23 */
24
25#include <sys/dsl_pool.h>
26#include <sys/dsl_dataset.h>
27#include <sys/dsl_prop.h>
28#include <sys/dsl_dir.h>
29#include <sys/dsl_synctask.h>
30#include <sys/dsl_scan.h>
31#include <sys/dnode.h>
32#include <sys/dmu_tx.h>
33#include <sys/dmu_objset.h>
34#include <sys/arc.h>
35#include <sys/zap.h>
36#include <sys/zio.h>
37#include <sys/zfs_context.h>
38#include <sys/fs/zfs.h>
39#include <sys/zfs_znode.h>
40#include <sys/spa_impl.h>
41#include <sys/dsl_deadlist.h>
42
43int zfs_no_write_throttle = 0;
44int zfs_write_limit_shift = 3; /* 1/8th of physical memory */
45int zfs_txg_synctime_ms = 1000; /* target millisecs to sync a txg */
46
47uint64_t zfs_write_limit_min = 32 << 20; /* min write limit is 32MB */
48uint64_t zfs_write_limit_max = 0; /* max data payload per txg */
49uint64_t zfs_write_limit_inflated = 0;
50uint64_t zfs_write_limit_override = 0;
51
52kmutex_t zfs_write_limit_lock;
53
54static pgcnt_t old_physmem = 0;
55
56SYSCTL_DECL(_vfs_zfs);
57TUNABLE_INT("vfs.zfs.no_write_throttle", &zfs_no_write_throttle);
58SYSCTL_INT(_vfs_zfs, OID_AUTO, no_write_throttle, CTLFLAG_RDTUN,
59 &zfs_no_write_throttle, 0, "");
60TUNABLE_INT("vfs.zfs.write_limit_shift", &zfs_write_limit_shift);
61SYSCTL_INT(_vfs_zfs, OID_AUTO, write_limit_shift, CTLFLAG_RDTUN,
62 &zfs_write_limit_shift, 0, "2^N of physical memory");
63SYSCTL_DECL(_vfs_zfs_txg);
64TUNABLE_INT("vfs.zfs.txg.synctime_ms", &zfs_txg_synctime_ms);
65SYSCTL_INT(_vfs_zfs_txg, OID_AUTO, synctime_ms, CTLFLAG_RDTUN,
66 &zfs_txg_synctime_ms, 0, "Target milliseconds to sync a txg");
67
68TUNABLE_QUAD("vfs.zfs.write_limit_min", &zfs_write_limit_min);
69SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_min, CTLFLAG_RDTUN,
70 &zfs_write_limit_min, 0, "Minimum write limit");
71TUNABLE_QUAD("vfs.zfs.write_limit_max", &zfs_write_limit_max);
72SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_max, CTLFLAG_RDTUN,
73 &zfs_write_limit_max, 0, "Maximum data payload per txg");
74TUNABLE_QUAD("vfs.zfs.write_limit_inflated", &zfs_write_limit_inflated);
75SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_inflated, CTLFLAG_RDTUN,
76 &zfs_write_limit_inflated, 0, "");
77TUNABLE_QUAD("vfs.zfs.write_limit_override", &zfs_write_limit_override);
78SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_override, CTLFLAG_RDTUN,
79 &zfs_write_limit_override, 0, "");
80
81int
82dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
83{
84 uint64_t obj;
85 int err;
86
87 err = zap_lookup(dp->dp_meta_objset,
88 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
89 name, sizeof (obj), 1, &obj);
90 if (err)
91 return (err);
92
93 return (dsl_dir_open_obj(dp, obj, name, dp, ddp));
94}
95
96static dsl_pool_t *
97dsl_pool_open_impl(spa_t *spa, uint64_t txg)
98{
99 dsl_pool_t *dp;
100 blkptr_t *bp = spa_get_rootblkptr(spa);
101
102 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
103 dp->dp_spa = spa;
104 dp->dp_meta_rootbp = *bp;
105 rw_init(&dp->dp_config_rwlock, NULL, RW_DEFAULT, NULL);
106 dp->dp_write_limit = zfs_write_limit_min;
107 txg_init(dp, txg);
108
109 txg_list_create(&dp->dp_dirty_datasets,
110 offsetof(dsl_dataset_t, ds_dirty_link));
111 txg_list_create(&dp->dp_dirty_dirs,
112 offsetof(dsl_dir_t, dd_dirty_link));
113 txg_list_create(&dp->dp_sync_tasks,
114 offsetof(dsl_sync_task_group_t, dstg_node));
115 list_create(&dp->dp_synced_datasets, sizeof (dsl_dataset_t),
116 offsetof(dsl_dataset_t, ds_synced_link));
117
118 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
119
120 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
121 1, 4, 0);
122
123 return (dp);
124}
125
126int
127dsl_pool_open(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
128{
129 int err;
130 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
131 dsl_dir_t *dd;
132 dsl_dataset_t *ds;
133 uint64_t obj;
134
135 rw_enter(&dp->dp_config_rwlock, RW_WRITER);
136 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
137 &dp->dp_meta_objset);
138 if (err)
139 goto out;
140
141 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
142 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
143 &dp->dp_root_dir_obj);
144 if (err)
145 goto out;
146
147 err = dsl_dir_open_obj(dp, dp->dp_root_dir_obj,
148 NULL, dp, &dp->dp_root_dir);
149 if (err)
150 goto out;
151
152 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
153 if (err)
154 goto out;
155
156 if (spa_version(spa) >= SPA_VERSION_ORIGIN) {
157 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
158 if (err)
159 goto out;
160 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
161 FTAG, &ds);
162 if (err == 0) {
163 err = dsl_dataset_hold_obj(dp,
164 ds->ds_phys->ds_prev_snap_obj, dp,
165 &dp->dp_origin_snap);
166 dsl_dataset_rele(ds, FTAG);
167 }
168 dsl_dir_close(dd, dp);
169 if (err)
170 goto out;
171 }
172
173 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
174 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
175 &dp->dp_free_dir);
176 if (err)
177 goto out;
178
179 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
180 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
181 if (err)
182 goto out;
183 VERIFY3U(0, ==, bpobj_open(&dp->dp_free_bpobj,
184 dp->dp_meta_objset, obj));
185 }
186
187 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
188 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
189 &dp->dp_tmp_userrefs_obj);
190 if (err == ENOENT)
191 err = 0;
192 if (err)
193 goto out;
194
195 err = dsl_scan_init(dp, txg);
196
197out:
198 rw_exit(&dp->dp_config_rwlock);
199 if (err)
200 dsl_pool_close(dp);
201 else
202 *dpp = dp;
203
204 return (err);
205}
206
207void
208dsl_pool_close(dsl_pool_t *dp)
209{
210 /* drop our references from dsl_pool_open() */
211
212 /*
213 * Since we held the origin_snap from "syncing" context (which
214 * includes pool-opening context), it actually only got a "ref"
215 * and not a hold, so just drop that here.
