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.
41 static int hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2);
43 RB_GENERATE2(hammer_und_rb_tree, hammer_undo, rb_node,
44 hammer_und_rb_compare, hammer_off_t, offset);
47 * Convert a zone-3 undo offset into a zone-2 buffer offset.
50 hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t zone3_off, int *errorp)
52 hammer_volume_t root_volume;
53 hammer_blockmap_t undomap __debugvar;
54 hammer_off_t result_offset;
57 KKASSERT((zone3_off & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_UNDO);
58 root_volume = hammer_get_root_volume(hmp, errorp);
61 undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
62 KKASSERT(HAMMER_ZONE_DECODE(undomap->alloc_offset) == HAMMER_ZONE_UNDO_INDEX);
63 KKASSERT(zone3_off < undomap->alloc_offset);
65 i = (zone3_off & HAMMER_OFF_SHORT_MASK) / HAMMER_LARGEBLOCK_SIZE;
66 result_offset = root_volume->ondisk->vol0_undo_array[i] +
67 (zone3_off & HAMMER_LARGEBLOCK_MASK64);
69 hammer_rel_volume(root_volume, 0);
70 return(result_offset);
74 * Generate UNDO record(s) for the block of data at the specified zone1
77 * The recovery code will execute UNDOs in reverse order, allowing overlaps.
78 * All the UNDOs are executed together so if we already laid one down we
79 * do not have to lay another one down for the same range.
81 * For HAMMER version 4+ UNDO a 512 byte boundary is enforced and a PAD
82 * will be laid down for any unused space. UNDO FIFO media structures
83 * will implement the hdr_seq field (it used to be reserved01), and
84 * both flush and recovery mechanics will be very different.
86 * WARNING! See also hammer_generate_redo() in hammer_redo.c
89 hammer_generate_undo(hammer_transaction_t trans,
90 hammer_off_t zone_off, void *base, int len)
93 hammer_volume_t root_volume;
94 hammer_blockmap_t undomap;
95 hammer_buffer_t buffer = NULL;
96 hammer_fifo_undo_t undo;
97 hammer_fifo_tail_t tail;
98 hammer_off_t next_offset;
106 * A SYNC record may be required before we can lay down a general
107 * UNDO. This ensures that the nominal recovery span contains
108 * at least one SYNC record telling the recovery code how far
109 * out-of-span it must go to run the REDOs.
111 if ((hmp->flags & HAMMER_MOUNT_REDO_SYNC) == 0 &&
112 hmp->version >= HAMMER_VOL_VERSION_FOUR) {
113 hammer_generate_redo_sync(trans);
117 * Enter the offset into our undo history. If there is an existing
118 * undo we do not have to generate a new one.
120 if (hammer_enter_undo_history(hmp, zone_off, len) == EALREADY)
123 root_volume = trans->rootvol;
124 undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
126 /* no undo recursion */
127 hammer_modify_volume(NULL, root_volume, NULL, 0);
128 hammer_lock_ex(&hmp->undo_lock);
130 /* undo had better not roll over (loose test) */
131 if (hammer_undo_space(trans) < len + HAMMER_BUFSIZE*3)
132 panic("hammer: insufficient undo FIFO space!");
135 * Loop until the undo for the entire range has been laid down.
139 * Fetch the layout offset in the UNDO FIFO, wrap it as
142 if (undomap->next_offset == undomap->alloc_offset) {
143 undomap->next_offset =
144 HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
146 next_offset = undomap->next_offset;
149 * This is a tail-chasing FIFO, when we hit the start of a new
150 * buffer we don't have to read it in.
152 if ((next_offset & HAMMER_BUFMASK) == 0) {
153 undo = hammer_bnew(hmp, next_offset, &error, &buffer);
154 hammer_format_undo(undo, hmp->undo_seqno ^ 0x40000000);
156 undo = hammer_bread(hmp, next_offset, &error, &buffer);
160 hammer_modify_buffer(NULL, buffer, NULL, 0);
163 * Calculate how big a media structure fits up to the next
164 * alignment point and how large a data payload we can
167 * If n calculates to 0 or negative there is no room for
168 * anything but a PAD.
170 bytes = HAMMER_UNDO_ALIGN -
171 ((int)next_offset & HAMMER_UNDO_MASK);
173 (int)sizeof(struct hammer_fifo_undo) -
174 (int)sizeof(struct hammer_fifo_tail);
177 * If available space is insufficient for any payload
178 * we have to lay down a PAD.
180 * The minimum PAD is 8 bytes and the head and tail will
181 * overlap each other in that case. PADs do not have
182 * sequence numbers or CRCs.
184 * A PAD may not start on a boundary. That is, every
185 * 512-byte block in the UNDO/REDO FIFO must begin with
186 * a record containing a sequence number.
