Message-id: <4B54D50C.90608@majjas.com>
Date:	Mon, 18 Jan 2010 16:39:24 -0500
From:	Michael Breuer <mbreuer@...jas.com>
To:	Jarek Poplawski <jarkao2@...il.com>
Cc:	Stephen Hemminger <shemminger@...ux-foundation.org>,
	David Miller <davem@...emloft.net>, akpm@...ux-foundation.org,
	flyboy@...il.com, linux-kernel@...r.kernel.org,
	netdev@...r.kernel.org
Subject: Re: [PATCH] af_packet: Don't use skb after dev_queue_xmit()

On 1/18/2010 4:25 PM, Jarek Poplawski wrote:
> On Mon, Jan 18, 2010 at 03:56:45PM -0500, Michael Breuer wrote:
>    
>> On 1/18/2010 3:46 PM, Jarek Poplawski wrote:
>>      
>>> On Mon, Jan 18, 2010 at 11:29:31AM -0500, Michael Breuer wrote:
>>>        
>>>> Ok - up on the two patches, no DMAR. Some early observations:
>>>>
>>>> 1. There's an early on MMAP oops (see below). This happens once, at
>>>> the completion of the transition to runlevel 5 (I've seen it
>>>> entering runlevel 3 as well). This does not recur when runlevels are
>>>> subsequently changed. I do not see this when running with DMAR
>>>> enabled.
>>>>          
>>> OK, you mentioned this oops (actually a warning only) happened during
>>> previous tests too.
>>>        
>> Yes - dk if it's significant or not. Only obvious difference between
>> DMAR and not.
>>      
> OK, let's try (as long as possible) if it can break so hard as with
> DMAR.
>
>    
>>>> 2. The dropped tx packet (DHCP) is a bit harder to recreate, but it
>>>> still happens.
>>>>          
>>> Btw, I guess you improved the test because you didn't mention it here,
>>> even after my explicit question?:
>>> http://permalink.gmane.org/gmane.linux.network/149171
>>>        
>> I had been focusing on the hangs - dhcp causing the initial crash
>> from December. After things stabilized with the af patch&  skb may
>> pull I started noticing the dropped tx packets. I reported the TX
>> loss on the 16th of January after confirming the issue.
>>      
> OK, but we need to establish some status quo after these patches
> before any new things (including DMAR), so I'd suggest trying this
> config really longer and harder.
>
>    
>>>> Interestingly, I initially saw no dropped packets
>>>> with ping - but after I went the DCHP route and eventually
>>>> reconnected, I could then cause dropped tx packets with ping. To
>>>> clarify:
>>>>
>>>> a) start throughput
>>>> b) ping device - no packet loss - this was true for the entire test run.
>>>> c) start throughput again
>>>> d) ping - no loss.
>>>> e) drop wifi on the device&   restart - first attempt worked. Repeat
>>>> attempt yielded the dropped DHCPOFFER packets. After about 6 tries,
>>>> the device reconnected to wifi.
>>>> f) ping again (after the reconnection) - packet loss rate about 80%.
>>>> g) simultaneously ping the wifi router - no loss.
>>>> h) After a while, packets are no longer dropped during ping. If I
>>>> manage to cause the dhcp drop again, and then ping after the device
>>>> finally reconnects, packet loss is significant for a while (maybe 30
>>>> sec to a minute). Then things return to normal. Note that the packet
>>>> loss continues even if the reported throughput drops to nil.
>>>> i) I can't cause the initial packet loss at RX rates below about
>>>> 30,000KBPS (as reported by nethogs).  At rates over 40 I can
>>>> reproduce this on this set of patches&   config about 60% of the
>>>> time.
>>>>          
>>> I forgot to mention, but did you try to check if these lost ping
>>> packets are "being dropped somewhere after wireshark sees them and
>>> before hitting the wire" like DHCPOFFER? Aren't there any sky2
>>> warnings/resets while this happens?
>>>
>>> Jarek P.
>>>        
>> Yes. There are no errors, and no statistics anywhere that I know to
>> look reflect the loss. Nothing in netstat; ethtool -S; etc. The only
>> loss reported is RX. The recent TX warnings/resets happened while
>> the machine was up for several days and while unattended and under
>> high RX load.
>>      
> Please check "tc -s qdisc" each time as well.
>
> Jarek P
>    

Some output from tc -s qdisc:

Before test:
qdisc pfifo_fast 0: dev eth0 root refcnt 2 bands 3 priomap  1 2 2 2 1 2 
0 0 1 1 1 1 1 1 1 1
  Sent 35279532 bytes 291080 pkt (dropped 0, overlimits 0 requeues 0)
  rate 0bit 0pps backlog 0b 0p requeues 0
qdisc pfifo_fast 0: dev eth1 root refcnt 2 bands 3 priomap  1 2 2 2 1 2 
0 0 1 1 1 1 1 1 1 1
  Sent 377308 bytes 3107 pkt (dropped 0, overlimits 0 requeues 0)
  rate 0bit 0pps backlog 0b 0p requeues 0

During test (after initial observed packet loss):
qdisc pfifo_fast 0: dev eth0 root refcnt 2 bands 3 priomap  1 2 2 2 1 2 
0 0 1 1 1 1 1 1 1 1
  Sent 123389424 bytes 1781403 pkt (dropped 0, overlimits 0 requeues 0)
  rate 0bit 0pps backlog 0b 0p requeues 0
qdisc pfifo_fast 0: dev eth1 root refcnt 2 bands 3 priomap  1 2 2 2 1 2 
0 0 1 1 1 1 1 1 1 1
  Sent 400862 bytes 3250 pkt (dropped 0, overlimits 0 requeues 0)
  rate 0bit 0pps backlog 0b 0p requeues 0

During test - while packet loss occuring:
qdisc pfifo_fast 0: dev eth0 root refcnt 2 bands 3 priomap  1 2 2 2 1 2 
0 0 1 1 1 1 1 1 1 1
  Sent 150518974 bytes 2138312 pkt (dropped 0, overlimits 0 requeues 0)
  rate 0bit 0pps backlog 0b 0p requeues 0
qdisc pfifo_fast 0: dev eth1 root refcnt 2 bands 3 priomap  1 2 2 2 1 2 
0 0 1 1 1 1 1 1 1 1
  Sent 422003 bytes 3432 pkt (dropped 0, overlimits 0 requeues 0)
  rate 0bit 0pps backlog 0b 0p requeues 0

After the conclusion of the test:
qdisc pfifo_fast 0: dev eth0 root refcnt 2 bands 3 priomap  1 2 2 2 1 2 
0 0 1 1 1 1 1 1 1 1
  Sent 244900497 bytes 3416350 pkt (dropped 0, overlimits 0 requeues 0)
  rate 0bit 0pps backlog 0b 0p requeues 0
qdisc pfifo_fast 0: dev eth1 root refcnt 2 bands 3 priomap  1 2 2 2 1 2 
0 0 1 1 1 1 1 1 1 1
  Sent 564380 bytes 4708 pkt (dropped 0, overlimits 0 requeues 0)
  rate 0bit 0pps backlog 0b 0p requeues 0


During the test, 8.9GB received; 232.9MB sent).

I also connected a second device through the wifi router. I was able to 
ping that device w/o loss while DHCP packets were being dropped to the 
other connected device.

Last note: just moved to 2.6.32.4 from .3 for this test (from git).

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists