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 $
15 * $DragonFly: src/sys/kern/vfs_bio.c,v 1.113 2008/07/18 00:01:11 dillon Exp $
19 * this file contains a new buffer I/O scheme implementing a coherent
20 * VM object and buffer cache scheme. Pains have been taken to make
21 * sure that the performance degradation associated with schemes such
22 * as this is not realized.
24 * Author: John S. Dyson
25 * Significant help during the development and debugging phases
26 * had been provided by David Greenman, also of the FreeBSD core team.
28 * see man buf(9) for more info.
31 #include <sys/param.h>
32 #include <sys/systm.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>
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>
58 #include <sys/thread2.h>
59 #include <sys/spinlock2.h>
60 #include <vm/vm_page2.h>
71 BQUEUE_NONE, /* not on any queue */
72 BQUEUE_LOCKED, /* locked buffers */
73 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
74 BQUEUE_DIRTY, /* B_DELWRI buffers */
75 BQUEUE_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
76 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
77 BQUEUE_EMPTY, /* empty buffer headers */
79 BUFFER_QUEUES /* number of buffer queues */
82 typedef enum bufq_type bufq_type_t;
84 #define BD_WAKE_SIZE 128
85 #define BD_WAKE_MASK (BD_WAKE_SIZE - 1)
87 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
89 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
91 struct buf *buf; /* buffer header pool */
93 static void vm_hold_free_pages(struct buf *bp, vm_offset_t from,
95 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
97 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
98 int pageno, vm_page_t m);
99 static void vfs_clean_pages(struct buf *bp);
100 static void vfs_setdirty(struct buf *bp);
101 static void vfs_vmio_release(struct buf *bp);
102 static int flushbufqueues(bufq_type_t q);
103 static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);
105 static void bd_signal(int totalspace);
106 static void buf_daemon(void);
107 static void buf_daemon_hw(void);
110 * bogus page -- for I/O to/from partially complete buffers
111 * this is a temporary solution to the problem, but it is not
112 * really that bad. it would be better to split the buffer
113 * for input in the case of buffers partially already in memory,
114 * but the code is intricate enough already.
116 vm_page_t bogus_page;
119 * These are all static, but make the ones we export globals so we do
120 * not need to use compiler magic.
122 int bufspace, maxbufspace,
123 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
124 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
125 static int lorunningspace, hirunningspace, runningbufreq;
126 int dirtybufspace, dirtybufspacehw, lodirtybufspace, hidirtybufspace;
127 int dirtybufcount, dirtybufcounthw;
128 int runningbufspace, runningbufcount;
129 static int getnewbufcalls;
130 static int getnewbufrestarts;
131 static int recoverbufcalls;
132 static int needsbuffer; /* locked by needsbuffer_spin */
133 static int bd_request; /* locked by needsbuffer_spin */
134 static int bd_request_hw; /* locked by needsbuffer_spin */
135 static u_int bd_wake_ary[BD_WAKE_SIZE];
136 static u_int bd_wake_index;
137 static struct spinlock needsbuffer_spin;
139 static struct thread *bufdaemon_td;
140 static struct thread *bufdaemonhw_td;
144 * Sysctls for operational control of the buffer cache.
146 SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
147 "Number of dirty buffers to flush before bufdaemon becomes inactive");
148 SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
149 "High watermark used to trigger explicit flushing of dirty buffers");
150 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
151 "Minimum amount of buffer space required for active I/O");
152 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
153 "Maximum amount of buffer space to usable for active I/O");
155 * Sysctls determining current state of the buffer cache.
157 SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
158 "Total number of buffers in buffer cache");
159 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
160 "Pending bytes of dirty buffers (all)");
161 SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
162 "Pending bytes of dirty buffers (heavy weight)");
163 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
164 "Pending number of dirty buffers");
165 SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
166 "Pending number of dirty buffers (heavy weight)");
167 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
168 "I/O bytes currently in progress due to asynchronous writes");
169 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
170 "I/O buffers currently in progress due to asynchronous writes");
171 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
172 "Hard limit on maximum amount of memory usable for buffer space");
173 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
174 "Soft limit on maximum amount of memory usable for buffer space");
175 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
176 "Minimum amount of memory to reserve for system buffer space");
177 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
178 "Amount of memory available for buffers");
179 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
180 0, "Maximum amount of memory reserved for buffers using malloc");
181 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
182 "Amount of memory left for buffers using malloc-scheme");
183 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
184 "New buffer header acquisition requests");
185 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
186 0, "New buffer header acquisition restarts");
187 SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
188 "Recover VM space in an emergency");
189 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
190 "Buffer acquisition restarts due to fragmented buffer map");
191 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
192 "Amount of time KVA space was deallocated in an arbitrary buffer");
193 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
194 "Amount of time buffer re-use operations were successful");
195 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
196 "sizeof(struct buf)");
198 char *buf_wmesg = BUF_WMESG;
200 extern int vm_swap_size;
202 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
203 #define VFS_BIO_NEED_UNUSED02 0x02
204 #define VFS_BIO_NEED_UNUSED04 0x04
205 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
210 * Called when buffer space is potentially available for recovery.
211 * getnewbuf() will block on this flag when it is unable to free
212 * sufficient buffer space. Buffer space becomes recoverable when
213 * bp's get placed back in the queues.
220 * If someone is waiting for BUF space, wake them up. Even
221 * though we haven't freed the kva space yet, the waiting
222 * process will be able to now.
224 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
225 spin_lock_wr(&needsbuffer_spin);
226 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
227 spin_unlock_wr(&needsbuffer_spin);
228 wakeup(&needsbuffer);
235 * Accounting for I/O in progress.
239 runningbufwakeup(struct buf *bp)
243 if ((totalspace = bp->b_runningbufspace) != 0) {
244 runningbufspace -= totalspace;
246 bp->b_runningbufspace = 0;
247 if (runningbufreq && runningbufspace <= lorunningspace) {
249 wakeup(&runningbufreq);
251 bd_signal(totalspace);
258 * Called when a buffer has been added to one of the free queues to
259 * account for the buffer and to wakeup anyone waiting for free buffers.
260 * This typically occurs when large amounts of metadata are being handled
261 * by the buffer cache ( else buffer space runs out first, usually ).
268 spin_lock_wr(&needsbuffer_spin);
269 needsbuffer &= ~VFS_BIO_NEED_ANY;
270 spin_unlock_wr(&needsbuffer_spin);
271 wakeup(&needsbuffer);
276 * waitrunningbufspace()
278 * Wait for the amount of running I/O to drop to a reasonable level.
280 * The caller may be using this function to block in a tight loop, we
281 * must block of runningbufspace is greater then the passed limit.
282 * And even with that it may not be enough, due to the presence of
283 * B_LOCKED dirty buffers, so also wait for at least one running buffer
287 waitrunningbufspace(int limit)
291 if (lorunningspace < limit)
292 lorun = lorunningspace;
297 if (runningbufspace > lorun) {
298 while (runningbufspace > lorun) {
300 tsleep(&runningbufreq, 0, "wdrain", 0);
302 } else if (runningbufspace) {
304 tsleep(&runningbufreq, 0, "wdrain2", 1);
310 * vfs_buf_test_cache:
312 * Called when a buffer is extended. This function clears the B_CACHE
313 * bit if the newly extended portion of the buffer does not contain
318 vfs_buf_test_cache(struct buf *bp,
319 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
322 if (bp->b_flags & B_CACHE) {
323 int base = (foff + off) & PAGE_MASK;
324 if (vm_page_is_valid(m, base, size) == 0)
325 bp->b_flags &= ~B_CACHE;
332 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
339 if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
342 if (bd_request == 0 &&
343 (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
344 dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
345 spin_lock_wr(&needsbuffer_spin);
347 spin_unlock_wr(&needsbuffer_spin);
350 if (bd_request_hw == 0 &&
351 (dirtybufspacehw > lodirtybufspace / 2 ||
352 dirtybufcounthw >= nbuf / 2)) {
353 spin_lock_wr(&needsbuffer_spin);
355 spin_unlock_wr(&needsbuffer_spin);
356 wakeup(&bd_request_hw);
363 * Get the buf_daemon heated up when the number of running and dirty
364 * buffers exceeds the mid-point.
373 mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;
375 totalspace = runningbufspace + dirtybufspace;
376 if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
378 mid2 = mid1 + (hidirtybufspace - mid1) / 2;
379 if (totalspace >= mid2)
380 return(totalspace - mid2);
388 * Wait for the buffer cache to flush (totalspace) bytes worth of
389 * buffers, then return.
391 * Regardless this function blocks while the number of dirty buffers
392 * exceeds hidirtybufspace.
395 bd_wait(int totalspace)
400 if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
403 while (totalspace > 0) {
406 if (totalspace > runningbufspace + dirtybufspace)
407 totalspace = runningbufspace + dirtybufspace;
408 count = totalspace / BKVASIZE;
409 if (count >= BD_WAKE_SIZE)
410 count = BD_WAKE_SIZE - 1;
411 i = (bd_wake_index + count) & BD_WAKE_MASK;
413 tsleep(&bd_wake_ary[i], 0, "flstik", hz);
416 totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
423 * This function is called whenever runningbufspace or dirtybufspace
424 * is reduced. Track threads waiting for run+dirty buffer I/O
428 bd_signal(int totalspace)
432 while (totalspace > 0) {
433 i = atomic_fetchadd_int(&bd_wake_index, 1);
435 if (bd_wake_ary[i]) {
437 wakeup(&bd_wake_ary[i]);
439 totalspace -= BKVASIZE;
446 * Load time initialisation of the buffer cache, called from machine
447 * dependant initialization code.
453 vm_offset_t bogus_offset;
456 spin_init(&needsbuffer_spin);
458 /* next, make a null set of free lists */
459 for (i = 0; i < BUFFER_QUEUES; i++)
460 TAILQ_INIT(&bufqueues[i]);
462 /* finally, initialize each buffer header and stick on empty q */
463 for (i = 0; i < nbuf; i++) {
465 bzero(bp, sizeof *bp);
466 bp->b_flags = B_INVAL; /* we're just an empty header */
467 bp->b_cmd = BUF_CMD_DONE;
468 bp->b_qindex = BQUEUE_EMPTY;
470 xio_init(&bp->b_xio);
473 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
477 * maxbufspace is the absolute maximum amount of buffer space we are
478 * allowed to reserve in KVM and in real terms. The absolute maximum
479 * is nominally used by buf_daemon. hibufspace is the nominal maximum
480 * used by most other processes. The differential is required to
481 * ensure that buf_daemon is able to run when other processes might
482 * be blocked waiting for buffer space.
484 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
485 * this may result in KVM fragmentation which is not handled optimally
488 maxbufspace = nbuf * BKVASIZE;
489 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
490 lobufspace = hibufspace - MAXBSIZE;
492 lorunningspace = 512 * 1024;
493 hirunningspace = 1024 * 1024;
496 * Limit the amount of malloc memory since it is wired permanently
497 * into the kernel space. Even though this is accounted for in
498 * the buffer allocation, we don't want the malloced region to grow
499 * uncontrolled. The malloc scheme improves memory utilization
500 * significantly on average (small) directories.
