_map is too fragmented ( space reservation fails )
1806 * If we have to flush dirty buffers ( but we try to avoid this )
1808 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1809 * Instead we ask the buf daemon to do it for us. We attempt to
1810 * avoid piecemeal wakeups of the pageout daemon.
1815 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1821 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1822 static int flushingbufs;
1825 * We can't afford to block since we might be holding a vnode lock,
1826 * which may prevent system daemons from running. We deal with
1827 * low-memory situations by proactively returning memory and running
1828 * async I/O rather then sync I/O.
1832 --getnewbufrestarts;
1834 ++getnewbufrestarts;
1837 * Setup for scan. If we do not have enough free buffers,
1838 * we setup a degenerate case that immediately fails. Note
1839 * that if we are specially marked process, we are allowed to
1840 * dip into our reserves.
1842 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1844 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1845 * However, there are a number of cases (defragging, reusing, ...)
1846 * where we cannot backup.
1848 nqindex = BQUEUE_EMPTYKVA;
1849 spin_lock_wr(&bufspin);
1850 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1854 * If no EMPTYKVA buffers and we are either
1855 * defragging or reusing, locate a CLEAN buffer
1856 * to free or reuse. If bufspace useage is low
1857 * skip this step so we can allocate a new buffer.
1859 if (defrag || bufspace >= lobufspace) {
1860 nqindex = BQUEUE_CLEAN;
1861 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1865 * If we could not find or were not allowed to reuse a
1866 * CLEAN buffer, check to see if it is ok to use an EMPTY
1867 * buffer. We can only use an EMPTY buffer if allocating
1868 * its KVA would not otherwise run us out of buffer space.
1870 if (nbp == NULL && defrag == 0 &&
1871 bufspace + maxsize < hibufspace) {
1872 nqindex = BQUEUE_EMPTY;
1873 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1878 * Run scan, possibly freeing data and/or kva mappings on the fly
1881 * WARNING! bufspin is held!
1883 while ((bp = nbp) != NULL) {
1884 int qindex = nqindex;
1886 nbp = TAILQ_NEXT(bp, b_freelist);
1889 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1890 * cycles through the queue twice before being selected.
1892 if (qindex == BQUEUE_CLEAN &&
1893 (bp->b_flags & B_AGE) == 0 && nbp) {
1894 bp->b_flags |= B_AGE;
1895 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1896 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1901 * Calculate next bp ( we can only use it if we do not block
1902 * or do other fancy things ).
1907 nqindex = BQUEUE_EMPTYKVA;
1908 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1911 case BQUEUE_EMPTYKVA:
1912 nqindex = BQUEUE_CLEAN;
1913 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1927 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1930 * Note: we no longer distinguish between VMIO and non-VMIO
1934 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1937 * If we are defragging then we need a buffer with
1938 * b_kvasize != 0. XXX this situation should no longer
1939 * occur, if defrag is non-zero the buffer's b_kvasize
1940 * should also be non-zero at this point. XXX
1942 if (defrag && bp->b_kvasize == 0) {
1943 kprintf("Warning: defrag empty buffer %p\n", bp);
1948 * Start freeing the bp. This is somewhat involved. nbp
1949 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1950 * on the clean list must be disassociated from their
1951 * current vnode. Buffers on the empty[kva] lists have
1952 * already been disassociated.
1955 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1956 spin_unlock_wr(&bufspin);
1957 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1958 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1961 if (bp->b_qindex != qindex) {
1962 spin_unlock_wr(&bufspin);
1963 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1967 bremfree_locked(bp);
1968 spin_unlock_wr(&bufspin);
1971 * Dependancies must be handled before we disassociate the
1974 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1975 * be immediately disassociated. HAMMER then becomes
1976 * responsible for releasing the buffer.
1978 * NOTE: bufspin is UNLOCKED now.
1980 if (LIST_FIRST(&bp->b_dep) != NULL) {
1984 if (bp->b_flags & B_LOCKED) {
1988 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1991 if (qindex == BQUEUE_CLEAN) {
1993 if (bp->b_flags & B_VMIO) {
1995 vfs_vmio_release(bp);
2004 * NOTE: nbp is now entirely invalid. We can only restart
2005 * the scan from this point on.
2007 * Get the rest of the buffer freed up. b_kva* is still
2008 * valid after this operation.
2011 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
2012 KKASSERT((bp->b_flags & B_HASHED) == 0);
2015 * critical section protection is not required when
2016 * scrapping a buffer's contents because it is already
2019 if (bp->b_bufsize) {
2025 bp->b_flags = B_BNOCLIP;
2026 bp->b_cmd = BUF_CMD_DONE;
2031 bp->b_xio.xio_npages = 0;
2032 bp->b_dirtyoff = bp->b_dirtyend = 0;
2033 bp->b_act_count = ACT_INIT;
2035 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2037 if (blkflags & GETBLK_BHEAVY)
2038 bp->b_flags |= B_HEAVY;
2041 * If we are defragging then free the buffer.
2044 bp->b_flags |= B_INVAL;
2052 * If we are overcomitted then recover the buffer and its
2053 * KVM space. This occurs in rare situations when multiple
2054 * processes are blocked in getnewbuf() or allocbuf().
2056 if (bufspace >= hibufspace)
2058 if (flushingbufs && bp->b_kvasize != 0) {
2059 bp->b_flags |= B_INVAL;
2064 if (bufspace < lobufspace)
2067 /* NOT REACHED, bufspin not held */
2071 * If we exhausted our list, sleep as appropriate. We may have to
2072 * wakeup various daemons and write out some dirty buffers.
2074 * Generally we are sleeping due to insufficient buffer space.
2076 * NOTE: bufspin is held if bp is NULL, else it is not held.
2082 spin_unlock_wr(&bufspin);
2084 flags = VFS_BIO_NEED_BUFSPACE;
2086 } else if (bufspace >= hibufspace) {
2088 flags = VFS_BIO_NEED_BUFSPACE;
2091 flags = VFS_BIO_NEED_ANY;
2094 needsbuffer |= flags;
2095 bd_speedup(); /* heeeelp */
2096 while (needsbuffer & flags) {
2097 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
2102 * We finally have a valid bp. We aren't quite out of the
2103 * woods, we still have to reserve kva space. In order
2104 * to keep fragmentation sane we only allocate kva in
2107 * (bufspin is not held)
2109 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2111 if (maxsize != bp->b_kvasize) {
2112 vm_offset_t addr = 0;
2118 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2119 vm_map_lock(&buffer_map);
2121 if (vm_map_findspace(&buffer_map,
2122 vm_map_min(&buffer_map), maxsize,
2123 maxsize, 0, &addr)) {
2125 * Uh oh. Buffer map is too fragmented. We
2126 * must defragment the map.
2128 vm_map_unlock(&buffer_map);
2129 vm_map_entry_release(count);
2132 bp->b_flags |= B_INVAL;
2138 vm_map_insert(&buffer_map, &count,
2140 addr, addr + maxsize,
2142 VM_PROT_ALL, VM_PROT_ALL,
2145 bp->b_kvabase = (caddr_t) addr;
2146 bp->b_kvasize = maxsize;
2147 bufspace += bp->b_kvasize;
2150 vm_map_unlock(&buffer_map);
2151 vm_map_entry_release(count);
2154 bp->b_data = bp->b_kvabase;
2160 * This routine is called in an emergency to recover VM pages from the
2161 * buffer cache by cashing in clean buffers. The idea is to recover
2162 * enough pages to be able to satisfy a stuck bio_page_alloc().
2165 recoverbufpages(void)
2172 spin_lock_wr(&bufspin);
2173 while (bytes < MAXBSIZE) {
2174 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2179 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2180 * cycles through the queue twice before being selected.
2182 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
2183 bp->b_flags |= B_AGE;
2184 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
2185 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
2193 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
2194 KKASSERT((bp->b_flags & B_DELWRI) == 0);
2197 * Start freeing the bp. This is somewhat involved.
2199 * Buffers on the clean list must be disassociated from
2200 * their current vnode
2203 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2204 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp);
2205 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
2208 if (bp->b_qindex != BQUEUE_CLEAN) {
2209 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp, bp->b_qindex);
2213 bremfree_locked(bp);
2214 spin_unlock_wr(&bufspin);
2217 * Dependancies must be handled before we disassociate the
2220 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2221 * be immediately disassociated. HAMMER then becomes
2222 * responsible for releasing the buffer.
2224 if (LIST_FIRST(&bp->b_dep) != NULL) {
2226 if (bp->b_flags & B_LOCKED) {
2228 spin_lock_wr(&bufspin);
2231 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2234 bytes += bp->b_bufsize;
2237 if (bp->b_flags & B_VMIO) {
2238 bp->b_flags |= B_DIRECT; /* try to free pages */
2239 vfs_vmio_release(bp);
2244 KKASSERT(bp->b_vp == NULL);
2245 KKASSERT((bp->b_flags & B_HASHED) == 0);
2248 * critical section protection is not required when
2249 * scrapping a buffer's contents because it is already
2256 bp->b_flags = B_BNOCLIP;
2257 bp->b_cmd = BUF_CMD_DONE;
2262 bp->b_xio.xio_npages = 0;
2263 bp->b_dirtyoff = bp->b_dirtyend = 0;
2265 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2267 bp->b_flags |= B_INVAL;
2270 spin_lock_wr(&bufspin);
2272 spin_unlock_wr(&bufspin);
2279 * Buffer flushing daemon. Buffers are normally flushed by the
2280 * update daemon but if it cannot keep up this process starts to
2281 * take the load in an attempt to prevent getnewbuf() from blocking.
2283 * Once a flush is initiated it does not stop until the number
2284 * of buffers falls below lodirtybuffers, but we will wake up anyone
2285 * waiting at the mid-point.
2288 static struct kproc_desc buf_kp = {
2293 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2294 kproc_start, &buf_kp)
2296 static struct kproc_desc bufhw_kp = {
2301 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2302 kproc_start, &bufhw_kp)
2310 * This process needs to be suspended prior to shutdown sync.
2312 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2313 bufdaemon_td, SHUTDOWN_PRI_LAST);
2314 curthread->td_flags |= TDF_SYSTHREAD;
2317 * This process is allowed to take the buffer cache to the limit
2322 kproc_suspend_loop();
2325 * Do the flush as long as the number of dirty buffers
2326 * (including those running) exceeds lodirtybufspace.
2328 * When flushing limit running I/O to hirunningspace
2329 * Do the flush. Limit the amount of in-transit I/O we
2330 * allow to build up, otherwise we would completely saturate
2331 * the I/O system. Wakeup any waiting processes before we
2332 * normally would so they can run in parallel with our drain.
2334 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2335 * but because we split the operation into two threads we
2336 * have to cut it in half for each thread.
2338 waitrunningbufspace();
2339 limit = lodirtybufspace / 2;
2340 while (runningbufspace + dirtybufspace > limit ||
2341 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
2342 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2344 if (runningbufspace < hirunningspace)
2346 waitrunningbufspace();
2350 * We reached our low water mark, reset the
2351 * request and sleep until we are needed again.
2352 * The sleep is just so the suspend code works.
2354 spin_lock_wr(&needsbuffer_spin);
2355 if (bd_request == 0) {
2356 ssleep(&bd_request, &needsbuffer_spin, 0,
2360 spin_unlock_wr(&needsbuffer_spin);
2370 * This process needs to be suspended prior to shutdown sync.
2372 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2373 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2374 curthread->td_flags |= TDF_SYSTHREAD;
2377 * This process is allowed to take the buffer cache to the limit
2382 kproc_suspend_loop();
2385 * Do the flush. Limit the amount of in-transit I/O we
2386 * allow to build up, otherwise we would completely saturate
2387 * the I/O system. Wakeup any waiting processes before we
2388 * normally would so they can run in parallel with our drain.
2390 * Once we decide to flush push the queued I/O up to
2391 * hirunningspace in order to trigger bursting by the bioq
2394 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2395 * but because we split the operation into two threads we
2396 * have to cut it in half for each thread.
2398 waitrunningbufspace();
2399 limit = lodirtybufspace / 2;
2400 while (runningbufspace + dirtybufspacehw > limit ||
2401 dirtybufcounthw >= nbuf / 2) {
2402 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2404 if (runningbufspace < hirunningspace)
2406 waitrunningbufspace();
2410 * We reached our low water mark, reset the
2411 * request and sleep until we are needed again.
2412 * The sleep is just so the suspend code works.
2414 spin_lock_wr(&needsbuffer_spin);
2415 if (bd_request_hw == 0) {
2416 ssleep(&bd_request_hw, &needsbuffer_spin, 0,
2420 spin_unlock_wr(&needsbuffer_spin);
2427 * Try to flush a buffer in the dirty queue. We must be careful to
2428 * free up B_INVAL buffers instead of write them, which NFS is
2429 * particularly sensitive to.
2431 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2432 * that we really want to try to get the buffer out and reuse it
2433 * due to the write load on the machine.
2436 flushbufqueues(bufq_type_t q)
2442 spin_lock_wr(&bufspin);
2445 bp = TAILQ_FIRST(&bufqueues[q]);
2447 KASSERT((bp->b_flags & B_DELWRI),
2448 ("unexpected clean buffer %p", bp));
2450 if (bp->b_flags & B_DELWRI) {
2451 if (bp->b_flags & B_INVAL) {
2452 spin_unlock_wr(&bufspin);
2454 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2455 panic("flushbufqueues: locked buf");
2461 if (LIST_FIRST(&bp->b_dep) != NULL &&
2462 (bp->b_flags & B_DEFERRED) == 0 &&
2463 buf_countdeps(bp, 0)) {
2464 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2465 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2467 bp->b_flags |= B_DEFERRED;
2468 bp = TAILQ_FIRST(&bufqueues[q]);
2473 * Only write it out if we can successfully lock
2474 * it. If the buffer has a dependancy,
2475 * buf_checkwrite must also return 0 for us to
2476 * be able to initate the write.
2478 * If the buffer is flagged B_ERROR it may be
2479 * requeued over and over again, we try to
2480 * avoid a live lock.
2482 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
2483 spin_unlock_wr(&bufspin);
2485 if (LIST_FIRST(&bp->b_dep) != NULL &&
2486 buf_checkwrite(bp)) {
2489 } else if (bp->b_flags & B_ERROR) {
2490 tsleep(bp, 0, "bioer", 1);
2491 bp->b_flags &= ~B_AGE;
2494 bp->b_flags |= B_AGE;
2501 bp = TAILQ_NEXT(bp, b_freelist);
2504 spin_unlock_wr(&bufspin);
2511 * Returns true if no I/O is needed to access the associated VM object.
2512 * This is like findblk except it also hunts around in the VM system for
2515 * Note that we ignore vm_page_free() races from interrupts against our
2516 * lookup, since if the caller is not protected our return value will not
2517 * be any more valid then otherwise once we exit the critical section.
2520 inmem(struct vnode *vp, off_t loffset)
2523 vm_offset_t toff, tinc, size;
2526 if (findblk(vp, loffset, FINDBLK_TEST))
2528 if (vp->v_mount == NULL)
2530 if ((obj = vp->v_object) == NULL)
2534 if (size > vp->v_mount->mnt_stat.f_iosize)
2535 size = vp->v_mount->mnt_stat.f_iosize;
2537 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2538 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2542 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2543 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2544 if (vm_page_is_valid(m,
2545 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2554 * Locate and return the specified buffer. Unless flagged otherwise,
2555 * a locked buffer will be returned if it exists or NULL if it does not.
2557 * findblk()'d buffers are still on the bufqueues and if you intend
2558 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2559 * and possibly do other stuff to it.
2561 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2562 * for locking the buffer and ensuring that it remains
2563 * the desired buffer after locking.
2565 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2566 * to acquire the lock we return NULL, even if the
2569 * (0) - Lock the buffer blocking.
2574 findblk(struct vnode *vp, off_t loffset, int flags)
2580 lkflags = LK_EXCLUSIVE;
2581 if (flags & FINDBLK_NBLOCK)
2582 lkflags |= LK_NOWAIT;
2585 lwkt_gettoken(&vlock, &vp->v_token);
2586 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2587 lwkt_reltoken(&vlock);
2588 if (bp == NULL || (flags & FINDBLK_TEST))
2590 if (BUF_LOCK(bp, lkflags)) {
2594 if (bp->b_vp == vp && bp->b_loffset == loffset)
2604 * Similar to getblk() except only returns the buffer if it is
2605 * B_CACHE and requires no other manipulation. Otherwise NULL
2608 * If B_RAM is set the buffer might be just fine, but we return
2609 * NULL anyway because we want the code to fall through to the
2610 * cluster read. Otherwise read-ahead breaks.
2613 getcacheblk(struct vnode *vp, off_t loffset)
2617 bp = findblk(vp, loffset, 0);
2619 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
2620 bp->b_flags &= ~B_AGE;
2633 * Get a block given a specified block and offset into a file/device.
2634 * B_INVAL may or may not be set on return. The caller should clear
2635 * B_INVAL prior to initiating a READ.
2637 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2638 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2639 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2640 * without doing any of those things the system will likely believe
2641 * the buffer to be valid (especially if it is not B_VMIO), and the
2642 * next getblk() will return the buffer with B_CACHE set.
2644 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2645 * an existing buffer.
2647 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2648 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2649 * and then cleared based on the backing VM. If the previous buffer is
2650 * non-0-sized but invalid, B_CACHE will be cleared.
2652 * If getblk() must create a new buffer, the new buffer is returned with
2653 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2654 * case it is returned with B_INVAL clear and B_CACHE set based on the
2657 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2658 * B_CACHE bit is clear.
2660 * What this means, basically, is that the caller should use B_CACHE to
2661 * determine whether the buffer is fully valid or not and should clear
2662 * B_INVAL prior to issuing a read. If the caller intends to validate
2663 * the buffer by loading its data area with something, the caller needs
2664 * to clear B_INVAL. If the caller does this without issuing an I/O,
2665 * the caller should set B_CACHE ( as an optimization ), else the caller
2666 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2667 * a write attempt or if it was a successfull read. If the caller
2668 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2669 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2673 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2674 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2679 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2682 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2686 if (size > MAXBSIZE)
2687 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2688 if (vp->v_object == NULL)
2689 panic("getblk: vnode %p has no object!", vp);
2692 if ((bp = findblk(vp, loffset, FINDBLK_TEST)) != NULL) {
2694 * The buffer was found in the cache, but we need to lock it.
2695 * Even with LK_NOWAIT the lockmgr may break our critical
2696 * section, so double-check the validity of the buffer
2697 * once the lock has been obtained.
2699 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2700 if (blkflags & GETBLK_NOWAIT)
2702 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2703 if (blkflags & GETBLK_PCATCH)
2704 lkflags |= LK_PCATCH;
2705 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2707 if (error == ENOLCK)
2711 /* buffer may have changed on us */
2715 * Once the buffer has been locked, make sure we didn't race
2716 * a buffer recyclement. Buffers that are no longer hashed
2717 * will have b_vp == NULL, so this takes care of that check
2720 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2721 kprintf("Warning buffer %p (vp %p loffset %lld) "
2723 bp, vp, (long long)loffset);
2729 * If SZMATCH any pre-existing buffer must be of the requested
2730 * size or NULL is returned. The caller absolutely does not
2731 * want getblk() to bwrite() the buffer on a size mismatch.
2733 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2739 * All vnode-based buffers must be backed by a VM object.
2741 KKASSERT(bp->b_flags & B_VMIO);
2742 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2743 bp->b_flags &= ~B_AGE;
2746 * Make sure that B_INVAL buffers do not have a cached
2747 * block number translation.
2749 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2750 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2751 " did not have cleared bio_offset cache\n",
2752 bp, vp, (long long)loffset);
2753 clearbiocache(&bp->b_bio2);
2757 * The buffer is locked. B_CACHE is cleared if the buffer is
2760 if (bp->b_flags & B_INVAL)
2761 bp->b_flags &= ~B_CACHE;
2765 * Any size inconsistancy with a dirty buffer or a buffer
2766 * with a softupdates dependancy must be resolved. Resizing
2767 * the buffer in such circumstances can lead to problems.
2769 * Dirty or dependant buffers are written synchronously.
2770 * Other types of buffers are simply released and
2771 * reconstituted as they may be backed by valid, dirty VM
2772 * pages (but not marked B_DELWRI).
2774 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2775 * and may be left over from a prior truncation (and thus
2776 * no longer represent the actual EOF point), so we
2777 * definitely do not want to B_NOCACHE the backing store.
