sys/vfs/hammer/hammer_prune.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_prune.c,v 1.6 2008/06/10 00:40:31 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38
  39 /*
  40  * Iterate through the specified range of object ids and remove any
  41  * deleted records that fall entirely within a prune modulo.
  42  *
  43  * A reverse iteration is used to prevent overlapping records from being
  44  * created during the iteration due to alignments.  This also allows us
  45  * to adjust alignments without blowing up the B-Tree.
  46  */
  47 static int check_prune(struct hammer_ioc_prune *prune, hammer_btree_elm_t elm,
  48                         int *realign_cre, int *realign_del);
  49 static int realign_prune(struct hammer_ioc_prune *prune, hammer_cursor_t cursor,
  50                         int realign_cre, int realign_del);
  51
  52 int
  53 hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
  54                  struct hammer_ioc_prune *prune)
  55 {
  56         struct hammer_cursor cursor;
  57         hammer_btree_elm_t elm;
  58         struct hammer_ioc_prune_elm *copy_elms;
  59         struct hammer_ioc_prune_elm *user_elms;
  60         int error;
  61         int isdir;
  62         int realign_cre;
  63         int realign_del;
  64         int elm_array_size;
  65
  66         if (prune->nelms < 0 || prune->nelms > HAMMER_MAX_PRUNE_ELMS)
  67                 return(EINVAL);
  68         if (prune->beg_localization > prune->end_localization)
  69                 return(EINVAL);
  70         if (prune->beg_localization == prune->end_localization) {
  71                 if (prune->beg_obj_id > prune->end_obj_id)
  72                         return(EINVAL);
  73                 /* key-space limitations - no check needed */
  74         }
  75         if ((prune->head.flags & HAMMER_IOC_PRUNE_ALL) && prune->nelms)
  76                 return(EINVAL);
  77
  78         prune->cur_localization = prune->end_localization;
  79         prune->cur_obj_id = prune->end_obj_id;
  80         prune->cur_key = HAMMER_MAX_KEY;
  81
  82         /*
  83          * Copy element array from userland
  84          */
  85         elm_array_size = sizeof(*copy_elms) * prune->nelms;
  86         user_elms = prune->elms;
  87         copy_elms = kmalloc(elm_array_size, M_TEMP, M_WAITOK);
  88         if ((error = copyin(user_elms, copy_elms, elm_array_size)) != 0)
  89                 goto failed;
  90         prune->elms = copy_elms;
  91
  92         /*
  93          * Scan backwards.  Retries typically occur if a deadlock is detected.
  94          */
  95 retry:
  96         error = hammer_init_cursor(trans, &cursor, NULL, NULL);
  97         if (error) {
  98                 hammer_done_cursor(&cursor);
  99                 goto failed;
 100         }
 101         cursor.key_beg.localization = prune->beg_localization;
 102         cursor.key_beg.obj_id = prune->beg_obj_id;
 103         cursor.key_beg.key = HAMMER_MIN_KEY;
 104         cursor.key_beg.create_tid = 1;
 105         cursor.key_beg.delete_tid = 0;
 106         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
 107         cursor.key_beg.obj_type = 0;
 108
 109         cursor.key_end.localization = prune->cur_localization;
 110         cursor.key_end.obj_id = prune->cur_obj_id;
 111         cursor.key_end.key = prune->cur_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 B-Tree code to clean up loose ends.
 122          */
 123         cursor.flags |= HAMMER_CURSOR_PRUNING;
 124
 125         hammer_sync_lock_sh(trans);
 126         error = hammer_btree_last(&cursor);
 127
 128         while (error == 0) {
 129                 /*
 130                  * Yield to more important tasks
 131                  */
 132                 if ((error = hammer_signal_check(trans->hmp)) != 0)
 133                         break;
 134                 if (trans->hmp->sync_lock.wanted) {
 135                         hammer_sync_unlock(trans);
 136                         tsleep(trans, 0, "hmrslo", hz / 10);
 137                         hammer_sync_lock_sh(trans);
 138                 }
 139                 if (trans->hmp->locked_dirty_count +
 140                     trans->hmp->io_running_count > hammer_limit_dirtybufs) {
 141                         hammer_sync_unlock(trans);
 142                         hammer_flusher_async(trans->hmp);
 143                         tsleep(trans, 0, "hmrslo", hz / 10);
 144                         hammer_sync_lock_sh(trans);
 145                 }
 146
 147                 /*
 148                  * Check for work
 149                  */
 150                 elm = &cursor.node->ondisk->elms[cursor.index];
 151                 prune->cur_localization = elm->base.localization;
 152                 prune->cur_obj_id = elm->base.obj_id;
 153                 prune->cur_key = elm->base.key;
 154
 155                 if (prune->stat_oldest_tid > elm->leaf.base.create_tid)
 156                         prune->stat_oldest_tid = elm->leaf.base.create_tid;
 157
 158                 if (hammer_debug_general & 0x0200) {
 159                         kprintf("check %016llx %016llx cre=%016llx del=%016llx\n",
 160                                         elm->base.obj_id,
 161                                         elm->base.key,
 162                                         elm->base.create_tid,
 163                                         elm->base.delete_tid);
 164                 }
 165
 166                 if (check_prune(prune, elm, &realign_cre, &realign_del) == 0) {
 167                         if (hammer_debug_general & 0x0200) {
 168                                 kprintf("check %016llx %016llx: DELETE\n",
 169                                         elm->base.obj_id, elm->base.key);
 170                         }
 171
 172                         /*
 173                          * NOTE: This can return EDEADLK
 174                          *
 175                          * Acquiring the sync lock guarantees that the
 176                          * operation will not cross a synchronization
 177                          * boundary (see the flusher).
