Re: [PATCH v2 12/12] iommu: Add helper to set iopf handler for domain

[Jason Gunthorpe <jgg@xxxxxxxx>]

On Fri, Aug 11, 2023 at 10:40:15AM +0800, Baolu Lu wrote:
> On 2023/8/11 3:18, Jason Gunthorpe wrote:
> > On Thu, Jul 27, 2023 at 01:48:37PM +0800, Lu Baolu wrote:
> > > To avoid open code everywhere.
> > > 
> > > Signed-off-by: Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx>
> > > ---
> > >   include/linux/iommu.h | 11 ++++++++++-
> > >   drivers/iommu/iommu.c | 20 ++++++++++++++++++--
> > >   2 files changed, 28 insertions(+), 3 deletions(-)
> > 
> > Seems like overkill at this point..
> > 
> > Also, I think this is probably upside down.
> > 
> > We want to create the domains as fault enabled in the first place.
> > 
> > A fault enabled domain should never be attached to something that
> > cannot support faults. It should also not support changing the fault
> > handler while it exists.
> > 
> > Thus at the creation point would be the time to supply the fault handler
> > as part of requesting faulting.
> > 
> > Taking an existing domain and making it faulting enabled is going to
> > be really messy in all the corner cases.
> 
> Yes. Agreed.
> 
> > 
> > My advice (and Robin will probably hate me), is to define a new op:
> > 
> > struct domain_alloc_paging_args {
> >         struct fault_handler *fault_handler;
> >         void *fault_data
> > };
> > 
> > struct iommu_domain *domain_alloc_paging2(struct device *dev, struct
> >         domain_alloc_paging_args *args)
> > 
> > The point would be to leave the majority of drivers using the
> > simplified, core assisted, domain_alloc_paging() interface and they
> > just don't have to touch any of this stuff at all.
> > 
> > Obviously if handler is given then the domain will be initialized as
> > faulting.
> 
> Perhaps we also need an internal helper for iommu drivers to check the
> iopf capability of the domain.

Yeah, maybe.

I've been mulling over this for a a bit here

Robin suggested to wrap everything in a arg to domain_alloc and build
a giant super multiplexor

I don't really like that because it makes it quite complicated for the
driver and multiplexor APIs are rarely good.

So for simple drivers I like the 'domain_alloc_paging' as the only op
they implement and it is obviously simple and hard to implement
wrong. Most drivers would do this.

We also need a:

struct iommu_domain *domain_alloc_sva(struct device *dev, struct mm_struct *mm)

So SVA can be fully setup at allocation time. SVA doesn't have any
legal permutations so it can be kept simple.

Then we need something to bundle:
 - Dirty tracking yes/no
 - The iommufd user space blob
 - Fault handling yes/no

For complex drivers.

So maybe we should just have a 3rd option

// I'm a complex driver and many people checked that I implemented
// this right:
struct domain_alloc_args {
       struct device *dev;
       unsigned int flags; // For requesting dirty tracking

       // alloc_domain_user interface
       struct iommu_domain *parent;
       void *user_data;
       size_t user_len;

       // Faulting
       struct fault_handler *fault_handler;
       void *fault_data;	
};
struct iommu_domain *domain_alloc(struct domain_alloc_args *args);

?

IDK, multiplexor APIs are rarely good, but maybe this is the right
direction?

Jason