sys/vfs/hammer/hammer_redo.c

   1 /*
   2  * Copyright (c) 2010 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 redo - REDO record support for the UNDO/REDO FIFO.
  37  *
  38  * See also hammer_undo.c
  39  */
  40
  41 #include "hammer.h"
  42
  43 RB_GENERATE2(hammer_redo_rb_tree, hammer_inode, rb_redonode,
  44              hammer_redo_rb_compare, hammer_off_t, redo_fifo_start);
  45
  46 /*
  47  * HAMMER version 4+ REDO support.
  48  *
  49  * REDO records are used to improve fsync() performance.  Instead of having
  50  * to go through a complete double-flush cycle involving at least two disk
  51  * synchronizations the fsync need only flush UNDO/REDO FIFO buffers through
  52  * the related REDO records, which is a single synchronization requiring
  53  * no track seeking.  If a recovery becomes necessary the recovery code
  54  * will generate logical data writes based on the REDO records encountered.
  55  * That is, the recovery code will UNDO any partial meta-data/data writes
  56  * at the raw disk block level and then REDO the data writes at the logical
  57  * level.
  58  */
  59 int
  60 hammer_generate_redo(hammer_transaction_t trans, hammer_inode_t ip,
  61                      hammer_off_t file_off, uint32_t flags,
  62                      void *base, int len)
  63 {
  64         hammer_mount_t hmp;
  65         hammer_volume_t root_volume;
  66         hammer_blockmap_t undomap;
  67         hammer_buffer_t buffer = NULL;
  68         hammer_fifo_redo_t redo;
  69         hammer_fifo_tail_t tail;
  70         hammer_off_t next_offset;
  71         int error;
  72         int bytes;
  73         int n;
  74
  75         /*
  76          * Setup
  77          */
  78         hmp = trans->hmp;
  79
  80         root_volume = trans->rootvol;
  81         undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
  82
  83         /*
  84          * No undo recursion when modifying the root volume
  85          */
  86         hammer_modify_volume_noundo(NULL, root_volume);
  87         hammer_lock_ex(&hmp->undo_lock);
  88
  89         /* undo had better not roll over (loose test) */
  90         if (hammer_undo_space(trans) < len + HAMMER_BUFSIZE*3)
  91                 hpanic("insufficient UNDO/REDO FIFO space for redo!");
  92
  93         /*
  94          * Loop until the undo for the entire range has been laid down.
  95          * Loop at least once (len might be 0 as a degenerate case).
  96          */
  97         for (;;) {
  98                 /*
  99                  * Fetch the layout offset in the UNDO FIFO, wrap it as
 100                  * necessary.
 101                  */
 102                 if (undomap->next_offset == undomap->alloc_offset) {
 103                         undomap->next_offset =
 104                                 HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
 105                 }
 106                 next_offset = undomap->next_offset;
 107
 108                 /*
 109                  * This is a tail-chasing FIFO, when we hit the start of a new
 110                  * buffer we don't have to read it in.
 111                  */
 112                 if ((next_offset & HAMMER_BUFMASK) == 0) {
 113                         redo = hammer_bnew(hmp, next_offset, &error, &buffer);
 114                         hammer_format_undo(redo, hmp->undo_seqno ^ 0x40000000);
 115                 } else {
 116                         redo = hammer_bread(hmp, next_offset, &error, &buffer);
 117                 }
 118                 if (error)
 119                         break;
 120                 hammer_modify_buffer_noundo(NULL, buffer);
 121
 122                 /*
 123                  * Calculate how big a media structure fits up to the next
 124                  * alignment point and how large a data payload we can
 125                  * accomodate.
 126                  *
 127                  * If n calculates to 0 or negative there is no room for
 128                  * anything but a PAD.
 129                  */
 130                 bytes = HAMMER_UNDO_ALIGN -
 131                         ((int)next_offset & HAMMER_UNDO_MASK);
 132                 n = bytes -
 133                     (int)sizeof(struct hammer_fifo_redo) -
 134                     (int)sizeof(struct hammer_fifo_tail);
 135
 136                 /*
 137                  * If available space is insufficient for any payload
 138                  * we have to lay down a PAD.
 139                  *
 140                  * The minimum PAD is 8 bytes and the head and tail will
 141                  * overlap each other in that case.  PADs do not have
 142                  * sequence numbers or CRCs.
