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 $
18 * this file contains a new buffer I/O scheme implementing a coherent
19 * VM object and buffer cache scheme. Pains have been taken to make
20 * sure that the performance degradation associated with schemes such
21 * as this is not realized.
23 * Author: John S. Dyson
24 * Significant help during the development and debugging phases
25 * had been provided by David Greenman, also of the FreeBSD core team.
27 * see man buf(9) for more info.
30 #include <sys/param.h>
31 #include <sys/systm.h>
34 #include <sys/devicestat.h>
35 #include <sys/eventhandler.h>
37 #include <sys/malloc.h>
38 #include <sys/mount.h>
39 #include <sys/kernel.h>
40 #include <sys/kthread.h>
42 #include <sys/reboot.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sysctl.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
47 #include <sys/dsched.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_kern.h>
51 #include <vm/vm_pageout.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_map.h>
56 #include <vm/vm_pager.h>
57 #include <vm/swap_pager.h>
60 #include <sys/thread2.h>
61 #include <sys/spinlock2.h>
62 #include <sys/mplock2.h>
63 #include <vm/vm_page2.h>
74 BQUEUE_NONE, /* not on any queue */
75 BQUEUE_LOCKED, /* locked buffers */
76 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
77 BQUEUE_DIRTY, /* B_DELWRI buffers */
78 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
79 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
80 BQUEUE_EMPTY, /* empty buffer headers */
82 BUFFER_QUEUES /* number of buffer queues */
85 typedef enum bufq_type bufq_type_t;
87 #define BD_WAKE_SIZE 16384
88 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
90 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
91 static struct spinlock bufqspin = SPINLOCK_INITIALIZER(&bufqspin);
92 static struct spinlock bufcspin = SPINLOCK_INITIALIZER(&bufcspin);
94 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
96 struct buf *buf; /* buffer header pool */
98 static void vfs_clean_pages(struct buf *bp);
99 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
100 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m);
101 static void vfs_vmio_release(struct buf *bp);
102 static int flushbufqueues(bufq_type_t q);
103 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);
105 static void bd_signal(int totalspace);
106 static void buf_daemon(void);
107 static void buf_daemon_hw(void);
110 * bogus page -- for I/O to/from partially complete buffers
111 * this is a temporary solution to the problem, but it is not
112 * really that bad. it would be better to split the buffer
113 * for input in the case of buffers partially already in memory,
114 * but the code is intricate enough already.
116 vm_page_t bogus_page;
119 * These are all static, but make the ones we export globals so we do
120 * not need to use compiler magic.
122 int bufspace; /* locked by buffer_map */
124 static int bufmallocspace; /* atomic ops */
125 int maxbufmallocspace, lobufspace, hibufspace;
126 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
127 static int lorunningspace;
128 static int hirunningspace;
129 static int runningbufreq; /* locked by bufcspin */
130 static int dirtybufspace; /* locked by bufcspin */
131 static int dirtybufcount; /* locked by bufcspin */
132 static int dirtybufspacehw; /* locked by bufcspin */
133 static int dirtybufcounthw; /* locked by bufcspin */
134 static int runningbufspace; /* locked by bufcspin */
135 static int runningbufcount; /* locked by bufcspin */
138 static int getnewbufcalls;
139 static int getnewbufrestarts;
140 static int recoverbufcalls;
141 static int needsbuffer; /* locked by bufcspin */
142 static int bd_request; /* locked by bufcspin */
143 static int bd_request_hw; /* locked by bufcspin */
144 static u_int bd_wake_ary[BD_WAKE_SIZE];
145 static u_int bd_wake_index;
146 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */
147 static int debug_commit;
149 static struct thread *bufdaemon_td;
150 static struct thread *bufdaemonhw_td;
151 static u_int lowmempgallocs;
152 static u_int lowmempgfails;
155 * Sysctls for operational control of the buffer cache.
157 SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
158 "Number of dirty buffers to flush before bufdaemon becomes inactive");
159 SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
160 "High watermark used to trigger explicit flushing of dirty buffers");
161 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
162 "Minimum amount of buffer space required for active I/O");
163 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
164 "Maximum amount of buffer space to usable for active I/O");
165 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0,
166 "Page allocations done during periods of very low free memory");
167 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0,
168 "Page allocations which failed during periods of very low free memory");
169 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
170 "Recycle pages to active or inactive queue transition pt 0-64");
172 * Sysctls determining current state of the buffer cache.
174 SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
175 "Total number of buffers in buffer cache");
176 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
177 "Pending bytes of dirty buffers (all)");
178 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
179 "Pending bytes of dirty buffers (heavy weight)");
180 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
181 "Pending number of dirty buffers");
182 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
183 "Pending number of dirty buffers (heavy weight)");
184 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
185 "I/O bytes currently in progress due to asynchronous writes");
186 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
187 "I/O buffers currently in progress due to asynchronous writes");
188 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
189 "Hard limit on maximum amount of memory usable for buffer space");
190 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
191 "Soft limit on maximum amount of memory usable for buffer space");
192 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
193 "Minimum amount of memory to reserve for system buffer space");
194 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
195 "Amount of memory available for buffers");
196 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
197 0, "Maximum amount of memory reserved for buffers using malloc");
198 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
199 "Amount of memory left for buffers using malloc-scheme");
200 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
201 "New buffer header acquisition requests");
202 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
203 0, "New buffer header acquisition restarts");
204 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
205 "Recover VM space in an emergency");
206 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
207 "Buffer acquisition restarts due to fragmented buffer map");
208 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
209 "Amount of time KVA space was deallocated in an arbitrary buffer");
210 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
211 "Amount of time buffer re-use operations were successful");
212 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, "");
213 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
214 "sizeof(struct buf)");
216 char *buf_wmesg = BUF_WMESG;
218 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
219 #define VFS_BIO_NEED_UNUSED02 0x02
220 #define VFS_BIO_NEED_UNUSED04 0x04
221 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
226 * Called when buffer space is potentially available for recovery.
227 * getnewbuf() will block on this flag when it is unable to free
228 * sufficient buffer space. Buffer space becomes recoverable when
229 * bp's get placed back in the queues.
235 * If someone is waiting for BUF space, wake them up. Even
236 * though we haven't freed the kva space yet, the waiting
237 * process will be able to now.
239 spin_lock(&bufcspin);
240 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
241 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
242 spin_unlock(&bufcspin);
243 wakeup(&needsbuffer);
245 spin_unlock(&bufcspin);
252 * Accounting for I/O in progress.
256 runningbufwakeup(struct buf *bp)
261 if ((totalspace = bp->b_runningbufspace) != 0) {
262 spin_lock(&bufcspin);
263 runningbufspace -= totalspace;
265 bp->b_runningbufspace = 0;
268 * see waitrunningbufspace() for limit test.
270 limit = hirunningspace * 4 / 6;
271 if (runningbufreq && runningbufspace <= limit) {
273 spin_unlock(&bufcspin);
274 wakeup(&runningbufreq);
276 spin_unlock(&bufcspin);
278 bd_signal(totalspace);
285 * Called when a buffer has been added to one of the free queues to
286 * account for the buffer and to wakeup anyone waiting for free buffers.
287 * This typically occurs when large amounts of metadata are being handled
288 * by the buffer cache ( else buffer space runs out first, usually ).
295 spin_lock(&bufcspin);
297 needsbuffer &= ~VFS_BIO_NEED_ANY;
298 spin_unlock(&bufcspin);
299 wakeup(&needsbuffer);
301 spin_unlock(&bufcspin);
306 * waitrunningbufspace()
308 * Wait for the amount of running I/O to drop to hirunningspace * 4 / 6.
309 * This is the point where write bursting stops so we don't want to wait
310 * for the running amount to drop below it (at least if we still want bioq
313 * The caller may be using this function to block in a tight loop, we
314 * must block while runningbufspace is greater then or equal to
315 * hirunningspace * 4 / 6.
317 * And even with that it may not be enough, due to the presence of
318 * B_LOCKED dirty buffers, so also wait for at least one running buffer
322 waitrunningbufspace(void)
324 int limit = hirunningspace * 4 / 6;
327 spin_lock(&bufcspin);
328 if (runningbufspace > limit) {
329 while (runningbufspace > limit) {
331 ssleep(&runningbufreq, &bufcspin, 0, "wdrn1", 0);
333 spin_unlock(&bufcspin);
334 } else if (runningbufspace > limit / 2) {
336 spin_unlock(&bufcspin);
337 tsleep(&dummy, 0, "wdrn2", 1);
339 spin_unlock(&bufcspin);
344 * buf_dirty_count_severe:
346 * Return true if we have too many dirty buffers.
349 buf_dirty_count_severe(void)
351 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
352 dirtybufcount >= nbuf / 2);
356 * Return true if the amount of running I/O is severe and BIOQ should
360 buf_runningbufspace_severe(void)
362 return (runningbufspace >= hirunningspace * 4 / 6);
366 * vfs_buf_test_cache:
368 * Called when a buffer is extended. This function clears the B_CACHE
369 * bit if the newly extended portion of the buffer does not contain
372 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
373 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
374 * them while a clean buffer was present.
378 vfs_buf_test_cache(struct buf *bp,
379 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
382 if (bp->b_flags & B_CACHE) {
383 int base = (foff + off) & PAGE_MASK;
384 if (vm_page_is_valid(m, base, size) == 0)
385 bp->b_flags &= ~B_CACHE;
392 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
401 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
404 if (bd_request == 0 &&
405 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
406 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
407 spin_lock(&bufcspin);
409 spin_unlock(&bufcspin);
412 if (bd_request_hw == 0 &&
413 (dirtybufspacehw > lodirtybufspace / 2 ||
414 dirtybufcounthw >= nbuf / 2)) {
415 spin_lock(&bufcspin);
417 spin_unlock(&bufcspin);
418 wakeup(&bd_request_hw);
425 * Get the buf_daemon heated up when the number of running and dirty
426 * buffers exceeds the mid-point.
428 * Return the total number of dirty bytes past the second mid point
429 * as a measure of how much excess dirty data there is in the system.
440 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
442 totalspace = runningbufspace + dirtybufspace;
443 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
445 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
446 if (totalspace >= mid2)
447 return(totalspace - mid2);
455 * Wait for the buffer cache to flush (totalspace) bytes worth of
456 * buffers, then return.
458 * Regardless this function blocks while the number of dirty buffers
459 * exceeds hidirtybufspace.
464 bd_wait(int totalspace)
469 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
472 while (totalspace > 0) {
474 if (totalspace > runningbufspace + dirtybufspace)
475 totalspace = runningbufspace + dirtybufspace;
476 count = totalspace / BKVASIZE;
477 if (count >= BD_WAKE_SIZE)
478 count = BD_WAKE_SIZE - 1;
480 spin_lock(&bufcspin);
481 i = (bd_wake_index + count) & BD_WAKE_MASK;
485 * This is not a strict interlock, so we play a bit loose
486 * with locking access to dirtybufspace*
488 tsleep_interlock(&bd_wake_ary[i], 0);
489 spin_unlock(&bufcspin);
490 tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz);
492 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
499 * This function is called whenever runningbufspace or dirtybufspace
500 * is reduced. Track threads waiting for run+dirty buffer I/O
506 bd_signal(int totalspace)
510 if (totalspace > 0) {
511 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
512 totalspace = BKVASIZE * BD_WAKE_SIZE;
513 spin_lock(&bufcspin);
514 while (totalspace > 0) {
517 if (bd_wake_ary[i]) {
519 spin_unlock(&bufcspin);
520 wakeup(&bd_wake_ary[i]);
521 spin_lock(&bufcspin);
523 totalspace -= BKVASIZE;
525 spin_unlock(&bufcspin);
530 * BIO tracking support routines.
532 * Release a ref on a bio_track. Wakeup requests are atomically released
533 * along with the last reference so bk_active will never wind up set to
540 bio_track_rel(struct bio_track *track)
548 active = track->bk_active;
549 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
553 * Full-on. Note that the wait flag is only atomically released on
554 * the 1->0 count transition.
556 * We check for a negative count transition using bit 30 since bit 31
557 * has a different meaning.
560 desired = (active & 0x7FFFFFFF) - 1;
562 desired |= active & 0x80000000;
563 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
564 if (desired & 0x40000000)
565 panic("bio_track_rel: bad count: %p\n", track);
566 if (active & 0x80000000)
570 active = track->bk_active;
575 * Wait for the tracking count to reach 0.
577 * Use atomic ops such that the wait flag is only set atomically when
578 * bk_active is non-zero.
583 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
592 if (track->bk_active == 0)
596 * Full-on. Note that the wait flag may only be atomically set if
597 * the active count is non-zero.
599 * NOTE: We cannot optimize active == desired since a wakeup could
600 * clear active prior to our tsleep_interlock().
603 while ((active = track->bk_active) != 0) {
605 desired = active | 0x80000000;
606 tsleep_interlock(track, slp_flags);
607 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
608 error = tsleep(track, slp_flags | PINTERLOCKED,
620 * Load time initialisation of the buffer cache, called from machine
621 * dependant initialization code.
627 vm_offset_t bogus_offset;
630 /* next, make a null set of free lists */
631 for (i = 0; i < BUFFER_QUEUES; i++)
632 TAILQ_INIT(&bufqueues[i]);
634 /* finally, initialize each buffer header and stick on empty q */
635 for (i = 0; i < nbuf; i++) {
637 bzero(bp, sizeof *bp);
638 bp->b_flags = B_INVAL; /* we're just an empty header */
639 bp->b_cmd = BUF_CMD_DONE;
640 bp->b_qindex = BQUEUE_EMPTY;
642 xio_init(&bp->b_xio);
644 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
648 * maxbufspace is the absolute maximum amount of buffer space we are
649 * allowed to reserve in KVM and in real terms. The absolute maximum
650 * is nominally used by buf_daemon. hibufspace is the nominal maximum
651 * used by most other processes. The differential is required to
652 * ensure that buf_daemon is able to run when other processes might
653 * be blocked waiting for buffer space.
655 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
656 * this may result in KVM fragmentation which is not handled optimally
659 maxbufspace = nbuf * BKVASIZE;
660 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
661 lobufspace = hibufspace - MAXBSIZE;
663 lorunningspace = 512 * 1024;
664 /* hirunningspace -- see below */
667 * Limit the amount of malloc memory since it is wired permanently
668 * into the kernel space. Even though this is accounted for in
669 * the buffer allocation, we don't want the malloced region to grow
670 * uncontrolled. The malloc scheme improves memory utilization
671 * significantly on average (small) directories.
673 maxbufmallocspace = hibufspace / 20;
676 * Reduce the chance of a deadlock occuring by limiting the number
677 * of delayed-write dirty buffers we allow to stack up.
679 * We don't want too much actually queued to the device at once
680 * (XXX this needs to be per-mount!), because the buffers will
681 * wind up locked for a very long period of time while the I/O
684 hidirtybufspace = hibufspace / 2; /* dirty + running */
685 hirunningspace = hibufspace / 16; /* locked & queued to device */
686 if (hirunningspace < 1024 * 1024)
687 hirunningspace = 1024 * 1024;
692 lodirtybufspace = hidirtybufspace / 2;
695 * Maximum number of async ops initiated per buf_daemon loop. This is
696 * somewhat of a hack at the moment, we really need to limit ourselves
697 * based on the number of bytes of I/O in-transit that were initiated
701 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
702 vm_object_hold(&kernel_object);
703 bogus_page = vm_page_alloc(&kernel_object,
704 (bogus_offset >> PAGE_SHIFT),
706 vm_object_drop(&kernel_object);
707 vmstats.v_wire_count++;
712 * Initialize the embedded bio structures, typically used by
713 * deprecated code which tries to allocate its own struct bufs.
716 initbufbio(struct buf *bp)
718 bp->b_bio1.bio_buf = bp;
719 bp->b_bio1.bio_prev = NULL;
720 bp->b_bio1.bio_offset = NOOFFSET;
721 bp->b_bio1.bio_next = &bp->b_bio2;
722 bp->b_bio1.bio_done = NULL;
723 bp->b_bio1.bio_flags = 0;
725 bp->b_bio2.bio_buf = bp;
726 bp->b_bio2.bio_prev = &bp->b_bio1;
727 bp->b_bio2.bio_offset = NOOFFSET;
728 bp->b_bio2.bio_next = NULL;
729 bp->b_bio2.bio_done = NULL;
730 bp->b_bio2.bio_flags = 0;
736 * Reinitialize the embedded bio structures as well as any additional
737 * translation cache layers.
740 reinitbufbio(struct buf *bp)
744 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
745 bio->bio_done = NULL;
746 bio->bio_offset = NOOFFSET;
751 * Undo the effects of an initbufbio().
754 uninitbufbio(struct buf *bp)
761 * Push another BIO layer onto an existing BIO and return it. The new
762 * BIO layer may already exist, holding cached translation data.
765 push_bio(struct bio *bio)
769 if ((nbio = bio->bio_next) == NULL) {
770 int index = bio - &bio->bio_buf->b_bio_array[0];
771 if (index >= NBUF_BIO - 1) {
772 panic("push_bio: too many layers bp %p\n",
775 nbio = &bio->bio_buf->b_bio_array[index + 1];
776 bio->bio_next = nbio;
777 nbio->bio_prev = bio;
778 nbio->bio_buf = bio->bio_buf;
779 nbio->bio_offset = NOOFFSET;
780 nbio->bio_done = NULL;
781 nbio->bio_next = NULL;
783 KKASSERT(nbio->bio_done == NULL);
788 * Pop a BIO translation layer, returning the previous layer. The
789 * must have been previously pushed.
792 pop_bio(struct bio *bio)
794 return(bio->bio_prev);
798 clearbiocache(struct bio *bio)
801 bio->bio_offset = NOOFFSET;
809 * Free the KVA allocation for buffer 'bp'.
811 * Must be called from a critical section as this is the only locking for
814 * Since this call frees up buffer space, we call bufspacewakeup().
819 bfreekva(struct buf *bp)
825 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
826 vm_map_lock(&buffer_map);
827 bufspace -= bp->b_kvasize;
828 vm_map_delete(&buffer_map,
829 (vm_offset_t) bp->b_kvabase,
830 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
833 vm_map_unlock(&buffer_map);
834 vm_map_entry_release(count);
836 bp->b_kvabase = NULL;
844 * Remove the buffer from the appropriate free list.
847 _bremfree(struct buf *bp)
849 if (bp->b_qindex != BQUEUE_NONE) {
850 KASSERT(BUF_REFCNTNB(bp) == 1,
851 ("bremfree: bp %p not locked",bp));
852 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
853 bp->b_qindex = BQUEUE_NONE;
855 if (BUF_REFCNTNB(bp) <= 1)
856 panic("bremfree: removing a buffer not on a queue");
861 bremfree(struct buf *bp)
863 spin_lock(&bufqspin);
865 spin_unlock(&bufqspin);
869 bremfree_locked(struct buf *bp)
877 * Get a buffer with the specified data. Look in the cache first. We
878 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
879 * is set, the buffer is valid and we do not have to do anything ( see
884 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
886 return (breadn(vp, loffset, size, NULL, NULL, 0, bpp));
890 * This version of bread issues any required I/O asyncnronously and
891 * makes a callback on completion.
893 * The callback must check whether BIO_DONE is set in the bio and issue
894 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing
895 * BIO_DONE and disposing of the I/O (bqrelse()ing it).
898 breadcb(struct vnode *vp, off_t loffset, int size,
899 void (*func)(struct bio *), void *arg)
903 bp = getblk(vp, loffset, size, 0, 0);
905 /* if not found in cache, do some I/O */
906 if ((bp->b_flags & B_CACHE) == 0) {
907 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
908 bp->b_cmd = BUF_CMD_READ;
909 bp->b_bio1.bio_done = func;
910 bp->b_bio1.bio_caller_info1.ptr = arg;
911 vfs_busy_pages(vp, bp);
913 vn_strategy(vp, &bp->b_bio1);
916 * Since we are issuing the callback synchronously it cannot
917 * race the BIO_DONE, so no need for atomic ops here.
919 /*bp->b_bio1.bio_done = func;*/
920 bp->b_bio1.bio_caller_info1.ptr = arg;
921 bp->b_bio1.bio_flags |= BIO_DONE;
931 * Operates like bread, but also starts asynchronous I/O on
932 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
933 * to initiating I/O . If B_CACHE is set, the buffer is valid
934 * and we do not have to do anything.
938 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
939 int *rabsize, int cnt, struct buf **bpp)
941 struct buf *bp, *rabp;
943 int rv = 0, readwait = 0;
945 *bpp = bp = getblk(vp, loffset, size, 0, 0);
947 /* if not found in cache, do some I/O */
948 if ((bp->b_flags & B_CACHE) == 0) {
949 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
950 bp->b_cmd = BUF_CMD_READ;
951 bp->b_bio1.bio_done = biodone_sync;
952 bp->b_bio1.bio_flags |= BIO_SYNC;
953 vfs_busy_pages(vp, bp);
954 vn_strategy(vp, &bp->b_bio1);
958 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
959 if (inmem(vp, *raoffset))
961 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
963 if ((rabp->b_flags & B_CACHE) == 0) {
964 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
965 rabp->b_cmd = BUF_CMD_READ;
966 vfs_busy_pages(vp, rabp);
968 vn_strategy(vp, &rabp->b_bio1);
974 rv = biowait(&bp->b_bio1, "biord");
981 * Synchronous write, waits for completion.
983 * Write, release buffer on completion. (Done by iodone
984 * if async). Do not bother writing anything if the buffer
987 * Note that we set B_CACHE here, indicating that buffer is
988 * fully valid and thus cacheable. This is true even of NFS
989 * now so we set it generally. This could be set either here
990 * or in biodone() since the I/O is synchronous. We put it
994 bwrite(struct buf *bp)
998 if (bp->b_flags & B_INVAL) {
1002 if (BUF_REFCNTNB(bp) == 0)
1003 panic("bwrite: buffer is not busy???");
1005 /* Mark the buffer clean */
1008 bp->b_flags &= ~(B_ERROR | B_EINTR);
1009 bp->b_flags |= B_CACHE;
1010 bp->b_cmd = BUF_CMD_WRITE;
1011 bp->b_bio1.bio_done = biodone_sync;
1012 bp->b_bio1.bio_flags |= BIO_SYNC;
1013 vfs_busy_pages(bp->b_vp, bp);
1016 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1017 * valid for vnode-backed buffers.
1019 bsetrunningbufspace(bp, bp->b_bufsize);
1020 vn_strategy(bp->b_vp, &bp->b_bio1);
1021 error = biowait(&bp->b_bio1, "biows");
1030 * Asynchronous write. Start output on a buffer, but do not wait for
1031 * it to complete. The buffer is released when the output completes.
1033 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1034 * B_INVAL buffers. Not us.
1037 bawrite(struct buf *bp)
1039 if (bp->b_flags & B_INVAL) {
1043 if (BUF_REFCNTNB(bp) == 0)
1044 panic("bwrite: buffer is not busy???");
1046 /* Mark the buffer clean */
1049 bp->b_flags &= ~(B_ERROR | B_EINTR);
1050 bp->b_flags |= B_CACHE;
1051 bp->b_cmd = BUF_CMD_WRITE;
1052 KKASSERT(bp->b_bio1.bio_done == NULL);
1053 vfs_busy_pages(bp->b_vp, bp);
1056 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1057 * valid for vnode-backed buffers.
1059 bsetrunningbufspace(bp, bp->b_bufsize);
1061 vn_strategy(bp->b_vp, &bp->b_bio1);
1067 * Ordered write. Start output on a buffer, and flag it so that the
1068 * device will write it in the order it was queued. The buffer is
1069 * released when the output completes. bwrite() ( or the VOP routine
1070 * anyway ) is responsible for handling B_INVAL buffers.
1073 bowrite(struct buf *bp)
1075 bp->b_flags |= B_ORDERED;
1083 * Delayed write. (Buffer is marked dirty). Do not bother writing
1084 * anything if the buffer is marked invalid.
1086 * Note that since the buffer must be completely valid, we can safely
1087 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1088 * biodone() in order to prevent getblk from writing the buffer
1089 * out synchronously.
1092 bdwrite(struct buf *bp)
1094 if (BUF_REFCNTNB(bp) == 0)
1095 panic("bdwrite: buffer is not busy");
1097 if (bp->b_flags & B_INVAL) {
1103 if (dsched_is_clear_buf_priv(bp))
1107 * Set B_CACHE, indicating that the buffer is fully valid. This is
1108 * true even of NFS now.
1110 bp->b_flags |= B_CACHE;
1113 * This bmap keeps the system from needing to do the bmap later,
1114 * perhaps when the system is attempting to do a sync. Since it
1115 * is likely that the indirect block -- or whatever other datastructure
1116 * that the filesystem needs is still in memory now, it is a good
1117 * thing to do this. Note also, that if the pageout daemon is
1118 * requesting a sync -- there might not be enough memory to do
1119 * the bmap then... So, this is important to do.
1121 if (bp->b_bio2.bio_offset == NOOFFSET) {
1122 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1123 NULL, NULL, BUF_CMD_WRITE);
1127 * Because the underlying pages may still be mapped and
1128 * writable trying to set the dirty buffer (b_dirtyoff/end)
1129 * range here will be inaccurate.
1131 * However, we must still clean the pages to satisfy the
1132 * vnode_pager and pageout daemon, so theythink the pages
1133 * have been "cleaned". What has really occured is that
1134 * they've been earmarked for later writing by the buffer
1137 * So we get the b_dirtyoff/end update but will not actually
1138 * depend on it (NFS that is) until the pages are busied for
1141 vfs_clean_pages(bp);
1145 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1146 * due to the softdep code.
1151 * Fake write - return pages to VM system as dirty, leave the buffer clean.
1152 * This is used by tmpfs.
1154 * It is important for any VFS using this routine to NOT use it for
1155 * IO_SYNC or IO_ASYNC operations which occur when the system really
1156 * wants to flush VM pages to backing store.
1159 buwrite(struct buf *bp)
1165 * Only works for VMIO buffers. If the buffer is already
1166 * marked for delayed-write we can't avoid the bdwrite().
1168 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) {
1174 * Set valid & dirty.
1176 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1177 m = bp->b_xio.xio_pages[i];
1178 vfs_dirty_one_page(bp, i, m);
1186 * Turn buffer into delayed write request by marking it B_DELWRI.
1187 * B_RELBUF and B_NOCACHE must be cleared.
1189 * We reassign the buffer to itself to properly update it in the
1190 * dirty/clean lists.
1192 * Must be called from a critical section.
1193 * The buffer must be on BQUEUE_NONE.
1196 bdirty(struct buf *bp)
1198 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1199 if (bp->b_flags & B_NOCACHE) {
1200 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1201 bp->b_flags &= ~B_NOCACHE;
1203 if (bp->b_flags & B_INVAL) {
1204 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1206 bp->b_flags &= ~B_RELBUF;
1208 if ((bp->b_flags & B_DELWRI) == 0) {
1209 lwkt_gettoken(&bp->b_vp->v_token);
1210 bp->b_flags |= B_DELWRI;
1212 lwkt_reltoken(&bp->b_vp->v_token);
1214 spin_lock(&bufcspin);
1216 dirtybufspace += bp->b_bufsize;
1217 if (bp->b_flags & B_HEAVY) {
1219 dirtybufspacehw += bp->b_bufsize;
1221 spin_unlock(&bufcspin);
1228 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1229 * needs to be flushed with a different buf_daemon thread to avoid
1230 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1233 bheavy(struct buf *bp)
1235 if ((bp->b_flags & B_HEAVY) == 0) {
1236 bp->b_flags |= B_HEAVY;
1237 if (bp->b_flags & B_DELWRI) {
1238 spin_lock(&bufcspin);
1240 dirtybufspacehw += bp->b_bufsize;
1241 spin_unlock(&bufcspin);
1249 * Clear B_DELWRI for buffer.
1251 * Must be called from a critical section.
1253 * The buffer is typically on BQUEUE_NONE but there is one case in
1254 * brelse() that calls this function after placing the buffer on
1255 * a different queue.
1260 bundirty(struct buf *bp)
1262 if (bp->b_flags & B_DELWRI) {
1263 lwkt_gettoken(&bp->b_vp->v_token);
1264 bp->b_flags &= ~B_DELWRI;
1266 lwkt_reltoken(&bp->b_vp->v_token);
1268 spin_lock(&bufcspin);
1270 dirtybufspace -= bp->b_bufsize;
1271 if (bp->b_flags & B_HEAVY) {
1273 dirtybufspacehw -= bp->b_bufsize;
1275 spin_unlock(&bufcspin);
1277 bd_signal(bp->b_bufsize);
1280 * Since it is now being written, we can clear its deferred write flag.
1282 bp->b_flags &= ~B_DEFERRED;
1286 * Set the b_runningbufspace field, used to track how much I/O is
1287 * in progress at any given moment.
