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);
556 * Transfer ownership to the kernel and initiate I/O.
559 io->hmp->io_running_space += io->bytes;
560 hammer_count_io_running_write += io->bytes;
562 hammer_io_flush_mark(io->volume);
565 /************************************************************************
567 ************************************************************************
569 * These routines deal with dependancies created when IO buffers get
570 * modified. The caller must call hammer_modify_*() on a referenced
571 * HAMMER structure prior to modifying its on-disk data.
573 * Any intent to modify an IO buffer acquires the related bp and imposes
574 * various write ordering dependancies.
578 * Mark a HAMMER structure as undergoing modification. Meta-data buffers
579 * are locked until the flusher can deal with them, pure data buffers
580 * can be written out.
584 hammer_io_modify(hammer_io_t io, int count)
587 * io->modify_refs must be >= 0
589 while (io->modify_refs < 0) {
591 tsleep(io, 0, "hmrmod", 0);
595 * Shortcut if nothing to do.
597 KKASSERT(io->lock.refs != 0 && io->bp != NULL);
598 io->modify_refs += count;
599 if (io->modified && io->released == 0)
602 hammer_lock_ex(&io->lock);
603 if (io->modified == 0) {
604 hammer_io_set_modlist(io);
609 BUF_KERNPROC(io->bp);
611 KKASSERT(io->modified != 0);
613 hammer_unlock(&io->lock);
618 hammer_io_modify_done(hammer_io_t io)
620 KKASSERT(io->modify_refs > 0);
622 if (io->modify_refs == 0 && io->waitmod) {
629 hammer_io_write_interlock(hammer_io_t io)
631 while (io->modify_refs != 0) {
633 tsleep(io, 0, "hmrmod", 0);
635 io->modify_refs = -1;
639 hammer_io_done_interlock(hammer_io_t io)
641 KKASSERT(io->modify_refs == -1);
650 * Caller intends to modify a volume's ondisk structure.
652 * This is only allowed if we are the flusher or we have a ref on the
656 hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
659 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
661 hammer_io_modify(&volume->io, 1);
663 intptr_t rel_offset = (intptr_t)base - (intptr_t)volume->ondisk;
664 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
665 hammer_generate_undo(trans,
666 HAMMER_ENCODE_RAW_VOLUME(volume->vol_no, rel_offset),
672 * Caller intends to modify a buffer's ondisk structure.
674 * This is only allowed if we are the flusher or we have a ref on the
678 hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
681 KKASSERT (trans == NULL || trans->sync_lock_refs > 0);
683 hammer_io_modify(&buffer->io, 1);
685 intptr_t rel_offset = (intptr_t)base - (intptr_t)buffer->ondisk;
686 KKASSERT((rel_offset & ~(intptr_t)HAMMER_BUFMASK) == 0);
687 hammer_generate_undo(trans,
688 buffer->zone2_offset + rel_offset,
694 hammer_modify_volume_done(hammer_volume_t volume)
696 hammer_io_modify_done(&volume->io);
700 hammer_modify_buffer_done(hammer_buffer_t buffer)
702 hammer_io_modify_done(&buffer->io);
706 * Mark an entity as not being dirty any more and finalize any
707 * delayed adjustments to the buffer.
709 * Delayed adjustments are an important performance enhancement, allowing
710 * us to avoid recalculating B-Tree node CRCs over and over again when
711 * making bulk-modifications to the B-Tree.
713 * If inval is non-zero delayed adjustments are ignored.
715 * This routine may dereference related btree nodes and cause the
716 * buffer to be dereferenced. The caller must own a reference on io.
719 hammer_io_clear_modify(struct hammer_io *io, int inval)
721 if (io->modified == 0)
725 * Take us off the mod-list and clear the modified bit.
