[v2] net/ixgbe: fix blocking system events

  From: Zhu Tao <taox.zhu@intel.com>

IXGBE link status task use rte alarm thread in old implementation.
Sometime ixgbe link status task takes up to 9 seconds. This will
severely affect the rte-alarm-thread-dependent a task in the
system, like interrupt or hotplug event. So replace with a
independent thread which has the same thread affinity settings
as rte interrupt.

Fixes: 0408f47b ("net/ixgbe: fix busy polling while fiber link update")
Cc: stable@dpdk.org

Signed-off-by: Zhu Tao <taox.zhu@intel.com>
---
 drivers/net/ixgbe/ixgbe_ethdev.c | 50 +++++++++++++++++++++++++++-----
 drivers/net/ixgbe/ixgbe_ethdev.h |  2 ++
 2 files changed, 44 insertions(+), 8 deletions(-)
  

> IXGBE link status task use rte alarm thread in old implementation.
> Sometime ixgbe link status task takes up to 9 seconds. This will
> severely affect the rte-alarm-thread-dependent a task in the
> system, like interrupt or hotplug event. So replace with a
> independent thread which has the same thread affinity settings
> as rte interrupt.
> 
> Fixes: 0408f47b ("net/ixgbe: fix busy polling while fiber link update")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Zhu Tao <taox.zhu@intel.com>
> ---

Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>

> 2.17.1
  

On 01/15, taox.zhu@intel.com wrote:
>From: Zhu Tao <taox.zhu@intel.com>
>
>IXGBE link status task use rte alarm thread in old implementation.
>Sometime ixgbe link status task takes up to 9 seconds. This will
>severely affect the rte-alarm-thread-dependent a task in the

What does "rte-alarm-thread-dependent a task" mean?

Thanks,
Xiaolong

>system, like interrupt or hotplug event. So replace with a
>independent thread which has the same thread affinity settings
>as rte interrupt.
>
>Fixes: 0408f47b ("net/ixgbe: fix busy polling while fiber link update")
>Cc: stable@dpdk.org
>
>Signed-off-by: Zhu Tao <taox.zhu@intel.com>
  

On 01/15, taox.zhu@intel.com wrote:
>From: Zhu Tao <taox.zhu@intel.com>
>
>IXGBE link status task use rte alarm thread in old implementation.

s/use/uses

>Sometime ixgbe link status task takes up to 9 seconds. This will
>severely affect the rte-alarm-thread-dependent a task in the
>system, like interrupt or hotplug event. So replace with a

s/a/an

>independent thread which has the same thread affinity settings
>as rte interrupt.
>
>Fixes: 0408f47b ("net/ixgbe: fix busy polling while fiber link update")

Should be:

Fixes: 0408f47ba4d6 ("net/ixgbe: fix busy polling while fiber link update")

>Cc: stable@dpdk.org
>

Applied to dpdk-next-net-intel with Konstantin's ack, Thanks.
  

On 2/15/2020 3:41 PM, Ye Xiaolong wrote:
> On 01/15, taox.zhu@intel.com wrote:
>> From: Zhu Tao <taox.zhu@intel.com>
>>
>> IXGBE link status task use rte alarm thread in old implementation.
> 
> s/use/uses
> 
>> Sometime ixgbe link status task takes up to 9 seconds. This will
>> severely affect the rte-alarm-thread-dependent a task in the
>> system, like interrupt or hotplug event. So replace with a
> 
> s/a/an
> 
>> independent thread which has the same thread affinity settings
>> as rte interrupt.
>>
>> Fixes: 0408f47b ("net/ixgbe: fix busy polling while fiber link update")
> 
> Should be:
> 
> Fixes: 0408f47ba4d6 ("net/ixgbe: fix busy polling while fiber link update")
> 
>> Cc: stable@dpdk.org
>>
> 
> Applied to dpdk-next-net-intel with Konstantin's ack, Thanks.
> 

Shared build is failing because of missing pthread library, fixing while merging
to next-net:

 diff --git a/drivers/net/ixgbe/Makefile b/drivers/net/ixgbe/Makefile
 index 8b9d51f75..aec56a680 100644
 --- a/drivers/net/ixgbe/Makefile
 +++ b/drivers/net/ixgbe/Makefile
 @@ -57,6 +57,7 @@ endif
  LDLIBS += -lrte_eal -lrte_mbuf -lrte_mempool -lrte_ring
  LDLIBS += -lrte_ethdev -lrte_net -lrte_kvargs -lrte_hash
  LDLIBS += -lrte_bus_pci
 +LDLIBS += -lpthread

