2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * $DragonFly: src/sys/vfs/hammer/hammer_io.c,v 1.55 2008/09/15 17:02:49 dillon Exp $
37 * IO Primitives and buffer cache management
39 * All major data-tracking structures in HAMMER contain a struct hammer_io
40 * which is used to manage their backing store. We use filesystem buffers
41 * for backing store and we leave them passively associated with their
44 * If the kernel tries to destroy a passively associated buf which we cannot
45 * yet let go we set B_LOCKED in the buffer and then actively released it
50 #include <sys/fcntl.h>
51 #include <sys/nlookup.h>
55 static void hammer_io_modify(hammer_io_t io, int count);
56 static void hammer_io_deallocate(struct buf *bp);
58 static void hammer_io_direct_read_complete(struct bio *nbio);
60 static void hammer_io_direct_write_complete(struct bio *nbio);
61 static int hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data);
62 static void hammer_io_set_modlist(struct hammer_io *io);
63 static void hammer_io_flush_mark(hammer_volume_t volume);
67 * Initialize a new, already-zero'd hammer_io structure, or reinitialize
68 * an existing hammer_io structure which may have switched to another type.
71 hammer_io_init(hammer_io_t io, hammer_volume_t volume, enum hammer_io_type type)
74 io->hmp = volume->io.hmp;
79 * Helper routine to disassociate a buffer cache buffer from an I/O
80 * structure. The buffer is unlocked and marked appropriate for reclamation.
82 * The io may have 0 or 1 references depending on who called us. The
83 * caller is responsible for dealing with the refs.
85 * This call can only be made when no action is required on the buffer.
87 * The caller must own the buffer and the IO must indicate that the
88 * structure no longer owns it (io.released != 0).
91 hammer_io_disassociate(hammer_io_structure_t iou)
93 struct buf *bp = iou->io.bp;
95 KKASSERT(iou->io.released);
96 KKASSERT(iou->io.modified == 0);
97 KKASSERT(LIST_FIRST(&bp->b_dep) == (void *)iou);
102 * If the buffer was locked someone wanted to get rid of it.
104 if (bp->b_flags & B_LOCKED) {
105 --hammer_count_io_locked;
106 bp->b_flags &= ~B_LOCKED;
108 if (iou->io.reclaim) {
109 bp->b_flags |= B_NOCACHE|B_RELBUF;
113 switch(iou->io.type) {
114 case HAMMER_STRUCTURE_VOLUME:
115 iou->volume.ondisk = NULL;
117 case HAMMER_STRUCTURE_DATA_BUFFER:
118 case HAMMER_STRUCTURE_META_BUFFER:
119 case HAMMER_STRUCTURE_UNDO_BUFFER:
120 iou->buffer.ondisk = NULL;
126 * Wait for any physical IO to complete
128 * XXX we aren't interlocked against a spinlock or anything so there
129 * is a small window in the interlock / io->running == 0 test.
132 hammer_io_wait(hammer_io_t io)
137 tsleep_interlock(io, 0);
138 if (io->running == 0)
140 tsleep(io, PINTERLOCKED, "hmrflw", hz);
141 if (io->running == 0)
148 * Wait for all hammer_io-initated write I/O's to complete. This is not
149 * supposed to count direct I/O's but some can leak through (for
150 * non-full-sized direct I/Os).
153 hammer_io_wait_all(hammer_mount_t hmp, const char *ident)
155 hammer_io_flush_sync(hmp);
157 while (hmp->io_running_space)
158 tsleep(&hmp->io_running_space, 0, ident, 0);
163 * Clear a flagged error condition on a I/O buffer. The caller must hold
164 * its own ref on the buffer.
167 hammer_io_clear_error(struct hammer_io *io)
171 hammer_unref(&io->lock);
172 KKASSERT(io->lock.refs > 0);
177 #define HAMMER_MAXRA 4
180 * Load bp for a HAMMER structure. The io must be exclusively locked by
183 * This routine is mostly used on meta-data and small-data blocks. Generally
184 * speaking HAMMER assumes some locality of reference and will cluster
187 * Note that clustering occurs at the device layer, not the logical layer.
188 * If the buffers do not apply to the current operation they may apply to
192 hammer_io_read(struct vnode *devvp, struct hammer_io *io, hammer_off_t limit)
197 if ((bp = io->bp) == NULL) {
198 hammer_count_io_running_read += io->bytes;
199 if (hammer_cluster_enable) {
200 error = cluster_read(devvp, limit,
201 io->offset, io->bytes,
203 HAMMER_CLUSTER_BUFS, &io->bp);
205 error = bread(devvp, io->offset, io->bytes, &io->bp);
207 hammer_stats_disk_read += io->bytes;
208 hammer_count_io_running_read -= io->bytes;
211 * The code generally assumes b_ops/b_dep has been set-up,
212 * even if we error out here.
