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
38 * The UNDO algorithm is trivial. The nominal UNDO range in the
39 * FIFO is determined by taking the first/next offset stored in
40 * the volume header. The next offset may not be correct since
41 * UNDO flushes are not required to flush the volume header, so
42 * the code also scans forward until it finds a discontinuous
45 * The UNDOs are then scanned and executed in reverse order. These
46 * UNDOs are effectively just data restorations based on HAMMER offsets.
50 * REDO records are laid down in the UNDO/REDO FIFO for nominal
51 * writes, truncations, and file extension ops. On a per-inode
52 * basis two types of REDO records are generated, REDO_WRITE
55 * Essentially the recovery block will contain UNDO records backing
56 * out partial operations and REDO records to regenerate those partial
57 * operations guaranteed by the filesystem during recovery.
59 * REDO generation is optional, and can also be started and then
60 * later stopped due to excessive write()s inbetween fsyncs, or not
61 * started at all. Because of this the recovery code must determine
62 * when REDOs are valid and when they are not. Additional records are
63 * generated to help figure it out.
65 * The REDO_TERM_WRITE and REDO_TERM_TRUNC records are generated
66 * during a flush cycle indicating which records the flush cycle
67 * has synched meta-data for, and HAMMER_REDO_SYNC is generated in
68 * each flush cycle to indicate how far back in the UNDO/REDO FIFO
69 * the recovery code must go to find the earliest applicable REDO
70 * record. Applicable REDO records can be far outside the nominal
71 * UNDO recovery range, for example if a write() lays down a REDO but
72 * the related file is not flushed for several cycles.
74 * The SYNC reference is to a point prior to the nominal UNDO FIFO
75 * range, creating an extended REDO range which must be scanned.
77 * Any REDO_WRITE/REDO_TRUNC encountered within the extended range
78 * which have no matching REDO_TERM_WRITE/REDO_TERM_TRUNC records
79 * prior to the start of the nominal UNDO range are applicable.
80 * That is, any REDO_TERM_* records in the extended range but not in
81 * the nominal undo range will mask any redo operations for prior REDO
82 * records. This is necessary because once the TERM is laid down
83 * followup operations may make additional changes to the related
84 * records but not necessarily record them as REDOs (because REDOs are
87 * REDO_TERM_WRITE/REDO_TERM_TRUNC records in the nominal UNDO range
88 * must be ignored since they represent meta-data flushes which are
89 * undone by the UNDOs in that nominal UNDO range by the recovery
90 * code. Only REDO_TERM_* records in the extended range but not
91 * in the nominal undo range are applicable.
93 * The REDO_SYNC record itself always exists in the nominal UNDO range
94 * (this is how the extended range is determined). For recovery
95 * purposes the most recent REDO_SYNC record is always used if several
98 * CRASHES DURING UNDO/REDO
100 * A crash during the UNDO phase requires no additional effort. The
101 * UNDOs will simply be re-run again. The state of the UNDO/REDO fifo
102 * remains unchanged and has no re-crash issues.
104 * A crash during the REDO phase is more complex because the REDOs
105 * run normal filesystem ops and generate additional UNDO/REDO records.
106 * REDO is disabled during REDO recovery and any SYNC records generated
107 * by flushes during REDO recovery must continue to reference the
108 * original extended range.
110 * If multiple crashes occur and the UNDO/REDO FIFO wraps, REDO recovery
111 * may become impossible. This is detected when the start of the
112 * extended range fails to have monotonically increasing sequence
113 * numbers leading into the nominal undo range.
119 * Each rterm entry has a list of fifo offsets indicating termination
120 * points. These are stripped as the scan progresses.
122 typedef struct hammer_rterm_entry {
123 struct hammer_rterm_entry *next;
124 hammer_off_t fifo_offset;
125 } *hammer_rterm_entry_t;
128 * rterm entries sorted in RB tree are indexed by objid, flags, and offset.
129 * TRUNC entries ignore the offset.
131 typedef struct hammer_rterm {
132 RB_ENTRY(hammer_rterm) rb_node;
134 u_int32_t redo_localization;
135 u_int32_t redo_flags;
136 hammer_off_t redo_offset;
137 hammer_rterm_entry_t term_list;
140 static int hammer_rterm_rb_cmp(hammer_rterm_t rt1, hammer_rterm_t rt2);
141 struct hammer_rterm_rb_tree;
142 RB_HEAD(hammer_rterm_rb_tree, hammer_rterm);
143 RB_PROTOTYPE(hammer_rterm_rb_tree, hammer_rterm, rb_node, hammer_rterm_rb_cmp);
145 static int hammer_check_tail_signature(hammer_fifo_tail_t tail,
146 hammer_off_t end_off);
147 static int hammer_check_head_signature(hammer_fifo_head_t head,
148 hammer_off_t beg_off);
149 static void hammer_recover_copy_undo(hammer_off_t undo_offset,
150 char *src, char *dst, int bytes);
151 static hammer_fifo_any_t hammer_recover_scan_fwd(hammer_mount_t hmp,
152 hammer_volume_t root_volume,
153 hammer_off_t *scan_offsetp,
154 int *errorp, struct hammer_buffer **bufferp);
155 static hammer_fifo_any_t hammer_recover_scan_rev(hammer_mount_t hmp,
156 hammer_volume_t root_volume,
157 hammer_off_t *scan_offsetp,
158 int *errorp, struct hammer_buffer **bufferp);
160 static void hammer_recover_debug_dump(int w, char *buf, int bytes);
162 static int hammer_recover_undo(hammer_mount_t hmp, hammer_volume_t root_volume,
163 hammer_fifo_undo_t undo);
164 static int hammer_recover_redo_rec(hammer_mount_t hmp,
165 struct hammer_rterm_rb_tree *root,
166 hammer_off_t redo_fifo_offset, hammer_fifo_redo_t redo);
167 static int hammer_recover_redo_run(hammer_mount_t hmp,
168 struct hammer_rterm_rb_tree *root,
169 hammer_off_t redo_fifo_offset, hammer_fifo_redo_t redo);
170 static void hammer_recover_redo_exec(hammer_mount_t hmp,
171 hammer_fifo_redo_t redo);
173 RB_GENERATE(hammer_rterm_rb_tree, hammer_rterm, rb_node, hammer_rterm_rb_cmp);
176 * Recover filesystem meta-data on mount. This procedure figures out the
177 * UNDO FIFO range and runs the UNDOs backwards. The FIFO pointers are not
178 * resynchronized by this procedure.
