2 * Copyright (c) 1994,1997 John S. Dyson
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice immediately at the beginning of the file, without modification,
10 * this list of conditions, and the following disclaimer.
11 * 2. Absolutely no warranty of function or purpose is made by the author
14 * $FreeBSD: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $
18 * this file contains a new buffer I/O scheme implementing a coherent
19 * VM object and buffer cache scheme. Pains have been taken to make
20 * sure that the performance degradation associated with schemes such
21 * as this is not realized.
23 * Author: John S. Dyson
24 * Significant help during the development and debugging phases
25 * had been provided by David Greenman, also of the FreeBSD core team.
27 * see man buf(9) for more info.
30 #include <sys/param.h>
31 #include <sys/systm.h>
34 #include <sys/devicestat.h>
35 #include <sys/eventhandler.h>
37 #include <sys/malloc.h>
38 #include <sys/mount.h>
39 #include <sys/kernel.h>
40 #include <sys/kthread.h>
42 #include <sys/reboot.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sysctl.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
47 #include <sys/dsched.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_kern.h>
51 #include <vm/vm_pageout.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_map.h>
56 #include <vm/vm_pager.h>
57 #include <vm/swap_pager.h>
60 #include <sys/thread2.h>
61 #include <sys/spinlock2.h>
62 #include <sys/mplock2.h>
63 #include <vm/vm_page2.h>
74 BQUEUE_NONE, /* not on any queue */
75 BQUEUE_LOCKED, /* locked buffers */
76 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
77 BQUEUE_DIRTY, /* B_DELWRI buffers */
78 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
79 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
80 BQUEUE_EMPTY, /* empty buffer headers */
82 BUFFER_QUEUES /* number of buffer queues */
85 typedef enum bufq_type bufq_type_t;
87 #define BD_WAKE_SIZE 16384
88 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
90 TAILQ_HEAD(bqueues, buf);
94 struct bqueues bufqueues[BUFFER_QUEUES];
97 struct bufpcpu bufpcpu[MAXCPU];
99 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
101 struct buf *buf; /* buffer header pool */
103 static void vfs_clean_pages(struct buf *bp);
104 static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
106 static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m);
108 static void vfs_vmio_release(struct buf *bp);
109 static int flushbufqueues(struct buf *marker, bufq_type_t q);
110 static vm_page_t bio_page_alloc(struct buf *bp, vm_object_t obj,
111 vm_pindex_t pg, int deficit);
113 static void bd_signal(long totalspace);
114 static void buf_daemon(void);
115 static void buf_daemon_hw(void);
118 * bogus page -- for I/O to/from partially complete buffers
119 * this is a temporary solution to the problem, but it is not
120 * really that bad. it would be better to split the buffer
121 * for input in the case of buffers partially already in memory,
122 * but the code is intricate enough already.
124 vm_page_t bogus_page;
127 * These are all static, but make the ones we export globals so we do
128 * not need to use compiler magic.
130 long bufspace; /* locked by buffer_map */
132 static long bufmallocspace; /* atomic ops */
133 long maxbufmallocspace, lobufspace, hibufspace;
134 static long bufreusecnt, bufdefragcnt, buffreekvacnt;
135 static long lorunningspace;
136 static long hirunningspace;
137 static long dirtykvaspace; /* atomic */
138 long dirtybufspace; /* atomic (global for systat) */
139 static long dirtybufcount; /* atomic */
140 static long dirtybufspacehw; /* atomic */
141 static long dirtybufcounthw; /* atomic */
142 static long runningbufspace; /* atomic */
143 static long runningbufcount; /* atomic */
144 long lodirtybufspace;
145 long hidirtybufspace;
146 static int getnewbufcalls;
147 static int getnewbufrestarts;
148 static int recoverbufcalls;
149 static int needsbuffer; /* atomic */
150 static int runningbufreq; /* atomic */
151 static int bd_request; /* atomic */
152 static int bd_request_hw; /* atomic */
153 static u_int bd_wake_ary[BD_WAKE_SIZE];
154 static u_int bd_wake_index;
155 static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */
156 static int debug_commit;
157 static int debug_bufbio;
159 static struct thread *bufdaemon_td;
160 static struct thread *bufdaemonhw_td;
161 static u_int lowmempgallocs;
162 static u_int lowmempgfails;
165 * Sysctls for operational control of the buffer cache.
167 SYSCTL_LONG(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
168 "Number of dirty buffers to flush before bufdaemon becomes inactive");
169 SYSCTL_LONG(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
170 "High watermark used to trigger explicit flushing of dirty buffers");
171 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
172 "Minimum amount of buffer space required for active I/O");
173 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
174 "Maximum amount of buffer space to usable for active I/O");
175 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0,
176 "Page allocations done during periods of very low free memory");
177 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0,
178 "Page allocations which failed during periods of very low free memory");
179 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
180 "Recycle pages to active or inactive queue transition pt 0-64");
182 * Sysctls determining current state of the buffer cache.
184 SYSCTL_LONG(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
185 "Total number of buffers in buffer cache");
186 SYSCTL_LONG(_vfs, OID_AUTO, dirtykvaspace, CTLFLAG_RD, &dirtykvaspace, 0,
187 "KVA reserved by dirty buffers (all)");
188 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
189 "Pending bytes of dirty buffers (all)");
190 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
191 "Pending bytes of dirty buffers (heavy weight)");
192 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
193 "Pending number of dirty buffers");
194 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
195 "Pending number of dirty buffers (heavy weight)");
196 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
197 "I/O bytes currently in progress due to asynchronous writes");
198 SYSCTL_LONG(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
199 "I/O buffers currently in progress due to asynchronous writes");
200 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
201 "Hard limit on maximum amount of memory usable for buffer space");
202 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
203 "Soft limit on maximum amount of memory usable for buffer space");
204 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
205 "Minimum amount of memory to reserve for system buffer space");
206 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
207 "Amount of memory available for buffers");
208 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
209 0, "Maximum amount of memory reserved for buffers using malloc");
210 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
211 "Amount of memory left for buffers using malloc-scheme");
212 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
213 "New buffer header acquisition requests");
214 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
215 0, "New buffer header acquisition restarts");
216 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
217 "Recover VM space in an emergency");
218 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
219 "Buffer acquisition restarts due to fragmented buffer map");
220 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
221 "Amount of time KVA space was deallocated in an arbitrary buffer");
222 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
223 "Amount of time buffer re-use operations were successful");
224 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, "");
225 SYSCTL_INT(_vfs, OID_AUTO, debug_bufbio, CTLFLAG_RW, &debug_bufbio, 0, "");
226 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
227 "sizeof(struct buf)");
229 char *buf_wmesg = BUF_WMESG;
231 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
232 #define VFS_BIO_NEED_UNUSED02 0x02
233 #define VFS_BIO_NEED_UNUSED04 0x04
234 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
239 * Called when buffer space is potentially available for recovery.
240 * getnewbuf() will block on this flag when it is unable to free
241 * sufficient buffer space. Buffer space becomes recoverable when
242 * bp's get placed back in the queues.
248 * If someone is waiting for BUF space, wake them up. Even
249 * though we haven't freed the kva space yet, the waiting
250 * process will be able to now.
253 int flags = needsbuffer;
255 if ((flags & VFS_BIO_NEED_BUFSPACE) == 0)
257 if (atomic_cmpset_int(&needsbuffer, flags,
258 flags & ~VFS_BIO_NEED_BUFSPACE)) {
259 wakeup(&needsbuffer);
269 * Accounting for I/O in progress.
273 runningbufwakeup(struct buf *bp)
278 if ((totalspace = bp->b_runningbufspace) != 0) {
279 atomic_add_long(&runningbufspace, -totalspace);
280 atomic_add_long(&runningbufcount, -1);
281 bp->b_runningbufspace = 0;
284 * see waitrunningbufspace() for limit test.
287 flags = runningbufreq;
291 if (atomic_cmpset_int(&runningbufreq, flags, 0)) {
292 wakeup(&runningbufreq);
297 bd_signal(totalspace);
304 * Called when a buffer has been added to one of the free queues to
305 * account for the buffer and to wakeup anyone waiting for free buffers.
306 * This typically occurs when large amounts of metadata are being handled
307 * by the buffer cache ( else buffer space runs out first, usually ).
318 if (atomic_cmpset_int(&needsbuffer, flags,
319 (flags & ~VFS_BIO_NEED_ANY))) {
320 wakeup(&needsbuffer);
328 * waitrunningbufspace()
330 * If runningbufspace exceeds 4/6 hirunningspace we block until
331 * runningbufspace drops to 3/6 hirunningspace. We also block if another
332 * thread blocked here in order to be fair, even if runningbufspace
333 * is now lower than the limit.
335 * The caller may be using this function to block in a tight loop, we
336 * must block while runningbufspace is greater than at least
337 * hirunningspace * 3 / 6.
340 waitrunningbufspace(void)
342 long limit = hirunningspace * 4 / 6;
345 while (runningbufspace > limit || runningbufreq) {
346 tsleep_interlock(&runningbufreq, 0);
347 flags = atomic_fetchadd_int(&runningbufreq, 1);
348 if (runningbufspace > limit || flags)
349 tsleep(&runningbufreq, PINTERLOCKED, "wdrn1", hz);
354 * buf_dirty_count_severe:
356 * Return true if we have too many dirty buffers.
359 buf_dirty_count_severe(void)
361 return (runningbufspace + dirtykvaspace >= hidirtybufspace ||
362 dirtybufcount >= nbuf / 2);
366 * Return true if the amount of running I/O is severe and BIOQ should
370 buf_runningbufspace_severe(void)
372 return (runningbufspace >= hirunningspace * 4 / 6);
376 * vfs_buf_test_cache:
378 * Called when a buffer is extended. This function clears the B_CACHE
379 * bit if the newly extended portion of the buffer does not contain
382 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
383 * cache buffers. The VM pages remain dirty, as someone had mmap()'d
384 * them while a clean buffer was present.
388 vfs_buf_test_cache(struct buf *bp,
389 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
392 if (bp->b_flags & B_CACHE) {
393 int base = (foff + off) & PAGE_MASK;
394 if (vm_page_is_valid(m, base, size) == 0)
395 bp->b_flags &= ~B_CACHE;
402 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
409 if (dirtykvaspace < lodirtybufspace && dirtybufcount < nbuf / 2)
412 if (bd_request == 0 &&
413 (dirtykvaspace > lodirtybufspace / 2 ||
414 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
415 if (atomic_fetchadd_int(&bd_request, 1) == 0)
418 if (bd_request_hw == 0 &&
419 (dirtykvaspace > lodirtybufspace / 2 ||
420 dirtybufcounthw >= nbuf / 2)) {
421 if (atomic_fetchadd_int(&bd_request_hw, 1) == 0)
422 wakeup(&bd_request_hw);
429 * Get the buf_daemon heated up when the number of running and dirty
430 * buffers exceeds the mid-point.
432 * Return the total number of dirty bytes past the second mid point
433 * as a measure of how much excess dirty data there is in the system.
442 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
444 totalspace = runningbufspace + dirtykvaspace;
445 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
447 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
448 if (totalspace >= mid2)
449 return(totalspace - mid2);
457 * Wait for the buffer cache to flush (totalspace) bytes worth of
458 * buffers, then return.
460 * Regardless this function blocks while the number of dirty buffers
461 * exceeds hidirtybufspace.
464 bd_wait(long totalspace)
471 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
474 while (totalspace > 0) {
478 * Order is important. Suppliers adjust bd_wake_index after
479 * updating runningbufspace/dirtykvaspace. We want to fetch
480 * bd_wake_index before accessing. Any error should thus
483 i = atomic_fetchadd_int(&bd_wake_index, 0);
484 if (totalspace > runningbufspace + dirtykvaspace)
485 totalspace = runningbufspace + dirtykvaspace;
486 count = totalspace / BKVASIZE;
487 if (count >= BD_WAKE_SIZE / 2)
488 count = BD_WAKE_SIZE / 2;
490 mi = i & BD_WAKE_MASK;
493 * This is not a strict interlock, so we play a bit loose
494 * with locking access to dirtybufspace*. We have to re-check
495 * bd_wake_index to ensure that it hasn't passed us.
497 tsleep_interlock(&bd_wake_ary[mi], 0);
498 atomic_add_int(&bd_wake_ary[mi], 1);
499 j = atomic_fetchadd_int(&bd_wake_index, 0);
500 if ((int)(i - j) >= 0)
501 tsleep(&bd_wake_ary[mi], PINTERLOCKED, "flstik", hz);
503 totalspace = runningbufspace + dirtykvaspace - hidirtybufspace;
510 * This function is called whenever runningbufspace or dirtykvaspace
511 * is reduced. Track threads waiting for run+dirty buffer I/O
515 bd_signal(long totalspace)
519 if (totalspace > 0) {
520 if (totalspace > BKVASIZE * BD_WAKE_SIZE)
521 totalspace = BKVASIZE * BD_WAKE_SIZE;
522 while (totalspace > 0) {
523 i = atomic_fetchadd_int(&bd_wake_index, 1);
525 if (atomic_readandclear_int(&bd_wake_ary[i]))
526 wakeup(&bd_wake_ary[i]);
527 totalspace -= BKVASIZE;
533 * BIO tracking support routines.
535 * Release a ref on a bio_track. Wakeup requests are atomically released
536 * along with the last reference so bk_active will never wind up set to
541 bio_track_rel(struct bio_track *track)
549 active = track->bk_active;
550 if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
554 * Full-on. Note that the wait flag is only atomically released on
555 * the 1->0 count transition.
557 * We check for a negative count transition using bit 30 since bit 31
558 * has a different meaning.
561 desired = (active & 0x7FFFFFFF) - 1;
563 desired |= active & 0x80000000;
564 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
565 if (desired & 0x40000000)
566 panic("bio_track_rel: bad count: %p", track);
567 if (active & 0x80000000)
571 active = track->bk_active;
576 * Wait for the tracking count to reach 0.
578 * Use atomic ops such that the wait flag is only set atomically when
579 * bk_active is non-zero.
582 bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
591 if (track->bk_active == 0)
595 * Full-on. Note that the wait flag may only be atomically set if
596 * the active count is non-zero.
598 * NOTE: We cannot optimize active == desired since a wakeup could
599 * clear active prior to our tsleep_interlock().
602 while ((active = track->bk_active) != 0) {
604 desired = active | 0x80000000;
605 tsleep_interlock(track, slp_flags);
606 if (atomic_cmpset_int(&track->bk_active, active, desired)) {
607 error = tsleep(track, slp_flags | PINTERLOCKED,
619 * Load time initialisation of the buffer cache, called from machine
620 * dependant initialization code.
624 bufinit(void *dummy __unused)
626 struct bufpcpu *pcpu;
628 vm_offset_t bogus_offset;
633 /* next, make a null set of free lists */
634 for (i = 0; i < ncpus; ++i) {
636 spin_init(&pcpu->spin, "bufinit");
637 for (j = 0; j < BUFFER_QUEUES; j++)
638 TAILQ_INIT(&pcpu->bufqueues[j]);
641 /* finally, initialize each buffer header and stick on empty q */
645 for (n = 0; n < nbuf; n++) {
647 bzero(bp, sizeof *bp);
648 bp->b_flags = B_INVAL; /* we're just an empty header */
649 bp->b_cmd = BUF_CMD_DONE;
650 bp->b_qindex = BQUEUE_EMPTY;
653 xio_init(&bp->b_xio);
655 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
663 * maxbufspace is the absolute maximum amount of buffer space we are
664 * allowed to reserve in KVM and in real terms. The absolute maximum
665 * is nominally used by buf_daemon. hibufspace is the nominal maximum
666 * used by most other processes. The differential is required to
667 * ensure that buf_daemon is able to run when other processes might
668 * be blocked waiting for buffer space.
670 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
671 * this may result in KVM fragmentation which is not handled optimally
674 maxbufspace = nbuf * BKVASIZE;
675 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
676 lobufspace = hibufspace - MAXBSIZE;
678 lorunningspace = 512 * 1024;
679 /* hirunningspace -- see below */
682 * Limit the amount of malloc memory since it is wired permanently
683 * into the kernel space. Even though this is accounted for in
684 * the buffer allocation, we don't want the malloced region to grow
685 * uncontrolled. The malloc scheme improves memory utilization
686 * significantly on average (small) directories.
688 maxbufmallocspace = hibufspace / 20;
691 * Reduce the chance of a deadlock occuring by limiting the number
692 * of delayed-write dirty buffers we allow to stack up.
694 * We don't want too much actually queued to the device at once
695 * (XXX this needs to be per-mount!), because the buffers will
696 * wind up locked for a very long period of time while the I/O
699 hidirtybufspace = hibufspace / 2; /* dirty + running */
700 hirunningspace = hibufspace / 16; /* locked & queued to device */
701 if (hirunningspace < 1024 * 1024)
702 hirunningspace = 1024 * 1024;
708 lodirtybufspace = hidirtybufspace / 2;
711 * Maximum number of async ops initiated per buf_daemon loop. This is
712 * somewhat of a hack at the moment, we really need to limit ourselves
713 * based on the number of bytes of I/O in-transit that were initiated
717 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
718 vm_object_hold(&kernel_object);
719 bogus_page = vm_page_alloc(&kernel_object,
720 (bogus_offset >> PAGE_SHIFT),
722 vm_object_drop(&kernel_object);
723 vmstats.v_wire_count++;
727 SYSINIT(do_bufinit, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, bufinit, NULL);
730 * Initialize the embedded bio structures, typically used by
731 * deprecated code which tries to allocate its own struct bufs.
734 initbufbio(struct buf *bp)
736 bp->b_bio1.bio_buf = bp;
737 bp->b_bio1.bio_prev = NULL;
738 bp->b_bio1.bio_offset = NOOFFSET;
739 bp->b_bio1.bio_next = &bp->b_bio2;
740 bp->b_bio1.bio_done = NULL;
741 bp->b_bio1.bio_flags = 0;
743 bp->b_bio2.bio_buf = bp;
744 bp->b_bio2.bio_prev = &bp->b_bio1;
745 bp->b_bio2.bio_offset = NOOFFSET;
746 bp->b_bio2.bio_next = NULL;
747 bp->b_bio2.bio_done = NULL;
748 bp->b_bio2.bio_flags = 0;
754 * Reinitialize the embedded bio structures as well as any additional
755 * translation cache layers.
758 reinitbufbio(struct buf *bp)
762 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
763 bio->bio_done = NULL;
764 bio->bio_offset = NOOFFSET;
769 * Undo the effects of an initbufbio().
772 uninitbufbio(struct buf *bp)
779 * Push another BIO layer onto an existing BIO and return it. The new
780 * BIO layer may already exist, holding cached translation data.
783 push_bio(struct bio *bio)
787 if ((nbio = bio->bio_next) == NULL) {
788 int index = bio - &bio->bio_buf->b_bio_array[0];
789 if (index >= NBUF_BIO - 1) {
790 panic("push_bio: too many layers bp %p",
793 nbio = &bio->bio_buf->b_bio_array[index + 1];
794 bio->bio_next = nbio;
795 nbio->bio_prev = bio;
796 nbio->bio_buf = bio->bio_buf;
797 nbio->bio_offset = NOOFFSET;
798 nbio->bio_done = NULL;
799 nbio->bio_next = NULL;
801 KKASSERT(nbio->bio_done == NULL);
806 * Pop a BIO translation layer, returning the previous layer. The
807 * must have been previously pushed.
810 pop_bio(struct bio *bio)
812 return(bio->bio_prev);
816 clearbiocache(struct bio *bio)
819 bio->bio_offset = NOOFFSET;
827 * Free the KVA allocation for buffer 'bp'.
829 * Must be called from a critical section as this is the only locking for
832 * Since this call frees up buffer space, we call bufspacewakeup().
835 bfreekva(struct buf *bp)
841 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
842 vm_map_lock(&buffer_map);
843 bufspace -= bp->b_kvasize;
844 vm_map_delete(&buffer_map,
845 (vm_offset_t) bp->b_kvabase,
846 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
849 vm_map_unlock(&buffer_map);
850 vm_map_entry_release(count);
852 bp->b_kvabase = NULL;
858 * Remove the buffer from the appropriate free list.
859 * (caller must be locked)
862 _bremfree(struct buf *bp)
864 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu];
866 if (bp->b_qindex != BQUEUE_NONE) {
867 KASSERT(BUF_REFCNTNB(bp) == 1,
868 ("bremfree: bp %p not locked",bp));
869 TAILQ_REMOVE(&pcpu->bufqueues[bp->b_qindex], bp, b_freelist);
870 bp->b_qindex = BQUEUE_NONE;
872 if (BUF_REFCNTNB(bp) <= 1)
873 panic("bremfree: removing a buffer not on a queue");
878 * bremfree() - must be called with a locked buffer
881 bremfree(struct buf *bp)
883 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu];
885 spin_lock(&pcpu->spin);
887 spin_unlock(&pcpu->spin);
891 * bremfree_locked - must be called with pcpu->spin locked
894 bremfree_locked(struct buf *bp)
900 * This version of bread issues any required I/O asyncnronously and
901 * makes a callback on completion.
903 * The callback must check whether BIO_DONE is set in the bio and issue
904 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing
905 * BIO_DONE and disposing of the I/O (bqrelse()ing it).
908 breadcb(struct vnode *vp, off_t loffset, int size,
909 void (*func)(struct bio *), void *arg)
913 bp = getblk(vp, loffset, size, 0, 0);
915 /* if not found in cache, do some I/O */
916 if ((bp->b_flags & B_CACHE) == 0) {
917 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
918 bp->b_cmd = BUF_CMD_READ;
919 bp->b_bio1.bio_done = func;
920 bp->b_bio1.bio_caller_info1.ptr = arg;
921 vfs_busy_pages(vp, bp);
923 vn_strategy(vp, &bp->b_bio1);
926 * Since we are issuing the callback synchronously it cannot
927 * race the BIO_DONE, so no need for atomic ops here.
929 /*bp->b_bio1.bio_done = func;*/
930 bp->b_bio1.bio_caller_info1.ptr = arg;
931 bp->b_bio1.bio_flags |= BIO_DONE;
939 * breadnx() - Terminal function for bread() and breadn().
941 * This function will start asynchronous I/O on read-ahead blocks as well
942 * as satisfy the primary request.
944 * We must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE is
945 * set, the buffer is valid and we do not have to do anything.
948 breadnx(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
949 int *rabsize, int cnt, struct buf **bpp)
951 struct buf *bp, *rabp;
953 int rv = 0, readwait = 0;
958 *bpp = bp = getblk(vp, loffset, size, 0, 0);
960 /* if not found in cache, do some I/O */
961 if ((bp->b_flags & B_CACHE) == 0) {
962 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
963 bp->b_cmd = BUF_CMD_READ;
964 bp->b_bio1.bio_done = biodone_sync;
965 bp->b_bio1.bio_flags |= BIO_SYNC;
966 vfs_busy_pages(vp, bp);
967 vn_strategy(vp, &bp->b_bio1);
971 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
972 if (inmem(vp, *raoffset))
974 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
976 if ((rabp->b_flags & B_CACHE) == 0) {
977 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
978 rabp->b_cmd = BUF_CMD_READ;
979 vfs_busy_pages(vp, rabp);
981 vn_strategy(vp, &rabp->b_bio1);
987 rv = biowait(&bp->b_bio1, "biord");
994 * Synchronous write, waits for completion.
996 * Write, release buffer on completion. (Done by iodone
997 * if async). Do not bother writing anything if the buffer
1000 * Note that we set B_CACHE here, indicating that buffer is
1001 * fully valid and thus cacheable. This is true even of NFS
1002 * now so we set it generally. This could be set either here
1003 * or in biodone() since the I/O is synchronous. We put it
1007 bwrite(struct buf *bp)
1011 if (bp->b_flags & B_INVAL) {
1015 if (BUF_REFCNTNB(bp) == 0)
1016 panic("bwrite: buffer is not busy???");
1019 * NOTE: We no longer mark the buffer clear prior to the vn_strategy()
1020 * call because it will remove the buffer from the vnode's
1021 * dirty buffer list prematurely and possibly cause filesystem
1022 * checks to race buffer flushes. This is now handled in
1025 * bundirty(bp); REMOVED
1028 bp->b_flags &= ~(B_ERROR | B_EINTR);
1029 bp->b_flags |= B_CACHE;
1030 bp->b_cmd = BUF_CMD_WRITE;
1031 bp->b_bio1.bio_done = biodone_sync;
1032 bp->b_bio1.bio_flags |= BIO_SYNC;
1033 vfs_busy_pages(bp->b_vp, bp);
1036 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1037 * valid for vnode-backed buffers.
1039 bsetrunningbufspace(bp, bp->b_bufsize);
1040 vn_strategy(bp->b_vp, &bp->b_bio1);
1041 error = biowait(&bp->b_bio1, "biows");
1050 * Asynchronous write. Start output on a buffer, but do not wait for
1051 * it to complete. The buffer is released when the output completes.
1053 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1054 * B_INVAL buffers. Not us.
1057 bawrite(struct buf *bp)
1059 if (bp->b_flags & B_INVAL) {
1063 if (BUF_REFCNTNB(bp) == 0)
1064 panic("bwrite: buffer is not busy???");
1067 * NOTE: We no longer mark the buffer clear prior to the vn_strategy()
1068 * call because it will remove the buffer from the vnode's
1069 * dirty buffer list prematurely and possibly cause filesystem
1070 * checks to race buffer flushes. This is now handled in
1073 * bundirty(bp); REMOVED
1075 bp->b_flags &= ~(B_ERROR | B_EINTR);
1076 bp->b_flags |= B_CACHE;
1077 bp->b_cmd = BUF_CMD_WRITE;
1078 KKASSERT(bp->b_bio1.bio_done == NULL);
1079 vfs_busy_pages(bp->b_vp, bp);
1082 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1083 * valid for vnode-backed buffers.
1085 bsetrunningbufspace(bp, bp->b_bufsize);
1087 vn_strategy(bp->b_vp, &bp->b_bio1);
1093 * Ordered write. Start output on a buffer, and flag it so that the
1094 * device will write it in the order it was queued. The buffer is
1095 * released when the output completes. bwrite() ( or the VOP routine
1096 * anyway ) is responsible for handling B_INVAL buffers.
1099 bowrite(struct buf *bp)
1101 bp->b_flags |= B_ORDERED;
1109 * Delayed write. (Buffer is marked dirty). Do not bother writing
1110 * anything if the buffer is marked invalid.
1112 * Note that since the buffer must be completely valid, we can safely
1113 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1114 * biodone() in order to prevent getblk from writing the buffer
1115 * out synchronously.
1118 bdwrite(struct buf *bp)
1120 if (BUF_REFCNTNB(bp) == 0)
1121 panic("bdwrite: buffer is not busy");
1123 if (bp->b_flags & B_INVAL) {
1129 if (dsched_is_clear_buf_priv(bp))
1133 * Set B_CACHE, indicating that the buffer is fully valid. This is
1134 * true even of NFS now.
1136 bp->b_flags |= B_CACHE;
1139 * This bmap keeps the system from needing to do the bmap later,
1140 * perhaps when the system is attempting to do a sync. Since it
1141 * is likely that the indirect block -- or whatever other datastructure
1142 * that the filesystem needs is still in memory now, it is a good
1143 * thing to do this. Note also, that if the pageout daemon is
1144 * requesting a sync -- there might not be enough memory to do
1145 * the bmap then... So, this is important to do.
1147 if (bp->b_bio2.bio_offset == NOOFFSET) {
1148 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1149 NULL, NULL, BUF_CMD_WRITE);
1153 * Because the underlying pages may still be mapped and
1154 * writable trying to set the dirty buffer (b_dirtyoff/end)
1155 * range here will be inaccurate.
1157 * However, we must still clean the pages to satisfy the
1158 * vnode_pager and pageout daemon, so theythink the pages
1159 * have been "cleaned". What has really occured is that
1160 * they've been earmarked for later writing by the buffer
1163 * So we get the b_dirtyoff/end update but will not actually
1164 * depend on it (NFS that is) until the pages are busied for
1167 vfs_clean_pages(bp);
1171 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1172 * due to the softdep code.
1177 * Fake write - return pages to VM system as dirty, leave the buffer clean.
1178 * This is used by tmpfs.
1180 * It is important for any VFS using this routine to NOT use it for
1181 * IO_SYNC or IO_ASYNC operations which occur when the system really
1182 * wants to flush VM pages to backing store.
1185 buwrite(struct buf *bp)
1191 * Only works for VMIO buffers. If the buffer is already
1192 * marked for delayed-write we can't avoid the bdwrite().
1194 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) {
1200 * Mark as needing a commit.
1202 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1203 m = bp->b_xio.xio_pages[i];
1204 vm_page_need_commit(m);
1212 * Turn buffer into delayed write request by marking it B_DELWRI.
1213 * B_RELBUF and B_NOCACHE must be cleared.
1215 * We reassign the buffer to itself to properly update it in the
1216 * dirty/clean lists.
1218 * Must be called from a critical section.
1219 * The buffer must be on BQUEUE_NONE.
1222 bdirty(struct buf *bp)
1224 KASSERT(bp->b_qindex == BQUEUE_NONE,
1225 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1226 if (bp->b_flags & B_NOCACHE) {
1227 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1228 bp->b_flags &= ~B_NOCACHE;
1230 if (bp->b_flags & B_INVAL) {
1231 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1233 bp->b_flags &= ~B_RELBUF;
1235 if ((bp->b_flags & B_DELWRI) == 0) {
1236 lwkt_gettoken(&bp->b_vp->v_token);
1237 bp->b_flags |= B_DELWRI;
1239 lwkt_reltoken(&bp->b_vp->v_token);
1241 atomic_add_long(&dirtybufcount, 1);
1242 atomic_add_long(&dirtykvaspace, bp->b_kvasize);
1243 atomic_add_long(&dirtybufspace, bp->b_bufsize);
1244 if (bp->b_flags & B_HEAVY) {
1245 atomic_add_long(&dirtybufcounthw, 1);
1246 atomic_add_long(&dirtybufspacehw, bp->b_bufsize);
1253 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1254 * needs to be flushed with a different buf_daemon thread to avoid
1255 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1258 bheavy(struct buf *bp)
1260 if ((bp->b_flags & B_HEAVY) == 0) {
1261 bp->b_flags |= B_HEAVY;
1262 if (bp->b_flags & B_DELWRI) {
1263 atomic_add_long(&dirtybufcounthw, 1);
1264 atomic_add_long(&dirtybufspacehw, bp->b_bufsize);
1272 * Clear B_DELWRI for buffer.