 143                  *
 144                  * A PAD may not start on a boundary.  That is, every
 145                  * 512-byte block in the UNDO/REDO FIFO must begin with
 146                  * a record containing a sequence number.
 147                  */
 148                 if (n <= 0) {
 149                         KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
 150                         KKASSERT(((int)next_offset & HAMMER_UNDO_MASK) != 0);
 151                         tail = (void *)((char *)redo + bytes - sizeof(*tail));
 152                         if ((void *)redo != (void *)tail) {
 153                                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 154                                 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
 155                                 tail->tail_size = bytes;
 156                         }
 157                         redo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
 158                         redo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
 159                         redo->head.hdr_size = bytes;
 160                         /* NO CRC OR SEQ NO */
 161                         undomap->next_offset += bytes;
 162                         hammer_modify_buffer_done(buffer);
 163                         hammer_stats_redo += bytes;
 164                         continue;
 165                 }
 166
 167                 /*
 168                  * When generating an inode-related REDO record we track
 169                  * the point in the UNDO/REDO FIFO containing the inode's
 170                  * earliest REDO record.  See hammer_generate_redo_sync().
 171                  *
 172                  * redo_fifo_next is cleared when an inode is staged to
 173                  * the backend and then used to determine how to reassign
 174                  * redo_fifo_start after the inode flush completes.
 175                  */
 176                 if (ip) {
 177                         redo->redo_objid = ip->obj_id;
 178                         redo->redo_localization = ip->obj_localization;
 179                         if ((ip->flags & HAMMER_INODE_RDIRTY) == 0) {
 180                                 ip->redo_fifo_start = next_offset;
 181                                 if (RB_INSERT(hammer_redo_rb_tree,
 182                                               &hmp->rb_redo_root, ip)) {
 183                                         hpanic("cannot insert inode %p on "
 184                                               "redo FIFO", ip);
 185                                 }
 186                                 ip->flags |= HAMMER_INODE_RDIRTY;
 187                         }
 188                         if (ip->redo_fifo_next == 0)
 189                                 ip->redo_fifo_next = next_offset;
 190                 } else {
 191                         redo->redo_objid = 0;
 192                         redo->redo_localization = 0;
 193                 }
 194
 195                 /*
 196                  * Calculate the actual payload and recalculate the size
 197                  * of the media structure as necessary.  If no data buffer
 198                  * is supplied there is no payload.
 199                  */
 200                 if (base == NULL) {
 201                         n = 0;
 202                 } else if (n > len) {
 203                         n = len;
 204                 }
 205                 bytes = ((n + HAMMER_HEAD_ALIGN_MASK) &
 206                          ~HAMMER_HEAD_ALIGN_MASK) +
 207                         (int)sizeof(struct hammer_fifo_redo) +
 208                         (int)sizeof(struct hammer_fifo_tail);
 209                 if (hammer_debug_general & 0x0080) {
 210                         hdkprintf("redo %016jx %d %d\n",
 211                                 (intmax_t)next_offset, bytes, n);
 212                 }
 213
 214                 redo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
 215                 redo->head.hdr_type = HAMMER_HEAD_TYPE_REDO;
 216                 redo->head.hdr_size = bytes;
 217                 redo->head.hdr_seq = hmp->undo_seqno++;
 218                 redo->head.hdr_crc = 0;
 219                 redo->redo_mtime = trans->time;
 220                 redo->redo_offset = file_off;
 221                 redo->redo_flags = flags;
 222
 223                 /*
 224                  * Incremental payload.  If no payload we throw the entire
 225                  * len into redo_data_bytes and will not loop.
 226                  */
 227                 if (base) {
 228                         redo->redo_data_bytes = n;
 229                         bcopy(base, redo + 1, n);
 230                         len -= n;
 231                         base = (char *)base + n;
 232                         file_off += n;
 233                 } else {
 234                         redo->redo_data_bytes = len;
 235                         file_off += len;
 236                         len = 0;
 237                 }
 238
 239                 tail = (void *)((char *)redo + bytes - sizeof(*tail));
 240                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 241                 tail->tail_type = HAMMER_HEAD_TYPE_REDO;
 242                 tail->tail_size = bytes;
 243
 244                 KKASSERT(bytes >= sizeof(redo->head));
 245                 hammer_crc_set_fifo_head(&redo->head, bytes);
 246                 undomap->next_offset += bytes;
 247                 hammer_stats_redo += 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                         redo = (void *)(tail + 1);
 266                         tail = (void *)((char *)redo + bytes - sizeof(*tail));
 267                         if ((void *)redo != (void *)tail) {
 268                                 tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 269                                 tail->tail_type = HAMMER_HEAD_TYPE_PAD;
 270                                 tail->tail_size = bytes;
 271                         }
 272                         redo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
 273                         redo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
 274                         redo->head.hdr_size = bytes;
 275                         /* NO CRC OR SEQ NO */
 276                 }
 277                 hammer_modify_buffer_done(buffer);
 278                 if (len == 0)
 279                         break;
 280         }
 281         hammer_modify_volume_done(root_volume);
 282         hammer_unlock(&hmp->undo_lock);
 283
 284         if (buffer)
 285                 hammer_rel_buffer(buffer, 0);
 286
 287         /*
 288          * Make sure the nominal undo span contains at least one REDO_SYNC,
 289          * otherwise the REDO recovery will not be triggered.
