module/zfs/zfeature.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 /*
  23  * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
  24  */
  25
  26 #include <sys/zfs_context.h>
  27 #include <sys/zfeature.h>
  28 #include <sys/dmu.h>
  29 #include <sys/nvpair.h>
  30 #include <sys/zap.h>
  31 #include <sys/dmu_tx.h>
  32 #include "zfeature_common.h"
  33 #include <sys/spa_impl.h>
  34
  35 /*
  36  * ZFS Feature Flags
  37  * -----------------
  38  *
  39  * ZFS feature flags are used to provide fine-grained versioning to the ZFS
  40  * on-disk format. Once enabled on a pool feature flags replace the old
  41  * spa_version() number.
  42  *
  43  * Each new on-disk format change will be given a uniquely identifying string
  44  * guid rather than a version number. This avoids the problem of different
  45  * organizations creating new on-disk formats with the same version number. To
  46  * keep feature guids unique they should consist of the reverse dns name of the
  47  * organization which implemented the feature and a short name for the feature,
  48  * separated by a colon (e.g. com.delphix:async_destroy).
  49  *
  50  * Reference Counts
  51  * ----------------
  52  *
  53  * Within each pool features can be in one of three states: disabled, enabled,
  54  * or active. These states are differentiated by a reference count stored on
  55  * disk for each feature:
  56  *
  57  *   1) If there is no reference count stored on disk the feature is disabled.
  58  *   2) If the reference count is 0 a system administrator has enabled the
  59  *      feature, but the feature has not been used yet, so no on-disk
  60  *      format changes have been made.
  61  *   3) If the reference count is greater than 0 the feature is active.
  62  *      The format changes required by the feature are currently on disk.
  63  *      Note that if the feature's format changes are reversed the feature
  64  *      may choose to set its reference count back to 0.
  65  *
  66  * Feature flags makes no differentiation between non-zero reference counts
  67  * for an active feature (e.g. a reference count of 1 means the same thing as a
  68  * reference count of 27834721), but feature implementations may choose to use
  69  * the reference count to store meaningful information. For example, a new RAID
  70  * implementation might set the reference count to the number of vdevs using
  71  * it. If all those disks are removed from the pool the feature goes back to
  72  * having a reference count of 0.
  73  *
  74  * It is the responsibility of the individual features to maintain a non-zero
  75  * reference count as long as the feature's format changes are present on disk.
  76  *
  77  * Dependencies
  78  * ------------
  79  *
  80  * Each feature may depend on other features. The only effect of this
  81  * relationship is that when a feature is enabled all of its dependencies are
  82  * automatically enabled as well. Any future work to support disabling of
  83  * features would need to ensure that features cannot be disabled if other
  84  * enabled features depend on them.
  85  *
  86  * On-disk Format
  87  * --------------
  88  *
  89  * When feature flags are enabled spa_version() is set to SPA_VERSION_FEATURES
  90  * (5000). In order for this to work the pool is automatically upgraded to
  91  * SPA_VERSION_BEFORE_FEATURES (28) first, so all pre-feature flags on disk
  92  * format changes will be in use.
  93  *
  94  * Information about features is stored in 3 ZAP objects in the pool's MOS.
  95  * These objects are linked to by the following names in the pool directory
  96  * object:
  97  *
  98  * 1) features_for_read: feature guid -> reference count
  99  *    Features needed to open the pool for reading.
 100  * 2) features_for_write: feature guid -> reference count
 101  *    Features needed to open the pool for writing.
 102  * 3) feature_descriptions: feature guid -> descriptive string
 103  *    A human readable string.
 104  *
 105  * All enabled features appear in either features_for_read or
 106  * features_for_write, but not both.
 107  *
 108  * To open a pool in read-only mode only the features listed in
 109  * features_for_read need to be supported.
 110  *
 111  * To open the pool in read-write mode features in both features_for_read and
 112  * features_for_write need to be supported.
 113  *
 114  * Some features may be required to read the ZAP objects containing feature
 115  * information. To allow software to check for compatibility with these features
 116  * before the pool is opened their names must be stored in the label in a
 117  * new "features_for_read" entry (note that features that are only required
 118  * to write to a pool never need to be stored in the label since the
 119  * features_for_write ZAP object can be read before the pool is written to).
