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 1.20 2004/03/11 20:14:46 hmp 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>
57 #include <vm/vm_page2.h>
59 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
61 struct bio_ops bioops; /* I/O operation notification */
63 struct buf *buf; /* buffer header pool */
64 struct swqueue bswlist;
66 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from,
68 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from,
70 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
71 int pageno, vm_page_t m);
72 static void vfs_clean_pages(struct buf * bp);
73 static void vfs_setdirty(struct buf *bp);
74 static void vfs_vmio_release(struct buf *bp);
75 static void vfs_backgroundwritedone(struct buf *bp);
76 static int flushbufqueues(void);
78 static int bd_request;
80 static void buf_daemon (void);
82 * bogus page -- for I/O to/from partially complete buffers
83 * this is a temporary solution to the problem, but it is not
84 * really that bad. it would be better to split the buffer
85 * for input in the case of buffers partially already in memory,
86 * but the code is intricate enough already.
89 int vmiodirenable = TRUE;
91 struct lwkt_token buftimetoken; /* Interlock on setting prio and timo */
93 static vm_offset_t bogus_offset;
95 static int bufspace, maxbufspace,
96 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
97 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
98 static int needsbuffer;
99 static int lorunningspace, hirunningspace, runningbufreq;
100 static int numdirtybuffers, lodirtybuffers, hidirtybuffers;
101 static int numfreebuffers, lofreebuffers, hifreebuffers;
102 static int getnewbufcalls;
103 static int getnewbufrestarts;
105 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD,
106 &numdirtybuffers, 0, "");
107 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW,
108 &lodirtybuffers, 0, "");
109 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW,
110 &hidirtybuffers, 0, "");
111 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD,
112 &numfreebuffers, 0, "");
113 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW,
114 &lofreebuffers, 0, "");
115 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW,
116 &hifreebuffers, 0, "");
117 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD,
118 &runningbufspace, 0, "");
119 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW,
120 &lorunningspace, 0, "");
121 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW,
122 &hirunningspace, 0, "");
123 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD,
124 &maxbufspace, 0, "");
125 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD,
127 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD,
129 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD,
131 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW,
132 &maxbufmallocspace, 0, "");
133 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD,
134 &bufmallocspace, 0, "");
135 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RW,
136 &getnewbufcalls, 0, "");
137 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RW,
138 &getnewbufrestarts, 0, "");
139 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW,
140 &vmiodirenable, 0, "");
141 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW,
142 &bufdefragcnt, 0, "");
143 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW,
144 &buffreekvacnt, 0, "");
145 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RW,
146 &bufreusecnt, 0, "");
149 * Disable background writes for now. There appear to be races in the
150 * flags tests and locking operations as well as races in the completion
151 * code modifying the original bp (origbp) without holding a lock, assuming
152 * splbio protection when there might not be splbio protection.
154 static int dobkgrdwrite = 0;
155 SYSCTL_INT(_debug, OID_AUTO, dobkgrdwrite, CTLFLAG_RW, &dobkgrdwrite, 0,
156 "Do background writes (honoring the BV_BKGRDWRITE flag)?");
158 static int bufhashmask;
159 static LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
160 struct bqueues bufqueues[BUFFER_QUEUES] = { { 0 } };
161 char *buf_wmesg = BUF_WMESG;
163 extern int vm_swap_size;
165 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
166 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
167 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
168 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
171 * Buffer hash table code. Note that the logical block scans linearly, which
172 * gives us some L1 cache locality.
177 bufhash(struct vnode *vnp, daddr_t bn)
179 return(&bufhashtbl[(((uintptr_t)(vnp) >> 7) + (int)bn) & bufhashmask]);
185 * If someone is blocked due to there being too many dirty buffers,
186 * and numdirtybuffers is now reasonable, wake them up.
190 numdirtywakeup(int level)
192 if (numdirtybuffers <= level) {
193 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
194 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
195 wakeup(&needsbuffer);
203 * Called when buffer space is potentially available for recovery.
204 * getnewbuf() will block on this flag when it is unable to free
205 * sufficient buffer space. Buffer space becomes recoverable when
206 * bp's get placed back in the queues.
213 * If someone is waiting for BUF space, wake them up. Even
214 * though we haven't freed the kva space yet, the waiting
215 * process will be able to now.
217 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
218 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
219 wakeup(&needsbuffer);
224 * runningbufwakeup() - in-progress I/O accounting.
228 runningbufwakeup(struct buf *bp)
230 if (bp->b_runningbufspace) {
231 runningbufspace -= bp->b_runningbufspace;
232 bp->b_runningbufspace = 0;
233 if (runningbufreq && runningbufspace <= lorunningspace) {
235 wakeup(&runningbufreq);
243 * Called when a buffer has been added to one of the free queues to
244 * account for the buffer and to wakeup anyone waiting for free buffers.
245 * This typically occurs when large amounts of metadata are being handled
246 * by the buffer cache ( else buffer space runs out first, usually ).
254 needsbuffer &= ~VFS_BIO_NEED_ANY;
255 if (numfreebuffers >= hifreebuffers)
256 needsbuffer &= ~VFS_BIO_NEED_FREE;
257 wakeup(&needsbuffer);
262 * waitrunningbufspace()
264 * runningbufspace is a measure of the amount of I/O currently
265 * running. This routine is used in async-write situations to
266 * prevent creating huge backups of pending writes to a device.
267 * Only asynchronous writes are governed by this function.
269 * Reads will adjust runningbufspace, but will not block based on it.
270 * The read load has a side effect of reducing the allowed write load.
272 * This does NOT turn an async write into a sync write. It waits
273 * for earlier writes to complete and generally returns before the
274 * caller's write has reached the device.
277 waitrunningbufspace(void)
279 while (runningbufspace > hirunningspace) {
282 s = splbio(); /* fix race against interrupt/biodone() */
284 tsleep(&runningbufreq, 0, "wdrain", 0);
290 * vfs_buf_test_cache:
292 * Called when a buffer is extended. This function clears the B_CACHE
293 * bit if the newly extended portion of the buffer does not contain
298 vfs_buf_test_cache(struct buf *bp,
299 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
302 if (bp->b_flags & B_CACHE) {
303 int base = (foff + off) & PAGE_MASK;
304 if (vm_page_is_valid(m, base, size) == 0)
305 bp->b_flags &= ~B_CACHE;
311 bd_wakeup(int dirtybuflevel)
313 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
320 * bd_speedup - speedup the buffer cache flushing code
331 * Initialize buffer headers and related structures.
335 bufhashinit(caddr_t vaddr)
337 /* first, make a null hash table */
338 for (bufhashmask = 8; bufhashmask < nbuf / 4; bufhashmask <<= 1)
340 bufhashtbl = (void *)vaddr;
341 vaddr = vaddr + sizeof(*bufhashtbl) * bufhashmask;
352 TAILQ_INIT(&bswlist);
353 LIST_INIT(&invalhash);
354 lwkt_token_init(&buftimetoken);
356 for (i = 0; i <= bufhashmask; i++)
357 LIST_INIT(&bufhashtbl[i]);
359 /* next, make a null set of free lists */
360 for (i = 0; i < BUFFER_QUEUES; i++)
361 TAILQ_INIT(&bufqueues[i]);
363 /* finally, initialize each buffer header and stick on empty q */
364 for (i = 0; i < nbuf; i++) {
366 bzero(bp, sizeof *bp);
367 bp->b_flags = B_INVAL; /* we're just an empty header */
369 bp->b_qindex = QUEUE_EMPTY;
371 LIST_INIT(&bp->b_dep);
373 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist);
374 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
378 * maxbufspace is the absolute maximum amount of buffer space we are
379 * allowed to reserve in KVM and in real terms. The absolute maximum
380 * is nominally used by buf_daemon. hibufspace is the nominal maximum
381 * used by most other processes. The differential is required to
382 * ensure that buf_daemon is able to run when other processes might
383 * be blocked waiting for buffer space.
385 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
386 * this may result in KVM fragmentation which is not handled optimally
389 maxbufspace = nbuf * BKVASIZE;
390 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
391 lobufspace = hibufspace - MAXBSIZE;
393 lorunningspace = 512 * 1024;
394 hirunningspace = 1024 * 1024;
397 * Limit the amount of malloc memory since it is wired permanently into
398 * the kernel space. Even though this is accounted for in the buffer
399 * allocation, we don't want the malloced region to grow uncontrolled.
400 * The malloc scheme improves memory utilization significantly on average
401 * (small) directories.
403 maxbufmallocspace = hibufspace / 20;
406 * Reduce the chance of a deadlock occuring by limiting the number
407 * of delayed-write dirty buffers we allow to stack up.
409 hidirtybuffers = nbuf / 4 + 20;
412 * To support extreme low-memory systems, make sure hidirtybuffers cannot
413 * eat up all available buffer space. This occurs when our minimum cannot
414 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
415 * BKVASIZE'd (8K) buffers.
417 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
418 hidirtybuffers >>= 1;
420 lodirtybuffers = hidirtybuffers / 2;
423 * Try to keep the number of free buffers in the specified range,
424 * and give special processes (e.g. like buf_daemon) access to an
427 lofreebuffers = nbuf / 18 + 5;
428 hifreebuffers = 2 * lofreebuffers;
429 numfreebuffers = nbuf;
432 * Maximum number of async ops initiated per buf_daemon loop. This is
433 * somewhat of a hack at the moment, we really need to limit ourselves
434 * based on the number of bytes of I/O in-transit that were initiated
438 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE);
439 bogus_page = vm_page_alloc(kernel_object,
440 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
442 vmstats.v_wire_count++;
447 * bfreekva() - free the kva allocation for a buffer.
449 * Must be called at splbio() or higher as this is the only locking for
452 * Since this call frees up buffer space, we call bufspacewakeup().
