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, hammer_get_volume()
103 * will 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_BIGBLOCK_SIZE / HAMMER_BUFSIZE);
230 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
231 (HAMMER_BIGBLOCK_SIZE / HAMMER_BUFSIZE);
237 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
239 volume->devvp->v_rdev->si_mountpoint = NULL;
240 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
241 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE, NULL);
242 vn_unlock(volume->devvp);
243 hammer_free_volume(volume);
249 * This is called for each volume when updating the mount point from
250 * read-write to read-only or vise-versa.
253 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
256 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
257 if (volume->io.hmp->ronly) {
258 /* do not call vinvalbuf */
259 VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
260 VOP_CLOSE(volume->devvp, FREAD|FWRITE, NULL);
262 /* do not call vinvalbuf */
263 VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
264 VOP_CLOSE(volume->devvp, FREAD, NULL);
266 vn_unlock(volume->devvp);
272 * Unload and free a HAMMER volume. Must return >= 0 to continue scan
273 * so returns -1 on failure.
276 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
278 hammer_mount_t hmp = volume->io.hmp;
279 int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
282 * Clean up the root volume pointer, which is held unlocked in hmp.
284 if (hmp->rootvol == volume)
288 * We must not flush a dirty buffer to disk on umount. It should
289 * have already been dealt with by the flusher, or we may be in
290 * catastrophic failure.
292 hammer_io_clear_modify(&volume->io, 1);
293 volume->io.waitdep = 1;
296 * Clean up the persistent ref ioerror might have on the volume
298 if (volume->io.ioerror)
299 hammer_io_clear_error_noassert(&volume->io);
302 * This should release the bp. Releasing the volume with flush set
303 * implies the interlock is set.
305 hammer_ref_interlock_true(&volume->io.lock);
306 hammer_rel_volume(volume, 1);
307 KKASSERT(volume->io.bp == NULL);
310 * There should be no references on the volume, no clusters, and
313 KKASSERT(hammer_norefs(&volume->io.lock));
315 volume->ondisk = NULL;
317 if (volume->devvp->v_rdev &&
318 volume->devvp->v_rdev->si_mountpoint == hmp->mp
320 volume->devvp->v_rdev->si_mountpoint = NULL;
324 * Make sure we don't sync anything to disk if we
325 * are in read-only mode (1) or critically-errored
326 * (2). Note that there may be dirty buffers in
327 * normal read-only mode from crash recovery.
329 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
330 vinvalbuf(volume->devvp, 0, 0, 0);
331 VOP_CLOSE(volume->devvp, FREAD, NULL);
332 vn_unlock(volume->devvp);
335 * Normal termination, save any dirty buffers
336 * (XXX there really shouldn't be any).
338 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
339 vinvalbuf(volume->devvp, V_SAVE, 0, 0);
340 VOP_CLOSE(volume->devvp, FREAD|FWRITE, NULL);
341 vn_unlock(volume->devvp);
346 * Destroy the structure
348 RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
349 hammer_free_volume(volume);
355 hammer_free_volume(hammer_volume_t volume)
357 hammer_mount_t hmp = volume->io.hmp;
359 if (volume->vol_name) {
360 kfree(volume->vol_name, hmp->m_misc);
361 volume->vol_name = NULL;
364 vrele(volume->devvp);
365 volume->devvp = NULL;
367 --hammer_count_volumes;
368 kfree(volume, hmp->m_misc);
372 * Get a HAMMER volume. The volume must already exist.
375 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
377 struct hammer_volume *volume;
380 * Locate the volume structure
382 volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
383 if (volume == NULL) {
389 * Reference the volume, load/check the data on the 0->1 transition.
390 * hammer_load_volume() will dispose of the interlock on return,
391 * and also clean up the ref count on error.
393 if (hammer_ref_interlock(&volume->io.lock)) {
394 *errorp = hammer_load_volume(volume);
398 KKASSERT(volume->ondisk);
405 hammer_ref_volume(hammer_volume_t volume)
410 * Reference the volume and deal with the check condition used to
411 * load its ondisk info.
413 if (hammer_ref_interlock(&volume->io.lock)) {
414 error = hammer_load_volume(volume);
416 KKASSERT(volume->ondisk);
423 * May be called without fs_token
426 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
428 hammer_volume_t volume;
430 volume = hmp->rootvol;
431 KKASSERT(volume != NULL);
434 * Reference the volume and deal with the check condition used to
435 * load its ondisk info.
437 if (hammer_ref_interlock(&volume->io.lock)) {
438 lwkt_gettoken(&volume->io.hmp->fs_token);
439 *errorp = hammer_load_volume(volume);
440 lwkt_reltoken(&volume->io.hmp->fs_token);
444 KKASSERT(volume->ondisk);
451 * Load a volume's on-disk information. The volume must be referenced and
452 * the interlock is held on call. The interlock will be released on return.
453 * The reference will also be released on return if an error occurs.
456 hammer_load_volume(hammer_volume_t volume)
460 if (volume->ondisk == NULL) {
461 error = hammer_io_read(volume->devvp, &volume->io,
464 volume->ondisk = (void *)volume->io.bp->b_data;
465 hammer_ref_interlock_done(&volume->io.lock);
467 hammer_rel_volume(volume, 1);
476 * Release a previously acquired reference on the volume.
478 * Volumes are not unloaded from memory during normal operation.
480 * May be called without fs_token
483 hammer_rel_volume(hammer_volume_t volume, int locked)
487 if (hammer_rel_interlock(&volume->io.lock, locked)) {
488 lwkt_gettoken(&volume->io.hmp->fs_token);
489 volume->ondisk = NULL;
490 bp = hammer_io_release(&volume->io, locked);
491 lwkt_reltoken(&volume->io.hmp->fs_token);
492 hammer_rel_interlock_done(&volume->io.lock, locked);
499 hammer_mountcheck_volumes(struct hammer_mount *hmp)
504 for (i = 0; i < hmp->nvolumes; ++i) {
505 vol = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, i);
512 /************************************************************************
514 ************************************************************************
516 * Manage buffers. Currently most blockmap-backed zones are direct-mapped
517 * to zone-2 buffer offsets, without a translation stage. However, the
518 * hammer_buffer structure is indexed by its zoneX_offset, not its
521 * The proper zone must be maintained throughout the code-base all the way
522 * through to the big-block allocator, or routines like hammer_del_buffers()
523 * will not be able to locate all potentially conflicting buffers.
527 * Helper function returns whether a zone offset can be directly translated
528 * to a raw buffer index or not. Really only the volume and undo zones
529 * can't be directly translated. Volumes are special-cased and undo zones
530 * shouldn't be aliased accessed in read-only mode.
532 * This function is ONLY used to detect aliased zones during a read-only
536 hammer_direct_zone(hammer_off_t buf_offset)
538 switch(HAMMER_ZONE_DECODE(buf_offset)) {
539 case HAMMER_ZONE_RAW_BUFFER_INDEX:
540 case HAMMER_ZONE_FREEMAP_INDEX:
541 case HAMMER_ZONE_BTREE_INDEX:
542 case HAMMER_ZONE_META_INDEX:
543 case HAMMER_ZONE_LARGE_DATA_INDEX:
544 case HAMMER_ZONE_SMALL_DATA_INDEX:
553 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
554 int bytes, int isnew, int *errorp)
556 hammer_buffer_t buffer;
557 hammer_volume_t volume;
558 hammer_off_t zone2_offset;
559 hammer_io_type_t iotype;
563 buf_offset &= ~HAMMER_BUFMASK64;
566 * Shortcut if the buffer is already cached
568 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
571 * Once refed the ondisk field will not be cleared by
572 * any other action. Shortcut the operation if the
573 * ondisk structure is valid.
576 if (hammer_ref_interlock(&buffer->io.lock) == 0) {
577 hammer_io_advance(&buffer->io);
578 KKASSERT(buffer->ondisk);
584 * 0->1 transition or defered 0->1 transition (CHECK),
585 * interlock now held. Shortcut if ondisk is already
588 atomic_add_int(&hammer_count_refedbufs, 1);
589 if (buffer->ondisk) {
590 hammer_io_advance(&buffer->io);
591 hammer_ref_interlock_done(&buffer->io.lock);
597 * The buffer is no longer loose if it has a ref, and
598 * cannot become loose once it gains a ref. Loose
599 * buffers will never be in a modified state. This should
600 * only occur on the 0->1 transition of refs.
602 * lose_list can be modified via a biodone() interrupt
603 * so the io_token must be held.
605 if (buffer->io.mod_root == &hmp->lose_root) {
606 lwkt_gettoken(&hmp->io_token);
607 if (buffer->io.mod_root == &hmp->lose_root) {
608 RB_REMOVE(hammer_mod_rb_tree,
609 buffer->io.mod_root, &buffer->io);
610 buffer->io.mod_root = NULL;
611 KKASSERT(buffer->io.modified == 0);
613 lwkt_reltoken(&hmp->io_token);
616 } else if (hmp->ronly && hammer_direct_zone(buf_offset)) {
618 * If this is a read-only mount there could be an alias
619 * in the raw-zone. If there is we use that buffer instead.
621 * rw mounts will not have aliases. Also note when going
622 * from ro -> rw the recovered raw buffers are flushed and
623 * reclaimed, so again there will not be any aliases once
626 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
627 hammer_xlate_to_zone2(buf_offset));
629 kprintf("HAMMER: recovered aliased %016jx\n",
630 (intmax_t)buf_offset);
636 * What is the buffer class?
638 zone = HAMMER_ZONE_DECODE(buf_offset);
641 case HAMMER_ZONE_LARGE_DATA_INDEX:
642 case HAMMER_ZONE_SMALL_DATA_INDEX:
643 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
645 case HAMMER_ZONE_UNDO_INDEX:
646 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
648 case HAMMER_ZONE_META_INDEX:
651 * NOTE: inode data and directory entries are placed in this
652 * zone. inode atime/mtime is updated in-place and thus
653 * buffers containing inodes must be synchronized as
654 * meta-buffers, same as buffers containing B-Tree info.
