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 1.76 2006/05/25 19:31:13 dillon Exp $
19 * this file contains a new buffer I/O scheme implementing a coherent
20 * VM object and buffer cache scheme. Pains have been taken to make
21 * sure that the performance degradation associated with schemes such
22 * as this is not realized.
24 * Author: John S. Dyson
25 * Significant help during the development and debugging phases
26 * had been provided by David Greenman, also of the FreeBSD core team.
28 * see man buf(9) for more info.
31 #include <sys/param.h>
32 #include <sys/systm.h>
35 #include <sys/eventhandler.h>
37 #include <sys/malloc.h>
38 #include <sys/mount.h>
39 #include <sys/kernel.h>
40 #include <sys/kthread.h>
42 #include <sys/reboot.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sysctl.h>
45 #include <sys/vmmeter.h>
46 #include <sys/vnode.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_kern.h>
51 #include <vm/vm_pageout.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_map.h>
58 #include <sys/thread2.h>
59 #include <vm/vm_page2.h>
69 #define BUFFER_QUEUES 6
71 BQUEUE_NONE, /* not on any queue */
72 BQUEUE_LOCKED, /* locked buffers */
73 BQUEUE_CLEAN, /* non-B_DELWRI buffers */
74 BQUEUE_DIRTY, /* B_DELWRI buffers */
75 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
76 BQUEUE_EMPTY /* empty buffer headers */
78 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
80 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
82 struct bio_ops bioops; /* I/O operation notification */
84 struct buf *buf; /* buffer header pool */
86 static void vm_hold_free_pages(struct buf *bp, vm_offset_t from,
88 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
90 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
91 int pageno, vm_page_t m);
92 static void vfs_clean_pages(struct buf *bp);
93 static void vfs_setdirty(struct buf *bp);
94 static void vfs_vmio_release(struct buf *bp);
95 static int flushbufqueues(void);
97 static int bd_request;
99 static void buf_daemon (void);
101 * bogus page -- for I/O to/from partially complete buffers
102 * this is a temporary solution to the problem, but it is not
103 * really that bad. it would be better to split the buffer
104 * for input in the case of buffers partially already in memory,
105 * but the code is intricate enough already.
107 vm_page_t bogus_page;
110 static int bufspace, maxbufspace,
111 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
112 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
113 static int needsbuffer;
114 static int lorunningspace, hirunningspace, runningbufreq;
115 static int numdirtybuffers, lodirtybuffers, hidirtybuffers;
116 static int numfreebuffers, lofreebuffers, hifreebuffers;
117 static int getnewbufcalls;
118 static int getnewbufrestarts;
121 * Sysctls for operational control of the buffer cache.
123 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
124 "Number of dirty buffers to flush before bufdaemon becomes inactive");
125 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
126 "High watermark used to trigger explicit flushing of dirty buffers");
127 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
128 "Low watermark for special reserve in low-memory situations");
129 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
130 "High watermark for special reserve in low-memory situations");
131 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
132 "Minimum amount of buffer space required for active I/O");
133 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
134 "Maximum amount of buffer space to usable for active I/O");
136 * Sysctls determining current state of the buffer cache.
138 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
139 "Pending number of dirty buffers");
140 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
141 "Number of free buffers on the buffer cache free list");
142 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
143 "I/O bytes currently in progress due to asynchronous writes");
144 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
145 "Hard limit on maximum amount of memory usable for buffer space");
146 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
147 "Soft limit on maximum amount of memory usable for buffer space");
148 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
149 "Minimum amount of memory to reserve for system buffer space");
150 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
151 "Amount of memory available for buffers");
152 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
153 0, "Maximum amount of memory reserved for buffers using malloc");
154 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
155 "Amount of memory left for buffers using malloc-scheme");
156 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
157 "New buffer header acquisition requests");
158 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
159 0, "New buffer header acquisition restarts");
160 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
161 "Buffer acquisition restarts due to fragmented buffer map");
162 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
163 "Amount of time KVA space was deallocated in an arbitrary buffer");
164 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
165 "Amount of time buffer re-use operations were successful");
166 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
167 "sizeof(struct buf)");
169 char *buf_wmesg = BUF_WMESG;
171 extern int vm_swap_size;
173 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
174 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
175 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
176 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
181 * If someone is blocked due to there being too many dirty buffers,
182 * and numdirtybuffers is now reasonable, wake them up.
186 numdirtywakeup(int level)
188 if (numdirtybuffers <= level) {
189 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
190 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
191 wakeup(&needsbuffer);
199 * Called when buffer space is potentially available for recovery.
200 * getnewbuf() will block on this flag when it is unable to free
201 * sufficient buffer space. Buffer space becomes recoverable when
202 * bp's get placed back in the queues.
209 * If someone is waiting for BUF space, wake them up. Even
210 * though we haven't freed the kva space yet, the waiting
211 * process will be able to now.
213 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
214 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
215 wakeup(&needsbuffer);
222 * Accounting for I/O in progress.
226 runningbufwakeup(struct buf *bp)
228 if (bp->b_runningbufspace) {
229 runningbufspace -= bp->b_runningbufspace;
230 bp->b_runningbufspace = 0;
231 if (runningbufreq && runningbufspace <= lorunningspace) {
233 wakeup(&runningbufreq);
241 * Called when a buffer has been added to one of the free queues to
242 * account for the buffer and to wakeup anyone waiting for free buffers.
243 * This typically occurs when large amounts of metadata are being handled
244 * by the buffer cache ( else buffer space runs out first, usually ).
252 needsbuffer &= ~VFS_BIO_NEED_ANY;
253 if (numfreebuffers >= hifreebuffers)
254 needsbuffer &= ~VFS_BIO_NEED_FREE;
255 wakeup(&needsbuffer);
260 * waitrunningbufspace()
262 * runningbufspace is a measure of the amount of I/O currently
263 * running. This routine is used in async-write situations to
264 * prevent creating huge backups of pending writes to a device.
265 * Only asynchronous writes are governed by this function.
267 * Reads will adjust runningbufspace, but will not block based on it.
268 * The read load has a side effect of reducing the allowed write load.
270 * This does NOT turn an async write into a sync write. It waits
271 * for earlier writes to complete and generally returns before the
272 * caller's write has reached the device.
275 waitrunningbufspace(void)
277 if (runningbufspace > hirunningspace) {
279 while (runningbufspace > hirunningspace) {
281 tsleep(&runningbufreq, 0, "wdrain", 0);
288 * vfs_buf_test_cache:
290 * Called when a buffer is extended. This function clears the B_CACHE
291 * bit if the newly extended portion of the buffer does not contain
296 vfs_buf_test_cache(struct buf *bp,
297 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
300 if (bp->b_flags & B_CACHE) {
301 int base = (foff + off) & PAGE_MASK;
302 if (vm_page_is_valid(m, base, size) == 0)
303 bp->b_flags &= ~B_CACHE;
310 * Wake up the buffer daemon if the number of outstanding dirty buffers
311 * is above specified threshold 'dirtybuflevel'.
313 * The buffer daemon is explicitly woken up when (a) the pending number
314 * of dirty buffers exceeds the recovery and stall mid-point value,
315 * (b) during bwillwrite() or (c) buf freelist was exhausted.
319 bd_wakeup(int dirtybuflevel)
321 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
330 * Speed up the buffer cache flushing process.
343 * Load time initialisation of the buffer cache, called from machine
344 * dependant initialization code.
350 vm_offset_t bogus_offset;
353 /* next, make a null set of free lists */
354 for (i = 0; i < BUFFER_QUEUES; i++)
355 TAILQ_INIT(&bufqueues[i]);
357 /* finally, initialize each buffer header and stick on empty q */
358 for (i = 0; i < nbuf; i++) {
360 bzero(bp, sizeof *bp);
361 bp->b_flags = B_INVAL; /* we're just an empty header */
362 bp->b_cmd = BUF_CMD_DONE;
363 bp->b_qindex = BQUEUE_EMPTY;
365 xio_init(&bp->b_xio);
366 LIST_INIT(&bp->b_dep);
368 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
372 * maxbufspace is the absolute maximum amount of buffer space we are
373 * allowed to reserve in KVM and in real terms. The absolute maximum
374 * is nominally used by buf_daemon. hibufspace is the nominal maximum
375 * used by most other processes. The differential is required to
376 * ensure that buf_daemon is able to run when other processes might
377 * be blocked waiting for buffer space.
379 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
380 * this may result in KVM fragmentation which is not handled optimally
383 maxbufspace = nbuf * BKVASIZE;
384 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
385 lobufspace = hibufspace - MAXBSIZE;
387 lorunningspace = 512 * 1024;
388 hirunningspace = 1024 * 1024;
391 * Limit the amount of malloc memory since it is wired permanently into
392 * the kernel space. Even though this is accounted for in the buffer
393 * allocation, we don't want the malloced region to grow uncontrolled.
394 * The malloc scheme improves memory utilization significantly on average
395 * (small) directories.
397 maxbufmallocspace = hibufspace / 20;
400 * Reduce the chance of a deadlock occuring by limiting the number
401 * of delayed-write dirty buffers we allow to stack up.
403 hidirtybuffers = nbuf / 4 + 20;
406 * To support extreme low-memory systems, make sure hidirtybuffers cannot
407 * eat up all available buffer space. This occurs when our minimum cannot
408 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
409 * BKVASIZE'd (8K) buffers.
411 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
412 hidirtybuffers >>= 1;
414 lodirtybuffers = hidirtybuffers / 2;
417 * Try to keep the number of free buffers in the specified range,
418 * and give special processes (e.g. like buf_daemon) access to an
421 lofreebuffers = nbuf / 18 + 5;
422 hifreebuffers = 2 * lofreebuffers;
423 numfreebuffers = nbuf;
426 * Maximum number of async ops initiated per buf_daemon loop. This is
427 * somewhat of a hack at the moment, we really need to limit ourselves
428 * based on the number of bytes of I/O in-transit that were initiated
432 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE);
433 bogus_page = vm_page_alloc(kernel_object,
434 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
436 vmstats.v_wire_count++;
441 * Initialize the embedded bio structures
444 initbufbio(struct buf *bp)
446 bp->b_bio1.bio_buf = bp;
447 bp->b_bio1.bio_prev = NULL;
448 bp->b_bio1.bio_offset = NOOFFSET;
449 bp->b_bio1.bio_next = &bp->b_bio2;
450 bp->b_bio1.bio_done = NULL;
452 bp->b_bio2.bio_buf = bp;
453 bp->b_bio2.bio_prev = &bp->b_bio1;
454 bp->b_bio2.bio_offset = NOOFFSET;
455 bp->b_bio2.bio_next = NULL;
456 bp->b_bio2.bio_done = NULL;
460 * Reinitialize the embedded bio structures as well as any additional
461 * translation cache layers.
