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 1.49.2.5 2008/09/15 17:10:03 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
);
65 * Initialize a new, already-zero'd hammer_io structure, or reinitialize
66 * an existing hammer_io structure which may have switched to another type.
69 hammer_io_init(hammer_io_t io
, hammer_mount_t hmp
, enum hammer_io_type type
)
76 * Helper routine to disassociate a buffer cache buffer from an I/O
77 * structure. The buffer is unlocked and marked appropriate for reclamation.
79 * The io may have 0 or 1 references depending on who called us. The
80 * caller is responsible for dealing with the refs.
82 * This call can only be made when no action is required on the buffer.
84 * The caller must own the buffer and the IO must indicate that the
85 * structure no longer owns it (io.released != 0).
88 hammer_io_disassociate(hammer_io_structure_t iou
)
90 struct buf
*bp
= iou
->io
.bp
;
92 KKASSERT(iou
->io
.released
);
93 KKASSERT(iou
->io
.modified
== 0);
94 KKASSERT(LIST_FIRST(&bp
->b_dep
) == (void *)iou
);
99 * If the buffer was locked someone wanted to get rid of it.
101 if (bp
->b_flags
& B_LOCKED
) {
102 --hammer_count_io_locked
;
103 bp
->b_flags
&= ~B_LOCKED
;
105 if (iou
->io
.reclaim
) {
106 bp
->b_flags
|= B_NOCACHE
|B_RELBUF
;
110 switch(iou
->io
.type
) {
111 case HAMMER_STRUCTURE_VOLUME
:
112 iou
->volume
.ondisk
= NULL
;
114 case HAMMER_STRUCTURE_DATA_BUFFER
:
115 case HAMMER_STRUCTURE_META_BUFFER
:
116 case HAMMER_STRUCTURE_UNDO_BUFFER
:
117 iou
->buffer
.ondisk
= NULL
;
123 * Wait for any physical IO to complete
126 hammer_io_wait(hammer_io_t io
)
130 tsleep_interlock(io
);
133 tsleep(io
, 0, "hmrflw", 0);
134 if (io
->running
== 0)
136 tsleep_interlock(io
);
138 if (io
->running
== 0)
146 * Wait for all hammer_io-initated write I/O's to complete. This is not
147 * supposed to count direct I/O's but some can leak through (for
148 * non-full-sized direct I/Os).
151 hammer_io_wait_all(hammer_mount_t hmp
, const char *ident
)
154 while (hmp
->io_running_space
)
155 tsleep(&hmp
->io_running_space
, 0, ident
, 0);
159 #define HAMMER_MAXRA 4
162 * Load bp for a HAMMER structure. The io must be exclusively locked by
165 * This routine is mostly used on meta-data and small-data blocks. Generally
166 * speaking HAMMER assumes some locality of reference and will cluster
169 * Note that clustering occurs at the device layer, not the logical layer.
170 * If the buffers do not apply to the current operation they may apply to
174 hammer_io_read(struct vnode
*devvp
, struct hammer_io
*io
, hammer_off_t limit
)
179 if ((bp
= io
->bp
) == NULL
) {
180 hammer_count_io_running_read
+= io
->bytes
;
181 if (hammer_cluster_enable
) {
182 error
= cluster_read(devvp
, limit
,
183 io
->offset
, io
->bytes
,
185 HAMMER_CLUSTER_BUFS
, &io
->bp
);
187 error
= bread(devvp
, io
->offset
, io
->bytes
, &io
->bp
);
189 hammer_stats_disk_read
+= io
->bytes
;
190 hammer_count_io_running_read
-= io
->bytes
;
193 * The code generally assumes b_ops/b_dep has been set-up,
194 * even if we error out here.
197 bp
->b_ops
= &hammer_bioops
;
198 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
199 LIST_INSERT_HEAD(&bp
->b_dep
, &io
->worklist
, node
);
201 KKASSERT(io
->modified
== 0);
202 KKASSERT(io
->running
== 0);
203 KKASSERT(io
->waiting
== 0);
204 io
->released
= 0; /* we hold an active lock on bp */
212 * Similar to hammer_io_read() but returns a zero'd out buffer instead.
213 * Must be called with the IO exclusively locked.
215 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background
216 * I/O by forcing the buffer to not be in a released state before calling
219 * This function will also mark the IO as modified but it will not
220 * increment the modify_refs count.
223 hammer_io_new(struct vnode
*devvp
, struct hammer_io
*io
)
227 if ((bp
= io
->bp
) == NULL
) {
228 io
->bp
= getblk(devvp
, io
->offset
, io
->bytes
, 0, 0);
230 bp
->b_ops
= &hammer_bioops
;
231 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
232 LIST_INSERT_HEAD(&bp
->b_dep
, &io
->worklist
, node
);
234 KKASSERT(io
->running
== 0);
244 hammer_io_modify(io
, 0);
250 * Remove potential device level aliases against buffers managed by high level
251 * vnodes. Aliases can also be created due to mixed buffer sizes.
