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 static long dirtybufspace; /* atomic */
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;
158 static struct thread *bufdaemon_td;
159 static struct thread *bufdaemonhw_td;
160 static u_int lowmempgallocs;
161 static u_int lowmempgfails;
164 * Sysctls for operational control of the buffer cache.
166 SYSCTL_LONG(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
167 "Number of dirty buffers to flush before bufdaemon becomes inactive");
168 SYSCTL_LONG(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
169 "High watermark used to trigger explicit flushing of dirty buffers");
170 SYSCTL_LONG(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
171 "Minimum amount of buffer space required for active I/O");
172 SYSCTL_LONG(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
173 "Maximum amount of buffer space to usable for active I/O");
174 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0,
175 "Page allocations done during periods of very low free memory");
176 SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0,
177 "Page allocations which failed during periods of very low free memory");
178 SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
179 "Recycle pages to active or inactive queue transition pt 0-64");
181 * Sysctls determining current state of the buffer cache.
183 SYSCTL_LONG(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
184 "Total number of buffers in buffer cache");
185 SYSCTL_LONG(_vfs, OID_AUTO, dirtykvaspace, CTLFLAG_RD, &dirtykvaspace, 0,
186 "KVA reserved by dirty buffers (all)");
187 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
188 "Pending bytes of dirty buffers (all)");
189 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
190 "Pending bytes of dirty buffers (heavy weight)");
191 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
192 "Pending number of dirty buffers");
193 SYSCTL_LONG(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
194 "Pending number of dirty buffers (heavy weight)");
195 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
196 "I/O bytes currently in progress due to asynchronous writes");
197 SYSCTL_LONG(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
198 "I/O buffers currently in progress due to asynchronous writes");
199 SYSCTL_LONG(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
200 "Hard limit on maximum amount of memory usable for buffer space");
201 SYSCTL_LONG(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
202 "Soft limit on maximum amount of memory usable for buffer space");
203 SYSCTL_LONG(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
204 "Minimum amount of memory to reserve for system buffer space");
205 SYSCTL_LONG(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
206 "Amount of memory available for buffers");
207 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
208 0, "Maximum amount of memory reserved for buffers using malloc");
209 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
210 "Amount of memory left for buffers using malloc-scheme");
211 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
212 "New buffer header acquisition requests");
213 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
214 0, "New buffer header acquisition restarts");
215 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
216 "Recover VM space in an emergency");
217 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
218 "Buffer acquisition restarts due to fragmented buffer map");
219 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
220 "Amount of time KVA space was deallocated in an arbitrary buffer");
221 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
222 "Amount of time buffer re-use operations were successful");
223 SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, "");
224 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
225 "sizeof(struct buf)");
227 char *buf_wmesg = BUF_WMESG;
229 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
230 #define VFS_BIO_NEED_UNUSED02 0x02
231 #define VFS_BIO_NEED_UNUSED04 0x04
232 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
237 * Called when buffer space is potentially available for recovery.
238 * getnewbuf() will block on this flag when it is unable to free
239 * sufficient buffer space. Buffer space becomes recoverable when
240 * bp's get placed back in the queues.
246 * If someone is waiting for BUF space, wake them up. Even
247 * though we haven't freed the kva space yet, the waiting
248 * process will be able to now.
251 int flags = needsbuffer;
253 if ((flags & VFS_BIO_NEED_BUFSPACE) == 0)
255 if (atomic_cmpset_int(&needsbuffer, flags,
256 flags & ~VFS_BIO_NEED_BUFSPACE)) {
257 wakeup(&needsbuffer);
267 * Accounting for I/O in progress.
271 runningbufwakeup(struct buf *bp)
277 if ((totalspace = bp->b_runningbufspace) != 0) {
278 atomic_add_long(&runningbufspace, -totalspace);
279 atomic_add_long(&runningbufcount, -1);
280 bp->b_runningbufspace = 0;
283 * see waitrunningbufspace() for limit test.
285 limit = hirunningspace * 3 / 6;
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.
625 struct bufpcpu *pcpu;
627 vm_offset_t bogus_offset;
632 /* next, make a null set of free lists */
633 for (i = 0; i < ncpus; ++i) {
635 spin_init(&pcpu->spin);
636 for (j = 0; j < BUFFER_QUEUES; j++)
637 TAILQ_INIT(&pcpu->bufqueues[j]);
640 /* finally, initialize each buffer header and stick on empty q */
644 for (n = 0; n < nbuf; n++) {
646 bzero(bp, sizeof *bp);
647 bp->b_flags = B_INVAL; /* we're just an empty header */
648 bp->b_cmd = BUF_CMD_DONE;
649 bp->b_qindex = BQUEUE_EMPTY;
652 xio_init(&bp->b_xio);
654 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
662 * maxbufspace is the absolute maximum amount of buffer space we are
663 * allowed to reserve in KVM and in real terms. The absolute maximum
664 * is nominally used by buf_daemon. hibufspace is the nominal maximum
665 * used by most other processes. The differential is required to
666 * ensure that buf_daemon is able to run when other processes might
667 * be blocked waiting for buffer space.
669 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
670 * this may result in KVM fragmentation which is not handled optimally
673 maxbufspace = nbuf * BKVASIZE;
674 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
675 lobufspace = hibufspace - MAXBSIZE;
677 lorunningspace = 512 * 1024;
678 /* hirunningspace -- see below */
681 * Limit the amount of malloc memory since it is wired permanently
682 * into the kernel space. Even though this is accounted for in
683 * the buffer allocation, we don't want the malloced region to grow
684 * uncontrolled. The malloc scheme improves memory utilization
685 * significantly on average (small) directories.
687 maxbufmallocspace = hibufspace / 20;
690 * Reduce the chance of a deadlock occuring by limiting the number
691 * of delayed-write dirty buffers we allow to stack up.
693 * We don't want too much actually queued to the device at once
694 * (XXX this needs to be per-mount!), because the buffers will
695 * wind up locked for a very long period of time while the I/O
698 hidirtybufspace = hibufspace / 2; /* dirty + running */
699 hirunningspace = hibufspace / 16; /* locked & queued to device */
700 if (hirunningspace < 1024 * 1024)
701 hirunningspace = 1024 * 1024;
707 lodirtybufspace = hidirtybufspace / 2;
710 * Maximum number of async ops initiated per buf_daemon loop. This is
711 * somewhat of a hack at the moment, we really need to limit ourselves
712 * based on the number of bytes of I/O in-transit that were initiated
716 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
717 vm_object_hold(&kernel_object);
718 bogus_page = vm_page_alloc(&kernel_object,
719 (bogus_offset >> PAGE_SHIFT),
721 vm_object_drop(&kernel_object);
722 vmstats.v_wire_count++;
727 * Initialize the embedded bio structures, typically used by
728 * deprecated code which tries to allocate its own struct bufs.
731 initbufbio(struct buf *bp)
733 bp->b_bio1.bio_buf = bp;
734 bp->b_bio1.bio_prev = NULL;
735 bp->b_bio1.bio_offset = NOOFFSET;
736 bp->b_bio1.bio_next = &bp->b_bio2;
737 bp->b_bio1.bio_done = NULL;
738 bp->b_bio1.bio_flags = 0;
740 bp->b_bio2.bio_buf = bp;
741 bp->b_bio2.bio_prev = &bp->b_bio1;
742 bp->b_bio2.bio_offset = NOOFFSET;
743 bp->b_bio2.bio_next = NULL;
744 bp->b_bio2.bio_done = NULL;
745 bp->b_bio2.bio_flags = 0;
751 * Reinitialize the embedded bio structures as well as any additional
752 * translation cache layers.
755 reinitbufbio(struct buf *bp)
759 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
760 bio->bio_done = NULL;
761 bio->bio_offset = NOOFFSET;
766 * Undo the effects of an initbufbio().
769 uninitbufbio(struct buf *bp)
776 * Push another BIO layer onto an existing BIO and return it. The new
777 * BIO layer may already exist, holding cached translation data.
780 push_bio(struct bio *bio)
784 if ((nbio = bio->bio_next) == NULL) {
785 int index = bio - &bio->bio_buf->b_bio_array[0];
786 if (index >= NBUF_BIO - 1) {
787 panic("push_bio: too many layers bp %p",
790 nbio = &bio->bio_buf->b_bio_array[index + 1];
791 bio->bio_next = nbio;
792 nbio->bio_prev = bio;
793 nbio->bio_buf = bio->bio_buf;
794 nbio->bio_offset = NOOFFSET;
795 nbio->bio_done = NULL;
796 nbio->bio_next = NULL;
798 KKASSERT(nbio->bio_done == NULL);
803 * Pop a BIO translation layer, returning the previous layer. The
804 * must have been previously pushed.
807 pop_bio(struct bio *bio)
809 return(bio->bio_prev);
813 clearbiocache(struct bio *bio)
816 bio->bio_offset = NOOFFSET;
824 * Free the KVA allocation for buffer 'bp'.
826 * Must be called from a critical section as this is the only locking for
829 * Since this call frees up buffer space, we call bufspacewakeup().
832 bfreekva(struct buf *bp)
838 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
839 vm_map_lock(&buffer_map);
840 bufspace -= bp->b_kvasize;
841 vm_map_delete(&buffer_map,
842 (vm_offset_t) bp->b_kvabase,
843 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
846 vm_map_unlock(&buffer_map);
847 vm_map_entry_release(count);
849 bp->b_kvabase = NULL;
855 * Remove the buffer from the appropriate free list.
856 * (caller must be locked)
859 _bremfree(struct buf *bp)
861 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu];
863 if (bp->b_qindex != BQUEUE_NONE) {
864 KASSERT(BUF_REFCNTNB(bp) == 1,
865 ("bremfree: bp %p not locked",bp));
866 TAILQ_REMOVE(&pcpu->bufqueues[bp->b_qindex], bp, b_freelist);
867 bp->b_qindex = BQUEUE_NONE;
869 if (BUF_REFCNTNB(bp) <= 1)
870 panic("bremfree: removing a buffer not on a queue");
875 * bremfree() - must be called with a locked buffer
878 bremfree(struct buf *bp)
880 struct bufpcpu *pcpu = &bufpcpu[bp->b_qcpu];
882 spin_lock(&pcpu->spin);
884 spin_unlock(&pcpu->spin);
888 * bremfree_locked - must be called with pcpu->spin locked
891 bremfree_locked(struct buf *bp)
897 * This version of bread issues any required I/O asyncnronously and
898 * makes a callback on completion.
900 * The callback must check whether BIO_DONE is set in the bio and issue
901 * the bpdone(bp, 0) if it isn't. The callback is responsible for clearing
902 * BIO_DONE and disposing of the I/O (bqrelse()ing it).
905 breadcb(struct vnode *vp, off_t loffset, int size,
906 void (*func)(struct bio *), void *arg)
910 bp = getblk(vp, loffset, size, 0, 0);
912 /* if not found in cache, do some I/O */
913 if ((bp->b_flags & B_CACHE) == 0) {
914 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
915 bp->b_cmd = BUF_CMD_READ;
916 bp->b_bio1.bio_done = func;
917 bp->b_bio1.bio_caller_info1.ptr = arg;
918 vfs_busy_pages(vp, bp);
920 vn_strategy(vp, &bp->b_bio1);
923 * Since we are issuing the callback synchronously it cannot
924 * race the BIO_DONE, so no need for atomic ops here.
926 /*bp->b_bio1.bio_done = func;*/
927 bp->b_bio1.bio_caller_info1.ptr = arg;
928 bp->b_bio1.bio_flags |= BIO_DONE;
936 * breadnx() - Terminal function for bread() and breadn().
938 * This function will start asynchronous I/O on read-ahead blocks as well
939 * as satisfy the primary request.
941 * We must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE is
942 * set, the buffer is valid and we do not have to do anything.
945 breadnx(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
946 int *rabsize, int cnt, struct buf **bpp)
948 struct buf *bp, *rabp;
950 int rv = 0, readwait = 0;
955 *bpp = bp = getblk(vp, loffset, size, 0, 0);
957 /* if not found in cache, do some I/O */
958 if ((bp->b_flags & B_CACHE) == 0) {
959 bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
960 bp->b_cmd = BUF_CMD_READ;
961 bp->b_bio1.bio_done = biodone_sync;
962 bp->b_bio1.bio_flags |= BIO_SYNC;
963 vfs_busy_pages(vp, bp);
964 vn_strategy(vp, &bp->b_bio1);
968 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
969 if (inmem(vp, *raoffset))
971 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
973 if ((rabp->b_flags & B_CACHE) == 0) {
974 rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
975 rabp->b_cmd = BUF_CMD_READ;
976 vfs_busy_pages(vp, rabp);
978 vn_strategy(vp, &rabp->b_bio1);
984 rv = biowait(&bp->b_bio1, "biord");
991 * Synchronous write, waits for completion.
993 * Write, release buffer on completion. (Done by iodone
994 * if async). Do not bother writing anything if the buffer
997 * Note that we set B_CACHE here, indicating that buffer is
998 * fully valid and thus cacheable. This is true even of NFS
999 * now so we set it generally. This could be set either here
1000 * or in biodone() since the I/O is synchronous. We put it
1004 bwrite(struct buf *bp)
1008 if (bp->b_flags & B_INVAL) {
1012 if (BUF_REFCNTNB(bp) == 0)
1013 panic("bwrite: buffer is not busy???");
1015 /* Mark the buffer clean */
1018 bp->b_flags &= ~(B_ERROR | B_EINTR);
1019 bp->b_flags |= B_CACHE;
1020 bp->b_cmd = BUF_CMD_WRITE;
1021 bp->b_bio1.bio_done = biodone_sync;
1022 bp->b_bio1.bio_flags |= BIO_SYNC;
1023 vfs_busy_pages(bp->b_vp, bp);
1026 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1027 * valid for vnode-backed buffers.
