How does the kernel detect a bus check HPMC?

[mvpel]

CNB: Yeah, I saw that wording in the crashinfo output, but whenever we ran the ts99 through MCA it spit out some rope number with the replacement recommendations, and HP support stopped there.

But after the second hardware replacement, we got to wondering about software causes, so I wanted to get a better handle on how things work down at this level so we can better characterize the problem.

[cnb]

Note: The PCI EH functionality is not supported on HP-UX 11i v1 OS.

http://bizsupport1.austin.hp.com/bc/docs/support/SupportManual/c02542174/c02542174.pdf

Rgds,

[cnb]

Yep.

I've seen *many* HPMC crashes resolved with O/S patches, Drivers and Firmware so don't overlook these when troubleshooting BUS_CHECK HPMC issues.


Rgds,

[mvpel]

Thanks for suggesting that! After a bit of rummaging, I found this in PHNE_27400:

----
( SR:8606287203 CR:JAGae51142 )
In some situations, the driver was posting an incorrect buffer address to the card causing a HPMC.

Resolution:
Driver has been modified to handle this case correctly.
----

So we'll start pulling on this thread and see where it leads us.

This description also helps me visualize how a driver could cause an HPMC, which I wasn't very clear on with all the emphasis on hardware.

[Laurent Menase]

you probably mean PHNE_28799

it is possible, but if it reoccurs you'll probably need to have the crash dump analyzed.

[Dennis Handly]

>so it's possible for software to manipulate the TLB directly

Only the kernel can do this with privileged instructions.

[mvpel]

We do have a crash dump, finally, and I've been doing a bit of digging. We'll see what turns up.

[mvpel]

Not much turned up, unfortunately. The crash dump suggested a hardware problem, as does the TS99 file, but on further investigation it turns out there's been a number of other instances of this type of crash on a variety of different systems, and it seems unlikely that all of them have a hardware problem.

[Laurent Menase]

so we are back on the need to have a true crash dump analyzis

are the panic stack all the same?

[mvpel]

We finally peeled enough of the onion to determine that the HPMC is caused by a reference to a page of shared memory mapped to the card's registers by the driver outside the range of the card's register space.

Per the PCI 2.1 spec, a card is not required to assert the device-selected pin to PCI bus address assertions outside its accepted range of registers, and the cards in question apparently don't.

We were able to reproduce the HPMC on Monday by poking the card's mapped IO memory in the wrong place.