Re: f45ec5ff16 ("userfaultfd: wp: support swap and page migration"): [ 140.777858] BUG: Bad rss-counter state mm:b278fc66 type:MM_ANONPAGES val:1

[Peter Xu <peterx@redhat.com>]

On Fri, Apr 10, 2020 at 08:38:05AM -0700, Matthew Wilcox wrote:
> On Fri, Apr 10, 2020 at 11:32:34AM -0400, Peter Xu wrote:
> > I'm still trying to digest on what's happened... It would be good too
> > if more information on the test could be given, e.g., what is the
> > behavior of trinity-c2. A reproducer is of course even better.
> 
> Trinity is a syscall fuzzer.  Don't expect what it's doing to make any
> sense, it's just executing syscalls at random.

OK thanks.

Though I just noticed that the original report is actually with some
attachments which I totally missed initially.  There's the config file
showing that we're with:

CONFIG_MIGRATION=y
CONFIG_MEMORY_FAILURE=y
CONFIG_DEVICE_PRIVATE=n

And even a reproducer.  However the reproducer script will fail at
wget, until I fixed it using:

initrd=openwrt-trinity-i386.cgz

to replace:

initrd=openwrt-i386-trinity.cgz

Then I can download the initrd and boot the VM with a decent QEMU.
However I didn't see any test running after the VM booted, and it will
reboot/shutdown after 100 sec without any error triggered (I believe
the rc.local tries to run something under /etc/kernel-tests/ but I'm
not sure it's running the right thing).

If there's any way to reproduce (I believe so because it can even
bisect in the original report, I just don't know how...), I'm thinking
maybe we can try to dump every swp entry change that could have been
touched in change_pte_range(), which is the only place that I thought
could be related to this in the commit, to see whether there's
anything suspecious.

8<----------------------------------------------
diff --git a/mm/mprotect.c b/mm/mprotect.c
index 1d823b050329..1b6daf7d03aa 100644
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -173,6 +173,8 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
                                newpte = pte_swp_clear_uffd_wp(newpte);
 
                        if (!pte_same(oldpte, newpte)) {
+                               pr_info("%s: Update swp entry, 0x%lx -> 0x%lx\n",
+                                       __func__, pte_val(oldpte), pte_val(newpte));
                                set_pte_at(vma->vm_mm, addr, pte, newpte);
                                pages++;
                        }

-- 
Peter Xu