Re: a crash when running strace from persistent memory

[Linus Torvalds]

On Thu, Sep 3, 2020 at 12:24 PM Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote:
>
> There's a bug when you run strace from dax-based filesystem.
>
> -- create real or emulated persistent memory device (/dev/pmem0)
> mkfs.ext2 /dev/pmem0
> -- mount it
> mount -t ext2 -o dax /dev/pmem0 /mnt/test
> -- copy the system to it (well, you can copy just a few files that are
>    needed for running strace and ls)
> cp -ax / /mnt/test
> -- bind the system directories
> mount --bind /dev /mnt/test/dev
> mount --bind /proc /mnt/test/proc
> mount --bind /sys /mnt/test/sys
> -- run strace on the ls command
> chroot /mnt/test/ strace /bin/ls
>
> You get this warning and ls is killed with SIGSEGV.
>
> I bisected the problem and it is caused by the commit
> 17839856fd588f4ab6b789f482ed3ffd7c403e1f (gup: document and work around
> "COW can break either way" issue). When I revert the patch (on the kernel
> 5.9-rc3), the bug goes away.

Funky. I really don't see how it could cause that, but we have the
UDDF issue too, so I'm guessing I will have to fix it the radical way
with Peter Xu's series based on my "rip out COW special cases" patch.

Or maybe I'm just using that as an excuse for really wanting to apply
that series.. Because we can't just revert that GUP commit due to
security concerns.

> [   84.191504] WARNING: CPU: 6 PID: 1350 at mm/memory.c:2486 wp_page_copy.cold+0xdb/0xf6

I'm assuming this is the WARN_ON_ONCE(1) on line 2482, and you have
some extra debug patch that causes that line to be off by 4? Because
at least for me, line 2486 is actually an empty line in v5.9-rc3.

That said, I really think this is a pre-existing race, and all the
"COW can break either way" patch does is change the timing (presumably
due to the actual pattern of actually doing the COW changing).

See commit c3e5ea6ee574 ("mm: avoid data corruption on CoW fault into
PFN-mapped VMA") for background.

Mikulas, can you check that everything works ok for that case if you
apply Peter's series? See

    https://lore.kernel.org/lkml/20200821234958.7896-1-peterx@xxxxxxxxxx/

or if you have 'b4' installed, use

    b4 am 20200821234958.7896-1-peterx@xxxxxxxxxx

to get the series..

                     Linus