On 30/03/2017 22:56, Mason wrote:

I've run into an issue.

If I boot the system with earlyprintk enabled (as I've
been doing throughout my dev), things work as expected.

But if I boot with earlyprintk disabled, then the system
does not "see" the PCIe board, because reading the vendor
ID returns 0xffffffff.

What we think is happening, is that when earlyprintk is
disabled, the system proceeds much faster through the
various inits, and the PCIe init happens when PCIe
link training has not completed yet.

If that is the case, then it seems I would need to check
the link state in my probe function.

I determined empirically that link training takes around
10-15 ms. Though I suppose this might depend on the
specific PCIe board? (I'm only considering x1 link.)

So I added an msleep(20); in the probe function, and in
the config_read callback, I check the link status on the
first read to the device.

Should I msleep(40) to be safe?

Regards.