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 1.96 2008/01/10 07:34:01 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 <sys/spinlock2.h>
60 #include <vm/vm_page2.h>
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_DIRTY_HW, /* B_DELWRI buffers - heavy weight */
76 BQUEUE_EMPTYKVA, /* empty buffer headers with KVA assignment */
77 BQUEUE_EMPTY, /* empty buffer headers */
79 BUFFER_QUEUES /* number of buffer queues */
82 typedef enum bufq_type bufq_type_t;
84 TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
86 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");
88 struct buf *buf; /* buffer header pool */
90 static void vm_hold_free_pages(struct buf *bp, vm_offset_t from,
92 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from,
94 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off,
95 int pageno, vm_page_t m);
96 static void vfs_clean_pages(struct buf *bp);
97 static void vfs_setdirty(struct buf *bp);
98 static void vfs_vmio_release(struct buf *bp);
99 static int flushbufqueues(bufq_type_t q);
101 static void buf_daemon(void);
102 static void buf_daemon_hw(void);
104 * bogus page -- for I/O to/from partially complete buffers
105 * this is a temporary solution to the problem, but it is not
106 * really that bad. it would be better to split the buffer
107 * for input in the case of buffers partially already in memory,
108 * but the code is intricate enough already.
110 vm_page_t bogus_page;
114 * These are all static, but make the ones we export globals so we do
115 * not need to use compiler magic.
117 int bufspace, maxbufspace,
118 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
119 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
120 static int lorunningspace, hirunningspace, runningbufreq;
121 int numdirtybuffers, numdirtybuffershw, lodirtybuffers, hidirtybuffers;
122 static int numfreebuffers, lofreebuffers, hifreebuffers;
123 static int getnewbufcalls;
124 static int getnewbufrestarts;
126 static int needsbuffer; /* locked by needsbuffer_spin */
127 static int bd_request; /* locked by needsbuffer_spin */
128 static int bd_request_hw; /* locked by needsbuffer_spin */
129 static struct spinlock needsbuffer_spin;
132 * Sysctls for operational control of the buffer cache.
134 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, &lodirtybuffers, 0,
135 "Number of dirty buffers to flush before bufdaemon becomes inactive");
136 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, &hidirtybuffers, 0,
137 "High watermark used to trigger explicit flushing of dirty buffers");
138 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, &lofreebuffers, 0,
139 "Low watermark for special reserve in low-memory situations");
140 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, &hifreebuffers, 0,
141 "High watermark for special reserve in low-memory situations");
142 SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
143 "Minimum amount of buffer space required for active I/O");
144 SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
145 "Maximum amount of buffer space to usable for active I/O");
147 * Sysctls determining current state of the buffer cache.
149 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
150 "Pending number of dirty buffers (all)");
151 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffershw, CTLFLAG_RD, &numdirtybuffershw, 0,
152 "Pending number of dirty buffers (heavy weight)");
153 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
154 "Number of free buffers on the buffer cache free list");
155 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
156 "I/O bytes currently in progress due to asynchronous writes");
157 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
158 "Hard limit on maximum amount of memory usable for buffer space");
159 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
160 "Soft limit on maximum amount of memory usable for buffer space");
161 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
162 "Minimum amount of memory to reserve for system buffer space");
163 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
164 "Amount of memory available for buffers");
165 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
166 0, "Maximum amount of memory reserved for buffers using malloc");
167 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
168 "Amount of memory left for buffers using malloc-scheme");
169 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
170 "New buffer header acquisition requests");
171 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
172 0, "New buffer header acquisition restarts");
173 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
174 "Buffer acquisition restarts due to fragmented buffer map");
175 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
176 "Amount of time KVA space was deallocated in an arbitrary buffer");
177 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
178 "Amount of time buffer re-use operations were successful");
179 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
180 "sizeof(struct buf)");
182 char *buf_wmesg = BUF_WMESG;
184 extern int vm_swap_size;
186 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
187 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
188 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
189 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
194 * If someone is blocked due to there being too many dirty buffers,
195 * and numdirtybuffers is now reasonable, wake them up.
200 if (numdirtybuffers <= (lodirtybuffers + hidirtybuffers) / 2) {
201 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
202 spin_lock_wr(&needsbuffer_spin);
203 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
204 spin_unlock_wr(&needsbuffer_spin);
205 wakeup(&needsbuffer);
213 * Called when buffer space is potentially available for recovery.
214 * getnewbuf() will block on this flag when it is unable to free
215 * sufficient buffer space. Buffer space becomes recoverable when
216 * bp's get placed back in the queues.
223 * If someone is waiting for BUF space, wake them up. Even
224 * though we haven't freed the kva space yet, the waiting
225 * process will be able to now.
227 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
228 spin_lock_wr(&needsbuffer_spin);
229 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
230 spin_unlock_wr(&needsbuffer_spin);
231 wakeup(&needsbuffer);
238 * Accounting for I/O in progress.
242 runningbufwakeup(struct buf *bp)
244 if (bp->b_runningbufspace) {
245 runningbufspace -= bp->b_runningbufspace;
246 bp->b_runningbufspace = 0;
247 if (runningbufreq && runningbufspace <= lorunningspace) {
249 wakeup(&runningbufreq);
257 * Called when a buffer has been added to one of the free queues to
258 * account for the buffer and to wakeup anyone waiting for free buffers.
259 * This typically occurs when large amounts of metadata are being handled
260 * by the buffer cache ( else buffer space runs out first, usually ).
268 spin_lock_wr(&needsbuffer_spin);
269 needsbuffer &= ~VFS_BIO_NEED_ANY;
270 if (numfreebuffers >= hifreebuffers)
271 needsbuffer &= ~VFS_BIO_NEED_FREE;
272 spin_unlock_wr(&needsbuffer_spin);
273 wakeup(&needsbuffer);
278 * waitrunningbufspace()
280 * runningbufspace is a measure of the amount of I/O currently
281 * running. This routine is used in async-write situations to
282 * prevent creating huge backups of pending writes to a device.
283 * Only asynchronous writes are governed by this function.
285 * Reads will adjust runningbufspace, but will not block based on it.
286 * The read load has a side effect of reducing the allowed write load.
288 * This does NOT turn an async write into a sync write. It waits
289 * for earlier writes to complete and generally returns before the
290 * caller's write has reached the device.
293 waitrunningbufspace(void)
295 if (runningbufspace > hirunningspace) {
297 while (runningbufspace > hirunningspace) {
299 tsleep(&runningbufreq, 0, "wdrain", 0);
306 * vfs_buf_test_cache:
308 * Called when a buffer is extended. This function clears the B_CACHE
309 * bit if the newly extended portion of the buffer does not contain
314 vfs_buf_test_cache(struct buf *bp,
315 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
318 if (bp->b_flags & B_CACHE) {
319 int base = (foff + off) & PAGE_MASK;
320 if (vm_page_is_valid(m, base, size) == 0)
321 bp->b_flags &= ~B_CACHE;
328 * Wake up the buffer daemon if the number of outstanding dirty buffers
329 * is above specified threshold 'dirtybuflevel'.
331 * The buffer daemons are explicitly woken up when (a) the pending number
332 * of dirty buffers exceeds the recovery and stall mid-point value,
333 * (b) during bwillwrite() or (c) buf freelist was exhausted.
335 * The buffer daemons will generally not stop flushing until the dirty
336 * buffer count goes below lodirtybuffers.
340 bd_wakeup(int dirtybuflevel)
342 if (bd_request == 0 && numdirtybuffers >= dirtybuflevel) {
343 spin_lock_wr(&needsbuffer_spin);
345 spin_unlock_wr(&needsbuffer_spin);
348 if (bd_request_hw == 0 && numdirtybuffershw >= dirtybuflevel) {
349 spin_lock_wr(&needsbuffer_spin);
351 spin_unlock_wr(&needsbuffer_spin);
352 wakeup(&bd_request_hw);
359 * Speed up the buffer cache flushing process.
372 * Load time initialisation of the buffer cache, called from machine
373 * dependant initialization code.
379 vm_offset_t bogus_offset;
382 spin_init(&needsbuffer_spin);
384 /* next, make a null set of free lists */
385 for (i = 0; i < BUFFER_QUEUES; i++)
386 TAILQ_INIT(&bufqueues[i]);
388 /* finally, initialize each buffer header and stick on empty q */
389 for (i = 0; i < nbuf; i++) {
391 bzero(bp, sizeof *bp);
392 bp->b_flags = B_INVAL; /* we're just an empty header */
393 bp->b_cmd = BUF_CMD_DONE;
394 bp->b_qindex = BQUEUE_EMPTY;
396 xio_init(&bp->b_xio);
399 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
403 * maxbufspace is the absolute maximum amount of buffer space we are
404 * allowed to reserve in KVM and in real terms. The absolute maximum
405 * is nominally used by buf_daemon. hibufspace is the nominal maximum
406 * used by most other processes. The differential is required to
407 * ensure that buf_daemon is able to run when other processes might
408 * be blocked waiting for buffer space.
410 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
411 * this may result in KVM fragmentation which is not handled optimally
414 maxbufspace = nbuf * BKVASIZE;
415 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
416 lobufspace = hibufspace - MAXBSIZE;
418 lorunningspace = 512 * 1024;
419 hirunningspace = 1024 * 1024;
422 * Limit the amount of malloc memory since it is wired permanently into
423 * the kernel space. Even though this is accounted for in the buffer
424 * allocation, we don't want the malloced region to grow uncontrolled.
425 * The malloc scheme improves memory utilization significantly on average
426 * (small) directories.
428 maxbufmallocspace = hibufspace / 20;
431 * Reduce the chance of a deadlock occuring by limiting the number
432 * of delayed-write dirty buffers we allow to stack up.
434 hidirtybuffers = nbuf / 4 + 20;
436 numdirtybuffershw = 0;
438 * To support extreme low-memory systems, make sure hidirtybuffers cannot
439 * eat up all available buffer space. This occurs when our minimum cannot
440 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
441 * BKVASIZE'd (8K) buffers.
443 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
444 hidirtybuffers >>= 1;
446 lodirtybuffers = hidirtybuffers / 2;
449 * Try to keep the number of free buffers in the specified range,
450 * and give special processes (e.g. like buf_daemon) access to an
453 lofreebuffers = nbuf / 18 + 5;
454 hifreebuffers = 2 * lofreebuffers;
455 numfreebuffers = nbuf;
458 * Maximum number of async ops initiated per buf_daemon loop. This is
459 * somewhat of a hack at the moment, we really need to limit ourselves
460 * based on the number of bytes of I/O in-transit that were initiated
464 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
465 bogus_page = vm_page_alloc(&kernel_object,
466 (bogus_offset >> PAGE_SHIFT),
468 vmstats.v_wire_count++;
473 * Initialize the embedded bio structures
476 initbufbio(struct buf *bp)
478 bp->b_bio1.bio_buf = bp;
479 bp->b_bio1.bio_prev = NULL;
480 bp->b_bio1.bio_offset = NOOFFSET;
481 bp->b_bio1.bio_next = &bp->b_bio2;
482 bp->b_bio1.bio_done = NULL;
484 bp->b_bio2.bio_buf = bp;
485 bp->b_bio2.bio_prev = &bp->b_bio1;
486 bp->b_bio2.bio_offset = NOOFFSET;
487 bp->b_bio2.bio_next = NULL;
488 bp->b_bio2.bio_done = NULL;
492 * Reinitialize the embedded bio structures as well as any additional
493 * translation cache layers.
496 reinitbufbio(struct buf *bp)
500 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
501 bio->bio_done = NULL;
502 bio->bio_offset = NOOFFSET;
507 * Push another BIO layer onto an existing BIO and return it. The new
508 * BIO layer may already exist, holding cached translation data.
511 push_bio(struct bio *bio)
515 if ((nbio = bio->bio_next) == NULL) {
516 int index = bio - &bio->bio_buf->b_bio_array[0];
517 if (index >= NBUF_BIO - 1) {
518 panic("push_bio: too many layers bp %p\n",
521 nbio = &bio->bio_buf->b_bio_array[index + 1];
522 bio->bio_next = nbio;
523 nbio->bio_prev = bio;
524 nbio->bio_buf = bio->bio_buf;
525 nbio->bio_offset = NOOFFSET;
526 nbio->bio_done = NULL;
527 nbio->bio_next = NULL;
529 KKASSERT(nbio->bio_done == NULL);
534 pop_bio(struct bio *bio)
540 clearbiocache(struct bio *bio)
543 bio->bio_offset = NOOFFSET;
551 * Free the KVA allocation for buffer 'bp'.
553 * Must be called from a critical section as this is the only locking for
556 * Since this call frees up buffer space, we call bufspacewakeup().
559 bfreekva(struct buf *bp)
565 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
566 vm_map_lock(&buffer_map);
567 bufspace -= bp->b_kvasize;
568 vm_map_delete(&buffer_map,
569 (vm_offset_t) bp->b_kvabase,
570 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
573 vm_map_unlock(&buffer_map);
574 vm_map_entry_release(count);
583 * Remove the buffer from the appropriate free list.
586 bremfree(struct buf *bp)
591 old_qindex = bp->b_qindex;
593 if (bp->b_qindex != BQUEUE_NONE) {
594 KASSERT(BUF_REFCNTNB(bp) == 1,
595 ("bremfree: bp %p not locked",bp));
596 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
597 bp->b_qindex = BQUEUE_NONE;
599 if (BUF_REFCNTNB(bp) <= 1)
600 panic("bremfree: removing a buffer not on a queue");
604 * Fixup numfreebuffers count. If the buffer is invalid or not
605 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
606 * the buffer was free and we must decrement numfreebuffers.
608 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
611 case BQUEUE_DIRTY_HW:
614 case BQUEUE_EMPTYKVA:
628 * Get a buffer with the specified data. Look in the cache first. We
629 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
630 * is set, the buffer is valid and we do not have to do anything ( see
634 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
638 bp = getblk(vp, loffset, size, 0, 0);
641 /* if not found in cache, do some I/O */
642 if ((bp->b_flags & B_CACHE) == 0) {
643 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp));
644 bp->b_flags &= ~(B_ERROR | B_INVAL);
645 bp->b_cmd = BUF_CMD_READ;
646 vfs_busy_pages(vp, bp);
647 vn_strategy(vp, &bp->b_bio1);
648 return (biowait(bp));
656 * Operates like bread, but also starts asynchronous I/O on
657 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
658 * to initiating I/O . If B_CACHE is set, the buffer is valid
659 * and we do not have to do anything.
662 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
663 int *rabsize, int cnt, struct buf **bpp)
665 struct buf *bp, *rabp;
667 int rv = 0, readwait = 0;
669 *bpp = bp = getblk(vp, loffset, size, 0, 0);
671 /* if not found in cache, do some I/O */
672 if ((bp->b_flags & B_CACHE) == 0) {
673 bp->b_flags &= ~(B_ERROR | B_INVAL);
674 bp->b_cmd = BUF_CMD_READ;
675 vfs_busy_pages(vp, bp);
676 vn_strategy(vp, &bp->b_bio1);
680 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
681 if (inmem(vp, *raoffset))
683 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
685 if ((rabp->b_flags & B_CACHE) == 0) {
686 rabp->b_flags |= B_ASYNC;
687 rabp->b_flags &= ~(B_ERROR | B_INVAL);
688 rabp->b_cmd = BUF_CMD_READ;
689 vfs_busy_pages(vp, rabp);
691 vn_strategy(vp, &rabp->b_bio1);
706 * Write, release buffer on completion. (Done by iodone
707 * if async). Do not bother writing anything if the buffer
710 * Note that we set B_CACHE here, indicating that buffer is
711 * fully valid and thus cacheable. This is true even of NFS
712 * now so we set it generally. This could be set either here
713 * or in biodone() since the I/O is synchronous. We put it
717 bwrite(struct buf *bp)
721 if (bp->b_flags & B_INVAL) {
726 oldflags = bp->b_flags;
728 if (BUF_REFCNTNB(bp) == 0)
729 panic("bwrite: buffer is not busy???");
732 /* Mark the buffer clean */
735 bp->b_flags &= ~B_ERROR;
736 bp->b_flags |= B_CACHE;
737 bp->b_cmd = BUF_CMD_WRITE;
738 vfs_busy_pages(bp->b_vp, bp);
741 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
742 * valid for vnode-backed buffers.
744 bp->b_runningbufspace = bp->b_bufsize;
745 runningbufspace += bp->b_runningbufspace;
748 if (oldflags & B_ASYNC)
750 vn_strategy(bp->b_vp, &bp->b_bio1);
752 if ((oldflags & B_ASYNC) == 0) {
753 int rtval = biowait(bp);
756 } else if ((oldflags & B_NOWDRAIN) == 0) {
758 * don't allow the async write to saturate the I/O
759 * system. Deadlocks can occur only if a device strategy
760 * routine (like in VN) turns around and issues another
761 * high-level write, in which case B_NOWDRAIN is expected
762 * to be set. Otherwise we will not deadlock here because
763 * we are blocking waiting for I/O that is already in-progress
766 waitrunningbufspace();
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_REFCNTNB(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_bio2.bio_offset == NOOFFSET) {
811 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
816 * Set the *dirty* buffer range based upon the VM system dirty pages.
821 * We need to do this here to satisfy the vnode_pager and the
822 * pageout daemon, so that it thinks that the pages have been
823 * "cleaned". Note that since the pages are in a delayed write
824 * buffer -- the VFS layer "will" see that the pages get written
825 * out on the next sync, or perhaps the cluster will be completed.
831 * Wakeup the buffer flushing daemon if we have a lot of dirty
832 * buffers (midpoint between our recovery point and our stall
835 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
838 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
839 * due to the softdep code.
846 * Turn buffer into delayed write request by marking it B_DELWRI.
847 * B_RELBUF and B_NOCACHE must be cleared.
849 * We reassign the buffer to itself to properly update it in the
852 * Since the buffer is not on a queue, we do not update the
853 * numfreebuffers count.
855 * Must be called from a critical section.
856 * The buffer must be on BQUEUE_NONE.
859 bdirty(struct buf *bp)
861 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
862 if (bp->b_flags & B_NOCACHE) {
863 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
864 bp->b_flags &= ~B_NOCACHE;
866 if (bp->b_flags & B_INVAL) {
867 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
869 bp->b_flags &= ~B_RELBUF;
871 if ((bp->b_flags & B_DELWRI) == 0) {
872 bp->b_flags |= B_DELWRI;
875 if (bp->b_flags & B_HEAVY)
877 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
882 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
883 * needs to be flushed with a different buf_daemon thread to avoid
884 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
887 bheavy(struct buf *bp)
889 if ((bp->b_flags & B_HEAVY) == 0) {
890 bp->b_flags |= B_HEAVY;
891 if (bp->b_flags & B_DELWRI)
899 * Clear B_DELWRI for buffer.
901 * Since the buffer is not on a queue, we do not update the numfreebuffers
904 * Must be called from a critical section.
906 * The buffer is typically on BQUEUE_NONE but there is one case in
907 * brelse() that calls this function after placing the buffer on
912 bundirty(struct buf *bp)
914 if (bp->b_flags & B_DELWRI) {
915 bp->b_flags &= ~B_DELWRI;
918 if (bp->b_flags & B_HEAVY)
923 * Since it is now being written, we can clear its deferred write flag.
925 bp->b_flags &= ~B_DEFERRED;
931 * Asynchronous write. Start output on a buffer, but do not wait for
932 * it to complete. The buffer is released when the output completes.
934 * bwrite() ( or the VOP routine anyway ) is responsible for handling
935 * B_INVAL buffers. Not us.
938 bawrite(struct buf *bp)
940 bp->b_flags |= B_ASYNC;
947 * Ordered write. Start output on a buffer, and flag it so that the
948 * device will write it in the order it was queued. The buffer is
949 * released when the output completes. bwrite() ( or the VOP routine
950 * anyway ) is responsible for handling B_INVAL buffers.
953 bowrite(struct buf *bp)
955 bp->b_flags |= B_ORDERED | B_ASYNC;
962 * Called prior to the locking of any vnodes when we are expecting to
963 * write. We do not want to starve the buffer cache with too many
964 * dirty buffers so we block here. By blocking prior to the locking
965 * of any vnodes we attempt to avoid the situation where a locked vnode
966 * prevents the various system daemons from flushing related buffers.
971 if (numdirtybuffers >= hidirtybuffers) {
972 while (numdirtybuffers >= hidirtybuffers) {
974 spin_lock_wr(&needsbuffer_spin);
975 if (numdirtybuffers >= hidirtybuffers) {
976 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
977 msleep(&needsbuffer, &needsbuffer_spin, 0,
980 spin_unlock_wr(&needsbuffer_spin);
986 * buf_dirty_count_severe:
988 * Return true if we have too many dirty buffers.
