Add new kstat for monitoring time in dmu_tx_assign
[zfs.git] / module / zfs / dsl_pool.c
index 85c745e..7795d80 100644 (file)
@@ -42,6 +42,7 @@
 #include <sys/dsl_deadlist.h>
 #include <sys/bptree.h>
 #include <sys/zfeature.h>
+#include <sys/zil_impl.h>
 
 int zfs_no_write_throttle = 0;
 int zfs_write_limit_shift = 3;                 /* 1/8th of physical memory */
@@ -57,6 +58,63 @@ kmutex_t zfs_write_limit_lock;
 
 static pgcnt_t old_physmem = 0;
 
+static void
+dsl_pool_tx_assign_init(dsl_pool_t *dp, unsigned int ndata)
+{
+       kstat_named_t *ks;
+       char name[KSTAT_STRLEN];
+       int i, data_size = ndata * sizeof(kstat_named_t);
+
+       (void) snprintf(name, KSTAT_STRLEN, "dmu_tx_assign-%s",
+                       spa_name(dp->dp_spa));
+
+       dp->dp_tx_assign_size = ndata;
+
+       if (data_size)
+               dp->dp_tx_assign_buckets = kmem_alloc(data_size, KM_SLEEP);
+       else
+               dp->dp_tx_assign_buckets = NULL;
+
+       for (i = 0; i < dp->dp_tx_assign_size; i++) {
+               ks = &dp->dp_tx_assign_buckets[i];
+               ks->data_type = KSTAT_DATA_UINT64;
+               ks->value.ui64 = 0;
+               (void) snprintf(ks->name, KSTAT_STRLEN, "%u us", 1 << i);
+       }
+
+       dp->dp_tx_assign_kstat = kstat_create("zfs", 0, name, "misc",
+           KSTAT_TYPE_NAMED, 0, KSTAT_FLAG_VIRTUAL);
+
+       if (dp->dp_tx_assign_kstat) {
+               dp->dp_tx_assign_kstat->ks_data = dp->dp_tx_assign_buckets;
+               dp->dp_tx_assign_kstat->ks_ndata = dp->dp_tx_assign_size;
+               dp->dp_tx_assign_kstat->ks_data_size = data_size;
+               kstat_install(dp->dp_tx_assign_kstat);
+       }
+}
+
+static void
+dsl_pool_tx_assign_destroy(dsl_pool_t *dp)
+{
+       if (dp->dp_tx_assign_buckets)
+               kmem_free(dp->dp_tx_assign_buckets,
+                         dp->dp_tx_assign_size * sizeof(kstat_named_t));
+
+       if (dp->dp_tx_assign_kstat)
+               kstat_delete(dp->dp_tx_assign_kstat);
+}
+
+void
+dsl_pool_tx_assign_add_usecs(dsl_pool_t *dp, uint64_t usecs)
+{
+       uint64_t idx = 0;
+
+       while (((1 << idx) < usecs) && (idx < dp->dp_tx_assign_size - 1))
+               idx++;
+
+       atomic_inc_64(&dp->dp_tx_assign_buckets[idx].value.ui64);
+}
+
 static int
 dsl_pool_txg_history_update(kstat_t *ksp, int rw)
 {
@@ -142,7 +200,7 @@ dsl_pool_txg_history_add(dsl_pool_t *dp, uint64_t txg)
 {
        txg_history_t *th, *rm;
 
-       th = kmem_zalloc(sizeof(txg_history_t), KM_SLEEP);
+       th = kmem_zalloc(sizeof(txg_history_t), KM_PUSHPAGE);
        mutex_init(&th->th_lock, NULL, MUTEX_DEFAULT, NULL);
        th->th_kstat.txg = txg;
        th->th_kstat.state = TXG_STATE_OPEN;
@@ -224,12 +282,12 @@ dsl_pool_open_impl(spa_t *spa, uint64_t txg)
 
        txg_list_create(&dp->dp_dirty_datasets,
            offsetof(dsl_dataset_t, ds_dirty_link));
+       txg_list_create(&dp->dp_dirty_zilogs,
+           offsetof(zilog_t, zl_dirty_link));
        txg_list_create(&dp->dp_dirty_dirs,
            offsetof(dsl_dir_t, dd_dirty_link));
        txg_list_create(&dp->dp_sync_tasks,
            offsetof(dsl_sync_task_group_t, dstg_node));
-       list_create(&dp->dp_synced_datasets, sizeof (dsl_dataset_t),
-           offsetof(dsl_dataset_t, ds_synced_link));
 
        mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
 
@@ -237,6 +295,7 @@ dsl_pool_open_impl(spa_t *spa, uint64_t txg)
            1, 4, 0);
 
        dsl_pool_txg_history_init(dp, txg);
+       dsl_pool_tx_assign_init(dp, 32);
 
        return (dp);
 }
@@ -321,6 +380,15 @@ dsl_pool_open(dsl_pool_t *dp)
                        goto out;
        }
 