656 iotype = HAMMER_STRUCTURE_META_BUFFER;
661 * Handle blockmap offset translations
663 if (zone >= HAMMER_ZONE_BTREE_INDEX) {
664 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
665 } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
666 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
668 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
669 zone2_offset = buf_offset;
676 * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
679 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
680 HAMMER_ZONE_RAW_BUFFER);
681 vol_no = HAMMER_VOL_DECODE(zone2_offset);
682 volume = hammer_get_volume(hmp, vol_no, errorp);
686 KKASSERT(zone2_offset < volume->maxbuf_off);
689 * Allocate a new buffer structure. We will check for races later.
691 ++hammer_count_buffers;
692 buffer = kmalloc(sizeof(*buffer), hmp->m_misc,
693 M_WAITOK|M_ZERO|M_USE_RESERVE);
694 buffer->zone2_offset = zone2_offset;
695 buffer->zoneX_offset = buf_offset;
697 hammer_io_init(&buffer->io, volume, iotype);
698 buffer->io.offset = volume->ondisk->vol_buf_beg +
699 (zone2_offset & HAMMER_OFF_SHORT_MASK);
700 buffer->io.bytes = bytes;
701 TAILQ_INIT(&buffer->clist);
702 hammer_ref_interlock_true(&buffer->io.lock);
705 * Insert the buffer into the RB tree and handle late collisions.
707 if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
708 hammer_rel_volume(volume, 0);
709 buffer->io.volume = NULL; /* safety */
710 if (hammer_rel_interlock(&buffer->io.lock, 1)) /* safety */
711 hammer_rel_interlock_done(&buffer->io.lock, 1);
712 --hammer_count_buffers;
713 kfree(buffer, hmp->m_misc);
716 atomic_add_int(&hammer_count_refedbufs, 1);
720 * The buffer is referenced and interlocked. Load the buffer
721 * if necessary. hammer_load_buffer() deals with the interlock
722 * and, if an error is returned, also deals with the ref.
724 if (buffer->ondisk == NULL) {
725 *errorp = hammer_load_buffer(buffer, isnew);
729 hammer_io_advance(&buffer->io);
730 hammer_ref_interlock_done(&buffer->io.lock);
737 * This is used by the direct-read code to deal with large-data buffers
738 * created by the reblocker and mirror-write code. The direct-read code
739 * bypasses the HAMMER buffer subsystem and so any aliased dirty or write-
740 * running hammer buffers must be fully synced to disk before we can issue
743 * This code path is not considered critical as only the rebocker and
744 * mirror-write code will create large-data buffers via the HAMMER buffer
745 * subsystem. They do that because they operate at the B-Tree level and
746 * do not access the vnode/inode structures.
749 hammer_sync_buffers(hammer_mount_t hmp, hammer_off_t base_offset, int bytes)
751 hammer_buffer_t buffer;
754 KKASSERT((base_offset & HAMMER_OFF_ZONE_MASK) ==
755 HAMMER_ZONE_LARGE_DATA);
758 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
760 if (buffer && (buffer->io.modified || buffer->io.running)) {
761 error = hammer_ref_buffer(buffer);
763 hammer_io_wait(&buffer->io);
764 if (buffer->io.modified) {
765 hammer_io_write_interlock(&buffer->io);
766 hammer_io_flush(&buffer->io, 0);
767 hammer_io_done_interlock(&buffer->io);
768 hammer_io_wait(&buffer->io);
770 hammer_rel_buffer(buffer, 0);
773 base_offset += HAMMER_BUFSIZE;
774 bytes -= HAMMER_BUFSIZE;
779 * Destroy all buffers covering the specified zoneX offset range. This
780 * is called when the related blockmap layer2 entry is freed or when
781 * a direct write bypasses our buffer/buffer-cache subsystem.
783 * The buffers may be referenced by the caller itself. Setting reclaim
784 * will cause the buffer to be destroyed when it's ref count reaches zero.
