HI Suzuki,
On Mon, 26 Jul 2021 at 17:01, Suzuki K Poulose suzuki.poulose@arm.com wrote:
Hi Mike,
On 26/07/2021 13:34, Mike Leach wrote:
Hi Suzuki,
On Fri, 23 Jul 2021 at 13:46, Suzuki K Poulose suzuki.poulose@arm.com wrote:
The TRBE driver marks the AUX buffer as TRUNCATED when we get an IRQ on FILL event. This has rather unwanted side-effect of the event being disabled when there may be more space in the ring buffer.
So, instead of TRUNCATE we need a different flag to indicate that the trace may have lost a few bytes (i.e from the point of generating the FILL event until the IRQ is consumed). Anyways, the userspace must use the size from RECORD_AUX headers to restrict the "trace" decoding.
Using PARTIAL flag causes the perf tool to generate the following warning:
Warning: AUX data had gaps in it XX times out of YY!
Are you running a KVM guest in the background?
which is pointlessly scary for a user. The other remaining options are :
- COLLISION - Use by SPE to indicate samples collided
- Add a new flag - Specifically for CoreSight, doesn't sound so good, if we can re-use something.
What is the user visible behaviour when using COLLISION?
If you meant a Warning from the perf tool (similar to TRUNCATE or PARTIAL), the answer is none. We could add one in the perf tool if you think this is necessary.
I do - the problem is that we have replaced a visible warning with a silent failure.
While we agree that the side effects of TRUNCATE mean it unfeasible as a solution here - at least the PARTIAL message does give some indication. The average perf user is going to rely on the output from the tool - if there is no warning they will assume all is good, but they have possible non-contiguous trace and no indication of such.
Since we are using a collision flag in a particular context - i.e. coresight trace - we have the chance to provide an appropriate message for this context.
The TRUNCATE warning is at least accurate - even if the KVM thing is something of a red herring.
Sorry - I meant PARTIAL here - but the comment stands otherwise.
It is easier to explain a "scary" warning, than try to debug someones problems if perf is silent or misleading when using the COLLISION flag.
The RECORD_AUX still has this flag. So, if someone really wanted to know how many times the TRBE fired the IRQ and thus potentially lost a few bytes of the trace, they could always look at this.
They could - but how would they know that they needed to - what indicators would they have that the trace was not continuous? The point of the perf tool is that it presents an accurate picture to the user, based on the data collected. Most users aren't going to start digging into the intricacies of the perf data file formats and nor should they have to.
Definitely this is not something similar to "TRUNCATED", which we realized the hard way, nor the PARTIAL. But the perf tool could report something similar. Please remember that the perf tool always uses the "size" field from the RECORD_AUX to limit the trace decoding.
So, I am not sure how this could create new problems.
There is no issue with decode - but if a user is investigating a problem using trace, they need to be aware that some trace might be dropped. That way they can take mitigating action.
Regards
Mike
Suzuki
Regards
Mike
Given that we don't already use the "COLLISION" flag, the above behavior can be notified using this flag for CoreSight.
Cc: Mathieu Poirier mathieu.poirier@linaro.org Cc: James Clark james.clark@arm.com Cc: Mike Leach mike.leach@linaro.org Cc: Anshuman Khandual anshuman.khandual@arm.com Cc: Leo Yan leo.yan@linaro.org Signed-off-by: Suzuki K Poulose suzuki.poulose@arm.com
drivers/hwtracing/coresight/coresight-trbe.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/drivers/hwtracing/coresight/coresight-trbe.c b/drivers/hwtracing/coresight/coresight-trbe.c index 503bea0137ae..d50f142e86d1 100644 --- a/drivers/hwtracing/coresight/coresight-trbe.c +++ b/drivers/hwtracing/coresight/coresight-trbe.c @@ -615,7 +615,7 @@ static unsigned long arm_trbe_update_buffer(struct coresight_device *csdev, * for correct size. Also, mark the buffer truncated. */ write = get_trbe_limit_pointer();
perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
perf_aux_output_flag(handle, PERF_AUX_FLAG_COLLISION); } offset = write - base;@@ -708,7 +708,7 @@ static void trbe_handle_overflow(struct perf_output_handle *handle) * collection upon the WRAP event, without stopping the source. */ perf_aux_output_flag(handle, PERF_AUX_FLAG_CORESIGHT_FORMAT_RAW |
PERF_AUX_FLAG_TRUNCATED);
PERF_AUX_FLAG_COLLISION); perf_aux_output_end(handle, size); event_data = perf_aux_output_begin(handle, event); if (!event_data) {-- 2.24.1