2 * Copyright (c) 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_recover.c,v 1.29 2008/07/26 05:36:21 dillon Exp $
39 static int hammer_check_tail_signature(hammer_fifo_tail_t tail,
40 hammer_off_t end_off);
41 static int hammer_check_head_signature(hammer_fifo_head_t head,
42 hammer_off_t beg_off);
43 static void hammer_recover_copy_undo(hammer_off_t undo_offset,
44 char *src, char *dst, int bytes);
45 static hammer_fifo_any_t hammer_recover_scan_fwd(hammer_mount_t hmp,
46 hammer_volume_t root_volume,
47 hammer_off_t *scan_offsetp,
48 int *errorp, struct hammer_buffer **bufferp);
49 static hammer_fifo_any_t hammer_recover_scan_rev(hammer_mount_t hmp,
50 hammer_volume_t root_volume,
51 hammer_off_t *scan_offsetp,
52 int *errorp, struct hammer_buffer **bufferp);
54 static void hammer_recover_debug_dump(int w, char *buf, int bytes);
56 static int hammer_recover_undo(hammer_mount_t hmp, hammer_volume_t root_volume,
57 hammer_fifo_undo_t undo);
60 * Recover filesystem meta-data on mount. This procedure figures out the
61 * UNDO FIFO range and runs the UNDOs backwards. The FIFO pointers are not
62 * resynchronized by this procedure.
64 * This procedure is run near the beginning of the mount sequence, before
65 * any B-Tree or high-level accesses are enabled, and is responsible for
66 * restoring the meta-data to a consistent state. High level HAMMER data
67 * structures (such as the B-Tree) cannot be accessed here.
69 * NOTE: No information from the root volume has been cached in the
70 * hammer_mount structure yet, so we need to access the root volume's
76 hammer_recover_stage1(hammer_mount_t hmp, hammer_volume_t root_volume)
78 hammer_blockmap_t rootmap;
79 hammer_buffer_t buffer;
80 hammer_off_t scan_offset;
81 hammer_off_t scan_offset_save;
83 hammer_fifo_any_t head;
84 hammer_off_t first_offset;
85 hammer_off_t last_offset;
90 * Examine the UNDO FIFO indices in the volume header.
92 rootmap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
93 first_offset = rootmap->first_offset;
94 last_offset = rootmap->next_offset;
98 if (first_offset > rootmap->alloc_offset ||
99 last_offset > rootmap->alloc_offset) {
100 kprintf("HAMMER(%s) Illegal UNDO FIFO index range "
101 "%016jx, %016jx limit %016jx\n",
102 root_volume->ondisk->vol_name,
103 (intmax_t)first_offset,
104 (intmax_t)last_offset,
105 (intmax_t)rootmap->alloc_offset);
111 * In HAMMER version 4+ filesystems the volume header does NOT
112 * contain definitive UNDO FIFO state. In particular, the
113 * rootmap->next_offset may not be indexed completely to the
114 * end of the active UNDO FIFO.
116 if (hmp->version >= HAMMER_VOL_VERSION_FOUR) {
118 * To find the definitive range we must first scan backwards
119 * from first_offset to locate the first real record and
120 * extract the sequence number from it. This record is not
121 * part of the active undo space.
123 scan_offset = first_offset;
127 head = hammer_recover_scan_rev(hmp, root_volume,
132 if (head->head.hdr_type != HAMMER_HEAD_TYPE_PAD) {
133 seqno = head->head.hdr_seq;
138 kprintf("HAMMER(%s) meta-data recovery failure "
139 "during seqno backscan\n",
140 root_volume->ondisk->vol_name);
145 * Scan forwards from first_offset and (seqno+1) looking
146 * for a sequence space discontinuity. This denotes the
147 * end of the active FIFO area.
149 * NOTE: For the case where the FIFO is empty the very first
150 * record we find will be discontinuous.
