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 hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2);
  42
  43 RB_GENERATE2(hammer_und_rb_tree, hammer_undo, rb_node,
  44              hammer_und_rb_compare, hammer_off_t, offset);
  45
  46 /*
  47  * Convert a zone-3 undo offset into a zone-2 buffer offset.
  48  */
  49 hammer_off_t
  50 hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t zone3_off, int *errorp)
  51 {
  52         hammer_volume_t root_volume;
  53         hammer_blockmap_t undomap __debugvar;
  54         hammer_off_t result_offset;
  55         int i;
  56
  57         KKASSERT((zone3_off & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_UNDO);
  58         root_volume = hammer_get_root_volume(hmp, errorp);
  59         if (*errorp)
  60                 return(0);
  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);
  64
  65         /*
  66          * undo offsets[i] in zone-2 +
  67          * big-block offset of zone-3 address
  68          * which results zone-2 address
  69          */
  70         i = (zone3_off & HAMMER_OFF_SHORT_MASK) / HAMMER_BIGBLOCK_SIZE;
  71         result_offset = root_volume->ondisk->vol0_undo_array[i] +
  72                         (zone3_off & HAMMER_BIGBLOCK_MASK64);
  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 FIFO space!");
 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 =
 149                                 HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
 150                 }
 151                 next_offset = undomap->next_offset;
 152
 153                 /*
 154                  * This is a tail-chasing FIFO, when we hit the start of a new
 155                  * buffer we don't have to read it in.
 156                  */
 157                 if ((next_offset & HAMMER_BUFMASK) == 0) {
 158                         undo = hammer_bnew(hmp, next_offset, &error, &buffer);
 159                         hammer_format_undo(undo, hmp->undo_seqno ^ 0x40000000);
 160                 } else {
 161                         undo = hammer_bread(hmp, next_offset, &error, &buffer);
 162                 }
 163                 if (error)
 164                         break;
 165                 /* no undo recursion */
 166                 hammer_modify_buffer_noundo(NULL, buffer);
 167
 168                 /*
 169                  * Calculate how big a media structure fits up to the next
 170                  * alignment point and how large a data payload we can
 171                  * accomodate.
 172                  *
 173                  * If n calculates to 0 or negative there is no room for
 174                  * anything but a PAD.
 175                  */
 176                 bytes = HAMMER_UNDO_ALIGN -
 177                         ((int)next_offset & HAMMER_UNDO_MASK);
 178                 n = bytes -
 179                     (int)sizeof(struct hammer_fifo_undo) -
 180                     (int)sizeof(struct hammer_fifo_tail);
 181
 182                 /*
 183                  * If available space is insufficient for any payload
 184                  * we have to lay down a PAD.
 185                  *
 186                  * The minimum PAD is 8 bytes and the head and tail will
 187                  * overlap each other in that case.  PADs do not have
 188                  * sequence numbers or CRCs.
 189                  *
 190                  * A PAD may not start on a boundary.  That is, every
 191                  * 512-byte block in the UNDO/REDO FIFO must begin with
 192                  * a record containing a sequence number.
 193                  */
 194                 if (n <= 0) {
 195                         KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
 196                         KKASSERT(((int)next_offset & HAMMER_UNDO_MASK) != 0);
 197                         tail = (void *)((char *)undo + bytes - sizeof(*tail));
 198                         if ((void *)undo != (void *)tail) {
 199                                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 200                                 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
 201                                 tail->tail_size = bytes;
 202                         }
 203                         undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
 204                         undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
 205                         undo->head.hdr_size = bytes;
 206                         /* NO CRC OR SEQ NO */
 207                         undomap->next_offset += bytes;
 208                         hammer_modify_buffer_done(buffer);
 209                         hammer_stats_undo += bytes;
 210                         continue;
 211                 }
 212
 213                 /*
 214                  * Calculate the actual payload and recalculate the size
 215                  * of the media structure as necessary.