2779 if (size != bp->b_bcount) {
2781 if (bp->b_flags & B_DELWRI) {
2782 bp->b_flags |= B_RELBUF;
2784 } else if (LIST_FIRST(&bp->b_dep)) {
2785 bp->b_flags |= B_RELBUF;
2788 bp->b_flags |= B_RELBUF;
2794 KKASSERT(size <= bp->b_kvasize);
2795 KASSERT(bp->b_loffset != NOOFFSET,
2796 ("getblk: no buffer offset"));
2799 * A buffer with B_DELWRI set and B_CACHE clear must
2800 * be committed before we can return the buffer in
2801 * order to prevent the caller from issuing a read
2802 * ( due to B_CACHE not being set ) and overwriting
2805 * Most callers, including NFS and FFS, need this to
2806 * operate properly either because they assume they
2807 * can issue a read if B_CACHE is not set, or because
2808 * ( for example ) an uncached B_DELWRI might loop due
2809 * to softupdates re-dirtying the buffer. In the latter
2810 * case, B_CACHE is set after the first write completes,
2811 * preventing further loops.
2813 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2814 * above while extending the buffer, we cannot allow the
2815 * buffer to remain with B_CACHE set after the write
2816 * completes or it will represent a corrupt state. To
2817 * deal with this we set B_NOCACHE to scrap the buffer
2820 * XXX Should this be B_RELBUF instead of B_NOCACHE?
2821 * I'm not even sure this state is still possible
2822 * now that getblk() writes out any dirty buffers
2825 * We might be able to do something fancy, like setting
2826 * B_CACHE in bwrite() except if B_DELWRI is already set,
2827 * so the below call doesn't set B_CACHE, but that gets real
2828 * confusing. This is much easier.
2831 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2833 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
2834 "and CACHE clear, b_flags %08x\n",
2835 bp, (intmax_t)bp->b_loffset, bp->b_flags);
2836 bp->b_flags |= B_NOCACHE;
2843 * Buffer is not in-core, create new buffer. The buffer
2844 * returned by getnewbuf() is locked. Note that the returned
2845 * buffer is also considered valid (not marked B_INVAL).
2847 * Calculating the offset for the I/O requires figuring out
2848 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2849 * the mount's f_iosize otherwise. If the vnode does not
2850 * have an associated mount we assume that the passed size is
2853 * Note that vn_isdisk() cannot be used here since it may
2854 * return a failure for numerous reasons. Note that the
2855 * buffer size may be larger then the block size (the caller
2856 * will use block numbers with the proper multiple). Beware
2857 * of using any v_* fields which are part of unions. In
2858 * particular, in DragonFly the mount point overloading
2859 * mechanism uses the namecache only and the underlying
2860 * directory vnode is not a special case.
2864 if (vp->v_type == VBLK || vp->v_type == VCHR)
2866 else if (vp->v_mount)
2867 bsize = vp->v_mount->mnt_stat.f_iosize;
2871 maxsize = size + (loffset & PAGE_MASK);
2872 maxsize = imax(maxsize, bsize);
2874 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
2876 if (slpflags || slptimeo)
2882 * Atomically insert the buffer into the hash, so that it can
2883 * be found by findblk().
2885 * If bgetvp() returns non-zero a collision occured, and the
2886 * bp will not be associated with the vnode.
2888 * Make sure the translation layer has been cleared.
2890 bp->b_loffset = loffset;
2891 bp->b_bio2.bio_offset = NOOFFSET;
2892 /* bp->b_bio2.bio_next = NULL; */
2894 if (bgetvp(vp, bp)) {
2895 bp->b_flags |= B_INVAL;
2901 * All vnode-based buffers must be backed by a VM object.
2903 KKASSERT(vp->v_object != NULL);
2904 bp->b_flags |= B_VMIO;
2905 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2917 * Reacquire a buffer that was previously released to the locked queue,
2918 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2919 * set B_LOCKED (which handles the acquisition race).
2921 * To this end, either B_LOCKED must be set or the dependancy list must be
2927 regetblk(struct buf *bp)
2929 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
2930 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2937 * Get an empty, disassociated buffer of given size. The buffer is
2938 * initially set to B_INVAL.
2940 * critical section protection is not required for the allocbuf()
2941 * call because races are impossible here.
2951 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2953 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2958 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2966 * This code constitutes the buffer memory from either anonymous system
2967 * memory (in the case of non-VMIO operations) or from an associated
2968 * VM object (in the case of VMIO operations). This code is able to
2969 * resize a buffer up or down.
2971 * Note that this code is tricky, and has many complications to resolve
2972 * deadlock or inconsistant data situations. Tread lightly!!!
2973 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2974 * the caller. Calling this code willy nilly can result in the loss of data.
2976 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2977 * B_CACHE for the non-VMIO case.
2979 * This routine does not need to be called from a critical section but you
2980 * must own the buffer.
2985 allocbuf(struct buf *bp, int size)
2987 int newbsize, mbsize;
2990 if (BUF_REFCNT(bp) == 0)
2991 panic("allocbuf: buffer not busy");
2993 if (bp->b_kvasize < size)
2994 panic("allocbuf: buffer too small");
2996 if ((bp->b_flags & B_VMIO) == 0) {
3000 * Just get anonymous memory from the kernel. Don't
3001 * mess with B_CACHE.
3003 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3004 if (bp->b_flags & B_MALLOC)
3007 newbsize = round_page(size);
3009 if (newbsize < bp->b_bufsize) {
3011 * Malloced buffers are not shrunk
3013 if (bp->b_flags & B_MALLOC) {
3015 bp->b_bcount = size;
3017 kfree(bp->b_data, M_BIOBUF);
3018 if (bp->b_bufsize) {
3019 bufmallocspace -= bp->b_bufsize;
3023 bp->b_data = bp->b_kvabase;
3025 bp->b_flags &= ~B_MALLOC;
3031 (vm_offset_t) bp->b_data + newbsize,
3032 (vm_offset_t) bp->b_data + bp->b_bufsize);
3033 } else if (newbsize > bp->b_bufsize) {
3035 * We only use malloced memory on the first allocation.
3036 * and revert to page-allocated memory when the buffer
3039 if ((bufmallocspace < maxbufmallocspace) &&
3040 (bp->b_bufsize == 0) &&
3041 (mbsize <= PAGE_SIZE/2)) {
3043 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3044 bp->b_bufsize = mbsize;
3045 bp->b_bcount = size;
3046 bp->b_flags |= B_MALLOC;
3047 bufmallocspace += mbsize;
3053 * If the buffer is growing on its other-than-first
3054 * allocation, then we revert to the page-allocation
3057 if (bp->b_flags & B_MALLOC) {
3058 origbuf = bp->b_data;
3059 origbufsize = bp->b_bufsize;
3060 bp->b_data = bp->b_kvabase;
3061 if (bp->b_bufsize) {
3062 bufmallocspace -= bp->b_bufsize;
3066 bp->b_flags &= ~B_MALLOC;
3067 newbsize = round_page(newbsize);
3071 (vm_offset_t) bp->b_data + bp->b_bufsize,
3072 (vm_offset_t) bp->b_data + newbsize);
3074 bcopy(origbuf, bp->b_data, origbufsize);
3075 kfree(origbuf, M_BIOBUF);
3082 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3083 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3084 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3085 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3087 if (bp->b_flags & B_MALLOC)
3088 panic("allocbuf: VMIO buffer can't be malloced");
3090 * Set B_CACHE initially if buffer is 0 length or will become
3093 if (size == 0 || bp->b_bufsize == 0)
3094 bp->b_flags |= B_CACHE;
3096 if (newbsize < bp->b_bufsize) {
3098 * DEV_BSIZE aligned new buffer size is less then the
3099 * DEV_BSIZE aligned existing buffer size. Figure out
3100 * if we have to remove any pages.