1274 * Must be called from a critical section.
1276 * The buffer is typically on BQUEUE_NONE but there is one case in
1277 * brelse() that calls this function after placing the buffer on
1278 * a different queue.
1281 bundirty(struct buf *bp)
1283 if (bp->b_flags & B_DELWRI) {
1284 lwkt_gettoken(&bp->b_vp->v_token);
1285 bp->b_flags &= ~B_DELWRI;
1287 lwkt_reltoken(&bp->b_vp->v_token);
1289 atomic_add_long(&dirtybufcount, -1);
1290 atomic_add_long(&dirtykvaspace, -bp->b_kvasize);
1291 atomic_add_long(&dirtybufspace, -bp->b_bufsize);
1292 if (bp->b_flags & B_HEAVY) {
1293 atomic_add_long(&dirtybufcounthw, -1);
1294 atomic_add_long(&dirtybufspacehw, -bp->b_bufsize);
1296 bd_signal(bp->b_bufsize);
1299 * Since it is now being written, we can clear its deferred write flag.
1301 bp->b_flags &= ~B_DEFERRED;
1305 * Set the b_runningbufspace field, used to track how much I/O is
1306 * in progress at any given moment.
1309 bsetrunningbufspace(struct buf *bp, int bytes)
1311 bp->b_runningbufspace = bytes;
1313 atomic_add_long(&runningbufspace, bytes);
1314 atomic_add_long(&runningbufcount, 1);
1321 * Release a busy buffer and, if requested, free its resources. The
1322 * buffer will be stashed in the appropriate bufqueue[] allowing it
1323 * to be accessed later as a cache entity or reused for other purposes.
1326 brelse(struct buf *bp)
1328 struct bufpcpu *pcpu;
1330 int saved_flags = bp->b_flags;
1333 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1334 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1337 * If B_NOCACHE is set we are being asked to destroy the buffer and
1338 * its backing store. Clear B_DELWRI.
1340 * B_NOCACHE is set in two cases: (1) when the caller really wants
1341 * to destroy the buffer and backing store and (2) when the caller
1342 * wants to destroy the buffer and backing store after a write
1345 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1349 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1351 * A re-dirtied buffer is only subject to destruction
1352 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1354 /* leave buffer intact */
1355 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1356 (bp->b_bufsize <= 0)) {
1358 * Either a failed read or we were asked to free or not
1359 * cache the buffer. This path is reached with B_DELWRI
1360 * set only if B_INVAL is already set. B_NOCACHE governs
1361 * backing store destruction.
1363 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1364 * buffer cannot be immediately freed.
1366 bp->b_flags |= B_INVAL;
1367 if (LIST_FIRST(&bp->b_dep) != NULL)
1369 if (bp->b_flags & B_DELWRI) {
1370 atomic_add_long(&dirtybufcount, -1);
1371 atomic_add_long(&dirtykvaspace, -bp->b_kvasize);
1372 atomic_add_long(&dirtybufspace, -bp->b_bufsize);
1373 if (bp->b_flags & B_HEAVY) {
1374 atomic_add_long(&dirtybufcounthw, -1);
1375 atomic_add_long(&dirtybufspacehw,
1378 bd_signal(bp->b_bufsize);
1380 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1384 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set,
1385 * or if b_refs is non-zero.
1387 * If vfs_vmio_release() is called with either bit set, the
1388 * underlying pages may wind up getting freed causing a previous
1389 * write (bdwrite()) to get 'lost' because pages associated with
1390 * a B_DELWRI bp are marked clean. Pages associated with a
1391 * B_LOCKED buffer may be mapped by the filesystem.
1393 * If we want to release the buffer ourselves (rather then the
1394 * originator asking us to release it), give the originator a
1395 * chance to countermand the release by setting B_LOCKED.
1397 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1398 * if B_DELWRI is set.
1400 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1401 * on pages to return pages to the VM page queues.
1403 if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) {
1404 bp->b_flags &= ~B_RELBUF;
1405 } else if (vm_page_count_min(0)) {
1406 if (LIST_FIRST(&bp->b_dep) != NULL)
1407 buf_deallocate(bp); /* can set B_LOCKED */
1408 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1409 bp->b_flags &= ~B_RELBUF;
1411 bp->b_flags |= B_RELBUF;
1415 * Make sure b_cmd is clear. It may have already been cleared by
1418 * At this point destroying the buffer is governed by the B_INVAL
1419 * or B_RELBUF flags.
1421 bp->b_cmd = BUF_CMD_DONE;
1422 dsched_exit_buf(bp);
1425 * VMIO buffer rundown. Make sure the VM page array is restored
1426 * after an I/O may have replaces some of the pages with bogus pages
1427 * in order to not destroy dirty pages in a fill-in read.
1429 * Note that due to the code above, if a buffer is marked B_DELWRI
1430 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1431 * B_INVAL may still be set, however.
1433 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1434 * but not the backing store. B_NOCACHE will destroy the backing
1437 * Note that dirty NFS buffers contain byte-granular write ranges
1438 * and should not be destroyed w/ B_INVAL even if the backing store
1441 if (bp->b_flags & B_VMIO) {
1443 * Rundown for VMIO buffers which are not dirty NFS buffers.
1455 * Get the base offset and length of the buffer. Note that
1456 * in the VMIO case if the buffer block size is not
1457 * page-aligned then b_data pointer may not be page-aligned.
1458 * But our b_xio.xio_pages array *IS* page aligned.
1460 * block sizes less then DEV_BSIZE (usually 512) are not
1461 * supported due to the page granularity bits (m->valid,
1462 * m->dirty, etc...).
1464 * See man buf(9) for more information
1467 resid = bp->b_bufsize;
1468 foff = bp->b_loffset;
1470 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1471 m = bp->b_xio.xio_pages[i];
1472 vm_page_flag_clear(m, PG_ZERO);
1474 * If we hit a bogus page, fixup *all* of them
1475 * now. Note that we left these pages wired
1476 * when we removed them so they had better exist,
1477 * and they cannot be ripped out from under us so
1478 * no critical section protection is necessary.
1480 if (m == bogus_page) {
1482 poff = OFF_TO_IDX(bp->b_loffset);
1484 vm_object_hold(obj);
1485 for (j = i; j < bp->b_xio.xio_npages; j++) {
1488 mtmp = bp->b_xio.xio_pages[j];
1489 if (mtmp == bogus_page) {
1490 mtmp = vm_page_lookup(obj, poff + j);
1492 panic("brelse: page missing");
1494 bp->b_xio.xio_pages[j] = mtmp;
1497 bp->b_flags &= ~B_HASBOGUS;
1498 vm_object_drop(obj);
1500 if ((bp->b_flags & B_INVAL) == 0) {
1501 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1502 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1504 m = bp->b_xio.xio_pages[i];
1508 * Invalidate the backing store if B_NOCACHE is set
1509 * (e.g. used with vinvalbuf()). If this is NFS
1510 * we impose a requirement that the block size be
1511 * a multiple of PAGE_SIZE and create a temporary
1512 * hack to basically invalidate the whole page. The
1513 * problem is that NFS uses really odd buffer sizes
1514 * especially when tracking piecemeal writes and
1515 * it also vinvalbuf()'s a lot, which would result
1516 * in only partial page validation and invalidation
1517 * here. If the file page is mmap()'d, however,
1518 * all the valid bits get set so after we invalidate
1519 * here we would end up with weird m->valid values
1520 * like 0xfc. nfs_getpages() can't handle this so
1521 * we clear all the valid bits for the NFS case
1522 * instead of just some of them.
1524 * The real bug is the VM system having to set m->valid
1525 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1526 * itself is an artifact of the whole 512-byte
1527 * granular mess that exists to support odd block
1528 * sizes and UFS meta-data block sizes (e.g. 6144).
1529 * A complete rewrite is required.
1533 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1534 int poffset = foff & PAGE_MASK;
1537 presid = PAGE_SIZE - poffset;
1538 if (bp->b_vp->v_tag == VT_NFS &&
1539 bp->b_vp->v_type == VREG) {
1541 } else if (presid > resid) {
1544 KASSERT(presid >= 0, ("brelse: extra page"));
1545 vm_page_set_invalid(m, poffset, presid);
1548 * Also make sure any swap cache is removed
1549 * as it is now stale (HAMMER in particular
1550 * uses B_NOCACHE to deal with buffer
1553 swap_pager_unswapped(m);
1555 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1556 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1558 if (bp->b_flags & (B_INVAL | B_RELBUF))
1559 vfs_vmio_release(bp);
1562 * Rundown for non-VMIO buffers.
1564 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1567 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1573 if (bp->b_qindex != BQUEUE_NONE)
1574 panic("brelse: free buffer onto another queue???");
1575 if (BUF_REFCNTNB(bp) > 1) {
1576 /* Temporary panic to verify exclusive locking */
1577 /* This panic goes away when we allow shared refs */
1578 panic("brelse: multiple refs");
1584 * Figure out the correct queue to place the cleaned up buffer on.
1585 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1586 * disassociated from their vnode.
1588 * Return the buffer to its original pcpu area
1590 pcpu = &bufpcpu[bp->b_qcpu];
1591 spin_lock(&pcpu->spin);
1593 if (bp->b_flags & B_LOCKED) {
1595 * Buffers that are locked are placed in the locked queue
1596 * immediately, regardless of their state.
1598 bp->b_qindex = BQUEUE_LOCKED;
1599 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1601 } else if (bp->b_bufsize == 0) {
1603 * Buffers with no memory. Due to conditionals near the top
1604 * of brelse() such buffers should probably already be
1605 * marked B_INVAL and disassociated from their vnode.
1607 bp->b_flags |= B_INVAL;
1608 KASSERT(bp->b_vp == NULL,
1609 ("bp1 %p flags %08x/%08x vnode %p "
1610 "unexpectededly still associated!",
1611 bp, saved_flags, bp->b_flags, bp->b_vp));
1612 KKASSERT((bp->b_flags & B_HASHED) == 0);
1613 if (bp->b_kvasize) {
1614 bp->b_qindex = BQUEUE_EMPTYKVA;
1616 bp->b_qindex = BQUEUE_EMPTY;
1618 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex],
1620 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1622 * Buffers with junk contents. Again these buffers had better
1623 * already be disassociated from their vnode.
1625 KASSERT(bp->b_vp == NULL,
1626 ("bp2 %p flags %08x/%08x vnode %p unexpectededly "
1627 "still associated!",
1628 bp, saved_flags, bp->b_flags, bp->b_vp));
1629 KKASSERT((bp->b_flags & B_HASHED) == 0);
1630 bp->b_flags |= B_INVAL;
1631 bp->b_qindex = BQUEUE_CLEAN;
1632 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex],
1636 * Remaining buffers. These buffers are still associated with
1639 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1641 bp->b_qindex = BQUEUE_DIRTY;
1642 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1645 case B_DELWRI | B_HEAVY:
1646 bp->b_qindex = BQUEUE_DIRTY_HW;
1647 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1652 * NOTE: Buffers are always placed at the end of the
1653 * queue. If B_AGE is not set the buffer will cycle
1654 * through the queue twice.
1656 bp->b_qindex = BQUEUE_CLEAN;
1657 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1662 spin_unlock(&pcpu->spin);
1665 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1666 * on the correct queue but we have not yet unlocked it.
1668 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1672 * The bp is on an appropriate queue unless locked. If it is not
1673 * locked or dirty we can wakeup threads waiting for buffer space.
1675 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1676 * if B_INVAL is set ).
1678 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1682 * Something we can maybe free or reuse
1684 if (bp->b_bufsize || bp->b_kvasize)
1688 * Clean up temporary flags and unlock the buffer.
1690 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
1697 * Release a buffer back to the appropriate queue but do not try to free
1698 * it. The buffer is expected to be used again soon.
1700 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1701 * biodone() to requeue an async I/O on completion. It is also used when
1702 * known good buffers need to be requeued but we think we may need the data
1705 * XXX we should be able to leave the B_RELBUF hint set on completion.
1708 bqrelse(struct buf *bp)
1710 struct bufpcpu *pcpu;
1712 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1713 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1715 if (bp->b_qindex != BQUEUE_NONE)
1716 panic("bqrelse: free buffer onto another queue???");
1717 if (BUF_REFCNTNB(bp) > 1) {
1718 /* do not release to free list */
1719 panic("bqrelse: multiple refs");
1723 buf_act_advance(bp);
1725 pcpu = &bufpcpu[bp->b_qcpu];
1726 spin_lock(&pcpu->spin);
1728 if (bp->b_flags & B_LOCKED) {
1730 * Locked buffers are released to the locked queue. However,
1731 * if the buffer is dirty it will first go into the dirty
1732 * queue and later on after the I/O completes successfully it
1733 * will be released to the locked queue.
1735 bp->b_qindex = BQUEUE_LOCKED;
1736 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1738 } else if (bp->b_flags & B_DELWRI) {
1739 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1740 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1741 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1743 } else if (vm_page_count_min(0)) {
1745 * We are too low on memory, we have to try to free the
1746 * buffer (most importantly: the wired pages making up its
1747 * backing store) *now*.
1749 spin_unlock(&pcpu->spin);
1753 bp->b_qindex = BQUEUE_CLEAN;
1754 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1757 spin_unlock(&pcpu->spin);
1760 * We have now placed the buffer on the proper queue, but have yet
1763 if ((bp->b_flags & B_LOCKED) == 0 &&
1764 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1769 * Something we can maybe free or reuse.
1771 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1775 * Final cleanup and unlock. Clear bits that are only used while a
1776 * buffer is actively locked.
1778 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
1779 dsched_exit_buf(bp);
1784 * Hold a buffer, preventing it from being reused. This will prevent
1785 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL
1786 * operations. If a B_INVAL operation occurs the buffer will remain held
1787 * but the underlying pages may get ripped out.
1789 * These functions are typically used in VOP_READ/VOP_WRITE functions
1790 * to hold a buffer during a copyin or copyout, preventing deadlocks
1791 * or recursive lock panics when read()/write() is used over mmap()'d
1794 * NOTE: bqhold() requires that the buffer be locked at the time of the
1795 * hold. bqdrop() has no requirements other than the buffer having
1796 * previously been held.
1799 bqhold(struct buf *bp)
1801 atomic_add_int(&bp->b_refs, 1);
1805 bqdrop(struct buf *bp)
1807 KKASSERT(bp->b_refs > 0);
1808 atomic_add_int(&bp->b_refs, -1);
1812 * Return backing pages held by the buffer 'bp' back to the VM system.