464 reinitbufbio(struct buf *bp)
468 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
469 bio->bio_done = NULL;
470 bio->bio_offset = NOOFFSET;
475 * Push another BIO layer onto an existing BIO and return it. The new
476 * BIO layer may already exist, holding cached translation data.
479 push_bio(struct bio *bio)
483 if ((nbio = bio->bio_next) == NULL) {
484 int index = bio - &bio->bio_buf->b_bio_array[0];
485 if (index >= NBUF_BIO) {
486 panic("push_bio: too many layers bp %p\n",
489 nbio = &bio->bio_buf->b_bio_array[index + 1];
490 bio->bio_next = nbio;
491 nbio->bio_prev = bio;
492 nbio->bio_buf = bio->bio_buf;
493 nbio->bio_offset = NOOFFSET;
494 nbio->bio_done = NULL;
495 nbio->bio_next = NULL;
497 KKASSERT(nbio->bio_done == NULL);
502 pop_bio(struct bio *bio)
508 clearbiocache(struct bio *bio)
511 bio->bio_offset = NOOFFSET;
519 * Free the KVA allocation for buffer 'bp'.
521 * Must be called from a critical section as this is the only locking for
524 * Since this call frees up buffer space, we call bufspacewakeup().
527 bfreekva(struct buf *bp)
533 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
534 vm_map_lock(buffer_map);
535 bufspace -= bp->b_kvasize;
536 vm_map_delete(buffer_map,
537 (vm_offset_t) bp->b_kvabase,
538 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
541 vm_map_unlock(buffer_map);
542 vm_map_entry_release(count);
551 * Remove the buffer from the appropriate free list.
554 bremfree(struct buf *bp)
559 old_qindex = bp->b_qindex;
561 if (bp->b_qindex != BQUEUE_NONE) {
562 KASSERT(BUF_REFCNTNB(bp) == 1,
563 ("bremfree: bp %p not locked",bp));
564 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
565 bp->b_qindex = BQUEUE_NONE;
567 if (BUF_REFCNTNB(bp) <= 1)
568 panic("bremfree: removing a buffer not on a queue");
572 * Fixup numfreebuffers count. If the buffer is invalid or not
573 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
574 * the buffer was free and we must decrement numfreebuffers.
576 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
581 case BQUEUE_EMPTYKVA:
595 * Get a buffer with the specified data. Look in the cache first. We
596 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
597 * is set, the buffer is valid and we do not have to do anything ( see
601 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
605 bp = getblk(vp, loffset, size, 0, 0);
608 /* if not found in cache, do some I/O */
609 if ((bp->b_flags & B_CACHE) == 0) {
610 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp));
611 bp->b_flags &= ~(B_ERROR | B_INVAL);
612 bp->b_cmd = BUF_CMD_READ;
613 vfs_busy_pages(vp, bp);
614 vn_strategy(vp, &bp->b_bio1);
615 return (biowait(bp));
623 * Operates like bread, but also starts asynchronous I/O on
624 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
625 * to initiating I/O . If B_CACHE is set, the buffer is valid
626 * and we do not have to do anything.
629 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
630 int *rabsize, int cnt, struct buf **bpp)
632 struct buf *bp, *rabp;
634 int rv = 0, readwait = 0;
636 *bpp = bp = getblk(vp, loffset, size, 0, 0);
638 /* if not found in cache, do some I/O */
639 if ((bp->b_flags & B_CACHE) == 0) {
640 bp->b_flags &= ~(B_ERROR | B_INVAL);
641 bp->b_cmd = BUF_CMD_READ;
642 vfs_busy_pages(vp, bp);
643 vn_strategy(vp, &bp->b_bio1);
647 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
648 if (inmem(vp, *raoffset))
650 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
652 if ((rabp->b_flags & B_CACHE) == 0) {
653 rabp->b_flags |= B_ASYNC;
654 rabp->b_flags &= ~(B_ERROR | B_INVAL);
655 rabp->b_cmd = BUF_CMD_READ;
656 vfs_busy_pages(vp, rabp);
658 vn_strategy(vp, &rabp->b_bio1);
673 * Write, release buffer on completion. (Done by iodone
674 * if async). Do not bother writing anything if the buffer
677 * Note that we set B_CACHE here, indicating that buffer is
678 * fully valid and thus cacheable. This is true even of NFS
679 * now so we set it generally. This could be set either here
680 * or in biodone() since the I/O is synchronous. We put it
684 bwrite(struct buf *bp)
688 if (bp->b_flags & B_INVAL) {
693 oldflags = bp->b_flags;
695 if (BUF_REFCNTNB(bp) == 0)
696 panic("bwrite: buffer is not busy???");
699 /* Mark the buffer clean */
702 bp->b_flags &= ~B_ERROR;
703 bp->b_flags |= B_CACHE;
704 bp->b_cmd = BUF_CMD_WRITE;
705 vfs_busy_pages(bp->b_vp, bp);
708 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
709 * valid for vnode-backed buffers.
711 bp->b_runningbufspace = bp->b_bufsize;
712 runningbufspace += bp->b_runningbufspace;
715 if (oldflags & B_ASYNC)
717 vn_strategy(bp->b_vp, &bp->b_bio1);
719 if ((oldflags & B_ASYNC) == 0) {
720 int rtval = biowait(bp);
723 } else if ((oldflags & B_NOWDRAIN) == 0) {
725 * don't allow the async write to saturate the I/O
726 * system. Deadlocks can occur only if a device strategy
727 * routine (like in VN) turns around and issues another
728 * high-level write, in which case B_NOWDRAIN is expected
729 * to be set. Otherwise we will not deadlock here because
730 * we are blocking waiting for I/O that is already in-progress
733 waitrunningbufspace();
742 * Delayed write. (Buffer is marked dirty). Do not bother writing
743 * anything if the buffer is marked invalid.
745 * Note that since the buffer must be completely valid, we can safely
746 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
747 * biodone() in order to prevent getblk from writing the buffer
751 bdwrite(struct buf *bp)
753 if (BUF_REFCNTNB(bp) == 0)
754 panic("bdwrite: buffer is not busy");
756 if (bp->b_flags & B_INVAL) {
763 * Set B_CACHE, indicating that the buffer is fully valid. This is
764 * true even of NFS now.
766 bp->b_flags |= B_CACHE;
769 * This bmap keeps the system from needing to do the bmap later,
770 * perhaps when the system is attempting to do a sync. Since it
771 * is likely that the indirect block -- or whatever other datastructure
772 * that the filesystem needs is still in memory now, it is a good
773 * thing to do this. Note also, that if the pageout daemon is
774 * requesting a sync -- there might not be enough memory to do
775 * the bmap then... So, this is important to do.
777 if (bp->b_bio2.bio_offset == NOOFFSET) {
778 VOP_BMAP(bp->b_vp, bp->b_loffset, NULL, &bp->b_bio2.bio_offset,
783 * Set the *dirty* buffer range based upon the VM system dirty pages.
788 * We need to do this here to satisfy the vnode_pager and the
789 * pageout daemon, so that it thinks that the pages have been
790 * "cleaned". Note that since the pages are in a delayed write
791 * buffer -- the VFS layer "will" see that the pages get written
792 * out on the next sync, or perhaps the cluster will be completed.
798 * Wakeup the buffer flushing daemon if we have a lot of dirty
799 * buffers (midpoint between our recovery point and our stall
802 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
805 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
806 * due to the softdep code.
813 * Turn buffer into delayed write request by marking it B_DELWRI.
814 * B_RELBUF and B_NOCACHE must be cleared.
816 * We reassign the buffer to itself to properly update it in the
819 * Since the buffer is not on a queue, we do not update the
820 * numfreebuffers count.
822 * Must be called from a critical section.
823 * The buffer must be on BQUEUE_NONE.
826 bdirty(struct buf *bp)
828 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
829 if (bp->b_flags & B_NOCACHE) {
830 printf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
831 bp->b_flags &= ~B_NOCACHE;
833 if (bp->b_flags & B_INVAL) {
834 printf("bdirty: warning, dirtying invalid buffer %p\n", bp);
836 bp->b_flags &= ~B_RELBUF;
838 if ((bp->b_flags & B_DELWRI) == 0) {
839 bp->b_flags |= B_DELWRI;
842 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
849 * Clear B_DELWRI for buffer.
851 * Since the buffer is not on a queue, we do not update the numfreebuffers
854 * Must be called from a critical section.
856 * The buffer is typically on BQUEUE_NONE but there is one case in
857 * brelse() that calls this function after placing the buffer on
862 bundirty(struct buf *bp)
864 if (bp->b_flags & B_DELWRI) {
865 bp->b_flags &= ~B_DELWRI;
868 numdirtywakeup(lodirtybuffers);
871 * Since it is now being written, we can clear its deferred write flag.
873 bp->b_flags &= ~B_DEFERRED;
879 * Asynchronous write. Start output on a buffer, but do not wait for
880 * it to complete. The buffer is released when the output completes.
882 * bwrite() ( or the VOP routine anyway ) is responsible for handling
883 * B_INVAL buffers. Not us.
886 bawrite(struct buf *bp)
888 bp->b_flags |= B_ASYNC;
895 * Ordered write. Start output on a buffer, and flag it so that the
896 * device will write it in the order it was queued. The buffer is
897 * released when the output completes. bwrite() ( or the VOP routine
898 * anyway ) is responsible for handling B_INVAL buffers.