3102 if (desiredpages < bp->b_xio.xio_npages) {
3103 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3105 * the page is not freed here -- it
3106 * is the responsibility of
3107 * vnode_pager_setsize
3109 m = bp->b_xio.xio_pages[i];
3110 KASSERT(m != bogus_page,
3111 ("allocbuf: bogus page found"));
3112 while (vm_page_sleep_busy(m, TRUE, "biodep"))
3115 bp->b_xio.xio_pages[i] = NULL;
3116 vm_page_unwire(m, 0);
3118 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3119 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3120 bp->b_xio.xio_npages = desiredpages;
3122 } else if (size > bp->b_bcount) {
3124 * We are growing the buffer, possibly in a
3125 * byte-granular fashion.
3133 * Step 1, bring in the VM pages from the object,
3134 * allocating them if necessary. We must clear
3135 * B_CACHE if these pages are not valid for the
3136 * range covered by the buffer.
3138 * critical section protection is required to protect
3139 * against interrupts unbusying and freeing pages
3140 * between our vm_page_lookup() and our
3141 * busycheck/wiring call.
3147 while (bp->b_xio.xio_npages < desiredpages) {
3151 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
3152 if ((m = vm_page_lookup(obj, pi)) == NULL) {
3154 * note: must allocate system pages
3155 * since blocking here could intefere
3156 * with paging I/O, no matter which
3159 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
3163 bp->b_flags &= ~B_CACHE;
3164 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3165 ++bp->b_xio.xio_npages;
3171 * We found a page. If we have to sleep on it,
3172 * retry because it might have gotten freed out
3175 * We can only test PG_BUSY here. Blocking on
3176 * m->busy might lead to a deadlock:
3178 * vm_fault->getpages->cluster_read->allocbuf
3182 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
3184 vm_page_flag_clear(m, PG_ZERO);
3186 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3187 ++bp->b_xio.xio_npages;
3188 if (bp->b_act_count < m->act_count)
3189 bp->b_act_count = m->act_count;
3194 * Step 2. We've loaded the pages into the buffer,
3195 * we have to figure out if we can still have B_CACHE
3196 * set. Note that B_CACHE is set according to the
3197 * byte-granular range ( bcount and size ), not the
3198 * aligned range ( newbsize ).
3200 * The VM test is against m->valid, which is DEV_BSIZE
3201 * aligned. Needless to say, the validity of the data
3202 * needs to also be DEV_BSIZE aligned. Note that this
3203 * fails with NFS if the server or some other client
3204 * extends the file's EOF. If our buffer is resized,
3205 * B_CACHE may remain set! XXX
3208 toff = bp->b_bcount;
3209 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3211 while ((bp->b_flags & B_CACHE) && toff < size) {
3214 if (tinc > (size - toff))
3217 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3225 bp->b_xio.xio_pages[pi]
3232 * Step 3, fixup the KVM pmap. Remember that
3233 * bp->b_data is relative to bp->b_loffset, but
3234 * bp->b_loffset may be offset into the first page.
3237 bp->b_data = (caddr_t)
3238 trunc_page((vm_offset_t)bp->b_data);
3240 (vm_offset_t)bp->b_data,
3241 bp->b_xio.xio_pages,
3242 bp->b_xio.xio_npages
3244 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3245 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3249 /* adjust space use on already-dirty buffer */
3250 if (bp->b_flags & B_DELWRI) {
3251 dirtybufspace += newbsize - bp->b_bufsize;
3252 if (bp->b_flags & B_HEAVY)
3253 dirtybufspacehw += newbsize - bp->b_bufsize;
3255 if (newbsize < bp->b_bufsize)
3257 bp->b_bufsize = newbsize; /* actual buffer allocation */
3258 bp->b_bcount = size; /* requested buffer size */
3265 * Wait for buffer I/O completion, returning error status. B_EINTR
3266 * is converted into an EINTR error but not cleared (since a chain
3267 * of biowait() calls may occur).
3269 * On return bpdone() will have been called but the buffer will remain
3270 * locked and will not have been brelse()'d.
3272 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3273 * likely still in progress on return.
3275 * NOTE! This operation is on a BIO, not a BUF.
3277 * NOTE! BIO_DONE is cleared by vn_strategy()
3282 _biowait(struct bio *bio, const char *wmesg, int to)
3284 struct buf *bp = bio->bio_buf;
3289 KKASSERT(bio == &bp->b_bio1);
3291 flags = bio->bio_flags;
3292 if (flags & BIO_DONE)
3294 tsleep_interlock(bio, 0);
3295 nflags = flags | BIO_WANT;
3296 tsleep_interlock(bio, 0);
3297 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3299 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3300 else if (bp->b_cmd == BUF_CMD_READ)
3301 error = tsleep(bio, PINTERLOCKED, "biord", to);
3303 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3305 kprintf("tsleep error biowait %d\n", error);
3315 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3316 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3317 if (bp->b_flags & B_EINTR)
3319 if (bp->b_flags & B_ERROR)
3320 return (bp->b_error ? bp->b_error : EIO);
3325 biowait(struct bio *bio, const char *wmesg)
3327 return(_biowait(bio, wmesg, 0));
3331 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3333 return(_biowait(bio, wmesg, to));
3337 * This associates a tracking count with an I/O. vn_strategy() and
3338 * dev_dstrategy() do this automatically but there are a few cases
3339 * where a vnode or device layer is bypassed when a block translation
3340 * is cached. In such cases bio_start_transaction() may be called on
3341 * the bypassed layers so the system gets an I/O in progress indication
3342 * for those higher layers.
3345 bio_start_transaction(struct bio *bio, struct bio_track *track)
3347 bio->bio_track = track;
3348 bio_track_ref(track);
3352 * Initiate I/O on a vnode.
3355 vn_strategy(struct vnode *vp, struct bio *bio)
3357 struct bio_track *track;
3359 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
3360 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
3361 track = &vp->v_track_read;
3363 track = &vp->v_track_write;
3364 KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3365 bio->bio_track = track;
3366 bio_track_ref(track);
3367 vop_strategy(*vp->v_ops, vp, bio);
3373 * Finish I/O on a buffer after all BIOs have been processed.
3374 * Called when the bio chain is exhausted or by biowait. If called
3375 * by biowait, elseit is typically 0.
3377 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3378 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3379 * assuming B_INVAL is clear.
3381 * For the VMIO case, we set B_CACHE if the op was a read and no
3382 * read error occured, or if the op was a write. B_CACHE is never
3383 * set if the buffer is invalid or otherwise uncacheable.
3385 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3386 * initiator to leave B_INVAL set to brelse the buffer out of existance
3387 * in the biodone routine.
