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 1.105 2008/06/19 23:27:35 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;
125 static int hidirtywait;
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);
214 * Heuristical wakeup for flstik. If the I/O is draining
215 * in large bursts we can allow new dirty buffers to be
216 * queued immediately.
220 wakeup(&hidirtybuffers);
228 * Called when buffer space is potentially available for recovery.
229 * getnewbuf() will block on this flag when it is unable to free
230 * sufficient buffer space. Buffer space becomes recoverable when
231 * bp's get placed back in the queues.
238 * If someone is waiting for BUF space, wake them up. Even
239 * though we haven't freed the kva space yet, the waiting
240 * process will be able to now.
242 if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
243 spin_lock_wr(&needsbuffer_spin);
244 needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
245 spin_unlock_wr(&needsbuffer_spin);
246 wakeup(&needsbuffer);
253 * Accounting for I/O in progress.
257 runningbufwakeup(struct buf *bp)
259 if (bp->b_runningbufspace) {
260 runningbufspace -= bp->b_runningbufspace;
262 bp->b_runningbufspace = 0;
263 if (runningbufreq && runningbufspace <= lorunningspace) {
265 wakeup(&runningbufreq);
274 * Called when a buffer has been added to one of the free queues to
275 * account for the buffer and to wakeup anyone waiting for free buffers.
276 * This typically occurs when large amounts of metadata are being handled
277 * by the buffer cache ( else buffer space runs out first, usually ).
285 spin_lock_wr(&needsbuffer_spin);
286 needsbuffer &= ~VFS_BIO_NEED_ANY;
287 if (numfreebuffers >= hifreebuffers)
288 needsbuffer &= ~VFS_BIO_NEED_FREE;
289 spin_unlock_wr(&needsbuffer_spin);
290 wakeup(&needsbuffer);
295 * waitrunningbufspace()
297 * runningbufspace is a measure of the amount of I/O currently
298 * running. This routine is used in async-write situations to
299 * prevent creating huge backups of pending writes to a device.
300 * Only asynchronous writes are governed by this function.
302 * Reads will adjust runningbufspace, but will not block based on it.
303 * The read load has a side effect of reducing the allowed write load.
305 * This does NOT turn an async write into a sync write. It waits
306 * for earlier writes to complete and generally returns before the
307 * caller's write has reached the device.
310 waitrunningbufspace(void)
312 if (runningbufspace > hirunningspace) {
314 while (runningbufspace > hirunningspace) {
316 tsleep(&runningbufreq, 0, "wdrain", 0);
323 * vfs_buf_test_cache:
325 * Called when a buffer is extended. This function clears the B_CACHE
326 * bit if the newly extended portion of the buffer does not contain
331 vfs_buf_test_cache(struct buf *bp,
332 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
335 if (bp->b_flags & B_CACHE) {
336 int base = (foff + off) & PAGE_MASK;
337 if (vm_page_is_valid(m, base, size) == 0)
338 bp->b_flags &= ~B_CACHE;
345 * Wake up the buffer daemon if the number of outstanding dirty buffers
346 * is above specified threshold 'dirtybuflevel'.
348 * The buffer daemons are explicitly woken up when (a) the pending number
349 * of dirty buffers exceeds the recovery and stall mid-point value,
350 * (b) during bwillwrite() or (c) buf freelist was exhausted.
352 * The buffer daemons will generally not stop flushing until the dirty
353 * buffer count goes below lodirtybuffers.
357 bd_wakeup(int dirtybuflevel)
359 if (bd_request == 0 && numdirtybuffers &&
360 runningbufcount + numdirtybuffers >= dirtybuflevel) {
361 spin_lock_wr(&needsbuffer_spin);
363 spin_unlock_wr(&needsbuffer_spin);
366 if (bd_request_hw == 0 && numdirtybuffershw &&
367 numdirtybuffershw >= dirtybuflevel) {
368 spin_lock_wr(&needsbuffer_spin);
370 spin_unlock_wr(&needsbuffer_spin);
371 wakeup(&bd_request_hw);
378 * Speed up the buffer cache flushing process.
391 * Load time initialisation of the buffer cache, called from machine
392 * dependant initialization code.
398 vm_offset_t bogus_offset;
401 spin_init(&needsbuffer_spin);
403 /* next, make a null set of free lists */
404 for (i = 0; i < BUFFER_QUEUES; i++)
405 TAILQ_INIT(&bufqueues[i]);
407 /* finally, initialize each buffer header and stick on empty q */
408 for (i = 0; i < nbuf; i++) {
410 bzero(bp, sizeof *bp);
411 bp->b_flags = B_INVAL; /* we're just an empty header */
412 bp->b_cmd = BUF_CMD_DONE;
413 bp->b_qindex = BQUEUE_EMPTY;
415 xio_init(&bp->b_xio);
418 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
422 * maxbufspace is the absolute maximum amount of buffer space we are
423 * allowed to reserve in KVM and in real terms. The absolute maximum
424 * is nominally used by buf_daemon. hibufspace is the nominal maximum
425 * used by most other processes. The differential is required to
426 * ensure that buf_daemon is able to run when other processes might
427 * be blocked waiting for buffer space.
429 * maxbufspace is based on BKVASIZE. Allocating buffers larger then
430 * this may result in KVM fragmentation which is not handled optimally
433 maxbufspace = nbuf * BKVASIZE;
434 hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
435 lobufspace = hibufspace - MAXBSIZE;
437 lorunningspace = 512 * 1024;
438 hirunningspace = 1024 * 1024;
441 * Limit the amount of malloc memory since it is wired permanently into
442 * the kernel space. Even though this is accounted for in the buffer
443 * allocation, we don't want the malloced region to grow uncontrolled.
444 * The malloc scheme improves memory utilization significantly on average
445 * (small) directories.
447 maxbufmallocspace = hibufspace / 20;
450 * Reduce the chance of a deadlock occuring by limiting the number
451 * of delayed-write dirty buffers we allow to stack up.
453 hidirtybuffers = nbuf / 4 + 20;
455 numdirtybuffershw = 0;
457 * To support extreme low-memory systems, make sure hidirtybuffers cannot
458 * eat up all available buffer space. This occurs when our minimum cannot
459 * be met. We try to size hidirtybuffers to 3/4 our buffer space assuming
460 * BKVASIZE'd (8K) buffers.
462 while (hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) {
463 hidirtybuffers >>= 1;
465 lodirtybuffers = hidirtybuffers / 2;
468 * Try to keep the number of free buffers in the specified range,
469 * and give special processes (e.g. like buf_daemon) access to an
472 lofreebuffers = nbuf / 18 + 5;
473 hifreebuffers = 2 * lofreebuffers;
474 numfreebuffers = nbuf;
477 * Maximum number of async ops initiated per buf_daemon loop. This is
478 * somewhat of a hack at the moment, we really need to limit ourselves
479 * based on the number of bytes of I/O in-transit that were initiated
483 bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
484 bogus_page = vm_page_alloc(&kernel_object,
485 (bogus_offset >> PAGE_SHIFT),
487 vmstats.v_wire_count++;
492 * Initialize the embedded bio structures
495 initbufbio(struct buf *bp)
497 bp->b_bio1.bio_buf = bp;
498 bp->b_bio1.bio_prev = NULL;
499 bp->b_bio1.bio_offset = NOOFFSET;
500 bp->b_bio1.bio_next = &bp->b_bio2;
501 bp->b_bio1.bio_done = NULL;
503 bp->b_bio2.bio_buf = bp;
504 bp->b_bio2.bio_prev = &bp->b_bio1;
505 bp->b_bio2.bio_offset = NOOFFSET;
506 bp->b_bio2.bio_next = NULL;
507 bp->b_bio2.bio_done = NULL;
511 * Reinitialize the embedded bio structures as well as any additional
512 * translation cache layers.
515 reinitbufbio(struct buf *bp)
519 for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
520 bio->bio_done = NULL;
521 bio->bio_offset = NOOFFSET;
526 * Push another BIO layer onto an existing BIO and return it. The new
527 * BIO layer may already exist, holding cached translation data.
530 push_bio(struct bio *bio)
534 if ((nbio = bio->bio_next) == NULL) {
535 int index = bio - &bio->bio_buf->b_bio_array[0];
536 if (index >= NBUF_BIO - 1) {
537 panic("push_bio: too many layers bp %p\n",
540 nbio = &bio->bio_buf->b_bio_array[index + 1];
541 bio->bio_next = nbio;
542 nbio->bio_prev = bio;
543 nbio->bio_buf = bio->bio_buf;
544 nbio->bio_offset = NOOFFSET;
545 nbio->bio_done = NULL;
546 nbio->bio_next = NULL;
548 KKASSERT(nbio->bio_done == NULL);
553 pop_bio(struct bio *bio)
559 clearbiocache(struct bio *bio)
562 bio->bio_offset = NOOFFSET;
570 * Free the KVA allocation for buffer 'bp'.
572 * Must be called from a critical section as this is the only locking for
575 * Since this call frees up buffer space, we call bufspacewakeup().
578 bfreekva(struct buf *bp)
584 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
585 vm_map_lock(&buffer_map);
586 bufspace -= bp->b_kvasize;
587 vm_map_delete(&buffer_map,
588 (vm_offset_t) bp->b_kvabase,
589 (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
592 vm_map_unlock(&buffer_map);
593 vm_map_entry_release(count);
602 * Remove the buffer from the appropriate free list.
605 bremfree(struct buf *bp)
610 old_qindex = bp->b_qindex;
612 if (bp->b_qindex != BQUEUE_NONE) {
613 KASSERT(BUF_REFCNTNB(bp) == 1,
614 ("bremfree: bp %p not locked",bp));
615 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
616 bp->b_qindex = BQUEUE_NONE;
618 if (BUF_REFCNTNB(bp) <= 1)
619 panic("bremfree: removing a buffer not on a queue");
623 * Fixup numfreebuffers count. If the buffer is invalid or not
624 * delayed-write, and it was on the EMPTY, LRU, or AGE queues,
625 * the buffer was free and we must decrement numfreebuffers.
627 if ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0) {
630 case BQUEUE_DIRTY_HW:
633 case BQUEUE_EMPTYKVA:
647 * Get a buffer with the specified data. Look in the cache first. We
648 * must clear B_ERROR and B_INVAL prior to initiating I/O. If B_CACHE
649 * is set, the buffer is valid and we do not have to do anything ( see
653 bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
657 bp = getblk(vp, loffset, size, 0, 0);
660 /* if not found in cache, do some I/O */
661 if ((bp->b_flags & B_CACHE) == 0) {
662 KASSERT(!(bp->b_flags & B_ASYNC), ("bread: illegal async bp %p", bp));
663 bp->b_flags &= ~(B_ERROR | B_INVAL);
664 bp->b_cmd = BUF_CMD_READ;
665 vfs_busy_pages(vp, bp);
666 vn_strategy(vp, &bp->b_bio1);
667 return (biowait(bp));
675 * Operates like bread, but also starts asynchronous I/O on
676 * read-ahead blocks. We must clear B_ERROR and B_INVAL prior
677 * to initiating I/O . If B_CACHE is set, the buffer is valid
678 * and we do not have to do anything.