152 * NOTE: Do not include trailing PADs in the scan range,
153 * and remember the returned scan_offset after a
154 * fwd iteration points to the end of the returned
157 kprintf("HAMMER(%s) meta-data recovery check seqno=%08x\n",
158 root_volume->ondisk->vol_name,
161 scan_offset = first_offset;
162 scan_offset_save = scan_offset;
165 head = hammer_recover_scan_fwd(hmp, root_volume,
170 if (head->head.hdr_type != HAMMER_HEAD_TYPE_PAD) {
171 if (seqno != head->head.hdr_seq) {
172 scan_offset = scan_offset_save;
175 scan_offset_save = scan_offset;
181 * If the forward scan is grossly ahead of last_offset
182 * then something is wrong. last_offset is supposed
185 if (last_offset >= scan_offset) {
186 bytes = last_offset - scan_offset;
188 bytes = rootmap->alloc_offset - scan_offset +
189 (last_offset & HAMMER_OFF_LONG_MASK);
192 (rootmap->alloc_offset & HAMMER_OFF_LONG_MASK) *
194 kprintf("HAMMER(%s) meta-data forward scan is "
195 "grossly beyond the last_offset in "
196 "the volume header, this can't be "
198 root_volume->ondisk->vol_name);
206 * Store the seqno. This will be the next seqno we lay down
207 * when generating new UNDOs.
209 hmp->undo_seqno = seqno;
211 kprintf("HAMMER(%s) meta-data recovery failure "
212 "during seqno fwdscan\n",
213 root_volume->ondisk->vol_name);
216 last_offset = scan_offset;
217 kprintf("HAMMER(%s) meta-data recovery range %016jx-%016jx "
218 "(invol %016jx) endseqno=%08x\n",
219 root_volume->ondisk->vol_name,
220 (intmax_t)first_offset,
221 (intmax_t)last_offset,
222 (intmax_t)rootmap->next_offset,
227 * Calculate the size of the active portion of the FIFO. If the
228 * FIFO is empty the filesystem is clean and no further action is
231 if (last_offset >= first_offset) {
232 bytes = last_offset - first_offset;
234 bytes = rootmap->alloc_offset - first_offset +
235 (last_offset & HAMMER_OFF_LONG_MASK);
242 kprintf("HAMMER(%s) Start meta-data recovery %016jx - %016jx "
243 "(%jd bytes of UNDO)%s\n",
244 root_volume->ondisk->vol_name,
245 (intmax_t)first_offset,
246 (intmax_t)last_offset,
248 (hmp->ronly ? " (RO)" : "(RW)"));
249 if (bytes > (rootmap->alloc_offset & HAMMER_OFF_LONG_MASK)) {
250 kprintf("Undo size is absurd, unable to mount\n");
256 * Scan the UNDOs backwards.
258 scan_offset = last_offset;
260 while ((int64_t)bytes > 0) {
261 KKASSERT(scan_offset != first_offset);
262 head = hammer_recover_scan_rev(hmp, root_volume,
263 &scan_offset, &error, &buffer);
266 error = hammer_recover_undo(hmp, root_volume, &head->undo);
268 kprintf("HAMMER(%s) UNDO record at %016jx failed\n",
269 root_volume->ondisk->vol_name,
270 (intmax_t)scan_offset - head->head.hdr_size);
273 bytes -= head->head.hdr_size;
276 * If too many dirty buffers have built up we have to flush'm
277 * out. As long as we do not flush out the volume header
278 * a crash here should not cause any problems.
280 * buffer must be released so the flush can assert that
281 * all buffers are idle.
283 if (hammer_flusher_meta_limit(hmp)) {
285 hammer_rel_buffer(buffer, 0);
288 if (hmp->ronly == 0) {
289 hammer_recover_flush_buffers(hmp, root_volume,
291 kprintf("HAMMER(%s) Continuing recovery\n",
292 root_volume->ondisk->vol_name);
294 kprintf("HAMMER(%s) Recovery failure: Insufficient buffer cache to hold dirty buffers on read-only mount!\n",
295 root_volume->ondisk->vol_name);
303 hammer_rel_buffer(buffer, 0);
308 * After completely flushing all the recovered buffers the volume
309 * header will also be flushed.
