RE: [PATCH] PCI: dwc: Added link up check in map_bus of dw_child_pcie_ops

[Z.q. Hou]

Hi Rob,

Thanks a lot for your comments!

> -----Original Message-----
> From: Rob Herring <robh@xxxxxxxxxx>
> Sent: 2020年9月18日 23:28
> To: Z.q. Hou <zhiqiang.hou@xxxxxxx>
> Cc: linux-kernel@xxxxxxxxxxxxxxx; PCI <linux-pci@xxxxxxxxxxxxxxx>; Lorenzo
> Pieralisi <lorenzo.pieralisi@xxxxxxx>; Bjorn Helgaas
> <bhelgaas@xxxxxxxxxx>; Gustavo Pimentel
> <gustavo.pimentel@xxxxxxxxxxxx>; Michael Walle <michael@xxxxxxxx>;
> Ard Biesheuvel <ardb@xxxxxxxxxx>
> Subject: Re: [PATCH] PCI: dwc: Added link up check in map_bus of
> dw_child_pcie_ops
> 
> On Fri, Sep 18, 2020 at 5:02 AM Z.q. Hou <zhiqiang.hou@xxxxxxx> wrote:
> >
> > Hi Rob,
> >
> > Thanks a lot for your comments!
> >
> > > -----Original Message-----
> > > From: Rob Herring <robh@xxxxxxxxxx>
> > > Sent: 2020年9月17日 4:29
> > > To: Z.q. Hou <zhiqiang.hou@xxxxxxx>
> > > Cc: linux-kernel@xxxxxxxxxxxxxxx; PCI <linux-pci@xxxxxxxxxxxxxxx>;
> > > Lorenzo Pieralisi <lorenzo.pieralisi@xxxxxxx>; Bjorn Helgaas
> > > <bhelgaas@xxxxxxxxxx>; Gustavo Pimentel
> > > <gustavo.pimentel@xxxxxxxxxxxx>; Michael Walle
> <michael@xxxxxxxx>;
> > > Ard Biesheuvel <ardb@xxxxxxxxxx>
> > > Subject: Re: [PATCH] PCI: dwc: Added link up check in map_bus of
> > > dw_child_pcie_ops
> > >
> > > On Tue, Sep 15, 2020 at 11:49 PM Zhiqiang Hou
> <Zhiqiang.Hou@xxxxxxx>
> > > wrote:
> > > >
> > > > From: Hou Zhiqiang <Zhiqiang.Hou@xxxxxxx>
> > > >
> > > > On NXP Layerscape platforms, it results in SError in the
> > > > enumeration of the PCIe controller, which is not connecting with
> > > > an Endpoint device. And it doesn't make sense to enumerate the
> > > > Endpoints when the PCIe link is down. So this patch added the link
> > > > up check to avoid to fire configuration transactions on link down bus.
> > >
> > > Michael reported the same issue as well.
> > >
> > > What happens if the link goes down between the check and the access?
> >
> > This patch cannot cover this case, and will get the SError.
> > But I think it makes sense to avoid firing transactions on link down bus.
> 
> That's impossible to do without a race even in h/w.

Agree.

> 
> > > It's a racy check. I'd like to find an alternative solution. It's
> > > even worse if Layerscape is used in ECAM mode. I looked at the EDK2
> > > setup for layerscape[1] and it looks like root ports are just skipped if link
> is down.
> > > Maybe a link down just never happens once up, but if so, then we
> > > only need to check it once and fail probe.
> >
> > Many customers connect the FPGA Endpoint, which may establish PCIe
> > link after the PCIe enumeration and then rescan the PCIe bus, so I
> > think it should not exit the probe of root port even if there is not link up
> during enumeration.
> 
> That's a good reason. I want to unify the behavior here as it varies per
> platform currently and wasn't sure which way to go.
> 
> 
> > > I've dug into this a bit more and am curious about the PCIE_ABSERR
> > > register setting which is set to:
> > >
> > > #define PCIE_ABSERR_SETTING 0x9401 /* Forward error of non-posted
> > > request */
> > >
> > > It seems to me this is not what we want at least for config
> > > accesses, but commit 84d897d6993 where this was added seems to say
> > > otherwise. Is it not possible to configure the response per access type?
> >
> > Thanks a lot for your investigation!
> > The story is like this: Some customers worry about these silent error
> > (DWC PCIe IP won't forward the error of outbound non-post request by
> > default), so we were pushed to enable the error forwarding to AXI in
> > the commit
> > 84d897d6993 as you saw. But it cannot differentiate the config
> > transactions from the MEM_rd, except the Vendor ID access, which is
> > controlled by a separate bit and it was set to not forward error of access
> of Vendor ID.
> > So we think it's okay to enable the error forwarding, the SError
> > should not occur, because after the enumeration it won't access the
> non-existent functions.
> 
> We've rejected upstream support for platforms aborting on config
> accesses[1]. I think there's clear consensus that aborting is the wrong
> behavior.
> 
> Do MEM_wr errors get forwarded? Seems like that would be enough. Also,
> wouldn't page faults catch most OOB accesses anyways? You need things
> page aligned anyways with an IOMMU and doing userspace access or guest
> assignment.

Yes, errors of MEM_wr can be forwarded.

> 
> Here's another idea, how about only enabling forwarding errors if the link is
> up? If really would need to be configured any time the link state changes
> rather than just at probe. I'm not sure if you have a way to disable it on link
> down though.

Dug deeper into this issue and found the setting of not forwarding error of non-existent Vender ID access counts on the link partner:
1. When there is a link partner (namely link up), it will return 0xffff when read non-existent function Vendor ID and won't forward error to AXI.
2. When no link partner (link down), it will forward the error of reading non-existent function Vendor ID to AXI and result in SError.

I think this is a DWC PCIe IP specific issue but not get feedback from design team.
I'm thinking to disable this error forwarding just like other platforms, since when these errors (UR, CA and CT) are detected, AER driver can also report the error and try to recover.

Thanks,
Zhiqiang

> 
> Rob