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         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);
  68
  69         hammer_rel_volume(root_volume, 0);
  70         return(result_offset);
  71 }
  72
  73 /*
  74  * Generate UNDO record(s) for the block of data at the specified zone1
  75  * or zone2 offset.
  76  *
  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.
  80  *
  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.
  85  *
  86  * WARNING!  See also hammer_generate_redo() in hammer_redo.c
  87  */
  88 int
  89 hammer_generate_undo(hammer_transaction_t trans,
  90                      hammer_off_t zone_off, void *base, int len)
  91 {
  92         hammer_mount_t hmp;
  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;
  99         int error;
 100         int bytes;
 101         int n;
 102
 103         hmp = trans->hmp;
 104
 105         /*
 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.
 110          */
 111         if ((hmp->flags & HAMMER_MOUNT_REDO_SYNC) == 0 &&
 112             hmp->version >= HAMMER_VOL_VERSION_FOUR) {
 113                 hammer_generate_redo_sync(trans);
 114         }
 115
 116         /*
 117          * Enter the offset into our undo history.  If there is an existing
 118          * undo we do not have to generate a new one.
 119          */
 120         if (hammer_enter_undo_history(hmp, zone_off, len) == EALREADY)
 121                 return(0);
 122
 123         root_volume = trans->rootvol;
 124         undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
 125
 126         /* no undo recursion */
 127         hammer_modify_volume(NULL, root_volume, NULL, 0);
 128         hammer_lock_ex(&hmp->undo_lock);
 129
 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!");
 133
 134         /*
 135          * Loop until the undo for the entire range has been laid down.
 136          */
 137         while (len) {
 138                 /*
 139                  * Fetch the layout offset in the UNDO FIFO, wrap it as
 140                  * necessary.
 141                  */
 142                 if (undomap->next_offset == undomap->alloc_offset) {
 143                         undomap->next_offset =
 144                                 HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
 145                 }
 146                 next_offset = undomap->next_offset;
 147
 148                 /*
 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.
 151                  */
 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);
 155                 } else {
 156                         undo = hammer_bread(hmp, next_offset, &error, &buffer);
 157                 }
 158                 if (error)
 159                         break;
 160                 hammer_modify_buffer(NULL, buffer, NULL, 0);
 161
 162                 /*
 163                  * Calculate how big a media structure fits up to the next
 164                  * alignment point and how large a data payload we can
 165                  * accomodate.
 166                  *
 167                  * If n calculates to 0 or negative there is no room for
 168                  * anything but a PAD.
 169                  */
 170                 bytes = HAMMER_UNDO_ALIGN -
 171                         ((int)next_offset & HAMMER_UNDO_MASK);
 172                 n = bytes -
 173                     (int)sizeof(struct hammer_fifo_undo) -
 174                     (int)sizeof(struct hammer_fifo_tail);
 175
 176                 /*
 177                  * If available space is insufficient for any payload
 178                  * we have to lay down a PAD.
 179                  *
 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.
 183                  *
 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.
 187                  */
 188                 if (n <= 0) {
 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;
 196                         }
 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;
 204                         continue;
 205                 }
 206
 207                 /*
 208                  * Calculate the actual payload and recalculate the size
 209                  * of the media structure as necessary.
 210                  */
 211                 if (n > len) {
 212                         n = len;
 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);
 217                 }
 218                 if (hammer_debug_general & 0x0080) {
 219                         kprintf("undo %016llx %d %d\n",
 220                                 (long long)next_offset, bytes, n);
 221                 }
 222
 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);
 231
 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;
 236
 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;
 242
 243                 /*
 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.
 251                  *
 252                  * This dummy PAD will be overwritten on the next undo so
 253                  * we do not adjust undomap->next_offset.
 254                  */
 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;
 265                         }
 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 */
 270                 }
 271                 hammer_modify_buffer_done(buffer);
 272
 273                 /*
 274                  * Adjust for loop
 275                  */
 276                 len -= n;
 277                 base = (char *)base + n;
 278                 zone_off += n;
 279         }
 280         hammer_modify_volume_done(root_volume);
 281         hammer_unlock(&hmp->undo_lock);
 282
 283         if (buffer)
 284                 hammer_rel_buffer(buffer, 0);
 285         return(error);
 286 }
 287
 288 /*
 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.
 291  *
 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
 294  * overwritten.
 295  *
 296  * The preformatted UNDO headers use the smallest possible sector size
 297  * (512) to ensure that any missed media writes are caught.
 298  *
 299  * NOTE: Also used by the REDO code.
 300  */
 301 void
 302 hammer_format_undo(void *base, u_int32_t seqno)
 303 {
 304         hammer_fifo_head_t head;
 305         hammer_fifo_tail_t tail;
 306         int i;
 307         int bytes = HAMMER_UNDO_ALIGN;
 308
 309         bzero(base, HAMMER_BUFSIZE);
 310
 311         for (i = 0; i < HAMMER_BUFSIZE; i += bytes) {
 312                 head = (void *)((char *)base + i);
 313                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 314
 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++;
 319                 head->hdr_crc = 0;
 320
 321                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 322                 tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
 323                 tail->tail_size = bytes;
 324
 325                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 326                              crc32(head + 1, bytes - sizeof(*head));
 327         }
 328 }
 329
 330 /*
 331  * HAMMER version 4+ conversion support.
 332  *
 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.
 337  */
 338 int
 339 hammer_upgrade_undo_4(hammer_transaction_t trans)
 340 {
 341         hammer_mount_t hmp;
 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;
 348         u_int32_t seqno;
 349         int error;
 350         int bytes;
 351
 352         hmp = trans->hmp;
 353
 354         root_volume = trans->rootvol;
 355
 356         /* no undo recursion */
 357         hammer_lock_ex(&hmp->undo_lock);
 358         hammer_modify_volume(NULL, root_volume, NULL, 0);
 359
 360         /*
 361          * Adjust the in-core undomap and the on-disk undomap.
 362          */
 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;
 367
 368         undomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
 369         undomap->next_offset = next_offset;
 370         undomap->first_offset = next_offset;
 371
 372         /*
 373          * Loop over the entire UNDO space creating DUMMY entries.  Sequence
 374          * numbers are assigned.
 375          */
 376         seqno = 0;
 377         bytes = HAMMER_UNDO_ALIGN;
 378
 379         while (next_offset != undomap->alloc_offset) {
 380                 head = hammer_bnew(hmp, next_offset, &error, &buffer);
 381                 if (error)
 382                         break;
 383                 hammer_modify_buffer(NULL, buffer, NULL, 0);
 384                 tail = (void *)((char *)head + bytes - sizeof(*tail));
 385
 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;
 390                 head->hdr_crc = 0;
 391
 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;
 396
 397                 head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
 398                              crc32(head + 1, bytes - sizeof(*head));
 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         kprintf("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                         (cundomap->next_offset & HAMMER_OFF_LONG_MASK);
 493         }
 494         max_bytes = cundomap->alloc_offset & HAMMER_OFF_SHORT_MASK;
 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 = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
 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 = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
 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 (((struct hammer_buffer *)io)->zoneX_offset == next_offset)
 539                 return(0);
 540         return(1);
 541 }
 542
 543 static int
 544 hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2)
 545 {
 546         if (node1->offset < node2->offset)
 547                 return(-1);
 548         if (node1->offset > node2->offset)
 549                 return(1);
 550         return(0);
 551 }
 552