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 1.34 2008/11/13 02:18:43 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_leaf_node(struct hammer_ioc_reblock *reblock, 54 hammer_cursor_t cursor, hammer_btree_elm_t elm); 55 static int hammer_reblock_int_node(struct hammer_ioc_reblock *reblock, 56 hammer_cursor_t cursor, hammer_btree_elm_t elm); 57 58 int 59 hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip, 60 struct hammer_ioc_reblock *reblock) 61 { 62 struct hammer_cursor cursor; 63 hammer_btree_elm_t elm; 64 int checkspace_count; 65 int error; 66 int seq; 67 int slop; 68 69 /* 70 * A fill level <= 20% is considered an emergency. free_level is 71 * inverted from fill_level. 72 */ 73 if (reblock->free_level >= HAMMER_LARGEBLOCK_SIZE * 8 / 10) 74 slop = HAMMER_CHKSPC_EMERGENCY; 75 else 76 slop = HAMMER_CHKSPC_REBLOCK; 77 78 if ((reblock->key_beg.localization | reblock->key_end.localization) & 79 HAMMER_LOCALIZE_PSEUDOFS_MASK) { 80 return(EINVAL); 81 } 82 if (reblock->key_beg.obj_id >= reblock->key_end.obj_id) 83 return(EINVAL); 84 if (reblock->free_level < 0) 85 return(EINVAL); 86 87 reblock->key_cur = reblock->key_beg; 88 reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK; 89 reblock->key_cur.localization += ip->obj_localization; 90 91 checkspace_count = 0; 92 seq = trans->hmp->flusher.act; 93 retry: 94 error = hammer_init_cursor(trans, &cursor, NULL, NULL); 95 if (error) { 96 hammer_done_cursor(&cursor); 97 goto failed; 98 } 99 cursor.key_beg.localization = reblock->key_cur.localization; 100 cursor.key_beg.obj_id = reblock->key_cur.obj_id; 101 cursor.key_beg.key = HAMMER_MIN_KEY; 102 cursor.key_beg.create_tid = 1; 103 cursor.key_beg.delete_tid = 0; 104 cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE; 105 cursor.key_beg.obj_type = 0; 106 107 cursor.key_end.localization = (reblock->key_end.localization & 108 HAMMER_LOCALIZE_MASK) + 109 ip->obj_localization; 110 cursor.key_end.obj_id = reblock->key_end.obj_id; 111 cursor.key_end.key = HAMMER_MAX_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 btree scan code to return internal nodes, 122 * so we can reblock them in addition to the leafs. Only specify it 123 * if we intend to reblock B-Tree nodes. 124 */ 125 if (reblock->head.flags & HAMMER_IOC_DO_BTREE) 126 cursor.flags |= HAMMER_CURSOR_REBLOCKING; 127 128 error = hammer_btree_first(&cursor); 129 while (error == 0) { 130 /* 131 * Internal or Leaf node 132 */ 133 elm = &cursor.node->ondisk->elms[cursor.index]; 134 reblock->key_cur.obj_id = elm->base.obj_id; 135 reblock->key_cur.localization = elm->base.localization; 136 137 /* 138 * Yield to more important tasks 139 */ 140 if ((error = hammer_signal_check(trans->hmp)) != 0) 141 break; 142 143 /* 144 * If there is insufficient free space it may be due to 145 * reserved bigblocks, which flushing might fix. 146 * 147 * WARNING: See warnings in hammer_unlock_cursor() function. 148 */ 149 if (hammer_checkspace(trans->hmp, slop)) { 150 if (++checkspace_count == 10) { 151 error = ENOSPC; 152 break; 153 } 154 hammer_unlock_cursor(&cursor); 155 hammer_flusher_wait(trans->hmp, seq); 156 hammer_lock_cursor(&cursor); 157 seq = hammer_flusher_async(trans->hmp, NULL); 158 continue; 159 } 160 161 /* 162 * Acquiring the sync_lock prevents the operation from 163 * crossing a synchronization boundary. 164 * 165 * NOTE: cursor.node may have changed on return. 166 * 167 * WARNING: See warnings in hammer_unlock_cursor() function. 168 */ 169 hammer_sync_lock_sh(trans); 170 error = hammer_reblock_helper(reblock, &cursor, elm); 171 hammer_sync_unlock(trans); 172 173 while (hammer_flusher_meta_halflimit(trans->hmp) || 174 hammer_flusher_undo_exhausted(trans, 2)) { 175 hammer_unlock_cursor(&cursor); 176 hammer_flusher_wait(trans->hmp, seq); 177 hammer_lock_cursor(&cursor); 178 seq = hammer_flusher_async_one(trans->hmp); 179 } 180 181 /* 182 * Setup for iteration, our cursor flags may be modified by 183 * other threads while we are unlocked. 