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 1.2 2003/06/17 04:28:41 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>
48 #include <vm/vm_param.h>
49 #include <vm/vm_kern.h>
50 #include <vm/vm_pageout.h>
51 #include <vm/vm_page.h>
52 #include <vm/vm_object.h>
53 #include <vm/vm_extern.h>
54 #include <vm/vm_map.h>
56 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
58 struct bio_ops bioops; /* I/O operation notification */
60 struct buf *buf; /* buffer header pool */
61 struct swqueue bswlist;
63 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from,
65 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from,
67 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
68 int pageno, vm_page_t m);
69 static void vfs_clean_pages(struct buf * bp);
70 static void vfs_setdirty(struct buf *bp);
71 static void vfs_vmio_release(struct buf *bp);
72 static void vfs_backgroundwritedone(struct buf *bp);
73 static int flushbufqueues(void);
75 static int bd_request;
77 static void buf_daemon __P((void));
79 * bogus page -- for I/O to/from partially complete buffers
80 * this is a temporary solution to the problem, but it is not
81 * really that bad. it would be better to split the buffer
82 * for input in the case of buffers partially already in memory,
83 * but the code is intricate enough already.
86 int vmiodirenable = TRUE;
88 static vm_offset_t bogus_offset;
90 static int bufspace, maxbufspace,
91 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
92 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
93 static int needsbuffer;
94 static int lorunningspace, hirunningspace, runningbufreq;
95 static int numdirtybuffers, lodirtybuffers, hidirtybuffers;
96 static int numfreebuffers, lofreebuffers, hifreebuffers;
97 static int getnewbufcalls;
98 static int getnewbufrestarts;
100 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD,
101 &numdirtybuffers, 0, "");
102 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW,
103 &lodirtybuffers, 0, "");
104 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW,
105 &hidirtybuffers, 0, "");
106 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD,
107 &numfreebuffers, 0, "");
108 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW,
109 &lofreebuffers, 0, "");
110 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW,
111 &hifreebuffers, 0, "");
112 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD,
113 &runningbufspace, 0, "");
114 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW,
115 &lorunningspace, 0, "");
116 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW,
117 &hirunningspace, 0, "");
118 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD,
119 &maxbufspace, 0, "");
120 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD,
122 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD,
124 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD,
126 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW,
127 &maxbufmallocspace, 0, "");
128 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD,
129 &bufmallocspace, 0, "");
130 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW,
131 &getnewbufcalls, 0, "");
132 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW,
133 &getnewbufrestarts, 0, "");
134 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW,
135 &vmiodirenable, 0, "");
136 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW,
137 &bufdefragcnt, 0, "");
138 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW,
139 &buffreekvacnt, 0, "");
140 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW,
141 &bufreusecnt, 0, "");
143 static int bufhashmask;
144 static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
145 struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } };
146 char *buf_wmesg = BUF_WMESG;
148 extern int vm_swap_size;
150 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
151 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
152 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
153 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
156 * Buffer hash table code. Note that the logical block scans linearly, which
157 * gives us some L1 cache locality.
162 bufhash(struct vnode *vnp, daddr_t bn)
164 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]);
170 * If someone is blocked due to there being too many dirty buffers,
171 * and numdirtybuffers is now reasonable, wake them up.
175 numdirtywakeup(int level)
177 if (numdirtybuffers <= level) {
178 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
179 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
180 wakeup(&needsbuffer);
188 * Called when buffer space is potentially available for recovery.
189 * getnewbuf() will block on this flag when it is unable to free
190 * sufficient buffer space. Buffer space becomes recoverable when
191 * bp's get placed back in the queues.
198 * If someone is waiting for BUF space, wake them up. Even
199 * though we haven't freed the kva space yet, the waiting
200 * process will be able to now.
202 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
203 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
204 wakeup(&needsbuffer);
209 * runningbufwakeup() - in-progress I/O accounting.
213 runningbufwakeup(struct buf *bp)
215 if (bp->b_runningbufspace) {
216 runningbufspace -= bp->b_runningbufspace;
217 bp->b_runningbufspace = 0;
218 if (runningbufreq && runningbufspace <= lorunningspace) {
220 wakeup(&runningbufreq);
228 * Called when a buffer has been added to one of the free queues to
229 * account for the buffer and to wakeup anyone waiting for free buffers.
230 * This typically occurs when large amounts of metadata are being handled
231 * by the buffer cache ( else buffer space runs out first, usually ).
239 needsbuffer &= ~VFS_BIO_NEED_ANY;
240 if (numfreebuffers >= hifreebuffers)
241 needsbuffer &= ~VFS_BIO_NEED_FREE;
242 wakeup(&needsbuffer);
247 * waitrunningbufspace()
249 * runningbufspace is a measure of the amount of I/O currently
250 * running. This routine is used in async-write situations to
251 * prevent creating huge backups of pending writes to a device.
252 * Only asynchronous writes are governed by this function.
254 * Reads will adjust runningbufspace, but will not block based on it.
255 * The read load has a side effect of reducing the allowed write load.
257 * This does NOT turn an async write into a sync write. It waits
258 * for earlier writes to complete and generally returns before the
259 * caller's write has reached the device.
262 waitrunningbufspace(void)
264 while (runningbufspace > hirunningspace) {
267 s = splbio(); /* fix race against interrupt/biodone() */
269 tsleep(&runningbufreq, PVM, "wdrain", 0);
275 * vfs_buf_test_cache:
277 * Called when a buffer is extended. This function clears the B_CACHE
278 * bit if the newly extended portion of the buffer does not contain
283 vfs_buf_test_cache(struct buf *bp,
284 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
287 if (bp->b_flags & B_CACHE) {
288 int base = (foff + off) & PAGE_MASK;
289 if (vm_page_is_valid(m, base, size) == 0)
290 bp->b_flags &= ~B_CACHE;
296 bd_wakeup(int dirtybuflevel)
298 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
305 * bd_speedup - speedup the buffer cache flushing code
316 * Initialize buffer headers and related structures.
320 bufhashinit(caddr_t vaddr)
322 /* first, make a null hash table */
323 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1)
325 bufhashtbl = (void *)vaddr;
326 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask;
337 TAILQ_INIT(&bswlist);
338 LIST_INIT(&invalhash);
339 simple_lock_init(&buftimelock);
341 for (i = 0; i <= bufhashmask; i++)
342 LIST_INIT(&bufhashtbl[i]);
344 /* next, make a null set of free lists */
345 for (i = 0; i < BUFFER_QUEUES; i++)
346 TAILQ_INIT(&bufqueues[i]);
348 /* finally, initialize each buffer header and stick on empty q */
349 for (i = 0; i < nbuf; i++) {
351 bzero(bp, sizeof *bp);
352 bp->b_flags = B_INVAL; /* we're just an empty header */
354 bp->b_rcred = NOCRED;
355 bp->b_wcred = NOCRED;
356 bp->b_qindex = QUEUE_EMPTY;
358 LIST_INIT(&bp->b_dep);
360 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
361 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
365 * maxbufspace is the absolute maximum amount of buffer space we are
366 * allowed to reserve in KVM and in real terms. The absolute maximum
367 * is nominally used by buf_daemon. hibufspace is the nominal maximum
368 * used by most other processes. The differential is required to
369 * ensure that buf_daemon is able to run when other processes might
370 * be blocked waiting for buffer space.
372 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
373 * this may result in KVM fragmentation which is not handled optimally
376 maxbufspace = nbuf * BKVASIZE;
377 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
378 lobufspace = hibufspace - MAXBSIZE;
380 lorunningspace = 512 * 1024;
381 hirunningspace = 1024 * 1024;
384 * Limit the amount of malloc memory since it is wired permanently into
385 * the kernel space. Even though this is accounted for in the buffer
386 * allocation, we don't want the malloced region to grow uncontrolled.
387 * The malloc scheme improves memory utilization significantly on average
388 * (small) directories.
390 maxbufmallocspace = hibufspace / 20;
393 * Reduce the chance of a deadlock occuring by limiting the number
394 * of delayed-write dirty buffers we allow to stack up.
396 hidirtybuffers = nbuf / 4 + 20;
399 * To support extreme low-memory systems, make sure hidirtybuffers cannot
400 * eat up all available buffer space. This occurs when our minimum cannot
401 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
402 * BKVASIZE'd (8K) buffers.
404 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
405 hidirtybuffers >>= 1;
407 lodirtybuffers = hidirtybuffers / 2;
410 * Try to keep the number of free buffers in the specified range,
411 * and give special processes (e.g. like buf_daemon) access to an
414 lofreebuffers = nbuf / 18 + 5;
415 hifreebuffers = 2 * lofreebuffers;
416 numfreebuffers = nbuf;
419 * Maximum number of async ops initiated per buf_daemon loop. This is
420 * somewhat of a hack at the moment, we really need to limit ourselves
421 * based on the number of bytes of I/O in-transit that were initiated
425 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE);
426 bogus_page = vm_page_alloc(kernel_object,
427 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
434 * bfreekva() - free the kva allocation for a buffer.
436 * Must be called at splbio() or higher as this is the only locking for
439 * Since this call frees up buffer space, we call bufspacewakeup().
442 bfreekva(struct buf * bp)
446 vm_map_lock(buffer_map);
447 bufspace -= bp->b_kvasize;
448 vm_map_delete(buffer_map,
449 (vm_offset_t) bp->b_kvabase,
450 (vm_offset_t) bp->b_kvabase + bp->b_kvasize
452 vm_map_unlock(buffer_map);
461 * Remove the buffer from the appropriate free list.
