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
35 * IO Primitives and buffer cache management
37 * All major data-tracking structures in HAMMER contain a struct hammer_io
38 * which is used to manage their backing store. We use filesystem buffers
39 * for backing store and we leave them passively associated with their
42 * If the kernel tries to destroy a passively associated buf which we cannot
43 * yet let go we set B_LOCKED in the buffer and then actively released it
46 * The io_token is required for anything which might race bioops and bio_done
47 * callbacks, with one exception: A successful hammer_try_interlock_norefs().
48 * the fs_token will be held in all other cases.
52 #include <sys/fcntl.h>
53 #include <sys/nlookup.h>
57 static void hammer_io_modify(hammer_io_t io, int count);
58 static void hammer_io_deallocate(struct buf *bp);
60 static void hammer_io_direct_read_complete(struct bio *nbio);
62 static void hammer_io_direct_write_complete(struct bio *nbio);
63 static int hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data);
64 static void hammer_io_set_modlist(struct hammer_io *io);
65 static void hammer_io_flush_mark(hammer_volume_t volume);
68 hammer_mod_rb_compare(hammer_io_t io1, hammer_io_t io2)
70 hammer_off_t io1_offset;
71 hammer_off_t io2_offset;
73 io1_offset = ((io1->offset & HAMMER_OFF_SHORT_MASK) << 8) |
74 HAMMER_VOL_DECODE(io1->offset);
75 io2_offset = ((io2->offset & HAMMER_OFF_SHORT_MASK) << 8) |
76 HAMMER_VOL_DECODE(io2->offset);
78 if (io1_offset < io2_offset)
80 if (io1_offset > io2_offset)
85 RB_GENERATE(hammer_mod_rb_tree, hammer_io, rb_node, hammer_mod_rb_compare);
88 * Initialize a new, already-zero'd hammer_io structure, or reinitialize
89 * an existing hammer_io structure which may have switched to another type.
92 hammer_io_init(hammer_io_t io, hammer_volume_t volume, enum hammer_io_type type)
95 io->hmp = volume->io.hmp;
100 * Helper routine to disassociate a buffer cache buffer from an I/O
101 * structure. The io must be interlocked and marked appropriately for
104 * The io must be in a released state with the io->bp owned and
105 * locked by the caller of this function. When not called from an
106 * io_deallocate() this cannot race an io_deallocate() since the
107 * kernel would be unable to get the buffer lock in that case.
108 * (The released state in this case means we own the bp, not the
109 * hammer_io structure).
111 * The io may have 0 or 1 references depending on who called us. The
112 * caller is responsible for dealing with the refs.
114 * This call can only be made when no action is required on the buffer.
116 * This function is guaranteed not to race against anything because we
117 * own both the io lock and the bp lock and are interlocked with no
121 hammer_io_disassociate(hammer_io_structure_t iou)
123 struct buf *bp = iou->io.bp;
125 KKASSERT(iou->io.released);
126 KKASSERT(iou->io.modified == 0);
127 KKASSERT(LIST_FIRST(&bp->b_dep) == (void *)iou);
132 * If the buffer was locked someone wanted to get rid of it.
134 if (bp->b_flags & B_LOCKED) {
135 atomic_add_int(&hammer_count_io_locked, -1);
136 bp->b_flags &= ~B_LOCKED;
138 if (iou->io.reclaim) {
139 bp->b_flags |= B_NOCACHE|B_RELBUF;
143 switch(iou->io.type) {
144 case HAMMER_STRUCTURE_VOLUME:
145 iou->volume.ondisk = NULL;
147 case HAMMER_STRUCTURE_DATA_BUFFER:
148 case HAMMER_STRUCTURE_META_BUFFER:
149 case HAMMER_STRUCTURE_UNDO_BUFFER:
150 iou->buffer.ondisk = NULL;
152 case HAMMER_STRUCTURE_DUMMY:
153 panic("hammer_io_disassociate: bad io type");
159 * Wait for any physical IO to complete
161 * XXX we aren't interlocked against a spinlock or anything so there
162 * is a small window in the interlock / io->running == 0 test.
165 hammer_io_wait(hammer_io_t io)
168 hammer_mount_t hmp = io->hmp;
170 lwkt_gettoken(&hmp->io_token);
171 while (io->running) {
173 tsleep_interlock(io, 0);
175 tsleep(io, PINTERLOCKED, "hmrflw", hz);
177 lwkt_reltoken(&hmp->io_token);
182 * Wait for all currently queued HAMMER-initiated I/Os to complete.
184 * This is not supposed to count direct I/O's but some can leak
185 * through (for non-full-sized direct I/Os).
188 hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush)
190 struct hammer_io iodummy;
194 * Degenerate case, no I/O is running
196 lwkt_gettoken(&hmp->io_token);
197 if (TAILQ_EMPTY(&hmp->iorun_list)) {
198 lwkt_reltoken(&hmp->io_token);
200 hammer_io_flush_sync(hmp);
203 bzero(&iodummy, sizeof(iodummy));
204 iodummy.type = HAMMER_STRUCTURE_DUMMY;
207 * Add placemarker and then wait until it becomes the head of
210 TAILQ_INSERT_TAIL(&hmp->iorun_list, &iodummy, iorun_entry);
211 while (TAILQ_FIRST(&hmp->iorun_list) != &iodummy) {
212 tsleep(&iodummy, 0, ident, 0);
216 * Chain in case several placemarkers are present.
218 TAILQ_REMOVE(&hmp->iorun_list, &iodummy, iorun_entry);
219 io = TAILQ_FIRST(&hmp->iorun_list);
220 if (io && io->type == HAMMER_STRUCTURE_DUMMY)
222 lwkt_reltoken(&hmp->io_token);
225 hammer_io_flush_sync(hmp);
229 * Clear a flagged error condition on a I/O buffer. The caller must hold
230 * its own ref on the buffer.
233 hammer_io_clear_error(struct hammer_io *io)
235 hammer_mount_t hmp = io->hmp;
237 lwkt_gettoken(&hmp->io_token);
240 hammer_rel(&io->lock);
241 KKASSERT(hammer_isactive(&io->lock));
243 lwkt_reltoken(&hmp->io_token);
247 hammer_io_clear_error_noassert(struct hammer_io *io)
249 hammer_mount_t hmp = io->hmp;
251 lwkt_gettoken(&hmp->io_token);
254 hammer_rel(&io->lock);
256 lwkt_reltoken(&hmp->io_token);
260 * This is an advisory function only which tells the buffer cache
261 * the bp is not a meta-data buffer, even though it is backed by
264 * This is used by HAMMER's reblocking code to avoid trying to
265 * swapcache the filesystem's data when it is read or written
266 * by the reblocking code.