+       if (spa_feature_is_active(dp->dp_spa,
+           &spa_feature_table[SPA_FEATURE_EMPTY_BPOBJ])) {
+               err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
+                   DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
+                   &dp->dp_empty_bpobj);
+               if (err != 0)
+                       goto out;
+       }
+
        err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
            DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
            &dp->dp_tmp_userrefs_obj);
@@ -362,13 +430,14 @@ dsl_pool_close(dsl_pool_t *dp)
                dmu_objset_evict(dp->dp_meta_objset);
 
        txg_list_destroy(&dp->dp_dirty_datasets);
+       txg_list_destroy(&dp->dp_dirty_zilogs);
        txg_list_destroy(&dp->dp_sync_tasks);
        txg_list_destroy(&dp->dp_dirty_dirs);
-       list_destroy(&dp->dp_synced_datasets);
 
        arc_flush(dp->dp_spa);
        txg_fini(dp);
        dsl_scan_fini(dp);
+       dsl_pool_tx_assign_destroy(dp);
        dsl_pool_txg_history_destroy(dp);
        rw_destroy(&dp->dp_config_rwlock);
        mutex_destroy(&dp->dp_lock);
@@ -395,7 +464,7 @@ dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
        /* create the pool directory */
        err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
            DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
-       ASSERT3U(err, ==, 0);
+       ASSERT0(err);
 
        /* Initialize scan structures */
        VERIFY3U(0, ==, dsl_scan_init(dp, txg));
@@ -445,6 +514,21 @@ dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
        return (dp);
 }
 
+/*
+ * Account for the meta-objset space in its placeholder dsl_dir.
+ */
+void
+dsl_pool_mos_diduse_space(dsl_pool_t *dp,
+    int64_t used, int64_t comp, int64_t uncomp)
+{
+       ASSERT3U(comp, ==, uncomp); /* it's all metadata */
+       mutex_enter(&dp->dp_lock);
+       dp->dp_mos_used_delta += used;
+       dp->dp_mos_compressed_delta += comp;
+       dp->dp_mos_uncompressed_delta += uncomp;
+       mutex_exit(&dp->dp_lock);
+}
+
 static int
 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
 {
@@ -463,11 +547,14 @@ dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
        dmu_tx_t *tx;
        dsl_dir_t *dd;
        dsl_dataset_t *ds;
-       dsl_sync_task_group_t *dstg;
        objset_t *mos = dp->dp_meta_objset;
        hrtime_t start, write_time;
        uint64_t data_written;
        int err;
+       list_t synced_datasets;
+
+       list_create(&synced_datasets, sizeof (dsl_dataset_t),
+           offsetof(dsl_dataset_t, ds_synced_link));
 
        /*
         * We need to copy dp_space_towrite() before doing
@@ -490,7 +577,7 @@ dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
                 * may sync newly-created datasets on pass 2.
                 */
                ASSERT(!list_link_active(&ds->ds_synced_link));
-               list_insert_tail(&dp->dp_synced_datasets, ds);
+               list_insert_tail(&synced_datasets, ds);
                dsl_dataset_sync(ds, zio, tx);
        }
        DTRACE_PROBE(pool_sync__1setup);
@@ -500,15 +587,20 @@ dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
        ASSERT(err == 0);
        DTRACE_PROBE(pool_sync__2rootzio);
 
-       for (ds = list_head(&dp->dp_synced_datasets); ds;
-           ds = list_next(&dp->dp_synced_datasets, ds))
+       /*
+        * After the data blocks have been written (ensured by the zio_wait()
+        * above), update the user/group space accounting.
+        */
+       for (ds = list_head(&synced_datasets); ds;
+           ds = list_next(&synced_datasets, ds))
                dmu_objset_do_userquota_updates(ds->ds_objset, tx);
 