184 */ 185 cursor.flags |= HAMMER_CURSOR_ATEDISK; 186 187 /* 188 * We allocate data buffers, which atm we don't track 189 * dirty levels for because we allow the kernel to write 190 * them. But if we allocate too many we can still deadlock 191 * the buffer cache. 192 * 193 * WARNING: See warnings in hammer_unlock_cursor() function. 194 * (The cursor's node and element may change!) 195 */ 196 if (bd_heatup()) { 197 hammer_unlock_cursor(&cursor); 198 bwillwrite(HAMMER_XBUFSIZE); 199 hammer_lock_cursor(&cursor); 200 } 201 202 if (error == 0) { 203 error = hammer_btree_iterate(&cursor); 204 } 205 206 } 207 if (error == ENOENT) 208 error = 0; 209 hammer_done_cursor(&cursor); 210 if (error == EWOULDBLOCK) { 211 hammer_flusher_sync(trans->hmp); 212 goto retry; 213 } 214 if (error == EDEADLK) 215 goto retry; 216 if (error == EINTR) { 217 reblock->head.flags |= HAMMER_IOC_HEAD_INTR; 218 error = 0; 219 } 220 failed: 221 reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK; 222 return(error); 223 } 224 225 /* 226 * Reblock the B-Tree (leaf) node, record, and/or data if necessary. 227 * 228 * XXX We have no visibility into internal B-Tree nodes at the moment, 229 * only leaf nodes. 230 */ 231 static int 232 hammer_reblock_helper(struct hammer_ioc_reblock *reblock, 233 hammer_cursor_t cursor, hammer_btree_elm_t elm) 234 { 235 hammer_mount_t hmp; 236 hammer_off_t tmp_offset; 237 struct hammer_btree_leaf_elm leaf; 238 int error; 239 int bytes; 240 int cur; 241 int iocflags; 242 243 error = 0; 244 hmp = cursor->trans->hmp; 245 246 /* 247 * Reblock data. Note that data embedded in a record is reblocked 248 * by the record reblock code. Data processing only occurs at leaf 249 * nodes and for RECORD element types. 250 */ 251 if (cursor->node->ondisk->type != HAMMER_BTREE_TYPE_LEAF) 252 goto skip; 253 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD) 254 return(0); 255 tmp_offset = elm->leaf.data_offset; 256 if (tmp_offset == 0) 257 goto skip; 258 if (error) 259 goto skip; 260 261 /* 262 * NOTE: Localization restrictions may also have been set-up, we can't 263 * just set the match flags willy-nilly here. 264 */ 265 switch(elm->leaf.base.rec_type) { 266 case HAMMER_RECTYPE_INODE: 267 iocflags = HAMMER_IOC_DO_INODES; 268 break; 269 case HAMMER_RECTYPE_EXT: 270 case HAMMER_RECTYPE_FIX: 271 case HAMMER_RECTYPE_PFS: 272 case HAMMER_RECTYPE_DIRENTRY: 273 iocflags = HAMMER_IOC_DO_DIRS; 274 break; 275 case HAMMER_RECTYPE_DATA: 276 case HAMMER_RECTYPE_DB: 277 iocflags = HAMMER_IOC_DO_DATA; 278 break; 279 default: 280 iocflags = 0; 281 break; 282 } 283 if (reblock->head.flags & iocflags) { 284 ++reblock->data_count; 285 reblock->data_byte_count += elm->leaf.data_len; 286 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error); 287 if (hammer_debug_general & 0x4000) 288 kprintf("D %6d/%d\n", bytes, reblock->free_level); 289 if (error == 0 && (cur == 0 || reblock->free_level == 0) && 290 bytes >= reblock->free_level) { 291 /* 292 * This is nasty, the uncache code may have to get 293 * vnode locks and because of that we can't hold 294 * the cursor locked. 295 * 296 * WARNING: See warnings in hammer_unlock_cursor() 297 * function. 298 */ 299 leaf = elm->leaf; 300 hammer_unlock_cursor(cursor); 301 hammer_io_direct_uncache(hmp, &leaf); 302 hammer_lock_cursor(cursor); 303 if (cursor->flags & HAMMER_CURSOR_RETEST) { 304 kprintf("hammer: retest after uncache\n"); 305 error = EDEADLK; 306 } else { 307 KKASSERT(bcmp(&elm->leaf, &leaf, sizeof(leaf)) == 0); 308 } 309 if (error == 0) 310 error = hammer_cursor_upgrade(cursor); 311 if (error == 0) { 312 error = hammer_reblock_data(reblock, 313 cursor, elm); 314 } 315 if (error == 0) { 316 ++reblock->data_moves; 317 reblock->data_byte_moves += elm->leaf.data_len; 318 } 319 } 320 } 321 322 skip: 323 /* 324 * Reblock a B-Tree internal or leaf node. A leaf node is reblocked 325 * on initial entry only (element 0). An internal node is reblocked 326 * when entered upward from its first leaf node only (also element 0). 