Re: [PATCH v2 23/24] iommu/arm-smmu: Allow non-strict in pgtable_quirks interface

From: Robin Murphy <robin.murphy@arm.com>
To: Will Deacon <will@kernel.org>
Cc: joro@8bytes.org, iommu@lists.linux-foundation.org,
	linux-arm-kernel@lists.infradead.org,
	linux-kernel@vger.kernel.org, suravee.suthikulpanit@amd.com,
	baolu.lu@linux.intel.com, john.garry@huawei.com,
	dianders@chromium.org
Subject: Re: [PATCH v2 23/24] iommu/arm-smmu: Allow non-strict in pgtable_quirks interface
Date: Tue, 3 Aug 2021 13:13:12 +0100	[thread overview]
Message-ID: <529750e7-89a0-cfb7-8cef-36bb78b06340@arm.com> (raw)
In-Reply-To: <20210803103559.GA30690@willie-the-truck>

On 2021-08-03 11:36, Will Deacon wrote:
> On Mon, Aug 02, 2021 at 03:15:50PM +0100, Robin Murphy wrote:
>> On 2021-08-02 14:04, Will Deacon wrote:
>>> On Wed, Jul 28, 2021 at 04:58:44PM +0100, Robin Murphy wrote:
>>>> To make io-pgtable aware of a flush queue being dynamically enabled,
>>>> allow IO_PGTABLE_QUIRK_NON_STRICT to be set even after a domain has been
>>>> attached to, and hook up the final piece of the puzzle in iommu-dma.
>>>>
>>>> Signed-off-by: Robin Murphy <robin.murphy@arm.com>
>>>> ---
>>>>    drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c | 15 +++++++++++++++
>>>>    drivers/iommu/arm/arm-smmu/arm-smmu.c       | 11 +++++++++++
>>>>    drivers/iommu/dma-iommu.c                   |  3 +++
>>>>    3 files changed, 29 insertions(+)
>>>>
>>>> diff --git a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
>>>> index 19400826eba7..40fa9cb382c3 100644
>>>> --- a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
>>>> +++ b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
>>>> @@ -2711,6 +2711,20 @@ static int arm_smmu_enable_nesting(struct iommu_domain *domain)
>>>>    	return ret;
>>>>    }
>>>> +static int arm_smmu_set_pgtable_quirks(struct iommu_domain *domain,
>>>> +		unsigned long quirks)
>>>> +{
>>>> +	struct arm_smmu_domain *smmu_domain = to_smmu_domain(domain);
>>>> +
>>>> +	if (quirks == IO_PGTABLE_QUIRK_NON_STRICT && smmu_domain->pgtbl_ops) {
>>>> +		struct io_pgtable *iop = io_pgtable_ops_to_pgtable(smmu_domain->pgtbl_ops);
>>>> +
>>>> +		iop->cfg.quirks |= IO_PGTABLE_QUIRK_NON_STRICT;
>>>> +		return 0;
>>>> +	}
>>>> +	return -EINVAL;
>>>> +}
>>>
>>> I don't see anything serialising this against a concurrent iommu_unmap(), so
>>> the ordering and atomicity looks quite suspicious to me here. I don't think
>>> it's just the page-table quirks either, but also setting cookie->fq_domain.
>>
>> Heh, I confess to very much taking the cheeky "let's say nothing and see
>> what Will thinks about concurrency" approach here :)
> 
> Damnit, I fell for that didn't I?
> 
> Overall, I'm really nervous about the concurrency here and think we'd be
> better off requiring the unbind as we have for the other domain changes.

Sure, the dynamic switch is what makes it ultimately work for Doug's 
use-case (where the unbind isn't viable), but I had every expectation 
that we might need to hold back those two patches for much deeper 
consideration. It's no accident that the first 22 patches can still be 
usefully applied without them!

In all honesty I don't really like this particular patch much, mostly 
because I increasingly dislike IO_PGTABLE_QUIRK_NON_STRICT at all, but 
since the interface was there it made it super easy to prove the 
concept. I have a more significant refactoring of the io-pgtable code 
sketched out in my mind already, it's just going to be more work.

>> The beauty of only allowing relaxation in the strict->non-strict direction
>> is that it shouldn't need serialising as such - it doesn't matter if the
>> update to cookie->fq_domain is observed between iommu_unmap() and
>> iommu_dma_free_iova(), since there's no correctness impact to queueing IOVAs
>> which may already have been invalidated and may or may not have been synced.
>> AFAICS the only condition which matters is that the setting of the
>> io-pgtable quirk must observe fq_domain being set. It feels like there must
>> be enough dependencies on the read side, but we might need an smp_wmb()
>> between the two in iommu_dma_init_fq()?
>>
>> I've also flip-flopped a bit on whether fq_domain needs to be accessed with
>> READ_ONCE/WRITE_ONCE - by the time of posting I'd convinced myself that it
>> was probably OK, but looking again now I suppose this wacky reordering is
>> theoretically possible:
>>
>>
>> 	iommu_dma_unmap() {
>> 		bool free_fq = cookie->fq_domain; // == false
>>
>> 		iommu_unmap();
>>
>> 		if (!cookie->fq_domain) // observes new non-NULL value
>> 			iommu_tlb_sync(); // skipped
>>
>> 		iommu_dma_free_iova { // inlined
>> 			if (free_fq) // false
>> 				queue_iova();
>> 			else
>> 				free_iova_fast(); // Uh-oh!
>> 		}
>> 	}
>>
>> so although I still can't see atomicity being a problem I guess we do need
>> it for the sake of reordering at least.
> 
> With your changes, I think quite a few things can go wrong.
> 
>    * cookie->fq_domain may be observed but iovad->fq could be NULL

Good point, I guess that already technically applies (if iovad->fq sat 
in a write buffer long enough), but it certainly becomes far easier to 
provoke. However a barrier after assigning fq_domain (as mentioned 
above) paired with the control dependency around the queue_iova() call 
would also fix that, right?

>    * We can miss the smp_wmb() in the pgtable code but end up deferring the
>      IOVA reclaim
>    * iommu_change_dev_def_domain() only holds the group mutex afaict, so can
>      possibly run concurrently with itself on the same domain?
>    * iommu_dma_init_fq() can flip the domain type back from
>      IOMMU_DOMAIN_DMA_FQ to IOMMU_DOMAIN_DMA on the error path
>    * set_pgtable_quirks() isn't atomic, which probably is ok for now, but the
>      moment we use it anywhere else it's dangerous

In other words, IO_PGTABLE_QUIRK_NON_STRICT is definitely the problem. 
I'll have a hack on that this afternoon, and if it starts to look 
rabbit-holey I'll split this bit off and post v3 of the rest of the series.

If all the io-pgtable and fq behaviour for a given call could be 
consistent based on a single READ_ONCE(cookie->fq_domain) in 
iommu_dma_unmap(), do you see any remaining issues other than the point 
above?

Thanks,
Robin.