x86: finding interrupts that aren't being accounted for?

[Rui Paulo]

> On Apr 6, 2015, at 13:38, Adrian Chadd <adrian at FreeBSD.org> wrote:
> 
> On 6 April 2015 at 12:18, John Baldwin <jhb at freebsd.org> wrote:
>> On Monday, April 06, 2015 12:21:29 AM Adrian Chadd wrote:
>>> Hi,
>>> 
>>> I have an .. odd problem on a Lenovo X230.
>>> 
>>> I just threw in a very old wifi card (Intel 3945) into the expresscard
>>> (pcie) slot. Now, we don't have any pcie-hp support in -HEAD just yet,
>>> but i wasn't expecting the system to crawl to a halt.
>>> 
>>> When I unplug it, everything returns to normal.
>>> 
>>> Other cards don't do this.
>>> 
>>> So, I figured it may be interrupt spam - but vmstat -ia shows no
>>> interrupts going crazy.
>>> 
>>> pmcstat -S CPU_CLK_UNHALTED_CORE -T -w 5 doesn't register anything
>>> either - only a handful of background samples.
>>> 
>>> However, /counter/ mode pmc tells a different story - pmcstat -s
>>> CPU_CLK_UNHALTED_CORE -w 1 shows all four cores going at 110% when the
>>> card is inserted, with brief periods of idle. Once I remove the card,
>>> the counters go back down to zero.
>>> 
>>> My working theory is: something is chewing CPU and it's likely
>>> interrupts, but if it is, it's something far, far earlier than the x86
>>> interrupt C code, which counts interrupts and spurious events.
>>> 
>>> So - has anyone diagnosed this stuff on FreeBSD/x86 before? I was kind
>>> of hoping we'd at least get accurate statistics about spurious
>>> interrupts, and if we don't, I'd like to understand why.
>>> 
>>> Thanks!
>> 
>> SMM?  Perhaps SMM doesn't hide itself from PMC counters (but it can hide itself
>> from samples).
>> 
>> If it is SMM there's not really anything you can do about it.  Try getting a
>> KTR_SCHED trace and looking at it in schedgraph.  When I've seen SMM isuses in
>> the past it shows up as hole in the graph where nothing happens in the system.
>> 
>> In your case you could perhaps be getting PCI errors that are triggering the
>> SMM handler.  Perhaps compare pciconf -le before and after to see if there are
>> any changes.
> 
> Hm, ok. Can we extract PCIe errors yet?

Yes, check pciconf.

--
Rui Paulo