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 1.3 2003/06/19 01:55:06 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 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
60 struct bio_ops bioops; /* I/O operation notification */
62 struct buf *buf; /* buffer header pool */
63 struct swqueue bswlist;
65 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from,
67 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from,
69 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
70 int pageno, vm_page_t m);
71 static void vfs_clean_pages(struct buf * bp);
72 static void vfs_setdirty(struct buf *bp);
73 static void vfs_vmio_release(struct buf *bp);
74 static void vfs_backgroundwritedone(struct buf *bp);
75 static int flushbufqueues(void);
77 static int bd_request;
79 static void buf_daemon __P((void));
81 * bogus page -- for I/O to/from partially complete buffers
82 * this is a temporary solution to the problem, but it is not
83 * really that bad. it would be better to split the buffer
84 * for input in the case of buffers partially already in memory,
85 * but the code is intricate enough already.
88 int vmiodirenable = TRUE;
90 static vm_offset_t bogus_offset;
92 static int bufspace, maxbufspace,
93 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
94 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
95 static int needsbuffer;
96 static int lorunningspace, hirunningspace, runningbufreq;
97 static int numdirtybuffers, lodirtybuffers, hidirtybuffers;
98 static int numfreebuffers, lofreebuffers, hifreebuffers;
99 static int getnewbufcalls;
100 static int getnewbufrestarts;
102 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD,
103 &numdirtybuffers, 0, "");
104 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW,
105 &lodirtybuffers, 0, "");
106 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW,
107 &hidirtybuffers, 0, "");
108 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD,
109 &numfreebuffers, 0, "");
110 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW,
111 &lofreebuffers, 0, "");
112 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW,
113 &hifreebuffers, 0, "");
114 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD,
115 &runningbufspace, 0, "");
116 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW,
117 &lorunningspace, 0, "");
118 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW,
119 &hirunningspace, 0, "");
120 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD,
121 &maxbufspace, 0, "");
122 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD,
124 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD,
126 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD,
128 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW,
129 &maxbufmallocspace, 0, "");
130 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD,
131 &bufmallocspace, 0, "");
132 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW,
133 &getnewbufcalls, 0, "");
134 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW,
135 &getnewbufrestarts, 0, "");
136 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW,
137 &vmiodirenable, 0, "");
138 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW,
139 &bufdefragcnt, 0, "");
140 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW,
141 &buffreekvacnt, 0, "");
142 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW,
143 &bufreusecnt, 0, "");
145 static int bufhashmask;
146 static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
147 struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } };
148 char *buf_wmesg = BUF_WMESG;
150 extern int vm_swap_size;
152 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
153 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
154 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
155 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
158 * Buffer hash table code. Note that the logical block scans linearly, which
159 * gives us some L1 cache locality.
164 bufhash(struct vnode *vnp, daddr_t bn)
166 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]);
172 * If someone is blocked due to there being too many dirty buffers,
173 * and numdirtybuffers is now reasonable, wake them up.
177 numdirtywakeup(int level)
179 if (numdirtybuffers <= level) {
180 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
181 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
182 wakeup(&needsbuffer);
190 * Called when buffer space is potentially available for recovery.
191 * getnewbuf() will block on this flag when it is unable to free
192 * sufficient buffer space. Buffer space becomes recoverable when
193 * bp's get placed back in the queues.
200 * If someone is waiting for BUF space, wake them up. Even
201 * though we haven't freed the kva space yet, the waiting
202 * process will be able to now.
204 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
205 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
206 wakeup(&needsbuffer);
211 * runningbufwakeup() - in-progress I/O accounting.
215 runningbufwakeup(struct buf *bp)
217 if (bp->b_runningbufspace) {
218 runningbufspace -= bp->b_runningbufspace;
219 bp->b_runningbufspace = 0;
220 if (runningbufreq && runningbufspace <= lorunningspace) {
222 wakeup(&runningbufreq);
230 * Called when a buffer has been added to one of the free queues to
231 * account for the buffer and to wakeup anyone waiting for free buffers.
232 * This typically occurs when large amounts of metadata are being handled
233 * by the buffer cache ( else buffer space runs out first, usually ).
241 needsbuffer &= ~VFS_BIO_NEED_ANY;
242 if (numfreebuffers >= hifreebuffers)
243 needsbuffer &= ~VFS_BIO_NEED_FREE;
244 wakeup(&needsbuffer);
249 * waitrunningbufspace()
251 * runningbufspace is a measure of the amount of I/O currently
252 * running. This routine is used in async-write situations to
253 * prevent creating huge backups of pending writes to a device.
254 * Only asynchronous writes are governed by this function.
256 * Reads will adjust runningbufspace, but will not block based on it.
257 * The read load has a side effect of reducing the allowed write load.
259 * This does NOT turn an async write into a sync write. It waits
260 * for earlier writes to complete and generally returns before the
261 * caller's write has reached the device.
264 waitrunningbufspace(void)
266 while (runningbufspace > hirunningspace) {
269 s = splbio(); /* fix race against interrupt/biodone() */
271 tsleep(&runningbufreq, PVM, "wdrain", 0);
277 * vfs_buf_test_cache:
279 * Called when a buffer is extended. This function clears the B_CACHE
280 * bit if the newly extended portion of the buffer does not contain
285 vfs_buf_test_cache(struct buf *bp,
286 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
289 if (bp->b_flags & B_CACHE) {
290 int base = (foff + off) & PAGE_MASK;
291 if (vm_page_is_valid(m, base, size) == 0)
292 bp->b_flags &= ~B_CACHE;
298 bd_wakeup(int dirtybuflevel)
300 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
307 * bd_speedup - speedup the buffer cache flushing code
318 * Initialize buffer headers and related structures.
322 bufhashinit(caddr_t vaddr)
324 /* first, make a null hash table */
325 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1)
327 bufhashtbl = (void *)vaddr;
328 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask;
339 TAILQ_INIT(&bswlist);
340 LIST_INIT(&invalhash);
341 simple_lock_init(&buftimelock);
343 for (i = 0; i <= bufhashmask; i++)
344 LIST_INIT(&bufhashtbl[i]);
346 /* next, make a null set of free lists */
347 for (i = 0; i < BUFFER_QUEUES; i++)
348 TAILQ_INIT(&bufqueues[i]);
350 /* finally, initialize each buffer header and stick on empty q */
351 for (i = 0; i < nbuf; i++) {
353 bzero(bp, sizeof *bp);
354 bp->b_flags = B_INVAL; /* we're just an empty header */
356 bp->b_rcred = NOCRED;
357 bp->b_wcred = NOCRED;
358 bp->b_qindex = QUEUE_EMPTY;
360 LIST_INIT(&bp->b_dep);
362 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
363 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
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),
436 * bfreekva() - free the kva allocation for a buffer.
438 * Must be called at splbio() or higher as this is the only locking for
441 * Since this call frees up buffer space, we call bufspacewakeup().
444 bfreekva(struct buf * bp)
448 vm_map_lock(buffer_map);
449 bufspace -= bp->b_kvasize;
450 vm_map_delete(buffer_map,
451 (vm_offset_t) bp->b_kvabase,
452 (vm_offset_t) bp->b_kvabase + bp->b_kvasize
454 vm_map_unlock(buffer_map);
463 * Remove the buffer from the appropriate free list.
466 bremfree(struct buf * bp)
469 int old_qindex = bp->b_qindex;
471 if (bp->b_qindex != QUEUE_NONE) {
472 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp));
473 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
474 bp->b_qindex = QUEUE_NONE;
476 if (BUF_REFCNT(bp) <= 1)
477 panic("bremfree: removing a buffer not on a queue");
481 * Fixup numfreebuffers count. If the buffer is invalid or not
482 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
483 * the buffer was free and we must decrement numfreebuffers.
485 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
502 * Get a buffer with the specified data. Look in the cache first. We
503 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
504 * is set, the buffer is valid and we do not have to do anything ( see
508 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
513 bp = getblk(vp, blkno, size, 0, 0);
516 /* if not found in cache, do some I/O */
517 if ((bp->b_flags & B_CACHE) == 0) {
519 curproc->p_stats->p_ru.ru_inblock++;
520 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp));
521 bp->b_flags |= B_READ;
522 bp->b_flags &= ~(B_ERROR | B_INVAL);
523 if (bp->b_rcred == NOCRED) {
528 vfs_busy_pages(bp, 0);
529 VOP_STRATEGY(vp, bp);
530 return (biowait(bp));
536 * Operates like bread, but also starts asynchronous I/O on
537 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
538 * to initiating I/O . If B_CACHE is set, the buffer is valid
539 * and we do not have to do anything.
