2 * Copyright (c) 1994,1997 John S. Dyson
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice immediately at the beginning of the file, without modification,
10 * this list of conditions, and the following disclaimer.
11 * 2. Absolutely no warranty of function or purpose is made by the author
14 * $FreeBSD: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $
15 * $DragonFly: src/sys/kern/vfs_bio.c,v 1.115 2008/08/13 11:02:31 swildner Exp $
19 * this file contains a new buffer I/O scheme implementing a coherent
20 * VM object and buffer cache scheme. Pains have been taken to make
21 * sure that the performance degradation associated with schemes such
22 * as this is not realized.
24 * Author: John S. Dyson
25 * Significant help during the development and debugging phases
26 * had been provided by David Greenman, also of the FreeBSD core team.
28 * see man buf(9) for more info.
31 #include <sys/param.h>
32 #include <sys/systm.h>
35 #include <sys/eventhandler.h>
37 #include <sys/malloc.h>
38 #include <sys/mount.h>
39 #include <sys/kernel.h>
40 #include <sys/kthread.h>
42 #include <sys/reboot.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sysctl.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
47 #include <sys/dsched.h>
50 #include <vm/vm_param.h>
51 #include <vm/vm_kern.h>
52 #include <vm/vm_pageout.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_map.h>
57 #include <vm/vm_pager.h>
58 #include <vm/swap_pager.h>
61 #include <sys/thread2.h>
62 #include <sys/spinlock2.h>
63 #include <sys/mplock2.h>
64 #include <vm/vm_page2.h>
75 BQUEUE_NONE, /* not on any queue */
76 BQUEUE_LOCKED, /* locked buffers */
77 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
78 BQUEUE_DIRTY, /* B_DELWRI buffers */
79 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
80 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
81 BQUEUE_EMPTY, /* empty buffer headers */
83 BUFFER_QUEUES /* number of buffer queues */
86 typedef enum bufq_type bufq_type_t;
88 #define BD_WAKE_SIZE 16384
89 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
91 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
92 static struct spinlock bufqspin = SPINLOCK_INITIALIZER(&bufqspin);
93 static struct spinlock bufcspin = SPINLOCK_INITIALIZER(&bufcspin);
95 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
97 struct buf *buf; /* buffer header pool */
99 static void vfs_clean_pages(struct buf *bp);
100 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
101 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m);
102 static void vfs_vmio_release(struct buf *bp);
103 static int flushbufqueues(bufq_type_t q);
104 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);
106 static void bd_signal(int totalspace);
107 static void buf_daemon(void);
108 static void buf_daemon_hw(void);
111 * bogus page -- for I/O to/from partially complete buffers
112 * this is a temporary solution to the problem, but it is not
113 * really that bad. it would be better to split the buffer
114 * for input in the case of buffers partially already in memory,
115 * but the code is intricate enough already.
117 vm_page_t bogus_page;
120 * These are all static, but make the ones we export globals so we do
121 * not need to use compiler magic.
123 int bufspace; /* locked by buffer_map */
125 static int bufmallocspace; /* atomic ops */
126 int maxbufmallocspace, lobufspace, hibufspace;
127 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
128 static int lorunningspace;
129 static int hirunningspace;
130 static int runningbufreq; /* locked by bufcspin */
131 static int dirtybufspace; /* locked by bufcspin */
132 static int dirtybufcount; /* locked by bufcspin */
133 static int dirtybufspacehw; /* locked by bufcspin */
134 static int dirtybufcounthw; /* locked by bufcspin */
135 static int runningbufspace; /* locked by bufcspin */
136 static int runningbufcount; /* locked by bufcspin */
139 static int getnewbufcalls;
140 static int getnewbufrestarts;
141 static int recoverbufcalls;
142 static int needsbuffer; /* locked by bufcspin */
143 static int bd_request; /* locked by bufcspin */
144 static int bd_request_hw; /* locked by bufcspin */
145 static u_int bd_wake_ary[BD_WAKE_SIZE];
146 static u_int bd_wake_index;
147 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */
148 static int debug_commit;
150 static struct thread *bufdaemon_td;
151 static struct thread *bufdaemonhw_td;
155 * Sysctls for operational control of the buffer cache.
157 SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
158 "Number of dirty buffers to flush before bufdaemon becomes inactive");
159 SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
160 "High watermark used to trigger explicit flushing of dirty buffers");
161 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
162 "Minimum amount of buffer space required for active I/O");
163 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
164 "Maximum amount of buffer space to usable for active I/O");
165 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
166 "Recycle pages to active or inactive queue transition pt 0-64");
168 * Sysctls determining current state of the buffer cache.
170 SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
171 "Total number of buffers in buffer cache");
172 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
173 "Pending bytes of dirty buffers (all)");
174 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
175 "Pending bytes of dirty buffers (heavy weight)");
176 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
177 "Pending number of dirty buffers");
178 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
179 "Pending number of dirty buffers (heavy weight)");
180 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
181 "I/O bytes currently in progress due to asynchronous writes");
182 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
183 "I/O buffers currently in progress due to asynchronous writes");
184 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
185 "Hard limit on maximum amount of memory usable for buffer space");
186 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
187 "Soft limit on maximum amount of memory usable for buffer space");
188 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
189 "Minimum amount of memory to reserve for system buffer space");
190 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
191 "Amount of memory available for buffers");
192 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
193 0, "Maximum amount of memory reserved for buffers using malloc");
194 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
195 "Amount of memory left for buffers using malloc-scheme");
196 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
197 "New buffer header acquisition requests");
198 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
199 0, "New buffer header acquisition restarts");
200 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
201 "Recover VM space in an emergency");
202 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
203 "Buffer acquisition restarts due to fragmented buffer map");
204 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
205 "Amount of time KVA space was deallocated in an arbitrary buffer");
206 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
207 "Amount of time buffer re-use operations were successful");
208 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, "");
209 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
210 "sizeof(struct buf)");
212 char *buf_wmesg = BUF_WMESG;
214 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
215 #define VFS_BIO_NEED_UNUSED02 0x02
216 #define VFS_BIO_NEED_UNUSED04 0x04
217 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
222 * Called when buffer space is potentially available for recovery.
223 * getnewbuf() will block on this flag when it is unable to free
224 * sufficient buffer space. Buffer space becomes recoverable when
225 * bp's get placed back in the queues.
231 * If someone is waiting for BUF space, wake them up. Even
232 * though we haven't freed the kva space yet, the waiting
233 * process will be able to now.
235 spin_lock(&bufcspin);
236 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
237 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
238 spin_unlock(&bufcspin);
239 wakeup(&needsbuffer);
241 spin_unlock(&bufcspin);
248 * Accounting for I/O in progress.
252 runningbufwakeup(struct buf *bp)
257 if ((totalspace = bp->b_runningbufspace) != 0) {
258 spin_lock(&bufcspin);
259 runningbufspace -= totalspace;
261 bp->b_runningbufspace = 0;
264 * see waitrunningbufspace() for limit test.
266 limit = hirunningspace * 4 / 6;
267 if (runningbufreq && runningbufspace <= limit) {
269 spin_unlock(&bufcspin);
270 wakeup(&runningbufreq);
272 spin_unlock(&bufcspin);
274 bd_signal(totalspace);
281 * Called when a buffer has been added to one of the free queues to
282 * account for the buffer and to wakeup anyone waiting for free buffers.
283 * This typically occurs when large amounts of metadata are being handled
284 * by the buffer cache ( else buffer space runs out first, usually ).
291 spin_lock(&bufcspin);
293 needsbuffer &= ~VFS_BIO_NEED_ANY;
294 spin_unlock(&bufcspin);
295 wakeup(&needsbuffer);
297 spin_unlock(&bufcspin);
302 * waitrunningbufspace()
304 * Wait for the amount of running I/O to drop to hirunningspace * 4 / 6.
305 * This is the point where write bursting stops so we don't want to wait
306 * for the running amount to drop below it (at least if we still want bioq
309 * The caller may be using this function to block in a tight loop, we
310 * must block while runningbufspace is greater then or equal to
311 * hirunningspace * 4 / 6.
313 * And even with that it may not be enough, due to the presence of
314 * B_LOCKED dirty buffers, so also wait for at least one running buffer
318 waitrunningbufspace(void)
320 int limit = hirunningspace * 4 / 6;
323 spin_lock(&bufcspin);
324 if (runningbufspace > limit) {
325 while (runningbufspace > limit) {
327 ssleep(&runningbufreq, &bufcspin, 0, "wdrn1", 0);
329 spin_unlock(&bufcspin);
330 } else if (runningbufspace > limit / 2) {
332 spin_unlock(&bufcspin);
333 tsleep(&dummy, 0, "wdrn2", 1);
335 spin_unlock(&bufcspin);
340 * buf_dirty_count_severe:
342 * Return true if we have too many dirty buffers.
345 buf_dirty_count_severe(void)
347 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
348 dirtybufcount >= nbuf / 2);
352 * Return true if the amount of running I/O is severe and BIOQ should
356 buf_runningbufspace_severe(void)
358 return (runningbufspace >= hirunningspace * 4 / 6);
362 * vfs_buf_test_cache:
364 * Called when a buffer is extended. This function clears the B_CACHE
365 * bit if the newly extended portion of the buffer does not contain
368 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
369 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
370 * them while a clean buffer was present.
374 vfs_buf_test_cache(struct buf *bp,
375 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
378 if (bp->b_flags & B_CACHE) {
379 int base = (foff + off) & PAGE_MASK;
380 if (vm_page_is_valid(m, base, size) == 0)
381 bp->b_flags &= ~B_CACHE;
388 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
397 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
400 if (bd_request == 0 &&
401 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
402 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
403 spin_lock(&bufcspin);
405 spin_unlock(&bufcspin);
408 if (bd_request_hw == 0 &&
409 (dirtybufspacehw > lodirtybufspace / 2 ||
410 dirtybufcounthw >= nbuf / 2)) {
411 spin_lock(&bufcspin);
413 spin_unlock(&bufcspin);
414 wakeup(&bd_request_hw);
421 * Get the buf_daemon heated up when the number of running and dirty
422 * buffers exceeds the mid-point.
424 * Return the total number of dirty bytes past the second mid point
425 * as a measure of how much excess dirty data there is in the system.
436 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
438 totalspace = runningbufspace + dirtybufspace;
439 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
441 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
442 if (totalspace >= mid2)
443 return(totalspace - mid2);
451 * Wait for the buffer cache to flush (totalspace) bytes worth of
452 * buffers, then return.
454 * Regardless this function blocks while the number of dirty buffers
455 * exceeds hidirtybufspace.
460 bd_wait(int totalspace)
465 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
468 while (totalspace > 0) {
470 if (totalspace > runningbufspace + dirtybufspace)
471 totalspace = runningbufspace + dirtybufspace;
472 count = totalspace / BKVASIZE;
473 if (count >= BD_WAKE_SIZE)
474 count = BD_WAKE_SIZE - 1;
476 spin_lock(&bufcspin);
477 i = (bd_wake_index + count) & BD_WAKE_MASK;
481 * This is not a strict interlock, so we play a bit loose
482 * with locking access to dirtybufspace*
484 tsleep_interlock(&bd_wake_ary[i], 0);
485 spin_unlock(&bufcspin);
486 tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz);
488 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
495 * This function is called whenever runningbufspace or dirtybufspace
496 * is reduced. Track threads waiting for run+dirty buffer I/O
502 bd_signal(int totalspace)
506 if (totalspace > 0) {
507 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
508 totalspace = BKVASIZE * BD_WAKE_SIZE;
509 spin_lock(&bufcspin);
510 while (totalspace > 0) {
513 if (bd_wake_ary[i]) {
515 spin_unlock(&bufcspin);
516 wakeup(&bd_wake_ary[i]);
517 spin_lock(&bufcspin);
519 totalspace -= BKVASIZE;
521 spin_unlock(&bufcspin);
526 * BIO tracking support routines.
528 * Release a ref on a bio_track. Wakeup requests are atomically released
529 * along with the last reference so bk_active will never wind up set to
536 bio_track_rel(struct bio_track *track)
544 active = track->bk_active;
545 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
549 * Full-on. Note that the wait flag is only atomically released on
550 * the 1->0 count transition.
552 * We check for a negative count transition using bit 30 since bit 31
553 * has a different meaning.
556 desired = (active & 0x7FFFFFFF) - 1;
558 desired |= active & 0x80000000;
559 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
560 if (desired & 0x40000000)
561 panic("bio_track_rel: bad count: %p\n", track);
562 if (active & 0x80000000)
566 active = track->bk_active;
571 * Wait for the tracking count to reach 0.
573 * Use atomic ops such that the wait flag is only set atomically when
574 * bk_active is non-zero.
579 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
588 if (track->bk_active == 0)
592 * Full-on. Note that the wait flag may only be atomically set if
593 * the active count is non-zero.
596 while ((active = track->bk_active) != 0) {
597 desired = active | 0x80000000;
598 tsleep_interlock(track, slp_flags);
599 if (active == desired ||
600 atomic_cmpset_int(&track->bk_active, active, desired)) {
601 error = tsleep(track, slp_flags | PINTERLOCKED,
613 * Load time initialisation of the buffer cache, called from machine
614 * dependant initialization code.
620 vm_offset_t bogus_offset;
623 /* next, make a null set of free lists */
624 for (i = 0; i < BUFFER_QUEUES; i++)
625 TAILQ_INIT(&bufqueues[i]);
627 /* finally, initialize each buffer header and stick on empty q */
628 for (i = 0; i < nbuf; i++) {
630 bzero(bp, sizeof *bp);
631 bp->b_flags = B_INVAL; /* we're just an empty header */
632 bp->b_cmd = BUF_CMD_DONE;
633 bp->b_qindex = BQUEUE_EMPTY;
635 xio_init(&bp->b_xio);
638 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
642 * maxbufspace is the absolute maximum amount of buffer space we are
643 * allowed to reserve in KVM and in real terms. The absolute maximum
644 * is nominally used by buf_daemon. hibufspace is the nominal maximum
645 * used by most other processes. The differential is required to
646 * ensure that buf_daemon is able to run when other processes might
647 * be blocked waiting for buffer space.
649 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
650 * this may result in KVM fragmentation which is not handled optimally
653 maxbufspace = nbuf * BKVASIZE;
654 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
655 lobufspace = hibufspace - MAXBSIZE;
657 lorunningspace = 512 * 1024;
658 /* hirunningspace -- see below */
661 * Limit the amount of malloc memory since it is wired permanently
662 * into the kernel space. Even though this is accounted for in
663 * the buffer allocation, we don't want the malloced region to grow
664 * uncontrolled. The malloc scheme improves memory utilization
665 * significantly on average (small) directories.
667 maxbufmallocspace = hibufspace / 20;
670 * Reduce the chance of a deadlock occuring by limiting the number
671 * of delayed-write dirty buffers we allow to stack up.
673 * We don't want too much actually queued to the device at once
674 * (XXX this needs to be per-mount!), because the buffers will
675 * wind up locked for a very long period of time while the I/O
678 hidirtybufspace = hibufspace / 2; /* dirty + running */
679 hirunningspace = hibufspace / 16; /* locked & queued to device */
680 if (hirunningspace < 1024 * 1024)
681 hirunningspace = 1024 * 1024;
686 lodirtybufspace = hidirtybufspace / 2;
689 * Maximum number of async ops initiated per buf_daemon loop. This is
690 * somewhat of a hack at the moment, we really need to limit ourselves
691 * based on the number of bytes of I/O in-transit that were initiated
695 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
696 bogus_page = vm_page_alloc(&kernel_object,
697 (bogus_offset >> PAGE_SHIFT),
699 vmstats.v_wire_count++;
704 * Initialize the embedded bio structures
707 initbufbio(struct buf *bp)
709 bp->b_bio1.bio_buf = bp;
710 bp->b_bio1.bio_prev = NULL;
711 bp->b_bio1.bio_offset = NOOFFSET;
712 bp->b_bio1.bio_next = &bp->b_bio2;
713 bp->b_bio1.bio_done = NULL;
714 bp->b_bio1.bio_flags = 0;
716 bp->b_bio2.bio_buf = bp;
717 bp->b_bio2.bio_prev = &bp->b_bio1;
718 bp->b_bio2.bio_offset = NOOFFSET;
719 bp->b_bio2.bio_next = NULL;
720 bp->b_bio2.bio_done = NULL;
721 bp->b_bio2.bio_flags = 0;
725 * Reinitialize the embedded bio structures as well as any additional
726 * translation cache layers.
729 reinitbufbio(struct buf *bp)
733 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
734 bio->bio_done = NULL;
735 bio->bio_offset = NOOFFSET;
740 * Push another BIO layer onto an existing BIO and return it. The new
741 * BIO layer may already exist, holding cached translation data.
744 push_bio(struct bio *bio)
748 if ((nbio = bio->bio_next) == NULL) {
749 int index = bio - &bio->bio_buf->b_bio_array[0];
750 if (index >= NBUF_BIO - 1) {
751 panic("push_bio: too many layers bp %p\n",
754 nbio = &bio->bio_buf->b_bio_array[index + 1];
755 bio->bio_next = nbio;
756 nbio->bio_prev = bio;
757 nbio->bio_buf = bio->bio_buf;
758 nbio->bio_offset = NOOFFSET;
759 nbio->bio_done = NULL;
760 nbio->bio_next = NULL;
762 KKASSERT(nbio->bio_done == NULL);
767 * Pop a BIO translation layer, returning the previous layer. The
768 * must have been previously pushed.
771 pop_bio(struct bio *bio)
773 return(bio->bio_prev);
777 clearbiocache(struct bio *bio)
780 bio->bio_offset = NOOFFSET;
788 * Free the KVA allocation for buffer 'bp'.
790 * Must be called from a critical section as this is the only locking for
793 * Since this call frees up buffer space, we call bufspacewakeup().
798 bfreekva(struct buf *bp)
804 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
805 vm_map_lock(&buffer_map);
806 bufspace -= bp->b_kvasize;
807 vm_map_delete(&buffer_map,
808 (vm_offset_t) bp->b_kvabase,
809 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
812 vm_map_unlock(&buffer_map);
813 vm_map_entry_release(count);
815 bp->b_kvabase = NULL;
823 * Remove the buffer from the appropriate free list.
826 _bremfree(struct buf *bp)
828 if (bp->b_qindex != BQUEUE_NONE) {
829 KASSERT(BUF_REFCNTNB(bp) == 1,
830 ("bremfree: bp %p not locked",bp));
831 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
832 bp->b_qindex = BQUEUE_NONE;
834 if (BUF_REFCNTNB(bp) <= 1)
835 panic("bremfree: removing a buffer not on a queue");
840 bremfree(struct buf *bp)
842 spin_lock(&bufqspin);
844 spin_unlock(&bufqspin);
848 bremfree_locked(struct buf *bp)
856 * Get a buffer with the specified data. Look in the cache first. We
857 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
858 * is set, the buffer is valid and we do not have to do anything ( see
864 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
868 bp = getblk(vp, loffset, size, 0, 0);
871 /* if not found in cache, do some I/O */
872 if ((bp->b_flags & B_CACHE) == 0) {
873 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
874 bp->b_cmd = BUF_CMD_READ;
875 bp->b_bio1.bio_done = biodone_sync;
876 bp->b_bio1.bio_flags |= BIO_SYNC;
877 vfs_busy_pages(vp, bp);
878 vn_strategy(vp, &bp->b_bio1);
879 return (biowait(&bp->b_bio1, "biord"));
887 * Operates like bread, but also starts asynchronous I/O on
888 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
889 * to initiating I/O . If B_CACHE is set, the buffer is valid
890 * and we do not have to do anything.
895 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
896 int *rabsize, int cnt, struct buf **bpp)
898 struct buf *bp, *rabp;
900 int rv = 0, readwait = 0;
902 *bpp = bp = getblk(vp, loffset, size, 0, 0);
904 /* if not found in cache, do some I/O */
905 if ((bp->b_flags & B_CACHE) == 0) {
906 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
907 bp->b_cmd = BUF_CMD_READ;
908 bp->b_bio1.bio_done = biodone_sync;
909 bp->b_bio1.bio_flags |= BIO_SYNC;
910 vfs_busy_pages(vp, bp);
911 vn_strategy(vp, &bp->b_bio1);
915 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
916 if (inmem(vp, *raoffset))
918 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
920 if ((rabp->b_flags & B_CACHE) == 0) {
921 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
922 rabp->b_cmd = BUF_CMD_READ;
923 vfs_busy_pages(vp, rabp);
925 vn_strategy(vp, &rabp->b_bio1);
931 rv = biowait(&bp->b_bio1, "biord");
938 * Synchronous write, waits for completion.
940 * Write, release buffer on completion. (Done by iodone
941 * if async). Do not bother writing anything if the buffer
944 * Note that we set B_CACHE here, indicating that buffer is
945 * fully valid and thus cacheable. This is true even of NFS
946 * now so we set it generally. This could be set either here
947 * or in biodone() since the I/O is synchronous. We put it
951 bwrite(struct buf *bp)
955 if (bp->b_flags & B_INVAL) {
959 if (BUF_REFCNTNB(bp) == 0)
960 panic("bwrite: buffer is not busy???");
962 /* Mark the buffer clean */
965 bp->b_flags &= ~(B_ERROR | B_EINTR);
966 bp->b_flags |= B_CACHE;
967 bp->b_cmd = BUF_CMD_WRITE;
968 bp->b_bio1.bio_done = biodone_sync;
969 bp->b_bio1.bio_flags |= BIO_SYNC;
970 vfs_busy_pages(bp->b_vp, bp);
973 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
974 * valid for vnode-backed buffers.
976 bsetrunningbufspace(bp, bp->b_bufsize);
977 vn_strategy(bp->b_vp, &bp->b_bio1);
978 error = biowait(&bp->b_bio1, "biows");
987 * Asynchronous write. Start output on a buffer, but do not wait for
988 * it to complete. The buffer is released when the output completes.
990 * bwrite() ( or the VOP routine anyway ) is responsible for handling
991 * B_INVAL buffers. Not us.
994 bawrite(struct buf *bp)
996 if (bp->b_flags & B_INVAL) {
1000 if (BUF_REFCNTNB(bp) == 0)
1001 panic("bwrite: buffer is not busy???");
1003 /* Mark the buffer clean */
1006 bp->b_flags &= ~(B_ERROR | B_EINTR);
1007 bp->b_flags |= B_CACHE;
1008 bp->b_cmd = BUF_CMD_WRITE;
1009 KKASSERT(bp->b_bio1.bio_done == NULL);
1010 vfs_busy_pages(bp->b_vp, bp);
1013 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1014 * valid for vnode-backed buffers.
1016 bsetrunningbufspace(bp, bp->b_bufsize);
1018 vn_strategy(bp->b_vp, &bp->b_bio1);
1024 * Ordered write. Start output on a buffer, and flag it so that the
1025 * device will write it in the order it was queued. The buffer is
1026 * released when the output completes. bwrite() ( or the VOP routine
1027 * anyway ) is responsible for handling B_INVAL buffers.
1030 bowrite(struct buf *bp)
1032 bp->b_flags |= B_ORDERED;
1040 * Delayed write. (Buffer is marked dirty). Do not bother writing
1041 * anything if the buffer is marked invalid.
1043 * Note that since the buffer must be completely valid, we can safely
1044 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1045 * biodone() in order to prevent getblk from writing the buffer
1046 * out synchronously.
1049 bdwrite(struct buf *bp)
1051 if (BUF_REFCNTNB(bp) == 0)
1052 panic("bdwrite: buffer is not busy");
1054 if (bp->b_flags & B_INVAL) {
1060 if (dsched_is_clear_buf_priv(bp))
1064 * Set B_CACHE, indicating that the buffer is fully valid. This is
1065 * true even of NFS now.
1067 bp->b_flags |= B_CACHE;
1070 * This bmap keeps the system from needing to do the bmap later,
1071 * perhaps when the system is attempting to do a sync. Since it
1072 * is likely that the indirect block -- or whatever other datastructure
1073 * that the filesystem needs is still in memory now, it is a good
1074 * thing to do this. Note also, that if the pageout daemon is
1075 * requesting a sync -- there might not be enough memory to do
1076 * the bmap then... So, this is important to do.
1078 if (bp->b_bio2.bio_offset == NOOFFSET) {
1079 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1080 NULL, NULL, BUF_CMD_WRITE);
1084 * Because the underlying pages may still be mapped and
1085 * writable trying to set the dirty buffer (b_dirtyoff/end)
1086 * range here will be inaccurate.
1088 * However, we must still clean the pages to satisfy the
1089 * vnode_pager and pageout daemon, so theythink the pages
1090 * have been "cleaned". What has really occured is that
1091 * they've been earmarked for later writing by the buffer
1094 * So we get the b_dirtyoff/end update but will not actually
1095 * depend on it (NFS that is) until the pages are busied for
1098 vfs_clean_pages(bp);
1102 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1103 * due to the softdep code.
1108 * Fake write - return pages to VM system as dirty, leave the buffer clean.
1109 * This is used by tmpfs.
1111 * It is important for any VFS using this routine to NOT use it for
1112 * IO_SYNC or IO_ASYNC operations which occur when the system really
1113 * wants to flush VM pages to backing store.
1116 buwrite(struct buf *bp)
1122 * Only works for VMIO buffers. If the buffer is already
1123 * marked for delayed-write we can't avoid the bdwrite().
1125 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) {
1131 * Set valid & dirty.
1133 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1134 m = bp->b_xio.xio_pages[i];
1135 vfs_dirty_one_page(bp, i, m);
1143 * Turn buffer into delayed write request by marking it B_DELWRI.
1144 * B_RELBUF and B_NOCACHE must be cleared.
1146 * We reassign the buffer to itself to properly update it in the
1147 * dirty/clean lists.
1149 * Must be called from a critical section.
1150 * The buffer must be on BQUEUE_NONE.
1153 bdirty(struct buf *bp)
1155 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1156 if (bp->b_flags & B_NOCACHE) {
1157 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1158 bp->b_flags &= ~B_NOCACHE;
1160 if (bp->b_flags & B_INVAL) {
1161 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1163 bp->b_flags &= ~B_RELBUF;
1165 if ((bp->b_flags & B_DELWRI) == 0) {
1166 lwkt_gettoken(&bp->b_vp->v_token);
1167 bp->b_flags |= B_DELWRI;
1169 lwkt_reltoken(&bp->b_vp->v_token);
1171 spin_lock(&bufcspin);
1173 dirtybufspace += bp->b_bufsize;
1174 if (bp->b_flags & B_HEAVY) {
1176 dirtybufspacehw += bp->b_bufsize;
1178 spin_unlock(&bufcspin);
1185 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1186 * needs to be flushed with a different buf_daemon thread to avoid
1187 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1190 bheavy(struct buf *bp)
1192 if ((bp->b_flags & B_HEAVY) == 0) {
1193 bp->b_flags |= B_HEAVY;
1194 if (bp->b_flags & B_DELWRI) {
1195 spin_lock(&bufcspin);
1197 dirtybufspacehw += bp->b_bufsize;
1198 spin_unlock(&bufcspin);
1206 * Clear B_DELWRI for buffer.
1208 * Must be called from a critical section.
1210 * The buffer is typically on BQUEUE_NONE but there is one case in
1211 * brelse() that calls this function after placing the buffer on
1212 * a different queue.
1217 bundirty(struct buf *bp)
1219 if (bp->b_flags & B_DELWRI) {
1220 lwkt_gettoken(&bp->b_vp->v_token);
1221 bp->b_flags &= ~B_DELWRI;
1223 lwkt_reltoken(&bp->b_vp->v_token);
1225 spin_lock(&bufcspin);
1227 dirtybufspace -= bp->b_bufsize;
1228 if (bp->b_flags & B_HEAVY) {
1230 dirtybufspacehw -= bp->b_bufsize;
1232 spin_unlock(&bufcspin);
1234 bd_signal(bp->b_bufsize);
1237 * Since it is now being written, we can clear its deferred write flag.
1239 bp->b_flags &= ~B_DEFERRED;
1243 * Set the b_runningbufspace field, used to track how much I/O is
1244 * in progress at any given moment.
1247 bsetrunningbufspace(struct buf *bp, int bytes)
1249 bp->b_runningbufspace = bytes;
1251 spin_lock(&bufcspin);
1252 runningbufspace += bytes;
1254 spin_unlock(&bufcspin);
1261 * Release a busy buffer and, if requested, free its resources. The
1262 * buffer will be stashed in the appropriate bufqueue[] allowing it
1263 * to be accessed later as a cache entity or reused for other purposes.
1268 brelse(struct buf *bp)
1271 int saved_flags = bp->b_flags;
1274 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1277 * If B_NOCACHE is set we are being asked to destroy the buffer and
1278 * its backing store. Clear B_DELWRI.
1280 * B_NOCACHE is set in two cases: (1) when the caller really wants
1281 * to destroy the buffer and backing store and (2) when the caller
1282 * wants to destroy the buffer and backing store after a write
1285 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1289 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1291 * A re-dirtied buffer is only subject to destruction
1292 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1294 /* leave buffer intact */
1295 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1296 (bp->b_bufsize <= 0)) {
1298 * Either a failed read or we were asked to free or not
1299 * cache the buffer. This path is reached with B_DELWRI
1300 * set only if B_INVAL is already set. B_NOCACHE governs
1301 * backing store destruction.
1303 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1304 * buffer cannot be immediately freed.
1306 bp->b_flags |= B_INVAL;
1307 if (LIST_FIRST(&bp->b_dep) != NULL)
1309 if (bp->b_flags & B_DELWRI) {
1310 spin_lock(&bufcspin);
1312 dirtybufspace -= bp->b_bufsize;
1313 if (bp->b_flags & B_HEAVY) {
1315 dirtybufspacehw -= bp->b_bufsize;
1317 spin_unlock(&bufcspin);
1319 bd_signal(bp->b_bufsize);
1321 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1325 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1326 * If vfs_vmio_release() is called with either bit set, the
1327 * underlying pages may wind up getting freed causing a previous
1328 * write (bdwrite()) to get 'lost' because pages associated with
1329 * a B_DELWRI bp are marked clean. Pages associated with a
1330 * B_LOCKED buffer may be mapped by the filesystem.
1332 * If we want to release the buffer ourselves (rather then the
1333 * originator asking us to release it), give the originator a
1334 * chance to countermand the release by setting B_LOCKED.
1336 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1337 * if B_DELWRI is set.
1339 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1340 * on pages to return pages to the VM page queues.
1342 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1343 bp->b_flags &= ~B_RELBUF;
1344 } else if (vm_page_count_severe()) {
1345 if (LIST_FIRST(&bp->b_dep) != NULL)
1346 buf_deallocate(bp); /* can set B_LOCKED */
1347 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1348 bp->b_flags &= ~B_RELBUF;
1350 bp->b_flags |= B_RELBUF;
1354 * Make sure b_cmd is clear. It may have already been cleared by
1357 * At this point destroying the buffer is governed by the B_INVAL
1358 * or B_RELBUF flags.
1360 bp->b_cmd = BUF_CMD_DONE;
1361 dsched_exit_buf(bp);
1364 * VMIO buffer rundown. Make sure the VM page array is restored
1365 * after an I/O may have replaces some of the pages with bogus pages
1366 * in order to not destroy dirty pages in a fill-in read.
1368 * Note that due to the code above, if a buffer is marked B_DELWRI
1369 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1370 * B_INVAL may still be set, however.
1372 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1373 * but not the backing store. B_NOCACHE will destroy the backing
1376 * Note that dirty NFS buffers contain byte-granular write ranges
1377 * and should not be destroyed w/ B_INVAL even if the backing store
1380 if (bp->b_flags & B_VMIO) {
1382 * Rundown for VMIO buffers which are not dirty NFS buffers.
1394 * Get the base offset and length of the buffer. Note that
1395 * in the VMIO case if the buffer block size is not
1396 * page-aligned then b_data pointer may not be page-aligned.
1397 * But our b_xio.xio_pages array *IS* page aligned.
1399 * block sizes less then DEV_BSIZE (usually 512) are not
1400 * supported due to the page granularity bits (m->valid,
1401 * m->dirty, etc...).
1403 * See man buf(9) for more information
1406 resid = bp->b_bufsize;
1407 foff = bp->b_loffset;
1409 lwkt_gettoken(&vm_token);
1410 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1411 m = bp->b_xio.xio_pages[i];
1412 vm_page_flag_clear(m, PG_ZERO);
1414 * If we hit a bogus page, fixup *all* of them
1415 * now. Note that we left these pages wired
1416 * when we removed them so they had better exist,
1417 * and they cannot be ripped out from under us so
1418 * no critical section protection is necessary.
1420 if (m == bogus_page) {
1422 poff = OFF_TO_IDX(bp->b_loffset);
1424 for (j = i; j < bp->b_xio.xio_npages; j++) {
1427 mtmp = bp->b_xio.xio_pages[j];
1428 if (mtmp == bogus_page) {
1429 mtmp = vm_page_lookup(obj, poff + j);
1431 panic("brelse: page missing");
1433 bp->b_xio.xio_pages[j] = mtmp;
1436 bp->b_flags &= ~B_HASBOGUS;
1438 if ((bp->b_flags & B_INVAL) == 0) {
1439 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1440 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1442 m = bp->b_xio.xio_pages[i];
1446 * Invalidate the backing store if B_NOCACHE is set
1447 * (e.g. used with vinvalbuf()). If this is NFS
1448 * we impose a requirement that the block size be
1449 * a multiple of PAGE_SIZE and create a temporary
1450 * hack to basically invalidate the whole page. The
1451 * problem is that NFS uses really odd buffer sizes
1452 * especially when tracking piecemeal writes and
1453 * it also vinvalbuf()'s a lot, which would result
1454 * in only partial page validation and invalidation
1455 * here. If the file page is mmap()'d, however,
1456 * all the valid bits get set so after we invalidate
1457 * here we would end up with weird m->valid values
1458 * like 0xfc. nfs_getpages() can't handle this so
1459 * we clear all the valid bits for the NFS case
1460 * instead of just some of them.
1462 * The real bug is the VM system having to set m->valid
1463 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1464 * itself is an artifact of the whole 512-byte
1465 * granular mess that exists to support odd block
1466 * sizes and UFS meta-data block sizes (e.g. 6144).
1467 * A complete rewrite is required.
1471 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1472 int poffset = foff & PAGE_MASK;
1475 presid = PAGE_SIZE - poffset;
1476 if (bp->b_vp->v_tag == VT_NFS &&
1477 bp->b_vp->v_type == VREG) {
1479 } else if (presid > resid) {
1482 KASSERT(presid >= 0, ("brelse: extra page"));
1483 vm_page_set_invalid(m, poffset, presid);
1486 * Also make sure any swap cache is removed
1487 * as it is now stale (HAMMER in particular
1488 * uses B_NOCACHE to deal with buffer
1491 swap_pager_unswapped(m);
1493 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1494 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1496 if (bp->b_flags & (B_INVAL | B_RELBUF))
1497 vfs_vmio_release(bp);
1498 lwkt_reltoken(&vm_token);
1501 * Rundown for non-VMIO buffers.
1503 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1506 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1512 if (bp->b_qindex != BQUEUE_NONE)
1513 panic("brelse: free buffer onto another queue???");
1514 if (BUF_REFCNTNB(bp) > 1) {
1515 /* Temporary panic to verify exclusive locking */
1516 /* This panic goes away when we allow shared refs */
1517 panic("brelse: multiple refs");
1523 * Figure out the correct queue to place the cleaned up buffer on.
1524 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1525 * disassociated from their vnode.
1527 spin_lock(&bufqspin);
1528 if (bp->b_flags & B_LOCKED) {
1530 * Buffers that are locked are placed in the locked queue
1531 * immediately, regardless of their state.
1533 bp->b_qindex = BQUEUE_LOCKED;
1534 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1535 } else if (bp->b_bufsize == 0) {
1537 * Buffers with no memory. Due to conditionals near the top
1538 * of brelse() such buffers should probably already be
1539 * marked B_INVAL and disassociated from their vnode.
1541 bp->b_flags |= B_INVAL;
1542 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1543 KKASSERT((bp->b_flags & B_HASHED) == 0);
1544 if (bp->b_kvasize) {
1545 bp->b_qindex = BQUEUE_EMPTYKVA;
1547 bp->b_qindex = BQUEUE_EMPTY;
1549 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1550 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1552 * Buffers with junk contents. Again these buffers had better
1553 * already be disassociated from their vnode.
1555 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1556 KKASSERT((bp->b_flags & B_HASHED) == 0);
1557 bp->b_flags |= B_INVAL;
1558 bp->b_qindex = BQUEUE_CLEAN;
1559 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1562 * Remaining buffers. These buffers are still associated with
1565 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1567 bp->b_qindex = BQUEUE_DIRTY;
1568 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1570 case B_DELWRI | B_HEAVY:
1571 bp->b_qindex = BQUEUE_DIRTY_HW;
1572 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1577 * NOTE: Buffers are always placed at the end of the
1578 * queue. If B_AGE is not set the buffer will cycle
1579 * through the queue twice.
1581 bp->b_qindex = BQUEUE_CLEAN;
1582 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1586 spin_unlock(&bufqspin);
1589 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1590 * on the correct queue.
1592 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1596 * The bp is on an appropriate queue unless locked. If it is not
1597 * locked or dirty we can wakeup threads waiting for buffer space.
1599 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1600 * if B_INVAL is set ).
1602 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1606 * Something we can maybe free or reuse
1608 if (bp->b_bufsize || bp->b_kvasize)
1612 * Clean up temporary flags and unlock the buffer.
1614 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
1621 * Release a buffer back to the appropriate queue but do not try to free
1622 * it. The buffer is expected to be used again soon.
1624 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1625 * biodone() to requeue an async I/O on completion. It is also used when
1626 * known good buffers need to be requeued but we think we may need the data
1629 * XXX we should be able to leave the B_RELBUF hint set on completion.
1634 bqrelse(struct buf *bp)
1636 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1638 if (bp->b_qindex != BQUEUE_NONE)
1639 panic("bqrelse: free buffer onto another queue???");
1640 if (BUF_REFCNTNB(bp) > 1) {
1641 /* do not release to free list */
1642 panic("bqrelse: multiple refs");
1646 buf_act_advance(bp);
1648 spin_lock(&bufqspin);
1649 if (bp->b_flags & B_LOCKED) {
1651 * Locked buffers are released to the locked queue. However,
1652 * if the buffer is dirty it will first go into the dirty
1653 * queue and later on after the I/O completes successfully it
1654 * will be released to the locked queue.
1656 bp->b_qindex = BQUEUE_LOCKED;
1657 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1658 } else if (bp->b_flags & B_DELWRI) {
1659 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1660 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1661 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1662 } else if (vm_page_count_severe()) {
1664 * We are too low on memory, we have to try to free the
1665 * buffer (most importantly: the wired pages making up its
1666 * backing store) *now*.
1668 spin_unlock(&bufqspin);
1672 bp->b_qindex = BQUEUE_CLEAN;
1673 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1675 spin_unlock(&bufqspin);
1677 if ((bp->b_flags & B_LOCKED) == 0 &&
1678 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1683 * Something we can maybe free or reuse.
1685 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1689 * Final cleanup and unlock. Clear bits that are only used while a
1690 * buffer is actively locked.
1692 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
1693 dsched_exit_buf(bp);
1700 * Return backing pages held by the buffer 'bp' back to the VM system
1701 * if possible. The pages are freed if they are no longer valid or
1702 * attempt to free if it was used for direct I/O otherwise they are
1703 * sent to the page cache.
1705 * Pages that were marked busy are left alone and skipped.
1707 * The KVA mapping (b_data) for the underlying pages is removed by
1711 vfs_vmio_release(struct buf *bp)
1716 lwkt_gettoken(&vm_token);
1717 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1718 m = bp->b_xio.xio_pages[i];
1719 bp->b_xio.xio_pages[i] = NULL;
1722 * The VFS is telling us this is not a meta-data buffer
1723 * even if it is backed by a block device.
1725 if (bp->b_flags & B_NOTMETA)
1726 vm_page_flag_set(m, PG_NOTMETA);
1729 * This is a very important bit of code. We try to track
1730 * VM page use whether the pages are wired into the buffer
1731 * cache or not. While wired into the buffer cache the
1732 * bp tracks the act_count.
1734 * We can choose to place unwired pages on the inactive
1735 * queue (0) or active queue (1). If we place too many
1736 * on the active queue the queue will cycle the act_count
1737 * on pages we'd like to keep, just from single-use pages
1738 * (such as when doing a tar-up or file scan).
1740 if (bp->b_act_count < vm_cycle_point)
1741 vm_page_unwire(m, 0);
1743 vm_page_unwire(m, 1);
1746 * We don't mess with busy pages, it is
1747 * the responsibility of the process that
1748 * busied the pages to deal with them.
1750 if ((m->flags & PG_BUSY) || (m->busy != 0))
1753 if (m->wire_count == 0) {
1754 vm_page_flag_clear(m, PG_ZERO);
1756 * Might as well free the page if we can and it has
1757 * no valid data. We also free the page if the
1758 * buffer was used for direct I/O.
1761 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1762 m->hold_count == 0) {
1764 vm_page_protect(m, VM_PROT_NONE);
1768 if (bp->b_flags & B_DIRECT) {
1769 vm_page_try_to_free(m);
1770 } else if (vm_page_count_severe()) {
1771 m->act_count = bp->b_act_count;
1772 vm_page_try_to_cache(m);
1774 m->act_count = bp->b_act_count;
1778 lwkt_reltoken(&vm_token);
1780 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
1781 bp->b_xio.xio_npages);
1782 if (bp->b_bufsize) {
1786 bp->b_xio.xio_npages = 0;
1787 bp->b_flags &= ~B_VMIO;
1788 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1796 * Implement clustered async writes for clearing out B_DELWRI buffers.
1797 * This is much better then the old way of writing only one buffer at
1798 * a time. Note that we may not be presented with the buffers in the
1799 * correct order, so we search for the cluster in both directions.
1801 * The buffer is locked on call.
1804 vfs_bio_awrite(struct buf *bp)
1808 off_t loffset = bp->b_loffset;
1809 struct vnode *vp = bp->b_vp;
1816 * right now we support clustered writing only to regular files. If
1817 * we find a clusterable block we could be in the middle of a cluster
1818 * rather then at the beginning.
1820 * NOTE: b_bio1 contains the logical loffset and is aliased
1821 * to b_loffset. b_bio2 contains the translated block number.
1823 if ((vp->v_type == VREG) &&
1824 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1825 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1827 size = vp->v_mount->mnt_stat.f_iosize;
1829 for (i = size; i < MAXPHYS; i += size) {
1830 if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) &&
1831 BUF_REFCNT(bpa) == 0 &&
1832 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1833 (B_DELWRI | B_CLUSTEROK)) &&
1834 (bpa->b_bufsize == size)) {
1835 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1836 (bpa->b_bio2.bio_offset !=
1837 bp->b_bio2.bio_offset + i))
1843 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1844 if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) &&
1845 BUF_REFCNT(bpa) == 0 &&
1846 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1847 (B_DELWRI | B_CLUSTEROK)) &&
1848 (bpa->b_bufsize == size)) {
1849 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1850 (bpa->b_bio2.bio_offset !=
1851 bp->b_bio2.bio_offset - j))
1861 * this is a possible cluster write
1863 if (nbytes != size) {
1865 nwritten = cluster_wbuild(vp, size,
1866 loffset - j, nbytes);
1872 * default (old) behavior, writing out only one block
1874 * XXX returns b_bufsize instead of b_bcount for nwritten?
1876 nwritten = bp->b_bufsize;
1886 * Find and initialize a new buffer header, freeing up existing buffers
1887 * in the bufqueues as necessary. The new buffer is returned locked.
1889 * Important: B_INVAL is not set. If the caller wishes to throw the
1890 * buffer away, the caller must set B_INVAL prior to calling brelse().
1893 * We have insufficient buffer headers
1894 * We have insufficient buffer space
1895 * buffer_map is too fragmented ( space reservation fails )
1896 * If we have to flush dirty buffers ( but we try to avoid this )
1898 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1899 * Instead we ask the buf daemon to do it for us. We attempt to
1900 * avoid piecemeal wakeups of the pageout daemon.
1905 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1911 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1912 static int flushingbufs;
1915 * We can't afford to block since we might be holding a vnode lock,
1916 * which may prevent system daemons from running. We deal with
1917 * low-memory situations by proactively returning memory and running
1918 * async I/O rather then sync I/O.
1922 --getnewbufrestarts;
1924 ++getnewbufrestarts;
1927 * Setup for scan. If we do not have enough free buffers,
1928 * we setup a degenerate case that immediately fails. Note
1929 * that if we are specially marked process, we are allowed to
1930 * dip into our reserves.
1932 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1934 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1935 * However, there are a number of cases (defragging, reusing, ...)
1936 * where we cannot backup.
1938 nqindex = BQUEUE_EMPTYKVA;
1939 spin_lock(&bufqspin);
1940 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1944 * If no EMPTYKVA buffers and we are either
1945 * defragging or reusing, locate a CLEAN buffer
1946 * to free or reuse. If bufspace useage is low
1947 * skip this step so we can allocate a new buffer.
1949 if (defrag || bufspace >= lobufspace) {
1950 nqindex = BQUEUE_CLEAN;
1951 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1955 * If we could not find or were not allowed to reuse a
1956 * CLEAN buffer, check to see if it is ok to use an EMPTY
1957 * buffer. We can only use an EMPTY buffer if allocating
1958 * its KVA would not otherwise run us out of buffer space.
1960 if (nbp == NULL && defrag == 0 &&
1961 bufspace + maxsize < hibufspace) {
1962 nqindex = BQUEUE_EMPTY;
1963 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1968 * Run scan, possibly freeing data and/or kva mappings on the fly
1971 * WARNING! bufqspin is held!
1973 while ((bp = nbp) != NULL) {
1974 int qindex = nqindex;
1976 nbp = TAILQ_NEXT(bp, b_freelist);
1979 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1980 * cycles through the queue twice before being selected.
1982 if (qindex == BQUEUE_CLEAN &&
1983 (bp->b_flags & B_AGE) == 0 && nbp) {
1984 bp->b_flags |= B_AGE;
1985 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1986 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1991 * Calculate next bp ( we can only use it if we do not block
1992 * or do other fancy things ).
1997 nqindex = BQUEUE_EMPTYKVA;
1998 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
2001 case BQUEUE_EMPTYKVA:
2002 nqindex = BQUEUE_CLEAN;
2003 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
2017 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2020 * Note: we no longer distinguish between VMIO and non-VMIO
2023 KASSERT((bp->b_flags & B_DELWRI) == 0,
2024 ("delwri buffer %p found in queue %d", bp, qindex));
2027 * Do not try to reuse a buffer with a non-zero b_refs.
2028 * This is an unsynchronized test. A synchronized test
2029 * is also performed after we lock the buffer.
2035 * If we are defragging then we need a buffer with
2036 * b_kvasize != 0. XXX this situation should no longer
2037 * occur, if defrag is non-zero the buffer's b_kvasize
2038 * should also be non-zero at this point. XXX
2040 if (defrag && bp->b_kvasize == 0) {
2041 kprintf("Warning: defrag empty buffer %p\n", bp);
2046 * Start freeing the bp. This is somewhat involved. nbp
2047 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
2048 * on the clean list must be disassociated from their
2049 * current vnode. Buffers on the empty[kva] lists have
2050 * already been disassociated.
2053 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2054 spin_unlock(&bufqspin);
2055 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
2058 if (bp->b_qindex != qindex) {
2059 spin_unlock(&bufqspin);
2060 kprintf("getnewbuf: warning, BUF_LOCK blocked "
2061 "unexpectedly on buf %p index %d->%d, "
2063 bp, qindex, bp->b_qindex);
2067 bremfree_locked(bp);
2068 spin_unlock(&bufqspin);
2071 * Dependancies must be handled before we disassociate the
2074 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2075 * be immediately disassociated. HAMMER then becomes
2076 * responsible for releasing the buffer.
2078 * NOTE: bufqspin is UNLOCKED now.
2080 if (LIST_FIRST(&bp->b_dep) != NULL) {
2082 if (bp->b_flags & B_LOCKED) {
2086 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2089 if (qindex == BQUEUE_CLEAN) {
2090 if (bp->b_flags & B_VMIO)
2091 vfs_vmio_release(bp);
2097 * NOTE: nbp is now entirely invalid. We can only restart
2098 * the scan from this point on.
2100 * Get the rest of the buffer freed up. b_kva* is still
2101 * valid after this operation.
2104 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
2105 KKASSERT((bp->b_flags & B_HASHED) == 0);
2108 * critical section protection is not required when
2109 * scrapping a buffer's contents because it is already
2115 bp->b_flags = B_BNOCLIP;
2116 bp->b_cmd = BUF_CMD_DONE;
2121 bp->b_xio.xio_npages = 0;
2122 bp->b_dirtyoff = bp->b_dirtyend = 0;
2123 bp->b_act_count = ACT_INIT;
2125 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2127 if (blkflags & GETBLK_BHEAVY)
2128 bp->b_flags |= B_HEAVY;
2131 * If we are defragging then free the buffer.
2134 bp->b_flags |= B_INVAL;
2142 * If we are overcomitted then recover the buffer and its
2143 * KVM space. This occurs in rare situations when multiple
2144 * processes are blocked in getnewbuf() or allocbuf().
2146 if (bufspace >= hibufspace)
2148 if (flushingbufs && bp->b_kvasize != 0) {
2149 bp->b_flags |= B_INVAL;
2154 if (bufspace < lobufspace)
2158 * The brelvp() above interlocked the buffer, test b_refs
2159 * to determine if the buffer can be reused. b_refs
2160 * interlocks lookup/blocking-lock operations and allowing
2161 * buffer reuse can create deadlocks depending on what
2162 * (vp,loffset) is assigned to the reused buffer (see getblk).
2165 bp->b_flags |= B_INVAL;
2172 /* NOT REACHED, bufqspin not held */
2176 * If we exhausted our list, sleep as appropriate. We may have to
2177 * wakeup various daemons and write out some dirty buffers.
2179 * Generally we are sleeping due to insufficient buffer space.
2181 * NOTE: bufqspin is held if bp is NULL, else it is not held.
2187 spin_unlock(&bufqspin);
2189 flags = VFS_BIO_NEED_BUFSPACE;
2191 } else if (bufspace >= hibufspace) {
2193 flags = VFS_BIO_NEED_BUFSPACE;
2196 flags = VFS_BIO_NEED_ANY;
2199 bd_speedup(); /* heeeelp */
2200 spin_lock(&bufcspin);
2201 needsbuffer |= flags;
2202 while (needsbuffer & flags) {
2203 if (ssleep(&needsbuffer, &bufcspin,
2204 slpflags, waitmsg, slptimeo)) {
2205 spin_unlock(&bufcspin);
2209 spin_unlock(&bufcspin);
2212 * We finally have a valid bp. We aren't quite out of the
2213 * woods, we still have to reserve kva space. In order
2214 * to keep fragmentation sane we only allocate kva in
2217 * (bufqspin is not held)
2219 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2221 if (maxsize != bp->b_kvasize) {
2222 vm_offset_t addr = 0;
2227 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2228 vm_map_lock(&buffer_map);
2230 if (vm_map_findspace(&buffer_map,
2231 vm_map_min(&buffer_map), maxsize,
2232 maxsize, 0, &addr)) {
2234 * Uh oh. Buffer map is too fragmented. We
2235 * must defragment the map.
2237 vm_map_unlock(&buffer_map);
2238 vm_map_entry_release(count);
2241 bp->b_flags |= B_INVAL;
2246 vm_map_insert(&buffer_map, &count,
2248 addr, addr + maxsize,
2250 VM_PROT_ALL, VM_PROT_ALL,
2253 bp->b_kvabase = (caddr_t) addr;
2254 bp->b_kvasize = maxsize;
2255 bufspace += bp->b_kvasize;
2258 vm_map_unlock(&buffer_map);
2259 vm_map_entry_release(count);
2261 bp->b_data = bp->b_kvabase;
2267 * This routine is called in an emergency to recover VM pages from the
2268 * buffer cache by cashing in clean buffers. The idea is to recover
2269 * enough pages to be able to satisfy a stuck bio_page_alloc().
2274 recoverbufpages(void)
2281 spin_lock(&bufqspin);
2282 while (bytes < MAXBSIZE) {
2283 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2288 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2289 * cycles through the queue twice before being selected.
2291 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
2292 bp->b_flags |= B_AGE;
2293 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
2294 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
2302 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
2303 KKASSERT((bp->b_flags & B_DELWRI) == 0);
2306 * Start freeing the bp. This is somewhat involved.
2308 * Buffers on the clean list must be disassociated from
2309 * their current vnode
2312 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2313 kprintf("recoverbufpages: warning, locked buf %p, "
2316 ssleep(&bd_request, &bufqspin, 0, "gnbxxx", hz / 100);
2319 if (bp->b_qindex != BQUEUE_CLEAN) {
2320 kprintf("recoverbufpages: warning, BUF_LOCK blocked "
2321 "unexpectedly on buf %p index %d, race "
2327 bremfree_locked(bp);
2328 spin_unlock(&bufqspin);
2331 * Dependancies must be handled before we disassociate the
2334 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2335 * be immediately disassociated. HAMMER then becomes
2336 * responsible for releasing the buffer.
2338 if (LIST_FIRST(&bp->b_dep) != NULL) {
2340 if (bp->b_flags & B_LOCKED) {
2342 spin_lock(&bufqspin);
2345 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2348 bytes += bp->b_bufsize;
2350 if (bp->b_flags & B_VMIO) {
2351 bp->b_flags |= B_DIRECT; /* try to free pages */
2352 vfs_vmio_release(bp);
2357 KKASSERT(bp->b_vp == NULL);
2358 KKASSERT((bp->b_flags & B_HASHED) == 0);
2361 * critical section protection is not required when
2362 * scrapping a buffer's contents because it is already
2368 bp->b_flags = B_BNOCLIP;
2369 bp->b_cmd = BUF_CMD_DONE;
2374 bp->b_xio.xio_npages = 0;
2375 bp->b_dirtyoff = bp->b_dirtyend = 0;
2377 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2379 bp->b_flags |= B_INVAL;
2382 spin_lock(&bufqspin);
2384 spin_unlock(&bufqspin);
2391 * Buffer flushing daemon. Buffers are normally flushed by the
2392 * update daemon but if it cannot keep up this process starts to
2393 * take the load in an attempt to prevent getnewbuf() from blocking.
2395 * Once a flush is initiated it does not stop until the number
2396 * of buffers falls below lodirtybuffers, but we will wake up anyone
2397 * waiting at the mid-point.
2400 static struct kproc_desc buf_kp = {
2405 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2406 kproc_start, &buf_kp)
2408 static struct kproc_desc bufhw_kp = {
2413 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2414 kproc_start, &bufhw_kp)
2425 * This process needs to be suspended prior to shutdown sync.
2427 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2428 bufdaemon_td, SHUTDOWN_PRI_LAST);
2429 curthread->td_flags |= TDF_SYSTHREAD;
2432 * This process is allowed to take the buffer cache to the limit
2435 kproc_suspend_loop();
2438 * Do the flush as long as the number of dirty buffers
2439 * (including those running) exceeds lodirtybufspace.
2441 * When flushing limit running I/O to hirunningspace
2442 * Do the flush. Limit the amount of in-transit I/O we
2443 * allow to build up, otherwise we would completely saturate
2444 * the I/O system. Wakeup any waiting processes before we
2445 * normally would so they can run in parallel with our drain.
2447 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2448 * but because we split the operation into two threads we
2449 * have to cut it in half for each thread.
2451 waitrunningbufspace();
2452 limit = lodirtybufspace / 2;
2453 while (runningbufspace + dirtybufspace > limit ||
2454 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
2455 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2457 if (runningbufspace < hirunningspace)
2459 waitrunningbufspace();
2463 * We reached our low water mark, reset the
2464 * request and sleep until we are needed again.
2465 * The sleep is just so the suspend code works.
2467 spin_lock(&bufcspin);
2468 if (bd_request == 0)
2469 ssleep(&bd_request, &bufcspin, 0, "psleep", hz);
2471 spin_unlock(&bufcspin);
2484 * This process needs to be suspended prior to shutdown sync.
2486 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2487 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2488 curthread->td_flags |= TDF_SYSTHREAD;
2491 * This process is allowed to take the buffer cache to the limit
2494 kproc_suspend_loop();
2497 * Do the flush. Limit the amount of in-transit I/O we
2498 * allow to build up, otherwise we would completely saturate
2499 * the I/O system. Wakeup any waiting processes before we
2500 * normally would so they can run in parallel with our drain.
2502 * Once we decide to flush push the queued I/O up to
2503 * hirunningspace in order to trigger bursting by the bioq
2506 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2507 * but because we split the operation into two threads we
2508 * have to cut it in half for each thread.
2510 waitrunningbufspace();
2511 limit = lodirtybufspace / 2;
2512 while (runningbufspace + dirtybufspacehw > limit ||
2513 dirtybufcounthw >= nbuf / 2) {
2514 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2516 if (runningbufspace < hirunningspace)
2518 waitrunningbufspace();
2522 * We reached our low water mark, reset the
2523 * request and sleep until we are needed again.
2524 * The sleep is just so the suspend code works.
2526 spin_lock(&bufcspin);
2527 if (bd_request_hw == 0)
2528 ssleep(&bd_request_hw, &bufcspin, 0, "psleep", hz);
2530 spin_unlock(&bufcspin);
2537 * Try to flush a buffer in the dirty queue. We must be careful to
2538 * free up B_INVAL buffers instead of write them, which NFS is
2539 * particularly sensitive to.
2541 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2542 * that we really want to try to get the buffer out and reuse it
2543 * due to the write load on the machine.
2545 * We must lock the buffer in order to check its validity before we
2546 * can mess with its contents. bufqspin isn't enough.
2549 flushbufqueues(bufq_type_t q)
2555 spin_lock(&bufqspin);
2558 bp = TAILQ_FIRST(&bufqueues[q]);
2560 if ((bp->b_flags & B_DELWRI) == 0) {
2561 kprintf("Unexpected clean buffer %p\n", bp);
2562 bp = TAILQ_NEXT(bp, b_freelist);
2565 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2566 bp = TAILQ_NEXT(bp, b_freelist);
2569 KKASSERT(bp->b_qindex == q);
2572 * Must recheck B_DELWRI after successfully locking
2575 if ((bp->b_flags & B_DELWRI) == 0) {
2577 bp = TAILQ_NEXT(bp, b_freelist);
2581 if (bp->b_flags & B_INVAL) {
2583 spin_unlock(&bufqspin);
2590 if (LIST_FIRST(&bp->b_dep) != NULL &&
2591 (bp->b_flags & B_DEFERRED) == 0 &&
2592 buf_countdeps(bp, 0)) {
2593 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2594 TAILQ_INSERT_TAIL(&bufqueues[q], bp, b_freelist);
2595 bp->b_flags |= B_DEFERRED;
2597 bp = TAILQ_FIRST(&bufqueues[q]);
2602 * If the buffer has a dependancy, buf_checkwrite() must
2603 * also return 0 for us to be able to initate the write.
2605 * If the buffer is flagged B_ERROR it may be requeued
2606 * over and over again, we try to avoid a live lock.
2608 * NOTE: buf_checkwrite is MPSAFE.
2610 spin_unlock(&bufqspin);
2613 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) {
2616 } else if (bp->b_flags & B_ERROR) {
2617 tsleep(bp, 0, "bioer", 1);
2618 bp->b_flags &= ~B_AGE;
2621 bp->b_flags |= B_AGE;
2628 spin_unlock(&bufqspin);
2635 * Returns true if no I/O is needed to access the associated VM object.
2636 * This is like findblk except it also hunts around in the VM system for
2639 * Note that we ignore vm_page_free() races from interrupts against our
2640 * lookup, since if the caller is not protected our return value will not
2641 * be any more valid then otherwise once we exit the critical section.
2644 inmem(struct vnode *vp, off_t loffset)
2647 vm_offset_t toff, tinc, size;
2650 if (findblk(vp, loffset, FINDBLK_TEST))
2652 if (vp->v_mount == NULL)
2654 if ((obj = vp->v_object) == NULL)
2658 if (size > vp->v_mount->mnt_stat.f_iosize)
2659 size = vp->v_mount->mnt_stat.f_iosize;
2661 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2662 lwkt_gettoken(&vm_token);
2663 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2664 lwkt_reltoken(&vm_token);
2668 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2669 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2670 if (vm_page_is_valid(m,
2671 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2680 * Locate and return the specified buffer. Unless flagged otherwise,
2681 * a locked buffer will be returned if it exists or NULL if it does not.
2683 * findblk()'d buffers are still on the bufqueues and if you intend
2684 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2685 * and possibly do other stuff to it.
2687 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2688 * for locking the buffer and ensuring that it remains
2689 * the desired buffer after locking.
2691 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2692 * to acquire the lock we return NULL, even if the
2695 * FINDBLK_REF - Returns the buffer ref'd, which prevents reuse
2696 * by getnewbuf() but does not prevent disassociation
2697 * while we are locked. Used to avoid deadlocks
2698 * against random (vp,loffset)s due to reassignment.
2700 * (0) - Lock the buffer blocking.
2705 findblk(struct vnode *vp, off_t loffset, int flags)
2710 lkflags = LK_EXCLUSIVE;
2711 if (flags & FINDBLK_NBLOCK)
2712 lkflags |= LK_NOWAIT;
2716 * Lookup. Ref the buf while holding v_token to prevent
2717 * reuse (but does not prevent diassociation).
2719 lwkt_gettoken(&vp->v_token);
2720 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2722 lwkt_reltoken(&vp->v_token);
2725 atomic_add_int(&bp->b_refs, 1);
2726 lwkt_reltoken(&vp->v_token);
2729 * If testing only break and return bp, do not lock.
2731 if (flags & FINDBLK_TEST)
2735 * Lock the buffer, return an error if the lock fails.
2736 * (only FINDBLK_NBLOCK can cause the lock to fail).
2738 if (BUF_LOCK(bp, lkflags)) {
2739 atomic_subtract_int(&bp->b_refs, 1);
2740 /* bp = NULL; not needed */
2745 * Revalidate the locked buf before allowing it to be
2748 if (bp->b_vp == vp && bp->b_loffset == loffset)
2750 atomic_subtract_int(&bp->b_refs, 1);
2757 if ((flags & FINDBLK_REF) == 0)
2758 atomic_subtract_int(&bp->b_refs, 1);
2763 unrefblk(struct buf *bp)
2765 atomic_subtract_int(&bp->b_refs, 1);
2771 * Similar to getblk() except only returns the buffer if it is
2772 * B_CACHE and requires no other manipulation. Otherwise NULL
2775 * If B_RAM is set the buffer might be just fine, but we return
2776 * NULL anyway because we want the code to fall through to the
2777 * cluster read. Otherwise read-ahead breaks.
2780 getcacheblk(struct vnode *vp, off_t loffset)
2784 bp = findblk(vp, loffset, 0);
2786 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
2787 bp->b_flags &= ~B_AGE;
2800 * Get a block given a specified block and offset into a file/device.
2801 * B_INVAL may or may not be set on return. The caller should clear
2802 * B_INVAL prior to initiating a READ.
2804 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2805 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2806 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2807 * without doing any of those things the system will likely believe
2808 * the buffer to be valid (especially if it is not B_VMIO), and the
2809 * next getblk() will return the buffer with B_CACHE set.
2811 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2812 * an existing buffer.
2814 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2815 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2816 * and then cleared based on the backing VM. If the previous buffer is
2817 * non-0-sized but invalid, B_CACHE will be cleared.
2819 * If getblk() must create a new buffer, the new buffer is returned with
2820 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2821 * case it is returned with B_INVAL clear and B_CACHE set based on the
2824 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2825 * B_CACHE bit is clear.
2827 * What this means, basically, is that the caller should use B_CACHE to
2828 * determine whether the buffer is fully valid or not and should clear
2829 * B_INVAL prior to issuing a read. If the caller intends to validate
2830 * the buffer by loading its data area with something, the caller needs
2831 * to clear B_INVAL. If the caller does this without issuing an I/O,
2832 * the caller should set B_CACHE ( as an optimization ), else the caller
2833 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2834 * a write attempt or if it was a successfull read. If the caller
2835 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2836 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2840 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2841 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2846 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2849 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2853 if (size > MAXBSIZE)
2854 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2855 if (vp->v_object == NULL)
2856 panic("getblk: vnode %p has no object!", vp);
2859 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) {
2861 * The buffer was found in the cache, but we need to lock it.
2862 * We must acquire a ref on the bp to prevent reuse, but
2863 * this will not prevent disassociation (brelvp()) so we
2864 * must recheck (vp,loffset) after acquiring the lock.
2866 * Without the ref the buffer could potentially be reused
2867 * before we acquire the lock and create a deadlock
2868 * situation between the thread trying to reuse the buffer
2869 * and us due to the fact that we would wind up blocking
2870 * on a random (vp,loffset).
2872 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2873 if (blkflags & GETBLK_NOWAIT) {
2877 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2878 if (blkflags & GETBLK_PCATCH)
2879 lkflags |= LK_PCATCH;
2880 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2883 if (error == ENOLCK)
2887 /* buffer may have changed on us */
2892 * Once the buffer has been locked, make sure we didn't race
2893 * a buffer recyclement. Buffers that are no longer hashed
2894 * will have b_vp == NULL, so this takes care of that check
2897 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2898 kprintf("Warning buffer %p (vp %p loffset %lld) "
2900 bp, vp, (long long)loffset);
2906 * If SZMATCH any pre-existing buffer must be of the requested
2907 * size or NULL is returned. The caller absolutely does not
2908 * want getblk() to bwrite() the buffer on a size mismatch.
2910 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2916 * All vnode-based buffers must be backed by a VM object.
2918 KKASSERT(bp->b_flags & B_VMIO);
2919 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2920 bp->b_flags &= ~B_AGE;
2923 * Make sure that B_INVAL buffers do not have a cached
2924 * block number translation.
2926 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2927 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2928 " did not have cleared bio_offset cache\n",
2929 bp, vp, (long long)loffset);
2930 clearbiocache(&bp->b_bio2);
2934 * The buffer is locked. B_CACHE is cleared if the buffer is
2937 if (bp->b_flags & B_INVAL)
2938 bp->b_flags &= ~B_CACHE;
2942 * Any size inconsistancy with a dirty buffer or a buffer
2943 * with a softupdates dependancy must be resolved. Resizing
2944 * the buffer in such circumstances can lead to problems.
2946 * Dirty or dependant buffers are written synchronously.
2947 * Other types of buffers are simply released and
2948 * reconstituted as they may be backed by valid, dirty VM
2949 * pages (but not marked B_DELWRI).
2951 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2952 * and may be left over from a prior truncation (and thus
2953 * no longer represent the actual EOF point), so we
2954 * definitely do not want to B_NOCACHE the backing store.
2956 if (size != bp->b_bcount) {
2957 if (bp->b_flags & B_DELWRI) {
2958 bp->b_flags |= B_RELBUF;
2960 } else if (LIST_FIRST(&bp->b_dep)) {
2961 bp->b_flags |= B_RELBUF;
2964 bp->b_flags |= B_RELBUF;
2969 KKASSERT(size <= bp->b_kvasize);
2970 KASSERT(bp->b_loffset != NOOFFSET,
2971 ("getblk: no buffer offset"));
2974 * A buffer with B_DELWRI set and B_CACHE clear must
2975 * be committed before we can return the buffer in
2976 * order to prevent the caller from issuing a read
2977 * ( due to B_CACHE not being set ) and overwriting
2980 * Most callers, including NFS and FFS, need this to
2981 * operate properly either because they assume they
2982 * can issue a read if B_CACHE is not set, or because
2983 * ( for example ) an uncached B_DELWRI might loop due
2984 * to softupdates re-dirtying the buffer. In the latter
2985 * case, B_CACHE is set after the first write completes,
2986 * preventing further loops.
2988 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2989 * above while extending the buffer, we cannot allow the
2990 * buffer to remain with B_CACHE set after the write
2991 * completes or it will represent a corrupt state. To
2992 * deal with this we set B_NOCACHE to scrap the buffer
2995 * XXX Should this be B_RELBUF instead of B_NOCACHE?
2996 * I'm not even sure this state is still possible
2997 * now that getblk() writes out any dirty buffers
3000 * We might be able to do something fancy, like setting
3001 * B_CACHE in bwrite() except if B_DELWRI is already set,
3002 * so the below call doesn't set B_CACHE, but that gets real
3003 * confusing. This is much easier.
3006 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3007 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
3008 "and CACHE clear, b_flags %08x\n",
3009 bp, (intmax_t)bp->b_loffset, bp->b_flags);
3010 bp->b_flags |= B_NOCACHE;
3016 * Buffer is not in-core, create new buffer. The buffer
3017 * returned by getnewbuf() is locked. Note that the returned
3018 * buffer is also considered valid (not marked B_INVAL).
3020 * Calculating the offset for the I/O requires figuring out
3021 * the block size. We use DEV_BSIZE for VBLK or VCHR and
3022 * the mount's f_iosize otherwise. If the vnode does not
3023 * have an associated mount we assume that the passed size is
3026 * Note that vn_isdisk() cannot be used here since it may
3027 * return a failure for numerous reasons. Note that the
3028 * buffer size may be larger then the block size (the caller
3029 * will use block numbers with the proper multiple). Beware
3030 * of using any v_* fields which are part of unions. In
3031 * particular, in DragonFly the mount point overloading
3032 * mechanism uses the namecache only and the underlying
3033 * directory vnode is not a special case.
3037 if (vp->v_type == VBLK || vp->v_type == VCHR)
3039 else if (vp->v_mount)
3040 bsize = vp->v_mount->mnt_stat.f_iosize;
3044 maxsize = size + (loffset & PAGE_MASK);
3045 maxsize = imax(maxsize, bsize);
3047 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
3049 if (slpflags || slptimeo)
3055 * Atomically insert the buffer into the hash, so that it can
3056 * be found by findblk().
3058 * If bgetvp() returns non-zero a collision occured, and the
3059 * bp will not be associated with the vnode.
3061 * Make sure the translation layer has been cleared.
3063 bp->b_loffset = loffset;
3064 bp->b_bio2.bio_offset = NOOFFSET;
3065 /* bp->b_bio2.bio_next = NULL; */
3067 if (bgetvp(vp, bp, size)) {
3068 bp->b_flags |= B_INVAL;
3074 * All vnode-based buffers must be backed by a VM object.
3076 KKASSERT(vp->v_object != NULL);
3077 bp->b_flags |= B_VMIO;
3078 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3082 KKASSERT(dsched_is_clear_buf_priv(bp));
3089 * Reacquire a buffer that was previously released to the locked queue,
3090 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
3091 * set B_LOCKED (which handles the acquisition race).
3093 * To this end, either B_LOCKED must be set or the dependancy list must be
3099 regetblk(struct buf *bp)
3101 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
3102 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
3109 * Get an empty, disassociated buffer of given size. The buffer is
3110 * initially set to B_INVAL.
3112 * critical section protection is not required for the allocbuf()
3113 * call because races are impossible here.
3123 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3125 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
3128 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3129 KKASSERT(dsched_is_clear_buf_priv(bp));
3137 * This code constitutes the buffer memory from either anonymous system
3138 * memory (in the case of non-VMIO operations) or from an associated
3139 * VM object (in the case of VMIO operations). This code is able to
3140 * resize a buffer up or down.
3142 * Note that this code is tricky, and has many complications to resolve
3143 * deadlock or inconsistant data situations. Tread lightly!!!
3144 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3145 * the caller. Calling this code willy nilly can result in the loss of
3148 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3149 * B_CACHE for the non-VMIO case.
3151 * This routine does not need to be called from a critical section but you
3152 * must own the buffer.
3157 allocbuf(struct buf *bp, int size)
3159 int newbsize, mbsize;
3162 if (BUF_REFCNT(bp) == 0)
3163 panic("allocbuf: buffer not busy");
3165 if (bp->b_kvasize < size)
3166 panic("allocbuf: buffer too small");
3168 if ((bp->b_flags & B_VMIO) == 0) {
3172 * Just get anonymous memory from the kernel. Don't
3173 * mess with B_CACHE.
3175 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3176 if (bp->b_flags & B_MALLOC)
3179 newbsize = round_page(size);
3181 if (newbsize < bp->b_bufsize) {
3183 * Malloced buffers are not shrunk
3185 if (bp->b_flags & B_MALLOC) {
3187 bp->b_bcount = size;
3189 kfree(bp->b_data, M_BIOBUF);
3190 if (bp->b_bufsize) {
3191 atomic_subtract_int(&bufmallocspace, bp->b_bufsize);
3195 bp->b_data = bp->b_kvabase;
3197 bp->b_flags &= ~B_MALLOC;
3203 (vm_offset_t) bp->b_data + newbsize,
3204 (vm_offset_t) bp->b_data + bp->b_bufsize);
3205 } else if (newbsize > bp->b_bufsize) {
3207 * We only use malloced memory on the first allocation.
3208 * and revert to page-allocated memory when the buffer
3211 if ((bufmallocspace < maxbufmallocspace) &&
3212 (bp->b_bufsize == 0) &&
3213 (mbsize <= PAGE_SIZE/2)) {
3215 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3216 bp->b_bufsize = mbsize;
3217 bp->b_bcount = size;
3218 bp->b_flags |= B_MALLOC;
3219 atomic_add_int(&bufmallocspace, mbsize);
3225 * If the buffer is growing on its other-than-first
3226 * allocation, then we revert to the page-allocation
3229 if (bp->b_flags & B_MALLOC) {
3230 origbuf = bp->b_data;
3231 origbufsize = bp->b_bufsize;
3232 bp->b_data = bp->b_kvabase;
3233 if (bp->b_bufsize) {
3234 atomic_subtract_int(&bufmallocspace,
3239 bp->b_flags &= ~B_MALLOC;
3240 newbsize = round_page(newbsize);
3244 (vm_offset_t) bp->b_data + bp->b_bufsize,
3245 (vm_offset_t) bp->b_data + newbsize);
3247 bcopy(origbuf, bp->b_data, origbufsize);
3248 kfree(origbuf, M_BIOBUF);
3255 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3256 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3257 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3258 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3260 if (bp->b_flags & B_MALLOC)
3261 panic("allocbuf: VMIO buffer can't be malloced");
3263 * Set B_CACHE initially if buffer is 0 length or will become
3266 if (size == 0 || bp->b_bufsize == 0)
3267 bp->b_flags |= B_CACHE;
3269 if (newbsize < bp->b_bufsize) {
3271 * DEV_BSIZE aligned new buffer size is less then the
3272 * DEV_BSIZE aligned existing buffer size. Figure out
3273 * if we have to remove any pages.
3275 if (desiredpages < bp->b_xio.xio_npages) {
3276 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3278 * the page is not freed here -- it
3279 * is the responsibility of
3280 * vnode_pager_setsize
3282 m = bp->b_xio.xio_pages[i];
3283 KASSERT(m != bogus_page,
3284 ("allocbuf: bogus page found"));
3285 while (vm_page_sleep_busy(m, TRUE, "biodep"))
3288 bp->b_xio.xio_pages[i] = NULL;
3289 vm_page_unwire(m, 0);
3291 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3292 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3293 bp->b_xio.xio_npages = desiredpages;
3295 } else if (size > bp->b_bcount) {
3297 * We are growing the buffer, possibly in a
3298 * byte-granular fashion.
3306 * Step 1, bring in the VM pages from the object,
3307 * allocating them if necessary. We must clear
3308 * B_CACHE if these pages are not valid for the
3309 * range covered by the buffer.
3311 * critical section protection is required to protect
3312 * against interrupts unbusying and freeing pages
3313 * between our vm_page_lookup() and our
3314 * busycheck/wiring call.
3319 lwkt_gettoken(&vm_token);
3320 while (bp->b_xio.xio_npages < desiredpages) {
3324 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
3325 if ((m = vm_page_lookup(obj, pi)) == NULL) {
3327 * note: must allocate system pages
3328 * since blocking here could intefere
3329 * with paging I/O, no matter which
3332 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
3336 vm_page_flag_clear(m, PG_ZERO);
3337 bp->b_flags &= ~B_CACHE;
3338 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3339 ++bp->b_xio.xio_npages;
3345 * We found a page. If we have to sleep on it,
3346 * retry because it might have gotten freed out
3349 * We can only test PG_BUSY here. Blocking on
3350 * m->busy might lead to a deadlock:
3352 * vm_fault->getpages->cluster_read->allocbuf
3356 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
3358 vm_page_flag_clear(m, PG_ZERO);
3360 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3361 ++bp->b_xio.xio_npages;
3362 if (bp->b_act_count < m->act_count)
3363 bp->b_act_count = m->act_count;
3365 lwkt_reltoken(&vm_token);
3368 * Step 2. We've loaded the pages into the buffer,
3369 * we have to figure out if we can still have B_CACHE
3370 * set. Note that B_CACHE is set according to the
3371 * byte-granular range ( bcount and size ), not the
3372 * aligned range ( newbsize ).
3374 * The VM test is against m->valid, which is DEV_BSIZE
3375 * aligned. Needless to say, the validity of the data
3376 * needs to also be DEV_BSIZE aligned. Note that this
3377 * fails with NFS if the server or some other client
3378 * extends the file's EOF. If our buffer is resized,
3379 * B_CACHE may remain set! XXX
3382 toff = bp->b_bcount;
3383 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3385 while ((bp->b_flags & B_CACHE) && toff < size) {
3388 if (tinc > (size - toff))
3391 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3399 bp->b_xio.xio_pages[pi]
3406 * Step 3, fixup the KVM pmap. Remember that
3407 * bp->b_data is relative to bp->b_loffset, but
3408 * bp->b_loffset may be offset into the first page.
3411 bp->b_data = (caddr_t)
3412 trunc_page((vm_offset_t)bp->b_data);
3414 (vm_offset_t)bp->b_data,
3415 bp->b_xio.xio_pages,
3416 bp->b_xio.xio_npages
3418 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3419 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3423 /* adjust space use on already-dirty buffer */
3424 if (bp->b_flags & B_DELWRI) {
3425 spin_lock(&bufcspin);
3426 dirtybufspace += newbsize - bp->b_bufsize;
3427 if (bp->b_flags & B_HEAVY)
3428 dirtybufspacehw += newbsize - bp->b_bufsize;
3429 spin_unlock(&bufcspin);
3431 if (newbsize < bp->b_bufsize)
3433 bp->b_bufsize = newbsize; /* actual buffer allocation */
3434 bp->b_bcount = size; /* requested buffer size */
3441 * Wait for buffer I/O completion, returning error status. B_EINTR
3442 * is converted into an EINTR error but not cleared (since a chain
3443 * of biowait() calls may occur).
3445 * On return bpdone() will have been called but the buffer will remain
3446 * locked and will not have been brelse()'d.
3448 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3449 * likely still in progress on return.
3451 * NOTE! This operation is on a BIO, not a BUF.
3453 * NOTE! BIO_DONE is cleared by vn_strategy()
3458 _biowait(struct bio *bio, const char *wmesg, int to)
3460 struct buf *bp = bio->bio_buf;
3465 KKASSERT(bio == &bp->b_bio1);
3467 flags = bio->bio_flags;
3468 if (flags & BIO_DONE)
3470 tsleep_interlock(bio, 0);
3471 nflags = flags | BIO_WANT;
3472 tsleep_interlock(bio, 0);
3473 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3475 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3476 else if (bp->b_cmd == BUF_CMD_READ)
3477 error = tsleep(bio, PINTERLOCKED, "biord", to);
3479 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3481 kprintf("tsleep error biowait %d\n", error);
3490 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3491 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3492 if (bp->b_flags & B_EINTR)
3494 if (bp->b_flags & B_ERROR)
3495 return (bp->b_error ? bp->b_error : EIO);
3500 biowait(struct bio *bio, const char *wmesg)
3502 return(_biowait(bio, wmesg, 0));
3506 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3508 return(_biowait(bio, wmesg, to));
3512 * This associates a tracking count with an I/O. vn_strategy() and
3513 * dev_dstrategy() do this automatically but there are a few cases
3514 * where a vnode or device layer is bypassed when a block translation
3515 * is cached. In such cases bio_start_transaction() may be called on
3516 * the bypassed layers so the system gets an I/O in progress indication
3517 * for those higher layers.
3520 bio_start_transaction(struct bio *bio, struct bio_track *track)
3522 bio->bio_track = track;
3523 if (dsched_is_clear_buf_priv(bio->bio_buf))
3524 dsched_new_buf(bio->bio_buf);
3525 bio_track_ref(track);
3529 * Initiate I/O on a vnode.
3531 * SWAPCACHE OPERATION:
3533 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately
3534 * devfs also uses b_vp for fake buffers so we also have to check
3535 * that B_PAGING is 0. In this case the passed 'vp' is probably the
3536 * underlying block device. The swap assignments are related to the
3537 * buffer cache buffer's b_vp, not the passed vp.
3539 * The passed vp == bp->b_vp only in the case where the strategy call
3540 * is made on the vp itself for its own buffers (a regular file or
3541 * block device vp). The filesystem usually then re-calls vn_strategy()
3542 * after translating the request to an underlying device.
3544 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the
3545 * underlying buffer cache buffers.
3547 * We can only deal with page-aligned buffers at the moment, because
3548 * we can't tell what the real dirty state for pages straddling a buffer
3551 * In order to call swap_pager_strategy() we must provide the VM object
3552 * and base offset for the underlying buffer cache pages so it can find
3556 vn_strategy(struct vnode *vp, struct bio *bio)
3558 struct bio_track *track;
3559 struct buf *bp = bio->bio_buf;
3561 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3564 * Set when an I/O is issued on the bp. Cleared by consumers
3565 * (aka HAMMER), allowing the consumer to determine if I/O had
3566 * actually occurred.
3568 bp->b_flags |= B_IODEBUG;
3571 * Handle the swap cache intercept.
3573 if (vn_cache_strategy(vp, bio))
3577 * Otherwise do the operation through the filesystem
3579 if (bp->b_cmd == BUF_CMD_READ)
3580 track = &vp->v_track_read;
3582 track = &vp->v_track_write;
3583 KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3584 bio->bio_track = track;
3585 if (dsched_is_clear_buf_priv(bio->bio_buf))
3586 dsched_new_buf(bio->bio_buf);
3587 bio_track_ref(track);
3588 vop_strategy(*vp->v_ops, vp, bio);
3592 vn_cache_strategy(struct vnode *vp, struct bio *bio)
3594 struct buf *bp = bio->bio_buf;
3601 * Is this buffer cache buffer suitable for reading from
3604 if (vm_swapcache_read_enable == 0 ||
3605 bp->b_cmd != BUF_CMD_READ ||
3606 ((bp->b_flags & B_CLUSTER) == 0 &&
3607 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) ||
3608 ((int)bp->b_loffset & PAGE_MASK) != 0 ||
3609 (bp->b_bcount & PAGE_MASK) != 0) {
3614 * Figure out the original VM object (it will match the underlying
3615 * VM pages). Note that swap cached data uses page indices relative
3616 * to that object, not relative to bio->bio_offset.
3618 if (bp->b_flags & B_CLUSTER)
3619 object = vp->v_object;
3621 object = bp->b_vp->v_object;
3624 * In order to be able to use the swap cache all underlying VM
3625 * pages must be marked as such, and we can't have any bogus pages.
3627 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
3628 m = bp->b_xio.xio_pages[i];
3629 if ((m->flags & PG_SWAPPED) == 0)
3631 if (m == bogus_page)
3636 * If we are good then issue the I/O using swap_pager_strategy()
3638 if (i == bp->b_xio.xio_npages) {
3639 m = bp->b_xio.xio_pages[0];
3640 nbio = push_bio(bio);
3641 nbio->bio_offset = ptoa(m->pindex);
3642 KKASSERT(m->object == object);
3643 swap_pager_strategy(object, nbio);
3652 * Finish I/O on a buffer after all BIOs have been processed.
3653 * Called when the bio chain is exhausted or by biowait. If called
3654 * by biowait, elseit is typically 0.
3656 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3657 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3658 * assuming B_INVAL is clear.
3660 * For the VMIO case, we set B_CACHE if the op was a read and no
3661 * read error occured, or if the op was a write. B_CACHE is never
3662 * set if the buffer is invalid or otherwise uncacheable.
3664 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3665 * initiator to leave B_INVAL set to brelse the buffer out of existance
3666 * in the biodone routine.
3669 bpdone(struct buf *bp, int elseit)
3673 KASSERT(BUF_REFCNTNB(bp) > 0,
3674 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3675 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3676 ("biodone: bp %p already done!", bp));
3679 * No more BIOs are left. All completion functions have been dealt
3680 * with, now we clean up the buffer.
3683 bp->b_cmd = BUF_CMD_DONE;
3686 * Only reads and writes are processed past this point.
3688 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3689 if (cmd == BUF_CMD_FREEBLKS)
3690 bp->b_flags |= B_NOCACHE;
3697 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3698 * a lot worse. XXX - move this above the clearing of b_cmd
3700 if (LIST_FIRST(&bp->b_dep) != NULL)
3701 buf_complete(bp); /* MPSAFE */
3704 * A failed write must re-dirty the buffer unless B_INVAL
3705 * was set. Only applicable to normal buffers (with VPs).
3706 * vinum buffers may not have a vp.
3708 if (cmd == BUF_CMD_WRITE &&
3709 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3710 bp->b_flags &= ~B_NOCACHE;
3715 if (bp->b_flags & B_VMIO) {
3721 struct vnode *vp = bp->b_vp;
3725 #if defined(VFS_BIO_DEBUG)
3726 if (vp->v_auxrefs == 0)
3727 panic("biodone: zero vnode hold count");
3728 if ((vp->v_flag & VOBJBUF) == 0)
3729 panic("biodone: vnode is not setup for merged cache");
3732 foff = bp->b_loffset;
3733 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3734 KASSERT(obj != NULL, ("biodone: missing VM object"));
3736 #if defined(VFS_BIO_DEBUG)
3737 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3738 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3739 obj->paging_in_progress, bp->b_xio.xio_npages);
3744 * Set B_CACHE if the op was a normal read and no error
3745 * occured. B_CACHE is set for writes in the b*write()
3748 iosize = bp->b_bcount - bp->b_resid;
3749 if (cmd == BUF_CMD_READ &&
3750 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3751 bp->b_flags |= B_CACHE;
3754 lwkt_gettoken(&vm_token);
3755 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3759 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3764 * cleanup bogus pages, restoring the originals. Since
3765 * the originals should still be wired, we don't have
3766 * to worry about interrupt/freeing races destroying
3767 * the VM object association.
3769 m = bp->b_xio.xio_pages[i];
3770 if (m == bogus_page) {
3772 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3774 panic("biodone: page disappeared");
3775 bp->b_xio.xio_pages[i] = m;
3776 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3777 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3779 #if defined(VFS_BIO_DEBUG)
3780 if (OFF_TO_IDX(foff) != m->pindex) {
3781 kprintf("biodone: foff(%lu)/m->pindex(%ld) "
3783 (unsigned long)foff, (long)m->pindex);
3788 * In the write case, the valid and clean bits are
3789 * already changed correctly (see bdwrite()), so we
3790 * only need to do this here in the read case.
3792 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3793 vfs_clean_one_page(bp, i, m);
3795 vm_page_flag_clear(m, PG_ZERO);
3798 * when debugging new filesystems or buffer I/O
3799 * methods, this is the most common error that pops
3800 * up. if you see this, you have not set the page
3801 * busy flag correctly!!!
3804 kprintf("biodone: page busy < 0, "
3805 "pindex: %d, foff: 0x(%x,%x), "
3806 "resid: %d, index: %d\n",
3807 (int) m->pindex, (int)(foff >> 32),
3808 (int) foff & 0xffffffff, resid, i);
3809 if (!vn_isdisk(vp, NULL))
3810 kprintf(" iosize: %ld, loffset: %lld, "
3811 "flags: 0x%08x, npages: %d\n",
3812 bp->b_vp->v_mount->mnt_stat.f_iosize,
3813 (long long)bp->b_loffset,
3814 bp->b_flags, bp->b_xio.xio_npages);
3816 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3817 (long long)bp->b_loffset,
3818 bp->b_flags, bp->b_xio.xio_npages);
3819 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3820 m->valid, m->dirty, m->wire_count);
3821 panic("biodone: page busy < 0");
3823 vm_page_io_finish(m);
3824 vm_object_pip_subtract(obj, 1);
3825 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3828 bp->b_flags &= ~B_HASBOGUS;
3830 vm_object_pip_wakeupn(obj, 0);
3831 lwkt_reltoken(&vm_token);
3835 * Finish up by releasing the buffer. There are no more synchronous
3836 * or asynchronous completions, those were handled by bio_done
3840 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
3851 biodone(struct bio *bio)
3853 struct buf *bp = bio->bio_buf;
3855 runningbufwakeup(bp);
3858 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3861 biodone_t *done_func;
3862 struct bio_track *track;
3865 * BIO tracking. Most but not all BIOs are tracked.
3867 if ((track = bio->bio_track) != NULL) {
3868 bio_track_rel(track);
3869 bio->bio_track = NULL;
3873 * A bio_done function terminates the loop. The function
3874 * will be responsible for any further chaining and/or
3875 * buffer management.
3877 * WARNING! The done function can deallocate the buffer!
3879 if ((done_func = bio->bio_done) != NULL) {
3880 bio->bio_done = NULL;
3884 bio = bio->bio_prev;
3888 * If we've run out of bio's do normal [a]synchronous completion.
3894 * Synchronous biodone - this terminates a synchronous BIO.
3896 * bpdone() is called with elseit=FALSE, leaving the buffer completed
3897 * but still locked. The caller must brelse() the buffer after waiting
3901 biodone_sync(struct bio *bio)
3903 struct buf *bp = bio->bio_buf;
3907 KKASSERT(bio == &bp->b_bio1);
3911 flags = bio->bio_flags;
3912 nflags = (flags | BIO_DONE) & ~BIO_WANT;
3914 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3915 if (flags & BIO_WANT)
3925 * This routine is called in lieu of iodone in the case of
3926 * incomplete I/O. This keeps the busy status for pages
3930 vfs_unbusy_pages(struct buf *bp)
3934 runningbufwakeup(bp);
3936 lwkt_gettoken(&vm_token);
3937 if (bp->b_flags & B_VMIO) {
3938 struct vnode *vp = bp->b_vp;
3943 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3944 vm_page_t m = bp->b_xio.xio_pages[i];
3947 * When restoring bogus changes the original pages
3948 * should still be wired, so we are in no danger of
3949 * losing the object association and do not need
3950 * critical section protection particularly.
3952 if (m == bogus_page) {
3953 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3955 panic("vfs_unbusy_pages: page missing");
3957 bp->b_xio.xio_pages[i] = m;
3958 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3959 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3961 vm_object_pip_subtract(obj, 1);
3962 vm_page_flag_clear(m, PG_ZERO);
3963 vm_page_io_finish(m);
3965 bp->b_flags &= ~B_HASBOGUS;
3966 vm_object_pip_wakeupn(obj, 0);
3968 lwkt_reltoken(&vm_token);
3974 * This routine is called before a device strategy routine.
3975 * It is used to tell the VM system that paging I/O is in
3976 * progress, and treat the pages associated with the buffer
3977 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3978 * flag is handled to make sure that the object doesn't become
3981 * Since I/O has not been initiated yet, certain buffer flags
3982 * such as B_ERROR or B_INVAL may be in an inconsistant state
3983 * and should be ignored.
3988 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3991 struct lwp *lp = curthread->td_lwp;
3994 * The buffer's I/O command must already be set. If reading,
3995 * B_CACHE must be 0 (double check against callers only doing
3996 * I/O when B_CACHE is 0).
3998 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3999 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
4001 if (bp->b_flags & B_VMIO) {
4004 lwkt_gettoken(&vm_token);
4007 KASSERT(bp->b_loffset != NOOFFSET,
4008 ("vfs_busy_pages: no buffer offset"));
4011 * Loop until none of the pages are busy.
4014 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4015 vm_page_t m = bp->b_xio.xio_pages[i];
4017 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
4022 * Setup for I/O, soft-busy the page right now because
4023 * the next loop may block.
4025 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4026 vm_page_t m = bp->b_xio.xio_pages[i];
4028 vm_page_flag_clear(m, PG_ZERO);
4029 if ((bp->b_flags & B_CLUSTER) == 0) {
4030 vm_object_pip_add(obj, 1);
4031 vm_page_io_start(m);
4036 * Adjust protections for I/O and do bogus-page mapping.
4037 * Assume that vm_page_protect() can block (it can block
4038 * if VM_PROT_NONE, don't take any chances regardless).
4040 * In particular note that for writes we must incorporate
4041 * page dirtyness from the VM system into the buffer's
4044 * For reads we theoretically must incorporate page dirtyness
4045 * from the VM system to determine if the page needs bogus
4046 * replacement, but we shortcut the test by simply checking
4047 * that all m->valid bits are set, indicating that the page
4048 * is fully valid and does not need to be re-read. For any
4049 * VM system dirtyness the page will also be fully valid
4050 * since it was mapped at one point.
4053 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4054 vm_page_t m = bp->b_xio.xio_pages[i];
4056 vm_page_flag_clear(m, PG_ZERO); /* XXX */
4057 if (bp->b_cmd == BUF_CMD_WRITE) {
4059 * When readying a vnode-backed buffer for
4060 * a write we must zero-fill any invalid
4061 * portions of the backing VM pages, mark
4062 * it valid and clear related dirty bits.
4064 * vfs_clean_one_page() incorporates any
4065 * VM dirtyness and updates the b_dirtyoff
4066 * range (after we've made the page RO).
4068 * It is also expected that the pmap modified
4069 * bit has already been cleared by the
4070 * vm_page_protect(). We may not be able
4071 * to clear all dirty bits for a page if it
4072 * was also memory mapped (NFS).
4074 * Finally be sure to unassign any swap-cache
4075 * backing store as it is now stale.
4077 vm_page_protect(m, VM_PROT_READ);
4078 vfs_clean_one_page(bp, i, m);
4079 swap_pager_unswapped(m);
4080 } else if (m->valid == VM_PAGE_BITS_ALL) {
4082 * When readying a vnode-backed buffer for
4083 * read we must replace any dirty pages with
4084 * a bogus page so dirty data is not destroyed
4085 * when filling gaps.
4087 * To avoid testing whether the page is
4088 * dirty we instead test that the page was
4089 * at some point mapped (m->valid fully
4090 * valid) with the understanding that
4091 * this also covers the dirty case.
4093 bp->b_xio.xio_pages[i] = bogus_page;
4094 bp->b_flags |= B_HASBOGUS;
4096 } else if (m->valid & m->dirty) {
4098 * This case should not occur as partial
4099 * dirtyment can only happen if the buffer
4100 * is B_CACHE, and this code is not entered
4101 * if the buffer is B_CACHE.
4103 kprintf("Warning: vfs_busy_pages - page not "
4104 "fully valid! loff=%jx bpf=%08x "
4105 "idx=%d val=%02x dir=%02x\n",
4106 (intmax_t)bp->b_loffset, bp->b_flags,
4107 i, m->valid, m->dirty);
4108 vm_page_protect(m, VM_PROT_NONE);
4111 * The page is not valid and can be made
4114 vm_page_protect(m, VM_PROT_NONE);
4118 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4119 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
4121 lwkt_reltoken(&vm_token);
4125 * This is the easiest place to put the process accounting for the I/O
4129 if (bp->b_cmd == BUF_CMD_READ)
4130 lp->lwp_ru.ru_inblock++;
4132 lp->lwp_ru.ru_oublock++;
4139 * Tell the VM system that the pages associated with this buffer
4140 * are clean. This is used for delayed writes where the data is
4141 * going to go to disk eventually without additional VM intevention.
4143 * Note that while we only really need to clean through to b_bcount, we
4144 * just go ahead and clean through to b_bufsize.
4147 vfs_clean_pages(struct buf *bp)
4152 if ((bp->b_flags & B_VMIO) == 0)
4155 KASSERT(bp->b_loffset != NOOFFSET,
4156 ("vfs_clean_pages: no buffer offset"));
4158 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4159 m = bp->b_xio.xio_pages[i];
4160 vfs_clean_one_page(bp, i, m);
4165 * vfs_clean_one_page:
4167 * Set the valid bits and clear the dirty bits in a page within a
4168 * buffer. The range is restricted to the buffer's size and the
4169 * buffer's logical offset might index into the first page.
4171 * The caller has busied or soft-busied the page and it is not mapped,
4172 * test and incorporate the dirty bits into b_dirtyoff/end before
4173 * clearing them. Note that we need to clear the pmap modified bits
4174 * after determining the the page was dirty, vm_page_set_validclean()
4175 * does not do it for us.
4177 * This routine is typically called after a read completes (dirty should
4178 * be zero in that case as we are not called on bogus-replace pages),
4179 * or before a write is initiated.
4182 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m)
4190 * Calculate offset range within the page but relative to buffer's
4191 * loffset. loffset might be offset into the first page.
4193 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4194 bcount = bp->b_bcount + xoff; /* offset adjusted */
4200 soff = (pageno << PAGE_SHIFT);
4201 eoff = soff + PAGE_SIZE;
4209 * Test dirty bits and adjust b_dirtyoff/end.
4211 * If dirty pages are incorporated into the bp any prior
4212 * B_NEEDCOMMIT state (NFS) must be cleared because the
4213 * caller has not taken into account the new dirty data.
4215 * If the page was memory mapped the dirty bits might go beyond the
4216 * end of the buffer, but we can't really make the assumption that
4217 * a file EOF straddles the buffer (even though this is the case for
4218 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
4219 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
4220 * This also saves some console spam.
4222 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK,
4223 * NFS can handle huge commits but not huge writes.
4225 vm_page_test_dirty(m);
4227 if ((bp->b_flags & B_NEEDCOMMIT) &&
4228 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) {
4230 kprintf("Warning: vfs_clean_one_page: bp %p "
4231 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT"
4232 " cmd %d vd %02x/%02x x/s/e %d %d %d "
4234 bp, (intmax_t)bp->b_loffset, bp->b_bcount,
4235 bp->b_flags, bp->b_cmd,
4236 m->valid, m->dirty, xoff, soff, eoff,
4237 bp->b_dirtyoff, bp->b_dirtyend);
4238 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
4240 print_backtrace(-1);
4243 * Only clear the pmap modified bits if ALL the dirty bits
4244 * are set, otherwise the system might mis-clear portions
4247 if (m->dirty == VM_PAGE_BITS_ALL &&
4248 (bp->b_flags & B_NEEDCOMMIT) == 0) {
4249 pmap_clear_modify(m);
4251 if (bp->b_dirtyoff > soff - xoff)
4252 bp->b_dirtyoff = soff - xoff;
4253 if (bp->b_dirtyend < eoff - xoff)
4254 bp->b_dirtyend = eoff - xoff;
4258 * Set related valid bits, clear related dirty bits.
4259 * Does not mess with the pmap modified bit.
4261 * WARNING! We cannot just clear all of m->dirty here as the
4262 * buffer cache buffers may use a DEV_BSIZE'd aligned
4263 * block size, or have an odd size (e.g. NFS at file EOF).
4264 * The putpages code can clear m->dirty to 0.
4266 * If a VOP_WRITE generates a buffer cache buffer which
4267 * covers the same space as mapped writable pages the
4268 * buffer flush might not be able to clear all the dirty
4269 * bits and still require a putpages from the VM system
4272 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff);
4276 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty.
4277 * The page data is assumed to be valid (there is no zeroing here).
4280 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m)
4288 * Calculate offset range within the page but relative to buffer's
4289 * loffset. loffset might be offset into the first page.
4291 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4292 bcount = bp->b_bcount + xoff; /* offset adjusted */
4298 soff = (pageno << PAGE_SHIFT);
4299 eoff = soff + PAGE_SIZE;
4305 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff);
4311 * Clear a buffer. This routine essentially fakes an I/O, so we need
4312 * to clear B_ERROR and B_INVAL.
4314 * Note that while we only theoretically need to clear through b_bcount,
4315 * we go ahead and clear through b_bufsize.
4319 vfs_bio_clrbuf(struct buf *bp)
4323 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
4324 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
4325 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4326 (bp->b_loffset & PAGE_MASK) == 0) {
4327 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4328 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
4332 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
4333 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
4334 bzero(bp->b_data, bp->b_bufsize);
4335 bp->b_xio.xio_pages[0]->valid |= mask;
4341 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
4342 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
4343 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
4344 ea = (caddr_t)(vm_offset_t)ulmin(
4345 (u_long)(vm_offset_t)ea,
4346 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
4347 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4348 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
4350 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
4351 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
4355 for (; sa < ea; sa += DEV_BSIZE, j++) {
4356 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
4357 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
4358 bzero(sa, DEV_BSIZE);
4361 bp->b_xio.xio_pages[i]->valid |= mask;
4362 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
4371 * vm_hold_load_pages:
4373 * Load pages into the buffer's address space. The pages are
4374 * allocated from the kernel object in order to reduce interference
4375 * with the any VM paging I/O activity. The range of loaded
4376 * pages will be wired.
4378 * If a page cannot be allocated, the 'pagedaemon' is woken up to
4379 * retrieve the full range (to - from) of pages.
4384 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4390 to = round_page(to);
4391 from = round_page(from);
4392 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4397 * Note: must allocate system pages since blocking here
4398 * could intefere with paging I/O, no matter which
4401 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT,
4402 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
4405 p->valid = VM_PAGE_BITS_ALL;
4406 vm_page_flag_clear(p, PG_ZERO);
4407 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
4408 bp->b_xio.xio_pages[index] = p;
4415 bp->b_xio.xio_npages = index;
4419 * Allocate pages for a buffer cache buffer.
4421 * Under extremely severe memory conditions even allocating out of the
4422 * system reserve can fail. If this occurs we must allocate out of the
4423 * interrupt reserve to avoid a deadlock with the pageout daemon.
4425 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
4426 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
4427 * against the pageout daemon if pages are not freed from other sources.
4433 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit)
4438 * Try a normal allocation, allow use of system reserve.
4440 lwkt_gettoken(&vm_token);
4441 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
4443 lwkt_reltoken(&vm_token);
4448 * The normal allocation failed and we clearly have a page
4449 * deficit. Try to reclaim some clean VM pages directly
4450 * from the buffer cache.
4452 vm_pageout_deficit += deficit;
4456 * We may have blocked, the caller will know what to do if the
4459 if (vm_page_lookup(obj, pg)) {
4460 lwkt_reltoken(&vm_token);
4465 * Allocate and allow use of the interrupt reserve.
4467 * If after all that we still can't allocate a VM page we are
4468 * in real trouble, but we slog on anyway hoping that the system
4471 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
4472 VM_ALLOC_INTERRUPT);
4474 if (vm_page_count_severe()) {
4475 kprintf("bio_page_alloc: WARNING emergency page "
4480 kprintf("bio_page_alloc: WARNING emergency page "
4481 "allocation failed\n");
4484 lwkt_reltoken(&vm_token);
4489 * vm_hold_free_pages:
4491 * Return pages associated with the buffer back to the VM system.
4493 * The range of pages underlying the buffer's address space will
4494 * be unmapped and un-wired.
4499 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4503 int index, newnpages;
4505 from = round_page(from);
4506 to = round_page(to);
4507 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4510 lwkt_gettoken(&vm_token);
4511 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4512 p = bp->b_xio.xio_pages[index];
4513 if (p && (index < bp->b_xio.xio_npages)) {
4515 kprintf("vm_hold_free_pages: doffset: %lld, "
4517 (long long)bp->b_bio2.bio_offset,
4518 (long long)bp->b_loffset);
4520 bp->b_xio.xio_pages[index] = NULL;
4523 vm_page_unwire(p, 0);
4527 bp->b_xio.xio_npages = newnpages;
4528 lwkt_reltoken(&vm_token);
4534 * Map a user buffer into KVM via a pbuf. On return the buffer's
4535 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4539 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4550 * bp had better have a command and it better be a pbuf.
4552 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4553 KKASSERT(bp->b_flags & B_PAGING);
4554 KKASSERT(bp->b_kvabase);
4560 * Map the user data into KVM. Mappings have to be page-aligned.
4562 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4565 vmprot = VM_PROT_READ;
4566 if (bp->b_cmd == BUF_CMD_READ)
4567 vmprot |= VM_PROT_WRITE;
4569 while (addr < udata + bytes) {
4571 * Do the vm_fault if needed; do the copy-on-write thing
4572 * when reading stuff off device into memory.
4574 * vm_fault_page*() returns a held VM page.
4576 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4577 va = trunc_page(va);
4579 m = vm_fault_page_quick(va, vmprot, &error);
4581 for (i = 0; i < pidx; ++i) {
4582 vm_page_unhold(bp->b_xio.xio_pages[i]);
4583 bp->b_xio.xio_pages[i] = NULL;
4587 bp->b_xio.xio_pages[pidx] = m;
4593 * Map the page array and set the buffer fields to point to
4594 * the mapped data buffer.
4596 if (pidx > btoc(MAXPHYS))
4597 panic("vmapbuf: mapped more than MAXPHYS");
4598 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4600 bp->b_xio.xio_npages = pidx;
4601 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4602 bp->b_bcount = bytes;
4603 bp->b_bufsize = bytes;
4610 * Free the io map PTEs associated with this IO operation.
4611 * We also invalidate the TLB entries and restore the original b_addr.
4614 vunmapbuf(struct buf *bp)
4619 KKASSERT(bp->b_flags & B_PAGING);
4621 npages = bp->b_xio.xio_npages;
4622 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4623 for (pidx = 0; pidx < npages; ++pidx) {
4624 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4625 bp->b_xio.xio_pages[pidx] = NULL;
4627 bp->b_xio.xio_npages = 0;
4628 bp->b_data = bp->b_kvabase;
4632 * Scan all buffers in the system and issue the callback.
4635 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4641 for (n = 0; n < nbuf; ++n) {
4642 if ((error = callback(&buf[n], info)) < 0) {
4652 * nestiobuf_iodone: biodone callback for nested buffers and propagate
4653 * completion to the master buffer.
4656 nestiobuf_iodone(struct bio *bio)
4659 struct buf *mbp, *bp;
4664 mbio = bio->bio_caller_info1.ptr;
4665 mbp = mbio->bio_buf;
4667 KKASSERT(bp->b_bcount <= bp->b_bufsize);
4668 KKASSERT(mbp != bp);
4670 error = bp->b_error;
4671 if (bp->b_error == 0 &&
4672 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
4674 * Not all got transfered, raise an error. We have no way to
4675 * propagate these conditions to mbp.
4680 donebytes = bp->b_bufsize;
4683 nestiobuf_done(mbio, donebytes, error);
4687 nestiobuf_done(struct bio *mbio, int donebytes, int error)
4691 mbp = mbio->bio_buf;
4693 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0);
4696 * If an error occured, propagate it to the master buffer.
4698 * Several biodone()s may wind up running concurrently so
4699 * use an atomic op to adjust b_flags.
4702 mbp->b_error = error;
4703 atomic_set_int(&mbp->b_flags, B_ERROR);
4707 * Decrement the operations in progress counter and terminate the
4708 * I/O if this was the last bit.
4710 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4717 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent
4718 * the mbio from being biodone()'d while we are still adding sub-bios to
4722 nestiobuf_init(struct bio *bio)
4724 bio->bio_driver_info = (void *)1;
4728 * The BIOs added to the nestedio have already been started, remove the
4729 * count that placeheld our mbio and biodone() it if the count would
4733 nestiobuf_start(struct bio *mbio)
4735 struct buf *mbp = mbio->bio_buf;
4738 * Decrement the operations in progress counter and terminate the
4739 * I/O if this was the last bit.
4741 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4742 if (mbp->b_flags & B_ERROR)
4743 mbp->b_resid = mbp->b_bcount;
4751 * Set an intermediate error prior to calling nestiobuf_start()
4754 nestiobuf_error(struct bio *mbio, int error)
4756 struct buf *mbp = mbio->bio_buf;
4759 mbp->b_error = error;
4760 atomic_set_int(&mbp->b_flags, B_ERROR);
4765 * nestiobuf_add: setup a "nested" buffer.
4767 * => 'mbp' is a "master" buffer which is being divided into sub pieces.
4768 * => 'bp' should be a buffer allocated by getiobuf.
4769 * => 'offset' is a byte offset in the master buffer.
4770 * => 'size' is a size in bytes of this nested buffer.
4773 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size)
4775 struct buf *mbp = mbio->bio_buf;
4776 struct vnode *vp = mbp->b_vp;
4778 KKASSERT(mbp->b_bcount >= offset + size);
4780 atomic_add_int((int *)&mbio->bio_driver_info, 1);
4782 /* kernel needs to own the lock for it to be released in biodone */
4785 bp->b_cmd = mbp->b_cmd;
4786 bp->b_bio1.bio_done = nestiobuf_iodone;
4787 bp->b_data = (char *)mbp->b_data + offset;
4788 bp->b_resid = bp->b_bcount = size;
4789 bp->b_bufsize = bp->b_bcount;
4791 bp->b_bio1.bio_track = NULL;
4792 bp->b_bio1.bio_caller_info1.ptr = mbio;
4796 * print out statistics from the current status of the buffer pool
4797 * this can be toggeled by the system control option debug.syncprt
4806 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
4807 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
4809 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
4811 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4814 spin_lock(&bufqspin);
4815 TAILQ_FOREACH(bp, dp, b_freelist) {
4816 counts[bp->b_bufsize/PAGE_SIZE]++;
4819 spin_unlock(&bufqspin);
4821 kprintf("%s: total-%d", bname[i], count);
4822 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4824 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
4832 DB_SHOW_COMMAND(buffer, db_show_buffer)
4835 struct buf *bp = (struct buf *)addr;
4838 db_printf("usage: show buffer <addr>\n");
4842 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4843 db_printf("b_cmd = %d\n", bp->b_cmd);
4844 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4845 "b_resid = %d\n, b_data = %p, "
4846 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4847 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4849 (long long)bp->b_bio2.bio_offset,
4850 (long long)(bp->b_bio2.bio_next ?
4851 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4852 if (bp->b_xio.xio_npages) {
4854 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4855 bp->b_xio.xio_npages);
4856 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4858 m = bp->b_xio.xio_pages[i];
4859 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4860 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4861 if ((i + 1) < bp->b_xio.xio_npages)