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);
94 struct bqueues bufqueues[BUFFER_QUEUES];
97 struct bufpcpu bufpcpu[MAXCPU];
99 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
101 struct buf *buf; /* buffer header pool */
103 static void vfs_clean_pages(struct buf *bp);
104 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
106 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m);
108 static void vfs_vmio_release(struct buf *bp);
109 static int flushbufqueues(struct buf *marker, bufq_type_t q);
110 static vm_page_t bio_page_alloc(struct buf *bp, vm_object_t obj,
111 vm_pindex_t pg, int deficit);
113 static void bd_signal(long totalspace);
114 static void buf_daemon(void);
115 static void buf_daemon_hw(void);
118 * bogus page -- for I/O to/from partially complete buffers
119 * this is a temporary solution to the problem, but it is not
120 * really that bad. it would be better to split the buffer
121 * for input in the case of buffers partially already in memory,
122 * but the code is intricate enough already.
124 vm_page_t bogus_page;
127 * These are all static, but make the ones we export globals so we do
128 * not need to use compiler magic.
130 long bufspace; /* locked by buffer_map */
132 static long bufmallocspace; /* atomic ops */
133 long maxbufmallocspace, lobufspace, hibufspace;
134 static long bufreusecnt, bufdefragcnt, buffreekvacnt;
135 static long lorunningspace;
136 static long hirunningspace;
137 static long dirtykvaspace; /* atomic */
138 long dirtybufspace; /* atomic (global for systat) */
139 static long dirtybufcount; /* atomic */
140 static long dirtybufspacehw; /* atomic */
141 static long dirtybufcounthw; /* atomic */
142 static long runningbufspace; /* atomic */
143 static long runningbufcount; /* atomic */
144 long lodirtybufspace;
145 long hidirtybufspace;
146 static int getnewbufcalls;
147 static int getnewbufrestarts;
148 static int recoverbufcalls;
149 static int needsbuffer; /* atomic */
150 static int runningbufreq; /* atomic */
151 static int bd_request; /* atomic */
152 static int bd_request_hw; /* atomic */
153 static u_int bd_wake_ary[BD_WAKE_SIZE];
154 static u_int bd_wake_index;
155 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */
156 static int debug_commit;
157 static int debug_bufbio;
159 static struct thread *bufdaemon_td;
160 static struct thread *bufdaemonhw_td;
161 static u_int lowmempgallocs;
162 static u_int lowmempgfails;
165 * Sysctls for operational control of the buffer cache.
167 SYSCTL_LONG(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
168 "Number of dirty buffers to flush before bufdaemon becomes inactive");
169 SYSCTL_LONG(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
170 "High watermark used to trigger explicit flushing of dirty buffers");
171 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
172 "Minimum amount of buffer space required for active I/O");
173 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
174 "Maximum amount of buffer space to usable for active I/O");
175 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0,
176 "Page allocations done during periods of very low free memory");
177 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0,
178 "Page allocations which failed during periods of very low free memory");
179 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
180 "Recycle pages to active or inactive queue transition pt 0-64");
182 * Sysctls determining current state of the buffer cache.
184 SYSCTL_LONG(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
185 "Total number of buffers in buffer cache");
186 SYSCTL_LONG(_vfs, OID_AUTO, dirtykvaspace, CTLFLAG_RD, &dirtykvaspace, 0,
187 "KVA reserved by dirty buffers (all)");
188 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
189 "Pending bytes of dirty buffers (all)");
190 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
191 "Pending bytes of dirty buffers (heavy weight)");
192 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
193 "Pending number of dirty buffers");
194 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
195 "Pending number of dirty buffers (heavy weight)");
196 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
197 "I/O bytes currently in progress due to asynchronous writes");
198 SYSCTL_LONG(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
199 "I/O buffers currently in progress due to asynchronous writes");
200 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
201 "Hard limit on maximum amount of memory usable for buffer space");
202 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
203 "Soft limit on maximum amount of memory usable for buffer space");
204 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
205 "Minimum amount of memory to reserve for system buffer space");
206 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
207 "Amount of memory available for buffers");
208 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
209 0, "Maximum amount of memory reserved for buffers using malloc");
210 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
211 "Amount of memory left for buffers using malloc-scheme");
212 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
213 "New buffer header acquisition requests");
214 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
215 0, "New buffer header acquisition restarts");
216 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
217 "Recover VM space in an emergency");
218 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
219 "Buffer acquisition restarts due to fragmented buffer map");
220 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
221 "Amount of time KVA space was deallocated in an arbitrary buffer");
222 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
223 "Amount of time buffer re-use operations were successful");
224 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, "");
225 SYSCTL_INT(_vfs, OID_AUTO, debug_bufbio, CTLFLAG_RW, &debug_bufbio, 0, "");
226 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
227 "sizeof(struct buf)");
229 char *buf_wmesg = BUF_WMESG;
231 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
232 #define VFS_BIO_NEED_UNUSED02 0x02
233 #define VFS_BIO_NEED_UNUSED04 0x04
234 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
239 * Called when buffer space is potentially available for recovery.
240 * getnewbuf() will block on this flag when it is unable to free
241 * sufficient buffer space. Buffer space becomes recoverable when
242 * bp's get placed back in the queues.
248 * If someone is waiting for BUF space, wake them up. Even
249 * though we haven't freed the kva space yet, the waiting
250 * process will be able to now.
253 int flags = needsbuffer;
255 if ((flags & VFS_BIO_NEED_BUFSPACE) == 0)
257 if (atomic_cmpset_int(&needsbuffer, flags,
258 flags & ~VFS_BIO_NEED_BUFSPACE)) {
259 wakeup(&needsbuffer);
269 * Accounting for I/O in progress.
273 runningbufwakeup(struct buf *bp)
279 if ((totalspace = bp->b_runningbufspace) != 0) {
280 atomic_add_long(&runningbufspace, -totalspace);
281 atomic_add_long(&runningbufcount, -1);
282 bp->b_runningbufspace = 0;
285 * see waitrunningbufspace() for limit test.
287 limit = hirunningspace * 3 / 6;
289 flags = runningbufreq;
293 if (atomic_cmpset_int(&runningbufreq, flags, 0)) {
294 wakeup(&runningbufreq);
299 bd_signal(totalspace);
306 * Called when a buffer has been added to one of the free queues to
307 * account for the buffer and to wakeup anyone waiting for free buffers.
308 * This typically occurs when large amounts of metadata are being handled
309 * by the buffer cache ( else buffer space runs out first, usually ).
320 if (atomic_cmpset_int(&needsbuffer, flags,
321 (flags & ~VFS_BIO_NEED_ANY))) {
322 wakeup(&needsbuffer);
330 * waitrunningbufspace()
332 * If runningbufspace exceeds 4/6 hirunningspace we block until
333 * runningbufspace drops to 3/6 hirunningspace. We also block if another
334 * thread blocked here in order to be fair, even if runningbufspace
335 * is now lower than the limit.
337 * The caller may be using this function to block in a tight loop, we
338 * must block while runningbufspace is greater than at least
339 * hirunningspace * 3 / 6.
342 waitrunningbufspace(void)
344 long limit = hirunningspace * 4 / 6;
347 while (runningbufspace > limit || runningbufreq) {
348 tsleep_interlock(&runningbufreq, 0);
349 flags = atomic_fetchadd_int(&runningbufreq, 1);
350 if (runningbufspace > limit || flags)
351 tsleep(&runningbufreq, PINTERLOCKED, "wdrn1", hz);
356 * buf_dirty_count_severe:
358 * Return true if we have too many dirty buffers.
361 buf_dirty_count_severe(void)
363 return (runningbufspace + dirtykvaspace >= hidirtybufspace ||
364 dirtybufcount >= nbuf / 2);
368 * Return true if the amount of running I/O is severe and BIOQ should
372 buf_runningbufspace_severe(void)
374 return (runningbufspace >= hirunningspace * 4 / 6);
378 * vfs_buf_test_cache:
380 * Called when a buffer is extended. This function clears the B_CACHE
381 * bit if the newly extended portion of the buffer does not contain
384 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
385 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
386 * them while a clean buffer was present.
390 vfs_buf_test_cache(struct buf *bp,
391 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
394 if (bp->b_flags & B_CACHE) {
395 int base = (foff + off) & PAGE_MASK;
396 if (vm_page_is_valid(m, base, size) == 0)
397 bp->b_flags &= ~B_CACHE;
404 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
411 if (dirtykvaspace < lodirtybufspace && dirtybufcount < nbuf / 2)
414 if (bd_request == 0 &&
415 (dirtykvaspace > lodirtybufspace / 2 ||
416 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
417 if (atomic_fetchadd_int(&bd_request, 1) == 0)
420 if (bd_request_hw == 0 &&
421 (dirtykvaspace > lodirtybufspace / 2 ||
422 dirtybufcounthw >= nbuf / 2)) {
423 if (atomic_fetchadd_int(&bd_request_hw, 1) == 0)
424 wakeup(&bd_request_hw);
431 * Get the buf_daemon heated up when the number of running and dirty
432 * buffers exceeds the mid-point.
434 * Return the total number of dirty bytes past the second mid point
435 * as a measure of how much excess dirty data there is in the system.
444 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
446 totalspace = runningbufspace + dirtykvaspace;
447 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
449 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
450 if (totalspace >= mid2)
451 return(totalspace - mid2);
459 * Wait for the buffer cache to flush (totalspace) bytes worth of
460 * buffers, then return.
462 * Regardless this function blocks while the number of dirty buffers
463 * exceeds hidirtybufspace.
466 bd_wait(long totalspace)
473 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
476 while (totalspace > 0) {
480 * Order is important. Suppliers adjust bd_wake_index after
481 * updating runningbufspace/dirtykvaspace. We want to fetch
482 * bd_wake_index before accessing. Any error should thus
485 i = atomic_fetchadd_int(&bd_wake_index, 0);
486 if (totalspace > runningbufspace + dirtykvaspace)
487 totalspace = runningbufspace + dirtykvaspace;
488 count = totalspace / BKVASIZE;
489 if (count >= BD_WAKE_SIZE / 2)
490 count = BD_WAKE_SIZE / 2;
492 mi = i & BD_WAKE_MASK;
495 * This is not a strict interlock, so we play a bit loose
496 * with locking access to dirtybufspace*. We have to re-check
497 * bd_wake_index to ensure that it hasn't passed us.
499 tsleep_interlock(&bd_wake_ary[mi], 0);
500 atomic_add_int(&bd_wake_ary[mi], 1);
501 j = atomic_fetchadd_int(&bd_wake_index, 0);
502 if ((int)(i - j) >= 0)
503 tsleep(&bd_wake_ary[mi], PINTERLOCKED, "flstik", hz);
505 totalspace = runningbufspace + dirtykvaspace - hidirtybufspace;
512 * This function is called whenever runningbufspace or dirtykvaspace
513 * is reduced. Track threads waiting for run+dirty buffer I/O
517 bd_signal(long totalspace)
521 if (totalspace > 0) {
522 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
523 totalspace = BKVASIZE * BD_WAKE_SIZE;
524 while (totalspace > 0) {
525 i = atomic_fetchadd_int(&bd_wake_index, 1);
527 if (atomic_readandclear_int(&bd_wake_ary[i]))
528 wakeup(&bd_wake_ary[i]);
529 totalspace -= BKVASIZE;
535 * BIO tracking support routines.
537 * Release a ref on a bio_track. Wakeup requests are atomically released
538 * along with the last reference so bk_active will never wind up set to
543 bio_track_rel(struct bio_track *track)
551 active = track->bk_active;
552 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
556 * Full-on. Note that the wait flag is only atomically released on
557 * the 1->0 count transition.
559 * We check for a negative count transition using bit 30 since bit 31
560 * has a different meaning.
563 desired = (active & 0x7FFFFFFF) - 1;
565 desired |= active & 0x80000000;
566 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
567 if (desired & 0x40000000)
568 panic("bio_track_rel: bad count: %p", track);
569 if (active & 0x80000000)
573 active = track->bk_active;
578 * Wait for the tracking count to reach 0.
580 * Use atomic ops such that the wait flag is only set atomically when
581 * bk_active is non-zero.
584 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
593 if (track->bk_active == 0)
597 * Full-on. Note that the wait flag may only be atomically set if
598 * the active count is non-zero.
600 * NOTE: We cannot optimize active == desired since a wakeup could
601 * clear active prior to our tsleep_interlock().
604 while ((active = track->bk_active) != 0) {
606 desired = active | 0x80000000;
607 tsleep_interlock(track, slp_flags);
608 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
609 error = tsleep(track, slp_flags | PINTERLOCKED,
621 * Load time initialisation of the buffer cache, called from machine
622 * dependant initialization code.
626 bufinit(void *dummy __unused)
628 struct bufpcpu *pcpu;
630 vm_offset_t bogus_offset;
635 /* next, make a null set of free lists */
636 for (i = 0; i < ncpus; ++i) {
638 spin_init(&pcpu->spin, "bufinit");
639 for (j = 0; j < BUFFER_QUEUES; j++)
640 TAILQ_INIT(&pcpu->bufqueues[j]);
643 /* finally, initialize each buffer header and stick on empty q */
647 for (n = 0; n < nbuf; n++) {
649 bzero(bp, sizeof *bp);
650 bp->b_flags = B_INVAL; /* we're just an empty header */
651 bp->b_cmd = BUF_CMD_DONE;
652 bp->b_qindex = BQUEUE_EMPTY;
655 xio_init(&bp->b_xio);
657 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
665 * maxbufspace is the absolute maximum amount of buffer space we are
666 * allowed to reserve in KVM and in real terms. The absolute maximum
667 * is nominally used by buf_daemon. hibufspace is the nominal maximum
668 * used by most other processes. The differential is required to
669 * ensure that buf_daemon is able to run when other processes might
670 * be blocked waiting for buffer space.
672 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
673 * this may result in KVM fragmentation which is not handled optimally
676 maxbufspace = nbuf * BKVASIZE;
677 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
678 lobufspace = hibufspace - MAXBSIZE;
680 lorunningspace = 512 * 1024;
681 /* hirunningspace -- see below */
684 * Limit the amount of malloc memory since it is wired permanently
685 * into the kernel space. Even though this is accounted for in
686 * the buffer allocation, we don't want the malloced region to grow
687 * uncontrolled. The malloc scheme improves memory utilization
688 * significantly on average (small) directories.
690 maxbufmallocspace = hibufspace / 20;
693 * Reduce the chance of a deadlock occuring by limiting the number
694 * of delayed-write dirty buffers we allow to stack up.
696 * We don't want too much actually queued to the device at once
697 * (XXX this needs to be per-mount!), because the buffers will
698 * wind up locked for a very long period of time while the I/O
701 hidirtybufspace = hibufspace / 2; /* dirty + running */
702 hirunningspace = hibufspace / 16; /* locked & queued to device */
703 if (hirunningspace < 1024 * 1024)
704 hirunningspace = 1024 * 1024;
710 lodirtybufspace = hidirtybufspace / 2;
713 * Maximum number of async ops initiated per buf_daemon loop. This is
714 * somewhat of a hack at the moment, we really need to limit ourselves
715 * based on the number of bytes of I/O in-transit that were initiated
719 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
720 vm_object_hold(&kernel_object);
721 bogus_page = vm_page_alloc(&kernel_object,
722 (bogus_offset >> PAGE_SHIFT),
724 vm_object_drop(&kernel_object);
725 vmstats.v_wire_count++;
729 SYSINIT(do_bufinit, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, bufinit, NULL);
732 * Initialize the embedded bio structures, typically used by
733 * deprecated code which tries to allocate its own struct bufs.
736 initbufbio(struct buf *bp)
738 bp->b_bio1.bio_buf = bp;
739 bp->b_bio1.bio_prev = NULL;
740 bp->b_bio1.bio_offset = NOOFFSET;
741 bp->b_bio1.bio_next = &bp->b_bio2;
742 bp->b_bio1.bio_done = NULL;
743 bp->b_bio1.bio_flags = 0;
745 bp->b_bio2.bio_buf = bp;
746 bp->b_bio2.bio_prev = &bp->b_bio1;
747 bp->b_bio2.bio_offset = NOOFFSET;
748 bp->b_bio2.bio_next = NULL;
749 bp->b_bio2.bio_done = NULL;
750 bp->b_bio2.bio_flags = 0;
756 * Reinitialize the embedded bio structures as well as any additional
757 * translation cache layers.
760 reinitbufbio(struct buf *bp)
764 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
765 bio->bio_done = NULL;
766 bio->bio_offset = NOOFFSET;
771 * Undo the effects of an initbufbio().
774 uninitbufbio(struct buf *bp)
781 * Push another BIO layer onto an existing BIO and return it. The new
782 * BIO layer may already exist, holding cached translation data.
785 push_bio(struct bio *bio)
789 if ((nbio = bio->bio_next) == NULL) {
790 int index = bio - &bio->bio_buf->b_bio_array[0];
791 if (index >= NBUF_BIO - 1) {
792 panic("push_bio: too many layers bp %p",
795 nbio = &bio->bio_buf->b_bio_array[index + 1];
796 bio->bio_next = nbio;
797 nbio->bio_prev = bio;
798 nbio->bio_buf = bio->bio_buf;
799 nbio->bio_offset = NOOFFSET;
800 nbio->bio_done = NULL;
801 nbio->bio_next = NULL;
803 KKASSERT(nbio->bio_done == NULL);
808 * Pop a BIO translation layer, returning the previous layer. The
809 * must have been previously pushed.
812 pop_bio(struct bio *bio)
814 return(bio->bio_prev);
818 clearbiocache(struct bio *bio)
821 bio->bio_offset = NOOFFSET;
829 * Free the KVA allocation for buffer 'bp'.
831 * Must be called from a critical section as this is the only locking for
834 * Since this call frees up buffer space, we call bufspacewakeup().
837 bfreekva(struct buf *bp)
843 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
844 vm_map_lock(&buffer_map);
845 bufspace -= bp->b_kvasize;
846 vm_map_delete(&buffer_map,
847 (vm_offset_t) bp->b_kvabase,
848 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
851 vm_map_unlock(&buffer_map);
852 vm_map_entry_release(count);
854 bp->b_kvabase = NULL;
860 * Remove the buffer from the appropriate free list.
861 * (caller must be locked)
864 _bremfree(struct buf *bp)
866 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu];
868 if (bp->b_qindex != BQUEUE_NONE) {
869 KASSERT(BUF_REFCNTNB(bp) == 1,
870 ("bremfree: bp %p not locked",bp));
871 TAILQ_REMOVE(&pcpu->bufqueues[bp->b_qindex], bp, b_freelist);
872 bp->b_qindex = BQUEUE_NONE;
874 if (BUF_REFCNTNB(bp) <= 1)
875 panic("bremfree: removing a buffer not on a queue");
880 * bremfree() - must be called with a locked buffer
883 bremfree(struct buf *bp)
885 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu];
887 spin_lock(&pcpu->spin);
889 spin_unlock(&pcpu->spin);
893 * bremfree_locked - must be called with pcpu->spin locked
896 bremfree_locked(struct buf *bp)
902 * This version of bread issues any required I/O asyncnronously and
903 * makes a callback on completion.
905 * The callback must check whether BIO_DONE is set in the bio and issue
906 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing
907 * BIO_DONE and disposing of the I/O (bqrelse()ing it).
910 breadcb(struct vnode *vp, off_t loffset, int size,
911 void (*func)(struct bio *), void *arg)
915 bp = getblk(vp, loffset, size, 0, 0);
917 /* if not found in cache, do some I/O */
918 if ((bp->b_flags & B_CACHE) == 0) {
919 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
920 bp->b_cmd = BUF_CMD_READ;
921 bp->b_bio1.bio_done = func;
922 bp->b_bio1.bio_caller_info1.ptr = arg;
923 vfs_busy_pages(vp, bp);
925 vn_strategy(vp, &bp->b_bio1);
928 * Since we are issuing the callback synchronously it cannot
929 * race the BIO_DONE, so no need for atomic ops here.
931 /*bp->b_bio1.bio_done = func;*/
932 bp->b_bio1.bio_caller_info1.ptr = arg;
933 bp->b_bio1.bio_flags |= BIO_DONE;
941 * breadnx() - Terminal function for bread() and breadn().
943 * This function will start asynchronous I/O on read-ahead blocks as well
944 * as satisfy the primary request.
946 * We must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE is
947 * set, the buffer is valid and we do not have to do anything.
950 breadnx(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
951 int *rabsize, int cnt, struct buf **bpp)
953 struct buf *bp, *rabp;
955 int rv = 0, readwait = 0;
960 *bpp = bp = getblk(vp, loffset, size, 0, 0);
962 /* if not found in cache, do some I/O */
963 if ((bp->b_flags & B_CACHE) == 0) {
964 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
965 bp->b_cmd = BUF_CMD_READ;
966 bp->b_bio1.bio_done = biodone_sync;
967 bp->b_bio1.bio_flags |= BIO_SYNC;
968 vfs_busy_pages(vp, bp);
969 vn_strategy(vp, &bp->b_bio1);
973 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
974 if (inmem(vp, *raoffset))
976 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
978 if ((rabp->b_flags & B_CACHE) == 0) {
979 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
980 rabp->b_cmd = BUF_CMD_READ;
981 vfs_busy_pages(vp, rabp);
983 vn_strategy(vp, &rabp->b_bio1);
989 rv = biowait(&bp->b_bio1, "biord");
996 * Synchronous write, waits for completion.
998 * Write, release buffer on completion. (Done by iodone
999 * if async). Do not bother writing anything if the buffer
1002 * Note that we set B_CACHE here, indicating that buffer is
1003 * fully valid and thus cacheable. This is true even of NFS
1004 * now so we set it generally. This could be set either here
1005 * or in biodone() since the I/O is synchronous. We put it
1009 bwrite(struct buf *bp)
1013 if (bp->b_flags & B_INVAL) {
1017 if (BUF_REFCNTNB(bp) == 0)
1018 panic("bwrite: buffer is not busy???");
1020 /* Mark the buffer clean */
1023 bp->b_flags &= ~(B_ERROR | B_EINTR);
1024 bp->b_flags |= B_CACHE;
1025 bp->b_cmd = BUF_CMD_WRITE;
1026 bp->b_bio1.bio_done = biodone_sync;
1027 bp->b_bio1.bio_flags |= BIO_SYNC;
1028 vfs_busy_pages(bp->b_vp, bp);
1031 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1032 * valid for vnode-backed buffers.
1034 bsetrunningbufspace(bp, bp->b_bufsize);
1035 vn_strategy(bp->b_vp, &bp->b_bio1);
1036 error = biowait(&bp->b_bio1, "biows");
1045 * Asynchronous write. Start output on a buffer, but do not wait for
1046 * it to complete. The buffer is released when the output completes.
1048 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1049 * B_INVAL buffers. Not us.
1052 bawrite(struct buf *bp)
1054 if (bp->b_flags & B_INVAL) {
1058 if (BUF_REFCNTNB(bp) == 0)
1059 panic("bwrite: buffer is not busy???");
1061 /* Mark the buffer clean */
1064 bp->b_flags &= ~(B_ERROR | B_EINTR);
1065 bp->b_flags |= B_CACHE;
1066 bp->b_cmd = BUF_CMD_WRITE;
1067 KKASSERT(bp->b_bio1.bio_done == NULL);
1068 vfs_busy_pages(bp->b_vp, bp);
1071 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1072 * valid for vnode-backed buffers.
1074 bsetrunningbufspace(bp, bp->b_bufsize);
1076 vn_strategy(bp->b_vp, &bp->b_bio1);
1082 * Ordered write. Start output on a buffer, and flag it so that the
1083 * device will write it in the order it was queued. The buffer is
1084 * released when the output completes. bwrite() ( or the VOP routine
1085 * anyway ) is responsible for handling B_INVAL buffers.
1088 bowrite(struct buf *bp)
1090 bp->b_flags |= B_ORDERED;
1098 * Delayed write. (Buffer is marked dirty). Do not bother writing
1099 * anything if the buffer is marked invalid.
1101 * Note that since the buffer must be completely valid, we can safely
1102 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1103 * biodone() in order to prevent getblk from writing the buffer
1104 * out synchronously.
1107 bdwrite(struct buf *bp)
1109 if (BUF_REFCNTNB(bp) == 0)
1110 panic("bdwrite: buffer is not busy");
1112 if (bp->b_flags & B_INVAL) {
1118 if (dsched_is_clear_buf_priv(bp))
1122 * Set B_CACHE, indicating that the buffer is fully valid. This is
1123 * true even of NFS now.
1125 bp->b_flags |= B_CACHE;
1128 * This bmap keeps the system from needing to do the bmap later,
1129 * perhaps when the system is attempting to do a sync. Since it
1130 * is likely that the indirect block -- or whatever other datastructure
1131 * that the filesystem needs is still in memory now, it is a good
1132 * thing to do this. Note also, that if the pageout daemon is
1133 * requesting a sync -- there might not be enough memory to do
1134 * the bmap then... So, this is important to do.
1136 if (bp->b_bio2.bio_offset == NOOFFSET) {
1137 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1138 NULL, NULL, BUF_CMD_WRITE);
1142 * Because the underlying pages may still be mapped and
1143 * writable trying to set the dirty buffer (b_dirtyoff/end)
1144 * range here will be inaccurate.
1146 * However, we must still clean the pages to satisfy the
1147 * vnode_pager and pageout daemon, so theythink the pages
1148 * have been "cleaned". What has really occured is that
1149 * they've been earmarked for later writing by the buffer
1152 * So we get the b_dirtyoff/end update but will not actually
1153 * depend on it (NFS that is) until the pages are busied for
1156 vfs_clean_pages(bp);
1160 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1161 * due to the softdep code.
1166 * Fake write - return pages to VM system as dirty, leave the buffer clean.
1167 * This is used by tmpfs.
1169 * It is important for any VFS using this routine to NOT use it for
1170 * IO_SYNC or IO_ASYNC operations which occur when the system really
1171 * wants to flush VM pages to backing store.
1174 buwrite(struct buf *bp)
1180 * Only works for VMIO buffers. If the buffer is already
1181 * marked for delayed-write we can't avoid the bdwrite().
1183 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) {
1189 * Mark as needing a commit.
1191 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1192 m = bp->b_xio.xio_pages[i];
1193 vm_page_need_commit(m);
1201 * Turn buffer into delayed write request by marking it B_DELWRI.
1202 * B_RELBUF and B_NOCACHE must be cleared.
1204 * We reassign the buffer to itself to properly update it in the
1205 * dirty/clean lists.
1207 * Must be called from a critical section.
1208 * The buffer must be on BQUEUE_NONE.
1211 bdirty(struct buf *bp)
1213 KASSERT(bp->b_qindex == BQUEUE_NONE,
1214 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1215 if (bp->b_flags & B_NOCACHE) {
1216 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1217 bp->b_flags &= ~B_NOCACHE;
1219 if (bp->b_flags & B_INVAL) {
1220 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1222 bp->b_flags &= ~B_RELBUF;
1224 if ((bp->b_flags & B_DELWRI) == 0) {
1225 lwkt_gettoken(&bp->b_vp->v_token);
1226 bp->b_flags |= B_DELWRI;
1228 lwkt_reltoken(&bp->b_vp->v_token);
1230 atomic_add_long(&dirtybufcount, 1);
1231 atomic_add_long(&dirtykvaspace, bp->b_kvasize);
1232 atomic_add_long(&dirtybufspace, bp->b_bufsize);
1233 if (bp->b_flags & B_HEAVY) {
1234 atomic_add_long(&dirtybufcounthw, 1);
1235 atomic_add_long(&dirtybufspacehw, bp->b_bufsize);
1242 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1243 * needs to be flushed with a different buf_daemon thread to avoid
1244 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1247 bheavy(struct buf *bp)
1249 if ((bp->b_flags & B_HEAVY) == 0) {
1250 bp->b_flags |= B_HEAVY;
1251 if (bp->b_flags & B_DELWRI) {
1252 atomic_add_long(&dirtybufcounthw, 1);
1253 atomic_add_long(&dirtybufspacehw, bp->b_bufsize);
1261 * Clear B_DELWRI for buffer.
1263 * Must be called from a critical section.
1265 * The buffer is typically on BQUEUE_NONE but there is one case in
1266 * brelse() that calls this function after placing the buffer on
1267 * a different queue.
1270 bundirty(struct buf *bp)
1272 if (bp->b_flags & B_DELWRI) {
1273 lwkt_gettoken(&bp->b_vp->v_token);
1274 bp->b_flags &= ~B_DELWRI;
1276 lwkt_reltoken(&bp->b_vp->v_token);
1278 atomic_add_long(&dirtybufcount, -1);
1279 atomic_add_long(&dirtykvaspace, -bp->b_kvasize);
1280 atomic_add_long(&dirtybufspace, -bp->b_bufsize);
1281 if (bp->b_flags & B_HEAVY) {
1282 atomic_add_long(&dirtybufcounthw, -1);
1283 atomic_add_long(&dirtybufspacehw, -bp->b_bufsize);
1285 bd_signal(bp->b_bufsize);
1288 * Since it is now being written, we can clear its deferred write flag.
1290 bp->b_flags &= ~B_DEFERRED;
1294 * Set the b_runningbufspace field, used to track how much I/O is
1295 * in progress at any given moment.
