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_rel(&io
->lock
);
207 KKASSERT(hammer_isactive(&io
->lock
));
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(hammer_isactive(&iou
->io
.lock
) == 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 must be interlocked with a refcount of zero. The hammer structure
384 * will remain interlocked on return.
386 * This routine may return a non-NULL bp to the caller for dispoal.
387 * The caller typically brelse()'s the bp.
389 * The bp may or may not still be passively associated with the IO. It
390 * will remain passively associated if it is unreleasable (e.g. a modified
393 * The only requirement here is that modified meta-data and volume-header
394 * buffer may NOT be disassociated from the IO structure, and consequently
395 * we also leave such buffers actively associated with the IO if they already
396 * are (since the kernel can't do anything with them anyway). Only the
397 * flusher is allowed to write such buffers out. Modified pure-data and
398 * undo buffers are returned to the kernel but left passively associated
399 * so we can track when the kernel writes the bp out.
402 hammer_io_release(struct hammer_io
*io
, int flush
)
404 union hammer_io_structure
*iou
= (void *)io
;
407 if ((bp
= io
->bp
) == NULL
)
411 * Try to flush a dirty IO to disk if asked to by the
412 * caller or if the kernel tried to flush the buffer in the past.
414 * Kernel-initiated flushes are only allowed for pure-data buffers.
415 * meta-data and volume buffers can only be flushed explicitly
420 hammer_io_flush(io
, 0);
421 } else if (bp
->b_flags
& B_LOCKED
) {
423 case HAMMER_STRUCTURE_DATA_BUFFER
:
424 hammer_io_flush(io
, 0);
426 case HAMMER_STRUCTURE_UNDO_BUFFER
:
427 hammer_io_flush(io
, hammer_undo_reclaim(io
));
432 } /* else no explicit request to flush the buffer */
436 * Wait for the IO to complete if asked to. This occurs when
437 * the buffer must be disposed of definitively during an umount
438 * or buffer invalidation.
440 if (io
->waitdep
&& io
->running
) {
445 * Return control of the buffer to the kernel (with the provisio
446 * that our bioops can override kernel decisions with regards to
449 if ((flush
|| io
->reclaim
) && io
->modified
== 0 && io
->running
== 0) {
451 * Always disassociate the bp if an explicit flush
452 * was requested and the IO completed with no error
453 * (so unmount can really clean up the structure).
461 hammer_io_disassociate((hammer_io_structure_t
)io
);
463 } else if (io
->modified
) {
465 * Only certain IO types can be released to the kernel if
466 * the buffer has been modified.
468 * volume and meta-data IO types may only be explicitly
472 case HAMMER_STRUCTURE_DATA_BUFFER
:
473 case HAMMER_STRUCTURE_UNDO_BUFFER
:
474 if (io
->released
== 0) {
482 bp
= NULL
; /* bp left associated */
483 } else if (io
->released
== 0) {
485 * Clean buffers can be generally released to the kernel.
486 * We leave the bp passively associated with the HAMMER
487 * structure and use bioops to disconnect it later on
488 * if the kernel wants to discard the buffer.
490 * We can steal the structure's ownership of the bp.
493 if (bp
->b_flags
& B_LOCKED
) {
494 hammer_io_disassociate(iou
);
498 hammer_io_disassociate(iou
);
501 /* return the bp (bp passively associated) */
506 * A released buffer is passively associate with our
507 * hammer_io structure. The kernel cannot destroy it
508 * without making a bioops call. If the kernel (B_LOCKED)
509 * or we (reclaim) requested that the buffer be destroyed
510 * we destroy it, otherwise we do a quick get/release to
511 * reset its position in the kernel's LRU list.
513 * Leaving the buffer passively associated allows us to
514 * use the kernel's LRU buffer flushing mechanisms rather
515 * then rolling our own.
517 * XXX there are two ways of doing this. We can re-acquire
518 * and passively release to reset the LRU, or not.
520 if (io
->running
== 0) {
522 if ((bp
->b_flags
& B_LOCKED
) || io
->reclaim
) {
523 hammer_io_disassociate(iou
);
526 /* return the bp (bp passively associated) */
530 * bp is left passively associated but we do not
531 * try to reacquire it. Interactions with the io
532 * structure will occur on completion of the bp's
542 * This routine is called with a locked IO when a flush is desired and
543 * no other references to the structure exists other then ours. This
544 * routine is ONLY called when HAMMER believes it is safe to flush a
545 * potentially modified buffer out.
548 hammer_io_flush(struct hammer_io
*io
, int reclaim
)
553 * Degenerate case - nothing to flush if nothing is dirty.