215 bp->b_ops = &hammer_bioops;
216 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
217 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
219 KKASSERT(io->modified == 0);
220 KKASSERT(io->running == 0);
221 KKASSERT(io->waiting == 0);
222 io->released = 0; /* we hold an active lock on bp */
230 * Similar to hammer_io_read() but returns a zero'd out buffer instead.
231 * Must be called with the IO exclusively locked.
233 * vfs_bio_clrbuf() is kinda nasty, enforce serialization against background
234 * I/O by forcing the buffer to not be in a released state before calling
237 * This function will also mark the IO as modified but it will not
238 * increment the modify_refs count.
241 hammer_io_new(struct vnode *devvp, struct hammer_io *io)
245 if ((bp = io->bp) == NULL) {
246 io->bp = getblk(devvp, io->offset, io->bytes, 0, 0);
248 bp->b_ops = &hammer_bioops;
249 KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
250 LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
252 KKASSERT(io->running == 0);
262 hammer_io_modify(io, 0);
268 * Advance the activity count on the underlying buffer because
269 * HAMMER does not getblk/brelse on every access.
272 hammer_io_advance(struct hammer_io *io)
275 buf_act_advance(io->bp);
279 * Remove potential device level aliases against buffers managed by high level
280 * vnodes. Aliases can also be created due to mixed buffer sizes or via
281 * direct access to the backing store device.
283 * This is nasty because the buffers are also VMIO-backed. Even if a buffer
284 * does not exist its backing VM pages might, and we have to invalidate
285 * those as well or a getblk() will reinstate them.
287 * Buffer cache buffers associated with hammer_buffers cannot be
291 hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset)
293 hammer_io_structure_t iou;
294 hammer_off_t phys_offset;
298 phys_offset = volume->ondisk->vol_buf_beg +
299 (zone2_offset & HAMMER_OFF_SHORT_MASK);
301 if ((bp = findblk(volume->devvp, phys_offset, FINDBLK_TEST)) != NULL)
302 bp = getblk(volume->devvp, phys_offset, bp->b_bufsize, 0, 0);
304 bp = getblk(volume->devvp, phys_offset, HAMMER_BUFSIZE, 0, 0);
305 if ((iou = (void *)LIST_FIRST(&bp->b_dep)) != NULL) {
307 hammer_ref(&iou->io.lock);
308 hammer_io_clear_modify(&iou->io, 1);
310 iou->io.released = 0;
314 KKASSERT(iou->io.lock.refs == 1);
315 hammer_rel_buffer(&iou->buffer, 0);
316 /*hammer_io_deallocate(bp);*/
321 KKASSERT((bp->b_flags & B_LOCKED) == 0);
323 bp->b_flags |= B_NOCACHE|B_RELBUF;
332 * This routine is called on the last reference to a hammer structure.
333 * The io is usually interlocked with io.loading and io.refs must be 1.
335 * This routine may return a non-NULL bp to the caller for dispoal. Disposal
336 * simply means the caller finishes decrementing the ref-count on the
337 * IO structure then brelse()'s the bp. The bp may or may not still be
338 * passively associated with the IO.
340 * The only requirement here is that modified meta-data and volume-header
341 * buffer may NOT be disassociated from the IO structure, and consequently
342 * we also leave such buffers actively associated with the IO if they already
343 * are (since the kernel can't do anything with them anyway). Only the
344 * flusher is allowed to write such buffers out. Modified pure-data and
345 * undo buffers are returned to the kernel but left passively associated
346 * so we can track when the kernel writes the bp out.
349 hammer_io_release(struct hammer_io *io, int flush)
351 union hammer_io_structure *iou = (void *)io;
354 if ((bp = io->bp) == NULL)
358 * Try to flush a dirty IO to disk if asked to by the
359 * caller or if the kernel tried to flush the buffer in the past.
361 * Kernel-initiated flushes are only allowed for pure-data buffers.
362 * meta-data and volume buffers can only be flushed explicitly
367 hammer_io_flush(io, 0);
368 } else if (bp->b_flags & B_LOCKED) {
370 case HAMMER_STRUCTURE_DATA_BUFFER:
371 hammer_io_flush(io, 0);
373 case HAMMER_STRUCTURE_UNDO_BUFFER:
374 hammer_io_flush(io, hammer_undo_reclaim(io));
379 } /* else no explicit request to flush the buffer */
383 * Wait for the IO to complete if asked to. This occurs when
384 * the buffer must be disposed of definitively during an umount
385 * or buffer invalidation.
387 if (io->waitdep && io->running) {
392 * Return control of the buffer to the kernel (with the provisio
393 * that our bioops can override kernel decisions with regards to
396 if ((flush || io->reclaim) && io->modified == 0 && io->running == 0) {
398 * Always disassociate the bp if an explicit flush
399 * was requested and the IO completed with no error
400 * (so unmount can really clean up the structure).
408 hammer_io_disassociate((hammer_io_structure_t)io);
410 } else if (io->modified) {
412 * Only certain IO types can be released to the kernel if
413 * the buffer has been modified.
