On 27/03/2024 18:17, Muhammad Usama Anjum wrote:
On 3/27/24 11:09 PM, Joao Martins wrote:
On 27/03/2024 17:49, Muhammad Usama Anjum wrote:
On 3/27/24 7:59 PM, Joao Martins wrote:
On 27/03/2024 11:49, Jason Gunthorpe wrote:
On Wed, Mar 27, 2024 at 03:14:25PM +0500, Muhammad Usama Anjum wrote:
On 3/26/24 8:03 PM, Jason Gunthorpe wrote: > On Tue, Mar 26, 2024 at 06:09:34PM +0500, Muhammad Usama Anjum wrote: >> Even after applying this config patch and following snippet (which doesn't >> terminate the program if mmap doesn't allocate exactly as the hint), I'm >> finding failed tests. >> >> @@ -1746,7 +1748,7 @@ FIXTURE_SETUP(iommufd_dirty_tracking) >> assert((uintptr_t)self->buffer % HUGEPAGE_SIZE == 0); >> vrc = mmap(self->buffer, variant->buffer_size, PROT_READ | PROT_WRITE, >> mmap_flags, -1, 0); >> - assert(vrc == self->buffer); >> + assert(vrc == self->buffer);// ??? >> >> On x86: >> # Totals: pass:176 fail:4 xfail:0 xpass:0 skip:0 error:0 >> On ARM64: >> # Totals: pass:166 fail:14 xfail:0 xpass:0 skip:0 error:0 >> >> The log files are attached. > > You probably don't have enough transparent huge pages available to the process > > echo 1024 > /proc/sys/vm/nr_hugepages After making huge pages available, the iommufd test always passed on x86. But there are still failures on arm64. I'm looking into the failures.
Oh that is really strange. Joao? Nicolin?
Definitely strange, I'll have a look.
So it set the expected number of dirty bits as that assert doesn't fail, but it is failing when we check that even bits are set but not odd ones. Like it's hasn't set those bits.
For mock tests there should be no difference between x86 and ARM assuming the typical 4K page-size. Maybe this is 64k base pages in ARM? That's the only thing that I can think of that affected mock domain.
The config is attached. The defaults are being used i.e., 4k page.
Looks like CONFIG_IOMMUFD_DRIVER is not defined :(
I'll retest with this config and update the patch to include it in the config fragment needed for this test. Once we add all required config options in config fragment, the test should never fail. Somehow this gets included in the x86, but not on ARM.
The option is automatically selected by an user of iova_bitmap_set(). x86 has AMD and Intel IOMMUs which use it and auto selected (ARM64 doesn't yet, but soon to be a reality). Some vfio drivers get it auto-selected it as well.
But I forgot to do the same thing for iommufd-test kconfig in commit a9af47e382a ("iommufd/selftest: Test IOMMU_HWPT_GET_DIRTY_BITMAP") where mock domain also uses iova_bitmap_set(). So your patch is likely not the right place to set CONFIG_IOMMUFD_DRIVER=y.
See Jason's snip on auto-selecting if IOMMUFD_TEST is set.