2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * $DragonFly: src/sys/vfs/hammer/hammer_io.c,v 1.55 2008/09/15 17:02:49 dillon Exp $
37 * IO Primitives and buffer cache management
39 * All major data-tracking structures in HAMMER contain a struct hammer_io
40 * which is used to manage their backing store. We use filesystem buffers
41 * for backing store and we leave them passively associated with their
44 * If the kernel tries to destroy a passively associated buf which we cannot
45 * yet let go we set B_LOCKED in the buffer and then actively released it
50 #include <sys/fcntl.h>
51 #include <sys/nlookup.h>
55 static void hammer_io_modify(hammer_io_t io
, int count
);
56 static void hammer_io_deallocate(struct buf
*bp
);
58 static void hammer_io_direct_read_complete(struct bio
*nbio
);
60 static void hammer_io_direct_write_complete(struct bio
*nbio
);
61 static int hammer_io_direct_uncache_callback(hammer_inode_t ip
, void *data
);
62 static void hammer_io_set_modlist(struct hammer_io
*io
);
63 static void hammer_io_flush_mark(hammer_volume_t volume
);
67 * Initialize a new, already-zero'd hammer_io structure, or reinitialize
68 * an existing hammer_io structure which may have switched to another type.
71 hammer_io_init(hammer_io_t io
, hammer_volume_t volume
, enum hammer_io_type type
)
74 io
->hmp
= volume
->io
.hmp
;
79 * Helper routine to disassociate a buffer cache buffer from an I/O
80 * structure. The buffer is unlocked and marked appropriate for reclamation.
82 * The io may have 0 or 1 references depending on who called us. The
83 * caller is responsible for dealing with the refs.
85 * This call can only be made when no action is required on the buffer.
87 * The caller must own the buffer and the IO must indicate that the
88 * structure no longer owns it (io.released != 0).
91 hammer_io_disassociate(hammer_io_structure_t iou
)
93 struct buf
*bp
= iou
->io
.bp
;
95 KKASSERT(iou
->io
.released
);
96 KKASSERT(iou
->io
.modified
== 0);
97 KKASSERT(LIST_FIRST(&bp
->b_dep
) == (void *)iou
);
102 * If the buffer was locked someone wanted to get rid of it.
104 if (bp
->b_flags
& B_LOCKED
) {
105 --hammer_count_io_locked
;
106 bp
->b_flags
&= ~B_LOCKED
;
108 if (iou
->io
.reclaim
) {
109 bp
->b_flags
|= B_NOCACHE
|B_RELBUF
;
113 switch(iou
->io
.type
) {
114 case HAMMER_STRUCTURE_VOLUME
:
115 iou
->volume
.ondisk
= NULL
;
117 case HAMMER_STRUCTURE_DATA_BUFFER
:
118 case HAMMER_STRUCTURE_META_BUFFER
:
119 case HAMMER_STRUCTURE_UNDO_BUFFER
:
120 iou
->buffer
.ondisk
= NULL
;
122 case HAMMER_STRUCTURE_DUMMY
:
123 panic("hammer_io_disassociate: bad io type");
129 * Wait for any physical IO to complete
131 * XXX we aren't interlocked against a spinlock or anything so there
132 * is a small window in the interlock / io->running == 0 test.
135 hammer_io_wait(hammer_io_t io
)
140 tsleep_interlock(io
, 0);
141 if (io
->running
== 0)
143 tsleep(io
, PINTERLOCKED
, "hmrflw", hz
);
144 if (io
->running
== 0)
151 * Wait for all currently queued HAMMER-initiated I/Os to complete.
153 * This is not supposed to count direct I/O's but some can leak
154 * through (for non-full-sized direct I/Os).
157 hammer_io_wait_all(hammer_mount_t hmp
, const char *ident
, int doflush
)
159 struct hammer_io iodummy
;
163 * Degenerate case, no I/O is running
166 if (TAILQ_EMPTY(&hmp
->iorun_list
)) {
169 hammer_io_flush_sync(hmp
);
172 bzero(&iodummy
, sizeof(iodummy
));
173 iodummy
.type
= HAMMER_STRUCTURE_DUMMY
;
176 * Add placemarker and then wait until it becomes the head of
179 TAILQ_INSERT_TAIL(&hmp
->iorun_list
, &iodummy
, iorun_entry
);
180 while (TAILQ_FIRST(&hmp
->iorun_list
) != &iodummy
) {
181 tsleep(&iodummy
, 0, ident
, 0);
185 * Chain in case several placemarkers are present.
187 TAILQ_REMOVE(&hmp
->iorun_list
, &iodummy
, iorun_entry
);
188 io
= TAILQ_FIRST(&hmp
->iorun_list
);
189 if (io
&& io
->type
== HAMMER_STRUCTURE_DUMMY
)
194 hammer_io_flush_sync(hmp
);
198 * Clear a flagged error condition on a I/O buffer. The caller must hold
199 * its own ref on the buffer.
202 hammer_io_clear_error(struct hammer_io
*io
)
206 hammer_unref(&io
->lock
);
207 KKASSERT(io
->lock
.refs
> 0);
212 * This is an advisory function only which tells the buffer cache
213 * the bp is not a meta-data buffer, even though it is backed by
216 * This is used by HAMMER's reblocking code to avoid trying to
217 * swapcache the filesystem's data when it is read or written
218 * by the reblocking code.
221 hammer_io_notmeta(hammer_buffer_t buffer
)
223 buffer
->io
.bp
->b_flags
|= B_NOTMETA
;
227 #define HAMMER_MAXRA 4
230 * Load bp for a HAMMER structure. The io must be exclusively locked by
233 * This routine is mostly used on meta-data and small-data blocks. Generally
234 * speaking HAMMER assumes some locality of reference and will cluster
237 * Note that clustering occurs at the device layer, not the logical layer.