253 * This is nasty because the buffers are also VMIO-backed. Even if a buffer
254 * does not exist its backing VM pages might, and we have to invalidate
255 * those as well or a getblk() will reinstate them.
258 hammer_io_inval(hammer_volume_t volume
, hammer_off_t zone2_offset
)
260 hammer_io_structure_t iou
;
261 hammer_off_t phys_offset
;
264 phys_offset
= volume
->ondisk
->vol_buf_beg
+
265 (zone2_offset
& HAMMER_OFF_SHORT_MASK
);
267 if ((bp
= findblk(volume
->devvp
, phys_offset
)) != NULL
)
268 bp
= getblk(volume
->devvp
, phys_offset
, bp
->b_bufsize
, 0, 0);
270 bp
= getblk(volume
->devvp
, phys_offset
, HAMMER_BUFSIZE
, 0, 0);
271 if ((iou
= (void *)LIST_FIRST(&bp
->b_dep
)) != NULL
) {
272 hammer_ref(&iou
->io
.lock
);
273 hammer_io_clear_modify(&iou
->io
, 1);
275 iou
->io
.released
= 0;
279 KKASSERT(iou
->io
.lock
.refs
== 1);
280 hammer_rel_buffer(&iou
->buffer
, 0);
281 /*hammer_io_deallocate(bp);*/
283 KKASSERT((bp
->b_flags
& B_LOCKED
) == 0);
285 bp
->b_flags
|= B_NOCACHE
|B_RELBUF
;
292 * This routine is called on the last reference to a hammer structure.
293 * The io is usually interlocked with io.loading and io.refs must be 1.
295 * This routine may return a non-NULL bp to the caller for dispoal. Disposal
296 * simply means the caller finishes decrementing the ref-count on the
297 * IO structure then brelse()'s the bp. The bp may or may not still be
298 * passively associated with the IO.
300 * The only requirement here is that modified meta-data and volume-header
301 * buffer may NOT be disassociated from the IO structure, and consequently
302 * we also leave such buffers actively associated with the IO if they already
303 * are (since the kernel can't do anything with them anyway). Only the
304 * flusher is allowed to write such buffers out. Modified pure-data and
305 * undo buffers are returned to the kernel but left passively associated
306 * so we can track when the kernel writes the bp out.
309 hammer_io_release(struct hammer_io
*io
, int flush
)
311 union hammer_io_structure
*iou
= (void *)io
;
314 if ((bp
= io
->bp
) == NULL
)
318 * Try to flush a dirty IO to disk if asked to by the
319 * caller or if the kernel tried to flush the buffer in the past.
321 * Kernel-initiated flushes are only allowed for pure-data buffers.
322 * meta-data and volume buffers can only be flushed explicitly
328 } else if (bp
->b_flags
& B_LOCKED
) {
330 case HAMMER_STRUCTURE_DATA_BUFFER
:
331 case HAMMER_STRUCTURE_UNDO_BUFFER
:
337 } /* else no explicit request to flush the buffer */
341 * Wait for the IO to complete if asked to. This occurs when
342 * the buffer must be disposed of definitively during an umount
343 * or buffer invalidation.
345 if (io
->waitdep
&& io
->running
) {
350 * Return control of the buffer to the kernel (with the provisio
351 * that our bioops can override kernel decisions with regards to
354 if ((flush
|| io
->reclaim
) && io
->modified
== 0 && io
->running
== 0) {
356 * Always disassociate the bp if an explicit flush
357 * was requested and the IO completed with no error
358 * (so unmount can really clean up the structure).
366 hammer_io_disassociate((hammer_io_structure_t
)io
);
368 } else if (io
->modified
) {
370 * Only certain IO types can be released to the kernel if
371 * the buffer has been modified.
373 * volume and meta-data IO types may only be explicitly
377 case HAMMER_STRUCTURE_DATA_BUFFER
:
378 case HAMMER_STRUCTURE_UNDO_BUFFER
:
379 if (io
->released
== 0) {
387 bp
= NULL
; /* bp left associated */
388 } else if (io
->released
== 0) {
390 * Clean buffers can be generally released to the kernel.
391 * We leave the bp passively associated with the HAMMER
392 * structure and use bioops to disconnect it later on
393 * if the kernel wants to discard the buffer.
395 * We can steal the structure's ownership of the bp.
398 if (bp
->b_flags
& B_LOCKED
) {
399 hammer_io_disassociate(iou
);
403 hammer_io_disassociate(iou
);
406 /* return the bp (bp passively associated) */
411 * A released buffer is passively associate with our
412 * hammer_io structure. The kernel cannot destroy it
413 * without making a bioops call. If the kernel (B_LOCKED)
414 * or we (reclaim) requested that the buffer be destroyed
415 * we destroy it, otherwise we do a quick get/release to
416 * reset its position in the kernel's LRU list.
418 * Leaving the buffer passively associated allows us to
419 * use the kernel's LRU buffer flushing mechanisms rather
420 * then rolling our own.