1029 bsetrunningbufspace(bp, bp->b_bufsize);
1030 vn_strategy(bp->b_vp, &bp->b_bio1);
1031 error = biowait(&bp->b_bio1, "biows");
1040 * Asynchronous write. Start output on a buffer, but do not wait for
1041 * it to complete. The buffer is released when the output completes.
1043 * bwrite() ( or the VOP routine anyway ) is responsible for handling
1044 * B_INVAL buffers. Not us.
1047 bawrite(struct buf *bp)
1049 if (bp->b_flags & B_INVAL) {
1053 if (BUF_REFCNTNB(bp) == 0)
1054 panic("bwrite: buffer is not busy???");
1056 /* Mark the buffer clean */
1059 bp->b_flags &= ~(B_ERROR | B_EINTR);
1060 bp->b_flags |= B_CACHE;
1061 bp->b_cmd = BUF_CMD_WRITE;
1062 KKASSERT(bp->b_bio1.bio_done == NULL);
1063 vfs_busy_pages(bp->b_vp, bp);
1066 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
1067 * valid for vnode-backed buffers.
1069 bsetrunningbufspace(bp, bp->b_bufsize);
1071 vn_strategy(bp->b_vp, &bp->b_bio1);
1077 * Ordered write. Start output on a buffer, and flag it so that the
1078 * device will write it in the order it was queued. The buffer is
1079 * released when the output completes. bwrite() ( or the VOP routine
1080 * anyway ) is responsible for handling B_INVAL buffers.
1083 bowrite(struct buf *bp)
1085 bp->b_flags |= B_ORDERED;
1093 * Delayed write. (Buffer is marked dirty). Do not bother writing
1094 * anything if the buffer is marked invalid.
1096 * Note that since the buffer must be completely valid, we can safely
1097 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
1098 * biodone() in order to prevent getblk from writing the buffer
1099 * out synchronously.
1102 bdwrite(struct buf *bp)
1104 if (BUF_REFCNTNB(bp) == 0)
1105 panic("bdwrite: buffer is not busy");
1107 if (bp->b_flags & B_INVAL) {
1113 if (dsched_is_clear_buf_priv(bp))
1117 * Set B_CACHE, indicating that the buffer is fully valid. This is
1118 * true even of NFS now.
1120 bp->b_flags |= B_CACHE;
1123 * This bmap keeps the system from needing to do the bmap later,
1124 * perhaps when the system is attempting to do a sync. Since it
1125 * is likely that the indirect block -- or whatever other datastructure
1126 * that the filesystem needs is still in memory now, it is a good
1127 * thing to do this. Note also, that if the pageout daemon is
1128 * requesting a sync -- there might not be enough memory to do
1129 * the bmap then... So, this is important to do.
1131 if (bp->b_bio2.bio_offset == NOOFFSET) {
1132 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1133 NULL, NULL, BUF_CMD_WRITE);
1137 * Because the underlying pages may still be mapped and
1138 * writable trying to set the dirty buffer (b_dirtyoff/end)
1139 * range here will be inaccurate.
1141 * However, we must still clean the pages to satisfy the
1142 * vnode_pager and pageout daemon, so theythink the pages
1143 * have been "cleaned". What has really occured is that
1144 * they've been earmarked for later writing by the buffer
1147 * So we get the b_dirtyoff/end update but will not actually
1148 * depend on it (NFS that is) until the pages are busied for
1151 vfs_clean_pages(bp);
1155 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
1156 * due to the softdep code.
1161 * Fake write - return pages to VM system as dirty, leave the buffer clean.
1162 * This is used by tmpfs.
1164 * It is important for any VFS using this routine to NOT use it for
1165 * IO_SYNC or IO_ASYNC operations which occur when the system really
1166 * wants to flush VM pages to backing store.
1169 buwrite(struct buf *bp)
1175 * Only works for VMIO buffers. If the buffer is already
1176 * marked for delayed-write we can't avoid the bdwrite().
1178 if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) {
1184 * Mark as needing a commit.
1186 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1187 m = bp->b_xio.xio_pages[i];
1188 vm_page_need_commit(m);
1196 * Turn buffer into delayed write request by marking it B_DELWRI.
1197 * B_RELBUF and B_NOCACHE must be cleared.
1199 * We reassign the buffer to itself to properly update it in the
1200 * dirty/clean lists.
1202 * Must be called from a critical section.
1203 * The buffer must be on BQUEUE_NONE.
1206 bdirty(struct buf *bp)
1208 KASSERT(bp->b_qindex == BQUEUE_NONE,
1209 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
1210 if (bp->b_flags & B_NOCACHE) {
1211 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
1212 bp->b_flags &= ~B_NOCACHE;
1214 if (bp->b_flags & B_INVAL) {
1215 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
1217 bp->b_flags &= ~B_RELBUF;
1219 if ((bp->b_flags & B_DELWRI) == 0) {
1220 lwkt_gettoken(&bp->b_vp->v_token);
1221 bp->b_flags |= B_DELWRI;
1223 lwkt_reltoken(&bp->b_vp->v_token);
1225 atomic_add_long(&dirtybufcount, 1);
1226 atomic_add_long(&dirtykvaspace, bp->b_kvasize);
1227 atomic_add_long(&dirtybufspace, bp->b_bufsize);
1228 if (bp->b_flags & B_HEAVY) {
1229 atomic_add_long(&dirtybufcounthw, 1);
1230 atomic_add_long(&dirtybufspacehw, bp->b_bufsize);
1237 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
1238 * needs to be flushed with a different buf_daemon thread to avoid
1239 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
1242 bheavy(struct buf *bp)
1244 if ((bp->b_flags & B_HEAVY) == 0) {
1245 bp->b_flags |= B_HEAVY;
1246 if (bp->b_flags & B_DELWRI) {
1247 atomic_add_long(&dirtybufcounthw, 1);
1248 atomic_add_long(&dirtybufspacehw, bp->b_bufsize);
1256 * Clear B_DELWRI for buffer.
1258 * Must be called from a critical section.
1260 * The buffer is typically on BQUEUE_NONE but there is one case in
1261 * brelse() that calls this function after placing the buffer on
1262 * a different queue.
1265 bundirty(struct buf *bp)
1267 if (bp->b_flags & B_DELWRI) {
1268 lwkt_gettoken(&bp->b_vp->v_token);
1269 bp->b_flags &= ~B_DELWRI;
1271 lwkt_reltoken(&bp->b_vp->v_token);
1273 atomic_add_long(&dirtybufcount, -1);
1274 atomic_add_long(&dirtykvaspace, -bp->b_kvasize);
1275 atomic_add_long(&dirtybufspace, -bp->b_bufsize);
1276 if (bp->b_flags & B_HEAVY) {
1277 atomic_add_long(&dirtybufcounthw, -1);
1278 atomic_add_long(&dirtybufspacehw, -bp->b_bufsize);
1280 bd_signal(bp->b_bufsize);
1283 * Since it is now being written, we can clear its deferred write flag.
1285 bp->b_flags &= ~B_DEFERRED;
1289 * Set the b_runningbufspace field, used to track how much I/O is
1290 * in progress at any given moment.
1293 bsetrunningbufspace(struct buf *bp, int bytes)
1295 bp->b_runningbufspace = bytes;
1297 atomic_add_long(&runningbufspace, bytes);
1298 atomic_add_long(&runningbufcount, 1);
1305 * Release a busy buffer and, if requested, free its resources. The
1306 * buffer will be stashed in the appropriate bufqueue[] allowing it
1307 * to be accessed later as a cache entity or reused for other purposes.
1310 brelse(struct buf *bp)
1312 struct bufpcpu *pcpu;
1314 int saved_flags = bp->b_flags;
1317 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1318 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1321 * If B_NOCACHE is set we are being asked to destroy the buffer and
1322 * its backing store. Clear B_DELWRI.
1324 * B_NOCACHE is set in two cases: (1) when the caller really wants
1325 * to destroy the buffer and backing store and (2) when the caller
1326 * wants to destroy the buffer and backing store after a write
1329 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1333 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1335 * A re-dirtied buffer is only subject to destruction
1336 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1338 /* leave buffer intact */
1339 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1340 (bp->b_bufsize <= 0)) {
1342 * Either a failed read or we were asked to free or not
1343 * cache the buffer. This path is reached with B_DELWRI
1344 * set only if B_INVAL is already set. B_NOCACHE governs
1345 * backing store destruction.
1347 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1348 * buffer cannot be immediately freed.
1350 bp->b_flags |= B_INVAL;
1351 if (LIST_FIRST(&bp->b_dep) != NULL)
1353 if (bp->b_flags & B_DELWRI) {
1354 atomic_add_long(&dirtybufcount, -1);
1355 atomic_add_long(&dirtykvaspace, -bp->b_kvasize);
1356 atomic_add_long(&dirtybufspace, -bp->b_bufsize);
1357 if (bp->b_flags & B_HEAVY) {
1358 atomic_add_long(&dirtybufcounthw, -1);
1359 atomic_add_long(&dirtybufspacehw,
1362 bd_signal(bp->b_bufsize);
1364 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1368 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set,
1369 * or if b_refs is non-zero.
1371 * If vfs_vmio_release() is called with either bit set, the
1372 * underlying pages may wind up getting freed causing a previous
1373 * write (bdwrite()) to get 'lost' because pages associated with
1374 * a B_DELWRI bp are marked clean. Pages associated with a
1375 * B_LOCKED buffer may be mapped by the filesystem.
1377 * If we want to release the buffer ourselves (rather then the
1378 * originator asking us to release it), give the originator a
1379 * chance to countermand the release by setting B_LOCKED.
1381 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1382 * if B_DELWRI is set.
1384 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1385 * on pages to return pages to the VM page queues.
1387 if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) {
1388 bp->b_flags &= ~B_RELBUF;
1389 } else if (vm_page_count_min(0)) {
1390 if (LIST_FIRST(&bp->b_dep) != NULL)
1391 buf_deallocate(bp); /* can set B_LOCKED */
1392 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1393 bp->b_flags &= ~B_RELBUF;
1395 bp->b_flags |= B_RELBUF;
1399 * Make sure b_cmd is clear. It may have already been cleared by
1402 * At this point destroying the buffer is governed by the B_INVAL
1403 * or B_RELBUF flags.
1405 bp->b_cmd = BUF_CMD_DONE;
1406 dsched_exit_buf(bp);
1409 * VMIO buffer rundown. Make sure the VM page array is restored
1410 * after an I/O may have replaces some of the pages with bogus pages
1411 * in order to not destroy dirty pages in a fill-in read.
1413 * Note that due to the code above, if a buffer is marked B_DELWRI
1414 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1415 * B_INVAL may still be set, however.
1417 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1418 * but not the backing store. B_NOCACHE will destroy the backing
1421 * Note that dirty NFS buffers contain byte-granular write ranges
1422 * and should not be destroyed w/ B_INVAL even if the backing store
1425 if (bp->b_flags & B_VMIO) {
1427 * Rundown for VMIO buffers which are not dirty NFS buffers.
1439 * Get the base offset and length of the buffer. Note that
1440 * in the VMIO case if the buffer block size is not
1441 * page-aligned then b_data pointer may not be page-aligned.
1442 * But our b_xio.xio_pages array *IS* page aligned.
1444 * block sizes less then DEV_BSIZE (usually 512) are not
1445 * supported due to the page granularity bits (m->valid,
1446 * m->dirty, etc...).
1448 * See man buf(9) for more information
1451 resid = bp->b_bufsize;
1452 foff = bp->b_loffset;
1454 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1455 m = bp->b_xio.xio_pages[i];
1456 vm_page_flag_clear(m, PG_ZERO);
1458 * If we hit a bogus page, fixup *all* of them
1459 * now. Note that we left these pages wired
1460 * when we removed them so they had better exist,
1461 * and they cannot be ripped out from under us so
1462 * no critical section protection is necessary.
1464 if (m == bogus_page) {
1466 poff = OFF_TO_IDX(bp->b_loffset);
1468 vm_object_hold(obj);
1469 for (j = i; j < bp->b_xio.xio_npages; j++) {
1472 mtmp = bp->b_xio.xio_pages[j];
1473 if (mtmp == bogus_page) {
1474 mtmp = vm_page_lookup(obj, poff + j);
1476 panic("brelse: page missing");
1478 bp->b_xio.xio_pages[j] = mtmp;
1481 bp->b_flags &= ~B_HASBOGUS;
1482 vm_object_drop(obj);
1484 if ((bp->b_flags & B_INVAL) == 0) {
1485 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1486 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1488 m = bp->b_xio.xio_pages[i];
1492 * Invalidate the backing store if B_NOCACHE is set
1493 * (e.g. used with vinvalbuf()). If this is NFS
1494 * we impose a requirement that the block size be
1495 * a multiple of PAGE_SIZE and create a temporary
1496 * hack to basically invalidate the whole page. The
1497 * problem is that NFS uses really odd buffer sizes
1498 * especially when tracking piecemeal writes and
1499 * it also vinvalbuf()'s a lot, which would result
1500 * in only partial page validation and invalidation
1501 * here. If the file page is mmap()'d, however,
1502 * all the valid bits get set so after we invalidate
1503 * here we would end up with weird m->valid values
1504 * like 0xfc. nfs_getpages() can't handle this so
1505 * we clear all the valid bits for the NFS case
1506 * instead of just some of them.
1508 * The real bug is the VM system having to set m->valid
1509 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1510 * itself is an artifact of the whole 512-byte
1511 * granular mess that exists to support odd block
1512 * sizes and UFS meta-data block sizes (e.g. 6144).
1513 * A complete rewrite is required.
1517 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1518 int poffset = foff & PAGE_MASK;
1521 presid = PAGE_SIZE - poffset;
1522 if (bp->b_vp->v_tag == VT_NFS &&
1523 bp->b_vp->v_type == VREG) {
1525 } else if (presid > resid) {
1528 KASSERT(presid >= 0, ("brelse: extra page"));
1529 vm_page_set_invalid(m, poffset, presid);
1532 * Also make sure any swap cache is removed
1533 * as it is now stale (HAMMER in particular
1534 * uses B_NOCACHE to deal with buffer
1537 swap_pager_unswapped(m);
1539 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1540 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1542 if (bp->b_flags & (B_INVAL | B_RELBUF))
1543 vfs_vmio_release(bp);
1546 * Rundown for non-VMIO buffers.
