[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <ZM2FrQcD8QzP6XM0@Asurada-Nvidia>
Date: Fri, 4 Aug 2023 16:11:41 -0700
From: Nicolin Chen <nicolinc@...dia.com>
To: Jason Gunthorpe <jgg@...dia.com>
CC: Michael Shavit <mshavit@...gle.com>, <iommu@...ts.linux.dev>,
<linux-arm-kernel@...ts.infradead.org>,
<linux-kernel@...r.kernel.org>, <will@...nel.org>,
<robin.murphy@....com>, <jean-philippe@...aro.org>
Subject: Re: [PATCH v4 6/8] iommu/arm-smmu-v3: Move CD table to
arm_smmu_master
On Fri, Aug 04, 2023 at 07:46:23PM -0300, Jason Gunthorpe wrote:
> On Fri, Aug 04, 2023 at 03:25:43PM -0700, Nicolin Chen wrote:
> > > @@ -2436,22 +2419,14 @@ static int arm_smmu_attach_dev(struct iommu_domain *domain, struct device *dev)
> > > if (!smmu_domain->smmu) {
> > > smmu_domain->smmu = smmu;
> > > ret = arm_smmu_domain_finalise(domain, master);
> > > - if (ret) {
> > > + if (ret)
> > > smmu_domain->smmu = NULL;
> > > - goto out_unlock;
> > > - }
> > > - } else if (smmu_domain->smmu != smmu) {
> > > + } else if (smmu_domain->smmu != smmu)
> > > ret = -EINVAL;
> > > - goto out_unlock;
> > > - } else if (smmu_domain->stage == ARM_SMMU_DOMAIN_S1 &&
> > > - master->ssid_bits != smmu_domain->cd_table.max_cds_bits) {
> > > - ret = -EINVAL;
> > > - goto out_unlock;
> > > - } else if (smmu_domain->stage == ARM_SMMU_DOMAIN_S1 &&
> > > - smmu_domain->cd_table.stall_enabled != master->stall_enabled) {
> > > - ret = -EINVAL;
> > > - goto out_unlock;
> > > - }
> >
> > ... then we remove this stall_enabled sanity also.
> >
> > This means a shared domain (holding a shared CD) being inserted
> > to two CD tables from two masters would have two different CDTE
> > configurations at the stall bit.
>
> I looked through the spec for a while and I thought this was fine..
>
> Stall is basically a master specific behavior on how to operate page
> faulting. It makes sense that it follows the master and the IOPTEs in
> the domain can be used with both the faulting and non-faulting page
> faulting path.
>
> I would expect the page faulting path to figure out what to (if there
> is anything special to do) do based on the master that triggered the
> fault, not based on the domain that received it.
Yea, I went through the spec too yet didn't find anything that
could block us. And there is no SW dependency on the STALL bit
of the CDTE: actually it has an inverse relationship with the
S1STALLD bit in the STE, so following the STE/cd_table/master
makes sense. So long as a master has its own cd_table holding
its own CDTE for a shared domain, HW CD caching should be fine
as well.
With that being said, I think mentioning this behavior change
in the commit log wouldn't hurt. Someday people might want to
check this out in case something breaks.
Thanks
Nic
Powered by blists - more mailing lists