555 if (io
->modified
== 0) {
560 KKASSERT(io
->modify_refs
<= 0);
563 * Acquire ownership of the bp, particularly before we clear our
566 * We are going to bawrite() this bp. Don't leave a window where
567 * io->released is set, we actually own the bp rather then our
573 /* BUF_KERNPROC(io->bp); */
574 /* io->released = 0; */
575 KKASSERT(io
->released
);
576 KKASSERT(io
->bp
== bp
);
582 if ((bp
->b_flags
& B_LOCKED
) == 0) {
583 bp
->b_flags
|= B_LOCKED
;
584 ++hammer_count_io_locked
;
589 * Acquire exclusive access to the bp and then clear the modified
590 * state of the buffer prior to issuing I/O to interlock any
591 * modifications made while the I/O is in progress. This shouldn't
592 * happen anyway but losing data would be worse. The modified bit
593 * will be rechecked after the IO completes.
595 * NOTE: This call also finalizes the buffer's content (inval == 0).
597 * This is only legal when lock.refs == 1 (otherwise we might clear
598 * the modified bit while there are still users of the cluster
599 * modifying the data).
601 * Do this before potentially blocking so any attempt to modify the
602 * ondisk while we are blocked blocks waiting for us.
604 hammer_ref(&io
->lock
);
605 hammer_io_clear_modify(io
, 0);
606 hammer_rel(&io
->lock
);
608 if (hammer_debug_io
& 0x0002)
609 kprintf("hammer io_write %016jx\n", bp
->b_bio1
.bio_offset
);
612 * Transfer ownership to the kernel and initiate I/O.
615 io
->hmp
->io_running_space
+= io
->bytes
;
616 TAILQ_INSERT_TAIL(&io
->hmp
->iorun_list
, io
, iorun_entry
);
617 hammer_count_io_running_write
+= io
->bytes
;
619 hammer_io_flush_mark(io
->volume
);
622 /************************************************************************
624 ************************************************************************
626 * These routines deal with dependancies created when IO buffers get
627 * modified. The caller must call hammer_modify_*() on a referenced
628 * HAMMER structure prior to modifying its on-disk data.
630 * Any intent to modify an IO buffer acquires the related bp and imposes
631 * various write ordering dependancies.
635 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
636 * are locked until the flusher can deal with them, pure data buffers
637 * can be written out.
641 hammer_io_modify(hammer_io_t io
, int count
)
644 * io->modify_refs must be >= 0
646 while (io
->modify_refs
< 0) {
648 tsleep(io
, 0, "hmrmod", 0);
652 * Shortcut if nothing to do.
654 KKASSERT(hammer_isactive(&io
->lock
) && io
->bp
!= NULL
);
655 io
->modify_refs
+= count
;
656 if (io
->modified
&& io
->released
== 0)
659 hammer_lock_ex(&io
->lock
);
660 if (io
->modified
== 0) {
661 hammer_io_set_modlist(io
);
666 BUF_KERNPROC(io
->bp
);
668 KKASSERT(io
->modified
!= 0);
670 hammer_unlock(&io
->lock
);
675 hammer_io_modify_done(hammer_io_t io
)
677 KKASSERT(io
->modify_refs
> 0);
679 if (io
->modify_refs
== 0 && io
->waitmod
) {
686 hammer_io_write_interlock(hammer_io_t io
)
688 while (io
->modify_refs
!= 0) {
690 tsleep(io
, 0, "hmrmod", 0);
692 io
->modify_refs
= -1;
696 hammer_io_done_interlock(hammer_io_t io
)
698 KKASSERT(io
->modify_refs
== -1);
707 * Caller intends to modify a volume's ondisk structure.
709 * This is only allowed if we are the flusher or we have a ref on the
713 hammer_modify_volume(hammer_transaction_t trans
, hammer_volume_t volume
,
716 KKASSERT (trans
== NULL
|| trans
->sync_lock_refs
> 0);
718 hammer_io_modify(&volume
->io
, 1);
720 intptr_t rel_offset
= (intptr_t)base
- (intptr_t)volume
->ondisk
;
721 KKASSERT((rel_offset
& ~(intptr_t)HAMMER_BUFMASK
) == 0);
722 hammer_generate_undo(trans
,
723 HAMMER_ENCODE_RAW_VOLUME(volume
->vol_no
, rel_offset
),
729 * Caller intends to modify a buffer's ondisk structure.
