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 1.34 2008/11/13 02:18:43 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         int slop;
  68
  69         /*
  70          * A fill level <= 20% is considered an emergency.  free_level is
  71          * inverted from fill_level.
  72          */
  73         if (reblock->free_level >= HAMMER_LARGEBLOCK_SIZE * 8 / 10)
  74                 slop = HAMMER_CHKSPC_EMERGENCY;
  75         else
  76                 slop = HAMMER_CHKSPC_REBLOCK;
  77
  78         if ((reblock->key_beg.localization | reblock->key_end.localization) &
  79             HAMMER_LOCALIZE_PSEUDOFS_MASK) {
  80                 return(EINVAL);
  81         }
  82         if (reblock->key_beg.obj_id >= reblock->key_end.obj_id)
  83                 return(EINVAL);
  84         if (reblock->free_level < 0)
  85                 return(EINVAL);
  86
  87         reblock->key_cur = reblock->key_beg;
  88         reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
  89         reblock->key_cur.localization += ip->obj_localization;
  90
  91         checkspace_count = 0;
  92         seq = trans->hmp->flusher.act;
  93 retry:
  94         error = hammer_init_cursor(trans, &cursor, NULL, NULL);
  95         if (error) {
  96                 hammer_done_cursor(&cursor);
  97                 goto failed;
  98         }
  99         cursor.key_beg.localization = reblock->key_cur.localization;
 100         cursor.key_beg.obj_id = reblock->key_cur.obj_id;
 101         cursor.key_beg.key = HAMMER_MIN_KEY;
 102         cursor.key_beg.create_tid = 1;
 103         cursor.key_beg.delete_tid = 0;
 104         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
 105         cursor.key_beg.obj_type = 0;
 106
 107         cursor.key_end.localization = (reblock->key_end.localization &
 108                                         HAMMER_LOCALIZE_MASK) +
 109                                       ip->obj_localization;
 110         cursor.key_end.obj_id = reblock->key_end.obj_id;
 111         cursor.key_end.key = HAMMER_MAX_KEY;
 112         cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
 113         cursor.key_end.delete_tid = 0;
 114         cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
 115         cursor.key_end.obj_type = 0;
 116
 117         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
 118         cursor.flags |= HAMMER_CURSOR_BACKEND;
 119
 120         /*
 121          * This flag allows the btree scan code to return internal nodes,
 122          * so we can reblock them in addition to the leafs.  Only specify it
 123          * if we intend to reblock B-Tree nodes.
 124          */
 125         if (reblock->head.flags & HAMMER_IOC_DO_BTREE)
 126                 cursor.flags |= HAMMER_CURSOR_REBLOCKING;
 127
 128         error = hammer_btree_first(&cursor);
 129         while (error == 0) {
 130                 /*
 131                  * Internal or Leaf node
 132                  */
 133                 elm = &cursor.node->ondisk->elms[cursor.index];
 134                 reblock->key_cur.obj_id = elm->base.obj_id;
 135                 reblock->key_cur.localization = elm->base.localization;
 136
 137                 /*
 138                  * Yield to more important tasks
 139                  */
 140                 if ((error = hammer_signal_check(trans->hmp)) != 0)
 141                         break;
 142
 143                 /*
 144                  * If there is insufficient free space it may be due to
 145                  * reserved bigblocks, which flushing might fix.
 146                  *
 147                  * WARNING: See warnings in hammer_unlock_cursor() function.
 148                  */
 149                 if (hammer_checkspace(trans->hmp, slop)) {
 150                         if (++checkspace_count == 10) {
 151                                 error = ENOSPC;
 152                                 break;
 153                         }
 154                         hammer_unlock_cursor(&cursor);
 155                         hammer_flusher_wait(trans->hmp, seq);
 156                         hammer_lock_cursor(&cursor);
 157                         seq = hammer_flusher_async(trans->hmp, NULL);
 158                         continue;
 159                 }
 160
 161                 /*
 162                  * Acquiring the sync_lock prevents the operation from
 163                  * crossing a synchronization boundary.