238 * If the buffers do not apply to the current operation they may apply to
242 hammer_io_read(struct vnode
*devvp
, struct hammer_io
*io
, hammer_off_t limit
)
247 if ((bp
= io
->bp
) == NULL
) {
248 hammer_count_io_running_read
+= io
->bytes
;
249 if (hammer_cluster_enable
) {
250 error
= cluster_read(devvp
, limit
,
251 io
->offset
, io
->bytes
,
253 HAMMER_CLUSTER_BUFS
, &io
->bp
);
255 error
= bread(devvp
, io
->offset
, io
->bytes
, &io
->bp
);
257 hammer_stats_disk_read
+= io
->bytes
;
258 hammer_count_io_running_read
-= io
->bytes
;
261 * The code generally assumes b_ops/b_dep has been set-up,
262 * even if we error out here.
265 bp
->b_ops
= &hammer_bioops
;
266 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
267 LIST_INSERT_HEAD(&bp
->b_dep
, &io
->worklist
, node
);
269 KKASSERT(io
->modified
== 0);
270 KKASSERT(io
->running
== 0);
271 KKASSERT(io
->waiting
== 0);
272 io
->released
= 0; /* we hold an active lock on bp */
280 * Similar to hammer_io_read() but returns a zero'd out buffer instead.
281 * Must be called with the IO exclusively locked.
283 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background
284 * I/O by forcing the buffer to not be in a released state before calling
287 * This function will also mark the IO as modified but it will not
288 * increment the modify_refs count.
291 hammer_io_new(struct vnode
*devvp
, struct hammer_io
*io
)
295 if ((bp
= io
->bp
) == NULL
) {
296 io
->bp
= getblk(devvp
, io
->offset
, io
->bytes
, 0, 0);
298 bp
->b_ops
= &hammer_bioops
;
299 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
300 LIST_INSERT_HEAD(&bp
->b_dep
, &io
->worklist
, node
);
302 KKASSERT(io
->running
== 0);
312 hammer_io_modify(io
, 0);
318 * Advance the activity count on the underlying buffer because
319 * HAMMER does not getblk/brelse on every access.
322 hammer_io_advance(struct hammer_io
*io
)
325 buf_act_advance(io
->bp
);
329 * Remove potential device level aliases against buffers managed by high level
330 * vnodes. Aliases can also be created due to mixed buffer sizes or via
331 * direct access to the backing store device.
333 * This is nasty because the buffers are also VMIO-backed. Even if a buffer
334 * does not exist its backing VM pages might, and we have to invalidate
335 * those as well or a getblk() will reinstate them.
337 * Buffer cache buffers associated with hammer_buffers cannot be
341 hammer_io_inval(hammer_volume_t volume
, hammer_off_t zone2_offset
)
343 hammer_io_structure_t iou
;
344 hammer_off_t phys_offset
;
348 phys_offset
= volume
->ondisk
->vol_buf_beg
+
349 (zone2_offset
& HAMMER_OFF_SHORT_MASK
);
351 if ((bp
= findblk(volume
->devvp
, phys_offset
, FINDBLK_TEST
)) != NULL
)
352 bp
= getblk(volume
->devvp
, phys_offset
, bp
->b_bufsize
, 0, 0);
354 bp
= getblk(volume
->devvp
, phys_offset
, HAMMER_BUFSIZE
, 0, 0);
355 if ((iou
= (void *)LIST_FIRST(&bp
->b_dep
)) != NULL
) {
357 hammer_ref(&iou
->io
.lock
);
358 hammer_io_clear_modify(&iou
->io
, 1);
360 iou
->io
.released
= 0;
364 KKASSERT(iou
->io
.lock
.refs
== 1);
365 hammer_rel_buffer(&iou
->buffer
, 0);
366 /*hammer_io_deallocate(bp);*/
371 KKASSERT((bp
->b_flags
& B_LOCKED
) == 0);
373 bp
->b_flags
|= B_NOCACHE
|B_RELBUF
;
382 * This routine is called on the last reference to a hammer structure.
383 * The io is usually interlocked with io.loading and io.refs must be 1.
385 * This routine may return a non-NULL bp to the caller for dispoal. Disposal
386 * simply means the caller finishes decrementing the ref-count on the
387 * IO structure then brelse()'s the bp. The bp may or may not still be
388 * passively associated with the IO.
390 * The only requirement here is that modified meta-data and volume-header
391 * buffer may NOT be disassociated from the IO structure, and consequently
392 * we also leave such buffers actively associated with the IO if they already
393 * are (since the kernel can't do anything with them anyway). Only the
394 * flusher is allowed to write such buffers out. Modified pure-data and
395 * undo buffers are returned to the kernel but left passively associated
396 * so we can track when the kernel writes the bp out.
399 hammer_io_release(struct hammer_io
*io
, int flush
)
401 union hammer_io_structure
*iou
= (void *)io
;
404 if ((bp
= io
->bp
) == NULL
)
408 * Try to flush a dirty IO to disk if asked to by the
409 * caller or if the kernel tried to flush the buffer in the past.
411 * Kernel-initiated flushes are only allowed for pure-data buffers.
412 * meta-data and volume buffers can only be flushed explicitly
417 hammer_io_flush(io
, 0);
418 } else if (bp
->b_flags
& B_LOCKED
) {
420 case HAMMER_STRUCTURE_DATA_BUFFER
:
421 hammer_io_flush(io
, 0);
423 case HAMMER_STRUCTURE_UNDO_BUFFER
:
424 hammer_io_flush(io
, hammer_undo_reclaim(io
));
429 } /* else no explicit request to flush the buffer */
433 * Wait for the IO to complete if asked to. This occurs when
434 * the buffer must be disposed of definitively during an umount
435 * or buffer invalidation.