216 */
217 if (dp->dp_origin_snap)
218 dsl_dataset_drop_ref(dp->dp_origin_snap, dp);
219 if (dp->dp_mos_dir)
220 dsl_dir_close(dp->dp_mos_dir, dp);
221 if (dp->dp_free_dir)
222 dsl_dir_close(dp->dp_free_dir, dp);
223 if (dp->dp_root_dir)
224 dsl_dir_close(dp->dp_root_dir, dp);
225
226 bpobj_close(&dp->dp_free_bpobj);
227
228 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
229 if (dp->dp_meta_objset)
230 dmu_objset_evict(dp->dp_meta_objset);
231
232 txg_list_destroy(&dp->dp_dirty_datasets);
233 txg_list_destroy(&dp->dp_sync_tasks);
234 txg_list_destroy(&dp->dp_dirty_dirs);
235 list_destroy(&dp->dp_synced_datasets);
236
237 arc_flush(dp->dp_spa);
238 txg_fini(dp);
239 dsl_scan_fini(dp);
240 rw_destroy(&dp->dp_config_rwlock);
241 mutex_destroy(&dp->dp_lock);
242 taskq_destroy(dp->dp_vnrele_taskq);
243 if (dp->dp_blkstats)
244 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
245 kmem_free(dp, sizeof (dsl_pool_t));
246}
247
248dsl_pool_t *
249dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
250{
251 int err;
252 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
253 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
254 objset_t *os;
255 dsl_dataset_t *ds;
256 uint64_t obj;
257
258 /* create and open the MOS (meta-objset) */
259 dp->dp_meta_objset = dmu_objset_create_impl(spa,
260 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
261
262 /* create the pool directory */
263 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
264 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
265 ASSERT3U(err, ==, 0);
266
267 /* Initialize scan structures */
268 VERIFY3U(0, ==, dsl_scan_init(dp, txg));
269
270 /* create and open the root dir */
271 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
272 VERIFY(0 == dsl_dir_open_obj(dp, dp->dp_root_dir_obj,
273 NULL, dp, &dp->dp_root_dir));
274
275 /* create and open the meta-objset dir */
276 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
277 VERIFY(0 == dsl_pool_open_special_dir(dp,
278 MOS_DIR_NAME, &dp->dp_mos_dir));
279
280 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
281 /* create and open the free dir */
282 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
283 FREE_DIR_NAME, tx);
284 VERIFY(0 == dsl_pool_open_special_dir(dp,
285 FREE_DIR_NAME, &dp->dp_free_dir));
286
287 /* create and open the free_bplist */
288 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
289 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
290 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
291 VERIFY3U(0, ==, bpobj_open(&dp->dp_free_bpobj,
292 dp->dp_meta_objset, obj));
293 }
294
295 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
296 dsl_pool_create_origin(dp, tx);
297
298 /* create the root dataset */
299 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
300
301 /* create the root objset */
302 VERIFY(0 == dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
303 os = dmu_objset_create_impl(dp->dp_spa, ds,
304 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
305#ifdef _KERNEL
306 zfs_create_fs(os, kcred, zplprops, tx);
307#endif
308 dsl_dataset_rele(ds, FTAG);
309
310 dmu_tx_commit(tx);
311
312 return (dp);
313}
314
315static int
316deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
317{
318 dsl_deadlist_t *dl = arg;
| 24 */
25
26#include <sys/dsl_pool.h>
27#include <sys/dsl_dataset.h>
28#include <sys/dsl_prop.h>
29#include <sys/dsl_dir.h>
30#include <sys/dsl_synctask.h>
31#include <sys/dsl_scan.h>
32#include <sys/dnode.h>
33#include <sys/dmu_tx.h>
34#include <sys/dmu_objset.h>
35#include <sys/arc.h>
36#include <sys/zap.h>
37#include <sys/zio.h>
38#include <sys/zfs_context.h>
39#include <sys/fs/zfs.h>
40#include <sys/zfs_znode.h>
41#include <sys/spa_impl.h>
42#include <sys/dsl_deadlist.h>
43
44int zfs_no_write_throttle = 0;
45int zfs_write_limit_shift = 3; /* 1/8th of physical memory */
46int zfs_txg_synctime_ms = 1000; /* target millisecs to sync a txg */
47
48uint64_t zfs_write_limit_min = 32 << 20; /* min write limit is 32MB */
49uint64_t zfs_write_limit_max = 0; /* max data payload per txg */
50uint64_t zfs_write_limit_inflated = 0;
51uint64_t zfs_write_limit_override = 0;
52
53kmutex_t zfs_write_limit_lock;
54
55static pgcnt_t old_physmem = 0;
56
57SYSCTL_DECL(_vfs_zfs);
58TUNABLE_INT("vfs.zfs.no_write_throttle", &zfs_no_write_throttle);
59SYSCTL_INT(_vfs_zfs, OID_AUTO, no_write_throttle, CTLFLAG_RDTUN,
60 &zfs_no_write_throttle, 0, "");
61TUNABLE_INT("vfs.zfs.write_limit_shift", &zfs_write_limit_shift);
62SYSCTL_INT(_vfs_zfs, OID_AUTO, write_limit_shift, CTLFLAG_RDTUN,
63 &zfs_write_limit_shift, 0, "2^N of physical memory");
64SYSCTL_DECL(_vfs_zfs_txg);
65TUNABLE_INT("vfs.zfs.txg.synctime_ms", &zfs_txg_synctime_ms);
66SYSCTL_INT(_vfs_zfs_txg, OID_AUTO, synctime_ms, CTLFLAG_RDTUN,
67 &zfs_txg_synctime_ms, 0, "Target milliseconds to sync a txg");
68
69TUNABLE_QUAD("vfs.zfs.write_limit_min", &zfs_write_limit_min);
70SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_min, CTLFLAG_RDTUN,
71 &zfs_write_limit_min, 0, "Minimum write limit");
72TUNABLE_QUAD("vfs.zfs.write_limit_max", &zfs_write_limit_max);
73SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_max, CTLFLAG_RDTUN,
74 &zfs_write_limit_max, 0, "Maximum data payload per txg");
75TUNABLE_QUAD("vfs.zfs.write_limit_inflated", &zfs_write_limit_inflated);
76SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_inflated, CTLFLAG_RDTUN,
77 &zfs_write_limit_inflated, 0, "");
78TUNABLE_QUAD("vfs.zfs.write_limit_override", &zfs_write_limit_override);
79SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_override, CTLFLAG_RDTUN,
80 &zfs_write_limit_override, 0, "");
81
82int
83dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
84{
85 uint64_t obj;
86 int err;
87
88 err = zap_lookup(dp->dp_meta_objset,
89 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
90 name, sizeof (obj), 1, &obj);
91 if (err)
92 return (err);
93
94 return (dsl_dir_open_obj(dp, obj, name, dp, ddp));
95}
96
97static dsl_pool_t *
98dsl_pool_open_impl(spa_t *spa, uint64_t txg)
99{
100 dsl_pool_t *dp;
101 blkptr_t *bp = spa_get_rootblkptr(spa);
102
103 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
104 dp->dp_spa = spa;
105 dp->dp_meta_rootbp = *bp;
106 rw_init(&dp->dp_config_rwlock, NULL, RW_DEFAULT, NULL);
107 dp->dp_write_limit = zfs_write_limit_min;
108 txg_init(dp, txg);
109
110 txg_list_create(&dp->dp_dirty_datasets,
111 offsetof(dsl_dataset_t, ds_dirty_link));
112 txg_list_create(&dp->dp_dirty_dirs,