681 breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
682 int *rabsize, int cnt, struct buf **bpp)
684 struct buf *bp, *rabp;
686 int rv = 0, readwait = 0;
688 *bpp = bp = getblk(vp, loffset, size, 0, 0);
690 /* if not found in cache, do some I/O */
691 if ((bp->b_flags & B_CACHE) == 0) {
692 bp->b_flags &= ~(B_ERROR | B_INVAL);
693 bp->b_cmd = BUF_CMD_READ;
694 vfs_busy_pages(vp, bp);
695 vn_strategy(vp, &bp->b_bio1);
699 for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
700 if (inmem(vp, *raoffset))
702 rabp = getblk(vp, *raoffset, *rabsize, 0, 0);
704 if ((rabp->b_flags & B_CACHE) == 0) {
705 rabp->b_flags |= B_ASYNC;
706 rabp->b_flags &= ~(B_ERROR | B_INVAL);
707 rabp->b_cmd = BUF_CMD_READ;
708 vfs_busy_pages(vp, rabp);
710 vn_strategy(vp, &rabp->b_bio1);
725 * Write, release buffer on completion. (Done by iodone
726 * if async). Do not bother writing anything if the buffer
729 * Note that we set B_CACHE here, indicating that buffer is
730 * fully valid and thus cacheable. This is true even of NFS
731 * now so we set it generally. This could be set either here
732 * or in biodone() since the I/O is synchronous. We put it
736 bwrite(struct buf *bp)
740 if (bp->b_flags & B_INVAL) {
745 oldflags = bp->b_flags;
747 if (BUF_REFCNTNB(bp) == 0)
748 panic("bwrite: buffer is not busy???");
751 /* Mark the buffer clean */
754 bp->b_flags &= ~B_ERROR;
755 bp->b_flags |= B_CACHE;
756 bp->b_cmd = BUF_CMD_WRITE;
757 vfs_busy_pages(bp->b_vp, bp);
760 * Normal bwrites pipeline writes. NOTE: b_bufsize is only
761 * valid for vnode-backed buffers.
763 bp->b_runningbufspace = bp->b_bufsize;
764 if (bp->b_runningbufspace) {
765 runningbufspace += bp->b_runningbufspace;
770 if (oldflags & B_ASYNC)
772 vn_strategy(bp->b_vp, &bp->b_bio1);
774 if ((oldflags & B_ASYNC) == 0) {
775 int rtval = biowait(bp);
785 * Delayed write. (Buffer is marked dirty). Do not bother writing
786 * anything if the buffer is marked invalid.
788 * Note that since the buffer must be completely valid, we can safely
789 * set B_CACHE. In fact, we have to set B_CACHE here rather then in
790 * biodone() in order to prevent getblk from writing the buffer
794 bdwrite(struct buf *bp)
796 if (BUF_REFCNTNB(bp) == 0)
797 panic("bdwrite: buffer is not busy");
799 if (bp->b_flags & B_INVAL) {
806 * Set B_CACHE, indicating that the buffer is fully valid. This is
807 * true even of NFS now.
809 bp->b_flags |= B_CACHE;
812 * This bmap keeps the system from needing to do the bmap later,
813 * perhaps when the system is attempting to do a sync. Since it
814 * is likely that the indirect block -- or whatever other datastructure
815 * that the filesystem needs is still in memory now, it is a good
816 * thing to do this. Note also, that if the pageout daemon is
817 * requesting a sync -- there might not be enough memory to do
818 * the bmap then... So, this is important to do.
820 if (bp->b_bio2.bio_offset == NOOFFSET) {
821 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
822 NULL, NULL, BUF_CMD_WRITE);
826 * Set the *dirty* buffer range based upon the VM system dirty pages.
831 * We need to do this here to satisfy the vnode_pager and the
832 * pageout daemon, so that it thinks that the pages have been
833 * "cleaned". Note that since the pages are in a delayed write
834 * buffer -- the VFS layer "will" see that the pages get written
835 * out on the next sync, or perhaps the cluster will be completed.
841 * Wakeup the buffer flushing daemon if we have a lot of dirty
842 * buffers (midpoint between our recovery point and our stall
845 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
848 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
849 * due to the softdep code.
856 * Turn buffer into delayed write request by marking it B_DELWRI.
857 * B_RELBUF and B_NOCACHE must be cleared.
859 * We reassign the buffer to itself to properly update it in the
862 * Since the buffer is not on a queue, we do not update the
863 * numfreebuffers count.
865 * Must be called from a critical section.
866 * The buffer must be on BQUEUE_NONE.
869 bdirty(struct buf *bp)
871 KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
872 if (bp->b_flags & B_NOCACHE) {
873 kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
874 bp->b_flags &= ~B_NOCACHE;
876 if (bp->b_flags & B_INVAL) {
877 kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
879 bp->b_flags &= ~B_RELBUF;
881 if ((bp->b_flags & B_DELWRI) == 0) {
882 bp->b_flags |= B_DELWRI;
885 if (bp->b_flags & B_HEAVY)
887 bd_wakeup((lodirtybuffers + hidirtybuffers) / 2);
892 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
893 * needs to be flushed with a different buf_daemon thread to avoid
894 * deadlocks. B_HEAVY also imposes restrictions in getnewbuf().
897 bheavy(struct buf *bp)
899 if ((bp->b_flags & B_HEAVY) == 0) {
900 bp->b_flags |= B_HEAVY;
901 if (bp->b_flags & B_DELWRI)
909 * Clear B_DELWRI for buffer.
911 * Since the buffer is not on a queue, we do not update the numfreebuffers
914 * Must be called from a critical section.
916 * The buffer is typically on BQUEUE_NONE but there is one case in
917 * brelse() that calls this function after placing the buffer on
922 bundirty(struct buf *bp)
924 if (bp->b_flags & B_DELWRI) {
925 bp->b_flags &= ~B_DELWRI;
928 if (bp->b_flags & B_HEAVY)
933 * Since it is now being written, we can clear its deferred write flag.
935 bp->b_flags &= ~B_DEFERRED;
941 * Asynchronous write. Start output on a buffer, but do not wait for
942 * it to complete. The buffer is released when the output completes.
944 * bwrite() ( or the VOP routine anyway ) is responsible for handling
945 * B_INVAL buffers. Not us.
948 bawrite(struct buf *bp)
950 bp->b_flags |= B_ASYNC;
957 * Ordered write. Start output on a buffer, and flag it so that the
958 * device will write it in the order it was queued. The buffer is
959 * released when the output completes. bwrite() ( or the VOP routine
960 * anyway ) is responsible for handling B_INVAL buffers.
963 bowrite(struct buf *bp)
965 bp->b_flags |= B_ORDERED | B_ASYNC;
972 * Called prior to the locking of any vnodes when we are expecting to
973 * write. We do not want to starve the buffer cache with too many
974 * dirty buffers so we block here. By blocking prior to the locking
975 * of any vnodes we attempt to avoid the situation where a locked vnode
976 * prevents the various system daemons from flushing related buffers.
981 int mid1 = hidirtybuffers / 2;
982 int mid2 = mid1 + hidirtybuffers / 4;
987 count = runningbufcount + numdirtybuffers;
990 * Nothing to do if nothing is stressed.
996 * Get the buffer daemon heated up
1000 while (count >= mid2) {
1002 * Start slowing down writes, down to 1 per second.
1004 if (count < hidirtybuffers) {
1005 delay = (count - mid2) * hz / (hidirtybuffers - mid2);
1006 delay = delay * 10 / (10 + priority);
1010 tsleep(&hidirtybuffers, 0, "flstik", delay);
1015 * Now we are really in trouble.
1018 spin_lock_wr(&needsbuffer_spin);
1019 count = runningbufcount + numdirtybuffers;
1020 if (count >= hidirtybuffers) {
1021 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1022 msleep(&needsbuffer, &needsbuffer_spin, 0, "flswai", 0);
1023 spin_unlock_wr(&needsbuffer_spin);
1025 count = runningbufcount + numdirtybuffers;
1028 /* FUTURE - maybe */
1029 else if (runningbufcount + numdirtybuffershw > hidirtybuffers / 2) {
1032 while (runningbufcount + numdirtybuffershw > hidirtybuffers) {
1033 needsbuffer |= VFS_BIO_NEED_DIRTYFLUSH;
1034 tsleep(&needsbuffer, slpflags, "newbuf",
1042 * buf_dirty_count_severe:
1044 * Return true if we have too many dirty buffers.
1047 buf_dirty_count_severe(void)
1049 return(runningbufcount + numdirtybuffers >= hidirtybuffers);
1055 * Release a busy buffer and, if requested, free its resources. The
1056 * buffer will be stashed in the appropriate bufqueue[] allowing it
1057 * to be accessed later as a cache entity or reused for other purposes.
1060 brelse(struct buf *bp)
1063 int saved_flags = bp->b_flags;
1066 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1071 * If B_NOCACHE is set we are being asked to destroy the buffer and
1072 * its backing store. Clear B_DELWRI.
1074 * B_NOCACHE is set in two cases: (1) when the caller really wants
1075 * to destroy the buffer and backing store and (2) when the caller
1076 * wants to destroy the buffer and backing store after a write
1079 if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
1083 if (bp->b_flags & B_LOCKED)
1084 bp->b_flags &= ~B_ERROR;
1087 * If a write error occurs and the caller does not want to throw
1088 * away the buffer, redirty the buffer. This will also clear
1091 if (bp->b_cmd == BUF_CMD_WRITE &&
1092 (bp->b_flags & (B_ERROR | B_INVAL)) == B_ERROR) {
1094 * Failed write, redirty. Must clear B_ERROR to prevent
1095 * pages from being scrapped. If B_INVAL is set then
1096 * this case is not run and the next case is run to
1097 * destroy the buffer. B_INVAL can occur if the buffer
1098 * is outside the range supported by the underlying device.
1100 bp->b_flags &= ~B_ERROR;
1102 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
1103 (bp->b_bufsize <= 0) || bp->b_cmd == BUF_CMD_FREEBLKS) {
1105 * Either a failed I/O or we were asked to free or not
1108 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
1109 * buffer cannot be immediately freed.
1111 bp->b_flags |= B_INVAL;
1112 if (LIST_FIRST(&bp->b_dep) != NULL)
1114 if (bp->b_flags & B_DELWRI) {
1116 if (bp->b_flags & B_HEAVY)
1117 --numdirtybuffershw;
1120 bp->b_flags &= ~(B_DELWRI | B_CACHE);
1124 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set.
1125 * If vfs_vmio_release() is called with either bit set, the
1126 * underlying pages may wind up getting freed causing a previous
1127 * write (bdwrite()) to get 'lost' because pages associated with
1128 * a B_DELWRI bp are marked clean. Pages associated with a
1129 * B_LOCKED buffer may be mapped by the filesystem.
1131 * If we want to release the buffer ourselves (rather then the
1132 * originator asking us to release it), give the originator a
1133 * chance to countermand the release by setting B_LOCKED.
1135 * We still allow the B_INVAL case to call vfs_vmio_release(), even
1136 * if B_DELWRI is set.
1138 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
1139 * on pages to return pages to the VM page queues.
1141 if (bp->b_flags & (B_DELWRI | B_LOCKED)) {
1142 bp->b_flags &= ~B_RELBUF;
1143 } else if (vm_page_count_severe()) {
1144 if (LIST_FIRST(&bp->b_dep) != NULL)
1146 if (bp->b_flags & (B_DELWRI | B_LOCKED))
1147 bp->b_flags &= ~B_RELBUF;
1149 bp->b_flags |= B_RELBUF;
1153 * At this point destroying the buffer is governed by the B_INVAL
1154 * or B_RELBUF flags.
1156 bp->b_cmd = BUF_CMD_DONE;
1159 * VMIO buffer rundown. Make sure the VM page array is restored
1160 * after an I/O may have replaces some of the pages with bogus pages
1161 * in order to not destroy dirty pages in a fill-in read.
