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 1.49.2.2 2008/07/18 00:21:09 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
254 hammer_io_inval(hammer_volume_t volume
, hammer_off_t zone2_offset
)
256 hammer_io_structure_t iou
;
257 hammer_off_t phys_offset
;
260 phys_offset
= volume
->ondisk
->vol_buf_beg
+
261 (zone2_offset
& HAMMER_OFF_SHORT_MASK
);
263 if ((bp
= findblk(volume
->devvp
, phys_offset
)) != NULL
) {
264 bp
= getblk(volume
->devvp
, phys_offset
, bp
->b_bufsize
, 0, 0);
265 if ((iou
= (void *)LIST_FIRST(&bp
->b_dep
)) != NULL
) {
266 hammer_io_clear_modify(&iou
->io
, 1);
269 hammer_io_deallocate(bp
);
271 KKASSERT((bp
->b_flags
& B_LOCKED
) == 0);
273 bp
->b_flags
|= B_NOCACHE
|B_RELBUF
;
281 * This routine is called on the last reference to a hammer structure.
282 * The io is usually interlocked with io.loading and io.refs must be 1.
284 * This routine may return a non-NULL bp to the caller for dispoal. Disposal
285 * simply means the caller finishes decrementing the ref-count on the
286 * IO structure then brelse()'s the bp. The bp may or may not still be
287 * passively associated with the IO.
289 * The only requirement here is that modified meta-data and volume-header
290 * buffer may NOT be disassociated from the IO structure, and consequently
291 * we also leave such buffers actively associated with the IO if they already
292 * are (since the kernel can't do anything with them anyway). Only the
293 * flusher is allowed to write such buffers out. Modified pure-data and
294 * undo buffers are returned to the kernel but left passively associated
295 * so we can track when the kernel writes the bp out.
298 hammer_io_release(struct hammer_io
*io
, int flush
)
300 union hammer_io_structure
*iou
= (void *)io
;
303 if ((bp
= io
->bp
) == NULL
)
307 * Try to flush a dirty IO to disk if asked to by the
308 * caller or if the kernel tried to flush the buffer in the past.
310 * Kernel-initiated flushes are only allowed for pure-data buffers.
311 * meta-data and volume buffers can only be flushed explicitly
317 } else if (bp
->b_flags
& B_LOCKED
) {
319 case HAMMER_STRUCTURE_DATA_BUFFER
:
320 case HAMMER_STRUCTURE_UNDO_BUFFER
:
326 } /* else no explicit request to flush the buffer */
330 * Wait for the IO to complete if asked to.
332 if (io
->waitdep
&& io
->running
) {
337 * Return control of the buffer to the kernel (with the provisio
338 * that our bioops can override kernel decisions with regards to
341 if ((flush
|| io
->reclaim
) && io
->modified
== 0 && io
->running
== 0) {
343 * Always disassociate the bp if an explicit flush
344 * was requested and the IO completed with no error
345 * (so unmount can really clean up the structure).
353 hammer_io_disassociate((hammer_io_structure_t
)io
);
355 } else if (io
->modified
) {
357 * Only certain IO types can be released to the kernel if
358 * the buffer has been modified.
360 * volume and meta-data IO types may only be explicitly
364 case HAMMER_STRUCTURE_DATA_BUFFER
:
365 case HAMMER_STRUCTURE_UNDO_BUFFER
:
366 if (io
->released
== 0) {
374 bp
= NULL
; /* bp left associated */
375 } else if (io
->released
== 0) {
377 * Clean buffers can be generally released to the kernel.
378 * We leave the bp passively associated with the HAMMER
379 * structure and use bioops to disconnect it later on
380 * if the kernel wants to discard the buffer.
382 * We can steal the structure's ownership of the bp.
385 if (bp
->b_flags
& B_LOCKED
) {
386 hammer_io_disassociate(iou
);
390 hammer_io_disassociate(iou
);
393 /* return the bp (bp passively associated) */
398 * A released buffer is passively associate with our
399 * hammer_io structure. The kernel cannot destroy it
400 * without making a bioops call. If the kernel (B_LOCKED)
401 * or we (reclaim) requested that the buffer be destroyed
402 * we destroy it, otherwise we do a quick get/release to
403 * reset its position in the kernel's LRU list.
