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 1.11 2008/07/01 17:30:42 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 prune_should_delete(struct hammer_ioc_prune *prune,
  48                                hammer_btree_leaf_elm_t elm);
  49 static void prune_check_nlinks(hammer_cursor_t cursor,
  50                                hammer_btree_leaf_elm_t elm);
  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_leaf_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 elm_array_size;
  63
  64         if (prune->nelms < 0 || prune->nelms > HAMMER_MAX_PRUNE_ELMS)
  65                 return(EINVAL);
  66         if ((prune->key_beg.localization | prune->key_end.localization) &
  67             HAMMER_LOCALIZE_PSEUDOFS_MASK) {
  68                 return(EINVAL);
  69         }
  70         if (prune->key_beg.localization > prune->key_end.localization)
  71                 return(EINVAL);
  72         if (prune->key_beg.localization == prune->key_end.localization) {
  73                 if (prune->key_beg.obj_id > prune->key_end.obj_id)
  74                         return(EINVAL);
  75                 /* key-space limitations - no check needed */
  76         }
  77         if ((prune->head.flags & HAMMER_IOC_PRUNE_ALL) && prune->nelms)
  78                 return(EINVAL);
  79
  80         prune->key_cur.localization = prune->key_end.localization +
  81                                       ip->obj_localization;
  82         prune->key_cur.obj_id = prune->key_end.obj_id;
  83         prune->key_cur.key = HAMMER_MAX_KEY;
  84
  85         /*
  86          * Copy element array from userland
  87          */
  88         elm_array_size = sizeof(*copy_elms) * prune->nelms;
  89         user_elms = prune->elms;
  90         copy_elms = kmalloc(elm_array_size, M_TEMP, M_WAITOK);
  91         if ((error = copyin(user_elms, copy_elms, elm_array_size)) != 0)
  92                 goto failed;
  93         prune->elms = copy_elms;
  94
  95         /*
  96          * Scan backwards.  Retries typically occur if a deadlock is detected.
  97          */
  98 retry:
  99         error = hammer_init_cursor(trans, &cursor, NULL, NULL);
 100         if (error) {
 101                 hammer_done_cursor(&cursor);
 102                 goto failed;
 103         }
 104         cursor.key_beg.localization = prune->key_beg.localization +
 105                                       ip->obj_localization;
 106         cursor.key_beg.obj_id = prune->key_beg.obj_id;
 107         cursor.key_beg.key = HAMMER_MIN_KEY;
 108         cursor.key_beg.create_tid = 1;
 109         cursor.key_beg.delete_tid = 0;
 110         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
 111         cursor.key_beg.obj_type = 0;
 112
 113         cursor.key_end.localization = prune->key_cur.localization;
 114         cursor.key_end.obj_id = prune->key_cur.obj_id;
 115         cursor.key_end.key = prune->key_cur.key;
 116         cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
 117         cursor.key_end.delete_tid = 0;
 118         cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
 119         cursor.key_end.obj_type = 0;
 120
 121         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
 122         cursor.flags |= HAMMER_CURSOR_BACKEND;
 123
 124         /*
 125          * This flag allows the B-Tree code to clean up loose ends.
 126          */
 127         cursor.flags |= HAMMER_CURSOR_PRUNING;
 128
 129         hammer_sync_lock_sh(trans);
 130         error = hammer_btree_last(&cursor);
 131
 132         while (error == 0) {
 133                 /*
 134                  * Check for work
 135                  */
 136                 elm = &cursor.node->ondisk->elms[cursor.index].leaf;
 137                 prune->key_cur = elm->base;
 138
 139                 /*
 140                  * Yield to more important tasks
 141                  */
 142                 if ((error = hammer_signal_check(trans->hmp)) != 0)
 143                         break;
 144                 if (trans->hmp->sync_lock.wanted) {
 145                         hammer_sync_unlock(trans);
 146                         tsleep(trans, 0, "hmrslo", hz / 10);
 147                         hammer_sync_lock_sh(trans);
 148                 }
 149                 if (hammer_flusher_meta_limit(trans->hmp) ||
 150                     hammer_flusher_undo_exhausted(trans, 2)) {
 151                         error = EWOULDBLOCK;
 152                         break;
 153                 }
 154
 155                 if (prune->stat_oldest_tid > elm->base.create_tid)
 156                         prune->stat_oldest_tid = elm->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 (prune_should_delete(prune, elm)) {
 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 {
 198                         /*
 199                          * Nothing to delete, but we may have to check other
 200                          * things.