268 * The caller has a ref on the buffer preventing the bp from
269 * being disassociated from it.
272 hammer_io_notmeta(hammer_buffer_t buffer)
274 if ((buffer->io.bp->b_flags & B_NOTMETA) == 0) {
275 hammer_mount_t hmp = buffer->io.hmp;
277 lwkt_gettoken(&hmp->io_token);
278 buffer->io.bp->b_flags |= B_NOTMETA;
279 lwkt_reltoken(&hmp->io_token);
284 * Load bp for a HAMMER structure. The io must be exclusively locked by
287 * This routine is mostly used on meta-data and small-data blocks. Generally
288 * speaking HAMMER assumes some locality of reference and will cluster.
290 * Note that the caller (hammer_ondisk.c) may place further restrictions
291 * on clusterability via the limit (in bytes). Typically large-data
292 * zones cannot be clustered due to their mixed buffer sizes. This is
293 * not an issue since such clustering occurs in hammer_vnops at the
294 * regular file layer, whereas this is the buffered block device layer.
296 * No I/O callbacks can occur while we hold the buffer locked.
299 hammer_io_read(struct vnode *devvp, struct hammer_io *io, int limit)
304 if ((bp = io->bp) == NULL) {
305 atomic_add_int(&hammer_count_io_running_read, io->bytes);
306 if (hammer_cluster_enable && limit > io->bytes) {
307 error = cluster_read(devvp, io->offset + limit,
308 io->offset, io->bytes,
313 error = bread(devvp, io->offset, io->bytes, &io->bp);
315 hammer_stats_disk_read += io->bytes;
316 atomic_add_int(&hammer_count_io_running_read, -io->bytes);
319 * The code generally assumes b_ops/b_dep has been set-up,
320 * even if we error out here.
323 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
324 const char *metatype;
327 case HAMMER_STRUCTURE_VOLUME:
330 case HAMMER_STRUCTURE_META_BUFFER:
331 switch(((struct hammer_buffer *)io)->
332 zoneX_offset & HAMMER_OFF_ZONE_MASK) {
333 case HAMMER_ZONE_BTREE:
336 case HAMMER_ZONE_META:
339 case HAMMER_ZONE_FREEMAP:
340 metatype = "freemap";
347 case HAMMER_STRUCTURE_DATA_BUFFER:
350 case HAMMER_STRUCTURE_UNDO_BUFFER:
354 metatype = "unknown";
357 kprintf("doff %016jx %s\n",
358 (intmax_t)bp->b_bio2.bio_offset,
361 bp->b_flags &= ~B_IODEBUG;
362 bp->b_ops = &hammer_bioops;
363 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
365 /* io->worklist is locked by the io lock */
366 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
368 KKASSERT(io->modified == 0);
369 KKASSERT(io->running == 0);
370 KKASSERT(io->waiting == 0);
371 io->released = 0; /* we hold an active lock on bp */
379 * Similar to hammer_io_read() but returns a zero'd out buffer instead.
380 * Must be called with the IO exclusively locked.
382 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background
383 * I/O by forcing the buffer to not be in a released state before calling
386 * This function will also mark the IO as modified but it will not
387 * increment the modify_refs count.
389 * No I/O callbacks can occur while we hold the buffer locked.
392 hammer_io_new(struct vnode *devvp, struct hammer_io *io)
396 if ((bp = io->bp) == NULL) {
397 io->bp = getblk(devvp, io->offset, io->bytes, 0, 0);
399 bp->b_ops = &hammer_bioops;
400 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
402 /* io->worklist is locked by the io lock */
403 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
405 KKASSERT(io->running == 0);
415 hammer_io_modify(io, 0);
421 * Advance the activity count on the underlying buffer because
422 * HAMMER does not getblk/brelse on every access.
424 * The io->bp cannot go away while the buffer is referenced.
427 hammer_io_advance(struct hammer_io *io)
430 buf_act_advance(io->bp);
434 * Remove potential device level aliases against buffers managed by high level
435 * vnodes. Aliases can also be created due to mixed buffer sizes or via
436 * direct access to the backing store device.
438 * This is nasty because the buffers are also VMIO-backed. Even if a buffer
439 * does not exist its backing VM pages might, and we have to invalidate
440 * those as well or a getblk() will reinstate them.
442 * Buffer cache buffers associated with hammer_buffers cannot be
446 hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset)
448 hammer_io_structure_t iou;
450 hammer_off_t phys_offset;
454 hmp = volume->io.hmp;
455 lwkt_gettoken(&hmp->io_token);
458 * If a device buffer already exists for the specified physical
459 * offset use that, otherwise instantiate a buffer to cover any
460 * related VM pages, set BNOCACHE, and brelse().
462 phys_offset = volume->ondisk->vol_buf_beg +
463 (zone2_offset & HAMMER_OFF_SHORT_MASK);
464 if ((bp = findblk(volume->devvp, phys_offset, 0)) != NULL)
467 bp = getblk(volume->devvp, phys_offset, HAMMER_BUFSIZE, 0, 0);
469 if ((iou = (void *)LIST_FIRST(&bp->b_dep)) != NULL) {
471 hammer_ref(&iou->io.lock);
472 hammer_io_clear_modify(&iou->io, 1);
474 iou->io.released = 0;
477 iou->io.waitdep = 1; /* XXX this is a fs_token field */
478 KKASSERT(hammer_isactive(&iou->io.lock) == 1);
479 hammer_rel_buffer(&iou->buffer, 0);
480 /*hammer_io_deallocate(bp);*/
485 KKASSERT((bp->b_flags & B_LOCKED) == 0);
487 bp->b_flags |= B_NOCACHE|B_RELBUF;
491 lwkt_reltoken(&hmp->io_token);
496 * This routine is called on the last reference to a hammer structure.
497 * The io must be interlocked with a refcount of zero. The hammer structure
498 * will remain interlocked on return.
500 * This routine may return a non-NULL bp to the caller for dispoal.
501 * The caller typically brelse()'s the bp.
