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
spa_features_check(spa_t * spa,boolean_t for_write,nvlist_t * unsup_feat,nvlist_t * enabled_feat)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 * Use an in-memory cache of feature refcounts for quick retrieval.
220 *
221 * Note: well-designed features will not need to use this; they should
222 * use spa_feature_is_enabled() and spa_feature_is_active() instead.
223 * However, this is non-static for zdb and zhack.
224 */
225 int
feature_get_refcount(spa_t * spa,zfeature_info_t * feature,uint64_t * res)226 feature_get_refcount(spa_t *spa, zfeature_info_t *feature, uint64_t *res)
227 {
228 ASSERT(VALID_FEATURE_FID(feature->fi_feature));
229 if (spa->spa_feat_refcount_cache[feature->fi_feature] ==
230 SPA_FEATURE_DISABLED) {
231 return (SET_ERROR(ENOTSUP));
232 }
233 *res = spa->spa_feat_refcount_cache[feature->fi_feature];
234 return (0);
235 }
236
237 /*
238 * Note: well-designed features will not need to use this; they should
239 * use spa_feature_is_enabled() and spa_feature_is_active() instead.
240 * However, this is non-static for zdb and zhack.
241 */
242 int
feature_get_refcount_from_disk(spa_t * spa,zfeature_info_t * feature,uint64_t * res)243 feature_get_refcount_from_disk(spa_t *spa, zfeature_info_t *feature,
244 uint64_t *res)
245 {
246 int err;
247 uint64_t refcount;
248 uint64_t zapobj = (feature->fi_flags & ZFEATURE_FLAG_READONLY_COMPAT) ?
249 spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
250
251 /*
252 * If the pool is currently being created, the feature objects may not
253 * have been allocated yet. Act as though all features are disabled.
254 */
255 if (zapobj == 0)
256 return (SET_ERROR(ENOTSUP));
257
258 err = zap_lookup(spa->spa_meta_objset, zapobj,
259 feature->fi_guid, sizeof (uint64_t), 1, &refcount);
260 if (err != 0) {
261 if (err == ENOENT)
262 return (SET_ERROR(ENOTSUP));
263 else
264 return (err);
265 }
266 *res = refcount;
267 return (0);
268 }
269
270
271 static int
feature_get_enabled_txg(spa_t * spa,zfeature_info_t * feature,uint64_t * res)272 feature_get_enabled_txg(spa_t *spa, zfeature_info_t *feature, uint64_t *res)
273 {
274 uint64_t enabled_txg_obj = spa->spa_feat_enabled_txg_obj;
275
276 ASSERT(zfeature_depends_on(feature->fi_feature,
277 SPA_FEATURE_ENABLED_TXG));
278
279 if (!spa_feature_is_enabled(spa, feature->fi_feature)) {
280 return (SET_ERROR(ENOTSUP));
281 }
282
283 ASSERT(enabled_txg_obj != 0);
284
285 VERIFY0(zap_lookup(spa->spa_meta_objset, spa->spa_feat_enabled_txg_obj,
286 feature->fi_guid, sizeof (uint64_t), 1, res));
287
288 return (0);
289 }
290
291 /*
292 * This function is non-static for zhack; it should otherwise not be used
293 * outside this file.
294 */
295 void
feature_sync(spa_t * spa,zfeature_info_t * feature,uint64_t refcount,dmu_tx_t * tx)296 feature_sync(spa_t *spa, zfeature_info_t *feature, uint64_t refcount,
297 dmu_tx_t *tx)
298 {
299 ASSERT(VALID_FEATURE_OR_NONE(feature->fi_feature));
300 uint64_t zapobj = (feature->fi_flags & ZFEATURE_FLAG_READONLY_COMPAT) ?
301 spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
302
303 VERIFY0(zap_update(spa->spa_meta_objset, zapobj, feature->fi_guid,
304 sizeof (uint64_t), 1, &refcount, tx));
305
306 /*
307 * feature_sync is called directly from zhack, allowing the
308 * creation of arbitrary features whose fi_feature field may
309 * be greater than SPA_FEATURES. When called from zhack, the
310 * zfeature_info_t object's fi_feature field will be set to
311 * SPA_FEATURE_NONE.
