sys/vfs/hammer/hammer_ioctl.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_ioctl.c,v 1.6 2008/03/19 20:18:17 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38
  39 static int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
  40                                 struct hammer_ioc_prune *prune);
  41 static int hammer_ioc_gethistory(hammer_transaction_t trans, hammer_inode_t ip,
  42                                 struct hammer_ioc_history *hist);
  43
  44 int
  45 hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
  46              struct ucred *cred)
  47 {
  48         struct hammer_transaction trans;
  49         int error;
  50
  51         error = suser_cred(cred, PRISON_ROOT);
  52
  53         hammer_start_transaction(&trans, ip->hmp);
  54
  55         switch(com) {
  56         case HAMMERIOC_PRUNE:
  57                 if (error == 0) {
  58                         error = hammer_ioc_prune(&trans, ip,
  59                                         (struct hammer_ioc_prune *)data);
  60                 }
  61                 break;
  62         case HAMMERIOC_GETHISTORY:
  63                 error = hammer_ioc_gethistory(&trans, ip,
  64                                         (struct hammer_ioc_history *)data);
  65                 break;
  66         case HAMMERIOC_REBLOCK:
  67                 error = hammer_ioc_reblock(&trans, ip,
  68                                         (struct hammer_ioc_reblock *)data);
  69                 break;
  70         default:
  71                 error = EOPNOTSUPP;
  72                 break;
  73         }
  74         hammer_commit_transaction(&trans);
  75         return (error);
  76 }
  77
  78 /*
  79  * Iterate through the specified range of object ids and remove any
  80  * deleted records that fall entirely within a prune modulo.
  81  *
  82  * A reverse iteration is used to prevent overlapping records from being
  83  * created during the iteration due to alignments.  This also allows us
  84  * to adjust alignments without blowing up the B-Tree.
  85  */
  86 static int check_prune(struct hammer_ioc_prune *prune, hammer_btree_elm_t elm,
  87                         int *realign_cre, int *realign_del);
  88 static int realign_prune(struct hammer_ioc_prune *prune, hammer_cursor_t cursor,
  89                         int realign_cre, int realign_del);
  90
  91 static int
  92 hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
  93                  struct hammer_ioc_prune *prune)
  94 {
  95         struct hammer_cursor cursor;
  96         hammer_btree_elm_t elm;
  97         int error;
  98         int isdir;
  99         int realign_cre;
 100         int realign_del;
 101
 102         if (prune->nelms < 0 || prune->nelms > HAMMER_MAX_PRUNE_ELMS)
 103                 return(EINVAL);
 104         if (prune->beg_obj_id >= prune->end_obj_id)
 105                 return(EINVAL);
 106         if ((prune->flags & HAMMER_IOC_PRUNE_ALL) && prune->nelms)
 107                 return(EINVAL);
 108
 109 retry:
 110         error = hammer_init_cursor(trans, &cursor, NULL);
 111         if (error) {
 112                 hammer_done_cursor(&cursor);
 113                 return(error);
 114         }
 115         cursor.key_beg.obj_id = prune->beg_obj_id;
 116         cursor.key_beg.key = HAMMER_MIN_KEY;
 117         cursor.key_beg.create_tid = 1;
 118         cursor.key_beg.delete_tid = 0;
 119         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
 120         cursor.key_beg.obj_type = 0;
 121
 122         cursor.key_end.obj_id = prune->cur_obj_id;
 123         cursor.key_end.key = prune->cur_key;
 124         cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
 125         cursor.key_end.delete_tid = 0;
 126         cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
 127         cursor.key_end.obj_type = 0;
 128
 129         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
 130
 131         error = hammer_btree_last(&cursor);
 132         while (error == 0) {
 133                 elm = &cursor.node->ondisk->elms[cursor.index];
 134                 prune->cur_obj_id = elm->base.obj_id;
 135                 prune->cur_key = elm->base.key;
 136
 137                 if (check_prune(prune, elm, &realign_cre, &realign_del) == 0) {
 138                         if (hammer_debug_general & 0x0200) {
 139                                 kprintf("check %016llx %016llx: DELETE\n",
 140                                         elm->base.obj_id, elm->base.key);
 141                         }
 142
 143                         /*
 144                          * NOTE: This can return EDEADLK
 145                          */
 146                         isdir = (elm->base.rec_type == HAMMER_RECTYPE_DIRENTRY);
