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>
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_buffer structures are indexed via their zoneX_offset, not
68 hammer_buf_rb_compare(hammer_buffer_t buf1
, hammer_buffer_t buf2
)
70 if (buf1
->zoneX_offset
< buf2
->zoneX_offset
)
72 if (buf1
->zoneX_offset
> buf2
->zoneX_offset
)
78 hammer_nod_rb_compare(hammer_node_t node1
, hammer_node_t node2
)
80 if (node1
->node_offset
< node2
->node_offset
)
82 if (node1
->node_offset
> node2
->node_offset
)
87 RB_GENERATE2(hammer_vol_rb_tree
, hammer_volume
, rb_node
,
88 hammer_vol_rb_compare
, int32_t, vol_no
);
89 RB_GENERATE2(hammer_buf_rb_tree
, hammer_buffer
, rb_node
,
90 hammer_buf_rb_compare
, hammer_off_t
, zoneX_offset
);
91 RB_GENERATE2(hammer_nod_rb_tree
, hammer_node
, rb_node
,
92 hammer_nod_rb_compare
, hammer_off_t
, node_offset
);
94 /************************************************************************
96 ************************************************************************
98 * Load a HAMMER volume by name. Returns 0 on success or a positive error
99 * code on failure. Volumes must be loaded at mount time, get_volume() will
100 * not load a new volume.
102 * Calls made to hammer_load_volume() or single-threaded
105 hammer_install_volume(struct hammer_mount
*hmp
, const char *volname
,
109 hammer_volume_t volume
;
110 struct hammer_volume_ondisk
*ondisk
;
111 struct nlookupdata nd
;
112 struct buf
*bp
= NULL
;
118 ronly
= ((mp
->mnt_flag
& MNT_RDONLY
) ? 1 : 0);
121 * Allocate a volume structure
123 ++hammer_count_volumes
;
124 volume
= kmalloc(sizeof(*volume
), hmp
->m_misc
, M_WAITOK
|M_ZERO
);
125 volume
->vol_name
= kstrdup(volname
, hmp
->m_misc
);
126 volume
->io
.hmp
= hmp
; /* bootstrap */
127 hammer_io_init(&volume
->io
, volume
, HAMMER_STRUCTURE_VOLUME
);
128 volume
->io
.offset
= 0LL;
129 volume
->io
.bytes
= HAMMER_BUFSIZE
;
132 * Get the device vnode
135 error
= nlookup_init(&nd
, volume
->vol_name
, UIO_SYSSPACE
, NLC_FOLLOW
);
137 error
= nlookup(&nd
);
139 error
= cache_vref(&nd
.nl_nch
, nd
.nl_cred
, &volume
->devvp
);
143 volume
->devvp
= devvp
;
147 if (vn_isdisk(volume
->devvp
, &error
)) {
148 error
= vfs_mountedon(volume
->devvp
);
152 count_udev(volume
->devvp
->v_umajor
, volume
->devvp
->v_uminor
) > 0) {
156 vn_lock(volume
->devvp
, LK_EXCLUSIVE
| LK_RETRY
);
157 error
= vinvalbuf(volume
->devvp
, V_SAVE
, 0, 0);
159 error
= VOP_OPEN(volume
->devvp
,
160 (ronly
? FREAD
: FREAD
|FWRITE
),
163 vn_unlock(volume
->devvp
);
166 hammer_free_volume(volume
);
169 volume
->devvp
->v_rdev
->si_mountpoint
= mp
;
173 * Extract the volume number from the volume header and do various
176 error
= bread(volume
->devvp
, 0LL, HAMMER_BUFSIZE
, &bp
);
179 ondisk
= (void *)bp
->b_data
;
180 if (ondisk
->vol_signature
!= HAMMER_FSBUF_VOLUME
) {
181 kprintf("hammer_mount: volume %s has an invalid header\n",
186 volume
->vol_no
= ondisk
->vol_no
;
187 volume
->buffer_base
= ondisk
->vol_buf_beg
;
188 volume
->vol_flags
= ondisk
->vol_flags
;
189 volume
->nblocks
= ondisk
->vol_nblocks
;
190 volume
->maxbuf_off
= HAMMER_ENCODE_RAW_BUFFER(volume
->vol_no
,
191 ondisk
->vol_buf_end
- ondisk
->vol_buf_beg
);
192 volume
->maxraw_off
= ondisk
->vol_buf_end
;
194 if (RB_EMPTY(&hmp
->rb_vols_root
)) {
195 hmp
->fsid
= ondisk
->vol_fsid
;
196 } else if (bcmp(&hmp
->fsid
, &ondisk
->vol_fsid
, sizeof(uuid_t
))) {
197 kprintf("hammer_mount: volume %s's fsid does not match "
198 "other volumes\n", volume
->vol_name
);
204 * Insert the volume structure into the red-black tree.
206 if (RB_INSERT(hammer_vol_rb_tree
, &hmp
->rb_vols_root
, volume
)) {
207 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
208 volume
->vol_name
, volume
->vol_no
);
213 * Set the root volume . HAMMER special cases rootvol the structure.
214 * We do not hold a ref because this would prevent related I/O
215 * from being flushed.
217 if (error
== 0 && ondisk
->vol_rootvol
== ondisk
->vol_no
) {
218 hmp
->rootvol
= volume
;
219 hmp
->nvolumes
= ondisk
->vol_count
;
224 hmp
->mp
->mnt_stat
.f_blocks
+= ondisk
->vol0_stat_bigblocks
*
225 (HAMMER_LARGEBLOCK_SIZE
/ HAMMER_BUFSIZE
);
226 hmp
->mp
->mnt_vstat
.f_blocks
+= ondisk
->vol0_stat_bigblocks
*
227 (HAMMER_LARGEBLOCK_SIZE
/ HAMMER_BUFSIZE
);
233 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
235 volume
->devvp
->v_rdev
->si_mountpoint
= NULL
;
236 VOP_CLOSE(volume
->devvp
, ronly
? FREAD
: FREAD
|FWRITE
);
237 hammer_free_volume(volume
);
243 * This is called for each volume when updating the mount point from
244 * read-write to read-only or vise-versa.
