module/zfs/dsl_deadlist.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2012 by Delphix. All rights reserved.
  24  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  25  */
  26
  27 #include <sys/dsl_dataset.h>
  28 #include <sys/dmu.h>
  29 #include <sys/refcount.h>
  30 #include <sys/zap.h>
  31 #include <sys/zfs_context.h>
  32 #include <sys/dsl_pool.h>
  33
  34 /*
  35  * Deadlist concurrency:
  36  *
  37  * Deadlists can only be modified from the syncing thread.
  38  *
  39  * Except for dsl_deadlist_insert(), it can only be modified with the
  40  * dp_config_rwlock held with RW_WRITER.
  41  *
  42  * The accessors (dsl_deadlist_space() and dsl_deadlist_space_range()) can
  43  * be called concurrently, from open context, with the dl_config_rwlock held
  44  * with RW_READER.
  45  *
  46  * Therefore, we only need to provide locking between dsl_deadlist_insert() and
  47  * the accessors, protecting:
  48  *     dl_phys->dl_used,comp,uncomp
  49  *     and protecting the dl_tree from being loaded.
  50  * The locking is provided by dl_lock.  Note that locking on the bpobj_t
  51  * provides its own locking, and dl_oldfmt is immutable.
  52  */
  53
  54 static int
  55 dsl_deadlist_compare(const void *arg1, const void *arg2)
  56 {
  57         const dsl_deadlist_entry_t *dle1 = arg1;
  58         const dsl_deadlist_entry_t *dle2 = arg2;
  59
  60         if (dle1->dle_mintxg < dle2->dle_mintxg)
  61                 return (-1);
  62         else if (dle1->dle_mintxg > dle2->dle_mintxg)
  63                 return (+1);
  64         else
  65                 return (0);
  66 }
  67
  68 static void
  69 dsl_deadlist_load_tree(dsl_deadlist_t *dl)
  70 {
  71         zap_cursor_t zc;
  72         zap_attribute_t za;
  73
  74         ASSERT(!dl->dl_oldfmt);
  75         if (dl->dl_havetree)
  76                 return;
  77
  78         avl_create(&dl->dl_tree, dsl_deadlist_compare,
  79             sizeof (dsl_deadlist_entry_t),
  80             offsetof(dsl_deadlist_entry_t, dle_node));
  81         for (zap_cursor_init(&zc, dl->dl_os, dl->dl_object);
  82             zap_cursor_retrieve(&zc, &za) == 0;
  83             zap_cursor_advance(&zc)) {
  84                 dsl_deadlist_entry_t *dle;
  85
  86                 dle = kmem_alloc(sizeof (*dle), KM_SLEEP);
  87                 dle->dle_mintxg = strtonum(za.za_name, NULL);
  88                 VERIFY3U(0, ==, bpobj_open(&dle->dle_bpobj, dl->dl_os,
  89                     za.za_first_integer));
  90                 avl_add(&dl->dl_tree, dle);
  91         }
  92         zap_cursor_fini(&zc);
  93         dl->dl_havetree = B_TRUE;
  94 }
  95
  96 void
  97 dsl_deadlist_open(dsl_deadlist_t *dl, objset_t *os, uint64_t object)
  98 {
  99         dmu_object_info_t doi;
 100
 101         mutex_init(&dl->dl_lock, NULL, MUTEX_DEFAULT, NULL);
 102         dl->dl_os = os;
 103         dl->dl_object = object;
 104         VERIFY3U(0, ==, dmu_bonus_hold(os, object, dl, &dl->dl_dbuf));
 105         dmu_object_info_from_db(dl->dl_dbuf, &doi);
 106         if (doi.doi_type == DMU_OT_BPOBJ) {
