[PATCH v5 14/29] iommufd/viommu: Add IOMMUFD_CMD_HW_QUEUE_ALLOC ioctl

[Nicolin Chen]

On Fri, May 30, 2025 at 02:40:37PM -0300, Jason Gunthorpe wrote:
> On Fri, May 30, 2025 at 10:38:24AM -0700, Nicolin Chen wrote:
> > On Fri, May 30, 2025 at 01:14:55PM -0300, Jason Gunthorpe wrote:
> > > On Sat, May 17, 2025 at 08:21:31PM -0700, Nicolin Chen wrote:
> > > > +	offset =
> > > > +		cmd->nesting_parent_iova - PAGE_ALIGN(cmd->nesting_parent_iova);
> > > > +	max_npages = DIV_ROUND_UP(offset + cmd->length, PAGE_SIZE);
> > > 
> > > This should probably be capped to PAGE_SIZE/sizeof(void *), return
> > > EINVAL if not
> > 
> > Hmm, mind elaborating where this PAGE_SIZE/sizeof comes from?
> 
> We can usually allocate up to a PAGE_SIZE without too much
> trouble. Beyond that it gets more likely to fail.

If PAGE_SIZE=4096, the upper limit for max_npages is 512, i.e. the
max size of a guest queue is 2MB? It seems to be too small, as the
VMM can use a larger huge page size to back the guest queue?

> > > > +	hw_queue->viommu = viommu;
> > > > +	refcount_inc(&viommu->obj.users);
> > > > +	hw_queue->length = cmd->length;
> > > > +	hw_queue->base_addr = cmd->nesting_parent_iova;
> > > 
> > > When the driver is running, which can be a source of bugs.
> > 
> > Hmm, I don't quite follow the "bugs" here. Any example?
> 
> Like if the driver thinks that hw_queue->length should be valid during
> init, it turns out it isn't.

Ah, I see. Yes.

Nicolin