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 1.67 2008/07/09 10:29:20 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>
48 static void hammer_free_volume(hammer_volume_t volume);
49 static int hammer_load_volume(hammer_volume_t volume);
50 static int hammer_load_buffer(hammer_buffer_t buffer, int isnew);
51 static int hammer_load_node(hammer_node_t node, int isnew);
54 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
56 if (vol1->vol_no < vol2->vol_no)
58 if (vol1->vol_no > vol2->vol_no)
64 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
66 if (buf1->zoneX_offset < buf2->zoneX_offset)
68 if (buf1->zoneX_offset > buf2->zoneX_offset)
74 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
76 if (node1->node_offset < node2->node_offset)
78 if (node1->node_offset > node2->node_offset)
83 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
84 hammer_vol_rb_compare, int32_t, vol_no);
85 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
86 hammer_buf_rb_compare, hammer_off_t, zoneX_offset);
87 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
88 hammer_nod_rb_compare, hammer_off_t, node_offset);
90 /************************************************************************
92 ************************************************************************
94 * Load a HAMMER volume by name. Returns 0 on success or a positive error
95 * code on failure. Volumes must be loaded at mount time, get_volume() will
96 * not load a new volume.
98 * Calls made to hammer_load_volume() or single-threaded
101 hammer_install_volume(struct hammer_mount *hmp, const char *volname)
104 hammer_volume_t volume;
105 struct hammer_volume_ondisk *ondisk;
106 struct nlookupdata nd;
107 struct buf *bp = NULL;
113 ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
116 * Allocate a volume structure
118 ++hammer_count_volumes;
119 volume = kmalloc(sizeof(*volume), M_HAMMER, M_WAITOK|M_ZERO);
120 volume->vol_name = kstrdup(volname, M_HAMMER);
121 hammer_io_init(&volume->io, hmp, HAMMER_STRUCTURE_VOLUME);
122 volume->io.offset = 0LL;
123 volume->io.bytes = HAMMER_BUFSIZE;
126 * Get the device vnode
128 error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
130 error = nlookup(&nd);
132 error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
135 if (vn_isdisk(volume->devvp, &error)) {
136 error = vfs_mountedon(volume->devvp);
140 count_udev(volume->devvp->v_umajor, volume->devvp->v_uminor) > 0) {
144 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
145 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
147 error = VOP_OPEN(volume->devvp,
148 (ronly ? FREAD : FREAD|FWRITE),
151 vn_unlock(volume->devvp);
154 hammer_free_volume(volume);
157 volume->devvp->v_rdev->si_mountpoint = mp;
161 * Extract the volume number from the volume header and do various
164 error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
167 ondisk = (void *)bp->b_data;
168 if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
169 kprintf("hammer_mount: volume %s has an invalid header\n",
174 volume->vol_no = ondisk->vol_no;
175 volume->buffer_base = ondisk->vol_buf_beg;
176 volume->vol_flags = ondisk->vol_flags;
177 volume->nblocks = ondisk->vol_nblocks;
178 volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
179 ondisk->vol_buf_end - ondisk->vol_buf_beg);
180 volume->maxraw_off = ondisk->vol_buf_end;
182 if (RB_EMPTY(&hmp->rb_vols_root)) {
183 hmp->fsid = ondisk->vol_fsid;
184 } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
185 kprintf("hammer_mount: volume %s's fsid does not match "
186 "other volumes\n", volume->vol_name);
192 * Insert the volume structure into the red-black tree.
194 if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
195 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
196 volume->vol_name, volume->vol_no);
201 * Set the root volume . HAMMER special cases rootvol the structure.
202 * We do not hold a ref because this would prevent related I/O
203 * from being flushed.
205 if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
206 hmp->rootvol = volume;
211 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
212 (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
213 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
214 (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
220 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
222 volume->devvp->v_rdev->si_mountpoint = NULL;
223 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
224 hammer_free_volume(volume);
230 * This is called for each volume when updating the mount point from
231 * read-write to read-only or vise-versa.
234 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
237 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
238 if (volume->io.hmp->ronly) {
239 /* do not call vinvalbuf */
240 VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
241 VOP_CLOSE(volume->devvp, FREAD|FWRITE);
243 /* do not call vinvalbuf */
244 VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
245 VOP_CLOSE(volume->devvp, FREAD);
247 vn_unlock(volume->devvp);
253 * Unload and free a HAMMER volume. Must return >= 0 to continue scan
254 * so returns -1 on failure.
257 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
259 struct hammer_mount *hmp = volume->io.hmp;
260 int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
264 * Clean up the root volume pointer, which is held unlocked in hmp.
266 if (hmp->rootvol == volume)
270 * Release our buffer and flush anything left in the buffer cache.
272 volume->io.waitdep = 1;
273 bp = hammer_io_release(&volume->io, 1);
274 hammer_io_clear_modlist(&volume->io);
277 * There should be no references on the volume, no clusters, and
280 KKASSERT(volume->io.lock.refs == 0);
284 volume->ondisk = NULL;
286 if (volume->devvp->v_rdev &&
287 volume->devvp->v_rdev->si_mountpoint == hmp->mp
289 volume->devvp->v_rdev->si_mountpoint = NULL;
292 vinvalbuf(volume->devvp, 0, 0, 0);
293 VOP_CLOSE(volume->devvp, FREAD);
295 vinvalbuf(volume->devvp, V_SAVE, 0, 0);
296 VOP_CLOSE(volume->devvp, FREAD|FWRITE);
301 * Destroy the structure
303 RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
304 hammer_free_volume(volume);
310 hammer_free_volume(hammer_volume_t volume)
312 if (volume->vol_name) {
313 kfree(volume->vol_name, M_HAMMER);
314 volume->vol_name = NULL;
317 vrele(volume->devvp);
318 volume->devvp = NULL;
320 --hammer_count_volumes;
321 kfree(volume, M_HAMMER);
325 * Get a HAMMER volume. The volume must already exist.
328 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
330 struct hammer_volume *volume;
333 * Locate the volume structure
335 volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
336 if (volume == NULL) {
340 hammer_ref(&volume->io.lock);
343 * Deal with on-disk info
345 if (volume->ondisk == NULL || volume->io.loading) {
346 *errorp = hammer_load_volume(volume);
348 hammer_rel_volume(volume, 1);
358 hammer_ref_volume(hammer_volume_t volume)
362 hammer_ref(&volume->io.lock);
365 * Deal with on-disk info
367 if (volume->ondisk == NULL || volume->io.loading) {
368 error = hammer_load_volume(volume);
370 hammer_rel_volume(volume, 1);
378 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
380 hammer_volume_t volume;
382 volume = hmp->rootvol;
383 KKASSERT(volume != NULL);
384 hammer_ref(&volume->io.lock);
387 * Deal with on-disk info
389 if (volume->ondisk == NULL || volume->io.loading) {
390 *errorp = hammer_load_volume(volume);
392 hammer_rel_volume(volume, 1);
402 * Load a volume's on-disk information. The volume must be referenced and
403 * not locked. We temporarily acquire an exclusive lock to interlock
404 * against releases or multiple get's.
407 hammer_load_volume(hammer_volume_t volume)
411 ++volume->io.loading;
412 hammer_lock_ex(&volume->io.lock);
414 if (volume->ondisk == NULL) {
415 error = hammer_io_read(volume->devvp, &volume->io,
418 volume->ondisk = (void *)volume->io.bp->b_data;
422 --volume->io.loading;
423 hammer_unlock(&volume->io.lock);
428 * Release a volume. Call hammer_io_release on the last reference. We have
429 * to acquire an exclusive lock to interlock against volume->ondisk tests
430 * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
433 * Volumes are not unloaded from memory during normal operation.
436 hammer_rel_volume(hammer_volume_t volume, int flush)
438 struct buf *bp = NULL;
441 if (volume->io.lock.refs == 1) {
442 ++volume->io.loading;
443 hammer_lock_ex(&volume->io.lock);
444 if (volume->io.lock.refs == 1) {
445 volume->ondisk = NULL;
446 bp = hammer_io_release(&volume->io, flush);
448 --volume->io.loading;
449 hammer_unlock(&volume->io.lock);
451 hammer_unref(&volume->io.lock);
457 /************************************************************************
459 ************************************************************************
461 * Manage buffers. Currently all blockmap-backed zones are translated
462 * to zone-2 buffer offsets.
465 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
466 int bytes, int isnew, int *errorp)
468 hammer_buffer_t buffer;
469 hammer_volume_t volume;
470 hammer_off_t zone2_offset;
471 hammer_io_type_t iotype;
475 buf_offset &= ~HAMMER_BUFMASK64;
478 * Shortcut if the buffer is already cached
480 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
482 if (buffer->io.lock.refs == 0)
483 ++hammer_count_refedbufs;
484 hammer_ref(&buffer->io.lock);
487 * Onced refed the ondisk field will not be cleared by
490 if (buffer->ondisk && buffer->io.loading == 0) {
496 * The buffer is no longer loose if it has a ref, and
497 * cannot become loose once it gains a ref. Loose
498 * buffers will never be in a modified state. This should
499 * only occur on the 0->1 transition of refs.
501 * lose_list can be modified via a biodone() interrupt.
503 if (buffer->io.mod_list == &hmp->lose_list) {
504 crit_enter(); /* biodone race against list */
505 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
508 buffer->io.mod_list = NULL;
509 KKASSERT(buffer->io.modified == 0);
515 * What is the buffer class?
517 zone = HAMMER_ZONE_DECODE(buf_offset);
520 case HAMMER_ZONE_LARGE_DATA_INDEX:
521 case HAMMER_ZONE_SMALL_DATA_INDEX:
522 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
524 case HAMMER_ZONE_UNDO_INDEX:
525 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
527 case HAMMER_ZONE_META_INDEX:
530 * NOTE: inode data and directory entries are placed in this
531 * zone. inode atime/mtime is updated in-place and thus
532 * buffers containing inodes must be synchronized as
533 * meta-buffers, same as buffers containing B-Tree info.
535 iotype = HAMMER_STRUCTURE_META_BUFFER;
540 * Handle blockmap offset translations
542 if (zone >= HAMMER_ZONE_BTREE_INDEX) {
543 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
544 } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
545 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
547 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
548 zone2_offset = buf_offset;
555 * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
558 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
559 HAMMER_ZONE_RAW_BUFFER);
560 vol_no = HAMMER_VOL_DECODE(zone2_offset);
561 volume = hammer_get_volume(hmp, vol_no, errorp);
565 KKASSERT(zone2_offset < volume->maxbuf_off);
568 * Allocate a new buffer structure. We will check for races later.
570 ++hammer_count_buffers;
571 buffer = kmalloc(sizeof(*buffer), M_HAMMER,
572 M_WAITOK|M_ZERO|M_USE_RESERVE);
573 buffer->zone2_offset = zone2_offset;
574 buffer->zoneX_offset = buf_offset;
575 buffer->volume = volume;
577 hammer_io_init(&buffer->io, hmp, iotype);
578 buffer->io.offset = volume->ondisk->vol_buf_beg +
579 (zone2_offset & HAMMER_OFF_SHORT_MASK);
580 buffer->io.bytes = bytes;
581 TAILQ_INIT(&buffer->clist);
582 hammer_ref(&buffer->io.lock);
585 * Insert the buffer into the RB tree and handle late collisions.
587 if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
588 hammer_unref(&buffer->io.lock);
589 --hammer_count_buffers;
590 kfree(buffer, M_HAMMER);
593 ++hammer_count_refedbufs;
597 * Deal with on-disk info and loading races.
599 if (buffer->ondisk == NULL || buffer->io.loading) {
600 *errorp = hammer_load_buffer(buffer, isnew);
602 hammer_rel_buffer(buffer, 1);
612 * Destroy all buffers covering the specified zoneX offset range. This
613 * is called when the related blockmap layer2 entry is freed or when
614 * a direct write bypasses our buffer/buffer-cache subsystem.
616 * The buffers may be referenced by the caller itself. Setting reclaim
617 * will cause the buffer to be destroyed when it's ref count reaches zero.