311 if (root_volume->io.recovered == 0) {
312 hammer_ref_volume(root_volume);
313 root_volume->io.recovered = 1;
317 * Finish up flushing (or discarding) recovered buffers. FIFO
318 * indices in the volume header are updated to the actual undo
319 * range but will not be collapsed until stage 2.
322 hammer_modify_volume(NULL, root_volume, NULL, 0);
323 rootmap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
324 rootmap->first_offset = first_offset;
325 rootmap->next_offset = last_offset;
326 hammer_modify_volume_done(root_volume);
328 hammer_recover_flush_buffers(hmp, root_volume, 1);
330 hammer_recover_flush_buffers(hmp, root_volume, -1);
332 kprintf("HAMMER(%s) End meta-data recovery\n",
333 root_volume->ondisk->vol_name);
338 * Execute redo operations
340 * This procedure is run at the end of the mount sequence, after the hammer
341 * mount structure has been completely initialized but before the filesystem
342 * goes live. It can access standard cursors, the B-Tree, flush the
343 * filesystem, and so forth.
345 * This code may only be called for read-write mounts or when a mount
346 * switches from read-only to read-write.
348 * The stage1 code will have already calculated the correct FIFO range
349 * and stored it in the rootmap.
352 hammer_recover_stage2(hammer_mount_t hmp, hammer_volume_t root_volume)
354 hammer_blockmap_t rootmap;
355 hammer_buffer_t buffer;
356 hammer_off_t scan_offset;
358 hammer_fifo_any_t head;
359 hammer_off_t first_offset;
360 hammer_off_t last_offset;
364 * Stage 2 can only be run on a RW mount, or when the mount is
365 * switched from RO to RW. It must be run only once.
367 KKASSERT(hmp->ronly == 0);
369 if (hmp->hflags & HMNT_STAGE2)
371 hmp->hflags |= HMNT_STAGE2;
374 * Examine the UNDO FIFO. If it is empty the filesystem is clean
375 * and no action need be taken.
377 rootmap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
378 first_offset = rootmap->first_offset;
379 last_offset = rootmap->next_offset;
380 if (first_offset == last_offset)
383 if (last_offset >= first_offset) {
384 bytes = last_offset - first_offset;
386 bytes = rootmap->alloc_offset - first_offset +
387 (last_offset & HAMMER_OFF_LONG_MASK);
389 kprintf("HAMMER(%s) Start redo recovery %016jx - %016jx "
390 "(%jd bytes of UNDO)%s\n",
391 root_volume->ondisk->vol_name,
392 (intmax_t)first_offset,
393 (intmax_t)last_offset,
395 (hmp->ronly ? " (RO)" : "(RW)"));
396 if (bytes > (rootmap->alloc_offset & HAMMER_OFF_LONG_MASK)) {
397 kprintf("Undo size is absurd, unable to mount\n");
402 * Scan the REDOs forwards.
404 scan_offset = first_offset;
408 KKASSERT(scan_offset != last_offset);
410 head = hammer_recover_scan_fwd(hmp, root_volume,
411 &scan_offset, &error, &buffer);
416 error = hammer_recover_redo(hmp, root_volume, &head->redo);
419 kprintf("HAMMER(%s) UNDO record at %016jx failed\n",
420 root_volume->ondisk->vol_name,
421 (intmax_t)scan_offset - head->head.hdr_size);
424 bytes -= head->head.hdr_size;
427 hammer_rel_buffer(buffer, 0);
432 * Finish up flushing (or discarding) recovered buffers by executing
433 * a normal flush cycle. Setting HMNT_UNDO_DIRTY bypasses degenerate
434 * case tests and forces the flush in order to update the FIFO indices.
436 * If a crash occurs during the flush the entire undo/redo will be
437 * re-run during recovery on the next mount.
440 if (rootmap->first_offset != rootmap->next_offset)
441 hmp->hflags |= HMNT_UNDO_DIRTY;
442 hammer_flusher_sync(hmp);
444 kprintf("HAMMER(%s) End redo recovery\n",
445 root_volume->ondisk->vol_name);
450 * Scan backwards from *scan_offsetp, return the FIFO record prior to the
451 * record at *scan_offsetp or NULL if an error occured.