 107                 dmu_buf_rele(dl->dl_dbuf, dl);
 108                 dl->dl_dbuf = NULL;
 109                 dl->dl_oldfmt = B_TRUE;
 110                 VERIFY3U(0, ==, bpobj_open(&dl->dl_bpobj, os, object));
 111                 return;
 112         }
 113
 114         dl->dl_oldfmt = B_FALSE;
 115         dl->dl_phys = dl->dl_dbuf->db_data;
 116         dl->dl_havetree = B_FALSE;
 117 }
 118
 119 void
 120 dsl_deadlist_close(dsl_deadlist_t *dl)
 121 {
 122         void *cookie = NULL;
 123         dsl_deadlist_entry_t *dle;
 124
 125         dl->dl_os = NULL;
 126
 127         if (dl->dl_oldfmt) {
 128                 dl->dl_oldfmt = B_FALSE;
 129                 bpobj_close(&dl->dl_bpobj);
 130                 return;
 131         }
 132
 133         if (dl->dl_havetree) {
 134                 while ((dle = avl_destroy_nodes(&dl->dl_tree, &cookie))
 135                     != NULL) {
 136                         bpobj_close(&dle->dle_bpobj);
 137                         kmem_free(dle, sizeof (*dle));
 138                 }
 139                 avl_destroy(&dl->dl_tree);
 140         }
 141         dmu_buf_rele(dl->dl_dbuf, dl);
 142         mutex_destroy(&dl->dl_lock);
 143         dl->dl_dbuf = NULL;
 144         dl->dl_phys = NULL;
 145 }
 146
 147 uint64_t
 148 dsl_deadlist_alloc(objset_t *os, dmu_tx_t *tx)
 149 {
 150         if (spa_version(dmu_objset_spa(os)) < SPA_VERSION_DEADLISTS)
 151                 return (bpobj_alloc(os, SPA_OLD_MAXBLOCKSIZE, tx));
 152         return (zap_create(os, DMU_OT_DEADLIST, DMU_OT_DEADLIST_HDR,
 153             sizeof (dsl_deadlist_phys_t), tx));
 154 }
 155
 156 void
 157 dsl_deadlist_free(objset_t *os, uint64_t dlobj, dmu_tx_t *tx)
 158 {
 159         dmu_object_info_t doi;
 160         zap_cursor_t zc;
 161         zap_attribute_t za;
 162
 163         VERIFY3U(0, ==, dmu_object_info(os, dlobj, &doi));
 164         if (doi.doi_type == DMU_OT_BPOBJ) {
 165                 bpobj_free(os, dlobj, tx);
 166                 return;
 167         }
 168
 169         for (zap_cursor_init(&zc, os, dlobj);
 170             zap_cursor_retrieve(&zc, &za) == 0;
 171             zap_cursor_advance(&zc)) {
 172                 uint64_t obj = za.za_first_integer;
 173                 if (obj == dmu_objset_pool(os)->dp_empty_bpobj)
 174                         bpobj_decr_empty(os, tx);
 175                 else
 176                         bpobj_free(os, obj, tx);
 177         }
 178         zap_cursor_fini(&zc);
 179         VERIFY3U(0, ==, dmu_object_free(os, dlobj, tx));
 180 }
 181
 182 static void
 183 dle_enqueue(dsl_deadlist_t *dl, dsl_deadlist_entry_t *dle,
 184     const blkptr_t *bp, dmu_tx_t *tx)
 185 {
 186         if (dle->dle_bpobj.bpo_object ==
 187             dmu_objset_pool(dl->dl_os)->dp_empty_bpobj) {
 188                 uint64_t obj = bpobj_alloc(dl->dl_os, SPA_OLD_MAXBLOCKSIZE, tx);
 189                 bpobj_close(&dle->dle_bpobj);
 190                 bpobj_decr_empty(dl->dl_os, tx);
 191                 VERIFY3U(0, ==, bpobj_open(&dle->dle_bpobj, dl->dl_os, obj));
 192                 VERIFY3U(0, ==, zap_update_int_key(dl->dl_os, dl->dl_object,
