DAOS-14561 vos: Add garbage collection metrics (#13244)

site_scons/site_tools/go_builder.py

@@ -70,6 +70,8 @@ def _check_go_version(context):
 
         # go version go1.2.3 Linux/amd64
         go_version = out.split(' ')[2].replace('go', '')
+        if '-' in go_version:
+            go_version = go_version.split('-')[0]
         if len([x for x, y in
                 zip(go_version.split('.'), MIN_GO_VERSION.split('.'))
                 if int(x) < int(y)]) > 0:

src/vos/tests/vts_gc.c

@@ -298,7 +298,7 @@ gc_key_test(void **state)
 	int		     rc;
 
 	rc = gc_key_run(args);
-	assert_int_equal(rc, 0);
+	assert_rc_equal(rc, 0);
 }
 
 static int

src/vos/vos_aggregate.c

@@ -2358,7 +2358,12 @@ vos_aggregate_pre_cb(daos_handle_t ih, vos_iter_entry_t *entry,
 	}
 
 	if (rc < 0) {
+		struct vos_agg_metrics *vam = agg_cont2metrics(cont);
+
 		D_ERROR("VOS aggregation failed: "DF_RC"\n", DP_RC(rc));
+		if (vam && vam->vam_fail_count)
+			d_tm_inc_counter(vam->vam_fail_count, 1);
+
 		return rc;
 	}
 
@@ -2431,7 +2436,11 @@ vos_aggregate_post_cb(daos_handle_t ih, vos_iter_entry_t *entry,
 		inc_agg_counter(agg_param, type, AGG_OP_DEL);
 		rc = 0;
 	} else if (rc != 0) {
+		struct vos_agg_metrics *vam = agg_cont2metrics(cont);
+
 		D_ERROR("VOS aggregation failed: %d\n", rc);
+		if (vam && vam->vam_fail_count)
+			d_tm_inc_counter(vam->vam_fail_count, 1);
 
 		/*
 		 * -DER_TX_BUSY error indicates current ilog aggregation
@@ -2442,8 +2451,6 @@ vos_aggregate_post_cb(daos_handle_t ih, vos_iter_entry_t *entry,
 		 * orphan the current entry due to incarnation log semantics.
 		 */
 		if (rc == -DER_TX_BUSY) {
-			struct vos_agg_metrics	*vam = agg_cont2metrics(cont);
-
 			agg_param->ap_in_progress = 1;
 			rc = 0;
 			switch (type) {
@@ -2736,6 +2743,13 @@ vos_aggregate(daos_handle_t coh, daos_epoch_range_t *epr,
 free_agg_data:
 	D_FREE(ad);
 
+	if (rc < 0) {
+		struct vos_agg_metrics *vam = agg_cont2metrics(cont);
+
+		if (vam && vam->vam_fail_count)
+			d_tm_inc_counter(vam->vam_fail_count, 1);
+	}
+
 	return rc;
 }
 

src/vos/vos_common.c

@@ -753,6 +753,12 @@ vos_metrics_alloc(const char *path, int tgt_id)
 	if (rc)
 		D_WARN("Failed to create 'merged_size' telemetry : "DF_RC"\n", DP_RC(rc));
 
+	/* VOS aggregation failed */
+	rc = d_tm_add_metric(&vam->vam_fail_count, D_TM_COUNTER, "aggregation failures", NULL,
+			     "%s/%s/fail_count/tgt_%u", path, VOS_AGG_DIR, tgt_id);
+	if (rc)
+		D_WARN("Failed to create 'fail_count' telemetry : "DF_RC"\n", DP_RC(rc));
+
 	/* Metrics related to VOS checkpointing */
 	vos_chkpt_metrics_init(&vp_metrics->vp_chkpt_metrics, path, tgt_id);
 
@@ -768,6 +774,21 @@ vos_metrics_alloc(const char *path, int tgt_id)
 	if (rc)
 		D_WARN("Failed to create 'nvme_used' telemetry : "DF_RC"\n", DP_RC(rc));
 