542 breadn(struct vnode * vp, daddr_t blkno, int size,
543 daddr_t * rablkno, int *rabsize,
544 int cnt, struct ucred * cred, struct buf ** bpp)
546 struct buf *bp, *rabp;
548 int rv = 0, readwait = 0;
550 *bpp = bp = getblk(vp, blkno, size, 0, 0);
552 /* if not found in cache, do some I/O */
553 if ((bp->b_flags & B_CACHE) == 0) {
555 curproc->p_stats->p_ru.ru_inblock++;
556 bp->b_flags |= B_READ;
557 bp->b_flags &= ~(B_ERROR | B_INVAL);
558 if (bp->b_rcred == NOCRED) {
563 vfs_busy_pages(bp, 0);
564 VOP_STRATEGY(vp, bp);
568 for (i = 0; i < cnt; i++, rablkno++, rabsize++) {
569 if (inmem(vp, *rablkno))
571 rabp = getblk(vp, *rablkno, *rabsize, 0, 0);
573 if ((rabp->b_flags & B_CACHE) == 0) {
575 curproc->p_stats->p_ru.ru_inblock++;
576 rabp->b_flags |= B_READ | B_ASYNC;
577 rabp->b_flags &= ~(B_ERROR | B_INVAL);
578 if (rabp->b_rcred == NOCRED) {
581 rabp->b_rcred = cred;
583 vfs_busy_pages(rabp, 0);
585 VOP_STRATEGY(vp, rabp);
598 * Write, release buffer on completion. (Done by iodone
599 * if async). Do not bother writing anything if the buffer
602 * Note that we set B_CACHE here, indicating that buffer is
603 * fully valid and thus cacheable. This is true even of NFS
604 * now so we set it generally. This could be set either here
605 * or in biodone() since the I/O is synchronous. We put it
609 bwrite(struct buf * bp)
614 if (bp->b_flags & B_INVAL) {
619 oldflags = bp->b_flags;
621 if (BUF_REFCNT(bp) == 0)
622 panic("bwrite: buffer is not busy???");
625 * If a background write is already in progress, delay
626 * writing this block if it is asynchronous. Otherwise
627 * wait for the background write to complete.
629 if (bp->b_xflags & BX_BKGRDINPROG) {
630 if (bp->b_flags & B_ASYNC) {
635 bp->b_xflags |= BX_BKGRDWAIT;
636 tsleep(&bp->b_xflags, PRIBIO, "biord", 0);
637 if (bp->b_xflags & BX_BKGRDINPROG)
638 panic("bwrite: still writing");
641 /* Mark the buffer clean */
645 * If this buffer is marked for background writing and we
646 * do not have to wait for it, make a copy and write the
647 * copy so as to leave this buffer ready for further use.
649 * This optimization eats a lot of memory. If we have a page
650 * or buffer shortfull we can't do it.
652 if ((bp->b_xflags & BX_BKGRDWRITE) &&
653 (bp->b_flags & B_ASYNC) &&
654 !vm_page_count_severe() &&
655 !buf_dirty_count_severe()) {
656 if (bp->b_flags & B_CALL)
657 panic("bwrite: need chained iodone");
659 /* get a new block */
660 newbp = geteblk(bp->b_bufsize);
662 /* set it to be identical to the old block */
663 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize);
664 bgetvp(bp->b_vp, newbp);
665 newbp->b_lblkno = bp->b_lblkno;
666 newbp->b_blkno = bp->b_blkno;
667 newbp->b_offset = bp->b_offset;
668 newbp->b_iodone = vfs_backgroundwritedone;
669 newbp->b_flags |= B_ASYNC | B_CALL;
670 newbp->b_flags &= ~B_INVAL;
672 /* move over the dependencies */
673 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps)
674 (*bioops.io_movedeps)(bp, newbp);
677 * Initiate write on the copy, release the original to
678 * the B_LOCKED queue so that it cannot go away until
679 * the background write completes. If not locked it could go
680 * away and then be reconstituted while it was being written.
681 * If the reconstituted buffer were written, we could end up
682 * with two background copies being written at the same time.
684 bp->b_xflags |= BX_BKGRDINPROG;
685 bp->b_flags |= B_LOCKED;
690 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
691 bp->b_flags |= B_WRITEINPROG | B_CACHE;
693 bp->b_vp->v_numoutput++;
694 vfs_busy_pages(bp, 1);
697 * Normal bwrites pipeline writes
699 bp->b_runningbufspace = bp->b_bufsize;
700 runningbufspace += bp->b_runningbufspace;
703 curproc->p_stats->p_ru.ru_oublock++;
705 if (oldflags & B_ASYNC)
707 VOP_STRATEGY(bp->b_vp, bp);
709 if ((oldflags & B_ASYNC) == 0) {
710 int rtval = biowait(bp);
713 } else if ((oldflags & B_NOWDRAIN) == 0) {
715 * don't allow the async write to saturate the I/O
716 * system. Deadlocks can occur only if a device strategy
717 * routine (like in VN) turns around and issues another
718 * high-level write, in which case B_NOWDRAIN is expected
719 * to be set. Otherwise we will not deadlock here because
720 * we are blocking waiting for I/O that is already in-progress
723 waitrunningbufspace();
730 * Complete a background write started from bwrite.
733 vfs_backgroundwritedone(bp)
739 * Find the original buffer that we are writing.
741 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL)
742 panic("backgroundwritedone: lost buffer");
744 * Process dependencies then return any unfinished ones.
746 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
747 (*bioops.io_complete)(bp);
748 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps)
749 (*bioops.io_movedeps)(bp, origbp);
751 * Clear the BX_BKGRDINPROG flag in the original buffer
752 * and awaken it if it is waiting for the write to complete.
753 * If BX_BKGRDINPROG is not set in the original buffer it must
754 * have been released and re-instantiated - which is not legal.
756 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), ("backgroundwritedone: lost buffer2"));
757 origbp->b_xflags &= ~BX_BKGRDINPROG;
758 if (origbp->b_xflags & BX_BKGRDWAIT) {
759 origbp->b_xflags &= ~BX_BKGRDWAIT;
760 wakeup(&origbp->b_xflags);
763 * Clear the B_LOCKED flag and remove it from the locked
764 * queue if it currently resides there.
766 origbp->b_flags &= ~B_LOCKED;
767 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
772 * This buffer is marked B_NOCACHE, so when it is released
773 * by biodone, it will be tossed. We mark it with B_READ
774 * to avoid biodone doing a second vwakeup.
776 bp->b_flags |= B_NOCACHE | B_READ;
777 bp->b_flags &= ~(B_CACHE | B_CALL | B_DONE);
783 * Delayed write. (Buffer is marked dirty). Do not bother writing
784 * anything if the buffer is marked invalid.
786 * Note that since the buffer must be completely valid, we can safely
787 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
788 * biodone() in order to prevent getblk from writing the buffer
792 bdwrite(struct buf * bp)
794 if (BUF_REFCNT(bp) == 0)
795 panic("bdwrite: buffer is not busy");
797 if (bp->b_flags & B_INVAL) {
804 * Set B_CACHE, indicating that the buffer is fully valid. This is
805 * true even of NFS now.
807 bp->b_flags |= B_CACHE;
810 * This bmap keeps the system from needing to do the bmap later,
811 * perhaps when the system is attempting to do a sync. Since it
812 * is likely that the indirect block -- or whatever other datastructure
813 * that the filesystem needs is still in memory now, it is a good
814 * thing to do this. Note also, that if the pageout daemon is
815 * requesting a sync -- there might not be enough memory to do
816 * the bmap then... So, this is important to do.
818 if (bp->b_lblkno == bp->b_blkno) {
819 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
823 * Set the *dirty* buffer range based upon the VM system dirty pages.
828 * We need to do this here to satisfy the vnode_pager and the
829 * pageout daemon, so that it thinks that the pages have been
830 * "cleaned". Note that since the pages are in a delayed write
831 * buffer -- the VFS layer "will" see that the pages get written
832 * out on the next sync, or perhaps the cluster will be completed.
838 * Wakeup the buffer flushing daemon if we have a lot of dirty
839 * buffers (midpoint between our recovery point and our stall
842 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
845 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
846 * due to the softdep code.
853 * Turn buffer into delayed write request. We must clear B_READ and
854 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
855 * itself to properly update it in the dirty/clean lists. We mark it
856 * B_DONE to ensure that any asynchronization of the buffer properly
857 * clears B_DONE ( else a panic will occur later ).
859 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
860 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
861 * should only be called if the buffer is known-good.
863 * Since the buffer is not on a queue, we do not update the numfreebuffers
866 * Must be called at splbio().
867 * The buffer must be on QUEUE_NONE.
873 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
874 bp->b_flags &= ~(B_READ|B_RELBUF);
876 if ((bp->b_flags & B_DELWRI) == 0) {
877 bp->b_flags |= B_DONE | B_DELWRI;
878 reassignbuf(bp, bp->b_vp);
880 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
887 * Clear B_DELWRI for buffer.
889 * Since the buffer is not on a queue, we do not update the numfreebuffers
892 * Must be called at splbio().
893 * The buffer must be on QUEUE_NONE.
900 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
902 if (bp->b_flags & B_DELWRI) {
903 bp->b_flags &= ~B_DELWRI;
904 reassignbuf(bp, bp->b_vp);
906 numdirtywakeup(lodirtybuffers);
909 * Since it is now being written, we can clear its deferred write flag.
911 bp->b_flags &= ~B_DEFERRED;
917 * Asynchronous write. Start output on a buffer, but do not wait for
918 * it to complete. The buffer is released when the output completes.
920 * bwrite() ( or the VOP routine anyway ) is responsible for handling
921 * B_INVAL buffers. Not us.
924 bawrite(struct buf * bp)
926 bp->b_flags |= B_ASYNC;
927 (void) VOP_BWRITE(bp->b_vp, bp);
933 * Ordered write. Start output on a buffer, and flag it so that the
934 * device will write it in the order it was queued. The buffer is
935 * released when the output completes. bwrite() ( or the VOP routine
936 * anyway ) is responsible for handling B_INVAL buffers.
939 bowrite(struct buf * bp)
941 bp->b_flags |= B_ORDERED | B_ASYNC;
942 return (VOP_BWRITE(bp->b_vp, bp));
948 * Called prior to the locking of any vnodes when we are expecting to
949 * write. We do not want to starve the buffer cache with too many
950 * dirty buffers so we block here. By blocking prior to the locking
951 * of any vnodes we attempt to avoid the situation where a locked vnode
952 * prevents the various system daemons from flushing related buffers.
958 if (numdirtybuffers >= hidirtybuffers) {
962 while (numdirtybuffers >= hidirtybuffers) {
964 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
965 tsleep(&needsbuffer, (PRIBIO + 4), "flswai", 0);
972 * Return true if we have too many dirty buffers.
975 buf_dirty_count_severe(void)
977 return(numdirtybuffers >= hidirtybuffers);
983 * Release a busy buffer and, if requested, free its resources. The
984 * buffer will be stashed in the appropriate bufqueue[] allowing it
985 * to be accessed later as a cache entity or reused for other purposes.
988 brelse(struct buf * bp)
992 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
996 if (bp->b_flags & B_LOCKED)
997 bp->b_flags &= ~B_ERROR;
999 if ((bp->b_flags & (B_READ | B_ERROR | B_INVAL)) == B_ERROR) {
1001 * Failed write, redirty. Must clear B_ERROR to prevent
1002 * pages from being scrapped. If B_INVAL is set then
1003 * this case is not run and the next case is run to
1004 * destroy the buffer. B_INVAL can occur if the buffer
1005 * is outside the range supported by the underlying device.
1007 bp->b_flags &= ~B_ERROR;
1009 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) ||
1010 (bp->b_bufsize <= 0)) {
1012 * Either a failed I/O or we were asked to free or not
1015 bp->b_flags |= B_INVAL;
1016 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
1017 (*bioops.io_deallocate)(bp);
1018 if (bp->b_flags & B_DELWRI) {
1020 numdirtywakeup(lodirtybuffers);
1022 bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF);
1023 if ((bp->b_flags & B_VMIO) == 0) {
1032 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1033 * is called with B_DELWRI set, the underlying pages may wind up
1034 * getting freed causing a previous write (bdwrite()) to get 'lost'
1035 * because pages associated with a B_DELWRI bp are marked clean.
1037 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1038 * if B_DELWRI is set.
1040 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1041 * on pages to return pages to the VM page queues.
1043 if (bp->b_flags & B_DELWRI)
1044 bp->b_flags &= ~B_RELBUF;
1045 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG))
1046 bp->b_flags |= B_RELBUF;
1049 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1050 * constituted, not even NFS buffers now. Two flags effect this. If
1051 * B_INVAL, the struct buf is invalidated but the VM object is kept
1052 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1054 * If B_ERROR or B_NOCACHE is set, pages in the VM object will be
1055 * invalidated. B_ERROR cannot be set for a failed write unless the
1056 * buffer is also B_INVAL because it hits the re-dirtying code above.
1058 * Normally we can do this whether a buffer is B_DELWRI or not. If
1059 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1060 * the commit state and we cannot afford to lose the buffer. If the
1061 * buffer has a background write in progress, we need to keep it
1062 * around to prevent it from being reconstituted and starting a second
1065 if ((bp->b_flags & B_VMIO)
1066 && !(bp->b_vp->v_tag == VT_NFS &&
1067 !vn_isdisk(bp->b_vp, NULL) &&
1068 (bp->b_flags & B_DELWRI))
1081 * Get the base offset and length of the buffer. Note that
1082 * in the VMIO case if the buffer block size is not
1083 * page-aligned then b_data pointer may not be page-aligned.
1084 * But our b_pages[] array *IS* page aligned.
1086 * block sizes less then DEV_BSIZE (usually 512) are not
1087 * supported due to the page granularity bits (m->valid,
1088 * m->dirty, etc...).
1090 * See man buf(9) for more information
1093 resid = bp->b_bufsize;
1094 foff = bp->b_offset;
1096 for (i = 0; i < bp->b_npages; i++) {
1098 vm_page_flag_clear(m, PG_ZERO);
1100 * If we hit a bogus page, fixup *all* of them
1103 if (m == bogus_page) {
1104 VOP_GETVOBJECT(vp, &obj);
1105 poff = OFF_TO_IDX(bp->b_offset);
1107 for (j = i; j < bp->b_npages; j++) {
1110 mtmp = bp->b_pages[j];
1111 if (mtmp == bogus_page) {
1112 mtmp = vm_page_lookup(obj, poff + j);
1114 panic("brelse: page missing\n");
1116 bp->b_pages[j] = mtmp;
1120 if ((bp->b_flags & B_INVAL) == 0) {
1121 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
1125 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1126 int poffset = foff & PAGE_MASK;
1127 int presid = resid > (PAGE_SIZE - poffset) ?
1128 (PAGE_SIZE - poffset) : resid;
1130 KASSERT(presid >= 0, ("brelse: extra page"));
1131 vm_page_set_invalid(m, poffset, presid);
1133 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1134 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1137 if (bp->b_flags & (B_INVAL | B_RELBUF))
1138 vfs_vmio_release(bp);
1140 } else if (bp->b_flags & B_VMIO) {
1142 if (bp->b_flags & (B_INVAL | B_RELBUF))
1143 vfs_vmio_release(bp);
1147 if (bp->b_qindex != QUEUE_NONE)
1148 panic("brelse: free buffer onto another queue???");
1149 if (BUF_REFCNT(bp) > 1) {
1150 /* Temporary panic to verify exclusive locking */
1151 /* This panic goes away when we allow shared refs */
1152 panic("brelse: multiple refs");
1153 /* do not release to free list */
1161 /* buffers with no memory */
1162 if (bp->b_bufsize == 0) {
1163 bp->b_flags |= B_INVAL;
1164 bp->b_xflags &= ~BX_BKGRDWRITE;
1165 if (bp->b_xflags & BX_BKGRDINPROG)
1166 panic("losing buffer 1");
1167 if (bp->b_kvasize) {
1168 bp->b_qindex = QUEUE_EMPTYKVA;
1170 bp->b_qindex = QUEUE_EMPTY;
1172 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1173 LIST_REMOVE(bp, b_hash);
1174 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1176 /* buffers with junk contents */
1177 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) {
1178 bp->b_flags |= B_INVAL;
1179 bp->b_xflags &= ~BX_BKGRDWRITE;
1180 if (bp->b_xflags & BX_BKGRDINPROG)
1181 panic("losing buffer 2");
1182 bp->b_qindex = QUEUE_CLEAN;
1183 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1184 LIST_REMOVE(bp, b_hash);
1185 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1188 /* buffers that are locked */
1189 } else if (bp->b_flags & B_LOCKED) {
1190 bp->b_qindex = QUEUE_LOCKED;
1191 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1193 /* remaining buffers */
1195 switch(bp->b_flags & (B_DELWRI|B_AGE)) {
1196 case B_DELWRI | B_AGE:
1197 bp->b_qindex = QUEUE_DIRTY;
1198 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1201 bp->b_qindex = QUEUE_DIRTY;
1202 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1205 bp->b_qindex = QUEUE_CLEAN;
1206 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1209 bp->b_qindex = QUEUE_CLEAN;
1210 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1216 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1217 * on the correct queue.
1219 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1223 * Fixup numfreebuffers count. The bp is on an appropriate queue
1224 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1225 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1226 * if B_INVAL is set ).
1229 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI))
1233 * Something we can maybe free or reuse
1235 if (bp->b_bufsize || bp->b_kvasize)
1240 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF |
1241 B_DIRECT | B_NOWDRAIN);
1246 * Release a buffer back to the appropriate queue but do not try to free
1247 * it. The buffer is expected to be used again soon.
1249 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1250 * biodone() to requeue an async I/O on completion. It is also used when
1251 * known good buffers need to be requeued but we think we may need the data
1254 * XXX we should be able to leave the B_RELBUF hint set on completion.
1257 bqrelse(struct buf * bp)
1263 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1265 if (bp->b_qindex != QUEUE_NONE)
1266 panic("bqrelse: free buffer onto another queue???");
1267 if (BUF_REFCNT(bp) > 1) {
1268 /* do not release to free list */
1269 panic("bqrelse: multiple refs");
1274 if (bp->b_flags & B_LOCKED) {
1275 bp->b_flags &= ~B_ERROR;
1276 bp->b_qindex = QUEUE_LOCKED;
1277 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1278 /* buffers with stale but valid contents */
1279 } else if (bp->b_flags & B_DELWRI) {
1280 bp->b_qindex = QUEUE_DIRTY;
1281 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1282 } else if (vm_page_count_severe()) {
1284 * We are too low on memory, we have to try to free the
1285 * buffer (most importantly: the wired pages making up its
1286 * backing store) *now*.
