Re: [PATCH v3] PCI: Workaround wrong flags completions for IDT switch

[Bjorn Helgaas]

On Wed, Jun 14, 2017 at 02:15:24PM -0400, James Puthukattukaran wrote:
> On 06/14/2017 07:32 AM, Bjorn Helgaas wrote:
> >On Tue, Jun 13, 2017 at 07:19:57PM -0400, James Puthukattukaran wrote:
> >>>On Jun 13, 2017, at 6:14 PM, Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> >>>>On Tue, Jun 13, 2017 at 02:30:55PM -0400, james puthukattukaran wrote:
> >>>>>On 6/13/2017 1:00 PM, Yinghai Lu wrote:
> >>>>>>On Mon, Jun 12, 2017 at 2:48 PM, Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> >>>>>>>On Fri, Jun 09, 2017 at 04:16:17PM -0700, Yinghai Lu wrote:
> >>>>>>>From: James Puthukattukaran <james.puthukattukaran@xxxxxxxxxx>
> >>>>>>>
> >>>>>>>The IDT switch incorrectly flags an ACS source violation on a read config
> >>>>>>>request to an end point device on the completion (IDT 89H32H8G3-YC,
> >>>>>>>errata #36) even though the PCI Express spec states that completions are
> >>>>>>>never affected by ACS source violation (PCI Spec 3.1, Section 6.12.1.1).
> >>>>>>Can you include a URL where this erratum is published?  If not, can
> >>>>>>you include the actual erratum text here?
> >>>>Here's the errata text
> >>>>------------------------------------
> >>>>Item #36 - Downstream port applies ACS Source Validation to
> >>>>Completions âSection 6.12.1.1" of the PCI Express Base
> >>>>Specification 3.1 states that completions are never affected by
> >>>>ACS Source Validation. However, completions received by a
> >>>>downstream port of the PCIe switch from a device that has not yet
> >>>>captured a PCIe bus number are incorrectly dropped by ACS source
> >>>>validation by the switch downstream port.
> >>>>
> >>>>Workaround: Issue a CfgWr1 to the downstream device before
> >>>>issuing the first CfgRd1 to the device.  This allows the
> >>>>downstream device to capture its bus number; ACS source
> >>>>validation no longer stops completions from being forwarded by
> >>>>the downstream port. It has been observed that Microsoft Windows
> >>>>implements this workaround already; however, some versions of
> >>>>Linux and other operating systems may not.
> >>>This doesn't mention anything about disabling ACS.  Issuing a
> >>>config write to devices downstream of an IDT bridge sounds simpler
> >>>than what this patch does.  Why don't you do that?
> >>The issue is how will we know is the config write succeeds if the
> >>device is not ready? I thought it was simpler to disable acs for the
> >>sake of the read and when we know that the device is ready ( returns
> >>vendor id from read), it's ready for subsequent config write.
> >If that's a problem, it sounds like the errata text is wrong or at
> >least incomplete.  If disabling ACS SV is required, the errata text
> >should mention it.
> >
> >But I don't think it is a problem.  Per PCIe r3.1, sec 2.3.2, if a
> >Root Complex receives a CRS completion for a Configuration Write, it
> >must re-issue the request.
> >
> >Or did you actually try that and find that it didn't work?
> I tried the write and it did not work. This could be because the
> root port gave up after getting a few CRS responses?
> I know that the IDT switch takes some time to be ready even to
> respond with CRS response (firmware initialization, etc). 

The ACS problem is not with completions for config reads to the switch
itself, so it doesn't matter how long it takes the switch to be ready.
The problem is how the switch handles completions for config reads to
devices *downstream* from the switch.

The workaround recommended by the erratum is to perform a config write
to the device *below* the switch before doing a config read.  E.g., in
pci_scan_device(), you would do something like this:

  struct pci_dev *pci_scan_device(...)
  {
    if (bus->self && bus->self->idt_acs_bug)
      pci_bus_write_config_word(bus, devfn, PCI_VENDOR_ID, 0);

    if (!pci_bus_read_dev_vendor_id(bus, devfn, &l, 60*1000))
      return NULL;

    ...

> The spec does not say how often the root complex would retry the
> request and  how long. So, I put in a 1 second delay before the
> write and the write seemed to latch/work.

That's interesting and suggests that you did actually do the config
write to the downstream device, not to the switch.  But you don't say
exactly how the write "did not work".  If the RC stopped retrying the
config write, I would expect it to at least report the error, e.g., by
logging a Completion Timeout error.  Then a subsequent read would
probably cause the ACS SV error because the device didn't accept the
write.

PCIe r3.1, sec 6.6.1, requires some delays between a Conventional
Reset and the first config requests.  I don't recall a place where we
explicitly do that delay, so that could conceivably account for the
device not accepting the write.  But I'd be surprised if we could boot
Linux and get to PCI enumeration that fast (on the order of 0.1 - 1
second).

In any case, I'm looking for a patch that checks for the broken IDT
devices, so we don't have to do all the workaround stuff for
everybody.

Bjorn