 164                  *
 165                  * NOTE: cursor.node may have changed on return.
 166                  *
 167                  * WARNING: See warnings in hammer_unlock_cursor() function.
 168                  */
 169                 hammer_sync_lock_sh(trans);
 170                 error = hammer_reblock_helper(reblock, &cursor, elm);
 171                 hammer_sync_unlock(trans);
 172
 173                 while (hammer_flusher_meta_halflimit(trans->hmp) ||
 174                        hammer_flusher_undo_exhausted(trans, 2)) {
 175                         hammer_unlock_cursor(&cursor);
 176                         hammer_flusher_wait(trans->hmp, seq);
 177                         hammer_lock_cursor(&cursor);
 178                         seq = hammer_flusher_async_one(trans->hmp);
 179                 }
 180
 181                 /*
 182                  * Setup for iteration, our cursor flags may be modified by
 183                  * other threads while we are unlocked.
 184                  */
 185                 cursor.flags |= HAMMER_CURSOR_ATEDISK;
 186
 187                 /*
 188                  * We allocate data buffers, which atm we don't track
 189                  * dirty levels for because we allow the kernel to write
 190                  * them.  But if we allocate too many we can still deadlock
 191                  * the buffer cache.
 192                  *
 193                  * WARNING: See warnings in hammer_unlock_cursor() function.
 194                  *          (The cursor's node and element may change!)
 195                  */
 196                 if (bd_heatup()) {
 197                         hammer_unlock_cursor(&cursor);
 198                         bwillwrite(HAMMER_XBUFSIZE);
 199                         hammer_lock_cursor(&cursor);
 200                 }
 201
 202                 if (error == 0) {
 203                         error = hammer_btree_iterate(&cursor);
 204                 }
 205
 206         }
 207         if (error == ENOENT)
 208                 error = 0;
 209         hammer_done_cursor(&cursor);
 210         if (error == EWOULDBLOCK) {
 211                 hammer_flusher_sync(trans->hmp);
 212                 goto retry;
 213         }
 214         if (error == EDEADLK)
 215                 goto retry;
 216         if (error == EINTR) {
 217                 reblock->head.flags |= HAMMER_IOC_HEAD_INTR;
 218                 error = 0;
 219         }
 220 failed:
 221         reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
 222         return(error);
 223 }
 224
 225 /*
 226  * Reblock the B-Tree (leaf) node, record, and/or data if necessary.
 227  *
 228  * XXX We have no visibility into internal B-Tree nodes at the moment,
 229  * only leaf nodes.
 230  */
 231 static int
 232 hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
 233                       hammer_cursor_t cursor, hammer_btree_elm_t elm)
 234 {
 235         hammer_mount_t hmp;
 236         hammer_off_t tmp_offset;
 237         struct hammer_btree_leaf_elm leaf;
 238         int error;
 239         int bytes;
 240         int cur;
 241         int iocflags;
 242
 243         error = 0;
 244         hmp = cursor->trans->hmp;
 245
 246         /*
 247          * Reblock data.  Note that data embedded in a record is reblocked
 248          * by the record reblock code.  Data processing only occurs at leaf
 249          * nodes and for RECORD element types.
 250          */
 251         if (cursor->node->ondisk->type != HAMMER_BTREE_TYPE_LEAF)
 252                 goto skip;
 253         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
 254                 return(0);
 255         tmp_offset = elm->leaf.data_offset;
 256         if (tmp_offset == 0)
 257                 goto skip;
 258         if (error)
 259                 goto skip;
 260
 261         /*
 262          * NOTE: Localization restrictions may also have been set-up, we can't
 263          *       just set the match flags willy-nilly here.
