numatop/0013-common-Fix-perf-init-for-large-systems.patch
Pingfan Liu 6a533d52ba add x86 zen support
Resolves: bz2088798
Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=2088798

Signed-off-by: Pingfan Liu <piliu@redhat.com>
2023-09-25 21:02:31 -04:00

59 lines
1.6 KiB
Diff

From b4543efe798bbc255519fdcec73484cbd43472d1 Mon Sep 17 00:00:00 2001
From: Sandipan Das <sandipan.das@amd.com>
Date: Mon, 18 Apr 2022 10:59:26 +0530
Subject: [PATCH 13/15] common: Fix perf init for large systems
Large systems with hundreds of CPUs can run into issues
during perf event initialization because of the default
resource limits for file descriptors. Set RLIMIT_NOFILE
explicitly to a fairly large value to avoid them.
Signed-off-by: Sandipan Das <sandipan.das@amd.com>
---
common/os/os_perf.c | 16 ++++++++++++++++
1 file changed, 16 insertions(+)
diff --git a/common/os/os_perf.c b/common/os/os_perf.c
index 44ca43d..49fdaaa 100644
--- a/common/os/os_perf.c
+++ b/common/os/os_perf.c
@@ -28,6 +28,7 @@
#include <inttypes.h>
#include <stdlib.h>
+#include <sys/resource.h>
#include <sys/types.h>
#include <unistd.h>
#include <string.h>
@@ -850,12 +851,27 @@ ll_init(pf_conf_t *conf)
int
os_perf_init(void)
{
+ struct rlimit limit;
int ringsize, size;
s_profiling_recbuf = NULL;
s_ll_recbuf = NULL;
s_partpause_enabled = B_FALSE;
+ /*
+ * Depending on the number of available CPUs in the system, the
+ * default fd limit may be exceeded. Set it to a large value to
+ * avoid running into problems.
+ */
+ limit.rlim_cur = 32768;
+ limit.rlim_max = 32768;
+
+ if (setrlimit(RLIMIT_NOFILE, &limit) < 0) {
+ exit_msg_put("Failed to setup perf!\n");
+ debug_print(NULL, 2, "os_perf_init failed\n");
+ return (-1);
+ }
+
ringsize = pf_ringsize_init();
size = ((ringsize / sizeof (pf_profiling_rbrec_t)) + 1) *
sizeof (pf_profiling_rec_t);
--
2.31.1