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 1.67 2006/04/30 17:22:17 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 <vm/vm_page2.h>
64 #define BUFFER_QUEUES 6
66 BQUEUE_NONE, /* not on any queue */
67 BQUEUE_LOCKED, /* locked buffers */
68 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
69 BQUEUE_DIRTY, /* B_DELWRI buffers */
70 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
71 BQUEUE_EMPTY /* empty buffer headers */
73 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
75 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
77 struct bio_ops bioops; /* I/O operation notification */
79 struct buf *buf; /* buffer header pool */
81 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from,
83 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from,
85 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
86 int pageno, vm_page_t m);
87 static void vfs_clean_pages(struct buf * bp);
88 static void vfs_setdirty(struct buf *bp);
89 static void vfs_vmio_release(struct buf *bp);
90 static int flushbufqueues(void);
92 static int bd_request;
94 static void buf_daemon (void);
96 * bogus page -- for I/O to/from partially complete buffers
97 * this is a temporary solution to the problem, but it is not
98 * really that bad. it would be better to split the buffer
99 * for input in the case of buffers partially already in memory,
100 * but the code is intricate enough already.
102 vm_page_t bogus_page;
105 static int bufspace, maxbufspace,
106 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
107 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
108 static int needsbuffer;
109 static int lorunningspace, hirunningspace, runningbufreq;
110 static int numdirtybuffers, lodirtybuffers, hidirtybuffers;
111 static int numfreebuffers, lofreebuffers, hifreebuffers;
112 static int getnewbufcalls;
113 static int getnewbufrestarts;
116 * Sysctls for operational control of the buffer cache.
118 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
119 "Number of dirty buffers to flush before bufdaemon becomes inactive");
120 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
121 "High watermark used to trigger explicit flushing of dirty buffers");
122 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
123 "Low watermark for special reserve in low-memory situations");
124 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
125 "High watermark for special reserve in low-memory situations");
126 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
127 "Minimum amount of buffer space required for active I/O");
128 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
129 "Maximum amount of buffer space to usable for active I/O");
131 * Sysctls determining current state of the buffer cache.
133 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
134 "Pending number of dirty buffers");
135 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
136 "Number of free buffers on the buffer cache free list");
137 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
138 "I/O bytes currently in progress due to asynchronous writes");
139 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
140 "Hard limit on maximum amount of memory usable for buffer space");
141 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
142 "Soft limit on maximum amount of memory usable for buffer space");
143 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
144 "Minimum amount of memory to reserve for system buffer space");
145 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
146 "Amount of memory available for buffers");
147 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
148 0, "Maximum amount of memory reserved for buffers using malloc");
149 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
150 "Amount of memory left for buffers using malloc-scheme");
151 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
152 "New buffer header acquisition requests");
153 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
154 0, "New buffer header acquisition restarts");
155 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
156 "Buffer acquisition restarts due to fragmented buffer map");
157 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
158 "Amount of time KVA space was deallocated in an arbitrary buffer");
159 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
160 "Amount of time buffer re-use operations were successful");
161 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
162 "sizeof(struct buf)");
164 char *buf_wmesg = BUF_WMESG;
166 extern int vm_swap_size;
168 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
169 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
170 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
171 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
176 * If someone is blocked due to there being too many dirty buffers,
177 * and numdirtybuffers is now reasonable, wake them up.
181 numdirtywakeup(int level)
183 if (numdirtybuffers <= level) {
184 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
185 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
186 wakeup(&needsbuffer);
194 * Called when buffer space is potentially available for recovery.
195 * getnewbuf() will block on this flag when it is unable to free
196 * sufficient buffer space. Buffer space becomes recoverable when
197 * bp's get placed back in the queues.
204 * If someone is waiting for BUF space, wake them up. Even
205 * though we haven't freed the kva space yet, the waiting
206 * process will be able to now.
208 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
209 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
210 wakeup(&needsbuffer);
217 * Accounting for I/O in progress.
221 runningbufwakeup(struct buf *bp)
223 if (bp->b_runningbufspace) {
224 runningbufspace -= bp->b_runningbufspace;
225 bp->b_runningbufspace = 0;
226 if (runningbufreq && runningbufspace <= lorunningspace) {
228 wakeup(&runningbufreq);
236 * Called when a buffer has been added to one of the free queues to
237 * account for the buffer and to wakeup anyone waiting for free buffers.
238 * This typically occurs when large amounts of metadata are being handled
239 * by the buffer cache ( else buffer space runs out first, usually ).
247 needsbuffer &= ~VFS_BIO_NEED_ANY;
248 if (numfreebuffers >= hifreebuffers)
249 needsbuffer &= ~VFS_BIO_NEED_FREE;
250 wakeup(&needsbuffer);
255 * waitrunningbufspace()
257 * runningbufspace is a measure of the amount of I/O currently
258 * running. This routine is used in async-write situations to
259 * prevent creating huge backups of pending writes to a device.
260 * Only asynchronous writes are governed by this function.
262 * Reads will adjust runningbufspace, but will not block based on it.
263 * The read load has a side effect of reducing the allowed write load.
265 * This does NOT turn an async write into a sync write. It waits
266 * for earlier writes to complete and generally returns before the
267 * caller's write has reached the device.
270 waitrunningbufspace(void)
272 if (runningbufspace > hirunningspace) {
274 while (runningbufspace > hirunningspace) {
276 tsleep(&runningbufreq, 0, "wdrain", 0);
283 * vfs_buf_test_cache:
285 * Called when a buffer is extended. This function clears the B_CACHE
286 * bit if the newly extended portion of the buffer does not contain
291 vfs_buf_test_cache(struct buf *bp,
292 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
295 if (bp->b_flags & B_CACHE) {
296 int base = (foff + off) & PAGE_MASK;
297 if (vm_page_is_valid(m, base, size) == 0)
298 bp->b_flags &= ~B_CACHE;
305 * Wake up the buffer daemon if the number of outstanding dirty buffers
306 * is above specified threshold 'dirtybuflevel'.
308 * The buffer daemon is explicitly woken up when (a) the pending number
309 * of dirty buffers exceeds the recovery and stall mid-point value,
310 * (b) during bwillwrite() or (c) buf freelist was exhausted.
314 bd_wakeup(int dirtybuflevel)
316 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
325 * Speed up the buffer cache flushing process.
338 * Load time initialisation of the buffer cache, called from machine
339 * dependant initialization code.
345 vm_offset_t bogus_offset;
348 /* next, make a null set of free lists */
349 for (i = 0; i < BUFFER_QUEUES; i++)
350 TAILQ_INIT(&bufqueues[i]);
352 /* finally, initialize each buffer header and stick on empty q */
353 for (i = 0; i < nbuf; i++) {
355 bzero(bp, sizeof *bp);
356 bp->b_flags = B_INVAL; /* we're just an empty header */
357 bp->b_cmd = BUF_CMD_DONE;
358 bp->b_qindex = BQUEUE_EMPTY;
360 xio_init(&bp->b_xio);
361 LIST_INIT(&bp->b_dep);
363 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
367 * maxbufspace is the absolute maximum amount of buffer space we are
368 * allowed to reserve in KVM and in real terms. The absolute maximum
369 * is nominally used by buf_daemon. hibufspace is the nominal maximum
370 * used by most other processes. The differential is required to
371 * ensure that buf_daemon is able to run when other processes might
372 * be blocked waiting for buffer space.
374 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
375 * this may result in KVM fragmentation which is not handled optimally
378 maxbufspace = nbuf * BKVASIZE;
379 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
380 lobufspace = hibufspace - MAXBSIZE;
382 lorunningspace = 512 * 1024;
383 hirunningspace = 1024 * 1024;
386 * Limit the amount of malloc memory since it is wired permanently into
387 * the kernel space. Even though this is accounted for in the buffer
388 * allocation, we don't want the malloced region to grow uncontrolled.
389 * The malloc scheme improves memory utilization significantly on average
390 * (small) directories.
392 maxbufmallocspace = hibufspace / 20;
395 * Reduce the chance of a deadlock occuring by limiting the number
396 * of delayed-write dirty buffers we allow to stack up.
398 hidirtybuffers = nbuf / 4 + 20;
401 * To support extreme low-memory systems, make sure hidirtybuffers cannot
402 * eat up all available buffer space. This occurs when our minimum cannot
403 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
404 * BKVASIZE'd (8K) buffers.
406 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
407 hidirtybuffers >>= 1;
409 lodirtybuffers = hidirtybuffers / 2;
412 * Try to keep the number of free buffers in the specified range,
413 * and give special processes (e.g. like buf_daemon) access to an
416 lofreebuffers = nbuf / 18 + 5;
417 hifreebuffers = 2 * lofreebuffers;
418 numfreebuffers = nbuf;
421 * Maximum number of async ops initiated per buf_daemon loop. This is
422 * somewhat of a hack at the moment, we really need to limit ourselves
423 * based on the number of bytes of I/O in-transit that were initiated
427 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE);
428 bogus_page = vm_page_alloc(kernel_object,
429 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
431 vmstats.v_wire_count++;
436 * Initialize the embedded bio structures
439 initbufbio(struct buf *bp)
441 bp->b_bio1.bio_buf = bp;
442 bp->b_bio1.bio_prev = NULL;
443 bp->b_bio1.bio_offset = NOOFFSET;
444 bp->b_bio1.bio_next = &bp->b_bio2;
445 bp->b_bio1.bio_done = NULL;
447 bp->b_bio2.bio_buf = bp;
448 bp->b_bio2.bio_prev = &bp->b_bio1;
449 bp->b_bio2.bio_offset = NOOFFSET;
450 bp->b_bio2.bio_next = NULL;
451 bp->b_bio2.bio_done = NULL;
455 * Reinitialize the embedded bio structures as well as any additional
456 * translation cache layers.
459 reinitbufbio(struct buf *bp)
463 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
464 bio->bio_done = NULL;
465 bio->bio_offset = NOOFFSET;
470 * Push another BIO layer onto an existing BIO and return it. The new
471 * BIO layer may already exist, holding cached translation data.
