Add txgs-<pool> kstat file
[zfs.git] / module / zfs / txg.c
index 5702a61..17494bc 100644 (file)
@@ -29,6 +29,7 @@
 #include <sys/dsl_pool.h>
 #include <sys/dsl_scan.h>
 #include <sys/callb.h>
+#include <sys/spa_impl.h>
 
 /*
  * Pool-wide transaction groups.
@@ -279,6 +280,8 @@ txg_rele_to_sync(txg_handle_t *th)
 static void
 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
 {
+       hrtime_t start;
+       txg_history_t *th;
        tx_state_t *tx = &dp->dp_tx;
        int g = txg & TXG_MASK;
        int c;
@@ -293,6 +296,15 @@ txg_quiesce(dsl_pool_t *dp, uint64_t txg)
        tx->tx_open_txg++;
 
        /*
+        * Measure how long the txg was open and replace the kstat.
+        */
+       th = dsl_pool_txg_history_get(dp, txg);
+       th->th_kstat.open_time = gethrtime() - th->th_kstat.birth;
+       th->th_kstat.state = TXG_STATE_QUIESCING;
+       dsl_pool_txg_history_put(th);
+       dsl_pool_txg_history_add(dp, tx->tx_open_txg);
+
+       /*
         * Now that we've incremented tx_open_txg, we can let threads
         * enter the next transaction group.
         */
@@ -302,6 +314,8 @@ txg_quiesce(dsl_pool_t *dp, uint64_t txg)
        /*
         * Quiesce the transaction group by waiting for everyone to txg_exit().
         */
+       start = gethrtime();
+
        for (c = 0; c < max_ncpus; c++) {
                tx_cpu_t *tc = &tx->tx_cpu[c];
                mutex_enter(&tc->tc_lock);
@@ -309,6 +323,13 @@ txg_quiesce(dsl_pool_t *dp, uint64_t txg)
                        cv_wait(&tc->tc_cv[g], &tc->tc_lock);
                mutex_exit(&tc->tc_lock);
        }
+
+       /*
+        * Measure how long the txg took to quiesce.
+        */
+       th = dsl_pool_txg_history_get(dp, txg);
+       th->th_kstat.quiesce_time = gethrtime() - start;
+       dsl_pool_txg_history_put(th);
 }
 
 static void
@@ -395,6 +416,8 @@ txg_sync_thread(dsl_pool_t *dp)
 
        start = delta = 0;
        for (;;) {
+               hrtime_t hrstart;
+               txg_history_t *th;
                uint64_t timer, timeout;
                uint64_t txg;
 
@@ -441,11 +464,17 @@ txg_sync_thread(dsl_pool_t *dp)
                tx->tx_syncing_txg = txg;
                cv_broadcast(&tx->tx_quiesce_more_cv);
 
+               th = dsl_pool_txg_history_get(dp, txg);
+               th->th_kstat.state = TXG_STATE_SYNCING;
+               vdev_get_stats(spa->spa_root_vdev, &th->th_vs1);
+               dsl_pool_txg_history_put(th);
+
                dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
                    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
                mutex_exit(&tx->tx_sync_lock);
 
                start = ddi_get_lbolt();
+               hrstart = gethrtime();
                spa_sync(spa, txg);
                delta = ddi_get_lbolt() - start;
 
@@ -458,6 +487,23 @@ txg_sync_thread(dsl_pool_t *dp)
                 * Dispatch commit callbacks to worker threads.
                 */
                txg_dispatch_callbacks(dp, txg);
+
+               /*
+                * Measure the txg sync time determine the amount of I/O done.
+                */
+               th = dsl_pool_txg_history_get(dp, txg);
+               vdev_get_stats(spa->spa_root_vdev, &th->th_vs2);
+               th->th_kstat.sync_time = gethrtime() - hrstart;
+               th->th_kstat.nread = th->th_vs2.vs_bytes[ZIO_TYPE_READ] -
+                   th->th_vs1.vs_bytes[ZIO_TYPE_READ];
+               th->th_kstat.nwritten = th->th_vs2.vs_bytes[ZIO_TYPE_WRITE] -
+                   th->th_vs1.vs_bytes[ZIO_TYPE_WRITE];
+               th->th_kstat.reads = th->th_vs2.vs_ops[ZIO_TYPE_READ] -
+                   th->th_vs1.vs_ops[ZIO_TYPE_READ];
+               th->th_kstat.writes = th->th_vs2.vs_ops[ZIO_TYPE_WRITE] -
+                   th->th_vs1.vs_ops[ZIO_TYPE_WRITE];
+               th->th_kstat.state = TXG_STATE_COMMITTED;
+               dsl_pool_txg_history_put(th);
        }
 }