1 /*
2 * Copyright (c) 2008 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 *
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
16 * distribution.
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.
20 *
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
32 * SUCH DAMAGE.
33 */
34
35 /*
36 * HAMMER undo - undo buffer/FIFO management.
37 */
38
39 #include "hammer.h"
40
41 static int
hammer_und_rb_compare(hammer_undo_t node1,hammer_undo_t node2)42 hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2)
43 {
44 if (node1->offset < node2->offset)
45 return(-1);
46 if (node1->offset > node2->offset)
47 return(1);
48 return(0);
49 }
50
51 RB_GENERATE2(hammer_und_rb_tree, hammer_undo, rb_node,
52 hammer_und_rb_compare, hammer_off_t, offset);
53
54 /*
55 * Convert a zone-3 undo offset into a zone-2 buffer offset.
56 */
57 hammer_off_t
hammer_undo_lookup(hammer_mount_t hmp,hammer_off_t zone3_off,int * errorp)58 hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t zone3_off, int *errorp)
59 {
60 hammer_volume_t root_volume;
61 hammer_blockmap_t undomap __debugvar;
62 hammer_off_t result_offset;
63
64 KKASSERT(hammer_is_zone_undo(zone3_off));
65 root_volume = hammer_get_root_volume(hmp, errorp);
66 if (*errorp)
67 return(0);
68 undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
69 KKASSERT(hammer_is_zone_undo(undomap->alloc_offset));
70 KKASSERT(zone3_off < undomap->alloc_offset);
71
72 result_offset = hammer_xlate_to_undo(root_volume->ondisk, zone3_off);
73
74 hammer_rel_volume(root_volume, 0);
75 return(result_offset);
76 }
77
78 /*
79 * Generate UNDO record(s) for the block of data at the specified zone1
80 * or zone2 offset.
81 *
82 * The recovery code will execute UNDOs in reverse order, allowing overlaps.
83 * All the UNDOs are executed together so if we already laid one down we
84 * do not have to lay another one down for the same range.
85 *
86 * For HAMMER version 4+ UNDO a 512 byte boundary is enforced and a PAD
87 * will be laid down for any unused space. UNDO FIFO media structures
88 * will implement the hdr_seq field (it used to be reserved01), and
89 * both flush and recovery mechanics will be very different.
90 *
91 * WARNING! See also hammer_generate_redo() in hammer_redo.c
92 */
93 int
hammer_generate_undo(hammer_transaction_t trans,hammer_off_t zone_off,void * base,int len)94 hammer_generate_undo(hammer_transaction_t trans,
95 hammer_off_t zone_off, void *base, int len)
96 {
97 hammer_mount_t hmp;
98 hammer_volume_t root_volume;
99 hammer_blockmap_t undomap;
100 hammer_buffer_t buffer = NULL;
101 hammer_fifo_undo_t undo;
102 hammer_fifo_tail_t tail;
103 hammer_off_t next_offset;
104 int error;
105 int bytes;
106 int n;
107
108 hmp = trans->hmp;
109
110 /*
111 * A SYNC record may be required before we can lay down a general
112 * UNDO. This ensures that the nominal recovery span contains
113 * at least one SYNC record telling the recovery code how far
114 * out-of-span it must go to run the REDOs.
115 */
116 if ((hmp->flags & HAMMER_MOUNT_REDO_SYNC) == 0 &&
117 hmp->version >= HAMMER_VOL_VERSION_FOUR) {
118 hammer_generate_redo_sync(trans);
119 }
120
121 /*
122 * Enter the offset into our undo history. If there is an existing
123 * undo we do not have to generate a new one.
124 */
125 if (hammer_enter_undo_history(hmp, zone_off, len) == EALREADY)
126 return(0);
127
128 root_volume = trans->rootvol;
129 undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
130
131 /* no undo recursion */
132 hammer_modify_volume_noundo(NULL, root_volume);
133 hammer_lock_ex(&hmp->undo_lock);
134
135 /* undo had better not roll over (loose test) */
136 if (hammer_undo_space(trans) < len + HAMMER_BUFSIZE*3)
137 hpanic("insufficient UNDO/REDO FIFO space for undo!");
138
139 /*
140 * Loop until the undo for the entire range has been laid down.
141 */
142 while (len) {
143 /*
144 * Fetch the layout offset in the UNDO FIFO, wrap it as
145 * necessary.
146 */
147 if (undomap->next_offset == undomap->alloc_offset)
148 undomap->next_offset = HAMMER_ENCODE_UNDO(0);
149 next_offset = undomap->next_offset;
150
151 /*
152 * This is a tail-chasing FIFO, when we hit the start of a new
153 * buffer we don't have to read it in.
