[PATCH v2 18/19] ARC: [plat-eznps] replace sync with proper cpu barrier

[Peter Zijlstra]

On Tue, Nov 17, 2015 at 07:02:00PM +0530, Vineet Gupta wrote:
> > You should never report more than PERF_MAX_STACK_DEPTH thingies anyway,
> > so once you've done that many loops, you're good to bail, right?
> 
> Yeah, although I need to ensure if arch code needs to check that. Plus the
> unwinder is also used for things like ps wchan, /proc/<pid>/stack, crash dump etc.

Ah, ok. On x86 we have a separate unwinder for perf. Trying to touch
userspace memory from NMI context is a tad tricky.

(imagine the NMI interrupting a page-fault and then itself triggering a
page-fault again)

> > You're likely to call this code from interrupt/NMI context, there is no
> > ^C or scheduling going to help you there.
> 
> For perf case, but there are other uses of unwinder as described above.
> 
> So in light of above, do u think that the cond_resched() + signal_pending check is
> not needed and the bail if over N iters will be fine !

It can't hurt, but you want to engineer it to be robust for the most
constrained environment, so I would say the bail over N iters had
better be robust, otherwise you're screwed anyhow ;-)