Re: [PATCH 2/2] iommufd/selftest: Test reserved regions near ULONG_MAX

[Nicolin Chen <nicolinc@nvidia.com>]

On Fri, Jul 18, 2025 at 07:23:25PM +0100, Robin Murphy wrote:
> On 2025-07-18 9:13 am, Nicolin Chen wrote:
> > On Thu, Jul 17, 2025 at 07:45:51PM -0700, Nicolin Chen wrote:
> > > On Thu, Jul 17, 2025 at 04:15:09PM -0300, Jason Gunthorpe wrote:
> > > 
> > > > +TEST_F(iommufd_ioas, reserved_overflow)
> > > > +{
> > > > +	struct iommu_test_cmd test_cmd = {
> > > > +		.size = sizeof(test_cmd),
> > > > +		.op = IOMMU_TEST_OP_ADD_RESERVED,
> > > > +		.id = self->ioas_id,
> > > > +		.add_reserved = { .start = 6,
> > > > +				  .length = 0xffffffffffff8001 },
> > > > +	};
> > > > +	__u64 iova;
> > > > +
> > > > +	ASSERT_EQ(0,
> > > > +		  ioctl(self->fd, _IOMMU_TEST_CMD(IOMMU_TEST_OP_ADD_RESERVED),
> > > > +			&test_cmd));
> > > > +	test_err_ioctl_ioas_map(ENOSPC, buffer, 0x5000, &iova);
> > > 
> > > When:
> > > PAGE_SIZE=SZ_64K = 0x10000
> > > MOCK_PAGE_SIZE = PAGE_SIZE / 2 = 0x8000
> > > 
> > > This likely fails the alignment test, returning -EINVAL instead:
> > > 
> > > # iommufd.c:988:reserved_overflow:Expected 28 (28) == errno (22)
> > > # reserved_overflow: Test failed
> > > #          FAIL  iommufd_ioas.mock_domain_limit.reserved_overflow
> > > 
> > > So, I think we'd have to pick a number aligned to MOCK_PAGE_SIZE?
> > > e.g. changing to 0x18000 for example can pass.
> > 
> > I realized that we can't change the number as it won't reproduce
> > on PAGE_SIZE=4K. So, perhaps it should just SKIP other page sizes
> > than 4K.
> 
> Shouldn't it work to parametrise both numbers accordingly, e.g. (if my
> Friday-evening maths holds up) reserve "1UL - MOCK_PAGE_SIZE * 16" then map
> "MOCK_PAGE_SIZE * 10"?

Ah, you are right, I forgot to change the reserved range.

0xfffffffffff80001 and 0x50000 could repro with 64K pages,
exactly what your math works :)

Thanks
Nicolin