189 KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
190 KKASSERT(((int)next_offset & HAMMER_UNDO_MASK) != 0);
191 tail = (void *)((char *)undo + bytes - sizeof(*tail));
192 if ((void *)undo != (void *)tail) {
193 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
194 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
195 tail->tail_size = bytes;
197 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
198 undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
199 undo->head.hdr_size = bytes;
200 /* NO CRC OR SEQ NO */
201 undomap->next_offset += bytes;
202 hammer_modify_buffer_done(buffer);
203 hammer_stats_undo += bytes;
208 * Calculate the actual payload and recalculate the size
209 * of the media structure as necessary.
213 bytes = ((n + HAMMER_HEAD_ALIGN_MASK) &
214 ~HAMMER_HEAD_ALIGN_MASK) +
215 (int)sizeof(struct hammer_fifo_undo) +
216 (int)sizeof(struct hammer_fifo_tail);
218 if (hammer_debug_general & 0x0080) {
219 kprintf("undo %016llx %d %d\n",
220 (long long)next_offset, bytes, n);
223 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
224 undo->head.hdr_type = HAMMER_HEAD_TYPE_UNDO;
225 undo->head.hdr_size = bytes;
226 undo->head.hdr_seq = hmp->undo_seqno++;
227 undo->head.hdr_crc = 0;
228 undo->undo_offset = zone_off;
229 undo->undo_data_bytes = n;
230 bcopy(base, undo + 1, n);
232 tail = (void *)((char *)undo + bytes - sizeof(*tail));
233 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
234 tail->tail_type = HAMMER_HEAD_TYPE_UNDO;
235 tail->tail_size = bytes;
237 KKASSERT(bytes >= sizeof(undo->head));
238 undo->head.hdr_crc = crc32(undo, HAMMER_FIFO_HEAD_CRCOFF) ^
239 crc32(&undo->head + 1, bytes - sizeof(undo->head));
240 undomap->next_offset += bytes;
241 hammer_stats_undo += bytes;
244 * Before we finish off the buffer we have to deal with any
245 * junk between the end of the media structure we just laid
246 * down and the UNDO alignment boundary. We do this by laying
247 * down a dummy PAD. Even though we will probably overwrite
248 * it almost immediately we have to do this so recovery runs
249 * can iterate the UNDO space without having to depend on
250 * the indices in the volume header.
252 * This dummy PAD will be overwritten on the next undo so
253 * we do not adjust undomap->next_offset.
255 bytes = HAMMER_UNDO_ALIGN -
256 ((int)undomap->next_offset & HAMMER_UNDO_MASK);
257 if (bytes != HAMMER_UNDO_ALIGN) {
258 KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
259 undo = (void *)(tail + 1);
260 tail = (void *)((char *)undo + bytes - sizeof(*tail));
261 if ((void *)undo != (void *)tail) {
262 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
263 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
264 tail->tail_size = bytes;
266 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
267 undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
268 undo->head.hdr_size = bytes;
269 /* NO CRC OR SEQ NO */
271 hammer_modify_buffer_done(buffer);
277 base = (char *)base + n;
280 hammer_modify_volume_done(root_volume);
281 hammer_unlock(&hmp->undo_lock);
284 hammer_rel_buffer(buffer, 0);
289 * Preformat a new UNDO block. We could read the old one in but we get
290 * better performance if we just pre-format a new one.
292 * The recovery code always works forwards so the caller just makes sure the
293 * seqno is not contiguous with prior UNDOs or ancient UNDOs now being
296 * The preformatted UNDO headers use the smallest possible sector size
297 * (512) to ensure that any missed media writes are caught.
299 * NOTE: Also used by the REDO code.
302 hammer_format_undo(void *base, u_int32_t seqno)
304 hammer_fifo_head_t head;
305 hammer_fifo_tail_t tail;
307 int bytes = HAMMER_UNDO_ALIGN;
309 bzero(base, HAMMER_BUFSIZE);
311 for (i = 0; i < HAMMER_BUFSIZE; i += bytes) {
312 head = (void *)((char *)base + i);
313 tail = (void *)((char *)head + bytes - sizeof(*tail));
315 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
316 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
317 head->hdr_size = bytes;
318 head->hdr_seq = seqno++;
321 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
322 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
323 tail->tail_size = bytes;
325 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
326 crc32(head + 1, bytes - sizeof(*head));
331 * HAMMER version 4+ conversion support.
333 * Convert a HAMMER version < 4 UNDO FIFO area to a 4+ UNDO FIFO area.
334 * The 4+ UNDO FIFO area is backwards compatible. The conversion is
335 * needed to initialize the sequence space and place headers on the
336 * new 512-byte undo boundary.
