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_ondisk.c,v 1.76 2008/08/29 20:19:08 dillon Exp $
37 * Manage HAMMER's on-disk structures. These routines are primarily
38 * responsible for interfacing with the kernel's I/O subsystem and for
39 * managing in-memory structures.
43 #include <sys/fcntl.h>
44 #include <sys/nlookup.h>
49 static void hammer_free_volume(hammer_volume_t volume);
50 static int hammer_load_volume(hammer_volume_t volume);
51 static int hammer_load_buffer(hammer_buffer_t buffer, int isnew);
52 static int hammer_load_node(hammer_transaction_t trans,
53 hammer_node_t node, int isnew);
54 static void _hammer_rel_node(hammer_node_t node, int locked);
57 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
59 if (vol1->vol_no < vol2->vol_no)
61 if (vol1->vol_no > vol2->vol_no)
67 * hammer_buffer structures are indexed via their zoneX_offset, not
71 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
73 if (buf1->zoneX_offset < buf2->zoneX_offset)
75 if (buf1->zoneX_offset > buf2->zoneX_offset)
81 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
83 if (node1->node_offset < node2->node_offset)
85 if (node1->node_offset > node2->node_offset)
90 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
91 hammer_vol_rb_compare, int32_t, vol_no);
92 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
93 hammer_buf_rb_compare, hammer_off_t, zoneX_offset);
94 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
95 hammer_nod_rb_compare, hammer_off_t, node_offset);
97 /************************************************************************
99 ************************************************************************
101 * Load a HAMMER volume by name. Returns 0 on success or a positive error
102 * code on failure. Volumes must be loaded at mount time, get_volume() will
103 * not load a new volume.
105 * The passed devvp is vref()'d but not locked. This function consumes the
106 * ref (typically by associating it with the volume structure).
108 * Calls made to hammer_load_volume() or single-threaded
111 hammer_install_volume(struct hammer_mount *hmp, const char *volname,
115 hammer_volume_t volume;
116 struct hammer_volume_ondisk *ondisk;
117 struct nlookupdata nd;
118 struct buf *bp = NULL;
124 ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
127 * Allocate a volume structure
129 ++hammer_count_volumes;
130 volume = kmalloc(sizeof(*volume), hmp->m_misc, M_WAITOK|M_ZERO);
131 volume->vol_name = kstrdup(volname, hmp->m_misc);
132 volume->io.hmp = hmp; /* bootstrap */
133 hammer_io_init(&volume->io, volume, HAMMER_STRUCTURE_VOLUME);
134 volume->io.offset = 0LL;
135 volume->io.bytes = HAMMER_BUFSIZE;
138 * Get the device vnode
141 error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
143 error = nlookup(&nd);
145 error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
149 volume->devvp = devvp;
153 if (vn_isdisk(volume->devvp, &error)) {
154 error = vfs_mountedon(volume->devvp);
157 if (error == 0 && vcount(volume->devvp) > 0)
160 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
161 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
163 error = VOP_OPEN(volume->devvp,
164 (ronly ? FREAD : FREAD|FWRITE),
167 vn_unlock(volume->devvp);
170 hammer_free_volume(volume);
173 volume->devvp->v_rdev->si_mountpoint = mp;
177 * Extract the volume number from the volume header and do various
180 error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
183 ondisk = (void *)bp->b_data;
184 if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
185 kprintf("hammer_mount: volume %s has an invalid header\n",
190 volume->vol_no = ondisk->vol_no;
191 volume->buffer_base = ondisk->vol_buf_beg;
192 volume->vol_flags = ondisk->vol_flags;
193 volume->nblocks = ondisk->vol_nblocks;
194 volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
195 ondisk->vol_buf_end - ondisk->vol_buf_beg);
196 volume->maxraw_off = ondisk->vol_buf_end;
198 if (RB_EMPTY(&hmp->rb_vols_root)) {
199 hmp->fsid = ondisk->vol_fsid;
200 } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
201 kprintf("hammer_mount: volume %s's fsid does not match "
202 "other volumes\n", volume->vol_name);
208 * Insert the volume structure into the red-black tree.
210 if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
211 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
212 volume->vol_name, volume->vol_no);
217 * Set the root volume . HAMMER special cases rootvol the structure.
218 * We do not hold a ref because this would prevent related I/O
219 * from being flushed.
221 if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
222 hmp->rootvol = volume;
223 hmp->nvolumes = ondisk->vol_count;
228 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
229 (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
230 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
231 (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
237 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
239 volume->devvp->v_rdev->si_mountpoint = NULL;
240 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
241 hammer_free_volume(volume);
247 * This is called for each volume when updating the mount point from
248 * read-write to read-only or vise-versa.
251 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
254 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
255 if (volume->io.hmp->ronly) {
256 /* do not call vinvalbuf */
257 VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
258 VOP_CLOSE(volume->devvp, FREAD|FWRITE);
260 /* do not call vinvalbuf */
261 VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
262 VOP_CLOSE(volume->devvp, FREAD);
264 vn_unlock(volume->devvp);
270 * Unload and free a HAMMER volume. Must return >= 0 to continue scan
271 * so returns -1 on failure.
274 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
276 hammer_mount_t hmp = volume->io.hmp;
277 int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
280 * Clean up the root volume pointer, which is held unlocked in hmp.