455 bfreekva(struct buf * bp)
461 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
462 vm_map_lock(buffer_map);
463 bufspace -= bp->b_kvasize;
464 vm_map_delete(buffer_map,
465 (vm_offset_t) bp->b_kvabase,
466 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
469 vm_map_unlock(buffer_map);
470 vm_map_entry_release(count);
479 * Remove the buffer from the appropriate free list.
482 bremfree(struct buf * bp)
485 int old_qindex = bp->b_qindex;
487 if (bp->b_qindex != QUEUE_NONE) {
488 KASSERT(BUF_REFCNT(bp) == 1, ("bremfree: bp %p not locked",bp));
489 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
490 bp->b_qindex = QUEUE_NONE;
492 if (BUF_REFCNT(bp) <= 1)
493 panic("bremfree: removing a buffer not on a queue");
497 * Fixup numfreebuffers count. If the buffer is invalid or not
498 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
499 * the buffer was free and we must decrement numfreebuffers.
501 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
518 * Get a buffer with the specified data. Look in the cache first. We
519 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
520 * is set, the buffer is valid and we do not have to do anything ( see
524 bread(struct vnode * vp, daddr_t blkno, int size, struct buf ** bpp)
528 bp = getblk(vp, blkno, size, 0, 0);
531 /* if not found in cache, do some I/O */
532 if ((bp->b_flags & B_CACHE) == 0) {
533 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp));
534 bp->b_flags |= B_READ;
535 bp->b_flags &= ~(B_ERROR | B_INVAL);
536 vfs_busy_pages(bp, 0);
537 VOP_STRATEGY(vp, bp);
538 return (biowait(bp));
544 * Operates like bread, but also starts asynchronous I/O on
545 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
546 * to initiating I/O . If B_CACHE is set, the buffer is valid
547 * and we do not have to do anything.
550 breadn(struct vnode * vp, daddr_t blkno, int size, daddr_t * rablkno,
551 int *rabsize, int cnt, struct buf ** bpp)
553 struct buf *bp, *rabp;
555 int rv = 0, readwait = 0;
557 *bpp = bp = getblk(vp, blkno, size, 0, 0);
559 /* if not found in cache, do some I/O */
560 if ((bp->b_flags & B_CACHE) == 0) {
561 bp->b_flags |= B_READ;
562 bp->b_flags &= ~(B_ERROR | B_INVAL);
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) {
574 rabp->b_flags |= B_READ | B_ASYNC;
575 rabp->b_flags &= ~(B_ERROR | B_INVAL);
576 vfs_busy_pages(rabp, 0);
578 VOP_STRATEGY(vp, rabp);
591 * Write, release buffer on completion. (Done by iodone
592 * if async). Do not bother writing anything if the buffer
595 * Note that we set B_CACHE here, indicating that buffer is
596 * fully valid and thus cacheable. This is true even of NFS
597 * now so we set it generally. This could be set either here
598 * or in biodone() since the I/O is synchronous. We put it
602 bwrite(struct buf * bp)
607 if (bp->b_flags & B_INVAL) {
612 oldflags = bp->b_flags;
614 if (BUF_REFCNT(bp) == 0)
615 panic("bwrite: buffer is not busy???");
618 * If a background write is already in progress, delay
619 * writing this block if it is asynchronous. Otherwise
620 * wait for the background write to complete.
622 if (bp->b_xflags & BX_BKGRDINPROG) {
623 if (bp->b_flags & B_ASYNC) {
628 bp->b_xflags |= BX_BKGRDWAIT;
629 tsleep(&bp->b_xflags, 0, "biord", 0);
630 if (bp->b_xflags & BX_BKGRDINPROG)
631 panic("bwrite: still writing");
634 /* Mark the buffer clean */
638 * If this buffer is marked for background writing and we
639 * do not have to wait for it, make a copy and write the
640 * copy so as to leave this buffer ready for further use.
642 * This optimization eats a lot of memory. If we have a page
643 * or buffer shortfull we can't do it.
646 (bp->b_xflags & BX_BKGRDWRITE) &&
647 (bp->b_flags & B_ASYNC) &&
648 !vm_page_count_severe() &&
649 !buf_dirty_count_severe()) {
650 if (bp->b_flags & B_CALL)
651 panic("bwrite: need chained iodone");
653 /* get a new block */
654 newbp = geteblk(bp->b_bufsize);
656 /* set it to be identical to the old block */
657 memcpy(newbp->b_data, bp->b_data, bp->b_bufsize);
658 bgetvp(bp->b_vp, newbp);
659 newbp->b_lblkno = bp->b_lblkno;
660 newbp->b_blkno = bp->b_blkno;
661 newbp->b_offset = bp->b_offset;
662 newbp->b_iodone = vfs_backgroundwritedone;
663 newbp->b_flags |= B_ASYNC | B_CALL;
664 newbp->b_flags &= ~B_INVAL;
666 /* move over the dependencies */
667 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps)
668 (*bioops.io_movedeps)(bp, newbp);
671 * Initiate write on the copy, release the original to
672 * the B_LOCKED queue so that it cannot go away until
673 * the background write completes. If not locked it could go
674 * away and then be reconstituted while it was being written.
675 * If the reconstituted buffer were written, we could end up
676 * with two background copies being written at the same time.
678 bp->b_xflags |= BX_BKGRDINPROG;
679 bp->b_flags |= B_LOCKED;
684 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
685 bp->b_flags |= B_WRITEINPROG | B_CACHE;
687 bp->b_vp->v_numoutput++;
688 vfs_busy_pages(bp, 1);
691 * Normal bwrites pipeline writes
693 bp->b_runningbufspace = bp->b_bufsize;
694 runningbufspace += bp->b_runningbufspace;
697 if (oldflags & B_ASYNC)
699 VOP_STRATEGY(bp->b_vp, bp);
701 if ((oldflags & B_ASYNC) == 0) {
702 int rtval = biowait(bp);
705 } else if ((oldflags & B_NOWDRAIN) == 0) {
707 * don't allow the async write to saturate the I/O
708 * system. Deadlocks can occur only if a device strategy
709 * routine (like in VN) turns around and issues another
710 * high-level write, in which case B_NOWDRAIN is expected
711 * to be set. Otherwise we will not deadlock here because
712 * we are blocking waiting for I/O that is already in-progress
715 waitrunningbufspace();
722 * Complete a background write started from bwrite.
725 vfs_backgroundwritedone(bp)
731 * Find the original buffer that we are writing.
733 if ((origbp = gbincore(bp->b_vp, bp->b_lblkno)) == NULL)
734 panic("backgroundwritedone: lost buffer");
736 * Process dependencies then return any unfinished ones.
738 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
739 (*bioops.io_complete)(bp);
740 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_movedeps)
741 (*bioops.io_movedeps)(bp, origbp);
743 * Clear the BX_BKGRDINPROG flag in the original buffer
744 * and awaken it if it is waiting for the write to complete.
745 * If BX_BKGRDINPROG is not set in the original buffer it must
746 * have been released and re-instantiated - which is not legal.
748 KASSERT((origbp->b_xflags & BX_BKGRDINPROG), ("backgroundwritedone: lost buffer2"));
749 origbp->b_xflags &= ~BX_BKGRDINPROG;
750 if (origbp->b_xflags & BX_BKGRDWAIT) {
751 origbp->b_xflags &= ~BX_BKGRDWAIT;
752 wakeup(&origbp->b_xflags);
755 * Clear the B_LOCKED flag and remove it from the locked
756 * queue if it currently resides there.
758 origbp->b_flags &= ~B_LOCKED;
759 if (BUF_LOCK(origbp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
764 * This buffer is marked B_NOCACHE, so when it is released
765 * by biodone, it will be tossed. We mark it with B_READ
766 * to avoid biodone doing a second vwakeup.
768 bp->b_flags |= B_NOCACHE | B_READ;
769 bp->b_flags &= ~(B_CACHE | B_CALL | B_DONE);
775 * Delayed write. (Buffer is marked dirty). Do not bother writing
776 * anything if the buffer is marked invalid.
778 * Note that since the buffer must be completely valid, we can safely
779 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
780 * biodone() in order to prevent getblk from writing the buffer
784 bdwrite(struct buf * bp)
786 if (BUF_REFCNT(bp) == 0)
787 panic("bdwrite: buffer is not busy");
789 if (bp->b_flags & B_INVAL) {
796 * Set B_CACHE, indicating that the buffer is fully valid. This is
797 * true even of NFS now.
799 bp->b_flags |= B_CACHE;
802 * This bmap keeps the system from needing to do the bmap later,
803 * perhaps when the system is attempting to do a sync. Since it
804 * is likely that the indirect block -- or whatever other datastructure
805 * that the filesystem needs is still in memory now, it is a good
806 * thing to do this. Note also, that if the pageout daemon is
807 * requesting a sync -- there might not be enough memory to do
808 * the bmap then... So, this is important to do.
810 if (bp->b_lblkno == bp->b_blkno) {
811 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
815 * Set the *dirty* buffer range based upon the VM system dirty pages.
820 * We need to do this here to satisfy the vnode_pager and the
821 * pageout daemon, so that it thinks that the pages have been
822 * "cleaned". Note that since the pages are in a delayed write
823 * buffer -- the VFS layer "will" see that the pages get written
824 * out on the next sync, or perhaps the cluster will be completed.
830 * Wakeup the buffer flushing daemon if we have a lot of dirty
831 * buffers (midpoint between our recovery point and our stall
834 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
837 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
838 * due to the softdep code.
845 * Turn buffer into delayed write request. We must clear B_READ and
846 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to
847 * itself to properly update it in the dirty/clean lists. We mark it
848 * B_DONE to ensure that any asynchronization of the buffer properly
849 * clears B_DONE ( else a panic will occur later ).
