Re: Legacy features in PCI Express devices

On 13/03/17 17:39, Mason wrote:

On 13/03/2017 18:12, Robin Murphy wrote:

quoted

On 13/03/17 16:10, Mason wrote:

quoted

There are two revisions of our PCI Express controller.

Rev 1 did not support the following features:

  1) legacy PCI interrupt delivery (INTx signals)
  2) I/O address space

Internally, someone stated that such missing support would prevent
some PCIe cards from working with our controller.

Are there really modern PCIe cards that require 1) and/or 2)
to function?

Can someone provide examples of such cards, so that I may test them
on both revisions?

I was told to check ath9k-based cards. Any other examples?

Looking around, I came across this thread:
http://lists.infradead.org/pipermail/linux-arm-kernel/2016-March/418254.html
"i.MX6 PCIe: Fix imx6_pcie_deassert_core_reset() polarity"

IIUC, although some PCIe boards do support MSI, the driver might not
put in the work to use that infrastructure, and instead reverts to
legacy interrupts. (So it is a SW issue, in a sense.)

Secondary to that category is endpoints which nominally support MSI, but
in a way which is unreliable or otherwise broken. My experience shows
that the Silicon Image SiI 3132 (as integrated on ARM Juno boards, but
seemingly also relatively common on 'generic' 2-port SATA cards) falls
into that category - using the command-line parameter to force MSIs
instead of legacy interrupts leads to the the machine barely reaching
userspace before something goes horribly wrong:

Do drivers typically support *both* MSI and INTx?

I'm not sure about "typically", but it certainly happens. For example,
the Intel e1000e NIC driver is one I know of which can fall back from
MSI-X to MSI to legacy dynamically.

Specifically, would the xhci driver support both?

Line 415 of xhci.c would appear to imply so.

If I remove MSI support from my kernel, I might be able to test
legacy interrupt support that way, right?

Indeed, disabling CONFIG_PCI_MSI should leave drivers with no other choice.

Robin.

Regards.