 147
 148                         error = hammer_delete_at_cursor(&cursor,
 149                                                         &prune->stat_bytes);
 150                         if (error)
 151                                 break;
 152
 153                         if (isdir)
 154                                 ++prune->stat_dirrecords;
 155                         else
 156                                 ++prune->stat_rawrecords;
 157                 } else if (realign_cre >= 0 || realign_del >= 0) {
 158                         error = realign_prune(prune, &cursor,
 159                                               realign_cre, realign_del);
 160                         if (error == 0) {
 161                                 cursor.flags |= HAMMER_CURSOR_ATEDISK;
 162                                 if (hammer_debug_general & 0x0200) {
 163                                         kprintf("check %016llx %016llx: "
 164                                                 "REALIGN\n",
 165                                                 elm->base.obj_id,
 166                                                 elm->base.key);
 167                                 }
 168                         }
 169                 } else {
 170                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
 171                         if (hammer_debug_general & 0x0100) {
 172                                 kprintf("check %016llx %016llx: SKIP\n",
 173                                         elm->base.obj_id, elm->base.key);
 174                         }
 175                 }
 176                 if (error == 0)
 177                         error = hammer_btree_iterate_reverse(&cursor);
 178         }
 179         if (error == ENOENT)
 180                 error = 0;
 181         hammer_done_cursor(&cursor);
 182         if (error == EDEADLK)
 183                 goto retry;
 184         return(error);
 185 }
 186
 187 /*
 188  * Check pruning list.  The list must be sorted in descending order.
 189  */
 190 static int
 191 check_prune(struct hammer_ioc_prune *prune, hammer_btree_elm_t elm,
 192             int *realign_cre, int *realign_del)
 193 {
 194         struct hammer_ioc_prune_elm *scan;
 195         int i;
 196
 197         *realign_cre = -1;
 198         *realign_del = -1;
 199
 200         /*
 201          * If pruning everything remove all records with a non-zero
 202          * delete_tid.
 203          */
 204         if (prune->flags & HAMMER_IOC_PRUNE_ALL) {
 205                 if (elm->base.delete_tid != 0)
 206                         return(0);
 207                 return(-1);
 208         }
 209
 210         for (i = 0; i < prune->nelms; ++i) {
 211                 scan = &prune->elms[i];
 212
 213                 /*
 214                  * Locate the scan index covering the create and delete TIDs.
 215                  */
 216                 if (*realign_cre < 0 &&
 217                     elm->base.create_tid >= scan->beg_tid &&
 218                     elm->base.create_tid < scan->end_tid) {
 219                         *realign_cre = i;
 220                 }
 221                 if (*realign_del < 0 && elm->base.delete_tid &&
 222                     elm->base.delete_tid > scan->beg_tid &&
 223                     elm->base.delete_tid <= scan->end_tid) {
 224                         *realign_del = i;
 225                 }
 226
 227                 /*
 228                  * Now check for loop termination.
 229                  */
 230                 if (elm->base.create_tid >= scan->end_tid ||
 231                     elm->base.delete_tid > scan->end_tid) {
 232                         break;
 233                 }
 234
 235                 /*
 236                  * Now determine if we can delete the record.
 237                  */
 238                 if (elm->base.delete_tid &&
 239                     elm->base.create_tid >= scan->beg_tid &&
 240                     elm->base.delete_tid <= scan->end_tid &&
 241                     elm->base.create_tid / scan->mod_tid ==
 242                     elm->base.delete_tid / scan->mod_tid) {
 243                         return(0);
 244                 }
 245         }
 246         return(-1);
 247 }
 248
 249 /*
 250  * Align the record to cover any gaps created through the deletion of
 251  * records within the pruning space.  If we were to just delete the records
 252  * there would be gaps which in turn would cause a snapshot that is NOT on
 253  * a pruning boundary to appear corrupt to the user.  Forcing alignment
 254  * of the create_tid and delete_tid for retained records 'reconnects'
 255  * the previously contiguous space, making it contiguous again after the
 256  * deletions.