851 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which
852 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty()
853 * should only be called if the buffer is known-good.
855 * Since the buffer is not on a queue, we do not update the numfreebuffers
858 * Must be called at splbio().
859 * The buffer must be on QUEUE_NONE.
865 KASSERT(bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
866 bp->b_flags &= ~(B_READ|B_RELBUF);
868 if ((bp->b_flags & B_DELWRI) == 0) {
869 bp->b_flags |= B_DONE | B_DELWRI;
870 reassignbuf(bp, bp->b_vp);
872 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
879 * Clear B_DELWRI for buffer.
881 * Since the buffer is not on a queue, we do not update the numfreebuffers
884 * Must be called at splbio().
885 * The buffer must be on QUEUE_NONE.
892 KASSERT(bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex));
894 if (bp->b_flags & B_DELWRI) {
895 bp->b_flags &= ~B_DELWRI;
896 reassignbuf(bp, bp->b_vp);
898 numdirtywakeup(lodirtybuffers);
901 * Since it is now being written, we can clear its deferred write flag.
903 bp->b_flags &= ~B_DEFERRED;
909 * Asynchronous write. Start output on a buffer, but do not wait for
910 * it to complete. The buffer is released when the output completes.
912 * bwrite() ( or the VOP routine anyway ) is responsible for handling
913 * B_INVAL buffers. Not us.
916 bawrite(struct buf * bp)
918 bp->b_flags |= B_ASYNC;
919 (void) VOP_BWRITE(bp->b_vp, bp);
925 * Ordered write. Start output on a buffer, and flag it so that the
926 * device will write it in the order it was queued. The buffer is
927 * released when the output completes. bwrite() ( or the VOP routine
928 * anyway ) is responsible for handling B_INVAL buffers.
931 bowrite(struct buf * bp)
933 bp->b_flags |= B_ORDERED | B_ASYNC;
934 return (VOP_BWRITE(bp->b_vp, bp));
940 * Called prior to the locking of any vnodes when we are expecting to
941 * write. We do not want to starve the buffer cache with too many
942 * dirty buffers so we block here. By blocking prior to the locking
943 * of any vnodes we attempt to avoid the situation where a locked vnode
944 * prevents the various system daemons from flushing related buffers.
950 if (numdirtybuffers >= hidirtybuffers) {
954 while (numdirtybuffers >= hidirtybuffers) {
956 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
957 tsleep(&needsbuffer, 0, "flswai", 0);
964 * Return true if we have too many dirty buffers.
967 buf_dirty_count_severe(void)
969 return(numdirtybuffers >= hidirtybuffers);
975 * Release a busy buffer and, if requested, free its resources. The
976 * buffer will be stashed in the appropriate bufqueue[] allowing it
977 * to be accessed later as a cache entity or reused for other purposes.
980 brelse(struct buf * bp)
984 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
988 if (bp->b_flags & B_LOCKED)
989 bp->b_flags &= ~B_ERROR;
991 if ((bp->b_flags & (B_READ | B_ERROR | B_INVAL)) == B_ERROR) {
993 * Failed write, redirty. Must clear B_ERROR to prevent
994 * pages from being scrapped. If B_INVAL is set then
995 * this case is not run and the next case is run to
996 * destroy the buffer. B_INVAL can occur if the buffer
997 * is outside the range supported by the underlying device.
999 bp->b_flags &= ~B_ERROR;
1001 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) ||
1002 (bp->b_bufsize <= 0)) {
1004 * Either a failed I/O or we were asked to free or not
1007 bp->b_flags |= B_INVAL;
1008 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
1009 (*bioops.io_deallocate)(bp);
1010 if (bp->b_flags & B_DELWRI) {
1012 numdirtywakeup(lodirtybuffers);
1014 bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF);
1015 if ((bp->b_flags & B_VMIO) == 0) {
1024 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1025 * is called with B_DELWRI set, the underlying pages may wind up
1026 * getting freed causing a previous write (bdwrite()) to get 'lost'
1027 * because pages associated with a B_DELWRI bp are marked clean.
1029 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1030 * if B_DELWRI is set.
1032 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1033 * on pages to return pages to the VM page queues.
1035 if (bp->b_flags & B_DELWRI)
1036 bp->b_flags &= ~B_RELBUF;
1037 else if (vm_page_count_severe() && !(bp->b_xflags & BX_BKGRDINPROG))
1038 bp->b_flags |= B_RELBUF;
1041 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer
1042 * constituted, not even NFS buffers now. Two flags effect this. If
1043 * B_INVAL, the struct buf is invalidated but the VM object is kept
1044 * around ( i.e. so it is trivial to reconstitute the buffer later ).
1046 * If B_ERROR or B_NOCACHE is set, pages in the VM object will be
1047 * invalidated. B_ERROR cannot be set for a failed write unless the
1048 * buffer is also B_INVAL because it hits the re-dirtying code above.
1050 * Normally we can do this whether a buffer is B_DELWRI or not. If
1051 * the buffer is an NFS buffer, it is tracking piecemeal writes or
1052 * the commit state and we cannot afford to lose the buffer. If the
1053 * buffer has a background write in progress, we need to keep it
1054 * around to prevent it from being reconstituted and starting a second
1057 if ((bp->b_flags & B_VMIO)
1058 && !(bp->b_vp->v_tag == VT_NFS &&
1059 !vn_isdisk(bp->b_vp, NULL) &&
1060 (bp->b_flags & B_DELWRI))
1073 * Get the base offset and length of the buffer. Note that
1074 * in the VMIO case if the buffer block size is not
1075 * page-aligned then b_data pointer may not be page-aligned.
1076 * But our b_pages[] array *IS* page aligned.
1078 * block sizes less then DEV_BSIZE (usually 512) are not
1079 * supported due to the page granularity bits (m->valid,
1080 * m->dirty, etc...).
1082 * See man buf(9) for more information
1085 resid = bp->b_bufsize;
1086 foff = bp->b_offset;
1088 for (i = 0; i < bp->b_npages; i++) {
1090 vm_page_flag_clear(m, PG_ZERO);
1092 * If we hit a bogus page, fixup *all* of them
1095 if (m == bogus_page) {
1096 VOP_GETVOBJECT(vp, &obj);
1097 poff = OFF_TO_IDX(bp->b_offset);
1099 for (j = i; j < bp->b_npages; j++) {
1102 mtmp = bp->b_pages[j];
1103 if (mtmp == bogus_page) {
1104 mtmp = vm_page_lookup(obj, poff + j);
1106 panic("brelse: page missing\n");
1108 bp->b_pages[j] = mtmp;
1112 if ((bp->b_flags & B_INVAL) == 0) {
1113 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
1117 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1118 int poffset = foff & PAGE_MASK;
1119 int presid = resid > (PAGE_SIZE - poffset) ?
1120 (PAGE_SIZE - poffset) : resid;
1122 KASSERT(presid >= 0, ("brelse: extra page"));
1123 vm_page_set_invalid(m, poffset, presid);
1125 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1126 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1129 if (bp->b_flags & (B_INVAL | B_RELBUF))
1130 vfs_vmio_release(bp);
1132 } else if (bp->b_flags & B_VMIO) {
1134 if (bp->b_flags & (B_INVAL | B_RELBUF))
1135 vfs_vmio_release(bp);
1139 if (bp->b_qindex != QUEUE_NONE)
1140 panic("brelse: free buffer onto another queue???");
1141 if (BUF_REFCNT(bp) > 1) {
1142 /* Temporary panic to verify exclusive locking */
1143 /* This panic goes away when we allow shared refs */
1144 panic("brelse: multiple refs");
1145 /* do not release to free list */
1153 /* buffers with no memory */
1154 if (bp->b_bufsize == 0) {
1155 bp->b_flags |= B_INVAL;
1156 bp->b_xflags &= ~BX_BKGRDWRITE;
1157 if (bp->b_xflags & BX_BKGRDINPROG)
1158 panic("losing buffer 1");
1159 if (bp->b_kvasize) {
1160 bp->b_qindex = QUEUE_EMPTYKVA;
1162 bp->b_qindex = QUEUE_EMPTY;
1164 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1165 LIST_REMOVE(bp, b_hash);
1166 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1168 /* buffers with junk contents */
1169 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) {
1170 bp->b_flags |= B_INVAL;
1171 bp->b_xflags &= ~BX_BKGRDWRITE;
1172 if (bp->b_xflags & BX_BKGRDINPROG)
1173 panic("losing buffer 2");
1174 bp->b_qindex = QUEUE_CLEAN;
1175 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1176 LIST_REMOVE(bp, b_hash);
1177 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1180 /* buffers that are locked */
1181 } else if (bp->b_flags & B_LOCKED) {
1182 bp->b_qindex = QUEUE_LOCKED;
1183 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1185 /* remaining buffers */
1187 switch(bp->b_flags & (B_DELWRI|B_AGE)) {
1188 case B_DELWRI | B_AGE:
1189 bp->b_qindex = QUEUE_DIRTY;
1190 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1193 bp->b_qindex = QUEUE_DIRTY;
1194 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1197 bp->b_qindex = QUEUE_CLEAN;
1198 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1201 bp->b_qindex = QUEUE_CLEAN;
1202 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1208 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1209 * on the correct queue.
1211 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1215 * Fixup numfreebuffers count. The bp is on an appropriate queue
1216 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1217 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1218 * if B_INVAL is set ).
1221 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI))
1225 * Something we can maybe free or reuse
1227 if (bp->b_bufsize || bp->b_kvasize)
1232 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF |
1233 B_DIRECT | B_NOWDRAIN);
1238 * Release a buffer back to the appropriate queue but do not try to free
1239 * it. The buffer is expected to be used again soon.