 193                     dle->dle_mintxg, obj, tx));
 194         }
 195         bpobj_enqueue(&dle->dle_bpobj, bp, tx);
 196 }
 197
 198 static void
 199 dle_enqueue_subobj(dsl_deadlist_t *dl, dsl_deadlist_entry_t *dle,
 200     uint64_t obj, dmu_tx_t *tx)
 201 {
 202         if (dle->dle_bpobj.bpo_object !=
 203             dmu_objset_pool(dl->dl_os)->dp_empty_bpobj) {
 204                 bpobj_enqueue_subobj(&dle->dle_bpobj, obj, tx);
 205         } else {
 206                 bpobj_close(&dle->dle_bpobj);
 207                 bpobj_decr_empty(dl->dl_os, tx);
 208                 VERIFY3U(0, ==, bpobj_open(&dle->dle_bpobj, dl->dl_os, obj));
 209                 VERIFY3U(0, ==, zap_update_int_key(dl->dl_os, dl->dl_object,
 210                     dle->dle_mintxg, obj, tx));
 211         }
 212 }
 213
 214 void
 215 dsl_deadlist_insert(dsl_deadlist_t *dl, const blkptr_t *bp, dmu_tx_t *tx)
 216 {
 217         dsl_deadlist_entry_t dle_tofind;
 218         dsl_deadlist_entry_t *dle;
 219         avl_index_t where;
 220
 221         if (dl->dl_oldfmt) {
 222                 bpobj_enqueue(&dl->dl_bpobj, bp, tx);
 223                 return;
 224         }
 225
 226         dsl_deadlist_load_tree(dl);
 227
 228         dmu_buf_will_dirty(dl->dl_dbuf, tx);
 229         mutex_enter(&dl->dl_lock);
 230         dl->dl_phys->dl_used +=
 231             bp_get_dsize_sync(dmu_objset_spa(dl->dl_os), bp);
 232         dl->dl_phys->dl_comp += BP_GET_PSIZE(bp);
 233         dl->dl_phys->dl_uncomp += BP_GET_UCSIZE(bp);
 234         mutex_exit(&dl->dl_lock);
 235
 236         dle_tofind.dle_mintxg = bp->blk_birth;
 237         dle = avl_find(&dl->dl_tree, &dle_tofind, &where);
 238         if (dle == NULL)
 239                 dle = avl_nearest(&dl->dl_tree, where, AVL_BEFORE);
 240         else
 241                 dle = AVL_PREV(&dl->dl_tree, dle);
 242
 243         if (dle == NULL) {
 244                 zfs_panic_recover("blkptr at %p has invalid BLK_BIRTH %llu",
 245                     bp, (longlong_t)bp->blk_birth);
 246                 dle = avl_first(&dl->dl_tree);
 247         }
 248
 249         ASSERT3P(dle, !=, NULL);
 250         dle_enqueue(dl, dle, bp, tx);
 251 }
 252
 253 /*
 254  * Insert new key in deadlist, which must be > all current entries.
 255  * mintxg is not inclusive.
 256  */
 257 void
 258 dsl_deadlist_add_key(dsl_deadlist_t *dl, uint64_t mintxg, dmu_tx_t *tx)
 259 {
 260         uint64_t obj;
 261         dsl_deadlist_entry_t *dle;
 262
 263         if (dl->dl_oldfmt)
 264                 return;
 265
 266         dsl_deadlist_load_tree(dl);
 267
 268         dle = kmem_alloc(sizeof (*dle), KM_SLEEP);
 269         dle->dle_mintxg = mintxg;
 270         obj = bpobj_alloc_empty(dl->dl_os, SPA_OLD_MAXBLOCKSIZE, tx);
 271         VERIFY3U(0, ==, bpobj_open(&dle->dle_bpobj, dl->dl_os, obj));
 272         avl_add(&dl->dl_tree, dle);
 273
 274         VERIFY3U(0, ==, zap_add_int_key(dl->dl_os, dl->dl_object,
 275             mintxg, obj, tx));
 276 }
 277
 278 /*
 279  * Remove this key, merging its entries into the previous key.