474 push_bio(struct bio *bio)
478 if ((nbio = bio->bio_next) == NULL) {
479 int index = bio - &bio->bio_buf->b_bio_array[0];
480 if (index >= NBUF_BIO) {
481 panic("push_bio: too many layers bp %p\n",
484 nbio = &bio->bio_buf->b_bio_array[index + 1];
485 bio->bio_next = nbio;
486 nbio->bio_prev = bio;
487 nbio->bio_buf = bio->bio_buf;
488 nbio->bio_offset = NOOFFSET;
489 nbio->bio_done = NULL;
490 nbio->bio_next = NULL;
492 KKASSERT(nbio->bio_done == NULL);
497 pop_bio(struct bio *bio)
503 clearbiocache(struct bio *bio)
506 bio->bio_offset = NOOFFSET;
514 * Free the KVA allocation for buffer 'bp'.
516 * Must be called from a critical section as this is the only locking for
519 * Since this call frees up buffer space, we call bufspacewakeup().
522 bfreekva(struct buf * bp)
528 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
529 vm_map_lock(buffer_map);
530 bufspace -= bp->b_kvasize;
531 vm_map_delete(buffer_map,
532 (vm_offset_t) bp->b_kvabase,
533 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
536 vm_map_unlock(buffer_map);
537 vm_map_entry_release(count);
546 * Remove the buffer from the appropriate free list.
549 bremfree(struct buf * bp)
554 old_qindex = bp->b_qindex;
556 if (bp->b_qindex != BQUEUE_NONE) {
557 KASSERT(BUF_REFCNTNB(bp) == 1,
558 ("bremfree: bp %p not locked",bp));
559 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
560 bp->b_qindex = BQUEUE_NONE;
562 if (BUF_REFCNTNB(bp) <= 1)
563 panic("bremfree: removing a buffer not on a queue");
567 * Fixup numfreebuffers count. If the buffer is invalid or not
568 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
569 * the buffer was free and we must decrement numfreebuffers.
571 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
576 case BQUEUE_EMPTYKVA:
590 * Get a buffer with the specified data. Look in the cache first. We
591 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
592 * is set, the buffer is valid and we do not have to do anything ( see
596 bread(struct vnode * vp, off_t loffset, int size, struct buf ** bpp)
600 bp = getblk(vp, loffset, size, 0, 0);
603 /* if not found in cache, do some I/O */
604 if ((bp->b_flags & B_CACHE) == 0) {
605 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp));
606 bp->b_flags &= ~(B_ERROR | B_INVAL);
607 bp->b_cmd = BUF_CMD_READ;
608 vfs_busy_pages(vp, bp);
609 vn_strategy(vp, &bp->b_bio1);
610 return (biowait(bp));
618 * Operates like bread, but also starts asynchronous I/O on
619 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
620 * to initiating I/O . If B_CACHE is set, the buffer is valid
621 * and we do not have to do anything.
624 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
625 int *rabsize, int cnt, struct buf ** bpp)
627 struct buf *bp, *rabp;
629 int rv = 0, readwait = 0;
631 *bpp = bp = getblk(vp, loffset, size, 0, 0);
633 /* if not found in cache, do some I/O */
634 if ((bp->b_flags & B_CACHE) == 0) {
635 bp->b_flags &= ~(B_ERROR | B_INVAL);
636 bp->b_cmd = BUF_CMD_READ;
637 vfs_busy_pages(vp, bp);
638 vn_strategy(vp, &bp->b_bio1);
642 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
643 if (inmem(vp, *raoffset))
645 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
647 if ((rabp->b_flags & B_CACHE) == 0) {
648 rabp->b_flags |= B_ASYNC;
649 rabp->b_flags &= ~(B_ERROR | B_INVAL);
650 rabp->b_cmd = BUF_CMD_READ;
651 vfs_busy_pages(vp, rabp);
653 vn_strategy(vp, &rabp->b_bio1);
668 * Write, release buffer on completion. (Done by iodone
669 * if async). Do not bother writing anything if the buffer
672 * Note that we set B_CACHE here, indicating that buffer is
673 * fully valid and thus cacheable. This is true even of NFS
674 * now so we set it generally. This could be set either here
675 * or in biodone() since the I/O is synchronous. We put it
679 bwrite(struct buf * bp)
683 if (bp->b_flags & B_INVAL) {
688 oldflags = bp->b_flags;
690 if (BUF_REFCNTNB(bp) == 0)
691 panic("bwrite: buffer is not busy???");
694 /* Mark the buffer clean */
697 bp->b_flags &= ~B_ERROR;
698 bp->b_flags |= B_CACHE;
699 bp->b_cmd = BUF_CMD_WRITE;
700 vfs_busy_pages(bp->b_vp, bp);
703 * Normal bwrites pipeline writes
705 bp->b_runningbufspace = bp->b_bufsize;
706 runningbufspace += bp->b_runningbufspace;
709 if (oldflags & B_ASYNC)
711 vn_strategy(bp->b_vp, &bp->b_bio1);
713 if ((oldflags & B_ASYNC) == 0) {
714 int rtval = biowait(bp);
717 } else if ((oldflags & B_NOWDRAIN) == 0) {
719 * don't allow the async write to saturate the I/O
720 * system. Deadlocks can occur only if a device strategy
721 * routine (like in VN) turns around and issues another
722 * high-level write, in which case B_NOWDRAIN is expected
723 * to be set. Otherwise we will not deadlock here because
724 * we are blocking waiting for I/O that is already in-progress
727 waitrunningbufspace();
736 * Delayed write. (Buffer is marked dirty). Do not bother writing
737 * anything if the buffer is marked invalid.
739 * Note that since the buffer must be completely valid, we can safely
740 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
741 * biodone() in order to prevent getblk from writing the buffer
745 bdwrite(struct buf *bp)
747 if (BUF_REFCNTNB(bp) == 0)
748 panic("bdwrite: buffer is not busy");
750 if (bp->b_flags & B_INVAL) {
757 * Set B_CACHE, indicating that the buffer is fully valid. This is
758 * true even of NFS now.
760 bp->b_flags |= B_CACHE;
763 * This bmap keeps the system from needing to do the bmap later,
764 * perhaps when the system is attempting to do a sync. Since it
765 * is likely that the indirect block -- or whatever other datastructure
766 * that the filesystem needs is still in memory now, it is a good
767 * thing to do this. Note also, that if the pageout daemon is
768 * requesting a sync -- there might not be enough memory to do
769 * the bmap then... So, this is important to do.
771 if (bp->b_bio2.bio_offset == NOOFFSET) {
772 VOP_BMAP(bp->b_vp, bp->b_loffset, NULL, &bp->b_bio2.bio_offset,
777 * Set the *dirty* buffer range based upon the VM system dirty pages.
782 * We need to do this here to satisfy the vnode_pager and the
783 * pageout daemon, so that it thinks that the pages have been
784 * "cleaned". Note that since the pages are in a delayed write
785 * buffer -- the VFS layer "will" see that the pages get written
786 * out on the next sync, or perhaps the cluster will be completed.
792 * Wakeup the buffer flushing daemon if we have a lot of dirty
793 * buffers (midpoint between our recovery point and our stall
796 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
799 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
800 * due to the softdep code.
807 * Turn buffer into delayed write request by marking it B_DELWRI.
808 * B_RELBUF and B_NOCACHE must be cleared.
810 * We reassign the buffer to itself to properly update it in the
813 * Since the buffer is not on a queue, we do not update the
814 * numfreebuffers count.
816 * Must be called from a critical section.
817 * The buffer must be on BQUEUE_NONE.
820 bdirty(struct buf *bp)
822 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
823 if (bp->b_flags & B_NOCACHE) {
824 printf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
825 bp->b_flags &= ~B_NOCACHE;
827 if (bp->b_flags & B_INVAL) {
828 printf("bdirty: warning, dirtying invalid buffer %p\n", bp);
830 bp->b_flags &= ~B_RELBUF;
832 if ((bp->b_flags & B_DELWRI) == 0) {
833 bp->b_flags |= B_DELWRI;
836 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
843 * Clear B_DELWRI for buffer.
845 * Since the buffer is not on a queue, we do not update the numfreebuffers
848 * Must be called from a critical section.
850 * The buffer is typically on BQUEUE_NONE but there is one case in
851 * brelse() that calls this function after placing the buffer on
856 bundirty(struct buf *bp)
858 if (bp->b_flags & B_DELWRI) {
859 bp->b_flags &= ~B_DELWRI;
862 numdirtywakeup(lodirtybuffers);
865 * Since it is now being written, we can clear its deferred write flag.
867 bp->b_flags &= ~B_DEFERRED;
873 * Asynchronous write. Start output on a buffer, but do not wait for
874 * it to complete. The buffer is released when the output completes.
876 * bwrite() ( or the VOP routine anyway ) is responsible for handling
877 * B_INVAL buffers. Not us.
880 bawrite(struct buf * bp)
882 bp->b_flags |= B_ASYNC;
883 (void) VOP_BWRITE(bp->b_vp, bp);
889 * Ordered write. Start output on a buffer, and flag it so that the
890 * device will write it in the order it was queued. The buffer is
891 * released when the output completes. bwrite() ( or the VOP routine
892 * anyway ) is responsible for handling B_INVAL buffers.
895 bowrite(struct buf * bp)
897 bp->b_flags |= B_ORDERED | B_ASYNC;
898 return (VOP_BWRITE(bp->b_vp, bp));
904 * Called prior to the locking of any vnodes when we are expecting to
905 * write. We do not want to starve the buffer cache with too many
906 * dirty buffers so we block here. By blocking prior to the locking
907 * of any vnodes we attempt to avoid the situation where a locked vnode
908 * prevents the various system daemons from flushing related buffers.
914 if (numdirtybuffers >= hidirtybuffers) {
916 while (numdirtybuffers >= hidirtybuffers) {
918 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
919 tsleep(&needsbuffer, 0, "flswai", 0);
926 * buf_dirty_count_severe:
928 * Return true if we have too many dirty buffers.
931 buf_dirty_count_severe(void)
933 return(numdirtybuffers >= hidirtybuffers);
939 * Release a busy buffer and, if requested, free its resources. The
940 * buffer will be stashed in the appropriate bufqueue[] allowing it
941 * to be accessed later as a cache entity or reused for other purposes.
