]> git.proxmox.com Git - mirror_zfs.git/blame - module/zfs/zfeature.c
Illumos #4101, #4102, #4103, #4105, #4106
[mirror_zfs.git] / module / zfs / zfeature.c
CommitLineData
9ae529ec
CS
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/*
2e528b49 23 * Copyright (c) 2013 by Delphix. All rights reserved.
9ae529ec
CS
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
163typedef 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 */
174boolean_t
175feature_is_supported(objset_t *os, uint64_t obj, uint64_t desc_obj,
b9b24bb4 176 nvlist_t *unsup_feat, nvlist_t *enabled_feat)
9ae529ec
CS
177{
178 boolean_t supported;
179 zap_cursor_t *zc;
180 zap_attribute_t *za;
181 char *buf;
182
d1d7e268
MK
183 zc = kmem_alloc(sizeof (zap_cursor_t), KM_SLEEP);
184 za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
9ae529ec
CS
185 buf = kmem_alloc(MAXPATHLEN, KM_SLEEP);
186
187 supported = B_TRUE;
188 for (zap_cursor_init(zc, os, obj);
189 zap_cursor_retrieve(zc, za) == 0;
190 zap_cursor_advance(zc)) {
191 ASSERT(za->za_integer_length == sizeof (uint64_t) &&
192 za->za_num_integers == 1);
193
b9b24bb4
CS
194 if (NULL != enabled_feat) {
195 fnvlist_add_uint64(enabled_feat, za->za_name,
196 za->za_first_integer);
197 }
198
9ae529ec
CS
199 if (za->za_first_integer != 0 &&
200 !zfeature_is_supported(za->za_name)) {
201 supported = B_FALSE;
202
b9b24bb4 203 if (NULL != unsup_feat) {
9ae529ec
CS
204 char *desc = "";
205
206 if (zap_lookup(os, desc_obj, za->za_name,
13a7ba1c 207 1, MAXPATHLEN, buf) == 0)
9ae529ec
CS
208 desc = buf;
209
210 VERIFY(nvlist_add_string(unsup_feat,
211 za->za_name, desc) == 0);
212 }
213 }
214 }
215 zap_cursor_fini(zc);
216
217 kmem_free(buf, MAXPATHLEN);
d1d7e268
MK
218 kmem_free(za, sizeof (zap_attribute_t));
219 kmem_free(zc, sizeof (zap_cursor_t));
9ae529ec
CS
220
221 return (supported);
222}
223
224static int
225feature_get_refcount(objset_t *os, uint64_t read_obj, uint64_t write_obj,
226 zfeature_info_t *feature, uint64_t *res)
227{
228 int err;
229 uint64_t refcount;
230 uint64_t zapobj = feature->fi_can_readonly ? write_obj : read_obj;
231
753c3839
MA
232 /*
233 * If the pool is currently being created, the feature objects may not
234 * have been allocated yet. Act as though all features are disabled.
235 */
236 if (zapobj == 0)
2e528b49 237 return (SET_ERROR(ENOTSUP));
9ae529ec
CS
238
239 err = zap_lookup(os, zapobj, feature->fi_guid, sizeof (uint64_t), 1,
240 &refcount);
241 if (err != 0) {
242 if (err == ENOENT)
2e528b49 243 return (SET_ERROR(ENOTSUP));
9ae529ec
CS
244 else
245 return (err);
246 }
247 *res = refcount;
248 return (0);
249}
250
251static int
252feature_do_action(objset_t *os, uint64_t read_obj, uint64_t write_obj,
253 uint64_t desc_obj, zfeature_info_t *feature, feature_action_t action,
254 dmu_tx_t *tx)
255{
256 int error;
257 uint64_t refcount;
258 uint64_t zapobj = feature->fi_can_readonly ? write_obj : read_obj;
259
260 ASSERT(0 != zapobj);
261 ASSERT(zfeature_is_valid_guid(feature->fi_guid));
262
263 error = zap_lookup(os, zapobj, feature->fi_guid,
264 sizeof (uint64_t), 1, &refcount);
265
266 /*
267 * If we can't ascertain the status of the specified feature, an I/O
268 * error occurred.
269 */
270 if (error != 0 && error != ENOENT)
271 return (error);
272
273 switch (action) {
274 case FEATURE_ACTION_ENABLE:
275 /*
276 * If the feature is already enabled, ignore the request.
