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 1.25 2008/07/09 10:29:20 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         int checkspace_count;
  66
  67         if ((reblock->key_beg.localization | reblock->key_end.localization) &
  68             HAMMER_LOCALIZE_PSEUDOFS_MASK) {
  69                 return(EINVAL);
  70         }
  71         if (reblock->key_beg.obj_id >= reblock->key_end.obj_id)
  72                 return(EINVAL);
  73         if (reblock->free_level < 0)
  74                 return(EINVAL);
  75
  76         reblock->key_cur = reblock->key_beg;
  77         reblock->key_cur.localization += ip->obj_localization;
  78
  79         checkspace_count = 0;
  80 retry:
  81         error = hammer_init_cursor(trans, &cursor, NULL, NULL);
  82         if (error) {
  83                 hammer_done_cursor(&cursor);
  84                 goto failed;
  85         }
  86         cursor.key_beg.localization = reblock->key_cur.localization;
  87         cursor.key_beg.obj_id = reblock->key_cur.obj_id;
  88         cursor.key_beg.key = HAMMER_MIN_KEY;
  89         cursor.key_beg.create_tid = 1;
  90         cursor.key_beg.delete_tid = 0;
  91         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
  92         cursor.key_beg.obj_type = 0;
  93
  94         cursor.key_end.localization = reblock->key_end.localization +
  95                                       ip->obj_localization;
  96         cursor.key_end.obj_id = reblock->key_end.obj_id;
  97         cursor.key_end.key = HAMMER_MAX_KEY;
  98         cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
  99         cursor.key_end.delete_tid = 0;
 100         cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
 101         cursor.key_end.obj_type = 0;
 102
 103         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
 104         cursor.flags |= HAMMER_CURSOR_BACKEND;
 105
 106         /*
 107          * This flag allows the btree scan code to return internal nodes,
 108          * so we can reblock them in addition to the leafs.  Only specify it
 109          * if we intend to reblock B-Tree nodes.
 110          */
 111         if (reblock->head.flags & HAMMER_IOC_DO_BTREE)
 112                 cursor.flags |= HAMMER_CURSOR_REBLOCKING;
 113
 114         error = hammer_btree_first(&cursor);
 115         while (error == 0) {
 116                 /*
 117                  * Internal or Leaf node
 118                  */
 119                 elm = &cursor.node->ondisk->elms[cursor.index];
 120                 reblock->key_cur.obj_id = elm->base.obj_id;
 121                 reblock->key_cur.localization = elm->base.localization;
 122
 123                 /*
 124                  * Yield to more important tasks
 125                  */
 126                 if ((error = hammer_signal_check(trans->hmp)) != 0)
 127                         break;
 128                 if (trans->hmp->sync_lock.wanted) {
 129                         tsleep(trans, 0, "hmrslo", hz / 10);
 130                 }
 131
 132                 /*
 133                  * If we build up too much meta-data we have to wait for
 134                  * a flush cycle.
 135                  */
 136                 if (hammer_flusher_meta_limit(trans->hmp) ||
 137                     hammer_flusher_undo_exhausted(trans, 2)) {
 138                         error = EWOULDBLOCK;
 139                         break;
 140                 }
 141
 142                 /*
 143                  * If there is insufficient free space it may be due to
 144                  * reserved bigblocks, which flushing might fix.
 145                  */
 146                 if (hammer_checkspace(trans->hmp, HAMMER_CHECKSPACE_SLOP_REBLOCK)) {
 147                         if (++checkspace_count == 10) {
 148                                 error = ENOSPC;
 149                         } else {
 150                                 error = EWOULDBLOCK;
 151                         }
 152                         break;
 153                 }
 154
 155                 /*
 156                  * Acquiring the sync_lock prevents the operation from
 157                  * crossing a synchronization boundary.
 158                  *
 159                  * NOTE: cursor.node may have changed on return.