405 * Leaving the buffer passively associated allows us to
406 * use the kernel's LRU buffer flushing mechanisms rather
407 * then rolling our own.
409 * XXX there are two ways of doing this. We can re-acquire
410 * and passively release to reset the LRU, or not.
412 if (io
->running
== 0) {
414 if ((bp
->b_flags
& B_LOCKED
) || io
->reclaim
) {
415 hammer_io_disassociate(iou
);
418 /* return the bp (bp passively associated) */
422 * bp is left passively associated but we do not
423 * try to reacquire it. Interactions with the io
424 * structure will occur on completion of the bp's
434 * This routine is called with a locked IO when a flush is desired and
435 * no other references to the structure exists other then ours. This
436 * routine is ONLY called when HAMMER believes it is safe to flush a
437 * potentially modified buffer out.
440 hammer_io_flush(struct hammer_io
*io
)
445 * Degenerate case - nothing to flush if nothing is dirty.
447 if (io
->modified
== 0) {
452 KKASSERT(io
->modify_refs
<= 0);
455 * Acquire ownership of the bp, particularly before we clear our
458 * We are going to bawrite() this bp. Don't leave a window where
459 * io->released is set, we actually own the bp rather then our
465 /* BUF_KERNPROC(io->bp); */
466 /* io->released = 0; */
467 KKASSERT(io
->released
);
468 KKASSERT(io
->bp
== bp
);
473 * Acquire exclusive access to the bp and then clear the modified
474 * state of the buffer prior to issuing I/O to interlock any
475 * modifications made while the I/O is in progress. This shouldn't
476 * happen anyway but losing data would be worse. The modified bit
477 * will be rechecked after the IO completes.
479 * NOTE: This call also finalizes the buffer's content (inval == 0).
481 * This is only legal when lock.refs == 1 (otherwise we might clear
482 * the modified bit while there are still users of the cluster
483 * modifying the data).
485 * Do this before potentially blocking so any attempt to modify the
486 * ondisk while we are blocked blocks waiting for us.
488 hammer_io_clear_modify(io
, 0);
491 * Transfer ownership to the kernel and initiate I/O.
494 io
->hmp
->io_running_space
+= io
->bytes
;
495 hammer_count_io_running_write
+= io
->bytes
;
499 /************************************************************************
501 ************************************************************************
503 * These routines deal with dependancies created when IO buffers get
504 * modified. The caller must call hammer_modify_*() on a referenced
505 * HAMMER structure prior to modifying its on-disk data.
507 * Any intent to modify an IO buffer acquires the related bp and imposes
508 * various write ordering dependancies.
512 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
513 * are locked until the flusher can deal with them, pure data buffers
514 * can be written out.
518 hammer_io_modify(hammer_io_t io
, int count
)
521 * io->modify_refs must be >= 0
523 while (io
->modify_refs
< 0) {
525 tsleep(io
, 0, "hmrmod", 0);
529 * Shortcut if nothing to do.
531 KKASSERT(io
->lock
.refs
!= 0 && io
->bp
!= NULL
);
532 io
->modify_refs
+= count
;
533 if (io
->modified
&& io
->released
== 0)
536 hammer_lock_ex(&io
->lock
);
537 if (io
->modified
== 0) {
538 hammer_io_set_modlist(io
);
543 BUF_KERNPROC(io
->bp
);
545 KKASSERT(io
->modified
!= 0);
547 hammer_unlock(&io
->lock
);
552 hammer_io_modify_done(hammer_io_t io
)
554 KKASSERT(io
->modify_refs
> 0);
556 if (io
->modify_refs
== 0 && io
->waitmod
) {
563 hammer_io_write_interlock(hammer_io_t io
)
565 while (io
->modify_refs
!= 0) {
567 tsleep(io
, 0, "hmrmod", 0);
569 io
->modify_refs
= -1;
573 hammer_io_done_interlock(hammer_io_t io
)
575 KKASSERT(io
->modify_refs
== -1);
584 * Caller intends to modify a volume's ondisk structure.
