lkml.org 
[lkml]   [2017]   [Mar]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: netlink: GPF in netlink_unicast
On 2017-03-07 14:23, Paul Moore wrote:
> On Tue, Mar 7, 2017 at 1:44 PM, Paul Moore <paul@paul-moore.com> wrote:
> > On Tue, Mar 7, 2017 at 10:55 AM, Richard Guy Briggs <rgb@redhat.com> wrote:
> >> On 2017-03-07 09:29, Paul Moore wrote:
> >>> On Mon, Mar 6, 2017 at 11:03 PM, Richard Guy Briggs <rgb@redhat.com> wrote:
> >>> > On 2017-03-06 10:10, Cong Wang wrote:
> >>> >> On Mon, Mar 6, 2017 at 2:54 AM, Dmitry Vyukov <dvyukov@google.com> wrote:
> >>> >> > Hello,
> >>> >> >
> >>> >> > I've got the following crash while running syzkaller fuzzer on
> >>> >> > net-next/8d70eeb84ab277377c017af6a21d0a337025dede:
> >>> >> >
> >>> >> > kasan: GPF could be caused by NULL-ptr deref or user memory access
> >>> >> > general protection fault: 0000 [#1] SMP KASAN
> >>> >> > Dumping ftrace buffer:
> >>> >> > (ftrace buffer empty)
> >>> >> > Modules linked in:
> >>> >> > CPU: 0 PID: 883 Comm: kauditd Not tainted 4.10.0+ #6
> >>> >> > Hardware name: Google Google Compute Engine/Google Compute Engine,
> >>> >> > BIOS Google 01/01/2011
> >>> >> > task: ffff8801d79f0240 task.stack: ffff8801d7a20000
> >>> >> > RIP: 0010:sock_sndtimeo include/net/sock.h:2162 [inline]
> >>> >> > RIP: 0010:netlink_unicast+0xdd/0x730 net/netlink/af_netlink.c:1249
> >>> >> > RSP: 0018:ffff8801d7a27c38 EFLAGS: 00010206
> >>> >> > RAX: 0000000000000056 RBX: ffff8801d7a27cd0 RCX: 0000000000000000
> >>> >> > RDX: 0000000000000000 RSI: 0000000000000000 RDI: 00000000000002b0
> >>> >> > RBP: ffff8801d7a27cf8 R08: ffffed00385cf286 R09: ffffed00385cf286
> >>> >> > R10: 0000000000000006 R11: ffffed00385cf285 R12: 0000000000000000
> >>> >> > R13: dffffc0000000000 R14: ffff8801c2fc3c80 R15: 00000000014000c0
> >>> >> > FS: 0000000000000000(0000) GS:ffff8801dbe00000(0000) knlGS:0000000000000000
> >>> >> > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> >>> >> > CR2: 0000000020cfd000 CR3: 00000001c758f000 CR4: 00000000001406f0
> >>> >> > Call Trace:
> >>> >> > kauditd_send_unicast_skb+0x3c/0x70 kernel/audit.c:482
> >>> >> > kauditd_thread+0x174/0xb00 kernel/audit.c:599
> >>> >> > kthread+0x326/0x3f0 kernel/kthread.c:229
> >>> >> > ret_from_fork+0x31/0x40 arch/x86/entry/entry_64.S:430
> >>> >> > Code: 44 89 fe e8 56 15 ff ff 8b 8d 70 ff ff ff 49 89 c6 31 c0 85 c9
> >>> >> > 75 27 e8 b2 b2 f4 fd 49 8d bc 24 b0 02 00 00 48 89 f8 48 c1 e8 03 <42>
> >>> >> > 80 3c 28 00 0f 85 37 06 00 00 49 8b 84 24 b0 02 00 00 4c 8d
> >>> >> > RIP: sock_sndtimeo include/net/sock.h:2162 [inline] RSP: ffff8801d7a27c38
> >>> >> > RIP: netlink_unicast+0xdd/0x730 net/netlink/af_netlink.c:1249 RSP:
> >>> >> > ffff8801d7a27c38
> >>> >> > ---[ end trace ad1bba9d457430b6 ]---
> >>> >> > Kernel panic - not syncing: Fatal exception
> >>> >> >
> >>> >> >
> >>> >> > This is not reproducible and seems to be caused by an elusive race.
> >>> >> > However, looking at the code I don't see any proper protection of
> >>> >> > audit_sock (other than the if (!audit_pid) which is obviously not
> >>> >> > enough to protect against races).
> >>> >>
> >>> >> audit_cmd_mutex is supposed to protect it, I think.
> >>> >> But kauditd_send_unicast_skb() seems not holding this mutex.
> >>> >
> >>> > Hmmmm, I wonder if it makes sense to wrap most of the contents of the
> >>> > outer while loop in kauditd_thread in the audit_cmd_mutex, or around the
> >>> > first two innter while loops and the "if (auditd)" condition after the
> >>> > "quick_loop:" label. The condition on auditd is supposed to catch that
> >>> > case. We don't want it locked while playing with the scheduler at the
> >>> > bottom of that function.
> >>>
> >>> Let me look into this and play around with a few things. I suspected
> >>> there might be a problem here, so I've got thoughts on how we might
> >>> resolve it; I just need to see code them up and see what option sucks
> >>> the least.
> >>>
> >>> FWIW Richard, yes wrapping most of kauditd_thread *should* resolve
> >>> this but it's pretty heavy handed and not my first choice.
> >>
> >> That's why the inner loops made a bit more sense since it wasn't really
> >> necessary and ran afoul of the scheduler anyways.
> >
> > One of my preferred options was to get us away from protecting
> > everything with the audit_cmd_mutex by creating a new locking approach
> > for the auditd connection state (using RCU/spinlocks since it rarely
> > changes in practice) and leaving the audit_cmd_mutex for it's
> > traditional role. This should minimize the performance impact of the
> > lock and clean things up a bit. I'm also moving all the auditd
> > connection state into a single struct (instead of several variables
> > associated only by convention) which moves us oh so slightly closer to
> > allowing multiple auditd connections (hey, it's something).
> >
> > It's taking a bit longer than expected as I'm dealing with a bit of a
> > head cold (or something) and my mind is far less than 100% at the
> > moment ...
>
> Ooof. I just noticed something, and maybe this is the fever talking,
> but why do we ever NULL out audit_sock and why are we bothering with
> those holds/puts? We create the audit netlink socket in
> audit_net_init() and it should remain valid until we kill it in
> audit_next_exit(); we sorta cheat on this now because we track the
> socket both in the per-netns audit_net struct as well as audit_sock,
> but that doesn't make our audit_sock manipulations right ...

At the moment, you are right, there is no reason to null audit_sock, and
not like auditd will appear on a different sock yet.

The only excuse I can give is that this was anticipating audit daemons
in more than one user namespace necessarily with their own network
namespaces. The AUDIT_GET, AUDIT_LIST_RULES commands are treated
properly since they use the per-netns audit_net struct and don't use the
primary queue. The AUDIT_USER_* messages are converted from their
originating namespaces ok, but will need to be tracked what network
namespace they came from for multiple audit daemons in the future.

> Man I hate this code. I *really* hate this code.
>
> paul moore

- RGB

--
Richard Guy Briggs <rgb@redhat.com>
Kernel Security Engineering, Base Operating Systems, Red Hat
Remote, Ottawa, Canada
Voice: +1.647.777.2635, Internal: (81) 32635

\
 
 \ /
  Last update: 2017-03-08 14:29    [W:0.211 / U:0.328 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site