Re: [PATCH V7 4/9] PCI: Enable 10-Bit Tag support for PCIe Endpoint devices

[Bjorn Helgaas <helgaas@kernel.org>]

On Thu, Aug 05, 2021 at 03:47:31PM +0800, Dongdong Liu wrote:
> Hi Bjorn
> 
> Many thanks for your review.
> On 2021/8/5 7:17, Bjorn Helgaas wrote:
> > On Wed, Aug 04, 2021 at 09:47:03PM +0800, Dongdong Liu wrote:
> > > 10-Bit Tag capability, introduced in PCIe-4.0 increases the total Tag
> > > field size from 8 bits to 10 bits.
> > > 
> > > PCIe spec 5.0 r1.0 section 2.2.6.2 "Considerations for Implementing
> > > 10-Bit Tag Capabilities" Implementation Note.
> > > For platforms where the RC supports 10-Bit Tag Completer capability,
> > > it is highly recommended for platform firmware or operating software
> > > that configures PCIe hierarchies to Set the 10-Bit Tag Requester Enable
> > > bit automatically in Endpoints with 10-Bit Tag Requester capability. This
> > > enables the important class of 10-Bit Tag capable adapters that send
> > > Memory Read Requests only to host memory.
> > 
> > Quoted material should be set off with a blank line before it and
> > indented by two spaces so it's clear exactly what comes from the spec
> > and what you've added.  For example, see
> > https://git.kernel.org/linus/ec411e02b7a2
> Good point, will fix.
> > 
> > We need to say why we assume it's safe to enable 10-bit tags for all
> > devices below a Root Port that supports them.  I think this has to do
> > with switches being required to forward 10-bit tags correctly even if
> > they were designed before 10-bit tags were added to the spec.
> 
> PCIe spec 5.0 r1.0 section 2.2.6.2 "Considerations for Implementing
> 10-Bit Tag Capabilities" Implementation Note:
> 
>   Switches that lack 10-Bit Tag Completer capability are still able to
>   forward NPRs and Completions carrying 10-Bit Tags correctly, since the
>   two new Tag bits are in TLP Header bits that were formerly Reserved,
>   and Switches are required to forward Reserved TLP Header bits without
>   modification. However, if such a Switch detects an error with an NPR
>   carrying a 10-Bit Tag, and that Switch handles the error by acting as
>   the Completer for the NPR, the resulting Completion will have an
>   invalid 10-Bit Tag. Thus, it is strongly recommended that Switches
>   between any components using 10-Bit Tags support 10-Bit Tag Completer
>   capability.  Note that Switches supporting 16.0 GT/s data rates or
>   greater must support 10-Bit Tag Completer capability.
> 
> This patch also consider to enable 10-Bit Tag for EP device need RP
> and Switch device support 10-Bit Tag Completer capability.
> > 
> > And it should call out any cases where it is *not* safe, e.g., if P2P
> > traffic is an issue.
> Yes, indeed.
> > 
> > If there are cases where we don't want to enable 10-bit tags, whether
> > it's to enable P2P traffic or merely to work around device defects,
> > that ability needs to be here from the beginning.  If somebody needs
> > to bisect with 10-bit tags disabled, we don't want a bisection hole
> > between this commit and the commit that adds the control.
> We provide sysfs file to disable 10-bit tag for P2P traffic when needed.
> The details see PATCH 7/8/9.

A mechanism for avoiding problems needs to be present from the very
beginning so there's no bisection hole.  It should not be added by a
future patch.

The sysfs file is a start, but if we run into an issue, it could mean
that we can't boot and run long enough to use sysfs to disable 10-bit
tags.  So I think we might need a kernel parameter that disables it
(and possibly other things like MPS optimization).

> Current we do not know the 10-bit tag defective devices, current may no
> need do as 8-bit tag does in quirk_no_ext_tags().