Re: [profile] amortize atomic hit count increments

[William Lee Irwin III <wli@holomorphy.com>]

William Lee Irwin III <wli@holomorphy.com> wrote:
>> read_proc_profile()
>>  does not flush the per-cpu hashtables because flushing may cause
>>  timeslice overrun on the systems where prof_buffer cacheline bounces
>>  are so problematic as to livelock the timer interrupt.

On Mon, Sep 13, 2004 at 10:05:21PM -0700, Andrew Morton wrote:
> That's a bit of a problem, isn't it?  As we can accumulate an arbitrarily
> large number of hits within the hash table is it not possible that the
> /proc/profile results could be grossly inaccurate?
> If you had two front-ends per cpu to the profiling buffer then the CPU
> which is running the /proc/profile read could tell all the other CPUs to
> flip to their alternate buffer and can then perform accumulation at its
> leisure.

This is superior to no flushing; I'll implement that and send out an
incremental update (or if preferred, an update of this patch).


On Mon, Sep 13, 2004 at 10:05:21PM -0700, Andrew Morton wrote:
> How does oprofile get around this?  I guess in most modes the CPUs are not
> synchronised.
> One wonders how long we should keep flogging the /prof/profile profiling
> code.  What systems are seeing this livelock?

The original bits were merely a consolidation extracted from a since-
dropped feature patch and an unrelated feature patch from mingo and
arjanv; this is an unrelated fix for SGI's stability issue on larger
Altixen. I personally intend to do no further adjustments.


-- wli