901 bowrite(struct buf *bp)
903 bp->b_flags |= B_ORDERED | B_ASYNC;
910 * Called prior to the locking of any vnodes when we are expecting to
911 * write. We do not want to starve the buffer cache with too many
912 * dirty buffers so we block here. By blocking prior to the locking
913 * of any vnodes we attempt to avoid the situation where a locked vnode
914 * prevents the various system daemons from flushing related buffers.
920 if (numdirtybuffers >= hidirtybuffers) {
922 while (numdirtybuffers >= hidirtybuffers) {
924 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
925 tsleep(&needsbuffer, 0, "flswai", 0);
932 * buf_dirty_count_severe:
934 * Return true if we have too many dirty buffers.
937 buf_dirty_count_severe(void)
939 return(numdirtybuffers >= hidirtybuffers);
945 * Release a busy buffer and, if requested, free its resources. The
946 * buffer will be stashed in the appropriate bufqueue[] allowing it
947 * to be accessed later as a cache entity or reused for other purposes.
950 brelse(struct buf *bp)
953 int saved_flags = bp->b_flags;
956 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
961 * If B_NOCACHE is set we are being asked to destroy the buffer and
962 * its backing store. Clear B_DELWRI.
964 * B_NOCACHE is set in two cases: (1) when the caller really wants
965 * to destroy the buffer and backing store and (2) when the caller
966 * wants to destroy the buffer and backing store after a write
969 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
973 if (bp->b_flags & B_LOCKED)
974 bp->b_flags &= ~B_ERROR;
977 * If a write error occurs and the caller does not want to throw
978 * away the buffer, redirty the buffer. This will also clear
981 if (bp->b_cmd == BUF_CMD_WRITE &&
982 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
984 * Failed write, redirty. Must clear B_ERROR to prevent
985 * pages from being scrapped. If B_INVAL is set then
986 * this case is not run and the next case is run to
987 * destroy the buffer. B_INVAL can occur if the buffer
988 * is outside the range supported by the underlying device.
990 bp->b_flags &= ~B_ERROR;
992 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
993 (bp->b_bufsize <= 0) || bp->b_cmd == BUF_CMD_FREEBLKS) {
995 * Either a failed I/O or we were asked to free or not
998 bp->b_flags |= B_INVAL;
999 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
1000 (*bioops.io_deallocate)(bp);
1001 if (bp->b_flags & B_DELWRI) {
1003 numdirtywakeup(lodirtybuffers);
1005 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1009 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release()
1010 * is called with B_DELWRI set, the underlying pages may wind up
1011 * getting freed causing a previous write (bdwrite()) to get 'lost'
1012 * because pages associated with a B_DELWRI bp are marked clean.
1014 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1015 * if B_DELWRI is set.
1017 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1018 * on pages to return pages to the VM page queues.
1020 if (bp->b_flags & B_DELWRI)
1021 bp->b_flags &= ~B_RELBUF;
1022 else if (vm_page_count_severe())
1023 bp->b_flags |= B_RELBUF;
1026 * At this point destroying the buffer is governed by the B_INVAL
1027 * or B_RELBUF flags.
1029 bp->b_cmd = BUF_CMD_DONE;
1032 * VMIO buffer rundown. Make sure the VM page array is restored
1033 * after an I/O may have replaces some of the pages with bogus pages
1034 * in order to not destroy dirty pages in a fill-in read.
1036 * Note that due to the code above, if a buffer is marked B_DELWRI
1037 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1038 * B_INVAL may still be set, however.
1040 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1041 * but not the backing store. B_NOCACHE will destroy the backing
1044 * Note that dirty NFS buffers contain byte-granular write ranges
1045 * and should not be destroyed w/ B_INVAL even if the backing store
1048 if (bp->b_flags & B_VMIO) {
1050 * Rundown for VMIO buffers which are not dirty NFS buffers.
1062 * Get the base offset and length of the buffer. Note that
1063 * in the VMIO case if the buffer block size is not
1064 * page-aligned then b_data pointer may not be page-aligned.
1065 * But our b_xio.xio_pages array *IS* page aligned.
1067 * block sizes less then DEV_BSIZE (usually 512) are not
1068 * supported due to the page granularity bits (m->valid,
1069 * m->dirty, etc...).
1071 * See man buf(9) for more information
1074 resid = bp->b_bufsize;
1075 foff = bp->b_loffset;
1077 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1078 m = bp->b_xio.xio_pages[i];
1079 vm_page_flag_clear(m, PG_ZERO);
1081 * If we hit a bogus page, fixup *all* of them
1082 * now. Note that we left these pages wired
1083 * when we removed them so they had better exist,
1084 * and they cannot be ripped out from under us so
1085 * no critical section protection is necessary.
1087 if (m == bogus_page) {
1089 poff = OFF_TO_IDX(bp->b_loffset);
1091 for (j = i; j < bp->b_xio.xio_npages; j++) {
1094 mtmp = bp->b_xio.xio_pages[j];
1095 if (mtmp == bogus_page) {
1096 mtmp = vm_page_lookup(obj, poff + j);
1098 panic("brelse: page missing");
1100 bp->b_xio.xio_pages[j] = mtmp;
1104 if ((bp->b_flags & B_INVAL) == 0) {
1105 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1106 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1108 m = bp->b_xio.xio_pages[i];
1112 * Invalidate the backing store if B_NOCACHE is set
1113 * (e.g. used with vinvalbuf()). If this is NFS
1114 * we impose a requirement that the block size be
1115 * a multiple of PAGE_SIZE and create a temporary
1116 * hack to basically invalidate the whole page. The
1117 * problem is that NFS uses really odd buffer sizes
1118 * especially when tracking piecemeal writes and
1119 * it also vinvalbuf()'s a lot, which would result
1120 * in only partial page validation and invalidation
1121 * here. If the file page is mmap()'d, however,
1122 * all the valid bits get set so after we invalidate
1123 * here we would end up with weird m->valid values
1124 * like 0xfc. nfs_getpages() can't handle this so
1125 * we clear all the valid bits for the NFS case
1126 * instead of just some of them.
1128 * The real bug is the VM system having to set m->valid
1129 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1130 * itself is an artifact of the whole 512-byte
1131 * granular mess that exists to support odd block
1132 * sizes and UFS meta-data block sizes (e.g. 6144).
1133 * A complete rewrite is required.
1135 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1136 int poffset = foff & PAGE_MASK;
1139 presid = PAGE_SIZE - poffset;
1140 if (bp->b_vp->v_tag == VT_NFS &&
1141 bp->b_vp->v_type == VREG) {
1143 } else if (presid > resid) {
1146 KASSERT(presid >= 0, ("brelse: extra page"));
1147 vm_page_set_invalid(m, poffset, presid);
1149 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1150 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1152 if (bp->b_flags & (B_INVAL | B_RELBUF))
1153 vfs_vmio_release(bp);
1156 * Rundown for non-VMIO buffers.
1158 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1161 printf("brelse bp %p %08x/%08x: Warning, caught and fixed brelvp bug\n", bp, saved_flags, bp->b_flags);
1170 if (bp->b_qindex != BQUEUE_NONE)
1171 panic("brelse: free buffer onto another queue???");
1172 if (BUF_REFCNTNB(bp) > 1) {
1173 /* Temporary panic to verify exclusive locking */
1174 /* This panic goes away when we allow shared refs */
1175 panic("brelse: multiple refs");
1176 /* do not release to free list */
1183 * Figure out the correct queue to place the cleaned up buffer on.
1184 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1185 * disassociated from their vnode.
1188 if (bp->b_bufsize == 0) {
1190 * Buffers with no memory. Due to conditionals near the top
1191 * of brelse() such buffers should probably already be
1192 * marked B_INVAL and disassociated from their vnode.
1194 bp->b_flags |= B_INVAL;
1195 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1196 KKASSERT((bp->b_flags & B_HASHED) == 0);
1197 if (bp->b_kvasize) {
1198 bp->b_qindex = BQUEUE_EMPTYKVA;
1200 bp->b_qindex = BQUEUE_EMPTY;
1202 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1203 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) {
1205 * Buffers with junk contents. Again these buffers had better
1206 * already be disassociated from their vnode.
1208 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1209 KKASSERT((bp->b_flags & B_HASHED) == 0);
1210 bp->b_flags |= B_INVAL;
1211 bp->b_qindex = BQUEUE_CLEAN;
1212 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1213 } else if (bp->b_flags & B_LOCKED) {
1215 * Buffers that are locked.
1217 bp->b_qindex = BQUEUE_LOCKED;
1218 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1221 * Remaining buffers. These buffers are still associated with
1224 switch(bp->b_flags & (B_DELWRI|B_AGE)) {
1225 case B_DELWRI | B_AGE:
1226 bp->b_qindex = BQUEUE_DIRTY;
1227 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1230 bp->b_qindex = BQUEUE_DIRTY;
1231 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1234 bp->b_qindex = BQUEUE_CLEAN;
1235 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1238 bp->b_qindex = BQUEUE_CLEAN;
1239 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1245 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1246 * on the correct queue.
1248 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1252 * Fixup numfreebuffers count. The bp is on an appropriate queue
1253 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1254 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1255 * if B_INVAL is set ).
1257 if ((bp->b_flags & B_LOCKED) == 0 && !(bp->b_flags & B_DELWRI))
1261 * Something we can maybe free or reuse
1263 if (bp->b_bufsize || bp->b_kvasize)
1267 * Clean up temporary flags and unlock the buffer.
1269 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF |
1270 B_DIRECT | B_NOWDRAIN);
1278 * Release a buffer back to the appropriate queue but do not try to free
1279 * it. The buffer is expected to be used again soon.
1281 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1282 * biodone() to requeue an async I/O on completion. It is also used when
1283 * known good buffers need to be requeued but we think we may need the data
1286 * XXX we should be able to leave the B_RELBUF hint set on completion.
1289 bqrelse(struct buf *bp)
1293 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1295 if (bp->b_qindex != BQUEUE_NONE)
1296 panic("bqrelse: free buffer onto another queue???");
1297 if (BUF_REFCNTNB(bp) > 1) {
1298 /* do not release to free list */
1299 panic("bqrelse: multiple refs");
1304 if (bp->b_flags & B_LOCKED) {
1305 bp->b_flags &= ~B_ERROR;
1306 bp->b_qindex = BQUEUE_LOCKED;
1307 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1308 /* buffers with stale but valid contents */
1309 } else if (bp->b_flags & B_DELWRI) {
1310 bp->b_qindex = BQUEUE_DIRTY;
1311 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1312 } else if (vm_page_count_severe()) {
1314 * We are too low on memory, we have to try to free the
1315 * buffer (most importantly: the wired pages making up its
1316 * backing store) *now*.
