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. Note that man buf(9) doesn't reflect
28 * the actual buf/bio implementation in DragonFly.
31 #include <sys/param.h>
32 #include <sys/systm.h>
35 #include <sys/devicestat.h>
36 #include <sys/eventhandler.h>
38 #include <sys/malloc.h>
39 #include <sys/mount.h>
40 #include <sys/kernel.h>
41 #include <sys/kthread.h>
43 #include <sys/reboot.h>
44 #include <sys/resourcevar.h>
45 #include <sys/sysctl.h>
46 #include <sys/vmmeter.h>
47 #include <sys/vnode.h>
48 #include <sys/dsched.h>
50 #include <vm/vm_param.h>
51 #include <vm/vm_kern.h>
52 #include <vm/vm_pageout.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_map.h>
57 #include <vm/vm_pager.h>
58 #include <vm/swap_pager.h>
61 #include <sys/spinlock2.h>
62 #include <vm/vm_page2.h>
73 BQUEUE_NONE, /* not on any queue */
74 BQUEUE_LOCKED, /* locked buffers */
75 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
76 BQUEUE_DIRTY, /* B_DELWRI buffers */
77 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
78 BQUEUE_EMPTY, /* empty buffer headers */
80 BUFFER_QUEUES /* number of buffer queues */
83 typedef enum bufq_type bufq_type_t;
85 #define BD_WAKE_SIZE 16384
86 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
88 TAILQ_HEAD(bqueues, buf);
92 struct bqueues bufqueues[BUFFER_QUEUES];
95 struct bufpcpu bufpcpu[MAXCPU];
97 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
99 struct buf *buf; /* buffer header pool */
101 static void vfs_clean_pages(struct buf *bp);
102 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
104 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m);
106 static void vfs_vmio_release(struct buf *bp);
107 static int flushbufqueues(struct buf *marker, bufq_type_t q);
108 static vm_page_t bio_page_alloc(struct buf *bp, vm_object_t obj,
109 vm_pindex_t pg, int deficit);
111 static void bd_signal(long totalspace);
112 static void buf_daemon(void);
113 static void buf_daemon_hw(void);
116 * bogus page -- for I/O to/from partially complete buffers
117 * this is a temporary solution to the problem, but it is not
118 * really that bad. it would be better to split the buffer
119 * for input in the case of buffers partially already in memory,
120 * but the code is intricate enough already.
122 vm_page_t bogus_page;
125 * These are all static, but make the ones we export globals so we do
126 * not need to use compiler magic.
128 long bufspace; /* atomic ops */
130 long maxbufmallocspace, lobufspace, hibufspace;
131 static long lorunningspace;
132 static long hirunningspace;
133 static long dirtykvaspace; /* atomic */
134 long dirtybufspace; /* atomic (global for systat) */
135 static long dirtybufcount; /* atomic */
136 static long dirtybufspacehw; /* atomic */
137 static long dirtybufcounthw; /* atomic */
138 static long runningbufspace; /* atomic */
139 static long runningbufcount; /* atomic */
140 long lodirtybufspace;
141 long hidirtybufspace;
142 static int getnewbufcalls;
143 static int needsbuffer; /* atomic */
144 static int runningbufreq; /* atomic */
145 static int bd_request; /* atomic */
146 static int bd_request_hw; /* atomic */
147 static u_int bd_wake_ary[BD_WAKE_SIZE];
148 static u_int bd_wake_index;
149 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */
150 static int debug_commit;
151 static int debug_bufbio;
152 static int debug_kvabio;
153 static long bufcache_bw = 200 * 1024 * 1024;
155 static struct thread *bufdaemon_td;
156 static struct thread *bufdaemonhw_td;
157 static u_int lowmempgallocs;
158 static u_int flushperqueue = 1024;
161 * Sysctls for operational control of the buffer cache.
163 SYSCTL_UINT(_vfs, OID_AUTO, flushperqueue, CTLFLAG_RW, &flushperqueue, 0,
164 "Number of buffers to flush from each per-cpu queue");
165 SYSCTL_LONG(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
166 "Number of dirty buffers to flush before bufdaemon becomes inactive");
167 SYSCTL_LONG(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
168 "High watermark used to trigger explicit flushing of dirty buffers");
169 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
170 "Minimum amount of buffer space required for active I/O");
171 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
172 "Maximum amount of buffer space to usable for active I/O");
173 SYSCTL_LONG(_vfs, OID_AUTO, bufcache_bw, CTLFLAG_RW, &bufcache_bw, 0,
174 "Buffer-cache -> VM page cache transfer bandwidth");
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, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
178 "Recycle pages to active or inactive queue transition pt 0-64");
180 * Sysctls determining current state of the buffer cache.
182 SYSCTL_LONG(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
183 "Total number of buffers in buffer cache");
184 SYSCTL_LONG(_vfs, OID_AUTO, dirtykvaspace, CTLFLAG_RD, &dirtykvaspace, 0,
185 "KVA reserved by dirty buffers (all)");
186 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
187 "Pending bytes of dirty buffers (all)");
188 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
189 "Pending bytes of dirty buffers (heavy weight)");
190 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
191 "Pending number of dirty buffers");
192 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
193 "Pending number of dirty buffers (heavy weight)");
194 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
195 "I/O bytes currently in progress due to asynchronous writes");
196 SYSCTL_LONG(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
197 "I/O buffers currently in progress due to asynchronous writes");
198 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
199 "Hard limit on maximum amount of memory usable for buffer space");
200 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
201 "Soft limit on maximum amount of memory usable for buffer space");
202 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
203 "Minimum amount of memory to reserve for system buffer space");
204 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
205 "Amount of memory available for buffers");
206 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
207 0, "Maximum amount of memory reserved for buffers using malloc");
208 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
209 "New buffer header acquisition requests");
210 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, "");
211 SYSCTL_INT(_vfs, OID_AUTO, debug_bufbio, CTLFLAG_RW, &debug_bufbio, 0, "");
212 SYSCTL_INT(_vfs, OID_AUTO, debug_kvabio, CTLFLAG_RW, &debug_kvabio, 0, "");
213 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
214 "sizeof(struct buf)");
216 char *buf_wmesg = BUF_WMESG;
218 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
219 #define VFS_BIO_NEED_UNUSED02 0x02
220 #define VFS_BIO_NEED_UNUSED04 0x04
221 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
224 * Called when buffer space is potentially available for recovery.
225 * getnewbuf() will block on this flag when it is unable to free
226 * sufficient buffer space. Buffer space becomes recoverable when
227 * bp's get placed back in the queues.
233 * If someone is waiting for BUF space, wake them up. Even
234 * though we haven't freed the kva space yet, the waiting
235 * process will be able to now.
238 int flags = needsbuffer;
240 if ((flags & VFS_BIO_NEED_BUFSPACE) == 0)
242 if (atomic_cmpset_int(&needsbuffer, flags,
243 flags & ~VFS_BIO_NEED_BUFSPACE)) {
244 wakeup(&needsbuffer);
254 * Accounting for I/O in progress.
258 runningbufwakeup(struct buf *bp)
263 if ((totalspace = bp->b_runningbufspace) != 0) {
264 atomic_add_long(&runningbufspace, -totalspace);
265 atomic_add_long(&runningbufcount, -1);
266 bp->b_runningbufspace = 0;
269 * see waitrunningbufspace() for limit test.
272 flags = runningbufreq;
276 if (atomic_cmpset_int(&runningbufreq, flags, 0)) {
277 wakeup(&runningbufreq);
282 bd_signal(totalspace);
289 * Called when a buffer has been added to one of the free queues to
290 * account for the buffer and to wakeup anyone waiting for free buffers.
291 * This typically occurs when large amounts of metadata are being handled
292 * by the buffer cache ( else buffer space runs out first, usually ).
303 if (atomic_cmpset_int(&needsbuffer, flags,
304 (flags & ~VFS_BIO_NEED_ANY))) {
305 wakeup(&needsbuffer);
313 * waitrunningbufspace()
315 * If runningbufspace exceeds 4/6 hirunningspace we block until
316 * runningbufspace drops to 3/6 hirunningspace. We also block if another
317 * thread blocked here in order to be fair, even if runningbufspace
318 * is now lower than the limit.
320 * The caller may be using this function to block in a tight loop, we
321 * must block while runningbufspace is greater than at least
322 * hirunningspace * 3 / 6.
325 waitrunningbufspace(void)
327 long limit = hirunningspace * 4 / 6;
330 while (runningbufspace > limit || runningbufreq) {
331 tsleep_interlock(&runningbufreq, 0);
332 flags = atomic_fetchadd_int(&runningbufreq, 1);
333 if (runningbufspace > limit || flags)
334 tsleep(&runningbufreq, PINTERLOCKED, "wdrn1", hz);
339 * buf_dirty_count_severe:
341 * Return true if we have too many dirty buffers.
344 buf_dirty_count_severe(void)
346 return (runningbufspace + dirtykvaspace >= hidirtybufspace ||
347 dirtybufcount >= nbuf / 2);
351 * Return true if the amount of running I/O is severe and BIOQ should
355 buf_runningbufspace_severe(void)
357 return (runningbufspace >= hirunningspace * 4 / 6);
361 * vfs_buf_test_cache:
363 * Called when a buffer is extended. This function clears the B_CACHE
364 * bit if the newly extended portion of the buffer does not contain
367 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
368 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
369 * them while a clean buffer was present.
373 vfs_buf_test_cache(struct buf *bp,
374 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
377 if (bp->b_flags & B_CACHE) {
378 int base = (foff + off) & PAGE_MASK;
379 if (vm_page_is_valid(m, base, size) == 0)
380 bp->b_flags &= ~B_CACHE;
387 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
394 if (dirtykvaspace < lodirtybufspace && dirtybufcount < nbuf / 2)
397 if (bd_request == 0 &&
398 (dirtykvaspace > lodirtybufspace / 2 ||
399 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
400 if (atomic_fetchadd_int(&bd_request, 1) == 0)
403 if (bd_request_hw == 0 &&
404 (dirtykvaspace > lodirtybufspace / 2 ||
405 dirtybufcounthw >= nbuf / 2)) {
406 if (atomic_fetchadd_int(&bd_request_hw, 1) == 0)
407 wakeup(&bd_request_hw);
414 * Get the buf_daemon heated up when the number of running and dirty
415 * buffers exceeds the mid-point.
417 * Return the total number of dirty bytes past the second mid point
418 * as a measure of how much excess dirty data there is in the system.
427 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
429 totalspace = runningbufspace + dirtykvaspace;
430 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
432 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
433 if (totalspace >= mid2)
434 return(totalspace - mid2);
442 * Wait for the buffer cache to flush (totalspace) bytes worth of
443 * buffers, then return.
445 * Regardless this function blocks while the number of dirty buffers
446 * exceeds hidirtybufspace.
449 bd_wait(long totalspace)
456 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
459 while (totalspace > 0) {
463 * Order is important. Suppliers adjust bd_wake_index after
464 * updating runningbufspace/dirtykvaspace. We want to fetch
465 * bd_wake_index before accessing. Any error should thus
468 i = atomic_fetchadd_int(&bd_wake_index, 0);
469 if (totalspace > runningbufspace + dirtykvaspace)
470 totalspace = runningbufspace + dirtykvaspace;
471 count = totalspace / MAXBSIZE;
472 if (count >= BD_WAKE_SIZE / 2)
473 count = BD_WAKE_SIZE / 2;
475 mi = i & BD_WAKE_MASK;
478 * This is not a strict interlock, so we play a bit loose
479 * with locking access to dirtybufspace*. We have to re-check
480 * bd_wake_index to ensure that it hasn't passed us.
482 tsleep_interlock(&bd_wake_ary[mi], 0);
483 atomic_add_int(&bd_wake_ary[mi], 1);
484 j = atomic_fetchadd_int(&bd_wake_index, 0);
485 if ((int)(i - j) >= 0)
486 tsleep(&bd_wake_ary[mi], PINTERLOCKED, "flstik", hz);
488 totalspace = runningbufspace + dirtykvaspace - hidirtybufspace;
495 * This function is called whenever runningbufspace or dirtykvaspace
496 * is reduced. Track threads waiting for run+dirty buffer I/O
500 bd_signal(long totalspace)
504 if (totalspace > 0) {
505 if (totalspace > MAXBSIZE * BD_WAKE_SIZE)
506 totalspace = MAXBSIZE * BD_WAKE_SIZE;
507 while (totalspace > 0) {
508 i = atomic_fetchadd_int(&bd_wake_index, 1);
510 if (atomic_readandclear_int(&bd_wake_ary[i]))
511 wakeup(&bd_wake_ary[i]);
512 totalspace -= MAXBSIZE;
518 * BIO tracking support routines.
520 * Release a ref on a bio_track. Wakeup requests are atomically released
521 * along with the last reference so bk_active will never wind up set to
526 bio_track_rel(struct bio_track *track)
534 active = track->bk_active;
535 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
539 * Full-on. Note that the wait flag is only atomically released on
540 * the 1->0 count transition.
542 * We check for a negative count transition using bit 30 since bit 31
543 * has a different meaning.
546 desired = (active & 0x7FFFFFFF) - 1;
548 desired |= active & 0x80000000;
549 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
550 if (desired & 0x40000000)
551 panic("bio_track_rel: bad count: %p", track);
552 if (active & 0x80000000)
556 active = track->bk_active;
561 * Wait for the tracking count to reach 0.
563 * Use atomic ops such that the wait flag is only set atomically when
564 * bk_active is non-zero.
567 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
576 if (track->bk_active == 0)
580 * Full-on. Note that the wait flag may only be atomically set if
581 * the active count is non-zero.
583 * NOTE: We cannot optimize active == desired since a wakeup could
584 * clear active prior to our tsleep_interlock().
587 while ((active = track->bk_active) != 0) {
589 desired = active | 0x80000000;
590 tsleep_interlock(track, slp_flags);
591 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
592 error = tsleep(track, slp_flags | PINTERLOCKED,
604 * Load time initialisation of the buffer cache, called from machine
605 * dependant initialization code.
609 bufinit(void *dummy __unused)
611 struct bufpcpu *pcpu;
613 vm_offset_t bogus_offset;
618 /* next, make a null set of free lists */
619 for (i = 0; i < ncpus; ++i) {
621 spin_init(&pcpu->spin, "bufinit");
622 for (j = 0; j < BUFFER_QUEUES; j++)
623 TAILQ_INIT(&pcpu->bufqueues[j]);
627 * Finally, initialize each buffer header and stick on empty q.
628 * Each buffer gets its own KVA reservation.
633 for (n = 0; n < nbuf; n++) {
635 bzero(bp, sizeof *bp);
636 bp->b_flags = B_INVAL; /* we're just an empty header */
637 bp->b_cmd = BUF_CMD_DONE;
638 bp->b_qindex = BQUEUE_EMPTY;
640 bp->b_kvabase = (void *)(vm_map_min(&buffer_map) +
642 bp->b_kvasize = MAXBSIZE;
644 xio_init(&bp->b_xio);
646 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
654 * maxbufspace is the absolute maximum amount of buffer space we are
655 * allowed to reserve in KVM and in real terms. The absolute maximum
656 * is nominally used by buf_daemon. hibufspace is the nominal maximum
657 * used by most other processes. The differential is required to
658 * ensure that buf_daemon is able to run when other processes might
659 * be blocked waiting for buffer space.
661 * Calculate hysteresis (lobufspace, hibufspace). Don't make it
662 * too large or we might lockup a cpu for too long a period of
663 * time in our tight loop.
665 maxbufspace = nbuf * NBUFCALCSIZE;
666 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
667 lobufspace = hibufspace * 7 / 8;
668 if (hibufspace - lobufspace > 64 * 1024 * 1024)
669 lobufspace = hibufspace - 64 * 1024 * 1024;
670 if (lobufspace > hibufspace - MAXBSIZE)
671 lobufspace = hibufspace - MAXBSIZE;
673 lorunningspace = 512 * 1024;
674 /* hirunningspace -- see below */
677 * Limit the amount of malloc memory since it is wired permanently
678 * into the kernel space. Even though this is accounted for in
679 * the buffer allocation, we don't want the malloced region to grow
680 * uncontrolled. The malloc scheme improves memory utilization
681 * significantly on average (small) directories.
683 maxbufmallocspace = hibufspace / 20;
686 * Reduce the chance of a deadlock occuring by limiting the number
687 * of delayed-write dirty buffers we allow to stack up.
689 * We don't want too much actually queued to the device at once
690 * (XXX this needs to be per-mount!), because the buffers will
691 * wind up locked for a very long period of time while the I/O
694 hidirtybufspace = hibufspace / 2; /* dirty + running */
695 hirunningspace = hibufspace / 16; /* locked & queued to device */
696 if (hirunningspace < 1024 * 1024)
697 hirunningspace = 1024 * 1024;
703 lodirtybufspace = hidirtybufspace / 2;
706 * Maximum number of async ops initiated per buf_daemon loop. This is
707 * somewhat of a hack at the moment, we really need to limit ourselves
708 * based on the number of bytes of I/O in-transit that were initiated
712 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE,
714 vm_object_hold(&kernel_object);
715 bogus_page = vm_page_alloc(&kernel_object,
716 (bogus_offset >> PAGE_SHIFT),
718 vm_object_drop(&kernel_object);
719 vmstats.v_wire_count++;
723 SYSINIT(do_bufinit, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, bufinit, NULL);
726 * Initialize the embedded bio structures, typically used by
727 * deprecated code which tries to allocate its own struct bufs.
730 initbufbio(struct buf *bp)
732 bp->b_bio1.bio_buf = bp;
733 bp->b_bio1.bio_prev = NULL;
734 bp->b_bio1.bio_offset = NOOFFSET;
735 bp->b_bio1.bio_next = &bp->b_bio2;
736 bp->b_bio1.bio_done = NULL;
737 bp->b_bio1.bio_flags = 0;
739 bp->b_bio2.bio_buf = bp;
740 bp->b_bio2.bio_prev = &bp->b_bio1;
741 bp->b_bio2.bio_offset = NOOFFSET;
742 bp->b_bio2.bio_next = NULL;
743 bp->b_bio2.bio_done = NULL;
744 bp->b_bio2.bio_flags = 0;
750 * Reinitialize the embedded bio structures as well as any additional
751 * translation cache layers.
754 reinitbufbio(struct buf *bp)
758 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
759 bio->bio_done = NULL;
760 bio->bio_offset = NOOFFSET;
765 * Undo the effects of an initbufbio().
768 uninitbufbio(struct buf *bp)
775 * Push another BIO layer onto an existing BIO and return it. The new
776 * BIO layer may already exist, holding cached translation data.
779 push_bio(struct bio *bio)
783 if ((nbio = bio->bio_next) == NULL) {
784 int index = bio - &bio->bio_buf->b_bio_array[0];
785 if (index >= NBUF_BIO - 1) {
786 panic("push_bio: too many layers %d for bp %p",
787 index, bio->bio_buf);
789 nbio = &bio->bio_buf->b_bio_array[index + 1];
790 bio->bio_next = nbio;
791 nbio->bio_prev = bio;
792 nbio->bio_buf = bio->bio_buf;
793 nbio->bio_offset = NOOFFSET;
794 nbio->bio_done = NULL;
795 nbio->bio_next = NULL;
797 KKASSERT(nbio->bio_done == NULL);
802 * Pop a BIO translation layer, returning the previous layer. The
803 * must have been previously pushed.
806 pop_bio(struct bio *bio)
808 return(bio->bio_prev);
812 clearbiocache(struct bio *bio)
815 bio->bio_offset = NOOFFSET;
821 * Remove the buffer from the appropriate free list.
822 * (caller must be locked)
825 _bremfree(struct buf *bp)
827 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu];
829 if (bp->b_qindex != BQUEUE_NONE) {
830 KASSERT(BUF_LOCKINUSE(bp), ("bremfree: bp %p not locked", bp));
831 TAILQ_REMOVE(&pcpu->bufqueues[bp->b_qindex], bp, b_freelist);
832 bp->b_qindex = BQUEUE_NONE;
834 if (!BUF_LOCKINUSE(bp))
835 panic("bremfree: removing a buffer not on a queue");
840 * bremfree() - must be called with a locked buffer
843 bremfree(struct buf *bp)
845 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu];
847 spin_lock(&pcpu->spin);
849 spin_unlock(&pcpu->spin);
853 * bremfree_locked - must be called with pcpu->spin locked
856 bremfree_locked(struct buf *bp)
862 * This version of bread issues any required I/O asyncnronously and
863 * makes a callback on completion.
865 * The callback must check whether BIO_DONE is set in the bio and issue
866 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing
867 * BIO_DONE and disposing of the I/O (bqrelse()ing it).
870 breadcb(struct vnode *vp, off_t loffset, int size, int bflags,
871 void (*func)(struct bio *), void *arg)
875 bp = getblk(vp, loffset, size, 0, 0);
877 /* if not found in cache, do some I/O */
878 if ((bp->b_flags & B_CACHE) == 0) {
879 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL | B_NOTMETA);
880 bp->b_flags |= bflags;
881 bp->b_cmd = BUF_CMD_READ;
882 bp->b_bio1.bio_done = func;
883 bp->b_bio1.bio_caller_info1.ptr = arg;
884 vfs_busy_pages(vp, bp);
886 vn_strategy(vp, &bp->b_bio1);
889 * Since we are issuing the callback synchronously it cannot
890 * race the BIO_DONE, so no need for atomic ops here.
892 /*bp->b_bio1.bio_done = func;*/
893 bp->b_bio1.bio_caller_info1.ptr = arg;
894 bp->b_bio1.bio_flags |= BIO_DONE;
902 * breadnx() - Terminal function for bread() and breadn().
904 * This function will start asynchronous I/O on read-ahead blocks as well
905 * as satisfy the primary request.
907 * We must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE is
908 * set, the buffer is valid and we do not have to do anything.
911 breadnx(struct vnode *vp, off_t loffset, int size, int bflags,
912 off_t *raoffset, int *rabsize,
913 int cnt, struct buf **bpp)
915 struct buf *bp, *rabp;
917 int rv = 0, readwait = 0;
918 int blkflags = (bflags & B_KVABIO) ? GETBLK_KVABIO : 0;
923 *bpp = bp = getblk(vp, loffset, size, blkflags, 0);
925 /* if not found in cache, do some I/O */
926 if ((bp->b_flags & B_CACHE) == 0) {
927 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL | B_NOTMETA);
928 bp->b_flags |= bflags;
929 bp->b_cmd = BUF_CMD_READ;
930 bp->b_bio1.bio_done = biodone_sync;
931 bp->b_bio1.bio_flags |= BIO_SYNC;
932 vfs_busy_pages(vp, bp);
933 vn_strategy(vp, &bp->b_bio1);
937 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
938 if (inmem(vp, *raoffset))
940 rabp = getblk(vp, *raoffset, *rabsize, GETBLK_KVABIO, 0);
942 if ((rabp->b_flags & B_CACHE) == 0) {
943 rabp->b_flags &= ~(B_ERROR | B_EINTR |
944 B_INVAL | B_NOTMETA);
945 rabp->b_flags |= (bflags & ~B_KVABIO);
946 rabp->b_cmd = BUF_CMD_READ;
947 vfs_busy_pages(vp, rabp);
949 vn_strategy(vp, &rabp->b_bio1);
955 rv = biowait(&bp->b_bio1, "biord");
962 * Synchronous write, waits for completion.
964 * Write, release buffer on completion. (Done by iodone
965 * if async). Do not bother writing anything if the buffer
968 * Note that we set B_CACHE here, indicating that buffer is
969 * fully valid and thus cacheable. This is true even of NFS
970 * now so we set it generally. This could be set either here
971 * or in biodone() since the I/O is synchronous. We put it
975 bwrite(struct buf *bp)
979 if (bp->b_flags & B_INVAL) {
983 if (BUF_LOCKINUSE(bp) == 0)
984 panic("bwrite: buffer is not busy???");
987 * NOTE: We no longer mark the buffer clear prior to the vn_strategy()
988 * call because it will remove the buffer from the vnode's
989 * dirty buffer list prematurely and possibly cause filesystem
990 * checks to race buffer flushes. This is now handled in
993 * bundirty(bp); REMOVED
996 bp->b_flags &= ~(B_ERROR | B_EINTR);
997 bp->b_flags |= B_CACHE;
998 bp->b_cmd = BUF_CMD_WRITE;
1000 bp->b_bio1.bio_done = biodone_sync;
1001 bp->b_bio1.bio_flags |= BIO_SYNC;
1002 vfs_busy_pages(bp->b_vp, bp);
1005 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1006 * valid for vnode-backed buffers.
1008 bsetrunningbufspace(bp, bp->b_bufsize);
1009 vn_strategy(bp->b_vp, &bp->b_bio1);
1010 error = biowait(&bp->b_bio1, "biows");
1019 * Asynchronous write. Start output on a buffer, but do not wait for
1020 * it to complete. The buffer is released when the output completes.
1022 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1023 * B_INVAL buffers. Not us.
1026 bawrite(struct buf *bp)
1028 if (bp->b_flags & B_INVAL) {
1032 if (BUF_LOCKINUSE(bp) == 0)
1033 panic("bawrite: buffer is not busy???");
1036 * NOTE: We no longer mark the buffer clear prior to the vn_strategy()
1037 * call because it will remove the buffer from the vnode's
1038 * dirty buffer list prematurely and possibly cause filesystem
1039 * checks to race buffer flushes. This is now handled in
1042 * bundirty(bp); REMOVED
1044 bp->b_flags &= ~(B_ERROR | B_EINTR);
1045 bp->b_flags |= B_CACHE;
1046 bp->b_cmd = BUF_CMD_WRITE;
1048 KKASSERT(bp->b_bio1.bio_done == NULL);
1049 vfs_busy_pages(bp->b_vp, bp);
1052 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1053 * valid for vnode-backed buffers.
1055 bsetrunningbufspace(bp, bp->b_bufsize);
1057 vn_strategy(bp->b_vp, &bp->b_bio1);
1063 * Delayed write. (Buffer is marked dirty). Do not bother writing
1064 * anything if the buffer is marked invalid.
1066 * Note that since the buffer must be completely valid, we can safely
1067 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1068 * biodone() in order to prevent getblk from writing the buffer
1069 * out synchronously.
1072 bdwrite(struct buf *bp)
1074 if (BUF_LOCKINUSE(bp) == 0)
1075 panic("bdwrite: buffer is not busy");
1077 if (bp->b_flags & B_INVAL) {
1083 dsched_buf_enter(bp); /* might stack */
1086 * Set B_CACHE, indicating that the buffer is fully valid. This is
1087 * true even of NFS now.
1089 bp->b_flags |= B_CACHE;
1092 * This bmap keeps the system from needing to do the bmap later,
1093 * perhaps when the system is attempting to do a sync. Since it
1094 * is likely that the indirect block -- or whatever other datastructure
1095 * that the filesystem needs is still in memory now, it is a good
1096 * thing to do this. Note also, that if the pageout daemon is
1097 * requesting a sync -- there might not be enough memory to do
1098 * the bmap then... So, this is important to do.
1100 if (bp->b_bio2.bio_offset == NOOFFSET) {
1101 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1102 NULL, NULL, BUF_CMD_WRITE);
1106 * Because the underlying pages may still be mapped and
1107 * writable trying to set the dirty buffer (b_dirtyoff/end)
1108 * range here will be inaccurate.
1110 * However, we must still clean the pages to satisfy the
1111 * vnode_pager and pageout daemon, so they think the pages
1112 * have been "cleaned". What has really occured is that
1113 * they've been earmarked for later writing by the buffer
1116 * So we get the b_dirtyoff/end update but will not actually
1117 * depend on it (NFS that is) until the pages are busied for
1120 vfs_clean_pages(bp);
1124 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1125 * due to the softdep code.
1130 * Fake write - return pages to VM system as dirty, leave the buffer clean.
1131 * This is used by tmpfs.
1133 * It is important for any VFS using this routine to NOT use it for
1134 * IO_SYNC or IO_ASYNC operations which occur when the system really
1135 * wants to flush VM pages to backing store.
1138 buwrite(struct buf *bp)
1144 * Only works for VMIO buffers. If the buffer is already
1145 * marked for delayed-write we can't avoid the bdwrite().
1147 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) {
1153 * Mark as needing a commit.
1155 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1156 m = bp->b_xio.xio_pages[i];
1157 vm_page_need_commit(m);
1165 * Turn buffer into delayed write request by marking it B_DELWRI.
1166 * B_RELBUF and B_NOCACHE must be cleared.
1168 * We reassign the buffer to itself to properly update it in the
1169 * dirty/clean lists.
1171 * Must be called from a critical section.
1172 * The buffer must be on BQUEUE_NONE.
1175 bdirty(struct buf *bp)
1177 KASSERT(bp->b_qindex == BQUEUE_NONE,
1178 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1179 if (bp->b_flags & B_NOCACHE) {
1180 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1181 bp->b_flags &= ~B_NOCACHE;
1183 if (bp->b_flags & B_INVAL) {
1184 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1186 bp->b_flags &= ~B_RELBUF;
1188 if ((bp->b_flags & B_DELWRI) == 0) {
1189 lwkt_gettoken(&bp->b_vp->v_token);
1190 bp->b_flags |= B_DELWRI;
1192 lwkt_reltoken(&bp->b_vp->v_token);
1194 atomic_add_long(&dirtybufcount, 1);
1195 atomic_add_long(&dirtykvaspace, bp->b_kvasize);
1196 atomic_add_long(&dirtybufspace, bp->b_bufsize);
1197 if (bp->b_flags & B_HEAVY) {
1198 atomic_add_long(&dirtybufcounthw, 1);
1199 atomic_add_long(&dirtybufspacehw, bp->b_bufsize);
1206 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1207 * needs to be flushed with a different buf_daemon thread to avoid
1208 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1211 bheavy(struct buf *bp)
1213 if ((bp->b_flags & B_HEAVY) == 0) {
1214 bp->b_flags |= B_HEAVY;
1215 if (bp->b_flags & B_DELWRI) {
1216 atomic_add_long(&dirtybufcounthw, 1);
1217 atomic_add_long(&dirtybufspacehw, bp->b_bufsize);
1225 * Clear B_DELWRI for buffer.
1227 * Must be called from a critical section.
1229 * The buffer is typically on BQUEUE_NONE but there is one case in
1230 * brelse() that calls this function after placing the buffer on
1231 * a different queue.
1234 bundirty(struct buf *bp)
1236 if (bp->b_flags & B_DELWRI) {
1237 lwkt_gettoken(&bp->b_vp->v_token);
1238 bp->b_flags &= ~B_DELWRI;
1240 lwkt_reltoken(&bp->b_vp->v_token);
1242 atomic_add_long(&dirtybufcount, -1);
1243 atomic_add_long(&dirtykvaspace, -bp->b_kvasize);
1244 atomic_add_long(&dirtybufspace, -bp->b_bufsize);
1245 if (bp->b_flags & B_HEAVY) {
1246 atomic_add_long(&dirtybufcounthw, -1);
1247 atomic_add_long(&dirtybufspacehw, -bp->b_bufsize);
1249 bd_signal(bp->b_bufsize);
1252 * Since it is now being written, we can clear its deferred write flag.
1254 bp->b_flags &= ~B_DEFERRED;
1258 * Set the b_runningbufspace field, used to track how much I/O is
1259 * in progress at any given moment.
1262 bsetrunningbufspace(struct buf *bp, int bytes)
1264 bp->b_runningbufspace = bytes;
1266 atomic_add_long(&runningbufspace, bytes);
1267 atomic_add_long(&runningbufcount, 1);
1274 * Release a busy buffer and, if requested, free its resources. The
1275 * buffer will be stashed in the appropriate bufqueue[] allowing it
1276 * to be accessed later as a cache entity or reused for other purposes.
1279 brelse(struct buf *bp)
1281 struct bufpcpu *pcpu;
1283 int saved_flags = bp->b_flags;
1286 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1287 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1290 * If B_NOCACHE is set we are being asked to destroy the buffer and
1291 * its backing store. Clear B_DELWRI.
1293 * B_NOCACHE is set in two cases: (1) when the caller really wants
1294 * to destroy the buffer and backing store and (2) when the caller
1295 * wants to destroy the buffer and backing store after a write
1298 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1302 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1304 * A re-dirtied buffer is only subject to destruction
1305 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1307 /* leave buffer intact */
1308 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1309 (bp->b_bufsize <= 0)) {
1311 * Either a failed read or we were asked to free or not
1312 * cache the buffer. This path is reached with B_DELWRI
1313 * set only if B_INVAL is already set. B_NOCACHE governs
1314 * backing store destruction.
1316 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1317 * buffer cannot be immediately freed.
1319 bp->b_flags |= B_INVAL;
1320 if (LIST_FIRST(&bp->b_dep) != NULL)
1322 if (bp->b_flags & B_DELWRI) {
1323 atomic_add_long(&dirtybufcount, -1);
1324 atomic_add_long(&dirtykvaspace, -bp->b_kvasize);
1325 atomic_add_long(&dirtybufspace, -bp->b_bufsize);
1326 if (bp->b_flags & B_HEAVY) {
1327 atomic_add_long(&dirtybufcounthw, -1);
1328 atomic_add_long(&dirtybufspacehw,
1331 bd_signal(bp->b_bufsize);
1333 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1337 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set,
1338 * or if b_refs is non-zero.
1340 * If vfs_vmio_release() is called with either bit set, the
1341 * underlying pages may wind up getting freed causing a previous
1342 * write (bdwrite()) to get 'lost' because pages associated with
1343 * a B_DELWRI bp are marked clean. Pages associated with a
1344 * B_LOCKED buffer may be mapped by the filesystem.
1346 * If we want to release the buffer ourselves (rather then the
1347 * originator asking us to release it), give the originator a
1348 * chance to countermand the release by setting B_LOCKED.
1350 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1351 * if B_DELWRI is set.
1353 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1354 * on pages to return pages to the VM page queues.
1356 if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) {
1357 bp->b_flags &= ~B_RELBUF;
1358 } else if (vm_page_count_min(0)) {
1359 if (LIST_FIRST(&bp->b_dep) != NULL)
1360 buf_deallocate(bp); /* can set B_LOCKED */
1361 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1362 bp->b_flags &= ~B_RELBUF;
1364 bp->b_flags |= B_RELBUF;
1368 * Make sure b_cmd is clear. It may have already been cleared by
1371 * At this point destroying the buffer is governed by the B_INVAL
1372 * or B_RELBUF flags.
1374 bp->b_cmd = BUF_CMD_DONE;
1375 dsched_buf_exit(bp);
1378 * VMIO buffer rundown. Make sure the VM page array is restored
1379 * after an I/O may have replaces some of the pages with bogus pages
1380 * in order to not destroy dirty pages in a fill-in read.
1382 * Note that due to the code above, if a buffer is marked B_DELWRI
1383 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1384 * B_INVAL may still be set, however.
1386 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1387 * but not the backing store. B_NOCACHE will destroy the backing
1390 * Note that dirty NFS buffers contain byte-granular write ranges
1391 * and should not be destroyed w/ B_INVAL even if the backing store
1394 if (bp->b_flags & B_VMIO) {
1396 * Rundown for VMIO buffers which are not dirty NFS buffers.
1408 * Get the base offset and length of the buffer. Note that
1409 * in the VMIO case if the buffer block size is not
1410 * page-aligned then b_data pointer may not be page-aligned.
1411 * But our b_xio.xio_pages array *IS* page aligned.
1413 * block sizes less then DEV_BSIZE (usually 512) are not
1414 * supported due to the page granularity bits (m->valid,
1415 * m->dirty, etc...).
1417 * See man buf(9) for more information
1420 resid = bp->b_bufsize;
1421 foff = bp->b_loffset;
1423 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1424 m = bp->b_xio.xio_pages[i];
1427 * If we hit a bogus page, fixup *all* of them
1428 * now. Note that we left these pages wired
1429 * when we removed them so they had better exist,
1430 * and they cannot be ripped out from under us so
1431 * no critical section protection is necessary.
1433 if (m == bogus_page) {
1435 poff = OFF_TO_IDX(bp->b_loffset);
1437 vm_object_hold(obj);
1438 for (j = i; j < bp->b_xio.xio_npages; j++) {
1441 mtmp = bp->b_xio.xio_pages[j];
1442 if (mtmp == bogus_page) {
1443 if ((bp->b_flags & B_HASBOGUS) == 0)
1444 panic("brelse: bp %p corrupt bogus", bp);
1445 mtmp = vm_page_lookup(obj, poff + j);
1447 panic("brelse: bp %p page %d missing", bp, j);
1448 bp->b_xio.xio_pages[j] = mtmp;
1451 vm_object_drop(obj);
1453 if ((bp->b_flags & B_HASBOGUS) ||
1454 (bp->b_flags & B_INVAL) == 0) {
1455 pmap_qenter_noinval(
1456 trunc_page((vm_offset_t)bp->b_data),
1457 bp->b_xio.xio_pages,
1458 bp->b_xio.xio_npages);
1459 bp->b_flags &= ~B_HASBOGUS;
1460 bp->b_flags |= B_KVABIO;
1463 m = bp->b_xio.xio_pages[i];
1467 * Invalidate the backing store if B_NOCACHE is set
1468 * (e.g. used with vinvalbuf()). If this is NFS
1469 * we impose a requirement that the block size be
1470 * a multiple of PAGE_SIZE and create a temporary
1471 * hack to basically invalidate the whole page. The
1472 * problem is that NFS uses really odd buffer sizes
1473 * especially when tracking piecemeal writes and
1474 * it also vinvalbuf()'s a lot, which would result
1475 * in only partial page validation and invalidation
1476 * here. If the file page is mmap()'d, however,
1477 * all the valid bits get set so after we invalidate
1478 * here we would end up with weird m->valid values
1479 * like 0xfc. nfs_getpages() can't handle this so
1480 * we clear all the valid bits for the NFS case
1481 * instead of just some of them.
1483 * The real bug is the VM system having to set m->valid
1484 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1485 * itself is an artifact of the whole 512-byte
1486 * granular mess that exists to support odd block
1487 * sizes and UFS meta-data block sizes (e.g. 6144).
1488 * A complete rewrite is required.
1492 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1493 int poffset = foff & PAGE_MASK;
1496 presid = PAGE_SIZE - poffset;
1497 if (bp->b_vp->v_tag == VT_NFS &&
1498 bp->b_vp->v_type == VREG) {
1500 } else if (presid > resid) {
1503 KASSERT(presid >= 0, ("brelse: extra page"));
1504 vm_page_set_invalid(m, poffset, presid);
1507 * Also make sure any swap cache is removed
1508 * as it is now stale (HAMMER in particular
1509 * uses B_NOCACHE to deal with buffer
1512 swap_pager_unswapped(m);
1514 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1515 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1517 if (bp->b_flags & (B_INVAL | B_RELBUF))
1518 vfs_vmio_release(bp);
1521 * Rundown for non-VMIO buffers.
1523 * XXX With B_MALLOC buffers removed, there should no longer
1524 * be any situation where brelse() is called on a non B_VMIO
1525 * buffer. Recommend assertion here. XXX
1527 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1530 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1536 if (bp->b_qindex != BQUEUE_NONE)
1537 panic("brelse: free buffer onto another queue???");
1540 * Figure out the correct queue to place the cleaned up buffer on.
1541 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1542 * disassociated from their vnode.
1544 * Return the buffer to its original pcpu area
1546 pcpu = &bufpcpu[bp->b_qcpu];
1547 spin_lock(&pcpu->spin);
1549 if (bp->b_flags & B_LOCKED) {
1551 * Buffers that are locked are placed in the locked queue
1552 * immediately, regardless of their state.
1554 bp->b_qindex = BQUEUE_LOCKED;
1555 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1557 } else if (bp->b_bufsize == 0) {
1559 * Buffers with no memory. Due to conditionals near the top
1560 * of brelse() such buffers should probably already be
1561 * marked B_INVAL and disassociated from their vnode.
1563 bp->b_flags |= B_INVAL;
1564 KASSERT(bp->b_vp == NULL,
1565 ("bp1 %p flags %08x/%08x vnode %p "
1566 "unexpectededly still associated!",
1567 bp, saved_flags, bp->b_flags, bp->b_vp));
1568 KKASSERT((bp->b_flags & B_HASHED) == 0);
1569 bp->b_qindex = BQUEUE_EMPTY;
1570 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex],
1572 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1574 * Buffers with junk contents. Again these buffers had better
1575 * already be disassociated from their vnode.
1577 KASSERT(bp->b_vp == NULL,
1578 ("bp2 %p flags %08x/%08x vnode %p unexpectededly "
1579 "still associated!",
1580 bp, saved_flags, bp->b_flags, bp->b_vp));
1581 KKASSERT((bp->b_flags & B_HASHED) == 0);
1582 bp->b_flags |= B_INVAL;
1583 bp->b_qindex = BQUEUE_CLEAN;
1584 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex],
1588 * Remaining buffers. These buffers are still associated with
1591 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1593 bp->b_qindex = BQUEUE_DIRTY;
1594 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1597 case B_DELWRI | B_HEAVY:
1598 bp->b_qindex = BQUEUE_DIRTY_HW;
1599 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1604 * NOTE: Buffers are always placed at the end of the
1605 * queue. If B_AGE is not set the buffer will cycle
1606 * through the queue twice.
1608 bp->b_qindex = BQUEUE_CLEAN;
1609 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1614 spin_unlock(&pcpu->spin);
1617 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1618 * on the correct queue but we have not yet unlocked it.
1620 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1624 * The bp is on an appropriate queue unless locked. If it is not
1625 * locked or dirty we can wakeup threads waiting for buffer space.
1627 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1628 * if B_INVAL is set ).
1630 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1634 * Something we can maybe free or reuse
1636 if (bp->b_bufsize || bp->b_kvasize)
1640 * Clean up temporary flags and unlock the buffer.
1642 bp->b_flags &= ~(B_NOCACHE | B_RELBUF | B_DIRECT);
1649 * Release a buffer back to the appropriate queue but do not try to free
1650 * it. The buffer is expected to be used again soon.
1652 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1653 * biodone() to requeue an async I/O on completion. It is also used when
1654 * known good buffers need to be requeued but we think we may need the data
1657 * XXX we should be able to leave the B_RELBUF hint set on completion.
1660 bqrelse(struct buf *bp)
1662 struct bufpcpu *pcpu;
1664 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1665 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1667 if (bp->b_qindex != BQUEUE_NONE)
1668 panic("bqrelse: free buffer onto another queue???");
1670 buf_act_advance(bp);
1672 pcpu = &bufpcpu[bp->b_qcpu];
1673 spin_lock(&pcpu->spin);
1675 if (bp->b_flags & B_LOCKED) {
1677 * Locked buffers are released to the locked queue. However,
1678 * if the buffer is dirty it will first go into the dirty
1679 * queue and later on after the I/O completes successfully it
1680 * will be released to the locked queue.
1682 bp->b_qindex = BQUEUE_LOCKED;
1683 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1685 } else if (bp->b_flags & B_DELWRI) {
1686 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1687 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1688 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1690 } else if (vm_page_count_min(0)) {
1692 * We are too low on memory, we have to try to free the
1693 * buffer (most importantly: the wired pages making up its
1694 * backing store) *now*.
1696 spin_unlock(&pcpu->spin);
1700 bp->b_qindex = BQUEUE_CLEAN;
1701 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1704 spin_unlock(&pcpu->spin);
1707 * We have now placed the buffer on the proper queue, but have yet
1710 if ((bp->b_flags & B_LOCKED) == 0 &&
1711 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1716 * Something we can maybe free or reuse.
1718 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1722 * Final cleanup and unlock. Clear bits that are only used while a
1723 * buffer is actively locked.
