[PATCH 2/3] coresight: Fail to open with return stacks if they are unavailable

[James Clark]

On 17/12/2021 11:52, Mike Leach wrote:
> Hi James
> 
> On Fri, 17 Dec 2021 at 09:41, James Clark <james.clark at arm.com> wrote:
>>
>>
>>
>> On 13/12/2021 09:48, Mike Leach wrote:
>>> Hi James,
>>>
>>> A couple of points - relating mainly to docs:
>>>
>>
>> Hi Mike,
>>
>> Thanks for the feedback. I was in the process of adding some docs and
>> ran into this https://lkml.org/lkml/2021/12/16/1087 so I went to fix
>> that first. Now I will add some more details and resubmit.
>>
>>> 1. Activating branch broadcast overrides any setting of return stack.
>>> As a minimum there needs to be a documentation update to reflect this
>>> -Setting both options is not prohibited in hardware - and in the case
>>> where we can use branch broadcast over a range both are then relevant.
>>>
>>
>> Do you mean that if branch broadcast and return stacks are both requested,
>> but branch broadcast is limited to a range, return stacks will only be
>> available outside that branch broadcast range? But if branch broadcast is
>> enabled for all ranges there will be no return stacks at all?
>>
> 
> Correct - you may want to branch broadcast over a small ranges, but
> otherwise use return stack to decrease the amount of trace in other
> areas.
> Once the complex config sets are accepted upstream then this set-up
> will be possible by writing a config to do this.
> 
>>> 2. A documented note to reflect that choosing this option will result
>>> in a significant increase in the amount of trace generated - possible
>>> danger of overflows, or less actual instructions covered. In addition
>>> perhaps documents could reflect the intended use-case for this option,
>>> given the disadvantages.
>>>
>>
>> Will do.
>>
>>> 3. Has this been tested in a situation where it will be of use?
>>> Testing against static code images will show the same decoded trace
>>> output as not using branch broadcast. (although the packet dumps will
>>> show additional output)>
>>> Given a primary use is for situations where code is patched or
>>> dynamically altered at runtime - then this can affect the full decode
>>> output. If the code is being patched to only alter the branch
>>> addresses then decode should work against static images.
>>> If, however, we are tracing code that adds in new branches, on top of
>>> NOPs for example, then the decoding against the original static image
>>> will be wrong, as the image will have the NOPs, rather than the branch
>>> instructions so the apparent location of E atoms will be in a
>>> different position to the actual code. Is there anything in perf that
>>> will ensure that the patched code is presented to the decoder?
>>>
>>> If there are potential decode issues - these too need documenting.
>>>
>>
>> I'm not sure this should be a blocking issue for this set. Branch broadcast
>> could already be enabled by setting the mode via sysfs. And the perf decode
>> part isn't necessarily a step in the workflow, maybe someone wants to gather
>> data for another tool.
>>
> 
> Someone could set this in sysfs - when collecting data via sysfs.  In
> this case they would not be using perf for decode anyway as you say.
> 
> What you have added here is a new method for perf to set this feature
> and perf always starts off with a clean configuration - then adds
> according to command line options.
> Therefore this will be the first time anyone will have been able to
> set this for a perf session.
> 
> It there are potential limitations, then we need to be clear about
> them - then users of perf, and other hypothetical tools can make a
> good choice about if this option is appropriate - and we avoid mailing
> list complaints (or at least can point to the docs) if users find
> issues that they were warned about!
> 
>> I will do some testing after this change though, but I imagine we would have
>> had issues reported it it wasn't working already which lowers the priority.
>>
> 
> really depends on what was tested! If it is being used over static
> code then the subsequent decode will be fine,
> 

I've posted v2 of this change with some docs changes. I also did some testing
around JITed code in java to see if there were any bad errors which I put
in the cover letter. Although I think the main point of adding this format
flag for perf is to open another avenue of testing or for people who just want
to look at the raw trace manually for now.

James

> Regards
> 
> Mike
> 
>> James
>>
>>> Other than the documents and testing,  I cannot see any issues with
>>> this patch set in terms of setting and enabling the option.
>>>
>>> Regards
>>>
>>> Mike
>>>
>>>
>>> On Fri, 10 Dec 2021 at 17:22, Mathieu Poirier
>>> <mathieu.poirier at linaro.org> wrote:
>>>>
>>>> Hi James,
>>>>
>>>> On Thu, Dec 09, 2021 at 11:13:55AM +0000, James Clark wrote:
>>>>>
>>>>>
>>>>> On 09/12/2021 11:00, Suzuki K Poulose wrote:
>>>>>> On 08/12/2021 16:09, James Clark wrote:
>>>>>>> Maintain consistency with the other options by failing to open when they
>>>>>>> aren't supported. For example ETM_OPT_TS, ETM_OPT_CTXTID2 and the newly
>>>>>>> added ETM_OPT_BRANCH_BROADCAST all return with -EINVAL if they are
>>>>>>> requested but not supported by hardware.
>>>>>>>
>>>>>>> The consequence of not doing this is that the user may not be
>>>>>>> aware that they are not enabling the feature as it is silently disabled.
>>>>>>>
>>>>>>> Signed-off-by: James Clark <james.clark at arm.com>
>>>>>>> ---
>>>>>>>   drivers/hwtracing/coresight/coresight-etm4x-core.c | 13 +++++++++----
>>>>>>>   1 file changed, 9 insertions(+), 4 deletions(-)
>>>>>>>
>>>>>>> diff --git a/drivers/hwtracing/coresight/coresight-etm4x-core.c b/drivers/hwtracing/coresight/coresight-etm4x-core.c
>>>>>>> index d2bafb50c66a..0a9bb943a5e5 100644
>>>>>>> --- a/drivers/hwtracing/coresight/coresight-etm4x-core.c
>>>>>>> +++ b/drivers/hwtracing/coresight/coresight-etm4x-core.c
>>>>>>> @@ -674,10 +674,15 @@ static int etm4_parse_event_config(struct coresight_device *csdev,
>>>>>>>       }
>>>>>>>         /* return stack - enable if selected and supported */
>>>>>>> -    if ((attr->config & BIT(ETM_OPT_RETSTK)) && drvdata->retstack)
>>>>>>> -        /* bit[12], Return stack enable bit */
>>>>>>> -        config->cfg |= BIT(12);
>>>>>>> -
>>>>>>> +    if (attr->config & BIT(ETM_OPT_RETSTK)) {
>>>>>>> +        if (!drvdata->retstack) {
>>>>>>> +            ret = -EINVAL;
>>>>>>> +            goto out;
>>>>>>> +        } else {
>>>>>>> +            /* bit[12], Return stack enable bit */
>>>>>>> +            config->cfg |= BIT(12);
>>>>>>> +        }
>>>>>>
>>>>>> nit: While at this, please could you change the hard coded value
>>>>>> to ETM4_CFG_BIT_RETSTK ?
>>>>>>
>>>>> I started changing them all because I had trouble searching for bits by name but then
>>>>> I thought it would snowball into a bigger change so I undid it.
>>>>>
>>>>> I think I'll just go and do it now if it's an issue here.
>>>>
>>>> I can apply this set right away and you send another patch to fix all hard coded
>>>> bitfields or you can send another revision with all 4 patches included in it
>>>> (bitfields fix plus these 3).  Just let me know what you want to do.  And next
>>>> time please add a cover letter.
>>>>
>>>> Thanks,
>>>> Mathieu
>>>>
>>>>>
>>>>>> Otherwise, looks good to me
>>>>>>
>>>>>> Suzuki
>>>
>>>
>>>
> 
> 
>