1322 bp->b_qindex = BQUEUE_CLEAN;
1323 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1326 if ((bp->b_flags & B_LOCKED) == 0 &&
1327 ((bp->b_flags & B_INVAL) || !(bp->b_flags & B_DELWRI))) {
1332 * Something we can maybe free or reuse.
1334 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1338 * Final cleanup and unlock. Clear bits that are only used while a
1339 * buffer is actively locked.
1341 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1349 * Return backing pages held by the buffer 'bp' back to the VM system
1350 * if possible. The pages are freed if they are no longer valid or
1351 * attempt to free if it was used for direct I/O otherwise they are
1352 * sent to the page cache.
1354 * Pages that were marked busy are left alone and skipped.
1356 * The KVA mapping (b_data) for the underlying pages is removed by
1360 vfs_vmio_release(struct buf *bp)
1366 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1367 m = bp->b_xio.xio_pages[i];
1368 bp->b_xio.xio_pages[i] = NULL;
1370 * In order to keep page LRU ordering consistent, put
1371 * everything on the inactive queue.
1373 vm_page_unwire(m, 0);
1375 * We don't mess with busy pages, it is
1376 * the responsibility of the process that
1377 * busied the pages to deal with them.
1379 if ((m->flags & PG_BUSY) || (m->busy != 0))
1382 if (m->wire_count == 0) {
1383 vm_page_flag_clear(m, PG_ZERO);
1385 * Might as well free the page if we can and it has
1386 * no valid data. We also free the page if the
1387 * buffer was used for direct I/O.
1389 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1390 m->hold_count == 0) {
1392 vm_page_protect(m, VM_PROT_NONE);
1394 } else if (bp->b_flags & B_DIRECT) {
1395 vm_page_try_to_free(m);
1396 } else if (vm_page_count_severe()) {
1397 vm_page_try_to_cache(m);
1402 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1403 if (bp->b_bufsize) {
1407 bp->b_xio.xio_npages = 0;
1408 bp->b_flags &= ~B_VMIO;
1416 * Implement clustered async writes for clearing out B_DELWRI buffers.
1417 * This is much better then the old way of writing only one buffer at
1418 * a time. Note that we may not be presented with the buffers in the
1419 * correct order, so we search for the cluster in both directions.
1421 * The buffer is locked on call.
1424 vfs_bio_awrite(struct buf *bp)
1428 off_t loffset = bp->b_loffset;
1429 struct vnode *vp = bp->b_vp;
1437 * right now we support clustered writing only to regular files. If
1438 * we find a clusterable block we could be in the middle of a cluster
1439 * rather then at the beginning.
1441 * NOTE: b_bio1 contains the logical loffset and is aliased
1442 * to b_loffset. b_bio2 contains the translated block number.
1444 if ((vp->v_type == VREG) &&
1445 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1446 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1448 size = vp->v_mount->mnt_stat.f_iosize;
1450 for (i = size; i < MAXPHYS; i += size) {
1451 if ((bpa = findblk(vp, loffset + i)) &&
1452 BUF_REFCNT(bpa) == 0 &&
1453 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1454 (B_DELWRI | B_CLUSTEROK)) &&
1455 (bpa->b_bufsize == size)) {
1456 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1457 (bpa->b_bio2.bio_offset !=
1458 bp->b_bio2.bio_offset + i))
1464 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1465 if ((bpa = findblk(vp, loffset - j)) &&
1466 BUF_REFCNT(bpa) == 0 &&
1467 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1468 (B_DELWRI | B_CLUSTEROK)) &&
1469 (bpa->b_bufsize == size)) {
1470 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1471 (bpa->b_bio2.bio_offset !=
1472 bp->b_bio2.bio_offset - j))
1481 * this is a possible cluster write
1483 if (nbytes != size) {
1485 nwritten = cluster_wbuild(vp, size,
1486 loffset - j, nbytes);
1493 bp->b_flags |= B_ASYNC;
1497 * default (old) behavior, writing out only one block
1499 * XXX returns b_bufsize instead of b_bcount for nwritten?
1501 nwritten = bp->b_bufsize;
1510 * Find and initialize a new buffer header, freeing up existing buffers
1511 * in the bufqueues as necessary. The new buffer is returned locked.
1513 * Important: B_INVAL is not set. If the caller wishes to throw the
1514 * buffer away, the caller must set B_INVAL prior to calling brelse().
1517 * We have insufficient buffer headers
1518 * We have insufficient buffer space
1519 * buffer_map is too fragmented ( space reservation fails )
1520 * If we have to flush dirty buffers ( but we try to avoid this )
1522 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1523 * Instead we ask the buf daemon to do it for us. We attempt to
1524 * avoid piecemeal wakeups of the pageout daemon.
1528 getnewbuf(int slpflag, int slptimeo, int size, int maxsize)
1534 static int flushingbufs;
1537 * We can't afford to block since we might be holding a vnode lock,
1538 * which may prevent system daemons from running. We deal with
1539 * low-memory situations by proactively returning memory and running
1540 * async I/O rather then sync I/O.
1544 --getnewbufrestarts;
1546 ++getnewbufrestarts;
1549 * Setup for scan. If we do not have enough free buffers,
1550 * we setup a degenerate case that immediately fails. Note
1551 * that if we are specially marked process, we are allowed to
1552 * dip into our reserves.
1554 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1556 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1557 * However, there are a number of cases (defragging, reusing, ...)
1558 * where we cannot backup.
1560 nqindex = BQUEUE_EMPTYKVA;
1561 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1565 * If no EMPTYKVA buffers and we are either
1566 * defragging or reusing, locate a CLEAN buffer
1567 * to free or reuse. If bufspace useage is low
1568 * skip this step so we can allocate a new buffer.
1570 if (defrag || bufspace >= lobufspace) {
1571 nqindex = BQUEUE_CLEAN;
1572 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1576 * If we could not find or were not allowed to reuse a
1577 * CLEAN buffer, check to see if it is ok to use an EMPTY
1578 * buffer. We can only use an EMPTY buffer if allocating
1579 * its KVA would not otherwise run us out of buffer space.
1581 if (nbp == NULL && defrag == 0 &&
1582 bufspace + maxsize < hibufspace) {
1583 nqindex = BQUEUE_EMPTY;
1584 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1589 * Run scan, possibly freeing data and/or kva mappings on the fly
1593 while ((bp = nbp) != NULL) {
1594 int qindex = nqindex;
1597 * Calculate next bp ( we can only use it if we do not block
1598 * or do other fancy things ).
1600 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1603 nqindex = BQUEUE_EMPTYKVA;
1604 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1607 case BQUEUE_EMPTYKVA:
1608 nqindex = BQUEUE_CLEAN;
1609 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1623 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistant queue %d bp %p", qindex, bp));
1626 * Note: we no longer distinguish between VMIO and non-VMIO
1630 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1633 * If we are defragging then we need a buffer with
1634 * b_kvasize != 0. XXX this situation should no longer
1635 * occur, if defrag is non-zero the buffer's b_kvasize
1636 * should also be non-zero at this point. XXX
1638 if (defrag && bp->b_kvasize == 0) {
1639 printf("Warning: defrag empty buffer %p\n", bp);
1644 * Start freeing the bp. This is somewhat involved. nbp
1645 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1646 * on the clean list must be disassociated from their
1647 * current vnode. Buffers on the empty[kva] lists have
1648 * already been disassociated.
1651 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1652 printf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1653 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1656 if (bp->b_qindex != qindex) {
1657 printf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1663 if (qindex == BQUEUE_CLEAN) {
1664 if (bp->b_flags & B_VMIO) {
1665 bp->b_flags &= ~B_ASYNC;
1666 vfs_vmio_release(bp);
1673 * NOTE: nbp is now entirely invalid. We can only restart
1674 * the scan from this point on.
1676 * Get the rest of the buffer freed up. b_kva* is still
1677 * valid after this operation.
1680 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1681 KKASSERT((bp->b_flags & B_HASHED) == 0);
1682 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate)
1683 (*bioops.io_deallocate)(bp);
1686 * critical section protection is not required when
1687 * scrapping a buffer's contents because it is already
1693 bp->b_flags = B_BNOCLIP;
1694 bp->b_cmd = BUF_CMD_DONE;
1699 bp->b_xio.xio_npages = 0;
1700 bp->b_dirtyoff = bp->b_dirtyend = 0;
1703 LIST_INIT(&bp->b_dep);
1706 * If we are defragging then free the buffer.
1709 bp->b_flags |= B_INVAL;
1717 * If we are overcomitted then recover the buffer and its
1718 * KVM space. This occurs in rare situations when multiple
1719 * processes are blocked in getnewbuf() or allocbuf().
1721 if (bufspace >= hibufspace)
1723 if (flushingbufs && bp->b_kvasize != 0) {
1724 bp->b_flags |= B_INVAL;
1729 if (bufspace < lobufspace)
1735 * If we exhausted our list, sleep as appropriate. We may have to
1736 * wakeup various daemons and write out some dirty buffers.
1738 * Generally we are sleeping due to insufficient buffer space.
1746 flags = VFS_BIO_NEED_BUFSPACE;
1748 } else if (bufspace >= hibufspace) {
1750 flags = VFS_BIO_NEED_BUFSPACE;
1753 flags = VFS_BIO_NEED_ANY;
1756 bd_speedup(); /* heeeelp */
1758 needsbuffer |= flags;
1759 while (needsbuffer & flags) {
1760 if (tsleep(&needsbuffer, slpflag, waitmsg, slptimeo))
1765 * We finally have a valid bp. We aren't quite out of the
1766 * woods, we still have to reserve kva space. In order
1767 * to keep fragmentation sane we only allocate kva in
1770 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1772 if (maxsize != bp->b_kvasize) {
1773 vm_offset_t addr = 0;
1778 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1779 vm_map_lock(buffer_map);
1781 if (vm_map_findspace(buffer_map,
1782 vm_map_min(buffer_map), maxsize,
1785 * Uh oh. Buffer map is too fragmented. We
1786 * must defragment the map.
1788 vm_map_unlock(buffer_map);
1789 vm_map_entry_release(count);
1792 bp->b_flags |= B_INVAL;
1797 vm_map_insert(buffer_map, &count,
1799 addr, addr + maxsize,
1800 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
1802 bp->b_kvabase = (caddr_t) addr;
1803 bp->b_kvasize = maxsize;
1804 bufspace += bp->b_kvasize;
1807 vm_map_unlock(buffer_map);
1808 vm_map_entry_release(count);
1810 bp->b_data = bp->b_kvabase;
1818 * Buffer flushing daemon. Buffers are normally flushed by the
1819 * update daemon but if it cannot keep up this process starts to
1820 * take the load in an attempt to prevent getnewbuf() from blocking.
1823 static struct thread *bufdaemonthread;
1825 static struct kproc_desc buf_kp = {
1830 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp)
1836 * This process needs to be suspended prior to shutdown sync.
1838 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
1839 bufdaemonthread, SHUTDOWN_PRI_LAST);
1842 * This process is allowed to take the buffer cache to the limit
1847 kproc_suspend_loop();
1850 * Do the flush. Limit the amount of in-transit I/O we
1851 * allow to build up, otherwise we would completely saturate
1852 * the I/O system. Wakeup any waiting processes before we
1853 * normally would so they can run in parallel with our drain.
1855 while (numdirtybuffers > lodirtybuffers) {
1856 if (flushbufqueues() == 0)
1858 waitrunningbufspace();
1859 numdirtywakeup((lodirtybuffers + hidirtybuffers) / 2);
1863 * Only clear bd_request if we have reached our low water
1864 * mark. The buf_daemon normally waits 5 seconds and
1865 * then incrementally flushes any dirty buffers that have
1866 * built up, within reason.
1868 * If we were unable to hit our low water mark and couldn't
1869 * find any flushable buffers, we sleep half a second.
1870 * Otherwise we loop immediately.
1872 if (numdirtybuffers <= lodirtybuffers) {
1874 * We reached our low water mark, reset the
1875 * request and sleep until we are needed again.
1876 * The sleep is just so the suspend code works.
1879 tsleep(&bd_request, 0, "psleep", hz);
1882 * We couldn't find any flushable dirty buffers but
1883 * still have too many dirty buffers, we
1884 * have to sleep and try again. (rare)
1886 tsleep(&bd_request, 0, "qsleep", hz / 2);
1894 * Try to flush a buffer in the dirty queue. We must be careful to
1895 * free up B_INVAL buffers instead of write them, which NFS is
1896 * particularly sensitive to.
1900 flushbufqueues(void)
1905 bp = TAILQ_FIRST(&bufqueues[BQUEUE_DIRTY]);
1908 KASSERT((bp->b_flags & B_DELWRI), ("unexpected clean buffer %p", bp));
1909 if (bp->b_flags & B_DELWRI) {
1910 if (bp->b_flags & B_INVAL) {
1911 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
1912 panic("flushbufqueues: locked buf");
1918 if (LIST_FIRST(&bp->b_dep) != NULL &&
1919 bioops.io_countdeps &&
1920 (bp->b_flags & B_DEFERRED) == 0 &&
1921 (*bioops.io_countdeps)(bp, 0)) {
1922 TAILQ_REMOVE(&bufqueues[BQUEUE_DIRTY],
1924 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY],
1926 bp->b_flags |= B_DEFERRED;
1927 bp = TAILQ_FIRST(&bufqueues[BQUEUE_DIRTY]);
1932 * Only write it out if we can successfully lock
1935 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
1941 bp = TAILQ_NEXT(bp, b_freelist);
1949 * Returns true if no I/O is needed to access the associated VM object.
1950 * This is like findblk except it also hunts around in the VM system for
1953 * Note that we ignore vm_page_free() races from interrupts against our
1954 * lookup, since if the caller is not protected our return value will not
1955 * be any more valid then otherwise once we exit the critical section.
1958 inmem(struct vnode *vp, off_t loffset)
1961 vm_offset_t toff, tinc, size;
1964 if (findblk(vp, loffset))
1966 if (vp->v_mount == NULL)
1968 if ((obj = vp->v_object) == NULL)
1972 if (size > vp->v_mount->mnt_stat.f_iosize)
1973 size = vp->v_mount->mnt_stat.f_iosize;
1975 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
1976 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
1980 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
1981 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
1982 if (vm_page_is_valid(m,
1983 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
1992 * Sets the dirty range for a buffer based on the status of the dirty
1993 * bits in the pages comprising the buffer.
1995 * The range is limited to the size of the buffer.
1997 * This routine is primarily used by NFS, but is generalized for the
2001 vfs_setdirty(struct buf *bp)
2007 * Degenerate case - empty buffer
2010 if (bp->b_bufsize == 0)
2014 * We qualify the scan for modified pages on whether the
2015 * object has been flushed yet. The OBJ_WRITEABLE flag
2016 * is not cleared simply by protecting pages off.
2019 if ((bp->b_flags & B_VMIO) == 0)
2022 object = bp->b_xio.xio_pages[0]->object;
2024 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2025 printf("Warning: object %p writeable but not mightbedirty\n", object);
2026 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2027 printf("Warning: object %p mightbedirty but not writeable\n", object);
2029 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2030 vm_offset_t boffset;
2031 vm_offset_t eoffset;
2034 * test the pages to see if they have been modified directly
2035 * by users through the VM system.
2037 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2038 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2039 vm_page_test_dirty(bp->b_xio.xio_pages[i]);
2043 * Calculate the encompassing dirty range, boffset and eoffset,
2044 * (eoffset - boffset) bytes.
2047 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2048 if (bp->b_xio.xio_pages[i]->dirty)
2051 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2053 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2054 if (bp->b_xio.xio_pages[i]->dirty) {
2058 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2061 * Fit it to the buffer.
2064 if (eoffset > bp->b_bcount)
2065 eoffset = bp->b_bcount;
2068 * If we have a good dirty range, merge with the existing
2072 if (boffset < eoffset) {
2073 if (bp->b_dirtyoff > boffset)
2074 bp->b_dirtyoff = boffset;
2075 if (bp->b_dirtyend < eoffset)
2076 bp->b_dirtyend = eoffset;
2084 * Locate and return the specified buffer, or NULL if the buffer does
2085 * not exist. Do not attempt to lock the buffer or manipulate it in
2086 * any way. The caller must validate that the correct buffer has been
2087 * obtain after locking it.
2090 findblk(struct vnode *vp, off_t loffset)
2095 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2103 * Get a block given a specified block and offset into a file/device.
2104 * B_INVAL may or may not be set on return. The caller should clear
2105 * B_INVAL prior to initiating a READ.
2107 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2108 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2109 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2110 * without doing any of those things the system will likely believe
2111 * the buffer to be valid (especially if it is not B_VMIO), and the
2112 * next getblk() will return the buffer with B_CACHE set.
2114 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2115 * an existing buffer.
2117 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2118 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2119 * and then cleared based on the backing VM. If the previous buffer is
2120 * non-0-sized but invalid, B_CACHE will be cleared.
2122 * If getblk() must create a new buffer, the new buffer is returned with
2123 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2124 * case it is returned with B_INVAL clear and B_CACHE set based on the
2127 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2128 * B_CACHE bit is clear.
2130 * What this means, basically, is that the caller should use B_CACHE to
2131 * determine whether the buffer is fully valid or not and should clear
2132 * B_INVAL prior to issuing a read. If the caller intends to validate
2133 * the buffer by loading its data area with something, the caller needs
2134 * to clear B_INVAL. If the caller does this without issuing an I/O,
2135 * the caller should set B_CACHE ( as an optimization ), else the caller
2136 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2137 * a write attempt or if it was a successfull read. If the caller
2138 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2139 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2142 getblk(struct vnode *vp, off_t loffset, int size, int slpflag, int slptimeo)
2146 if (size > MAXBSIZE)
2147 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2148 if (vp->v_object == NULL)
2149 panic("getblk: vnode %p has no object!", vp);
2154 * Block if we are low on buffers. Certain processes are allowed
2155 * to completely exhaust the buffer cache.
2157 * If this check ever becomes a bottleneck it may be better to
2158 * move it into the else, when findblk() fails. At the moment
2159 * it isn't a problem.
2161 * XXX remove, we cannot afford to block anywhere if holding a vnode
2162 * lock in low-memory situation, so take it to the max.
2164 if (numfreebuffers == 0) {
2167 needsbuffer |= VFS_BIO_NEED_ANY;
2168 tsleep(&needsbuffer, slpflag, "newbuf", slptimeo);
2171 if ((bp = findblk(vp, loffset))) {
2173 * The buffer was found in the cache, but we need to lock it.
2174 * Even with LK_NOWAIT the lockmgr may break our critical
2175 * section, so double-check the validity of the buffer
2176 * once the lock has been obtained.
2178 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2179 int lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2180 if (slpflag & PCATCH)
2181 lkflags |= LK_PCATCH;
2182 if (BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo) ==
2191 * Once the buffer has been locked, make sure we didn't race
2192 * a buffer recyclement. Buffers that are no longer hashed
2193 * will have b_vp == NULL, so this takes care of that check
2196 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2197 printf("Warning buffer %p (vp %p loffset %lld) was recycled\n", bp, vp, loffset);
2203 * All vnode-based buffers must be backed by a VM object.
2205 KKASSERT(bp->b_flags & B_VMIO);
2206 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2209 * Make sure that B_INVAL buffers do not have a cached
2210 * block number translation.
2212 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2213 printf("Warning invalid buffer %p (vp %p loffset %lld) did not have cleared bio_offset cache\n", bp, vp, loffset);
2214 clearbiocache(&bp->b_bio2);
2218 * The buffer is locked. B_CACHE is cleared if the buffer is
2221 if (bp->b_flags & B_INVAL)
2222 bp->b_flags &= ~B_CACHE;
2226 * Any size inconsistancy with a dirty buffer or a buffer
2227 * with a softupdates dependancy must be resolved. Resizing
2228 * the buffer in such circumstances can lead to problems.