 216                  */
 217                 if (n > len) {
 218                         n = len;
 219                         bytes = ((n + HAMMER_HEAD_ALIGN_MASK) &
 220                                  ~HAMMER_HEAD_ALIGN_MASK) +
 221                                 (int)sizeof(struct hammer_fifo_undo) +
 222                                 (int)sizeof(struct hammer_fifo_tail);
 223                 }
 224                 if (hammer_debug_general & 0x0080) {
 225                         hdkprintf("undo %016llx %d %d\n",
 226                                 (long long)next_offset, bytes, n);
 227                 }
 228
 229                 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
 230                 undo->head.hdr_type = HAMMER_HEAD_TYPE_UNDO;
 231                 undo->head.hdr_size = bytes;
 232                 undo->head.hdr_seq = hmp->undo_seqno++;
 233                 undo->head.hdr_crc = 0;
 234                 undo->undo_offset = zone_off;
 235                 undo->undo_data_bytes = n;
 236                 bcopy(base, undo + 1, n);
 237
 238                 tail = (void *)((char *)undo + bytes - sizeof(*tail));
 239                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 240                 tail->tail_type = HAMMER_HEAD_TYPE_UNDO;
 241                 tail->tail_size = bytes;
 242
 243                 KKASSERT(bytes >= sizeof(undo->head));
 244                 undo->head.hdr_crc = crc32(undo, HAMMER_FIFO_HEAD_CRCOFF) ^
 245                              crc32(&undo->head + 1, bytes - sizeof(undo->head));
 246                 undomap->next_offset += bytes;
 247                 hammer_stats_undo += bytes;
 248
 249                 /*
 250                  * Before we finish off the buffer we have to deal with any
 251                  * junk between the end of the media structure we just laid
 252                  * down and the UNDO alignment boundary.  We do this by laying
 253                  * down a dummy PAD.  Even though we will probably overwrite
 254                  * it almost immediately we have to do this so recovery runs
 255                  * can iterate the UNDO space without having to depend on
 256                  * the indices in the volume header.
 257                  *
 258                  * This dummy PAD will be overwritten on the next undo so
 259                  * we do not adjust undomap->next_offset.
 260                  */
 261                 bytes = HAMMER_UNDO_ALIGN -
 262                         ((int)undomap->next_offset & HAMMER_UNDO_MASK);
 263                 if (bytes != HAMMER_UNDO_ALIGN) {
 264                         KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
 265                         undo = (void *)(tail + 1);
 266                         tail = (void *)((char *)undo + bytes - sizeof(*tail));
 267                         if ((void *)undo != (void *)tail) {
 268                                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 269                                 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
 270                                 tail->tail_size = bytes;
 271                         }
 272                         undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
 273                         undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
 274                         undo->head.hdr_size = bytes;
 275                         /* NO CRC OR SEQ NO */
 276                 }
 277                 hammer_modify_buffer_done(buffer);
 278
 279                 /*
 280                  * Adjust for loop
 281                  */
 282                 len -= n;
 283                 base = (char *)base + n;
 284                 zone_off += n;
 285         }
 286         hammer_modify_volume_done(root_volume);
 287         hammer_unlock(&hmp->undo_lock);
 288
 289         if (buffer)
 290                 hammer_rel_buffer(buffer, 0);
 291         return(error);
 292 }
 293
 294 /*
 295  * Preformat a new UNDO block.  We could read the old one in but we get
 296  * better performance if we just pre-format a new one.
 297  *
 298  * The recovery code always works forwards so the caller just makes sure the
 299  * seqno is not contiguous with prior UNDOs or ancient UNDOs now being
 300  * overwritten.
 301  *
 302  * The preformatted UNDO headers use the smallest possible sector size
 303  * (512) to ensure that any missed media writes are caught.
 304  *
 305  * NOTE: Also used by the REDO code.
