Re: [PATCH v5 1/1] selftests/mm/uffd: Refactor non-composite global vars into struct

From: Peter Xu
Date: Tue Jun 17 2025 - 13:27:42 EST


On Tue, Jun 17, 2025 at 11:52:52AM -0400, Peter Xu wrote:
> On Mon, Jun 16, 2025 at 05:26:18PM -0700, Andrew Morton wrote:
> > On Mon, 16 Jun 2025 15:34:06 +0530 Ujwal Kundur <ujwal.kundur@xxxxxxxxx> wrote:
> >
> > > Refactor macros and non-composite global variable definitions into a
> > > struct that is defined at the start of a test and is passed around
> > > instead of relying on global vars.
> >
> > Well I guess that's nicer.
> >
> > > 5 files changed, 616 insertions(+), 542 deletions(-)
> >
> > It needs to be!
> >
> > Thanks, I'll queue it for testing while Peter thinks about it :)
>
> I didn't pay much attention on this one as I saw Brandan was actively
> reviewing it, which was great.
>
> This is definitely an improvement to the test. Thanks both!

I did give it a quick run today, but I found I hit this:

$ ./uffd-unit-tests
Testing UFFDIO_API (with syscall)... done
Testing UFFDIO_API (with /dev/userfaultfd)... done
Testing register-ioctls on anon... done
ERROR: munmap (errno=22, @uffd-common.c:277)

IIUC it's because after moving most globals to stack, they are not properly
zero-initialized anymore. In this case it failed at MEM_ANON of
register-ioctls test, trying to munmap() some address that will start to be
random garbage since it's on the stack. So maybe we at least need
something like this?

diff --git a/tools/testing/selftests/mm/uffd-unit-tests.c b/tools/testing/selftests/mm/uffd-unit-tests.c
index bed96f41c578..0b66ca3e7b82 100644
--- a/tools/testing/selftests/mm/uffd-unit-tests.c
+++ b/tools/testing/selftests/mm/uffd-unit-tests.c
@@ -1744,7 +1744,7 @@ int main(int argc, char *argv[])
mem_type = &mem_types[j];

/* Initialize global test options */
- uffd_global_test_opts_t gopts;
+ uffd_global_test_opts_t gopts = { 0 };

gopts.map_shared = mem_type->shared;
uffd_test_ops = mem_type->mem_ops;

Even with that, it fails somewhere later.

Ujwal, can you reproduce these issues and have a look?

Thanks,

--
Peter Xu