2230 if (size != bp->b_bcount) {
2231 if (bp->b_flags & B_DELWRI) {
2232 bp->b_flags |= B_NOCACHE;
2234 } else if (LIST_FIRST(&bp->b_dep)) {
2235 bp->b_flags |= B_NOCACHE;
2238 bp->b_flags |= B_RELBUF;
2243 KKASSERT(size <= bp->b_kvasize);
2244 KASSERT(bp->b_loffset != NOOFFSET,
2245 ("getblk: no buffer offset"));
2248 * A buffer with B_DELWRI set and B_CACHE clear must
2249 * be committed before we can return the buffer in
2250 * order to prevent the caller from issuing a read
2251 * ( due to B_CACHE not being set ) and overwriting
2254 * Most callers, including NFS and FFS, need this to
2255 * operate properly either because they assume they
2256 * can issue a read if B_CACHE is not set, or because
2257 * ( for example ) an uncached B_DELWRI might loop due
2258 * to softupdates re-dirtying the buffer. In the latter
2259 * case, B_CACHE is set after the first write completes,
2260 * preventing further loops.
2262 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2263 * above while extending the buffer, we cannot allow the
2264 * buffer to remain with B_CACHE set after the write
2265 * completes or it will represent a corrupt state. To
2266 * deal with this we set B_NOCACHE to scrap the buffer
2269 * We might be able to do something fancy, like setting
2270 * B_CACHE in bwrite() except if B_DELWRI is already set,
2271 * so the below call doesn't set B_CACHE, but that gets real
2272 * confusing. This is much easier.
2275 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2276 bp->b_flags |= B_NOCACHE;
2283 * Buffer is not in-core, create new buffer. The buffer
2284 * returned by getnewbuf() is locked. Note that the returned
2285 * buffer is also considered valid (not marked B_INVAL).
2287 * Calculating the offset for the I/O requires figuring out
2288 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2289 * the mount's f_iosize otherwise. If the vnode does not
2290 * have an associated mount we assume that the passed size is
2293 * Note that vn_isdisk() cannot be used here since it may
2294 * return a failure for numerous reasons. Note that the
2295 * buffer size may be larger then the block size (the caller
2296 * will use block numbers with the proper multiple). Beware
2297 * of using any v_* fields which are part of unions. In
2298 * particular, in DragonFly the mount point overloading
2299 * mechanism is such that the underlying directory (with a
2300 * non-NULL v_mountedhere) is not a special case.
2304 if (vp->v_type == VBLK || vp->v_type == VCHR)
2306 else if (vp->v_mount)
2307 bsize = vp->v_mount->mnt_stat.f_iosize;
2311 maxsize = size + (loffset & PAGE_MASK);
2312 maxsize = imax(maxsize, bsize);
2314 if ((bp = getnewbuf(slpflag, slptimeo, size, maxsize)) == NULL) {
2315 if (slpflag || slptimeo) {
2323 * This code is used to make sure that a buffer is not
2324 * created while the getnewbuf routine is blocked.
2325 * This can be a problem whether the vnode is locked or not.
2326 * If the buffer is created out from under us, we have to
2327 * throw away the one we just created. There is now window
2328 * race because we are safely running in a critical section
2329 * from the point of the duplicate buffer creation through
2330 * to here, and we've locked the buffer.
2332 if (findblk(vp, loffset)) {
2333 bp->b_flags |= B_INVAL;
2339 * Insert the buffer into the hash, so that it can
2340 * be found by findblk().
2342 * Make sure the translation layer has been cleared.
2344 bp->b_loffset = loffset;
2345 bp->b_bio2.bio_offset = NOOFFSET;
2346 /* bp->b_bio2.bio_next = NULL; */
2351 * All vnode-based buffers must be backed by a VM object.
2353 KKASSERT(vp->v_object != NULL);
2354 bp->b_flags |= B_VMIO;
2355 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2367 * Get an empty, disassociated buffer of given size. The buffer is
2368 * initially set to B_INVAL.
2370 * critical section protection is not required for the allocbuf()
2371 * call because races are impossible here.
2379 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2382 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2386 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2394 * This code constitutes the buffer memory from either anonymous system
2395 * memory (in the case of non-VMIO operations) or from an associated
2396 * VM object (in the case of VMIO operations). This code is able to
2397 * resize a buffer up or down.
2399 * Note that this code is tricky, and has many complications to resolve
2400 * deadlock or inconsistant data situations. Tread lightly!!!
2401 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2402 * the caller. Calling this code willy nilly can result in the loss of data.
2404 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2405 * B_CACHE for the non-VMIO case.
2407 * This routine does not need to be called from a critical section but you
2408 * must own the buffer.
2411 allocbuf(struct buf *bp, int size)
2413 int newbsize, mbsize;
2416 if (BUF_REFCNT(bp) == 0)
2417 panic("allocbuf: buffer not busy");
2419 if (bp->b_kvasize < size)
2420 panic("allocbuf: buffer too small");
2422 if ((bp->b_flags & B_VMIO) == 0) {
2426 * Just get anonymous memory from the kernel. Don't
2427 * mess with B_CACHE.
2429 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2430 if (bp->b_flags & B_MALLOC)
2433 newbsize = round_page(size);
2435 if (newbsize < bp->b_bufsize) {
2437 * Malloced buffers are not shrunk
2439 if (bp->b_flags & B_MALLOC) {
2441 bp->b_bcount = size;
2443 free(bp->b_data, M_BIOBUF);
2444 if (bp->b_bufsize) {
2445 bufmallocspace -= bp->b_bufsize;
2449 bp->b_data = bp->b_kvabase;
2451 bp->b_flags &= ~B_MALLOC;
2457 (vm_offset_t) bp->b_data + newbsize,
2458 (vm_offset_t) bp->b_data + bp->b_bufsize);
2459 } else if (newbsize > bp->b_bufsize) {
2461 * We only use malloced memory on the first allocation.
2462 * and revert to page-allocated memory when the buffer
2465 if ((bufmallocspace < maxbufmallocspace) &&
2466 (bp->b_bufsize == 0) &&
2467 (mbsize <= PAGE_SIZE/2)) {
2469 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK);
2470 bp->b_bufsize = mbsize;
2471 bp->b_bcount = size;
2472 bp->b_flags |= B_MALLOC;
2473 bufmallocspace += mbsize;
2479 * If the buffer is growing on its other-than-first
2480 * allocation, then we revert to the page-allocation
2483 if (bp->b_flags & B_MALLOC) {
2484 origbuf = bp->b_data;
2485 origbufsize = bp->b_bufsize;
2486 bp->b_data = bp->b_kvabase;
2487 if (bp->b_bufsize) {
2488 bufmallocspace -= bp->b_bufsize;
2492 bp->b_flags &= ~B_MALLOC;
2493 newbsize = round_page(newbsize);
2497 (vm_offset_t) bp->b_data + bp->b_bufsize,
2498 (vm_offset_t) bp->b_data + newbsize);
2500 bcopy(origbuf, bp->b_data, origbufsize);
2501 free(origbuf, M_BIOBUF);
2508 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2509 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
2510 newbsize + PAGE_MASK) >> PAGE_SHIFT;
2511 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
2513 if (bp->b_flags & B_MALLOC)
2514 panic("allocbuf: VMIO buffer can't be malloced");
2516 * Set B_CACHE initially if buffer is 0 length or will become
2519 if (size == 0 || bp->b_bufsize == 0)
2520 bp->b_flags |= B_CACHE;
2522 if (newbsize < bp->b_bufsize) {
2524 * DEV_BSIZE aligned new buffer size is less then the
2525 * DEV_BSIZE aligned existing buffer size. Figure out
2526 * if we have to remove any pages.
2528 if (desiredpages < bp->b_xio.xio_npages) {
2529 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
2531 * the page is not freed here -- it
2532 * is the responsibility of
2533 * vnode_pager_setsize
2535 m = bp->b_xio.xio_pages[i];
2536 KASSERT(m != bogus_page,
2537 ("allocbuf: bogus page found"));
2538 while (vm_page_sleep_busy(m, TRUE, "biodep"))
2541 bp->b_xio.xio_pages[i] = NULL;
2542 vm_page_unwire(m, 0);
2544 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2545 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
2546 bp->b_xio.xio_npages = desiredpages;
2548 } else if (size > bp->b_bcount) {
2550 * We are growing the buffer, possibly in a
2551 * byte-granular fashion.
2559 * Step 1, bring in the VM pages from the object,
2560 * allocating them if necessary. We must clear
2561 * B_CACHE if these pages are not valid for the
2562 * range covered by the buffer.
2564 * critical section protection is required to protect
2565 * against interrupts unbusying and freeing pages
2566 * between our vm_page_lookup() and our
2567 * busycheck/wiring call.
2573 while (bp->b_xio.xio_npages < desiredpages) {
2577 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
2578 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2580 * note: must allocate system pages
2581 * since blocking here could intefere
2582 * with paging I/O, no matter which
2585 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
2588 vm_pageout_deficit += desiredpages -
2589 bp->b_xio.xio_npages;
2593 bp->b_flags &= ~B_CACHE;
2594 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2595 ++bp->b_xio.xio_npages;
2601 * We found a page. If we have to sleep on it,
2602 * retry because it might have gotten freed out
2605 * We can only test PG_BUSY here. Blocking on
2606 * m->busy might lead to a deadlock:
2608 * vm_fault->getpages->cluster_read->allocbuf
2612 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
2616 * We have a good page. Should we wakeup the
2619 if ((curthread != pagethread) &&
2620 ((m->queue - m->pc) == PQ_CACHE) &&
2621 ((vmstats.v_free_count + vmstats.v_cache_count) <
2622 (vmstats.v_free_min + vmstats.v_cache_min))) {
2623 pagedaemon_wakeup();
2625 vm_page_flag_clear(m, PG_ZERO);
2627 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2628 ++bp->b_xio.xio_npages;
2633 * Step 2. We've loaded the pages into the buffer,
2634 * we have to figure out if we can still have B_CACHE
2635 * set. Note that B_CACHE is set according to the
2636 * byte-granular range ( bcount and size ), not the
2637 * aligned range ( newbsize ).