1241 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1242 * biodone() to requeue an async I/O on completion. It is also used when
1243 * known good buffers need to be requeued but we think we may need the data
1246 * XXX we should be able to leave the B_RELBUF hint set on completion.
1249 bqrelse(struct buf * bp)
1255 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1257 if (bp->b_qindex != QUEUE_NONE)
1258 panic("bqrelse: free buffer onto another queue???");
1259 if (BUF_REFCNT(bp) > 1) {
1260 /* do not release to free list */
1261 panic("bqrelse: multiple refs");
1266 if (bp->b_flags & B_LOCKED) {
1267 bp->b_flags &= ~B_ERROR;
1268 bp->b_qindex = QUEUE_LOCKED;
1269 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist);
1270 /* buffers with stale but valid contents */
1271 } else if (bp->b_flags & B_DELWRI) {
1272 bp->b_qindex = QUEUE_DIRTY;
1273 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY], bp, b_freelist);
1274 } else if (vm_page_count_severe()) {
1276 * We are too low on memory, we have to try to free the
1277 * buffer (most importantly: the wired pages making up its
1278 * backing store) *now*.
1284 bp->b_qindex = QUEUE_CLEAN;
1285 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_CLEAN], bp, b_freelist);
1288 if ((bp->b_flags & B_LOCKED) == 0 &&
1289 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) {
1294 * Something we can maybe free or reuse.
1296 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1301 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1306 vfs_vmio_release(bp)
1313 for (i = 0; i < bp->b_npages; i++) {
1315 bp->b_pages[i] = NULL;
1317 * In order to keep page LRU ordering consistent, put
1318 * everything on the inactive queue.
1320 vm_page_unwire(m, 0);
1322 * We don't mess with busy pages, it is
1323 * the responsibility of the process that
1324 * busied the pages to deal with them.
1326 if ((m->flags & PG_BUSY) || (m->busy != 0))
1329 if (m->wire_count == 0) {
1330 vm_page_flag_clear(m, PG_ZERO);
1332 * Might as well free the page if we can and it has
1333 * no valid data. We also free the page if the
1334 * buffer was used for direct I/O.
1336 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) {
1338 vm_page_protect(m, VM_PROT_NONE);
1340 } else if (bp->b_flags & B_DIRECT) {
1341 vm_page_try_to_free(m);
1342 } else if (vm_page_count_severe()) {
1343 vm_page_try_to_cache(m);
1348 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
1349 if (bp->b_bufsize) {
1354 bp->b_flags &= ~B_VMIO;
1360 * Check to see if a block is currently memory resident.
1363 gbincore(struct vnode * vp, daddr_t blkno)
1366 struct bufhashhdr *bh;
1368 bh = bufhash(vp, blkno);
1370 /* Search hash chain */
1371 LIST_FOREACH(bp, bh, b_hash) {
1373 if (bp->b_vp == vp && bp->b_lblkno == blkno &&
1374 (bp->b_flags & B_INVAL) == 0) {
1384 * Implement clustered async writes for clearing out B_DELWRI buffers.
1385 * This is much better then the old way of writing only one buffer at
1386 * a time. Note that we may not be presented with the buffers in the
1387 * correct order, so we search for the cluster in both directions.
1390 vfs_bio_awrite(struct buf * bp)
1394 daddr_t lblkno = bp->b_lblkno;
1395 struct vnode *vp = bp->b_vp;
1405 * right now we support clustered writing only to regular files. If
1406 * we find a clusterable block we could be in the middle of a cluster
1407 * rather then at the beginning.
1409 if ((vp->v_type == VREG) &&
1410 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1411 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1413 size = vp->v_mount->mnt_stat.f_iosize;
1414 maxcl = MAXPHYS / size;
1416 for (i = 1; i < maxcl; i++) {
1417 if ((bpa = gbincore(vp, lblkno + i)) &&
1418 BUF_REFCNT(bpa) == 0 &&
1419 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1420 (B_DELWRI | B_CLUSTEROK)) &&
1421 (bpa->b_bufsize == size)) {
1422 if ((bpa->b_blkno == bpa->b_lblkno) ||
1424 bp->b_blkno + ((i * size) >> DEV_BSHIFT)))
1430 for (j = 1; i + j <= maxcl && j <= lblkno; j++) {
1431 if ((bpa = gbincore(vp, lblkno - j)) &&
1432 BUF_REFCNT(bpa) == 0 &&
1433 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1434 (B_DELWRI | B_CLUSTEROK)) &&
1435 (bpa->b_bufsize == size)) {
1436 if ((bpa->b_blkno == bpa->b_lblkno) ||
1438 bp->b_blkno - ((j * size) >> DEV_BSHIFT)))
1447 * this is a possible cluster write
1450 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl);
1456 BUF_LOCK(bp, LK_EXCLUSIVE);
1458 bp->b_flags |= B_ASYNC;
1462 * default (old) behavior, writing out only one block
1464 * XXX returns b_bufsize instead of b_bcount for nwritten?
1466 nwritten = bp->b_bufsize;
1467 (void) VOP_BWRITE(bp->b_vp, bp);
1475 * Find and initialize a new buffer header, freeing up existing buffers
1476 * in the bufqueues as necessary. The new buffer is returned locked.
1478 * Important: B_INVAL is not set. If the caller wishes to throw the
1479 * buffer away, the caller must set B_INVAL prior to calling brelse().
1482 * We have insufficient buffer headers
1483 * We have insufficient buffer space
1484 * buffer_map is too fragmented ( space reservation fails )
1485 * If we have to flush dirty buffers ( but we try to avoid this )
1487 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1488 * Instead we ask the buf daemon to do it for us. We attempt to
1489 * avoid piecemeal wakeups of the pageout daemon.
1493 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1499 static int flushingbufs;
1502 * We can't afford to block since we might be holding a vnode lock,
1503 * which may prevent system daemons from running. We deal with
1504 * low-memory situations by proactively returning memory and running
1505 * async I/O rather then sync I/O.
1509 --getnewbufrestarts;
1511 ++getnewbufrestarts;
1514 * Setup for scan. If we do not have enough free buffers,
1515 * we setup a degenerate case that immediately fails. Note
1516 * that if we are specially marked process, we are allowed to
1517 * dip into our reserves.
1519 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1521 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1522 * However, there are a number of cases (defragging, reusing, ...)
1523 * where we cannot backup.
1525 nqindex = QUEUE_EMPTYKVA;
1526 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA]);
1530 * If no EMPTYKVA buffers and we are either
1531 * defragging or reusing, locate a CLEAN buffer
1532 * to free or reuse. If bufspace useage is low
1533 * skip this step so we can allocate a new buffer.
1535 if (defrag || bufspace >= lobufspace) {
1536 nqindex = QUEUE_CLEAN;
1537 nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN]);
1541 * If we could not find or were not allowed to reuse a
1542 * CLEAN buffer, check to see if it is ok to use an EMPTY
1543 * buffer. We can only use an EMPTY buffer if allocating
1544 * its KVA would not otherwise run us out of buffer space.
1546 if (nbp == NULL && defrag == 0 &&
1547 bufspace + maxsize < hibufspace) {
1548 nqindex = QUEUE_EMPTY;
1549 nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
1554 * Run scan, possibly freeing data and/or kva mappings on the fly
1558 while ((bp = nbp) != NULL) {
1559 int qindex = nqindex;
1562 * Calculate next bp ( we can only use it if we do not block
1563 * or do other fancy things ).
1565 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1568 nqindex = QUEUE_EMPTYKVA;
1569 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTYKVA])))
1572 case QUEUE_EMPTYKVA:
1573 nqindex = QUEUE_CLEAN;
1574 if ((nbp = TAILQ_FIRST(&bufqueues[QUEUE_CLEAN])))
1588 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1591 * Note: we no longer distinguish between VMIO and non-VMIO
1595 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1598 * If we are defragging then we need a buffer with
1599 * b_kvasize != 0. XXX this situation should no longer
1600 * occur, if defrag is non-zero the buffer's b_kvasize
1601 * should also be non-zero at this point. XXX
1603 if (defrag && bp->b_kvasize == 0) {
1604 printf("Warning: defrag empty buffer %p\n", bp);
1609 * Start freeing the bp. This is somewhat involved. nbp
1610 * remains valid only for QUEUE_EMPTY[KVA] bp's.
1613 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1614 panic("getnewbuf: locked buf");
1617 if (qindex == QUEUE_CLEAN) {
1618 if (bp->b_flags & B_VMIO) {
1619 bp->b_flags &= ~B_ASYNC;
1620 vfs_vmio_release(bp);
1627 * NOTE: nbp is now entirely invalid. We can only restart
1628 * the scan from this point on.
1630 * Get the rest of the buffer freed up. b_kva* is still
1631 * valid after this operation.
1634 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
1635 (*bioops.io_deallocate)(bp);
1636 if (bp->b_xflags & BX_BKGRDINPROG)
1637 panic("losing buffer 3");
1638 LIST_REMOVE(bp, b_hash);
1639 LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1648 bp->b_blkno = bp->b_lblkno = 0;
1649 bp->b_offset = NOOFFSET;
1655 bp->b_dirtyoff = bp->b_dirtyend = 0;
1657 LIST_INIT(&bp->b_dep);
1660 * If we are defragging then free the buffer.
1663 bp->b_flags |= B_INVAL;
1671 * If we are overcomitted then recover the buffer and its
1672 * KVM space. This occurs in rare situations when multiple
1673 * processes are blocked in getnewbuf() or allocbuf().
1675 if (bufspace >= hibufspace)
1677 if (flushingbufs && bp->b_kvasize != 0) {
1678 bp->b_flags |= B_INVAL;
1683 if (bufspace < lobufspace)
1689 * If we exhausted our list, sleep as appropriate. We may have to
1690 * wakeup various daemons and write out some dirty buffers.
