It was found that when the perf-record command was running on large
system with many cores and GHES enabled, the performance of the
workload being monitored will slow down significantly. For example,
# ./hackbench
Time: 0.047
# perf record ./hackbench
Time: 21.392
The big slowdown was traced to the fact that too much time (more
than 1ms) was spent in each invocation of ghes_notify_nmi(). The
perf-record command will cause NMI to be generated periodically at
high frequency. If many cores are doing that, it will cause a long
queue to form in the global ghes_nmi_lock lock.
This patch limits only up to 2 cpus to be working on ghes_notify_nmi()
at any given instance. This will significantly reduce the time spent
by each CPU in ghes_notify_nmi(). With that patch, the perf slowdown
was significantly reduced.
# perf record ./hackbench
Time: 0.133
Signed-off-by: Waiman Long <Waiman.Long@xxxxxx>
---
drivers/acpi/apei/ghes.c | 15 +++++++++++++++
1 files changed, 15 insertions(+), 0 deletions(-)
diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c
index e82d097..132a014 100644
--- a/drivers/acpi/apei/ghes.c
+++ b/drivers/acpi/apei/ghes.c
@@ -803,6 +803,20 @@ static int ghes_notify_nmi(unsigned int cmd, struct pt_regs *regs)
int sev, sev_global = -1;
int ret = NMI_DONE;
+ /* # of CPUs processing NMI */
+ static atomic_t ghes_nmi_cnt = ATOMIC_INIT(0);
+
+ /*
+ * Multiple CPUs may get NMI and come here to check GHES. If one has
+ * gotten the lock and is doing the check, there is no point in having
+ * more than one other CPU waiting just in case the active CPU has
+ * past the point where the NMI was generated. So we set a limit of
+ * only allowing 2 CPUs working here. The rests can just return.
+ */
+ if (atomic_add_return(1, &ghes_nmi_cnt) > 2) {
+ atomic_dec(&ghes_nmi_cnt);
+ return ret;
+ }
raw_spin_lock(&ghes_nmi_lock);
list_for_each_entry_rcu(ghes, &ghes_nmi, list) {
if (ghes_read_estatus(ghes, 1)) {
@@ -863,6 +877,7 @@ next:
#endif
out:
+ atomic_dec(&ghes_nmi_cnt);
raw_spin_unlock(&ghes_nmi_lock);
return ret;
}
--
1.7.1
--
To unsubscribe from this list: send the line "unsubscribe linux-acpi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