2639 * The VM test is against m->valid, which is DEV_BSIZE
2640 * aligned. Needless to say, the validity of the data
2641 * needs to also be DEV_BSIZE aligned. Note that this
2642 * fails with NFS if the server or some other client
2643 * extends the file's EOF. If our buffer is resized,
2644 * B_CACHE may remain set! XXX
2647 toff = bp->b_bcount;
2648 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
2650 while ((bp->b_flags & B_CACHE) && toff < size) {
2653 if (tinc > (size - toff))
2656 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
2664 bp->b_xio.xio_pages[pi]
2671 * Step 3, fixup the KVM pmap. Remember that
2672 * bp->b_data is relative to bp->b_loffset, but
2673 * bp->b_loffset may be offset into the first page.
2676 bp->b_data = (caddr_t)
2677 trunc_page((vm_offset_t)bp->b_data);
2679 (vm_offset_t)bp->b_data,
2680 bp->b_xio.xio_pages,
2681 bp->b_xio.xio_npages
2683 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2684 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
2687 if (newbsize < bp->b_bufsize)
2689 bp->b_bufsize = newbsize; /* actual buffer allocation */
2690 bp->b_bcount = size; /* requested buffer size */
2697 * Wait for buffer I/O completion, returning error status. The buffer
2698 * is left locked on return. B_EINTR is converted into an EINTR error
2701 * NOTE! The original b_cmd is lost on return, since b_cmd will be
2702 * set to BUF_CMD_DONE.
2705 biowait(struct buf *bp)
2708 while (bp->b_cmd != BUF_CMD_DONE) {
2709 if (bp->b_cmd == BUF_CMD_READ)
2710 tsleep(bp, 0, "biord", 0);
2712 tsleep(bp, 0, "biowr", 0);
2715 if (bp->b_flags & B_EINTR) {
2716 bp->b_flags &= ~B_EINTR;
2719 if (bp->b_flags & B_ERROR) {
2720 return (bp->b_error ? bp->b_error : EIO);
2727 * This associates a tracking count with an I/O. vn_strategy() and
2728 * dev_dstrategy() do this automatically but there are a few cases
2729 * where a vnode or device layer is bypassed when a block translation
2730 * is cached. In such cases bio_start_transaction() may be called on
2731 * the bypassed layers so the system gets an I/O in progress indication
2732 * for those higher layers.
2735 bio_start_transaction(struct bio *bio, struct bio_track *track)
2737 bio->bio_track = track;
2738 atomic_add_int(&track->bk_active, 1);
2742 * Initiate I/O on a vnode.
2745 vn_strategy(struct vnode *vp, struct bio *bio)
2747 struct bio_track *track;
2749 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
2750 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
2751 track = &vp->v_track_read;
2753 track = &vp->v_track_write;
2754 bio->bio_track = track;
2755 atomic_add_int(&track->bk_active, 1);
2756 vop_strategy(*vp->v_ops, vp, bio);
2763 * Finish I/O on a buffer, optionally calling a completion function.
2764 * This is usually called from an interrupt so process blocking is
2767 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
2768 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
2769 * assuming B_INVAL is clear.
2771 * For the VMIO case, we set B_CACHE if the op was a read and no
2772 * read error occured, or if the op was a write. B_CACHE is never
2773 * set if the buffer is invalid or otherwise uncacheable.
2775 * biodone does not mess with B_INVAL, allowing the I/O routine or the
2776 * initiator to leave B_INVAL set to brelse the buffer out of existance
2777 * in the biodone routine.
2780 biodone(struct bio *bio)
2782 struct buf *bp = bio->bio_buf;
2787 KASSERT(BUF_REFCNTNB(bp) > 0,
2788 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
2789 KASSERT(bp->b_cmd != BUF_CMD_DONE,
2790 ("biodone: bp %p already done!", bp));
2792 runningbufwakeup(bp);
2795 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
2798 biodone_t *done_func;
2799 struct bio_track *track;
2802 * BIO tracking. Most but not all BIOs are tracked.
2804 if ((track = bio->bio_track) != NULL) {
2805 atomic_subtract_int(&track->bk_active, 1);
2806 if (track->bk_active < 0) {
2807 panic("biodone: bad active count bio %p\n",
2810 if (track->bk_waitflag) {
2811 track->bk_waitflag = 0;
2814 bio->bio_track = NULL;
2818 * A bio_done function terminates the loop. The function
2819 * will be responsible for any further chaining and/or
2820 * buffer management.
2822 * WARNING! The done function can deallocate the buffer!
2824 if ((done_func = bio->bio_done) != NULL) {
2825 bio->bio_done = NULL;
2830 bio = bio->bio_prev;
2834 bp->b_cmd = BUF_CMD_DONE;
2837 * Only reads and writes are processed past this point.
2839 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
2846 * Warning: softupdates may re-dirty the buffer.
2848 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete)
2849 (*bioops.io_complete)(bp);
2851 if (bp->b_flags & B_VMIO) {
2857 struct vnode *vp = bp->b_vp;
2861 #if defined(VFS_BIO_DEBUG)
2862 if (vp->v_holdcnt == 0)
2863 panic("biodone: zero vnode hold count");
2864 if ((vp->v_flag & VOBJBUF) == 0)
2865 panic("biodone: vnode is not setup for merged cache");
2868 foff = bp->b_loffset;
2869 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
2870 KASSERT(obj != NULL, ("biodone: missing VM object"));
2872 #if defined(VFS_BIO_DEBUG)
2873 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
2874 printf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
2875 obj->paging_in_progress, bp->b_xio.xio_npages);
2880 * Set B_CACHE if the op was a normal read and no error
2881 * occured. B_CACHE is set for writes in the b*write()
2884 iosize = bp->b_bcount - bp->b_resid;
2885 if (cmd == BUF_CMD_READ && (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
2886 bp->b_flags |= B_CACHE;
2889 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2893 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
2898 * cleanup bogus pages, restoring the originals. Since
2899 * the originals should still be wired, we don't have
2900 * to worry about interrupt/freeing races destroying
2901 * the VM object association.
2903 m = bp->b_xio.xio_pages[i];
2904 if (m == bogus_page) {
2906 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
2908 panic("biodone: page disappeared");
2909 bp->b_xio.xio_pages[i] = m;
2910 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
2911 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
2913 #if defined(VFS_BIO_DEBUG)
2914 if (OFF_TO_IDX(foff) != m->pindex) {
2916 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
2917 (unsigned long)foff, m->pindex);
2922 * In the write case, the valid and clean bits are
2923 * already changed correctly ( see bdwrite() ), so we
2924 * only need to do this here in the read case.
2926 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
2927 vfs_page_set_valid(bp, foff, i, m);
2929 vm_page_flag_clear(m, PG_ZERO);
2932 * when debugging new filesystems or buffer I/O methods, this
2933 * is the most common error that pops up. if you see this, you
2934 * have not set the page busy flag correctly!!!
2937 printf("biodone: page busy < 0, "
2938 "pindex: %d, foff: 0x(%x,%x), "
2939 "resid: %d, index: %d\n",
2940 (int) m->pindex, (int)(foff >> 32),
2941 (int) foff & 0xffffffff, resid, i);
2942 if (!vn_isdisk(vp, NULL))
2943 printf(" iosize: %ld, loffset: %lld, flags: 0x%08x, npages: %d\n",
2944 bp->b_vp->v_mount->mnt_stat.f_iosize,
2946 bp->b_flags, bp->b_xio.xio_npages);
2948 printf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
2950 bp->b_flags, bp->b_xio.xio_npages);
2951 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
2952 m->valid, m->dirty, m->wire_count);
2953 panic("biodone: page busy < 0");
2955 vm_page_io_finish(m);
2956 vm_object_pip_subtract(obj, 1);
2957 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
2961 vm_object_pip_wakeupn(obj, 0);
2965 * For asynchronous completions, release the buffer now. The brelse
2966 * will do a wakeup there if necessary - so no need to do a wakeup
2967 * here in the async case. The sync case always needs to do a wakeup.
2970 if (bp->b_flags & B_ASYNC) {
2971 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
2984 * This routine is called in lieu of iodone in the case of
2985 * incomplete I/O. This keeps the busy status for pages
2989 vfs_unbusy_pages(struct buf *bp)
2993 runningbufwakeup(bp);
2994 if (bp->b_flags & B_VMIO) {
2995 struct vnode *vp = bp->b_vp;
3000 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3001 vm_page_t m = bp->b_xio.xio_pages[i];
3004 * When restoring bogus changes the original pages
3005 * should still be wired, so we are in no danger of
3006 * losing the object association and do not need
3007 * critical section protection particularly.
3009 if (m == bogus_page) {
3010 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3012 panic("vfs_unbusy_pages: page missing");
3014 bp->b_xio.xio_pages[i] = m;
3015 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3016 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3018 vm_object_pip_subtract(obj, 1);
3019 vm_page_flag_clear(m, PG_ZERO);
3020 vm_page_io_finish(m);
3022 vm_object_pip_wakeupn(obj, 0);
3027 * vfs_page_set_valid:
3029 * Set the valid bits in a page based on the supplied offset. The
3030 * range is restricted to the buffer's size.
3032 * This routine is typically called after a read completes.
3035 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3037 vm_ooffset_t soff, eoff;
3040 * Start and end offsets in buffer. eoff - soff may not cross a
3041 * page boundry or cross the end of the buffer. The end of the
3042 * buffer, in this case, is our file EOF, not the allocation size
3046 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3047 if (eoff > bp->b_loffset + bp->b_bcount)
3048 eoff = bp->b_loffset + bp->b_bcount;
3051 * Set valid range. This is typically the entire buffer and thus the
3055 vm_page_set_validclean(
3057 (vm_offset_t) (soff & PAGE_MASK),
3058 (vm_offset_t) (eoff - soff)
3066 * This routine is called before a device strategy routine.
3067 * It is used to tell the VM system that paging I/O is in
3068 * progress, and treat the pages associated with the buffer
3069 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3070 * flag is handled to make sure that the object doesn't become
3073 * Since I/O has not been initiated yet, certain buffer flags
3074 * such as B_ERROR or B_INVAL may be in an inconsistant state
3075 * and should be ignored.
