On Thu, Mar 07, 2024 at 09:31:12PM +0100, Niklas Cassel wrote:

On Mon, Mar 04, 2024 at 02:52:16PM +0530, Manivannan Sadhasivam wrote:

quoted

The DWC glue drivers requiring an active reference clock from the PCIe host
for initializing their PCIe EP core, set a flag called 'core_init_notifier'
to let DWC driver know that these drivers need a special attention during
initialization. In these drivers, access to the hw registers (like DBI)
before receiving the active refclk from host will result in access failure
and also could cause a whole system hang.

But the current DWC EP driver doesn't honor the requirements of the drivers
setting 'core_init_notifier' flag and tries to access the DBI registers
during dw_pcie_ep_init(). This causes the system hang for glue drivers such
as Tegra194 and Qcom EP as they depend on refclk from host and have set the
above mentioned flag.

To workaround this issue, users of the affected platforms have to maintain
the dependency with the PCIe host by booting the PCIe EP after host boot.
But this won't provide a good user experience, since PCIe EP is _one_ of
the features of those platforms and it doesn't make sense to delay the
whole platform booting due to PCIe requiring active refclk.

So to fix this issue, let's move all the DBI access from
dw_pcie_ep_init() in the DWC EP driver to the dw_pcie_ep_init_complete()
API. This API will only be called by the drivers setting
'core_init_notifier' flag once refclk is received from host. For the rest
of the drivers that gets the refclk locally, this API will be called
within dw_pcie_ep_init().

Fixes: e966f7390da9 ("PCI: dwc: Refactor core initialization code for EP mode")
Co-developed-by: Vidya Sagar <redacted>
Signed-off-by: Vidya Sagar <redacted>
Reviewed-by: Frank Li <Frank.Li@nxp.com>
Signed-off-by: Manivannan Sadhasivam <redacted>
---

I'm not sure if the Fixes tag is stictly correct, since there is
nothing wrong with the commit that the Fixes-tag is referencing.

No. The commit was intented to move all the DBI accesses to
dw_pcie_ep_init_complete(), but it left few things like ep_init() callback that
could access the DBI registers. One may argue that the none of the drivers at
that time were accessing DBI registers in that callback etc... but I used that
commit as a fixes tag for the sake of backporting. Otherwise, I don't see how we
can easily backport this patch.

What this patch addresses is an additional use-case/feature,
which allows you to start the EP-side before the RC-side.

However, I'm guessing that you kept the Fixes-tag such that this
patch will get backported. However, this patch is number 4/10 in
the patch series. If this is a strict fix that you want backported,
and it does not depend on any of the previous patches (it doesn't
seem that way), then I think that you should have put it as patch
1/10 in the series.

Not strictly required. Usually the fixes are added first for the ease of merging
as you said, but here I intend to merge this series as it is and it is not
fixing anything in the ongoing release. But, if I happen to respin, I may
reorder so that this can get merged early in next release cycle (this series is
going to miss 6.9 anyway).

Patch ordering aside:
Reviewed-by: Niklas Cassel <cassel@kernel.org>

Thanks!

- Mani

-- 
மணிவண்ணன் சதாசிவம்