180 * This procedure is run near the beginning of the mount sequence, before
181 * any B-Tree or high-level accesses are enabled, and is responsible for
182 * restoring the meta-data to a consistent state. High level HAMMER data
183 * structures (such as the B-Tree) cannot be accessed here.
185 * NOTE: No information from the root volume has been cached in the
186 * hammer_mount structure yet, so we need to access the root volume's
192 hammer_recover_stage1(hammer_mount_t hmp, hammer_volume_t root_volume)
194 hammer_blockmap_t rootmap;
195 hammer_buffer_t buffer;
196 hammer_off_t scan_offset;
197 hammer_off_t scan_offset_save;
199 hammer_fifo_any_t head;
200 hammer_off_t first_offset;
201 hammer_off_t last_offset;
204 int degenerate_case = 0;
207 * Examine the UNDO FIFO indices in the volume header.
209 rootmap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
210 first_offset = rootmap->first_offset;
211 last_offset = rootmap->next_offset;
215 hmp->recover_stage2_offset = 0;
217 if (first_offset > rootmap->alloc_offset ||
218 last_offset > rootmap->alloc_offset) {
219 kprintf("HAMMER(%s) Illegal UNDO FIFO index range "
220 "%016jx, %016jx limit %016jx\n",
221 root_volume->ondisk->vol_name,
222 (intmax_t)first_offset,
223 (intmax_t)last_offset,
224 (intmax_t)rootmap->alloc_offset);
230 * In HAMMER version 4+ filesystems the volume header does NOT
231 * contain definitive UNDO FIFO state. In particular, the
232 * rootmap->next_offset may not be indexed completely to the
233 * end of the active UNDO FIFO.
235 if (hmp->version >= HAMMER_VOL_VERSION_FOUR) {
237 * To find the definitive range we must first scan backwards
238 * from first_offset to locate the first real record and
239 * extract the sequence number from it. This record is not
240 * part of the active undo space.
242 scan_offset = first_offset;
246 head = hammer_recover_scan_rev(hmp, root_volume,
251 if (head->head.hdr_type != HAMMER_HEAD_TYPE_PAD) {
252 seqno = head->head.hdr_seq;
257 kprintf("HAMMER(%s) recovery failure "
258 "during seqno backscan\n",
259 root_volume->ondisk->vol_name);
264 * Scan forwards from first_offset and (seqno+1) looking
265 * for a sequence space discontinuity. This denotes the
266 * end of the active FIFO area.
268 * NOTE: For the case where the FIFO is empty the very first
269 * record we find will be discontinuous.
271 * NOTE: Do not include trailing PADs in the scan range,
272 * and remember the returned scan_offset after a
273 * fwd iteration points to the end of the returned
276 kprintf("HAMMER(%s) recovery check seqno=%08x\n",
277 root_volume->ondisk->vol_name,
280 scan_offset = first_offset;
281 scan_offset_save = scan_offset;
283 hmp->recover_stage2_seqno = seqno;
286 head = hammer_recover_scan_fwd(hmp, root_volume,
291 if (head->head.hdr_type != HAMMER_HEAD_TYPE_PAD) {
292 if (seqno != head->head.hdr_seq) {
293 scan_offset = scan_offset_save;
296 scan_offset_save = scan_offset;
302 * If the forward scan is grossly ahead of last_offset
303 * then something is wrong. last_offset is supposed
306 if (last_offset >= scan_offset) {
307 bytes = last_offset - scan_offset;
309 bytes = rootmap->alloc_offset - scan_offset +
310 (last_offset & HAMMER_OFF_LONG_MASK);
313 (rootmap->alloc_offset & HAMMER_OFF_LONG_MASK) *
315 kprintf("HAMMER(%s) recovery forward scan is "
316 "grossly beyond the last_offset in "
317 "the volume header, this can't be "
319 root_volume->ondisk->vol_name);
327 * Store the seqno. This will be the next seqno we lay down
328 * when generating new UNDOs.
330 hmp->undo_seqno = seqno;
332 kprintf("HAMMER(%s) recovery failure "
333 "during seqno fwdscan\n",
334 root_volume->ondisk->vol_name);
337 last_offset = scan_offset;
338 kprintf("HAMMER(%s) recovery range %016jx-%016jx\n"
339 "HAMMER(%s) recovery nexto %016jx endseqno=%08x\n",
340 root_volume->ondisk->vol_name,
341 (intmax_t)first_offset,
342 (intmax_t)last_offset,
343 root_volume->ondisk->vol_name,
344 (intmax_t)rootmap->next_offset,
349 * Calculate the size of the active portion of the FIFO. If the
350 * FIFO is empty the filesystem is clean and no further action is
353 if (last_offset >= first_offset) {
354 bytes = last_offset - first_offset;
356 bytes = rootmap->alloc_offset - first_offset +
357 (last_offset & HAMMER_OFF_LONG_MASK);
365 kprintf("HAMMER(%s) recovery undo %016jx-%016jx (%jd bytes)%s\n",
366 root_volume->ondisk->vol_name,
367 (intmax_t)first_offset,
368 (intmax_t)last_offset,
370 (hmp->ronly ? " (RO)" : "(RW)"));
371 if (bytes > (rootmap->alloc_offset & HAMMER_OFF_LONG_MASK)) {
372 kprintf("Undo size is absurd, unable to mount\n");
378 * Scan the UNDOs backwards.
380 scan_offset = last_offset;
382 while ((int64_t)bytes > 0) {
383 KKASSERT(scan_offset != first_offset);
384 head = hammer_recover_scan_rev(hmp, root_volume,
385 &scan_offset, &error, &buffer);
392 error = hammer_recover_undo(hmp, root_volume, &head->undo);
394 kprintf("HAMMER(%s) UNDO record at %016jx failed\n",
395 root_volume->ondisk->vol_name,
396 (intmax_t)scan_offset - head->head.hdr_size);
401 * The first REDO_SYNC record encountered (scanning backwards)
402 * enables REDO processing.
404 if (head->head.hdr_type == HAMMER_HEAD_TYPE_REDO &&
405 head->redo.redo_flags == HAMMER_REDO_SYNC) {
406 if (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_REQ) {
407 kprintf("HAMMER(%s) Ignoring extra REDO_SYNC "
408 "records in UNDO/REDO FIFO.\n",
409 root_volume->ondisk->vol_name
412 hmp->flags |= HAMMER_MOUNT_REDO_RECOVERY_REQ;
413 hmp->recover_stage2_offset =
414 head->redo.redo_offset;
415 kprintf("HAMMER(%s) Found REDO_SYNC %016jx\n",
416 root_volume->ondisk->vol_name,
417 (intmax_t)head->redo.redo_offset);
421 bytes -= head->head.hdr_size;
424 * If too many dirty buffers have built up we have to flush'm
425 * out. As long as we do not flush out the volume header
426 * a crash here should not cause any problems.
428 * buffer must be released so the flush can assert that
429 * all buffers are idle.
431 if (hammer_flusher_meta_limit(hmp)) {
433 hammer_rel_buffer(buffer, 0);
436 if (hmp->ronly == 0) {
437 hammer_recover_flush_buffers(hmp, root_volume,
439 kprintf("HAMMER(%s) Continuing recovery\n",
440 root_volume->ondisk->vol_name);
442 kprintf("HAMMER(%s) Recovery failure: Insufficient buffer cache to hold dirty buffers on read-only mount!\n",
443 root_volume->ondisk->vol_name);
449 KKASSERT(error || bytes == 0);
452 hammer_rel_buffer(buffer, 0);
457 * After completely flushing all the recovered buffers the volume
458 * header will also be flushed.
460 if (root_volume->io.recovered == 0) {
461 hammer_ref_volume(root_volume);
462 root_volume->io.recovered = 1;
466 * Finish up flushing (or discarding) recovered buffers. FIFO
467 * indices in the volume header are updated to the actual undo
468 * range but will not be collapsed until stage 2.
471 hammer_modify_volume(NULL, root_volume, NULL, 0);
472 rootmap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
473 rootmap->first_offset = first_offset;
474 rootmap->next_offset = last_offset;
475 hammer_modify_volume_done(root_volume);
477 hammer_recover_flush_buffers(hmp, root_volume, 1);
479 hammer_recover_flush_buffers(hmp, root_volume, -1);
481 if (degenerate_case == 0) {
482 kprintf("HAMMER(%s) recovery complete\n",
483 root_volume->ondisk->vol_name);
485 kprintf("HAMMER(%s) mounted clean, no recovery needed\n",
486 root_volume->ondisk->vol_name);
492 * Execute redo operations
494 * This procedure is run at the end of the mount sequence, after the hammer
495 * mount structure has been completely initialized but before the filesystem
496 * goes live. It can access standard cursors, the B-Tree, flush the
497 * filesystem, and so forth.
499 * This code may only be called for read-write mounts or when a mount
500 * switches from read-only to read-write. vnodes may or may not be present.
502 * The stage1 code will have already calculated the correct FIFO range
503 * for the nominal UNDO FIFO and stored it in the rootmap. The extended
504 * range for REDO is stored in hmp->recover_stage2_offset.
507 hammer_recover_stage2(hammer_mount_t hmp, hammer_volume_t root_volume)
509 hammer_blockmap_t rootmap;
510 hammer_buffer_t buffer;
511 hammer_off_t scan_offset;
512 hammer_off_t oscan_offset;
514 hammer_off_t ext_bytes;
515 hammer_fifo_any_t head;
516 hammer_off_t first_offset;
517 hammer_off_t last_offset;
518 hammer_off_t ext_offset;
519 struct hammer_rterm_rb_tree rterm_root;
526 * Stage 2 can only be run on a RW mount, or when the mount is
527 * switched from RO to RW.
529 KKASSERT(hmp->ronly == 0);
530 RB_INIT(&rterm_root);
533 * Examine the UNDO FIFO. If it is empty the filesystem is clean
534 * and no action need be taken.
536 rootmap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
537 first_offset = rootmap->first_offset;
538 last_offset = rootmap->next_offset;
539 if (first_offset == last_offset) {
540 KKASSERT((hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_REQ) == 0);
545 * Stage2 must only be run once, and will not be run at all
546 * if Stage1 did not find a REDO_SYNC record.
551 if ((hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_REQ) == 0)
553 hmp->flags &= ~HAMMER_MOUNT_REDO_RECOVERY_REQ;
554 hmp->flags |= HAMMER_MOUNT_REDO_RECOVERY_RUN;
555 ext_offset = hmp->recover_stage2_offset;
556 if (ext_offset == 0) {
557 kprintf("HAMMER(%s) REDO stage specified but no REDO_SYNC "
558 "offset, ignoring\n",
559 root_volume->ondisk->vol_name);
564 * Calculate nominal UNDO range (this is not yet the extended
567 if (last_offset >= first_offset) {
568 bytes = last_offset - first_offset;
570 bytes = rootmap->alloc_offset - first_offset +
571 (last_offset & HAMMER_OFF_LONG_MASK);
573 kprintf("HAMMER(%s) recovery redo %016jx-%016jx (%jd bytes)%s\n",
574 root_volume->ondisk->vol_name,
575 (intmax_t)first_offset,
576 (intmax_t)last_offset,
578 (hmp->ronly ? " (RO)" : "(RW)"));
580 if (bytes > (rootmap->alloc_offset & HAMMER_OFF_LONG_MASK)) {
581 kprintf("Undo size is absurd, unable to mount\n");
587 * Scan the REDOs backwards collecting REDO_TERM_* information.