1725 bp->b_flags &= ~(B_NOCACHE | B_RELBUF);
1726 dsched_buf_exit(bp);
1731 * Hold a buffer, preventing it from being reused. This will prevent
1732 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL
1733 * operations. If a B_INVAL operation occurs the buffer will remain held
1734 * but the underlying pages may get ripped out.
1736 * These functions are typically used in VOP_READ/VOP_WRITE functions
1737 * to hold a buffer during a copyin or copyout, preventing deadlocks
1738 * or recursive lock panics when read()/write() is used over mmap()'d
1741 * NOTE: bqhold() requires that the buffer be locked at the time of the
1742 * hold. bqdrop() has no requirements other than the buffer having
1743 * previously been held.
1746 bqhold(struct buf *bp)
1748 atomic_add_int(&bp->b_refs, 1);
1752 bqdrop(struct buf *bp)
1754 KKASSERT(bp->b_refs > 0);
1755 atomic_add_int(&bp->b_refs, -1);
1759 * Return backing pages held by the buffer 'bp' back to the VM system.
1760 * This routine is called when the bp is invalidated, released, or
1763 * The KVA mapping (b_data) for the underlying pages is removed by
1766 * WARNING! This routine is integral to the low memory critical path
1767 * when a buffer is B_RELBUF'd. If the system has a severe page
1768 * deficit we need to get the page(s) onto the PQ_FREE or PQ_CACHE
1769 * queues so they can be reused in the current pageout daemon
1773 vfs_vmio_release(struct buf *bp)
1778 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1779 m = bp->b_xio.xio_pages[i];
1780 bp->b_xio.xio_pages[i] = NULL;
1783 * We need to own the page in order to safely unwire it.
1785 vm_page_busy_wait(m, FALSE, "vmiopg");
1788 * The VFS is telling us this is not a meta-data buffer
1789 * even if it is backed by a block device.
1791 if (bp->b_flags & B_NOTMETA)
1792 vm_page_flag_set(m, PG_NOTMETA);
1795 * This is a very important bit of code. We try to track
1796 * VM page use whether the pages are wired into the buffer
1797 * cache or not. While wired into the buffer cache the
1798 * bp tracks the act_count.
1800 * We can choose to place unwired pages on the inactive
1801 * queue (0) or active queue (1). If we place too many
1802 * on the active queue the queue will cycle the act_count
1803 * on pages we'd like to keep, just from single-use pages
1804 * (such as when doing a tar-up or file scan).
1806 if (bp->b_act_count < vm_cycle_point)
1807 vm_page_unwire(m, 0);
1809 vm_page_unwire(m, 1);
1812 * If the wire_count has dropped to 0 we may need to take
1813 * further action before unbusying the page.
1815 * WARNING: vm_page_try_*() also checks PG_NEED_COMMIT for us.
1817 if (m->wire_count == 0) {
1818 if (bp->b_flags & B_DIRECT) {
1820 * Attempt to free the page if B_DIRECT is
1821 * set, the caller does not desire the page
1825 vm_page_try_to_free(m);
1826 } else if ((bp->b_flags & B_NOTMETA) ||
1827 vm_page_count_min(0)) {
1829 * Attempt to move the page to PQ_CACHE
1830 * if B_NOTMETA is set. This flag is set
1831 * by HAMMER to remove one of the two pages
1832 * present when double buffering is enabled.
1834 * Attempt to move the page to PQ_CACHE
1835 * If we have a severe page deficit. This
1836 * will cause buffer cache operations related
1837 * to pageouts to recycle the related pages
1838 * in order to avoid a low memory deadlock.
1840 m->act_count = bp->b_act_count;
1841 vm_page_try_to_cache(m);
1844 * Nominal case, leave the page on the
1845 * queue the original unwiring placed it on
1846 * (active or inactive).
1848 m->act_count = bp->b_act_count;
1857 * Zero out the pmap pte's for the mapping, but don't bother
1858 * invalidating the TLB. The range will be properly invalidating
1859 * when new pages are entered into the mapping.
1861 * This in particular reduces tmpfs tear-down overhead and reduces
1862 * buffer cache re-use overhead (one invalidation sequence instead
1863 * of two per re-use).
1865 pmap_qremove_noinval(trunc_page((vm_offset_t) bp->b_data),
1866 bp->b_xio.xio_npages);
1867 CPUMASK_ASSZERO(bp->b_cpumask);
1868 if (bp->b_bufsize) {
1869 atomic_add_long(&bufspace, -bp->b_bufsize);
1873 bp->b_xio.xio_npages = 0;
1874 bp->b_flags &= ~B_VMIO;
1875 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1881 * Find and initialize a new buffer header, freeing up existing buffers
1882 * in the bufqueues as necessary. The new buffer is returned locked.
1884 * Important: B_INVAL is not set. If the caller wishes to throw the
1885 * buffer away, the caller must set B_INVAL prior to calling brelse().
1888 * We have insufficient buffer headers
1889 * We have insufficient buffer space
1891 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1892 * Instead we ask the buf daemon to do it for us. We attempt to
1893 * avoid piecemeal wakeups of the pageout daemon.
1896 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1898 struct bufpcpu *pcpu;
1903 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1904 int maxloops = 200000;
1905 int restart_reason = 0;
1906 struct buf *restart_bp = NULL;
1907 static char flushingbufs[MAXCPU];
1911 * We can't afford to block since we might be holding a vnode lock,
1912 * which may prevent system daemons from running. We deal with
1913 * low-memory situations by proactively returning memory and running
1914 * async I/O rather then sync I/O.
1918 nqcpu = mycpu->gd_cpuid;
1919 flushingp = &flushingbufs[nqcpu];
1921 if (bufspace < lobufspace)
1924 if (debug_bufbio && --maxloops == 0)
1925 panic("getnewbuf, excessive loops on cpu %d restart %d (%p)",
1926 mycpu->gd_cpuid, restart_reason, restart_bp);
1929 * Setup for scan. If we do not have enough free buffers,
1930 * we setup a degenerate case that immediately fails. Note
1931 * that if we are specially marked process, we are allowed to
1932 * dip into our reserves.
1934 * The scanning sequence is nominally: EMPTY->CLEAN
1936 pcpu = &bufpcpu[nqcpu];
1937 spin_lock(&pcpu->spin);
1940 * Prime the scan for this cpu. Locate the first buffer to
1941 * check. If we are flushing buffers we must skip the
1944 nqindex = BQUEUE_EMPTY;
1945 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTY]);
1946 if (nbp == NULL || *flushingp) {
1947 nqindex = BQUEUE_CLEAN;
1948 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]);
1952 * Run scan, possibly freeing data and/or kva mappings on the fly,
1955 * WARNING! spin is held!
1957 while ((bp = nbp) != NULL) {
1958 int qindex = nqindex;
1960 nbp = TAILQ_NEXT(bp, b_freelist);
1963 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1964 * cycles through the queue twice before being selected.
1966 if (qindex == BQUEUE_CLEAN &&
1967 (bp->b_flags & B_AGE) == 0 && nbp) {
1968 bp->b_flags |= B_AGE;
1969 TAILQ_REMOVE(&pcpu->bufqueues[qindex],
1971 TAILQ_INSERT_TAIL(&pcpu->bufqueues[qindex],
1977 * Calculate next bp ( we can only use it if we do not block
1978 * or do other fancy things ).
1983 nqindex = BQUEUE_CLEAN;
1984 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN])))
1998 KASSERT(bp->b_qindex == qindex,
1999 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2002 * Note: we no longer distinguish between VMIO and non-VMIO
2005 KASSERT((bp->b_flags & B_DELWRI) == 0,
2006 ("delwri buffer %p found in queue %d", bp, qindex));
2009 * Do not try to reuse a buffer with a non-zero b_refs.
2010 * This is an unsynchronized test. A synchronized test
2011 * is also performed after we lock the buffer.
2017 * Start freeing the bp. This is somewhat involved. nbp
2018 * remains valid only for BQUEUE_EMPTY bp's. Buffers
2019 * on the clean list must be disassociated from their
2020 * current vnode. Buffers on the empty lists have
2021 * already been disassociated.
2023 * b_refs is checked after locking along with queue changes.
2024 * We must check here to deal with zero->nonzero transitions
2025 * made by the owner of the buffer lock, which is used by
2026 * VFS's to hold the buffer while issuing an unlocked
2027 * uiomove()s. We cannot invalidate the buffer's pages
2028 * for this case. Once we successfully lock a buffer the
2029 * only 0->1 transitions of b_refs will occur via findblk().
2031 * We must also check for queue changes after successful
2032 * locking as the current lock holder may dispose of the
2033 * buffer and change its queue.
2035 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2036 spin_unlock(&pcpu->spin);
2037 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100);
2042 if (bp->b_qindex != qindex || bp->b_refs) {
2043 spin_unlock(&pcpu->spin);
2049 bremfree_locked(bp);
2050 spin_unlock(&pcpu->spin);
2053 * Dependancies must be handled before we disassociate the
2056 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2057 * be immediately disassociated. HAMMER then becomes
2058 * responsible for releasing the buffer.
2060 * NOTE: spin is UNLOCKED now.
2062 if (LIST_FIRST(&bp->b_dep) != NULL) {
2064 if (bp->b_flags & B_LOCKED) {
2070 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2074 * CLEAN buffers have content or associations that must be
2075 * cleaned out if not repurposing.
2077 if (qindex == BQUEUE_CLEAN) {
2078 if (bp->b_flags & B_VMIO)
2079 vfs_vmio_release(bp);
2085 * NOTE: nbp is now entirely invalid. We can only restart
2086 * the scan from this point on.
2088 * Get the rest of the buffer freed up. b_kva* is still
2089 * valid after this operation.
2091 KASSERT(bp->b_vp == NULL,
2092 ("bp3 %p flags %08x vnode %p qindex %d "
2093 "unexpectededly still associated!",
2094 bp, bp->b_flags, bp->b_vp, qindex));
2095 KKASSERT((bp->b_flags & B_HASHED) == 0);
2100 if (bp->b_flags & (B_VNDIRTY | B_VNCLEAN | B_HASHED)) {
2101 kprintf("getnewbuf: caught bug vp queue "
2102 "%p/%08x qidx %d\n",
2103 bp, bp->b_flags, qindex);
2106 bp->b_flags = B_BNOCLIP;
2107 bp->b_cmd = BUF_CMD_DONE;
2112 bp->b_xio.xio_npages = 0;
2113 bp->b_dirtyoff = bp->b_dirtyend = 0;
2114 bp->b_act_count = ACT_INIT;
2116 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2118 if (blkflags & GETBLK_BHEAVY)
2119 bp->b_flags |= B_HEAVY;
2121 if (bufspace >= hibufspace)
2123 if (bufspace < lobufspace)
2126 bp->b_flags |= B_INVAL;
2134 * b_refs can transition to a non-zero value while we hold
2135 * the buffer locked due to a findblk(). Our brelvp() above
2136 * interlocked any future possible transitions due to
2139 * If we find b_refs to be non-zero we can destroy the
2140 * buffer's contents but we cannot yet reuse the buffer.
2143 bp->b_flags |= B_INVAL;
2152 * We found our buffer!
2158 * If we exhausted our list, iterate other cpus. If that fails,
2159 * sleep as appropriate. We may have to wakeup various daemons
2160 * and write out some dirty buffers.
2162 * Generally we are sleeping due to insufficient buffer space.
2164 * NOTE: spin is held if bp is NULL, else it is not held.
2170 spin_unlock(&pcpu->spin);
2172 nqcpu = (nqcpu + 1) % ncpus;
2173 if (nqcpu != mycpu->gd_cpuid) {
2179 if (bufspace >= hibufspace) {
2181 flags = VFS_BIO_NEED_BUFSPACE;
2184 flags = VFS_BIO_NEED_ANY;
2187 bd_speedup(); /* heeeelp */
2188 atomic_set_int(&needsbuffer, flags);
2189 while (needsbuffer & flags) {
2192 tsleep_interlock(&needsbuffer, 0);
2193 value = atomic_fetchadd_int(&needsbuffer, 0);
2194 if (value & flags) {
2195 if (tsleep(&needsbuffer, PINTERLOCKED|slpflags,
2196 waitmsg, slptimeo)) {
2203 * We finally have a valid bp. Reset b_data.
2205 * (spin is not held)
2207 bp->b_data = bp->b_kvabase;
2215 * Buffer flushing daemon. Buffers are normally flushed by the
2216 * update daemon but if it cannot keep up this process starts to
2217 * take the load in an attempt to prevent getnewbuf() from blocking.
2219 * Once a flush is initiated it does not stop until the number
2220 * of buffers falls below lodirtybuffers, but we will wake up anyone
2221 * waiting at the mid-point.
2223 static struct kproc_desc buf_kp = {
2228 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2229 kproc_start, &buf_kp);
2231 static struct kproc_desc bufhw_kp = {
2236 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2237 kproc_start, &bufhw_kp);
2240 buf_daemon1(struct thread *td, int queue, int (*buf_limit_fn)(long),
2246 marker = kmalloc(sizeof(*marker), M_BIOBUF, M_WAITOK | M_ZERO);
2247 marker->b_flags |= B_MARKER;
2248 marker->b_qindex = BQUEUE_NONE;
2252 * This process needs to be suspended prior to shutdown sync.
2254 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2255 td, SHUTDOWN_PRI_LAST);
2256 curthread->td_flags |= TDF_SYSTHREAD;
2259 * This process is allowed to take the buffer cache to the limit
2262 kproc_suspend_loop();
2265 * Do the flush as long as the number of dirty buffers
2266 * (including those running) exceeds lodirtybufspace.
2268 * When flushing limit running I/O to hirunningspace
2269 * Do the flush. Limit the amount of in-transit I/O we
2270 * allow to build up, otherwise we would completely saturate
2271 * the I/O system. Wakeup any waiting processes before we
2272 * normally would so they can run in parallel with our drain.
2274 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2275 * but because we split the operation into two threads we
2276 * have to cut it in half for each thread.
2278 waitrunningbufspace();
2279 limit = lodirtybufspace / 2;
2280 while (buf_limit_fn(limit)) {
2281 if (flushbufqueues(marker, queue) == 0)
2283 if (runningbufspace < hirunningspace)
2285 waitrunningbufspace();
2289 * We reached our low water mark, reset the
2290 * request and sleep until we are needed again.
2291 * The sleep is just so the suspend code works.
