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 1.17 2008/06/09 04:19:10 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->beg_obj_id >= reblock->end_obj_id)
  67                 return(EINVAL);
  68         if (reblock->free_level < 0)
  69                 return(EINVAL);
  70
  71         reblock->cur_obj_id = reblock->beg_obj_id;
  72         reblock->cur_localization = reblock->beg_localization;
  73
  74 retry:
  75         error = hammer_init_cursor(trans, &cursor, NULL, NULL);
  76         if (error) {
  77                 hammer_done_cursor(&cursor);
  78                 return(error);
  79         }
  80         cursor.key_beg.localization = reblock->cur_localization;
  81         cursor.key_beg.obj_id = reblock->cur_obj_id;
  82         cursor.key_beg.key = HAMMER_MIN_KEY;
  83         cursor.key_beg.create_tid = 1;
  84         cursor.key_beg.delete_tid = 0;
  85         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
  86         cursor.key_beg.obj_type = 0;
  87
  88         cursor.key_end.localization = reblock->end_localization;
  89         cursor.key_end.obj_id = reblock->end_obj_id;
  90         cursor.key_end.key = HAMMER_MAX_KEY;
  91         cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
  92         cursor.key_end.delete_tid = 0;
  93         cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
  94         cursor.key_end.obj_type = 0;
  95
  96         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
  97         cursor.flags |= HAMMER_CURSOR_BACKEND;
  98
  99         /*
 100          * This flag allows the btree scan code to return internal nodes,
 101          * so we can reblock them in addition to the leafs.  Only specify it
 102          * if we intend to reblock B-Tree nodes.
 103          */
 104         if (reblock->head.flags & HAMMER_IOC_DO_BTREE)
 105                 cursor.flags |= HAMMER_CURSOR_REBLOCKING;
 106
 107         error = hammer_btree_first(&cursor);
 108         while (error == 0) {
 109                 /*
 110                  * Internal or Leaf node
 111                  */
 112                 elm = &cursor.node->ondisk->elms[cursor.index];
 113                 reblock->cur_obj_id = elm->base.obj_id;
 114                 reblock->cur_localization = elm->base.localization;
 115
 116                 /*
 117                  * Yield to more important tasks
 118                  */
 119                 if ((error = hammer_signal_check(trans->hmp)) != 0)
 120                         break;
 121                 if (trans->hmp->sync_lock.wanted) {
 122                         tsleep(trans, 0, "hmrslo", hz / 10);
 123                 }
 124                 if (trans->hmp->locked_dirty_count +
 125                     trans->hmp->io_running_count > hammer_limit_dirtybufs) {
 126                         hammer_flusher_async(trans->hmp);
 127                         tsleep(trans, 0, "hmrslo", hz / 10);
 128                 }
 129
 130                 /*
 131                  * Acquiring the sync_lock prevents the operation from
 132                  * crossing a synchronization boundary.
 133                  *
 134                  * NOTE: cursor.node may have changed on return.
 135                  */
 136                 hammer_sync_lock_sh(trans);
 137                 error = hammer_reblock_helper(reblock, &cursor, elm);
 138                 hammer_sync_unlock(trans);
 139                 if (error == 0) {
 140                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
 141                         error = hammer_btree_iterate(&cursor);
 142                 }
 143         }
 144         if (error == ENOENT)
 145                 error = 0;
 146         hammer_done_cursor(&cursor);
 147         if (error == EDEADLK)
 148                 goto retry;
 149         if (error == EINTR) {
 150                 reblock->head.flags |= HAMMER_IOC_HEAD_INTR;
 151                 error = 0;
 152         }
 153         return(error);
 154 }
 155
 156 /*
 157  * Reblock the B-Tree (leaf) node, record, and/or data if necessary.
 158  *
 159  * XXX We have no visibility into internal B-Tree nodes at the moment,
 160  * only leaf nodes.
 161  */
 162 static int
 163 hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
 164                       hammer_cursor_t cursor, hammer_btree_elm_t elm)
 165 {
 166         hammer_off_t tmp_offset;
 167         int error;
 168         int zone;
 169         int bytes;
 170         int cur;
 171
 172         error = 0;
 173
 174         /*
 175          * Reblock data.  Note that data embedded in a record is reblocked
 176          * by the record reblock code.  Data processing only occurs at leaf
 177          * nodes and for RECORD element types.