588 * This information is only collected for the extended range,
589 * non-inclusive of any TERMs in the nominal UNDO range.
591 * If the stage2 extended range is inside the nominal undo range
592 * we have nothing to scan.
594 * This must fit in memory!
596 if (first_offset < last_offset) {
598 * [ first_offset........last_offset ]
600 if (ext_offset < first_offset) {
602 ext_bytes = first_offset - ext_offset;
603 } else if (ext_offset > last_offset) {
605 ext_bytes = (rootmap->alloc_offset - ext_offset) +
606 (first_offset & HAMMER_OFF_LONG_MASK);
608 ext_bytes = -(ext_offset - first_offset);
613 * [......last_offset first_offset.....]
615 if (ext_offset < last_offset) {
616 ext_bytes = -((rootmap->alloc_offset - first_offset) +
617 (ext_offset & HAMMER_OFF_LONG_MASK));
619 } else if (ext_offset > first_offset) {
620 ext_bytes = -(ext_offset - first_offset);
623 ext_bytes = first_offset - ext_offset;
629 scan_offset = first_offset;
630 kprintf("HAMMER(%s) Find extended redo %016jx, %jd extbytes\n",
631 root_volume->ondisk->vol_name,
632 (intmax_t)ext_offset,
633 (intmax_t)ext_bytes);
634 seqno = hmp->recover_stage2_seqno - 1;
636 head = hammer_recover_scan_rev(hmp, root_volume,
641 if (head->head.hdr_type != HAMMER_HEAD_TYPE_PAD) {
642 if (head->head.hdr_seq != seqno) {
646 error = hammer_recover_redo_rec(
648 scan_offset, &head->redo);
651 if (scan_offset == ext_offset)
655 kprintf("HAMMER(%s) Find extended redo failed %d, "
656 "unable to run REDO\n",
657 root_volume->ondisk->vol_name,
662 kprintf("HAMMER(%s) Embeded extended redo %016jx, "
664 root_volume->ondisk->vol_name,
665 (intmax_t)ext_offset,
666 (intmax_t)ext_bytes);
670 * Scan the REDO forwards through the entire extended range.
671 * Anything with a previously recorded matching TERM is discarded.
673 scan_offset = ext_offset;
677 * NOTE: when doing a forward scan the returned scan_offset is
678 * for the record following the returned record, so we
679 * have to play a bit.
681 while ((int64_t)bytes > 0) {
682 KKASSERT(scan_offset != last_offset);
684 oscan_offset = scan_offset;
685 head = hammer_recover_scan_fwd(hmp, root_volume,
686 &scan_offset, &error, &buffer);
690 error = hammer_recover_redo_run(hmp, &rterm_root,
691 oscan_offset, &head->redo);
693 kprintf("HAMMER(%s) UNDO record at %016jx failed\n",
694 root_volume->ondisk->vol_name,
695 (intmax_t)scan_offset - head->head.hdr_size);
698 bytes -= head->head.hdr_size;
700 KKASSERT(error || bytes == 0);
702 hammer_rel_buffer(buffer, 0);
711 hammer_rterm_t rterm;
712 hammer_rterm_entry_t rte;
714 while ((rterm = RB_ROOT(&rterm_root)) != NULL) {
715 RB_REMOVE(hammer_rterm_rb_tree, &rterm_root, rterm);
716 while ((rte = rterm->term_list) != NULL) {
717 rterm->term_list = rte->next;
718 kfree(rte, hmp->m_misc);
720 kfree(rterm, hmp->m_misc);
725 * Finish up flushing (or discarding) recovered buffers by executing
726 * a normal flush cycle. Setting HMNT_UNDO_DIRTY bypasses degenerate
727 * case tests and forces the flush in order to update the FIFO indices.
729 * If a crash occurs during the flush the entire undo/redo will be
730 * re-run during recovery on the next mount.
733 if (rootmap->first_offset != rootmap->next_offset)
734 hmp->hflags |= HMNT_UNDO_DIRTY;
735 hammer_flusher_sync(hmp);
738 hmp->flags &= ~HAMMER_MOUNT_REDO_RECOVERY_RUN;
740 kprintf("HAMMER(%s) End redo recovery\n",
741 root_volume->ondisk->vol_name);
747 * Scan backwards from *scan_offsetp, return the FIFO record prior to the
748 * record at *scan_offsetp or NULL if an error occured.
750 * On return *scan_offsetp will be the offset of the returned record.
