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