786 * Return 0 on success, EAGAIN if some buffers could not be destroyed due
787 * to additional references held by other threads, or some other (typically
791 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
792 hammer_off_t zone2_offset, int bytes,
793 int report_conflicts)
795 hammer_buffer_t buffer;
796 hammer_volume_t volume;
801 vol_no = HAMMER_VOL_DECODE(zone2_offset);
802 volume = hammer_get_volume(hmp, vol_no, &ret_error);
803 KKASSERT(ret_error == 0);
806 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
809 error = hammer_ref_buffer(buffer);
810 if (hammer_debug_general & 0x20000) {
811 kprintf("hammer: delbufr %016jx "
813 (intmax_t)buffer->zoneX_offset,
815 hammer_oneref(&buffer->io.lock));
817 if (error == 0 && !hammer_oneref(&buffer->io.lock)) {
819 hammer_rel_buffer(buffer, 0);
822 KKASSERT(buffer->zone2_offset == zone2_offset);
823 hammer_io_clear_modify(&buffer->io, 1);
824 buffer->io.reclaim = 1;
825 buffer->io.waitdep = 1;
826 KKASSERT(buffer->io.volume == volume);
827 hammer_rel_buffer(buffer, 0);
830 error = hammer_io_inval(volume, zone2_offset);
834 if (report_conflicts ||
835 (hammer_debug_general & 0x8000)) {
836 kprintf("hammer_del_buffers: unable to "
837 "invalidate %016llx buffer=%p rep=%d\n",
838 (long long)base_offset,
839 buffer, report_conflicts);
842 base_offset += HAMMER_BUFSIZE;
843 zone2_offset += HAMMER_BUFSIZE;
844 bytes -= HAMMER_BUFSIZE;
846 hammer_rel_volume(volume, 0);
851 * Given a referenced and interlocked buffer load/validate the data.
853 * The buffer interlock will be released on return. If an error is
854 * returned the buffer reference will also be released (and the buffer
855 * pointer will thus be stale).
858 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
860 hammer_volume_t volume;
864 * Load the buffer's on-disk info
866 volume = buffer->io.volume;
868 if (hammer_debug_io & 0x0004) {
869 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
870 (long long)buffer->zoneX_offset,
871 (long long)buffer->zone2_offset,
872 isnew, buffer->ondisk);
875 if (buffer->ondisk == NULL) {
877 * Issue the read or generate a new buffer. When reading
878 * the limit argument controls any read-ahead clustering
879 * hammer_io_read() is allowed to do.
881 * We cannot read-ahead in the large-data zone and we cannot
882 * cross a big-block boundary as the next big-block might
883 * use a different buffer size.
886 error = hammer_io_new(volume->devvp, &buffer->io);
887 } else if ((buffer->zoneX_offset & HAMMER_OFF_ZONE_MASK) ==
888 HAMMER_ZONE_LARGE_DATA) {
889 error = hammer_io_read(volume->devvp, &buffer->io,
894 limit = (buffer->zone2_offset +
895 HAMMER_BIGBLOCK_MASK64) &
896 ~HAMMER_BIGBLOCK_MASK64;
897 limit -= buffer->zone2_offset;
898 error = hammer_io_read(volume->devvp, &buffer->io,
902 buffer->ondisk = (void *)buffer->io.bp->b_data;
904 error = hammer_io_new(volume->devvp, &buffer->io);
909 hammer_io_advance(&buffer->io);
910 hammer_ref_interlock_done(&buffer->io.lock);
912 hammer_rel_buffer(buffer, 1);
918 * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
919 * This routine is only called during unmount or when a volume is
922 * If data != NULL, it specifies a volume whoose buffers should
926 hammer_unload_buffer(hammer_buffer_t buffer, void *data)
928 struct hammer_volume *volume = (struct hammer_volume *) data;
931 * If volume != NULL we are only interested in unloading buffers
932 * associated with a particular volume.
934 if (volume != NULL && volume != buffer->io.volume)
938 * Clean up the persistent ref ioerror might have on the buffer
939 * and acquire a ref. Expect a 0->1 transition.
941 if (buffer->io.ioerror) {
942 hammer_io_clear_error_noassert(&buffer->io);
943 atomic_add_int(&hammer_count_refedbufs, -1);
945 hammer_ref_interlock_true(&buffer->io.lock);
946 atomic_add_int(&hammer_count_refedbufs, 1);
949 * We must not flush a dirty buffer to disk on umount. It should
950 * have already been dealt with by the flusher, or we may be in
951 * catastrophic failure.
953 * We must set waitdep to ensure that a running buffer is waited
954 * on and released prior to us trying to unload the volume.
956 hammer_io_clear_modify(&buffer->io, 1);
957 hammer_flush_buffer_nodes(buffer);
958 buffer->io.waitdep = 1;
959 hammer_rel_buffer(buffer, 1);
964 * Reference a buffer that is either already referenced or via a specially
965 * handled pointer (aka cursor->buffer).
968 hammer_ref_buffer(hammer_buffer_t buffer)
975 * Acquire a ref, plus the buffer will be interlocked on the
978 locked = hammer_ref_interlock(&buffer->io.lock);
979 hmp = buffer->io.hmp;
982 * At this point a biodone() will not touch the buffer other then
983 * incidental bits. However, lose_list can be modified via
984 * a biodone() interrupt.
986 * No longer loose. lose_list requires the io_token.
988 if (buffer->io.mod_root == &hmp->lose_root) {
989 lwkt_gettoken(&hmp->io_token);
990 if (buffer->io.mod_root == &hmp->lose_root) {
991 RB_REMOVE(hammer_mod_rb_tree,
992 buffer->io.mod_root, &buffer->io);
993 buffer->io.mod_root = NULL;
995 lwkt_reltoken(&hmp->io_token);
999 atomic_add_int(&hammer_count_refedbufs, 1);
1000 error = hammer_load_buffer(buffer, 0);
1001 /* NOTE: on error the buffer pointer is stale */
1009 * Release a reference on the buffer. On the 1->0 transition the
1010 * underlying IO will be released but the data reference is left
1013 * Only destroy the structure itself if the related buffer cache buffer
1014 * was disassociated from it. This ties the management of the structure
1015 * to the buffer cache subsystem. buffer->ondisk determines whether the
1016 * embedded io is referenced or not.