464 bremfree(struct buf * bp)
467 int old_qindex = bp->b_qindex;
469 if (bp->b_qindex != QUEUE_NONE) {
470 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp));
471 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
472 bp->b_qindex = QUEUE_NONE;
474 if (BUF_REFCNT(bp) <= 1)
475 panic("bremfree: removing a buffer not on a queue");
479 * Fixup numfreebuffers count. If the buffer is invalid or not
480 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
481 * the buffer was free and we must decrement numfreebuffers.
483 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
500 * Get a buffer with the specified data. Look in the cache first. We
501 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
502 * is set, the buffer is valid and we do not have to do anything ( see
506 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
511 bp = getblk(vp, blkno, size, 0, 0);
514 /* if not found in cache, do some I/O */
515 if ((bp->b_flags & B_CACHE) == 0) {
517 curproc->p_stats->p_ru.ru_inblock++;
518 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp));
519 bp->b_flags |= B_READ;
520 bp->b_flags &= ~(B_ERROR | B_INVAL);
521 if (bp->b_rcred == NOCRED) {
526 vfs_busy_pages(bp, 0);
527 VOP_STRATEGY(vp, bp);
528 return (biowait(bp));
534 * Operates like bread, but also starts asynchronous I/O on
535 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
536 * to initiating I/O . If B_CACHE is set, the buffer is valid
537 * and we do not have to do anything.
540 breadn(struct vnode * vp, daddr_t blkno, int size,
541 daddr_t * rablkno, int *rabsize,
542 int cnt, struct ucred * cred, struct buf ** bpp)
544 struct buf *bp, *rabp;
546 int rv = 0, readwait = 0;
548 *bpp = bp = getblk(vp, blkno, size, 0, 0);
550 /* if not found in cache, do some I/O */
551 if ((bp->b_flags & B_CACHE) == 0) {
553 curproc->p_stats->p_ru.ru_inblock++;
554 bp->b_flags |= B_READ;
555 bp->b_flags &= ~(B_ERROR | B_INVAL);
556 if (bp->b_rcred == NOCRED) {
561 vfs_busy_pages(bp, 0);
562 VOP_STRATEGY(vp, bp);
566 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
567 if (inmem(vp, *rablkno))
569 rabp = getblk(vp, *rablkno, *rabsize, 0, 0);
571 if ((rabp->b_flags & B_CACHE) == 0) {
573 curproc->p_stats->p_ru.ru_inblock++;
574 rabp->b_flags |= B_READ | B_ASYNC;
575 rabp->b_flags &= ~(B_ERROR | B_INVAL);
576 if (rabp->b_rcred == NOCRED) {
579 rabp->b_rcred = cred;
581 vfs_busy_pages(rabp, 0);
583 VOP_STRATEGY(vp, rabp);
596 * Write, release buffer on completion. (Done by iodone
597 * if async). Do not bother writing anything if the buffer
600 * Note that we set B_CACHE here, indicating that buffer is
601 * fully valid and thus cacheable. This is true even of NFS
602 * now so we set it generally. This could be set either here
603 * or in biodone() since the I/O is synchronous. We put it
607 bwrite(struct buf * bp)
612 if (bp->b_flags & B_INVAL) {
617 oldflags = bp->b_flags;
619 if (BUF_REFCNT(bp) == 0)
620 panic("bwrite: buffer is not busy???");
623 * If a background write is already in progress, delay
624 * writing this block if it is asynchronous. Otherwise
625 * wait for the background write to complete.
627 if (bp->b_xflags & BX_BKGRDINPROG) {
628 if (bp->b_flags & B_ASYNC) {
633 bp->b_xflags |= BX_BKGRDWAIT;
634 tsleep(&bp->b_xflags, PRIBIO, "biord", 0);
635 if (bp->b_xflags & BX_BKGRDINPROG)
636 panic("bwrite: still writing");
639 /* Mark the buffer clean */
643 * If this buffer is marked for background writing and we
644 * do not have to wait for it, make a copy and write the
645 * copy so as to leave this buffer ready for further use.
647 * This optimization eats a lot of memory. If we have a page
648 * or buffer shortfull we can't do it.
650 if ((bp->b_xflags & BX_BKGRDWRITE) &&
651 (bp->b_flags & B_ASYNC) &&
652 !vm_page_count_severe() &&
653 !buf_dirty_count_severe()) {
654 if (bp->b_flags & B_CALL)
655 panic("bwrite: need chained iodone");
657 /* get a new block */
658 newbp = geteblk(bp->b_bufsize);
660 /* set it to be identical to the old block */
661 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize);
662 bgetvp(bp->b_vp, newbp);
663 newbp->b_lblkno = bp->b_lblkno;
664 newbp->b_blkno = bp->b_blkno;
665 newbp->b_offset = bp->b_offset;
666 newbp->b_iodone = vfs_backgroundwritedone;
667 newbp->b_flags |= B_ASYNC | B_CALL;
668 newbp->b_flags &= ~B_INVAL;
670 /* move over the dependencies */
671 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps)
672 (*bioops.io_movedeps)(bp, newbp);
675 * Initiate write on the copy, release the original to
676 * the B_LOCKED queue so that it cannot go away until
677 * the background write completes. If not locked it could go
678 * away and then be reconstituted while it was being written.
679 * If the reconstituted buffer were written, we could end up
680 * with two background copies being written at the same time.
682 bp->b_xflags |= BX_BKGRDINPROG;
683 bp->b_flags |= B_LOCKED;
688 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
689 bp->b_flags |= B_WRITEINPROG | B_CACHE;
691 bp->b_vp->v_numoutput++;
692 vfs_busy_pages(bp, 1);
695 * Normal bwrites pipeline writes
697 bp->b_runningbufspace = bp->b_bufsize;
698 runningbufspace += bp->b_runningbufspace;
701 curproc->p_stats->p_ru.ru_oublock++;
703 if (oldflags & B_ASYNC)
705 VOP_STRATEGY(bp->b_vp, bp);
707 if ((oldflags & B_ASYNC) == 0) {
708 int rtval = biowait(bp);
711 } else if ((oldflags & B_NOWDRAIN) == 0) {
713 * don't allow the async write to saturate the I/O
714 * system. Deadlocks can occur only if a device strategy
715 * routine (like in VN) turns around and issues another
716 * high-level write, in which case B_NOWDRAIN is expected
717 * to be set. Otherwise we will not deadlock here because
718 * we are blocking waiting for I/O that is already in-progress
721 waitrunningbufspace();
728 * Complete a background write started from bwrite.
731 vfs_backgroundwritedone(bp)
737 * Find the original buffer that we are writing.
739 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL)
740 panic("backgroundwritedone: lost buffer");
742 * Process dependencies then return any unfinished ones.
744 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
745 (*bioops.io_complete)(bp);
746 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps)
747 (*bioops.io_movedeps)(bp, origbp);
749 * Clear the BX_BKGRDINPROG flag in the original buffer
750 * and awaken it if it is waiting for the write to complete.
751 * If BX_BKGRDINPROG is not set in the original buffer it must
752 * have been released and re-instantiated - which is not legal.
754 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), ("backgroundwritedone: lost buffer2"));
755 origbp->b_xflags &= ~BX_BKGRDINPROG;
756 if (origbp->b_xflags & BX_BKGRDWAIT) {
757 origbp->b_xflags &= ~BX_BKGRDWAIT;
758 wakeup(&origbp->b_xflags);
761 * Clear the B_LOCKED flag and remove it from the locked
762 * queue if it currently resides there.
764 origbp->b_flags &= ~B_LOCKED;
765 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
770 * This buffer is marked B_NOCACHE, so when it is released
771 * by biodone, it will be tossed. We mark it with B_READ
772 * to avoid biodone doing a second vwakeup.
774 bp->b_flags |= B_NOCACHE | B_READ;
775 bp->b_flags &= ~(B_CACHE | B_CALL | B_DONE);
781 * Delayed write. (Buffer is marked dirty). Do not bother writing
782 * anything if the buffer is marked invalid.
784 * Note that since the buffer must be completely valid, we can safely
785 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
786 * biodone() in order to prevent getblk from writing the buffer
790 bdwrite(struct buf * bp)
792 if (BUF_REFCNT(bp) == 0)
793 panic("bdwrite: buffer is not busy");
795 if (bp->b_flags & B_INVAL) {
802 * Set B_CACHE, indicating that the buffer is fully valid. This is
803 * true even of NFS now.
805 bp->b_flags |= B_CACHE;
808 * This bmap keeps the system from needing to do the bmap later,
809 * perhaps when the system is attempting to do a sync. Since it
810 * is likely that the indirect block -- or whatever other datastructure
811 * that the filesystem needs is still in memory now, it is a good
812 * thing to do this. Note also, that if the pageout daemon is
813 * requesting a sync -- there might not be enough memory to do
814 * the bmap then... So, this is important to do.
816 if (bp->b_lblkno == bp->b_blkno) {
817 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
821 * Set the *dirty* buffer range based upon the VM system dirty pages.
826 * We need to do this here to satisfy the vnode_pager and the
827 * pageout daemon, so that it thinks that the pages have been
828 * "cleaned". Note that since the pages are in a delayed write
829 * buffer -- the VFS layer "will" see that the pages get written
830 * out on the next sync, or perhaps the cluster will be completed.
