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);
59 static void hammer_indirect_callback(struct bio *bio);
61 static void hammer_io_direct_read_complete(struct bio *nbio);
63 static void hammer_io_direct_write_complete(struct bio *nbio);
64 static int hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data);
65 static void hammer_io_set_modlist(struct hammer_io *io);
66 static void hammer_io_flush_mark(hammer_volume_t volume);
69 hammer_mod_rb_compare(hammer_io_t io1, hammer_io_t io2)
71 hammer_off_t io1_offset;
72 hammer_off_t io2_offset;
74 io1_offset = ((io1->offset & HAMMER_OFF_SHORT_MASK) << 8) |
75 HAMMER_VOL_DECODE(io1->offset);
76 io2_offset = ((io2->offset & HAMMER_OFF_SHORT_MASK) << 8) |
77 HAMMER_VOL_DECODE(io2->offset);
79 if (io1_offset < io2_offset)
81 if (io1_offset > io2_offset)
86 RB_GENERATE(hammer_mod_rb_tree, hammer_io, rb_node, hammer_mod_rb_compare);
89 * Initialize a new, already-zero'd hammer_io structure, or reinitialize
90 * an existing hammer_io structure which may have switched to another type.
93 hammer_io_init(hammer_io_t io, hammer_volume_t volume, enum hammer_io_type type)
96 io->hmp = volume->io.hmp;
101 * Helper routine to disassociate a buffer cache buffer from an I/O
102 * structure. The io must be interlocked and marked appropriately for
105 * The io must be in a released state with the io->bp owned and
106 * locked by the caller of this function. When not called from an
107 * io_deallocate() this cannot race an io_deallocate() since the
108 * kernel would be unable to get the buffer lock in that case.
109 * (The released state in this case means we own the bp, not the
110 * hammer_io structure).
112 * The io may have 0 or 1 references depending on who called us. The
113 * caller is responsible for dealing with the refs.
115 * This call can only be made when no action is required on the buffer.
117 * This function is guaranteed not to race against anything because we
118 * own both the io lock and the bp lock and are interlocked with no
122 hammer_io_disassociate(hammer_io_structure_t iou)
124 struct buf *bp = iou->io.bp;
126 KKASSERT(iou->io.released);
127 KKASSERT(iou->io.modified == 0);
128 KKASSERT(LIST_FIRST(&bp->b_dep) == (void *)iou);
133 * If the buffer was locked someone wanted to get rid of it.
135 if (bp->b_flags & B_LOCKED) {
136 atomic_add_int(&hammer_count_io_locked, -1);
137 bp->b_flags &= ~B_LOCKED;
139 if (iou->io.reclaim) {
140 bp->b_flags |= B_NOCACHE|B_RELBUF;
144 switch(iou->io.type) {
145 case HAMMER_STRUCTURE_VOLUME:
146 iou->volume.ondisk = NULL;
148 case HAMMER_STRUCTURE_DATA_BUFFER:
149 case HAMMER_STRUCTURE_META_BUFFER:
150 case HAMMER_STRUCTURE_UNDO_BUFFER:
151 iou->buffer.ondisk = NULL;
153 case HAMMER_STRUCTURE_DUMMY:
154 panic("hammer_io_disassociate: bad io type");
160 * Wait for any physical IO to complete
162 * XXX we aren't interlocked against a spinlock or anything so there
163 * is a small window in the interlock / io->running == 0 test.
166 hammer_io_wait(hammer_io_t io)
169 hammer_mount_t hmp = io->hmp;
171 lwkt_gettoken(&hmp->io_token);
172 while (io->running) {
174 tsleep_interlock(io, 0);
176 tsleep(io, PINTERLOCKED, "hmrflw", hz);
178 lwkt_reltoken(&hmp->io_token);
183 * Wait for all currently queued HAMMER-initiated I/Os to complete.
185 * This is not supposed to count direct I/O's but some can leak
186 * through (for non-full-sized direct I/Os).
189 hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush)
191 struct hammer_io iodummy;
195 * Degenerate case, no I/O is running
197 lwkt_gettoken(&hmp->io_token);
198 if (TAILQ_EMPTY(&hmp->iorun_list)) {
199 lwkt_reltoken(&hmp->io_token);
201 hammer_io_flush_sync(hmp);
204 bzero(&iodummy, sizeof(iodummy));
205 iodummy.type = HAMMER_STRUCTURE_DUMMY;
208 * Add placemarker and then wait until it becomes the head of
211 TAILQ_INSERT_TAIL(&hmp->iorun_list, &iodummy, iorun_entry);
212 while (TAILQ_FIRST(&hmp->iorun_list) != &iodummy) {
213 tsleep(&iodummy, 0, ident, 0);
217 * Chain in case several placemarkers are present.
219 TAILQ_REMOVE(&hmp->iorun_list, &iodummy, iorun_entry);
220 io = TAILQ_FIRST(&hmp->iorun_list);
221 if (io && io->type == HAMMER_STRUCTURE_DUMMY)
223 lwkt_reltoken(&hmp->io_token);
226 hammer_io_flush_sync(hmp);
230 * Clear a flagged error condition on a I/O buffer. The caller must hold
231 * its own ref on the buffer.
234 hammer_io_clear_error(struct hammer_io *io)
236 hammer_mount_t hmp = io->hmp;
238 lwkt_gettoken(&hmp->io_token);
241 hammer_rel(&io->lock);
242 KKASSERT(hammer_isactive(&io->lock));
244 lwkt_reltoken(&hmp->io_token);
248 hammer_io_clear_error_noassert(struct hammer_io *io)
250 hammer_mount_t hmp = io->hmp;
252 lwkt_gettoken(&hmp->io_token);
255 hammer_rel(&io->lock);
257 lwkt_reltoken(&hmp->io_token);
261 * This is an advisory function only which tells the buffer cache
262 * the bp is not a meta-data buffer, even though it is backed by
265 * This is used by HAMMER's reblocking code to avoid trying to
266 * swapcache the filesystem's data when it is read or written
267 * by the reblocking code.
269 * The caller has a ref on the buffer preventing the bp from
270 * being disassociated from it.
273 hammer_io_notmeta(hammer_buffer_t buffer)
275 if ((buffer->io.bp->b_flags & B_NOTMETA) == 0) {
276 hammer_mount_t hmp = buffer->io.hmp;
278 lwkt_gettoken(&hmp->io_token);
279 buffer->io.bp->b_flags |= B_NOTMETA;
280 lwkt_reltoken(&hmp->io_token);
285 * Load bp for a HAMMER structure. The io must be exclusively locked by
288 * This routine is mostly used on meta-data and small-data blocks. Generally
289 * speaking HAMMER assumes some locality of reference and will cluster.