1163 * Note that due to the code above, if a buffer is marked B_DELWRI
1164 * then the B_RELBUF and B_NOCACHE bits will always be clear.
1165 * B_INVAL may still be set, however.
1167 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
1168 * but not the backing store. B_NOCACHE will destroy the backing
1171 * Note that dirty NFS buffers contain byte-granular write ranges
1172 * and should not be destroyed w/ B_INVAL even if the backing store
1175 if (bp->b_flags & B_VMIO) {
1177 * Rundown for VMIO buffers which are not dirty NFS buffers.
1189 * Get the base offset and length of the buffer. Note that
1190 * in the VMIO case if the buffer block size is not
1191 * page-aligned then b_data pointer may not be page-aligned.
1192 * But our b_xio.xio_pages array *IS* page aligned.
1194 * block sizes less then DEV_BSIZE (usually 512) are not
1195 * supported due to the page granularity bits (m->valid,
1196 * m->dirty, etc...).
1198 * See man buf(9) for more information
1201 resid = bp->b_bufsize;
1202 foff = bp->b_loffset;
1204 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1205 m = bp->b_xio.xio_pages[i];
1206 vm_page_flag_clear(m, PG_ZERO);
1208 * If we hit a bogus page, fixup *all* of them
1209 * now. Note that we left these pages wired
1210 * when we removed them so they had better exist,
1211 * and they cannot be ripped out from under us so
1212 * no critical section protection is necessary.
1214 if (m == bogus_page) {
1216 poff = OFF_TO_IDX(bp->b_loffset);
1218 for (j = i; j < bp->b_xio.xio_npages; j++) {
1221 mtmp = bp->b_xio.xio_pages[j];
1222 if (mtmp == bogus_page) {
1223 mtmp = vm_page_lookup(obj, poff + j);
1225 panic("brelse: page missing");
1227 bp->b_xio.xio_pages[j] = mtmp;
1231 if ((bp->b_flags & B_INVAL) == 0) {
1232 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
1233 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1235 m = bp->b_xio.xio_pages[i];
1239 * Invalidate the backing store if B_NOCACHE is set
1240 * (e.g. used with vinvalbuf()). If this is NFS
1241 * we impose a requirement that the block size be
1242 * a multiple of PAGE_SIZE and create a temporary
1243 * hack to basically invalidate the whole page. The
1244 * problem is that NFS uses really odd buffer sizes
1245 * especially when tracking piecemeal writes and
1246 * it also vinvalbuf()'s a lot, which would result
1247 * in only partial page validation and invalidation
1248 * here. If the file page is mmap()'d, however,
1249 * all the valid bits get set so after we invalidate
1250 * here we would end up with weird m->valid values
1251 * like 0xfc. nfs_getpages() can't handle this so
1252 * we clear all the valid bits for the NFS case
1253 * instead of just some of them.
1255 * The real bug is the VM system having to set m->valid
1256 * to VM_PAGE_BITS_ALL for faulted-in pages, which
1257 * itself is an artifact of the whole 512-byte
1258 * granular mess that exists to support odd block
1259 * sizes and UFS meta-data block sizes (e.g. 6144).
1260 * A complete rewrite is required.
1262 if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
1263 int poffset = foff & PAGE_MASK;
1266 presid = PAGE_SIZE - poffset;
1267 if (bp->b_vp->v_tag == VT_NFS &&
1268 bp->b_vp->v_type == VREG) {
1270 } else if (presid > resid) {
1273 KASSERT(presid >= 0, ("brelse: extra page"));
1274 vm_page_set_invalid(m, poffset, presid);
1276 resid -= PAGE_SIZE - (foff & PAGE_MASK);
1277 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
1279 if (bp->b_flags & (B_INVAL | B_RELBUF))
1280 vfs_vmio_release(bp);
1283 * Rundown for non-VMIO buffers.
1285 if (bp->b_flags & (B_INVAL | B_RELBUF)) {
1288 kprintf("brelse bp %p %08x/%08x: Warning, caught and fixed brelvp bug\n", bp, saved_flags, bp->b_flags);
1292 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1298 if (bp->b_qindex != BQUEUE_NONE)
1299 panic("brelse: free buffer onto another queue???");
1300 if (BUF_REFCNTNB(bp) > 1) {
1301 /* Temporary panic to verify exclusive locking */
1302 /* This panic goes away when we allow shared refs */
1303 panic("brelse: multiple refs");
1304 /* do not release to free list */
1311 * Figure out the correct queue to place the cleaned up buffer on.
1312 * Buffers placed in the EMPTY or EMPTYKVA had better already be
1313 * disassociated from their vnode.
1315 if (bp->b_flags & B_LOCKED) {
1317 * Buffers that are locked are placed in the locked queue
1318 * immediately, regardless of their state.
1320 bp->b_qindex = BQUEUE_LOCKED;
1321 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1322 } else if (bp->b_bufsize == 0) {
1324 * Buffers with no memory. Due to conditionals near the top
1325 * of brelse() such buffers should probably already be
1326 * marked B_INVAL and disassociated from their vnode.
1328 bp->b_flags |= B_INVAL;
1329 KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1330 KKASSERT((bp->b_flags & B_HASHED) == 0);
1331 if (bp->b_kvasize) {
1332 bp->b_qindex = BQUEUE_EMPTYKVA;
1334 bp->b_qindex = BQUEUE_EMPTY;
1336 TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
1337 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) {
1339 * Buffers with junk contents. Again these buffers had better
1340 * already be disassociated from their vnode.
1342 KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
1343 KKASSERT((bp->b_flags & B_HASHED) == 0);
1344 bp->b_flags |= B_INVAL;
1345 bp->b_qindex = BQUEUE_CLEAN;
1346 TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1349 * Remaining buffers. These buffers are still associated with
1352 switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
1354 bp->b_qindex = BQUEUE_DIRTY;
1355 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
1357 case B_DELWRI | B_HEAVY:
1358 bp->b_qindex = BQUEUE_DIRTY_HW;
1359 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
1364 * NOTE: Buffers are always placed at the end of the
1365 * queue. If B_AGE is not set the buffer will cycle
1366 * through the queue twice.
1368 bp->b_qindex = BQUEUE_CLEAN;
1369 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1375 * If B_INVAL, clear B_DELWRI. We've already placed the buffer
1376 * on the correct queue.
1378 if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
1382 * Fixup numfreebuffers count. The bp is on an appropriate queue
1383 * unless locked. We then bump numfreebuffers if it is not B_DELWRI.
1384 * We've already handled the B_INVAL case ( B_DELWRI will be clear
1385 * if B_INVAL is set ).
1387 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
1391 * Something we can maybe free or reuse
1393 if (bp->b_bufsize || bp->b_kvasize)
1397 * Clean up temporary flags and unlock the buffer.
1399 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT);
1407 * Release a buffer back to the appropriate queue but do not try to free
1408 * it. The buffer is expected to be used again soon.
1410 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by
1411 * biodone() to requeue an async I/O on completion. It is also used when
1412 * known good buffers need to be requeued but we think we may need the data
1415 * XXX we should be able to leave the B_RELBUF hint set on completion.
1418 bqrelse(struct buf *bp)
1422 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));
1424 if (bp->b_qindex != BQUEUE_NONE)
1425 panic("bqrelse: free buffer onto another queue???");
1426 if (BUF_REFCNTNB(bp) > 1) {
1427 /* do not release to free list */
1428 panic("bqrelse: multiple refs");
1433 if (bp->b_flags & B_LOCKED) {
1435 * Locked buffers are released to the locked queue. However,
1436 * if the buffer is dirty it will first go into the dirty
1437 * queue and later on after the I/O completes successfully it
1438 * will be released to the locked queue.
1440 bp->b_flags &= ~B_ERROR;
1441 bp->b_qindex = BQUEUE_LOCKED;
1442 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
1443 } else if (bp->b_flags & B_DELWRI) {
1444 bp->b_qindex = (bp->b_flags & B_HEAVY) ?
1445 BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
1446 TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
1447 } else if (vm_page_count_severe()) {
1449 * We are too low on memory, we have to try to free the
1450 * buffer (most importantly: the wired pages making up its
1451 * backing store) *now*.
1457 bp->b_qindex = BQUEUE_CLEAN;
1458 TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
1461 if ((bp->b_flags & B_LOCKED) == 0 &&
1462 ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
1467 * Something we can maybe free or reuse.
1469 if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
1473 * Final cleanup and unlock. Clear bits that are only used while a
1474 * buffer is actively locked.
1476 bp->b_flags &= ~(B_ORDERED | B_ASYNC | B_NOCACHE | B_RELBUF);
1484 * Return backing pages held by the buffer 'bp' back to the VM system
1485 * if possible. The pages are freed if they are no longer valid or
1486 * attempt to free if it was used for direct I/O otherwise they are
1487 * sent to the page cache.
1489 * Pages that were marked busy are left alone and skipped.
1491 * The KVA mapping (b_data) for the underlying pages is removed by
1495 vfs_vmio_release(struct buf *bp)
1501 for (i = 0; i < bp->b_xio.xio_npages; i++) {
1502 m = bp->b_xio.xio_pages[i];
1503 bp->b_xio.xio_pages[i] = NULL;
1505 * In order to keep page LRU ordering consistent, put
1506 * everything on the inactive queue.
1508 vm_page_unwire(m, 0);
1510 * We don't mess with busy pages, it is
1511 * the responsibility of the process that
1512 * busied the pages to deal with them.
1514 if ((m->flags & PG_BUSY) || (m->busy != 0))
1517 if (m->wire_count == 0) {
1518 vm_page_flag_clear(m, PG_ZERO);
1520 * Might as well free the page if we can and it has
1521 * no valid data. We also free the page if the
1522 * buffer was used for direct I/O.
1524 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
1525 m->hold_count == 0) {
1527 vm_page_protect(m, VM_PROT_NONE);
1529 } else if (bp->b_flags & B_DIRECT) {
1530 vm_page_try_to_free(m);
1531 } else if (vm_page_count_severe()) {
1532 vm_page_try_to_cache(m);
1537 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
1538 if (bp->b_bufsize) {
1542 bp->b_xio.xio_npages = 0;
1543 bp->b_flags &= ~B_VMIO;
1544 KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
1552 * Implement clustered async writes for clearing out B_DELWRI buffers.
1553 * This is much better then the old way of writing only one buffer at
1554 * a time. Note that we may not be presented with the buffers in the
1555 * correct order, so we search for the cluster in both directions.
1557 * The buffer is locked on call.
1560 vfs_bio_awrite(struct buf *bp)
1564 off_t loffset = bp->b_loffset;
1565 struct vnode *vp = bp->b_vp;
1573 * right now we support clustered writing only to regular files. If
1574 * we find a clusterable block we could be in the middle of a cluster
1575 * rather then at the beginning.
1577 * NOTE: b_bio1 contains the logical loffset and is aliased
1578 * to b_loffset. b_bio2 contains the translated block number.