503 * The bp may or may not still be passively associated with the IO. It
504 * will remain passively associated if it is unreleasable (e.g. a modified
507 * The only requirement here is that modified meta-data and volume-header
508 * buffer may NOT be disassociated from the IO structure, and consequently
509 * we also leave such buffers actively associated with the IO if they already
510 * are (since the kernel can't do anything with them anyway). Only the
511 * flusher is allowed to write such buffers out. Modified pure-data and
512 * undo buffers are returned to the kernel but left passively associated
513 * so we can track when the kernel writes the bp out.
516 hammer_io_release(struct hammer_io *io, int flush)
518 union hammer_io_structure *iou = (void *)io;
521 if ((bp = io->bp) == NULL)
525 * Try to flush a dirty IO to disk if asked to by the
526 * caller or if the kernel tried to flush the buffer in the past.
528 * Kernel-initiated flushes are only allowed for pure-data buffers.
529 * meta-data and volume buffers can only be flushed explicitly
534 hammer_io_flush(io, 0);
535 } else if (bp->b_flags & B_LOCKED) {
537 case HAMMER_STRUCTURE_DATA_BUFFER:
538 hammer_io_flush(io, 0);
540 case HAMMER_STRUCTURE_UNDO_BUFFER:
541 hammer_io_flush(io, hammer_undo_reclaim(io));
546 } /* else no explicit request to flush the buffer */
550 * Wait for the IO to complete if asked to. This occurs when
551 * the buffer must be disposed of definitively during an umount
552 * or buffer invalidation.
554 if (io->waitdep && io->running) {
559 * Return control of the buffer to the kernel (with the provisio
560 * that our bioops can override kernel decisions with regards to
563 if ((flush || io->reclaim) && io->modified == 0 && io->running == 0) {
565 * Always disassociate the bp if an explicit flush
566 * was requested and the IO completed with no error
567 * (so unmount can really clean up the structure).
575 hammer_io_disassociate((hammer_io_structure_t)io);
577 } else if (io->modified) {
579 * Only certain IO types can be released to the kernel if
580 * the buffer has been modified.
582 * volume and meta-data IO types may only be explicitly
586 case HAMMER_STRUCTURE_DATA_BUFFER:
587 case HAMMER_STRUCTURE_UNDO_BUFFER:
588 if (io->released == 0) {
596 bp = NULL; /* bp left associated */
597 } else if (io->released == 0) {
599 * Clean buffers can be generally released to the kernel.
600 * We leave the bp passively associated with the HAMMER
601 * structure and use bioops to disconnect it later on
602 * if the kernel wants to discard the buffer.
604 * We can steal the structure's ownership of the bp.
607 if (bp->b_flags & B_LOCKED) {
608 hammer_io_disassociate(iou);
612 hammer_io_disassociate(iou);
615 /* return the bp (bp passively associated) */
620 * A released buffer is passively associate with our
621 * hammer_io structure. The kernel cannot destroy it
622 * without making a bioops call. If the kernel (B_LOCKED)
623 * or we (reclaim) requested that the buffer be destroyed
624 * we destroy it, otherwise we do a quick get/release to
625 * reset its position in the kernel's LRU list.
627 * Leaving the buffer passively associated allows us to
628 * use the kernel's LRU buffer flushing mechanisms rather
629 * then rolling our own.
631 * XXX there are two ways of doing this. We can re-acquire
632 * and passively release to reset the LRU, or not.
634 if (io->running == 0) {
636 if ((bp->b_flags & B_LOCKED) || io->reclaim) {
637 hammer_io_disassociate(iou);
640 /* return the bp (bp passively associated) */
644 * bp is left passively associated but we do not
645 * try to reacquire it. Interactions with the io
646 * structure will occur on completion of the bp's
656 * This routine is called with a locked IO when a flush is desired and
657 * no other references to the structure exists other then ours. This
658 * routine is ONLY called when HAMMER believes it is safe to flush a
659 * potentially modified buffer out.
661 * The locked io or io reference prevents a flush from being initiated
665 hammer_io_flush(struct hammer_io *io, int reclaim)
671 * Degenerate case - nothing to flush if nothing is dirty.
673 if (io->modified == 0)
677 KKASSERT(io->modify_refs <= 0);
680 * Acquire ownership of the bp, particularly before we clear our
683 * We are going to bawrite() this bp. Don't leave a window where
684 * io->released is set, we actually own the bp rather then our
687 * The io_token should not be required here as only
693 /* BUF_KERNPROC(io->bp); */
694 /* io->released = 0; */
695 KKASSERT(io->released);
696 KKASSERT(io->bp == bp);
703 if ((bp->b_flags & B_LOCKED) == 0) {
704 bp->b_flags |= B_LOCKED;
705 atomic_add_int(&hammer_count_io_locked, 1);
710 * Acquire exclusive access to the bp and then clear the modified
711 * state of the buffer prior to issuing I/O to interlock any
712 * modifications made while the I/O is in progress. This shouldn't
713 * happen anyway but losing data would be worse. The modified bit
714 * will be rechecked after the IO completes.
716 * NOTE: This call also finalizes the buffer's content (inval == 0).
718 * This is only legal when lock.refs == 1 (otherwise we might clear
719 * the modified bit while there are still users of the cluster
720 * modifying the data).
722 * Do this before potentially blocking so any attempt to modify the
723 * ondisk while we are blocked blocks waiting for us.
725 hammer_ref(&io->lock);
726 hammer_io_clear_modify(io, 0);
727 hammer_rel(&io->lock);
729 if (hammer_debug_io & 0x0002)
730 kprintf("hammer io_write %016jx\n", bp->b_bio1.bio_offset);
733 * Transfer ownership to the kernel and initiate I/O.
735 * NOTE: We do not hold io_token so an atomic op is required to
736 * update io_running_space.
739 atomic_add_int(&hmp->io_running_space, io->bytes);
740 atomic_add_int(&hammer_count_io_running_write, io->bytes);
741 lwkt_gettoken(&hmp->io_token);
742 TAILQ_INSERT_TAIL(&hmp->iorun_list, io, iorun_entry);
743 lwkt_reltoken(&hmp->io_token);
745 hammer_io_flush_mark(io->volume);
748 /************************************************************************
750 ************************************************************************
752 * These routines deal with dependancies created when IO buffers get
753 * modified. The caller must call hammer_modify_*() on a referenced
754 * HAMMER structure prior to modifying its on-disk data.
756 * Any intent to modify an IO buffer acquires the related bp and imposes
757 * various write ordering dependancies.
