Received-SPF: pass (google.com: domain of linux-kernel+bounces-116393-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Precedence: bulk
MIME-Version: 1.0
References: <CAAd53p52fi_wr3Js9Rqct+i1D3rjrnVZ6tBN=uHqThM7UvzXQA@mail.gmail.com>
 <20240322164311.GA1367151@bhelgaas>
In-Reply-To: <20240322164311.GA1367151@bhelgaas>
From: Kai-Heng Feng <kai.heng.feng@canonical.com>
Date: Mon, 25 Mar 2024 10:02:27 +0800
Message-ID: <CAAd53p53NjYsYctrb6oknXVTMEq919nfVvrK17EBXwsGCxvtbA@mail.gmail.com>
Subject: Re: [PATCH v2] mmc: sdhci-pci-gli: GL975x: Mask rootport's replay
 timer timeout during suspend
To: Bjorn Helgaas <helgaas@kernel.org>
Cc: adrian.hunter@intel.com, ulf.hansson@linaro.org, 
	Victor Shih <victor.shih@genesyslogic.com.tw>, Ben Chuang <benchuanggli@gmail.com>, 
	linux-mmc@vger.kernel.org, linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Sat, Mar 23, 2024 at 12:43=E2=80=AFAM Bjorn Helgaas <helgaas@kernel.org>=
 wrote:
>
> On Thu, Mar 21, 2024 at 06:05:33PM +0800, Kai-Heng Feng wrote:
> > On Sat, Jan 20, 2024 at 6:41=E2=80=AFAM Bjorn Helgaas <helgaas@kernel.o=
rg> wrote:
> > > On Thu, Jan 18, 2024 at 02:40:50PM +0800, Kai-Heng Feng wrote:
> > > > On Sat, Jan 13, 2024 at 1:37=E2=80=AFAM Bjorn Helgaas <helgaas@kern=
el.org> wrote:
> > > > > On Fri, Jan 12, 2024 at 01:14:42PM +0800, Kai-Heng Feng wrote:
> > > > > > On Sat, Jan 6, 2024 at 5:19=E2=80=AFAM Bjorn Helgaas <helgaas@k=
ernel.org> wrote:
> > > > > > > On Thu, Dec 21, 2023 at 11:21:47AM +0800, Kai-Heng Feng wrote=
:
> > > > > > > > Spamming `lspci -vv` can still observe the replay timer tim=
eout error
> > > > > > > > even after commit 015c9cbcf0ad ("mmc: sdhci-pci-gli: GL9750=
: Mask the
> > > > > > > > replay timer timeout of AER"), albeit with a lower reproduc=
e rate.
> > > > > > >
> > > > > > > I'm not sure what this is telling me.  By "spamming `lspci -v=
v`, do
> > > > > > > you mean that if you run lspci continually, you still see Rep=
lay Timer
> > > > > > > Timeout logged, e.g.,
> > > > > > >
> > > > > > >   CESta:        ... Timeout+
> > > > > >
> > > > > > Yes it's logged and the AER IRQ is raised.
> > > > >
> > > > > IIUC the AER IRQ is the important thing.
> > > > >
> > > > > Neither 015c9cbcf0ad nor this patch affects logging in
> > > > > PCI_ERR_COR_STATUS, so the lspci output won't change and mentioni=
ng it
> > > > > here doesn't add useful information.
> > > >
> > > > You are right. That's just a way to access config space to reproduc=
e
> > > > the issue.
> > >
> > > Oh, I think I completely misunderstood you!  I thought you were sayin=
g
> > > that suspending the device caused the PCI_ERR_COR_REP_TIMER error, an=
d
> > > you happened to see that it was logged when you ran lspci.
> >
> > Both running lspci and suspending the device can observe the error,
> > because both are accessing the config space.
> >
> > > But I guess you mean that running lspci actually *causes* the error?
> > > I.e., lspci does a config access while we're suspending the device
> > > causes the error, and the config access itself causes the error, whic=
h
> > > causes the ERR_COR message and ultimately the AER interrupt, and that
> > > interrupt prevents the system suspend.
> >
> > My point was that any kind of PCI config access can cause the error.
> > Using lspci is just make the error more easier to reproduce.
> >
> > > If that's the case, I wonder if this is a generic problem that could
> > > happen with *any* device, not just GL975x.
> >
> > For now, it's just GL975x.
> >
> > > What power state do we put the GL975x in during system suspend?
> > > D3hot?  D3cold?  Is there anything that prevents config access while
> > > we suspend it?
> >
> > The target device state is D3hot.
> > However, the issue happens when the devices is in D0, when the PCI
> > core is saving the device's config space.
> >
> > So I think the issue isn't related to the device state.
> >
> > > We do have dev->block_cfg_access, and there's a comment that says
> > > "we're required to prevent config accesses during D-state
> > > transitions," but I don't see it being used during D-state
> > > transitions.
> >
> > Yes, there isn't any D-state change happens here.
>
> So the timeout happens sometimes on any config accesses to the device,
> no matter what the power state is?

Yes.

> If that's the case, why do the
> masking in the suspend/resume callbacks?

Because there's no functional impact when the error happens, other
than suspend/resume.

>
> If it's not related to a power state change, it sounds like something
> that should be a quirk or done at probe time.

Sure, I'll change that to be done at probe time.

Kai-Heng

>
> Bjorn