The script I mentioned in that other mail should help with this:
https://github.com/bjackman/linux/blob/github-base/.github/scripts/run_local...
That's my hacky attempt at a "hermetic" runner for these tests, it ought to let you get the exact Kconfig, userspace, and QEMU command that I used when I hit the issue.
The GitHub workflow definition shows how to get its dependencies installed on a Debian-alike:
https://github.com/bjackman/linux/blob/github-base/.github/workflows/test.ya...
Thanks for sharing this, I was able to reproduce and fix the problem. Turns out char *area_src_alias, area_dst_alias, area_remap were left uninitialized in uffd_test_ctx_init while it was working before this refactor because of the global nature of these variables. Using virtme-ng also helps me ensure that the next spin on this patch will work :)
Ideally the test changes should run the same before/after that series applied. Meanwhile, no conflict expected between the two, hence no worry on the order to land.
Sorry, I meant to link the uffd_poison_test changes [1], I've rebased on top of those changes.
Will send a v6 soon.
[1] https://lore.kernel.org/all/20250620150058.1729489-1-peterx@redhat.com/
Thanks, Ujwal