Hi Rafael, Stephen & Bjorn,
Glad to see you all in this thread that talks about:
* adding S2RAM support to a PCIe controller driver
* by taking into account that the PCI clock must be
{enabled before,disabled after} the PCI IP itself
* and that it requires some tweaking in the clock driver to
promote the suspend/resume() callbacks to the NOIRQ phase
(reference there [1]).
Stephen, Rafael answered here to your remark (in thread [1]) about the
NOIRQ promotion (see below).
Bjorn, there is a question for you below about the need for a PCI
controller driver to suspend/resume in the NOIRQ phase.
Rafael, thanks for the explanation of what the PM core sequences really
are, I would need you to confirm my approach that promotes the clock
suspend/resume() callbacks to the NOIRQ phase, or otherwise give me
pointers to an alternate solution (also below).
"Rafael J. Wysocki" <rjw@xxxxxxxxxxxxx> wrote on Tue, 18 Dec 2018
11:54:43 +0100:
> On Monday, December 17, 2018 3:54:26 PM CET Miquel Raynal wrote:
> > Hi Rafael,
> >
> > "Rafael J. Wysocki" <rjw@xxxxxxxxxxxxx> wrote on Thu, 13 Dec 2018
> > 22:50:51 +0100:
> >
> > > On Thursday, December 13, 2018 3:30:00 PM CET Miquel Raynal wrote:
> > > > Hi Lorenzo,
> > > >
> > > > > > If that's really the case, then I can see how one device and it's
> > > > > > children are suspended and the irq for it is disabled but the providing
> > > > > > devices (clk, regulator, bus controller, etc.) are still fully active
> > > > > > and not suspended but in fact completely usable and able to service
> > > > > > interrupts. If that all makes sense, then I would answer the question
> > > > > > with a definitive "yes it's all fine" because the clk consumer could be
> > > > > > in the NOIRQ phase of its suspend but the clk provider wouldn't have
> > > > > > even started suspending yet when clk_disable_unprepare() is called.
> > > > >
> > > > > That's a very good summary and address my concern, I still question this
> > > > > patch correctness (and many others that carry out clk operations in S2R
> > > > > NOIRQ phase), they may work but do not tell me they are rock solid given
> > > > > your accurate summary above.
> > > >
> > > > I understand your concern but I don't see any alternative right now
> > > > and a deep rework of the PM core to respect such dependency is not
> > > > something that can be done in a reasonable amount of time.
> > >
> > > Maybe you don't need to rework anything. :-)
> > >
> > > Have you considered using device links?
> >
> > Absolutely, yes :) I am actively working on it in parallel, you can
> > check the third version there [1]. Stephen Boyd has a slightly
> > different idea of how it should be done, I will propose a v4 this week,
> > I can add you in copy if you are interested!
> >
> > Anyway, there is one thing that is still missing:
> > * Let's have device A that requests clock B
> > * With the device link series, A is linked (as a child) to B.
> > * A suspend/resume hooks handle things in the NOIRQ phase.
>
> Why do you need them to run in the "noirq" phase in the first place?
I suppose (and I would like Bjorn to validate my thoughts) that this
is a limitation imposed by the PCI core, as described in this
commit:
commit ab14d45ea58eae67c739e4ba01871cae7b6c4586
Author: Thomas Petazzoni <thomas.petazzoni@xxxxxxxxxxxxxxxxxx>
Date: Tue Mar 17 15:55:45 2015 +0100
PCI: mvebu: Add suspend/resume support
Add suspend/resume support for the mvebu PCIe host driver. Without this
commit, the system will panic at resume time when PCIe devices are
connected.
Note that we have to use the ->suspend_noirq() and ->resume_noirq() hooks,
because at resume time, the PCI fixups are done at ->resume_noirq() time,
so the PCIe controller has to be ready at this point.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@xxxxxxxxxxxxxxxxxx>
Signed-off-by: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
Acked-by: Jason Cooper <jason@xxxxxxxxxxxxxx>
>
> > * B suspend/resume hooks handle things in the default phase.
> >
> > What I expected during a suspend:
> > 1/ ->suspend_noirq(device A)
> > 2/ ->suspend(clock B)
>
> This expectation is not in agreement with the documented suspend code flow,
> however.
>
> Each phase of it is carried out for *all* devices completely before getting
> to the next phase, "prepare" first, then "suspend", "suspend_late" and
> "suspend_noirq", in this order.
Thanks for clarifying, now it is clear and it also answers Stephen
remark in the related thread [1]:
[PATCH 2/4] clk: mvebu: armada-37xx-periph: change
suspend/resume time
Stephen, said:
"This seems sad that the PM core can't "priority boost" any
suspend/resume callbacks of a device that doesn't have noirq callbacks
when a device that depends on it from the device link perspective does
have noirq callbacks."
>
> > Unfortunately, device links do not seem to enforce any priority between
> > phases (default/late/noirq) and what happens is:
> > 1/ ->suspend(B)
> > 2/ ->suspend_noirq(A)
> > Which has no sense in my case. Hence, I had to request the clock
> > suspend/resume callbacks to be upgraded to the NOIRQ phase as well (I
> > don't have a better solution for now). This is still under discussion
> > in a thread you have been recently added to by Bjorn, see [2].
> >
> > So when I told you I was not confident in "reworking the PM core to
> > respect such dependency", this is what I was referring to. I am
> > definitely ready to help, but I don't feel I can do it alone.
> >
> > [1] https://www.spinics.net/lists/linux-clk/msg32824.html
> > [2] https://marc.info/?l=linux-pm&m=154465198510735&w=2
>
> The rework you seem to be talking about is not possible, I'm afraid.
>
Ok, then do you agree that the only solution in this case is what I
propose in thread [1], ie. promoting the clock suspend/resume callbacks
to the NOIRQ phase in order to ensure that they will run first (once
device links will be merged too) ?
[1] https://www.spinics.net/lists/linux-clk/msg32537.html
Thank you very much for helping,
Miquèl