1692 * Generally we are sleeping due to insufficient buffer space.
1700 flags = VFS_BIO_NEED_BUFSPACE;
1702 } else if (bufspace >= hibufspace) {
1704 flags = VFS_BIO_NEED_BUFSPACE;
1707 flags = VFS_BIO_NEED_ANY;
1710 bd_speedup(); /* heeeelp */
1712 needsbuffer |= flags;
1713 while (needsbuffer & flags) {
1714 if (tsleep(&needsbuffer, slpflag, waitmsg, slptimeo))
1719 * We finally have a valid bp. We aren't quite out of the
1720 * woods, we still have to reserve kva space. In order
1721 * to keep fragmentation sane we only allocate kva in
1724 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1726 if (maxsize != bp->b_kvasize) {
1727 vm_offset_t addr = 0;
1732 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1733 vm_map_lock(buffer_map);
1735 if (vm_map_findspace(buffer_map,
1736 vm_map_min(buffer_map), maxsize,
1739 * Uh oh. Buffer map is to fragmented. We
1740 * must defragment the map.
1742 vm_map_unlock(buffer_map);
1743 vm_map_entry_release(count);
1746 bp->b_flags |= B_INVAL;
1751 vm_map_insert(buffer_map, &count,
1753 addr, addr + maxsize,
1754 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
1756 bp->b_kvabase = (caddr_t) addr;
1757 bp->b_kvasize = maxsize;
1758 bufspace += bp->b_kvasize;
1761 vm_map_unlock(buffer_map);
1762 vm_map_entry_release(count);
1764 bp->b_data = bp->b_kvabase;
1772 * buffer flushing daemon. Buffers are normally flushed by the
1773 * update daemon but if it cannot keep up this process starts to
1774 * take the load in an attempt to prevent getnewbuf() from blocking.
1777 static struct thread *bufdaemonthread;
1779 static struct kproc_desc buf_kp = {
1784 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
1792 * This process needs to be suspended prior to shutdown sync.
1794 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
1795 bufdaemonthread, SHUTDOWN_PRI_LAST);
1798 * This process is allowed to take the buffer cache to the limit
1803 kproc_suspend_loop();
1806 * Do the flush. Limit the amount of in-transit I/O we
1807 * allow to build up, otherwise we would completely saturate
1808 * the I/O system. Wakeup any waiting processes before we
1809 * normally would so they can run in parallel with our drain.
1811 while (numdirtybuffers > lodirtybuffers) {
1812 if (flushbufqueues() == 0)
1814 waitrunningbufspace();
1815 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
1819 * Only clear bd_request if we have reached our low water
1820 * mark. The buf_daemon normally waits 5 seconds and
1821 * then incrementally flushes any dirty buffers that have
1822 * built up, within reason.
1824 * If we were unable to hit our low water mark and couldn't
1825 * find any flushable buffers, we sleep half a second.
1826 * Otherwise we loop immediately.
1828 if (numdirtybuffers <= lodirtybuffers) {
1830 * We reached our low water mark, reset the
1831 * request and sleep until we are needed again.
1832 * The sleep is just so the suspend code works.
1835 tsleep(&bd_request, 0, "psleep", hz);
1838 * We couldn't find any flushable dirty buffers but
1839 * still have too many dirty buffers, we
1840 * have to sleep and try again. (rare)
1842 tsleep(&bd_request, 0, "qsleep", hz / 2);
1850 * Try to flush a buffer in the dirty queue. We must be careful to
1851 * free up B_INVAL buffers instead of write them, which NFS is
1852 * particularly sensitive to.
1856 flushbufqueues(void)
1861 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
1864 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp));
1865 if ((bp->b_flags & B_DELWRI) != 0 &&
1866 (bp->b_xflags & BX_BKGRDINPROG) == 0) {
1867 if (bp->b_flags & B_INVAL) {
1868 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1869 panic("flushbufqueues: locked buf");
1875 if (LIST_FIRST(&bp->b_dep) != NULL &&
1876 bioops.io_countdeps &&
1877 (bp->b_flags & B_DEFERRED) == 0 &&
1878 (*bioops.io_countdeps)(bp, 0)) {
1879 TAILQ_REMOVE(&bufqueues[QUEUE_DIRTY],
1881 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_DIRTY],
1883 bp->b_flags |= B_DEFERRED;
1884 bp = TAILQ_FIRST(&bufqueues[QUEUE_DIRTY]);
1891 bp = TAILQ_NEXT(bp, b_freelist);
1897 * Check to see if a block is currently memory resident.
1900 incore(struct vnode * vp, daddr_t blkno)
1905 bp = gbincore(vp, blkno);
1911 * Returns true if no I/O is needed to access the
1912 * associated VM object. This is like incore except
1913 * it also hunts around in the VM system for the data.
1917 inmem(struct vnode * vp, daddr_t blkno)
1920 vm_offset_t toff, tinc, size;
1924 if (incore(vp, blkno))
1926 if (vp->v_mount == NULL)
1928 if (VOP_GETVOBJECT(vp, &obj) != 0 || (vp->v_flag & VOBJBUF) == 0)
1932 if (size > vp->v_mount->mnt_stat.f_iosize)
1933 size = vp->v_mount->mnt_stat.f_iosize;
1934 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize;
1936 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
1937 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff));
1941 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK))
1942 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK);
1943 if (vm_page_is_valid(m,
1944 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0)
1953 * Sets the dirty range for a buffer based on the status of the dirty
1954 * bits in the pages comprising the buffer.
1956 * The range is limited to the size of the buffer.
1958 * This routine is primarily used by NFS, but is generalized for the
1962 vfs_setdirty(struct buf *bp)
1968 * Degenerate case - empty buffer
1971 if (bp->b_bufsize == 0)
1975 * We qualify the scan for modified pages on whether the
1976 * object has been flushed yet. The OBJ_WRITEABLE flag
1977 * is not cleared simply by protecting pages off.
1980 if ((bp->b_flags & B_VMIO) == 0)
1983 object = bp->b_pages[0]->object;
1985 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
1986 printf("Warning: object %p writeable but not mightbedirty\n", object);
1987 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
1988 printf("Warning: object %p mightbedirty but not writeable\n", object);
1990 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
1991 vm_offset_t boffset;
1992 vm_offset_t eoffset;
1995 * test the pages to see if they have been modified directly
1996 * by users through the VM system.
1998 for (i = 0; i < bp->b_npages; i++) {
1999 vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
2000 vm_page_test_dirty(bp->b_pages[i]);
2004 * Calculate the encompassing dirty range, boffset and eoffset,
2005 * (eoffset - boffset) bytes.
2008 for (i = 0; i < bp->b_npages; i++) {
2009 if (bp->b_pages[i]->dirty)
2012 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2014 for (i = bp->b_npages - 1; i >= 0; --i) {
2015 if (bp->b_pages[i]->dirty) {
2019 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK);
2022 * Fit it to the buffer.
2025 if (eoffset > bp->b_bcount)
2026 eoffset = bp->b_bcount;
2029 * If we have a good dirty range, merge with the existing
2033 if (boffset < eoffset) {
2034 if (bp->b_dirtyoff > boffset)
2035 bp->b_dirtyoff = boffset;
2036 if (bp->b_dirtyend < eoffset)
2037 bp->b_dirtyend = eoffset;
2045 * Get a block given a specified block and offset into a file/device.
2046 * The buffers B_DONE bit will be cleared on return, making it almost
2047 * ready for an I/O initiation. B_INVAL may or may not be set on
2048 * return. The caller should clear B_INVAL prior to initiating a
2051 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2052 * an existing buffer.
2054 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2055 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2056 * and then cleared based on the backing VM. If the previous buffer is
2057 * non-0-sized but invalid, B_CACHE will be cleared.
2059 * If getblk() must create a new buffer, the new buffer is returned with
2060 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2061 * case it is returned with B_INVAL clear and B_CACHE set based on the
2064 * getblk() also forces a VOP_BWRITE() for any B_DELWRI buffer whos
2065 * B_CACHE bit is clear.
2067 * What this means, basically, is that the caller should use B_CACHE to
2068 * determine whether the buffer is fully valid or not and should clear
2069 * B_INVAL prior to issuing a read. If the caller intends to validate
2070 * the buffer by loading its data area with something, the caller needs
2071 * to clear B_INVAL. If the caller does this without issuing an I/O,
2072 * the caller should set B_CACHE ( as an optimization ), else the caller
2073 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2074 * a write attempt or if it was a successfull read. If the caller
2075 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2076 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2079 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo)
2083 struct bufhashhdr *bh;
2085 if (size > MAXBSIZE)
2086 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE);
2091 * Block if we are low on buffers. Certain processes are allowed
2092 * to completely exhaust the buffer cache.
2094 * If this check ever becomes a bottleneck it may be better to
2095 * move it into the else, when gbincore() fails. At the moment
2096 * it isn't a problem.
2098 * XXX remove, we cannot afford to block anywhere if holding a vnode
2099 * lock in low-memory situation, so take it to the max.
2101 if (numfreebuffers == 0) {
2104 needsbuffer |= VFS_BIO_NEED_ANY;
2105 tsleep(&needsbuffer, slpflag, "newbuf", slptimeo);
2108 if ((bp = gbincore(vp, blkno))) {
2110 * Buffer is in-core. If the buffer is not busy, it must
2114 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2115 if (BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL,
2116 "getblk", slpflag, slptimeo) == ENOLCK)
2119 return (struct buf *) NULL;
2123 * The buffer is locked. B_CACHE is cleared if the buffer is
2124 * invalid. Ohterwise, for a non-VMIO buffer, B_CACHE is set
2125 * and for a VMIO buffer B_CACHE is adjusted according to the
2128 if (bp->b_flags & B_INVAL)
2129 bp->b_flags &= ~B_CACHE;
2130 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0)
2131 bp->b_flags |= B_CACHE;
2135 * check for size inconsistancies for non-VMIO case.