453 * On return *scan_offsetp will be the offset of the returned record.
456 hammer_recover_scan_rev(hammer_mount_t hmp, hammer_volume_t root_volume,
457 hammer_off_t *scan_offsetp,
458 int *errorp, struct hammer_buffer **bufferp)
460 hammer_off_t scan_offset;
461 hammer_blockmap_t rootmap;
462 hammer_fifo_any_t head;
463 hammer_fifo_tail_t tail;
465 rootmap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
466 scan_offset = *scan_offsetp;
468 if (hammer_debug_general & 0x0080)
469 kprintf("rev scan_offset %016jx\n", (intmax_t)scan_offset);
470 if (scan_offset == HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0))
471 scan_offset = rootmap->alloc_offset;
472 if (scan_offset - sizeof(*tail) <
473 HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0)) {
474 kprintf("HAMMER(%s) UNDO record at %016jx FIFO underflow\n",
475 root_volume->ondisk->vol_name,
476 (intmax_t)scan_offset);
480 tail = hammer_bread(hmp, scan_offset - sizeof(*tail),
483 kprintf("HAMMER(%s) Unable to read UNDO TAIL "
485 root_volume->ondisk->vol_name,
486 (intmax_t)scan_offset - sizeof(*tail));
490 if (hammer_check_tail_signature(tail, scan_offset) != 0) {
491 kprintf("HAMMER(%s) Illegal UNDO TAIL signature "
493 root_volume->ondisk->vol_name,
494 (intmax_t)scan_offset - sizeof(*tail));
498 head = (void *)((char *)tail + sizeof(*tail) - tail->tail_size);
499 *scan_offsetp = scan_offset - head->head.hdr_size;
505 * Scan forwards from *scan_offsetp, return the FIFO record or NULL if
508 * On return *scan_offsetp will be the offset of the record following
509 * the returned record.
512 hammer_recover_scan_fwd(hammer_mount_t hmp, hammer_volume_t root_volume,
513 hammer_off_t *scan_offsetp,
514 int *errorp, struct hammer_buffer **bufferp)
516 hammer_off_t scan_offset;
517 hammer_blockmap_t rootmap;
518 hammer_fifo_any_t head;
520 rootmap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
521 scan_offset = *scan_offsetp;
523 if (hammer_debug_general & 0x0080)
524 kprintf("fwd scan_offset %016jx\n", (intmax_t)scan_offset);
525 if (scan_offset == rootmap->alloc_offset)
526 scan_offset = HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
528 head = hammer_bread(hmp, scan_offset, errorp, bufferp);
530 kprintf("HAMMER(%s) Unable to read UNDO HEAD at %016jx\n",
531 root_volume->ondisk->vol_name,
532 (intmax_t)scan_offset);
536 if (hammer_check_head_signature(&head->head, scan_offset) != 0) {
537 kprintf("HAMMER(%s) Illegal UNDO TAIL signature "
539 root_volume->ondisk->vol_name,
540 (intmax_t)scan_offset);
544 scan_offset += head->head.hdr_size;
545 if (scan_offset == rootmap->alloc_offset)
546 scan_offset = HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
547 *scan_offsetp = scan_offset;
553 * Helper function for hammer_check_{head,tail}_signature(). Check stuff
554 * once the head and tail has been established.
556 * This function validates the entire FIFO record wrapper.
560 _hammer_check_signature(hammer_fifo_head_t head, hammer_fifo_tail_t tail,
561 hammer_off_t beg_off)
563 hammer_off_t end_off;
568 * Check signatures. The tail signature is allowed to be the
569 * head signature only for 8-byte PADs.