3390 bpdone(struct buf *bp, int elseit)
3394 KASSERT(BUF_REFCNTNB(bp) > 0,
3395 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3396 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3397 ("biodone: bp %p already done!", bp));
3400 * No more BIOs are left. All completion functions have been dealt
3401 * with, now we clean up the buffer.
3404 bp->b_cmd = BUF_CMD_DONE;
3407 * Only reads and writes are processed past this point.
3409 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3410 if (cmd == BUF_CMD_FREEBLKS)
3411 bp->b_flags |= B_NOCACHE;
3418 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3419 * a lot worse. XXX - move this above the clearing of b_cmd
3421 if (LIST_FIRST(&bp->b_dep) != NULL)
3425 * A failed write must re-dirty the buffer unless B_INVAL
3426 * was set. Only applicable to normal buffers (with VPs).
3427 * vinum buffers may not have a vp.
3429 if (cmd == BUF_CMD_WRITE &&
3430 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3431 bp->b_flags &= ~B_NOCACHE;
3436 if (bp->b_flags & B_VMIO) {
3442 struct vnode *vp = bp->b_vp;
3446 #if defined(VFS_BIO_DEBUG)
3447 if (vp->v_auxrefs == 0)
3448 panic("biodone: zero vnode hold count");
3449 if ((vp->v_flag & VOBJBUF) == 0)
3450 panic("biodone: vnode is not setup for merged cache");
3453 foff = bp->b_loffset;
3454 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3455 KASSERT(obj != NULL, ("biodone: missing VM object"));
3457 #if defined(VFS_BIO_DEBUG)
3458 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3459 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3460 obj->paging_in_progress, bp->b_xio.xio_npages);
3465 * Set B_CACHE if the op was a normal read and no error
3466 * occured. B_CACHE is set for writes in the b*write()
3469 iosize = bp->b_bcount - bp->b_resid;
3470 if (cmd == BUF_CMD_READ &&
3471 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3472 bp->b_flags |= B_CACHE;
3477 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3481 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3486 * cleanup bogus pages, restoring the originals. Since
3487 * the originals should still be wired, we don't have
3488 * to worry about interrupt/freeing races destroying
3489 * the VM object association.
3491 m = bp->b_xio.xio_pages[i];
3492 if (m == bogus_page) {
3494 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3496 panic("biodone: page disappeared");
3497 bp->b_xio.xio_pages[i] = m;
3498 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3499 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3501 #if defined(VFS_BIO_DEBUG)
3502 if (OFF_TO_IDX(foff) != m->pindex) {
3503 kprintf("biodone: foff(%lu)/m->pindex(%ld) "
3505 (unsigned long)foff, (long)m->pindex);
3510 * In the write case, the valid and clean bits are
3511 * already changed correctly (see bdwrite()), so we
3512 * only need to do this here in the read case.
3514 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3515 vfs_clean_one_page(bp, i, m);
3517 vm_page_flag_clear(m, PG_ZERO);
3520 * when debugging new filesystems or buffer I/O
3521 * methods, this is the most common error that pops
3522 * up. if you see this, you have not set the page
3523 * busy flag correctly!!!
3526 kprintf("biodone: page busy < 0, "
3527 "pindex: %d, foff: 0x(%x,%x), "
3528 "resid: %d, index: %d\n",
3529 (int) m->pindex, (int)(foff >> 32),
3530 (int) foff & 0xffffffff, resid, i);
3531 if (!vn_isdisk(vp, NULL))
3532 kprintf(" iosize: %ld, loffset: %lld, "
3533 "flags: 0x%08x, npages: %d\n",
3534 bp->b_vp->v_mount->mnt_stat.f_iosize,
3535 (long long)bp->b_loffset,
3536 bp->b_flags, bp->b_xio.xio_npages);
3538 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3539 (long long)bp->b_loffset,
3540 bp->b_flags, bp->b_xio.xio_npages);
3541 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3542 m->valid, m->dirty, m->wire_count);
3543 panic("biodone: page busy < 0");
3545 vm_page_io_finish(m);
3546 vm_object_pip_subtract(obj, 1);
3547 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3551 vm_object_pip_wakeupn(obj, 0);
3557 * Finish up by releasing the buffer. There are no more synchronous
3558 * or asynchronous completions, those were handled by bio_done
3562 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
3573 biodone(struct bio *bio)
3575 struct buf *bp = bio->bio_buf;
3577 runningbufwakeup(bp);
3580 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3583 biodone_t *done_func;
3584 struct bio_track *track;
3587 * BIO tracking. Most but not all BIOs are tracked.
3589 if ((track = bio->bio_track) != NULL) {
3590 bio_track_rel(track);
3591 bio->bio_track = NULL;
3595 * A bio_done function terminates the loop. The function
3596 * will be responsible for any further chaining and/or
3597 * buffer management.
3599 * WARNING! The done function can deallocate the buffer!
3601 if ((done_func = bio->bio_done) != NULL) {
3602 bio->bio_done = NULL;
3606 bio = bio->bio_prev;
3610 * If we've run out of bio's do normal [a]synchronous completion.
3616 * Synchronous biodone - this terminates a synchronous BIO.
3618 * bpdone() is called with elseit=FALSE, leaving the buffer completed
3619 * but still locked. The caller must brelse() the buffer after waiting
3623 biodone_sync(struct bio *bio)
3625 struct buf *bp = bio->bio_buf;
3629 KKASSERT(bio == &bp->b_bio1);
3633 flags = bio->bio_flags;
3634 nflags = (flags | BIO_DONE) & ~BIO_WANT;
3636 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3637 if (flags & BIO_WANT)
3647 * This routine is called in lieu of iodone in the case of
3648 * incomplete I/O. This keeps the busy status for pages
3652 vfs_unbusy_pages(struct buf *bp)
3656 runningbufwakeup(bp);
3657 if (bp->b_flags & B_VMIO) {
3658 struct vnode *vp = bp->b_vp;
3663 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3664 vm_page_t m = bp->b_xio.xio_pages[i];
3667 * When restoring bogus changes the original pages
3668 * should still be wired, so we are in no danger of
3669 * losing the object association and do not need
3670 * critical section protection particularly.
3672 if (m == bogus_page) {
3673 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3675 panic("vfs_unbusy_pages: page missing");
3677 bp->b_xio.xio_pages[i] = m;
3678 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3679 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3681 vm_object_pip_subtract(obj, 1);
3682 vm_page_flag_clear(m, PG_ZERO);
3683 vm_page_io_finish(m);
3685 vm_object_pip_wakeupn(obj, 0);
3692 * This routine is called before a device strategy routine.
3693 * It is used to tell the VM system that paging I/O is in
3694 * progress, and treat the pages associated with the buffer
3695 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3696 * flag is handled to make sure that the object doesn't become
3699 * Since I/O has not been initiated yet, certain buffer flags
3700 * such as B_ERROR or B_INVAL may be in an inconsistant state
3701 * and should be ignored.
3704 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3707 struct lwp *lp = curthread->td_lwp;
3710 * The buffer's I/O command must already be set. If reading,
3711 * B_CACHE must be 0 (double check against callers only doing
3712 * I/O when B_CACHE is 0).
3714 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3715 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3717 if (bp->b_flags & B_VMIO) {
3721 KASSERT(bp->b_loffset != NOOFFSET,
3722 ("vfs_busy_pages: no buffer offset"));
3725 * Loop until none of the pages are busy.
3728 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3729 vm_page_t m = bp->b_xio.xio_pages[i];
3731 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3736 * Setup for I/O, soft-busy the page right now because
3737 * the next loop may block.
3739 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3740 vm_page_t m = bp->b_xio.xio_pages[i];
3742 vm_page_flag_clear(m, PG_ZERO);
3743 if ((bp->b_flags & B_CLUSTER) == 0) {
3744 vm_object_pip_add(obj, 1);
3745 vm_page_io_start(m);
3750 * Adjust protections for I/O and do bogus-page mapping.
3751 * Assume that vm_page_protect() can block (it can block
3752 * if VM_PROT_NONE, don't take any chances regardless).
3754 * In particularly note that for writes we must incorporate
3755 * page dirtyness from the VM system into the buffer's
3758 * For reads we theoretically must incorporate page dirtyness
3759 * from the VM system to determine if the page needs bogus
3760 * replacement, but we shortcut the test by simply checking
3761 * that all m->valid bits are set, indicating that the page
3762 * is fully valid and does not need to be re-read. For any
3763 * VM system dirtyness the page will also be fully valid
3764 * since it was mapped at one point.
3767 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3768 vm_page_t m = bp->b_xio.xio_pages[i];
3770 vm_page_flag_clear(m, PG_ZERO); /* XXX */
3771 if (bp->b_cmd == BUF_CMD_WRITE) {
3773 * When readying a vnode-backed buffer for
3774 * a write we must zero-fill any invalid
3775 * portions of the backing VM pages, mark
3776 * it valid and clear related dirty bits.
3778 * vfs_clean_one_page() incorporates any
3779 * VM dirtyness and updates the b_dirtyoff
3780 * range (after we've made the page RO).
3782 * It is also expected that the pmap modified
3783 * bit has already been cleared by the
3784 * vm_page_protect(). We may not be able
3785 * to clear all dirty bits for a page if it
3786 * was also memory mapped (NFS).
3788 vm_page_protect(m, VM_PROT_READ);
3789 vfs_clean_one_page(bp, i, m);
3790 } else if (m->valid == VM_PAGE_BITS_ALL) {
3792 * When readying a vnode-backed buffer for
3793 * read we must replace any dirty pages with
3794 * a bogus page so dirty data is not destroyed
3795 * when filling gaps.
3797 * To avoid testing whether the page is
3798 * dirty we instead test that the page was
3799 * at some point mapped (m->valid fully
3800 * valid) with the understanding that
3801 * this also covers the dirty case.
3803 bp->b_xio.xio_pages[i] = bogus_page;
3805 } else if (m->valid & m->dirty) {
3807 * This case should not occur as partial
3808 * dirtyment can only happen if the buffer
3809 * is B_CACHE, and this code is not entered
3810 * if the buffer is B_CACHE.
3812 kprintf("Warning: vfs_busy_pages - page not "
3813 "fully valid! loff=%jx bpf=%08x "
3814 "idx=%d val=%02x dir=%02x\n",
3815 (intmax_t)bp->b_loffset, bp->b_flags,
3816 i, m->valid, m->dirty);
3817 vm_page_protect(m, VM_PROT_NONE);
3820 * The page is not valid and can be made
3823 vm_page_protect(m, VM_PROT_NONE);
3827 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3828 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3833 * This is the easiest place to put the process accounting for the I/O
3837 if (bp->b_cmd == BUF_CMD_READ)
3838 lp->lwp_ru.ru_inblock++;
3840 lp->lwp_ru.ru_oublock++;
3847 * Tell the VM system that the pages associated with this buffer
3848 * are clean. This is used for delayed writes where the data is
3849 * going to go to disk eventually without additional VM intevention.
3851 * Note that while we only really need to clean through to b_bcount, we
3852 * just go ahead and clean through to b_bufsize.
3855 vfs_clean_pages(struct buf *bp)
3860 if ((bp->b_flags & B_VMIO) == 0)
3863 KASSERT(bp->b_loffset != NOOFFSET,
3864 ("vfs_clean_pages: no buffer offset"));
3866 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3867 m = bp->b_xio.xio_pages[i];
3868 vfs_clean_one_page(bp, i, m);
3873 * vfs_clean_one_page:
3875 * Set the valid bits and clear the dirty bits in a page within a
3876 * buffer. The range is restricted to the buffer's size and the
3877 * buffer's logical offset might index into the first page.
3879 * The caller has busied or soft-busied the page and it is not mapped,
3880 * test and incorporate the dirty bits into b_dirtyoff/end before
3881 * clearing them. Note that we need to clear the pmap modified bits
3882 * after determining the the page was dirty, vm_page_set_validclean()
3883 * does not do it for us.
3885 * This routine is typically called after a read completes (dirty should
3886 * be zero in that case as we are not called on bogus-replace pages),
3887 * or before a write is initiated.
3890 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m)
3898 * Calculate offset range within the page but relative to buffer's
3899 * loffset. loffset might be offset into the first page.
3901 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
3902 bcount = bp->b_bcount + xoff; /* offset adjusted */
3908 soff = (pageno << PAGE_SHIFT);
3909 eoff = soff + PAGE_SIZE;
3917 * Test dirty bits and adjust b_dirtyoff/end.
3919 * If dirty pages are incorporated into the bp any prior
3920 * B_NEEDCOMMIT state (NFS) must be cleared because the
3921 * caller has not taken into account the new dirty data.
3923 * If the page was memory mapped the dirty bits might go beyond the
3924 * end of the buffer, but we can't really make the assumption that
3925 * a file EOF straddles the buffer (even though this is the case for
3926 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
3927 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
3928 * This also saves some console spam.
3930 vm_page_test_dirty(m);
3932 pmap_clear_modify(m);
3933 if ((bp->b_flags & B_NEEDCOMMIT) &&
3934 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) {
3935 kprintf("Warning: vfs_clean_one_page: bp %p "
3936 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT\n",
3937 bp, (intmax_t)bp->b_loffset, bp->b_bcount,
3939 bp->b_flags &= ~B_NEEDCOMMIT;
3941 if (bp->b_dirtyoff > soff - xoff)
3942 bp->b_dirtyoff = soff - xoff;
3943 if (bp->b_dirtyend < eoff - xoff)
3944 bp->b_dirtyend = eoff - xoff;
3948 * Set related valid bits, clear related dirty bits.
3949 * Does not mess with the pmap modified bit.
3951 * WARNING! We cannot just clear all of m->dirty here as the
3952 * buffer cache buffers may use a DEV_BSIZE'd aligned
3953 * block size, or have an odd size (e.g. NFS at file EOF).
3954 * The putpages code can clear m->dirty to 0.
3956 * If a VOP_WRITE generates a buffer cache buffer which
3957 * covers the same space as mapped writable pages the
3958 * buffer flush might not be able to clear all the dirty
3959 * bits and still require a putpages from the VM system
3962 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff);
3968 * Clear a buffer. This routine essentially fakes an I/O, so we need
3969 * to clear B_ERROR and B_INVAL.
