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 1.62 2008/06/21 20:21:58 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->fsid_udev = dev2udev(vn_todev(volume->devvp));
212 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
213 (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
214 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
215 (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
221 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
223 volume->devvp->v_rdev->si_mountpoint = NULL;
224 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
225 hammer_free_volume(volume);
231 * This is called for each volume when updating the mount point from
232 * read-write to read-only or vise-versa.
235 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
238 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
239 if (volume->io.hmp->ronly) {
240 /* do not call vinvalbuf */
241 VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
242 VOP_CLOSE(volume->devvp, FREAD|FWRITE);
244 /* do not call vinvalbuf */
245 VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
246 VOP_CLOSE(volume->devvp, FREAD);
248 vn_unlock(volume->devvp);
254 * Unload and free a HAMMER volume. Must return >= 0 to continue scan
255 * so returns -1 on failure.
258 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
260 struct hammer_mount *hmp = volume->io.hmp;
261 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 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);
282 volume->ondisk = NULL;
284 if (volume->devvp->v_rdev &&
285 volume->devvp->v_rdev->si_mountpoint == hmp->mp
287 volume->devvp->v_rdev->si_mountpoint = NULL;
290 vinvalbuf(volume->devvp, 0, 0, 0);
291 VOP_CLOSE(volume->devvp, FREAD);
293 vinvalbuf(volume->devvp, V_SAVE, 0, 0);
294 VOP_CLOSE(volume->devvp, FREAD|FWRITE);
299 * Destroy the structure
301 RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
302 hammer_free_volume(volume);
308 hammer_free_volume(hammer_volume_t volume)
310 if (volume->vol_name) {
311 kfree(volume->vol_name, M_HAMMER);
312 volume->vol_name = NULL;
315 vrele(volume->devvp);
316 volume->devvp = NULL;
318 --hammer_count_volumes;
319 kfree(volume, M_HAMMER);
323 * Get a HAMMER volume. The volume must already exist.
326 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
328 struct hammer_volume *volume;
331 * Locate the volume structure
333 volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
334 if (volume == NULL) {
338 hammer_ref(&volume->io.lock);
341 * Deal with on-disk info
343 if (volume->ondisk == NULL || volume->io.loading) {
344 *errorp = hammer_load_volume(volume);
346 hammer_rel_volume(volume, 1);
356 hammer_ref_volume(hammer_volume_t volume)
360 hammer_ref(&volume->io.lock);
363 * Deal with on-disk info
365 if (volume->ondisk == NULL || volume->io.loading) {
366 error = hammer_load_volume(volume);
368 hammer_rel_volume(volume, 1);
376 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
378 hammer_volume_t volume;
380 volume = hmp->rootvol;
381 KKASSERT(volume != NULL);
382 hammer_ref(&volume->io.lock);
385 * Deal with on-disk info
387 if (volume->ondisk == NULL || volume->io.loading) {
388 *errorp = hammer_load_volume(volume);
390 hammer_rel_volume(volume, 1);
400 * Load a volume's on-disk information. The volume must be referenced and
401 * not locked. We temporarily acquire an exclusive lock to interlock
402 * against releases or multiple get's.
405 hammer_load_volume(hammer_volume_t volume)
409 ++volume->io.loading;
410 hammer_lock_ex(&volume->io.lock);
412 if (volume->ondisk == NULL) {
413 error = hammer_io_read(volume->devvp, &volume->io,
416 volume->ondisk = (void *)volume->io.bp->b_data;
420 --volume->io.loading;
421 hammer_unlock(&volume->io.lock);
426 * Release a volume. Call hammer_io_release on the last reference. We have
427 * to acquire an exclusive lock to interlock against volume->ondisk tests
428 * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
431 * Volumes are not unloaded from memory during normal operation.
434 hammer_rel_volume(hammer_volume_t volume, int flush)
437 if (volume->io.lock.refs == 1) {
438 ++volume->io.loading;
439 hammer_lock_ex(&volume->io.lock);
440 if (volume->io.lock.refs == 1) {
441 volume->ondisk = NULL;
442 hammer_io_release(&volume->io, flush);
444 --volume->io.loading;
445 hammer_unlock(&volume->io.lock);
447 hammer_unref(&volume->io.lock);
451 /************************************************************************
453 ************************************************************************
455 * Manage buffers. Currently all blockmap-backed zones are translated
456 * to zone-2 buffer offsets.