1298 bsetrunningbufspace(struct buf *bp, int bytes)
1300 bp->b_runningbufspace = bytes;
1302 atomic_add_long(&runningbufspace, bytes);
1303 atomic_add_long(&runningbufcount, 1);
1310 * Release a busy buffer and, if requested, free its resources. The
1311 * buffer will be stashed in the appropriate bufqueue[] allowing it
1312 * to be accessed later as a cache entity or reused for other purposes.
1315 brelse(struct buf *bp)
1317 struct bufpcpu *pcpu;
1319 int saved_flags = bp->b_flags;
1322 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1323 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1326 * If B_NOCACHE is set we are being asked to destroy the buffer and
1327 * its backing store. Clear B_DELWRI.
1329 * B_NOCACHE is set in two cases: (1) when the caller really wants
1330 * to destroy the buffer and backing store and (2) when the caller
1331 * wants to destroy the buffer and backing store after a write
1334 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1338 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1340 * A re-dirtied buffer is only subject to destruction
1341 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1343 /* leave buffer intact */
1344 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1345 (bp->b_bufsize <= 0)) {
1347 * Either a failed read or we were asked to free or not
1348 * cache the buffer. This path is reached with B_DELWRI
1349 * set only if B_INVAL is already set. B_NOCACHE governs
1350 * backing store destruction.
1352 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1353 * buffer cannot be immediately freed.
1355 bp->b_flags |= B_INVAL;
1356 if (LIST_FIRST(&bp->b_dep) != NULL)
1358 if (bp->b_flags & B_DELWRI) {
1359 atomic_add_long(&dirtybufcount, -1);
1360 atomic_add_long(&dirtykvaspace, -bp->b_kvasize);
1361 atomic_add_long(&dirtybufspace, -bp->b_bufsize);
1362 if (bp->b_flags & B_HEAVY) {
1363 atomic_add_long(&dirtybufcounthw, -1);
1364 atomic_add_long(&dirtybufspacehw,
1367 bd_signal(bp->b_bufsize);
1369 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1373 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set,
1374 * or if b_refs is non-zero.
1376 * If vfs_vmio_release() is called with either bit set, the
1377 * underlying pages may wind up getting freed causing a previous
1378 * write (bdwrite()) to get 'lost' because pages associated with
1379 * a B_DELWRI bp are marked clean. Pages associated with a
1380 * B_LOCKED buffer may be mapped by the filesystem.
1382 * If we want to release the buffer ourselves (rather then the
1383 * originator asking us to release it), give the originator a
1384 * chance to countermand the release by setting B_LOCKED.
1386 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1387 * if B_DELWRI is set.
1389 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1390 * on pages to return pages to the VM page queues.
1392 if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) {
1393 bp->b_flags &= ~B_RELBUF;
1394 } else if (vm_page_count_min(0)) {
1395 if (LIST_FIRST(&bp->b_dep) != NULL)
1396 buf_deallocate(bp); /* can set B_LOCKED */
1397 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1398 bp->b_flags &= ~B_RELBUF;
1400 bp->b_flags |= B_RELBUF;
1404 * Make sure b_cmd is clear. It may have already been cleared by
1407 * At this point destroying the buffer is governed by the B_INVAL
1408 * or B_RELBUF flags.
1410 bp->b_cmd = BUF_CMD_DONE;
1411 dsched_exit_buf(bp);
1414 * VMIO buffer rundown. Make sure the VM page array is restored
1415 * after an I/O may have replaces some of the pages with bogus pages
1416 * in order to not destroy dirty pages in a fill-in read.
1418 * Note that due to the code above, if a buffer is marked B_DELWRI
1419 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1420 * B_INVAL may still be set, however.
1422 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1423 * but not the backing store. B_NOCACHE will destroy the backing
1426 * Note that dirty NFS buffers contain byte-granular write ranges
1427 * and should not be destroyed w/ B_INVAL even if the backing store
1430 if (bp->b_flags & B_VMIO) {
1432 * Rundown for VMIO buffers which are not dirty NFS buffers.
1444 * Get the base offset and length of the buffer. Note that
1445 * in the VMIO case if the buffer block size is not
1446 * page-aligned then b_data pointer may not be page-aligned.
1447 * But our b_xio.xio_pages array *IS* page aligned.
1449 * block sizes less then DEV_BSIZE (usually 512) are not
1450 * supported due to the page granularity bits (m->valid,
1451 * m->dirty, etc...).
1453 * See man buf(9) for more information
1456 resid = bp->b_bufsize;
1457 foff = bp->b_loffset;
1459 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1460 m = bp->b_xio.xio_pages[i];
1461 vm_page_flag_clear(m, PG_ZERO);
1463 * If we hit a bogus page, fixup *all* of them
1464 * now. Note that we left these pages wired
1465 * when we removed them so they had better exist,
1466 * and they cannot be ripped out from under us so
1467 * no critical section protection is necessary.
1469 if (m == bogus_page) {
1471 poff = OFF_TO_IDX(bp->b_loffset);
1473 vm_object_hold(obj);
1474 for (j = i; j < bp->b_xio.xio_npages; j++) {
1477 mtmp = bp->b_xio.xio_pages[j];
1478 if (mtmp == bogus_page) {
1479 mtmp = vm_page_lookup(obj, poff + j);
1481 panic("brelse: page missing");
1483 bp->b_xio.xio_pages[j] = mtmp;
1486 bp->b_flags &= ~B_HASBOGUS;
1487 vm_object_drop(obj);
1489 if ((bp->b_flags & B_INVAL) == 0) {
1490 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1491 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1493 m = bp->b_xio.xio_pages[i];
1497 * Invalidate the backing store if B_NOCACHE is set
1498 * (e.g. used with vinvalbuf()). If this is NFS
1499 * we impose a requirement that the block size be
1500 * a multiple of PAGE_SIZE and create a temporary
1501 * hack to basically invalidate the whole page. The
1502 * problem is that NFS uses really odd buffer sizes
1503 * especially when tracking piecemeal writes and
1504 * it also vinvalbuf()'s a lot, which would result
1505 * in only partial page validation and invalidation
1506 * here. If the file page is mmap()'d, however,
1507 * all the valid bits get set so after we invalidate
1508 * here we would end up with weird m->valid values
1509 * like 0xfc. nfs_getpages() can't handle this so
1510 * we clear all the valid bits for the NFS case
1511 * instead of just some of them.
1513 * The real bug is the VM system having to set m->valid
1514 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1515 * itself is an artifact of the whole 512-byte
1516 * granular mess that exists to support odd block
1517 * sizes and UFS meta-data block sizes (e.g. 6144).
1518 * A complete rewrite is required.
1522 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1523 int poffset = foff & PAGE_MASK;
1526 presid = PAGE_SIZE - poffset;
1527 if (bp->b_vp->v_tag == VT_NFS &&
1528 bp->b_vp->v_type == VREG) {
1530 } else if (presid > resid) {
1533 KASSERT(presid >= 0, ("brelse: extra page"));
1534 vm_page_set_invalid(m, poffset, presid);
1537 * Also make sure any swap cache is removed
1538 * as it is now stale (HAMMER in particular
1539 * uses B_NOCACHE to deal with buffer
1542 swap_pager_unswapped(m);
1544 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1545 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1547 if (bp->b_flags & (B_INVAL | B_RELBUF))
1548 vfs_vmio_release(bp);
1551 * Rundown for non-VMIO buffers.
1553 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1556 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1562 if (bp->b_qindex != BQUEUE_NONE)
1563 panic("brelse: free buffer onto another queue???");
1564 if (BUF_REFCNTNB(bp) > 1) {
1565 /* Temporary panic to verify exclusive locking */
1566 /* This panic goes away when we allow shared refs */
1567 panic("brelse: multiple refs");
1573 * Figure out the correct queue to place the cleaned up buffer on.
1574 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1575 * disassociated from their vnode.
1577 * Return the buffer to its original pcpu area
1579 pcpu = &bufpcpu[bp->b_qcpu];
1580 spin_lock(&pcpu->spin);
1582 if (bp->b_flags & B_LOCKED) {
1584 * Buffers that are locked are placed in the locked queue
1585 * immediately, regardless of their state.
1587 bp->b_qindex = BQUEUE_LOCKED;
1588 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1590 } else if (bp->b_bufsize == 0) {
1592 * Buffers with no memory. Due to conditionals near the top
1593 * of brelse() such buffers should probably already be
1594 * marked B_INVAL and disassociated from their vnode.
1596 bp->b_flags |= B_INVAL;
1597 KASSERT(bp->b_vp == NULL,
1598 ("bp1 %p flags %08x/%08x vnode %p "
1599 "unexpectededly still associated!",
1600 bp, saved_flags, bp->b_flags, bp->b_vp));
1601 KKASSERT((bp->b_flags & B_HASHED) == 0);
1602 if (bp->b_kvasize) {
1603 bp->b_qindex = BQUEUE_EMPTYKVA;
1605 bp->b_qindex = BQUEUE_EMPTY;
1607 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex],
1609 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1611 * Buffers with junk contents. Again these buffers had better
1612 * already be disassociated from their vnode.
1614 KASSERT(bp->b_vp == NULL,
1615 ("bp2 %p flags %08x/%08x vnode %p unexpectededly "
1616 "still associated!",
1617 bp, saved_flags, bp->b_flags, bp->b_vp));
1618 KKASSERT((bp->b_flags & B_HASHED) == 0);
1619 bp->b_flags |= B_INVAL;
1620 bp->b_qindex = BQUEUE_CLEAN;
1621 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex],
1625 * Remaining buffers. These buffers are still associated with
1628 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1630 bp->b_qindex = BQUEUE_DIRTY;
1631 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1634 case B_DELWRI | B_HEAVY:
1635 bp->b_qindex = BQUEUE_DIRTY_HW;
1636 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1641 * NOTE: Buffers are always placed at the end of the
1642 * queue. If B_AGE is not set the buffer will cycle
1643 * through the queue twice.
1645 bp->b_qindex = BQUEUE_CLEAN;
1646 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1651 spin_unlock(&pcpu->spin);
1654 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1655 * on the correct queue but we have not yet unlocked it.
1657 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1661 * The bp is on an appropriate queue unless locked. If it is not
1662 * locked or dirty we can wakeup threads waiting for buffer space.
1664 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1665 * if B_INVAL is set ).
1667 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1671 * Something we can maybe free or reuse
1673 if (bp->b_bufsize || bp->b_kvasize)
1677 * Clean up temporary flags and unlock the buffer.
1679 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
1686 * Release a buffer back to the appropriate queue but do not try to free
1687 * it. The buffer is expected to be used again soon.
1689 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1690 * biodone() to requeue an async I/O on completion. It is also used when
1691 * known good buffers need to be requeued but we think we may need the data
1694 * XXX we should be able to leave the B_RELBUF hint set on completion.
1697 bqrelse(struct buf *bp)
1699 struct bufpcpu *pcpu;
1701 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1702 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1704 if (bp->b_qindex != BQUEUE_NONE)
1705 panic("bqrelse: free buffer onto another queue???");
1706 if (BUF_REFCNTNB(bp) > 1) {
1707 /* do not release to free list */
1708 panic("bqrelse: multiple refs");
1712 buf_act_advance(bp);
1714 pcpu = &bufpcpu[bp->b_qcpu];
1715 spin_lock(&pcpu->spin);
1717 if (bp->b_flags & B_LOCKED) {
1719 * Locked buffers are released to the locked queue. However,
1720 * if the buffer is dirty it will first go into the dirty
1721 * queue and later on after the I/O completes successfully it
1722 * will be released to the locked queue.
1724 bp->b_qindex = BQUEUE_LOCKED;
1725 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1727 } else if (bp->b_flags & B_DELWRI) {
1728 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1729 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1730 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1732 } else if (vm_page_count_min(0)) {
1734 * We are too low on memory, we have to try to free the
1735 * buffer (most importantly: the wired pages making up its
1736 * backing store) *now*.
1738 spin_unlock(&pcpu->spin);
1742 bp->b_qindex = BQUEUE_CLEAN;
1743 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1746 spin_unlock(&pcpu->spin);
1749 * We have now placed the buffer on the proper queue, but have yet
1752 if ((bp->b_flags & B_LOCKED) == 0 &&
1753 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1758 * Something we can maybe free or reuse.
1760 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1764 * Final cleanup and unlock. Clear bits that are only used while a
1765 * buffer is actively locked.
1767 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
1768 dsched_exit_buf(bp);
1773 * Hold a buffer, preventing it from being reused. This will prevent
1774 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL
1775 * operations. If a B_INVAL operation occurs the buffer will remain held
1776 * but the underlying pages may get ripped out.
1778 * These functions are typically used in VOP_READ/VOP_WRITE functions
1779 * to hold a buffer during a copyin or copyout, preventing deadlocks
1780 * or recursive lock panics when read()/write() is used over mmap()'d
1783 * NOTE: bqhold() requires that the buffer be locked at the time of the
1784 * hold. bqdrop() has no requirements other than the buffer having
1785 * previously been held.
1788 bqhold(struct buf *bp)
1790 atomic_add_int(&bp->b_refs, 1);
1794 bqdrop(struct buf *bp)
1796 KKASSERT(bp->b_refs > 0);
1797 atomic_add_int(&bp->b_refs, -1);
1801 * Return backing pages held by the buffer 'bp' back to the VM system.
1802 * This routine is called when the bp is invalidated, released, or
1805 * The KVA mapping (b_data) for the underlying pages is removed by
1808 * WARNING! This routine is integral to the low memory critical path
1809 * when a buffer is B_RELBUF'd. If the system has a severe page
1810 * deficit we need to get the page(s) onto the PQ_FREE or PQ_CACHE
1811 * queues so they can be reused in the current pageout daemon
1815 vfs_vmio_release(struct buf *bp)
1820 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1821 m = bp->b_xio.xio_pages[i];
1822 bp->b_xio.xio_pages[i] = NULL;
1825 * We need to own the page in order to safely unwire it.
1827 vm_page_busy_wait(m, FALSE, "vmiopg");
1830 * The VFS is telling us this is not a meta-data buffer
1831 * even if it is backed by a block device.
1833 if (bp->b_flags & B_NOTMETA)
1834 vm_page_flag_set(m, PG_NOTMETA);
1837 * This is a very important bit of code. We try to track
1838 * VM page use whether the pages are wired into the buffer
1839 * cache or not. While wired into the buffer cache the
1840 * bp tracks the act_count.
1842 * We can choose to place unwired pages on the inactive
1843 * queue (0) or active queue (1). If we place too many
1844 * on the active queue the queue will cycle the act_count
1845 * on pages we'd like to keep, just from single-use pages
1846 * (such as when doing a tar-up or file scan).
