Re: [PATCH] tracing: Fix race when concurrently splice_read trace_pipe

On Sun, 13 Aug 2023 09:41:33 -0700
Linus Torvalds [off-list ref] wrote:

So if somebody really *really* want sendfile to work for this case, then you

 (a) do need to fix the race in tracing_splice_read_pipe (which you
should do anyway, since you can obviously always use splice() itself,
not sendfile()).

AND

 (b) figure out when 'splice_write()' will always succeed fully and
convince people that we can do that "extended version" of sendfile().

No big deal. I really don't care about splicing trace_pipe anyway. That was
added by others, but the trace_pipe_raw is what really should be used.

Zheng's race needs to be fixed regardless, but I just wanted to make sure
this wasn't some kind of hidden regression, as there were patches to
trace_pipe to make sendfile get fixed in the past.

a29054d9478d0 ("tracing: Fix crash from reading trace_pipe with sendfile")

But that too was triggered by cat. If cat no longer uses sendfile for
trace_pipe, I'm fine with it.

-- Steve