415 * volume and meta-data IO types may only be explicitly
419 case HAMMER_STRUCTURE_DATA_BUFFER:
420 case HAMMER_STRUCTURE_UNDO_BUFFER:
421 if (io->released == 0) {
429 bp = NULL; /* bp left associated */
430 } else if (io->released == 0) {
432 * Clean buffers can be generally released to the kernel.
433 * We leave the bp passively associated with the HAMMER
434 * structure and use bioops to disconnect it later on
435 * if the kernel wants to discard the buffer.
437 * We can steal the structure's ownership of the bp.
440 if (bp->b_flags & B_LOCKED) {
441 hammer_io_disassociate(iou);
445 hammer_io_disassociate(iou);
448 /* return the bp (bp passively associated) */
453 * A released buffer is passively associate with our
454 * hammer_io structure. The kernel cannot destroy it
455 * without making a bioops call. If the kernel (B_LOCKED)
456 * or we (reclaim) requested that the buffer be destroyed
457 * we destroy it, otherwise we do a quick get/release to
458 * reset its position in the kernel's LRU list.
460 * Leaving the buffer passively associated allows us to
461 * use the kernel's LRU buffer flushing mechanisms rather
462 * then rolling our own.
464 * XXX there are two ways of doing this. We can re-acquire
465 * and passively release to reset the LRU, or not.
467 if (io->running == 0) {
469 if ((bp->b_flags & B_LOCKED) || io->reclaim) {
470 hammer_io_disassociate(iou);
473 /* return the bp (bp passively associated) */
477 * bp is left passively associated but we do not
478 * try to reacquire it. Interactions with the io
479 * structure will occur on completion of the bp's
489 * This routine is called with a locked IO when a flush is desired and
490 * no other references to the structure exists other then ours. This
491 * routine is ONLY called when HAMMER believes it is safe to flush a
492 * potentially modified buffer out.
495 hammer_io_flush(struct hammer_io *io, int reclaim)
500 * Degenerate case - nothing to flush if nothing is dirty.
502 if (io->modified == 0) {
507 KKASSERT(io->modify_refs <= 0);
510 * Acquire ownership of the bp, particularly before we clear our
513 * We are going to bawrite() this bp. Don't leave a window where
514 * io->released is set, we actually own the bp rather then our
520 /* BUF_KERNPROC(io->bp); */
521 /* io->released = 0; */
522 KKASSERT(io->released);
523 KKASSERT(io->bp == bp);
529 if ((bp->b_flags & B_LOCKED) == 0) {
530 bp->b_flags |= B_LOCKED;
531 ++hammer_count_io_locked;
536 * Acquire exclusive access to the bp and then clear the modified
537 * state of the buffer prior to issuing I/O to interlock any
538 * modifications made while the I/O is in progress. This shouldn't
539 * happen anyway but losing data would be worse. The modified bit
540 * will be rechecked after the IO completes.
542 * NOTE: This call also finalizes the buffer's content (inval == 0).
544 * This is only legal when lock.refs == 1 (otherwise we might clear
545 * the modified bit while there are still users of the cluster
546 * modifying the data).
548 * Do this before potentially blocking so any attempt to modify the
549 * ondisk while we are blocked blocks waiting for us.
551 hammer_ref(&io->lock);
552 hammer_io_clear_modify(io, 0);
553 hammer_unref(&io->lock);
555 if (hammer_debug_io & 0x0002)
556 kprintf("hammer io_write %016jx\n", bp->b_bio1.bio_offset);
559 * Transfer ownership to the kernel and initiate I/O.
562 io->hmp->io_running_space += io->bytes;
563 hammer_count_io_running_write += io->bytes;
565 hammer_io_flush_mark(io->volume);
568 /************************************************************************
570 ************************************************************************
572 * These routines deal with dependancies created when IO buffers get
573 * modified. The caller must call hammer_modify_*() on a referenced
574 * HAMMER structure prior to modifying its on-disk data.
576 * Any intent to modify an IO buffer acquires the related bp and imposes
577 * various write ordering dependancies.
581 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
582 * are locked until the flusher can deal with them, pure data buffers
583 * can be written out.
587 hammer_io_modify(hammer_io_t io, int count)
590 * io->modify_refs must be >= 0
592 while (io->modify_refs < 0) {
594 tsleep(io, 0, "hmrmod", 0);
598 * Shortcut if nothing to do.