502 maxbufmallocspace = hibufspace / 20;
505 * Reduce the chance of a deadlock occuring by limiting the number
506 * of delayed-write dirty buffers we allow to stack up.
508 hidirtybufspace = hibufspace / 2;
512 lodirtybufspace = hidirtybufspace / 2;
515 * Maximum number of async ops initiated per buf_daemon loop. This is
516 * somewhat of a hack at the moment, we really need to limit ourselves
517 * based on the number of bytes of I/O in-transit that were initiated
521 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
522 bogus_page = vm_page_alloc(&kernel_object,
523 (bogus_offset >> PAGE_SHIFT),
525 vmstats.v_wire_count++;
530 * Initialize the embedded bio structures
533 initbufbio(struct buf *bp)
535 bp->b_bio1.bio_buf = bp;
536 bp->b_bio1.bio_prev = NULL;
537 bp->b_bio1.bio_offset = NOOFFSET;
538 bp->b_bio1.bio_next = &bp->b_bio2;
539 bp->b_bio1.bio_done = NULL;
541 bp->b_bio2.bio_buf = bp;
542 bp->b_bio2.bio_prev = &bp->b_bio1;
543 bp->b_bio2.bio_offset = NOOFFSET;
544 bp->b_bio2.bio_next = NULL;
545 bp->b_bio2.bio_done = NULL;
549 * Reinitialize the embedded bio structures as well as any additional
550 * translation cache layers.
553 reinitbufbio(struct buf *bp)
557 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
558 bio->bio_done = NULL;
559 bio->bio_offset = NOOFFSET;
564 * Push another BIO layer onto an existing BIO and return it. The new
565 * BIO layer may already exist, holding cached translation data.
568 push_bio(struct bio *bio)
572 if ((nbio = bio->bio_next) == NULL) {
573 int index = bio - &bio->bio_buf->b_bio_array[0];
574 if (index >= NBUF_BIO - 1) {
575 panic("push_bio: too many layers bp %p\n",
578 nbio = &bio->bio_buf->b_bio_array[index + 1];
579 bio->bio_next = nbio;
580 nbio->bio_prev = bio;
581 nbio->bio_buf = bio->bio_buf;
582 nbio->bio_offset = NOOFFSET;
583 nbio->bio_done = NULL;
584 nbio->bio_next = NULL;
586 KKASSERT(nbio->bio_done == NULL);
591 * Pop a BIO translation layer, returning the previous layer. The
592 * must have been previously pushed.
595 pop_bio(struct bio *bio)
597 return(bio->bio_prev);
601 clearbiocache(struct bio *bio)
604 bio->bio_offset = NOOFFSET;
612 * Free the KVA allocation for buffer 'bp'.
614 * Must be called from a critical section as this is the only locking for
617 * Since this call frees up buffer space, we call bufspacewakeup().
620 bfreekva(struct buf *bp)
626 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
627 vm_map_lock(&buffer_map);
628 bufspace -= bp->b_kvasize;
629 vm_map_delete(&buffer_map,
630 (vm_offset_t) bp->b_kvabase,
631 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
634 vm_map_unlock(&buffer_map);
635 vm_map_entry_release(count);
644 * Remove the buffer from the appropriate free list.
647 bremfree(struct buf *bp)
652 old_qindex = bp->b_qindex;
654 if (bp->b_qindex != BQUEUE_NONE) {
655 KASSERT(BUF_REFCNTNB(bp) == 1,
656 ("bremfree: bp %p not locked",bp));
657 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
658 bp->b_qindex = BQUEUE_NONE;
660 if (BUF_REFCNTNB(bp) <= 1)
661 panic("bremfree: removing a buffer not on a queue");
671 * Get a buffer with the specified data. Look in the cache first. We
672 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
673 * is set, the buffer is valid and we do not have to do anything ( see
677 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
681 bp = getblk(vp, loffset, size, 0, 0);
684 /* if not found in cache, do some I/O */
685 if ((bp->b_flags & B_CACHE) == 0) {
686 KASSERT(!(bp->b_flags & B_ASYNC),
687 ("bread: illegal async bp %p", bp));
688 bp->b_flags &= ~(B_ERROR | B_INVAL);
689 bp->b_cmd = BUF_CMD_READ;
690 vfs_busy_pages(vp, bp);
691 vn_strategy(vp, &bp->b_bio1);
692 return (biowait(bp));
700 * Operates like bread, but also starts asynchronous I/O on
701 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
702 * to initiating I/O . If B_CACHE is set, the buffer is valid
703 * and we do not have to do anything.
706 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
707 int *rabsize, int cnt, struct buf **bpp)
709 struct buf *bp, *rabp;
711 int rv = 0, readwait = 0;
713 *bpp = bp = getblk(vp, loffset, size, 0, 0);
715 /* if not found in cache, do some I/O */
716 if ((bp->b_flags & B_CACHE) == 0) {
717 bp->b_flags &= ~(B_ERROR | B_INVAL);
718 bp->b_cmd = BUF_CMD_READ;
719 vfs_busy_pages(vp, bp);
720 vn_strategy(vp, &bp->b_bio1);
724 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
725 if (inmem(vp, *raoffset))
727 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
729 if ((rabp->b_flags & B_CACHE) == 0) {
730 rabp->b_flags |= B_ASYNC;
731 rabp->b_flags &= ~(B_ERROR | B_INVAL);
732 rabp->b_cmd = BUF_CMD_READ;
733 vfs_busy_pages(vp, rabp);
735 vn_strategy(vp, &rabp->b_bio1);
750 * Write, release buffer on completion. (Done by iodone
751 * if async). Do not bother writing anything if the buffer
754 * Note that we set B_CACHE here, indicating that buffer is
755 * fully valid and thus cacheable. This is true even of NFS
756 * now so we set it generally. This could be set either here
757 * or in biodone() since the I/O is synchronous. We put it
761 bwrite(struct buf *bp)
765 if (bp->b_flags & B_INVAL) {
770 oldflags = bp->b_flags;
772 if (BUF_REFCNTNB(bp) == 0)
773 panic("bwrite: buffer is not busy???");
776 /* Mark the buffer clean */
779 bp->b_flags &= ~B_ERROR;
780 bp->b_flags |= B_CACHE;
781 bp->b_cmd = BUF_CMD_WRITE;
782 vfs_busy_pages(bp->b_vp, bp);
785 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
786 * valid for vnode-backed buffers.
788 bp->b_runningbufspace = bp->b_bufsize;
789 if (bp->b_runningbufspace) {
790 runningbufspace += bp->b_runningbufspace;
795 if (oldflags & B_ASYNC)
797 vn_strategy(bp->b_vp, &bp->b_bio1);
799 if ((oldflags & B_ASYNC) == 0) {
800 int rtval = biowait(bp);
810 * Delayed write. (Buffer is marked dirty). Do not bother writing
811 * anything if the buffer is marked invalid.
813 * Note that since the buffer must be completely valid, we can safely
814 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
815 * biodone() in order to prevent getblk from writing the buffer
819 bdwrite(struct buf *bp)
821 if (BUF_REFCNTNB(bp) == 0)
822 panic("bdwrite: buffer is not busy");
824 if (bp->b_flags & B_INVAL) {
831 * Set B_CACHE, indicating that the buffer is fully valid. This is
832 * true even of NFS now.
834 bp->b_flags |= B_CACHE;
837 * This bmap keeps the system from needing to do the bmap later,
838 * perhaps when the system is attempting to do a sync. Since it
839 * is likely that the indirect block -- or whatever other datastructure
840 * that the filesystem needs is still in memory now, it is a good
841 * thing to do this. Note also, that if the pageout daemon is
842 * requesting a sync -- there might not be enough memory to do
843 * the bmap then... So, this is important to do.
845 if (bp->b_bio2.bio_offset == NOOFFSET) {
846 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
847 NULL, NULL, BUF_CMD_WRITE);
851 * Set the *dirty* buffer range based upon the VM system dirty pages.
856 * We need to do this here to satisfy the vnode_pager and the
857 * pageout daemon, so that it thinks that the pages have been
858 * "cleaned". Note that since the pages are in a delayed write
859 * buffer -- the VFS layer "will" see that the pages get written
860 * out on the next sync, or perhaps the cluster will be completed.
866 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
867 * due to the softdep code.
874 * Turn buffer into delayed write request by marking it B_DELWRI.
875 * B_RELBUF and B_NOCACHE must be cleared.
877 * We reassign the buffer to itself to properly update it in the
880 * Must be called from a critical section.
881 * The buffer must be on BQUEUE_NONE.
884 bdirty(struct buf *bp)
886 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
887 if (bp->b_flags & B_NOCACHE) {
888 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
889 bp->b_flags &= ~B_NOCACHE;
891 if (bp->b_flags & B_INVAL) {
892 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
894 bp->b_flags &= ~B_RELBUF;
896 if ((bp->b_flags & B_DELWRI) == 0) {
897 bp->b_flags |= B_DELWRI;
900 dirtybufspace += bp->b_bufsize;
901 if (bp->b_flags & B_HEAVY) {
903 dirtybufspacehw += bp->b_bufsize;
910 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
911 * needs to be flushed with a different buf_daemon thread to avoid
912 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
915 bheavy(struct buf *bp)
917 if ((bp->b_flags & B_HEAVY) == 0) {
918 bp->b_flags |= B_HEAVY;
919 if (bp->b_flags & B_DELWRI) {
921 dirtybufspacehw += bp->b_bufsize;
929 * Clear B_DELWRI for buffer.
931 * Must be called from a critical section.
933 * The buffer is typically on BQUEUE_NONE but there is one case in
934 * brelse() that calls this function after placing the buffer on
939 bundirty(struct buf *bp)
941 if (bp->b_flags & B_DELWRI) {
942 bp->b_flags &= ~B_DELWRI;
945 dirtybufspace -= bp->b_bufsize;
946 if (bp->b_flags & B_HEAVY) {
948 dirtybufspacehw -= bp->b_bufsize;
950 bd_signal(bp->b_bufsize);
953 * Since it is now being written, we can clear its deferred write flag.
955 bp->b_flags &= ~B_DEFERRED;
961 * Asynchronous write. Start output on a buffer, but do not wait for
962 * it to complete. The buffer is released when the output completes.
964 * bwrite() ( or the VOP routine anyway ) is responsible for handling
965 * B_INVAL buffers. Not us.
968 bawrite(struct buf *bp)
970 bp->b_flags |= B_ASYNC;
977 * Ordered write. Start output on a buffer, and flag it so that the
978 * device will write it in the order it was queued. The buffer is
979 * released when the output completes. bwrite() ( or the VOP routine
980 * anyway ) is responsible for handling B_INVAL buffers.
983 bowrite(struct buf *bp)
985 bp->b_flags |= B_ORDERED | B_ASYNC;
990 * buf_dirty_count_severe:
992 * Return true if we have too many dirty buffers.
995 buf_dirty_count_severe(void)
997 return (runningbufspace + dirtybufspace >= hidirtybufspace ||
998 dirtybufcount >= nbuf / 2);
1004 * Release a busy buffer and, if requested, free its resources. The
1005 * buffer will be stashed in the appropriate bufqueue[] allowing it
1006 * to be accessed later as a cache entity or reused for other purposes.