291 * Note that the caller (hammer_ondisk.c) may place further restrictions
292 * on clusterability via the limit (in bytes). Typically large-data
293 * zones cannot be clustered due to their mixed buffer sizes. This is
294 * not an issue since such clustering occurs in hammer_vnops at the
295 * regular file layer, whereas this is the buffered block device layer.
297 * No I/O callbacks can occur while we hold the buffer locked.
300 hammer_io_read(struct vnode *devvp, struct hammer_io *io, int limit)
305 if ((bp = io->bp) == NULL) {
306 atomic_add_int(&hammer_count_io_running_read, io->bytes);
307 if (hammer_cluster_enable && limit > io->bytes) {
308 error = cluster_read(devvp, io->offset + limit,
309 io->offset, io->bytes,
314 error = bread(devvp, io->offset, io->bytes, &io->bp);
316 hammer_stats_disk_read += io->bytes;
317 atomic_add_int(&hammer_count_io_running_read, -io->bytes);
320 * The code generally assumes b_ops/b_dep has been set-up,
321 * even if we error out here.
324 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IODEBUG)) {
325 const char *metatype;
328 case HAMMER_STRUCTURE_VOLUME:
331 case HAMMER_STRUCTURE_META_BUFFER:
332 switch(((struct hammer_buffer *)io)->
333 zoneX_offset & HAMMER_OFF_ZONE_MASK) {
334 case HAMMER_ZONE_BTREE:
337 case HAMMER_ZONE_META:
340 case HAMMER_ZONE_FREEMAP:
341 metatype = "freemap";
348 case HAMMER_STRUCTURE_DATA_BUFFER:
351 case HAMMER_STRUCTURE_UNDO_BUFFER:
355 metatype = "unknown";
358 kprintf("doff %016jx %s\n",
359 (intmax_t)bp->b_bio2.bio_offset,
362 bp->b_flags &= ~B_IODEBUG;
363 bp->b_ops = &hammer_bioops;
364 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
366 /* io->worklist is locked by the io lock */
367 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
369 KKASSERT(io->modified == 0);
370 KKASSERT(io->running == 0);
371 KKASSERT(io->waiting == 0);
372 io->released = 0; /* we hold an active lock on bp */
380 * Similar to hammer_io_read() but returns a zero'd out buffer instead.
381 * Must be called with the IO exclusively locked.
383 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background
384 * I/O by forcing the buffer to not be in a released state before calling
387 * This function will also mark the IO as modified but it will not
388 * increment the modify_refs count.
390 * No I/O callbacks can occur while we hold the buffer locked.
393 hammer_io_new(struct vnode *devvp, struct hammer_io *io)
397 if ((bp = io->bp) == NULL) {
398 io->bp = getblk(devvp, io->offset, io->bytes, 0, 0);
400 bp->b_ops = &hammer_bioops;
401 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
403 /* io->worklist is locked by the io lock */
404 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
406 KKASSERT(io->running == 0);
416 hammer_io_modify(io, 0);
422 * Advance the activity count on the underlying buffer because
423 * HAMMER does not getblk/brelse on every access.
425 * The io->bp cannot go away while the buffer is referenced.
428 hammer_io_advance(struct hammer_io *io)
431 buf_act_advance(io->bp);
435 * Remove potential device level aliases against buffers managed by high level
436 * vnodes. Aliases can also be created due to mixed buffer sizes or via
437 * direct access to the backing store device.
439 * This is nasty because the buffers are also VMIO-backed. Even if a buffer
440 * does not exist its backing VM pages might, and we have to invalidate
441 * those as well or a getblk() will reinstate them.
443 * Buffer cache buffers associated with hammer_buffers cannot be
447 hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset)
449 hammer_io_structure_t iou;
451 hammer_off_t phys_offset;
455 hmp = volume->io.hmp;
456 lwkt_gettoken(&hmp->io_token);
459 * If a device buffer already exists for the specified physical
460 * offset use that, otherwise instantiate a buffer to cover any
461 * related VM pages, set BNOCACHE, and brelse().
463 phys_offset = volume->ondisk->vol_buf_beg +
464 (zone2_offset & HAMMER_OFF_SHORT_MASK);
465 if ((bp = findblk(volume->devvp, phys_offset, 0)) != NULL)
468 bp = getblk(volume->devvp, phys_offset, HAMMER_BUFSIZE, 0, 0);
470 if ((iou = (void *)LIST_FIRST(&bp->b_dep)) != NULL) {
472 hammer_ref(&iou->io.lock);
473 hammer_io_clear_modify(&iou->io, 1);
475 iou->io.released = 0;
478 iou->io.waitdep = 1; /* XXX this is a fs_token field */
479 KKASSERT(hammer_isactive(&iou->io.lock) == 1);
480 hammer_rel_buffer(&iou->buffer, 0);
481 /*hammer_io_deallocate(bp);*/
486 KKASSERT((bp->b_flags & B_LOCKED) == 0);
488 bp->b_flags |= B_NOCACHE|B_RELBUF;
492 lwkt_reltoken(&hmp->io_token);
497 * This routine is called on the last reference to a hammer structure.
498 * The io must be interlocked with a refcount of zero. The hammer structure
499 * will remain interlocked on return.
501 * This routine may return a non-NULL bp to the caller for dispoal.
502 * The caller typically brelse()'s the bp.
504 * The bp may or may not still be passively associated with the IO. It
505 * will remain passively associated if it is unreleasable (e.g. a modified
508 * The only requirement here is that modified meta-data and volume-header
509 * buffer may NOT be disassociated from the IO structure, and consequently
510 * we also leave such buffers actively associated with the IO if they already
511 * are (since the kernel can't do anything with them anyway). Only the
512 * flusher is allowed to write such buffers out. Modified pure-data and
513 * undo buffers are returned to the kernel but left passively associated
514 * so we can track when the kernel writes the bp out.
517 hammer_io_release(struct hammer_io *io, int flush)
519 union hammer_io_structure *iou = (void *)io;
522 if ((bp = io->bp) == NULL)
526 * Try to flush a dirty IO to disk if asked to by the
527 * caller or if the kernel tried to flush the buffer in the past.
529 * Kernel-initiated flushes are only allowed for pure-data buffers.
530 * meta-data and volume buffers can only be flushed explicitly
535 hammer_io_flush(io, 0);
536 } else if (bp->b_flags & B_LOCKED) {
538 case HAMMER_STRUCTURE_DATA_BUFFER:
539 hammer_io_flush(io, 0);
541 case HAMMER_STRUCTURE_UNDO_BUFFER:
542 hammer_io_flush(io, hammer_undo_reclaim(io));
547 } /* else no explicit request to flush the buffer */
551 * Wait for the IO to complete if asked to. This occurs when
552 * the buffer must be disposed of definitively during an umount
553 * or buffer invalidation.