1580 if ((vp->v_type == VREG) &&
1581 (vp->v_mount != 0) && /* Only on nodes that have the size info */
1582 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {
1584 size = vp->v_mount->mnt_stat.f_iosize;
1586 for (i = size; i < MAXPHYS; i += size) {
1587 if ((bpa = findblk(vp, loffset + i)) &&
1588 BUF_REFCNT(bpa) == 0 &&
1589 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1590 (B_DELWRI | B_CLUSTEROK)) &&
1591 (bpa->b_bufsize == size)) {
1592 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1593 (bpa->b_bio2.bio_offset !=
1594 bp->b_bio2.bio_offset + i))
1600 for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
1601 if ((bpa = findblk(vp, loffset - j)) &&
1602 BUF_REFCNT(bpa) == 0 &&
1603 ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
1604 (B_DELWRI | B_CLUSTEROK)) &&
1605 (bpa->b_bufsize == size)) {
1606 if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
1607 (bpa->b_bio2.bio_offset !=
1608 bp->b_bio2.bio_offset - j))
1617 * this is a possible cluster write
1619 if (nbytes != size) {
1621 nwritten = cluster_wbuild(vp, size,
1622 loffset - j, nbytes);
1629 bp->b_flags |= B_ASYNC;
1633 * default (old) behavior, writing out only one block
1635 * XXX returns b_bufsize instead of b_bcount for nwritten?
1637 nwritten = bp->b_bufsize;
1646 * Find and initialize a new buffer header, freeing up existing buffers
1647 * in the bufqueues as necessary. The new buffer is returned locked.
1649 * Important: B_INVAL is not set. If the caller wishes to throw the
1650 * buffer away, the caller must set B_INVAL prior to calling brelse().
1653 * We have insufficient buffer headers
1654 * We have insufficient buffer space
1655 * buffer_map is too fragmented ( space reservation fails )
1656 * If we have to flush dirty buffers ( but we try to avoid this )
1658 * To avoid VFS layer recursion we do not flush dirty buffers ourselves.
1659 * Instead we ask the buf daemon to do it for us. We attempt to
1660 * avoid piecemeal wakeups of the pageout daemon.
1664 getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
1670 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
1671 static int flushingbufs;
1674 * We can't afford to block since we might be holding a vnode lock,
1675 * which may prevent system daemons from running. We deal with
1676 * low-memory situations by proactively returning memory and running
1677 * async I/O rather then sync I/O.
1681 --getnewbufrestarts;
1683 ++getnewbufrestarts;
1686 * Setup for scan. If we do not have enough free buffers,
1687 * we setup a degenerate case that immediately fails. Note
1688 * that if we are specially marked process, we are allowed to
1689 * dip into our reserves.
1691 * The scanning sequence is nominally: EMPTY->EMPTYKVA->CLEAN
1693 * We start with EMPTYKVA. If the list is empty we backup to EMPTY.
1694 * However, there are a number of cases (defragging, reusing, ...)
1695 * where we cannot backup.
1697 nqindex = BQUEUE_EMPTYKVA;
1698 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);
1702 * If no EMPTYKVA buffers and we are either
1703 * defragging or reusing, locate a CLEAN buffer
1704 * to free or reuse. If bufspace useage is low
1705 * skip this step so we can allocate a new buffer.
1707 if (defrag || bufspace >= lobufspace) {
1708 nqindex = BQUEUE_CLEAN;
1709 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
1713 * If we could not find or were not allowed to reuse a
1714 * CLEAN buffer, check to see if it is ok to use an EMPTY
1715 * buffer. We can only use an EMPTY buffer if allocating
1716 * its KVA would not otherwise run us out of buffer space.
1718 if (nbp == NULL && defrag == 0 &&
1719 bufspace + maxsize < hibufspace) {
1720 nqindex = BQUEUE_EMPTY;
1721 nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
1726 * Run scan, possibly freeing data and/or kva mappings on the fly
1730 while ((bp = nbp) != NULL) {
1731 int qindex = nqindex;
1733 nbp = TAILQ_NEXT(bp, b_freelist);
1736 * BQUEUE_CLEAN - B_AGE special case. If not set the bp
1737 * cycles through the queue twice before being selected.
1739 if (qindex == BQUEUE_CLEAN &&
1740 (bp->b_flags & B_AGE) == 0 && nbp) {
1741 bp->b_flags |= B_AGE;
1742 TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
1743 TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
1748 * Calculate next bp ( we can only use it if we do not block
1749 * or do other fancy things ).
1754 nqindex = BQUEUE_EMPTYKVA;
1755 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
1758 case BQUEUE_EMPTYKVA:
1759 nqindex = BQUEUE_CLEAN;
1760 if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
1774 KASSERT(bp->b_qindex == qindex, ("getnewbuf: inconsistent queue %d bp %p", qindex, bp));
1777 * Note: we no longer distinguish between VMIO and non-VMIO
1781 KASSERT((bp->b_flags & B_DELWRI) == 0, ("delwri buffer %p found in queue %d", bp, qindex));
1784 * If we are defragging then we need a buffer with
1785 * b_kvasize != 0. XXX this situation should no longer
1786 * occur, if defrag is non-zero the buffer's b_kvasize
1787 * should also be non-zero at this point. XXX
1789 if (defrag && bp->b_kvasize == 0) {
1790 kprintf("Warning: defrag empty buffer %p\n", bp);
1795 * Start freeing the bp. This is somewhat involved. nbp
1796 * remains valid only for BQUEUE_EMPTY[KVA] bp's. Buffers
1797 * on the clean list must be disassociated from their
1798 * current vnode. Buffers on the empty[kva] lists have
1799 * already been disassociated.
1802 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1803 kprintf("getnewbuf: warning, locked buf %p, race corrected\n", bp);
1804 tsleep(&bd_request, 0, "gnbxxx", hz / 100);
1807 if (bp->b_qindex != qindex) {
1808 kprintf("getnewbuf: warning, BUF_LOCK blocked unexpectedly on buf %p index %d->%d, race corrected\n", bp, qindex, bp->b_qindex);
1815 * Dependancies must be handled before we disassociate the
1818 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
1819 * be immediately disassociated. HAMMER then becomes
1820 * responsible for releasing the buffer.
1822 if (LIST_FIRST(&bp->b_dep) != NULL) {
1824 if (bp->b_flags & B_LOCKED) {
1828 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1831 if (qindex == BQUEUE_CLEAN) {
1832 if (bp->b_flags & B_VMIO) {
1833 bp->b_flags &= ~B_ASYNC;
1834 vfs_vmio_release(bp);
1841 * NOTE: nbp is now entirely invalid. We can only restart
1842 * the scan from this point on.
1844 * Get the rest of the buffer freed up. b_kva* is still
1845 * valid after this operation.
1848 KASSERT(bp->b_vp == NULL, ("bp3 %p flags %08x vnode %p qindex %d unexpectededly still associated!", bp, bp->b_flags, bp->b_vp, qindex));
1849 KKASSERT((bp->b_flags & B_HASHED) == 0);
1852 * critical section protection is not required when
1853 * scrapping a buffer's contents because it is already
1859 bp->b_flags = B_BNOCLIP;
1860 bp->b_cmd = BUF_CMD_DONE;
1865 bp->b_xio.xio_npages = 0;
1866 bp->b_dirtyoff = bp->b_dirtyend = 0;
1868 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
1870 if (blkflags & GETBLK_BHEAVY)
1871 bp->b_flags |= B_HEAVY;
1874 * If we are defragging then free the buffer.
1877 bp->b_flags |= B_INVAL;
1885 * If we are overcomitted then recover the buffer and its
1886 * KVM space. This occurs in rare situations when multiple
1887 * processes are blocked in getnewbuf() or allocbuf().
1889 if (bufspace >= hibufspace)
1891 if (flushingbufs && bp->b_kvasize != 0) {
1892 bp->b_flags |= B_INVAL;
1897 if (bufspace < lobufspace)
1903 * If we exhausted our list, sleep as appropriate. We may have to
1904 * wakeup various daemons and write out some dirty buffers.
1906 * Generally we are sleeping due to insufficient buffer space.
1914 flags = VFS_BIO_NEED_BUFSPACE;
1916 } else if (bufspace >= hibufspace) {
1918 flags = VFS_BIO_NEED_BUFSPACE;
1921 flags = VFS_BIO_NEED_ANY;
1924 needsbuffer |= flags;
1925 bd_speedup(); /* heeeelp */
1926 while (needsbuffer & flags) {
1927 if (tsleep(&needsbuffer, slpflags, waitmsg, slptimeo))
1932 * We finally have a valid bp. We aren't quite out of the
1933 * woods, we still have to reserve kva space. In order
1934 * to keep fragmentation sane we only allocate kva in
1937 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;
1939 if (maxsize != bp->b_kvasize) {
1940 vm_offset_t addr = 0;
1945 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1946 vm_map_lock(&buffer_map);
1948 if (vm_map_findspace(&buffer_map,
1949 vm_map_min(&buffer_map), maxsize,
1952 * Uh oh. Buffer map is too fragmented. We
1953 * must defragment the map.
1955 vm_map_unlock(&buffer_map);
1956 vm_map_entry_release(count);
1959 bp->b_flags |= B_INVAL;
1964 vm_map_insert(&buffer_map, &count,
1966 addr, addr + maxsize,
1968 VM_PROT_ALL, VM_PROT_ALL,
1971 bp->b_kvabase = (caddr_t) addr;
1972 bp->b_kvasize = maxsize;
1973 bufspace += bp->b_kvasize;
1976 vm_map_unlock(&buffer_map);
1977 vm_map_entry_release(count);
1979 bp->b_data = bp->b_kvabase;
1987 * Buffer flushing daemon. Buffers are normally flushed by the
1988 * update daemon but if it cannot keep up this process starts to
1989 * take the load in an attempt to prevent getnewbuf() from blocking.
1991 * Once a flush is initiated it does not stop until the number
1992 * of buffers falls below lodirtybuffers, but we will wake up anyone
1993 * waiting at the mid-point.
1996 static struct thread *bufdaemon_td;
1997 static struct thread *bufdaemonhw_td;
1999 static struct kproc_desc buf_kp = {
2004 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2005 kproc_start, &buf_kp)
2007 static struct kproc_desc bufhw_kp = {
2012 SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
2013 kproc_start, &bufhw_kp)
2019 * This process needs to be suspended prior to shutdown sync.
2021 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2022 bufdaemon_td, SHUTDOWN_PRI_LAST);
2025 * This process is allowed to take the buffer cache to the limit
2030 kproc_suspend_loop();
2033 * Do the flush. Limit the amount of in-transit I/O we
2034 * allow to build up, otherwise we would completely saturate
2035 * the I/O system. Wakeup any waiting processes before we
2036 * normally would so they can run in parallel with our drain.
2038 while (numdirtybuffers > lodirtybuffers) {
2039 if (flushbufqueues(BQUEUE_DIRTY) == 0)
2041 waitrunningbufspace();
2044 if (runningbufcount + numdirtybuffers > lodirtybuffers) {
2045 waitrunningbufspace();
2050 * Only clear bd_request if we have reached our low water
2051 * mark. The buf_daemon normally waits 5 seconds and
2052 * then incrementally flushes any dirty buffers that have
2053 * built up, within reason.
2055 * If we were unable to hit our low water mark and couldn't
2056 * find any flushable buffers, we sleep half a second.
2057 * Otherwise we loop immediately.
2059 if (runningbufcount + numdirtybuffers <= lodirtybuffers) {
2061 * We reached our low water mark, reset the
2062 * request and sleep until we are needed again.
2063 * The sleep is just so the suspend code works.
2065 spin_lock_wr(&needsbuffer_spin);
2067 msleep(&bd_request, &needsbuffer_spin, 0,
2069 spin_unlock_wr(&needsbuffer_spin);
2072 * We couldn't find any flushable dirty buffers but
2073 * still have too many dirty buffers, we
2074 * have to sleep and try again. (rare)
2076 tsleep(&bd_request, 0, "qsleep", hz / 2);
2085 * This process needs to be suspended prior to shutdown sync.
