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 1.31 2008/07/13 09:32:48 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 checkspace_count;
  65         int error;
  66         int seq;
  67
  68         if ((reblock->key_beg.localization | reblock->key_end.localization) &
  69             HAMMER_LOCALIZE_PSEUDOFS_MASK) {
  70                 return(EINVAL);
  71         }
  72         if (reblock->key_beg.obj_id >= reblock->key_end.obj_id)
  73                 return(EINVAL);
  74         if (reblock->free_level < 0)
  75                 return(EINVAL);
  76
  77         reblock->key_cur = reblock->key_beg;
  78         reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
  79         reblock->key_cur.localization += ip->obj_localization;
  80
  81         checkspace_count = 0;
  82         seq = trans->hmp->flusher.act;
  83 retry:
  84         error = hammer_init_cursor(trans, &cursor, NULL, NULL);
  85         if (error) {
  86                 hammer_done_cursor(&cursor);
  87                 goto failed;
  88         }
  89         cursor.key_beg.localization = reblock->key_cur.localization;
  90         cursor.key_beg.obj_id = reblock->key_cur.obj_id;
  91         cursor.key_beg.key = HAMMER_MIN_KEY;
  92         cursor.key_beg.create_tid = 1;
  93         cursor.key_beg.delete_tid = 0;
  94         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
  95         cursor.key_beg.obj_type = 0;
  96
  97         cursor.key_end.localization = (reblock->key_end.localization &
  98                                         HAMMER_LOCALIZE_MASK) +
  99                                       ip->obj_localization;
 100         cursor.key_end.obj_id = reblock->key_end.obj_id;
 101         cursor.key_end.key = HAMMER_MAX_KEY;
 102         cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
 103         cursor.key_end.delete_tid = 0;
 104         cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
 105         cursor.key_end.obj_type = 0;
 106
 107         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
 108         cursor.flags |= HAMMER_CURSOR_BACKEND;
 109
 110         /*
 111          * This flag allows the btree scan code to return internal nodes,
 112          * so we can reblock them in addition to the leafs.  Only specify it
 113          * if we intend to reblock B-Tree nodes.
 114          */
 115         if (reblock->head.flags & HAMMER_IOC_DO_BTREE)
 116                 cursor.flags |= HAMMER_CURSOR_REBLOCKING;
 117
 118         error = hammer_btree_first(&cursor);
 119         while (error == 0) {
 120                 /*
 121                  * Internal or Leaf node
 122                  */
 123                 elm = &cursor.node->ondisk->elms[cursor.index];
 124                 reblock->key_cur.obj_id = elm->base.obj_id;
 125                 reblock->key_cur.localization = elm->base.localization;
 126
 127                 /*
 128                  * Yield to more important tasks
 129                  */
 130                 if ((error = hammer_signal_check(trans->hmp)) != 0)
 131                         break;
 132
 133                 /*
 134                  * If there is insufficient free space it may be due to
 135                  * reserved bigblocks, which flushing might fix.
 136                  */
 137                 if (hammer_checkspace(trans->hmp, HAMMER_CHKSPC_REBLOCK)) {
 138                         if (++checkspace_count == 10) {
 139                                 error = ENOSPC;
 140                                 break;
 141                         }
 142                         hammer_unlock_cursor(&cursor, 0);
 143                         hammer_flusher_wait(trans->hmp, seq);
 144                         hammer_lock_cursor(&cursor, 0);
 145                         seq = hammer_flusher_async(trans->hmp, NULL);
 146                 }
 147
 148                 /*
 149                  * We allocate data buffers, which atm we don't track
 150                  * dirty levels for because we allow the kernel to write
 151                  * them.  But if we allocate too many we can still deadlock
 152                  * the buffer cache.
 153                  */
 154                 if (bd_heatup()) {
 155                         hammer_unlock_cursor(&cursor, 0);
 156                         bwillwrite(HAMMER_BUFSIZE);
 157                         hammer_lock_cursor(&cursor, 0);
 158                 }
 159
 160                 /*
 161                  * Acquiring the sync_lock prevents the operation from
 162                  * crossing a synchronization boundary.
 163                  *
 164                  * NOTE: cursor.node may have changed on return.
