Fix EventPipe sample profiler resolution on Windows. #58997
Conversation
If sample profiler was initially created before sampling has been enabled, happens when setting up sessions during startup, sample_profiler_load_dependecies won't get called, meaning we won't setup timeBeginPeriod/timeEndPeriod on Windows, and won't adjust default scheduling resolution to 1ms staying on default 16ms. There is a ref count check in current sample_profiler_enable checking if we should call sample_profiler_load_dependecies (_ref_count == 0), but in case where we can't start profiling _can_start_sampling == false, we won't call sample_profiler_enable but always increase _ref_count that in turn will prevent calls to sample_profiler_load_dependecies when enabling sample profiler and that in turn won't call timeBeginPeriod staying on default scheduling resolution. Fix is to always call sample_profiler_load_dependecies making sure we will setup needed dependencies when _ref_count == 0.
ghost
added
the
|
//CC @josalem |
josalem
left a comment
josalem
left a comment
There was a problem hiding this comment.
It's worth pointing out that this only affects Windows tracing sessions that suspend the runtime on startup to capture startup events. This should affect any other scenarios.
Thanks @lateralusX
Yes, it only happens if there are sessions created before EventPipe init finished (as part of doing suspended startup), but also applies to the default file based session controlled through COMPlus_EnableEventPipe environment variable. |
|
/backport to release/6.0 |
|
Started backporting to release/6.0: https://github.com/dotnet/runtime/actions/runs/1230885023 |
2 participants
If sample profiler was initially created before sampling has been enabled, happens when setting up sessions during startup,
sample_profiler_load_dependecies won't get called, meaning we won't setup timeBeginPeriod/timeEndPeriod on Windows, and won't adjust default scheduling resolution to 1ms staying on default 16ms.
There is a ref count check in current sample_profiler_enable checking if we should call sample_profiler_load_dependecies (_ref_count == 0), but in case where we can't start profiling _can_start_sampling == false, we won't call sample_profiler_enable but always increase _ref_count that in turn will prevent calls to sample_profiler_load_dependecies when enabling sample profiler and that in turn won't call timeBeginPeriod staying on default scheduling resolution.
Fix is to always call sample_profiler_load_dependecies making sure we will setup needed dependencies when _ref_count == 0.