1292 bp->b_qindex = QUEUE_CLEAN;
1293 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1296 if ((bp->b_flags & B_LOCKED) == 0 &&
1297 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) {
1302 * Something we can maybe free or reuse.
1304 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1309 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1314 vfs_vmio_release(bp)
1321 for (i = 0; i < bp->b_npages; i++) {
1323 bp->b_pages[i] = NULL;
1325 * In order to keep page LRU ordering consistent, put
1326 * everything on the inactive queue.
1328 vm_page_unwire(m, 0);
1330 * We don't mess with busy pages, it is
1331 * the responsibility of the process that
1332 * busied the pages to deal with them.
1334 if ((m->flags & PG_BUSY) || (m->busy != 0))
1337 if (m->wire_count == 0) {
1338 vm_page_flag_clear(m, PG_ZERO);
1340 * Might as well free the page if we can and it has
1341 * no valid data. We also free the page if the
1342 * buffer was used for direct I/O.
1344 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) {
1346 vm_page_protect(m, VM_PROT_NONE);
1348 } else if (bp->b_flags & B_DIRECT) {
1349 vm_page_try_to_free(m);
1350 } else if (vm_page_count_severe()) {
1351 vm_page_try_to_cache(m);
1356 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1357 if (bp->b_bufsize) {
1362 bp->b_flags &= ~B_VMIO;
1368 * Check to see if a block is currently memory resident.
1371 gbincore(struct vnode * vp, daddr_t blkno)
1374 struct bufhashhdr *bh;
1376 bh = bufhash(vp, blkno);
1378 /* Search hash chain */
1379 LIST_FOREACH(bp, bh, b_hash) {
1381 if (bp->b_vp == vp && bp->b_lblkno == blkno &&
1382 (bp->b_flags & B_INVAL) == 0) {
1392 * Implement clustered async writes for clearing out B_DELWRI buffers.
1393 * This is much better then the old way of writing only one buffer at
1394 * a time. Note that we may not be presented with the buffers in the
1395 * correct order, so we search for the cluster in both directions.
1398 vfs_bio_awrite(struct buf * bp)
1402 daddr_t lblkno = bp->b_lblkno;
1403 struct vnode *vp = bp->b_vp;
1413 * right now we support clustered writing only to regular files. If
1414 * we find a clusterable block we could be in the middle of a cluster
1415 * rather then at the beginning.
1417 if ((vp->v_type == VREG) &&
1418 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1419 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1421 size = vp->v_mount->mnt_stat.f_iosize;
1422 maxcl = MAXPHYS / size;
1424 for (i = 1; i < maxcl; i++) {
1425 if ((bpa = gbincore(vp, lblkno + i)) &&
1426 BUF_REFCNT(bpa) == 0 &&
1427 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1428 (B_DELWRI | B_CLUSTEROK)) &&
1429 (bpa->b_bufsize == size)) {
1430 if ((bpa->b_blkno == bpa->b_lblkno) ||
1432 bp->b_blkno + ((i * size) >> DEV_BSHIFT)))
1438 for (j = 1; i + j <= maxcl && j <= lblkno; j++) {
1439 if ((bpa = gbincore(vp, lblkno - j)) &&
1440 BUF_REFCNT(bpa) == 0 &&
1441 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1442 (B_DELWRI | B_CLUSTEROK)) &&
1443 (bpa->b_bufsize == size)) {
1444 if ((bpa->b_blkno == bpa->b_lblkno) ||
1446 bp->b_blkno - ((j * size) >> DEV_BSHIFT)))
1455 * this is a possible cluster write
1458 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1464 BUF_LOCK(bp, LK_EXCLUSIVE);
1466 bp->b_flags |= B_ASYNC;
1470 * default (old) behavior, writing out only one block
1472 * XXX returns b_bufsize instead of b_bcount for nwritten?
1474 nwritten = bp->b_bufsize;
1475 (void) VOP_BWRITE(bp->b_vp, bp);
1483 * Find and initialize a new buffer header, freeing up existing buffers
1484 * in the bufqueues as necessary. The new buffer is returned locked.
1486 * Important: B_INVAL is not set. If the caller wishes to throw the
1487 * buffer away, the caller must set B_INVAL prior to calling brelse().
1490 * We have insufficient buffer headers
1491 * We have insufficient buffer space
1492 * buffer_map is too fragmented ( space reservation fails )
1493 * If we have to flush dirty buffers ( but we try to avoid this )
1495 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1496 * Instead we ask the buf daemon to do it for us. We attempt to
1497 * avoid piecemeal wakeups of the pageout daemon.
1501 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1507 static int flushingbufs;
1510 * We can't afford to block since we might be holding a vnode lock,
1511 * which may prevent system daemons from running. We deal with
1512 * low-memory situations by proactively returning memory and running
1513 * async I/O rather then sync I/O.
1517 --getnewbufrestarts;
1519 ++getnewbufrestarts;
1522 * Setup for scan. If we do not have enough free buffers,
1523 * we setup a degenerate case that immediately fails. Note
1524 * that if we are specially marked process, we are allowed to
1525 * dip into our reserves.
1527 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1529 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1530 * However, there are a number of cases (defragging, reusing, ...)
1531 * where we cannot backup.
1533 nqindex = QUEUE_EMPTYKVA;
1534 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1538 * If no EMPTYKVA buffers and we are either
1539 * defragging or reusing, locate a CLEAN buffer
1540 * to free or reuse. If bufspace useage is low
1541 * skip this step so we can allocate a new buffer.
1543 if (defrag || bufspace >= lobufspace) {
1544 nqindex = QUEUE_CLEAN;
1545 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1549 * If we could not find or were not allowed to reuse a
1550 * CLEAN buffer, check to see if it is ok to use an EMPTY
1551 * buffer. We can only use an EMPTY buffer if allocating
1552 * its KVA would not otherwise run us out of buffer space.
1554 if (nbp == NULL && defrag == 0 &&
1555 bufspace + maxsize < hibufspace) {
1556 nqindex = QUEUE_EMPTY;
1557 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1562 * Run scan, possibly freeing data and/or kva mappings on the fly
1566 while ((bp = nbp) != NULL) {
1567 int qindex = nqindex;
1570 * Calculate next bp ( we can only use it if we do not block
1571 * or do other fancy things ).
1573 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1576 nqindex = QUEUE_EMPTYKVA;
1577 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1580 case QUEUE_EMPTYKVA:
1581 nqindex = QUEUE_CLEAN;
1582 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1596 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1599 * Note: we no longer distinguish between VMIO and non-VMIO
1603 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1606 * If we are defragging then we need a buffer with
1607 * b_kvasize != 0. XXX this situation should no longer
1608 * occur, if defrag is non-zero the buffer's b_kvasize
1609 * should also be non-zero at this point. XXX
1611 if (defrag && bp->b_kvasize == 0) {
1612 printf("Warning: defrag empty buffer %p\n", bp);
1617 * Start freeing the bp. This is somewhat involved. nbp
1618 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1621 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1622 panic("getnewbuf: locked buf");
1625 if (qindex == QUEUE_CLEAN) {
1626 if (bp->b_flags & B_VMIO) {
1627 bp->b_flags &= ~B_ASYNC;
1628 vfs_vmio_release(bp);
1635 * NOTE: nbp is now entirely invalid. We can only restart
1636 * the scan from this point on.
1638 * Get the rest of the buffer freed up. b_kva* is still
1639 * valid after this operation.
1642 if (bp->b_rcred != NOCRED) {
1643 crfree(bp->b_rcred);
1644 bp->b_rcred = NOCRED;
1646 if (bp->b_wcred != NOCRED) {
1647 crfree(bp->b_wcred);
1648 bp->b_wcred = NOCRED;
1650 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
1651 (*bioops.io_deallocate)(bp);
1652 if (bp->b_xflags & BX_BKGRDINPROG)
1653 panic("losing buffer 3");
1654 LIST_REMOVE(bp, b_hash);
1655 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1664 bp->b_blkno = bp->b_lblkno = 0;
1665 bp->b_offset = NOOFFSET;
1671 bp->b_dirtyoff = bp->b_dirtyend = 0;
1673 LIST_INIT(&bp->b_dep);
1676 * If we are defragging then free the buffer.
1679 bp->b_flags |= B_INVAL;
1687 * If we are overcomitted then recover the buffer and its
1688 * KVM space. This occurs in rare situations when multiple
1689 * processes are blocked in getnewbuf() or allocbuf().
1691 if (bufspace >= hibufspace)
1693 if (flushingbufs && bp->b_kvasize != 0) {
1694 bp->b_flags |= B_INVAL;
1699 if (bufspace < lobufspace)
1705 * If we exhausted our list, sleep as appropriate. We may have to
1706 * wakeup various daemons and write out some dirty buffers.
1708 * Generally we are sleeping due to insufficient buffer space.