620 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
621 hammer_off_t zone2_offset, int bytes)
623 hammer_buffer_t buffer;
624 hammer_volume_t volume;
628 vol_no = HAMMER_VOL_DECODE(zone2_offset);
629 volume = hammer_get_volume(hmp, vol_no, &error);
630 KKASSERT(error == 0);
633 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
636 KKASSERT(buffer->zone2_offset == zone2_offset);
637 hammer_io_clear_modify(&buffer->io, 1);
638 buffer->io.reclaim = 1;
639 KKASSERT(buffer->volume == volume);
640 if (buffer->io.lock.refs == 0)
641 hammer_unload_buffer(buffer, NULL);
643 hammer_io_inval(volume, zone2_offset);
645 base_offset += HAMMER_BUFSIZE;
646 zone2_offset += HAMMER_BUFSIZE;
647 bytes -= HAMMER_BUFSIZE;
649 hammer_rel_volume(volume, 0);
653 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
655 hammer_volume_t volume;
659 * Load the buffer's on-disk info
661 volume = buffer->volume;
662 ++buffer->io.loading;
663 hammer_lock_ex(&buffer->io.lock);
665 if (hammer_debug_io & 0x0001) {
666 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
667 buffer->zoneX_offset, buffer->zone2_offset, isnew,
671 if (buffer->ondisk == NULL) {
673 error = hammer_io_new(volume->devvp, &buffer->io);
675 error = hammer_io_read(volume->devvp, &buffer->io,
679 buffer->ondisk = (void *)buffer->io.bp->b_data;
681 error = hammer_io_new(volume->devvp, &buffer->io);
685 --buffer->io.loading;
686 hammer_unlock(&buffer->io.lock);
691 * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
694 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
696 ++hammer_count_refedbufs;
697 hammer_ref(&buffer->io.lock);
698 hammer_flush_buffer_nodes(buffer);
699 KKASSERT(buffer->io.lock.refs == 1);
700 hammer_rel_buffer(buffer, 2);
705 * Reference a buffer that is either already referenced or via a specially
706 * handled pointer (aka cursor->buffer).
709 hammer_ref_buffer(hammer_buffer_t buffer)
713 if (buffer->io.lock.refs == 0)
714 ++hammer_count_refedbufs;
715 hammer_ref(&buffer->io.lock);
718 * At this point a biodone() will not touch the buffer other then
719 * incidental bits. However, lose_list can be modified via
720 * a biodone() interrupt.
724 if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
726 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
727 buffer->io.mod_list = NULL;
731 if (buffer->ondisk == NULL || buffer->io.loading) {
732 error = hammer_load_buffer(buffer, 0);
734 hammer_rel_buffer(buffer, 1);
736 * NOTE: buffer pointer can become stale after
747 * Release a buffer. We have to deal with several places where
748 * another thread can ref the buffer.
750 * Only destroy the structure itself if the related buffer cache buffer
751 * was disassociated from it. This ties the management of the structure
752 * to the buffer cache subsystem. buffer->ondisk determines whether the
753 * embedded io is referenced or not.
756 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
758 hammer_volume_t volume;
759 struct buf *bp = NULL;
763 if (buffer->io.lock.refs == 1) {
764 ++buffer->io.loading; /* force interlock check */
765 hammer_lock_ex(&buffer->io.lock);
766 if (buffer->io.lock.refs == 1) {
767 bp = hammer_io_release(&buffer->io, flush);
769 if (buffer->io.lock.refs == 1)
770 --hammer_count_refedbufs;
772 if (buffer->io.bp == NULL &&
773 buffer->io.lock.refs == 1) {
777 * NOTE: It is impossible for any associated
778 * B-Tree nodes to have refs if the buffer
779 * has no additional refs.
781 RB_REMOVE(hammer_buf_rb_tree,
782 &buffer->io.hmp->rb_bufs_root,
784 volume = buffer->volume;
785 buffer->volume = NULL; /* sanity */
786 hammer_rel_volume(volume, 0);
787 hammer_io_clear_modlist(&buffer->io);
788 hammer_flush_buffer_nodes(buffer);
789 KKASSERT(TAILQ_EMPTY(&buffer->clist));
793 --buffer->io.loading;
794 hammer_unlock(&buffer->io.lock);
796 hammer_unref(&buffer->io.lock);
801 --hammer_count_buffers;
802 kfree(buffer, M_HAMMER);
807 * Access the filesystem buffer containing the specified hammer offset.
808 * buf_offset is a conglomeration of the volume number and vol_buf_beg
809 * relative buffer offset. It must also have bit 55 set to be valid.