761 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
762 * are locked until the flusher can deal with them, pure data buffers
763 * can be written out.
765 * The referenced io prevents races.
769 hammer_io_modify(hammer_io_t io, int count)
772 * io->modify_refs must be >= 0
774 while (io->modify_refs < 0) {
776 tsleep(io, 0, "hmrmod", 0);
780 * Shortcut if nothing to do.
782 KKASSERT(hammer_isactive(&io->lock) && io->bp != NULL);
783 io->modify_refs += count;
784 if (io->modified && io->released == 0)
788 * NOTE: It is important not to set the modified bit
789 * until after we have acquired the bp or we risk
790 * racing against checkwrite.
792 hammer_lock_ex(&io->lock);
795 BUF_KERNPROC(io->bp);
798 if (io->modified == 0) {
799 hammer_io_set_modlist(io);
802 hammer_unlock(&io->lock);
807 hammer_io_modify_done(hammer_io_t io)
809 KKASSERT(io->modify_refs > 0);
811 if (io->modify_refs == 0 && io->waitmod) {
818 * The write interlock blocks other threads trying to modify a buffer
819 * (they block in hammer_io_modify()) after us, or blocks us while other
820 * threads are in the middle of modifying a buffer.
822 * The caller also has a ref on the io, however if we are not careful
823 * we will race bioops callbacks (checkwrite). To deal with this
824 * we must at least acquire and release the io_token, and it is probably
825 * better to hold it through the setting of modify_refs.
828 hammer_io_write_interlock(hammer_io_t io)
830 hammer_mount_t hmp = io->hmp;
832 lwkt_gettoken(&hmp->io_token);
833 while (io->modify_refs != 0) {
835 tsleep(io, 0, "hmrmod", 0);
837 io->modify_refs = -1;
838 lwkt_reltoken(&hmp->io_token);
842 hammer_io_done_interlock(hammer_io_t io)
844 KKASSERT(io->modify_refs == -1);
853 * Caller intends to modify a volume's ondisk structure.
855 * This is only allowed if we are the flusher or we have a ref on the
859 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
862 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
864 hammer_io_modify(&volume->io, 1);
866 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk;
867 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
868 hammer_generate_undo(trans,
869 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset),
875 * Caller intends to modify a buffer's ondisk structure.
877 * This is only allowed if we are the flusher or we have a ref on the
881 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
884 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
886 hammer_io_modify(&buffer->io, 1);
888 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk;
889 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
890 hammer_generate_undo(trans,
891 buffer->zone2_offset + rel_offset,
897 hammer_modify_volume_done(hammer_volume_t volume)
899 hammer_io_modify_done(&volume->io);
903 hammer_modify_buffer_done(hammer_buffer_t buffer)
905 hammer_io_modify_done(&buffer->io);
909 * Mark an entity as not being dirty any more and finalize any
910 * delayed adjustments to the buffer.
912 * Delayed adjustments are an important performance enhancement, allowing
913 * us to avoid recalculating B-Tree node CRCs over and over again when
914 * making bulk-modifications to the B-Tree.
916 * If inval is non-zero delayed adjustments are ignored.
918 * This routine may dereference related btree nodes and cause the
919 * buffer to be dereferenced. The caller must own a reference on io.
922 hammer_io_clear_modify(struct hammer_io *io, int inval)
927 * io_token is needed to avoid races on mod_root
929 if (io->modified == 0)
932 lwkt_gettoken(&hmp->io_token);
933 if (io->modified == 0) {
934 lwkt_reltoken(&hmp->io_token);
939 * Take us off the mod-list and clear the modified bit.
941 KKASSERT(io->mod_root != NULL);
942 if (io->mod_root == &io->hmp->volu_root ||
943 io->mod_root == &io->hmp->meta_root) {
944 io->hmp->locked_dirty_space -= io->bytes;
945 atomic_add_int(&hammer_count_dirtybufspace, -io->bytes);
947 RB_REMOVE(hammer_mod_rb_tree, io->mod_root, io);
951 lwkt_reltoken(&hmp->io_token);
954 * If this bit is not set there are no delayed adjustments.
961 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference
962 * on the node (& underlying buffer). Release the node after clearing
965 if (io->type == HAMMER_STRUCTURE_META_BUFFER) {
966 hammer_buffer_t buffer = (void *)io;
970 TAILQ_FOREACH(node, &buffer->clist, entry) {
971 if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0)
973 node->flags &= ~HAMMER_NODE_NEEDSCRC;
974 KKASSERT(node->ondisk);
976 node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE);
977 hammer_rel_node(node);
981 /* caller must still have ref on io */
982 KKASSERT(hammer_isactive(&io->lock));
986 * Clear the IO's modify list. Even though the IO is no longer modified
987 * it may still be on the lose_root. This routine is called just before
988 * the governing hammer_buffer is destroyed.
990 * mod_root requires io_token protection.
993 hammer_io_clear_modlist(struct hammer_io *io)
995 hammer_mount_t hmp = io->hmp;
997 KKASSERT(io->modified == 0);
999 lwkt_gettoken(&hmp->io_token);
1001 KKASSERT(io->mod_root == &io->hmp->lose_root);
1002 RB_REMOVE(hammer_mod_rb_tree, io->mod_root, io);
1003 io->mod_root = NULL;
1005 lwkt_reltoken(&hmp->io_token);
1010 hammer_io_set_modlist(struct hammer_io *io)
1012 struct hammer_mount *hmp = io->hmp;
1014 lwkt_gettoken(&hmp->io_token);
1015 KKASSERT(io->mod_root == NULL);
1018 case HAMMER_STRUCTURE_VOLUME:
1019 io->mod_root = &hmp->volu_root;
1020 hmp->locked_dirty_space += io->bytes;
1021 atomic_add_int(&hammer_count_dirtybufspace, io->bytes);
1023 case HAMMER_STRUCTURE_META_BUFFER:
1024 io->mod_root = &hmp->meta_root;
1025 hmp->locked_dirty_space += io->bytes;
1026 atomic_add_int(&hammer_count_dirtybufspace, io->bytes);
1028 case HAMMER_STRUCTURE_UNDO_BUFFER:
1029 io->mod_root = &hmp->undo_root;
1031 case HAMMER_STRUCTURE_DATA_BUFFER:
1032 io->mod_root = &hmp->data_root;
1034 case HAMMER_STRUCTURE_DUMMY:
1035 panic("hammer_io_set_modlist: bad io type");
1036 break; /* NOT REACHED */
1038 if (RB_INSERT(hammer_mod_rb_tree, io->mod_root, io)) {
1039 panic("hammer_io_set_modlist: duplicate entry");
1042 lwkt_reltoken(&hmp->io_token);
1045 /************************************************************************
1047 ************************************************************************
1052 * Pre-IO initiation kernel callback - cluster build only
1054 * bioops callback - hold io_token
1057 hammer_io_start(struct buf *bp)
1059 /* nothing to do, so io_token not needed */
1063 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT!