1716 flags = VFS_BIO_NEED_BUFSPACE;
1718 } else if (bufspace >= hibufspace) {
1720 flags = VFS_BIO_NEED_BUFSPACE;
1723 flags = VFS_BIO_NEED_ANY;
1726 bd_speedup(); /* heeeelp */
1728 needsbuffer |= flags;
1729 while (needsbuffer & flags) {
1730 if (tsleep(&needsbuffer, (PRIBIO + 4) | slpflag,
1736 * We finally have a valid bp. We aren't quite out of the
1737 * woods, we still have to reserve kva space. In order
1738 * to keep fragmentation sane we only allocate kva in
1741 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1743 if (maxsize != bp->b_kvasize) {
1744 vm_offset_t addr = 0;
1748 vm_map_lock(buffer_map);
1750 if (vm_map_findspace(buffer_map,
1751 vm_map_min(buffer_map), maxsize, &addr)) {
1753 * Uh oh. Buffer map is to fragmented. We
1754 * must defragment the map.
1756 vm_map_unlock(buffer_map);
1759 bp->b_flags |= B_INVAL;
1764 vm_map_insert(buffer_map, NULL, 0,
1765 addr, addr + maxsize,
1766 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
1768 bp->b_kvabase = (caddr_t) addr;
1769 bp->b_kvasize = maxsize;
1770 bufspace += bp->b_kvasize;
1773 vm_map_unlock(buffer_map);
1775 bp->b_data = bp->b_kvabase;
1783 * buffer flushing daemon. Buffers are normally flushed by the
1784 * update daemon but if it cannot keep up this process starts to
1785 * take the load in an attempt to prevent getnewbuf() from blocking.
1788 static struct proc *bufdaemonproc;
1790 static struct kproc_desc buf_kp = {
1795 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
1803 * This process needs to be suspended prior to shutdown sync.
1805 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, bufdaemonproc,
1809 * This process is allowed to take the buffer cache to the limit
1814 kproc_suspend_loop(bufdaemonproc);
1817 * Do the flush. Limit the amount of in-transit I/O we
1818 * allow to build up, otherwise we would completely saturate
1819 * the I/O system. Wakeup any waiting processes before we
1820 * normally would so they can run in parallel with our drain.
1822 while (numdirtybuffers > lodirtybuffers) {
1823 if (flushbufqueues() == 0)
1825 waitrunningbufspace();
1826 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
1830 * Only clear bd_request if we have reached our low water
1831 * mark. The buf_daemon normally waits 5 seconds and
1832 * then incrementally flushes any dirty buffers that have
1833 * built up, within reason.
1835 * If we were unable to hit our low water mark and couldn't
1836 * find any flushable buffers, we sleep half a second.
1837 * Otherwise we loop immediately.
1839 if (numdirtybuffers <= lodirtybuffers) {
1841 * We reached our low water mark, reset the
1842 * request and sleep until we are needed again.
1843 * The sleep is just so the suspend code works.
1846 tsleep(&bd_request, PVM, "psleep", hz);
1849 * We couldn't find any flushable dirty buffers but
1850 * still have too many dirty buffers, we
1851 * have to sleep and try again. (rare)
1853 tsleep(&bd_request, PVM, "qsleep", hz / 2);
1861 * Try to flush a buffer in the dirty queue. We must be careful to
1862 * free up B_INVAL buffers instead of write them, which NFS is
1863 * particularly sensitive to.
1867 flushbufqueues(void)
1872 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
1875 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp));
1876 if ((bp->b_flags & B_DELWRI) != 0 &&
1877 (bp->b_xflags & BX_BKGRDINPROG) == 0) {
1878 if (bp->b_flags & B_INVAL) {
1879 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1880 panic("flushbufqueues: locked buf");
1886 if (LIST_FIRST(&bp->b_dep) != NULL &&
1887 bioops.io_countdeps &&
1888 (bp->b_flags & B_DEFERRED) == 0 &&
1889 (*bioops.io_countdeps)(bp, 0)) {
1890 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY],
1892 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY],
1894 bp->b_flags |= B_DEFERRED;
1895 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
1902 bp = TAILQ_NEXT(bp, b_freelist);
1908 * Check to see if a block is currently memory resident.
1911 incore(struct vnode * vp, daddr_t blkno)
1916 bp = gbincore(vp, blkno);
1922 * Returns true if no I/O is needed to access the
1923 * associated VM object. This is like incore except
1924 * it also hunts around in the VM system for the data.
1928 inmem(struct vnode * vp, daddr_t blkno)
1931 vm_offset_t toff, tinc, size;
1935 if (incore(vp, blkno))
1937 if (vp->v_mount == NULL)
1939 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0)
1943 if (size > vp->v_mount->mnt_stat.f_iosize)
1944 size = vp->v_mount->mnt_stat.f_iosize;
1945 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
1947 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
1948 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
1952 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
1953 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
1954 if (vm_page_is_valid(m,
1955 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
1964 * Sets the dirty range for a buffer based on the status of the dirty
1965 * bits in the pages comprising the buffer.
1967 * The range is limited to the size of the buffer.
1969 * This routine is primarily used by NFS, but is generalized for the
1973 vfs_setdirty(struct buf *bp)
1979 * Degenerate case - empty buffer
1982 if (bp->b_bufsize == 0)
1986 * We qualify the scan for modified pages on whether the
1987 * object has been flushed yet. The OBJ_WRITEABLE flag
1988 * is not cleared simply by protecting pages off.
1991 if ((bp->b_flags & B_VMIO) == 0)
1994 object = bp->b_pages[0]->object;
1996 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
1997 printf("Warning: object %p writeable but not mightbedirty\n", object);
1998 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
1999 printf("Warning: object %p mightbedirty but not writeable\n", object);
2001 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2002 vm_offset_t boffset;
2003 vm_offset_t eoffset;
2006 * test the pages to see if they have been modified directly
2007 * by users through the VM system.
2009 for (i = 0; i < bp->b_npages; i++) {
2010 vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
2011 vm_page_test_dirty(bp->b_pages[i]);
2015 * Calculate the encompassing dirty range, boffset and eoffset,
2016 * (eoffset - boffset) bytes.
2019 for (i = 0; i < bp->b_npages; i++) {
2020 if (bp->b_pages[i]->dirty)
2023 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2025 for (i = bp->b_npages - 1; i >= 0; --i) {
2026 if (bp->b_pages[i]->dirty) {
2030 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2033 * Fit it to the buffer.
2036 if (eoffset > bp->b_bcount)
2037 eoffset = bp->b_bcount;
2040 * If we have a good dirty range, merge with the existing
2044 if (boffset < eoffset) {
2045 if (bp->b_dirtyoff > boffset)
2046 bp->b_dirtyoff = boffset;
2047 if (bp->b_dirtyend < eoffset)
2048 bp->b_dirtyend = eoffset;
2056 * Get a block given a specified block and offset into a file/device.
2057 * The buffers B_DONE bit will be cleared on return, making it almost
2058 * ready for an I/O initiation. B_INVAL may or may not be set on
2059 * return. The caller should clear B_INVAL prior to initiating a
2062 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2063 * an existing buffer.
2065 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2066 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2067 * and then cleared based on the backing VM. If the previous buffer is
2068 * non-0-sized but invalid, B_CACHE will be cleared.
2070 * If getblk() must create a new buffer, the new buffer is returned with
2071 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2072 * case it is returned with B_INVAL clear and B_CACHE set based on the
2075 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos
2076 * B_CACHE bit is clear.
2078 * What this means, basically, is that the caller should use B_CACHE to
2079 * determine whether the buffer is fully valid or not and should clear
2080 * B_INVAL prior to issuing a read. If the caller intends to validate
2081 * the buffer by loading its data area with something, the caller needs
2082 * to clear B_INVAL. If the caller does this without issuing an I/O,
2083 * the caller should set B_CACHE ( as an optimization ), else the caller
2084 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2085 * a write attempt or if it was a successfull read. If the caller
2086 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2087 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2090 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo)
2094 struct bufhashhdr *bh;
2096 if (size > MAXBSIZE)
2097 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2102 * Block if we are low on buffers. Certain processes are allowed
2103 * to completely exhaust the buffer cache.
2105 * If this check ever becomes a bottleneck it may be better to
2106 * move it into the else, when gbincore() fails. At the moment
2107 * it isn't a problem.
2109 * XXX remove, we cannot afford to block anywhere if holding a vnode
2110 * lock in low-memory situation, so take it to the max.
2112 if (numfreebuffers == 0) {
2115 needsbuffer |= VFS_BIO_NEED_ANY;
2116 tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf",
2120 if ((bp = gbincore(vp, blkno))) {
2122 * Buffer is in-core. If the buffer is not busy, it must
2126 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2127 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL,
2128 "getblk", slpflag, slptimeo) == ENOLCK)
2131 return (struct buf *) NULL;
2135 * The buffer is locked. B_CACHE is cleared if the buffer is
2136 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set
2137 * and for a VMIO buffer B_CACHE is adjusted according to the
2140 if (bp->b_flags & B_INVAL)
2141 bp->b_flags &= ~B_CACHE;
2142 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2143 bp->b_flags |= B_CACHE;
2147 * check for size inconsistancies for non-VMIO case.