944 brelse(struct buf * bp)
947 int saved_flags = bp->b_flags;
950 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
954 if ((bp->b_flags & (B_NOCACHE|B_DIRTY)) == (B_NOCACHE|B_DIRTY)) {
955 printf("warning: buf %p marked dirty & B_NOCACHE, clearing B_NOCACHE\n", bp);
956 bp->b_flags &= ~B_NOCACHE;
959 if (bp->b_flags & B_LOCKED)
960 bp->b_flags &= ~B_ERROR;
962 if (bp->b_cmd == BUF_CMD_WRITE &&
963 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
965 * Failed write, redirty. Must clear B_ERROR to prevent
966 * pages from being scrapped. If B_INVAL is set then
967 * this case is not run and the next case is run to
968 * destroy the buffer. B_INVAL can occur if the buffer
969 * is outside the range supported by the underlying device.
971 bp->b_flags &= ~B_ERROR;
973 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
974 (bp->b_bufsize <= 0) || bp->b_cmd == BUF_CMD_FREEBLKS) {
976 * Either a failed I/O or we were asked to free or not
979 bp->b_flags |= B_INVAL;
980 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
981 (*bioops.io_deallocate)(bp);
982 if (bp->b_flags & B_DELWRI) {
984 numdirtywakeup(lodirtybuffers);
986 bp->b_flags &= ~(B_DELWRI | B_CACHE);
990 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
991 * is called with B_DELWRI set, the underlying pages may wind up
992 * getting freed causing a previous write (bdwrite()) to get 'lost'
993 * because pages associated with a B_DELWRI bp are marked clean.
995 * We still allow the B_INVAL case to call vfs_vmio_release(), even
996 * if B_DELWRI is set.
998 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
999 * on pages to return pages to the VM page queues.
1001 if (bp->b_flags & B_DELWRI)
1002 bp->b_flags &= ~B_RELBUF;
1003 else if (vm_page_count_severe())
1004 bp->b_flags |= B_RELBUF;
1007 * At this point destroying the buffer is governed by the B_INVAL
1008 * or B_RELBUF flags.
1010 bp->b_cmd = BUF_CMD_DONE;
1013 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1014 * constituted, not even NFS buffers now. Two flags effect this. If
1015 * B_INVAL, the struct buf is invalidated but the VM object is kept
1016 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1018 * If B_ERROR or B_NOCACHE is set, pages in the VM object will be
1019 * invalidated. B_ERROR cannot be set for a failed write unless the
1020 * buffer is also B_INVAL because it hits the re-dirtying code above.
1022 * Normally we can do this whether a buffer is B_DELWRI or not. If
1023 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1024 * the commit state and we cannot afford to lose the buffer. If the
1025 * buffer has a background write in progress, we need to keep it
1026 * around to prevent it from being reconstituted and starting a second
1029 if ((bp->b_flags & B_VMIO)
1030 && !(bp->b_vp->v_tag == VT_NFS &&
1031 !vn_isdisk(bp->b_vp, NULL) &&
1032 (bp->b_flags & B_DELWRI))
1035 * Rundown for VMIO buffers which are not dirty NFS buffers.
1047 * Get the base offset and length of the buffer. Note that
1048 * in the VMIO case if the buffer block size is not
1049 * page-aligned then b_data pointer may not be page-aligned.
1050 * But our b_xio.xio_pages array *IS* page aligned.
1052 * block sizes less then DEV_BSIZE (usually 512) are not
1053 * supported due to the page granularity bits (m->valid,
1054 * m->dirty, etc...).
1056 * See man buf(9) for more information
1059 resid = bp->b_bufsize;
1060 foff = bp->b_loffset;
1062 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1063 m = bp->b_xio.xio_pages[i];
1064 vm_page_flag_clear(m, PG_ZERO);
1066 * If we hit a bogus page, fixup *all* of them
1067 * now. Note that we left these pages wired
1068 * when we removed them so they had better exist,
1069 * and they cannot be ripped out from under us so
1070 * no critical section protection is necessary.
1072 if (m == bogus_page) {
1074 poff = OFF_TO_IDX(bp->b_loffset);
1076 for (j = i; j < bp->b_xio.xio_npages; j++) {
1079 mtmp = bp->b_xio.xio_pages[j];
1080 if (mtmp == bogus_page) {
1081 mtmp = vm_page_lookup(obj, poff + j);
1083 panic("brelse: page missing");
1085 bp->b_xio.xio_pages[j] = mtmp;
1089 if ((bp->b_flags & B_INVAL) == 0) {
1090 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1091 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1093 m = bp->b_xio.xio_pages[i];
1097 * Invalidate the backing store if B_NOCACHE is set
1098 * (e.g. used with vinvalbuf()). If this is NFS
1099 * we impose a requirement that the block size be
1100 * a multiple of PAGE_SIZE and create a temporary
1101 * hack to basically invalidate the whole page. The
1102 * problem is that NFS uses really odd buffer sizes
1103 * especially when tracking piecemeal writes and
1104 * it also vinvalbuf()'s a lot, which would result
1105 * in only partial page validation and invalidation
1106 * here. If the file page is mmap()'d, however,
1107 * all the valid bits get set so after we invalidate
1108 * here we would end up with weird m->valid values
1109 * like 0xfc. nfs_getpages() can't handle this so
1110 * we clear all the valid bits for the NFS case
1111 * instead of just some of them.
1113 * The real bug is the VM system having to set m->valid
1114 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1115 * itself is an artifact of the whole 512-byte
1116 * granular mess that exists to support odd block
1117 * sizes and UFS meta-data block sizes (e.g. 6144).
1118 * A complete rewrite is required.
1120 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1121 int poffset = foff & PAGE_MASK;
1124 presid = PAGE_SIZE - poffset;
1125 if (bp->b_vp->v_tag == VT_NFS &&
1126 bp->b_vp->v_type == VREG) {
1128 } else if (presid > resid) {
1131 KASSERT(presid >= 0, ("brelse: extra page"));
1132 vm_page_set_invalid(m, poffset, presid);
1134 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1135 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1137 if (bp->b_flags & (B_INVAL | B_RELBUF))
1138 vfs_vmio_release(bp);
1139 } else if (bp->b_flags & B_VMIO) {
1141 * Rundown for VMIO buffers which are dirty NFS buffers. Such
1142 * buffers contain tracking ranges for NFS and cannot normally
1143 * be released. Due to the dirty check above this series of
1144 * conditionals, B_RELBUF probably will never be set in this
1147 if (bp->b_flags & (B_INVAL | B_RELBUF))
1148 vfs_vmio_release(bp);
1151 * Rundown for non-VMIO buffers.
1153 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1156 printf("brelse bp %p %08x/%08x: Warning, caught and fixed brelvp bug\n", bp, saved_flags, bp->b_flags);
1165 if (bp->b_qindex != BQUEUE_NONE)
1166 panic("brelse: free buffer onto another queue???");
1167 if (BUF_REFCNTNB(bp) > 1) {
1168 /* Temporary panic to verify exclusive locking */
1169 /* This panic goes away when we allow shared refs */
1170 panic("brelse: multiple refs");
1171 /* do not release to free list */
1178 * Figure out the correct queue to place the cleaned up buffer on.
1179 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1180 * disassociated from their vnode.
1183 if (bp->b_bufsize == 0) {
1185 * Buffers with no memory. Due to conditionals near the top
1186 * of brelse() such buffers should probably already be
1187 * marked B_INVAL and disassociated from their vnode.
1189 bp->b_flags |= B_INVAL;
1190 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1191 KKASSERT((bp->b_flags & B_HASHED) == 0);
1192 if (bp->b_kvasize) {
1193 bp->b_qindex = BQUEUE_EMPTYKVA;
1195 bp->b_qindex = BQUEUE_EMPTY;
1197 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1198 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) {
1200 * Buffers with junk contents. Again these buffers had better
1201 * already be disassociated from their vnode.
1203 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1204 KKASSERT((bp->b_flags & B_HASHED) == 0);
1205 bp->b_flags |= B_INVAL;
1206 bp->b_qindex = BQUEUE_CLEAN;
1207 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1208 } else if (bp->b_flags & B_LOCKED) {
1210 * Buffers that are locked.
1212 bp->b_qindex = BQUEUE_LOCKED;
1213 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1216 * Remaining buffers. These buffers are still associated with
1219 switch(bp->b_flags & (B_DELWRI|B_AGE)) {
1220 case B_DELWRI | B_AGE:
1221 bp->b_qindex = BQUEUE_DIRTY;
1222 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1225 bp->b_qindex = BQUEUE_DIRTY;
1226 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1229 bp->b_qindex = BQUEUE_CLEAN;
1230 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1233 bp->b_qindex = BQUEUE_CLEAN;
1234 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1240 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1241 * on the correct queue.
1243 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1247 * Fixup numfreebuffers count. The bp is on an appropriate queue
1248 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1249 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1250 * if B_INVAL is set ).
1252 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI))
1256 * Something we can maybe free or reuse
1258 if (bp->b_bufsize || bp->b_kvasize)
1262 * Clean up temporary flags and unlock the buffer.
1264 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF |
1265 B_DIRECT | B_NOWDRAIN);
1273 * Release a buffer back to the appropriate queue but do not try to free
1274 * it. The buffer is expected to be used again soon.
1276 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1277 * biodone() to requeue an async I/O on completion. It is also used when
1278 * known good buffers need to be requeued but we think we may need the data
1281 * XXX we should be able to leave the B_RELBUF hint set on completion.
1284 bqrelse(struct buf * bp)
1288 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1290 if (bp->b_qindex != BQUEUE_NONE)
1291 panic("bqrelse: free buffer onto another queue???");
1292 if (BUF_REFCNTNB(bp) > 1) {
1293 /* do not release to free list */
1294 panic("bqrelse: multiple refs");
1299 if (bp->b_flags & B_LOCKED) {
1300 bp->b_flags &= ~B_ERROR;
1301 bp->b_qindex = BQUEUE_LOCKED;
1302 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1303 /* buffers with stale but valid contents */
1304 } else if (bp->b_flags & B_DELWRI) {
1305 bp->b_qindex = BQUEUE_DIRTY;
1306 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1307 } else if (vm_page_count_severe()) {
1309 * We are too low on memory, we have to try to free the
1310 * buffer (most importantly: the wired pages making up its
1311 * backing store) *now*.
