Re: [PATCH] perf/ring_buffer: ensure atomicity and order of updates

[Mark Rutland]

On Fri, May 11, 2018 at 06:22:29PM +0200, Peter Zijlstra wrote:
> On Fri, May 11, 2018 at 11:59:32AM +0100, Mark Rutland wrote:
> > READ_ONCE() and WRITE_ONCE() "helpfully" make a silent fallback to a
> > memcpy in this case, so we're broken today, regardless of this change.
> > 
> > I suspect that in practice we get single-copy-atomicity for the 32-bit
> > halves, and sessions likely produce less than 4GiB of ringbuffer data,
> > so failures would be rare.
> 
> This should not be a problem because of the 32bit adress space limit,
> which would necessarily limit us to the low word.

For the wrapped values, yes.

I thought that the head and tail values were meant to be free-running,
but I can't see where I got that idea from now that I've gone digging
again.

> Also note that in perf_output_put_handle(), where we write ->data_head,
> the store is from an 'unsigned long'. So on 32bit that will result in a
> zero high word. Similarly, in __perf_output_begin() we read ->data_tail
> into an unsigned long, which will discard the high word.

Ah, that's a fair point. So it's just compat userspace that this is
potentially borked for. ;)

> So userspace should always read (head) a zero high word, irrespective of
> a split store (2x32bit), and the kernel will disregard the high word on
> reading (tail), irrespective of what userspace put there.
> 
> This is all a bit subtle, and could probably use a comment, but it ought
> to work..

It would be nice to guarantee that we don't lose 32-bit atomicity by
virtue of {READ,WRITE}_ONCE() falling back to memcpy in this case, so
maybe we should wrap this in some helpers.

I'll see if I can come up with something which isn't hideous, or I might
just pretend I never stumbled across this. :)

Thanks,
Mark.