555 if (io->waitdep && io->running) {
560 * Return control of the buffer to the kernel (with the provisio
561 * that our bioops can override kernel decisions with regards to
564 if ((flush || io->reclaim) && io->modified == 0 && io->running == 0) {
566 * Always disassociate the bp if an explicit flush
567 * was requested and the IO completed with no error
568 * (so unmount can really clean up the structure).
576 hammer_io_disassociate((hammer_io_structure_t)io);
578 } else if (io->modified) {
580 * Only certain IO types can be released to the kernel if
581 * the buffer has been modified.
583 * volume and meta-data IO types may only be explicitly
587 case HAMMER_STRUCTURE_DATA_BUFFER:
588 case HAMMER_STRUCTURE_UNDO_BUFFER:
589 if (io->released == 0) {
597 bp = NULL; /* bp left associated */
598 } else if (io->released == 0) {
600 * Clean buffers can be generally released to the kernel.
601 * We leave the bp passively associated with the HAMMER
602 * structure and use bioops to disconnect it later on
603 * if the kernel wants to discard the buffer.
605 * We can steal the structure's ownership of the bp.
608 if (bp->b_flags & B_LOCKED) {
609 hammer_io_disassociate(iou);
613 hammer_io_disassociate(iou);
616 /* return the bp (bp passively associated) */
621 * A released buffer is passively associate with our
622 * hammer_io structure. The kernel cannot destroy it
623 * without making a bioops call. If the kernel (B_LOCKED)
624 * or we (reclaim) requested that the buffer be destroyed
625 * we destroy it, otherwise we do a quick get/release to
626 * reset its position in the kernel's LRU list.
628 * Leaving the buffer passively associated allows us to
629 * use the kernel's LRU buffer flushing mechanisms rather
630 * then rolling our own.
632 * XXX there are two ways of doing this. We can re-acquire
633 * and passively release to reset the LRU, or not.
635 if (io->running == 0) {
637 if ((bp->b_flags & B_LOCKED) || io->reclaim) {
638 hammer_io_disassociate(iou);
641 /* return the bp (bp passively associated) */
645 * bp is left passively associated but we do not
646 * try to reacquire it. Interactions with the io
647 * structure will occur on completion of the bp's
657 * This routine is called with a locked IO when a flush is desired and
658 * no other references to the structure exists other then ours. This
659 * routine is ONLY called when HAMMER believes it is safe to flush a
660 * potentially modified buffer out.
662 * The locked io or io reference prevents a flush from being initiated
666 hammer_io_flush(struct hammer_io *io, int reclaim)
672 * Degenerate case - nothing to flush if nothing is dirty.
674 if (io->modified == 0)
678 KKASSERT(io->modify_refs <= 0);
681 * Acquire ownership of the bp, particularly before we clear our
684 * We are going to bawrite() this bp. Don't leave a window where
685 * io->released is set, we actually own the bp rather then our
688 * The io_token should not be required here as only
694 /* BUF_KERNPROC(io->bp); */
695 /* io->released = 0; */
696 KKASSERT(io->released);
697 KKASSERT(io->bp == bp);
704 if ((bp->b_flags & B_LOCKED) == 0) {
705 bp->b_flags |= B_LOCKED;
706 atomic_add_int(&hammer_count_io_locked, 1);
711 * Acquire exclusive access to the bp and then clear the modified
712 * state of the buffer prior to issuing I/O to interlock any
713 * modifications made while the I/O is in progress. This shouldn't
714 * happen anyway but losing data would be worse. The modified bit
715 * will be rechecked after the IO completes.
717 * NOTE: This call also finalizes the buffer's content (inval == 0).
719 * This is only legal when lock.refs == 1 (otherwise we might clear
720 * the modified bit while there are still users of the cluster
721 * modifying the data).
723 * Do this before potentially blocking so any attempt to modify the
724 * ondisk while we are blocked blocks waiting for us.
726 hammer_ref(&io->lock);
727 hammer_io_clear_modify(io, 0);
728 hammer_rel(&io->lock);
730 if (hammer_debug_io & 0x0002)
731 kprintf("hammer io_write %016jx\n", bp->b_bio1.bio_offset);
734 * Transfer ownership to the kernel and initiate I/O.
736 * NOTE: We do not hold io_token so an atomic op is required to
737 * update io_running_space.
740 atomic_add_int(&hmp->io_running_space, io->bytes);
741 atomic_add_int(&hammer_count_io_running_write, io->bytes);
742 lwkt_gettoken(&hmp->io_token);
743 TAILQ_INSERT_TAIL(&hmp->iorun_list, io, iorun_entry);
744 lwkt_reltoken(&hmp->io_token);
746 hammer_io_flush_mark(io->volume);
749 /************************************************************************
751 ************************************************************************
753 * These routines deal with dependancies created when IO buffers get
754 * modified. The caller must call hammer_modify_*() on a referenced
755 * HAMMER structure prior to modifying its on-disk data.
757 * Any intent to modify an IO buffer acquires the related bp and imposes
758 * various write ordering dependancies.
762 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
763 * are locked until the flusher can deal with them, pure data buffers
764 * can be written out.
766 * The referenced io prevents races.
770 hammer_io_modify(hammer_io_t io, int count)
773 * io->modify_refs must be >= 0
775 while (io->modify_refs < 0) {
777 tsleep(io, 0, "hmrmod", 0);
781 * Shortcut if nothing to do.
783 KKASSERT(hammer_isactive(&io->lock) && io->bp != NULL);
784 io->modify_refs += count;
785 if (io->modified && io->released == 0)
789 * NOTE: It is important not to set the modified bit
790 * until after we have acquired the bp or we risk
791 * racing against checkwrite.
793 hammer_lock_ex(&io->lock);
796 BUF_KERNPROC(io->bp);
799 if (io->modified == 0) {
800 hammer_io_set_modlist(io);
803 hammer_unlock(&io->lock);
808 hammer_io_modify_done(hammer_io_t io)
810 KKASSERT(io->modify_refs > 0);
812 if (io->modify_refs == 0 && io->waitmod) {
819 * The write interlock blocks other threads trying to modify a buffer
820 * (they block in hammer_io_modify()) after us, or blocks us while other
821 * threads are in the middle of modifying a buffer.
823 * The caller also has a ref on the io, however if we are not careful
824 * we will race bioops callbacks (checkwrite). To deal with this
825 * we must at least acquire and release the io_token, and it is probably
826 * better to hold it through the setting of modify_refs.
829 hammer_io_write_interlock(hammer_io_t io)
831 hammer_mount_t hmp = io->hmp;
833 lwkt_gettoken(&hmp->io_token);
834 while (io->modify_refs != 0) {
836 tsleep(io, 0, "hmrmod", 0);
838 io->modify_refs = -1;
839 lwkt_reltoken(&hmp->io_token);
843 hammer_io_done_interlock(hammer_io_t io)
845 KKASSERT(io->modify_refs == -1);
854 * Caller intends to modify a volume's ondisk structure.