113 offsetof(dsl_dir_t, dd_dirty_link));
114 txg_list_create(&dp->dp_sync_tasks,
115 offsetof(dsl_sync_task_group_t, dstg_node));
116 list_create(&dp->dp_synced_datasets, sizeof (dsl_dataset_t),
117 offsetof(dsl_dataset_t, ds_synced_link));
118
119 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
120
121 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
122 1, 4, 0);
123
124 return (dp);
125}
126
127int
128dsl_pool_open(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
129{
130 int err;
131 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
132 dsl_dir_t *dd;
133 dsl_dataset_t *ds;
134 uint64_t obj;
135
136 rw_enter(&dp->dp_config_rwlock, RW_WRITER);
137 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
138 &dp->dp_meta_objset);
139 if (err)
140 goto out;
141
142 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
143 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
144 &dp->dp_root_dir_obj);
145 if (err)
146 goto out;
147
148 err = dsl_dir_open_obj(dp, dp->dp_root_dir_obj,
149 NULL, dp, &dp->dp_root_dir);
150 if (err)
151 goto out;
152
153 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
154 if (err)
155 goto out;
156
157 if (spa_version(spa) >= SPA_VERSION_ORIGIN) {
158 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
159 if (err)
160 goto out;
161 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
162 FTAG, &ds);
163 if (err == 0) {
164 err = dsl_dataset_hold_obj(dp,
165 ds->ds_phys->ds_prev_snap_obj, dp,
166 &dp->dp_origin_snap);
167 dsl_dataset_rele(ds, FTAG);
168 }
169 dsl_dir_close(dd, dp);
170 if (err)
171 goto out;
172 }
173
174 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
175 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
176 &dp->dp_free_dir);
177 if (err)
178 goto out;
179
180 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
181 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
182 if (err)
183 goto out;
184 VERIFY3U(0, ==, bpobj_open(&dp->dp_free_bpobj,
185 dp->dp_meta_objset, obj));
186 }
187
188 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
189 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
190 &dp->dp_tmp_userrefs_obj);
191 if (err == ENOENT)
192 err = 0;
193 if (err)
194 goto out;
195
196 err = dsl_scan_init(dp, txg);
197
198out:
199 rw_exit(&dp->dp_config_rwlock);
200 if (err)
201 dsl_pool_close(dp);
202 else
203 *dpp = dp;
204
205 return (err);
206}
207
208void
209dsl_pool_close(dsl_pool_t *dp)
210{
211 /* drop our references from dsl_pool_open() */
212
213 /*
214 * Since we held the origin_snap from "syncing" context (which
215 * includes pool-opening context), it actually only got a "ref"
216 * and not a hold, so just drop that here.
217 */
218 if (dp->dp_origin_snap)
219 dsl_dataset_drop_ref(dp->dp_origin_snap, dp);
220 if (dp->dp_mos_dir)
221 dsl_dir_close(dp->dp_mos_dir, dp);
222 if (dp->dp_free_dir)
223 dsl_dir_close(dp->dp_free_dir, dp);
224 if (dp->dp_root_dir)
225 dsl_dir_close(dp->dp_root_dir, dp);
226
227 bpobj_close(&dp->dp_free_bpobj);
228
229 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
230 if (dp->dp_meta_objset)
231 dmu_objset_evict(dp->dp_meta_objset);
232
233 txg_list_destroy(&dp->dp_dirty_datasets);
234 txg_list_destroy(&dp->dp_sync_tasks);
235 txg_list_destroy(&dp->dp_dirty_dirs);
236 list_destroy(&dp->dp_synced_datasets);
237
238 arc_flush(dp->dp_spa);
239 txg_fini(dp);
240 dsl_scan_fini(dp);
241 rw_destroy(&dp->dp_config_rwlock);
242 mutex_destroy(&dp->dp_lock);
243 taskq_destroy(dp->dp_vnrele_taskq);
244 if (dp->dp_blkstats)
245 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
246 kmem_free(dp, sizeof (dsl_pool_t));
247}
248
249dsl_pool_t *
250dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
251{
252 int err;
253 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
254 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
255 objset_t *os;
256 dsl_dataset_t *ds;
257 uint64_t obj;
258
259 /* create and open the MOS (meta-objset) */
260 dp->dp_meta_objset = dmu_objset_create_impl(spa,
261 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
262
263 /* create the pool directory */
264 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
265 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
266 ASSERT3U(err, ==, 0);
267
268 /* Initialize scan structures */
269 VERIFY3U(0, ==, dsl_scan_init(dp, txg));
270
271 /* create and open the root dir */
272 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
273 VERIFY(0 == dsl_dir_open_obj(dp, dp->dp_root_dir_obj,
274 NULL, dp, &dp->dp_root_dir));
275
276 /* create and open the meta-objset dir */
277 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
278 VERIFY(0 == dsl_pool_open_special_dir(dp,
279 MOS_DIR_NAME, &dp->dp_mos_dir));
280
281 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
282 /* create and open the free dir */
283 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
284 FREE_DIR_NAME, tx);
285 VERIFY(0 == dsl_pool_open_special_dir(dp,
286 FREE_DIR_NAME, &dp->dp_free_dir));
287
288 /* create and open the free_bplist */
289 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
290 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
291 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
292 VERIFY3U(0, ==, bpobj_open(&dp->dp_free_bpobj,
293 dp->dp_meta_objset, obj));
294 }
295
296 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
297 dsl_pool_create_origin(dp, tx);
298
299 /* create the root dataset */
300 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
301
302 /* create the root objset */
303 VERIFY(0 == dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
304 os = dmu_objset_create_impl(dp->dp_spa, ds,
305 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
306#ifdef _KERNEL
307 zfs_create_fs(os, kcred, zplprops, tx);
308#endif
309 dsl_dataset_rele(ds, FTAG);
310
311 dmu_tx_commit(tx);
312
313 return (dp);
314}
315
316static int
317deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
318{
319 dsl_deadlist_t *dl = arg;
|
320 return (0);
321}
322
323void
324dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
325{
326 zio_t *zio;
327 dmu_tx_t *tx;
328 dsl_dir_t *dd;
329 dsl_dataset_t *ds;
330 dsl_sync_task_group_t *dstg;
331 objset_t *mos = dp->dp_meta_objset;
332 hrtime_t start, write_time;
333 uint64_t data_written;
334 int err;
335
336 /*
337 * We need to copy dp_space_towrite() before doing
338 * dsl_sync_task_group_sync(), because
339 * dsl_dataset_snapshot_reserve_space() will increase
340 * dp_space_towrite but not actually write anything.
341 */
342 data_written = dp->dp_space_towrite[txg & TXG_MASK];
343
344 tx = dmu_tx_create_assigned(dp, txg);
345
346 dp->dp_read_overhead = 0;
347 start = gethrtime();
348
349 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
350 while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
351 /*
352 * We must not sync any non-MOS datasets twice, because
353 * we may have taken a snapshot of them. However, we
354 * may sync newly-created datasets on pass 2.
355 */
356 ASSERT(!list_link_active(&ds->ds_synced_link));
357 list_insert_tail(&dp->dp_synced_datasets, ds);
358 dsl_dataset_sync(ds, zio, tx);
359 }
360 DTRACE_PROBE(pool_sync__1setup);
361 err = zio_wait(zio);
362
363 write_time = gethrtime() - start;
364 ASSERT(err == 0);
365 DTRACE_PROBE(pool_sync__2rootzio);
366
367 for (ds = list_head(&dp->dp_synced_datasets); ds;
368 ds = list_next(&dp->dp_synced_datasets, ds))
369 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
370
371 /*
372 * Sync the datasets again to push out the changes due to
373 * userspace updates. This must be done before we process the
374 * sync tasks, because that could cause a snapshot of a dataset
375 * whose ds_bp will be rewritten when we do this 2nd sync.
376 */
377 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
378 while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
379 ASSERT(list_link_active(&ds->ds_synced_link));
380 dmu_buf_rele(ds->ds_dbuf, ds);
381 dsl_dataset_sync(ds, zio, tx);
382 }
383 err = zio_wait(zio);
384
385 /*
386 * Move dead blocks from the pending deadlist to the on-disk
387 * deadlist.
388 */
389 for (ds = list_head(&dp->dp_synced_datasets); ds;
390 ds = list_next(&dp->dp_synced_datasets, ds)) {
391 bplist_iterate(&ds->ds_pending_deadlist,
392 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
393 }
394
395 while (dstg = txg_list_remove(&dp->dp_sync_tasks, txg)) {
396 /*
397 * No more sync tasks should have been added while we
398 * were syncing.
399 */
400 ASSERT(spa_sync_pass(dp->dp_spa) == 1);
401 dsl_sync_task_group_sync(dstg, tx);
402 }
403 DTRACE_PROBE(pool_sync__3task);
404
405 start = gethrtime();
406 while (dd = txg_list_remove(&dp->dp_dirty_dirs, txg))
407 dsl_dir_sync(dd, tx);
408 write_time += gethrtime() - start;
409
410 start = gethrtime();
411 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
412 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
413 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
414 dmu_objset_sync(mos, zio, tx);
415 err = zio_wait(zio);
416 ASSERT(err == 0);
417 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
418 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
419 }
420 write_time += gethrtime() - start;
421 DTRACE_PROBE2(pool_sync__4io, hrtime_t, write_time,
422 hrtime_t, dp->dp_read_overhead);
423 write_time -= dp->dp_read_overhead;
424
425 dmu_tx_commit(tx);
426
427 dp->dp_space_towrite[txg & TXG_MASK] = 0;
428 ASSERT(dp->dp_tempreserved[txg & TXG_MASK] == 0);
429
430 /*
431 * If the write limit max has not been explicitly set, set it
432 * to a fraction of available physical memory (default 1/8th).
433 * Note that we must inflate the limit because the spa
434 * inflates write sizes to account for data replication.
435 * Check this each sync phase to catch changing memory size.
436 */
437 if (physmem != old_physmem && zfs_write_limit_shift) {
438 mutex_enter(&zfs_write_limit_lock);
439 old_physmem = physmem;
440 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
441 zfs_write_limit_inflated = MAX(zfs_write_limit_min,
442 spa_get_asize(dp->dp_spa, zfs_write_limit_max));
443 mutex_exit(&zfs_write_limit_lock);
444 }
445
446 /*
447 * Attempt to keep the sync time consistent by adjusting the
448 * amount of write traffic allowed into each transaction group.
449 * Weight the throughput calculation towards the current value:
450 * thru = 3/4 old_thru + 1/4 new_thru
451 *
452 * Note: write_time is in nanosecs, so write_time/MICROSEC
453 * yields millisecs
454 */
455 ASSERT(zfs_write_limit_min > 0);
456 if (data_written > zfs_write_limit_min / 8 && write_time > MICROSEC) {
457 uint64_t throughput = data_written / (write_time / MICROSEC);
458
459 if (dp->dp_throughput)
460 dp->dp_throughput = throughput / 4 +
461 3 * dp->dp_throughput / 4;
462 else
463 dp->dp_throughput = throughput;
464 dp->dp_write_limit = MIN(zfs_write_limit_inflated,
465 MAX(zfs_write_limit_min,
466 dp->dp_throughput * zfs_txg_synctime_ms));
467 }
468}
469
470void
471dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
472{
473 dsl_dataset_t *ds;
474 objset_t *os;
475
476 while (ds = list_head(&dp->dp_synced_datasets)) {
477 list_remove(&dp->dp_synced_datasets, ds);
478 os = ds->ds_objset;
479 zil_clean(os->os_zil, txg);
480 ASSERT(!dmu_objset_is_dirty(os, txg));
481 dmu_buf_rele(ds->ds_dbuf, ds);
482 }
483 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
484}
485
486/*
487 * TRUE if the current thread is the tx_sync_thread or if we
488 * are being called from SPA context during pool initialization.
489 */
490int
491dsl_pool_sync_context(dsl_pool_t *dp)
492{
493 return (curthread == dp->dp_tx.tx_sync_thread ||
494 spa_get_dsl(dp->dp_spa) == NULL);
495}
496
497uint64_t
498dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
499{
500 uint64_t space, resv;
501
502 /*
503 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
504 * efficiency.
505 * XXX The intent log is not accounted for, so it must fit
506 * within this slop.
507 *
508 * If we're trying to assess whether it's OK to do a free,
509 * cut the reservation in half to allow forward progress
510 * (e.g. make it possible to rm(1) files from a full pool).
511 */
512 space = spa_get_dspace(dp->dp_spa);
513 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
514 if (netfree)
515 resv >>= 1;
516
517 return (space - resv);
518}
519
520int
521dsl_pool_tempreserve_space(dsl_pool_t *dp, uint64_t space, dmu_tx_t *tx)
522{
523 uint64_t reserved = 0;
524 uint64_t write_limit = (zfs_write_limit_override ?
