On Wed, 28 Jun 2017 12:32:50 +0100
Mark Rutland [off-list ref] wrote:

On Wed, Jun 28, 2017 at 12:26:02PM +0100, Mark Rutland wrote:

quoted

On Tue, Jun 27, 2017 at 04:07:58PM -0500, Kim Phillips wrote:

quoted

FWIW, there is also this one I saw with mmap-pages set to 5
(pages), which gets rounded up to 8 pages:

Sorry, *what* does the rounding upwards? Userspace, perf core, or the
driver? Where?

SPE implementations may vary from the minimum buffer alignment of the
smallest available page size, so I left the bts userspace tool code's
upwards-rounding code intact for now.  

I'll take this opportunity to submit the SPE perf tool patch in the
form of a reply to this email:  Look for the rounding code in
tools/perf/arch/arm64/util/arm-spe.c:arm_spe_recording_options().

quoted

That's worrying. I'll see if I can reproduce this.

Actually, this might be down to the IDX2OFF() macro being borked for non
power-of-two buffer sizes.

Do you have Will's latest fixes? In his tree there's a commit:

  4f331cd62531dce2 ("squash! drivers/perf: Add support for ARMv8.2 Statistical Profiling Extension")

... which should fix the IDX2OFF() bug.

yes, I've been running with that squash! commit since a couple of days
after I noticed it over a week ago.

It's be good to reproduce the issue if we can, regardless.

FWIW, I couldn't the little I tried today.

Kim