1009 brelse(struct buf *bp)
1012 int saved_flags = bp->b_flags;
1015 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1020 * If B_NOCACHE is set we are being asked to destroy the buffer and
1021 * its backing store. Clear B_DELWRI.
1023 * B_NOCACHE is set in two cases: (1) when the caller really wants
1024 * to destroy the buffer and backing store and (2) when the caller
1025 * wants to destroy the buffer and backing store after a write
1028 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1032 if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
1034 * A re-dirtied buffer is only subject to destruction
1035 * by B_INVAL. B_ERROR and B_NOCACHE are ignored.
1037 /* leave buffer intact */
1038 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1039 (bp->b_bufsize <= 0)) {
1041 * Either a failed read or we were asked to free or not
1042 * cache the buffer. This path is reached with B_DELWRI
1043 * set only if B_INVAL is already set. B_NOCACHE governs
1044 * backing store destruction.
1046 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1047 * buffer cannot be immediately freed.
1049 bp->b_flags |= B_INVAL;
1050 if (LIST_FIRST(&bp->b_dep) != NULL)
1052 if (bp->b_flags & B_DELWRI) {
1054 dirtybufspace -= bp->b_bufsize;
1055 if (bp->b_flags & B_HEAVY) {
1057 dirtybufspacehw -= bp->b_bufsize;
1059 bd_signal(bp->b_bufsize);
1061 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1065 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1066 * If vfs_vmio_release() is called with either bit set, the
1067 * underlying pages may wind up getting freed causing a previous
1068 * write (bdwrite()) to get 'lost' because pages associated with
1069 * a B_DELWRI bp are marked clean. Pages associated with a
1070 * B_LOCKED buffer may be mapped by the filesystem.
1072 * If we want to release the buffer ourselves (rather then the
1073 * originator asking us to release it), give the originator a
1074 * chance to countermand the release by setting B_LOCKED.
1076 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1077 * if B_DELWRI is set.
1079 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1080 * on pages to return pages to the VM page queues.
1082 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1083 bp->b_flags &= ~B_RELBUF;
1084 } else if (vm_page_count_severe()) {
1085 if (LIST_FIRST(&bp->b_dep) != NULL)
1086 buf_deallocate(bp); /* can set B_LOCKED */
1087 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1088 bp->b_flags &= ~B_RELBUF;
1090 bp->b_flags |= B_RELBUF;
1094 * Make sure b_cmd is clear. It may have already been cleared by
1097 * At this point destroying the buffer is governed by the B_INVAL
1098 * or B_RELBUF flags.
1100 bp->b_cmd = BUF_CMD_DONE;
1103 * VMIO buffer rundown. Make sure the VM page array is restored
1104 * after an I/O may have replaces some of the pages with bogus pages
1105 * in order to not destroy dirty pages in a fill-in read.
1107 * Note that due to the code above, if a buffer is marked B_DELWRI
1108 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1109 * B_INVAL may still be set, however.
1111 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1112 * but not the backing store. B_NOCACHE will destroy the backing
1115 * Note that dirty NFS buffers contain byte-granular write ranges
1116 * and should not be destroyed w/ B_INVAL even if the backing store
1119 if (bp->b_flags & B_VMIO) {
1121 * Rundown for VMIO buffers which are not dirty NFS buffers.
1133 * Get the base offset and length of the buffer. Note that
1134 * in the VMIO case if the buffer block size is not
1135 * page-aligned then b_data pointer may not be page-aligned.
1136 * But our b_xio.xio_pages array *IS* page aligned.
1138 * block sizes less then DEV_BSIZE (usually 512) are not
1139 * supported due to the page granularity bits (m->valid,
1140 * m->dirty, etc...).
1142 * See man buf(9) for more information
1145 resid = bp->b_bufsize;
1146 foff = bp->b_loffset;
1148 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1149 m = bp->b_xio.xio_pages[i];
1150 vm_page_flag_clear(m, PG_ZERO);
1152 * If we hit a bogus page, fixup *all* of them
1153 * now. Note that we left these pages wired
1154 * when we removed them so they had better exist,
1155 * and they cannot be ripped out from under us so
1156 * no critical section protection is necessary.
1158 if (m == bogus_page) {
1160 poff = OFF_TO_IDX(bp->b_loffset);
1162 for (j = i; j < bp->b_xio.xio_npages; j++) {
1165 mtmp = bp->b_xio.xio_pages[j];
1166 if (mtmp == bogus_page) {
1167 mtmp = vm_page_lookup(obj, poff + j);
1169 panic("brelse: page missing");
1171 bp->b_xio.xio_pages[j] = mtmp;
1175 if ((bp->b_flags & B_INVAL) == 0) {
1176 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1177 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1179 m = bp->b_xio.xio_pages[i];
1183 * Invalidate the backing store if B_NOCACHE is set
1184 * (e.g. used with vinvalbuf()). If this is NFS
1185 * we impose a requirement that the block size be
1186 * a multiple of PAGE_SIZE and create a temporary
1187 * hack to basically invalidate the whole page. The
1188 * problem is that NFS uses really odd buffer sizes
1189 * especially when tracking piecemeal writes and
1190 * it also vinvalbuf()'s a lot, which would result
1191 * in only partial page validation and invalidation
1192 * here. If the file page is mmap()'d, however,
1193 * all the valid bits get set so after we invalidate
1194 * here we would end up with weird m->valid values
1195 * like 0xfc. nfs_getpages() can't handle this so
1196 * we clear all the valid bits for the NFS case
1197 * instead of just some of them.
1199 * The real bug is the VM system having to set m->valid
1200 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1201 * itself is an artifact of the whole 512-byte
1202 * granular mess that exists to support odd block
1203 * sizes and UFS meta-data block sizes (e.g. 6144).
1204 * A complete rewrite is required.
1206 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1207 int poffset = foff & PAGE_MASK;
1210 presid = PAGE_SIZE - poffset;
1211 if (bp->b_vp->v_tag == VT_NFS &&
1212 bp->b_vp->v_type == VREG) {
1214 } else if (presid > resid) {
1217 KASSERT(presid >= 0, ("brelse: extra page"));
1218 vm_page_set_invalid(m, poffset, presid);
1220 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1221 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1223 if (bp->b_flags & (B_INVAL | B_RELBUF))
1224 vfs_vmio_release(bp);
1227 * Rundown for non-VMIO buffers.
1229 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1232 kprintf("brelse bp %p %08x/%08x: Warning, caught and fixed brelvp bug\n", bp, saved_flags, bp->b_flags);
1236 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1242 if (bp->b_qindex != BQUEUE_NONE)
1243 panic("brelse: free buffer onto another queue???");
1244 if (BUF_REFCNTNB(bp) > 1) {
1245 /* Temporary panic to verify exclusive locking */
1246 /* This panic goes away when we allow shared refs */
1247 panic("brelse: multiple refs");
1248 /* do not release to free list */
1255 * Figure out the correct queue to place the cleaned up buffer on.
1256 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1257 * disassociated from their vnode.
1259 if (bp->b_flags & B_LOCKED) {
1261 * Buffers that are locked are placed in the locked queue
1262 * immediately, regardless of their state.
1264 bp->b_qindex = BQUEUE_LOCKED;
1265 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1266 } else if (bp->b_bufsize == 0) {
1268 * Buffers with no memory. Due to conditionals near the top
1269 * of brelse() such buffers should probably already be
1270 * marked B_INVAL and disassociated from their vnode.
1272 bp->b_flags |= B_INVAL;
1273 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1274 KKASSERT((bp->b_flags & B_HASHED) == 0);
1275 if (bp->b_kvasize) {
1276 bp->b_qindex = BQUEUE_EMPTYKVA;
1278 bp->b_qindex = BQUEUE_EMPTY;
1280 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1281 } else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
1283 * Buffers with junk contents. Again these buffers had better
1284 * already be disassociated from their vnode.
1286 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1287 KKASSERT((bp->b_flags & B_HASHED) == 0);
1288 bp->b_flags |= B_INVAL;
1289 bp->b_qindex = BQUEUE_CLEAN;
1290 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1293 * Remaining buffers. These buffers are still associated with
1296 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1298 bp->b_qindex = BQUEUE_DIRTY;
1299 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1301 case B_DELWRI | B_HEAVY:
1302 bp->b_qindex = BQUEUE_DIRTY_HW;
1303 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1308 * NOTE: Buffers are always placed at the end of the
1309 * queue. If B_AGE is not set the buffer will cycle
1310 * through the queue twice.
1312 bp->b_qindex = BQUEUE_CLEAN;
1313 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1319 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1320 * on the correct queue.
1322 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1326 * The bp is on an appropriate queue unless locked. If it is not
1327 * locked or dirty we can wakeup threads waiting for buffer space.
1329 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1330 * if B_INVAL is set ).
1332 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1336 * Something we can maybe free or reuse
1338 if (bp->b_bufsize || bp->b_kvasize)
1342 * Clean up temporary flags and unlock the buffer.
1344 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
1352 * Release a buffer back to the appropriate queue but do not try to free
1353 * it. The buffer is expected to be used again soon.
1355 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1356 * biodone() to requeue an async I/O on completion. It is also used when
1357 * known good buffers need to be requeued but we think we may need the data
1360 * XXX we should be able to leave the B_RELBUF hint set on completion.
1363 bqrelse(struct buf *bp)
1367 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1369 if (bp->b_qindex != BQUEUE_NONE)
1370 panic("bqrelse: free buffer onto another queue???");
1371 if (BUF_REFCNTNB(bp) > 1) {
1372 /* do not release to free list */
1373 panic("bqrelse: multiple refs");
1378 if (bp->b_flags & B_LOCKED) {
1380 * Locked buffers are released to the locked queue. However,
1381 * if the buffer is dirty it will first go into the dirty
1382 * queue and later on after the I/O completes successfully it
1383 * will be released to the locked queue.
1385 bp->b_qindex = BQUEUE_LOCKED;
1386 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1387 } else if (bp->b_flags & B_DELWRI) {
1388 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1389 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1390 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1391 } else if (vm_page_count_severe()) {
1393 * We are too low on memory, we have to try to free the
1394 * buffer (most importantly: the wired pages making up its
1395 * backing store) *now*.
1401 bp->b_qindex = BQUEUE_CLEAN;
1402 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1405 if ((bp->b_flags & B_LOCKED) == 0 &&
1406 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1411 * Something we can maybe free or reuse.
1413 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1417 * Final cleanup and unlock. Clear bits that are only used while a
1418 * buffer is actively locked.
1420 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF);
1428 * Return backing pages held by the buffer 'bp' back to the VM system
1429 * if possible. The pages are freed if they are no longer valid or
1430 * attempt to free if it was used for direct I/O otherwise they are
1431 * sent to the page cache.
1433 * Pages that were marked busy are left alone and skipped.
1435 * The KVA mapping (b_data) for the underlying pages is removed by
1439 vfs_vmio_release(struct buf *bp)
1445 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1446 m = bp->b_xio.xio_pages[i];
1447 bp->b_xio.xio_pages[i] = NULL;
1449 * In order to keep page LRU ordering consistent, put
1450 * everything on the inactive queue.