1813 * This routine is called when the bp is invalidated, released, or
1816 * The KVA mapping (b_data) for the underlying pages is removed by
1819 * WARNING! This routine is integral to the low memory critical path
1820 * when a buffer is B_RELBUF'd. If the system has a severe page
1821 * deficit we need to get the page(s) onto the PQ_FREE or PQ_CACHE
1822 * queues so they can be reused in the current pageout daemon
1826 vfs_vmio_release(struct buf *bp)
1831 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1832 m = bp->b_xio.xio_pages[i];
1833 bp->b_xio.xio_pages[i] = NULL;
1836 * We need to own the page in order to safely unwire it.
1838 vm_page_busy_wait(m, FALSE, "vmiopg");
1841 * The VFS is telling us this is not a meta-data buffer
1842 * even if it is backed by a block device.
1844 if (bp->b_flags & B_NOTMETA)
1845 vm_page_flag_set(m, PG_NOTMETA);
1848 * This is a very important bit of code. We try to track
1849 * VM page use whether the pages are wired into the buffer
1850 * cache or not. While wired into the buffer cache the
1851 * bp tracks the act_count.
1853 * We can choose to place unwired pages on the inactive
1854 * queue (0) or active queue (1). If we place too many
1855 * on the active queue the queue will cycle the act_count
1856 * on pages we'd like to keep, just from single-use pages
1857 * (such as when doing a tar-up or file scan).
1859 if (bp->b_act_count < vm_cycle_point)
1860 vm_page_unwire(m, 0);
1862 vm_page_unwire(m, 1);
1865 * If the wire_count has dropped to 0 we may need to take
1866 * further action before unbusying the page.
1868 * WARNING: vm_page_try_*() also checks PG_NEED_COMMIT for us.
1870 if (m->wire_count == 0) {
1871 vm_page_flag_clear(m, PG_ZERO);
1873 if (bp->b_flags & B_DIRECT) {
1875 * Attempt to free the page if B_DIRECT is
1876 * set, the caller does not desire the page
1880 vm_page_try_to_free(m);
1881 } else if ((bp->b_flags & B_NOTMETA) ||
1882 vm_page_count_min(0)) {
1884 * Attempt to move the page to PQ_CACHE
1885 * if B_NOTMETA is set. This flag is set
1886 * by HAMMER to remove one of the two pages
1887 * present when double buffering is enabled.
1889 * Attempt to move the page to PQ_CACHE
1890 * If we have a severe page deficit. This
1891 * will cause buffer cache operations related
1892 * to pageouts to recycle the related pages
1893 * in order to avoid a low memory deadlock.
1895 m->act_count = bp->b_act_count;
1897 vm_page_try_to_cache(m);
1900 * Nominal case, leave the page on the
1901 * queue the original unwiring placed it on
1902 * (active or inactive).
1904 m->act_count = bp->b_act_count;
1912 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
1913 bp->b_xio.xio_npages);
1914 if (bp->b_bufsize) {
1918 bp->b_xio.xio_npages = 0;
1919 bp->b_flags &= ~B_VMIO;
1920 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1926 * Find and initialize a new buffer header, freeing up existing buffers
1927 * in the bufqueues as necessary. The new buffer is returned locked.
1929 * Important: B_INVAL is not set. If the caller wishes to throw the
1930 * buffer away, the caller must set B_INVAL prior to calling brelse().
1933 * We have insufficient buffer headers
1934 * We have insufficient buffer space
1935 * buffer_map is too fragmented ( space reservation fails )
1936 * If we have to flush dirty buffers ( but we try to avoid this )
1938 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1939 * Instead we ask the buf daemon to do it for us. We attempt to
1940 * avoid piecemeal wakeups of the pageout daemon.
1943 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1945 struct bufpcpu *pcpu;
1951 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1952 int maxloops = 200000;
1953 int restart_reason = 0;
1954 struct buf *restart_bp = NULL;
1955 static int flushingbufs;
1958 * We can't afford to block since we might be holding a vnode lock,
1959 * which may prevent system daemons from running. We deal with
1960 * low-memory situations by proactively returning memory and running
1961 * async I/O rather then sync I/O.
1965 --getnewbufrestarts;
1966 nqcpu = mycpu->gd_cpuid;
1968 ++getnewbufrestarts;
1970 if (debug_bufbio && --maxloops == 0)
1971 panic("getnewbuf, excessive loops on cpu %d restart %d (%p)",
1972 mycpu->gd_cpuid, restart_reason, restart_bp);
1975 * Setup for scan. If we do not have enough free buffers,
1976 * we setup a degenerate case that immediately fails. Note
1977 * that if we are specially marked process, we are allowed to
1978 * dip into our reserves.
1980 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1982 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1983 * However, there are a number of cases (defragging, reusing, ...)
1984 * where we cannot backup.
1986 pcpu = &bufpcpu[nqcpu];
1987 nqindex = BQUEUE_EMPTYKVA;
1988 spin_lock(&pcpu->spin);
1990 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTYKVA]);
1994 * If no EMPTYKVA buffers and we are either
1995 * defragging or reusing, locate a CLEAN buffer
1996 * to free or reuse. If bufspace useage is low
1997 * skip this step so we can allocate a new buffer.
1999 if (defrag || bufspace >= lobufspace) {
2000 nqindex = BQUEUE_CLEAN;
2001 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]);
2005 * If we could not find or were not allowed to reuse a
2006 * CLEAN buffer, check to see if it is ok to use an EMPTY
2007 * buffer. We can only use an EMPTY buffer if allocating
2008 * its KVA would not otherwise run us out of buffer space.
2010 if (nbp == NULL && defrag == 0 &&
2011 bufspace + maxsize < hibufspace) {
2012 nqindex = BQUEUE_EMPTY;
2013 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTY]);
2018 * Run scan, possibly freeing data and/or kva mappings on the fly
2021 * WARNING! spin is held!
2023 while ((bp = nbp) != NULL) {
2024 int qindex = nqindex;
2026 nbp = TAILQ_NEXT(bp, b_freelist);
2029 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2030 * cycles through the queue twice before being selected.
2032 if (qindex == BQUEUE_CLEAN &&
2033 (bp->b_flags & B_AGE) == 0 && nbp) {
2034 bp->b_flags |= B_AGE;
2035 TAILQ_REMOVE(&pcpu->bufqueues[qindex],
2037 TAILQ_INSERT_TAIL(&pcpu->bufqueues[qindex],
2043 * Calculate next bp ( we can only use it if we do not block
2044 * or do other fancy things ).
2049 nqindex = BQUEUE_EMPTYKVA;
2050 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTYKVA])))
2053 case BQUEUE_EMPTYKVA:
2054 nqindex = BQUEUE_CLEAN;
2055 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN])))
2069 KASSERT(bp->b_qindex == qindex,
2070 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2073 * Note: we no longer distinguish between VMIO and non-VMIO
2076 KASSERT((bp->b_flags & B_DELWRI) == 0,
2077 ("delwri buffer %p found in queue %d", bp, qindex));
2080 * Do not try to reuse a buffer with a non-zero b_refs.
2081 * This is an unsynchronized test. A synchronized test
2082 * is also performed after we lock the buffer.
2088 * If we are defragging then we need a buffer with
2089 * b_kvasize != 0. XXX this situation should no longer
2090 * occur, if defrag is non-zero the buffer's b_kvasize
2091 * should also be non-zero at this point. XXX
2093 if (defrag && bp->b_kvasize == 0) {
2094 kprintf("Warning: defrag empty buffer %p\n", bp);
2099 * Start freeing the bp. This is somewhat involved. nbp
2100 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
2101 * on the clean list must be disassociated from their
2102 * current vnode. Buffers on the empty[kva] lists have
2103 * already been disassociated.
2105 * b_refs is checked after locking along with queue changes.
2106 * We must check here to deal with zero->nonzero transitions
2107 * made by the owner of the buffer lock, which is used by
2108 * VFS's to hold the buffer while issuing an unlocked
2109 * uiomove()s. We cannot invalidate the buffer's pages
2110 * for this case. Once we successfully lock a buffer the
2111 * only 0->1 transitions of b_refs will occur via findblk().
2113 * We must also check for queue changes after successful
2114 * locking as the current lock holder may dispose of the
2115 * buffer and change its queue.
2117 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2118 spin_unlock(&pcpu->spin);
2119 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100);
2124 if (bp->b_qindex != qindex || bp->b_refs) {
2125 spin_unlock(&pcpu->spin);
2131 bremfree_locked(bp);
2132 spin_unlock(&pcpu->spin);
2135 * Dependancies must be handled before we disassociate the
2138 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2139 * be immediately disassociated. HAMMER then becomes
2140 * responsible for releasing the buffer.
2142 * NOTE: spin is UNLOCKED now.
2144 if (LIST_FIRST(&bp->b_dep) != NULL) {
2146 if (bp->b_flags & B_LOCKED) {
2152 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2155 if (qindex == BQUEUE_CLEAN) {
2156 if (bp->b_flags & B_VMIO)
2157 vfs_vmio_release(bp);
2163 * NOTE: nbp is now entirely invalid. We can only restart
2164 * the scan from this point on.
2166 * Get the rest of the buffer freed up. b_kva* is still
2167 * valid after this operation.
2169 KASSERT(bp->b_vp == NULL,
2170 ("bp3 %p flags %08x vnode %p qindex %d "
2171 "unexpectededly still associated!",
2172 bp, bp->b_flags, bp->b_vp, qindex));
2173 KKASSERT((bp->b_flags & B_HASHED) == 0);
2176 * critical section protection is not required when
2177 * scrapping a buffer's contents because it is already
2183 if (bp->b_flags & (B_VNDIRTY | B_VNCLEAN | B_HASHED)) {
2184 kprintf("getnewbuf: caught bug vp queue "
2185 "%p/%08x qidx %d\n",
2186 bp, bp->b_flags, qindex);
2189 bp->b_flags = B_BNOCLIP;
2190 bp->b_cmd = BUF_CMD_DONE;
2195 bp->b_xio.xio_npages = 0;
2196 bp->b_dirtyoff = bp->b_dirtyend = 0;
2197 bp->b_act_count = ACT_INIT;
2199 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2201 if (blkflags & GETBLK_BHEAVY)
2202 bp->b_flags |= B_HEAVY;
2205 * If we are defragging then free the buffer.
2208 bp->b_flags |= B_INVAL;
2218 * If we are overcomitted then recover the buffer and its
2219 * KVM space. This occurs in rare situations when multiple
2220 * processes are blocked in getnewbuf() or allocbuf().
2222 * On 64-bit systems BKVASIZE == MAXBSIZE and overcommit
2223 * should not be possible.
2225 if (bufspace >= hibufspace)
2227 if (BKVASIZE != MAXBSIZE) {
2228 if (flushingbufs && bp->b_kvasize != 0) {
2229 bp->b_flags |= B_INVAL;
2237 if (bufspace < lobufspace)
2241 * b_refs can transition to a non-zero value while we hold
2242 * the buffer locked due to a findblk(). Our brelvp() above
2243 * interlocked any future possible transitions due to
2246 * If we find b_refs to be non-zero we can destroy the
2247 * buffer's contents but we cannot yet reuse the buffer.
2250 bp->b_flags |= B_INVAL;
2251 if (BKVASIZE != MAXBSIZE)
2259 /* NOT REACHED, spin not held */
2263 * If we exhausted our list, iterate other cpus. If that fails,
2264 * sleep as appropriate. We may have to wakeup various daemons
2265 * and write out some dirty buffers.
2267 * Generally we are sleeping due to insufficient buffer space.
2269 * NOTE: spin is held if bp is NULL, else it is not held.
2275 spin_unlock(&pcpu->spin);
2277 nqcpu = (nqcpu + 1) % ncpus;
2278 if (nqcpu != mycpu->gd_cpuid) {
2285 flags = VFS_BIO_NEED_BUFSPACE;
2287 } else if (bufspace >= hibufspace) {
2289 flags = VFS_BIO_NEED_BUFSPACE;
2292 flags = VFS_BIO_NEED_ANY;
2295 bd_speedup(); /* heeeelp */
2296 atomic_set_int(&needsbuffer, flags);
2297 while (needsbuffer & flags) {
2300 tsleep_interlock(&needsbuffer, 0);
2301 value = atomic_fetchadd_int(&needsbuffer, 0);
2302 if (value & flags) {
2303 if (tsleep(&needsbuffer, PINTERLOCKED|slpflags,
2304 waitmsg, slptimeo)) {
2311 * We finally have a valid bp. We aren't quite out of the
2312 * woods, we still have to reserve kva space. In order
2313 * to keep fragmentation sane we only allocate kva in
2316 * (spin is not held)
2318 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2320 if (maxsize != bp->b_kvasize) {
2321 vm_offset_t addr = 0;
2326 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2327 vm_map_lock(&buffer_map);
2329 if (vm_map_findspace(&buffer_map,
2330 vm_map_min(&buffer_map), maxsize,
2331 maxsize, 0, &addr)) {
2333 * Uh oh. Buffer map is too fragmented. We
2334 * must defragment the map.
2336 vm_map_unlock(&buffer_map);
2337 vm_map_entry_release(count);
2340 bp->b_flags |= B_INVAL;
2347 vm_map_insert(&buffer_map, &count,
2349 0, addr, addr + maxsize,
2351 VM_PROT_ALL, VM_PROT_ALL,
2354 bp->b_kvabase = (caddr_t) addr;
2355 bp->b_kvasize = maxsize;
2356 bufspace += bp->b_kvasize;
2359 vm_map_unlock(&buffer_map);
2360 vm_map_entry_release(count);
2362 bp->b_data = bp->b_kvabase;
2370 * Buffer flushing daemon. Buffers are normally flushed by the
2371 * update daemon but if it cannot keep up this process starts to
2372 * take the load in an attempt to prevent getnewbuf() from blocking.
2374 * Once a flush is initiated it does not stop until the number
2375 * of buffers falls below lodirtybuffers, but we will wake up anyone
2376 * waiting at the mid-point.
2378 static struct kproc_desc buf_kp = {
2383 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2384 kproc_start, &buf_kp);
2386 static struct kproc_desc bufhw_kp = {
2391 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2392 kproc_start, &bufhw_kp);
2395 buf_daemon1(struct thread *td, int queue, int (*buf_limit_fn)(long),
2401 marker = kmalloc(sizeof(*marker), M_BIOBUF, M_WAITOK | M_ZERO);
2402 marker->b_flags |= B_MARKER;
2403 marker->b_qindex = BQUEUE_NONE;
2407 * This process needs to be suspended prior to shutdown sync.
2409 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2410 td, SHUTDOWN_PRI_LAST);
2411 curthread->td_flags |= TDF_SYSTHREAD;
2414 * This process is allowed to take the buffer cache to the limit
2417 kproc_suspend_loop();
2420 * Do the flush as long as the number of dirty buffers
2421 * (including those running) exceeds lodirtybufspace.
2423 * When flushing limit running I/O to hirunningspace
2424 * Do the flush. Limit the amount of in-transit I/O we
2425 * allow to build up, otherwise we would completely saturate
2426 * the I/O system. Wakeup any waiting processes before we
2427 * normally would so they can run in parallel with our drain.
2429 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2430 * but because we split the operation into two threads we
2431 * have to cut it in half for each thread.