727 KKASSERT(io->mod_list != NULL);
728 if (io->mod_list == &io->hmp->volu_list ||
729 io->mod_list == &io->hmp->meta_list) {
730 io->hmp->locked_dirty_space -= io->bytes;
731 hammer_count_dirtybufspace -= io->bytes;
733 TAILQ_REMOVE(io->mod_list, io, mod_entry);
738 * If this bit is not set there are no delayed adjustments.
745 * Finalize requested CRCs. The NEEDSCRC flag also holds a reference
746 * on the node (& underlying buffer). Release the node after clearing
749 if (io->type == HAMMER_STRUCTURE_META_BUFFER) {
750 hammer_buffer_t buffer = (void *)io;
754 TAILQ_FOREACH(node, &buffer->clist, entry) {
755 if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0)
757 node->flags &= ~HAMMER_NODE_NEEDSCRC;
758 KKASSERT(node->ondisk);
760 node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE);
761 hammer_rel_node(node);
765 /* caller must still have ref on io */
766 KKASSERT(io->lock.refs > 0);
770 * Clear the IO's modify list. Even though the IO is no longer modified
771 * it may still be on the lose_list. This routine is called just before
772 * the governing hammer_buffer is destroyed.
775 hammer_io_clear_modlist(struct hammer_io *io)
777 KKASSERT(io->modified == 0);
779 crit_enter(); /* biodone race against list */
780 KKASSERT(io->mod_list == &io->hmp->lose_list);
781 TAILQ_REMOVE(io->mod_list, io, mod_entry);
788 hammer_io_set_modlist(struct hammer_io *io)
790 struct hammer_mount *hmp = io->hmp;
792 KKASSERT(io->mod_list == NULL);
795 case HAMMER_STRUCTURE_VOLUME:
796 io->mod_list = &hmp->volu_list;
797 hmp->locked_dirty_space += io->bytes;
798 hammer_count_dirtybufspace += io->bytes;
800 case HAMMER_STRUCTURE_META_BUFFER:
801 io->mod_list = &hmp->meta_list;
802 hmp->locked_dirty_space += io->bytes;
803 hammer_count_dirtybufspace += io->bytes;
805 case HAMMER_STRUCTURE_UNDO_BUFFER:
806 io->mod_list = &hmp->undo_list;
808 case HAMMER_STRUCTURE_DATA_BUFFER:
809 io->mod_list = &hmp->data_list;
812 TAILQ_INSERT_TAIL(io->mod_list, io, mod_entry);
815 /************************************************************************
817 ************************************************************************
822 * Pre-IO initiation kernel callback - cluster build only
825 hammer_io_start(struct buf *bp)
830 * Post-IO completion kernel callback - MAY BE CALLED FROM INTERRUPT!
832 * NOTE: HAMMER may modify a buffer after initiating I/O. The modified bit
833 * may also be set if we were marking a cluster header open. Only remove
834 * our dependancy if the modified bit is clear.
837 hammer_io_complete(struct buf *bp)
839 union hammer_io_structure *iou = (void *)LIST_FIRST(&bp->b_dep);
841 KKASSERT(iou->io.released == 1);
844 * Deal with people waiting for I/O to drain
846 if (iou->io.running) {
848 * Deal with critical write errors. Once a critical error
849 * has been flagged in hmp the UNDO FIFO will not be updated.