1290 bsetrunningbufspace(struct buf *bp, int bytes)
1292 bp->b_runningbufspace = bytes;
1294 spin_lock(&bufcspin);
1295 runningbufspace += bytes;
1297 spin_unlock(&bufcspin);
1304 * Release a busy buffer and, if requested, free its resources. The
1305 * buffer will be stashed in the appropriate bufqueue[] allowing it
1306 * to be accessed later as a cache entity or reused for other purposes.
1311 brelse(struct buf *bp)
1314 int saved_flags = bp->b_flags;
1317 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1320 * If B_NOCACHE is set we are being asked to destroy the buffer and
1321 * its backing store. Clear B_DELWRI.
1323 * B_NOCACHE is set in two cases: (1) when the caller really wants
1324 * to destroy the buffer and backing store and (2) when the caller
1325 * wants to destroy the buffer and backing store after a write
1328 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1332 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1334 * A re-dirtied buffer is only subject to destruction
1335 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1337 /* leave buffer intact */
1338 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1339 (bp->b_bufsize <= 0)) {
1341 * Either a failed read or we were asked to free or not
1342 * cache the buffer. This path is reached with B_DELWRI
1343 * set only if B_INVAL is already set. B_NOCACHE governs
1344 * backing store destruction.
1346 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1347 * buffer cannot be immediately freed.
1349 bp->b_flags |= B_INVAL;
1350 if (LIST_FIRST(&bp->b_dep) != NULL)
1352 if (bp->b_flags & B_DELWRI) {
1353 spin_lock(&bufcspin);
1355 dirtybufspace -= bp->b_bufsize;
1356 if (bp->b_flags & B_HEAVY) {
1358 dirtybufspacehw -= bp->b_bufsize;
1360 spin_unlock(&bufcspin);
1362 bd_signal(bp->b_bufsize);
1364 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1368 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set,
1369 * or if b_refs is non-zero.
1371 * If vfs_vmio_release() is called with either bit set, the
1372 * underlying pages may wind up getting freed causing a previous
1373 * write (bdwrite()) to get 'lost' because pages associated with
1374 * a B_DELWRI bp are marked clean. Pages associated with a
1375 * B_LOCKED buffer may be mapped by the filesystem.
1377 * If we want to release the buffer ourselves (rather then the
1378 * originator asking us to release it), give the originator a
1379 * chance to countermand the release by setting B_LOCKED.
1381 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1382 * if B_DELWRI is set.
1384 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1385 * on pages to return pages to the VM page queues.
1387 if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) {
1388 bp->b_flags &= ~B_RELBUF;
1389 } else if (vm_page_count_severe()) {
1390 if (LIST_FIRST(&bp->b_dep) != NULL)
1391 buf_deallocate(bp); /* can set B_LOCKED */
1392 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1393 bp->b_flags &= ~B_RELBUF;
1395 bp->b_flags |= B_RELBUF;
1399 * Make sure b_cmd is clear. It may have already been cleared by
1402 * At this point destroying the buffer is governed by the B_INVAL
1403 * or B_RELBUF flags.
1405 bp->b_cmd = BUF_CMD_DONE;
1406 dsched_exit_buf(bp);
1409 * VMIO buffer rundown. Make sure the VM page array is restored
1410 * after an I/O may have replaces some of the pages with bogus pages
1411 * in order to not destroy dirty pages in a fill-in read.
1413 * Note that due to the code above, if a buffer is marked B_DELWRI
1414 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1415 * B_INVAL may still be set, however.
1417 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1418 * but not the backing store. B_NOCACHE will destroy the backing
1421 * Note that dirty NFS buffers contain byte-granular write ranges
1422 * and should not be destroyed w/ B_INVAL even if the backing store
1425 if (bp->b_flags & B_VMIO) {
1427 * Rundown for VMIO buffers which are not dirty NFS buffers.
1439 * Get the base offset and length of the buffer. Note that
1440 * in the VMIO case if the buffer block size is not
1441 * page-aligned then b_data pointer may not be page-aligned.
1442 * But our b_xio.xio_pages array *IS* page aligned.
1444 * block sizes less then DEV_BSIZE (usually 512) are not
1445 * supported due to the page granularity bits (m->valid,
1446 * m->dirty, etc...).
1448 * See man buf(9) for more information
1451 resid = bp->b_bufsize;
1452 foff = bp->b_loffset;
1454 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1455 m = bp->b_xio.xio_pages[i];
1456 vm_page_flag_clear(m, PG_ZERO);
1458 * If we hit a bogus page, fixup *all* of them
1459 * now. Note that we left these pages wired
1460 * when we removed them so they had better exist,
1461 * and they cannot be ripped out from under us so
1462 * no critical section protection is necessary.
1464 if (m == bogus_page) {
1466 poff = OFF_TO_IDX(bp->b_loffset);
1468 vm_object_hold(obj);
1469 for (j = i; j < bp->b_xio.xio_npages; j++) {
1472 mtmp = bp->b_xio.xio_pages[j];
1473 if (mtmp == bogus_page) {
1474 mtmp = vm_page_lookup(obj, poff + j);
1476 panic("brelse: page missing");
1478 bp->b_xio.xio_pages[j] = mtmp;
1481 bp->b_flags &= ~B_HASBOGUS;
1482 vm_object_drop(obj);
1484 if ((bp->b_flags & B_INVAL) == 0) {
1485 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1486 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1488 m = bp->b_xio.xio_pages[i];
1492 * Invalidate the backing store if B_NOCACHE is set
1493 * (e.g. used with vinvalbuf()). If this is NFS
1494 * we impose a requirement that the block size be
1495 * a multiple of PAGE_SIZE and create a temporary
1496 * hack to basically invalidate the whole page. The
1497 * problem is that NFS uses really odd buffer sizes
1498 * especially when tracking piecemeal writes and
1499 * it also vinvalbuf()'s a lot, which would result
1500 * in only partial page validation and invalidation
1501 * here. If the file page is mmap()'d, however,
1502 * all the valid bits get set so after we invalidate
1503 * here we would end up with weird m->valid values
1504 * like 0xfc. nfs_getpages() can't handle this so
1505 * we clear all the valid bits for the NFS case
1506 * instead of just some of them.
1508 * The real bug is the VM system having to set m->valid
1509 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1510 * itself is an artifact of the whole 512-byte
1511 * granular mess that exists to support odd block
1512 * sizes and UFS meta-data block sizes (e.g. 6144).
1513 * A complete rewrite is required.
1517 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1518 int poffset = foff & PAGE_MASK;
1521 presid = PAGE_SIZE - poffset;
1522 if (bp->b_vp->v_tag == VT_NFS &&
1523 bp->b_vp->v_type == VREG) {
1525 } else if (presid > resid) {
1528 KASSERT(presid >= 0, ("brelse: extra page"));
1529 vm_page_set_invalid(m, poffset, presid);
1532 * Also make sure any swap cache is removed
1533 * as it is now stale (HAMMER in particular
1534 * uses B_NOCACHE to deal with buffer
1537 swap_pager_unswapped(m);
1539 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1540 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1542 if (bp->b_flags & (B_INVAL | B_RELBUF))
1543 vfs_vmio_release(bp);
1546 * Rundown for non-VMIO buffers.
1548 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1551 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1557 if (bp->b_qindex != BQUEUE_NONE)
1558 panic("brelse: free buffer onto another queue???");
1559 if (BUF_REFCNTNB(bp) > 1) {
1560 /* Temporary panic to verify exclusive locking */
1561 /* This panic goes away when we allow shared refs */
1562 panic("brelse: multiple refs");
1568 * Figure out the correct queue to place the cleaned up buffer on.
1569 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1570 * disassociated from their vnode.
1572 spin_lock(&bufqspin);
1573 if (bp->b_flags & B_LOCKED) {
1575 * Buffers that are locked are placed in the locked queue
1576 * immediately, regardless of their state.
1578 bp->b_qindex = BQUEUE_LOCKED;
1579 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1580 } else if (bp->b_bufsize == 0) {
1582 * Buffers with no memory. Due to conditionals near the top
1583 * of brelse() such buffers should probably already be
1584 * marked B_INVAL and disassociated from their vnode.
1586 bp->b_flags |= B_INVAL;
1587 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1588 KKASSERT((bp->b_flags & B_HASHED) == 0);
1589 if (bp->b_kvasize) {
1590 bp->b_qindex = BQUEUE_EMPTYKVA;
1592 bp->b_qindex = BQUEUE_EMPTY;
1594 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1595 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1597 * Buffers with junk contents. Again these buffers had better
1598 * already be disassociated from their vnode.
1600 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1601 KKASSERT((bp->b_flags & B_HASHED) == 0);
1602 bp->b_flags |= B_INVAL;
1603 bp->b_qindex = BQUEUE_CLEAN;
1604 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1607 * Remaining buffers. These buffers are still associated with
1610 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1612 bp->b_qindex = BQUEUE_DIRTY;
1613 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1615 case B_DELWRI | B_HEAVY:
1616 bp->b_qindex = BQUEUE_DIRTY_HW;
1617 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1622 * NOTE: Buffers are always placed at the end of the
1623 * queue. If B_AGE is not set the buffer will cycle
1624 * through the queue twice.
1626 bp->b_qindex = BQUEUE_CLEAN;
1627 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1631 spin_unlock(&bufqspin);
1634 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1635 * on the correct queue.
1637 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1641 * The bp is on an appropriate queue unless locked. If it is not
1642 * locked or dirty we can wakeup threads waiting for buffer space.
1644 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1645 * if B_INVAL is set ).
1647 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1651 * Something we can maybe free or reuse
1653 if (bp->b_bufsize || bp->b_kvasize)
1657 * Clean up temporary flags and unlock the buffer.
1659 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
1666 * Release a buffer back to the appropriate queue but do not try to free
1667 * it. The buffer is expected to be used again soon.
1669 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1670 * biodone() to requeue an async I/O on completion. It is also used when
1671 * known good buffers need to be requeued but we think we may need the data
1674 * XXX we should be able to leave the B_RELBUF hint set on completion.
1679 bqrelse(struct buf *bp)
1681 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1683 if (bp->b_qindex != BQUEUE_NONE)
1684 panic("bqrelse: free buffer onto another queue???");
1685 if (BUF_REFCNTNB(bp) > 1) {
1686 /* do not release to free list */
1687 panic("bqrelse: multiple refs");
1691 buf_act_advance(bp);
1693 spin_lock(&bufqspin);
1694 if (bp->b_flags & B_LOCKED) {
1696 * Locked buffers are released to the locked queue. However,
1697 * if the buffer is dirty it will first go into the dirty
1698 * queue and later on after the I/O completes successfully it
1699 * will be released to the locked queue.
1701 bp->b_qindex = BQUEUE_LOCKED;
1702 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1703 } else if (bp->b_flags & B_DELWRI) {
1704 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1705 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1706 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1707 } else if (vm_page_count_severe()) {
1709 * We are too low on memory, we have to try to free the
1710 * buffer (most importantly: the wired pages making up its
1711 * backing store) *now*.
1713 spin_unlock(&bufqspin);
1717 bp->b_qindex = BQUEUE_CLEAN;
1718 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1720 spin_unlock(&bufqspin);
1722 if ((bp->b_flags & B_LOCKED) == 0 &&
1723 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1728 * Something we can maybe free or reuse.
1730 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1734 * Final cleanup and unlock. Clear bits that are only used while a
1735 * buffer is actively locked.
1737 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
1738 dsched_exit_buf(bp);
1743 * Hold a buffer, preventing it from being reused. This will prevent
1744 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL
1745 * operations. If a B_INVAL operation occurs the buffer will remain held
1746 * but the underlying pages may get ripped out.
1748 * These functions are typically used in VOP_READ/VOP_WRITE functions
1749 * to hold a buffer during a copyin or copyout, preventing deadlocks
1750 * or recursive lock panics when read()/write() is used over mmap()'d
1753 * NOTE: bqhold() requires that the buffer be locked at the time of the
1754 * hold. bqdrop() has no requirements other than the buffer having
1755 * previously been held.
1758 bqhold(struct buf *bp)
1760 atomic_add_int(&bp->b_refs, 1);
1764 bqdrop(struct buf *bp)
1766 KKASSERT(bp->b_refs > 0);
1767 atomic_add_int(&bp->b_refs, -1);
1773 * Return backing pages held by the buffer 'bp' back to the VM system
1774 * if possible. The pages are freed if they are no longer valid or
1775 * attempt to free if it was used for direct I/O otherwise they are
1776 * sent to the page cache.
1778 * Pages that were marked busy are left alone and skipped.
1780 * The KVA mapping (b_data) for the underlying pages is removed by
1784 vfs_vmio_release(struct buf *bp)
1789 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1790 m = bp->b_xio.xio_pages[i];
1791 bp->b_xio.xio_pages[i] = NULL;
1793 vm_page_busy_wait(m, FALSE, "vmiopg");
1796 * The VFS is telling us this is not a meta-data buffer
1797 * even if it is backed by a block device.
1799 if (bp->b_flags & B_NOTMETA)
1800 vm_page_flag_set(m, PG_NOTMETA);
1803 * This is a very important bit of code. We try to track
1804 * VM page use whether the pages are wired into the buffer
1805 * cache or not. While wired into the buffer cache the
1806 * bp tracks the act_count.
1808 * We can choose to place unwired pages on the inactive
1809 * queue (0) or active queue (1). If we place too many
1810 * on the active queue the queue will cycle the act_count
1811 * on pages we'd like to keep, just from single-use pages
1812 * (such as when doing a tar-up or file scan).
1814 if (bp->b_act_count < vm_cycle_point)
1815 vm_page_unwire(m, 0);
1817 vm_page_unwire(m, 1);
1820 * We don't mess with busy pages, it is the responsibility
1821 * of the process that busied the pages to deal with them.
1823 * However, the caller may have marked the page invalid and
1824 * we must still make sure the page is no longer mapped.
1826 if ((m->flags & PG_BUSY) || (m->busy != 0)) {
1827 vm_page_protect(m, VM_PROT_NONE);
1832 if (m->wire_count == 0) {
1833 vm_page_flag_clear(m, PG_ZERO);
1835 * Might as well free the page if we can and it has
1836 * no valid data. We also free the page if the
1837 * buffer was used for direct I/O.
1840 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1841 m->hold_count == 0) {
1842 vm_page_protect(m, VM_PROT_NONE);
1847 * Cache the page if we are really low on free
1850 * Also bypass the active and inactive queues
1851 * if B_NOTMETA is set. This flag is set by HAMMER
1852 * on a regular file buffer when double buffering
1853 * is enabled or on a block device buffer representing
1854 * file data when double buffering is not enabled.
1855 * The flag prevents two copies of the same data from
1856 * being cached for long periods of time.
