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 1.102 2008/05/09 07:24:45 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;
113 * These are all static, but make the ones we export globals so we do
114 * not need to use compiler magic.
116 int bufspace, maxbufspace,
117 bufmallocspace, maxbufmallocspace, lobufspace, hibufspace;
118 static int bufreusecnt, bufdefragcnt, buffreekvacnt;
119 static int lorunningspace, hirunningspace, runningbufreq;
120 int numdirtybuffers, numdirtybuffershw, lodirtybuffers, hidirtybuffers;
121 int runningbufspace, runningbufcount;
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, nbuf, CTLFLAG_RD, &nbuf, 0,
150 "Total number of buffers in buffer cache");
151 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, &numdirtybuffers, 0,
152 "Pending number of dirty buffers (all)");
153 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffershw, CTLFLAG_RD, &numdirtybuffershw, 0,
154 "Pending number of dirty buffers (heavy weight)");
155 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0,
156 "Number of free buffers on the buffer cache free list");
157 SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
158 "I/O bytes currently in progress due to asynchronous writes");
159 SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
160 "I/O buffers currently in progress due to asynchronous writes");
161 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
162 "Hard limit on maximum amount of memory usable for buffer space");
163 SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
164 "Soft limit on maximum amount of memory usable for buffer space");
165 SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
166 "Minimum amount of memory to reserve for system buffer space");
167 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
168 "Amount of memory available for buffers");
169 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
170 0, "Maximum amount of memory reserved for buffers using malloc");
171 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
172 "Amount of memory left for buffers using malloc-scheme");
173 SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
174 "New buffer header acquisition requests");
175 SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
176 0, "New buffer header acquisition restarts");
177 SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
178 "Buffer acquisition restarts due to fragmented buffer map");
179 SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
180 "Amount of time KVA space was deallocated in an arbitrary buffer");
181 SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
182 "Amount of time buffer re-use operations were successful");
183 SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
184 "sizeof(struct buf)");
186 char *buf_wmesg = BUF_WMESG;
188 extern int vm_swap_size;
190 #define VFS_BIO_NEED_ANY 0x01 /* any freeable buffer */
191 #define VFS_BIO_NEED_DIRTYFLUSH 0x02 /* waiting for dirty buffer flush */
192 #define VFS_BIO_NEED_FREE 0x04 /* wait for free bufs, hi hysteresis */
193 #define VFS_BIO_NEED_BUFSPACE 0x08 /* wait for buf space, lo hysteresis */
198 * If someone is blocked due to there being too many dirty buffers,
199 * and numdirtybuffers is now reasonable, wake them up.
204 if (runningbufcount + numdirtybuffers <=
205 (lodirtybuffers + hidirtybuffers) / 2) {
206 if (needsbuffer & VFS_BIO_NEED_DIRTYFLUSH) {
207 spin_lock_wr(&needsbuffer_spin);
208 needsbuffer &= ~VFS_BIO_NEED_DIRTYFLUSH;
209 spin_unlock_wr(&needsbuffer_spin);
210 wakeup(&needsbuffer);
218 * Called when buffer space is potentially available for recovery.
219 * getnewbuf() will block on this flag when it is unable to free
220 * sufficient buffer space. Buffer space becomes recoverable when
221 * bp's get placed back in the queues.
228 * If someone is waiting for BUF space, wake them up. Even
229 * though we haven't freed the kva space yet, the waiting
230 * process will be able to now.
232 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
233 spin_lock_wr(&needsbuffer_spin);
234 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
235 spin_unlock_wr(&needsbuffer_spin);
236 wakeup(&needsbuffer);
243 * Accounting for I/O in progress.
247 runningbufwakeup(struct buf *bp)
249 if (bp->b_runningbufspace) {
250 runningbufspace -= bp->b_runningbufspace;
252 bp->b_runningbufspace = 0;
253 if (runningbufreq && runningbufspace <= lorunningspace) {
255 wakeup(&runningbufreq);
264 * Called when a buffer has been added to one of the free queues to
265 * account for the buffer and to wakeup anyone waiting for free buffers.
266 * This typically occurs when large amounts of metadata are being handled
267 * by the buffer cache ( else buffer space runs out first, usually ).
275 spin_lock_wr(&needsbuffer_spin);
276 needsbuffer &= ~VFS_BIO_NEED_ANY;
277 if (numfreebuffers >= hifreebuffers)
278 needsbuffer &= ~VFS_BIO_NEED_FREE;
279 spin_unlock_wr(&needsbuffer_spin);
280 wakeup(&needsbuffer);
285 * waitrunningbufspace()
287 * runningbufspace is a measure of the amount of I/O currently
288 * running. This routine is used in async-write situations to
289 * prevent creating huge backups of pending writes to a device.
290 * Only asynchronous writes are governed by this function.
292 * Reads will adjust runningbufspace, but will not block based on it.
293 * The read load has a side effect of reducing the allowed write load.
295 * This does NOT turn an async write into a sync write. It waits
296 * for earlier writes to complete and generally returns before the
297 * caller's write has reached the device.
300 waitrunningbufspace(void)
302 if (runningbufspace > hirunningspace) {
304 while (runningbufspace > hirunningspace) {
306 tsleep(&runningbufreq, 0, "wdrain", 0);
313 * vfs_buf_test_cache:
315 * Called when a buffer is extended. This function clears the B_CACHE
316 * bit if the newly extended portion of the buffer does not contain
321 vfs_buf_test_cache(struct buf *bp,
322 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
325 if (bp->b_flags & B_CACHE) {
326 int base = (foff + off) & PAGE_MASK;
327 if (vm_page_is_valid(m, base, size) == 0)
328 bp->b_flags &= ~B_CACHE;
335 * Wake up the buffer daemon if the number of outstanding dirty buffers
336 * is above specified threshold 'dirtybuflevel'.
338 * The buffer daemons are explicitly woken up when (a) the pending number
339 * of dirty buffers exceeds the recovery and stall mid-point value,
340 * (b) during bwillwrite() or (c) buf freelist was exhausted.
342 * The buffer daemons will generally not stop flushing until the dirty
343 * buffer count goes below lodirtybuffers.
347 bd_wakeup(int dirtybuflevel)
349 if (bd_request == 0 && numdirtybuffers &&
350 runningbufcount + numdirtybuffers >= dirtybuflevel) {
351 spin_lock_wr(&needsbuffer_spin);
353 spin_unlock_wr(&needsbuffer_spin);
356 if (bd_request_hw == 0 && numdirtybuffershw &&
357 numdirtybuffershw >= dirtybuflevel) {
358 spin_lock_wr(&needsbuffer_spin);
360 spin_unlock_wr(&needsbuffer_spin);
361 wakeup(&bd_request_hw);
368 * Speed up the buffer cache flushing process.
381 * Load time initialisation of the buffer cache, called from machine
382 * dependant initialization code.
388 vm_offset_t bogus_offset;
391 spin_init(&needsbuffer_spin);
393 /* next, make a null set of free lists */
394 for (i = 0; i < BUFFER_QUEUES; i++)
395 TAILQ_INIT(&bufqueues[i]);
397 /* finally, initialize each buffer header and stick on empty q */
398 for (i = 0; i < nbuf; i++) {
400 bzero(bp, sizeof *bp);
401 bp->b_flags = B_INVAL; /* we're just an empty header */
402 bp->b_cmd = BUF_CMD_DONE;
403 bp->b_qindex = BQUEUE_EMPTY;
405 xio_init(&bp->b_xio);
408 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
412 * maxbufspace is the absolute maximum amount of buffer space we are
413 * allowed to reserve in KVM and in real terms. The absolute maximum
414 * is nominally used by buf_daemon. hibufspace is the nominal maximum
415 * used by most other processes. The differential is required to
416 * ensure that buf_daemon is able to run when other processes might
417 * be blocked waiting for buffer space.
419 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
420 * this may result in KVM fragmentation which is not handled optimally
423 maxbufspace = nbuf * BKVASIZE;
424 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
425 lobufspace = hibufspace - MAXBSIZE;
427 lorunningspace = 512 * 1024;
428 hirunningspace = 1024 * 1024;
431 * Limit the amount of malloc memory since it is wired permanently into
432 * the kernel space. Even though this is accounted for in the buffer
433 * allocation, we don't want the malloced region to grow uncontrolled.
434 * The malloc scheme improves memory utilization significantly on average
435 * (small) directories.
437 maxbufmallocspace = hibufspace / 20;
440 * Reduce the chance of a deadlock occuring by limiting the number
441 * of delayed-write dirty buffers we allow to stack up.
443 hidirtybuffers = nbuf / 4 + 20;
445 numdirtybuffershw = 0;
447 * To support extreme low-memory systems, make sure hidirtybuffers cannot
448 * eat up all available buffer space. This occurs when our minimum cannot
449 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
450 * BKVASIZE'd (8K) buffers.
452 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
453 hidirtybuffers >>= 1;
455 lodirtybuffers = hidirtybuffers / 2;
458 * Try to keep the number of free buffers in the specified range,
459 * and give special processes (e.g. like buf_daemon) access to an
462 lofreebuffers = nbuf / 18 + 5;
463 hifreebuffers = 2 * lofreebuffers;
464 numfreebuffers = nbuf;
467 * Maximum number of async ops initiated per buf_daemon loop. This is
468 * somewhat of a hack at the moment, we really need to limit ourselves
469 * based on the number of bytes of I/O in-transit that were initiated
473 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
474 bogus_page = vm_page_alloc(&kernel_object,
475 (bogus_offset >> PAGE_SHIFT),
477 vmstats.v_wire_count++;
482 * Initialize the embedded bio structures
485 initbufbio(struct buf *bp)
487 bp->b_bio1.bio_buf = bp;
488 bp->b_bio1.bio_prev = NULL;
489 bp->b_bio1.bio_offset = NOOFFSET;
490 bp->b_bio1.bio_next = &bp->b_bio2;
491 bp->b_bio1.bio_done = NULL;
493 bp->b_bio2.bio_buf = bp;
494 bp->b_bio2.bio_prev = &bp->b_bio1;
495 bp->b_bio2.bio_offset = NOOFFSET;
496 bp->b_bio2.bio_next = NULL;
497 bp->b_bio2.bio_done = NULL;
501 * Reinitialize the embedded bio structures as well as any additional
502 * translation cache layers.
505 reinitbufbio(struct buf *bp)
509 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
510 bio->bio_done = NULL;
511 bio->bio_offset = NOOFFSET;
516 * Push another BIO layer onto an existing BIO and return it. The new
517 * BIO layer may already exist, holding cached translation data.
520 push_bio(struct bio *bio)
524 if ((nbio = bio->bio_next) == NULL) {
525 int index = bio - &bio->bio_buf->b_bio_array[0];
526 if (index >= NBUF_BIO - 1) {
527 panic("push_bio: too many layers bp %p\n",
530 nbio = &bio->bio_buf->b_bio_array[index + 1];
531 bio->bio_next = nbio;
532 nbio->bio_prev = bio;
533 nbio->bio_buf = bio->bio_buf;
534 nbio->bio_offset = NOOFFSET;
535 nbio->bio_done = NULL;
536 nbio->bio_next = NULL;
538 KKASSERT(nbio->bio_done == NULL);
543 pop_bio(struct bio *bio)
549 clearbiocache(struct bio *bio)
552 bio->bio_offset = NOOFFSET;
560 * Free the KVA allocation for buffer 'bp'.
562 * Must be called from a critical section as this is the only locking for
565 * Since this call frees up buffer space, we call bufspacewakeup().
568 bfreekva(struct buf *bp)
574 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
575 vm_map_lock(&buffer_map);
576 bufspace -= bp->b_kvasize;
577 vm_map_delete(&buffer_map,
578 (vm_offset_t) bp->b_kvabase,
579 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
582 vm_map_unlock(&buffer_map);
583 vm_map_entry_release(count);
592 * Remove the buffer from the appropriate free list.
595 bremfree(struct buf *bp)
600 old_qindex = bp->b_qindex;
602 if (bp->b_qindex != BQUEUE_NONE) {
603 KASSERT(BUF_REFCNTNB(bp) == 1,
604 ("bremfree: bp %p not locked",bp));
605 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
606 bp->b_qindex = BQUEUE_NONE;
608 if (BUF_REFCNTNB(bp) <= 1)
609 panic("bremfree: removing a buffer not on a queue");
613 * Fixup numfreebuffers count. If the buffer is invalid or not
614 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
615 * the buffer was free and we must decrement numfreebuffers.
617 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
620 case BQUEUE_DIRTY_HW:
623 case BQUEUE_EMPTYKVA:
637 * Get a buffer with the specified data. Look in the cache first. We
638 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
639 * is set, the buffer is valid and we do not have to do anything ( see
643 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
647 bp = getblk(vp, loffset, size, 0, 0);
650 /* if not found in cache, do some I/O */
651 if ((bp->b_flags & B_CACHE) == 0) {
652 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp));
653 bp->b_flags &= ~(B_ERROR | B_INVAL);
654 bp->b_cmd = BUF_CMD_READ;
655 vfs_busy_pages(vp, bp);
656 vn_strategy(vp, &bp->b_bio1);
657 return (biowait(bp));
665 * Operates like bread, but also starts asynchronous I/O on
666 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
667 * to initiating I/O . If B_CACHE is set, the buffer is valid
668 * and we do not have to do anything.
671 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
672 int *rabsize, int cnt, struct buf **bpp)
674 struct buf *bp, *rabp;
676 int rv = 0, readwait = 0;
678 *bpp = bp = getblk(vp, loffset, size, 0, 0);
680 /* if not found in cache, do some I/O */
681 if ((bp->b_flags & B_CACHE) == 0) {
682 bp->b_flags &= ~(B_ERROR | B_INVAL);
683 bp->b_cmd = BUF_CMD_READ;
684 vfs_busy_pages(vp, bp);
685 vn_strategy(vp, &bp->b_bio1);
689 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
690 if (inmem(vp, *raoffset))
692 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
694 if ((rabp->b_flags & B_CACHE) == 0) {
695 rabp->b_flags |= B_ASYNC;
696 rabp->b_flags &= ~(B_ERROR | B_INVAL);
697 rabp->b_cmd = BUF_CMD_READ;
698 vfs_busy_pages(vp, rabp);
700 vn_strategy(vp, &rabp->b_bio1);
715 * Write, release buffer on completion. (Done by iodone
716 * if async). Do not bother writing anything if the buffer
719 * Note that we set B_CACHE here, indicating that buffer is
720 * fully valid and thus cacheable. This is true even of NFS
721 * now so we set it generally. This could be set either here
722 * or in biodone() since the I/O is synchronous. We put it
726 bwrite(struct buf *bp)
730 if (bp->b_flags & B_INVAL) {
735 oldflags = bp->b_flags;
737 if (BUF_REFCNTNB(bp) == 0)
738 panic("bwrite: buffer is not busy???");
741 /* Mark the buffer clean */
744 bp->b_flags &= ~B_ERROR;
745 bp->b_flags |= B_CACHE;
746 bp->b_cmd = BUF_CMD_WRITE;
747 vfs_busy_pages(bp->b_vp, bp);
750 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
751 * valid for vnode-backed buffers.
753 bp->b_runningbufspace = bp->b_bufsize;
754 if (bp->b_runningbufspace) {
755 runningbufspace += bp->b_runningbufspace;
760 if (oldflags & B_ASYNC)
762 vn_strategy(bp->b_vp, &bp->b_bio1);
764 if ((oldflags & B_ASYNC) == 0) {
765 int rtval = biowait(bp);
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 (runningbufcount + numdirtybuffers >= hidirtybuffers / 2) {
973 while (runningbufcount + numdirtybuffers >= hidirtybuffers) {
975 spin_lock_wr(&needsbuffer_spin);
976 if (runningbufcount + numdirtybuffers >=
978 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
979 msleep(&needsbuffer, &needsbuffer_spin, 0,
982 spin_unlock_wr(&needsbuffer_spin);
987 else if (runningbufcount + numdirtybuffershw > hidirtybuffers / 2) {
990 while (runningbufcount + numdirtybuffershw > hidirtybuffers) {
991 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
992 tsleep(&needsbuffer, slpflags, "newbuf",
1000 * buf_dirty_count_severe:
1002 * Return true if we have too many dirty buffers.
1005 buf_dirty_count_severe(void)
1007 return(runningbufcount + numdirtybuffers >= hidirtybuffers);
1013 * Release a busy buffer and, if requested, free its resources. The
1014 * buffer will be stashed in the appropriate bufqueue[] allowing it
1015 * to be accessed later as a cache entity or reused for other purposes.
1018 brelse(struct buf *bp)
1021 int saved_flags = bp->b_flags;
1024 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1029 * If B_NOCACHE is set we are being asked to destroy the buffer and
1030 * its backing store. Clear B_DELWRI.
1032 * B_NOCACHE is set in two cases: (1) when the caller really wants
1033 * to destroy the buffer and backing store and (2) when the caller
1034 * wants to destroy the buffer and backing store after a write
1037 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1041 if (bp->b_flags & B_LOCKED)
1042 bp->b_flags &= ~B_ERROR;
1045 * If a write error occurs and the caller does not want to throw
1046 * away the buffer, redirty the buffer. This will also clear
1049 if (bp->b_cmd == BUF_CMD_WRITE &&
1050 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
1052 * Failed write, redirty. Must clear B_ERROR to prevent
1053 * pages from being scrapped. If B_INVAL is set then
1054 * this case is not run and the next case is run to
1055 * destroy the buffer. B_INVAL can occur if the buffer
1056 * is outside the range supported by the underlying device.
1058 bp->b_flags &= ~B_ERROR;
1060 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1061 (bp->b_bufsize <= 0) || bp->b_cmd == BUF_CMD_FREEBLKS) {
1063 * Either a failed I/O or we were asked to free or not
1066 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1067 * buffer cannot be immediately freed.
1069 bp->b_flags |= B_INVAL;
1070 if (LIST_FIRST(&bp->b_dep) != NULL)
1072 if (bp->b_flags & B_DELWRI) {
1074 if (bp->b_flags & B_HEAVY)
1075 --numdirtybuffershw;
1078 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1082 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1083 * If vfs_vmio_release() is called with either bit set, the
1084 * underlying pages may wind up getting freed causing a previous
1085 * write (bdwrite()) to get 'lost' because pages associated with
1086 * a B_DELWRI bp are marked clean. Pages associated with a
1087 * B_LOCKED buffer may be mapped by the filesystem.
1089 * If we want to release the buffer ourselves (rather then the
1090 * originator asking us to release it), give the originator a
1091 * chance to countermand the release by setting B_LOCKED.
1093 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1094 * if B_DELWRI is set.
1096 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1097 * on pages to return pages to the VM page queues.
1099 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1100 bp->b_flags &= ~B_RELBUF;
1101 } else if (vm_page_count_severe()) {
1102 if (LIST_FIRST(&bp->b_dep) != NULL)
1104 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1105 bp->b_flags &= ~B_RELBUF;
1107 bp->b_flags |= B_RELBUF;
1111 * At this point destroying the buffer is governed by the B_INVAL
1112 * or B_RELBUF flags.
1114 bp->b_cmd = BUF_CMD_DONE;
1117 * VMIO buffer rundown. Make sure the VM page array is restored
1118 * after an I/O may have replaces some of the pages with bogus pages
1119 * in order to not destroy dirty pages in a fill-in read.
1121 * Note that due to the code above, if a buffer is marked B_DELWRI
1122 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1123 * B_INVAL may still be set, however.
1125 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1126 * but not the backing store. B_NOCACHE will destroy the backing
1129 * Note that dirty NFS buffers contain byte-granular write ranges
1130 * and should not be destroyed w/ B_INVAL even if the backing store
1133 if (bp->b_flags & B_VMIO) {
1135 * Rundown for VMIO buffers which are not dirty NFS buffers.