1848 if (bp->b_act_count < vm_cycle_point)
1849 vm_page_unwire(m, 0);
1851 vm_page_unwire(m, 1);
1854 * If the wire_count has dropped to 0 we may need to take
1855 * further action before unbusying the page.
1857 * WARNING: vm_page_try_*() also checks PG_NEED_COMMIT for us.
1859 if (m->wire_count == 0) {
1860 vm_page_flag_clear(m, PG_ZERO);
1862 if (bp->b_flags & B_DIRECT) {
1864 * Attempt to free the page if B_DIRECT is
1865 * set, the caller does not desire the page
1869 vm_page_try_to_free(m);
1870 } else if ((bp->b_flags & B_NOTMETA) ||
1871 vm_page_count_min(0)) {
1873 * Attempt to move the page to PQ_CACHE
1874 * if B_NOTMETA is set. This flag is set
1875 * by HAMMER to remove one of the two pages
1876 * present when double buffering is enabled.
1878 * Attempt to move the page to PQ_CACHE
1879 * If we have a severe page deficit. This
1880 * will cause buffer cache operations related
1881 * to pageouts to recycle the related pages
1882 * in order to avoid a low memory deadlock.
1884 m->act_count = bp->b_act_count;
1886 vm_page_try_to_cache(m);
1889 * Nominal case, leave the page on the
1890 * queue the original unwiring placed it on
1891 * (active or inactive).
1893 m->act_count = bp->b_act_count;
1901 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
1902 bp->b_xio.xio_npages);
1903 if (bp->b_bufsize) {
1907 bp->b_xio.xio_npages = 0;
1908 bp->b_flags &= ~B_VMIO;
1909 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1915 * Find and initialize a new buffer header, freeing up existing buffers
1916 * in the bufqueues as necessary. The new buffer is returned locked.
1918 * Important: B_INVAL is not set. If the caller wishes to throw the
1919 * buffer away, the caller must set B_INVAL prior to calling brelse().
1922 * We have insufficient buffer headers
1923 * We have insufficient buffer space
1924 * buffer_map is too fragmented ( space reservation fails )
1925 * If we have to flush dirty buffers ( but we try to avoid this )
1927 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1928 * Instead we ask the buf daemon to do it for us. We attempt to
1929 * avoid piecemeal wakeups of the pageout daemon.
1932 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1934 struct bufpcpu *pcpu;
1940 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1941 int maxloops = 200000;
1942 int restart_reason = 0;
1943 struct buf *restart_bp = NULL;
1944 static int flushingbufs;
1947 * We can't afford to block since we might be holding a vnode lock,
1948 * which may prevent system daemons from running. We deal with
1949 * low-memory situations by proactively returning memory and running
1950 * async I/O rather then sync I/O.
1954 --getnewbufrestarts;
1955 nqcpu = mycpu->gd_cpuid;
1957 ++getnewbufrestarts;
1959 if (debug_bufbio && --maxloops == 0)
1960 panic("getnewbuf, excessive loops on cpu %d restart %d (%p)",
1961 mycpu->gd_cpuid, restart_reason, restart_bp);
1964 * Setup for scan. If we do not have enough free buffers,
1965 * we setup a degenerate case that immediately fails. Note
1966 * that if we are specially marked process, we are allowed to
1967 * dip into our reserves.
1969 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1971 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1972 * However, there are a number of cases (defragging, reusing, ...)
1973 * where we cannot backup.
1975 pcpu = &bufpcpu[nqcpu];
1976 nqindex = BQUEUE_EMPTYKVA;
1977 spin_lock(&pcpu->spin);
1979 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTYKVA]);
1983 * If no EMPTYKVA buffers and we are either
1984 * defragging or reusing, locate a CLEAN buffer
1985 * to free or reuse. If bufspace useage is low
1986 * skip this step so we can allocate a new buffer.
1988 if (defrag || bufspace >= lobufspace) {
1989 nqindex = BQUEUE_CLEAN;
1990 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]);
1994 * If we could not find or were not allowed to reuse a
1995 * CLEAN buffer, check to see if it is ok to use an EMPTY
1996 * buffer. We can only use an EMPTY buffer if allocating
1997 * its KVA would not otherwise run us out of buffer space.
1999 if (nbp == NULL && defrag == 0 &&
2000 bufspace + maxsize < hibufspace) {
2001 nqindex = BQUEUE_EMPTY;
2002 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTY]);
2007 * Run scan, possibly freeing data and/or kva mappings on the fly
2010 * WARNING! spin is held!
2012 while ((bp = nbp) != NULL) {
2013 int qindex = nqindex;
2015 nbp = TAILQ_NEXT(bp, b_freelist);
2018 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2019 * cycles through the queue twice before being selected.
2021 if (qindex == BQUEUE_CLEAN &&
2022 (bp->b_flags & B_AGE) == 0 && nbp) {
2023 bp->b_flags |= B_AGE;
2024 TAILQ_REMOVE(&pcpu->bufqueues[qindex],
2026 TAILQ_INSERT_TAIL(&pcpu->bufqueues[qindex],
2032 * Calculate next bp ( we can only use it if we do not block
2033 * or do other fancy things ).
2038 nqindex = BQUEUE_EMPTYKVA;
2039 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTYKVA])))
2042 case BQUEUE_EMPTYKVA:
2043 nqindex = BQUEUE_CLEAN;
2044 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN])))
2058 KASSERT(bp->b_qindex == qindex,
2059 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2062 * Note: we no longer distinguish between VMIO and non-VMIO
2065 KASSERT((bp->b_flags & B_DELWRI) == 0,
2066 ("delwri buffer %p found in queue %d", bp, qindex));
2069 * Do not try to reuse a buffer with a non-zero b_refs.
2070 * This is an unsynchronized test. A synchronized test
2071 * is also performed after we lock the buffer.
2077 * If we are defragging then we need a buffer with
2078 * b_kvasize != 0. XXX this situation should no longer
2079 * occur, if defrag is non-zero the buffer's b_kvasize
2080 * should also be non-zero at this point. XXX
2082 if (defrag && bp->b_kvasize == 0) {
2083 kprintf("Warning: defrag empty buffer %p\n", bp);
2088 * Start freeing the bp. This is somewhat involved. nbp
2089 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
2090 * on the clean list must be disassociated from their
2091 * current vnode. Buffers on the empty[kva] lists have
2092 * already been disassociated.
2094 * b_refs is checked after locking along with queue changes.
2095 * We must check here to deal with zero->nonzero transitions
2096 * made by the owner of the buffer lock, which is used by
2097 * VFS's to hold the buffer while issuing an unlocked
2098 * uiomove()s. We cannot invalidate the buffer's pages
2099 * for this case. Once we successfully lock a buffer the
2100 * only 0->1 transitions of b_refs will occur via findblk().
2102 * We must also check for queue changes after successful
2103 * locking as the current lock holder may dispose of the
2104 * buffer and change its queue.
2106 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2107 spin_unlock(&pcpu->spin);
2108 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100);
2113 if (bp->b_qindex != qindex || bp->b_refs) {
2114 spin_unlock(&pcpu->spin);
2120 bremfree_locked(bp);
2121 spin_unlock(&pcpu->spin);
2124 * Dependancies must be handled before we disassociate the
2127 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2128 * be immediately disassociated. HAMMER then becomes
2129 * responsible for releasing the buffer.
2131 * NOTE: spin is UNLOCKED now.
2133 if (LIST_FIRST(&bp->b_dep) != NULL) {
2135 if (bp->b_flags & B_LOCKED) {
2141 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2144 if (qindex == BQUEUE_CLEAN) {
2145 if (bp->b_flags & B_VMIO)
2146 vfs_vmio_release(bp);
2152 * NOTE: nbp is now entirely invalid. We can only restart
2153 * the scan from this point on.
2155 * Get the rest of the buffer freed up. b_kva* is still
2156 * valid after this operation.
2158 KASSERT(bp->b_vp == NULL,
2159 ("bp3 %p flags %08x vnode %p qindex %d "
2160 "unexpectededly still associated!",
2161 bp, bp->b_flags, bp->b_vp, qindex));
2162 KKASSERT((bp->b_flags & B_HASHED) == 0);
2165 * critical section protection is not required when
2166 * scrapping a buffer's contents because it is already
2172 if (bp->b_flags & (B_VNDIRTY | B_VNCLEAN | B_HASHED)) {
2173 kprintf("getnewbuf: caught bug vp queue "
2174 "%p/%08x qidx %d\n",
2175 bp, bp->b_flags, qindex);
2178 bp->b_flags = B_BNOCLIP;
2179 bp->b_cmd = BUF_CMD_DONE;
2184 bp->b_xio.xio_npages = 0;
2185 bp->b_dirtyoff = bp->b_dirtyend = 0;
2186 bp->b_act_count = ACT_INIT;
2188 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2190 if (blkflags & GETBLK_BHEAVY)
2191 bp->b_flags |= B_HEAVY;
2194 * If we are defragging then free the buffer.
2197 bp->b_flags |= B_INVAL;
2207 * If we are overcomitted then recover the buffer and its
2208 * KVM space. This occurs in rare situations when multiple
2209 * processes are blocked in getnewbuf() or allocbuf().
2211 * On 64-bit systems BKVASIZE == MAXBSIZE and overcommit
2212 * should not be possible.
2214 if (bufspace >= hibufspace)
2216 if (BKVASIZE != MAXBSIZE) {
2217 if (flushingbufs && bp->b_kvasize != 0) {
2218 bp->b_flags |= B_INVAL;
2226 if (bufspace < lobufspace)
2230 * b_refs can transition to a non-zero value while we hold
2231 * the buffer locked due to a findblk(). Our brelvp() above
2232 * interlocked any future possible transitions due to
2235 * If we find b_refs to be non-zero we can destroy the
2236 * buffer's contents but we cannot yet reuse the buffer.
2239 bp->b_flags |= B_INVAL;
2240 if (BKVASIZE != MAXBSIZE)
2248 /* NOT REACHED, spin not held */
2252 * If we exhausted our list, iterate other cpus. If that fails,
2253 * sleep as appropriate. We may have to wakeup various daemons
2254 * and write out some dirty buffers.
2256 * Generally we are sleeping due to insufficient buffer space.
2258 * NOTE: spin is held if bp is NULL, else it is not held.
2264 spin_unlock(&pcpu->spin);
2266 nqcpu = (nqcpu + 1) % ncpus;
2267 if (nqcpu != mycpu->gd_cpuid) {
2274 flags = VFS_BIO_NEED_BUFSPACE;
2276 } else if (bufspace >= hibufspace) {
2278 flags = VFS_BIO_NEED_BUFSPACE;
2281 flags = VFS_BIO_NEED_ANY;
2284 bd_speedup(); /* heeeelp */
2285 atomic_set_int(&needsbuffer, flags);
2286 while (needsbuffer & flags) {
2289 tsleep_interlock(&needsbuffer, 0);
2290 value = atomic_fetchadd_int(&needsbuffer, 0);
2291 if (value & flags) {
2292 if (tsleep(&needsbuffer, PINTERLOCKED|slpflags,
2293 waitmsg, slptimeo)) {
2300 * We finally have a valid bp. We aren't quite out of the
2301 * woods, we still have to reserve kva space. In order
2302 * to keep fragmentation sane we only allocate kva in
2305 * (spin is not held)
2307 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2309 if (maxsize != bp->b_kvasize) {
2310 vm_offset_t addr = 0;
2315 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2316 vm_map_lock(&buffer_map);
2318 if (vm_map_findspace(&buffer_map,
2319 vm_map_min(&buffer_map), maxsize,
2320 maxsize, 0, &addr)) {
2322 * Uh oh. Buffer map is too fragmented. We
2323 * must defragment the map.
2325 vm_map_unlock(&buffer_map);
2326 vm_map_entry_release(count);
2329 bp->b_flags |= B_INVAL;
2336 vm_map_insert(&buffer_map, &count,
2338 addr, addr + maxsize,
2340 VM_PROT_ALL, VM_PROT_ALL,
2343 bp->b_kvabase = (caddr_t) addr;
2344 bp->b_kvasize = maxsize;
2345 bufspace += bp->b_kvasize;
2348 vm_map_unlock(&buffer_map);
2349 vm_map_entry_release(count);
2351 bp->b_data = bp->b_kvabase;
2359 * Buffer flushing daemon. Buffers are normally flushed by the
2360 * update daemon but if it cannot keep up this process starts to
2361 * take the load in an attempt to prevent getnewbuf() from blocking.
2363 * Once a flush is initiated it does not stop until the number
2364 * of buffers falls below lodirtybuffers, but we will wake up anyone
2365 * waiting at the mid-point.
2367 static struct kproc_desc buf_kp = {
2372 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2373 kproc_start, &buf_kp)
2375 static struct kproc_desc bufhw_kp = {
2380 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2381 kproc_start, &bufhw_kp)
2384 buf_daemon1(struct thread *td, int queue, int (*buf_limit_fn)(long),
2390 marker = kmalloc(sizeof(*marker), M_BIOBUF, M_WAITOK | M_ZERO);
2391 marker->b_flags |= B_MARKER;
2392 marker->b_qindex = BQUEUE_NONE;
2396 * This process needs to be suspended prior to shutdown sync.
2398 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2399 td, SHUTDOWN_PRI_LAST);
2400 curthread->td_flags |= TDF_SYSTHREAD;
2403 * This process is allowed to take the buffer cache to the limit
2406 kproc_suspend_loop();
2409 * Do the flush as long as the number of dirty buffers
2410 * (including those running) exceeds lodirtybufspace.
2412 * When flushing limit running I/O to hirunningspace
2413 * Do the flush. Limit the amount of in-transit I/O we
2414 * allow to build up, otherwise we would completely saturate
2415 * the I/O system. Wakeup any waiting processes before we
2416 * normally would so they can run in parallel with our drain.
2418 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2419 * but because we split the operation into two threads we
2420 * have to cut it in half for each thread.
2422 waitrunningbufspace();
2423 limit = lodirtybufspace / 2;
2424 while (buf_limit_fn(limit)) {
2425 if (flushbufqueues(marker, queue) == 0)
2427 if (runningbufspace < hirunningspace)
2429 waitrunningbufspace();
2433 * We reached our low water mark, reset the
2434 * request and sleep until we are needed again.
2435 * The sleep is just so the suspend code works.