836 * Wakeup the buffer flushing daemon if we have a lot of dirty
837 * buffers (midpoint between our recovery point and our stall
840 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
843 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
844 * due to the softdep code.
851 * Turn buffer into delayed write request. We must clear B_READ and
852 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
853 * itself to properly update it in the dirty/clean lists. We mark it
854 * B_DONE to ensure that any asynchronization of the buffer properly
855 * clears B_DONE ( else a panic will occur later ).
857 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
858 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
859 * should only be called if the buffer is known-good.
861 * Since the buffer is not on a queue, we do not update the numfreebuffers
864 * Must be called at splbio().
865 * The buffer must be on QUEUE_NONE.
871 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
872 bp->b_flags &= ~(B_READ|B_RELBUF);
874 if ((bp->b_flags & B_DELWRI) == 0) {
875 bp->b_flags |= B_DONE | B_DELWRI;
876 reassignbuf(bp, bp->b_vp);
878 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
885 * Clear B_DELWRI for buffer.
887 * Since the buffer is not on a queue, we do not update the numfreebuffers
890 * Must be called at splbio().
891 * The buffer must be on QUEUE_NONE.
898 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
900 if (bp->b_flags & B_DELWRI) {
901 bp->b_flags &= ~B_DELWRI;
902 reassignbuf(bp, bp->b_vp);
904 numdirtywakeup(lodirtybuffers);
907 * Since it is now being written, we can clear its deferred write flag.
909 bp->b_flags &= ~B_DEFERRED;
915 * Asynchronous write. Start output on a buffer, but do not wait for
916 * it to complete. The buffer is released when the output completes.
918 * bwrite() ( or the VOP routine anyway ) is responsible for handling
919 * B_INVAL buffers. Not us.
922 bawrite(struct buf * bp)
924 bp->b_flags |= B_ASYNC;
925 (void) VOP_BWRITE(bp->b_vp, bp);
931 * Ordered write. Start output on a buffer, and flag it so that the
932 * device will write it in the order it was queued. The buffer is
933 * released when the output completes. bwrite() ( or the VOP routine
934 * anyway ) is responsible for handling B_INVAL buffers.
937 bowrite(struct buf * bp)
939 bp->b_flags |= B_ORDERED | B_ASYNC;
940 return (VOP_BWRITE(bp->b_vp, bp));
946 * Called prior to the locking of any vnodes when we are expecting to
947 * write. We do not want to starve the buffer cache with too many
948 * dirty buffers so we block here. By blocking prior to the locking
949 * of any vnodes we attempt to avoid the situation where a locked vnode
950 * prevents the various system daemons from flushing related buffers.
956 if (numdirtybuffers >= hidirtybuffers) {
960 while (numdirtybuffers >= hidirtybuffers) {
962 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
963 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0);
970 * Return true if we have too many dirty buffers.
973 buf_dirty_count_severe(void)
975 return(numdirtybuffers >= hidirtybuffers);
981 * Release a busy buffer and, if requested, free its resources. The
982 * buffer will be stashed in the appropriate bufqueue[] allowing it
983 * to be accessed later as a cache entity or reused for other purposes.
986 brelse(struct buf * bp)
990 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
994 if (bp->b_flags & B_LOCKED)
995 bp->b_flags &= ~B_ERROR;
997 if ((bp->b_flags & (B_READ | B_ERROR | B_INVAL)) == B_ERROR) {
999 * Failed write, redirty. Must clear B_ERROR to prevent
1000 * pages from being scrapped. If B_INVAL is set then
1001 * this case is not run and the next case is run to
1002 * destroy the buffer. B_INVAL can occur if the buffer
1003 * is outside the range supported by the underlying device.
1005 bp->b_flags &= ~B_ERROR;
1007 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) ||
1008 (bp->b_bufsize <= 0)) {
1010 * Either a failed I/O or we were asked to free or not
1013 bp->b_flags |= B_INVAL;
1014 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
1015 (*bioops.io_deallocate)(bp);
1016 if (bp->b_flags & B_DELWRI) {
1018 numdirtywakeup(lodirtybuffers);
1020 bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF);
1021 if ((bp->b_flags & B_VMIO) == 0) {
1030 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1031 * is called with B_DELWRI set, the underlying pages may wind up
1032 * getting freed causing a previous write (bdwrite()) to get 'lost'
1033 * because pages associated with a B_DELWRI bp are marked clean.
1035 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1036 * if B_DELWRI is set.
1038 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1039 * on pages to return pages to the VM page queues.
1041 if (bp->b_flags & B_DELWRI)
1042 bp->b_flags &= ~B_RELBUF;
1043 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG))
1044 bp->b_flags |= B_RELBUF;
1047 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1048 * constituted, not even NFS buffers now. Two flags effect this. If
1049 * B_INVAL, the struct buf is invalidated but the VM object is kept
1050 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1052 * If B_ERROR or B_NOCACHE is set, pages in the VM object will be
1053 * invalidated. B_ERROR cannot be set for a failed write unless the
1054 * buffer is also B_INVAL because it hits the re-dirtying code above.
1056 * Normally we can do this whether a buffer is B_DELWRI or not. If
1057 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1058 * the commit state and we cannot afford to lose the buffer. If the
1059 * buffer has a background write in progress, we need to keep it
1060 * around to prevent it from being reconstituted and starting a second
1063 if ((bp->b_flags & B_VMIO)
1064 && !(bp->b_vp->v_tag == VT_NFS &&
1065 !vn_isdisk(bp->b_vp, NULL) &&
1066 (bp->b_flags & B_DELWRI))
1079 * Get the base offset and length of the buffer. Note that
1080 * in the VMIO case if the buffer block size is not
1081 * page-aligned then b_data pointer may not be page-aligned.
1082 * But our b_pages[] array *IS* page aligned.
1084 * block sizes less then DEV_BSIZE (usually 512) are not
1085 * supported due to the page granularity bits (m->valid,
1086 * m->dirty, etc...).
1088 * See man buf(9) for more information
1091 resid = bp->b_bufsize;
1092 foff = bp->b_offset;
1094 for (i = 0; i < bp->b_npages; i++) {
1096 vm_page_flag_clear(m, PG_ZERO);
1098 * If we hit a bogus page, fixup *all* of them
1101 if (m == bogus_page) {
1102 VOP_GETVOBJECT(vp, &obj);
1103 poff = OFF_TO_IDX(bp->b_offset);
1105 for (j = i; j < bp->b_npages; j++) {
1108 mtmp = bp->b_pages[j];
1109 if (mtmp == bogus_page) {
1110 mtmp = vm_page_lookup(obj, poff + j);
1112 panic("brelse: page missing\n");
1114 bp->b_pages[j] = mtmp;
1118 if ((bp->b_flags & B_INVAL) == 0) {
1119 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
1123 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1124 int poffset = foff & PAGE_MASK;
1125 int presid = resid > (PAGE_SIZE - poffset) ?
1126 (PAGE_SIZE - poffset) : resid;
1128 KASSERT(presid >= 0, ("brelse: extra page"));
1129 vm_page_set_invalid(m, poffset, presid);
1131 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1132 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1135 if (bp->b_flags & (B_INVAL | B_RELBUF))
1136 vfs_vmio_release(bp);
1138 } else if (bp->b_flags & B_VMIO) {
1140 if (bp->b_flags & (B_INVAL | B_RELBUF))
1141 vfs_vmio_release(bp);
1145 if (bp->b_qindex != QUEUE_NONE)
1146 panic("brelse: free buffer onto another queue???");
1147 if (BUF_REFCNT(bp) > 1) {
1148 /* Temporary panic to verify exclusive locking */
1149 /* This panic goes away when we allow shared refs */
1150 panic("brelse: multiple refs");
1151 /* do not release to free list */
1159 /* buffers with no memory */
1160 if (bp->b_bufsize == 0) {
1161 bp->b_flags |= B_INVAL;
1162 bp->b_xflags &= ~BX_BKGRDWRITE;
1163 if (bp->b_xflags & BX_BKGRDINPROG)
1164 panic("losing buffer 1");
1165 if (bp->b_kvasize) {
1166 bp->b_qindex = QUEUE_EMPTYKVA;
1168 bp->b_qindex = QUEUE_EMPTY;
1170 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1171 LIST_REMOVE(bp, b_hash);
1172 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1174 /* buffers with junk contents */
1175 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) {
1176 bp->b_flags |= B_INVAL;
1177 bp->b_xflags &= ~BX_BKGRDWRITE;
1178 if (bp->b_xflags & BX_BKGRDINPROG)
1179 panic("losing buffer 2");
1180 bp->b_qindex = QUEUE_CLEAN;
1181 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1182 LIST_REMOVE(bp, b_hash);
1183 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1186 /* buffers that are locked */
1187 } else if (bp->b_flags & B_LOCKED) {
1188 bp->b_qindex = QUEUE_LOCKED;
1189 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1191 /* remaining buffers */
1193 switch(bp->b_flags & (B_DELWRI|B_AGE)) {
1194 case B_DELWRI | B_AGE:
1195 bp->b_qindex = QUEUE_DIRTY;
1196 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1199 bp->b_qindex = QUEUE_DIRTY;
1200 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1203 bp->b_qindex = QUEUE_CLEAN;
1204 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1207 bp->b_qindex = QUEUE_CLEAN;
1208 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1214 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1215 * on the correct queue.
1217 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1221 * Fixup numfreebuffers count. The bp is on an appropriate queue
1222 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1223 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1224 * if B_INVAL is set ).