850 * That way crash recover will give us a consistent
853 * Because of this we can throw away failed UNDO buffers. If
854 * we throw away META or DATA buffers we risk corrupting
855 * the now read-only version of the filesystem visible to
856 * the user. Clear B_ERROR so the buffer is not re-dirtied
857 * by the kernel and ref the io so it doesn't get thrown
860 if (bp->b_flags & B_ERROR) {
861 hammer_critical_error(iou->io.hmp, NULL, bp->b_error,
862 "while flushing meta-data");
863 switch(iou->io.type) {
864 case HAMMER_STRUCTURE_UNDO_BUFFER:
867 if (iou->io.ioerror == 0) {
869 if (iou->io.lock.refs == 0)
870 ++hammer_count_refedbufs;
871 hammer_ref(&iou->io.lock);
875 bp->b_flags &= ~B_ERROR;
878 hammer_io_set_modlist(&iou->io);
879 iou->io.modified = 1;
882 hammer_stats_disk_write += iou->io.bytes;
883 hammer_count_io_running_write -= iou->io.bytes;
884 iou->io.hmp->io_running_space -= iou->io.bytes;
885 if (iou->io.hmp->io_running_space == 0)
886 wakeup(&iou->io.hmp->io_running_space);
887 KKASSERT(iou->io.hmp->io_running_space >= 0);
890 hammer_stats_disk_read += iou->io.bytes;
893 if (iou->io.waiting) {
899 * If B_LOCKED is set someone wanted to deallocate the bp at some
900 * point, do it now if refs has become zero.
902 if ((bp->b_flags & B_LOCKED) && iou->io.lock.refs == 0) {
903 KKASSERT(iou->io.modified == 0);
904 --hammer_count_io_locked;
905 bp->b_flags &= ~B_LOCKED;
906 hammer_io_deallocate(bp);
907 /* structure may be dead now */
912 * Callback from kernel when it wishes to deallocate a passively
913 * associated structure. This mostly occurs with clean buffers
914 * but it may be possible for a holding structure to be marked dirty
915 * while its buffer is passively associated. The caller owns the bp.
917 * If we cannot disassociate we set B_LOCKED to prevent the buffer
918 * from getting reused.
920 * WARNING: Because this can be called directly by getnewbuf we cannot
921 * recurse into the tree. If a bp cannot be immediately disassociated
922 * our only recourse is to set B_LOCKED.
924 * WARNING: This may be called from an interrupt via hammer_io_complete()
927 hammer_io_deallocate(struct buf *bp)
929 hammer_io_structure_t iou = (void *)LIST_FIRST(&bp->b_dep);
931 KKASSERT((bp->b_flags & B_LOCKED) == 0 && iou->io.running == 0);
932 if (iou->io.lock.refs > 0 || iou->io.modified) {
934 * It is not legal to disassociate a modified buffer. This
935 * case really shouldn't ever occur.
937 bp->b_flags |= B_LOCKED;
938 ++hammer_count_io_locked;
941 * Disassociate the BP. If the io has no refs left we
942 * have to add it to the loose list.
944 hammer_io_disassociate(iou);
945 if (iou->io.type != HAMMER_STRUCTURE_VOLUME) {
946 KKASSERT(iou->io.bp == NULL);
947 KKASSERT(iou->io.mod_list == NULL);
948 crit_enter(); /* biodone race against list */
949 iou->io.mod_list = &iou->io.hmp->lose_list;
950 TAILQ_INSERT_TAIL(iou->io.mod_list, &iou->io, mod_entry);
957 hammer_io_fsync(struct vnode *vp)
963 * NOTE: will not be called unless we tell the kernel about the
964 * bioops. Unused... we use the mount's VFS_SYNC instead.
967 hammer_io_sync(struct mount *mp)
973 hammer_io_movedeps(struct buf *bp1, struct buf *bp2)
978 * I/O pre-check for reading and writing. HAMMER only uses this for
979 * B_CACHE buffers so checkread just shouldn't happen, but if it does
982 * Writing is a different case. We don't want the kernel to try to write
983 * out a buffer that HAMMER may be modifying passively or which has a
984 * dependancy. In addition, kernel-demanded writes can only proceed for
985 * certain types of buffers (i.e. UNDO and DATA types). Other dirty
986 * buffer types can only be explicitly written by the flusher.
988 * checkwrite will only be called for bdwrite()n buffers. If we return
989 * success the kernel is guaranteed to initiate the buffer write.
992 hammer_io_checkread(struct buf *bp)
998 hammer_io_checkwrite(struct buf *bp)
1000 hammer_io_t io = (void *)LIST_FIRST(&bp->b_dep);
1003 * This shouldn't happen under normal operation.
