On 11/4/2020 2:37 AM, Bjorn Helgaas wrote:
On Tue, Nov 03, 2020 at 08:57:01AM +0530, Vidya Sagar wrote:
On 11/3/2020 4:32 AM, Bjorn Helgaas wrote:
On Thu, Oct 29, 2020 at 11:09:59AM +0530, Vidya Sagar wrote:
DesignWare core has a TLP digest (TD) override bit in one of the control
registers of ATU. This bit also needs to be programmed for proper ECRC
functionality. This is currently identified as an issue with DesignWare
IP version 4.90a. This patch does the required programming in ATU upon
querying the system policy for ECRC.
I guess this is a hardware defect, right?
Yes. This is common across all DWC implementations (version 4.90 precisely)
How much of a problem would it be if we instead added a "no_ecrc"
quirk for this hardware so we never enabled ECRC?
Well, on Tegra for some of the high fidelity use cases, ECRC is required to
be turned on and if it can be done safely with these patches, why shouldn't
we not enable ECRC at all?
IIUC, the current Linux support of ECRC is a single choice at
boot-time: by default ECRC is not enabled, but if you boot with
"pci=ecrc=on", we turn on ECRC for every device.
That seems like the minimal support, but I think the spec allows ECRC
to be enabled selectively, on individual devices. I can imagine a
sysfs knob that would allow us to enable/disable ECRC per-device at
run-time.
If we had such a sysfs knob, it would be pretty ugly and maybe
impractical to work around this hardware issue. So I'm a little bit
hesitant to add functionality that might have to be removed in the
future.
Agree with this. But since it is a boot-time choice at this point, I think
we can still go ahead with this approach to have a working ECRC mechanism
right? I don't see any sysfs knob for AER controlling at this point.
I don't want to do anything that will prevent us from adding
per-device ECRC control in the future.
My concern is that if we add a run-time sysfs knob in the future, the
user experience on this hardware will be poor because there's no
convenient path to twiddle the PCIE_ATU_TD bit when the generic code
changes the AER Control bit.
Agree.
Can we add it to the documentation that run time changing of ECRC settings
are not supported on this (i.e. Tegra194) platform (or for that matter on
any SoC with PCIe based on DesignWare core version 4.90A). By 'not
supported', I meant that the ECRC digest part may not work but the normal
functionality will continue to work without reporting any AER errors. I
tried to emulate the following scenarios on Tegra194 silicon and here are my
observations.
FWIW, I'm referring to the PCIe spec Rev 5.0 Ver 1.0 (22 May 2019)
What is the failure mode in these scenarios:
- User boots with defaults, ECRC is disabled.
- User enables ECRC via sysfs.
- What happens here? ECRC is enabled via AER Control but not via
DWC TD bit. I assume ECRC doesn't work. Does it cause PCIe
errors (malformed TLP, etc)?
Since DWC TD bit is not programmed, for all the transactions that go through
ATU, TLP Digest won't get generated (although AER registers indicate that
ECRC should get generated).
As per the spec section "2.7.1 ECRC Rules"
---
If a device Function is enabled to generate ECRC, it must calculate and
apply ECRC for all TLPs originated by the Function
---
So the RP would be violating the PCIe spec, but it doesn't result in any
error because the same section has the following rule as well
---
If a device Function is enabled to check ECRC, it must do so for all TLPs
with ECRC where the device is the ultimate PCI Express Receiver
Note that it is still possible for the Function to receive TLPs without
ECRC, and these are processed normally - this is not an error
---
so, even if the EP has ECRC check enabled, because of the above rule, it
just processes those packets without ECRC as normal packets.
Basically, whoever is enabling ECRC run time gets cheated as transactions
routed through ATU don't really have ECRC digest
Or this one:
- User boots with "pci=ecrc=on", ECRC is enabled.
- ECRC works fine.
- User disables ECRC via sysfs.
- What happens here? ECRC is disabled via AER Control, but DWC TD
bit thinks it's still enabled.
In this case, the EP doesn't have ECRC check enabled but receives TLPs with
ECRC digest. This again won't result in any error because of the section
"2.2.3 TLP Digest Rules" which says
---
If an intermediate or ultimate PCI Express Receiver of the TLP does not
support ECRC checking, the Receiver must ignore the TLP Digest
---
So the EP just ignores the TLP Digest and process the TLP normally.
Although functionality wise there is no issue here, there could be some
impact on the perf because of the extra DWord data. This is again debatable
as the perf/data path is typically from EP's DMA engine to host system
memory and not config/BAR accesses.
If you enabled ECRC unconditionally on DWC and the sysfs knob had no
effect, I'd be OK with that. I'm more worried about what happens when
the AER bit and the DWC TD bit are set so they don't match. What is
the failure mode there?
Based on the above experiments, we can as well keep the DWC TD bit
programmed unconditionally and it won't lead to any errors. As mentioned
before, the only downside of it is the extra DWord in each ATU routed TLP
which may load the bus (in downstream direction) with no real benefit as
such.