422 * XXX there are two ways of doing this. We can re-acquire
423 * and passively release to reset the LRU, or not.
425 if (io
->running
== 0) {
427 if ((bp
->b_flags
& B_LOCKED
) || io
->reclaim
) {
428 hammer_io_disassociate(iou
);
431 /* return the bp (bp passively associated) */
435 * bp is left passively associated but we do not
436 * try to reacquire it. Interactions with the io
437 * structure will occur on completion of the bp's
447 * This routine is called with a locked IO when a flush is desired and
448 * no other references to the structure exists other then ours. This
449 * routine is ONLY called when HAMMER believes it is safe to flush a
450 * potentially modified buffer out.
453 hammer_io_flush(struct hammer_io
*io
)
458 * Degenerate case - nothing to flush if nothing is dirty.
460 if (io
->modified
== 0) {
465 KKASSERT(io
->modify_refs
<= 0);
468 * Acquire ownership of the bp, particularly before we clear our
471 * We are going to bawrite() this bp. Don't leave a window where
472 * io->released is set, we actually own the bp rather then our
478 /* BUF_KERNPROC(io->bp); */
479 /* io->released = 0; */
480 KKASSERT(io
->released
);
481 KKASSERT(io
->bp
== bp
);
486 * Acquire exclusive access to the bp and then clear the modified
487 * state of the buffer prior to issuing I/O to interlock any
488 * modifications made while the I/O is in progress. This shouldn't
489 * happen anyway but losing data would be worse. The modified bit
490 * will be rechecked after the IO completes.
492 * NOTE: This call also finalizes the buffer's content (inval == 0).
494 * This is only legal when lock.refs == 1 (otherwise we might clear
495 * the modified bit while there are still users of the cluster
496 * modifying the data).
498 * Do this before potentially blocking so any attempt to modify the
499 * ondisk while we are blocked blocks waiting for us.
501 hammer_ref(&io
->lock
);
502 hammer_io_clear_modify(io
, 0);
503 hammer_unref(&io
->lock
);
506 * Transfer ownership to the kernel and initiate I/O.
509 io
->hmp
->io_running_space
+= io
->bytes
;
510 hammer_count_io_running_write
+= io
->bytes
;
514 /************************************************************************
516 ************************************************************************
518 * These routines deal with dependancies created when IO buffers get
519 * modified. The caller must call hammer_modify_*() on a referenced
520 * HAMMER structure prior to modifying its on-disk data.
522 * Any intent to modify an IO buffer acquires the related bp and imposes
523 * various write ordering dependancies.
527 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
528 * are locked until the flusher can deal with them, pure data buffers
529 * can be written out.
533 hammer_io_modify(hammer_io_t io
, int count
)
536 * io->modify_refs must be >= 0
538 while (io
->modify_refs
< 0) {
540 tsleep(io
, 0, "hmrmod", 0);
544 * Shortcut if nothing to do.
546 KKASSERT(io
->lock
.refs
!= 0 && io
->bp
!= NULL
);
547 io
->modify_refs
+= count
;
548 if (io
->modified
&& io
->released
== 0)
551 hammer_lock_ex(&io
->lock
);
552 if (io
->modified
== 0) {
553 hammer_io_set_modlist(io
);
558 BUF_KERNPROC(io
->bp
);
560 KKASSERT(io
->modified
!= 0);
562 hammer_unlock(&io
->lock
);
567 hammer_io_modify_done(hammer_io_t io
)
569 KKASSERT(io
->modify_refs
> 0);
571 if (io
->modify_refs
== 0 && io
->waitmod
) {
578 hammer_io_write_interlock(hammer_io_t io
)
580 while (io
->modify_refs
!= 0) {
582 tsleep(io
, 0, "hmrmod", 0);
584 io
->modify_refs
= -1;
588 hammer_io_done_interlock(hammer_io_t io
)
590 KKASSERT(io
->modify_refs
== -1);
599 * Caller intends to modify a volume's ondisk structure.
601 * This is only allowed if we are the flusher or we have a ref on the
605 hammer_modify_volume(hammer_transaction_t trans
, hammer_volume_t volume
,
608 KKASSERT (trans
== NULL
|| trans
->sync_lock_refs
> 0);
610 hammer_io_modify(&volume
->io
, 1);
612 intptr_t rel_offset
= (intptr_t)base
- (intptr_t)volume
->ondisk
;
613 KKASSERT((rel_offset
& ~(intptr_t)HAMMER_BUFMASK
) == 0);
614 hammer_generate_undo(trans
, &volume
->io
,
615 HAMMER_ENCODE_RAW_VOLUME(volume
->vol_no
, rel_offset
),
621 * Caller intends to modify a buffer's ondisk structure.