2433 waitrunningbufspace();
2434 limit = lodirtybufspace / 2;
2435 while (buf_limit_fn(limit)) {
2436 if (flushbufqueues(marker, queue) == 0)
2438 if (runningbufspace < hirunningspace)
2440 waitrunningbufspace();
2444 * We reached our low water mark, reset the
2445 * request and sleep until we are needed again.
2446 * The sleep is just so the suspend code works.
2448 tsleep_interlock(bd_req, 0);
2449 if (atomic_swap_int(bd_req, 0) == 0)
2450 tsleep(bd_req, PINTERLOCKED, "psleep", hz);
2453 /*kfree(marker, M_BIOBUF);*/
2457 buf_daemon_limit(long limit)
2459 return (runningbufspace + dirtykvaspace > limit ||
2460 dirtybufcount - dirtybufcounthw >= nbuf / 2);
2464 buf_daemon_hw_limit(long limit)
2466 return (runningbufspace + dirtykvaspace > limit ||
2467 dirtybufcounthw >= nbuf / 2);
2473 buf_daemon1(bufdaemon_td, BQUEUE_DIRTY, buf_daemon_limit,
2480 buf_daemon1(bufdaemonhw_td, BQUEUE_DIRTY_HW, buf_daemon_hw_limit,
2487 * Try to flush a buffer in the dirty queue. We must be careful to
2488 * free up B_INVAL buffers instead of write them, which NFS is
2489 * particularly sensitive to.
2491 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2492 * that we really want to try to get the buffer out and reuse it
2493 * due to the write load on the machine.
2495 * We must lock the buffer in order to check its validity before we
2496 * can mess with its contents. spin isn't enough.
2499 flushbufqueues(struct buf *marker, bufq_type_t q)
2501 struct bufpcpu *pcpu;
2504 int lcpu = marker->b_qcpu;
2506 KKASSERT(marker->b_qindex == BQUEUE_NONE);
2507 KKASSERT(marker->b_flags & B_MARKER);
2511 * Spinlock needed to perform operations on the queue and may be
2512 * held through a non-blocking BUF_LOCK(), but cannot be held when
2513 * BUF_UNLOCK()ing or through any other major operation.
2515 pcpu = &bufpcpu[marker->b_qcpu];
2516 spin_lock(&pcpu->spin);
2517 marker->b_qindex = q;
2518 TAILQ_INSERT_HEAD(&pcpu->bufqueues[q], marker, b_freelist);
2521 while ((bp = TAILQ_NEXT(bp, b_freelist)) != NULL) {
2523 * NOTE: spinlock is always held at the top of the loop
2525 if (bp->b_flags & B_MARKER)
2527 if ((bp->b_flags & B_DELWRI) == 0) {
2528 kprintf("Unexpected clean buffer %p\n", bp);
2531 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
2533 KKASSERT(bp->b_qcpu == marker->b_qcpu && bp->b_qindex == q);
2536 * Once the buffer is locked we will have no choice but to
2537 * unlock the spinlock around a later BUF_UNLOCK and re-set
2538 * bp = marker when looping. Move the marker now to make
2541 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist);
2542 TAILQ_INSERT_AFTER(&pcpu->bufqueues[q], bp, marker, b_freelist);
2545 * Must recheck B_DELWRI after successfully locking
2548 if ((bp->b_flags & B_DELWRI) == 0) {
2549 spin_unlock(&pcpu->spin);
2551 spin_lock(&pcpu->spin);
2557 * Remove the buffer from its queue. We still own the
2563 * Disposing of an invalid buffer counts as a flush op
2565 if (bp->b_flags & B_INVAL) {
2566 spin_unlock(&pcpu->spin);
2568 spin_lock(&pcpu->spin);
2574 * Release the spinlock for the more complex ops we
2575 * are now going to do.
2577 spin_unlock(&pcpu->spin);
2581 * This is a bit messy
2583 if (LIST_FIRST(&bp->b_dep) != NULL &&
2584 (bp->b_flags & B_DEFERRED) == 0 &&
2585 buf_countdeps(bp, 0)) {
2586 spin_lock(&pcpu->spin);
2587 TAILQ_INSERT_TAIL(&pcpu->bufqueues[q], bp, b_freelist);
2589 bp->b_flags |= B_DEFERRED;
2590 spin_unlock(&pcpu->spin);
2592 spin_lock(&pcpu->spin);
2598 * spinlock not held here.
2600 * If the buffer has a dependancy, buf_checkwrite() must
2601 * also return 0 for us to be able to initate the write.
2603 * If the buffer is flagged B_ERROR it may be requeued
2604 * over and over again, we try to avoid a live lock.
2606 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) {
2608 } else if (bp->b_flags & B_ERROR) {
2609 tsleep(bp, 0, "bioer", 1);
2610 bp->b_flags &= ~B_AGE;
2613 bp->b_flags |= B_AGE;
2616 spin_lock(&pcpu->spin);
2621 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist);
2622 marker->b_qindex = BQUEUE_NONE;
2623 spin_unlock(&pcpu->spin);
2626 * Advance the marker to be fair.
2628 marker->b_qcpu = (marker->b_qcpu + 1) % ncpus;
2630 if (marker->b_qcpu != lcpu)
2640 * Returns true if no I/O is needed to access the associated VM object.
2641 * This is like findblk except it also hunts around in the VM system for
2644 * Note that we ignore vm_page_free() races from interrupts against our
2645 * lookup, since if the caller is not protected our return value will not
2646 * be any more valid then otherwise once we exit the critical section.
2649 inmem(struct vnode *vp, off_t loffset)
2652 vm_offset_t toff, tinc, size;
2656 if (findblk(vp, loffset, FINDBLK_TEST))
2658 if (vp->v_mount == NULL)
2660 if ((obj = vp->v_object) == NULL)
2664 if (size > vp->v_mount->mnt_stat.f_iosize)
2665 size = vp->v_mount->mnt_stat.f_iosize;
2667 vm_object_hold(obj);
2668 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2669 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2675 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2676 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2677 if (vm_page_is_valid(m,
2678 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) {
2683 vm_object_drop(obj);
2690 * Locate and return the specified buffer. Unless flagged otherwise,
2691 * a locked buffer will be returned if it exists or NULL if it does not.
2693 * findblk()'d buffers are still on the bufqueues and if you intend
2694 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2695 * and possibly do other stuff to it.
2697 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2698 * for locking the buffer and ensuring that it remains
2699 * the desired buffer after locking.
2701 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2702 * to acquire the lock we return NULL, even if the
2705 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal
2706 * reuse by getnewbuf() but does not prevent
2707 * disassociation (B_INVAL). Used to avoid deadlocks
2708 * against random (vp,loffset)s due to reassignment.
2710 * (0) - Lock the buffer blocking.
2713 findblk(struct vnode *vp, off_t loffset, int flags)
2718 lkflags = LK_EXCLUSIVE;
2719 if (flags & FINDBLK_NBLOCK)
2720 lkflags |= LK_NOWAIT;
2724 * Lookup. Ref the buf while holding v_token to prevent
2725 * reuse (but does not prevent diassociation).
2727 lwkt_gettoken_shared(&vp->v_token);
2728 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2730 lwkt_reltoken(&vp->v_token);
2734 lwkt_reltoken(&vp->v_token);
2737 * If testing only break and return bp, do not lock.
2739 if (flags & FINDBLK_TEST)
2743 * Lock the buffer, return an error if the lock fails.
2744 * (only FINDBLK_NBLOCK can cause the lock to fail).
2746 if (BUF_LOCK(bp, lkflags)) {
2747 atomic_subtract_int(&bp->b_refs, 1);
2748 /* bp = NULL; not needed */
2753 * Revalidate the locked buf before allowing it to be
2756 if (bp->b_vp == vp && bp->b_loffset == loffset)
2758 atomic_subtract_int(&bp->b_refs, 1);
2765 if ((flags & FINDBLK_REF) == 0)
2766 atomic_subtract_int(&bp->b_refs, 1);
2773 * Similar to getblk() except only returns the buffer if it is
2774 * B_CACHE and requires no other manipulation. Otherwise NULL
2775 * is returned. NULL is also returned if GETBLK_NOWAIT is set
2776 * and the getblk() would block.
2778 * If B_RAM is set the buffer might be just fine, but we return
2779 * NULL anyway because we want the code to fall through to the
2780 * cluster read. Otherwise read-ahead breaks.
2782 * If blksize is 0 the buffer cache buffer must already be fully
2785 * If blksize is non-zero getblk() will be used, allowing a buffer
2786 * to be reinstantiated from its VM backing store. The buffer must
2787 * still be fully cached after reinstantiation to be returned.
2790 getcacheblk(struct vnode *vp, off_t loffset, int blksize, int blkflags)
2793 int fndflags = (blkflags & GETBLK_NOWAIT) ? FINDBLK_NBLOCK : 0;
2796 bp = getblk(vp, loffset, blksize, blkflags, 0);
2798 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2800 bp->b_flags &= ~B_AGE;
2807 bp = findblk(vp, loffset, fndflags);
2809 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2811 bp->b_flags &= ~B_AGE;
2825 * Get a block given a specified block and offset into a file/device.
2826 * B_INVAL may or may not be set on return. The caller should clear
2827 * B_INVAL prior to initiating a READ.
2829 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2830 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2831 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2832 * without doing any of those things the system will likely believe
2833 * the buffer to be valid (especially if it is not B_VMIO), and the
2834 * next getblk() will return the buffer with B_CACHE set.
2836 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2837 * an existing buffer.
2839 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2840 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2841 * and then cleared based on the backing VM. If the previous buffer is
2842 * non-0-sized but invalid, B_CACHE will be cleared.
2844 * If getblk() must create a new buffer, the new buffer is returned with
2845 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2846 * case it is returned with B_INVAL clear and B_CACHE set based on the
2849 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2850 * B_CACHE bit is clear.
2852 * What this means, basically, is that the caller should use B_CACHE to
2853 * determine whether the buffer is fully valid or not and should clear
2854 * B_INVAL prior to issuing a read. If the caller intends to validate
2855 * the buffer by loading its data area with something, the caller needs
2856 * to clear B_INVAL. If the caller does this without issuing an I/O,
2857 * the caller should set B_CACHE ( as an optimization ), else the caller
2858 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2859 * a write attempt or if it was a successfull read. If the caller
2860 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2861 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2865 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2866 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2869 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2872 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2876 if (size > MAXBSIZE)
2877 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2878 if (vp->v_object == NULL)
2879 panic("getblk: vnode %p has no object!", vp);
2882 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) {
2884 * The buffer was found in the cache, but we need to lock it.
2885 * We must acquire a ref on the bp to prevent reuse, but
2886 * this will not prevent disassociation (brelvp()) so we
2887 * must recheck (vp,loffset) after acquiring the lock.
2889 * Without the ref the buffer could potentially be reused
2890 * before we acquire the lock and create a deadlock
2891 * situation between the thread trying to reuse the buffer
2892 * and us due to the fact that we would wind up blocking
2893 * on a random (vp,loffset).
2895 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2896 if (blkflags & GETBLK_NOWAIT) {
2900 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2901 if (blkflags & GETBLK_PCATCH)
2902 lkflags |= LK_PCATCH;
2903 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2906 if (error == ENOLCK)
2910 /* buffer may have changed on us */
2915 * Once the buffer has been locked, make sure we didn't race
2916 * a buffer recyclement. Buffers that are no longer hashed
2917 * will have b_vp == NULL, so this takes care of that check
2920 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2921 kprintf("Warning buffer %p (vp %p loffset %lld) "
2923 bp, vp, (long long)loffset);
2929 * If SZMATCH any pre-existing buffer must be of the requested
2930 * size or NULL is returned. The caller absolutely does not
2931 * want getblk() to bwrite() the buffer on a size mismatch.
2933 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2939 * All vnode-based buffers must be backed by a VM object.
2941 KKASSERT(bp->b_flags & B_VMIO);
2942 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2943 bp->b_flags &= ~B_AGE;
2946 * Make sure that B_INVAL buffers do not have a cached
2947 * block number translation.
2949 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2950 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2951 " did not have cleared bio_offset cache\n",
2952 bp, vp, (long long)loffset);
2953 clearbiocache(&bp->b_bio2);
2957 * The buffer is locked. B_CACHE is cleared if the buffer is
2960 if (bp->b_flags & B_INVAL)
2961 bp->b_flags &= ~B_CACHE;
2965 * Any size inconsistancy with a dirty buffer or a buffer
2966 * with a softupdates dependancy must be resolved. Resizing
2967 * the buffer in such circumstances can lead to problems.
2969 * Dirty or dependant buffers are written synchronously.
2970 * Other types of buffers are simply released and
2971 * reconstituted as they may be backed by valid, dirty VM
2972 * pages (but not marked B_DELWRI).
2974 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2975 * and may be left over from a prior truncation (and thus
2976 * no longer represent the actual EOF point), so we
2977 * definitely do not want to B_NOCACHE the backing store.
2979 if (size != bp->b_bcount) {
2980 if (bp->b_flags & B_DELWRI) {
2981 bp->b_flags |= B_RELBUF;
2983 } else if (LIST_FIRST(&bp->b_dep)) {
2984 bp->b_flags |= B_RELBUF;
2987 bp->b_flags |= B_RELBUF;
2992 KKASSERT(size <= bp->b_kvasize);
2993 KASSERT(bp->b_loffset != NOOFFSET,
2994 ("getblk: no buffer offset"));
2997 * A buffer with B_DELWRI set and B_CACHE clear must
2998 * be committed before we can return the buffer in
2999 * order to prevent the caller from issuing a read
3000 * ( due to B_CACHE not being set ) and overwriting
3003 * Most callers, including NFS and FFS, need this to
3004 * operate properly either because they assume they
3005 * can issue a read if B_CACHE is not set, or because
3006 * ( for example ) an uncached B_DELWRI might loop due
3007 * to softupdates re-dirtying the buffer. In the latter
3008 * case, B_CACHE is set after the first write completes,
3009 * preventing further loops.
3011 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
3012 * above while extending the buffer, we cannot allow the
3013 * buffer to remain with B_CACHE set after the write
3014 * completes or it will represent a corrupt state. To
3015 * deal with this we set B_NOCACHE to scrap the buffer
3018 * XXX Should this be B_RELBUF instead of B_NOCACHE?
3019 * I'm not even sure this state is still possible
3020 * now that getblk() writes out any dirty buffers
3023 * We might be able to do something fancy, like setting
3024 * B_CACHE in bwrite() except if B_DELWRI is already set,
3025 * so the below call doesn't set B_CACHE, but that gets real
3026 * confusing. This is much easier.
3029 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
3030 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
3031 "and CACHE clear, b_flags %08x\n",
3032 bp, (uintmax_t)bp->b_loffset, bp->b_flags);
3033 bp->b_flags |= B_NOCACHE;
3039 * Buffer is not in-core, create new buffer. The buffer
3040 * returned by getnewbuf() is locked. Note that the returned
3041 * buffer is also considered valid (not marked B_INVAL).