 178                          */
 179                         isdir = (elm->base.rec_type == HAMMER_RECTYPE_DIRENTRY);
 180
 181                         error = hammer_delete_at_cursor(&cursor,
 182                                                         &prune->stat_bytes);
 183                         if (error)
 184                                 break;
 185
 186                         if (isdir)
 187                                 ++prune->stat_dirrecords;
 188                         else
 189                                 ++prune->stat_rawrecords;
 190
 191                         /*
 192                          * The current record might now be the one after
 193                          * the one we deleted, set ATEDISK to force us
 194                          * to skip it (since we are iterating backwards).
 195                          */
 196                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
 197                 } else if (realign_cre >= 0 || realign_del >= 0) {
 198                         error = realign_prune(prune, &cursor,
 199                                               realign_cre, realign_del);
 200                         if (error == 0) {
 201                                 cursor.flags |= HAMMER_CURSOR_ATEDISK;
 202                                 if (hammer_debug_general & 0x0200) {
 203                                         kprintf("check %016llx %016llx: "
 204                                                 "REALIGN\n",
 205                                                 elm->base.obj_id,
 206                                                 elm->base.key);
 207                                 }
 208                         }
 209                 } else {
 210                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
 211                         if (hammer_debug_general & 0x0100) {
 212                                 kprintf("check %016llx %016llx: SKIP\n",
 213                                         elm->base.obj_id, elm->base.key);
 214                         }
 215                 }
 216                 ++prune->stat_scanrecords;
 217                 error = hammer_btree_iterate_reverse(&cursor);
 218         }
 219         hammer_sync_unlock(trans);
 220         if (error == ENOENT)
 221                 error = 0;
 222         hammer_done_cursor(&cursor);
 223         if (error == EDEADLK)
 224                 goto retry;
 225         if (error == EINTR) {
 226                 prune->head.flags |= HAMMER_IOC_HEAD_INTR;
 227                 error = 0;
 228         }
 229 failed:
 230         prune->elms = user_elms;
 231         kfree(copy_elms, M_TEMP);
 232         return(error);
 233 }
 234
 235 /*
 236  * Check pruning list.  The list must be sorted in descending order.
 237  */
 238 static int
 239 check_prune(struct hammer_ioc_prune *prune, hammer_btree_elm_t elm,
 240             int *realign_cre, int *realign_del)
 241 {
 242         struct hammer_ioc_prune_elm *scan;
 243         int i;
 244
 245         *realign_cre = -1;
 246         *realign_del = -1;
 247
 248         /*
 249          * If pruning everything remove all records with a non-zero
 250          * delete_tid.
 251          */
 252         if (prune->head.flags & HAMMER_IOC_PRUNE_ALL) {
 253                 if (elm->base.delete_tid != 0)
 254                         return(0);
 255                 return(-1);
 256         }
 257
 258         for (i = 0; i < prune->nelms; ++i) {
 259                 scan = &prune->elms[i];
 260
 261                 /*
 262                  * Locate the scan index covering the create and delete TIDs.
 263                  */
 264                 if (*realign_cre < 0 &&
 265                     elm->base.create_tid >= scan->beg_tid &&
 266                     elm->base.create_tid < scan->end_tid) {
 267                         *realign_cre = i;
 268                 }
 269                 if (*realign_del < 0 && elm->base.delete_tid &&
 270                     elm->base.delete_tid > scan->beg_tid &&
 271                     elm->base.delete_tid <= scan->end_tid) {
 272                         *realign_del = i;
 273                 }
 274
 275                 /*
 276                  * Now check for loop termination.