312 */
313 if (feature->fi_feature != SPA_FEATURE_NONE) {
314 uint64_t *refcount_cache =
315 &spa->spa_feat_refcount_cache[feature->fi_feature];
316 #ifdef atomic_swap_64
317 VERIFY3U(*refcount_cache, ==,
318 atomic_swap_64(refcount_cache, refcount));
319 #else
320 *refcount_cache = refcount;
321 #endif
322 }
323
324 if (refcount == 0)
325 spa_deactivate_mos_feature(spa, feature->fi_guid);
326 else if (feature->fi_flags & ZFEATURE_FLAG_MOS)
327 spa_activate_mos_feature(spa, feature->fi_guid, tx);
328 }
329
330 /*
331 * This function is non-static for zhack; it should otherwise not be used
332 * outside this file.
333 */
334 void
feature_enable_sync(spa_t * spa,zfeature_info_t * feature,dmu_tx_t * tx)335 feature_enable_sync(spa_t *spa, zfeature_info_t *feature, dmu_tx_t *tx)
336 {
337 uint64_t initial_refcount =
338 (feature->fi_flags & ZFEATURE_FLAG_ACTIVATE_ON_ENABLE) ? 1 : 0;
339 uint64_t zapobj = (feature->fi_flags & ZFEATURE_FLAG_READONLY_COMPAT) ?
340 spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
341
342 ASSERT(0 != zapobj);
343 ASSERT(zfeature_is_valid_guid(feature->fi_guid));
344 ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
345
346 /*
347 * If the feature is already enabled, ignore the request.
348 */
349 if (zap_contains(spa->spa_meta_objset, zapobj, feature->fi_guid) == 0)
350 return;
351
352 for (int i = 0; feature->fi_depends[i] != SPA_FEATURE_NONE; i++)
353 spa_feature_enable(spa, feature->fi_depends[i], tx);
354
355 VERIFY0(zap_update(spa->spa_meta_objset, spa->spa_feat_desc_obj,
356 feature->fi_guid, 1, strlen(feature->fi_desc) + 1,
357 feature->fi_desc, tx));
358
359 feature_sync(spa, feature, initial_refcount, tx);
360
361 if (spa_feature_is_enabled(spa, SPA_FEATURE_ENABLED_TXG)) {
362 uint64_t enabling_txg = dmu_tx_get_txg(tx);
363
364 if (spa->spa_feat_enabled_txg_obj == 0ULL) {
365 spa->spa_feat_enabled_txg_obj =
366 zap_create_link(spa->spa_meta_objset,
367 DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
368 DMU_POOL_FEATURE_ENABLED_TXG, tx);
369 }
370 spa_feature_incr(spa, SPA_FEATURE_ENABLED_TXG, tx);
371
372 VERIFY0(zap_add(spa->spa_meta_objset,
373 spa->spa_feat_enabled_txg_obj, feature->fi_guid,
374 sizeof (uint64_t), 1, &enabling_txg, tx));
375 }
376 }
377
378 static void
feature_do_action(spa_t * spa,spa_feature_t fid,feature_action_t action,dmu_tx_t * tx)379 feature_do_action(spa_t *spa, spa_feature_t fid, feature_action_t action,
380 dmu_tx_t *tx)
381 {
382 uint64_t refcount;
383 zfeature_info_t *feature = &spa_feature_table[fid];
384 uint64_t zapobj = (feature->fi_flags & ZFEATURE_FLAG_READONLY_COMPAT) ?
385 spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
386
387 ASSERT(VALID_FEATURE_FID(fid));
388 ASSERT(0 != zapobj);
389 ASSERT(zfeature_is_valid_guid(feature->fi_guid));
390
391 ASSERT(dmu_tx_is_syncing(tx));
392 ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
393
394 VERIFY3U(feature_get_refcount(spa, feature, &refcount), !=, ENOTSUP);
395
396 switch (action) {
397 case FEATURE_ACTION_INCR:
398 VERIFY3U(refcount, !=, UINT64_MAX);
399 refcount++;
400 break;
401 case FEATURE_ACTION_DECR:
402 VERIFY3U(refcount, !=, 0);
403 refcount--;
404 break;
405 default:
406 ASSERT(0);
407 break;
408 }
409
410 feature_sync(spa, feature, refcount, tx);
411 }
412
413 void
spa_feature_create_zap_objects(spa_t * spa,dmu_tx_t * tx)414 spa_feature_create_zap_objects(spa_t *spa, dmu_tx_t *tx)
415 {
416 /*
417 * We create feature flags ZAP objects in two instances: during pool
418 * creation and during pool upgrade.