3043 * Calculating the offset for the I/O requires figuring out
3044 * the block size. We use DEV_BSIZE for VBLK or VCHR and
3045 * the mount's f_iosize otherwise. If the vnode does not
3046 * have an associated mount we assume that the passed size is
3049 * Note that vn_isdisk() cannot be used here since it may
3050 * return a failure for numerous reasons. Note that the
3051 * buffer size may be larger then the block size (the caller
3052 * will use block numbers with the proper multiple). Beware
3053 * of using any v_* fields which are part of unions. In
3054 * particular, in DragonFly the mount point overloading
3055 * mechanism uses the namecache only and the underlying
3056 * directory vnode is not a special case.
3060 if (vp->v_type == VBLK || vp->v_type == VCHR)
3062 else if (vp->v_mount)
3063 bsize = vp->v_mount->mnt_stat.f_iosize;
3067 maxsize = size + (loffset & PAGE_MASK);
3068 maxsize = imax(maxsize, bsize);
3070 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
3072 if (slpflags || slptimeo)
3078 * Atomically insert the buffer into the hash, so that it can
3079 * be found by findblk().
3081 * If bgetvp() returns non-zero a collision occured, and the
3082 * bp will not be associated with the vnode.
3084 * Make sure the translation layer has been cleared.
3086 bp->b_loffset = loffset;
3087 bp->b_bio2.bio_offset = NOOFFSET;
3088 /* bp->b_bio2.bio_next = NULL; */
3090 if (bgetvp(vp, bp, size)) {
3091 bp->b_flags |= B_INVAL;
3097 * All vnode-based buffers must be backed by a VM object.
3099 KKASSERT(vp->v_object != NULL);
3100 bp->b_flags |= B_VMIO;
3101 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3105 KKASSERT(dsched_is_clear_buf_priv(bp));
3112 * Reacquire a buffer that was previously released to the locked queue,
3113 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
3114 * set B_LOCKED (which handles the acquisition race).
3116 * To this end, either B_LOCKED must be set or the dependancy list must be
3120 regetblk(struct buf *bp)
3122 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
3123 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
3130 * Get an empty, disassociated buffer of given size. The buffer is
3131 * initially set to B_INVAL.
3133 * critical section protection is not required for the allocbuf()
3134 * call because races are impossible here.
3142 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3144 while ((bp = getnewbuf(0, 0, size, maxsize)) == NULL)
3147 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3148 KKASSERT(dsched_is_clear_buf_priv(bp));
3156 * This code constitutes the buffer memory from either anonymous system
3157 * memory (in the case of non-VMIO operations) or from an associated
3158 * VM object (in the case of VMIO operations). This code is able to
3159 * resize a buffer up or down.
3161 * Note that this code is tricky, and has many complications to resolve
3162 * deadlock or inconsistant data situations. Tread lightly!!!
3163 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3164 * the caller. Calling this code willy nilly can result in the loss of
3167 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3168 * B_CACHE for the non-VMIO case.
3170 * This routine does not need to be called from a critical section but you
3171 * must own the buffer.
3174 allocbuf(struct buf *bp, int size)
3176 int newbsize, mbsize;
3179 if (BUF_REFCNT(bp) == 0)
3180 panic("allocbuf: buffer not busy");
3182 if (bp->b_kvasize < size)
3183 panic("allocbuf: buffer too small");
3185 if ((bp->b_flags & B_VMIO) == 0) {
3189 * Just get anonymous memory from the kernel. Don't
3190 * mess with B_CACHE.
3192 mbsize = roundup2(size, DEV_BSIZE);
3193 if (bp->b_flags & B_MALLOC)
3196 newbsize = round_page(size);
3198 if (newbsize < bp->b_bufsize) {
3200 * Malloced buffers are not shrunk
3202 if (bp->b_flags & B_MALLOC) {
3204 bp->b_bcount = size;
3206 kfree(bp->b_data, M_BIOBUF);
3207 if (bp->b_bufsize) {
3208 atomic_subtract_long(&bufmallocspace, bp->b_bufsize);
3212 bp->b_data = bp->b_kvabase;
3214 bp->b_flags &= ~B_MALLOC;
3220 (vm_offset_t) bp->b_data + newbsize,
3221 (vm_offset_t) bp->b_data + bp->b_bufsize);
3222 } else if (newbsize > bp->b_bufsize) {
3224 * We only use malloced memory on the first allocation.
3225 * and revert to page-allocated memory when the buffer
3228 if ((bufmallocspace < maxbufmallocspace) &&
3229 (bp->b_bufsize == 0) &&
3230 (mbsize <= PAGE_SIZE/2)) {
3232 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3233 bp->b_bufsize = mbsize;
3234 bp->b_bcount = size;
3235 bp->b_flags |= B_MALLOC;
3236 atomic_add_long(&bufmallocspace, mbsize);
3242 * If the buffer is growing on its other-than-first
3243 * allocation, then we revert to the page-allocation
3246 if (bp->b_flags & B_MALLOC) {
3247 origbuf = bp->b_data;
3248 origbufsize = bp->b_bufsize;
3249 bp->b_data = bp->b_kvabase;
3250 if (bp->b_bufsize) {
3251 atomic_subtract_long(&bufmallocspace,
3256 bp->b_flags &= ~B_MALLOC;
3257 newbsize = round_page(newbsize);
3261 (vm_offset_t) bp->b_data + bp->b_bufsize,
3262 (vm_offset_t) bp->b_data + newbsize);
3264 bcopy(origbuf, bp->b_data, origbufsize);
3265 kfree(origbuf, M_BIOBUF);
3272 newbsize = roundup2(size, DEV_BSIZE);
3273 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3274 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3275 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3277 if (bp->b_flags & B_MALLOC)
3278 panic("allocbuf: VMIO buffer can't be malloced");
3280 * Set B_CACHE initially if buffer is 0 length or will become
3283 if (size == 0 || bp->b_bufsize == 0)
3284 bp->b_flags |= B_CACHE;
3286 if (newbsize < bp->b_bufsize) {
3288 * DEV_BSIZE aligned new buffer size is less then the
3289 * DEV_BSIZE aligned existing buffer size. Figure out
3290 * if we have to remove any pages.
3292 if (desiredpages < bp->b_xio.xio_npages) {
3293 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3295 * the page is not freed here -- it
3296 * is the responsibility of
3297 * vnode_pager_setsize
3299 m = bp->b_xio.xio_pages[i];
3300 KASSERT(m != bogus_page,
3301 ("allocbuf: bogus page found"));
3302 vm_page_busy_wait(m, TRUE, "biodep");
3303 bp->b_xio.xio_pages[i] = NULL;
3304 vm_page_unwire(m, 0);
3307 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3308 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3309 bp->b_xio.xio_npages = desiredpages;
3311 } else if (size > bp->b_bcount) {
3313 * We are growing the buffer, possibly in a
3314 * byte-granular fashion.
3322 * Step 1, bring in the VM pages from the object,
3323 * allocating them if necessary. We must clear
3324 * B_CACHE if these pages are not valid for the
3325 * range covered by the buffer.
3327 * critical section protection is required to protect
3328 * against interrupts unbusying and freeing pages
3329 * between our vm_page_lookup() and our
3330 * busycheck/wiring call.
3335 vm_object_hold(obj);
3336 while (bp->b_xio.xio_npages < desiredpages) {
3341 pi = OFF_TO_IDX(bp->b_loffset) +
3342 bp->b_xio.xio_npages;
3345 * Blocking on m->busy might lead to a
3348 * vm_fault->getpages->cluster_read->allocbuf
3350 m = vm_page_lookup_busy_try(obj, pi, FALSE,
3353 vm_page_sleep_busy(m, FALSE, "pgtblk");
3358 * note: must allocate system pages
3359 * since blocking here could intefere
3360 * with paging I/O, no matter which
3363 m = bio_page_alloc(bp, obj, pi, desiredpages - bp->b_xio.xio_npages);
3366 vm_page_flag_clear(m, PG_ZERO);
3368 bp->b_flags &= ~B_CACHE;
3369 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3370 ++bp->b_xio.xio_npages;
3376 * We found a page and were able to busy it.
3378 vm_page_flag_clear(m, PG_ZERO);
3381 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3382 ++bp->b_xio.xio_npages;
3383 if (bp->b_act_count < m->act_count)
3384 bp->b_act_count = m->act_count;
3386 vm_object_drop(obj);
3389 * Step 2. We've loaded the pages into the buffer,
3390 * we have to figure out if we can still have B_CACHE
3391 * set. Note that B_CACHE is set according to the
3392 * byte-granular range ( bcount and size ), not the
3393 * aligned range ( newbsize ).
3395 * The VM test is against m->valid, which is DEV_BSIZE
3396 * aligned. Needless to say, the validity of the data
3397 * needs to also be DEV_BSIZE aligned. Note that this
3398 * fails with NFS if the server or some other client
3399 * extends the file's EOF. If our buffer is resized,
3400 * B_CACHE may remain set! XXX
3403 toff = bp->b_bcount;
3404 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3406 while ((bp->b_flags & B_CACHE) && toff < size) {
3409 if (tinc > (size - toff))
3412 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3420 bp->b_xio.xio_pages[pi]
3427 * Step 3, fixup the KVM pmap. Remember that
3428 * bp->b_data is relative to bp->b_loffset, but
3429 * bp->b_loffset may be offset into the first page.
3432 bp->b_data = (caddr_t)
3433 trunc_page((vm_offset_t)bp->b_data);
3435 (vm_offset_t)bp->b_data,
3436 bp->b_xio.xio_pages,
3437 bp->b_xio.xio_npages
3439 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3440 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3444 /* adjust space use on already-dirty buffer */
3445 if (bp->b_flags & B_DELWRI) {
3446 /* dirtykvaspace unchanged */
3447 atomic_add_long(&dirtybufspace, newbsize - bp->b_bufsize);
3448 if (bp->b_flags & B_HEAVY) {
3449 atomic_add_long(&dirtybufspacehw,
3450 newbsize - bp->b_bufsize);
3453 if (newbsize < bp->b_bufsize)
3455 bp->b_bufsize = newbsize; /* actual buffer allocation */
3456 bp->b_bcount = size; /* requested buffer size */
3463 * Wait for buffer I/O completion, returning error status. B_EINTR
3464 * is converted into an EINTR error but not cleared (since a chain
3465 * of biowait() calls may occur).
3467 * On return bpdone() will have been called but the buffer will remain
3468 * locked and will not have been brelse()'d.
3470 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3471 * likely still in progress on return.
3473 * NOTE! This operation is on a BIO, not a BUF.
3475 * NOTE! BIO_DONE is cleared by vn_strategy()
3478 _biowait(struct bio *bio, const char *wmesg, int to)
3480 struct buf *bp = bio->bio_buf;
3485 KKASSERT(bio == &bp->b_bio1);
3487 flags = bio->bio_flags;
3488 if (flags & BIO_DONE)
3490 nflags = flags | BIO_WANT;
3491 tsleep_interlock(bio, 0);
3492 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3494 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3495 else if (bp->b_cmd == BUF_CMD_READ)
3496 error = tsleep(bio, PINTERLOCKED, "biord", to);
3498 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3500 kprintf("tsleep error biowait %d\n", error);
3509 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3510 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3511 if (bp->b_flags & B_EINTR)
3513 if (bp->b_flags & B_ERROR)
3514 return (bp->b_error ? bp->b_error : EIO);
3519 biowait(struct bio *bio, const char *wmesg)
3521 return(_biowait(bio, wmesg, 0));
3525 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3527 return(_biowait(bio, wmesg, to));
3531 * This associates a tracking count with an I/O. vn_strategy() and
3532 * dev_dstrategy() do this automatically but there are a few cases
3533 * where a vnode or device layer is bypassed when a block translation
3534 * is cached. In such cases bio_start_transaction() may be called on
3535 * the bypassed layers so the system gets an I/O in progress indication
3536 * for those higher layers.
3539 bio_start_transaction(struct bio *bio, struct bio_track *track)
3541 bio->bio_track = track;
3542 if (dsched_is_clear_buf_priv(bio->bio_buf))
3543 dsched_new_buf(bio->bio_buf);
3544 bio_track_ref(track);
3548 * Initiate I/O on a vnode.
3550 * SWAPCACHE OPERATION:
3552 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately
3553 * devfs also uses b_vp for fake buffers so we also have to check
3554 * that B_PAGING is 0. In this case the passed 'vp' is probably the
3555 * underlying block device. The swap assignments are related to the
3556 * buffer cache buffer's b_vp, not the passed vp.
3558 * The passed vp == bp->b_vp only in the case where the strategy call
3559 * is made on the vp itself for its own buffers (a regular file or
3560 * block device vp). The filesystem usually then re-calls vn_strategy()
3561 * after translating the request to an underlying device.
3563 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the
3564 * underlying buffer cache buffers.
3566 * We can only deal with page-aligned buffers at the moment, because
3567 * we can't tell what the real dirty state for pages straddling a buffer
3570 * In order to call swap_pager_strategy() we must provide the VM object
3571 * and base offset for the underlying buffer cache pages so it can find
3575 vn_strategy(struct vnode *vp, struct bio *bio)
3577 struct bio_track *track;
3578 struct buf *bp = bio->bio_buf;
3580 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3583 * Set when an I/O is issued on the bp. Cleared by consumers
3584 * (aka HAMMER), allowing the consumer to determine if I/O had
3585 * actually occurred.
3587 bp->b_flags |= B_IODEBUG;
3590 * Handle the swap cache intercept.
3592 if (vn_cache_strategy(vp, bio))
3596 * Otherwise do the operation through the filesystem
3598 if (bp->b_cmd == BUF_CMD_READ)
3599 track = &vp->v_track_read;
3601 track = &vp->v_track_write;
3602 KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3603 bio->bio_track = track;
3604 if (dsched_is_clear_buf_priv(bio->bio_buf))
3605 dsched_new_buf(bio->bio_buf);
3606 bio_track_ref(track);
3607 vop_strategy(*vp->v_ops, vp, bio);
3610 static void vn_cache_strategy_callback(struct bio *bio);
3613 vn_cache_strategy(struct vnode *vp, struct bio *bio)
3615 struct buf *bp = bio->bio_buf;
3622 * Stop using swapcache if paniced, dumping, or dumped
3624 if (panicstr || dumping)
3628 * Is this buffer cache buffer suitable for reading from
3631 if (vm_swapcache_read_enable == 0 ||
3632 bp->b_cmd != BUF_CMD_READ ||
3633 ((bp->b_flags & B_CLUSTER) == 0 &&
3634 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) ||
3635 ((int)bp->b_loffset & PAGE_MASK) != 0 ||
3636 (bp->b_bcount & PAGE_MASK) != 0) {
3641 * Figure out the original VM object (it will match the underlying
3642 * VM pages). Note that swap cached data uses page indices relative
3643 * to that object, not relative to bio->bio_offset.
3645 if (bp->b_flags & B_CLUSTER)
3646 object = vp->v_object;
3648 object = bp->b_vp->v_object;
3651 * In order to be able to use the swap cache all underlying VM
3652 * pages must be marked as such, and we can't have any bogus pages.
3654 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
3655 m = bp->b_xio.xio_pages[i];
3656 if ((m->flags & PG_SWAPPED) == 0)
3658 if (m == bogus_page)
3663 * If we are good then issue the I/O using swap_pager_strategy().
3665 * We can only do this if the buffer actually supports object-backed
3666 * I/O. If it doesn't npages will be 0.