2293 tsleep_interlock(bd_req, 0);
2294 if (atomic_swap_int(bd_req, 0) == 0)
2295 tsleep(bd_req, PINTERLOCKED, "psleep", hz);
2298 /*kfree(marker, M_BIOBUF);*/
2302 buf_daemon_limit(long limit)
2304 return (runningbufspace + dirtykvaspace > limit ||
2305 dirtybufcount - dirtybufcounthw >= nbuf / 2);
2309 buf_daemon_hw_limit(long limit)
2311 return (runningbufspace + dirtykvaspace > limit ||
2312 dirtybufcounthw >= nbuf / 2);
2318 buf_daemon1(bufdaemon_td, BQUEUE_DIRTY, buf_daemon_limit,
2325 buf_daemon1(bufdaemonhw_td, BQUEUE_DIRTY_HW, buf_daemon_hw_limit,
2330 * Flush up to (flushperqueue) buffers in the dirty queue. Each cpu has a
2331 * localized version of the queue. Each call made to this function iterates
2332 * to another cpu. It is desireable to flush several buffers from the same
2333 * cpu's queue at once, as these are likely going to be linear.
2335 * We must be careful to free up B_INVAL buffers instead of write them, which
2336 * NFS is particularly sensitive to.
2338 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate that we
2339 * really want to try to get the buffer out and reuse it due to the write
2340 * load on the machine.
2342 * We must lock the buffer in order to check its validity before we can mess
2343 * with its contents. spin isn't enough.
2346 flushbufqueues(struct buf *marker, bufq_type_t q)
2348 struct bufpcpu *pcpu;
2351 u_int loops = flushperqueue;
2352 int lcpu = marker->b_qcpu;
2354 KKASSERT(marker->b_qindex == BQUEUE_NONE);
2355 KKASSERT(marker->b_flags & B_MARKER);
2359 * Spinlock needed to perform operations on the queue and may be
2360 * held through a non-blocking BUF_LOCK(), but cannot be held when
2361 * BUF_UNLOCK()ing or through any other major operation.
2363 pcpu = &bufpcpu[marker->b_qcpu];
2364 spin_lock(&pcpu->spin);
2365 marker->b_qindex = q;
2366 TAILQ_INSERT_HEAD(&pcpu->bufqueues[q], marker, b_freelist);
2369 while ((bp = TAILQ_NEXT(bp, b_freelist)) != NULL) {
2371 * NOTE: spinlock is always held at the top of the loop
2373 if (bp->b_flags & B_MARKER)
2375 if ((bp->b_flags & B_DELWRI) == 0) {
2376 kprintf("Unexpected clean buffer %p\n", bp);
2379 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
2381 KKASSERT(bp->b_qcpu == marker->b_qcpu && bp->b_qindex == q);
2384 * Once the buffer is locked we will have no choice but to
2385 * unlock the spinlock around a later BUF_UNLOCK and re-set
2386 * bp = marker when looping. Move the marker now to make
2389 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist);
2390 TAILQ_INSERT_AFTER(&pcpu->bufqueues[q], bp, marker, b_freelist);
2393 * Must recheck B_DELWRI after successfully locking
2396 if ((bp->b_flags & B_DELWRI) == 0) {
2397 spin_unlock(&pcpu->spin);
2399 spin_lock(&pcpu->spin);
2405 * Remove the buffer from its queue. We still own the
2411 * Disposing of an invalid buffer counts as a flush op
2413 if (bp->b_flags & B_INVAL) {
2414 spin_unlock(&pcpu->spin);
2420 * Release the spinlock for the more complex ops we
2421 * are now going to do.
2423 spin_unlock(&pcpu->spin);
2427 * This is a bit messy
2429 if (LIST_FIRST(&bp->b_dep) != NULL &&
2430 (bp->b_flags & B_DEFERRED) == 0 &&
2431 buf_countdeps(bp, 0)) {
2432 spin_lock(&pcpu->spin);
2433 TAILQ_INSERT_TAIL(&pcpu->bufqueues[q], bp, b_freelist);
2435 bp->b_flags |= B_DEFERRED;
2436 spin_unlock(&pcpu->spin);
2438 spin_lock(&pcpu->spin);
2444 * spinlock not held here.
2446 * If the buffer has a dependancy, buf_checkwrite() must
2447 * also return 0 for us to be able to initate the write.
2449 * If the buffer is flagged B_ERROR it may be requeued
2450 * over and over again, we try to avoid a live lock.
2452 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) {
2454 } else if (bp->b_flags & B_ERROR) {
2455 tsleep(bp, 0, "bioer", 1);
2456 bp->b_flags &= ~B_AGE;
2459 bp->b_flags |= B_AGE | B_KVABIO;
2462 /* bp invalid but needs to be NULL-tested if we break out */
2464 spin_lock(&pcpu->spin);
2470 /* bp is invalid here but can be NULL-tested to advance */
2472 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist);
2473 marker->b_qindex = BQUEUE_NONE;
2474 spin_unlock(&pcpu->spin);
2477 * Advance the marker to be fair.
2479 marker->b_qcpu = (marker->b_qcpu + 1) % ncpus;
2481 if (marker->b_qcpu != lcpu)
2491 * Returns true if no I/O is needed to access the associated VM object.
2492 * This is like findblk except it also hunts around in the VM system for
2495 * Note that we ignore vm_page_free() races from interrupts against our
2496 * lookup, since if the caller is not protected our return value will not
2497 * be any more valid then otherwise once we exit the critical section.
2500 inmem(struct vnode *vp, off_t loffset)
2503 vm_offset_t toff, tinc, size;
2507 if (findblk(vp, loffset, FINDBLK_TEST))
2509 if (vp->v_mount == NULL)
2511 if ((obj = vp->v_object) == NULL)
2515 if (size > vp->v_mount->mnt_stat.f_iosize)
2516 size = vp->v_mount->mnt_stat.f_iosize;
2518 vm_object_hold(obj);
2519 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2520 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2526 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2527 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2528 if (vm_page_is_valid(m,
2529 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) {
2534 vm_object_drop(obj);
2541 * Locate and return the specified buffer. Unless flagged otherwise,
2542 * a locked buffer will be returned if it exists or NULL if it does not.
2544 * findblk()'d buffers are still on the bufqueues and if you intend
2545 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2546 * and possibly do other stuff to it.
2548 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2549 * for locking the buffer and ensuring that it remains
2550 * the desired buffer after locking.
2552 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2553 * to acquire the lock we return NULL, even if the
2556 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal
2557 * reuse by getnewbuf() but does not prevent
2558 * disassociation (B_INVAL). Used to avoid deadlocks
2559 * against random (vp,loffset)s due to reassignment.
2561 * FINDBLK_KVABIO - Only applicable when returning a locked buffer.
2562 * Indicates that the caller supports B_KVABIO.
2564 * (0) - Lock the buffer blocking.
2567 findblk(struct vnode *vp, off_t loffset, int flags)
2572 lkflags = LK_EXCLUSIVE;
2573 if (flags & FINDBLK_NBLOCK)
2574 lkflags |= LK_NOWAIT;
2578 * Lookup. Ref the buf while holding v_token to prevent
2579 * reuse (but does not prevent diassociation).
2581 lwkt_gettoken_shared(&vp->v_token);
2582 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2584 lwkt_reltoken(&vp->v_token);
2588 lwkt_reltoken(&vp->v_token);
2591 * If testing only break and return bp, do not lock.
2593 if (flags & FINDBLK_TEST)
2597 * Lock the buffer, return an error if the lock fails.
2598 * (only FINDBLK_NBLOCK can cause the lock to fail).
2600 if (BUF_LOCK(bp, lkflags)) {
2601 atomic_subtract_int(&bp->b_refs, 1);
2602 /* bp = NULL; not needed */
2607 * Revalidate the locked buf before allowing it to be
2610 * B_KVABIO is only set/cleared when locking. When
2611 * clearing B_KVABIO, we must ensure that the buffer
2612 * is synchronized to all cpus.
2614 if (bp->b_vp == vp && bp->b_loffset == loffset) {
2615 if (flags & FINDBLK_KVABIO)
2616 bp->b_flags |= B_KVABIO;
2621 atomic_subtract_int(&bp->b_refs, 1);
2628 if ((flags & FINDBLK_REF) == 0)
2629 atomic_subtract_int(&bp->b_refs, 1);
2636 * Similar to getblk() except only returns the buffer if it is
2637 * B_CACHE and requires no other manipulation. Otherwise NULL
2638 * is returned. NULL is also returned if GETBLK_NOWAIT is set
2639 * and the getblk() would block.
2641 * If B_RAM is set the buffer might be just fine, but we return
2642 * NULL anyway because we want the code to fall through to the
2643 * cluster read to issue more read-aheads. Otherwise read-ahead breaks.
2645 * If blksize is 0 the buffer cache buffer must already be fully
2648 * If blksize is non-zero getblk() will be used, allowing a buffer
2649 * to be reinstantiated from its VM backing store. The buffer must
2650 * still be fully cached after reinstantiation to be returned.
2653 getcacheblk(struct vnode *vp, off_t loffset, int blksize, int blkflags)
2658 if (blkflags & GETBLK_NOWAIT)
2659 fndflags |= FINDBLK_NBLOCK;
2660 if (blkflags & GETBLK_KVABIO)
2661 fndflags |= FINDBLK_KVABIO;
2664 bp = getblk(vp, loffset, blksize, blkflags, 0);
2666 if ((bp->b_flags & (B_INVAL | B_CACHE)) == B_CACHE) {
2667 bp->b_flags &= ~B_AGE;
2668 if (bp->b_flags & B_RAM) {
2678 bp = findblk(vp, loffset, fndflags);
2680 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2682 bp->b_flags &= ~B_AGE;
2696 * Get a block given a specified block and offset into a file/device.
2697 * B_INVAL may or may not be set on return. The caller should clear
2698 * B_INVAL prior to initiating a READ.
2700 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2701 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2702 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2703 * without doing any of those things the system will likely believe
2704 * the buffer to be valid (especially if it is not B_VMIO), and the
2705 * next getblk() will return the buffer with B_CACHE set.
2707 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2708 * an existing buffer.
2710 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2711 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2712 * and then cleared based on the backing VM. If the previous buffer is
2713 * non-0-sized but invalid, B_CACHE will be cleared.
2715 * If getblk() must create a new buffer, the new buffer is returned with
2716 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2717 * case it is returned with B_INVAL clear and B_CACHE set based on the
2720 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2721 * B_CACHE bit is clear.
2723 * What this means, basically, is that the caller should use B_CACHE to
2724 * determine whether the buffer is fully valid or not and should clear
2725 * B_INVAL prior to issuing a read. If the caller intends to validate
2726 * the buffer by loading its data area with something, the caller needs
2727 * to clear B_INVAL. If the caller does this without issuing an I/O,
2728 * the caller should set B_CACHE ( as an optimization ), else the caller
2729 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2730 * a write attempt or if it was a successfull read. If the caller
2731 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2732 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2736 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2737 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2740 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2743 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2747 if (size > MAXBSIZE)
2748 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2749 if (vp->v_object == NULL)
2750 panic("getblk: vnode %p has no object!", vp);
2753 * NOTE: findblk does not try to resolve KVABIO in REF-only mode.
2754 * we still have to handle that ourselves.
2757 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) {
2759 * The buffer was found in the cache, but we need to lock it.
2760 * We must acquire a ref on the bp to prevent reuse, but
2761 * this will not prevent disassociation (brelvp()) so we
2762 * must recheck (vp,loffset) after acquiring the lock.
2764 * Without the ref the buffer could potentially be reused
2765 * before we acquire the lock and create a deadlock
2766 * situation between the thread trying to reuse the buffer
2767 * and us due to the fact that we would wind up blocking
2768 * on a random (vp,loffset).
2770 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2771 if (blkflags & GETBLK_NOWAIT) {
2775 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2776 if (blkflags & GETBLK_PCATCH)
2777 lkflags |= LK_PCATCH;
2778 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2781 if (error == ENOLCK)
2785 /* buffer may have changed on us */
2790 * Once the buffer has been locked, make sure we didn't race
2791 * a buffer recyclement. Buffers that are no longer hashed
2792 * will have b_vp == NULL, so this takes care of that check
2795 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2797 kprintf("Warning buffer %p (vp %p loffset %lld) "
2799 bp, vp, (long long)loffset);
2806 * If SZMATCH any pre-existing buffer must be of the requested
2807 * size or NULL is returned. The caller absolutely does not
2808 * want getblk() to bwrite() the buffer on a size mismatch.
2810 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2816 * All vnode-based buffers must be backed by a VM object.
2818 * Set B_KVABIO for any incidental work, we will fix it
2821 KKASSERT(bp->b_flags & B_VMIO);
2822 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2823 bp->b_flags &= ~B_AGE;
2824 bp->b_flags |= B_KVABIO;
2827 * Make sure that B_INVAL buffers do not have a cached
2828 * block number translation.
2830 if ((bp->b_flags & B_INVAL) &&
2831 (bp->b_bio2.bio_offset != NOOFFSET)) {
2832 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2833 " did not have cleared bio_offset cache\n",
2834 bp, vp, (long long)loffset);
2835 clearbiocache(&bp->b_bio2);
2839 * The buffer is locked. B_CACHE is cleared if the buffer is
2842 * After the bremfree(), disposals must use b[q]relse().
2844 if (bp->b_flags & B_INVAL)
2845 bp->b_flags &= ~B_CACHE;
2849 * Any size inconsistancy with a dirty buffer or a buffer
2850 * with a softupdates dependancy must be resolved. Resizing
2851 * the buffer in such circumstances can lead to problems.
2853 * Dirty or dependant buffers are written synchronously.
2854 * Other types of buffers are simply released and
2855 * reconstituted as they may be backed by valid, dirty VM
2856 * pages (but not marked B_DELWRI).
2858 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2859 * and may be left over from a prior truncation (and thus
2860 * no longer represent the actual EOF point), so we
2861 * definitely do not want to B_NOCACHE the backing store.
2863 if (size != bp->b_bcount) {
2864 if (bp->b_flags & B_DELWRI) {
2865 bp->b_flags |= B_RELBUF;
2867 } else if (LIST_FIRST(&bp->b_dep)) {
2868 bp->b_flags |= B_RELBUF;
2871 bp->b_flags |= B_RELBUF;
2876 KKASSERT(size <= bp->b_kvasize);
2877 KASSERT(bp->b_loffset != NOOFFSET,
2878 ("getblk: no buffer offset"));
2881 * A buffer with B_DELWRI set and B_CACHE clear must
2882 * be committed before we can return the buffer in
2883 * order to prevent the caller from issuing a read
2884 * ( due to B_CACHE not being set ) and overwriting
2887 * Most callers, including NFS and FFS, need this to
2888 * operate properly either because they assume they
2889 * can issue a read if B_CACHE is not set, or because
2890 * ( for example ) an uncached B_DELWRI might loop due
2891 * to softupdates re-dirtying the buffer. In the latter
2892 * case, B_CACHE is set after the first write completes,
2893 * preventing further loops.