600 KKASSERT(io->lock.refs != 0 && io->bp != NULL);
601 io->modify_refs += count;
602 if (io->modified && io->released == 0)
605 hammer_lock_ex(&io->lock);
606 if (io->modified == 0) {
607 hammer_io_set_modlist(io);
612 BUF_KERNPROC(io->bp);
614 KKASSERT(io->modified != 0);
616 hammer_unlock(&io->lock);
621 hammer_io_modify_done(hammer_io_t io)
623 KKASSERT(io->modify_refs > 0);
625 if (io->modify_refs == 0 && io->waitmod) {
632 hammer_io_write_interlock(hammer_io_t io)
634 while (io->modify_refs != 0) {
636 tsleep(io, 0, "hmrmod", 0);
638 io->modify_refs = -1;
642 hammer_io_done_interlock(hammer_io_t io)
644 KKASSERT(io->modify_refs == -1);
653 * Caller intends to modify a volume's ondisk structure.
655 * This is only allowed if we are the flusher or we have a ref on the
659 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
662 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
664 hammer_io_modify(&volume->io, 1);
666 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk;
667 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
668 hammer_generate_undo(trans,
669 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset),
675 * Caller intends to modify a buffer's ondisk structure.
677 * This is only allowed if we are the flusher or we have a ref on the
681 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
684 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
686 hammer_io_modify(&buffer->io, 1);
688 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk;
689 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
690 hammer_generate_undo(trans,
691 buffer->zone2_offset + rel_offset,
697 hammer_modify_volume_done(hammer_volume_t volume)
699 hammer_io_modify_done(&volume->io);
703 hammer_modify_buffer_done(hammer_buffer_t buffer)
705 hammer_io_modify_done(&buffer->io);
709 * Mark an entity as not being dirty any more and finalize any
710 * delayed adjustments to the buffer.
712 * Delayed adjustments are an important performance enhancement, allowing
713 * us to avoid recalculating B-Tree node CRCs over and over again when
714 * making bulk-modifications to the B-Tree.
716 * If inval is non-zero delayed adjustments are ignored.
718 * This routine may dereference related btree nodes and cause the
719 * buffer to be dereferenced. The caller must own a reference on io.
722 hammer_io_clear_modify(struct hammer_io *io, int inval)
724 if (io->modified == 0)
728 * Take us off the mod-list and clear the modified bit.
730 KKASSERT(io->mod_list != NULL);
731 if (io->mod_list == &io->hmp->volu_list ||
732 io->mod_list == &io->hmp->meta_list) {
733 io->hmp->locked_dirty_space -= io->bytes;
734 hammer_count_dirtybufspace -= io->bytes;
736 TAILQ_REMOVE(io->mod_list, io, mod_entry);
741 * If this bit is not set there are no delayed adjustments.
748 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference
749 * on the node (& underlying buffer). Release the node after clearing
752 if (io->type == HAMMER_STRUCTURE_META_BUFFER) {
753 hammer_buffer_t buffer = (void *)io;
757 TAILQ_FOREACH(node, &buffer->clist, entry) {
758 if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0)
760 node->flags &= ~HAMMER_NODE_NEEDSCRC;
761 KKASSERT(node->ondisk);
763 node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE);
764 hammer_rel_node(node);
768 /* caller must still have ref on io */
769 KKASSERT(io->lock.refs > 0);
773 * Clear the IO's modify list. Even though the IO is no longer modified
774 * it may still be on the lose_list. This routine is called just before
775 * the governing hammer_buffer is destroyed.
778 hammer_io_clear_modlist(struct hammer_io *io)
780 KKASSERT(io->modified == 0);
782 crit_enter(); /* biodone race against list */
783 KKASSERT(io->mod_list == &io->hmp->lose_list);
784 TAILQ_REMOVE(io->mod_list, io, mod_entry);
791 hammer_io_set_modlist(struct hammer_io *io)
793 struct hammer_mount *hmp = io->hmp;
795 KKASSERT(io->mod_list == NULL);
798 case HAMMER_STRUCTURE_VOLUME:
799 io->mod_list = &hmp->volu_list;
800 hmp->locked_dirty_space += io->bytes;
801 hammer_count_dirtybufspace += io->bytes;
803 case HAMMER_STRUCTURE_META_BUFFER:
804 io->mod_list = &hmp->meta_list;
805 hmp->locked_dirty_space += io->bytes;
806 hammer_count_dirtybufspace += io->bytes;
808 case HAMMER_STRUCTURE_UNDO_BUFFER:
809 io->mod_list = &hmp->undo_list;
811 case HAMMER_STRUCTURE_DATA_BUFFER:
812 io->mod_list = &hmp->data_list;
815 TAILQ_INSERT_TAIL(io->mod_list, io, mod_entry);
818 /************************************************************************
820 ************************************************************************
825 * Pre-IO initiation kernel callback - cluster build only
828 hammer_io_start(struct buf *bp)
833 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT!
835 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
836 * may also be set if we were marking a cluster header open. Only remove
837 * our dependancy if the modified bit is clear.
840 hammer_io_complete(struct buf *bp)
842 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep);
844 KKASSERT(iou->io.released == 1);
847 * Deal with people waiting for I/O to drain
849 if (iou->io.running) {
851 * Deal with critical write errors. Once a critical error
852 * has been flagged in hmp the UNDO FIFO will not be updated.