 306  */
 307 void
 308 hammer_format_undo(void *base, u_int32_t seqno)
 309 {
 310         hammer_fifo_head_t head;
 311         hammer_fifo_tail_t tail;
 312         int i;
 313         int bytes = HAMMER_UNDO_ALIGN;
 314
 315         bzero(base, HAMMER_BUFSIZE);
 316
 317         for (i = 0; i < HAMMER_BUFSIZE; i += bytes) {
 318                 head = (void *)((char *)base + i);
 319                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 320
 321                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 322                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 323                 head->hdr_size = bytes;
 324                 head->hdr_seq = seqno++;
 325                 head->hdr_crc = 0;
 326
 327                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 328                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 329                 tail->tail_size = bytes;
 330
 331                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 332                              crc32(head + 1, bytes - sizeof(*head));
 333         }
 334 }
 335
 336 /*
 337  * HAMMER version 4+ conversion support.
 338  *
 339  * Convert a HAMMER version < 4 UNDO FIFO area to a 4+ UNDO FIFO area.
 340  * The 4+ UNDO FIFO area is backwards compatible.  The conversion is
 341  * needed to initialize the sequence space and place headers on the
 342  * new 512-byte undo boundary.
 343  */
 344 int
 345 hammer_upgrade_undo_4(hammer_transaction_t trans)
 346 {
 347         hammer_mount_t hmp;
 348         hammer_volume_t root_volume;
 349         hammer_blockmap_t undomap;
 350         hammer_buffer_t buffer = NULL;
 351         hammer_fifo_head_t head;
 352         hammer_fifo_tail_t tail;
 353         hammer_off_t next_offset;
 354         u_int32_t seqno;
 355         int error;
 356         int bytes;
 357
 358         hmp = trans->hmp;
 359
 360         root_volume = trans->rootvol;
 361
 362         /* no undo recursion */
 363         hammer_lock_ex(&hmp->undo_lock);
 364         hammer_modify_volume_noundo(NULL, root_volume);
 365
 366         /*
 367          * Adjust the in-core undomap and the on-disk undomap.
 368          */
 369         next_offset = HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
 370         undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
 371         undomap->next_offset = next_offset;
 372         undomap->first_offset = next_offset;
 373
 374         undomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 375         undomap->next_offset = next_offset;
 376         undomap->first_offset = next_offset;
 377
 378         /*
 379          * Loop over the entire UNDO space creating DUMMY entries.  Sequence
 380          * numbers are assigned.
 381          */
 382         seqno = 0;
 383         bytes = HAMMER_UNDO_ALIGN;
 384
 385         while (next_offset != undomap->alloc_offset) {
 386                 head = hammer_bnew(hmp, next_offset, &error, &buffer);
 387                 if (error)
 388                         break;
 389                 hammer_modify_buffer_noundo(NULL, buffer);
 390                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 391
 392                 head->hdr_signature = HAMMER_HEAD_SIGNATURE;
 393                 head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
 394                 head->hdr_size = bytes;
 395                 head->hdr_seq = seqno;
 396                 head->hdr_crc = 0;
 397
 398                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 399                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 400                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 401                 tail->tail_size = bytes;
 402
 403                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 404                              crc32(head + 1, bytes - sizeof(*head));
 405                 hammer_modify_buffer_done(buffer);
 406
 407                 hammer_stats_undo += bytes;
 408                 next_offset += HAMMER_UNDO_ALIGN;
 409                 ++seqno;
 410         }
 411
 412         /*
 413          * The sequence number will be the next sequence number to lay down.
 414          */
 415         hmp->undo_seqno = seqno;
 416         hmkprintf(hmp, "version upgrade seqno start %08x\n", seqno);
 417
 418         hammer_modify_volume_done(root_volume);
 419         hammer_unlock(&hmp->undo_lock);
 420
 421         if (buffer)
 422                 hammer_rel_buffer(buffer, 0);
 423         return (error);
 424 }
 425
 426 /*
 427  * UNDO HISTORY API
 428  *
 429  * It is not necessary to layout an undo record for the same address space
 430  * multiple times.  Maintain a cache of recent undo's.
 431  */
 432
 433 /*
 434  * Enter an undo into the history.  Return EALREADY if the request completely
 435  * covers a previous request.
