Dear Adam,
On the mainline u-boot we found a commit which helps to fix a problem regarding PCIe and the Armada 370: http://git.denx.de/?p=u-boot.git;a=commit;h=6bbe0924a799d33c1a8c9de38b60a5e0...
We currently facing a similar problem with the Armada 385 and PCIe. The problem also appears with two different Atheros chipsets one (AR9280) gets never detected and the other one gets detected most of the time but not always (QCA9882). With Intel modules on the other hand, we can't reproduce the issue. It's the same behaviour as described in the commit above, the link training fails and therefore the link is signalized as down. If a link is successfully established (QCA9882) we don't see any errors during communication.
Do you know if we have the same possibility on the Armada 385 to tune the PCI Express pulse width as you proposed it for the Armada 370? Unfortunately the PCIe Phy seems to be quite different to the one from the Armada 370.
Best regards Stefan