1452 vm_page_unwire(m, 0);
1454 * We don't mess with busy pages, it is
1455 * the responsibility of the process that
1456 * busied the pages to deal with them.
1458 if ((m->flags & PG_BUSY) || (m->busy != 0))
1461 if (m->wire_count == 0) {
1462 vm_page_flag_clear(m, PG_ZERO);
1464 * Might as well free the page if we can and it has
1465 * no valid data. We also free the page if the
1466 * buffer was used for direct I/O.
1468 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1469 m->hold_count == 0) {
1471 vm_page_protect(m, VM_PROT_NONE);
1473 } else if (bp->b_flags & B_DIRECT) {
1474 vm_page_try_to_free(m);
1475 } else if (vm_page_count_severe()) {
1476 vm_page_try_to_cache(m);
1481 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1482 if (bp->b_bufsize) {
1486 bp->b_xio.xio_npages = 0;
1487 bp->b_flags &= ~B_VMIO;
1488 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1496 * Implement clustered async writes for clearing out B_DELWRI buffers.
1497 * This is much better then the old way of writing only one buffer at
1498 * a time. Note that we may not be presented with the buffers in the
1499 * correct order, so we search for the cluster in both directions.
1501 * The buffer is locked on call.
1504 vfs_bio_awrite(struct buf *bp)
1508 off_t loffset = bp->b_loffset;
1509 struct vnode *vp = bp->b_vp;
1517 * right now we support clustered writing only to regular files. If
1518 * we find a clusterable block we could be in the middle of a cluster
1519 * rather then at the beginning.
1521 * NOTE: b_bio1 contains the logical loffset and is aliased
1522 * to b_loffset. b_bio2 contains the translated block number.
1524 if ((vp->v_type == VREG) &&
1525 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1526 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1528 size = vp->v_mount->mnt_stat.f_iosize;
1530 for (i = size; i < MAXPHYS; i += size) {
1531 if ((bpa = findblk(vp, loffset + i)) &&
1532 BUF_REFCNT(bpa) == 0 &&
1533 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1534 (B_DELWRI | B_CLUSTEROK)) &&
1535 (bpa->b_bufsize == size)) {
1536 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1537 (bpa->b_bio2.bio_offset !=
1538 bp->b_bio2.bio_offset + i))
1544 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1545 if ((bpa = findblk(vp, loffset - j)) &&
1546 BUF_REFCNT(bpa) == 0 &&
1547 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1548 (B_DELWRI | B_CLUSTEROK)) &&
1549 (bpa->b_bufsize == size)) {
1550 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1551 (bpa->b_bio2.bio_offset !=
1552 bp->b_bio2.bio_offset - j))
1561 * this is a possible cluster write
1563 if (nbytes != size) {
1565 nwritten = cluster_wbuild(vp, size,
1566 loffset - j, nbytes);
1573 bp->b_flags |= B_ASYNC;
1577 * default (old) behavior, writing out only one block
1579 * XXX returns b_bufsize instead of b_bcount for nwritten?
1581 nwritten = bp->b_bufsize;
1590 * Find and initialize a new buffer header, freeing up existing buffers
1591 * in the bufqueues as necessary. The new buffer is returned locked.
1593 * Important: B_INVAL is not set. If the caller wishes to throw the
1594 * buffer away, the caller must set B_INVAL prior to calling brelse().
1597 * We have insufficient buffer headers
1598 * We have insufficient buffer space
1599 * buffer_map is too fragmented ( space reservation fails )
1600 * If we have to flush dirty buffers ( but we try to avoid this )
1602 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1603 * Instead we ask the buf daemon to do it for us. We attempt to
1604 * avoid piecemeal wakeups of the pageout daemon.
1608 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1614 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1615 static int flushingbufs;
1618 * We can't afford to block since we might be holding a vnode lock,
1619 * which may prevent system daemons from running. We deal with
1620 * low-memory situations by proactively returning memory and running
1621 * async I/O rather then sync I/O.
1625 --getnewbufrestarts;
1627 ++getnewbufrestarts;
1630 * Setup for scan. If we do not have enough free buffers,
1631 * we setup a degenerate case that immediately fails. Note
1632 * that if we are specially marked process, we are allowed to
1633 * dip into our reserves.
1635 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1637 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1638 * However, there are a number of cases (defragging, reusing, ...)
1639 * where we cannot backup.
1641 nqindex = BQUEUE_EMPTYKVA;
1642 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1646 * If no EMPTYKVA buffers and we are either
1647 * defragging or reusing, locate a CLEAN buffer
1648 * to free or reuse. If bufspace useage is low
1649 * skip this step so we can allocate a new buffer.
1651 if (defrag || bufspace >= lobufspace) {
1652 nqindex = BQUEUE_CLEAN;
1653 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1657 * If we could not find or were not allowed to reuse a
1658 * CLEAN buffer, check to see if it is ok to use an EMPTY
1659 * buffer. We can only use an EMPTY buffer if allocating
1660 * its KVA would not otherwise run us out of buffer space.
1662 if (nbp == NULL && defrag == 0 &&
1663 bufspace + maxsize < hibufspace) {
1664 nqindex = BQUEUE_EMPTY;
1665 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1670 * Run scan, possibly freeing data and/or kva mappings on the fly
1674 while ((bp = nbp) != NULL) {
1675 int qindex = nqindex;
1677 nbp = TAILQ_NEXT(bp, b_freelist);
1680 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1681 * cycles through the queue twice before being selected.
1683 if (qindex == BQUEUE_CLEAN &&
1684 (bp->b_flags & B_AGE) == 0 && nbp) {
1685 bp->b_flags |= B_AGE;
1686 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1687 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1692 * Calculate next bp ( we can only use it if we do not block
1693 * or do other fancy things ).
1698 nqindex = BQUEUE_EMPTYKVA;
1699 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1702 case BQUEUE_EMPTYKVA:
1703 nqindex = BQUEUE_CLEAN;
1704 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1718 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1721 * Note: we no longer distinguish between VMIO and non-VMIO
1725 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1728 * If we are defragging then we need a buffer with
1729 * b_kvasize != 0. XXX this situation should no longer
1730 * occur, if defrag is non-zero the buffer's b_kvasize
1731 * should also be non-zero at this point. XXX
1733 if (defrag && bp->b_kvasize == 0) {
1734 kprintf("Warning: defrag empty buffer %p\n", bp);
1739 * Start freeing the bp. This is somewhat involved. nbp
1740 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1741 * on the clean list must be disassociated from their
1742 * current vnode. Buffers on the empty[kva] lists have
1743 * already been disassociated.
1746 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1747 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1748 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1751 if (bp->b_qindex != qindex) {
1752 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1759 * Dependancies must be handled before we disassociate the
1762 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1763 * be immediately disassociated. HAMMER then becomes
1764 * responsible for releasing the buffer.
1766 if (LIST_FIRST(&bp->b_dep) != NULL) {
1768 if (bp->b_flags & B_LOCKED) {
1772 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1775 if (qindex == BQUEUE_CLEAN) {
1776 if (bp->b_flags & B_VMIO) {
1777 bp->b_flags &= ~B_ASYNC;
1778 vfs_vmio_release(bp);
1785 * NOTE: nbp is now entirely invalid. We can only restart
1786 * the scan from this point on.
1788 * Get the rest of the buffer freed up. b_kva* is still
1789 * valid after this operation.
1792 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1793 KKASSERT((bp->b_flags & B_HASHED) == 0);
1796 * critical section protection is not required when
1797 * scrapping a buffer's contents because it is already
1803 bp->b_flags = B_BNOCLIP;
1804 bp->b_cmd = BUF_CMD_DONE;
1809 bp->b_xio.xio_npages = 0;
1810 bp->b_dirtyoff = bp->b_dirtyend = 0;
1812 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1814 if (blkflags & GETBLK_BHEAVY)
1815 bp->b_flags |= B_HEAVY;
1818 * If we are defragging then free the buffer.
1821 bp->b_flags |= B_INVAL;
1829 * If we are overcomitted then recover the buffer and its
1830 * KVM space. This occurs in rare situations when multiple
1831 * processes are blocked in getnewbuf() or allocbuf().
1833 if (bufspace >= hibufspace)
1835 if (flushingbufs && bp->b_kvasize != 0) {
1836 bp->b_flags |= B_INVAL;
1841 if (bufspace < lobufspace)
1847 * If we exhausted our list, sleep as appropriate. We may have to
1848 * wakeup various daemons and write out some dirty buffers.
1850 * Generally we are sleeping due to insufficient buffer space.
1858 flags = VFS_BIO_NEED_BUFSPACE;
1860 } else if (bufspace >= hibufspace) {
1862 flags = VFS_BIO_NEED_BUFSPACE;
1865 flags = VFS_BIO_NEED_ANY;
1868 needsbuffer |= flags;
1869 bd_speedup(); /* heeeelp */
1870 while (needsbuffer & flags) {
1871 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
1876 * We finally have a valid bp. We aren't quite out of the
1877 * woods, we still have to reserve kva space. In order
1878 * to keep fragmentation sane we only allocate kva in
1881 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1883 if (maxsize != bp->b_kvasize) {
1884 vm_offset_t addr = 0;
1889 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1890 vm_map_lock(&buffer_map);
1892 if (vm_map_findspace(&buffer_map,
1893 vm_map_min(&buffer_map), maxsize,
1896 * Uh oh. Buffer map is too fragmented. We
1897 * must defragment the map.
1899 vm_map_unlock(&buffer_map);
1900 vm_map_entry_release(count);
1903 bp->b_flags |= B_INVAL;
1908 vm_map_insert(&buffer_map, &count,
1910 addr, addr + maxsize,
1912 VM_PROT_ALL, VM_PROT_ALL,
1915 bp->b_kvabase = (caddr_t) addr;
1916 bp->b_kvasize = maxsize;
1917 bufspace += bp->b_kvasize;
1920 vm_map_unlock(&buffer_map);
1921 vm_map_entry_release(count);
1923 bp->b_data = bp->b_kvabase;
1929 * This routine is called in an emergency to recover VM pages from the
1930 * buffer cache by cashing in clean buffers. The idea is to recover
1931 * enough pages to be able to satisfy a stuck bio_page_alloc().
1934 recoverbufpages(void)
1941 while (bytes < MAXBSIZE) {
1942 bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1947 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1948 * cycles through the queue twice before being selected.
1950 if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
1951 bp->b_flags |= B_AGE;
1952 TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1953 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
1961 KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
1962 KKASSERT((bp->b_flags & B_DELWRI) == 0);
1965 * Start freeing the bp. This is somewhat involved.
1967 * Buffers on the clean list must be disassociated from
1968 * their current vnode
1971 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1972 kprintf("recoverbufpages: warning, locked buf %p, race corrected\n", bp);
1973 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1976 if (bp->b_qindex != BQUEUE_CLEAN) {
1977 kprintf("recoverbufpages: warning, BUF_LOCK blocked unexpectedly on buf %p index %d, race corrected\n", bp, bp->b_qindex);
1984 * Dependancies must be handled before we disassociate the
1987 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1988 * be immediately disassociated. HAMMER then becomes
1989 * responsible for releasing the buffer.
1991 if (LIST_FIRST(&bp->b_dep) != NULL) {
1993 if (bp->b_flags & B_LOCKED) {
1997 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2000 bytes += bp->b_bufsize;
2002 if (bp->b_flags & B_VMIO) {
2003 bp->b_flags &= ~B_ASYNC;
2004 bp->b_flags |= B_DIRECT; /* try to free pages */
2005 vfs_vmio_release(bp);
2010 KKASSERT(bp->b_vp == NULL);
2011 KKASSERT((bp->b_flags & B_HASHED) == 0);
2014 * critical section protection is not required when
2015 * scrapping a buffer's contents because it is already
2021 bp->b_flags = B_BNOCLIP;
2022 bp->b_cmd = BUF_CMD_DONE;
2027 bp->b_xio.xio_npages = 0;
2028 bp->b_dirtyoff = bp->b_dirtyend = 0;
2030 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
2032 bp->b_flags |= B_INVAL;
2042 * Buffer flushing daemon. Buffers are normally flushed by the
2043 * update daemon but if it cannot keep up this process starts to
2044 * take the load in an attempt to prevent getnewbuf() from blocking.
2046 * Once a flush is initiated it does not stop until the number
2047 * of buffers falls below lodirtybuffers, but we will wake up anyone
2048 * waiting at the mid-point.
2051 static struct kproc_desc buf_kp = {
2056 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2057 kproc_start, &buf_kp)
2059 static struct kproc_desc bufhw_kp = {
2064 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2065 kproc_start, &bufhw_kp)
2073 * This process needs to be suspended prior to shutdown sync.
2075 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2076 bufdaemon_td, SHUTDOWN_PRI_LAST);
2077 curthread->td_flags |= TDF_SYSTHREAD;
2080 * This process is allowed to take the buffer cache to the limit
2085 kproc_suspend_loop();
2088 * Do the flush. Limit the amount of in-transit I/O we
2089 * allow to build up, otherwise we would completely saturate
2090 * the I/O system. Wakeup any waiting processes before we
2091 * normally would so they can run in parallel with our drain.
2093 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2094 * but because we split the operation into two threads we
2095 * have to cut it in half for each thread.
2097 limit = lodirtybufspace / 2;
2098 waitrunningbufspace(limit);
2099 while (runningbufspace + dirtybufspace > limit ||
2100 dirtybufcount - dirtybufcounthw >= nbuf / 2) {
2101 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2103 waitrunningbufspace(limit);
2107 * We reached our low water mark, reset the
2108 * request and sleep until we are needed again.
2109 * The sleep is just so the suspend code works.
2111 spin_lock_wr(&needsbuffer_spin);
2112 if (bd_request == 0) {
2113 msleep(&bd_request, &needsbuffer_spin, 0,
2117 spin_unlock_wr(&needsbuffer_spin);
2127 * This process needs to be suspended prior to shutdown sync.
2129 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2130 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2131 curthread->td_flags |= TDF_SYSTHREAD;
2134 * This process is allowed to take the buffer cache to the limit
2139 kproc_suspend_loop();
2142 * Do the flush. Limit the amount of in-transit I/O we
2143 * allow to build up, otherwise we would completely saturate
2144 * the I/O system. Wakeup any waiting processes before we
2145 * normally would so they can run in parallel with our drain.
2147 * Our aggregate normal+HW lo water mark is lodirtybufspace,
2148 * but because we split the operation into two threads we
2149 * have to cut it in half for each thread.
2151 limit = lodirtybufspace / 2;
2152 waitrunningbufspace(limit);
2153 while (runningbufspace + dirtybufspacehw > limit ||
2154 dirtybufcounthw >= nbuf / 2) {
2155 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2157 waitrunningbufspace(limit);
2161 * We reached our low water mark, reset the
2162 * request and sleep until we are needed again.
2163 * The sleep is just so the suspend code works.
2165 spin_lock_wr(&needsbuffer_spin);
2166 if (bd_request_hw == 0) {
2167 msleep(&bd_request_hw, &needsbuffer_spin, 0,
2171 spin_unlock_wr(&needsbuffer_spin);
2178 * Try to flush a buffer in the dirty queue. We must be careful to
2179 * free up B_INVAL buffers instead of write them, which NFS is
2180 * particularly sensitive to.
2182 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2183 * that we really want to try to get the buffer out and reuse it
2184 * due to the write load on the machine.
2188 flushbufqueues(bufq_type_t q)
2193 bp = TAILQ_FIRST(&bufqueues[q]);
2195 KASSERT((bp->b_flags & B_DELWRI),
2196 ("unexpected clean buffer %p", bp));
2198 if (bp->b_flags & B_DELWRI) {
2199 if (bp->b_flags & B_INVAL) {
2200 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2201 panic("flushbufqueues: locked buf");
2207 if (LIST_FIRST(&bp->b_dep) != NULL &&
2208 (bp->b_flags & B_DEFERRED) == 0 &&
2209 buf_countdeps(bp, 0)) {
2210 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2211 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2213 bp->b_flags |= B_DEFERRED;
2214 bp = TAILQ_FIRST(&bufqueues[q]);
2219 * Only write it out if we can successfully lock
2220 * it. If the buffer has a dependancy,
2221 * buf_checkwrite must also return 0 for us to
2222 * be able to initate the write.
2224 * If the buffer is flagged B_ERROR it may be
2225 * requeued over and over again, we try to
2226 * avoid a live lock.
2228 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
2229 if (LIST_FIRST(&bp->b_dep) != NULL &&
2230 buf_checkwrite(bp)) {
2233 } else if (bp->b_flags & B_ERROR) {
2234 tsleep(bp, 0, "bioer", 1);
2235 bp->b_flags &= ~B_AGE;
2238 bp->b_flags |= B_AGE;
2245 bp = TAILQ_NEXT(bp, b_freelist);
2253 * Returns true if no I/O is needed to access the associated VM object.
2254 * This is like findblk except it also hunts around in the VM system for
2257 * Note that we ignore vm_page_free() races from interrupts against our
2258 * lookup, since if the caller is not protected our return value will not
2259 * be any more valid then otherwise once we exit the critical section.
2262 inmem(struct vnode *vp, off_t loffset)
2265 vm_offset_t toff, tinc, size;
2268 if (findblk(vp, loffset))
2270 if (vp->v_mount == NULL)
2272 if ((obj = vp->v_object) == NULL)
2276 if (size > vp->v_mount->mnt_stat.f_iosize)
2277 size = vp->v_mount->mnt_stat.f_iosize;
2279 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2280 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2284 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2285 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2286 if (vm_page_is_valid(m,
2287 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2296 * Sets the dirty range for a buffer based on the status of the dirty
2297 * bits in the pages comprising the buffer.
2299 * The range is limited to the size of the buffer.
2301 * This routine is primarily used by NFS, but is generalized for the
2305 vfs_setdirty(struct buf *bp)
2311 * Degenerate case - empty buffer
2314 if (bp->b_bufsize == 0)
2318 * We qualify the scan for modified pages on whether the
2319 * object has been flushed yet. The OBJ_WRITEABLE flag
2320 * is not cleared simply by protecting pages off.
2323 if ((bp->b_flags & B_VMIO) == 0)
2326 object = bp->b_xio.xio_pages[0]->object;
2328 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2329 kprintf("Warning: object %p writeable but not mightbedirty\n", object);
2330 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2331 kprintf("Warning: object %p mightbedirty but not writeable\n", object);
2333 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2334 vm_offset_t boffset;
2335 vm_offset_t eoffset;
2338 * test the pages to see if they have been modified directly
2339 * by users through the VM system.
2341 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2342 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2343 vm_page_test_dirty(bp->b_xio.xio_pages[i]);
2347 * Calculate the encompassing dirty range, boffset and eoffset,
2348 * (eoffset - boffset) bytes.
2351 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2352 if (bp->b_xio.xio_pages[i]->dirty)
2355 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2357 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2358 if (bp->b_xio.xio_pages[i]->dirty) {
2362 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2365 * Fit it to the buffer.
2368 if (eoffset > bp->b_bcount)
2369 eoffset = bp->b_bcount;
2372 * If we have a good dirty range, merge with the existing
2376 if (boffset < eoffset) {
2377 if (bp->b_dirtyoff > boffset)
2378 bp->b_dirtyoff = boffset;
2379 if (bp->b_dirtyend < eoffset)
2380 bp->b_dirtyend = eoffset;
2388 * Locate and return the specified buffer, or NULL if the buffer does
2389 * not exist. Do not attempt to lock the buffer or manipulate it in
2390 * any way. The caller must validate that the correct buffer has been
2391 * obtain after locking it.
2394 findblk(struct vnode *vp, off_t loffset)
2399 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2407 * Get a block given a specified block and offset into a file/device.
2408 * B_INVAL may or may not be set on return. The caller should clear
2409 * B_INVAL prior to initiating a READ.
2411 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2412 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2413 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2414 * without doing any of those things the system will likely believe
2415 * the buffer to be valid (especially if it is not B_VMIO), and the
2416 * next getblk() will return the buffer with B_CACHE set.
2418 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2419 * an existing buffer.
2421 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2422 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2423 * and then cleared based on the backing VM. If the previous buffer is
2424 * non-0-sized but invalid, B_CACHE will be cleared.
2426 * If getblk() must create a new buffer, the new buffer is returned with
2427 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2428 * case it is returned with B_INVAL clear and B_CACHE set based on the
2431 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2432 * B_CACHE bit is clear.
2434 * What this means, basically, is that the caller should use B_CACHE to
2435 * determine whether the buffer is fully valid or not and should clear
2436 * B_INVAL prior to issuing a read. If the caller intends to validate
2437 * the buffer by loading its data area with something, the caller needs
2438 * to clear B_INVAL. If the caller does this without issuing an I/O,
2439 * the caller should set B_CACHE ( as an optimization ), else the caller
2440 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2441 * a write attempt or if it was a successfull read. If the caller
2442 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2443 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2447 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2448 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2451 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2454 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2457 if (size > MAXBSIZE)
2458 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2459 if (vp->v_object == NULL)
2460 panic("getblk: vnode %p has no object!", vp);
2464 if ((bp = findblk(vp, loffset))) {
2466 * The buffer was found in the cache, but we need to lock it.
2467 * Even with LK_NOWAIT the lockmgr may break our critical
2468 * section, so double-check the validity of the buffer
2469 * once the lock has been obtained.
2471 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2472 if (blkflags & GETBLK_NOWAIT) {
2476 int lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2477 if (blkflags & GETBLK_PCATCH)
2478 lkflags |= LK_PCATCH;
2479 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2481 if (error == ENOLCK)
2489 * Once the buffer has been locked, make sure we didn't race
2490 * a buffer recyclement. Buffers that are no longer hashed
2491 * will have b_vp == NULL, so this takes care of that check
2494 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2495 kprintf("Warning buffer %p (vp %p loffset %lld) was recycled\n", bp, vp, loffset);
2501 * If SZMATCH any pre-existing buffer must be of the requested
2502 * size or NULL is returned. The caller absolutely does not
2503 * want getblk() to bwrite() the buffer on a size mismatch.