1227 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI))
1231 * Something we can maybe free or reuse
1233 if (bp->b_bufsize || bp->b_kvasize)
1238 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF |
1239 B_DIRECT | B_NOWDRAIN);
1244 * Release a buffer back to the appropriate queue but do not try to free
1245 * it. The buffer is expected to be used again soon.
1247 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1248 * biodone() to requeue an async I/O on completion. It is also used when
1249 * known good buffers need to be requeued but we think we may need the data
1252 * XXX we should be able to leave the B_RELBUF hint set on completion.
1255 bqrelse(struct buf * bp)
1261 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1263 if (bp->b_qindex != QUEUE_NONE)
1264 panic("bqrelse: free buffer onto another queue???");
1265 if (BUF_REFCNT(bp) > 1) {
1266 /* do not release to free list */
1267 panic("bqrelse: multiple refs");
1272 if (bp->b_flags & B_LOCKED) {
1273 bp->b_flags &= ~B_ERROR;
1274 bp->b_qindex = QUEUE_LOCKED;
1275 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1276 /* buffers with stale but valid contents */
1277 } else if (bp->b_flags & B_DELWRI) {
1278 bp->b_qindex = QUEUE_DIRTY;
1279 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1280 } else if (vm_page_count_severe()) {
1282 * We are too low on memory, we have to try to free the
1283 * buffer (most importantly: the wired pages making up its
1284 * backing store) *now*.
1290 bp->b_qindex = QUEUE_CLEAN;
1291 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1294 if ((bp->b_flags & B_LOCKED) == 0 &&
1295 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) {
1300 * Something we can maybe free or reuse.
1302 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1307 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1312 vfs_vmio_release(bp)
1319 for (i = 0; i < bp->b_npages; i++) {
1321 bp->b_pages[i] = NULL;
1323 * In order to keep page LRU ordering consistent, put
1324 * everything on the inactive queue.
1326 vm_page_unwire(m, 0);
1328 * We don't mess with busy pages, it is
1329 * the responsibility of the process that
1330 * busied the pages to deal with them.
1332 if ((m->flags & PG_BUSY) || (m->busy != 0))
1335 if (m->wire_count == 0) {
1336 vm_page_flag_clear(m, PG_ZERO);
1338 * Might as well free the page if we can and it has
1339 * no valid data. We also free the page if the
1340 * buffer was used for direct I/O.
1342 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) {
1344 vm_page_protect(m, VM_PROT_NONE);
1346 } else if (bp->b_flags & B_DIRECT) {
1347 vm_page_try_to_free(m);
1348 } else if (vm_page_count_severe()) {
1349 vm_page_try_to_cache(m);
1354 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1355 if (bp->b_bufsize) {
1360 bp->b_flags &= ~B_VMIO;
1366 * Check to see if a block is currently memory resident.
1369 gbincore(struct vnode * vp, daddr_t blkno)
1372 struct bufhashhdr *bh;
1374 bh = bufhash(vp, blkno);
1376 /* Search hash chain */
1377 LIST_FOREACH(bp, bh, b_hash) {
1379 if (bp->b_vp == vp && bp->b_lblkno == blkno &&
1380 (bp->b_flags & B_INVAL) == 0) {
1390 * Implement clustered async writes for clearing out B_DELWRI buffers.
1391 * This is much better then the old way of writing only one buffer at
1392 * a time. Note that we may not be presented with the buffers in the
1393 * correct order, so we search for the cluster in both directions.
1396 vfs_bio_awrite(struct buf * bp)
1400 daddr_t lblkno = bp->b_lblkno;
1401 struct vnode *vp = bp->b_vp;
1411 * right now we support clustered writing only to regular files. If
1412 * we find a clusterable block we could be in the middle of a cluster
1413 * rather then at the beginning.
1415 if ((vp->v_type == VREG) &&
1416 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1417 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1419 size = vp->v_mount->mnt_stat.f_iosize;
1420 maxcl = MAXPHYS / size;
1422 for (i = 1; i < maxcl; i++) {
1423 if ((bpa = gbincore(vp, lblkno + i)) &&
1424 BUF_REFCNT(bpa) == 0 &&
1425 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1426 (B_DELWRI | B_CLUSTEROK)) &&
1427 (bpa->b_bufsize == size)) {
1428 if ((bpa->b_blkno == bpa->b_lblkno) ||
1430 bp->b_blkno + ((i * size) >> DEV_BSHIFT)))
1436 for (j = 1; i + j <= maxcl && j <= lblkno; j++) {
1437 if ((bpa = gbincore(vp, lblkno - j)) &&
1438 BUF_REFCNT(bpa) == 0 &&
1439 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1440 (B_DELWRI | B_CLUSTEROK)) &&
1441 (bpa->b_bufsize == size)) {
1442 if ((bpa->b_blkno == bpa->b_lblkno) ||
1444 bp->b_blkno - ((j * size) >> DEV_BSHIFT)))
1453 * this is a possible cluster write
1456 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1462 BUF_LOCK(bp, LK_EXCLUSIVE);
1464 bp->b_flags |= B_ASYNC;
1468 * default (old) behavior, writing out only one block
1470 * XXX returns b_bufsize instead of b_bcount for nwritten?
1472 nwritten = bp->b_bufsize;
1473 (void) VOP_BWRITE(bp->b_vp, bp);
1481 * Find and initialize a new buffer header, freeing up existing buffers
1482 * in the bufqueues as necessary. The new buffer is returned locked.
1484 * Important: B_INVAL is not set. If the caller wishes to throw the
1485 * buffer away, the caller must set B_INVAL prior to calling brelse().
1488 * We have insufficient buffer headers
1489 * We have insufficient buffer space
1490 * buffer_map is too fragmented ( space reservation fails )
1491 * If we have to flush dirty buffers ( but we try to avoid this )
1493 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1494 * Instead we ask the buf daemon to do it for us. We attempt to
1495 * avoid piecemeal wakeups of the pageout daemon.
1499 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1505 static int flushingbufs;
1508 * We can't afford to block since we might be holding a vnode lock,
1509 * which may prevent system daemons from running. We deal with
1510 * low-memory situations by proactively returning memory and running
1511 * async I/O rather then sync I/O.
1515 --getnewbufrestarts;
1517 ++getnewbufrestarts;
1520 * Setup for scan. If we do not have enough free buffers,
1521 * we setup a degenerate case that immediately fails. Note
1522 * that if we are specially marked process, we are allowed to
1523 * dip into our reserves.
1525 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1527 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1528 * However, there are a number of cases (defragging, reusing, ...)
1529 * where we cannot backup.
1531 nqindex = QUEUE_EMPTYKVA;
1532 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1536 * If no EMPTYKVA buffers and we are either
1537 * defragging or reusing, locate a CLEAN buffer
1538 * to free or reuse. If bufspace useage is low
1539 * skip this step so we can allocate a new buffer.
1541 if (defrag || bufspace >= lobufspace) {
1542 nqindex = QUEUE_CLEAN;
1543 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1547 * If we could not find or were not allowed to reuse a
1548 * CLEAN buffer, check to see if it is ok to use an EMPTY
1549 * buffer. We can only use an EMPTY buffer if allocating
1550 * its KVA would not otherwise run us out of buffer space.
1552 if (nbp == NULL && defrag == 0 &&
1553 bufspace + maxsize < hibufspace) {
1554 nqindex = QUEUE_EMPTY;
1555 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1560 * Run scan, possibly freeing data and/or kva mappings on the fly
1564 while ((bp = nbp) != NULL) {
1565 int qindex = nqindex;
1568 * Calculate next bp ( we can only use it if we do not block
1569 * or do other fancy things ).
1571 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1574 nqindex = QUEUE_EMPTYKVA;
1575 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1578 case QUEUE_EMPTYKVA:
1579 nqindex = QUEUE_CLEAN;
1580 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1594 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1597 * Note: we no longer distinguish between VMIO and non-VMIO
1601 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1604 * If we are defragging then we need a buffer with
1605 * b_kvasize != 0. XXX this situation should no longer
1606 * occur, if defrag is non-zero the buffer's b_kvasize
1607 * should also be non-zero at this point. XXX
1609 if (defrag && bp->b_kvasize == 0) {
1610 printf("Warning: defrag empty buffer %p\n", bp);
1615 * Start freeing the bp. This is somewhat involved. nbp
1616 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1619 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1620 panic("getnewbuf: locked buf");
1623 if (qindex == QUEUE_CLEAN) {
1624 if (bp->b_flags & B_VMIO) {
1625 bp->b_flags &= ~B_ASYNC;
1626 vfs_vmio_release(bp);
1633 * NOTE: nbp is now entirely invalid. We can only restart
1634 * the scan from this point on.
1636 * Get the rest of the buffer freed up. b_kva* is still
1637 * valid after this operation.
1640 if (bp->b_rcred != NOCRED) {
1641 crfree(bp->b_rcred);
1642 bp->b_rcred = NOCRED;
1644 if (bp->b_wcred != NOCRED) {
1645 crfree(bp->b_wcred);
1646 bp->b_wcred = NOCRED;
1648 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
1649 (*bioops.io_deallocate)(bp);
1650 if (bp->b_xflags & BX_BKGRDINPROG)
1651 panic("losing buffer 3");
1652 LIST_REMOVE(bp, b_hash);
1653 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1662 bp->b_blkno = bp->b_lblkno = 0;
1663 bp->b_offset = NOOFFSET;
1669 bp->b_dirtyoff = bp->b_dirtyend = 0;
1671 LIST_INIT(&bp->b_dep);
1674 * If we are defragging then free the buffer.