327 * Further revisits of the internal node (index > 0) are ignored. 328 */ 329 tmp_offset = cursor->node->node_offset; 330 if (cursor->index == 0 && 331 error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) { 332 ++reblock->btree_count; 333 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error); 334 if (hammer_debug_general & 0x4000) 335 kprintf("B %6d/%d\n", bytes, reblock->free_level); 336 if (error == 0 && (cur == 0 || reblock->free_level == 0) && 337 bytes >= reblock->free_level) { 338 error = hammer_cursor_upgrade(cursor); 339 if (error == 0) { 340 if (cursor->parent) 341 elm = &cursor->parent->ondisk->elms[cursor->parent_index]; 342 else 343 elm = NULL; 344 switch(cursor->node->ondisk->type) { 345 case HAMMER_BTREE_TYPE_LEAF: 346 error = hammer_reblock_leaf_node( 347 reblock, cursor, elm); 348 break; 349 case HAMMER_BTREE_TYPE_INTERNAL: 350 error = hammer_reblock_int_node( 351 reblock, cursor, elm); 352 break; 353 default: 354 panic("Illegal B-Tree node type"); 355 } 356 } 357 if (error == 0) { 358 ++reblock->btree_moves; 359 } 360 } 361 } 362 363 hammer_cursor_downgrade(cursor); 364 return(error); 365 } 366 367 /* 368 * Reblock a record's data. Both the B-Tree element and record pointers 369 * to the data must be adjusted. 370 */ 371 static int 372 hammer_reblock_data(struct hammer_ioc_reblock *reblock, 373 hammer_cursor_t cursor, hammer_btree_elm_t elm) 374 { 375 struct hammer_buffer *data_buffer = NULL; 376 hammer_off_t ndata_offset; 377 int error; 378 void *ndata; 379 380 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA | 381 HAMMER_CURSOR_GET_LEAF); 382 if (error) 383 return (error); 384 ndata = hammer_alloc_data(cursor->trans, elm->leaf.data_len, 385 elm->leaf.base.rec_type, 386 &ndata_offset, &data_buffer, 387 0, &error); 388 if (error) 389 goto done; 390 391 /* 392 * Move the data 393 */ 394 hammer_modify_buffer(cursor->trans, data_buffer, NULL, 0); 395 bcopy(cursor->data, ndata, elm->leaf.data_len); 396 hammer_modify_buffer_done(data_buffer); 397 398 hammer_blockmap_free(cursor->trans, 399 elm->leaf.data_offset, elm->leaf.data_len); 400 401 hammer_modify_node(cursor->trans, cursor->node, 402 &elm->leaf.data_offset, sizeof(hammer_off_t)); 403 elm->leaf.data_offset = ndata_offset; 404 hammer_modify_node_done(cursor->node); 405 406 done: 407 if (data_buffer) 408 hammer_rel_buffer(data_buffer, 0); 409 return (error); 410 } 411 412 /* 413 * Reblock a B-Tree leaf node. The parent must be adjusted to point to 414 * the new copy of the leaf node. 415 * 416 * elm is a pointer to the parent element pointing at cursor.node. 417 */ 418 static int 419 hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock, 420 hammer_cursor_t cursor, hammer_btree_elm_t elm) 421 { 422 hammer_node_t onode; 423 hammer_node_t nnode; 424 int error; 425 426 /* 427 * Don't supply a hint when allocating the leaf. Fills are done 428 * from the leaf upwards. 429 */ 430 onode = cursor->node; 431 nnode = hammer_alloc_btree(cursor->trans, 0, &error); 432 433 if (nnode == NULL) 434 return (error); 435 436 /* 437 * Move the node 438 */ 439 hammer_lock_ex(&nnode->lock); 440 hammer_modify_node_noundo(cursor->trans, nnode); 441 bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk)); 442 443 if (elm) { 444 /* 445 * We are not the root of the B-Tree 446 */ 447 hammer_modify_node(cursor->trans, cursor->parent, 448 &elm->internal.subtree_offset, 449 sizeof(elm->internal.subtree_offset)); 450 elm->internal.subtree_offset = nnode->node_offset; 451 hammer_modify_node_done(cursor->parent); 452 } else { 453 /* 454 * We are the