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 1.9 2008/04/25 21:49:49 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_done_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 (prune->stat_oldest_tid > elm->leaf.base.create_tid)
 138                         prune->stat_oldest_tid = elm->leaf.base.create_tid;
 139
 140                 if (check_prune(prune, elm, &realign_cre, &realign_del) == 0) {
 141                         if (hammer_debug_general & 0x0200) {
 142                                 kprintf("check %016llx %016llx: DELETE\n",
 143                                         elm->base.obj_id, elm->base.key);
 144                         }
 145
 146                         /*
 147                          * NOTE: This can return EDEADLK
 148                          */
 149                         isdir = (elm->base.rec_type == HAMMER_RECTYPE_DIRENTRY);
 150
 151                         error = hammer_delete_at_cursor(&cursor,
 152                                                         &prune->stat_bytes);
 153                         if (error)
 154                                 break;
 155
 156                         if (isdir)
 157                                 ++prune->stat_dirrecords;
 158                         else
 159                                 ++prune->stat_rawrecords;
 160                 } else if (realign_cre >= 0 || realign_del >= 0) {
 161                         error = realign_prune(prune, &cursor,
 162                                               realign_cre, realign_del);
 163                         if (error == 0) {
 164                                 cursor.flags |= HAMMER_CURSOR_ATEDISK;
 165                                 if (hammer_debug_general & 0x0200) {
 166                                         kprintf("check %016llx %016llx: "
 167                                                 "REALIGN\n",
 168                                                 elm->base.obj_id,
 169                                                 elm->base.key);
 170                                 }
 171                         }
 172                 } else {
 173                         cursor.flags |= HAMMER_CURSOR_ATEDISK;
 174                         if (hammer_debug_general & 0x0100) {
 175                                 kprintf("check %016llx %016llx: SKIP\n",
 176                                         elm->base.obj_id, elm->base.key);
 177                         }
 178                 }
 179                 error = hammer_signal_check(trans->hmp);
 180                 if (error == 0)
 181                         error = hammer_btree_iterate_reverse(&cursor);
 182         }
 183         if (error == ENOENT)
 184                 error = 0;
 185         hammer_done_cursor(&cursor);
 186         if (error == EDEADLK)
 187                 goto retry;
 188         return(error);
 189 }
 190
 191 /*
 192  * Check pruning list.  The list must be sorted in descending order.
 193  */
 194 static int
 195 check_prune(struct hammer_ioc_prune *prune, hammer_btree_elm_t elm,
 196             int *realign_cre, int *realign_del)
 197 {
 198         struct hammer_ioc_prune_elm *scan;
 199         int i;
 200
 201         *realign_cre = -1;
 202         *realign_del = -1;
 203
 204         /*
 205          * If pruning everything remove all records with a non-zero
 206          * delete_tid.
 207          */
 208         if (prune->flags & HAMMER_IOC_PRUNE_ALL) {
 209                 if (elm->base.delete_tid != 0)
 210                         return(0);
 211                 return(-1);
 212         }
 213
 214         for (i = 0; i < prune->nelms; ++i) {
 215                 scan = &prune->elms[i];
 216
 217                 /*
 218                  * Locate the scan index covering the create and delete TIDs.
 219                  */
 220                 if (*realign_cre < 0 &&
 221                     elm->base.create_tid >= scan->beg_tid &&
 222                     elm->base.create_tid < scan->end_tid) {
 223                         *realign_cre = i;
 224                 }
 225                 if (*realign_del < 0 && elm->base.delete_tid &&
 226                     elm->base.delete_tid > scan->beg_tid &&
 227                     elm->base.delete_tid <= scan->end_tid) {
 228                         *realign_del = i;
 229                 }
 230
 231                 /*
 232                  * Now check for loop termination.
 233                  */
 234                 if (elm->base.create_tid >= scan->end_tid ||
 235                     elm->base.delete_tid > scan->end_tid) {
 236                         break;
 237                 }
 238
 239                 /*
 240                  * Now determine if we can delete the record.
 241                  */
 242                 if (elm->base.delete_tid &&
 243                     elm->base.create_tid >= scan->beg_tid &&
 244                     elm->base.delete_tid <= scan->end_tid &&
 245                     elm->base.create_tid / scan->mod_tid ==
 246                     elm->base.delete_tid / scan->mod_tid) {
 247                         return(0);
 248                 }
 249         }
 250         return(-1);
 251 }
 252
 253 /*
 254  * Align the record to cover any gaps created through the deletion of
 255  * records within the pruning space.  If we were to just delete the records
 256  * there would be gaps which in turn would cause a snapshot that is NOT on
 257  * a pruning boundary to appear corrupt to the user.  Forcing alignment
 258  * of the create_tid and delete_tid for retained records 'reconnects'
 259  * the previously contiguous space, making it contiguous again after the
 260  * deletions.