 257  *
 258  * The use of a reverse iteration allows us to safely align the records and
 259  * related elements without creating temporary overlaps.  XXX we should
 260  * add ordering dependancies for record buffers to guarantee consistency
 261  * during recovery.
 262  */
 263 static int
 264 realign_prune(struct hammer_ioc_prune *prune,
 265               hammer_cursor_t cursor, int realign_cre, int realign_del)
 266 {
 267         hammer_btree_elm_t elm;
 268         hammer_tid_t delta;
 269         hammer_tid_t mod;
 270         hammer_tid_t tid;
 271         int error;
 272
 273         hammer_cursor_downgrade(cursor);
 274
 275         elm = &cursor->node->ondisk->elms[cursor->index];
 276         ++prune->stat_realignments;
 277
 278         /*
 279          * Align the create_tid.  By doing a reverse iteration we guarantee
 280          * that all records after our current record have already been
 281          * aligned, allowing us to safely correct the right-hand-boundary
 282          * (because no record to our right if otherwise exactly matching
 283          * will have a create_tid to the left of our aligned create_tid).
 284          *
 285          * Ordering is important here XXX but disk write ordering for
 286          * inter-cluster corrections is not currently guaranteed.
 287          */
 288         error = 0;
 289         if (realign_cre >= 0) {
 290                 mod = prune->elms[realign_cre].mod_tid;
 291                 delta = elm->leaf.base.create_tid % mod;
 292                 if (delta) {
 293                         tid = elm->leaf.base.create_tid - delta + mod;
 294
 295                         /* can EDEADLK */
 296                         error = hammer_btree_correct_rhb(cursor, tid + 1);
 297                         if (error == 0) {
 298                                 error = hammer_btree_extract(cursor,
 299                                                      HAMMER_CURSOR_GET_RECORD);
 300                         }
 301                         if (error == 0) {
 302                                 /* can EDEADLK */
 303                                 error = hammer_cursor_upgrade(cursor);
 304                         }
 305                         if (error == 0) {
 306                                 hammer_modify_buffer(cursor->trans,
 307                                                      cursor->record_buffer,
 308                                                      NULL, 0);
 309                                 cursor->record->base.base.create_tid = tid;
 310                                 hammer_modify_node(cursor->trans, cursor->node,
 311                                                    elm, sizeof(*elm));
 312                                 elm->leaf.base.create_tid = tid;
 313                         }
 314                 }
 315         }
 316
 317         /*
 318          * Align the delete_tid.  This only occurs if the record is historical
 319          * was deleted at some point.  Realigning the delete_tid does not
 320          * move the record within the B-Tree but may cause it to temporarily
 321          * overlap a record that has not yet been pruned.
 322          */
 323         if (error == 0 && realign_del >= 0) {
 324                 mod = prune->elms[realign_del].mod_tid;
 325                 delta = elm->leaf.base.delete_tid % mod;
 326                 hammer_modify_node(cursor->trans, cursor->node,
 327                                    elm, sizeof(*elm));
 328                 if (delta) {
 329                         error = hammer_btree_extract(cursor,
 330                                                      HAMMER_CURSOR_GET_RECORD);
 331                         if (error == 0) {
 332                                 elm->leaf.base.delete_tid =
 333                                                 elm->leaf.base.delete_tid -
 334                                                 delta + mod;
 335                                 hammer_modify_buffer(cursor->trans, cursor->record_buffer, &cursor->record->base.base.delete_tid, sizeof(hammer_tid_t));
 336                                 cursor->record->base.base.delete_tid =
 337                                                 elm->leaf.base.delete_tid;
 338                         }
 339                 }
 340         }
 341         return (error);
 342 }
 343
 344 /*
 345  * Iterate through an object's inode or an object's records and record
 346  * modification TIDs.
 347  */
 348 static void add_history(hammer_inode_t ip, struct hammer_ioc_history *hist,
 349                         hammer_btree_elm_t elm);
 350
 351 static
 352 int
 353 hammer_ioc_gethistory(hammer_transaction_t trans, hammer_inode_t ip,
 354                       struct hammer_ioc_history *hist)
 355 {
 356         struct hammer_cursor cursor;
 357         hammer_btree_elm_t elm;
 358         int error;
 359
 360         /*
 361          * Validate the structure and initialize for return.
