Re: [nfsd4] potentially hardware breaking regression in 4.14-rc and 4.13.11

[Kees Cook]

On Fri, Nov 17, 2017 at 11:03 AM, Patrick McLean <chutzpah@xxxxxxxxxx> wrote:
> On 2017-11-16 04:54 PM, Kees Cook wrote:
>> On Mon, Nov 13, 2017 at 2:48 PM, Patrick McLean <chutzpah@xxxxxxxxxx> wrote:
>>> On 2017-11-11 09:31 AM, Linus Torvalds wrote:
>>>> Boris Lukashev points out that Patrick should probably check a newer
>>>> version of gcc.
>>>>
>>>> I looked around, and in one of the emails, Patrick said:
>>>>
>>>>   "No changes, both the working and broken kernels were built with
>>>>    distro-provided gcc 5.4.0 and binutils 2.28.1"
>>>>
>>>> and gcc-5.4.0 is certainly not very recent. It's not _ancient_, but
>>>> it's a bug-fix release to a pretty old branch that is not exactly new.
>>>>
>>>> It would probably be good to check if the problems persist with gcc
>>>> 6.x or 7.x.. I have no idea which gcc version the randstruct people
>>>> tend to use themselves.
>>>
>>> I just tested it with gcc 7.2, and was able to reproduce the NULL
>>> pointer dereference, the backtrace looks slightly different this time.
>>>
>>> I will also test with binutils 2.29, though I doubt that will make any
>>> difference.
>>>
>>>> [   56.165181] BUG: unable to handle kernel NULL pointer dereference at 0000000000000560
>>>> [   56.166563] IP: vfs_statfs+0x7c/0xc0
>>>> [   56.167249] PGD 0 P4D 0
>>>> [   56.167860] Oops: 0000 [#1] SMP
>>>> [   56.176478] Modules linked in: ipt_MASQUERADE nf_nat_masquerade_ipv4 xt_multiport xt_addrtype iptable_mangle iptable>
>>>> [   56.180227] CPU: 0 PID: 3985 Comm: nfsd Tainted: G           O    4.14.0-git-kratos-1 #1
>>>> [   56.181728] Hardware name: TYAN S5510/S5510, BIOS V2.02 03/12/2013
>>>> [   56.182729] task: ffff88040c412a00 task.stack: ffffc90002c18000
>>>> [   56.183629] RIP: 0010:vfs_statfs+0x7c/0xc0
>>>> [   56.184341] RSP: 0018:ffffc90002c1bb28 EFLAGS: 00010202
>>>> [   56.185143] RAX: 0000000000000000 RBX: ffffc90002c1bbf0 RCX: 0000000000000020
>>>> [   56.186085] RDX: 0000000000001801 RSI: 0000000000001801 RDI: 0000000000000000
>>>> [   56.187066] RBP: ffffc90002c1bbc0 R08: ffffffffffffff00 R09: 00000000000000ff
>>>> [   56.188268] R10: 000000000038be3a R11: ffff880408b18258 R12: 0000000000000000
>>>> [   56.189336] R13: ffff88040c23ad00 R14: ffff88040b874000 R15: ffffc90002c1bbf0
>>>> [   56.190444] FS:  0000000000000000(0000) GS:ffff88041fc00000(0000) knlGS:0000000000000000
>>>> [   56.191876] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>>>> [   56.192843] CR2: 0000000000000560 CR3: 0000000001e0a002 CR4: 00000000001606f0
>>>> [   56.193898] Call Trace:
>>>> [   56.194510]  nfsd4_encode_fattr+0x201/0x1f90
>>>> [   56.195267]  ? generic_permission+0x12c/0x1a0
>>>> [   56.196025]  nfsd4_encode_getattr+0x25/0x30
>>>> [   56.196753]  nfsd4_encode_operation+0x98/0x1b0
>>>> [   56.197526]  nfsd4_proc_compound+0x2a0/0x5e0
>>>> [   56.198268]  nfsd_dispatch+0xe8/0x220
>>>> [   56.198968]  svc_process_common+0x475/0x640
>>>> [   56.199696]  ? nfsd_destroy+0x60/0x60
>>>> [   56.200404]  svc_process+0xf2/0x1a0
>>>> [   56.201079]  nfsd+0xe3/0x150
>>>> [   56.201706]  kthread+0x117/0x130
>>>> [   56.202354]  ? kthread_create_on_node+0x40/0x40
>>>> [   56.203100]  ret_from_fork+0x25/0x30
