Re: [BUG] kernel BUG at fs/userfaultfd.c:385 after 04f5866e41fb

From: Oleg Nesterov <oleg@redhat.com>
To: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Michal Hocko <mhocko@suse.com>, linux-mm <linux-mm@kvack.org>,
	Andrea Arcangeli <aarcange@redhat.com>,
	Peter Xu <peterx@redhat.com>, Mike Rapoport <rppt@linux.ibm.com>,
	Jann Horn <jannh@google.com>, Jason Gunthorpe <jgg@mellanox.com>,
	Andrew Morton <akpm@linux-foundation.org>
Subject: Re: [BUG] kernel BUG at fs/userfaultfd.c:385 after 04f5866e41fb
Date: Mon, 19 Aug 2019 14:48:33 +0200	[thread overview]
Message-ID: <20190819124832.GA15044@redhat.com> (raw)
In-Reply-To: <3b521a8c-586f-251e-f486-d71ff094b8e9@huawei.com>

Kefeng, et al, sorry I am travelling till the next monday, can't even
open my laptop too often.

On 08/16, Kefeng Wang wrote:
>
> The patch do fix the UAF and avoid panic, and it doesn't seem to cause new issue,
> even if there are some another issue, it can be fixed later :)

Yes... but if we want a hot fix without any understanding why does it
help in this particular case we can simply add mmget_still_valid() into
handle_userfault(), I am almost sure this should equally "fix" (hide) the
problem.

But to clarify, let me repeat that I still think the patch I sent makes
sense anyway; __handle_mm_fault() is possible even after do_coredump()
sets mm->core_state and after userfaultfd_unregister().

> >> [   67.430243] RIP: 0010:copy_user_handle_tail+0x2/0x10
> >> [   67.431586] Code: c3 0f 1f 80 00 00 00 00 66 66 90 83 fa 40 0f 82 70 ff ff ff 89 d1 f3 a4 31 c0 66 66 90 c3 66 2e 0f 1f 84 00 00 00 00 00 89 d1 <f3> a4 89 c8 66 66 90 c3 66 0f 1f 44 00 00 66 66 90 83 fa 08 0f 82
> >> [   67.436978] RSP: 0018:ffff8883c4e8f908 EFLAGS: 00010246
> >> [   67.438743] RAX: 0000000000000001 RBX: 0000000020ffd000 RCX: 0000000000001000
> >> [   67.441101] RDX: 0000000000001000 RSI: 0000000020ffd000 RDI: ffff8883c0aa4000
> >> [   67.442865] RBP: 0000000000001000 R08: ffffed1078154a00 R09: 0000000000000000
> >> [   67.444534] R10: 0000000000000200 R11: ffffed10781549ff R12: ffff8883c0aa4000
> >> [   67.446216] R13: ffff8883c6096000 R14: ffff88837721f838 R15: ffff8883c6096000
> >> [   67.448388]  _copy_from_user+0xa1/0xd0
> >> [   67.449655]  mcopy_atomic+0xb3d/0x1380
> >> [   67.450991]  ? lock_downgrade+0x3a0/0x3a0
> >> [   67.452337]  ? mm_alloc_pmd+0x130/0x130
> >> [   67.453618]  ? __might_fault+0x7d/0xe0
> >> [   67.454980]  userfaultfd_ioctl+0x14a2/0x1c30
> >
> > This must not be called after __fput(). So I think there is something else,
> > may by just an unbalanced userfaultfd_ctx_put(). I dunno, I know nothing
> > about usefaultfd.
>
> There are different processes, maybe some concurrency problems.

and this is what I think we need to understand...

> > It would be nice to understand what this reproducer does...
>
> I tried strace -f the reproducer, but can't find any useful info.

Could you send me the output of strace -f ? Not sure it will help and
again, I am not sure I will have a chance to read it this week.

Perhaps it is possible to translate this test-case to C somehow?

Oleg.