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 1.15 2008/05/15 03:36:40 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                  * NOTE: cursor.node may have changed on return.
 104                  */
 105                 hammer_lock_ex(&trans->hmp->sync_lock);
 106                 error = hammer_reblock_helper(reblock, &cursor, elm);
 107                 hammer_unlock(&trans->hmp->sync_lock);
 108                 if (error == 0) {
 109                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
 110                         error = hammer_btree_iterate(&cursor);
 111                 }
 112
 113                 /*
 114                  * Bad hack for now, don't blow out the kernel's buffer
 115                  * cache.  NOTE: We still hold locks on the cursor, we
 116                  * cannot call the flusher synchronously.
 117                  */
 118                 if (trans->hmp->locked_dirty_count > hammer_limit_dirtybufs) {
 119                         hammer_flusher_async(trans->hmp);
 120                         tsleep(trans, 0, "hmrslo", hz / 10);
 121                 }
 122                 if (error == 0)
 123                         error = hammer_signal_check(trans->hmp);
 124         }
 125         if (error == ENOENT)
 126                 error = 0;
 127         hammer_done_cursor(&cursor);
 128         if (error == EDEADLK)
 129                 goto retry;
 130         if (error == EINTR) {
 131                 reblock->head.flags |= HAMMER_IOC_HEAD_INTR;
 132                 error = 0;
 133         }
 134         return(error);
 135 }
 136
 137 /*
 138  * Reblock the B-Tree (leaf) node, record, and/or data if necessary.
 139  *
 140  * XXX We have no visibility into internal B-Tree nodes at the moment,
 141  * only leaf nodes.
 142  */
 143 static int
 144 hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
 145                       hammer_cursor_t cursor, hammer_btree_elm_t elm)
 146 {
 147         hammer_off_t tmp_offset;
 148         int error;
 149         int zone;
 150         int bytes;
 151         int cur;
 152
 153         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
 154                 return(0);
 155         error = 0;
 156
 157         /*
 158          * Reblock data.  Note that data embedded in a record is reblocked
 159          * by the record reblock code.
 160          */
 161         tmp_offset = elm->leaf.data_offset;
 162         zone = HAMMER_ZONE_DECODE(tmp_offset);          /* can be 0 */
 163         if ((zone == HAMMER_ZONE_SMALL_DATA_INDEX ||
 164              zone == HAMMER_ZONE_LARGE_DATA_INDEX) &&
 165             error == 0 && (reblock->head.flags & HAMMER_IOC_DO_DATA)) {
 166                 ++reblock->data_count;
 167                 reblock->data_byte_count += elm->leaf.data_len;
 168                 bytes = hammer_blockmap_getfree(cursor->trans->hmp, tmp_offset,
 169                                                 &cur, &error);
 170                 if (error == 0 && cur == 0 && bytes >= reblock->free_level) {
 171                         if (hammer_debug_general & 0x4000)
 172                                 kprintf("%6d ", bytes);
 173                         error = hammer_cursor_upgrade(cursor);
 174                         if (error == 0) {
 175                                 error = hammer_reblock_data(reblock,
 176                                                             cursor, elm);
 177                         }
 178                         if (error == 0) {
 179                                 ++reblock->data_moves;
 180                                 reblock->data_byte_moves += elm->leaf.data_len;
 181                         }
 182                 }
 183         }
 184
 185         /*
 186          * Reblock a B-Tree node.  Adjust elm to point at the parent's
 187          * leaf entry.
 188          */
 189         tmp_offset = cursor->node->node_offset;
 190         zone = HAMMER_ZONE_DECODE(tmp_offset);
 191         if (zone == HAMMER_ZONE_BTREE_INDEX && cursor->index == 0 &&
 192             error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) {
 193                 ++reblock->btree_count;
 194                 bytes = hammer_blockmap_getfree(cursor->trans->hmp, tmp_offset,
 195                                                 &cur, &error);
 196                 if (error == 0 && cur == 0 && bytes >= reblock->free_level) {
 197                         if (hammer_debug_general & 0x4000)
 198                                 kprintf("%6d ", bytes);
 199                         error = hammer_cursor_upgrade(cursor);
 200                         if (error == 0) {
 201                                 if (cursor->parent)
 202                                         elm = &cursor->parent->ondisk->elms[cursor->parent_index];
 203                                 else
 204                                         elm = NULL;
 205                                 error = hammer_reblock_node(reblock,
 206                                                             cursor, elm);
 207                         }
 208                         if (error == 0) {
 209                                 ++reblock->btree_moves;
 210                         }
 211                 }
 212         }
 213
 214         hammer_cursor_downgrade(cursor);
 215         return(error);
 216 }
 217
 218 /*
 219  * Reblock a record's data.  Both the B-Tree element and record pointers
 220  * to the data must be adjusted.