853 * That way crash recover will give us a consistent
856 * Because of this we can throw away failed UNDO buffers. If
857 * we throw away META or DATA buffers we risk corrupting
858 * the now read-only version of the filesystem visible to
859 * the user. Clear B_ERROR so the buffer is not re-dirtied
860 * by the kernel and ref the io so it doesn't get thrown
863 if (bp->b_flags & B_ERROR) {
864 hammer_critical_error(iou->io.hmp, NULL, bp->b_error,
865 "while flushing meta-data");
866 switch(iou->io.type) {
867 case HAMMER_STRUCTURE_UNDO_BUFFER:
870 if (iou->io.ioerror == 0) {
872 if (iou->io.lock.refs == 0)
873 ++hammer_count_refedbufs;
874 hammer_ref(&iou->io.lock);
878 bp->b_flags &= ~B_ERROR;
881 hammer_io_set_modlist(&iou->io);
882 iou->io.modified = 1;
885 hammer_stats_disk_write += iou->io.bytes;
886 hammer_count_io_running_write -= iou->io.bytes;
887 iou->io.hmp->io_running_space -= iou->io.bytes;
888 if (iou->io.hmp->io_running_space == 0)
889 wakeup(&iou->io.hmp->io_running_space);
890 KKASSERT(iou->io.hmp->io_running_space >= 0);
893 hammer_stats_disk_read += iou->io.bytes;
896 if (iou->io.waiting) {
902 * If B_LOCKED is set someone wanted to deallocate the bp at some
903 * point, do it now if refs has become zero.
905 if ((bp->b_flags & B_LOCKED) && iou->io.lock.refs == 0) {
906 KKASSERT(iou->io.modified == 0);
907 --hammer_count_io_locked;
908 bp->b_flags &= ~B_LOCKED;
909 hammer_io_deallocate(bp);
910 /* structure may be dead now */
915 * Callback from kernel when it wishes to deallocate a passively
916 * associated structure. This mostly occurs with clean buffers
917 * but it may be possible for a holding structure to be marked dirty
918 * while its buffer is passively associated. The caller owns the bp.
920 * If we cannot disassociate we set B_LOCKED to prevent the buffer
921 * from getting reused.
923 * WARNING: Because this can be called directly by getnewbuf we cannot
924 * recurse into the tree. If a bp cannot be immediately disassociated
925 * our only recourse is to set B_LOCKED.
927 * WARNING: This may be called from an interrupt via hammer_io_complete()
930 hammer_io_deallocate(struct buf *bp)
932 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep);
934 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0);
935 if (iou->io.lock.refs > 0 || iou->io.modified) {
937 * It is not legal to disassociate a modified buffer. This
938 * case really shouldn't ever occur.
940 bp->b_flags |= B_LOCKED;
941 ++hammer_count_io_locked;
944 * Disassociate the BP. If the io has no refs left we
945 * have to add it to the loose list.
947 hammer_io_disassociate(iou);
948 if (iou->io.type != HAMMER_STRUCTURE_VOLUME) {
949 KKASSERT(iou->io.bp == NULL);
950 KKASSERT(iou->io.mod_list == NULL);
951 crit_enter(); /* biodone race against list */
952 iou->io.mod_list = &iou->io.hmp->lose_list;
953 TAILQ_INSERT_TAIL(iou->io.mod_list, &iou->io, mod_entry);
960 hammer_io_fsync(struct vnode *vp)
966 * NOTE: will not be called unless we tell the kernel about the
967 * bioops. Unused... we use the mount's VFS_SYNC instead.
970 hammer_io_sync(struct mount *mp)
976 hammer_io_movedeps(struct buf *bp1, struct buf *bp2)
981 * I/O pre-check for reading and writing. HAMMER only uses this for
982 * B_CACHE buffers so checkread just shouldn't happen, but if it does
985 * Writing is a different case. We don't want the kernel to try to write
986 * out a buffer that HAMMER may be modifying passively or which has a
987 * dependancy. In addition, kernel-demanded writes can only proceed for
988 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
989 * buffer types can only be explicitly written by the flusher.
991 * checkwrite will only be called for bdwrite()n buffers. If we return
992 * success the kernel is guaranteed to initiate the buffer write.
995 hammer_io_checkread(struct buf *bp)
1001 hammer_io_checkwrite(struct buf *bp)
1003 hammer_io_t io = (void *)LIST_FIRST(&bp->b_dep);
1006 * This shouldn't happen under normal operation.
1008 if (io->type == HAMMER_STRUCTURE_VOLUME ||
1009 io->type == HAMMER_STRUCTURE_META_BUFFER) {
1011 panic("hammer_io_checkwrite: illegal buffer");
1012 if ((bp->b_flags & B_LOCKED) == 0) {
1013 bp->b_flags |= B_LOCKED;
1014 ++hammer_count_io_locked;
1020 * We can only clear the modified bit if the IO is not currently
1021 * undergoing modification. Otherwise we may miss changes.