1019 hammer_rel_buffer(hammer_buffer_t buffer, int locked)
1021 hammer_volume_t volume;
1023 struct buf *bp = NULL;
1026 hmp = buffer->io.hmp;
1028 if (hammer_rel_interlock(&buffer->io.lock, locked) == 0)
1032 * hammer_count_refedbufs accounting. Decrement if we are in
1033 * the error path or if CHECK is clear.
1035 * If we are not in the error path and CHECK is set the caller
1036 * probably just did a hammer_ref() and didn't account for it,
1037 * so we don't account for the loss here.
1039 if (locked || (buffer->io.lock.refs & HAMMER_REFS_CHECK) == 0)
1040 atomic_add_int(&hammer_count_refedbufs, -1);
1043 * If the caller locked us or the normal released transitions
1044 * from 1->0 (and acquired the lock) attempt to release the
1045 * io. If the called locked us we tell hammer_io_release()
1046 * to flush (which would be the unload or failure path).
1048 bp = hammer_io_release(&buffer->io, locked);
1051 * If the buffer has no bp association and no refs we can destroy
1054 * NOTE: It is impossible for any associated B-Tree nodes to have
1055 * refs if the buffer has no additional refs.
1057 if (buffer->io.bp == NULL && hammer_norefs(&buffer->io.lock)) {
1058 RB_REMOVE(hammer_buf_rb_tree,
1059 &buffer->io.hmp->rb_bufs_root,
1061 volume = buffer->io.volume;
1062 buffer->io.volume = NULL; /* sanity */
1063 hammer_rel_volume(volume, 0);
1064 hammer_io_clear_modlist(&buffer->io);
1065 hammer_flush_buffer_nodes(buffer);
1066 KKASSERT(TAILQ_EMPTY(&buffer->clist));
1073 hammer_rel_interlock_done(&buffer->io.lock, locked);
1077 --hammer_count_buffers;
1078 kfree(buffer, hmp->m_misc);
1083 * Access the filesystem buffer containing the specified hammer offset.
1084 * buf_offset is a conglomeration of the volume number and vol_buf_beg
1085 * relative buffer offset. It must also have bit 55 set to be valid.
1086 * (see hammer_off_t in hammer_disk.h).
1088 * Any prior buffer in *bufferp will be released and replaced by the
1091 * NOTE: The buffer is indexed via its zoneX_offset but we allow the
1092 * passed cached *bufferp to match against either zoneX or zone2.
1096 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1097 int *errorp, struct hammer_buffer **bufferp)
1099 hammer_buffer_t buffer;
1100 int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
1102 buf_offset &= ~HAMMER_BUFMASK64;
1103 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
1106 if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
1107 buffer->zoneX_offset != buf_offset)) {
1109 hammer_rel_buffer(buffer, 0);
1110 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
1117 * Return a pointer to the buffer data.
1122 return((char *)buffer->ondisk + xoff);
1126 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
1127 int *errorp, struct hammer_buffer **bufferp)
1129 return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
1133 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1134 int *errorp, struct hammer_buffer **bufferp)
1136 bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1137 return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
1141 * Access the filesystem buffer containing the specified hammer offset.
1142 * No disk read operation occurs. The result buffer may contain garbage.
1144 * Any prior buffer in *bufferp will be released and replaced by the
1147 * This function marks the buffer dirty but does not increment its
1148 * modify_refs count.
1152 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1153 int *errorp, struct hammer_buffer **bufferp)
1155 hammer_buffer_t buffer;
1156 int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
1158 buf_offset &= ~HAMMER_BUFMASK64;
1161 if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
1162 buffer->zoneX_offset != buf_offset)) {
1164 hammer_rel_buffer(buffer, 0);
1165 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
1172 * Return a pointer to the buffer data.
1177 return((char *)buffer->ondisk + xoff);
1181 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
1182 int *errorp, struct hammer_buffer **bufferp)
1184 return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
1188 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1189 int *errorp, struct hammer_buffer **bufferp)
1191 bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1192 return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
1195 /************************************************************************
1197 ************************************************************************
1199 * Manage B-Tree nodes. B-Tree nodes represent the primary indexing
1200 * method used by the HAMMER filesystem.
1202 * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
1203 * associated with its buffer, and will only referenced the buffer while
1204 * the node itself is referenced.
1206 * A hammer_node can also be passively associated with other HAMMER
1207 * structures, such as inodes, while retaining 0 references. These
1208 * associations can be cleared backwards using a pointer-to-pointer in
1211 * This allows the HAMMER implementation to cache hammer_nodes long-term
1212 * and short-cut a great deal of the infrastructure's complexity. In
1213 * most cases a cached node can be reacquired without having to dip into
1214 * either the buffer or cluster management code.
