Message-ID: <Y2UkAKfHefdYW59A@orome>
Date:   Fri, 4 Nov 2022 15:38:56 +0100
From:   Thierry Reding <thierry.reding@...il.com>
To:     Robin Murphy <robin.murphy@....com>
Cc:     Ulf Hansson <ulf.hansson@...aro.org>,
        Prathamesh Shete <pshete@...dia.com>, joro@...tes.org,
        adrian.hunter@...el.com, jonathanh@...dia.com,
        p.zabel@...gutronix.de, linux-mmc@...r.kernel.org,
        linux-tegra@...r.kernel.org, linux-kernel@...r.kernel.org,
        will@...nel.org, iommu@...ts.linux.dev, anrao@...dia.com,
        smangipudi@...dia.com, kyarlagadda@...dia.com,
        Thierry Reding <treding@...dia.com>
Subject: Re: [PATCH v10 1/4] iommu: Always define struct iommu_fwspec

On Thu, Nov 03, 2022 at 05:35:19PM +0000, Robin Murphy wrote:
> On 2022-11-03 14:55, Ulf Hansson wrote:
> > On Thu, 3 Nov 2022 at 15:01, Thierry Reding <thierry.reding@...il.com> wrote:
> > > 
> > > On Thu, Nov 03, 2022 at 12:23:20PM +0000, Robin Murphy wrote:
> > > > On 2022-11-03 04:38, Prathamesh Shete wrote:
> > > > > In order to fully make use of the !IOMMU_API stub functions, make the
> > > > > struct iommu_fwspec always available so that users of the stubs can keep
> > > > > using the structure's internals without causing compile failures.
> > > > 
> > > > I'm really in two minds about this... fwspecs are an internal detail of the
> > > > IOMMU API that are meant to be private between individual drivers and
> > > > firmware code, so anything poking at them arguably does and should depend on
> > > > CONFIG_IOMMU_API. It looks like the stub for dev_iommu_fwspec_get() was only
> > > > added for the sake of one driver that was misusing it where it really wanted
> > > > device_iommu_mapped(), and has since been fixed, so if anything my
> > > > preference would be to remove that stub :/
> > > 
> > > Tegra has been using this type of weak dependency on IOMMU_API mainly in
> > > order to allow building without the IOMMU support on some old platforms
> > > where people may actually care about the kernel size (Tegra20 systems
> > > were sometimes severely constrained and don't have anything that we'd
> > > call an IOMMU today).
> > > 
> > > We have similar stubs in place for most other major subsystems in order
> > > to allow code to simply compile out if the subsystem is disabled, which
> > > is quite convenient for sharing code between platforms that may want a
> > > given feature and other platforms that may not want it, without causing
> > > too much of a hassle with compile-testing.
> > 
> > I agree with the above.
> > 
> > Moreover, the stubs make the code more portable/scalable and so it
> > becomes easier to maintain.
> 
> Are you suggesting that having the same thing open-coded slightly
> differently (with bugs) in 8 different places is somehow more maintainable
> than abstracting it into a single centralised implementation?
> 
> Is it "easier to maintain" when already seemingly every thing I try to clean
> up or refactor in the IOMMU API at the moment is stymied by finding Tegra
> drivers doing unexpected (and often questionable) things? Is it "more
> scalable" to make it even easier for people to copy questionable code
> without a second thought, leaving API maintainers to play an ever-expanding
> game of whack-a-mole to clean it up? No. No it chuffing well isn't :(
> 
> > > > I don't technically have much objection to this patch in isolation, but what
> > > > I don't like is the direction of travel it implies. I see the anti-pattern
> > > > is only spread across Tegra drivers, making Tegra-specific assumptions, so
> > > > in my view the best answer would be to abstract that fwpsec dependency into
> > > > a single Tegra-specific helper, which would better represent the nature of
> > > > what's really going on here.
> > > 
> > > I don't see how this is an anti-pattern. It might not be common for
> > > drivers to need to reach into iommu_fwspec, so that might indeed be
> > > specific to Tegra (for whatever reason our IP seems to want extra
> > > flexibility), but the general pattern of using stubs is wide-spread,
> > > so I don't see why IOMMU_API would need to be special.
> > 
> > Again, I agree.
> 
> The anti-pattern is reaching into some other driver's private data assuming
> a particular format, with zero indication of the huge degree of assumption
> involved, and half the time not even checking that what's being dereferenced
> is valid.

If this is really driver private data that nobody else should be
accessing, then this certainly is lacking documentation. And quite
frankly, perhaps it should really be hidden from other drivers in
that case.

Remember the reason why we want to make this change is because we can
already get access to all this data if we depend on IOMMU_API, while all
the rest is also already available for !IOMMU_API configurations. This
change removes that inconsistency.

> > Moreover, a "git grep CONFIG_IOMMU_API" indicates that the problem
> > isn't specific to Tegra. The "#ifdef CONFIG_IOMMU_API" seems to be
> > sprinkled across the kernel. I think it would be nice if we could
> > improve the situation. So far, using stubs along with what the
> > $subject patch proposes, seems to me to be the best approach.
> 
> Yes, there is plenty of code through the tree that is only relevant to the
> IOMMU API and would be a complete waste of space without it, that is not the
> point in question here. Grep for dev_iommu_fwspec_get; outside
> drivers/iommu, the only users are IOMMU-API-specific parts of ACPI code, as
> intended, plus 8 random Tegra drivers.
> 
> Now, there does happen to be a tacit contract between the ACPI IORT code and
> the Arm SMMU drivers for how SMMU StreamIDs are encoded in their respective
> fwspecs, but it was never intended for wider consumption.

Again, if iommu_fwspec and its accessors were somehow hidden, or if it
was documented that this was not meant for wider consumption, then
perhaps we wouldn't have started using it in the first place.

>                                                           If Tegra drivers
> want to have a special relationship with arm-smmu then fair enough, but they
> can do the same as MSM and formalise it somewhere that the SMMU driver
> maintainers are at least aware of, rather than holding the whole generic
> IOMMU API hostage.
> 
> Since apparently it wasn't clear, what I was proposing is a driver helper at
> least something like this:
> 
> int tegra_arm_smmu_streamid(struct device *dev)
> {
> #ifdef CONFIG_IOMMU_API
> 	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev)
> 
> 	if (fwspec && fwspec->num_ids == 1)
> 		return fwspec->ids[0] & 0xffff;
> #endif
> 	return -EINVAL;
> }
> 
> Now THAT is scalable and maintainable; any number of random drivers can call
> it without any preconditions, it's a lot clearer what's going on, and I
> won't have to swear profusely while herding patches through half a dozen
> different trees if, when my ops rework gets to the point of refactoring
> iommu_fwspec with dev_iommu, it ends up changing anything significant.

I don't have any objection to making that change. It's not like we're
doing all of this just for fun. We need to do this in order for the
hardware to work correctly. If the above is an acceptable way to do this
and preferable to using the more generic API then we can move ahead with
that.

In order to keep things moving, do we want to merge the change as-is for
now and then subsequently do a pass over all Tegra drivers and use a
common function for stream ID access? Or do you want me to make that
change prior to getting the SDHCI series merged?

Thierry

Download attachment "signature.asc" of type "application/pgp-signature" (834 bytes)

Powered by blists - more mailing lists