On Tue, May 14, 2013 at 5:33 PM, Eric W. Biederman <ebiederm at xmission.com> wrote: > Dave Lloyd <dave at davelloyd.com> writes: > >> On Tue, May 14, 2013 at 5:01 PM, Eric W. Biederman >> <ebiederm at xmission.com> wrote: >> >>> >>> Yes this does seem to be all over the place, and memory corruption >>> probably caused by ongoing-dma seems like a reasonable hypothesis. >> >> Thank goodness it's not just me! :-) > > It is a classic issue, although I suspect something is unique in your > setup because it has (to my knowledge) not been a widespread problem for > years. It could certainly be buggy hardware. Other details include: Kernel 3.0.29.0 and we are also using infiniband (which I believe I found a reference to the Mellanox hardware potentially causing this issue unless the driver was unloaded before reboot with kexec). The potential issue with unloading the IB drivers doesn't bug me nearly as much as not unloading pata_amd and pata_acpi causing the ACPI Error messages upon reboot with kexec. I'm inclined to chalk the ACPI Error mesages up to potentially buggy BIOS/hardware from the vendor since pata_amd and pata_acpi are in wide use and I would expect to see more issues reported were there truly an issue with rebooting with kexec and not unloading pata_amd and pata_acpi. >>> The easy first thing to try is to remove all of your kernel modules >>> before you reboot with kexec. Not infrequently the module remove path >>> is better tested than the device shutdown path. >> >> I'm trying this now. In one panic, the pte referenced was >> 0x100010000000000 which sure looks a whole like someone wrote his >> registers in there. It certainly doesn't look like a valid pte. >> >> So far, unloading pata_acpi and pata_amd seem to have eliminated the >> ACPI exception messages. I believe that this resets the device >> properly. Unfortunately, it looks like lots of drivers don't implement >> the pci_driver->shutdown call, so it would make sense that this is a >> relatively widespread problem. > > Most devices don't leave dma setup if you reboot, and even more the > generic pci clears the bus master DMA bit which shuts down a lot more > dma. > > So the actual lack of a shutdown method is not as much of an issue as it > might appear. Interesting. Thanks for the information. I have to admit, this part of the kernel is a bit of a mystery to me. --dlloyd