 264          */
 265         switch(elm->leaf.base.rec_type) {
 266         case HAMMER_RECTYPE_INODE:
 267         case HAMMER_RECTYPE_SNAPSHOT:
 268         case HAMMER_RECTYPE_CONFIG:
 269                 iocflags = HAMMER_IOC_DO_INODES;
 270                 break;
 271         case HAMMER_RECTYPE_EXT:
 272         case HAMMER_RECTYPE_FIX:
 273         case HAMMER_RECTYPE_PFS:
 274         case HAMMER_RECTYPE_DIRENTRY:
 275                 iocflags = HAMMER_IOC_DO_DIRS;
 276                 break;
 277         case HAMMER_RECTYPE_DATA:
 278         case HAMMER_RECTYPE_DB:
 279                 iocflags = HAMMER_IOC_DO_DATA;
 280                 break;
 281         default:
 282                 iocflags = 0;
 283                 break;
 284         }
 285         if (reblock->head.flags & iocflags) {
 286                 ++reblock->data_count;
 287                 reblock->data_byte_count += elm->leaf.data_len;
 288                 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
 289                 if (hammer_debug_general & 0x4000)
 290                         kprintf("D %6d/%d\n", bytes, reblock->free_level);
 291                 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
 292                     bytes >= reblock->free_level) {
 293                         /*
 294                          * This is nasty, the uncache code may have to get
 295                          * vnode locks and because of that we can't hold
 296                          * the cursor locked.
 297                          *
 298                          * WARNING: See warnings in hammer_unlock_cursor()
 299                          *          function.
 300                          */
 301                         leaf = elm->leaf;
 302                         hammer_unlock_cursor(cursor);
 303                         hammer_io_direct_uncache(hmp, &leaf);
 304                         hammer_lock_cursor(cursor);
 305                         if (cursor->flags & HAMMER_CURSOR_RETEST) {
 306                                 kprintf("hammer: retest after uncache\n");
 307                                 error = EDEADLK;
 308                         } else {
 309                                 KKASSERT(bcmp(&elm->leaf, &leaf, sizeof(leaf)) == 0);
 310                         }
 311                         if (error == 0)
 312                                 error = hammer_cursor_upgrade(cursor);
 313                         if (error == 0) {
 314                                 error = hammer_reblock_data(reblock,
 315                                                             cursor, elm);
 316                         }
 317                         if (error == 0) {
 318                                 ++reblock->data_moves;
 319                                 reblock->data_byte_moves += elm->leaf.data_len;
 320                         }
 321                 }
 322         }
 323
 324 skip:
 325         /*
 326          * Reblock a B-Tree internal or leaf node.  A leaf node is reblocked
 327          * on initial entry only (element 0).  An internal node is reblocked
 328          * when entered upward from its first leaf node only (also element 0).
 329          * Further revisits of the internal node (index > 0) are ignored.
 330          */
 331         tmp_offset = cursor->node->node_offset;
 332         if (cursor->index == 0 &&
 333             error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) {
 334                 ++reblock->btree_count;
 335                 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
 336                 if (hammer_debug_general & 0x4000)
 337                         kprintf("B %6d/%d\n", bytes, reblock->free_level);
 338                 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
 339                     bytes >= reblock->free_level) {
 340                         error = hammer_cursor_upgrade(cursor);
 341                         if (error == 0) {
 342                                 if (cursor->parent)
 343                                         elm = &cursor->parent->ondisk->elms[cursor->parent_index];
 344                                 else
 345                                         elm = NULL;
 346                                 switch(cursor->node->ondisk->type) {
 347                                 case HAMMER_BTREE_TYPE_LEAF:
 348                                         error = hammer_reblock_leaf_node(
 349                                                         reblock, cursor, elm);
 350                                         break;
 351                                 case HAMMER_BTREE_TYPE_INTERNAL:
 352                                         error = hammer_reblock_int_node(
 353                                                         reblock, cursor, elm);
 354                                         break;
 355                                 default:
 356                                         panic("Illegal B-Tree node type");
 357                                 }
 358                         }
 359                         if (error == 0) {
 360                                 ++reblock->btree_moves;
 361                         }
 362                 }
 363         }
 364
 365         hammer_cursor_downgrade(cursor);
 366         return(error);
 367 }
 368
 369 /*
 370  * Reblock a record's data.  Both the B-Tree element and record pointers
 371  * to the data must be adjusted.