+	/* VOS space SCM total metric */
+	rc = d_tm_add_metric(&vsm->vsm_scm_total, D_TM_GAUGE, "SCM space total", "bytes",
+			     "%s/%s/scm_total/tgt_%u", path, VOS_SPACE_DIR, tgt_id);
+	if (rc)
+		D_WARN("Failed to create 'scm_total' telemetry : " DF_RC "\n", DP_RC(rc));
+
+	/* VOS space NVME total metric */
+	rc = d_tm_add_metric(&vsm->vsm_nvme_total, D_TM_GAUGE, "NVME space total", "bytes",
+			     "%s/%s/nvme_total/tgt_%u", path, VOS_SPACE_DIR, tgt_id);
+	if (rc)
+		D_WARN("Failed to create 'nvme_total' telemetry : " DF_RC "\n", DP_RC(rc));
+
+	/** garbage collection metrics */
+	vos_gc_metrics_init(&vp_metrics->vp_gc_metrics, path, tgt_id);
+
 	/* Initialize the vos_space_metrics timeout counter */
 	vsm->vsm_last_update_ts = 0;
 

src/vos/vos_gc.c

@@ -657,6 +657,33 @@ gc_get_container(struct vos_pool *pool)
 	return cont;
 }
 
+static void
+gc_update_stats(struct vos_pool *pool)
+{
+	struct vos_gc_stat    *stat  = &pool->vp_gc_stat;
+	struct vos_gc_stat    *gstat = &pool->vp_gc_stat_global;
+	struct vos_gc_metrics *vgm;
+
+	if (pool->vp_metrics != NULL) {
+		vgm = &pool->vp_metrics->vp_gc_metrics;
+		d_tm_inc_counter(vgm->vgm_cont_del, stat->gs_conts);
+		d_tm_inc_counter(vgm->vgm_obj_del, stat->gs_objs);
+		d_tm_inc_counter(vgm->vgm_dkey_del, stat->gs_dkeys);
+		d_tm_inc_counter(vgm->vgm_akey_del, stat->gs_akeys);
+		d_tm_inc_counter(vgm->vgm_ev_del, stat->gs_recxs);
+		d_tm_inc_counter(vgm->vgm_sv_del, stat->gs_singvs);
+	}
+
+	gstat->gs_conts += stat->gs_conts;
+	gstat->gs_objs += stat->gs_objs;
+	gstat->gs_dkeys += stat->gs_dkeys;
+	gstat->gs_akeys += stat->gs_akeys;
+	gstat->gs_recxs += stat->gs_recxs;
+	gstat->gs_singvs += stat->gs_singvs;
+
+	memset(stat, 0, sizeof(*stat));
+}
+
 /**
  * Run garbage collector for a pool, it returns if all @credits are consumed
  * or there is nothing to be reclaimed.
@@ -671,7 +698,7 @@ gc_reclaim_pool(struct vos_pool *pool, int *credits, bool *empty_ret)
 
 	if (pool->vp_dying) {
 		*empty_ret = true;
-		return 0;
+		D_GOTO(done, rc = 0);
 	}
 
 	/* take an extra ref to avoid concurrent container destroy/free */
@@ -684,7 +711,8 @@ gc_reclaim_pool(struct vos_pool *pool, int *credits, bool *empty_ret)
 			DP_UUID(pool->vp_id), DP_RC(rc));
 		if (cont != NULL)
 			vos_cont_decref(cont);
-		return rc;
+		*empty_ret = false;
+		goto done;
 	}
 
 	*empty_ret = false;
@@ -781,6 +809,9 @@ gc_reclaim_pool(struct vos_pool *pool, int *credits, bool *empty_ret)
 	if (cont != NULL)
 		vos_cont_decref(cont);
 
