[PATCH] powerpc/perf: Fix oops when grouping different pmu events

[PATCH] powerpc/perf: Fix oops when grouping different pmu events

[Ravi Bangoria]

When user tries to group imc (In-Memory Collections) event with
normal event, (sometime) kernel crashes with following log:

    Faulting instruction address: 0x00000000
    [link register   ] c00000000010ce88 power_check_constraints+0x128/0x980
    ...
    c00000000010e238 power_pmu_event_init+0x268/0x6f0
    c0000000002dc60c perf_try_init_event+0xdc/0x1a0
    c0000000002dce88 perf_event_alloc+0x7b8/0xac0
    c0000000002e92e0 SyS_perf_event_open+0x530/0xda0
    c00000000000b004 system_call+0x38/0xe0

'event_base' field of 'struct hw_perf_event' is used as flags for
normal hw events and used as memory address for imc events. While
grouping these two types of events, collect_events() tries to
interpret imc 'event_base' as a flag, which causes a corruption
resulting in a crash.

Consider only those events which belongs to 'perf_hw_context' in
collect_events().

Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
---
 arch/powerpc/perf/core-book3s.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c
index 9e3da16..1538129 100644
--- a/arch/powerpc/perf/core-book3s.c
+++ b/arch/powerpc/perf/core-book3s.c
@@ -1415,7 +1415,7 @@ static int collect_events(struct perf_event *group, int max_count,
 	int n = 0;
 	struct perf_event *event;
 
-	if (!is_software_event(group)) {
+	if (group->pmu->task_ctx_nr == perf_hw_context) {
 		if (n >= max_count)
 			return -1;
 		ctrs[n] = group;
@@ -1423,7 +1423,7 @@ static int collect_events(struct perf_event *group, int max_count,
 		events[n++] = group->hw.config;
 	}
 	list_for_each_entry(event, &group->sibling_list, group_entry) {
-		if (!is_software_event(event) &&
+		if (event->pmu->task_ctx_nr == perf_hw_context &&
 		    event->state != PERF_EVENT_STATE_OFF) {
 			if (n >= max_count)
 				return -1;
-- 
1.8.3.1

Re: [PATCH] powerpc/perf: Fix oops when grouping different pmu events

[Madhavan Srinivasan]

On Thursday 30 November 2017 02:03 PM, Ravi Bangoria wrote:
> When user tries to group imc (In-Memory Collections) event with
> normal event, (sometime) kernel crashes with following log:
>
>      Faulting instruction address: 0x00000000
>      [link register   ] c00000000010ce88 power_check_constraints+0x128/0x980
>      ...
>      c00000000010e238 power_pmu_event_init+0x268/0x6f0
>      c0000000002dc60c perf_try_init_event+0xdc/0x1a0
>      c0000000002dce88 perf_event_alloc+0x7b8/0xac0
>      c0000000002e92e0 SyS_perf_event_open+0x530/0xda0
>      c00000000000b004 system_call+0x38/0xe0
>
> 'event_base' field of 'struct hw_perf_event' is used as flags for
> normal hw events and used as memory address for imc events. While
> grouping these two types of events, collect_events() tries to
> interpret imc 'event_base' as a flag, which causes a corruption
> resulting in a crash.
>
> Consider only those events which belongs to 'perf_hw_context' in
> collect_events().

Reviewed-By: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>

>
> Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
> ---
>   arch/powerpc/perf/core-book3s.c | 4 ++--
>   1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c
> index 9e3da16..1538129 100644
> --- a/arch/powerpc/perf/core-book3s.c
> +++ b/arch/powerpc/perf/core-book3s.c
> @@ -1415,7 +1415,7 @@ static int collect_events(struct perf_event *group, int max_count,
>   	int n = 0;
>   	struct perf_event *event;
>
> -	if (!is_software_event(group)) {
> +	if (group->pmu->task_ctx_nr == perf_hw_context) {
>   		if (n >= max_count)
>   			return -1;
>   		ctrs[n] = group;
> @@ -1423,7 +1423,7 @@ static int collect_events(struct perf_event *group, int max_count,
>   		events[n++] = group->hw.config;
>   	}
>   	list_for_each_entry(event, &group->sibling_list, group_entry) {
> -		if (!is_software_event(event) &&
> +		if (event->pmu->task_ctx_nr == perf_hw_context &&
>   		    event->state != PERF_EVENT_STATE_OFF) {
>   			if (n >= max_count)
>   				return -1;

Re: powerpc/perf: Fix oops when grouping different pmu events

[Michael Ellerman]

On Thu, 2017-11-30 at 08:33:22 UTC, Ravi Bangoria wrote:
> When user tries to group imc (In-Memory Collections) event with
> normal event, (sometime) kernel crashes with following log:
> 
>     Faulting instruction address: 0x00000000
>     [link register   ] c00000000010ce88 power_check_constraints+0x128/0x980
>     ...
>     c00000000010e238 power_pmu_event_init+0x268/0x6f0
>     c0000000002dc60c perf_try_init_event+0xdc/0x1a0
>     c0000000002dce88 perf_event_alloc+0x7b8/0xac0
>     c0000000002e92e0 SyS_perf_event_open+0x530/0xda0
>     c00000000000b004 system_call+0x38/0xe0
> 
> 'event_base' field of 'struct hw_perf_event' is used as flags for
> normal hw events and used as memory address for imc events. While
> grouping these two types of events, collect_events() tries to
> interpret imc 'event_base' as a flag, which causes a corruption
> resulting in a crash.
> 
> Consider only those events which belongs to 'perf_hw_context' in
> collect_events().
> 
> Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
> Reviewed-By: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>

Applied to powerpc fixes, thanks.

https://git.kernel.org/powerpc/c/5aa04b3eb6fca63d2e9827be656dca

cheers