1005 if (io->type == HAMMER_STRUCTURE_VOLUME ||
1006 io->type == HAMMER_STRUCTURE_META_BUFFER) {
1008 panic("hammer_io_checkwrite: illegal buffer");
1009 if ((bp->b_flags & B_LOCKED) == 0) {
1010 bp->b_flags |= B_LOCKED;
1011 ++hammer_count_io_locked;
1017 * We can only clear the modified bit if the IO is not currently
1018 * undergoing modification. Otherwise we may miss changes.
1020 * Only data and undo buffers can reach here. These buffers do
1021 * not have terminal crc functions but we temporarily reference
1022 * the IO anyway, just in case.
1024 if (io->modify_refs == 0 && io->modified) {
1025 hammer_ref(&io->lock);
1026 hammer_io_clear_modify(io, 0);
1027 hammer_unref(&io->lock);
1028 } else if (io->modified) {
1029 KKASSERT(io->type == HAMMER_STRUCTURE_DATA_BUFFER);
1033 * The kernel is going to start the IO, set io->running.
1035 KKASSERT(io->running == 0);
1037 io->hmp->io_running_space += io->bytes;
1038 hammer_count_io_running_write += io->bytes;
1043 * Return non-zero if we wish to delay the kernel's attempt to flush
1044 * this buffer to disk.
1047 hammer_io_countdeps(struct buf *bp, int n)
1052 struct bio_ops hammer_bioops = {
1053 .io_start = hammer_io_start,
1054 .io_complete = hammer_io_complete,
1055 .io_deallocate = hammer_io_deallocate,
1056 .io_fsync = hammer_io_fsync,
1057 .io_sync = hammer_io_sync,
1058 .io_movedeps = hammer_io_movedeps,
1059 .io_countdeps = hammer_io_countdeps,
1060 .io_checkread = hammer_io_checkread,
1061 .io_checkwrite = hammer_io_checkwrite,
1064 /************************************************************************
1066 ************************************************************************
1068 * These functions operate directly on the buffer cache buffer associated
1069 * with a front-end vnode rather then a back-end device vnode.
1073 * Read a buffer associated with a front-end vnode directly from the
1074 * disk media. The bio may be issued asynchronously. If leaf is non-NULL
1075 * we validate the CRC.
1077 * We must check for the presence of a HAMMER buffer to handle the case
1078 * where the reblocker has rewritten the data (which it does via the HAMMER
1079 * buffer system, not via the high-level vnode buffer cache), but not yet
1080 * committed the buffer to the media.
1083 hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1084 hammer_btree_leaf_elm_t leaf)
1086 hammer_off_t buf_offset;
1087 hammer_off_t zone2_offset;
1088 hammer_volume_t volume;
1094 buf_offset = bio->bio_offset;
1095 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) ==
1096 HAMMER_ZONE_LARGE_DATA);
1099 * The buffer cache may have an aliased buffer (the reblocker can
1100 * write them). If it does we have to sync any dirty data before
1101 * we can build our direct-read. This is a non-critical code path.
1104 hammer_sync_buffers(hmp, buf_offset, bp->b_bufsize);
1107 * Resolve to a zone-2 offset. The conversion just requires
1108 * munging the top 4 bits but we want to abstract it anyway
1109 * so the blockmap code can verify the zone assignment.
1111 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
1114 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
1115 HAMMER_ZONE_RAW_BUFFER);
1118 * Resolve volume and raw-offset for 3rd level bio. The
1119 * offset will be specific to the volume.
1121 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1122 volume = hammer_get_volume(hmp, vol_no, &error);
1123 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1130 nbio = push_bio(bio);
1131 nbio->bio_offset = volume->ondisk->vol_buf_beg +
1132 (zone2_offset & HAMMER_OFF_SHORT_MASK);
1135 * XXX disabled - our CRC check doesn't work if the OS
1136 * does bogus_page replacement on the direct-read.