 362          */
 363         if (hist->beg_tid > hist->end_tid)
 364                 return(EINVAL);
 365         if (hist->flags & HAMMER_IOC_HISTORY_ATKEY) {
 366                 if (hist->key > hist->nxt_key)
 367                         return(EINVAL);
 368         }
 369
 370         hist->obj_id = ip->obj_id;
 371         hist->count = 0;
 372         hist->nxt_tid = hist->end_tid;
 373         hist->flags &= ~HAMMER_IOC_HISTORY_NEXT_TID;
 374         hist->flags &= ~HAMMER_IOC_HISTORY_NEXT_KEY;
 375         hist->flags &= ~HAMMER_IOC_HISTORY_EOF;
 376         hist->flags &= ~HAMMER_IOC_HISTORY_UNSYNCED;
 377         if ((ip->flags & HAMMER_INODE_MODMASK) & ~HAMMER_INODE_ITIMES)
 378                 hist->flags |= HAMMER_IOC_HISTORY_UNSYNCED;
 379
 380         /*
 381          * Setup the cursor.  We can't handle undeletable records
 382          * (create_tid of 0) at the moment.  A create_tid of 0 has
 383          * a special meaning and cannot be specified in the cursor.
 384          */
 385         error = hammer_init_cursor(trans, &cursor, &ip->cache[0]);
 386         if (error) {
 387                 hammer_done_cursor(&cursor);
 388                 return(error);
 389         }
 390
 391         cursor.key_beg.obj_id = hist->obj_id;
 392         cursor.key_beg.create_tid = hist->beg_tid;
 393         cursor.key_beg.delete_tid = 0;
 394         cursor.key_beg.obj_type = 0;
 395         if (cursor.key_beg.create_tid == HAMMER_MIN_TID)
 396                 cursor.key_beg.create_tid = 1;
 397
 398         cursor.key_end.obj_id = hist->obj_id;
 399         cursor.key_end.create_tid = hist->end_tid;
 400         cursor.key_end.delete_tid = 0;
 401         cursor.key_end.obj_type = 0;
 402
 403         cursor.flags |= HAMMER_CURSOR_END_EXCLUSIVE;
 404
 405         if (hist->flags & HAMMER_IOC_HISTORY_ATKEY) {
 406                 /*
 407                  * key-range within the file.  For a regular file the
 408                  * on-disk key represents BASE+LEN, not BASE, so the
 409                  * first possible record containing the offset 'key'
 410                  * has an on-disk key of (key + 1).
 411                  */
 412                 cursor.key_beg.key = hist->key;
 413                 cursor.key_end.key = HAMMER_MAX_KEY;
 414
 415                 switch(ip->ino_rec.base.base.obj_type) {
 416                 case HAMMER_OBJTYPE_REGFILE:
 417                         ++cursor.key_beg.key;
 418                         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
 419                         break;
 420                 case HAMMER_OBJTYPE_DIRECTORY:
 421                         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
 422                         break;
 423                 case HAMMER_OBJTYPE_DBFILE:
 424                         cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
 425                         break;
 426                 default:
 427                         error = EINVAL;
 428                         break;
 429                 }
 430                 cursor.key_end.rec_type = cursor.key_beg.rec_type;
 431         } else {
 432                 /*
 433                  * The inode itself.
 434                  */
 435                 cursor.key_beg.key = 0;
 436                 cursor.key_end.key = 0;
 437                 cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 438                 cursor.key_end.rec_type = HAMMER_RECTYPE_INODE;
 439         }
 440
 441         error = hammer_btree_first(&cursor);
 442         while (error == 0) {
 443                 elm = &cursor.node->ondisk->elms[cursor.index];
 444
 445                 add_history(ip, hist, elm);
 446                 if (hist->flags & (HAMMER_IOC_HISTORY_NEXT_TID |
 447                                   HAMMER_IOC_HISTORY_NEXT_KEY |
 448                                   HAMMER_IOC_HISTORY_EOF)) {
 449                         break;
 450                 }
 451                 error = hammer_btree_iterate(&cursor);
 452         }
 453         if (error == ENOENT) {
 454                 hist->flags |= HAMMER_IOC_HISTORY_EOF;
 455                 error = 0;
 456         }
 457         hammer_done_cursor(&cursor);
 458         return(error);
 459 }
 460
 461 /*
 462  * Add the scanned element to the ioctl return structure.  Some special
 463  * casing is required for regular files to accomodate how data ranges are
 464  * stored on-disk.