731 * This is only allowed if we are the flusher or we have a ref on the
735 hammer_modify_buffer(hammer_transaction_t trans
, hammer_buffer_t buffer
,
738 KKASSERT (trans
== NULL
|| trans
->sync_lock_refs
> 0);
740 hammer_io_modify(&buffer
->io
, 1);
742 intptr_t rel_offset
= (intptr_t)base
- (intptr_t)buffer
->ondisk
;
743 KKASSERT((rel_offset
& ~(intptr_t)HAMMER_BUFMASK
) == 0);
744 hammer_generate_undo(trans
,
745 buffer
->zone2_offset
+ rel_offset
,
751 hammer_modify_volume_done(hammer_volume_t volume
)
753 hammer_io_modify_done(&volume
->io
);
757 hammer_modify_buffer_done(hammer_buffer_t buffer
)
759 hammer_io_modify_done(&buffer
->io
);
763 * Mark an entity as not being dirty any more and finalize any
764 * delayed adjustments to the buffer.
766 * Delayed adjustments are an important performance enhancement, allowing
767 * us to avoid recalculating B-Tree node CRCs over and over again when
768 * making bulk-modifications to the B-Tree.
770 * If inval is non-zero delayed adjustments are ignored.
772 * This routine may dereference related btree nodes and cause the
773 * buffer to be dereferenced. The caller must own a reference on io.
776 hammer_io_clear_modify(struct hammer_io
*io
, int inval
)
778 if (io
->modified
== 0)
782 * Take us off the mod-list and clear the modified bit.
784 KKASSERT(io
->mod_list
!= NULL
);
785 if (io
->mod_list
== &io
->hmp
->volu_list
||
786 io
->mod_list
== &io
->hmp
->meta_list
) {
787 io
->hmp
->locked_dirty_space
-= io
->bytes
;
788 hammer_count_dirtybufspace
-= io
->bytes
;
790 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
795 * If this bit is not set there are no delayed adjustments.
802 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference
803 * on the node (& underlying buffer). Release the node after clearing
806 if (io
->type
== HAMMER_STRUCTURE_META_BUFFER
) {
807 hammer_buffer_t buffer
= (void *)io
;
811 TAILQ_FOREACH(node
, &buffer
->clist
, entry
) {
812 if ((node
->flags
& HAMMER_NODE_NEEDSCRC
) == 0)
814 node
->flags
&= ~HAMMER_NODE_NEEDSCRC
;
815 KKASSERT(node
->ondisk
);
817 node
->ondisk
->crc
= crc32(&node
->ondisk
->crc
+ 1, HAMMER_BTREE_CRCSIZE
);
818 hammer_rel_node(node
);
822 /* caller must still have ref on io */
823 KKASSERT(hammer_isactive(&io
->lock
));
827 * Clear the IO's modify list. Even though the IO is no longer modified
828 * it may still be on the lose_list. This routine is called just before
829 * the governing hammer_buffer is destroyed.
832 hammer_io_clear_modlist(struct hammer_io
*io
)
834 KKASSERT(io
->modified
== 0);
836 crit_enter(); /* biodone race against list */
837 KKASSERT(io
->mod_list
== &io
->hmp
->lose_list
);
838 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
845 hammer_io_set_modlist(struct hammer_io
*io
)
847 struct hammer_mount
*hmp
= io
->hmp
;
849 KKASSERT(io
->mod_list
== NULL
);
852 case HAMMER_STRUCTURE_VOLUME
:
853 io
->mod_list
= &hmp
->volu_list
;
854 hmp
->locked_dirty_space
+= io
->bytes
;
855 hammer_count_dirtybufspace
+= io
->bytes
;
857 case HAMMER_STRUCTURE_META_BUFFER
:
858 io
->mod_list
= &hmp
->meta_list
;
859 hmp
->locked_dirty_space
+= io
->bytes
;
860 hammer_count_dirtybufspace
+= io
->bytes
;
862 case HAMMER_STRUCTURE_UNDO_BUFFER
:
863 io
->mod_list
= &hmp
->undo_list
;
865 case HAMMER_STRUCTURE_DATA_BUFFER
:
866 io
->mod_list
= &hmp
->data_list
;
868 case HAMMER_STRUCTURE_DUMMY
:
869 panic("hammer_io_disassociate: bad io type");
872 TAILQ_INSERT_TAIL(io
->mod_list
, io
, mod_entry
);
875 /************************************************************************
877 ************************************************************************
882 * Pre-IO initiation kernel callback - cluster build only
885 hammer_io_start(struct buf
*bp
)
890 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT!
892 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
893 * may also be set if we were marking a cluster header open. Only remove
894 * our dependancy if the modified bit is clear.
897 hammer_io_complete(struct buf
*bp
)
899 union hammer_io_structure
*iou
= (void *)LIST_FIRST(&bp
->b_dep
);
900 struct hammer_io
*ionext
;
902 KKASSERT(iou
->io
.released
== 1);
905 * Deal with people waiting for I/O to drain
907 if (iou
->io
.running
) {
909 * Deal with critical write errors. Once a critical error
910 * has been flagged in hmp the UNDO FIFO will not be updated.
