[RFC PATCH 2/2] PCI: rockchip-host: drop wait on PERST# toggle

[Geraldo Nascimento]

On Mon, Nov 03, 2025 at 12:10:38PM -0600, Bjorn Helgaas wrote:
> On Mon, Nov 03, 2025 at 03:27:25AM -0300, Geraldo Nascimento wrote:
> > With this change PCIe will complete link-training with a known quirky
> > device - Samsung OEM PM981a SSD. This is completely against the PCIe
> > spec and yet it works as long as the power regulator for 3v3 PCIe
> > power is not defined as always-on or boot-on.
> 
> What is against the spec?  In what way is this SSD "known quirky"?  Is
> there a published erratum for it?

Hi Bjorn!

My proposed change itself is against the spec.

This SSD is known to simply not complete initial link-training with
Rockchip-IP PCIe. This is not officialy documented but based on
reports across boards (like the Armbian one).

I think it's the other way around though, it's the Rockchip-IP PCIe
controller that is quirky somehow and doesn't play nice with
many, many devices.

> 
> Removing this delay might make this SSD work, but if this delay is
> required per PCIe spec, how can we be confident that other devices
> will still work?

We really can't be sure there will be no side-effects to other devices.

> 
> Reports of devices that still work is not really enough to move this
> from the "hack that makes one device work" column to the "safe and
> effective for all devices" column.

I agree, and I knew from the start I would not get encouragements from
the community relative to this change. Still, it seemed selfish not to
share the workaround.

> 
> It's easy to see how *lack* of a delay can break something, but much
> harder to imagine how *removing* a delay can make something work.
> Devices must be able to tolerate pretty much arbitrary delays.

That said, the problem of quirky Rockchip-IP PCIe remains. The fact
that removing the delay makes it work with devices that previously
did not complete initial link training must at least point to a
resolution somehow, even if the present change is unacceptable.

I'll continue to hack around this, see if I can pinpoint exactly why
that delay screws up my initial link training.

Thanks,
Geraldo Nascimento