525 zfs_write_limit_override : dp->dp_write_limit);
526
527 if (zfs_no_write_throttle) {
528 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK],
529 space);
530 return (0);
531 }
532
533 /*
534 * Check to see if we have exceeded the maximum allowed IO for
535 * this transaction group. We can do this without locks since
536 * a little slop here is ok. Note that we do the reserved check
537 * with only half the requested reserve: this is because the
538 * reserve requests are worst-case, and we really don't want to
539 * throttle based off of worst-case estimates.
540 */
541 if (write_limit > 0) {
542 reserved = dp->dp_space_towrite[tx->tx_txg & TXG_MASK]
543 + dp->dp_tempreserved[tx->tx_txg & TXG_MASK] / 2;
544
545 if (reserved && reserved > write_limit)
546 return (ERESTART);
547 }
548
549 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], space);
550
551 /*
552 * If this transaction group is over 7/8ths capacity, delay
553 * the caller 1 clock tick. This will slow down the "fill"
554 * rate until the sync process can catch up with us.
555 */
556 if (reserved && reserved > (write_limit - (write_limit >> 3)))
557 txg_delay(dp, tx->tx_txg, 1);
558
559 return (0);
560}
561
562void
563dsl_pool_tempreserve_clear(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
564{
565 ASSERT(dp->dp_tempreserved[tx->tx_txg & TXG_MASK] >= space);
566 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], -space);
567}
568
569void
570dsl_pool_memory_pressure(dsl_pool_t *dp)
571{
572 uint64_t space_inuse = 0;
573 int i;
574
575 if (dp->dp_write_limit == zfs_write_limit_min)
576 return;
577
578 for (i = 0; i < TXG_SIZE; i++) {
579 space_inuse += dp->dp_space_towrite[i];
580 space_inuse += dp->dp_tempreserved[i];
581 }
582 dp->dp_write_limit = MAX(zfs_write_limit_min,
583 MIN(dp->dp_write_limit, space_inuse / 4));
584}
585
586void
587dsl_pool_willuse_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
588{
589 if (space > 0) {
590 mutex_enter(&dp->dp_lock);
591 dp->dp_space_towrite[tx->tx_txg & TXG_MASK] += space;
592 mutex_exit(&dp->dp_lock);
593 }
594}
595
596/* ARGSUSED */
597static int
598upgrade_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
599{
600 dmu_tx_t *tx = arg;
601 dsl_dataset_t *ds, *prev = NULL;
602 int err;
603 dsl_pool_t *dp = spa_get_dsl(spa);
604
605 err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
606 if (err)
607 return (err);
608
609 while (ds->ds_phys->ds_prev_snap_obj != 0) {
610 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
611 FTAG, &prev);
612 if (err) {
613 dsl_dataset_rele(ds, FTAG);
614 return (err);
615 }
616
617 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
618 break;
619 dsl_dataset_rele(ds, FTAG);
620 ds = prev;
621 prev = NULL;
622 }
623
624 if (prev == NULL) {
625 prev = dp->dp_origin_snap;
626
627 /*
628 * The $ORIGIN can't have any data, or the accounting
629 * will be wrong.
630 */
631 ASSERT(prev->ds_phys->ds_bp.blk_birth == 0);
632
633 /* The origin doesn't get attached to itself */
634 if (ds->ds_object == prev->ds_object) {
635 dsl_dataset_rele(ds, FTAG);
636 return (0);
637 }
638
639 dmu_buf_will_dirty(ds->ds_dbuf, tx);
640 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
641 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
642
643 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
644 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
645
646 dmu_buf_will_dirty(prev->ds_dbuf, tx);
647 prev->ds_phys->ds_num_children++;
648
649 if (ds->ds_phys->ds_next_snap_obj == 0) {
650 ASSERT(ds->ds_prev == NULL);
651 VERIFY(0 == dsl_dataset_hold_obj(dp,
652 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
653 }
654 }
655
656 ASSERT(ds->ds_dir->dd_phys->dd_origin_obj == prev->ds_object);
657 ASSERT(ds->ds_phys->ds_prev_snap_obj == prev->ds_object);
658
659 if (prev->ds_phys->ds_next_clones_obj == 0) {
660 dmu_buf_will_dirty(prev->ds_dbuf, tx);
661 prev->ds_phys->ds_next_clones_obj =
662 zap_create(dp->dp_meta_objset,
663 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
664 }
665 VERIFY(0 == zap_add_int(dp->dp_meta_objset,
666 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
667
668 dsl_dataset_rele(ds, FTAG);
669 if (prev != dp->dp_origin_snap)
670 dsl_dataset_rele(prev, FTAG);
671 return (0);
672}
673
674void
675dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
676{
677 ASSERT(dmu_tx_is_syncing(tx));
678 ASSERT(dp->dp_origin_snap != NULL);
679
680 VERIFY3U(0, ==, dmu_objset_find_spa(dp->dp_spa, NULL, upgrade_clones_cb,
681 tx, DS_FIND_CHILDREN));
682}
683
684/* ARGSUSED */
685static int
686upgrade_dir_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
687{
688 dmu_tx_t *tx = arg;
689 dsl_dataset_t *ds;
690 dsl_pool_t *dp = spa_get_dsl(spa);
691 objset_t *mos = dp->dp_meta_objset;
692
693 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
694
695 if (ds->ds_dir->dd_phys->dd_origin_obj) {
696 dsl_dataset_t *origin;
697
698 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
699 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
700
701 if (origin->ds_dir->dd_phys->dd_clones == 0) {
702 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
703 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
704 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
705 }
706
707 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
708 origin->ds_dir->dd_phys->dd_clones, dsobj, tx));
709
710 dsl_dataset_rele(origin, FTAG);
711 }
712
713 dsl_dataset_rele(ds, FTAG);
714 return (0);
715}
716
717void
718dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
719{
720 ASSERT(dmu_tx_is_syncing(tx));
721 uint64_t obj;
722
723 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
724 VERIFY(0 == dsl_pool_open_special_dir(dp,
725 FREE_DIR_NAME, &dp->dp_free_dir));
726
727 /*
728 * We can't use bpobj_alloc(), because spa_version() still
729 * returns the old version, and we need a new-version bpobj with
730 * subobj support. So call dmu_object_alloc() directly.
731 */
732 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
733 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
734 VERIFY3U(0, ==, zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
735 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
736 VERIFY3U(0, ==, bpobj_open(&dp->dp_free_bpobj,
737 dp->dp_meta_objset, obj));
738
739 VERIFY3U(0, ==, dmu_objset_find_spa(dp->dp_spa, NULL,
740 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
741}
742
743void
744dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
745{
746 uint64_t dsobj;
747 dsl_dataset_t *ds;
748
749 ASSERT(dmu_tx_is_syncing(tx));
750 ASSERT(dp->dp_origin_snap == NULL);
751
752 /* create the origin dir, ds, & snap-ds */
753 rw_enter(&dp->dp_config_rwlock, RW_WRITER);
754 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
755 NULL, 0, kcred, tx);
756 VERIFY(0 == dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
757 dsl_dataset_snapshot_sync(ds, ORIGIN_DIR_NAME, tx);
758 VERIFY(0 == dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
759 dp, &dp->dp_origin_snap));
760 dsl_dataset_rele(ds, FTAG);
761 rw_exit(&dp->dp_config_rwlock);
762}
763
764taskq_t *
765dsl_pool_vnrele_taskq(dsl_pool_t *dp)
766{
767 return (dp->dp_vnrele_taskq);
768}
769
770/*
771 * Walk through the pool-wide zap object of temporary snapshot user holds
772 * and release them.