1548 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1551 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1557 if (bp->b_qindex != BQUEUE_NONE)
1558 panic("brelse: free buffer onto another queue???");
1559 if (BUF_REFCNTNB(bp) > 1) {
1560 /* Temporary panic to verify exclusive locking */
1561 /* This panic goes away when we allow shared refs */
1562 panic("brelse: multiple refs");
1568 * Figure out the correct queue to place the cleaned up buffer on.
1569 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1570 * disassociated from their vnode.
1572 * Return the buffer to its original pcpu area
1574 pcpu = &bufpcpu[bp->b_qcpu];
1575 spin_lock(&pcpu->spin);
1577 if (bp->b_flags & B_LOCKED) {
1579 * Buffers that are locked are placed in the locked queue
1580 * immediately, regardless of their state.
1582 bp->b_qindex = BQUEUE_LOCKED;
1583 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1585 } else if (bp->b_bufsize == 0) {
1587 * Buffers with no memory. Due to conditionals near the top
1588 * of brelse() such buffers should probably already be
1589 * marked B_INVAL and disassociated from their vnode.
1591 bp->b_flags |= B_INVAL;
1592 KASSERT(bp->b_vp == NULL,
1593 ("bp1 %p flags %08x/%08x vnode %p "
1594 "unexpectededly still associated!",
1595 bp, saved_flags, bp->b_flags, bp->b_vp));
1596 KKASSERT((bp->b_flags & B_HASHED) == 0);
1597 if (bp->b_kvasize) {
1598 bp->b_qindex = BQUEUE_EMPTYKVA;
1600 bp->b_qindex = BQUEUE_EMPTY;
1602 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex],
1604 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1606 * Buffers with junk contents. Again these buffers had better
1607 * already be disassociated from their vnode.
1609 KASSERT(bp->b_vp == NULL,
1610 ("bp2 %p flags %08x/%08x vnode %p unexpectededly "
1611 "still associated!",
1612 bp, saved_flags, bp->b_flags, bp->b_vp));
1613 KKASSERT((bp->b_flags & B_HASHED) == 0);
1614 bp->b_flags |= B_INVAL;
1615 bp->b_qindex = BQUEUE_CLEAN;
1616 TAILQ_INSERT_HEAD(&pcpu->bufqueues[bp->b_qindex],
1620 * Remaining buffers. These buffers are still associated with
1623 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1625 bp->b_qindex = BQUEUE_DIRTY;
1626 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1629 case B_DELWRI | B_HEAVY:
1630 bp->b_qindex = BQUEUE_DIRTY_HW;
1631 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1636 * NOTE: Buffers are always placed at the end of the
1637 * queue. If B_AGE is not set the buffer will cycle
1638 * through the queue twice.
1640 bp->b_qindex = BQUEUE_CLEAN;
1641 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1646 spin_unlock(&pcpu->spin);
1649 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1650 * on the correct queue but we have not yet unlocked it.
1652 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1656 * The bp is on an appropriate queue unless locked. If it is not
1657 * locked or dirty we can wakeup threads waiting for buffer space.
1659 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1660 * if B_INVAL is set ).
1662 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1666 * Something we can maybe free or reuse
1668 if (bp->b_bufsize || bp->b_kvasize)
1672 * Clean up temporary flags and unlock the buffer.
1674 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
1681 * Release a buffer back to the appropriate queue but do not try to free
1682 * it. The buffer is expected to be used again soon.
1684 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1685 * biodone() to requeue an async I/O on completion. It is also used when
1686 * known good buffers need to be requeued but we think we may need the data
1689 * XXX we should be able to leave the B_RELBUF hint set on completion.
1692 bqrelse(struct buf *bp)
1694 struct bufpcpu *pcpu;
1696 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)),
1697 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1699 if (bp->b_qindex != BQUEUE_NONE)
1700 panic("bqrelse: free buffer onto another queue???");
1701 if (BUF_REFCNTNB(bp) > 1) {
1702 /* do not release to free list */
1703 panic("bqrelse: multiple refs");
1707 buf_act_advance(bp);
1709 pcpu = &bufpcpu[bp->b_qcpu];
1710 spin_lock(&pcpu->spin);
1712 if (bp->b_flags & B_LOCKED) {
1714 * Locked buffers are released to the locked queue. However,
1715 * if the buffer is dirty it will first go into the dirty
1716 * queue and later on after the I/O completes successfully it
1717 * will be released to the locked queue.
1719 bp->b_qindex = BQUEUE_LOCKED;
1720 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1722 } else if (bp->b_flags & B_DELWRI) {
1723 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1724 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1725 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1727 } else if (vm_page_count_min(0)) {
1729 * We are too low on memory, we have to try to free the
1730 * buffer (most importantly: the wired pages making up its
1731 * backing store) *now*.
1733 spin_unlock(&pcpu->spin);
1737 bp->b_qindex = BQUEUE_CLEAN;
1738 TAILQ_INSERT_TAIL(&pcpu->bufqueues[bp->b_qindex],
1741 spin_unlock(&pcpu->spin);
1744 * We have now placed the buffer on the proper queue, but have yet
1747 if ((bp->b_flags & B_LOCKED) == 0 &&
1748 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1753 * Something we can maybe free or reuse.
1755 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1759 * Final cleanup and unlock. Clear bits that are only used while a
1760 * buffer is actively locked.
1762 bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
1763 dsched_exit_buf(bp);
1768 * Hold a buffer, preventing it from being reused. This will prevent
1769 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL
1770 * operations. If a B_INVAL operation occurs the buffer will remain held
1771 * but the underlying pages may get ripped out.
1773 * These functions are typically used in VOP_READ/VOP_WRITE functions
1774 * to hold a buffer during a copyin or copyout, preventing deadlocks
1775 * or recursive lock panics when read()/write() is used over mmap()'d
1778 * NOTE: bqhold() requires that the buffer be locked at the time of the
1779 * hold. bqdrop() has no requirements other than the buffer having
1780 * previously been held.
1783 bqhold(struct buf *bp)
1785 atomic_add_int(&bp->b_refs, 1);
1789 bqdrop(struct buf *bp)
1791 KKASSERT(bp->b_refs > 0);
1792 atomic_add_int(&bp->b_refs, -1);
1796 * Return backing pages held by the buffer 'bp' back to the VM system.
1797 * This routine is called when the bp is invalidated, released, or
1800 * The KVA mapping (b_data) for the underlying pages is removed by
1803 * WARNING! This routine is integral to the low memory critical path
1804 * when a buffer is B_RELBUF'd. If the system has a severe page
1805 * deficit we need to get the page(s) onto the PQ_FREE or PQ_CACHE
1806 * queues so they can be reused in the current pageout daemon
1810 vfs_vmio_release(struct buf *bp)
1815 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1816 m = bp->b_xio.xio_pages[i];
1817 bp->b_xio.xio_pages[i] = NULL;
1820 * We need to own the page in order to safely unwire it.
1822 vm_page_busy_wait(m, FALSE, "vmiopg");
1825 * The VFS is telling us this is not a meta-data buffer
1826 * even if it is backed by a block device.
1828 if (bp->b_flags & B_NOTMETA)
1829 vm_page_flag_set(m, PG_NOTMETA);
1832 * This is a very important bit of code. We try to track
1833 * VM page use whether the pages are wired into the buffer
1834 * cache or not. While wired into the buffer cache the
1835 * bp tracks the act_count.
1837 * We can choose to place unwired pages on the inactive
1838 * queue (0) or active queue (1). If we place too many
1839 * on the active queue the queue will cycle the act_count
1840 * on pages we'd like to keep, just from single-use pages
1841 * (such as when doing a tar-up or file scan).
1843 if (bp->b_act_count < vm_cycle_point)
1844 vm_page_unwire(m, 0);
1846 vm_page_unwire(m, 1);
1849 * If the wire_count has dropped to 0 we may need to take
1850 * further action before unbusying the page.
1852 * WARNING: vm_page_try_*() also checks PG_NEED_COMMIT for us.
1854 if (m->wire_count == 0) {
1855 vm_page_flag_clear(m, PG_ZERO);
1857 if (bp->b_flags & B_DIRECT) {
1859 * Attempt to free the page if B_DIRECT is
1860 * set, the caller does not desire the page
1864 vm_page_try_to_free(m);
1865 } else if ((bp->b_flags & B_NOTMETA) ||
1866 vm_page_count_min(0)) {
1868 * Attempt to move the page to PQ_CACHE
1869 * if B_NOTMETA is set. This flag is set
1870 * by HAMMER to remove one of the two pages
1871 * present when double buffering is enabled.
1873 * Attempt to move the page to PQ_CACHE
1874 * If we have a severe page deficit. This
1875 * will cause buffer cache operations related
1876 * to pageouts to recycle the related pages
1877 * in order to avoid a low memory deadlock.
1879 m->act_count = bp->b_act_count;
1881 vm_page_try_to_cache(m);
1884 * Nominal case, leave the page on the
1885 * queue the original unwiring placed it on
1886 * (active or inactive).
1888 m->act_count = bp->b_act_count;
1896 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
1897 bp->b_xio.xio_npages);
1898 if (bp->b_bufsize) {
1902 bp->b_xio.xio_npages = 0;
1903 bp->b_flags &= ~B_VMIO;
1904 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1910 * Find and initialize a new buffer header, freeing up existing buffers
1911 * in the bufqueues as necessary. The new buffer is returned locked.
1913 * Important: B_INVAL is not set. If the caller wishes to throw the
1914 * buffer away, the caller must set B_INVAL prior to calling brelse().
1917 * We have insufficient buffer headers
1918 * We have insufficient buffer space
1919 * buffer_map is too fragmented ( space reservation fails )
1920 * If we have to flush dirty buffers ( but we try to avoid this )
1922 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1923 * Instead we ask the buf daemon to do it for us. We attempt to
1924 * avoid piecemeal wakeups of the pageout daemon.
1927 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1929 struct bufpcpu *pcpu;
1935 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1936 static int flushingbufs;
1939 * We can't afford to block since we might be holding a vnode lock,
1940 * which may prevent system daemons from running. We deal with
1941 * low-memory situations by proactively returning memory and running
1942 * async I/O rather then sync I/O.
1946 --getnewbufrestarts;
1947 nqcpu = mycpu->gd_cpuid;
1949 ++getnewbufrestarts;
1952 * Setup for scan. If we do not have enough free buffers,
1953 * we setup a degenerate case that immediately fails. Note
1954 * that if we are specially marked process, we are allowed to
1955 * dip into our reserves.
1957 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1959 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1960 * However, there are a number of cases (defragging, reusing, ...)
1961 * where we cannot backup.
1963 pcpu = &bufpcpu[nqcpu];
1964 nqindex = BQUEUE_EMPTYKVA;
1965 spin_lock(&pcpu->spin);
1967 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTYKVA]);
1971 * If no EMPTYKVA buffers and we are either
1972 * defragging or reusing, locate a CLEAN buffer
1973 * to free or reuse. If bufspace useage is low
1974 * skip this step so we can allocate a new buffer.
1976 if (defrag || bufspace >= lobufspace) {
1977 nqindex = BQUEUE_CLEAN;
1978 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN]);
1982 * If we could not find or were not allowed to reuse a
1983 * CLEAN buffer, check to see if it is ok to use an EMPTY
1984 * buffer. We can only use an EMPTY buffer if allocating
1985 * its KVA would not otherwise run us out of buffer space.
1987 if (nbp == NULL && defrag == 0 &&
1988 bufspace + maxsize < hibufspace) {
1989 nqindex = BQUEUE_EMPTY;
1990 nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTY]);
1995 * Run scan, possibly freeing data and/or kva mappings on the fly
1998 * WARNING! spin is held!
2000 while ((bp = nbp) != NULL) {
2001 int qindex = nqindex;
2003 nbp = TAILQ_NEXT(bp, b_freelist);
2006 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
2007 * cycles through the queue twice before being selected.
2009 if (qindex == BQUEUE_CLEAN &&
2010 (bp->b_flags & B_AGE) == 0 && nbp) {
2011 bp->b_flags |= B_AGE;
2012 TAILQ_REMOVE(&pcpu->bufqueues[qindex],
2014 TAILQ_INSERT_TAIL(&pcpu->bufqueues[qindex],
2020 * Calculate next bp ( we can only use it if we do not block
2021 * or do other fancy things ).
2026 nqindex = BQUEUE_EMPTYKVA;
2027 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_EMPTYKVA])))
2030 case BQUEUE_EMPTYKVA:
2031 nqindex = BQUEUE_CLEAN;
2032 if ((nbp = TAILQ_FIRST(&pcpu->bufqueues[BQUEUE_CLEAN])))
2046 KASSERT(bp->b_qindex == qindex,
2047 ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
2050 * Note: we no longer distinguish between VMIO and non-VMIO
2053 KASSERT((bp->b_flags & B_DELWRI) == 0,
2054 ("delwri buffer %p found in queue %d", bp, qindex));
2057 * Do not try to reuse a buffer with a non-zero b_refs.
2058 * This is an unsynchronized test. A synchronized test
2059 * is also performed after we lock the buffer.
2065 * If we are defragging then we need a buffer with
2066 * b_kvasize != 0. XXX this situation should no longer
2067 * occur, if defrag is non-zero the buffer's b_kvasize
2068 * should also be non-zero at this point. XXX
2070 if (defrag && bp->b_kvasize == 0) {
2071 kprintf("Warning: defrag empty buffer %p\n", bp);
2076 * Start freeing the bp. This is somewhat involved. nbp
2077 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
2078 * on the clean list must be disassociated from their
2079 * current vnode. Buffers on the empty[kva] lists have
2080 * already been disassociated.
2082 * b_refs is checked after locking along with queue changes.