1138 if (leaf && hammer_verify_data) {
1139 nbio->bio_done = hammer_io_direct_read_complete;
1140 nbio->bio_caller_info1.uvalue32 = leaf->data_crc;
1143 hammer_stats_disk_read += bp->b_bufsize;
1144 vn_strategy(volume->devvp, nbio);
1146 hammer_rel_volume(volume, 0);
1149 kprintf("hammer_direct_read: failed @ %016llx\n",
1150 (long long)zone2_offset);
1151 bp->b_error = error;
1152 bp->b_flags |= B_ERROR;
1160 * On completion of the BIO this callback must check the data CRC
1161 * and chain to the previous bio.
1165 hammer_io_direct_read_complete(struct bio *nbio)
1169 u_int32_t rec_crc = nbio->bio_caller_info1.uvalue32;
1172 if (crc32(bp->b_data, bp->b_bufsize) != rec_crc) {
1173 kprintf("HAMMER: data_crc error @%016llx/%d\n",
1174 nbio->bio_offset, bp->b_bufsize);
1175 if (hammer_debug_critical)
1176 Debugger("data_crc on read");
1177 bp->b_flags |= B_ERROR;
1180 obio = pop_bio(nbio);
1186 * Write a buffer associated with a front-end vnode directly to the
1187 * disk media. The bio may be issued asynchronously.
1189 * The BIO is associated with the specified record and RECF_DIRECT_IO
1190 * is set. The recorded is added to its object.
1193 hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record,
1196 hammer_btree_leaf_elm_t leaf = &record->leaf;
1197 hammer_off_t buf_offset;
1198 hammer_off_t zone2_offset;
1199 hammer_volume_t volume;
1200 hammer_buffer_t buffer;
1207 buf_offset = leaf->data_offset;
1209 KKASSERT(buf_offset > HAMMER_ZONE_BTREE);
1210 KKASSERT(bio->bio_buf->b_cmd == BUF_CMD_WRITE);
1212 if ((buf_offset & HAMMER_BUFMASK) == 0 &&
1213 leaf->data_len >= HAMMER_BUFSIZE) {
1215 * We are using the vnode's bio to write directly to the
1216 * media, any hammer_buffer at the same zone-X offset will
1217 * now have stale data.
1219 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, &error);
1220 vol_no = HAMMER_VOL_DECODE(zone2_offset);
1221 volume = hammer_get_volume(hmp, vol_no, &error);
1223 if (error == 0 && zone2_offset >= volume->maxbuf_off)
1227 KKASSERT((bp->b_bufsize & HAMMER_BUFMASK) == 0);
1229 hammer_del_buffers(hmp, buf_offset,
1230 zone2_offset, bp->b_bufsize);
1234 * Second level bio - cached zone2 offset.
1236 * (We can put our bio_done function in either the
1237 * 2nd or 3rd level).
1239 nbio = push_bio(bio);
1240 nbio->bio_offset = zone2_offset;
1241 nbio->bio_done = hammer_io_direct_write_complete;
1242 nbio->bio_caller_info1.ptr = record;
1243 record->zone2_offset = zone2_offset;
1244 record->flags |= HAMMER_RECF_DIRECT_IO |
1245 HAMMER_RECF_DIRECT_INVAL;
1248 * Third level bio - raw offset specific to the
1251 zone2_offset &= HAMMER_OFF_SHORT_MASK;
1252 nbio = push_bio(nbio);
1253 nbio->bio_offset = volume->ondisk->vol_buf_beg +
1255 hammer_stats_disk_write += bp->b_bufsize;
1256 vn_strategy(volume->devvp, nbio);
1257 hammer_io_flush_mark(volume);
1259 hammer_rel_volume(volume, 0);
1262 * Must fit in a standard HAMMER buffer. In this case all
1263 * consumers use the HAMMER buffer system and RECF_DIRECT_IO
1264 * does not need to be set-up.