773 */
774void
775dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
776{
777 zap_attribute_t za;
778 zap_cursor_t zc;
779 objset_t *mos = dp->dp_meta_objset;
780 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
781
782 if (zapobj == 0)
783 return;
784 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
785
786 for (zap_cursor_init(&zc, mos, zapobj);
787 zap_cursor_retrieve(&zc, &za) == 0;
788 zap_cursor_advance(&zc)) {
789 char *htag;
790 uint64_t dsobj;
791
792 htag = strchr(za.za_name, '-');
793 *htag = '\0';
794 ++htag;
795 dsobj = strtonum(za.za_name, NULL);
796 (void) dsl_dataset_user_release_tmp(dp, dsobj, htag, B_FALSE);
797 }
798 zap_cursor_fini(&zc);
799}
800
801/*
802 * Create the pool-wide zap object for storing temporary snapshot holds.
803 */
804void
805dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
806{
807 objset_t *mos = dp->dp_meta_objset;
808
809 ASSERT(dp->dp_tmp_userrefs_obj == 0);
810 ASSERT(dmu_tx_is_syncing(tx));
811
812 dp->dp_tmp_userrefs_obj = zap_create(mos, DMU_OT_USERREFS,
813 DMU_OT_NONE, 0, tx);
814
815 VERIFY(zap_add(mos, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS,
816 sizeof (uint64_t), 1, &dp->dp_tmp_userrefs_obj, tx) == 0);
817}
818
819static int
820dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
821 const char *tag, uint64_t *now, dmu_tx_t *tx, boolean_t holding)
822{
823 objset_t *mos = dp->dp_meta_objset;
824 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
825 char *name;
826 int error;
827
828 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
829 ASSERT(dmu_tx_is_syncing(tx));
830
831 /*
832 * If the pool was created prior to SPA_VERSION_USERREFS, the
833 * zap object for temporary holds might not exist yet.
834 */
835 if (zapobj == 0) {
836 if (holding) {
837 dsl_pool_user_hold_create_obj(dp, tx);
838 zapobj = dp->dp_tmp_userrefs_obj;
839 } else {
840 return (ENOENT);
841 }
842 }
843
844 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
845 if (holding)
846 error = zap_add(mos, zapobj, name, 8, 1, now, tx);
847 else
848 error = zap_remove(mos, zapobj, name, tx);
849 strfree(name);
850
851 return (error);
852}
853
854/*
855 * Add a temporary hold for the given dataset object and tag.
856 */
857int
858dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
859 uint64_t *now, dmu_tx_t *tx)
860{
861 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
862}
863
864/*
865 * Release a temporary hold for the given dataset object and tag.
866 */
867int
868dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
869 dmu_tx_t *tx)
870{
871 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
872 tx, B_FALSE));
873}
| 324 return (0);
325}
326
327void
328dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
329{
330 zio_t *zio;
331 dmu_tx_t *tx;
332 dsl_dir_t *dd;
333 dsl_dataset_t *ds;
334 dsl_sync_task_group_t *dstg;
335 objset_t *mos = dp->dp_meta_objset;
336 hrtime_t start, write_time;
337 uint64_t data_written;
338 int err;
339
340 /*
341 * We need to copy dp_space_towrite() before doing
342 * dsl_sync_task_group_sync(), because
343 * dsl_dataset_snapshot_reserve_space() will increase
344 * dp_space_towrite but not actually write anything.
345 */
346 data_written = dp->dp_space_towrite[txg & TXG_MASK];
347
348 tx = dmu_tx_create_assigned(dp, txg);
349
350 dp->dp_read_overhead = 0;
351 start = gethrtime();
352
353 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
354 while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
355 /*
356 * We must not sync any non-MOS datasets twice, because
357 * we may have taken a snapshot of them. However, we
358 * may sync newly-created datasets on pass 2.
359 */
360 ASSERT(!list_link_active(&ds->ds_synced_link));
361 list_insert_tail(&dp->dp_synced_datasets, ds);
362 dsl_dataset_sync(ds, zio, tx);
363 }
364 DTRACE_PROBE(pool_sync__1setup);
365 err = zio_wait(zio);
366
367 write_time = gethrtime() - start;
368 ASSERT(err == 0);
369 DTRACE_PROBE(pool_sync__2rootzio);
370
371 for (ds = list_head(&dp->dp_synced_datasets); ds;
372 ds = list_next(&dp->dp_synced_datasets, ds))
373 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
374
375 /*
376 * Sync the datasets again to push out the changes due to
377 * userspace updates. This must be done before we process the
378 * sync tasks, because that could cause a snapshot of a dataset
379 * whose ds_bp will be rewritten when we do this 2nd sync.
380 */
381 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
382 while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
383 ASSERT(list_link_active(&ds->ds_synced_link));
384 dmu_buf_rele(ds->ds_dbuf, ds);
385 dsl_dataset_sync(ds, zio, tx);
386 }
387 err = zio_wait(zio);
388
389 /*
390 * Move dead blocks from the pending deadlist to the on-disk
391 * deadlist.
392 */
393 for (ds = list_head(&dp->dp_synced_datasets); ds;
394 ds = list_next(&dp->dp_synced_datasets, ds)) {
395 bplist_iterate(&ds->ds_pending_deadlist,
396 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
397 }
398
399 while (dstg = txg_list_remove(&dp->dp_sync_tasks, txg)) {
400 /*
401 * No more sync tasks should have been added while we
402 * were syncing.
403 */
404 ASSERT(spa_sync_pass(dp->dp_spa) == 1);
405 dsl_sync_task_group_sync(dstg, tx);
406 }
407 DTRACE_PROBE(pool_sync__3task);
408
409 start = gethrtime();
410 while (dd = txg_list_remove(&dp->dp_dirty_dirs, txg))
411 dsl_dir_sync(dd, tx);
412 write_time += gethrtime() - start;
413
414 start = gethrtime();
415 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
416 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
417 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
418 dmu_objset_sync(mos, zio, tx);
419 err = zio_wait(zio);
420 ASSERT(err == 0);
421 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
422 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
423 }
424 write_time += gethrtime() - start;
425 DTRACE_PROBE2(pool_sync__4io, hrtime_t, write_time,
426 hrtime_t, dp->dp_read_overhead);
427 write_time -= dp->dp_read_overhead;
428
429 dmu_tx_commit(tx);
430
431 dp->dp_space_towrite[txg & TXG_MASK] = 0;
432 ASSERT(dp->dp_tempreserved[txg & TXG_MASK] == 0);
433
434 /*
435 * If the write limit max has not been explicitly set, set it
436 * to a fraction of available physical memory (default 1/8th).