2895 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2896 * above while extending the buffer, we cannot allow the
2897 * buffer to remain with B_CACHE set after the write
2898 * completes or it will represent a corrupt state. To
2899 * deal with this we set B_NOCACHE to scrap the buffer
2902 * XXX Should this be B_RELBUF instead of B_NOCACHE?
2903 * I'm not even sure this state is still possible
2904 * now that getblk() writes out any dirty buffers
2907 * We might be able to do something fancy, like setting
2908 * B_CACHE in bwrite() except if B_DELWRI is already set,
2909 * so the below call doesn't set B_CACHE, but that gets real
2910 * confusing. This is much easier.
2912 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2913 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
2914 "and CACHE clear, b_flags %08x\n",
2915 bp, (uintmax_t)bp->b_loffset, bp->b_flags);
2916 bp->b_flags |= B_NOCACHE;
2922 * Buffer is not in-core, create new buffer. The buffer
2923 * returned by getnewbuf() is locked. Note that the returned
2924 * buffer is also considered valid (not marked B_INVAL).
2926 * Calculating the offset for the I/O requires figuring out
2927 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2928 * the mount's f_iosize otherwise. If the vnode does not
2929 * have an associated mount we assume that the passed size is
2932 * Note that vn_isdisk() cannot be used here since it may
2933 * return a failure for numerous reasons. Note that the
2934 * buffer size may be larger then the block size (the caller
2935 * will use block numbers with the proper multiple). Beware
2936 * of using any v_* fields which are part of unions. In
2937 * particular, in DragonFly the mount point overloading
2938 * mechanism uses the namecache only and the underlying
2939 * directory vnode is not a special case.
2943 if (vp->v_type == VBLK || vp->v_type == VCHR)
2945 else if (vp->v_mount)
2946 bsize = vp->v_mount->mnt_stat.f_iosize;
2950 maxsize = size + (loffset & PAGE_MASK);
2951 maxsize = imax(maxsize, bsize);
2953 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
2955 if (slpflags || slptimeo)
2961 * Atomically insert the buffer into the hash, so that it can
2962 * be found by findblk().
2964 * If bgetvp() returns non-zero a collision occured, and the
2965 * bp will not be associated with the vnode.
2967 * Make sure the translation layer has been cleared.
2969 bp->b_loffset = loffset;
2970 bp->b_bio2.bio_offset = NOOFFSET;
2971 /* bp->b_bio2.bio_next = NULL; */
2973 if (bgetvp(vp, bp, size)) {
2974 bp->b_flags |= B_INVAL;
2980 * All vnode-based buffers must be backed by a VM object.
2982 * Set B_KVABIO for incidental work
2984 KKASSERT(vp->v_object != NULL);
2985 bp->b_flags |= B_VMIO | B_KVABIO;
2986 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2992 * Do the nasty smp broadcast (if the buffer needs it) when KVABIO
2995 if (bp && (blkflags & GETBLK_KVABIO) == 0) {
3004 * Reacquire a buffer that was previously released to the locked queue,
3005 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
3006 * set B_LOCKED (which handles the acquisition race).
3008 * To this end, either B_LOCKED must be set or the dependancy list must be
3012 regetblk(struct buf *bp)
3014 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
3015 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
3022 * This code constitutes the buffer memory from either anonymous system
3023 * memory (in the case of non-VMIO operations) or from an associated
3024 * VM object (in the case of VMIO operations). This code is able to
3025 * resize a buffer up or down.
3027 * Note that this code is tricky, and has many complications to resolve
3028 * deadlock or inconsistant data situations. Tread lightly!!!
3029 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3030 * the caller. Calling this code willy nilly can result in the loss of
3033 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3034 * B_CACHE for the non-VMIO case.
3036 * This routine does not need to be called from a critical section but you
3037 * must own the buffer.
3040 allocbuf(struct buf *bp, int size)
3047 if (BUF_LOCKINUSE(bp) == 0)
3048 panic("allocbuf: buffer not busy");
3050 if (bp->b_kvasize < size)
3051 panic("allocbuf: buffer too small");
3053 KKASSERT(bp->b_flags & B_VMIO);
3055 newbsize = roundup2(size, DEV_BSIZE);
3056 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3057 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3058 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3061 * Set B_CACHE initially if buffer is 0 length or will become
3064 if (size == 0 || bp->b_bufsize == 0)
3065 bp->b_flags |= B_CACHE;
3067 if (newbsize < bp->b_bufsize) {
3069 * DEV_BSIZE aligned new buffer size is less then the
3070 * DEV_BSIZE aligned existing buffer size. Figure out
3071 * if we have to remove any pages.
3073 if (desiredpages < bp->b_xio.xio_npages) {
3074 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3076 * the page is not freed here -- it
3077 * is the responsibility of
3078 * vnode_pager_setsize
3080 m = bp->b_xio.xio_pages[i];
3081 KASSERT(m != bogus_page,
3082 ("allocbuf: bogus page found"));
3083 vm_page_busy_wait(m, TRUE, "biodep");
3084 bp->b_xio.xio_pages[i] = NULL;
3085 vm_page_unwire(m, 0);
3088 pmap_qremove_noinval((vm_offset_t)
3089 trunc_page((vm_offset_t)bp->b_data) +
3090 (desiredpages << PAGE_SHIFT),
3091 (bp->b_xio.xio_npages - desiredpages));
3092 bp->b_xio.xio_npages = desiredpages;
3095 * Don't bother invalidating the pmap changes
3096 * (which wastes global SMP invalidation IPIs)
3097 * when setting the size to 0. This case occurs
3098 * when called via getnewbuf() during buffer
3101 if (desiredpages == 0) {
3102 CPUMASK_ASSZERO(bp->b_cpumask);
3107 } else if (size > bp->b_bcount) {
3109 * We are growing the buffer, possibly in a
3110 * byte-granular fashion.
3118 * Step 1, bring in the VM pages from the object,
3119 * allocating them if necessary. We must clear
3120 * B_CACHE if these pages are not valid for the
3121 * range covered by the buffer.
3126 vm_object_hold(obj);
3127 while (bp->b_xio.xio_npages < desiredpages) {
3132 pi = OFF_TO_IDX(bp->b_loffset) +
3133 bp->b_xio.xio_npages;
3136 * Blocking on m->busy_count might lead to a
3139 * vm_fault->getpages->cluster_read->allocbuf
3141 m = vm_page_lookup_busy_try(obj, pi, FALSE,
3144 vm_page_sleep_busy(m, FALSE, "pgtblk");
3149 * note: must allocate system pages
3150 * since blocking here could intefere
3151 * with paging I/O, no matter which
3154 m = bio_page_alloc(bp, obj, pi,
3156 bp->b_xio.xio_npages);
3160 bp->b_flags &= ~B_CACHE;
3161 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3162 ++bp->b_xio.xio_npages;
3168 * We found a page and were able to busy it.
3172 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3173 ++bp->b_xio.xio_npages;
3174 if (bp->b_act_count < m->act_count)
3175 bp->b_act_count = m->act_count;
3177 vm_object_drop(obj);
3180 * Step 2. We've loaded the pages into the buffer,
3181 * we have to figure out if we can still have B_CACHE
3182 * set. Note that B_CACHE is set according to the
3183 * byte-granular range ( bcount and size ), not the
3184 * aligned range ( newbsize ).
3186 * The VM test is against m->valid, which is DEV_BSIZE
3187 * aligned. Needless to say, the validity of the data
3188 * needs to also be DEV_BSIZE aligned. Note that this
3189 * fails with NFS if the server or some other client
3190 * extends the file's EOF. If our buffer is resized,
3191 * B_CACHE may remain set! XXX
3194 toff = bp->b_bcount;
3195 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3197 while ((bp->b_flags & B_CACHE) && toff < size) {
3200 if (tinc > (size - toff))
3203 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3211 bp->b_xio.xio_pages[pi]
3218 * Step 3, fixup the KVM pmap. Remember that
3219 * bp->b_data is relative to bp->b_loffset, but
3220 * bp->b_loffset may be offset into the first page.
3222 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data);
3223 pmap_qenter_noinval((vm_offset_t)bp->b_data,
3224 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3225 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3226 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3229 atomic_add_long(&bufspace, newbsize - bp->b_bufsize);
3231 /* adjust space use on already-dirty buffer */
3232 if (bp->b_flags & B_DELWRI) {
3233 /* dirtykvaspace unchanged */
3234 atomic_add_long(&dirtybufspace, newbsize - bp->b_bufsize);
3235 if (bp->b_flags & B_HEAVY) {
3236 atomic_add_long(&dirtybufspacehw,
3237 newbsize - bp->b_bufsize);
3240 bp->b_bufsize = newbsize; /* actual buffer allocation */
3241 bp->b_bcount = size; /* requested buffer size */
3248 * Wait for buffer I/O completion, returning error status. B_EINTR
3249 * is converted into an EINTR error but not cleared (since a chain
3250 * of biowait() calls may occur).
3252 * On return bpdone() will have been called but the buffer will remain
3253 * locked and will not have been brelse()'d.
3255 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3256 * likely still in progress on return.
3258 * NOTE! This operation is on a BIO, not a BUF.
3260 * NOTE! BIO_DONE is cleared by vn_strategy()
3263 _biowait(struct bio *bio, const char *wmesg, int to)
3265 struct buf *bp = bio->bio_buf;
3270 KKASSERT(bio == &bp->b_bio1);
3272 flags = bio->bio_flags;
3273 if (flags & BIO_DONE)
3275 nflags = flags | BIO_WANT;
3276 tsleep_interlock(bio, 0);
3277 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3279 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3280 else if (bp->b_cmd == BUF_CMD_READ)
3281 error = tsleep(bio, PINTERLOCKED, "biord", to);
3283 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3285 kprintf("tsleep error biowait %d\n", error);
3294 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3295 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3296 if (bp->b_flags & B_EINTR)
3298 if (bp->b_flags & B_ERROR)
3299 return (bp->b_error ? bp->b_error : EIO);
3304 biowait(struct bio *bio, const char *wmesg)
3306 return(_biowait(bio, wmesg, 0));
3310 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3312 return(_biowait(bio, wmesg, to));
3316 * This associates a tracking count with an I/O. vn_strategy() and
3317 * dev_dstrategy() do this automatically but there are a few cases
3318 * where a vnode or device layer is bypassed when a block translation
3319 * is cached. In such cases bio_start_transaction() may be called on
3320 * the bypassed layers so the system gets an I/O in progress indication
3321 * for those higher layers.
3324 bio_start_transaction(struct bio *bio, struct bio_track *track)
3326 bio->bio_track = track;
3327 bio_track_ref(track);
3328 dsched_buf_enter(bio->bio_buf); /* might stack */
3332 * Initiate I/O on a vnode.
3334 * SWAPCACHE OPERATION:
3336 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately
3337 * devfs also uses b_vp for fake buffers so we also have to check
3338 * that B_PAGING is 0. In this case the passed 'vp' is probably the
3339 * underlying block device. The swap assignments are related to the
3340 * buffer cache buffer's b_vp, not the passed vp.
3342 * The passed vp == bp->b_vp only in the case where the strategy call
3343 * is made on the vp itself for its own buffers (a regular file or
3344 * block device vp). The filesystem usually then re-calls vn_strategy()
3345 * after translating the request to an underlying device.
3347 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the
3348 * underlying buffer cache buffers.
3350 * We can only deal with page-aligned buffers at the moment, because
3351 * we can't tell what the real dirty state for pages straddling a buffer
3354 * In order to call swap_pager_strategy() we must provide the VM object
3355 * and base offset for the underlying buffer cache pages so it can find
3359 vn_strategy(struct vnode *vp, struct bio *bio)
3361 struct bio_track *track;
3362 struct buf *bp = bio->bio_buf;
3364 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3367 * Set when an I/O is issued on the bp. Cleared by consumers
3368 * (aka HAMMER), allowing the consumer to determine if I/O had
3369 * actually occurred.
3371 bp->b_flags |= B_IOISSUED;
3374 * Handle the swapcache intercept.
3376 * NOTE: The swapcache itself always supports KVABIO and will
3377 * do the right thing if its underlying devices do not.
3379 if (vn_cache_strategy(vp, bio))
3383 * If the vnode does not support KVABIO and the buffer is using
3384 * KVABIO, we must synchronize b_data to all cpus before dispatching.
3386 if ((vp->v_flag & VKVABIO) == 0 && (bp->b_flags & B_KVABIO))
3390 * Otherwise do the operation through the filesystem
3392 if (bp->b_cmd == BUF_CMD_READ)
3393 track = &vp->v_track_read;
3395 track = &vp->v_track_write;
3396 KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3397 bio->bio_track = track;
3398 bio_track_ref(track);
3399 dsched_buf_enter(bp); /* might stack */
3400 vop_strategy(*vp->v_ops, vp, bio);
3404 * vn_cache_strategy()
3406 * Returns 1 if the interrupt was successful, 0 if not.
3408 * NOTE: This function supports the KVABIO API wherein b_data might not
3409 * be synchronized to the current cpu.
3411 static void vn_cache_strategy_callback(struct bio *bio);
3414 vn_cache_strategy(struct vnode *vp, struct bio *bio)
3416 struct buf *bp = bio->bio_buf;
3423 * Stop using swapcache if paniced, dumping, or dumped
3425 if (panicstr || dumping)
3429 * Is this buffer cache buffer suitable for reading from
3432 if (vm_swapcache_read_enable == 0 ||
3433 bp->b_cmd != BUF_CMD_READ ||
3434 ((bp->b_flags & B_CLUSTER) == 0 &&
3435 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) ||
3436 ((int)bp->b_loffset & PAGE_MASK) != 0 ||
3437 (bp->b_bcount & PAGE_MASK) != 0) {
3442 * Figure out the original VM object (it will match the underlying
3443 * VM pages). Note that swap cached data uses page indices relative
3444 * to that object, not relative to bio->bio_offset.
3446 if (bp->b_flags & B_CLUSTER)
3447 object = vp->v_object;
3449 object = bp->b_vp->v_object;
3452 * In order to be able to use the swap cache all underlying VM
3453 * pages must be marked as such, and we can't have any bogus pages.
3455 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
3456 m = bp->b_xio.xio_pages[i];
3457 if ((m->flags & PG_SWAPPED) == 0)
3459 if (m == bogus_page)
3464 * If we are good then issue the I/O using swap_pager_strategy().
3466 * We can only do this if the buffer actually supports object-backed
3467 * I/O. If it doesn't npages will be 0.
