Re: Process-wide watchpoints

[Dmitry Vyukov <dvyukov@google.com>]

On Thu, Nov 12, 2020 at 11:31 AM Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Thu, Nov 12, 2020 at 08:46:23AM +0100, Dmitry Vyukov wrote:
>
> > for sampling race detection),
> > number of threads in the process can be up to, say, ~~10K and the
> > watchpoint is intended to be set for a very brief period of time
> > (~~few ms).
>
> Performance is a consideration here, doing lots of IPIs in such a short
> window, on potentially large machines is a DoS risk.
>
> > This can be done today with both perf_event_open and ptrace.
> > However, the problem is that both APIs work on a single thread level
> > (? perf_event_open can be inherited by children, but not for existing
> > siblings). So doing this would require iterating over, say, 10K
>
> One way would be to create the event before the process starts spawning
> threads and keeping it disabled. Then every thread will inherit it, but
> it'll be inactive.
>
> > I see at least one potential problem: what do we do if some sibling
> > thread already has all 4 watchpoints consumed?
>
> That would be immediately avoided by this, since it will have the
> watchpoint reserved per inheriting the event.
>
> Then you can do ioctl(PERF_EVENT_IOC_{MODIFY_ATTRIBUTES,ENABLE,DISABLE})
> to update the watch location and enable/disable it. This _will_ indeed
> result in a shitload of IPIs if the threads are active, but it should
> work.

Aha! That's the possibility I missed.
We will try to prototype this and get back with more questions if/when
we have them.
Thanks!