2437 tsleep_interlock(bd_req, 0);
2438 if (atomic_swap_int(bd_req, 0) == 0)
2439 tsleep(bd_req, PINTERLOCKED, "psleep", hz);
2442 /*kfree(marker, M_BIOBUF);*/
2446 buf_daemon_limit(long limit)
2448 return (runningbufspace + dirtykvaspace > limit ||
2449 dirtybufcount - dirtybufcounthw >= nbuf / 2);
2453 buf_daemon_hw_limit(long limit)
2455 return (runningbufspace + dirtykvaspace > limit ||
2456 dirtybufcounthw >= nbuf / 2);
2462 buf_daemon1(bufdaemon_td, BQUEUE_DIRTY, buf_daemon_limit,
2469 buf_daemon1(bufdaemonhw_td, BQUEUE_DIRTY_HW, buf_daemon_hw_limit,
2476 * Try to flush a buffer in the dirty queue. We must be careful to
2477 * free up B_INVAL buffers instead of write them, which NFS is
2478 * particularly sensitive to.
2480 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2481 * that we really want to try to get the buffer out and reuse it
2482 * due to the write load on the machine.
2484 * We must lock the buffer in order to check its validity before we
2485 * can mess with its contents. spin isn't enough.
2488 flushbufqueues(struct buf *marker, bufq_type_t q)
2490 struct bufpcpu *pcpu;
2493 int lcpu = marker->b_qcpu;
2495 KKASSERT(marker->b_qindex == BQUEUE_NONE);
2496 KKASSERT(marker->b_flags & B_MARKER);
2500 * Spinlock needed to perform operations on the queue and may be
2501 * held through a non-blocking BUF_LOCK(), but cannot be held when
2502 * BUF_UNLOCK()ing or through any other major operation.
2504 pcpu = &bufpcpu[marker->b_qcpu];
2505 spin_lock(&pcpu->spin);
2506 marker->b_qindex = q;
2507 TAILQ_INSERT_HEAD(&pcpu->bufqueues[q], marker, b_freelist);
2510 while ((bp = TAILQ_NEXT(bp, b_freelist)) != NULL) {
2512 * NOTE: spinlock is always held at the top of the loop
2514 if (bp->b_flags & B_MARKER)
2516 if ((bp->b_flags & B_DELWRI) == 0) {
2517 kprintf("Unexpected clean buffer %p\n", bp);
2520 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
2522 KKASSERT(bp->b_qcpu == marker->b_qcpu && bp->b_qindex == q);
2525 * Once the buffer is locked we will have no choice but to
2526 * unlock the spinlock around a later BUF_UNLOCK and re-set
2527 * bp = marker when looping. Move the marker now to make
2530 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist);
2531 TAILQ_INSERT_AFTER(&pcpu->bufqueues[q], bp, marker, b_freelist);
2534 * Must recheck B_DELWRI after successfully locking
2537 if ((bp->b_flags & B_DELWRI) == 0) {
2538 spin_unlock(&pcpu->spin);
2540 spin_lock(&pcpu->spin);
2546 * Remove the buffer from its queue. We still own the
2552 * Disposing of an invalid buffer counts as a flush op
2554 if (bp->b_flags & B_INVAL) {
2555 spin_unlock(&pcpu->spin);
2557 spin_lock(&pcpu->spin);
2563 * Release the spinlock for the more complex ops we
2564 * are now going to do.
2566 spin_unlock(&pcpu->spin);
2570 * This is a bit messy
2572 if (LIST_FIRST(&bp->b_dep) != NULL &&
2573 (bp->b_flags & B_DEFERRED) == 0 &&
2574 buf_countdeps(bp, 0)) {
2575 spin_lock(&pcpu->spin);
2576 TAILQ_INSERT_TAIL(&pcpu->bufqueues[q], bp, b_freelist);
2578 bp->b_flags |= B_DEFERRED;
2579 spin_unlock(&pcpu->spin);
2581 spin_lock(&pcpu->spin);
2587 * spinlock not held here.
2589 * If the buffer has a dependancy, buf_checkwrite() must
2590 * also return 0 for us to be able to initate the write.
2592 * If the buffer is flagged B_ERROR it may be requeued
2593 * over and over again, we try to avoid a live lock.
2595 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) {
2597 } else if (bp->b_flags & B_ERROR) {
2598 tsleep(bp, 0, "bioer", 1);
2599 bp->b_flags &= ~B_AGE;
2602 bp->b_flags |= B_AGE;
2605 spin_lock(&pcpu->spin);
2610 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist);
2611 marker->b_qindex = BQUEUE_NONE;
2612 spin_unlock(&pcpu->spin);
2615 * Advance the marker to be fair.
2617 marker->b_qcpu = (marker->b_qcpu + 1) % ncpus;
2619 if (marker->b_qcpu != lcpu)
2629 * Returns true if no I/O is needed to access the associated VM object.
2630 * This is like findblk except it also hunts around in the VM system for
2633 * Note that we ignore vm_page_free() races from interrupts against our
2634 * lookup, since if the caller is not protected our return value will not
2635 * be any more valid then otherwise once we exit the critical section.
2638 inmem(struct vnode *vp, off_t loffset)
2641 vm_offset_t toff, tinc, size;
2645 if (findblk(vp, loffset, FINDBLK_TEST))
2647 if (vp->v_mount == NULL)
2649 if ((obj = vp->v_object) == NULL)
2653 if (size > vp->v_mount->mnt_stat.f_iosize)
2654 size = vp->v_mount->mnt_stat.f_iosize;
2656 vm_object_hold(obj);
2657 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2658 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2664 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2665 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2666 if (vm_page_is_valid(m,
2667 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) {
2672 vm_object_drop(obj);
2679 * Locate and return the specified buffer. Unless flagged otherwise,
2680 * a locked buffer will be returned if it exists or NULL if it does not.
2682 * findblk()'d buffers are still on the bufqueues and if you intend
2683 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2684 * and possibly do other stuff to it.
2686 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2687 * for locking the buffer and ensuring that it remains
2688 * the desired buffer after locking.
2690 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2691 * to acquire the lock we return NULL, even if the
2694 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal
2695 * reuse by getnewbuf() but does not prevent
2696 * disassociation (B_INVAL). Used to avoid deadlocks
2697 * against random (vp,loffset)s due to reassignment.
2699 * (0) - Lock the buffer blocking.
2702 findblk(struct vnode *vp, off_t loffset, int flags)
2707 lkflags = LK_EXCLUSIVE;
2708 if (flags & FINDBLK_NBLOCK)
2709 lkflags |= LK_NOWAIT;
2713 * Lookup. Ref the buf while holding v_token to prevent
2714 * reuse (but does not prevent diassociation).
2716 lwkt_gettoken_shared(&vp->v_token);
2717 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2719 lwkt_reltoken(&vp->v_token);
2723 lwkt_reltoken(&vp->v_token);
2726 * If testing only break and return bp, do not lock.
2728 if (flags & FINDBLK_TEST)
2732 * Lock the buffer, return an error if the lock fails.
2733 * (only FINDBLK_NBLOCK can cause the lock to fail).
2735 if (BUF_LOCK(bp, lkflags)) {
2736 atomic_subtract_int(&bp->b_refs, 1);
2737 /* bp = NULL; not needed */
2742 * Revalidate the locked buf before allowing it to be
2745 if (bp->b_vp == vp && bp->b_loffset == loffset)
2747 atomic_subtract_int(&bp->b_refs, 1);
2754 if ((flags & FINDBLK_REF) == 0)
2755 atomic_subtract_int(&bp->b_refs, 1);
2762 * Similar to getblk() except only returns the buffer if it is
2763 * B_CACHE and requires no other manipulation. Otherwise NULL
2764 * is returned. NULL is also returned if GETBLK_NOWAIT is set
2765 * and the getblk() would block.
2767 * If B_RAM is set the buffer might be just fine, but we return
2768 * NULL anyway because we want the code to fall through to the
2769 * cluster read. Otherwise read-ahead breaks.
2771 * If blksize is 0 the buffer cache buffer must already be fully
2774 * If blksize is non-zero getblk() will be used, allowing a buffer
2775 * to be reinstantiated from its VM backing store. The buffer must
2776 * still be fully cached after reinstantiation to be returned.
2779 getcacheblk(struct vnode *vp, off_t loffset, int blksize, int blkflags)
2782 int fndflags = (blkflags & GETBLK_NOWAIT) ? FINDBLK_NBLOCK : 0;
2785 bp = getblk(vp, loffset, blksize, blkflags, 0);
2787 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2789 bp->b_flags &= ~B_AGE;
2796 bp = findblk(vp, loffset, fndflags);
2798 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2800 bp->b_flags &= ~B_AGE;
2814 * Get a block given a specified block and offset into a file/device.
2815 * B_INVAL may or may not be set on return. The caller should clear
2816 * B_INVAL prior to initiating a READ.
2818 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2819 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2820 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2821 * without doing any of those things the system will likely believe
2822 * the buffer to be valid (especially if it is not B_VMIO), and the
2823 * next getblk() will return the buffer with B_CACHE set.
2825 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2826 * an existing buffer.
2828 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2829 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2830 * and then cleared based on the backing VM. If the previous buffer is
2831 * non-0-sized but invalid, B_CACHE will be cleared.
2833 * If getblk() must create a new buffer, the new buffer is returned with
2834 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2835 * case it is returned with B_INVAL clear and B_CACHE set based on the
2838 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2839 * B_CACHE bit is clear.
2841 * What this means, basically, is that the caller should use B_CACHE to
2842 * determine whether the buffer is fully valid or not and should clear
2843 * B_INVAL prior to issuing a read. If the caller intends to validate
2844 * the buffer by loading its data area with something, the caller needs
2845 * to clear B_INVAL. If the caller does this without issuing an I/O,
2846 * the caller should set B_CACHE ( as an optimization ), else the caller
2847 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2848 * a write attempt or if it was a successfull read. If the caller
2849 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2850 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2854 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2855 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2858 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2861 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2865 if (size > MAXBSIZE)
2866 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2867 if (vp->v_object == NULL)
2868 panic("getblk: vnode %p has no object!", vp);
2871 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) {
2873 * The buffer was found in the cache, but we need to lock it.
2874 * We must acquire a ref on the bp to prevent reuse, but
2875 * this will not prevent disassociation (brelvp()) so we
2876 * must recheck (vp,loffset) after acquiring the lock.
2878 * Without the ref the buffer could potentially be reused
2879 * before we acquire the lock and create a deadlock
2880 * situation between the thread trying to reuse the buffer
2881 * and us due to the fact that we would wind up blocking
2882 * on a random (vp,loffset).
2884 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2885 if (blkflags & GETBLK_NOWAIT) {
2889 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2890 if (blkflags & GETBLK_PCATCH)
2891 lkflags |= LK_PCATCH;
2892 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2895 if (error == ENOLCK)
2899 /* buffer may have changed on us */
2904 * Once the buffer has been locked, make sure we didn't race
2905 * a buffer recyclement. Buffers that are no longer hashed
2906 * will have b_vp == NULL, so this takes care of that check
2909 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2910 kprintf("Warning buffer %p (vp %p loffset %lld) "
2912 bp, vp, (long long)loffset);
2918 * If SZMATCH any pre-existing buffer must be of the requested
2919 * size or NULL is returned. The caller absolutely does not
2920 * want getblk() to bwrite() the buffer on a size mismatch.
2922 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2928 * All vnode-based buffers must be backed by a VM object.
2930 KKASSERT(bp->b_flags & B_VMIO);
2931 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2932 bp->b_flags &= ~B_AGE;
2935 * Make sure that B_INVAL buffers do not have a cached
2936 * block number translation.
2938 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2939 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2940 " did not have cleared bio_offset cache\n",
2941 bp, vp, (long long)loffset);
2942 clearbiocache(&bp->b_bio2);
2946 * The buffer is locked. B_CACHE is cleared if the buffer is
2949 if (bp->b_flags & B_INVAL)
2950 bp->b_flags &= ~B_CACHE;
2954 * Any size inconsistancy with a dirty buffer or a buffer
2955 * with a softupdates dependancy must be resolved. Resizing
2956 * the buffer in such circumstances can lead to problems.
2958 * Dirty or dependant buffers are written synchronously.
2959 * Other types of buffers are simply released and
2960 * reconstituted as they may be backed by valid, dirty VM
2961 * pages (but not marked B_DELWRI).
2963 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2964 * and may be left over from a prior truncation (and thus
2965 * no longer represent the actual EOF point), so we
2966 * definitely do not want to B_NOCACHE the backing store.
2968 if (size != bp->b_bcount) {
2969 if (bp->b_flags & B_DELWRI) {
2970 bp->b_flags |= B_RELBUF;
2972 } else if (LIST_FIRST(&bp->b_dep)) {
2973 bp->b_flags |= B_RELBUF;
2976 bp->b_flags |= B_RELBUF;
2981 KKASSERT(size <= bp->b_kvasize);
2982 KASSERT(bp->b_loffset != NOOFFSET,
2983 ("getblk: no buffer offset"));
2986 * A buffer with B_DELWRI set and B_CACHE clear must
2987 * be committed before we can return the buffer in
2988 * order to prevent the caller from issuing a read
2989 * ( due to B_CACHE not being set ) and overwriting
2992 * Most callers, including NFS and FFS, need this to
2993 * operate properly either because they assume they
2994 * can issue a read if B_CACHE is not set, or because
2995 * ( for example ) an uncached B_DELWRI might loop due
2996 * to softupdates re-dirtying the buffer. In the latter
2997 * case, B_CACHE is set after the first write completes,
2998 * preventing further loops.
3000 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
3001 * above while extending the buffer, we cannot allow the
3002 * buffer to remain with B_CACHE set after the write
3003 * completes or it will represent a corrupt state. To
3004 * deal with this we set B_NOCACHE to scrap the buffer
3007 * XXX Should this be B_RELBUF instead of B_NOCACHE?
3008 * I'm not even sure this state is still possible
3009 * now that getblk() writes out any dirty buffers
3012 * We might be able to do something fancy, like setting
3013 * B_CACHE in bwrite() except if B_DELWRI is already set,
3014 * so the below call doesn't set B_CACHE, but that gets real
3015 * confusing. This is much easier.
3018 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3019 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
3020 "and CACHE clear, b_flags %08x\n",
3021 bp, (uintmax_t)bp->b_loffset, bp->b_flags);
3022 bp->b_flags |= B_NOCACHE;
3028 * Buffer is not in-core, create new buffer. The buffer
3029 * returned by getnewbuf() is locked. Note that the returned
3030 * buffer is also considered valid (not marked B_INVAL).
3032 * Calculating the offset for the I/O requires figuring out
3033 * the block size. We use DEV_BSIZE for VBLK or VCHR and
3034 * the mount's f_iosize otherwise. If the vnode does not
3035 * have an associated mount we assume that the passed size is
3038 * Note that vn_isdisk() cannot be used here since it may
3039 * return a failure for numerous reasons. Note that the
3040 * buffer size may be larger then the block size (the caller
3041 * will use block numbers with the proper multiple). Beware
3042 * of using any v_* fields which are part of unions. In
3043 * particular, in DragonFly the mount point overloading
3044 * mechanism uses the namecache only and the underlying
3045 * directory vnode is not a special case.