 165                  */
 166                 hammer_sync_lock_sh(trans);
 167                 error = hammer_reblock_helper(reblock, &cursor, elm);
 168                 hammer_sync_unlock(trans);
 169
 170                 while (hammer_flusher_meta_halflimit(trans->hmp) ||
 171                        hammer_flusher_undo_exhausted(trans, 2)) {
 172                         hammer_unlock_cursor(&cursor, 0);
 173                         hammer_flusher_wait(trans->hmp, seq);
 174                         hammer_lock_cursor(&cursor, 0);
 175                         seq = hammer_flusher_async_one(trans->hmp);
 176                 }
 177                 if (error == 0) {
 178                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
 179                         error = hammer_btree_iterate(&cursor);
 180                 }
 181
 182         }
 183         if (error == ENOENT)
 184                 error = 0;
 185         hammer_done_cursor(&cursor);
 186         if (error == EWOULDBLOCK) {
 187                 hammer_flusher_sync(trans->hmp);
 188                 goto retry;
 189         }
 190         if (error == EDEADLK)
 191                 goto retry;
 192         if (error == EINTR) {
 193                 reblock->head.flags |= HAMMER_IOC_HEAD_INTR;
 194                 error = 0;
 195         }
 196 failed:
 197         reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
 198         return(error);
 199 }
 200
 201 /*
 202  * Reblock the B-Tree (leaf) node, record, and/or data if necessary.
 203  *
 204  * XXX We have no visibility into internal B-Tree nodes at the moment,
 205  * only leaf nodes.
 206  */
 207 static int
 208 hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
 209                       hammer_cursor_t cursor, hammer_btree_elm_t elm)
 210 {
 211         hammer_mount_t hmp;
 212         hammer_off_t tmp_offset;
 213         int error;
 214         int bytes;
 215         int cur;
 216         int iocflags;
 217
 218         error = 0;
 219         hmp = cursor->trans->hmp;
 220
 221         /*
 222          * Reblock data.  Note that data embedded in a record is reblocked
 223          * by the record reblock code.  Data processing only occurs at leaf
 224          * nodes and for RECORD element types.
 225          */
 226         if (cursor->node->ondisk->type != HAMMER_BTREE_TYPE_LEAF)
 227                 goto skip;
 228         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
 229                 return(0);
 230         tmp_offset = elm->leaf.data_offset;
 231         if (tmp_offset == 0)
 232                 goto skip;
 233         if (error)
 234                 goto skip;
 235
 236         /*
 237          * NOTE: Localization restrictions may also have been set-up, we can't
 238          * just set the match flags willy-nilly here.
 239          */
 240         switch(elm->leaf.base.rec_type) {
 241         case HAMMER_RECTYPE_INODE:
 242                 iocflags = HAMMER_IOC_DO_INODES;
 243                 break;
 244         case HAMMER_RECTYPE_EXT:
 245         case HAMMER_RECTYPE_FIX:
 246         case HAMMER_RECTYPE_PFS:
 247         case HAMMER_RECTYPE_DIRENTRY:
 248                 iocflags = HAMMER_IOC_DO_DIRS;
 249                 break;
 250         case HAMMER_RECTYPE_DATA:
 251         case HAMMER_RECTYPE_DB:
 252                 iocflags = HAMMER_IOC_DO_DATA;
 253                 break;
 254         default:
 255                 iocflags = 0;
 256                 break;
 257         }
 258         if (reblock->head.flags & iocflags) {
 259                 ++reblock->data_count;
 260                 reblock->data_byte_count += elm->leaf.data_len;
 261                 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
 262                 if (hammer_debug_general & 0x4000)
 263                         kprintf("D %6d/%d\n", bytes, reblock->free_level);
 264                 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
 265                     bytes >= reblock->free_level) {
 266                         hammer_io_direct_uncache(hmp, &elm->leaf);
 267                         error = hammer_cursor_upgrade(cursor);
 268                         if (error == 0) {
 269                                 error = hammer_reblock_data(reblock,
 270                                                             cursor, elm);
 271                         }
 272                         if (error == 0) {
 273                                 ++reblock->data_moves;
 274                                 reblock->data_byte_moves += elm->leaf.data_len;
 275                         }
 276                 }
 277         }
 278
 279 skip:
 280         /*
 281          * Reblock a B-Tree internal or leaf node.
