Re: [v4] PCI: improve host drivers compile test coverage

[Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>]

On Mon, Jun 18, 2018 at 06:53:10PM +0200, Jan Kiszka wrote:
> On 2018-06-18 14:52, Lorenzo Pieralisi wrote:
> > On Mon, Jun 18, 2018 at 01:42:25PM +0200, Jan Kiszka wrote:
> >> On 2018-06-18 11:32, Lorenzo Pieralisi wrote:
> >>> On Fri, Jun 15, 2018 at 12:34:51PM -0600, Rob Herring wrote:
> >>>> On Fri, Jun 15, 2018 at 11:55 AM, Scott Branden
> >>>> <scott.branden@broadcom.com> wrote:
> >>>>> Hi Lorenzo,
> >>>>>
> >>>>>
> >>>>>
> >>>>> On 18-06-15 05:58 AM, Lorenzo Pieralisi wrote:
> >>>>>>
> >>>>>> On Wed, Jun 13, 2018 at 11:11:46AM -0700, Guenter Roeck wrote:
> >>>>>>>
> >>>>>>> On Wed, Jun 13, 2018 at 11:09:35AM +0100, Lorenzo Pieralisi wrote:
> >>>>>>>>
> >>>>>>>> [+Jan, Ley Foon, RMK]
> >>>>>>>>
> >>>>>>>> On Tue, Jun 12, 2018 at 10:02:29AM -0700, Guenter Roeck wrote:
> >>>>>>>>>
> >>>>>>>>> On Thu, Apr 05, 2018 at 02:31:54PM -0500, Rob Herring wrote:
> >>>>>>>>>>
> >>>>>>>>>> Add COMPILE_TEST on driver config options with it. Some ARM drivers
> >>>>>>>>>> still have arch dependencies, so we have to keep those dependent on
> >>>>>>>>>> ARM.
> >>>>>>>>>>
> >>>>>>>>>> Cc: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
> >>>>>>>>>> Cc: Bjorn Helgaas <bhelgaas@google.com>
> >>>>>>>>>> Cc: linux-pci@vger.kernel.org
> >>>>>>>>>> Signed-off-by: Rob Herring <robh@kernel.org>
> >>>>>>>>>
> >>>>>>>>> This patch has the undesirable side effect that it selects PCI_DOMAINS
> >>>>>>>>> for
> >>>>>>>>> sparc32:allmodconfig, which in turn results in
> >>>>>>>>>
> >>>>>>>>> drivers/ata/pata_ali.c: In function 'ali_init_chipset':
> >>>>>>>>> drivers/ata/pata_ali.c:469:38: error:
> >>>>>>>>>         implicit declaration of function 'pci_domain_nr'; did you mean
> >>>>>>>>> 'pci_iomap_wc'?
> >>>>>>>>>
> >>>>>>>>> Unfortunately, 37bd62d224c82 ("PCI: Enable PCI_DOMAINS along with
> >>>>>>>>> generic
> >>>>>>>>> PCI host controller") has pretty much the same result. No idea how to
> >>>>>>>>> fix
> >>>>>>>>> the problem, so I won't even try.
> >>>>>>>>
> >>>>>>>> Sorry about that.
> >>>>>>>>
> >>>>>>>> One option would consist in removing all PCI_DOMAINS selection from
> >>>>>>>> drivers/pci/controller/Kconfig and delegate it to arches even though
> >>>>>>>> this would force PCI_DOMAINS selection on all ARM platforms (it is
> >>>>>>>> already selected for ARCH_MULTIPLATFORM).
> >>>>>>>>
> >>>>>>>> Or we add back arch dependency to the relevant host bridges.
> >>>>>>>>
> >>>>>>>> Everything else I have in mind seems overkill to me given that this
> >>>>>>>> patch was added to improve test coverage (we could add a default
> >>>>>>>> pci_domain_nr() stub - weak or #define - that returns 0 in case arches
> >>>>>>>> do not provide an implementation but do we really want to do that ?).
> >>>>>>>>
> >>>>>>>> Thoughts appreciated.
> >>>>>>>>
> >>>>>>>  From the definition of PCI_DOMAINS, I suspect the original idea was that
> >>>>>>> drivers should depend on it, not select it. Especially auto-selecting
> >>>>>>> it with PCI_HOST_GENERIC seems like a bad idea to me. However, that is
> >>>>>>> just me. I'll leave it up to Bjorn to decide what if anything he wants
> >>>>>>> to do about it.
> >>>>>>
> >>>>>> Here is a patch that should reinstate the previous behaviour but
> >>>>>> it will make PCI_DOMAINS a visible option on ARM 32-bit systems; whether
> >>>>>> that's acceptable that's the question I need to answer, it should
> >>>>>> honour old configs and it does not force PCI_DOMAINS selection on
> >>>>>> non-DT arch/arm PCI host controllers (that do not need PCI_DOMAINS
> >>>>>> anyway so I suspect that enabling it on all ARM 32-bit platforms
> >>>>>> should not break anything but I preferred to be cautious).
> >>>>>
> >>>>> I think this change will also require a patch enabling CONFIG_PCI_DOMAINS in
> >>>>> multi_v7_defconfig and iproc_defconfig at the very least?
> >>>>
> >>>> Perhaps the sub-arches that want this should select it. It is more a
> >>>> platform option/decision more than a controller option.
> >>>
> >>> Yes, that makes sense, I assume ARCH_VIRT is a sensible choice for
> >>> virtual machines configuration, Jan ?
> >>>
> >>> I will add a PCI_DOMAINS selection to all (ARM) arches that select
> >>> PCI_DOMAINS in drivers/pci/controller/Kconfig.
> >>>
> >>> For Jailhouse configurations I need Jan's input, I assume adding
> >>> the selection to ARCH_VIRT is the correct way forward, please let
> >>> me know asap.
> >>
> >> So far, there is no need on ARM or ARM64 declare a special platform in
> >> order to run as Jailhouse (secondary) guest.
> >>
> >> My original patch was just about making PCI_DOMAINS manually
> >> configurable, which would have been fine for our use case.
> > 
> > I need more details on your system configuration to make sure we can fix
> > this in a way that does not upset anybody; I am not a big fan of making
> > PCI_DOMAINS visible since it affects other platforms and it is different
> > from how it is managed for all other arches, so please provide details.
> 
> Our setup is as follows: A platform, e.g. Jetson TK1 or TX1/2, already
> has one PCI host controller. When enabling Jailhouse on it (from within
> a running Linux), this adds the generic PCI host controller as virtual
> one. So we need to configure the system to support both controllers and
> PCI domains. But, otherwise, the system does not look different from a
> physical one.

Well, it looks like the only sensible way forward is to amend the
patch below and make PCI_DOMAINS a visible option on ARM and still
select it from the sub-arch entries that need it:

https://marc.info/?l=linux-pci&m=152932092612352&w=2

If there are other ideas I am happy to hear them.

Lorenzo