 465  */
 466 static void
 467 add_history(hammer_inode_t ip, struct hammer_ioc_history *hist,
 468             hammer_btree_elm_t elm)
 469 {
 470         if (elm->base.btype != HAMMER_BTREE_TYPE_RECORD)
 471                 return;
 472         if ((hist->flags & HAMMER_IOC_HISTORY_ATKEY) &&
 473             ip->ino_rec.base.base.obj_type == HAMMER_OBJTYPE_REGFILE) {
 474                 /*
 475                  * Adjust nxt_key
 476                  */
 477                 if (hist->nxt_key > elm->leaf.base.key - elm->leaf.data_len &&
 478                     hist->key < elm->leaf.base.key - elm->leaf.data_len) {
 479                         hist->nxt_key = elm->leaf.base.key - elm->leaf.data_len;
 480                 }
 481                 if (hist->nxt_key > elm->leaf.base.key)
 482                         hist->nxt_key = elm->leaf.base.key;
 483
 484                 /*
 485                  * Record is beyond MAXPHYS, there won't be any more records
 486                  * in the iteration covering the requested offset (key).
 487                  */
 488                 if (elm->leaf.base.key >= MAXPHYS &&
 489                     elm->leaf.base.key - MAXPHYS > hist->key) {
 490                         hist->flags |= HAMMER_IOC_HISTORY_NEXT_KEY;
 491                 }
 492
 493                 /*
 494                  * Data-range of record does not cover the key.
 495                  */
 496                 if (elm->leaf.base.key - elm->leaf.data_len > hist->key)
 497                         return;
 498
 499         } else if (hist->flags & HAMMER_IOC_HISTORY_ATKEY) {
 500                 /*
 501                  * Adjust nxt_key
 502                  */
 503                 if (hist->nxt_key > elm->leaf.base.key &&
 504                     hist->key < elm->leaf.base.key) {
 505                         hist->nxt_key = elm->leaf.base.key;
 506                 }
 507
 508                 /*
 509                  * Record is beyond the requested key.
 510                  */
 511                 if (elm->leaf.base.key > hist->key)
 512                         hist->flags |= HAMMER_IOC_HISTORY_NEXT_KEY;
 513         }
 514
 515         /*
 516          * Add create_tid if it is in-bounds.
 517          */
 518         if ((hist->count == 0 ||
 519              elm->leaf.base.create_tid != hist->tid_ary[hist->count - 1]) &&
 520             elm->leaf.base.create_tid >= hist->beg_tid &&
 521             elm->leaf.base.create_tid < hist->end_tid) {
 522                 if (hist->count == HAMMER_MAX_HISTORY_ELMS) {
 523                         hist->nxt_tid = elm->leaf.base.create_tid;
 524                         hist->flags |= HAMMER_IOC_HISTORY_NEXT_TID;
 525                         return;
 526                 }
 527                 hist->tid_ary[hist->count++] = elm->leaf.base.create_tid;
 528         }
 529
 530         /*
 531          * Add delete_tid if it is in-bounds.  Note that different portions
 532          * of the history may have overlapping data ranges with different
 533          * delete_tid's.  If this case occurs the delete_tid may match the
 534          * create_tid of a following record.  XXX
 535          *
 536          *      [        ]
 537          *            [     ]
 538          */
 539         if (elm->leaf.base.delete_tid &&
 540             elm->leaf.base.delete_tid >= hist->beg_tid &&
 541             elm->leaf.base.delete_tid < hist->end_tid) {
 542                 if (hist->count == HAMMER_MAX_HISTORY_ELMS) {
 543                         hist->nxt_tid = elm->leaf.base.delete_tid;
 544                         hist->flags |= HAMMER_IOC_HISTORY_NEXT_TID;
 545                         return;
 546                 }
 547                 hist->tid_ary[hist->count++] = elm->leaf.base.delete_tid;
 548         }
 549 }
 550