437 if (io
->waitdep
&& io
->running
) {
442 * Return control of the buffer to the kernel (with the provisio
443 * that our bioops can override kernel decisions with regards to
446 if ((flush
|| io
->reclaim
) && io
->modified
== 0 && io
->running
== 0) {
448 * Always disassociate the bp if an explicit flush
449 * was requested and the IO completed with no error
450 * (so unmount can really clean up the structure).
458 hammer_io_disassociate((hammer_io_structure_t
)io
);
460 } else if (io
->modified
) {
462 * Only certain IO types can be released to the kernel if
463 * the buffer has been modified.
465 * volume and meta-data IO types may only be explicitly
469 case HAMMER_STRUCTURE_DATA_BUFFER
:
470 case HAMMER_STRUCTURE_UNDO_BUFFER
:
471 if (io
->released
== 0) {
479 bp
= NULL
; /* bp left associated */
480 } else if (io
->released
== 0) {
482 * Clean buffers can be generally released to the kernel.
483 * We leave the bp passively associated with the HAMMER
484 * structure and use bioops to disconnect it later on
485 * if the kernel wants to discard the buffer.
487 * We can steal the structure's ownership of the bp.
490 if (bp
->b_flags
& B_LOCKED
) {
491 hammer_io_disassociate(iou
);
495 hammer_io_disassociate(iou
);
498 /* return the bp (bp passively associated) */
503 * A released buffer is passively associate with our
504 * hammer_io structure. The kernel cannot destroy it
505 * without making a bioops call. If the kernel (B_LOCKED)
506 * or we (reclaim) requested that the buffer be destroyed
507 * we destroy it, otherwise we do a quick get/release to
508 * reset its position in the kernel's LRU list.
510 * Leaving the buffer passively associated allows us to
511 * use the kernel's LRU buffer flushing mechanisms rather
512 * then rolling our own.
514 * XXX there are two ways of doing this. We can re-acquire
515 * and passively release to reset the LRU, or not.
517 if (io
->running
== 0) {
519 if ((bp
->b_flags
& B_LOCKED
) || io
->reclaim
) {
520 hammer_io_disassociate(iou
);
523 /* return the bp (bp passively associated) */
527 * bp is left passively associated but we do not
528 * try to reacquire it. Interactions with the io
529 * structure will occur on completion of the bp's
539 * This routine is called with a locked IO when a flush is desired and
540 * no other references to the structure exists other then ours. This
541 * routine is ONLY called when HAMMER believes it is safe to flush a
542 * potentially modified buffer out.
545 hammer_io_flush(struct hammer_io
*io
, int reclaim
)
550 * Degenerate case - nothing to flush if nothing is dirty.
552 if (io
->modified
== 0) {
557 KKASSERT(io
->modify_refs
<= 0);
560 * Acquire ownership of the bp, particularly before we clear our
563 * We are going to bawrite() this bp. Don't leave a window where
564 * io->released is set, we actually own the bp rather then our
570 /* BUF_KERNPROC(io->bp); */
571 /* io->released = 0; */
572 KKASSERT(io
->released
);
573 KKASSERT(io
->bp
== bp
);
579 if ((bp
->b_flags
& B_LOCKED
) == 0) {
580 bp
->b_flags
|= B_LOCKED
;
581 ++hammer_count_io_locked
;
586 * Acquire exclusive access to the bp and then clear the modified
587 * state of the buffer prior to issuing I/O to interlock any
588 * modifications made while the I/O is in progress. This shouldn't
589 * happen anyway but losing data would be worse. The modified bit
590 * will be rechecked after the IO completes.
592 * NOTE: This call also finalizes the buffer's content (inval == 0).
594 * This is only legal when lock.refs == 1 (otherwise we might clear
595 * the modified bit while there are still users of the cluster
596 * modifying the data).
598 * Do this before potentially blocking so any attempt to modify the
599 * ondisk while we are blocked blocks waiting for us.
601 hammer_ref(&io
->lock
);
602 hammer_io_clear_modify(io
, 0);
603 hammer_unref(&io
->lock
);
605 if (hammer_debug_io
& 0x0002)
606 kprintf("hammer io_write %016jx\n", bp
->b_bio1
.bio_offset
);
609 * Transfer ownership to the kernel and initiate I/O.
612 io
->hmp
->io_running_space
+= io
->bytes
;
613 TAILQ_INSERT_TAIL(&io
->hmp
->iorun_list
, io
, iorun_entry
);
614 hammer_count_io_running_write
+= io
->bytes
;
616 hammer_io_flush_mark(io
->volume
);
619 /************************************************************************
621 ************************************************************************
623 * These routines deal with dependancies created when IO buffers get
624 * modified. The caller must call hammer_modify_*() on a referenced
625 * HAMMER structure prior to modifying its on-disk data.
627 * Any intent to modify an IO buffer acquires the related bp and imposes
628 * various write ordering dependancies.
632 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
633 * are locked until the flusher can deal with them, pure data buffers
634 * can be written out.
638 hammer_io_modify(hammer_io_t io
, int count
)
641 * io->modify_refs must be >= 0
643 while (io
->modify_refs
< 0) {
645 tsleep(io
, 0, "hmrmod", 0);
649 * Shortcut if nothing to do.
651 KKASSERT(io
->lock
.refs
!= 0 && io
->bp
!= NULL
);
652 io
->modify_refs
+= count
;
653 if (io
->modified
&& io
->released
== 0)
656 hammer_lock_ex(&io
->lock
);
657 if (io
->modified
== 0) {
658 hammer_io_set_modlist(io
);
663 BUF_KERNPROC(io
->bp
);
665 KKASSERT(io
->modified
!= 0);
667 hammer_unlock(&io
->lock
);
672 hammer_io_modify_done(hammer_io_t io
)
674 KKASSERT(io
->modify_refs
> 0);
676 if (io
->modify_refs
== 0 && io
->waitmod
) {
683 hammer_io_write_interlock(hammer_io_t io
)
685 while (io
->modify_refs
!= 0) {
687 tsleep(io
, 0, "hmrmod", 0);
689 io
->modify_refs
= -1;
693 hammer_io_done_interlock(hammer_io_t io
)
695 KKASSERT(io
->modify_refs
== -1);
704 * Caller intends to modify a volume's ondisk structure.