 160                  */
 161                 hammer_sync_lock_sh(trans);
 162                 error = hammer_reblock_helper(reblock, &cursor, elm);
 163                 hammer_sync_unlock(trans);
 164                 if (error == 0) {
 165                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
 166                         error = hammer_btree_iterate(&cursor);
 167                 }
 168
 169         }
 170         if (error == ENOENT)
 171                 error = 0;
 172         hammer_done_cursor(&cursor);
 173         if (error == EWOULDBLOCK) {
 174                 hammer_flusher_sync(trans->hmp);
 175                 goto retry;
 176         }
 177         if (error == EDEADLK)
 178                 goto retry;
 179         if (error == EINTR) {
 180                 reblock->head.flags |= HAMMER_IOC_HEAD_INTR;
 181                 error = 0;
 182         }
 183 failed:
 184         reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
 185         return(error);
 186 }
 187
 188 /*
 189  * Reblock the B-Tree (leaf) node, record, and/or data if necessary.
 190  *
 191  * XXX We have no visibility into internal B-Tree nodes at the moment,
 192  * only leaf nodes.
 193  */
 194 static int
 195 hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
 196                       hammer_cursor_t cursor, hammer_btree_elm_t elm)
 197 {
 198         hammer_mount_t hmp;
 199         hammer_off_t tmp_offset;
 200         int error;
 201         int bytes;
 202         int cur;
 203         int iocflags;
 204
 205         error = 0;
 206         hmp = cursor->trans->hmp;
 207
 208         /*
 209          * Reblock data.  Note that data embedded in a record is reblocked
 210          * by the record reblock code.  Data processing only occurs at leaf
 211          * nodes and for RECORD element types.
 212          */
 213         if (cursor->node->ondisk->type != HAMMER_BTREE_TYPE_LEAF)
 214                 goto skip;
 215         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
 216                 return(0);
 217         tmp_offset = elm->leaf.data_offset;
 218         if (tmp_offset == 0)
 219                 goto skip;
 220         if (error)
 221                 goto skip;
 222
 223         /*
 224          * NOTE: Localization restrictions may also have been set-up, we can't
 225          * just set the match flags willy-nilly here.
 226          */
 227         switch(elm->leaf.base.rec_type) {
 228         case HAMMER_RECTYPE_INODE:
 229                 iocflags = HAMMER_IOC_DO_INODES;
 230                 break;
 231         case HAMMER_RECTYPE_EXT:
 232         case HAMMER_RECTYPE_FIX:
 233         case HAMMER_RECTYPE_PFS:
 234         case HAMMER_RECTYPE_DIRENTRY:
 235                 iocflags = HAMMER_IOC_DO_DIRS;
 236                 break;
 237         case HAMMER_RECTYPE_DATA:
 238         case HAMMER_RECTYPE_DB:
 239                 iocflags = HAMMER_IOC_DO_DATA;
 240                 break;
 241         default:
 242                 iocflags = 0;
 243                 break;
 244         }
 245         if (reblock->head.flags & iocflags) {
 246                 ++reblock->data_count;
 247                 reblock->data_byte_count += elm->leaf.data_len;
 248                 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
 249                 if (hammer_debug_general & 0x4000)
 250                         kprintf("D %6d/%d\n", bytes, reblock->free_level);
 251                 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
 252                     bytes >= reblock->free_level) {
 253                         hammer_io_direct_uncache(hmp, &elm->leaf);
 254                         error = hammer_cursor_upgrade(cursor);
 255                         if (error == 0) {
 256                                 error = hammer_reblock_data(reblock,
 257                                                             cursor, elm);
 258                         }
 259                         if (error == 0) {
 260                                 ++reblock->data_moves;
 261                                 reblock->data_byte_moves += elm->leaf.data_len;
 262                         }
 263                 }
 264         }
 265
 266 skip:
 267         /*
 268          * Reblock a B-Tree internal or leaf node.
