Re: [PATCH v3 6/6] perf cs-etm: Synchronize instruction sample with the thread stack

[Mathieu Poirier]

On Sat, Oct 05, 2019 at 05:16:14PM +0800, Leo Yan wrote:
> The synthesized flow use 'tidq->packet' for instruction samples; on the
> other hand, 'tidp->prev_packet' is used to generate the thread stack and
> the branch samples, this results in the instruction samples using one
> packet ahead than thread stack and branch samples ('tidp->prev_packet'
> vs 'tidq->packet').
> 
> This leads to an instruction's callchain error as shows in below
> example:
> 
>   main  1579        100      instructions:
>   	ffff000010214854 perf_event_update_userpage+0x4c ([kernel.kallsyms])
>   	ffff000010214850 perf_event_update_userpage+0x48 ([kernel.kallsyms])
>   	ffff000010219360 perf_swevent_add+0x88 ([kernel.kallsyms])
>   	ffff0000102135f4 event_sched_in.isra.57+0xbc ([kernel.kallsyms])
>   	ffff0000102137a0 group_sched_in+0x60 ([kernel.kallsyms])
>   	ffff000010213b84 flexible_sched_in+0xfc ([kernel.kallsyms])
>   	ffff00001020c0b4 visit_groups_merge+0x12c ([kernel.kallsyms])
> 
> In the callchain log, for the two continuous lines the up line contains
> one child function info and the followed line contains the caller
> function info, and so forth.  So the first two lines are:
> 
>   perf_event_update_userpage+0x4c  => the sampled instruction
>   perf_event_update_userpage+0x48  => the parent function's calling
> 
> The child function and parent function both are the same function
> perf_event_update_userpage(), but this isn't a recursive function, thus
> the sequence for perf_event_update_userpage() calling itself shouldn't
> never happen.  This callchain error is caused by the instruction sample
> using an ahead packet than the thread stack, the thread stack is deferred
> to process the new packet and misses to pop stack if it is just a return
> packet.
> 
> To fix this issue, we can simply change to use 'tidq->prev_packet' to
> generate the instruction samples, this allows the thread stack to push
> and pop synchronously with instruction sample.  Finally, the callchain
> can be displayed correctly as below:
> 
>   main  1579        100      instructions:
> 	ffff000010214854 perf_event_update_userpage+0x4c ([kernel.kallsyms])
> 	ffff000010219360 perf_swevent_add+0x88 ([kernel.kallsyms])
> 	ffff0000102135f4 event_sched_in.isra.57+0xbc ([kernel.kallsyms])
> 	ffff0000102137a0 group_sched_in+0x60 ([kernel.kallsyms])
> 	ffff000010213b84 flexible_sched_in+0xfc ([kernel.kallsyms])
> 	ffff00001020c0b4 visit_groups_merge+0x12c ([kernel.kallsyms])
> 
> Signed-off-by: Leo Yan <leo.yan@xxxxxxxxxx>
> ---
>  tools/perf/util/cs-etm.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c
> index 56e501cd2f5f..fa969dcb45d2 100644
> --- a/tools/perf/util/cs-etm.c
> +++ b/tools/perf/util/cs-etm.c
> @@ -1419,7 +1419,7 @@ static int cs_etm__sample(struct cs_etm_queue *etmq,
>  	struct cs_etm_packet *tmp;
>  	int ret;
>  	u8 trace_chan_id = tidq->trace_chan_id;
> -	u64 instrs_executed = tidq->packet->instr_count;
> +	u64 instrs_executed = tidq->prev_packet->instr_count;
>  
>  	tidq->period_instructions += instrs_executed;
>  
> @@ -1450,7 +1450,7 @@ static int cs_etm__sample(struct cs_etm_queue *etmq,
>  		 */
>  		s64 offset = (instrs_executed - instrs_over - 1);
>  		u64 addr = cs_etm__instr_addr(etmq, trace_chan_id,
> -					      tidq->packet, offset);
> +					      tidq->prev_packet, offset);

I have tested this set in --per-thread mode and things are working as
advertised. Did you see how things look like in CPU-wide scenarios?

Thanks,
Mathieu

>  
>  		ret = cs_etm__synth_instruction_sample(
>  			etmq, tidq, addr, etm->instructions_sample_period);
> -- 
> 2.17.1
>