1317 bp->b_qindex = BQUEUE_CLEAN;
1318 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1321 if ((bp->b_flags & B_LOCKED) == 0 &&
1322 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) {
1327 * Something we can maybe free or reuse.
1329 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1333 * Final cleanup and unlock. Clear bits that are only used while a
1334 * buffer is actively locked.
1336 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1344 * Return backing pages held by the buffer 'bp' back to the VM system
1345 * if possible. The pages are freed if they are no longer valid or
1346 * attempt to free if it was used for direct I/O otherwise they are
1347 * sent to the page cache.
1349 * Pages that were marked busy are left alone and skipped.
1351 * The KVA mapping (b_data) for the underlying pages is removed by
1355 vfs_vmio_release(struct buf *bp)
1361 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1362 m = bp->b_xio.xio_pages[i];
1363 bp->b_xio.xio_pages[i] = NULL;
1365 * In order to keep page LRU ordering consistent, put
1366 * everything on the inactive queue.
1368 vm_page_unwire(m, 0);
1370 * We don't mess with busy pages, it is
1371 * the responsibility of the process that
1372 * busied the pages to deal with them.
1374 if ((m->flags & PG_BUSY) || (m->busy != 0))
1377 if (m->wire_count == 0) {
1378 vm_page_flag_clear(m, PG_ZERO);
1380 * Might as well free the page if we can and it has
1381 * no valid data. We also free the page if the
1382 * buffer was used for direct I/O.
1384 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1385 m->hold_count == 0) {
1387 vm_page_protect(m, VM_PROT_NONE);
1389 } else if (bp->b_flags & B_DIRECT) {
1390 vm_page_try_to_free(m);
1391 } else if (vm_page_count_severe()) {
1392 vm_page_try_to_cache(m);
1397 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1398 if (bp->b_bufsize) {
1402 bp->b_xio.xio_npages = 0;
1403 bp->b_flags &= ~B_VMIO;
1411 * Implement clustered async writes for clearing out B_DELWRI buffers.
1412 * This is much better then the old way of writing only one buffer at
1413 * a time. Note that we may not be presented with the buffers in the
1414 * correct order, so we search for the cluster in both directions.
1416 * The buffer is locked on call.
1419 vfs_bio_awrite(struct buf *bp)
1423 off_t loffset = bp->b_loffset;
1424 struct vnode *vp = bp->b_vp;
1432 * right now we support clustered writing only to regular files. If
1433 * we find a clusterable block we could be in the middle of a cluster
1434 * rather then at the beginning.
1436 * NOTE: b_bio1 contains the logical loffset and is aliased
1437 * to b_loffset. b_bio2 contains the translated block number.
1439 if ((vp->v_type == VREG) &&
1440 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1441 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1443 size = vp->v_mount->mnt_stat.f_iosize;
1445 for (i = size; i < MAXPHYS; i += size) {
1446 if ((bpa = findblk(vp, loffset + i)) &&
1447 BUF_REFCNT(bpa) == 0 &&
1448 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1449 (B_DELWRI | B_CLUSTEROK)) &&
1450 (bpa->b_bufsize == size)) {
1451 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1452 (bpa->b_bio2.bio_offset !=
1453 bp->b_bio2.bio_offset + i))
1459 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1460 if ((bpa = findblk(vp, loffset - j)) &&
1461 BUF_REFCNT(bpa) == 0 &&
1462 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1463 (B_DELWRI | B_CLUSTEROK)) &&
1464 (bpa->b_bufsize == size)) {
1465 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1466 (bpa->b_bio2.bio_offset !=
1467 bp->b_bio2.bio_offset - j))
1476 * this is a possible cluster write
1478 if (nbytes != size) {
1480 nwritten = cluster_wbuild(vp, size,
1481 loffset - j, nbytes);
1488 bp->b_flags |= B_ASYNC;
1492 * default (old) behavior, writing out only one block
1494 * XXX returns b_bufsize instead of b_bcount for nwritten?
1496 nwritten = bp->b_bufsize;
1497 (void) VOP_BWRITE(bp->b_vp, bp);
1505 * Find and initialize a new buffer header, freeing up existing buffers
1506 * in the bufqueues as necessary. The new buffer is returned locked.
1508 * Important: B_INVAL is not set. If the caller wishes to throw the
1509 * buffer away, the caller must set B_INVAL prior to calling brelse().
1512 * We have insufficient buffer headers
1513 * We have insufficient buffer space
1514 * buffer_map is too fragmented ( space reservation fails )
1515 * If we have to flush dirty buffers ( but we try to avoid this )
1517 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1518 * Instead we ask the buf daemon to do it for us. We attempt to
1519 * avoid piecemeal wakeups of the pageout daemon.
1523 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1529 static int flushingbufs;
1532 * We can't afford to block since we might be holding a vnode lock,
1533 * which may prevent system daemons from running. We deal with
1534 * low-memory situations by proactively returning memory and running
1535 * async I/O rather then sync I/O.
1539 --getnewbufrestarts;
1541 ++getnewbufrestarts;
1544 * Setup for scan. If we do not have enough free buffers,
1545 * we setup a degenerate case that immediately fails. Note
1546 * that if we are specially marked process, we are allowed to
1547 * dip into our reserves.
1549 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1551 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1552 * However, there are a number of cases (defragging, reusing, ...)
1553 * where we cannot backup.
1555 nqindex = BQUEUE_EMPTYKVA;
1556 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1560 * If no EMPTYKVA buffers and we are either
1561 * defragging or reusing, locate a CLEAN buffer
1562 * to free or reuse. If bufspace useage is low
1563 * skip this step so we can allocate a new buffer.
1565 if (defrag || bufspace >= lobufspace) {
1566 nqindex = BQUEUE_CLEAN;
1567 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1571 * If we could not find or were not allowed to reuse a
1572 * CLEAN buffer, check to see if it is ok to use an EMPTY
1573 * buffer. We can only use an EMPTY buffer if allocating
1574 * its KVA would not otherwise run us out of buffer space.
1576 if (nbp == NULL && defrag == 0 &&
1577 bufspace + maxsize < hibufspace) {
1578 nqindex = BQUEUE_EMPTY;
1579 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1584 * Run scan, possibly freeing data and/or kva mappings on the fly
1588 while ((bp = nbp) != NULL) {
1589 int qindex = nqindex;
1592 * Calculate next bp ( we can only use it if we do not block
1593 * or do other fancy things ).
1595 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1598 nqindex = BQUEUE_EMPTYKVA;
1599 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1602 case BQUEUE_EMPTYKVA:
1603 nqindex = BQUEUE_CLEAN;
1604 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1618 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1621 * Note: we no longer distinguish between VMIO and non-VMIO
1625 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1628 * If we are defragging then we need a buffer with
1629 * b_kvasize != 0. XXX this situation should no longer
1630 * occur, if defrag is non-zero the buffer's b_kvasize
1631 * should also be non-zero at this point. XXX
1633 if (defrag && bp->b_kvasize == 0) {
1634 printf("Warning: defrag empty buffer %p\n", bp);
1639 * Start freeing the bp. This is somewhat involved. nbp
1640 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1641 * on the clean list must be disassociated from their
1642 * current vnode. Buffers on the empty[kva] lists have
1643 * already been disassociated.
1646 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1647 printf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1648 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1651 if (bp->b_qindex != qindex) {
1652 printf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1658 if (qindex == BQUEUE_CLEAN) {
1659 if (bp->b_flags & B_VMIO) {
1660 bp->b_flags &= ~B_ASYNC;
1661 vfs_vmio_release(bp);
1668 * NOTE: nbp is now entirely invalid. We can only restart
1669 * the scan from this point on.
1671 * Get the rest of the buffer freed up. b_kva* is still
1672 * valid after this operation.
1675 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1676 KKASSERT((bp->b_flags & B_HASHED) == 0);
1677 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
1678 (*bioops.io_deallocate)(bp);
1681 * critical section protection is not required when
1682 * scrapping a buffer's contents because it is already
1689 bp->b_cmd = BUF_CMD_DONE;
1695 bp->b_xio.xio_npages = 0;
1696 bp->b_dirtyoff = bp->b_dirtyend = 0;
1699 LIST_INIT(&bp->b_dep);
1702 * If we are defragging then free the buffer.
1705 bp->b_flags |= B_INVAL;
1713 * If we are overcomitted then recover the buffer and its
1714 * KVM space. This occurs in rare situations when multiple
1715 * processes are blocked in getnewbuf() or allocbuf().
1717 if (bufspace >= hibufspace)
1719 if (flushingbufs && bp->b_kvasize != 0) {
1720 bp->b_flags |= B_INVAL;
1725 if (bufspace < lobufspace)
1731 * If we exhausted our list, sleep as appropriate. We may have to
1732 * wakeup various daemons and write out some dirty buffers.
1734 * Generally we are sleeping due to insufficient buffer space.
1742 flags = VFS_BIO_NEED_BUFSPACE;
1744 } else if (bufspace >= hibufspace) {
1746 flags = VFS_BIO_NEED_BUFSPACE;
1749 flags = VFS_BIO_NEED_ANY;
1752 bd_speedup(); /* heeeelp */
1754 needsbuffer |= flags;
1755 while (needsbuffer & flags) {
1756 if (tsleep(&needsbuffer, slpflag, waitmsg, slptimeo))
1761 * We finally have a valid bp. We aren't quite out of the
1762 * woods, we still have to reserve kva space. In order
1763 * to keep fragmentation sane we only allocate kva in
1766 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1768 if (maxsize != bp->b_kvasize) {
1769 vm_offset_t addr = 0;
1774 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1775 vm_map_lock(buffer_map);
1777 if (vm_map_findspace(buffer_map,
1778 vm_map_min(buffer_map), maxsize,
1781 * Uh oh. Buffer map is too fragmented. We
1782 * must defragment the map.