623 * This is only allowed if we are the flusher or we have a ref on the
627 hammer_modify_buffer(hammer_transaction_t trans
, hammer_buffer_t buffer
,
630 KKASSERT (trans
== NULL
|| trans
->sync_lock_refs
> 0);
632 hammer_io_modify(&buffer
->io
, 1);
634 intptr_t rel_offset
= (intptr_t)base
- (intptr_t)buffer
->ondisk
;
635 KKASSERT((rel_offset
& ~(intptr_t)HAMMER_BUFMASK
) == 0);
636 hammer_generate_undo(trans
, &buffer
->io
,
637 buffer
->zone2_offset
+ rel_offset
,
643 hammer_modify_volume_done(hammer_volume_t volume
)
645 hammer_io_modify_done(&volume
->io
);
649 hammer_modify_buffer_done(hammer_buffer_t buffer
)
651 hammer_io_modify_done(&buffer
->io
);
655 * Mark an entity as not being dirty any more and finalize any
656 * delayed adjustments to the buffer.
658 * Delayed adjustments are an important performance enhancement, allowing
659 * us to avoid recalculating B-Tree node CRCs over and over again when
660 * making bulk-modifications to the B-Tree.
662 * If inval is non-zero delayed adjustments are ignored.
664 * This routine may dereference related btree nodes and cause the
665 * buffer to be dereferenced. The caller must own a reference on io.
668 hammer_io_clear_modify(struct hammer_io
*io
, int inval
)
670 if (io
->modified
== 0)
674 * Take us off the mod-list and clear the modified bit.
676 KKASSERT(io
->mod_list
!= NULL
);
677 if (io
->mod_list
== &io
->hmp
->volu_list
||
678 io
->mod_list
== &io
->hmp
->meta_list
) {
679 io
->hmp
->locked_dirty_space
-= io
->bytes
;
680 hammer_count_dirtybufspace
-= io
->bytes
;
682 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
687 * If this bit is not set there are no delayed adjustments.
694 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference
695 * on the node (& underlying buffer). Release the node after clearing
698 if (io
->type
== HAMMER_STRUCTURE_META_BUFFER
) {
699 hammer_buffer_t buffer
= (void *)io
;
703 TAILQ_FOREACH(node
, &buffer
->clist
, entry
) {
704 if ((node
->flags
& HAMMER_NODE_NEEDSCRC
) == 0)
706 node
->flags
&= ~HAMMER_NODE_NEEDSCRC
;
707 KKASSERT(node
->ondisk
);
709 node
->ondisk
->crc
= crc32(&node
->ondisk
->crc
+ 1, HAMMER_BTREE_CRCSIZE
);
710 hammer_rel_node(node
);
714 /* caller must still have ref on io */
715 KKASSERT(io
->lock
.refs
> 0);
719 * Clear the IO's modify list. Even though the IO is no longer modified
720 * it may still be on the lose_list. This routine is called just before
721 * the governing hammer_buffer is destroyed.
724 hammer_io_clear_modlist(struct hammer_io
*io
)
726 KKASSERT(io
->modified
== 0);
728 crit_enter(); /* biodone race against list */
729 KKASSERT(io
->mod_list
== &io
->hmp
->lose_list
);
730 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
737 hammer_io_set_modlist(struct hammer_io
*io
)
739 struct hammer_mount
*hmp
= io
->hmp
;
741 KKASSERT(io
->mod_list
== NULL
);
744 case HAMMER_STRUCTURE_VOLUME
:
745 io
->mod_list
= &hmp
->volu_list
;
746 hmp
->locked_dirty_space
+= io
->bytes
;
747 hammer_count_dirtybufspace
+= io
->bytes
;
749 case HAMMER_STRUCTURE_META_BUFFER
:
750 io
->mod_list
= &hmp
->meta_list
;
751 hmp
->locked_dirty_space
+= io
->bytes
;
752 hammer_count_dirtybufspace
+= io
->bytes
;
754 case HAMMER_STRUCTURE_UNDO_BUFFER
:
755 io
->mod_list
= &hmp
->undo_list
;
757 case HAMMER_STRUCTURE_DATA_BUFFER
:
758 io
->mod_list
= &hmp
->data_list
;
761 TAILQ_INSERT_TAIL(io
->mod_list
, io
, mod_entry
);
764 /************************************************************************
766 ************************************************************************
771 * Pre-IO initiation kernel callback - cluster build only
774 hammer_io_start(struct buf
*bp
)
779 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT!
781 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
782 * may also be set if we were marking a cluster header open. Only remove
783 * our dependancy if the modified bit is clear.
786 hammer_io_complete(struct buf
*bp
)
788 union hammer_io_structure
*iou
= (void *)LIST_FIRST(&bp
->b_dep
);
790 KKASSERT(iou
->io
.released
== 1);
793 * Deal with people waiting for I/O to drain
795 if (iou
->io
.running
) {
797 * Deal with critical write errors. Once a critical error
798 * has been flagged in hmp the UNDO FIFO will not be updated.
