From: Willy Tarreau <w@1wt.eu>
Date: Thu, 13 Mar 2025 14:18:58 +0000 (+0100)
Subject: MINOR: cpu-topo: ignore excess of too small clusters
X-Git-Tag: v3.2-dev8~50
X-Git-Url: http://git.ipfire.org/?a=commitdiff_plain;h=72633666065c045ef67135fae611ae150dbc37e5;p=thirdparty%2Fhaproxy.git

MINOR: cpu-topo: ignore excess of too small clusters

On some Arm systems (typically A76/N1) where CPUs can be associated in
pairs, clusters are reported while they have no incidence on I/O etc.
Yet it's possible to have tens of clusters of 2 CPUs each, which is
counter productive since it does not even allow to start enough threads.

Let's detect this situation as soon as there are at least 4 clusters
having each 2 CPUs or less, which is already very suspcious. In this
case, all these clusters will be reset as meaningless. In the worst
case if needed they'll be re-assigned based on L2/L3.
---

diff --git a/src/cpu_topo.c b/src/cpu_topo.c
index dc1d446e1..d4bdd9f7e 100644
--- a/src/cpu_topo.c
+++ b/src/cpu_topo.c
@@ -531,6 +531,7 @@ void cpu_fixup_topology(void)
 	int cpu, cpu2;
 	int curr_id, prev_id;
 	int min_id, neg;
+	int cl_cpu, small_cl;
 
 	/* fill the package id, node id and thread_id. First we'll build a bitmap
 	 * of all unassigned ones so that we can spot the lowest unassigned one
@@ -601,6 +602,73 @@ void cpu_fixup_topology(void)
 		}
 	}
 
+	/* Some machines (typically ARM cortex A76 and Neoverse-N1) report 1
+	 * cluster per pair of cores due to the internal architecture. While
+	 * this can occasionally make sense (i.e. big.LITTLE etc), when there
+	 * are many clusters of few cores, this is totally pointless. Here
+	 * we'll check if there are at least 4 2-cpu clusters, and if so, all
+	 * the 2-cpu clusters will be cancelled.
+	 */
+	cpu_reorder_by_cluster(ha_cpu_topo, cpu_topo_maxcpus);
+
+	curr_id = -1;
+	cl_cpu = small_cl = 0;
+	for (cpu = cpu2 = 0; cpu <= cpu_topo_lastcpu; cpu++) {
+		if (ha_cpu_topo[cpu].cl_gid < 0)
+			continue;
+
+		if (ha_cpu_topo[cpu].st & HA_CPU_F_EXCL_MASK)
+			continue;
+
+		if (ha_cpu_topo[cpu].cl_gid != curr_id) {
+			if (curr_id >= 0 && cl_cpu <= 2)
+				small_cl++;
+			cl_cpu = 0;
+			cpu2 = cpu;
+			curr_id = ha_cpu_topo[cpu].cl_gid;
+		}
+		cl_cpu++;
+	}
+
+	/* last one */
+	if (cl_cpu && cl_cpu <= 2)
+		small_cl++;
+
+	/* here we have the number of small clusters (<=2 cpu) in small_cl */
+	if (small_cl >= 4) {
+		for (cpu = cpu2 = 0; cpu <= cpu_topo_lastcpu; cpu++) {
+			if (ha_cpu_topo[cpu].cl_gid < 0)
+				continue;
+
+			if (ha_cpu_topo[cpu].st & HA_CPU_F_EXCL_MASK)
+				continue;
+
+			if (ha_cpu_topo[cpu].cl_gid != curr_id) {
+				if (curr_id >= 0 && cl_cpu <= 2) {
+					/* small cluster found for curr_id */
+					while (cpu2 < cpu) {
+						if (ha_cpu_topo[cpu2].cl_gid == curr_id)
+							ha_cpu_topo[cpu2].cl_gid = -1;
+						cpu2++;
+					}
+				}
+				cl_cpu = 0;
+				cpu2 = cpu;
+				curr_id = ha_cpu_topo[cpu].cl_gid;
+			}
+			cl_cpu++;
+		}
+
+		/* handle the last cluster */
+		while (curr_id >= 0 && cl_cpu <= 2 && cpu2 < cpu) {
+			if (ha_cpu_topo[cpu2].cl_gid == curr_id)
+				ha_cpu_topo[cpu2].cl_gid = -1;
+			cpu2++;
+		}
+	}
+
+	cpu_reorder_by_index(ha_cpu_topo, cpu_topo_maxcpus);
+
 	/* assign capacity if not filled, based on the number of threads on the
 	 * core: in a same package, SMT-capable cores are generally those
 	 * optimized for performers while non-SMT ones are generally those