Date:	Thu, 7 Jan 2016 11:33:36 +0800
From:	zhuyj <zyjzyj2000@...il.com>
To:	"Tantilov, Emil S" <emil.s.tantilov@...el.com>,
	Michal Kubecek <mkubecek@...e.cz>,
	Jay Vosburgh <jay.vosburgh@...onical.com>
Cc:	"vfalico@...il.com" <vfalico@...il.com>,
	"gospo@...ulusnetworks.com" <gospo@...ulusnetworks.com>,
	"netdev@...r.kernel.org" <netdev@...r.kernel.org>,
	"Shteinbock, Boris (Wind River)" <boris.shteinbock@...driver.com>
Subject: Re: [PATCH 1/1] bonding: restrict up state in 802.3ad mode

On 01/07/2016 10:43 AM, Tantilov, Emil S wrote:
>> -----Original Message-----
>> From: zhuyj [mailto:zyjzyj2000@...il.com]
>> Sent: Tuesday, January 05, 2016 7:05 PM
>> To: Tantilov, Emil S; Michal Kubecek; Jay Vosburgh
>> Cc: vfalico@...il.com; gospo@...ulusnetworks.com; netdev@...r.kernel.org;
>> Shteinbock, Boris (Wind River)
>> Subject: Re: [PATCH 1/1] bonding: restrict up state in 802.3ad mode
>>
>> On 01/06/2016 09:26 AM, Tantilov, Emil S wrote:
>>>> -----Original Message-----
>>>> From: netdev-owner@...r.kernel.org [mailto:netdev-owner@...r.kernel.org]
>> On
>>>> Behalf Of zhuyj
>>>> Sent: Monday, December 28, 2015 1:19 AM
>>>> To: Michal Kubecek; Jay Vosburgh
>>>> Cc: vfalico@...il.com; gospo@...ulusnetworks.com;
>> netdev@...r.kernel.org;
>>>> Shteinbock, Boris (Wind River)
>>>> Subject: Re: [PATCH 1/1] bonding: restrict up state in 802.3ad mode
>>>>
>>>> On 12/28/2015 04:43 PM, Michal Kubecek wrote:
>>>>> On Thu, Dec 17, 2015 at 01:57:16PM -0800, Jay Vosburgh wrote:
>>>>>> <zyjzyj2000@...il.com> wrote:
>>>>>>> In 802.3ad mode, the speed and duplex is needed. But in some NIC,
>>>>>>> there is a time span between NIC up state and getting speed and
>> duplex.
>>>>>>> As such, sometimes a slave in 802.3ad mode is in up state without
>>>>>>> speed and duplex. This will make bonding in 802.3ad mode can not
>>>>>>> work well.
>>>>>>> To make bonding driver be compatible with more NICs, it is
>>>>>>> necessary to restrict the up state in 802.3ad mode.
>>>>>> 	What device is this?  It seems a bit odd that an Ethernet device
>>>>>> can be carrier up but not have the duplex and speed available.
>>>>> ...
>>>>>> 	In general, though, bonding expects a speed or duplex change to
>>>>>> be announced via a NETDEV_UPDATE or NETDEV_UP notifier, which would
>>>>>> propagate to the 802.3ad logic.
>>>>>>
>>>>>> 	If the device here is going carrier up prior to having speed or
>>>>>> duplex available, then maybe it should call netdev_state_change() when
>>>>>> the duplex and speed are available, or delay calling
>> netif_carrier_on().
>>>>> I have encountered this problem (NIC having carrier on before being
>> able
>>>>> to detect speed/duplex and driver not notifying when speed/duplex
>>>>> becomes available) with netxen cards earlier. But it was eventually
>>>>> fixed in the driver by commit 9d01412ae76f ("netxen: Fix link event
>>>>> handling.") so this example rather supports what you said.
>>>>>
>>>>>                                                              Michal
>> Kubecek
>>>> Thanks a lot.
>>>> I checked the commit 9d01412ae76f ("netxen: Fix link event
>>>> handling."). The symptoms are the same with mine.
>>>>
>>>> The root cause is different. In my problem, the root cause is that LINKS
>>>> register[]  can not provide link_up and link_speed at the same time.
>>>> There is a time span between link_up and link_speed.
>>> The LINK_UP and LINK_SPEED bits in the LINKS register for ixgbe HW are
>> updated
>>> simultaneously. Do you have any proof to show the delay you are referring
>> to
>>> as I am sure our HW engineers would like to know about it.
>> Sorry. I can not reproduce this problem locally. What I have is the
>> feedback from the customer.
> So you are assuming that there is a delay due to the issue you are seeing?

Sure. Before I get the further feedback from the customer, I can not 
make further conclusion.
My patch is based on the feedback from the customer.