 120  * To save space in the label features must be explicitly marked as needing to
 121  * be written to the label. Also, reference counts are not stored in the label,
 122  * instead any feature whose reference count drops to 0 is removed from the
 123  * label.
 124  *
 125  * Adding New Features
 126  * -------------------
 127  *
 128  * Features must be registered in zpool_feature_init() function in
 129  * zfeature_common.c using the zfeature_register() function. This function
 130  * has arguments to specify if the feature should be stored in the
 131  * features_for_read or features_for_write ZAP object and if it needs to be
 132  * written to the label when active.
 133  *
 134  * Once a feature is registered it will appear as a "feature@<feature name>"
 135  * property which can be set by an administrator. Feature implementors should
 136  * use the spa_feature_is_enabled() and spa_feature_is_active() functions to
 137  * query the state of a feature and the spa_feature_incr() and
 138  * spa_feature_decr() functions to change an enabled feature's reference count.
 139  * Reference counts may only be updated in the syncing context.
 140  *
 141  * Features may not perform enable-time initialization. Instead, any such
 142  * initialization should occur when the feature is first used. This design
 143  * enforces that on-disk changes be made only when features are used. Code
 144  * should only check if a feature is enabled using spa_feature_is_enabled(),
 145  * not by relying on any feature specific metadata existing. If a feature is
 146  * enabled, but the feature's metadata is not on disk yet then it should be
 147  * created as needed.
 148  *
 149  * As an example, consider the com.delphix:async_destroy feature. This feature
 150  * relies on the existence of a bptree in the MOS that store blocks for
 151  * asynchronous freeing. This bptree is not created when async_destroy is
 152  * enabled. Instead, when a dataset is destroyed spa_feature_is_enabled() is
 153  * called to check if async_destroy is enabled. If it is and the bptree object
 154  * does not exist yet, the bptree object is created as part of the dataset
 155  * destroy and async_destroy's reference count is incremented to indicate it
 156  * has made an on-disk format change. Later, after the destroyed dataset's
 157  * blocks have all been asynchronously freed there is no longer any use for the
 158  * bptree object, so it is destroyed and async_destroy's reference count is
 159  * decremented back to 0 to indicate that it has undone its on-disk format
 160  * changes.
 161  */
 162
 163 typedef enum {
 164         FEATURE_ACTION_INCR,
 165         FEATURE_ACTION_DECR,
 166 } feature_action_t;
 167
 168 /*
 169  * Checks that the active features in the pool are supported by
 170  * this software.  Adds each unsupported feature (name -> description) to
 171  * the supplied nvlist.
 172  */
 173 boolean_t
 174 spa_features_check(spa_t *spa, boolean_t for_write,
 175     nvlist_t *unsup_feat, nvlist_t *enabled_feat)
 176 {
 177         objset_t *os = spa->spa_meta_objset;
 178         boolean_t supported;
 179         zap_cursor_t *zc;
 180         zap_attribute_t *za;
 181         uint64_t obj = for_write ?
 182             spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
 183         char *buf;
 184
 185         zc = kmem_alloc(sizeof (zap_cursor_t), KM_SLEEP);
 186         za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
 187         buf = kmem_alloc(MAXPATHLEN, KM_SLEEP);
 188
 189         supported = B_TRUE;
 190         for (zap_cursor_init(zc, os, obj);
 191             zap_cursor_retrieve(zc, za) == 0;
 192             zap_cursor_advance(zc)) {
 193                 ASSERT(za->za_integer_length == sizeof (uint64_t) &&
 194                     za->za_num_integers == 1);
 195
 196                 if (NULL != enabled_feat) {
 197                         fnvlist_add_uint64(enabled_feat, za->za_name,
 198                             za->za_first_integer);
 199                 }
 200
 201                 if (za->za_first_integer != 0 &&
 202                     !zfeature_is_supported(za->za_name)) {
 203                         supported = B_FALSE;
 204
 205                         if (NULL != unsup_feat) {
 206                                 char *desc = "";
 207
 208                                 if (zap_lookup(os, spa->spa_feat_desc_obj,
 209                                     za->za_name, 1, MAXPATHLEN, buf) == 0)
 210                                         desc = buf;
 211
 212                                 VERIFY(nvlist_add_string(unsup_feat,
 213                                     za->za_name, desc) == 0);
 214                         }
 215                 }
 216         }
 217         zap_cursor_fini(zc);
 218
 219         kmem_free(buf, MAXPATHLEN);
 220         kmem_free(za, sizeof (zap_attribute_t));
 221         kmem_free(zc, sizeof (zap_cursor_t));
 222
 223         return (supported);
 224 }
 225
 226 /*
 227  * Use an in-memory cache of feature refcounts for quick retrieval.
 228  *
 229  * Note: well-designed features will not need to use this; they should
 230  * use spa_feature_is_enabled() and spa_feature_is_active() instead.
 231  * However, this is non-static for zdb, zhack, and spa_add_feature_stats().
 232  */
 233 int
 234 feature_get_refcount(spa_t *spa, zfeature_info_t *feature, uint64_t *res)
 235 {
 236         ASSERT(VALID_FEATURE_FID(feature->fi_feature));
 237         if (spa->spa_feat_refcount_cache[feature->fi_feature] ==
 238             SPA_FEATURE_DISABLED) {
 239                 return (SET_ERROR(ENOTSUP));
 240         }
 241         *res = spa->spa_feat_refcount_cache[feature->fi_feature];
 242         return (0);
 243 }
 244
 245 /*
 246  * Note: well-designed features will not need to use this; they should
 247  * use spa_feature_is_enabled() and spa_feature_is_active() instead.
 248  * However, this is non-static for zdb and zhack.
 249  */
 250 int
 251 feature_get_refcount_from_disk(spa_t *spa, zfeature_info_t *feature,
 252     uint64_t *res)
 253 {
 254         int err;
 255         uint64_t refcount;
 256         uint64_t zapobj = (feature->fi_flags & ZFEATURE_FLAG_READONLY_COMPAT) ?
 257             spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
 258
 259         /*
 260          * If the pool is currently being created, the feature objects may not
 261          * have been allocated yet.  Act as though all features are disabled.
 262          */
 263         if (zapobj == 0)
 264                 return (SET_ERROR(ENOTSUP));
 265
 266         err = zap_lookup(spa->spa_meta_objset, zapobj,
 267             feature->fi_guid, sizeof (uint64_t), 1, &refcount);
 268         if (err != 0) {
 269                 if (err == ENOENT)
 270                         return (SET_ERROR(ENOTSUP));
 271                 else
 272                         return (err);
 273         }
 274         *res = refcount;
 275         return (0);
 276 }
 277
 278
 279 static int
 280 feature_get_enabled_txg(spa_t *spa, zfeature_info_t *feature, uint64_t *res) {
 281         ASSERTV(uint64_t enabled_txg_obj = spa->spa_feat_enabled_txg_obj);
 282
 283         ASSERT(zfeature_depends_on(feature->fi_feature,
 284             SPA_FEATURE_ENABLED_TXG));
 285
 286         if (!spa_feature_is_enabled(spa, feature->fi_feature)) {
 287                 return (SET_ERROR(ENOTSUP));
 288         }
 289
 290         ASSERT(enabled_txg_obj != 0);
 291
 292         VERIFY0(zap_lookup(spa->spa_meta_objset, spa->spa_feat_enabled_txg_obj,
 293             feature->fi_guid, sizeof (uint64_t), 1, res));
 294
 295         return (0);
 296 }
 297
 298 /*
 299  * This function is non-static for zhack; it should otherwise not be used
 300  * outside this file.
 301  */
 302 void
 303 feature_sync(spa_t *spa, zfeature_info_t *feature, uint64_t refcount,
 304     dmu_tx_t *tx)
 305 {
 306         uint64_t zapobj;
 307
 308         ASSERT(VALID_FEATURE_OR_NONE(feature->fi_feature));
 309         zapobj = (feature->fi_flags & ZFEATURE_FLAG_READONLY_COMPAT) ?
 310             spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
 311         VERIFY0(zap_update(spa->spa_meta_objset, zapobj, feature->fi_guid,
 312             sizeof (uint64_t), 1, &refcount, tx));
 313
 314         /*
 315          * feature_sync is called directly from zhack, allowing the
 316          * creation of arbitrary features whose fi_feature field may
 317          * be greater than SPA_FEATURES. When called from zhack, the
 318          * zfeature_info_t object's fi_feature field will be set to
 319          * SPA_FEATURE_NONE.
