Print this page
4171 clean up spa_feature_*() interfaces
4172 implement extensible_dataset feature for use by other zpool features
Reviewed by: Max Grossman <max.grossman@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>


 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_ENABLE,
 165         FEATURE_ACTION_INCR,
 166         FEATURE_ACTION_DECR,
 167 } feature_action_t;
 168 
 169 /*
 170  * Checks that the features active in the specified object are supported by
 171  * this software.  Adds each unsupported feature (name -> description) to
 172  * the supplied nvlist.
 173  */
 174 boolean_t
 175 feature_is_supported(objset_t *os, uint64_t obj, uint64_t desc_obj,
 176     nvlist_t *unsup_feat, nvlist_t *enabled_feat)
 177 {

 178         boolean_t supported;
 179         zap_cursor_t zc;
 180         zap_attribute_t za;


 181 
 182         supported = B_TRUE;
 183         for (zap_cursor_init(&zc, os, obj);
 184             zap_cursor_retrieve(&zc, &za) == 0;
 185             zap_cursor_advance(&zc)) {
 186                 ASSERT(za.za_integer_length == sizeof (uint64_t) &&
 187                     za.za_num_integers == 1);
 188 
 189                 if (NULL != enabled_feat) {
 190                         fnvlist_add_uint64(enabled_feat, za.za_name,
 191                             za.za_first_integer);
 192                 }
 193 
 194                 if (za.za_first_integer != 0 &&
 195                     !zfeature_is_supported(za.za_name)) {
 196                         supported = B_FALSE;
 197 
 198                         if (NULL != unsup_feat) {
 199                                 char *desc = "";
 200                                 char buf[MAXPATHLEN];
 201 
 202                                 if (zap_lookup(os, desc_obj, za.za_name,
 203                                     1, sizeof (buf), buf) == 0)
 204                                         desc = buf;
 205 
 206                                 VERIFY(nvlist_add_string(unsup_feat, za.za_name,
 207                                     desc) == 0);
 208                         }
 209                 }
 210         }
 211         zap_cursor_fini(&zc);
 212 
 213         return (supported);
 214 }
 215 
 216 static int
 217 feature_get_refcount(objset_t *os, uint64_t read_obj, uint64_t write_obj,
 218     zfeature_info_t *feature, uint64_t *res)




 219 {
 220         int err;
 221         uint64_t refcount;
 222         uint64_t zapobj = feature->fi_can_readonly ? write_obj : read_obj;

 223 
 224         /*
 225          * If the pool is currently being created, the feature objects may not
 226          * have been allocated yet.  Act as though all features are disabled.
 227          */
 228         if (zapobj == 0)
 229                 return (SET_ERROR(ENOTSUP));
 230 
 231         err = zap_lookup(os, zapobj, feature->fi_guid, sizeof (uint64_t), 1,
 232             &refcount);
 233         if (err != 0) {
 234                 if (err == ENOENT)
 235                         return (SET_ERROR(ENOTSUP));
 236                 else
 237                         return (err);
 238         }
 239         *res = refcount;
 240         return (0);
 241 }
 242 
 243 static int
 244 feature_do_action(objset_t *os, uint64_t read_obj, uint64_t write_obj,
 245     uint64_t desc_obj, zfeature_info_t *feature, feature_action_t action,



 246     dmu_tx_t *tx)
 247 {
 248         int error;
 249         uint64_t refcount;
 250         uint64_t zapobj = feature->fi_can_readonly ? write_obj : read_obj;
 251 
 252         ASSERT(0 != zapobj);
 253         ASSERT(zfeature_is_valid_guid(feature->fi_guid));
 254 
 255         error = zap_lookup(os, zapobj, feature->fi_guid,
 256             sizeof (uint64_t), 1, &refcount);



 257 
 258         /*
 259          * If we can't ascertain the status of the specified feature, an I/O
 260          * error occurred.
 261          */
 262         if (error != 0 && error != ENOENT)
 263                 return (error);