247 hammer_adjust_volume_mode(hammer_volume_t volume
, void *data __unused
)
250 vn_lock(volume
->devvp
, LK_EXCLUSIVE
| LK_RETRY
);
251 if (volume
->io
.hmp
->ronly
) {
252 /* do not call vinvalbuf */
253 VOP_OPEN(volume
->devvp
, FREAD
, FSCRED
, NULL
);
254 VOP_CLOSE(volume
->devvp
, FREAD
|FWRITE
);
256 /* do not call vinvalbuf */
257 VOP_OPEN(volume
->devvp
, FREAD
|FWRITE
, FSCRED
, NULL
);
258 VOP_CLOSE(volume
->devvp
, FREAD
);
260 vn_unlock(volume
->devvp
);
266 * Unload and free a HAMMER volume. Must return >= 0 to continue scan
267 * so returns -1 on failure.
270 hammer_unload_volume(hammer_volume_t volume
, void *data __unused
)
272 hammer_mount_t hmp
= volume
->io
.hmp
;
273 int ronly
= ((hmp
->mp
->mnt_flag
& MNT_RDONLY
) ? 1 : 0);
277 * Clean up the root volume pointer, which is held unlocked in hmp.
279 if (hmp
->rootvol
== volume
)
283 * We must not flush a dirty buffer to disk on umount. It should
284 * have already been dealt with by the flusher, or we may be in
285 * catastrophic failure.
287 hammer_io_clear_modify(&volume
->io
, 1);
288 volume
->io
.waitdep
= 1;
289 bp
= hammer_io_release(&volume
->io
, 1);
292 * Clean up the persistent ref ioerror might have on the volume
294 if (volume
->io
.ioerror
) {
295 volume
->io
.ioerror
= 0;
296 hammer_unref(&volume
->io
.lock
);
300 * There should be no references on the volume, no clusters, and
303 KKASSERT(volume
->io
.lock
.refs
== 0);
307 volume
->ondisk
= NULL
;
309 if (volume
->devvp
->v_rdev
&&
310 volume
->devvp
->v_rdev
->si_mountpoint
== hmp
->mp
312 volume
->devvp
->v_rdev
->si_mountpoint
= NULL
;
316 * Make sure we don't sync anything to disk if we
317 * are in read-only mode (1) or critically-errored
318 * (2). Note that there may be dirty buffers in
319 * normal read-only mode from crash recovery.
321 vinvalbuf(volume
->devvp
, 0, 0, 0);
322 VOP_CLOSE(volume
->devvp
, FREAD
);
325 * Normal termination, save any dirty buffers
326 * (XXX there really shouldn't be any).
328 vinvalbuf(volume
->devvp
, V_SAVE
, 0, 0);
329 VOP_CLOSE(volume
->devvp
, FREAD
|FWRITE
);
334 * Destroy the structure
336 RB_REMOVE(hammer_vol_rb_tree
, &hmp
->rb_vols_root
, volume
);
337 hammer_free_volume(volume
);
343 hammer_free_volume(hammer_volume_t volume
)
345 hammer_mount_t hmp
= volume
->io
.hmp
;
347 if (volume
->vol_name
) {
348 kfree(volume
->vol_name
, hmp
->m_misc
);
349 volume
->vol_name
= NULL
;
352 vrele(volume
->devvp
);
353 volume
->devvp
= NULL
;
355 --hammer_count_volumes
;
356 kfree(volume
, hmp
->m_misc
);
360 * Get a HAMMER volume. The volume must already exist.
363 hammer_get_volume(struct hammer_mount
*hmp
, int32_t vol_no
, int *errorp
)
365 struct hammer_volume
*volume
;
368 * Locate the volume structure
370 volume
= RB_LOOKUP(hammer_vol_rb_tree
, &hmp
->rb_vols_root
, vol_no
);
371 if (volume
== NULL
) {
375 hammer_ref(&volume
->io
.lock
);
378 * Deal with on-disk info
380 if (volume
->ondisk
== NULL
|| volume
->io
.loading
) {
381 *errorp
= hammer_load_volume(volume
);
383 hammer_rel_volume(volume
, 1);
393 hammer_ref_volume(hammer_volume_t volume
)
397 hammer_ref(&volume
->io
.lock
);
400 * Deal with on-disk info
402 if (volume
->ondisk
== NULL
|| volume
->io
.loading
) {
403 error
= hammer_load_volume(volume
);
405 hammer_rel_volume(volume
, 1);
413 hammer_get_root_volume(struct hammer_mount
*hmp
, int *errorp
)
415 hammer_volume_t volume
;
417 volume
= hmp
->rootvol
;
418 KKASSERT(volume
!= NULL
);
419 hammer_ref(&volume
->io
.lock
);
422 * Deal with on-disk info
424 if (volume
->ondisk
== NULL
|| volume
->io
.loading
) {
425 *errorp
= hammer_load_volume(volume
);
427 hammer_rel_volume(volume
, 1);
437 * Load a volume's on-disk information. The volume must be referenced and
438 * not locked. We temporarily acquire an exclusive lock to interlock
439 * against releases or multiple get's.