911 * That way crash recover will give us a consistent
914 * Because of this we can throw away failed UNDO buffers. If
915 * we throw away META or DATA buffers we risk corrupting
916 * the now read-only version of the filesystem visible to
917 * the user. Clear B_ERROR so the buffer is not re-dirtied
918 * by the kernel and ref the io so it doesn't get thrown
921 if (bp
->b_flags
& B_ERROR
) {
922 hammer_critical_error(iou
->io
.hmp
, NULL
, bp
->b_error
,
923 "while flushing meta-data");
924 switch(iou
->io
.type
) {
925 case HAMMER_STRUCTURE_UNDO_BUFFER
:
928 if (iou
->io
.ioerror
== 0) {
930 hammer_ref(&iou
->io
.lock
);
934 bp
->b_flags
&= ~B_ERROR
;
937 hammer_io_set_modlist(&iou
->io
);
938 iou
->io
.modified
= 1;
941 hammer_stats_disk_write
+= iou
->io
.bytes
;
942 hammer_count_io_running_write
-= iou
->io
.bytes
;
943 iou
->io
.hmp
->io_running_space
-= iou
->io
.bytes
;
944 KKASSERT(iou
->io
.hmp
->io_running_space
>= 0);
948 * Remove from iorun list and wakeup any multi-io waiter(s).
950 if (TAILQ_FIRST(&iou
->io
.hmp
->iorun_list
) == &iou
->io
) {
951 ionext
= TAILQ_NEXT(&iou
->io
, iorun_entry
);
952 if (ionext
&& ionext
->type
== HAMMER_STRUCTURE_DUMMY
)
955 TAILQ_REMOVE(&iou
->io
.hmp
->iorun_list
, &iou
->io
, iorun_entry
);
957 hammer_stats_disk_read
+= iou
->io
.bytes
;
960 if (iou
->io
.waiting
) {
966 * If B_LOCKED is set someone wanted to deallocate the bp at some
967 * point, try to do it now. The operation will fail if there are
968 * refs or if hammer_io_deallocate() is unable to gain the
971 if (bp
->b_flags
& B_LOCKED
) {
972 --hammer_count_io_locked
;
973 bp
->b_flags
&= ~B_LOCKED
;
974 hammer_io_deallocate(bp
);
975 /* structure may be dead now */
980 * Callback from kernel when it wishes to deallocate a passively
981 * associated structure. This mostly occurs with clean buffers
982 * but it may be possible for a holding structure to be marked dirty
983 * while its buffer is passively associated. The caller owns the bp.
985 * If we cannot disassociate we set B_LOCKED to prevent the buffer
986 * from getting reused.
988 * WARNING: Because this can be called directly by getnewbuf we cannot
989 * recurse into the tree. If a bp cannot be immediately disassociated
990 * our only recourse is to set B_LOCKED.
992 * WARNING: This may be called from an interrupt via hammer_io_complete()
995 hammer_io_deallocate(struct buf
*bp
)
997 hammer_io_structure_t iou
= (void *)LIST_FIRST(&bp
->b_dep
);
999 KKASSERT((bp
->b_flags
& B_LOCKED
) == 0 && iou
->io
.running
== 0);
1000 if (hammer_try_interlock_norefs(&iou
->io
.lock
) == 0) {
1002 * We cannot safely disassociate a bp from a referenced
1003 * or interlocked HAMMER structure.
1005 bp
->b_flags
|= B_LOCKED
;
1006 ++hammer_count_io_locked
;
1007 } else if (iou
->io
.modified
) {
1009 * It is not legal to disassociate a modified buffer. This
1010 * case really shouldn't ever occur.
1012 bp
->b_flags
|= B_LOCKED
;
1013 ++hammer_count_io_locked
;
1014 hammer_put_interlock(&iou
->io
.lock
, 0);
1017 * Disassociate the BP. If the io has no refs left we
1018 * have to add it to the loose list.
1020 hammer_io_disassociate(iou
);
1021 if (iou
->io
.type
!= HAMMER_STRUCTURE_VOLUME
) {
1022 KKASSERT(iou
->io
.bp
== NULL
);
1023 KKASSERT(iou
->io
.mod_list
== NULL
);
1024 crit_enter(); /* biodone race against list */
1025 iou
->io
.mod_list
= &iou
->io
.hmp
->lose_list
;
1026 TAILQ_INSERT_TAIL(iou
->io
.mod_list
, &iou
->io
, mod_entry
);
1029 hammer_put_interlock(&iou
->io
.lock
, 1);
1034 hammer_io_fsync(struct vnode
*vp
)
1040 * NOTE: will not be called unless we tell the kernel about the
1041 * bioops. Unused... we use the mount's VFS_SYNC instead.