586 * This is only allowed if we are the flusher or we have a ref on the
590 hammer_modify_volume(hammer_transaction_t trans
, hammer_volume_t volume
,
593 KKASSERT (trans
== NULL
|| trans
->sync_lock_refs
> 0);
595 hammer_io_modify(&volume
->io
, 1);
597 intptr_t rel_offset
= (intptr_t)base
- (intptr_t)volume
->ondisk
;
598 KKASSERT((rel_offset
& ~(intptr_t)HAMMER_BUFMASK
) == 0);
599 hammer_generate_undo(trans
, &volume
->io
,
600 HAMMER_ENCODE_RAW_VOLUME(volume
->vol_no
, rel_offset
),
606 * Caller intends to modify a buffer's ondisk structure.
608 * This is only allowed if we are the flusher or we have a ref on the
612 hammer_modify_buffer(hammer_transaction_t trans
, hammer_buffer_t buffer
,
615 KKASSERT (trans
== NULL
|| trans
->sync_lock_refs
> 0);
617 hammer_io_modify(&buffer
->io
, 1);
619 intptr_t rel_offset
= (intptr_t)base
- (intptr_t)buffer
->ondisk
;
620 KKASSERT((rel_offset
& ~(intptr_t)HAMMER_BUFMASK
) == 0);
621 hammer_generate_undo(trans
, &buffer
->io
,
622 buffer
->zone2_offset
+ rel_offset
,
628 hammer_modify_volume_done(hammer_volume_t volume
)
630 hammer_io_modify_done(&volume
->io
);
634 hammer_modify_buffer_done(hammer_buffer_t buffer
)
636 hammer_io_modify_done(&buffer
->io
);
640 * Mark an entity as not being dirty any more and finalize any
641 * delayed adjustments to the buffer.
643 * Delayed adjustments are an important performance enhancement, allowing
644 * us to avoid recalculating B-Tree node CRCs over and over again when
645 * making bulk-modifications to the B-Tree.
647 * If inval is non-zero delayed adjustments are ignored.
650 hammer_io_clear_modify(struct hammer_io
*io
, int inval
)
652 if (io
->modified
== 0)
656 * Take us off the mod-list and clear the modified bit.
658 KKASSERT(io
->mod_list
!= NULL
);
659 if (io
->mod_list
== &io
->hmp
->volu_list
||
660 io
->mod_list
== &io
->hmp
->meta_list
) {
661 io
->hmp
->locked_dirty_space
-= io
->bytes
;
662 hammer_count_dirtybufspace
-= io
->bytes
;
664 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
669 * If this bit is not set there are no delayed adjustments.
676 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference
677 * on the node (& underlying buffer). Release the node after clearing
680 if (io
->type
== HAMMER_STRUCTURE_META_BUFFER
) {
681 hammer_buffer_t buffer
= (void *)io
;
685 TAILQ_FOREACH(node
, &buffer
->clist
, entry
) {
686 if ((node
->flags
& HAMMER_NODE_NEEDSCRC
) == 0)
688 node
->flags
&= ~HAMMER_NODE_NEEDSCRC
;
689 KKASSERT(node
->ondisk
);
691 node
->ondisk
->crc
= crc32(&node
->ondisk
->crc
+ 1, HAMMER_BTREE_CRCSIZE
);
692 hammer_rel_node(node
);
700 * Clear the IO's modify list. Even though the IO is no longer modified
701 * it may still be on the lose_list. This routine is called just before
702 * the governing hammer_buffer is destroyed.
