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
;
126 * Wait for any physical IO to complete
128 * XXX we aren't interlocked against a spinlock or anything so there
129 * is a small window in the interlock / io->running == 0 test.
132 hammer_io_wait(hammer_io_t io
)
137 tsleep_interlock(io
, 0);
138 if (io
->running
== 0)
140 tsleep(io
, PINTERLOCKED
, "hmrflw", hz
);
141 if (io
->running
== 0)
148 * Wait for all hammer_io-initated write I/O's to complete. This is not
149 * supposed to count direct I/O's but some can leak through (for
150 * non-full-sized direct I/Os).
153 hammer_io_wait_all(hammer_mount_t hmp
, const char *ident
)
155 hammer_io_flush_sync(hmp
);
157 while (hmp
->io_running_space
)
158 tsleep(&hmp
->io_running_space
, 0, ident
, 0);
162 #define HAMMER_MAXRA 4
165 * Load bp for a HAMMER structure. The io must be exclusively locked by
168 * This routine is mostly used on meta-data and small-data blocks. Generally
169 * speaking HAMMER assumes some locality of reference and will cluster
172 * Note that clustering occurs at the device layer, not the logical layer.
173 * If the buffers do not apply to the current operation they may apply to
177 hammer_io_read(struct vnode
*devvp
, struct hammer_io
*io
, hammer_off_t limit
)
182 if ((bp
= io
->bp
) == NULL
) {
183 hammer_count_io_running_read
+= io
->bytes
;
184 if (hammer_cluster_enable
) {
185 error
= cluster_read(devvp
, limit
,
186 io
->offset
, io
->bytes
,
188 HAMMER_CLUSTER_BUFS
, &io
->bp
);
190 error
= bread(devvp
, io
->offset
, io
->bytes
, &io
->bp
);
192 hammer_stats_disk_read
+= io
->bytes
;
193 hammer_count_io_running_read
-= io
->bytes
;
196 * The code generally assumes b_ops/b_dep has been set-up,
197 * even if we error out here.
200 bp
->b_ops
= &hammer_bioops
;
201 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
202 LIST_INSERT_HEAD(&bp
->b_dep
, &io
->worklist
, node
);
204 KKASSERT(io
->modified
== 0);
205 KKASSERT(io
->running
== 0);
206 KKASSERT(io
->waiting
== 0);
207 io
->released
= 0; /* we hold an active lock on bp */
215 * Similar to hammer_io_read() but returns a zero'd out buffer instead.
216 * Must be called with the IO exclusively locked.
218 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background
219 * I/O by forcing the buffer to not be in a released state before calling
222 * This function will also mark the IO as modified but it will not
223 * increment the modify_refs count.
226 hammer_io_new(struct vnode
*devvp
, struct hammer_io
*io
)
230 if ((bp
= io
->bp
) == NULL
) {
231 io
->bp
= getblk(devvp
, io
->offset
, io
->bytes
, 0, 0);
233 bp
->b_ops
= &hammer_bioops
;
234 KKASSERT(LIST_FIRST(&bp
->b_dep
) == NULL
);
235 LIST_INSERT_HEAD(&bp
->b_dep
, &io
->worklist
, node
);
237 KKASSERT(io
->running
== 0);
247 hammer_io_modify(io
, 0);
253 * Remove potential device level aliases against buffers managed by high level
254 * vnodes. Aliases can also be created due to mixed buffer sizes or via
255 * direct access to the backing store device.
257 * This is nasty because the buffers are also VMIO-backed. Even if a buffer
258 * does not exist its backing VM pages might, and we have to invalidate
259 * those as well or a getblk() will reinstate them.
261 * Buffer cache buffers associated with hammer_buffers cannot be
265 hammer_io_inval(hammer_volume_t volume
, hammer_off_t zone2_offset
)
267 hammer_io_structure_t iou
;
268 hammer_off_t phys_offset
;
272 phys_offset
= volume
->ondisk
->vol_buf_beg
+
273 (zone2_offset
& HAMMER_OFF_SHORT_MASK
);
275 if ((bp
= findblk(volume
->devvp
, phys_offset
, FINDBLK_TEST
)) != NULL
)
276 bp
= getblk(volume
->devvp
, phys_offset
, bp
->b_bufsize
, 0, 0);
278 bp
= getblk(volume
->devvp
, phys_offset
, HAMMER_BUFSIZE
, 0, 0);
279 if ((iou
= (void *)LIST_FIRST(&bp
->b_dep
)) != NULL
) {
281 hammer_ref(&iou
->io
.lock
);
282 hammer_io_clear_modify(&iou
->io
, 1);
284 iou
->io
.released
= 0;
288 KKASSERT(iou
->io
.lock
.refs
== 1);
289 hammer_rel_buffer(&iou
->buffer
, 0);
290 /*hammer_io_deallocate(bp);*/
295 KKASSERT((bp
->b_flags
& B_LOCKED
) == 0);
297 bp
->b_flags
|= B_NOCACHE
|B_RELBUF
;
306 * This routine is called on the last reference to a hammer structure.
307 * The io is usually interlocked with io.loading and io.refs must be 1.
309 * This routine may return a non-NULL bp to the caller for dispoal. Disposal
310 * simply means the caller finishes decrementing the ref-count on the
311 * IO structure then brelse()'s the bp. The bp may or may not still be
312 * passively associated with the IO.
314 * The only requirement here is that modified meta-data and volume-header
315 * buffer may NOT be disassociated from the IO structure, and consequently
316 * we also leave such buffers actively associated with the IO if they already
317 * are (since the kernel can't do anything with them anyway). Only the
318 * flusher is allowed to write such buffers out. Modified pure-data and
319 * undo buffers are returned to the kernel but left passively associated
320 * so we can track when the kernel writes the bp out.
323 hammer_io_release(struct hammer_io
*io
, int flush
)
325 union hammer_io_structure
*iou
= (void *)io
;
328 if ((bp
= io
->bp
) == NULL
)
332 * Try to flush a dirty IO to disk if asked to by the
333 * caller or if the kernel tried to flush the buffer in the past.
335 * Kernel-initiated flushes are only allowed for pure-data buffers.
336 * meta-data and volume buffers can only be flushed explicitly
341 hammer_io_flush(io
, 0);
342 } else if (bp
->b_flags
& B_LOCKED
) {
344 case HAMMER_STRUCTURE_DATA_BUFFER
:
345 hammer_io_flush(io
, 0);
347 case HAMMER_STRUCTURE_UNDO_BUFFER
:
348 hammer_io_flush(io
, hammer_undo_reclaim(io
));
353 } /* else no explicit request to flush the buffer */
357 * Wait for the IO to complete if asked to. This occurs when
358 * the buffer must be disposed of definitively during an umount
359 * or buffer invalidation.
361 if (io
->waitdep
&& io
->running
) {
366 * Return control of the buffer to the kernel (with the provisio
367 * that our bioops can override kernel decisions with regards to
370 if ((flush
|| io
->reclaim
) && io
->modified
== 0 && io
->running
== 0) {
372 * Always disassociate the bp if an explicit flush
373 * was requested and the IO completed with no error
374 * (so unmount can really clean up the structure).
382 hammer_io_disassociate((hammer_io_structure_t
)io
);
384 } else if (io
->modified
) {
386 * Only certain IO types can be released to the kernel if
387 * the buffer has been modified.
389 * volume and meta-data IO types may only be explicitly
393 case HAMMER_STRUCTURE_DATA_BUFFER
:
394 case HAMMER_STRUCTURE_UNDO_BUFFER
:
395 if (io
->released
== 0) {
403 bp
= NULL
; /* bp left associated */
404 } else if (io
->released
== 0) {
406 * Clean buffers can be generally released to the kernel.
407 * We leave the bp passively associated with the HAMMER
408 * structure and use bioops to disconnect it later on
409 * if the kernel wants to discard the buffer.
411 * We can steal the structure's ownership of the bp.
414 if (bp
->b_flags
& B_LOCKED
) {
415 hammer_io_disassociate(iou
);
419 hammer_io_disassociate(iou
);
422 /* return the bp (bp passively associated) */
427 * A released buffer is passively associate with our
428 * hammer_io structure. The kernel cannot destroy it
429 * without making a bioops call. If the kernel (B_LOCKED)
430 * or we (reclaim) requested that the buffer be destroyed
431 * we destroy it, otherwise we do a quick get/release to
432 * reset its position in the kernel's LRU list.
434 * Leaving the buffer passively associated allows us to
435 * use the kernel's LRU buffer flushing mechanisms rather
436 * then rolling our own.
438 * XXX there are two ways of doing this. We can re-acquire
439 * and passively release to reset the LRU, or not.
441 if (io
->running
== 0) {
443 if ((bp
->b_flags
& B_LOCKED
) || io
->reclaim
) {
444 hammer_io_disassociate(iou
);
447 /* return the bp (bp passively associated) */
451 * bp is left passively associated but we do not
452 * try to reacquire it. Interactions with the io
453 * structure will occur on completion of the bp's
463 * This routine is called with a locked IO when a flush is desired and
464 * no other references to the structure exists other then ours. This
465 * routine is ONLY called when HAMMER believes it is safe to flush a
466 * potentially modified buffer out.
469 hammer_io_flush(struct hammer_io
*io
, int reclaim
)
474 * Degenerate case - nothing to flush if nothing is dirty.
476 if (io
->modified
== 0) {
481 KKASSERT(io
->modify_refs
<= 0);
484 * Acquire ownership of the bp, particularly before we clear our
487 * We are going to bawrite() this bp. Don't leave a window where
488 * io->released is set, we actually own the bp rather then our
494 /* BUF_KERNPROC(io->bp); */
495 /* io->released = 0; */
496 KKASSERT(io
->released
);
497 KKASSERT(io
->bp
== bp
);
503 if ((bp
->b_flags
& B_LOCKED
) == 0) {
504 bp
->b_flags
|= B_LOCKED
;
505 ++hammer_count_io_locked
;
510 * Acquire exclusive access to the bp and then clear the modified
511 * state of the buffer prior to issuing I/O to interlock any
512 * modifications made while the I/O is in progress. This shouldn't
513 * happen anyway but losing data would be worse. The modified bit
514 * will be rechecked after the IO completes.
