sys/vfs/hammer/hammer_reblock.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_reblock.c,v 1.21 2008/06/24 17:38:17 dillon Exp $
  35  */
  36 /*
  37  * HAMMER reblocker - This code frees up fragmented physical space
  38  *
  39  * HAMMER only keeps track of free space on a big-block basis.  A big-block
  40  * containing holes can only be freed by migrating the remaining data in
  41  * that big-block into a new big-block, then freeing the big-block.
  42  *
  43  * This function is called from an ioctl or via the hammer support thread.
  44  */
  45
  46 #include "hammer.h"
  47
  48 static int hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
  49                                  hammer_cursor_t cursor,
  50                                  hammer_btree_elm_t elm);
  51 static int hammer_reblock_data(struct hammer_ioc_reblock *reblock,
  52                                 hammer_cursor_t cursor, hammer_btree_elm_t elm);
  53 static int hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
  54                                 hammer_cursor_t cursor, hammer_btree_elm_t elm);
  55 static int hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
  56                                 hammer_cursor_t cursor, hammer_btree_elm_t elm);
  57
  58 int
  59 hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
  60                struct hammer_ioc_reblock *reblock)
  61 {
  62         struct hammer_cursor cursor;
  63         hammer_btree_elm_t elm;
  64         int error;
  65
  66         if ((reblock->key_beg.localization | reblock->key_end.localization) &
  67             HAMMER_LOCALIZE_PSEUDOFS_MASK) {
  68                 return(EINVAL);
  69         }
  70         if (reblock->key_beg.obj_id >= reblock->key_end.obj_id)
  71                 return(EINVAL);
  72         if (reblock->free_level < 0)
  73                 return(EINVAL);
  74
  75         reblock->key_cur = reblock->key_beg;
  76         reblock->key_cur.localization += ip->obj_localization;
  77
  78 retry:
  79         error = hammer_init_cursor(trans, &cursor, NULL, NULL);
  80         if (error) {
  81                 hammer_done_cursor(&cursor);
  82                 goto failed;
  83         }
  84         cursor.key_beg.localization = reblock->key_cur.localization;
  85         cursor.key_beg.obj_id = reblock->key_cur.obj_id;
  86         cursor.key_beg.key = HAMMER_MIN_KEY;
  87         cursor.key_beg.create_tid = 1;
  88         cursor.key_beg.delete_tid = 0;
  89         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
  90         cursor.key_beg.obj_type = 0;
  91
  92         cursor.key_end.localization = reblock->key_end.localization +
  93                                       ip->obj_localization;
  94         cursor.key_end.obj_id = reblock->key_end.obj_id;
  95         cursor.key_end.key = HAMMER_MAX_KEY;
  96         cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
  97         cursor.key_end.delete_tid = 0;
  98         cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
  99         cursor.key_end.obj_type = 0;
 100
 101         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
 102         cursor.flags |= HAMMER_CURSOR_BACKEND;
 103
 104         /*
 105          * This flag allows the btree scan code to return internal nodes,
 106          * so we can reblock them in addition to the leafs.  Only specify it
 107          * if we intend to reblock B-Tree nodes.
 108          */
 109         if (reblock->head.flags & HAMMER_IOC_DO_BTREE)
 110                 cursor.flags |= HAMMER_CURSOR_REBLOCKING;
 111
 112         error = hammer_btree_first(&cursor);
 113         while (error == 0) {
 114                 /*
 115                  * Internal or Leaf node
 116                  */
 117                 elm = &cursor.node->ondisk->elms[cursor.index];
 118                 reblock->key_cur.obj_id = elm->base.obj_id;
 119                 reblock->key_cur.localization = elm->base.localization;
 120
 121                 /*
 122                  * Yield to more important tasks
 123                  */
 124                 if ((error = hammer_signal_check(trans->hmp)) != 0)
 125                         break;
 126                 if (trans->hmp->sync_lock.wanted) {
 127                         tsleep(trans, 0, "hmrslo", hz / 10);
 128                 }
 129                 if (trans->hmp->locked_dirty_count +
 130                     trans->hmp->io_running_count > hammer_limit_dirtybufs) {
 131                         hammer_flusher_async(trans->hmp);
 132                         tsleep(trans, 0, "hmrslo", hz / 10);
 133                 }
 134
 135                 /*
 136                  * Acquiring the sync_lock prevents the operation from
 137                  * crossing a synchronization boundary.
 138                  *
 139                  * NOTE: cursor.node may have changed on return.
 140                  */
 141                 hammer_sync_lock_sh(trans);
 142                 error = hammer_reblock_helper(reblock, &cursor, elm);
 143                 hammer_sync_unlock(trans);
 144                 if (error == 0) {
 145                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
 146                         error = hammer_btree_iterate(&cursor);
 147                 }
 148         }
 149         if (error == ENOENT)
 150                 error = 0;
 151         hammer_done_cursor(&cursor);
 152         if (error == EDEADLK)
 153                 goto retry;
 154         if (error == EINTR) {
 155                 reblock->head.flags |= HAMMER_IOC_HEAD_INTR;
 156                 error = 0;
 157         }
 158 failed:
 159         reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
 160         return(error);
 161 }
 162
 163 /*
 164  * Reblock the B-Tree (leaf) node, record, and/or data if necessary.
 165  *
 166  * XXX We have no visibility into internal B-Tree nodes at the moment,
 167  * only leaf nodes.
 168  */
 169 static int
 170 hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
 171                       hammer_cursor_t cursor, hammer_btree_elm_t elm)
 172 {
 173         hammer_mount_t hmp;
 174         hammer_off_t tmp_offset;
 175         int error;
 176         int bytes;
 177         int cur;
 178         int iocflags;
 179
 180         error = 0;
 181         hmp = cursor->trans->hmp;
 182
 183         /*
 184          * Reblock data.  Note that data embedded in a record is reblocked
 185          * by the record reblock code.  Data processing only occurs at leaf
 186          * nodes and for RECORD element types.
 187          */
 188         if (cursor->node->ondisk->type != HAMMER_BTREE_TYPE_LEAF)
 189                 goto skip;
 190         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
 191                 return(0);
 192         tmp_offset = elm->leaf.data_offset;
 193         if (tmp_offset == 0)
 194                 goto skip;
 195         if (error)
 196                 goto skip;
 197
 198         /*
 199          * NOTE: Localization restrictions may also have been set-up, we can't
 200          * just set the match flags willy-nilly here.
 201          */
 202         switch(elm->leaf.base.rec_type) {
 203         case HAMMER_RECTYPE_INODE:
 204                 iocflags = HAMMER_IOC_DO_INODES;
 205                 break;
 206         case HAMMER_RECTYPE_EXT:
 207         case HAMMER_RECTYPE_FIX:
 208         case HAMMER_RECTYPE_DIRENTRY:
 209                 iocflags = HAMMER_IOC_DO_DIRS;
 210                 break;
 211         case HAMMER_RECTYPE_DATA:
 212         case HAMMER_RECTYPE_DB:
 213                 iocflags = HAMMER_IOC_DO_DATA;
 214                 break;
 215         default:
 216                 iocflags = 0;
 217                 break;
 218         }
 219         if (reblock->head.flags & iocflags) {
 220                 ++reblock->data_count;
 221                 reblock->data_byte_count += elm->leaf.data_len;
 222                 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
 223                 if (hammer_debug_general & 0x4000)
 224                         kprintf("D %6d/%d\n", bytes, reblock->free_level);
 225                 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
 226                     bytes >= reblock->free_level) {
 227                         hammer_io_direct_uncache(hmp, &elm->leaf);
 228                         error = hammer_cursor_upgrade(cursor);
 229                         if (error == 0) {
 230                                 error = hammer_reblock_data(reblock,
 231                                                             cursor, elm);
 232                         }
 233                         if (error == 0) {
 234                                 ++reblock->data_moves;
 235                                 reblock->data_byte_moves += elm->leaf.data_len;
 236                         }
 237                 }
 238         }
 239
 240 skip:
 241         /*
 242          * Reblock a B-Tree internal or leaf node.