799 * That way crash recover will give us a consistent
802 * Because of this we can throw away failed UNDO buffers. If
803 * we throw away META or DATA buffers we risk corrupting
804 * the now read-only version of the filesystem visible to
805 * the user. Clear B_ERROR so the buffer is not re-dirtied
806 * by the kernel and ref the io so it doesn't get thrown
809 if (bp
->b_flags
& B_ERROR
) {
810 hammer_critical_error(iou
->io
.hmp
, NULL
, bp
->b_error
,
811 "while flushing meta-data");
812 switch(iou
->io
.type
) {
813 case HAMMER_STRUCTURE_UNDO_BUFFER
:
816 if (iou
->io
.ioerror
== 0) {
818 if (iou
->io
.lock
.refs
== 0)
819 ++hammer_count_refedbufs
;
820 hammer_ref(&iou
->io
.lock
);
824 bp
->b_flags
&= ~B_ERROR
;
827 hammer_io_set_modlist(&iou
->io
);
828 iou
->io
.modified
= 1;
831 hammer_stats_disk_write
+= iou
->io
.bytes
;
832 hammer_count_io_running_write
-= iou
->io
.bytes
;
833 iou
->io
.hmp
->io_running_space
-= iou
->io
.bytes
;
834 if (iou
->io
.hmp
->io_running_space
== 0)
835 wakeup(&iou
->io
.hmp
->io_running_space
);
836 KKASSERT(iou
->io
.hmp
->io_running_space
>= 0);
839 hammer_stats_disk_read
+= iou
->io
.bytes
;
842 if (iou
->io
.waiting
) {
848 * If B_LOCKED is set someone wanted to deallocate the bp at some
849 * point, do it now if refs has become zero.
851 if ((bp
->b_flags
& B_LOCKED
) && iou
->io
.lock
.refs
== 0) {
852 KKASSERT(iou
->io
.modified
== 0);
853 --hammer_count_io_locked
;
854 bp
->b_flags
&= ~B_LOCKED
;
855 hammer_io_deallocate(bp
);
856 /* structure may be dead now */
861 * Callback from kernel when it wishes to deallocate a passively
862 * associated structure. This mostly occurs with clean buffers
863 * but it may be possible for a holding structure to be marked dirty
864 * while its buffer is passively associated. The caller owns the bp.
866 * If we cannot disassociate we set B_LOCKED to prevent the buffer
867 * from getting reused.
869 * WARNING: Because this can be called directly by getnewbuf we cannot
870 * recurse into the tree. If a bp cannot be immediately disassociated
871 * our only recourse is to set B_LOCKED.
873 * WARNING: This may be called from an interrupt via hammer_io_complete()
876 hammer_io_deallocate(struct buf
*bp
)
878 hammer_io_structure_t iou
= (void *)LIST_FIRST(&bp
->b_dep
);
880 KKASSERT((bp
->b_flags
& B_LOCKED
) == 0 && iou
->io
.running
== 0);
881 if (iou
->io
.lock
.refs
> 0 || iou
->io
.modified
) {
883 * It is not legal to disassociate a modified buffer. This
884 * case really shouldn't ever occur.
886 bp
->b_flags
|= B_LOCKED
;
887 ++hammer_count_io_locked
;
890 * Disassociate the BP. If the io has no refs left we
891 * have to add it to the loose list.
893 hammer_io_disassociate(iou
);
894 if (iou
->io
.type
!= HAMMER_STRUCTURE_VOLUME
) {
895 KKASSERT(iou
->io
.bp
== NULL
);
896 KKASSERT(iou
->io
.mod_list
== NULL
);
897 crit_enter(); /* biodone race against list */
898 iou
->io
.mod_list
= &iou
->io
.hmp
->lose_list
;
899 TAILQ_INSERT_TAIL(iou
->io
.mod_list
, &iou
->io
, mod_entry
);
906 hammer_io_fsync(struct vnode
*vp
)
912 * NOTE: will not be called unless we tell the kernel about the
913 * bioops. Unused... we use the mount's VFS_SYNC instead.
916 hammer_io_sync(struct mount
*mp
)
922 hammer_io_movedeps(struct buf
*bp1
, struct buf
*bp2
)
927 * I/O pre-check for reading and writing. HAMMER only uses this for
928 * B_CACHE buffers so checkread just shouldn't happen, but if it does
931 * Writing is a different case. We don't want the kernel to try to write
932 * out a buffer that HAMMER may be modifying passively or which has a
933 * dependancy. In addition, kernel-demanded writes can only proceed for
934 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
935 * buffer types can only be explicitly written by the flusher.
937 * checkwrite will only be called for bdwrite()n buffers. If we return
938 * success the kernel is guaranteed to initiate the buffer write.
941 hammer_io_checkread(struct buf
*bp
)
947 hammer_io_checkwrite(struct buf
*bp
)
949 hammer_io_t io
= (void *)LIST_FIRST(&bp
->b_dep
);
952 * This shouldn't happen under normal operation.