442 hammer_load_volume(hammer_volume_t volume
)
446 ++volume
->io
.loading
;
447 hammer_lock_ex(&volume
->io
.lock
);
449 if (volume
->ondisk
== NULL
) {
450 error
= hammer_io_read(volume
->devvp
, &volume
->io
,
453 volume
->ondisk
= (void *)volume
->io
.bp
->b_data
;
457 --volume
->io
.loading
;
458 hammer_unlock(&volume
->io
.lock
);
463 * Release a volume. Call hammer_io_release on the last reference. We have
464 * to acquire an exclusive lock to interlock against volume->ondisk tests
465 * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
468 * Volumes are not unloaded from memory during normal operation.
471 hammer_rel_volume(hammer_volume_t volume
, int flush
)
473 struct buf
*bp
= NULL
;
476 if (volume
->io
.lock
.refs
== 1) {
477 ++volume
->io
.loading
;
478 hammer_lock_ex(&volume
->io
.lock
);
479 if (volume
->io
.lock
.refs
== 1) {
480 volume
->ondisk
= NULL
;
481 bp
= hammer_io_release(&volume
->io
, flush
);
483 --volume
->io
.loading
;
484 hammer_unlock(&volume
->io
.lock
);
486 hammer_unref(&volume
->io
.lock
);
493 hammer_mountcheck_volumes(struct hammer_mount
*hmp
)
498 for (i
= 0; i
< hmp
->nvolumes
; ++i
) {
499 vol
= RB_LOOKUP(hammer_vol_rb_tree
, &hmp
->rb_vols_root
, i
);
506 /************************************************************************
508 ************************************************************************
510 * Manage buffers. Currently all blockmap-backed zones are direct-mapped
511 * to zone-2 buffer offsets, without a translation stage. However, the
512 * hammer_buffer structure is indexed by its zoneX_offset, not its
515 * The proper zone must be maintained throughout the code-base all the way
516 * through to the big-block allocator, or routines like hammer_del_buffers()
517 * will not be able to locate all potentially conflicting buffers.
520 hammer_get_buffer(hammer_mount_t hmp
, hammer_off_t buf_offset
,
521 int bytes
, int isnew
, int *errorp
)
523 hammer_buffer_t buffer
;
524 hammer_volume_t volume
;
525 hammer_off_t zone2_offset
;
526 hammer_io_type_t iotype
;
530 buf_offset
&= ~HAMMER_BUFMASK64
;
533 * Shortcut if the buffer is already cached
535 buffer
= RB_LOOKUP(hammer_buf_rb_tree
, &hmp
->rb_bufs_root
, buf_offset
);
537 if (buffer
->io
.lock
.refs
== 0)
538 ++hammer_count_refedbufs
;
539 hammer_ref(&buffer
->io
.lock
);
542 * Once refed the ondisk field will not be cleared by
545 if (buffer
->ondisk
&& buffer
->io
.loading
== 0) {
551 * The buffer is no longer loose if it has a ref, and
552 * cannot become loose once it gains a ref. Loose
553 * buffers will never be in a modified state. This should
554 * only occur on the 0->1 transition of refs.
556 * lose_list can be modified via a biodone() interrupt.
558 if (buffer
->io
.mod_list
== &hmp
->lose_list
) {
559 crit_enter(); /* biodone race against list */
560 TAILQ_REMOVE(buffer
->io
.mod_list
, &buffer
->io
,
563 buffer
->io
.mod_list
= NULL
;
564 KKASSERT(buffer
->io
.modified
== 0);
570 * What is the buffer class?
572 zone
= HAMMER_ZONE_DECODE(buf_offset
);
575 case HAMMER_ZONE_LARGE_DATA_INDEX
:
576 case HAMMER_ZONE_SMALL_DATA_INDEX
:
577 iotype
= HAMMER_STRUCTURE_DATA_BUFFER
;
579 case HAMMER_ZONE_UNDO_INDEX
:
580 iotype
= HAMMER_STRUCTURE_UNDO_BUFFER
;
582 case HAMMER_ZONE_META_INDEX
:
585 * NOTE: inode data and directory entries are placed in this
586 * zone. inode atime/mtime is updated in-place and thus
587 * buffers containing inodes must be synchronized as
588 * meta-buffers, same as buffers containing B-Tree info.
590 iotype
= HAMMER_STRUCTURE_META_BUFFER
;
595 * Handle blockmap offset translations
597 if (zone
>= HAMMER_ZONE_BTREE_INDEX
) {
598 zone2_offset
= hammer_blockmap_lookup(hmp
, buf_offset
, errorp
);
599 } else if (zone
== HAMMER_ZONE_UNDO_INDEX
) {
600 zone2_offset
= hammer_undo_lookup(hmp
, buf_offset
, errorp
);
602 KKASSERT(zone
== HAMMER_ZONE_RAW_BUFFER_INDEX
);
603 zone2_offset
= buf_offset
;
610 * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
613 KKASSERT((zone2_offset
& HAMMER_OFF_ZONE_MASK
) ==
614 HAMMER_ZONE_RAW_BUFFER
);
615 vol_no
= HAMMER_VOL_DECODE(zone2_offset
);
616 volume
= hammer_get_volume(hmp
, vol_no
, errorp
);
620 KKASSERT(zone2_offset
< volume
->maxbuf_off
);
623 * Allocate a new buffer structure. We will check for races later.
625 ++hammer_count_buffers
;
626 buffer
= kmalloc(sizeof(*buffer
), hmp
->m_misc
,
627 M_WAITOK
|M_ZERO
|M_USE_RESERVE
);
628 buffer
->zone2_offset
= zone2_offset
;
629 buffer
->zoneX_offset
= buf_offset
;
631 hammer_io_init(&buffer
->io
, volume
, iotype
);
632 buffer
->io
.offset
= volume
->ondisk
->vol_buf_beg
+
633 (zone2_offset
& HAMMER_OFF_SHORT_MASK
);
634 buffer
->io
.bytes
= bytes
;
635 TAILQ_INIT(&buffer
->clist
);
636 hammer_ref(&buffer
->io
.lock
);
639 * Insert the buffer into the RB tree and handle late collisions.
