[PATCH V7 00/11] Support for generic ACPI based PCI host controller

[Ard Biesheuvel]

On 20 May 2016 at 06:41, Jon Masters <jcm at redhat.com> wrote:
> Hi Tomasz, all,
>
> On 05/11/2016 07:08 AM, Tomasz Nowicki wrote:
>
>> On 11.05.2016 12:41, Gabriele Paoloni wrote:
>
>>>> v6 -> v7
>>>> - drop quirks handling
>>>
>>> Maybe I missed something in the v6 discussion thread; when was it
>>> decided to drop quirk handling?
>>
>> I had such requests in previous series.
>
> A quick note on quirk handling. This, I believe, applies post-merge of
> the base infrastructure, which I realize will likely not have quirks.
>
> We've some "gen1" ARMv8 server platforms where we end up doing quirks
> (for things like forcing 32-bit config space accessors and the like) due
> to people repurposing existing embedded PCIe IP blocks or using them for
> the first time (especially in servers), and those being involved in the
> design not necessarily seeing this problem ahead of time, or not
> realizing that it would be an issue for servers. In the early days of
> ARM server designs 3-4 years ago, many of us had never really played
> with ECAM or realized how modern topologies are built.
>
> Anyway. We missed this one in our SBSA requirements. They say (words to
> the effect of) "thou shalt do PCIe the way it is done on servers" but
> they aren't prescriptive, and they don't tell people how that actually
> is in reality. That is being fixed. A lot of things are happening behind
> the scenes - especially with third party IP block providers (all of whom
> myself and others are speaking with directly about this) - to ensure
> that the next wave of designs won't repeat these mistakes. We don't have
> a time machine, but we can contain this from becoming an ongoing mess
> for upstream, and we will do so. It won't be a zoo.
>
> Various proposals have arisen for how to handle quirks in the longer
> term, including elaborate frameworks and tables to describe them
> generically. I would like to caution against such approaches, especially
> in the case that they deviate from practice on x86, or prior to being
> standardized fully with other Operating System vendors. I don't expect
> there to be too many more than the existing initial set of quirks we
> have seen posted. A number of "future" server SoCs have already been
> fixed prior to silicon, and new design starts are being warned not to
> make this a problem for us to have to clean up later.
>
> So, I would like to suggest that the eventual framework mirror the
> existing approach on x86 systems (matching DMI, etc.) and not be made
> into some kind of generic, utopia. This is a case where we want there to
> be pain involved (and upstream patches required) when people screw up,
> so that they have a level of pain in response to ever making this
> mistake in the future. If we try to create too grand a generic scheme
> and make it too easy to handle this kind of situation beyond the small
> number of existing offenders, we undermine efforts to force vendors to
> ensure that their IP blocks are compliant going forward.
>

I understand that there is a desire from the RedHat side to mimic x86
as closely as possible, but I never saw any technical justification
for that. DMI contains strings that are visible to userland, and you
effectively lock those down to certain values just so that the kernel
can distinguish a broken PCIe root complex from a working one. Linux
on x86 had no choice, since the overwhelming majority of existing
hardware misrepresented itself as generic, and DMI was the only thing
available to actually distinguish these broken implementations from
one another. This does not mean we should allow and/or encourage this
first gen hardware to misrepresent non-compliant hardware as compliant
as well.

Since you are talking to all the people involved, how about you
convince them to put something in the ACPI tables that allows the
kernel to distinguish those non-standard PCIe implementations from
hardware that is really generic? This way, we can sidestep the quirks
debate entirely, since it will simply be a different device as far as
the kernel is concerned. This is no worse than a quirk from a
practical point of view, since an older OS will be equally unable to
run on newer hardware, but it is arguably more true to the standards
compliance you tend to preach about, especially since this small pool
of third party IP could potentially be identified directly rather than
based on some divination of the SoC we may or may not be running on. I
am also convinced that adding support for an additional HID() to the
ACPI ECAM driver with some special config space handling wired in is
an easier sell upstream than making the same ugly mess x86 has had to
make because they did not have any choice to begin with.

If we do need a quirks handling mechanism, I still don't see how the
x86 situation extrapolates to ARM. ACPI offers plenty of ways for a
SoC vendor to identify the make and particular revision, and quirks
could be keyed off of that.

-- 
Ard.