2150 if (bp->b_bcount != size) {
2151 if ((bp->b_flags & B_VMIO) == 0 ||
2152 (size > bp->b_kvasize)) {
2153 if (bp->b_flags & B_DELWRI) {
2154 bp->b_flags |= B_NOCACHE;
2155 VOP_BWRITE(bp->b_vp, bp);
2157 if ((bp->b_flags & B_VMIO) &&
2158 (LIST_FIRST(&bp->b_dep) == NULL)) {
2159 bp->b_flags |= B_RELBUF;
2162 bp->b_flags |= B_NOCACHE;
2163 VOP_BWRITE(bp->b_vp, bp);
2171 * If the size is inconsistant in the VMIO case, we can resize
2172 * the buffer. This might lead to B_CACHE getting set or
2173 * cleared. If the size has not changed, B_CACHE remains
2174 * unchanged from its previous state.
2177 if (bp->b_bcount != size)
2180 KASSERT(bp->b_offset != NOOFFSET,
2181 ("getblk: no buffer offset"));
2184 * A buffer with B_DELWRI set and B_CACHE clear must
2185 * be committed before we can return the buffer in
2186 * order to prevent the caller from issuing a read
2187 * ( due to B_CACHE not being set ) and overwriting
2190 * Most callers, including NFS and FFS, need this to
2191 * operate properly either because they assume they
2192 * can issue a read if B_CACHE is not set, or because
2193 * ( for example ) an uncached B_DELWRI might loop due
2194 * to softupdates re-dirtying the buffer. In the latter
2195 * case, B_CACHE is set after the first write completes,
2196 * preventing further loops.
2198 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2199 * above while extending the buffer, we cannot allow the
2200 * buffer to remain with B_CACHE set after the write
2201 * completes or it will represent a corrupt state. To
2202 * deal with this we set B_NOCACHE to scrap the buffer
2205 * We might be able to do something fancy, like setting
2206 * B_CACHE in bwrite() except if B_DELWRI is already set,
2207 * so the below call doesn't set B_CACHE, but that gets real
2208 * confusing. This is much easier.
2211 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2212 bp->b_flags |= B_NOCACHE;
2213 VOP_BWRITE(bp->b_vp, bp);
2218 bp->b_flags &= ~B_DONE;
2221 * Buffer is not in-core, create new buffer. The buffer
2222 * returned by getnewbuf() is locked. Note that the returned
2223 * buffer is also considered valid (not marked B_INVAL).
2225 int bsize, maxsize, vmio;
2228 if (vn_isdisk(vp, NULL))
2230 else if (vp->v_mountedhere)
2231 bsize = vp->v_mountedhere->mnt_stat.f_iosize;
2232 else if (vp->v_mount)
2233 bsize = vp->v_mount->mnt_stat.f_iosize;
2237 offset = (off_t)blkno * bsize;
2238 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF);
2239 maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2240 maxsize = imax(maxsize, bsize);
2242 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) {
2243 if (slpflag || slptimeo) {
2251 * This code is used to make sure that a buffer is not
2252 * created while the getnewbuf routine is blocked.
2253 * This can be a problem whether the vnode is locked or not.
2254 * If the buffer is created out from under us, we have to
2255 * throw away the one we just created. There is now window
2256 * race because we are safely running at splbio() from the
2257 * point of the duplicate buffer creation through to here,
2258 * and we've locked the buffer.
2260 if (gbincore(vp, blkno)) {
2261 bp->b_flags |= B_INVAL;
2267 * Insert the buffer into the hash, so that it can
2268 * be found by incore.
2270 bp->b_blkno = bp->b_lblkno = blkno;
2271 bp->b_offset = offset;
2274 LIST_REMOVE(bp, b_hash);
2275 bh = bufhash(vp, blkno);
2276 LIST_INSERT_HEAD(bh, bp, b_hash);
2279 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
2280 * buffer size starts out as 0, B_CACHE will be set by
2281 * allocbuf() for the VMIO case prior to it testing the
2282 * backing store for validity.
2286 bp->b_flags |= B_VMIO;
2287 #if defined(VFS_BIO_DEBUG)
2288 if (vp->v_type != VREG && vp->v_type != VBLK)
2289 printf("getblk: vmioing file type %d???\n", vp->v_type);
2292 bp->b_flags &= ~B_VMIO;
2298 bp->b_flags &= ~B_DONE;
2304 * Get an empty, disassociated buffer of given size. The buffer is initially
2314 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2317 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0);
2320 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2326 * This code constitutes the buffer memory from either anonymous system
2327 * memory (in the case of non-VMIO operations) or from an associated
2328 * VM object (in the case of VMIO operations). This code is able to
2329 * resize a buffer up or down.
2331 * Note that this code is tricky, and has many complications to resolve
2332 * deadlock or inconsistant data situations. Tread lightly!!!
2333 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2334 * the caller. Calling this code willy nilly can result in the loss of data.
2336 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2337 * B_CACHE for the non-VMIO case.
2341 allocbuf(struct buf *bp, int size)
2343 int newbsize, mbsize;
2346 if (BUF_REFCNT(bp) == 0)
2347 panic("allocbuf: buffer not busy");
2349 if (bp->b_kvasize < size)
2350 panic("allocbuf: buffer too small");
2352 if ((bp->b_flags & B_VMIO) == 0) {
2356 * Just get anonymous memory from the kernel. Don't
2357 * mess with B_CACHE.
2359 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2360 #if !defined(NO_B_MALLOC)
2361 if (bp->b_flags & B_MALLOC)
2365 newbsize = round_page(size);
2367 if (newbsize < bp->b_bufsize) {
2368 #if !defined(NO_B_MALLOC)
2370 * malloced buffers are not shrunk
2372 if (bp->b_flags & B_MALLOC) {
2374 bp->b_bcount = size;
2376 free(bp->b_data, M_BIOBUF);
2377 if (bp->b_bufsize) {
2378 bufmallocspace -= bp->b_bufsize;
2382 bp->b_data = bp->b_kvabase;
2384 bp->b_flags &= ~B_MALLOC;
2391 (vm_offset_t) bp->b_data + newbsize,
2392 (vm_offset_t) bp->b_data + bp->b_bufsize);
2393 } else if (newbsize > bp->b_bufsize) {
2394 #if !defined(NO_B_MALLOC)
2396 * We only use malloced memory on the first allocation.
2397 * and revert to page-allocated memory when the buffer
2400 if ( (bufmallocspace < maxbufmallocspace) &&
2401 (bp->b_bufsize == 0) &&
2402 (mbsize <= PAGE_SIZE/2)) {
2404 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2405 bp->b_bufsize = mbsize;
2406 bp->b_bcount = size;
2407 bp->b_flags |= B_MALLOC;
2408 bufmallocspace += mbsize;
2414 #if !defined(NO_B_MALLOC)
2416 * If the buffer is growing on its other-than-first allocation,
2417 * then we revert to the page-allocation scheme.
2419 if (bp->b_flags & B_MALLOC) {
2420 origbuf = bp->b_data;
2421 origbufsize = bp->b_bufsize;
2422 bp->b_data = bp->b_kvabase;
2423 if (bp->b_bufsize) {
2424 bufmallocspace -= bp->b_bufsize;
2428 bp->b_flags &= ~B_MALLOC;
2429 newbsize = round_page(newbsize);
2434 (vm_offset_t) bp->b_data + bp->b_bufsize,
2435 (vm_offset_t) bp->b_data + newbsize);
2436 #if !defined(NO_B_MALLOC)
2438 bcopy(origbuf, bp->b_data, origbufsize);
2439 free(origbuf, M_BIOBUF);
2447 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2448 desiredpages = (size == 0) ? 0 :
2449 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2451 #if !defined(NO_B_MALLOC)
2452 if (bp->b_flags & B_MALLOC)
2453 panic("allocbuf: VMIO buffer can't be malloced");
2456 * Set B_CACHE initially if buffer is 0 length or will become
2459 if (size == 0 || bp->b_bufsize == 0)
2460 bp->b_flags |= B_CACHE;
2462 if (newbsize < bp->b_bufsize) {
2464 * DEV_BSIZE aligned new buffer size is less then the
2465 * DEV_BSIZE aligned existing buffer size. Figure out
2466 * if we have to remove any pages.
2468 if (desiredpages < bp->b_npages) {
2469 for (i = desiredpages; i < bp->b_npages; i++) {
2471 * the page is not freed here -- it
2472 * is the responsibility of
2473 * vnode_pager_setsize
2476 KASSERT(m != bogus_page,
2477 ("allocbuf: bogus page found"));
2478 while (vm_page_sleep_busy(m, TRUE, "biodep"))
2481 bp->b_pages[i] = NULL;
2482 vm_page_unwire(m, 0);
2484 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2485 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2486 bp->b_npages = desiredpages;
2488 } else if (size > bp->b_bcount) {
2490 * We are growing the buffer, possibly in a
2491 * byte-granular fashion.
2499 * Step 1, bring in the VM pages from the object,
2500 * allocating them if necessary. We must clear
2501 * B_CACHE if these pages are not valid for the
2502 * range covered by the buffer.
