Use a pool of buffers to be used as per-thread trace buffers

We used to allocate a new buffer each time the per-thread trace buffer
was full. The old buffer was enqueued in a task queue which is then
flushed by a thread pool worker. Thread pool worker can be slow to flush
and release these buffers and it increases the memory usage especially
on heavy apps with multiple threads. This causes LMKs and OOMs. It is
not great in general to have unbounded memory usage for tracing. So this
CL allocates a pool of buffers that would be used as a per-thread buffer.
If there's no free buffer, we flush the buffer at the head of the task
queue into the file and use that buffer.

On google maps app, the earlier approach causes LMKs for traces greater
than 45-60 seconds. With this approach, we can collect traces of a much
longer duration. The performance impact is not too noticeable. In the
earlier case a large memory usage was slowing down the app too.

Bug: 259258187
Test: art/test.py --trace --stream
Change-Id: Ic02f26e2cd14b9ffb1a046e119e03fb4cd022305

0 files changed, 0 insertions, 0 deletions