Skip to contentSkip to conversation

interconnect: Aggregate bandwidth votes for unused nodes in sync_state()

PATCH In review Odelu Kukatla 3 participants 3 messages Jun 19, 2021 – Jun 25, 2021

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 interconnect: Aggregate bandwidth votes for unused nodes in sync_state()

When removing the initial bandwidth votes in sync_state(), make sure to call the aggregate() function for nodes which don't have any clients yet. aggregate_requests() does not invoke aggregate() for unused nodes.

---
 drivers/interconnect/core.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
drivers/interconnect/core.c +10 -1
index 8a1e70e..1d9a00a 100644
@@ -1106,7 +1106,16 @@ void icc_sync_state(struct device *dev)
 		dev_dbg(p->dev, "interconnect provider is in synced state\n");
 		list_for_each_entry(n, &p->nodes, node_list) {
 			if (n->init_avg || n->init_peak) {
-				aggregate_requests(n);
+				if (hlist_empty(&n->req_list)) {
+					if (p->pre_aggregate)
+						p->pre_aggregate(n);
+
+					p->aggregate(n, 0, 0, 0, &n->avg_bw,
+						&n->peak_bw);
+				} else {
+					aggregate_requests(n);
+				}
+
 				p->set(n, n);
 			}
 		}
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
#1 lore raw
Quoted 6 lines · Odelu Kukatla

When removing the initial bandwidth votes in sync_state(), make sure to call the aggregate() function for nodes which don't have any clients yet. aggregate_requests() does not invoke aggregate() for unused nodes.

It seems this should have a 'Fixes' tag for b1d681d8d324 ("interconnect:
Add sync state support")', to make sure the change makes it into the stable trees.

Quoted 9 lines
---
 drivers/interconnect/core.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
drivers/interconnect/core.c +0 -0
index 8a1e70e..1d9a00a 100644

drivers/interconnect/core.c void icc_sync_state(struct device *dev) · line 1109

Quoted 5 lines
drivers/interconnect/core.c +1 -1
@@ -1106,7 +1106,16 @@ void icc_sync_state(struct device *dev)
 		dev_dbg(p->dev, "interconnect provider is in synced state\n");
 		list_for_each_entry(n, &p->nodes, node_list) {
 			if (n->init_avg || n->init_peak) {
-				aggregate_requests(n);
+				if (hlist_empty(&n->req_list)) {

nit: consider handling the common case in the 'if' branch and the exception of the initial votes in 'else'.

Quoted 12 lines
drivers/interconnect/core.c +9 -0
+					if (p->pre_aggregate)
+						p->pre_aggregate(n);
+
+					p->aggregate(n, 0, 0, 0, &n->avg_bw,
+						&n->peak_bw);
+				} else {
+					aggregate_requests(n);
+				}
+
 				p->set(n, n);
 			}
 		}
#2 lore raw
In reply to Odelu Kukatla
Quoted 23 lines · Odelu Kukatla

When removing the initial bandwidth votes in sync_state(), make sure to call the aggregate() function for nodes which don't have any clients yet. aggregate_requests() does not invoke aggregate() for unused nodes.

---
  drivers/interconnect/core.c | 11 ++++++++++-
  1 file changed, 10 insertions(+), 1 deletion(-)
drivers/interconnect/core.c +3 -1
index 8a1e70e..1d9a00a 100644
@@ -1106,7 +1106,16 @@ void icc_sync_state(struct device *dev)
  		dev_dbg(p->dev, "interconnect provider is in synced state\n");
  		list_for_each_entry(n, &p->nodes, node_list) {
  			if (n->init_avg || n->init_peak) {
-				aggregate_requests(n);
+				if (hlist_empty(&n->req_list)) {
+					if (p->pre_aggregate)
+						p->pre_aggregate(n);

drivers/interconnect/core.c void icc_sync_state(struct device *dev) · line 1121

Quoted 10 lines
drivers/interconnect/core.c +7 -0
+
+					p->aggregate(n, 0, 0, 0, &n->avg_bw,
+						&n->peak_bw);
+				} else {
+					aggregate_requests(n);
+				}
+
  				p->set(n, n);
  			}
  		}

I ran into this issue myself last week. There is an alternative fix that I think is a little cleaner. The reason we need the aggregate() call here is that the icc-rpmh providers only add BCMs to the commit list in the aggregate() callback. If aggregate() isn't called, then we don't commit anything to HW. But, if we instead add BCMs to the commit list in pre_aggregate(), then the existing aggregate_requests() call is sufficient, since that always calls pre_aggregate() even when req_list is empty. That means we'd need to update icc_node_add() to call pre_aggregate() before aggregate(), otherwise we wouldn't commit the initial floors. But I think always calling pre_aggregate() before aggregate() is a reasonable requirement.

I've posted these changes in a patch series along with a separate sync-state-related fix. The solutions could co-exist, but they are largely redundant.