1044 hammer_io_sync(struct mount
*mp
)
1050 hammer_io_movedeps(struct buf
*bp1
, struct buf
*bp2
)
1055 * I/O pre-check for reading and writing. HAMMER only uses this for
1056 * B_CACHE buffers so checkread just shouldn't happen, but if it does
1059 * Writing is a different case. We don't want the kernel to try to write
1060 * out a buffer that HAMMER may be modifying passively or which has a
1061 * dependancy. In addition, kernel-demanded writes can only proceed for
1062 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
1063 * buffer types can only be explicitly written by the flusher.
1065 * checkwrite will only be called for bdwrite()n buffers. If we return
1066 * success the kernel is guaranteed to initiate the buffer write.
1069 hammer_io_checkread(struct buf
*bp
)
1075 hammer_io_checkwrite(struct buf
*bp
)
1077 hammer_io_t io
= (void *)LIST_FIRST(&bp
->b_dep
);
1080 * This shouldn't happen under normal operation.
1082 if (io
->type
== HAMMER_STRUCTURE_VOLUME
||
1083 io
->type
== HAMMER_STRUCTURE_META_BUFFER
) {
1085 panic("hammer_io_checkwrite: illegal buffer");
1086 if ((bp
->b_flags
& B_LOCKED
) == 0) {
1087 bp
->b_flags
|= B_LOCKED
;
1088 ++hammer_count_io_locked
;
1094 * We can only clear the modified bit if the IO is not currently
1095 * undergoing modification. Otherwise we may miss changes.
1097 * Only data and undo buffers can reach here. These buffers do
1098 * not have terminal crc functions but we temporarily reference
1099 * the IO anyway, just in case.
1101 if (io
->modify_refs
== 0 && io
->modified
) {
1102 hammer_ref(&io
->lock
);
1103 hammer_io_clear_modify(io
, 0);
1104 hammer_rel(&io
->lock
);
1105 } else if (io
->modified
) {
1106 KKASSERT(io
->type
== HAMMER_STRUCTURE_DATA_BUFFER
);
1110 * The kernel is going to start the IO, set io->running.
1112 KKASSERT(io
->running
== 0);
1114 io
->hmp
->io_running_space
+= io
->bytes
;
1115 TAILQ_INSERT_TAIL(&io
->hmp
->iorun_list
, io
, iorun_entry
);
1116 hammer_count_io_running_write
+= io
->bytes
;
1121 * Return non-zero if we wish to delay the kernel's attempt to flush
1122 * this buffer to disk.
1125 hammer_io_countdeps(struct buf
*bp
, int n
)
1130 struct bio_ops hammer_bioops
= {
1131 .io_start
= hammer_io_start
,
1132 .io_complete
= hammer_io_complete
,
1133 .io_deallocate
= hammer_io_deallocate
,
1134 .io_fsync
= hammer_io_fsync
,
1135 .io_sync
= hammer_io_sync
,
1136 .io_movedeps
= hammer_io_movedeps
,
1137 .io_countdeps
= hammer_io_countdeps
,
1138 .io_checkread
= hammer_io_checkread
,
1139 .io_checkwrite
= hammer_io_checkwrite
,
1142 /************************************************************************
1144 ************************************************************************
1146 * These functions operate directly on the buffer cache buffer associated
1147 * with a front-end vnode rather then a back-end device vnode.
1151 * Read a buffer associated with a front-end vnode directly from the
1152 * disk media. The bio may be issued asynchronously. If leaf is non-NULL
1153 * we validate the CRC.
1155 * We must check for the presence of a HAMMER buffer to handle the case
1156 * where the reblocker has rewritten the data (which it does via the HAMMER
1157 * buffer system, not via the high-level vnode buffer cache), but not yet
1158 * committed the buffer to the media.
1161 hammer_io_direct_read(hammer_mount_t hmp
, struct bio
*bio
,
1162 hammer_btree_leaf_elm_t leaf
)
1164 hammer_off_t buf_offset
;
1165 hammer_off_t zone2_offset
;
1166 hammer_volume_t volume
;
1172 buf_offset
= bio
->bio_offset
;
1173 KKASSERT((buf_offset
& HAMMER_OFF_ZONE_MASK
) ==
1174 HAMMER_ZONE_LARGE_DATA
);
1177 * The buffer cache may have an aliased buffer (the reblocker can
1178 * write them). If it does we have to sync any dirty data before
1179 * we can build our direct-read. This is a non-critical code path.
