Re: [PATCH v2] tracing: Introduce pipe_cpumask to avoid race on trace_pipes

[Google]

On Fri, 18 Aug 2023 11:53:22 -0400
Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:

> On Fri, 18 Aug 2023 23:23:01 +0900
> Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx> wrote:
> 
> > It uses trace_pipe_raw. I guess if splice(from trace_pipe_raw to virtio-serial)
> > returns -1 and errno == EAGAIN, the trace data will be lost?
> 
> It shouldn't. If it does, then there's likely a bug. The code will block
> and if an interrupt comes in it will return immediately without reading
> from the buffer.
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/trace/trace.c#n8262
> 
> I don't see where it would return -EINTR and consume data, but I may be
> missing something.

Hmm, I suspect the case if the spilice_to_pipe() returns -EAGAIN.

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/trace/trace.c#n8491

It seems not handling such case.

Anyway, I also think something wrong in virtio-serial (or misusing?), since
it can not read anything from the host sometimes. I just setup the virtio-trace
with below patch (ignore EAGAIN).