3049 if (vp->v_type == VBLK || vp->v_type == VCHR)
3051 else if (vp->v_mount)
3052 bsize = vp->v_mount->mnt_stat.f_iosize;
3056 maxsize = size + (loffset & PAGE_MASK);
3057 maxsize = imax(maxsize, bsize);
3059 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
3061 if (slpflags || slptimeo)
3067 * Atomically insert the buffer into the hash, so that it can
3068 * be found by findblk().
3070 * If bgetvp() returns non-zero a collision occured, and the
3071 * bp will not be associated with the vnode.
3073 * Make sure the translation layer has been cleared.
3075 bp->b_loffset = loffset;
3076 bp->b_bio2.bio_offset = NOOFFSET;
3077 /* bp->b_bio2.bio_next = NULL; */
3079 if (bgetvp(vp, bp, size)) {
3080 bp->b_flags |= B_INVAL;
3086 * All vnode-based buffers must be backed by a VM object.
3088 KKASSERT(vp->v_object != NULL);
3089 bp->b_flags |= B_VMIO;
3090 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3094 KKASSERT(dsched_is_clear_buf_priv(bp));
3101 * Reacquire a buffer that was previously released to the locked queue,
3102 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
3103 * set B_LOCKED (which handles the acquisition race).
3105 * To this end, either B_LOCKED must be set or the dependancy list must be
3109 regetblk(struct buf *bp)
3111 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
3112 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
3119 * Get an empty, disassociated buffer of given size. The buffer is
3120 * initially set to B_INVAL.
3122 * critical section protection is not required for the allocbuf()
3123 * call because races are impossible here.
3131 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3133 while ((bp = getnewbuf(0, 0, size, maxsize)) == NULL)
3136 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3137 KKASSERT(dsched_is_clear_buf_priv(bp));
3145 * This code constitutes the buffer memory from either anonymous system
3146 * memory (in the case of non-VMIO operations) or from an associated
3147 * VM object (in the case of VMIO operations). This code is able to
3148 * resize a buffer up or down.
3150 * Note that this code is tricky, and has many complications to resolve
3151 * deadlock or inconsistant data situations. Tread lightly!!!
3152 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3153 * the caller. Calling this code willy nilly can result in the loss of
3156 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3157 * B_CACHE for the non-VMIO case.
3159 * This routine does not need to be called from a critical section but you
3160 * must own the buffer.
3163 allocbuf(struct buf *bp, int size)
3165 int newbsize, mbsize;
3168 if (BUF_REFCNT(bp) == 0)
3169 panic("allocbuf: buffer not busy");
3171 if (bp->b_kvasize < size)
3172 panic("allocbuf: buffer too small");
3174 if ((bp->b_flags & B_VMIO) == 0) {
3178 * Just get anonymous memory from the kernel. Don't
3179 * mess with B_CACHE.
3181 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3182 if (bp->b_flags & B_MALLOC)
3185 newbsize = round_page(size);
3187 if (newbsize < bp->b_bufsize) {
3189 * Malloced buffers are not shrunk
3191 if (bp->b_flags & B_MALLOC) {
3193 bp->b_bcount = size;
3195 kfree(bp->b_data, M_BIOBUF);
3196 if (bp->b_bufsize) {
3197 atomic_subtract_long(&bufmallocspace, bp->b_bufsize);
3201 bp->b_data = bp->b_kvabase;
3203 bp->b_flags &= ~B_MALLOC;
3209 (vm_offset_t) bp->b_data + newbsize,
3210 (vm_offset_t) bp->b_data + bp->b_bufsize);
3211 } else if (newbsize > bp->b_bufsize) {
3213 * We only use malloced memory on the first allocation.
3214 * and revert to page-allocated memory when the buffer
3217 if ((bufmallocspace < maxbufmallocspace) &&
3218 (bp->b_bufsize == 0) &&
3219 (mbsize <= PAGE_SIZE/2)) {
3221 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3222 bp->b_bufsize = mbsize;
3223 bp->b_bcount = size;
3224 bp->b_flags |= B_MALLOC;
3225 atomic_add_long(&bufmallocspace, mbsize);
3231 * If the buffer is growing on its other-than-first
3232 * allocation, then we revert to the page-allocation
3235 if (bp->b_flags & B_MALLOC) {
3236 origbuf = bp->b_data;
3237 origbufsize = bp->b_bufsize;
3238 bp->b_data = bp->b_kvabase;
3239 if (bp->b_bufsize) {
3240 atomic_subtract_long(&bufmallocspace,
3245 bp->b_flags &= ~B_MALLOC;
3246 newbsize = round_page(newbsize);
3250 (vm_offset_t) bp->b_data + bp->b_bufsize,
3251 (vm_offset_t) bp->b_data + newbsize);
3253 bcopy(origbuf, bp->b_data, origbufsize);
3254 kfree(origbuf, M_BIOBUF);
3261 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3262 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3263 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3264 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3266 if (bp->b_flags & B_MALLOC)
3267 panic("allocbuf: VMIO buffer can't be malloced");
3269 * Set B_CACHE initially if buffer is 0 length or will become
3272 if (size == 0 || bp->b_bufsize == 0)
3273 bp->b_flags |= B_CACHE;
3275 if (newbsize < bp->b_bufsize) {
3277 * DEV_BSIZE aligned new buffer size is less then the
3278 * DEV_BSIZE aligned existing buffer size. Figure out
3279 * if we have to remove any pages.
3281 if (desiredpages < bp->b_xio.xio_npages) {
3282 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3284 * the page is not freed here -- it
3285 * is the responsibility of
3286 * vnode_pager_setsize
3288 m = bp->b_xio.xio_pages[i];
3289 KASSERT(m != bogus_page,
3290 ("allocbuf: bogus page found"));
3291 vm_page_busy_wait(m, TRUE, "biodep");
3292 bp->b_xio.xio_pages[i] = NULL;
3293 vm_page_unwire(m, 0);
3296 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3297 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3298 bp->b_xio.xio_npages = desiredpages;
3300 } else if (size > bp->b_bcount) {
3302 * We are growing the buffer, possibly in a
3303 * byte-granular fashion.
3311 * Step 1, bring in the VM pages from the object,
3312 * allocating them if necessary. We must clear
3313 * B_CACHE if these pages are not valid for the
3314 * range covered by the buffer.
3316 * critical section protection is required to protect
3317 * against interrupts unbusying and freeing pages
3318 * between our vm_page_lookup() and our
3319 * busycheck/wiring call.
3324 vm_object_hold(obj);
3325 while (bp->b_xio.xio_npages < desiredpages) {
3330 pi = OFF_TO_IDX(bp->b_loffset) +
3331 bp->b_xio.xio_npages;
3334 * Blocking on m->busy might lead to a
3337 * vm_fault->getpages->cluster_read->allocbuf
3339 m = vm_page_lookup_busy_try(obj, pi, FALSE,
3342 vm_page_sleep_busy(m, FALSE, "pgtblk");
3347 * note: must allocate system pages
3348 * since blocking here could intefere
3349 * with paging I/O, no matter which
3352 m = bio_page_alloc(bp, obj, pi, desiredpages - bp->b_xio.xio_npages);
3355 vm_page_flag_clear(m, PG_ZERO);
3357 bp->b_flags &= ~B_CACHE;
3358 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3359 ++bp->b_xio.xio_npages;
3365 * We found a page and were able to busy it.
3367 vm_page_flag_clear(m, PG_ZERO);
3370 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3371 ++bp->b_xio.xio_npages;
3372 if (bp->b_act_count < m->act_count)
3373 bp->b_act_count = m->act_count;
3375 vm_object_drop(obj);
3378 * Step 2. We've loaded the pages into the buffer,
3379 * we have to figure out if we can still have B_CACHE
3380 * set. Note that B_CACHE is set according to the
3381 * byte-granular range ( bcount and size ), not the
3382 * aligned range ( newbsize ).
3384 * The VM test is against m->valid, which is DEV_BSIZE
3385 * aligned. Needless to say, the validity of the data
3386 * needs to also be DEV_BSIZE aligned. Note that this
3387 * fails with NFS if the server or some other client
3388 * extends the file's EOF. If our buffer is resized,
3389 * B_CACHE may remain set! XXX
3392 toff = bp->b_bcount;
3393 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3395 while ((bp->b_flags & B_CACHE) && toff < size) {
3398 if (tinc > (size - toff))
3401 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3409 bp->b_xio.xio_pages[pi]
3416 * Step 3, fixup the KVM pmap. Remember that
3417 * bp->b_data is relative to bp->b_loffset, but
3418 * bp->b_loffset may be offset into the first page.
3421 bp->b_data = (caddr_t)
3422 trunc_page((vm_offset_t)bp->b_data);
3424 (vm_offset_t)bp->b_data,
3425 bp->b_xio.xio_pages,
3426 bp->b_xio.xio_npages
3428 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3429 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3433 /* adjust space use on already-dirty buffer */
3434 if (bp->b_flags & B_DELWRI) {
3435 /* dirtykvaspace unchanged */
3436 atomic_add_long(&dirtybufspace, newbsize - bp->b_bufsize);
3437 if (bp->b_flags & B_HEAVY) {
3438 atomic_add_long(&dirtybufspacehw,
3439 newbsize - bp->b_bufsize);
3442 if (newbsize < bp->b_bufsize)
3444 bp->b_bufsize = newbsize; /* actual buffer allocation */
3445 bp->b_bcount = size; /* requested buffer size */
3452 * Wait for buffer I/O completion, returning error status. B_EINTR
3453 * is converted into an EINTR error but not cleared (since a chain
3454 * of biowait() calls may occur).
3456 * On return bpdone() will have been called but the buffer will remain
3457 * locked and will not have been brelse()'d.
3459 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3460 * likely still in progress on return.
3462 * NOTE! This operation is on a BIO, not a BUF.
3464 * NOTE! BIO_DONE is cleared by vn_strategy()
3467 _biowait(struct bio *bio, const char *wmesg, int to)
3469 struct buf *bp = bio->bio_buf;
3474 KKASSERT(bio == &bp->b_bio1);
3476 flags = bio->bio_flags;
3477 if (flags & BIO_DONE)
3479 nflags = flags | BIO_WANT;
3480 tsleep_interlock(bio, 0);
3481 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3483 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3484 else if (bp->b_cmd == BUF_CMD_READ)
3485 error = tsleep(bio, PINTERLOCKED, "biord", to);
3487 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3489 kprintf("tsleep error biowait %d\n", error);
3498 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3499 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3500 if (bp->b_flags & B_EINTR)
3502 if (bp->b_flags & B_ERROR)
3503 return (bp->b_error ? bp->b_error : EIO);
3508 biowait(struct bio *bio, const char *wmesg)
3510 return(_biowait(bio, wmesg, 0));
3514 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3516 return(_biowait(bio, wmesg, to));
3520 * This associates a tracking count with an I/O. vn_strategy() and
3521 * dev_dstrategy() do this automatically but there are a few cases
3522 * where a vnode or device layer is bypassed when a block translation
3523 * is cached. In such cases bio_start_transaction() may be called on
3524 * the bypassed layers so the system gets an I/O in progress indication
3525 * for those higher layers.
3528 bio_start_transaction(struct bio *bio, struct bio_track *track)
3530 bio->bio_track = track;
3531 if (dsched_is_clear_buf_priv(bio->bio_buf))
3532 dsched_new_buf(bio->bio_buf);
3533 bio_track_ref(track);
3537 * Initiate I/O on a vnode.
3539 * SWAPCACHE OPERATION:
3541 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately
3542 * devfs also uses b_vp for fake buffers so we also have to check
3543 * that B_PAGING is 0. In this case the passed 'vp' is probably the
3544 * underlying block device. The swap assignments are related to the
3545 * buffer cache buffer's b_vp, not the passed vp.
3547 * The passed vp == bp->b_vp only in the case where the strategy call
3548 * is made on the vp itself for its own buffers (a regular file or
3549 * block device vp). The filesystem usually then re-calls vn_strategy()
3550 * after translating the request to an underlying device.
3552 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the
3553 * underlying buffer cache buffers.
3555 * We can only deal with page-aligned buffers at the moment, because
3556 * we can't tell what the real dirty state for pages straddling a buffer
3559 * In order to call swap_pager_strategy() we must provide the VM object
3560 * and base offset for the underlying buffer cache pages so it can find
3564 vn_strategy(struct vnode *vp, struct bio *bio)
3566 struct bio_track *track;
3567 struct buf *bp = bio->bio_buf;
3569 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3572 * Set when an I/O is issued on the bp. Cleared by consumers
3573 * (aka HAMMER), allowing the consumer to determine if I/O had
3574 * actually occurred.
3576 bp->b_flags |= B_IODEBUG;
3579 * Handle the swap cache intercept.
3581 if (vn_cache_strategy(vp, bio))
3585 * Otherwise do the operation through the filesystem
3587 if (bp->b_cmd == BUF_CMD_READ)
3588 track = &vp->v_track_read;
3590 track = &vp->v_track_write;
3591 KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3592 bio->bio_track = track;
3593 if (dsched_is_clear_buf_priv(bio->bio_buf))
3594 dsched_new_buf(bio->bio_buf);
3595 bio_track_ref(track);
3596 vop_strategy(*vp->v_ops, vp, bio);
3599 static void vn_cache_strategy_callback(struct bio *bio);
3602 vn_cache_strategy(struct vnode *vp, struct bio *bio)
3604 struct buf *bp = bio->bio_buf;
3611 * Is this buffer cache buffer suitable for reading from
3614 if (vm_swapcache_read_enable == 0 ||
3615 bp->b_cmd != BUF_CMD_READ ||
3616 ((bp->b_flags & B_CLUSTER) == 0 &&
3617 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) ||
3618 ((int)bp->b_loffset & PAGE_MASK) != 0 ||
3619 (bp->b_bcount & PAGE_MASK) != 0) {
3624 * Figure out the original VM object (it will match the underlying
3625 * VM pages). Note that swap cached data uses page indices relative
3626 * to that object, not relative to bio->bio_offset.
3628 if (bp->b_flags & B_CLUSTER)
3629 object = vp->v_object;
3631 object = bp->b_vp->v_object;
3634 * In order to be able to use the swap cache all underlying VM
3635 * pages must be marked as such, and we can't have any bogus pages.
3637 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
3638 m = bp->b_xio.xio_pages[i];
3639 if ((m->flags & PG_SWAPPED) == 0)
3641 if (m == bogus_page)
3646 * If we are good then issue the I/O using swap_pager_strategy().
3648 * We can only do this if the buffer actually supports object-backed
3649 * I/O. If it doesn't npages will be 0.
