[RFC PATCH v2 2/2] x86/perf/amd: Resolve NMI latency issues when multiple PMCs are active

[Lendacky, Thomas]

On AMD processors, the detection of an overflowed PMC counter in the NMI
handler relies on the current value of the PMC. So, for example, to check
for overflow on a 48-bit counter, bit 47 is checked to see if it is 1 (not
overflowed) or 0 (overflowed).

When the perf NMI handler executes it does not know in advance which PMC
counters have overflowed. As such, the NMI handler will process all active
PMC counters that have overflowed. NMI latency in newer AMD processors can
result in multiple overflowed PMC counters being processed in one NMI and
then a subsequent NMI, that does not appear to be a back-to-back NMI, not
finding any PMC counters that have overflowed. This may appear to be an
unhandled NMI resulting in either a panic or a series of messages,
depending on how the kernel was configured.

To mitigate this issue, add an AMD handle_irq callback function,
amd_pmu_handle_irq(), that will invoke the common x86_pmu_handle_irq()
function and upon return perform some additional processing that will
indicate if the NMI has been handled or would have been handled had an
earlier NMI not handled the overflowed PMC. Using a per-CPU variable, a
minimum value of the number of active PMCs or 2 will be set whenever a
PMC is active. This is used to indicate the possible number of NMIs that
can still occur. The value of 2 is used for when an NMI does not arrive
at the LAPIC in time to be collapsed into an already pending NMI. Each
time the function is called without having handled an overflowed counter,
the per-CPU value is checked. If the value is non-zero, it is decremented
and the NMI indicates that it handled the NMI. If the value is zero, then
the NMI indicates that it did not handle the NMI.

This issue is different from a previous issue related to perf NMIs that
was fixed in commit:
  63e6be6d98e1 ("perf, x86: Catch spurious interrupts after disabling counters")

The difference here is that the NMI latency can contribute to what appear
to be spurious NMIs during the handling of PMC counter overflow while the
counter is active as opposed to when the counter is being disabled.

Cc: <stable@xxxxxxxxxxxxxxx> # 4.14.x-
Signed-off-by: Tom Lendacky <thomas.lendacky@xxxxxxx>
---
 arch/x86/events/amd/core.c |   58 +++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 57 insertions(+), 1 deletion(-)

diff --git a/arch/x86/events/amd/core.c b/arch/x86/events/amd/core.c
index 732402dcffdc..3bc7a92de8af 100644
--- a/arch/x86/events/amd/core.c
+++ b/arch/x86/events/amd/core.c
@@ -5,9 +5,12 @@
 #include <linux/slab.h>
 #include <linux/delay.h>
 #include <asm/apicdef.h>
+#include <asm/nmi.h>
 
 #include "../perf_event.h"
 
+static DEFINE_PER_CPU(unsigned int, perf_nmi_counter);
+
 static __initconst const u64 amd_hw_cache_event_ids
 				[PERF_COUNT_HW_CACHE_MAX]
 				[PERF_COUNT_HW_CACHE_OP_MAX]
@@ -508,6 +511,59 @@ void amd_pmu_disable_all(void)
 	}
 }
 
+/*
+ * Because of NMI latency, if multiple PMC counters are active or other sources
+ * of NMIs are received, the perf NMI handler can handle one or more overflowed
+ * PMC counters outside of the NMI associated with the PMC overflow. If the NMI
+ * doesn't arrive at the LAPIC in time to become a pending NMI, then the kernel
+ * back-to-back NMI support won't be active. This PMC handler needs to take into
+ * account that this can occur, otherwise this could result in unknown NMI
+ * messages being issued. Examples of this is PMC overflow while in the NMI
+ * handler when multiple PMCs are active or PMC overflow while handling some
+ * other source of an NMI.
+ *
+ * Attempt to mitigate this by using the number of active PMCs to determine
+ * whether to return NMI_HANDLED if the perf NMI handler did not handle/reset
+ * any PMCs. The per-CPU perf_nmi_counter variable is set to a minimum of the
+ * number of active PMCs or 2. The value of 2 is used in case an NMI does not
+ * arrive at the LAPIC in time to be collapsed into an already pending NMI.
+ */
+static int amd_pmu_handle_irq(struct pt_regs *regs)
+{
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	int active, handled;
+
+	active = __bitmap_weight(cpuc->active_mask, X86_PMC_IDX_MAX);
+	handled = x86_pmu_handle_irq(regs);
+
+	/*
+	 * If no counters are active, reset perf_nmi_counter and return
+	 * NMI_DONE
+	 */
+	if (!active) {
+		this_cpu_write(perf_nmi_counter, 0);
+		return NMI_DONE;
+	}
+
+	/*
+	 * If a counter was handled, record the number of possible remaining
+	 * NMIs that can occur.
+	 */
+	if (handled) {
+		this_cpu_write(perf_nmi_counter,
+			       min_t(unsigned int, 2, active));
+
+		return handled;
+	}
+
+	if (!this_cpu_read(perf_nmi_counter))
+		return NMI_DONE;
+
+	this_cpu_dec(perf_nmi_counter);
+
+	return NMI_HANDLED;
+}
+
 static struct event_constraint *
 amd_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
 			  struct perf_event *event)
@@ -700,7 +756,7 @@ static ssize_t amd_event_sysfs_show(char *page, u64 config)
 
 static __initconst const struct x86_pmu amd_pmu = {
 	.name			= "AMD",
-	.handle_irq		= x86_pmu_handle_irq,
+	.handle_irq		= amd_pmu_handle_irq,
 	.disable_all		= amd_pmu_disable_all,
 	.enable_all		= x86_pmu_enable_all,
 	.enable			= x86_pmu_enable_event,