pci_restore_state
Hello.
Currently pci_restore_state() (drivers/pci/pci.c) restores the PCI state by copying in the saved PCI configuration space, a dword at a time, starting from 0 up to 15.
This causes a crash when my Macbook resumes from sleep (specifically, when restoring the configuration space of the PCI bridge).
Reading the PCI specification, the register at dword 1 (ie: bytes 4-7) is split half and half between status and command words. The command word effectively controls the way which the PCI device interacts with the system. If it is 0, the device is logically disconnected from the bus (PCI Local Bus Specification Revision 2.2, 6.2.2 "Device Control").
When a device first powers up, the command register value is normally zero (and is zero in my specific case).
The problem with the way that the PCI state is currently restored is that the write to the command register logically reconnects the device to the bus before the rest of the configuration space has been filled in.
My Macbook crashes on resume.
If I reverse the for loop to start from 15 and count down to 0, then the majority of the configuration space is filled in _before_ the command word is modified. No crash.
About dword 0 -- this dword contains only the vendor and product ID codes which will always be set by the device (as far as I know) so it's not a problem to not have written these before restoring the value of the command register. I'm not sure they even ever need to be written.
I am essentially requesting that this piece of code be changed in the stock kernel as I changed it (described above). It makes sleep work on my Macbook and may fix some others too (and/or make more reliable some which already work).
Cheers
Quoted 22 lines · On Sun, 04 Jun 2006 06:13:30 -0400 Ryan Lortie <[email protected]> wrote:
Currently
pci_restore_state()(drivers/pci/pci.c) restores the PCI state by copying in the saved PCI configuration space, a dword at a time, starting from 0 up to 15.This causes a crash when my Macbook resumes from sleep (specifically, when restoring the configuration space of the PCI bridge).
Reading the PCI specification, the register at dword 1 (ie: bytes 4-7) is split half and half between status and command words. The command word effectively controls the way which the PCI device interacts with the system. If it is 0, the device is logically disconnected from the bus (PCI Local Bus Specification Revision 2.2, 6.2.2 "Device Control").
When a device first powers up, the command register value is normally zero (and is zero in my specific case).
The problem with the way that the PCI state is currently restored is that the write to the command register logically reconnects the device to the bus before the rest of the configuration space has been filled in.
Quoted 5 lines
My Macbook crashes on resume.
If I reverse the for loop to start from 15 and count down to 0, then the majority of the configuration space is filled in _before_ the command word is modified. No crash.
We have a patch pending which will do that.
The present plan will be to get this into 2.6.18.
Quoted 9 lines · Andrew Morton
Quoted 3 lines · On Sun, 04 Jun 2006 06:13:30 -0400 Ryan Lortie <[email protected]> wrote:
If I reverse the for loop to start from 15 and count down to 0, then the majority of the configuration space is filled in _before_ the command word is modified. No crash.
We have a patch pending which will do that.
We really shouldn't be writing to the BIST register, at least...
Also, I don't quite see the point of writing to the read-only registers such as vendor and device ID.
Paul.
Quoted 10 lines · Paul Mackerras
Quoted 8 lines · Andrew Morton
Quoted 3 lines · On Sun, 04 Jun 2006 06:13:30 -0400 Ryan Lortie <[email protected]> wrote:
If I reverse the for loop to start from 15 and count down to 0, then the majority of the configuration space is filled in _before_ the command word is modified. No crash.
We have a patch pending which will do that.
Quoted 8 lines
We really shouldn't be writing to the BIST register, at least...
Also, I don't quite see the point of writing to the read-only registers such as vendor and device ID.
Paul.
Any comments on this patch as an alternative solution?
http://marc.theaimsgroup.com/?l=linux-kernel&m=114949711413176&w=2
Thanks,
Adam