2505 if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
2512 * All vnode-based buffers must be backed by a VM object.
2514 KKASSERT(bp->b_flags & B_VMIO);
2515 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2516 bp->b_flags &= ~B_AGE;
2519 * Make sure that B_INVAL buffers do not have a cached
2520 * block number translation.
2522 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2523 kprintf("Warning invalid buffer %p (vp %p loffset %lld) did not have cleared bio_offset cache\n", bp, vp, loffset);
2524 clearbiocache(&bp->b_bio2);
2528 * The buffer is locked. B_CACHE is cleared if the buffer is
2531 if (bp->b_flags & B_INVAL)
2532 bp->b_flags &= ~B_CACHE;
2536 * Any size inconsistancy with a dirty buffer or a buffer
2537 * with a softupdates dependancy must be resolved. Resizing
2538 * the buffer in such circumstances can lead to problems.
2540 if (size != bp->b_bcount) {
2541 if (bp->b_flags & B_DELWRI) {
2542 bp->b_flags |= B_NOCACHE;
2544 } else if (LIST_FIRST(&bp->b_dep)) {
2545 bp->b_flags |= B_NOCACHE;
2548 bp->b_flags |= B_RELBUF;
2553 KKASSERT(size <= bp->b_kvasize);
2554 KASSERT(bp->b_loffset != NOOFFSET,
2555 ("getblk: no buffer offset"));
2558 * A buffer with B_DELWRI set and B_CACHE clear must
2559 * be committed before we can return the buffer in
2560 * order to prevent the caller from issuing a read
2561 * ( due to B_CACHE not being set ) and overwriting
2564 * Most callers, including NFS and FFS, need this to
2565 * operate properly either because they assume they
2566 * can issue a read if B_CACHE is not set, or because
2567 * ( for example ) an uncached B_DELWRI might loop due
2568 * to softupdates re-dirtying the buffer. In the latter
2569 * case, B_CACHE is set after the first write completes,
2570 * preventing further loops.
2572 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2573 * above while extending the buffer, we cannot allow the
2574 * buffer to remain with B_CACHE set after the write
2575 * completes or it will represent a corrupt state. To
2576 * deal with this we set B_NOCACHE to scrap the buffer
2579 * We might be able to do something fancy, like setting
2580 * B_CACHE in bwrite() except if B_DELWRI is already set,
2581 * so the below call doesn't set B_CACHE, but that gets real
2582 * confusing. This is much easier.
2585 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2586 bp->b_flags |= B_NOCACHE;
2593 * Buffer is not in-core, create new buffer. The buffer
2594 * returned by getnewbuf() is locked. Note that the returned
2595 * buffer is also considered valid (not marked B_INVAL).
2597 * Calculating the offset for the I/O requires figuring out
2598 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2599 * the mount's f_iosize otherwise. If the vnode does not
2600 * have an associated mount we assume that the passed size is
2603 * Note that vn_isdisk() cannot be used here since it may
2604 * return a failure for numerous reasons. Note that the
2605 * buffer size may be larger then the block size (the caller
2606 * will use block numbers with the proper multiple). Beware
2607 * of using any v_* fields which are part of unions. In
2608 * particular, in DragonFly the mount point overloading
2609 * mechanism uses the namecache only and the underlying
2610 * directory vnode is not a special case.
2614 if (vp->v_type == VBLK || vp->v_type == VCHR)
2616 else if (vp->v_mount)
2617 bsize = vp->v_mount->mnt_stat.f_iosize;
2621 maxsize = size + (loffset & PAGE_MASK);
2622 maxsize = imax(maxsize, bsize);
2624 if ((bp = getnewbuf(blkflags, slptimeo, size, maxsize)) == NULL) {
2625 if (slpflags || slptimeo) {
2633 * This code is used to make sure that a buffer is not
2634 * created while the getnewbuf routine is blocked.
2635 * This can be a problem whether the vnode is locked or not.
2636 * If the buffer is created out from under us, we have to
2637 * throw away the one we just created. There is no window
2638 * race because we are safely running in a critical section
2639 * from the point of the duplicate buffer creation through
2640 * to here, and we've locked the buffer.
2642 if (findblk(vp, loffset)) {
2643 bp->b_flags |= B_INVAL;
2649 * Insert the buffer into the hash, so that it can
2650 * be found by findblk().
2652 * Make sure the translation layer has been cleared.
2654 bp->b_loffset = loffset;
2655 bp->b_bio2.bio_offset = NOOFFSET;
2656 /* bp->b_bio2.bio_next = NULL; */
2661 * All vnode-based buffers must be backed by a VM object.
2663 KKASSERT(vp->v_object != NULL);
2664 bp->b_flags |= B_VMIO;
2665 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2677 * Reacquire a buffer that was previously released to the locked queue,
2678 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2679 * set B_LOCKED (which handles the acquisition race).
2681 * To this end, either B_LOCKED must be set or the dependancy list must be
2685 regetblk(struct buf *bp)
2687 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
2688 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2697 * Get an empty, disassociated buffer of given size. The buffer is
2698 * initially set to B_INVAL.
2700 * critical section protection is not required for the allocbuf()
2701 * call because races are impossible here.
2709 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2712 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2716 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2724 * This code constitutes the buffer memory from either anonymous system
2725 * memory (in the case of non-VMIO operations) or from an associated
2726 * VM object (in the case of VMIO operations). This code is able to
2727 * resize a buffer up or down.
2729 * Note that this code is tricky, and has many complications to resolve
2730 * deadlock or inconsistant data situations. Tread lightly!!!
2731 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2732 * the caller. Calling this code willy nilly can result in the loss of data.
2734 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2735 * B_CACHE for the non-VMIO case.
2737 * This routine does not need to be called from a critical section but you
2738 * must own the buffer.
2741 allocbuf(struct buf *bp, int size)
2743 int newbsize, mbsize;
2746 if (BUF_REFCNT(bp) == 0)
2747 panic("allocbuf: buffer not busy");
2749 if (bp->b_kvasize < size)
2750 panic("allocbuf: buffer too small");
2752 if ((bp->b_flags & B_VMIO) == 0) {
2756 * Just get anonymous memory from the kernel. Don't
2757 * mess with B_CACHE.
2759 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2760 if (bp->b_flags & B_MALLOC)
2763 newbsize = round_page(size);
2765 if (newbsize < bp->b_bufsize) {
2767 * Malloced buffers are not shrunk
2769 if (bp->b_flags & B_MALLOC) {
2771 bp->b_bcount = size;
2773 kfree(bp->b_data, M_BIOBUF);
2774 if (bp->b_bufsize) {
2775 bufmallocspace -= bp->b_bufsize;
2779 bp->b_data = bp->b_kvabase;
2781 bp->b_flags &= ~B_MALLOC;
2787 (vm_offset_t) bp->b_data + newbsize,
2788 (vm_offset_t) bp->b_data + bp->b_bufsize);
2789 } else if (newbsize > bp->b_bufsize) {
2791 * We only use malloced memory on the first allocation.
2792 * and revert to page-allocated memory when the buffer
2795 if ((bufmallocspace < maxbufmallocspace) &&
2796 (bp->b_bufsize == 0) &&
2797 (mbsize <= PAGE_SIZE/2)) {
2799 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
2800 bp->b_bufsize = mbsize;
2801 bp->b_bcount = size;
2802 bp->b_flags |= B_MALLOC;
2803 bufmallocspace += mbsize;
2809 * If the buffer is growing on its other-than-first
2810 * allocation, then we revert to the page-allocation
2813 if (bp->b_flags & B_MALLOC) {
2814 origbuf = bp->b_data;
2815 origbufsize = bp->b_bufsize;
2816 bp->b_data = bp->b_kvabase;
2817 if (bp->b_bufsize) {
2818 bufmallocspace -= bp->b_bufsize;
2822 bp->b_flags &= ~B_MALLOC;
2823 newbsize = round_page(newbsize);
2827 (vm_offset_t) bp->b_data + bp->b_bufsize,
2828 (vm_offset_t) bp->b_data + newbsize);
2830 bcopy(origbuf, bp->b_data, origbufsize);
2831 kfree(origbuf, M_BIOBUF);
2838 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2839 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
2840 newbsize + PAGE_MASK) >> PAGE_SHIFT;
2841 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
2843 if (bp->b_flags & B_MALLOC)
2844 panic("allocbuf: VMIO buffer can't be malloced");
2846 * Set B_CACHE initially if buffer is 0 length or will become
2849 if (size == 0 || bp->b_bufsize == 0)
2850 bp->b_flags |= B_CACHE;
2852 if (newbsize < bp->b_bufsize) {
2854 * DEV_BSIZE aligned new buffer size is less then the
2855 * DEV_BSIZE aligned existing buffer size. Figure out
2856 * if we have to remove any pages.
2858 if (desiredpages < bp->b_xio.xio_npages) {
2859 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
2861 * the page is not freed here -- it
2862 * is the responsibility of
2863 * vnode_pager_setsize
2865 m = bp->b_xio.xio_pages[i];
2866 KASSERT(m != bogus_page,
2867 ("allocbuf: bogus page found"));
2868 while (vm_page_sleep_busy(m, TRUE, "biodep"))
2871 bp->b_xio.xio_pages[i] = NULL;
2872 vm_page_unwire(m, 0);
2874 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2875 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
2876 bp->b_xio.xio_npages = desiredpages;
2878 } else if (size > bp->b_bcount) {
2880 * We are growing the buffer, possibly in a
2881 * byte-granular fashion.
2889 * Step 1, bring in the VM pages from the object,
2890 * allocating them if necessary. We must clear
2891 * B_CACHE if these pages are not valid for the
2892 * range covered by the buffer.
2894 * critical section protection is required to protect
2895 * against interrupts unbusying and freeing pages
2896 * between our vm_page_lookup() and our
2897 * busycheck/wiring call.
2903 while (bp->b_xio.xio_npages < desiredpages) {
2907 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
2908 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2910 * note: must allocate system pages
2911 * since blocking here could intefere
2912 * with paging I/O, no matter which
2915 m = bio_page_alloc(obj, pi, desiredpages - bp->b_xio.xio_npages);
2919 bp->b_flags &= ~B_CACHE;
2920 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2921 ++bp->b_xio.xio_npages;
2927 * We found a page. If we have to sleep on it,
2928 * retry because it might have gotten freed out
2931 * We can only test PG_BUSY here. Blocking on
2932 * m->busy might lead to a deadlock:
2934 * vm_fault->getpages->cluster_read->allocbuf
2938 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
2942 * We have a good page. Should we wakeup the
2945 if ((curthread != pagethread) &&
2946 ((m->queue - m->pc) == PQ_CACHE) &&
2947 ((vmstats.v_free_count + vmstats.v_cache_count) <
2948 (vmstats.v_free_min + vmstats.v_cache_min))) {
2949 pagedaemon_wakeup();
2951 vm_page_flag_clear(m, PG_ZERO);
2953 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2954 ++bp->b_xio.xio_npages;
2959 * Step 2. We've loaded the pages into the buffer,
2960 * we have to figure out if we can still have B_CACHE
2961 * set. Note that B_CACHE is set according to the
2962 * byte-granular range ( bcount and size ), not the
2963 * aligned range ( newbsize ).