2087 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
2088 bufdaemonhw_td, SHUTDOWN_PRI_LAST);
2091 * This process is allowed to take the buffer cache to the limit
2096 kproc_suspend_loop();
2099 * Do the flush. Limit the amount of in-transit I/O we
2100 * allow to build up, otherwise we would completely saturate
2101 * the I/O system. Wakeup any waiting processes before we
2102 * normally would so they can run in parallel with our drain.
2104 while (numdirtybuffershw > lodirtybuffers) {
2105 if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
2107 waitrunningbufspace();
2110 if (runningbufcount + numdirtybuffershw > lodirtybuffers) {
2111 waitrunningbufspace();
2115 * Only clear bd_request if we have reached our low water
2116 * mark. The buf_daemon normally waits 5 seconds and
2117 * then incrementally flushes any dirty buffers that have
2118 * built up, within reason.
2120 * If we were unable to hit our low water mark and couldn't
2121 * find any flushable buffers, we sleep half a second.
2122 * Otherwise we loop immediately.
2124 if (runningbufcount + numdirtybuffershw <= lodirtybuffers) {
2126 * We reached our low water mark, reset the
2127 * request and sleep until we are needed again.
2128 * The sleep is just so the suspend code works.
2130 spin_lock_wr(&needsbuffer_spin);
2132 msleep(&bd_request_hw, &needsbuffer_spin, 0,
2134 spin_unlock_wr(&needsbuffer_spin);
2137 * We couldn't find any flushable dirty buffers but
2138 * still have too many dirty buffers, we
2139 * have to sleep and try again. (rare)
2141 tsleep(&bd_request_hw, 0, "qsleep", hz / 2);
2149 * Try to flush a buffer in the dirty queue. We must be careful to
2150 * free up B_INVAL buffers instead of write them, which NFS is
2151 * particularly sensitive to.
2153 * B_RELBUF may only be set by VFSs. We do set B_AGE to indicate
2154 * that we really want to try to get the buffer out and reuse it
2155 * due to the write load on the machine.
2159 flushbufqueues(bufq_type_t q)
2164 bp = TAILQ_FIRST(&bufqueues[q]);
2167 KASSERT((bp->b_flags & B_DELWRI),
2168 ("unexpected clean buffer %p", bp));
2170 if (bp->b_flags & B_DELWRI) {
2171 if (bp->b_flags & B_INVAL) {
2172 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
2173 panic("flushbufqueues: locked buf");
2179 if (LIST_FIRST(&bp->b_dep) != NULL &&
2180 (bp->b_flags & B_DEFERRED) == 0 &&
2181 buf_countdeps(bp, 0)) {
2182 TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
2183 TAILQ_INSERT_TAIL(&bufqueues[q], bp,
2185 bp->b_flags |= B_DEFERRED;
2186 bp = TAILQ_FIRST(&bufqueues[q]);
2191 * Only write it out if we can successfully lock
2192 * it. If the buffer has a dependancy,
2193 * buf_checkwrite must also return 0.
2195 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
2196 if (LIST_FIRST(&bp->b_dep) != NULL &&
2197 buf_checkwrite(bp)) {
2201 bp->b_flags |= B_AGE;
2208 bp = TAILQ_NEXT(bp, b_freelist);
2216 * Returns true if no I/O is needed to access the associated VM object.
2217 * This is like findblk except it also hunts around in the VM system for
2220 * Note that we ignore vm_page_free() races from interrupts against our
2221 * lookup, since if the caller is not protected our return value will not
2222 * be any more valid then otherwise once we exit the critical section.
2225 inmem(struct vnode *vp, off_t loffset)
2228 vm_offset_t toff, tinc, size;
2231 if (findblk(vp, loffset))
2233 if (vp->v_mount == NULL)
2235 if ((obj = vp->v_object) == NULL)
2239 if (size > vp->v_mount->mnt_stat.f_iosize)
2240 size = vp->v_mount->mnt_stat.f_iosize;
2242 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
2243 m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
2247 if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
2248 tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
2249 if (vm_page_is_valid(m,
2250 (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
2259 * Sets the dirty range for a buffer based on the status of the dirty
2260 * bits in the pages comprising the buffer.
2262 * The range is limited to the size of the buffer.
2264 * This routine is primarily used by NFS, but is generalized for the
2268 vfs_setdirty(struct buf *bp)
2274 * Degenerate case - empty buffer
2277 if (bp->b_bufsize == 0)
2281 * We qualify the scan for modified pages on whether the
2282 * object has been flushed yet. The OBJ_WRITEABLE flag
2283 * is not cleared simply by protecting pages off.
2286 if ((bp->b_flags & B_VMIO) == 0)
2289 object = bp->b_xio.xio_pages[0]->object;
2291 if ((object->flags & OBJ_WRITEABLE) && !(object->flags & OBJ_MIGHTBEDIRTY))
2292 kprintf("Warning: object %p writeable but not mightbedirty\n", object);
2293 if (!(object->flags & OBJ_WRITEABLE) && (object->flags & OBJ_MIGHTBEDIRTY))
2294 kprintf("Warning: object %p mightbedirty but not writeable\n", object);
2296 if (object->flags & (OBJ_MIGHTBEDIRTY|OBJ_CLEANING)) {
2297 vm_offset_t boffset;
2298 vm_offset_t eoffset;
2301 * test the pages to see if they have been modified directly
2302 * by users through the VM system.
2304 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2305 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
2306 vm_page_test_dirty(bp->b_xio.xio_pages[i]);
2310 * Calculate the encompassing dirty range, boffset and eoffset,
2311 * (eoffset - boffset) bytes.
2314 for (i = 0; i < bp->b_xio.xio_npages; i++) {
2315 if (bp->b_xio.xio_pages[i]->dirty)
2318 boffset = (i << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2320 for (i = bp->b_xio.xio_npages - 1; i >= 0; --i) {
2321 if (bp->b_xio.xio_pages[i]->dirty) {
2325 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_loffset & PAGE_MASK);
2328 * Fit it to the buffer.
2331 if (eoffset > bp->b_bcount)
2332 eoffset = bp->b_bcount;
2335 * If we have a good dirty range, merge with the existing
2339 if (boffset < eoffset) {
2340 if (bp->b_dirtyoff > boffset)
2341 bp->b_dirtyoff = boffset;
2342 if (bp->b_dirtyend < eoffset)
2343 bp->b_dirtyend = eoffset;
2351 * Locate and return the specified buffer, or NULL if the buffer does
2352 * not exist. Do not attempt to lock the buffer or manipulate it in
2353 * any way. The caller must validate that the correct buffer has been
2354 * obtain after locking it.
2357 findblk(struct vnode *vp, off_t loffset)
2362 bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
2370 * Get a block given a specified block and offset into a file/device.
2371 * B_INVAL may or may not be set on return. The caller should clear
2372 * B_INVAL prior to initiating a READ.
2374 * IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
2375 * IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
2376 * OR SET B_INVAL BEFORE RETIRING IT. If you retire a getblk'd buffer
2377 * without doing any of those things the system will likely believe
2378 * the buffer to be valid (especially if it is not B_VMIO), and the
2379 * next getblk() will return the buffer with B_CACHE set.
2381 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
2382 * an existing buffer.
2384 * For a VMIO buffer, B_CACHE is modified according to the backing VM.
2385 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
2386 * and then cleared based on the backing VM. If the previous buffer is
2387 * non-0-sized but invalid, B_CACHE will be cleared.
2389 * If getblk() must create a new buffer, the new buffer is returned with
2390 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
2391 * case it is returned with B_INVAL clear and B_CACHE set based on the
2394 * getblk() also forces a bwrite() for any B_DELWRI buffer whos
2395 * B_CACHE bit is clear.
2397 * What this means, basically, is that the caller should use B_CACHE to
2398 * determine whether the buffer is fully valid or not and should clear
2399 * B_INVAL prior to issuing a read. If the caller intends to validate
2400 * the buffer by loading its data area with something, the caller needs
2401 * to clear B_INVAL. If the caller does this without issuing an I/O,
2402 * the caller should set B_CACHE ( as an optimization ), else the caller
2403 * should issue the I/O and biodone() will set B_CACHE if the I/O was
2404 * a write attempt or if it was a successfull read. If the caller
2405 * intends to issue a READ, the caller must clear B_INVAL and B_ERROR
2406 * prior to issuing the READ. biodone() will *not* clear B_INVAL.
2410 * GETBLK_PCATCH - catch signal if blocked, can cause NULL return
2411 * GETBLK_BHEAVY - heavy-weight buffer cache buffer
2414 getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
2417 int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
2420 if (size > MAXBSIZE)
2421 panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
2422 if (vp->v_object == NULL)
2423 panic("getblk: vnode %p has no object!", vp);
2427 if ((bp = findblk(vp, loffset))) {
2429 * The buffer was found in the cache, but we need to lock it.
2430 * Even with LK_NOWAIT the lockmgr may break our critical
2431 * section, so double-check the validity of the buffer
2432 * once the lock has been obtained.
2434 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
2435 int lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
2436 if (blkflags & GETBLK_PCATCH)
2437 lkflags |= LK_PCATCH;
2438 error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
2440 if (error == ENOLCK)
2448 * Once the buffer has been locked, make sure we didn't race
2449 * a buffer recyclement. Buffers that are no longer hashed
2450 * will have b_vp == NULL, so this takes care of that check
2453 if (bp->b_vp != vp || bp->b_loffset != loffset) {
2454 kprintf("Warning buffer %p (vp %p loffset %lld) was recycled\n", bp, vp, loffset);
2460 * All vnode-based buffers must be backed by a VM object.
2462 KKASSERT(bp->b_flags & B_VMIO);
2463 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2464 bp->b_flags &= ~B_AGE;
2467 * Make sure that B_INVAL buffers do not have a cached
2468 * block number translation.
2470 if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
2471 kprintf("Warning invalid buffer %p (vp %p loffset %lld) did not have cleared bio_offset cache\n", bp, vp, loffset);
2472 clearbiocache(&bp->b_bio2);
2476 * The buffer is locked. B_CACHE is cleared if the buffer is
2479 if (bp->b_flags & B_INVAL)
2480 bp->b_flags &= ~B_CACHE;
2484 * Any size inconsistancy with a dirty buffer or a buffer
2485 * with a softupdates dependancy must be resolved. Resizing
2486 * the buffer in such circumstances can lead to problems.
2488 if (size != bp->b_bcount) {
2489 if (bp->b_flags & B_DELWRI) {
2490 bp->b_flags |= B_NOCACHE;
2492 } else if (LIST_FIRST(&bp->b_dep)) {
2493 bp->b_flags |= B_NOCACHE;
2496 bp->b_flags |= B_RELBUF;
2501 KKASSERT(size <= bp->b_kvasize);
2502 KASSERT(bp->b_loffset != NOOFFSET,
2503 ("getblk: no buffer offset"));
2506 * A buffer with B_DELWRI set and B_CACHE clear must
2507 * be committed before we can return the buffer in
2508 * order to prevent the caller from issuing a read
2509 * ( due to B_CACHE not being set ) and overwriting
2512 * Most callers, including NFS and FFS, need this to
2513 * operate properly either because they assume they
2514 * can issue a read if B_CACHE is not set, or because
2515 * ( for example ) an uncached B_DELWRI might loop due
2516 * to softupdates re-dirtying the buffer. In the latter
2517 * case, B_CACHE is set after the first write completes,
2518 * preventing further loops.
2520 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE
2521 * above while extending the buffer, we cannot allow the
2522 * buffer to remain with B_CACHE set after the write
2523 * completes or it will represent a corrupt state. To
2524 * deal with this we set B_NOCACHE to scrap the buffer
2527 * We might be able to do something fancy, like setting
2528 * B_CACHE in bwrite() except if B_DELWRI is already set,
2529 * so the below call doesn't set B_CACHE, but that gets real
2530 * confusing. This is much easier.
2533 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
2534 bp->b_flags |= B_NOCACHE;
2541 * Buffer is not in-core, create new buffer. The buffer
2542 * returned by getnewbuf() is locked. Note that the returned
2543 * buffer is also considered valid (not marked B_INVAL).
2545 * Calculating the offset for the I/O requires figuring out
2546 * the block size. We use DEV_BSIZE for VBLK or VCHR and
2547 * the mount's f_iosize otherwise. If the vnode does not
2548 * have an associated mount we assume that the passed size is
2551 * Note that vn_isdisk() cannot be used here since it may
2552 * return a failure for numerous reasons. Note that the
2553 * buffer size may be larger then the block size (the caller
2554 * will use block numbers with the proper multiple). Beware
2555 * of using any v_* fields which are part of unions. In
2556 * particular, in DragonFly the mount point overloading
2557 * mechanism uses the namecache only and the underlying
2558 * directory vnode is not a special case.
2562 if (vp->v_type == VBLK || vp->v_type == VCHR)
2564 else if (vp->v_mount)
2565 bsize = vp->v_mount->mnt_stat.f_iosize;
2569 maxsize = size + (loffset & PAGE_MASK);
2570 maxsize = imax(maxsize, bsize);
2572 if ((bp = getnewbuf(blkflags, slptimeo, size, maxsize)) == NULL) {
2573 if (slpflags || slptimeo) {
2581 * This code is used to make sure that a buffer is not
2582 * created while the getnewbuf routine is blocked.
2583 * This can be a problem whether the vnode is locked or not.
2584 * If the buffer is created out from under us, we have to
2585 * throw away the one we just created. There is no window
2586 * race because we are safely running in a critical section
2587 * from the point of the duplicate buffer creation through
2588 * to here, and we've locked the buffer.
2590 if (findblk(vp, loffset)) {
2591 bp->b_flags |= B_INVAL;
2597 * Insert the buffer into the hash, so that it can
2598 * be found by findblk().
2600 * Make sure the translation layer has been cleared.
2602 bp->b_loffset = loffset;
2603 bp->b_bio2.bio_offset = NOOFFSET;
2604 /* bp->b_bio2.bio_next = NULL; */
2609 * All vnode-based buffers must be backed by a VM object.
2611 KKASSERT(vp->v_object != NULL);
2612 bp->b_flags |= B_VMIO;
2613 KKASSERT(bp->b_cmd == BUF_CMD_DONE);
2625 * Reacquire a buffer that was previously released to the locked queue,
2626 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
2627 * set B_LOCKED (which handles the acquisition race).
2629 * To this end, either B_LOCKED must be set or the dependancy list must be
2633 regetblk(struct buf *bp)
2635 KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
2636 BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
2645 * Get an empty, disassociated buffer of given size. The buffer is
2646 * initially set to B_INVAL.
2648 * critical section protection is not required for the allocbuf()
2649 * call because races are impossible here.
2657 maxsize = (size + BKVAMASK) & ~BKVAMASK;
2660 while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
2664 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */
2672 * This code constitutes the buffer memory from either anonymous system
2673 * memory (in the case of non-VMIO operations) or from an associated
2674 * VM object (in the case of VMIO operations). This code is able to
2675 * resize a buffer up or down.
2677 * Note that this code is tricky, and has many complications to resolve
2678 * deadlock or inconsistant data situations. Tread lightly!!!
2679 * There are B_CACHE and B_DELWRI interactions that must be dealt with by
2680 * the caller. Calling this code willy nilly can result in the loss of data.
2682 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with
2683 * B_CACHE for the non-VMIO case.
2685 * This routine does not need to be called from a critical section but you
2686 * must own the buffer.
2689 allocbuf(struct buf *bp, int size)
2691 int newbsize, mbsize;
2694 if (BUF_REFCNT(bp) == 0)
2695 panic("allocbuf: buffer not busy");
2697 if (bp->b_kvasize < size)
2698 panic("allocbuf: buffer too small");
2700 if ((bp->b_flags & B_VMIO) == 0) {
2704 * Just get anonymous memory from the kernel. Don't
2705 * mess with B_CACHE.
2707 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2708 if (bp->b_flags & B_MALLOC)
2711 newbsize = round_page(size);
2713 if (newbsize < bp->b_bufsize) {
2715 * Malloced buffers are not shrunk
2717 if (bp->b_flags & B_MALLOC) {
2719 bp->b_bcount = size;
2721 kfree(bp->b_data, M_BIOBUF);
2722 if (bp->b_bufsize) {
2723 bufmallocspace -= bp->b_bufsize;
2727 bp->b_data = bp->b_kvabase;
2729 bp->b_flags &= ~B_MALLOC;
2735 (vm_offset_t) bp->b_data + newbsize,
2736 (vm_offset_t) bp->b_data + bp->b_bufsize);
2737 } else if (newbsize > bp->b_bufsize) {
2739 * We only use malloced memory on the first allocation.
2740 * and revert to page-allocated memory when the buffer
2743 if ((bufmallocspace < maxbufmallocspace) &&
2744 (bp->b_bufsize == 0) &&
2745 (mbsize <= PAGE_SIZE/2)) {
2747 bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
2748 bp->b_bufsize = mbsize;
2749 bp->b_bcount = size;
2750 bp->b_flags |= B_MALLOC;
2751 bufmallocspace += mbsize;
2757 * If the buffer is growing on its other-than-first
2758 * allocation, then we revert to the page-allocation
2761 if (bp->b_flags & B_MALLOC) {
2762 origbuf = bp->b_data;
2763 origbufsize = bp->b_bufsize;
2764 bp->b_data = bp->b_kvabase;
2765 if (bp->b_bufsize) {
2766 bufmallocspace -= bp->b_bufsize;
2770 bp->b_flags &= ~B_MALLOC;
2771 newbsize = round_page(newbsize);
2775 (vm_offset_t) bp->b_data + bp->b_bufsize,
2776 (vm_offset_t) bp->b_data + newbsize);
2778 bcopy(origbuf, bp->b_data, origbufsize);
2779 kfree(origbuf, M_BIOBUF);
2786 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
2787 desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
2788 newbsize + PAGE_MASK) >> PAGE_SHIFT;
2789 KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);
2791 if (bp->b_flags & B_MALLOC)
2792 panic("allocbuf: VMIO buffer can't be malloced");
2794 * Set B_CACHE initially if buffer is 0 length or will become
2797 if (size == 0 || bp->b_bufsize == 0)
2798 bp->b_flags |= B_CACHE;
2800 if (newbsize < bp->b_bufsize) {
2802 * DEV_BSIZE aligned new buffer size is less then the
2803 * DEV_BSIZE aligned existing buffer size. Figure out
2804 * if we have to remove any pages.
2806 if (desiredpages < bp->b_xio.xio_npages) {
2807 for (i = desiredpages; i < bp->b_xio.xio_npages; i++) {
2809 * the page is not freed here -- it
2810 * is the responsibility of
2811 * vnode_pager_setsize
2813 m = bp->b_xio.xio_pages[i];
2814 KASSERT(m != bogus_page,
2815 ("allocbuf: bogus page found"));
2816 while (vm_page_sleep_busy(m, TRUE, "biodep"))
2819 bp->b_xio.xio_pages[i] = NULL;
2820 vm_page_unwire(m, 0);
2822 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) +
2823 (desiredpages << PAGE_SHIFT), (bp->b_xio.xio_npages - desiredpages));
2824 bp->b_xio.xio_npages = desiredpages;
2826 } else if (size > bp->b_bcount) {
2828 * We are growing the buffer, possibly in a
2829 * byte-granular fashion.
2837 * Step 1, bring in the VM pages from the object,
2838 * allocating them if necessary. We must clear
2839 * B_CACHE if these pages are not valid for the
2840 * range covered by the buffer.
2842 * critical section protection is required to protect
2843 * against interrupts unbusying and freeing pages
2844 * between our vm_page_lookup() and our
2845 * busycheck/wiring call.
2851 while (bp->b_xio.xio_npages < desiredpages) {
2855 pi = OFF_TO_IDX(bp->b_loffset) + bp->b_xio.xio_npages;
2856 if ((m = vm_page_lookup(obj, pi)) == NULL) {
2858 * note: must allocate system pages
2859 * since blocking here could intefere
2860 * with paging I/O, no matter which
2863 m = vm_page_alloc(obj, pi, VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
2866 vm_pageout_deficit += desiredpages -
2867 bp->b_xio.xio_npages;
2871 bp->b_flags &= ~B_CACHE;
2872 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2873 ++bp->b_xio.xio_npages;
2879 * We found a page. If we have to sleep on it,
2880 * retry because it might have gotten freed out
2883 * We can only test PG_BUSY here. Blocking on
2884 * m->busy might lead to a deadlock:
2886 * vm_fault->getpages->cluster_read->allocbuf
2890 if (vm_page_sleep_busy(m, FALSE, "pgtblk"))
2894 * We have a good page. Should we wakeup the
2897 if ((curthread != pagethread) &&
2898 ((m->queue - m->pc) == PQ_CACHE) &&
2899 ((vmstats.v_free_count + vmstats.v_cache_count) <
2900 (vmstats.v_free_min + vmstats.v_cache_min))) {
2901 pagedaemon_wakeup();
2903 vm_page_flag_clear(m, PG_ZERO);
2905 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
2906 ++bp->b_xio.xio_npages;
2911 * Step 2. We've loaded the pages into the buffer,
2912 * we have to figure out if we can still have B_CACHE
2913 * set. Note that B_CACHE is set according to the
2914 * byte-granular range ( bcount and size ), not the
2915 * aligned range ( newbsize ).
2917 * The VM test is against m->valid, which is DEV_BSIZE
2918 * aligned. Needless to say, the validity of the data
2919 * needs to also be DEV_BSIZE aligned. Note that this
2920 * fails with NFS if the server or some other client
2921 * extends the file's EOF. If our buffer is resized,
2922 * B_CACHE may remain set! XXX
2925 toff = bp->b_bcount;
2926 tinc = PAGE_SIZE - ((bp->b_loffset + toff) & PAGE_MASK);
2928 while ((bp->b_flags & B_CACHE) && toff < size) {
2931 if (tinc > (size - toff))
2934 pi = ((bp->b_loffset & PAGE_MASK) + toff) >>
2942 bp->b_xio.xio_pages[pi]
2949 * Step 3, fixup the KVM pmap. Remember that
2950 * bp->b_data is relative to bp->b_loffset, but
2951 * bp->b_loffset may be offset into the first page.
2954 bp->b_data = (caddr_t)
2955 trunc_page((vm_offset_t)bp->b_data);
2957 (vm_offset_t)bp->b_data,
2958 bp->b_xio.xio_pages,
2959 bp->b_xio.xio_npages
2961 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data |
2962 (vm_offset_t)(bp->b_loffset & PAGE_MASK));
2965 if (newbsize < bp->b_bufsize)
2967 bp->b_bufsize = newbsize; /* actual buffer allocation */
2968 bp->b_bcount = size; /* requested buffer size */
2975 * Wait for buffer I/O completion, returning error status. The buffer
2976 * is left locked on return. B_EINTR is converted into an EINTR error
2979 * NOTE! The original b_cmd is lost on return, since b_cmd will be
2980 * set to BUF_CMD_DONE.
