On Tue, 14 May 2019 16:58:59 +0200 Marek Behún marek.behun@nic.cz wrote:
The ofdata_to_platdata method for this driver returns -ENODEV if link is down for a given bus, for example if there is no device in the slot. This causes the uclass_{first,next}_device to return NULL for this bus in pci-uclass.c:pci_init, which of course stops probing of buses which come after.
So if the slot on the first bus is empty on Turris Omnia, and the slot on second bus has a device connected, the device is not probed in U-Boot. On Turris Omnia the PCIe devices have to be probed in U-Boot to work correctly in Linux. Therefore we need this fix.
...
if (!mvebu_pcie_link_up(pcie)) { debug("%s: %s - down\n", __func__, pcie->name);
ret = -ENODEV;
The problem is how uclass_{first,next}_device functions work.
They use helpers uclass_find_{first,next}_device, which iterate all devices. But uclass_{first,next}_device functions also use uclass_get_device_tail on the device returned by helper. This function can return NULL if device failed to probe, which causes the iteration to stop.
Wouldn't it be better if uclass_next_device tried iterating via uclass_find_next_device until a device was found which probed successfully?
I don't know if this would break other things in U-Boot though.
Marek