2965 * The VM test is against m->valid, which is DEV_BSIZE
2966 * aligned. Needless to say, the validity of the data
2967 * needs to also be DEV_BSIZE aligned. Note that this
2968 * fails with NFS if the server or some other client
2969 * extends the file's EOF. If our buffer is resized,
2970 * B_CACHE may remain set! XXX
2973 toff = bp->b_bcount;
2974 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
2976 while ((bp->b_flags & B_CACHE) && toff < size) {
2979 if (tinc > (size - toff))
2982 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
2990 bp->b_xio.xio_pages[pi]
2997 * Step 3, fixup the KVM pmap. Remember that
2998 * bp->b_data is relative to bp->b_loffset, but
2999 * bp->b_loffset may be offset into the first page.
3002 bp->b_data = (caddr_t)
3003 trunc_page((vm_offset_t)bp->b_data);
3005 (vm_offset_t)bp->b_data,
3006 bp->b_xio.xio_pages,
3007 bp->b_xio.xio_npages
3009 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
3010 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
3014 /* adjust space use on already-dirty buffer */
3015 if (bp->b_flags & B_DELWRI) {
3016 dirtybufspace += newbsize - bp->b_bufsize;
3017 if (bp->b_flags & B_HEAVY)
3018 dirtybufspacehw += newbsize - bp->b_bufsize;
3020 if (newbsize < bp->b_bufsize)
3022 bp->b_bufsize = newbsize; /* actual buffer allocation */
3023 bp->b_bcount = size; /* requested buffer size */
3030 * Wait for buffer I/O completion, returning error status. The buffer
3031 * is left locked on return. B_EINTR is converted into an EINTR error
3034 * NOTE! The original b_cmd is lost on return, since b_cmd will be
3035 * set to BUF_CMD_DONE.
3038 biowait(struct buf *bp)
3041 while (bp->b_cmd != BUF_CMD_DONE) {
3042 if (bp->b_cmd == BUF_CMD_READ)
3043 tsleep(bp, 0, "biord", 0);
3045 tsleep(bp, 0, "biowr", 0);
3048 if (bp->b_flags & B_EINTR) {
3049 bp->b_flags &= ~B_EINTR;
3052 if (bp->b_flags & B_ERROR) {
3053 return (bp->b_error ? bp->b_error : EIO);
3060 * This associates a tracking count with an I/O. vn_strategy() and
3061 * dev_dstrategy() do this automatically but there are a few cases
3062 * where a vnode or device layer is bypassed when a block translation
3063 * is cached. In such cases bio_start_transaction() may be called on
3064 * the bypassed layers so the system gets an I/O in progress indication
3065 * for those higher layers.
3068 bio_start_transaction(struct bio *bio, struct bio_track *track)
3070 bio->bio_track = track;
3071 atomic_add_int(&track->bk_active, 1);
3075 * Initiate I/O on a vnode.
3078 vn_strategy(struct vnode *vp, struct bio *bio)
3080 struct bio_track *track;
3082 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
3083 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
3084 track = &vp->v_track_read;
3086 track = &vp->v_track_write;
3087 bio->bio_track = track;
3088 atomic_add_int(&track->bk_active, 1);
3089 vop_strategy(*vp->v_ops, vp, bio);
3096 * Finish I/O on a buffer, optionally calling a completion function.
3097 * This is usually called from an interrupt so process blocking is
3100 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3101 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3102 * assuming B_INVAL is clear.
3104 * For the VMIO case, we set B_CACHE if the op was a read and no
3105 * read error occured, or if the op was a write. B_CACHE is never
3106 * set if the buffer is invalid or otherwise uncacheable.
3108 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3109 * initiator to leave B_INVAL set to brelse the buffer out of existance
3110 * in the biodone routine.
3113 biodone(struct bio *bio)
3115 struct buf *bp = bio->bio_buf;
3120 KASSERT(BUF_REFCNTNB(bp) > 0,
3121 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3122 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3123 ("biodone: bp %p already done!", bp));
3125 runningbufwakeup(bp);
3128 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3131 biodone_t *done_func;
3132 struct bio_track *track;
3135 * BIO tracking. Most but not all BIOs are tracked.
3137 if ((track = bio->bio_track) != NULL) {
3138 atomic_subtract_int(&track->bk_active, 1);
3139 if (track->bk_active < 0) {
3140 panic("biodone: bad active count bio %p\n",
3143 if (track->bk_waitflag) {
3144 track->bk_waitflag = 0;
3147 bio->bio_track = NULL;
3151 * A bio_done function terminates the loop. The function
3152 * will be responsible for any further chaining and/or
3153 * buffer management.
3155 * WARNING! The done function can deallocate the buffer!
3157 if ((done_func = bio->bio_done) != NULL) {
3158 bio->bio_done = NULL;
3163 bio = bio->bio_prev;
3167 bp->b_cmd = BUF_CMD_DONE;
3170 * Only reads and writes are processed past this point.
3172 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3173 if (cmd == BUF_CMD_FREEBLKS)
3174 bp->b_flags |= B_NOCACHE;
3181 * Warning: softupdates may re-dirty the buffer, and HAMMER can do
3182 * a lot worse. XXX - move this above the clearing of b_cmd
3184 if (LIST_FIRST(&bp->b_dep) != NULL)
3188 * A failed write must re-dirty the buffer unless B_INVAL
3191 if (cmd == BUF_CMD_WRITE &&
3192 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
3193 bp->b_flags &= ~B_NOCACHE;
3198 if (bp->b_flags & B_VMIO) {
3204 struct vnode *vp = bp->b_vp;
3208 #if defined(VFS_BIO_DEBUG)
3209 if (vp->v_auxrefs == 0)
3210 panic("biodone: zero vnode hold count");
3211 if ((vp->v_flag & VOBJBUF) == 0)
3212 panic("biodone: vnode is not setup for merged cache");
3215 foff = bp->b_loffset;
3216 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3217 KASSERT(obj != NULL, ("biodone: missing VM object"));
3219 #if defined(VFS_BIO_DEBUG)
3220 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3221 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3222 obj->paging_in_progress, bp->b_xio.xio_npages);
3227 * Set B_CACHE if the op was a normal read and no error
3228 * occured. B_CACHE is set for writes in the b*write()
3231 iosize = bp->b_bcount - bp->b_resid;
3232 if (cmd == BUF_CMD_READ && (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3233 bp->b_flags |= B_CACHE;
3236 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3240 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3245 * cleanup bogus pages, restoring the originals. Since
3246 * the originals should still be wired, we don't have
3247 * to worry about interrupt/freeing races destroying
3248 * the VM object association.
3250 m = bp->b_xio.xio_pages[i];
3251 if (m == bogus_page) {
3253 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3255 panic("biodone: page disappeared");
3256 bp->b_xio.xio_pages[i] = m;
3257 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3258 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3260 #if defined(VFS_BIO_DEBUG)
3261 if (OFF_TO_IDX(foff) != m->pindex) {
3263 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
3264 (unsigned long)foff, m->pindex);
3269 * In the write case, the valid and clean bits are
3270 * already changed correctly ( see bdwrite() ), so we
3271 * only need to do this here in the read case.
3273 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3274 vfs_page_set_valid(bp, foff, i, m);
3276 vm_page_flag_clear(m, PG_ZERO);
3279 * when debugging new filesystems or buffer I/O methods, this
3280 * is the most common error that pops up. if you see this, you
3281 * have not set the page busy flag correctly!!!
3284 kprintf("biodone: page busy < 0, "
3285 "pindex: %d, foff: 0x(%x,%x), "
3286 "resid: %d, index: %d\n",
3287 (int) m->pindex, (int)(foff >> 32),
3288 (int) foff & 0xffffffff, resid, i);
3289 if (!vn_isdisk(vp, NULL))
3290 kprintf(" iosize: %ld, loffset: %lld, flags: 0x%08x, npages: %d\n",
3291 bp->b_vp->v_mount->mnt_stat.f_iosize,
3293 bp->b_flags, bp->b_xio.xio_npages);
3295 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3297 bp->b_flags, bp->b_xio.xio_npages);
3298 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3299 m->valid, m->dirty, m->wire_count);
3300 panic("biodone: page busy < 0");
3302 vm_page_io_finish(m);
3303 vm_object_pip_subtract(obj, 1);
3304 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3308 vm_object_pip_wakeupn(obj, 0);
3312 * For asynchronous completions, release the buffer now. The brelse
3313 * will do a wakeup there if necessary - so no need to do a wakeup
3314 * here in the async case. The sync case always needs to do a wakeup.
3317 if (bp->b_flags & B_ASYNC) {
3318 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
3331 * This routine is called in lieu of iodone in the case of
3332 * incomplete I/O. This keeps the busy status for pages
3336 vfs_unbusy_pages(struct buf *bp)
3340 runningbufwakeup(bp);
3341 if (bp->b_flags & B_VMIO) {
3342 struct vnode *vp = bp->b_vp;
3347 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3348 vm_page_t m = bp->b_xio.xio_pages[i];
3351 * When restoring bogus changes the original pages
3352 * should still be wired, so we are in no danger of
3353 * losing the object association and do not need
3354 * critical section protection particularly.
3356 if (m == bogus_page) {
3357 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3359 panic("vfs_unbusy_pages: page missing");
3361 bp->b_xio.xio_pages[i] = m;
3362 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3363 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3365 vm_object_pip_subtract(obj, 1);
3366 vm_page_flag_clear(m, PG_ZERO);
3367 vm_page_io_finish(m);
3369 vm_object_pip_wakeupn(obj, 0);
3374 * vfs_page_set_valid:
3376 * Set the valid bits in a page based on the supplied offset. The
3377 * range is restricted to the buffer's size.
3379 * This routine is typically called after a read completes.
3382 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3384 vm_ooffset_t soff, eoff;
3387 * Start and end offsets in buffer. eoff - soff may not cross a
3388 * page boundry or cross the end of the buffer. The end of the
3389 * buffer, in this case, is our file EOF, not the allocation size
3393 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3394 if (eoff > bp->b_loffset + bp->b_bcount)
3395 eoff = bp->b_loffset + bp->b_bcount;
3398 * Set valid range. This is typically the entire buffer and thus the
3402 vm_page_set_validclean(
3404 (vm_offset_t) (soff & PAGE_MASK),
3405 (vm_offset_t) (eoff - soff)
3413 * This routine is called before a device strategy routine.
3414 * It is used to tell the VM system that paging I/O is in
3415 * progress, and treat the pages associated with the buffer
3416 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3417 * flag is handled to make sure that the object doesn't become
3420 * Since I/O has not been initiated yet, certain buffer flags
3421 * such as B_ERROR or B_INVAL may be in an inconsistant state
3422 * and should be ignored.
3425 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3428 struct lwp *lp = curthread->td_lwp;
3431 * The buffer's I/O command must already be set. If reading,
3432 * B_CACHE must be 0 (double check against callers only doing
3433 * I/O when B_CACHE is 0).
