Any ideas about a crash on reboot with igb and intel_iommu?

Any ideas about a crash on reboot with igb and intel_iommu?

[Roland Dreier]

Hi everyone,

I've been asked to help debug a strange crash, and I'm wondering if
anyone has seen something similar.  The setup is a bit awkward because
this is happening in manufacturing burn-in and we have not reproduced it
in the lab yet, so my ability to do specific experiments is still
limited.

Anyway, we have a fairly standard two-socket Xeon server product that
passes all tests with Nehalem CPUs.  However, when we use Westmere CPUs
(which also requires a new BIOS of course), some fraction of the systems
are crashing during burn-in, which basically runs a cycle where it runs
CPU and memory stress tests and then reboots the system for the next
round of tests.  The crash is happening on reboot, and unfortunately I
only have a bunch of pictures of the traceback output, but we've seen
multiple cases where the system is crashing with a traceback like:

  rb_erase
  __free_iova
  flush_unmaps
  intel_unmap_page
  igb_clean_rx_ring
  igb_down
  igb_close
  __igb_shutdown
  igb_shutdown
  pci_device_shutdown
  device_shutdown
  kernel_restart_prepare
  kernel_restart
  sys_reboot

The newest kernel they've been able to try is 2.6.30.9, but from looking
at the kernel changelogs for igb and intel_iommu at least, I don't see
anything particularly promising that was fixed since then.

One other data point is that enabling the BIOS option "maximize memory
under 4GB" (which apparently just allocates less space for PCI BARs
below 4GB) seems to make this crash go away again.

Anyway, does this tickle anyone's memory?  I'm trying to get a better
handle on things, but if this has been seen before, I'd sure love to
skip some of the pain of debugging this.

Thanks,
  Roland
-- 
Roland Dreier <rolandd@cisco.com> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html

Re: Any ideas about a crash on reboot with igb and intel_iommu?

[Roland Dreier]

 > One other data point is that enabling the BIOS option "maximize memory
 > under 4GB" (which apparently just allocates less space for PCI BARs
 > below 4GB) seems to make this crash go away again.

Never mind about this, apparently the crashes do actually happen even
after changing the BIOS config.
-- 
Roland Dreier <rolandd@cisco.com> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html

Re: Any ideas about a crash on reboot with igb and intel_iommu?

[Roland Dreier]

So actually I found the following change that went into 2.6.31:

    commit 91615f765a2935b6cbae424b9eee1585ed681ae6
    Author: Jesse Brandeburg <jesse.brandeburg@intel.com>
    Date:   Tue Jun 30 12:45:15 2009 +0000

        igb: fix unmap length bug
        
        driver was mixing NET_IP_ALIGN count bytes in map/unmap calls
        unevenly. Only map the bytes that the hardware might dma into

igb in 2.6.30.y is doing pci_map_single(<some length>) and doing
pci_unmap_single(<some other length>).

However I haven't been able to provoke the crash yet, even by bouncing
an igb interface (with VT-d turned on and with CONFIG_DMAR_DEFAULT_ON enabled).
Does this make sense as the sort of thing that might corrupt the iova
rbtree and lead to a crash in the __free_iova / rb_erase code?

Thanks,
  Roland
-- 
Roland Dreier <rolandd@cisco.com> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html