 372  */
 373 static int
 374 hammer_reblock_data(struct hammer_ioc_reblock *reblock,
 375                     hammer_cursor_t cursor, hammer_btree_elm_t elm)
 376 {
 377         struct hammer_buffer *data_buffer = NULL;
 378         hammer_off_t ndata_offset;
 379         int error;
 380         void *ndata;
 381
 382         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA |
 383                                              HAMMER_CURSOR_GET_LEAF);
 384         if (error)
 385                 return (error);
 386         ndata = hammer_alloc_data(cursor->trans, elm->leaf.data_len,
 387                                   elm->leaf.base.rec_type,
 388                                   &ndata_offset, &data_buffer,
 389                                   0, &error);
 390         if (error)
 391                 goto done;
 392
 393         /*
 394          * Move the data
 395          */
 396         hammer_modify_buffer(cursor->trans, data_buffer, NULL, 0);
 397         bcopy(cursor->data, ndata, elm->leaf.data_len);
 398         hammer_modify_buffer_done(data_buffer);
 399
 400         hammer_blockmap_free(cursor->trans,
 401                              elm->leaf.data_offset, elm->leaf.data_len);
 402
 403         hammer_modify_node(cursor->trans, cursor->node,
 404                            &elm->leaf.data_offset, sizeof(hammer_off_t));
 405         elm->leaf.data_offset = ndata_offset;
 406         hammer_modify_node_done(cursor->node);
 407
 408 done:
 409         if (data_buffer)
 410                 hammer_rel_buffer(data_buffer, 0);
 411         return (error);
 412 }
 413
 414 /*
 415  * Reblock a B-Tree leaf node.  The parent must be adjusted to point to
 416  * the new copy of the leaf node.
 417  *
 418  * elm is a pointer to the parent element pointing at cursor.node.
 419  */
 420 static int
 421 hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
 422                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
 423 {
 424         hammer_node_t onode;
 425         hammer_node_t nnode;
 426         int error;
 427
 428         /*
 429          * Don't supply a hint when allocating the leaf.  Fills are done
 430          * from the leaf upwards.
 431          */
 432         onode = cursor->node;
 433         nnode = hammer_alloc_btree(cursor->trans, 0, &error);
 434
 435         if (nnode == NULL)
 436                 return (error);
 437
 438         /*
 439          * Move the node
 440          */
 441         hammer_lock_ex(&nnode->lock);
 442         hammer_modify_node_noundo(cursor->trans, nnode);
 443         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 444
 445         if (elm) {
 446                 /*
 447                  * We are not the root of the B-Tree
 448                  */
 449                 hammer_modify_node(cursor->trans, cursor->parent,
 450                                    &elm->internal.subtree_offset,
 451                                    sizeof(elm->internal.subtree_offset));
 452                 elm->internal.subtree_offset = nnode->node_offset;
 453                 hammer_modify_node_done(cursor->parent);
 454         } else {
 455                 /*
 456                  * We are the root of the B-Tree
 457                  */
 458                 hammer_volume_t volume;
 459
 460                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 461                 KKASSERT(error == 0);
 462
 463                 hammer_modify_volume_field(cursor->trans, volume,
 464                                            vol0_btree_root);
 465                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 466                 hammer_modify_volume_done(volume);
 467                 hammer_rel_volume(volume, 0);
 468         }
 469
 470         hammer_cursor_replaced_node(onode, nnode);
 471         hammer_delete_node(cursor->trans, onode);
 472
 473         if (hammer_debug_general & 0x4000) {
 474                 kprintf("REBLOCK LNODE %016llx -> %016llx\n",
 475                         (long long)onode->node_offset,
 476                         (long long)nnode->node_offset);
 477         }
 478         hammer_modify_node_done(nnode);
 479         cursor->node = nnode;
 480
 481         hammer_unlock(&onode->lock);
 482         hammer_rel_node(onode);
 483
 484         return (error);
 485 }
 486
 487 /*
 488  * Reblock a B-Tree internal node.  The parent must be adjusted to point to
 489  * the new copy of the internal node, and the node's children's parent
 490  * pointers must also be adjusted to point to the new copy.