459 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
460 int bytes, int isnew, int *errorp)
462 hammer_buffer_t buffer;
463 hammer_volume_t volume;
464 hammer_off_t zone2_offset;
465 hammer_io_type_t iotype;
469 buf_offset &= ~HAMMER_BUFMASK64;
472 * Shortcut if the buffer is already cached
474 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
476 if (buffer->io.lock.refs == 0)
477 ++hammer_count_refedbufs;
478 hammer_ref(&buffer->io.lock);
481 * Onced refed the ondisk field will not be cleared by
484 if (buffer->ondisk && buffer->io.loading == 0) {
490 * The buffer is no longer loose if it has a ref, and
491 * cannot become loose once it gains a ref. Loose
492 * buffers will never be in a modified state. This should
493 * only occur on the 0->1 transition of refs.
495 * lose_list can be modified via a biodone() interrupt.
497 if (buffer->io.mod_list == &hmp->lose_list) {
498 crit_enter(); /* biodone race against list */
499 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
502 buffer->io.mod_list = NULL;
503 KKASSERT(buffer->io.modified == 0);
509 * What is the buffer class?
511 zone = HAMMER_ZONE_DECODE(buf_offset);
514 case HAMMER_ZONE_LARGE_DATA_INDEX:
515 case HAMMER_ZONE_SMALL_DATA_INDEX:
516 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
518 case HAMMER_ZONE_UNDO_INDEX:
519 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
521 case HAMMER_ZONE_META_INDEX:
524 * NOTE: inode data and directory entries are placed in this
525 * zone. inode atime/mtime is updated in-place and thus
526 * buffers containing inodes must be synchronized as
527 * meta-buffers, same as buffers containing B-Tree info.
529 iotype = HAMMER_STRUCTURE_META_BUFFER;
534 * Handle blockmap offset translations
536 if (zone >= HAMMER_ZONE_BTREE_INDEX) {
537 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
538 } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
539 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
541 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
542 zone2_offset = buf_offset;
549 * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
552 KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
553 HAMMER_ZONE_RAW_BUFFER);
554 vol_no = HAMMER_VOL_DECODE(zone2_offset);
555 volume = hammer_get_volume(hmp, vol_no, errorp);
559 KKASSERT(zone2_offset < volume->maxbuf_off);
562 * Allocate a new buffer structure. We will check for races later.
564 ++hammer_count_buffers;
565 buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
566 buffer->zone2_offset = zone2_offset;
567 buffer->zoneX_offset = buf_offset;
568 buffer->volume = volume;
570 hammer_io_init(&buffer->io, hmp, iotype);
571 buffer->io.offset = volume->ondisk->vol_buf_beg +
572 (zone2_offset & HAMMER_OFF_SHORT_MASK);
573 buffer->io.bytes = bytes;
574 TAILQ_INIT(&buffer->clist);
575 hammer_ref(&buffer->io.lock);
578 * Insert the buffer into the RB tree and handle late collisions.
580 if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
581 hammer_unref(&buffer->io.lock);
582 --hammer_count_buffers;
583 kfree(buffer, M_HAMMER);
586 ++hammer_count_refedbufs;
590 * Deal with on-disk info and loading races.
592 if (buffer->ondisk == NULL || buffer->io.loading) {
593 *errorp = hammer_load_buffer(buffer, isnew);
595 hammer_rel_buffer(buffer, 1);
605 * Destroy all buffers covering the specified zoneX offset range. This
606 * is called when the related blockmap layer2 entry is freed or when
607 * a direct write bypasses our buffer/buffer-cache subsystem.
609 * The buffers may be referenced by the caller itself. Setting reclaim
610 * will cause the buffer to be destroyed when it's ref count reaches zero.