1858 if (bp->b_flags & B_DIRECT) {
1860 vm_page_try_to_free(m);
1861 } else if ((bp->b_flags & B_NOTMETA) ||
1862 vm_page_count_severe()) {
1863 m->act_count = bp->b_act_count;
1865 vm_page_try_to_cache(m);
1867 m->act_count = bp->b_act_count;
1875 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
1876 bp->b_xio.xio_npages);
1877 if (bp->b_bufsize) {
1881 bp->b_xio.xio_npages = 0;
1882 bp->b_flags &= ~B_VMIO;
1883 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1891 * Implement clustered async writes for clearing out B_DELWRI buffers.
1892 * This is much better then the old way of writing only one buffer at
1893 * a time. Note that we may not be presented with the buffers in the
1894 * correct order, so we search for the cluster in both directions.
1896 * The buffer is locked on call.
1899 vfs_bio_awrite(struct buf *bp)
1903 off_t loffset = bp->b_loffset;
1904 struct vnode *vp = bp->b_vp;
1911 * right now we support clustered writing only to regular files. If
1912 * we find a clusterable block we could be in the middle of a cluster
1913 * rather then at the beginning.
1915 * NOTE: b_bio1 contains the logical loffset and is aliased
1916 * to b_loffset. b_bio2 contains the translated block number.
1918 if ((vp->v_type == VREG) &&
1919 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1920 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1922 size = vp->v_mount->mnt_stat.f_iosize;
1924 for (i = size; i < MAXPHYS; i += size) {
1925 if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) &&
1926 BUF_REFCNT(bpa) == 0 &&
1927 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1928 (B_DELWRI | B_CLUSTEROK)) &&
1929 (bpa->b_bufsize == size)) {
1930 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1931 (bpa->b_bio2.bio_offset !=
1932 bp->b_bio2.bio_offset + i))
1938 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1939 if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) &&
1940 BUF_REFCNT(bpa) == 0 &&
1941 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1942 (B_DELWRI | B_CLUSTEROK)) &&
1943 (bpa->b_bufsize == size)) {
1944 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1945 (bpa->b_bio2.bio_offset !=
1946 bp->b_bio2.bio_offset - j))
1956 * this is a possible cluster write
1958 if (nbytes != size) {
1960 nwritten = cluster_wbuild(vp, size,
1961 loffset - j, nbytes);
1967 * default (old) behavior, writing out only one block
1969 * XXX returns b_bufsize instead of b_bcount for nwritten?
1971 nwritten = bp->b_bufsize;
1981 * Find and initialize a new buffer header, freeing up existing buffers
1982 * in the bufqueues as necessary. The new buffer is returned locked.
1984 * Important: B_INVAL is not set. If the caller wishes to throw the
1985 * buffer away, the caller must set B_INVAL prior to calling brelse().
1988 * We have insufficient buffer headers
1989 * We have insufficient buffer space
1990 * buffer_map is too fragmented ( space reservation fails )
1991 * If we have to flush dirty buffers ( but we try to avoid this )
1993 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1994 * Instead we ask the buf daemon to do it for us. We attempt to
1995 * avoid piecemeal wakeups of the pageout daemon.
2000 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
2006 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2007 static int flushingbufs;
2010 * We can't afford to block since we might be holding a vnode lock,
2011 * which may prevent system daemons from running. We deal with
2012 * low-memory situations by proactively returning memory and running
2013 * async I/O rather then sync I/O.
2017 --getnewbufrestarts;
2019 ++getnewbufrestarts;
2022 * Setup for scan. If we do not have enough free buffers,
2023 * we setup a degenerate case that immediately fails. Note
2024 * that if we are specially marked process, we are allowed to
2025 * dip into our reserves.
2027 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
2029 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
2030 * However, there are a number of cases (defragging, reusing, ...)
2031 * where we cannot backup.
2033 nqindex = BQUEUE_EMPTYKVA;
2034 spin_lock(&bufqspin);
2035 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
2039 * If no EMPTYKVA buffers and we are either
2040 * defragging or reusing, locate a CLEAN buffer
2041 * to free or reuse. If bufspace useage is low
2042 * skip this step so we can allocate a new buffer.
2044 if (defrag || bufspace >= lobufspace) {
2045 nqindex = BQUEUE_CLEAN;
2046 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2050 * If we could not find or were not allowed to reuse a
2051 * CLEAN buffer, check to see if it is ok to use an EMPTY
2052 * buffer. We can only use an EMPTY buffer if allocating
2053 * its KVA would not otherwise run us out of buffer space.
2055 if (nbp == NULL && defrag == 0 &&
2056 bufspace + maxsize < hibufspace) {
2057 nqindex = BQUEUE_EMPTY;
2058 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
2063 * Run scan, possibly freeing data and/or kva mappings on the fly
2066 * WARNING! bufqspin is held!
2068 while ((bp = nbp) != NULL) {
2069 int qindex = nqindex;
2071 nbp = TAILQ_NEXT(bp, b_freelist);
2074 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2075 * cycles through the queue twice before being selected.
2077 if (qindex == BQUEUE_CLEAN &&
2078 (bp->b_flags & B_AGE) == 0 && nbp) {
2079 bp->b_flags |= B_AGE;
2080 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
2081 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
2086 * Calculate next bp ( we can only use it if we do not block
2087 * or do other fancy things ).
2092 nqindex = BQUEUE_EMPTYKVA;
2093 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
2096 case BQUEUE_EMPTYKVA:
2097 nqindex = BQUEUE_CLEAN;
2098 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
2112 KASSERT(bp->b_qindex == qindex,
2113 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2116 * Note: we no longer distinguish between VMIO and non-VMIO
2119 KASSERT((bp->b_flags & B_DELWRI) == 0,
2120 ("delwri buffer %p found in queue %d", bp, qindex));
2123 * Do not try to reuse a buffer with a non-zero b_refs.
2124 * This is an unsynchronized test. A synchronized test
2125 * is also performed after we lock the buffer.
2131 * If we are defragging then we need a buffer with
2132 * b_kvasize != 0. XXX this situation should no longer
2133 * occur, if defrag is non-zero the buffer's b_kvasize
2134 * should also be non-zero at this point. XXX
2136 if (defrag && bp->b_kvasize == 0) {
2137 kprintf("Warning: defrag empty buffer %p\n", bp);
2142 * Start freeing the bp. This is somewhat involved. nbp
2143 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
2144 * on the clean list must be disassociated from their
2145 * current vnode. Buffers on the empty[kva] lists have
2146 * already been disassociated.
2148 * b_refs is checked after locking along with queue changes.
2149 * We must check here to deal with zero->nonzero transitions
2150 * made by the owner of the buffer lock, which is used by
2151 * VFS's to hold the buffer while issuing an unlocked
2152 * uiomove()s. We cannot invalidate the buffer's pages
2153 * for this case. Once we successfully lock a buffer the
2154 * only 0->1 transitions of b_refs will occur via findblk().
2156 * We must also check for queue changes after successful
2157 * locking as the current lock holder may dispose of the
2158 * buffer and change its queue.
2160 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2161 spin_unlock(&bufqspin);
2162 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100);
2165 if (bp->b_qindex != qindex || bp->b_refs) {
2166 spin_unlock(&bufqspin);
2170 bremfree_locked(bp);
2171 spin_unlock(&bufqspin);
2174 * Dependancies must be handled before we disassociate the
2177 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2178 * be immediately disassociated. HAMMER then becomes
2179 * responsible for releasing the buffer.
2181 * NOTE: bufqspin is UNLOCKED now.
2183 if (LIST_FIRST(&bp->b_dep) != NULL) {
2185 if (bp->b_flags & B_LOCKED) {
2189 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2192 if (qindex == BQUEUE_CLEAN) {
2193 if (bp->b_flags & B_VMIO)
2194 vfs_vmio_release(bp);
2200 * NOTE: nbp is now entirely invalid. We can only restart
2201 * the scan from this point on.
2203 * Get the rest of the buffer freed up. b_kva* is still
2204 * valid after this operation.
2206 KASSERT(bp->b_vp == NULL,
2207 ("bp3 %p flags %08x vnode %p qindex %d "
2208 "unexpectededly still associated!",
2209 bp, bp->b_flags, bp->b_vp, qindex));
2210 KKASSERT((bp->b_flags & B_HASHED) == 0);
2213 * critical section protection is not required when
2214 * scrapping a buffer's contents because it is already
2220 bp->b_flags = B_BNOCLIP;
2221 bp->b_cmd = BUF_CMD_DONE;
2226 bp->b_xio.xio_npages = 0;
2227 bp->b_dirtyoff = bp->b_dirtyend = 0;
2228 bp->b_act_count = ACT_INIT;
2230 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2232 if (blkflags & GETBLK_BHEAVY)
2233 bp->b_flags |= B_HEAVY;
2236 * If we are defragging then free the buffer.
2239 bp->b_flags |= B_INVAL;
2247 * If we are overcomitted then recover the buffer and its
2248 * KVM space. This occurs in rare situations when multiple
2249 * processes are blocked in getnewbuf() or allocbuf().
2251 if (bufspace >= hibufspace)
2253 if (flushingbufs && bp->b_kvasize != 0) {
2254 bp->b_flags |= B_INVAL;
2259 if (bufspace < lobufspace)
2263 * b_refs can transition to a non-zero value while we hold
2264 * the buffer locked due to a findblk(). Our brelvp() above
2265 * interlocked any future possible transitions due to
2268 * If we find b_refs to be non-zero we can destroy the
2269 * buffer's contents but we cannot yet reuse the buffer.
2272 bp->b_flags |= B_INVAL;
2278 /* NOT REACHED, bufqspin not held */
2282 * If we exhausted our list, sleep as appropriate. We may have to
2283 * wakeup various daemons and write out some dirty buffers.
2285 * Generally we are sleeping due to insufficient buffer space.
2287 * NOTE: bufqspin is held if bp is NULL, else it is not held.
2293 spin_unlock(&bufqspin);
2295 flags = VFS_BIO_NEED_BUFSPACE;
2297 } else if (bufspace >= hibufspace) {
2299 flags = VFS_BIO_NEED_BUFSPACE;
2302 flags = VFS_BIO_NEED_ANY;
2305 bd_speedup(); /* heeeelp */
2306 spin_lock(&bufcspin);
2307 needsbuffer |= flags;
2308 while (needsbuffer & flags) {
2309 if (ssleep(&needsbuffer, &bufcspin,
2310 slpflags, waitmsg, slptimeo)) {
2311 spin_unlock(&bufcspin);
2315 spin_unlock(&bufcspin);
2318 * We finally have a valid bp. We aren't quite out of the
2319 * woods, we still have to reserve kva space. In order
2320 * to keep fragmentation sane we only allocate kva in
2323 * (bufqspin is not held)
2325 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2327 if (maxsize != bp->b_kvasize) {
2328 vm_offset_t addr = 0;
2333 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2334 vm_map_lock(&buffer_map);
2336 if (vm_map_findspace(&buffer_map,
2337 vm_map_min(&buffer_map), maxsize,
2338 maxsize, 0, &addr)) {
2340 * Uh oh. Buffer map is too fragmented. We
2341 * must defragment the map.
2343 vm_map_unlock(&buffer_map);
2344 vm_map_entry_release(count);
2347 bp->b_flags |= B_INVAL;
2352 vm_map_insert(&buffer_map, &count,
2354 addr, addr + maxsize,
2356 VM_PROT_ALL, VM_PROT_ALL,
2359 bp->b_kvabase = (caddr_t) addr;
2360 bp->b_kvasize = maxsize;
2361 bufspace += bp->b_kvasize;
2364 vm_map_unlock(&buffer_map);
2365 vm_map_entry_release(count);
2367 bp->b_data = bp->b_kvabase;
2373 * This routine is called in an emergency to recover VM pages from the
2374 * buffer cache by cashing in clean buffers. The idea is to recover
2375 * enough pages to be able to satisfy a stuck bio_page_alloc().
2380 recoverbufpages(void)
2387 spin_lock(&bufqspin);
2388 while (bytes < MAXBSIZE) {
2389 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
2394 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2395 * cycles through the queue twice before being selected.
2397 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
2398 bp->b_flags |= B_AGE;
2399 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
2400 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
2408 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
2409 KKASSERT((bp->b_flags & B_DELWRI) == 0);
2412 * Start freeing the bp. This is somewhat involved.
2414 * Buffers on the clean list must be disassociated from
2415 * their current vnode
2418 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2419 kprintf("recoverbufpages: warning, locked buf %p, "
2422 ssleep(&bd_request, &bufqspin, 0, "gnbxxx", hz / 100);
2425 if (bp->b_qindex != BQUEUE_CLEAN) {
2426 kprintf("recoverbufpages: warning, BUF_LOCK blocked "
2427 "unexpectedly on buf %p index %d, race "
2433 bremfree_locked(bp);
2434 spin_unlock(&bufqspin);
2437 * Dependancies must be handled before we disassociate the
2440 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2441 * be immediately disassociated. HAMMER then becomes
2442 * responsible for releasing the buffer.
2444 if (LIST_FIRST(&bp->b_dep) != NULL) {
2446 if (bp->b_flags & B_LOCKED) {
2448 spin_lock(&bufqspin);
2451 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2454 bytes += bp->b_bufsize;
2456 if (bp->b_flags & B_VMIO) {
2457 bp->b_flags |= B_DIRECT; /* try to free pages */
2458 vfs_vmio_release(bp);
2463 KKASSERT(bp->b_vp == NULL);
2464 KKASSERT((bp->b_flags & B_HASHED) == 0);
2467 * critical section protection is not required when
2468 * scrapping a buffer's contents because it is already
2474 bp->b_flags = B_BNOCLIP;
2475 bp->b_cmd = BUF_CMD_DONE;
2480 bp->b_xio.xio_npages = 0;
2481 bp->b_dirtyoff = bp->b_dirtyend = 0;
2483 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2485 bp->b_flags |= B_INVAL;
2488 spin_lock(&bufqspin);
2490 spin_unlock(&bufqspin);
2497 * Buffer flushing daemon. Buffers are normally flushed by the
2498 * update daemon but if it cannot keep up this process starts to
2499 * take the load in an attempt to prevent getnewbuf() from blocking.
2501 * Once a flush is initiated it does not stop until the number
2502 * of buffers falls below lodirtybuffers, but we will wake up anyone
2503 * waiting at the mid-point.
2506 static struct kproc_desc buf_kp = {
2511 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2512 kproc_start, &buf_kp)
2514 static struct kproc_desc bufhw_kp = {
2519 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2520 kproc_start, &bufhw_kp)
2531 * This process needs to be suspended prior to shutdown sync.
2533 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2534 bufdaemon_td, SHUTDOWN_PRI_LAST);
2535 curthread->td_flags |= TDF_SYSTHREAD;
2538 * This process is allowed to take the buffer cache to the limit
2541 kproc_suspend_loop();
2544 * Do the flush as long as the number of dirty buffers
2545 * (including those running) exceeds lodirtybufspace.
2547 * When flushing limit running I/O to hirunningspace
2548 * Do the flush. Limit the amount of in-transit I/O we
2549 * allow to build up, otherwise we would completely saturate
2550 * the I/O system. Wakeup any waiting processes before we
2551 * normally would so they can run in parallel with our drain.
2553 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2554 * but because we split the operation into two threads we
2555 * have to cut it in half for each thread.
2557 waitrunningbufspace();
2558 limit = lodirtybufspace / 2;
2559 while (runningbufspace + dirtybufspace > limit ||
2560 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
2561 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2563 if (runningbufspace < hirunningspace)
2565 waitrunningbufspace();
2569 * We reached our low water mark, reset the
2570 * request and sleep until we are needed again.
2571 * The sleep is just so the suspend code works.
2573 spin_lock(&bufcspin);
2574 if (bd_request == 0)
2575 ssleep(&bd_request, &bufcspin, 0, "psleep", hz);
2577 spin_unlock(&bufcspin);
2590 * This process needs to be suspended prior to shutdown sync.
2592 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2593 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2594 curthread->td_flags |= TDF_SYSTHREAD;
2597 * This process is allowed to take the buffer cache to the limit
2600 kproc_suspend_loop();
2603 * Do the flush. Limit the amount of in-transit I/O we
2604 * allow to build up, otherwise we would completely saturate
2605 * the I/O system. Wakeup any waiting processes before we
2606 * normally would so they can run in parallel with our drain.
2608 * Once we decide to flush push the queued I/O up to
2609 * hirunningspace in order to trigger bursting by the bioq
2612 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2613 * but because we split the operation into two threads we
2614 * have to cut it in half for each thread.
2616 waitrunningbufspace();
2617 limit = lodirtybufspace / 2;
2618 while (runningbufspace + dirtybufspacehw > limit ||
2619 dirtybufcounthw >= nbuf / 2) {
2620 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2622 if (runningbufspace < hirunningspace)
2624 waitrunningbufspace();
2628 * We reached our low water mark, reset the
2629 * request and sleep until we are needed again.
2630 * The sleep is just so the suspend code works.
2632 spin_lock(&bufcspin);
2633 if (bd_request_hw == 0)
2634 ssleep(&bd_request_hw, &bufcspin, 0, "psleep", hz);
2636 spin_unlock(&bufcspin);
2643 * Try to flush a buffer in the dirty queue. We must be careful to
2644 * free up B_INVAL buffers instead of write them, which NFS is
2645 * particularly sensitive to.
2647 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2648 * that we really want to try to get the buffer out and reuse it
2649 * due to the write load on the machine.
2651 * We must lock the buffer in order to check its validity before we
2652 * can mess with its contents. bufqspin isn't enough.
2655 flushbufqueues(bufq_type_t q)
2661 spin_lock(&bufqspin);
2664 bp = TAILQ_FIRST(&bufqueues[q]);
2666 if ((bp->b_flags & B_DELWRI) == 0) {
2667 kprintf("Unexpected clean buffer %p\n", bp);
2668 bp = TAILQ_NEXT(bp, b_freelist);
2671 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2672 bp = TAILQ_NEXT(bp, b_freelist);
2675 KKASSERT(bp->b_qindex == q);
2678 * Must recheck B_DELWRI after successfully locking
2681 if ((bp->b_flags & B_DELWRI) == 0) {
2683 bp = TAILQ_NEXT(bp, b_freelist);
2687 if (bp->b_flags & B_INVAL) {
2689 spin_unlock(&bufqspin);
2696 spin_unlock(&bufqspin);
2699 if (LIST_FIRST(&bp->b_dep) != NULL &&
2700 (bp->b_flags & B_DEFERRED) == 0 &&
2701 buf_countdeps(bp, 0)) {
2702 spin_lock(&bufqspin);
2704 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2705 TAILQ_INSERT_TAIL(&bufqueues[q], bp, b_freelist);
2706 bp->b_flags |= B_DEFERRED;
2708 bp = TAILQ_FIRST(&bufqueues[q]);
2713 * If the buffer has a dependancy, buf_checkwrite() must
2714 * also return 0 for us to be able to initate the write.
2716 * If the buffer is flagged B_ERROR it may be requeued
2717 * over and over again, we try to avoid a live lock.
2719 * NOTE: buf_checkwrite is MPSAFE.
2721 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) {
2724 } else if (bp->b_flags & B_ERROR) {
2725 tsleep(bp, 0, "bioer", 1);
2726 bp->b_flags &= ~B_AGE;
2729 bp->b_flags |= B_AGE;
2736 spin_unlock(&bufqspin);
2743 * Returns true if no I/O is needed to access the associated VM object.
2744 * This is like findblk except it also hunts around in the VM system for
2747 * Note that we ignore vm_page_free() races from interrupts against our
2748 * lookup, since if the caller is not protected our return value will not
2749 * be any more valid then otherwise once we exit the critical section.
2752 inmem(struct vnode *vp, off_t loffset)
2755 vm_offset_t toff, tinc, size;
2759 if (findblk(vp, loffset, FINDBLK_TEST))
2761 if (vp->v_mount == NULL)
2763 if ((obj = vp->v_object) == NULL)
2767 if (size > vp->v_mount->mnt_stat.f_iosize)
2768 size = vp->v_mount->mnt_stat.f_iosize;
2770 vm_object_hold(obj);
2771 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2772 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2778 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2779 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2780 if (vm_page_is_valid(m,
2781 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) {
2786 vm_object_drop(obj);
2793 * Locate and return the specified buffer. Unless flagged otherwise,
2794 * a locked buffer will be returned if it exists or NULL if it does not.
2796 * findblk()'d buffers are still on the bufqueues and if you intend
2797 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2798 * and possibly do other stuff to it.
2800 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2801 * for locking the buffer and ensuring that it remains
2802 * the desired buffer after locking.
2804 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2805 * to acquire the lock we return NULL, even if the
2808 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal
2809 * reuse by getnewbuf() but does not prevent
2810 * disassociation (B_INVAL). Used to avoid deadlocks
2811 * against random (vp,loffset)s due to reassignment.
2813 * (0) - Lock the buffer blocking.
2818 findblk(struct vnode *vp, off_t loffset, int flags)
2823 lkflags = LK_EXCLUSIVE;
2824 if (flags & FINDBLK_NBLOCK)
2825 lkflags |= LK_NOWAIT;
2829 * Lookup. Ref the buf while holding v_token to prevent
2830 * reuse (but does not prevent diassociation).
2832 lwkt_gettoken(&vp->v_token);
2833 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2835 lwkt_reltoken(&vp->v_token);
2839 lwkt_reltoken(&vp->v_token);
2842 * If testing only break and return bp, do not lock.
2844 if (flags & FINDBLK_TEST)
2848 * Lock the buffer, return an error if the lock fails.
2849 * (only FINDBLK_NBLOCK can cause the lock to fail).
2851 if (BUF_LOCK(bp, lkflags)) {
2852 atomic_subtract_int(&bp->b_refs, 1);
2853 /* bp = NULL; not needed */
2858 * Revalidate the locked buf before allowing it to be
2861 if (bp->b_vp == vp && bp->b_loffset == loffset)
2863 atomic_subtract_int(&bp->b_refs, 1);
2870 if ((flags & FINDBLK_REF) == 0)
2871 atomic_subtract_int(&bp->b_refs, 1);
2878 * Similar to getblk() except only returns the buffer if it is
2879 * B_CACHE and requires no other manipulation. Otherwise NULL
2882 * If B_RAM is set the buffer might be just fine, but we return
2883 * NULL anyway because we want the code to fall through to the
2884 * cluster read. Otherwise read-ahead breaks.
2886 * If blksize is 0 the buffer cache buffer must already be fully
2889 * If blksize is non-zero getblk() will be used, allowing a buffer
2890 * to be reinstantiated from its VM backing store. The buffer must
2891 * still be fully cached after reinstantiation to be returned.
2894 getcacheblk(struct vnode *vp, off_t loffset, int blksize)
2899 bp = getblk(vp, loffset, blksize, 0, 0);
2901 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2903 bp->b_flags &= ~B_AGE;
2910 bp = findblk(vp, loffset, 0);
2912 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2914 bp->b_flags &= ~B_AGE;
2928 * Get a block given a specified block and offset into a file/device.
2929 * B_INVAL may or may not be set on return. The caller should clear
2930 * B_INVAL prior to initiating a READ.
2932 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2933 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2934 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2935 * without doing any of those things the system will likely believe
2936 * the buffer to be valid (especially if it is not B_VMIO), and the
2937 * next getblk() will return the buffer with B_CACHE set.
2939 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2940 * an existing buffer.
2942 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2943 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2944 * and then cleared based on the backing VM. If the previous buffer is
2945 * non-0-sized but invalid, B_CACHE will be cleared.
2947 * If getblk() must create a new buffer, the new buffer is returned with
2948 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2949 * case it is returned with B_INVAL clear and B_CACHE set based on the
2952 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2953 * B_CACHE bit is clear.
2955 * What this means, basically, is that the caller should use B_CACHE to
2956 * determine whether the buffer is fully valid or not and should clear
2957 * B_INVAL prior to issuing a read. If the caller intends to validate
2958 * the buffer by loading its data area with something, the caller needs
2959 * to clear B_INVAL. If the caller does this without issuing an I/O,
2960 * the caller should set B_CACHE ( as an optimization ), else the caller
2961 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2962 * a write attempt or if it was a successfull read. If the caller
2963 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2964 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2968 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2969 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2974 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2977 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2981 if (size > MAXBSIZE)
2982 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2983 if (vp->v_object == NULL)
2984 panic("getblk: vnode %p has no object!", vp);
2987 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) {
2989 * The buffer was found in the cache, but we need to lock it.
2990 * We must acquire a ref on the bp to prevent reuse, but
2991 * this will not prevent disassociation (brelvp()) so we
2992 * must recheck (vp,loffset) after acquiring the lock.
2994 * Without the ref the buffer could potentially be reused
2995 * before we acquire the lock and create a deadlock
2996 * situation between the thread trying to reuse the buffer
2997 * and us due to the fact that we would wind up blocking
2998 * on a random (vp,loffset).
3000 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
3001 if (blkflags & GETBLK_NOWAIT) {
3005 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
3006 if (blkflags & GETBLK_PCATCH)
3007 lkflags |= LK_PCATCH;
3008 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
3011 if (error == ENOLCK)
3015 /* buffer may have changed on us */
3020 * Once the buffer has been locked, make sure we didn't race
3021 * a buffer recyclement. Buffers that are no longer hashed
3022 * will have b_vp == NULL, so this takes care of that check
3025 if (bp->b_vp != vp || bp->b_loffset != loffset) {
3026 kprintf("Warning buffer %p (vp %p loffset %lld) "
3028 bp, vp, (long long)loffset);
3034 * If SZMATCH any pre-existing buffer must be of the requested
3035 * size or NULL is returned. The caller absolutely does not
3036 * want getblk() to bwrite() the buffer on a size mismatch.
3038 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
3044 * All vnode-based buffers must be backed by a VM object.
3046 KKASSERT(bp->b_flags & B_VMIO);
3047 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3048 bp->b_flags &= ~B_AGE;
3051 * Make sure that B_INVAL buffers do not have a cached
3052 * block number translation.
3054 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
3055 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
3056 " did not have cleared bio_offset cache\n",
3057 bp, vp, (long long)loffset);
3058 clearbiocache(&bp->b_bio2);
3062 * The buffer is locked. B_CACHE is cleared if the buffer is
3065 if (bp->b_flags & B_INVAL)
3066 bp->b_flags &= ~B_CACHE;
3070 * Any size inconsistancy with a dirty buffer or a buffer
3071 * with a softupdates dependancy must be resolved. Resizing
3072 * the buffer in such circumstances can lead to problems.
3074 * Dirty or dependant buffers are written synchronously.
3075 * Other types of buffers are simply released and
3076 * reconstituted as they may be backed by valid, dirty VM
3077 * pages (but not marked B_DELWRI).
3079 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
3080 * and may be left over from a prior truncation (and thus
3081 * no longer represent the actual EOF point), so we
3082 * definitely do not want to B_NOCACHE the backing store.
3084 if (size != bp->b_bcount) {
3085 if (bp->b_flags & B_DELWRI) {
3086 bp->b_flags |= B_RELBUF;
3088 } else if (LIST_FIRST(&bp->b_dep)) {
3089 bp->b_flags |= B_RELBUF;
3092 bp->b_flags |= B_RELBUF;
3097 KKASSERT(size <= bp->b_kvasize);
3098 KASSERT(bp->b_loffset != NOOFFSET,
3099 ("getblk: no buffer offset"));
3102 * A buffer with B_DELWRI set and B_CACHE clear must
3103 * be committed before we can return the buffer in
3104 * order to prevent the caller from issuing a read
3105 * ( due to B_CACHE not being set ) and overwriting
3108 * Most callers, including NFS and FFS, need this to
3109 * operate properly either because they assume they
3110 * can issue a read if B_CACHE is not set, or because
3111 * ( for example ) an uncached B_DELWRI might loop due
3112 * to softupdates re-dirtying the buffer. In the latter
3113 * case, B_CACHE is set after the first write completes,
3114 * preventing further loops.
3116 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
3117 * above while extending the buffer, we cannot allow the
3118 * buffer to remain with B_CACHE set after the write
3119 * completes or it will represent a corrupt state. To
3120 * deal with this we set B_NOCACHE to scrap the buffer
3123 * XXX Should this be B_RELBUF instead of B_NOCACHE?
3124 * I'm not even sure this state is still possible
3125 * now that getblk() writes out any dirty buffers
3128 * We might be able to do something fancy, like setting
3129 * B_CACHE in bwrite() except if B_DELWRI is already set,
3130 * so the below call doesn't set B_CACHE, but that gets real
3131 * confusing. This is much easier.
3134 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3135 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
3136 "and CACHE clear, b_flags %08x\n",
3137 bp, (intmax_t)bp->b_loffset, bp->b_flags);
3138 bp->b_flags |= B_NOCACHE;
3144 * Buffer is not in-core, create new buffer. The buffer
3145 * returned by getnewbuf() is locked. Note that the returned
3146 * buffer is also considered valid (not marked B_INVAL).