 261  *
 262  * The use of a reverse iteration allows us to safely align the records and
 263  * related elements without creating temporary overlaps.  XXX we should
 264  * add ordering dependancies for record buffers to guarantee consistency
 265  * during recovery.
 266  */
 267 static int
 268 realign_prune(struct hammer_ioc_prune *prune,
 269               hammer_cursor_t cursor, int realign_cre, int realign_del)
 270 {
 271         hammer_btree_elm_t elm;
 272         hammer_tid_t delta;
 273         hammer_tid_t mod;
 274         hammer_tid_t tid;
 275         int error;
 276
 277         hammer_cursor_downgrade(cursor);
 278
 279         elm = &cursor->node->ondisk->elms[cursor->index];
 280         ++prune->stat_realignments;
 281
 282         /*
 283          * Align the create_tid.  By doing a reverse iteration we guarantee
 284          * that all records after our current record have already been
 285          * aligned, allowing us to safely correct the right-hand-boundary
 286          * (because no record to our right if otherwise exactly matching
 287          * will have a create_tid to the left of our aligned create_tid).
 288          *
 289          * Ordering is important here XXX but disk write ordering for
 290          * inter-cluster corrections is not currently guaranteed.
 291          */
 292         error = 0;
 293         if (realign_cre >= 0) {
 294                 mod = prune->elms[realign_cre].mod_tid;
 295                 delta = elm->leaf.base.create_tid % mod;
 296                 if (delta) {
 297                         tid = elm->leaf.base.create_tid - delta + mod;
 298
 299                         /* can EDEADLK */
 300                         error = hammer_btree_correct_rhb(cursor, tid + 1);
 301                         if (error == 0) {
 302                                 error = hammer_btree_extract(cursor,
 303                                                      HAMMER_CURSOR_GET_RECORD);
 304                         }
 305                         if (error == 0) {
 306                                 /* can EDEADLK */
 307                                 error = hammer_cursor_upgrade(cursor);
 308                         }
 309                         if (error == 0) {
 310                                 hammer_modify_buffer(cursor->trans,
 311                                                      cursor->record_buffer,
 312                                                      NULL, 0);
 313                                 cursor->record->base.base.create_tid = tid;
 314                                 hammer_modify_buffer_done(cursor->record_buffer);
 315                                 hammer_modify_node(cursor->trans, cursor->node,
 316                                                    &elm->leaf.base.create_tid,
 317                                                    sizeof(elm->leaf.base.create_tid));
 318                                 elm->leaf.base.create_tid = tid;
 319                                 hammer_modify_node_done(cursor->node);
 320                         }
 321                 }
 322         }
 323
 324         /*
 325          * Align the delete_tid.  This only occurs if the record is historical
 326          * was deleted at some point.  Realigning the delete_tid does not
 327          * move the record within the B-Tree but may cause it to temporarily
 328          * overlap a record that has not yet been pruned.
 329          */
 330         if (error == 0 && realign_del >= 0) {
 331                 mod = prune->elms[realign_del].mod_tid;
 332                 delta = elm->leaf.base.delete_tid % mod;
 333                 if (delta) {
 334                         error = hammer_btree_extract(cursor,
 335                                                      HAMMER_CURSOR_GET_RECORD);
 336                         if (error == 0) {
 337                                 hammer_modify_node(cursor->trans, cursor->node,
 338                                                    &elm->leaf.base.delete_tid,
 339                                                    sizeof(elm->leaf.base.delete_tid));
 340                                 elm->leaf.base.delete_tid =
 341                                                 elm->leaf.base.delete_tid -
 342                                                 delta + mod;
 343                                 hammer_modify_node_done(cursor->node);
 344                                 hammer_modify_buffer(cursor->trans, cursor->record_buffer, &cursor->record->base.base.delete_tid, sizeof(hammer_tid_t));
 345                                 cursor->record->base.base.delete_tid =
 346                                                 elm->leaf.base.delete_tid;
 347                                 hammer_modify_buffer_done(cursor->record_buffer);
 348                         }
 349                 }
 350         }
 351         return (error);
 352 }
 353
 354 /*
 355  * Iterate through an object's inode or an object's records and record
 356  * modification TIDs.