613 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
614 hammer_off_t zone2_offset, int bytes)
616 hammer_buffer_t buffer;
617 hammer_volume_t volume;
621 vol_no = HAMMER_VOL_DECODE(zone2_offset);
622 volume = hammer_get_volume(hmp, vol_no, &error);
623 KKASSERT(error == 0);
626 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
629 KKASSERT(buffer->zone2_offset == zone2_offset);
630 hammer_io_clear_modify(&buffer->io, 1);
631 buffer->io.reclaim = 1;
632 KKASSERT(buffer->volume == volume);
633 if (buffer->io.lock.refs == 0)
634 hammer_unload_buffer(buffer, NULL);
636 hammer_io_inval(volume, zone2_offset);
638 base_offset += HAMMER_BUFSIZE;
639 zone2_offset += HAMMER_BUFSIZE;
640 bytes -= HAMMER_BUFSIZE;
642 hammer_rel_volume(volume, 0);
646 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
648 hammer_volume_t volume;
652 * Load the buffer's on-disk info
654 volume = buffer->volume;
655 ++buffer->io.loading;
656 hammer_lock_ex(&buffer->io.lock);
658 if (hammer_debug_io & 0x0001) {
659 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
660 buffer->zoneX_offset, buffer->zone2_offset, isnew,
664 if (buffer->ondisk == NULL) {
666 error = hammer_io_new(volume->devvp, &buffer->io);
668 error = hammer_io_read(volume->devvp, &buffer->io,
672 buffer->ondisk = (void *)buffer->io.bp->b_data;
674 error = hammer_io_new(volume->devvp, &buffer->io);
678 --buffer->io.loading;
679 hammer_unlock(&buffer->io.lock);
684 * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
687 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
689 ++hammer_count_refedbufs;
690 hammer_ref(&buffer->io.lock);
691 hammer_flush_buffer_nodes(buffer);
692 KKASSERT(buffer->io.lock.refs == 1);
693 hammer_rel_buffer(buffer, 2);
698 * Reference a buffer that is either already referenced or via a specially
699 * handled pointer (aka cursor->buffer).
702 hammer_ref_buffer(hammer_buffer_t buffer)
706 if (buffer->io.lock.refs == 0)
707 ++hammer_count_refedbufs;
708 hammer_ref(&buffer->io.lock);
711 * At this point a biodone() will not touch the buffer other then
712 * incidental bits. However, lose_list can be modified via
713 * a biodone() interrupt.
717 if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
719 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
720 buffer->io.mod_list = NULL;
724 if (buffer->ondisk == NULL || buffer->io.loading) {
725 error = hammer_load_buffer(buffer, 0);
727 hammer_rel_buffer(buffer, 1);
729 * NOTE: buffer pointer can become stale after
740 * Release a buffer. We have to deal with several places where
741 * another thread can ref the buffer.
743 * Only destroy the structure itself if the related buffer cache buffer
744 * was disassociated from it. This ties the management of the structure
745 * to the buffer cache subsystem. buffer->ondisk determines whether the
746 * embedded io is referenced or not.
749 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
751 hammer_volume_t volume;
755 if (buffer->io.lock.refs == 1) {
756 ++buffer->io.loading; /* force interlock check */
757 hammer_lock_ex(&buffer->io.lock);
758 if (buffer->io.lock.refs == 1) {
759 hammer_io_release(&buffer->io, flush);
761 if (buffer->io.lock.refs == 1)
762 --hammer_count_refedbufs;
764 if (buffer->io.bp == NULL &&
765 buffer->io.lock.refs == 1) {
769 * NOTE: It is impossible for any associated
770 * B-Tree nodes to have refs if the buffer
771 * has no additional refs.
773 RB_REMOVE(hammer_buf_rb_tree,
774 &buffer->io.hmp->rb_bufs_root,
776 volume = buffer->volume;
777 buffer->volume = NULL; /* sanity */
778 hammer_rel_volume(volume, 0);
779 hammer_io_clear_modlist(&buffer->io);
780 hammer_flush_buffer_nodes(buffer);
781 KKASSERT(TAILQ_EMPTY(&buffer->clist));
785 --buffer->io.loading;
786 hammer_unlock(&buffer->io.lock);
788 hammer_unref(&buffer->io.lock);
791 --hammer_count_buffers;
792 kfree(buffer, M_HAMMER);
797 * Access the filesystem buffer containing the specified hammer offset.