1182 hammer_sync_buffers(hmp
, buf_offset
, bp
->b_bufsize
);
1185 * Resolve to a zone-2 offset. The conversion just requires
1186 * munging the top 4 bits but we want to abstract it anyway
1187 * so the blockmap code can verify the zone assignment.
1189 zone2_offset
= hammer_blockmap_lookup(hmp
, buf_offset
, &error
);
1192 KKASSERT((zone2_offset
& HAMMER_OFF_ZONE_MASK
) ==
1193 HAMMER_ZONE_RAW_BUFFER
);
1196 * Resolve volume and raw-offset for 3rd level bio. The
1197 * offset will be specific to the volume.
1199 vol_no
= HAMMER_VOL_DECODE(zone2_offset
);
1200 volume
= hammer_get_volume(hmp
, vol_no
, &error
);
1201 if (error
== 0 && zone2_offset
>= volume
->maxbuf_off
)
1208 nbio
= push_bio(bio
);
1209 nbio
->bio_offset
= volume
->ondisk
->vol_buf_beg
+
1210 (zone2_offset
& HAMMER_OFF_SHORT_MASK
);
1213 * XXX disabled - our CRC check doesn't work if the OS
1214 * does bogus_page replacement on the direct-read.
1216 if (leaf
&& hammer_verify_data
) {
1217 nbio
->bio_done
= hammer_io_direct_read_complete
;
1218 nbio
->bio_caller_info1
.uvalue32
= leaf
->data_crc
;
1221 hammer_stats_disk_read
+= bp
->b_bufsize
;
1222 vn_strategy(volume
->devvp
, nbio
);
1224 hammer_rel_volume(volume
, 0);
1227 kprintf("hammer_direct_read: failed @ %016llx\n",
1228 (long long)zone2_offset
);
1229 bp
->b_error
= error
;
1230 bp
->b_flags
|= B_ERROR
;
1238 * On completion of the BIO this callback must check the data CRC
1239 * and chain to the previous bio.
1243 hammer_io_direct_read_complete(struct bio
*nbio
)
1247 u_int32_t rec_crc
= nbio
->bio_caller_info1
.uvalue32
;
1250 if (crc32(bp
->b_data
, bp
->b_bufsize
) != rec_crc
) {
1251 kprintf("HAMMER: data_crc error @%016llx/%d\n",
1252 nbio
->bio_offset
, bp
->b_bufsize
);
1253 if (hammer_debug_critical
)
1254 Debugger("data_crc on read");
1255 bp
->b_flags
|= B_ERROR
;
1258 obio
= pop_bio(nbio
);
1264 * Write a buffer associated with a front-end vnode directly to the
1265 * disk media. The bio may be issued asynchronously.
1267 * The BIO is associated with the specified record and RECF_DIRECT_IO
1268 * is set. The recorded is added to its object.
1271 hammer_io_direct_write(hammer_mount_t hmp
, struct bio
*bio
,
1272 hammer_record_t record
)
1274 hammer_btree_leaf_elm_t leaf
= &record
->leaf
;
1275 hammer_off_t buf_offset
;
1276 hammer_off_t zone2_offset
;
1277 hammer_volume_t volume
;
1278 hammer_buffer_t buffer
;
1285 buf_offset
= leaf
->data_offset
;
1287 KKASSERT(buf_offset
> HAMMER_ZONE_BTREE
);
1288 KKASSERT(bio
->bio_buf
->b_cmd
== BUF_CMD_WRITE
);
1291 * Issue or execute the I/O. The new memory record must replace
1292 * the old one before the I/O completes, otherwise a reaquisition of
1293 * the buffer will load the old media data instead of the new.
1295 if ((buf_offset
& HAMMER_BUFMASK
) == 0 &&
1296 leaf
->data_len
>= HAMMER_BUFSIZE
) {
1298 * We are using the vnode's bio to write directly to the
1299 * media, any hammer_buffer at the same zone-X offset will
1300 * now have stale data.
1302 zone2_offset
= hammer_blockmap_lookup(hmp
, buf_offset
, &error
);
1303 vol_no
= HAMMER_VOL_DECODE(zone2_offset
);
1304 volume
= hammer_get_volume(hmp
, vol_no
, &error
);
1306 if (error
== 0 && zone2_offset
>= volume
->maxbuf_off
)
1310 KKASSERT((bp
->b_bufsize
& HAMMER_BUFMASK
) == 0);
1312 hammer_del_buffers(hmp, buf_offset,
1313 zone2_offset, bp->b_bufsize);
1317 * Second level bio - cached zone2 offset.
1319 * (We can put our bio_done function in either the
1320 * 2nd or 3rd level).
