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 1.14 2008/05/12 21:17:18 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_node(struct hammer_ioc_reblock *reblock,
  54                                 hammer_cursor_t cursor, hammer_btree_elm_t elm);
  55
  56 int
  57 hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
  58                struct hammer_ioc_reblock *reblock)
  59 {
  60         struct hammer_cursor cursor;
  61         hammer_btree_elm_t elm;
  62         int error;
  63
  64         if (reblock->beg_obj_id >= reblock->end_obj_id)
  65                 return(EINVAL);
  66         if (reblock->free_level < 0)
  67                 return(EINVAL);
  68
  69         reblock->cur_obj_id = reblock->beg_obj_id;
  70
  71 retry:
  72         error = hammer_init_cursor(trans, &cursor, NULL, 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         cursor.flags |= HAMMER_CURSOR_BACKEND;
  93
  94         error = hammer_btree_first(&cursor);
  95         while (error == 0) {
  96                 elm = &cursor.node->ondisk->elms[cursor.index];
  97                 reblock->cur_obj_id = elm->base.obj_id;
  98
  99                 /*
 100                  * Acquiring the sync_lock prevents the operation from
 101                  * crossing a synchronization boundary.
 102                  */
 103                 hammer_lock_ex(&trans->hmp->sync_lock);
 104                 error = hammer_reblock_helper(reblock, &cursor, elm);
 105                 hammer_unlock(&trans->hmp->sync_lock);
 106                 if (error == 0) {
 107                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
 108                         error = hammer_btree_iterate(&cursor);
 109                 }
 110
 111                 /*
 112                  * Bad hack for now, don't blow out the kernel's buffer
 113                  * cache.  NOTE: We still hold locks on the cursor, we
 114                  * cannot call the flusher synchronously.
 115                  */
 116                 if (trans->hmp->locked_dirty_count > hammer_limit_dirtybufs) {
 117                         hammer_flusher_async(trans->hmp);
 118                         tsleep(trans, 0, "hmrslo", hz / 10);
 119                 }
 120                 if (error == 0)
 121                         error = hammer_signal_check(trans->hmp);
 122         }
 123         if (error == ENOENT)
 124                 error = 0;
 125         hammer_done_cursor(&cursor);
 126         if (error == EDEADLK)
 127                 goto retry;
 128         if (error == EINTR) {
 129                 reblock->head.flags |= HAMMER_IOC_HEAD_INTR;
 130                 error = 0;
 131         }
 132         return(error);
 133 }
 134
 135 /*
 136  * Reblock the B-Tree (leaf) node, record, and/or data if necessary.
 137  *
 138  * XXX We have no visibility into internal B-Tree nodes at the moment,
 139  * only leaf nodes.
 140  */
 141 static int
 142 hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
 143                       hammer_cursor_t cursor, hammer_btree_elm_t elm)
 144 {
 145         hammer_off_t tmp_offset;
 146         int error;
 147         int zone;
 148         int bytes;
 149         int cur;
 150
 151         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
 152                 return(0);
 153         error = 0;
 154
 155         /*
 156          * Reblock data.  Note that data embedded in a record is reblocked
 157          * by the record reblock code.
 158          */
 159         tmp_offset = elm->leaf.data_offset;
 160         zone = HAMMER_ZONE_DECODE(tmp_offset);          /* can be 0 */
 161         if ((zone == HAMMER_ZONE_SMALL_DATA_INDEX ||
 162              zone == HAMMER_ZONE_LARGE_DATA_INDEX) &&
 163             error == 0 && (reblock->head.flags & HAMMER_IOC_DO_DATA)) {
 164                 ++reblock->data_count;
 165                 reblock->data_byte_count += elm->leaf.data_len;
 166                 bytes = hammer_blockmap_getfree(cursor->trans->hmp, tmp_offset,
 167                                                 &cur, &error);
 168                 if (error == 0 && cur == 0 && bytes >= reblock->free_level) {
 169                         if (hammer_debug_general & 0x4000)
 170                                 kprintf("%6d ", bytes);
 171                         error = hammer_cursor_upgrade(cursor);
 172                         if (error == 0) {
 173                                 error = hammer_reblock_data(reblock,
 174                                                             cursor, elm);
 175                         }
 176                         if (error == 0) {
 177                                 ++reblock->data_moves;
 178                                 reblock->data_byte_moves += elm->leaf.data_len;
 179                         }
 180                 }
 181         }
 182
 183         /*
 184          * Reblock a B-Tree node.  Adjust elm to point at the parent's
 185          * leaf entry.