1147 * Get the base offset and length of the buffer. Note that
1148 * in the VMIO case if the buffer block size is not
1149 * page-aligned then b_data pointer may not be page-aligned.
1150 * But our b_xio.xio_pages array *IS* page aligned.
1152 * block sizes less then DEV_BSIZE (usually 512) are not
1153 * supported due to the page granularity bits (m->valid,
1154 * m->dirty, etc...).
1156 * See man buf(9) for more information
1159 resid = bp->b_bufsize;
1160 foff = bp->b_loffset;
1162 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1163 m = bp->b_xio.xio_pages[i];
1164 vm_page_flag_clear(m, PG_ZERO);
1166 * If we hit a bogus page, fixup *all* of them
1167 * now. Note that we left these pages wired
1168 * when we removed them so they had better exist,
1169 * and they cannot be ripped out from under us so
1170 * no critical section protection is necessary.
1172 if (m == bogus_page) {
1174 poff = OFF_TO_IDX(bp->b_loffset);
1176 for (j = i; j < bp->b_xio.xio_npages; j++) {
1179 mtmp = bp->b_xio.xio_pages[j];
1180 if (mtmp == bogus_page) {
1181 mtmp = vm_page_lookup(obj, poff + j);
1183 panic("brelse: page missing");
1185 bp->b_xio.xio_pages[j] = mtmp;
1189 if ((bp->b_flags & B_INVAL) == 0) {
1190 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1191 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1193 m = bp->b_xio.xio_pages[i];
1197 * Invalidate the backing store if B_NOCACHE is set
1198 * (e.g. used with vinvalbuf()). If this is NFS
1199 * we impose a requirement that the block size be
1200 * a multiple of PAGE_SIZE and create a temporary
1201 * hack to basically invalidate the whole page. The
1202 * problem is that NFS uses really odd buffer sizes
1203 * especially when tracking piecemeal writes and
1204 * it also vinvalbuf()'s a lot, which would result
1205 * in only partial page validation and invalidation
1206 * here. If the file page is mmap()'d, however,
1207 * all the valid bits get set so after we invalidate
1208 * here we would end up with weird m->valid values
1209 * like 0xfc. nfs_getpages() can't handle this so
1210 * we clear all the valid bits for the NFS case
1211 * instead of just some of them.
1213 * The real bug is the VM system having to set m->valid
1214 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1215 * itself is an artifact of the whole 512-byte
1216 * granular mess that exists to support odd block
1217 * sizes and UFS meta-data block sizes (e.g. 6144).
1218 * A complete rewrite is required.
1220 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1221 int poffset = foff & PAGE_MASK;
1224 presid = PAGE_SIZE - poffset;
1225 if (bp->b_vp->v_tag == VT_NFS &&
1226 bp->b_vp->v_type == VREG) {
1228 } else if (presid > resid) {
1231 KASSERT(presid >= 0, ("brelse: extra page"));
1232 vm_page_set_invalid(m, poffset, presid);
1234 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1235 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1237 if (bp->b_flags & (B_INVAL | B_RELBUF))
1238 vfs_vmio_release(bp);
1241 * Rundown for non-VMIO buffers.
1243 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1246 kprintf("brelse bp %p %08x/%08x: Warning, caught and fixed brelvp bug\n", bp, saved_flags, bp->b_flags);
1250 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1256 if (bp->b_qindex != BQUEUE_NONE)
1257 panic("brelse: free buffer onto another queue???");
1258 if (BUF_REFCNTNB(bp) > 1) {
1259 /* Temporary panic to verify exclusive locking */
1260 /* This panic goes away when we allow shared refs */
1261 panic("brelse: multiple refs");
1262 /* do not release to free list */
1269 * Figure out the correct queue to place the cleaned up buffer on.
1270 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1271 * disassociated from their vnode.
1273 if (bp->b_flags & B_LOCKED) {
1275 * Buffers that are locked are placed in the locked queue
1276 * immediately, regardless of their state.
1278 bp->b_qindex = BQUEUE_LOCKED;
1279 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1280 } else if (bp->b_bufsize == 0) {
1282 * Buffers with no memory. Due to conditionals near the top
1283 * of brelse() such buffers should probably already be
1284 * marked B_INVAL and disassociated from their vnode.
1286 bp->b_flags |= B_INVAL;
1287 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1288 KKASSERT((bp->b_flags & B_HASHED) == 0);
1289 if (bp->b_kvasize) {
1290 bp->b_qindex = BQUEUE_EMPTYKVA;
1292 bp->b_qindex = BQUEUE_EMPTY;
1294 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1295 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) {
1297 * Buffers with junk contents. Again these buffers had better
1298 * already be disassociated from their vnode.
1300 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1301 KKASSERT((bp->b_flags & B_HASHED) == 0);
1302 bp->b_flags |= B_INVAL;
1303 bp->b_qindex = BQUEUE_CLEAN;
1304 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1307 * Remaining buffers. These buffers are still associated with
1310 switch(bp->b_flags & (B_DELWRI|B_HEAVY|B_AGE)) {
1311 case B_DELWRI | B_AGE:
1312 bp->b_qindex = BQUEUE_DIRTY;
1313 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1316 bp->b_qindex = BQUEUE_DIRTY;
1317 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1319 case B_DELWRI | B_HEAVY | B_AGE:
1320 bp->b_qindex = BQUEUE_DIRTY_HW;
1321 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_DIRTY_HW], bp,
1324 case B_DELWRI | B_HEAVY:
1325 bp->b_qindex = BQUEUE_DIRTY_HW;
1326 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1329 case B_HEAVY | B_AGE:
1331 bp->b_qindex = BQUEUE_CLEAN;
1332 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1335 bp->b_qindex = BQUEUE_CLEAN;
1336 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1342 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1343 * on the correct queue.
1345 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1349 * Fixup numfreebuffers count. The bp is on an appropriate queue
1350 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1351 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1352 * if B_INVAL is set ).
1354 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1358 * Something we can maybe free or reuse
1360 if (bp->b_bufsize || bp->b_kvasize)
1364 * Clean up temporary flags and unlock the buffer.
1366 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF |
1375 * Release a buffer back to the appropriate queue but do not try to free
1376 * it. The buffer is expected to be used again soon.
1378 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1379 * biodone() to requeue an async I/O on completion. It is also used when
1380 * known good buffers need to be requeued but we think we may need the data
1383 * XXX we should be able to leave the B_RELBUF hint set on completion.
1386 bqrelse(struct buf *bp)
1390 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1392 if (bp->b_qindex != BQUEUE_NONE)
1393 panic("bqrelse: free buffer onto another queue???");
1394 if (BUF_REFCNTNB(bp) > 1) {
1395 /* do not release to free list */
1396 panic("bqrelse: multiple refs");
1401 if (bp->b_flags & B_LOCKED) {
1403 * Locked buffers are released to the locked queue. However,
1404 * if the buffer is dirty it will first go into the dirty
1405 * queue and later on after the I/O completes successfully it
1406 * will be released to the locked queue.
1408 bp->b_flags &= ~B_ERROR;
1409 bp->b_qindex = BQUEUE_LOCKED;
1410 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1411 } else if (bp->b_flags & B_DELWRI) {
1412 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1413 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1414 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1415 } else if (vm_page_count_severe()) {
1417 * We are too low on memory, we have to try to free the
1418 * buffer (most importantly: the wired pages making up its
1419 * backing store) *now*.
1425 bp->b_qindex = BQUEUE_CLEAN;
1426 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1429 if ((bp->b_flags & B_LOCKED) == 0 &&
1430 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1435 * Something we can maybe free or reuse.
1437 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1441 * Final cleanup and unlock. Clear bits that are only used while a
1442 * buffer is actively locked.
1444 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF);
1452 * Return backing pages held by the buffer 'bp' back to the VM system
1453 * if possible. The pages are freed if they are no longer valid or
1454 * attempt to free if it was used for direct I/O otherwise they are
1455 * sent to the page cache.
1457 * Pages that were marked busy are left alone and skipped.
1459 * The KVA mapping (b_data) for the underlying pages is removed by
1463 vfs_vmio_release(struct buf *bp)
1469 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1470 m = bp->b_xio.xio_pages[i];
1471 bp->b_xio.xio_pages[i] = NULL;
1473 * In order to keep page LRU ordering consistent, put
1474 * everything on the inactive queue.
1476 vm_page_unwire(m, 0);
1478 * We don't mess with busy pages, it is
1479 * the responsibility of the process that
1480 * busied the pages to deal with them.
1482 if ((m->flags & PG_BUSY) || (m->busy != 0))
1485 if (m->wire_count == 0) {
1486 vm_page_flag_clear(m, PG_ZERO);
1488 * Might as well free the page if we can and it has
1489 * no valid data. We also free the page if the
1490 * buffer was used for direct I/O.
1492 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1493 m->hold_count == 0) {
1495 vm_page_protect(m, VM_PROT_NONE);
1497 } else if (bp->b_flags & B_DIRECT) {
1498 vm_page_try_to_free(m);
1499 } else if (vm_page_count_severe()) {
1500 vm_page_try_to_cache(m);
1505 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1506 if (bp->b_bufsize) {
1510 bp->b_xio.xio_npages = 0;
1511 bp->b_flags &= ~B_VMIO;
1512 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1520 * Implement clustered async writes for clearing out B_DELWRI buffers.
1521 * This is much better then the old way of writing only one buffer at
1522 * a time. Note that we may not be presented with the buffers in the
1523 * correct order, so we search for the cluster in both directions.
1525 * The buffer is locked on call.
1528 vfs_bio_awrite(struct buf *bp)
1532 off_t loffset = bp->b_loffset;
1533 struct vnode *vp = bp->b_vp;
1541 * right now we support clustered writing only to regular files. If
1542 * we find a clusterable block we could be in the middle of a cluster
1543 * rather then at the beginning.
1545 * NOTE: b_bio1 contains the logical loffset and is aliased
1546 * to b_loffset. b_bio2 contains the translated block number.