641 if (RB_INSERT(hammer_buf_rb_tree
, &hmp
->rb_bufs_root
, buffer
)) {
642 hammer_unref(&buffer
->io
.lock
); /* safety */
643 --hammer_count_buffers
;
644 hammer_rel_volume(volume
, 0);
645 buffer
->io
.volume
= NULL
; /* safety */
646 kfree(buffer
, hmp
->m_misc
);
649 ++hammer_count_refedbufs
;
653 * Deal with on-disk info and loading races.
655 if (buffer
->ondisk
== NULL
|| buffer
->io
.loading
) {
656 *errorp
= hammer_load_buffer(buffer
, isnew
);
658 hammer_rel_buffer(buffer
, 1);
668 * This is used by the direct-read code to deal with large-data buffers
669 * created by the reblocker and mirror-write code. The direct-read code
670 * bypasses the HAMMER buffer subsystem and so any aliased dirty or write-
671 * running hammer buffers must be fully synced to disk before we can issue
674 * This code path is not considered critical as only the rebocker and
675 * mirror-write code will create large-data buffers via the HAMMER buffer
676 * subsystem. They do that because they operate at the B-Tree level and
677 * do not access the vnode/inode structures.
680 hammer_sync_buffers(hammer_mount_t hmp
, hammer_off_t base_offset
, int bytes
)
682 hammer_buffer_t buffer
;
685 KKASSERT((base_offset
& HAMMER_OFF_ZONE_MASK
) ==
686 HAMMER_ZONE_LARGE_DATA
);
689 buffer
= RB_LOOKUP(hammer_buf_rb_tree
, &hmp
->rb_bufs_root
,
691 if (buffer
&& (buffer
->io
.modified
|| buffer
->io
.running
)) {
692 error
= hammer_ref_buffer(buffer
);
694 hammer_io_wait(&buffer
->io
);
695 if (buffer
->io
.modified
) {
696 hammer_io_write_interlock(&buffer
->io
);
697 hammer_io_flush(&buffer
->io
);
698 hammer_io_done_interlock(&buffer
->io
);
699 hammer_io_wait(&buffer
->io
);
701 hammer_rel_buffer(buffer
, 0);
704 base_offset
+= HAMMER_BUFSIZE
;
705 bytes
-= HAMMER_BUFSIZE
;
710 * Destroy all buffers covering the specified zoneX offset range. This
711 * is called when the related blockmap layer2 entry is freed or when
712 * a direct write bypasses our buffer/buffer-cache subsystem.
714 * The buffers may be referenced by the caller itself. Setting reclaim
715 * will cause the buffer to be destroyed when it's ref count reaches zero.
717 * Return 0 on success, EAGAIN if some buffers could not be destroyed due
718 * to additional references held by other threads, or some other (typically
722 hammer_del_buffers(hammer_mount_t hmp
, hammer_off_t base_offset
,
723 hammer_off_t zone2_offset
, int bytes
,
724 int report_conflicts
)
726 hammer_buffer_t buffer
;
727 hammer_volume_t volume
;
732 vol_no
= HAMMER_VOL_DECODE(zone2_offset
);
733 volume
= hammer_get_volume(hmp
, vol_no
, &ret_error
);
734 KKASSERT(ret_error
== 0);
737 buffer
= RB_LOOKUP(hammer_buf_rb_tree
, &hmp
->rb_bufs_root
,
740 error
= hammer_ref_buffer(buffer
);
741 if (error
== 0 && buffer
->io
.lock
.refs
!= 1) {
743 hammer_rel_buffer(buffer
, 0);
746 KKASSERT(buffer
->zone2_offset
== zone2_offset
);
747 hammer_io_clear_modify(&buffer
->io
, 1);
748 buffer
->io
.reclaim
= 1;
749 buffer
->io
.waitdep
= 1;
750 KKASSERT(buffer
->io
.volume
== volume
);
751 hammer_rel_buffer(buffer
, 0);
754 error
= hammer_io_inval(volume
, zone2_offset
);
758 if (report_conflicts
|| (hammer_debug_general
& 0x8000))
759 kprintf("hammer_del_buffers: unable to invalidate %016llx buffer=%p rep=%d\n", base_offset
, buffer
, report_conflicts
);
761 base_offset
+= HAMMER_BUFSIZE
;
762 zone2_offset
+= HAMMER_BUFSIZE
;
763 bytes
-= HAMMER_BUFSIZE
;
765 hammer_rel_volume(volume
, 0);
770 hammer_load_buffer(hammer_buffer_t buffer
, int isnew
)
772 hammer_volume_t volume
;
776 * Load the buffer's on-disk info
778 volume
= buffer
->io
.volume
;
779 ++buffer
->io
.loading
;
780 hammer_lock_ex(&buffer
->io
.lock
);
782 if (hammer_debug_io
& 0x0001) {
783 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
784 buffer
->zoneX_offset
, buffer
->zone2_offset
, isnew
,
788 if (buffer
->ondisk
== NULL
) {
790 error
= hammer_io_new(volume
->devvp
, &buffer
->io
);
792 error
= hammer_io_read(volume
->devvp
, &buffer
->io
,
796 buffer
->ondisk
= (void *)buffer
->io
.bp
->b_data
;
798 error
= hammer_io_new(volume
->devvp
, &buffer
->io
);
802 --buffer
->io
.loading
;
803 hammer_unlock(&buffer
->io
.lock
);
808 * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
809 * This routine is only called during unmount.
812 hammer_unload_buffer(hammer_buffer_t buffer
, void *data __unused
)
815 * Clean up the persistent ref ioerror might have on the buffer
816 * and acquire a ref (steal ioerror's if we can).