706 * This is only allowed if we are the flusher or we have a ref on the
710 hammer_modify_volume(hammer_transaction_t trans
, hammer_volume_t volume
,
713 KKASSERT (trans
== NULL
|| trans
->sync_lock_refs
> 0);
715 hammer_io_modify(&volume
->io
, 1);
717 intptr_t rel_offset
= (intptr_t)base
- (intptr_t)volume
->ondisk
;
718 KKASSERT((rel_offset
& ~(intptr_t)HAMMER_BUFMASK
) == 0);
719 hammer_generate_undo(trans
,
720 HAMMER_ENCODE_RAW_VOLUME(volume
->vol_no
, rel_offset
),
726 * Caller intends to modify a buffer's ondisk structure.
728 * This is only allowed if we are the flusher or we have a ref on the
732 hammer_modify_buffer(hammer_transaction_t trans
, hammer_buffer_t buffer
,
735 KKASSERT (trans
== NULL
|| trans
->sync_lock_refs
> 0);
737 hammer_io_modify(&buffer
->io
, 1);
739 intptr_t rel_offset
= (intptr_t)base
- (intptr_t)buffer
->ondisk
;
740 KKASSERT((rel_offset
& ~(intptr_t)HAMMER_BUFMASK
) == 0);
741 hammer_generate_undo(trans
,
742 buffer
->zone2_offset
+ rel_offset
,
748 hammer_modify_volume_done(hammer_volume_t volume
)
750 hammer_io_modify_done(&volume
->io
);
754 hammer_modify_buffer_done(hammer_buffer_t buffer
)
756 hammer_io_modify_done(&buffer
->io
);
760 * Mark an entity as not being dirty any more and finalize any
761 * delayed adjustments to the buffer.
763 * Delayed adjustments are an important performance enhancement, allowing
764 * us to avoid recalculating B-Tree node CRCs over and over again when
765 * making bulk-modifications to the B-Tree.
767 * If inval is non-zero delayed adjustments are ignored.
769 * This routine may dereference related btree nodes and cause the
770 * buffer to be dereferenced. The caller must own a reference on io.
773 hammer_io_clear_modify(struct hammer_io
*io
, int inval
)
775 if (io
->modified
== 0)
779 * Take us off the mod-list and clear the modified bit.
781 KKASSERT(io
->mod_list
!= NULL
);
782 if (io
->mod_list
== &io
->hmp
->volu_list
||
783 io
->mod_list
== &io
->hmp
->meta_list
) {
784 io
->hmp
->locked_dirty_space
-= io
->bytes
;
785 hammer_count_dirtybufspace
-= io
->bytes
;
787 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
792 * If this bit is not set there are no delayed adjustments.
799 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference
800 * on the node (& underlying buffer). Release the node after clearing
803 if (io
->type
== HAMMER_STRUCTURE_META_BUFFER
) {
804 hammer_buffer_t buffer
= (void *)io
;
808 TAILQ_FOREACH(node
, &buffer
->clist
, entry
) {
809 if ((node
->flags
& HAMMER_NODE_NEEDSCRC
) == 0)
811 node
->flags
&= ~HAMMER_NODE_NEEDSCRC
;
812 KKASSERT(node
->ondisk
);
814 node
->ondisk
->crc
= crc32(&node
->ondisk
->crc
+ 1, HAMMER_BTREE_CRCSIZE
);
815 hammer_rel_node(node
);
819 /* caller must still have ref on io */
820 KKASSERT(io
->lock
.refs
> 0);
824 * Clear the IO's modify list. Even though the IO is no longer modified
825 * it may still be on the lose_list. This routine is called just before
826 * the governing hammer_buffer is destroyed.
829 hammer_io_clear_modlist(struct hammer_io
*io
)
831 KKASSERT(io
->modified
== 0);
833 crit_enter(); /* biodone race against list */
834 KKASSERT(io
->mod_list
== &io
->hmp
->lose_list
);
835 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
842 hammer_io_set_modlist(struct hammer_io
*io
)
844 struct hammer_mount
*hmp
= io
->hmp
;
846 KKASSERT(io
->mod_list
== NULL
);
849 case HAMMER_STRUCTURE_VOLUME
:
850 io
->mod_list
= &hmp
->volu_list
;
851 hmp
->locked_dirty_space
+= io
->bytes
;
852 hammer_count_dirtybufspace
+= io
->bytes
;
854 case HAMMER_STRUCTURE_META_BUFFER
:
855 io
->mod_list
= &hmp
->meta_list
;
856 hmp
->locked_dirty_space
+= io
->bytes
;
857 hammer_count_dirtybufspace
+= io
->bytes
;
859 case HAMMER_STRUCTURE_UNDO_BUFFER
:
860 io
->mod_list
= &hmp
->undo_list
;
862 case HAMMER_STRUCTURE_DATA_BUFFER
:
863 io
->mod_list
= &hmp
->data_list
;
865 case HAMMER_STRUCTURE_DUMMY
:
866 panic("hammer_io_disassociate: bad io type");
869 TAILQ_INSERT_TAIL(io
->mod_list
, io
, mod_entry
);
872 /************************************************************************
874 ************************************************************************
879 * Pre-IO initiation kernel callback - cluster build only
882 hammer_io_start(struct buf
*bp
)
887 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT!
889 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
890 * may also be set if we were marking a cluster header open. Only remove
891 * our dependancy if the modified bit is clear.
