Re: perf: rdpmc and PERF_EVENT_IOC_RESET

[Peter Zijlstra]

On Wed, Aug 17, 2016 at 12:53:01AM -0400, Vince Weaver wrote:
> Hello
> 
> so using rdpmc() and the mmap page to do fast perf_event reads seems to 
> interact poorly with the PERF_EVENT_IOC_RESET ioctl.

Hurm.. I never considered using RESET with rdpmc(). The typical usage of
rdpmc() I considered is something like:

	u64 start = rdpmc();

	/* code goes here */

	delta = rdpmc() - start;


That said; I don't object to fixing the interaction if it doesn't have
downsides.

> From what I can tell, on reset event->count is set to zero, but 
> event->hw.prev_count is not, so the userpg->offset field ends up negative 
> and weird things happen.

_perf_event_reset()
  perf_event_read() /* updates prev_count */
  local64_set(count, 0);


after that, like you say: userpg->offset = perf_event_count() -
prev_count, which does indeed end up negative.

But the rdpmc user function should:

u64 mmap_read_self(void *addr)
{
        struct perf_event_mmap_page *pc = addr;
        u32 seq, idx, time_mult = 0, time_shift = 0, width = 0;
        u64 count, cyc = 0, time_offset = 0, enabled, running, delta;
        s64 pmc = 0;

        do {
                seq = pc->lock;
                barrier();

		/* enabled/running muck */

                idx = pc->index;
                count = pc->offset;
                if (pc->cap_user_rdpmc && idx) {
                        width = pc->pmc_width;
                        pmc = rdpmc(idx - 1);
                }

                barrier();
        } while (pc->lock != seq);

        if (idx) {
                pmc <<= 64 - width;
                pmc >>= 64 - width; /* shift right signed */
                count += pmc;
        }

	/* more enabled/running muck */

        return count;
}

Which sign-extends the RDPMC result and adds it to userpg->offset,
resulting in a positive number again.


So I think it should work..