+done:
+	gc_update_stats(pool);
+
 	return rc;
 }
 
@@ -916,7 +947,7 @@ gc_have_pool(struct vos_pool *pool)
 static void
 gc_log_pool(struct vos_pool *pool)
 {
-	struct vos_gc_stat *stat = &pool->vp_gc_stat;
+	struct vos_gc_stat *stat = &pool->vp_gc_stat_global;
 
 	D_DEBUG(DB_TRACE,
 		"Pool="DF_UUID", GC reclaimed:\n"
@@ -1106,6 +1137,9 @@ int
 vos_gc_pool(daos_handle_t poh, int credits, int (*yield_func)(void *arg),
 	    void *yield_arg)
 {
+	struct d_tm_node_t      *duration = NULL;
+	struct d_tm_node_t      *tight    = NULL;
+	struct d_tm_node_t      *slack    = NULL;
 	struct vos_pool		*pool = vos_hdl2pool(poh);
 	struct vos_tls		*tls  = vos_tls_get(pool->vp_sysdb);
 	struct vos_gc_param	 param;
@@ -1131,24 +1165,44 @@ vos_gc_pool(daos_handle_t poh, int credits, int (*yield_func)(void *arg),
 
 	tls->vtl_gc_running++;
 
+	if (pool->vp_metrics != NULL) {
+		duration = pool->vp_metrics->vp_gc_metrics.vgm_duration;
+		slack    = pool->vp_metrics->vp_gc_metrics.vgm_slack_cnt;
+		tight    = pool->vp_metrics->vp_gc_metrics.vgm_tight_cnt;
+	}
+
 	while (1) {
 		int	creds = param.vgc_credits;
 
+		d_tm_mark_duration_start(duration, D_TM_CLOCK_THREAD_CPUTIME);
+		if (creds == GC_CREDS_TIGHT)
+			d_tm_inc_counter(tight, 1);
+		else
+			d_tm_inc_counter(slack, 1);
+
 		if (credits > 0 && (credits - total) < creds)
 			creds = credits - total;
 
 		total += creds;
 		rc = vos_gc_pool_tight(poh, &creds);
+
 		if (rc) {
 			D_ERROR("GC pool failed: " DF_RC "\n", DP_RC(rc));
+			d_tm_mark_duration_end(duration);
 			break;
 		}
 		total -= creds; /* subtract the remainded credits */
-		if (creds != 0)
+		if (creds != 0) {
+			d_tm_mark_duration_end(duration);
 			break; /* reclaimed everything */
+		}
 
-		if (credits > 0 && total >= credits)
+		if (credits > 0 && total >= credits) {
+			d_tm_mark_duration_end(duration);
 			break; /* consumed all credits */
+		}
+
+		d_tm_mark_duration_end(duration);
 
 		if (vos_gc_yield(&param)) {
 			D_DEBUG(DB_TRACE, "GC pool run aborted\n");
@@ -1199,3 +1253,65 @@ vos_flush_pool(daos_handle_t poh, bool force, uint32_t nr_flush, uint32_t *nr_fl
 