 280  */
 281 void
 282 dsl_deadlist_remove_key(dsl_deadlist_t *dl, uint64_t mintxg, dmu_tx_t *tx)
 283 {
 284         dsl_deadlist_entry_t dle_tofind;
 285         dsl_deadlist_entry_t *dle, *dle_prev;
 286
 287         if (dl->dl_oldfmt)
 288                 return;
 289
 290         dsl_deadlist_load_tree(dl);
 291
 292         dle_tofind.dle_mintxg = mintxg;
 293         dle = avl_find(&dl->dl_tree, &dle_tofind, NULL);
 294         dle_prev = AVL_PREV(&dl->dl_tree, dle);
 295
 296         dle_enqueue_subobj(dl, dle_prev, dle->dle_bpobj.bpo_object, tx);
 297
 298         avl_remove(&dl->dl_tree, dle);
 299         bpobj_close(&dle->dle_bpobj);
 300         kmem_free(dle, sizeof (*dle));
 301
 302         VERIFY3U(0, ==, zap_remove_int(dl->dl_os, dl->dl_object, mintxg, tx));
 303 }
 304
 305 /*
 306  * Walk ds's snapshots to regenerate generate ZAP & AVL.
 307  */
 308 static void
 309 dsl_deadlist_regenerate(objset_t *os, uint64_t dlobj,
 310     uint64_t mrs_obj, dmu_tx_t *tx)
 311 {
 312         dsl_deadlist_t dl;
 313         dsl_pool_t *dp = dmu_objset_pool(os);
 314
 315         dsl_deadlist_open(&dl, os, dlobj);
 316         if (dl.dl_oldfmt) {
 317                 dsl_deadlist_close(&dl);
 318                 return;
 319         }
 320
 321         while (mrs_obj != 0) {
 322                 dsl_dataset_t *ds;
 323                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, mrs_obj, FTAG, &ds));
 324                 dsl_deadlist_add_key(&dl,
 325                     dsl_dataset_phys(ds)->ds_prev_snap_txg, tx);
 326                 mrs_obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
 327                 dsl_dataset_rele(ds, FTAG);
 328         }
 329         dsl_deadlist_close(&dl);
 330 }
 331
 332 uint64_t
 333 dsl_deadlist_clone(dsl_deadlist_t *dl, uint64_t maxtxg,
 334     uint64_t mrs_obj, dmu_tx_t *tx)
 335 {
 336         dsl_deadlist_entry_t *dle;
 337         uint64_t newobj;
 338
 339         newobj = dsl_deadlist_alloc(dl->dl_os, tx);
 340
 341         if (dl->dl_oldfmt) {
 342                 dsl_deadlist_regenerate(dl->dl_os, newobj, mrs_obj, tx);
 343                 return (newobj);
 344         }
 345
 346         dsl_deadlist_load_tree(dl);
 347
 348         for (dle = avl_first(&dl->dl_tree); dle;
 349             dle = AVL_NEXT(&dl->dl_tree, dle)) {
 350                 uint64_t obj;
 351
 352                 if (dle->dle_mintxg >= maxtxg)
 353                         break;
 354
 355                 obj = bpobj_alloc_empty(dl->dl_os, SPA_OLD_MAXBLOCKSIZE, tx);
 356                 VERIFY3U(0, ==, zap_add_int_key(dl->dl_os, newobj,
 357                     dle->dle_mintxg, obj, tx));
 358         }
 359         return (newobj);
 360 }
 361
 362 void
 363 dsl_deadlist_space(dsl_deadlist_t *dl,
 364     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
 365 {
 366         if (dl->dl_oldfmt) {
 367                 VERIFY3U(0, ==, bpobj_space(&dl->dl_bpobj,
 368                     usedp, compp, uncompp));
 369                 return;
 370         }
 371
 372         mutex_enter(&dl->dl_lock);
 373         *usedp = dl->dl_phys->dl_used;
 374         *compp = dl->dl_phys->dl_comp;
 375         *uncompp = dl->dl_phys->dl_uncomp;
 376         mutex_exit(&dl->dl_lock);
 377 }
 378
 379 /*
 380  * return space used in the range (mintxg, maxtxg].