282 if (hmp->rootvol == volume)
286 * We must not flush a dirty buffer to disk on umount. It should
287 * have already been dealt with by the flusher, or we may be in
288 * catastrophic failure.
290 hammer_io_clear_modify(&volume->io, 1);
291 volume->io.waitdep = 1;
294 * Clean up the persistent ref ioerror might have on the volume
296 if (volume->io.ioerror)
297 hammer_io_clear_error_noassert(&volume->io);
300 * This should release the bp. Releasing the volume with flush set
301 * implies the interlock is set.
303 hammer_ref_interlock_true(&volume->io.lock);
304 hammer_rel_volume(volume, 1);
305 KKASSERT(volume->io.bp == NULL);
308 * There should be no references on the volume, no clusters, and
311 KKASSERT(hammer_norefs(&volume->io.lock));
313 volume->ondisk = NULL;
315 if (volume->devvp->v_rdev &&
316 volume->devvp->v_rdev->si_mountpoint == hmp->mp
318 volume->devvp->v_rdev->si_mountpoint = NULL;
322 * Make sure we don't sync anything to disk if we
323 * are in read-only mode (1) or critically-errored
324 * (2). Note that there may be dirty buffers in
325 * normal read-only mode from crash recovery.
327 vinvalbuf(volume->devvp, 0, 0, 0);
328 VOP_CLOSE(volume->devvp, FREAD);
331 * Normal termination, save any dirty buffers
332 * (XXX there really shouldn't be any).
334 vinvalbuf(volume->devvp, V_SAVE, 0, 0);
335 VOP_CLOSE(volume->devvp, FREAD|FWRITE);
340 * Destroy the structure
342 RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
343 hammer_free_volume(volume);
349 hammer_free_volume(hammer_volume_t volume)
351 hammer_mount_t hmp = volume->io.hmp;
353 if (volume->vol_name) {
354 kfree(volume->vol_name, hmp->m_misc);
355 volume->vol_name = NULL;
358 vrele(volume->devvp);
359 volume->devvp = NULL;
361 --hammer_count_volumes;
362 kfree(volume, hmp->m_misc);
366 * Get a HAMMER volume. The volume must already exist.
369 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
371 struct hammer_volume *volume;
374 * Locate the volume structure
376 volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
377 if (volume == NULL) {
383 * Reference the volume, load/check the data on the 0->1 transition.
384 * hammer_load_volume() will dispose of the interlock on return,
385 * and also clean up the ref count on error.
387 if (hammer_ref_interlock(&volume->io.lock)) {
388 *errorp = hammer_load_volume(volume);
392 KKASSERT(volume->ondisk);
399 hammer_ref_volume(hammer_volume_t volume)
404 * Reference the volume and deal with the check condition used to
405 * load its ondisk info.
407 if (hammer_ref_interlock(&volume->io.lock)) {
408 error = hammer_load_volume(volume);
410 KKASSERT(volume->ondisk);
417 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
419 hammer_volume_t volume;
421 volume = hmp->rootvol;
422 KKASSERT(volume != NULL);
425 * Reference the volume and deal with the check condition used to
426 * load its ondisk info.
428 if (hammer_ref_interlock(&volume->io.lock)) {
429 *errorp = hammer_load_volume(volume);
433 KKASSERT(volume->ondisk);
440 * Load a volume's on-disk information. The volume must be referenced and
441 * the interlock is held on call. The interlock will be released on return.
442 * The reference will also be released on return if an error occurs.
445 hammer_load_volume(hammer_volume_t volume)
449 if (volume->ondisk == NULL) {
450 error = hammer_io_read(volume->devvp, &volume->io,
453 volume->ondisk = (void *)volume->io.bp->b_data;
454 hammer_ref_interlock_done(&volume->io.lock);
456 hammer_rel_volume(volume, 1);
465 * Release a previously acquired reference on the volume.
467 * Volumes are not unloaded from memory during normal operation.
470 hammer_rel_volume(hammer_volume_t volume, int locked)
474 if (hammer_rel_interlock(&volume->io.lock, locked)) {
475 volume->ondisk = NULL;
476 bp = hammer_io_release(&volume->io, locked);
477 hammer_rel_interlock_done(&volume->io.lock, locked);
484 hammer_mountcheck_volumes(struct hammer_mount *hmp)
489 for (i = 0; i < hmp->nvolumes; ++i) {
490 vol = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, i);
497 /************************************************************************
499 ************************************************************************
501 * Manage buffers. Currently most blockmap-backed zones are direct-mapped
502 * to zone-2 buffer offsets, without a translation stage. However, the
503 * hammer_buffer structure is indexed by its zoneX_offset, not its
506 * The proper zone must be maintained throughout the code-base all the way
507 * through to the big-block allocator, or routines like hammer_del_buffers()
508 * will not be able to locate all potentially conflicting buffers.
512 * Helper function returns whether a zone offset can be directly translated
513 * to a raw buffer index or not. Really only the volume and undo zones
514 * can't be directly translated. Volumes are special-cased and undo zones
515 * shouldn't be aliased accessed in read-only mode.
517 * This function is ONLY used to detect aliased zones during a read-only
521 hammer_direct_zone(hammer_off_t buf_offset)
523 switch(HAMMER_ZONE_DECODE(buf_offset)) {
524 case HAMMER_ZONE_RAW_BUFFER_INDEX:
525 case HAMMER_ZONE_FREEMAP_INDEX:
526 case HAMMER_ZONE_BTREE_INDEX:
527 case HAMMER_ZONE_META_INDEX:
528 case HAMMER_ZONE_LARGE_DATA_INDEX:
529 case HAMMER_ZONE_SMALL_DATA_INDEX:
538 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
539 int bytes, int isnew, int *errorp)
541 hammer_buffer_t buffer;
542 hammer_volume_t volume;
543 hammer_off_t zone2_offset;
544 hammer_io_type_t iotype;
548 buf_offset &= ~HAMMER_BUFMASK64;
551 * Shortcut if the buffer is already cached
553 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
556 * Once refed the ondisk field will not be cleared by
557 * any other action. Shortcut the operation if the
558 * ondisk structure is valid.
561 if (hammer_ref_interlock(&buffer->io.lock) == 0) {
562 hammer_io_advance(&buffer->io);
563 KKASSERT(buffer->ondisk);
569 * 0->1 transition or defered 0->1 transition (CHECK),
570 * interlock now held. Shortcut if ondisk is already
573 atomic_add_int(&hammer_count_refedbufs, 1);
574 if (buffer->ondisk) {
575 hammer_io_advance(&buffer->io);
576 hammer_ref_interlock_done(&buffer->io.lock);
582 * The buffer is no longer loose if it has a ref, and
583 * cannot become loose once it gains a ref. Loose
584 * buffers will never be in a modified state. This should
585 * only occur on the 0->1 transition of refs.
587 * lose_list can be modified via a biodone() interrupt
588 * so the io_token must be held.
590 if (buffer->io.mod_root == &hmp->lose_root) {
591 lwkt_gettoken(&hmp->io_token);
592 if (buffer->io.mod_root == &hmp->lose_root) {
593 RB_REMOVE(hammer_mod_rb_tree,
594 buffer->io.mod_root, &buffer->io);
595 buffer->io.mod_root = NULL;
596 KKASSERT(buffer->io.modified == 0);
598 lwkt_reltoken(&hmp->io_token);
601 } else if (hmp->ronly && hammer_direct_zone(buf_offset)) {
603 * If this is a read-only mount there could be an alias
604 * in the raw-zone. If there is we use that buffer instead.
606 * rw mounts will not have aliases. Also note when going
607 * from ro -> rw the recovered raw buffers are flushed and
608 * reclaimed, so again there will not be any aliases once
611 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
612 (buf_offset & ~HAMMER_OFF_ZONE_MASK) |
613 HAMMER_ZONE_RAW_BUFFER);
615 kprintf("HAMMER: recovered aliased %016jx\n",
616 (intmax_t)buf_offset);
622 * What is the buffer class?
624 zone = HAMMER_ZONE_DECODE(buf_offset);
627 case HAMMER_ZONE_LARGE_DATA_INDEX:
628 case HAMMER_ZONE_SMALL_DATA_INDEX:
629 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
631 case HAMMER_ZONE_UNDO_INDEX:
632 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
634 case HAMMER_ZONE_META_INDEX:
637 * NOTE: inode data and directory entries are placed in this
638 * zone. inode atime/mtime is updated in-place and thus
639 * buffers containing inodes must be synchronized as
640 * meta-buffers, same as buffers containing B-Tree info.
642 iotype = HAMMER_STRUCTURE_META_BUFFER;
647 * Handle blockmap offset translations
649 if (zone >= HAMMER_ZONE_BTREE_INDEX) {
650 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
651 } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
652 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
654 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
655 zone2_offset = buf_offset;
662 * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
665 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
666 HAMMER_ZONE_RAW_BUFFER);
667 vol_no = HAMMER_VOL_DECODE(zone2_offset);
668 volume = hammer_get_volume(hmp, vol_no, errorp);
672 KKASSERT(zone2_offset < volume->maxbuf_off);
675 * Allocate a new buffer structure. We will check for races later.
677 ++hammer_count_buffers;
678 buffer = kmalloc(sizeof(*buffer), hmp->m_misc,
679 M_WAITOK|M_ZERO|M_USE_RESERVE);
680 buffer->zone2_offset = zone2_offset;
681 buffer->zoneX_offset = buf_offset;
683 hammer_io_init(&buffer->io, volume, iotype);
684 buffer->io.offset = volume->ondisk->vol_buf_beg +
685 (zone2_offset & HAMMER_OFF_SHORT_MASK);
686 buffer->io.bytes = bytes;
687 TAILQ_INIT(&buffer->clist);
688 hammer_ref_interlock_true(&buffer->io.lock);
691 * Insert the buffer into the RB tree and handle late collisions.
693 if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
694 hammer_rel_volume(volume, 0);
695 buffer->io.volume = NULL; /* safety */
696 if (hammer_rel_interlock(&buffer->io.lock, 1)) /* safety */
697 hammer_rel_interlock_done(&buffer->io.lock, 1);
698 --hammer_count_buffers;
699 kfree(buffer, hmp->m_misc);
702 atomic_add_int(&hammer_count_refedbufs, 1);
706 * The buffer is referenced and interlocked. Load the buffer
707 * if necessary. hammer_load_buffer() deals with the interlock
708 * and, if an error is returned, also deals with the ref.
710 if (buffer->ondisk == NULL) {
711 *errorp = hammer_load_buffer(buffer, isnew);
715 hammer_io_advance(&buffer->io);
716 hammer_ref_interlock_done(&buffer->io.lock);
723 * This is used by the direct-read code to deal with large-data buffers
724 * created by the reblocker and mirror-write code. The direct-read code
725 * bypasses the HAMMER buffer subsystem and so any aliased dirty or write-
726 * running hammer buffers must be fully synced to disk before we can issue
729 * This code path is not considered critical as only the rebocker and
730 * mirror-write code will create large-data buffers via the HAMMER buffer
731 * subsystem. They do that because they operate at the B-Tree level and
732 * do not access the vnode/inode structures.
735 hammer_sync_buffers(hammer_mount_t hmp, hammer_off_t base_offset, int bytes)
737 hammer_buffer_t buffer;
740 KKASSERT((base_offset & HAMMER_OFF_ZONE_MASK) ==
741 HAMMER_ZONE_LARGE_DATA);
744 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
746 if (buffer && (buffer->io.modified || buffer->io.running)) {
747 error = hammer_ref_buffer(buffer);
749 hammer_io_wait(&buffer->io);
750 if (buffer->io.modified) {
751 hammer_io_write_interlock(&buffer->io);
752 hammer_io_flush(&buffer->io, 0);
753 hammer_io_done_interlock(&buffer->io);
754 hammer_io_wait(&buffer->io);
756 hammer_rel_buffer(buffer, 0);
759 base_offset += HAMMER_BUFSIZE;
760 bytes -= HAMMER_BUFSIZE;
765 * Destroy all buffers covering the specified zoneX offset range. This
766 * is called when the related blockmap layer2 entry is freed or when
767 * a direct write bypasses our buffer/buffer-cache subsystem.
769 * The buffers may be referenced by the caller itself. Setting reclaim
770 * will cause the buffer to be destroyed when it's ref count reaches zero.
772 * Return 0 on success, EAGAIN if some buffers could not be destroyed due
773 * to additional references held by other threads, or some other (typically
777 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
778 hammer_off_t zone2_offset, int bytes,
779 int report_conflicts)
781 hammer_buffer_t buffer;
782 hammer_volume_t volume;
787 vol_no = HAMMER_VOL_DECODE(zone2_offset);
788 volume = hammer_get_volume(hmp, vol_no, &ret_error);
789 KKASSERT(ret_error == 0);
792 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
795 error = hammer_ref_buffer(buffer);
796 if (hammer_debug_general & 0x20000) {
797 kprintf("hammer: delbufr %016jx "
799 (intmax_t)buffer->zoneX_offset,
801 hammer_oneref(&buffer->io.lock));
803 if (error == 0 && !hammer_oneref(&buffer->io.lock)) {
805 hammer_rel_buffer(buffer, 0);
808 KKASSERT(buffer->zone2_offset == zone2_offset);
809 hammer_io_clear_modify(&buffer->io, 1);
810 buffer->io.reclaim = 1;
811 buffer->io.waitdep = 1;
812 KKASSERT(buffer->io.volume == volume);
813 hammer_rel_buffer(buffer, 0);
816 error = hammer_io_inval(volume, zone2_offset);
820 if (report_conflicts ||
821 (hammer_debug_general & 0x8000)) {
822 kprintf("hammer_del_buffers: unable to "
823 "invalidate %016llx buffer=%p rep=%d\n",
824 (long long)base_offset,
825 buffer, report_conflicts);
828 base_offset += HAMMER_BUFSIZE;
829 zone2_offset += HAMMER_BUFSIZE;
830 bytes -= HAMMER_BUFSIZE;
832 hammer_rel_volume(volume, 0);
837 * Given a referenced and interlocked buffer load/validate the data.
839 * The buffer interlock will be released on return. If an error is
840 * returned the buffer reference will also be released (and the buffer
841 * pointer will thus be stale).
844 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
846 hammer_volume_t volume;
850 * Load the buffer's on-disk info
852 volume = buffer->io.volume;
854 if (hammer_debug_io & 0x0004) {
855 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
856 (long long)buffer->zoneX_offset,
857 (long long)buffer->zone2_offset,
858 isnew, buffer->ondisk);
861 if (buffer->ondisk == NULL) {
863 * Issue the read or generate a new buffer. When reading
864 * the limit argument controls any read-ahead clustering
865 * hammer_io_read() is allowed to do.
867 * We cannot read-ahead in the large-data zone and we cannot
868 * cross a largeblock boundary as the next largeblock might
869 * use a different buffer size.
872 error = hammer_io_new(volume->devvp, &buffer->io);
873 } else if ((buffer->zoneX_offset & HAMMER_OFF_ZONE_MASK) ==
874 HAMMER_ZONE_LARGE_DATA) {
875 error = hammer_io_read(volume->devvp, &buffer->io,
880 limit = (buffer->zone2_offset +
881 HAMMER_LARGEBLOCK_MASK64) &
882 ~HAMMER_LARGEBLOCK_MASK64;
883 limit -= buffer->zone2_offset;
884 error = hammer_io_read(volume->devvp, &buffer->io,
888 buffer->ondisk = (void *)buffer->io.bp->b_data;
890 error = hammer_io_new(volume->devvp, &buffer->io);
895 hammer_io_advance(&buffer->io);
896 hammer_ref_interlock_done(&buffer->io.lock);
898 hammer_rel_buffer(buffer, 1);
904 * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
905 * This routine is only called during unmount or when a volume is
908 * If data != NULL, it specifies a volume whoose buffers should
912 hammer_unload_buffer(hammer_buffer_t buffer, void *data)
914 struct hammer_volume *volume = (struct hammer_volume *) data;
917 * If volume != NULL we are only interested in unloading buffers
918 * associated with a particular volume.
