Re: [PATCH 3/4] iommu/io-pgtable-arm: Rationalise TCR handling

[Robin Murphy <robin.murphy@arm.com>]

On 20/08/2019 17:07, Will Deacon wrote:
> On Tue, Aug 20, 2019 at 04:25:56PM +0100, Robin Murphy wrote:
>> On 20/08/2019 11:31, Will Deacon wrote:
>>> On Mon, Aug 19, 2019 at 07:19:30PM +0100, Robin Murphy wrote:
>>>> Although it's conceptually nice for the io_pgtable_cfg to provide a
>>>> standard VMSA TCR value, the reality is that no VMSA-compliant IOMMU
>>>> looks exactly like an Arm CPU, and they all have various other TCR
>>>> controls which io-pgtable can't be expected to understand. Thus since
>>>> there is an expectation that drivers will have to add to the given TCR
>>>> value anyway, let's strip it down to just the essentials that are
>>>> directly relevant to io-pgatble's inner workings - namely the address
>>>> sizes, walk attributes, and where appropriate, format selection.
>>>>
>>>> Signed-off-by: Robin Murphy <robin.murphy@arm.com>
>>>> ---
>>>>    drivers/iommu/arm-smmu-v3.c        | 7 +------
>>>>    drivers/iommu/arm-smmu.c           | 1 +
>>>>    drivers/iommu/arm-smmu.h           | 2 ++
>>>>    drivers/iommu/io-pgtable-arm-v7s.c | 6 ++----
>>>>    drivers/iommu/io-pgtable-arm.c     | 4 ----
>>>>    drivers/iommu/qcom_iommu.c         | 2 +-
>>>>    6 files changed, 7 insertions(+), 15 deletions(-)
>>>
>>> Hmm, so I'm a bit nervous about this one since I think we really should
>>> be providing a TCR with EPD1 set if we're only giving you TTBR0. Relying
>>> on the driver to do this worries me. See my comments on the next patch.
>>
>> The whole idea is that we already know we can't provide a *complete* TCR
>> value (not least because anything above bit 31 is the wild west), thus
>> there's really no point in io-pgtable trying to provide anything other than
>> the parts it definitely controls. It makes sense to provide this partial TCR
>> value "as if" for TTBR0, since that's the most common case, but ultimately
>> io-pgatble doesn't know (or need to) which TTBR the caller intends to
>> actually use for this table. Even if the caller *is* allocating it for
>> TTBR0, io-pgtable doesn't know that they haven't got something live in TTBR1
>> already, so it still wouldn't be in a position to make the EPD1 call either
>> way.
> 
> Ok, but the driver can happily rewrite/ignore what it gets back. I suppose
> an alternative would be scrapped the 'u64 tcr' and instead having a bunch
> of named bitfields for the stuff we're actually providing, although I'd
> still like EPDx to be in there.

I like the bitfield idea; it would certainly emphasise the "you have to 
do something more with this" angle that I'm pushing towards here, but 
still leave things framed in TCR terms without having to go to some more 
general abstraction. It really doesn't play into your EPD argument 
though - such a config would be providing TxSZ/TGx/IRGNx/ORGNx/SHx, but 
EPDy, for y = !x. For a driver to understand that and do the right thing 
with it is even more involved than for the driver to just set EPD1 by 
itself anyway.

>> Ultimately, it's the IOMMU drivers who decide what they put in which TTBR,
>> so it's the IOMMU drivers which have to take responsibility for EPD*. Sure
>> you can worry about it, but you can equally worry about them them
>> misprogramming the ASID or anything else...
> 
> I find the EPDx bits particularly dangerous because:
> 
>    - They're easily overlooked
>    - Clobbering TTBR1 with 0x0 doesn't disable walks via TTBR1 as you might
>      reasonably expect

(FWIW I'm not sure that that is a reasonable expectation, at least for 
anyone savvy enough to be programming an MMU in the first place. There 
are plenty of systems with RAM at 0x0)

>    - If you do the above without EPD, the breakage will be subtle
> 
> and given that I don't see any real downsides to us providing a default TCR
> value with EPD set appropriately, then I think we should do that. I'd be
> happy to revisit the decision later on if it's getting the way of a real
> use-case, but it feels like we're throwing the baby out with the bathwater
> at the moment and I'd rather do this incrementally based on actual need.
The downside is maintaining extra complexity for the sake of a 
theoretical concern which hasn't been borne out in practice, with a 
promise of yet more complexity down the line. Moving the 3 babies which 
anyone acknowledges into their own private baths takes a whole -8 lines 
of code to implement. That said, if we don't go down the bitfield route, 
then I *can* leave the default TCR having EPD1 set if that makes you 
feel warm and safe, but it will still be resoundingly ignored.

If only LPAE had created these bits as enables rather than disables then 
things would be logical and we could all be happy, but here we are...

Robin.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel