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/devicestat.h>
36 #include <sys/eventhandler.h>
38 #include <sys/malloc.h>
39 #include <sys/mount.h>
40 #include <sys/kernel.h>
41 #include <sys/kthread.h>
43 #include <sys/reboot.h>
44 #include <sys/resourcevar.h>
45 #include <sys/sysctl.h>
46 #include <sys/vmmeter.h>
47 #include <sys/vnode.h>
48 #include <sys/dsched.h>
51 #include <vm/vm_param.h>
52 #include <vm/vm_kern.h>
53 #include <vm/vm_pageout.h>
54 #include <vm/vm_page.h>
55 #include <vm/vm_object.h>
56 #include <vm/vm_extern.h>
57 #include <vm/vm_map.h>
58 #include <vm/vm_pager.h>
59 #include <vm/swap_pager.h>
62 #include <sys/thread2.h>
63 #include <sys/spinlock2.h>
64 #include <sys/mplock2.h>
65 #include <vm/vm_page2.h>
76 BQUEUE_NONE, /* not on any queue */
77 BQUEUE_LOCKED, /* locked buffers */
78 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
79 BQUEUE_DIRTY, /* B_DELWRI buffers */
80 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
81 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
82 BQUEUE_EMPTY, /* empty buffer headers */
84 BUFFER_QUEUES /* number of buffer queues */
87 typedef enum bufq_type bufq_type_t;
89 #define BD_WAKE_SIZE 16384
90 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
92 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
93 static struct spinlock bufqspin = SPINLOCK_INITIALIZER(&bufqspin);
94 static struct spinlock bufcspin = SPINLOCK_INITIALIZER(&bufcspin);
96 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
98 struct buf *buf; /* buffer header pool */
100 static void vfs_clean_pages(struct buf *bp);
101 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
102 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m);
103 static void vfs_vmio_release(struct buf *bp);
104 static int flushbufqueues(bufq_type_t q);
105 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);
107 static void bd_signal(int totalspace);
108 static void buf_daemon(void);
109 static void buf_daemon_hw(void);
112 * bogus page -- for I/O to/from partially complete buffers
113 * this is a temporary solution to the problem, but it is not
114 * really that bad. it would be better to split the buffer
115 * for input in the case of buffers partially already in memory,
116 * but the code is intricate enough already.
118 vm_page_t bogus_page;
121 * These are all static, but make the ones we export globals so we do
122 * not need to use compiler magic.
124 int bufspace; /* locked by buffer_map */
126 static int bufmallocspace; /* atomic ops */
127 int maxbufmallocspace, lobufspace, hibufspace;
128 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
129 static int lorunningspace;
130 static int hirunningspace;
131 static int runningbufreq; /* locked by bufcspin */
132 static int dirtybufspace; /* locked by bufcspin */
133 static int dirtybufcount; /* locked by bufcspin */
134 static int dirtybufspacehw; /* locked by bufcspin */
135 static int dirtybufcounthw; /* locked by bufcspin */
136 static int runningbufspace; /* locked by bufcspin */
137 static int runningbufcount; /* locked by bufcspin */
140 static int getnewbufcalls;
141 static int getnewbufrestarts;
142 static int recoverbufcalls;
143 static int needsbuffer; /* locked by bufcspin */
144 static int bd_request; /* locked by bufcspin */
145 static int bd_request_hw; /* locked by bufcspin */
146 static u_int bd_wake_ary[BD_WAKE_SIZE];
147 static u_int bd_wake_index;
148 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */
149 static int debug_commit;
151 static struct thread *bufdaemon_td;
152 static struct thread *bufdaemonhw_td;
153 static u_int lowmempgallocs;
154 static u_int lowmempgfails;
157 * Sysctls for operational control of the buffer cache.
159 SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
160 "Number of dirty buffers to flush before bufdaemon becomes inactive");
161 SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
162 "High watermark used to trigger explicit flushing of dirty buffers");
163 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
164 "Minimum amount of buffer space required for active I/O");
165 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
166 "Maximum amount of buffer space to usable for active I/O");
167 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0,
168 "Page allocations done during periods of very low free memory");
169 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0,
170 "Page allocations which failed during periods of very low free memory");
171 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
172 "Recycle pages to active or inactive queue transition pt 0-64");
174 * Sysctls determining current state of the buffer cache.
176 SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
177 "Total number of buffers in buffer cache");
178 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
179 "Pending bytes of dirty buffers (all)");
180 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
181 "Pending bytes of dirty buffers (heavy weight)");
182 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
183 "Pending number of dirty buffers");
184 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
185 "Pending number of dirty buffers (heavy weight)");
186 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
187 "I/O bytes currently in progress due to asynchronous writes");
188 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
189 "I/O buffers currently in progress due to asynchronous writes");
190 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
191 "Hard limit on maximum amount of memory usable for buffer space");
192 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
193 "Soft limit on maximum amount of memory usable for buffer space");
194 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
195 "Minimum amount of memory to reserve for system buffer space");
196 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
197 "Amount of memory available for buffers");
198 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
199 0, "Maximum amount of memory reserved for buffers using malloc");
200 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
201 "Amount of memory left for buffers using malloc-scheme");
202 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
203 "New buffer header acquisition requests");
204 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
205 0, "New buffer header acquisition restarts");
206 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
207 "Recover VM space in an emergency");
208 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
209 "Buffer acquisition restarts due to fragmented buffer map");
210 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
211 "Amount of time KVA space was deallocated in an arbitrary buffer");
212 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
213 "Amount of time buffer re-use operations were successful");
214 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, "");
215 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
216 "sizeof(struct buf)");
218 char *buf_wmesg = BUF_WMESG;
220 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
221 #define VFS_BIO_NEED_UNUSED02 0x02
222 #define VFS_BIO_NEED_UNUSED04 0x04
223 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
228 * Called when buffer space is potentially available for recovery.
229 * getnewbuf() will block on this flag when it is unable to free
230 * sufficient buffer space. Buffer space becomes recoverable when
231 * bp's get placed back in the queues.
237 * If someone is waiting for BUF space, wake them up. Even
238 * though we haven't freed the kva space yet, the waiting
239 * process will be able to now.
241 spin_lock(&bufcspin);
242 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
243 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
244 spin_unlock(&bufcspin);
245 wakeup(&needsbuffer);
247 spin_unlock(&bufcspin);
254 * Accounting for I/O in progress.
258 runningbufwakeup(struct buf *bp)
263 if ((totalspace = bp->b_runningbufspace) != 0) {
264 spin_lock(&bufcspin);
265 runningbufspace -= totalspace;
267 bp->b_runningbufspace = 0;
270 * see waitrunningbufspace() for limit test.
272 limit = hirunningspace * 4 / 6;
273 if (runningbufreq && runningbufspace <= limit) {
275 spin_unlock(&bufcspin);
276 wakeup(&runningbufreq);
278 spin_unlock(&bufcspin);
280 bd_signal(totalspace);
287 * Called when a buffer has been added to one of the free queues to
288 * account for the buffer and to wakeup anyone waiting for free buffers.
289 * This typically occurs when large amounts of metadata are being handled
290 * by the buffer cache ( else buffer space runs out first, usually ).
297 spin_lock(&bufcspin);
299 needsbuffer &= ~VFS_BIO_NEED_ANY;
300 spin_unlock(&bufcspin);
301 wakeup(&needsbuffer);
303 spin_unlock(&bufcspin);
308 * waitrunningbufspace()
310 * Wait for the amount of running I/O to drop to hirunningspace * 4 / 6.
311 * This is the point where write bursting stops so we don't want to wait
312 * for the running amount to drop below it (at least if we still want bioq
315 * The caller may be using this function to block in a tight loop, we
316 * must block while runningbufspace is greater then or equal to
317 * hirunningspace * 4 / 6.
319 * And even with that it may not be enough, due to the presence of
320 * B_LOCKED dirty buffers, so also wait for at least one running buffer
324 waitrunningbufspace(void)
326 int limit = hirunningspace * 4 / 6;
329 spin_lock(&bufcspin);
330 if (runningbufspace > limit) {
331 while (runningbufspace > limit) {
333 ssleep(&runningbufreq, &bufcspin, 0, "wdrn1", 0);
335 spin_unlock(&bufcspin);
336 } else if (runningbufspace > limit / 2) {
338 spin_unlock(&bufcspin);
339 tsleep(&dummy, 0, "wdrn2", 1);
341 spin_unlock(&bufcspin);
346 * buf_dirty_count_severe:
348 * Return true if we have too many dirty buffers.
351 buf_dirty_count_severe(void)
353 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
354 dirtybufcount >= nbuf / 2);
358 * Return true if the amount of running I/O is severe and BIOQ should
362 buf_runningbufspace_severe(void)
364 return (runningbufspace >= hirunningspace * 4 / 6);
368 * vfs_buf_test_cache:
370 * Called when a buffer is extended. This function clears the B_CACHE
371 * bit if the newly extended portion of the buffer does not contain
374 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
375 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
376 * them while a clean buffer was present.
380 vfs_buf_test_cache(struct buf *bp,
381 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
384 if (bp->b_flags & B_CACHE) {
385 int base = (foff + off) & PAGE_MASK;
386 if (vm_page_is_valid(m, base, size) == 0)
387 bp->b_flags &= ~B_CACHE;
394 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
403 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
406 if (bd_request == 0 &&
407 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
408 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
409 spin_lock(&bufcspin);
411 spin_unlock(&bufcspin);
414 if (bd_request_hw == 0 &&
415 (dirtybufspacehw > lodirtybufspace / 2 ||
416 dirtybufcounthw >= nbuf / 2)) {
417 spin_lock(&bufcspin);
419 spin_unlock(&bufcspin);
420 wakeup(&bd_request_hw);
427 * Get the buf_daemon heated up when the number of running and dirty
428 * buffers exceeds the mid-point.
430 * Return the total number of dirty bytes past the second mid point
431 * as a measure of how much excess dirty data there is in the system.
442 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
444 totalspace = runningbufspace + dirtybufspace;
445 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
447 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
448 if (totalspace >= mid2)
449 return(totalspace - mid2);
457 * Wait for the buffer cache to flush (totalspace) bytes worth of
458 * buffers, then return.
460 * Regardless this function blocks while the number of dirty buffers
461 * exceeds hidirtybufspace.
466 bd_wait(int totalspace)
471 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
474 while (totalspace > 0) {
476 if (totalspace > runningbufspace + dirtybufspace)
477 totalspace = runningbufspace + dirtybufspace;
478 count = totalspace / BKVASIZE;
479 if (count >= BD_WAKE_SIZE)
480 count = BD_WAKE_SIZE - 1;
482 spin_lock(&bufcspin);
483 i = (bd_wake_index + count) & BD_WAKE_MASK;
487 * This is not a strict interlock, so we play a bit loose
488 * with locking access to dirtybufspace*
490 tsleep_interlock(&bd_wake_ary[i], 0);
491 spin_unlock(&bufcspin);
492 tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz);
494 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
501 * This function is called whenever runningbufspace or dirtybufspace
502 * is reduced. Track threads waiting for run+dirty buffer I/O
508 bd_signal(int totalspace)
512 if (totalspace > 0) {
513 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
514 totalspace = BKVASIZE * BD_WAKE_SIZE;
515 spin_lock(&bufcspin);
516 while (totalspace > 0) {
519 if (bd_wake_ary[i]) {
521 spin_unlock(&bufcspin);
522 wakeup(&bd_wake_ary[i]);
523 spin_lock(&bufcspin);
525 totalspace -= BKVASIZE;
527 spin_unlock(&bufcspin);
532 * BIO tracking support routines.
534 * Release a ref on a bio_track. Wakeup requests are atomically released
535 * along with the last reference so bk_active will never wind up set to
542 bio_track_rel(struct bio_track *track)
550 active = track->bk_active;
551 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
555 * Full-on. Note that the wait flag is only atomically released on
556 * the 1->0 count transition.
558 * We check for a negative count transition using bit 30 since bit 31
559 * has a different meaning.
562 desired = (active & 0x7FFFFFFF) - 1;
564 desired |= active & 0x80000000;
565 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
566 if (desired & 0x40000000)
567 panic("bio_track_rel: bad count: %p\n", track);
568 if (active & 0x80000000)
572 active = track->bk_active;
577 * Wait for the tracking count to reach 0.
579 * Use atomic ops such that the wait flag is only set atomically when
580 * bk_active is non-zero.
585 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
594 if (track->bk_active == 0)
598 * Full-on. Note that the wait flag may only be atomically set if
599 * the active count is non-zero.
601 * NOTE: We cannot optimize active == desired since a wakeup could
602 * clear active prior to our tsleep_interlock().
605 while ((active = track->bk_active) != 0) {
607 desired = active | 0x80000000;
608 tsleep_interlock(track, slp_flags);
609 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
610 error = tsleep(track, slp_flags | PINTERLOCKED,
622 * Load time initialisation of the buffer cache, called from machine
623 * dependant initialization code.
629 vm_offset_t bogus_offset;
632 /* next, make a null set of free lists */
633 for (i = 0; i < BUFFER_QUEUES; i++)
634 TAILQ_INIT(&bufqueues[i]);
636 /* finally, initialize each buffer header and stick on empty q */
637 for (i = 0; i < nbuf; i++) {
639 bzero(bp, sizeof *bp);
640 bp->b_flags = B_INVAL; /* we're just an empty header */
641 bp->b_cmd = BUF_CMD_DONE;
642 bp->b_qindex = BQUEUE_EMPTY;
644 xio_init(&bp->b_xio);
646 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
650 * maxbufspace is the absolute maximum amount of buffer space we are
651 * allowed to reserve in KVM and in real terms. The absolute maximum
652 * is nominally used by buf_daemon. hibufspace is the nominal maximum
653 * used by most other processes. The differential is required to
654 * ensure that buf_daemon is able to run when other processes might
655 * be blocked waiting for buffer space.
657 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
658 * this may result in KVM fragmentation which is not handled optimally
661 maxbufspace = nbuf * BKVASIZE;
662 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
663 lobufspace = hibufspace - MAXBSIZE;
665 lorunningspace = 512 * 1024;
666 /* hirunningspace -- see below */
669 * Limit the amount of malloc memory since it is wired permanently
670 * into the kernel space. Even though this is accounted for in
671 * the buffer allocation, we don't want the malloced region to grow
672 * uncontrolled. The malloc scheme improves memory utilization
673 * significantly on average (small) directories.
675 maxbufmallocspace = hibufspace / 20;
678 * Reduce the chance of a deadlock occuring by limiting the number
679 * of delayed-write dirty buffers we allow to stack up.
681 * We don't want too much actually queued to the device at once
682 * (XXX this needs to be per-mount!), because the buffers will
683 * wind up locked for a very long period of time while the I/O
686 hidirtybufspace = hibufspace / 2; /* dirty + running */
687 hirunningspace = hibufspace / 16; /* locked & queued to device */
688 if (hirunningspace < 1024 * 1024)
689 hirunningspace = 1024 * 1024;
694 lodirtybufspace = hidirtybufspace / 2;
697 * Maximum number of async ops initiated per buf_daemon loop. This is
698 * somewhat of a hack at the moment, we really need to limit ourselves
699 * based on the number of bytes of I/O in-transit that were initiated
703 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
704 vm_object_hold(&kernel_object);
705 bogus_page = vm_page_alloc(&kernel_object,
706 (bogus_offset >> PAGE_SHIFT),
708 vm_object_drop(&kernel_object);
709 vmstats.v_wire_count++;
714 * Initialize the embedded bio structures, typically used by
715 * deprecated code which tries to allocate its own struct bufs.
718 initbufbio(struct buf *bp)
720 bp->b_bio1.bio_buf = bp;
721 bp->b_bio1.bio_prev = NULL;
722 bp->b_bio1.bio_offset = NOOFFSET;
723 bp->b_bio1.bio_next = &bp->b_bio2;
724 bp->b_bio1.bio_done = NULL;
725 bp->b_bio1.bio_flags = 0;
727 bp->b_bio2.bio_buf = bp;
728 bp->b_bio2.bio_prev = &bp->b_bio1;
729 bp->b_bio2.bio_offset = NOOFFSET;
730 bp->b_bio2.bio_next = NULL;
731 bp->b_bio2.bio_done = NULL;
732 bp->b_bio2.bio_flags = 0;
738 * Reinitialize the embedded bio structures as well as any additional
739 * translation cache layers.
742 reinitbufbio(struct buf *bp)
746 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
747 bio->bio_done = NULL;
748 bio->bio_offset = NOOFFSET;
753 * Undo the effects of an initbufbio().
756 uninitbufbio(struct buf *bp)
763 * Push another BIO layer onto an existing BIO and return it. The new
764 * BIO layer may already exist, holding cached translation data.
767 push_bio(struct bio *bio)
771 if ((nbio = bio->bio_next) == NULL) {
772 int index = bio - &bio->bio_buf->b_bio_array[0];
773 if (index >= NBUF_BIO - 1) {
774 panic("push_bio: too many layers bp %p\n",
777 nbio = &bio->bio_buf->b_bio_array[index + 1];
778 bio->bio_next = nbio;
779 nbio->bio_prev = bio;
780 nbio->bio_buf = bio->bio_buf;
781 nbio->bio_offset = NOOFFSET;
782 nbio->bio_done = NULL;
783 nbio->bio_next = NULL;
785 KKASSERT(nbio->bio_done == NULL);
790 * Pop a BIO translation layer, returning the previous layer. The
791 * must have been previously pushed.
794 pop_bio(struct bio *bio)
796 return(bio->bio_prev);
800 clearbiocache(struct bio *bio)
803 bio->bio_offset = NOOFFSET;
811 * Free the KVA allocation for buffer 'bp'.
813 * Must be called from a critical section as this is the only locking for
816 * Since this call frees up buffer space, we call bufspacewakeup().
821 bfreekva(struct buf *bp)
827 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
828 vm_map_lock(&buffer_map);
829 bufspace -= bp->b_kvasize;
830 vm_map_delete(&buffer_map,
831 (vm_offset_t) bp->b_kvabase,
832 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
835 vm_map_unlock(&buffer_map);
836 vm_map_entry_release(count);
838 bp->b_kvabase = NULL;
846 * Remove the buffer from the appropriate free list.
849 _bremfree(struct buf *bp)
851 if (bp->b_qindex != BQUEUE_NONE) {
852 KASSERT(BUF_REFCNTNB(bp) == 1,
853 ("bremfree: bp %p not locked",bp));
854 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
855 bp->b_qindex = BQUEUE_NONE;
857 if (BUF_REFCNTNB(bp) <= 1)
858 panic("bremfree: removing a buffer not on a queue");
863 bremfree(struct buf *bp)
865 spin_lock(&bufqspin);
867 spin_unlock(&bufqspin);
871 bremfree_locked(struct buf *bp)
879 * Get a buffer with the specified data. Look in the cache first. We
880 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
881 * is set, the buffer is valid and we do not have to do anything ( see
886 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
888 return (breadn(vp, loffset, size, NULL, NULL, 0, bpp));
892 * This version of bread issues any required I/O asyncnronously and
893 * makes a callback on completion.
895 * The callback must check whether BIO_DONE is set in the bio and issue
896 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing
897 * BIO_DONE and disposing of the I/O (bqrelse()ing it).
900 breadcb(struct vnode *vp, off_t loffset, int size,
901 void (*func)(struct bio *), void *arg)
905 bp = getblk(vp, loffset, size, 0, 0);
907 /* if not found in cache, do some I/O */
908 if ((bp->b_flags & B_CACHE) == 0) {
909 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
910 bp->b_cmd = BUF_CMD_READ;
911 bp->b_bio1.bio_done = func;
912 bp->b_bio1.bio_caller_info1.ptr = arg;
913 vfs_busy_pages(vp, bp);
915 vn_strategy(vp, &bp->b_bio1);
918 * Since we are issuing the callback synchronously it cannot
919 * race the BIO_DONE, so no need for atomic ops here.
921 /*bp->b_bio1.bio_done = func;*/
922 bp->b_bio1.bio_caller_info1.ptr = arg;
923 bp->b_bio1.bio_flags |= BIO_DONE;
933 * Operates like bread, but also starts asynchronous I/O on
934 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
935 * to initiating I/O . If B_CACHE is set, the buffer is valid
936 * and we do not have to do anything.
940 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
941 int *rabsize, int cnt, struct buf **bpp)
943 struct buf *bp, *rabp;
945 int rv = 0, readwait = 0;
947 *bpp = bp = getblk(vp, loffset, size, 0, 0);
949 /* if not found in cache, do some I/O */
950 if ((bp->b_flags & B_CACHE) == 0) {
951 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
952 bp->b_cmd = BUF_CMD_READ;
953 bp->b_bio1.bio_done = biodone_sync;
954 bp->b_bio1.bio_flags |= BIO_SYNC;
955 vfs_busy_pages(vp, bp);
956 vn_strategy(vp, &bp->b_bio1);
960 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
961 if (inmem(vp, *raoffset))
963 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
965 if ((rabp->b_flags & B_CACHE) == 0) {
966 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
967 rabp->b_cmd = BUF_CMD_READ;
968 vfs_busy_pages(vp, rabp);
970 vn_strategy(vp, &rabp->b_bio1);
976 rv = biowait(&bp->b_bio1, "biord");
983 * Synchronous write, waits for completion.
985 * Write, release buffer on completion. (Done by iodone
986 * if async). Do not bother writing anything if the buffer
989 * Note that we set B_CACHE here, indicating that buffer is
990 * fully valid and thus cacheable. This is true even of NFS
991 * now so we set it generally. This could be set either here
992 * or in biodone() since the I/O is synchronous. We put it
996 bwrite(struct buf *bp)
1000 if (bp->b_flags & B_INVAL) {
1004 if (BUF_REFCNTNB(bp) == 0)
1005 panic("bwrite: buffer is not busy???");
1007 /* Mark the buffer clean */
1010 bp->b_flags &= ~(B_ERROR | B_EINTR);
1011 bp->b_flags |= B_CACHE;
1012 bp->b_cmd = BUF_CMD_WRITE;
1013 bp->b_bio1.bio_done = biodone_sync;
1014 bp->b_bio1.bio_flags |= BIO_SYNC;
1015 vfs_busy_pages(bp->b_vp, bp);
1018 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1019 * valid for vnode-backed buffers.
1021 bsetrunningbufspace(bp, bp->b_bufsize);
1022 vn_strategy(bp->b_vp, &bp->b_bio1);
1023 error = biowait(&bp->b_bio1, "biows");
1032 * Asynchronous write. Start output on a buffer, but do not wait for
1033 * it to complete. The buffer is released when the output completes.
1035 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1036 * B_INVAL buffers. Not us.
1039 bawrite(struct buf *bp)
1041 if (bp->b_flags & B_INVAL) {
1045 if (BUF_REFCNTNB(bp) == 0)
1046 panic("bwrite: buffer is not busy???");
1048 /* Mark the buffer clean */
1051 bp->b_flags &= ~(B_ERROR | B_EINTR);
1052 bp->b_flags |= B_CACHE;
1053 bp->b_cmd = BUF_CMD_WRITE;
1054 KKASSERT(bp->b_bio1.bio_done == NULL);
1055 vfs_busy_pages(bp->b_vp, bp);
1058 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1059 * valid for vnode-backed buffers.
1061 bsetrunningbufspace(bp, bp->b_bufsize);
1063 vn_strategy(bp->b_vp, &bp->b_bio1);
1069 * Ordered write. Start output on a buffer, and flag it so that the
1070 * device will write it in the order it was queued. The buffer is
1071 * released when the output completes. bwrite() ( or the VOP routine
1072 * anyway ) is responsible for handling B_INVAL buffers.
1075 bowrite(struct buf *bp)
1077 bp->b_flags |= B_ORDERED;
1085 * Delayed write. (Buffer is marked dirty). Do not bother writing
1086 * anything if the buffer is marked invalid.
1088 * Note that since the buffer must be completely valid, we can safely
1089 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1090 * biodone() in order to prevent getblk from writing the buffer
1091 * out synchronously.
1094 bdwrite(struct buf *bp)
1096 if (BUF_REFCNTNB(bp) == 0)
1097 panic("bdwrite: buffer is not busy");
1099 if (bp->b_flags & B_INVAL) {
1105 if (dsched_is_clear_buf_priv(bp))
1109 * Set B_CACHE, indicating that the buffer is fully valid. This is
1110 * true even of NFS now.
1112 bp->b_flags |= B_CACHE;
1115 * This bmap keeps the system from needing to do the bmap later,
1116 * perhaps when the system is attempting to do a sync. Since it
1117 * is likely that the indirect block -- or whatever other datastructure
1118 * that the filesystem needs is still in memory now, it is a good
1119 * thing to do this. Note also, that if the pageout daemon is
1120 * requesting a sync -- there might not be enough memory to do
1121 * the bmap then... So, this is important to do.
1123 if (bp->b_bio2.bio_offset == NOOFFSET) {
1124 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1125 NULL, NULL, BUF_CMD_WRITE);
1129 * Because the underlying pages may still be mapped and
1130 * writable trying to set the dirty buffer (b_dirtyoff/end)
1131 * range here will be inaccurate.
1133 * However, we must still clean the pages to satisfy the
1134 * vnode_pager and pageout daemon, so theythink the pages
1135 * have been "cleaned". What has really occured is that
1136 * they've been earmarked for later writing by the buffer
1139 * So we get the b_dirtyoff/end update but will not actually
1140 * depend on it (NFS that is) until the pages are busied for
1143 vfs_clean_pages(bp);
1147 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1148 * due to the softdep code.
1153 * Fake write - return pages to VM system as dirty, leave the buffer clean.
1154 * This is used by tmpfs.
1156 * It is important for any VFS using this routine to NOT use it for
1157 * IO_SYNC or IO_ASYNC operations which occur when the system really
1158 * wants to flush VM pages to backing store.
1161 buwrite(struct buf *bp)
1167 * Only works for VMIO buffers. If the buffer is already
1168 * marked for delayed-write we can't avoid the bdwrite().
1170 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) {
1176 * Set valid & dirty.
1178 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1179 m = bp->b_xio.xio_pages[i];
1180 vfs_dirty_one_page(bp, i, m);
1188 * Turn buffer into delayed write request by marking it B_DELWRI.
1189 * B_RELBUF and B_NOCACHE must be cleared.
1191 * We reassign the buffer to itself to properly update it in the
1192 * dirty/clean lists.
1194 * Must be called from a critical section.
1195 * The buffer must be on BQUEUE_NONE.
1198 bdirty(struct buf *bp)
1200 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1201 if (bp->b_flags & B_NOCACHE) {
1202 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1203 bp->b_flags &= ~B_NOCACHE;
1205 if (bp->b_flags & B_INVAL) {
1206 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1208 bp->b_flags &= ~B_RELBUF;
1210 if ((bp->b_flags & B_DELWRI) == 0) {
1211 lwkt_gettoken(&bp->b_vp->v_token);
1212 bp->b_flags |= B_DELWRI;
1214 lwkt_reltoken(&bp->b_vp->v_token);
1216 spin_lock(&bufcspin);
1218 dirtybufspace += bp->b_bufsize;
1219 if (bp->b_flags & B_HEAVY) {
1221 dirtybufspacehw += bp->b_bufsize;
1223 spin_unlock(&bufcspin);
1230 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1231 * needs to be flushed with a different buf_daemon thread to avoid
1232 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1235 bheavy(struct buf *bp)
1237 if ((bp->b_flags & B_HEAVY) == 0) {
1238 bp->b_flags |= B_HEAVY;
1239 if (bp->b_flags & B_DELWRI) {
1240 spin_lock(&bufcspin);
1242 dirtybufspacehw += bp->b_bufsize;
1243 spin_unlock(&bufcspin);
1251 * Clear B_DELWRI for buffer.
1253 * Must be called from a critical section.
1255 * The buffer is typically on BQUEUE_NONE but there is one case in
1256 * brelse() that calls this function after placing the buffer on
1257 * a different queue.
1262 bundirty(struct buf *bp)
1264 if (bp->b_flags & B_DELWRI) {
1265 lwkt_gettoken(&bp->b_vp->v_token);
1266 bp->b_flags &= ~B_DELWRI;
1268 lwkt_reltoken(&bp->b_vp->v_token);
1270 spin_lock(&bufcspin);
1272 dirtybufspace -= bp->b_bufsize;
1273 if (bp->b_flags & B_HEAVY) {
1275 dirtybufspacehw -= bp->b_bufsize;
1277 spin_unlock(&bufcspin);
1279 bd_signal(bp->b_bufsize);
1282 * Since it is now being written, we can clear its deferred write flag.
1284 bp->b_flags &= ~B_DEFERRED;
1288 * Set the b_runningbufspace field, used to track how much I/O is
1289 * in progress at any given moment.
1292 bsetrunningbufspace(struct buf *bp, int bytes)
1294 bp->b_runningbufspace = bytes;
1296 spin_lock(&bufcspin);
1297 runningbufspace += bytes;
1299 spin_unlock(&bufcspin);
1306 * Release a busy buffer and, if requested, free its resources. The
1307 * buffer will be stashed in the appropriate bufqueue[] allowing it
1308 * to be accessed later as a cache entity or reused for other purposes.
1313 brelse(struct buf *bp)
1316 int saved_flags = bp->b_flags;
1319 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1322 * If B_NOCACHE is set we are being asked to destroy the buffer and
1323 * its backing store. Clear B_DELWRI.
1325 * B_NOCACHE is set in two cases: (1) when the caller really wants
1326 * to destroy the buffer and backing store and (2) when the caller
1327 * wants to destroy the buffer and backing store after a write
1330 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1334 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1336 * A re-dirtied buffer is only subject to destruction
1337 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1339 /* leave buffer intact */
1340 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1341 (bp->b_bufsize <= 0)) {
1343 * Either a failed read or we were asked to free or not
1344 * cache the buffer. This path is reached with B_DELWRI
1345 * set only if B_INVAL is already set. B_NOCACHE governs
1346 * backing store destruction.
1348 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1349 * buffer cannot be immediately freed.
1351 bp->b_flags |= B_INVAL;
1352 if (LIST_FIRST(&bp->b_dep) != NULL)
1354 if (bp->b_flags & B_DELWRI) {
1355 spin_lock(&bufcspin);
1357 dirtybufspace -= bp->b_bufsize;
1358 if (bp->b_flags & B_HEAVY) {
1360 dirtybufspacehw -= bp->b_bufsize;
1362 spin_unlock(&bufcspin);
1364 bd_signal(bp->b_bufsize);
1366 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1370 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set,
1371 * or if b_refs is non-zero.
1373 * If vfs_vmio_release() is called with either bit set, the
1374 * underlying pages may wind up getting freed causing a previous
1375 * write (bdwrite()) to get 'lost' because pages associated with
1376 * a B_DELWRI bp are marked clean. Pages associated with a
1377 * B_LOCKED buffer may be mapped by the filesystem.
1379 * If we want to release the buffer ourselves (rather then the
1380 * originator asking us to release it), give the originator a
1381 * chance to countermand the release by setting B_LOCKED.
1383 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1384 * if B_DELWRI is set.
1386 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1387 * on pages to return pages to the VM page queues.
1389 if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) {
1390 bp->b_flags &= ~B_RELBUF;
1391 } else if (vm_page_count_severe()) {
1392 if (LIST_FIRST(&bp->b_dep) != NULL)
1393 buf_deallocate(bp); /* can set B_LOCKED */
1394 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1395 bp->b_flags &= ~B_RELBUF;
1397 bp->b_flags |= B_RELBUF;
1401 * Make sure b_cmd is clear. It may have already been cleared by
1404 * At this point destroying the buffer is governed by the B_INVAL
1405 * or B_RELBUF flags.
1407 bp->b_cmd = BUF_CMD_DONE;
1408 dsched_exit_buf(bp);
1411 * VMIO buffer rundown. Make sure the VM page array is restored
1412 * after an I/O may have replaces some of the pages with bogus pages
1413 * in order to not destroy dirty pages in a fill-in read.
1415 * Note that due to the code above, if a buffer is marked B_DELWRI
1416 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1417 * B_INVAL may still be set, however.
1419 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1420 * but not the backing store. B_NOCACHE will destroy the backing
1423 * Note that dirty NFS buffers contain byte-granular write ranges
1424 * and should not be destroyed w/ B_INVAL even if the backing store
1427 if (bp->b_flags & B_VMIO) {
1429 * Rundown for VMIO buffers which are not dirty NFS buffers.
1441 * Get the base offset and length of the buffer. Note that
1442 * in the VMIO case if the buffer block size is not
1443 * page-aligned then b_data pointer may not be page-aligned.
1444 * But our b_xio.xio_pages array *IS* page aligned.
1446 * block sizes less then DEV_BSIZE (usually 512) are not
1447 * supported due to the page granularity bits (m->valid,
1448 * m->dirty, etc...).
1450 * See man buf(9) for more information
1453 resid = bp->b_bufsize;
1454 foff = bp->b_loffset;
1456 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1457 m = bp->b_xio.xio_pages[i];
1458 vm_page_flag_clear(m, PG_ZERO);
1460 * If we hit a bogus page, fixup *all* of them
1461 * now. Note that we left these pages wired
1462 * when we removed them so they had better exist,
1463 * and they cannot be ripped out from under us so
1464 * no critical section protection is necessary.
1466 if (m == bogus_page) {
1468 poff = OFF_TO_IDX(bp->b_loffset);
1470 vm_object_hold(obj);
1471 for (j = i; j < bp->b_xio.xio_npages; j++) {
1474 mtmp = bp->b_xio.xio_pages[j];
1475 if (mtmp == bogus_page) {
1476 mtmp = vm_page_lookup(obj, poff + j);
1478 panic("brelse: page missing");
1480 bp->b_xio.xio_pages[j] = mtmp;
1483 bp->b_flags &= ~B_HASBOGUS;
1484 vm_object_drop(obj);
1486 if ((bp->b_flags & B_INVAL) == 0) {
1487 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1488 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1490 m = bp->b_xio.xio_pages[i];
1494 * Invalidate the backing store if B_NOCACHE is set
1495 * (e.g. used with vinvalbuf()). If this is NFS
1496 * we impose a requirement that the block size be
1497 * a multiple of PAGE_SIZE and create a temporary
1498 * hack to basically invalidate the whole page. The
1499 * problem is that NFS uses really odd buffer sizes
1500 * especially when tracking piecemeal writes and
1501 * it also vinvalbuf()'s a lot, which would result
1502 * in only partial page validation and invalidation
1503 * here. If the file page is mmap()'d, however,
1504 * all the valid bits get set so after we invalidate
1505 * here we would end up with weird m->valid values
1506 * like 0xfc. nfs_getpages() can't handle this so
1507 * we clear all the valid bits for the NFS case
1508 * instead of just some of them.
1510 * The real bug is the VM system having to set m->valid
1511 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1512 * itself is an artifact of the whole 512-byte
1513 * granular mess that exists to support odd block
1514 * sizes and UFS meta-data block sizes (e.g. 6144).
1515 * A complete rewrite is required.
1519 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1520 int poffset = foff & PAGE_MASK;
1523 presid = PAGE_SIZE - poffset;
1524 if (bp->b_vp->v_tag == VT_NFS &&
1525 bp->b_vp->v_type == VREG) {
1527 } else if (presid > resid) {
1530 KASSERT(presid >= 0, ("brelse: extra page"));
1531 vm_page_set_invalid(m, poffset, presid);
1534 * Also make sure any swap cache is removed
1535 * as it is now stale (HAMMER in particular
1536 * uses B_NOCACHE to deal with buffer
1539 swap_pager_unswapped(m);
1541 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1542 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1544 if (bp->b_flags & (B_INVAL | B_RELBUF))
1545 vfs_vmio_release(bp);
1548 * Rundown for non-VMIO buffers.
1550 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1553 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1559 if (bp->b_qindex != BQUEUE_NONE)
1560 panic("brelse: free buffer onto another queue???");
1561 if (BUF_REFCNTNB(bp) > 1) {
1562 /* Temporary panic to verify exclusive locking */
1563 /* This panic goes away when we allow shared refs */
1564 panic("brelse: multiple refs");
1570 * Figure out the correct queue to place the cleaned up buffer on.
1571 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1572 * disassociated from their vnode.
1574 spin_lock(&bufqspin);
1575 if (bp->b_flags & B_LOCKED) {
1577 * Buffers that are locked are placed in the locked queue
1578 * immediately, regardless of their state.
1580 bp->b_qindex = BQUEUE_LOCKED;
1581 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1582 } else if (bp->b_bufsize == 0) {
1584 * Buffers with no memory. Due to conditionals near the top
1585 * of brelse() such buffers should probably already be
1586 * marked B_INVAL and disassociated from their vnode.
1588 bp->b_flags |= B_INVAL;
1589 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1590 KKASSERT((bp->b_flags & B_HASHED) == 0);
1591 if (bp->b_kvasize) {
1592 bp->b_qindex = BQUEUE_EMPTYKVA;
1594 bp->b_qindex = BQUEUE_EMPTY;
1596 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1597 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1599 * Buffers with junk contents. Again these buffers had better
1600 * already be disassociated from their vnode.
1602 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1603 KKASSERT((bp->b_flags & B_HASHED) == 0);
1604 bp->b_flags |= B_INVAL;
1605 bp->b_qindex = BQUEUE_CLEAN;
1606 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1609 * Remaining buffers. These buffers are still associated with
1612 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1614 bp->b_qindex = BQUEUE_DIRTY;
1615 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1617 case B_DELWRI | B_HEAVY:
1618 bp->b_qindex = BQUEUE_DIRTY_HW;
1619 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1624 * NOTE: Buffers are always placed at the end of the
1625 * queue. If B_AGE is not set the buffer will cycle
1626 * through the queue twice.
1628 bp->b_qindex = BQUEUE_CLEAN;
1629 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1633 spin_unlock(&bufqspin);
1636 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1637 * on the correct queue.
1639 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1643 * The bp is on an appropriate queue unless locked. If it is not
1644 * locked or dirty we can wakeup threads waiting for buffer space.
1646 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1647 * if B_INVAL is set ).
1649 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1653 * Something we can maybe free or reuse
1655 if (bp->b_bufsize || bp->b_kvasize)
1659 * Clean up temporary flags and unlock the buffer.
1661 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
1668 * Release a buffer back to the appropriate queue but do not try to free
1669 * it. The buffer is expected to be used again soon.
1671 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1672 * biodone() to requeue an async I/O on completion. It is also used when
1673 * known good buffers need to be requeued but we think we may need the data
1676 * XXX we should be able to leave the B_RELBUF hint set on completion.
1681 bqrelse(struct buf *bp)
1683 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1685 if (bp->b_qindex != BQUEUE_NONE)
1686 panic("bqrelse: free buffer onto another queue???");
1687 if (BUF_REFCNTNB(bp) > 1) {
1688 /* do not release to free list */
1689 panic("bqrelse: multiple refs");
1693 buf_act_advance(bp);
1695 spin_lock(&bufqspin);
1696 if (bp->b_flags & B_LOCKED) {
1698 * Locked buffers are released to the locked queue. However,
1699 * if the buffer is dirty it will first go into the dirty
1700 * queue and later on after the I/O completes successfully it
1701 * will be released to the locked queue.
1703 bp->b_qindex = BQUEUE_LOCKED;
1704 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1705 } else if (bp->b_flags & B_DELWRI) {
1706 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1707 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1708 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1709 } else if (vm_page_count_severe()) {
1711 * We are too low on memory, we have to try to free the
1712 * buffer (most importantly: the wired pages making up its
1713 * backing store) *now*.
1715 spin_unlock(&bufqspin);
1719 bp->b_qindex = BQUEUE_CLEAN;
1720 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1722 spin_unlock(&bufqspin);
1724 if ((bp->b_flags & B_LOCKED) == 0 &&
1725 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1730 * Something we can maybe free or reuse.
1732 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1736 * Final cleanup and unlock. Clear bits that are only used while a
1737 * buffer is actively locked.
1739 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
1740 dsched_exit_buf(bp);
1745 * Hold a buffer, preventing it from being reused. This will prevent
1746 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL
1747 * operations. If a B_INVAL operation occurs the buffer will remain held
1748 * but the underlying pages may get ripped out.
1750 * These functions are typically used in VOP_READ/VOP_WRITE functions
1751 * to hold a buffer during a copyin or copyout, preventing deadlocks
1752 * or recursive lock panics when read()/write() is used over mmap()'d
1755 * NOTE: bqhold() requires that the buffer be locked at the time of the
1756 * hold. bqdrop() has no requirements other than the buffer having
1757 * previously been held.
1760 bqhold(struct buf *bp)
1762 atomic_add_int(&bp->b_refs, 1);
1766 bqdrop(struct buf *bp)
1768 KKASSERT(bp->b_refs > 0);
1769 atomic_add_int(&bp->b_refs, -1);
1775 * Return backing pages held by the buffer 'bp' back to the VM system
1776 * if possible. The pages are freed if they are no longer valid or
1777 * attempt to free if it was used for direct I/O otherwise they are
1778 * sent to the page cache.
1780 * Pages that were marked busy are left alone and skipped.
1782 * The KVA mapping (b_data) for the underlying pages is removed by
1786 vfs_vmio_release(struct buf *bp)
1791 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1792 m = bp->b_xio.xio_pages[i];
1793 bp->b_xio.xio_pages[i] = NULL;
1795 vm_page_busy_wait(m, FALSE, "vmiopg");
1798 * The VFS is telling us this is not a meta-data buffer
1799 * even if it is backed by a block device.
1801 if (bp->b_flags & B_NOTMETA)
1802 vm_page_flag_set(m, PG_NOTMETA);
1805 * This is a very important bit of code. We try to track
1806 * VM page use whether the pages are wired into the buffer
1807 * cache or not. While wired into the buffer cache the
1808 * bp tracks the act_count.
1810 * We can choose to place unwired pages on the inactive
1811 * queue (0) or active queue (1). If we place too many
1812 * on the active queue the queue will cycle the act_count
1813 * on pages we'd like to keep, just from single-use pages
1814 * (such as when doing a tar-up or file scan).
1816 if (bp->b_act_count < vm_cycle_point)
1817 vm_page_unwire(m, 0);
1819 vm_page_unwire(m, 1);
1822 * We don't mess with busy pages, it is the responsibility
1823 * of the process that busied the pages to deal with them.
1825 * However, the caller may have marked the page invalid and
1826 * we must still make sure the page is no longer mapped.
1828 if ((m->flags & PG_BUSY) || (m->busy != 0)) {
1829 vm_page_protect(m, VM_PROT_NONE);
1834 if (m->wire_count == 0) {
1835 vm_page_flag_clear(m, PG_ZERO);
1837 * Might as well free the page if we can and it has
1838 * no valid data. We also free the page if the
1839 * buffer was used for direct I/O.
1842 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1843 m->hold_count == 0) {
1844 vm_page_protect(m, VM_PROT_NONE);
1849 * Cache the page if we are really low on free
1852 * Also bypass the active and inactive queues
1853 * if B_NOTMETA is set. This flag is set by HAMMER
1854 * on a regular file buffer when double buffering
1855 * is enabled or on a block device buffer representing
1856 * file data when double buffering is not enabled.
1857 * The flag prevents two copies of the same data from
1858 * being cached for long periods of time.
1860 if (bp->b_flags & B_DIRECT) {
1862 vm_page_try_to_free(m);
1863 } else if ((bp->b_flags & B_NOTMETA) ||
1864 vm_page_count_severe()) {
1865 m->act_count = bp->b_act_count;
1867 vm_page_try_to_cache(m);
1869 m->act_count = bp->b_act_count;
1877 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
1878 bp->b_xio.xio_npages);
1879 if (bp->b_bufsize) {
1883 bp->b_xio.xio_npages = 0;
1884 bp->b_flags &= ~B_VMIO;
1885 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1893 * Implement clustered async writes for clearing out B_DELWRI buffers.
1894 * This is much better then the old way of writing only one buffer at
1895 * a time. Note that we may not be presented with the buffers in the
1896 * correct order, so we search for the cluster in both directions.
1898 * The buffer is locked on call.
1901 vfs_bio_awrite(struct buf *bp)
1905 off_t loffset = bp->b_loffset;
1906 struct vnode *vp = bp->b_vp;
1913 * right now we support clustered writing only to regular files. If
1914 * we find a clusterable block we could be in the middle of a cluster
1915 * rather then at the beginning.
1917 * NOTE: b_bio1 contains the logical loffset and is aliased
1918 * to b_loffset. b_bio2 contains the translated block number.
1920 if ((vp->v_type == VREG) &&
1921 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1922 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1924 size = vp->v_mount->mnt_stat.f_iosize;
1926 for (i = size; i < MAXPHYS; i += size) {
1927 if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) &&
1928 BUF_REFCNT(bpa) == 0 &&
1929 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1930 (B_DELWRI | B_CLUSTEROK)) &&
1931 (bpa->b_bufsize == size)) {
1932 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1933 (bpa->b_bio2.bio_offset !=
1934 bp->b_bio2.bio_offset + i))
1940 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1941 if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) &&
1942 BUF_REFCNT(bpa) == 0 &&
1943 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1944 (B_DELWRI | B_CLUSTEROK)) &&
1945 (bpa->b_bufsize == size)) {
1946 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1947 (bpa->b_bio2.bio_offset !=
1948 bp->b_bio2.bio_offset - j))
1958 * this is a possible cluster write
1960 if (nbytes != size) {
1962 nwritten = cluster_wbuild(vp, size,
1963 loffset - j, nbytes);
1969 * default (old) behavior, writing out only one block
1971 * XXX returns b_bufsize instead of b_bcount for nwritten?
1973 nwritten = bp->b_bufsize;
1983 * Find and initialize a new buffer header, freeing up existing buffers
1984 * in the bufqueues as necessary. The new buffer is returned locked.
1986 * Important: B_INVAL is not set. If the caller wishes to throw the
1987 * buffer away, the caller must set B_INVAL prior to calling brelse().
1990 * We have insufficient buffer headers
1991 * We have insufficient buffer space
1992 * buffer_map is too fragmented ( space reservation fails )
1993 * If we have to flush dirty buffers ( but we try to avoid this )
1995 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1996 * Instead we ask the buf daemon to do it for us. We attempt to
1997 * avoid piecemeal wakeups of the pageout daemon.
2002 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
2008 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2009 static int flushingbufs;
2012 * We can't afford to block since we might be holding a vnode lock,
2013 * which may prevent system daemons from running. We deal with
2014 * low-memory situations by proactively returning memory and running
2015 * async I/O rather then sync I/O.
2019 --getnewbufrestarts;
2021 ++getnewbufrestarts;
2024 * Setup for scan. If we do not have enough free buffers,
2025 * we setup a degenerate case that immediately fails. Note
2026 * that if we are specially marked process, we are allowed to
2027 * dip into our reserves.
2029 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
2031 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
2032 * However, there are a number of cases (defragging, reusing, ...)
2033 * where we cannot backup.
2035 nqindex = BQUEUE_EMPTYKVA;
2036 spin_lock(&bufqspin);
2037 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
2041 * If no EMPTYKVA buffers and we are either
2042 * defragging or reusing, locate a CLEAN buffer
2043 * to free or reuse. If bufspace useage is low
2044 * skip this step so we can allocate a new buffer.
2046 if (defrag || bufspace >= lobufspace) {
2047 nqindex = BQUEUE_CLEAN;
2048 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2052 * If we could not find or were not allowed to reuse a
2053 * CLEAN buffer, check to see if it is ok to use an EMPTY
2054 * buffer. We can only use an EMPTY buffer if allocating
2055 * its KVA would not otherwise run us out of buffer space.
2057 if (nbp == NULL && defrag == 0 &&
2058 bufspace + maxsize < hibufspace) {
2059 nqindex = BQUEUE_EMPTY;
2060 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
2065 * Run scan, possibly freeing data and/or kva mappings on the fly
2068 * WARNING! bufqspin is held!
2070 while ((bp = nbp) != NULL) {
2071 int qindex = nqindex;
2073 nbp = TAILQ_NEXT(bp, b_freelist);
2076 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2077 * cycles through the queue twice before being selected.
2079 if (qindex == BQUEUE_CLEAN &&
2080 (bp->b_flags & B_AGE) == 0 && nbp) {
2081 bp->b_flags |= B_AGE;
2082 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
2083 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
2088 * Calculate next bp ( we can only use it if we do not block
2089 * or do other fancy things ).
2094 nqindex = BQUEUE_EMPTYKVA;
2095 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
2098 case BQUEUE_EMPTYKVA:
2099 nqindex = BQUEUE_CLEAN;
2100 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
2114 KASSERT(bp->b_qindex == qindex,
2115 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2118 * Note: we no longer distinguish between VMIO and non-VMIO
2121 KASSERT((bp->b_flags & B_DELWRI) == 0,
2122 ("delwri buffer %p found in queue %d", bp, qindex));
2125 * Do not try to reuse a buffer with a non-zero b_refs.
2126 * This is an unsynchronized test. A synchronized test
2127 * is also performed after we lock the buffer.
2133 * If we are defragging then we need a buffer with
2134 * b_kvasize != 0. XXX this situation should no longer
2135 * occur, if defrag is non-zero the buffer's b_kvasize
2136 * should also be non-zero at this point. XXX
2138 if (defrag && bp->b_kvasize == 0) {
2139 kprintf("Warning: defrag empty buffer %p\n", bp);
2144 * Start freeing the bp. This is somewhat involved. nbp
2145 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
2146 * on the clean list must be disassociated from their
2147 * current vnode. Buffers on the empty[kva] lists have
2148 * already been disassociated.
2150 * b_refs is checked after locking along with queue changes.
2151 * We must check here to deal with zero->nonzero transitions
2152 * made by the owner of the buffer lock, which is used by
2153 * VFS's to hold the buffer while issuing an unlocked
2154 * uiomove()s. We cannot invalidate the buffer's pages
2155 * for this case. Once we successfully lock a buffer the
2156 * only 0->1 transitions of b_refs will occur via findblk().
2158 * We must also check for queue changes after successful
2159 * locking as the current lock holder may dispose of the
2160 * buffer and change its queue.
2162 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2163 spin_unlock(&bufqspin);
2164 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100);
2167 if (bp->b_qindex != qindex || bp->b_refs) {
2168 spin_unlock(&bufqspin);
2172 bremfree_locked(bp);
2173 spin_unlock(&bufqspin);
2176 * Dependancies must be handled before we disassociate the
2179 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2180 * be immediately disassociated. HAMMER then becomes
2181 * responsible for releasing the buffer.
2183 * NOTE: bufqspin is UNLOCKED now.
2185 if (LIST_FIRST(&bp->b_dep) != NULL) {
2187 if (bp->b_flags & B_LOCKED) {
2191 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2194 if (qindex == BQUEUE_CLEAN) {
2195 if (bp->b_flags & B_VMIO)
2196 vfs_vmio_release(bp);
2202 * NOTE: nbp is now entirely invalid. We can only restart
2203 * the scan from this point on.
2205 * Get the rest of the buffer freed up. b_kva* is still
2206 * valid after this operation.
2208 KASSERT(bp->b_vp == NULL,
2209 ("bp3 %p flags %08x vnode %p qindex %d "
2210 "unexpectededly still associated!",
2211 bp, bp->b_flags, bp->b_vp, qindex));
2212 KKASSERT((bp->b_flags & B_HASHED) == 0);
2215 * critical section protection is not required when
2216 * scrapping a buffer's contents because it is already
2222 bp->b_flags = B_BNOCLIP;
2223 bp->b_cmd = BUF_CMD_DONE;
2228 bp->b_xio.xio_npages = 0;
2229 bp->b_dirtyoff = bp->b_dirtyend = 0;
2230 bp->b_act_count = ACT_INIT;
2232 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2234 if (blkflags & GETBLK_BHEAVY)
2235 bp->b_flags |= B_HEAVY;
2238 * If we are defragging then free the buffer.
2241 bp->b_flags |= B_INVAL;
2249 * If we are overcomitted then recover the buffer and its
2250 * KVM space. This occurs in rare situations when multiple
2251 * processes are blocked in getnewbuf() or allocbuf().
2253 if (bufspace >= hibufspace)
2255 if (flushingbufs && bp->b_kvasize != 0) {
2256 bp->b_flags |= B_INVAL;
2261 if (bufspace < lobufspace)
2265 * b_refs can transition to a non-zero value while we hold
2266 * the buffer locked due to a findblk(). Our brelvp() above
2267 * interlocked any future possible transitions due to
2270 * If we find b_refs to be non-zero we can destroy the
2271 * buffer's contents but we cannot yet reuse the buffer.
2274 bp->b_flags |= B_INVAL;
2280 /* NOT REACHED, bufqspin not held */
2284 * If we exhausted our list, sleep as appropriate. We may have to
2285 * wakeup various daemons and write out some dirty buffers.
2287 * Generally we are sleeping due to insufficient buffer space.
2289 * NOTE: bufqspin is held if bp is NULL, else it is not held.
2295 spin_unlock(&bufqspin);
2297 flags = VFS_BIO_NEED_BUFSPACE;
2299 } else if (bufspace >= hibufspace) {
2301 flags = VFS_BIO_NEED_BUFSPACE;
2304 flags = VFS_BIO_NEED_ANY;
2307 bd_speedup(); /* heeeelp */
2308 spin_lock(&bufcspin);
2309 needsbuffer |= flags;
2310 while (needsbuffer & flags) {
2311 if (ssleep(&needsbuffer, &bufcspin,
2312 slpflags, waitmsg, slptimeo)) {
2313 spin_unlock(&bufcspin);
2317 spin_unlock(&bufcspin);
2320 * We finally have a valid bp. We aren't quite out of the
2321 * woods, we still have to reserve kva space. In order
2322 * to keep fragmentation sane we only allocate kva in
2325 * (bufqspin is not held)
2327 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2329 if (maxsize != bp->b_kvasize) {
2330 vm_offset_t addr = 0;
2335 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2336 vm_map_lock(&buffer_map);
2338 if (vm_map_findspace(&buffer_map,
2339 vm_map_min(&buffer_map), maxsize,
2340 maxsize, 0, &addr)) {
2342 * Uh oh. Buffer map is too fragmented. We
2343 * must defragment the map.
2345 vm_map_unlock(&buffer_map);
2346 vm_map_entry_release(count);
2349 bp->b_flags |= B_INVAL;
2354 vm_map_insert(&buffer_map, &count,
2356 addr, addr + maxsize,
2358 VM_PROT_ALL, VM_PROT_ALL,
2361 bp->b_kvabase = (caddr_t) addr;
2362 bp->b_kvasize = maxsize;
2363 bufspace += bp->b_kvasize;
2366 vm_map_unlock(&buffer_map);
2367 vm_map_entry_release(count);
2369 bp->b_data = bp->b_kvabase;
2375 * This routine is called in an emergency to recover VM pages from the
2376 * buffer cache by cashing in clean buffers. The idea is to recover
2377 * enough pages to be able to satisfy a stuck bio_page_alloc().
2382 recoverbufpages(void)
2389 spin_lock(&bufqspin);
2390 while (bytes < MAXBSIZE) {
2391 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2396 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2397 * cycles through the queue twice before being selected.
2399 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
2400 bp->b_flags |= B_AGE;
2401 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
2402 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
2410 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
2411 KKASSERT((bp->b_flags & B_DELWRI) == 0);
2414 * Start freeing the bp. This is somewhat involved.
2416 * Buffers on the clean list must be disassociated from
2417 * their current vnode
2420 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2421 kprintf("recoverbufpages: warning, locked buf %p, "
2424 ssleep(&bd_request, &bufqspin, 0, "gnbxxx", hz / 100);
2427 if (bp->b_qindex != BQUEUE_CLEAN) {
2428 kprintf("recoverbufpages: warning, BUF_LOCK blocked "
2429 "unexpectedly on buf %p index %d, race "
2435 bremfree_locked(bp);
2436 spin_unlock(&bufqspin);
2439 * Dependancies must be handled before we disassociate the
2442 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2443 * be immediately disassociated. HAMMER then becomes
2444 * responsible for releasing the buffer.
2446 if (LIST_FIRST(&bp->b_dep) != NULL) {
2448 if (bp->b_flags & B_LOCKED) {
2450 spin_lock(&bufqspin);
2453 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2456 bytes += bp->b_bufsize;
2458 if (bp->b_flags & B_VMIO) {
2459 bp->b_flags |= B_DIRECT; /* try to free pages */
2460 vfs_vmio_release(bp);
2465 KKASSERT(bp->b_vp == NULL);
2466 KKASSERT((bp->b_flags & B_HASHED) == 0);
2469 * critical section protection is not required when
2470 * scrapping a buffer's contents because it is already
2476 bp->b_flags = B_BNOCLIP;
2477 bp->b_cmd = BUF_CMD_DONE;
2482 bp->b_xio.xio_npages = 0;
2483 bp->b_dirtyoff = bp->b_dirtyend = 0;
2485 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2487 bp->b_flags |= B_INVAL;
2490 spin_lock(&bufqspin);
2492 spin_unlock(&bufqspin);
2499 * Buffer flushing daemon. Buffers are normally flushed by the
2500 * update daemon but if it cannot keep up this process starts to
2501 * take the load in an attempt to prevent getnewbuf() from blocking.
2503 * Once a flush is initiated it does not stop until the number
2504 * of buffers falls below lodirtybuffers, but we will wake up anyone
2505 * waiting at the mid-point.
2508 static struct kproc_desc buf_kp = {
2513 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2514 kproc_start, &buf_kp)
2516 static struct kproc_desc bufhw_kp = {
2521 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2522 kproc_start, &bufhw_kp)
2533 * This process needs to be suspended prior to shutdown sync.
2535 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2536 bufdaemon_td, SHUTDOWN_PRI_LAST);
2537 curthread->td_flags |= TDF_SYSTHREAD;
2540 * This process is allowed to take the buffer cache to the limit
2543 kproc_suspend_loop();
2546 * Do the flush as long as the number of dirty buffers
2547 * (including those running) exceeds lodirtybufspace.
2549 * When flushing limit running I/O to hirunningspace
2550 * Do the flush. Limit the amount of in-transit I/O we
2551 * allow to build up, otherwise we would completely saturate
2552 * the I/O system. Wakeup any waiting processes before we
2553 * normally would so they can run in parallel with our drain.
2555 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2556 * but because we split the operation into two threads we
2557 * have to cut it in half for each thread.
2559 waitrunningbufspace();
2560 limit = lodirtybufspace / 2;
2561 while (runningbufspace + dirtybufspace > limit ||
2562 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
2563 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2565 if (runningbufspace < hirunningspace)
2567 waitrunningbufspace();
2571 * We reached our low water mark, reset the
2572 * request and sleep until we are needed again.
2573 * The sleep is just so the suspend code works.
2575 spin_lock(&bufcspin);
2576 if (bd_request == 0)
2577 ssleep(&bd_request, &bufcspin, 0, "psleep", hz);
2579 spin_unlock(&bufcspin);
2592 * This process needs to be suspended prior to shutdown sync.
2594 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2595 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2596 curthread->td_flags |= TDF_SYSTHREAD;
2599 * This process is allowed to take the buffer cache to the limit
2602 kproc_suspend_loop();
2605 * Do the flush. Limit the amount of in-transit I/O we
2606 * allow to build up, otherwise we would completely saturate
2607 * the I/O system. Wakeup any waiting processes before we
2608 * normally would so they can run in parallel with our drain.
2610 * Once we decide to flush push the queued I/O up to
2611 * hirunningspace in order to trigger bursting by the bioq
2614 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2615 * but because we split the operation into two threads we
2616 * have to cut it in half for each thread.
2618 waitrunningbufspace();
2619 limit = lodirtybufspace / 2;
2620 while (runningbufspace + dirtybufspacehw > limit ||
2621 dirtybufcounthw >= nbuf / 2) {
2622 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2624 if (runningbufspace < hirunningspace)
2626 waitrunningbufspace();
2630 * We reached our low water mark, reset the
2631 * request and sleep until we are needed again.
2632 * The sleep is just so the suspend code works.
2634 spin_lock(&bufcspin);
2635 if (bd_request_hw == 0)
2636 ssleep(&bd_request_hw, &bufcspin, 0, "psleep", hz);
2638 spin_unlock(&bufcspin);
2645 * Try to flush a buffer in the dirty queue. We must be careful to
2646 * free up B_INVAL buffers instead of write them, which NFS is
2647 * particularly sensitive to.
2649 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2650 * that we really want to try to get the buffer out and reuse it
2651 * due to the write load on the machine.
2653 * We must lock the buffer in order to check its validity before we
2654 * can mess with its contents. bufqspin isn't enough.
2657 flushbufqueues(bufq_type_t q)
2663 spin_lock(&bufqspin);
2666 bp = TAILQ_FIRST(&bufqueues[q]);
2668 if ((bp->b_flags & B_DELWRI) == 0) {
2669 kprintf("Unexpected clean buffer %p\n", bp);
2670 bp = TAILQ_NEXT(bp, b_freelist);
2673 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2674 bp = TAILQ_NEXT(bp, b_freelist);
2677 KKASSERT(bp->b_qindex == q);
2680 * Must recheck B_DELWRI after successfully locking
2683 if ((bp->b_flags & B_DELWRI) == 0) {
2685 bp = TAILQ_NEXT(bp, b_freelist);
2689 if (bp->b_flags & B_INVAL) {
2691 spin_unlock(&bufqspin);
2698 spin_unlock(&bufqspin);
2701 if (LIST_FIRST(&bp->b_dep) != NULL &&
2702 (bp->b_flags & B_DEFERRED) == 0 &&
2703 buf_countdeps(bp, 0)) {
2704 spin_lock(&bufqspin);
2706 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2707 TAILQ_INSERT_TAIL(&bufqueues[q], bp, b_freelist);
2708 bp->b_flags |= B_DEFERRED;
2710 bp = TAILQ_FIRST(&bufqueues[q]);
2715 * If the buffer has a dependancy, buf_checkwrite() must
2716 * also return 0 for us to be able to initate the write.
2718 * If the buffer is flagged B_ERROR it may be requeued
2719 * over and over again, we try to avoid a live lock.
2721 * NOTE: buf_checkwrite is MPSAFE.
2723 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) {
2726 } else if (bp->b_flags & B_ERROR) {
2727 tsleep(bp, 0, "bioer", 1);
2728 bp->b_flags &= ~B_AGE;
2731 bp->b_flags |= B_AGE;
2738 spin_unlock(&bufqspin);
2745 * Returns true if no I/O is needed to access the associated VM object.
2746 * This is like findblk except it also hunts around in the VM system for
2749 * Note that we ignore vm_page_free() races from interrupts against our
2750 * lookup, since if the caller is not protected our return value will not
2751 * be any more valid then otherwise once we exit the critical section.
2754 inmem(struct vnode *vp, off_t loffset)
2757 vm_offset_t toff, tinc, size;
2761 if (findblk(vp, loffset, FINDBLK_TEST))
2763 if (vp->v_mount == NULL)
2765 if ((obj = vp->v_object) == NULL)
2769 if (size > vp->v_mount->mnt_stat.f_iosize)
2770 size = vp->v_mount->mnt_stat.f_iosize;
2772 vm_object_hold(obj);
2773 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2774 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2780 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2781 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2782 if (vm_page_is_valid(m,
2783 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) {
2788 vm_object_drop(obj);
2795 * Locate and return the specified buffer. Unless flagged otherwise,
2796 * a locked buffer will be returned if it exists or NULL if it does not.
2798 * findblk()'d buffers are still on the bufqueues and if you intend
2799 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2800 * and possibly do other stuff to it.
2802 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2803 * for locking the buffer and ensuring that it remains
2804 * the desired buffer after locking.
2806 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2807 * to acquire the lock we return NULL, even if the
2810 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal
2811 * reuse by getnewbuf() but does not prevent
2812 * disassociation (B_INVAL). Used to avoid deadlocks
2813 * against random (vp,loffset)s due to reassignment.
2815 * (0) - Lock the buffer blocking.
2820 findblk(struct vnode *vp, off_t loffset, int flags)
2825 lkflags = LK_EXCLUSIVE;
2826 if (flags & FINDBLK_NBLOCK)
2827 lkflags |= LK_NOWAIT;
2831 * Lookup. Ref the buf while holding v_token to prevent
2832 * reuse (but does not prevent diassociation).
2834 lwkt_gettoken(&vp->v_token);
2835 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2837 lwkt_reltoken(&vp->v_token);
2841 lwkt_reltoken(&vp->v_token);
2844 * If testing only break and return bp, do not lock.
2846 if (flags & FINDBLK_TEST)
2850 * Lock the buffer, return an error if the lock fails.
2851 * (only FINDBLK_NBLOCK can cause the lock to fail).
2853 if (BUF_LOCK(bp, lkflags)) {
2854 atomic_subtract_int(&bp->b_refs, 1);
2855 /* bp = NULL; not needed */
2860 * Revalidate the locked buf before allowing it to be
2863 if (bp->b_vp == vp && bp->b_loffset == loffset)
2865 atomic_subtract_int(&bp->b_refs, 1);
2872 if ((flags & FINDBLK_REF) == 0)
2873 atomic_subtract_int(&bp->b_refs, 1);
2880 * Similar to getblk() except only returns the buffer if it is
2881 * B_CACHE and requires no other manipulation. Otherwise NULL
2884 * If B_RAM is set the buffer might be just fine, but we return
2885 * NULL anyway because we want the code to fall through to the
2886 * cluster read. Otherwise read-ahead breaks.
2888 * If blksize is 0 the buffer cache buffer must already be fully
2891 * If blksize is non-zero getblk() will be used, allowing a buffer
2892 * to be reinstantiated from its VM backing store. The buffer must
2893 * still be fully cached after reinstantiation to be returned.
2896 getcacheblk(struct vnode *vp, off_t loffset, int blksize)
2901 bp = getblk(vp, loffset, blksize, 0, 0);
2903 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2905 bp->b_flags &= ~B_AGE;
2912 bp = findblk(vp, loffset, 0);
2914 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2916 bp->b_flags &= ~B_AGE;
2930 * Get a block given a specified block and offset into a file/device.
2931 * B_INVAL may or may not be set on return. The caller should clear
2932 * B_INVAL prior to initiating a READ.
2934 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2935 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2936 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2937 * without doing any of those things the system will likely believe
2938 * the buffer to be valid (especially if it is not B_VMIO), and the
2939 * next getblk() will return the buffer with B_CACHE set.
2941 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2942 * an existing buffer.
2944 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2945 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2946 * and then cleared based on the backing VM. If the previous buffer is
2947 * non-0-sized but invalid, B_CACHE will be cleared.
2949 * If getblk() must create a new buffer, the new buffer is returned with
2950 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2951 * case it is returned with B_INVAL clear and B_CACHE set based on the
2954 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2955 * B_CACHE bit is clear.
2957 * What this means, basically, is that the caller should use B_CACHE to
2958 * determine whether the buffer is fully valid or not and should clear
2959 * B_INVAL prior to issuing a read. If the caller intends to validate
2960 * the buffer by loading its data area with something, the caller needs
2961 * to clear B_INVAL. If the caller does this without issuing an I/O,
2962 * the caller should set B_CACHE ( as an optimization ), else the caller
2963 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2964 * a write attempt or if it was a successfull read. If the caller
2965 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2966 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2970 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2971 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2976 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2979 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2983 if (size > MAXBSIZE)
2984 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2985 if (vp->v_object == NULL)
2986 panic("getblk: vnode %p has no object!", vp);
2989 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) {
2991 * The buffer was found in the cache, but we need to lock it.
2992 * We must acquire a ref on the bp to prevent reuse, but
2993 * this will not prevent disassociation (brelvp()) so we
2994 * must recheck (vp,loffset) after acquiring the lock.
2996 * Without the ref the buffer could potentially be reused
2997 * before we acquire the lock and create a deadlock
2998 * situation between the thread trying to reuse the buffer
2999 * and us due to the fact that we would wind up blocking
3000 * on a random (vp,loffset).
3002 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
3003 if (blkflags & GETBLK_NOWAIT) {
3007 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
3008 if (blkflags & GETBLK_PCATCH)
3009 lkflags |= LK_PCATCH;
3010 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
3013 if (error == ENOLCK)
3017 /* buffer may have changed on us */
3022 * Once the buffer has been locked, make sure we didn't race
3023 * a buffer recyclement. Buffers that are no longer hashed
3024 * will have b_vp == NULL, so this takes care of that check
3027 if (bp->b_vp != vp || bp->b_loffset != loffset) {
3028 kprintf("Warning buffer %p (vp %p loffset %lld) "
3030 bp, vp, (long long)loffset);
3036 * If SZMATCH any pre-existing buffer must be of the requested
3037 * size or NULL is returned. The caller absolutely does not
3038 * want getblk() to bwrite() the buffer on a size mismatch.
3040 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
3046 * All vnode-based buffers must be backed by a VM object.
3048 KKASSERT(bp->b_flags & B_VMIO);
3049 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3050 bp->b_flags &= ~B_AGE;
3053 * Make sure that B_INVAL buffers do not have a cached
3054 * block number translation.
3056 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
3057 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
3058 " did not have cleared bio_offset cache\n",
3059 bp, vp, (long long)loffset);
3060 clearbiocache(&bp->b_bio2);
3064 * The buffer is locked. B_CACHE is cleared if the buffer is
3067 if (bp->b_flags & B_INVAL)
3068 bp->b_flags &= ~B_CACHE;
3072 * Any size inconsistancy with a dirty buffer or a buffer
3073 * with a softupdates dependancy must be resolved. Resizing
3074 * the buffer in such circumstances can lead to problems.
3076 * Dirty or dependant buffers are written synchronously.
3077 * Other types of buffers are simply released and
3078 * reconstituted as they may be backed by valid, dirty VM
3079 * pages (but not marked B_DELWRI).
3081 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
3082 * and may be left over from a prior truncation (and thus
3083 * no longer represent the actual EOF point), so we
3084 * definitely do not want to B_NOCACHE the backing store.
3086 if (size != bp->b_bcount) {
3087 if (bp->b_flags & B_DELWRI) {
3088 bp->b_flags |= B_RELBUF;
3090 } else if (LIST_FIRST(&bp->b_dep)) {
3091 bp->b_flags |= B_RELBUF;
3094 bp->b_flags |= B_RELBUF;
3099 KKASSERT(size <= bp->b_kvasize);
3100 KASSERT(bp->b_loffset != NOOFFSET,
3101 ("getblk: no buffer offset"));
3104 * A buffer with B_DELWRI set and B_CACHE clear must
3105 * be committed before we can return the buffer in
3106 * order to prevent the caller from issuing a read
3107 * ( due to B_CACHE not being set ) and overwriting
3110 * Most callers, including NFS and FFS, need this to
3111 * operate properly either because they assume they
3112 * can issue a read if B_CACHE is not set, or because
3113 * ( for example ) an uncached B_DELWRI might loop due
3114 * to softupdates re-dirtying the buffer. In the latter
3115 * case, B_CACHE is set after the first write completes,
3116 * preventing further loops.
3118 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
3119 * above while extending the buffer, we cannot allow the
3120 * buffer to remain with B_CACHE set after the write
3121 * completes or it will represent a corrupt state. To
3122 * deal with this we set B_NOCACHE to scrap the buffer
3125 * XXX Should this be B_RELBUF instead of B_NOCACHE?
3126 * I'm not even sure this state is still possible
3127 * now that getblk() writes out any dirty buffers
3130 * We might be able to do something fancy, like setting
3131 * B_CACHE in bwrite() except if B_DELWRI is already set,
3132 * so the below call doesn't set B_CACHE, but that gets real
3133 * confusing. This is much easier.
3136 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3137 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
3138 "and CACHE clear, b_flags %08x\n",
3139 bp, (intmax_t)bp->b_loffset, bp->b_flags);
3140 bp->b_flags |= B_NOCACHE;
3146 * Buffer is not in-core, create new buffer. The buffer
3147 * returned by getnewbuf() is locked. Note that the returned
3148 * buffer is also considered valid (not marked B_INVAL).
3150 * Calculating the offset for the I/O requires figuring out
3151 * the block size. We use DEV_BSIZE for VBLK or VCHR and
3152 * the mount's f_iosize otherwise. If the vnode does not
3153 * have an associated mount we assume that the passed size is
3156 * Note that vn_isdisk() cannot be used here since it may
3157 * return a failure for numerous reasons. Note that the
3158 * buffer size may be larger then the block size (the caller
3159 * will use block numbers with the proper multiple). Beware
3160 * of using any v_* fields which are part of unions. In
3161 * particular, in DragonFly the mount point overloading
3162 * mechanism uses the namecache only and the underlying
3163 * directory vnode is not a special case.
3167 if (vp->v_type == VBLK || vp->v_type == VCHR)
3169 else if (vp->v_mount)
3170 bsize = vp->v_mount->mnt_stat.f_iosize;
3174 maxsize = size + (loffset & PAGE_MASK);
3175 maxsize = imax(maxsize, bsize);
3177 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
3179 if (slpflags || slptimeo)
3185 * Atomically insert the buffer into the hash, so that it can
3186 * be found by findblk().
3188 * If bgetvp() returns non-zero a collision occured, and the
3189 * bp will not be associated with the vnode.
3191 * Make sure the translation layer has been cleared.
3193 bp->b_loffset = loffset;
3194 bp->b_bio2.bio_offset = NOOFFSET;
3195 /* bp->b_bio2.bio_next = NULL; */
3197 if (bgetvp(vp, bp, size)) {
3198 bp->b_flags |= B_INVAL;
3204 * All vnode-based buffers must be backed by a VM object.
3206 KKASSERT(vp->v_object != NULL);
3207 bp->b_flags |= B_VMIO;
3208 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3212 KKASSERT(dsched_is_clear_buf_priv(bp));
3219 * Reacquire a buffer that was previously released to the locked queue,
3220 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
3221 * set B_LOCKED (which handles the acquisition race).
3223 * To this end, either B_LOCKED must be set or the dependancy list must be
3229 regetblk(struct buf *bp)
3231 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
3232 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
3239 * Get an empty, disassociated buffer of given size. The buffer is
3240 * initially set to B_INVAL.
3242 * critical section protection is not required for the allocbuf()
3243 * call because races are impossible here.
3253 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3255 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
3258 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3259 KKASSERT(dsched_is_clear_buf_priv(bp));
3267 * This code constitutes the buffer memory from either anonymous system
3268 * memory (in the case of non-VMIO operations) or from an associated
3269 * VM object (in the case of VMIO operations). This code is able to
3270 * resize a buffer up or down.
3272 * Note that this code is tricky, and has many complications to resolve
3273 * deadlock or inconsistant data situations. Tread lightly!!!
3274 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3275 * the caller. Calling this code willy nilly can result in the loss of
3278 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3279 * B_CACHE for the non-VMIO case.
3281 * This routine does not need to be called from a critical section but you
3282 * must own the buffer.
3287 allocbuf(struct buf *bp, int size)
3289 int newbsize, mbsize;
3292 if (BUF_REFCNT(bp) == 0)
3293 panic("allocbuf: buffer not busy");
3295 if (bp->b_kvasize < size)
3296 panic("allocbuf: buffer too small");
3298 if ((bp->b_flags & B_VMIO) == 0) {
3302 * Just get anonymous memory from the kernel. Don't
3303 * mess with B_CACHE.
3305 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3306 if (bp->b_flags & B_MALLOC)
3309 newbsize = round_page(size);
3311 if (newbsize < bp->b_bufsize) {
3313 * Malloced buffers are not shrunk
3315 if (bp->b_flags & B_MALLOC) {
3317 bp->b_bcount = size;
3319 kfree(bp->b_data, M_BIOBUF);
3320 if (bp->b_bufsize) {
3321 atomic_subtract_int(&bufmallocspace, bp->b_bufsize);
3325 bp->b_data = bp->b_kvabase;
3327 bp->b_flags &= ~B_MALLOC;
3333 (vm_offset_t) bp->b_data + newbsize,
3334 (vm_offset_t) bp->b_data + bp->b_bufsize);
3335 } else if (newbsize > bp->b_bufsize) {
3337 * We only use malloced memory on the first allocation.
3338 * and revert to page-allocated memory when the buffer
3341 if ((bufmallocspace < maxbufmallocspace) &&
3342 (bp->b_bufsize == 0) &&
3343 (mbsize <= PAGE_SIZE/2)) {
3345 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3346 bp->b_bufsize = mbsize;
3347 bp->b_bcount = size;
3348 bp->b_flags |= B_MALLOC;
3349 atomic_add_int(&bufmallocspace, mbsize);
3355 * If the buffer is growing on its other-than-first
3356 * allocation, then we revert to the page-allocation
3359 if (bp->b_flags & B_MALLOC) {
3360 origbuf = bp->b_data;
3361 origbufsize = bp->b_bufsize;
3362 bp->b_data = bp->b_kvabase;
3363 if (bp->b_bufsize) {
3364 atomic_subtract_int(&bufmallocspace,
3369 bp->b_flags &= ~B_MALLOC;
3370 newbsize = round_page(newbsize);
3374 (vm_offset_t) bp->b_data + bp->b_bufsize,
3375 (vm_offset_t) bp->b_data + newbsize);
3377 bcopy(origbuf, bp->b_data, origbufsize);
3378 kfree(origbuf, M_BIOBUF);
3385 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3386 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3387 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3388 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3390 if (bp->b_flags & B_MALLOC)
3391 panic("allocbuf: VMIO buffer can't be malloced");
3393 * Set B_CACHE initially if buffer is 0 length or will become
3396 if (size == 0 || bp->b_bufsize == 0)
3397 bp->b_flags |= B_CACHE;
3399 if (newbsize < bp->b_bufsize) {
3401 * DEV_BSIZE aligned new buffer size is less then the
3402 * DEV_BSIZE aligned existing buffer size. Figure out
3403 * if we have to remove any pages.
3405 if (desiredpages < bp->b_xio.xio_npages) {
3406 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3408 * the page is not freed here -- it
3409 * is the responsibility of
3410 * vnode_pager_setsize
3412 m = bp->b_xio.xio_pages[i];
3413 KASSERT(m != bogus_page,
3414 ("allocbuf: bogus page found"));
3415 vm_page_busy_wait(m, TRUE, "biodep");
3416 bp->b_xio.xio_pages[i] = NULL;
3417 vm_page_unwire(m, 0);
3420 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3421 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3422 bp->b_xio.xio_npages = desiredpages;
3424 } else if (size > bp->b_bcount) {
3426 * We are growing the buffer, possibly in a
3427 * byte-granular fashion.
3435 * Step 1, bring in the VM pages from the object,
3436 * allocating them if necessary. We must clear
3437 * B_CACHE if these pages are not valid for the
3438 * range covered by the buffer.
3440 * critical section protection is required to protect
3441 * against interrupts unbusying and freeing pages
3442 * between our vm_page_lookup() and our
3443 * busycheck/wiring call.
3448 vm_object_hold(obj);
3449 while (bp->b_xio.xio_npages < desiredpages) {
3454 pi = OFF_TO_IDX(bp->b_loffset) +
3455 bp->b_xio.xio_npages;
3458 * Blocking on m->busy might lead to a
3461 * vm_fault->getpages->cluster_read->allocbuf
3463 m = vm_page_lookup_busy_try(obj, pi, FALSE,
3466 vm_page_sleep_busy(m, FALSE, "pgtblk");
3471 * note: must allocate system pages
3472 * since blocking here could intefere
3473 * with paging I/O, no matter which
3476 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
3479 vm_page_flag_clear(m, PG_ZERO);
3481 bp->b_flags &= ~B_CACHE;
3482 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3483 ++bp->b_xio.xio_npages;
3489 * We found a page and were able to busy it.
3491 vm_page_flag_clear(m, PG_ZERO);
3494 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3495 ++bp->b_xio.xio_npages;
3496 if (bp->b_act_count < m->act_count)
3497 bp->b_act_count = m->act_count;
3499 vm_object_drop(obj);
3502 * Step 2. We've loaded the pages into the buffer,
3503 * we have to figure out if we can still have B_CACHE
3504 * set. Note that B_CACHE is set according to the
3505 * byte-granular range ( bcount and size ), not the
3506 * aligned range ( newbsize ).
3508 * The VM test is against m->valid, which is DEV_BSIZE
3509 * aligned. Needless to say, the validity of the data
3510 * needs to also be DEV_BSIZE aligned. Note that this
3511 * fails with NFS if the server or some other client
3512 * extends the file's EOF. If our buffer is resized,
3513 * B_CACHE may remain set! XXX
3516 toff = bp->b_bcount;
3517 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3519 while ((bp->b_flags & B_CACHE) && toff < size) {
3522 if (tinc > (size - toff))
3525 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3533 bp->b_xio.xio_pages[pi]
3540 * Step 3, fixup the KVM pmap. Remember that
3541 * bp->b_data is relative to bp->b_loffset, but
3542 * bp->b_loffset may be offset into the first page.
3545 bp->b_data = (caddr_t)
3546 trunc_page((vm_offset_t)bp->b_data);
3548 (vm_offset_t)bp->b_data,
3549 bp->b_xio.xio_pages,
3550 bp->b_xio.xio_npages
3552 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3553 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3557 /* adjust space use on already-dirty buffer */
3558 if (bp->b_flags & B_DELWRI) {
3559 spin_lock(&bufcspin);
3560 dirtybufspace += newbsize - bp->b_bufsize;
3561 if (bp->b_flags & B_HEAVY)
3562 dirtybufspacehw += newbsize - bp->b_bufsize;
3563 spin_unlock(&bufcspin);
3565 if (newbsize < bp->b_bufsize)
3567 bp->b_bufsize = newbsize; /* actual buffer allocation */
3568 bp->b_bcount = size; /* requested buffer size */
3575 * Wait for buffer I/O completion, returning error status. B_EINTR
3576 * is converted into an EINTR error but not cleared (since a chain
3577 * of biowait() calls may occur).
3579 * On return bpdone() will have been called but the buffer will remain
3580 * locked and will not have been brelse()'d.
3582 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3583 * likely still in progress on return.
3585 * NOTE! This operation is on a BIO, not a BUF.
3587 * NOTE! BIO_DONE is cleared by vn_strategy()
3592 _biowait(struct bio *bio, const char *wmesg, int to)
3594 struct buf *bp = bio->bio_buf;
3599 KKASSERT(bio == &bp->b_bio1);
3601 flags = bio->bio_flags;
3602 if (flags & BIO_DONE)
3604 nflags = flags | BIO_WANT;
3605 tsleep_interlock(bio, 0);
3606 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3608 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3609 else if (bp->b_cmd == BUF_CMD_READ)
3610 error = tsleep(bio, PINTERLOCKED, "biord", to);
3612 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3614 kprintf("tsleep error biowait %d\n", error);
3623 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3624 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3625 if (bp->b_flags & B_EINTR)
3627 if (bp->b_flags & B_ERROR)
3628 return (bp->b_error ? bp->b_error : EIO);
3633 biowait(struct bio *bio, const char *wmesg)
3635 return(_biowait(bio, wmesg, 0));
3639 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3641 return(_biowait(bio, wmesg, to));
3645 * This associates a tracking count with an I/O. vn_strategy() and
3646 * dev_dstrategy() do this automatically but there are a few cases
3647 * where a vnode or device layer is bypassed when a block translation
3648 * is cached. In such cases bio_start_transaction() may be called on
3649 * the bypassed layers so the system gets an I/O in progress indication
3650 * for those higher layers.
3653 bio_start_transaction(struct bio *bio, struct bio_track *track)
3655 bio->bio_track = track;
3656 if (dsched_is_clear_buf_priv(bio->bio_buf))
3657 dsched_new_buf(bio->bio_buf);
3658 bio_track_ref(track);
3662 * Initiate I/O on a vnode.
3664 * SWAPCACHE OPERATION:
3666 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately
3667 * devfs also uses b_vp for fake buffers so we also have to check
3668 * that B_PAGING is 0. In this case the passed 'vp' is probably the
3669 * underlying block device. The swap assignments are related to the
3670 * buffer cache buffer's b_vp, not the passed vp.
3672 * The passed vp == bp->b_vp only in the case where the strategy call
3673 * is made on the vp itself for its own buffers (a regular file or
3674 * block device vp). The filesystem usually then re-calls vn_strategy()
3675 * after translating the request to an underlying device.
3677 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the
3678 * underlying buffer cache buffers.
3680 * We can only deal with page-aligned buffers at the moment, because
3681 * we can't tell what the real dirty state for pages straddling a buffer
3684 * In order to call swap_pager_strategy() we must provide the VM object
3685 * and base offset for the underlying buffer cache pages so it can find
3689 vn_strategy(struct vnode *vp, struct bio *bio)
3691 struct bio_track *track;
3692 struct buf *bp = bio->bio_buf;
3694 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3697 * Set when an I/O is issued on the bp. Cleared by consumers
3698 * (aka HAMMER), allowing the consumer to determine if I/O had
3699 * actually occurred.
3701 bp->b_flags |= B_IODEBUG;
3704 * Handle the swap cache intercept.
3706 if (vn_cache_strategy(vp, bio))
3710 * Otherwise do the operation through the filesystem
3712 if (bp->b_cmd == BUF_CMD_READ)
3713 track = &vp->v_track_read;
3715 track = &vp->v_track_write;
3716 KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3717 bio->bio_track = track;
3718 if (dsched_is_clear_buf_priv(bio->bio_buf))
3719 dsched_new_buf(bio->bio_buf);
3720 bio_track_ref(track);
3721 vop_strategy(*vp->v_ops, vp, bio);
3724 static void vn_cache_strategy_callback(struct bio *bio);
3727 vn_cache_strategy(struct vnode *vp, struct bio *bio)
3729 struct buf *bp = bio->bio_buf;
3736 * Is this buffer cache buffer suitable for reading from
3739 if (vm_swapcache_read_enable == 0 ||
3740 bp->b_cmd != BUF_CMD_READ ||
3741 ((bp->b_flags & B_CLUSTER) == 0 &&
3742 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) ||
3743 ((int)bp->b_loffset & PAGE_MASK) != 0 ||
3744 (bp->b_bcount & PAGE_MASK) != 0) {
3749 * Figure out the original VM object (it will match the underlying
3750 * VM pages). Note that swap cached data uses page indices relative
3751 * to that object, not relative to bio->bio_offset.
3753 if (bp->b_flags & B_CLUSTER)
3754 object = vp->v_object;
3756 object = bp->b_vp->v_object;
3759 * In order to be able to use the swap cache all underlying VM
3760 * pages must be marked as such, and we can't have any bogus pages.
3762 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
3763 m = bp->b_xio.xio_pages[i];
3764 if ((m->flags & PG_SWAPPED) == 0)
3766 if (m == bogus_page)
3771 * If we are good then issue the I/O using swap_pager_strategy().
3773 * We can only do this if the buffer actually supports object-backed
3774 * I/O. If it doesn't npages will be 0.
3776 if (i && i == bp->b_xio.xio_npages) {
3777 m = bp->b_xio.xio_pages[0];
3778 nbio = push_bio(bio);
3779 nbio->bio_done = vn_cache_strategy_callback;
3780 nbio->bio_offset = ptoa(m->pindex);
3781 KKASSERT(m->object == object);
3782 swap_pager_strategy(object, nbio);
3789 * This is a bit of a hack but since the vn_cache_strategy() function can
3790 * override a VFS's strategy function we must make sure that the bio, which
3791 * is probably bio2, doesn't leak an unexpected offset value back to the
3792 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the
3793 * bio went through its own file strategy function and the the bio2 offset
3794 * is a cached disk offset when, in fact, it isn't.
3797 vn_cache_strategy_callback(struct bio *bio)
3799 bio->bio_offset = NOOFFSET;
3800 biodone(pop_bio(bio));
3806 * Finish I/O on a buffer after all BIOs have been processed.
3807 * Called when the bio chain is exhausted or by biowait. If called
3808 * by biowait, elseit is typically 0.
3810 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3811 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3812 * assuming B_INVAL is clear.
3814 * For the VMIO case, we set B_CACHE if the op was a read and no
3815 * read error occured, or if the op was a write. B_CACHE is never
3816 * set if the buffer is invalid or otherwise uncacheable.
3818 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3819 * initiator to leave B_INVAL set to brelse the buffer out of existance
3820 * in the biodone routine.
3823 bpdone(struct buf *bp, int elseit)
3827 KASSERT(BUF_REFCNTNB(bp) > 0,
3828 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3829 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3830 ("biodone: bp %p already done!", bp));
3833 * No more BIOs are left. All completion functions have been dealt
3834 * with, now we clean up the buffer.
3837 bp->b_cmd = BUF_CMD_DONE;
3840 * Only reads and writes are processed past this point.
3842 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3843 if (cmd == BUF_CMD_FREEBLKS)
3844 bp->b_flags |= B_NOCACHE;
3851 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3852 * a lot worse. XXX - move this above the clearing of b_cmd
3854 if (LIST_FIRST(&bp->b_dep) != NULL)
3855 buf_complete(bp); /* MPSAFE */
3858 * A failed write must re-dirty the buffer unless B_INVAL
3859 * was set. Only applicable to normal buffers (with VPs).
3860 * vinum buffers may not have a vp.
3862 if (cmd == BUF_CMD_WRITE &&
3863 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3864 bp->b_flags &= ~B_NOCACHE;
3869 if (bp->b_flags & B_VMIO) {
3875 struct vnode *vp = bp->b_vp;
3879 #if defined(VFS_BIO_DEBUG)
3880 if (vp->v_auxrefs == 0)
3881 panic("biodone: zero vnode hold count");
3882 if ((vp->v_flag & VOBJBUF) == 0)
3883 panic("biodone: vnode is not setup for merged cache");
3886 foff = bp->b_loffset;
3887 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3888 KASSERT(obj != NULL, ("biodone: missing VM object"));
3890 #if defined(VFS_BIO_DEBUG)
3891 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3892 kprintf("biodone: paging in progress(%d) < "
3893 "bp->b_xio.xio_npages(%d)\n",
3894 obj->paging_in_progress,
3895 bp->b_xio.xio_npages);
3900 * Set B_CACHE if the op was a normal read and no error
3901 * occured. B_CACHE is set for writes in the b*write()
3904 iosize = bp->b_bcount - bp->b_resid;
3905 if (cmd == BUF_CMD_READ &&
3906 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3907 bp->b_flags |= B_CACHE;
3910 vm_object_hold(obj);
3911 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3915 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3920 * cleanup bogus pages, restoring the originals. Since
3921 * the originals should still be wired, we don't have
3922 * to worry about interrupt/freeing races destroying
3923 * the VM object association.
3925 m = bp->b_xio.xio_pages[i];
3926 if (m == bogus_page) {
3928 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3930 panic("biodone: page disappeared");
3931 bp->b_xio.xio_pages[i] = m;
3932 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3933 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3935 #if defined(VFS_BIO_DEBUG)
3936 if (OFF_TO_IDX(foff) != m->pindex) {
3937 kprintf("biodone: foff(%lu)/m->pindex(%ld) "
3939 (unsigned long)foff, (long)m->pindex);
3944 * In the write case, the valid and clean bits are
3945 * already changed correctly (see bdwrite()), so we
3946 * only need to do this here in the read case.
3948 vm_page_busy_wait(m, FALSE, "bpdpgw");
3949 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3950 vfs_clean_one_page(bp, i, m);
3952 vm_page_flag_clear(m, PG_ZERO);
3955 * when debugging new filesystems or buffer I/O
3956 * methods, this is the most common error that pops
3957 * up. if you see this, you have not set the page
3958 * busy flag correctly!!!
3961 kprintf("biodone: page busy < 0, "
3962 "pindex: %d, foff: 0x(%x,%x), "
3963 "resid: %d, index: %d\n",
3964 (int) m->pindex, (int)(foff >> 32),
3965 (int) foff & 0xffffffff, resid, i);
3966 if (!vn_isdisk(vp, NULL))
3967 kprintf(" iosize: %ld, loffset: %lld, "
3968 "flags: 0x%08x, npages: %d\n",
3969 bp->b_vp->v_mount->mnt_stat.f_iosize,
3970 (long long)bp->b_loffset,
3971 bp->b_flags, bp->b_xio.xio_npages);
3973 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3974 (long long)bp->b_loffset,
3975 bp->b_flags, bp->b_xio.xio_npages);
3976 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3977 m->valid, m->dirty, m->wire_count);
3978 panic("biodone: page busy < 0");
3980 vm_page_io_finish(m);
3982 vm_object_pip_wakeup(obj);
3983 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3986 bp->b_flags &= ~B_HASBOGUS;
3987 vm_object_drop(obj);
3991 * Finish up by releasing the buffer. There are no more synchronous
3992 * or asynchronous completions, those were handled by bio_done
3996 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
4007 biodone(struct bio *bio)
4009 struct buf *bp = bio->bio_buf;
4011 runningbufwakeup(bp);
4014 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
4017 biodone_t *done_func;
4018 struct bio_track *track;
4021 * BIO tracking. Most but not all BIOs are tracked.
4023 if ((track = bio->bio_track) != NULL) {
4024 bio_track_rel(track);
4025 bio->bio_track = NULL;
4029 * A bio_done function terminates the loop. The function
4030 * will be responsible for any further chaining and/or
4031 * buffer management.
4033 * WARNING! The done function can deallocate the buffer!
4035 if ((done_func = bio->bio_done) != NULL) {
4036 bio->bio_done = NULL;
4040 bio = bio->bio_prev;
4044 * If we've run out of bio's do normal [a]synchronous completion.
4050 * Synchronous biodone - this terminates a synchronous BIO.
4052 * bpdone() is called with elseit=FALSE, leaving the buffer completed
4053 * but still locked. The caller must brelse() the buffer after waiting
4057 biodone_sync(struct bio *bio)
4059 struct buf *bp = bio->bio_buf;
4063 KKASSERT(bio == &bp->b_bio1);
4067 flags = bio->bio_flags;
4068 nflags = (flags | BIO_DONE) & ~BIO_WANT;
4070 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
4071 if (flags & BIO_WANT)
4081 * This routine is called in lieu of iodone in the case of
4082 * incomplete I/O. This keeps the busy status for pages
4086 vfs_unbusy_pages(struct buf *bp)
4090 runningbufwakeup(bp);
4092 if (bp->b_flags & B_VMIO) {
4093 struct vnode *vp = bp->b_vp;
4097 vm_object_hold(obj);
4099 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4100 vm_page_t m = bp->b_xio.xio_pages[i];
4103 * When restoring bogus changes the original pages
4104 * should still be wired, so we are in no danger of
4105 * losing the object association and do not need
4106 * critical section protection particularly.
4108 if (m == bogus_page) {
4109 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
4111 panic("vfs_unbusy_pages: page missing");
4113 bp->b_xio.xio_pages[i] = m;
4114 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4115 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
4117 vm_page_busy_wait(m, FALSE, "bpdpgw");
4118 vm_page_flag_clear(m, PG_ZERO);
4119 vm_page_io_finish(m);
4121 vm_object_pip_wakeup(obj);
4123 bp->b_flags &= ~B_HASBOGUS;
4124 vm_object_drop(obj);
4131 * This routine is called before a device strategy routine.
4132 * It is used to tell the VM system that paging I/O is in
4133 * progress, and treat the pages associated with the buffer
4134 * almost as being PG_BUSY. Also the object 'paging_in_progress'
4135 * flag is handled to make sure that the object doesn't become
4138 * Since I/O has not been initiated yet, certain buffer flags
4139 * such as B_ERROR or B_INVAL may be in an inconsistant state
4140 * and should be ignored.
4145 vfs_busy_pages(struct vnode *vp, struct buf *bp)
4148 struct lwp *lp = curthread->td_lwp;
4151 * The buffer's I/O command must already be set. If reading,
4152 * B_CACHE must be 0 (double check against callers only doing
4153 * I/O when B_CACHE is 0).
4155 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4156 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
4158 if (bp->b_flags & B_VMIO) {
4162 KASSERT(bp->b_loffset != NOOFFSET,
4163 ("vfs_busy_pages: no buffer offset"));
4166 * Busy all the pages. We have to busy them all at once
4167 * to avoid deadlocks.
4170 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4171 vm_page_t m = bp->b_xio.xio_pages[i];
4173 if (vm_page_busy_try(m, FALSE)) {
4174 vm_page_sleep_busy(m, FALSE, "vbpage");
4176 vm_page_wakeup(bp->b_xio.xio_pages[i]);
4182 * Setup for I/O, soft-busy the page right now because
4183 * the next loop may block.
4185 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4186 vm_page_t m = bp->b_xio.xio_pages[i];
4188 vm_page_flag_clear(m, PG_ZERO);
4189 if ((bp->b_flags & B_CLUSTER) == 0) {
4190 vm_object_pip_add(obj, 1);
4191 vm_page_io_start(m);
4196 * Adjust protections for I/O and do bogus-page mapping.
4197 * Assume that vm_page_protect() can block (it can block
4198 * if VM_PROT_NONE, don't take any chances regardless).
4200 * In particular note that for writes we must incorporate
4201 * page dirtyness from the VM system into the buffer's
4204 * For reads we theoretically must incorporate page dirtyness
4205 * from the VM system to determine if the page needs bogus
4206 * replacement, but we shortcut the test by simply checking
4207 * that all m->valid bits are set, indicating that the page
4208 * is fully valid and does not need to be re-read. For any
4209 * VM system dirtyness the page will also be fully valid
4210 * since it was mapped at one point.
4213 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4214 vm_page_t m = bp->b_xio.xio_pages[i];
4216 vm_page_flag_clear(m, PG_ZERO); /* XXX */
4217 if (bp->b_cmd == BUF_CMD_WRITE) {
4219 * When readying a vnode-backed buffer for
4220 * a write we must zero-fill any invalid
4221 * portions of the backing VM pages, mark
4222 * it valid and clear related dirty bits.
4224 * vfs_clean_one_page() incorporates any
4225 * VM dirtyness and updates the b_dirtyoff
4226 * range (after we've made the page RO).
4228 * It is also expected that the pmap modified
4229 * bit has already been cleared by the
4230 * vm_page_protect(). We may not be able
4231 * to clear all dirty bits for a page if it
4232 * was also memory mapped (NFS).
4234 * Finally be sure to unassign any swap-cache
4235 * backing store as it is now stale.
4237 vm_page_protect(m, VM_PROT_READ);
4238 vfs_clean_one_page(bp, i, m);
4239 swap_pager_unswapped(m);
4240 } else if (m->valid == VM_PAGE_BITS_ALL) {
4242 * When readying a vnode-backed buffer for
4243 * read we must replace any dirty pages with
4244 * a bogus page so dirty data is not destroyed
4245 * when filling gaps.
4247 * To avoid testing whether the page is
4248 * dirty we instead test that the page was
4249 * at some point mapped (m->valid fully
4250 * valid) with the understanding that
4251 * this also covers the dirty case.
4253 bp->b_xio.xio_pages[i] = bogus_page;
4254 bp->b_flags |= B_HASBOGUS;
4256 } else if (m->valid & m->dirty) {
4258 * This case should not occur as partial
4259 * dirtyment can only happen if the buffer
4260 * is B_CACHE, and this code is not entered
4261 * if the buffer is B_CACHE.
4263 kprintf("Warning: vfs_busy_pages - page not "
4264 "fully valid! loff=%jx bpf=%08x "
4265 "idx=%d val=%02x dir=%02x\n",
4266 (intmax_t)bp->b_loffset, bp->b_flags,
4267 i, m->valid, m->dirty);
4268 vm_page_protect(m, VM_PROT_NONE);
4271 * The page is not valid and can be made
4274 vm_page_protect(m, VM_PROT_NONE);
4279 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4280 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
4285 * This is the easiest place to put the process accounting for the I/O
4289 if (bp->b_cmd == BUF_CMD_READ)
4290 lp->lwp_ru.ru_inblock++;
4292 lp->lwp_ru.ru_oublock++;
4297 * Tell the VM system that the pages associated with this buffer
4298 * are clean. This is used for delayed writes where the data is
4299 * going to go to disk eventually without additional VM intevention.
4301 * NOTE: While we only really need to clean through to b_bcount, we
4302 * just go ahead and clean through to b_bufsize.
4305 vfs_clean_pages(struct buf *bp)
4310 if ((bp->b_flags & B_VMIO) == 0)
4313 KASSERT(bp->b_loffset != NOOFFSET,
4314 ("vfs_clean_pages: no buffer offset"));
4316 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4317 m = bp->b_xio.xio_pages[i];
4318 vfs_clean_one_page(bp, i, m);
4323 * vfs_clean_one_page:
4325 * Set the valid bits and clear the dirty bits in a page within a
4326 * buffer. The range is restricted to the buffer's size and the
4327 * buffer's logical offset might index into the first page.
4329 * The caller has busied or soft-busied the page and it is not mapped,
4330 * test and incorporate the dirty bits into b_dirtyoff/end before
4331 * clearing them. Note that we need to clear the pmap modified bits
4332 * after determining the the page was dirty, vm_page_set_validclean()
4333 * does not do it for us.
4335 * This routine is typically called after a read completes (dirty should
4336 * be zero in that case as we are not called on bogus-replace pages),
4337 * or before a write is initiated.
4340 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m)
4348 * Calculate offset range within the page but relative to buffer's
4349 * loffset. loffset might be offset into the first page.
4351 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4352 bcount = bp->b_bcount + xoff; /* offset adjusted */
4358 soff = (pageno << PAGE_SHIFT);
4359 eoff = soff + PAGE_SIZE;
4367 * Test dirty bits and adjust b_dirtyoff/end.
4369 * If dirty pages are incorporated into the bp any prior
4370 * B_NEEDCOMMIT state (NFS) must be cleared because the
4371 * caller has not taken into account the new dirty data.
4373 * If the page was memory mapped the dirty bits might go beyond the
4374 * end of the buffer, but we can't really make the assumption that
4375 * a file EOF straddles the buffer (even though this is the case for
4376 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
4377 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
4378 * This also saves some console spam.
4380 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK,
4381 * NFS can handle huge commits but not huge writes.
4383 vm_page_test_dirty(m);
4385 if ((bp->b_flags & B_NEEDCOMMIT) &&
4386 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) {
4388 kprintf("Warning: vfs_clean_one_page: bp %p "
4389 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT"
4390 " cmd %d vd %02x/%02x x/s/e %d %d %d "
4392 bp, (intmax_t)bp->b_loffset, bp->b_bcount,
4393 bp->b_flags, bp->b_cmd,
4394 m->valid, m->dirty, xoff, soff, eoff,
4395 bp->b_dirtyoff, bp->b_dirtyend);
4396 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
4398 print_backtrace(-1);
4401 * Only clear the pmap modified bits if ALL the dirty bits
4402 * are set, otherwise the system might mis-clear portions
4405 if (m->dirty == VM_PAGE_BITS_ALL &&
4406 (bp->b_flags & B_NEEDCOMMIT) == 0) {
4407 pmap_clear_modify(m);
4409 if (bp->b_dirtyoff > soff - xoff)
4410 bp->b_dirtyoff = soff - xoff;
4411 if (bp->b_dirtyend < eoff - xoff)
4412 bp->b_dirtyend = eoff - xoff;
4416 * Set related valid bits, clear related dirty bits.
4417 * Does not mess with the pmap modified bit.
4419 * WARNING! We cannot just clear all of m->dirty here as the
4420 * buffer cache buffers may use a DEV_BSIZE'd aligned
4421 * block size, or have an odd size (e.g. NFS at file EOF).
4422 * The putpages code can clear m->dirty to 0.
4424 * If a VOP_WRITE generates a buffer cache buffer which
4425 * covers the same space as mapped writable pages the
4426 * buffer flush might not be able to clear all the dirty
4427 * bits and still require a putpages from the VM system
4430 * WARNING! vm_page_set_validclean() currently assumes vm_token
4431 * is held. The page might not be busied (bdwrite() case).
4432 * XXX remove this comment once we've validated that this
4433 * is no longer an issue.
4435 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff);
4439 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty.
4440 * The page data is assumed to be valid (there is no zeroing here).
4443 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m)
4451 * Calculate offset range within the page but relative to buffer's
4452 * loffset. loffset might be offset into the first page.
4454 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4455 bcount = bp->b_bcount + xoff; /* offset adjusted */
4461 soff = (pageno << PAGE_SHIFT);
4462 eoff = soff + PAGE_SIZE;
4468 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff);
4474 * Clear a buffer. This routine essentially fakes an I/O, so we need
4475 * to clear B_ERROR and B_INVAL.
4477 * Note that while we only theoretically need to clear through b_bcount,
4478 * we go ahead and clear through b_bufsize.
4482 vfs_bio_clrbuf(struct buf *bp)
4486 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
4487 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
4488 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4489 (bp->b_loffset & PAGE_MASK) == 0) {
4490 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4491 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
4495 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
4496 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
4497 bzero(bp->b_data, bp->b_bufsize);
4498 bp->b_xio.xio_pages[0]->valid |= mask;
4504 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
4505 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
4506 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
4507 ea = (caddr_t)(vm_offset_t)ulmin(
4508 (u_long)(vm_offset_t)ea,
4509 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
4510 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4511 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
4513 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
4514 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
4518 for (; sa < ea; sa += DEV_BSIZE, j++) {
4519 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
4520 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
4521 bzero(sa, DEV_BSIZE);
4524 bp->b_xio.xio_pages[i]->valid |= mask;
4525 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
4534 * vm_hold_load_pages:
4536 * Load pages into the buffer's address space. The pages are
4537 * allocated from the kernel object in order to reduce interference
4538 * with the any VM paging I/O activity. The range of loaded
4539 * pages will be wired.
4541 * If a page cannot be allocated, the 'pagedaemon' is woken up to
4542 * retrieve the full range (to - from) of pages.
4547 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4553 to = round_page(to);
4554 from = round_page(from);
4555 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4560 * Note: must allocate system pages since blocking here
4561 * could intefere with paging I/O, no matter which
4564 vm_object_hold(&kernel_object);
4565 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT,
4566 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
4567 vm_object_drop(&kernel_object);
4570 p->valid = VM_PAGE_BITS_ALL;
4571 vm_page_flag_clear(p, PG_ZERO);
4572 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
4573 bp->b_xio.xio_pages[index] = p;
4580 bp->b_xio.xio_npages = index;
4584 * Allocate pages for a buffer cache buffer.
4586 * Under extremely severe memory conditions even allocating out of the
4587 * system reserve can fail. If this occurs we must allocate out of the
4588 * interrupt reserve to avoid a deadlock with the pageout daemon.
4590 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
4591 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
4592 * against the pageout daemon if pages are not freed from other sources.
4598 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit)
4602 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj));
4605 * Try a normal allocation, allow use of system reserve.
4607 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
4612 * The normal allocation failed and we clearly have a page
4613 * deficit. Try to reclaim some clean VM pages directly
4614 * from the buffer cache.
4616 vm_pageout_deficit += deficit;
4620 * We may have blocked, the caller will know what to do if the
4623 if (vm_page_lookup(obj, pg)) {
4628 * Allocate and allow use of the interrupt reserve.
4630 * If after all that we still can't allocate a VM page we are
4631 * in real trouble, but we slog on anyway hoping that the system
4634 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
4635 VM_ALLOC_INTERRUPT);
4637 if (vm_page_count_severe()) {
4639 vm_wait(hz / 20 + 1);
4642 kprintf("bio_page_alloc: Memory exhausted during bufcache "
4643 "page allocation\n");
4651 * vm_hold_free_pages:
4653 * Return pages associated with the buffer back to the VM system.
4655 * The range of pages underlying the buffer's address space will
4656 * be unmapped and un-wired.
4661 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4665 int index, newnpages;
4667 from = round_page(from);
4668 to = round_page(to);
4669 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4672 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4673 p = bp->b_xio.xio_pages[index];
4674 if (p && (index < bp->b_xio.xio_npages)) {
4676 kprintf("vm_hold_free_pages: doffset: %lld, "
4678 (long long)bp->b_bio2.bio_offset,
4679 (long long)bp->b_loffset);
4681 bp->b_xio.xio_pages[index] = NULL;
4683 vm_page_busy_wait(p, FALSE, "vmhldpg");
4684 vm_page_unwire(p, 0);
4688 bp->b_xio.xio_npages = newnpages;
4694 * Map a user buffer into KVM via a pbuf. On return the buffer's
4695 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4699 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4710 * bp had better have a command and it better be a pbuf.
4712 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4713 KKASSERT(bp->b_flags & B_PAGING);
4714 KKASSERT(bp->b_kvabase);
4720 * Map the user data into KVM. Mappings have to be page-aligned.
4722 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4725 vmprot = VM_PROT_READ;
4726 if (bp->b_cmd == BUF_CMD_READ)
4727 vmprot |= VM_PROT_WRITE;
4729 while (addr < udata + bytes) {
4731 * Do the vm_fault if needed; do the copy-on-write thing
4732 * when reading stuff off device into memory.
4734 * vm_fault_page*() returns a held VM page.
4736 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4737 va = trunc_page(va);
4739 m = vm_fault_page_quick(va, vmprot, &error);
4741 for (i = 0; i < pidx; ++i) {
4742 vm_page_unhold(bp->b_xio.xio_pages[i]);
4743 bp->b_xio.xio_pages[i] = NULL;
4747 bp->b_xio.xio_pages[pidx] = m;
4753 * Map the page array and set the buffer fields to point to
4754 * the mapped data buffer.
4756 if (pidx > btoc(MAXPHYS))
4757 panic("vmapbuf: mapped more than MAXPHYS");
4758 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4760 bp->b_xio.xio_npages = pidx;
4761 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4762 bp->b_bcount = bytes;
4763 bp->b_bufsize = bytes;
4770 * Free the io map PTEs associated with this IO operation.
4771 * We also invalidate the TLB entries and restore the original b_addr.
4774 vunmapbuf(struct buf *bp)
4779 KKASSERT(bp->b_flags & B_PAGING);
4781 npages = bp->b_xio.xio_npages;
4782 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4783 for (pidx = 0; pidx < npages; ++pidx) {
4784 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4785 bp->b_xio.xio_pages[pidx] = NULL;
4787 bp->b_xio.xio_npages = 0;
4788 bp->b_data = bp->b_kvabase;
4792 * Scan all buffers in the system and issue the callback.
4795 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4801 for (n = 0; n < nbuf; ++n) {
4802 if ((error = callback(&buf[n], info)) < 0) {
4812 * nestiobuf_iodone: biodone callback for nested buffers and propagate
4813 * completion to the master buffer.
4816 nestiobuf_iodone(struct bio *bio)
4819 struct buf *mbp, *bp;
4820 struct devstat *stats;
4825 mbio = bio->bio_caller_info1.ptr;
4826 stats = bio->bio_caller_info2.ptr;
4827 mbp = mbio->bio_buf;
4829 KKASSERT(bp->b_bcount <= bp->b_bufsize);
4830 KKASSERT(mbp != bp);
4832 error = bp->b_error;
4833 if (bp->b_error == 0 &&
4834 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
4836 * Not all got transfered, raise an error. We have no way to
4837 * propagate these conditions to mbp.
4842 donebytes = bp->b_bufsize;
4846 nestiobuf_done(mbio, donebytes, error, stats);
4850 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats)
4854 mbp = mbio->bio_buf;
4856 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0);
4859 * If an error occured, propagate it to the master buffer.
4861 * Several biodone()s may wind up running concurrently so
4862 * use an atomic op to adjust b_flags.
4865 mbp->b_error = error;
4866 atomic_set_int(&mbp->b_flags, B_ERROR);
4870 * Decrement the operations in progress counter and terminate the
4871 * I/O if this was the last bit.
4873 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4876 devstat_end_transaction_buf(stats, mbp);
4882 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent
4883 * the mbio from being biodone()'d while we are still adding sub-bios to
4887 nestiobuf_init(struct bio *bio)
4889 bio->bio_driver_info = (void *)1;
4893 * The BIOs added to the nestedio have already been started, remove the
4894 * count that placeheld our mbio and biodone() it if the count would
4898 nestiobuf_start(struct bio *mbio)
4900 struct buf *mbp = mbio->bio_buf;
4903 * Decrement the operations in progress counter and terminate the
4904 * I/O if this was the last bit.
4906 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4907 if (mbp->b_flags & B_ERROR)
4908 mbp->b_resid = mbp->b_bcount;
4916 * Set an intermediate error prior to calling nestiobuf_start()
4919 nestiobuf_error(struct bio *mbio, int error)
4921 struct buf *mbp = mbio->bio_buf;
4924 mbp->b_error = error;
4925 atomic_set_int(&mbp->b_flags, B_ERROR);
4930 * nestiobuf_add: setup a "nested" buffer.
4932 * => 'mbp' is a "master" buffer which is being divided into sub pieces.
4933 * => 'bp' should be a buffer allocated by getiobuf.
4934 * => 'offset' is a byte offset in the master buffer.
4935 * => 'size' is a size in bytes of this nested buffer.
4938 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats)
4940 struct buf *mbp = mbio->bio_buf;
4941 struct vnode *vp = mbp->b_vp;
4943 KKASSERT(mbp->b_bcount >= offset + size);
4945 atomic_add_int((int *)&mbio->bio_driver_info, 1);
4947 /* kernel needs to own the lock for it to be released in biodone */
4950 bp->b_cmd = mbp->b_cmd;
4951 bp->b_bio1.bio_done = nestiobuf_iodone;
4952 bp->b_data = (char *)mbp->b_data + offset;
4953 bp->b_resid = bp->b_bcount = size;
4954 bp->b_bufsize = bp->b_bcount;
4956 bp->b_bio1.bio_track = NULL;
4957 bp->b_bio1.bio_caller_info1.ptr = mbio;
4958 bp->b_bio1.bio_caller_info2.ptr = stats;
4962 * print out statistics from the current status of the buffer pool
4963 * this can be toggeled by the system control option debug.syncprt
4972 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
4973 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
4975 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
4977 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4980 spin_lock(&bufqspin);
4981 TAILQ_FOREACH(bp, dp, b_freelist) {
4982 counts[bp->b_bufsize/PAGE_SIZE]++;
4985 spin_unlock(&bufqspin);
4987 kprintf("%s: total-%d", bname[i], count);
4988 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4990 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
4998 DB_SHOW_COMMAND(buffer, db_show_buffer)
5001 struct buf *bp = (struct buf *)addr;
5004 db_printf("usage: show buffer <addr>\n");
5008 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
5009 db_printf("b_cmd = %d\n", bp->b_cmd);
5010 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
5011 "b_resid = %d\n, b_data = %p, "
5012 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
5013 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5015 (long long)bp->b_bio2.bio_offset,
5016 (long long)(bp->b_bio2.bio_next ?
5017 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
5018 if (bp->b_xio.xio_npages) {
5020 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
5021 bp->b_xio.xio_npages);
5022 for (i = 0; i < bp->b_xio.xio_npages; i++) {
5024 m = bp->b_xio.xio_pages[i];
5025 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
5026 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
5027 if ((i + 1) < bp->b_xio.xio_npages)