Re: [RFC PATCH] perf/x86: improve the event scheduling to avoid unnecessary pmu_stop/start

[Peter Zijlstra]

On Thu, Feb 10, 2022 at 12:39:30PM +0800, Wen Yang wrote:
> This issue has been there for a long time, we could reproduce it as follows:
> 
> 1, run a script that periodically collects perf data, eg:
> while true
> do
>     perf stat -e cache-misses,cache-misses,cache-misses -C 1 sleep 2
>     perf stat -e cache-misses -C 1 sleep 2
>     sleep 1
> done
> 
> 2, run another one to capture the IPC, eg:
> perf stat -e cycles:D,instructions:D  -C 1 -I 1000
> 
> Then we could observe that the counter used by cycles:D changes frequently:
> crash> struct  cpu_hw_events.n_events,assign,event_list,events  ffff88bf7f44f420
>   n_events = 3
>   assign = {33, 1, 32, 0, 1, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
>   event_list = {0xffff88bf77b85000, 0xffff88b72db82000, 0xffff88b72db85800, 0xffff88ff6cfcb000, 0xffff88ff609f1800, 0xffff88ff609f1800, 0xffff88ff5f46a800, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}
>   events = {0x0, 0xffff88b72db82000, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0xffff88b72db85800, 0xffff88bf77b85000, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}
> 
> crash> struct  cpu_hw_events.n_events,assign,event_list,events  ffff88bf7f44f420
>   n_events = 6
>   assign = {33, 3, 32, 0, 1, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
>   event_list = {0xffff88bf77b85000, 0xffff88b72db82000, 0xffff88b72db85800, 0xffff88bf46c34000, 0xffff88bf46c35000, 0xffff88bf46c30000, 0xffff88ff5f46a800, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}
>   events = {0xffff88bf46c34000, 0xffff88bf46c35000, 0xffff88bf46c30000, 0xffff88b72db82000, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0xffff88b72db85800, 0xffff88bf77b85000, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}

The above is unreadable; consider rewriting it in something other than a
raw data dump.

> The reason is that NMI watchdog permanently consumes one FP,
> so cycles can only use one GP, and its hweight is 5,
> but the hweight of other events (cache misses) is 4,
> so the counter used by cycles will be frequently taken away,
> resulting in unnecessary pmu_stop/start.

And the solution..... ? Or should I try and reverse engineer from the
proposed patch? If I do, how do I know it is what you intended?