kernel - Fix excessive mbuf use in nfs_realign()
[dragonfly.git] / sys / vfs / hammer / hammer_undo.c
CommitLineData
bf686dbe
MD
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 *
cdb6e4e6 34 * $DragonFly: src/sys/vfs/hammer/hammer_undo.c,v 1.20 2008/07/18 00:19:53 dillon Exp $
bf686dbe
MD
35 */
36
37/*
38 * HAMMER undo - undo buffer/FIFO management.
39 */
40
41#include "hammer.h"
42
e8599db1
MD
43static int hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2);
44
45RB_GENERATE2(hammer_und_rb_tree, hammer_undo, rb_node,
46 hammer_und_rb_compare, hammer_off_t, offset);
47
bf686dbe
MD
48/*
49 * Convert a zone-3 undo offset into a zone-2 buffer offset.
50 */
51hammer_off_t
52hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t zone3_off, int *errorp)
53{
54 hammer_volume_t root_volume;
55 hammer_blockmap_t undomap;
bf686dbe
MD
56 hammer_off_t result_offset;
57 int i;
58
59 KKASSERT((zone3_off & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_UNDO);
60 root_volume = hammer_get_root_volume(hmp, errorp);
61 if (*errorp)
62 return(0);
0729c8c8 63 undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
bf686dbe
MD
64 KKASSERT(HAMMER_ZONE_DECODE(undomap->alloc_offset) == HAMMER_ZONE_UNDO_INDEX);
65 KKASSERT (zone3_off < undomap->alloc_offset);
66
67 i = (zone3_off & HAMMER_OFF_SHORT_MASK) / HAMMER_LARGEBLOCK_SIZE;
cb51be26 68 result_offset = root_volume->ondisk->vol0_undo_array[i] +
bf686dbe
MD
69 (zone3_off & HAMMER_LARGEBLOCK_MASK64);
70
71 hammer_rel_volume(root_volume, 0);
72 return(result_offset);
73}
74
75/*
02428fb6 76 * Generate UNDO record(s) for the block of data at the specified zone1
f90dde4c 77 * or zone2 offset.
e8599db1
MD
78 *
79 * The recovery code will execute UNDOs in reverse order, allowing overlaps.
80 * All the UNDOs are executed together so if we already laid one down we
81 * do not have to lay another one down for the same range.
02428fb6
MD
82 *
83 * For HAMMER version 4+ UNDO a 512 byte boundary is enforced and a PAD
84 * will be laid down for any unused space. UNDO FIFO media structures
85 * will implement the hdr_seq field (it used to be reserved01), and
86 * both flush and recovery mechanics will be very different.
47f363f1
MD
87 *
88 * WARNING! See also hammer_generate_redo() in hammer_redo.c
bf686dbe
MD
89 */
90int
02428fb6 91hammer_generate_undo(hammer_transaction_t trans,
f90dde4c 92 hammer_off_t zone_off, void *base, int len)
bf686dbe 93{
d99d6bf5 94 hammer_mount_t hmp;
bf686dbe 95 hammer_volume_t root_volume;
bf686dbe
MD
96 hammer_blockmap_t undomap;
97 hammer_buffer_t buffer = NULL;
bf686dbe
MD
98 hammer_fifo_undo_t undo;
99 hammer_fifo_tail_t tail;
100 hammer_off_t next_offset;
bf686dbe
MD
101 int error;
102 int bytes;
02428fb6 103 int n;
bf686dbe 104
d99d6bf5
MD
105 hmp = trans->hmp;
106
47f363f1
MD
107 /*
108 * A SYNC record may be required before we can lay down a general
109 * UNDO. This ensures that the nominal recovery span contains
110 * at least one SYNC record telling the recovery code how far
111 * out-of-span it must go to run the REDOs.
112 */
113 if ((hmp->flags & HAMMER_MOUNT_REDO_SYNC) == 0 &&
114 hmp->version >= HAMMER_VOL_VERSION_FOUR) {
115 hammer_generate_redo_sync(trans);
116 }
117
e8599db1
MD
118 /*
119 * Enter the offset into our undo history. If there is an existing
120 * undo we do not have to generate a new one.