516 * NOTE: This call also finalizes the buffer's content (inval == 0).
518 * This is only legal when lock.refs == 1 (otherwise we might clear
519 * the modified bit while there are still users of the cluster
520 * modifying the data).
522 * Do this before potentially blocking so any attempt to modify the
523 * ondisk while we are blocked blocks waiting for us.
525 hammer_ref(&io
->lock
);
526 hammer_io_clear_modify(io
, 0);
527 hammer_unref(&io
->lock
);
530 * Transfer ownership to the kernel and initiate I/O.
533 io
->hmp
->io_running_space
+= io
->bytes
;
534 hammer_count_io_running_write
+= io
->bytes
;
536 hammer_io_flush_mark(io
->volume
);
539 /************************************************************************
541 ************************************************************************
543 * These routines deal with dependancies created when IO buffers get
544 * modified. The caller must call hammer_modify_*() on a referenced
545 * HAMMER structure prior to modifying its on-disk data.
547 * Any intent to modify an IO buffer acquires the related bp and imposes
548 * various write ordering dependancies.
552 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
553 * are locked until the flusher can deal with them, pure data buffers
554 * can be written out.
558 hammer_io_modify(hammer_io_t io
, int count
)
561 * io->modify_refs must be >= 0
563 while (io
->modify_refs
< 0) {
565 tsleep(io
, 0, "hmrmod", 0);
569 * Shortcut if nothing to do.
571 KKASSERT(io
->lock
.refs
!= 0 && io
->bp
!= NULL
);
572 io
->modify_refs
+= count
;
573 if (io
->modified
&& io
->released
== 0)
576 hammer_lock_ex(&io
->lock
);
577 if (io
->modified
== 0) {
578 hammer_io_set_modlist(io
);
583 BUF_KERNPROC(io
->bp
);
585 KKASSERT(io
->modified
!= 0);
587 hammer_unlock(&io
->lock
);
592 hammer_io_modify_done(hammer_io_t io
)
594 KKASSERT(io
->modify_refs
> 0);
596 if (io
->modify_refs
== 0 && io
->waitmod
) {
603 hammer_io_write_interlock(hammer_io_t io
)
605 while (io
->modify_refs
!= 0) {
607 tsleep(io
, 0, "hmrmod", 0);
609 io
->modify_refs
= -1;
613 hammer_io_done_interlock(hammer_io_t io
)
615 KKASSERT(io
->modify_refs
== -1);
624 * Caller intends to modify a volume's ondisk structure.
626 * This is only allowed if we are the flusher or we have a ref on the
630 hammer_modify_volume(hammer_transaction_t trans
, hammer_volume_t volume
,
633 KKASSERT (trans
== NULL
|| trans
->sync_lock_refs
> 0);
635 hammer_io_modify(&volume
->io
, 1);
637 intptr_t rel_offset
= (intptr_t)base
- (intptr_t)volume
->ondisk
;
638 KKASSERT((rel_offset
& ~(intptr_t)HAMMER_BUFMASK
) == 0);
639 hammer_generate_undo(trans
, &volume
->io
,
640 HAMMER_ENCODE_RAW_VOLUME(volume
->vol_no
, rel_offset
),
646 * Caller intends to modify a buffer's ondisk structure.
648 * This is only allowed if we are the flusher or we have a ref on the
652 hammer_modify_buffer(hammer_transaction_t trans
, hammer_buffer_t buffer
,
655 KKASSERT (trans
== NULL
|| trans
->sync_lock_refs
> 0);
657 hammer_io_modify(&buffer
->io
, 1);
659 intptr_t rel_offset
= (intptr_t)base
- (intptr_t)buffer
->ondisk
;
660 KKASSERT((rel_offset
& ~(intptr_t)HAMMER_BUFMASK
) == 0);
661 hammer_generate_undo(trans
, &buffer
->io
,
662 buffer
->zone2_offset
+ rel_offset
,
668 hammer_modify_volume_done(hammer_volume_t volume
)
670 hammer_io_modify_done(&volume
->io
);
674 hammer_modify_buffer_done(hammer_buffer_t buffer
)
676 hammer_io_modify_done(&buffer
->io
);
680 * Mark an entity as not being dirty any more and finalize any
681 * delayed adjustments to the buffer.
683 * Delayed adjustments are an important performance enhancement, allowing
684 * us to avoid recalculating B-Tree node CRCs over and over again when
685 * making bulk-modifications to the B-Tree.
687 * If inval is non-zero delayed adjustments are ignored.
689 * This routine may dereference related btree nodes and cause the
690 * buffer to be dereferenced. The caller must own a reference on io.
693 hammer_io_clear_modify(struct hammer_io
*io
, int inval
)
695 if (io
->modified
== 0)
699 * Take us off the mod-list and clear the modified bit.