2138 if (bp->b_bcount != size) {
2139 if ((bp->b_flags & B_VMIO) == 0 ||
2140 (size > bp->b_kvasize)) {
2141 if (bp->b_flags & B_DELWRI) {
2142 bp->b_flags |= B_NOCACHE;
2143 VOP_BWRITE(bp->b_vp, bp);
2145 if ((bp->b_flags & B_VMIO) &&
2146 (LIST_FIRST(&bp->b_dep) == NULL)) {
2147 bp->b_flags |= B_RELBUF;
2150 bp->b_flags |= B_NOCACHE;
2151 VOP_BWRITE(bp->b_vp, bp);
2159 * If the size is inconsistant in the VMIO case, we can resize
2160 * the buffer. This might lead to B_CACHE getting set or
2161 * cleared. If the size has not changed, B_CACHE remains
2162 * unchanged from its previous state.
2165 if (bp->b_bcount != size)
2168 KASSERT(bp->b_offset != NOOFFSET,
2169 ("getblk: no buffer offset"));
2172 * A buffer with B_DELWRI set and B_CACHE clear must
2173 * be committed before we can return the buffer in
2174 * order to prevent the caller from issuing a read
2175 * ( due to B_CACHE not being set ) and overwriting
2178 * Most callers, including NFS and FFS, need this to
2179 * operate properly either because they assume they
2180 * can issue a read if B_CACHE is not set, or because
2181 * ( for example ) an uncached B_DELWRI might loop due
2182 * to softupdates re-dirtying the buffer. In the latter
2183 * case, B_CACHE is set after the first write completes,
2184 * preventing further loops.
2186 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2187 * above while extending the buffer, we cannot allow the
2188 * buffer to remain with B_CACHE set after the write
2189 * completes or it will represent a corrupt state. To
2190 * deal with this we set B_NOCACHE to scrap the buffer
2193 * We might be able to do something fancy, like setting
2194 * B_CACHE in bwrite() except if B_DELWRI is already set,
2195 * so the below call doesn't set B_CACHE, but that gets real
2196 * confusing. This is much easier.
2199 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2200 bp->b_flags |= B_NOCACHE;
2201 VOP_BWRITE(bp->b_vp, bp);
2206 bp->b_flags &= ~B_DONE;
2209 * Buffer is not in-core, create new buffer. The buffer
2210 * returned by getnewbuf() is locked. Note that the returned
2211 * buffer is also considered valid (not marked B_INVAL).
2213 int bsize, maxsize, vmio;
2216 if (vn_isdisk(vp, NULL))
2218 else if (vp->v_mountedhere)
2219 bsize = vp->v_mountedhere->mnt_stat.f_iosize;
2220 else if (vp->v_mount)
2221 bsize = vp->v_mount->mnt_stat.f_iosize;
2225 offset = (off_t)blkno * bsize;
2226 vmio = (VOP_GETVOBJECT(vp, NULL) == 0) && (vp->v_flag & VOBJBUF);
2227 maxsize = vmio ? size + (offset & PAGE_MASK) : size;
2228 maxsize = imax(maxsize, bsize);
2230 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) {
2231 if (slpflag || slptimeo) {
2239 * This code is used to make sure that a buffer is not
2240 * created while the getnewbuf routine is blocked.
2241 * This can be a problem whether the vnode is locked or not.
2242 * If the buffer is created out from under us, we have to
2243 * throw away the one we just created. There is now window
2244 * race because we are safely running at splbio() from the
2245 * point of the duplicate buffer creation through to here,
2246 * and we've locked the buffer.
2248 if (gbincore(vp, blkno)) {
2249 bp->b_flags |= B_INVAL;
2255 * Insert the buffer into the hash, so that it can
2256 * be found by incore.
2258 bp->b_blkno = bp->b_lblkno = blkno;
2259 bp->b_offset = offset;
2262 LIST_REMOVE(bp, b_hash);
2263 bh = bufhash(vp, blkno);
2264 LIST_INSERT_HEAD(bh, bp, b_hash);
2267 * set B_VMIO bit. allocbuf() the buffer bigger. Since the
2268 * buffer size starts out as 0, B_CACHE will be set by
2269 * allocbuf() for the VMIO case prior to it testing the
2270 * backing store for validity.
2274 bp->b_flags |= B_VMIO;
2275 #if defined(VFS_BIO_DEBUG)
2276 if (vn_canvmio(vp) != TRUE)
2277 printf("getblk: vmioing file type %d???\n", vp->v_type);
2280 bp->b_flags &= ~B_VMIO;
2286 bp->b_flags &= ~B_DONE;
2292 * Get an empty, disassociated buffer of given size. The buffer is initially
2302 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2305 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0);
2308 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2314 * This code constitutes the buffer memory from either anonymous system
2315 * memory (in the case of non-VMIO operations) or from an associated
2316 * VM object (in the case of VMIO operations). This code is able to
2317 * resize a buffer up or down.
2319 * Note that this code is tricky, and has many complications to resolve
2320 * deadlock or inconsistant data situations. Tread lightly!!!
2321 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2322 * the caller. Calling this code willy nilly can result in the loss of data.
2324 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2325 * B_CACHE for the non-VMIO case.
2329 allocbuf(struct buf *bp, int size)
2331 int newbsize, mbsize;
2334 if (BUF_REFCNT(bp) == 0)
2335 panic("allocbuf: buffer not busy");
2337 if (bp->b_kvasize < size)
2338 panic("allocbuf: buffer too small");
2340 if ((bp->b_flags & B_VMIO) == 0) {
2344 * Just get anonymous memory from the kernel. Don't
2345 * mess with B_CACHE.
2347 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2348 #if !defined(NO_B_MALLOC)
2349 if (bp->b_flags & B_MALLOC)
2353 newbsize = round_page(size);
2355 if (newbsize < bp->b_bufsize) {
2356 #if !defined(NO_B_MALLOC)
2358 * malloced buffers are not shrunk
2360 if (bp->b_flags & B_MALLOC) {
2362 bp->b_bcount = size;
2364 free(bp->b_data, M_BIOBUF);
2365 if (bp->b_bufsize) {
2366 bufmallocspace -= bp->b_bufsize;
2370 bp->b_data = bp->b_kvabase;
2372 bp->b_flags &= ~B_MALLOC;
2379 (vm_offset_t) bp->b_data + newbsize,
2380 (vm_offset_t) bp->b_data + bp->b_bufsize);
2381 } else if (newbsize > bp->b_bufsize) {
2382 #if !defined(NO_B_MALLOC)
2384 * We only use malloced memory on the first allocation.
2385 * and revert to page-allocated memory when the buffer
2388 if ( (bufmallocspace < maxbufmallocspace) &&
2389 (bp->b_bufsize == 0) &&
2390 (mbsize <= PAGE_SIZE/2)) {
2392 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2393 bp->b_bufsize = mbsize;
2394 bp->b_bcount = size;
2395 bp->b_flags |= B_MALLOC;
2396 bufmallocspace += mbsize;
2402 #if !defined(NO_B_MALLOC)
2404 * If the buffer is growing on its other-than-first allocation,
2405 * then we revert to the page-allocation scheme.
2407 if (bp->b_flags & B_MALLOC) {
2408 origbuf = bp->b_data;
2409 origbufsize = bp->b_bufsize;
2410 bp->b_data = bp->b_kvabase;
2411 if (bp->b_bufsize) {
2412 bufmallocspace -= bp->b_bufsize;
2416 bp->b_flags &= ~B_MALLOC;
2417 newbsize = round_page(newbsize);
2422 (vm_offset_t) bp->b_data + bp->b_bufsize,
2423 (vm_offset_t) bp->b_data + newbsize);
2424 #if !defined(NO_B_MALLOC)
2426 bcopy(origbuf, bp->b_data, origbufsize);
2427 free(origbuf, M_BIOBUF);
2435 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2436 desiredpages = (size == 0) ? 0 :
2437 num_pages((bp->b_offset & PAGE_MASK) + newbsize);
2439 #if !defined(NO_B_MALLOC)
2440 if (bp->b_flags & B_MALLOC)
2441 panic("allocbuf: VMIO buffer can't be malloced");
2444 * Set B_CACHE initially if buffer is 0 length or will become
2447 if (size == 0 || bp->b_bufsize == 0)
2448 bp->b_flags |= B_CACHE;
2450 if (newbsize < bp->b_bufsize) {
2452 * DEV_BSIZE aligned new buffer size is less then the
2453 * DEV_BSIZE aligned existing buffer size. Figure out
2454 * if we have to remove any pages.
2456 if (desiredpages < bp->b_npages) {
2457 for (i = desiredpages; i < bp->b_npages; i++) {
2459 * the page is not freed here -- it
2460 * is the responsibility of
2461 * vnode_pager_setsize
2464 KASSERT(m != bogus_page,
2465 ("allocbuf: bogus page found"));
2466 while (vm_page_sleep_busy(m, TRUE, "biodep"))
2469 bp->b_pages[i] = NULL;
2470 vm_page_unwire(m, 0);
2472 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2473 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages));
2474 bp->b_npages = desiredpages;
2476 } else if (size > bp->b_bcount) {
2478 * We are growing the buffer, possibly in a
2479 * byte-granular fashion.
2487 * Step 1, bring in the VM pages from the object,
2488 * allocating them if necessary. We must clear
2489 * B_CACHE if these pages are not valid for the
2490 * range covered by the buffer.