 357  */
 358 static void add_history(hammer_inode_t ip, struct hammer_ioc_history *hist,
 359                         hammer_btree_elm_t elm);
 360
 361 static
 362 int
 363 hammer_ioc_gethistory(hammer_transaction_t trans, hammer_inode_t ip,
 364                       struct hammer_ioc_history *hist)
 365 {
 366         struct hammer_cursor cursor;
 367         hammer_btree_elm_t elm;
 368         int error;
 369
 370         /*
 371          * Validate the structure and initialize for return.
 372          */
 373         if (hist->beg_tid > hist->end_tid)
 374                 return(EINVAL);
 375         if (hist->flags & HAMMER_IOC_HISTORY_ATKEY) {
 376                 if (hist->key > hist->nxt_key)
 377                         return(EINVAL);
 378         }
 379
 380         hist->obj_id = ip->obj_id;
 381         hist->count = 0;
 382         hist->nxt_tid = hist->end_tid;
 383         hist->flags &= ~HAMMER_IOC_HISTORY_NEXT_TID;
 384         hist->flags &= ~HAMMER_IOC_HISTORY_NEXT_KEY;
 385         hist->flags &= ~HAMMER_IOC_HISTORY_EOF;
 386         hist->flags &= ~HAMMER_IOC_HISTORY_UNSYNCED;
 387         if ((ip->flags & HAMMER_INODE_MODMASK) & ~HAMMER_INODE_ITIMES)
 388                 hist->flags |= HAMMER_IOC_HISTORY_UNSYNCED;
 389
 390         /*
 391          * Setup the cursor.  We can't handle undeletable records
 392          * (create_tid of 0) at the moment.  A create_tid of 0 has
 393          * a special meaning and cannot be specified in the cursor.
 394          */
 395         error = hammer_init_cursor(trans, &cursor, &ip->cache[0]);
 396         if (error) {
 397                 hammer_done_cursor(&cursor);
 398                 return(error);
 399         }
 400
 401         cursor.key_beg.obj_id = hist->obj_id;
 402         cursor.key_beg.create_tid = hist->beg_tid;
 403         cursor.key_beg.delete_tid = 0;
 404         cursor.key_beg.obj_type = 0;
 405         if (cursor.key_beg.create_tid == HAMMER_MIN_TID)
 406                 cursor.key_beg.create_tid = 1;
 407
 408         cursor.key_end.obj_id = hist->obj_id;
 409         cursor.key_end.create_tid = hist->end_tid;
 410         cursor.key_end.delete_tid = 0;
 411         cursor.key_end.obj_type = 0;
 412
 413         cursor.flags |= HAMMER_CURSOR_END_EXCLUSIVE;
 414
 415         if (hist->flags & HAMMER_IOC_HISTORY_ATKEY) {
 416                 /*
 417                  * key-range within the file.  For a regular file the
 418                  * on-disk key represents BASE+LEN, not BASE, so the
 419                  * first possible record containing the offset 'key'
 420                  * has an on-disk key of (key + 1).
 421                  */
 422                 cursor.key_beg.key = hist->key;
 423                 cursor.key_end.key = HAMMER_MAX_KEY;
 424
 425                 switch(ip->ino_rec.base.base.obj_type) {
 426                 case HAMMER_OBJTYPE_REGFILE:
 427                         ++cursor.key_beg.key;
 428                         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
 429                         break;
 430                 case HAMMER_OBJTYPE_DIRECTORY:
 431                         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
 432                         break;
 433                 case HAMMER_OBJTYPE_DBFILE:
 434                         cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
 435                         break;
 436                 default:
 437                         error = EINVAL;
 438                         break;
 439                 }
 440                 cursor.key_end.rec_type = cursor.key_beg.rec_type;
 441         } else {
 442                 /*
 443                  * The inode itself.
 444                  */
 445                 cursor.key_beg.key = 0;
 446                 cursor.key_end.key = 0;
 447                 cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
 448                 cursor.key_end.rec_type = HAMMER_RECTYPE_INODE;
 449         }
 450
 451         error = hammer_btree_first(&cursor);
 452         while (error == 0) {
 453                 elm = &cursor.node->ondisk->elms[cursor.index];
 454
 455                 add_history(ip, hist, elm);
 456                 if (hist->flags & (HAMMER_IOC_HISTORY_NEXT_TID |
 457                                   HAMMER_IOC_HISTORY_NEXT_KEY |
 458                                   HAMMER_IOC_HISTORY_EOF)) {
 459                         break;
 460                 }
 461                 error = hammer_btree_iterate(&cursor);
 462         }
 463         if (error == ENOENT) {
 464                 hist->flags |= HAMMER_IOC_HISTORY_EOF;
 465                 error = 0;
 466         }
 467         hammer_done_cursor(&cursor);
 468         return(error);
 469 }
 470
 471 /*
 472  * Add the scanned element to the ioctl return structure.  Some special
 473  * casing is required for regular files to accomodate how data ranges are
 474  * stored on-disk.
