Message-ID: <20161017183209.GA18465@jcartwri.amer.corp.natinst.com>
Date:   Mon, 17 Oct 2016 13:32:09 -0500
From:   Julia Cartwright <julia@...com>
To:     Alexander Duyck <alexander.duyck@...il.com>
CC:     "Koehrer Mathias (ETAS/ESW5)" <mathias.koehrer@...s.com>,
        Bjorn Helgaas <helgaas@...nel.org>,
        "linux-rt-users@...r.kernel.org" <linux-rt-users@...r.kernel.org>,
        Sebastian Andrzej Siewior <sebastian.siewior@...utronix.de>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        "intel-wired-lan@...ts.osuosl.org" <intel-wired-lan@...ts.osuosl.org>,
        Matthew Garrett <mjg59@...eos.com>,
        Bjorn Helgaas <bhelgaas@...gle.com>,
        Greg <gvrose8192@...il.com>, <linux-pci@...r.kernel.org>
Subject: Re: Kernel 4.6.7-rt13: Intel Ethernet driver igb causes huge
 latencies in cyclictest

+linux-pci

On Mon, Oct 17, 2016 at 08:39:40AM -0700, Alexander Duyck wrote:
> On Mon, Oct 17, 2016 at 8:00 AM, Koehrer Mathias (ETAS/ESW5)
> <mathias.koehrer@...s.com> wrote:
> > Hi Julia!
> >> > > Have you tested on a vanilla (non-RT) kernel?  I doubt there is
> >> > > anything RT specific about what you are seeing, but it might be nice
> >> > > to get confirmation.  Also, bisection would probably be easier if you confirm on a
> >> vanilla kernel.
> >> > >
> >> > > I find it unlikely that it's a kernel config option that changed
> >> > > which regressed you, but instead was a code change to a driver.
> >> > > Which driver is now the question, and the surface area is still big
> >> > > (processor mapping attributes for this region, PCI root complex configuration,
> >> PCI brige configuration, igb driver itself, etc.).
> >> > >
> >> > > Big enough that I'd recommend a bisection.  It looks like a
> >> > > bisection between 3.18 and 4.8 would take you about 18 tries to narrow down,
> >> assuming all goes well.
> >> > >
> >> >
> >> > I have now repeated my tests using the vanilla kernel.
> >> > There I got the very same issue.
> >> > Using kernel 4.0 is fine, however starting with kernel 4.1, the issue appears.
> >>
> >> Great, thanks for confirming!  That helps narrow things down quite a bit.
> >>
> >> > Here is my exact (reproducible) test description:
> >> > I applied the following patch to the kernel to get the igb trace.
> >> > This patch instruments the igb_rd32() function to measure the call to
> >> > readl() which is used to access registers of the igb NIC.
> >>
> >> I took your test setup and ran it between 4.0 and 4.1 on the hardware on my desk,
> >> which is an Atom-based board with dual I210s, however I didn't see much
> >> difference.
> >>
> >> However, it's a fairly simple board, with a much simpler PCI topology than your
> >> workstation.  I'll see if I can find some other hardware to test on.
> >>
> >> [..]
> >> > This means, that I think that some other stuff in kernel 4.1 has
> >> > changed, which has impact on the igb accesses.
> >> >
> >> > Any idea what component could cause this kind of issue?
> >>
> >> Can you continue your bisection using 'git bisect'?  You've already narrowed it down
> >> between 4.0 and 4.1, so you're well on your way.
> >>
> >
> > OK - done.
> > And finally I was successful!
> > The following git commit is the one that is causing the trouble!
> > (The full commit is in the attachment).
> > +++++++++++++++++++++ BEGIN +++++++++++++++++++++++++++
> > commit 387d37577fdd05e9472c20885464c2a53b3c945f
> > Author: Matthew Garrett <mjg59@...eos.com>
> > Date:   Tue Apr 7 11:07:00 2015 -0700
> >
> >     PCI: Don't clear ASPM bits when the FADT declares it's unsupported
> >
> >     Communications with a hardware vendor confirm that the expected behaviour
> >     on systems that set the FADT ASPM disable bit but which still grant full
> >     PCIe control is for the OS to leave any BIOS configuration intact and
> >     refuse to touch the ASPM bits.  This mimics the behaviour of Windows.
> >
> >     Signed-off-by: Matthew Garrett <mjg59@...eos.com>
> >     Signed-off-by: Bjorn Helgaas <bhelgaas@...gle.com>
> > +++++++++++++++++++++ HEADER +++++++++++++++++++++++++++
> >
> > The only files that are modified by this commit are
> > drivers/acpi/pci_root.c
> > drivers/pci/pcie/aspm.c
> > include/linux/pci-aspm.h
> >
> > This is all generic PCIe stuff - however I do not really understand what
> > the changes of the commit are...
> >
> > In my setup I am using a dual port igb Ethernet adapter.
> > This has an onboard PCIe switch and it might be that the configuration of this
> > PCIe switch on the Intel board is causing the trouble.
> >
> > Please see also the output of "lspci -v" in the attachment.
> > The relevant PCI address of the NIC is 04:00.0 / 04:00.1
> >
> > Any feedback on this is welcome!
> >
> > Thanks
> >
> > Mathias
>
> Hi Mathias,
>
> If you could set the output of lspci -vvv it might be more useful as
> most of the configuration data isn't present in the lspci dump you had
> attached.  Specifically if you could do this for the working case and
> the non-working case we could verify if this issue is actually due to
> the ASPM configuration on the device.
>
> Also one thing you might try is booting your kernel with the kernel
> parameter "pcie_aspm=off".  It sounds like the extra latency is likely
> due to your platform enabling ASPM on the device and this in turn will
> add latency if the PCIe link is disabled when you attempt to perform a
> read as it takes some time to bring the PCIe link up when in L1 state.

So if we assume it's this commit causing the regression, then it's safe
to assume that this system's BIOS is claiming to not support ASPM in the
FADT, but the BIOS is leaving ASPM configured in some way on the
relevant devices.

Also, unfortunately, taking a look at the code which handles
"pcie_aspm=off", it won't be sufficient to disable ASPM on these
this system, as disabling these states is skipped when the FADT doesn't
advertise ASPM support.

What would be needed is an option like "force", but which force
_disables_ ASPM.  "force-disable", maybe.

   Julia

Powered by blists - more mailing lists