 	return rc;
 }
+
+#define VOS_GC_DIR "vos_gc"
+void
+vos_gc_metrics_init(struct vos_gc_metrics *vgm, const char *path, int tgt_id)
+{
+	int rc;
+
+	/* GC slice duration */
+	rc = d_tm_add_metric(&vgm->vgm_duration, D_TM_DURATION | D_TM_CLOCK_THREAD_CPUTIME,
+			     "GC slice duration", NULL, "%s/%s/duration/tgt_%u", path, VOS_GC_DIR,
+			     tgt_id);
+	if (rc)
+		D_WARN("Failed to create 'duration' telemetry: " DF_RC "\n", DP_RC(rc));
+
+	/* GC container deletion */
+	rc = d_tm_add_metric(&vgm->vgm_cont_del, D_TM_COUNTER, "GC containers deleted", NULL,
+			     "%s/%s/cont_del/tgt_%u", path, VOS_GC_DIR, tgt_id);
+	if (rc)
+		D_WARN("Failed to create 'cont_del' telemetry: " DF_RC "\n", DP_RC(rc));
+
+	/* GC object deletion */
+	rc = d_tm_add_metric(&vgm->vgm_obj_del, D_TM_COUNTER, "GC objects deleted", NULL,
+			     "%s/%s/obj_del/tgt_%u", path, VOS_GC_DIR, tgt_id);
+	if (rc)
+		D_WARN("Failed to create 'obj_del' telemetry: " DF_RC "\n", DP_RC(rc));
+
+	/* GC dkey deletion */
+	rc = d_tm_add_metric(&vgm->vgm_dkey_del, D_TM_COUNTER, "GC dkeys deleted", NULL,
+			     "%s/%s/dkey_del/tgt_%u", path, VOS_GC_DIR, tgt_id);
+	if (rc)
+		D_WARN("Failed to create 'dkey_del' telemetry: " DF_RC "\n", DP_RC(rc));
+
+	/* GC akey deletion */
+	rc = d_tm_add_metric(&vgm->vgm_akey_del, D_TM_COUNTER, "GC akeys deleted", NULL,
+			     "%s/%s/akey_del/tgt_%u", path, VOS_GC_DIR, tgt_id);
+	if (rc)
+		D_WARN("Failed to create 'akey_del' telemetry: " DF_RC "\n", DP_RC(rc));
+
+	/* GC ev deletion */
+	rc = d_tm_add_metric(&vgm->vgm_ev_del, D_TM_COUNTER, "GC ev deleted", NULL,
+			     "%s/%s/ev_del/tgt_%u", path, VOS_GC_DIR, tgt_id);
+	if (rc)
+		D_WARN("Failed to create 'ev_del' telemetry: " DF_RC "\n", DP_RC(rc));
+
+	/* GC sv deletion */
+	rc = d_tm_add_metric(&vgm->vgm_sv_del, D_TM_COUNTER, "GC sv deleted", NULL,
+			     "%s/%s/sv_del/tgt_%u", path, VOS_GC_DIR, tgt_id);
+	if (rc)
+		D_WARN("Failed to create 'sv_del' telemetry: " DF_RC "\n", DP_RC(rc));
+
+	/* GC slack mode runs */
+	rc = d_tm_add_metric(&vgm->vgm_slack_cnt, D_TM_COUNTER, "GC slack mode count", NULL,
+			     "%s/%s/slack_cnt/tgt_%u", path, VOS_GC_DIR, tgt_id);
+	if (rc)
+		D_WARN("Failed to create 'slack_cnt' telemetry: " DF_RC "\n", DP_RC(rc));
+
+	/* GC tight mode runs */
+	rc = d_tm_add_metric(&vgm->vgm_tight_cnt, D_TM_COUNTER, "GC tight mode count", NULL,
+			     "%s/%s/tight_cnt/tgt_%u", path, VOS_GC_DIR, tgt_id);
+	if (rc)
+		D_WARN("Failed to create 'tight_cnt' telemetry: " DF_RC "\n", DP_RC(rc));
+}

src/vos/vos_internal.h

@@ -186,6 +186,19 @@ struct vos_agg_metrics {
 	struct d_tm_node_t	*vam_del_ev;		/* Deleted EV records */
 	struct d_tm_node_t	*vam_merge_recs;	/* Total merged EV records */
 	struct d_tm_node_t	*vam_merge_size;	/* Total merged size */
+	struct d_tm_node_t	*vam_fail_count;	/* Aggregation failed */
+};
+
+struct vos_gc_metrics {
+	struct d_tm_node_t *vgm_duration;  /* Duration of each gc scan */
+	struct d_tm_node_t *vgm_cont_del;  /* containers reclaimed */
+	struct d_tm_node_t *vgm_obj_del;   /* objects reclaimed */
+	struct d_tm_node_t *vgm_dkey_del;  /* dkeys reclaimed */
+	struct d_tm_node_t *vgm_akey_del;  /* akeys reclaimed */
+	struct d_tm_node_t *vgm_ev_del;    /* EV records reclaimed */
+	struct d_tm_node_t *vgm_sv_del;    /* SV records reclaimed */
+	struct d_tm_node_t *vgm_slack_cnt; /* Slack mode count */
+	struct d_tm_node_t *vgm_tight_cnt; /* Tight mode count */
 };
 