810 * (see hammer_off_t in hammer_disk.h).
812 * Any prior buffer in *bufferp will be released and replaced by the
817 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
818 int *errorp, struct hammer_buffer **bufferp)
820 hammer_buffer_t buffer;
821 int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
823 buf_offset &= ~HAMMER_BUFMASK64;
824 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
827 if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
828 buffer->zoneX_offset != buf_offset)) {
830 hammer_rel_buffer(buffer, 0);
831 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
838 * Return a pointer to the buffer data.
843 return((char *)buffer->ondisk + xoff);
847 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
848 int *errorp, struct hammer_buffer **bufferp)
850 return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
854 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
855 int *errorp, struct hammer_buffer **bufferp)
857 bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
858 return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
862 * Access the filesystem buffer containing the specified hammer offset.
863 * No disk read operation occurs. The result buffer may contain garbage.
865 * Any prior buffer in *bufferp will be released and replaced by the
868 * This function marks the buffer dirty but does not increment its
873 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
874 int *errorp, struct hammer_buffer **bufferp)
876 hammer_buffer_t buffer;
877 int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
879 buf_offset &= ~HAMMER_BUFMASK64;
882 if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
883 buffer->zoneX_offset != buf_offset)) {
885 hammer_rel_buffer(buffer, 0);
886 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
893 * Return a pointer to the buffer data.
898 return((char *)buffer->ondisk + xoff);
902 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
903 int *errorp, struct hammer_buffer **bufferp)
905 return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
909 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
910 int *errorp, struct hammer_buffer **bufferp)
912 bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
913 return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
916 /************************************************************************
918 ************************************************************************
920 * Manage B-Tree nodes. B-Tree nodes represent the primary indexing
921 * method used by the HAMMER filesystem.
923 * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
924 * associated with its buffer, and will only referenced the buffer while
925 * the node itself is referenced.
927 * A hammer_node can also be passively associated with other HAMMER
928 * structures, such as inodes, while retaining 0 references. These
929 * associations can be cleared backwards using a pointer-to-pointer in
932 * This allows the HAMMER implementation to cache hammer_nodes long-term
933 * and short-cut a great deal of the infrastructure's complexity. In
934 * most cases a cached node can be reacquired without having to dip into
935 * either the buffer or cluster management code.
937 * The caller must pass a referenced cluster on call and will retain
938 * ownership of the reference on return. The node will acquire its own
939 * additional references, if necessary.
942 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
943 int isnew, int *errorp)
947 KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
950 * Locate the structure, allocating one if necessary.
953 node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
955 ++hammer_count_nodes;
956 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO|M_USE_RESERVE);
957 node->node_offset = node_offset;
959 TAILQ_INIT(&node->cursor_list);
960 TAILQ_INIT(&node->cache_list);
961 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
962 --hammer_count_nodes;
963 kfree(node, M_HAMMER);
967 hammer_ref(&node->lock);
971 *errorp = hammer_load_node(node, isnew);
973 hammer_rel_node(node);
980 * Reference an already-referenced node.
983 hammer_ref_node(hammer_node_t node)
985 KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
986 hammer_ref(&node->lock);
990 * Load a node's on-disk data reference.
993 hammer_load_node(hammer_node_t node, int isnew)
995 hammer_buffer_t buffer;
996 hammer_off_t buf_offset;
1001 hammer_lock_ex(&node->lock);
1002 if (node->ondisk == NULL) {
1004 * This is a little confusing but the jist is that
1005 * node->buffer determines whether the node is on
1006 * the buffer's clist and node->ondisk determines
1007 * whether the buffer is referenced.
1009 * We could be racing a buffer release, in which case
1010 * node->buffer may become NULL while we are blocked
1011 * referencing the buffer.