 201                          */
 202                         prune_check_nlinks(&cursor, elm);
 203                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
 204                         if (hammer_debug_general & 0x0100) {
 205                                 kprintf("check %016llx %016llx: SKIP\n",
 206                                         elm->base.obj_id, elm->base.key);
 207                         }
 208                 }
 209                 ++prune->stat_scanrecords;
 210                 error = hammer_btree_iterate_reverse(&cursor);
 211         }
 212         hammer_sync_unlock(trans);
 213         if (error == ENOENT)
 214                 error = 0;
 215         hammer_done_cursor(&cursor);
 216         if (error == EWOULDBLOCK) {
 217                 hammer_flusher_sync(trans->hmp);
 218                 goto retry;
 219         }
 220         if (error == EDEADLK)
 221                 goto retry;
 222         if (error == EINTR) {
 223                 prune->head.flags |= HAMMER_IOC_HEAD_INTR;
 224                 error = 0;
 225         }
 226 failed:
 227         prune->key_cur.localization &= HAMMER_LOCALIZE_MASK;
 228         prune->elms = user_elms;
 229         kfree(copy_elms, M_TEMP);
 230         return(error);
 231 }
 232
 233 /*
 234  * Check pruning list.  The list must be sorted in descending order.
 235  *
 236  * Return non-zero if the record should be deleted.
 237  */
 238 static int
 239 prune_should_delete(struct hammer_ioc_prune *prune, hammer_btree_leaf_elm_t elm)
 240 {
 241         struct hammer_ioc_prune_elm *scan;
 242         int i;
 243
 244         /*
 245          * If pruning everything remove all records with a non-zero
 246          * delete_tid.
 247          */
 248         if (prune->head.flags & HAMMER_IOC_PRUNE_ALL) {
 249                 if (elm->base.delete_tid != 0)
 250                         return(1);
 251                 return(0);
 252         }
 253
 254         for (i = 0; i < prune->nelms; ++i) {
 255                 scan = &prune->elms[i];
 256
 257                 /*
 258                  * Check for loop termination.
 259                  */
 260                 if (elm->base.create_tid >= scan->end_tid ||
 261                     elm->base.delete_tid > scan->end_tid) {
 262                         break;
 263                 }
 264
 265                 /*
 266                  * Determine if we can delete the record.
 267                  */
 268                 if (elm->base.delete_tid &&
 269                     elm->base.create_tid >= scan->beg_tid &&
 270                     elm->base.delete_tid <= scan->end_tid &&
 271                     (elm->base.create_tid - scan->beg_tid) / scan->mod_tid ==
 272                     (elm->base.delete_tid - scan->beg_tid) / scan->mod_tid) {
 273                         return(1);
 274                 }
 275         }
 276         return(0);
 277 }
 278
 279 /*
 280  * Dangling inodes can occur if processes are holding open descriptors on
 281  * deleted files as-of when a machine crashes.  When we find one simply
 282  * acquire the inode and release it.  The inode handling code will then
 283  * do the right thing.
 284  */
 285 static
 286 void
 287 prune_check_nlinks(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm)
 288 {
 289         hammer_inode_t ip;
 290         int error;
 291
 292         if (elm->base.rec_type != HAMMER_RECTYPE_INODE)
 293                 return;
 294         if (elm->base.delete_tid != 0)
 295                 return;
 296         if (hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA))
 297                 return;
 298         if (cursor->data->inode.nlinks)
 299                 return;
 300         hammer_cursor_downgrade(cursor);
 301         hammer_sync_unlock(cursor->trans);
 302         ip = hammer_get_inode(cursor->trans, NULL, elm->base.obj_id,
 303                       HAMMER_MAX_TID,
 304                       elm->base.localization & HAMMER_LOCALIZE_PSEUDOFS_MASK,
 305                       0, &error);
 306         if (ip) {
 307                 kprintf("pruning disconnected inode %016llx\n",
 308                         elm->base.obj_id);
 309                 hammer_rel_inode(ip, 0);
 310         } else {
 311                 kprintf("unable to prune disconnected inode %016llx\n",
 312                         elm->base.obj_id);
 313         }
 314         hammer_sync_lock_sh(cursor->trans);
 315 }
 316
 317 #if 0
 318
 319 /*
 320  * NOTE: THIS CODE HAS BEEN REMOVED!  Pruning no longer attempts to realign
 321  *       adjacent records because it seriously interferes with every
 322  *       mirroring algorithm I could come up with.