1548 if ((vp->v_type == VREG) &&
1549 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1550 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1552 size = vp->v_mount->mnt_stat.f_iosize;
1554 for (i = size; i < MAXPHYS; i += size) {
1555 if ((bpa = findblk(vp, loffset + i)) &&
1556 BUF_REFCNT(bpa) == 0 &&
1557 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1558 (B_DELWRI | B_CLUSTEROK)) &&
1559 (bpa->b_bufsize == size)) {
1560 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1561 (bpa->b_bio2.bio_offset !=
1562 bp->b_bio2.bio_offset + i))
1568 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1569 if ((bpa = findblk(vp, loffset - j)) &&
1570 BUF_REFCNT(bpa) == 0 &&
1571 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1572 (B_DELWRI | B_CLUSTEROK)) &&
1573 (bpa->b_bufsize == size)) {
1574 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1575 (bpa->b_bio2.bio_offset !=
1576 bp->b_bio2.bio_offset - j))
1585 * this is a possible cluster write
1587 if (nbytes != size) {
1589 nwritten = cluster_wbuild(vp, size,
1590 loffset - j, nbytes);
1597 bp->b_flags |= B_ASYNC;
1601 * default (old) behavior, writing out only one block
1603 * XXX returns b_bufsize instead of b_bcount for nwritten?
1605 nwritten = bp->b_bufsize;
1614 * Find and initialize a new buffer header, freeing up existing buffers
1615 * in the bufqueues as necessary. The new buffer is returned locked.
1617 * Important: B_INVAL is not set. If the caller wishes to throw the
1618 * buffer away, the caller must set B_INVAL prior to calling brelse().
1621 * We have insufficient buffer headers
1622 * We have insufficient buffer space
1623 * buffer_map is too fragmented ( space reservation fails )
1624 * If we have to flush dirty buffers ( but we try to avoid this )
1626 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1627 * Instead we ask the buf daemon to do it for us. We attempt to
1628 * avoid piecemeal wakeups of the pageout daemon.
1632 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1638 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1639 static int flushingbufs;
1642 * We can't afford to block since we might be holding a vnode lock,
1643 * which may prevent system daemons from running. We deal with
1644 * low-memory situations by proactively returning memory and running
1645 * async I/O rather then sync I/O.
1649 --getnewbufrestarts;
1651 ++getnewbufrestarts;
1654 * Setup for scan. If we do not have enough free buffers,
1655 * we setup a degenerate case that immediately fails. Note
1656 * that if we are specially marked process, we are allowed to
1657 * dip into our reserves.
1659 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1661 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1662 * However, there are a number of cases (defragging, reusing, ...)
1663 * where we cannot backup.
1665 nqindex = BQUEUE_EMPTYKVA;
1666 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1670 * If no EMPTYKVA buffers and we are either
1671 * defragging or reusing, locate a CLEAN buffer
1672 * to free or reuse. If bufspace useage is low
1673 * skip this step so we can allocate a new buffer.
1675 if (defrag || bufspace >= lobufspace) {
1676 nqindex = BQUEUE_CLEAN;
1677 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1681 * If we could not find or were not allowed to reuse a
1682 * CLEAN buffer, check to see if it is ok to use an EMPTY
1683 * buffer. We can only use an EMPTY buffer if allocating
1684 * its KVA would not otherwise run us out of buffer space.
1686 if (nbp == NULL && defrag == 0 &&
1687 bufspace + maxsize < hibufspace) {
1688 nqindex = BQUEUE_EMPTY;
1689 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1694 * Run scan, possibly freeing data and/or kva mappings on the fly
1698 while ((bp = nbp) != NULL) {
1699 int qindex = nqindex;
1702 * Calculate next bp ( we can only use it if we do not block
1703 * or do other fancy things ).
1705 if ((nbp = TAILQ_NEXT(bp, b_freelist)) == NULL) {
1708 nqindex = BQUEUE_EMPTYKVA;
1709 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1712 case BQUEUE_EMPTYKVA:
1713 nqindex = BQUEUE_CLEAN;
1714 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1728 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1731 * Note: we no longer distinguish between VMIO and non-VMIO
1735 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1738 * If we are defragging then we need a buffer with
1739 * b_kvasize != 0. XXX this situation should no longer
1740 * occur, if defrag is non-zero the buffer's b_kvasize
1741 * should also be non-zero at this point. XXX
1743 if (defrag && bp->b_kvasize == 0) {
1744 kprintf("Warning: defrag empty buffer %p\n", bp);
1749 * Start freeing the bp. This is somewhat involved. nbp
1750 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1751 * on the clean list must be disassociated from their
1752 * current vnode. Buffers on the empty[kva] lists have
1753 * already been disassociated.
1756 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1757 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1758 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1761 if (bp->b_qindex != qindex) {
1762 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1769 * Dependancies must be handled before we disassociate the
1772 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1773 * be immediately disassociated. HAMMER then becomes
1774 * responsible for releasing the buffer.
1776 if (LIST_FIRST(&bp->b_dep) != NULL) {
1778 if (bp->b_flags & B_LOCKED) {
1782 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1785 if (qindex == BQUEUE_CLEAN) {
1786 if (bp->b_flags & B_VMIO) {
1787 bp->b_flags &= ~B_ASYNC;
1788 vfs_vmio_release(bp);
1795 * NOTE: nbp is now entirely invalid. We can only restart
1796 * the scan from this point on.
1798 * Get the rest of the buffer freed up. b_kva* is still
1799 * valid after this operation.
1802 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1803 KKASSERT((bp->b_flags & B_HASHED) == 0);
1806 * critical section protection is not required when
1807 * scrapping a buffer's contents because it is already
1813 bp->b_flags = B_BNOCLIP;
1814 bp->b_cmd = BUF_CMD_DONE;
1819 bp->b_xio.xio_npages = 0;
1820 bp->b_dirtyoff = bp->b_dirtyend = 0;
1823 if (blkflags & GETBLK_BHEAVY)
1824 bp->b_flags |= B_HEAVY;
1827 * If we are defragging then free the buffer.
1830 bp->b_flags |= B_INVAL;
1838 * If we are overcomitted then recover the buffer and its
1839 * KVM space. This occurs in rare situations when multiple
1840 * processes are blocked in getnewbuf() or allocbuf().
1842 if (bufspace >= hibufspace)
1844 if (flushingbufs && bp->b_kvasize != 0) {
1845 bp->b_flags |= B_INVAL;
1850 if (bufspace < lobufspace)
1856 * If we exhausted our list, sleep as appropriate. We may have to
1857 * wakeup various daemons and write out some dirty buffers.
1859 * Generally we are sleeping due to insufficient buffer space.
1867 flags = VFS_BIO_NEED_BUFSPACE;
1869 } else if (bufspace >= hibufspace) {
1871 flags = VFS_BIO_NEED_BUFSPACE;
1874 flags = VFS_BIO_NEED_ANY;
1877 needsbuffer |= flags;
1878 bd_speedup(); /* heeeelp */
1879 while (needsbuffer & flags) {
1880 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
1885 * We finally have a valid bp. We aren't quite out of the
1886 * woods, we still have to reserve kva space. In order
1887 * to keep fragmentation sane we only allocate kva in
1890 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1892 if (maxsize != bp->b_kvasize) {
1893 vm_offset_t addr = 0;
1898 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1899 vm_map_lock(&buffer_map);
1901 if (vm_map_findspace(&buffer_map,
1902 vm_map_min(&buffer_map), maxsize,
1905 * Uh oh. Buffer map is too fragmented. We
1906 * must defragment the map.
1908 vm_map_unlock(&buffer_map);
1909 vm_map_entry_release(count);
1912 bp->b_flags |= B_INVAL;
1917 vm_map_insert(&buffer_map, &count,
1919 addr, addr + maxsize,
1921 VM_PROT_ALL, VM_PROT_ALL,
1924 bp->b_kvabase = (caddr_t) addr;
1925 bp->b_kvasize = maxsize;
1926 bufspace += bp->b_kvasize;
1929 vm_map_unlock(&buffer_map);
1930 vm_map_entry_release(count);
1932 bp->b_data = bp->b_kvabase;
1940 * Buffer flushing daemon. Buffers are normally flushed by the
1941 * update daemon but if it cannot keep up this process starts to
1942 * take the load in an attempt to prevent getnewbuf() from blocking.
1944 * Once a flush is initiated it does not stop until the number
1945 * of buffers falls below lodirtybuffers, but we will wake up anyone
1946 * waiting at the mid-point.
1949 static struct thread *bufdaemon_td;
1950 static struct thread *bufdaemonhw_td;
1952 static struct kproc_desc buf_kp = {
1957 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
1958 kproc_start, &buf_kp)
1960 static struct kproc_desc bufhw_kp = {
1965 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
1966 kproc_start, &bufhw_kp)
1972 * This process needs to be suspended prior to shutdown sync.
1974 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
1975 bufdaemon_td, SHUTDOWN_PRI_LAST);
1978 * This process is allowed to take the buffer cache to the limit
1983 kproc_suspend_loop();
1986 * Do the flush. Limit the amount of in-transit I/O we
1987 * allow to build up, otherwise we would completely saturate
1988 * the I/O system. Wakeup any waiting processes before we
1989 * normally would so they can run in parallel with our drain.
1991 while (numdirtybuffers > lodirtybuffers) {
1992 if (flushbufqueues(BQUEUE_DIRTY) == 0)
1994 waitrunningbufspace();
1997 if (runningbufcount + numdirtybuffers > lodirtybuffers) {
1998 waitrunningbufspace();
2003 * Only clear bd_request if we have reached our low water
2004 * mark. The buf_daemon normally waits 5 seconds and
2005 * then incrementally flushes any dirty buffers that have
2006 * built up, within reason.
2008 * If we were unable to hit our low water mark and couldn't
2009 * find any flushable buffers, we sleep half a second.
2010 * Otherwise we loop immediately.