 264 
 265         switch (action) {
 266         case FEATURE_ACTION_ENABLE:


 267                 /*
 268                  * If the feature is already enabled, ignore the request.
 269                  */
 270                 if (error == 0)
 271                         return (0);
 272                 refcount = 0;
 273                 break;




























 274         case FEATURE_ACTION_INCR:
 275                 if (error == ENOENT)
 276                         return (SET_ERROR(ENOTSUP));
 277                 if (refcount == UINT64_MAX)
 278                         return (SET_ERROR(EOVERFLOW));
 279                 refcount++;
 280                 break;
 281         case FEATURE_ACTION_DECR:
 282                 if (error == ENOENT)
 283                         return (SET_ERROR(ENOTSUP));
 284                 if (refcount == 0)
 285                         return (SET_ERROR(EOVERFLOW));
 286                 refcount--;
 287                 break;
 288         default:
 289                 ASSERT(0);
 290                 break;
 291         }
 292 
 293         if (action == FEATURE_ACTION_ENABLE) {
 294                 int i;
 295 
 296                 for (i = 0; feature->fi_depends[i] != NULL; i++) {
 297                         zfeature_info_t *dep = feature->fi_depends[i];
 298 
 299                         error = feature_do_action(os, read_obj, write_obj,
 300                             desc_obj, dep, FEATURE_ACTION_ENABLE, tx);
 301                         if (error != 0)
 302                                 return (error);
 303                 }
 304         }
 305 
 306         error = zap_update(os, zapobj, feature->fi_guid,
 307             sizeof (uint64_t), 1, &refcount, tx);
 308         if (error != 0)
 309                 return (error);
 310 
 311         if (action == FEATURE_ACTION_ENABLE) {
 312                 error = zap_update(os, desc_obj,
 313                     feature->fi_guid, 1, strlen(feature->fi_desc) + 1,
 314                     feature->fi_desc, tx);
 315                 if (error != 0)
 316                         return (error);
 317         }
 318 
 319         if (action == FEATURE_ACTION_INCR && refcount == 1 && feature->fi_mos) {
 320                 spa_activate_mos_feature(dmu_objset_spa(os), feature->fi_guid);
 321         }
 322 
 323         if (action == FEATURE_ACTION_DECR && refcount == 0) {
 324                 spa_deactivate_mos_feature(dmu_objset_spa(os),
 325                     feature->fi_guid);
 326         }
 327 
 328         return (0);
 329 }
 330 
 331 void
 332 spa_feature_create_zap_objects(spa_t *spa, dmu_tx_t *tx)
 333 {
 334         /*
 335          * We create feature flags ZAP objects in two instances: during pool
 336          * creation and during pool upgrade.
 337          */
 338         ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)) || (!spa->spa_sync_on &&
 339             tx->tx_txg == TXG_INITIAL));
 340 
 341         spa->spa_feat_for_read_obj = zap_create_link(spa->spa_meta_objset,
 342             DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
 343             DMU_POOL_FEATURES_FOR_READ, tx);
 344         spa->spa_feat_for_write_obj = zap_create_link(spa->spa_meta_objset,
 345             DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
 346             DMU_POOL_FEATURES_FOR_WRITE, tx);
 347         spa->spa_feat_desc_obj = zap_create_link(spa->spa_meta_objset,
 348             DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
 349             DMU_POOL_FEATURE_DESCRIPTIONS, tx);
 350 }
 351 
 352 /*
 353  * Enable any required dependencies, then enable the requested feature.
 354  */
 355 void
 356 spa_feature_enable(spa_t *spa, zfeature_info_t *feature, dmu_tx_t *tx)
 357 {
 358         ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
 359         VERIFY3U(0, ==, feature_do_action(spa->spa_meta_objset,
 360             spa->spa_feat_for_read_obj, spa->spa_feat_for_write_obj,
 361             spa->spa_feat_desc_obj, feature, FEATURE_ACTION_ENABLE, tx));
 362 }
 363 
 364 void
 365 spa_feature_incr(spa_t *spa, zfeature_info_t *feature, dmu_tx_t *tx)
 366 {
 367         ASSERT(dmu_tx_is_syncing(tx));
 368         ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
 369         VERIFY3U(0, ==, feature_do_action(spa->spa_meta_objset,
 370             spa->spa_feat_for_read_obj, spa->spa_feat_for_write_obj,
 371             spa->spa_feat_desc_obj, feature, FEATURE_ACTION_INCR, tx));
 372 }
 373 
 374 void
 375 spa_feature_decr(spa_t *spa, zfeature_info_t *feature, dmu_tx_t *tx)
 376 {
 377         ASSERT(dmu_tx_is_syncing(tx));
 378         ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
 379         VERIFY3U(0, ==, feature_do_action(spa->spa_meta_objset,
 380             spa->spa_feat_for_read_obj, spa->spa_feat_for_write_obj,
 381             spa->spa_feat_desc_obj, feature, FEATURE_ACTION_DECR, tx));
 382 }
 383 
 384 /*
 385  * This interface is for debugging only. Normal consumers should use
 386  * spa_feature_is_enabled/spa_feature_is_active.
 387  */
 388 int
 389 spa_feature_get_refcount(spa_t *spa, zfeature_info_t *feature)
 390 {
 391         int err;
 392         uint64_t refcount;
 393 
 394         if (spa_version(spa) < SPA_VERSION_FEATURES)
 395                 return (B_FALSE);
 396 
 397         err = feature_get_refcount(spa->spa_meta_objset,
 398             spa->spa_feat_for_read_obj, spa->spa_feat_for_write_obj,
 399             feature, &refcount);
 400         ASSERT(err == 0 || err == ENOTSUP);
 401         return (err == 0 ? refcount : 0);
 402 }
 403 
 404 boolean_t
 405 spa_feature_is_enabled(spa_t *spa, zfeature_info_t *feature)
 406 {
 407         int err;
 408         uint64_t refcount;
 409 

 410         if (spa_version(spa) < SPA_VERSION_FEATURES)
 411                 return (B_FALSE);
 412 
 413         err = feature_get_refcount(spa->spa_meta_objset,
 414             spa->spa_feat_for_read_obj, spa->spa_feat_for_write_obj,
 415             feature, &refcount);
 416         ASSERT(err == 0 || err == ENOTSUP);
 417         return (err == 0);
 418 }
 419 
 420 boolean_t
 421 spa_feature_is_active(spa_t *spa, zfeature_info_t *feature)
 422 {
 423         int err;
 424         uint64_t refcount;
 425 

 426         if (spa_version(spa) < SPA_VERSION_FEATURES)
 427                 return (B_FALSE);
 428 
 429         err = feature_get_refcount(spa->spa_meta_objset,
 430             spa->spa_feat_for_read_obj, spa->spa_feat_for_write_obj,
 431             feature, &refcount);
 432         ASSERT(err == 0 || err == ENOTSUP);
 433         return (err == 0 && refcount > 0);
 434 }


 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 
 184         supported = B_TRUE;
 185         for (zap_cursor_init(&zc, os, obj);
 186             zap_cursor_retrieve(&zc, &za) == 0;
 187             zap_cursor_advance(&zc)) {
 188                 ASSERT(za.za_integer_length == sizeof (uint64_t) &&
 189                     za.za_num_integers == 1);
 190 
 191                 if (NULL != enabled_feat) {
 192                         fnvlist_add_uint64(enabled_feat, za.za_name,
 193                             za.za_first_integer);
 194                 }
 195 
 196                 if (za.za_first_integer != 0 &&
 197                     !zfeature_is_supported(za.za_name)) {
 198                         supported = B_FALSE;
 199 
 200                         if (NULL != unsup_feat) {
 201                                 char *desc = "";
 202                                 char buf[MAXPATHLEN];
 203 
 204                                 if (zap_lookup(os, spa->spa_feat_desc_obj,
 205                                     za.za_name, 1, sizeof (buf), buf) == 0)
 206                                         desc = buf;
 207 
 208                                 VERIFY(nvlist_add_string(unsup_feat, za.za_name,
 209                                     desc) == 0);
 210                         }
 211                 }
 212         }
 213         zap_cursor_fini(&zc);
 214 
 215         return (supported);
 216 }
 217 
 218 /*
 219  * Note: well-designed features will not need to use this; they should
 220  * use spa_feature_is_enabled() and spa_feature_is_active() instead.
 221  * However, this is non-static for zdb and zhack.
 222  */
 223 int
 224 feature_get_refcount(spa_t *spa, zfeature_info_t *feature, uint64_t *res)
 225 {
 226         int err;
 227         uint64_t refcount;
 228         uint64_t zapobj = feature->fi_can_readonly ?
 229             spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
 230 
 231         /*
 232          * If the pool is currently being created, the feature objects may not
 233          * have been allocated yet.  Act as though all features are disabled.
 234          */
 235         if (zapobj == 0)
 236                 return (SET_ERROR(ENOTSUP));
 237 
 238         err = zap_lookup(spa->spa_meta_objset, zapobj,
 239             feature->fi_guid, sizeof (uint64_t), 1, &refcount);
 240         if (err != 0) {
 241                 if (err == ENOENT)
 242                         return (SET_ERROR(ENOTSUP));
 243                 else
 244                         return (err);
 245         }
 246         *res = refcount;
 247         return (0);
 248 }
 249 
 250 /*
 251  * This function is non-static for zhack; it should otherwise not be used
 252  * outside this file.
 253  */
 254 void
 255 feature_sync(spa_t *spa, zfeature_info_t *feature, uint64_t refcount,
 256     dmu_tx_t *tx)
 257 {
 258         uint64_t zapobj = feature->fi_can_readonly ?
 259             spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;

 260 
 261         VERIFY0(zap_update(spa->spa_meta_objset, zapobj, feature->fi_guid,
 262             sizeof (uint64_t), 1, &refcount, tx));
 263 
 264         if (refcount == 0)
 265                 spa_deactivate_mos_feature(spa, feature->fi_guid);
 266         else if (feature->fi_mos)
 267                 spa_activate_mos_feature(spa, feature->fi_guid);
 268 }
 269 
 270 /*
 271  * This function is non-static for zhack; it should otherwise not be used
 272  * outside this file.
 273  */
 274 void
 275 feature_enable_sync(spa_t *spa, zfeature_info_t *feature, dmu_tx_t *tx)
 276 {
 277         uint64_t zapobj = feature->fi_can_readonly ?
 278             spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
 279 
 280         ASSERT(0 != zapobj);
 281         ASSERT(zfeature_is_valid_guid(feature->fi_guid));
 282         ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
 283 
 284         /*
 285          * If the feature is already enabled, ignore the request.
 286          */
 287         if (zap_contains(spa->spa_meta_objset, zapobj, feature->fi_guid) == 0)
 288                 return;
 289 
 290         for (int i = 0; feature->fi_depends[i] != SPA_FEATURE_NONE; i++)
 291                 spa_feature_enable(spa, feature->fi_depends[i], tx);
 292 
 293         VERIFY0(zap_update(spa->spa_meta_objset, spa->spa_feat_desc_obj,
 294             feature->fi_guid, 1, strlen(feature->fi_desc) + 1,
 295             feature->fi_desc, tx));
 296         feature_sync(spa, feature, 0, tx);
 297 }
 298 
 299 static void
 300 feature_do_action(spa_t *spa, spa_feature_t fid, feature_action_t action,
 301     dmu_tx_t *tx)
 302 {
 303         uint64_t refcount;
 304         zfeature_info_t *feature = &spa_feature_table[fid];
 305         uint64_t zapobj = feature->fi_can_readonly ?
 306             spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
 307 
 308         ASSERT3U(fid, <, SPA_FEATURES);
 309         ASSERT(0 != zapobj);
 310         ASSERT(zfeature_is_valid_guid(feature->fi_guid));
 311 
 312         ASSERT(dmu_tx_is_syncing(tx));
 313         ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
 314 
 315         VERIFY0(zap_lookup(spa->spa_meta_objset, zapobj, feature->fi_guid,
 316             sizeof (uint64_t), 1, &refcount));
 317 
 318         switch (action) {
 319         case FEATURE_ACTION_INCR:
 320                 VERIFY3U(refcount, !=, UINT64_MAX);



