Messages in this thread | | | Date | Mon, 5 Jan 2015 12:32:52 +0000 | From | Will Deacon <> | Subject | Re: [PATCH] Revert 9fc2105aeaaf56b0cf75296a84702d0f9e64437b to fix pyaudio (and probably more) |
| |
Hi Ted,
On Mon, Jan 05, 2015 at 01:34:36AM +0000, Theodore Ts'o wrote: > On Sun, Jan 04, 2015 at 09:26:59PM +0000, Russell King - ARM Linux wrote: > > > > With the revert in place, we now have insanely small bogomips values > > reported via /proc/cpuinfo when hardware timers are used. That needs > > fixing. > > Why does it need to be fixed? > > It's clear that there are applications that are working OK with the > existing value,
I'm not sure it is that clear -- the reported regression was on a processor that doesn't use the timer-backed delay loop, so the bogomips value will essentially be restored by reverting the patch.
The issue comes on newer CPUs, where there will now be a very small bogomips value reported and (to my knowledge) nobody has yet tried running some affected applications there to see if they can cope.
> and if you change it to fix it for some new applications, but it breaks > for others, then have you considered defining a new interface (perhaps > exported via sysfs) that exports a "sane" value and document that new > applications shoud use the new interface. > > Or if the answer is that no one should be using the bogomips field at > all, then just document *that*, and then leave it be, so that existing > applications don't break.
It never hurts to document our assumptions or anticipated/preferred use-cases but in this case I think bogomips is difficult enough to use on any half-recent SoCs that most developers have either (a) found another way to do what they want (perf counters, clock_gettime) or (b) stopped bothering to guess the CPU frequency when it's not actually needed, so I don't *think* that new applications are such an issue.
Cheers,
Will
| |