 475  */
 476 static void
 477 add_history(hammer_inode_t ip, struct hammer_ioc_history *hist,
 478             hammer_btree_elm_t elm)
 479 {
 480         if (elm->base.btype != HAMMER_BTREE_TYPE_RECORD)
 481                 return;
 482         if ((hist->flags & HAMMER_IOC_HISTORY_ATKEY) &&
 483             ip->ino_rec.base.base.obj_type == HAMMER_OBJTYPE_REGFILE) {
 484                 /*
 485                  * Adjust nxt_key
 486                  */
 487                 if (hist->nxt_key > elm->leaf.base.key - elm->leaf.data_len &&
 488                     hist->key < elm->leaf.base.key - elm->leaf.data_len) {
 489                         hist->nxt_key = elm->leaf.base.key - elm->leaf.data_len;
 490                 }
 491                 if (hist->nxt_key > elm->leaf.base.key)
 492                         hist->nxt_key = elm->leaf.base.key;
 493
 494                 /*
 495                  * Record is beyond MAXPHYS, there won't be any more records
 496                  * in the iteration covering the requested offset (key).
 497                  */
 498                 if (elm->leaf.base.key >= MAXPHYS &&
 499                     elm->leaf.base.key - MAXPHYS > hist->key) {
 500                         hist->flags |= HAMMER_IOC_HISTORY_NEXT_KEY;
 501                 }
 502
 503                 /*
 504                  * Data-range of record does not cover the key.
 505                  */
 506                 if (elm->leaf.base.key - elm->leaf.data_len > hist->key)
 507                         return;
 508
 509         } else if (hist->flags & HAMMER_IOC_HISTORY_ATKEY) {
 510                 /*
 511                  * Adjust nxt_key
 512                  */
 513                 if (hist->nxt_key > elm->leaf.base.key &&
 514                     hist->key < elm->leaf.base.key) {
 515                         hist->nxt_key = elm->leaf.base.key;
 516                 }
 517
 518                 /*
 519                  * Record is beyond the requested key.
 520                  */
 521                 if (elm->leaf.base.key > hist->key)
 522                         hist->flags |= HAMMER_IOC_HISTORY_NEXT_KEY;
 523         }
 524
 525         /*
 526          * Add create_tid if it is in-bounds.
 527          */
 528         if ((hist->count == 0 ||
 529              elm->leaf.base.create_tid != hist->tid_ary[hist->count - 1]) &&
 530             elm->leaf.base.create_tid >= hist->beg_tid &&
 531             elm->leaf.base.create_tid < hist->end_tid) {
 532                 if (hist->count == HAMMER_MAX_HISTORY_ELMS) {
 533                         hist->nxt_tid = elm->leaf.base.create_tid;
 534                         hist->flags |= HAMMER_IOC_HISTORY_NEXT_TID;
 535                         return;
 536                 }
 537                 hist->tid_ary[hist->count++] = elm->leaf.base.create_tid;
 538         }
 539
 540         /*
 541          * Add delete_tid if it is in-bounds.  Note that different portions
 542          * of the history may have overlapping data ranges with different
 543          * delete_tid's.  If this case occurs the delete_tid may match the
 544          * create_tid of a following record.  XXX
 545          *
 546          *      [        ]
 547          *            [     ]
 548          */
 549         if (elm->leaf.base.delete_tid &&
 550             elm->leaf.base.delete_tid >= hist->beg_tid &&
 551             elm->leaf.base.delete_tid < hist->end_tid) {
 552                 if (hist->count == HAMMER_MAX_HISTORY_ELMS) {
 553                         hist->nxt_tid = elm->leaf.base.delete_tid;
 554                         hist->flags |= HAMMER_IOC_HISTORY_NEXT_TID;
 555                         return;
 556                 }
 557                 hist->tid_ary[hist->count++] = elm->leaf.base.delete_tid;
 558         }
 559 }
 560