3148 * Calculating the offset for the I/O requires figuring out
3149 * the block size. We use DEV_BSIZE for VBLK or VCHR and
3150 * the mount's f_iosize otherwise. If the vnode does not
3151 * have an associated mount we assume that the passed size is
3154 * Note that vn_isdisk() cannot be used here since it may
3155 * return a failure for numerous reasons. Note that the
3156 * buffer size may be larger then the block size (the caller
3157 * will use block numbers with the proper multiple). Beware
3158 * of using any v_* fields which are part of unions. In
3159 * particular, in DragonFly the mount point overloading
3160 * mechanism uses the namecache only and the underlying
3161 * directory vnode is not a special case.
3165 if (vp->v_type == VBLK || vp->v_type == VCHR)
3167 else if (vp->v_mount)
3168 bsize = vp->v_mount->mnt_stat.f_iosize;
3172 maxsize = size + (loffset & PAGE_MASK);
3173 maxsize = imax(maxsize, bsize);
3175 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
3177 if (slpflags || slptimeo)
3183 * Atomically insert the buffer into the hash, so that it can
3184 * be found by findblk().
3186 * If bgetvp() returns non-zero a collision occured, and the
3187 * bp will not be associated with the vnode.
3189 * Make sure the translation layer has been cleared.
3191 bp->b_loffset = loffset;
3192 bp->b_bio2.bio_offset = NOOFFSET;
3193 /* bp->b_bio2.bio_next = NULL; */
3195 if (bgetvp(vp, bp, size)) {
3196 bp->b_flags |= B_INVAL;
3202 * All vnode-based buffers must be backed by a VM object.
3204 KKASSERT(vp->v_object != NULL);
3205 bp->b_flags |= B_VMIO;
3206 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3210 KKASSERT(dsched_is_clear_buf_priv(bp));
3217 * Reacquire a buffer that was previously released to the locked queue,
3218 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
3219 * set B_LOCKED (which handles the acquisition race).
3221 * To this end, either B_LOCKED must be set or the dependancy list must be
3227 regetblk(struct buf *bp)
3229 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
3230 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
3237 * Get an empty, disassociated buffer of given size. The buffer is
3238 * initially set to B_INVAL.
3240 * critical section protection is not required for the allocbuf()
3241 * call because races are impossible here.
3251 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3253 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
3256 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3257 KKASSERT(dsched_is_clear_buf_priv(bp));
3265 * This code constitutes the buffer memory from either anonymous system
3266 * memory (in the case of non-VMIO operations) or from an associated
3267 * VM object (in the case of VMIO operations). This code is able to
3268 * resize a buffer up or down.
3270 * Note that this code is tricky, and has many complications to resolve
3271 * deadlock or inconsistant data situations. Tread lightly!!!
3272 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3273 * the caller. Calling this code willy nilly can result in the loss of
3276 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3277 * B_CACHE for the non-VMIO case.
3279 * This routine does not need to be called from a critical section but you
3280 * must own the buffer.
3285 allocbuf(struct buf *bp, int size)
3287 int newbsize, mbsize;
3290 if (BUF_REFCNT(bp) == 0)
3291 panic("allocbuf: buffer not busy");
3293 if (bp->b_kvasize < size)
3294 panic("allocbuf: buffer too small");
3296 if ((bp->b_flags & B_VMIO) == 0) {
3300 * Just get anonymous memory from the kernel. Don't
3301 * mess with B_CACHE.
3303 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3304 if (bp->b_flags & B_MALLOC)
3307 newbsize = round_page(size);
3309 if (newbsize < bp->b_bufsize) {
3311 * Malloced buffers are not shrunk
3313 if (bp->b_flags & B_MALLOC) {
3315 bp->b_bcount = size;
3317 kfree(bp->b_data, M_BIOBUF);
3318 if (bp->b_bufsize) {
3319 atomic_subtract_int(&bufmallocspace, bp->b_bufsize);
3323 bp->b_data = bp->b_kvabase;
3325 bp->b_flags &= ~B_MALLOC;
3331 (vm_offset_t) bp->b_data + newbsize,
3332 (vm_offset_t) bp->b_data + bp->b_bufsize);
3333 } else if (newbsize > bp->b_bufsize) {
3335 * We only use malloced memory on the first allocation.
3336 * and revert to page-allocated memory when the buffer
3339 if ((bufmallocspace < maxbufmallocspace) &&
3340 (bp->b_bufsize == 0) &&
3341 (mbsize <= PAGE_SIZE/2)) {
3343 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3344 bp->b_bufsize = mbsize;
3345 bp->b_bcount = size;
3346 bp->b_flags |= B_MALLOC;
3347 atomic_add_int(&bufmallocspace, mbsize);
3353 * If the buffer is growing on its other-than-first
3354 * allocation, then we revert to the page-allocation
3357 if (bp->b_flags & B_MALLOC) {
3358 origbuf = bp->b_data;
3359 origbufsize = bp->b_bufsize;
3360 bp->b_data = bp->b_kvabase;
3361 if (bp->b_bufsize) {
3362 atomic_subtract_int(&bufmallocspace,
3367 bp->b_flags &= ~B_MALLOC;
3368 newbsize = round_page(newbsize);
3372 (vm_offset_t) bp->b_data + bp->b_bufsize,
3373 (vm_offset_t) bp->b_data + newbsize);
3375 bcopy(origbuf, bp->b_data, origbufsize);
3376 kfree(origbuf, M_BIOBUF);
3383 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3384 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3385 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3386 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3388 if (bp->b_flags & B_MALLOC)
3389 panic("allocbuf: VMIO buffer can't be malloced");
3391 * Set B_CACHE initially if buffer is 0 length or will become
3394 if (size == 0 || bp->b_bufsize == 0)
3395 bp->b_flags |= B_CACHE;
3397 if (newbsize < bp->b_bufsize) {
3399 * DEV_BSIZE aligned new buffer size is less then the
3400 * DEV_BSIZE aligned existing buffer size. Figure out
3401 * if we have to remove any pages.
3403 if (desiredpages < bp->b_xio.xio_npages) {
3404 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3406 * the page is not freed here -- it
3407 * is the responsibility of
3408 * vnode_pager_setsize
3410 m = bp->b_xio.xio_pages[i];
3411 KASSERT(m != bogus_page,
3412 ("allocbuf: bogus page found"));
3413 vm_page_busy_wait(m, TRUE, "biodep");
3414 bp->b_xio.xio_pages[i] = NULL;
3415 vm_page_unwire(m, 0);
3418 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3419 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3420 bp->b_xio.xio_npages = desiredpages;
3422 } else if (size > bp->b_bcount) {
3424 * We are growing the buffer, possibly in a
3425 * byte-granular fashion.
3433 * Step 1, bring in the VM pages from the object,
3434 * allocating them if necessary. We must clear
3435 * B_CACHE if these pages are not valid for the
3436 * range covered by the buffer.
3438 * critical section protection is required to protect
3439 * against interrupts unbusying and freeing pages
3440 * between our vm_page_lookup() and our
3441 * busycheck/wiring call.
3446 vm_object_hold(obj);
3447 while (bp->b_xio.xio_npages < desiredpages) {
3452 pi = OFF_TO_IDX(bp->b_loffset) +
3453 bp->b_xio.xio_npages;
3456 * Blocking on m->busy might lead to a
3459 * vm_fault->getpages->cluster_read->allocbuf
3461 m = vm_page_lookup_busy_try(obj, pi, FALSE,
3464 vm_page_sleep_busy(m, FALSE, "pgtblk");
3469 * note: must allocate system pages
3470 * since blocking here could intefere
3471 * with paging I/O, no matter which
3474 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
3477 vm_page_flag_clear(m, PG_ZERO);
3479 bp->b_flags &= ~B_CACHE;
3480 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3481 ++bp->b_xio.xio_npages;
3487 * We found a page and were able to busy it.
3489 vm_page_flag_clear(m, PG_ZERO);
3492 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3493 ++bp->b_xio.xio_npages;
3494 if (bp->b_act_count < m->act_count)
3495 bp->b_act_count = m->act_count;
3497 vm_object_drop(obj);
3500 * Step 2. We've loaded the pages into the buffer,
3501 * we have to figure out if we can still have B_CACHE
3502 * set. Note that B_CACHE is set according to the
3503 * byte-granular range ( bcount and size ), not the
3504 * aligned range ( newbsize ).
3506 * The VM test is against m->valid, which is DEV_BSIZE
3507 * aligned. Needless to say, the validity of the data
3508 * needs to also be DEV_BSIZE aligned. Note that this
3509 * fails with NFS if the server or some other client
3510 * extends the file's EOF. If our buffer is resized,
3511 * B_CACHE may remain set! XXX
3514 toff = bp->b_bcount;
3515 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3517 while ((bp->b_flags & B_CACHE) && toff < size) {
3520 if (tinc > (size - toff))
3523 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3531 bp->b_xio.xio_pages[pi]
3538 * Step 3, fixup the KVM pmap. Remember that
3539 * bp->b_data is relative to bp->b_loffset, but
3540 * bp->b_loffset may be offset into the first page.
3543 bp->b_data = (caddr_t)
3544 trunc_page((vm_offset_t)bp->b_data);
3546 (vm_offset_t)bp->b_data,
3547 bp->b_xio.xio_pages,
3548 bp->b_xio.xio_npages
3550 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3551 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3555 /* adjust space use on already-dirty buffer */
3556 if (bp->b_flags & B_DELWRI) {
3557 spin_lock(&bufcspin);
3558 dirtybufspace += newbsize - bp->b_bufsize;
3559 if (bp->b_flags & B_HEAVY)
3560 dirtybufspacehw += newbsize - bp->b_bufsize;
3561 spin_unlock(&bufcspin);
3563 if (newbsize < bp->b_bufsize)
3565 bp->b_bufsize = newbsize; /* actual buffer allocation */
3566 bp->b_bcount = size; /* requested buffer size */
3573 * Wait for buffer I/O completion, returning error status. B_EINTR
3574 * is converted into an EINTR error but not cleared (since a chain
3575 * of biowait() calls may occur).
3577 * On return bpdone() will have been called but the buffer will remain
3578 * locked and will not have been brelse()'d.
3580 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3581 * likely still in progress on return.
3583 * NOTE! This operation is on a BIO, not a BUF.
3585 * NOTE! BIO_DONE is cleared by vn_strategy()
3590 _biowait(struct bio *bio, const char *wmesg, int to)
3592 struct buf *bp = bio->bio_buf;
3597 KKASSERT(bio == &bp->b_bio1);
3599 flags = bio->bio_flags;
3600 if (flags & BIO_DONE)
3602 nflags = flags | BIO_WANT;
3603 tsleep_interlock(bio, 0);
3604 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3606 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3607 else if (bp->b_cmd == BUF_CMD_READ)
3608 error = tsleep(bio, PINTERLOCKED, "biord", to);
3610 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3612 kprintf("tsleep error biowait %d\n", error);
3621 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3622 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3623 if (bp->b_flags & B_EINTR)
3625 if (bp->b_flags & B_ERROR)
3626 return (bp->b_error ? bp->b_error : EIO);
3631 biowait(struct bio *bio, const char *wmesg)
3633 return(_biowait(bio, wmesg, 0));
3637 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3639 return(_biowait(bio, wmesg, to));
3643 * This associates a tracking count with an I/O. vn_strategy() and
3644 * dev_dstrategy() do this automatically but there are a few cases
3645 * where a vnode or device layer is bypassed when a block translation
3646 * is cached. In such cases bio_start_transaction() may be called on
3647 * the bypassed layers so the system gets an I/O in progress indication
3648 * for those higher layers.
3651 bio_start_transaction(struct bio *bio, struct bio_track *track)
3653 bio->bio_track = track;
3654 if (dsched_is_clear_buf_priv(bio->bio_buf))
3655 dsched_new_buf(bio->bio_buf);
3656 bio_track_ref(track);
3660 * Initiate I/O on a vnode.
3662 * SWAPCACHE OPERATION:
3664 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately
3665 * devfs also uses b_vp for fake buffers so we also have to check
3666 * that B_PAGING is 0. In this case the passed 'vp' is probably the
3667 * underlying block device. The swap assignments are related to the
3668 * buffer cache buffer's b_vp, not the passed vp.
3670 * The passed vp == bp->b_vp only in the case where the strategy call
3671 * is made on the vp itself for its own buffers (a regular file or
3672 * block device vp). The filesystem usually then re-calls vn_strategy()
3673 * after translating the request to an underlying device.
3675 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the
3676 * underlying buffer cache buffers.
3678 * We can only deal with page-aligned buffers at the moment, because
3679 * we can't tell what the real dirty state for pages straddling a buffer
3682 * In order to call swap_pager_strategy() we must provide the VM object
3683 * and base offset for the underlying buffer cache pages so it can find
3687 vn_strategy(struct vnode *vp, struct bio *bio)
3689 struct bio_track *track;
3690 struct buf *bp = bio->bio_buf;
3692 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3695 * Set when an I/O is issued on the bp. Cleared by consumers
3696 * (aka HAMMER), allowing the consumer to determine if I/O had
3697 * actually occurred.
3699 bp->b_flags |= B_IODEBUG;
3702 * Handle the swap cache intercept.
3704 if (vn_cache_strategy(vp, bio))
3708 * Otherwise do the operation through the filesystem
3710 if (bp->b_cmd == BUF_CMD_READ)
3711 track = &vp->v_track_read;
3713 track = &vp->v_track_write;
3714 KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3715 bio->bio_track = track;
3716 if (dsched_is_clear_buf_priv(bio->bio_buf))
3717 dsched_new_buf(bio->bio_buf);
3718 bio_track_ref(track);
3719 vop_strategy(*vp->v_ops, vp, bio);
3722 static void vn_cache_strategy_callback(struct bio *bio);
3725 vn_cache_strategy(struct vnode *vp, struct bio *bio)
3727 struct buf *bp = bio->bio_buf;
3734 * Is this buffer cache buffer suitable for reading from
3737 if (vm_swapcache_read_enable == 0 ||
3738 bp->b_cmd != BUF_CMD_READ ||
3739 ((bp->b_flags & B_CLUSTER) == 0 &&
3740 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) ||
3741 ((int)bp->b_loffset & PAGE_MASK) != 0 ||
3742 (bp->b_bcount & PAGE_MASK) != 0) {
3747 * Figure out the original VM object (it will match the underlying
3748 * VM pages). Note that swap cached data uses page indices relative
3749 * to that object, not relative to bio->bio_offset.