3078 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3081 struct proc *p = curthread->td_proc;
3084 * The buffer's I/O command must already be set. If reading,
3085 * B_CACHE must be 0 (double check against callers only doing
3086 * I/O when B_CACHE is 0).
3088 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3089 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3091 if (bp->b_flags & B_VMIO) {
3096 foff = bp->b_loffset;
3097 KASSERT(bp->b_loffset != NOOFFSET,
3098 ("vfs_busy_pages: no buffer offset"));
3102 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3103 vm_page_t m = bp->b_xio.xio_pages[i];
3104 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3109 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3110 vm_page_t m = bp->b_xio.xio_pages[i];
3112 vm_page_flag_clear(m, PG_ZERO);
3113 if ((bp->b_flags & B_CLUSTER) == 0) {
3114 vm_object_pip_add(obj, 1);
3115 vm_page_io_start(m);
3119 * When readying a vnode-backed buffer for a write
3120 * we must zero-fill any invalid portions of the
3123 * When readying a vnode-backed buffer for a read
3124 * we must replace any dirty pages with a bogus
3125 * page so we do not destroy dirty data when
3126 * filling in gaps. Dirty pages might not
3127 * necessarily be marked dirty yet, so use m->valid
3128 * as a reasonable test.
3130 * Bogus page replacement is, uh, bogus. We need
3131 * to find a better way.
3133 vm_page_protect(m, VM_PROT_NONE);
3134 if (bp->b_cmd == BUF_CMD_WRITE) {
3135 vfs_page_set_valid(bp, foff, i, m);
3136 } else if (m->valid == VM_PAGE_BITS_ALL) {
3137 bp->b_xio.xio_pages[i] = bogus_page;
3140 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3143 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3144 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3148 * This is the easiest place to put the process accounting for the I/O
3152 if (bp->b_cmd == BUF_CMD_READ)
3153 p->p_stats->p_ru.ru_inblock++;
3155 p->p_stats->p_ru.ru_oublock++;
3162 * Tell the VM system that the pages associated with this buffer
3163 * are clean. This is used for delayed writes where the data is
3164 * going to go to disk eventually without additional VM intevention.
3166 * Note that while we only really need to clean through to b_bcount, we
3167 * just go ahead and clean through to b_bufsize.
3170 vfs_clean_pages(struct buf *bp)
3174 if (bp->b_flags & B_VMIO) {
3177 foff = bp->b_loffset;
3178 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset"));
3179 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3180 vm_page_t m = bp->b_xio.xio_pages[i];
3181 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3182 vm_ooffset_t eoff = noff;
3184 if (eoff > bp->b_loffset + bp->b_bufsize)
3185 eoff = bp->b_loffset + bp->b_bufsize;
3186 vfs_page_set_valid(bp, foff, i, m);
3187 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3194 * vfs_bio_set_validclean:
3196 * Set the range within the buffer to valid and clean. The range is
3197 * relative to the beginning of the buffer, b_loffset. Note that
3198 * b_loffset itself may be offset from the beginning of the first page.
3202 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3204 if (bp->b_flags & B_VMIO) {
3209 * Fixup base to be relative to beginning of first page.
3210 * Set initial n to be the maximum number of bytes in the
3211 * first page that can be validated.
3214 base += (bp->b_loffset & PAGE_MASK);
3215 n = PAGE_SIZE - (base & PAGE_MASK);
3217 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) {
3218 vm_page_t m = bp->b_xio.xio_pages[i];
3223 vm_page_set_validclean(m, base & PAGE_MASK, n);
3234 * Clear a buffer. This routine essentially fakes an I/O, so we need
3235 * to clear B_ERROR and B_INVAL.
3237 * Note that while we only theoretically need to clear through b_bcount,
3238 * we go ahead and clear through b_bufsize.
3242 vfs_bio_clrbuf(struct buf *bp)
3246 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3247 bp->b_flags &= ~(B_INVAL|B_ERROR);
3248 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3249 (bp->b_loffset & PAGE_MASK) == 0) {
3250 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3251 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3255 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3256 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3257 bzero(bp->b_data, bp->b_bufsize);
3258 bp->b_xio.xio_pages[0]->valid |= mask;
3263 ea = sa = bp->b_data;
3264 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3265 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3266 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3267 ea = (caddr_t)(vm_offset_t)ulmin(
3268 (u_long)(vm_offset_t)ea,
3269 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3270 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3271 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3273 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3274 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3278 for (; sa < ea; sa += DEV_BSIZE, j++) {
3279 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3280 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3281 bzero(sa, DEV_BSIZE);
3284 bp->b_xio.xio_pages[i]->valid |= mask;
3285 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
3294 * vm_hold_load_pages:
3296 * Load pages into the buffer's address space. The pages are
3297 * allocated from the kernel object in order to reduce interference
3298 * with the any VM paging I/O activity. The range of loaded
3299 * pages will be wired.
3301 * If a page cannot be allocated, the 'pagedaemon' is woken up to
3302 * retrieve the full range (to - from) of pages.
3306 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3312 to = round_page(to);
3313 from = round_page(from);
3314 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3316 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3321 * Note: must allocate system pages since blocking here
3322 * could intefere with paging I/O, no matter which
3325 p = vm_page_alloc(kernel_object,
3326 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT),
3327 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
3329 vm_pageout_deficit += (to - from) >> PAGE_SHIFT;
3334 p->valid = VM_PAGE_BITS_ALL;
3335 vm_page_flag_clear(p, PG_ZERO);
3336 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3337 bp->b_xio.xio_pages[index] = p;
3340 bp->b_xio.xio_npages = index;
3344 * vm_hold_free_pages:
3346 * Return pages associated with the buffer back to the VM system.
3348 * The range of pages underlying the buffer's address space will
3349 * be unmapped and un-wired.
3352 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3356 int index, newnpages;
3358 from = round_page(from);
3359 to = round_page(to);
3360 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3362 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3363 p = bp->b_xio.xio_pages[index];
3364 if (p && (index < bp->b_xio.xio_npages)) {
3366 printf("vm_hold_free_pages: doffset: %lld, loffset: %lld\n",
3367 bp->b_bio2.bio_offset, bp->b_loffset);
3369 bp->b_xio.xio_pages[index] = NULL;
3372 vm_page_unwire(p, 0);
3376 bp->b_xio.xio_npages = newnpages;
3382 * Map a user buffer into KVM via a pbuf. On return the buffer's
3383 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
3387 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
3397 * bp had better have a command and it better be a pbuf.
3399 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3400 KKASSERT(bp->b_flags & B_PAGING);
3406 * Map the user data into KVM. Mappings have to be page-aligned.
3408 addr = (caddr_t)trunc_page((vm_offset_t)udata);
3411 vmprot = VM_PROT_READ;
3412 if (bp->b_cmd == BUF_CMD_READ)
3413 vmprot |= VM_PROT_WRITE;
3415 while (addr < udata + bytes) {
3417 * Do the vm_fault if needed; do the copy-on-write thing
3418 * when reading stuff off device into memory.
3422 i = vm_fault_quick(addr, vmprot);
3424 i = vm_fault_quick(udata, vmprot);
3426 for (i = 0; i < pidx; ++i) {
3427 vm_page_unhold(bp->b_xio.xio_pages[i]);
3428 bp->b_xio.xio_pages[i] = NULL;
3434 * Extract from current process's address map. Since the
3435 * fault succeeded, an empty page indicates a race.
3437 pa = pmap_extract(&curproc->p_vmspace->vm_pmap, (vm_offset_t)addr);
3439 printf("vmapbuf: warning, race against user address during I/O");
3442 m = PHYS_TO_VM_PAGE(pa);
3444 bp->b_xio.xio_pages[pidx] = m;
3450 * Map the page array and set the buffer fields to point to
3451 * the mapped data buffer.
3453 if (pidx > btoc(MAXPHYS))
3454 panic("vmapbuf: mapped more than MAXPHYS");
3455 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
3457 bp->b_xio.xio_npages = pidx;
3458 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
3459 bp->b_bcount = bytes;
3460 bp->b_bufsize = bytes;
3467 * Free the io map PTEs associated with this IO operation.
3468 * We also invalidate the TLB entries and restore the original b_addr.
3471 vunmapbuf(struct buf *bp)
3476 KKASSERT(bp->b_flags & B_PAGING);
3478 npages = bp->b_xio.xio_npages;
3479 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3480 for (pidx = 0; pidx < npages; ++pidx) {
3481 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
3482 bp->b_xio.xio_pages[pidx] = NULL;
3484 bp->b_xio.xio_npages = 0;
3485 bp->b_data = bp->b_kvabase;
3489 * Scan all buffers in the system and issue the callback.
3492 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
3498 for (n = 0; n < nbuf; ++n) {
3499 if ((error = callback(&buf[n], info)) < 0) {
3509 * print out statistics from the current status of the buffer pool
3510 * this can be toggeled by the system control option debug.syncprt
3519 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
3520 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
3522 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
3524 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
3527 TAILQ_FOREACH(bp, dp, b_freelist) {
3528 counts[bp->b_bufsize/PAGE_SIZE]++;
3532 printf("%s: total-%d", bname[i], count);
3533 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
3535 printf(", %d-%d", j * PAGE_SIZE, counts[j]);
3543 DB_SHOW_COMMAND(buffer, db_show_buffer)
3546 struct buf *bp = (struct buf *)addr;
3549 db_printf("usage: show buffer <addr>\n");
3553 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3554 db_printf("b_cmd = %d\n", bp->b_cmd);
3555 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
3556 "b_resid = %d\n, b_data = %p, "
3557 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
3558 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3559 bp->b_data, bp->b_bio2.bio_offset, (bp->b_bio2.bio_next ? bp->b_bio2.bio_next->bio_offset : (off_t)-1));
3560 if (bp->b_xio.xio_npages) {
3562 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
3563 bp->b_xio.xio_npages);
3564 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3566 m = bp->b_xio.xio_pages[i];
3567 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3568 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3569 if ((i + 1) < bp->b_xio.xio_npages)