 /*
@@ -200,10 +213,14 @@ struct vos_chkpt_metrics {
 };
 
 void vos_chkpt_metrics_init(struct vos_chkpt_metrics *vc_metrics, const char *path, int tgt_id);
+void
+vos_gc_metrics_init(struct vos_gc_metrics *vc_metrics, const char *path, int tgt_id);
 
 struct vos_space_metrics {
 	struct d_tm_node_t	*vsm_scm_used;		/* SCM space used */
 	struct d_tm_node_t	*vsm_nvme_used;		/* NVMe space used */
+	struct d_tm_node_t      *vsm_scm_total;         /* SCM space total */
+	struct d_tm_node_t      *vsm_nvme_total;        /* NVMe space total */
 	uint64_t		 vsm_last_update_ts;	/* Timeout counter */
 };
 
@@ -219,6 +236,7 @@ struct vos_rh_metrics {
 struct vos_pool_metrics {
 	void			*vp_vea_metrics;
 	struct vos_agg_metrics	 vp_agg_metrics;
+	struct vos_gc_metrics    vp_gc_metrics;
 	struct vos_space_metrics vp_space_metrics;
 	struct vos_chkpt_metrics vp_chkpt_metrics;
 	struct vos_rh_metrics	 vp_rh_metrics;
@@ -255,6 +273,8 @@ struct vos_pool {
 	/** btr handle for the container table */
 	daos_handle_t		vp_cont_th;
 	/** GC statistics of this pool */
+	struct vos_gc_stat       vp_gc_stat_global;
+	/** GC per slice statistics of this pool */
 	struct vos_gc_stat	vp_gc_stat;
 	/** link chain on vos_tls::vtl_gc_pools */
 	d_list_t		vp_gc_link;

src/vos/vos_pool.c

@@ -1494,7 +1494,7 @@ vos_pool_query(daos_handle_t poh, vos_pool_info_t *pinfo)
 
 	D_ASSERT(pinfo != NULL);
 	pinfo->pif_cont_nr = pool_df->pd_cont_nr;
-	pinfo->pif_gc_stat = pool->vp_gc_stat;
+	pinfo->pif_gc_stat = pool->vp_gc_stat_global;
 
 	rc = vos_space_query(pool, &pinfo->pif_space, true);
 	if (rc)
@@ -1552,7 +1552,7 @@ vos_pool_ctl(daos_handle_t poh, enum vos_pool_opc opc, void *param)
 	default:
 		return -DER_NOSYS;
 	case VOS_PO_CTL_RESET_GC:
-		memset(&pool->vp_gc_stat, 0, sizeof(pool->vp_gc_stat));
+		memset(&pool->vp_gc_stat_global, 0, sizeof(pool->vp_gc_stat_global));
 		break;
 	case VOS_PO_CTL_SET_POLICY:
 		if (param == NULL)

src/vos/vos_space.c

@@ -393,6 +393,12 @@ vos_space_update_metrics(struct vos_pool *pool)
 	if (!vpm)
 		return;
 
+	if (vpm->vp_space_metrics.vsm_last_update_ts == 0) {
+		/* Set the constant values */
+		d_tm_set_gauge(vpm->vp_space_metrics.vsm_scm_total, pool->vp_pool_df->pd_scm_sz);
+		d_tm_set_gauge(vpm->vp_space_metrics.vsm_nvme_total, pool->vp_pool_df->pd_nvme_sz);
+	}
+
 	now = daos_gettime_coarse();
 	if (now < vpm->vp_space_metrics.vsm_last_update_ts + VOS_SPACE_METRICS_INTV) {
 		return;