1784 vm_map_unlock(buffer_map);
1785 vm_map_entry_release(count);
1788 bp->b_flags |= B_INVAL;
1793 vm_map_insert(buffer_map, &count,
1795 addr, addr + maxsize,
1796 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
1798 bp->b_kvabase = (caddr_t) addr;
1799 bp->b_kvasize = maxsize;
1800 bufspace += bp->b_kvasize;
1803 vm_map_unlock(buffer_map);
1804 vm_map_entry_release(count);
1806 bp->b_data = bp->b_kvabase;
1814 * Buffer flushing daemon. Buffers are normally flushed by the
1815 * update daemon but if it cannot keep up this process starts to
1816 * take the load in an attempt to prevent getnewbuf() from blocking.
1819 static struct thread *bufdaemonthread;
1821 static struct kproc_desc buf_kp = {
1826 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
1832 * This process needs to be suspended prior to shutdown sync.
1834 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
1835 bufdaemonthread, SHUTDOWN_PRI_LAST);
1838 * This process is allowed to take the buffer cache to the limit
1843 kproc_suspend_loop();
1846 * Do the flush. Limit the amount of in-transit I/O we
1847 * allow to build up, otherwise we would completely saturate
1848 * the I/O system. Wakeup any waiting processes before we
1849 * normally would so they can run in parallel with our drain.
1851 while (numdirtybuffers > lodirtybuffers) {
1852 if (flushbufqueues() == 0)
1854 waitrunningbufspace();
1855 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
1859 * Only clear bd_request if we have reached our low water
1860 * mark. The buf_daemon normally waits 5 seconds and
1861 * then incrementally flushes any dirty buffers that have
1862 * built up, within reason.
1864 * If we were unable to hit our low water mark and couldn't
1865 * find any flushable buffers, we sleep half a second.
1866 * Otherwise we loop immediately.
1868 if (numdirtybuffers <= lodirtybuffers) {
1870 * We reached our low water mark, reset the
1871 * request and sleep until we are needed again.
1872 * The sleep is just so the suspend code works.
1875 tsleep(&bd_request, 0, "psleep", hz);
1878 * We couldn't find any flushable dirty buffers but
1879 * still have too many dirty buffers, we
1880 * have to sleep and try again. (rare)
1882 tsleep(&bd_request, 0, "qsleep", hz / 2);
1890 * Try to flush a buffer in the dirty queue. We must be careful to
1891 * free up B_INVAL buffers instead of write them, which NFS is
1892 * particularly sensitive to.
1896 flushbufqueues(void)
1901 bp = TAILQ_FIRST(&bufqueues[BQUEUE_DIRTY]);
1904 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp));
1905 if (bp->b_flags & B_DELWRI) {
1906 if (bp->b_flags & B_INVAL) {
1907 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1908 panic("flushbufqueues: locked buf");
1914 if (LIST_FIRST(&bp->b_dep) != NULL &&
1915 bioops.io_countdeps &&
1916 (bp->b_flags & B_DEFERRED) == 0 &&
1917 (*bioops.io_countdeps)(bp, 0)) {
1918 TAILQ_REMOVE(&bufqueues[BQUEUE_DIRTY],
1920 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY],
1922 bp->b_flags |= B_DEFERRED;
1923 bp = TAILQ_FIRST(&bufqueues[BQUEUE_DIRTY]);
1928 * Only write it out if we can successfully lock
1931 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
1937 bp = TAILQ_NEXT(bp, b_freelist);
1945 * Returns true if no I/O is needed to access the associated VM object.
1946 * This is like findblk except it also hunts around in the VM system for
1949 * Note that we ignore vm_page_free() races from interrupts against our
1950 * lookup, since if the caller is not protected our return value will not
1951 * be any more valid then otherwise once we exit the critical section.
1954 inmem(struct vnode *vp, off_t loffset)
1957 vm_offset_t toff, tinc, size;
1960 if (findblk(vp, loffset))
1962 if (vp->v_mount == NULL)
1964 if ((obj = vp->v_object) == NULL)
1968 if (size > vp->v_mount->mnt_stat.f_iosize)
1969 size = vp->v_mount->mnt_stat.f_iosize;
1971 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
1972 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
1976 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
1977 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
1978 if (vm_page_is_valid(m,
1979 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
1988 * Sets the dirty range for a buffer based on the status of the dirty
1989 * bits in the pages comprising the buffer.
1991 * The range is limited to the size of the buffer.
1993 * This routine is primarily used by NFS, but is generalized for the
1997 vfs_setdirty(struct buf *bp)
2003 * Degenerate case - empty buffer
2006 if (bp->b_bufsize == 0)
2010 * We qualify the scan for modified pages on whether the
2011 * object has been flushed yet. The OBJ_WRITEABLE flag
2012 * is not cleared simply by protecting pages off.
2015 if ((bp->b_flags & B_VMIO) == 0)
2018 object = bp->b_xio.xio_pages[0]->object;
2020 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2021 printf("Warning: object %p writeable but not mightbedirty\n", object);
2022 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2023 printf("Warning: object %p mightbedirty but not writeable\n", object);
2025 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2026 vm_offset_t boffset;
2027 vm_offset_t eoffset;
2030 * test the pages to see if they have been modified directly
2031 * by users through the VM system.
2033 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2034 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2035 vm_page_test_dirty(bp->b_xio.xio_pages[i]);
2039 * Calculate the encompassing dirty range, boffset and eoffset,
2040 * (eoffset - boffset) bytes.
2043 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2044 if (bp->b_xio.xio_pages[i]->dirty)
2047 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2049 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2050 if (bp->b_xio.xio_pages[i]->dirty) {
2054 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2057 * Fit it to the buffer.
2060 if (eoffset > bp->b_bcount)
2061 eoffset = bp->b_bcount;
2064 * If we have a good dirty range, merge with the existing
2068 if (boffset < eoffset) {
2069 if (bp->b_dirtyoff > boffset)
2070 bp->b_dirtyoff = boffset;
2071 if (bp->b_dirtyend < eoffset)
2072 bp->b_dirtyend = eoffset;
2080 * Locate and return the specified buffer, or NULL if the buffer does
2081 * not exist. Do not attempt to lock the buffer or manipulate it in
2082 * any way. The caller must validate that the correct buffer has been
2083 * obtain after locking it.
2086 findblk(struct vnode *vp, off_t loffset)
2091 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2099 * Get a block given a specified block and offset into a file/device.
2100 * B_INVAL may or may not be set on return. The caller should clear
2101 * B_INVAL prior to initiating a READ.
2103 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2104 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2105 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2106 * without doing any of those things the system will likely believe
2107 * the buffer to be valid (especially if it is not B_VMIO), and the
2108 * next getblk() will return the buffer with B_CACHE set.
2110 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2111 * an existing buffer.
2113 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2114 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2115 * and then cleared based on the backing VM. If the previous buffer is
2116 * non-0-sized but invalid, B_CACHE will be cleared.
2118 * If getblk() must create a new buffer, the new buffer is returned with
2119 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2120 * case it is returned with B_INVAL clear and B_CACHE set based on the
2123 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos
2124 * B_CACHE bit is clear.
2126 * What this means, basically, is that the caller should use B_CACHE to
2127 * determine whether the buffer is fully valid or not and should clear
2128 * B_INVAL prior to issuing a read. If the caller intends to validate
2129 * the buffer by loading its data area with something, the caller needs
2130 * to clear B_INVAL. If the caller does this without issuing an I/O,
2131 * the caller should set B_CACHE ( as an optimization ), else the caller
2132 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2133 * a write attempt or if it was a successfull read. If the caller
2134 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2135 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2138 getblk(struct vnode *vp, off_t loffset, int size, int slpflag, int slptimeo)
2142 if (size > MAXBSIZE)
2143 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2144 if (vp->v_object == NULL)
2145 panic("getblk: vnode %p has no object!", vp);
2150 * Block if we are low on buffers. Certain processes are allowed
2151 * to completely exhaust the buffer cache.
2153 * If this check ever becomes a bottleneck it may be better to
2154 * move it into the else, when findblk() fails. At the moment
2155 * it isn't a problem.
2157 * XXX remove, we cannot afford to block anywhere if holding a vnode
2158 * lock in low-memory situation, so take it to the max.
2160 if (numfreebuffers == 0) {
2163 needsbuffer |= VFS_BIO_NEED_ANY;
2164 tsleep(&needsbuffer, slpflag, "newbuf", slptimeo);
2167 if ((bp = findblk(vp, loffset))) {
2169 * The buffer was found in the cache, but we need to lock it.
2170 * Even with LK_NOWAIT the lockmgr may break our critical
2171 * section, so double-check the validity of the buffer
2172 * once the lock has been obtained.
2174 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2175 int lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2176 if (slpflag & PCATCH)
2177 lkflags |= LK_PCATCH;
2178 if (BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo) ==
2187 * Once the buffer has been locked, make sure we didn't race
2188 * a buffer recyclement. Buffers that are no longer hashed
2189 * will have b_vp == NULL, so this takes care of that check
2192 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2193 printf("Warning buffer %p (vp %p loffset %lld) was recycled\n", bp, vp, loffset);
2199 * All vnode-based buffers must be backed by a VM object.
2201 KKASSERT(bp->b_flags & B_VMIO);
2202 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2205 * Make sure that B_INVAL buffers do not have a cached
2206 * block number translation.
2208 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2209 printf("Warning invalid buffer %p (vp %p loffset %lld) did not have cleared bio_offset cache\n", bp, vp, loffset);
2210 clearbiocache(&bp->b_bio2);
2214 * The buffer is locked. B_CACHE is cleared if the buffer is
2217 if (bp->b_flags & B_INVAL)
2218 bp->b_flags &= ~B_CACHE;
2222 * Any size inconsistancy with a dirty buffer or a buffer
2223 * with a softupdates dependancy must be resolved. Resizing
2224 * the buffer in such circumstances can lead to problems.