2983 biowait(struct buf *bp)
2986 while (bp->b_cmd != BUF_CMD_DONE) {
2987 if (bp->b_cmd == BUF_CMD_READ)
2988 tsleep(bp, 0, "biord", 0);
2990 tsleep(bp, 0, "biowr", 0);
2993 if (bp->b_flags & B_EINTR) {
2994 bp->b_flags &= ~B_EINTR;
2997 if (bp->b_flags & B_ERROR) {
2998 return (bp->b_error ? bp->b_error : EIO);
3005 * This associates a tracking count with an I/O. vn_strategy() and
3006 * dev_dstrategy() do this automatically but there are a few cases
3007 * where a vnode or device layer is bypassed when a block translation
3008 * is cached. In such cases bio_start_transaction() may be called on
3009 * the bypassed layers so the system gets an I/O in progress indication
3010 * for those higher layers.
3013 bio_start_transaction(struct bio *bio, struct bio_track *track)
3015 bio->bio_track = track;
3016 atomic_add_int(&track->bk_active, 1);
3020 * Initiate I/O on a vnode.
3023 vn_strategy(struct vnode *vp, struct bio *bio)
3025 struct bio_track *track;
3027 KKASSERT(bio->bio_buf->b_cmd != BUF_CMD_DONE);
3028 if (bio->bio_buf->b_cmd == BUF_CMD_READ)
3029 track = &vp->v_track_read;
3031 track = &vp->v_track_write;
3032 bio->bio_track = track;
3033 atomic_add_int(&track->bk_active, 1);
3034 vop_strategy(*vp->v_ops, vp, bio);
3041 * Finish I/O on a buffer, optionally calling a completion function.
3042 * This is usually called from an interrupt so process blocking is
3045 * biodone is also responsible for setting B_CACHE in a B_VMIO bp.
3046 * In a non-VMIO bp, B_CACHE will be set on the next getblk()
3047 * assuming B_INVAL is clear.
3049 * For the VMIO case, we set B_CACHE if the op was a read and no
3050 * read error occured, or if the op was a write. B_CACHE is never
3051 * set if the buffer is invalid or otherwise uncacheable.
3053 * biodone does not mess with B_INVAL, allowing the I/O routine or the
3054 * initiator to leave B_INVAL set to brelse the buffer out of existance
3055 * in the biodone routine.
3058 biodone(struct bio *bio)
3060 struct buf *bp = bio->bio_buf;
3065 KASSERT(BUF_REFCNTNB(bp) > 0,
3066 ("biodone: bp %p not busy %d", bp, BUF_REFCNTNB(bp)));
3067 KASSERT(bp->b_cmd != BUF_CMD_DONE,
3068 ("biodone: bp %p already done!", bp));
3070 runningbufwakeup(bp);
3073 * Run up the chain of BIO's. Leave b_cmd intact for the duration.
3076 biodone_t *done_func;
3077 struct bio_track *track;
3080 * BIO tracking. Most but not all BIOs are tracked.
3082 if ((track = bio->bio_track) != NULL) {
3083 atomic_subtract_int(&track->bk_active, 1);
3084 if (track->bk_active < 0) {
3085 panic("biodone: bad active count bio %p\n",
3088 if (track->bk_waitflag) {
3089 track->bk_waitflag = 0;
3092 bio->bio_track = NULL;
3096 * A bio_done function terminates the loop. The function
3097 * will be responsible for any further chaining and/or
3098 * buffer management.
3100 * WARNING! The done function can deallocate the buffer!
3102 if ((done_func = bio->bio_done) != NULL) {
3103 bio->bio_done = NULL;
3108 bio = bio->bio_prev;
3112 bp->b_cmd = BUF_CMD_DONE;
3115 * Only reads and writes are processed past this point.
3117 if (cmd != BUF_CMD_READ && cmd != BUF_CMD_WRITE) {
3124 * Warning: softupdates may re-dirty the buffer.
3126 if (LIST_FIRST(&bp->b_dep) != NULL)
3129 if (bp->b_flags & B_VMIO) {
3135 struct vnode *vp = bp->b_vp;
3139 #if defined(VFS_BIO_DEBUG)
3140 if (vp->v_auxrefs == 0)
3141 panic("biodone: zero vnode hold count");
3142 if ((vp->v_flag & VOBJBUF) == 0)
3143 panic("biodone: vnode is not setup for merged cache");
3146 foff = bp->b_loffset;
3147 KASSERT(foff != NOOFFSET, ("biodone: no buffer offset"));
3148 KASSERT(obj != NULL, ("biodone: missing VM object"));
3150 #if defined(VFS_BIO_DEBUG)
3151 if (obj->paging_in_progress < bp->b_xio.xio_npages) {
3152 kprintf("biodone: paging in progress(%d) < bp->b_xio.xio_npages(%d)\n",
3153 obj->paging_in_progress, bp->b_xio.xio_npages);
3158 * Set B_CACHE if the op was a normal read and no error
3159 * occured. B_CACHE is set for writes in the b*write()
3162 iosize = bp->b_bcount - bp->b_resid;
3163 if (cmd == BUF_CMD_READ && (bp->b_flags & (B_INVAL|B_NOCACHE|B_ERROR)) == 0) {
3164 bp->b_flags |= B_CACHE;
3167 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3171 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff;
3176 * cleanup bogus pages, restoring the originals. Since
3177 * the originals should still be wired, we don't have
3178 * to worry about interrupt/freeing races destroying
3179 * the VM object association.
3181 m = bp->b_xio.xio_pages[i];
3182 if (m == bogus_page) {
3184 m = vm_page_lookup(obj, OFF_TO_IDX(foff));
3186 panic("biodone: page disappeared");
3187 bp->b_xio.xio_pages[i] = m;
3188 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3189 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3191 #if defined(VFS_BIO_DEBUG)
3192 if (OFF_TO_IDX(foff) != m->pindex) {
3194 "biodone: foff(%lu)/m->pindex(%d) mismatch\n",
3195 (unsigned long)foff, m->pindex);
3200 * In the write case, the valid and clean bits are
3201 * already changed correctly ( see bdwrite() ), so we
3202 * only need to do this here in the read case.
3204 if (cmd == BUF_CMD_READ && !bogusflag && resid > 0) {
3205 vfs_page_set_valid(bp, foff, i, m);
3207 vm_page_flag_clear(m, PG_ZERO);
3210 * when debugging new filesystems or buffer I/O methods, this
3211 * is the most common error that pops up. if you see this, you
3212 * have not set the page busy flag correctly!!!
3215 kprintf("biodone: page busy < 0, "
3216 "pindex: %d, foff: 0x(%x,%x), "
3217 "resid: %d, index: %d\n",
3218 (int) m->pindex, (int)(foff >> 32),
3219 (int) foff & 0xffffffff, resid, i);
3220 if (!vn_isdisk(vp, NULL))
3221 kprintf(" iosize: %ld, loffset: %lld, flags: 0x%08x, npages: %d\n",
3222 bp->b_vp->v_mount->mnt_stat.f_iosize,
3224 bp->b_flags, bp->b_xio.xio_npages);
3226 kprintf(" VDEV, loffset: %lld, flags: 0x%08x, npages: %d\n",
3228 bp->b_flags, bp->b_xio.xio_npages);
3229 kprintf(" valid: 0x%x, dirty: 0x%x, wired: %d\n",
3230 m->valid, m->dirty, m->wire_count);
3231 panic("biodone: page busy < 0");
3233 vm_page_io_finish(m);
3234 vm_object_pip_subtract(obj, 1);
3235 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3239 vm_object_pip_wakeupn(obj, 0);
3243 * For asynchronous completions, release the buffer now. The brelse
3244 * will do a wakeup there if necessary - so no need to do a wakeup
3245 * here in the async case. The sync case always needs to do a wakeup.
3248 if (bp->b_flags & B_ASYNC) {
3249 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0)
3262 * This routine is called in lieu of iodone in the case of
3263 * incomplete I/O. This keeps the busy status for pages
3267 vfs_unbusy_pages(struct buf *bp)
3271 runningbufwakeup(bp);
3272 if (bp->b_flags & B_VMIO) {
3273 struct vnode *vp = bp->b_vp;
3278 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3279 vm_page_t m = bp->b_xio.xio_pages[i];
3282 * When restoring bogus changes the original pages
3283 * should still be wired, so we are in no danger of
3284 * losing the object association and do not need
3285 * critical section protection particularly.
3287 if (m == bogus_page) {
3288 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_loffset) + i);
3290 panic("vfs_unbusy_pages: page missing");
3292 bp->b_xio.xio_pages[i] = m;
3293 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3294 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3296 vm_object_pip_subtract(obj, 1);
3297 vm_page_flag_clear(m, PG_ZERO);
3298 vm_page_io_finish(m);
3300 vm_object_pip_wakeupn(obj, 0);
3305 * vfs_page_set_valid:
3307 * Set the valid bits in a page based on the supplied offset. The
3308 * range is restricted to the buffer's size.
3310 * This routine is typically called after a read completes.
3313 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m)
3315 vm_ooffset_t soff, eoff;
3318 * Start and end offsets in buffer. eoff - soff may not cross a
3319 * page boundry or cross the end of the buffer. The end of the
3320 * buffer, in this case, is our file EOF, not the allocation size
3324 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3325 if (eoff > bp->b_loffset + bp->b_bcount)
3326 eoff = bp->b_loffset + bp->b_bcount;
3329 * Set valid range. This is typically the entire buffer and thus the
3333 vm_page_set_validclean(
3335 (vm_offset_t) (soff & PAGE_MASK),
3336 (vm_offset_t) (eoff - soff)
3344 * This routine is called before a device strategy routine.
3345 * It is used to tell the VM system that paging I/O is in
3346 * progress, and treat the pages associated with the buffer
3347 * almost as being PG_BUSY. Also the object 'paging_in_progress'
3348 * flag is handled to make sure that the object doesn't become
3351 * Since I/O has not been initiated yet, certain buffer flags
3352 * such as B_ERROR or B_INVAL may be in an inconsistant state
3353 * and should be ignored.
3356 vfs_busy_pages(struct vnode *vp, struct buf *bp)
3359 struct lwp *lp = curthread->td_lwp;
3362 * The buffer's I/O command must already be set. If reading,
3363 * B_CACHE must be 0 (double check against callers only doing
3364 * I/O when B_CACHE is 0).
3366 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3367 KKASSERT(bp->b_cmd == BUF_CMD_WRITE || (bp->b_flags & B_CACHE) == 0);
3369 if (bp->b_flags & B_VMIO) {
3374 foff = bp->b_loffset;
3375 KASSERT(bp->b_loffset != NOOFFSET,
3376 ("vfs_busy_pages: no buffer offset"));
3380 * Loop until none of the pages are busy.
3383 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3384 vm_page_t m = bp->b_xio.xio_pages[i];
3386 if (vm_page_sleep_busy(m, FALSE, "vbpage"))
3391 * Setup for I/O, soft-busy the page right now because
3392 * the next loop may block.
3394 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3395 vm_page_t m = bp->b_xio.xio_pages[i];
3397 vm_page_flag_clear(m, PG_ZERO);
3398 if ((bp->b_flags & B_CLUSTER) == 0) {
3399 vm_object_pip_add(obj, 1);
3400 vm_page_io_start(m);
3405 * Adjust protections for I/O and do bogus-page mapping.
3406 * Assume that vm_page_protect() can block (it can block
3407 * if VM_PROT_NONE, don't take any chances regardless).