856 * This is only allowed if we are the flusher or we have a ref on the
860 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
863 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
865 hammer_io_modify(&volume->io, 1);
867 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk;
868 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
869 hammer_generate_undo(trans,
870 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset),
876 * Caller intends to modify a buffer's ondisk structure.
878 * This is only allowed if we are the flusher or we have a ref on the
882 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
885 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
887 hammer_io_modify(&buffer->io, 1);
889 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk;
890 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
891 hammer_generate_undo(trans,
892 buffer->zone2_offset + rel_offset,
898 hammer_modify_volume_done(hammer_volume_t volume)
900 hammer_io_modify_done(&volume->io);
904 hammer_modify_buffer_done(hammer_buffer_t buffer)
906 hammer_io_modify_done(&buffer->io);
910 * Mark an entity as not being dirty any more and finalize any
911 * delayed adjustments to the buffer.
913 * Delayed adjustments are an important performance enhancement, allowing
914 * us to avoid recalculating B-Tree node CRCs over and over again when
915 * making bulk-modifications to the B-Tree.
917 * If inval is non-zero delayed adjustments are ignored.
919 * This routine may dereference related btree nodes and cause the
920 * buffer to be dereferenced. The caller must own a reference on io.
923 hammer_io_clear_modify(struct hammer_io *io, int inval)
928 * io_token is needed to avoid races on mod_root
930 if (io->modified == 0)
933 lwkt_gettoken(&hmp->io_token);
934 if (io->modified == 0) {
935 lwkt_reltoken(&hmp->io_token);
940 * Take us off the mod-list and clear the modified bit.
942 KKASSERT(io->mod_root != NULL);
943 if (io->mod_root == &io->hmp->volu_root ||
944 io->mod_root == &io->hmp->meta_root) {
945 io->hmp->locked_dirty_space -= io->bytes;
946 atomic_add_int(&hammer_count_dirtybufspace, -io->bytes);
948 RB_REMOVE(hammer_mod_rb_tree, io->mod_root, io);
952 lwkt_reltoken(&hmp->io_token);
955 * If this bit is not set there are no delayed adjustments.
962 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference
963 * on the node (& underlying buffer). Release the node after clearing
966 if (io->type == HAMMER_STRUCTURE_META_BUFFER) {
967 hammer_buffer_t buffer = (void *)io;
971 TAILQ_FOREACH(node, &buffer->clist, entry) {
972 if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0)
974 node->flags &= ~HAMMER_NODE_NEEDSCRC;
975 KKASSERT(node->ondisk);
977 node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE);
978 hammer_rel_node(node);
982 /* caller must still have ref on io */
983 KKASSERT(hammer_isactive(&io->lock));
987 * Clear the IO's modify list. Even though the IO is no longer modified
988 * it may still be on the lose_root. This routine is called just before
989 * the governing hammer_buffer is destroyed.
991 * mod_root requires io_token protection.
994 hammer_io_clear_modlist(struct hammer_io *io)
996 hammer_mount_t hmp = io->hmp;
998 KKASSERT(io->modified == 0);
1000 lwkt_gettoken(&hmp->io_token);
1002 KKASSERT(io->mod_root == &io->hmp->lose_root);
1003 RB_REMOVE(hammer_mod_rb_tree, io->mod_root, io);
1004 io->mod_root = NULL;
1006 lwkt_reltoken(&hmp->io_token);
1011 hammer_io_set_modlist(struct hammer_io *io)
1013 struct hammer_mount *hmp = io->hmp;
1015 lwkt_gettoken(&hmp->io_token);
1016 KKASSERT(io->mod_root == NULL);
1019 case HAMMER_STRUCTURE_VOLUME:
1020 io->mod_root = &hmp->volu_root;
1021 hmp->locked_dirty_space += io->bytes;
1022 atomic_add_int(&hammer_count_dirtybufspace, io->bytes);
1024 case HAMMER_STRUCTURE_META_BUFFER:
1025 io->mod_root = &hmp->meta_root;
1026 hmp->locked_dirty_space += io->bytes;
1027 atomic_add_int(&hammer_count_dirtybufspace, io->bytes);
1029 case HAMMER_STRUCTURE_UNDO_BUFFER:
1030 io->mod_root = &hmp->undo_root;
1032 case HAMMER_STRUCTURE_DATA_BUFFER:
1033 io->mod_root = &hmp->data_root;
1035 case HAMMER_STRUCTURE_DUMMY:
1036 panic("hammer_io_set_modlist: bad io type");
1037 break; /* NOT REACHED */
1039 if (RB_INSERT(hammer_mod_rb_tree, io->mod_root, io)) {
1040 panic("hammer_io_set_modlist: duplicate entry");
1043 lwkt_reltoken(&hmp->io_token);
1046 /************************************************************************
1048 ************************************************************************
1053 * Pre-IO initiation kernel callback - cluster build only
1055 * bioops callback - hold io_token
1058 hammer_io_start(struct buf *bp)
1060 /* nothing to do, so io_token not needed */
1064 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT!
1066 * NOTE: HAMMER may modify a data buffer after we have initiated write
1069 * NOTE: MPSAFE callback
1071 * bioops callback - hold io_token
1074 hammer_io_complete(struct buf *bp)
1076 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep);
1077 struct hammer_mount *hmp = iou->io.hmp;
1078 struct hammer_io *ionext;
1080 lwkt_gettoken(&hmp->io_token);
1082 KKASSERT(iou->io.released == 1);
1085 * Deal with people waiting for I/O to drain
1087 if (iou->io.running) {
1089 * Deal with critical write errors. Once a critical error
1090 * has been flagged in hmp the UNDO FIFO will not be updated.
1091 * That way crash recover will give us a consistent
1094 * Because of this we can throw away failed UNDO buffers. If
1095 * we throw away META or DATA buffers we risk corrupting
1096 * the now read-only version of the filesystem visible to
1097 * the user. Clear B_ERROR so the buffer is not re-dirtied
1098 * by the kernel and ref the io so it doesn't get thrown
1101 if (bp->b_flags & B_ERROR) {
1102 lwkt_gettoken(&hmp->fs_token);
1103 hammer_critical_error(hmp, NULL, bp->b_error,
1104 "while flushing meta-data");
1105 lwkt_reltoken(&hmp->fs_token);
1107 switch(iou->io.type) {
1108 case HAMMER_STRUCTURE_UNDO_BUFFER:
1111 if (iou->io.ioerror == 0) {
1112 iou->io.ioerror = 1;
1113 hammer_ref(&iou->io.lock);
1117 bp->b_flags &= ~B_ERROR;
1120 hammer_io_set_modlist(&iou->io);
1121 iou->io.modified = 1;
1124 hammer_stats_disk_write += iou->io.bytes;
1125 atomic_add_int(&hammer_count_io_running_write, -iou->io.bytes);
1126 atomic_add_int(&hmp->io_running_space, -iou->io.bytes);
1127 if (hmp->io_running_wakeup &&
1128 hmp->io_running_space < hammer_limit_running_io / 2) {
1129 hmp->io_running_wakeup = 0;
1130 wakeup(&hmp->io_running_wakeup);
1132 KKASSERT(hmp->io_running_space >= 0);
1133 iou->io.running = 0;
1136 * Remove from iorun list and wakeup any multi-io waiter(s).
