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 1.3 2008/03/19 20:18:17 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_record(struct hammer_ioc_reblock *reblock,
  54                                 hammer_cursor_t cursor, hammer_btree_elm_t elm);
  55 static int hammer_reblock_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 retry:
  72         error = hammer_init_cursor(trans, &cursor, NULL);
  73         if (error) {
  74                 hammer_done_cursor(&cursor);
  75                 return(error);
  76         }
  77         cursor.key_beg.obj_id = reblock->cur_obj_id;
  78         cursor.key_beg.key = HAMMER_MIN_KEY;
  79         cursor.key_beg.create_tid = 1;
  80         cursor.key_beg.delete_tid = 0;
  81         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
  82         cursor.key_beg.obj_type = 0;
  83
  84         cursor.key_end.obj_id = reblock->end_obj_id;
  85         cursor.key_end.key = HAMMER_MAX_KEY;
  86         cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
  87         cursor.key_end.delete_tid = 0;
  88         cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
  89         cursor.key_end.obj_type = 0;
  90
  91         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
  92
  93         error = hammer_btree_first(&cursor);
  94         while (error == 0) {
  95                 elm = &cursor.node->ondisk->elms[cursor.index];
  96                 reblock->cur_obj_id = elm->base.obj_id;
  97
  98                 error = hammer_reblock_helper(reblock, &cursor, elm);
  99                 if (error == 0) {
 100                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
 101                         error = hammer_btree_iterate(&cursor);
 102                 }
 103         }
 104         if (error == ENOENT)
 105                 error = 0;
 106         hammer_done_cursor(&cursor);
 107         if (error == EDEADLK)
 108                 goto retry;
 109         return(error);
 110 }
 111
 112 /*
 113  * Reblock the B-Tree (leaf) node, record, and/or data if necessary.
 114  *
 115  * XXX We have no visibility into internal B-Tree nodes at the moment.
 116  */
 117 static int
 118 hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
 119                       hammer_cursor_t cursor, hammer_btree_elm_t elm)
 120 {
 121         hammer_off_t tmp_offset;
 122         int error;
 123         int zone;
 124         int bytes;
 125         int cur;
 126
 127         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
 128                 return(0);
 129         error = 0;
 130
 131         /*
 132          * Reblock data.  Note that data embedded in a record is reblocked
 133          * by the record reblock code.
 134          */
 135         tmp_offset = elm->leaf.data_offset;
 136         zone = HAMMER_ZONE_DECODE(tmp_offset);          /* can be 0 */
 137         if ((zone == HAMMER_ZONE_SMALL_DATA_INDEX ||
 138              zone == HAMMER_ZONE_LARGE_DATA_INDEX) && error == 0) {
 139                 ++reblock->data_count;
 140                 reblock->data_byte_count += elm->leaf.data_len;
 141                 bytes = hammer_blockmap_getfree(cursor->trans->hmp, tmp_offset,
 142                                                 &cur, &error);
 143                 if (error == 0 && cur == 0 && bytes > reblock->free_level) {
 144                         kprintf("%6d ", bytes);
 145                         error = hammer_cursor_upgrade(cursor);
 146                         if (error == 0) {
 147                                 error = hammer_reblock_data(reblock,
 148                                                             cursor, elm);
 149                         }
 150                         if (error == 0) {
 151                                 ++reblock->data_moves;
 152                                 reblock->data_byte_moves += elm->leaf.data_len;
 153                         }
 154                 }
 155         }
 156
 157         /*
 158          * Reblock a record
 159          */
 160         tmp_offset = elm->leaf.rec_offset;
 161         zone = HAMMER_ZONE_DECODE(tmp_offset);
 162         if (zone == HAMMER_ZONE_RECORD_INDEX && error == 0) {
 163                 ++reblock->record_count;
 164                 bytes = hammer_blockmap_getfree(cursor->trans->hmp, tmp_offset,
 165                                                 &cur, &error);
 166                 if (error == 0 && cur == 0 && bytes > reblock->free_level) {
 167                         kprintf("%6d ", bytes);
 168                         error = hammer_cursor_upgrade(cursor);
 169                         if (error == 0) {
 170                                 error = hammer_reblock_record(reblock,
 171                                                               cursor, elm);
 172                         }
 173                         if (error == 0) {
 174                                 ++reblock->record_moves;
 175                         }
 176                 }
 177         }
 178
 179         /*
 180          * Reblock a B-Tree node.  Adjust elm to point at the parent's
 181          * leaf entry.