954 if (io
->type
== HAMMER_STRUCTURE_VOLUME
||
955 io
->type
== HAMMER_STRUCTURE_META_BUFFER
) {
957 panic("hammer_io_checkwrite: illegal buffer");
958 if ((bp
->b_flags
& B_LOCKED
) == 0) {
959 bp
->b_flags
|= B_LOCKED
;
960 ++hammer_count_io_locked
;
966 * We can only clear the modified bit if the IO is not currently
967 * undergoing modification. Otherwise we may miss changes.
969 * Only data and undo buffers can reach here. These buffers do
970 * not have terminal crc functions but we temporarily reference
971 * the IO anyway, just in case.
973 if (io
->modify_refs
== 0 && io
->modified
) {
974 hammer_ref(&io
->lock
);
975 hammer_io_clear_modify(io
, 0);
976 hammer_unref(&io
->lock
);
977 } else if (io
->modified
) {
978 KKASSERT(io
->type
== HAMMER_STRUCTURE_DATA_BUFFER
);
982 * The kernel is going to start the IO, set io->running.
984 KKASSERT(io
->running
== 0);
986 io
->hmp
->io_running_space
+= io
->bytes
;
987 hammer_count_io_running_write
+= io
->bytes
;
992 * Return non-zero if we wish to delay the kernel's attempt to flush
993 * this buffer to disk.
996 hammer_io_countdeps(struct buf
*bp
, int n
)
1001 struct bio_ops hammer_bioops
= {
1002 .io_start
= hammer_io_start
,
1003 .io_complete
= hammer_io_complete
,
1004 .io_deallocate
= hammer_io_deallocate
,
1005 .io_fsync
= hammer_io_fsync
,
1006 .io_sync
= hammer_io_sync
,
1007 .io_movedeps
= hammer_io_movedeps
,
1008 .io_countdeps
= hammer_io_countdeps
,
1009 .io_checkread
= hammer_io_checkread
,
1010 .io_checkwrite
= hammer_io_checkwrite
,
1013 /************************************************************************
1015 ************************************************************************
1017 * These functions operate directly on the buffer cache buffer associated
1018 * with a front-end vnode rather then a back-end device vnode.
1022 * Read a buffer associated with a front-end vnode directly from the
1023 * disk media. The bio may be issued asynchronously. If leaf is non-NULL
1024 * we validate the CRC.
1026 * We must check for the presence of a HAMMER buffer to handle the case
1027 * where the reblocker has rewritten the data (which it does via the HAMMER
1028 * buffer system, not via the high-level vnode buffer cache), but not yet
1029 * committed the buffer to the media.
1032 hammer_io_direct_read(hammer_mount_t hmp
, struct bio
*bio
,
1033 hammer_btree_leaf_elm_t leaf
)
1035 hammer_off_t buf_offset
;
1036 hammer_off_t zone2_offset
;
1037 hammer_volume_t volume
;
1043 buf_offset
= bio
->bio_offset
;
1044 KKASSERT((buf_offset
& HAMMER_OFF_ZONE_MASK
) ==
1045 HAMMER_ZONE_LARGE_DATA
);
1048 * The buffer cache may have an aliased buffer (the reblocker can
1049 * write them). If it does we have to sync any dirty data before
1050 * we can build our direct-read. This is a non-critical code path.
1053 hammer_sync_buffers(hmp
, buf_offset
, bp
->b_bufsize
);
1056 * Resolve to a zone-2 offset. The conversion just requires
1057 * munging the top 4 bits but we want to abstract it anyway
1058 * so the blockmap code can verify the zone assignment.
1060 zone2_offset
= hammer_blockmap_lookup(hmp
, buf_offset
, &error
);
1063 KKASSERT((zone2_offset
& HAMMER_OFF_ZONE_MASK
) ==
1064 HAMMER_ZONE_RAW_BUFFER
);
1067 * Resolve volume and raw-offset for 3rd level bio. The
1068 * offset will be specific to the volume.
1070 vol_no
= HAMMER_VOL_DECODE(zone2_offset
);
1071 volume
= hammer_get_volume(hmp
, vol_no
, &error
);
1072 if (error
== 0 && zone2_offset
>= volume
->maxbuf_off
)
1079 nbio
= push_bio(bio
);
1080 nbio
->bio_offset
= volume
->ondisk
->vol_buf_beg
+
1081 (zone2_offset
& HAMMER_OFF_SHORT_MASK
);
1084 * XXX disabled - our CRC check doesn't work if the OS
1085 * does bogus_page replacement on the direct-read.
1087 if (leaf
&& hammer_verify_data
) {
1088 nbio
->bio_done
= hammer_io_direct_read_complete
;
1089 nbio
->bio_caller_info1
.uvalue32
= leaf
->data_crc
;
1092 hammer_stats_disk_read
+= bp
->b_bufsize
;
1093 vn_strategy(volume
->devvp
, nbio
);
1095 hammer_rel_volume(volume
, 0);
1098 kprintf("hammer_direct_read: failed @ %016llx\n",
1100 bp
->b_error
= error
;
1101 bp
->b_flags
|= B_ERROR
;
1109 * On completion of the BIO this callback must check the data CRC
1110 * and chain to the previous bio.