1138 if (TAILQ_FIRST(&hmp->iorun_list) == &iou->io) {
1139 ionext = TAILQ_NEXT(&iou->io, iorun_entry);
1140 if (ionext && ionext->type == HAMMER_STRUCTURE_DUMMY)
1143 TAILQ_REMOVE(&hmp->iorun_list, &iou->io, iorun_entry);
1145 hammer_stats_disk_read += iou->io.bytes;
1148 if (iou->io.waiting) {
1149 iou->io.waiting = 0;
1154 * If B_LOCKED is set someone wanted to deallocate the bp at some
1155 * point, try to do it now. The operation will fail if there are
1156 * refs or if hammer_io_deallocate() is unable to gain the
1159 if (bp->b_flags & B_LOCKED) {
1160 atomic_add_int(&hammer_count_io_locked, -1);
1161 bp->b_flags &= ~B_LOCKED;
1162 hammer_io_deallocate(bp);
1163 /* structure may be dead now */
1165 lwkt_reltoken(&hmp->io_token);
1169 * Callback from kernel when it wishes to deallocate a passively
1170 * associated structure. This mostly occurs with clean buffers
1171 * but it may be possible for a holding structure to be marked dirty
1172 * while its buffer is passively associated. The caller owns the bp.
1174 * If we cannot disassociate we set B_LOCKED to prevent the buffer
1175 * from getting reused.
1177 * WARNING: Because this can be called directly by getnewbuf we cannot
1178 * recurse into the tree. If a bp cannot be immediately disassociated
1179 * our only recourse is to set B_LOCKED.
1181 * WARNING: This may be called from an interrupt via hammer_io_complete()
1183 * bioops callback - hold io_token
1186 hammer_io_deallocate(struct buf *bp)
1188 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep);
1193 lwkt_gettoken(&hmp->io_token);
1195 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0);
1196 if (hammer_try_interlock_norefs(&iou->io.lock) == 0) {
1198 * We cannot safely disassociate a bp from a referenced
1199 * or interlocked HAMMER structure.
1201 bp->b_flags |= B_LOCKED;
1202 atomic_add_int(&hammer_count_io_locked, 1);
1203 } else if (iou->io.modified) {
1205 * It is not legal to disassociate a modified buffer. This
1206 * case really shouldn't ever occur.
1208 bp->b_flags |= B_LOCKED;
1209 atomic_add_int(&hammer_count_io_locked, 1);
1210 hammer_put_interlock(&iou->io.lock, 0);
1213 * Disassociate the BP. If the io has no refs left we
1214 * have to add it to the loose list. The kernel has
1215 * locked the buffer and therefore our io must be
1216 * in a released state.
1218 hammer_io_disassociate(iou);
1219 if (iou->io.type != HAMMER_STRUCTURE_VOLUME) {
1220 KKASSERT(iou->io.bp == NULL);
1221 KKASSERT(iou->io.mod_root == NULL);
1222 iou->io.mod_root = &hmp->lose_root;
1223 if (RB_INSERT(hammer_mod_rb_tree, iou->io.mod_root,
1225 panic("hammer_io_deallocate: duplicate entry");
1228 hammer_put_interlock(&iou->io.lock, 1);
1230 lwkt_reltoken(&hmp->io_token);
1234 * bioops callback - hold io_token
1237 hammer_io_fsync(struct vnode *vp)
1239 /* nothing to do, so io_token not needed */
1244 * NOTE: will not be called unless we tell the kernel about the
1245 * bioops. Unused... we use the mount's VFS_SYNC instead.
1247 * bioops callback - hold io_token
1250 hammer_io_sync(struct mount *mp)
1252 /* nothing to do, so io_token not needed */
1257 * bioops callback - hold io_token
1260 hammer_io_movedeps(struct buf *bp1, struct buf *bp2)
1262 /* nothing to do, so io_token not needed */
1266 * I/O pre-check for reading and writing. HAMMER only uses this for
1267 * B_CACHE buffers so checkread just shouldn't happen, but if it does
1270 * Writing is a different case. We don't want the kernel to try to write
1271 * out a buffer that HAMMER may be modifying passively or which has a
1272 * dependancy. In addition, kernel-demanded writes can only proceed for
1273 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
1274 * buffer types can only be explicitly written by the flusher.
1276 * checkwrite will only be called for bdwrite()n buffers. If we return
1277 * success the kernel is guaranteed to initiate the buffer write.
1279 * bioops callback - hold io_token
1282 hammer_io_checkread(struct buf *bp)
1284 /* nothing to do, so io_token not needed */
1289 * The kernel is asking us whether it can write out a dirty buffer or not.
1291 * bioops callback - hold io_token
1294 hammer_io_checkwrite(struct buf *bp)
1296 hammer_io_t io = (void *)LIST_FIRST(&bp->b_dep);
1297 hammer_mount_t hmp = io->hmp;
1300 * This shouldn't happen under normal operation.
1302 lwkt_gettoken(&hmp->io_token);
1303 if (io->type == HAMMER_STRUCTURE_VOLUME ||
1304 io->type == HAMMER_STRUCTURE_META_BUFFER) {
1306 panic("hammer_io_checkwrite: illegal buffer");
1307 if ((bp->b_flags & B_LOCKED) == 0) {
1308 bp->b_flags |= B_LOCKED;
1309 atomic_add_int(&hammer_count_io_locked, 1);
1311 lwkt_reltoken(&hmp->io_token);
1316 * We have to be able to interlock the IO to safely modify any
1317 * of its fields without holding the fs_token. If we can't lock
1318 * it then we are racing someone.
1320 * Our ownership of the bp lock prevents the io from being ripped
1321 * out from under us.
1323 if (hammer_try_interlock_norefs(&io->lock) == 0) {
1324 bp->b_flags |= B_LOCKED;
1325 atomic_add_int(&hammer_count_io_locked, 1);
1326 lwkt_reltoken(&hmp->io_token);
1331 * The modified bit must be cleared prior to the initiation of
1332 * any IO (returning 0 initiates the IO). Because this is a
1333 * normal data buffer hammer_io_clear_modify() runs through a
1334 * simple degenerate case.
1336 * Return 0 will cause the kernel to initiate the IO, and we
1337 * must normally clear the modified bit before we begin. If
1338 * the io has modify_refs we do not clear the modified bit,
1339 * otherwise we may miss changes.