753 hammer_recover_scan_rev(hammer_mount_t hmp, hammer_volume_t root_volume,
754 hammer_off_t *scan_offsetp,
755 int *errorp, struct hammer_buffer **bufferp)
757 hammer_off_t scan_offset;
758 hammer_blockmap_t rootmap;
759 hammer_fifo_any_t head;
760 hammer_fifo_tail_t tail;
762 rootmap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
763 scan_offset = *scan_offsetp;
765 if (hammer_debug_general & 0x0080)
766 kprintf("rev scan_offset %016jx\n", (intmax_t)scan_offset);
767 if (scan_offset == HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0))
768 scan_offset = rootmap->alloc_offset;
769 if (scan_offset - sizeof(*tail) <
770 HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0)) {
771 kprintf("HAMMER(%s) UNDO record at %016jx FIFO underflow\n",
772 root_volume->ondisk->vol_name,
773 (intmax_t)scan_offset);
777 tail = hammer_bread(hmp, scan_offset - sizeof(*tail),
780 kprintf("HAMMER(%s) Unable to read UNDO TAIL "
782 root_volume->ondisk->vol_name,
783 (intmax_t)scan_offset - sizeof(*tail));
787 if (hammer_check_tail_signature(tail, scan_offset) != 0) {
788 kprintf("HAMMER(%s) Illegal UNDO TAIL signature "
790 root_volume->ondisk->vol_name,
791 (intmax_t)scan_offset - sizeof(*tail));
795 head = (void *)((char *)tail + sizeof(*tail) - tail->tail_size);
796 *scan_offsetp = scan_offset - head->head.hdr_size;
802 * Scan forwards from *scan_offsetp, return the FIFO record or NULL if
805 * On return *scan_offsetp will be the offset of the record following
806 * the returned record.
809 hammer_recover_scan_fwd(hammer_mount_t hmp, hammer_volume_t root_volume,
810 hammer_off_t *scan_offsetp,
811 int *errorp, struct hammer_buffer **bufferp)
813 hammer_off_t scan_offset;
814 hammer_blockmap_t rootmap;
815 hammer_fifo_any_t head;
817 rootmap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
818 scan_offset = *scan_offsetp;
820 if (hammer_debug_general & 0x0080)
821 kprintf("fwd scan_offset %016jx\n", (intmax_t)scan_offset);
822 if (scan_offset == rootmap->alloc_offset)
823 scan_offset = HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
825 head = hammer_bread(hmp, scan_offset, errorp, bufferp);
827 kprintf("HAMMER(%s) Unable to read UNDO HEAD at %016jx\n",
828 root_volume->ondisk->vol_name,
829 (intmax_t)scan_offset);
833 if (hammer_check_head_signature(&head->head, scan_offset) != 0) {
834 kprintf("HAMMER(%s) Illegal UNDO TAIL signature "
836 root_volume->ondisk->vol_name,
837 (intmax_t)scan_offset);
841 scan_offset += head->head.hdr_size;
842 if (scan_offset == rootmap->alloc_offset)
843 scan_offset = HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
844 *scan_offsetp = scan_offset;
850 * Helper function for hammer_check_{head,tail}_signature(). Check stuff
851 * once the head and tail has been established.
853 * This function validates the entire FIFO record wrapper.
857 _hammer_check_signature(hammer_fifo_head_t head, hammer_fifo_tail_t tail,
858 hammer_off_t beg_off)
860 hammer_off_t end_off;
865 * Check signatures. The tail signature is allowed to be the
866 * head signature only for 8-byte PADs.
868 if (head->hdr_signature != HAMMER_HEAD_SIGNATURE) {
869 kprintf("HAMMER: FIFO record bad head signature "
875 if (head->hdr_size < HAMMER_HEAD_ALIGN ||
876 (head->hdr_size & HAMMER_HEAD_ALIGN_MASK)) {
877 kprintf("HAMMER: FIFO record unaligned or bad size"
883 end_off = beg_off + head->hdr_size;
885 if (head->hdr_type != HAMMER_HEAD_TYPE_PAD ||
886 (size_t)(end_off - beg_off) != sizeof(*tail)) {
887 if (head->hdr_type != tail->tail_type) {
888 kprintf("HAMMER: FIFO record head/tail type mismatch "
889 "%04x %04x at %016jx\n",
890 head->hdr_type, tail->tail_type,
894 if (head->hdr_size != tail->tail_size) {
895 kprintf("HAMMER: FIFO record head/tail size mismatch "
896 "%04x %04x at %016jx\n",
897 head->hdr_size, tail->tail_size,
901 if (tail->tail_signature != HAMMER_TAIL_SIGNATURE) {
902 kprintf("HAMMER: FIFO record bad tail signature "
904 tail->tail_signature,
911 * Non-PAD records must have a CRC and must be sized at
912 * least large enough to fit the head and tail.
914 if (head->hdr_type != HAMMER_HEAD_TYPE_PAD) {
915 crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
916 crc32(head + 1, head->hdr_size - sizeof(*head));
917 if (head->hdr_crc != crc) {
918 kprintf("HAMMER: FIFO record CRC failed %08x %08x "
924 if (head->hdr_size < sizeof(*head) + sizeof(*tail)) {
925 kprintf("HAMMER: FIFO record too small "
936 bytes = head->hdr_size;
937 tail = (void *)((char *)head + bytes - sizeof(*tail));
938 if (tail->tail_size != head->hdr_size) {
939 kprintf("HAMMER: Bad tail size %04x vs %04x at %016jx\n",
940 tail->tail_size, head->hdr_size,
944 if (tail->tail_type != head->hdr_type) {
945 kprintf("HAMMER: Bad tail type %04x vs %04x at %016jx\n",
946 tail->tail_type, head->hdr_type,
955 * Check that the FIFO record is in-bounds given the head and the
958 * Also checks that the head and tail structures agree with each other,
959 * but does not check beyond the signature, type, and size.
962 hammer_check_head_signature(hammer_fifo_head_t head, hammer_off_t beg_off)
964 hammer_fifo_tail_t tail;
965 hammer_off_t end_off;
968 * head overlaps buffer boundary. This could be a PAD so only
969 * check the minimum PAD size here.
971 if (((beg_off + sizeof(*tail) - 1) ^ (beg_off)) & ~HAMMER_BUFMASK64)
975 * Calculate the ending offset and make sure the record does
976 * not cross a buffer boundary.