 323  *
 324  *       This means that historical accesses beyond the first snapshot
 325  *       softlink should be on snapshot boundaries only.  Historical
 326  *       accesses from "now" to the first snapshot softlink continue to
 327  *       be fine-grained.
 328  *
 329  * NOTE: It also looks like there's a bug in the removed code.  It is believed
 330  *       that create_tid can sometimes get set to 0xffffffffffffffff.  Just as
 331  *       well we no longer try to do this fancy shit.  Probably the attempt to
 332  *       correct the rhb is blowing up the cursor's indexing or addressing mapping.
 333  *
 334  * Align the record to cover any gaps created through the deletion of
 335  * records within the pruning space.  If we were to just delete the records
 336  * there would be gaps which in turn would cause a snapshot that is NOT on
 337  * a pruning boundary to appear corrupt to the user.  Forcing alignment
 338  * of the create_tid and delete_tid for retained records 'reconnects'
 339  * the previously contiguous space, making it contiguous again after the
 340  * deletions.
 341  *
 342  * The use of a reverse iteration allows us to safely align the records and
 343  * related elements without creating temporary overlaps.  XXX we should
 344  * add ordering dependancies for record buffers to guarantee consistency
 345  * during recovery.
 346  */
 347 static int
 348 realign_prune(struct hammer_ioc_prune *prune,
 349               hammer_cursor_t cursor, int realign_cre, int realign_del)
 350 {
 351         struct hammer_ioc_prune_elm *scan;
 352         hammer_btree_elm_t elm;
 353         hammer_tid_t delta;
 354         hammer_tid_t tid;
 355         int error;
 356
 357         hammer_cursor_downgrade(cursor);
 358
 359         elm = &cursor->node->ondisk->elms[cursor->index];
 360         ++prune->stat_realignments;
 361
 362         /*
 363          * Align the create_tid.  By doing a reverse iteration we guarantee
 364          * that all records after our current record have already been
 365          * aligned, allowing us to safely correct the right-hand-boundary
 366          * (because no record to our right is otherwise exactly matching
 367          * will have a create_tid to the left of our aligned create_tid).
 368          */
 369         error = 0;
 370         if (realign_cre >= 0) {
 371                 scan = &prune->elms[realign_cre];
 372
 373                 delta = (elm->leaf.base.create_tid - scan->beg_tid) %
 374                         scan->mod_tid;
 375                 if (delta) {
 376                         tid = elm->leaf.base.create_tid - delta + scan->mod_tid;
 377
 378                         /* can EDEADLK */
 379                         error = hammer_btree_correct_rhb(cursor, tid + 1);
 380                         if (error == 0) {
 381                                 error = hammer_btree_extract(cursor,
 382                                                      HAMMER_CURSOR_GET_LEAF);
 383                         }
 384                         if (error == 0) {
 385                                 /* can EDEADLK */
 386                                 error = hammer_cursor_upgrade(cursor);
 387                         }
 388                         if (error == 0) {
 389                                 hammer_modify_node(cursor->trans, cursor->node,
 390                                             &elm->leaf.base.create_tid,
 391                                             sizeof(elm->leaf.base.create_tid));
 392                                 elm->leaf.base.create_tid = tid;
 393                                 hammer_modify_node_done(cursor->node);
 394                         }
 395                 }
 396         }
 397
 398         /*
 399          * Align the delete_tid.  This only occurs if the record is historical
 400          * was deleted at some point.  Realigning the delete_tid does not
 401          * move the record within the B-Tree but may cause it to temporarily
 402          * overlap a record that has not yet been pruned.
 403          */
 404         if (error == 0 && realign_del >= 0) {
 405                 scan = &prune->elms[realign_del];
 406
 407                 delta = (elm->leaf.base.delete_tid - scan->beg_tid) %
 408                         scan->mod_tid;
 409                 if (delta) {
 410                         error = hammer_btree_extract(cursor,
 411                                                      HAMMER_CURSOR_GET_LEAF);
 412                         if (error == 0) {
 413                                 hammer_modify_node(cursor->trans, cursor->node,
 414                                             &elm->leaf.base.delete_tid,
 415                                             sizeof(elm->leaf.base.delete_tid));
 416                                 elm->leaf.base.delete_tid =
 417                                             elm->leaf.base.delete_tid -
 418                                             delta + scan->mod_tid;
 419                                 hammer_modify_node_done(cursor->node);
 420                         }
 421                 }
 422         }
 423         return (error);
 424 }
 425
 426 #endif