1114 hammer_io_direct_read_complete(struct bio
*nbio
)
1118 u_int32_t rec_crc
= nbio
->bio_caller_info1
.uvalue32
;
1121 if (crc32(bp
->b_data
, bp
->b_bufsize
) != rec_crc
) {
1122 kprintf("HAMMER: data_crc error @%016llx/%d\n",
1123 nbio
->bio_offset
, bp
->b_bufsize
);
1124 if (hammer_debug_debug
)
1126 bp
->b_flags
|= B_ERROR
;
1129 obio
= pop_bio(nbio
);
1135 * Write a buffer associated with a front-end vnode directly to the
1136 * disk media. The bio may be issued asynchronously.
1138 * The BIO is associated with the specified record and RECF_DIRECT_IO
1139 * is set. The recorded is added to its object.
1142 hammer_io_direct_write(hammer_mount_t hmp
, hammer_record_t record
,
1145 hammer_btree_leaf_elm_t leaf
= &record
->leaf
;
1146 hammer_off_t buf_offset
;
1147 hammer_off_t zone2_offset
;
1148 hammer_volume_t volume
;
1149 hammer_buffer_t buffer
;
1156 buf_offset
= leaf
->data_offset
;
1158 KKASSERT(buf_offset
> HAMMER_ZONE_BTREE
);
1159 KKASSERT(bio
->bio_buf
->b_cmd
== BUF_CMD_WRITE
);
1161 if ((buf_offset
& HAMMER_BUFMASK
) == 0 &&
1162 leaf
->data_len
>= HAMMER_BUFSIZE
) {
1164 * We are using the vnode's bio to write directly to the
1165 * media, any hammer_buffer at the same zone-X offset will
1166 * now have stale data.
1168 zone2_offset
= hammer_blockmap_lookup(hmp
, buf_offset
, &error
);
1169 vol_no
= HAMMER_VOL_DECODE(zone2_offset
);
1170 volume
= hammer_get_volume(hmp
, vol_no
, &error
);
1172 if (error
== 0 && zone2_offset
>= volume
->maxbuf_off
)
1176 KKASSERT((bp
->b_bufsize
& HAMMER_BUFMASK
) == 0);
1178 hammer_del_buffers(hmp, buf_offset,
1179 zone2_offset, bp->b_bufsize);
1183 * Second level bio - cached zone2 offset.
1185 * (We can put our bio_done function in either the
1186 * 2nd or 3rd level).
1188 nbio
= push_bio(bio
);
1189 nbio
->bio_offset
= zone2_offset
;
1190 nbio
->bio_done
= hammer_io_direct_write_complete
;
1191 nbio
->bio_caller_info1
.ptr
= record
;
1192 record
->zone2_offset
= zone2_offset
;
1193 record
->flags
|= HAMMER_RECF_DIRECT_IO
|
1194 HAMMER_RECF_DIRECT_INVAL
;
1197 * Third level bio - raw offset specific to the
1200 zone2_offset
&= HAMMER_OFF_SHORT_MASK
;
1201 nbio
= push_bio(nbio
);
1202 nbio
->bio_offset
= volume
->ondisk
->vol_buf_beg
+
1204 hammer_stats_disk_write
+= bp
->b_bufsize
;
1205 vn_strategy(volume
->devvp
, nbio
);
1207 hammer_rel_volume(volume
, 0);
1210 * Must fit in a standard HAMMER buffer. In this case all
1211 * consumers use the HAMMER buffer system and RECF_DIRECT_IO
1212 * does not need to be set-up.
1214 KKASSERT(((buf_offset
^ (buf_offset
+ leaf
->data_len
- 1)) & ~HAMMER_BUFMASK64
) == 0);
1216 ptr
= hammer_bread(hmp
, buf_offset
, &error
, &buffer
);
1219 bp
->b_flags
|= B_AGE
;
1220 hammer_io_modify(&buffer
->io
, 1);
1221 bcopy(bp
->b_data
, ptr
, leaf
->data_len
);
1222 hammer_io_modify_done(&buffer
->io
);
1223 hammer_rel_buffer(buffer
, 0);
1230 * The record is all setup now, add it. Potential conflics
1231 * have already been dealt with.
1233 error
= hammer_mem_add(record
);
1234 KKASSERT(error
== 0);
1237 * Major suckage occured.
1239 kprintf("hammer_direct_write: failed @ %016llx\n",
1244 bp
->b_flags
|= B_ERROR
;
1246 record
->flags
|= HAMMER_RECF_DELETED_FE
;
1247 hammer_rel_mem_record(record
);
1253 * On completion of the BIO this callback must disconnect
1254 * it from the hammer_record and chain to the previous bio.
1256 * An I/O error forces the mount to read-only. Data buffers
1257 * are not B_LOCKED like meta-data buffers are, so we have to
1258 * throw the buffer away to prevent the kernel from retrying.