818 if (buffer
->io
.ioerror
) {
819 buffer
->io
.ioerror
= 0;
821 if (buffer
->io
.lock
.refs
== 0)
822 ++hammer_count_refedbufs
;
823 hammer_ref(&buffer
->io
.lock
);
827 * We must not flush a dirty buffer to disk on umount. It should
828 * have already been dealt with by the flusher, or we may be in
829 * catastrophic failure.
831 hammer_io_clear_modify(&buffer
->io
, 1);
832 hammer_flush_buffer_nodes(buffer
);
833 KKASSERT(buffer
->io
.lock
.refs
== 1);
834 hammer_rel_buffer(buffer
, 2);
839 * Reference a buffer that is either already referenced or via a specially
840 * handled pointer (aka cursor->buffer).
843 hammer_ref_buffer(hammer_buffer_t buffer
)
847 if (buffer
->io
.lock
.refs
== 0)
848 ++hammer_count_refedbufs
;
849 hammer_ref(&buffer
->io
.lock
);
852 * At this point a biodone() will not touch the buffer other then
853 * incidental bits. However, lose_list can be modified via
854 * a biodone() interrupt.
858 if (buffer
->io
.mod_list
== &buffer
->io
.hmp
->lose_list
) {
860 TAILQ_REMOVE(buffer
->io
.mod_list
, &buffer
->io
, mod_entry
);
861 buffer
->io
.mod_list
= NULL
;
865 if (buffer
->ondisk
== NULL
|| buffer
->io
.loading
) {
866 error
= hammer_load_buffer(buffer
, 0);
868 hammer_rel_buffer(buffer
, 1);
870 * NOTE: buffer pointer can become stale after
881 * Release a buffer. We have to deal with several places where
882 * another thread can ref the buffer.
884 * Only destroy the structure itself if the related buffer cache buffer
885 * was disassociated from it. This ties the management of the structure
886 * to the buffer cache subsystem. buffer->ondisk determines whether the
887 * embedded io is referenced or not.
890 hammer_rel_buffer(hammer_buffer_t buffer
, int flush
)
892 hammer_volume_t volume
;
894 struct buf
*bp
= NULL
;
897 hmp
= buffer
->io
.hmp
;
900 if (buffer
->io
.lock
.refs
== 1) {
901 ++buffer
->io
.loading
; /* force interlock check */
902 hammer_lock_ex(&buffer
->io
.lock
);
903 if (buffer
->io
.lock
.refs
== 1) {
904 bp
= hammer_io_release(&buffer
->io
, flush
);
906 if (buffer
->io
.lock
.refs
== 1)
907 --hammer_count_refedbufs
;
909 if (buffer
->io
.bp
== NULL
&&
910 buffer
->io
.lock
.refs
== 1) {
914 * NOTE: It is impossible for any associated
915 * B-Tree nodes to have refs if the buffer
916 * has no additional refs.
918 RB_REMOVE(hammer_buf_rb_tree
,
919 &buffer
->io
.hmp
->rb_bufs_root
,
921 volume
= buffer
->io
.volume
;
922 buffer
->io
.volume
= NULL
; /* sanity */
923 hammer_rel_volume(volume
, 0);
924 hammer_io_clear_modlist(&buffer
->io
);
925 hammer_flush_buffer_nodes(buffer
);
926 KKASSERT(TAILQ_EMPTY(&buffer
->clist
));
930 --buffer
->io
.loading
;
931 hammer_unlock(&buffer
->io
.lock
);
933 hammer_unref(&buffer
->io
.lock
);
938 --hammer_count_buffers
;
939 kfree(buffer
, hmp
->m_misc
);
944 * Access the filesystem buffer containing the specified hammer offset.
945 * buf_offset is a conglomeration of the volume number and vol_buf_beg
946 * relative buffer offset. It must also have bit 55 set to be valid.
947 * (see hammer_off_t in hammer_disk.h).
949 * Any prior buffer in *bufferp will be released and replaced by the
952 * NOTE: The buffer is indexed via its zoneX_offset but we allow the
953 * passed cached *bufferp to match against either zoneX or zone2.
957 _hammer_bread(hammer_mount_t hmp
, hammer_off_t buf_offset
, int bytes
,
958 int *errorp
, struct hammer_buffer
**bufferp
)
960 hammer_buffer_t buffer
;
961 int32_t xoff
= (int32_t)buf_offset
& HAMMER_BUFMASK
;
963 buf_offset
&= ~HAMMER_BUFMASK64
;
964 KKASSERT((buf_offset
& HAMMER_OFF_ZONE_MASK
) != 0);
967 if (buffer
== NULL
|| (buffer
->zone2_offset
!= buf_offset
&&
968 buffer
->zoneX_offset
!= buf_offset
)) {
970 hammer_rel_buffer(buffer
, 0);
971 buffer
= hammer_get_buffer(hmp
, buf_offset
, bytes
, 0, errorp
);
978 * Return a pointer to the buffer data.
983 return((char *)buffer
->ondisk
+ xoff
);
987 hammer_bread(hammer_mount_t hmp
, hammer_off_t buf_offset
,
988 int *errorp
, struct hammer_buffer
**bufferp
)
990 return(_hammer_bread(hmp
, buf_offset
, HAMMER_BUFSIZE
, errorp
, bufferp
));
994 hammer_bread_ext(hammer_mount_t hmp
, hammer_off_t buf_offset
, int bytes
,
995 int *errorp
, struct hammer_buffer
**bufferp
)
997 bytes
= (bytes
+ HAMMER_BUFMASK
) & ~HAMMER_BUFMASK
;
998 return(_hammer_bread(hmp
, buf_offset
, bytes
, errorp
, bufferp
));
1002 * Access the filesystem buffer containing the specified hammer offset.
1003 * No disk read operation occurs. The result buffer may contain garbage.