2506 VOP_GETVOBJECT(vp, &obj);
2508 while (bp->b_npages < desiredpages) {
2512 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2513 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2515 * note: must allocate system pages
2516 * since blocking here could intefere
2517 * with paging I/O, no matter which
2520 m = vm_page_alloc(obj, pi, VM_ALLOC_SYSTEM);
2523 vm_pageout_deficit += desiredpages - bp->b_npages;
2527 bp->b_flags &= ~B_CACHE;
2528 bp->b_pages[bp->b_npages] = m;
2535 * We found a page. If we have to sleep on it,
2536 * retry because it might have gotten freed out
2539 * We can only test PG_BUSY here. Blocking on
2540 * m->busy might lead to a deadlock:
2542 * vm_fault->getpages->cluster_read->allocbuf
2546 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
2550 * We have a good page. Should we wakeup the
2553 if ((curproc != pageproc) &&
2554 ((m->queue - m->pc) == PQ_CACHE) &&
2555 ((cnt.v_free_count + cnt.v_cache_count) <
2556 (cnt.v_free_min + cnt.v_cache_min))) {
2557 pagedaemon_wakeup();
2559 vm_page_flag_clear(m, PG_ZERO);
2561 bp->b_pages[bp->b_npages] = m;
2566 * Step 2. We've loaded the pages into the buffer,
2567 * we have to figure out if we can still have B_CACHE
2568 * set. Note that B_CACHE is set according to the
2569 * byte-granular range ( bcount and size ), new the
2570 * aligned range ( newbsize ).
2572 * The VM test is against m->valid, which is DEV_BSIZE
2573 * aligned. Needless to say, the validity of the data
2574 * needs to also be DEV_BSIZE aligned. Note that this
2575 * fails with NFS if the server or some other client
2576 * extends the file's EOF. If our buffer is resized,
2577 * B_CACHE may remain set! XXX
2580 toff = bp->b_bcount;
2581 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2583 while ((bp->b_flags & B_CACHE) && toff < size) {
2586 if (tinc > (size - toff))
2589 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2604 * Step 3, fixup the KVM pmap. Remember that
2605 * bp->b_data is relative to bp->b_offset, but
2606 * bp->b_offset may be offset into the first page.
2609 bp->b_data = (caddr_t)
2610 trunc_page((vm_offset_t)bp->b_data);
2612 (vm_offset_t)bp->b_data,
2616 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2617 (vm_offset_t)(bp->b_offset & PAGE_MASK));
2620 if (newbsize < bp->b_bufsize)
2622 bp->b_bufsize = newbsize; /* actual buffer allocation */
2623 bp->b_bcount = size; /* requested buffer size */
2630 * Wait for buffer I/O completion, returning error status. The buffer
2631 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR
2632 * error and cleared.
2635 biowait(register struct buf * bp)
2640 while ((bp->b_flags & B_DONE) == 0) {
2641 #if defined(NO_SCHEDULE_MODS)
2642 tsleep(bp, PRIBIO, "biowait", 0);
2644 if (bp->b_flags & B_READ)
2645 tsleep(bp, PRIBIO, "biord", 0);
2647 tsleep(bp, PRIBIO, "biowr", 0);
2651 if (bp->b_flags & B_EINTR) {
2652 bp->b_flags &= ~B_EINTR;
2655 if (bp->b_flags & B_ERROR) {
2656 return (bp->b_error ? bp->b_error : EIO);
2665 * Finish I/O on a buffer, optionally calling a completion function.
2666 * This is usually called from an interrupt so process blocking is
2669 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
2670 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
2671 * assuming B_INVAL is clear.
2673 * For the VMIO case, we set B_CACHE if the op was a read and no
2674 * read error occured, or if the op was a write. B_CACHE is never
2675 * set if the buffer is invalid or otherwise uncacheable.
2677 * biodone does not mess with B_INVAL, allowing the I/O routine or the
2678 * initiator to leave B_INVAL set to brelse the buffer out of existance
2679 * in the biodone routine.
2682 biodone(register struct buf * bp)
2688 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp)));
2689 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
2691 bp->b_flags |= B_DONE;
2692 runningbufwakeup(bp);
2694 if (bp->b_flags & B_FREEBUF) {
2700 if ((bp->b_flags & B_READ) == 0) {
2704 /* call optional completion function if requested */
2705 if (bp->b_flags & B_CALL) {
2706 bp->b_flags &= ~B_CALL;
2707 (*bp->b_iodone) (bp);
2711 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
2712 (*bioops.io_complete)(bp);
2714 if (bp->b_flags & B_VMIO) {
2720 struct vnode *vp = bp->b_vp;
2722 error = VOP_GETVOBJECT(vp, &obj);
2724 #if defined(VFS_BIO_DEBUG)
2725 if (vp->v_usecount == 0) {
2726 panic("biodone: zero vnode ref count");
2730 panic("biodone: missing VM object");
2733 if ((vp->v_flag & VOBJBUF) == 0) {
2734 panic("biodone: vnode is not setup for merged cache");
2738 foff = bp->b_offset;
2739 KASSERT(bp->b_offset != NOOFFSET,
2740 ("biodone: no buffer offset"));
2743 panic("biodone: no object");
2745 #if defined(VFS_BIO_DEBUG)
2746 if (obj->paging_in_progress < bp->b_npages) {
2747 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
2748 obj->paging_in_progress, bp->b_npages);
2753 * Set B_CACHE if the op was a normal read and no error
2754 * occured. B_CACHE is set for writes in the b*write()
2757 iosize = bp->b_bcount - bp->b_resid;
2758 if ((bp->b_flags & (B_READ|B_FREEBUF|B_INVAL|B_NOCACHE|B_ERROR)) == B_READ) {
2759 bp->b_flags |= B_CACHE;
2762 for (i = 0; i < bp->b_npages; i++) {
2766 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2771 * cleanup bogus pages, restoring the originals
2774 if (m == bogus_page) {
2776 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2778 panic("biodone: page disappeared");
2780 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2782 #if defined(VFS_BIO_DEBUG)
2783 if (OFF_TO_IDX(foff) != m->pindex) {
2785 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
2786 (unsigned long)foff, m->pindex);
2791 * In the write case, the valid and clean bits are
2792 * already changed correctly ( see bdwrite() ), so we
2793 * only need to do this here in the read case.
2795 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) {
2796 vfs_page_set_valid(bp, foff, i, m);
2798 vm_page_flag_clear(m, PG_ZERO);
2801 * when debugging new filesystems or buffer I/O methods, this
2802 * is the most common error that pops up. if you see this, you
2803 * have not set the page busy flag correctly!!!
2806 printf("biodone: page busy < 0, "
2807 "pindex: %d, foff: 0x(%x,%x), "
2808 "resid: %d, index: %d\n",
2809 (int) m->pindex, (int)(foff >> 32),
2810 (int) foff & 0xffffffff, resid, i);
2811 if (!vn_isdisk(vp, NULL))
2812 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n",
2813 bp->b_vp->v_mount->mnt_stat.f_iosize,
2815 bp->b_flags, bp->b_npages);
2817 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n",
2819 bp->b_flags, bp->b_npages);
2820 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
2821 m->valid, m->dirty, m->wire_count);
2822 panic("biodone: page busy < 0\n");
2824 vm_page_io_finish(m);
2825 vm_object_pip_subtract(obj, 1);
2826 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2830 vm_object_pip_wakeupn(obj, 0);
2834 * For asynchronous completions, release the buffer now. The brelse
2835 * will do a wakeup there if necessary - so no need to do a wakeup
2836 * here in the async case. The sync case always needs to do a wakeup.
2839 if (bp->b_flags & B_ASYNC) {
2840 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
2851 * This routine is called in lieu of iodone in the case of
2852 * incomplete I/O. This keeps the busy status for pages
2856 vfs_unbusy_pages(struct buf * bp)
2860 runningbufwakeup(bp);
2861 if (bp->b_flags & B_VMIO) {
2862 struct vnode *vp = bp->b_vp;
2865 VOP_GETVOBJECT(vp, &obj);
2867 for (i = 0; i < bp->b_npages; i++) {
2868 vm_page_t m = bp->b_pages[i];
2870 if (m == bogus_page) {
2871 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
2873 panic("vfs_unbusy_pages: page missing\n");
2876 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2878 vm_object_pip_subtract(obj, 1);
2879 vm_page_flag_clear(m, PG_ZERO);
2880 vm_page_io_finish(m);
2882 vm_object_pip_wakeupn(obj, 0);
2887 * vfs_page_set_valid:
2889 * Set the valid bits in a page based on the supplied offset. The
2890 * range is restricted to the buffer's size.
2892 * This routine is typically called after a read completes.
2895 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
2897 vm_ooffset_t soff, eoff;
2900 * Start and end offsets in buffer. eoff - soff may not cross a
2901 * page boundry or cross the end of the buffer. The end of the
2902 * buffer, in this case, is our file EOF, not the allocation size
2906 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2907 if (eoff > bp->b_offset + bp->b_bcount)
2908 eoff = bp->b_offset + bp->b_bcount;
2911 * Set valid range. This is typically the entire buffer and thus the
2915 vm_page_set_validclean(
2917 (vm_offset_t) (soff & PAGE_MASK),
2918 (vm_offset_t) (eoff - soff)
2924 * This routine is called before a device strategy routine.
2925 * It is used to tell the VM system that paging I/O is in
2926 * progress, and treat the pages associated with the buffer
2927 * almost as being PG_BUSY. Also the object paging_in_progress
2928 * flag is handled to make sure that the object doesn't become
2931 * Since I/O has not been initiated yet, certain buffer flags
2932 * such as B_ERROR or B_INVAL may be in an inconsistant state
2933 * and should be ignored.
