Re: [RFC PATCH v2 4/9] iommu/arm-smmu-v3-sva: Allocate new ASID from installed_smmus

From: Michael Shavit
Date: Wed Aug 23 2023 - 03:26:42 EST


On Tue, Aug 22, 2023 at 9:19 PM Jason Gunthorpe <jgg@xxxxxxxxxx> wrote:
>
> On Tue, Aug 22, 2023 at 06:57:00PM +0800, Michael Shavit wrote:
> > Pick an ASID that is within the supported range of all SMMUs that the
> > domain is installed to.
> >
> > Signed-off-by: Michael Shavit <mshavit@xxxxxxxxxx>
> > ---
> >
> > (no changes since v1)
> >
> > .../iommu/arm/arm-smmu-v3/arm-smmu-v3-sva.c | 23 +++++++++++++++----
> > 1 file changed, 19 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3-sva.c b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3-sva.c
> > index fe88a7880ad57..92d2f8c4e90a8 100644
> > --- a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3-sva.c
> > +++ b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3-sva.c
> > @@ -66,6 +66,20 @@ static int arm_smmu_write_ctx_desc_devices(struct arm_smmu_domain *smmu_domain,
> > return ret;
> > }
> >
> > +static u32 arm_smmu_domain_max_asid_bits(struct arm_smmu_domain *smmu_domain)
> > +{
> > + struct arm_smmu_master *master;
> > + unsigned long flags;
> > + u32 asid_bits = 16;
> > +
> > + spin_lock_irqsave(&smmu_domain->devices_lock, flags);
> > + list_for_each_entry(master, &smmu_domain->devices,
> > + domain_head)
> > + asid_bits = min(asid_bits, master->smmu->asid_bits);
> > + spin_unlock_irqrestore(&smmu_domain->devices_lock, flags);
> > + return asid_bits;
> > +}
>
> I still don't like this, it is not locked properly. You release the
> devices_lock which means the max_asid could change before we get to
> arm_smmu_write_ctx_desc()

Good point.

> If you want to take this shortcut temporarily then a global max_asid
> is probably a better plan. Change it to a per-master allocation later
> to remove that.

Two options there:
1. When allocating a new ASID in arm_smmu_share_asid, limit ourselves
to 8-bit-wide ASIDs regardless of whether all the installed SMMUs
support 16bit ASIDs.
2. In addition, also use a maximum 8-bit-wide ASID when allocating
asids in arm_smmu_domain_finalise_s1.

The first one has minimal impact since arm_smmu_share_asid is
supposedly rare, and is a simple replacement for this patch.

The second one is more intrusive since we'd be limiting the number of
dma/unmanaged domains to a fairly small number, but it has the
advantage of allowing those domains to always successfully attach to
masters belonging to SMMUs with different asid_bits values (without
having to re-allocate a new ASID for the domain arm_smmu_share_asid
style). Where-as this series simply fails to attach the domain in such
scenarios.