xref: /illumos-gate/usr/src/uts/common/fs/zfs/zfeature.c (revision 2acef22db7808606888f8f92715629ff3ba555b9)
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) 2013 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 
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 }
408