 178          */
 179         if (cursor->node->ondisk->type != HAMMER_BTREE_TYPE_LEAF)
 180                 goto skip;
 181         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
 182                 return(0);
 183         tmp_offset = elm->leaf.data_offset;
 184         zone = HAMMER_ZONE_DECODE(tmp_offset);          /* can be 0 */
 185         if ((zone == HAMMER_ZONE_SMALL_DATA_INDEX ||
 186              zone == HAMMER_ZONE_LARGE_DATA_INDEX) &&
 187             error == 0 && (reblock->head.flags & (HAMMER_IOC_DO_DATA | HAMMER_IOC_DO_INODES))) {
 188                 ++reblock->data_count;
 189                 reblock->data_byte_count += elm->leaf.data_len;
 190                 bytes = hammer_blockmap_getfree(cursor->trans->hmp, tmp_offset,
 191                                                 &cur, &error);
 192                 if (hammer_debug_general & 0x4000)
 193                         kprintf("D %6d/%d\n", bytes, reblock->free_level);
 194                 if (error == 0 && cur == 0 && bytes >= reblock->free_level) {
 195                         error = hammer_cursor_upgrade(cursor);
 196                         if (error == 0) {
 197                                 error = hammer_reblock_data(reblock,
 198                                                             cursor, elm);
 199                         }
 200                         if (error == 0) {
 201                                 ++reblock->data_moves;
 202                                 reblock->data_byte_moves += elm->leaf.data_len;
 203                         }
 204                 }
 205         }
 206
 207 skip:
 208         /*
 209          * Reblock a B-Tree internal or leaf node.
 210          */
 211         tmp_offset = cursor->node->node_offset;
 212         zone = HAMMER_ZONE_DECODE(tmp_offset);
 213         if (zone == HAMMER_ZONE_BTREE_INDEX && cursor->index == 0 &&
 214             error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) {
 215                 ++reblock->btree_count;
 216                 bytes = hammer_blockmap_getfree(cursor->trans->hmp, tmp_offset,
 217                                                 &cur, &error);
 218                 if (hammer_debug_general & 0x4000)
 219                         kprintf("B %6d/%d\n", bytes, reblock->free_level);
 220                 if (error == 0 && cur == 0 && bytes >= reblock->free_level) {
 221                         error = hammer_cursor_upgrade(cursor);
 222                         if (error == 0) {
 223                                 if (cursor->parent)
 224                                         elm = &cursor->parent->ondisk->elms[cursor->parent_index];
 225                                 else
 226                                         elm = NULL;
 227                                 switch(cursor->node->ondisk->type) {
 228                                 case HAMMER_BTREE_TYPE_LEAF:
 229                                         error = hammer_reblock_leaf_node(
 230                                                         reblock, cursor, elm);
 231                                         break;
 232                                 case HAMMER_BTREE_TYPE_INTERNAL:
 233                                         error = hammer_reblock_int_node(
 234                                                         reblock, cursor, elm);
 235                                         break;
 236                                 default:
 237                                         panic("Illegal B-Tree node type");
 238                                 }
 239                         }
 240                         if (error == 0) {
 241                                 ++reblock->btree_moves;
 242                         }
 243                 }
 244         }
 245
 246         hammer_cursor_downgrade(cursor);
 247         return(error);
 248 }
 249
 250 /*
 251  * Reblock a record's data.  Both the B-Tree element and record pointers
 252  * to the data must be adjusted.
 253  */
 254 static int
 255 hammer_reblock_data(struct hammer_ioc_reblock *reblock,
 256                     hammer_cursor_t cursor, hammer_btree_elm_t elm)
 257 {
 258         struct hammer_buffer *data_buffer = NULL;
 259         hammer_off_t ndata_offset;
 260         int error;
 261         void *ndata;
 262
 263         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA |
 264                                              HAMMER_CURSOR_GET_LEAF);
 265         if (error)
 266                 return (error);
 267         ndata = hammer_alloc_data(cursor->trans, elm->leaf.data_len,
 268                                   &ndata_offset, &data_buffer, &error);
 269         if (error)
 270                 goto done;
 271
 272         /*
 273          * Move the data
 274          */
 275         hammer_modify_buffer(cursor->trans, data_buffer, NULL, 0);
 276         bcopy(cursor->data, ndata, elm->leaf.data_len);
 277         hammer_modify_buffer_done(data_buffer);
 278
 279         hammer_blockmap_free(cursor->trans,
 280                              elm->leaf.data_offset, elm->leaf.data_len);
 281
 282         hammer_modify_node(cursor->trans, cursor->node,
 283                            &elm->leaf.data_offset, sizeof(hammer_off_t));
 284         elm->leaf.data_offset = ndata_offset;
 285         hammer_modify_node_done(cursor->node);
 286
 287 done:
 288         if (data_buffer)
 289                 hammer_rel_buffer(data_buffer, 0);
 290         return (error);
 291 }
 292
 293 /*
 294  * Reblock a B-Tree leaf node.  The parent must be adjusted to point to
 295  * the new copy of the leaf node.
 296  *
 297  * elm is a pointer to the parent element pointing at cursor.node.