991 buf_dirty_count_severe(void)
993 return(numdirtybuffers >= hidirtybuffers);
999 * Release a busy buffer and, if requested, free its resources. The
1000 * buffer will be stashed in the appropriate bufqueue[] allowing it
1001 * to be accessed later as a cache entity or reused for other purposes.
1004 brelse(struct buf *bp)
1007 int saved_flags = bp->b_flags;
1010 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1015 * If B_NOCACHE is set we are being asked to destroy the buffer and
1016 * its backing store. Clear B_DELWRI.
1018 * B_NOCACHE is set in two cases: (1) when the caller really wants
1019 * to destroy the buffer and backing store and (2) when the caller
1020 * wants to destroy the buffer and backing store after a write
1023 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1027 if (bp->b_flags & B_LOCKED)
1028 bp->b_flags &= ~B_ERROR;
1031 * If a write error occurs and the caller does not want to throw
1032 * away the buffer, redirty the buffer. This will also clear
1035 if (bp->b_cmd == BUF_CMD_WRITE &&
1036 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
1038 * Failed write, redirty. Must clear B_ERROR to prevent
1039 * pages from being scrapped. If B_INVAL is set then
1040 * this case is not run and the next case is run to
1041 * destroy the buffer. B_INVAL can occur if the buffer
1042 * is outside the range supported by the underlying device.
1044 bp->b_flags &= ~B_ERROR;
1046 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1047 (bp->b_bufsize <= 0) || bp->b_cmd == BUF_CMD_FREEBLKS) {
1049 * Either a failed I/O or we were asked to free or not
1052 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1053 * buffer cannot be immediately freed.
1055 bp->b_flags |= B_INVAL;
1056 if (LIST_FIRST(&bp->b_dep) != NULL)
1058 if (bp->b_flags & B_DELWRI) {
1060 if (bp->b_flags & B_HEAVY)
1061 --numdirtybuffershw;
1064 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1068 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1069 * If vfs_vmio_release() is called with either bit set, the
1070 * underlying pages may wind up getting freed causing a previous
1071 * write (bdwrite()) to get 'lost' because pages associated with
1072 * a B_DELWRI bp are marked clean. Pages associated with a
1073 * B_LOCKED buffer may be mapped by the filesystem.
1075 * If we want to release the buffer ourselves (rather then the
1076 * originator asking us to release it), give the originator a
1077 * chance to countermand the release by setting B_LOCKED.
1079 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1080 * if B_DELWRI is set.
1082 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1083 * on pages to return pages to the VM page queues.
1085 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1086 bp->b_flags &= ~B_RELBUF;
1087 } else if (vm_page_count_severe()) {
1089 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1090 bp->b_flags &= ~B_RELBUF;
1092 bp->b_flags |= B_RELBUF;
1096 * At this point destroying the buffer is governed by the B_INVAL
1097 * or B_RELBUF flags.
1099 bp->b_cmd = BUF_CMD_DONE;
1102 * VMIO buffer rundown. Make sure the VM page array is restored
1103 * after an I/O may have replaces some of the pages with bogus pages
1104 * in order to not destroy dirty pages in a fill-in read.
1106 * Note that due to the code above, if a buffer is marked B_DELWRI
1107 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1108 * B_INVAL may still be set, however.
1110 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1111 * but not the backing store. B_NOCACHE will destroy the backing
1114 * Note that dirty NFS buffers contain byte-granular write ranges
1115 * and should not be destroyed w/ B_INVAL even if the backing store
1118 if (bp->b_flags & B_VMIO) {
1120 * Rundown for VMIO buffers which are not dirty NFS buffers.
1132 * Get the base offset and length of the buffer. Note that
1133 * in the VMIO case if the buffer block size is not
1134 * page-aligned then b_data pointer may not be page-aligned.
1135 * But our b_xio.xio_pages array *IS* page aligned.
1137 * block sizes less then DEV_BSIZE (usually 512) are not
1138 * supported due to the page granularity bits (m->valid,
1139 * m->dirty, etc...).
1141 * See man buf(9) for more information
1144 resid = bp->b_bufsize;
1145 foff = bp->b_loffset;
1147 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1148 m = bp->b_xio.xio_pages[i];
1149 vm_page_flag_clear(m, PG_ZERO);
1151 * If we hit a bogus page, fixup *all* of them
1152 * now. Note that we left these pages wired
1153 * when we removed them so they had better exist,
1154 * and they cannot be ripped out from under us so
1155 * no critical section protection is necessary.
1157 if (m == bogus_page) {
1159 poff = OFF_TO_IDX(bp->b_loffset);
1161 for (j = i; j < bp->b_xio.xio_npages; j++) {
1164 mtmp = bp->b_xio.xio_pages[j];
1165 if (mtmp == bogus_page) {
1166 mtmp = vm_page_lookup(obj, poff + j);
1168 panic("brelse: page missing");
1170 bp->b_xio.xio_pages[j] = mtmp;
1174 if ((bp->b_flags & B_INVAL) == 0) {
1175 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1176 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1178 m = bp->b_xio.xio_pages[i];
1182 * Invalidate the backing store if B_NOCACHE is set
1183 * (e.g. used with vinvalbuf()). If this is NFS
1184 * we impose a requirement that the block size be
1185 * a multiple of PAGE_SIZE and create a temporary
1186 * hack to basically invalidate the whole page. The
1187 * problem is that NFS uses really odd buffer sizes
1188 * especially when tracking piecemeal writes and
1189 * it also vinvalbuf()'s a lot, which would result
1190 * in only partial page validation and invalidation
1191 * here. If the file page is mmap()'d, however,
1192 * all the valid bits get set so after we invalidate
1193 * here we would end up with weird m->valid values
1194 * like 0xfc. nfs_getpages() can't handle this so
1195 * we clear all the valid bits for the NFS case
1196 * instead of just some of them.
1198 * The real bug is the VM system having to set m->valid
1199 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1200 * itself is an artifact of the whole 512-byte
1201 * granular mess that exists to support odd block
1202 * sizes and UFS meta-data block sizes (e.g. 6144).
1203 * A complete rewrite is required.
1205 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1206 int poffset = foff & PAGE_MASK;
1209 presid = PAGE_SIZE - poffset;
1210 if (bp->b_vp->v_tag == VT_NFS &&
1211 bp->b_vp->v_type == VREG) {
1213 } else if (presid > resid) {
1216 KASSERT(presid >= 0, ("brelse: extra page"));
1217 vm_page_set_invalid(m, poffset, presid);
1219 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1220 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1222 if (bp->b_flags & (B_INVAL | B_RELBUF))
1223 vfs_vmio_release(bp);
1226 * Rundown for non-VMIO buffers.
1228 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1231 kprintf("brelse bp %p %08x/%08x: Warning, caught and fixed brelvp bug\n", bp, saved_flags, bp->b_flags);
1240 if (bp->b_qindex != BQUEUE_NONE)
1241 panic("brelse: free buffer onto another queue???");
1242 if (BUF_REFCNTNB(bp) > 1) {
1243 /* Temporary panic to verify exclusive locking */
1244 /* This panic goes away when we allow shared refs */
1245 panic("brelse: multiple refs");
1246 /* do not release to free list */
1253 * Figure out the correct queue to place the cleaned up buffer on.
1254 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1255 * disassociated from their vnode.
1257 if (bp->b_flags & B_LOCKED) {
1259 * Buffers that are locked are placed in the locked queue
1260 * immediately, regardless of their state.
1262 bp->b_qindex = BQUEUE_LOCKED;
1263 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1264 } else if (bp->b_bufsize == 0) {
1266 * Buffers with no memory. Due to conditionals near the top
1267 * of brelse() such buffers should probably already be
1268 * marked B_INVAL and disassociated from their vnode.
1270 bp->b_flags |= B_INVAL;
1271 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1272 KKASSERT((bp->b_flags & B_HASHED) == 0);
1273 if (bp->b_kvasize) {
1274 bp->b_qindex = BQUEUE_EMPTYKVA;
1276 bp->b_qindex = BQUEUE_EMPTY;
1278 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1279 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) {
1281 * Buffers with junk contents. Again these buffers had better
1282 * already be disassociated from their vnode.
1284 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1285 KKASSERT((bp->b_flags & B_HASHED) == 0);
1286 bp->b_flags |= B_INVAL;
1287 bp->b_qindex = BQUEUE_CLEAN;
1288 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1291 * Remaining buffers. These buffers are still associated with
1294 switch(bp->b_flags & (B_DELWRI|B_HEAVY|B_AGE)) {
1295 case B_DELWRI | B_AGE:
1296 bp->b_qindex = BQUEUE_DIRTY;
1297 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1300 bp->b_qindex = BQUEUE_DIRTY;
1301 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1303 case B_DELWRI | B_HEAVY | B_AGE:
1304 bp->b_qindex = BQUEUE_DIRTY_HW;
1305 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_DIRTY_HW], bp,
1308 case B_DELWRI | B_HEAVY:
1309 bp->b_qindex = BQUEUE_DIRTY_HW;
1310 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1313 case B_HEAVY | B_AGE:
1315 bp->b_qindex = BQUEUE_CLEAN;
1316 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1319 bp->b_qindex = BQUEUE_CLEAN;
1320 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1326 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1327 * on the correct queue.
1329 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1333 * Fixup numfreebuffers count. The bp is on an appropriate queue
1334 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1335 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1336 * if B_INVAL is set ).
1338 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1342 * Something we can maybe free or reuse
1344 if (bp->b_bufsize || bp->b_kvasize)
1348 * Clean up temporary flags and unlock the buffer.
1350 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF |
1351 B_DIRECT | B_NOWDRAIN);
1359 * Release a buffer back to the appropriate queue but do not try to free
1360 * it. The buffer is expected to be used again soon.
1362 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1363 * biodone() to requeue an async I/O on completion. It is also used when
1364 * known good buffers need to be requeued but we think we may need the data
1367 * XXX we should be able to leave the B_RELBUF hint set on completion.
1370 bqrelse(struct buf *bp)
1374 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1376 if (bp->b_qindex != BQUEUE_NONE)
1377 panic("bqrelse: free buffer onto another queue???");
1378 if (BUF_REFCNTNB(bp) > 1) {
1379 /* do not release to free list */
1380 panic("bqrelse: multiple refs");
1385 if (bp->b_flags & B_LOCKED) {
1387 * Locked buffers are released to the locked queue. However,
1388 * if the buffer is dirty it will first go into the dirty
1389 * queue and later on after the I/O completes successfully it
1390 * will be released to the locked queue.