701 KKASSERT(io
->mod_list
!= NULL
);
702 if (io
->mod_list
== &io
->hmp
->volu_list
||
703 io
->mod_list
== &io
->hmp
->meta_list
) {
704 io
->hmp
->locked_dirty_space
-= io
->bytes
;
705 hammer_count_dirtybufspace
-= io
->bytes
;
707 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
712 * If this bit is not set there are no delayed adjustments.
719 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference
720 * on the node (& underlying buffer). Release the node after clearing
723 if (io
->type
== HAMMER_STRUCTURE_META_BUFFER
) {
724 hammer_buffer_t buffer
= (void *)io
;
728 TAILQ_FOREACH(node
, &buffer
->clist
, entry
) {
729 if ((node
->flags
& HAMMER_NODE_NEEDSCRC
) == 0)
731 node
->flags
&= ~HAMMER_NODE_NEEDSCRC
;
732 KKASSERT(node
->ondisk
);
734 node
->ondisk
->crc
= crc32(&node
->ondisk
->crc
+ 1, HAMMER_BTREE_CRCSIZE
);
735 hammer_rel_node(node
);
739 /* caller must still have ref on io */
740 KKASSERT(io
->lock
.refs
> 0);
744 * Clear the IO's modify list. Even though the IO is no longer modified
745 * it may still be on the lose_list. This routine is called just before
746 * the governing hammer_buffer is destroyed.
749 hammer_io_clear_modlist(struct hammer_io
*io
)
751 KKASSERT(io
->modified
== 0);
753 crit_enter(); /* biodone race against list */
754 KKASSERT(io
->mod_list
== &io
->hmp
->lose_list
);
755 TAILQ_REMOVE(io
->mod_list
, io
, mod_entry
);
762 hammer_io_set_modlist(struct hammer_io
*io
)
764 struct hammer_mount
*hmp
= io
->hmp
;
766 KKASSERT(io
->mod_list
== NULL
);
769 case HAMMER_STRUCTURE_VOLUME
:
770 io
->mod_list
= &hmp
->volu_list
;
771 hmp
->locked_dirty_space
+= io
->bytes
;
772 hammer_count_dirtybufspace
+= io
->bytes
;
774 case HAMMER_STRUCTURE_META_BUFFER
:
775 io
->mod_list
= &hmp
->meta_list
;
776 hmp
->locked_dirty_space
+= io
->bytes
;
777 hammer_count_dirtybufspace
+= io
->bytes
;
779 case HAMMER_STRUCTURE_UNDO_BUFFER
:
780 io
->mod_list
= &hmp
->undo_list
;
782 case HAMMER_STRUCTURE_DATA_BUFFER
:
783 io
->mod_list
= &hmp
->data_list
;
786 TAILQ_INSERT_TAIL(io
->mod_list
, io
, mod_entry
);
789 /************************************************************************
791 ************************************************************************
796 * Pre-IO initiation kernel callback - cluster build only
799 hammer_io_start(struct buf
*bp
)
804 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT!
806 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
807 * may also be set if we were marking a cluster header open. Only remove
808 * our dependancy if the modified bit is clear.
811 hammer_io_complete(struct buf
*bp
)
813 union hammer_io_structure
*iou
= (void *)LIST_FIRST(&bp
->b_dep
);
815 KKASSERT(iou
->io
.released
== 1);
818 * Deal with people waiting for I/O to drain
820 if (iou
->io
.running
) {
822 * Deal with critical write errors. Once a critical error
823 * has been flagged in hmp the UNDO FIFO will not be updated.
824 * That way crash recover will give us a consistent
827 * Because of this we can throw away failed UNDO buffers. If
828 * we throw away META or DATA buffers we risk corrupting
829 * the now read-only version of the filesystem visible to
830 * the user. Clear B_ERROR so the buffer is not re-dirtied
831 * by the kernel and ref the io so it doesn't get thrown
834 if (bp
->b_flags
& B_ERROR
) {
835 hammer_critical_error(iou
->io
.hmp
, NULL
, bp
->b_error
,
836 "while flushing meta-data");
837 switch(iou
->io
.type
) {
838 case HAMMER_STRUCTURE_UNDO_BUFFER
:
841 if (iou
->io
.ioerror
== 0) {
843 if (iou
->io
.lock
.refs
== 0)
844 ++hammer_count_refedbufs
;
845 hammer_ref(&iou
->io
.lock
);
849 bp
->b_flags
&= ~B_ERROR
;
852 hammer_io_set_modlist(&iou
->io
);
853 iou
->io
.modified
= 1;
856 hammer_stats_disk_write
+= iou
->io
.bytes
;
857 hammer_count_io_running_write
-= iou
->io
.bytes
;
858 iou
->io
.hmp
->io_running_space
-= iou
->io
.bytes
;
859 if (iou
->io
.hmp
->io_running_space
== 0)
860 wakeup(&iou
->io
.hmp
->io_running_space
);
861 KKASSERT(iou
->io
.hmp
->io_running_space
>= 0);
864 hammer_stats_disk_read
+= iou
->io
.bytes
;
867 if (iou
->io
.waiting
) {
873 * If B_LOCKED is set someone wanted to deallocate the bp at some
874 * point, do it now if refs has become zero.