1065 * NOTE: HAMMER may modify a data buffer after we have initiated write
1068 * NOTE: MPSAFE callback
1070 * bioops callback - hold io_token
1073 hammer_io_complete(struct buf *bp)
1075 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep);
1076 struct hammer_mount *hmp = iou->io.hmp;
1077 struct hammer_io *ionext;
1079 lwkt_gettoken(&hmp->io_token);
1081 KKASSERT(iou->io.released == 1);
1084 * Deal with people waiting for I/O to drain
1086 if (iou->io.running) {
1088 * Deal with critical write errors. Once a critical error
1089 * has been flagged in hmp the UNDO FIFO will not be updated.
1090 * That way crash recover will give us a consistent
1093 * Because of this we can throw away failed UNDO buffers. If
1094 * we throw away META or DATA buffers we risk corrupting
1095 * the now read-only version of the filesystem visible to
1096 * the user. Clear B_ERROR so the buffer is not re-dirtied
1097 * by the kernel and ref the io so it doesn't get thrown
1100 if (bp->b_flags & B_ERROR) {
1101 lwkt_gettoken(&hmp->fs_token);
1102 hammer_critical_error(hmp, NULL, bp->b_error,
1103 "while flushing meta-data");
1104 lwkt_reltoken(&hmp->fs_token);
1106 switch(iou->io.type) {
1107 case HAMMER_STRUCTURE_UNDO_BUFFER:
1110 if (iou->io.ioerror == 0) {
1111 iou->io.ioerror = 1;
1112 hammer_ref(&iou->io.lock);
1116 bp->b_flags &= ~B_ERROR;
1119 hammer_io_set_modlist(&iou->io);
1120 iou->io.modified = 1;
1123 hammer_stats_disk_write += iou->io.bytes;
1124 atomic_add_int(&hammer_count_io_running_write, -iou->io.bytes);
1125 atomic_add_int(&hmp->io_running_space, -iou->io.bytes);
1126 if (hmp->io_running_wakeup &&
1127 hmp->io_running_space < hammer_limit_running_io / 2) {
1128 hmp->io_running_wakeup = 0;
1129 wakeup(&hmp->io_running_wakeup);
1131 KKASSERT(hmp->io_running_space >= 0);
1132 iou->io.running = 0;
1135 * Remove from iorun list and wakeup any multi-io waiter(s).
1137 if (TAILQ_FIRST(&hmp->iorun_list) == &iou->io) {
1138 ionext = TAILQ_NEXT(&iou->io, iorun_entry);
1139 if (ionext && ionext->type == HAMMER_STRUCTURE_DUMMY)
1142 TAILQ_REMOVE(&hmp->iorun_list, &iou->io, iorun_entry);
1144 hammer_stats_disk_read += iou->io.bytes;
1147 if (iou->io.waiting) {
1148 iou->io.waiting = 0;
1153 * If B_LOCKED is set someone wanted to deallocate the bp at some
1154 * point, try to do it now. The operation will fail if there are
1155 * refs or if hammer_io_deallocate() is unable to gain the
1158 if (bp->b_flags & B_LOCKED) {
1159 atomic_add_int(&hammer_count_io_locked, -1);
1160 bp->b_flags &= ~B_LOCKED;
1161 hammer_io_deallocate(bp);
1162 /* structure may be dead now */
1164 lwkt_reltoken(&hmp->io_token);
1168 * Callback from kernel when it wishes to deallocate a passively
1169 * associated structure. This mostly occurs with clean buffers
1170 * but it may be possible for a holding structure to be marked dirty
1171 * while its buffer is passively associated. The caller owns the bp.
1173 * If we cannot disassociate we set B_LOCKED to prevent the buffer
1174 * from getting reused.
1176 * WARNING: Because this can be called directly by getnewbuf we cannot
1177 * recurse into the tree. If a bp cannot be immediately disassociated
1178 * our only recourse is to set B_LOCKED.
1180 * WARNING: This may be called from an interrupt via hammer_io_complete()
1182 * bioops callback - hold io_token
1185 hammer_io_deallocate(struct buf *bp)
1187 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep);
1192 lwkt_gettoken(&hmp->io_token);
1194 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0);
1195 if (hammer_try_interlock_norefs(&iou->io.lock) == 0) {
1197 * We cannot safely disassociate a bp from a referenced
1198 * or interlocked HAMMER structure.
1200 bp->b_flags |= B_LOCKED;
1201 atomic_add_int(&hammer_count_io_locked, 1);
1202 } else if (iou->io.modified) {
1204 * It is not legal to disassociate a modified buffer. This
1205 * case really shouldn't ever occur.
1207 bp->b_flags |= B_LOCKED;
1208 atomic_add_int(&hammer_count_io_locked, 1);
1209 hammer_put_interlock(&iou->io.lock, 0);
1212 * Disassociate the BP. If the io has no refs left we
1213 * have to add it to the loose list. The kernel has
1214 * locked the buffer and therefore our io must be
1215 * in a released state.
1217 hammer_io_disassociate(iou);
1218 if (iou->io.type != HAMMER_STRUCTURE_VOLUME) {
1219 KKASSERT(iou->io.bp == NULL);
1220 KKASSERT(iou->io.mod_root == NULL);
1221 iou->io.mod_root = &hmp->lose_root;
1222 if (RB_INSERT(hammer_mod_rb_tree, iou->io.mod_root,
1224 panic("hammer_io_deallocate: duplicate entry");
1227 hammer_put_interlock(&iou->io.lock, 1);
1229 lwkt_reltoken(&hmp->io_token);
1233 * bioops callback - hold io_token
1236 hammer_io_fsync(struct vnode *vp)
1238 /* nothing to do, so io_token not needed */
1243 * NOTE: will not be called unless we tell the kernel about the
1244 * bioops. Unused... we use the mount's VFS_SYNC instead.