3410 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3411 vm_page_t m = bp->b_xio.xio_pages[i];
3414 * When readying a vnode-backed buffer for a write
3415 * we must zero-fill any invalid portions of the
3418 * When readying a vnode-backed buffer for a read
3419 * we must replace any dirty pages with a bogus
3420 * page so we do not destroy dirty data when
3421 * filling in gaps. Dirty pages might not
3422 * necessarily be marked dirty yet, so use m->valid
3423 * as a reasonable test.
3425 * Bogus page replacement is, uh, bogus. We need
3426 * to find a better way.
3428 if (bp->b_cmd == BUF_CMD_WRITE) {
3429 vm_page_protect(m, VM_PROT_READ);
3430 vfs_page_set_valid(bp, foff, i, m);
3431 } else if (m->valid == VM_PAGE_BITS_ALL) {
3432 bp->b_xio.xio_pages[i] = bogus_page;
3435 vm_page_protect(m, VM_PROT_NONE);
3437 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3440 pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
3441 bp->b_xio.xio_pages, bp->b_xio.xio_npages);
3445 * This is the easiest place to put the process accounting for the I/O
3449 if (bp->b_cmd == BUF_CMD_READ)
3450 lp->lwp_ru.ru_inblock++;
3452 lp->lwp_ru.ru_oublock++;
3459 * Tell the VM system that the pages associated with this buffer
3460 * are clean. This is used for delayed writes where the data is
3461 * going to go to disk eventually without additional VM intevention.
3463 * Note that while we only really need to clean through to b_bcount, we
3464 * just go ahead and clean through to b_bufsize.
3467 vfs_clean_pages(struct buf *bp)
3471 if (bp->b_flags & B_VMIO) {
3474 foff = bp->b_loffset;
3475 KASSERT(foff != NOOFFSET, ("vfs_clean_pages: no buffer offset"));
3476 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3477 vm_page_t m = bp->b_xio.xio_pages[i];
3478 vm_ooffset_t noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
3479 vm_ooffset_t eoff = noff;
3481 if (eoff > bp->b_loffset + bp->b_bufsize)
3482 eoff = bp->b_loffset + bp->b_bufsize;
3483 vfs_page_set_valid(bp, foff, i, m);
3484 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */
3491 * vfs_bio_set_validclean:
3493 * Set the range within the buffer to valid and clean. The range is
3494 * relative to the beginning of the buffer, b_loffset. Note that
3495 * b_loffset itself may be offset from the beginning of the first page.
3499 vfs_bio_set_validclean(struct buf *bp, int base, int size)
3501 if (bp->b_flags & B_VMIO) {
3506 * Fixup base to be relative to beginning of first page.
3507 * Set initial n to be the maximum number of bytes in the
3508 * first page that can be validated.
3511 base += (bp->b_loffset & PAGE_MASK);
3512 n = PAGE_SIZE - (base & PAGE_MASK);
3514 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_xio.xio_npages; ++i) {
3515 vm_page_t m = bp->b_xio.xio_pages[i];
3520 vm_page_set_validclean(m, base & PAGE_MASK, n);
3531 * Clear a buffer. This routine essentially fakes an I/O, so we need
3532 * to clear B_ERROR and B_INVAL.
3534 * Note that while we only theoretically need to clear through b_bcount,
3535 * we go ahead and clear through b_bufsize.
3539 vfs_bio_clrbuf(struct buf *bp)
3543 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) {
3544 bp->b_flags &= ~(B_INVAL|B_ERROR);
3545 if ((bp->b_xio.xio_npages == 1) && (bp->b_bufsize < PAGE_SIZE) &&
3546 (bp->b_loffset & PAGE_MASK) == 0) {
3547 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1;
3548 if ((bp->b_xio.xio_pages[0]->valid & mask) == mask) {
3552 if (((bp->b_xio.xio_pages[0]->flags & PG_ZERO) == 0) &&
3553 ((bp->b_xio.xio_pages[0]->valid & mask) == 0)) {
3554 bzero(bp->b_data, bp->b_bufsize);
3555 bp->b_xio.xio_pages[0]->valid |= mask;
3560 ea = sa = bp->b_data;
3561 for(i=0;i<bp->b_xio.xio_npages;i++,sa=ea) {
3562 int j = ((vm_offset_t)sa & PAGE_MASK) / DEV_BSIZE;
3563 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE);
3564 ea = (caddr_t)(vm_offset_t)ulmin(
3565 (u_long)(vm_offset_t)ea,
3566 (u_long)(vm_offset_t)bp->b_data + bp->b_bufsize);
3567 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j;
3568 if ((bp->b_xio.xio_pages[i]->valid & mask) == mask)
3570 if ((bp->b_xio.xio_pages[i]->valid & mask) == 0) {
3571 if ((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) {
3575 for (; sa < ea; sa += DEV_BSIZE, j++) {
3576 if (((bp->b_xio.xio_pages[i]->flags & PG_ZERO) == 0) &&
3577 (bp->b_xio.xio_pages[i]->valid & (1<<j)) == 0)
3578 bzero(sa, DEV_BSIZE);
3581 bp->b_xio.xio_pages[i]->valid |= mask;
3582 vm_page_flag_clear(bp->b_xio.xio_pages[i], PG_ZERO);
3591 * vm_hold_load_pages:
3593 * Load pages into the buffer's address space. The pages are
3594 * allocated from the kernel object in order to reduce interference
3595 * with the any VM paging I/O activity. The range of loaded
3596 * pages will be wired.
3598 * If a page cannot be allocated, the 'pagedaemon' is woken up to
3599 * retrieve the full range (to - from) of pages.
3603 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3609 to = round_page(to);
3610 from = round_page(from);
3611 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3613 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3618 * Note: must allocate system pages since blocking here
3619 * could intefere with paging I/O, no matter which
3622 p = vm_page_alloc(&kernel_object,
3624 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM);
3626 vm_pageout_deficit += (to - from) >> PAGE_SHIFT;
3631 p->valid = VM_PAGE_BITS_ALL;
3632 vm_page_flag_clear(p, PG_ZERO);
3633 pmap_kenter(pg, VM_PAGE_TO_PHYS(p));
3634 bp->b_xio.xio_pages[index] = p;
3637 bp->b_xio.xio_npages = index;
3641 * vm_hold_free_pages:
3643 * Return pages associated with the buffer back to the VM system.
3645 * The range of pages underlying the buffer's address space will
3646 * be unmapped and un-wired.
3649 vm_hold_free_pages(struct buf *bp, vm_offset_t from, vm_offset_t to)
3653 int index, newnpages;
3655 from = round_page(from);
3656 to = round_page(to);
3657 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT;
3659 for (pg = from; pg < to; pg += PAGE_SIZE, index++) {
3660 p = bp->b_xio.xio_pages[index];
3661 if (p && (index < bp->b_xio.xio_npages)) {
3663 kprintf("vm_hold_free_pages: doffset: %lld, loffset: %lld\n",
3664 bp->b_bio2.bio_offset, bp->b_loffset);
3666 bp->b_xio.xio_pages[index] = NULL;
3669 vm_page_unwire(p, 0);
3673 bp->b_xio.xio_npages = newnpages;
3679 * Map a user buffer into KVM via a pbuf. On return the buffer's
3680 * b_data, b_bufsize, and b_bcount will be set, and its XIO page array
3684 vmapbuf(struct buf *bp, caddr_t udata, int bytes)
3695 * bp had better have a command and it better be a pbuf.
3697 KKASSERT(bp->b_cmd != BUF_CMD_DONE);
3698 KKASSERT(bp->b_flags & B_PAGING);
3704 * Map the user data into KVM. Mappings have to be page-aligned.
3706 addr = (caddr_t)trunc_page((vm_offset_t)udata);
3709 vmprot = VM_PROT_READ;
3710 if (bp->b_cmd == BUF_CMD_READ)
3711 vmprot |= VM_PROT_WRITE;
3713 while (addr < udata + bytes) {
3715 * Do the vm_fault if needed; do the copy-on-write thing
3716 * when reading stuff off device into memory.
3718 * vm_fault_page*() returns a held VM page.
3720 va = (addr >= udata) ? (vm_offset_t)addr : (vm_offset_t)udata;
3721 va = trunc_page(va);
3723 m = vm_fault_page_quick(va, vmprot, &error);
3725 for (i = 0; i < pidx; ++i) {
3726 vm_page_unhold(bp->b_xio.xio_pages[i]);
3727 bp->b_xio.xio_pages[i] = NULL;
3731 bp->b_xio.xio_pages[pidx] = m;
3737 * Map the page array and set the buffer fields to point to
3738 * the mapped data buffer.
3740 if (pidx > btoc(MAXPHYS))
3741 panic("vmapbuf: mapped more than MAXPHYS");
3742 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_xio.xio_pages, pidx);
3744 bp->b_xio.xio_npages = pidx;
3745 bp->b_data = bp->b_kvabase + ((int)(intptr_t)udata & PAGE_MASK);
3746 bp->b_bcount = bytes;
3747 bp->b_bufsize = bytes;
3754 * Free the io map PTEs associated with this IO operation.
3755 * We also invalidate the TLB entries and restore the original b_addr.
3758 vunmapbuf(struct buf *bp)
3763 KKASSERT(bp->b_flags & B_PAGING);
3765 npages = bp->b_xio.xio_npages;
3766 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages);
3767 for (pidx = 0; pidx < npages; ++pidx) {
3768 vm_page_unhold(bp->b_xio.xio_pages[pidx]);
3769 bp->b_xio.xio_pages[pidx] = NULL;
3771 bp->b_xio.xio_npages = 0;
3772 bp->b_data = bp->b_kvabase;
3776 * Scan all buffers in the system and issue the callback.
3779 scan_all_buffers(int (*callback)(struct buf *, void *), void *info)
3785 for (n = 0; n < nbuf; ++n) {
3786 if ((error = callback(&buf[n], info)) < 0) {
3796 * print out statistics from the current status of the buffer pool
3797 * this can be toggeled by the system control option debug.syncprt
3806 int counts[(MAXBSIZE / PAGE_SIZE) + 1];
3807 static char *bname[3] = { "LOCKED", "LRU", "AGE" };
3809 for (dp = bufqueues, i = 0; dp < &bufqueues[3]; dp++, i++) {
3811 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
3814 TAILQ_FOREACH(bp, dp, b_freelist) {
3815 counts[bp->b_bufsize/PAGE_SIZE]++;
3819 kprintf("%s: total-%d", bname[i], count);
3820 for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
3822 kprintf(", %d-%d", j * PAGE_SIZE, counts[j]);
3830 DB_SHOW_COMMAND(buffer, db_show_buffer)
3833 struct buf *bp = (struct buf *)addr;
3836 db_printf("usage: show buffer <addr>\n");
3840 db_printf("b_flags = 0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS);
3841 db_printf("b_cmd = %d\n", bp->b_cmd);
3842 db_printf("b_error = %d, b_bufsize = %d, b_bcount = %d, "
3843 "b_resid = %d\n, b_data = %p, "
3844 "bio_offset(disk) = %lld, bio_offset(phys) = %lld\n",
3845 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid,
3846 bp->b_data, bp->b_bio2.bio_offset, (bp->b_bio2.bio_next ? bp->b_bio2.bio_next->bio_offset : (off_t)-1));
3847 if (bp->b_xio.xio_npages) {
3849 db_printf("b_xio.xio_npages = %d, pages(OBJ, IDX, PA): ",
3850 bp->b_xio.xio_npages);
3851 for (i = 0; i < bp->b_xio.xio_npages; i++) {
3853 m = bp->b_xio.xio_pages[i];
3854 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object,
3855 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m));
3856 if ((i + 1) < bp->b_xio.xio_npages)