978 end_off = beg_off + head->hdr_size;
979 if ((beg_off ^ (end_off - 1)) & ~HAMMER_BUFMASK64)
981 tail = (void *)((char *)head + head->hdr_size - sizeof(*tail));
982 return (_hammer_check_signature(head, tail, beg_off));
986 * Check that the FIFO record is in-bounds given the tail and the
987 * hammer offset. The offset is pointing at the ending boundary of the
990 * Also checks that the head and tail structures agree with each other,
991 * but does not check beyond the signature, type, and size.
994 hammer_check_tail_signature(hammer_fifo_tail_t tail, hammer_off_t end_off)
996 hammer_fifo_head_t head;
997 hammer_off_t beg_off;
1000 * tail overlaps buffer boundary
1002 if (((end_off - sizeof(*tail)) ^ (end_off - 1)) & ~HAMMER_BUFMASK64)
1006 * Calculate the begining offset and make sure the record does
1007 * not cross a buffer boundary.
1009 beg_off = end_off - tail->tail_size;
1010 if ((beg_off ^ (end_off - 1)) & ~HAMMER_BUFMASK64)
1012 head = (void *)((char *)tail + sizeof(*tail) - tail->tail_size);
1013 return (_hammer_check_signature(head, tail, beg_off));
1017 hammer_recover_undo(hammer_mount_t hmp, hammer_volume_t root_volume,
1018 hammer_fifo_undo_t undo)
1020 hammer_volume_t volume;
1021 hammer_buffer_t buffer;
1022 hammer_off_t buf_offset;
1030 * Only process UNDO records. Flag if we find other records to
1031 * optimize stage2 recovery.
1033 if (undo->head.hdr_type != HAMMER_HEAD_TYPE_UNDO)
1037 * Validate the UNDO record.
1039 bytes = undo->head.hdr_size - sizeof(*undo) -
1040 sizeof(struct hammer_fifo_tail);
1041 if (bytes < 0 || undo->undo_data_bytes < 0 ||
1042 undo->undo_data_bytes > bytes) {
1043 kprintf("HAMMER: Corrupt UNDO record, undo_data_bytes %d/%d\n",
1044 undo->undo_data_bytes, bytes);
1048 bytes = undo->undo_data_bytes;
1051 * The undo offset may only be a zone-1 or zone-2 offset.
1053 * Currently we only support a zone-1 offset representing the
1056 zone = HAMMER_ZONE_DECODE(undo->undo_offset);
1057 offset = undo->undo_offset & HAMMER_BUFMASK;
1059 if (offset + bytes > HAMMER_BUFSIZE) {
1060 kprintf("HAMMER: Corrupt UNDO record, bad offset\n");
1065 case HAMMER_ZONE_RAW_VOLUME_INDEX:
1066 vol_no = HAMMER_VOL_DECODE(undo->undo_offset);
1067 volume = hammer_get_volume(hmp, vol_no, &error);
1068 if (volume == NULL) {
1069 kprintf("HAMMER: UNDO record, "
1070 "cannot access volume %d\n", vol_no);
1073 hammer_modify_volume(NULL, volume, NULL, 0);
1074 hammer_recover_copy_undo(undo->undo_offset,
1076 (char *)volume->ondisk + offset,
1078 hammer_modify_volume_done(volume);
1081 * Multiple modifications may be made to the same buffer.
1082 * Also, the volume header cannot be written out until
1083 * everything else has been flushed. This also
1084 * covers the read-only case by preventing the kernel from
1085 * flushing the buffer.
1087 if (volume->io.recovered == 0)
1088 volume->io.recovered = 1;
1090 hammer_rel_volume(volume, 0);
1092 case HAMMER_ZONE_RAW_BUFFER_INDEX:
1093 buf_offset = undo->undo_offset & ~HAMMER_BUFMASK64;
1094 buffer = hammer_get_buffer(hmp, buf_offset, HAMMER_BUFSIZE,
1096 if (buffer == NULL) {
1097 kprintf("HAMMER: UNDO record, "
1098 "cannot access buffer %016jx\n",
1099 (intmax_t)undo->undo_offset);
1102 hammer_modify_buffer(NULL, buffer, NULL, 0);
1103 hammer_recover_copy_undo(undo->undo_offset,
1105 (char *)buffer->ondisk + offset,
1107 hammer_modify_buffer_done(buffer);
1110 * Multiple modifications may be made to the same buffer,
1111 * improve performance by delaying the flush. This also
1112 * covers the read-only case by preventing the kernel from
1113 * flushing the buffer.
1115 if (buffer->io.recovered == 0)
1116 buffer->io.recovered = 1;
1118 hammer_rel_buffer(buffer, 0);
1121 kprintf("HAMMER: Corrupt UNDO record\n");
1128 hammer_recover_copy_undo(hammer_off_t undo_offset,
1129 char *src, char *dst, int bytes)
1131 if (hammer_debug_general & 0x0080) {
1132 kprintf("UNDO %016jx: %d\n",
1133 (intmax_t)undo_offset, bytes);
1136 kprintf("UNDO %016jx:", (intmax_t)undo_offset);
1137 hammer_recover_debug_dump(22, dst, bytes);
1138 kprintf("%22s", "to:");
1139 hammer_recover_debug_dump(22, src, bytes);
1141 bcopy(src, dst, bytes);
1145 * Record HAMMER_REDO_TERM_WRITE and HAMMER_REDO_TERM_TRUNC operations
1146 * during the backwards scan of the extended UNDO/REDO FIFO. This scan
1147 * does not include the nominal UNDO range, just the extended range.