 277                  */
 278                 if (elm->base.create_tid >= scan->end_tid ||
 279                     elm->base.delete_tid > scan->end_tid) {
 280                         break;
 281                 }
 282
 283                 /*
 284                  * Now determine if we can delete the record.
 285                  */
 286                 if (elm->base.delete_tid &&
 287                     elm->base.create_tid >= scan->beg_tid &&
 288                     elm->base.delete_tid <= scan->end_tid &&
 289                     (elm->base.create_tid - scan->beg_tid) / scan->mod_tid ==
 290                     (elm->base.delete_tid - scan->beg_tid) / scan->mod_tid) {
 291                         return(0);
 292                 }
 293         }
 294         return(-1);
 295 }
 296
 297 /*
 298  * Align the record to cover any gaps created through the deletion of
 299  * records within the pruning space.  If we were to just delete the records
 300  * there would be gaps which in turn would cause a snapshot that is NOT on
 301  * a pruning boundary to appear corrupt to the user.  Forcing alignment
 302  * of the create_tid and delete_tid for retained records 'reconnects'
 303  * the previously contiguous space, making it contiguous again after the
 304  * deletions.
 305  *
 306  * The use of a reverse iteration allows us to safely align the records and
 307  * related elements without creating temporary overlaps.  XXX we should
 308  * add ordering dependancies for record buffers to guarantee consistency
 309  * during recovery.
 310  */
 311 static int
 312 realign_prune(struct hammer_ioc_prune *prune,
 313               hammer_cursor_t cursor, int realign_cre, int realign_del)
 314 {
 315         struct hammer_ioc_prune_elm *scan;
 316         hammer_btree_elm_t elm;
 317         hammer_tid_t delta;
 318         hammer_tid_t tid;
 319         int error;
 320
 321         hammer_cursor_downgrade(cursor);
 322
 323         elm = &cursor->node->ondisk->elms[cursor->index];
 324         ++prune->stat_realignments;
 325
 326         /*
 327          * Align the create_tid.  By doing a reverse iteration we guarantee
 328          * that all records after our current record have already been
 329          * aligned, allowing us to safely correct the right-hand-boundary
 330          * (because no record to our right is otherwise exactly matching
 331          * will have a create_tid to the left of our aligned create_tid).
 332          */
 333         error = 0;
 334         if (realign_cre >= 0) {
 335                 scan = &prune->elms[realign_cre];
 336
 337                 delta = (elm->leaf.base.create_tid - scan->beg_tid) %
 338                         scan->mod_tid;
 339                 if (delta) {
 340                         tid = elm->leaf.base.create_tid - delta + scan->mod_tid;
 341
 342                         /* can EDEADLK */
 343                         error = hammer_btree_correct_rhb(cursor, tid + 1);
 344                         if (error == 0) {
 345                                 error = hammer_btree_extract(cursor,
 346                                                      HAMMER_CURSOR_GET_LEAF);
 347                         }
 348                         if (error == 0) {
 349                                 /* can EDEADLK */
 350                                 error = hammer_cursor_upgrade(cursor);
 351                         }
 352                         if (error == 0) {
 353                                 hammer_modify_node(cursor->trans, cursor->node,
 354                                             &elm->leaf.base.create_tid,
 355                                             sizeof(elm->leaf.base.create_tid));
 356                                 elm->leaf.base.create_tid = tid;
 357                                 hammer_modify_node_done(cursor->node);
 358                         }
 359                 }
 360         }
 361
 362         /*
 363          * Align the delete_tid.  This only occurs if the record is historical
 364          * was deleted at some point.  Realigning the delete_tid does not
 365          * move the record within the B-Tree but may cause it to temporarily
 366          * overlap a record that has not yet been pruned.
 367          */
 368         if (error == 0 && realign_del >= 0) {
 369                 scan = &prune->elms[realign_del];
 370
 371                 delta = (elm->leaf.base.delete_tid - scan->beg_tid) %
 372                         scan->mod_tid;
 373                 if (delta) {
 374                         error = hammer_btree_extract(cursor,
 375                                                      HAMMER_CURSOR_GET_LEAF);
 376                         if (error == 0) {
 377                                 hammer_modify_node(cursor->trans, cursor->node,
 378                                             &elm->leaf.base.delete_tid,
 379                                             sizeof(elm->leaf.base.delete_tid));
 380                                 elm->leaf.base.delete_tid =
 381                                             elm->leaf.base.delete_tid -
 382                                             delta + scan->mod_tid;
 383                                 hammer_modify_node_done(cursor->node);
 384                         }
 385                 }
 386         }
 387         return (error);
 388 }
 389