1322 nbio
= push_bio(bio
);
1323 nbio
->bio_offset
= zone2_offset
;
1324 nbio
->bio_done
= hammer_io_direct_write_complete
;
1325 nbio
->bio_caller_info1
.ptr
= record
;
1326 record
->zone2_offset
= zone2_offset
;
1327 record
->flags
|= HAMMER_RECF_DIRECT_IO
|
1328 HAMMER_RECF_DIRECT_INVAL
;
1331 * Third level bio - raw offset specific to the
1334 zone2_offset
&= HAMMER_OFF_SHORT_MASK
;
1335 nbio
= push_bio(nbio
);
1336 nbio
->bio_offset
= volume
->ondisk
->vol_buf_beg
+
1338 hammer_stats_disk_write
+= bp
->b_bufsize
;
1339 hammer_ip_replace_bulk(hmp
, record
);
1340 vn_strategy(volume
->devvp
, nbio
);
1341 hammer_io_flush_mark(volume
);
1343 hammer_rel_volume(volume
, 0);
1346 * Must fit in a standard HAMMER buffer. In this case all
1347 * consumers use the HAMMER buffer system and RECF_DIRECT_IO
1348 * does not need to be set-up.
1350 KKASSERT(((buf_offset
^ (buf_offset
+ leaf
->data_len
- 1)) & ~HAMMER_BUFMASK64
) == 0);
1352 ptr
= hammer_bread(hmp
, buf_offset
, &error
, &buffer
);
1355 bp
->b_flags
|= B_AGE
;
1356 hammer_io_modify(&buffer
->io
, 1);
1357 bcopy(bp
->b_data
, ptr
, leaf
->data_len
);
1358 hammer_io_modify_done(&buffer
->io
);
1359 hammer_rel_buffer(buffer
, 0);
1361 hammer_ip_replace_bulk(hmp
, record
);
1367 * Major suckage occured. Also note: The record was
1368 * never added to the tree so we do not have to worry
1369 * about the backend.
1371 kprintf("hammer_direct_write: failed @ %016llx\n",
1372 (long long)leaf
->data_offset
);
1376 bp
->b_flags
|= B_ERROR
;
1378 record
->flags
|= HAMMER_RECF_DELETED_FE
;
1379 hammer_rel_mem_record(record
);
1385 * On completion of the BIO this callback must disconnect
1386 * it from the hammer_record and chain to the previous bio.
1388 * An I/O error forces the mount to read-only. Data buffers
1389 * are not B_LOCKED like meta-data buffers are, so we have to
1390 * throw the buffer away to prevent the kernel from retrying.
1394 hammer_io_direct_write_complete(struct bio
*nbio
)
1398 hammer_record_t record
= nbio
->bio_caller_info1
.ptr
;
1401 obio
= pop_bio(nbio
);
1402 if (bp
->b_flags
& B_ERROR
) {
1403 hammer_critical_error(record
->ip
->hmp
, record
->ip
,
1405 "while writing bulk data");
1406 bp
->b_flags
|= B_INVAL
;
1410 KKASSERT(record
!= NULL
);
1411 KKASSERT(record
->flags
& HAMMER_RECF_DIRECT_IO
);
1412 if (record
->flags
& HAMMER_RECF_DIRECT_WAIT
) {
1413 record
->flags
&= ~(HAMMER_RECF_DIRECT_IO
|
1414 HAMMER_RECF_DIRECT_WAIT
);
1415 /* record can disappear once DIRECT_IO flag is cleared */
1416 wakeup(&record
->flags
);
1418 record
->flags
&= ~HAMMER_RECF_DIRECT_IO
;
1419 /* record can disappear once DIRECT_IO flag is cleared */
1425 * This is called before a record is either committed to the B-Tree
1426 * or destroyed, to resolve any associated direct-IO.
1428 * (1) We must wait for any direct-IO related to the record to complete.
1430 * (2) We must remove any buffer cache aliases for data accessed via
1431 * leaf->data_offset or zone2_offset so non-direct-IO consumers
1432 * (the mirroring and reblocking code) do not see stale data.
1435 hammer_io_direct_wait(hammer_record_t record
)
1438 * Wait for I/O to complete
1440 if (record
->flags
& HAMMER_RECF_DIRECT_IO
) {
1442 while (record
->flags
& HAMMER_RECF_DIRECT_IO
) {
1443 record
->flags
|= HAMMER_RECF_DIRECT_WAIT
;
1444 tsleep(&record
->flags
, 0, "hmdiow", 0);
1450 * Invalidate any related buffer cache aliases associated with the
1451 * backing device. This is needed because the buffer cache buffer
1452 * for file data is associated with the file vnode, not the backing
1455 * XXX I do not think this case can occur any more now that
1456 * reservations ensure that all such buffers are removed before
1457 * an area can be reused.