798 * buf_offset is a conglomeration of the volume number and vol_buf_beg
799 * relative buffer offset. It must also have bit 55 set to be valid.
800 * (see hammer_off_t in hammer_disk.h).
802 * Any prior buffer in *bufferp will be released and replaced by the
807 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
808 int *errorp, struct hammer_buffer **bufferp)
810 hammer_buffer_t buffer;
811 int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
813 buf_offset &= ~HAMMER_BUFMASK64;
814 KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
817 if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
818 buffer->zoneX_offset != buf_offset)) {
820 hammer_rel_buffer(buffer, 0);
821 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
828 * Return a pointer to the buffer data.
833 return((char *)buffer->ondisk + xoff);
837 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
838 int *errorp, struct hammer_buffer **bufferp)
840 return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
844 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
845 int *errorp, struct hammer_buffer **bufferp)
847 bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
848 return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
852 * Access the filesystem buffer containing the specified hammer offset.
853 * No disk read operation occurs. The result buffer may contain garbage.
855 * Any prior buffer in *bufferp will be released and replaced by the
858 * This function marks the buffer dirty but does not increment its
863 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
864 int *errorp, struct hammer_buffer **bufferp)
866 hammer_buffer_t buffer;
867 int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
869 buf_offset &= ~HAMMER_BUFMASK64;
872 if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
873 buffer->zoneX_offset != buf_offset)) {
875 hammer_rel_buffer(buffer, 0);
876 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
883 * Return a pointer to the buffer data.
888 return((char *)buffer->ondisk + xoff);
892 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
893 int *errorp, struct hammer_buffer **bufferp)
895 return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
899 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
900 int *errorp, struct hammer_buffer **bufferp)
902 bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
903 return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
906 /************************************************************************
908 ************************************************************************
910 * Manage B-Tree nodes. B-Tree nodes represent the primary indexing
911 * method used by the HAMMER filesystem.
913 * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
914 * associated with its buffer, and will only referenced the buffer while
915 * the node itself is referenced.
917 * A hammer_node can also be passively associated with other HAMMER
918 * structures, such as inodes, while retaining 0 references. These
919 * associations can be cleared backwards using a pointer-to-pointer in
922 * This allows the HAMMER implementation to cache hammer_nodes long-term
923 * and short-cut a great deal of the infrastructure's complexity. In
924 * most cases a cached node can be reacquired without having to dip into
925 * either the buffer or cluster management code.
927 * The caller must pass a referenced cluster on call and will retain
928 * ownership of the reference on return. The node will acquire its own
929 * additional references, if necessary.
932 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
933 int isnew, int *errorp)
937 KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
940 * Locate the structure, allocating one if necessary.
943 node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
945 ++hammer_count_nodes;
946 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
947 node->node_offset = node_offset;
949 TAILQ_INIT(&node->cache_list);
950 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
951 --hammer_count_nodes;
952 kfree(node, M_HAMMER);
956 hammer_ref(&node->lock);
960 *errorp = hammer_load_node(node, isnew);
962 hammer_rel_node(node);
969 * Reference an already-referenced node.
972 hammer_ref_node(hammer_node_t node)
974 KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
975 hammer_ref(&node->lock);
979 * Load a node's on-disk data reference.
982 hammer_load_node(hammer_node_t node, int isnew)
984 hammer_buffer_t buffer;
985 hammer_off_t buf_offset;
990 hammer_lock_ex(&node->lock);
991 if (node->ondisk == NULL) {
993 * This is a little confusing but the jist is that
994 * node->buffer determines whether the node is on
995 * the buffer's clist and node->ondisk determines
996 * whether the buffer is referenced.
998 * We could be racing a buffer release, in which case
999 * node->buffer may become NULL while we are blocked
1000 * referencing the buffer.