1023 * Only data and undo buffers can reach here. These buffers do
1024 * not have terminal crc functions but we temporarily reference
1025 * the IO anyway, just in case.
1027 if (io->modify_refs == 0 && io->modified) {
1028 hammer_ref(&io->lock);
1029 hammer_io_clear_modify(io, 0);
1030 hammer_unref(&io->lock);
1031 } else if (io->modified) {
1032 KKASSERT(io->type == HAMMER_STRUCTURE_DATA_BUFFER);
1036 * The kernel is going to start the IO, set io->running.
1038 KKASSERT(io->running == 0);
1040 io->hmp->io_running_space += io->bytes;
1041 hammer_count_io_running_write += io->bytes;
1046 * Return non-zero if we wish to delay the kernel's attempt to flush
1047 * this buffer to disk.
1050 hammer_io_countdeps(struct buf *bp, int n)
1055 struct bio_ops hammer_bioops = {
1056 .io_start = hammer_io_start,
1057 .io_complete = hammer_io_complete,
1058 .io_deallocate = hammer_io_deallocate,
1059 .io_fsync = hammer_io_fsync,
1060 .io_sync = hammer_io_sync,
1061 .io_movedeps = hammer_io_movedeps,
1062 .io_countdeps = hammer_io_countdeps,
1063 .io_checkread = hammer_io_checkread,
1064 .io_checkwrite = hammer_io_checkwrite,
1067 /************************************************************************
1069 ************************************************************************
1071 * These functions operate directly on the buffer cache buffer associated
1072 * with a front-end vnode rather then a back-end device vnode.
1076 * Read a buffer associated with a front-end vnode directly from the
1077 * disk media. The bio may be issued asynchronously. If leaf is non-NULL
1078 * we validate the CRC.
1080 * We must check for the presence of a HAMMER buffer to handle the case
1081 * where the reblocker has rewritten the data (which it does via the HAMMER
1082 * buffer system, not via the high-level vnode buffer cache), but not yet
1083 * committed the buffer to the media.
1086 hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1087 hammer_btree_leaf_elm_t leaf)
1089 hammer_off_t buf_offset;
1090 hammer_off_t zone2_offset;
1091 hammer_volume_t volume;
1097 buf_offset = bio->bio_offset;
1098 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) ==
1099 HAMMER_ZONE_LARGE_DATA);
1102 * The buffer cache may have an aliased buffer (the reblocker can
1103 * write them). If it does we have to sync any dirty data before
1104 * we can build our direct-read. This is a non-critical code path.
1107 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize);
1110 * Resolve to a zone-2 offset. The conversion just requires
1111 * munging the top 4 bits but we want to abstract it anyway
1112 * so the blockmap code can verify the zone assignment.
1114 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
1117 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
1118 HAMMER_ZONE_RAW_BUFFER);
1121 * Resolve volume and raw-offset for 3rd level bio. The
1122 * offset will be specific to the volume.
1124 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1125 volume = hammer_get_volume(hmp, vol_no, &error);
1126 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1133 nbio = push_bio(bio);
1134 nbio->bio_offset = volume->ondisk->vol_buf_beg +
1135 (zone2_offset & HAMMER_OFF_SHORT_MASK);
1138 * XXX disabled - our CRC check doesn't work if the OS
1139 * does bogus_page replacement on the direct-read.
1141 if (leaf && hammer_verify_data) {
1142 nbio->bio_done = hammer_io_direct_read_complete;
1143 nbio->bio_caller_info1.uvalue32 = leaf->data_crc;
1146 hammer_stats_disk_read += bp->b_bufsize;
1147 vn_strategy(volume->devvp, nbio);
1149 hammer_rel_volume(volume, 0);
1152 kprintf("hammer_direct_read: failed @ %016llx\n",
1153 (long long)zone2_offset);
1154 bp->b_error = error;
1155 bp->b_flags |= B_ERROR;
1163 * On completion of the BIO this callback must check the data CRC
1164 * and chain to the previous bio.
1168 hammer_io_direct_read_complete(struct bio *nbio)
1172 u_int32_t rec_crc = nbio->bio_caller_info1.uvalue32;
1175 if (crc32(bp->b_data, bp->b_bufsize) != rec_crc) {
1176 kprintf("HAMMER: data_crc error @%016llx/%d\n",
1177 nbio->bio_offset, bp->b_bufsize);
1178 if (hammer_debug_critical)
1179 Debugger("data_crc on read");
1180 bp->b_flags |= B_ERROR;
1183 obio = pop_bio(nbio);
1189 * Write a buffer associated with a front-end vnode directly to the
1190 * disk media. The bio may be issued asynchronously.
1192 * The BIO is associated with the specified record and RECF_DIRECT_IO
1193 * is set. The recorded is added to its object.