 243          */
 244         tmp_offset = cursor->node->node_offset;
 245         if (cursor->index == 0 &&
 246             error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) {
 247                 ++reblock->btree_count;
 248                 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
 249                 if (hammer_debug_general & 0x4000)
 250                         kprintf("B %6d/%d\n", bytes, reblock->free_level);
 251                 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
 252                     bytes >= reblock->free_level) {
 253                         error = hammer_cursor_upgrade(cursor);
 254                         if (error == 0) {
 255                                 if (cursor->parent)
 256                                         elm = &cursor->parent->ondisk->elms[cursor->parent_index];
 257                                 else
 258                                         elm = NULL;
 259                                 switch(cursor->node->ondisk->type) {
 260                                 case HAMMER_BTREE_TYPE_LEAF:
 261                                         error = hammer_reblock_leaf_node(
 262                                                         reblock, cursor, elm);
 263                                         break;
 264                                 case HAMMER_BTREE_TYPE_INTERNAL:
 265                                         error = hammer_reblock_int_node(
 266                                                         reblock, cursor, elm);
 267                                         break;
 268                                 default:
 269                                         panic("Illegal B-Tree node type");
 270                                 }
 271                         }
 272                         if (error == 0) {
 273                                 ++reblock->btree_moves;
 274                         }
 275                 }
 276         }
 277
 278         hammer_cursor_downgrade(cursor);
 279         return(error);
 280 }
 281
 282 /*
 283  * Reblock a record's data.  Both the B-Tree element and record pointers
 284  * to the data must be adjusted.
 285  */
 286 static int
 287 hammer_reblock_data(struct hammer_ioc_reblock *reblock,
 288                     hammer_cursor_t cursor, hammer_btree_elm_t elm)
 289 {
 290         struct hammer_buffer *data_buffer = NULL;
 291         hammer_off_t ndata_offset;
 292         int error;
 293         void *ndata;
 294
 295         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA |
 296                                              HAMMER_CURSOR_GET_LEAF);
 297         if (error)
 298                 return (error);
 299         ndata = hammer_alloc_data(cursor->trans, elm->leaf.data_len,
 300                                   elm->leaf.base.rec_type,
 301                                   &ndata_offset, &data_buffer, &error);
 302         if (error)
 303                 goto done;
 304
 305         /*
 306          * Move the data
 307          */
 308         hammer_modify_buffer(cursor->trans, data_buffer, NULL, 0);
 309         bcopy(cursor->data, ndata, elm->leaf.data_len);
 310         hammer_modify_buffer_done(data_buffer);
 311
 312         hammer_blockmap_free(cursor->trans,
 313                              elm->leaf.data_offset, elm->leaf.data_len);
 314
 315         hammer_modify_node(cursor->trans, cursor->node,
 316                            &elm->leaf.data_offset, sizeof(hammer_off_t));
 317         elm->leaf.data_offset = ndata_offset;
 318         hammer_modify_node_done(cursor->node);
 319
 320 done:
 321         if (data_buffer)
 322                 hammer_rel_buffer(data_buffer, 0);
 323         return (error);
 324 }
 325
 326 /*
 327  * Reblock a B-Tree leaf node.  The parent must be adjusted to point to
 328  * the new copy of the leaf node.
 329  *
 330  * elm is a pointer to the parent element pointing at cursor.node.