1216 * The caller must pass a referenced cluster on call and will retain
1217 * ownership of the reference on return. The node will acquire its own
1218 * additional references, if necessary.
1221 hammer_get_node(hammer_transaction_t trans, hammer_off_t node_offset,
1222 int isnew, int *errorp)
1224 hammer_mount_t hmp = trans->hmp;
1228 KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
1231 * Locate the structure, allocating one if necessary.
1234 node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
1236 ++hammer_count_nodes;
1237 node = kmalloc(sizeof(*node), hmp->m_misc, M_WAITOK|M_ZERO|M_USE_RESERVE);
1238 node->node_offset = node_offset;
1240 TAILQ_INIT(&node->cursor_list);
1241 TAILQ_INIT(&node->cache_list);
1242 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
1243 --hammer_count_nodes;
1244 kfree(node, hmp->m_misc);
1247 doload = hammer_ref_interlock_true(&node->lock);
1249 doload = hammer_ref_interlock(&node->lock);
1252 *errorp = hammer_load_node(trans, node, isnew);
1253 trans->flags |= HAMMER_TRANSF_DIDIO;
1257 KKASSERT(node->ondisk);
1259 hammer_io_advance(&node->buffer->io);
1265 * Reference an already-referenced node. 0->1 transitions should assert
1266 * so we do not have to deal with hammer_ref() setting CHECK.
1269 hammer_ref_node(hammer_node_t node)
1271 KKASSERT(hammer_isactive(&node->lock) && node->ondisk != NULL);
1272 hammer_ref(&node->lock);
1276 * Load a node's on-disk data reference. Called with the node referenced
1279 * On return the node interlock will be unlocked. If a non-zero error code
1280 * is returned the node will also be dereferenced (and the caller's pointer
1284 hammer_load_node(hammer_transaction_t trans, hammer_node_t node, int isnew)
1286 hammer_buffer_t buffer;
1287 hammer_off_t buf_offset;
1291 if (node->ondisk == NULL) {
1293 * This is a little confusing but the jist is that
1294 * node->buffer determines whether the node is on
1295 * the buffer's clist and node->ondisk determines
1296 * whether the buffer is referenced.
1298 * We could be racing a buffer release, in which case
1299 * node->buffer may become NULL while we are blocked
1300 * referencing the buffer.
1302 if ((buffer = node->buffer) != NULL) {
1303 error = hammer_ref_buffer(buffer);
1304 if (error == 0 && node->buffer == NULL) {
1305 TAILQ_INSERT_TAIL(&buffer->clist,
1307 node->buffer = buffer;
1310 buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1311 buffer = hammer_get_buffer(node->hmp, buf_offset,
1312 HAMMER_BUFSIZE, 0, &error);
1314 KKASSERT(error == 0);
1315 TAILQ_INSERT_TAIL(&buffer->clist,
1317 node->buffer = buffer;
1322 node->ondisk = (void *)((char *)buffer->ondisk +
1323 (node->node_offset & HAMMER_BUFMASK));
1326 * Check CRC. NOTE: Neither flag is set and the CRC is not
1327 * generated on new B-Tree nodes.
1330 (node->flags & HAMMER_NODE_CRCANY) == 0) {
1331 if (hammer_crc_test_btree(node->ondisk) == 0) {
1332 if (hammer_debug_critical)
1333 Debugger("CRC FAILED: B-TREE NODE");
1334 node->flags |= HAMMER_NODE_CRCBAD;
1336 node->flags |= HAMMER_NODE_CRCGOOD;
1340 if (node->flags & HAMMER_NODE_CRCBAD) {
1341 if (trans->flags & HAMMER_TRANSF_CRCDOM)
1348 _hammer_rel_node(node, 1);
1350 hammer_ref_interlock_done(&node->lock);
1356 * Safely reference a node, interlock against flushes via the IO subsystem.
1359 hammer_ref_node_safe(hammer_transaction_t trans, hammer_node_cache_t cache,
1367 doload = hammer_ref_interlock(&node->lock);
1369 *errorp = hammer_load_node(trans, node, 0);
1373 KKASSERT(node->ondisk);
1374 if (node->flags & HAMMER_NODE_CRCBAD) {
1375 if (trans->flags & HAMMER_TRANSF_CRCDOM)
1379 _hammer_rel_node(node, 0);
1392 * Release a hammer_node. On the last release the node dereferences
1393 * its underlying buffer and may or may not be destroyed.
1395 * If locked is non-zero the passed node has been interlocked by the
1396 * caller and we are in the failure/unload path, otherwise it has not and
1397 * we are doing a normal release.
1399 * This function will dispose of the interlock and the reference.
1400 * On return the node pointer is stale.