1677 bp->b_flags |= B_INVAL;
1685 * If we are overcomitted then recover the buffer and its
1686 * KVM space. This occurs in rare situations when multiple
1687 * processes are blocked in getnewbuf() or allocbuf().
1689 if (bufspace >= hibufspace)
1691 if (flushingbufs && bp->b_kvasize != 0) {
1692 bp->b_flags |= B_INVAL;
1697 if (bufspace < lobufspace)
1703 * If we exhausted our list, sleep as appropriate. We may have to
1704 * wakeup various daemons and write out some dirty buffers.
1706 * Generally we are sleeping due to insufficient buffer space.
1714 flags = VFS_BIO_NEED_BUFSPACE;
1716 } else if (bufspace >= hibufspace) {
1718 flags = VFS_BIO_NEED_BUFSPACE;
1721 flags = VFS_BIO_NEED_ANY;
1724 bd_speedup(); /* heeeelp */
1726 needsbuffer |= flags;
1727 while (needsbuffer & flags) {
1728 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag,
1734 * We finally have a valid bp. We aren't quite out of the
1735 * woods, we still have to reserve kva space. In order
1736 * to keep fragmentation sane we only allocate kva in
1739 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1741 if (maxsize != bp->b_kvasize) {
1742 vm_offset_t addr = 0;
1746 vm_map_lock(buffer_map);
1748 if (vm_map_findspace(buffer_map,
1749 vm_map_min(buffer_map), maxsize, &addr)) {
1751 * Uh oh. Buffer map is to fragmented. We
1752 * must defragment the map.
1754 vm_map_unlock(buffer_map);
1757 bp->b_flags |= B_INVAL;
1762 vm_map_insert(buffer_map, NULL, 0,
1763 addr, addr + maxsize,
1764 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
1766 bp->b_kvabase = (caddr_t) addr;
1767 bp->b_kvasize = maxsize;
1768 bufspace += bp->b_kvasize;
1771 vm_map_unlock(buffer_map);
1773 bp->b_data = bp->b_kvabase;
1781 * buffer flushing daemon. Buffers are normally flushed by the
1782 * update daemon but if it cannot keep up this process starts to
1783 * take the load in an attempt to prevent getnewbuf() from blocking.
1786 static struct proc *bufdaemonproc;
1788 static struct kproc_desc buf_kp = {
1793 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
1801 * This process needs to be suspended prior to shutdown sync.
1803 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, bufdaemonproc,
1807 * This process is allowed to take the buffer cache to the limit
1812 kproc_suspend_loop(bufdaemonproc);
1815 * Do the flush. Limit the amount of in-transit I/O we
1816 * allow to build up, otherwise we would completely saturate
1817 * the I/O system. Wakeup any waiting processes before we
1818 * normally would so they can run in parallel with our drain.
1820 while (numdirtybuffers > lodirtybuffers) {
1821 if (flushbufqueues() == 0)
1823 waitrunningbufspace();
1824 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
1828 * Only clear bd_request if we have reached our low water
1829 * mark. The buf_daemon normally waits 5 seconds and
1830 * then incrementally flushes any dirty buffers that have
1831 * built up, within reason.
1833 * If we were unable to hit our low water mark and couldn't
1834 * find any flushable buffers, we sleep half a second.
1835 * Otherwise we loop immediately.
1837 if (numdirtybuffers <= lodirtybuffers) {
1839 * We reached our low water mark, reset the
1840 * request and sleep until we are needed again.
1841 * The sleep is just so the suspend code works.
1844 tsleep(&bd_request, PVM, "psleep", hz);
1847 * We couldn't find any flushable dirty buffers but
1848 * still have too many dirty buffers, we
1849 * have to sleep and try again. (rare)
1851 tsleep(&bd_request, PVM, "qsleep", hz / 2);
1859 * Try to flush a buffer in the dirty queue. We must be careful to
1860 * free up B_INVAL buffers instead of write them, which NFS is
1861 * particularly sensitive to.
1865 flushbufqueues(void)
1870 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
1873 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp));
1874 if ((bp->b_flags & B_DELWRI) != 0 &&
1875 (bp->b_xflags & BX_BKGRDINPROG) == 0) {
1876 if (bp->b_flags & B_INVAL) {
1877 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1878 panic("flushbufqueues: locked buf");
1884 if (LIST_FIRST(&bp->b_dep) != NULL &&
1885 bioops.io_countdeps &&
1886 (bp->b_flags & B_DEFERRED) == 0 &&
1887 (*bioops.io_countdeps)(bp, 0)) {
1888 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY],
1890 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY],
1892 bp->b_flags |= B_DEFERRED;
1893 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
1900 bp = TAILQ_NEXT(bp, b_freelist);
1906 * Check to see if a block is currently memory resident.
1909 incore(struct vnode * vp, daddr_t blkno)
1914 bp = gbincore(vp, blkno);
1920 * Returns true if no I/O is needed to access the
1921 * associated VM object. This is like incore except
1922 * it also hunts around in the VM system for the data.
1926 inmem(struct vnode * vp, daddr_t blkno)
1929 vm_offset_t toff, tinc, size;
1933 if (incore(vp, blkno))
1935 if (vp->v_mount == NULL)
1937 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0)
1941 if (size > vp->v_mount->mnt_stat.f_iosize)
1942 size = vp->v_mount->mnt_stat.f_iosize;
1943 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
1945 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
1946 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
1950 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
1951 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
1952 if (vm_page_is_valid(m,
1953 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
1962 * Sets the dirty range for a buffer based on the status of the dirty
1963 * bits in the pages comprising the buffer.
1965 * The range is limited to the size of the buffer.
1967 * This routine is primarily used by NFS, but is generalized for the
1971 vfs_setdirty(struct buf *bp)
1977 * Degenerate case - empty buffer
1980 if (bp->b_bufsize == 0)
1984 * We qualify the scan for modified pages on whether the
1985 * object has been flushed yet. The OBJ_WRITEABLE flag
1986 * is not cleared simply by protecting pages off.
1989 if ((bp->b_flags & B_VMIO) == 0)
1992 object = bp->b_pages[0]->object;
1994 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
1995 printf("Warning: object %p writeable but not mightbedirty\n", object);
1996 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
1997 printf("Warning: object %p mightbedirty but not writeable\n", object);
1999 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2000 vm_offset_t boffset;
2001 vm_offset_t eoffset;
2004 * test the pages to see if they have been modified directly
2005 * by users through the VM system.
2007 for (i = 0; i < bp->b_npages; i++) {
2008 vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
2009 vm_page_test_dirty(bp->b_pages[i]);
2013 * Calculate the encompassing dirty range, boffset and eoffset,
2014 * (eoffset - boffset) bytes.
2017 for (i = 0; i < bp->b_npages; i++) {
2018 if (bp->b_pages[i]->dirty)
2021 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2023 for (i = bp->b_npages - 1; i >= 0; --i) {
2024 if (bp->b_pages[i]->dirty) {
2028 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2031 * Fit it to the buffer.
2034 if (eoffset > bp->b_bcount)
2035 eoffset = bp->b_bcount;
2038 * If we have a good dirty range, merge with the existing
2042 if (boffset < eoffset) {
2043 if (bp->b_dirtyoff > boffset)
2044 bp->b_dirtyoff = boffset;
2045 if (bp->b_dirtyend < eoffset)
2046 bp->b_dirtyend = eoffset;
2054 * Get a block given a specified block and offset into a file/device.
2055 * The buffers B_DONE bit will be cleared on return, making it almost
2056 * ready for an I/O initiation. B_INVAL may or may not be set on
2057 * return. The caller should clear B_INVAL prior to initiating a
2060 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2061 * an existing buffer.
2063 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2064 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2065 * and then cleared based on the backing VM. If the previous buffer is
2066 * non-0-sized but invalid, B_CACHE will be cleared.
2068 * If getblk() must create a new buffer, the new buffer is returned with
2069 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2070 * case it is returned with B_INVAL clear and B_CACHE set based on the
2073 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos
2074 * B_CACHE bit is clear.
2076 * What this means, basically, is that the caller should use B_CACHE to
2077 * determine whether the buffer is fully valid or not and should clear
2078 * B_INVAL prior to issuing a read. If the caller intends to validate
2079 * the buffer by loading its data area with something, the caller needs
2080 * to clear B_INVAL. If the caller does this without issuing an I/O,
2081 * the caller should set B_CACHE ( as an optimization ), else the caller
2082 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2083 * a write attempt or if it was a successfull read. If the caller
2084 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2085 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2088 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo)
2092 struct bufhashhdr *bh;
2094 if (size > MAXBSIZE)
2095 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2100 * Block if we are low on buffers. Certain processes are allowed
2101 * to completely exhaust the buffer cache.
2103 * If this check ever becomes a bottleneck it may be better to
2104 * move it into the else, when gbincore() fails. At the moment
2105 * it isn't a problem.
2107 * XXX remove, we cannot afford to block anywhere if holding a vnode
2108 * lock in low-memory situation, so take it to the max.