1002 if ((buffer = node->buffer) != NULL) {
1003 error = hammer_ref_buffer(buffer);
1004 if (error == 0 && node->buffer == NULL) {
1005 TAILQ_INSERT_TAIL(&buffer->clist,
1007 node->buffer = buffer;
1010 buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1011 buffer = hammer_get_buffer(node->hmp, buf_offset,
1012 HAMMER_BUFSIZE, 0, &error);
1014 KKASSERT(error == 0);
1015 TAILQ_INSERT_TAIL(&buffer->clist,
1017 node->buffer = buffer;
1022 node->ondisk = (void *)((char *)buffer->ondisk +
1023 (node->node_offset & HAMMER_BUFMASK));
1025 (node->flags & HAMMER_NODE_CRCGOOD) == 0) {
1026 if (hammer_crc_test_btree(node->ondisk) == 0)
1027 Debugger("CRC FAILED: B-TREE NODE");
1028 node->flags |= HAMMER_NODE_CRCGOOD;
1033 hammer_unlock(&node->lock);
1038 * Safely reference a node, interlock against flushes via the IO subsystem.
1041 hammer_ref_node_safe(struct hammer_mount *hmp, hammer_node_cache_t cache,
1048 hammer_ref(&node->lock);
1052 *errorp = hammer_load_node(node, 0);
1054 hammer_rel_node(node);
1064 * Release a hammer_node. On the last release the node dereferences
1065 * its underlying buffer and may or may not be destroyed.
1068 hammer_rel_node(hammer_node_t node)
1070 hammer_buffer_t buffer;
1073 * If this isn't the last ref just decrement the ref count and
1076 if (node->lock.refs > 1) {
1077 hammer_unref(&node->lock);
1082 * If there is no ondisk info or no buffer the node failed to load,
1083 * remove the last reference and destroy the node.
1085 if (node->ondisk == NULL) {
1086 hammer_unref(&node->lock);
1087 hammer_flush_node(node);
1088 /* node is stale now */
1093 * Do not disassociate the node from the buffer if it represents
1094 * a modified B-Tree node that still needs its crc to be generated.
1096 if (node->flags & HAMMER_NODE_NEEDSCRC)
1100 * Do final cleanups and then either destroy the node and leave it
1101 * passively cached. The buffer reference is removed regardless.
1103 buffer = node->buffer;
1104 node->ondisk = NULL;
1106 if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1107 hammer_unref(&node->lock);
1108 hammer_rel_buffer(buffer, 0);
1115 hammer_unref(&node->lock);
1116 hammer_flush_node(node);
1118 hammer_rel_buffer(buffer, 0);
1122 * Free space on-media associated with a B-Tree node.
1125 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1127 KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1128 node->flags |= HAMMER_NODE_DELETED;
1129 hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1133 * Passively cache a referenced hammer_node. The caller may release
1134 * the node on return.
1137 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1140 * If the node is being deleted, don't cache it!
1142 if (node->flags & HAMMER_NODE_DELETED)
1144 if (cache->node == node)
1147 hammer_uncache_node(cache);
1148 if (node->flags & HAMMER_NODE_DELETED)
1151 TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1155 hammer_uncache_node(hammer_node_cache_t cache)
1159 if ((node = cache->node) != NULL) {
1160 TAILQ_REMOVE(&node->cache_list, cache, entry);
1162 if (TAILQ_EMPTY(&node->cache_list))
1163 hammer_flush_node(node);
1168 * Remove a node's cache references and destroy the node if it has no
1169 * other references or backing store.
1172 hammer_flush_node(hammer_node_t node)
1174 hammer_node_cache_t cache;
1175 hammer_buffer_t buffer;
1177 while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1178 TAILQ_REMOVE(&node->cache_list, cache, entry);
1181 if (node->lock.refs == 0 && node->ondisk == NULL) {
1182 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1183 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1184 if ((buffer = node->buffer) != NULL) {
1185 node->buffer = NULL;
1186 TAILQ_REMOVE(&buffer->clist, node, entry);
1187 /* buffer is unreferenced because ondisk is NULL */
1189 --hammer_count_nodes;
1190 kfree(node, M_HAMMER);
1195 * Flush passively cached B-Tree nodes associated with this buffer.
1196 * This is only called when the buffer is about to be destroyed, so
1197 * none of the nodes should have any references. The buffer is locked.
1199 * We may be interlocked with the buffer.