1403 _hammer_rel_node(hammer_node_t node, int locked)
1405 hammer_buffer_t buffer;
1408 * Deref the node. If this isn't the 1->0 transition we're basically
1409 * done. If locked is non-zero this function will just deref the
1410 * locked node and return TRUE, otherwise it will deref the locked
1411 * node and either lock and return TRUE on the 1->0 transition or
1412 * not lock and return FALSE.
1414 if (hammer_rel_interlock(&node->lock, locked) == 0)
1418 * Either locked was non-zero and we are interlocked, or the
1419 * hammer_rel_interlock() call returned non-zero and we are
1422 * The ref-count must still be decremented if locked != 0 so
1423 * the cleanup required still varies a bit.
1425 * hammer_flush_node() when called with 1 or 2 will dispose of
1426 * the lock and possible ref-count.
1428 if (node->ondisk == NULL) {
1429 hammer_flush_node(node, locked + 1);
1430 /* node is stale now */
1435 * Do not disassociate the node from the buffer if it represents
1436 * a modified B-Tree node that still needs its crc to be generated.
1438 if (node->flags & HAMMER_NODE_NEEDSCRC) {
1439 hammer_rel_interlock_done(&node->lock, locked);
1444 * Do final cleanups and then either destroy the node and leave it
1445 * passively cached. The buffer reference is removed regardless.
1447 buffer = node->buffer;
1448 node->ondisk = NULL;
1450 if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1454 hammer_rel_interlock_done(&node->lock, locked);
1459 hammer_flush_node(node, locked + 1);
1463 hammer_rel_buffer(buffer, 0);
1467 hammer_rel_node(hammer_node_t node)
1469 _hammer_rel_node(node, 0);
1473 * Free space on-media associated with a B-Tree node.
1476 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1478 KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1479 node->flags |= HAMMER_NODE_DELETED;
1480 hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1484 * Passively cache a referenced hammer_node. The caller may release
1485 * the node on return.
1488 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1491 * If the node doesn't exist, or is being deleted, don't cache it!
1493 * The node can only ever be NULL in the I/O failure path.
1495 if (node == NULL || (node->flags & HAMMER_NODE_DELETED))
1497 if (cache->node == node)
1500 hammer_uncache_node(cache);
1501 if (node->flags & HAMMER_NODE_DELETED)
1504 TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1508 hammer_uncache_node(hammer_node_cache_t cache)
1512 if ((node = cache->node) != NULL) {
1513 TAILQ_REMOVE(&node->cache_list, cache, entry);
1515 if (TAILQ_EMPTY(&node->cache_list))
1516 hammer_flush_node(node, 0);
1521 * Remove a node's cache references and destroy the node if it has no
1522 * other references or backing store.
1524 * locked == 0 Normal unlocked operation
1525 * locked == 1 Call hammer_rel_interlock_done(..., 0);
1526 * locked == 2 Call hammer_rel_interlock_done(..., 1);
1528 * XXX for now this isn't even close to being MPSAFE so the refs check
1532 hammer_flush_node(hammer_node_t node, int locked)
1534 hammer_node_cache_t cache;
1535 hammer_buffer_t buffer;
1536 hammer_mount_t hmp = node->hmp;
1539 while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1540 TAILQ_REMOVE(&node->cache_list, cache, entry);
1545 * NOTE: refs is predisposed if another thread is blocking and
1546 * will be larger than 0 in that case. We aren't MPSAFE
1549 if (node->ondisk == NULL && hammer_norefs(&node->lock)) {
1550 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1551 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1552 if ((buffer = node->buffer) != NULL) {
1553 node->buffer = NULL;
1554 TAILQ_REMOVE(&buffer->clist, node, entry);
1555 /* buffer is unreferenced because ondisk is NULL */
1563 * Deal with the interlock if locked == 1 or locked == 2.
1566 hammer_rel_interlock_done(&node->lock, locked - 1);
1569 * Destroy if requested
1572 --hammer_count_nodes;
1573 kfree(node, hmp->m_misc);
1578 * Flush passively cached B-Tree nodes associated with this buffer.
1579 * This is only called when the buffer is about to be destroyed, so
1580 * none of the nodes should have any references. The buffer is locked.
1582 * We may be interlocked with the buffer.
1585 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1589 while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1590 KKASSERT(node->ondisk == NULL);
1591 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1593 if (hammer_try_interlock_norefs(&node->lock)) {
1594 hammer_ref(&node->lock);
1595 node->flags |= HAMMER_NODE_FLUSH;
1596 _hammer_rel_node(node, 1);
1598 KKASSERT(node->buffer != NULL);
1599 buffer = node->buffer;
1600 node->buffer = NULL;
1601 TAILQ_REMOVE(&buffer->clist, node, entry);
1602 /* buffer is unreferenced because ondisk is NULL */
1608 /************************************************************************
1610 ************************************************************************/
1613 * Allocate a B-Tree node.