1196 hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record,
1199 hammer_btree_leaf_elm_t leaf = &record->leaf;
1200 hammer_off_t buf_offset;
1201 hammer_off_t zone2_offset;
1202 hammer_volume_t volume;
1203 hammer_buffer_t buffer;
1210 buf_offset = leaf->data_offset;
1212 KKASSERT(buf_offset > HAMMER_ZONE_BTREE);
1213 KKASSERT(bio->bio_buf->b_cmd == BUF_CMD_WRITE);
1215 if ((buf_offset & HAMMER_BUFMASK) == 0 &&
1216 leaf->data_len >= HAMMER_BUFSIZE) {
1218 * We are using the vnode's bio to write directly to the
1219 * media, any hammer_buffer at the same zone-X offset will
1220 * now have stale data.
1222 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
1223 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1224 volume = hammer_get_volume(hmp, vol_no, &error);
1226 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1230 KKASSERT((bp->b_bufsize & HAMMER_BUFMASK) == 0);
1232 hammer_del_buffers(hmp, buf_offset,
1233 zone2_offset, bp->b_bufsize);
1237 * Second level bio - cached zone2 offset.
1239 * (We can put our bio_done function in either the
1240 * 2nd or 3rd level).
1242 nbio = push_bio(bio);
1243 nbio->bio_offset = zone2_offset;
1244 nbio->bio_done = hammer_io_direct_write_complete;
1245 nbio->bio_caller_info1.ptr = record;
1246 record->zone2_offset = zone2_offset;
1247 record->flags |= HAMMER_RECF_DIRECT_IO |
1248 HAMMER_RECF_DIRECT_INVAL;
1251 * Third level bio - raw offset specific to the
1254 zone2_offset &= HAMMER_OFF_SHORT_MASK;
1255 nbio = push_bio(nbio);
1256 nbio->bio_offset = volume->ondisk->vol_buf_beg +
1258 hammer_stats_disk_write += bp->b_bufsize;
1259 vn_strategy(volume->devvp, nbio);
1260 hammer_io_flush_mark(volume);
1262 hammer_rel_volume(volume, 0);
1265 * Must fit in a standard HAMMER buffer. In this case all
1266 * consumers use the HAMMER buffer system and RECF_DIRECT_IO
1267 * does not need to be set-up.
1269 KKASSERT(((buf_offset ^ (buf_offset + leaf->data_len - 1)) & ~HAMMER_BUFMASK64) == 0);
1271 ptr = hammer_bread(hmp, buf_offset, &error, &buffer);
1274 bp->b_flags |= B_AGE;
1275 hammer_io_modify(&buffer->io, 1);
1276 bcopy(bp->b_data, ptr, leaf->data_len);
1277 hammer_io_modify_done(&buffer->io);
1278 hammer_rel_buffer(buffer, 0);
1285 * The record is all setup now, add it. Potential conflics
1286 * have already been dealt with.
1288 error = hammer_mem_add(record);
1289 KKASSERT(error == 0);
1292 * Major suckage occured. Also note: The record was never added
1293 * to the tree so we do not have to worry about the backend.
1295 kprintf("hammer_direct_write: failed @ %016llx\n",
1296 (long long)leaf->data_offset);
1300 bp->b_flags |= B_ERROR;
1302 record->flags |= HAMMER_RECF_DELETED_FE;
1303 hammer_rel_mem_record(record);
1309 * On completion of the BIO this callback must disconnect
1310 * it from the hammer_record and chain to the previous bio.
1312 * An I/O error forces the mount to read-only. Data buffers
1313 * are not B_LOCKED like meta-data buffers are, so we have to
1314 * throw the buffer away to prevent the kernel from retrying.
1318 hammer_io_direct_write_complete(struct bio *nbio)
1322 hammer_record_t record = nbio->bio_caller_info1.ptr;
1325 obio = pop_bio(nbio);
1326 if (bp->b_flags & B_ERROR) {
1327 hammer_critical_error(record->ip->hmp, record->ip,
1329 "while writing bulk data");
1330 bp->b_flags |= B_INVAL;
1334 KKASSERT(record != NULL);
1335 KKASSERT(record->flags & HAMMER_RECF_DIRECT_IO);
1336 if (record->flags & HAMMER_RECF_DIRECT_WAIT) {
1337 record->flags &= ~(HAMMER_RECF_DIRECT_IO |
1338 HAMMER_RECF_DIRECT_WAIT);
1339 /* record can disappear once DIRECT_IO flag is cleared */
1340 wakeup(&record->flags);
1342 record->flags &= ~HAMMER_RECF_DIRECT_IO;
1343 /* record can disappear once DIRECT_IO flag is cleared */
1349 * This is called before a record is either committed to the B-Tree
1350 * or destroyed, to resolve any associated direct-IO.
1352 * (1) We must wait for any direct-IO related to the record to complete.
1354 * (2) We must remove any buffer cache aliases for data accessed via
1355 * leaf->data_offset or zone2_offset so non-direct-IO consumers
1356 * (the mirroring and reblocking code) do not see stale data.