154 */
155 if ((next_offset & HAMMER_BUFMASK) == 0) {
156 undo = hammer_bnew(hmp, next_offset, &error, &buffer);
157 hammer_format_undo(hmp, undo,
158 hmp->undo_seqno ^ 0x40000000);
159 } else {
160 undo = hammer_bread(hmp, next_offset, &error, &buffer);
161 }
162 if (error)
163 break;
164 /* no undo recursion */
165 hammer_modify_buffer_noundo(NULL, buffer);
166
167 /*
168 * Calculate how big a media structure fits up to the next
169 * alignment point and how large a data payload we can
170 * accomodate.
171 *
172 * If n calculates to 0 or negative there is no room for
173 * anything but a PAD.
174 */
175 bytes = HAMMER_UNDO_ALIGN -
176 ((int)next_offset & HAMMER_UNDO_MASK);
177 n = bytes -
178 (int)sizeof(struct hammer_fifo_undo) -
179 (int)sizeof(struct hammer_fifo_tail);
180
181 /*
182 * If available space is insufficient for any payload
183 * we have to lay down a PAD.
184 *
185 * The minimum PAD is 8 bytes and the head and tail will
186 * overlap each other in that case. PADs do not have
187 * sequence numbers or CRCs.
188 *
189 * A PAD may not start on a boundary. That is, every
190 * 512-byte block in the UNDO/REDO FIFO must begin with
191 * a record containing a sequence number.
192 */
193 if (n <= 0) {
194 KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
195 KKASSERT(((int)next_offset & HAMMER_UNDO_MASK) != 0);
196 tail = (void *)((char *)undo + bytes - sizeof(*tail));
197 if ((void *)undo != (void *)tail) {
198 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
199 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
200 tail->tail_size = bytes;
201 }
202 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
203 undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
204 undo->head.hdr_size = bytes;
205 /* NO CRC OR SEQ NO */
206 undomap->next_offset += bytes;
207 hammer_modify_buffer_done(buffer);
208 hammer_stats_undo += bytes;
209 continue;
210 }
211
212 /*
213 * Calculate the actual payload and recalculate the size
214 * of the media structure as necessary.
215 */
216 if (n > len) {
217 n = len;
218 bytes = HAMMER_HEAD_DOALIGN(n) +
219 (int)sizeof(struct hammer_fifo_undo) +
220 (int)sizeof(struct hammer_fifo_tail);
221 }
222 if (hammer_debug_general & 0x0080) {
223 hdkprintf("undo %016jx %d %d\n",
224 (intmax_t)next_offset, bytes, n);
225 }
226
227 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
228 undo->head.hdr_type = HAMMER_HEAD_TYPE_UNDO;
229 undo->head.hdr_size = bytes;
230 undo->head.hdr_seq = hmp->undo_seqno++;
231 undo->head.hdr_crc = 0;
232 undo->undo_offset = zone_off;
233 undo->undo_data_bytes = n;
234 bcopy(base, undo + 1, n);
235
236 tail = (void *)((char *)undo + bytes - sizeof(*tail));
237 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
238 tail->tail_type = HAMMER_HEAD_TYPE_UNDO;
239 tail->tail_size = bytes;
240
241 KKASSERT(bytes >= sizeof(undo->head));
242 hammer_crc_set_fifo_head(hmp->version, &undo->head, bytes);
243 undomap->next_offset += bytes;
244 hammer_stats_undo += bytes;
245
246 /*
247 * Before we finish off the buffer we have to deal with any
248 * junk between the end of the media structure we just laid
249 * down and the UNDO alignment boundary. We do this by laying
250 * down a dummy PAD. Even though we will probably overwrite
251 * it almost immediately we have to do this so recovery runs
252 * can iterate the UNDO space without having to depend on
253 * the indices in the volume header.
254 *
255 * This dummy PAD will be overwritten on the next undo so
256 * we do not adjust undomap->next_offset.
257 */
258 bytes = HAMMER_UNDO_ALIGN -
259 ((int)undomap->next_offset & HAMMER_UNDO_MASK);
260 if (bytes != HAMMER_UNDO_ALIGN) {
261 KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
262 undo = (void *)(tail + 1);
263 tail = (void *)((char *)undo + bytes - sizeof(*tail));
264 if ((void *)undo != (void *)tail) {
265 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
266 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
267 tail->tail_size = bytes;
268 }
269 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
270 undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
271 undo->head.hdr_size = bytes;
272 /* NO CRC OR SEQ NO */
273 }
274 hammer_modify_buffer_done(buffer);
275
276 /*
277 * Adjust for loop
278 */
279 len -= n;
280 base = (char *)base + n;
281 zone_off += n;
282 }
283 hammer_modify_volume_done(root_volume);
284 hammer_unlock(&hmp->undo_lock);
285
286 if (buffer)
287 hammer_rel_buffer(buffer, 0);
288 return(error);
289 }
290
291 /*
292 * Preformat a new UNDO block. We could read the old one in but we get
293 * better performance if we just pre-format a new one.