 320          */
 321         if (feature->fi_feature != SPA_FEATURE_NONE) {
 322                 uint64_t *refcount_cache =
 323                     &spa->spa_feat_refcount_cache[feature->fi_feature];
 324                 VERIFY3U(*refcount_cache, ==,
 325                     atomic_swap_64(refcount_cache, refcount));
 326         }
 327
 328         if (refcount == 0)
 329                 spa_deactivate_mos_feature(spa, feature->fi_guid);
 330         else if (feature->fi_flags & ZFEATURE_FLAG_MOS)
 331                 spa_activate_mos_feature(spa, feature->fi_guid, tx);
 332 }
 333
 334 /*
 335  * This function is non-static for zhack; it should otherwise not be used
 336  * outside this file.
 337  */
 338 void
 339 feature_enable_sync(spa_t *spa, zfeature_info_t *feature, dmu_tx_t *tx)
 340 {
 341         uint64_t initial_refcount =
 342             (feature->fi_flags & ZFEATURE_FLAG_ACTIVATE_ON_ENABLE) ? 1 : 0;
 343         uint64_t zapobj = (feature->fi_flags & ZFEATURE_FLAG_READONLY_COMPAT) ?
 344             spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
 345         int i;
 346
 347         ASSERT(0 != zapobj);
 348         ASSERT(zfeature_is_valid_guid(feature->fi_guid));
 349         ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
 350
 351         /*
 352          * If the feature is already enabled, ignore the request.
 353          */
 354         if (zap_contains(spa->spa_meta_objset, zapobj, feature->fi_guid) == 0)
 355                 return;
 356
 357         for (i = 0; feature->fi_depends[i] != SPA_FEATURE_NONE; i++)
 358                 spa_feature_enable(spa, feature->fi_depends[i], tx);
 359
 360         VERIFY0(zap_update(spa->spa_meta_objset, spa->spa_feat_desc_obj,
 361             feature->fi_guid, 1, strlen(feature->fi_desc) + 1,
 362             feature->fi_desc, tx));
 363
 364         feature_sync(spa, feature, initial_refcount, tx);
 365
 366         if (spa_feature_is_enabled(spa, SPA_FEATURE_ENABLED_TXG)) {
 367                 uint64_t enabling_txg = dmu_tx_get_txg(tx);
 368
 369                 if (spa->spa_feat_enabled_txg_obj == 0ULL) {
 370                         spa->spa_feat_enabled_txg_obj =
 371                             zap_create_link(spa->spa_meta_objset,
 372                             DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
 373                             DMU_POOL_FEATURE_ENABLED_TXG, tx);
 374                 }
 375                 spa_feature_incr(spa, SPA_FEATURE_ENABLED_TXG, tx);
 376
 377                 VERIFY0(zap_add(spa->spa_meta_objset,
 378                     spa->spa_feat_enabled_txg_obj, feature->fi_guid,
 379                     sizeof (uint64_t), 1, &enabling_txg, tx));
 380         }
 381 }
 382
 383 static void
 384 feature_do_action(spa_t *spa, spa_feature_t fid, feature_action_t action,
 385     dmu_tx_t *tx)
 386 {
 387         uint64_t refcount = 0;
 388         zfeature_info_t *feature = &spa_feature_table[fid];
 389         ASSERTV(uint64_t zapobj =
 390             (feature->fi_flags & ZFEATURE_FLAG_READONLY_COMPAT) ?
 391             spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj);
 392
 393         ASSERT(VALID_FEATURE_FID(fid));
 394         ASSERT(0 != zapobj);
 395         ASSERT(zfeature_is_valid_guid(feature->fi_guid));
 396
 397         ASSERT(dmu_tx_is_syncing(tx));
 398         ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
 399
 400         VERIFY3U(feature_get_refcount(spa, feature, &refcount), !=, ENOTSUP);
 401
 402         switch (action) {
 403         case FEATURE_ACTION_INCR:
 404                 VERIFY3U(refcount, !=, UINT64_MAX);
 405                 refcount++;
 406                 break;
 407         case FEATURE_ACTION_DECR:
 408                 VERIFY3U(refcount, !=, 0);
 409                 refcount--;
 410                 break;
 411         default:
 412                 ASSERT(0);
 413                 break;
 414         }
 415
 416         feature_sync(spa, feature, refcount, tx);
 417 }
 418
 419 void
 420 spa_feature_create_zap_objects(spa_t *spa, dmu_tx_t *tx)
 421 {
 422         /*
 423          * We create feature flags ZAP objects in two instances: during pool
 424          * creation and during pool upgrade.