 298  */
 299 static int
 300 hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
 301                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
 302 {
 303         hammer_node_t onode;
 304         hammer_node_t nnode;
 305         int error;
 306
 307         onode = cursor->node;
 308         nnode = hammer_alloc_btree(cursor->trans, &error);
 309
 310         if (nnode == NULL)
 311                 return (error);
 312
 313         /*
 314          * Move the node
 315          */
 316         hammer_lock_ex(&nnode->lock);
 317         hammer_modify_node_noundo(cursor->trans, nnode);
 318         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 319
 320         if (elm) {
 321                 /*
 322                  * We are not the root of the B-Tree
 323                  */
 324                 hammer_modify_node(cursor->trans, cursor->parent,
 325                                    &elm->internal.subtree_offset,
 326                                    sizeof(elm->internal.subtree_offset));
 327                 elm->internal.subtree_offset = nnode->node_offset;
 328                 hammer_modify_node_done(cursor->parent);
 329         } else {
 330                 /*
 331                  * We are the root of the B-Tree
 332                  */
 333                 hammer_volume_t volume;
 334
 335                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 336                 KKASSERT(error == 0);
 337
 338                 hammer_modify_volume_field(cursor->trans, volume,
 339                                            vol0_btree_root);
 340                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 341                 hammer_modify_volume_done(volume);
 342                 hammer_rel_volume(volume, 0);
 343         }
 344
 345         hammer_delete_node(cursor->trans, onode);
 346
 347         if (hammer_debug_general & 0x4000) {
 348                 kprintf("REBLOCK LNODE %016llx -> %016llx\n",
 349                         onode->node_offset, nnode->node_offset);
 350         }
 351         hammer_modify_node_done(nnode);
 352         cursor->node = nnode;
 353
 354         hammer_unlock(&onode->lock);
 355         hammer_rel_node(onode);
 356
 357         return (error);
 358 }
 359
 360 /*
 361  * Reblock a B-Tree internal node.  The parent must be adjusted to point to
 362  * the new copy of the internal node, and the node's children's parent
 363  * pointers must also be adjusted to point to the new copy.
 364  *
 365  * elm is a pointer to the parent element pointing at cursor.node.
 366  */
 367 static int
 368 hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
 369                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
 370 {
 371         hammer_node_locklist_t locklist = NULL;
 372         hammer_node_t onode;
 373         hammer_node_t nnode;
 374         int error;
 375         int i;
 376
 377         error = hammer_btree_lock_children(cursor, &locklist);
 378         if (error)
 379                 goto done;
 380
 381         onode = cursor->node;
 382         nnode = hammer_alloc_btree(cursor->trans, &error);
 383
 384         if (nnode == NULL)
 385                 goto done;
 386
 387         /*
 388          * Move the node.  Adjust the parent's pointer to us first.
 389          */
 390         hammer_lock_ex(&nnode->lock);
 391         hammer_modify_node_noundo(cursor->trans, nnode);
 392         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 393
 394         if (elm) {
 395                 /*
 396                  * We are not the root of the B-Tree
 397                  */
 398                 hammer_modify_node(cursor->trans, cursor->parent,
 399                                    &elm->internal.subtree_offset,
 400                                    sizeof(elm->internal.subtree_offset));
 401                 elm->internal.subtree_offset = nnode->node_offset;
 402                 hammer_modify_node_done(cursor->parent);
 403         } else {
 404                 /*
 405                  * We are the root of the B-Tree
 406                  */
 407                 hammer_volume_t volume;
 408
 409                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 410                 KKASSERT(error == 0);
 411
 412                 hammer_modify_volume_field(cursor->trans, volume,
 413                                            vol0_btree_root);
 414                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 415                 hammer_modify_volume_done(volume);
 416                 hammer_rel_volume(volume, 0);
 417         }
 418
 419         /*
 420          * Now adjust our children's pointers to us.
 421          */
 422         for (i = 0; i < nnode->ondisk->count; ++i) {
 423                 elm = &nnode->ondisk->elms[i];
 424                 error = btree_set_parent(cursor->trans, nnode, elm);
 425                 if (error)
 426                         panic("reblock internal node: fixup problem");
 427         }
 428
 429         /*
 430          * Clean up.
 431          *
 432          * The new node replaces the current node in the cursor.  The cursor
 433          * expects it to be locked so leave it locked.  Discard onode.
 434          */
 435         hammer_delete_node(cursor->trans, onode);
 436
 437         if (hammer_debug_general & 0x4000) {
 438                 kprintf("REBLOCK INODE %016llx -> %016llx\n",
 439                         onode->node_offset, nnode->node_offset);
 440         }
 441         hammer_modify_node_done(nnode);
 442         cursor->node = nnode;
 443
 444         hammer_unlock(&onode->lock);
 445         hammer_rel_node(onode);
 446
 447 done:
 448         hammer_btree_unlock_children(&locklist);
 449         return (error);
 450 }
 451