1392 bp->b_flags &= ~B_ERROR;
1393 bp->b_qindex = BQUEUE_LOCKED;
1394 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1395 } else if (bp->b_flags & B_DELWRI) {
1396 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1397 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1398 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1399 } else if (vm_page_count_severe()) {
1401 * We are too low on memory, we have to try to free the
1402 * buffer (most importantly: the wired pages making up its
1403 * backing store) *now*.
1409 bp->b_qindex = BQUEUE_CLEAN;
1410 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1413 if ((bp->b_flags & B_LOCKED) == 0 &&
1414 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1419 * Something we can maybe free or reuse.
1421 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1425 * Final cleanup and unlock. Clear bits that are only used while a
1426 * buffer is actively locked.
1428 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1436 * Return backing pages held by the buffer 'bp' back to the VM system
1437 * if possible. The pages are freed if they are no longer valid or
1438 * attempt to free if it was used for direct I/O otherwise they are
1439 * sent to the page cache.
1441 * Pages that were marked busy are left alone and skipped.
1443 * The KVA mapping (b_data) for the underlying pages is removed by
1447 vfs_vmio_release(struct buf *bp)
1453 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1454 m = bp->b_xio.xio_pages[i];
1455 bp->b_xio.xio_pages[i] = NULL;
1457 * In order to keep page LRU ordering consistent, put
1458 * everything on the inactive queue.
1460 vm_page_unwire(m, 0);
1462 * We don't mess with busy pages, it is
1463 * the responsibility of the process that
1464 * busied the pages to deal with them.
1466 if ((m->flags & PG_BUSY) || (m->busy != 0))
1469 if (m->wire_count == 0) {
1470 vm_page_flag_clear(m, PG_ZERO);
1472 * Might as well free the page if we can and it has
1473 * no valid data. We also free the page if the
1474 * buffer was used for direct I/O.
1476 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1477 m->hold_count == 0) {
1479 vm_page_protect(m, VM_PROT_NONE);
1481 } else if (bp->b_flags & B_DIRECT) {
1482 vm_page_try_to_free(m);
1483 } else if (vm_page_count_severe()) {
1484 vm_page_try_to_cache(m);
1489 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1490 if (bp->b_bufsize) {
1494 bp->b_xio.xio_npages = 0;
1495 bp->b_flags &= ~B_VMIO;
1503 * Implement clustered async writes for clearing out B_DELWRI buffers.
1504 * This is much better then the old way of writing only one buffer at
1505 * a time. Note that we may not be presented with the buffers in the
1506 * correct order, so we search for the cluster in both directions.
1508 * The buffer is locked on call.
1511 vfs_bio_awrite(struct buf *bp)
1515 off_t loffset = bp->b_loffset;
1516 struct vnode *vp = bp->b_vp;
1524 * right now we support clustered writing only to regular files. If
1525 * we find a clusterable block we could be in the middle of a cluster
1526 * rather then at the beginning.
1528 * NOTE: b_bio1 contains the logical loffset and is aliased
1529 * to b_loffset. b_bio2 contains the translated block number.
1531 if ((vp->v_type == VREG) &&
1532 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1533 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1535 size = vp->v_mount->mnt_stat.f_iosize;
1537 for (i = size; i < MAXPHYS; i += size) {
1538 if ((bpa = findblk(vp, loffset + i)) &&
1539 BUF_REFCNT(bpa) == 0 &&
1540 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1541 (B_DELWRI | B_CLUSTEROK)) &&
1542 (bpa->b_bufsize == size)) {
1543 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1544 (bpa->b_bio2.bio_offset !=
1545 bp->b_bio2.bio_offset + i))
1551 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1552 if ((bpa = findblk(vp, loffset - j)) &&
1553 BUF_REFCNT(bpa) == 0 &&
1554 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1555 (B_DELWRI | B_CLUSTEROK)) &&
1556 (bpa->b_bufsize == size)) {
1557 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1558 (bpa->b_bio2.bio_offset !=
1559 bp->b_bio2.bio_offset - j))
1568 * this is a possible cluster write
1570 if (nbytes != size) {
1572 nwritten = cluster_wbuild(vp, size,
1573 loffset - j, nbytes);
1580 bp->b_flags |= B_ASYNC;
1584 * default (old) behavior, writing out only one block
1586 * XXX returns b_bufsize instead of b_bcount for nwritten?
1588 nwritten = bp->b_bufsize;
1597 * Find and initialize a new buffer header, freeing up existing buffers
1598 * in the bufqueues as necessary. The new buffer is returned locked.
1600 * Important: B_INVAL is not set. If the caller wishes to throw the
1601 * buffer away, the caller must set B_INVAL prior to calling brelse().
1604 * We have insufficient buffer headers
1605 * We have insufficient buffer space
1606 * buffer_map is too fragmented ( space reservation fails )
1607 * If we have to flush dirty buffers ( but we try to avoid this )
1609 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1610 * Instead we ask the buf daemon to do it for us. We attempt to
1611 * avoid piecemeal wakeups of the pageout daemon.
1615 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1621 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1622 static int flushingbufs;
1625 * We can't afford to block since we might be holding a vnode lock,
1626 * which may prevent system daemons from running. We deal with
1627 * low-memory situations by proactively returning memory and running
1628 * async I/O rather then sync I/O.
1632 --getnewbufrestarts;
1634 ++getnewbufrestarts;
1637 * Setup for scan. If we do not have enough free buffers,
1638 * we setup a degenerate case that immediately fails. Note
1639 * that if we are specially marked process, we are allowed to
1640 * dip into our reserves.
1642 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1644 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1645 * However, there are a number of cases (defragging, reusing, ...)
1646 * where we cannot backup.
1648 nqindex = BQUEUE_EMPTYKVA;
1649 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1653 * If no EMPTYKVA buffers and we are either
1654 * defragging or reusing, locate a CLEAN buffer
1655 * to free or reuse. If bufspace useage is low
1656 * skip this step so we can allocate a new buffer.
1658 if (defrag || bufspace >= lobufspace) {
1659 nqindex = BQUEUE_CLEAN;
1660 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1664 * If we could not find or were not allowed to reuse a
1665 * CLEAN buffer, check to see if it is ok to use an EMPTY
1666 * buffer. We can only use an EMPTY buffer if allocating
1667 * its KVA would not otherwise run us out of buffer space.
1669 if (nbp == NULL && defrag == 0 &&
1670 bufspace + maxsize < hibufspace) {
1671 nqindex = BQUEUE_EMPTY;
1672 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1677 * Run scan, possibly freeing data and/or kva mappings on the fly
1681 while ((bp = nbp) != NULL) {
1682 int qindex = nqindex;
1685 * Calculate next bp ( we can only use it if we do not block
1686 * or do other fancy things ).
1688 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1691 nqindex = BQUEUE_EMPTYKVA;
1692 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1695 case BQUEUE_EMPTYKVA:
1696 nqindex = BQUEUE_CLEAN;
1697 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1711 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1714 * Note: we no longer distinguish between VMIO and non-VMIO
1718 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1721 * If we are defragging then we need a buffer with
1722 * b_kvasize != 0. XXX this situation should no longer
1723 * occur, if defrag is non-zero the buffer's b_kvasize
1724 * should also be non-zero at this point. XXX
1726 if (defrag && bp->b_kvasize == 0) {
1727 kprintf("Warning: defrag empty buffer %p\n", bp);
1732 * Start freeing the bp. This is somewhat involved. nbp
1733 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1734 * on the clean list must be disassociated from their
1735 * current vnode. Buffers on the empty[kva] lists have
1736 * already been disassociated.
1739 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1740 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1741 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1744 if (bp->b_qindex != qindex) {
1745 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1752 * Dependancies must be handled before we disassociate the
1755 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1756 * be immediately disassociated. HAMMER then becomes
1757 * responsible for releasing the buffer.
1759 if (LIST_FIRST(&bp->b_dep) != NULL) {
1761 if (bp->b_flags & B_LOCKED) {
1765 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1768 if (qindex == BQUEUE_CLEAN) {
1769 if (bp->b_flags & B_VMIO) {
1770 bp->b_flags &= ~B_ASYNC;
1771 vfs_vmio_release(bp);
1778 * NOTE: nbp is now entirely invalid. We can only restart
1779 * the scan from this point on.
1781 * Get the rest of the buffer freed up. b_kva* is still
1782 * valid after this operation.
1785 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1786 KKASSERT((bp->b_flags & B_HASHED) == 0);
1789 * critical section protection is not required when
1790 * scrapping a buffer's contents because it is already
1796 bp->b_flags = B_BNOCLIP;
1797 bp->b_cmd = BUF_CMD_DONE;
1802 bp->b_xio.xio_npages = 0;
1803 bp->b_dirtyoff = bp->b_dirtyend = 0;
1806 if (blkflags & GETBLK_BHEAVY)
1807 bp->b_flags |= B_HEAVY;
1810 * If we are defragging then free the buffer.
1813 bp->b_flags |= B_INVAL;
1821 * If we are overcomitted then recover the buffer and its
1822 * KVM space. This occurs in rare situations when multiple
1823 * processes are blocked in getnewbuf() or allocbuf().
1825 if (bufspace >= hibufspace)
1827 if (flushingbufs && bp->b_kvasize != 0) {
1828 bp->b_flags |= B_INVAL;
1833 if (bufspace < lobufspace)
1839 * If we exhausted our list, sleep as appropriate. We may have to
1840 * wakeup various daemons and write out some dirty buffers.
1842 * Generally we are sleeping due to insufficient buffer space.
1850 flags = VFS_BIO_NEED_BUFSPACE;
1852 } else if (bufspace >= hibufspace) {
1854 flags = VFS_BIO_NEED_BUFSPACE;
1857 flags = VFS_BIO_NEED_ANY;
1860 needsbuffer |= flags;
1861 bd_speedup(); /* heeeelp */
1862 while (needsbuffer & flags) {
1863 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
1868 * We finally have a valid bp. We aren't quite out of the
1869 * woods, we still have to reserve kva space. In order
1870 * to keep fragmentation sane we only allocate kva in
1873 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1875 if (maxsize != bp->b_kvasize) {
1876 vm_offset_t addr = 0;
1881 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1882 vm_map_lock(&buffer_map);
1884 if (vm_map_findspace(&buffer_map,
1885 vm_map_min(&buffer_map), maxsize,
1888 * Uh oh. Buffer map is too fragmented. We
1889 * must defragment the map.