1013 if ((buffer = node->buffer) != NULL) {
1014 error = hammer_ref_buffer(buffer);
1015 if (error == 0 && node->buffer == NULL) {
1016 TAILQ_INSERT_TAIL(&buffer->clist,
1018 node->buffer = buffer;
1021 buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1022 buffer = hammer_get_buffer(node->hmp, buf_offset,
1023 HAMMER_BUFSIZE, 0, &error);
1025 KKASSERT(error == 0);
1026 TAILQ_INSERT_TAIL(&buffer->clist,
1028 node->buffer = buffer;
1033 node->ondisk = (void *)((char *)buffer->ondisk +
1034 (node->node_offset & HAMMER_BUFMASK));
1036 (node->flags & HAMMER_NODE_CRCGOOD) == 0) {
1037 if (hammer_crc_test_btree(node->ondisk) == 0)
1038 Debugger("CRC FAILED: B-TREE NODE");
1039 node->flags |= HAMMER_NODE_CRCGOOD;
1044 hammer_unlock(&node->lock);
1049 * Safely reference a node, interlock against flushes via the IO subsystem.
1052 hammer_ref_node_safe(struct hammer_mount *hmp, hammer_node_cache_t cache,
1059 hammer_ref(&node->lock);
1063 *errorp = hammer_load_node(node, 0);
1065 hammer_rel_node(node);
1075 * Release a hammer_node. On the last release the node dereferences
1076 * its underlying buffer and may or may not be destroyed.
1079 hammer_rel_node(hammer_node_t node)
1081 hammer_buffer_t buffer;
1084 * If this isn't the last ref just decrement the ref count and
1087 if (node->lock.refs > 1) {
1088 hammer_unref(&node->lock);
1093 * If there is no ondisk info or no buffer the node failed to load,
1094 * remove the last reference and destroy the node.
1096 if (node->ondisk == NULL) {
1097 hammer_unref(&node->lock);
1098 hammer_flush_node(node);
1099 /* node is stale now */
1104 * Do not disassociate the node from the buffer if it represents
1105 * a modified B-Tree node that still needs its crc to be generated.
1107 if (node->flags & HAMMER_NODE_NEEDSCRC)
1111 * Do final cleanups and then either destroy the node and leave it
1112 * passively cached. The buffer reference is removed regardless.
1114 buffer = node->buffer;
1115 node->ondisk = NULL;
1117 if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1118 hammer_unref(&node->lock);
1119 hammer_rel_buffer(buffer, 0);
1126 hammer_unref(&node->lock);
1127 hammer_flush_node(node);
1129 hammer_rel_buffer(buffer, 0);
1133 * Free space on-media associated with a B-Tree node.
1136 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1138 KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1139 node->flags |= HAMMER_NODE_DELETED;
1140 hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1144 * Passively cache a referenced hammer_node. The caller may release
1145 * the node on return.
1148 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1151 * If the node is being deleted, don't cache it!
1153 if (node->flags & HAMMER_NODE_DELETED)
1155 if (cache->node == node)
1158 hammer_uncache_node(cache);
1159 if (node->flags & HAMMER_NODE_DELETED)
1162 TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1166 hammer_uncache_node(hammer_node_cache_t cache)
1170 if ((node = cache->node) != NULL) {
1171 TAILQ_REMOVE(&node->cache_list, cache, entry);
1173 if (TAILQ_EMPTY(&node->cache_list))
1174 hammer_flush_node(node);
1179 * Remove a node's cache references and destroy the node if it has no
1180 * other references or backing store.
1183 hammer_flush_node(hammer_node_t node)
1185 hammer_node_cache_t cache;
1186 hammer_buffer_t buffer;
1188 while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1189 TAILQ_REMOVE(&node->cache_list, cache, entry);
1192 if (node->lock.refs == 0 && node->ondisk == NULL) {
1193 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1194 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1195 if ((buffer = node->buffer) != NULL) {
1196 node->buffer = NULL;
1197 TAILQ_REMOVE(&buffer->clist, node, entry);
1198 /* buffer is unreferenced because ondisk is NULL */
1200 --hammer_count_nodes;
1201 kfree(node, M_HAMMER);
1206 * Flush passively cached B-Tree nodes associated with this buffer.
1207 * This is only called when the buffer is about to be destroyed, so
1208 * none of the nodes should have any references. The buffer is locked.
1210 * We may be interlocked with the buffer.