1202 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1206 while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1207 KKASSERT(node->ondisk == NULL);
1208 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1210 if (node->lock.refs == 0) {
1211 hammer_ref(&node->lock);
1212 node->flags |= HAMMER_NODE_FLUSH;
1213 hammer_rel_node(node);
1215 KKASSERT(node->loading != 0);
1216 KKASSERT(node->buffer != NULL);
1217 buffer = node->buffer;
1218 node->buffer = NULL;
1219 TAILQ_REMOVE(&buffer->clist, node, entry);
1220 /* buffer is unreferenced because ondisk is NULL */
1226 /************************************************************************
1228 ************************************************************************/
1231 * Allocate a B-Tree node.
1234 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1236 hammer_buffer_t buffer = NULL;
1237 hammer_node_t node = NULL;
1238 hammer_off_t node_offset;
1240 node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1241 sizeof(struct hammer_node_ondisk),
1244 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1245 hammer_modify_node_noundo(trans, node);
1246 bzero(node->ondisk, sizeof(*node->ondisk));
1247 hammer_modify_node_done(node);
1250 hammer_rel_buffer(buffer, 0);
1255 * Allocate data. If the address of a data buffer is supplied then
1256 * any prior non-NULL *data_bufferp will be released and *data_bufferp
1257 * will be set to the related buffer. The caller must release it when
1258 * finally done. The initial *data_bufferp should be set to NULL by
1261 * The caller is responsible for making hammer_modify*() calls on the
1265 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1266 u_int16_t rec_type, hammer_off_t *data_offsetp,
1267 struct hammer_buffer **data_bufferp, int *errorp)
1277 case HAMMER_RECTYPE_INODE:
1278 case HAMMER_RECTYPE_DIRENTRY:
1279 case HAMMER_RECTYPE_EXT:
1280 case HAMMER_RECTYPE_FIX:
1281 zone = HAMMER_ZONE_META_INDEX;
1283 case HAMMER_RECTYPE_DATA:
1284 case HAMMER_RECTYPE_DB:
1285 if (data_len <= HAMMER_BUFSIZE / 2) {
1286 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1288 data_len = (data_len + HAMMER_BUFMASK) &
1290 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1294 panic("hammer_alloc_data: rec_type %04x unknown",
1296 zone = 0; /* NOT REACHED */
1299 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1304 if (*errorp == 0 && data_bufferp) {
1306 data = hammer_bread_ext(trans->hmp, *data_offsetp,
1307 data_len, errorp, data_bufferp);
1308 KKASSERT(*errorp == 0);
1315 KKASSERT(*errorp == 0);
1320 * Sync dirty buffers to the media and clean-up any loose ends.
1322 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1323 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1326 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1328 struct hammer_sync_info info;
1331 info.waitfor = waitfor;
1332 if (waitfor == MNT_WAIT) {
1333 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1334 hammer_sync_scan1, hammer_sync_scan2, &info);
1336 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1337 hammer_sync_scan1, hammer_sync_scan2, &info);
1343 * Filesystem sync. If doing a synchronous sync make a second pass on
1344 * the vnodes in case any were already flushing during the first pass,
1345 * and activate the flusher twice (the second time brings the UNDO FIFO's
1346 * start position up to the end position after the first call).
1349 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1351 struct hammer_sync_info info;
1354 info.waitfor = MNT_NOWAIT;
1355 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1356 hammer_sync_scan1, hammer_sync_scan2, &info);
1357 if (info.error == 0 && waitfor == MNT_WAIT) {
1358 info.waitfor = waitfor;
1359 vmntvnodescan(hmp->mp, VMSC_GETVP,
1360 hammer_sync_scan1, hammer_sync_scan2, &info);
1362 if (waitfor == MNT_WAIT) {
1363 hammer_flusher_sync(hmp);
1364 hammer_flusher_sync(hmp);
1366 hammer_flusher_async(hmp);
1372 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1374 struct hammer_inode *ip;
1377 if (vp->v_type == VNON || ip == NULL ||
1378 ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1379 RB_EMPTY(&vp->v_rbdirty_tree))) {
1386 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1388 struct hammer_sync_info *info = data;
1389 struct hammer_inode *ip;
1393 if (vp->v_type == VNON || vp->v_type == VBAD ||
1394 ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1395 RB_EMPTY(&vp->v_rbdirty_tree))) {
1398 error = VOP_FSYNC(vp, info->waitfor);
1400 info->error = error;