894 hammer_io_complete(struct buf
*bp
)
896 union hammer_io_structure
*iou
= (void *)LIST_FIRST(&bp
->b_dep
);
897 struct hammer_io
*ionext
;
899 KKASSERT(iou
->io
.released
== 1);
902 * Deal with people waiting for I/O to drain
904 if (iou
->io
.running
) {
906 * Deal with critical write errors. Once a critical error
907 * has been flagged in hmp the UNDO FIFO will not be updated.
908 * That way crash recover will give us a consistent
911 * Because of this we can throw away failed UNDO buffers. If
912 * we throw away META or DATA buffers we risk corrupting
913 * the now read-only version of the filesystem visible to
914 * the user. Clear B_ERROR so the buffer is not re-dirtied
915 * by the kernel and ref the io so it doesn't get thrown
918 if (bp
->b_flags
& B_ERROR
) {
919 hammer_critical_error(iou
->io
.hmp
, NULL
, bp
->b_error
,
920 "while flushing meta-data");
921 switch(iou
->io
.type
) {
922 case HAMMER_STRUCTURE_UNDO_BUFFER
:
925 if (iou
->io
.ioerror
== 0) {
927 if (iou
->io
.lock
.refs
== 0)
928 ++hammer_count_refedbufs
;
929 hammer_ref(&iou
->io
.lock
);
933 bp
->b_flags
&= ~B_ERROR
;
936 hammer_io_set_modlist(&iou
->io
);
937 iou
->io
.modified
= 1;
940 hammer_stats_disk_write
+= iou
->io
.bytes
;
941 hammer_count_io_running_write
-= iou
->io
.bytes
;
942 iou
->io
.hmp
->io_running_space
-= iou
->io
.bytes
;
943 KKASSERT(iou
->io
.hmp
->io_running_space
>= 0);
947 * Remove from iorun list and wakeup any multi-io waiter(s).
949 if (TAILQ_FIRST(&iou
->io
.hmp
->iorun_list
) == &iou
->io
) {
950 ionext
= TAILQ_NEXT(&iou
->io
, iorun_entry
);
951 if (ionext
&& ionext
->type
== HAMMER_STRUCTURE_DUMMY
)
954 TAILQ_REMOVE(&iou
->io
.hmp
->iorun_list
, &iou
->io
, iorun_entry
);
956 hammer_stats_disk_read
+= iou
->io
.bytes
;
959 if (iou
->io
.waiting
) {
965 * If B_LOCKED is set someone wanted to deallocate the bp at some
966 * point, do it now if refs has become zero.
968 if ((bp
->b_flags
& B_LOCKED
) && iou
->io
.lock
.refs
== 0) {
969 KKASSERT(iou
->io
.modified
== 0);
970 --hammer_count_io_locked
;
971 bp
->b_flags
&= ~B_LOCKED
;
972 hammer_io_deallocate(bp
);
973 /* structure may be dead now */
978 * Callback from kernel when it wishes to deallocate a passively
979 * associated structure. This mostly occurs with clean buffers
980 * but it may be possible for a holding structure to be marked dirty
981 * while its buffer is passively associated. The caller owns the bp.
983 * If we cannot disassociate we set B_LOCKED to prevent the buffer
984 * from getting reused.
986 * WARNING: Because this can be called directly by getnewbuf we cannot
987 * recurse into the tree. If a bp cannot be immediately disassociated
988 * our only recourse is to set B_LOCKED.
990 * WARNING: This may be called from an interrupt via hammer_io_complete()
993 hammer_io_deallocate(struct buf
*bp
)
995 hammer_io_structure_t iou
= (void *)LIST_FIRST(&bp
->b_dep
);
997 KKASSERT((bp
->b_flags
& B_LOCKED
) == 0 && iou
->io
.running
== 0);
998 if (iou
->io
.lock
.refs
> 0 || iou
->io
.modified
) {
1000 * It is not legal to disassociate a modified buffer. This
1001 * case really shouldn't ever occur.
1003 bp
->b_flags
|= B_LOCKED
;
1004 ++hammer_count_io_locked
;
1007 * Disassociate the BP. If the io has no refs left we
1008 * have to add it to the loose list.
1010 hammer_io_disassociate(iou
);
1011 if (iou
->io
.type
!= HAMMER_STRUCTURE_VOLUME
) {
1012 KKASSERT(iou
->io
.bp
== NULL
);
1013 KKASSERT(iou
->io
.mod_list
== NULL
);
1014 crit_enter(); /* biodone race against list */
1015 iou
->io
.mod_list
= &iou
->io
.hmp
->lose_list
;
1016 TAILQ_INSERT_TAIL(iou
->io
.mod_list
, &iou
->io
, mod_entry
);
1023 hammer_io_fsync(struct vnode
*vp
)
1029 * NOTE: will not be called unless we tell the kernel about the
1030 * bioops. Unused... we use the mount's VFS_SYNC instead.
1033 hammer_io_sync(struct mount
*mp
)
1039 hammer_io_movedeps(struct buf
*bp1
, struct buf
*bp2
)
1044 * I/O pre-check for reading and writing. HAMMER only uses this for
1045 * B_CACHE buffers so checkread just shouldn't happen, but if it does
1048 * Writing is a different case. We don't want the kernel to try to write
1049 * out a buffer that HAMMER may be modifying passively or which has a
1050 * dependancy. In addition, kernel-demanded writes can only proceed for
1051 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
1052 * buffer types can only be explicitly written by the flusher.
1054 * checkwrite will only be called for bdwrite()n buffers. If we return
1055 * success the kernel is guaranteed to initiate the buffer write.
1058 hammer_io_checkread(struct buf
*bp
)
1064 hammer_io_checkwrite(struct buf
*bp
)
1066 hammer_io_t io
= (void *)LIST_FIRST(&bp
->b_dep
);
1069 * This shouldn't happen under normal operation.