1246 * bioops callback - hold io_token
1249 hammer_io_sync(struct mount *mp)
1251 /* nothing to do, so io_token not needed */
1256 * bioops callback - hold io_token
1259 hammer_io_movedeps(struct buf *bp1, struct buf *bp2)
1261 /* nothing to do, so io_token not needed */
1265 * I/O pre-check for reading and writing. HAMMER only uses this for
1266 * B_CACHE buffers so checkread just shouldn't happen, but if it does
1269 * Writing is a different case. We don't want the kernel to try to write
1270 * out a buffer that HAMMER may be modifying passively or which has a
1271 * dependancy. In addition, kernel-demanded writes can only proceed for
1272 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
1273 * buffer types can only be explicitly written by the flusher.
1275 * checkwrite will only be called for bdwrite()n buffers. If we return
1276 * success the kernel is guaranteed to initiate the buffer write.
1278 * bioops callback - hold io_token
1281 hammer_io_checkread(struct buf *bp)
1283 /* nothing to do, so io_token not needed */
1288 * The kernel is asking us whether it can write out a dirty buffer or not.
1290 * bioops callback - hold io_token
1293 hammer_io_checkwrite(struct buf *bp)
1295 hammer_io_t io = (void *)LIST_FIRST(&bp->b_dep);
1296 hammer_mount_t hmp = io->hmp;
1299 * This shouldn't happen under normal operation.
1301 lwkt_gettoken(&hmp->io_token);
1302 if (io->type == HAMMER_STRUCTURE_VOLUME ||
1303 io->type == HAMMER_STRUCTURE_META_BUFFER) {
1305 panic("hammer_io_checkwrite: illegal buffer");
1306 if ((bp->b_flags & B_LOCKED) == 0) {
1307 bp->b_flags |= B_LOCKED;
1308 atomic_add_int(&hammer_count_io_locked, 1);
1310 lwkt_reltoken(&hmp->io_token);
1315 * We have to be able to interlock the IO to safely modify any
1316 * of its fields without holding the fs_token. If we can't lock
1317 * it then we are racing someone.
1319 * Our ownership of the bp lock prevents the io from being ripped
1320 * out from under us.
1322 if (hammer_try_interlock_norefs(&io->lock) == 0) {
1323 bp->b_flags |= B_LOCKED;
1324 atomic_add_int(&hammer_count_io_locked, 1);
1325 lwkt_reltoken(&hmp->io_token);
1330 * The modified bit must be cleared prior to the initiation of
1331 * any IO (returning 0 initiates the IO). Because this is a
1332 * normal data buffer hammer_io_clear_modify() runs through a
1333 * simple degenerate case.
1335 * Return 0 will cause the kernel to initiate the IO, and we
1336 * must normally clear the modified bit before we begin. If
1337 * the io has modify_refs we do not clear the modified bit,
1338 * otherwise we may miss changes.
1340 * Only data and undo buffers can reach here. These buffers do
1341 * not have terminal crc functions but we temporarily reference
1342 * the IO anyway, just in case.
1344 if (io->modify_refs == 0 && io->modified) {
1345 hammer_ref(&io->lock);
1346 hammer_io_clear_modify(io, 0);
1347 hammer_rel(&io->lock);
1348 } else if (io->modified) {
1349 KKASSERT(io->type == HAMMER_STRUCTURE_DATA_BUFFER);
1353 * The kernel is going to start the IO, set io->running.
1355 KKASSERT(io->running == 0);
1357 atomic_add_int(&io->hmp->io_running_space, io->bytes);
1358 atomic_add_int(&hammer_count_io_running_write, io->bytes);
1359 TAILQ_INSERT_TAIL(&io->hmp->iorun_list, io, iorun_entry);
1361 hammer_put_interlock(&io->lock, 1);
1362 lwkt_reltoken(&hmp->io_token);
1368 * Return non-zero if we wish to delay the kernel's attempt to flush
1369 * this buffer to disk.
1371 * bioops callback - hold io_token
1374 hammer_io_countdeps(struct buf *bp, int n)
1376 /* nothing to do, so io_token not needed */
1380 struct bio_ops hammer_bioops = {
1381 .io_start = hammer_io_start,
1382 .io_complete = hammer_io_complete,
1383 .io_deallocate = hammer_io_deallocate,
1384 .io_fsync = hammer_io_fsync,
1385 .io_sync = hammer_io_sync,
1386 .io_movedeps = hammer_io_movedeps,
1387 .io_countdeps = hammer_io_countdeps,
1388 .io_checkread = hammer_io_checkread,
1389 .io_checkwrite = hammer_io_checkwrite,
1392 /************************************************************************
1394 ************************************************************************
1396 * These functions operate directly on the buffer cache buffer associated
1397 * with a front-end vnode rather then a back-end device vnode.
1401 * Read a buffer associated with a front-end vnode directly from the
1402 * disk media. The bio may be issued asynchronously. If leaf is non-NULL
1403 * we validate the CRC.
1405 * We must check for the presence of a HAMMER buffer to handle the case
1406 * where the reblocker has rewritten the data (which it does via the HAMMER
1407 * buffer system, not via the high-level vnode buffer cache), but not yet
1408 * committed the buffer to the media.
1411 hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1412 hammer_btree_leaf_elm_t leaf)
1414 hammer_off_t buf_offset;
1415 hammer_off_t zone2_offset;
1416 hammer_volume_t volume;
1422 buf_offset = bio->bio_offset;
1423 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) ==
1424 HAMMER_ZONE_LARGE_DATA);
1427 * The buffer cache may have an aliased buffer (the reblocker can
1428 * write them). If it does we have to sync any dirty data before
1429 * we can build our direct-read. This is a non-critical code path.
1432 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize);
1435 * Resolve to a zone-2 offset. The conversion just requires
1436 * munging the top 4 bits but we want to abstract it anyway
1437 * so the blockmap code can verify the zone assignment.
1439 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
1442 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
1443 HAMMER_ZONE_RAW_BUFFER);
1446 * Resolve volume and raw-offset for 3rd level bio. The
1447 * offset will be specific to the volume.
1449 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1450 volume = hammer_get_volume(hmp, vol_no, &error);
1451 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1458 nbio = push_bio(bio);
1459 nbio->bio_offset = volume->ondisk->vol_buf_beg +
1460 (zone2_offset & HAMMER_OFF_SHORT_MASK);
1463 * XXX disabled - our CRC check doesn't work if the OS
1464 * does bogus_page replacement on the direct-read.