2083 * We must check here to deal with zero->nonzero transitions
2084 * made by the owner of the buffer lock, which is used by
2085 * VFS's to hold the buffer while issuing an unlocked
2086 * uiomove()s. We cannot invalidate the buffer's pages
2087 * for this case. Once we successfully lock a buffer the
2088 * only 0->1 transitions of b_refs will occur via findblk().
2090 * We must also check for queue changes after successful
2091 * locking as the current lock holder may dispose of the
2092 * buffer and change its queue.
2094 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
2095 spin_unlock(&pcpu->spin);
2096 tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100);
2099 if (bp->b_qindex != qindex || bp->b_refs) {
2100 spin_unlock(&pcpu->spin);
2104 bremfree_locked(bp);
2105 spin_unlock(&pcpu->spin);
2108 * Dependancies must be handled before we disassociate the
2111 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
2112 * be immediately disassociated. HAMMER then becomes
2113 * responsible for releasing the buffer.
2115 * NOTE: spin is UNLOCKED now.
2117 if (LIST_FIRST(&bp->b_dep) != NULL) {
2119 if (bp->b_flags & B_LOCKED) {
2123 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2126 if (qindex == BQUEUE_CLEAN) {
2127 if (bp->b_flags & B_VMIO)
2128 vfs_vmio_release(bp);
2134 * NOTE: nbp is now entirely invalid. We can only restart
2135 * the scan from this point on.
2137 * Get the rest of the buffer freed up. b_kva* is still
2138 * valid after this operation.
2140 KASSERT(bp->b_vp == NULL,
2141 ("bp3 %p flags %08x vnode %p qindex %d "
2142 "unexpectededly still associated!",
2143 bp, bp->b_flags, bp->b_vp, qindex));
2144 KKASSERT((bp->b_flags & B_HASHED) == 0);
2147 * critical section protection is not required when
2148 * scrapping a buffer's contents because it is already
2154 bp->b_flags = B_BNOCLIP;
2155 bp->b_cmd = BUF_CMD_DONE;
2160 bp->b_xio.xio_npages = 0;
2161 bp->b_dirtyoff = bp->b_dirtyend = 0;
2162 bp->b_act_count = ACT_INIT;
2164 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2166 if (blkflags & GETBLK_BHEAVY)
2167 bp->b_flags |= B_HEAVY;
2170 * If we are defragging then free the buffer.
2173 bp->b_flags |= B_INVAL;
2181 * If we are overcomitted then recover the buffer and its
2182 * KVM space. This occurs in rare situations when multiple
2183 * processes are blocked in getnewbuf() or allocbuf().
2185 * (We don't have to recover the KVM space if
2186 * BKVASIZE == MAXBSIZE)
2188 if (bufspace >= hibufspace)
2190 if (flushingbufs && bp->b_kvasize != 0) {
2191 bp->b_flags |= B_INVAL;
2192 if (BKVASIZE != MAXBSIZE)
2197 if (bufspace < lobufspace)
2201 * b_refs can transition to a non-zero value while we hold
2202 * the buffer locked due to a findblk(). Our brelvp() above
2203 * interlocked any future possible transitions due to
2206 * If we find b_refs to be non-zero we can destroy the
2207 * buffer's contents but we cannot yet reuse the buffer.
2210 bp->b_flags |= B_INVAL;
2211 if (BKVASIZE != MAXBSIZE)
2217 /* NOT REACHED, spin not held */
2221 * If we exhausted our list, iterate other cpus. If that fails,
2222 * sleep as appropriate. We may have to wakeup various daemons
2223 * and write out some dirty buffers.
2225 * Generally we are sleeping due to insufficient buffer space.
2227 * NOTE: spin is held if bp is NULL, else it is not held.
2233 spin_unlock(&pcpu->spin);
2235 nqcpu = (nqcpu + 1) % ncpus;
2236 if (nqcpu != mycpu->gd_cpuid)
2240 flags = VFS_BIO_NEED_BUFSPACE;
2242 } else if (bufspace >= hibufspace) {
2244 flags = VFS_BIO_NEED_BUFSPACE;
2247 flags = VFS_BIO_NEED_ANY;
2250 bd_speedup(); /* heeeelp */
2251 atomic_set_int(&needsbuffer, flags);
2252 while (needsbuffer & flags) {
2255 tsleep_interlock(&needsbuffer, 0);
2256 value = atomic_fetchadd_int(&needsbuffer, 0);
2257 if (value & flags) {
2258 if (tsleep(&needsbuffer, PINTERLOCKED|slpflags,
2259 waitmsg, slptimeo)) {
2266 * We finally have a valid bp. We aren't quite out of the
2267 * woods, we still have to reserve kva space. In order
2268 * to keep fragmentation sane we only allocate kva in
2271 * (spin is not held)
2273 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
2275 if (maxsize != bp->b_kvasize) {
2276 vm_offset_t addr = 0;
2281 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2282 vm_map_lock(&buffer_map);
2284 if (vm_map_findspace(&buffer_map,
2285 vm_map_min(&buffer_map), maxsize,
2286 maxsize, 0, &addr)) {
2288 * Uh oh. Buffer map is too fragmented. We
2289 * must defragment the map.
2291 vm_map_unlock(&buffer_map);
2292 vm_map_entry_release(count);
2295 bp->b_flags |= B_INVAL;
2300 vm_map_insert(&buffer_map, &count,
2302 addr, addr + maxsize,
2304 VM_PROT_ALL, VM_PROT_ALL,
2307 bp->b_kvabase = (caddr_t) addr;
2308 bp->b_kvasize = maxsize;
2309 bufspace += bp->b_kvasize;
2312 vm_map_unlock(&buffer_map);
2313 vm_map_entry_release(count);
2315 bp->b_data = bp->b_kvabase;
2323 * Buffer flushing daemon. Buffers are normally flushed by the
2324 * update daemon but if it cannot keep up this process starts to
2325 * take the load in an attempt to prevent getnewbuf() from blocking.
2327 * Once a flush is initiated it does not stop until the number
2328 * of buffers falls below lodirtybuffers, but we will wake up anyone
2329 * waiting at the mid-point.
2331 static struct kproc_desc buf_kp = {
2336 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2337 kproc_start, &buf_kp)
2339 static struct kproc_desc bufhw_kp = {
2344 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2345 kproc_start, &bufhw_kp)
2348 buf_daemon1(struct thread *td, int queue, int (*buf_limit_fn)(long),
2354 marker = kmalloc(sizeof(*marker), M_BIOBUF, M_WAITOK | M_ZERO);
2355 marker->b_flags |= B_MARKER;
2356 marker->b_qindex = BQUEUE_NONE;
2360 * This process needs to be suspended prior to shutdown sync.
2362 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2363 td, SHUTDOWN_PRI_LAST);
2364 curthread->td_flags |= TDF_SYSTHREAD;
2367 * This process is allowed to take the buffer cache to the limit
2370 kproc_suspend_loop();
2373 * Do the flush as long as the number of dirty buffers
2374 * (including those running) exceeds lodirtybufspace.
2376 * When flushing limit running I/O to hirunningspace
2377 * Do the flush. Limit the amount of in-transit I/O we
2378 * allow to build up, otherwise we would completely saturate
2379 * the I/O system. Wakeup any waiting processes before we
2380 * normally would so they can run in parallel with our drain.
2382 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2383 * but because we split the operation into two threads we
2384 * have to cut it in half for each thread.
2386 waitrunningbufspace();
2387 limit = lodirtybufspace / 2;
2388 while (buf_limit_fn(limit)) {
2389 if (flushbufqueues(marker, queue) == 0)
2391 if (runningbufspace < hirunningspace)
2393 waitrunningbufspace();
2397 * We reached our low water mark, reset the
2398 * request and sleep until we are needed again.
2399 * The sleep is just so the suspend code works.
2401 tsleep_interlock(bd_req, 0);
2402 if (atomic_swap_int(bd_req, 0) == 0)
2403 tsleep(bd_req, PINTERLOCKED, "psleep", hz);
2406 /*kfree(marker, M_BIOBUF);*/
2410 buf_daemon_limit(long limit)
2412 return (runningbufspace + dirtykvaspace > limit ||
2413 dirtybufcount - dirtybufcounthw >= nbuf / 2);
2417 buf_daemon_hw_limit(long limit)
2419 return (runningbufspace + dirtykvaspace > limit ||
2420 dirtybufcounthw >= nbuf / 2);
2426 buf_daemon1(bufdaemon_td, BQUEUE_DIRTY, buf_daemon_limit,
2433 buf_daemon1(bufdaemonhw_td, BQUEUE_DIRTY_HW, buf_daemon_hw_limit,
2440 * Try to flush a buffer in the dirty queue. We must be careful to
2441 * free up B_INVAL buffers instead of write them, which NFS is
2442 * particularly sensitive to.
2444 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2445 * that we really want to try to get the buffer out and reuse it
2446 * due to the write load on the machine.
2448 * We must lock the buffer in order to check its validity before we
2449 * can mess with its contents. spin isn't enough.
2452 flushbufqueues(struct buf *marker, bufq_type_t q)
2454 struct bufpcpu *pcpu;
2457 int lcpu = marker->b_qcpu;
2459 KKASSERT(marker->b_qindex == BQUEUE_NONE);
2460 KKASSERT(marker->b_flags & B_MARKER);
2464 * Spinlock needed to perform operations on the queue and may be
2465 * held through a non-blocking BUF_LOCK(), but cannot be held when
2466 * BUF_UNLOCK()ing or through any other major operation.
2468 pcpu = &bufpcpu[marker->b_qcpu];
2469 spin_lock(&pcpu->spin);
2470 marker->b_qindex = q;
2471 TAILQ_INSERT_HEAD(&pcpu->bufqueues[q], marker, b_freelist);
2474 while ((bp = TAILQ_NEXT(bp, b_freelist)) != NULL) {
2476 * NOTE: spinlock is always held at the top of the loop
2478 if (bp->b_flags & B_MARKER)
2480 if ((bp->b_flags & B_DELWRI) == 0) {
2481 kprintf("Unexpected clean buffer %p\n", bp);
2484 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
2486 KKASSERT(bp->b_qcpu == marker->b_qcpu && bp->b_qindex == q);
2489 * Once the buffer is locked we will have no choice but to
2490 * unlock the spinlock around a later BUF_UNLOCK and re-set
2491 * bp = marker when looping. Move the marker now to make
2494 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist);
2495 TAILQ_INSERT_AFTER(&pcpu->bufqueues[q], bp, marker, b_freelist);
2498 * Must recheck B_DELWRI after successfully locking
2501 if ((bp->b_flags & B_DELWRI) == 0) {
2502 spin_unlock(&pcpu->spin);
2504 spin_lock(&pcpu->spin);
2510 * Remove the buffer from its queue. We still own the
2516 * Disposing of an invalid buffer counts as a flush op
2518 if (bp->b_flags & B_INVAL) {
2519 spin_unlock(&pcpu->spin);
2521 spin_lock(&pcpu->spin);
2527 * Release the spinlock for the more complex ops we
2528 * are now going to do.
2530 spin_unlock(&pcpu->spin);
2534 * This is a bit messy
2536 if (LIST_FIRST(&bp->b_dep) != NULL &&
2537 (bp->b_flags & B_DEFERRED) == 0 &&
2538 buf_countdeps(bp, 0)) {
2539 spin_lock(&pcpu->spin);
2540 TAILQ_INSERT_TAIL(&pcpu->bufqueues[q], bp, b_freelist);
2542 bp->b_flags |= B_DEFERRED;
2543 spin_unlock(&pcpu->spin);
2545 spin_lock(&pcpu->spin);
2551 * spinlock not held here.
2553 * If the buffer has a dependancy, buf_checkwrite() must
2554 * also return 0 for us to be able to initate the write.
2556 * If the buffer is flagged B_ERROR it may be requeued
2557 * over and over again, we try to avoid a live lock.
2559 if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) {
2561 } else if (bp->b_flags & B_ERROR) {
2562 tsleep(bp, 0, "bioer", 1);
2563 bp->b_flags &= ~B_AGE;
2566 bp->b_flags |= B_AGE;
2569 spin_lock(&pcpu->spin);
2574 TAILQ_REMOVE(&pcpu->bufqueues[q], marker, b_freelist);
2575 marker->b_qindex = BQUEUE_NONE;
2576 spin_unlock(&pcpu->spin);
2579 * Advance the marker to be fair.
2581 marker->b_qcpu = (marker->b_qcpu + 1) % ncpus;
2583 if (marker->b_qcpu != lcpu)
2593 * Returns true if no I/O is needed to access the associated VM object.
2594 * This is like findblk except it also hunts around in the VM system for
2597 * Note that we ignore vm_page_free() races from interrupts against our
2598 * lookup, since if the caller is not protected our return value will not
2599 * be any more valid then otherwise once we exit the critical section.
2602 inmem(struct vnode *vp, off_t loffset)
2605 vm_offset_t toff, tinc, size;
2609 if (findblk(vp, loffset, FINDBLK_TEST))
2611 if (vp->v_mount == NULL)
2613 if ((obj = vp->v_object) == NULL)
2617 if (size > vp->v_mount->mnt_stat.f_iosize)
2618 size = vp->v_mount->mnt_stat.f_iosize;
2620 vm_object_hold(obj);
2621 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2622 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2628 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2629 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2630 if (vm_page_is_valid(m,
2631 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0) {
2636 vm_object_drop(obj);
2643 * Locate and return the specified buffer. Unless flagged otherwise,
2644 * a locked buffer will be returned if it exists or NULL if it does not.
2646 * findblk()'d buffers are still on the bufqueues and if you intend
2647 * to use your (locked NON-TEST) buffer you need to bremfree(bp)
2648 * and possibly do other stuff to it.
2650 * FINDBLK_TEST - Do not lock the buffer. The caller is responsible
2651 * for locking the buffer and ensuring that it remains
2652 * the desired buffer after locking.
2654 * FINDBLK_NBLOCK - Lock the buffer non-blocking. If we are unable
2655 * to acquire the lock we return NULL, even if the
2658 * FINDBLK_REF - Returns the buffer ref'd, which prevents normal
2659 * reuse by getnewbuf() but does not prevent
2660 * disassociation (B_INVAL). Used to avoid deadlocks
2661 * against random (vp,loffset)s due to reassignment.
2663 * (0) - Lock the buffer blocking.
2666 findblk(struct vnode *vp, off_t loffset, int flags)
2671 lkflags = LK_EXCLUSIVE;
2672 if (flags & FINDBLK_NBLOCK)
2673 lkflags |= LK_NOWAIT;
2677 * Lookup. Ref the buf while holding v_token to prevent
2678 * reuse (but does not prevent diassociation).
2680 lwkt_gettoken_shared(&vp->v_token);
2681 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2683 lwkt_reltoken(&vp->v_token);
2687 lwkt_reltoken(&vp->v_token);
2690 * If testing only break and return bp, do not lock.
2692 if (flags & FINDBLK_TEST)
2696 * Lock the buffer, return an error if the lock fails.
2697 * (only FINDBLK_NBLOCK can cause the lock to fail).
2699 if (BUF_LOCK(bp, lkflags)) {
2700 atomic_subtract_int(&bp->b_refs, 1);
2701 /* bp = NULL; not needed */
2706 * Revalidate the locked buf before allowing it to be
2709 if (bp->b_vp == vp && bp->b_loffset == loffset)
2711 atomic_subtract_int(&bp->b_refs, 1);
2718 if ((flags & FINDBLK_REF) == 0)
2719 atomic_subtract_int(&bp->b_refs, 1);
2726 * Similar to getblk() except only returns the buffer if it is
2727 * B_CACHE and requires no other manipulation. Otherwise NULL
2730 * If B_RAM is set the buffer might be just fine, but we return
2731 * NULL anyway because we want the code to fall through to the
2732 * cluster read. Otherwise read-ahead breaks.
2734 * If blksize is 0 the buffer cache buffer must already be fully
2737 * If blksize is non-zero getblk() will be used, allowing a buffer
2738 * to be reinstantiated from its VM backing store. The buffer must
2739 * still be fully cached after reinstantiation to be returned.
2742 getcacheblk(struct vnode *vp, off_t loffset, int blksize, int blkflags)
2745 int fndflags = (blkflags & GETBLK_NOWAIT) ? FINDBLK_NBLOCK : 0;
2748 bp = getblk(vp, loffset, blksize, blkflags, 0);
2750 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2752 bp->b_flags &= ~B_AGE;
2759 bp = findblk(vp, loffset, fndflags);
2761 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
2763 bp->b_flags &= ~B_AGE;
2777 * Get a block given a specified block and offset into a file/device.
2778 * B_INVAL may or may not be set on return. The caller should clear
2779 * B_INVAL prior to initiating a READ.
2781 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2782 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2783 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2784 * without doing any of those things the system will likely believe
2785 * the buffer to be valid (especially if it is not B_VMIO), and the
2786 * next getblk() will return the buffer with B_CACHE set.
2788 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2789 * an existing buffer.
2791 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2792 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2793 * and then cleared based on the backing VM. If the previous buffer is
2794 * non-0-sized but invalid, B_CACHE will be cleared.
2796 * If getblk() must create a new buffer, the new buffer is returned with
2797 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2798 * case it is returned with B_INVAL clear and B_CACHE set based on the
2801 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2802 * B_CACHE bit is clear.
2804 * What this means, basically, is that the caller should use B_CACHE to
2805 * determine whether the buffer is fully valid or not and should clear
2806 * B_INVAL prior to issuing a read. If the caller intends to validate
2807 * the buffer by loading its data area with something, the caller needs
2808 * to clear B_INVAL. If the caller does this without issuing an I/O,
2809 * the caller should set B_CACHE ( as an optimization ), else the caller
2810 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2811 * a write attempt or if it was a successfull read. If the caller
2812 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2813 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2817 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2818 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2821 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2824 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2828 if (size > MAXBSIZE)
2829 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2830 if (vp->v_object == NULL)
2831 panic("getblk: vnode %p has no object!", vp);
2834 if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) {
2836 * The buffer was found in the cache, but we need to lock it.
2837 * We must acquire a ref on the bp to prevent reuse, but
2838 * this will not prevent disassociation (brelvp()) so we
2839 * must recheck (vp,loffset) after acquiring the lock.
2841 * Without the ref the buffer could potentially be reused
2842 * before we acquire the lock and create a deadlock
2843 * situation between the thread trying to reuse the buffer
2844 * and us due to the fact that we would wind up blocking
2845 * on a random (vp,loffset).
2847 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2848 if (blkflags & GETBLK_NOWAIT) {
2852 lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2853 if (blkflags & GETBLK_PCATCH)
2854 lkflags |= LK_PCATCH;
2855 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2858 if (error == ENOLCK)
2862 /* buffer may have changed on us */
2867 * Once the buffer has been locked, make sure we didn't race
2868 * a buffer recyclement. Buffers that are no longer hashed
2869 * will have b_vp == NULL, so this takes care of that check
2872 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2873 kprintf("Warning buffer %p (vp %p loffset %lld) "
2875 bp, vp, (long long)loffset);
2881 * If SZMATCH any pre-existing buffer must be of the requested
2882 * size or NULL is returned. The caller absolutely does not
2883 * want getblk() to bwrite() the buffer on a size mismatch.
2885 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2891 * All vnode-based buffers must be backed by a VM object.
2893 KKASSERT(bp->b_flags & B_VMIO);
2894 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2895 bp->b_flags &= ~B_AGE;
2898 * Make sure that B_INVAL buffers do not have a cached
2899 * block number translation.
2901 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2902 kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
2903 " did not have cleared bio_offset cache\n",
2904 bp, vp, (long long)loffset);
2905 clearbiocache(&bp->b_bio2);
2909 * The buffer is locked. B_CACHE is cleared if the buffer is
2912 if (bp->b_flags & B_INVAL)
2913 bp->b_flags &= ~B_CACHE;
2917 * Any size inconsistancy with a dirty buffer or a buffer
2918 * with a softupdates dependancy must be resolved. Resizing
2919 * the buffer in such circumstances can lead to problems.
2921 * Dirty or dependant buffers are written synchronously.
2922 * Other types of buffers are simply released and
2923 * reconstituted as they may be backed by valid, dirty VM
2924 * pages (but not marked B_DELWRI).
2926 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
2927 * and may be left over from a prior truncation (and thus
2928 * no longer represent the actual EOF point), so we
2929 * definitely do not want to B_NOCACHE the backing store.
2931 if (size != bp->b_bcount) {
2932 if (bp->b_flags & B_DELWRI) {
2933 bp->b_flags |= B_RELBUF;
2935 } else if (LIST_FIRST(&bp->b_dep)) {
2936 bp->b_flags |= B_RELBUF;
2939 bp->b_flags |= B_RELBUF;
2944 KKASSERT(size <= bp->b_kvasize);
2945 KASSERT(bp->b_loffset != NOOFFSET,
2946 ("getblk: no buffer offset"));
2949 * A buffer with B_DELWRI set and B_CACHE clear must
2950 * be committed before we can return the buffer in
2951 * order to prevent the caller from issuing a read
2952 * ( due to B_CACHE not being set ) and overwriting
2955 * Most callers, including NFS and FFS, need this to
2956 * operate properly either because they assume they
2957 * can issue a read if B_CACHE is not set, or because
2958 * ( for example ) an uncached B_DELWRI might loop due
2959 * to softupdates re-dirtying the buffer. In the latter
2960 * case, B_CACHE is set after the first write completes,
2961 * preventing further loops.
2963 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2964 * above while extending the buffer, we cannot allow the
2965 * buffer to remain with B_CACHE set after the write
2966 * completes or it will represent a corrupt state. To
2967 * deal with this we set B_NOCACHE to scrap the buffer
2970 * XXX Should this be B_RELBUF instead of B_NOCACHE?
2971 * I'm not even sure this state is still possible
2972 * now that getblk() writes out any dirty buffers
2975 * We might be able to do something fancy, like setting
2976 * B_CACHE in bwrite() except if B_DELWRI is already set,
2977 * so the below call doesn't set B_CACHE, but that gets real
2978 * confusing. This is much easier.
2981 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2982 kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
2983 "and CACHE clear, b_flags %08x\n",
2984 bp, (uintmax_t)bp->b_loffset, bp->b_flags);
2985 bp->b_flags |= B_NOCACHE;
2991 * Buffer is not in-core, create new buffer. The buffer
2992 * returned by getnewbuf() is locked. Note that the returned
2993 * buffer is also considered valid (not marked B_INVAL).
2995 * Calculating the offset for the I/O requires figuring out
2996 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2997 * the mount's f_iosize otherwise. If the vnode does not
2998 * have an associated mount we assume that the passed size is
3001 * Note that vn_isdisk() cannot be used here since it may
3002 * return a failure for numerous reasons. Note that the
3003 * buffer size may be larger then the block size (the caller
3004 * will use block numbers with the proper multiple). Beware
3005 * of using any v_* fields which are part of unions. In
3006 * particular, in DragonFly the mount point overloading
3007 * mechanism uses the namecache only and the underlying
3008 * directory vnode is not a special case.
3012 if (vp->v_type == VBLK || vp->v_type == VCHR)
3014 else if (vp->v_mount)
3015 bsize = vp->v_mount->mnt_stat.f_iosize;
3019 maxsize = size + (loffset & PAGE_MASK);
3020 maxsize = imax(maxsize, bsize);
3022 bp = getnewbuf(blkflags, slptimeo, size, maxsize);
3024 if (slpflags || slptimeo)
3030 * Atomically insert the buffer into the hash, so that it can
3031 * be found by findblk().
3033 * If bgetvp() returns non-zero a collision occured, and the
3034 * bp will not be associated with the vnode.
3036 * Make sure the translation layer has been cleared.
3038 bp->b_loffset = loffset;
3039 bp->b_bio2.bio_offset = NOOFFSET;
3040 /* bp->b_bio2.bio_next = NULL; */
3042 if (bgetvp(vp, bp, size)) {
3043 bp->b_flags |= B_INVAL;
3049 * All vnode-based buffers must be backed by a VM object.
3051 KKASSERT(vp->v_object != NULL);
3052 bp->b_flags |= B_VMIO;
3053 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3057 KKASSERT(dsched_is_clear_buf_priv(bp));
3064 * Reacquire a buffer that was previously released to the locked queue,
3065 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
3066 * set B_LOCKED (which handles the acquisition race).
3068 * To this end, either B_LOCKED must be set or the dependancy list must be
3072 regetblk(struct buf *bp)
3074 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
3075 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
3082 * Get an empty, disassociated buffer of given size. The buffer is
3083 * initially set to B_INVAL.
3085 * critical section protection is not required for the allocbuf()
3086 * call because races are impossible here.
3094 maxsize = (size + BKVAMASK) & ~BKVAMASK;
3096 while ((bp = getnewbuf(0, 0, size, maxsize)) == NULL)
3099 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
3100 KKASSERT(dsched_is_clear_buf_priv(bp));
3108 * This code constitutes the buffer memory from either anonymous system
3109 * memory (in the case of non-VMIO operations) or from an associated
3110 * VM object (in the case of VMIO operations). This code is able to
3111 * resize a buffer up or down.
3113 * Note that this code is tricky, and has many complications to resolve
3114 * deadlock or inconsistant data situations. Tread lightly!!!
3115 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
3116 * the caller. Calling this code willy nilly can result in the loss of
3119 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
3120 * B_CACHE for the non-VMIO case.
3122 * This routine does not need to be called from a critical section but you
3123 * must own the buffer.
3126 allocbuf(struct buf *bp, int size)
3128 int newbsize, mbsize;
3131 if (BUF_REFCNT(bp) == 0)
3132 panic("allocbuf: buffer not busy");
3134 if (bp->b_kvasize < size)
3135 panic("allocbuf: buffer too small");
3137 if ((bp->b_flags & B_VMIO) == 0) {
3141 * Just get anonymous memory from the kernel. Don't
3142 * mess with B_CACHE.
3144 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3145 if (bp->b_flags & B_MALLOC)
3148 newbsize = round_page(size);
3150 if (newbsize < bp->b_bufsize) {
3152 * Malloced buffers are not shrunk
3154 if (bp->b_flags & B_MALLOC) {
3156 bp->b_bcount = size;
3158 kfree(bp->b_data, M_BIOBUF);
3159 if (bp->b_bufsize) {
3160 atomic_subtract_long(&bufmallocspace, bp->b_bufsize);
3164 bp->b_data = bp->b_kvabase;
3166 bp->b_flags &= ~B_MALLOC;
3172 (vm_offset_t) bp->b_data + newbsize,
3173 (vm_offset_t) bp->b_data + bp->b_bufsize);
3174 } else if (newbsize > bp->b_bufsize) {
3176 * We only use malloced memory on the first allocation.
3177 * and revert to page-allocated memory when the buffer
3180 if ((bufmallocspace < maxbufmallocspace) &&
3181 (bp->b_bufsize == 0) &&
3182 (mbsize <= PAGE_SIZE/2)) {
3184 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
3185 bp->b_bufsize = mbsize;
3186 bp->b_bcount = size;
3187 bp->b_flags |= B_MALLOC;
3188 atomic_add_long(&bufmallocspace, mbsize);
3194 * If the buffer is growing on its other-than-first
3195 * allocation, then we revert to the page-allocation
3198 if (bp->b_flags & B_MALLOC) {
3199 origbuf = bp->b_data;
3200 origbufsize = bp->b_bufsize;
3201 bp->b_data = bp->b_kvabase;
3202 if (bp->b_bufsize) {
3203 atomic_subtract_long(&bufmallocspace,
3208 bp->b_flags &= ~B_MALLOC;
3209 newbsize = round_page(newbsize);
3213 (vm_offset_t) bp->b_data + bp->b_bufsize,
3214 (vm_offset_t) bp->b_data + newbsize);
3216 bcopy(origbuf, bp->b_data, origbufsize);
3217 kfree(origbuf, M_BIOBUF);
3224 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
3225 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
3226 newbsize + PAGE_MASK) >> PAGE_SHIFT;
3227 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
3229 if (bp->b_flags & B_MALLOC)
3230 panic("allocbuf: VMIO buffer can't be malloced");
3232 * Set B_CACHE initially if buffer is 0 length or will become
3235 if (size == 0 || bp->b_bufsize == 0)
3236 bp->b_flags |= B_CACHE;
3238 if (newbsize < bp->b_bufsize) {
3240 * DEV_BSIZE aligned new buffer size is less then the
3241 * DEV_BSIZE aligned existing buffer size. Figure out
3242 * if we have to remove any pages.