2012 if (runningbufcount + numdirtybuffers <= lodirtybuffers) {
2014 * We reached our low water mark, reset the
2015 * request and sleep until we are needed again.
2016 * The sleep is just so the suspend code works.
2018 spin_lock_wr(&needsbuffer_spin);
2020 msleep(&bd_request, &needsbuffer_spin, 0,
2022 spin_unlock_wr(&needsbuffer_spin);
2025 * We couldn't find any flushable dirty buffers but
2026 * still have too many dirty buffers, we
2027 * have to sleep and try again. (rare)
2029 tsleep(&bd_request, 0, "qsleep", hz / 2);
2038 * This process needs to be suspended prior to shutdown sync.
2040 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2041 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2044 * This process is allowed to take the buffer cache to the limit
2049 kproc_suspend_loop();
2052 * Do the flush. Limit the amount of in-transit I/O we
2053 * allow to build up, otherwise we would completely saturate
2054 * the I/O system. Wakeup any waiting processes before we
2055 * normally would so they can run in parallel with our drain.
2057 while (numdirtybuffershw > lodirtybuffers) {
2058 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2060 waitrunningbufspace();
2063 if (runningbufcount + numdirtybuffershw > lodirtybuffers) {
2064 waitrunningbufspace();
2068 * Only clear bd_request if we have reached our low water
2069 * mark. The buf_daemon normally waits 5 seconds and
2070 * then incrementally flushes any dirty buffers that have
2071 * built up, within reason.
2073 * If we were unable to hit our low water mark and couldn't
2074 * find any flushable buffers, we sleep half a second.
2075 * Otherwise we loop immediately.
2077 if (runningbufcount + numdirtybuffershw <= lodirtybuffers) {
2079 * We reached our low water mark, reset the
2080 * request and sleep until we are needed again.
2081 * The sleep is just so the suspend code works.
2083 spin_lock_wr(&needsbuffer_spin);
2085 msleep(&bd_request_hw, &needsbuffer_spin, 0,
2087 spin_unlock_wr(&needsbuffer_spin);
2090 * We couldn't find any flushable dirty buffers but
2091 * still have too many dirty buffers, we
2092 * have to sleep and try again. (rare)
2094 tsleep(&bd_request_hw, 0, "qsleep", hz / 2);
2102 * Try to flush a buffer in the dirty queue. We must be careful to
2103 * free up B_INVAL buffers instead of write them, which NFS is
2104 * particularly sensitive to.
2108 flushbufqueues(bufq_type_t q)
2113 bp = TAILQ_FIRST(&bufqueues[q]);
2116 KASSERT((bp->b_flags & B_DELWRI),
2117 ("unexpected clean buffer %p", bp));
2118 if (bp->b_flags & B_DELWRI) {
2119 if (bp->b_flags & B_INVAL) {
2120 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2121 panic("flushbufqueues: locked buf");
2127 if (LIST_FIRST(&bp->b_dep) != NULL &&
2128 (bp->b_flags & B_DEFERRED) == 0 &&
2129 buf_countdeps(bp, 0)) {
2130 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2131 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2133 bp->b_flags |= B_DEFERRED;
2134 bp = TAILQ_FIRST(&bufqueues[q]);
2139 * Only write it out if we can successfully lock
2140 * it. If the buffer has a dependancy,
2141 * buf_checkwrite must also return 0.
2143 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
2144 if (LIST_FIRST(&bp->b_dep) != NULL &&
2145 buf_checkwrite(bp)) {
2155 bp = TAILQ_NEXT(bp, b_freelist);
2163 * Returns true if no I/O is needed to access the associated VM object.
2164 * This is like findblk except it also hunts around in the VM system for
2167 * Note that we ignore vm_page_free() races from interrupts against our
2168 * lookup, since if the caller is not protected our return value will not
2169 * be any more valid then otherwise once we exit the critical section.
2172 inmem(struct vnode *vp, off_t loffset)
2175 vm_offset_t toff, tinc, size;
2178 if (findblk(vp, loffset))
2180 if (vp->v_mount == NULL)
2182 if ((obj = vp->v_object) == NULL)
2186 if (size > vp->v_mount->mnt_stat.f_iosize)
2187 size = vp->v_mount->mnt_stat.f_iosize;
2189 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2190 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2194 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2195 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2196 if (vm_page_is_valid(m,
2197 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2206 * Sets the dirty range for a buffer based on the status of the dirty
2207 * bits in the pages comprising the buffer.
2209 * The range is limited to the size of the buffer.
2211 * This routine is primarily used by NFS, but is generalized for the
2215 vfs_setdirty(struct buf *bp)
2221 * Degenerate case - empty buffer
2224 if (bp->b_bufsize == 0)
2228 * We qualify the scan for modified pages on whether the
2229 * object has been flushed yet. The OBJ_WRITEABLE flag
2230 * is not cleared simply by protecting pages off.
2233 if ((bp->b_flags & B_VMIO) == 0)
2236 object = bp->b_xio.xio_pages[0]->object;
2238 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2239 kprintf("Warning: object %p writeable but not mightbedirty\n", object);
2240 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2241 kprintf("Warning: object %p mightbedirty but not writeable\n", object);
2243 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2244 vm_offset_t boffset;
2245 vm_offset_t eoffset;
2248 * test the pages to see if they have been modified directly
2249 * by users through the VM system.
2251 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2252 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2253 vm_page_test_dirty(bp->b_xio.xio_pages[i]);
2257 * Calculate the encompassing dirty range, boffset and eoffset,
2258 * (eoffset - boffset) bytes.
2261 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2262 if (bp->b_xio.xio_pages[i]->dirty)
2265 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2267 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2268 if (bp->b_xio.xio_pages[i]->dirty) {
2272 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2275 * Fit it to the buffer.
2278 if (eoffset > bp->b_bcount)
2279 eoffset = bp->b_bcount;
2282 * If we have a good dirty range, merge with the existing
2286 if (boffset < eoffset) {
2287 if (bp->b_dirtyoff > boffset)
2288 bp->b_dirtyoff = boffset;
2289 if (bp->b_dirtyend < eoffset)
2290 bp->b_dirtyend = eoffset;
2298 * Locate and return the specified buffer, or NULL if the buffer does
2299 * not exist. Do not attempt to lock the buffer or manipulate it in
2300 * any way. The caller must validate that the correct buffer has been
2301 * obtain after locking it.
2304 findblk(struct vnode *vp, off_t loffset)
2309 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2317 * Get a block given a specified block and offset into a file/device.
2318 * B_INVAL may or may not be set on return. The caller should clear
2319 * B_INVAL prior to initiating a READ.
2321 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2322 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2323 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2324 * without doing any of those things the system will likely believe
2325 * the buffer to be valid (especially if it is not B_VMIO), and the
2326 * next getblk() will return the buffer with B_CACHE set.
2328 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2329 * an existing buffer.
2331 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2332 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2333 * and then cleared based on the backing VM. If the previous buffer is
2334 * non-0-sized but invalid, B_CACHE will be cleared.
2336 * If getblk() must create a new buffer, the new buffer is returned with
2337 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2338 * case it is returned with B_INVAL clear and B_CACHE set based on the
2341 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2342 * B_CACHE bit is clear.
2344 * What this means, basically, is that the caller should use B_CACHE to
2345 * determine whether the buffer is fully valid or not and should clear
2346 * B_INVAL prior to issuing a read. If the caller intends to validate
2347 * the buffer by loading its data area with something, the caller needs
2348 * to clear B_INVAL. If the caller does this without issuing an I/O,
2349 * the caller should set B_CACHE ( as an optimization ), else the caller
2350 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2351 * a write attempt or if it was a successfull read. If the caller
2352 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2353 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2357 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2358 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2361 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2364 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2366 if (size > MAXBSIZE)
2367 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2368 if (vp->v_object == NULL)
2369 panic("getblk: vnode %p has no object!", vp);
2373 if ((bp = findblk(vp, loffset))) {
2375 * The buffer was found in the cache, but we need to lock it.
2376 * Even with LK_NOWAIT the lockmgr may break our critical
2377 * section, so double-check the validity of the buffer
2378 * once the lock has been obtained.
2380 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2381 int lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2382 if (blkflags & GETBLK_PCATCH)
2383 lkflags |= LK_PCATCH;
2384 if (BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo) ==
2393 * Once the buffer has been locked, make sure we didn't race
2394 * a buffer recyclement. Buffers that are no longer hashed
2395 * will have b_vp == NULL, so this takes care of that check
2398 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2399 kprintf("Warning buffer %p (vp %p loffset %lld) was recycled\n", bp, vp, loffset);
2405 * All vnode-based buffers must be backed by a VM object.
2407 KKASSERT(bp->b_flags & B_VMIO);
2408 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2411 * Make sure that B_INVAL buffers do not have a cached
2412 * block number translation.
2414 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2415 kprintf("Warning invalid buffer %p (vp %p loffset %lld) did not have cleared bio_offset cache\n", bp, vp, loffset);
2416 clearbiocache(&bp->b_bio2);
2420 * The buffer is locked. B_CACHE is cleared if the buffer is
2423 if (bp->b_flags & B_INVAL)
2424 bp->b_flags &= ~B_CACHE;
2428 * Any size inconsistancy with a dirty buffer or a buffer
2429 * with a softupdates dependancy must be resolved. Resizing
2430 * the buffer in such circumstances can lead to problems.
2432 if (size != bp->b_bcount) {
2433 if (bp->b_flags & B_DELWRI) {
2434 bp->b_flags |= B_NOCACHE;
2436 } else if (LIST_FIRST(&bp->b_dep)) {
2437 bp->b_flags |= B_NOCACHE;
2440 bp->b_flags |= B_RELBUF;
2445 KKASSERT(size <= bp->b_kvasize);
2446 KASSERT(bp->b_loffset != NOOFFSET,
2447 ("getblk: no buffer offset"));
2450 * A buffer with B_DELWRI set and B_CACHE clear must
2451 * be committed before we can return the buffer in
2452 * order to prevent the caller from issuing a read
2453 * ( due to B_CACHE not being set ) and overwriting
2456 * Most callers, including NFS and FFS, need this to
2457 * operate properly either because they assume they
2458 * can issue a read if B_CACHE is not set, or because
2459 * ( for example ) an uncached B_DELWRI might loop due
2460 * to softupdates re-dirtying the buffer. In the latter
2461 * case, B_CACHE is set after the first write completes,
2462 * preventing further loops.