 282          */
 283         tmp_offset = cursor->node->node_offset;
 284         if (cursor->index == 0 &&
 285             error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) {
 286                 ++reblock->btree_count;
 287                 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
 288                 if (hammer_debug_general & 0x4000)
 289                         kprintf("B %6d/%d\n", bytes, reblock->free_level);
 290                 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
 291                     bytes >= reblock->free_level) {
 292                         error = hammer_cursor_upgrade(cursor);
 293                         if (error == 0) {
 294                                 if (cursor->parent)
 295                                         elm = &cursor->parent->ondisk->elms[cursor->parent_index];
 296                                 else
 297                                         elm = NULL;
 298                                 switch(cursor->node->ondisk->type) {
 299                                 case HAMMER_BTREE_TYPE_LEAF:
 300                                         error = hammer_reblock_leaf_node(
 301                                                         reblock, cursor, elm);
 302                                         break;
 303                                 case HAMMER_BTREE_TYPE_INTERNAL:
 304                                         error = hammer_reblock_int_node(
 305                                                         reblock, cursor, elm);
 306                                         break;
 307                                 default:
 308                                         panic("Illegal B-Tree node type");
 309                                 }
 310                         }
 311                         if (error == 0) {
 312                                 ++reblock->btree_moves;
 313                         }
 314                 }
 315         }
 316
 317         hammer_cursor_downgrade(cursor);
 318         return(error);
 319 }
 320
 321 /*
 322  * Reblock a record's data.  Both the B-Tree element and record pointers
 323  * to the data must be adjusted.
 324  */
 325 static int
 326 hammer_reblock_data(struct hammer_ioc_reblock *reblock,
 327                     hammer_cursor_t cursor, hammer_btree_elm_t elm)
 328 {
 329         struct hammer_buffer *data_buffer = NULL;
 330         hammer_off_t ndata_offset;
 331         int error;
 332         void *ndata;
 333
 334         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA |
 335                                              HAMMER_CURSOR_GET_LEAF);
 336         if (error)
 337                 return (error);
 338         ndata = hammer_alloc_data(cursor->trans, elm->leaf.data_len,
 339                                   elm->leaf.base.rec_type,
 340                                   &ndata_offset, &data_buffer, &error);
 341         if (error)
 342                 goto done;
 343
 344         /*
 345          * Move the data
 346          */
 347         hammer_modify_buffer(cursor->trans, data_buffer, NULL, 0);
 348         bcopy(cursor->data, ndata, elm->leaf.data_len);
 349         hammer_modify_buffer_done(data_buffer);
 350
 351         hammer_blockmap_free(cursor->trans,
 352                              elm->leaf.data_offset, elm->leaf.data_len);
 353
 354         hammer_modify_node(cursor->trans, cursor->node,
 355                            &elm->leaf.data_offset, sizeof(hammer_off_t));
 356         elm->leaf.data_offset = ndata_offset;
 357         hammer_modify_node_done(cursor->node);
 358
 359 done:
 360         if (data_buffer)
 361                 hammer_rel_buffer(data_buffer, 0);
 362         return (error);
 363 }
 364
 365 /*
 366  * Reblock a B-Tree leaf node.  The parent must be adjusted to point to
 367  * the new copy of the leaf node.
 368  *
 369  * elm is a pointer to the parent element pointing at cursor.node.