339 hammer_upgrade_undo_4(hammer_transaction_t trans)
342 hammer_volume_t root_volume;
343 hammer_blockmap_t undomap;
344 hammer_buffer_t buffer = NULL;
345 hammer_fifo_head_t head;
346 hammer_fifo_tail_t tail;
347 hammer_off_t next_offset;
354 root_volume = trans->rootvol;
356 /* no undo recursion */
357 hammer_lock_ex(&hmp->undo_lock);
358 hammer_modify_volume(NULL, root_volume, NULL, 0);
361 * Adjust the in-core undomap and the on-disk undomap.
363 next_offset = HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
364 undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
365 undomap->next_offset = next_offset;
366 undomap->first_offset = next_offset;
368 undomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
369 undomap->next_offset = next_offset;
370 undomap->first_offset = next_offset;
373 * Loop over the entire UNDO space creating DUMMY entries. Sequence
374 * numbers are assigned.
377 bytes = HAMMER_UNDO_ALIGN;
379 while (next_offset != undomap->alloc_offset) {
380 head = hammer_bnew(hmp, next_offset, &error, &buffer);
383 hammer_modify_buffer(NULL, buffer, NULL, 0);
384 tail = (void *)((char *)head + bytes - sizeof(*tail));
386 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
387 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
388 head->hdr_size = bytes;
389 head->hdr_seq = seqno;
392 tail = (void *)((char *)head + bytes - sizeof(*tail));
393 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
394 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
395 tail->tail_size = bytes;
397 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
398 crc32(head + 1, bytes - sizeof(*head));
399 hammer_modify_buffer_done(buffer);
401 hammer_stats_undo += bytes;
402 next_offset += HAMMER_UNDO_ALIGN;
407 * The sequence number will be the next sequence number to lay down.
409 hmp->undo_seqno = seqno;
410 kprintf("version upgrade seqno start %08x\n", seqno);
412 hammer_modify_volume_done(root_volume);
413 hammer_unlock(&hmp->undo_lock);
416 hammer_rel_buffer(buffer, 0);
423 * It is not necessary to layout an undo record for the same address space
424 * multiple times. Maintain a cache of recent undo's.
428 * Enter an undo into the history. Return EALREADY if the request completely
429 * covers a previous request.
432 hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset, int bytes)
435 hammer_undo_t onode __debugvar;
437 node = RB_LOOKUP(hammer_und_rb_tree, &hmp->rb_undo_root, offset);
439 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
440 TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
441 if (bytes <= node->bytes)
446 if (hmp->undo_alloc != HAMMER_MAX_UNDOS) {
447 node = &hmp->undos[hmp->undo_alloc++];
449 node = TAILQ_FIRST(&hmp->undo_lru_list);
450 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
451 RB_REMOVE(hammer_und_rb_tree, &hmp->rb_undo_root, node);
453 node->offset = offset;
455 TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
456 onode = RB_INSERT(hammer_und_rb_tree, &hmp->rb_undo_root, node);
457 KKASSERT(onode == NULL);
462 hammer_clear_undo_history(hammer_mount_t hmp)
464 RB_INIT(&hmp->rb_undo_root);
465 TAILQ_INIT(&hmp->undo_lru_list);
470 * Return how much of the undo FIFO has been used
472 * The calculation includes undo FIFO space still reserved from a previous
473 * flush (because it will still be run on recovery if a crash occurs and
474 * we can't overwrite it yet).
477 hammer_undo_used(hammer_transaction_t trans)
479 hammer_blockmap_t cundomap;
480 hammer_blockmap_t dundomap;
481 int64_t max_bytes __debugvar;
484 cundomap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
485 dundomap = &trans->rootvol->ondisk->
486 vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
488 if (dundomap->first_offset <= cundomap->next_offset) {
489 bytes = cundomap->next_offset - dundomap->first_offset;
491 bytes = cundomap->alloc_offset - dundomap->first_offset +
492 (cundomap->next_offset & HAMMER_OFF_LONG_MASK);
494 max_bytes = cundomap->alloc_offset & HAMMER_OFF_SHORT_MASK;
495 KKASSERT(bytes <= max_bytes);
500 * Return how much of the undo FIFO is available for new records.
503 hammer_undo_space(hammer_transaction_t trans)
505 hammer_blockmap_t rootmap;
508 rootmap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
509 max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
510 return(max_bytes - hammer_undo_used(trans));
514 hammer_undo_max(hammer_mount_t hmp)
516 hammer_blockmap_t rootmap;
519 rootmap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
520 max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
526 * Returns 1 if the undo buffer should be reclaimed on release. The
527 * only undo buffer we do NOT want to reclaim is the one at the current
531 hammer_undo_reclaim(hammer_io_t io)
533 hammer_blockmap_t undomap;
534 hammer_off_t next_offset;
536 undomap = &io->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
537 next_offset = undomap->next_offset & ~HAMMER_BUFMASK64;
538 if (((struct hammer_buffer *)io)->zoneX_offset == next_offset)
544 hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2)
546 if (node1->offset < node2->offset)
548 if (node1->offset > node2->offset)