2464 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2465 * above while extending the buffer, we cannot allow the
2466 * buffer to remain with B_CACHE set after the write
2467 * completes or it will represent a corrupt state. To
2468 * deal with this we set B_NOCACHE to scrap the buffer
2471 * We might be able to do something fancy, like setting
2472 * B_CACHE in bwrite() except if B_DELWRI is already set,
2473 * so the below call doesn't set B_CACHE, but that gets real
2474 * confusing. This is much easier.
2477 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2478 bp->b_flags |= B_NOCACHE;
2485 * Buffer is not in-core, create new buffer. The buffer
2486 * returned by getnewbuf() is locked. Note that the returned
2487 * buffer is also considered valid (not marked B_INVAL).
2489 * Calculating the offset for the I/O requires figuring out
2490 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2491 * the mount's f_iosize otherwise. If the vnode does not
2492 * have an associated mount we assume that the passed size is
2495 * Note that vn_isdisk() cannot be used here since it may
2496 * return a failure for numerous reasons. Note that the
2497 * buffer size may be larger then the block size (the caller
2498 * will use block numbers with the proper multiple). Beware
2499 * of using any v_* fields which are part of unions. In
2500 * particular, in DragonFly the mount point overloading
2501 * mechanism uses the namecache only and the underlying
2502 * directory vnode is not a special case.
2506 if (vp->v_type == VBLK || vp->v_type == VCHR)
2508 else if (vp->v_mount)
2509 bsize = vp->v_mount->mnt_stat.f_iosize;
2513 maxsize = size + (loffset & PAGE_MASK);
2514 maxsize = imax(maxsize, bsize);
2516 if ((bp = getnewbuf(blkflags, slptimeo, size, maxsize)) == NULL) {
2517 if (slpflags || slptimeo) {
2525 * This code is used to make sure that a buffer is not
2526 * created while the getnewbuf routine is blocked.
2527 * This can be a problem whether the vnode is locked or not.
2528 * If the buffer is created out from under us, we have to
2529 * throw away the one we just created. There is no window
2530 * race because we are safely running in a critical section
2531 * from the point of the duplicate buffer creation through
2532 * to here, and we've locked the buffer.
2534 if (findblk(vp, loffset)) {
2535 bp->b_flags |= B_INVAL;
2541 * Insert the buffer into the hash, so that it can
2542 * be found by findblk().
2544 * Make sure the translation layer has been cleared.
2546 bp->b_loffset = loffset;
2547 bp->b_bio2.bio_offset = NOOFFSET;
2548 /* bp->b_bio2.bio_next = NULL; */
2553 * All vnode-based buffers must be backed by a VM object.
2555 KKASSERT(vp->v_object != NULL);
2556 bp->b_flags |= B_VMIO;
2557 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2569 * Reacquire a buffer that was previously released to the locked queue,
2570 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2571 * set B_LOCKED (which handles the acquisition race).
2573 * To this end, either B_LOCKED must be set or the dependancy list must be
2577 regetblk(struct buf *bp)
2579 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
2580 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2589 * Get an empty, disassociated buffer of given size. The buffer is
2590 * initially set to B_INVAL.
2592 * critical section protection is not required for the allocbuf()
2593 * call because races are impossible here.
2601 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2604 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2608 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2616 * This code constitutes the buffer memory from either anonymous system
2617 * memory (in the case of non-VMIO operations) or from an associated
2618 * VM object (in the case of VMIO operations). This code is able to
2619 * resize a buffer up or down.
2621 * Note that this code is tricky, and has many complications to resolve
2622 * deadlock or inconsistant data situations. Tread lightly!!!
2623 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2624 * the caller. Calling this code willy nilly can result in the loss of data.
2626 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2627 * B_CACHE for the non-VMIO case.
2629 * This routine does not need to be called from a critical section but you
2630 * must own the buffer.
2633 allocbuf(struct buf *bp, int size)
2635 int newbsize, mbsize;
2638 if (BUF_REFCNT(bp) == 0)
2639 panic("allocbuf: buffer not busy");
2641 if (bp->b_kvasize < size)
2642 panic("allocbuf: buffer too small");
2644 if ((bp->b_flags & B_VMIO) == 0) {
2648 * Just get anonymous memory from the kernel. Don't
2649 * mess with B_CACHE.
2651 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2652 if (bp->b_flags & B_MALLOC)
2655 newbsize = round_page(size);
2657 if (newbsize < bp->b_bufsize) {
2659 * Malloced buffers are not shrunk
2661 if (bp->b_flags & B_MALLOC) {
2663 bp->b_bcount = size;
2665 kfree(bp->b_data, M_BIOBUF);
2666 if (bp->b_bufsize) {
2667 bufmallocspace -= bp->b_bufsize;
2671 bp->b_data = bp->b_kvabase;
2673 bp->b_flags &= ~B_MALLOC;
2679 (vm_offset_t) bp->b_data + newbsize,
2680 (vm_offset_t) bp->b_data + bp->b_bufsize);
2681 } else if (newbsize > bp->b_bufsize) {
2683 * We only use malloced memory on the first allocation.
2684 * and revert to page-allocated memory when the buffer
2687 if ((bufmallocspace < maxbufmallocspace) &&
2688 (bp->b_bufsize == 0) &&
2689 (mbsize <= PAGE_SIZE/2)) {
2691 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
2692 bp->b_bufsize = mbsize;
2693 bp->b_bcount = size;
2694 bp->b_flags |= B_MALLOC;
2695 bufmallocspace += mbsize;
2701 * If the buffer is growing on its other-than-first
2702 * allocation, then we revert to the page-allocation
2705 if (bp->b_flags & B_MALLOC) {
2706 origbuf = bp->b_data;
2707 origbufsize = bp->b_bufsize;
2708 bp->b_data = bp->b_kvabase;
2709 if (bp->b_bufsize) {
2710 bufmallocspace -= bp->b_bufsize;
2714 bp->b_flags &= ~B_MALLOC;
2715 newbsize = round_page(newbsize);
2719 (vm_offset_t) bp->b_data + bp->b_bufsize,
2720 (vm_offset_t) bp->b_data + newbsize);
2722 bcopy(origbuf, bp->b_data, origbufsize);
2723 kfree(origbuf, M_BIOBUF);
2730 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2731 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
2732 newbsize + PAGE_MASK) >> PAGE_SHIFT;
2733 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
2735 if (bp->b_flags & B_MALLOC)
2736 panic("allocbuf: VMIO buffer can't be malloced");
2738 * Set B_CACHE initially if buffer is 0 length or will become
2741 if (size == 0 || bp->b_bufsize == 0)
2742 bp->b_flags |= B_CACHE;
2744 if (newbsize < bp->b_bufsize) {
2746 * DEV_BSIZE aligned new buffer size is less then the
2747 * DEV_BSIZE aligned existing buffer size. Figure out
2748 * if we have to remove any pages.
2750 if (desiredpages < bp->b_xio.xio_npages) {
2751 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
2753 * the page is not freed here -- it
2754 * is the responsibility of
2755 * vnode_pager_setsize
2757 m = bp->b_xio.xio_pages[i];
2758 KASSERT(m != bogus_page,
2759 ("allocbuf: bogus page found"));
2760 while (vm_page_sleep_busy(m, TRUE, "biodep"))
2763 bp->b_xio.xio_pages[i] = NULL;
2764 vm_page_unwire(m, 0);
2766 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2767 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
2768 bp->b_xio.xio_npages = desiredpages;
2770 } else if (size > bp->b_bcount) {
2772 * We are growing the buffer, possibly in a
2773 * byte-granular fashion.
2781 * Step 1, bring in the VM pages from the object,
2782 * allocating them if necessary. We must clear
2783 * B_CACHE if these pages are not valid for the
2784 * range covered by the buffer.
2786 * critical section protection is required to protect
2787 * against interrupts unbusying and freeing pages
2788 * between our vm_page_lookup() and our
2789 * busycheck/wiring call.
2795 while (bp->b_xio.xio_npages < desiredpages) {
2799 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
2800 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2802 * note: must allocate system pages
2803 * since blocking here could intefere
2804 * with paging I/O, no matter which
2807 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
2810 vm_pageout_deficit += desiredpages -
2811 bp->b_xio.xio_npages;
2815 bp->b_flags &= ~B_CACHE;
2816 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2817 ++bp->b_xio.xio_npages;
2823 * We found a page. If we have to sleep on it,
2824 * retry because it might have gotten freed out
2827 * We can only test PG_BUSY here. Blocking on
2828 * m->busy might lead to a deadlock:
2830 * vm_fault->getpages->cluster_read->allocbuf
2834 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
2838 * We have a good page. Should we wakeup the
2841 if ((curthread != pagethread) &&
2842 ((m->queue - m->pc) == PQ_CACHE) &&
2843 ((vmstats.v_free_count + vmstats.v_cache_count) <
2844 (vmstats.v_free_min + vmstats.v_cache_min))) {
2845 pagedaemon_wakeup();
2847 vm_page_flag_clear(m, PG_ZERO);
2849 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2850 ++bp->b_xio.xio_npages;
2855 * Step 2. We've loaded the pages into the buffer,
2856 * we have to figure out if we can still have B_CACHE
2857 * set. Note that B_CACHE is set according to the
2858 * byte-granular range ( bcount and size ), not the
2859 * aligned range ( newbsize ).