1150 hammer_recover_redo_rec(hammer_mount_t hmp, struct hammer_rterm_rb_tree *root,
1151 hammer_off_t scan_offset, hammer_fifo_redo_t redo)
1153 hammer_rterm_t rterm;
1154 hammer_rterm_t nrterm;
1155 hammer_rterm_entry_t rte;
1157 if (redo->head.hdr_type != HAMMER_HEAD_TYPE_REDO)
1159 if (redo->redo_flags != HAMMER_REDO_TERM_WRITE &&
1160 redo->redo_flags != HAMMER_REDO_TERM_TRUNC) {
1164 nrterm = kmalloc(sizeof(*nrterm), hmp->m_misc, M_WAITOK|M_ZERO);
1165 nrterm->redo_objid = redo->redo_objid;
1166 nrterm->redo_localization = redo->redo_localization;
1167 nrterm->redo_flags = redo->redo_flags;
1168 nrterm->redo_offset = redo->redo_offset;
1170 rterm = RB_INSERT(hammer_rterm_rb_tree, root, nrterm);
1172 kfree(nrterm, hmp->m_misc);
1176 kprintf("record record %016jx objid %016jx offset %016jx flags %08x\n",
1177 (intmax_t)scan_offset,
1178 (intmax_t)redo->redo_objid,
1179 (intmax_t)redo->redo_offset,
1180 (int)redo->redo_flags);
1183 * Scan in reverse order, rte prepended, so the rte list will be
1186 rte = kmalloc(sizeof(*rte), hmp->m_misc, M_WAITOK|M_ZERO);
1187 rte->fifo_offset = scan_offset;
1188 rte->next = rterm->term_list;
1189 rterm->term_list = rte;
1195 * Execute HAMMER_REDO_WRITE and HAMMER_REDO_TRUNC operations during
1196 * the forwards scan of the entire extended UNDO/REDO FIFO range.
1198 * Records matching previously recorded TERMs have already been committed
1202 hammer_recover_redo_run(hammer_mount_t hmp, struct hammer_rterm_rb_tree *root,
1203 hammer_off_t scan_offset, hammer_fifo_redo_t redo)
1205 struct hammer_rterm rtval;
1206 hammer_rterm_t rterm;
1207 hammer_rterm_entry_t rte;
1209 if (redo->head.hdr_type != HAMMER_HEAD_TYPE_REDO)
1212 switch(redo->redo_flags) {
1213 case HAMMER_REDO_WRITE:
1214 case HAMMER_REDO_TRUNC:
1216 * We hit a REDO request. The REDO request is only executed
1217 * if there is no matching TERM.
1219 bzero(&rtval, sizeof(rtval));
1220 rtval.redo_objid = redo->redo_objid;
1221 rtval.redo_localization = redo->redo_localization;
1222 rtval.redo_offset = redo->redo_offset;
1223 rtval.redo_flags = (redo->redo_flags == HAMMER_REDO_WRITE) ?
1224 HAMMER_REDO_TERM_WRITE :
1225 HAMMER_REDO_TERM_TRUNC;
1227 rterm = RB_FIND(hammer_rterm_rb_tree, root, &rtval);
1229 kprintf("ignore record %016jx objid %016jx "
1230 "offset %016jx flags %08x\n",
1231 (intmax_t)scan_offset,
1232 (intmax_t)redo->redo_objid,
1233 (intmax_t)redo->redo_offset,
1234 (int)redo->redo_flags);
1238 kprintf("run record %016jx objid %016jx "
1239 "offset %016jx flags %08x\n",
1240 (intmax_t)scan_offset,
1241 (intmax_t)redo->redo_objid,
1242 (intmax_t)redo->redo_offset,
1243 (int)redo->redo_flags);
1246 * Redo stage2 can access a live filesystem, acquire the
1249 hammer_recover_redo_exec(hmp, redo);
1251 case HAMMER_REDO_TERM_WRITE:
1252 case HAMMER_REDO_TERM_TRUNC:
1254 * As we encounter TERMs in the forward scan we remove
1255 * them. Once the forward scan hits the nominal undo range
1256 * there will be no more recorded TERMs.
1258 bzero(&rtval, sizeof(rtval));
1259 rtval.redo_objid = redo->redo_objid;
1260 rtval.redo_localization = redo->redo_localization;
1261 rtval.redo_flags = redo->redo_flags;
1262 rtval.redo_offset = redo->redo_offset;
1264 rterm = RB_FIND(hammer_rterm_rb_tree, root, &rtval);
1266 if ((rte = rterm->term_list) != NULL) {
1267 KKASSERT(rte->fifo_offset == scan_offset);
1268 rterm->term_list = rte->next;
1269 kfree(rte, hmp->m_misc);
1278 hammer_recover_redo_exec(hammer_mount_t hmp, hammer_fifo_redo_t redo)
1280 struct hammer_transaction trans;
1282 struct hammer_inode *ip;
1283 struct vnode *vp = NULL;
1286 hammer_start_transaction(&trans, hmp);
1288 ip = hammer_get_inode(&trans, NULL, redo->redo_objid,
1289 HAMMER_MAX_TID, redo->redo_localization,
1292 kprintf("unable to find objid %016jx lo %08x\n",
1293 (intmax_t)redo->redo_objid, redo->redo_localization);
1296 error = hammer_get_vnode(ip, &vp);
1298 kprintf("unable to acquire vnode for %016jx lo %08x\n",
1299 (intmax_t)redo->redo_objid, redo->redo_localization);
1303 switch(redo->redo_flags) {
1304 case HAMMER_REDO_WRITE:
1305 error = VOP_OPEN(vp, FREAD|FWRITE, proc0.p_ucred, NULL);
1307 kprintf("vn_rdwr open returned %d\n", error);
1311 error = vn_rdwr(UIO_WRITE, vp, (void *)(redo + 1),
1312 redo->redo_data_bytes,
1313 redo->redo_offset, UIO_SYSSPACE,
1314 0, proc0.p_ucred, NULL);
1315 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1317 kprintf("vn_rdwr write returned %d\n", error);
1318 VOP_CLOSE(vp, FREAD|FWRITE);
1320 case HAMMER_REDO_TRUNC:
1321 kprintf("setattr offset %016jx error %d\n",
1322 (intmax_t)redo->redo_offset, error);
1324 va.va_size = redo->redo_offset;
1325 error = VOP_SETATTR(vp, &va, proc0.p_ucred);
1327 kprintf("stattr returned %d\n", error);
1332 hammer_rel_inode(ip, 0);
1334 hammer_done_transaction(&trans);
1338 * RB tree compare function. Note that REDO_TERM_TRUNC ops ignore
1341 * WRITE@0 TERM@0 WRITE@0 .... (no TERM@0) etc.