3651 if (i && i == bp->b_xio.xio_npages) {
3652 m = bp->b_xio.xio_pages[0];
3653 nbio = push_bio(bio);
3654 nbio->bio_done = vn_cache_strategy_callback;
3655 nbio->bio_offset = ptoa(m->pindex);
3656 KKASSERT(m->object == object);
3657 swap_pager_strategy(object, nbio);
3664 * This is a bit of a hack but since the vn_cache_strategy() function can
3665 * override a VFS's strategy function we must make sure that the bio, which
3666 * is probably bio2, doesn't leak an unexpected offset value back to the
3667 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the
3668 * bio went through its own file strategy function and the the bio2 offset
3669 * is a cached disk offset when, in fact, it isn't.
3672 vn_cache_strategy_callback(struct bio *bio)
3674 bio->bio_offset = NOOFFSET;
3675 biodone(pop_bio(bio));
3681 * Finish I/O on a buffer after all BIOs have been processed.
3682 * Called when the bio chain is exhausted or by biowait. If called
3683 * by biowait, elseit is typically 0.
3685 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3686 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3687 * assuming B_INVAL is clear.
3689 * For the VMIO case, we set B_CACHE if the op was a read and no
3690 * read error occured, or if the op was a write. B_CACHE is never
3691 * set if the buffer is invalid or otherwise uncacheable.
3693 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3694 * initiator to leave B_INVAL set to brelse the buffer out of existance
3695 * in the biodone routine.
3698 bpdone(struct buf *bp, int elseit)
3702 KASSERT(BUF_REFCNTNB(bp) > 0,
3703 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3704 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3705 ("biodone: bp %p already done!", bp));
3708 * No more BIOs are left. All completion functions have been dealt
3709 * with, now we clean up the buffer.
3712 bp->b_cmd = BUF_CMD_DONE;
3715 * Only reads and writes are processed past this point.
3717 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3718 if (cmd == BUF_CMD_FREEBLKS)
3719 bp->b_flags |= B_NOCACHE;
3726 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3727 * a lot worse. XXX - move this above the clearing of b_cmd
3729 if (LIST_FIRST(&bp->b_dep) != NULL)
3733 * A failed write must re-dirty the buffer unless B_INVAL
3734 * was set. Only applicable to normal buffers (with VPs).
3735 * vinum buffers may not have a vp.
3737 if (cmd == BUF_CMD_WRITE &&
3738 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3739 bp->b_flags &= ~B_NOCACHE;
3744 if (bp->b_flags & B_VMIO) {
3750 struct vnode *vp = bp->b_vp;
3754 #if defined(VFS_BIO_DEBUG)
3755 if (vp->v_auxrefs == 0)
3756 panic("biodone: zero vnode hold count");
3757 if ((vp->v_flag & VOBJBUF) == 0)
3758 panic("biodone: vnode is not setup for merged cache");
3761 foff = bp->b_loffset;
3762 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3763 KASSERT(obj != NULL, ("biodone: missing VM object"));
3765 #if defined(VFS_BIO_DEBUG)
3766 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3767 kprintf("biodone: paging in progress(%d) < "
3768 "bp->b_xio.xio_npages(%d)\n",
3769 obj->paging_in_progress,
3770 bp->b_xio.xio_npages);
3775 * Set B_CACHE if the op was a normal read and no error
3776 * occured. B_CACHE is set for writes in the b*write()
3779 iosize = bp->b_bcount - bp->b_resid;
3780 if (cmd == BUF_CMD_READ &&
3781 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3782 bp->b_flags |= B_CACHE;
3785 vm_object_hold(obj);
3786 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3790 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3795 * cleanup bogus pages, restoring the originals. Since
3796 * the originals should still be wired, we don't have
3797 * to worry about interrupt/freeing races destroying
3798 * the VM object association.
3800 m = bp->b_xio.xio_pages[i];
3801 if (m == bogus_page) {
3803 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3805 panic("biodone: page disappeared");
3806 bp->b_xio.xio_pages[i] = m;
3807 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3808 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3810 #if defined(VFS_BIO_DEBUG)
3811 if (OFF_TO_IDX(foff) != m->pindex) {
3812 kprintf("biodone: foff(%lu)/m->pindex(%ld) "
3814 (unsigned long)foff, (long)m->pindex);
3819 * In the write case, the valid and clean bits are
3820 * already changed correctly (see bdwrite()), so we
3821 * only need to do this here in the read case.
3823 vm_page_busy_wait(m, FALSE, "bpdpgw");
3824 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3825 vfs_clean_one_page(bp, i, m);
3827 vm_page_flag_clear(m, PG_ZERO);
3830 * when debugging new filesystems or buffer I/O
3831 * methods, this is the most common error that pops
3832 * up. if you see this, you have not set the page
3833 * busy flag correctly!!!
3836 kprintf("biodone: page busy < 0, "
3837 "pindex: %d, foff: 0x(%x,%x), "
3838 "resid: %d, index: %d\n",
3839 (int) m->pindex, (int)(foff >> 32),
3840 (int) foff & 0xffffffff, resid, i);
3841 if (!vn_isdisk(vp, NULL))
3842 kprintf(" iosize: %ld, loffset: %lld, "
3843 "flags: 0x%08x, npages: %d\n",
3844 bp->b_vp->v_mount->mnt_stat.f_iosize,
3845 (long long)bp->b_loffset,
3846 bp->b_flags, bp->b_xio.xio_npages);
3848 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3849 (long long)bp->b_loffset,
3850 bp->b_flags, bp->b_xio.xio_npages);
3851 kprintf(" valid: 0x%x, dirty: 0x%x, "
3855 panic("biodone: page busy < 0");
3857 vm_page_io_finish(m);
3859 vm_object_pip_wakeup(obj);
3860 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3863 bp->b_flags &= ~B_HASBOGUS;
3864 vm_object_drop(obj);
3868 * Finish up by releasing the buffer. There are no more synchronous
3869 * or asynchronous completions, those were handled by bio_done
3873 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
3884 biodone(struct bio *bio)
3886 struct buf *bp = bio->bio_buf;
3888 runningbufwakeup(bp);
3891 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3894 biodone_t *done_func;
3895 struct bio_track *track;
3898 * BIO tracking. Most but not all BIOs are tracked.
3900 if ((track = bio->bio_track) != NULL) {
3901 bio_track_rel(track);
3902 bio->bio_track = NULL;
3906 * A bio_done function terminates the loop. The function
3907 * will be responsible for any further chaining and/or
3908 * buffer management.
3910 * WARNING! The done function can deallocate the buffer!
3912 if ((done_func = bio->bio_done) != NULL) {
3913 bio->bio_done = NULL;
3917 bio = bio->bio_prev;
3921 * If we've run out of bio's do normal [a]synchronous completion.
3927 * Synchronous biodone - this terminates a synchronous BIO.
3929 * bpdone() is called with elseit=FALSE, leaving the buffer completed
3930 * but still locked. The caller must brelse() the buffer after waiting
3934 biodone_sync(struct bio *bio)
3936 struct buf *bp = bio->bio_buf;
3940 KKASSERT(bio == &bp->b_bio1);
3944 flags = bio->bio_flags;
3945 nflags = (flags | BIO_DONE) & ~BIO_WANT;
3947 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3948 if (flags & BIO_WANT)
3958 * This routine is called in lieu of iodone in the case of
3959 * incomplete I/O. This keeps the busy status for pages
3963 vfs_unbusy_pages(struct buf *bp)
3967 runningbufwakeup(bp);
3969 if (bp->b_flags & B_VMIO) {
3970 struct vnode *vp = bp->b_vp;
3974 vm_object_hold(obj);
3976 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3977 vm_page_t m = bp->b_xio.xio_pages[i];
3980 * When restoring bogus changes the original pages
3981 * should still be wired, so we are in no danger of
3982 * losing the object association and do not need
3983 * critical section protection particularly.
3985 if (m == bogus_page) {
3986 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3988 panic("vfs_unbusy_pages: page missing");
3990 bp->b_xio.xio_pages[i] = m;
3991 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3992 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3994 vm_page_busy_wait(m, FALSE, "bpdpgw");
3995 vm_page_flag_clear(m, PG_ZERO);
3996 vm_page_io_finish(m);
3998 vm_object_pip_wakeup(obj);
4000 bp->b_flags &= ~B_HASBOGUS;
4001 vm_object_drop(obj);
4008 * This routine is called before a device strategy routine.
4009 * It is used to tell the VM system that paging I/O is in
4010 * progress, and treat the pages associated with the buffer
4011 * almost as being PG_BUSY. Also the object 'paging_in_progress'
4012 * flag is handled to make sure that the object doesn't become
4015 * Since I/O has not been initiated yet, certain buffer flags
4016 * such as B_ERROR or B_INVAL may be in an inconsistant state
4017 * and should be ignored.
4020 vfs_busy_pages(struct vnode *vp, struct buf *bp)
4023 struct lwp *lp = curthread->td_lwp;
4026 * The buffer's I/O command must already be set. If reading,
4027 * B_CACHE must be 0 (double check against callers only doing
4028 * I/O when B_CACHE is 0).
4030 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4031 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
4033 if (bp->b_flags & B_VMIO) {
4037 KASSERT(bp->b_loffset != NOOFFSET,
4038 ("vfs_busy_pages: no buffer offset"));
4041 * Busy all the pages. We have to busy them all at once
4042 * to avoid deadlocks.
4045 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4046 vm_page_t m = bp->b_xio.xio_pages[i];
4048 if (vm_page_busy_try(m, FALSE)) {
4049 vm_page_sleep_busy(m, FALSE, "vbpage");
4051 vm_page_wakeup(bp->b_xio.xio_pages[i]);
4057 * Setup for I/O, soft-busy the page right now because
4058 * the next loop may block.
4060 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4061 vm_page_t m = bp->b_xio.xio_pages[i];
4063 vm_page_flag_clear(m, PG_ZERO);
4064 if ((bp->b_flags & B_CLUSTER) == 0) {
4065 vm_object_pip_add(obj, 1);
4066 vm_page_io_start(m);
4071 * Adjust protections for I/O and do bogus-page mapping.
4072 * Assume that vm_page_protect() can block (it can block
4073 * if VM_PROT_NONE, don't take any chances regardless).
4075 * In particular note that for writes we must incorporate
4076 * page dirtyness from the VM system into the buffer's
4079 * For reads we theoretically must incorporate page dirtyness
4080 * from the VM system to determine if the page needs bogus
4081 * replacement, but we shortcut the test by simply checking
4082 * that all m->valid bits are set, indicating that the page
4083 * is fully valid and does not need to be re-read. For any
4084 * VM system dirtyness the page will also be fully valid
4085 * since it was mapped at one point.
4088 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4089 vm_page_t m = bp->b_xio.xio_pages[i];
4091 vm_page_flag_clear(m, PG_ZERO); /* XXX */
4092 if (bp->b_cmd == BUF_CMD_WRITE) {
4094 * When readying a vnode-backed buffer for
4095 * a write we must zero-fill any invalid
4096 * portions of the backing VM pages, mark
4097 * it valid and clear related dirty bits.
4099 * vfs_clean_one_page() incorporates any
4100 * VM dirtyness and updates the b_dirtyoff
4101 * range (after we've made the page RO).
4103 * It is also expected that the pmap modified
4104 * bit has already been cleared by the
4105 * vm_page_protect(). We may not be able
4106 * to clear all dirty bits for a page if it
4107 * was also memory mapped (NFS).
4109 * Finally be sure to unassign any swap-cache
4110 * backing store as it is now stale.
4112 vm_page_protect(m, VM_PROT_READ);
4113 vfs_clean_one_page(bp, i, m);
4114 swap_pager_unswapped(m);
4115 } else if (m->valid == VM_PAGE_BITS_ALL) {
4117 * When readying a vnode-backed buffer for
4118 * read we must replace any dirty pages with
4119 * a bogus page so dirty data is not destroyed
4120 * when filling gaps.
4122 * To avoid testing whether the page is
4123 * dirty we instead test that the page was
4124 * at some point mapped (m->valid fully
4125 * valid) with the understanding that
4126 * this also covers the dirty case.
4128 bp->b_xio.xio_pages[i] = bogus_page;
4129 bp->b_flags |= B_HASBOGUS;
4131 } else if (m->valid & m->dirty) {
4133 * This case should not occur as partial
4134 * dirtyment can only happen if the buffer
4135 * is B_CACHE, and this code is not entered
4136 * if the buffer is B_CACHE.
4138 kprintf("Warning: vfs_busy_pages - page not "
4139 "fully valid! loff=%jx bpf=%08x "
4140 "idx=%d val=%02x dir=%02x\n",
4141 (uintmax_t)bp->b_loffset, bp->b_flags,
4142 i, m->valid, m->dirty);
4143 vm_page_protect(m, VM_PROT_NONE);
4146 * The page is not valid and can be made
4149 vm_page_protect(m, VM_PROT_NONE);
4154 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4155 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
4160 * This is the easiest place to put the process accounting for the I/O
4164 if (bp->b_cmd == BUF_CMD_READ)
4165 lp->lwp_ru.ru_inblock++;
4167 lp->lwp_ru.ru_oublock++;
4172 * Tell the VM system that the pages associated with this buffer
4173 * are clean. This is used for delayed writes where the data is
4174 * going to go to disk eventually without additional VM intevention.
4176 * NOTE: While we only really need to clean through to b_bcount, we
4177 * just go ahead and clean through to b_bufsize.
4180 vfs_clean_pages(struct buf *bp)
4185 if ((bp->b_flags & B_VMIO) == 0)
4188 KASSERT(bp->b_loffset != NOOFFSET,
4189 ("vfs_clean_pages: no buffer offset"));
4191 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4192 m = bp->b_xio.xio_pages[i];
4193 vfs_clean_one_page(bp, i, m);
4198 * vfs_clean_one_page:
4200 * Set the valid bits and clear the dirty bits in a page within a
4201 * buffer. The range is restricted to the buffer's size and the
4202 * buffer's logical offset might index into the first page.
4204 * The caller has busied or soft-busied the page and it is not mapped,
4205 * test and incorporate the dirty bits into b_dirtyoff/end before
4206 * clearing them. Note that we need to clear the pmap modified bits
4207 * after determining the the page was dirty, vm_page_set_validclean()
4208 * does not do it for us.
4210 * This routine is typically called after a read completes (dirty should
4211 * be zero in that case as we are not called on bogus-replace pages),
4212 * or before a write is initiated.
4215 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m)
4223 * Calculate offset range within the page but relative to buffer's
4224 * loffset. loffset might be offset into the first page.