1071 if (io
->type
== HAMMER_STRUCTURE_VOLUME
||
1072 io
->type
== HAMMER_STRUCTURE_META_BUFFER
) {
1074 panic("hammer_io_checkwrite: illegal buffer");
1075 if ((bp
->b_flags
& B_LOCKED
) == 0) {
1076 bp
->b_flags
|= B_LOCKED
;
1077 ++hammer_count_io_locked
;
1083 * We can only clear the modified bit if the IO is not currently
1084 * undergoing modification. Otherwise we may miss changes.
1086 * Only data and undo buffers can reach here. These buffers do
1087 * not have terminal crc functions but we temporarily reference
1088 * the IO anyway, just in case.
1090 if (io
->modify_refs
== 0 && io
->modified
) {
1091 hammer_ref(&io
->lock
);
1092 hammer_io_clear_modify(io
, 0);
1093 hammer_unref(&io
->lock
);
1094 } else if (io
->modified
) {
1095 KKASSERT(io
->type
== HAMMER_STRUCTURE_DATA_BUFFER
);
1099 * The kernel is going to start the IO, set io->running.
1101 KKASSERT(io
->running
== 0);
1103 io
->hmp
->io_running_space
+= io
->bytes
;
1104 TAILQ_INSERT_TAIL(&io
->hmp
->iorun_list
, io
, iorun_entry
);
1105 hammer_count_io_running_write
+= io
->bytes
;
1110 * Return non-zero if we wish to delay the kernel's attempt to flush
1111 * this buffer to disk.
1114 hammer_io_countdeps(struct buf
*bp
, int n
)
1119 struct bio_ops hammer_bioops
= {
1120 .io_start
= hammer_io_start
,
1121 .io_complete
= hammer_io_complete
,
1122 .io_deallocate
= hammer_io_deallocate
,
1123 .io_fsync
= hammer_io_fsync
,
1124 .io_sync
= hammer_io_sync
,
1125 .io_movedeps
= hammer_io_movedeps
,
1126 .io_countdeps
= hammer_io_countdeps
,
1127 .io_checkread
= hammer_io_checkread
,
1128 .io_checkwrite
= hammer_io_checkwrite
,
1131 /************************************************************************
1133 ************************************************************************
1135 * These functions operate directly on the buffer cache buffer associated
1136 * with a front-end vnode rather then a back-end device vnode.
1140 * Read a buffer associated with a front-end vnode directly from the
1141 * disk media. The bio may be issued asynchronously. If leaf is non-NULL
1142 * we validate the CRC.
1144 * We must check for the presence of a HAMMER buffer to handle the case
1145 * where the reblocker has rewritten the data (which it does via the HAMMER
1146 * buffer system, not via the high-level vnode buffer cache), but not yet
1147 * committed the buffer to the media.
1150 hammer_io_direct_read(hammer_mount_t hmp
, struct bio
*bio
,
1151 hammer_btree_leaf_elm_t leaf
)
1153 hammer_off_t buf_offset
;
1154 hammer_off_t zone2_offset
;
1155 hammer_volume_t volume
;
1161 buf_offset
= bio
->bio_offset
;
1162 KKASSERT((buf_offset
& HAMMER_OFF_ZONE_MASK
) ==
1163 HAMMER_ZONE_LARGE_DATA
);
1166 * The buffer cache may have an aliased buffer (the reblocker can
1167 * write them). If it does we have to sync any dirty data before
1168 * we can build our direct-read. This is a non-critical code path.
1171 hammer_sync_buffers(hmp
, buf_offset
, bp
->b_bufsize
);
1174 * Resolve to a zone-2 offset. The conversion just requires
1175 * munging the top 4 bits but we want to abstract it anyway
1176 * so the blockmap code can verify the zone assignment.
1178 zone2_offset
= hammer_blockmap_lookup(hmp
, buf_offset
, &error
);
1181 KKASSERT((zone2_offset
& HAMMER_OFF_ZONE_MASK
) ==
1182 HAMMER_ZONE_RAW_BUFFER
);
1185 * Resolve volume and raw-offset for 3rd level bio. The
1186 * offset will be specific to the volume.
1188 vol_no
= HAMMER_VOL_DECODE(zone2_offset
);
1189 volume
= hammer_get_volume(hmp
, vol_no
, &error
);
1190 if (error
== 0 && zone2_offset
>= volume
->maxbuf_off
)
1197 nbio
= push_bio(bio
);
1198 nbio
->bio_offset
= volume
->ondisk
->vol_buf_beg
+
1199 (zone2_offset
& HAMMER_OFF_SHORT_MASK
);
1202 * XXX disabled - our CRC check doesn't work if the OS
1203 * does bogus_page replacement on the direct-read.
1205 if (leaf
&& hammer_verify_data
) {
1206 nbio
->bio_done
= hammer_io_direct_read_complete
;
1207 nbio
->bio_caller_info1
.uvalue32
= leaf
->data_crc
;
1210 hammer_stats_disk_read
+= bp
->b_bufsize
;
1211 vn_strategy(volume
->devvp
, nbio
);
1213 hammer_rel_volume(volume
, 0);
1216 kprintf("hammer_direct_read: failed @ %016llx\n",
1217 (long long)zone2_offset
);
1218 bp
->b_error
= error
;
1219 bp
->b_flags
|= B_ERROR
;
1227 * On completion of the BIO this callback must check the data CRC
1228 * and chain to the previous bio.
1232 hammer_io_direct_read_complete(struct bio
*nbio
)
1236 u_int32_t rec_crc
= nbio
->bio_caller_info1
.uvalue32
;
1239 if (crc32(bp
->b_data
, bp
->b_bufsize
) != rec_crc
) {
1240 kprintf("HAMMER: data_crc error @%016llx/%d\n",
1241 nbio
->bio_offset
, bp
->b_bufsize
);
1242 if (hammer_debug_critical
)
1243 Debugger("data_crc on read");
1244 bp
->b_flags
|= B_ERROR
;
1247 obio
= pop_bio(nbio
);
1253 * Write a buffer associated with a front-end vnode directly to the
1254 * disk media. The bio may be issued asynchronously.