 186          */
 187         tmp_offset = cursor->node->node_offset;
 188         zone = HAMMER_ZONE_DECODE(tmp_offset);
 189         if (zone == HAMMER_ZONE_BTREE_INDEX && cursor->index == 0 &&
 190             error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) {
 191                 ++reblock->btree_count;
 192                 bytes = hammer_blockmap_getfree(cursor->trans->hmp, tmp_offset,
 193                                                 &cur, &error);
 194                 if (error == 0 && cur == 0 && bytes >= reblock->free_level) {
 195                         if (hammer_debug_general & 0x4000)
 196                                 kprintf("%6d ", bytes);
 197                         error = hammer_cursor_upgrade(cursor);
 198                         if (error == 0) {
 199                                 if (cursor->parent)
 200                                         elm = &cursor->parent->ondisk->elms[cursor->parent_index];
 201                                 else
 202                                         elm = NULL;
 203                                 error = hammer_reblock_node(reblock,
 204                                                             cursor, elm);
 205                         }
 206                         if (error == 0) {
 207                                 ++reblock->btree_moves;
 208                         }
 209                 }
 210         }
 211
 212         hammer_cursor_downgrade(cursor);
 213         return(error);
 214 }
 215
 216 /*
 217  * Reblock a record's data.  Both the B-Tree element and record pointers
 218  * to the data must be adjusted.
 219  */
 220 static int
 221 hammer_reblock_data(struct hammer_ioc_reblock *reblock,
 222                     hammer_cursor_t cursor, hammer_btree_elm_t elm)
 223 {
 224         struct hammer_buffer *data_buffer = NULL;
 225         hammer_off_t ndata_offset;
 226         int error;
 227         void *ndata;
 228
 229         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA |
 230                                              HAMMER_CURSOR_GET_LEAF);
 231         if (error)
 232                 return (error);
 233         ndata = hammer_alloc_data(cursor->trans, elm->leaf.data_len,
 234                                   &ndata_offset, &data_buffer, &error);
 235         if (error)
 236                 goto done;
 237
 238         /*
 239          * Move the data
 240          */
 241         hammer_modify_buffer(cursor->trans, data_buffer, NULL, 0);
 242         bcopy(cursor->data, ndata, elm->leaf.data_len);
 243         hammer_modify_buffer_done(data_buffer);
 244
 245         hammer_blockmap_free(cursor->trans,
 246                              elm->leaf.data_offset, elm->leaf.data_len);
 247
 248         hammer_modify_node(cursor->trans, cursor->node,
 249                            &elm->leaf.data_offset, sizeof(hammer_off_t));
 250         elm->leaf.data_offset = ndata_offset;
 251         hammer_modify_node_done(cursor->node);
 252
 253 done:
 254         if (data_buffer)
 255                 hammer_rel_buffer(data_buffer, 0);
 256         return (error);
 257 }
 258
 259 /*
 260  * Reblock a B-Tree (leaf) node.  The parent must be adjusted to point to
 261  * the new copy of the leaf node.  elm is a pointer to the parent element
 262  * pointing at cursor.node.
 263  *
 264  * XXX reblock internal nodes too.
 265  */
 266 static int
 267 hammer_reblock_node(struct hammer_ioc_reblock *reblock,
 268                     hammer_cursor_t cursor, hammer_btree_elm_t elm)
 269 {
 270         hammer_node_t onode;
 271         hammer_node_t nnode;
 272         int error;
 273
 274         onode = cursor->node;
 275         nnode = hammer_alloc_btree(cursor->trans, &error);
 276         hammer_lock_ex(&nnode->lock);
 277
 278         hammer_modify_node_noundo(cursor->trans, nnode);
 279
 280         if (nnode == NULL)
 281                 return (error);
 282
 283         /*
 284          * Move the node
 285          */
 286         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 287
 288         if (elm) {
 289                 /*
 290                  * We are not the root of the B-Tree
 291                  */
 292                 hammer_modify_node(cursor->trans, cursor->parent,
 293                                    &elm->internal.subtree_offset,
 294                                    sizeof(elm->internal.subtree_offset));
 295                 elm->internal.subtree_offset = nnode->node_offset;
 296                 hammer_modify_node_done(cursor->parent);
 297         } else {
 298                 /*
 299                  * We are the root of the B-Tree
 300                  */
 301                 hammer_volume_t volume;
 302
 303                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 304                 KKASSERT(error == 0);
 305
 306                 hammer_modify_volume_field(cursor->trans, volume,
 307                                            vol0_btree_root);
 308                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 309                 hammer_modify_volume_done(volume);
 310                 hammer_rel_volume(volume, 0);
 311         }
 312
 313         hammer_delete_node(cursor->trans, onode);
 314
 315         if (hammer_debug_general & 0x4000) {
 316                 kprintf("REBLOCK NODE %016llx -> %016llx\n",
 317                         onode->node_offset, nnode->node_offset);
 318         }
 319         hammer_modify_node_done(nnode);
 320
 321         cursor->node = nnode;
 322         hammer_rel_node(onode);
 323
 324         return (error);
 325 }
 326