1341 * Only data and undo buffers can reach here. These buffers do
1342 * not have terminal crc functions but we temporarily reference
1343 * the IO anyway, just in case.
1345 if (io->modify_refs == 0 && io->modified) {
1346 hammer_ref(&io->lock);
1347 hammer_io_clear_modify(io, 0);
1348 hammer_rel(&io->lock);
1349 } else if (io->modified) {
1350 KKASSERT(io->type == HAMMER_STRUCTURE_DATA_BUFFER);
1354 * The kernel is going to start the IO, set io->running.
1356 KKASSERT(io->running == 0);
1358 atomic_add_int(&io->hmp->io_running_space, io->bytes);
1359 atomic_add_int(&hammer_count_io_running_write, io->bytes);
1360 TAILQ_INSERT_TAIL(&io->hmp->iorun_list, io, iorun_entry);
1362 hammer_put_interlock(&io->lock, 1);
1363 lwkt_reltoken(&hmp->io_token);
1369 * Return non-zero if we wish to delay the kernel's attempt to flush
1370 * this buffer to disk.
1372 * bioops callback - hold io_token
1375 hammer_io_countdeps(struct buf *bp, int n)
1377 /* nothing to do, so io_token not needed */
1381 struct bio_ops hammer_bioops = {
1382 .io_start = hammer_io_start,
1383 .io_complete = hammer_io_complete,
1384 .io_deallocate = hammer_io_deallocate,
1385 .io_fsync = hammer_io_fsync,
1386 .io_sync = hammer_io_sync,
1387 .io_movedeps = hammer_io_movedeps,
1388 .io_countdeps = hammer_io_countdeps,
1389 .io_checkread = hammer_io_checkread,
1390 .io_checkwrite = hammer_io_checkwrite,
1393 /************************************************************************
1395 ************************************************************************
1397 * These functions operate directly on the buffer cache buffer associated
1398 * with a front-end vnode rather then a back-end device vnode.
1402 * Read a buffer associated with a front-end vnode directly from the
1403 * disk media. The bio may be issued asynchronously. If leaf is non-NULL
1404 * we validate the CRC.
1406 * We must check for the presence of a HAMMER buffer to handle the case
1407 * where the reblocker has rewritten the data (which it does via the HAMMER
1408 * buffer system, not via the high-level vnode buffer cache), but not yet
1409 * committed the buffer to the media.
1412 hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1413 hammer_btree_leaf_elm_t leaf)
1415 hammer_off_t buf_offset;
1416 hammer_off_t zone2_offset;
1417 hammer_volume_t volume;
1423 buf_offset = bio->bio_offset;
1424 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) ==
1425 HAMMER_ZONE_LARGE_DATA);
1428 * The buffer cache may have an aliased buffer (the reblocker can
1429 * write them). If it does we have to sync any dirty data before
1430 * we can build our direct-read. This is a non-critical code path.
1433 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize);
1436 * Resolve to a zone-2 offset. The conversion just requires
1437 * munging the top 4 bits but we want to abstract it anyway
1438 * so the blockmap code can verify the zone assignment.
1440 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
1443 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
1444 HAMMER_ZONE_RAW_BUFFER);
1447 * Resolve volume and raw-offset for 3rd level bio. The
1448 * offset will be specific to the volume.
1450 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1451 volume = hammer_get_volume(hmp, vol_no, &error);
1452 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1459 nbio = push_bio(bio);
1460 nbio->bio_offset = volume->ondisk->vol_buf_beg +
1461 (zone2_offset & HAMMER_OFF_SHORT_MASK);
1464 * XXX disabled - our CRC check doesn't work if the OS
1465 * does bogus_page replacement on the direct-read.
1467 if (leaf && hammer_verify_data) {
1468 nbio->bio_done = hammer_io_direct_read_complete;
1469 nbio->bio_caller_info1.uvalue32 = leaf->data_crc;
1472 hammer_stats_disk_read += bp->b_bufsize;
1473 vn_strategy(volume->devvp, nbio);
1475 hammer_rel_volume(volume, 0);
1478 kprintf("hammer_direct_read: failed @ %016llx\n",
1479 (long long)zone2_offset);
1480 bp->b_error = error;
1481 bp->b_flags |= B_ERROR;
1488 * This works similarly to hammer_io_direct_read() except instead of
1489 * directly reading from the device into the bio we instead indirectly
1490 * read through the device's buffer cache and then copy the data into
1493 * If leaf is non-NULL and validation is enabled, the CRC will be checked.
1495 * This routine also executes asynchronously. It allows hammer strategy
1496 * calls to operate asynchronously when in double_buffer mode (in addition
1497 * to operating asynchronously when in normal mode).
1500 hammer_io_indirect_read(hammer_mount_t hmp, struct bio *bio,
1501 hammer_btree_leaf_elm_t leaf)
1503 hammer_off_t buf_offset;
1504 hammer_off_t zone2_offset;
1505 hammer_volume_t volume;
1510 buf_offset = bio->bio_offset;
1511 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) ==
1512 HAMMER_ZONE_LARGE_DATA);
1515 * The buffer cache may have an aliased buffer (the reblocker can
1516 * write them). If it does we have to sync any dirty data before
1517 * we can build our direct-read. This is a non-critical code path.
1520 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize);
1523 * Resolve to a zone-2 offset. The conversion just requires
1524 * munging the top 4 bits but we want to abstract it anyway
1525 * so the blockmap code can verify the zone assignment.
1527 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
1530 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
1531 HAMMER_ZONE_RAW_BUFFER);
1534 * Resolve volume and raw-offset for 3rd level bio. The
1535 * offset will be specific to the volume.
1537 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1538 volume = hammer_get_volume(hmp, vol_no, &error);
1539 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1544 * Convert to the raw volume->devvp offset and acquire
1545 * the buf, issuing async I/O if necessary.
1547 buf_offset = volume->ondisk->vol_buf_beg +
1548 (zone2_offset & HAMMER_OFF_SHORT_MASK);
1550 if (leaf && hammer_verify_data) {
1551 bio->bio_caller_info1.uvalue32 = leaf->data_crc;
1552 bio->bio_caller_info2.index = 1;
1554 bio->bio_caller_info2.index = 0;
1556 breadcb(volume->devvp, buf_offset, bp->b_bufsize,
1557 hammer_indirect_callback, bio);
1559 hammer_rel_volume(volume, 0);
1562 kprintf("hammer_direct_read: failed @ %016llx\n",
1563 (long long)zone2_offset);
1564 bp->b_error = error;
1565 bp->b_flags |= B_ERROR;
1572 * Indirect callback on completion. bio/bp specify the device-backed
1573 * buffer. bio->bio_caller_info1.ptr holds obio.
1575 * obio/obp is the original regular file buffer. obio->bio_caller_info*
1576 * contains the crc specification.