3244 if (desiredpages < bp->b_xio.xio_npages) {
3245 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
3247 * the page is not freed here -- it
3248 * is the responsibility of
3249 * vnode_pager_setsize
3251 m = bp->b_xio.xio_pages[i];
3252 KASSERT(m != bogus_page,
3253 ("allocbuf: bogus page found"));
3254 vm_page_busy_wait(m, TRUE, "biodep");
3255 bp->b_xio.xio_pages[i] = NULL;
3256 vm_page_unwire(m, 0);
3259 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
3260 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
3261 bp->b_xio.xio_npages = desiredpages;
3263 } else if (size > bp->b_bcount) {
3265 * We are growing the buffer, possibly in a
3266 * byte-granular fashion.
3274 * Step 1, bring in the VM pages from the object,
3275 * allocating them if necessary. We must clear
3276 * B_CACHE if these pages are not valid for the
3277 * range covered by the buffer.
3279 * critical section protection is required to protect
3280 * against interrupts unbusying and freeing pages
3281 * between our vm_page_lookup() and our
3282 * busycheck/wiring call.
3287 vm_object_hold(obj);
3288 while (bp->b_xio.xio_npages < desiredpages) {
3293 pi = OFF_TO_IDX(bp->b_loffset) +
3294 bp->b_xio.xio_npages;
3297 * Blocking on m->busy might lead to a
3300 * vm_fault->getpages->cluster_read->allocbuf
3302 m = vm_page_lookup_busy_try(obj, pi, FALSE,
3305 vm_page_sleep_busy(m, FALSE, "pgtblk");
3310 * note: must allocate system pages
3311 * since blocking here could intefere
3312 * with paging I/O, no matter which
3315 m = bio_page_alloc(bp, obj, pi, desiredpages - bp->b_xio.xio_npages);
3318 vm_page_flag_clear(m, PG_ZERO);
3320 bp->b_flags &= ~B_CACHE;
3321 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3322 ++bp->b_xio.xio_npages;
3328 * We found a page and were able to busy it.
3330 vm_page_flag_clear(m, PG_ZERO);
3333 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
3334 ++bp->b_xio.xio_npages;
3335 if (bp->b_act_count < m->act_count)
3336 bp->b_act_count = m->act_count;
3338 vm_object_drop(obj);
3341 * Step 2. We've loaded the pages into the buffer,
3342 * we have to figure out if we can still have B_CACHE
3343 * set. Note that B_CACHE is set according to the
3344 * byte-granular range ( bcount and size ), not the
3345 * aligned range ( newbsize ).
3347 * The VM test is against m->valid, which is DEV_BSIZE
3348 * aligned. Needless to say, the validity of the data
3349 * needs to also be DEV_BSIZE aligned. Note that this
3350 * fails with NFS if the server or some other client
3351 * extends the file's EOF. If our buffer is resized,
3352 * B_CACHE may remain set! XXX
3355 toff = bp->b_bcount;
3356 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
3358 while ((bp->b_flags & B_CACHE) && toff < size) {
3361 if (tinc > (size - toff))
3364 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
3372 bp->b_xio.xio_pages[pi]
3379 * Step 3, fixup the KVM pmap. Remember that
3380 * bp->b_data is relative to bp->b_loffset, but
3381 * bp->b_loffset may be offset into the first page.
3384 bp->b_data = (caddr_t)
3385 trunc_page((vm_offset_t)bp->b_data);
3387 (vm_offset_t)bp->b_data,
3388 bp->b_xio.xio_pages,
3389 bp->b_xio.xio_npages
3391 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3392 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3396 /* adjust space use on already-dirty buffer */
3397 if (bp->b_flags & B_DELWRI) {
3398 /* dirtykvaspace unchanged */
3399 atomic_add_long(&dirtybufspace, newbsize - bp->b_bufsize);
3400 if (bp->b_flags & B_HEAVY) {
3401 atomic_add_long(&dirtybufspacehw,
3402 newbsize - bp->b_bufsize);
3405 if (newbsize < bp->b_bufsize)
3407 bp->b_bufsize = newbsize; /* actual buffer allocation */
3408 bp->b_bcount = size; /* requested buffer size */
3415 * Wait for buffer I/O completion, returning error status. B_EINTR
3416 * is converted into an EINTR error but not cleared (since a chain
3417 * of biowait() calls may occur).
3419 * On return bpdone() will have been called but the buffer will remain
3420 * locked and will not have been brelse()'d.
3422 * NOTE! If a timeout is specified and ETIMEDOUT occurs the I/O is
3423 * likely still in progress on return.
3425 * NOTE! This operation is on a BIO, not a BUF.
3427 * NOTE! BIO_DONE is cleared by vn_strategy()
3430 _biowait(struct bio *bio, const char *wmesg, int to)
3432 struct buf *bp = bio->bio_buf;
3437 KKASSERT(bio == &bp->b_bio1);
3439 flags = bio->bio_flags;
3440 if (flags & BIO_DONE)
3442 nflags = flags | BIO_WANT;
3443 tsleep_interlock(bio, 0);
3444 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3446 error = tsleep(bio, PINTERLOCKED, wmesg, to);
3447 else if (bp->b_cmd == BUF_CMD_READ)
3448 error = tsleep(bio, PINTERLOCKED, "biord", to);
3450 error = tsleep(bio, PINTERLOCKED, "biowr", to);
3452 kprintf("tsleep error biowait %d\n", error);
3461 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
3462 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
3463 if (bp->b_flags & B_EINTR)
3465 if (bp->b_flags & B_ERROR)
3466 return (bp->b_error ? bp->b_error : EIO);
3471 biowait(struct bio *bio, const char *wmesg)
3473 return(_biowait(bio, wmesg, 0));
3477 biowait_timeout(struct bio *bio, const char *wmesg, int to)
3479 return(_biowait(bio, wmesg, to));
3483 * This associates a tracking count with an I/O. vn_strategy() and
3484 * dev_dstrategy() do this automatically but there are a few cases
3485 * where a vnode or device layer is bypassed when a block translation
3486 * is cached. In such cases bio_start_transaction() may be called on
3487 * the bypassed layers so the system gets an I/O in progress indication
3488 * for those higher layers.
3491 bio_start_transaction(struct bio *bio, struct bio_track *track)
3493 bio->bio_track = track;
3494 if (dsched_is_clear_buf_priv(bio->bio_buf))
3495 dsched_new_buf(bio->bio_buf);
3496 bio_track_ref(track);
3500 * Initiate I/O on a vnode.
3502 * SWAPCACHE OPERATION:
3504 * Real buffer cache buffers have a non-NULL bp->b_vp. Unfortunately
3505 * devfs also uses b_vp for fake buffers so we also have to check
3506 * that B_PAGING is 0. In this case the passed 'vp' is probably the
3507 * underlying block device. The swap assignments are related to the
3508 * buffer cache buffer's b_vp, not the passed vp.
3510 * The passed vp == bp->b_vp only in the case where the strategy call
3511 * is made on the vp itself for its own buffers (a regular file or
3512 * block device vp). The filesystem usually then re-calls vn_strategy()
3513 * after translating the request to an underlying device.
3515 * Cluster buffers set B_CLUSTER and the passed vp is the vp of the
3516 * underlying buffer cache buffers.
3518 * We can only deal with page-aligned buffers at the moment, because
3519 * we can't tell what the real dirty state for pages straddling a buffer
3522 * In order to call swap_pager_strategy() we must provide the VM object
3523 * and base offset for the underlying buffer cache pages so it can find
3527 vn_strategy(struct vnode *vp, struct bio *bio)
3529 struct bio_track *track;
3530 struct buf *bp = bio->bio_buf;
3532 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3535 * Set when an I/O is issued on the bp. Cleared by consumers
3536 * (aka HAMMER), allowing the consumer to determine if I/O had
3537 * actually occurred.
3539 bp->b_flags |= B_IODEBUG;
3542 * Handle the swap cache intercept.
3544 if (vn_cache_strategy(vp, bio))
3548 * Otherwise do the operation through the filesystem
3550 if (bp->b_cmd == BUF_CMD_READ)
3551 track = &vp->v_track_read;
3553 track = &vp->v_track_write;
3554 KKASSERT((bio->bio_flags & BIO_DONE) == 0);
3555 bio->bio_track = track;
3556 if (dsched_is_clear_buf_priv(bio->bio_buf))
3557 dsched_new_buf(bio->bio_buf);
3558 bio_track_ref(track);
3559 vop_strategy(*vp->v_ops, vp, bio);
3562 static void vn_cache_strategy_callback(struct bio *bio);
3565 vn_cache_strategy(struct vnode *vp, struct bio *bio)
3567 struct buf *bp = bio->bio_buf;
3574 * Is this buffer cache buffer suitable for reading from
3577 if (vm_swapcache_read_enable == 0 ||
3578 bp->b_cmd != BUF_CMD_READ ||
3579 ((bp->b_flags & B_CLUSTER) == 0 &&
3580 (bp->b_vp == NULL || (bp->b_flags & B_PAGING))) ||
3581 ((int)bp->b_loffset & PAGE_MASK) != 0 ||
3582 (bp->b_bcount & PAGE_MASK) != 0) {
3587 * Figure out the original VM object (it will match the underlying
3588 * VM pages). Note that swap cached data uses page indices relative
3589 * to that object, not relative to bio->bio_offset.
3591 if (bp->b_flags & B_CLUSTER)
3592 object = vp->v_object;
3594 object = bp->b_vp->v_object;
3597 * In order to be able to use the swap cache all underlying VM
3598 * pages must be marked as such, and we can't have any bogus pages.
3600 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
3601 m = bp->b_xio.xio_pages[i];
3602 if ((m->flags & PG_SWAPPED) == 0)
3604 if (m == bogus_page)
3609 * If we are good then issue the I/O using swap_pager_strategy().
3611 * We can only do this if the buffer actually supports object-backed
3612 * I/O. If it doesn't npages will be 0.
3614 if (i && i == bp->b_xio.xio_npages) {
3615 m = bp->b_xio.xio_pages[0];
3616 nbio = push_bio(bio);
3617 nbio->bio_done = vn_cache_strategy_callback;
3618 nbio->bio_offset = ptoa(m->pindex);
3619 KKASSERT(m->object == object);
3620 swap_pager_strategy(object, nbio);
3627 * This is a bit of a hack but since the vn_cache_strategy() function can
3628 * override a VFS's strategy function we must make sure that the bio, which
3629 * is probably bio2, doesn't leak an unexpected offset value back to the
3630 * filesystem. The filesystem (e.g. UFS) might otherwise assume that the
3631 * bio went through its own file strategy function and the the bio2 offset
3632 * is a cached disk offset when, in fact, it isn't.
3635 vn_cache_strategy_callback(struct bio *bio)
3637 bio->bio_offset = NOOFFSET;
3638 biodone(pop_bio(bio));
3644 * Finish I/O on a buffer after all BIOs have been processed.
3645 * Called when the bio chain is exhausted or by biowait. If called
3646 * by biowait, elseit is typically 0.
3648 * bpdone is also responsible for setting B_CACHE in a B_VMIO bp.
3649 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3650 * assuming B_INVAL is clear.
3652 * For the VMIO case, we set B_CACHE if the op was a read and no
3653 * read error occured, or if the op was a write. B_CACHE is never
3654 * set if the buffer is invalid or otherwise uncacheable.
3656 * bpdone does not mess with B_INVAL, allowing the I/O routine or the
3657 * initiator to leave B_INVAL set to brelse the buffer out of existance
3658 * in the biodone routine.
3661 bpdone(struct buf *bp, int elseit)
3665 KASSERT(BUF_REFCNTNB(bp) > 0,
3666 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3667 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3668 ("biodone: bp %p already done!", bp));
3671 * No more BIOs are left. All completion functions have been dealt
3672 * with, now we clean up the buffer.
3675 bp->b_cmd = BUF_CMD_DONE;
3678 * Only reads and writes are processed past this point.
3680 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3681 if (cmd == BUF_CMD_FREEBLKS)
3682 bp->b_flags |= B_NOCACHE;
3689 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3690 * a lot worse. XXX - move this above the clearing of b_cmd
3692 if (LIST_FIRST(&bp->b_dep) != NULL)
3696 * A failed write must re-dirty the buffer unless B_INVAL
3697 * was set. Only applicable to normal buffers (with VPs).
3698 * vinum buffers may not have a vp.
3700 if (cmd == BUF_CMD_WRITE &&
3701 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3702 bp->b_flags &= ~B_NOCACHE;
3707 if (bp->b_flags & B_VMIO) {
3713 struct vnode *vp = bp->b_vp;
3717 #if defined(VFS_BIO_DEBUG)
3718 if (vp->v_auxrefs == 0)
3719 panic("biodone: zero vnode hold count");
3720 if ((vp->v_flag & VOBJBUF) == 0)
3721 panic("biodone: vnode is not setup for merged cache");
3724 foff = bp->b_loffset;
3725 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3726 KASSERT(obj != NULL, ("biodone: missing VM object"));
3728 #if defined(VFS_BIO_DEBUG)
3729 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3730 kprintf("biodone: paging in progress(%d) < "
3731 "bp->b_xio.xio_npages(%d)\n",
3732 obj->paging_in_progress,
3733 bp->b_xio.xio_npages);
3738 * Set B_CACHE if the op was a normal read and no error
3739 * occured. B_CACHE is set for writes in the b*write()
3742 iosize = bp->b_bcount - bp->b_resid;
3743 if (cmd == BUF_CMD_READ &&
3744 (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3745 bp->b_flags |= B_CACHE;
3748 vm_object_hold(obj);
3749 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3753 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3758 * cleanup bogus pages, restoring the originals. Since
3759 * the originals should still be wired, we don't have
3760 * to worry about interrupt/freeing races destroying
3761 * the VM object association.
