commit e61fdc062aa0af94d045ce7d7372cf5195ee92ea
author Primiano Tucci <primiano@google.com> Wed Aug 09 19:53:09 2023 +0100
committer Primiano Tucci <primiano@google.com> Wed Aug 09 19:53:09 2023 +0100
tree 7a89ed29b00a70e5db9b0e656e95e6109cae91db
parent fc215b9b6a0df329b099f37d6d40ab07d5ab8542

probes: fix premature flush ack if a data source acks synchronously

While iterating on the num_data_sources we populate the
pending_flushes_ map. The intended use of that map is:
- Every time a data source acks the flush, we clear
  the correponding entry in the map.
- Once the map is empty, we know all data sources have acked,
  and we can give the cumulative Flush response to traced.
- HOWEVER: if a data source invokes the flush_callback directly
  in the it->second->Flush(...) call we end up nesting a call
  to OnDataSourceFlushComplete while still in the first
  iteration of the Flush loop.
- The final effect is that while we are in the first iteration
  of the Flush loop, we temporarily empty the map and send a
  confirmation straight-away to traced (which then will snapshot
  the trace ahead of time) while we are still flushing.

Bug: 295189870
Bug: 290213426
Test: unfortunately this bug is so subtle and required the alignment
      on so many unfavorable conditions that I struggle to write a
      meaningful test for it.
      PerfettoFtraceIntegrationTest.TestFtraceFlush was intending to
      unveil this sort of bug, but didn't have all the triggering
      conditions.
Change-Id: I84a3f1a89b0c7dddffaef6cccca8f21febf5519a