2226 if (size != bp->b_bcount) {
2227 if (bp->b_flags & B_DELWRI) {
2228 bp->b_flags |= B_NOCACHE;
2229 VOP_BWRITE(bp->b_vp, bp);
2230 } else if (LIST_FIRST(&bp->b_dep)) {
2231 bp->b_flags |= B_NOCACHE;
2232 VOP_BWRITE(bp->b_vp, bp);
2234 bp->b_flags |= B_RELBUF;
2239 KKASSERT(size <= bp->b_kvasize);
2240 KASSERT(bp->b_loffset != NOOFFSET,
2241 ("getblk: no buffer offset"));
2244 * A buffer with B_DELWRI set and B_CACHE clear must
2245 * be committed before we can return the buffer in
2246 * order to prevent the caller from issuing a read
2247 * ( due to B_CACHE not being set ) and overwriting
2250 * Most callers, including NFS and FFS, need this to
2251 * operate properly either because they assume they
2252 * can issue a read if B_CACHE is not set, or because
2253 * ( for example ) an uncached B_DELWRI might loop due
2254 * to softupdates re-dirtying the buffer. In the latter
2255 * case, B_CACHE is set after the first write completes,
2256 * preventing further loops.
2258 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2259 * above while extending the buffer, we cannot allow the
2260 * buffer to remain with B_CACHE set after the write
2261 * completes or it will represent a corrupt state. To
2262 * deal with this we set B_NOCACHE to scrap the buffer
2265 * We might be able to do something fancy, like setting
2266 * B_CACHE in bwrite() except if B_DELWRI is already set,
2267 * so the below call doesn't set B_CACHE, but that gets real
2268 * confusing. This is much easier.
2271 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2272 bp->b_flags |= B_NOCACHE;
2273 VOP_BWRITE(bp->b_vp, bp);
2279 * Buffer is not in-core, create new buffer. The buffer
2280 * returned by getnewbuf() is locked. Note that the returned
2281 * buffer is also considered valid (not marked B_INVAL).
2283 * Calculating the offset for the I/O requires figuring out
2284 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2285 * the mount's f_iosize otherwise. If the vnode does not
2286 * have an associated mount we assume that the passed size is
2289 * Note that vn_isdisk() cannot be used here since it may
2290 * return a failure for numerous reasons. Note that the
2291 * buffer size may be larger then the block size (the caller
2292 * will use block numbers with the proper multiple). Beware
2293 * of using any v_* fields which are part of unions. In
2294 * particular, in DragonFly the mount point overloading
2295 * mechanism is such that the underlying directory (with a
2296 * non-NULL v_mountedhere) is not a special case.
2300 if (vp->v_type == VBLK || vp->v_type == VCHR)
2302 else if (vp->v_mount)
2303 bsize = vp->v_mount->mnt_stat.f_iosize;
2307 maxsize = size + (loffset & PAGE_MASK);
2308 maxsize = imax(maxsize, bsize);
2310 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) {
2311 if (slpflag || slptimeo) {
2319 * This code is used to make sure that a buffer is not
2320 * created while the getnewbuf routine is blocked.
2321 * This can be a problem whether the vnode is locked or not.
2322 * If the buffer is created out from under us, we have to
2323 * throw away the one we just created. There is now window
2324 * race because we are safely running in a critical section
2325 * from the point of the duplicate buffer creation through
2326 * to here, and we've locked the buffer.
2328 if (findblk(vp, loffset)) {
2329 bp->b_flags |= B_INVAL;
2335 * Insert the buffer into the hash, so that it can
2336 * be found by findblk().
2338 * Make sure the translation layer has been cleared.
2340 bp->b_loffset = loffset;
2341 bp->b_bio2.bio_offset = NOOFFSET;
2342 /* bp->b_bio2.bio_next = NULL; */
2347 * All vnode-based buffers must be backed by a VM object.
2349 KKASSERT(vp->v_object != NULL);
2350 bp->b_flags |= B_VMIO;
2351 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2363 * Get an empty, disassociated buffer of given size. The buffer is
2364 * initially set to B_INVAL.
2366 * critical section protection is not required for the allocbuf()
2367 * call because races are impossible here.
2375 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2378 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2382 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2390 * This code constitutes the buffer memory from either anonymous system
2391 * memory (in the case of non-VMIO operations) or from an associated
2392 * VM object (in the case of VMIO operations). This code is able to
2393 * resize a buffer up or down.
2395 * Note that this code is tricky, and has many complications to resolve
2396 * deadlock or inconsistant data situations. Tread lightly!!!
2397 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2398 * the caller. Calling this code willy nilly can result in the loss of data.
2400 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2401 * B_CACHE for the non-VMIO case.
2403 * This routine does not need to be called from a critical section but you
2404 * must own the buffer.
2407 allocbuf(struct buf *bp, int size)
2409 int newbsize, mbsize;
2412 if (BUF_REFCNT(bp) == 0)
2413 panic("allocbuf: buffer not busy");
2415 if (bp->b_kvasize < size)
2416 panic("allocbuf: buffer too small");
2418 if ((bp->b_flags & B_VMIO) == 0) {
2422 * Just get anonymous memory from the kernel. Don't
2423 * mess with B_CACHE.
2425 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2426 if (bp->b_flags & B_MALLOC)
2429 newbsize = round_page(size);
2431 if (newbsize < bp->b_bufsize) {
2433 * malloced buffers are not shrunk
2435 if (bp->b_flags & B_MALLOC) {
2437 bp->b_bcount = size;
2439 free(bp->b_data, M_BIOBUF);
2440 if (bp->b_bufsize) {
2441 bufmallocspace -= bp->b_bufsize;
2445 bp->b_data = bp->b_kvabase;
2447 bp->b_flags &= ~B_MALLOC;
2453 (vm_offset_t) bp->b_data + newbsize,
2454 (vm_offset_t) bp->b_data + bp->b_bufsize);
2455 } else if (newbsize > bp->b_bufsize) {
2457 * We only use malloced memory on the first allocation.
2458 * and revert to page-allocated memory when the buffer
2461 if ((bufmallocspace < maxbufmallocspace) &&
2462 (bp->b_bufsize == 0) &&
2463 (mbsize <= PAGE_SIZE/2)) {
2465 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2466 bp->b_bufsize = mbsize;
2467 bp->b_bcount = size;
2468 bp->b_flags |= B_MALLOC;
2469 bufmallocspace += mbsize;
2475 * If the buffer is growing on its other-than-first
2476 * allocation, then we revert to the page-allocation
2479 if (bp->b_flags & B_MALLOC) {
2480 origbuf = bp->b_data;
2481 origbufsize = bp->b_bufsize;
2482 bp->b_data = bp->b_kvabase;
2483 if (bp->b_bufsize) {
2484 bufmallocspace -= bp->b_bufsize;
2488 bp->b_flags &= ~B_MALLOC;
2489 newbsize = round_page(newbsize);
2493 (vm_offset_t) bp->b_data + bp->b_bufsize,
2494 (vm_offset_t) bp->b_data + newbsize);
2496 bcopy(origbuf, bp->b_data, origbufsize);
2497 free(origbuf, M_BIOBUF);
2504 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2505 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
2506 newbsize + PAGE_MASK) >> PAGE_SHIFT;
2507 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
2509 if (bp->b_flags & B_MALLOC)
2510 panic("allocbuf: VMIO buffer can't be malloced");
2512 * Set B_CACHE initially if buffer is 0 length or will become
2515 if (size == 0 || bp->b_bufsize == 0)
2516 bp->b_flags |= B_CACHE;
2518 if (newbsize < bp->b_bufsize) {
2520 * DEV_BSIZE aligned new buffer size is less then the
2521 * DEV_BSIZE aligned existing buffer size. Figure out
2522 * if we have to remove any pages.
2524 if (desiredpages < bp->b_xio.xio_npages) {
2525 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
2527 * the page is not freed here -- it
2528 * is the responsibility of
2529 * vnode_pager_setsize
2531 m = bp->b_xio.xio_pages[i];
2532 KASSERT(m != bogus_page,
2533 ("allocbuf: bogus page found"));
2534 while (vm_page_sleep_busy(m, TRUE, "biodep"))
2537 bp->b_xio.xio_pages[i] = NULL;
2538 vm_page_unwire(m, 0);
2540 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2541 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
2542 bp->b_xio.xio_npages = desiredpages;
2544 } else if (size > bp->b_bcount) {
2546 * We are growing the buffer, possibly in a
2547 * byte-granular fashion.
2555 * Step 1, bring in the VM pages from the object,
2556 * allocating them if necessary. We must clear
2557 * B_CACHE if these pages are not valid for the
2558 * range covered by the buffer.
2560 * critical section protection is required to protect
2561 * against interrupts unbusying and freeing pages
2562 * between our vm_page_lookup() and our
2563 * busycheck/wiring call.
2569 while (bp->b_xio.xio_npages < desiredpages) {
2573 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
2574 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2576 * note: must allocate system pages
2577 * since blocking here could intefere
2578 * with paging I/O, no matter which
2581 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
2584 vm_pageout_deficit += desiredpages -
2585 bp->b_xio.xio_npages;
2589 bp->b_flags &= ~B_CACHE;
2590 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2591 ++bp->b_xio.xio_npages;
2597 * We found a page. If we have to sleep on it,
2598 * retry because it might have gotten freed out
2601 * We can only test PG_BUSY here. Blocking on
2602 * m->busy might lead to a deadlock:
2604 * vm_fault->getpages->cluster_read->allocbuf
2608 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
2612 * We have a good page. Should we wakeup the
2615 if ((curthread != pagethread) &&
2616 ((m->queue - m->pc) == PQ_CACHE) &&
2617 ((vmstats.v_free_count + vmstats.v_cache_count) <
2618 (vmstats.v_free_min + vmstats.v_cache_min))) {
2619 pagedaemon_wakeup();
2621 vm_page_flag_clear(m, PG_ZERO);
2623 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2624 ++bp->b_xio.xio_npages;
2629 * Step 2. We've loaded the pages into the buffer,
2630 * we have to figure out if we can still have B_CACHE
2631 * set. Note that B_CACHE is set according to the
2632 * byte-granular range ( bcount and size ), not the
2633 * aligned range ( newbsize ).
