11.0 stuck on high network load

[Julien Charbon]

 Hi Slawa,

On 9/25/16 2:46 PM, Slawa Olhovchenkov wrote:
> On Fri, Sep 23, 2016 at 11:01:43PM +0300, Slawa Olhovchenkov wrote:
>> On Wed, Sep 21, 2016 at 11:25:18PM +0200, Julien Charbon wrote:
>>>
>>> On 9/21/16 9:51 PM, Slawa Olhovchenkov wrote:
>>>> On Wed, Sep 21, 2016 at 09:11:24AM +0200, Julien Charbon wrote:
>>>>>  You can also use Dtrace and lockstat (especially with the lockstat -s
>>>>> option):
>>>>>
>>>>> https://wiki.freebsd.org/DTrace/One-Liners#Kernel_Locks
>>>>> https://www.freebsd.org/cgi/man.cgi?query=lockstat&manpath=FreeBSD+11.0-RELEASE
>>>>>
>>>>>  But I am less familiar with Dtrace/lockstat tools.
>>>>
>>>> I am still use old kernel and got lockdown again.
>>>> Try using lockstat (I am save more output), interesting may be next:
>>>>
>>>> R/W writer spin on writer: 190019 events in 1.070 seconds (177571 events/sec)
>>>>
>>>> -------------------------------------------------------------------------------
>>>> Count indv cuml rcnt     nsec Lock                   Caller                  
>>>> 140839  74%  74% 0.00    24659 tcpinp                 tcp_tw_2msl_scan+0xc6   
>>>>
>>>>       nsec ------ Time Distribution ------ count     Stack                   
>>>>       4096 |                               913       tcp_twstart+0xa3        
>>>>       8192 |@@@@@@@@@@@@                   58191     tcp_do_segment+0x201f   
>>>>      16384 |@@@@@@                         29594     tcp_input+0xe1c         
>>>>      32768 |@@@@                           23447     ip_input+0x15f          
>>>>      65536 |@@@                            16197     
>>>>     131072 |@                              8674      
>>>>     262144 |                               3358      
>>>>     524288 |                               456       
>>>>    1048576 |                               9         
>>>> -------------------------------------------------------------------------------
>>>> Count indv cuml rcnt     nsec Lock                   Caller                  
>>>> 49180  26% 100% 0.00    15929 tcpinp                 tcp_tw_2msl_scan+0xc6   
>>>>
>>>>       nsec ------ Time Distribution ------ count     Stack                   
>>>>       4096 |                               157       pfslowtimo+0x54         
>>>>       8192 |@@@@@@@@@@@@@@@                24796     softclock_call_cc+0x179 
>>>>      16384 |@@@@@@                         11223     softclock+0x44          
>>>>      32768 |@@@@                           7426      intr_event_execute_handlers+0x95
>>>>      65536 |@@                             3918      
>>>>     131072 |                               1363      
>>>>     262144 |                               278       
>>>>     524288 |                               19        
>>>> -------------------------------------------------------------------------------
>>>
>>>  This is interesting, it seems that you have two call paths competing
>>> for INP locks here:
>>>
>>>  - pfslowtimo()/tcp_tw_2msl_scan(reuse=0) and
>>>
>>>  - tcp_input()/tcp_twstart()/tcp_tw_2msl_scan(reuse=1)
>>
>> My current hypothesis:
>>
>> nginx do write() (or may be close()?) to socket, kernel lock
>> first inp in V_twq_2msl, happen callout for pfslowtimo() on the same
>> CPU core and tcp_tw_2msl_scan infinity locked on same inp.
>>
>> In this case you modification can't help, before next try we need some
>> like yeld().
> 
> Or may be locks leaks.
> Or both.

 Actually one extra debug thing you can do is launching lockstat with
below extra options:

 -H     For Hold lock stats
 -P     To get the overall time
 -s 20  To get the stackstrace

 To see who is holding the INP lock for so long.  Thanks to Hiren for
pointing the -H option.

--
Julien