3668 if (i && i == bp->b_xio.xio_npages) {
3669 m = bp->b_xio.xio_pages[0];
3670 nbio = push_bio(bio);
3671 nbio->bio_done = vn_cache_strategy_callback;
3672 nbio->bio_offset = ptoa(m->pindex);
3673 KKASSERT(m->object == object);
3674 swap_pager_strategy(object, nbio);
3681 * This is a bit of a hack but since the vn_cache_strategy() function can
3682 * override a VFS's strategy function we must make sure that the bio, which
3683 * is probably bio2, doesn't leak an unexpected offset value back to the
3684 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the
3685 * bio went through its own file strategy function and the the bio2 offset
3686 * is a cached disk offset when, in fact, it isn't.
3689 vn_cache_strategy_callback(struct bio *bio)
3691 bio->bio_offset = NOOFFSET;
3692 biodone(pop_bio(bio));
3698 * Finish I/O on a buffer after all BIOs have been processed.
3699 * Called when the bio chain is exhausted or by biowait. If called
3700 * by biowait, elseit is typically 0.
3702 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3703 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3704 * assuming B_INVAL is clear.
3706 * For the VMIO case, we set B_CACHE if the op was a read and no
3707 * read error occured, or if the op was a write. B_CACHE is never
3708 * set if the buffer is invalid or otherwise uncacheable.
3710 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3711 * initiator to leave B_INVAL set to brelse the buffer out of existance
3712 * in the biodone routine.
3714 * bpdone is responsible for calling bundirty() on the buffer after a
3715 * successful write. We previously did this prior to initiating the
3716 * write under the assumption that the buffer might be dirtied again
3717 * while the write was in progress, however doing it before-hand creates
3718 * a race condition prior to the call to vn_strategy() where the
3719 * filesystem may not be aware that a dirty buffer is present.
3720 * It should not be possible for the buffer or its underlying pages to
3721 * be redirtied prior to bpdone()'s unbusying of the underlying VM
3725 bpdone(struct buf *bp, int elseit)
3729 KASSERT(BUF_REFCNTNB(bp) > 0,
3730 ("bpdone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3731 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3732 ("bpdone: bp %p already done!", bp));
3735 * No more BIOs are left. All completion functions have been dealt
3736 * with, now we clean up the buffer.
3739 bp->b_cmd = BUF_CMD_DONE;
3742 * Only reads and writes are processed past this point.
3744 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3745 if (cmd == BUF_CMD_FREEBLKS)
3746 bp->b_flags |= B_NOCACHE;
3753 * A failed write must re-dirty the buffer unless B_INVAL
3756 * A successful write must clear the dirty flag. This is done after
3757 * the write to ensure that the buffer remains on the vnode's dirty
3758 * list for filesystem interlocks / checks until the write is actually
3759 * complete. HAMMER2 is sensitive to this issue.
3761 * Only applicable to normal buffers (with VPs). vinum buffers may
3764 * Must be done prior to calling buf_complete() as the callback might
3765 * re-dirty the buffer.
3767 if (cmd == BUF_CMD_WRITE) {
3768 if ((bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3769 bp->b_flags &= ~B_NOCACHE;
3779 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3780 * a lot worse. XXX - move this above the clearing of b_cmd
3782 if (LIST_FIRST(&bp->b_dep) != NULL)
3785 if (bp->b_flags & B_VMIO) {
3791 struct vnode *vp = bp->b_vp;
3795 #if defined(VFS_BIO_DEBUG)
3796 if (vp->v_auxrefs == 0)
3797 panic("bpdone: zero vnode hold count");
3798 if ((vp->v_flag & VOBJBUF) == 0)
3799 panic("bpdone: vnode is not setup for merged cache");
3802 foff = bp->b_loffset;
3803 KASSERT(foff != NOOFFSET, ("bpdone: no buffer offset"));
3804 KASSERT(obj != NULL, ("bpdone: missing VM object"));
3806 #if defined(VFS_BIO_DEBUG)
3807 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3808 kprintf("bpdone: paging in progress(%d) < "
3809 "bp->b_xio.xio_npages(%d)\n",
3810 obj->paging_in_progress,
3811 bp->b_xio.xio_npages);
3816 * Set B_CACHE if the op was a normal read and no error
3817 * occured. B_CACHE is set for writes in the b*write()
3820 iosize = bp->b_bcount - bp->b_resid;
3821 if (cmd == BUF_CMD_READ &&
3822 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3823 bp->b_flags |= B_CACHE;
3826 vm_object_hold(obj);
3827 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3831 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3836 * cleanup bogus pages, restoring the originals. Since
3837 * the originals should still be wired, we don't have
3838 * to worry about interrupt/freeing races destroying
3839 * the VM object association.
3841 m = bp->b_xio.xio_pages[i];
3842 if (m == bogus_page) {
3844 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3846 panic("bpdone: page disappeared");
3847 bp->b_xio.xio_pages[i] = m;
3848 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3849 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3851 #if defined(VFS_BIO_DEBUG)
3852 if (OFF_TO_IDX(foff) != m->pindex) {
3853 kprintf("bpdone: foff(%lu)/m->pindex(%ld) "
3855 (unsigned long)foff, (long)m->pindex);
3860 * In the write case, the valid and clean bits are
3861 * already changed correctly (see bdwrite()), so we
3862 * only need to do this here in the read case.
3864 vm_page_busy_wait(m, FALSE, "bpdpgw");
3865 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3866 vfs_clean_one_page(bp, i, m);
3868 vm_page_flag_clear(m, PG_ZERO);
3871 * when debugging new filesystems or buffer I/O
3872 * methods, this is the most common error that pops
3873 * up. if you see this, you have not set the page
3874 * busy flag correctly!!!
3877 kprintf("bpdone: page busy < 0, "
3878 "pindex: %d, foff: 0x(%x,%x), "
3879 "resid: %d, index: %d\n",
3880 (int) m->pindex, (int)(foff >> 32),
3881 (int) foff & 0xffffffff, resid, i);
3882 if (!vn_isdisk(vp, NULL))
3883 kprintf(" iosize: %ld, loffset: %lld, "
3884 "flags: 0x%08x, npages: %d\n",
3885 bp->b_vp->v_mount->mnt_stat.f_iosize,
3886 (long long)bp->b_loffset,
3887 bp->b_flags, bp->b_xio.xio_npages);
3889 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3890 (long long)bp->b_loffset,
3891 bp->b_flags, bp->b_xio.xio_npages);
3892 kprintf(" valid: 0x%x, dirty: 0x%x, "
3896 panic("bpdone: page busy < 0");
3898 vm_page_io_finish(m);
3900 vm_object_pip_wakeup(obj);
3901 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3904 bp->b_flags &= ~B_HASBOGUS;
3905 vm_object_drop(obj);
3909 * Finish up by releasing the buffer. There are no more synchronous
3910 * or asynchronous completions, those were handled by bio_done
3914 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
3925 biodone(struct bio *bio)
3927 struct buf *bp = bio->bio_buf;
3929 runningbufwakeup(bp);
3932 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3935 biodone_t *done_func;
3936 struct bio_track *track;
3939 * BIO tracking. Most but not all BIOs are tracked.
3941 if ((track = bio->bio_track) != NULL) {
3942 bio_track_rel(track);
3943 bio->bio_track = NULL;
3947 * A bio_done function terminates the loop. The function
3948 * will be responsible for any further chaining and/or
3949 * buffer management.
3951 * WARNING! The done function can deallocate the buffer!
3953 if ((done_func = bio->bio_done) != NULL) {
3954 bio->bio_done = NULL;
3958 bio = bio->bio_prev;
3962 * If we've run out of bio's do normal [a]synchronous completion.
3968 * Synchronous biodone - this terminates a synchronous BIO.
3970 * bpdone() is called with elseit=FALSE, leaving the buffer completed
3971 * but still locked. The caller must brelse() the buffer after waiting
3975 biodone_sync(struct bio *bio)
3977 struct buf *bp = bio->bio_buf;
3981 KKASSERT(bio == &bp->b_bio1);
3985 flags = bio->bio_flags;
3986 nflags = (flags | BIO_DONE) & ~BIO_WANT;
3988 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3989 if (flags & BIO_WANT)
3999 * This routine is called in lieu of iodone in the case of
4000 * incomplete I/O. This keeps the busy status for pages
4004 vfs_unbusy_pages(struct buf *bp)
4008 runningbufwakeup(bp);
4010 if (bp->b_flags & B_VMIO) {
4011 struct vnode *vp = bp->b_vp;
4015 vm_object_hold(obj);
4017 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4018 vm_page_t m = bp->b_xio.xio_pages[i];
4021 * When restoring bogus changes the original pages
4022 * should still be wired, so we are in no danger of
4023 * losing the object association and do not need
4024 * critical section protection particularly.
4026 if (m == bogus_page) {
4027 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
4029 panic("vfs_unbusy_pages: page missing");
4031 bp->b_xio.xio_pages[i] = m;
4032 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4033 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
4035 vm_page_busy_wait(m, FALSE, "bpdpgw");
4036 vm_page_flag_clear(m, PG_ZERO);
4037 vm_page_io_finish(m);
4039 vm_object_pip_wakeup(obj);
4041 bp->b_flags &= ~B_HASBOGUS;
4042 vm_object_drop(obj);
4049 * This routine is called before a device strategy routine.
4050 * It is used to tell the VM system that paging I/O is in
4051 * progress, and treat the pages associated with the buffer
4052 * almost as being PG_BUSY. Also the object 'paging_in_progress'
4053 * flag is handled to make sure that the object doesn't become
4056 * Since I/O has not been initiated yet, certain buffer flags
4057 * such as B_ERROR or B_INVAL may be in an inconsistant state
4058 * and should be ignored.
4061 vfs_busy_pages(struct vnode *vp, struct buf *bp)
4064 struct lwp *lp = curthread->td_lwp;
4067 * The buffer's I/O command must already be set. If reading,
4068 * B_CACHE must be 0 (double check against callers only doing
4069 * I/O when B_CACHE is 0).
4071 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4072 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
4074 if (bp->b_flags & B_VMIO) {
4078 KASSERT(bp->b_loffset != NOOFFSET,
4079 ("vfs_busy_pages: no buffer offset"));
4082 * Busy all the pages. We have to busy them all at once
4083 * to avoid deadlocks.
4086 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4087 vm_page_t m = bp->b_xio.xio_pages[i];
4089 if (vm_page_busy_try(m, FALSE)) {
4090 vm_page_sleep_busy(m, FALSE, "vbpage");
4092 vm_page_wakeup(bp->b_xio.xio_pages[i]);
4098 * Setup for I/O, soft-busy the page right now because
4099 * the next loop may block.
4101 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4102 vm_page_t m = bp->b_xio.xio_pages[i];
4104 vm_page_flag_clear(m, PG_ZERO);
4105 if ((bp->b_flags & B_CLUSTER) == 0) {
4106 vm_object_pip_add(obj, 1);
4107 vm_page_io_start(m);
4112 * Adjust protections for I/O and do bogus-page mapping.
4113 * Assume that vm_page_protect() can block (it can block
4114 * if VM_PROT_NONE, don't take any chances regardless).
4116 * In particular note that for writes we must incorporate
4117 * page dirtyness from the VM system into the buffer's
4120 * For reads we theoretically must incorporate page dirtyness
4121 * from the VM system to determine if the page needs bogus
4122 * replacement, but we shortcut the test by simply checking
4123 * that all m->valid bits are set, indicating that the page
4124 * is fully valid and does not need to be re-read. For any
4125 * VM system dirtyness the page will also be fully valid
4126 * since it was mapped at one point.
4129 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4130 vm_page_t m = bp->b_xio.xio_pages[i];
4132 vm_page_flag_clear(m, PG_ZERO); /* XXX */
4133 if (bp->b_cmd == BUF_CMD_WRITE) {
4135 * When readying a vnode-backed buffer for
4136 * a write we must zero-fill any invalid
4137 * portions of the backing VM pages, mark
4138 * it valid and clear related dirty bits.
4140 * vfs_clean_one_page() incorporates any
4141 * VM dirtyness and updates the b_dirtyoff
4142 * range (after we've made the page RO).
4144 * It is also expected that the pmap modified
4145 * bit has already been cleared by the
4146 * vm_page_protect(). We may not be able
4147 * to clear all dirty bits for a page if it
4148 * was also memory mapped (NFS).
4150 * Finally be sure to unassign any swap-cache
4151 * backing store as it is now stale.
4153 vm_page_protect(m, VM_PROT_READ);
4154 vfs_clean_one_page(bp, i, m);
4155 swap_pager_unswapped(m);
4156 } else if (m->valid == VM_PAGE_BITS_ALL) {
4158 * When readying a vnode-backed buffer for
4159 * read we must replace any dirty pages with
4160 * a bogus page so dirty data is not destroyed
4161 * when filling gaps.
4163 * To avoid testing whether the page is
4164 * dirty we instead test that the page was
4165 * at some point mapped (m->valid fully
4166 * valid) with the understanding that
4167 * this also covers the dirty case.
4169 bp->b_xio.xio_pages[i] = bogus_page;
4170 bp->b_flags |= B_HASBOGUS;
4172 } else if (m->valid & m->dirty) {
4174 * This case should not occur as partial
4175 * dirtyment can only happen if the buffer
4176 * is B_CACHE, and this code is not entered
4177 * if the buffer is B_CACHE.
4179 kprintf("Warning: vfs_busy_pages - page not "
4180 "fully valid! loff=%jx bpf=%08x "
4181 "idx=%d val=%02x dir=%02x\n",
4182 (uintmax_t)bp->b_loffset, bp->b_flags,
4183 i, m->valid, m->dirty);
4184 vm_page_protect(m, VM_PROT_NONE);
4187 * The page is not valid and can be made
4190 vm_page_protect(m, VM_PROT_NONE);
4195 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4196 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
4201 * This is the easiest place to put the process accounting for the I/O
4205 if (bp->b_cmd == BUF_CMD_READ)
4206 lp->lwp_ru.ru_inblock++;
4208 lp->lwp_ru.ru_oublock++;
4213 * Tell the VM system that the pages associated with this buffer
4214 * are clean. This is used for delayed writes where the data is
4215 * going to go to disk eventually without additional VM intevention.
4217 * NOTE: While we only really need to clean through to b_bcount, we
4218 * just go ahead and clean through to b_bufsize.
4221 vfs_clean_pages(struct buf *bp)
4226 if ((bp->b_flags & B_VMIO) == 0)
4229 KASSERT(bp->b_loffset != NOOFFSET,
4230 ("vfs_clean_pages: no buffer offset"));
4232 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4233 m = bp->b_xio.xio_pages[i];
4234 vfs_clean_one_page(bp, i, m);
4239 * vfs_clean_one_page:
4241 * Set the valid bits and clear the dirty bits in a page within a
4242 * buffer. The range is restricted to the buffer's size and the
4243 * buffer's logical offset might index into the first page.
4245 * The caller has busied or soft-busied the page and it is not mapped,
4246 * test and incorporate the dirty bits into b_dirtyoff/end before
4247 * clearing them. Note that we need to clear the pmap modified bits
4248 * after determining the the page was dirty, vm_page_set_validclean()
4249 * does not do it for us.
