The lock in the “TRACE_EVENT” interface causes the running thread to sleeping about perfetto HOT 1 CLOSED

Hi!

(1) Why does a normally running thread switch to sleeping state? The reason is lock?

In this case, it looks like the reason is a lock

(2) Is there a lock mechanism in the TRACE_EVENT interface? But I didn't find any lock.

Yes, there are some locks, I'm afraid:

• Every thread normally writes to its own shared memory chunk, so it doesn't need to take any lock. When the memory chunk is finished though (every 1Kb), it does take this lock to acquire another chunk:

  perfetto/src/tracing/core/shared_memory_arbiter_impl.cc

  Line 113 in 13e65b2

  std::unique_lock<std::mutex> scoped_lock(lock_);

• The first time each thread sees a Track

  perfetto/include/perfetto/tracing/internal/track_event_internal.h

  Line 252 in 13e65b2

  // If the given track hasn't been seen by the trace writer yet, write a

  it will take a lock:

  perfetto/include/perfetto/tracing/track.h

  Line 364 in 13e65b2

  std::lock_guard<std::mutex> lock(mutex_);

Most of the times these are not taken and tracing goes through the fast path, but sometimes TRACE_EVENT can actually take a lock.

(3) How to avoid thread switching from running to sleeping in a friendly way when using TRACE_EVENT?

I'm afraid there's no solution. Perfetto has to perform cross thread synchronization from time to time and that involves taking a lock in the current implementation.

Please, let me know if you have any other questions.

from perfetto.