705 hammer_io_clear_modlist(struct hammer_io
*io
)
707 KKASSERT(io
->modified
== 0);
709 crit_enter(); /* biodone race against list */
710 KKASSERT(io
->mod_list
== &io
->hmp
->lose_list
);
711 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
718 hammer_io_set_modlist(struct hammer_io
*io
)
720 struct hammer_mount
*hmp
= io
->hmp
;
722 KKASSERT(io
->mod_list
== NULL
);
725 case HAMMER_STRUCTURE_VOLUME
:
726 io
->mod_list
= &hmp
->volu_list
;
727 hmp
->locked_dirty_space
+= io
->bytes
;
728 hammer_count_dirtybufspace
+= io
->bytes
;
730 case HAMMER_STRUCTURE_META_BUFFER
:
731 io
->mod_list
= &hmp
->meta_list
;
732 hmp
->locked_dirty_space
+= io
->bytes
;
733 hammer_count_dirtybufspace
+= io
->bytes
;
735 case HAMMER_STRUCTURE_UNDO_BUFFER
:
736 io
->mod_list
= &hmp
->undo_list
;
738 case HAMMER_STRUCTURE_DATA_BUFFER
:
739 io
->mod_list
= &hmp
->data_list
;
742 TAILQ_INSERT_TAIL(io
->mod_list
, io
, mod_entry
);
745 /************************************************************************
747 ************************************************************************
752 * Pre-IO initiation kernel callback - cluster build only
755 hammer_io_start(struct buf
*bp
)
760 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT!
762 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
763 * may also be set if we were marking a cluster header open. Only remove
764 * our dependancy if the modified bit is clear.
767 hammer_io_complete(struct buf
*bp
)
769 union hammer_io_structure
*iou
= (void *)LIST_FIRST(&bp
->b_dep
);
771 KKASSERT(iou
->io
.released
== 1);
774 * Deal with people waiting for I/O to drain
776 if (iou
->io
.running
) {
778 * Deal with critical write errors. Once a critical error
779 * has been flagged in hmp the UNDO FIFO will not be updated.
780 * That way crash recover will give us a consistent
783 * Because of this we can throw away failed UNDO buffers. If
784 * we throw away META or DATA buffers we risk corrupting
785 * the now read-only version of the filesystem visible to
786 * the user. Clear B_ERROR so the buffer is not re-dirtied
787 * by the kernel and ref the io so it doesn't get thrown
790 if (bp
->b_flags
& B_ERROR
) {
791 hammer_critical_error(iou
->io
.hmp
, NULL
, bp
->b_error
,
792 "while flushing meta-data");
793 switch(iou
->io
.type
) {
794 case HAMMER_STRUCTURE_UNDO_BUFFER
:
797 if (iou
->io
.ioerror
== 0) {
799 if (iou
->io
.lock
.refs
== 0)
800 ++hammer_count_refedbufs
;
801 hammer_ref(&iou
->io
.lock
);
805 bp
->b_flags
&= ~B_ERROR
;
808 hammer_io_set_modlist(&iou
->io
);
809 iou
->io
.modified
= 1;
812 hammer_stats_disk_write
+= iou
->io
.bytes
;
813 hammer_count_io_running_write
-= iou
->io
.bytes
;
814 iou
->io
.hmp
->io_running_space
-= iou
->io
.bytes
;
815 if (iou
->io
.hmp
->io_running_space
== 0)
816 wakeup(&iou
->io
.hmp
->io_running_space
);
817 KKASSERT(iou
->io
.hmp
->io_running_space
>= 0);
820 hammer_stats_disk_read
+= iou
->io
.bytes
;
823 if (iou
->io
.waiting
) {
829 * If B_LOCKED is set someone wanted to deallocate the bp at some
830 * point, do it now if refs has become zero.
832 if ((bp
->b_flags
& B_LOCKED
) && iou
->io
.lock
.refs
== 0) {
833 KKASSERT(iou
->io
.modified
== 0);
834 --hammer_count_io_locked
;
835 bp
->b_flags
&= ~B_LOCKED
;
836 hammer_io_deallocate(bp
);
837 /* structure may be dead now */
842 * Callback from kernel when it wishes to deallocate a passively
843 * associated structure. This mostly occurs with clean buffers
844 * but it may be possible for a holding structure to be marked dirty
845 * while its buffer is passively associated. The caller owns the bp.
847 * If we cannot disassociate we set B_LOCKED to prevent the buffer
848 * from getting reused.