571 if (head->hdr_signature != HAMMER_HEAD_SIGNATURE) {
572 kprintf("HAMMER: FIFO record bad head signature "
578 if (head->hdr_size < HAMMER_HEAD_ALIGN ||
579 (head->hdr_size & HAMMER_HEAD_ALIGN_MASK)) {
580 kprintf("HAMMER: FIFO record unaligned or bad size"
586 end_off = beg_off + head->hdr_size;
588 if (head->hdr_type != HAMMER_HEAD_TYPE_PAD ||
589 (size_t)(end_off - beg_off) != sizeof(*tail)) {
590 if (head->hdr_type != tail->tail_type) {
591 kprintf("HAMMER: FIFO record head/tail type mismatch "
592 "%04x %04x at %016jx\n",
593 head->hdr_type, tail->tail_type,
597 if (head->hdr_size != tail->tail_size) {
598 kprintf("HAMMER: FIFO record head/tail size mismatch "
599 "%04x %04x at %016jx\n",
600 head->hdr_size, tail->tail_size,
604 if (tail->tail_signature != HAMMER_TAIL_SIGNATURE) {
605 kprintf("HAMMER: FIFO record bad tail signature "
607 tail->tail_signature,
614 * Non-PAD records must have a CRC and must be sized at
615 * least large enough to fit the head and tail.
617 if (head->hdr_type != HAMMER_HEAD_TYPE_PAD) {
618 crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
619 crc32(head + 1, head->hdr_size - sizeof(*head));
620 if (head->hdr_crc != crc) {
621 kprintf("HAMMER: FIFO record CRC failed %08x %08x "
627 if (head->hdr_size < sizeof(*head) + sizeof(*tail)) {
628 kprintf("HAMMER: FIFO record too small "
639 bytes = head->hdr_size;
640 tail = (void *)((char *)head + bytes - sizeof(*tail));
641 if (tail->tail_size != head->hdr_size) {
642 kprintf("HAMMER: Bad tail size %04x vs %04x at %016jx\n",
643 tail->tail_size, head->hdr_size,
647 if (tail->tail_type != head->hdr_type) {
648 kprintf("HAMMER: Bad tail type %04x vs %04x at %016jx\n",
649 tail->tail_type, head->hdr_type,
658 * Check that the FIFO record is in-bounds given the head and the
661 * Also checks that the head and tail structures agree with each other,
662 * but does not check beyond the signature, type, and size.
665 hammer_check_head_signature(hammer_fifo_head_t head, hammer_off_t beg_off)
667 hammer_fifo_tail_t tail;
668 hammer_off_t end_off;
671 * head overlaps buffer boundary. This could be a PAD so only
672 * check the minimum PAD size here.
674 if (((beg_off + sizeof(*tail) - 1) ^ (beg_off)) & ~HAMMER_BUFMASK64)
678 * Calculate the ending offset and make sure the record does
679 * not cross a buffer boundary.
681 end_off = beg_off + head->hdr_size;
682 if ((beg_off ^ (end_off - 1)) & ~HAMMER_BUFMASK64)
684 tail = (void *)((char *)head + head->hdr_size - sizeof(*tail));
685 return (_hammer_check_signature(head, tail, beg_off));
689 * Check that the FIFO record is in-bounds given the tail and the
690 * hammer offset. The offset is pointing at the ending boundary of the
693 * Also checks that the head and tail structures agree with each other,
694 * but does not check beyond the signature, type, and size.
697 hammer_check_tail_signature(hammer_fifo_tail_t tail, hammer_off_t end_off)
699 hammer_fifo_head_t head;
700 hammer_off_t beg_off;
703 * tail overlaps buffer boundary
705 if (((end_off - sizeof(*tail)) ^ (end_off - 1)) & ~HAMMER_BUFMASK64)
709 * Calculate the begining offset and make sure the record does
710 * not cross a buffer boundary.
712 beg_off = end_off - tail->tail_size;
713 if ((beg_off ^ (end_off - 1)) & ~HAMMER_BUFMASK64)
715 head = (void *)((char *)tail + sizeof(*tail) - tail->tail_size);
716 return (_hammer_check_signature(head, tail, beg_off));
720 hammer_recover_undo(hammer_mount_t hmp, hammer_volume_t root_volume,
721 hammer_fifo_undo_t undo)
723 hammer_volume_t volume;
724 hammer_buffer_t buffer;
725 hammer_off_t buf_offset;
733 * Only process UNDO records. Flag if we find other records to
734 * optimize stage2 recovery.