1256 * The BIO is associated with the specified record and RECF_DIRECT_IO
1257 * is set. The recorded is added to its object.
1260 hammer_io_direct_write(hammer_mount_t hmp
, struct bio
*bio
,
1261 hammer_record_t record
)
1263 hammer_btree_leaf_elm_t leaf
= &record
->leaf
;
1264 hammer_off_t buf_offset
;
1265 hammer_off_t zone2_offset
;
1266 hammer_volume_t volume
;
1267 hammer_buffer_t buffer
;
1274 buf_offset
= leaf
->data_offset
;
1276 KKASSERT(buf_offset
> HAMMER_ZONE_BTREE
);
1277 KKASSERT(bio
->bio_buf
->b_cmd
== BUF_CMD_WRITE
);
1280 * Issue or execute the I/O. The new memory record must replace
1281 * the old one before the I/O completes, otherwise a reaquisition of
1282 * the buffer will load the old media data instead of the new.
1284 if ((buf_offset
& HAMMER_BUFMASK
) == 0 &&
1285 leaf
->data_len
>= HAMMER_BUFSIZE
) {
1287 * We are using the vnode's bio to write directly to the
1288 * media, any hammer_buffer at the same zone-X offset will
1289 * now have stale data.
1291 zone2_offset
= hammer_blockmap_lookup(hmp
, buf_offset
, &error
);
1292 vol_no
= HAMMER_VOL_DECODE(zone2_offset
);
1293 volume
= hammer_get_volume(hmp
, vol_no
, &error
);
1295 if (error
== 0 && zone2_offset
>= volume
->maxbuf_off
)
1299 KKASSERT((bp
->b_bufsize
& HAMMER_BUFMASK
) == 0);
1301 hammer_del_buffers(hmp, buf_offset,
1302 zone2_offset, bp->b_bufsize);
1306 * Second level bio - cached zone2 offset.
1308 * (We can put our bio_done function in either the
1309 * 2nd or 3rd level).
1311 nbio
= push_bio(bio
);
1312 nbio
->bio_offset
= zone2_offset
;
1313 nbio
->bio_done
= hammer_io_direct_write_complete
;
1314 nbio
->bio_caller_info1
.ptr
= record
;
1315 record
->zone2_offset
= zone2_offset
;
1316 record
->flags
|= HAMMER_RECF_DIRECT_IO
|
1317 HAMMER_RECF_DIRECT_INVAL
;
1320 * Third level bio - raw offset specific to the
1323 zone2_offset
&= HAMMER_OFF_SHORT_MASK
;
1324 nbio
= push_bio(nbio
);
1325 nbio
->bio_offset
= volume
->ondisk
->vol_buf_beg
+
1327 hammer_stats_disk_write
+= bp
->b_bufsize
;
1328 hammer_ip_replace_bulk(hmp
, record
);
1329 vn_strategy(volume
->devvp
, nbio
);
1330 hammer_io_flush_mark(volume
);
1332 hammer_rel_volume(volume
, 0);
1335 * Must fit in a standard HAMMER buffer. In this case all
1336 * consumers use the HAMMER buffer system and RECF_DIRECT_IO
1337 * does not need to be set-up.
1339 KKASSERT(((buf_offset
^ (buf_offset
+ leaf
->data_len
- 1)) & ~HAMMER_BUFMASK64
) == 0);
1341 ptr
= hammer_bread(hmp
, buf_offset
, &error
, &buffer
);
1344 bp
->b_flags
|= B_AGE
;
1345 hammer_io_modify(&buffer
->io
, 1);
1346 bcopy(bp
->b_data
, ptr
, leaf
->data_len
);
1347 hammer_io_modify_done(&buffer
->io
);
1348 hammer_rel_buffer(buffer
, 0);
1350 hammer_ip_replace_bulk(hmp
, record
);
1356 * Major suckage occured. Also note: The record was
1357 * never added to the tree so we do not have to worry
1358 * about the backend.
1360 kprintf("hammer_direct_write: failed @ %016llx\n",
1361 (long long)leaf
->data_offset
);
1365 bp
->b_flags
|= B_ERROR
;
1367 record
->flags
|= HAMMER_RECF_DELETED_FE
;
1368 hammer_rel_mem_record(record
);
1374 * On completion of the BIO this callback must disconnect
1375 * it from the hammer_record and chain to the previous bio.
1377 * An I/O error forces the mount to read-only. Data buffers
1378 * are not B_LOCKED like meta-data buffers are, so we have to
1379 * throw the buffer away to prevent the kernel from retrying.
1383 hammer_io_direct_write_complete(struct bio
*nbio
)
1387 hammer_record_t record
= nbio
->bio_caller_info1
.ptr
;
1390 obio
= pop_bio(nbio
);
1391 if (bp
->b_flags
& B_ERROR
) {
1392 hammer_critical_error(record
->ip
->hmp
, record
->ip
,
1394 "while writing bulk data");
1395 bp
->b_flags
|= B_INVAL
;
1399 KKASSERT(record
!= NULL
);
1400 KKASSERT(record
->flags
& HAMMER_RECF_DIRECT_IO
);
1401 if (record
->flags
& HAMMER_RECF_DIRECT_WAIT
) {
1402 record
->flags
&= ~(HAMMER_RECF_DIRECT_IO
|
1403 HAMMER_RECF_DIRECT_WAIT
);
1404 /* record can disappear once DIRECT_IO flag is cleared */
1405 wakeup(&record
->flags
);
1407 record
->flags
&= ~HAMMER_RECF_DIRECT_IO
;
1408 /* record can disappear once DIRECT_IO flag is cleared */
1414 * This is called before a record is either committed to the B-Tree
1415 * or destroyed, to resolve any associated direct-IO.
