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  * $DragonFly: src/sys/vfs/hammer/hammer_undo.c,v 1.12 2008/05/04 19:57:42 dillon Exp $
  35  */
  36
  37 /*
  38  * HAMMER undo - undo buffer/FIFO management.
  39  */
  40
  41 #include "hammer.h"
  42
  43 static int hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2);
  44
  45 RB_GENERATE2(hammer_und_rb_tree, hammer_undo, rb_node,
  46              hammer_und_rb_compare, hammer_off_t, offset);
  47
  48 /*
  49  * Convert a zone-3 undo offset into a zone-2 buffer offset.
  50  */
  51 hammer_off_t
  52 hammer_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;
  56         struct hammer_blockmap_layer2 *layer2;
  57         hammer_off_t result_offset;
  58         int i;
  59
  60         KKASSERT((zone3_off & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_UNDO);
  61         root_volume = hammer_get_root_volume(hmp, errorp);
  62         if (*errorp)
  63                 return(0);
  64         undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
  65         KKASSERT(HAMMER_ZONE_DECODE(undomap->alloc_offset) == HAMMER_ZONE_UNDO_INDEX);
  66         KKASSERT (zone3_off < undomap->alloc_offset);
  67
  68         i = (zone3_off & HAMMER_OFF_SHORT_MASK) / HAMMER_LARGEBLOCK_SIZE;
  69         layer2 = &root_volume->ondisk->vol0_undo_array[i];
  70         result_offset = layer2->u.phys_offset +
  71                         (zone3_off & HAMMER_LARGEBLOCK_MASK64);
  72
  73         hammer_rel_volume(root_volume, 0);
  74         return(result_offset);
  75 }
  76
  77 /*
  78  * Generate an UNDO record for the block of data at the specified zone1
  79  * or zone2 offset.
  80  *
  81  * The recovery code will execute UNDOs in reverse order, allowing overlaps.
  82  * All the UNDOs are executed together so if we already laid one down we
  83  * do not have to lay another one down for the same range.
  84  */
  85 int
  86 hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
  87                      hammer_off_t zone_off, void *base, int len)
  88 {
  89         hammer_volume_t root_volume;
  90         hammer_volume_ondisk_t ondisk;
  91         hammer_blockmap_t undomap;
  92         hammer_buffer_t buffer = NULL;
  93         hammer_fifo_undo_t undo;
  94         hammer_fifo_tail_t tail;
  95         hammer_off_t next_offset;
  96         int error;
  97         int bytes;
  98
  99         /*
 100          * Enter the offset into our undo history.  If there is an existing
 101          * undo we do not have to generate a new one.
 102          */
 103         if (hammer_enter_undo_history(trans->hmp, zone_off, len) == EALREADY)
 104                 return(0);
 105
 106         root_volume = trans->rootvol;
 107         ondisk = root_volume->ondisk;
 108         undomap = &trans->hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
 109
 110         /* no undo recursion */
 111         hammer_modify_volume(NULL, root_volume, NULL, 0);
 112
 113 again:
 114         /*
 115          * Allocate space in the FIFO
 116          */
 117         bytes = ((len + HAMMER_HEAD_ALIGN_MASK) & ~HAMMER_HEAD_ALIGN_MASK) +
 118                 sizeof(struct hammer_fifo_undo) +
 119                 sizeof(struct hammer_fifo_tail);
 120         if (hammer_undo_space(trans->hmp) < bytes + HAMMER_BUFSIZE*2)
 121                 panic("hammer: insufficient undo FIFO space!");
 122
 123         next_offset = undomap->next_offset;
 124
 125         /*
 126          * Wrap next_offset
 127          */
 128         if (undomap->next_offset == undomap->alloc_offset) {
 129                 next_offset = HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
 130                 undomap->next_offset = next_offset;
 131                 kprintf("undo zone's next_offset wrapped\n");
 132         }
 133
 134         undo = hammer_bread(trans->hmp, next_offset, &error, &buffer);
 135         hammer_modify_buffer(NULL, buffer, NULL, 0);
 136
 137         /*
 138          * We raced another thread, try again.
 139          */
 140         if (undomap->next_offset != next_offset) {
 141                 hammer_modify_buffer_done(buffer);
 142                 goto again;
 143         }
 144
 145         /*
 146          * The FIFO entry would cross a buffer boundary, PAD to the end
 147          * of the buffer and try again.  Due to our data alignment, the
 148          * worst case (smallest) PAD record is 8 bytes.  PAD records only
 149          * populate the first 8 bytes of hammer_fifo_head and the tail may
 150          * be at the same offset as the head.
 151          */
 152         if ((next_offset ^ (next_offset + bytes)) & ~HAMMER_BUFMASK64) {
 153                 bytes = HAMMER_BUFSIZE - ((int)next_offset & HAMMER_BUFMASK);
 154                 tail = (void *)((char *)undo + bytes - sizeof(*tail));
 155                 if ((void *)undo != (void *)tail) {
 156                         tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 157                         tail->tail_type = HAMMER_HEAD_TYPE_PAD;
 158                         tail->tail_size = bytes;
 159                 }
 160                 undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
 161                 undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
 162                 undo->head.hdr_size = bytes;
 163                 undomap->next_offset += bytes;
 164                 hammer_modify_buffer_done(buffer);
 165                 goto again;
 166         }
 167         if (hammer_debug_general & 0x0080)
 168                 kprintf("undo %016llx %d %d\n", next_offset, bytes, len);
 169
 170         /*
 171          * We're good, create the entry.