1891 vm_map_unlock(&buffer_map);
1892 vm_map_entry_release(count);
1895 bp->b_flags |= B_INVAL;
1900 vm_map_insert(&buffer_map, &count,
1902 addr, addr + maxsize,
1904 VM_PROT_ALL, VM_PROT_ALL,
1907 bp->b_kvabase = (caddr_t) addr;
1908 bp->b_kvasize = maxsize;
1909 bufspace += bp->b_kvasize;
1912 vm_map_unlock(&buffer_map);
1913 vm_map_entry_release(count);
1915 bp->b_data = bp->b_kvabase;
1923 * Buffer flushing daemon. Buffers are normally flushed by the
1924 * update daemon but if it cannot keep up this process starts to
1925 * take the load in an attempt to prevent getnewbuf() from blocking.
1927 * Once a flush is initiated it does not stop until the number
1928 * of buffers falls below lodirtybuffers, but we will wake up anyone
1929 * waiting at the mid-point.
1932 static struct thread *bufdaemon_td;
1933 static struct thread *bufdaemonhw_td;
1935 static struct kproc_desc buf_kp = {
1940 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
1941 kproc_start, &buf_kp)
1943 static struct kproc_desc bufhw_kp = {
1948 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
1949 kproc_start, &bufhw_kp)
1955 * This process needs to be suspended prior to shutdown sync.
1957 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
1958 bufdaemon_td, SHUTDOWN_PRI_LAST);
1961 * This process is allowed to take the buffer cache to the limit
1966 kproc_suspend_loop();
1969 * Do the flush. Limit the amount of in-transit I/O we
1970 * allow to build up, otherwise we would completely saturate
1971 * the I/O system. Wakeup any waiting processes before we
1972 * normally would so they can run in parallel with our drain.
1974 while (numdirtybuffers > lodirtybuffers) {
1975 if (flushbufqueues(BQUEUE_DIRTY) == 0)
1977 waitrunningbufspace();
1983 * Only clear bd_request if we have reached our low water
1984 * mark. The buf_daemon normally waits 5 seconds and
1985 * then incrementally flushes any dirty buffers that have
1986 * built up, within reason.
1988 * If we were unable to hit our low water mark and couldn't
1989 * find any flushable buffers, we sleep half a second.
1990 * Otherwise we loop immediately.
1992 if (numdirtybuffers <= lodirtybuffers) {
1994 * We reached our low water mark, reset the
1995 * request and sleep until we are needed again.
1996 * The sleep is just so the suspend code works.
1998 spin_lock_wr(&needsbuffer_spin);
2000 msleep(&bd_request, &needsbuffer_spin, 0,
2002 spin_unlock_wr(&needsbuffer_spin);
2005 * We couldn't find any flushable dirty buffers but
2006 * still have too many dirty buffers, we
2007 * have to sleep and try again. (rare)
2009 tsleep(&bd_request, 0, "qsleep", hz / 2);
2018 * This process needs to be suspended prior to shutdown sync.
2020 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2021 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2024 * This process is allowed to take the buffer cache to the limit
2029 kproc_suspend_loop();
2032 * Do the flush. Limit the amount of in-transit I/O we
2033 * allow to build up, otherwise we would completely saturate
2034 * the I/O system. Wakeup any waiting processes before we
2035 * normally would so they can run in parallel with our drain.
2037 while (numdirtybuffershw > lodirtybuffers) {
2038 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2040 waitrunningbufspace();
2045 * Only clear bd_request if we have reached our low water
2046 * mark. The buf_daemon normally waits 5 seconds and
2047 * then incrementally flushes any dirty buffers that have
2048 * built up, within reason.
2050 * If we were unable to hit our low water mark and couldn't
2051 * find any flushable buffers, we sleep half a second.
2052 * Otherwise we loop immediately.
2054 if (numdirtybuffershw <= lodirtybuffers) {
2056 * We reached our low water mark, reset the
2057 * request and sleep until we are needed again.
2058 * The sleep is just so the suspend code works.
2060 spin_lock_wr(&needsbuffer_spin);
2062 msleep(&bd_request_hw, &needsbuffer_spin, 0,
2064 spin_unlock_wr(&needsbuffer_spin);
2067 * We couldn't find any flushable dirty buffers but
2068 * still have too many dirty buffers, we
2069 * have to sleep and try again. (rare)
2071 tsleep(&bd_request_hw, 0, "qsleep", hz / 2);
2079 * Try to flush a buffer in the dirty queue. We must be careful to
2080 * free up B_INVAL buffers instead of write them, which NFS is
2081 * particularly sensitive to.
2085 flushbufqueues(bufq_type_t q)
2090 bp = TAILQ_FIRST(&bufqueues[q]);
2093 KASSERT((bp->b_flags & B_DELWRI),
2094 ("unexpected clean buffer %p", bp));
2095 if (bp->b_flags & B_DELWRI) {
2096 if (bp->b_flags & B_INVAL) {
2097 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2098 panic("flushbufqueues: locked buf");
2104 if (LIST_FIRST(&bp->b_dep) != NULL &&
2105 (bp->b_flags & B_DEFERRED) == 0 &&
2106 buf_countdeps(bp, 0)) {
2107 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2108 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2110 bp->b_flags |= B_DEFERRED;
2111 bp = TAILQ_FIRST(&bufqueues[q]);
2116 * Only write it out if we can successfully lock
2117 * it. If the buffer has a dependancy,
2118 * buf_checkwrite must also return 0.
2120 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
2121 if (LIST_FIRST(&bp->b_dep) != NULL &&
2122 buf_checkwrite(bp)) {
2132 bp = TAILQ_NEXT(bp, b_freelist);
2140 * Returns true if no I/O is needed to access the associated VM object.
2141 * This is like findblk except it also hunts around in the VM system for
2144 * Note that we ignore vm_page_free() races from interrupts against our
2145 * lookup, since if the caller is not protected our return value will not
2146 * be any more valid then otherwise once we exit the critical section.
2149 inmem(struct vnode *vp, off_t loffset)
2152 vm_offset_t toff, tinc, size;
2155 if (findblk(vp, loffset))
2157 if (vp->v_mount == NULL)
2159 if ((obj = vp->v_object) == NULL)
2163 if (size > vp->v_mount->mnt_stat.f_iosize)
2164 size = vp->v_mount->mnt_stat.f_iosize;
2166 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2167 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2171 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2172 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2173 if (vm_page_is_valid(m,
2174 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2183 * Sets the dirty range for a buffer based on the status of the dirty
2184 * bits in the pages comprising the buffer.
2186 * The range is limited to the size of the buffer.
2188 * This routine is primarily used by NFS, but is generalized for the
2192 vfs_setdirty(struct buf *bp)
2198 * Degenerate case - empty buffer
2201 if (bp->b_bufsize == 0)
2205 * We qualify the scan for modified pages on whether the
2206 * object has been flushed yet. The OBJ_WRITEABLE flag
2207 * is not cleared simply by protecting pages off.
2210 if ((bp->b_flags & B_VMIO) == 0)
2213 object = bp->b_xio.xio_pages[0]->object;
2215 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2216 kprintf("Warning: object %p writeable but not mightbedirty\n", object);
2217 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2218 kprintf("Warning: object %p mightbedirty but not writeable\n", object);
2220 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2221 vm_offset_t boffset;
2222 vm_offset_t eoffset;
2225 * test the pages to see if they have been modified directly
2226 * by users through the VM system.
2228 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2229 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2230 vm_page_test_dirty(bp->b_xio.xio_pages[i]);
2234 * Calculate the encompassing dirty range, boffset and eoffset,
2235 * (eoffset - boffset) bytes.
2238 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2239 if (bp->b_xio.xio_pages[i]->dirty)
2242 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2244 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2245 if (bp->b_xio.xio_pages[i]->dirty) {
2249 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2252 * Fit it to the buffer.
2255 if (eoffset > bp->b_bcount)
2256 eoffset = bp->b_bcount;
2259 * If we have a good dirty range, merge with the existing
2263 if (boffset < eoffset) {
2264 if (bp->b_dirtyoff > boffset)
2265 bp->b_dirtyoff = boffset;
2266 if (bp->b_dirtyend < eoffset)
2267 bp->b_dirtyend = eoffset;
2275 * Locate and return the specified buffer, or NULL if the buffer does
2276 * not exist. Do not attempt to lock the buffer or manipulate it in
2277 * any way. The caller must validate that the correct buffer has been
2278 * obtain after locking it.
2281 findblk(struct vnode *vp, off_t loffset)
2286 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2294 * Get a block given a specified block and offset into a file/device.
2295 * B_INVAL may or may not be set on return. The caller should clear
2296 * B_INVAL prior to initiating a READ.
2298 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2299 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2300 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2301 * without doing any of those things the system will likely believe
2302 * the buffer to be valid (especially if it is not B_VMIO), and the
2303 * next getblk() will return the buffer with B_CACHE set.
2305 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2306 * an existing buffer.
2308 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2309 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2310 * and then cleared based on the backing VM. If the previous buffer is
2311 * non-0-sized but invalid, B_CACHE will be cleared.
2313 * If getblk() must create a new buffer, the new buffer is returned with
2314 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2315 * case it is returned with B_INVAL clear and B_CACHE set based on the
2318 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2319 * B_CACHE bit is clear.
2321 * What this means, basically, is that the caller should use B_CACHE to
2322 * determine whether the buffer is fully valid or not and should clear
2323 * B_INVAL prior to issuing a read. If the caller intends to validate
2324 * the buffer by loading its data area with something, the caller needs
2325 * to clear B_INVAL. If the caller does this without issuing an I/O,
2326 * the caller should set B_CACHE ( as an optimization ), else the caller
2327 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2328 * a write attempt or if it was a successfull read. If the caller
2329 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2330 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2334 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2335 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2338 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2341 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2343 if (size > MAXBSIZE)
2344 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2345 if (vp->v_object == NULL)
2346 panic("getblk: vnode %p has no object!", vp);
2350 if ((bp = findblk(vp, loffset))) {
2352 * The buffer was found in the cache, but we need to lock it.
2353 * Even with LK_NOWAIT the lockmgr may break our critical
2354 * section, so double-check the validity of the buffer
2355 * once the lock has been obtained.
2357 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2358 int lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2359 if (blkflags & GETBLK_PCATCH)
2360 lkflags |= LK_PCATCH;
2361 if (BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo) ==
2370 * Once the buffer has been locked, make sure we didn't race
2371 * a buffer recyclement. Buffers that are no longer hashed
2372 * will have b_vp == NULL, so this takes care of that check
2375 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2376 kprintf("Warning buffer %p (vp %p loffset %lld) was recycled\n", bp, vp, loffset);
2382 * All vnode-based buffers must be backed by a VM object.
2384 KKASSERT(bp->b_flags & B_VMIO);
2385 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2388 * Make sure that B_INVAL buffers do not have a cached
2389 * block number translation.
2391 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2392 kprintf("Warning invalid buffer %p (vp %p loffset %lld) did not have cleared bio_offset cache\n", bp, vp, loffset);
2393 clearbiocache(&bp->b_bio2);
2397 * The buffer is locked. B_CACHE is cleared if the buffer is
2400 if (bp->b_flags & B_INVAL)
2401 bp->b_flags &= ~B_CACHE;
2405 * Any size inconsistancy with a dirty buffer or a buffer
2406 * with a softupdates dependancy must be resolved. Resizing
2407 * the buffer in such circumstances can lead to problems.
2409 if (size != bp->b_bcount) {
2410 if (bp->b_flags & B_DELWRI) {
2411 bp->b_flags |= B_NOCACHE;
2413 } else if (LIST_FIRST(&bp->b_dep)) {
2414 bp->b_flags |= B_NOCACHE;
2417 bp->b_flags |= B_RELBUF;
2422 KKASSERT(size <= bp->b_kvasize);
2423 KASSERT(bp->b_loffset != NOOFFSET,
2424 ("getblk: no buffer offset"));
2427 * A buffer with B_DELWRI set and B_CACHE clear must
2428 * be committed before we can return the buffer in
2429 * order to prevent the caller from issuing a read
2430 * ( due to B_CACHE not being set ) and overwriting
2433 * Most callers, including NFS and FFS, need this to
2434 * operate properly either because they assume they
2435 * can issue a read if B_CACHE is not set, or because
2436 * ( for example ) an uncached B_DELWRI might loop due
2437 * to softupdates re-dirtying the buffer. In the latter
2438 * case, B_CACHE is set after the first write completes,
2439 * preventing further loops.
2441 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2442 * above while extending the buffer, we cannot allow the
2443 * buffer to remain with B_CACHE set after the write
2444 * completes or it will represent a corrupt state. To
2445 * deal with this we set B_NOCACHE to scrap the buffer
2448 * We might be able to do something fancy, like setting
2449 * B_CACHE in bwrite() except if B_DELWRI is already set,
2450 * so the below call doesn't set B_CACHE, but that gets real
2451 * confusing. This is much easier.
2454 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2455 bp->b_flags |= B_NOCACHE;
2462 * Buffer is not in-core, create new buffer. The buffer
2463 * returned by getnewbuf() is locked. Note that the returned
2464 * buffer is also considered valid (not marked B_INVAL).
2466 * Calculating the offset for the I/O requires figuring out
2467 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2468 * the mount's f_iosize otherwise. If the vnode does not
2469 * have an associated mount we assume that the passed size is
2472 * Note that vn_isdisk() cannot be used here since it may
2473 * return a failure for numerous reasons. Note that the
2474 * buffer size may be larger then the block size (the caller
2475 * will use block numbers with the proper multiple). Beware
2476 * of using any v_* fields which are part of unions. In
2477 * particular, in DragonFly the mount point overloading
2478 * mechanism uses the namecache only and the underlying
2479 * directory vnode is not a special case.
2484 * Don't let heavy weight buffers deadlock us.
2486 if ((blkflags & GETBLK_BHEAVY) &&
2487 numdirtybuffershw > hidirtybuffers) {
2488 while (numdirtybuffershw > hidirtybuffers) {
2489 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
2490 tsleep(&needsbuffer, slpflags, "newbuf",
2496 if (vp->v_type == VBLK || vp->v_type == VCHR)
2498 else if (vp->v_mount)
2499 bsize = vp->v_mount->mnt_stat.f_iosize;
2503 maxsize = size + (loffset & PAGE_MASK);
2504 maxsize = imax(maxsize, bsize);
2506 if ((bp = getnewbuf(blkflags, slptimeo, size, maxsize)) == NULL) {
2507 if (slpflags || slptimeo) {
2515 * This code is used to make sure that a buffer is not
2516 * created while the getnewbuf routine is blocked.
2517 * This can be a problem whether the vnode is locked or not.
2518 * If the buffer is created out from under us, we have to
2519 * throw away the one we just created. There is no window
2520 * race because we are safely running in a critical section
2521 * from the point of the duplicate buffer creation through
2522 * to here, and we've locked the buffer.
2524 if (findblk(vp, loffset)) {
2525 bp->b_flags |= B_INVAL;
2531 * Insert the buffer into the hash, so that it can
2532 * be found by findblk().
2534 * Make sure the translation layer has been cleared.
2536 bp->b_loffset = loffset;
2537 bp->b_bio2.bio_offset = NOOFFSET;
2538 /* bp->b_bio2.bio_next = NULL; */
2543 * All vnode-based buffers must be backed by a VM object.
2545 KKASSERT(vp->v_object != NULL);
2546 bp->b_flags |= B_VMIO;
2547 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2559 * Reacquire a buffer that was previously released to the locked queue,
2560 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2561 * set B_LOCKED (which handles the acquisition race).
2563 * To this end, either B_LOCKED must be set or the dependancy list must be
2567 regetblk(struct buf *bp)
2569 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
2570 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2579 * Get an empty, disassociated buffer of given size. The buffer is
2580 * initially set to B_INVAL.
2582 * critical section protection is not required for the allocbuf()
2583 * call because races are impossible here.
2591 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2594 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2598 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2606 * This code constitutes the buffer memory from either anonymous system
2607 * memory (in the case of non-VMIO operations) or from an associated
2608 * VM object (in the case of VMIO operations). This code is able to
2609 * resize a buffer up or down.
2611 * Note that this code is tricky, and has many complications to resolve
2612 * deadlock or inconsistant data situations. Tread lightly!!!
2613 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2614 * the caller. Calling this code willy nilly can result in the loss of data.
2616 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2617 * B_CACHE for the non-VMIO case.
2619 * This routine does not need to be called from a critical section but you
2620 * must own the buffer.
2623 allocbuf(struct buf *bp, int size)
2625 int newbsize, mbsize;
2628 if (BUF_REFCNT(bp) == 0)
2629 panic("allocbuf: buffer not busy");
2631 if (bp->b_kvasize < size)
2632 panic("allocbuf: buffer too small");
2634 if ((bp->b_flags & B_VMIO) == 0) {
2638 * Just get anonymous memory from the kernel. Don't
2639 * mess with B_CACHE.
2641 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2642 if (bp->b_flags & B_MALLOC)
2645 newbsize = round_page(size);
2647 if (newbsize < bp->b_bufsize) {
2649 * Malloced buffers are not shrunk
2651 if (bp->b_flags & B_MALLOC) {
2653 bp->b_bcount = size;
2655 kfree(bp->b_data, M_BIOBUF);
2656 if (bp->b_bufsize) {
2657 bufmallocspace -= bp->b_bufsize;
2661 bp->b_data = bp->b_kvabase;
2663 bp->b_flags &= ~B_MALLOC;
2669 (vm_offset_t) bp->b_data + newbsize,
2670 (vm_offset_t) bp->b_data + bp->b_bufsize);
2671 } else if (newbsize > bp->b_bufsize) {
2673 * We only use malloced memory on the first allocation.
2674 * and revert to page-allocated memory when the buffer
2677 if ((bufmallocspace < maxbufmallocspace) &&
2678 (bp->b_bufsize == 0) &&
2679 (mbsize <= PAGE_SIZE/2)) {
2681 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
2682 bp->b_bufsize = mbsize;
2683 bp->b_bcount = size;
2684 bp->b_flags |= B_MALLOC;
2685 bufmallocspace += mbsize;
2691 * If the buffer is growing on its other-than-first
2692 * allocation, then we revert to the page-allocation
2695 if (bp->b_flags & B_MALLOC) {
2696 origbuf = bp->b_data;
2697 origbufsize = bp->b_bufsize;
2698 bp->b_data = bp->b_kvabase;
2699 if (bp->b_bufsize) {
2700 bufmallocspace -= bp->b_bufsize;
2704 bp->b_flags &= ~B_MALLOC;
2705 newbsize = round_page(newbsize);
2709 (vm_offset_t) bp->b_data + bp->b_bufsize,
2710 (vm_offset_t) bp->b_data + newbsize);
2712 bcopy(origbuf, bp->b_data, origbufsize);
2713 kfree(origbuf, M_BIOBUF);
2720 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2721 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
2722 newbsize + PAGE_MASK) >> PAGE_SHIFT;
2723 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
2725 if (bp->b_flags & B_MALLOC)
2726 panic("allocbuf: VMIO buffer can't be malloced");
2728 * Set B_CACHE initially if buffer is 0 length or will become
2731 if (size == 0 || bp->b_bufsize == 0)
2732 bp->b_flags |= B_CACHE;
2734 if (newbsize < bp->b_bufsize) {
2736 * DEV_BSIZE aligned new buffer size is less then the
2737 * DEV_BSIZE aligned existing buffer size. Figure out
2738 * if we have to remove any pages.
2740 if (desiredpages < bp->b_xio.xio_npages) {
2741 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
2743 * the page is not freed here -- it
2744 * is the responsibility of
2745 * vnode_pager_setsize
2747 m = bp->b_xio.xio_pages[i];
2748 KASSERT(m != bogus_page,
2749 ("allocbuf: bogus page found"));
2750 while (vm_page_sleep_busy(m, TRUE, "biodep"))
2753 bp->b_xio.xio_pages[i] = NULL;
2754 vm_page_unwire(m, 0);
2756 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2757 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
2758 bp->b_xio.xio_npages = desiredpages;
2760 } else if (size > bp->b_bcount) {
2762 * We are growing the buffer, possibly in a
2763 * byte-granular fashion.
2771 * Step 1, bring in the VM pages from the object,
2772 * allocating them if necessary. We must clear
2773 * B_CACHE if these pages are not valid for the
2774 * range covered by the buffer.
2776 * critical section protection is required to protect
2777 * against interrupts unbusying and freeing pages
2778 * between our vm_page_lookup() and our
2779 * busycheck/wiring call.
2785 while (bp->b_xio.xio_npages < desiredpages) {
2789 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
2790 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2792 * note: must allocate system pages
2793 * since blocking here could intefere
2794 * with paging I/O, no matter which
2797 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
2800 vm_pageout_deficit += desiredpages -
2801 bp->b_xio.xio_npages;
2805 bp->b_flags &= ~B_CACHE;
2806 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2807 ++bp->b_xio.xio_npages;
2813 * We found a page. If we have to sleep on it,
2814 * retry because it might have gotten freed out
2817 * We can only test PG_BUSY here. Blocking on
2818 * m->busy might lead to a deadlock:
2820 * vm_fault->getpages->cluster_read->allocbuf
2824 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
2828 * We have a good page. Should we wakeup the
2831 if ((curthread != pagethread) &&
2832 ((m->queue - m->pc) == PQ_CACHE) &&
2833 ((vmstats.v_free_count + vmstats.v_cache_count) <
2834 (vmstats.v_free_min + vmstats.v_cache_min))) {
2835 pagedaemon_wakeup();
2837 vm_page_flag_clear(m, PG_ZERO);
2839 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2840 ++bp->b_xio.xio_npages;
2845 * Step 2. We've loaded the pages into the buffer,
2846 * we have to figure out if we can still have B_CACHE
2847 * set. Note that B_CACHE is set according to the
2848 * byte-granular range ( bcount and size ), not the
2849 * aligned range ( newbsize ).
2851 * The VM test is against m->valid, which is DEV_BSIZE
2852 * aligned. Needless to say, the validity of the data
2853 * needs to also be DEV_BSIZE aligned. Note that this
2854 * fails with NFS if the server or some other client
2855 * extends the file's EOF. If our buffer is resized,
2856 * B_CACHE may remain set! XXX
2859 toff = bp->b_bcount;
2860 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
2862 while ((bp->b_flags & B_CACHE) && toff < size) {
2865 if (tinc > (size - toff))
2868 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
2876 bp->b_xio.xio_pages[pi]
2883 * Step 3, fixup the KVM pmap. Remember that
2884 * bp->b_data is relative to bp->b_loffset, but
2885 * bp->b_loffset may be offset into the first page.
2888 bp->b_data = (caddr_t)
2889 trunc_page((vm_offset_t)bp->b_data);
2891 (vm_offset_t)bp->b_data,
2892 bp->b_xio.xio_pages,
2893 bp->b_xio.xio_npages
2895 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2896 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
2899 if (newbsize < bp->b_bufsize)
2901 bp->b_bufsize = newbsize; /* actual buffer allocation */
2902 bp->b_bcount = size; /* requested buffer size */
2909 * Wait for buffer I/O completion, returning error status. The buffer
2910 * is left locked on return. B_EINTR is converted into an EINTR error
2913 * NOTE! The original b_cmd is lost on return, since b_cmd will be
2914 * set to BUF_CMD_DONE.
2917 biowait(struct buf *bp)
2920 while (bp->b_cmd != BUF_CMD_DONE) {
2921 if (bp->b_cmd == BUF_CMD_READ)
2922 tsleep(bp, 0, "biord", 0);
2924 tsleep(bp, 0, "biowr", 0);
2927 if (bp->b_flags & B_EINTR) {
2928 bp->b_flags &= ~B_EINTR;
2931 if (bp->b_flags & B_ERROR) {
2932 return (bp->b_error ? bp->b_error : EIO);
2939 * This associates a tracking count with an I/O. vn_strategy() and
2940 * dev_dstrategy() do this automatically but there are a few cases
2941 * where a vnode or device layer is bypassed when a block translation
2942 * is cached. In such cases bio_start_transaction() may be called on
2943 * the bypassed layers so the system gets an I/O in progress indication
2944 * for those higher layers.
2947 bio_start_transaction(struct bio *bio, struct bio_track *track)
2949 bio->bio_track = track;
2950 atomic_add_int(&track->bk_active, 1);
2954 * Initiate I/O on a vnode.
2957 vn_strategy(struct vnode *vp, struct bio *bio)
2959 struct bio_track *track;
2961 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
2962 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
2963 track = &vp->v_track_read;
2965 track = &vp->v_track_write;
2966 bio->bio_track = track;
2967 atomic_add_int(&track->bk_active, 1);
2968 vop_strategy(*vp->v_ops, vp, bio);
2975 * Finish I/O on a buffer, optionally calling a completion function.
2976 * This is usually called from an interrupt so process blocking is
2979 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
2980 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
2981 * assuming B_INVAL is clear.
2983 * For the VMIO case, we set B_CACHE if the op was a read and no
2984 * read error occured, or if the op was a write. B_CACHE is never
2985 * set if the buffer is invalid or otherwise uncacheable.
2987 * biodone does not mess with B_INVAL, allowing the I/O routine or the
2988 * initiator to leave B_INVAL set to brelse the buffer out of existance
2989 * in the biodone routine.
2992 biodone(struct bio *bio)
2994 struct buf *bp = bio->bio_buf;
2999 KASSERT(BUF_REFCNTNB(bp) > 0,
3000 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3001 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3002 ("biodone: bp %p already done!", bp));
3004 runningbufwakeup(bp);
3007 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3010 biodone_t *done_func;
3011 struct bio_track *track;
3014 * BIO tracking. Most but not all BIOs are tracked.
3016 if ((track = bio->bio_track) != NULL) {
3017 atomic_subtract_int(&track->bk_active, 1);
3018 if (track->bk_active < 0) {
3019 panic("biodone: bad active count bio %p\n",
3022 if (track->bk_waitflag) {
3023 track->bk_waitflag = 0;
3026 bio->bio_track = NULL;
3030 * A bio_done function terminates the loop. The function
3031 * will be responsible for any further chaining and/or
3032 * buffer management.
3034 * WARNING! The done function can deallocate the buffer!
3036 if ((done_func = bio->bio_done) != NULL) {
3037 bio->bio_done = NULL;
3042 bio = bio->bio_prev;
3046 bp->b_cmd = BUF_CMD_DONE;
3049 * Only reads and writes are processed past this point.
3051 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3058 * Warning: softupdates may re-dirty the buffer.
3060 if (LIST_FIRST(&bp->b_dep) != NULL)
3063 if (bp->b_flags & B_VMIO) {
3069 struct vnode *vp = bp->b_vp;
3073 #if defined(VFS_BIO_DEBUG)
3074 if (vp->v_auxrefs == 0)
3075 panic("biodone: zero vnode hold count");
3076 if ((vp->v_flag & VOBJBUF) == 0)
3077 panic("biodone: vnode is not setup for merged cache");
3080 foff = bp->b_loffset;
3081 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3082 KASSERT(obj != NULL, ("biodone: missing VM object"));
3084 #if defined(VFS_BIO_DEBUG)
3085 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3086 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3087 obj->paging_in_progress, bp->b_xio.xio_npages);
3092 * Set B_CACHE if the op was a normal read and no error
3093 * occured. B_CACHE is set for writes in the b*write()
3096 iosize = bp->b_bcount - bp->b_resid;
3097 if (cmd == BUF_CMD_READ && (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3098 bp->b_flags |= B_CACHE;
3101 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3105 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3110 * cleanup bogus pages, restoring the originals. Since
3111 * the originals should still be wired, we don't have
3112 * to worry about interrupt/freeing races destroying
3113 * the VM object association.
3115 m = bp->b_xio.xio_pages[i];
3116 if (m == bogus_page) {
3118 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3120 panic("biodone: page disappeared");
3121 bp->b_xio.xio_pages[i] = m;
3122 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3123 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3125 #if defined(VFS_BIO_DEBUG)
3126 if (OFF_TO_IDX(foff) != m->pindex) {
3128 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
3129 (unsigned long)foff, m->pindex);
3134 * In the write case, the valid and clean bits are
3135 * already changed correctly ( see bdwrite() ), so we
3136 * only need to do this here in the read case.
3138 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3139 vfs_page_set_valid(bp, foff, i, m);
3141 vm_page_flag_clear(m, PG_ZERO);
3144 * when debugging new filesystems or buffer I/O methods, this
3145 * is the most common error that pops up. if you see this, you
3146 * have not set the page busy flag correctly!!!
3149 kprintf("biodone: page busy < 0, "
3150 "pindex: %d, foff: 0x(%x,%x), "
3151 "resid: %d, index: %d\n",
3152 (int) m->pindex, (int)(foff >> 32),
3153 (int) foff & 0xffffffff, resid, i);
3154 if (!vn_isdisk(vp, NULL))
3155 kprintf(" iosize: %ld, loffset: %lld, flags: 0x%08x, npages: %d\n",
3156 bp->b_vp->v_mount->mnt_stat.f_iosize,
3158 bp->b_flags, bp->b_xio.xio_npages);
3160 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3162 bp->b_flags, bp->b_xio.xio_npages);
3163 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3164 m->valid, m->dirty, m->wire_count);
3165 panic("biodone: page busy < 0");
3167 vm_page_io_finish(m);
3168 vm_object_pip_subtract(obj, 1);
3169 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3173 vm_object_pip_wakeupn(obj, 0);
3177 * For asynchronous completions, release the buffer now. The brelse
3178 * will do a wakeup there if necessary - so no need to do a wakeup
3179 * here in the async case. The sync case always needs to do a wakeup.
3182 if (bp->b_flags & B_ASYNC) {
3183 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
3196 * This routine is called in lieu of iodone in the case of
3197 * incomplete I/O. This keeps the busy status for pages
3201 vfs_unbusy_pages(struct buf *bp)
3205 runningbufwakeup(bp);
3206 if (bp->b_flags & B_VMIO) {
3207 struct vnode *vp = bp->b_vp;
3212 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3213 vm_page_t m = bp->b_xio.xio_pages[i];
3216 * When restoring bogus changes the original pages
3217 * should still be wired, so we are in no danger of
3218 * losing the object association and do not need
3219 * critical section protection particularly.
3221 if (m == bogus_page) {
3222 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3224 panic("vfs_unbusy_pages: page missing");
3226 bp->b_xio.xio_pages[i] = m;
3227 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3228 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3230 vm_object_pip_subtract(obj, 1);
3231 vm_page_flag_clear(m, PG_ZERO);
3232 vm_page_io_finish(m);
3234 vm_object_pip_wakeupn(obj, 0);
3239 * vfs_page_set_valid:
3241 * Set the valid bits in a page based on the supplied offset. The
3242 * range is restricted to the buffer's size.
3244 * This routine is typically called after a read completes.