1359 hammer_io_direct_wait(hammer_record_t record)
1362 * Wait for I/O to complete
1364 if (record->flags & HAMMER_RECF_DIRECT_IO) {
1366 while (record->flags & HAMMER_RECF_DIRECT_IO) {
1367 record->flags |= HAMMER_RECF_DIRECT_WAIT;
1368 tsleep(&record->flags, 0, "hmdiow", 0);
1374 * Invalidate any related buffer cache aliases associated with the
1375 * backing device. This is needed because the buffer cache buffer
1376 * for file data is associated with the file vnode, not the backing
1379 * XXX I do not think this case can occur any more now that
1380 * reservations ensure that all such buffers are removed before
1381 * an area can be reused.
1383 if (record->flags & HAMMER_RECF_DIRECT_INVAL) {
1384 KKASSERT(record->leaf.data_offset);
1385 hammer_del_buffers(record->ip->hmp, record->leaf.data_offset,
1386 record->zone2_offset, record->leaf.data_len,
1388 record->flags &= ~HAMMER_RECF_DIRECT_INVAL;
1393 * This is called to remove the second-level cached zone-2 offset from
1394 * frontend buffer cache buffers, now stale due to a data relocation.
1395 * These offsets are generated by cluster_read() via VOP_BMAP, or directly
1396 * by hammer_vop_strategy_read().
1398 * This is rather nasty because here we have something like the reblocker
1399 * scanning the raw B-Tree with no held references on anything, really,
1400 * other then a shared lock on the B-Tree node, and we have to access the
1401 * frontend's buffer cache to check for and clean out the association.
1402 * Specifically, if the reblocker is moving data on the disk, these cached
1403 * offsets will become invalid.
1405 * Only data record types associated with the large-data zone are subject
1406 * to direct-io and need to be checked.
1410 hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf)
1412 struct hammer_inode_info iinfo;
1415 if (leaf->base.rec_type != HAMMER_RECTYPE_DATA)
1417 zone = HAMMER_ZONE_DECODE(leaf->data_offset);
1418 if (zone != HAMMER_ZONE_LARGE_DATA_INDEX)
1420 iinfo.obj_id = leaf->base.obj_id;
1421 iinfo.obj_asof = 0; /* unused */
1422 iinfo.obj_localization = leaf->base.localization &
1423 HAMMER_LOCALIZE_PSEUDOFS_MASK;
1424 iinfo.u.leaf = leaf;
1425 hammer_scan_inode_snapshots(hmp, &iinfo,
1426 hammer_io_direct_uncache_callback,
1431 hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data)
1433 hammer_inode_info_t iinfo = data;
1434 hammer_off_t data_offset;
1435 hammer_off_t file_offset;
1442 data_offset = iinfo->u.leaf->data_offset;
1443 file_offset = iinfo->u.leaf->base.key - iinfo->u.leaf->data_len;
1444 blksize = iinfo->u.leaf->data_len;
1445 KKASSERT((blksize & HAMMER_BUFMASK) == 0);
1447 hammer_ref(&ip->lock);
1448 if (hammer_get_vnode(ip, &vp) == 0) {
1449 if ((bp = findblk(ip->vp, file_offset, FINDBLK_TEST)) != NULL &&
1450 bp->b_bio2.bio_offset != NOOFFSET) {
1451 bp = getblk(ip->vp, file_offset, blksize, 0, 0);
1452 bp->b_bio2.bio_offset = NOOFFSET;
1457 hammer_rel_inode(ip, 0);
1463 * This function is called when writes may have occured on the volume,
1464 * indicating that the device may be holding cached writes.
1467 hammer_io_flush_mark(hammer_volume_t volume)
1469 volume->vol_flags |= HAMMER_VOLF_NEEDFLUSH;
1473 * This function ensures that the device has flushed any cached writes out.
1476 hammer_io_flush_sync(hammer_mount_t hmp)
1478 hammer_volume_t volume;
1479 struct buf *bp_base = NULL;
1482 RB_FOREACH(volume, hammer_vol_rb_tree, &hmp->rb_vols_root) {
1483 if (volume->vol_flags & HAMMER_VOLF_NEEDFLUSH) {
1484 volume->vol_flags &= ~HAMMER_VOLF_NEEDFLUSH;
1486 bp->b_bio1.bio_offset = 0;
1489 bp->b_cmd = BUF_CMD_FLUSH;
1490 bp->b_bio1.bio_caller_info1.cluster_head = bp_base;
1491 bp->b_bio1.bio_done = biodone_sync;
1492 bp->b_bio1.bio_flags |= BIO_SYNC;
1494 vn_strategy(volume->devvp, &bp->b_bio1);
1497 while ((bp = bp_base) != NULL) {
1498 bp_base = bp->b_bio1.bio_caller_info1.cluster_head;
1499 biowait(&bp->b_bio1, "hmrFLS");