2635 * The VM test is against m->valid, which is DEV_BSIZE
2636 * aligned. Needless to say, the validity of the data
2637 * needs to also be DEV_BSIZE aligned. Note that this
2638 * fails with NFS if the server or some other client
2639 * extends the file's EOF. If our buffer is resized,
2640 * B_CACHE may remain set! XXX
2643 toff = bp->b_bcount;
2644 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
2646 while ((bp->b_flags & B_CACHE) && toff < size) {
2649 if (tinc > (size - toff))
2652 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
2660 bp->b_xio.xio_pages[pi]
2667 * Step 3, fixup the KVM pmap. Remember that
2668 * bp->b_data is relative to bp->b_loffset, but
2669 * bp->b_loffset may be offset into the first page.
2672 bp->b_data = (caddr_t)
2673 trunc_page((vm_offset_t)bp->b_data);
2675 (vm_offset_t)bp->b_data,
2676 bp->b_xio.xio_pages,
2677 bp->b_xio.xio_npages
2679 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2680 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
2683 if (newbsize < bp->b_bufsize)
2685 bp->b_bufsize = newbsize; /* actual buffer allocation */
2686 bp->b_bcount = size; /* requested buffer size */
2693 * Wait for buffer I/O completion, returning error status. The buffer
2694 * is left locked on return. B_EINTR is converted into an EINTR error
2697 * NOTE! The original b_cmd is lost on return, since b_cmd will be
2698 * set to BUF_CMD_DONE.
2701 biowait(struct buf * bp)
2704 while (bp->b_cmd != BUF_CMD_DONE) {
2705 if (bp->b_cmd == BUF_CMD_READ)
2706 tsleep(bp, 0, "biord", 0);
2708 tsleep(bp, 0, "biowr", 0);
2711 if (bp->b_flags & B_EINTR) {
2712 bp->b_flags &= ~B_EINTR;
2715 if (bp->b_flags & B_ERROR) {
2716 return (bp->b_error ? bp->b_error : EIO);
2723 * This associates a tracking count with an I/O. vn_strategy() and
2724 * dev_dstrategy() do this automatically but there are a few cases
2725 * where a vnode or device layer is bypassed when a block translation
2726 * is cached. In such cases bio_start_transaction() may be called on
2727 * the bypassed layers so the system gets an I/O in progress indication
2728 * for those higher layers.
2731 bio_start_transaction(struct bio *bio, struct bio_track *track)
2733 bio->bio_track = track;
2734 atomic_add_int(&track->bk_active, 1);
2738 * Initiate I/O on a vnode.
2741 vn_strategy(struct vnode *vp, struct bio *bio)
2743 struct bio_track *track;
2745 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
2746 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
2747 track = &vp->v_track_read;
2749 track = &vp->v_track_write;
2750 bio->bio_track = track;
2751 atomic_add_int(&track->bk_active, 1);
2752 vop_strategy(*vp->v_ops, vp, bio);
2759 * Finish I/O on a buffer, optionally calling a completion function.
2760 * This is usually called from an interrupt so process blocking is
2763 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
2764 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
2765 * assuming B_INVAL is clear.
2767 * For the VMIO case, we set B_CACHE if the op was a read and no
2768 * read error occured, or if the op was a write. B_CACHE is never
2769 * set if the buffer is invalid or otherwise uncacheable.
2771 * biodone does not mess with B_INVAL, allowing the I/O routine or the
2772 * initiator to leave B_INVAL set to brelse the buffer out of existance
2773 * in the biodone routine.
2776 biodone(struct bio *bio)
2778 struct buf *bp = bio->bio_buf;
2783 KASSERT(BUF_REFCNTNB(bp) > 0,
2784 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
2785 KASSERT(bp->b_cmd != BUF_CMD_DONE,
2786 ("biodone: bp %p already done!", bp));
2788 runningbufwakeup(bp);
2791 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
2794 biodone_t *done_func;
2795 struct bio_track *track;
2798 * BIO tracking. Most but not all BIOs are tracked.
2800 if ((track = bio->bio_track) != NULL) {
2801 atomic_subtract_int(&track->bk_active, 1);
2802 if (track->bk_active < 0) {
2803 panic("biodone: bad active count bio %p\n",
2806 if (track->bk_waitflag) {
2807 track->bk_waitflag = 0;
2810 bio->bio_track = NULL;
2814 * A bio_done function terminates the loop. The function
2815 * will be responsible for any further chaining and/or
2816 * buffer management.
2818 * WARNING! The done function can deallocate the buffer!
2820 if ((done_func = bio->bio_done) != NULL) {
2821 bio->bio_done = NULL;
2826 bio = bio->bio_prev;
2830 bp->b_cmd = BUF_CMD_DONE;
2833 * Only reads and writes are processed past this point.
2835 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
2842 * Warning: softupdates may re-dirty the buffer.
2844 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
2845 (*bioops.io_complete)(bp);
2847 if (bp->b_flags & B_VMIO) {
2853 struct vnode *vp = bp->b_vp;
2857 #if defined(VFS_BIO_DEBUG)
2858 if (vp->v_holdcnt == 0)
2859 panic("biodone: zero vnode hold count");
2860 if ((vp->v_flag & VOBJBUF) == 0)
2861 panic("biodone: vnode is not setup for merged cache");
2864 foff = bp->b_loffset;
2865 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
2866 KASSERT(obj != NULL, ("biodone: missing VM object"));
2868 #if defined(VFS_BIO_DEBUG)
2869 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
2870 printf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
2871 obj->paging_in_progress, bp->b_xio.xio_npages);
2876 * Set B_CACHE if the op was a normal read and no error
2877 * occured. B_CACHE is set for writes in the b*write()
2880 iosize = bp->b_bcount - bp->b_resid;
2881 if (cmd == BUF_CMD_READ && (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
2882 bp->b_flags |= B_CACHE;
2885 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2889 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2894 * cleanup bogus pages, restoring the originals. Since
2895 * the originals should still be wired, we don't have
2896 * to worry about interrupt/freeing races destroying
2897 * the VM object association.
2899 m = bp->b_xio.xio_pages[i];
2900 if (m == bogus_page) {
2902 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2904 panic("biodone: page disappeared");
2905 bp->b_xio.xio_pages[i] = m;
2906 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2907 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
2909 #if defined(VFS_BIO_DEBUG)
2910 if (OFF_TO_IDX(foff) != m->pindex) {
2912 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
2913 (unsigned long)foff, m->pindex);
2918 * In the write case, the valid and clean bits are
2919 * already changed correctly ( see bdwrite() ), so we
2920 * only need to do this here in the read case.
2922 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
2923 vfs_page_set_valid(bp, foff, i, m);
2925 vm_page_flag_clear(m, PG_ZERO);
2928 * when debugging new filesystems or buffer I/O methods, this
2929 * is the most common error that pops up. if you see this, you
2930 * have not set the page busy flag correctly!!!
2933 printf("biodone: page busy < 0, "
2934 "pindex: %d, foff: 0x(%x,%x), "
2935 "resid: %d, index: %d\n",
2936 (int) m->pindex, (int)(foff >> 32),
2937 (int) foff & 0xffffffff, resid, i);
2938 if (!vn_isdisk(vp, NULL))
2939 printf(" iosize: %ld, loffset: %lld, flags: 0x%08x, npages: %d\n",
2940 bp->b_vp->v_mount->mnt_stat.f_iosize,
2942 bp->b_flags, bp->b_xio.xio_npages);
2944 printf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
2946 bp->b_flags, bp->b_xio.xio_npages);
2947 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
2948 m->valid, m->dirty, m->wire_count);
2949 panic("biodone: page busy < 0");
2951 vm_page_io_finish(m);
2952 vm_object_pip_subtract(obj, 1);
2953 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2957 vm_object_pip_wakeupn(obj, 0);
2961 * For asynchronous completions, release the buffer now. The brelse
2962 * will do a wakeup there if necessary - so no need to do a wakeup
2963 * here in the async case. The sync case always needs to do a wakeup.
2966 if (bp->b_flags & B_ASYNC) {
2967 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
2980 * This routine is called in lieu of iodone in the case of
2981 * incomplete I/O. This keeps the busy status for pages
2985 vfs_unbusy_pages(struct buf *bp)
2989 runningbufwakeup(bp);
2990 if (bp->b_flags & B_VMIO) {
2991 struct vnode *vp = bp->b_vp;
2996 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2997 vm_page_t m = bp->b_xio.xio_pages[i];
3000 * When restoring bogus changes the original pages
3001 * should still be wired, so we are in no danger of
3002 * losing the object association and do not need
3003 * critical section protection particularly.
3005 if (m == bogus_page) {
3006 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3008 panic("vfs_unbusy_pages: page missing");
3010 bp->b_xio.xio_pages[i] = m;
3011 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3012 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3014 vm_object_pip_subtract(obj, 1);
3015 vm_page_flag_clear(m, PG_ZERO);
3016 vm_page_io_finish(m);
3018 vm_object_pip_wakeupn(obj, 0);
3023 * vfs_page_set_valid:
3025 * Set the valid bits in a page based on the supplied offset. The
3026 * range is restricted to the buffer's size.
3028 * This routine is typically called after a read completes.
3031 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3033 vm_ooffset_t soff, eoff;
3036 * Start and end offsets in buffer. eoff - soff may not cross a
3037 * page boundry or cross the end of the buffer. The end of the
3038 * buffer, in this case, is our file EOF, not the allocation size
3042 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3043 if (eoff > bp->b_loffset + bp->b_bcount)
3044 eoff = bp->b_loffset + bp->b_bcount;
3047 * Set valid range. This is typically the entire buffer and thus the
3051 vm_page_set_validclean(
3053 (vm_offset_t) (soff & PAGE_MASK),
3054 (vm_offset_t) (eoff - soff)
3062 * This routine is called before a device strategy routine.
3063 * It is used to tell the VM system that paging I/O is in
3064 * progress, and treat the pages associated with the buffer
3065 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3066 * flag is handled to make sure that the object doesn't become
3069 * Since I/O has not been initiated yet, certain buffer flags
3070 * such as B_ERROR or B_INVAL may be in an inconsistant state
3071 * and should be ignored.
