Re: [PATCH] USB:Fix ehci infinite suspend-resume loop issue in zhaoxin

From: Alan Stern
Date: Tue Apr 05 2022 - 19:17:48 EST

On Wed, Mar 16, 2022 at 10:18:39AM +0800, WeitaoWang-oc@xxxxxxxxxxx wrote:
> On 2022/3/15 11:18, Alan Stern wrote:
> > On Tue, Mar 15, 2022 at 08:39:09PM +0800, WeitaoWang-oc@xxxxxxxxxxx wrote:
> > > I have encountered the following situation if EHCI runtime suspend is
> > > enabled by default.

Some things about this still confuse me...

> > > 1.Wake from system to disk and boot OS.
> >
> > You're talking about resuming after hibernation, right?
> You're right.
> > > 2.EHCI will entry runtime suspend after enumerated by driver during boot
> > > phase of suspend to disk
> >
> > I'm not sure what you mean by "boot phase of suspend to disk". This is
> > while the restore kernel is starting up at the beginning of resume from
> > hibernation, right?
> >
> You understood exactly what I was saying.

Okay, so we're waking up from hibernation.

> > > 3.EHCI will be placed to freeze state and ehci_resume is called after image
> > > is loaded.
> >
> > ehci_resume is called to leave runtime suspend. Going into the freeze
> > state doesn't require any changes.

In fact, the resume kernel doesn't call ehci_resume at all. Here's what
it does:

The resume kernel boots;

If your patch causes STS_PCD to be set at this point, the flag
should get cleared shortly afterward by ehci_irq;

ehci-hcd goes into runtime suspend;

The kernel reads the system image that was stored earlier when
hibernation began;

After the image is loaded, the system goes into the freeze
state (this does not call any routines in ehci-hcd);

The resume kernel transfers control to the image kernel;

Now the image kernel is running. It goes into the restore state, which
involves calling ehci_resume. Presumably your patch cases the STS_PCD
flag to get set at this point.

But that's all! The system is now back up, out of hibernation, and
running normally. There are no more calls to check_root_hub_suspended

> > > 4.If PCD flag is set(caused by patch), then HCD_FLAG_RH_RUNNING will be set.
> > >
> > > 5.Pci_pm_freeze_noirq is called to check ehci root hub state and return
> > > value is -EBUSY. which will cause
> > > quiesce phase of suspend to disk fail.
> >
> > You're talking about check_root_hub_suspended() in hcd-pci.c, right?
> >
> It's right.

So how can this happen, given that the image kernel doesn't call

Alan Stern