sbin/hammer: Add hammer strip command
[dragonfly.git] / sys / vfs / hammer / hammer_undo.c
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
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
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_ZONE_DECODE(undomap->alloc_offset) == HAMMER_ZONE_UNDO_INDEX);
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
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(undo, hmp->undo_seqno ^ 0x40000000);
158                 } else {
159                         undo = hammer_bread(hmp, next_offset, &error, &buffer);
160                 }
161                 if (error)
162                         break;
163                 /* no undo recursion */
164                 hammer_modify_buffer_noundo(NULL, buffer);
165
166                 /*
167                  * Calculate how big a media structure fits up to the next
168                  * alignment point and how large a data payload we can
169                  * accomodate.
170                  *
171                  * If n calculates to 0 or negative there is no room for
172                  * anything but a PAD.
173                  */
174                 bytes = HAMMER_UNDO_ALIGN -
175                         ((int)next_offset & HAMMER_UNDO_MASK);
176                 n = bytes -
177                     (int)sizeof(struct hammer_fifo_undo) -
178                     (int)sizeof(struct hammer_fifo_tail);
179
180                 /*
181                  * If available space is insufficient for any payload
182                  * we have to lay down a PAD.
183                  *
184                  * The minimum PAD is 8 bytes and the head and tail will
185                  * overlap each other in that case.  PADs do not have
186                  * sequence numbers or CRCs.
187                  *
188                  * A PAD may not start on a boundary.  That is, every
189                  * 512-byte block in the UNDO/REDO FIFO must begin with
190                  * a record containing a sequence number.
191                  */
192                 if (n <= 0) {
193                         KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
194                         KKASSERT(((int)next_offset & HAMMER_UNDO_MASK) != 0);
195                         tail = (void *)((char *)undo + bytes - sizeof(*tail));
196                         if ((void *)undo != (void *)tail) {
197                                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
198                                 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
199                                 tail->tail_size = bytes;
200                         }
201                         undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
202                         undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
203                         undo->head.hdr_size = bytes;
204                         /* NO CRC OR SEQ NO */
205                         undomap->next_offset += bytes;
206                         hammer_modify_buffer_done(buffer);
207                         hammer_stats_undo += bytes;
208                         continue;
209                 }
210
211                 /*
212                  * Calculate the actual payload and recalculate the size
213                  * of the media structure as necessary.
214                  */
215                 if (n > len) {
216                         n = len;
217                         bytes = HAMMER_HEAD_DOALIGN(n) +
218                                 (int)sizeof(struct hammer_fifo_undo) +
219                                 (int)sizeof(struct hammer_fifo_tail);
220                 }
221                 if (hammer_debug_general & 0x0080) {
222                         hdkprintf("undo %016jx %d %d\n",
223                                 (intmax_t)next_offset, bytes, n);
224                 }
225
226                 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
227                 undo->head.hdr_type = HAMMER_HEAD_TYPE_UNDO;
228                 undo->head.hdr_size = bytes;
229                 undo->head.hdr_seq = hmp->undo_seqno++;
230                 undo->head.hdr_crc = 0;
231                 undo->undo_offset = zone_off;
232                 undo->undo_data_bytes = n;
233                 bcopy(base, undo + 1, n);
234
235                 tail = (void *)((char *)undo + bytes - sizeof(*tail));
236                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
237                 tail->tail_type = HAMMER_HEAD_TYPE_UNDO;
238                 tail->tail_size = bytes;
239
240                 KKASSERT(bytes >= sizeof(undo->head));
241                 hammer_crc_set_fifo_head(&undo->head, bytes);
242                 undomap->next_offset += bytes;
243                 hammer_stats_undo += bytes;
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);
284
285         if (buffer)
286                 hammer_rel_buffer(buffer, 0);
287         return(error);
288 }
289
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.
297  *
298  * The preformatted UNDO headers use the smallest possible sector size
299  * (512) to ensure that any missed media writes are caught.
300  *
301  * NOTE: Also used by the REDO code.
302  */
303 void
304 hammer_format_undo(void *base, uint32_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                 hammer_crc_set_fifo_head(head, bytes);
328         }
329 }
330
331 /*
332  * HAMMER version 4+ conversion support.
333  *
334  * Convert a HAMMER version < 4 UNDO FIFO area to a 4+ UNDO FIFO area.
335  * The 4+ UNDO FIFO area is backwards compatible.  The conversion is
336  * needed to initialize the sequence space and place headers on the
337  * new 512-byte undo boundary.
338  */
339 int
340 hammer_upgrade_undo_4(hammer_transaction_t trans)
341 {
342         hammer_mount_t hmp;
343         hammer_volume_t root_volume;
344         hammer_blockmap_t undomap;
345         hammer_buffer_t buffer = NULL;
346         hammer_fifo_head_t head;
347         hammer_fifo_tail_t tail;
348         hammer_off_t next_offset;
349         uint32_t seqno;
350         int error;
351         int bytes;
352
353         hmp = trans->hmp;
354
355         root_volume = trans->rootvol;
356
357         /* no undo recursion */
358         hammer_lock_ex(&hmp->undo_lock);
359         hammer_modify_volume_noundo(NULL, root_volume);
360
361         /*
362          * Adjust the in-core undomap and the on-disk undomap.