1344 hammer_rterm_rb_cmp(hammer_rterm_t rt1, hammer_rterm_t rt2)
1346 if (rt1->redo_objid < rt2->redo_objid)
1348 if (rt1->redo_objid > rt2->redo_objid)
1350 if (rt1->redo_localization < rt2->redo_localization)
1352 if (rt1->redo_localization > rt2->redo_localization)
1354 if (rt1->redo_flags < rt2->redo_flags)
1356 if (rt1->redo_flags > rt2->redo_flags)
1358 if (rt1->redo_flags != HAMMER_REDO_TERM_TRUNC) {
1359 if (rt1->redo_offset < rt2->redo_offset)
1361 if (rt1->redo_offset > rt2->redo_offset)
1370 hammer_recover_debug_dump(int w, char *buf, int bytes)
1374 for (i = 0; i < bytes; ++i) {
1375 if (i && (i & 15) == 0)
1376 kprintf("\n%*.*s", w, w, "");
1377 kprintf(" %02x", (unsigned char)buf[i]);
1385 * Flush recovered buffers from recovery operations. The call to this
1386 * routine may be delayed if a read-only mount was made and then later
1387 * upgraded to read-write. This routine is also called when unmounting
1388 * a read-only mount to clean out recovered (dirty) buffers which we
1389 * couldn't flush (because the mount is read-only).
1391 * The volume header is always written last. The UNDO FIFO will be forced
1392 * to zero-length by setting next_offset to first_offset. This leaves the
1393 * (now stale) UNDO information used to recover the disk available for
1394 * forensic analysis.
1396 * final is typically 0 or 1. The volume header is only written if final
1397 * is 1. If final is -1 the recovered buffers are discarded instead of
1398 * written and root_volume can also be passed as NULL in that case.
1400 static int hammer_recover_flush_volume_callback(hammer_volume_t, void *);
1401 static int hammer_recover_flush_buffer_callback(hammer_buffer_t, void *);
1404 hammer_recover_flush_buffers(hammer_mount_t hmp, hammer_volume_t root_volume,
1408 * Flush the buffers out asynchronously, wait for all the I/O to
1409 * complete, then do it again to destroy the buffer cache buffer
1410 * so it doesn't alias something later on.
1412 RB_SCAN(hammer_buf_rb_tree, &hmp->rb_bufs_root, NULL,
1413 hammer_recover_flush_buffer_callback, &final);
1414 hammer_io_wait_all(hmp, "hmrrcw", 1);
1415 RB_SCAN(hammer_buf_rb_tree, &hmp->rb_bufs_root, NULL,
1416 hammer_recover_flush_buffer_callback, &final);
1419 * Flush all volume headers except the root volume. If final < 0
1420 * we discard all volume headers including the root volume.
1423 RB_SCAN(hammer_vol_rb_tree, &hmp->rb_vols_root, NULL,
1424 hammer_recover_flush_volume_callback, root_volume);
1426 RB_SCAN(hammer_vol_rb_tree, &hmp->rb_vols_root, NULL,
1427 hammer_recover_flush_volume_callback, NULL);
1431 * Finalize the root volume header.
1433 if (root_volume && root_volume->io.recovered && final > 0) {
1434 hammer_io_wait_all(hmp, "hmrflx", 1);
1435 root_volume->io.recovered = 0;
1436 hammer_io_flush(&root_volume->io, 0);
1437 hammer_rel_volume(root_volume, 0);
1438 hammer_io_wait_all(hmp, "hmrfly", 1);
1443 * Callback to flush volume headers. If discarding data will be NULL and
1444 * all volume headers (including the root volume) will be discarded.
1445 * Otherwise data is the root_volume and we flush all volume headers
1446 * EXCEPT the root_volume.
1448 * Clear any I/O error or modified condition when discarding buffers to
1449 * clean up the reference count, otherwise the buffer may have extra refs
1454 hammer_recover_flush_volume_callback(hammer_volume_t volume, void *data)
1456 hammer_volume_t root_volume = data;
1458 if (volume->io.recovered && volume != root_volume) {
1459 volume->io.recovered = 0;
1460 if (root_volume != NULL) {
1461 hammer_io_flush(&volume->io, 0);
1463 hammer_io_clear_error(&volume->io);
1464 hammer_io_clear_modify(&volume->io, 1);
1466 hammer_rel_volume(volume, 0);
1472 * Flush or discard recovered I/O buffers.
1474 * Clear any I/O error or modified condition when discarding buffers to
1475 * clean up the reference count, otherwise the buffer may have extra refs
1480 hammer_recover_flush_buffer_callback(hammer_buffer_t buffer, void *data)
1482 int final = *(int *)data;
1485 if (buffer->io.recovered) {
1486 buffer->io.recovered = 0;
1487 buffer->io.reclaim = 1;
1489 hammer_io_clear_error(&buffer->io);
1490 hammer_io_clear_modify(&buffer->io, 1);
1492 hammer_io_flush(&buffer->io, 0);
1494 hammer_rel_buffer(buffer, 0);
1496 flush = hammer_ref_interlock(&buffer->io.lock);
1498 ++hammer_count_refedbufs;
1501 hammer_io_clear_error(&buffer->io);
1502 hammer_io_clear_modify(&buffer->io, 1);
1504 KKASSERT(hammer_oneref(&buffer->io.lock));
1505 buffer->io.reclaim = 1;
1506 hammer_rel_buffer(buffer, flush);