920 if (volume != NULL && volume != buffer->io.volume)
924 * Clean up the persistent ref ioerror might have on the buffer
925 * and acquire a ref. Expect a 0->1 transition.
927 if (buffer->io.ioerror) {
928 hammer_io_clear_error_noassert(&buffer->io);
929 atomic_add_int(&hammer_count_refedbufs, -1);
931 hammer_ref_interlock_true(&buffer->io.lock);
932 atomic_add_int(&hammer_count_refedbufs, 1);
935 * We must not flush a dirty buffer to disk on umount. It should
936 * have already been dealt with by the flusher, or we may be in
937 * catastrophic failure.
939 * We must set waitdep to ensure that a running buffer is waited
940 * on and released prior to us trying to unload the volume.
942 hammer_io_clear_modify(&buffer->io, 1);
943 hammer_flush_buffer_nodes(buffer);
944 buffer->io.waitdep = 1;
945 hammer_rel_buffer(buffer, 1);
950 * Reference a buffer that is either already referenced or via a specially
951 * handled pointer (aka cursor->buffer).
954 hammer_ref_buffer(hammer_buffer_t buffer)
961 * Acquire a ref, plus the buffer will be interlocked on the
964 locked = hammer_ref_interlock(&buffer->io.lock);
965 hmp = buffer->io.hmp;
968 * At this point a biodone() will not touch the buffer other then
969 * incidental bits. However, lose_list can be modified via
970 * a biodone() interrupt.
972 * No longer loose. lose_list requires the io_token.
974 if (buffer->io.mod_root == &hmp->lose_root) {
975 lwkt_gettoken(&hmp->io_token);
976 if (buffer->io.mod_root == &hmp->lose_root) {
977 RB_REMOVE(hammer_mod_rb_tree,
978 buffer->io.mod_root, &buffer->io);
979 buffer->io.mod_root = NULL;
981 lwkt_reltoken(&hmp->io_token);
985 atomic_add_int(&hammer_count_refedbufs, 1);
986 error = hammer_load_buffer(buffer, 0);
987 /* NOTE: on error the buffer pointer is stale */
995 * Release a reference on the buffer. On the 1->0 transition the
996 * underlying IO will be released but the data reference is left
999 * Only destroy the structure itself if the related buffer cache buffer
1000 * was disassociated from it. This ties the management of the structure
1001 * to the buffer cache subsystem. buffer->ondisk determines whether the
1002 * embedded io is referenced or not.
1005 hammer_rel_buffer(hammer_buffer_t buffer, int locked)
1007 hammer_volume_t volume;
1009 struct buf *bp = NULL;
1012 hmp = buffer->io.hmp;
1014 if (hammer_rel_interlock(&buffer->io.lock, locked) == 0)
1018 * hammer_count_refedbufs accounting. Decrement if we are in
1019 * the error path or if CHECK is clear.
1021 * If we are not in the error path and CHECK is set the caller
1022 * probably just did a hammer_ref() and didn't account for it,
1023 * so we don't account for the loss here.
1025 if (locked || (buffer->io.lock.refs & HAMMER_REFS_CHECK) == 0)
1026 atomic_add_int(&hammer_count_refedbufs, -1);
1029 * If the caller locked us or the normal released transitions
1030 * from 1->0 (and acquired the lock) attempt to release the
1031 * io. If the called locked us we tell hammer_io_release()
1032 * to flush (which would be the unload or failure path).
1034 bp = hammer_io_release(&buffer->io, locked);
1037 * If the buffer has no bp association and no refs we can destroy
1040 * NOTE: It is impossible for any associated B-Tree nodes to have
1041 * refs if the buffer has no additional refs.
1043 if (buffer->io.bp == NULL && hammer_norefs(&buffer->io.lock)) {
1044 RB_REMOVE(hammer_buf_rb_tree,
1045 &buffer->io.hmp->rb_bufs_root,
1047 volume = buffer->io.volume;
1048 buffer->io.volume = NULL; /* sanity */
1049 hammer_rel_volume(volume, 0);
1050 hammer_io_clear_modlist(&buffer->io);
1051 hammer_flush_buffer_nodes(buffer);
1052 KKASSERT(TAILQ_EMPTY(&buffer->clist));
1059 hammer_rel_interlock_done(&buffer->io.lock, locked);
1063 --hammer_count_buffers;
1064 kfree(buffer, hmp->m_misc);
1069 * Access the filesystem buffer containing the specified hammer offset.
1070 * buf_offset is a conglomeration of the volume number and vol_buf_beg
1071 * relative buffer offset. It must also have bit 55 set to be valid.
1072 * (see hammer_off_t in hammer_disk.h).
1074 * Any prior buffer in *bufferp will be released and replaced by the
1077 * NOTE: The buffer is indexed via its zoneX_offset but we allow the
1078 * passed cached *bufferp to match against either zoneX or zone2.
1082 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1083 int *errorp, struct hammer_buffer **bufferp)
1085 hammer_buffer_t buffer;
1086 int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
1088 buf_offset &= ~HAMMER_BUFMASK64;
1089 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
1092 if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
1093 buffer->zoneX_offset != buf_offset)) {
1095 hammer_rel_buffer(buffer, 0);
1096 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
1103 * Return a pointer to the buffer data.
