Skip to contentSkip to conversation

Bug in pci_restore_msi_state

DISCUSSION Eric W. Biederman 1 participant 1 message Mar 22, 2007 1 message not in the archive

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw

I couldn't find the original thread so I am restarting one adding I hope the appropriate people to the cc list.

Andrew thanks for forwarding this one. I finally see enough of this to see what is going on.

The question is about the following sequence from libata.

   pci_enable_msi();
   pci_save_state(pdev);
   pci_disable_device(pdev);
   
   pci_restore_state(pdev);
   pci_enable_device(pdev);

Should this be allowed?

Specifically should pci_disable_device() be allowed when we have device resources requested and potentially in use.

Specifically should pci_enable_device() be expected to cope with the case where we are already using some a pci devices resources?

....

If this sequence is sane pci_enable_device needs to not reset dev->irq In this case I guess it is acpi_pci_irq_enable doing that.

If this sequence is not sane we can just put outlaw calling pci_disable_device when msi is enabled and fix it that way.

The suggestion below of setting msi_enabled to 0. While it will avoid the immediate problem, leaves us still having requested the msi irq (which won't work now) and so we won't be in a state where we can do useful things after resume, plus we would wind up leaking the msi state.

I think fixing pci_enable_device to cope when a device is already in use is likely to be the path of least resistance.

If suspend to disk needs to cope with any pci resource assignment changes then we need to free all irqs and all pci bars before the suspend and reallocate them after the resume, and we should outlaw pci_disable_device with pci device resources in use.

Eric

Quoted 57 lines · Andrew Morton

Begin forwarded message:

Date: Tue, 20 Mar 2007 22:07:59 +0100
From: Thomas Meyer <[email protected]>
To: Linux Kernel Mailing List <[email protected]>
Cc: Pekka Enberg <[email protected]>
Subject: Re: Bug in pci_restore_msi_state

Quoted 13 lines · Pekka Enberg
Quoted 8 lines · Thomas Meyer

Hello everybody.

I get this bug after suspending to disk twice:

http://m3y3r.de/bilder/Bug-pci_restore_msi_state.png

This happens with current git head
cd05a1f818073a623455a58e756c5b419fc98db9.

If you know a kernel that works, please consider doing git bisect:

http://www.kernel.org/pub/software/scm/git/docs/howto/isolate-bugs-with-bisect.txt

Quoted 2 lines

Ok. The error happens in the call of pcibios_enable_device in the function do_pci_enable_device (drivers/pci/pci.c):

Before the call of pcibios_enable_device:

do_pci_enable_device: pci_dev= c1a40000
do_pci_enable_device: irq= 218
do_pci_enable_device: msi_enabled= 1

after the call:
ACPI: PCI Interrupt 0000:00:1f.2[B] -> GSI 19 (level, low) -> IRQ 19

do_pci_enable_device: pci_dev= c1a40000
do_pci_enable_device: irq= 19
do_pci_enable_device: msi_enabled= 1

So the function acpi_pci_irq_enable in drivers/acpi/pci_irq.c needs to be fixed:

a.) acpi_pci_irq_enable should care about msi setups and set the flag msi_enabled to 0 or
b.) acpi_pci_irq_enable should care about msi setups and should touch the dev->irq field.

Please comment on this.

with kind regards
thomas