3763 m = bp->b_xio.xio_pages[i];
3764 if (m == bogus_page) {
3766 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3768 panic("biodone: page disappeared");
3769 bp->b_xio.xio_pages[i] = m;
3770 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3771 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3773 #if defined(VFS_BIO_DEBUG)
3774 if (OFF_TO_IDX(foff) != m->pindex) {
3775 kprintf("biodone: foff(%lu)/m->pindex(%ld) "
3777 (unsigned long)foff, (long)m->pindex);
3782 * In the write case, the valid and clean bits are
3783 * already changed correctly (see bdwrite()), so we
3784 * only need to do this here in the read case.
3786 vm_page_busy_wait(m, FALSE, "bpdpgw");
3787 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3788 vfs_clean_one_page(bp, i, m);
3790 vm_page_flag_clear(m, PG_ZERO);
3793 * when debugging new filesystems or buffer I/O
3794 * methods, this is the most common error that pops
3795 * up. if you see this, you have not set the page
3796 * busy flag correctly!!!
3799 kprintf("biodone: page busy < 0, "
3800 "pindex: %d, foff: 0x(%x,%x), "
3801 "resid: %d, index: %d\n",
3802 (int) m->pindex, (int)(foff >> 32),
3803 (int) foff & 0xffffffff, resid, i);
3804 if (!vn_isdisk(vp, NULL))
3805 kprintf(" iosize: %ld, loffset: %lld, "
3806 "flags: 0x%08x, npages: %d\n",
3807 bp->b_vp->v_mount->mnt_stat.f_iosize,
3808 (long long)bp->b_loffset,
3809 bp->b_flags, bp->b_xio.xio_npages);
3811 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3812 (long long)bp->b_loffset,
3813 bp->b_flags, bp->b_xio.xio_npages);
3814 kprintf(" valid: 0x%x, dirty: 0x%x, "
3818 panic("biodone: page busy < 0");
3820 vm_page_io_finish(m);
3822 vm_object_pip_wakeup(obj);
3823 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3826 bp->b_flags &= ~B_HASBOGUS;
3827 vm_object_drop(obj);
3831 * Finish up by releasing the buffer. There are no more synchronous
3832 * or asynchronous completions, those were handled by bio_done
3836 if (bp->b_flags & (B_NOCACHE|B_INVAL|B_ERROR|B_RELBUF))
3847 biodone(struct bio *bio)
3849 struct buf *bp = bio->bio_buf;
3851 runningbufwakeup(bp);
3854 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3857 biodone_t *done_func;
3858 struct bio_track *track;
3861 * BIO tracking. Most but not all BIOs are tracked.
3863 if ((track = bio->bio_track) != NULL) {
3864 bio_track_rel(track);
3865 bio->bio_track = NULL;
3869 * A bio_done function terminates the loop. The function
3870 * will be responsible for any further chaining and/or
3871 * buffer management.
3873 * WARNING! The done function can deallocate the buffer!
3875 if ((done_func = bio->bio_done) != NULL) {
3876 bio->bio_done = NULL;
3880 bio = bio->bio_prev;
3884 * If we've run out of bio's do normal [a]synchronous completion.
3890 * Synchronous biodone - this terminates a synchronous BIO.
3892 * bpdone() is called with elseit=FALSE, leaving the buffer completed
3893 * but still locked. The caller must brelse() the buffer after waiting
3897 biodone_sync(struct bio *bio)
3899 struct buf *bp = bio->bio_buf;
3903 KKASSERT(bio == &bp->b_bio1);
3907 flags = bio->bio_flags;
3908 nflags = (flags | BIO_DONE) & ~BIO_WANT;
3910 if (atomic_cmpset_int(&bio->bio_flags, flags, nflags)) {
3911 if (flags & BIO_WANT)
3921 * This routine is called in lieu of iodone in the case of
3922 * incomplete I/O. This keeps the busy status for pages
3926 vfs_unbusy_pages(struct buf *bp)
3930 runningbufwakeup(bp);
3932 if (bp->b_flags & B_VMIO) {
3933 struct vnode *vp = bp->b_vp;
3937 vm_object_hold(obj);
3939 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3940 vm_page_t m = bp->b_xio.xio_pages[i];
3943 * When restoring bogus changes the original pages
3944 * should still be wired, so we are in no danger of
3945 * losing the object association and do not need
3946 * critical section protection particularly.
3948 if (m == bogus_page) {
3949 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3951 panic("vfs_unbusy_pages: page missing");
3953 bp->b_xio.xio_pages[i] = m;
3954 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3955 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3957 vm_page_busy_wait(m, FALSE, "bpdpgw");
3958 vm_page_flag_clear(m, PG_ZERO);
3959 vm_page_io_finish(m);
3961 vm_object_pip_wakeup(obj);
3963 bp->b_flags &= ~B_HASBOGUS;
3964 vm_object_drop(obj);
3971 * This routine is called before a device strategy routine.
3972 * It is used to tell the VM system that paging I/O is in
3973 * progress, and treat the pages associated with the buffer
3974 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3975 * flag is handled to make sure that the object doesn't become
3978 * Since I/O has not been initiated yet, certain buffer flags
3979 * such as B_ERROR or B_INVAL may be in an inconsistant state
3980 * and should be ignored.
3983 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3986 struct lwp *lp = curthread->td_lwp;
3989 * The buffer's I/O command must already be set. If reading,
3990 * B_CACHE must be 0 (double check against callers only doing
3991 * I/O when B_CACHE is 0).
3993 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3994 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3996 if (bp->b_flags & B_VMIO) {
4000 KASSERT(bp->b_loffset != NOOFFSET,
4001 ("vfs_busy_pages: no buffer offset"));
4004 * Busy all the pages. We have to busy them all at once
4005 * to avoid deadlocks.
4008 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4009 vm_page_t m = bp->b_xio.xio_pages[i];
4011 if (vm_page_busy_try(m, FALSE)) {
4012 vm_page_sleep_busy(m, FALSE, "vbpage");
4014 vm_page_wakeup(bp->b_xio.xio_pages[i]);
4020 * Setup for I/O, soft-busy the page right now because
4021 * the next loop may block.
4023 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4024 vm_page_t m = bp->b_xio.xio_pages[i];
4026 vm_page_flag_clear(m, PG_ZERO);
4027 if ((bp->b_flags & B_CLUSTER) == 0) {
4028 vm_object_pip_add(obj, 1);
4029 vm_page_io_start(m);
4034 * Adjust protections for I/O and do bogus-page mapping.
4035 * Assume that vm_page_protect() can block (it can block
4036 * if VM_PROT_NONE, don't take any chances regardless).
4038 * In particular note that for writes we must incorporate
4039 * page dirtyness from the VM system into the buffer's
4042 * For reads we theoretically must incorporate page dirtyness
4043 * from the VM system to determine if the page needs bogus
4044 * replacement, but we shortcut the test by simply checking
4045 * that all m->valid bits are set, indicating that the page
4046 * is fully valid and does not need to be re-read. For any
4047 * VM system dirtyness the page will also be fully valid
4048 * since it was mapped at one point.
4051 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4052 vm_page_t m = bp->b_xio.xio_pages[i];
4054 vm_page_flag_clear(m, PG_ZERO); /* XXX */
4055 if (bp->b_cmd == BUF_CMD_WRITE) {
4057 * When readying a vnode-backed buffer for
4058 * a write we must zero-fill any invalid
4059 * portions of the backing VM pages, mark
4060 * it valid and clear related dirty bits.
4062 * vfs_clean_one_page() incorporates any
4063 * VM dirtyness and updates the b_dirtyoff
4064 * range (after we've made the page RO).
4066 * It is also expected that the pmap modified
4067 * bit has already been cleared by the
4068 * vm_page_protect(). We may not be able
4069 * to clear all dirty bits for a page if it
4070 * was also memory mapped (NFS).
4072 * Finally be sure to unassign any swap-cache
4073 * backing store as it is now stale.
4075 vm_page_protect(m, VM_PROT_READ);
4076 vfs_clean_one_page(bp, i, m);
4077 swap_pager_unswapped(m);
4078 } else if (m->valid == VM_PAGE_BITS_ALL) {
4080 * When readying a vnode-backed buffer for
4081 * read we must replace any dirty pages with
4082 * a bogus page so dirty data is not destroyed
4083 * when filling gaps.
4085 * To avoid testing whether the page is
4086 * dirty we instead test that the page was
4087 * at some point mapped (m->valid fully
4088 * valid) with the understanding that
4089 * this also covers the dirty case.
4091 bp->b_xio.xio_pages[i] = bogus_page;
4092 bp->b_flags |= B_HASBOGUS;
4094 } else if (m->valid & m->dirty) {
4096 * This case should not occur as partial
4097 * dirtyment can only happen if the buffer
4098 * is B_CACHE, and this code is not entered
4099 * if the buffer is B_CACHE.
4101 kprintf("Warning: vfs_busy_pages - page not "
4102 "fully valid! loff=%jx bpf=%08x "
4103 "idx=%d val=%02x dir=%02x\n",
4104 (uintmax_t)bp->b_loffset, bp->b_flags,
4105 i, m->valid, m->dirty);
4106 vm_page_protect(m, VM_PROT_NONE);
4109 * The page is not valid and can be made
4112 vm_page_protect(m, VM_PROT_NONE);
4117 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
4118 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
4123 * This is the easiest place to put the process accounting for the I/O
4127 if (bp->b_cmd == BUF_CMD_READ)
4128 lp->lwp_ru.ru_inblock++;
4130 lp->lwp_ru.ru_oublock++;
4135 * Tell the VM system that the pages associated with this buffer
4136 * are clean. This is used for delayed writes where the data is
4137 * going to go to disk eventually without additional VM intevention.
4139 * NOTE: While we only really need to clean through to b_bcount, we
4140 * just go ahead and clean through to b_bufsize.
4143 vfs_clean_pages(struct buf *bp)
4148 if ((bp->b_flags & B_VMIO) == 0)
4151 KASSERT(bp->b_loffset != NOOFFSET,
4152 ("vfs_clean_pages: no buffer offset"));
4154 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4155 m = bp->b_xio.xio_pages[i];
4156 vfs_clean_one_page(bp, i, m);
4161 * vfs_clean_one_page:
4163 * Set the valid bits and clear the dirty bits in a page within a
4164 * buffer. The range is restricted to the buffer's size and the
4165 * buffer's logical offset might index into the first page.
4167 * The caller has busied or soft-busied the page and it is not mapped,
4168 * test and incorporate the dirty bits into b_dirtyoff/end before
4169 * clearing them. Note that we need to clear the pmap modified bits
4170 * after determining the the page was dirty, vm_page_set_validclean()
4171 * does not do it for us.
4173 * This routine is typically called after a read completes (dirty should
4174 * be zero in that case as we are not called on bogus-replace pages),
4175 * or before a write is initiated.
4178 vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m)
4186 * Calculate offset range within the page but relative to buffer's
4187 * loffset. loffset might be offset into the first page.
4189 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4190 bcount = bp->b_bcount + xoff; /* offset adjusted */
4196 soff = (pageno << PAGE_SHIFT);
4197 eoff = soff + PAGE_SIZE;
4205 * Test dirty bits and adjust b_dirtyoff/end.
4207 * If dirty pages are incorporated into the bp any prior
4208 * B_NEEDCOMMIT state (NFS) must be cleared because the
4209 * caller has not taken into account the new dirty data.
4211 * If the page was memory mapped the dirty bits might go beyond the
4212 * end of the buffer, but we can't really make the assumption that
4213 * a file EOF straddles the buffer (even though this is the case for
4214 * NFS if B_NEEDCOMMIT is also set). So for the purposes of clearing
4215 * B_NEEDCOMMIT we only test the dirty bits covered by the buffer.
4216 * This also saves some console spam.
4218 * When clearing B_NEEDCOMMIT we must also clear B_CLUSTEROK,
4219 * NFS can handle huge commits but not huge writes.
4221 vm_page_test_dirty(m);
4223 if ((bp->b_flags & B_NEEDCOMMIT) &&
4224 (m->dirty & vm_page_bits(soff & PAGE_MASK, eoff - soff))) {
4226 kprintf("Warning: vfs_clean_one_page: bp %p "
4227 "loff=%jx,%d flgs=%08x clr B_NEEDCOMMIT"
4228 " cmd %d vd %02x/%02x x/s/e %d %d %d "
4230 bp, (uintmax_t)bp->b_loffset, bp->b_bcount,
4231 bp->b_flags, bp->b_cmd,
4232 m->valid, m->dirty, xoff, soff, eoff,
4233 bp->b_dirtyoff, bp->b_dirtyend);
4234 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
4236 print_backtrace(-1);
4239 * Only clear the pmap modified bits if ALL the dirty bits
4240 * are set, otherwise the system might mis-clear portions
4243 if (m->dirty == VM_PAGE_BITS_ALL &&
4244 (bp->b_flags & B_NEEDCOMMIT) == 0) {
4245 pmap_clear_modify(m);
4247 if (bp->b_dirtyoff > soff - xoff)
4248 bp->b_dirtyoff = soff - xoff;
4249 if (bp->b_dirtyend < eoff - xoff)
4250 bp->b_dirtyend = eoff - xoff;
4254 * Set related valid bits, clear related dirty bits.
4255 * Does not mess with the pmap modified bit.