1616 hammer_alloc_btree(hammer_transaction_t trans, hammer_off_t hint, int *errorp)
1618 hammer_buffer_t buffer = NULL;
1619 hammer_node_t node = NULL;
1620 hammer_off_t node_offset;
1622 node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1623 sizeof(struct hammer_node_ondisk),
1626 node = hammer_get_node(trans, node_offset, 1, errorp);
1627 hammer_modify_node_noundo(trans, node);
1628 bzero(node->ondisk, sizeof(*node->ondisk));
1629 hammer_modify_node_done(node);
1632 hammer_rel_buffer(buffer, 0);
1637 * Allocate data. If the address of a data buffer is supplied then
1638 * any prior non-NULL *data_bufferp will be released and *data_bufferp
1639 * will be set to the related buffer. The caller must release it when
1640 * finally done. The initial *data_bufferp should be set to NULL by
1643 * The caller is responsible for making hammer_modify*() calls on the
1647 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1648 u_int16_t rec_type, hammer_off_t *data_offsetp,
1649 struct hammer_buffer **data_bufferp,
1650 hammer_off_t hint, int *errorp)
1660 case HAMMER_RECTYPE_INODE:
1661 case HAMMER_RECTYPE_DIRENTRY:
1662 case HAMMER_RECTYPE_EXT:
1663 case HAMMER_RECTYPE_FIX:
1664 case HAMMER_RECTYPE_PFS:
1665 case HAMMER_RECTYPE_SNAPSHOT:
1666 case HAMMER_RECTYPE_CONFIG:
1667 zone = HAMMER_ZONE_META_INDEX;
1669 case HAMMER_RECTYPE_DATA:
1670 case HAMMER_RECTYPE_DB:
1672 * This is an exceptional case. HAMMER usually
1673 * uses HAMMER_ZONE_LARGE_DATA when the data length
1674 * is >=HAMMER_BUFSIZE, but not 1/2 of that. Mirror
1675 * write code seems to be the only case that allocates
1676 * HAMMER_RECTYPE_DATA via this function.
1678 * When data_len is >HAMMER_BUFSIZE/2 it uses
1679 * HAMMER_ZONE_LARGE_DATA but data_len is also rounded
1680 * up so it doesn't make much difference from the
1681 * normal way of using this zone.
1683 * Also note hammer_vop_strategy_write() could have
1684 * rounded up storage allocation size of the original
1685 * mirror source to fs block size when it was written
1686 * if the file size was >HAMMER_BUFSIZE/2.
1688 if (data_len <= HAMMER_BUFSIZE / 2) {
1689 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1691 data_len = (data_len + HAMMER_BUFMASK) &
1693 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1697 panic("hammer_alloc_data: rec_type %04x unknown",
1699 zone = 0; /* NOT REACHED */
1702 *data_offsetp = hammer_blockmap_alloc(trans, zone, data_len,
1707 if (*errorp == 0 && data_bufferp) {
1709 data = hammer_bread_ext(trans->hmp, *data_offsetp,
1710 data_len, errorp, data_bufferp);
1721 * Sync dirty buffers to the media and clean-up any loose ends.
1723 * These functions do not start the flusher going, they simply
1724 * queue everything up to the flusher.
1726 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1729 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1731 struct hammer_sync_info info;
1734 info.waitfor = waitfor;
1735 if (waitfor == MNT_WAIT) {
1736 vsyncscan(hmp->mp, VMSC_GETVP | VMSC_ONEPASS,
1737 hammer_sync_scan2, &info);
1739 vsyncscan(hmp->mp, VMSC_GETVP | VMSC_ONEPASS | VMSC_NOWAIT,
1740 hammer_sync_scan2, &info);
1746 * Filesystem sync. If doing a synchronous sync make a second pass on
1747 * the vnodes in case any were already flushing during the first pass,
1748 * and activate the flusher twice (the second time brings the UNDO FIFO's
1749 * start position up to the end position after the first call).
1751 * If doing a lazy sync make just one pass on the vnode list, ignoring
1752 * any new vnodes added to the list while the sync is in progress.
1755 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1757 struct hammer_sync_info info;
1761 if (waitfor & MNT_LAZY)
1762 flags |= VMSC_ONEPASS;
1765 info.waitfor = MNT_NOWAIT;
1766 vsyncscan(hmp->mp, flags | VMSC_NOWAIT, hammer_sync_scan2, &info);
1768 if (info.error == 0 && (waitfor & MNT_WAIT)) {
1769 info.waitfor = waitfor;
1770 vsyncscan(hmp->mp, flags, hammer_sync_scan2, &info);
1772 if (waitfor == MNT_WAIT) {
1773 hammer_flusher_sync(hmp);
1774 hammer_flusher_sync(hmp);
1776 hammer_flusher_async(hmp, NULL);
1777 hammer_flusher_async(hmp, NULL);
1783 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1785 struct hammer_sync_info *info = data;
1786 struct hammer_inode *ip;
1792 if (vp->v_type == VNON || vp->v_type == VBAD) {
1796 if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1797 RB_EMPTY(&vp->v_rbdirty_tree)) {
1801 error = VOP_FSYNC(vp, MNT_NOWAIT, 0);
1803 info->error = error;