 436  */
 437 int
 438 hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset, int bytes)
 439 {
 440         hammer_undo_t node;
 441         hammer_undo_t onode __debugvar;
 442
 443         node = RB_LOOKUP(hammer_und_rb_tree, &hmp->rb_undo_root, offset);
 444         if (node) {
 445                 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
 446                 TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
 447                 if (bytes <= node->bytes)
 448                         return(EALREADY);
 449                 node->bytes = bytes;
 450                 return(0);
 451         }
 452         if (hmp->undo_alloc != HAMMER_MAX_UNDOS) {
 453                 node = &hmp->undos[hmp->undo_alloc++];
 454         } else {
 455                 node = TAILQ_FIRST(&hmp->undo_lru_list);
 456                 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
 457                 RB_REMOVE(hammer_und_rb_tree, &hmp->rb_undo_root, node);
 458         }
 459         node->offset = offset;
 460         node->bytes = bytes;
 461         TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
 462         onode = RB_INSERT(hammer_und_rb_tree, &hmp->rb_undo_root, node);
 463         KKASSERT(onode == NULL);
 464         return(0);
 465 }
 466
 467 void
 468 hammer_clear_undo_history(hammer_mount_t hmp)
 469 {
 470         RB_INIT(&hmp->rb_undo_root);
 471         TAILQ_INIT(&hmp->undo_lru_list);
 472         hmp->undo_alloc = 0;
 473 }
 474
 475 /*
 476  * Return how much of the undo FIFO has been used
 477  *
 478  * The calculation includes undo FIFO space still reserved from a previous
 479  * flush (because it will still be run on recovery if a crash occurs and
 480  * we can't overwrite it yet).
 481  */
 482 int64_t
 483 hammer_undo_used(hammer_transaction_t trans)
 484 {
 485         hammer_blockmap_t cundomap;
 486         hammer_blockmap_t dundomap;
 487         int64_t max_bytes __debugvar;
 488         int64_t bytes;
 489
 490         cundomap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
 491         dundomap = &trans->rootvol->ondisk->
 492                                 vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 493
 494         if (dundomap->first_offset <= cundomap->next_offset) {
 495                 bytes = cundomap->next_offset - dundomap->first_offset;
 496         } else {
 497                 bytes = cundomap->alloc_offset - dundomap->first_offset +
 498                         (cundomap->next_offset & HAMMER_OFF_LONG_MASK);
 499         }
 500         max_bytes = cundomap->alloc_offset & HAMMER_OFF_SHORT_MASK;
 501         KKASSERT(bytes <= max_bytes);
 502         return(bytes);
 503 }
 504
 505 /*
 506  * Return how much of the undo FIFO is available for new records.
 507  */
 508 int64_t
 509 hammer_undo_space(hammer_transaction_t trans)
 510 {
 511         hammer_blockmap_t rootmap;
 512         int64_t max_bytes;
 513
 514         rootmap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
 515         max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
 516         return(max_bytes - hammer_undo_used(trans));
 517 }
 518
 519 int64_t
 520 hammer_undo_max(hammer_mount_t hmp)
 521 {
 522         hammer_blockmap_t rootmap;
 523         int64_t max_bytes;
 524
 525         rootmap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
 526         max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
 527
 528         return(max_bytes);
 529 }
 530
 531 /*
 532  * Returns 1 if the undo buffer should be reclaimed on release.  The
 533  * only undo buffer we do NOT want to reclaim is the one at the current
 534  * append offset.
 535  */
 536 int
 537 hammer_undo_reclaim(hammer_io_t io)
 538 {
 539         hammer_blockmap_t undomap;
 540         hammer_off_t next_offset;
 541
 542         undomap = &io->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
 543         next_offset = undomap->next_offset & ~HAMMER_BUFMASK64;
 544         if (((struct hammer_buffer *)io)->zoneX_offset == next_offset)
 545                 return(0);
 546         return(1);
 547 }
 548
 549 static int
 550 hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2)
 551 {
 552         if (node1->offset < node2->offset)
 553                 return(-1);
 554         if (node1->offset > node2->offset)
 555                 return(1);
 556         return(0);
 557 }
 558