root of the B-Tree 455 */ 456 hammer_volume_t volume; 457 458 volume = hammer_get_root_volume(cursor->trans->hmp, &error); 459 KKASSERT(error == 0); 460 461 hammer_modify_volume_field(cursor->trans, volume, 462 vol0_btree_root); 463 volume->ondisk->vol0_btree_root = nnode->node_offset; 464 hammer_modify_volume_done(volume); 465 hammer_rel_volume(volume, 0); 466 } 467 468 hammer_cursor_replaced_node(onode, nnode); 469 hammer_delete_node(cursor->trans, onode); 470 471 if (hammer_debug_general & 0x4000) { 472 kprintf("REBLOCK LNODE %016llx -> %016llx\n", 473 (long long)onode->node_offset, 474 (long long)nnode->node_offset); 475 } 476 hammer_modify_node_done(nnode); 477 cursor->node = nnode; 478 479 hammer_unlock(&onode->lock); 480 hammer_rel_node(onode); 481 482 return (error); 483 } 484 485 /* 486 * Reblock a B-Tree internal node. The parent must be adjusted to point to 487 * the new copy of the internal node, and the node's children's parent 488 * pointers must also be adjusted to point to the new copy. 489 * 490 * elm is a pointer to the parent element pointing at cursor.node. 491 */ 492 static int 493 hammer_reblock_int_node(struct hammer_ioc_reblock *reblock, 494 hammer_cursor_t cursor, hammer_btree_elm_t elm) 495 { 496 struct hammer_node_lock lockroot; 497 hammer_node_t onode; 498 hammer_node_t nnode; 499 hammer_off_t hint; 500 int error; 501 int i; 502 503 hammer_node_lock_init(&lockroot, cursor->node); 504 error = hammer_btree_lock_children(cursor, 1, &lockroot); 505 if (error) 506 goto done; 507 508 /* 509 * The internal node is visited after recursing through its 510 * first element. Use the subtree offset allocated for that 511 * element as a hint for allocating the internal node. 512 */ 513 onode = cursor->node; 514 if (onode->ondisk->count) 515 hint = onode->ondisk->elms[0].internal.subtree_offset; 516 else 517 hint = 0; 518 nnode = hammer_alloc_btree(cursor->trans, hint, &error); 519 520 if (nnode == NULL) 521 goto done; 522 523 /* 524 * Move the node. Adjust the parent's pointer to us first. 525 */ 526 hammer_lock_ex(&nnode->lock); 527 hammer_modify_node_noundo(cursor->trans, nnode); 528 bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk)); 529 530 if (elm) { 531 /* 532 * We are not the root of the B-Tree 533 */ 534 hammer_modify_node(cursor->trans, cursor->parent, 535 &elm->internal.subtree_offset, 536 sizeof(elm->internal.subtree_offset)); 537 elm->internal.subtree_offset = nnode->node_offset; 538 hammer_modify_node_done(cursor->parent); 539 } else { 540 /* 541 * We are the root of the B-Tree 542 */ 543 hammer_volume_t volume; 544 545 volume = hammer_get_root_volume(cursor->trans->hmp, &error); 546 KKASSERT(error == 0); 547 548 hammer_modify_volume_field(cursor->trans, volume, 549 vol0_btree_root); 550 volume->ondisk->vol0_btree_root = nnode->node_offset; 551 hammer_modify_volume_done(volume); 552 hammer_rel_volume(volume, 0); 553 } 554 555 /* 556 * Now adjust our children's pointers to us. 557 */ 558 for (i = 0; i < nnode->ondisk->count; ++i) { 559 elm = &nnode->ondisk->elms[i]; 560 error = btree_set_parent(cursor->trans, nnode, elm); 561 if (error) 562 panic("reblock internal node: fixup problem"); 563 } 564 565 /* 566 * Clean up. 567 * 568 * The new node replaces the current node in the cursor. The cursor 569 * expects it to be locked so leave it locked. Discard onode. 570 */ 571 hammer_cursor_replaced_node(onode, nnode); 572 hammer_delete_node(cursor->trans, onode); 573 574 if (hammer_debug_general & 0x4000) { 575 kprintf("REBLOCK INODE %016llx -> %016llx\n", 576 (long long)onode->node_offset, 577 (long long)nnode->node_offset); 578 } 579 hammer_modify_node_done(nnode); 580 cursor->node = nnode; 581 582 hammer_unlock(&onode->lock); 583 hammer_rel_node(onode); 584 585 done: 586 hammer_btree_unlock_children(cursor, &lockroot); 587 return (error); 588 } 589 590