3435 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3436 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3438 if (bp->b_flags & B_VMIO) {
3443 foff = bp->b_loffset;
3444 KASSERT(bp->b_loffset != NOOFFSET,
3445 ("vfs_busy_pages: no buffer offset"));
3449 * Loop until none of the pages are busy.
3452 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3453 vm_page_t m = bp->b_xio.xio_pages[i];
3455 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3460 * Setup for I/O, soft-busy the page right now because
3461 * the next loop may block.
3463 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3464 vm_page_t m = bp->b_xio.xio_pages[i];
3466 vm_page_flag_clear(m, PG_ZERO);
3467 if ((bp->b_flags & B_CLUSTER) == 0) {
3468 vm_object_pip_add(obj, 1);
3469 vm_page_io_start(m);
3474 * Adjust protections for I/O and do bogus-page mapping.
3475 * Assume that vm_page_protect() can block (it can block
3476 * if VM_PROT_NONE, don't take any chances regardless).
3479 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3480 vm_page_t m = bp->b_xio.xio_pages[i];
3483 * When readying a vnode-backed buffer for a write
3484 * we must zero-fill any invalid portions of the
3487 * When readying a vnode-backed buffer for a read
3488 * we must replace any dirty pages with a bogus
3489 * page so we do not destroy dirty data when
3490 * filling in gaps. Dirty pages might not
3491 * necessarily be marked dirty yet, so use m->valid
3492 * as a reasonable test.
3494 * Bogus page replacement is, uh, bogus. We need
3495 * to find a better way.
3497 if (bp->b_cmd == BUF_CMD_WRITE) {
3498 vm_page_protect(m, VM_PROT_READ);
3499 vfs_page_set_valid(bp, foff, i, m);
3500 } else if (m->valid == VM_PAGE_BITS_ALL) {
3501 bp->b_xio.xio_pages[i] = bogus_page;
3504 vm_page_protect(m, VM_PROT_NONE);
3506 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3509 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3510 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3514 * This is the easiest place to put the process accounting for the I/O
3518 if (bp->b_cmd == BUF_CMD_READ)
3519 lp->lwp_ru.ru_inblock++;
3521 lp->lwp_ru.ru_oublock++;
3528 * Tell the VM system that the pages associated with this buffer
3529 * are clean. This is used for delayed writes where the data is
3530 * going to go to disk eventually without additional VM intevention.
3532 * Note that while we only really need to clean through to b_bcount, we
3533 * just go ahead and clean through to b_bufsize.
3536 vfs_clean_pages(struct buf *bp)
3540 if (bp->b_flags & B_VMIO) {
3543 foff = bp->b_loffset;
3544 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset"));
3545 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3546 vm_page_t m = bp->b_xio.xio_pages[i];
3547 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3548 vm_ooffset_t eoff = noff;
3550 if (eoff > bp->b_loffset + bp->b_bufsize)
3551 eoff = bp->b_loffset + bp->b_bufsize;
3552 vfs_page_set_valid(bp, foff, i, m);
3553 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3560 * vfs_bio_set_validclean:
3562 * Set the range within the buffer to valid and clean. The range is
3563 * relative to the beginning of the buffer, b_loffset. Note that
3564 * b_loffset itself may be offset from the beginning of the first page.
3568 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3570 if (bp->b_flags & B_VMIO) {
3575 * Fixup base to be relative to beginning of first page.
3576 * Set initial n to be the maximum number of bytes in the
3577 * first page that can be validated.
3580 base += (bp->b_loffset & PAGE_MASK);
3581 n = PAGE_SIZE - (base & PAGE_MASK);
3583 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) {
3584 vm_page_t m = bp->b_xio.xio_pages[i];
3589 vm_page_set_validclean(m, base & PAGE_MASK, n);
3600 * Clear a buffer. This routine essentially fakes an I/O, so we need
3601 * to clear B_ERROR and B_INVAL.
3603 * Note that while we only theoretically need to clear through b_bcount,
3604 * we go ahead and clear through b_bufsize.
3608 vfs_bio_clrbuf(struct buf *bp)
3612 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3613 bp->b_flags &= ~(B_INVAL|B_ERROR);
3614 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3615 (bp->b_loffset & PAGE_MASK) == 0) {
3616 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3617 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3621 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3622 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3623 bzero(bp->b_data, bp->b_bufsize);
3624 bp->b_xio.xio_pages[0]->valid |= mask;
3629 ea = sa = bp->b_data;
3630 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3631 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3632 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3633 ea = (caddr_t)(vm_offset_t)ulmin(
3634 (u_long)(vm_offset_t)ea,
3635 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3636 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3637 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3639 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3640 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3644 for (; sa < ea; sa += DEV_BSIZE, j++) {
3645 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3646 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3647 bzero(sa, DEV_BSIZE);
3650 bp->b_xio.xio_pages[i]->valid |= mask;
3651 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
3660 * vm_hold_load_pages:
3662 * Load pages into the buffer's address space. The pages are
3663 * allocated from the kernel object in order to reduce interference
3664 * with the any VM paging I/O activity. The range of loaded
3665 * pages will be wired.
3667 * If a page cannot be allocated, the 'pagedaemon' is woken up to
3668 * retrieve the full range (to - from) of pages.
3672 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3678 to = round_page(to);
3679 from = round_page(from);
3680 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3685 * Note: must allocate system pages since blocking here
3686 * could intefere with paging I/O, no matter which
3689 p = bio_page_alloc(&kernel_object, pg >> PAGE_SHIFT,
3690 (vm_pindex_t)((to - pg) >> PAGE_SHIFT));
3693 p->valid = VM_PAGE_BITS_ALL;
3694 vm_page_flag_clear(p, PG_ZERO);
3695 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3696 bp->b_xio.xio_pages[index] = p;
3703 bp->b_xio.xio_npages = index;
3707 * Allocate pages for a buffer cache buffer.
3709 * Under extremely severe memory conditions even allocating out of the
3710 * system reserve can fail. If this occurs we must allocate out of the
3711 * interrupt reserve to avoid a deadlock with the pageout daemon.
3713 * The pageout daemon can run (putpages -> VOP_WRITE -> getblk -> allocbuf).
3714 * If the buffer cache's vm_page_alloc() fails a vm_wait() can deadlock
3715 * against the pageout daemon if pages are not freed from other sources.
3719 bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit)
3724 * Try a normal allocation, allow use of system reserve.
3726 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
3731 * The normal allocation failed and we clearly have a page
3732 * deficit. Try to reclaim some clean VM pages directly
3733 * from the buffer cache.
3735 vm_pageout_deficit += deficit;
3739 * We may have blocked, the caller will know what to do if the
3742 if (vm_page_lookup(obj, pg))
3746 * Allocate and allow use of the interrupt reserve.
3748 * If after all that we still can't allocate a VM page we are
3749 * in real trouble, but we slog on anyway hoping that the system
3752 p = vm_page_alloc(obj, pg, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
3753 VM_ALLOC_INTERRUPT);
3755 if (vm_page_count_severe()) {
3756 kprintf("bio_page_alloc: WARNING emergency page "
3761 kprintf("bio_page_alloc: WARNING emergency page "
3762 "allocation failed\n");
3769 * vm_hold_free_pages:
3771 * Return pages associated with the buffer back to the VM system.
3773 * The range of pages underlying the buffer's address space will
3774 * be unmapped and un-wired.
3777 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3781 int index, newnpages;
3783 from = round_page(from);
3784 to = round_page(to);
3785 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3787 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3788 p = bp->b_xio.xio_pages[index];
3789 if (p && (index < bp->b_xio.xio_npages)) {
3791 kprintf("vm_hold_free_pages: doffset: %lld, loffset: %lld\n",
3792 bp->b_bio2.bio_offset, bp->b_loffset);
3794 bp->b_xio.xio_pages[index] = NULL;
3797 vm_page_unwire(p, 0);
3801 bp->b_xio.xio_npages = newnpages;
3807 * Map a user buffer into KVM via a pbuf. On return the buffer's
3808 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
3812 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
3823 * bp had better have a command and it better be a pbuf.
3825 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3826 KKASSERT(bp->b_flags & B_PAGING);
3832 * Map the user data into KVM. Mappings have to be page-aligned.
3834 addr = (caddr_t)trunc_page((vm_offset_t)udata);
3837 vmprot = VM_PROT_READ;
3838 if (bp->b_cmd == BUF_CMD_READ)
3839 vmprot |= VM_PROT_WRITE;
3841 while (addr < udata + bytes) {
3843 * Do the vm_fault if needed; do the copy-on-write thing
3844 * when reading stuff off device into memory.
3846 * vm_fault_page*() returns a held VM page.
3848 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
3849 va = trunc_page(va);
3851 m = vm_fault_page_quick(va, vmprot, &error);
3853 for (i = 0; i < pidx; ++i) {
3854 vm_page_unhold(bp->b_xio.xio_pages[i]);
3855 bp->b_xio.xio_pages[i] = NULL;
3859 bp->b_xio.xio_pages[pidx] = m;
3865 * Map the page array and set the buffer fields to point to
3866 * the mapped data buffer.
3868 if (pidx > btoc(MAXPHYS))
3869 panic("vmapbuf: mapped more than MAXPHYS");
3870 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
3872 bp->b_xio.xio_npages = pidx;
3873 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
3874 bp->b_bcount = bytes;
3875 bp->b_bufsize = bytes;
3882 * Free the io map PTEs associated with this IO operation.
3883 * We also invalidate the TLB entries and restore the original b_addr.
3886 vunmapbuf(struct buf *bp)
3891 KKASSERT(bp->b_flags & B_PAGING);
3893 npages = bp->b_xio.xio_npages;
3894 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3895 for (pidx = 0; pidx < npages; ++pidx) {
3896 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
3897 bp->b_xio.xio_pages[pidx] = NULL;
3899 bp->b_xio.xio_npages = 0;
3900 bp->b_data = bp->b_kvabase;
3904 * Scan all buffers in the system and issue the callback.
3907 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
3913 for (n = 0; n < nbuf; ++n) {
3914 if ((error = callback(&buf[n], info)) < 0) {
3924 * print out statistics from the current status of the buffer pool
3925 * this can be toggeled by the system control option debug.syncprt
3934 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
3935 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
3937 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
3939 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
3942 TAILQ_FOREACH(bp, dp, b_freelist) {
3943 counts[bp->b_bufsize/PAGE_SIZE]++;
3947 kprintf("%s: total-%d", bname[i], count);
3948 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
3950 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
3958 DB_SHOW_COMMAND(buffer, db_show_buffer)
3961 struct buf *bp = (struct buf *)addr;
3964 db_printf("usage: show buffer <addr>\n");
3968 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3969 db_printf("b_cmd = %d\n", bp->b_cmd);
3970 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
3971 "b_resid = %d\n, b_data = %p, "
3972 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
3973 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3974 bp->b_data, bp->b_bio2.bio_offset, (bp->b_bio2.bio_next ? bp->b_bio2.bio_next->bio_offset : (off_t)-1));
3975 if (bp->b_xio.xio_npages) {
3977 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
3978 bp->b_xio.xio_npages);
3979 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3981 m = bp->b_xio.xio_pages[i];
3982 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3983 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3984 if ((i + 1) < bp->b_xio.xio_npages)