1466 if (leaf && hammer_verify_data) {
1467 nbio->bio_done = hammer_io_direct_read_complete;
1468 nbio->bio_caller_info1.uvalue32 = leaf->data_crc;
1471 hammer_stats_disk_read += bp->b_bufsize;
1472 vn_strategy(volume->devvp, nbio);
1474 hammer_rel_volume(volume, 0);
1477 kprintf("hammer_direct_read: failed @ %016llx\n",
1478 (long long)zone2_offset);
1479 bp->b_error = error;
1480 bp->b_flags |= B_ERROR;
1488 * On completion of the BIO this callback must check the data CRC
1489 * and chain to the previous bio.
1491 * MPSAFE - since we do not modify and hammer_records we do not need
1494 * NOTE: MPSAFE callback
1498 hammer_io_direct_read_complete(struct bio *nbio)
1502 u_int32_t rec_crc = nbio->bio_caller_info1.uvalue32;
1505 if (crc32(bp->b_data, bp->b_bufsize) != rec_crc) {
1506 kprintf("HAMMER: data_crc error @%016llx/%d\n",
1507 nbio->bio_offset, bp->b_bufsize);
1508 if (hammer_debug_critical)
1509 Debugger("data_crc on read");
1510 bp->b_flags |= B_ERROR;
1513 obio = pop_bio(nbio);
1519 * Write a buffer associated with a front-end vnode directly to the
1520 * disk media. The bio may be issued asynchronously.
1522 * The BIO is associated with the specified record and RECG_DIRECT_IO
1523 * is set. The recorded is added to its object.
1526 hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio,
1527 hammer_record_t record)
1529 hammer_btree_leaf_elm_t leaf = &record->leaf;
1530 hammer_off_t buf_offset;
1531 hammer_off_t zone2_offset;
1532 hammer_volume_t volume;
1533 hammer_buffer_t buffer;
1540 buf_offset = leaf->data_offset;
1542 KKASSERT(buf_offset > HAMMER_ZONE_BTREE);
1543 KKASSERT(bio->bio_buf->b_cmd == BUF_CMD_WRITE);
1546 * Issue or execute the I/O. The new memory record must replace
1547 * the old one before the I/O completes, otherwise a reaquisition of
1548 * the buffer will load the old media data instead of the new.
1550 if ((buf_offset & HAMMER_BUFMASK) == 0 &&
1551 leaf->data_len >= HAMMER_BUFSIZE) {
1553 * We are using the vnode's bio to write directly to the
1554 * media, any hammer_buffer at the same zone-X offset will
1555 * now have stale data.
1557 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
1558 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1559 volume = hammer_get_volume(hmp, vol_no, &error);
1561 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1565 KKASSERT((bp->b_bufsize & HAMMER_BUFMASK) == 0);
1567 hammer_del_buffers(hmp, buf_offset,
1568 zone2_offset, bp->b_bufsize);
1572 * Second level bio - cached zone2 offset.
1574 * (We can put our bio_done function in either the
1575 * 2nd or 3rd level).
1577 nbio = push_bio(bio);
1578 nbio->bio_offset = zone2_offset;
1579 nbio->bio_done = hammer_io_direct_write_complete;
1580 nbio->bio_caller_info1.ptr = record;
1581 record->zone2_offset = zone2_offset;
1582 record->gflags |= HAMMER_RECG_DIRECT_IO |
1583 HAMMER_RECG_DIRECT_INVAL;
1586 * Third level bio - raw offset specific to the
1589 zone2_offset &= HAMMER_OFF_SHORT_MASK;
1590 nbio = push_bio(nbio);
1591 nbio->bio_offset = volume->ondisk->vol_buf_beg +
1593 hammer_stats_disk_write += bp->b_bufsize;
1594 hammer_ip_replace_bulk(hmp, record);
1595 vn_strategy(volume->devvp, nbio);
1596 hammer_io_flush_mark(volume);
1598 hammer_rel_volume(volume, 0);
1601 * Must fit in a standard HAMMER buffer. In this case all
1602 * consumers use the HAMMER buffer system and RECG_DIRECT_IO
1603 * does not need to be set-up.
1605 KKASSERT(((buf_offset ^ (buf_offset + leaf->data_len - 1)) & ~HAMMER_BUFMASK64) == 0);
1607 ptr = hammer_bread(hmp, buf_offset, &error, &buffer);
1610 bp->b_flags |= B_AGE;
1611 hammer_io_modify(&buffer->io, 1);
1612 bcopy(bp->b_data, ptr, leaf->data_len);
1613 hammer_io_modify_done(&buffer->io);
1614 hammer_rel_buffer(buffer, 0);
1616 hammer_ip_replace_bulk(hmp, record);
1622 * Major suckage occured. Also note: The record was
1623 * never added to the tree so we do not have to worry
1624 * about the backend.
1626 kprintf("hammer_direct_write: failed @ %016llx\n",
1627 (long long)leaf->data_offset);
1631 bp->b_flags |= B_ERROR;
1633 record->flags |= HAMMER_RECF_DELETED_FE;
1634 hammer_rel_mem_record(record);
1640 * On completion of the BIO this callback must disconnect
1641 * it from the hammer_record and chain to the previous bio.
1643 * An I/O error forces the mount to read-only. Data buffers
1644 * are not B_LOCKED like meta-data buffers are, so we have to
1645 * throw the buffer away to prevent the kernel from retrying.
1647 * NOTE: MPSAFE callback, only modify fields we have explicit
1648 * access to (the bp and the record->gflags).
1652 hammer_io_direct_write_complete(struct bio *nbio)
1656 hammer_record_t record;
1659 record = nbio->bio_caller_info1.ptr;
1660 KKASSERT(record != NULL);
1661 hmp = record->ip->hmp;
1663 lwkt_gettoken(&hmp->io_token);
1666 obio = pop_bio(nbio);
1667 if (bp->b_flags & B_ERROR) {
1668 lwkt_gettoken(&hmp->fs_token);
1669 hammer_critical_error(hmp, record->ip,
1671 "while writing bulk data");
1672 lwkt_reltoken(&hmp->fs_token);
1673 bp->b_flags |= B_INVAL;
1677 KKASSERT(record->gflags & HAMMER_RECG_DIRECT_IO);
1678 if (record->gflags & HAMMER_RECG_DIRECT_WAIT) {
1679 record->gflags &= ~(HAMMER_RECG_DIRECT_IO |
1680 HAMMER_RECG_DIRECT_WAIT);
1681 /* record can disappear once DIRECT_IO flag is cleared */
1682 wakeup(&record->flags);
1684 record->gflags &= ~HAMMER_RECG_DIRECT_IO;
1685 /* record can disappear once DIRECT_IO flag is cleared */
1687 lwkt_reltoken(&hmp->io_token);
1692 * This is called before a record is either committed to the B-Tree
1693 * or destroyed, to resolve any associated direct-IO.