1578 * We are responsible for calling bpdone() and bqrelse() on bio/bp, and
1579 * for calling biodone() on obio.
1582 hammer_indirect_callback(struct bio *bio)
1584 struct buf *bp = bio->bio_buf;
1589 * If BIO_DONE is already set the device buffer was already
1590 * fully valid (B_CACHE). If it is not set then I/O was issued
1591 * and we have to run I/O completion as the last bio.
1593 * Nobody is waiting for our device I/O to complete, we are
1594 * responsible for bqrelse()ing it which means we also have to do
1595 * the equivalent of biowait() and clear BIO_DONE (which breadcb()
1598 * Any preexisting device buffer should match the requested size,
1599 * but due to bigblock recycling and other factors there is some
1600 * fragility there, so we assert that the device buffer covers
1603 if ((bio->bio_flags & BIO_DONE) == 0)
1605 bio->bio_flags &= ~(BIO_DONE | BIO_SYNC);
1607 obio = bio->bio_caller_info1.ptr;
1608 obp = obio->bio_buf;
1610 if (bp->b_flags & B_ERROR) {
1611 obp->b_flags |= B_ERROR;
1612 obp->b_error = bp->b_error;
1613 } else if (obio->bio_caller_info2.index &&
1614 obio->bio_caller_info1.uvalue32 !=
1615 crc32(bp->b_data, bp->b_bufsize)) {
1616 obp->b_flags |= B_ERROR;
1619 KKASSERT(bp->b_bufsize >= obp->b_bufsize);
1620 bcopy(bp->b_data, obp->b_data, obp->b_bufsize);
1622 obp->b_flags |= B_AGE;
1630 * On completion of the BIO this callback must check the data CRC
1631 * and chain to the previous bio.
1633 * MPSAFE - since we do not modify and hammer_records we do not need
1636 * NOTE: MPSAFE callback
1640 hammer_io_direct_read_complete(struct bio *nbio)
1644 u_int32_t rec_crc = nbio->bio_caller_info1.uvalue32;
1647 if (crc32(bp->b_data, bp->b_bufsize) != rec_crc) {
1648 kprintf("HAMMER: data_crc error @%016llx/%d\n",
1649 nbio->bio_offset, bp->b_bufsize);
1650 if (hammer_debug_critical)
1651 Debugger("data_crc on read");
1652 bp->b_flags |= B_ERROR;
1655 obio = pop_bio(nbio);
1661 * Write a buffer associated with a front-end vnode directly to the
1662 * disk media. The bio may be issued asynchronously.
1664 * The BIO is associated with the specified record and RECG_DIRECT_IO
1665 * is set. The recorded is added to its object.
1668 hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio,
1669 hammer_record_t record)
1671 hammer_btree_leaf_elm_t leaf = &record->leaf;
1672 hammer_off_t buf_offset;
1673 hammer_off_t zone2_offset;
1674 hammer_volume_t volume;
1675 hammer_buffer_t buffer;
1682 buf_offset = leaf->data_offset;
1684 KKASSERT(buf_offset > HAMMER_ZONE_BTREE);
1685 KKASSERT(bio->bio_buf->b_cmd == BUF_CMD_WRITE);
1688 * Issue or execute the I/O. The new memory record must replace
1689 * the old one before the I/O completes, otherwise a reaquisition of
1690 * the buffer will load the old media data instead of the new.
1692 if ((buf_offset & HAMMER_BUFMASK) == 0 &&
1693 leaf->data_len >= HAMMER_BUFSIZE) {
1695 * We are using the vnode's bio to write directly to the
1696 * media, any hammer_buffer at the same zone-X offset will
1697 * now have stale data.
1699 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
1700 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1701 volume = hammer_get_volume(hmp, vol_no, &error);
1703 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1707 KKASSERT((bp->b_bufsize & HAMMER_BUFMASK) == 0);
1709 hammer_del_buffers(hmp, buf_offset,
1710 zone2_offset, bp->b_bufsize);
1714 * Second level bio - cached zone2 offset.
1716 * (We can put our bio_done function in either the
1717 * 2nd or 3rd level).
1719 nbio = push_bio(bio);
1720 nbio->bio_offset = zone2_offset;
1721 nbio->bio_done = hammer_io_direct_write_complete;
1722 nbio->bio_caller_info1.ptr = record;
1723 record->zone2_offset = zone2_offset;
1724 record->gflags |= HAMMER_RECG_DIRECT_IO |
1725 HAMMER_RECG_DIRECT_INVAL;
1728 * Third level bio - raw offset specific to the
1731 zone2_offset &= HAMMER_OFF_SHORT_MASK;
1732 nbio = push_bio(nbio);
1733 nbio->bio_offset = volume->ondisk->vol_buf_beg +
1735 hammer_stats_disk_write += bp->b_bufsize;
1736 hammer_ip_replace_bulk(hmp, record);
1737 vn_strategy(volume->devvp, nbio);
1738 hammer_io_flush_mark(volume);
1740 hammer_rel_volume(volume, 0);
1743 * Must fit in a standard HAMMER buffer. In this case all
1744 * consumers use the HAMMER buffer system and RECG_DIRECT_IO
1745 * does not need to be set-up.
1747 KKASSERT(((buf_offset ^ (buf_offset + leaf->data_len - 1)) & ~HAMMER_BUFMASK64) == 0);
1749 ptr = hammer_bread(hmp, buf_offset, &error, &buffer);
1752 bp->b_flags |= B_AGE;
1753 hammer_io_modify(&buffer->io, 1);
1754 bcopy(bp->b_data, ptr, leaf->data_len);
1755 hammer_io_modify_done(&buffer->io);
1756 hammer_rel_buffer(buffer, 0);
1758 hammer_ip_replace_bulk(hmp, record);
1764 * Major suckage occured. Also note: The record was
1765 * never added to the tree so we do not have to worry
1766 * about the backend.
1768 kprintf("hammer_direct_write: failed @ %016llx\n",
1769 (long long)leaf->data_offset);
1773 bp->b_flags |= B_ERROR;
1775 record->flags |= HAMMER_RECF_DELETED_FE;
1776 hammer_rel_mem_record(record);
1782 * On completion of the BIO this callback must disconnect
1783 * it from the hammer_record and chain to the previous bio.
1785 * An I/O error forces the mount to read-only. Data buffers
1786 * are not B_LOCKED like meta-data buffers are, so we have to
1787 * throw the buffer away to prevent the kernel from retrying.
1789 * NOTE: MPSAFE callback, only modify fields we have explicit
1790 * access to (the bp and the record->gflags).
