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
36 * HAMMER undo - undo buffer/FIFO management.
42 hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2)
44 if (node1->offset < node2->offset)
46 if (node1->offset > node2->offset)
51 RB_GENERATE2(hammer_und_rb_tree, hammer_undo, rb_node,
52 hammer_und_rb_compare, hammer_off_t, offset);
55 * Convert a zone-3 undo offset into a zone-2 buffer offset.
58 hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t zone3_off, int *errorp)
60 hammer_volume_t root_volume;
61 hammer_blockmap_t undomap __debugvar;
62 hammer_off_t result_offset;
65 KKASSERT(hammer_is_zone_undo(zone3_off));
66 root_volume = hammer_get_root_volume(hmp, errorp);
69 undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
70 KKASSERT(HAMMER_ZONE_DECODE(undomap->alloc_offset) == HAMMER_ZONE_UNDO_INDEX);
71 KKASSERT(zone3_off < undomap->alloc_offset);
74 * undo offsets[i] in zone-2 +
75 * big-block offset of zone-3 address
76 * which results zone-2 address
78 i = (zone3_off & HAMMER_OFF_SHORT_MASK) / HAMMER_BIGBLOCK_SIZE;
79 result_offset = root_volume->ondisk->vol0_undo_array[i] +
80 (zone3_off & HAMMER_BIGBLOCK_MASK64);
82 hammer_rel_volume(root_volume, 0);
83 return(result_offset);
87 * Generate UNDO record(s) for the block of data at the specified zone1
90 * The recovery code will execute UNDOs in reverse order, allowing overlaps.
91 * All the UNDOs are executed together so if we already laid one down we
92 * do not have to lay another one down for the same range.
94 * For HAMMER version 4+ UNDO a 512 byte boundary is enforced and a PAD
95 * will be laid down for any unused space. UNDO FIFO media structures
96 * will implement the hdr_seq field (it used to be reserved01), and
97 * both flush and recovery mechanics will be very different.
99 * WARNING! See also hammer_generate_redo() in hammer_redo.c
102 hammer_generate_undo(hammer_transaction_t trans,
103 hammer_off_t zone_off, void *base, int len)
106 hammer_volume_t root_volume;
107 hammer_blockmap_t undomap;
108 hammer_buffer_t buffer = NULL;
109 hammer_fifo_undo_t undo;
110 hammer_fifo_tail_t tail;
111 hammer_off_t next_offset;
119 * A SYNC record may be required before we can lay down a general
120 * UNDO. This ensures that the nominal recovery span contains
121 * at least one SYNC record telling the recovery code how far
122 * out-of-span it must go to run the REDOs.
124 if ((hmp->flags & HAMMER_MOUNT_REDO_SYNC) == 0 &&
125 hmp->version >= HAMMER_VOL_VERSION_FOUR) {
126 hammer_generate_redo_sync(trans);
130 * Enter the offset into our undo history. If there is an existing
131 * undo we do not have to generate a new one.
133 if (hammer_enter_undo_history(hmp, zone_off, len) == EALREADY)
136 root_volume = trans->rootvol;
137 undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
139 /* no undo recursion */
140 hammer_modify_volume_noundo(NULL, root_volume);
141 hammer_lock_ex(&hmp->undo_lock);
143 /* undo had better not roll over (loose test) */
144 if (hammer_undo_space(trans) < len + HAMMER_BUFSIZE*3)
145 hpanic("insufficient undo FIFO space!");
148 * Loop until the undo for the entire range has been laid down.
152 * Fetch the layout offset in the UNDO FIFO, wrap it as
155 if (undomap->next_offset == undomap->alloc_offset) {
156 undomap->next_offset =
157 HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
159 next_offset = undomap->next_offset;
162 * This is a tail-chasing FIFO, when we hit the start of a new
163 * buffer we don't have to read it in.
165 if ((next_offset & HAMMER_BUFMASK) == 0) {
166 undo = hammer_bnew(hmp, next_offset, &error, &buffer);
167 hammer_format_undo(undo, hmp->undo_seqno ^ 0x40000000);
169 undo = hammer_bread(hmp, next_offset, &error, &buffer);
173 /* no undo recursion */
174 hammer_modify_buffer_noundo(NULL, buffer);
177 * Calculate how big a media structure fits up to the next
178 * alignment point and how large a data payload we can
181 * If n calculates to 0 or negative there is no room for
182 * anything but a PAD.
184 bytes = HAMMER_UNDO_ALIGN -
185 ((int)next_offset & HAMMER_UNDO_MASK);
187 (int)sizeof(struct hammer_fifo_undo) -
188 (int)sizeof(struct hammer_fifo_tail);
191 * If available space is insufficient for any payload
192 * we have to lay down a PAD.
194 * The minimum PAD is 8 bytes and the head and tail will
195 * overlap each other in that case. PADs do not have
196 * sequence numbers or CRCs.