4251 * This routine is typically called after a read completes (dirty should
4252 * be zero in that case as we are not called on bogus-replace pages),
4253 * or before a write is initiated.
4256 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m)
4264 * Calculate offset range within the page but relative to buffer's
4265 * loffset. loffset might be offset into the first page.
4267 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4268 bcount = bp->b_bcount + xoff; /* offset adjusted */
4274 soff = (pageno << PAGE_SHIFT);
4275 eoff = soff + PAGE_SIZE;
4283 * Test dirty bits and adjust b_dirtyoff/end.
4285 * If dirty pages are incorporated into the bp any prior
4286 * B_NEEDCOMMIT state (NFS) must be cleared because the
4287 * caller has not taken into account the new dirty data.
4289 * If the page was memory mapped the dirty bits might go beyond the
4290 * end of the buffer, but we can't really make the assumption that
4291 * a file EOF straddles the buffer (even though this is the case for
4292 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
4293 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
4294 * This also saves some console spam.
4296 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK,
4297 * NFS can handle huge commits but not huge writes.
4299 vm_page_test_dirty(m);
4301 if ((bp->b_flags & B_NEEDCOMMIT) &&
4302 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) {
4304 kprintf("Warning: vfs_clean_one_page: bp %p "
4305 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT"
4306 " cmd %d vd %02x/%02x x/s/e %d %d %d "
4308 bp, (uintmax_t)bp->b_loffset, bp->b_bcount,
4309 bp->b_flags, bp->b_cmd,
4310 m->valid, m->dirty, xoff, soff, eoff,
4311 bp->b_dirtyoff, bp->b_dirtyend);
4312 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
4314 print_backtrace(-1);
4317 * Only clear the pmap modified bits if ALL the dirty bits
4318 * are set, otherwise the system might mis-clear portions
4321 if (m->dirty == VM_PAGE_BITS_ALL &&
4322 (bp->b_flags & B_NEEDCOMMIT) == 0) {
4323 pmap_clear_modify(m);
4325 if (bp->b_dirtyoff > soff - xoff)
4326 bp->b_dirtyoff = soff - xoff;
4327 if (bp->b_dirtyend < eoff - xoff)
4328 bp->b_dirtyend = eoff - xoff;
4332 * Set related valid bits, clear related dirty bits.
4333 * Does not mess with the pmap modified bit.
4335 * WARNING! We cannot just clear all of m->dirty here as the
4336 * buffer cache buffers may use a DEV_BSIZE'd aligned
4337 * block size, or have an odd size (e.g. NFS at file EOF).
4338 * The putpages code can clear m->dirty to 0.
4340 * If a VOP_WRITE generates a buffer cache buffer which
4341 * covers the same space as mapped writable pages the
4342 * buffer flush might not be able to clear all the dirty
4343 * bits and still require a putpages from the VM system
4346 * WARNING! vm_page_set_validclean() currently assumes vm_token
4347 * is held. The page might not be busied (bdwrite() case).
4348 * XXX remove this comment once we've validated that this
4349 * is no longer an issue.
4351 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff);
4356 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty.
4357 * The page data is assumed to be valid (there is no zeroing here).
4360 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m)
4368 * Calculate offset range within the page but relative to buffer's
4369 * loffset. loffset might be offset into the first page.
4371 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4372 bcount = bp->b_bcount + xoff; /* offset adjusted */
4378 soff = (pageno << PAGE_SHIFT);
4379 eoff = soff + PAGE_SIZE;
4385 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff);
4392 * Clear a buffer. This routine essentially fakes an I/O, so we need
4393 * to clear B_ERROR and B_INVAL.
4395 * Note that while we only theoretically need to clear through b_bcount,
4396 * we go ahead and clear through b_bufsize.
4400 vfs_bio_clrbuf(struct buf *bp)
4404 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
4405 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
4406 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4407 (bp->b_loffset & PAGE_MASK) == 0) {
4408 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4409 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
4413 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
4414 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
4415 bzero(bp->b_data, bp->b_bufsize);
4416 bp->b_xio.xio_pages[0]->valid |= mask;
4422 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
4423 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
4424 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
4425 ea = (caddr_t)(vm_offset_t)ulmin(
4426 (u_long)(vm_offset_t)ea,
4427 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
4428 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4429 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
4431 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
4432 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
4436 for (; sa < ea; sa += DEV_BSIZE, j++) {
4437 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
4438 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
4439 bzero(sa, DEV_BSIZE);
4442 bp->b_xio.xio_pages[i]->valid |= mask;
4443 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
4452 * vm_hold_load_pages:
4454 * Load pages into the buffer's address space. The pages are
4455 * allocated from the kernel object in order to reduce interference
4456 * with the any VM paging I/O activity. The range of loaded
4457 * pages will be wired.
4459 * If a page cannot be allocated, the 'pagedaemon' is woken up to
4460 * retrieve the full range (to - from) of pages.
4463 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4469 to = round_page(to);
4470 from = round_page(from);
4471 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4476 * Note: must allocate system pages since blocking here
4477 * could intefere with paging I/O, no matter which
4480 vm_object_hold(&kernel_object);
4481 p = bio_page_alloc(bp, &kernel_object, pg >> PAGE_SHIFT,
4482 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
4483 vm_object_drop(&kernel_object);
4486 p->valid = VM_PAGE_BITS_ALL;
4487 vm_page_flag_clear(p, PG_ZERO);
4488 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
4489 bp->b_xio.xio_pages[index] = p;
4496 bp->b_xio.xio_npages = index;
4500 * Allocate a page for a buffer cache buffer.
4502 * If NULL is returned the caller is expected to retry (typically check if
4503 * the page already exists on retry before trying to allocate one).
4505 * NOTE! Low-memory handling is dealt with in b[q]relse(), not here. This
4506 * function will use the system reserve with the hope that the page
4507 * allocations can be returned to PQ_CACHE/PQ_FREE when the caller
4508 * is done with the buffer.
4510 * NOTE! However, TMPFS is a special case because flushing a dirty buffer
4511 * to TMPFS doesn't clean the page. For TMPFS, only the pagedaemon
4512 * is capable of retiring pages (to swap). For TMPFS we don't dig
4513 * into the system reserve because doing so could stall out pretty
4514 * much every process running on the system.
4518 bio_page_alloc(struct buf *bp, vm_object_t obj, vm_pindex_t pg, int deficit)
4520 int vmflags = VM_ALLOC_NORMAL | VM_ALLOC_NULL_OK;
4523 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj));
4526 * Try a normal allocation first.
4528 p = vm_page_alloc(obj, pg, vmflags);
4531 if (vm_page_lookup(obj, pg))
4533 vm_pageout_deficit += deficit;
4536 * Try again, digging into the system reserve.
4538 * Trying to recover pages from the buffer cache here can deadlock
4539 * against other threads trying to busy underlying pages so we
4540 * depend on the code in brelse() and bqrelse() to free/cache the
4541 * underlying buffer cache pages when memory is low.
4543 if (curthread->td_flags & TDF_SYSTHREAD)
4544 vmflags |= VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT;
4545 else if (bp->b_vp && bp->b_vp->v_tag == VT_TMPFS)
4548 vmflags |= VM_ALLOC_SYSTEM;
4550 /*recoverbufpages();*/
4551 p = vm_page_alloc(obj, pg, vmflags);
4554 if (vm_page_lookup(obj, pg))
4558 * Wait for memory to free up and try again
4560 if (vm_page_count_severe())
4562 vm_wait(hz / 20 + 1);
4564 p = vm_page_alloc(obj, pg, vmflags);
4567 if (vm_page_lookup(obj, pg))
4571 * Ok, now we are really in trouble.
4574 static struct krate biokrate = { .freq = 1 };
4575 krateprintf(&biokrate,
4576 "Warning: bio_page_alloc: memory exhausted "
4577 "during bufcache page allocation from %s\n",
4578 curthread->td_comm);
4580 if (curthread->td_flags & TDF_SYSTHREAD)
4581 vm_wait(hz / 20 + 1);
4583 vm_wait(hz / 2 + 1);
4588 * vm_hold_free_pages:
4590 * Return pages associated with the buffer back to the VM system.
4592 * The range of pages underlying the buffer's address space will
4593 * be unmapped and un-wired.
4596 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4600 int index, newnpages;
4602 from = round_page(from);
4603 to = round_page(to);
4604 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4607 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4608 p = bp->b_xio.xio_pages[index];
4609 if (p && (index < bp->b_xio.xio_npages)) {
4611 kprintf("vm_hold_free_pages: doffset: %lld, "
4613 (long long)bp->b_bio2.bio_offset,
4614 (long long)bp->b_loffset);
4616 bp->b_xio.xio_pages[index] = NULL;
4618 vm_page_busy_wait(p, FALSE, "vmhldpg");
4619 vm_page_unwire(p, 0);
4623 bp->b_xio.xio_npages = newnpages;
4629 * Map a user buffer into KVM via a pbuf. On return the buffer's
4630 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4634 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4645 * bp had better have a command and it better be a pbuf.
4647 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4648 KKASSERT(bp->b_flags & B_PAGING);
4649 KKASSERT(bp->b_kvabase);
4655 * Map the user data into KVM. Mappings have to be page-aligned.
4657 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4660 vmprot = VM_PROT_READ;
4661 if (bp->b_cmd == BUF_CMD_READ)
4662 vmprot |= VM_PROT_WRITE;
4664 while (addr < udata + bytes) {
4666 * Do the vm_fault if needed; do the copy-on-write thing
4667 * when reading stuff off device into memory.
4669 * vm_fault_page*() returns a held VM page.
4671 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4672 va = trunc_page(va);
4674 m = vm_fault_page_quick(va, vmprot, &error);
4676 for (i = 0; i < pidx; ++i) {
4677 vm_page_unhold(bp->b_xio.xio_pages[i]);
4678 bp->b_xio.xio_pages[i] = NULL;
4682 bp->b_xio.xio_pages[pidx] = m;
4688 * Map the page array and set the buffer fields to point to
4689 * the mapped data buffer.
4691 if (pidx > btoc(MAXPHYS))
4692 panic("vmapbuf: mapped more than MAXPHYS");
4693 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4695 bp->b_xio.xio_npages = pidx;
4696 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4697 bp->b_bcount = bytes;
4698 bp->b_bufsize = bytes;
4705 * Free the io map PTEs associated with this IO operation.
4706 * We also invalidate the TLB entries and restore the original b_addr.
4709 vunmapbuf(struct buf *bp)
4714 KKASSERT(bp->b_flags & B_PAGING);
4716 npages = bp->b_xio.xio_npages;
4717 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4718 for (pidx = 0; pidx < npages; ++pidx) {
4719 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4720 bp->b_xio.xio_pages[pidx] = NULL;
4722 bp->b_xio.xio_npages = 0;
4723 bp->b_data = bp->b_kvabase;
4727 * Scan all buffers in the system and issue the callback.
4730 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4736 for (n = 0; n < nbuf; ++n) {
4737 if ((error = callback(&buf[n], info)) < 0) {
4747 * nestiobuf_iodone: biodone callback for nested buffers and propagate
4748 * completion to the master buffer.
4751 nestiobuf_iodone(struct bio *bio)
4754 struct buf *mbp, *bp;
4755 struct devstat *stats;
4760 mbio = bio->bio_caller_info1.ptr;
4761 stats = bio->bio_caller_info2.ptr;
4762 mbp = mbio->bio_buf;
4764 KKASSERT(bp->b_bcount <= bp->b_bufsize);
4765 KKASSERT(mbp != bp);
4767 error = bp->b_error;
4768 if (bp->b_error == 0 &&
4769 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
4771 * Not all got transfered, raise an error. We have no way to
4772 * propagate these conditions to mbp.
4777 donebytes = bp->b_bufsize;
4781 nestiobuf_done(mbio, donebytes, error, stats);
4785 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats)
4789 mbp = mbio->bio_buf;
4791 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0);
4794 * If an error occured, propagate it to the master buffer.
4796 * Several biodone()s may wind up running concurrently so
4797 * use an atomic op to adjust b_flags.
4800 mbp->b_error = error;
4801 atomic_set_int(&mbp->b_flags, B_ERROR);
4805 * Decrement the operations in progress counter and terminate the
4806 * I/O if this was the last bit.
4808 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4811 devstat_end_transaction_buf(stats, mbp);
4817 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent
4818 * the mbio from being biodone()'d while we are still adding sub-bios to
4822 nestiobuf_init(struct bio *bio)
4824 bio->bio_driver_info = (void *)1;
4828 * The BIOs added to the nestedio have already been started, remove the
4829 * count that placeheld our mbio and biodone() it if the count would
4833 nestiobuf_start(struct bio *mbio)
4835 struct buf *mbp = mbio->bio_buf;
4838 * Decrement the operations in progress counter and terminate the
4839 * I/O if this was the last bit.
4841 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4842 if (mbp->b_flags & B_ERROR)
4843 mbp->b_resid = mbp->b_bcount;
4851 * Set an intermediate error prior to calling nestiobuf_start()
4854 nestiobuf_error(struct bio *mbio, int error)
4856 struct buf *mbp = mbio->bio_buf;
4859 mbp->b_error = error;
4860 atomic_set_int(&mbp->b_flags, B_ERROR);
4865 * nestiobuf_add: setup a "nested" buffer.
4867 * => 'mbp' is a "master" buffer which is being divided into sub pieces.
4868 * => 'bp' should be a buffer allocated by getiobuf.
4869 * => 'offset' is a byte offset in the master buffer.
4870 * => 'size' is a size in bytes of this nested buffer.
4873 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats)
4875 struct buf *mbp = mbio->bio_buf;
4876 struct vnode *vp = mbp->b_vp;
4878 KKASSERT(mbp->b_bcount >= offset + size);
4880 atomic_add_int((int *)&mbio->bio_driver_info, 1);
4882 /* kernel needs to own the lock for it to be released in biodone */
4885 bp->b_cmd = mbp->b_cmd;
4886 bp->b_bio1.bio_done = nestiobuf_iodone;
4887 bp->b_data = (char *)mbp->b_data + offset;
4888 bp->b_resid = bp->b_bcount = size;
4889 bp->b_bufsize = bp->b_bcount;
4891 bp->b_bio1.bio_track = NULL;
4892 bp->b_bio1.bio_caller_info1.ptr = mbio;
4893 bp->b_bio1.bio_caller_info2.ptr = stats;
4898 DB_SHOW_COMMAND(buffer, db_show_buffer)
4901 struct buf *bp = (struct buf *)addr;
4904 db_printf("usage: show buffer <addr>\n");
4908 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4909 db_printf("b_cmd = %d\n", bp->b_cmd);
4910 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4911 "b_resid = %d\n, b_data = %p, "
4912 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4913 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4915 (long long)bp->b_bio2.bio_offset,
4916 (long long)(bp->b_bio2.bio_next ?
4917 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4918 if (bp->b_xio.xio_npages) {
4920 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4921 bp->b_xio.xio_npages);
4922 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4924 m = bp->b_xio.xio_pages[i];
4925 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4926 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4927 if ((i + 1) < bp->b_xio.xio_npages)