Re: BUG in find_pid_ns

From: Eric W. Biederman
Date: Mon Feb 18 2013 - 23:38:30 EST


Sasha Levin <sasha.levin@xxxxxxxxxx> writes:

> On 02/17/2013 07:17 PM, ebiederm@xxxxxxxxxxxx wrote:
>> The bad pointer value is 0xfffffffffffffff0. Hmm.
>>
>> If you have the failure location correct it looks like a corrupted hash
>> entry was found while following the hash chain.
>>
>> It looks like the memory has been set to -16 -EBUSY? Weird.
>>
>> It smells like something is stomping on the memory of a struct pid, with
>> the same hash value and thus in the same hash chain as the current pid.
>>
>> Can you reproduce this?
>
> I've just reproduced it again:
>
> [ 2404.518957] BUG: unable to handle kernel paging request at fffffffffffffff0
> [ 2404.520024] IP: [<ffffffff81131d50>] find_pid_ns+0x110/0x1f0
> [ 2404.520024] PGD 5429067 PUD 542b067 PMD 0
> [ 2404.520024] Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
> [ 2404.520024] Dumping ftrace buffer:
> [ 2404.520024] (ftrace buffer empty)
> [ 2404.520024] Modules linked in:
> [ 2404.520024] CPU 3
> [ 2404.520024] Pid: 6890, comm: trinity Tainted: G W 3.8.0-rc7-next-20130215-sasha-00027-gb399f44-dirty #288
> [ 2404.520024] RIP: 0010:[<ffffffff81131d50>] [<ffffffff81131d50>] find_pid_ns+0x110/0x1f0
> [ 2404.520024] RSP: 0018:ffff8800af1dfe18 EFLAGS: 00010286
> [ 2404.520024] RAX: 0000000000000001 RBX: 0000000000004b72 RCX: 0000000000000000
> [ 2404.520024] RDX: 0000000000000001 RSI: ffffffff85466e40 RDI: 0000000000000286
> [ 2404.520024] RBP: ffff8800af1dfe48 R08: 0000000000000001 R09: 0000000000000001
> [ 2404.520024] R10: 0000000000000001 R11: 0000000000000000 R12: ffffffff85466460
> [ 2404.520024] R13: ffff8800bf8d3ef8 R14: fffffffffffffff0 R15: ffff8800a43d9a40
> [ 2404.520024] FS: 00007f8300f79700(0000) GS:ffff8800bbc00000(0000) knlGS:0000000000000000
> [ 2404.520024] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [ 2404.520024] CR2: fffffffffffffff0 CR3: 00000000af0b7000 CR4: 00000000000406e0
> [ 2404.520024] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> [ 2404.520024] DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
> [ 2404.520024] Process trinity (pid: 6890, threadinfo ffff8800af1de000, task ffff8800b060b000)
> [ 2404.520024] Stack:
> [ 2404.520024] ffffffff85466e40 0000000000004b72 ffff8800af1dfed8 0000000000000000
> [ 2404.520024] 0000000000000003 20c49ba5e353f7cf ffff8800af1dfe58 ffffffff81131e5c
> [ 2404.520024] ffff8800af1dfec8 ffffffff8112400f ffffffff81123f9c 0000000000000000
> [ 2404.520024] Call Trace:
> [ 2404.520024] [<ffffffff81131e5c>] find_vpid+0x2c/0x30
> [ 2404.520024] [<ffffffff8112400f>] kill_something_info+0x9f/0x270
> [ 2404.673395] [<ffffffff81123f9c>] ? kill_something_info+0x2c/0x270
> [ 2404.673395] [<ffffffff81125e38>] sys_kill+0x88/0xa0
> [ 2404.673395] [<ffffffff8107ad34>] ? syscall_trace_enter+0x24/0x2e0
> [ 2404.694324] [<ffffffff811813b8>] ? trace_hardirqs_on_caller+0x128/0x160
> [ 2404.694324] [<ffffffff83d96275>] ? tracesys+0x7e/0xe6
> [ 2404.694324] [<ffffffff83d962d8>] tracesys+0xe1/0xe6
> [ 2404.694324] Code: 4d 8b 75 00 e8 b2 0e 00 00 85 c0 0f 84 d2 00 00 00 80 3d fa 17 d5 04 00 0f 85 c5 00 00 00 e9 93 00 00 00 0f
> 1f 84 00 00 00 00 00 <41> 39 1e 75 2b 4d 39 66 08 75 25 41 8b 84 24 20 08 00 00 48 c1
> [ 2404.733487] RIP [<ffffffff81131d50>] find_pid_ns+0x110/0x1f0
> [ 2404.740299] RSP <ffff8800af1dfe18>
> [ 2404.740299] CR2: fffffffffffffff0
> [ 2404.740299] ---[ end trace 9f8bc22bbe4fe990 ]---
>
> I'm not sure what debug info I could throw in which will be helpful. Dump
> the entire chain or table if 'pnr' happens to look odd?
>
>> Memory corruption is hard to trace down with just a single data point.
>>
>> Looking a little closer Sasha you have rewritten
>> hlist_for_each_entry_rcu, and that seems to be the most recent patch
>> dealing with pids, and we are failing in hlist_for_each_entry_rcu.
>>
>> I haven't looked at your patch in enough detail to know if you have
>> missed something or not, but a brand new patch and a brand new failure
>> certainly look suspicious at first glance.
>
> Agreed, I've also took a second look at it when this BUG popped up. What
> surprises me about it is that if the new iteration is broken, the kernel
> would spectacularly break in a bunch of places instead of failing in the
> exact same place twice.
>
> Not ignoring the possibility it's broken though.

I don't see any obvious problems with your code however. struct upid is:

struct upid {
/* Try to keep pid_chain in the same cacheline as nr for find_vpid */
int nr;
struct pid_namespace *ns;
struct hlist_node pid_chain;
};

Which puts pid_chain at offset 16. nr is at offset 0 and is the first
field we access.

Trying to access a upid at address fffffffffffffff0 looks for all the
world like hlist_entry_safe decremented a NULL pointer by 16 bytebs.

I suggest you take a look at the assembly. Upon occassion gcc
optmizations get smart and optimize out things we would prefer they
kept.

As for the rest perhaps there is something odd about the optimization
of pid_nr_ns or perhaps kill_something_info is the only frequent test
where we try to acess something that does not exist.

Eric

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/