Re: [PATCH v7 0/5] Add support for debugfs based RAS DES feature in PCIe DW

[Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>]

On Tue, Feb 25, 2025 at 03:35:25PM +0100, Niklas Cassel wrote:
> On Tue, Feb 25, 2025 at 01:58:35PM +0530, Manivannan Sadhasivam wrote:
> > On Mon, Feb 24, 2025 at 06:08:26PM +0100, Niklas Cassel wrote:
> > > Hello Shradha,
> > > 
> > > On Fri, Feb 21, 2025 at 06:45:43PM +0530, Shradha Todi wrote:
> > > > DesignWare controller provides a vendor specific extended capability
> > > > called RASDES as an IP feature. This extended capability  provides
> > > > hardware information like:
> > > >  - Debug registers to know the state of the link or controller. 
> > > >  - Error injection mechanisms to inject various PCIe errors including
> > > >    sequence number, CRC
> > > >  - Statistical counters to know how many times a particular event
> > > >    occurred
> > > > 
> > > > However, in Linux we do not have any generic or custom support to be
> > > > able to use this feature in an efficient manner. This is the reason we
> > > > are proposing this framework. Debug and bring up time of high-speed IPs
> > > > are highly dependent on costlier hardware analyzers and this solution
> > > > will in some ways help to reduce the HW analyzer usage.
> > > > 
> > > > The debugfs entries can be used to get information about underlying
> > > > hardware and can be shared with user space. Separate debugfs entries has
> > > > been created to cater to all the DES hooks provided by the controller.
> > > > The debugfs entries interacts with the RASDES registers in the required
> > > > sequence and provides the meaningful data to the user. This eases the
> > > > effort to understand and use the register information for debugging.
> > > > 
> > > > This series creates a generic debugfs framework for DesignWare PCIe
> > > > controllers where other debug features apart from RASDES can also be
> > > > added as and when required.
> > > > 
> > > > v7:
> > > >     - Moved the patches to make finding VSEC IDs common from Mani's patchset [1]
> > > >       into this series to remove dependancy as discussed
> > > >     - Addressed style related change requests from v6
> > > 
> > > I tested this series, and one thing that I noticed:
> > > 
> > > # for f in /sys/kernel/debug/dwc_pcie_a40000000.pcie/rasdes_event_counter/*/counter_enable; do echo 1 > $f; done
> > > 
> > > # grep "" /sys/kernel/debug/dwc_pcie_a40000000.pcie/rasdes_event_counter/*/* | grep Disabled
> > > /sys/kernel/debug/dwc_pcie_a40000000.pcie/rasdes_event_counter/ctl_skp_os_parity_err/counter_enable:Counter Disabled
> > > /sys/kernel/debug/dwc_pcie_a40000000.pcie/rasdes_event_counter/deskew_uncompleted_err/counter_enable:Counter Disabled
> > > /sys/kernel/debug/dwc_pcie_a40000000.pcie/rasdes_event_counter/framing_err_in_l0/counter_enable:Counter Disabled
> > > /sys/kernel/debug/dwc_pcie_a40000000.pcie/rasdes_event_counter/margin_crc_parity_err/counter_enable:Counter Disabled
> > > /sys/kernel/debug/dwc_pcie_a40000000.pcie/rasdes_event_counter/retimer_parity_err_1st/counter_enable:Counter Disabled
> > > /sys/kernel/debug/dwc_pcie_a40000000.pcie/rasdes_event_counter/retimer_parity_err_2nd/counter_enable:Counter Disabled
> > > 
> > > that there are some events that cannot be enabled when testing on my platform,
> > > rk3588, perhaps this is because my version of the DWC IP does not have these
> > > events.
> > > 
> > > (Because all the other events can be enabled successfully:
> > > # grep "" /sys/kernel/debug/dwc_pcie_a40000000.pcie/rasdes_event_counter/*/* | grep Enabled | wc -l
> > > 29
> > > )
> > > 
> > > 
> > > So the question is, how do we want to handle that?
> > >
> > 
> > This is a really good question.
> >  
> > > E.g. counter_enable_write() could theoretically read back the
> > > dw_pcie_readl_dbi(pci, rinfo->ras_cap_offset + RAS_DES_EVENT_COUNTER_CTRL_REG);
> > > register after doing the
> > > ww_pcie_writel_dbi(pci, rinfo->ras_cap_offset + RAS_DES_EVENT_COUNTER_CTRL_REG, val);
> > > 
> > > to actually check if it could enable the event.
> > > 
> > > If counter_enable_write() could not enable the specific event, should it
> > > perhaps return a failure to user space?
> > > 
> > 
> > Yes, it would be appropriate to return -EOPNOTSUPP in that case. But I'd like to
> > merge this series asap. So this patch can come on top of this series.
> 
> I agree that returning an error is probably the nicest thing.
> 
> However, this series has been picked up already :)
> 
> Is there anyone who volunteers on implementing the proposed feature?
> 

I did and submitted the fix [1].

> If you have time, it would be interesting to see if you see the same behavior
> on QCOM SoCs. (Assuming that your DWC PCIe controller does not implement all
> events that Samsung DWC PCIe controller does.)
> 

Yeah, I observed the same behavior on the Qcom platform, though only 2 counters
were not supported. But I also noticed a null ptr dereference due to refclk
dependency, so submitted a fix for that also.

- Mani

[1] https://lore.kernel.org/linux-pci/20250225171239.19574-1-manivannan.sadhasivam@linaro.org/

-- 
மணிவண்ணன் சதாசிவம்