 269          */
 270         tmp_offset = cursor->node->node_offset;
 271         if (cursor->index == 0 &&
 272             error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) {
 273                 ++reblock->btree_count;
 274                 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
 275                 if (hammer_debug_general & 0x4000)
 276                         kprintf("B %6d/%d\n", bytes, reblock->free_level);
 277                 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
 278                     bytes >= reblock->free_level) {
 279                         error = hammer_cursor_upgrade(cursor);
 280                         if (error == 0) {
 281                                 if (cursor->parent)
 282                                         elm = &cursor->parent->ondisk->elms[cursor->parent_index];
 283                                 else
 284                                         elm = NULL;
 285                                 switch(cursor->node->ondisk->type) {
 286                                 case HAMMER_BTREE_TYPE_LEAF:
 287                                         error = hammer_reblock_leaf_node(
 288                                                         reblock, cursor, elm);
 289                                         break;
 290                                 case HAMMER_BTREE_TYPE_INTERNAL:
 291                                         error = hammer_reblock_int_node(
 292                                                         reblock, cursor, elm);
 293                                         break;
 294                                 default:
 295                                         panic("Illegal B-Tree node type");
 296                                 }
 297                         }
 298                         if (error == 0) {
 299                                 ++reblock->btree_moves;
 300                         }
 301                 }
 302         }
 303
 304         hammer_cursor_downgrade(cursor);
 305         return(error);
 306 }
 307
 308 /*
 309  * Reblock a record's data.  Both the B-Tree element and record pointers
 310  * to the data must be adjusted.
 311  */
 312 static int
 313 hammer_reblock_data(struct hammer_ioc_reblock *reblock,
 314                     hammer_cursor_t cursor, hammer_btree_elm_t elm)
 315 {
 316         struct hammer_buffer *data_buffer = NULL;
 317         hammer_off_t ndata_offset;
 318         int error;
 319         void *ndata;
 320
 321         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA |
 322                                              HAMMER_CURSOR_GET_LEAF);
 323         if (error)
 324                 return (error);
 325         ndata = hammer_alloc_data(cursor->trans, elm->leaf.data_len,
 326                                   elm->leaf.base.rec_type,
 327                                   &ndata_offset, &data_buffer, &error);
 328         if (error)
 329                 goto done;
 330
 331         /*
 332          * Move the data
 333          */
 334         hammer_modify_buffer(cursor->trans, data_buffer, NULL, 0);
 335         bcopy(cursor->data, ndata, elm->leaf.data_len);
 336         hammer_modify_buffer_done(data_buffer);
 337
 338         hammer_blockmap_free(cursor->trans,
 339                              elm->leaf.data_offset, elm->leaf.data_len);
 340
 341         hammer_modify_node(cursor->trans, cursor->node,
 342                            &elm->leaf.data_offset, sizeof(hammer_off_t));
 343         elm->leaf.data_offset = ndata_offset;
 344         hammer_modify_node_done(cursor->node);
 345
 346 done:
 347         if (data_buffer)
 348                 hammer_rel_buffer(data_buffer, 0);
 349         return (error);
 350 }
 351
 352 /*
 353  * Reblock a B-Tree leaf node.  The parent must be adjusted to point to
 354  * the new copy of the leaf node.
 355  *
 356  * elm is a pointer to the parent element pointing at cursor.node.