1005 * Any prior buffer in *bufferp will be released and replaced by the
1008 * This function marks the buffer dirty but does not increment its
1009 * modify_refs count.
1013 _hammer_bnew(hammer_mount_t hmp
, hammer_off_t buf_offset
, int bytes
,
1014 int *errorp
, struct hammer_buffer
**bufferp
)
1016 hammer_buffer_t buffer
;
1017 int32_t xoff
= (int32_t)buf_offset
& HAMMER_BUFMASK
;
1019 buf_offset
&= ~HAMMER_BUFMASK64
;
1022 if (buffer
== NULL
|| (buffer
->zone2_offset
!= buf_offset
&&
1023 buffer
->zoneX_offset
!= buf_offset
)) {
1025 hammer_rel_buffer(buffer
, 0);
1026 buffer
= hammer_get_buffer(hmp
, buf_offset
, bytes
, 1, errorp
);
1033 * Return a pointer to the buffer data.
1038 return((char *)buffer
->ondisk
+ xoff
);
1042 hammer_bnew(hammer_mount_t hmp
, hammer_off_t buf_offset
,
1043 int *errorp
, struct hammer_buffer
**bufferp
)
1045 return(_hammer_bnew(hmp
, buf_offset
, HAMMER_BUFSIZE
, errorp
, bufferp
));
1049 hammer_bnew_ext(hammer_mount_t hmp
, hammer_off_t buf_offset
, int bytes
,
1050 int *errorp
, struct hammer_buffer
**bufferp
)
1052 bytes
= (bytes
+ HAMMER_BUFMASK
) & ~HAMMER_BUFMASK
;
1053 return(_hammer_bnew(hmp
, buf_offset
, bytes
, errorp
, bufferp
));
1056 /************************************************************************
1058 ************************************************************************
1060 * Manage B-Tree nodes. B-Tree nodes represent the primary indexing
1061 * method used by the HAMMER filesystem.
1063 * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
1064 * associated with its buffer, and will only referenced the buffer while
1065 * the node itself is referenced.
1067 * A hammer_node can also be passively associated with other HAMMER
1068 * structures, such as inodes, while retaining 0 references. These
1069 * associations can be cleared backwards using a pointer-to-pointer in
1072 * This allows the HAMMER implementation to cache hammer_nodes long-term
1073 * and short-cut a great deal of the infrastructure's complexity. In
1074 * most cases a cached node can be reacquired without having to dip into
1075 * either the buffer or cluster management code.
1077 * The caller must pass a referenced cluster on call and will retain
1078 * ownership of the reference on return. The node will acquire its own
1079 * additional references, if necessary.
1082 hammer_get_node(hammer_transaction_t trans
, hammer_off_t node_offset
,
1083 int isnew
, int *errorp
)
1085 hammer_mount_t hmp
= trans
->hmp
;
1088 KKASSERT((node_offset
& HAMMER_OFF_ZONE_MASK
) == HAMMER_ZONE_BTREE
);
1091 * Locate the structure, allocating one if necessary.
1094 node
= RB_LOOKUP(hammer_nod_rb_tree
, &hmp
->rb_nods_root
, node_offset
);
1096 ++hammer_count_nodes
;
1097 node
= kmalloc(sizeof(*node
), hmp
->m_misc
, M_WAITOK
|M_ZERO
|M_USE_RESERVE
);
1098 node
->node_offset
= node_offset
;
1100 TAILQ_INIT(&node
->cursor_list
);
1101 TAILQ_INIT(&node
->cache_list
);
1102 if (RB_INSERT(hammer_nod_rb_tree
, &hmp
->rb_nods_root
, node
)) {
1103 --hammer_count_nodes
;
1104 kfree(node
, hmp
->m_misc
);
1108 hammer_ref(&node
->lock
);
1112 *errorp
= hammer_load_node(node
, isnew
);
1113 trans
->flags
|= HAMMER_TRANSF_DIDIO
;
1116 hammer_rel_node(node
);
1123 * Reference an already-referenced node.
1126 hammer_ref_node(hammer_node_t node
)
1128 KKASSERT(node
->lock
.refs
> 0 && node
->ondisk
!= NULL
);
1129 hammer_ref(&node
->lock
);
1133 * Load a node's on-disk data reference.
1136 hammer_load_node(hammer_node_t node
, int isnew
)
1138 hammer_buffer_t buffer
;
1139 hammer_off_t buf_offset
;
1144 hammer_lock_ex(&node
->lock
);
1145 if (node
->ondisk
== NULL
) {
1147 * This is a little confusing but the jist is that
1148 * node->buffer determines whether the node is on
1149 * the buffer's clist and node->ondisk determines
1150 * whether the buffer is referenced.
1152 * We could be racing a buffer release, in which case
1153 * node->buffer may become NULL while we are blocked
1154 * referencing the buffer.
1156 if ((buffer
= node
->buffer
) != NULL
) {
1157 error
= hammer_ref_buffer(buffer
);
1158 if (error
== 0 && node
->buffer
== NULL
) {
1159 TAILQ_INSERT_TAIL(&buffer
->clist
,
1161 node
->buffer
= buffer
;
1164 buf_offset
= node
->node_offset
& ~HAMMER_BUFMASK64
;
1165 buffer
= hammer_get_buffer(node
->hmp
, buf_offset
,
1166 HAMMER_BUFSIZE
, 0, &error
);
1168 KKASSERT(error
== 0);
1169 TAILQ_INSERT_TAIL(&buffer
->clist
,
1171 node
->buffer
= buffer
;
1176 node
->ondisk
= (void *)((char *)buffer
->ondisk
+
1177 (node
->node_offset
& HAMMER_BUFMASK
));
1179 (node
->flags
& HAMMER_NODE_CRCGOOD
) == 0) {
1180 if (hammer_crc_test_btree(node
->ondisk
) == 0)
1181 Debugger("CRC FAILED: B-TREE NODE");
1182 node
->flags
|= HAMMER_NODE_CRCGOOD
;
1187 hammer_unlock(&node
->lock
);
1192 * Safely reference a node, interlock against flushes via the IO subsystem.