850 * WARNING: Because this can be called directly by getnewbuf we cannot
851 * recurse into the tree. If a bp cannot be immediately disassociated
852 * our only recourse is to set B_LOCKED.
854 * WARNING: This may be called from an interrupt via hammer_io_complete()
857 hammer_io_deallocate(struct buf
*bp
)
859 hammer_io_structure_t iou
= (void *)LIST_FIRST(&bp
->b_dep
);
861 KKASSERT((bp
->b_flags
& B_LOCKED
) == 0 && iou
->io
.running
== 0);
862 if (iou
->io
.lock
.refs
> 0 || iou
->io
.modified
) {
864 * It is not legal to disassociate a modified buffer. This
865 * case really shouldn't ever occur.
867 bp
->b_flags
|= B_LOCKED
;
868 ++hammer_count_io_locked
;
871 * Disassociate the BP. If the io has no refs left we
872 * have to add it to the loose list.
874 hammer_io_disassociate(iou
);
875 if (iou
->io
.type
!= HAMMER_STRUCTURE_VOLUME
) {
876 KKASSERT(iou
->io
.bp
== NULL
);
877 KKASSERT(iou
->io
.mod_list
== NULL
);
878 crit_enter(); /* biodone race against list */
879 iou
->io
.mod_list
= &iou
->io
.hmp
->lose_list
;
880 TAILQ_INSERT_TAIL(iou
->io
.mod_list
, &iou
->io
, mod_entry
);
887 hammer_io_fsync(struct vnode
*vp
)
893 * NOTE: will not be called unless we tell the kernel about the
894 * bioops. Unused... we use the mount's VFS_SYNC instead.
897 hammer_io_sync(struct mount
*mp
)
903 hammer_io_movedeps(struct buf
*bp1
, struct buf
*bp2
)
908 * I/O pre-check for reading and writing. HAMMER only uses this for
909 * B_CACHE buffers so checkread just shouldn't happen, but if it does
912 * Writing is a different case. We don't want the kernel to try to write
913 * out a buffer that HAMMER may be modifying passively or which has a
914 * dependancy. In addition, kernel-demanded writes can only proceed for
915 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
916 * buffer types can only be explicitly written by the flusher.
918 * checkwrite will only be called for bdwrite()n buffers. If we return
919 * success the kernel is guaranteed to initiate the buffer write.
922 hammer_io_checkread(struct buf
*bp
)
928 hammer_io_checkwrite(struct buf
*bp
)
930 hammer_io_t io
= (void *)LIST_FIRST(&bp
->b_dep
);
933 * This shouldn't happen under normal operation.
935 if (io
->type
== HAMMER_STRUCTURE_VOLUME
||
936 io
->type
== HAMMER_STRUCTURE_META_BUFFER
) {
938 panic("hammer_io_checkwrite: illegal buffer");
939 if ((bp
->b_flags
& B_LOCKED
) == 0) {
940 bp
->b_flags
|= B_LOCKED
;
941 ++hammer_count_io_locked
;
947 * We can only clear the modified bit if the IO is not currently
948 * undergoing modification. Otherwise we may miss changes.
950 if (io
->modify_refs
== 0 && io
->modified
)
951 hammer_io_clear_modify(io
, 0);
954 * The kernel is going to start the IO, set io->running.
956 KKASSERT(io
->running
== 0);
958 io
->hmp
->io_running_space
+= io
->bytes
;
959 hammer_count_io_running_write
+= io
->bytes
;
964 * Return non-zero if we wish to delay the kernel's attempt to flush
965 * this buffer to disk.
968 hammer_io_countdeps(struct buf
*bp
, int n
)
973 struct bio_ops hammer_bioops
= {
974 .io_start
= hammer_io_start
,
975 .io_complete
= hammer_io_complete
,
976 .io_deallocate
= hammer_io_deallocate
,
977 .io_fsync
= hammer_io_fsync
,
978 .io_sync
= hammer_io_sync
,
979 .io_movedeps
= hammer_io_movedeps
,
980 .io_countdeps
= hammer_io_countdeps
,
981 .io_checkread
= hammer_io_checkread
,
982 .io_checkwrite
= hammer_io_checkwrite
,
985 /************************************************************************
987 ************************************************************************
989 * These functions operate directly on the buffer cache buffer associated
990 * with a front-end vnode rather then a back-end device vnode.
