[PATCH v4 1/5] dt-bindings: PCI: brcmstb: brcm,{enable-l1ss,completion-timeout-us} props

[Bjorn Helgaas]

On Wed, May 03, 2023 at 10:38:57AM -0400, Jim Quinlan wrote:
> On Sun, Apr 30, 2023 at 3:10 PM Bjorn Helgaas <helgaas at kernel.org> wrote:
> > On Fri, Apr 28, 2023 at 06:34:55PM -0400, Jim Quinlan wrote:
> > > brcm,enable-l1ss (bool):
> > >
> > >   The Broadcom STB/CM PCIe HW -- a core that is also used by RPi SOCs --
> > >   requires the driver probe() to deliberately place the HW one of three
> > >   CLKREQ# modes:
> > >
> > >   (a) CLKREQ# driven by the RC unconditionally
> > >   (b) CLKREQ# driven by the EP for ASPM L0s, L1
> > >   (c) Bidirectional CLKREQ#, as used for L1 Substates (L1SS).
> > >
> > >   The HW+driver can tell the difference between downstream devices that
> > >   need (a) and (b), but does not know when to configure (c).  All devices
> > >   should work fine when the driver chooses (a) or (b), but (c) may be
> > >   desired to realize the extra power savings that L1SS offers.  So we
> > >   introduce the boolean "brcm,enable-l1ss" property to inform the driver
> > >   that (c) is desired.  Setting this property only makes sense when the
> > >   downstream device is L1SS-capable and the OS is configured to activate
> > >   this mode (e.g. policy==superpowersave).
> ...

> > What bad things would happen if the driver always configured (c)?
>
> Well, our driver has traditionally only supported (b) and our
> existing boards have been designed with this in mind.  I would not
> want to switch modes w'o the user/customer/engineer opting-in to do
> so.  Further, the PCIe HW engineer told me defaulting to (c) was a
> bad idea and was "asking for trouble".  Note that the commit's
> comment has that warning about L1SS mode not meeting this 400ns
> spec, and I suspect that many of our existing designs have bumped
> into that.
> 
> But to answer your question, I haven't found a scenario that did not
> work by setting mode (c).  That doesn't mean they are not out there.
> 
> > Other platforms don't require this, and having to edit the DT
> > based on what PCIe device is plugged in seems wrong.  If brcmstb
> > does need it, that suggests a hardware defect.  If we need this to
> > work around a defect, that's OK, but we should acknowledge the
> > defect so we can stop using this for future hardware that doesn't
> > need it.
> 
> All devices should work w/o the user having to change the DT.  Only
> if they desire L1SS must they add the "brcm,enable-l1ss" property.

I thought the DT was supposed to describe properties of the
*hardware*, but this seems more like "use this untested clkreq
configuration," which maybe could be done via a module parameter?

Whatever the mechanism, it looks like patch 2/5 makes brcmstb
advertise the appropriate ASPM and L1SS stuff in the PCIe and L1SS
Capabilities so the OS will do the right thing without any core
changes.

> > Maybe the name should be more specific to CLKREQ#, since this
> > doesn't actually *enable* L1SS; apparently it's just one of the
> > pieces needed to enable L1SS?
> 
> The other pieces are:  (a) policy == POWERSUPERSAVE and (b) an
> L1SS-capable device, which seem unrelated and are out of the scope
> of the driver.

Right.  Of course, if ASPM and L1SS support are advertised, the OS can
still choose whether to enable them, and that choice can change at
run-time.

> The RPi Raspian folks have been using "brcm,enable-l1ss"  for a
> while now and I would prefer to keep that name for compatibility.

BTW, the DT comment in the patch refers to PCIe Mini CEM .0 sec
3.2.5.2.5.  I think the correct section is 3.2.5.2.2 (at least in the
r2.1 spec).

There's also a footnote to the effect that T_CRLon is allowed to
exceed 400ns when LTR is supported and enabled.  L1.2 requires LTR, so
if L1.2 is the case where brcmstb exceeds 400ns, that might not be a
problem.

Bjorn