736 if (undo->head.hdr_type != HAMMER_HEAD_TYPE_UNDO) {
737 if (undo->head.hdr_type == HAMMER_HEAD_TYPE_REDO)
738 hmp->hflags |= HMNT_HASREDO;
743 * Validate the UNDO record.
745 bytes = undo->head.hdr_size - sizeof(*undo) -
746 sizeof(struct hammer_fifo_tail);
747 if (bytes < 0 || undo->undo_data_bytes < 0 ||
748 undo->undo_data_bytes > bytes) {
749 kprintf("HAMMER: Corrupt UNDO record, undo_data_bytes %d/%d\n",
750 undo->undo_data_bytes, bytes);
754 bytes = undo->undo_data_bytes;
757 * The undo offset may only be a zone-1 or zone-2 offset.
759 * Currently we only support a zone-1 offset representing the
762 zone = HAMMER_ZONE_DECODE(undo->undo_offset);
763 offset = undo->undo_offset & HAMMER_BUFMASK;
765 if (offset + bytes > HAMMER_BUFSIZE) {
766 kprintf("HAMMER: Corrupt UNDO record, bad offset\n");
771 case HAMMER_ZONE_RAW_VOLUME_INDEX:
772 vol_no = HAMMER_VOL_DECODE(undo->undo_offset);
773 volume = hammer_get_volume(hmp, vol_no, &error);
774 if (volume == NULL) {
775 kprintf("HAMMER: UNDO record, "
776 "cannot access volume %d\n", vol_no);
779 hammer_modify_volume(NULL, volume, NULL, 0);
780 hammer_recover_copy_undo(undo->undo_offset,
782 (char *)volume->ondisk + offset,
784 hammer_modify_volume_done(volume);
787 * Multiple modifications may be made to the same buffer.
788 * Also, the volume header cannot be written out until
789 * everything else has been flushed. This also
790 * covers the read-only case by preventing the kernel from
791 * flushing the buffer.
793 if (volume->io.recovered == 0)
794 volume->io.recovered = 1;
796 hammer_rel_volume(volume, 0);
798 case HAMMER_ZONE_RAW_BUFFER_INDEX:
799 buf_offset = undo->undo_offset & ~HAMMER_BUFMASK64;
800 buffer = hammer_get_buffer(hmp, buf_offset, HAMMER_BUFSIZE,
802 if (buffer == NULL) {
803 kprintf("HAMMER: UNDO record, "
804 "cannot access buffer %016jx\n",
805 (intmax_t)undo->undo_offset);
808 hammer_modify_buffer(NULL, buffer, NULL, 0);
809 hammer_recover_copy_undo(undo->undo_offset,
811 (char *)buffer->ondisk + offset,
813 hammer_modify_buffer_done(buffer);
816 * Multiple modifications may be made to the same buffer,
817 * improve performance by delaying the flush. This also
818 * covers the read-only case by preventing the kernel from
819 * flushing the buffer.
821 if (buffer->io.recovered == 0)
822 buffer->io.recovered = 1;
824 hammer_rel_buffer(buffer, 0);
827 kprintf("HAMMER: Corrupt UNDO record\n");
834 hammer_recover_copy_undo(hammer_off_t undo_offset,
835 char *src, char *dst, int bytes)
837 if (hammer_debug_general & 0x0080) {
838 kprintf("UNDO %016jx: %d\n",
839 (intmax_t)undo_offset, bytes);
842 kprintf("UNDO %016jx:", (intmax_t)undo_offset);
843 hammer_recover_debug_dump(22, dst, bytes);
844 kprintf("%22s", "to:");
845 hammer_recover_debug_dump(22, src, bytes);
847 bcopy(src, dst, bytes);
853 hammer_recover_debug_dump(int w, char *buf, int bytes)
857 for (i = 0; i < bytes; ++i) {
858 if (i && (i & 15) == 0)
859 kprintf("\n%*.*s", w, w, "");
860 kprintf(" %02x", (unsigned char)buf[i]);
868 * Flush recovered buffers from recovery operations. The call to this
869 * routine may be delayed if a read-only mount was made and then later
870 * upgraded to read-write. This routine is also called when unmounting
871 * a read-only mount to clean out recovered (dirty) buffers which we
872 * couldn't flush (because the mount is read-only).