3469 if (i && i == bp->b_xio.xio_npages) {
3470 m = bp->b_xio.xio_pages[0];
3471 nbio = push_bio(bio);
3472 nbio->bio_done = vn_cache_strategy_callback;
3473 nbio->bio_offset = ptoa(m->pindex);
3474 KKASSERT(m->object == object);
3475 swap_pager_strategy(object, nbio);
3482 * This is a bit of a hack but since the vn_cache_strategy() function can
3483 * override a VFS's strategy function we must make sure that the bio, which
3484 * is probably bio2, doesn't leak an unexpected offset value back to the
3485 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the
3486 * bio went through its own file strategy function and the the bio2 offset
3487 * is a cached disk offset when, in fact, it isn't.
3490 vn_cache_strategy_callback(struct bio *bio)
3492 bio->bio_offset = NOOFFSET;
3493 biodone(pop_bio(bio));
3499 * Finish I/O on a buffer after all BIOs have been processed.
3500 * Called when the bio chain is exhausted or by biowait. If called
3501 * by biowait, elseit is typically 0.
3503 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3504 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3505 * assuming B_INVAL is clear.
3507 * For the VMIO case, we set B_CACHE if the op was a read and no
3508 * read error occured, or if the op was a write. B_CACHE is never
3509 * set if the buffer is invalid or otherwise uncacheable.
3511 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3512 * initiator to leave B_INVAL set to brelse the buffer out of existance
3513 * in the biodone routine.
3515 * bpdone is responsible for calling bundirty() on the buffer after a
3516 * successful write. We previously did this prior to initiating the
3517 * write under the assumption that the buffer might be dirtied again
3518 * while the write was in progress, however doing it before-hand creates
3519 * a race condition prior to the call to vn_strategy() where the
3520 * filesystem may not be aware that a dirty buffer is present.
3521 * It should not be possible for the buffer or its underlying pages to
3522 * be redirtied prior to bpdone()'s unbusying of the underlying VM
3526 bpdone(struct buf *bp, int elseit)
3530 KASSERT(BUF_LOCKINUSE(bp), ("bpdone: bp %p not busy", bp));
3531 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3532 ("bpdone: bp %p already done!", bp));
3535 * No more BIOs are left. All completion functions have been dealt
3536 * with, now we clean up the buffer.
3539 bp->b_cmd = BUF_CMD_DONE;
3542 * Only reads and writes are processed past this point.
3544 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3545 if (cmd == BUF_CMD_FREEBLKS)
3546 bp->b_flags |= B_NOCACHE;
3553 * A failed write must re-dirty the buffer unless B_INVAL
3556 * A successful write must clear the dirty flag. This is done after
3557 * the write to ensure that the buffer remains on the vnode's dirty
3558 * list for filesystem interlocks / checks until the write is actually
3559 * complete. HAMMER2 is sensitive to this issue.
3561 * Only applicable to normal buffers (with VPs). vinum buffers may
3564 * Must be done prior to calling buf_complete() as the callback might
3565 * re-dirty the buffer.
3567 if (cmd == BUF_CMD_WRITE) {
3568 if ((bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3569 bp->b_flags &= ~B_NOCACHE;
3579 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3580 * a lot worse. XXX - move this above the clearing of b_cmd
3582 if (LIST_FIRST(&bp->b_dep) != NULL)
3585 if (bp->b_flags & B_VMIO) {
3591 struct vnode *vp = bp->b_vp;
3595 #if defined(VFS_BIO_DEBUG)
3596 if (vp->v_auxrefs == 0)
3597 panic("bpdone: zero vnode hold count");
3598 if ((vp->v_flag & VOBJBUF) == 0)
3599 panic("bpdone: vnode is not setup for merged cache");
3602 foff = bp->b_loffset;
3603 KASSERT(foff != NOOFFSET, ("bpdone: no buffer offset"));
3604 KASSERT(obj != NULL, ("bpdone: missing VM object"));
3606 #if defined(VFS_BIO_DEBUG)
3607 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3608 kprintf("bpdone: paging in progress(%d) < "
3609 "bp->b_xio.xio_npages(%d)\n",
3610 obj->paging_in_progress,
3611 bp->b_xio.xio_npages);
3616 * Set B_CACHE if the op was a normal read and no error
3617 * occured. B_CACHE is set for writes in the b*write()
3620 iosize = bp->b_bcount - bp->b_resid;
3621 if (cmd == BUF_CMD_READ &&
3622 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3623 bp->b_flags |= B_CACHE;
3626 vm_object_hold(obj);
3627 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3631 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3636 * cleanup bogus pages, restoring the originals. Since
3637 * the originals should still be wired, we don't have
3638 * to worry about interrupt/freeing races destroying
3639 * the VM object association.
3641 m = bp->b_xio.xio_pages[i];
3642 if (m == bogus_page) {
3643 if ((bp->b_flags & B_HASBOGUS) == 0)
3644 panic("bpdone: bp %p corrupt bogus", bp);
3645 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3647 panic("bpdone: page disappeared");
3648 bp->b_xio.xio_pages[i] = m;
3653 #if defined(VFS_BIO_DEBUG)
3654 if (OFF_TO_IDX(foff) != m->pindex) {
3655 kprintf("bpdone: foff(%lu)/m->pindex(%ld) "
3657 (unsigned long)foff, (long)m->pindex);
3662 * In the write case, the valid and clean bits are
3663 * already changed correctly (see bdwrite()), so we
3664 * only need to do this here in the read case.
3666 vm_page_busy_wait(m, FALSE, "bpdpgw");
3667 if (cmd == BUF_CMD_READ && isbogus == 0 && resid > 0)
3668 vfs_clean_one_page(bp, i, m);
3671 * when debugging new filesystems or buffer I/O
3672 * methods, this is the most common error that pops
3673 * up. if you see this, you have not set the page
3674 * busy flag correctly!!!
3676 if ((m->busy_count & PBUSY_MASK) == 0) {
3677 kprintf("bpdone: page busy < 0, "
3678 "pindex: %d, foff: 0x(%x,%x), "
3679 "resid: %d, index: %d\n",
3680 (int) m->pindex, (int)(foff >> 32),
3681 (int) foff & 0xffffffff, resid, i);
3682 if (!vn_isdisk(vp, NULL))
3683 kprintf(" iosize: %ld, loffset: %lld, "
3684 "flags: 0x%08x, npages: %d\n",
3685 bp->b_vp->v_mount->mnt_stat.f_iosize,
3686 (long long)bp->b_loffset,
3687 bp->b_flags, bp->b_xio.xio_npages);
3689 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3690 (long long)bp->b_loffset,
3691 bp->b_flags, bp->b_xio.xio_npages);
3692 kprintf(" valid: 0x%x, dirty: 0x%x, "
3696 panic("bpdone: page busy < 0");
3698 vm_page_io_finish(m);
3700 vm_object_pip_wakeup(obj);
3701 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3704 if (bp->b_flags & B_HASBOGUS) {
3705 pmap_qenter_noinval(trunc_page((vm_offset_t)bp->b_data),
3706 bp->b_xio.xio_pages,
3707 bp->b_xio.xio_npages);
3708 bp->b_flags &= ~B_HASBOGUS;
3711 vm_object_drop(obj);
3715 * Finish up by releasing the buffer. There are no more synchronous
3716 * or asynchronous completions, those were handled by bio_done
3720 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
3731 biodone(struct bio *bio)
3733 struct buf *bp = bio->bio_buf;
3735 runningbufwakeup(bp);
3738 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3741 biodone_t *done_func;
3742 struct bio_track *track;
3745 * BIO tracking. Most but not all BIOs are tracked.
3747 if ((track = bio->bio_track) != NULL) {
3748 bio_track_rel(track);
3749 bio->bio_track = NULL;
3753 * A bio_done function terminates the loop. The function
3754 * will be responsible for any further chaining and/or
3755 * buffer management.
3757 * WARNING! The done function can deallocate the buffer!
3759 if ((done_func = bio->bio_done) != NULL) {
3760 bio->bio_done = NULL;
3764 bio = bio->bio_prev;
3768 * If we've run out of bio's do normal [a]synchronous completion.
3774 * Synchronous biodone - this terminates a synchronous BIO.
3776 * bpdone() is called with elseit=FALSE, leaving the buffer completed
3777 * but still locked. The caller must brelse() the buffer after waiting
3781 biodone_sync(struct bio *bio)
3783 struct buf *bp = bio->bio_buf;
3787 KKASSERT(bio == &bp->b_bio1);
3791 flags = bio->bio_flags;
3792 nflags = (flags | BIO_DONE) & ~BIO_WANT;
3794 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3795 if (flags & BIO_WANT)
3805 * This routine is called in lieu of iodone in the case of
3806 * incomplete I/O. This keeps the busy status for pages
3810 vfs_unbusy_pages(struct buf *bp)
3814 runningbufwakeup(bp);
3816 if (bp->b_flags & B_VMIO) {
3817 struct vnode *vp = bp->b_vp;
3821 vm_object_hold(obj);
3823 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3824 vm_page_t m = bp->b_xio.xio_pages[i];
3827 * When restoring bogus changes the original pages
3828 * should still be wired, so we are in no danger of
3829 * losing the object association and do not need
3830 * critical section protection particularly.
3832 if (m == bogus_page) {
3833 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3835 panic("vfs_unbusy_pages: page missing");
3837 bp->b_xio.xio_pages[i] = m;
3839 vm_page_busy_wait(m, FALSE, "bpdpgw");
3840 vm_page_io_finish(m);
3842 vm_object_pip_wakeup(obj);
3844 if (bp->b_flags & B_HASBOGUS) {
3845 pmap_qenter_noinval(trunc_page((vm_offset_t)bp->b_data),
3846 bp->b_xio.xio_pages,
3847 bp->b_xio.xio_npages);
3848 bp->b_flags &= ~B_HASBOGUS;
3851 vm_object_drop(obj);
3858 * This routine is called before a device strategy routine.
3859 * It is used to tell the VM system that paging I/O is in
3860 * progress, and treat the pages associated with the buffer
3861 * almost as being PBUSY_LOCKED. Also the object 'paging_in_progress'
3862 * flag is handled to make sure that the object doesn't become
3865 * Since I/O has not been initiated yet, certain buffer flags
3866 * such as B_ERROR or B_INVAL may be in an inconsistant state
3867 * and should be ignored.
3870 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3873 struct lwp *lp = curthread->td_lwp;
3876 * The buffer's I/O command must already be set. If reading,
3877 * B_CACHE must be 0 (double check against callers only doing
3878 * I/O when B_CACHE is 0).
3880 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3881 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3883 if (bp->b_flags & B_VMIO) {
3887 KASSERT(bp->b_loffset != NOOFFSET,
3888 ("vfs_busy_pages: no buffer offset"));
3891 * Busy all the pages. We have to busy them all at once
3892 * to avoid deadlocks.
3895 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3896 vm_page_t m = bp->b_xio.xio_pages[i];
3898 if (vm_page_busy_try(m, FALSE)) {
3899 vm_page_sleep_busy(m, FALSE, "vbpage");
3901 vm_page_wakeup(bp->b_xio.xio_pages[i]);
3907 * Setup for I/O, soft-busy the page right now because
3908 * the next loop may block.
3910 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3911 vm_page_t m = bp->b_xio.xio_pages[i];
3913 if ((bp->b_flags & B_CLUSTER) == 0) {
3914 vm_object_pip_add(obj, 1);
3915 vm_page_io_start(m);
3920 * Adjust protections for I/O and do bogus-page mapping.
3921 * Assume that vm_page_protect() can block (it can block
3922 * if VM_PROT_NONE, don't take any chances regardless).
3924 * In particular note that for writes we must incorporate
3925 * page dirtyness from the VM system into the buffer's
3928 * For reads we theoretically must incorporate page dirtyness
3929 * from the VM system to determine if the page needs bogus
3930 * replacement, but we shortcut the test by simply checking
3931 * that all m->valid bits are set, indicating that the page
3932 * is fully valid and does not need to be re-read. For any
3933 * VM system dirtyness the page will also be fully valid
3934 * since it was mapped at one point.
3937 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3938 vm_page_t m = bp->b_xio.xio_pages[i];
3940 if (bp->b_cmd == BUF_CMD_WRITE) {
3942 * When readying a vnode-backed buffer for
3943 * a write we must zero-fill any invalid
3944 * portions of the backing VM pages, mark
3945 * it valid and clear related dirty bits.
3947 * vfs_clean_one_page() incorporates any
3948 * VM dirtyness and updates the b_dirtyoff
3949 * range (after we've made the page RO).
3951 * It is also expected that the pmap modified
3952 * bit has already been cleared by the
3953 * vm_page_protect(). We may not be able
3954 * to clear all dirty bits for a page if it
3955 * was also memory mapped (NFS).
3957 * Finally be sure to unassign any swap-cache
3958 * backing store as it is now stale.
3960 vm_page_protect(m, VM_PROT_READ);
3961 vfs_clean_one_page(bp, i, m);
3962 swap_pager_unswapped(m);
3963 } else if (m->valid == VM_PAGE_BITS_ALL) {
3965 * When readying a vnode-backed buffer for
3966 * read we must replace any dirty pages with
3967 * a bogus page so dirty data is not destroyed
3968 * when filling gaps.
3970 * To avoid testing whether the page is
3971 * dirty we instead test that the page was
3972 * at some point mapped (m->valid fully
3973 * valid) with the understanding that
3974 * this also covers the dirty case.
3976 bp->b_xio.xio_pages[i] = bogus_page;
3977 bp->b_flags |= B_HASBOGUS;
3979 } else if (m->valid & m->dirty) {
3981 * This case should not occur as partial
3982 * dirtyment can only happen if the buffer
3983 * is B_CACHE, and this code is not entered
3984 * if the buffer is B_CACHE.
3986 kprintf("Warning: vfs_busy_pages - page not "
3987 "fully valid! loff=%jx bpf=%08x "
3988 "idx=%d val=%02x dir=%02x\n",
3989 (uintmax_t)bp->b_loffset, bp->b_flags,
3990 i, m->valid, m->dirty);
3991 vm_page_protect(m, VM_PROT_NONE);
3994 * The page is not valid and can be made
3997 vm_page_protect(m, VM_PROT_NONE);
4002 pmap_qenter_noinval(trunc_page((vm_offset_t)bp->b_data),
4003 bp->b_xio.xio_pages,
4004 bp->b_xio.xio_npages);
4010 * This is the easiest place to put the process accounting for the I/O
4014 if (bp->b_cmd == BUF_CMD_READ)
4015 lp->lwp_ru.ru_inblock++;
4017 lp->lwp_ru.ru_oublock++;
4022 * Tell the VM system that the pages associated with this buffer
4023 * are clean. This is used for delayed writes where the data is
4024 * going to go to disk eventually without additional VM intevention.
4026 * NOTE: While we only really need to clean through to b_bcount, we
4027 * just go ahead and clean through to b_bufsize.