1266 KKASSERT(((buf_offset ^ (buf_offset + leaf->data_len - 1)) & ~HAMMER_BUFMASK64) == 0);
1268 ptr = hammer_bread(hmp, buf_offset, &error, &buffer);
1271 bp->b_flags |= B_AGE;
1272 hammer_io_modify(&buffer->io, 1);
1273 bcopy(bp->b_data, ptr, leaf->data_len);
1274 hammer_io_modify_done(&buffer->io);
1275 hammer_rel_buffer(buffer, 0);
1282 * The record is all setup now, add it. Potential conflics
1283 * have already been dealt with.
1285 error = hammer_mem_add(record);
1286 KKASSERT(error == 0);
1289 * Major suckage occured. Also note: The record was never added
1290 * to the tree so we do not have to worry about the backend.
1292 kprintf("hammer_direct_write: failed @ %016llx\n",
1293 (long long)leaf->data_offset);
1297 bp->b_flags |= B_ERROR;
1299 record->flags |= HAMMER_RECF_DELETED_FE;
1300 hammer_rel_mem_record(record);
1306 * On completion of the BIO this callback must disconnect
1307 * it from the hammer_record and chain to the previous bio.
1309 * An I/O error forces the mount to read-only. Data buffers
1310 * are not B_LOCKED like meta-data buffers are, so we have to
1311 * throw the buffer away to prevent the kernel from retrying.
1315 hammer_io_direct_write_complete(struct bio *nbio)
1319 hammer_record_t record = nbio->bio_caller_info1.ptr;
1322 obio = pop_bio(nbio);
1323 if (bp->b_flags & B_ERROR) {
1324 hammer_critical_error(record->ip->hmp, record->ip,
1326 "while writing bulk data");
1327 bp->b_flags |= B_INVAL;
1331 KKASSERT(record != NULL);
1332 KKASSERT(record->flags & HAMMER_RECF_DIRECT_IO);
1333 if (record->flags & HAMMER_RECF_DIRECT_WAIT) {
1334 record->flags &= ~(HAMMER_RECF_DIRECT_IO |
1335 HAMMER_RECF_DIRECT_WAIT);
1336 /* record can disappear once DIRECT_IO flag is cleared */
1337 wakeup(&record->flags);
1339 record->flags &= ~HAMMER_RECF_DIRECT_IO;
1340 /* record can disappear once DIRECT_IO flag is cleared */
1346 * This is called before a record is either committed to the B-Tree
1347 * or destroyed, to resolve any associated direct-IO.
1349 * (1) We must wait for any direct-IO related to the record to complete.
1351 * (2) We must remove any buffer cache aliases for data accessed via
1352 * leaf->data_offset or zone2_offset so non-direct-IO consumers
1353 * (the mirroring and reblocking code) do not see stale data.
1356 hammer_io_direct_wait(hammer_record_t record)
1359 * Wait for I/O to complete
1361 if (record->flags & HAMMER_RECF_DIRECT_IO) {
1363 while (record->flags & HAMMER_RECF_DIRECT_IO) {
1364 record->flags |= HAMMER_RECF_DIRECT_WAIT;
1365 tsleep(&record->flags, 0, "hmdiow", 0);
1371 * Invalidate any related buffer cache aliases associated with the
1372 * backing device. This is needed because the buffer cache buffer
1373 * for file data is associated with the file vnode, not the backing
1376 * XXX I do not think this case can occur any more now that
1377 * reservations ensure that all such buffers are removed before
1378 * an area can be reused.
1380 if (record->flags & HAMMER_RECF_DIRECT_INVAL) {
1381 KKASSERT(record->leaf.data_offset);
1382 hammer_del_buffers(record->ip->hmp, record->leaf.data_offset,
1383 record->zone2_offset, record->leaf.data_len,
1385 record->flags &= ~HAMMER_RECF_DIRECT_INVAL;
1390 * This is called to remove the second-level cached zone-2 offset from
1391 * frontend buffer cache buffers, now stale due to a data relocation.
