Re: [PATCH] iommu: iommu_group_alloc_default_domain: Fix ops->default_domain codepath

[Ovidiu Panait]

Hi Jason,

On 23.01.2024 19:49, Jason Gunthorpe wrote:
> CAUTION: This email comes from a non Wind River email account!
> Do not click links or open attachments unless you recognize the sender and know the content is safe.
>
> On Tue, Jan 23, 2024 at 06:58:29PM +0200, ovidiu.panait@xxxxxxxxxxxxx wrote:
> > From: Ovidiu Panait <ovidiu.panait@xxxxxxxxxxxxx>
> >
> > Since commit [1], FSL_PAMU initialization on t4240rdb board fails
> > with: "fsl-pamu-domain: pamu_domain_init: Can't register iommu device".
> >
> > Commit [1] changed the behavior for drivers that set ops->default_domain,
> > as the function iommu_group_alloc_default_domain() is now called with
> > "req_type == IOMMU_DOMAIN_IDENTITY", rather than "req_type == 0". This
> > causes the following EINVAL condition to be hit during FSL_PAMU
> > initialization:
> >    ...
> >         if (ops->default_domain) {
> >                 if (req_type)
> >                         return ERR_PTR(-EINVAL);
> >                 return ops->default_domain;
> >         }
> >    ...
> >
> > To avoid this issue when ops->default_domain != NULL, remove the check for
> > non-zero req_type and always return the proper default domain, as set in
> > the driver.
> >
> >    [1] 0f6a90436a57 ("iommu: Do not use IOMMU_DOMAIN_DMA [...]")
> >
> > Fixes: 0f6a90436a57 ("iommu: Do not use IOMMU_DOMAIN_DMA if CONFIG_IOMMU_DMA is not enabled")
> > Signed-off-by: Ovidiu Panait <ovidiu.panait@xxxxxxxxxxxxx>
> > ---
> >   drivers/iommu/iommu.c | 5 +----
> >   1 file changed, 1 insertion(+), 4 deletions(-)
> Oh, that is a neat combination of things
>
> Removing the test will cause problems for the sys flow - I think the
> right thing is to get the correct req_type earlier. paml wants to use
> a specific domain type so it should not be switching to identity at
> all.
>
> Does the below work for you?
Yes, this works, thanks a lot!
> Also, what is your interest here? I have been wanting to delete this
> driver, would your system still be OK for your usage if it was
> removed?
CONFIG_FSL_PAMU is enabled by default in corenet64_smp_defconfig, and it 
looks
like many platforms are based on this config. I actually noticed this issue
while boot testing the mainline kernel on the t4240rdb board with the 
default
corenet64_smp_defconfig.

In this particular case, the board still boots fine with CONFIG_FSL_PAMU 
disabled,
but I'm not sure how/if removing the PAMU driver will impact other 
platforms.

Ovidiu
> Thanks,
> Jason
>
> diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
> index 68e648b5576706..7fa3ccdc3286e6 100644
> --- a/drivers/iommu/iommu.c
> +++ b/drivers/iommu/iommu.c
> @@ -1799,7 +1799,7 @@ iommu_group_alloc_default_domain(struct iommu_group *group, int req_type)
>           * domain. Do not use in new drivers.
>           */
>          if (ops->default_domain) {
> -               if (req_type)
> +               if (req_type != ops->default_domain->type)
>                          return ERR_PTR(-EINVAL);
>                  return ops->default_domain;
>          }
> @@ -1871,10 +1871,18 @@ static int iommu_get_def_domain_type(struct iommu_group *group,
>          const struct iommu_ops *ops = dev_iommu_ops(dev);
>          int type;
>
> -       if (!ops->def_domain_type)
> -               return cur_type;
> -
> -       type = ops->def_domain_type(dev);
> +       if (ops->default_domain) {
> +               /*
> +                * Drivers that declare a global static default_domain will
> +                * always choose that.
> +                */
> +               type = ops->default_domain->type;
> +       } else {
> +               if (ops->def_domain_type)
> +                       type = ops->def_domain_type(dev);
> +               else
> +                       type = cur_type;
> +       }
>          if (!type || cur_type == type)
>                  return cur_type;
>          if (!cur_type)