1108 return((char *)buffer->ondisk + xoff);
1112 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
1113 int *errorp, struct hammer_buffer **bufferp)
1115 return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
1119 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1120 int *errorp, struct hammer_buffer **bufferp)
1122 bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1123 return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
1127 * Access the filesystem buffer containing the specified hammer offset.
1128 * No disk read operation occurs. The result buffer may contain garbage.
1130 * Any prior buffer in *bufferp will be released and replaced by the
1133 * This function marks the buffer dirty but does not increment its
1134 * modify_refs count.
1138 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1139 int *errorp, struct hammer_buffer **bufferp)
1141 hammer_buffer_t buffer;
1142 int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
1144 buf_offset &= ~HAMMER_BUFMASK64;
1147 if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
1148 buffer->zoneX_offset != buf_offset)) {
1150 hammer_rel_buffer(buffer, 0);
1151 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
1158 * Return a pointer to the buffer data.
1163 return((char *)buffer->ondisk + xoff);
1167 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
1168 int *errorp, struct hammer_buffer **bufferp)
1170 return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
1174 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1175 int *errorp, struct hammer_buffer **bufferp)
1177 bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1178 return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
1181 /************************************************************************
1183 ************************************************************************
1185 * Manage B-Tree nodes. B-Tree nodes represent the primary indexing
1186 * method used by the HAMMER filesystem.
1188 * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
1189 * associated with its buffer, and will only referenced the buffer while
1190 * the node itself is referenced.
1192 * A hammer_node can also be passively associated with other HAMMER
1193 * structures, such as inodes, while retaining 0 references. These
1194 * associations can be cleared backwards using a pointer-to-pointer in
1197 * This allows the HAMMER implementation to cache hammer_nodes long-term
1198 * and short-cut a great deal of the infrastructure's complexity. In
1199 * most cases a cached node can be reacquired without having to dip into
1200 * either the buffer or cluster management code.
1202 * The caller must pass a referenced cluster on call and will retain
1203 * ownership of the reference on return. The node will acquire its own
1204 * additional references, if necessary.
1207 hammer_get_node(hammer_transaction_t trans, hammer_off_t node_offset,
1208 int isnew, int *errorp)
1210 hammer_mount_t hmp = trans->hmp;
1214 KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
1217 * Locate the structure, allocating one if necessary.
1220 node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
1222 ++hammer_count_nodes;
1223 node = kmalloc(sizeof(*node), hmp->m_misc, M_WAITOK|M_ZERO|M_USE_RESERVE);
1224 node->node_offset = node_offset;
1226 TAILQ_INIT(&node->cursor_list);
1227 TAILQ_INIT(&node->cache_list);
1228 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
1229 --hammer_count_nodes;
1230 kfree(node, hmp->m_misc);
1233 doload = hammer_ref_interlock_true(&node->lock);
1235 doload = hammer_ref_interlock(&node->lock);
1238 *errorp = hammer_load_node(trans, node, isnew);
1239 trans->flags |= HAMMER_TRANSF_DIDIO;
1243 KKASSERT(node->ondisk);
1245 hammer_io_advance(&node->buffer->io);
1251 * Reference an already-referenced node. 0->1 transitions should assert
1252 * so we do not have to deal with hammer_ref() setting CHECK.
1255 hammer_ref_node(hammer_node_t node)
1257 KKASSERT(hammer_isactive(&node->lock) && node->ondisk != NULL);
1258 hammer_ref(&node->lock);
1262 * Load a node's on-disk data reference. Called with the node referenced
1265 * On return the node interlock will be unlocked. If a non-zero error code
1266 * is returned the node will also be dereferenced (and the caller's pointer
1270 hammer_load_node(hammer_transaction_t trans, hammer_node_t node, int isnew)
1272 hammer_buffer_t buffer;
1273 hammer_off_t buf_offset;
1277 if (node->ondisk == NULL) {
1279 * This is a little confusing but the jist is that
1280 * node->buffer determines whether the node is on
1281 * the buffer's clist and node->ondisk determines
1282 * whether the buffer is referenced.
1284 * We could be racing a buffer release, in which case
1285 * node->buffer may become NULL while we are blocked
1286 * referencing the buffer.
1288 if ((buffer = node->buffer) != NULL) {
1289 error = hammer_ref_buffer(buffer);
1290 if (error == 0 && node->buffer == NULL) {
1291 TAILQ_INSERT_TAIL(&buffer->clist,
1293 node->buffer = buffer;
1296 buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1297 buffer = hammer_get_buffer(node->hmp, buf_offset,
1298 HAMMER_BUFSIZE, 0, &error);
1300 KKASSERT(error == 0);
1301 TAILQ_INSERT_TAIL(&buffer->clist,
1303 node->buffer = buffer;
1308 node->ondisk = (void *)((char *)buffer->ondisk +
1309 (node->node_offset & HAMMER_BUFMASK));
1312 * Check CRC. NOTE: Neither flag is set and the CRC is not
1313 * generated on new B-Tree nodes.
