Re: [RFC] audit: avoid soft lockup in audit_log_start()

[Konstantin Khlebnikov <khlebnikov@openvz.org>]

Luiz Capitulino wrote:
> On Mon, 09 Sep 2013 19:19:14 +0400
> Konstantin Khlebnikov<khlebnikov@openvz.org>  wrote:
>
>> Luiz Capitulino wrote:
>>> On Mon, 09 Sep 2013 18:32:13 +0400
>>> Konstantin Khlebnikov<khlebnikov@openvz.org>   wrote:
>>>
>>>> Luiz Capitulino wrote:
>>>>> I'm getting the following soft lockup:
>>>>>
>>>>> CPU: 6 PID: 2278 Comm: killall5 Tainted: GF            3.11.0-rc7+ #1
>>>>> Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
>>>>>     0000000000000099 ffff88011fd83de8 ffffffff815324df 0000000000002800
>>>>>     ffffffff817d48f9 ffff88011fd83e68 ffffffff8152e669 ffff88011fd83e68
>>>>>     ffffffff00000008 ffff88011fd83e78 ffff88011fd83e18 0000004081dac040
>>>>> Call Trace:
>>>>>     <IRQ>     [<ffffffff815324df>] dump_stack+0x46/0x58
>>>>>     [<ffffffff8152e669>] panic+0xbb/0x1c4
>>>>>     [<ffffffff810d03c3>] watchdog_timer_fn+0x163/0x170
>>>>>     [<ffffffff8106c691>] __run_hrtimer+0x81/0x1c0
>>>>>     [<ffffffff810d0260>] ? watchdog+0x30/0x30
>>>>>     [<ffffffff8106cea7>] hrtimer_interrupt+0x107/0x240
>>>>>     [<ffffffff8102f61b>] local_apic_timer_interrupt+0x3b/0x60
>>>>>     [<ffffffff81542465>] smp_apic_timer_interrupt+0x45/0x60
>>>>>     [<ffffffff8154124a>] apic_timer_interrupt+0x6a/0x70
>>>>>     <EOI>     [<ffffffff810c2f5f>] ? audit_log_start+0xbf/0x430
>>>>>     [<ffffffff810c2fe7>] ? audit_log_start+0x147/0x430
>>>>>     [<ffffffff81079030>] ? try_to_wake_up+0x2a0/0x2a0
>>>>>     [<ffffffff810c86be>] audit_log_exit+0x6ae/0xc30
>>>>>     [<ffffffff81188662>] ? __alloc_fd+0x42/0x100
>>>>>     [<ffffffff810c98e7>] __audit_syscall_exit+0x257/0x2b0
>>>>>     [<ffffffff81540794>] sysret_audit+0x17/0x21
>>>>>
>>>>> The reproducer is somewhat unusual:
>>>>>
>>>>>     1. Install RHEL6.5 (maybe a similar older user-space will do)
>>>>>     2. Boot the just installed system
>>>>>     3. In this first boot you'll meet the firstboot script, which
>>>>>        will do some setup and (depending on your answers) it will
>>>>>        reboot the machine
>>>>>     4. During that first reboot the system hangs while terminating
>>>>>        all processes:
>>>>>
>>>>>           Sending all processes the TERM signal...
>>>>>
>>>>>        It's when the soft lockup above happens. And yes, I managed
>>>>>        to get this with latest upstream kernel (HEAD fa8218def1b1)
>>>>>
>>>>> I'm reproducing it on a VM, but the first report was on bare-metal.
>>>>>
>>>>> This is what is happening:
>>>>>
>>>>>     1. audit_log_start() is called
>>>>>     2. As we have SKBs waiting in audit_skb_queue and all conditions
>>>>>        evaluate to true, we sleep in wait_for_auditd()
>>>>>     3. Go to 2, until sleep_time gets negative and audit_log_start()
>>>>>        just busy-waits
>>>>>
>>>>> Now, *why* this is happening is a mistery to me. I tried debugging
>>>>> it, but all I could find is that at some point the kauditd thread
>>>>> never wakes up after having called schedule(). I even tried waking
>>>>> it up before calling wait_for_auditd(), but it didn't.
>>>>
>>>> We run into the same problem in rhel6 kernel.
>>>>
>>>> "readahead-collector" uses audit interface and sometimes stuck in 'stopped' state.
>>>
>>> Yes, please also see:
>>>
>>>    http://marc.info/?l=linux-kernel&m=137818375024600&w=2
>>>
>>>> After commit 829199197a430dade2519d54f5545c4a094393b8
>>>> (which was backported by RH into their kernel)
>>>> audit emiters will block forever if userspace daemon cannot handle backlog.
>>>> That commit just breaks timeout condition, after timeout waiting loop turns
>>>> into busy loop until deamon dies or returns back to work.
>>>>
>>>> this trivial patch should fix that problem
>>>>
>>>> --- a/kernel/audit.c
>>>> +++ b/kernel/audit.c
>>>> @@ -1215,9 +1215,10 @@ struct audit_buffer *audit_log_start(struct audit_context *ctx, gfp_t gfp_mask,
>>>>
>>>>                            sleep_time = timeout_start + audit_backlog_wait_time -
>>>>                                            jiffies;
>>>> -                       if ((long)sleep_time>   0)
>>>> +                       if ((long)sleep_time>   0) {
>>>>                                    wait_for_auditd(sleep_time);
>>>> -                       continue;
>>>> +                               continue;
>>>> +                       }
>>>>                    }
>>>
>>>
>>> Chuck Anderson posted a similar fix:
>>>
>>>    http://marc.info/?l=linux-kernel&m=137817994623832&w=2
>>>
>>> I still get a hang for around a minute with Chuck's fix, I believe I'll
>>> get the same thing with yours.
>>
>> Yep, this is normal behaviour -- default audit_backlog_wait_time is 60 seconds.
>
> Yes, I know that's the cause but I don't call it normal behavior to
> be unable to use my system during 60 seconds.

This stuff was designed for auditing, so it shouldn't loose messages too easily.
Meanwhile 'readahead' reuses it for a different purpose without proper tuning.
I think kernel should provide interface for tuning that timeout and let 'readahead'
set it to zero or some small value.

>
>> But this logic is really strange, I don't see any interface for tuning that timeout
>> and seems like kernel set it to zero after first disaster and never recovers it back.