198 * A PAD may not start on a boundary. That is, every
199 * 512-byte block in the UNDO/REDO FIFO must begin with
200 * a record containing a sequence number.
203 KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
204 KKASSERT(((int)next_offset & HAMMER_UNDO_MASK) != 0);
205 tail = (void *)((char *)undo + bytes - sizeof(*tail));
206 if ((void *)undo != (void *)tail) {
207 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
208 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
209 tail->tail_size = bytes;
211 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
212 undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
213 undo->head.hdr_size = bytes;
214 /* NO CRC OR SEQ NO */
215 undomap->next_offset += bytes;
216 hammer_modify_buffer_done(buffer);
217 hammer_stats_undo += bytes;
222 * Calculate the actual payload and recalculate the size
223 * of the media structure as necessary.
227 bytes = ((n + HAMMER_HEAD_ALIGN_MASK) &
228 ~HAMMER_HEAD_ALIGN_MASK) +
229 (int)sizeof(struct hammer_fifo_undo) +
230 (int)sizeof(struct hammer_fifo_tail);
232 if (hammer_debug_general & 0x0080) {
233 hdkprintf("undo %016llx %d %d\n",
234 (long long)next_offset, bytes, n);
237 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
238 undo->head.hdr_type = HAMMER_HEAD_TYPE_UNDO;
239 undo->head.hdr_size = bytes;
240 undo->head.hdr_seq = hmp->undo_seqno++;
241 undo->head.hdr_crc = 0;
242 undo->undo_offset = zone_off;
243 undo->undo_data_bytes = n;
244 bcopy(base, undo + 1, n);
246 tail = (void *)((char *)undo + bytes - sizeof(*tail));
247 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
248 tail->tail_type = HAMMER_HEAD_TYPE_UNDO;
249 tail->tail_size = bytes;
251 KKASSERT(bytes >= sizeof(undo->head));
252 undo->head.hdr_crc = crc32(undo, HAMMER_FIFO_HEAD_CRCOFF) ^
253 crc32(&undo->head + 1, bytes - sizeof(undo->head));
254 undomap->next_offset += bytes;
255 hammer_stats_undo += bytes;
258 * Before we finish off the buffer we have to deal with any
259 * junk between the end of the media structure we just laid
260 * down and the UNDO alignment boundary. We do this by laying
261 * down a dummy PAD. Even though we will probably overwrite
262 * it almost immediately we have to do this so recovery runs
263 * can iterate the UNDO space without having to depend on
264 * the indices in the volume header.
266 * This dummy PAD will be overwritten on the next undo so
267 * we do not adjust undomap->next_offset.
269 bytes = HAMMER_UNDO_ALIGN -
270 ((int)undomap->next_offset & HAMMER_UNDO_MASK);
271 if (bytes != HAMMER_UNDO_ALIGN) {
272 KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
273 undo = (void *)(tail + 1);
274 tail = (void *)((char *)undo + bytes - sizeof(*tail));
275 if ((void *)undo != (void *)tail) {
276 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
277 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
278 tail->tail_size = bytes;
280 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
281 undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
282 undo->head.hdr_size = bytes;
283 /* NO CRC OR SEQ NO */
285 hammer_modify_buffer_done(buffer);
291 base = (char *)base + n;
294 hammer_modify_volume_done(root_volume);
295 hammer_unlock(&hmp->undo_lock);
298 hammer_rel_buffer(buffer, 0);
303 * Preformat a new UNDO block. We could read the old one in but we get
304 * better performance if we just pre-format a new one.
306 * The recovery code always works forwards so the caller just makes sure the
307 * seqno is not contiguous with prior UNDOs or ancient UNDOs now being
310 * The preformatted UNDO headers use the smallest possible sector size
311 * (512) to ensure that any missed media writes are caught.
313 * NOTE: Also used by the REDO code.
316 hammer_format_undo(void *base, uint32_t seqno)
318 hammer_fifo_head_t head;
319 hammer_fifo_tail_t tail;
321 int bytes = HAMMER_UNDO_ALIGN;
323 bzero(base, HAMMER_BUFSIZE);
325 for (i = 0; i < HAMMER_BUFSIZE; i += bytes) {
326 head = (void *)((char *)base + i);
327 tail = (void *)((char *)head + bytes - sizeof(*tail));
329 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
330 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
331 head->hdr_size = bytes;
332 head->hdr_seq = seqno++;
335 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
336 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
337 tail->tail_size = bytes;
339 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
340 crc32(head + 1, bytes - sizeof(*head));
345 * HAMMER version 4+ conversion support.
347 * Convert a HAMMER version < 4 UNDO FIFO area to a 4+ UNDO FIFO area.
348 * The 4+ UNDO FIFO area is backwards compatible. The conversion is
349 * needed to initialize the sequence space and place headers on the
350 * new 512-byte undo boundary.
