On 09/06/2012 08:58 PM, Shawn Guo wrote:

On Thu, Sep 06, 2012 at 04:35:25PM -0600, Stephen Warren wrote:

quoted

I believe this patch is causing issues initializing PCI-Express on Tegra.

In next-20120906, I cold-booted 10 times. 3 times, PCIe initialized OK,
and 7 times, the driver timed out in arch/arm/mach-tegra/pcie.c function
tegra_pcie_check_link(). With this patch reverted, another 10 cold boot
attempts all succeeded just fine. Similarly, the regression appeared in
next-20120905, and I isolated it to arch/arm/kernel/, and this is the
only patch in that directory between next-20120904 and next-20120905.

Do you have any idea what the problem might be?

Looking at the timestamps in dmesg in the failing case, the driver is
waiting the expected (per pcie.c driver code) 1.2 seconds before giving
up on the port, although I suppose if the kernel's idea of real-time is
off, then the dmesg log timestamps might be off too.

Just for identifying the problem, can you test the following change to
see if it fixes the failure.

Yes, that does solve the problem (well, with s/late_init/late_initcall/).

As you imply, that change wouldn't help if cpu-tegra.c was built as a
module. So, is this something you can work around in your patch?

Thanks for the quick response.