1392 * These offsets are generated by cluster_read() via VOP_BMAP, or directly
1393 * by hammer_vop_strategy_read().
1395 * This is rather nasty because here we have something like the reblocker
1396 * scanning the raw B-Tree with no held references on anything, really,
1397 * other then a shared lock on the B-Tree node, and we have to access the
1398 * frontend's buffer cache to check for and clean out the association.
1399 * Specifically, if the reblocker is moving data on the disk, these cached
1400 * offsets will become invalid.
1402 * Only data record types associated with the large-data zone are subject
1403 * to direct-io and need to be checked.
1407 hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf)
1409 struct hammer_inode_info iinfo;
1412 if (leaf->base.rec_type != HAMMER_RECTYPE_DATA)
1414 zone = HAMMER_ZONE_DECODE(leaf->data_offset);
1415 if (zone != HAMMER_ZONE_LARGE_DATA_INDEX)
1417 iinfo.obj_id = leaf->base.obj_id;
1418 iinfo.obj_asof = 0; /* unused */
1419 iinfo.obj_localization = leaf->base.localization &
1420 HAMMER_LOCALIZE_PSEUDOFS_MASK;
1421 iinfo.u.leaf = leaf;
1422 hammer_scan_inode_snapshots(hmp, &iinfo,
1423 hammer_io_direct_uncache_callback,
1428 hammer_io_direct_uncache_callback(hammer_inode_t ip, void *data)
1430 hammer_inode_info_t iinfo = data;
1431 hammer_off_t data_offset;
1432 hammer_off_t file_offset;
1439 data_offset = iinfo->u.leaf->data_offset;
1440 file_offset = iinfo->u.leaf->base.key - iinfo->u.leaf->data_len;
1441 blksize = iinfo->u.leaf->data_len;
1442 KKASSERT((blksize & HAMMER_BUFMASK) == 0);
1444 hammer_ref(&ip->lock);
1445 if (hammer_get_vnode(ip, &vp) == 0) {
1446 if ((bp = findblk(ip->vp, file_offset, FINDBLK_TEST)) != NULL &&
1447 bp->b_bio2.bio_offset != NOOFFSET) {
1448 bp = getblk(ip->vp, file_offset, blksize, 0, 0);
1449 bp->b_bio2.bio_offset = NOOFFSET;
1454 hammer_rel_inode(ip, 0);
1460 * This function is called when writes may have occured on the volume,
1461 * indicating that the device may be holding cached writes.
1464 hammer_io_flush_mark(hammer_volume_t volume)
1466 volume->vol_flags |= HAMMER_VOLF_NEEDFLUSH;
1470 * This function ensures that the device has flushed any cached writes out.
1473 hammer_io_flush_sync(hammer_mount_t hmp)
1475 hammer_volume_t volume;
1476 struct buf *bp_base = NULL;
1479 RB_FOREACH(volume, hammer_vol_rb_tree, &hmp->rb_vols_root) {
1480 if (volume->vol_flags & HAMMER_VOLF_NEEDFLUSH) {
1481 volume->vol_flags &= ~HAMMER_VOLF_NEEDFLUSH;
1483 bp->b_bio1.bio_offset = 0;
1486 bp->b_cmd = BUF_CMD_FLUSH;
1487 bp->b_bio1.bio_caller_info1.cluster_head = bp_base;
1488 bp->b_bio1.bio_done = biodone_sync;
1489 bp->b_bio1.bio_flags |= BIO_SYNC;
1491 vn_strategy(volume->devvp, &bp->b_bio1);
1494 while ((bp = bp_base) != NULL) {
1495 bp_base = bp->b_bio1.bio_caller_info1.cluster_head;
1496 biowait(&bp->b_bio1, "hmrFLS");