4226 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4227 bcount = bp->b_bcount + xoff; /* offset adjusted */
4233 soff = (pageno << PAGE_SHIFT);
4234 eoff = soff + PAGE_SIZE;
4242 * Test dirty bits and adjust b_dirtyoff/end.
4244 * If dirty pages are incorporated into the bp any prior
4245 * B_NEEDCOMMIT state (NFS) must be cleared because the
4246 * caller has not taken into account the new dirty data.
4248 * If the page was memory mapped the dirty bits might go beyond the
4249 * end of the buffer, but we can't really make the assumption that
4250 * a file EOF straddles the buffer (even though this is the case for
4251 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
4252 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
4253 * This also saves some console spam.
4255 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK,
4256 * NFS can handle huge commits but not huge writes.
4258 vm_page_test_dirty(m);
4260 if ((bp->b_flags & B_NEEDCOMMIT) &&
4261 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) {
4263 kprintf("Warning: vfs_clean_one_page: bp %p "
4264 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT"
4265 " cmd %d vd %02x/%02x x/s/e %d %d %d "
4267 bp, (uintmax_t)bp->b_loffset, bp->b_bcount,
4268 bp->b_flags, bp->b_cmd,
4269 m->valid, m->dirty, xoff, soff, eoff,
4270 bp->b_dirtyoff, bp->b_dirtyend);
4271 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
4273 print_backtrace(-1);
4276 * Only clear the pmap modified bits if ALL the dirty bits
4277 * are set, otherwise the system might mis-clear portions
4280 if (m->dirty == VM_PAGE_BITS_ALL &&
4281 (bp->b_flags & B_NEEDCOMMIT) == 0) {
4282 pmap_clear_modify(m);
4284 if (bp->b_dirtyoff > soff - xoff)
4285 bp->b_dirtyoff = soff - xoff;
4286 if (bp->b_dirtyend < eoff - xoff)
4287 bp->b_dirtyend = eoff - xoff;
4291 * Set related valid bits, clear related dirty bits.
4292 * Does not mess with the pmap modified bit.
4294 * WARNING! We cannot just clear all of m->dirty here as the
4295 * buffer cache buffers may use a DEV_BSIZE'd aligned
4296 * block size, or have an odd size (e.g. NFS at file EOF).
4297 * The putpages code can clear m->dirty to 0.
4299 * If a VOP_WRITE generates a buffer cache buffer which
4300 * covers the same space as mapped writable pages the
4301 * buffer flush might not be able to clear all the dirty
4302 * bits and still require a putpages from the VM system
4305 * WARNING! vm_page_set_validclean() currently assumes vm_token
4306 * is held. The page might not be busied (bdwrite() case).
4307 * XXX remove this comment once we've validated that this
4308 * is no longer an issue.
4310 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff);
4315 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty.
4316 * The page data is assumed to be valid (there is no zeroing here).
4319 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m)
4327 * Calculate offset range within the page but relative to buffer's
4328 * loffset. loffset might be offset into the first page.
4330 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4331 bcount = bp->b_bcount + xoff; /* offset adjusted */
4337 soff = (pageno << PAGE_SHIFT);
4338 eoff = soff + PAGE_SIZE;
4344 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff);
4351 * Clear a buffer. This routine essentially fakes an I/O, so we need
4352 * to clear B_ERROR and B_INVAL.
4354 * Note that while we only theoretically need to clear through b_bcount,
4355 * we go ahead and clear through b_bufsize.
4359 vfs_bio_clrbuf(struct buf *bp)
4363 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
4364 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
4365 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4366 (bp->b_loffset & PAGE_MASK) == 0) {
4367 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4368 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
4372 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
4373 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
4374 bzero(bp->b_data, bp->b_bufsize);
4375 bp->b_xio.xio_pages[0]->valid |= mask;
4381 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
4382 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
4383 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
4384 ea = (caddr_t)(vm_offset_t)ulmin(
4385 (u_long)(vm_offset_t)ea,
4386 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
4387 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4388 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
4390 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
4391 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
4395 for (; sa < ea; sa += DEV_BSIZE, j++) {
4396 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
4397 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
4398 bzero(sa, DEV_BSIZE);
4401 bp->b_xio.xio_pages[i]->valid |= mask;
4402 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
4411 * vm_hold_load_pages:
4413 * Load pages into the buffer's address space. The pages are
4414 * allocated from the kernel object in order to reduce interference
4415 * with the any VM paging I/O activity. The range of loaded
4416 * pages will be wired.
4418 * If a page cannot be allocated, the 'pagedaemon' is woken up to
4419 * retrieve the full range (to - from) of pages.
4422 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4428 to = round_page(to);
4429 from = round_page(from);
4430 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4435 * Note: must allocate system pages since blocking here
4436 * could intefere with paging I/O, no matter which
4439 vm_object_hold(&kernel_object);
4440 p = bio_page_alloc(bp, &kernel_object, pg >> PAGE_SHIFT,
4441 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
4442 vm_object_drop(&kernel_object);
4445 p->valid = VM_PAGE_BITS_ALL;
4446 vm_page_flag_clear(p, PG_ZERO);
4447 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
4448 bp->b_xio.xio_pages[index] = p;
4455 bp->b_xio.xio_npages = index;
4459 * Allocate a page for a buffer cache buffer.
4461 * If NULL is returned the caller is expected to retry (typically check if
4462 * the page already exists on retry before trying to allocate one).
4464 * NOTE! Low-memory handling is dealt with in b[q]relse(), not here. This
4465 * function will use the system reserve with the hope that the page
4466 * allocations can be returned to PQ_CACHE/PQ_FREE when the caller
4467 * is done with the buffer.
4469 * NOTE! However, TMPFS is a special case because flushing a dirty buffer
4470 * to TMPFS doesn't clean the page. For TMPFS, only the pagedaemon
4471 * is capable of retiring pages (to swap). For TMPFS we don't dig
4472 * into the system reserve because doing so could stall out pretty
4473 * much every process running on the system.
4477 bio_page_alloc(struct buf *bp, vm_object_t obj, vm_pindex_t pg, int deficit)
4479 int vmflags = VM_ALLOC_NORMAL | VM_ALLOC_NULL_OK;
4482 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj));
4485 * Try a normal allocation first.
4487 p = vm_page_alloc(obj, pg, vmflags);
4490 if (vm_page_lookup(obj, pg))
4492 vm_pageout_deficit += deficit;
4495 * Try again, digging into the system reserve.
4497 * Trying to recover pages from the buffer cache here can deadlock
4498 * against other threads trying to busy underlying pages so we
4499 * depend on the code in brelse() and bqrelse() to free/cache the
4500 * underlying buffer cache pages when memory is low.
4502 if (curthread->td_flags & TDF_SYSTHREAD)
4503 vmflags |= VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT;
4504 else if (bp->b_vp && bp->b_vp->v_tag == VT_TMPFS)
4507 vmflags |= VM_ALLOC_SYSTEM;
4509 /*recoverbufpages();*/
4510 p = vm_page_alloc(obj, pg, vmflags);
4513 if (vm_page_lookup(obj, pg))
4517 * Wait for memory to free up and try again
4519 if (vm_page_count_severe())
4521 vm_wait(hz / 20 + 1);
4523 p = vm_page_alloc(obj, pg, vmflags);
4526 if (vm_page_lookup(obj, pg))
4530 * Ok, now we are really in trouble.
4533 static struct krate biokrate = { .freq = 1 };
4534 krateprintf(&biokrate,
4535 "Warning: bio_page_alloc: memory exhausted "
4536 "during bufcache page allocation from %s\n",
4537 curthread->td_comm);
4539 if (curthread->td_flags & TDF_SYSTHREAD)
4540 vm_wait(hz / 20 + 1);
4542 vm_wait(hz / 2 + 1);
4547 * vm_hold_free_pages:
4549 * Return pages associated with the buffer back to the VM system.
4551 * The range of pages underlying the buffer's address space will
4552 * be unmapped and un-wired.
4555 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4559 int index, newnpages;
4561 from = round_page(from);
4562 to = round_page(to);
4563 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4566 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4567 p = bp->b_xio.xio_pages[index];
4568 if (p && (index < bp->b_xio.xio_npages)) {
4570 kprintf("vm_hold_free_pages: doffset: %lld, "
4572 (long long)bp->b_bio2.bio_offset,
4573 (long long)bp->b_loffset);
4575 bp->b_xio.xio_pages[index] = NULL;
4577 vm_page_busy_wait(p, FALSE, "vmhldpg");
4578 vm_page_unwire(p, 0);
4582 bp->b_xio.xio_npages = newnpages;
4588 * Map a user buffer into KVM via a pbuf. On return the buffer's
4589 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4593 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4604 * bp had better have a command and it better be a pbuf.
4606 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4607 KKASSERT(bp->b_flags & B_PAGING);
4608 KKASSERT(bp->b_kvabase);
4614 * Map the user data into KVM. Mappings have to be page-aligned.
4616 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4619 vmprot = VM_PROT_READ;
4620 if (bp->b_cmd == BUF_CMD_READ)
4621 vmprot |= VM_PROT_WRITE;
4623 while (addr < udata + bytes) {
4625 * Do the vm_fault if needed; do the copy-on-write thing
4626 * when reading stuff off device into memory.
4628 * vm_fault_page*() returns a held VM page.
4630 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4631 va = trunc_page(va);
4633 m = vm_fault_page_quick(va, vmprot, &error);
4635 for (i = 0; i < pidx; ++i) {
4636 vm_page_unhold(bp->b_xio.xio_pages[i]);
4637 bp->b_xio.xio_pages[i] = NULL;
4641 bp->b_xio.xio_pages[pidx] = m;
4647 * Map the page array and set the buffer fields to point to
4648 * the mapped data buffer.
4650 if (pidx > btoc(MAXPHYS))
4651 panic("vmapbuf: mapped more than MAXPHYS");
4652 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4654 bp->b_xio.xio_npages = pidx;
4655 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4656 bp->b_bcount = bytes;
4657 bp->b_bufsize = bytes;
4664 * Free the io map PTEs associated with this IO operation.
4665 * We also invalidate the TLB entries and restore the original b_addr.
4668 vunmapbuf(struct buf *bp)
4673 KKASSERT(bp->b_flags & B_PAGING);
4675 npages = bp->b_xio.xio_npages;
4676 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4677 for (pidx = 0; pidx < npages; ++pidx) {
4678 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4679 bp->b_xio.xio_pages[pidx] = NULL;
4681 bp->b_xio.xio_npages = 0;
4682 bp->b_data = bp->b_kvabase;
4686 * Scan all buffers in the system and issue the callback.
4689 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4695 for (n = 0; n < nbuf; ++n) {
4696 if ((error = callback(&buf[n], info)) < 0) {
4706 * nestiobuf_iodone: biodone callback for nested buffers and propagate
4707 * completion to the master buffer.
4710 nestiobuf_iodone(struct bio *bio)
4713 struct buf *mbp, *bp;
4714 struct devstat *stats;
4719 mbio = bio->bio_caller_info1.ptr;
4720 stats = bio->bio_caller_info2.ptr;
4721 mbp = mbio->bio_buf;
4723 KKASSERT(bp->b_bcount <= bp->b_bufsize);
4724 KKASSERT(mbp != bp);
4726 error = bp->b_error;
4727 if (bp->b_error == 0 &&
4728 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
4730 * Not all got transfered, raise an error. We have no way to
4731 * propagate these conditions to mbp.
4736 donebytes = bp->b_bufsize;
4740 nestiobuf_done(mbio, donebytes, error, stats);
4744 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats)
4748 mbp = mbio->bio_buf;
4750 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0);
4753 * If an error occured, propagate it to the master buffer.
4755 * Several biodone()s may wind up running concurrently so
4756 * use an atomic op to adjust b_flags.
4759 mbp->b_error = error;
4760 atomic_set_int(&mbp->b_flags, B_ERROR);
4764 * Decrement the operations in progress counter and terminate the
4765 * I/O if this was the last bit.
4767 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4770 devstat_end_transaction_buf(stats, mbp);
4776 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent
4777 * the mbio from being biodone()'d while we are still adding sub-bios to
4781 nestiobuf_init(struct bio *bio)
4783 bio->bio_driver_info = (void *)1;
4787 * The BIOs added to the nestedio have already been started, remove the
4788 * count that placeheld our mbio and biodone() it if the count would
4792 nestiobuf_start(struct bio *mbio)
4794 struct buf *mbp = mbio->bio_buf;
4797 * Decrement the operations in progress counter and terminate the
4798 * I/O if this was the last bit.
4800 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4801 if (mbp->b_flags & B_ERROR)
4802 mbp->b_resid = mbp->b_bcount;
4810 * Set an intermediate error prior to calling nestiobuf_start()
4813 nestiobuf_error(struct bio *mbio, int error)
4815 struct buf *mbp = mbio->bio_buf;
4818 mbp->b_error = error;
4819 atomic_set_int(&mbp->b_flags, B_ERROR);
4824 * nestiobuf_add: setup a "nested" buffer.
4826 * => 'mbp' is a "master" buffer which is being divided into sub pieces.
4827 * => 'bp' should be a buffer allocated by getiobuf.
4828 * => 'offset' is a byte offset in the master buffer.
4829 * => 'size' is a size in bytes of this nested buffer.
4832 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats)
4834 struct buf *mbp = mbio->bio_buf;
4835 struct vnode *vp = mbp->b_vp;
4837 KKASSERT(mbp->b_bcount >= offset + size);
4839 atomic_add_int((int *)&mbio->bio_driver_info, 1);
4841 /* kernel needs to own the lock for it to be released in biodone */
4844 bp->b_cmd = mbp->b_cmd;
4845 bp->b_bio1.bio_done = nestiobuf_iodone;
4846 bp->b_data = (char *)mbp->b_data + offset;
4847 bp->b_resid = bp->b_bcount = size;
4848 bp->b_bufsize = bp->b_bcount;
4850 bp->b_bio1.bio_track = NULL;
4851 bp->b_bio1.bio_caller_info1.ptr = mbio;
4852 bp->b_bio1.bio_caller_info2.ptr = stats;
4857 DB_SHOW_COMMAND(buffer, db_show_buffer)
4860 struct buf *bp = (struct buf *)addr;
4863 db_printf("usage: show buffer <addr>\n");
4867 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4868 db_printf("b_cmd = %d\n", bp->b_cmd);
4869 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4870 "b_resid = %d\n, b_data = %p, "
4871 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4872 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4874 (long long)bp->b_bio2.bio_offset,
4875 (long long)(bp->b_bio2.bio_next ?
4876 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4877 if (bp->b_xio.xio_npages) {
4879 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4880 bp->b_xio.xio_npages);
4881 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4883 m = bp->b_xio.xio_pages[i];
4884 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4885 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4886 if ((i + 1) < bp->b_xio.xio_npages)