[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <CAKHBV27tdN6Wqovy41iOQ3dzSPWD4XSHgJgyBPoW_ttDCD3eyw@mail.gmail.com>
Date: Wed, 23 Aug 2023 15:26:00 +0800
From: Michael Shavit <mshavit@...gle.com>
To: Jason Gunthorpe <jgg@...dia.com>
Cc: iommu@...ts.linux.dev, linux-arm-kernel@...ts.infradead.org,
linux-kernel@...r.kernel.org, nicolinc@...dia.com,
tina.zhang@...el.com, jean-philippe@...aro.org, will@...nel.org,
robin.murphy@....com, Dawei Li <set_pte_at@...look.com>,
Joerg Roedel <joro@...tes.org>,
"Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>,
Lu Baolu <baolu.lu@...ux.intel.com>,
Mark Brown <broonie@...nel.org>
Subject: Re: [RFC PATCH v2 4/9] iommu/arm-smmu-v3-sva: Allocate new ASID from installed_smmus
On Tue, Aug 22, 2023 at 9:19 PM Jason Gunthorpe <jgg@...dia.com> 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@...gle.com>
> > ---
> >
> > (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.
Powered by blists - more mailing lists