[PATCH] powerpc/perf: prevent mixed EBB and non-EBB events

EBB events must be under exclusive groups, so there is no mix of EBB and
non-EBB events on the same PMU. This requirement worked fine as perf core
would not allow other pinned events to be scheduled together with exclusive
events.

This assumption was broken by commit 1908dc911792 ("perf: Tweak
perf_event_attr::exclusive semantics").

After that, the test cpu_event_pinned_vs_ebb_test started succeeding after
read_events, but worse, the task would not have given access to PMC1, so
when it tried to write to it, it was killed with "illegal instruction".

Preventing mixed EBB and non-EBB events from being add to the same PMU will
just revert to the previous behavior and the test will succeed.

Fixes: 1908dc911792 (perf: Tweak perf_event_attr::exclusive semantics)
Signed-off-by: Thadeu Lima de Souza Cascardo <redacted>
---
 arch/powerpc/perf/core-book3s.c | 20 ++++++++++++++++----
 1 file changed, 16 insertions(+), 4 deletions(-)

diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c
index 43599e671d38..d767f7944f85 100644
--- a/arch/powerpc/perf/core-book3s.c
+++ b/arch/powerpc/perf/core-book3s.c

@@ -1010,9 +1010,25 @@ static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
 			  int n_prev, int n_new)
 {
 	int eu = 0, ek = 0, eh = 0;
+	bool ebb = false;
 	int i, n, first;
 	struct perf_event *event;
 
+	n = n_prev + n_new;
+	if (n <= 1)
+		return 0;
+
+	first = 1;
+	for (i = 0; i < n; ++i) {
+		event = ctrs[i];
+		if (first) {
+			ebb = is_ebb_event(event);
+			first = 0;
+		} else if (is_ebb_event(event) != ebb) {
+			return -EAGAIN;
+		}
+	}
+
 	/*
 	 * If the PMU we're on supports per event exclude settings then we
 	 * don't need to do any of this logic. NB. This assumes no PMU has both

@@ -1021,10 +1037,6 @@ static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
 	if (ppmu->flags & PPMU_ARCH_207S)
 		return 0;
 
-	n = n_prev + n_new;
-	if (n <= 1)
-		return 0;
-
 	first = 1;
 	for (i = 0; i < n; ++i) {
 		if (cflags[i] & PPMU_LIMITED_PMC_OK) {

-- 
2.27.0

On 24-Feb-2021, at 5:51 PM, Thadeu Lima de Souza Cascardo [off-list ref] wrote:

EBB events must be under exclusive groups, so there is no mix of EBB and
non-EBB events on the same PMU. This requirement worked fine as perf core
would not allow other pinned events to be scheduled together with exclusive
events.

This assumption was broken by commit 1908dc911792 ("perf: Tweak
perf_event_attr::exclusive semantics").

After that, the test cpu_event_pinned_vs_ebb_test started succeeding after
read_events, but worse, the task would not have given access to PMC1, so
when it tried to write to it, it was killed with "illegal instruction".

Preventing mixed EBB and non-EBB events from being add to the same PMU will
just revert to the previous behavior and the test will succeed.

Hi,

Thanks for checking this. I checked your patch which is fixing “check_excludes” to make
sure all events must agree on EBB. But in the PMU group constraints, we already have check for
EBB events. This is in arch/powerpc/perf/isa207-common.c ( isa207_get_constraint function ).

<<>>
mask  |= CNST_EBB_VAL(ebb);
value |= CNST_EBB_MASK;
<<>>

But the above setting for mask and value is interchanged. We actually need to fix here.

Below patch should fix this:

diff --git a/arch/powerpc/perf/isa207-common.c b/arch/powerpc/perf/isa207-common.c
index e4f577da33d8..8b5eeb6fb2fb 100644
--- a/arch/powerpc/perf/isa207-common.c
+++ b/arch/powerpc/perf/isa207-common.c

@@ -447,8 +447,8 @@ int isa207_get_constraint(u64 event, unsigned long *maskp, unsigned long *valp,
         * EBB events are pinned & exclusive, so this should never actually
         * hit, but we leave it as a fallback in case.
         */
-       mask  |= CNST_EBB_VAL(ebb);
-       value |= CNST_EBB_MASK;
+       mask  |= CNST_EBB_MASK;
+       value |= CNST_EBB_VAL(ebb);
 
        *maskp = mask;
        *valp = value;

Can you please try with this patch.

Thanks
Athira

On 05-Mar-2021, at 11:20 AM, Athira Rajeev [off-list ref] wrote:

quoted

On 24-Feb-2021, at 5:51 PM, Thadeu Lima de Souza Cascardo [off-list ref] wrote:

EBB events must be under exclusive groups, so there is no mix of EBB and
non-EBB events on the same PMU. This requirement worked fine as perf core
would not allow other pinned events to be scheduled together with exclusive
events.

This assumption was broken by commit 1908dc911792 ("perf: Tweak
perf_event_attr::exclusive semantics").

After that, the test cpu_event_pinned_vs_ebb_test started succeeding after
read_events, but worse, the task would not have given access to PMC1, so
when it tried to write to it, it was killed with "illegal instruction".

Preventing mixed EBB and non-EBB events from being add to the same PMU will
just revert to the previous behavior and the test will succeed.

Hi,

Thanks for checking this. I checked your patch which is fixing “check_excludes” to make
sure all events must agree on EBB. But in the PMU group constraints, we already have check for
EBB events. This is in arch/powerpc/perf/isa207-common.c ( isa207_get_constraint function ).

<<>>
mask  |= CNST_EBB_VAL(ebb);
value |= CNST_EBB_MASK;
<<>>

But the above setting for mask and value is interchanged. We actually need to fix here.

Hi,

I have sent a patch for fixing this EBB mask/value setting.
This is the link to patch:

powerpc/perf: Fix PMU constraint check for EBB events
https://patchwork.ozlabs.org/project/linuxppc-dev/list/?series=237669

Thanks
Athira