437 * Note that we must inflate the limit because the spa
438 * inflates write sizes to account for data replication.
439 * Check this each sync phase to catch changing memory size.
440 */
441 if (physmem != old_physmem && zfs_write_limit_shift) {
442 mutex_enter(&zfs_write_limit_lock);
443 old_physmem = physmem;
444 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
445 zfs_write_limit_inflated = MAX(zfs_write_limit_min,
446 spa_get_asize(dp->dp_spa, zfs_write_limit_max));
447 mutex_exit(&zfs_write_limit_lock);
448 }
449
450 /*
451 * Attempt to keep the sync time consistent by adjusting the
452 * amount of write traffic allowed into each transaction group.
453 * Weight the throughput calculation towards the current value:
454 * thru = 3/4 old_thru + 1/4 new_thru
455 *
456 * Note: write_time is in nanosecs, so write_time/MICROSEC
457 * yields millisecs
458 */
459 ASSERT(zfs_write_limit_min > 0);
460 if (data_written > zfs_write_limit_min / 8 && write_time > MICROSEC) {
461 uint64_t throughput = data_written / (write_time / MICROSEC);
462
463 if (dp->dp_throughput)
464 dp->dp_throughput = throughput / 4 +
465 3 * dp->dp_throughput / 4;
466 else
467 dp->dp_throughput = throughput;
468 dp->dp_write_limit = MIN(zfs_write_limit_inflated,
469 MAX(zfs_write_limit_min,
470 dp->dp_throughput * zfs_txg_synctime_ms));
471 }
472}
473
474void
475dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
476{
477 dsl_dataset_t *ds;
478 objset_t *os;
479
480 while (ds = list_head(&dp->dp_synced_datasets)) {
481 list_remove(&dp->dp_synced_datasets, ds);
482 os = ds->ds_objset;
483 zil_clean(os->os_zil, txg);
484 ASSERT(!dmu_objset_is_dirty(os, txg));
485 dmu_buf_rele(ds->ds_dbuf, ds);
486 }
487 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
488}
489
490/*
491 * TRUE if the current thread is the tx_sync_thread or if we
492 * are being called from SPA context during pool initialization.
493 */
494int
495dsl_pool_sync_context(dsl_pool_t *dp)
496{
497 return (curthread == dp->dp_tx.tx_sync_thread ||
498 spa_get_dsl(dp->dp_spa) == NULL);
499}
500
501uint64_t
502dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
503{
504 uint64_t space, resv;
505
506 /*
507 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
508 * efficiency.
509 * XXX The intent log is not accounted for, so it must fit
510 * within this slop.
511 *
512 * If we're trying to assess whether it's OK to do a free,
513 * cut the reservation in half to allow forward progress
514 * (e.g. make it possible to rm(1) files from a full pool).
515 */
516 space = spa_get_dspace(dp->dp_spa);
517 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
518 if (netfree)
519 resv >>= 1;
520
521 return (space - resv);
522}
523
524int
525dsl_pool_tempreserve_space(dsl_pool_t *dp, uint64_t space, dmu_tx_t *tx)
526{
527 uint64_t reserved = 0;
528 uint64_t write_limit = (zfs_write_limit_override ?
529 zfs_write_limit_override : dp->dp_write_limit);
530
531 if (zfs_no_write_throttle) {
532 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK],
533 space);
534 return (0);
535 }
536
537 /*
538 * Check to see if we have exceeded the maximum allowed IO for
539 * this transaction group. We can do this without locks since
540 * a little slop here is ok. Note that we do the reserved check
541 * with only half the requested reserve: this is because the
542 * reserve requests are worst-case, and we really don't want to
543 * throttle based off of worst-case estimates.
544 */
545 if (write_limit > 0) {
546 reserved = dp->dp_space_towrite[tx->tx_txg & TXG_MASK]
547 + dp->dp_tempreserved[tx->tx_txg & TXG_MASK] / 2;
548
549 if (reserved && reserved > write_limit)
550 return (ERESTART);
551 }
552
553 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], space);
554
555 /*
556 * If this transaction group is over 7/8ths capacity, delay
557 * the caller 1 clock tick. This will slow down the "fill"
558 * rate until the sync process can catch up with us.
559 */
560 if (reserved && reserved > (write_limit - (write_limit >> 3)))
561 txg_delay(dp, tx->tx_txg, 1);
562
563 return (0);
564}
565
566void
567dsl_pool_tempreserve_clear(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
568{
569 ASSERT(dp->dp_tempreserved[tx->tx_txg & TXG_MASK] >= space);
570 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], -space);
571}
572
573void
574dsl_pool_memory_pressure(dsl_pool_t *dp)
575{
576 uint64_t space_inuse = 0;
577 int i;
578
579 if (dp->dp_write_limit == zfs_write_limit_min)
580 return;
581
582 for (i = 0; i < TXG_SIZE; i++) {
583 space_inuse += dp->dp_space_towrite[i];
584 space_inuse += dp->dp_tempreserved[i];
585 }
586 dp->dp_write_limit = MAX(zfs_write_limit_min,
587 MIN(dp->dp_write_limit, space_inuse / 4));
588}
589
590void
591dsl_pool_willuse_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
592{
593 if (space > 0) {
594 mutex_enter(&dp->dp_lock);
595 dp->dp_space_towrite[tx->tx_txg & TXG_MASK] += space;
596 mutex_exit(&dp->dp_lock);
597 }
598}
599
600/* ARGSUSED */
601static int
602upgrade_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
603{
604 dmu_tx_t *tx = arg;
605 dsl_dataset_t *ds, *prev = NULL;
606 int err;
607 dsl_pool_t *dp = spa_get_dsl(spa);
608
609 err = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
610 if (err)
611 return (err);
612
613 while (ds->ds_phys->ds_prev_snap_obj != 0) {
614 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
615 FTAG, &prev);
616 if (err) {
617 dsl_dataset_rele(ds, FTAG);
618 return (err);
619 }
620
621 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
622 break;
623 dsl_dataset_rele(ds, FTAG);
624 ds = prev;
625 prev = NULL;
626 }
627
628 if (prev == NULL) {
629 prev = dp->dp_origin_snap;
630
631 /*
632 * The $ORIGIN can't have any data, or the accounting
633 * will be wrong.