4030 vfs_clean_pages(struct buf *bp)
4035 if ((bp->b_flags & B_VMIO) == 0)
4038 KASSERT(bp->b_loffset != NOOFFSET,
4039 ("vfs_clean_pages: no buffer offset"));
4041 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4042 m = bp->b_xio.xio_pages[i];
4043 vfs_clean_one_page(bp, i, m);
4048 * vfs_clean_one_page:
4050 * Set the valid bits and clear the dirty bits in a page within a
4051 * buffer. The range is restricted to the buffer's size and the
4052 * buffer's logical offset might index into the first page.
4054 * The caller has busied or soft-busied the page and it is not mapped,
4055 * test and incorporate the dirty bits into b_dirtyoff/end before
4056 * clearing them. Note that we need to clear the pmap modified bits
4057 * after determining the the page was dirty, vm_page_set_validclean()
4058 * does not do it for us.
4060 * This routine is typically called after a read completes (dirty should
4061 * be zero in that case as we are not called on bogus-replace pages),
4062 * or before a write is initiated.
4065 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m)
4073 * Calculate offset range within the page but relative to buffer's
4074 * loffset. loffset might be offset into the first page.
4076 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4077 bcount = bp->b_bcount + xoff; /* offset adjusted */
4083 soff = (pageno << PAGE_SHIFT);
4084 eoff = soff + PAGE_SIZE;
4092 * Test dirty bits and adjust b_dirtyoff/end.
4094 * If dirty pages are incorporated into the bp any prior
4095 * B_NEEDCOMMIT state (NFS) must be cleared because the
4096 * caller has not taken into account the new dirty data.
4098 * If the page was memory mapped the dirty bits might go beyond the
4099 * end of the buffer, but we can't really make the assumption that
4100 * a file EOF straddles the buffer (even though this is the case for
4101 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
4102 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
4103 * This also saves some console spam.
4105 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK,
4106 * NFS can handle huge commits but not huge writes.
4108 vm_page_test_dirty(m);
4110 if ((bp->b_flags & B_NEEDCOMMIT) &&
4111 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) {
4113 kprintf("Warning: vfs_clean_one_page: bp %p "
4114 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT"
4115 " cmd %d vd %02x/%02x x/s/e %d %d %d "
4117 bp, (uintmax_t)bp->b_loffset, bp->b_bcount,
4118 bp->b_flags, bp->b_cmd,
4119 m->valid, m->dirty, xoff, soff, eoff,
4120 bp->b_dirtyoff, bp->b_dirtyend);
4121 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
4123 print_backtrace(-1);
4126 * Only clear the pmap modified bits if ALL the dirty bits
4127 * are set, otherwise the system might mis-clear portions
4130 if (m->dirty == VM_PAGE_BITS_ALL &&
4131 (bp->b_flags & B_NEEDCOMMIT) == 0) {
4132 pmap_clear_modify(m);
4134 if (bp->b_dirtyoff > soff - xoff)
4135 bp->b_dirtyoff = soff - xoff;
4136 if (bp->b_dirtyend < eoff - xoff)
4137 bp->b_dirtyend = eoff - xoff;
4141 * Set related valid bits, clear related dirty bits.
4142 * Does not mess with the pmap modified bit.
4144 * WARNING! We cannot just clear all of m->dirty here as the
4145 * buffer cache buffers may use a DEV_BSIZE'd aligned
4146 * block size, or have an odd size (e.g. NFS at file EOF).
4147 * The putpages code can clear m->dirty to 0.
4149 * If a VOP_WRITE generates a buffer cache buffer which
4150 * covers the same space as mapped writable pages the
4151 * buffer flush might not be able to clear all the dirty
4152 * bits and still require a putpages from the VM system
4155 * WARNING! vm_page_set_validclean() currently assumes vm_token
4156 * is held. The page might not be busied (bdwrite() case).
4157 * XXX remove this comment once we've validated that this
4158 * is no longer an issue.
4160 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff);
4165 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty.
4166 * The page data is assumed to be valid (there is no zeroing here).
4169 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m)
4177 * Calculate offset range within the page but relative to buffer's
4178 * loffset. loffset might be offset into the first page.
4180 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4181 bcount = bp->b_bcount + xoff; /* offset adjusted */
4187 soff = (pageno << PAGE_SHIFT);
4188 eoff = soff + PAGE_SIZE;
4194 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff);
4201 * Clear a buffer. This routine essentially fakes an I/O, so we need
4202 * to clear B_ERROR and B_INVAL.
4204 * Note that while we only theoretically need to clear through b_bcount,
4205 * we go ahead and clear through b_bufsize.
4208 vfs_bio_clrbuf(struct buf *bp)
4212 KKASSERT(bp->b_flags & B_VMIO);
4214 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
4217 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4218 (bp->b_loffset & PAGE_MASK) == 0) {
4219 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4220 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
4224 if ((bp->b_xio.xio_pages[0]->valid & mask) == 0) {
4225 bzero(bp->b_data, bp->b_bufsize);
4226 bp->b_xio.xio_pages[0]->valid |= mask;
4232 for(i = 0; i < bp->b_xio.xio_npages; i++, sa=ea) {
4233 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
4234 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
4235 ea = (caddr_t)(vm_offset_t)ulmin(
4236 (u_long)(vm_offset_t)ea,
4237 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
4238 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4239 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
4241 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
4244 for (; sa < ea; sa += DEV_BSIZE, j++) {
4245 if ((bp->b_xio.xio_pages[i]->valid &
4247 bzero(sa, DEV_BSIZE);
4251 bp->b_xio.xio_pages[i]->valid |= mask;
4257 * Allocate a page for a buffer cache buffer.
4259 * If NULL is returned the caller is expected to retry (typically check if
4260 * the page already exists on retry before trying to allocate one).
4262 * NOTE! Low-memory handling is dealt with in b[q]relse(), not here. This
4263 * function will use the system reserve with the hope that the page
4264 * allocations can be returned to PQ_CACHE/PQ_FREE when the caller
4265 * is done with the buffer.
4267 * NOTE! However, TMPFS is a special case because flushing a dirty buffer
4268 * to TMPFS doesn't clean the page. For TMPFS, only the pagedaemon
4269 * is capable of retiring pages (to swap). For TMPFS we don't dig
4270 * into the system reserve because doing so could stall out pretty
4271 * much every process running on the system.
4275 bio_page_alloc(struct buf *bp, vm_object_t obj, vm_pindex_t pg, int deficit)
4277 int vmflags = VM_ALLOC_NORMAL | VM_ALLOC_NULL_OK;
4280 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj));
4285 vmflags |= VM_ALLOC_CPU(obj->pg_color % ncpus);
4288 * Try a normal allocation first.
4290 p = vm_page_alloc(obj, pg, vmflags);
4293 if (vm_page_lookup(obj, pg))
4295 vm_pageout_deficit += deficit;
4298 * Try again, digging into the system reserve.
4300 * Trying to recover pages from the buffer cache here can deadlock
4301 * against other threads trying to busy underlying pages so we
4302 * depend on the code in brelse() and bqrelse() to free/cache the
4303 * underlying buffer cache pages when memory is low.
4305 if (curthread->td_flags & TDF_SYSTHREAD)
4306 vmflags |= VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT;
4307 else if (bp->b_vp && bp->b_vp->v_tag == VT_TMPFS)
4310 vmflags |= VM_ALLOC_SYSTEM;
4312 /*recoverbufpages();*/
4313 p = vm_page_alloc(obj, pg, vmflags);
4316 if (vm_page_lookup(obj, pg))
4320 * Wait for memory to free up and try again
4322 if (vm_page_count_severe())
4324 vm_wait(hz / 20 + 1);
4326 p = vm_page_alloc(obj, pg, vmflags);
4329 if (vm_page_lookup(obj, pg))
4333 * Ok, now we are really in trouble.
4336 static struct krate biokrate = { .freq = 1 };
4337 krateprintf(&biokrate,
4338 "Warning: bio_page_alloc: memory exhausted "
4339 "during buffer cache page allocation from %s\n",
4340 curthread->td_comm);
4342 if (curthread->td_flags & TDF_SYSTHREAD)
4343 vm_wait(hz / 20 + 1);
4345 vm_wait(hz / 2 + 1);
4350 * The buffer's mapping has changed. Adjust the buffer's memory
4351 * synchronization. The caller is the exclusive holder of the buffer
4352 * and has set or cleared B_KVABIO according to preference.
4354 * WARNING! If the caller is using B_KVABIO mode, this function will
4355 * not map the data to the current cpu. The caller must also
4356 * call bkvasync(bp).
4359 bkvareset(struct buf *bp)
4361 if (bp->b_flags & B_KVABIO) {
4362 CPUMASK_ASSZERO(bp->b_cpumask);
4364 CPUMASK_ORMASK(bp->b_cpumask, smp_active_mask);
4371 * The buffer will be used by the caller on the caller's cpu, synchronize
4372 * its data to the current cpu. Caller must control the buffer by holding
4373 * its lock, but calling cpu does not necessarily have to be the owner of
4374 * the lock (i.e. HAMMER2's concurrent I/O accessors).
4376 * If B_KVABIO is not set, the buffer is already fully synchronized.
4379 bkvasync(struct buf *bp)
4381 int cpuid = mycpu->gd_cpuid;
4384 if ((bp->b_flags & B_KVABIO) &&
4385 CPUMASK_TESTBIT(bp->b_cpumask, cpuid) == 0) {
4387 while (bdata < bp->b_data + bp->b_bufsize) {
4389 bdata += PAGE_SIZE -
4390 ((intptr_t)bdata & PAGE_MASK);
4392 ATOMIC_CPUMASK_ORBIT(bp->b_cpumask, cpuid);
4397 * The buffer will be used by a subsystem that does not understand
4398 * the KVABIO API. Make sure its data is synchronized to all cpus.
4400 * If B_KVABIO is not set, the buffer is already fully synchronized.
4402 * NOTE! This is the only safe way to clear B_KVABIO on a buffer.
4405 bkvasync_all(struct buf *bp)
4407 if (debug_kvabio > 0) {
4409 print_backtrace(10);
4412 if ((bp->b_flags & B_KVABIO) &&
4413 CPUMASK_CMPMASKNEQ(bp->b_cpumask, smp_active_mask)) {
4416 ATOMIC_CPUMASK_ORMASK(bp->b_cpumask, smp_active_mask);
4418 bp->b_flags &= ~B_KVABIO;
4422 * Scan all buffers in the system and issue the callback.
4425 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4431 for (n = 0; n < nbuf; ++n) {
4432 if ((error = callback(&buf[n], info)) < 0) {
4442 * nestiobuf_iodone: biodone callback for nested buffers and propagate
4443 * completion to the master buffer.
4446 nestiobuf_iodone(struct bio *bio)
4449 struct buf *mbp, *bp;
4450 struct devstat *stats;
4455 mbio = bio->bio_caller_info1.ptr;
4456 stats = bio->bio_caller_info2.ptr;
4457 mbp = mbio->bio_buf;
4459 KKASSERT(bp->b_bcount <= bp->b_bufsize);
4460 KKASSERT(mbp != bp);
4462 error = bp->b_error;
4463 if (bp->b_error == 0 &&
4464 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
4466 * Not all got transfered, raise an error. We have no way to
4467 * propagate these conditions to mbp.
4472 donebytes = bp->b_bufsize;
4476 nestiobuf_done(mbio, donebytes, error, stats);
4480 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats)
4484 mbp = mbio->bio_buf;
4486 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0);
4489 * If an error occured, propagate it to the master buffer.
4491 * Several biodone()s may wind up running concurrently so
4492 * use an atomic op to adjust b_flags.
4495 mbp->b_error = error;
4496 atomic_set_int(&mbp->b_flags, B_ERROR);
4500 * Decrement the operations in progress counter and terminate the
4501 * I/O if this was the last bit.
4503 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4506 devstat_end_transaction_buf(stats, mbp);
4512 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent
4513 * the mbio from being biodone()'d while we are still adding sub-bios to
4517 nestiobuf_init(struct bio *bio)
4519 bio->bio_driver_info = (void *)1;
4523 * The BIOs added to the nestedio have already been started, remove the
4524 * count that placeheld our mbio and biodone() it if the count would
4528 nestiobuf_start(struct bio *mbio)
4530 struct buf *mbp = mbio->bio_buf;
4533 * Decrement the operations in progress counter and terminate the
4534 * I/O if this was the last bit.
4536 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4537 if (mbp->b_flags & B_ERROR)
4538 mbp->b_resid = mbp->b_bcount;
4546 * Set an intermediate error prior to calling nestiobuf_start()
4549 nestiobuf_error(struct bio *mbio, int error)
4551 struct buf *mbp = mbio->bio_buf;
4554 mbp->b_error = error;
4555 atomic_set_int(&mbp->b_flags, B_ERROR);
4560 * nestiobuf_add: setup a "nested" buffer.
4562 * => 'mbp' is a "master" buffer which is being divided into sub pieces.
4563 * => 'bp' should be a buffer allocated by getiobuf.
4564 * => 'offset' is a byte offset in the master buffer.
4565 * => 'size' is a size in bytes of this nested buffer.
4568 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats)
4570 struct buf *mbp = mbio->bio_buf;
4571 struct vnode *vp = mbp->b_vp;
4573 KKASSERT(mbp->b_bcount >= offset + size);
4575 atomic_add_int((int *)&mbio->bio_driver_info, 1);
4577 /* kernel needs to own the lock for it to be released in biodone */
4580 bp->b_cmd = mbp->b_cmd;
4581 bp->b_bio1.bio_done = nestiobuf_iodone;
4582 bp->b_data = (char *)mbp->b_data + offset;
4583 bp->b_resid = bp->b_bcount = size;
4584 bp->b_bufsize = bp->b_bcount;
4586 bp->b_bio1.bio_track = NULL;
4587 bp->b_bio1.bio_caller_info1.ptr = mbio;
4588 bp->b_bio1.bio_caller_info2.ptr = stats;
4593 DB_SHOW_COMMAND(buffer, db_show_buffer)
4596 struct buf *bp = (struct buf *)addr;
4599 db_printf("usage: show buffer <addr>\n");
4603 db_printf("b_flags = 0x%pb%i\n", PRINT_BUF_FLAGS, bp->b_flags);
4604 db_printf("b_cmd = %d\n", bp->b_cmd);
4605 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4606 "b_resid = %d\n, b_data = %p, "
4607 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4608 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4610 (long long)bp->b_bio2.bio_offset,
4611 (long long)(bp->b_bio2.bio_next ?
4612 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4613 if (bp->b_xio.xio_npages) {
4615 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4616 bp->b_xio.xio_npages);
4617 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4619 m = bp->b_xio.xio_pages[i];
4620 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4621 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4622 if ((i + 1) < bp->b_xio.xio_npages)