lkml.org 
[lkml]   [2019]   [Dec]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH] pcie: Add quirk for the Arm Neoverse N1SDP platform
    On Thu, Dec 12, 2019 at 11:05:31AM +0000, Andre Przywara wrote:
    > On 11/12/2019 20:17, Bjorn Helgaas wrote:
    > > On Wed, Dec 11, 2019 at 11:00:49AM +0000, Andre Przywara wrote:
    > >> On Tue, 10 Dec 2019 08:41:15 -0600
    > >> Bjorn Helgaas <helgaas@kernel.org> wrote:
    > >>> On Mon, Dec 09, 2019 at 04:06:38PM +0000, Andre Przywara wrote:
    > >>>> From: Deepak Pandey <Deepak.Pandey@arm.com>
    > >>>>
    > >>>> The Arm N1SDP SoC suffers from some PCIe integration issues, most
    > >>>> prominently config space accesses to not existing BDFs being answered
    > >>>> with a bus abort, resulting in an SError.
    > >>>
    > >>> Can we tease this apart a little more? Linux doesn't program all the
    > >>> bits that control error signaling, so even on hardware that works
    > >>> perfectly, much of this behavior is determined by what firmware did.
    > >>> I wonder if Linux could be more careful about this.
    > >>>
    > >>> "Bus abort" is not a term used in PCIe.
    > >>
    > >> Yes, sorry, that was my sloppy term, also aiming more at the CPU
    > >> side of the bus, between the cores and the RC.
    > >>
    > >>> IIUC, a config read to a
    > >>> device that doesn't exist should terminate with an Unsupported Request
    > >>> completion, e.g., see the implementation note in PCIe r5.0 sec 2.3.1.
    > >>
    > >> Yes, that's what Lorenzo mentioned as well.
    > >>
    > >>> The UR should be an uncorrectable non-fatal error (Table 6-5), and
    > >>> Figures 6-2 and 6-3 show how it should be handled and when it should
    > >>> be signaled as a system error. In case you don't have a copy of the
    > >>> spec, I extracted those two figures and put them at [1].
    > >>
    > >> Thanks for that.
    > >> So in the last few months we tossed several ideas around how to
    > >> work-around this without kernel intervention, all of them turned out
    > >> to be not working. There are indeed registers in the RC that
    > >> influence error reporting to the CPU side, but even if we could
    > >> suppress (or catch) the SError, we can't recover and fixup the read
    > >> transaction to the CPU. Even Lorenzo gave up on this ;-) As far as I
    > >> understood this, there are gates missing which are supposed to
    > >> translate this specific UR into a valid "all-1s" response.
    > >
    > > But the commit log says firmware scanned the bus (catching the
    > > SErrors). Shouldn't Linux be able to catch them the same way?
    >
    > Not really. The scanning is done by the SCP management processor,
    > which is a Cortex-M class core on the same bus. So it's a simple,
    > single core running very early after power-on, when the actual AP
    > cores are still off. The SError handler is set to just increase a
    > value, then to return. This value is then checked before and after a
    > config space access for a given BDF:
    > https://git.linaro.org/landing-teams/working/arm/n1sdp-pcie-quirk.git/tree/scp
    >
    > On the AP cores that run Linux later on this is quite different: The
    > SError is asynchronous, imprecise (inexact) and has no syndrome
    > information. That means we can't attribute this anymore to the
    > faulting instruction, we don't even know if it happened due to this
    > config space access. The CPU might have executed later instructions
    > already, so the state is broken at this point. SError basically
    > means: the system is screwed up. Because this is quite common for
    > SErrors, we don't even allow to register SError handlers in arm64
    > Linux.
    >
    > So even if we could somehow handle this is in Linux, it would be a
    > much greater and intrusive hack, so I'd rather stick with this
    > version.

    The problem is that from a PCIe point of view, UR is something we
    should be able to tolerate. It happens during enumeration and also
    during hotplug. It definitely does not mean "the system is screwed up
    and must be rebooted."

    To go back to Figure 6-3, I'm getting the impression that the "System
    Error" shown at the top is *not* the "SError" you're referring to. If
    they were the same, the Root Control enable bits should gate it, but
    according to your lspci, those enable bits are cleared, yet you still
    take SErrors.

    SError is asynchronous and imprecise. Is there no way to do the
    config access in a way that makes it precise, by adding some kind of
    sync? There's no reason we can't single-thread config accesses and
    maybe even MMIO/IO port accesses as well if necessary.

    Bjorn

    \
     
     \ /
      Last update: 2019-12-13 22:07    [W:3.745 / U:0.004 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site