2494 VOP_GETVOBJECT(vp, &obj);
2496 while (bp->b_npages < desiredpages) {
2500 pi = OFF_TO_IDX(bp->b_offset) + bp->b_npages;
2501 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2503 * note: must allocate system pages
2504 * since blocking here could intefere
2505 * with paging I/O, no matter which
2508 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
2511 vm_pageout_deficit += desiredpages - bp->b_npages;
2515 bp->b_flags &= ~B_CACHE;
2516 bp->b_pages[bp->b_npages] = m;
2523 * We found a page. If we have to sleep on it,
2524 * retry because it might have gotten freed out
2527 * We can only test PG_BUSY here. Blocking on
2528 * m->busy might lead to a deadlock:
2530 * vm_fault->getpages->cluster_read->allocbuf
2534 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
2538 * We have a good page. Should we wakeup the
2541 if ((curthread != pagethread) &&
2542 ((m->queue - m->pc) == PQ_CACHE) &&
2543 ((vmstats.v_free_count + vmstats.v_cache_count) <
2544 (vmstats.v_free_min + vmstats.v_cache_min))) {
2545 pagedaemon_wakeup();
2547 vm_page_flag_clear(m, PG_ZERO);
2549 bp->b_pages[bp->b_npages] = m;
2554 * Step 2. We've loaded the pages into the buffer,
2555 * we have to figure out if we can still have B_CACHE
2556 * set. Note that B_CACHE is set according to the
2557 * byte-granular range ( bcount and size ), new the
2558 * aligned range ( newbsize ).
2560 * The VM test is against m->valid, which is DEV_BSIZE
2561 * aligned. Needless to say, the validity of the data
2562 * needs to also be DEV_BSIZE aligned. Note that this
2563 * fails with NFS if the server or some other client
2564 * extends the file's EOF. If our buffer is resized,
2565 * B_CACHE may remain set! XXX
2568 toff = bp->b_bcount;
2569 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK);
2571 while ((bp->b_flags & B_CACHE) && toff < size) {
2574 if (tinc > (size - toff))
2577 pi = ((bp->b_offset & PAGE_MASK) + toff) >>
2592 * Step 3, fixup the KVM pmap. Remember that
2593 * bp->b_data is relative to bp->b_offset, but
2594 * bp->b_offset may be offset into the first page.
2597 bp->b_data = (caddr_t)
2598 trunc_page((vm_offset_t)bp->b_data);
2600 (vm_offset_t)bp->b_data,
2604 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2605 (vm_offset_t)(bp->b_offset & PAGE_MASK));
2608 if (newbsize < bp->b_bufsize)
2610 bp->b_bufsize = newbsize; /* actual buffer allocation */
2611 bp->b_bcount = size; /* requested buffer size */
2618 * Wait for buffer I/O completion, returning error status. The buffer
2619 * is left locked and B_DONE on return. B_EINTR is converted into a EINTR
2620 * error and cleared.
2623 biowait(struct buf * bp)
2628 while ((bp->b_flags & B_DONE) == 0) {
2629 #if defined(NO_SCHEDULE_MODS)
2630 tsleep(bp, 0, "biowait", 0);
2632 if (bp->b_flags & B_READ)
2633 tsleep(bp, 0, "biord", 0);
2635 tsleep(bp, 0, "biowr", 0);
2639 if (bp->b_flags & B_EINTR) {
2640 bp->b_flags &= ~B_EINTR;
2643 if (bp->b_flags & B_ERROR) {
2644 return (bp->b_error ? bp->b_error : EIO);
2653 * Finish I/O on a buffer, optionally calling a completion function.
2654 * This is usually called from an interrupt so process blocking is
2657 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
2658 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
2659 * assuming B_INVAL is clear.
2661 * For the VMIO case, we set B_CACHE if the op was a read and no
2662 * read error occured, or if the op was a write. B_CACHE is never
2663 * set if the buffer is invalid or otherwise uncacheable.
2665 * biodone does not mess with B_INVAL, allowing the I/O routine or the
2666 * initiator to leave B_INVAL set to brelse the buffer out of existance
2667 * in the biodone routine.
2670 biodone(struct buf * bp)
2676 KASSERT(BUF_REFCNT(bp) > 0, ("biodone: bp %p not busy %d", bp, BUF_REFCNT(bp)));
2677 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp));
2679 bp->b_flags |= B_DONE;
2680 runningbufwakeup(bp);
2682 if (bp->b_flags & B_FREEBUF) {
2688 if ((bp->b_flags & B_READ) == 0) {
2692 /* call optional completion function if requested */
2693 if (bp->b_flags & B_CALL) {
2694 bp->b_flags &= ~B_CALL;
2695 (*bp->b_iodone) (bp);
2699 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
2700 (*bioops.io_complete)(bp);
2702 if (bp->b_flags & B_VMIO) {
2708 struct vnode *vp = bp->b_vp;
2710 error = VOP_GETVOBJECT(vp, &obj);
2712 #if defined(VFS_BIO_DEBUG)
2713 if (vp->v_usecount == 0) {
2714 panic("biodone: zero vnode ref count");
2718 panic("biodone: missing VM object");
2721 if ((vp->v_flag & VOBJBUF) == 0) {
2722 panic("biodone: vnode is not setup for merged cache");
2726 foff = bp->b_offset;
2727 KASSERT(bp->b_offset != NOOFFSET,
2728 ("biodone: no buffer offset"));
2731 panic("biodone: no object");
2733 #if defined(VFS_BIO_DEBUG)
2734 if (obj->paging_in_progress < bp->b_npages) {
2735 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n",
2736 obj->paging_in_progress, bp->b_npages);
2741 * Set B_CACHE if the op was a normal read and no error
2742 * occured. B_CACHE is set for writes in the b*write()
2745 iosize = bp->b_bcount - bp->b_resid;
2746 if ((bp->b_flags & (B_READ|B_FREEBUF|B_INVAL|B_NOCACHE|B_ERROR)) == B_READ) {
2747 bp->b_flags |= B_CACHE;
2750 for (i = 0; i < bp->b_npages; i++) {
2754 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2759 * cleanup bogus pages, restoring the originals
2762 if (m == bogus_page) {
2764 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2766 panic("biodone: page disappeared");
2768 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2770 #if defined(VFS_BIO_DEBUG)
2771 if (OFF_TO_IDX(foff) != m->pindex) {
2773 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
2774 (unsigned long)foff, m->pindex);
2779 * In the write case, the valid and clean bits are
2780 * already changed correctly ( see bdwrite() ), so we
2781 * only need to do this here in the read case.
2783 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) {
2784 vfs_page_set_valid(bp, foff, i, m);
2786 vm_page_flag_clear(m, PG_ZERO);
2789 * when debugging new filesystems or buffer I/O methods, this
2790 * is the most common error that pops up. if you see this, you
2791 * have not set the page busy flag correctly!!!
2794 printf("biodone: page busy < 0, "
2795 "pindex: %d, foff: 0x(%x,%x), "
2796 "resid: %d, index: %d\n",
2797 (int) m->pindex, (int)(foff >> 32),
2798 (int) foff & 0xffffffff, resid, i);
2799 if (!vn_isdisk(vp, NULL))
2800 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n",
2801 bp->b_vp->v_mount->mnt_stat.f_iosize,
2803 bp->b_flags, bp->b_npages);
2805 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n",
2807 bp->b_flags, bp->b_npages);
2808 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
2809 m->valid, m->dirty, m->wire_count);
2810 panic("biodone: page busy < 0\n");
2812 vm_page_io_finish(m);
2813 vm_object_pip_subtract(obj, 1);
2814 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2818 vm_object_pip_wakeupn(obj, 0);
2822 * For asynchronous completions, release the buffer now. The brelse
2823 * will do a wakeup there if necessary - so no need to do a wakeup
2824 * here in the async case. The sync case always needs to do a wakeup.
2827 if (bp->b_flags & B_ASYNC) {
2828 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
2839 * This routine is called in lieu of iodone in the case of
2840 * incomplete I/O. This keeps the busy status for pages
2844 vfs_unbusy_pages(struct buf * bp)
2848 runningbufwakeup(bp);
2849 if (bp->b_flags & B_VMIO) {
2850 struct vnode *vp = bp->b_vp;
2853 VOP_GETVOBJECT(vp, &obj);
2855 for (i = 0; i < bp->b_npages; i++) {
2856 vm_page_t m = bp->b_pages[i];
2858 if (m == bogus_page) {
2859 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i);
2861 panic("vfs_unbusy_pages: page missing\n");
2864 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2866 vm_object_pip_subtract(obj, 1);
2867 vm_page_flag_clear(m, PG_ZERO);
2868 vm_page_io_finish(m);
2870 vm_object_pip_wakeupn(obj, 0);
2875 * vfs_page_set_valid:
2877 * Set the valid bits in a page based on the supplied offset. The
2878 * range is restricted to the buffer's size.
2880 * This routine is typically called after a read completes.
2883 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
2885 vm_ooffset_t soff, eoff;
2888 * Start and end offsets in buffer. eoff - soff may not cross a
2889 * page boundry or cross the end of the buffer. The end of the
2890 * buffer, in this case, is our file EOF, not the allocation size
2894 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2895 if (eoff > bp->b_offset + bp->b_bcount)
2896 eoff = bp->b_offset + bp->b_bcount;
2899 * Set valid range. This is typically the entire buffer and thus the
2903 vm_page_set_validclean(
2905 (vm_offset_t) (soff & PAGE_MASK),
2906 (vm_offset_t) (eoff - soff)
2912 * This routine is called before a device strategy routine.
2913 * It is used to tell the VM system that paging I/O is in
2914 * progress, and treat the pages associated with the buffer
2915 * almost as being PG_BUSY. Also the object paging_in_progress
2916 * flag is handled to make sure that the object doesn't become
2919 * Since I/O has not been initiated yet, certain buffer flags
2920 * such as B_ERROR or B_INVAL may be in an inconsistant state
2921 * and should be ignored.