634 */
635 ASSERT(prev->ds_phys->ds_bp.blk_birth == 0);
636
637 /* The origin doesn't get attached to itself */
638 if (ds->ds_object == prev->ds_object) {
639 dsl_dataset_rele(ds, FTAG);
640 return (0);
641 }
642
643 dmu_buf_will_dirty(ds->ds_dbuf, tx);
644 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
645 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
646
647 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
648 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
649
650 dmu_buf_will_dirty(prev->ds_dbuf, tx);
651 prev->ds_phys->ds_num_children++;
652
653 if (ds->ds_phys->ds_next_snap_obj == 0) {
654 ASSERT(ds->ds_prev == NULL);
655 VERIFY(0 == dsl_dataset_hold_obj(dp,
656 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
657 }
658 }
659
660 ASSERT(ds->ds_dir->dd_phys->dd_origin_obj == prev->ds_object);
661 ASSERT(ds->ds_phys->ds_prev_snap_obj == prev->ds_object);
662
663 if (prev->ds_phys->ds_next_clones_obj == 0) {
664 dmu_buf_will_dirty(prev->ds_dbuf, tx);
665 prev->ds_phys->ds_next_clones_obj =
666 zap_create(dp->dp_meta_objset,
667 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
668 }
669 VERIFY(0 == zap_add_int(dp->dp_meta_objset,
670 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
671
672 dsl_dataset_rele(ds, FTAG);
673 if (prev != dp->dp_origin_snap)
674 dsl_dataset_rele(prev, FTAG);
675 return (0);
676}
677
678void
679dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
680{
681 ASSERT(dmu_tx_is_syncing(tx));
682 ASSERT(dp->dp_origin_snap != NULL);
683
684 VERIFY3U(0, ==, dmu_objset_find_spa(dp->dp_spa, NULL, upgrade_clones_cb,
685 tx, DS_FIND_CHILDREN));
686}
687
688/* ARGSUSED */
689static int
690upgrade_dir_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
691{
692 dmu_tx_t *tx = arg;
693 dsl_dataset_t *ds;
694 dsl_pool_t *dp = spa_get_dsl(spa);
695 objset_t *mos = dp->dp_meta_objset;
696
697 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
698
699 if (ds->ds_dir->dd_phys->dd_origin_obj) {
700 dsl_dataset_t *origin;
701
702 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
703 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
704
705 if (origin->ds_dir->dd_phys->dd_clones == 0) {
706 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
707 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
708 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
709 }
710
711 VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
712 origin->ds_dir->dd_phys->dd_clones, dsobj, tx));
713
714 dsl_dataset_rele(origin, FTAG);
715 }
716
717 dsl_dataset_rele(ds, FTAG);
718 return (0);
719}
720
721void
722dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
723{
724 ASSERT(dmu_tx_is_syncing(tx));
725 uint64_t obj;
726
727 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
728 VERIFY(0 == dsl_pool_open_special_dir(dp,
729 FREE_DIR_NAME, &dp->dp_free_dir));
730
731 /*
732 * We can't use bpobj_alloc(), because spa_version() still
733 * returns the old version, and we need a new-version bpobj with
734 * subobj support. So call dmu_object_alloc() directly.
735 */
736 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
737 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
738 VERIFY3U(0, ==, zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
739 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
740 VERIFY3U(0, ==, bpobj_open(&dp->dp_free_bpobj,
741 dp->dp_meta_objset, obj));
742
743 VERIFY3U(0, ==, dmu_objset_find_spa(dp->dp_spa, NULL,
744 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
745}
746
747void
748dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
749{
750 uint64_t dsobj;
751 dsl_dataset_t *ds;
752
753 ASSERT(dmu_tx_is_syncing(tx));
754 ASSERT(dp->dp_origin_snap == NULL);
755
756 /* create the origin dir, ds, & snap-ds */
757 rw_enter(&dp->dp_config_rwlock, RW_WRITER);
758 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
759 NULL, 0, kcred, tx);
760 VERIFY(0 == dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
761 dsl_dataset_snapshot_sync(ds, ORIGIN_DIR_NAME, tx);
762 VERIFY(0 == dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
763 dp, &dp->dp_origin_snap));
764 dsl_dataset_rele(ds, FTAG);
765 rw_exit(&dp->dp_config_rwlock);
766}
767
768taskq_t *
769dsl_pool_vnrele_taskq(dsl_pool_t *dp)
770{
771 return (dp->dp_vnrele_taskq);
772}
773
774/*
775 * Walk through the pool-wide zap object of temporary snapshot user holds
776 * and release them.
777 */
778void
779dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
780{
781 zap_attribute_t za;
782 zap_cursor_t zc;
783 objset_t *mos = dp->dp_meta_objset;
784 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
785
786 if (zapobj == 0)
787 return;
788 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
789
790 for (zap_cursor_init(&zc, mos, zapobj);
791 zap_cursor_retrieve(&zc, &za) == 0;
792 zap_cursor_advance(&zc)) {
793 char *htag;
794 uint64_t dsobj;
795
796 htag = strchr(za.za_name, '-');
797 *htag = '\0';
798 ++htag;
799 dsobj = strtonum(za.za_name, NULL);
800 (void) dsl_dataset_user_release_tmp(dp, dsobj, htag, B_FALSE);
801 }
802 zap_cursor_fini(&zc);
803}
804
805/*
806 * Create the pool-wide zap object for storing temporary snapshot holds.
807 */
808void
809dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
810{
811 objset_t *mos = dp->dp_meta_objset;
812
813 ASSERT(dp->dp_tmp_userrefs_obj == 0);
814 ASSERT(dmu_tx_is_syncing(tx));
815
816 dp->dp_tmp_userrefs_obj = zap_create(mos, DMU_OT_USERREFS,
817 DMU_OT_NONE, 0, tx);
818
819 VERIFY(zap_add(mos, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS,
820 sizeof (uint64_t), 1, &dp->dp_tmp_userrefs_obj, tx) == 0);
821}
822
823static int
824dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
825 const char *tag, uint64_t *now, dmu_tx_t *tx, boolean_t holding)
826{
827 objset_t *mos = dp->dp_meta_objset;
828 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
829 char *name;
830 int error;
831
832 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
833 ASSERT(dmu_tx_is_syncing(tx));
834
835 /*
836 * If the pool was created prior to SPA_VERSION_USERREFS, the
837 * zap object for temporary holds might not exist yet.
838 */
839 if (zapobj == 0) {
840 if (holding) {
841 dsl_pool_user_hold_create_obj(dp, tx);
842 zapobj = dp->dp_tmp_userrefs_obj;
843 } else {
844 return (ENOENT);
845 }
846 }
847
848 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
849 if (holding)
850 error = zap_add(mos, zapobj, name, 8, 1, now, tx);
851 else
852 error = zap_remove(mos, zapobj, name, tx);
853 strfree(name);
854
855 return (error);
856}
857
858/*
859 * Add a temporary hold for the given dataset object and tag.
860 */
861int
862dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
863 uint64_t *now, dmu_tx_t *tx)
864{
865 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
866}
867
868/*
869 * Release a temporary hold for the given dataset object and tag.
870 */
871int
872dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
873 dmu_tx_t *tx)
874{
875 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
876 tx, B_FALSE));
877}
|