353 hammer_upgrade_undo_4(hammer_transaction_t trans)
356 hammer_volume_t root_volume;
357 hammer_blockmap_t undomap;
358 hammer_buffer_t buffer = NULL;
359 hammer_fifo_head_t head;
360 hammer_fifo_tail_t tail;
361 hammer_off_t next_offset;
368 root_volume = trans->rootvol;
370 /* no undo recursion */
371 hammer_lock_ex(&hmp->undo_lock);
372 hammer_modify_volume_noundo(NULL, root_volume);
375 * Adjust the in-core undomap and the on-disk undomap.
377 next_offset = HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
378 undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
379 undomap->next_offset = next_offset;
380 undomap->first_offset = next_offset;
382 undomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
383 undomap->next_offset = next_offset;
384 undomap->first_offset = next_offset;
387 * Loop over the entire UNDO space creating DUMMY entries. Sequence
388 * numbers are assigned.
391 bytes = HAMMER_UNDO_ALIGN;
393 while (next_offset != undomap->alloc_offset) {
394 head = hammer_bnew(hmp, next_offset, &error, &buffer);
397 hammer_modify_buffer_noundo(NULL, buffer);
398 tail = (void *)((char *)head + bytes - sizeof(*tail));
400 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
401 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
402 head->hdr_size = bytes;
403 head->hdr_seq = seqno;
406 tail = (void *)((char *)head + bytes - sizeof(*tail));
407 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
408 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
409 tail->tail_size = bytes;
411 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
412 crc32(head + 1, bytes - sizeof(*head));
413 hammer_modify_buffer_done(buffer);
415 hammer_stats_undo += bytes;
416 next_offset += HAMMER_UNDO_ALIGN;
421 * The sequence number will be the next sequence number to lay down.
423 hmp->undo_seqno = seqno;
424 hmkprintf(hmp, "version upgrade seqno start %08x\n", seqno);
426 hammer_modify_volume_done(root_volume);
427 hammer_unlock(&hmp->undo_lock);
430 hammer_rel_buffer(buffer, 0);
437 * It is not necessary to layout an undo record for the same address space
438 * multiple times. Maintain a cache of recent undo's.
442 * Enter an undo into the history. Return EALREADY if the request completely
443 * covers a previous request.
446 hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset, int bytes)
449 hammer_undo_t onode __debugvar;
451 node = RB_LOOKUP(hammer_und_rb_tree, &hmp->rb_undo_root, offset);
453 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
454 TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
455 if (bytes <= node->bytes)
460 if (hmp->undo_alloc != HAMMER_MAX_UNDOS) {
461 node = &hmp->undos[hmp->undo_alloc++];
463 node = TAILQ_FIRST(&hmp->undo_lru_list);
464 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
465 RB_REMOVE(hammer_und_rb_tree, &hmp->rb_undo_root, node);
467 node->offset = offset;
469 TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
470 onode = RB_INSERT(hammer_und_rb_tree, &hmp->rb_undo_root, node);
471 KKASSERT(onode == NULL);
476 hammer_clear_undo_history(hammer_mount_t hmp)
478 RB_INIT(&hmp->rb_undo_root);
479 TAILQ_INIT(&hmp->undo_lru_list);
484 * Return how much of the undo FIFO has been used
486 * The calculation includes undo FIFO space still reserved from a previous
487 * flush (because it will still be run on recovery if a crash occurs and
488 * we can't overwrite it yet).
491 hammer_undo_used(hammer_transaction_t trans)
493 hammer_blockmap_t cundomap;
494 hammer_blockmap_t dundomap;
495 int64_t max_bytes __debugvar;
498 cundomap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
499 dundomap = &trans->rootvol->ondisk->
500 vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
502 if (dundomap->first_offset <= cundomap->next_offset) {
503 bytes = cundomap->next_offset - dundomap->first_offset;
505 bytes = cundomap->alloc_offset - dundomap->first_offset +
506 (cundomap->next_offset & HAMMER_OFF_LONG_MASK);
508 max_bytes = cundomap->alloc_offset & HAMMER_OFF_SHORT_MASK;
509 KKASSERT(bytes <= max_bytes);
514 * Return how much of the undo FIFO is available for new records.
517 hammer_undo_space(hammer_transaction_t trans)
519 hammer_blockmap_t rootmap;
522 rootmap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
523 max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
524 return(max_bytes - hammer_undo_used(trans));
528 hammer_undo_max(hammer_mount_t hmp)
530 hammer_blockmap_t rootmap;
533 rootmap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
534 max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
540 * Returns 1 if the undo buffer should be reclaimed on release. The
541 * only undo buffer we do NOT want to reclaim is the one at the current
545 hammer_undo_reclaim(hammer_io_t io)
547 hammer_blockmap_t undomap;
548 hammer_off_t next_offset;
550 undomap = &io->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
551 next_offset = undomap->next_offset & ~HAMMER_BUFMASK64;
552 if (HAMMER_ITOB(io)->zoneX_offset == next_offset)