 172          */
 173         undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
 174         undo->head.hdr_type = HAMMER_HEAD_TYPE_UNDO;
 175         undo->head.hdr_size = bytes;
 176         undo->head.reserved01 = 0;
 177         undo->head.hdr_crc = 0;
 178         undo->undo_offset = zone_off;
 179         undo->undo_data_bytes = len;
 180         bcopy(base, undo + 1, len);
 181
 182         tail = (void *)((char *)undo + bytes - sizeof(*tail));
 183         tail->tail_signature = HAMMER_TAIL_SIGNATURE;
 184         tail->tail_type = HAMMER_HEAD_TYPE_UNDO;
 185         tail->tail_size = bytes;
 186
 187         undo->head.hdr_crc = crc32(undo, bytes);
 188         undomap->next_offset += bytes;
 189
 190         hammer_modify_buffer_done(buffer);
 191         hammer_modify_volume_done(root_volume);
 192
 193         if (buffer)
 194                 hammer_rel_buffer(buffer, 0);
 195         return(error);
 196 }
 197
 198 /*
 199  * UNDO HISTORY API
 200  *
 201  * It is not necessary to layout an undo record for the same address space
 202  * multiple times.  Maintain a cache of recent undo's.
 203  */
 204
 205 /*
 206  * Enter an undo into the history.  Return EALREADY if the request completely
 207  * covers a previous request.
 208  */
 209 int
 210 hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset, int bytes)
 211 {
 212         hammer_undo_t node;
 213         hammer_undo_t onode;
 214
 215         node = RB_LOOKUP(hammer_und_rb_tree, &hmp->rb_undo_root, offset);
 216         if (node) {
 217                 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
 218                 TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
 219                 if (bytes <= node->bytes)
 220                         return(EALREADY);
 221                 node->bytes = bytes;
 222                 return(0);
 223         }
 224         if (hmp->undo_alloc != HAMMER_MAX_UNDOS) {
 225                 node = &hmp->undos[hmp->undo_alloc++];
 226         } else {
 227                 node = TAILQ_FIRST(&hmp->undo_lru_list);
 228                 TAILQ_REMOVE(&hmp->undo_lru_list, node, lru_entry);
 229                 RB_REMOVE(hammer_und_rb_tree, &hmp->rb_undo_root, node);
 230         }
 231         node->offset = offset;
 232         node->bytes = bytes;
 233         TAILQ_INSERT_TAIL(&hmp->undo_lru_list, node, lru_entry);
 234         onode = RB_INSERT(hammer_und_rb_tree, &hmp->rb_undo_root, node);
 235         KKASSERT(onode == NULL);
 236         return(0);
 237 }
 238
 239 void
 240 hammer_clear_undo_history(hammer_mount_t hmp)
 241 {
 242         RB_INIT(&hmp->rb_undo_root);
 243         TAILQ_INIT(&hmp->undo_lru_list);
 244         hmp->undo_alloc = 0;
 245 }
 246
 247 /*
 248  * Misc helper routines.  Return available space and total space.
 249  */
 250 int64_t
 251 hammer_undo_used(hammer_mount_t hmp)
 252 {
 253         hammer_blockmap_t rootmap;
 254         int64_t max_bytes;
 255         int64_t bytes;
 256
 257         rootmap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
 258
 259         if (rootmap->first_offset <= rootmap->next_offset) {
 260                 bytes = rootmap->next_offset - rootmap->first_offset;
 261         } else {
 262                 bytes = rootmap->alloc_offset - rootmap->first_offset +
 263                         (rootmap->next_offset & HAMMER_OFF_LONG_MASK);
 264         }
 265         max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
 266         KKASSERT(bytes <= max_bytes);
 267         return(bytes);
 268 }
 269
 270 int64_t
 271 hammer_undo_space(hammer_mount_t hmp)
 272 {
 273         hammer_blockmap_t rootmap;
 274         int64_t max_bytes;
 275
 276         rootmap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
 277         max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
 278         return(max_bytes - hammer_undo_used(hmp));
 279 }
 280
 281 int64_t
 282 hammer_undo_max(hammer_mount_t hmp)
 283 {
 284         hammer_blockmap_t rootmap;
 285         int64_t max_bytes;
 286
 287         rootmap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
 288         max_bytes = rootmap->alloc_offset & HAMMER_OFF_SHORT_MASK;
 289
 290         return(max_bytes);
 291 }
 292
 293 static int
 294 hammer_und_rb_compare(hammer_undo_t node1, hammer_undo_t node2)
 295 {
 296         if (node1->offset < node2->offset)
 297                 return(-1);
 298         if (node1->offset > node2->offset)
 299                 return(1);
 300         return(0);
 301 }
 302