3751 if (bp->b_flags & B_CLUSTER)
3752 object = vp->v_object;
3754 object = bp->b_vp->v_object;
3757 * In order to be able to use the swap cache all underlying VM
3758 * pages must be marked as such, and we can't have any bogus pages.
3760 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
3761 m = bp->b_xio.xio_pages[i];
3762 if ((m->flags & PG_SWAPPED) == 0)
3764 if (m == bogus_page)
3769 * If we are good then issue the I/O using swap_pager_strategy().
3771 * We can only do this if the buffer actually supports object-backed
3772 * I/O. If it doesn't npages will be 0.
3774 if (i && i == bp->b_xio.xio_npages) {
3775 m = bp->b_xio.xio_pages[0];
3776 nbio = push_bio(bio);
3777 nbio->bio_done = vn_cache_strategy_callback;
3778 nbio->bio_offset = ptoa(m->pindex);
3779 KKASSERT(m->object == object);
3780 swap_pager_strategy(object, nbio);
3787 * This is a bit of a hack but since the vn_cache_strategy() function can
3788 * override a VFS's strategy function we must make sure that the bio, which
3789 * is probably bio2, doesn't leak an unexpected offset value back to the
3790 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the
3791 * bio went through its own file strategy function and the the bio2 offset
3792 * is a cached disk offset when, in fact, it isn't.
3795 vn_cache_strategy_callback(struct bio *bio)
3797 bio->bio_offset = NOOFFSET;
3798 biodone(pop_bio(bio));
3804 * Finish I/O on a buffer after all BIOs have been processed.
3805 * Called when the bio chain is exhausted or by biowait. If called
3806 * by biowait, elseit is typically 0.
3808 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3809 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3810 * assuming B_INVAL is clear.
3812 * For the VMIO case, we set B_CACHE if the op was a read and no
3813 * read error occured, or if the op was a write. B_CACHE is never
3814 * set if the buffer is invalid or otherwise uncacheable.
3816 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3817 * initiator to leave B_INVAL set to brelse the buffer out of existance
3818 * in the biodone routine.
3821 bpdone(struct buf *bp, int elseit)
3825 KASSERT(BUF_REFCNTNB(bp) > 0,
3826 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3827 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3828 ("biodone: bp %p already done!", bp));
3831 * No more BIOs are left. All completion functions have been dealt
3832 * with, now we clean up the buffer.
3835 bp->b_cmd = BUF_CMD_DONE;
3838 * Only reads and writes are processed past this point.
3840 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3841 if (cmd == BUF_CMD_FREEBLKS)
3842 bp->b_flags |= B_NOCACHE;
3849 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3850 * a lot worse. XXX - move this above the clearing of b_cmd
3852 if (LIST_FIRST(&bp->b_dep) != NULL)
3853 buf_complete(bp); /* MPSAFE */
3856 * A failed write must re-dirty the buffer unless B_INVAL
3857 * was set. Only applicable to normal buffers (with VPs).
3858 * vinum buffers may not have a vp.
3860 if (cmd == BUF_CMD_WRITE &&
3861 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3862 bp->b_flags &= ~B_NOCACHE;
3867 if (bp->b_flags & B_VMIO) {
3873 struct vnode *vp = bp->b_vp;
3877 #if defined(VFS_BIO_DEBUG)
3878 if (vp->v_auxrefs == 0)
3879 panic("biodone: zero vnode hold count");
3880 if ((vp->v_flag & VOBJBUF) == 0)
3881 panic("biodone: vnode is not setup for merged cache");
3884 foff = bp->b_loffset;
3885 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3886 KASSERT(obj != NULL, ("biodone: missing VM object"));
3888 #if defined(VFS_BIO_DEBUG)
3889 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3890 kprintf("biodone: paging in progress(%d) < "
3891 "bp->b_xio.xio_npages(%d)\n",
3892 obj->paging_in_progress,
3893 bp->b_xio.xio_npages);
3898 * Set B_CACHE if the op was a normal read and no error
3899 * occured. B_CACHE is set for writes in the b*write()
3902 iosize = bp->b_bcount - bp->b_resid;
3903 if (cmd == BUF_CMD_READ &&
3904 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3905 bp->b_flags |= B_CACHE;
3908 vm_object_hold(obj);
3909 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3913 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3918 * cleanup bogus pages, restoring the originals. Since
3919 * the originals should still be wired, we don't have
3920 * to worry about interrupt/freeing races destroying
3921 * the VM object association.
3923 m = bp->b_xio.xio_pages[i];
3924 if (m == bogus_page) {
3926 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3928 panic("biodone: page disappeared");
3929 bp->b_xio.xio_pages[i] = m;
3930 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3931 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3933 #if defined(VFS_BIO_DEBUG)
3934 if (OFF_TO_IDX(foff) != m->pindex) {
3935 kprintf("biodone: foff(%lu)/m->pindex(%ld) "
3937 (unsigned long)foff, (long)m->pindex);
3942 * In the write case, the valid and clean bits are
3943 * already changed correctly (see bdwrite()), so we
3944 * only need to do this here in the read case.
3946 vm_page_busy_wait(m, FALSE, "bpdpgw");
3947 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3948 vfs_clean_one_page(bp, i, m);
3950 vm_page_flag_clear(m, PG_ZERO);
3953 * when debugging new filesystems or buffer I/O
3954 * methods, this is the most common error that pops
3955 * up. if you see this, you have not set the page
3956 * busy flag correctly!!!
3959 kprintf("biodone: page busy < 0, "
3960 "pindex: %d, foff: 0x(%x,%x), "
3961 "resid: %d, index: %d\n",
3962 (int) m->pindex, (int)(foff >> 32),
3963 (int) foff & 0xffffffff, resid, i);
3964 if (!vn_isdisk(vp, NULL))
3965 kprintf(" iosize: %ld, loffset: %lld, "
3966 "flags: 0x%08x, npages: %d\n",
3967 bp->b_vp->v_mount->mnt_stat.f_iosize,
3968 (long long)bp->b_loffset,
3969 bp->b_flags, bp->b_xio.xio_npages);
3971 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3972 (long long)bp->b_loffset,
3973 bp->b_flags, bp->b_xio.xio_npages);
3974 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3975 m->valid, m->dirty, m->wire_count);
3976 panic("biodone: page busy < 0");
3978 vm_page_io_finish(m);
3980 vm_object_pip_wakeup(obj);
3981 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3984 bp->b_flags &= ~B_HASBOGUS;
3985 vm_object_drop(obj);
3989 * Finish up by releasing the buffer. There are no more synchronous
3990 * or asynchronous completions, those were handled by bio_done
3994 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
4005 biodone(struct bio *bio)
4007 struct buf *bp = bio->bio_buf;
4009 runningbufwakeup(bp);
4012 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
4015 biodone_t *done_func;
4016 struct bio_track *track;
4019 * BIO tracking. Most but not all BIOs are tracked.
4021 if ((track = bio->bio_track) != NULL) {
4022 bio_track_rel(track);
4023 bio->bio_track = NULL;
4027 * A bio_done function terminates the loop. The function
4028 * will be responsible for any further chaining and/or
4029 * buffer management.
4031 * WARNING! The done function can deallocate the buffer!
4033 if ((done_func = bio->bio_done) != NULL) {
4034 bio->bio_done = NULL;
4038 bio = bio->bio_prev;
4042 * If we've run out of bio's do normal [a]synchronous completion.
4048 * Synchronous biodone - this terminates a synchronous BIO.
4050 * bpdone() is called with elseit=FALSE, leaving the buffer completed
4051 * but still locked. The caller must brelse() the buffer after waiting
4055 biodone_sync(struct bio *bio)
4057 struct buf *bp = bio->bio_buf;
4061 KKASSERT(bio == &bp->b_bio1);
4065 flags = bio->bio_flags;
4066 nflags = (flags | BIO_DONE) & ~BIO_WANT;
4068 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
4069 if (flags & BIO_WANT)
4079 * This routine is called in lieu of iodone in the case of
4080 * incomplete I/O. This keeps the busy status for pages
4084 vfs_unbusy_pages(struct buf *bp)
4088 runningbufwakeup(bp);
4090 if (bp->b_flags & B_VMIO) {
4091 struct vnode *vp = bp->b_vp;
4095 vm_object_hold(obj);
4097 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4098 vm_page_t m = bp->b_xio.xio_pages[i];
4101 * When restoring bogus changes the original pages
4102 * should still be wired, so we are in no danger of
4103 * losing the object association and do not need
4104 * critical section protection particularly.
4106 if (m == bogus_page) {
4107 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
4109 panic("vfs_unbusy_pages: page missing");
4111 bp->b_xio.xio_pages[i] = m;
4112 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4113 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
4115 vm_page_busy_wait(m, FALSE, "bpdpgw");
4116 vm_page_flag_clear(m, PG_ZERO);
4117 vm_page_io_finish(m);
4119 vm_object_pip_wakeup(obj);
4121 bp->b_flags &= ~B_HASBOGUS;
4122 vm_object_drop(obj);
4129 * This routine is called before a device strategy routine.
4130 * It is used to tell the VM system that paging I/O is in
4131 * progress, and treat the pages associated with the buffer
4132 * almost as being PG_BUSY. Also the object 'paging_in_progress'
4133 * flag is handled to make sure that the object doesn't become
4136 * Since I/O has not been initiated yet, certain buffer flags
4137 * such as B_ERROR or B_INVAL may be in an inconsistant state
4138 * and should be ignored.
4143 vfs_busy_pages(struct vnode *vp, struct buf *bp)
4146 struct lwp *lp = curthread->td_lwp;
4149 * The buffer's I/O command must already be set. If reading,
4150 * B_CACHE must be 0 (double check against callers only doing
4151 * I/O when B_CACHE is 0).
4153 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4154 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
4156 if (bp->b_flags & B_VMIO) {
4160 KASSERT(bp->b_loffset != NOOFFSET,
4161 ("vfs_busy_pages: no buffer offset"));
4164 * Busy all the pages. We have to busy them all at once
4165 * to avoid deadlocks.
4168 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4169 vm_page_t m = bp->b_xio.xio_pages[i];
4171 if (vm_page_busy_try(m, FALSE)) {
4172 vm_page_sleep_busy(m, FALSE, "vbpage");
4174 vm_page_wakeup(bp->b_xio.xio_pages[i]);
4180 * Setup for I/O, soft-busy the page right now because
4181 * the next loop may block.
4183 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4184 vm_page_t m = bp->b_xio.xio_pages[i];
4186 vm_page_flag_clear(m, PG_ZERO);
4187 if ((bp->b_flags & B_CLUSTER) == 0) {
4188 vm_object_pip_add(obj, 1);
4189 vm_page_io_start(m);
4194 * Adjust protections for I/O and do bogus-page mapping.
4195 * Assume that vm_page_protect() can block (it can block
4196 * if VM_PROT_NONE, don't take any chances regardless).
4198 * In particular note that for writes we must incorporate
4199 * page dirtyness from the VM system into the buffer's
4202 * For reads we theoretically must incorporate page dirtyness
4203 * from the VM system to determine if the page needs bogus
4204 * replacement, but we shortcut the test by simply checking
4205 * that all m->valid bits are set, indicating that the page
4206 * is fully valid and does not need to be re-read. For any
4207 * VM system dirtyness the page will also be fully valid
4208 * since it was mapped at one point.
4211 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4212 vm_page_t m = bp->b_xio.xio_pages[i];
4214 vm_page_flag_clear(m, PG_ZERO); /* XXX */
4215 if (bp->b_cmd == BUF_CMD_WRITE) {
4217 * When readying a vnode-backed buffer for
4218 * a write we must zero-fill any invalid
4219 * portions of the backing VM pages, mark
4220 * it valid and clear related dirty bits.
4222 * vfs_clean_one_page() incorporates any
4223 * VM dirtyness and updates the b_dirtyoff
4224 * range (after we've made the page RO).
4226 * It is also expected that the pmap modified
4227 * bit has already been cleared by the
4228 * vm_page_protect(). We may not be able
4229 * to clear all dirty bits for a page if it
4230 * was also memory mapped (NFS).
4232 * Finally be sure to unassign any swap-cache
4233 * backing store as it is now stale.
4235 vm_page_protect(m, VM_PROT_READ);
4236 vfs_clean_one_page(bp, i, m);
4237 swap_pager_unswapped(m);
4238 } else if (m->valid == VM_PAGE_BITS_ALL) {
4240 * When readying a vnode-backed buffer for
4241 * read we must replace any dirty pages with
4242 * a bogus page so dirty data is not destroyed
4243 * when filling gaps.
4245 * To avoid testing whether the page is
4246 * dirty we instead test that the page was
4247 * at some point mapped (m->valid fully
4248 * valid) with the understanding that
4249 * this also covers the dirty case.
4251 bp->b_xio.xio_pages[i] = bogus_page;
4252 bp->b_flags |= B_HASBOGUS;
4254 } else if (m->valid & m->dirty) {
4256 * This case should not occur as partial
4257 * dirtyment can only happen if the buffer
4258 * is B_CACHE, and this code is not entered
4259 * if the buffer is B_CACHE.
4261 kprintf("Warning: vfs_busy_pages - page not "
4262 "fully valid! loff=%jx bpf=%08x "
4263 "idx=%d val=%02x dir=%02x\n",
4264 (intmax_t)bp->b_loffset, bp->b_flags,
4265 i, m->valid, m->dirty);
4266 vm_page_protect(m, VM_PROT_NONE);
4269 * The page is not valid and can be made
4272 vm_page_protect(m, VM_PROT_NONE);
4277 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4278 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
4283 * This is the easiest place to put the process accounting for the I/O
4287 if (bp->b_cmd == BUF_CMD_READ)
4288 lp->lwp_ru.ru_inblock++;
4290 lp->lwp_ru.ru_oublock++;
4295 * Tell the VM system that the pages associated with this buffer
4296 * are clean. This is used for delayed writes where the data is
4297 * going to go to disk eventually without additional VM intevention.
4299 * NOTE: While we only really need to clean through to b_bcount, we
4300 * just go ahead and clean through to b_bufsize.
4303 vfs_clean_pages(struct buf *bp)
4308 if ((bp->b_flags & B_VMIO) == 0)
4311 KASSERT(bp->b_loffset != NOOFFSET,
4312 ("vfs_clean_pages: no buffer offset"));
4314 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4315 m = bp->b_xio.xio_pages[i];
4316 vfs_clean_one_page(bp, i, m);
4321 * vfs_clean_one_page:
4323 * Set the valid bits and clear the dirty bits in a page within a
4324 * buffer. The range is restricted to the buffer's size and the
4325 * buffer's logical offset might index into the first page.