1195 hammer_ref_node_safe(struct hammer_mount
*hmp
, hammer_node_cache_t cache
,
1202 hammer_ref(&node
->lock
);
1206 *errorp
= hammer_load_node(node
, 0);
1208 hammer_rel_node(node
);
1218 * Release a hammer_node. On the last release the node dereferences
1219 * its underlying buffer and may or may not be destroyed.
1222 hammer_rel_node(hammer_node_t node
)
1224 hammer_buffer_t buffer
;
1227 * If this isn't the last ref just decrement the ref count and
1230 if (node
->lock
.refs
> 1) {
1231 hammer_unref(&node
->lock
);
1236 * If there is no ondisk info or no buffer the node failed to load,
1237 * remove the last reference and destroy the node.
1239 if (node
->ondisk
== NULL
) {
1240 hammer_unref(&node
->lock
);
1241 hammer_flush_node(node
);
1242 /* node is stale now */
1247 * Do not disassociate the node from the buffer if it represents
1248 * a modified B-Tree node that still needs its crc to be generated.
1250 if (node
->flags
& HAMMER_NODE_NEEDSCRC
)
1254 * Do final cleanups and then either destroy the node and leave it
1255 * passively cached. The buffer reference is removed regardless.
1257 buffer
= node
->buffer
;
1258 node
->ondisk
= NULL
;
1260 if ((node
->flags
& HAMMER_NODE_FLUSH
) == 0) {
1261 hammer_unref(&node
->lock
);
1262 hammer_rel_buffer(buffer
, 0);
1269 hammer_unref(&node
->lock
);
1270 hammer_flush_node(node
);
1272 hammer_rel_buffer(buffer
, 0);
1276 * Free space on-media associated with a B-Tree node.
1279 hammer_delete_node(hammer_transaction_t trans
, hammer_node_t node
)
1281 KKASSERT((node
->flags
& HAMMER_NODE_DELETED
) == 0);
1282 node
->flags
|= HAMMER_NODE_DELETED
;
1283 hammer_blockmap_free(trans
, node
->node_offset
, sizeof(*node
->ondisk
));
1287 * Passively cache a referenced hammer_node. The caller may release
1288 * the node on return.
1291 hammer_cache_node(hammer_node_cache_t cache
, hammer_node_t node
)
1294 * If the node doesn't exist, or is being deleted, don't cache it!
1296 * The node can only ever be NULL in the I/O failure path.
1298 if (node
== NULL
|| (node
->flags
& HAMMER_NODE_DELETED
))
1300 if (cache
->node
== node
)
1303 hammer_uncache_node(cache
);
1304 if (node
->flags
& HAMMER_NODE_DELETED
)
1307 TAILQ_INSERT_TAIL(&node
->cache_list
, cache
, entry
);
1311 hammer_uncache_node(hammer_node_cache_t cache
)
1315 if ((node
= cache
->node
) != NULL
) {
1316 TAILQ_REMOVE(&node
->cache_list
, cache
, entry
);
1318 if (TAILQ_EMPTY(&node
->cache_list
))
1319 hammer_flush_node(node
);
1324 * Remove a node's cache references and destroy the node if it has no
1325 * other references or backing store.
1328 hammer_flush_node(hammer_node_t node
)
1330 hammer_node_cache_t cache
;
1331 hammer_buffer_t buffer
;
1332 hammer_mount_t hmp
= node
->hmp
;
1334 while ((cache
= TAILQ_FIRST(&node
->cache_list
)) != NULL
) {
1335 TAILQ_REMOVE(&node
->cache_list
, cache
, entry
);
1338 if (node
->lock
.refs
== 0 && node
->ondisk
== NULL
) {
1339 KKASSERT((node
->flags
& HAMMER_NODE_NEEDSCRC
) == 0);
1340 RB_REMOVE(hammer_nod_rb_tree
, &node
->hmp
->rb_nods_root
, node
);
1341 if ((buffer
= node
->buffer
) != NULL
) {
1342 node
->buffer
= NULL
;
1343 TAILQ_REMOVE(&buffer
->clist
, node
, entry
);
1344 /* buffer is unreferenced because ondisk is NULL */
1346 --hammer_count_nodes
;
1347 kfree(node
, hmp
->m_misc
);
1352 * Flush passively cached B-Tree nodes associated with this buffer.
1353 * This is only called when the buffer is about to be destroyed, so
1354 * none of the nodes should have any references. The buffer is locked.
1356 * We may be interlocked with the buffer.
1359 hammer_flush_buffer_nodes(hammer_buffer_t buffer
)
1363 while ((node
= TAILQ_FIRST(&buffer
->clist
)) != NULL
) {
1364 KKASSERT(node
->ondisk
== NULL
);
1365 KKASSERT((node
->flags
& HAMMER_NODE_NEEDSCRC
) == 0);
1367 if (node
->lock
.refs
== 0) {
1368 hammer_ref(&node
->lock
);
1369 node
->flags
|= HAMMER_NODE_FLUSH
;
1370 hammer_rel_node(node
);
1372 KKASSERT(node
->loading
!= 0);
1373 KKASSERT(node
->buffer
!= NULL
);
1374 buffer
= node
->buffer
;
1375 node
->buffer
= NULL
;
1376 TAILQ_REMOVE(&buffer
->clist
, node
, entry
);
1377 /* buffer is unreferenced because ondisk is NULL */
1383 /************************************************************************
1385 ************************************************************************/
1388 * Allocate a B-Tree node.