1459 if (record
->flags
& HAMMER_RECF_DIRECT_INVAL
) {
1460 KKASSERT(record
->leaf
.data_offset
);
1461 hammer_del_buffers(record
->ip
->hmp
, record
->leaf
.data_offset
,
1462 record
->zone2_offset
, record
->leaf
.data_len
,
1464 record
->flags
&= ~HAMMER_RECF_DIRECT_INVAL
;
1469 * This is called to remove the second-level cached zone-2 offset from
1470 * frontend buffer cache buffers, now stale due to a data relocation.
1471 * These offsets are generated by cluster_read() via VOP_BMAP, or directly
1472 * by hammer_vop_strategy_read().
1474 * This is rather nasty because here we have something like the reblocker
1475 * scanning the raw B-Tree with no held references on anything, really,
1476 * other then a shared lock on the B-Tree node, and we have to access the
1477 * frontend's buffer cache to check for and clean out the association.
1478 * Specifically, if the reblocker is moving data on the disk, these cached
1479 * offsets will become invalid.
1481 * Only data record types associated with the large-data zone are subject
1482 * to direct-io and need to be checked.
1486 hammer_io_direct_uncache(hammer_mount_t hmp
, hammer_btree_leaf_elm_t leaf
)
1488 struct hammer_inode_info iinfo
;
1491 if (leaf
->base
.rec_type
!= HAMMER_RECTYPE_DATA
)
1493 zone
= HAMMER_ZONE_DECODE(leaf
->data_offset
);
1494 if (zone
!= HAMMER_ZONE_LARGE_DATA_INDEX
)
1496 iinfo
.obj_id
= leaf
->base
.obj_id
;
1497 iinfo
.obj_asof
= 0; /* unused */
1498 iinfo
.obj_localization
= leaf
->base
.localization
&
1499 HAMMER_LOCALIZE_PSEUDOFS_MASK
;
1500 iinfo
.u
.leaf
= leaf
;
1501 hammer_scan_inode_snapshots(hmp
, &iinfo
,
1502 hammer_io_direct_uncache_callback
,
1507 hammer_io_direct_uncache_callback(hammer_inode_t ip
, void *data
)
1509 hammer_inode_info_t iinfo
= data
;
1510 hammer_off_t data_offset
;
1511 hammer_off_t file_offset
;
1518 data_offset
= iinfo
->u
.leaf
->data_offset
;
1519 file_offset
= iinfo
->u
.leaf
->base
.key
- iinfo
->u
.leaf
->data_len
;
1520 blksize
= iinfo
->u
.leaf
->data_len
;
1521 KKASSERT((blksize
& HAMMER_BUFMASK
) == 0);
1523 hammer_ref(&ip
->lock
);
1524 if (hammer_get_vnode(ip
, &vp
) == 0) {
1525 if ((bp
= findblk(ip
->vp
, file_offset
, FINDBLK_TEST
)) != NULL
&&
1526 bp
->b_bio2
.bio_offset
!= NOOFFSET
) {
1527 bp
= getblk(ip
->vp
, file_offset
, blksize
, 0, 0);
1528 bp
->b_bio2
.bio_offset
= NOOFFSET
;
1533 hammer_rel_inode(ip
, 0);
1539 * This function is called when writes may have occured on the volume,
1540 * indicating that the device may be holding cached writes.
1543 hammer_io_flush_mark(hammer_volume_t volume
)
1545 volume
->vol_flags
|= HAMMER_VOLF_NEEDFLUSH
;
1549 * This function ensures that the device has flushed any cached writes out.
1552 hammer_io_flush_sync(hammer_mount_t hmp
)
1554 hammer_volume_t volume
;
1555 struct buf
*bp_base
= NULL
;
1558 RB_FOREACH(volume
, hammer_vol_rb_tree
, &hmp
->rb_vols_root
) {
1559 if (volume
->vol_flags
& HAMMER_VOLF_NEEDFLUSH
) {
1560 volume
->vol_flags
&= ~HAMMER_VOLF_NEEDFLUSH
;
1562 bp
->b_bio1
.bio_offset
= 0;
1565 bp
->b_cmd
= BUF_CMD_FLUSH
;
1566 bp
->b_bio1
.bio_caller_info1
.cluster_head
= bp_base
;
1567 bp
->b_bio1
.bio_done
= biodone_sync
;
1568 bp
->b_bio1
.bio_flags
|= BIO_SYNC
;
1570 vn_strategy(volume
->devvp
, &bp
->b_bio1
);
1573 while ((bp
= bp_base
) != NULL
) {
1574 bp_base
= bp
->b_bio1
.bio_caller_info1
.cluster_head
;
1575 biowait(&bp
->b_bio1
, "hmrFLS");