Re: [PATCH V4] POWER: perf_event: Skip updating kernel counters ifregister value shrinks

From: Eric B Munson
Date: Wed Apr 27 2011 - 08:19:14 EST


On Wed, 27 Apr 2011, David Laight wrote:

> I keep telling Eric that the code below is incorrect
> modulo arithimetic...

But it isn't, and it doesn't have trouble with 2^32 - 1. Here is one done by
hand:

Counter is at 0xffffffff and is rolled over to 0x101 (258 counted items so
that we miss the test for rollback).

0x00000000ffffffff (Remember counters are 32 bit, but we store them in 64)
-0x0000000000000101
=0xffffffff00000102

After the mask we have 0x00000000000102, the actual difference between the
counters.

>
> > +static u64 check_and_compute_delta(u64 prev, u64 val)
> > +{
> > + u64 delta = (val - prev) & 0xfffffffful;
> > +
> > + /*
> > + * POWER7 can roll back counter values, if the new value is
> smaller
> > + * than the previous value it will cause the delta and the
> counter to
> > + * have bogus values unless we rolled a counter over. If a
> coutner is
> > + * rolled back, it will be smaller, but within 256, which is the
> maximum
> > + * number of events to rollback at once. If we dectect a
> rollback
> > + * return 0. This can lead to a small lack of precision in the
> > + * counters.
> > + */
> > + if (prev > val && (prev - val) < 256)
> > + delta = 0;
> > +
> > + return delta;
>
> The code should detect rollback by looking at the value of 'delta'
> otherwise there are horrid end effects near 2^32-1.
>
> For instance:
> u32 delta = val - prev;
> return delta & 0x80000000 ? 0 : delta;
>
>
> David
>
>
>
>

Attachment: signature.asc
Description: Digital signature