294 *
295 * The recovery code always works forwards so the caller just makes sure the
296 * seqno is not contiguous with prior UNDOs or ancient UNDOs now being
297 * overwritten.
298 *
299 * The preformatted UNDO headers use the smallest possible sector size
300 * (512) to ensure that any missed media writes are caught.
301 *
302 * NOTE: Also used by the REDO code.
303 */
304 void
hammer_format_undo(hammer_mount_t hmp,void * base,uint32_t seqno)305 hammer_format_undo(hammer_mount_t hmp, void *base, uint32_t seqno)
306 {
307 hammer_fifo_head_t head;
308 hammer_fifo_tail_t tail;
309 int i;
310 int bytes = HAMMER_UNDO_ALIGN;
311
312 bzero(base, HAMMER_BUFSIZE);
313
314 for (i = 0; i < HAMMER_BUFSIZE; i += bytes) {
315 head = (void *)((char *)base + i);
316 tail = (void *)((char *)head + bytes - sizeof(*tail));
317
318 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
319 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
320 head->hdr_size = bytes;
321 head->hdr_seq = seqno++;
322 head->hdr_crc = 0;
323
324 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
325 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
326 tail->tail_size = bytes;
327
328 hammer_crc_set_fifo_head(hmp->version, head, bytes);
329 }
330 }
331
332 /*
333 * HAMMER version 4+ conversion support.
334 *
335 * Convert a HAMMER version < 4 UNDO FIFO area to a 4+ UNDO FIFO area.
336 * The 4+ UNDO FIFO area is backwards compatible. The conversion is
337 * needed to initialize the sequence space and place headers on the
338 * new 512-byte undo boundary.
339 */
340 int
hammer_upgrade_undo_4(hammer_transaction_t trans)341 hammer_upgrade_undo_4(hammer_transaction_t trans)
342 {
343 hammer_mount_t hmp;
344 hammer_volume_t root_volume;
345 hammer_blockmap_t undomap;
346 hammer_buffer_t buffer = NULL;
347 hammer_fifo_head_t head;
348 hammer_fifo_tail_t tail;
349 hammer_off_t next_offset;
350 uint32_t seqno;
351 int error;
352 int bytes;
353
354 hmp = trans->hmp;
355
356 root_volume = trans->rootvol;
357
358 /* no undo recursion */
359 hammer_lock_ex(&hmp->undo_lock);
360 hammer_modify_volume_noundo(NULL, root_volume);
361
362 /*
363 * Adjust the in-core undomap and the on-disk undomap.
364 */
365 next_offset = HAMMER_ENCODE_UNDO(0);
366 undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
367 undomap->next_offset = next_offset;
368 undomap->first_offset = next_offset;
369
370 undomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
371 undomap->next_offset = next_offset;
372 undomap->first_offset = next_offset;
373
374 /*
375 * Loop over the entire UNDO space creating DUMMY entries. Sequence
376 * numbers are assigned.
377 */
378 seqno = 0;
379 bytes = HAMMER_UNDO_ALIGN;
380
381 while (next_offset != undomap->alloc_offset) {
382 head = hammer_bnew(hmp, next_offset, &error, &buffer);
383 if (error)
384 break;
385 hammer_modify_buffer_noundo(NULL, buffer);
386 tail = (void *)((char *)head + bytes - sizeof(*tail));
387
388 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
389 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
390 head->hdr_size = bytes;
391 head->hdr_seq = seqno;
392 head->hdr_crc = 0;
393
394 tail = (void *)((char *)head + bytes - sizeof(*tail));
395 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
396 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
397 tail->tail_size = bytes;
398
399 hammer_crc_set_fifo_head(hmp->version, head, bytes);
400 hammer_modify_buffer_done(buffer);
401
402 hammer_stats_undo += bytes;
403 next_offset += HAMMER_UNDO_ALIGN;
404 ++seqno;
405 }
406
407 /*
408 * The sequence number will be the next sequence number to lay down.
