On 07/23/2014 01:42 PM, Jason Gunthorpe wrote:
On Tue, Jul 22, 2014 at 05:52:00PM -0600, Bjorn Helgaas wrote:
If there is a hardware defect, a PCI quirk is a reasonable way to work
around it, since that's the main purpose of quirks. fixup_mpss_256()
is an example of something that sounds superficially similar.
It was my suggestion to engage the PCI-E tuning code. By my
understanding the HW bug is that read response segmentation at the
host bridge does not work - so all read requests from any downstream
device must have responses that fit within a single packet.
In the case of Keystone PCI, when I set the MRSS to 256 in the EP, PCI
controller is able to function properly. Keystone spec says.
• Maximum outbound payload size of 128 bytes
• Maximum inbound payload size of 256 bytes
• Maximum remote read request size of 256 bytes
I am interpreting the "Maximum remote read request size" to indicate it
can's handle if it exceeds the limit. It has an outbound payload size of
128 bytes. So in this case a read request would results in 2 completion
packets. So it seems to be able to segment up to maximum 256 bytes of
read request. Where do I find the requirement in PCI spec that "read
response segmentation at the host bridge does not work" ?
This is completely against how the spec envisions things working,
segmentation is a mandatory function. As you point out there is no
parameter bounding the maximum read request size that a completer will
accept.
So, the only fix is that every downstream device must always have a
MRSS set to less than the MPS of the host bridge.
Why this can't be the default behavior in the PCI core? Any cons?
Murali
Which means the tuning code must be involved somehow, as that code
controls the MRSS of unrelated devices...
Regards,
Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