121 */
d99d6bf5 122 if (hammer_enter_undo_history(hmp, zone_off, len) == EALREADY)
e8599db1 123 return(0);
e8599db1 124
b58c6388 125 root_volume = trans->rootvol;
d99d6bf5 126 undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
36f82b23
MD
127
128 /* no undo recursion */
129 hammer_modify_volume(NULL, root_volume, NULL, 0);
d99d6bf5 130 hammer_lock_ex(&hmp->undo_lock);
f90dde4c 131
02428fb6
MD
132 /* undo had better not roll over (loose test) */
133 if (hammer_undo_space(trans) < len + HAMMER_BUFSIZE*3)
134 panic("hammer: insufficient undo FIFO space!");
bf686dbe
MD
135
136 /*
02428fb6 137 * Loop until the undo for the entire range has been laid down.
bf686dbe 138 */
02428fb6
MD
139 while (len) {
140 /*
141 * Fetch the layout offset in the UNDO FIFO, wrap it as
142 * necessary.
143 */
144 if (undomap->next_offset == undomap->alloc_offset) {
145 undomap->next_offset =
146 HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
147 }
148 next_offset = undomap->next_offset;
bf686dbe 149
02428fb6
MD
150 /*
151 * This is a tail-chasing FIFO, when we hit the start of a new
152 * buffer we don't have to read it in.
153 */
154 if ((next_offset & HAMMER_BUFMASK) == 0) {
155 undo = hammer_bnew(hmp, next_offset, &error, &buffer);
156 hammer_format_undo(undo, hmp->undo_seqno ^ 0x40000000);
157 } else {
158 undo = hammer_bread(hmp, next_offset, &error, &buffer);
159 }
160 if (error)
161 break;
162 hammer_modify_buffer(NULL, buffer, NULL, 0);
cdb6e4e6 163
02428fb6
MD
164 /*
165 * Calculate how big a media structure fits up to the next
166 * alignment point and how large a data payload we can
167 * accomodate.
168 *
169 * If n calculates to 0 or negative there is no room for
170 * anything but a PAD.
171 */
172 bytes = HAMMER_UNDO_ALIGN -
173 ((int)next_offset & HAMMER_UNDO_MASK);
174 n = bytes -
175 (int)sizeof(struct hammer_fifo_undo) -
176 (int)sizeof(struct hammer_fifo_tail);
d99d6bf5 177
02428fb6
MD
178 /*
179 * If available space is insufficient for any payload
180 * we have to lay down a PAD.
181 *
182 * The minimum PAD is 8 bytes and the head and tail will
183 * overlap each other in that case. PADs do not have
184 * sequence numbers or CRCs.
185 *
186 * A PAD may not start on a boundary. That is, every
187 * 512-byte block in the UNDO/REDO FIFO must begin with
188 * a record containing a sequence number.
189 */
190 if (n <= 0) {
191 KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
192 KKASSERT(((int)next_offset & HAMMER_UNDO_MASK) != 0);
193 tail = (void *)((char *)undo + bytes - sizeof(*tail));
194 if ((void *)undo != (void *)tail) {
195 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
196 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
197 tail->tail_size = bytes;
198 }
199 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
200 undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
201 undo->head.hdr_size = bytes;
202 /* NO CRC OR SEQ NO */
203 undomap->next_offset += bytes;
204 hammer_modify_buffer_done(buffer);
205 hammer_stats_undo += bytes;
206 continue;
207 }
c9b9e29d 208
02428fb6
MD
209 /*
210 * Calculate the actual payload and recalculate the size
211 * of the media structure as necessary.
212 */
213 if (n > len) {
214 n = len;
215 bytes = ((n + HAMMER_HEAD_ALIGN_MASK) &
216 ~HAMMER_HEAD_ALIGN_MASK) +
217 (int)sizeof(struct hammer_fifo_undo) +
218 (int)sizeof(struct hammer_fifo_tail);
bf686dbe 219 }
02428fb6
MD
220 if (hammer_debug_general & 0x0080) {
221 kprintf("undo %016llx %d %d\n",
222 (long long)next_offset, bytes, n);
223 }
224
bf686dbe 225 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
02428fb6 226 undo->head.hdr_type = HAMMER_HEAD_TYPE_UNDO;
bf686dbe 227 undo->head.hdr_size = bytes;
02428fb6
MD
228 undo->head.hdr_seq = hmp->undo_seqno++;
229 undo->head.hdr_crc = 0;
230 undo->undo_offset = zone_off;
231 undo->undo_data_bytes = n;
232 bcopy(base, undo + 1, n);
233
234 tail = (void *)((char *)undo + bytes - sizeof(*tail));
235 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
236 tail->tail_type = HAMMER_HEAD_TYPE_UNDO;
237 tail->tail_size = bytes;
238
239 KKASSERT(bytes >= sizeof(undo->head));
240 undo->head.hdr_crc = crc32(undo, HAMMER_FIFO_HEAD_CRCOFF) ^
241 crc32(&undo->head + 1, bytes - sizeof(undo->head));
bf686dbe 242 undomap->next_offset += bytes;
89e744ce 243 hammer_stats_undo += bytes;
02428fb6
MD
244
245 /*
246 * Before we finish off the buffer we have to deal with any
247 * junk between the end of the media structure we just laid
248 * down and the UNDO alignment boundary. We do this by laying
249 * down a dummy PAD. Even though we will probably overwrite
250 * it almost immediately we have to do this so recovery runs
251 * can iterate the UNDO space without having to depend on
252 * the indices in the volume header.