876 if ((bp
->b_flags
& B_LOCKED
) && iou
->io
.lock
.refs
== 0) {
877 KKASSERT(iou
->io
.modified
== 0);
878 --hammer_count_io_locked
;
879 bp
->b_flags
&= ~B_LOCKED
;
880 hammer_io_deallocate(bp
);
881 /* structure may be dead now */
886 * Callback from kernel when it wishes to deallocate a passively
887 * associated structure. This mostly occurs with clean buffers
888 * but it may be possible for a holding structure to be marked dirty
889 * while its buffer is passively associated. The caller owns the bp.
891 * If we cannot disassociate we set B_LOCKED to prevent the buffer
892 * from getting reused.
894 * WARNING: Because this can be called directly by getnewbuf we cannot
895 * recurse into the tree. If a bp cannot be immediately disassociated
896 * our only recourse is to set B_LOCKED.
898 * WARNING: This may be called from an interrupt via hammer_io_complete()
901 hammer_io_deallocate(struct buf
*bp
)
903 hammer_io_structure_t iou
= (void *)LIST_FIRST(&bp
->b_dep
);
905 KKASSERT((bp
->b_flags
& B_LOCKED
) == 0 && iou
->io
.running
== 0);
906 if (iou
->io
.lock
.refs
> 0 || iou
->io
.modified
) {
908 * It is not legal to disassociate a modified buffer. This
909 * case really shouldn't ever occur.
911 bp
->b_flags
|= B_LOCKED
;
912 ++hammer_count_io_locked
;
915 * Disassociate the BP. If the io has no refs left we
916 * have to add it to the loose list.
918 hammer_io_disassociate(iou
);
919 if (iou
->io
.type
!= HAMMER_STRUCTURE_VOLUME
) {
920 KKASSERT(iou
->io
.bp
== NULL
);
921 KKASSERT(iou
->io
.mod_list
== NULL
);
922 crit_enter(); /* biodone race against list */
923 iou
->io
.mod_list
= &iou
->io
.hmp
->lose_list
;
924 TAILQ_INSERT_TAIL(iou
->io
.mod_list
, &iou
->io
, mod_entry
);
931 hammer_io_fsync(struct vnode
*vp
)
937 * NOTE: will not be called unless we tell the kernel about the
938 * bioops. Unused... we use the mount's VFS_SYNC instead.
941 hammer_io_sync(struct mount
*mp
)
947 hammer_io_movedeps(struct buf
*bp1
, struct buf
*bp2
)
952 * I/O pre-check for reading and writing. HAMMER only uses this for
953 * B_CACHE buffers so checkread just shouldn't happen, but if it does
956 * Writing is a different case. We don't want the kernel to try to write
957 * out a buffer that HAMMER may be modifying passively or which has a
958 * dependancy. In addition, kernel-demanded writes can only proceed for
959 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
960 * buffer types can only be explicitly written by the flusher.
962 * checkwrite will only be called for bdwrite()n buffers. If we return
963 * success the kernel is guaranteed to initiate the buffer write.
966 hammer_io_checkread(struct buf
*bp
)
972 hammer_io_checkwrite(struct buf
*bp
)
974 hammer_io_t io
= (void *)LIST_FIRST(&bp
->b_dep
);
977 * This shouldn't happen under normal operation.
979 if (io
->type
== HAMMER_STRUCTURE_VOLUME
||
980 io
->type
== HAMMER_STRUCTURE_META_BUFFER
) {
982 panic("hammer_io_checkwrite: illegal buffer");
983 if ((bp
->b_flags
& B_LOCKED
) == 0) {
984 bp
->b_flags
|= B_LOCKED
;
985 ++hammer_count_io_locked
;
991 * We can only clear the modified bit if the IO is not currently
992 * undergoing modification. Otherwise we may miss changes.
994 * Only data and undo buffers can reach here. These buffers do
995 * not have terminal crc functions but we temporarily reference
996 * the IO anyway, just in case.
998 if (io
->modify_refs
== 0 && io
->modified
) {
999 hammer_ref(&io
->lock
);
1000 hammer_io_clear_modify(io
, 0);
1001 hammer_unref(&io
->lock
);
1002 } else if (io
->modified
) {
1003 KKASSERT(io
->type
== HAMMER_STRUCTURE_DATA_BUFFER
);
1007 * The kernel is going to start the IO, set io->running.
1009 KKASSERT(io
->running
== 0);
1011 io
->hmp
->io_running_space
+= io
->bytes
;
1012 hammer_count_io_running_write
+= io
->bytes
;
1017 * Return non-zero if we wish to delay the kernel's attempt to flush
1018 * this buffer to disk.
1021 hammer_io_countdeps(struct buf
*bp
, int n
)
1026 struct bio_ops hammer_bioops
= {
1027 .io_start
= hammer_io_start
,
1028 .io_complete
= hammer_io_complete
,
1029 .io_deallocate
= hammer_io_deallocate
,
1030 .io_fsync
= hammer_io_fsync
,
1031 .io_sync
= hammer_io_sync
,
1032 .io_movedeps
= hammer_io_movedeps
,
1033 .io_countdeps
= hammer_io_countdeps
,
1034 .io_checkread
= hammer_io_checkread
,
1035 .io_checkwrite
= hammer_io_checkwrite
,
1038 /************************************************************************
1040 ************************************************************************
1042 * These functions operate directly on the buffer cache buffer associated
1043 * with a front-end vnode rather then a back-end device vnode.