409 */
410 hmp->undo_seqno = seqno;
411 hmkprintf(hmp, "version upgrade seqno start %08x\n", seqno);
412
413 hammer_modify_volume_done(root_volume);
414 hammer_unlock(&hmp->undo_lock);
415
416 if (buffer)
417 hammer_rel_buffer(buffer, 0);
418 return (error);
419 }
420
421 /*
422 * UNDO HISTORY API
423 *
424 * It is not necessary to layout an undo record for the same address space
425 * multiple times. Maintain a cache of recent undo's.
426 */
427
428 /*
429 * Enter an undo into the history. Return EALREADY if the request completely
430 * covers a previous request.
431 */
432 int
hammer_enter_undo_history(hammer_mount_t hmp,hammer_off_t offset,int bytes)433 hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset, int bytes)
434 {
435 hammer_undo_t node;
436 hammer_undo_t onode __debugvar;
437
438 node = RB_LOOKUP(hammer_und_rb_tree, &hmp->rb_undo_root, offset);
439 if (node) {
440 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
441 TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
442 if (bytes <= node->bytes)
443 return(EALREADY);
444 node->bytes = bytes;
445 return(0);
446 }
447 if (hmp->undo_alloc != HAMMER_MAX_UNDOS) {
448 node = &hmp->undos[hmp->undo_alloc++];
449 } else {
450 node = TAILQ_FIRST(&hmp->undo_lru_list);
451 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
452 RB_REMOVE(hammer_und_rb_tree, &hmp->rb_undo_root, node);
453 }
454 node->offset = offset;
455 node->bytes = bytes;
456 TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
457 onode = RB_INSERT(hammer_und_rb_tree, &hmp->rb_undo_root, node);
458 KKASSERT(onode == NULL);
459 return(0);
460 }
461
462 void
hammer_clear_undo_history(hammer_mount_t hmp)463 hammer_clear_undo_history(hammer_mount_t hmp)
464 {
465 RB_INIT(&hmp->rb_undo_root);
466 TAILQ_INIT(&hmp->undo_lru_list);
467 hmp->undo_alloc = 0;
468 }
469
470 /*
471 * Return how much of the undo FIFO has been used
472 *
473 * The calculation includes undo FIFO space still reserved from a previous
474 * flush (because it will still be run on recovery if a crash occurs and
475 * we can't overwrite it yet).
476 */
477 int64_t
hammer_undo_used(hammer_transaction_t trans)478 hammer_undo_used(hammer_transaction_t trans)
479 {
480 hammer_blockmap_t cundomap;
481 hammer_blockmap_t dundomap;
482 int64_t max_bytes __debugvar;
483 int64_t bytes;
484
485 cundomap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
486 dundomap = &trans->rootvol->ondisk->
487 vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
488
489 if (dundomap->first_offset <= cundomap->next_offset) {
490 bytes = cundomap->next_offset - dundomap->first_offset;
491 } else {
492 bytes = cundomap->alloc_offset - dundomap->first_offset +
493 HAMMER_OFF_LONG_ENCODE(cundomap->next_offset);
494 }
495 max_bytes = HAMMER_OFF_SHORT_ENCODE(cundomap->alloc_offset);
496 KKASSERT(bytes <= max_bytes);
497 return(bytes);
498 }
499
500 /*
501 * Return how much of the undo FIFO is available for new records.
502 */
503 int64_t
hammer_undo_space(hammer_transaction_t trans)504 hammer_undo_space(hammer_transaction_t trans)
505 {
506 hammer_blockmap_t rootmap;
507 int64_t max_bytes;
508
509 rootmap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
510 max_bytes = HAMMER_OFF_SHORT_ENCODE(rootmap->alloc_offset);
511 return(max_bytes - hammer_undo_used(trans));
512 }
513
514 int64_t
hammer_undo_max(hammer_mount_t hmp)515 hammer_undo_max(hammer_mount_t hmp)
516 {
517 hammer_blockmap_t rootmap;
518 int64_t max_bytes;
519
520 rootmap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
521 max_bytes = HAMMER_OFF_SHORT_ENCODE(rootmap->alloc_offset);
522
523 return(max_bytes);
524 }
525
526 /*
527 * Returns 1 if the undo buffer should be reclaimed on release. The
528 * only undo buffer we do NOT want to reclaim is the one at the current
529 * append offset.
530 */
531 int
hammer_undo_reclaim(hammer_io_t io)532 hammer_undo_reclaim(hammer_io_t io)
533 {
534 hammer_blockmap_t undomap;
535 hammer_off_t next_offset;
536
537 undomap = &io->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
538 next_offset = undomap->next_offset & ~HAMMER_BUFMASK64;
539 if (HAMMER_ITOB(io)->zoneX_offset == next_offset)
540 return(0);
541 return(1);
542 }
543