1391 hammer_alloc_btree(hammer_transaction_t trans
, int *errorp
)
1393 hammer_buffer_t buffer
= NULL
;
1394 hammer_node_t node
= NULL
;
1395 hammer_off_t node_offset
;
1397 node_offset
= hammer_blockmap_alloc(trans
, HAMMER_ZONE_BTREE_INDEX
,
1398 sizeof(struct hammer_node_ondisk
),
1401 node
= hammer_get_node(trans
, node_offset
, 1, errorp
);
1402 hammer_modify_node_noundo(trans
, node
);
1403 bzero(node
->ondisk
, sizeof(*node
->ondisk
));
1404 hammer_modify_node_done(node
);
1407 hammer_rel_buffer(buffer
, 0);
1412 * Allocate data. If the address of a data buffer is supplied then
1413 * any prior non-NULL *data_bufferp will be released and *data_bufferp
1414 * will be set to the related buffer. The caller must release it when
1415 * finally done. The initial *data_bufferp should be set to NULL by
1418 * The caller is responsible for making hammer_modify*() calls on the
1422 hammer_alloc_data(hammer_transaction_t trans
, int32_t data_len
,
1423 u_int16_t rec_type
, hammer_off_t
*data_offsetp
,
1424 struct hammer_buffer
**data_bufferp
, int *errorp
)
1434 case HAMMER_RECTYPE_INODE
:
1435 case HAMMER_RECTYPE_DIRENTRY
:
1436 case HAMMER_RECTYPE_EXT
:
1437 case HAMMER_RECTYPE_FIX
:
1438 case HAMMER_RECTYPE_PFS
:
1439 zone
= HAMMER_ZONE_META_INDEX
;
1441 case HAMMER_RECTYPE_DATA
:
1442 case HAMMER_RECTYPE_DB
:
1443 if (data_len
<= HAMMER_BUFSIZE
/ 2) {
1444 zone
= HAMMER_ZONE_SMALL_DATA_INDEX
;
1446 data_len
= (data_len
+ HAMMER_BUFMASK
) &
1448 zone
= HAMMER_ZONE_LARGE_DATA_INDEX
;
1452 panic("hammer_alloc_data: rec_type %04x unknown",
1454 zone
= 0; /* NOT REACHED */
1457 *data_offsetp
= hammer_blockmap_alloc(trans
, zone
,
1462 if (*errorp
== 0 && data_bufferp
) {
1464 data
= hammer_bread_ext(trans
->hmp
, *data_offsetp
,
1465 data_len
, errorp
, data_bufferp
);
1476 * Sync dirty buffers to the media and clean-up any loose ends.
1478 * These functions do not start the flusher going, they simply
1479 * queue everything up to the flusher.
1481 static int hammer_sync_scan1(struct mount
*mp
, struct vnode
*vp
, void *data
);
1482 static int hammer_sync_scan2(struct mount
*mp
, struct vnode
*vp
, void *data
);
1485 hammer_queue_inodes_flusher(hammer_mount_t hmp
, int waitfor
)
1487 struct hammer_sync_info info
;
1490 info
.waitfor
= waitfor
;
1491 if (waitfor
== MNT_WAIT
) {
1492 vmntvnodescan(hmp
->mp
, VMSC_GETVP
|VMSC_ONEPASS
,
1493 hammer_sync_scan1
, hammer_sync_scan2
, &info
);
1495 vmntvnodescan(hmp
->mp
, VMSC_GETVP
|VMSC_ONEPASS
|VMSC_NOWAIT
,
1496 hammer_sync_scan1
, hammer_sync_scan2
, &info
);
1502 * Filesystem sync. If doing a synchronous sync make a second pass on
1503 * the vnodes in case any were already flushing during the first pass,
1504 * and activate the flusher twice (the second time brings the UNDO FIFO's
1505 * start position up to the end position after the first call).
1508 hammer_sync_hmp(hammer_mount_t hmp
, int waitfor
)
1510 struct hammer_sync_info info
;
1513 info
.waitfor
= MNT_NOWAIT
;
1514 vmntvnodescan(hmp
->mp
, VMSC_GETVP
|VMSC_NOWAIT
,
1515 hammer_sync_scan1
, hammer_sync_scan2
, &info
);
1516 if (info
.error
== 0 && waitfor
== MNT_WAIT
) {
1517 info
.waitfor
= waitfor
;
1518 vmntvnodescan(hmp
->mp
, VMSC_GETVP
,
1519 hammer_sync_scan1
, hammer_sync_scan2
, &info
);
1521 if (waitfor
== MNT_WAIT
) {
1522 hammer_flusher_sync(hmp
);
1523 hammer_flusher_sync(hmp
);
1525 hammer_flusher_async(hmp
, NULL
);
1526 hammer_flusher_async(hmp
, NULL
);
1532 hammer_sync_scan1(struct mount
*mp
, struct vnode
*vp
, void *data
)
1534 struct hammer_inode
*ip
;
1537 if (vp
->v_type
== VNON
|| ip
== NULL
||
1538 ((ip
->flags
& HAMMER_INODE_MODMASK
) == 0 &&
1539 RB_EMPTY(&vp
->v_rbdirty_tree
))) {
1546 hammer_sync_scan2(struct mount
*mp
, struct vnode
*vp
, void *data
)
1548 struct hammer_sync_info
*info
= data
;
1549 struct hammer_inode
*ip
;
1553 if (vp
->v_type
== VNON
|| vp
->v_type
== VBAD
||
1554 ((ip
->flags
& HAMMER_INODE_MODMASK
) == 0 &&
1555 RB_EMPTY(&vp
->v_rbdirty_tree
))) {
1558 error
= VOP_FSYNC(vp
, MNT_NOWAIT
);
1560 info
->error
= error
;