4257 * WARNING! We cannot just clear all of m->dirty here as the
4258 * buffer cache buffers may use a DEV_BSIZE'd aligned
4259 * block size, or have an odd size (e.g. NFS at file EOF).
4260 * The putpages code can clear m->dirty to 0.
4262 * If a VOP_WRITE generates a buffer cache buffer which
4263 * covers the same space as mapped writable pages the
4264 * buffer flush might not be able to clear all the dirty
4265 * bits and still require a putpages from the VM system
4268 * WARNING! vm_page_set_validclean() currently assumes vm_token
4269 * is held. The page might not be busied (bdwrite() case).
4270 * XXX remove this comment once we've validated that this
4271 * is no longer an issue.
4273 vm_page_set_validclean(m, soff & PAGE_MASK, eoff - soff);
4278 * Similar to vfs_clean_one_page() but sets the bits to valid and dirty.
4279 * The page data is assumed to be valid (there is no zeroing here).
4282 vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m)
4290 * Calculate offset range within the page but relative to buffer's
4291 * loffset. loffset might be offset into the first page.
4293 xoff = (int)bp->b_loffset & PAGE_MASK; /* loffset offset into pg 0 */
4294 bcount = bp->b_bcount + xoff; /* offset adjusted */
4300 soff = (pageno << PAGE_SHIFT);
4301 eoff = soff + PAGE_SIZE;
4307 vm_page_set_validdirty(m, soff & PAGE_MASK, eoff - soff);
4314 * Clear a buffer. This routine essentially fakes an I/O, so we need
4315 * to clear B_ERROR and B_INVAL.
4317 * Note that while we only theoretically need to clear through b_bcount,
4318 * we go ahead and clear through b_bufsize.
4322 vfs_bio_clrbuf(struct buf *bp)
4326 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
4327 bp->b_flags &= ~(B_INVAL | B_EINTR | B_ERROR);
4328 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
4329 (bp->b_loffset & PAGE_MASK) == 0) {
4330 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
4331 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
4335 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
4336 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
4337 bzero(bp->b_data, bp->b_bufsize);
4338 bp->b_xio.xio_pages[0]->valid |= mask;
4344 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
4345 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
4346 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
4347 ea = (caddr_t)(vm_offset_t)ulmin(
4348 (u_long)(vm_offset_t)ea,
4349 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
4350 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
4351 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
4353 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
4354 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
4358 for (; sa < ea; sa += DEV_BSIZE, j++) {
4359 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
4360 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
4361 bzero(sa, DEV_BSIZE);
4364 bp->b_xio.xio_pages[i]->valid |= mask;
4365 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
4374 * vm_hold_load_pages:
4376 * Load pages into the buffer's address space. The pages are
4377 * allocated from the kernel object in order to reduce interference
4378 * with the any VM paging I/O activity. The range of loaded
4379 * pages will be wired.
4381 * If a page cannot be allocated, the 'pagedaemon' is woken up to
4382 * retrieve the full range (to - from) of pages.
4385 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4391 to = round_page(to);
4392 from = round_page(from);
4393 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4398 * Note: must allocate system pages since blocking here
4399 * could intefere with paging I/O, no matter which
4402 vm_object_hold(&kernel_object);
4403 p = bio_page_alloc(bp, &kernel_object, pg >> PAGE_SHIFT,
4404 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
4405 vm_object_drop(&kernel_object);
4408 p->valid = VM_PAGE_BITS_ALL;
4409 vm_page_flag_clear(p, PG_ZERO);
4410 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
4411 bp->b_xio.xio_pages[index] = p;
4418 bp->b_xio.xio_npages = index;
4422 * Allocate a page for a buffer cache buffer.
4424 * If NULL is returned the caller is expected to retry (typically check if
4425 * the page already exists on retry before trying to allocate one).
4427 * NOTE! Low-memory handling is dealt with in b[q]relse(), not here. This
4428 * function will use the system reserve with the hope that the page
4429 * allocations can be returned to PQ_CACHE/PQ_FREE when the caller
4430 * is done with the buffer.
4432 * NOTE! However, TMPFS is a special case because flushing a dirty buffer
4433 * to TMPFS doesn't clean the page. For TMPFS, only the pagedaemon
4434 * is capable of retiring pages (to swap). For TMPFS we don't dig
4435 * into the system reserve because doing so could stall out pretty
4436 * much every process running on the system.
4440 bio_page_alloc(struct buf *bp, vm_object_t obj, vm_pindex_t pg, int deficit)
4442 int vmflags = VM_ALLOC_NORMAL | VM_ALLOC_NULL_OK;
4445 ASSERT_LWKT_TOKEN_HELD(vm_object_token(obj));
4448 * Try a normal allocation first.
4450 p = vm_page_alloc(obj, pg, vmflags);
4453 if (vm_page_lookup(obj, pg))
4455 vm_pageout_deficit += deficit;
4458 * Try again, digging into the system reserve.
4460 * Trying to recover pages from the buffer cache here can deadlock
4461 * against other threads trying to busy underlying pages so we
4462 * depend on the code in brelse() and bqrelse() to free/cache the
4463 * underlying buffer cache pages when memory is low.
4465 if (curthread->td_flags & TDF_SYSTHREAD)
4466 vmflags |= VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT;
4467 else if (bp->b_vp && bp->b_vp->v_tag == VT_TMPFS)
4470 vmflags |= VM_ALLOC_SYSTEM;
4472 /*recoverbufpages();*/
4473 p = vm_page_alloc(obj, pg, vmflags);
4476 if (vm_page_lookup(obj, pg))
4480 * Wait for memory to free up and try again
4482 if (vm_page_count_severe())
4484 vm_wait(hz / 20 + 1);
4486 p = vm_page_alloc(obj, pg, vmflags);
4489 if (vm_page_lookup(obj, pg))
4493 * Ok, now we are really in trouble.
4496 static struct krate biokrate = { .freq = 1 };
4497 krateprintf(&biokrate,
4498 "Warning: bio_page_alloc: memory exhausted "
4499 "during bufcache page allocation from %s\n",
4500 curthread->td_comm);
4502 if (curthread->td_flags & TDF_SYSTHREAD)
4503 vm_wait(hz / 20 + 1);
4505 vm_wait(hz / 2 + 1);
4510 * vm_hold_free_pages:
4512 * Return pages associated with the buffer back to the VM system.
4514 * The range of pages underlying the buffer's address space will
4515 * be unmapped and un-wired.
4518 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
4522 int index, newnpages;
4524 from = round_page(from);
4525 to = round_page(to);
4526 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
4529 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
4530 p = bp->b_xio.xio_pages[index];
4531 if (p && (index < bp->b_xio.xio_npages)) {
4533 kprintf("vm_hold_free_pages: doffset: %lld, "
4535 (long long)bp->b_bio2.bio_offset,
4536 (long long)bp->b_loffset);
4538 bp->b_xio.xio_pages[index] = NULL;
4540 vm_page_busy_wait(p, FALSE, "vmhldpg");
4541 vm_page_unwire(p, 0);
4545 bp->b_xio.xio_npages = newnpages;
4551 * Map a user buffer into KVM via a pbuf. On return the buffer's
4552 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
4556 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
4567 * bp had better have a command and it better be a pbuf.
4569 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
4570 KKASSERT(bp->b_flags & B_PAGING);
4571 KKASSERT(bp->b_kvabase);
4577 * Map the user data into KVM. Mappings have to be page-aligned.
4579 addr = (caddr_t)trunc_page((vm_offset_t)udata);
4582 vmprot = VM_PROT_READ;
4583 if (bp->b_cmd == BUF_CMD_READ)
4584 vmprot |= VM_PROT_WRITE;
4586 while (addr < udata + bytes) {
4588 * Do the vm_fault if needed; do the copy-on-write thing
4589 * when reading stuff off device into memory.
4591 * vm_fault_page*() returns a held VM page.
4593 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
4594 va = trunc_page(va);
4596 m = vm_fault_page_quick(va, vmprot, &error);
4598 for (i = 0; i < pidx; ++i) {
4599 vm_page_unhold(bp->b_xio.xio_pages[i]);
4600 bp->b_xio.xio_pages[i] = NULL;
4604 bp->b_xio.xio_pages[pidx] = m;
4610 * Map the page array and set the buffer fields to point to
4611 * the mapped data buffer.
4613 if (pidx > btoc(MAXPHYS))
4614 panic("vmapbuf: mapped more than MAXPHYS");
4615 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
4617 bp->b_xio.xio_npages = pidx;
4618 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
4619 bp->b_bcount = bytes;
4620 bp->b_bufsize = bytes;
4627 * Free the io map PTEs associated with this IO operation.
4628 * We also invalidate the TLB entries and restore the original b_addr.
4631 vunmapbuf(struct buf *bp)
4636 KKASSERT(bp->b_flags & B_PAGING);
4638 npages = bp->b_xio.xio_npages;
4639 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
4640 for (pidx = 0; pidx < npages; ++pidx) {
4641 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
4642 bp->b_xio.xio_pages[pidx] = NULL;
4644 bp->b_xio.xio_npages = 0;
4645 bp->b_data = bp->b_kvabase;
4649 * Scan all buffers in the system and issue the callback.
4652 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
4658 for (n = 0; n < nbuf; ++n) {
4659 if ((error = callback(&buf[n], info)) < 0) {
4669 * nestiobuf_iodone: biodone callback for nested buffers and propagate
4670 * completion to the master buffer.
4673 nestiobuf_iodone(struct bio *bio)
4676 struct buf *mbp, *bp;
4677 struct devstat *stats;
4682 mbio = bio->bio_caller_info1.ptr;
4683 stats = bio->bio_caller_info2.ptr;
4684 mbp = mbio->bio_buf;
4686 KKASSERT(bp->b_bcount <= bp->b_bufsize);
4687 KKASSERT(mbp != bp);
4689 error = bp->b_error;
4690 if (bp->b_error == 0 &&
4691 (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
4693 * Not all got transfered, raise an error. We have no way to
4694 * propagate these conditions to mbp.
4699 donebytes = bp->b_bufsize;
4703 nestiobuf_done(mbio, donebytes, error, stats);
4707 nestiobuf_done(struct bio *mbio, int donebytes, int error, struct devstat *stats)
4711 mbp = mbio->bio_buf;
4713 KKASSERT((int)(intptr_t)mbio->bio_driver_info > 0);
4716 * If an error occured, propagate it to the master buffer.
4718 * Several biodone()s may wind up running concurrently so
4719 * use an atomic op to adjust b_flags.
4722 mbp->b_error = error;
4723 atomic_set_int(&mbp->b_flags, B_ERROR);
4727 * Decrement the operations in progress counter and terminate the
4728 * I/O if this was the last bit.
4730 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4733 devstat_end_transaction_buf(stats, mbp);
4739 * Initialize a nestiobuf for use. Set an initial count of 1 to prevent
4740 * the mbio from being biodone()'d while we are still adding sub-bios to
4744 nestiobuf_init(struct bio *bio)
4746 bio->bio_driver_info = (void *)1;
4750 * The BIOs added to the nestedio have already been started, remove the
4751 * count that placeheld our mbio and biodone() it if the count would
4755 nestiobuf_start(struct bio *mbio)
4757 struct buf *mbp = mbio->bio_buf;
4760 * Decrement the operations in progress counter and terminate the
4761 * I/O if this was the last bit.
4763 if (atomic_fetchadd_int((int *)&mbio->bio_driver_info, -1) == 1) {
4764 if (mbp->b_flags & B_ERROR)
4765 mbp->b_resid = mbp->b_bcount;
4773 * Set an intermediate error prior to calling nestiobuf_start()
4776 nestiobuf_error(struct bio *mbio, int error)
4778 struct buf *mbp = mbio->bio_buf;
4781 mbp->b_error = error;
4782 atomic_set_int(&mbp->b_flags, B_ERROR);
4787 * nestiobuf_add: setup a "nested" buffer.
4789 * => 'mbp' is a "master" buffer which is being divided into sub pieces.
4790 * => 'bp' should be a buffer allocated by getiobuf.
4791 * => 'offset' is a byte offset in the master buffer.
4792 * => 'size' is a size in bytes of this nested buffer.
4795 nestiobuf_add(struct bio *mbio, struct buf *bp, int offset, size_t size, struct devstat *stats)
4797 struct buf *mbp = mbio->bio_buf;
4798 struct vnode *vp = mbp->b_vp;
4800 KKASSERT(mbp->b_bcount >= offset + size);
4802 atomic_add_int((int *)&mbio->bio_driver_info, 1);
4804 /* kernel needs to own the lock for it to be released in biodone */
4807 bp->b_cmd = mbp->b_cmd;
4808 bp->b_bio1.bio_done = nestiobuf_iodone;
4809 bp->b_data = (char *)mbp->b_data + offset;
4810 bp->b_resid = bp->b_bcount = size;
4811 bp->b_bufsize = bp->b_bcount;
4813 bp->b_bio1.bio_track = NULL;
4814 bp->b_bio1.bio_caller_info1.ptr = mbio;
4815 bp->b_bio1.bio_caller_info2.ptr = stats;
4820 DB_SHOW_COMMAND(buffer, db_show_buffer)
4823 struct buf *bp = (struct buf *)addr;
4826 db_printf("usage: show buffer <addr>\n");
4830 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
4831 db_printf("b_cmd = %d\n", bp->b_cmd);
4832 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
4833 "b_resid = %d\n, b_data = %p, "
4834 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
4835 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
4837 (long long)bp->b_bio2.bio_offset,
4838 (long long)(bp->b_bio2.bio_next ?
4839 bp->b_bio2.bio_next->bio_offset : (off_t)-1));
4840 if (bp->b_xio.xio_npages) {
4842 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
4843 bp->b_xio.xio_npages);
4844 for (i = 0; i < bp->b_xio.xio_npages; i++) {
4846 m = bp->b_xio.xio_pages[i];
4847 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
4848 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
4849 if ((i + 1) < bp->b_xio.xio_npages)