2924 vfs_busy_pages(struct buf * bp, int clear_modify)
2928 if (bp->b_flags & B_VMIO) {
2929 struct vnode *vp = bp->b_vp;
2933 VOP_GETVOBJECT(vp, &obj);
2934 foff = bp->b_offset;
2935 KASSERT(bp->b_offset != NOOFFSET,
2936 ("vfs_busy_pages: no buffer offset"));
2940 for (i = 0; i < bp->b_npages; i++) {
2941 vm_page_t m = bp->b_pages[i];
2942 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
2947 for (i = 0; i < bp->b_npages; i++) {
2948 vm_page_t m = bp->b_pages[i];
2950 vm_page_flag_clear(m, PG_ZERO);
2951 if ((bp->b_flags & B_CLUSTER) == 0) {
2952 vm_object_pip_add(obj, 1);
2953 vm_page_io_start(m);
2957 * When readying a buffer for a read ( i.e
2958 * clear_modify == 0 ), it is important to do
2959 * bogus_page replacement for valid pages in
2960 * partially instantiated buffers. Partially
2961 * instantiated buffers can, in turn, occur when
2962 * reconstituting a buffer from its VM backing store
2963 * base. We only have to do this if B_CACHE is
2964 * clear ( which causes the I/O to occur in the
2965 * first place ). The replacement prevents the read
2966 * I/O from overwriting potentially dirty VM-backed
2967 * pages. XXX bogus page replacement is, uh, bogus.
2968 * It may not work properly with small-block devices.
2969 * We need to find a better way.
2972 vm_page_protect(m, VM_PROT_NONE);
2974 vfs_page_set_valid(bp, foff, i, m);
2975 else if (m->valid == VM_PAGE_BITS_ALL &&
2976 (bp->b_flags & B_CACHE) == 0) {
2977 bp->b_pages[i] = bogus_page;
2980 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2983 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages);
2987 * This is the easiest place to put the process accounting for the I/O
2993 if ((p = curthread->td_proc) != NULL) {
2994 if (bp->b_flags & B_READ)
2995 p->p_stats->p_ru.ru_inblock++;
2997 p->p_stats->p_ru.ru_oublock++;
3003 * Tell the VM system that the pages associated with this buffer
3004 * are clean. This is used for delayed writes where the data is
3005 * going to go to disk eventually without additional VM intevention.
3007 * Note that while we only really need to clean through to b_bcount, we
3008 * just go ahead and clean through to b_bufsize.
3011 vfs_clean_pages(struct buf * bp)
3015 if (bp->b_flags & B_VMIO) {
3018 foff = bp->b_offset;
3019 KASSERT(bp->b_offset != NOOFFSET,
3020 ("vfs_clean_pages: no buffer offset"));
3021 for (i = 0; i < bp->b_npages; i++) {
3022 vm_page_t m = bp->b_pages[i];
3023 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3024 vm_ooffset_t eoff = noff;
3026 if (eoff > bp->b_offset + bp->b_bufsize)
3027 eoff = bp->b_offset + bp->b_bufsize;
3028 vfs_page_set_valid(bp, foff, i, m);
3029 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3036 * vfs_bio_set_validclean:
3038 * Set the range within the buffer to valid and clean. The range is
3039 * relative to the beginning of the buffer, b_offset. Note that b_offset
3040 * itself may be offset from the beginning of the first page.
3044 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3046 if (bp->b_flags & B_VMIO) {
3051 * Fixup base to be relative to beginning of first page.
3052 * Set initial n to be the maximum number of bytes in the
3053 * first page that can be validated.
3056 base += (bp->b_offset & PAGE_MASK);
3057 n = PAGE_SIZE - (base & PAGE_MASK);
3059 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) {
3060 vm_page_t m = bp->b_pages[i];
3065 vm_page_set_validclean(m, base & PAGE_MASK, n);
3076 * clear a buffer. This routine essentially fakes an I/O, so we need
3077 * to clear B_ERROR and B_INVAL.
3079 * Note that while we only theoretically need to clear through b_bcount,
3080 * we go ahead and clear through b_bufsize.
3084 vfs_bio_clrbuf(struct buf *bp)
3088 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3089 bp->b_flags &= ~(B_INVAL|B_ERROR);
3090 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3091 (bp->b_offset & PAGE_MASK) == 0) {
3092 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3093 if ((bp->b_pages[0]->valid & mask) == mask) {
3097 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) &&
3098 ((bp->b_pages[0]->valid & mask) == 0)) {
3099 bzero(bp->b_data, bp->b_bufsize);
3100 bp->b_pages[0]->valid |= mask;
3105 ea = sa = bp->b_data;
3106 for(i=0;i<bp->b_npages;i++,sa=ea) {
3107 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3108 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3109 ea = (caddr_t)(vm_offset_t)ulmin(
3110 (u_long)(vm_offset_t)ea,
3111 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3112 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3113 if ((bp->b_pages[i]->valid & mask) == mask)
3115 if ((bp->b_pages[i]->valid & mask) == 0) {
3116 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) {
3120 for (; sa < ea; sa += DEV_BSIZE, j++) {
3121 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) &&
3122 (bp->b_pages[i]->valid & (1<<j)) == 0)
3123 bzero(sa, DEV_BSIZE);
3126 bp->b_pages[i]->valid |= mask;
3127 vm_page_flag_clear(bp->b_pages[i], PG_ZERO);
3136 * vm_hold_load_pages and vm_hold_unload pages get pages into
3137 * a buffers address space. The pages are anonymous and are
3138 * not associated with a file object.
3141 vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3147 to = round_page(to);
3148 from = round_page(from);
3149 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3151 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3156 * note: must allocate system pages since blocking here
3157 * could intefere with paging I/O, no matter which
3160 p = vm_page_alloc(kernel_object,
3161 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3162 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
3164 vm_pageout_deficit += (to - from) >> PAGE_SHIFT;
3169 p->valid = VM_PAGE_BITS_ALL;
3170 vm_page_flag_clear(p, PG_ZERO);
3171 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3172 bp->b_pages[index] = p;
3175 bp->b_npages = index;
3179 vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to)
3183 int index, newnpages;
3185 from = round_page(from);
3186 to = round_page(to);
3187 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3189 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3190 p = bp->b_pages[index];
3191 if (p && (index < bp->b_npages)) {
3193 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n",
3194 bp->b_blkno, bp->b_lblkno);
3196 bp->b_pages[index] = NULL;
3199 vm_page_unwire(p, 0);
3203 bp->b_npages = newnpages;
3207 * Map an IO request into kernel virtual address space.
3209 * All requests are (re)mapped into kernel VA space.
3210 * Notice that we use b_bufsize for the size of the buffer
3211 * to be mapped. b_bcount might be modified by the driver.
3214 vmapbuf(struct buf *bp)
3216 caddr_t addr, v, kva;
3222 if ((bp->b_flags & B_PHYS) == 0)
3224 if (bp->b_bufsize < 0)
3226 for (v = bp->b_saveaddr,
3227 addr = (caddr_t)trunc_page((vm_offset_t)bp->b_data),
3229 addr < bp->b_data + bp->b_bufsize;
3230 addr += PAGE_SIZE, v += PAGE_SIZE, pidx++) {
3232 * Do the vm_fault if needed; do the copy-on-write thing
3233 * when reading stuff off device into memory.
3236 i = vm_fault_quick((addr >= bp->b_data) ? addr : bp->b_data,
3237 (bp->b_flags&B_READ)?(VM_PROT_READ|VM_PROT_WRITE):VM_PROT_READ);
3239 for (i = 0; i < pidx; ++i) {
3240 vm_page_unhold(bp->b_pages[i]);
3241 bp->b_pages[i] = NULL;
3247 * WARNING! If sparc support is MFCd in the future this will
3248 * have to be changed from pmap_kextract() to pmap_extract()
3252 #error "If MFCing sparc support use pmap_extract"
3254 pa = pmap_kextract((vm_offset_t)addr);
3256 printf("vmapbuf: warning, race against user address during I/O");
3259 m = PHYS_TO_VM_PAGE(pa);
3261 bp->b_pages[pidx] = m;
3263 if (pidx > btoc(MAXPHYS))
3264 panic("vmapbuf: mapped more than MAXPHYS");
3265 pmap_qenter((vm_offset_t)bp->b_saveaddr, bp->b_pages, pidx);
3267 kva = bp->b_saveaddr;
3268 bp->b_npages = pidx;
3269 bp->b_saveaddr = bp->b_data;
3270 bp->b_data = kva + (((vm_offset_t) bp->b_data) & PAGE_MASK);
3275 * Free the io map PTEs associated with this IO operation.
3276 * We also invalidate the TLB entries and restore the original b_addr.
3286 if ((bp->b_flags & B_PHYS) == 0)
3289 npages = bp->b_npages;
3290 pmap_qremove(trunc_page((vm_offset_t)bp->b_data),
3293 for (pidx = 0; pidx < npages; pidx++)
3294 vm_page_unhold(*m++);
3296 bp->b_data = bp->b_saveaddr;
3299 #include "opt_ddb.h"
3301 #include <ddb/ddb.h>
3303 DB_SHOW_COMMAND(buffer, db_show_buffer)
3306 struct buf *bp = (struct buf *)addr;
3309 db_printf("usage: show buffer <addr>\n");
3313 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3314 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, "
3315 "b_resid = %ld\nb_dev = (%d,%d), b_data = %p, "
3316 "b_blkno = %d, b_pblkno = %d\n",
3317 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3318 major(bp->b_dev), minor(bp->b_dev),
3319 bp->b_data, bp->b_blkno, bp->b_pblkno);
3322 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages);
3323 for (i = 0; i < bp->b_npages; i++) {
3326 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3327 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3328 if ((i + 1) < bp->b_npages)