253 *
254 * This dummy PAD will be overwritten on the next undo so
255 * we do not adjust undomap->next_offset.
256 */
257 bytes = HAMMER_UNDO_ALIGN -
258 ((int)undomap->next_offset & HAMMER_UNDO_MASK);
259 if (bytes != HAMMER_UNDO_ALIGN) {
260 KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
261 undo = (void *)(tail + 1);
262 tail = (void *)((char *)undo + bytes - sizeof(*tail));
263 if ((void *)undo != (void *)tail) {
264 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
265 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
266 tail->tail_size = bytes;
267 }
268 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
269 undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
270 undo->head.hdr_size = bytes;
271 /* NO CRC OR SEQ NO */
272 }
273 hammer_modify_buffer_done(buffer);
274
275 /*
276 * Adjust for loop
277 */
278 len -= n;
279 base = (char *)base + n;
280 zone_off += n;
281 }
282 hammer_modify_volume_done(root_volume);
283 hammer_unlock(&hmp->undo_lock);
02428fb6
MD
284
285 if (buffer)
286 hammer_rel_buffer(buffer, 0);
287 return(error);
288}
289
02428fb6
MD
290/*
291 * Preformat a new UNDO block. We could read the old one in but we get
292 * better performance if we just pre-format a new one.
293 *
294 * The recovery code always works forwards so the caller just makes sure the
295 * seqno is not contiguous with prior UNDOs or ancient UNDOs now being
296 * overwritten.
6048b411
MD
297 *
298 * The preformatted UNDO headers use the smallest possible sector size
299 * (512) to ensure that any missed media writes are caught.
47f363f1
MD
300 *
301 * NOTE: Also used by the REDO code.
02428fb6 302 */
02428fb6
MD
303void
304hammer_format_undo(void *base, u_int32_t seqno)
305{
306 hammer_fifo_head_t head;
307 hammer_fifo_tail_t tail;
308 int i;
309 int bytes = HAMMER_UNDO_ALIGN;
310
311 bzero(base, HAMMER_BUFSIZE);
312
313 for (i = 0; i < HAMMER_BUFSIZE; i += bytes) {
314 head = (void *)((char *)base + i);
315 tail = (void *)((char *)head + bytes - sizeof(*tail));
316
317 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
318 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
319 head->hdr_size = bytes;
320 head->hdr_seq = seqno++;
321 head->hdr_crc = 0;
322
323 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
324 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
325 tail->tail_size = bytes;
326
327 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
328 crc32(head + 1, bytes - sizeof(*head));
bf686dbe 329 }
02428fb6
MD
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 */
340int
341hammer_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 u_int32_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(NULL, root_volume, NULL, 0);
361
362 /*
363 * Adjust the in-core undomap and the on-disk undomap.
364 */
365 next_offset = HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 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(NULL, buffer, NULL, 0);
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 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
400 crc32(head + 1, bytes - sizeof(*head));
401 hammer_modify_buffer_done(buffer);
402
403 hammer_stats_undo += bytes;
404 next_offset += HAMMER_UNDO_ALIGN;
405 ++seqno;
973c11b9 406 }
bf686dbe
MD
407
408 /*
02428fb6 409 * The sequence number will be the next sequence number to lay down.
bf686dbe 410 */
02428fb6
MD
411 hmp->undo_seqno = seqno;
412 kprintf("version upgrade seqno start %08x\n", seqno);
c9b9e29d 413
10a5d1ba 414 hammer_modify_volume_done(root_volume);
d99d6bf5
MD
415 hammer_unlock(&hmp->undo_lock);
416
bf686dbe
MD
417 if (buffer)
418 hammer_rel_buffer(buffer, 0);
02428fb6 419 return (error);
bf686dbe
MD
420}
421
e8599db1
MD
422/*
423 * UNDO HISTORY API
424 *
425 * It is not necessary to layout an undo record for the same address space
426 * multiple times. Maintain a cache of recent undo's.