419 */
420 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)) || (!spa->spa_sync_on &&
421 tx->tx_txg == TXG_INITIAL));
422
423 spa->spa_feat_for_read_obj = zap_create_link(spa->spa_meta_objset,
424 DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
425 DMU_POOL_FEATURES_FOR_READ, tx);
426 spa->spa_feat_for_write_obj = zap_create_link(spa->spa_meta_objset,
427 DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
428 DMU_POOL_FEATURES_FOR_WRITE, tx);
429 spa->spa_feat_desc_obj = zap_create_link(spa->spa_meta_objset,
430 DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
431 DMU_POOL_FEATURE_DESCRIPTIONS, tx);
432 }
433
434 /*
435 * Enable any required dependencies, then enable the requested feature.
436 */
437 void
spa_feature_enable(spa_t * spa,spa_feature_t fid,dmu_tx_t * tx)438 spa_feature_enable(spa_t *spa, spa_feature_t fid, dmu_tx_t *tx)
439 {
440 ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
441 ASSERT(VALID_FEATURE_FID(fid));
442 feature_enable_sync(spa, &spa_feature_table[fid], tx);
443 }
444
445 void
spa_feature_incr(spa_t * spa,spa_feature_t fid,dmu_tx_t * tx)446 spa_feature_incr(spa_t *spa, spa_feature_t fid, dmu_tx_t *tx)
447 {
448 feature_do_action(spa, fid, FEATURE_ACTION_INCR, tx);
449 }
450
451 void
spa_feature_decr(spa_t * spa,spa_feature_t fid,dmu_tx_t * tx)452 spa_feature_decr(spa_t *spa, spa_feature_t fid, dmu_tx_t *tx)
453 {
454 feature_do_action(spa, fid, FEATURE_ACTION_DECR, tx);
455 }
456
457 boolean_t
spa_feature_is_enabled(spa_t * spa,spa_feature_t fid)458 spa_feature_is_enabled(spa_t *spa, spa_feature_t fid)
459 {
460 int err;
461 uint64_t refcount;
462
463 ASSERT(VALID_FEATURE_FID(fid));
464 if (spa_version(spa) < SPA_VERSION_FEATURES)
465 return (B_FALSE);
466
467 err = feature_get_refcount(spa, &spa_feature_table[fid], &refcount);
468 ASSERT(err == 0 || err == ENOTSUP);
469 return (err == 0);
470 }
471
472 boolean_t
spa_feature_is_active(spa_t * spa,spa_feature_t fid)473 spa_feature_is_active(spa_t *spa, spa_feature_t fid)
474 {
475 int err;
476 uint64_t refcount;
477
478 ASSERT(VALID_FEATURE_FID(fid));
479 if (spa_version(spa) < SPA_VERSION_FEATURES)
480 return (B_FALSE);
481
482 err = feature_get_refcount(spa, &spa_feature_table[fid], &refcount);
483 ASSERT(err == 0 || err == ENOTSUP);
484 return (err == 0 && refcount > 0);
485 }
486
487 /*
488 * For the feature specified by fid (which must depend on
489 * SPA_FEATURE_ENABLED_TXG), return the TXG at which it was enabled in the
490 * OUT txg argument.
491 *
492 * Returns B_TRUE if the feature is enabled, in which case txg will be filled
493 * with the transaction group in which the specified feature was enabled.
494 * Returns B_FALSE otherwise (i.e. if the feature is not enabled).
495 */
496 boolean_t
spa_feature_enabled_txg(spa_t * spa,spa_feature_t fid,uint64_t * txg)497 spa_feature_enabled_txg(spa_t *spa, spa_feature_t fid, uint64_t *txg)
498 {
499 int err;
500
501 ASSERT(VALID_FEATURE_FID(fid));
502 if (spa_version(spa) < SPA_VERSION_FEATURES)
503 return (B_FALSE);
504
505 err = feature_get_enabled_txg(spa, &spa_feature_table[fid], txg);
506 ASSERT(err == 0 || err == ENOTSUP);
507
508 return (err == 0);
509 }
510