1794 hammer_io_direct_write_complete(struct bio *nbio)
1798 hammer_record_t record;
1801 record = nbio->bio_caller_info1.ptr;
1802 KKASSERT(record != NULL);
1803 hmp = record->ip->hmp;
1805 lwkt_gettoken(&hmp->io_token);
1808 obio = pop_bio(nbio);
1809 if (bp->b_flags & B_ERROR) {
1810 lwkt_gettoken(&hmp->fs_token);
1811 hammer_critical_error(hmp, record->ip,
1813 "while writing bulk data");
1814 lwkt_reltoken(&hmp->fs_token);
1815 bp->b_flags |= B_INVAL;
1819 KKASSERT(record->gflags & HAMMER_RECG_DIRECT_IO);
1820 if (record->gflags & HAMMER_RECG_DIRECT_WAIT) {
1821 record->gflags &= ~(HAMMER_RECG_DIRECT_IO |
1822 HAMMER_RECG_DIRECT_WAIT);
1823 /* record can disappear once DIRECT_IO flag is cleared */
1824 wakeup(&record->flags);
1826 record->gflags &= ~HAMMER_RECG_DIRECT_IO;
1827 /* record can disappear once DIRECT_IO flag is cleared */
1829 lwkt_reltoken(&hmp->io_token);
1834 * This is called before a record is either committed to the B-Tree
1835 * or destroyed, to resolve any associated direct-IO.
1837 * (1) We must wait for any direct-IO related to the record to complete.
1839 * (2) We must remove any buffer cache aliases for data accessed via
1840 * leaf->data_offset or zone2_offset so non-direct-IO consumers
1841 * (the mirroring and reblocking code) do not see stale data.
1844 hammer_io_direct_wait(hammer_record_t record)
1846 hammer_mount_t hmp = record->ip->hmp;
1849 * Wait for I/O to complete
1851 if (record->gflags & HAMMER_RECG_DIRECT_IO) {
1852 lwkt_gettoken(&hmp->io_token);
1853 while (record->gflags & HAMMER_RECG_DIRECT_IO) {
1854 record->gflags |= HAMMER_RECG_DIRECT_WAIT;
1855 tsleep(&record->flags, 0, "hmdiow", 0);
1857 lwkt_reltoken(&hmp->io_token);
1861 * Invalidate any related buffer cache aliases associated with the
1862 * backing device. This is needed because the buffer cache buffer
1863 * for file data is associated with the file vnode, not the backing
1866 * XXX I do not think this case can occur any more now that
1867 * reservations ensure that all such buffers are removed before
1868 * an area can be reused.
1870 if (record->gflags & HAMMER_RECG_DIRECT_INVAL) {
1871 KKASSERT(record->leaf.data_offset);
1872 hammer_del_buffers(hmp, record->leaf.data_offset,
1873 record->zone2_offset, record->leaf.data_len,
1875 record->gflags &= ~HAMMER_RECG_DIRECT_INVAL;
1880 * This is called to remove the second-level cached zone-2 offset from
1881 * frontend buffer cache buffers, now stale due to a data relocation.
1882 * These offsets are generated by cluster_read() via VOP_BMAP, or directly
1883 * by hammer_vop_strategy_read().
1885 * This is rather nasty because here we have something like the reblocker
1886 * scanning the raw B-Tree with no held references on anything, really,
1887 * other then a shared lock on the B-Tree node, and we have to access the
1888 * frontend's buffer cache to check for and clean out the association.
1889 * Specifically, if the reblocker is moving data on the disk, these cached
1890 * offsets will become invalid.
1892 * Only data record types associated with the large-data zone are subject
1893 * to direct-io and need to be checked.
1897 hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf)
1899 struct hammer_inode_info iinfo;
1902 if (leaf->base.rec_type != HAMMER_RECTYPE_DATA)
1904 zone = HAMMER_ZONE_DECODE(leaf->data_offset);
1905 if (zone != HAMMER_ZONE_LARGE_DATA_INDEX)
1907 iinfo.obj_id = leaf->base.obj_id;
1908 iinfo.obj_asof = 0; /* unused */
1909 iinfo.obj_localization = leaf->base.localization &
1910 HAMMER_LOCALIZE_PSEUDOFS_MASK;
1911 iinfo.u.leaf = leaf;
1912 hammer_scan_inode_snapshots(hmp, &iinfo,
1913 hammer_io_direct_uncache_callback,
1918 hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data)
1920 hammer_inode_info_t iinfo = data;
1921 hammer_off_t data_offset;
1922 hammer_off_t file_offset;
1929 data_offset = iinfo->u.leaf->data_offset;
1930 file_offset = iinfo->u.leaf->base.key - iinfo->u.leaf->data_len;
1931 blksize = iinfo->u.leaf->data_len;
1932 KKASSERT((blksize & HAMMER_BUFMASK) == 0);
1935 * Warning: FINDBLK_TEST return stable storage but not stable
1936 * contents. It happens to be ok in this case.
1938 hammer_ref(&ip->lock);
1939 if (hammer_get_vnode(ip, &vp) == 0) {
1940 if ((bp = findblk(ip->vp, file_offset, FINDBLK_TEST)) != NULL &&
1941 bp->b_bio2.bio_offset != NOOFFSET) {
1942 bp = getblk(ip->vp, file_offset, blksize, 0, 0);
1943 bp->b_bio2.bio_offset = NOOFFSET;
1948 hammer_rel_inode(ip, 0);
1954 * This function is called when writes may have occured on the volume,
1955 * indicating that the device may be holding cached writes.
1958 hammer_io_flush_mark(hammer_volume_t volume)
1960 atomic_set_int(&volume->vol_flags, HAMMER_VOLF_NEEDFLUSH);
1964 * This function ensures that the device has flushed any cached writes out.
1967 hammer_io_flush_sync(hammer_mount_t hmp)
1969 hammer_volume_t volume;
1970 struct buf *bp_base = NULL;
1973 RB_FOREACH(volume, hammer_vol_rb_tree, &hmp->rb_vols_root) {
1974 if (volume->vol_flags & HAMMER_VOLF_NEEDFLUSH) {
1975 atomic_clear_int(&volume->vol_flags,
1976 HAMMER_VOLF_NEEDFLUSH);
1978 bp->b_bio1.bio_offset = 0;
1981 bp->b_cmd = BUF_CMD_FLUSH;
1982 bp->b_bio1.bio_caller_info1.cluster_head = bp_base;
1983 bp->b_bio1.bio_done = biodone_sync;
1984 bp->b_bio1.bio_flags |= BIO_SYNC;
1986 vn_strategy(volume->devvp, &bp->b_bio1);
1989 while ((bp = bp_base) != NULL) {
1990 bp_base = bp->b_bio1.bio_caller_info1.cluster_head;
1991 biowait(&bp->b_bio1, "hmrFLS");
1997 * Limit the amount of backlog which we allow to build up
2000 hammer_io_limit_backlog(hammer_mount_t hmp)
2002 while (hmp->io_running_space > hammer_limit_running_io) {
2003 hmp->io_running_wakeup = 1;
2004 tsleep(&hmp->io_running_wakeup, 0, "hmiolm", hz / 10);