427 */
428
429/*
430 * Enter an undo into the history. Return EALREADY if the request completely
431 * covers a previous request.
432 */
433int
434hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset, int bytes)
435{
436 hammer_undo_t node;
437 hammer_undo_t onode;
438
439 node = RB_LOOKUP(hammer_und_rb_tree, &hmp->rb_undo_root, offset);
440 if (node) {
441 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
442 TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
443 if (bytes <= node->bytes)
444 return(EALREADY);
445 node->bytes = bytes;
446 return(0);
447 }
448 if (hmp->undo_alloc != HAMMER_MAX_UNDOS) {
449 node = &hmp->undos[hmp->undo_alloc++];
450 } else {
451 node = TAILQ_FIRST(&hmp->undo_lru_list);
452 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
453 RB_REMOVE(hammer_und_rb_tree, &hmp->rb_undo_root, node);
454 }
455 node->offset = offset;
456 node->bytes = bytes;
457 TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
458 onode = RB_INSERT(hammer_und_rb_tree, &hmp->rb_undo_root, node);
459 KKASSERT(onode == NULL);
460 return(0);
461}
462
463void
464hammer_clear_undo_history(hammer_mount_t hmp)
465{
466 RB_INIT(&hmp->rb_undo_root);
467 TAILQ_INIT(&hmp->undo_lru_list);
468 hmp->undo_alloc = 0;
469}
470
471/*
06ad81ff
MD
472 * Return how much of the undo FIFO has been used
473 *
474 * The calculation includes undo FIFO space still reserved from a previous
475 * flush (because it will still be run on recovery if a crash occurs and
476 * we can't overwrite it yet).
e8599db1 477 */
1f07f686 478int64_t
06ad81ff 479hammer_undo_used(hammer_transaction_t trans)
1f07f686 480{
06ad81ff
MD
481 hammer_blockmap_t cundomap;
482 hammer_blockmap_t dundomap;
1f07f686 483 int64_t max_bytes;
c9b9e29d 484 int64_t bytes;
1f07f686 485
06ad81ff
MD
486 cundomap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
487 dundomap = &trans->rootvol->ondisk->
488 vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
1f07f686 489
06ad81ff
MD
490 if (dundomap->first_offset <= cundomap->next_offset) {
491 bytes = cundomap->next_offset - dundomap->first_offset;
1f07f686 492 } else {
06ad81ff
MD
493 bytes = cundomap->alloc_offset - dundomap->first_offset +
494 (cundomap->next_offset & HAMMER_OFF_LONG_MASK);
1f07f686 495 }
06ad81ff 496 max_bytes = cundomap->alloc_offset & HAMMER_OFF_SHORT_MASK;
c9b9e29d
MD
497 KKASSERT(bytes <= max_bytes);
498 return(bytes);
499}
500
06ad81ff
MD
501/*
502 * Return how much of the undo FIFO is available for new records.
503 */
c9b9e29d 504int64_t
06ad81ff 505hammer_undo_space(hammer_transaction_t trans)
c9b9e29d
MD
506{
507 hammer_blockmap_t rootmap;
508 int64_t max_bytes;
509
06ad81ff 510 rootmap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
c9b9e29d 511 max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
06ad81ff 512 return(max_bytes - hammer_undo_used(trans));
1f07f686
MD
513}
514
515int64_t
516hammer_undo_max(hammer_mount_t hmp)
517{
518 hammer_blockmap_t rootmap;
519 int64_t max_bytes;
520
521 rootmap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
c9b9e29d 522 max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
1f07f686
MD
523
524 return(max_bytes);
525}
526
710733a6
MD
527/*
528 * Returns 1 if the undo buffer should be reclaimed on release. The
529 * only undo buffer we do NOT want to reclaim is the one at the current
530 * append offset.
531 */
532int
533hammer_undo_reclaim(hammer_io_t io)
534{
535 hammer_blockmap_t undomap;
536 hammer_off_t next_offset;
537
538 undomap = &io->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
539 next_offset = undomap->next_offset & ~HAMMER_BUFMASK64;
540 if (((struct hammer_buffer *)io)->zoneX_offset == next_offset)
541 return(0);
542 return(1);
543}
544
e8599db1
MD
545static int
546hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2)
547{
548 if (node1->offset < node2->offset)
549 return(-1);
550 if (node1->offset > node2->offset)
551 return(1);
552 return(0);
553}
554