 370  */
 371 static int
 372 hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
 373                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
 374 {
 375         hammer_node_t onode;
 376         hammer_node_t nnode;
 377         int error;
 378
 379         onode = cursor->node;
 380         nnode = hammer_alloc_btree(cursor->trans, &error);
 381
 382         if (nnode == NULL)
 383                 return (error);
 384
 385         /*
 386          * Move the node
 387          */
 388         hammer_lock_ex(&nnode->lock);
 389         hammer_modify_node_noundo(cursor->trans, nnode);
 390         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 391
 392         if (elm) {
 393                 /*
 394                  * We are not the root of the B-Tree
 395                  */
 396                 hammer_modify_node(cursor->trans, cursor->parent,
 397                                    &elm->internal.subtree_offset,
 398                                    sizeof(elm->internal.subtree_offset));
 399                 elm->internal.subtree_offset = nnode->node_offset;
 400                 hammer_modify_node_done(cursor->parent);
 401         } else {
 402                 /*
 403                  * We are the root of the B-Tree
 404                  */
 405                 hammer_volume_t volume;
 406
 407                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 408                 KKASSERT(error == 0);
 409
 410                 hammer_modify_volume_field(cursor->trans, volume,
 411                                            vol0_btree_root);
 412                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 413                 hammer_modify_volume_done(volume);
 414                 hammer_rel_volume(volume, 0);
 415         }
 416
 417         hammer_cursor_replaced_node(onode, nnode);
 418         hammer_delete_node(cursor->trans, onode);
 419
 420         if (hammer_debug_general & 0x4000) {
 421                 kprintf("REBLOCK LNODE %016llx -> %016llx\n",
 422                         onode->node_offset, nnode->node_offset);
 423         }
 424         hammer_modify_node_done(nnode);
 425         cursor->node = nnode;
 426
 427         hammer_unlock(&onode->lock);
 428         hammer_rel_node(onode);
 429
 430         return (error);
 431 }
 432
 433 /*
 434  * Reblock a B-Tree internal node.  The parent must be adjusted to point to
 435  * the new copy of the internal node, and the node's children's parent
 436  * pointers must also be adjusted to point to the new copy.
 437  *
 438  * elm is a pointer to the parent element pointing at cursor.node.
 439  */
 440 static int
 441 hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
 442                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
 443 {
 444         hammer_node_locklist_t locklist = NULL;
 445         hammer_node_t onode;
 446         hammer_node_t nnode;
 447         int error;
 448         int i;
 449
 450         error = hammer_btree_lock_children(cursor, &locklist);
 451         if (error)
 452                 goto done;
 453
 454         onode = cursor->node;
 455         nnode = hammer_alloc_btree(cursor->trans, &error);
 456
 457         if (nnode == NULL)
 458                 goto done;
 459
 460         /*
 461          * Move the node.  Adjust the parent's pointer to us first.
 462          */
 463         hammer_lock_ex(&nnode->lock);
 464         hammer_modify_node_noundo(cursor->trans, nnode);
 465         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 466
 467         if (elm) {
 468                 /*
 469                  * We are not the root of the B-Tree
 470                  */
 471                 hammer_modify_node(cursor->trans, cursor->parent,
 472                                    &elm->internal.subtree_offset,
 473                                    sizeof(elm->internal.subtree_offset));
 474                 elm->internal.subtree_offset = nnode->node_offset;
 475                 hammer_modify_node_done(cursor->parent);
 476         } else {
 477                 /*
 478                  * We are the root of the B-Tree
 479                  */
 480                 hammer_volume_t volume;
 481
 482                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 483                 KKASSERT(error == 0);
 484
 485                 hammer_modify_volume_field(cursor->trans, volume,
 486                                            vol0_btree_root);
 487                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 488                 hammer_modify_volume_done(volume);
 489                 hammer_rel_volume(volume, 0);
 490         }
 491
 492         /*
 493          * Now adjust our children's pointers to us.
 494          */
 495         for (i = 0; i < nnode->ondisk->count; ++i) {
 496                 elm = &nnode->ondisk->elms[i];
 497                 error = btree_set_parent(cursor->trans, nnode, elm);
 498                 if (error)
 499                         panic("reblock internal node: fixup problem");
 500         }
 501
 502         /*
 503          * Clean up.
 504          *
 505          * The new node replaces the current node in the cursor.  The cursor
 506          * expects it to be locked so leave it locked.  Discard onode.
 507          */
 508         hammer_cursor_replaced_node(onode, nnode);
 509         hammer_delete_node(cursor->trans, onode);
 510
 511         if (hammer_debug_general & 0x4000) {
 512                 kprintf("REBLOCK INODE %016llx -> %016llx\n",
 513                         onode->node_offset, nnode->node_offset);
 514         }
 515         hammer_modify_node_done(nnode);
 516         cursor->node = nnode;
 517
 518         hammer_unlock(&onode->lock);
 519         hammer_rel_node(onode);
 520
 521 done:
 522         hammer_btree_unlock_children(&locklist);
 523         return (error);
 524 }
 525