 357  */
 358 static int
 359 hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
 360                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
 361 {
 362         hammer_node_t onode;
 363         hammer_node_t nnode;
 364         int error;
 365
 366         onode = cursor->node;
 367         nnode = hammer_alloc_btree(cursor->trans, &error);
 368
 369         if (nnode == NULL)
 370                 return (error);
 371
 372         /*
 373          * Move the node
 374          */
 375         hammer_lock_ex(&nnode->lock);
 376         hammer_modify_node_noundo(cursor->trans, nnode);
 377         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 378
 379         if (elm) {
 380                 /*
 381                  * We are not the root of the B-Tree
 382                  */
 383                 hammer_modify_node(cursor->trans, cursor->parent,
 384                                    &elm->internal.subtree_offset,
 385                                    sizeof(elm->internal.subtree_offset));
 386                 elm->internal.subtree_offset = nnode->node_offset;
 387                 hammer_modify_node_done(cursor->parent);
 388         } else {
 389                 /*
 390                  * We are the root of the B-Tree
 391                  */
 392                 hammer_volume_t volume;
 393
 394                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 395                 KKASSERT(error == 0);
 396
 397                 hammer_modify_volume_field(cursor->trans, volume,
 398                                            vol0_btree_root);
 399                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 400                 hammer_modify_volume_done(volume);
 401                 hammer_rel_volume(volume, 0);
 402         }
 403
 404         hammer_cursor_replaced_node(onode, nnode);
 405         hammer_delete_node(cursor->trans, onode);
 406
 407         if (hammer_debug_general & 0x4000) {
 408                 kprintf("REBLOCK LNODE %016llx -> %016llx\n",
 409                         onode->node_offset, nnode->node_offset);
 410         }
 411         hammer_modify_node_done(nnode);
 412         cursor->node = nnode;
 413
 414         hammer_unlock(&onode->lock);
 415         hammer_rel_node(onode);
 416
 417         return (error);
 418 }
 419
 420 /*
 421  * Reblock a B-Tree internal node.  The parent must be adjusted to point to
 422  * the new copy of the internal node, and the node's children's parent
 423  * pointers must also be adjusted to point to the new copy.
 424  *
 425  * elm is a pointer to the parent element pointing at cursor.node.
 426  */
 427 static int
 428 hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
 429                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
 430 {
 431         hammer_node_locklist_t locklist = NULL;
 432         hammer_node_t onode;
 433         hammer_node_t nnode;
 434         int error;
 435         int i;
 436
 437         error = hammer_btree_lock_children(cursor, &locklist);
 438         if (error)
 439                 goto done;
 440
 441         onode = cursor->node;
 442         nnode = hammer_alloc_btree(cursor->trans, &error);
 443
 444         if (nnode == NULL)
 445                 goto done;
 446
 447         /*
 448          * Move the node.  Adjust the parent's pointer to us first.
 449          */
 450         hammer_lock_ex(&nnode->lock);
 451         hammer_modify_node_noundo(cursor->trans, nnode);
 452         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 453
 454         if (elm) {
 455                 /*
 456                  * We are not the root of the B-Tree
 457                  */
 458                 hammer_modify_node(cursor->trans, cursor->parent,
 459                                    &elm->internal.subtree_offset,
 460                                    sizeof(elm->internal.subtree_offset));
 461                 elm->internal.subtree_offset = nnode->node_offset;
 462                 hammer_modify_node_done(cursor->parent);
 463         } else {
 464                 /*
 465                  * We are the root of the B-Tree
 466                  */
 467                 hammer_volume_t volume;
 468
 469                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 470                 KKASSERT(error == 0);
 471
 472                 hammer_modify_volume_field(cursor->trans, volume,
 473                                            vol0_btree_root);
 474                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 475                 hammer_modify_volume_done(volume);
 476                 hammer_rel_volume(volume, 0);
 477         }
 478
 479         /*
 480          * Now adjust our children's pointers to us.
 481          */
 482         for (i = 0; i < nnode->ondisk->count; ++i) {
 483                 elm = &nnode->ondisk->elms[i];
 484                 error = btree_set_parent(cursor->trans, nnode, elm);
 485                 if (error)
 486                         panic("reblock internal node: fixup problem");
 487         }
 488
 489         /*
 490          * Clean up.
 491          *
 492          * The new node replaces the current node in the cursor.  The cursor
 493          * expects it to be locked so leave it locked.  Discard onode.
 494          */
 495         hammer_cursor_replaced_node(onode, nnode);
 496         hammer_delete_node(cursor->trans, onode);
 497
 498         if (hammer_debug_general & 0x4000) {
 499                 kprintf("REBLOCK INODE %016llx -> %016llx\n",
 500                         onode->node_offset, nnode->node_offset);
 501         }
 502         hammer_modify_node_done(nnode);
 503         cursor->node = nnode;
 504
 505         hammer_unlock(&onode->lock);
 506         hammer_rel_node(onode);
 507
 508 done:
 509         hammer_btree_unlock_children(&locklist);
 510         return (error);
 511 }
 512