3247 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3249 vm_ooffset_t soff, eoff;
3252 * Start and end offsets in buffer. eoff - soff may not cross a
3253 * page boundry or cross the end of the buffer. The end of the
3254 * buffer, in this case, is our file EOF, not the allocation size
3258 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3259 if (eoff > bp->b_loffset + bp->b_bcount)
3260 eoff = bp->b_loffset + bp->b_bcount;
3263 * Set valid range. This is typically the entire buffer and thus the
3267 vm_page_set_validclean(
3269 (vm_offset_t) (soff & PAGE_MASK),
3270 (vm_offset_t) (eoff - soff)
3278 * This routine is called before a device strategy routine.
3279 * It is used to tell the VM system that paging I/O is in
3280 * progress, and treat the pages associated with the buffer
3281 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3282 * flag is handled to make sure that the object doesn't become
3285 * Since I/O has not been initiated yet, certain buffer flags
3286 * such as B_ERROR or B_INVAL may be in an inconsistant state
3287 * and should be ignored.
3290 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3293 struct lwp *lp = curthread->td_lwp;
3296 * The buffer's I/O command must already be set. If reading,
3297 * B_CACHE must be 0 (double check against callers only doing
3298 * I/O when B_CACHE is 0).
3300 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3301 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3303 if (bp->b_flags & B_VMIO) {
3308 foff = bp->b_loffset;
3309 KASSERT(bp->b_loffset != NOOFFSET,
3310 ("vfs_busy_pages: no buffer offset"));
3314 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3315 vm_page_t m = bp->b_xio.xio_pages[i];
3316 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3321 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3322 vm_page_t m = bp->b_xio.xio_pages[i];
3324 vm_page_flag_clear(m, PG_ZERO);
3325 if ((bp->b_flags & B_CLUSTER) == 0) {
3326 vm_object_pip_add(obj, 1);
3327 vm_page_io_start(m);
3331 * When readying a vnode-backed buffer for a write
3332 * we must zero-fill any invalid portions of the
3335 * When readying a vnode-backed buffer for a read
3336 * we must replace any dirty pages with a bogus
3337 * page so we do not destroy dirty data when
3338 * filling in gaps. Dirty pages might not
3339 * necessarily be marked dirty yet, so use m->valid
3340 * as a reasonable test.
3342 * Bogus page replacement is, uh, bogus. We need
3343 * to find a better way.
3345 vm_page_protect(m, VM_PROT_NONE);
3346 if (bp->b_cmd == BUF_CMD_WRITE) {
3347 vfs_page_set_valid(bp, foff, i, m);
3348 } else if (m->valid == VM_PAGE_BITS_ALL) {
3349 bp->b_xio.xio_pages[i] = bogus_page;
3352 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3355 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3356 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3360 * This is the easiest place to put the process accounting for the I/O
3364 if (bp->b_cmd == BUF_CMD_READ)
3365 lp->lwp_ru.ru_inblock++;
3367 lp->lwp_ru.ru_oublock++;
3374 * Tell the VM system that the pages associated with this buffer
3375 * are clean. This is used for delayed writes where the data is
3376 * going to go to disk eventually without additional VM intevention.
3378 * Note that while we only really need to clean through to b_bcount, we
3379 * just go ahead and clean through to b_bufsize.
3382 vfs_clean_pages(struct buf *bp)
3386 if (bp->b_flags & B_VMIO) {
3389 foff = bp->b_loffset;
3390 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset"));
3391 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3392 vm_page_t m = bp->b_xio.xio_pages[i];
3393 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3394 vm_ooffset_t eoff = noff;
3396 if (eoff > bp->b_loffset + bp->b_bufsize)
3397 eoff = bp->b_loffset + bp->b_bufsize;
3398 vfs_page_set_valid(bp, foff, i, m);
3399 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3406 * vfs_bio_set_validclean:
3408 * Set the range within the buffer to valid and clean. The range is
3409 * relative to the beginning of the buffer, b_loffset. Note that
3410 * b_loffset itself may be offset from the beginning of the first page.
3414 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3416 if (bp->b_flags & B_VMIO) {
3421 * Fixup base to be relative to beginning of first page.
3422 * Set initial n to be the maximum number of bytes in the
3423 * first page that can be validated.
3426 base += (bp->b_loffset & PAGE_MASK);
3427 n = PAGE_SIZE - (base & PAGE_MASK);
3429 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) {
3430 vm_page_t m = bp->b_xio.xio_pages[i];
3435 vm_page_set_validclean(m, base & PAGE_MASK, n);
3446 * Clear a buffer. This routine essentially fakes an I/O, so we need
3447 * to clear B_ERROR and B_INVAL.
3449 * Note that while we only theoretically need to clear through b_bcount,
3450 * we go ahead and clear through b_bufsize.
3454 vfs_bio_clrbuf(struct buf *bp)
3458 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3459 bp->b_flags &= ~(B_INVAL|B_ERROR);
3460 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3461 (bp->b_loffset & PAGE_MASK) == 0) {
3462 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3463 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3467 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3468 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3469 bzero(bp->b_data, bp->b_bufsize);
3470 bp->b_xio.xio_pages[0]->valid |= mask;
3475 ea = sa = bp->b_data;
3476 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3477 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3478 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3479 ea = (caddr_t)(vm_offset_t)ulmin(
3480 (u_long)(vm_offset_t)ea,
3481 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3482 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3483 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3485 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3486 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3490 for (; sa < ea; sa += DEV_BSIZE, j++) {
3491 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3492 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3493 bzero(sa, DEV_BSIZE);
3496 bp->b_xio.xio_pages[i]->valid |= mask;
3497 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
3506 * vm_hold_load_pages:
3508 * Load pages into the buffer's address space. The pages are
3509 * allocated from the kernel object in order to reduce interference
3510 * with the any VM paging I/O activity. The range of loaded
3511 * pages will be wired.
3513 * If a page cannot be allocated, the 'pagedaemon' is woken up to
3514 * retrieve the full range (to - from) of pages.
3518 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3524 to = round_page(to);
3525 from = round_page(from);
3526 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3528 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3533 * Note: must allocate system pages since blocking here
3534 * could intefere with paging I/O, no matter which
3537 p = vm_page_alloc(&kernel_object,
3539 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
3541 vm_pageout_deficit += (to - from) >> PAGE_SHIFT;
3546 p->valid = VM_PAGE_BITS_ALL;
3547 vm_page_flag_clear(p, PG_ZERO);
3548 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3549 bp->b_xio.xio_pages[index] = p;
3552 bp->b_xio.xio_npages = index;
3556 * vm_hold_free_pages:
3558 * Return pages associated with the buffer back to the VM system.
3560 * The range of pages underlying the buffer's address space will
3561 * be unmapped and un-wired.
3564 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3568 int index, newnpages;
3570 from = round_page(from);
3571 to = round_page(to);
3572 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3574 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3575 p = bp->b_xio.xio_pages[index];
3576 if (p && (index < bp->b_xio.xio_npages)) {
3578 kprintf("vm_hold_free_pages: doffset: %lld, loffset: %lld\n",
3579 bp->b_bio2.bio_offset, bp->b_loffset);
3581 bp->b_xio.xio_pages[index] = NULL;
3584 vm_page_unwire(p, 0);
3588 bp->b_xio.xio_npages = newnpages;
3594 * Map a user buffer into KVM via a pbuf. On return the buffer's
3595 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
3599 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
3610 * bp had better have a command and it better be a pbuf.
3612 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3613 KKASSERT(bp->b_flags & B_PAGING);
3619 * Map the user data into KVM. Mappings have to be page-aligned.
3621 addr = (caddr_t)trunc_page((vm_offset_t)udata);
3624 vmprot = VM_PROT_READ;
3625 if (bp->b_cmd == BUF_CMD_READ)
3626 vmprot |= VM_PROT_WRITE;
3628 while (addr < udata + bytes) {
3630 * Do the vm_fault if needed; do the copy-on-write thing
3631 * when reading stuff off device into memory.
3633 * vm_fault_page*() returns a held VM page.
3635 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
3636 va = trunc_page(va);
3638 m = vm_fault_page_quick(va, vmprot, &error);
3640 for (i = 0; i < pidx; ++i) {
3641 vm_page_unhold(bp->b_xio.xio_pages[i]);
3642 bp->b_xio.xio_pages[i] = NULL;
3646 bp->b_xio.xio_pages[pidx] = m;
3652 * Map the page array and set the buffer fields to point to
3653 * the mapped data buffer.
3655 if (pidx > btoc(MAXPHYS))
3656 panic("vmapbuf: mapped more than MAXPHYS");
3657 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
3659 bp->b_xio.xio_npages = pidx;
3660 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
3661 bp->b_bcount = bytes;
3662 bp->b_bufsize = bytes;
3669 * Free the io map PTEs associated with this IO operation.
3670 * We also invalidate the TLB entries and restore the original b_addr.
3673 vunmapbuf(struct buf *bp)
3678 KKASSERT(bp->b_flags & B_PAGING);
3680 npages = bp->b_xio.xio_npages;
3681 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3682 for (pidx = 0; pidx < npages; ++pidx) {
3683 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
3684 bp->b_xio.xio_pages[pidx] = NULL;
3686 bp->b_xio.xio_npages = 0;
3687 bp->b_data = bp->b_kvabase;
3691 * Scan all buffers in the system and issue the callback.
3694 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
3700 for (n = 0; n < nbuf; ++n) {
3701 if ((error = callback(&buf[n], info)) < 0) {
3711 * print out statistics from the current status of the buffer pool
3712 * this can be toggeled by the system control option debug.syncprt
3721 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
3722 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
3724 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
3726 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
3729 TAILQ_FOREACH(bp, dp, b_freelist) {
3730 counts[bp->b_bufsize/PAGE_SIZE]++;
3734 kprintf("%s: total-%d", bname[i], count);
3735 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
3737 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
3745 DB_SHOW_COMMAND(buffer, db_show_buffer)
3748 struct buf *bp = (struct buf *)addr;
3751 db_printf("usage: show buffer <addr>\n");
3755 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3756 db_printf("b_cmd = %d\n", bp->b_cmd);
3757 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
3758 "b_resid = %d\n, b_data = %p, "
3759 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
3760 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3761 bp->b_data, bp->b_bio2.bio_offset, (bp->b_bio2.bio_next ? bp->b_bio2.bio_next->bio_offset : (off_t)-1));
3762 if (bp->b_xio.xio_npages) {
3764 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
3765 bp->b_xio.xio_npages);
3766 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3768 m = bp->b_xio.xio_pages[i];
3769 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3770 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3771 if ((i + 1) < bp->b_xio.xio_npages)