On Oct 29, 2008, at 10:05 AM, Peter Tyser wrote:
Shouldn't we only be doing this for an agent? Also is the right place to enable it? Just wondering if board code should have more flexibility here.
I was under the impression that if a PCIe interface was configured as root complex its CFG_READY bit would still need to be set in order for an endpoint device to perform config cycles to it (eg to see if the root complex device was sharing its memory, or if the root complex's shared memory was prefetchable, etc). So I thought it made sense to unconditionally set the CFG_READY bit for PCIe interfaces. Let me know if I'm missing something here (I thought some non-fsl chips could generate config cycles in endpoint mode)...
No Freescale part that I'm aware of (ppc based) has ever allowed inbound config cycles if the device is in host mode/root complex.
Alright.
Looking at section 4.4.3.6 of the MPC8548 manual I thought that the ACL bit could be set on a host interface based on the value of the cfg_cpu_boot reset configuration bit. Looking over the manual a bit more closely, it looks like if a PCI interface is configured as host it can't respond to inbound config cycles period (Table 16-65 of the 8548 manual). So for PCI hosts, ACL does not need to be cleared as you suggested.
As far as having more flexibility, I couldn't think of a reason that boards would want to have more fine grained control of this. Is there an example you can think of where a board would want to do it differently?
Its a timing issue if we are talking about using this for "agent" support. The idea is once any cpu local software has run and setup things than it sets the ACL bit to allow an external master access. Its not clear when that point might be as we can't predicate what the local CPU SW wants setup before allowing access.
It seems like 99% of the users could do any custom configuration of inbound windows between the calls to fsl_pci_setup_inbound_windows() and fsl_pci_init() in pci_init_board(). If its a concern I can pull the configuration unlocking into a separate function to to allow for odd situations. That would allow a board to call it at any point in the boot process (eg last_stage_init()).
If you are only running your devices/boards in host mode this patch is not needed.
Most of our Freescale boards require this change as they are PMCs, XMCs, cPCI, etc boards which can be configured as host/agent/root complex/endpoint dynamically depending on a carrier card, chassis, etc. We also have a card with 2 Freescale CPUs connected via PCIe which requires the change.
How about I pull the changes into a separate function and add checks to ensure the the interface is host before allowing inbound config cycles? I'd only add a call to the new function to the two boards I recently added support for (assuming they are accepted:) and leave it to other boards to add a call to it if they need it.
perfect.
- k