 331  */
 332 static int
 333 hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
 334                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
 335 {
 336         hammer_node_t onode;
 337         hammer_node_t nnode;
 338         int error;
 339
 340         onode = cursor->node;
 341         nnode = hammer_alloc_btree(cursor->trans, &error);
 342
 343         if (nnode == NULL)
 344                 return (error);
 345
 346         /*
 347          * Move the node
 348          */
 349         hammer_lock_ex(&nnode->lock);
 350         hammer_modify_node_noundo(cursor->trans, nnode);
 351         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 352
 353         if (elm) {
 354                 /*
 355                  * We are not the root of the B-Tree
 356                  */
 357                 hammer_modify_node(cursor->trans, cursor->parent,
 358                                    &elm->internal.subtree_offset,
 359                                    sizeof(elm->internal.subtree_offset));
 360                 elm->internal.subtree_offset = nnode->node_offset;
 361                 hammer_modify_node_done(cursor->parent);
 362         } else {
 363                 /*
 364                  * We are the root of the B-Tree
 365                  */
 366                 hammer_volume_t volume;
 367
 368                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 369                 KKASSERT(error == 0);
 370
 371                 hammer_modify_volume_field(cursor->trans, volume,
 372                                            vol0_btree_root);
 373                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 374                 hammer_modify_volume_done(volume);
 375                 hammer_rel_volume(volume, 0);
 376         }
 377
 378         hammer_delete_node(cursor->trans, onode);
 379
 380         if (hammer_debug_general & 0x4000) {
 381                 kprintf("REBLOCK LNODE %016llx -> %016llx\n",
 382                         onode->node_offset, nnode->node_offset);
 383         }
 384         hammer_modify_node_done(nnode);
 385         cursor->node = nnode;
 386
 387         hammer_unlock(&onode->lock);
 388         hammer_rel_node(onode);
 389
 390         return (error);
 391 }
 392
 393 /*
 394  * Reblock a B-Tree internal node.  The parent must be adjusted to point to
 395  * the new copy of the internal node, and the node's children's parent
 396  * pointers must also be adjusted to point to the new copy.
 397  *
 398  * elm is a pointer to the parent element pointing at cursor.node.
 399  */
 400 static int
 401 hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
 402                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
 403 {
 404         hammer_node_locklist_t locklist = NULL;
 405         hammer_node_t onode;
 406         hammer_node_t nnode;
 407         int error;
 408         int i;
 409
 410         error = hammer_btree_lock_children(cursor, &locklist);
 411         if (error)
 412                 goto done;
 413
 414         onode = cursor->node;
 415         nnode = hammer_alloc_btree(cursor->trans, &error);
 416
 417         if (nnode == NULL)
 418                 goto done;
 419
 420         /*
 421          * Move the node.  Adjust the parent's pointer to us first.
 422          */
 423         hammer_lock_ex(&nnode->lock);
 424         hammer_modify_node_noundo(cursor->trans, nnode);
 425         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 426
 427         if (elm) {
 428                 /*
 429                  * We are not the root of the B-Tree
 430                  */
 431                 hammer_modify_node(cursor->trans, cursor->parent,
 432                                    &elm->internal.subtree_offset,
 433                                    sizeof(elm->internal.subtree_offset));
 434                 elm->internal.subtree_offset = nnode->node_offset;
 435                 hammer_modify_node_done(cursor->parent);
 436         } else {
 437                 /*
 438                  * We are the root of the B-Tree
 439                  */
 440                 hammer_volume_t volume;
 441
 442                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 443                 KKASSERT(error == 0);
 444
 445                 hammer_modify_volume_field(cursor->trans, volume,
 446                                            vol0_btree_root);
 447                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 448                 hammer_modify_volume_done(volume);
 449                 hammer_rel_volume(volume, 0);
 450         }
 451
 452         /*
 453          * Now adjust our children's pointers to us.
 454          */
 455         for (i = 0; i < nnode->ondisk->count; ++i) {
 456                 elm = &nnode->ondisk->elms[i];
 457                 error = btree_set_parent(cursor->trans, nnode, elm);
 458                 if (error)
 459                         panic("reblock internal node: fixup problem");
 460         }
 461
 462         /*
 463          * Clean up.
 464          *
 465          * The new node replaces the current node in the cursor.  The cursor
 466          * expects it to be locked so leave it locked.  Discard onode.
 467          */
 468         hammer_delete_node(cursor->trans, onode);
 469
 470         if (hammer_debug_general & 0x4000) {
 471                 kprintf("REBLOCK INODE %016llx -> %016llx\n",
 472                         onode->node_offset, nnode->node_offset);
 473         }
 474         hammer_modify_node_done(nnode);
 475         cursor->node = nnode;
 476
 477         hammer_unlock(&onode->lock);
 478         hammer_rel_node(onode);
 479
 480 done:
 481         hammer_btree_unlock_children(&locklist);
 482         return (error);
 483 }
 484