277 */
278 if (error == 0)
279 return (0);
280 refcount = 0;
281 break;
282 case FEATURE_ACTION_INCR:
283 if (error == ENOENT)
2e528b49 284 return (SET_ERROR(ENOTSUP));
9ae529ec 285 if (refcount == UINT64_MAX)
2e528b49 286 return (SET_ERROR(EOVERFLOW));
9ae529ec
CS
287 refcount++;
288 break;
289 case FEATURE_ACTION_DECR:
290 if (error == ENOENT)
2e528b49 291 return (SET_ERROR(ENOTSUP));
9ae529ec 292 if (refcount == 0)
2e528b49 293 return (SET_ERROR(EOVERFLOW));
9ae529ec
CS
294 refcount--;
295 break;
296 default:
297 ASSERT(0);
298 break;
299 }
300
301 if (action == FEATURE_ACTION_ENABLE) {
302 int i;
303
304 for (i = 0; feature->fi_depends[i] != NULL; i++) {
305 zfeature_info_t *dep = feature->fi_depends[i];
306
307 error = feature_do_action(os, read_obj, write_obj,
308 desc_obj, dep, FEATURE_ACTION_ENABLE, tx);
309 if (error != 0)
310 return (error);
311 }
312 }
313
314 error = zap_update(os, zapobj, feature->fi_guid,
315 sizeof (uint64_t), 1, &refcount, tx);
316 if (error != 0)
317 return (error);
318
319 if (action == FEATURE_ACTION_ENABLE) {
320 error = zap_update(os, desc_obj,
321 feature->fi_guid, 1, strlen(feature->fi_desc) + 1,
322 feature->fi_desc, tx);
323 if (error != 0)
324 return (error);
325 }
326
327 if (action == FEATURE_ACTION_INCR && refcount == 1 && feature->fi_mos) {
328 spa_activate_mos_feature(dmu_objset_spa(os), feature->fi_guid);
329 }
330
331 if (action == FEATURE_ACTION_DECR && refcount == 0) {
332 spa_deactivate_mos_feature(dmu_objset_spa(os),
333 feature->fi_guid);
334 }
335
336 return (0);
337}
338
339void
340spa_feature_create_zap_objects(spa_t *spa, dmu_tx_t *tx)
341{
342 /*
343 * We create feature flags ZAP objects in two instances: during pool
344 * creation and during pool upgrade.
345 */
346 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)) || (!spa->spa_sync_on &&
347 tx->tx_txg == TXG_INITIAL));
348
349 spa->spa_feat_for_read_obj = zap_create_link(spa->spa_meta_objset,
350 DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
351 DMU_POOL_FEATURES_FOR_READ, tx);
352 spa->spa_feat_for_write_obj = zap_create_link(spa->spa_meta_objset,
353 DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
354 DMU_POOL_FEATURES_FOR_WRITE, tx);
355 spa->spa_feat_desc_obj = zap_create_link(spa->spa_meta_objset,
356 DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
357 DMU_POOL_FEATURE_DESCRIPTIONS, tx);
358}
359
360/*
361 * Enable any required dependencies, then enable the requested feature.
362 */
363void
364spa_feature_enable(spa_t *spa, zfeature_info_t *feature, dmu_tx_t *tx)
365{
366 ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
367 VERIFY3U(0, ==, feature_do_action(spa->spa_meta_objset,
368 spa->spa_feat_for_read_obj, spa->spa_feat_for_write_obj,
369 spa->spa_feat_desc_obj, feature, FEATURE_ACTION_ENABLE, tx));
370}
371
9ae529ec
CS
372void
373spa_feature_incr(spa_t *spa, zfeature_info_t *feature, dmu_tx_t *tx)
374{
93cf2076 375 ASSERT(dmu_tx_is_syncing(tx));
9ae529ec
CS
376 ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
377 VERIFY3U(0, ==, feature_do_action(spa->spa_meta_objset,
378 spa->spa_feat_for_read_obj, spa->spa_feat_for_write_obj,
379 spa->spa_feat_desc_obj, feature, FEATURE_ACTION_INCR, tx));
380}
381
9ae529ec
CS
382void
383spa_feature_decr(spa_t *spa, zfeature_info_t *feature, dmu_tx_t *tx)
384{
93cf2076 385 ASSERT(dmu_tx_is_syncing(tx));
9ae529ec
CS
386 ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
387 VERIFY3U(0, ==, feature_do_action(spa->spa_meta_objset,
388 spa->spa_feat_for_read_obj, spa->spa_feat_for_write_obj,
389 spa->spa_feat_desc_obj, feature, FEATURE_ACTION_DECR, tx));
390}
391
93cf2076
GW
392/*
393 * This interface is for debugging only. Normal consumers should use
394 * spa_feature_is_enabled/spa_feature_is_active.
395 */
396int
397spa_feature_get_refcount(spa_t *spa, zfeature_info_t *feature)
398{
399 int err;
400 uint64_t refcount = 0;
401
402 if (spa_version(spa) < SPA_VERSION_FEATURES)
403 return (B_FALSE);
404
405 err = feature_get_refcount(spa->spa_meta_objset,
406 spa->spa_feat_for_read_obj, spa->spa_feat_for_write_obj,
407 feature, &refcount);
408 ASSERT(err == 0 || err == ENOTSUP);
409 return (err == 0 ? refcount : 0);
410}
411
9ae529ec
CS
412boolean_t
413spa_feature_is_enabled(spa_t *spa, zfeature_info_t *feature)
414{
415 int err;
416 uint64_t refcount = 0;
417
418 if (spa_version(spa) < SPA_VERSION_FEATURES)
419 return (B_FALSE);
420
421 err = feature_get_refcount(spa->spa_meta_objset,
422 spa->spa_feat_for_read_obj, spa->spa_feat_for_write_obj,
423 feature, &refcount);
424 ASSERT(err == 0 || err == ENOTSUP);
425 return (err == 0);
426}
427
428boolean_t
429spa_feature_is_active(spa_t *spa, zfeature_info_t *feature)
430{
431 int err;
432 uint64_t refcount = 0;
433
434 if (spa_version(spa) < SPA_VERSION_FEATURES)
435 return (B_FALSE);
436
437 err = feature_get_refcount(spa->spa_meta_objset,
438 spa->spa_feat_for_read_obj, spa->spa_feat_for_write_obj,
439 feature, &refcount);
440 ASSERT(err == 0 || err == ENOTSUP);
441 return (err == 0 && refcount > 0);
442}