1262 hammer_io_direct_write_complete(struct bio
*nbio
)
1266 hammer_record_t record
= nbio
->bio_caller_info1
.ptr
;
1269 obio
= pop_bio(nbio
);
1270 if (bp
->b_flags
& B_ERROR
) {
1271 hammer_critical_error(record
->ip
->hmp
, record
->ip
,
1273 "while writing bulk data");
1274 bp
->b_flags
|= B_INVAL
;
1278 KKASSERT(record
!= NULL
);
1279 KKASSERT(record
->flags
& HAMMER_RECF_DIRECT_IO
);
1280 record
->flags
&= ~HAMMER_RECF_DIRECT_IO
;
1281 if (record
->flags
& HAMMER_RECF_DIRECT_WAIT
) {
1282 record
->flags
&= ~HAMMER_RECF_DIRECT_WAIT
;
1283 wakeup(&record
->flags
);
1289 * This is called before a record is either committed to the B-Tree
1290 * or destroyed, to resolve any associated direct-IO.
1292 * (1) We must wait for any direct-IO related to the record to complete.
1294 * (2) We must remove any buffer cache aliases for data accessed via
1295 * leaf->data_offset or zone2_offset so non-direct-IO consumers
1296 * (the mirroring and reblocking code) do not see stale data.
1299 hammer_io_direct_wait(hammer_record_t record
)
1302 * Wait for I/O to complete
1304 if (record
->flags
& HAMMER_RECF_DIRECT_IO
) {
1306 while (record
->flags
& HAMMER_RECF_DIRECT_IO
) {
1307 record
->flags
|= HAMMER_RECF_DIRECT_WAIT
;
1308 tsleep(&record
->flags
, 0, "hmdiow", 0);
1314 * Invalidate any related buffer cache aliases.
1316 if (record
->flags
& HAMMER_RECF_DIRECT_INVAL
) {
1317 KKASSERT(record
->leaf
.data_offset
);
1318 hammer_del_buffers(record
->ip
->hmp
,
1319 record
->leaf
.data_offset
,
1320 record
->zone2_offset
,
1321 record
->leaf
.data_len
);
1322 record
->flags
&= ~HAMMER_RECF_DIRECT_INVAL
;
1327 * This is called to remove the second-level cached zone-2 offset from
1328 * frontend buffer cache buffers, now stale due to a data relocation.
1329 * These offsets are generated by cluster_read() via VOP_BMAP, or directly
1330 * by hammer_vop_strategy_read().
1332 * This is rather nasty because here we have something like the reblocker
1333 * scanning the raw B-Tree with no held references on anything, really,
1334 * other then a shared lock on the B-Tree node, and we have to access the
1335 * frontend's buffer cache to check for and clean out the association.
1336 * Specifically, if the reblocker is moving data on the disk, these cached
1337 * offsets will become invalid.
1339 * Only data record types associated with the large-data zone are subject
1340 * to direct-io and need to be checked.
1344 hammer_io_direct_uncache(hammer_mount_t hmp
, hammer_btree_leaf_elm_t leaf
)
1346 struct hammer_inode_info iinfo
;
1349 if (leaf
->base
.rec_type
!= HAMMER_RECTYPE_DATA
)
1351 zone
= HAMMER_ZONE_DECODE(leaf
->data_offset
);
1352 if (zone
!= HAMMER_ZONE_LARGE_DATA_INDEX
)
1354 iinfo
.obj_id
= leaf
->base
.obj_id
;
1355 iinfo
.obj_asof
= 0; /* unused */
1356 iinfo
.obj_localization
= leaf
->base
.localization
&
1357 HAMMER_LOCALIZE_PSEUDOFS_MASK
;
1358 iinfo
.u
.leaf
= leaf
;
1359 hammer_scan_inode_snapshots(hmp
, &iinfo
,
1360 hammer_io_direct_uncache_callback
,
1365 hammer_io_direct_uncache_callback(hammer_inode_t ip
, void *data
)
1367 hammer_inode_info_t iinfo
= data
;
1368 hammer_off_t data_offset
;
1369 hammer_off_t file_offset
;
1376 data_offset
= iinfo
->u
.leaf
->data_offset
;
1377 file_offset
= iinfo
->u
.leaf
->base
.key
- iinfo
->u
.leaf
->data_len
;
1378 blksize
= iinfo
->u
.leaf
->data_len
;
1379 KKASSERT((blksize
& HAMMER_BUFMASK
) == 0);
1381 hammer_ref(&ip
->lock
);
1382 if (hammer_get_vnode(ip
, &vp
) == 0) {
1383 if ((bp
= findblk(ip
->vp
, file_offset
)) != NULL
&&
1384 bp
->b_bio2
.bio_offset
!= NOOFFSET
) {
1385 bp
= getblk(ip
->vp
, file_offset
, blksize
, 0, 0);
1386 bp
->b_bio2
.bio_offset
= NOOFFSET
;
1391 hammer_rel_inode(ip
, 0);