2861 * The VM test is against m->valid, which is DEV_BSIZE
2862 * aligned. Needless to say, the validity of the data
2863 * needs to also be DEV_BSIZE aligned. Note that this
2864 * fails with NFS if the server or some other client
2865 * extends the file's EOF. If our buffer is resized,
2866 * B_CACHE may remain set! XXX
2869 toff = bp->b_bcount;
2870 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
2872 while ((bp->b_flags & B_CACHE) && toff < size) {
2875 if (tinc > (size - toff))
2878 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
2886 bp->b_xio.xio_pages[pi]
2893 * Step 3, fixup the KVM pmap. Remember that
2894 * bp->b_data is relative to bp->b_loffset, but
2895 * bp->b_loffset may be offset into the first page.
2898 bp->b_data = (caddr_t)
2899 trunc_page((vm_offset_t)bp->b_data);
2901 (vm_offset_t)bp->b_data,
2902 bp->b_xio.xio_pages,
2903 bp->b_xio.xio_npages
2905 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2906 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
2909 if (newbsize < bp->b_bufsize)
2911 bp->b_bufsize = newbsize; /* actual buffer allocation */
2912 bp->b_bcount = size; /* requested buffer size */
2919 * Wait for buffer I/O completion, returning error status. The buffer
2920 * is left locked on return. B_EINTR is converted into an EINTR error
2923 * NOTE! The original b_cmd is lost on return, since b_cmd will be
2924 * set to BUF_CMD_DONE.
2927 biowait(struct buf *bp)
2930 while (bp->b_cmd != BUF_CMD_DONE) {
2931 if (bp->b_cmd == BUF_CMD_READ)
2932 tsleep(bp, 0, "biord", 0);
2934 tsleep(bp, 0, "biowr", 0);
2937 if (bp->b_flags & B_EINTR) {
2938 bp->b_flags &= ~B_EINTR;
2941 if (bp->b_flags & B_ERROR) {
2942 return (bp->b_error ? bp->b_error : EIO);
2949 * This associates a tracking count with an I/O. vn_strategy() and
2950 * dev_dstrategy() do this automatically but there are a few cases
2951 * where a vnode or device layer is bypassed when a block translation
2952 * is cached. In such cases bio_start_transaction() may be called on
2953 * the bypassed layers so the system gets an I/O in progress indication
2954 * for those higher layers.
2957 bio_start_transaction(struct bio *bio, struct bio_track *track)
2959 bio->bio_track = track;
2960 atomic_add_int(&track->bk_active, 1);
2964 * Initiate I/O on a vnode.
2967 vn_strategy(struct vnode *vp, struct bio *bio)
2969 struct bio_track *track;
2971 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
2972 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
2973 track = &vp->v_track_read;
2975 track = &vp->v_track_write;
2976 bio->bio_track = track;
2977 atomic_add_int(&track->bk_active, 1);
2978 vop_strategy(*vp->v_ops, vp, bio);
2985 * Finish I/O on a buffer, optionally calling a completion function.
2986 * This is usually called from an interrupt so process blocking is
2989 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
2990 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
2991 * assuming B_INVAL is clear.
2993 * For the VMIO case, we set B_CACHE if the op was a read and no
2994 * read error occured, or if the op was a write. B_CACHE is never
2995 * set if the buffer is invalid or otherwise uncacheable.
2997 * biodone does not mess with B_INVAL, allowing the I/O routine or the
2998 * initiator to leave B_INVAL set to brelse the buffer out of existance
2999 * in the biodone routine.
3002 biodone(struct bio *bio)
3004 struct buf *bp = bio->bio_buf;
3009 KASSERT(BUF_REFCNTNB(bp) > 0,
3010 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3011 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3012 ("biodone: bp %p already done!", bp));
3014 runningbufwakeup(bp);
3017 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3020 biodone_t *done_func;
3021 struct bio_track *track;
3024 * BIO tracking. Most but not all BIOs are tracked.
3026 if ((track = bio->bio_track) != NULL) {
3027 atomic_subtract_int(&track->bk_active, 1);
3028 if (track->bk_active < 0) {
3029 panic("biodone: bad active count bio %p\n",
3032 if (track->bk_waitflag) {
3033 track->bk_waitflag = 0;
3036 bio->bio_track = NULL;
3040 * A bio_done function terminates the loop. The function
3041 * will be responsible for any further chaining and/or
3042 * buffer management.
3044 * WARNING! The done function can deallocate the buffer!
3046 if ((done_func = bio->bio_done) != NULL) {
3047 bio->bio_done = NULL;
3052 bio = bio->bio_prev;
3056 bp->b_cmd = BUF_CMD_DONE;
3059 * Only reads and writes are processed past this point.
3061 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3068 * Warning: softupdates may re-dirty the buffer.
3070 if (LIST_FIRST(&bp->b_dep) != NULL)
3073 if (bp->b_flags & B_VMIO) {
3079 struct vnode *vp = bp->b_vp;
3083 #if defined(VFS_BIO_DEBUG)
3084 if (vp->v_auxrefs == 0)
3085 panic("biodone: zero vnode hold count");
3086 if ((vp->v_flag & VOBJBUF) == 0)
3087 panic("biodone: vnode is not setup for merged cache");
3090 foff = bp->b_loffset;
3091 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3092 KASSERT(obj != NULL, ("biodone: missing VM object"));
3094 #if defined(VFS_BIO_DEBUG)
3095 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3096 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3097 obj->paging_in_progress, bp->b_xio.xio_npages);
3102 * Set B_CACHE if the op was a normal read and no error
3103 * occured. B_CACHE is set for writes in the b*write()
3106 iosize = bp->b_bcount - bp->b_resid;
3107 if (cmd == BUF_CMD_READ && (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3108 bp->b_flags |= B_CACHE;
3111 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3115 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3120 * cleanup bogus pages, restoring the originals. Since
3121 * the originals should still be wired, we don't have
3122 * to worry about interrupt/freeing races destroying
3123 * the VM object association.
3125 m = bp->b_xio.xio_pages[i];
3126 if (m == bogus_page) {
3128 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3130 panic("biodone: page disappeared");
3131 bp->b_xio.xio_pages[i] = m;
3132 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3133 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3135 #if defined(VFS_BIO_DEBUG)
3136 if (OFF_TO_IDX(foff) != m->pindex) {
3138 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
3139 (unsigned long)foff, m->pindex);
3144 * In the write case, the valid and clean bits are
3145 * already changed correctly ( see bdwrite() ), so we
3146 * only need to do this here in the read case.
3148 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3149 vfs_page_set_valid(bp, foff, i, m);
3151 vm_page_flag_clear(m, PG_ZERO);
3154 * when debugging new filesystems or buffer I/O methods, this
3155 * is the most common error that pops up. if you see this, you
3156 * have not set the page busy flag correctly!!!
3159 kprintf("biodone: page busy < 0, "
3160 "pindex: %d, foff: 0x(%x,%x), "
3161 "resid: %d, index: %d\n",
3162 (int) m->pindex, (int)(foff >> 32),
3163 (int) foff & 0xffffffff, resid, i);
3164 if (!vn_isdisk(vp, NULL))
3165 kprintf(" iosize: %ld, loffset: %lld, flags: 0x%08x, npages: %d\n",
3166 bp->b_vp->v_mount->mnt_stat.f_iosize,
3168 bp->b_flags, bp->b_xio.xio_npages);
3170 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3172 bp->b_flags, bp->b_xio.xio_npages);
3173 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3174 m->valid, m->dirty, m->wire_count);
3175 panic("biodone: page busy < 0");
3177 vm_page_io_finish(m);
3178 vm_object_pip_subtract(obj, 1);
3179 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3183 vm_object_pip_wakeupn(obj, 0);
3187 * For asynchronous completions, release the buffer now. The brelse
3188 * will do a wakeup there if necessary - so no need to do a wakeup
3189 * here in the async case. The sync case always needs to do a wakeup.
3192 if (bp->b_flags & B_ASYNC) {
3193 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
3206 * This routine is called in lieu of iodone in the case of
3207 * incomplete I/O. This keeps the busy status for pages
3211 vfs_unbusy_pages(struct buf *bp)
3215 runningbufwakeup(bp);
3216 if (bp->b_flags & B_VMIO) {
3217 struct vnode *vp = bp->b_vp;
3222 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3223 vm_page_t m = bp->b_xio.xio_pages[i];
3226 * When restoring bogus changes the original pages
3227 * should still be wired, so we are in no danger of
3228 * losing the object association and do not need
3229 * critical section protection particularly.
3231 if (m == bogus_page) {
3232 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3234 panic("vfs_unbusy_pages: page missing");
3236 bp->b_xio.xio_pages[i] = m;
3237 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3238 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3240 vm_object_pip_subtract(obj, 1);
3241 vm_page_flag_clear(m, PG_ZERO);
3242 vm_page_io_finish(m);
3244 vm_object_pip_wakeupn(obj, 0);
3249 * vfs_page_set_valid:
3251 * Set the valid bits in a page based on the supplied offset. The
3252 * range is restricted to the buffer's size.
3254 * This routine is typically called after a read completes.
3257 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3259 vm_ooffset_t soff, eoff;
3262 * Start and end offsets in buffer. eoff - soff may not cross a
3263 * page boundry or cross the end of the buffer. The end of the
3264 * buffer, in this case, is our file EOF, not the allocation size
3268 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3269 if (eoff > bp->b_loffset + bp->b_bcount)
3270 eoff = bp->b_loffset + bp->b_bcount;
3273 * Set valid range. This is typically the entire buffer and thus the
3277 vm_page_set_validclean(
3279 (vm_offset_t) (soff & PAGE_MASK),
3280 (vm_offset_t) (eoff - soff)
3288 * This routine is called before a device strategy routine.
3289 * It is used to tell the VM system that paging I/O is in
3290 * progress, and treat the pages associated with the buffer
3291 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3292 * flag is handled to make sure that the object doesn't become
3295 * Since I/O has not been initiated yet, certain buffer flags
3296 * such as B_ERROR or B_INVAL may be in an inconsistant state
3297 * and should be ignored.
3300 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3303 struct lwp *lp = curthread->td_lwp;
3306 * The buffer's I/O command must already be set. If reading,
3307 * B_CACHE must be 0 (double check against callers only doing
3308 * I/O when B_CACHE is 0).