4327 * The caller has busied or soft-busied the page and it is not mapped,
4328 * test and incorporate the dirty bits into b_dirtyoff/end before
4329 * clearing them. Note that we need to clear the pmap modified bits
4330 * after determining the the page was dirty, vm_page_set_validclean()
4331 * does not do it for us.
4333 * This routine is typically called after a read completes (dirty should
4334 * be zero in that case as we are not called on bogus-replace pages),
4335 * or before a write is initiated.
4338 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m)
4346 * Calculate offset range within the page but relative to buffer's
4347 * loffset. loffset might be offset into the first page.
4349 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4350 bcount = bp->b_bcount + xoff; /* offset adjusted */
4356 soff = (pageno << PAGE_SHIFT);
4357 eoff = soff + PAGE_SIZE;
4365 * Test dirty bits and adjust b_dirtyoff/end.
4367 * If dirty pages are incorporated into the bp any prior
4368 * B_NEEDCOMMIT state (NFS) must be cleared because the
4369 * caller has not taken into account the new dirty data.
4371 * If the page was memory mapped the dirty bits might go beyond the
4372 * end of the buffer, but we can't really make the assumption that
4373 * a file EOF straddles the buffer (even though this is the case for
4374 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
4375 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
4376 * This also saves some console spam.
4378 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK,
4379 * NFS can handle huge commits but not huge writes.
4381 vm_page_test_dirty(m);
4383 if ((bp->b_flags & B_NEEDCOMMIT) &&
4384 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) {
4386 kprintf("Warning: vfs_clean_one_page: bp %p "
4387 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT"
4388 " cmd %d vd %02x/%02x x/s/e %d %d %d "
4390 bp, (intmax_t)bp->b_loffset, bp->b_bcount,
4391 bp->b_flags, bp->b_cmd,
4392 m->valid, m->dirty, xoff, soff, eoff,
4393 bp->b_dirtyoff, bp->b_dirtyend);
4394 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
4396 print_backtrace(-1);
4399 * Only clear the pmap modified bits if ALL the dirty bits
4400 * are set, otherwise the system might mis-clear portions
4403 if (m->dirty == VM_PAGE_BITS_ALL &&
4404 (bp->b_flags & B_NEEDCOMMIT) == 0) {
4405 pmap_clear_modify(m);
4407 if (bp->b_dirtyoff > soff - xoff)
4408 bp->b_dirtyoff = soff - xoff;
4409 if (bp->b_dirtyend < eoff - xoff)
4410 bp->b_dirtyend = eoff - xoff;
4414 * Set related valid bits, clear related dirty bits.
4415 * Does not mess with the pmap modified bit.
4417 * WARNING! We cannot just clear all of m->dirty here as the
4418 * buffer cache buffers may use a DEV_BSIZE'd aligned
4419 * block size, or have an odd size (e.g. NFS at file EOF).
4420 * The putpages code can clear m->dirty to 0.
4422 * If a VOP_WRITE generates a buffer cache buffer which
4423 * covers the same space as mapped writable pages the
4424 * buffer flush might not be able to clear all the dirty
4425 * bits and still require a putpages from the VM system
4428 * WARNING! vm_page_set_validclean() currently assumes vm_token
4429 * is held. The page might not be busied (bdwrite() case).
4430 * XXX remove this comment once we've validated that this
4431 * is no longer an issue.
4433 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff);
4437 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty.
4438 * The page data is assumed to be valid (there is no zeroing here).
4441 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m)
4449 * Calculate offset range within the page but relative to buffer's
4450 * loffset. loffset might be offset into the first page.
4452 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4453 bcount = bp->b_bcount + xoff; /* offset adjusted */
4459 soff = (pageno << PAGE_SHIFT);
4460 eoff = soff + PAGE_SIZE;
4466 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff);
4472 * Clear a buffer. This routine essentially fakes an I/O, so we need
4473 * to clear B_ERROR and B_INVAL.
4475 * Note that while we only theoretically need to clear through b_bcount,
4476 * we go ahead and clear through b_bufsize.
4480 vfs_bio_clrbuf(struct buf *bp)
4484 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
4485 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
4486 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4487 (bp->b_loffset & PAGE_MASK) == 0) {
4488 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4489 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
4493 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
4494 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
4495 bzero(bp->b_data, bp->b_bufsize);
4496 bp->b_xio.xio_pages[0]->valid |= mask;
4502 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
4503 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
4504 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
4505 ea = (caddr_t)(vm_offset_t)ulmin(
4506 (u_long)(vm_offset_t)ea,
4507 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
4508 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4509 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
4511 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
4512 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
4516 for (; sa < ea; sa += DEV_BSIZE, j++) {
4517 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
4518 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
4519 bzero(sa, DEV_BSIZE);
4522 bp->b_xio.xio_pages[i]->valid |= mask;
4523 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
4532 * vm_hold_load_pages:
4534 * Load pages into the buffer's address space. The pages are
4535 * allocated from the kernel object in order to reduce interference
4536 * with the any VM paging I/O activity. The range of loaded
4537 * pages will be wired.
4539 * If a page cannot be allocated, the 'pagedaemon' is woken up to
4540 * retrieve the full range (to - from) of pages.
4545 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4551 to = round_page(to);
4552 from = round_page(from);
4553 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4558 * Note: must allocate system pages since blocking here
4559 * could intefere with paging I/O, no matter which
4562 vm_object_hold(&kernel_object);
4563 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT,
4564 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
4565 vm_object_drop(&kernel_object);
4568 p->valid = VM_PAGE_BITS_ALL;
4569 vm_page_flag_clear(p, PG_ZERO);
4570 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
4571 bp->b_xio.xio_pages[index] = p;
4578 bp->b_xio.xio_npages = index;
4582 * Allocate pages for a buffer cache buffer.
4584 * Under extremely severe memory conditions even allocating out of the
4585 * system reserve can fail. If this occurs we must allocate out of the
4586 * interrupt reserve to avoid a deadlock with the pageout daemon.
4588 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
4589 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
4590 * against the pageout daemon if pages are not freed from other sources.
4596 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit)
4600 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj));
4603 * Try a normal allocation, allow use of system reserve.
4605 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
4610 * The normal allocation failed and we clearly have a page
4611 * deficit. Try to reclaim some clean VM pages directly
4612 * from the buffer cache.
4614 vm_pageout_deficit += deficit;
4618 * We may have blocked, the caller will know what to do if the
4621 if (vm_page_lookup(obj, pg)) {
4626 * Only system threads can use the interrupt reserve
4628 if ((curthread->td_flags & TDF_SYSTHREAD) == 0) {
4635 * Allocate and allow use of the interrupt reserve.
4637 * If after all that we still can't allocate a VM page we are
4638 * in real trouble, but we slog on anyway hoping that the system
4641 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
4642 VM_ALLOC_INTERRUPT);
4644 if (vm_page_count_severe()) {
4646 vm_wait(hz / 20 + 1);
4649 kprintf("bio_page_alloc: Memory exhausted during bufcache "
4650 "page allocation\n");
4658 * vm_hold_free_pages:
4660 * Return pages associated with the buffer back to the VM system.
4662 * The range of pages underlying the buffer's address space will
4663 * be unmapped and un-wired.
4668 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4672 int index, newnpages;
4674 from = round_page(from);
4675 to = round_page(to);
4676 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4679 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4680 p = bp->b_xio.xio_pages[index];
4681 if (p && (index < bp->b_xio.xio_npages)) {
4683 kprintf("vm_hold_free_pages: doffset: %lld, "
4685 (long long)bp->b_bio2.bio_offset,
4686 (long long)bp->b_loffset);
4688 bp->b_xio.xio_pages[index] = NULL;
4690 vm_page_busy_wait(p, FALSE, "vmhldpg");
4691 vm_page_unwire(p, 0);
4695 bp->b_xio.xio_npages = newnpages;
4701 * Map a user buffer into KVM via a pbuf. On return the buffer's
4702 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4706 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4717 * bp had better have a command and it better be a pbuf.
4719 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4720 KKASSERT(bp->b_flags & B_PAGING);
4721 KKASSERT(bp->b_kvabase);
4727 * Map the user data into KVM. Mappings have to be page-aligned.
4729 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4732 vmprot = VM_PROT_READ;
4733 if (bp->b_cmd == BUF_CMD_READ)
4734 vmprot |= VM_PROT_WRITE;
4736 while (addr < udata + bytes) {
4738 * Do the vm_fault if needed; do the copy-on-write thing
4739 * when reading stuff off device into memory.
4741 * vm_fault_page*() returns a held VM page.
4743 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4744 va = trunc_page(va);
4746 m = vm_fault_page_quick(va, vmprot, &error);
4748 for (i = 0; i < pidx; ++i) {
4749 vm_page_unhold(bp->b_xio.xio_pages[i]);
4750 bp->b_xio.xio_pages[i] = NULL;
4754 bp->b_xio.xio_pages[pidx] = m;
4760 * Map the page array and set the buffer fields to point to
4761 * the mapped data buffer.
4763 if (pidx > btoc(MAXPHYS))
4764 panic("vmapbuf: mapped more than MAXPHYS");
4765 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4767 bp->b_xio.xio_npages = pidx;
4768 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4769 bp->b_bcount = bytes;
4770 bp->b_bufsize = bytes;
4777 * Free the io map PTEs associated with this IO operation.
4778 * We also invalidate the TLB entries and restore the original b_addr.
4781 vunmapbuf(struct buf *bp)
4786 KKASSERT(bp->b_flags & B_PAGING);
4788 npages = bp->b_xio.xio_npages;
4789 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4790 for (pidx = 0; pidx < npages; ++pidx) {
4791 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4792 bp->b_xio.xio_pages[pidx] = NULL;
4794 bp->b_xio.xio_npages = 0;
4795 bp->b_data = bp->b_kvabase;
4799 * Scan all buffers in the system and issue the callback.
4802 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4808 for (n = 0; n < nbuf; ++n) {
4809 if ((error = callback(&buf[n], info)) < 0) {
4819 * nestiobuf_iodone: biodone callback for nested buffers and propagate
4820 * completion to the master buffer.
4823 nestiobuf_iodone(struct bio *bio)
4826 struct buf *mbp, *bp;
4827 struct devstat *stats;
4832 mbio = bio->bio_caller_info1.ptr;
4833 stats = bio->bio_caller_info2.ptr;
4834 mbp = mbio->bio_buf;
4836 KKASSERT(bp->b_bcount <= bp->b_bufsize);
4837 KKASSERT(mbp != bp);
4839 error = bp->b_error;
4840 if (bp->b_error == 0 &&
4841 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
4843 * Not all got transfered, raise an error. We have no way to
4844 * propagate these conditions to mbp.
4849 donebytes = bp->b_bufsize;
4853 nestiobuf_done(mbio, donebytes, error, stats);
4857 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats)
4861 mbp = mbio->bio_buf;
4863 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0);
4866 * If an error occured, propagate it to the master buffer.
4868 * Several biodone()s may wind up running concurrently so
4869 * use an atomic op to adjust b_flags.
4872 mbp->b_error = error;
4873 atomic_set_int(&mbp->b_flags, B_ERROR);
4877 * Decrement the operations in progress counter and terminate the
4878 * I/O if this was the last bit.
4880 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4883 devstat_end_transaction_buf(stats, mbp);
4889 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent
4890 * the mbio from being biodone()'d while we are still adding sub-bios to
4894 nestiobuf_init(struct bio *bio)
4896 bio->bio_driver_info = (void *)1;
4900 * The BIOs added to the nestedio have already been started, remove the
4901 * count that placeheld our mbio and biodone() it if the count would
4905 nestiobuf_start(struct bio *mbio)
4907 struct buf *mbp = mbio->bio_buf;
4910 * Decrement the operations in progress counter and terminate the
4911 * I/O if this was the last bit.
4913 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4914 if (mbp->b_flags & B_ERROR)
4915 mbp->b_resid = mbp->b_bcount;
4923 * Set an intermediate error prior to calling nestiobuf_start()
4926 nestiobuf_error(struct bio *mbio, int error)
4928 struct buf *mbp = mbio->bio_buf;
4931 mbp->b_error = error;
4932 atomic_set_int(&mbp->b_flags, B_ERROR);
4937 * nestiobuf_add: setup a "nested" buffer.
4939 * => 'mbp' is a "master" buffer which is being divided into sub pieces.
4940 * => 'bp' should be a buffer allocated by getiobuf.
4941 * => 'offset' is a byte offset in the master buffer.
4942 * => 'size' is a size in bytes of this nested buffer.
4945 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats)
4947 struct buf *mbp = mbio->bio_buf;
4948 struct vnode *vp = mbp->b_vp;
4950 KKASSERT(mbp->b_bcount >= offset + size);
4952 atomic_add_int((int *)&mbio->bio_driver_info, 1);
4954 /* kernel needs to own the lock for it to be released in biodone */
4957 bp->b_cmd = mbp->b_cmd;
4958 bp->b_bio1.bio_done = nestiobuf_iodone;
4959 bp->b_data = (char *)mbp->b_data + offset;
4960 bp->b_resid = bp->b_bcount = size;
4961 bp->b_bufsize = bp->b_bcount;
4963 bp->b_bio1.bio_track = NULL;
4964 bp->b_bio1.bio_caller_info1.ptr = mbio;
4965 bp->b_bio1.bio_caller_info2.ptr = stats;
4969 * print out statistics from the current status of the buffer pool
4970 * this can be toggeled by the system control option debug.syncprt
4979 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
4980 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
4982 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
4984 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4987 spin_lock(&bufqspin);
4988 TAILQ_FOREACH(bp, dp, b_freelist) {
4989 counts[bp->b_bufsize/PAGE_SIZE]++;
4992 spin_unlock(&bufqspin);
4994 kprintf("%s: total-%d", bname[i], count);
4995 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
4997 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
5005 DB_SHOW_COMMAND(buffer, db_show_buffer)
5008 struct buf *bp = (struct buf *)addr;
5011 db_printf("usage: show buffer <addr>\n");
5015 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
5016 db_printf("b_cmd = %d\n", bp->b_cmd);
5017 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
5018 "b_resid = %d\n, b_data = %p, "
5019 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
5020 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
5022 (long long)bp->b_bio2.bio_offset,
5023 (long long)(bp->b_bio2.bio_next ?
5024 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
5025 if (bp->b_xio.xio_npages) {
5027 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
5028 bp->b_xio.xio_npages);
5029 for (i = 0; i < bp->b_xio.xio_npages; i++) {
5031 m = bp->b_xio.xio_pages[i];
5032 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
5033 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
5034 if ((i + 1) < bp->b_xio.xio_npages)