994 * Read a buffer associated with a front-end vnode directly from the
995 * disk media. The bio may be issued asynchronously. If leaf is non-NULL
996 * we validate the CRC.
998 * A second-level bio already resolved to a zone-2 offset (typically by
999 * the BMAP code, or by a previous hammer_io_direct_write()), is passed.
1001 * We must check for the presence of a HAMMER buffer to handle the case
1002 * where the reblocker has rewritten the data (which it does via the HAMMER
1003 * buffer system, not via the high-level vnode buffer cache), but not yet
1004 * committed the buffer to the media.
1007 hammer_io_direct_read(hammer_mount_t hmp
, struct bio
*bio
,
1008 hammer_btree_leaf_elm_t leaf
)
1010 hammer_off_t buf_offset
;
1011 hammer_off_t zone2_offset
;
1012 hammer_volume_t volume
;
1018 buf_offset
= bio
->bio_offset
;
1019 KKASSERT((buf_offset
& HAMMER_OFF_ZONE_MASK
) ==
1020 HAMMER_ZONE_LARGE_DATA
);
1023 * The buffer cache may have an aliased buffer (the reblocker can
1024 * write them). If it does we have to sync any dirty data before
1025 * we can build our direct-read. This is a non-critical code path.
1028 hammer_sync_buffers(hmp
, buf_offset
, bp
->b_bufsize
);
1031 * Resolve to a zone-2 offset. The conversion just requires
1032 * munging the top 4 bits but we want to abstract it anyway
1033 * so the blockmap code can verify the zone assignment.
1035 zone2_offset
= hammer_blockmap_lookup(hmp
, buf_offset
, &error
);
1038 KKASSERT((zone2_offset
& HAMMER_OFF_ZONE_MASK
) ==
1039 HAMMER_ZONE_RAW_BUFFER
);
1042 * Resolve volume and raw-offset for 3rd level bio. The
1043 * offset will be specific to the volume.
1045 vol_no
= HAMMER_VOL_DECODE(zone2_offset
);
1046 volume
= hammer_get_volume(hmp
, vol_no
, &error
);
1047 if (error
== 0 && zone2_offset
>= volume
->maxbuf_off
)
1051 zone2_offset
&= HAMMER_OFF_SHORT_MASK
;
1053 nbio
= push_bio(bio
);
1054 nbio
->bio_offset
= volume
->ondisk
->vol_buf_beg
+
1058 * XXX disabled - our CRC check doesn't work if the OS
1059 * does bogus_page replacement on the direct-read.
1061 if (leaf
&& hammer_verify_data
) {
1062 nbio
->bio_done
= hammer_io_direct_read_complete
;
1063 nbio
->bio_caller_info1
.uvalue32
= leaf
->data_crc
;
1066 hammer_stats_disk_read
+= bp
->b_bufsize
;
1067 vn_strategy(volume
->devvp
, nbio
);
1069 hammer_rel_volume(volume
, 0);
1072 kprintf("hammer_direct_read: failed @ %016llx\n",
1074 bp
->b_error
= error
;
1075 bp
->b_flags
|= B_ERROR
;
1083 * On completion of the BIO this callback must check the data CRC
1084 * and chain to the previous bio.
1088 hammer_io_direct_read_complete(struct bio
*nbio
)
1092 u_int32_t rec_crc
= nbio
->bio_caller_info1
.uvalue32
;
1095 if (crc32(bp
->b_data
, bp
->b_bufsize
) != rec_crc
) {
1096 kprintf("HAMMER: data_crc error @%016llx/%d\n",
1097 nbio
->bio_offset
, bp
->b_bufsize
);
1098 if (hammer_debug_debug
)
1100 bp
->b_flags
|= B_ERROR
;
1103 obio
= pop_bio(nbio
);
1109 * Write a buffer associated with a front-end vnode directly to the
1110 * disk media. The bio may be issued asynchronously.