2110 if (numfreebuffers == 0) {
2113 needsbuffer |= VFS_BIO_NEED_ANY;
2114 tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf",
2118 if ((bp = gbincore(vp, blkno))) {
2120 * Buffer is in-core. If the buffer is not busy, it must
2124 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2125 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL,
2126 "getblk", slpflag, slptimeo) == ENOLCK)
2129 return (struct buf *) NULL;
2133 * The buffer is locked. B_CACHE is cleared if the buffer is
2134 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set
2135 * and for a VMIO buffer B_CACHE is adjusted according to the
2138 if (bp->b_flags & B_INVAL)
2139 bp->b_flags &= ~B_CACHE;
2140 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2141 bp->b_flags |= B_CACHE;
2145 * check for size inconsistancies for non-VMIO case.
2148 if (bp->b_bcount != size) {
2149 if ((bp->b_flags & B_VMIO) == 0 ||
2150 (size > bp->b_kvasize)) {
2151 if (bp->b_flags & B_DELWRI) {
2152 bp->b_flags |= B_NOCACHE;
2153 VOP_BWRITE(bp->b_vp, bp);
2155 if ((bp->b_flags & B_VMIO) &&
2156 (LIST_FIRST(&bp->b_dep) == NULL)) {
2157 bp->b_flags |= B_RELBUF;
2160 bp->b_flags |= B_NOCACHE;
2161 VOP_BWRITE(bp->b_vp, bp);
2169 * If the size is inconsistant in the VMIO case, we can resize
2170 * the buffer. This might lead to B_CACHE getting set or
2171 * cleared. If the size has not changed, B_CACHE remains
2172 * unchanged from its previous state.
2175 if (bp->b_bcount != size)
2178 KASSERT(bp->b_offset != NOOFFSET,
2179 ("getblk: no buffer offset"));
2182 * A buffer with B_DELWRI set and B_CACHE clear must
2183 * be committed before we can return the buffer in
2184 * order to prevent the caller from issuing a read
2185 * ( due to B_CACHE not being set ) and overwriting
2188 * Most callers, including NFS and FFS, need this to
2189 * operate properly either because they assume they
2190 * can issue a read if B_CACHE is not set, or because
2191 * ( for example ) an uncached B_DELWRI might loop due
2192 * to softupdates re-dirtying the buffer. In the latter
2193 * case, B_CACHE is set after the first write completes,
2194 * preventing further loops.
2196 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2197 * above while extending the buffer, we cannot allow the
2198 * buffer to remain with B_CACHE set after the write
2199 * completes or it will represent a corrupt state. To
2200 * deal with this we set B_NOCACHE to scrap the buffer
2203 * We might be able to do something fancy, like setting
2204 * B_CACHE in bwrite() except if B_DELWRI is already set,
2205 * so the below call doesn't set B_CACHE, but that gets real
2206 * confusing. This is much easier.
2209 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2210 bp->b_flags |= B_NOCACHE;
2211 VOP_BWRITE(bp->b_vp, bp);
2216 bp->b_flags &= ~B_DONE;
2219 * Buffer is not in-core, create new buffer. The buffer
2220 * returned by getnewbuf() is locked. Note that the returned
2221 * buffer is also considered valid (not marked B_INVAL).
2223 int bsize, maxsize, vmio;
2226 if (vn_isdisk(vp, NULL))
2228 else if (vp->v_mountedhere)
2229 bsize = vp->v_mountedhere->mnt_stat.f_iosize;
2230 else if (vp->v_mount)
2231 bsize = vp->v_mount->mnt_stat.f_iosize;
2235 offset = (off_t)blkno * bsize;
2236 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF);
2237 maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2238 maxsize = imax(maxsize, bsize);
2240 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) {
2241 if (slpflag || slptimeo) {
2249 * This code is used to make sure that a buffer is not
2250 * created while the getnewbuf routine is blocked.
2251 * This can be a problem whether the vnode is locked or not.
2252 * If the buffer is created out from under us, we have to
2253 * throw away the one we just created. There is now window
2254 * race because we are safely running at splbio() from the
2255 * point of the duplicate buffer creation through to here,
2256 * and we've locked the buffer.
2258 if (gbincore(vp, blkno)) {
2259 bp->b_flags |= B_INVAL;
2265 * Insert the buffer into the hash, so that it can
2266 * be found by incore.
2268 bp->b_blkno = bp->b_lblkno = blkno;
2269 bp->b_offset = offset;
2272 LIST_REMOVE(bp, b_hash);
2273 bh = bufhash(vp, blkno);
2274 LIST_INSERT_HEAD(bh, bp, b_hash);
2277 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
2278 * buffer size starts out as 0, B_CACHE will be set by
2279 * allocbuf() for the VMIO case prior to it testing the
2280 * backing store for validity.
2284 bp->b_flags |= B_VMIO;
2285 #if defined(VFS_BIO_DEBUG)
2286 if (vp->v_type != VREG && vp->v_type != VBLK)
2287 printf("getblk: vmioing file type %d???\n", vp->v_type);
2290 bp->b_flags &= ~B_VMIO;
2296 bp->b_flags &= ~B_DONE;
2302 * Get an empty, disassociated buffer of given size. The buffer is initially
2312 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2315 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0);
2318 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2324 * This code constitutes the buffer memory from either anonymous system
2325 * memory (in the case of non-VMIO operations) or from an associated
2326 * VM object (in the case of VMIO operations). This code is able to
2327 * resize a buffer up or down.
2329 * Note that this code is tricky, and has many complications to resolve
2330 * deadlock or inconsistant data situations. Tread lightly!!!
2331 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2332 * the caller. Calling this code willy nilly can result in the loss of data.
2334 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2335 * B_CACHE for the non-VMIO case.
2339 allocbuf(struct buf *bp, int size)
2341 int newbsize, mbsize;
2344 if (BUF_REFCNT(bp) == 0)
2345 panic("allocbuf: buffer not busy");
2347 if (bp->b_kvasize < size)
2348 panic("allocbuf: buffer too small");
2350 if ((bp->b_flags & B_VMIO) == 0) {
2354 * Just get anonymous memory from the kernel. Don't
2355 * mess with B_CACHE.
2357 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2358 #if !defined(NO_B_MALLOC)
2359 if (bp->b_flags & B_MALLOC)
2363 newbsize = round_page(size);
2365 if (newbsize < bp->b_bufsize) {
2366 #if !defined(NO_B_MALLOC)
2368 * malloced buffers are not shrunk
2370 if (bp->b_flags & B_MALLOC) {
2372 bp->b_bcount = size;
2374 free(bp->b_data, M_BIOBUF);
2375 if (bp->b_bufsize) {
2376 bufmallocspace -= bp->b_bufsize;
2380 bp->b_data = bp->b_kvabase;
2382 bp->b_flags &= ~B_MALLOC;
2389 (vm_offset_t) bp->b_data + newbsize,
2390 (vm_offset_t) bp->b_data + bp->b_bufsize);
2391 } else if (newbsize > bp->b_bufsize) {
2392 #if !defined(NO_B_MALLOC)
2394 * We only use malloced memory on the first allocation.
2395 * and revert to page-allocated memory when the buffer
2398 if ( (bufmallocspace < maxbufmallocspace) &&
2399 (bp->b_bufsize == 0) &&
2400 (mbsize <= PAGE_SIZE/2)) {
2402 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2403 bp->b_bufsize = mbsize;
2404 bp->b_bcount = size;
2405 bp->b_flags |= B_MALLOC;
2406 bufmallocspace += mbsize;
2412 #if !defined(NO_B_MALLOC)
2414 * If the buffer is growing on its other-than-first allocation,
2415 * then we revert to the page-allocation scheme.
2417 if (bp->b_flags & B_MALLOC) {
2418 origbuf = bp->b_data;
2419 origbufsize = bp->b_bufsize;
2420 bp->b_data = bp->b_kvabase;
2421 if (bp->b_bufsize) {
2422 bufmallocspace -= bp->b_bufsize;
2426 bp->b_flags &= ~B_MALLOC;
2427 newbsize = round_page(newbsize);
2432 (vm_offset_t) bp->b_data + bp->b_bufsize,
2433 (vm_offset_t) bp->b_data + newbsize);
2434 #if !defined(NO_B_MALLOC)
2436 bcopy(origbuf, bp->b_data, origbufsize);
2437 free(origbuf, M_BIOBUF);
2445 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2446 desiredpages = (size == 0) ? 0 :
2447 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2449 #if !defined(NO_B_MALLOC)
2450 if (bp->b_flags & B_MALLOC)
2451 panic("allocbuf: VMIO buffer can't be malloced");
2454 * Set B_CACHE initially if buffer is 0 length or will become
2457 if (size == 0 || bp->b_bufsize == 0)
2458 bp->b_flags |= B_CACHE;
2460 if (newbsize < bp->b_bufsize) {
2462 * DEV_BSIZE aligned new buffer size is less then the
2463 * DEV_BSIZE aligned existing buffer size. Figure out
2464 * if we have to remove any pages.
2466 if (desiredpages < bp->b_npages) {
2467 for (i = desiredpages; i < bp->b_npages; i++) {
2469 * the page is not freed here -- it
2470 * is the responsibility of
2471 * vnode_pager_setsize
2474 KASSERT(m != bogus_page,
2475 ("allocbuf: bogus page found"));
2476 while (vm_page_sleep_busy(m, TRUE, "biodep"))
2479 bp->b_pages[i] = NULL;
2480 vm_page_unwire(m, 0);
2482 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2483 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2484 bp->b_npages = desiredpages;
2486 } else if (size > bp->b_bcount) {
2488 * We are growing the buffer, possibly in a
2489 * byte-granular fashion.
2497 * Step 1, bring in the VM pages from the object,
2498 * allocating them if necessary. We must clear
2499 * B_CACHE if these pages are not valid for the
2500 * range covered by the buffer.
2504 VOP_GETVOBJECT(vp, &obj);
2506 while (bp->b_npages < desiredpages) {
2510 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2511 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2513 * note: must allocate system pages
2514 * since blocking here could intefere
2515 * with paging I/O, no matter which
2518 m = vm_page_alloc(obj, pi, VM_ALLOC_SYSTEM);
2521 vm_pageout_deficit += desiredpages - bp->b_npages;
2525 bp->b_flags &= ~B_CACHE;
2526 bp->b_pages[bp->b_npages] = m;
2533 * We found a page. If we have to sleep on it,
2534 * retry because it might have gotten freed out
2537 * We can only test PG_BUSY here. Blocking on
2538 * m->busy might lead to a deadlock:
2540 * vm_fault->getpages->cluster_read->allocbuf
2544 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
2548 * We have a good page. Should we wakeup the
2551 if ((curproc != pageproc) &&
2552 ((m->queue - m->pc) == PQ_CACHE) &&
2553 ((cnt.v_free_count + cnt.v_cache_count) <
2554 (cnt.v_free_min + cnt.v_cache_min))) {
2555 pagedaemon_wakeup();
2557 vm_page_flag_clear(m, PG_ZERO);
2559 bp->b_pages[bp->b_npages] = m;
2564 * Step 2. We've loaded the pages into the buffer,
2565 * we have to figure out if we can still have B_CACHE
2566 * set. Note that B_CACHE is set according to the
2567 * byte-granular range ( bcount and size ), new the
2568 * aligned range ( newbsize ).
2570 * The VM test is against m->valid, which is DEV_BSIZE
2571 * aligned. Needless to say, the validity of the data
2572 * needs to also be DEV_BSIZE aligned. Note that this
2573 * fails with NFS if the server or some other client
2574 * extends the file's EOF. If our buffer is resized,
2575 * B_CACHE may remain set! XXX
2578 toff = bp->b_bcount;
2579 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2581 while ((bp->b_flags & B_CACHE) && toff < size) {
2584 if (tinc > (size - toff))
2587 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2602 * Step 3, fixup the KVM pmap. Remember that
2603 * bp->b_data is relative to bp->b_offset, but
2604 * bp->b_offset may be offset into the first page.
2607 bp->b_data = (caddr_t)
2608 trunc_page((vm_offset_t)bp->b_data);
2610 (vm_offset_t)bp->b_data,
2614 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2615 (vm_offset_t)(bp->b_offset & PAGE_MASK));
2618 if (newbsize < bp->b_bufsize)
2620 bp->b_bufsize = newbsize; /* actual buffer allocation */
2621 bp->b_bcount = size; /* requested buffer size */
2628 * Wait for buffer I/O completion, returning error status. The buffer
2629 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR
2630 * error and cleared.
2633 biowait(register struct buf * bp)
2638 while ((bp->b_flags & B_DONE) == 0) {
2639 #if defined(NO_SCHEDULE_MODS)
2640 tsleep(bp, PRIBIO, "biowait", 0);
2642 if (bp->b_flags & B_READ)
2643 tsleep(bp, PRIBIO, "biord", 0);
2645 tsleep(bp, PRIBIO, "biowr", 0);
2649 if (bp->b_flags & B_EINTR) {
2650 bp->b_flags &= ~B_EINTR;
2653 if (bp->b_flags & B_ERROR) {
2654 return (bp->b_error ? bp->b_error : EIO);
2663 * Finish I/O on a buffer, optionally calling a completion function.
2664 * This is usually called from an interrupt so process blocking is
2667 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
2668 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
2669 * assuming B_INVAL is clear.
2671 * For the VMIO case, we set B_CACHE if the op was a read and no
2672 * read error occured, or if the op was a write. B_CACHE is never
2673 * set if the buffer is invalid or otherwise uncacheable.
2675 * biodone does not mess with B_INVAL, allowing the I/O routine or the
2676 * initiator to leave B_INVAL set to brelse the buffer out of existance
2677 * in the biodone routine.
2680 biodone(register struct buf * bp)
2686 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp)));
2687 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
2689 bp->b_flags |= B_DONE;
2690 runningbufwakeup(bp);
2692 if (bp->b_flags & B_FREEBUF) {
2698 if ((bp->b_flags & B_READ) == 0) {
2702 /* call optional completion function if requested */
2703 if (bp->b_flags & B_CALL) {
2704 bp->b_flags &= ~B_CALL;
2705 (*bp->b_iodone) (bp);
2709 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
2710 (*bioops.io_complete)(bp);
2712 if (bp->b_flags & B_VMIO) {
2718 struct vnode *vp = bp->b_vp;
2720 error = VOP_GETVOBJECT(vp, &obj);
2722 #if defined(VFS_BIO_DEBUG)
2723 if (vp->v_usecount == 0) {
2724 panic("biodone: zero vnode ref count");
2728 panic("biodone: missing VM object");
2731 if ((vp->v_flag & VOBJBUF) == 0) {
2732 panic("biodone: vnode is not setup for merged cache");
2736 foff = bp->b_offset;
2737 KASSERT(bp->b_offset != NOOFFSET,
2738 ("biodone: no buffer offset"));
2741 panic("biodone: no object");
2743 #if defined(VFS_BIO_DEBUG)
2744 if (obj->paging_in_progress < bp->b_npages) {
2745 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
2746 obj->paging_in_progress, bp->b_npages);
2751 * Set B_CACHE if the op was a normal read and no error
2752 * occured. B_CACHE is set for writes in the b*write()
2755 iosize = bp->b_bcount - bp->b_resid;
2756 if ((bp->b_flags & (B_READ|B_FREEBUF|B_INVAL|B_NOCACHE|B_ERROR)) == B_READ) {
2757 bp->b_flags |= B_CACHE;
2760 for (i = 0; i < bp->b_npages; i++) {
2764 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2769 * cleanup bogus pages, restoring the originals
2772 if (m == bogus_page) {
2774 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2776 panic("biodone: page disappeared");
2778 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2780 #if defined(VFS_BIO_DEBUG)
2781 if (OFF_TO_IDX(foff) != m->pindex) {
2783 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
2784 (unsigned long)foff, m->pindex);
2789 * In the write case, the valid and clean bits are
2790 * already changed correctly ( see bdwrite() ), so we
2791 * only need to do this here in the read case.
2793 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) {
2794 vfs_page_set_valid(bp, foff, i, m);
2796 vm_page_flag_clear(m, PG_ZERO);
2799 * when debugging new filesystems or buffer I/O methods, this
2800 * is the most common error that pops up. if you see this, you
2801 * have not set the page busy flag correctly!!!
2804 printf("biodone: page busy < 0, "
2805 "pindex: %d, foff: 0x(%x,%x), "
2806 "resid: %d, index: %d\n",
2807 (int) m->pindex, (int)(foff >> 32),
2808 (int) foff & 0xffffffff, resid, i);
2809 if (!vn_isdisk(vp, NULL))
2810 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n",
2811 bp->b_vp->v_mount->mnt_stat.f_iosize,
2813 bp->b_flags, bp->b_npages);
2815 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n",
2817 bp->b_flags, bp->b_npages);
2818 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
2819 m->valid, m->dirty, m->wire_count);
2820 panic("biodone: page busy < 0\n");
2822 vm_page_io_finish(m);
2823 vm_object_pip_subtract(obj, 1);
2824 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2828 vm_object_pip_wakeupn(obj, 0);
2832 * For asynchronous completions, release the buffer now. The brelse
2833 * will do a wakeup there if necessary - so no need to do a wakeup
2834 * here in the async case. The sync case always needs to do a wakeup.
2837 if (bp->b_flags & B_ASYNC) {
2838 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
2849 * This routine is called in lieu of iodone in the case of
2850 * incomplete I/O. This keeps the busy status for pages
2854 vfs_unbusy_pages(struct buf * bp)
2858 runningbufwakeup(bp);
2859 if (bp->b_flags & B_VMIO) {
2860 struct vnode *vp = bp->b_vp;
2863 VOP_GETVOBJECT(vp, &obj);
2865 for (i = 0; i < bp->b_npages; i++) {
2866 vm_page_t m = bp->b_pages[i];
2868 if (m == bogus_page) {
2869 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
2871 panic("vfs_unbusy_pages: page missing\n");
2874 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2876 vm_object_pip_subtract(obj, 1);
2877 vm_page_flag_clear(m, PG_ZERO);
2878 vm_page_io_finish(m);
2880 vm_object_pip_wakeupn(obj, 0);
2885 * vfs_page_set_valid:
2887 * Set the valid bits in a page based on the supplied offset. The
2888 * range is restricted to the buffer's size.
2890 * This routine is typically called after a read completes.
2893 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
2895 vm_ooffset_t soff, eoff;
2898 * Start and end offsets in buffer. eoff - soff may not cross a
2899 * page boundry or cross the end of the buffer. The end of the
2900 * buffer, in this case, is our file EOF, not the allocation size
2904 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2905 if (eoff > bp->b_offset + bp->b_bcount)
2906 eoff = bp->b_offset + bp->b_bcount;
2909 * Set valid range. This is typically the entire buffer and thus the
2913 vm_page_set_validclean(
2915 (vm_offset_t) (soff & PAGE_MASK),
2916 (vm_offset_t) (eoff - soff)
2922 * This routine is called before a device strategy routine.
2923 * It is used to tell the VM system that paging I/O is in
2924 * progress, and treat the pages associated with the buffer
2925 * almost as being PG_BUSY. Also the object paging_in_progress
2926 * flag is handled to make sure that the object doesn't become
2929 * Since I/O has not been initiated yet, certain buffer flags
2930 * such as B_ERROR or B_INVAL may be in an inconsistant state
2931 * and should be ignored.
2934 vfs_busy_pages(struct buf * bp, int clear_modify)
2938 if (bp->b_flags & B_VMIO) {
2939 struct vnode *vp = bp->b_vp;
2943 VOP_GETVOBJECT(vp, &obj);
2944 foff = bp->b_offset;
2945 KASSERT(bp->b_offset != NOOFFSET,
2946 ("vfs_busy_pages: no buffer offset"));
2950 for (i = 0; i < bp->b_npages; i++) {
2951 vm_page_t m = bp->b_pages[i];
2952 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
2957 for (i = 0; i < bp->b_npages; i++) {
2958 vm_page_t m = bp->b_pages[i];
2960 vm_page_flag_clear(m, PG_ZERO);
2961 if ((bp->b_flags & B_CLUSTER) == 0) {
2962 vm_object_pip_add(obj, 1);
2963 vm_page_io_start(m);
2967 * When readying a buffer for a read ( i.e
2968 * clear_modify == 0 ), it is important to do
2969 * bogus_page replacement for valid pages in
2970 * partially instantiated buffers. Partially
2971 * instantiated buffers can, in turn, occur when
2972 * reconstituting a buffer from its VM backing store
2973 * base. We only have to do this if B_CACHE is
2974 * clear ( which causes the I/O to occur in the
2975 * first place ). The replacement prevents the read
2976 * I/O from overwriting potentially dirty VM-backed
2977 * pages. XXX bogus page replacement is, uh, bogus.
2978 * It may not work properly with small-block devices.
2979 * We need to find a better way.
2982 vm_page_protect(m, VM_PROT_NONE);
2984 vfs_page_set_valid(bp, foff, i, m);
2985 else if (m->valid == VM_PAGE_BITS_ALL &&
2986 (bp->b_flags & B_CACHE) == 0) {
2987 bp->b_pages[i] = bogus_page;
2990 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2993 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2998 * Tell the VM system that the pages associated with this buffer
2999 * are clean. This is used for delayed writes where the data is
3000 * going to go to disk eventually without additional VM intevention.
3002 * Note that while we only really need to clean through to b_bcount, we
3003 * just go ahead and clean through to b_bufsize.
3006 vfs_clean_pages(struct buf * bp)
3010 if (bp->b_flags & B_VMIO) {
3013 foff = bp->b_offset;
3014 KASSERT(bp->b_offset != NOOFFSET,
3015 ("vfs_clean_pages: no buffer offset"));
3016 for (i = 0; i < bp->b_npages; i++) {
3017 vm_page_t m = bp->b_pages[i];
3018 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3019 vm_ooffset_t eoff = noff;
3021 if (eoff > bp->b_offset + bp->b_bufsize)
3022 eoff = bp->b_offset + bp->b_bufsize;
3023 vfs_page_set_valid(bp, foff, i, m);
3024 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3031 * vfs_bio_set_validclean:
3033 * Set the range within the buffer to valid and clean. The range is
3034 * relative to the beginning of the buffer, b_offset. Note that b_offset
3035 * itself may be offset from the beginning of the first page.
3039 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3041 if (bp->b_flags & B_VMIO) {
3046 * Fixup base to be relative to beginning of first page.
3047 * Set initial n to be the maximum number of bytes in the
3048 * first page that can be validated.
3051 base += (bp->b_offset & PAGE_MASK);
3052 n = PAGE_SIZE - (base & PAGE_MASK);
3054 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3055 vm_page_t m = bp->b_pages[i];
3060 vm_page_set_validclean(m, base & PAGE_MASK, n);
3071 * clear a buffer. This routine essentially fakes an I/O, so we need
3072 * to clear B_ERROR and B_INVAL.
3074 * Note that while we only theoretically need to clear through b_bcount,
3075 * we go ahead and clear through b_bufsize.
3079 vfs_bio_clrbuf(struct buf *bp)
3083 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3084 bp->b_flags &= ~(B_INVAL|B_ERROR);
3085 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3086 (bp->b_offset & PAGE_MASK) == 0) {
3087 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3088 if ((bp->b_pages[0]->valid & mask) == mask) {
3092 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3093 ((bp->b_pages[0]->valid & mask) == 0)) {
3094 bzero(bp->b_data, bp->b_bufsize);
3095 bp->b_pages[0]->valid |= mask;
3100 ea = sa = bp->b_data;
3101 for(i=0;i<bp->b_npages;i++,sa=ea) {
3102 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3103 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3104 ea = (caddr_t)(vm_offset_t)ulmin(
3105 (u_long)(vm_offset_t)ea,
3106 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3107 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3108 if ((bp->b_pages[i]->valid & mask) == mask)
3110 if ((bp->b_pages[i]->valid & mask) == 0) {
3111 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) {
3115 for (; sa < ea; sa += DEV_BSIZE, j++) {
3116 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3117 (bp->b_pages[i]->valid & (1<<j)) == 0)
3118 bzero(sa, DEV_BSIZE);
3121 bp->b_pages[i]->valid |= mask;
3122 vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
3131 * vm_hold_load_pages and vm_hold_unload pages get pages into
3132 * a buffers address space. The pages are anonymous and are
3133 * not associated with a file object.
3136 vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3142 to = round_page(to);
3143 from = round_page(from);
3144 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3146 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3151 * note: must allocate system pages since blocking here
3152 * could intefere with paging I/O, no matter which
3155 p = vm_page_alloc(kernel_object,
3156 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3159 vm_pageout_deficit += (to - from) >> PAGE_SHIFT;
3164 p->valid = VM_PAGE_BITS_ALL;
3165 vm_page_flag_clear(p, PG_ZERO);
3166 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3167 bp->b_pages[index] = p;
3170 bp->b_npages = index;
3174 vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3178 int index, newnpages;
3180 from = round_page(from);
3181 to = round_page(to);
3182 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3184 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3185 p = bp->b_pages[index];
3186 if (p && (index < bp->b_npages)) {
3188 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n",
3189 bp->b_blkno, bp->b_lblkno);
3191 bp->b_pages[index] = NULL;
3194 vm_page_unwire(p, 0);
3198 bp->b_npages = newnpages;
3202 * Map an IO request into kernel virtual address space.
3204 * All requests are (re)mapped into kernel VA space.
3205 * Notice that we use b_bufsize for the size of the buffer
3206 * to be mapped. b_bcount might be modified by the driver.
3209 vmapbuf(struct buf *bp)
3211 caddr_t addr, v, kva;
3217 if ((bp->b_flags & B_PHYS) == 0)
3219 if (bp->b_bufsize < 0)
3221 for (v = bp->b_saveaddr,
3222 addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data),
3224 addr < bp->b_data + bp->b_bufsize;
3225 addr += PAGE_SIZE, v += PAGE_SIZE, pidx++) {
3227 * Do the vm_fault if needed; do the copy-on-write thing
3228 * when reading stuff off device into memory.
3231 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data,
3232 (bp->b_flags&B_READ)?(VM_PROT_READ|VM_PROT_WRITE):VM_PROT_READ);
3234 for (i = 0; i < pidx; ++i) {
3235 vm_page_unhold(bp->b_pages[i]);
3236 bp->b_pages[i] = NULL;
3242 * WARNING! If sparc support is MFCd in the future this will
3243 * have to be changed from pmap_kextract() to pmap_extract()
3247 #error "If MFCing sparc support use pmap_extract"
3249 pa = pmap_kextract((vm_offset_t)addr);
3251 printf("vmapbuf: warning, race against user address during I/O");
3254 m = PHYS_TO_VM_PAGE(pa);
3256 bp->b_pages[pidx] = m;
3258 if (pidx > btoc(MAXPHYS))
3259 panic("vmapbuf: mapped more than MAXPHYS");
3260 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3262 kva = bp->b_saveaddr;
3263 bp->b_npages = pidx;
3264 bp->b_saveaddr = bp->b_data;
3265 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3270 * Free the io map PTEs associated with this IO operation.
3271 * We also invalidate the TLB entries and restore the original b_addr.
3275 register struct buf *bp;
3281 if ((bp->b_flags & B_PHYS) == 0)
3284 npages = bp->b_npages;
3285 pmap_qremove(trunc_page((vm_offset_t)bp->b_data),
3288 for (pidx = 0; pidx < npages; pidx++)
3289 vm_page_unhold(*m++);
3291 bp->b_data = bp->b_saveaddr;
3294 #include "opt_ddb.h"
3296 #include <ddb/ddb.h>
3298 DB_SHOW_COMMAND(buffer, db_show_buffer)
3301 struct buf *bp = (struct buf *)addr;
3304 db_printf("usage: show buffer <addr>\n");
3308 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3309 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, "
3310 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, "
3311 "b_blkno = %d, b_pblkno = %d\n",
3312 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3313 major(bp->b_dev), minor(bp->b_dev),
3314 bp->b_data, bp->b_blkno, bp->b_pblkno);
3317 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3318 for (i = 0; i < bp->b_npages; i++) {
3321 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3322 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3323 if ((i + 1) < bp->b_npages)