 425          */
 426         ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)) || (!spa->spa_sync_on &&
 427             tx->tx_txg == TXG_INITIAL));
 428
 429         spa->spa_feat_for_read_obj = zap_create_link(spa->spa_meta_objset,
 430             DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
 431             DMU_POOL_FEATURES_FOR_READ, tx);
 432         spa->spa_feat_for_write_obj = zap_create_link(spa->spa_meta_objset,
 433             DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
 434             DMU_POOL_FEATURES_FOR_WRITE, tx);
 435         spa->spa_feat_desc_obj = zap_create_link(spa->spa_meta_objset,
 436             DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
 437             DMU_POOL_FEATURE_DESCRIPTIONS, tx);
 438 }
 439
 440 /*
 441  * Enable any required dependencies, then enable the requested feature.
 442  */
 443 void
 444 spa_feature_enable(spa_t *spa, spa_feature_t fid, dmu_tx_t *tx)
 445 {
 446         ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
 447         ASSERT(VALID_FEATURE_FID(fid));
 448         feature_enable_sync(spa, &spa_feature_table[fid], tx);
 449 }
 450
 451 void
 452 spa_feature_incr(spa_t *spa, spa_feature_t fid, dmu_tx_t *tx)
 453 {
 454         feature_do_action(spa, fid, FEATURE_ACTION_INCR, tx);
 455 }
 456
 457 void
 458 spa_feature_decr(spa_t *spa, spa_feature_t fid, dmu_tx_t *tx)
 459 {
 460         feature_do_action(spa, fid, FEATURE_ACTION_DECR, tx);
 461 }
 462
 463 boolean_t
 464 spa_feature_is_enabled(spa_t *spa, spa_feature_t fid)
 465 {
 466         int err;
 467         uint64_t refcount = 0;
 468
 469         ASSERT(VALID_FEATURE_FID(fid));
 470         if (spa_version(spa) < SPA_VERSION_FEATURES)
 471                 return (B_FALSE);
 472
 473         err = feature_get_refcount(spa, &spa_feature_table[fid], &refcount);
 474         ASSERT(err == 0 || err == ENOTSUP);
 475         return (err == 0);
 476 }
 477
 478 boolean_t
 479 spa_feature_is_active(spa_t *spa, spa_feature_t fid)
 480 {
 481         int err;
 482         uint64_t refcount = 0;
 483
 484         ASSERT(VALID_FEATURE_FID(fid));
 485         if (spa_version(spa) < SPA_VERSION_FEATURES)
 486                 return (B_FALSE);
 487
 488         err = feature_get_refcount(spa, &spa_feature_table[fid], &refcount);
 489         ASSERT(err == 0 || err == ENOTSUP);
 490         return (err == 0 && refcount > 0);
 491 }
 492
 493 /*
 494  * For the feature specified by fid (which must depend on
 495  * SPA_FEATURE_ENABLED_TXG), return the TXG at which it was enabled in the
 496  * OUT txg argument.
 497  *
 498  * Returns B_TRUE if the feature is enabled, in which case txg will be filled
 499  * with the transaction group in which the specified feature was enabled.
 500  * Returns B_FALSE otherwise (i.e. if the feature is not enabled).
 501  */
 502 boolean_t
 503 spa_feature_enabled_txg(spa_t *spa, spa_feature_t fid, uint64_t *txg) {
 504         int err;
 505
 506         ASSERT(VALID_FEATURE_FID(fid));
 507         if (spa_version(spa) < SPA_VERSION_FEATURES)
 508                 return (B_FALSE);
 509
 510         err = feature_get_enabled_txg(spa, &spa_feature_table[fid], txg);
 511         ASSERT(err == 0 || err == ENOTSUP);
 512
 513         return (err == 0);
 514 }