1047 * Read a buffer associated with a front-end vnode directly from the
1048 * disk media. The bio may be issued asynchronously. If leaf is non-NULL
1049 * we validate the CRC.
1051 * We must check for the presence of a HAMMER buffer to handle the case
1052 * where the reblocker has rewritten the data (which it does via the HAMMER
1053 * buffer system, not via the high-level vnode buffer cache), but not yet
1054 * committed the buffer to the media.
1057 hammer_io_direct_read(hammer_mount_t hmp
, struct bio
*bio
,
1058 hammer_btree_leaf_elm_t leaf
)
1060 hammer_off_t buf_offset
;
1061 hammer_off_t zone2_offset
;
1062 hammer_volume_t volume
;
1068 buf_offset
= bio
->bio_offset
;
1069 KKASSERT((buf_offset
& HAMMER_OFF_ZONE_MASK
) ==
1070 HAMMER_ZONE_LARGE_DATA
);
1073 * The buffer cache may have an aliased buffer (the reblocker can
1074 * write them). If it does we have to sync any dirty data before
1075 * we can build our direct-read. This is a non-critical code path.
1078 hammer_sync_buffers(hmp
, buf_offset
, bp
->b_bufsize
);
1081 * Resolve to a zone-2 offset. The conversion just requires
1082 * munging the top 4 bits but we want to abstract it anyway
1083 * so the blockmap code can verify the zone assignment.
1085 zone2_offset
= hammer_blockmap_lookup(hmp
, buf_offset
, &error
);
1088 KKASSERT((zone2_offset
& HAMMER_OFF_ZONE_MASK
) ==
1089 HAMMER_ZONE_RAW_BUFFER
);
1092 * Resolve volume and raw-offset for 3rd level bio. The
1093 * offset will be specific to the volume.
1095 vol_no
= HAMMER_VOL_DECODE(zone2_offset
);
1096 volume
= hammer_get_volume(hmp
, vol_no
, &error
);
1097 if (error
== 0 && zone2_offset
>= volume
->maxbuf_off
)
1104 nbio
= push_bio(bio
);
1105 nbio
->bio_offset
= volume
->ondisk
->vol_buf_beg
+
1106 (zone2_offset
& HAMMER_OFF_SHORT_MASK
);
1109 * XXX disabled - our CRC check doesn't work if the OS
1110 * does bogus_page replacement on the direct-read.
1112 if (leaf
&& hammer_verify_data
) {
1113 nbio
->bio_done
= hammer_io_direct_read_complete
;
1114 nbio
->bio_caller_info1
.uvalue32
= leaf
->data_crc
;
1117 hammer_stats_disk_read
+= bp
->b_bufsize
;
1118 vn_strategy(volume
->devvp
, nbio
);
1120 hammer_rel_volume(volume
, 0);
1123 kprintf("hammer_direct_read: failed @ %016llx\n",
1124 (long long)zone2_offset
);
1125 bp
->b_error
= error
;
1126 bp
->b_flags
|= B_ERROR
;
1134 * On completion of the BIO this callback must check the data CRC
1135 * and chain to the previous bio.
1139 hammer_io_direct_read_complete(struct bio
*nbio
)
1143 u_int32_t rec_crc
= nbio
->bio_caller_info1
.uvalue32
;
1146 if (crc32(bp
->b_data
, bp
->b_bufsize
) != rec_crc
) {
1147 kprintf("HAMMER: data_crc error @%016llx/%d\n",
1148 nbio
->bio_offset
, bp
->b_bufsize
);
1149 if (hammer_debug_debug
)
1151 bp
->b_flags
|= B_ERROR
;
1154 obio
= pop_bio(nbio
);
1160 * Write a buffer associated with a front-end vnode directly to the
1161 * disk media. The bio may be issued asynchronously.
1163 * The BIO is associated with the specified record and RECF_DIRECT_IO
1164 * is set. The recorded is added to its object.
1167 hammer_io_direct_write(hammer_mount_t hmp
, hammer_record_t record
,
1170 hammer_btree_leaf_elm_t leaf
= &record
->leaf
;
1171 hammer_off_t buf_offset
;
1172 hammer_off_t zone2_offset
;
1173 hammer_volume_t volume
;
1174 hammer_buffer_t buffer
;
1181 buf_offset
= leaf
->data_offset
;
1183 KKASSERT(buf_offset
> HAMMER_ZONE_BTREE
);
1184 KKASSERT(bio
->bio_buf
->b_cmd
== BUF_CMD_WRITE
);
1186 if ((buf_offset
& HAMMER_BUFMASK
) == 0 &&
1187 leaf
->data_len
>= HAMMER_BUFSIZE
) {
1189 * We are using the vnode's bio to write directly to the
1190 * media, any hammer_buffer at the same zone-X offset will
1191 * now have stale data.