1112 * The BIO is associated with the specified record and RECF_DIRECT_IO
1116 hammer_io_direct_write(hammer_mount_t hmp
, hammer_record_t record
,
1119 hammer_btree_leaf_elm_t leaf
= &record
->leaf
;
1120 hammer_off_t buf_offset
;
1121 hammer_off_t zone2_offset
;
1122 hammer_volume_t volume
;
1123 hammer_buffer_t buffer
;
1130 buf_offset
= leaf
->data_offset
;
1132 KKASSERT(buf_offset
> HAMMER_ZONE_BTREE
);
1133 KKASSERT(bio
->bio_buf
->b_cmd
== BUF_CMD_WRITE
);
1135 if ((buf_offset
& HAMMER_BUFMASK
) == 0 &&
1136 leaf
->data_len
>= HAMMER_BUFSIZE
) {
1138 * We are using the vnode's bio to write directly to the
1139 * media, any hammer_buffer at the same zone-X offset will
1140 * now have stale data.
1142 zone2_offset
= hammer_blockmap_lookup(hmp
, buf_offset
, &error
);
1143 vol_no
= HAMMER_VOL_DECODE(zone2_offset
);
1144 volume
= hammer_get_volume(hmp
, vol_no
, &error
);
1146 if (error
== 0 && zone2_offset
>= volume
->maxbuf_off
)
1150 KKASSERT((bp
->b_bufsize
& HAMMER_BUFMASK
) == 0);
1151 hammer_del_buffers(hmp
, buf_offset
,
1152 zone2_offset
, bp
->b_bufsize
);
1155 * Second level bio - cached zone2 offset.
1157 * (We can put our bio_done function in either the
1158 * 2nd or 3rd level).
1160 nbio
= push_bio(bio
);
1161 nbio
->bio_offset
= zone2_offset
;
1162 nbio
->bio_done
= hammer_io_direct_write_complete
;
1163 nbio
->bio_caller_info1
.ptr
= record
;
1164 record
->flags
|= HAMMER_RECF_DIRECT_IO
;
1167 * Third level bio - raw offset specific to the
1170 zone2_offset
&= HAMMER_OFF_SHORT_MASK
;
1171 nbio
= push_bio(nbio
);
1172 nbio
->bio_offset
= volume
->ondisk
->vol_buf_beg
+
1174 hammer_stats_disk_write
+= bp
->b_bufsize
;
1175 vn_strategy(volume
->devvp
, nbio
);
1177 hammer_rel_volume(volume
, 0);
1180 * Must fit in a standard HAMMER buffer. In this case all
1181 * consumers use the HAMMER buffer system and RECF_DIRECT_IO
1182 * does not need to be set-up.
1184 KKASSERT(((buf_offset
^ (buf_offset
+ leaf
->data_len
- 1)) & ~HAMMER_BUFMASK64
) == 0);
1186 ptr
= hammer_bread(hmp
, buf_offset
, &error
, &buffer
);
1189 bp
->b_flags
|= B_AGE
;
1190 hammer_io_modify(&buffer
->io
, 1);
1191 bcopy(bp
->b_data
, ptr
, leaf
->data_len
);
1192 hammer_io_modify_done(&buffer
->io
);
1193 hammer_rel_buffer(buffer
, 0);
1199 kprintf("hammer_direct_write: failed @ %016llx\n",
1204 bp
->b_flags
|= B_ERROR
;
1211 * On completion of the BIO this callback must disconnect
1212 * it from the hammer_record and chain to the previous bio.
1214 * An I/O error forces the mount to read-only. Data buffers
1215 * are not B_LOCKED like meta-data buffers are, so we have to
1216 * throw the buffer away to prevent the kernel from retrying.