874 * The volume header is always written last. The UNDO FIFO will be forced
875 * to zero-length by setting next_offset to first_offset. This leaves the
876 * (now stale) UNDO information used to recover the disk available for
879 * final is typically 0 or 1. The volume header is only written if final
880 * is 1. If final is -1 the recovered buffers are discarded instead of
881 * written and root_volume can also be passed as NULL in that case.
883 static int hammer_recover_flush_volume_callback(hammer_volume_t, void *);
884 static int hammer_recover_flush_buffer_callback(hammer_buffer_t, void *);
887 hammer_recover_flush_buffers(hammer_mount_t hmp, hammer_volume_t root_volume,
891 * Flush the buffers out asynchronously, wait for all the I/O to
892 * complete, then do it again to destroy the buffer cache buffer
893 * so it doesn't alias something later on.
895 RB_SCAN(hammer_buf_rb_tree, &hmp->rb_bufs_root, NULL,
896 hammer_recover_flush_buffer_callback, &final);
897 hammer_io_wait_all(hmp, "hmrrcw");
898 RB_SCAN(hammer_buf_rb_tree, &hmp->rb_bufs_root, NULL,
899 hammer_recover_flush_buffer_callback, &final);
902 * Flush all volume headers except the root volume. If final < 0
903 * we discard all volume headers including the root volume.
906 RB_SCAN(hammer_vol_rb_tree, &hmp->rb_vols_root, NULL,
907 hammer_recover_flush_volume_callback, root_volume);
909 RB_SCAN(hammer_vol_rb_tree, &hmp->rb_vols_root, NULL,
910 hammer_recover_flush_volume_callback, NULL);
914 * Finalize the root volume header.
916 if (root_volume && root_volume->io.recovered && final > 0) {
918 while (hmp->io_running_space > 0)
919 tsleep(&hmp->io_running_space, 0, "hmrflx", 0);
921 root_volume->io.recovered = 0;
922 hammer_io_flush(&root_volume->io, 0);
923 hammer_rel_volume(root_volume, 0);
928 * Callback to flush volume headers. If discarding data will be NULL and
929 * all volume headers (including the root volume) will be discarded.
930 * Otherwise data is the root_volume and we flush all volume headers
931 * EXCEPT the root_volume.
933 * Clear any I/O error or modified condition when discarding buffers to
934 * clean up the reference count, otherwise the buffer may have extra refs
939 hammer_recover_flush_volume_callback(hammer_volume_t volume, void *data)
941 hammer_volume_t root_volume = data;
943 if (volume->io.recovered && volume != root_volume) {
944 volume->io.recovered = 0;
945 if (root_volume != NULL) {
946 hammer_io_flush(&volume->io, 0);
948 hammer_io_clear_error(&volume->io);
949 hammer_io_clear_modify(&volume->io, 1);
951 hammer_rel_volume(volume, 0);
957 * Flush or discard recovered I/O buffers.
959 * Clear any I/O error or modified condition when discarding buffers to
960 * clean up the reference count, otherwise the buffer may have extra refs
965 hammer_recover_flush_buffer_callback(hammer_buffer_t buffer, void *data)
967 int final = *(int *)data;
969 if (buffer->io.recovered) {
970 buffer->io.recovered = 0;
971 buffer->io.reclaim = 1;
973 hammer_io_clear_error(&buffer->io);
974 hammer_io_clear_modify(&buffer->io, 1);
976 hammer_io_flush(&buffer->io, 0);
978 hammer_rel_buffer(buffer, 0);
980 if (buffer->io.lock.refs == 0)
981 ++hammer_count_refedbufs;
982 hammer_ref(&buffer->io.lock);
984 hammer_io_clear_error(&buffer->io);
985 hammer_io_clear_modify(&buffer->io, 1);
987 KKASSERT(buffer->io.lock.refs == 1);
988 buffer->io.reclaim = 1;
989 hammer_rel_buffer(buffer, 1);