1193 zone2_offset
= hammer_blockmap_lookup(hmp
, buf_offset
, &error
);
1194 vol_no
= HAMMER_VOL_DECODE(zone2_offset
);
1195 volume
= hammer_get_volume(hmp
, vol_no
, &error
);
1197 if (error
== 0 && zone2_offset
>= volume
->maxbuf_off
)
1201 KKASSERT((bp
->b_bufsize
& HAMMER_BUFMASK
) == 0);
1203 hammer_del_buffers(hmp, buf_offset,
1204 zone2_offset, bp->b_bufsize);
1208 * Second level bio - cached zone2 offset.
1210 * (We can put our bio_done function in either the
1211 * 2nd or 3rd level).
1213 nbio
= push_bio(bio
);
1214 nbio
->bio_offset
= zone2_offset
;
1215 nbio
->bio_done
= hammer_io_direct_write_complete
;
1216 nbio
->bio_caller_info1
.ptr
= record
;
1217 record
->zone2_offset
= zone2_offset
;
1218 record
->flags
|= HAMMER_RECF_DIRECT_IO
|
1219 HAMMER_RECF_DIRECT_INVAL
;
1222 * Third level bio - raw offset specific to the
1225 zone2_offset
&= HAMMER_OFF_SHORT_MASK
;
1226 nbio
= push_bio(nbio
);
1227 nbio
->bio_offset
= volume
->ondisk
->vol_buf_beg
+
1229 hammer_stats_disk_write
+= bp
->b_bufsize
;
1230 vn_strategy(volume
->devvp
, nbio
);
1231 hammer_io_flush_mark(volume
);
1233 hammer_rel_volume(volume
, 0);
1236 * Must fit in a standard HAMMER buffer. In this case all
1237 * consumers use the HAMMER buffer system and RECF_DIRECT_IO
1238 * does not need to be set-up.
1240 KKASSERT(((buf_offset
^ (buf_offset
+ leaf
->data_len
- 1)) & ~HAMMER_BUFMASK64
) == 0);
1242 ptr
= hammer_bread(hmp
, buf_offset
, &error
, &buffer
);
1245 bp
->b_flags
|= B_AGE
;
1246 hammer_io_modify(&buffer
->io
, 1);
1247 bcopy(bp
->b_data
, ptr
, leaf
->data_len
);
1248 hammer_io_modify_done(&buffer
->io
);
1249 hammer_rel_buffer(buffer
, 0);
1256 * The record is all setup now, add it. Potential conflics
1257 * have already been dealt with.
1259 error
= hammer_mem_add(record
);
1260 KKASSERT(error
== 0);
1263 * Major suckage occured. Also note: The record was never added
1264 * to the tree so we do not have to worry about the backend.
1266 kprintf("hammer_direct_write: failed @ %016llx\n",
1267 (long long)leaf
->data_offset
);
1271 bp
->b_flags
|= B_ERROR
;
1273 record
->flags
|= HAMMER_RECF_DELETED_FE
;
1274 hammer_rel_mem_record(record
);
1280 * On completion of the BIO this callback must disconnect
1281 * it from the hammer_record and chain to the previous bio.
1283 * An I/O error forces the mount to read-only. Data buffers
1284 * are not B_LOCKED like meta-data buffers are, so we have to
1285 * throw the buffer away to prevent the kernel from retrying.
1289 hammer_io_direct_write_complete(struct bio
*nbio
)
1293 hammer_record_t record
= nbio
->bio_caller_info1
.ptr
;
1296 obio
= pop_bio(nbio
);
1297 if (bp
->b_flags
& B_ERROR
) {
1298 hammer_critical_error(record
->ip
->hmp
, record
->ip
,
1300 "while writing bulk data");
1301 bp
->b_flags
|= B_INVAL
;
1305 KKASSERT(record
!= NULL
);
1306 KKASSERT(record
->flags
& HAMMER_RECF_DIRECT_IO
);
1307 if (record
->flags
& HAMMER_RECF_DIRECT_WAIT
) {
1308 record
->flags
&= ~(HAMMER_RECF_DIRECT_IO
|
1309 HAMMER_RECF_DIRECT_WAIT
);
1310 /* record can disappear once DIRECT_IO flag is cleared */
1311 wakeup(&record
->flags
);
1313 record
->flags
&= ~HAMMER_RECF_DIRECT_IO
;
1314 /* record can disappear once DIRECT_IO flag is cleared */
1320 * This is called before a record is either committed to the B-Tree
1321 * or destroyed, to resolve any associated direct-IO.
1323 * (1) We must wait for any direct-IO related to the record to complete.
1325 * (2) We must remove any buffer cache aliases for data accessed via
1326 * leaf->data_offset or zone2_offset so non-direct-IO consumers
1327 * (the mirroring and reblocking code) do not see stale data.