1316 (node->flags & HAMMER_NODE_CRCANY) == 0) {
1317 if (hammer_crc_test_btree(node->ondisk) == 0) {
1318 if (hammer_debug_critical)
1319 Debugger("CRC FAILED: B-TREE NODE");
1320 node->flags |= HAMMER_NODE_CRCBAD;
1322 node->flags |= HAMMER_NODE_CRCGOOD;
1326 if (node->flags & HAMMER_NODE_CRCBAD) {
1327 if (trans->flags & HAMMER_TRANSF_CRCDOM)
1334 _hammer_rel_node(node, 1);
1336 hammer_ref_interlock_done(&node->lock);
1342 * Safely reference a node, interlock against flushes via the IO subsystem.
1345 hammer_ref_node_safe(hammer_transaction_t trans, hammer_node_cache_t cache,
1353 doload = hammer_ref_interlock(&node->lock);
1355 *errorp = hammer_load_node(trans, node, 0);
1359 KKASSERT(node->ondisk);
1360 if (node->flags & HAMMER_NODE_CRCBAD) {
1361 if (trans->flags & HAMMER_TRANSF_CRCDOM)
1365 _hammer_rel_node(node, 0);
1378 * Release a hammer_node. On the last release the node dereferences
1379 * its underlying buffer and may or may not be destroyed.
1381 * If locked is non-zero the passed node has been interlocked by the
1382 * caller and we are in the failure/unload path, otherwise it has not and
1383 * we are doing a normal release.
1385 * This function will dispose of the interlock and the reference.
1386 * On return the node pointer is stale.
1389 _hammer_rel_node(hammer_node_t node, int locked)
1391 hammer_buffer_t buffer;
1394 * Deref the node. If this isn't the 1->0 transition we're basically
1395 * done. If locked is non-zero this function will just deref the
1396 * locked node and return TRUE, otherwise it will deref the locked
1397 * node and either lock and return TRUE on the 1->0 transition or
1398 * not lock and return FALSE.
1400 if (hammer_rel_interlock(&node->lock, locked) == 0)
1404 * Either locked was non-zero and we are interlocked, or the
1405 * hammer_rel_interlock() call returned non-zero and we are
1408 * The ref-count must still be decremented if locked != 0 so
1409 * the cleanup required still varies a bit.
1411 * hammer_flush_node() when called with 1 or 2 will dispose of
1412 * the lock and possible ref-count.
1414 if (node->ondisk == NULL) {
1415 hammer_flush_node(node, locked + 1);
1416 /* node is stale now */
1421 * Do not disassociate the node from the buffer if it represents
1422 * a modified B-Tree node that still needs its crc to be generated.
1424 if (node->flags & HAMMER_NODE_NEEDSCRC) {
1425 hammer_rel_interlock_done(&node->lock, locked);
1430 * Do final cleanups and then either destroy the node and leave it
1431 * passively cached. The buffer reference is removed regardless.
1433 buffer = node->buffer;
1434 node->ondisk = NULL;
1436 if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1440 hammer_rel_interlock_done(&node->lock, locked);
1445 hammer_flush_node(node, locked + 1);
1449 hammer_rel_buffer(buffer, 0);
1453 hammer_rel_node(hammer_node_t node)
1455 _hammer_rel_node(node, 0);
1459 * Free space on-media associated with a B-Tree node.
1462 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1464 KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1465 node->flags |= HAMMER_NODE_DELETED;
1466 hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1470 * Passively cache a referenced hammer_node. The caller may release
1471 * the node on return.
1474 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1477 * If the node doesn't exist, or is being deleted, don't cache it!
1479 * The node can only ever be NULL in the I/O failure path.
1481 if (node == NULL || (node->flags & HAMMER_NODE_DELETED))
1483 if (cache->node == node)
1486 hammer_uncache_node(cache);
1487 if (node->flags & HAMMER_NODE_DELETED)
1490 TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1494 hammer_uncache_node(hammer_node_cache_t cache)
1498 if ((node = cache->node) != NULL) {
1499 TAILQ_REMOVE(&node->cache_list, cache, entry);
1501 if (TAILQ_EMPTY(&node->cache_list))
1502 hammer_flush_node(node, 0);
1507 * Remove a node's cache references and destroy the node if it has no
1508 * other references or backing store.
1510 * locked == 0 Normal unlocked operation
1511 * locked == 1 Call hammer_rel_interlock_done(..., 0);
1512 * locked == 2 Call hammer_rel_interlock_done(..., 1);
1514 * XXX for now this isn't even close to being MPSAFE so the refs check
1518 hammer_flush_node(hammer_node_t node, int locked)
1520 hammer_node_cache_t cache;
1521 hammer_buffer_t buffer;
1522 hammer_mount_t hmp = node->hmp;
1525 while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1526 TAILQ_REMOVE(&node->cache_list, cache, entry);
1531 * NOTE: refs is predisposed if another thread is blocking and
1532 * will be larger than 0 in that case. We aren't MPSAFE
1535 if (node->ondisk == NULL && hammer_norefs(&node->lock)) {
1536 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1537 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1538 if ((buffer = node->buffer) != NULL) {
1539 node->buffer = NULL;
1540 TAILQ_REMOVE(&buffer->clist, node, entry);
1541 /* buffer is unreferenced because ondisk is NULL */
1549 * Deal with the interlock if locked == 1 or locked == 2.
1552 hammer_rel_interlock_done(&node->lock, locked - 1);
1555 * Destroy if requested
1558 --hammer_count_nodes;
1559 kfree(node, hmp->m_misc);
1564 * Flush passively cached B-Tree nodes associated with this buffer.