3971 * Note that while we only theoretically need to clear through b_bcount,
3972 * we go ahead and clear through b_bufsize.
3976 vfs_bio_clrbuf(struct buf *bp)
3980 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3981 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
3982 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3983 (bp->b_loffset & PAGE_MASK) == 0) {
3984 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3985 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3989 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3990 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3991 bzero(bp->b_data, bp->b_bufsize);
3992 bp->b_xio.xio_pages[0]->valid |= mask;
3998 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3999 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
4000 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
4001 ea = (caddr_t)(vm_offset_t)ulmin(
4002 (u_long)(vm_offset_t)ea,
4003 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
4004 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4005 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
4007 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
4008 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
4012 for (; sa < ea; sa += DEV_BSIZE, j++) {
4013 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
4014 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
4015 bzero(sa, DEV_BSIZE);
4018 bp->b_xio.xio_pages[i]->valid |= mask;
4019 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
4028 * vm_hold_load_pages:
4030 * Load pages into the buffer's address space. The pages are
4031 * allocated from the kernel object in order to reduce interference
4032 * with the any VM paging I/O activity. The range of loaded
4033 * pages will be wired.
4035 * If a page cannot be allocated, the 'pagedaemon' is woken up to
4036 * retrieve the full range (to - from) of pages.
4040 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4046 to = round_page(to);
4047 from = round_page(from);
4048 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4053 * Note: must allocate system pages since blocking here
4054 * could intefere with paging I/O, no matter which
4057 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT,
4058 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
4061 p->valid = VM_PAGE_BITS_ALL;
4062 vm_page_flag_clear(p, PG_ZERO);
4063 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
4064 bp->b_xio.xio_pages[index] = p;
4071 bp->b_xio.xio_npages = index;
4075 * Allocate pages for a buffer cache buffer.
4077 * Under extremely severe memory conditions even allocating out of the
4078 * system reserve can fail. If this occurs we must allocate out of the
4079 * interrupt reserve to avoid a deadlock with the pageout daemon.
4081 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
4082 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
4083 * against the pageout daemon if pages are not freed from other sources.
4087 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit)
4092 * Try a normal allocation, allow use of system reserve.
4094 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
4099 * The normal allocation failed and we clearly have a page
4100 * deficit. Try to reclaim some clean VM pages directly
4101 * from the buffer cache.
4103 vm_pageout_deficit += deficit;
4107 * We may have blocked, the caller will know what to do if the
4110 if (vm_page_lookup(obj, pg))
4114 * Allocate and allow use of the interrupt reserve.
4116 * If after all that we still can't allocate a VM page we are
4117 * in real trouble, but we slog on anyway hoping that the system
4120 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
4121 VM_ALLOC_INTERRUPT);
4123 if (vm_page_count_severe()) {
4124 kprintf("bio_page_alloc: WARNING emergency page "
4129 kprintf("bio_page_alloc: WARNING emergency page "
4130 "allocation failed\n");
4137 * vm_hold_free_pages:
4139 * Return pages associated with the buffer back to the VM system.
4141 * The range of pages underlying the buffer's address space will
4142 * be unmapped and un-wired.
4145 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4149 int index, newnpages;
4151 from = round_page(from);
4152 to = round_page(to);
4153 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4155 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4156 p = bp->b_xio.xio_pages[index];
4157 if (p && (index < bp->b_xio.xio_npages)) {
4159 kprintf("vm_hold_free_pages: doffset: %lld, "
4161 (long long)bp->b_bio2.bio_offset,
4162 (long long)bp->b_loffset);
4164 bp->b_xio.xio_pages[index] = NULL;
4167 vm_page_unwire(p, 0);
4171 bp->b_xio.xio_npages = newnpages;
4177 * Map a user buffer into KVM via a pbuf. On return the buffer's
4178 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4182 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4193 * bp had better have a command and it better be a pbuf.
4195 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4196 KKASSERT(bp->b_flags & B_PAGING);
4202 * Map the user data into KVM. Mappings have to be page-aligned.
4204 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4207 vmprot = VM_PROT_READ;
4208 if (bp->b_cmd == BUF_CMD_READ)
4209 vmprot |= VM_PROT_WRITE;
4211 while (addr < udata + bytes) {
4213 * Do the vm_fault if needed; do the copy-on-write thing
4214 * when reading stuff off device into memory.
4216 * vm_fault_page*() returns a held VM page.
4218 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4219 va = trunc_page(va);
4221 m = vm_fault_page_quick(va, vmprot, &error);
4223 for (i = 0; i < pidx; ++i) {
4224 vm_page_unhold(bp->b_xio.xio_pages[i]);
4225 bp->b_xio.xio_pages[i] = NULL;
4229 bp->b_xio.xio_pages[pidx] = m;
4235 * Map the page array and set the buffer fields to point to
4236 * the mapped data buffer.
4238 if (pidx > btoc(MAXPHYS))
4239 panic("vmapbuf: mapped more than MAXPHYS");
4240 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4242 bp->b_xio.xio_npages = pidx;
4243 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4244 bp->b_bcount = bytes;
4245 bp->b_bufsize = bytes;
4252 * Free the io map PTEs associated with this IO operation.
4253 * We also invalidate the TLB entries and restore the original b_addr.
4256 vunmapbuf(struct buf *bp)
4261 KKASSERT(bp->b_flags & B_PAGING);
4263 npages = bp->b_xio.xio_npages;
4264 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4265 for (pidx = 0; pidx < npages; ++pidx) {
4266 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4267 bp->b_xio.xio_pages[pidx] = NULL;
4269 bp->b_xio.xio_npages = 0;
4270 bp->b_data = bp->b_kvabase;
4274 * Scan all buffers in the system and issue the callback.
4277 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4283 for (n = 0; n < nbuf; ++n) {
4284 if ((error = callback(&buf[n], info)) < 0) {
4294 * print out statistics from the current status of the buffer pool
4295 * this can be toggeled by the system control option debug.syncprt
4304 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
4305 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
4307 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
4309 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4312 TAILQ_FOREACH(bp, dp, b_freelist) {
4313 counts[bp->b_bufsize/PAGE_SIZE]++;
4317 kprintf("%s: total-%d", bname[i], count);
4318 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4320 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
4328 DB_SHOW_COMMAND(buffer, db_show_buffer)
4331 struct buf *bp = (struct buf *)addr;
4334 db_printf("usage: show buffer <addr>\n");
4338 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4339 db_printf("b_cmd = %d\n", bp->b_cmd);
4340 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4341 "b_resid = %d\n, b_data = %p, "
4342 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4343 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4345 (long long)bp->b_bio2.bio_offset,
4346 (long long)(bp->b_bio2.bio_next ?
4347 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4348 if (bp->b_xio.xio_npages) {
4350 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4351 bp->b_xio.xio_npages);
4352 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4354 m = bp->b_xio.xio_pages[i];
4355 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4356 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4357 if ((i + 1) < bp->b_xio.xio_npages)