1330 hammer_io_direct_wait(hammer_record_t record
)
1333 * Wait for I/O to complete
1335 if (record
->flags
& HAMMER_RECF_DIRECT_IO
) {
1337 while (record
->flags
& HAMMER_RECF_DIRECT_IO
) {
1338 record
->flags
|= HAMMER_RECF_DIRECT_WAIT
;
1339 tsleep(&record
->flags
, 0, "hmdiow", 0);
1345 * Invalidate any related buffer cache aliases associated with the
1346 * backing device. This is needed because the buffer cache buffer
1347 * for file data is associated with the file vnode, not the backing
1350 * XXX I do not think this case can occur any more now that
1351 * reservations ensure that all such buffers are removed before
1352 * an area can be reused.
1354 if (record
->flags
& HAMMER_RECF_DIRECT_INVAL
) {
1355 KKASSERT(record
->leaf
.data_offset
);
1356 hammer_del_buffers(record
->ip
->hmp
, record
->leaf
.data_offset
,
1357 record
->zone2_offset
, record
->leaf
.data_len
,
1359 record
->flags
&= ~HAMMER_RECF_DIRECT_INVAL
;
1364 * This is called to remove the second-level cached zone-2 offset from
1365 * frontend buffer cache buffers, now stale due to a data relocation.
1366 * These offsets are generated by cluster_read() via VOP_BMAP, or directly
1367 * by hammer_vop_strategy_read().
1369 * This is rather nasty because here we have something like the reblocker
1370 * scanning the raw B-Tree with no held references on anything, really,
1371 * other then a shared lock on the B-Tree node, and we have to access the
1372 * frontend's buffer cache to check for and clean out the association.
1373 * Specifically, if the reblocker is moving data on the disk, these cached
1374 * offsets will become invalid.
1376 * Only data record types associated with the large-data zone are subject
1377 * to direct-io and need to be checked.
1381 hammer_io_direct_uncache(hammer_mount_t hmp
, hammer_btree_leaf_elm_t leaf
)
1383 struct hammer_inode_info iinfo
;
1386 if (leaf
->base
.rec_type
!= HAMMER_RECTYPE_DATA
)
1388 zone
= HAMMER_ZONE_DECODE(leaf
->data_offset
);
1389 if (zone
!= HAMMER_ZONE_LARGE_DATA_INDEX
)
1391 iinfo
.obj_id
= leaf
->base
.obj_id
;
1392 iinfo
.obj_asof
= 0; /* unused */
1393 iinfo
.obj_localization
= leaf
->base
.localization
&
1394 HAMMER_LOCALIZE_PSEUDOFS_MASK
;
1395 iinfo
.u
.leaf
= leaf
;
1396 hammer_scan_inode_snapshots(hmp
, &iinfo
,
1397 hammer_io_direct_uncache_callback
,
1402 hammer_io_direct_uncache_callback(hammer_inode_t ip
, void *data
)
1404 hammer_inode_info_t iinfo
= data
;
1405 hammer_off_t data_offset
;
1406 hammer_off_t file_offset
;
1413 data_offset
= iinfo
->u
.leaf
->data_offset
;
1414 file_offset
= iinfo
->u
.leaf
->base
.key
- iinfo
->u
.leaf
->data_len
;
1415 blksize
= iinfo
->u
.leaf
->data_len
;
1416 KKASSERT((blksize
& HAMMER_BUFMASK
) == 0);
1418 hammer_ref(&ip
->lock
);
1419 if (hammer_get_vnode(ip
, &vp
) == 0) {
1420 if ((bp
= findblk(ip
->vp
, file_offset
, FINDBLK_TEST
)) != NULL
&&
1421 bp
->b_bio2
.bio_offset
!= NOOFFSET
) {
1422 bp
= getblk(ip
->vp
, file_offset
, blksize
, 0, 0);
1423 bp
->b_bio2
.bio_offset
= NOOFFSET
;
1428 hammer_rel_inode(ip
, 0);
1434 * This function is called when writes may have occured on the volume,
1435 * indicating that the device may be holding cached writes.
1438 hammer_io_flush_mark(hammer_volume_t volume
)
1440 volume
->vol_flags
|= HAMMER_VOLF_NEEDFLUSH
;
1444 * This function ensures that the device has flushed any cached writes out.
1447 hammer_io_flush_sync(hammer_mount_t hmp
)
1449 hammer_volume_t volume
;
1450 struct buf
*bp_base
= NULL
;
1453 RB_FOREACH(volume
, hammer_vol_rb_tree
, &hmp
->rb_vols_root
) {
1454 if (volume
->vol_flags
& HAMMER_VOLF_NEEDFLUSH
) {
1455 volume
->vol_flags
&= ~HAMMER_VOLF_NEEDFLUSH
;
1457 bp
->b_bio1
.bio_offset
= 0;
1460 bp
->b_cmd
= BUF_CMD_FLUSH
;
1461 bp
->b_bio1
.bio_caller_info1
.cluster_head
= bp_base
;
1462 bp
->b_bio1
.bio_done
= biodone_sync
;
1463 bp
->b_bio1
.bio_flags
|= BIO_SYNC
;
1465 vn_strategy(volume
->devvp
, &bp
->b_bio1
);
1468 while ((bp
= bp_base
) != NULL
) {
1469 bp_base
= bp
->b_bio1
.bio_caller_info1
.cluster_head
;
1470 biowait(&bp
->b_bio1
, "hmrFLS");