1417 * (1) We must wait for any direct-IO related to the record to complete.
1419 * (2) We must remove any buffer cache aliases for data accessed via
1420 * leaf->data_offset or zone2_offset so non-direct-IO consumers
1421 * (the mirroring and reblocking code) do not see stale data.
1424 hammer_io_direct_wait(hammer_record_t record
)
1427 * Wait for I/O to complete
1429 if (record
->flags
& HAMMER_RECF_DIRECT_IO
) {
1431 while (record
->flags
& HAMMER_RECF_DIRECT_IO
) {
1432 record
->flags
|= HAMMER_RECF_DIRECT_WAIT
;
1433 tsleep(&record
->flags
, 0, "hmdiow", 0);
1439 * Invalidate any related buffer cache aliases associated with the
1440 * backing device. This is needed because the buffer cache buffer
1441 * for file data is associated with the file vnode, not the backing
1444 * XXX I do not think this case can occur any more now that
1445 * reservations ensure that all such buffers are removed before
1446 * an area can be reused.
1448 if (record
->flags
& HAMMER_RECF_DIRECT_INVAL
) {
1449 KKASSERT(record
->leaf
.data_offset
);
1450 hammer_del_buffers(record
->ip
->hmp
, record
->leaf
.data_offset
,
1451 record
->zone2_offset
, record
->leaf
.data_len
,
1453 record
->flags
&= ~HAMMER_RECF_DIRECT_INVAL
;
1458 * This is called to remove the second-level cached zone-2 offset from
1459 * frontend buffer cache buffers, now stale due to a data relocation.
1460 * These offsets are generated by cluster_read() via VOP_BMAP, or directly
1461 * by hammer_vop_strategy_read().
1463 * This is rather nasty because here we have something like the reblocker
1464 * scanning the raw B-Tree with no held references on anything, really,
1465 * other then a shared lock on the B-Tree node, and we have to access the
1466 * frontend's buffer cache to check for and clean out the association.
1467 * Specifically, if the reblocker is moving data on the disk, these cached
1468 * offsets will become invalid.
1470 * Only data record types associated with the large-data zone are subject
1471 * to direct-io and need to be checked.
1475 hammer_io_direct_uncache(hammer_mount_t hmp
, hammer_btree_leaf_elm_t leaf
)
1477 struct hammer_inode_info iinfo
;
1480 if (leaf
->base
.rec_type
!= HAMMER_RECTYPE_DATA
)
1482 zone
= HAMMER_ZONE_DECODE(leaf
->data_offset
);
1483 if (zone
!= HAMMER_ZONE_LARGE_DATA_INDEX
)
1485 iinfo
.obj_id
= leaf
->base
.obj_id
;
1486 iinfo
.obj_asof
= 0; /* unused */
1487 iinfo
.obj_localization
= leaf
->base
.localization
&
1488 HAMMER_LOCALIZE_PSEUDOFS_MASK
;
1489 iinfo
.u
.leaf
= leaf
;
1490 hammer_scan_inode_snapshots(hmp
, &iinfo
,
1491 hammer_io_direct_uncache_callback
,
1496 hammer_io_direct_uncache_callback(hammer_inode_t ip
, void *data
)
1498 hammer_inode_info_t iinfo
= data
;
1499 hammer_off_t data_offset
;
1500 hammer_off_t file_offset
;
1507 data_offset
= iinfo
->u
.leaf
->data_offset
;
1508 file_offset
= iinfo
->u
.leaf
->base
.key
- iinfo
->u
.leaf
->data_len
;
1509 blksize
= iinfo
->u
.leaf
->data_len
;
1510 KKASSERT((blksize
& HAMMER_BUFMASK
) == 0);
1512 hammer_ref(&ip
->lock
);
1513 if (hammer_get_vnode(ip
, &vp
) == 0) {
1514 if ((bp
= findblk(ip
->vp
, file_offset
, FINDBLK_TEST
)) != NULL
&&
1515 bp
->b_bio2
.bio_offset
!= NOOFFSET
) {
1516 bp
= getblk(ip
->vp
, file_offset
, blksize
, 0, 0);
1517 bp
->b_bio2
.bio_offset
= NOOFFSET
;
1522 hammer_rel_inode(ip
, 0);
1528 * This function is called when writes may have occured on the volume,
1529 * indicating that the device may be holding cached writes.
1532 hammer_io_flush_mark(hammer_volume_t volume
)
1534 volume
->vol_flags
|= HAMMER_VOLF_NEEDFLUSH
;
1538 * This function ensures that the device has flushed any cached writes out.
1541 hammer_io_flush_sync(hammer_mount_t hmp
)
1543 hammer_volume_t volume
;
1544 struct buf
*bp_base
= NULL
;
1547 RB_FOREACH(volume
, hammer_vol_rb_tree
, &hmp
->rb_vols_root
) {
1548 if (volume
->vol_flags
& HAMMER_VOLF_NEEDFLUSH
) {
1549 volume
->vol_flags
&= ~HAMMER_VOLF_NEEDFLUSH
;
1551 bp
->b_bio1
.bio_offset
= 0;
1554 bp
->b_cmd
= BUF_CMD_FLUSH
;
1555 bp
->b_bio1
.bio_caller_info1
.cluster_head
= bp_base
;
1556 bp
->b_bio1
.bio_done
= biodone_sync
;
1557 bp
->b_bio1
.bio_flags
|= BIO_SYNC
;
1559 vn_strategy(volume
->devvp
, &bp
->b_bio1
);
1562 while ((bp
= bp_base
) != NULL
) {
1563 bp_base
= bp
->b_bio1
.bio_caller_info1
.cluster_head
;
1564 biowait(&bp
->b_bio1
, "hmrFLS");