1220 hammer_io_direct_write_complete(struct bio
*nbio
)
1223 hammer_record_t record
= nbio
->bio_caller_info1
.ptr
;
1225 obio
= pop_bio(nbio
);
1226 if (obio
->bio_buf
->b_flags
& B_ERROR
) {
1227 hammer_critical_error(record
->ip
->hmp
, record
->ip
,
1228 obio
->bio_buf
->b_error
,
1229 "while writing bulk data");
1230 obio
->bio_buf
->b_flags
|= B_INVAL
;
1233 KKASSERT(record
!= NULL
&& (record
->flags
& HAMMER_RECF_DIRECT_IO
));
1234 record
->flags
&= ~HAMMER_RECF_DIRECT_IO
;
1235 if (record
->flags
& HAMMER_RECF_DIRECT_WAIT
) {
1236 record
->flags
&= ~HAMMER_RECF_DIRECT_WAIT
;
1237 wakeup(&record
->flags
);
1243 * This is called before a record is either committed to the B-Tree
1244 * or destroyed, to resolve any associated direct-IO. We must
1245 * ensure that the data is available on-media to other consumers
1246 * such as the reblocker or mirroring code.
1248 * Note that other consumers might access the data via the block
1249 * device's buffer cache and not the high level vnode's buffer cache.
1252 hammer_io_direct_wait(hammer_record_t record
)
1255 while (record
->flags
& HAMMER_RECF_DIRECT_IO
) {
1256 record
->flags
|= HAMMER_RECF_DIRECT_WAIT
;
1257 tsleep(&record
->flags
, 0, "hmdiow", 0);
1263 * This is called to remove the second-level cached zone-2 offset from
1264 * frontend buffer cache buffers, now stale due to a data relocation.
1265 * These offsets are generated by cluster_read() via VOP_BMAP, or directly
1266 * by hammer_vop_strategy_read().
1268 * This is rather nasty because here we have something like the reblocker
1269 * scanning the raw B-Tree with no held references on anything, really,
1270 * other then a shared lock on the B-Tree node, and we have to access the
1271 * frontend's buffer cache to check for and clean out the association.
1272 * Specifically, if the reblocker is moving data on the disk, these cached
1273 * offsets will become invalid.
1275 * Only data record types associated with the large-data zone are subject
1276 * to direct-io and need to be checked.
1280 hammer_io_direct_uncache(hammer_mount_t hmp
, hammer_btree_leaf_elm_t leaf
)
1282 struct hammer_inode_info iinfo
;
1285 if (leaf
->base
.rec_type
!= HAMMER_RECTYPE_DATA
)
1287 zone
= HAMMER_ZONE_DECODE(leaf
->data_offset
);
1288 if (zone
!= HAMMER_ZONE_LARGE_DATA_INDEX
)
1290 iinfo
.obj_id
= leaf
->base
.obj_id
;
1291 iinfo
.obj_asof
= 0; /* unused */
1292 iinfo
.obj_localization
= leaf
->base
.localization
&
1293 HAMMER_LOCALIZE_PSEUDOFS_MASK
;
1294 iinfo
.u
.leaf
= leaf
;
1295 hammer_scan_inode_snapshots(hmp
, &iinfo
,
1296 hammer_io_direct_uncache_callback
,
1301 hammer_io_direct_uncache_callback(hammer_inode_t ip
, void *data
)
1303 hammer_inode_info_t iinfo
= data
;
1304 hammer_off_t data_offset
;
1305 hammer_off_t file_offset
;
1312 data_offset
= iinfo
->u
.leaf
->data_offset
;
1313 file_offset
= iinfo
->u
.leaf
->base
.key
- iinfo
->u
.leaf
->data_len
;
1314 blksize
= iinfo
->u
.leaf
->data_len
;
1315 KKASSERT((blksize
& HAMMER_BUFMASK
) == 0);
1317 hammer_ref(&ip
->lock
);
1318 if (hammer_get_vnode(ip
, &vp
) == 0) {
1319 if ((bp
= findblk(ip
->vp
, file_offset
)) != NULL
&&
1320 bp
->b_bio2
.bio_offset
!= NOOFFSET
) {
1321 bp
= getblk(ip
->vp
, file_offset
, blksize
, 0, 0);
1322 bp
->b_bio2
.bio_offset
= NOOFFSET
;
1327 hammer_rel_inode(ip
, 0);