1213 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1217 while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1218 KKASSERT(node->ondisk == NULL);
1219 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1221 if (node->lock.refs == 0) {
1222 hammer_ref(&node->lock);
1223 node->flags |= HAMMER_NODE_FLUSH;
1224 hammer_rel_node(node);
1226 KKASSERT(node->loading != 0);
1227 KKASSERT(node->buffer != NULL);
1228 buffer = node->buffer;
1229 node->buffer = NULL;
1230 TAILQ_REMOVE(&buffer->clist, node, entry);
1231 /* buffer is unreferenced because ondisk is NULL */
1237 /************************************************************************
1239 ************************************************************************/
1242 * Allocate a B-Tree node.
1245 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1247 hammer_buffer_t buffer = NULL;
1248 hammer_node_t node = NULL;
1249 hammer_off_t node_offset;
1251 node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1252 sizeof(struct hammer_node_ondisk),
1255 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1256 hammer_modify_node_noundo(trans, node);
1257 bzero(node->ondisk, sizeof(*node->ondisk));
1258 hammer_modify_node_done(node);
1261 hammer_rel_buffer(buffer, 0);
1266 * Allocate data. If the address of a data buffer is supplied then
1267 * any prior non-NULL *data_bufferp will be released and *data_bufferp
1268 * will be set to the related buffer. The caller must release it when
1269 * finally done. The initial *data_bufferp should be set to NULL by
1272 * The caller is responsible for making hammer_modify*() calls on the
1276 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1277 u_int16_t rec_type, hammer_off_t *data_offsetp,
1278 struct hammer_buffer **data_bufferp, int *errorp)
1288 case HAMMER_RECTYPE_INODE:
1289 case HAMMER_RECTYPE_DIRENTRY:
1290 case HAMMER_RECTYPE_EXT:
1291 case HAMMER_RECTYPE_FIX:
1292 case HAMMER_RECTYPE_PFS:
1293 zone = HAMMER_ZONE_META_INDEX;
1295 case HAMMER_RECTYPE_DATA:
1296 case HAMMER_RECTYPE_DB:
1297 if (data_len <= HAMMER_BUFSIZE / 2) {
1298 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1300 data_len = (data_len + HAMMER_BUFMASK) &
1302 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1306 panic("hammer_alloc_data: rec_type %04x unknown",
1308 zone = 0; /* NOT REACHED */
1311 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1316 if (*errorp == 0 && data_bufferp) {
1318 data = hammer_bread_ext(trans->hmp, *data_offsetp,
1319 data_len, errorp, data_bufferp);
1320 KKASSERT(*errorp == 0);
1327 KKASSERT(*errorp == 0);
1332 * Sync dirty buffers to the media and clean-up any loose ends.
1334 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1335 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1338 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1340 struct hammer_sync_info info;
1343 info.waitfor = waitfor;
1344 if (waitfor == MNT_WAIT) {
1345 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1346 hammer_sync_scan1, hammer_sync_scan2, &info);
1348 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1349 hammer_sync_scan1, hammer_sync_scan2, &info);
1355 * Filesystem sync. If doing a synchronous sync make a second pass on
1356 * the vnodes in case any were already flushing during the first pass,
1357 * and activate the flusher twice (the second time brings the UNDO FIFO's
1358 * start position up to the end position after the first call).
1361 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1363 struct hammer_sync_info info;
1366 info.waitfor = MNT_NOWAIT;
1367 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1368 hammer_sync_scan1, hammer_sync_scan2, &info);
1369 if (info.error == 0 && waitfor == MNT_WAIT) {
1370 info.waitfor = waitfor;
1371 vmntvnodescan(hmp->mp, VMSC_GETVP,
1372 hammer_sync_scan1, hammer_sync_scan2, &info);
1374 if (waitfor == MNT_WAIT) {
1375 hammer_flusher_sync(hmp);
1376 hammer_flusher_sync(hmp);
1378 hammer_flusher_async(hmp);
1384 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1386 struct hammer_inode *ip;
1389 if (vp->v_type == VNON || ip == NULL ||
1390 ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1391 RB_EMPTY(&vp->v_rbdirty_tree))) {
1398 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1400 struct hammer_sync_info *info = data;
1401 struct hammer_inode *ip;
1405 if (vp->v_type == VNON || vp->v_type == VBAD ||
1406 ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1407 RB_EMPTY(&vp->v_rbdirty_tree))) {
1410 error = VOP_FSYNC(vp, info->waitfor);
1412 info->error = error;