 182          */
 183         tmp_offset = cursor->node->node_offset;
 184         zone = HAMMER_ZONE_DECODE(tmp_offset);
 185         if (zone == HAMMER_ZONE_BTREE_INDEX && error == 0 &&
 186             cursor->index == 0) {
 187                 ++reblock->btree_count;
 188                 bytes = hammer_blockmap_getfree(cursor->trans->hmp, tmp_offset,
 189                                                 &cur, &error);
 190                 if (error == 0 && cur == 0 && bytes > reblock->free_level) {
 191                         kprintf("%6d ", bytes);
 192                         error = hammer_cursor_upgrade(cursor);
 193                         if (error == 0) {
 194                                 if (cursor->parent)
 195                                         elm = &cursor->parent->ondisk->elms[cursor->parent_index];
 196                                 else
 197                                         elm = NULL;
 198                                 error = hammer_reblock_node(reblock,
 199                                                             cursor, elm);
 200                         }
 201                         if (error == 0) {
 202                                 ++reblock->btree_moves;
 203                         }
 204                 }
 205         }
 206
 207         hammer_cursor_downgrade(cursor);
 208         return(error);
 209 }
 210
 211 /*
 212  * Reblock a record's data.  Both the B-Tree element and record pointers
 213  * to the data must be adjusted.
 214  */
 215 static int
 216 hammer_reblock_data(struct hammer_ioc_reblock *reblock,
 217                     hammer_cursor_t cursor, hammer_btree_elm_t elm)
 218 {
 219         struct hammer_buffer *data_buffer = NULL;
 220         hammer_off_t ndata_offset;
 221         int error;
 222         void *ndata;
 223
 224         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA |
 225                                              HAMMER_CURSOR_GET_RECORD);
 226         if (error)
 227                 return (error);
 228         ndata = hammer_alloc_data(cursor->trans, elm->leaf.data_len,
 229                                   &ndata_offset, &data_buffer, &error);
 230         if (error)
 231                 goto done;
 232
 233         /*
 234          * Move the data
 235          */
 236         bcopy(cursor->data, ndata, elm->leaf.data_len);
 237         hammer_modify_node(cursor->trans, cursor->node,
 238                            &elm->leaf.data_offset, sizeof(hammer_off_t));
 239         hammer_modify_record(cursor->trans, cursor->record_buffer,
 240                              &cursor->record->base.data_off,
 241                              sizeof(hammer_off_t));
 242
 243         hammer_blockmap_free(cursor->trans,
 244                              elm->leaf.data_offset, elm->leaf.data_len);
 245
 246         cursor->record->base.data_off = ndata_offset;
 247         elm->leaf.data_offset = ndata_offset;
 248
 249 done:
 250         if (data_buffer)
 251                 hammer_rel_buffer(data_buffer, 0);
 252         return (error);
 253 }
 254
 255 /*
 256  * Reblock a record.  The B-Tree must be adjusted to point to the new record
 257  * and the existing record must be physically destroyed so a FS rebuild
 258  * does not see two versions of the same record.