 381  * Includes maxtxg, does not include mintxg.
 382  * mintxg and maxtxg must both be keys in the deadlist (unless maxtxg is
 383  * larger than any bp in the deadlist (eg. UINT64_MAX)).
 384  */
 385 void
 386 dsl_deadlist_space_range(dsl_deadlist_t *dl, uint64_t mintxg, uint64_t maxtxg,
 387     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
 388 {
 389         dsl_deadlist_entry_t *dle;
 390         dsl_deadlist_entry_t dle_tofind;
 391         avl_index_t where;
 392
 393         if (dl->dl_oldfmt) {
 394                 VERIFY3U(0, ==, bpobj_space_range(&dl->dl_bpobj,
 395                     mintxg, maxtxg, usedp, compp, uncompp));
 396                 return;
 397         }
 398
 399         *usedp = *compp = *uncompp = 0;
 400
 401         mutex_enter(&dl->dl_lock);
 402         dsl_deadlist_load_tree(dl);
 403         dle_tofind.dle_mintxg = mintxg;
 404         dle = avl_find(&dl->dl_tree, &dle_tofind, &where);
 405         /*
 406          * If we don't find this mintxg, there shouldn't be anything
 407          * after it either.
 408          */
 409         ASSERT(dle != NULL ||
 410             avl_nearest(&dl->dl_tree, where, AVL_AFTER) == NULL);
 411
 412         for (; dle && dle->dle_mintxg < maxtxg;
 413             dle = AVL_NEXT(&dl->dl_tree, dle)) {
 414                 uint64_t used, comp, uncomp;
 415
 416                 VERIFY3U(0, ==, bpobj_space(&dle->dle_bpobj,
 417                     &used, &comp, &uncomp));
 418
 419                 *usedp += used;
 420                 *compp += comp;
 421                 *uncompp += uncomp;
 422         }
 423         mutex_exit(&dl->dl_lock);
 424 }
 425
 426 static void
 427 dsl_deadlist_insert_bpobj(dsl_deadlist_t *dl, uint64_t obj, uint64_t birth,
 428     dmu_tx_t *tx)
 429 {
 430         dsl_deadlist_entry_t dle_tofind;
 431         dsl_deadlist_entry_t *dle;
 432         avl_index_t where;
 433         uint64_t used, comp, uncomp;
 434         bpobj_t bpo;
 435
 436         VERIFY3U(0, ==, bpobj_open(&bpo, dl->dl_os, obj));
 437         VERIFY3U(0, ==, bpobj_space(&bpo, &used, &comp, &uncomp));
 438         bpobj_close(&bpo);
 439
 440         dsl_deadlist_load_tree(dl);
 441
 442         dmu_buf_will_dirty(dl->dl_dbuf, tx);
 443         mutex_enter(&dl->dl_lock);
 444         dl->dl_phys->dl_used += used;
 445         dl->dl_phys->dl_comp += comp;
 446         dl->dl_phys->dl_uncomp += uncomp;
 447         mutex_exit(&dl->dl_lock);
 448
 449         dle_tofind.dle_mintxg = birth;
 450         dle = avl_find(&dl->dl_tree, &dle_tofind, &where);
 451         if (dle == NULL)
 452                 dle = avl_nearest(&dl->dl_tree, where, AVL_BEFORE);
 453         dle_enqueue_subobj(dl, dle, obj, tx);
 454 }
 455
 456 static int
 457 dsl_deadlist_insert_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
 458 {
 459         dsl_deadlist_t *dl = arg;
 460         dsl_deadlist_insert(dl, bp, tx);
 461         return (0);
 462 }
 463
 464 /*
 465  * Merge the deadlist pointed to by 'obj' into dl.  obj will be left as
 466  * an empty deadlist.