1695 * (1) We must wait for any direct-IO related to the record to complete.
1697 * (2) We must remove any buffer cache aliases for data accessed via
1698 * leaf->data_offset or zone2_offset so non-direct-IO consumers
1699 * (the mirroring and reblocking code) do not see stale data.
1702 hammer_io_direct_wait(hammer_record_t record)
1704 hammer_mount_t hmp = record->ip->hmp;
1707 * Wait for I/O to complete
1709 if (record->gflags & HAMMER_RECG_DIRECT_IO) {
1710 lwkt_gettoken(&hmp->io_token);
1711 while (record->gflags & HAMMER_RECG_DIRECT_IO) {
1712 record->gflags |= HAMMER_RECG_DIRECT_WAIT;
1713 tsleep(&record->flags, 0, "hmdiow", 0);
1715 lwkt_reltoken(&hmp->io_token);
1719 * Invalidate any related buffer cache aliases associated with the
1720 * backing device. This is needed because the buffer cache buffer
1721 * for file data is associated with the file vnode, not the backing
1724 * XXX I do not think this case can occur any more now that
1725 * reservations ensure that all such buffers are removed before
1726 * an area can be reused.
1728 if (record->gflags & HAMMER_RECG_DIRECT_INVAL) {
1729 KKASSERT(record->leaf.data_offset);
1730 hammer_del_buffers(hmp, record->leaf.data_offset,
1731 record->zone2_offset, record->leaf.data_len,
1733 record->gflags &= ~HAMMER_RECG_DIRECT_INVAL;
1738 * This is called to remove the second-level cached zone-2 offset from
1739 * frontend buffer cache buffers, now stale due to a data relocation.
1740 * These offsets are generated by cluster_read() via VOP_BMAP, or directly
1741 * by hammer_vop_strategy_read().
1743 * This is rather nasty because here we have something like the reblocker
1744 * scanning the raw B-Tree with no held references on anything, really,
1745 * other then a shared lock on the B-Tree node, and we have to access the
1746 * frontend's buffer cache to check for and clean out the association.
1747 * Specifically, if the reblocker is moving data on the disk, these cached
1748 * offsets will become invalid.
1750 * Only data record types associated with the large-data zone are subject
1751 * to direct-io and need to be checked.
1755 hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf)
1757 struct hammer_inode_info iinfo;
1760 if (leaf->base.rec_type != HAMMER_RECTYPE_DATA)
1762 zone = HAMMER_ZONE_DECODE(leaf->data_offset);
1763 if (zone != HAMMER_ZONE_LARGE_DATA_INDEX)
1765 iinfo.obj_id = leaf->base.obj_id;
1766 iinfo.obj_asof = 0; /* unused */
1767 iinfo.obj_localization = leaf->base.localization &
1768 HAMMER_LOCALIZE_PSEUDOFS_MASK;
1769 iinfo.u.leaf = leaf;
1770 hammer_scan_inode_snapshots(hmp, &iinfo,
1771 hammer_io_direct_uncache_callback,
1776 hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data)
1778 hammer_inode_info_t iinfo = data;
1779 hammer_off_t data_offset;
1780 hammer_off_t file_offset;
1787 data_offset = iinfo->u.leaf->data_offset;
1788 file_offset = iinfo->u.leaf->base.key - iinfo->u.leaf->data_len;
1789 blksize = iinfo->u.leaf->data_len;
1790 KKASSERT((blksize & HAMMER_BUFMASK) == 0);
1793 * Warning: FINDBLK_TEST return stable storage but not stable
1794 * contents. It happens to be ok in this case.
1796 hammer_ref(&ip->lock);
1797 if (hammer_get_vnode(ip, &vp) == 0) {
1798 if ((bp = findblk(ip->vp, file_offset, FINDBLK_TEST)) != NULL &&
1799 bp->b_bio2.bio_offset != NOOFFSET) {
1800 bp = getblk(ip->vp, file_offset, blksize, 0, 0);
1801 bp->b_bio2.bio_offset = NOOFFSET;
1806 hammer_rel_inode(ip, 0);
1812 * This function is called when writes may have occured on the volume,
1813 * indicating that the device may be holding cached writes.
1816 hammer_io_flush_mark(hammer_volume_t volume)
1818 atomic_set_int(&volume->vol_flags, HAMMER_VOLF_NEEDFLUSH);
1822 * This function ensures that the device has flushed any cached writes out.
1825 hammer_io_flush_sync(hammer_mount_t hmp)
1827 hammer_volume_t volume;
1828 struct buf *bp_base = NULL;
1831 RB_FOREACH(volume, hammer_vol_rb_tree, &hmp->rb_vols_root) {
1832 if (volume->vol_flags & HAMMER_VOLF_NEEDFLUSH) {
1833 atomic_clear_int(&volume->vol_flags,
1834 HAMMER_VOLF_NEEDFLUSH);
1836 bp->b_bio1.bio_offset = 0;
1839 bp->b_cmd = BUF_CMD_FLUSH;
1840 bp->b_bio1.bio_caller_info1.cluster_head = bp_base;
1841 bp->b_bio1.bio_done = biodone_sync;
1842 bp->b_bio1.bio_flags |= BIO_SYNC;
1844 vn_strategy(volume->devvp, &bp->b_bio1);
1847 while ((bp = bp_base) != NULL) {
1848 bp_base = bp->b_bio1.bio_caller_info1.cluster_head;
1849 biowait(&bp->b_bio1, "hmrFLS");
1855 * Limit the amount of backlog which we allow to build up
1858 hammer_io_limit_backlog(hammer_mount_t hmp)
1860 while (hmp->io_running_space > hammer_limit_running_io) {
1861 hmp->io_running_wakeup = 1;
1862 tsleep(&hmp->io_running_wakeup, 0, "hmiolm", hz / 10);