[PATCH 1/7] General notification queue with user mmap()'able ring buffer

[David Howells]

Greg KH <gregkh at linuxfoundation.org> wrote:

> So, if that's all that needs to be fixed, can you use the same
> buffer/code if that patch is merged?

I really don't know.  The perf code is complex, partially in hardware drivers
and is tricky to understand - though a chunk of that is the "aux" buffer part;
PeterZ used words like "special" and "magic" and the comments in the code talk
about the hardware writing into the buffer.

__perf_output_begin() does not appear to be SMP safe.  It uses local_cmpxchg()
and local_add() which on x86 lack the LOCK prefix.

stracing the perf command on my test machine, it calls perf_event_open(2) four
times and mmap's each fd it gets back.  I'm guessing that each one maps a
separate buffer for each CPU.

So to use watch_queue based on perf's buffering, you would have to have a
(2^N)+1 pages-sized buffer for each CPU.  So that would be a minimum of 64K of
unswappable memory for my desktop machine, say).  Multiply that by each
process that wants to listen for events...

What I'm aiming for is something that has a single buffer used by all CPUs for
each instance of /dev/watch_queue opened and I'd also like to avoid having to
allocate the metadata page and the aux buffer to save space.  This is locked
memory and cannot be swapped.

Also, perf has to leave a gap in the ring because it uses CIRC_SPACE(), though
that's a minor detail that I guess can't be fixed now.

I'm also slightly concerned that __perf_output_begin() doesn't check if
rb->user->tail has got ahead of rb->user->head or that it's lagging too far
behind.  I doubt it's a serious problem for the kernel since it won't write
outside of the buffer, but userspace might screw up.  I think the worst that
will happen is that userspace will get confused.

One thing I would like is to waive the 2^N size requirement.  I understand
*why* we do that, but I wonder how expensive DIV instructions are for
relatively small divisors.

David