>>>> [   56.203774] Code: d6 89 d6 81 ce 00 04 00 00 f6 c1 08 0f 45 d6 89 d6 81 ce 00 08 00 00 f6 c1 10 0f 45 d6 89 d6 81 ce>
>>>> [   56.206289] RIP: vfs_statfs+0x7c/0xc0 RSP: ffffc90002c1bb28
>>>> [   56.207110] CR2: 0000000000000560
>>>> [   56.207763] ---[ end trace d452986a80f64aaa ]---
>>>
>>>> On Sat, Nov 11, 2017 at 8:13 AM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>>>>>
>>>>> I'll take a closer look at this and see if I can provide something to
>>>>> narrow it down.
>>
>> How reliable is this crash? The best idea I have to isolate it would
>> be to bisect the additions of the __randomize_layout markings on
>> various structures. I would start with the ones Al is most upset to
>> see randomized. ;)
>
> It's pretty reliable, once I get a bad seed I can reproduce the crash
> pretty quickly.
>
>>
>> All that said, I'd like to better understand the BIOS side of this a
>> little better. In the first email in this thread, you showed two BUGs
>> separated by a little time, which implies to me that the NULL deref
>> and the BIOS no longer POSTing are separate (though seemingly related)
>> issues. Have you had machines survive the BUG without blowing up the
>> BIOS?
>
> We had 3 machines die due to the BIOS issue (all of them pretty quickly
> with the bad-seed kernel). All the dead machines had the same
> motherboard model. I have not managed to reproduce the issue again on
> the machine I restored via the IPMI interface, I suspect that it may be
> a bug in the BIOS that was fixed in a more recent version.
>
>>
>> I'm still trying to wrap my head around how the BIOS could be blowing
>> up. I assume there's some magic memory address that is getting poked
>> as a result of some struct randomization bug, so tracking that down
>> should be possible assuming you can stand reflashing your BIOS across
>> the bisects.
>
> That is our theory, some magic memory address that caused an overwrite
> of the flash where the BIOS code is stored. We are working under the
> assumption that it was fixed in a more recent BIOS update, since I have
> not managed to reproduce the issue on the resurrected machine.

Okay, well that's certainly better than having to reflash at every
bisection step! :)

>> For the first step, I'd try a revert of
>> 9225331b310821760f39ba55b00b8973602adbb5, which enables a large
>> portion of struct randomization. If that doesn't change things, I can
>> provide a series that reverts 3859a271a003aba01e45b85c9d8b355eb7bf25f9
>> and then re-applies __randomize_layout one structure per patch, and
>> you could bisect that?
>
> Sure, I can bisect that.

Okay, that should at least let us know if this is a specific struct
that is not expecting to get randomized, or if there is some deeper
flaw. Here's the tree, based on 4.14:
https://git.kernel.org/pub/scm/linux/kernel/git/kees/linux.git/log/?h=kspp/randstruct/bisection

With commit d9e12200852d, all randomization selections are reverted. I
would expect this to be a "good" kernel for the bisect.

The very end of the series (commit d893c17b3146), everything is back
to being randomized. I would expect this to be a "bad" kernel.

Each step between those two commits adds randomization to a single
struct (with the filesystem stuff near the front).

Here's hoping it'll be something obvious. :) Thanks for taking the
time to debug this!

-Kees

-- 
Kees Cook
Pixel Security