1565 * This is only called when the buffer is about to be destroyed, so
1566 * none of the nodes should have any references. The buffer is locked.
1568 * We may be interlocked with the buffer.
1571 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1575 while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1576 KKASSERT(node->ondisk == NULL);
1577 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1579 if (hammer_try_interlock_norefs(&node->lock)) {
1580 hammer_ref(&node->lock);
1581 node->flags |= HAMMER_NODE_FLUSH;
1582 _hammer_rel_node(node, 1);
1584 KKASSERT(node->buffer != NULL);
1585 buffer = node->buffer;
1586 node->buffer = NULL;
1587 TAILQ_REMOVE(&buffer->clist, node, entry);
1588 /* buffer is unreferenced because ondisk is NULL */
1594 /************************************************************************
1596 ************************************************************************/
1599 * Allocate a B-Tree node.
1602 hammer_alloc_btree(hammer_transaction_t trans, hammer_off_t hint, int *errorp)
1604 hammer_buffer_t buffer = NULL;
1605 hammer_node_t node = NULL;
1606 hammer_off_t node_offset;
1608 node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1609 sizeof(struct hammer_node_ondisk),
1612 node = hammer_get_node(trans, node_offset, 1, errorp);
1613 hammer_modify_node_noundo(trans, node);
1614 bzero(node->ondisk, sizeof(*node->ondisk));
1615 hammer_modify_node_done(node);
1618 hammer_rel_buffer(buffer, 0);
1623 * Allocate data. If the address of a data buffer is supplied then
1624 * any prior non-NULL *data_bufferp will be released and *data_bufferp
1625 * will be set to the related buffer. The caller must release it when
1626 * finally done. The initial *data_bufferp should be set to NULL by
1629 * The caller is responsible for making hammer_modify*() calls on the
1633 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1634 u_int16_t rec_type, hammer_off_t *data_offsetp,
1635 struct hammer_buffer **data_bufferp,
1636 hammer_off_t hint, int *errorp)
1646 case HAMMER_RECTYPE_INODE:
1647 case HAMMER_RECTYPE_DIRENTRY:
1648 case HAMMER_RECTYPE_EXT:
1649 case HAMMER_RECTYPE_FIX:
1650 case HAMMER_RECTYPE_PFS:
1651 case HAMMER_RECTYPE_SNAPSHOT:
1652 case HAMMER_RECTYPE_CONFIG:
1653 zone = HAMMER_ZONE_META_INDEX;
1655 case HAMMER_RECTYPE_DATA:
1656 case HAMMER_RECTYPE_DB:
1657 if (data_len <= HAMMER_BUFSIZE / 2) {
1658 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1660 data_len = (data_len + HAMMER_BUFMASK) &
1662 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1666 panic("hammer_alloc_data: rec_type %04x unknown",
1668 zone = 0; /* NOT REACHED */
1671 *data_offsetp = hammer_blockmap_alloc(trans, zone, data_len,
1676 if (*errorp == 0 && data_bufferp) {
1678 data = hammer_bread_ext(trans->hmp, *data_offsetp,
1679 data_len, errorp, data_bufferp);
1690 * Sync dirty buffers to the media and clean-up any loose ends.
1692 * These functions do not start the flusher going, they simply
1693 * queue everything up to the flusher.
1695 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1696 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1699 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1701 struct hammer_sync_info info;
1704 info.waitfor = waitfor;
1705 if (waitfor == MNT_WAIT) {
1706 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1707 hammer_sync_scan1, hammer_sync_scan2, &info);
1709 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1710 hammer_sync_scan1, hammer_sync_scan2, &info);
1716 * Filesystem sync. If doing a synchronous sync make a second pass on
1717 * the vnodes in case any were already flushing during the first pass,
1718 * and activate the flusher twice (the second time brings the UNDO FIFO's
1719 * start position up to the end position after the first call).
1721 * If doing a lazy sync make just one pass on the vnode list, ignoring
1722 * any new vnodes added to the list while the sync is in progress.
1725 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1727 struct hammer_sync_info info;
1731 if (waitfor & MNT_LAZY)
1732 flags |= VMSC_ONEPASS;
1735 info.waitfor = MNT_NOWAIT;
1736 vmntvnodescan(hmp->mp, flags | VMSC_NOWAIT,
1737 hammer_sync_scan1, hammer_sync_scan2, &info);
1739 if (info.error == 0 && (waitfor & MNT_WAIT)) {
1740 info.waitfor = waitfor;
1741 vmntvnodescan(hmp->mp, flags,
1742 hammer_sync_scan1, hammer_sync_scan2, &info);
1744 if (waitfor == MNT_WAIT) {
1745 hammer_flusher_sync(hmp);
1746 hammer_flusher_sync(hmp);
1748 hammer_flusher_async(hmp, NULL);
1749 hammer_flusher_async(hmp, NULL);
1755 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1757 struct hammer_inode *ip;
1760 if (vp->v_type == VNON || ip == NULL ||
1761 ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1762 RB_EMPTY(&vp->v_rbdirty_tree))) {
1769 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1771 struct hammer_sync_info *info = data;
1772 struct hammer_inode *ip;
1776 if (vp->v_type == VNON || vp->v_type == VBAD ||
1777 ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1778 RB_EMPTY(&vp->v_rbdirty_tree))) {
1781 error = VOP_FSYNC(vp, MNT_NOWAIT, 0);
1783 info->error = error;