363          */
364         next_offset = HAMMER_ENCODE_UNDO(0);
365         undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
366         undomap->next_offset = next_offset;
367         undomap->first_offset = next_offset;
368
369         undomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
370         undomap->next_offset = next_offset;
371         undomap->first_offset = next_offset;
372
373         /*
374          * Loop over the entire UNDO space creating DUMMY entries.  Sequence
375          * numbers are assigned.
376          */
377         seqno = 0;
378         bytes = HAMMER_UNDO_ALIGN;
379
380         while (next_offset != undomap->alloc_offset) {
381                 head = hammer_bnew(hmp, next_offset, &error, &buffer);
382                 if (error)
383                         break;
384                 hammer_modify_buffer_noundo(NULL, buffer);
385                 tail = (void *)((char *)head + bytes - sizeof(*tail));
386
387                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
388                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
389                 head->hdr_size = bytes;
390                 head->hdr_seq = seqno;
391                 head->hdr_crc = 0;
392
393                 tail = (void *)((char *)head + bytes - sizeof(*tail));
394                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
395                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
396                 tail->tail_size = bytes;
397
398                 hammer_crc_set_fifo_head(head, bytes);
399                 hammer_modify_buffer_done(buffer);
400
401                 hammer_stats_undo += bytes;
402                 next_offset += HAMMER_UNDO_ALIGN;
403                 ++seqno;
404         }
405
406         /*
407          * The sequence number will be the next sequence number to lay down.
408          */
409         hmp->undo_seqno = seqno;
410         hmkprintf(hmp, "version upgrade seqno start %08x\n", seqno);
411
412         hammer_modify_volume_done(root_volume);
413         hammer_unlock(&hmp->undo_lock);
414
415         if (buffer)
416                 hammer_rel_buffer(buffer, 0);
417         return (error);
418 }
419
420 /*
421  * UNDO HISTORY API
422  *
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.
425  */
426
427 /*
428  * Enter an undo into the history.  Return EALREADY if the request completely
429  * covers a previous request.
430  */
431 int
432 hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset, int bytes)
433 {
434         hammer_undo_t node;
435         hammer_undo_t onode __debugvar;
436
437         node = RB_LOOKUP(hammer_und_rb_tree, &hmp->rb_undo_root, offset);
438         if (node) {
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)
442                         return(EALREADY);
443                 node->bytes = bytes;
444                 return(0);
445         }
446         if (hmp->undo_alloc != HAMMER_MAX_UNDOS) {
447                 node = &hmp->undos[hmp->undo_alloc++];
448         } else {
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);
452         }
453         node->offset = offset;
454         node->bytes = bytes;
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);
458         return(0);
459 }
460
461 void
462 hammer_clear_undo_history(hammer_mount_t hmp)
463 {
464         RB_INIT(&hmp->rb_undo_root);
465         TAILQ_INIT(&hmp->undo_lru_list);
466         hmp->undo_alloc = 0;
467 }
468
469 /*
470  * Return how much of the undo FIFO has been used
471  *
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).
475  */
476 int64_t
477 hammer_undo_used(hammer_transaction_t trans)
478 {
479         hammer_blockmap_t cundomap;
480         hammer_blockmap_t dundomap;
481         int64_t max_bytes __debugvar;
482         int64_t bytes;
483
484         cundomap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
485         dundomap = &trans->rootvol->ondisk->
486                                 vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
487
488         if (dundomap->first_offset <= cundomap->next_offset) {
489                 bytes = cundomap->next_offset - dundomap->first_offset;
490         } else {
491                 bytes = cundomap->alloc_offset - dundomap->first_offset +
492                         HAMMER_OFF_LONG_ENCODE(cundomap->next_offset);
493         }
494         max_bytes = HAMMER_OFF_SHORT_ENCODE(cundomap->alloc_offset);
495         KKASSERT(bytes <= max_bytes);
496         return(bytes);
497 }
498
499 /*
500  * Return how much of the undo FIFO is available for new records.
501  */
502 int64_t
503 hammer_undo_space(hammer_transaction_t trans)
504 {
505         hammer_blockmap_t rootmap;
506         int64_t max_bytes;
507
508         rootmap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
509         max_bytes = HAMMER_OFF_SHORT_ENCODE(rootmap->alloc_offset);
510         return(max_bytes - hammer_undo_used(trans));
511 }
512
513 int64_t
514 hammer_undo_max(hammer_mount_t hmp)
515 {
516         hammer_blockmap_t rootmap;
517         int64_t max_bytes;
518
519         rootmap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
520         max_bytes = HAMMER_OFF_SHORT_ENCODE(rootmap->alloc_offset);
521
522         return(max_bytes);
523 }
524
525 /*
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
528  * append offset.
529  */
530 int
531 hammer_undo_reclaim(hammer_io_t io)
532 {
533         hammer_blockmap_t undomap;
534         hammer_off_t next_offset;
535
536         undomap = &io->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
537         next_offset = undomap->next_offset & ~HAMMER_BUFMASK64;
538         if (HAMMER_ITOB(io)->zoneX_offset == next_offset)
539                 return(0);
540         return(1);
541 }