 221  */
 222 static int
 223 hammer_reblock_data(struct hammer_ioc_reblock *reblock,
 224                     hammer_cursor_t cursor, hammer_btree_elm_t elm)
 225 {
 226         struct hammer_buffer *data_buffer = NULL;
 227         hammer_off_t ndata_offset;
 228         int error;
 229         void *ndata;
 230
 231         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA |
 232                                              HAMMER_CURSOR_GET_LEAF);
 233         if (error)
 234                 return (error);
 235         ndata = hammer_alloc_data(cursor->trans, elm->leaf.data_len,
 236                                   &ndata_offset, &data_buffer, &error);
 237         if (error)
 238                 goto done;
 239
 240         /*
 241          * Move the data
 242          */
 243         hammer_modify_buffer(cursor->trans, data_buffer, NULL, 0);
 244         bcopy(cursor->data, ndata, elm->leaf.data_len);
 245         hammer_modify_buffer_done(data_buffer);
 246
 247         hammer_blockmap_free(cursor->trans,
 248                              elm->leaf.data_offset, elm->leaf.data_len);
 249
 250         hammer_modify_node(cursor->trans, cursor->node,
 251                            &elm->leaf.data_offset, sizeof(hammer_off_t));
 252         elm->leaf.data_offset = ndata_offset;
 253         hammer_modify_node_done(cursor->node);
 254
 255 done:
 256         if (data_buffer)
 257                 hammer_rel_buffer(data_buffer, 0);
 258         return (error);
 259 }
 260
 261 /*
 262  * Reblock a B-Tree (leaf) node.  The parent must be adjusted to point to
 263  * the new copy of the leaf node.  elm is a pointer to the parent element
 264  * pointing at cursor.node.
 265  *
 266  * XXX reblock internal nodes too.
 267  */
 268 static int
 269 hammer_reblock_node(struct hammer_ioc_reblock *reblock,
 270                     hammer_cursor_t cursor, hammer_btree_elm_t elm)
 271 {
 272         hammer_node_t onode;
 273         hammer_node_t nnode;
 274         int error;
 275
 276         onode = cursor->node;
 277         nnode = hammer_alloc_btree(cursor->trans, &error);
 278
 279         if (nnode == NULL)
 280                 return (error);
 281
 282         /*
 283          * Move the node
 284          */
 285         hammer_lock_ex(&nnode->lock);
 286         hammer_modify_node_noundo(cursor->trans, nnode);
 287         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 288
 289         if (elm) {
 290                 /*
 291                  * We are not the root of the B-Tree
 292                  */
 293                 hammer_modify_node(cursor->trans, cursor->parent,
 294                                    &elm->internal.subtree_offset,
 295                                    sizeof(elm->internal.subtree_offset));
 296                 elm->internal.subtree_offset = nnode->node_offset;
 297                 hammer_modify_node_done(cursor->parent);
 298         } else {
 299                 /*
 300                  * We are the root of the B-Tree
 301                  */
 302                 hammer_volume_t volume;
 303
 304                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 305                 KKASSERT(error == 0);
 306
 307                 hammer_modify_volume_field(cursor->trans, volume,
 308                                            vol0_btree_root);
 309                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 310                 hammer_modify_volume_done(volume);
 311                 hammer_rel_volume(volume, 0);
 312         }
 313
 314         hammer_delete_node(cursor->trans, onode);
 315
 316         if (hammer_debug_general & 0x4000) {
 317                 kprintf("REBLOCK NODE %016llx -> %016llx\n",
 318                         onode->node_offset, nnode->node_offset);
 319         }
 320         hammer_modify_node_done(nnode);
 321         cursor->node = nnode;
 322
 323         hammer_unlock(&onode->lock);
 324         hammer_rel_node(onode);
 325
 326         return (error);
 327 }
 328