 290          */
 291         if ((hmp->flags & HAMMER_MOUNT_REDO_SYNC) == 0 &&
 292             flags != HAMMER_REDO_SYNC) {
 293                 hammer_generate_redo_sync(trans);
 294         }
 295
 296         return(error);
 297 }
 298
 299 /*
 300  * Generate a REDO SYNC record.  At least one such record must be generated
 301  * in the nominal recovery span for the recovery code to be able to run
 302  * REDOs outside of the span.
 303  *
 304  * The SYNC record contains the aggregate earliest UNDO/REDO FIFO offset
 305  * for all inodes with active REDOs.  This changes dynamically as inodes
 306  * get flushed.
 307  *
 308  * During recovery stage2 any new flush cycles must specify the original
 309  * redo sync offset.  That way a crash will re-run the REDOs, at least
 310  * up to the point where the UNDO FIFO does not overwrite the area.
 311  */
 312 void
 313 hammer_generate_redo_sync(hammer_transaction_t trans)
 314 {
 315         hammer_mount_t hmp = trans->hmp;
 316         hammer_inode_t ip;
 317         hammer_off_t redo_fifo_start;
 318
 319         if (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) {
 320                 ip = NULL;
 321                 redo_fifo_start = hmp->recover_stage2_offset;
 322         } else {
 323                 ip = RB_FIRST(hammer_redo_rb_tree, &hmp->rb_redo_root);
 324                 if (ip)
 325                         redo_fifo_start = ip->redo_fifo_start;
 326                 else
 327                         redo_fifo_start = 0;
 328         }
 329         if (redo_fifo_start) {
 330                 if (hammer_debug_io & 0x0004) {
 331                         hdkprintf("SYNC IP %p %016jx\n",
 332                                 ip, (intmax_t)redo_fifo_start);
 333                 }
 334                 hammer_generate_redo(trans, NULL, redo_fifo_start,
 335                                      HAMMER_REDO_SYNC, NULL, 0);
 336                 trans->hmp->flags |= HAMMER_MOUNT_REDO_SYNC;
 337         }
 338 }
 339
 340 /*
 341  * This is called when an inode is queued to the backend.
 342  */
 343 void
 344 hammer_redo_fifo_start_flush(hammer_inode_t ip)
 345 {
 346         ip->redo_fifo_next = 0;
 347 }
 348
 349 /*
 350  * This is called when an inode backend flush is finished.  We have to make
 351  * sure that RDIRTY is not set unless dirty bufs are present.  Dirty bufs
 352  * can get destroyed through operations such as truncations and leave
 353  * us with a stale redo_fifo_next.
 354  */
 355 void
 356 hammer_redo_fifo_end_flush(hammer_inode_t ip)
 357 {
 358         hammer_mount_t hmp = ip->hmp;
 359
 360         if (ip->flags & HAMMER_INODE_RDIRTY) {
 361                 RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
 362                 ip->flags &= ~HAMMER_INODE_RDIRTY;
 363         }
 364         if ((ip->flags & HAMMER_INODE_BUFS) == 0)
 365                 ip->redo_fifo_next = 0;
 366         if (ip->redo_fifo_next) {
 367                 ip->redo_fifo_start = ip->redo_fifo_next;
 368                 if (RB_INSERT(hammer_redo_rb_tree, &hmp->rb_redo_root, ip)) {
 369                         hpanic("cannot reinsert inode %p on redo FIFO", ip);
 370                 }
 371                 ip->flags |= HAMMER_INODE_RDIRTY;
 372         }
 373 }