 467  */
 468 void
 469 dsl_deadlist_merge(dsl_deadlist_t *dl, uint64_t obj, dmu_tx_t *tx)
 470 {
 471         zap_cursor_t zc;
 472         zap_attribute_t za;
 473         dmu_buf_t *bonus;
 474         dsl_deadlist_phys_t *dlp;
 475         dmu_object_info_t doi;
 476
 477         VERIFY3U(0, ==, dmu_object_info(dl->dl_os, obj, &doi));
 478         if (doi.doi_type == DMU_OT_BPOBJ) {
 479                 bpobj_t bpo;
 480                 VERIFY3U(0, ==, bpobj_open(&bpo, dl->dl_os, obj));
 481                 VERIFY3U(0, ==, bpobj_iterate(&bpo,
 482                     dsl_deadlist_insert_cb, dl, tx));
 483                 bpobj_close(&bpo);
 484                 return;
 485         }
 486
 487         for (zap_cursor_init(&zc, dl->dl_os, obj);
 488             zap_cursor_retrieve(&zc, &za) == 0;
 489             zap_cursor_advance(&zc)) {
 490                 uint64_t mintxg = strtonum(za.za_name, NULL);
 491                 dsl_deadlist_insert_bpobj(dl, za.za_first_integer, mintxg, tx);
 492                 VERIFY3U(0, ==, zap_remove_int(dl->dl_os, obj, mintxg, tx));
 493         }
 494         zap_cursor_fini(&zc);
 495
 496         VERIFY3U(0, ==, dmu_bonus_hold(dl->dl_os, obj, FTAG, &bonus));
 497         dlp = bonus->db_data;
 498         dmu_buf_will_dirty(bonus, tx);
 499         bzero(dlp, sizeof (*dlp));
 500         dmu_buf_rele(bonus, FTAG);
 501 }
 502
 503 /*
 504  * Remove entries on dl that are >= mintxg, and put them on the bpobj.
 505  */
 506 void
 507 dsl_deadlist_move_bpobj(dsl_deadlist_t *dl, bpobj_t *bpo, uint64_t mintxg,
 508     dmu_tx_t *tx)
 509 {
 510         dsl_deadlist_entry_t dle_tofind;
 511         dsl_deadlist_entry_t *dle;
 512         avl_index_t where;
 513
 514         ASSERT(!dl->dl_oldfmt);
 515         dmu_buf_will_dirty(dl->dl_dbuf, tx);
 516         dsl_deadlist_load_tree(dl);
 517
 518         dle_tofind.dle_mintxg = mintxg;
 519         dle = avl_find(&dl->dl_tree, &dle_tofind, &where);
 520         if (dle == NULL)
 521                 dle = avl_nearest(&dl->dl_tree, where, AVL_AFTER);
 522         while (dle) {
 523                 uint64_t used, comp, uncomp;
 524                 dsl_deadlist_entry_t *dle_next;
 525
 526                 bpobj_enqueue_subobj(bpo, dle->dle_bpobj.bpo_object, tx);
 527
 528                 VERIFY3U(0, ==, bpobj_space(&dle->dle_bpobj,
 529                     &used, &comp, &uncomp));
 530                 mutex_enter(&dl->dl_lock);
 531                 ASSERT3U(dl->dl_phys->dl_used, >=, used);
 532                 ASSERT3U(dl->dl_phys->dl_comp, >=, comp);
 533                 ASSERT3U(dl->dl_phys->dl_uncomp, >=, uncomp);
 534                 dl->dl_phys->dl_used -= used;
 535                 dl->dl_phys->dl_comp -= comp;
 536                 dl->dl_phys->dl_uncomp -= uncomp;
 537                 mutex_exit(&dl->dl_lock);
 538
 539                 VERIFY3U(0, ==, zap_remove_int(dl->dl_os, dl->dl_object,
 540                     dle->dle_mintxg, tx));
 541
 542                 dle_next = AVL_NEXT(&dl->dl_tree, dle);
 543                 avl_remove(&dl->dl_tree, dle);
 544                 bpobj_close(&dle->dle_bpobj);
 545                 kmem_free(dle, sizeof (*dle));
 546                 dle = dle_next;
 547         }
 548 }