>
>> Settings for eth0:
>>     Supported ports: [ TP ]
>>     Supported link modes:   100baseT/Full
>>                             1000baseT/Full
>>                             10000baseT/Full
>>     Supported pause frame use: No
>>     Supports auto-negotiation: Yes
>>     Advertised link modes:  100baseT/Full
>>                             1000baseT/Full
>>                             10000baseT/Full
>>     Advertised pause frame use: No
>>     Advertised auto-negotiation: Yes
>>     Speed: Unknown!
>>     Duplex: Unknown! (255)
>>     Port: Twisted Pair
>>     PHYAD: 0
>>     Transceiver: external
>>     Auto-negotiation: on
>>     MDI-X: Unknown
>>     Supports Wake-on: d
>>     Wake-on: d
>>     Current message level: 0x00000007 (7)
>>                    drv probe link
>>     Link detected: yes
> The speed and the link state here are reported from
> different sources:
Sure. 
ixgbe_get_settings->hw->mac.ops.check_link(X540)->ixgbe_check_mac_link_generic
In this function ixgbe_check_mac_link_generic, the register IXGBE_LINKS 
is checked. link_up and
link_speed is got from this register.

>
>>     Link detected: yes
> Comes from a netif_carrier_ok() check. This is done via ethtool_op_get_link()
>
> Only the speed is reported through the LINKS register - that is why it is reported
> as "Unknown" - in other words link_up is false.
Sorry. I do not agree with you.

static inline bool netif_carrier_ok(const struct net_device *dev)
{
         return !test_bit(__LINK_STATE_NOCARRIER, &dev->state);
}

netif_carrier_ok will check __LINK_STATE_NOCARRIER. This 
__LINK_STATE_NOCARRIER is set by netif_carrier_on.

/**
  *      netif_carrier_on - set carrier
  *      @dev: network device
  *
  * Device has detected that carrier.
  */
void netif_carrier_on(struct net_device *dev)
{
         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
                 if (dev->reg_state == NETREG_UNINITIALIZED)
                         return;
                 atomic_inc(&dev->carrier_changes);
                 linkwatch_fire_event(dev);
                 if (netif_running(dev))
                         __netdev_watchdog_up(dev);
         }
}

In ixgbe driver, in ixgbe_main.c +6506, this function 
ixgbe_watchdog_link_is_up runs
netif_carrier_on function.

ixgbe_watchdog_link_is_up is in service_task. If 
IXGBE_FLAG_NEED_LINK_UPDATE is set in adapter->flags,
the function ixgbe_watchdog_link_is_up will run every 100ms.

IXGBE_FLAG_NEED_LINK_UPDATE is set in ixgbe_check_lsc in x540. This 
function ixgbe_check_lsc is in irq handler.
link_up will trigger it.

As such, link_up will trriger ixgbe_check_lsc to set 
IXGBE_FLAG_NEED_LINK_UPDATE in adapter->flags. In the end,
service_task will check the register IXGBE_LINKS every 100ms.

So ixgbe_get_settings and netif_carrier_ok travel different paths to the 
function ixgbe_check_mac_link_generic.
And the time span between ixgbe_get_settings and netif_carrier_ok is 
very tiny, about 100ms. So we can treat it simultaneous.

>
> This is a trace from the case where the bonding driver reports 0 Mbps:
>
>     kworker/u48:1-27950 [010] ....  6493.084916: ixgbe_service_task: eth1: link_speed = 80, link_up = false
>     kworker/u48:1-27950 [011] ....  6493.184894: ixgbe_service_task: eth1: link_speed = 80, link_up = false
>     kworker/u48:1-27950 [000] ....  6494.439883: ixgbe_service_task: eth1: link_speed = 80, link_up = true
>     kworker/u48:1-27950 [000] ....  6494.464204: ixgbe_service_task: eth1: NIC Link is Up 10 Gbps, Flow Control: RX/TX
>       kworker/0:2-1926  [000] ....  6494.464249: ixgbe_get_settings: eth1: link_speed = 80, link_up = false
>    NetworkManager-3819  [008] ....  6494.464484: ixgbe_get_settings: eth1: link_speed = 80, link_up = false
>     kworker/u48:1-27950 [007] ....  6494.496886: bond_mii_monitor: bond0: link status definitely up for interface eth1, 0 Mbps full duplex
>    NetworkManager-3819  [008] ....  6494.496967: ixgbe_get_settings: eth1: link_speed = 80, link_up = false
>     kworker/u48:1-27950 [008] ....  6495.288798: ixgbe_service_task: eth1: link_speed = 80, link_up = true
>     kworker/u48:1-27950 [008] ....  6495.388806: ixgbe_service_task: eth1: link_speed = 80, link_up = true
>
> As you can see the link is initially established, but then lost and if just so happens that the
> bonding driver is checking it at that time it will report 0 Mbps.
Thanks for your reply. I will delve into the source code.

Best Regards!
Zhu Yanjun
>
> I will give your patch a try and see if it helps in this situation.
>
> Thanks,
> Emil
>

--
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