        /*
         * Sync the datasets again to push out the changes due to
         * userspace updates.  This must be done before we process the
-        * sync tasks, because that could cause a snapshot of a dataset
-        * whose ds_bp will be rewritten when we do this 2nd sync.
+        * sync tasks, so that any snapshots will have the correct
+        * user accounting information (and we won't get confused
+        * about which blocks are part of the snapshot).
         */
        zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
        while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg))) {
@@ -519,30 +611,42 @@ dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
        err = zio_wait(zio);
 
        /*
-        * Move dead blocks from the pending deadlist to the on-disk
-        * deadlist.
+        * Now that the datasets have been completely synced, we can
+        * clean up our in-memory structures accumulated while syncing:
+        *
+        *  - move dead blocks from the pending deadlist to the on-disk deadlist
+        *  - clean up zil records
+        *  - release hold from dsl_dataset_dirty()
         */
-       for (ds = list_head(&dp->dp_synced_datasets); ds;
-           ds = list_next(&dp->dp_synced_datasets, ds)) {
+       while ((ds = list_remove_head(&synced_datasets))) {
+               ASSERTV(objset_t *os = ds->ds_objset);
                bplist_iterate(&ds->ds_pending_deadlist,
                    deadlist_enqueue_cb, &ds->ds_deadlist, tx);
+               ASSERT(!dmu_objset_is_dirty(os, txg));
+               dmu_buf_rele(ds->ds_dbuf, ds);
        }
 
-       while ((dstg = txg_list_remove(&dp->dp_sync_tasks, txg))) {
-               /*
-                * No more sync tasks should have been added while we
-                * were syncing.
-                */
-               ASSERT(spa_sync_pass(dp->dp_spa) == 1);
-               dsl_sync_task_group_sync(dstg, tx);
-       }
-       DTRACE_PROBE(pool_sync__3task);
-
        start = gethrtime();
        while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)))
                dsl_dir_sync(dd, tx);
        write_time += gethrtime() - start;
 
+       /*
+        * The MOS's space is accounted for in the pool/$MOS
+        * (dp_mos_dir).  We can't modify the mos while we're syncing
+        * it, so we remember the deltas and apply them here.
+        */
+       if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
+           dp->dp_mos_uncompressed_delta != 0) {
+               dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
+                   dp->dp_mos_used_delta,
+                   dp->dp_mos_compressed_delta,
+                   dp->dp_mos_uncompressed_delta, tx);
+               dp->dp_mos_used_delta = 0;
+               dp->dp_mos_compressed_delta = 0;
+               dp->dp_mos_uncompressed_delta = 0;
+       }
+
        start = gethrtime();
        if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
            list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
@@ -558,6 +662,27 @@ dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
            hrtime_t, dp->dp_read_overhead);
        write_time -= dp->dp_read_overhead;
 
+       /*
+        * If we modify a dataset in the same txg that we want to destroy it,
+        * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
+        * dsl_dir_destroy_check() will fail if there are unexpected holds.
+        * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
+        * and clearing the hold on it) before we process the sync_tasks.
+        * The MOS data dirtied by the sync_tasks will be synced on the next
+        * pass.
+        */
+       DTRACE_PROBE(pool_sync__3task);
+       if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
+               dsl_sync_task_group_t *dstg;
+               /*
+                * No more sync tasks should have been added while we
+                * were syncing.
+                */
+               ASSERT(spa_sync_pass(dp->dp_spa) == 1);
+               while ((dstg = txg_list_remove(&dp->dp_sync_tasks, txg)))
+                       dsl_sync_task_group_sync(dstg, tx);
+       }
+
        dmu_tx_commit(tx);
 
        dp->dp_space_towrite[txg & TXG_MASK] = 0;
@@ -606,15 +731,14 @@ dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
 void
 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
 {
+       zilog_t *zilog;
        dsl_dataset_t *ds;
-       objset_t *os;
 
-       while ((ds = list_head(&dp->dp_synced_datasets))) {
-               list_remove(&dp->dp_synced_datasets, ds);
-               os = ds->ds_objset;
-               zil_clean(os->os_zil, txg);
-               ASSERT(!dmu_objset_is_dirty(os, txg));
-               dmu_buf_rele(ds->ds_dbuf, ds);
+       while ((zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg))) {
+               ds = dmu_objset_ds(zilog->zl_os);
+               zil_clean(zilog, txg);
+               ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
+               dmu_buf_rele(ds->ds_dbuf, zilog);
        }
        ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
 }