3074 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3077 struct proc *p = curthread->td_proc;
3080 * The buffer's I/O command must already be set. If reading,
3081 * B_CACHE must be 0 (double check against callers only doing
3082 * I/O when B_CACHE is 0).
3084 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3085 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3087 if (bp->b_flags & B_VMIO) {
3092 foff = bp->b_loffset;
3093 KASSERT(bp->b_loffset != NOOFFSET,
3094 ("vfs_busy_pages: no buffer offset"));
3098 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3099 vm_page_t m = bp->b_xio.xio_pages[i];
3100 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3105 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3106 vm_page_t m = bp->b_xio.xio_pages[i];
3108 vm_page_flag_clear(m, PG_ZERO);
3109 if ((bp->b_flags & B_CLUSTER) == 0) {
3110 vm_object_pip_add(obj, 1);
3111 vm_page_io_start(m);
3115 * When readying a vnode-backed buffer for a write
3116 * we must zero-fill any invalid portions of the
3119 * When readying a vnode-backed buffer for a read
3120 * we must replace any dirty pages with a bogus
3121 * page so we do not destroy dirty data when
3122 * filling in gaps. Dirty pages might not
3123 * necessarily be marked dirty yet, so use m->valid
3124 * as a reasonable test.
3126 * Bogus page replacement is, uh, bogus. We need
3127 * to find a better way.
3129 vm_page_protect(m, VM_PROT_NONE);
3130 if (bp->b_cmd == BUF_CMD_WRITE) {
3131 vfs_page_set_valid(bp, foff, i, m);
3132 } else if (m->valid == VM_PAGE_BITS_ALL) {
3133 bp->b_xio.xio_pages[i] = bogus_page;
3136 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3139 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3140 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3144 * This is the easiest place to put the process accounting for the I/O
3148 if (bp->b_cmd == BUF_CMD_READ)
3149 p->p_stats->p_ru.ru_inblock++;
3151 p->p_stats->p_ru.ru_oublock++;
3158 * Tell the VM system that the pages associated with this buffer
3159 * are clean. This is used for delayed writes where the data is
3160 * going to go to disk eventually without additional VM intevention.
3162 * Note that while we only really need to clean through to b_bcount, we
3163 * just go ahead and clean through to b_bufsize.
3166 vfs_clean_pages(struct buf *bp)
3170 if (bp->b_flags & B_VMIO) {
3173 foff = bp->b_loffset;
3174 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset"));
3175 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3176 vm_page_t m = bp->b_xio.xio_pages[i];
3177 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3178 vm_ooffset_t eoff = noff;
3180 if (eoff > bp->b_loffset + bp->b_bufsize)
3181 eoff = bp->b_loffset + bp->b_bufsize;
3182 vfs_page_set_valid(bp, foff, i, m);
3183 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3190 * vfs_bio_set_validclean:
3192 * Set the range within the buffer to valid and clean. The range is
3193 * relative to the beginning of the buffer, b_loffset. Note that
3194 * b_loffset itself may be offset from the beginning of the first page.
3198 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3200 if (bp->b_flags & B_VMIO) {
3205 * Fixup base to be relative to beginning of first page.
3206 * Set initial n to be the maximum number of bytes in the
3207 * first page that can be validated.
3210 base += (bp->b_loffset & PAGE_MASK);
3211 n = PAGE_SIZE - (base & PAGE_MASK);
3213 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) {
3214 vm_page_t m = bp->b_xio.xio_pages[i];
3219 vm_page_set_validclean(m, base & PAGE_MASK, n);
3230 * Clear a buffer. This routine essentially fakes an I/O, so we need
3231 * to clear B_ERROR and B_INVAL.
3233 * Note that while we only theoretically need to clear through b_bcount,
3234 * we go ahead and clear through b_bufsize.
3238 vfs_bio_clrbuf(struct buf *bp)
3242 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3243 bp->b_flags &= ~(B_INVAL|B_ERROR);
3244 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3245 (bp->b_loffset & PAGE_MASK) == 0) {
3246 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3247 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3251 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3252 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3253 bzero(bp->b_data, bp->b_bufsize);
3254 bp->b_xio.xio_pages[0]->valid |= mask;
3259 ea = sa = bp->b_data;
3260 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3261 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3262 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3263 ea = (caddr_t)(vm_offset_t)ulmin(
3264 (u_long)(vm_offset_t)ea,
3265 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3266 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3267 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3269 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3270 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3274 for (; sa < ea; sa += DEV_BSIZE, j++) {
3275 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3276 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3277 bzero(sa, DEV_BSIZE);
3280 bp->b_xio.xio_pages[i]->valid |= mask;
3281 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
3290 * vm_hold_load_pages:
3292 * Load pages into the buffer's address space. The pages are
3293 * allocated from the kernel object in order to reduce interference
3294 * with the any VM paging I/O activity. The range of loaded
3295 * pages will be wired.
3297 * If a page cannot be allocated, the 'pagedaemon' is woken up to
3298 * retrieve the full range (to - from) of pages.
3302 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3308 to = round_page(to);
3309 from = round_page(from);
3310 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3312 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3317 * Note: must allocate system pages since blocking here
3318 * could intefere with paging I/O, no matter which
3321 p = vm_page_alloc(kernel_object,
3322 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3323 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
3325 vm_pageout_deficit += (to - from) >> PAGE_SHIFT;
3330 p->valid = VM_PAGE_BITS_ALL;
3331 vm_page_flag_clear(p, PG_ZERO);
3332 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3333 bp->b_xio.xio_pages[index] = p;
3336 bp->b_xio.xio_npages = index;
3340 * vm_hold_free_pages:
3342 * Return pages associated with the buffer back to the VM system.
3344 * The range of pages underlying the buffer's address space will
3345 * be unmapped and un-wired.
3348 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3352 int index, newnpages;
3354 from = round_page(from);
3355 to = round_page(to);
3356 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3358 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3359 p = bp->b_xio.xio_pages[index];
3360 if (p && (index < bp->b_xio.xio_npages)) {
3362 printf("vm_hold_free_pages: doffset: %lld, loffset: %lld\n",
3363 bp->b_bio2.bio_offset, bp->b_loffset);
3365 bp->b_xio.xio_pages[index] = NULL;
3368 vm_page_unwire(p, 0);
3372 bp->b_xio.xio_npages = newnpages;
3378 * Map an IO request into kernel virtual address space.
3380 * All requests are (re)mapped into kernel VA space.
3381 * Notice that we use b_bufsize for the size of the buffer
3382 * to be mapped. b_bcount might be modified by the driver.
3385 vmapbuf(struct buf *bp)
3387 caddr_t addr, v, kva;
3394 * bp had better have a command
3396 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3398 if (bp->b_bufsize < 0)
3400 for (v = bp->b_saveaddr,
3401 addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data),
3403 addr < bp->b_data + bp->b_bufsize;
3404 addr += PAGE_SIZE, v += PAGE_SIZE, pidx++) {
3406 * Do the vm_fault if needed; do the copy-on-write thing
3407 * when reading stuff off device into memory.
3410 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data,
3411 (bp->b_cmd == BUF_CMD_READ)?(VM_PROT_READ|VM_PROT_WRITE):VM_PROT_READ);
3413 for (i = 0; i < pidx; ++i) {
3414 vm_page_unhold(bp->b_xio.xio_pages[i]);
3415 bp->b_xio.xio_pages[i] = NULL;
3421 * WARNING! If sparc support is MFCd in the future this will
3422 * have to be changed from pmap_kextract() to pmap_extract()
3426 #error "If MFCing sparc support use pmap_extract"
3428 pa = pmap_kextract((vm_offset_t)addr);
3430 printf("vmapbuf: warning, race against user address during I/O");
3433 m = PHYS_TO_VM_PAGE(pa);
3435 bp->b_xio.xio_pages[pidx] = m;
3437 if (pidx > btoc(MAXPHYS))
3438 panic("vmapbuf: mapped more than MAXPHYS");
3439 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_xio.xio_pages, pidx);
3441 kva = bp->b_saveaddr;
3442 bp->b_xio.xio_npages = pidx;
3443 bp->b_saveaddr = bp->b_data;
3444 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3451 * Free the io map PTEs associated with this IO operation.
3452 * We also invalidate the TLB entries and restore the original b_addr.
3455 vunmapbuf(struct buf *bp)
3461 npages = bp->b_xio.xio_npages;
3462 pmap_qremove(trunc_page((vm_offset_t)bp->b_data),
3464 m = bp->b_xio.xio_pages;
3465 for (pidx = 0; pidx < npages; pidx++)
3466 vm_page_unhold(*m++);
3468 bp->b_data = bp->b_saveaddr;
3472 * Scan all buffers in the system and issue the callback.
3475 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
3481 for (n = 0; n < nbuf; ++n) {
3482 if ((error = callback(&buf[n], info)) < 0) {
3492 * print out statistics from the current status of the buffer pool
3493 * this can be toggeled by the system control option debug.syncprt
3502 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
3503 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
3505 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
3507 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
3510 TAILQ_FOREACH(bp, dp, b_freelist) {
3511 counts[bp->b_bufsize/PAGE_SIZE]++;
3515 printf("%s: total-%d", bname[i], count);
3516 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
3518 printf(", %d-%d", j * PAGE_SIZE, counts[j]);
3524 #include "opt_ddb.h"
3526 #include <ddb/ddb.h>
3528 DB_SHOW_COMMAND(buffer, db_show_buffer)
3531 struct buf *bp = (struct buf *)addr;
3534 db_printf("usage: show buffer <addr>\n");
3538 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3539 db_printf("b_cmd = %d\n", bp->b_cmd);
3540 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
3541 "b_resid = %d\n, b_data = %p, "
3542 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
3543 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3544 bp->b_data, bp->b_bio2.bio_offset, (bp->b_bio2.bio_next ? bp->b_bio2.bio_next->bio_offset : (off_t)-1));
3545 if (bp->b_xio.xio_npages) {
3547 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
3548 bp->b_xio.xio_npages);
3549 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3551 m = bp->b_xio.xio_pages[i];
3552 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3553 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3554 if ((i + 1) < bp->b_xio.xio_npages)