2936 vfs_busy_pages(struct buf * bp, int clear_modify)
2940 if (bp->b_flags & B_VMIO) {
2941 struct vnode *vp = bp->b_vp;
2945 VOP_GETVOBJECT(vp, &obj);
2946 foff = bp->b_offset;
2947 KASSERT(bp->b_offset != NOOFFSET,
2948 ("vfs_busy_pages: no buffer offset"));
2952 for (i = 0; i < bp->b_npages; i++) {
2953 vm_page_t m = bp->b_pages[i];
2954 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
2959 for (i = 0; i < bp->b_npages; i++) {
2960 vm_page_t m = bp->b_pages[i];
2962 vm_page_flag_clear(m, PG_ZERO);
2963 if ((bp->b_flags & B_CLUSTER) == 0) {
2964 vm_object_pip_add(obj, 1);
2965 vm_page_io_start(m);
2969 * When readying a buffer for a read ( i.e
2970 * clear_modify == 0 ), it is important to do
2971 * bogus_page replacement for valid pages in
2972 * partially instantiated buffers. Partially
2973 * instantiated buffers can, in turn, occur when
2974 * reconstituting a buffer from its VM backing store
2975 * base. We only have to do this if B_CACHE is
2976 * clear ( which causes the I/O to occur in the
2977 * first place ). The replacement prevents the read
2978 * I/O from overwriting potentially dirty VM-backed
2979 * pages. XXX bogus page replacement is, uh, bogus.
2980 * It may not work properly with small-block devices.
2981 * We need to find a better way.
2984 vm_page_protect(m, VM_PROT_NONE);
2986 vfs_page_set_valid(bp, foff, i, m);
2987 else if (m->valid == VM_PAGE_BITS_ALL &&
2988 (bp->b_flags & B_CACHE) == 0) {
2989 bp->b_pages[i] = bogus_page;
2992 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2995 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
3000 * Tell the VM system that the pages associated with this buffer
3001 * are clean. This is used for delayed writes where the data is
3002 * going to go to disk eventually without additional VM intevention.
3004 * Note that while we only really need to clean through to b_bcount, we
3005 * just go ahead and clean through to b_bufsize.
3008 vfs_clean_pages(struct buf * bp)
3012 if (bp->b_flags & B_VMIO) {
3015 foff = bp->b_offset;
3016 KASSERT(bp->b_offset != NOOFFSET,
3017 ("vfs_clean_pages: no buffer offset"));
3018 for (i = 0; i < bp->b_npages; i++) {
3019 vm_page_t m = bp->b_pages[i];
3020 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3021 vm_ooffset_t eoff = noff;
3023 if (eoff > bp->b_offset + bp->b_bufsize)
3024 eoff = bp->b_offset + bp->b_bufsize;
3025 vfs_page_set_valid(bp, foff, i, m);
3026 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3033 * vfs_bio_set_validclean:
3035 * Set the range within the buffer to valid and clean. The range is
3036 * relative to the beginning of the buffer, b_offset. Note that b_offset
3037 * itself may be offset from the beginning of the first page.
3041 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3043 if (bp->b_flags & B_VMIO) {
3048 * Fixup base to be relative to beginning of first page.
3049 * Set initial n to be the maximum number of bytes in the
3050 * first page that can be validated.
3053 base += (bp->b_offset & PAGE_MASK);
3054 n = PAGE_SIZE - (base & PAGE_MASK);
3056 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3057 vm_page_t m = bp->b_pages[i];
3062 vm_page_set_validclean(m, base & PAGE_MASK, n);
3073 * clear a buffer. This routine essentially fakes an I/O, so we need
3074 * to clear B_ERROR and B_INVAL.
3076 * Note that while we only theoretically need to clear through b_bcount,
3077 * we go ahead and clear through b_bufsize.
3081 vfs_bio_clrbuf(struct buf *bp)
3085 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3086 bp->b_flags &= ~(B_INVAL|B_ERROR);
3087 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3088 (bp->b_offset & PAGE_MASK) == 0) {
3089 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3090 if ((bp->b_pages[0]->valid & mask) == mask) {
3094 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3095 ((bp->b_pages[0]->valid & mask) == 0)) {
3096 bzero(bp->b_data, bp->b_bufsize);
3097 bp->b_pages[0]->valid |= mask;
3102 ea = sa = bp->b_data;
3103 for(i=0;i<bp->b_npages;i++,sa=ea) {
3104 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3105 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3106 ea = (caddr_t)(vm_offset_t)ulmin(
3107 (u_long)(vm_offset_t)ea,
3108 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3109 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3110 if ((bp->b_pages[i]->valid & mask) == mask)
3112 if ((bp->b_pages[i]->valid & mask) == 0) {
3113 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) {
3117 for (; sa < ea; sa += DEV_BSIZE, j++) {
3118 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3119 (bp->b_pages[i]->valid & (1<<j)) == 0)
3120 bzero(sa, DEV_BSIZE);
3123 bp->b_pages[i]->valid |= mask;
3124 vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
3133 * vm_hold_load_pages and vm_hold_unload pages get pages into
3134 * a buffers address space. The pages are anonymous and are
3135 * not associated with a file object.
3138 vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3144 to = round_page(to);
3145 from = round_page(from);
3146 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3148 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3153 * note: must allocate system pages since blocking here
3154 * could intefere with paging I/O, no matter which
3157 p = vm_page_alloc(kernel_object,
3158 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3161 vm_pageout_deficit += (to - from) >> PAGE_SHIFT;
3166 p->valid = VM_PAGE_BITS_ALL;
3167 vm_page_flag_clear(p, PG_ZERO);
3168 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3169 bp->b_pages[index] = p;
3172 bp->b_npages = index;
3176 vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3180 int index, newnpages;
3182 from = round_page(from);
3183 to = round_page(to);
3184 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3186 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3187 p = bp->b_pages[index];
3188 if (p && (index < bp->b_npages)) {
3190 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n",
3191 bp->b_blkno, bp->b_lblkno);
3193 bp->b_pages[index] = NULL;
3196 vm_page_unwire(p, 0);
3200 bp->b_npages = newnpages;
3204 * Map an IO request into kernel virtual address space.
3206 * All requests are (re)mapped into kernel VA space.
3207 * Notice that we use b_bufsize for the size of the buffer
3208 * to be mapped. b_bcount might be modified by the driver.
3211 vmapbuf(struct buf *bp)
3213 caddr_t addr, v, kva;
3219 if ((bp->b_flags & B_PHYS) == 0)
3221 if (bp->b_bufsize < 0)
3223 for (v = bp->b_saveaddr,
3224 addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data),
3226 addr < bp->b_data + bp->b_bufsize;
3227 addr += PAGE_SIZE, v += PAGE_SIZE, pidx++) {
3229 * Do the vm_fault if needed; do the copy-on-write thing
3230 * when reading stuff off device into memory.
3233 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data,
3234 (bp->b_flags&B_READ)?(VM_PROT_READ|VM_PROT_WRITE):VM_PROT_READ);
3236 for (i = 0; i < pidx; ++i) {
3237 vm_page_unhold(bp->b_pages[i]);
3238 bp->b_pages[i] = NULL;
3244 * WARNING! If sparc support is MFCd in the future this will
3245 * have to be changed from pmap_kextract() to pmap_extract()
3249 #error "If MFCing sparc support use pmap_extract"
3251 pa = pmap_kextract((vm_offset_t)addr);
3253 printf("vmapbuf: warning, race against user address during I/O");
3256 m = PHYS_TO_VM_PAGE(pa);
3258 bp->b_pages[pidx] = m;
3260 if (pidx > btoc(MAXPHYS))
3261 panic("vmapbuf: mapped more than MAXPHYS");
3262 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3264 kva = bp->b_saveaddr;
3265 bp->b_npages = pidx;
3266 bp->b_saveaddr = bp->b_data;
3267 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3272 * Free the io map PTEs associated with this IO operation.
3273 * We also invalidate the TLB entries and restore the original b_addr.
3277 register struct buf *bp;
3283 if ((bp->b_flags & B_PHYS) == 0)
3286 npages = bp->b_npages;
3287 pmap_qremove(trunc_page((vm_offset_t)bp->b_data),
3290 for (pidx = 0; pidx < npages; pidx++)
3291 vm_page_unhold(*m++);
3293 bp->b_data = bp->b_saveaddr;
3296 #include "opt_ddb.h"
3298 #include <ddb/ddb.h>
3300 DB_SHOW_COMMAND(buffer, db_show_buffer)
3303 struct buf *bp = (struct buf *)addr;
3306 db_printf("usage: show buffer <addr>\n");
3310 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3311 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, "
3312 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, "
3313 "b_blkno = %d, b_pblkno = %d\n",
3314 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3315 major(bp->b_dev), minor(bp->b_dev),
3316 bp->b_data, bp->b_blkno, bp->b_pblkno);
3319 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3320 for (i = 0; i < bp->b_npages; i++) {
3323 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3324 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3325 if ((i + 1) < bp->b_npages)