 259  */
 260 static int
 261 hammer_reblock_record(struct hammer_ioc_reblock *reblock,
 262                       hammer_cursor_t cursor, hammer_btree_elm_t elm)
 263 {
 264         struct hammer_buffer *rec_buffer = NULL;
 265         hammer_off_t nrec_offset;
 266         hammer_off_t ndata_offset;
 267         hammer_record_ondisk_t orec;
 268         hammer_record_ondisk_t nrec;
 269         int error;
 270         int inline_data;
 271
 272         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_RECORD);
 273         if (error)
 274                 return (error);
 275
 276         nrec = hammer_alloc_record(cursor->trans, &nrec_offset,
 277                                    elm->leaf.base.rec_type, &rec_buffer,
 278                                    0, NULL, NULL, &error);
 279         if (error)
 280                 goto done;
 281
 282         /*
 283          * Move the record.  Check for an inline data reference and move that
 284          * too if necessary.
 285          */
 286         orec = cursor->record;
 287         bcopy(orec, nrec, sizeof(*nrec));
 288
 289         if ((orec->base.data_off & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_RECORD) {
 290                 ndata_offset = orec->base.data_off - elm->leaf.rec_offset;
 291                 KKASSERT(ndata_offset < sizeof(*nrec));
 292                 ndata_offset += nrec_offset;
 293                 inline_data = 1;
 294         } else {
 295                 ndata_offset = 0;
 296                 inline_data = 0;
 297         }
 298
 299         hammer_modify_record(cursor->trans, cursor->record_buffer,
 300                              &orec->base.base.rec_type,
 301                              sizeof(orec->base.base.rec_type));
 302         orec->base.base.rec_type |= HAMMER_RECTYPE_MOVED;
 303
 304         hammer_blockmap_free(cursor->trans,
 305                              elm->leaf.rec_offset, sizeof(*nrec));
 306
 307         kprintf("REBLOCK RECD %016llx -> %016llx\n",
 308                 elm->leaf.rec_offset, nrec_offset);
 309
 310         hammer_modify_node(cursor->trans, cursor->node,
 311                            &elm->leaf.rec_offset, sizeof(hammer_off_t));
 312         elm->leaf.rec_offset = nrec_offset;
 313         if (inline_data) {
 314                 hammer_modify_node(cursor->trans, cursor->node,
 315                                  &elm->leaf.data_offset, sizeof(hammer_off_t));
 316                 elm->leaf.data_offset = ndata_offset;
 317         }
 318
 319 done:
 320         if (rec_buffer)
 321                 hammer_rel_buffer(rec_buffer, 0);
 322         return (error);
 323 }
 324
 325 /*
 326  * Reblock a B-Tree (leaf) node.  The parent must be adjusted to point to
 327  * the new copy of the leaf node.  elm is a pointer to the parent element
 328  * pointing at cursor.node.
 329  *
 330  * XXX reblock internal nodes too.
 331  */
 332 static int
 333 hammer_reblock_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         hammer_lock_ex(&nnode->lock);
 343
 344         if (nnode == NULL)
 345                 return (error);
 346
 347         /*
 348          * Move the node
 349          */
 350         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 351
 352         if (elm) {
 353                 /*
 354                  * We are not the root of the B-Tree
 355                  */
 356                 hammer_modify_node(cursor->trans, cursor->parent,
 357                                    &elm->internal.subtree_offset,
 358                                    sizeof(elm->internal.subtree_offset));
 359                 elm->internal.subtree_offset = nnode->node_offset;
 360         } else {
 361                 /*
 362                  * We are the root of the B-Tree
 363                  */
 364                 hammer_volume_t volume;
 365
 366                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 367                 KKASSERT(error == 0);
 368
 369                 hammer_modify_volume(cursor->trans, volume,
 370                                      &volume->ondisk->vol0_btree_root,
 371                                      sizeof(hammer_off_t));
 372                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 373                 hammer_rel_volume(volume, 0);
 374         }
 375
 376         hammer_delete_node(cursor->trans, onode);
 377
 378         kprintf("REBLOCK NODE %016llx -> %016llx\n",
 379                 onode->node_offset, nnode->node_offset);
 380
 381         cursor->node = nnode;
 382         hammer_rel_node(onode);
 383
 384         return (error);
 385 }
 386