Re: [RFC V1 4/4] iommufd: update VA

[Jason Gunthorpe <jgg@nvidia.com>]

On Mon, Aug 05, 2024 at 03:02:22PM -0400, Steven Sistare wrote:
> On 7/22/2024 11:51 AM, Jason Gunthorpe wrote:
> > On Sat, Jul 20, 2024 at 11:56:45AM -0700, Steve Sistare wrote:
> > > @@ -458,6 +459,76 @@ static void change_mm(struct iopt_pages *pages)
> > >   	free_uid(old_user);
> > >   }
> > > +#define MAX_UMAP_ENTRIES 1024
> > 
> > Only 170 of these fit in a 4k page :\
> > 
> > Are you worried about the high order allocation this could trigger?
> 
> Should I worry?

Probably? Experiance shows busy servers cannot do a high order
allocation, even order 4 is likely to fail due to fragmentation. If
you imagine you will need more than 1-4 pages of this then it would
probably be a good idea to avoid the higher order allocation.,

> I define a max to guard against a DOS from userland requesting a huge number
> of entries and causing the driver to allocate a huge amount of memory.
> 
> The use cases I have seen only have dozens of entries.  Do you have a figure
> in mind for the upper bound?

Then it is probably fine as is

Jason