 321                 refcount++;
 322                 break;
 323         case FEATURE_ACTION_DECR:
 324                 VERIFY3U(refcount, !=, 0);



 325                 refcount--;
 326                 break;
 327         default:
 328                 ASSERT(0);
 329                 break;
 330         }
 331 
 332         feature_sync(spa, feature, refcount, tx);



































 333 }
 334 
 335 void
 336 spa_feature_create_zap_objects(spa_t *spa, dmu_tx_t *tx)
 337 {
 338         /*
 339          * We create feature flags ZAP objects in two instances: during pool
 340          * creation and during pool upgrade.
 341          */
 342         ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)) || (!spa->spa_sync_on &&
 343             tx->tx_txg == TXG_INITIAL));
 344 
 345         spa->spa_feat_for_read_obj = zap_create_link(spa->spa_meta_objset,
 346             DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
 347             DMU_POOL_FEATURES_FOR_READ, tx);
 348         spa->spa_feat_for_write_obj = zap_create_link(spa->spa_meta_objset,
 349             DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
 350             DMU_POOL_FEATURES_FOR_WRITE, tx);
 351         spa->spa_feat_desc_obj = zap_create_link(spa->spa_meta_objset,
 352             DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
 353             DMU_POOL_FEATURE_DESCRIPTIONS, tx);
 354 }
 355 
 356 /*
 357  * Enable any required dependencies, then enable the requested feature.
 358  */
 359 void
 360 spa_feature_enable(spa_t *spa, spa_feature_t fid, dmu_tx_t *tx)
 361 {
 362         ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
 363         ASSERT3U(fid, <, SPA_FEATURES);
 364         feature_enable_sync(spa, &spa_feature_table[fid], tx);

 365 }
 366 
 367 void
 368 spa_feature_incr(spa_t *spa, spa_feature_t fid, dmu_tx_t *tx)
 369 {
 370         feature_do_action(spa, fid, FEATURE_ACTION_INCR, tx);




 371 }
 372 
 373 void
 374 spa_feature_decr(spa_t *spa, spa_feature_t fid, dmu_tx_t *tx)
 375 {
 376         feature_do_action(spa, fid, FEATURE_ACTION_DECR, tx);




 377 }
 378 




















 379 boolean_t
 380 spa_feature_is_enabled(spa_t *spa, spa_feature_t fid)
 381 {
 382         int err;
 383         uint64_t refcount;
 384 
 385         ASSERT3U(fid, <, SPA_FEATURES);
 386         if (spa_version(spa) < SPA_VERSION_FEATURES)
 387                 return (B_FALSE);
 388 
 389         err = feature_get_refcount(spa, &spa_feature_table[fid], &refcount);


 390         ASSERT(err == 0 || err == ENOTSUP);
 391         return (err == 0);
 392 }
 393 
 394 boolean_t
 395 spa_feature_is_active(spa_t *spa, spa_feature_t fid)
 396 {
 397         int err;
 398         uint64_t refcount;
 399 
 400         ASSERT3U(fid, <, SPA_FEATURES);
 401         if (spa_version(spa) < SPA_VERSION_FEATURES)
 402                 return (B_FALSE);
 403 
 404         err = feature_get_refcount(spa, &spa_feature_table[fid], &refcount);


 405         ASSERT(err == 0 || err == ENOTSUP);
 406         return (err == 0 && refcount > 0);
 407 }