Re: [patch] x86, perf: avoid spamming kernel log for bts buffer failure

[Peter Zijlstra <peterz@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 1758 bytes --]

On Wed, Jul 02, 2014 at 03:16:40PM +0200, Stephane Eranian wrote:
> On Tue, Jul 1, 2014 at 11:34 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> > On Mon, Jun 30, 2014 at 04:04:08PM -0700, David Rientjes wrote:
> >> It's unnecessary to excessively spam the kernel log anytime the BTS buffer
> >> cannot be allocated, so make this allocation __GFP_NOWARN.
> >>
> >> The user probably will want to at least find some artifact that the
> >> allocation has failed in the past, probably due to fragmentation because
> >> of its large size, when it's not allocated at bootstrap.  Thus, add a
> >> WARN_ONCE() so something is left behind for them to understand why perf
> >> commnads that require PEBS is not working properly.
> >
> > Can you elaborate a bit under which conditions this triggered? Typically
> > we should be doing fairly well allocating such buffers with GFP_KERNEL,
> > that should allow things like compaction to run and create higher order
> > pages.
> >
> I think this triggers when you have fragmented memory and you have
> perf_events active and inactive (i.e., 0 users = no nmi watchdog) frequently.
> Each first user invokes the reserve_ds() function to reserve DS, PEBS, BTS.

Right, that'd suck. I suppose we could also change that to allocate the
DS resources on first demand and never free them again.

So only allocate the PEBS buffer when we create the first PEBS event,
and idem for the BTS muck.

> The reason for BTS rather then PEBS is the size of the allocation.
> PEBS allocates one page, i.e., less likely to get a failure than BTS
> which allocates 4 pages, I think.

Sure..

> David and I discussed this. He can probably add more background
> info, if needed.

It would still be good to see why compaction etc is failing.

[-- Attachment #2: Type: application/pgp-signature, Size: 836 bytes --]