3310 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3311 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3313 if (bp->b_flags & B_VMIO) {
3318 foff = bp->b_loffset;
3319 KASSERT(bp->b_loffset != NOOFFSET,
3320 ("vfs_busy_pages: no buffer offset"));
3324 * Loop until none of the pages are busy.
3327 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3328 vm_page_t m = bp->b_xio.xio_pages[i];
3330 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3335 * Setup for I/O, soft-busy the page right now because
3336 * the next loop may block.
3338 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3339 vm_page_t m = bp->b_xio.xio_pages[i];
3341 vm_page_flag_clear(m, PG_ZERO);
3342 if ((bp->b_flags & B_CLUSTER) == 0) {
3343 vm_object_pip_add(obj, 1);
3344 vm_page_io_start(m);
3349 * Adjust protections for I/O and do bogus-page mapping.
3350 * Assume that vm_page_protect() can block (it can block
3351 * if VM_PROT_NONE, don't take any chances regardless).
3354 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3355 vm_page_t m = bp->b_xio.xio_pages[i];
3358 * When readying a vnode-backed buffer for a write
3359 * we must zero-fill any invalid portions of the
3362 * When readying a vnode-backed buffer for a read
3363 * we must replace any dirty pages with a bogus
3364 * page so we do not destroy dirty data when
3365 * filling in gaps. Dirty pages might not
3366 * necessarily be marked dirty yet, so use m->valid
3367 * as a reasonable test.
3369 * Bogus page replacement is, uh, bogus. We need
3370 * to find a better way.
3372 if (bp->b_cmd == BUF_CMD_WRITE) {
3373 vm_page_protect(m, VM_PROT_READ);
3374 vfs_page_set_valid(bp, foff, i, m);
3375 } else if (m->valid == VM_PAGE_BITS_ALL) {
3376 bp->b_xio.xio_pages[i] = bogus_page;
3379 vm_page_protect(m, VM_PROT_NONE);
3381 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3384 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3385 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3389 * This is the easiest place to put the process accounting for the I/O
3393 if (bp->b_cmd == BUF_CMD_READ)
3394 lp->lwp_ru.ru_inblock++;
3396 lp->lwp_ru.ru_oublock++;
3403 * Tell the VM system that the pages associated with this buffer
3404 * are clean. This is used for delayed writes where the data is
3405 * going to go to disk eventually without additional VM intevention.
3407 * Note that while we only really need to clean through to b_bcount, we
3408 * just go ahead and clean through to b_bufsize.
3411 vfs_clean_pages(struct buf *bp)
3415 if (bp->b_flags & B_VMIO) {
3418 foff = bp->b_loffset;
3419 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset"));
3420 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3421 vm_page_t m = bp->b_xio.xio_pages[i];
3422 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3423 vm_ooffset_t eoff = noff;
3425 if (eoff > bp->b_loffset + bp->b_bufsize)
3426 eoff = bp->b_loffset + bp->b_bufsize;
3427 vfs_page_set_valid(bp, foff, i, m);
3428 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3435 * vfs_bio_set_validclean:
3437 * Set the range within the buffer to valid and clean. The range is
3438 * relative to the beginning of the buffer, b_loffset. Note that
3439 * b_loffset itself may be offset from the beginning of the first page.
3443 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3445 if (bp->b_flags & B_VMIO) {
3450 * Fixup base to be relative to beginning of first page.
3451 * Set initial n to be the maximum number of bytes in the
3452 * first page that can be validated.
3455 base += (bp->b_loffset & PAGE_MASK);
3456 n = PAGE_SIZE - (base & PAGE_MASK);
3458 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) {
3459 vm_page_t m = bp->b_xio.xio_pages[i];
3464 vm_page_set_validclean(m, base & PAGE_MASK, n);
3475 * Clear a buffer. This routine essentially fakes an I/O, so we need
3476 * to clear B_ERROR and B_INVAL.
3478 * Note that while we only theoretically need to clear through b_bcount,
3479 * we go ahead and clear through b_bufsize.
3483 vfs_bio_clrbuf(struct buf *bp)
3487 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3488 bp->b_flags &= ~(B_INVAL|B_ERROR);
3489 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3490 (bp->b_loffset & PAGE_MASK) == 0) {
3491 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3492 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3496 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3497 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3498 bzero(bp->b_data, bp->b_bufsize);
3499 bp->b_xio.xio_pages[0]->valid |= mask;
3504 ea = sa = bp->b_data;
3505 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3506 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3507 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3508 ea = (caddr_t)(vm_offset_t)ulmin(
3509 (u_long)(vm_offset_t)ea,
3510 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3511 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3512 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3514 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3515 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3519 for (; sa < ea; sa += DEV_BSIZE, j++) {
3520 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3521 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3522 bzero(sa, DEV_BSIZE);
3525 bp->b_xio.xio_pages[i]->valid |= mask;
3526 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
3535 * vm_hold_load_pages:
3537 * Load pages into the buffer's address space. The pages are
3538 * allocated from the kernel object in order to reduce interference
3539 * with the any VM paging I/O activity. The range of loaded
3540 * pages will be wired.
3542 * If a page cannot be allocated, the 'pagedaemon' is woken up to
3543 * retrieve the full range (to - from) of pages.
3547 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3553 to = round_page(to);
3554 from = round_page(from);
3555 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3557 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3562 * Note: must allocate system pages since blocking here
3563 * could intefere with paging I/O, no matter which
3566 p = vm_page_alloc(&kernel_object,
3568 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
3570 vm_pageout_deficit += (to - from) >> PAGE_SHIFT;
3575 p->valid = VM_PAGE_BITS_ALL;
3576 vm_page_flag_clear(p, PG_ZERO);
3577 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3578 bp->b_xio.xio_pages[index] = p;
3581 bp->b_xio.xio_npages = index;
3585 * vm_hold_free_pages:
3587 * Return pages associated with the buffer back to the VM system.
3589 * The range of pages underlying the buffer's address space will
3590 * be unmapped and un-wired.
3593 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3597 int index, newnpages;
3599 from = round_page(from);
3600 to = round_page(to);
3601 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3603 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3604 p = bp->b_xio.xio_pages[index];
3605 if (p && (index < bp->b_xio.xio_npages)) {
3607 kprintf("vm_hold_free_pages: doffset: %lld, loffset: %lld\n",
3608 bp->b_bio2.bio_offset, bp->b_loffset);
3610 bp->b_xio.xio_pages[index] = NULL;
3613 vm_page_unwire(p, 0);
3617 bp->b_xio.xio_npages = newnpages;
3623 * Map a user buffer into KVM via a pbuf. On return the buffer's
3624 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
3628 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
3639 * bp had better have a command and it better be a pbuf.
3641 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3642 KKASSERT(bp->b_flags & B_PAGING);
3648 * Map the user data into KVM. Mappings have to be page-aligned.
3650 addr = (caddr_t)trunc_page((vm_offset_t)udata);
3653 vmprot = VM_PROT_READ;
3654 if (bp->b_cmd == BUF_CMD_READ)
3655 vmprot |= VM_PROT_WRITE;
3657 while (addr < udata + bytes) {
3659 * Do the vm_fault if needed; do the copy-on-write thing
3660 * when reading stuff off device into memory.
3662 * vm_fault_page*() returns a held VM page.
3664 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
3665 va = trunc_page(va);
3667 m = vm_fault_page_quick(va, vmprot, &error);
3669 for (i = 0; i < pidx; ++i) {
3670 vm_page_unhold(bp->b_xio.xio_pages[i]);
3671 bp->b_xio.xio_pages[i] = NULL;
3675 bp->b_xio.xio_pages[pidx] = m;
3681 * Map the page array and set the buffer fields to point to
3682 * the mapped data buffer.
3684 if (pidx > btoc(MAXPHYS))
3685 panic("vmapbuf: mapped more than MAXPHYS");
3686 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
3688 bp->b_xio.xio_npages = pidx;
3689 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
3690 bp->b_bcount = bytes;
3691 bp->b_bufsize = bytes;
3698 * Free the io map PTEs associated with this IO operation.
3699 * We also invalidate the TLB entries and restore the original b_addr.
3702 vunmapbuf(struct buf *bp)
3707 KKASSERT(bp->b_flags & B_PAGING);
3709 npages = bp->b_xio.xio_npages;
3710 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3711 for (pidx = 0; pidx < npages; ++pidx) {
3712 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
3713 bp->b_xio.xio_pages[pidx] = NULL;
3715 bp->b_xio.xio_npages = 0;
3716 bp->b_data = bp->b_kvabase;
3720 * Scan all buffers in the system and issue the callback.
3723 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
3729 for (n = 0; n < nbuf; ++n) {
3730 if ((error = callback(&buf[n], info)) < 0) {
3740 * print out statistics from the current status of the buffer pool
3741 * this can be toggeled by the system control option debug.syncprt
3750 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
3751 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
3753 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
3755 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
3758 TAILQ_FOREACH(bp, dp, b_freelist) {
3759 counts[bp->b_bufsize/PAGE_SIZE]++;
3763 kprintf("%s: total-%d", bname[i], count);
3764 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
3766 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
3774 DB_SHOW_COMMAND(buffer, db_show_buffer)
3777 struct buf *bp = (struct buf *)addr;
3780 db_printf("usage: show buffer <addr>\n");
3784 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3785 db_printf("b_cmd = %d\n", bp->b_cmd);
3786 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
3787 "b_resid = %d\n, b_data = %p, "
3788 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
3789 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3790 bp->b_data, bp->b_bio2.bio_offset, (bp->b_bio2.bio_next ? bp->b_bio2.bio_next->bio_offset : (off_t)-1));
3791 if (bp->b_xio.xio_npages) {
3793 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
3794 bp->b_xio.xio_npages);
3795 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3797 m = bp->b_xio.xio_pages[i];
3798 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3799 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3800 if ((i + 1) < bp->b_xio.xio_npages)