Re: [PATCH V2 2/2] firmware: arm_scmi: Skip adding bad duplicates

[Cristian Marussi <cristian.marussi@arm.com>]

On Wed, Sep 04, 2024 at 08:43:24AM +0530, Sibi Sankar wrote:
> Ensure that the bad duplicates reported by the platform firmware doesn't
> get added to the opp-tables.
> 

Hi Sibi,

so if the idea is to make the code more robust when FW sends BAD
duplicates, you necessarily need to properly drop opps in opp_count too.

One other option would be to just loop with xa_for_each BUT opp_count is
used in a number of places...so first of all let's try drop count properly.

Can you try this patch down below, instead of your patch.
If it solves, I will send a patch (after testing it a bit more :D)

Thanks,
Cristian

P.S.: thanks for spotting this, I forgot NOT to trust FW replies as usual :P

--->8---
diff --git a/drivers/firmware/arm_scmi/perf.c b/drivers/firmware/arm_scmi/perf.c
index 397a39729e29..cbac29792d1e 100644
--- a/drivers/firmware/arm_scmi/perf.c
+++ b/drivers/firmware/arm_scmi/perf.c
@@ -340,7 +340,7 @@ static int iter_perf_levels_update_state(struct scmi_iterator_state *st,
 	return 0;
 }
 
-static inline void
+static inline int
 process_response_opp(struct device *dev, struct scmi_perf_domain_info *dom,
 		     struct scmi_opp *opp, unsigned int loop_idx,
 		     const struct scmi_msg_resp_perf_describe_levels *r)
@@ -353,12 +353,16 @@ process_response_opp(struct device *dev, struct scmi_perf_domain_info *dom,
 		le16_to_cpu(r->opp[loop_idx].transition_latency_us);
 
 	ret = xa_insert(&dom->opps_by_lvl, opp->perf, opp, GFP_KERNEL);
-	if (ret)
+	if (ret) {
 		dev_warn(dev, "Failed to add opps_by_lvl at %d for %s - ret:%d\n",
 			 opp->perf, dom->name, ret);
+		return ret;
+	}
+
+	return 0;
 }
 
-static inline void
+static inline int
 process_response_opp_v4(struct device *dev, struct scmi_perf_domain_info *dom,
 			struct scmi_opp *opp, unsigned int loop_idx,
 			const struct scmi_msg_resp_perf_describe_levels_v4 *r)
@@ -371,9 +375,11 @@ process_response_opp_v4(struct device *dev, struct scmi_perf_domain_info *dom,
 		le16_to_cpu(r->opp[loop_idx].transition_latency_us);
 
 	ret = xa_insert(&dom->opps_by_lvl, opp->perf, opp, GFP_KERNEL);
-	if (ret)
+	if (ret) {
 		dev_warn(dev, "Failed to add opps_by_lvl at %d for %s - ret:%d\n",
 			 opp->perf, dom->name, ret);
+		return ret;
+	}
 
 	/* Note that PERF v4 reports always five 32-bit words */
 	opp->indicative_freq = le32_to_cpu(r->opp[loop_idx].indicative_freq);
@@ -382,13 +388,21 @@ process_response_opp_v4(struct device *dev, struct scmi_perf_domain_info *dom,
 
 		ret = xa_insert(&dom->opps_by_idx, opp->level_index, opp,
 				GFP_KERNEL);
-		if (ret)
+		if (ret) {
 			dev_warn(dev,
 				 "Failed to add opps_by_idx at %d for %s - ret:%d\n",
 				 opp->level_index, dom->name, ret);
 
+			/* Cleanup by_lvl too */
+			xa_erase(&dom->opps_by_lvl, opp->perf);
+
+			return ret;
+		}
+
 		hash_add(dom->opps_by_freq, &opp->hash, opp->indicative_freq);
 	}
+
+	return 0;
 }
 
 static int
@@ -396,16 +410,22 @@ iter_perf_levels_process_response(const struct scmi_protocol_handle *ph,
 				  const void *response,
 				  struct scmi_iterator_state *st, void *priv)
 {
+	int ret;
 	struct scmi_opp *opp;
 	struct scmi_perf_ipriv *p = priv;
 
 	opp = &p->perf_dom->opp[st->desc_index + st->loop_idx];
 	if (PROTOCOL_REV_MAJOR(p->version) <= 0x3)
-		process_response_opp(ph->dev, p->perf_dom, opp, st->loop_idx,
-				     response);
+		ret = process_response_opp(ph->dev, p->perf_dom, opp,
+					   st->loop_idx, response);
 	else
-		process_response_opp_v4(ph->dev, p->perf_dom, opp, st->loop_idx,
-					response);
+		ret = process_response_opp_v4(ph->dev, p->perf_dom, opp,
+					      st->loop_idx, response);
+
+	/* Skip BAD duplicates received from firmware */
+	if (ret)
+		return ret == -EBUSY ? 0 : ret;
+
 	p->perf_dom->opp_count++;
 
 	dev_dbg(ph->dev, "Level %d Power %d Latency %dus Ifreq %d Index %d\n",
---8<---