 491  *
 492  * elm is a pointer to the parent element pointing at cursor.node.
 493  */
 494 static int
 495 hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
 496                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
 497 {
 498         struct hammer_node_lock lockroot;
 499         hammer_node_t onode;
 500         hammer_node_t nnode;
 501         hammer_off_t hint;
 502         int error;
 503         int i;
 504
 505         hammer_node_lock_init(&lockroot, cursor->node);
 506         error = hammer_btree_lock_children(cursor, 1, &lockroot);
 507         if (error)
 508                 goto done;
 509
 510         /*
 511          * The internal node is visited after recursing through its
 512          * first element.  Use the subtree offset allocated for that
 513          * element as a hint for allocating the internal node.
 514          */
 515         onode = cursor->node;
 516         if (onode->ondisk->count)
 517                 hint = onode->ondisk->elms[0].internal.subtree_offset;
 518         else
 519                 hint = 0;
 520         nnode = hammer_alloc_btree(cursor->trans, hint, &error);
 521
 522         if (nnode == NULL)
 523                 goto done;
 524
 525         /*
 526          * Move the node.  Adjust the parent's pointer to us first.
 527          */
 528         hammer_lock_ex(&nnode->lock);
 529         hammer_modify_node_noundo(cursor->trans, nnode);
 530         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 531
 532         if (elm) {
 533                 /*
 534                  * We are not the root of the B-Tree
 535                  */
 536                 hammer_modify_node(cursor->trans, cursor->parent,
 537                                    &elm->internal.subtree_offset,
 538                                    sizeof(elm->internal.subtree_offset));
 539                 elm->internal.subtree_offset = nnode->node_offset;
 540                 hammer_modify_node_done(cursor->parent);
 541         } else {
 542                 /*
 543                  * We are the root of the B-Tree
 544                  */
 545                 hammer_volume_t volume;
 546
 547                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 548                 KKASSERT(error == 0);
 549
 550                 hammer_modify_volume_field(cursor->trans, volume,
 551                                            vol0_btree_root);
 552                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 553                 hammer_modify_volume_done(volume);
 554                 hammer_rel_volume(volume, 0);
 555         }
 556
 557         /*
 558          * Now adjust our children's pointers to us.
 559          */
 560         for (i = 0; i < nnode->ondisk->count; ++i) {
 561                 elm = &nnode->ondisk->elms[i];
 562                 error = btree_set_parent(cursor->trans, nnode, elm);
 563                 if (error)
 564                         panic("reblock internal node: fixup problem");
 565         }
 566
 567         /*
 568          * Clean up.
 569          *
 570          * The new node replaces the current node in the cursor.  The cursor
 571          * expects it to be locked so leave it locked.  Discard onode.
 572          */
 573         hammer_cursor_replaced_node(onode, nnode);
 574         hammer_delete_node(cursor->trans, onode);
 575
 576         if (hammer_debug_general & 0x4000) {
 577                 kprintf("REBLOCK INODE %016llx -> %016llx\n",
 578                         (long long)onode->node_offset,
 579                         (long long)nnode->node_offset);
 580         }
 581         hammer_modify_node_done(nnode);
 582         cursor->node = nnode;
 583
 584         hammer_unlock(&onode->lock);
 585         hammer_rel_node(onode);
 586
 587 done:
 588         hammer_btree_unlock_children(cursor, &lockroot);
 589         return (error);
 590 }
 591