  #
  # Add extra flags for base driver files (also known as shared code)
  

17/02/2020 14:01, Ferruh Yigit:
> On 2/15/2020 3:41 PM, Ye Xiaolong wrote:
> > On 01/15, taox.zhu@intel.com wrote:
> >> From: Zhu Tao <taox.zhu@intel.com>
> >>
> >> IXGBE link status task use rte alarm thread in old implementation.
> > 
> > s/use/uses
> > 
> >> Sometime ixgbe link status task takes up to 9 seconds. This will
> >> severely affect the rte-alarm-thread-dependent a task in the
> >> system, like interrupt or hotplug event. So replace with a
> > 
> > s/a/an
> > 
> >> independent thread which has the same thread affinity settings
> >> as rte interrupt.
> >>
> >> Fixes: 0408f47b ("net/ixgbe: fix busy polling while fiber link update")
> > 
> > Should be:
> > 
> > Fixes: 0408f47ba4d6 ("net/ixgbe: fix busy polling while fiber link update")
> > 
> >> Cc: stable@dpdk.org
> >>
> > 
> > Applied to dpdk-next-net-intel with Konstantin's ack, Thanks.
> > 
> 
> Shared build is failing because of missing pthread library, fixing while merging
> to next-net:

One more thing looks strange in this patch:
	ixgbe_dev_cannel_link_thread
Should it be
	ixgbe_dev_cancel_link_thread
?

Note: I looked at it because I am not sure multiplying the interrupt
threads is a good idea.
Basically the link status management is too long in this driver.
Instead of fixing the root cause, you move the annoying workload
somewhere else. But it is still there...

Please could you work on a long term fix?
  

Hi Thomas
 
Thank you for your correction spelling error of 'cancel'.

Indeed it is not the best solution by creating a thread. I refer to the same solution with Linux kernel driver. Linux kernel driver manages link status by using a thread. Maybe we can figure out another better solution to fix this problem but it may take much more time. At this time, 20.02 formal release is coming and this problem affect some basic library. 
Tks for your understanding.


BR,
Zhu, Tao


> -----Original Message-----
> From: Thomas Monjalon [mailto:thomas@monjalon.net]
> Sent: Thursday, February 20, 2020 11:37 PM
> To: Ye, Xiaolong <xiaolong.ye@intel.com>; Zhu, TaoX <taox.zhu@intel.com>;
> Yigit, Ferruh <ferruh.yigit@intel.com>
> Cc: dev@dpdk.org; Ananyev, Konstantin <konstantin.ananyev@intel.com>;
> Lu, Wenzhuo <wenzhuo.lu@intel.com>; stable@dpdk.org
> Subject: Re: [dpdk-dev] [PATCH v2] net/ixgbe: fix blocking system events
> 
> 17/02/2020 14:01, Ferruh Yigit:
> > On 2/15/2020 3:41 PM, Ye Xiaolong wrote:
> > > On 01/15, taox.zhu@intel.com wrote:
> > >> From: Zhu Tao <taox.zhu@intel.com>
> > >>
> > >> IXGBE link status task use rte alarm thread in old implementation.
> > >
> > > s/use/uses
> > >
> > >> Sometime ixgbe link status task takes up to 9 seconds. This will
> > >> severely affect the rte-alarm-thread-dependent a task in the
> > >> system, like interrupt or hotplug event. So replace with a
> > >
> > > s/a/an
> > >
> > >> independent thread which has the same thread affinity settings as
> > >> rte interrupt.
> > >>
> > >> Fixes: 0408f47b ("net/ixgbe: fix busy polling while fiber link
> > >> update")
> > >
> > > Should be:
> > >
> > > Fixes: 0408f47ba4d6 ("net/ixgbe: fix busy polling while fiber link
> > > update")
> > >
> > >> Cc: stable@dpdk.org
> > >>
> > >
> > > Applied to dpdk-next-net-intel with Konstantin's ack, Thanks.
> > >
> >
> > Shared build is failing because of missing pthread library, fixing
> > while merging to next-net:
> 
> One more thing looks strange in this patch:
> 	ixgbe_dev_cannel_link_thread
> Should it be
> 	ixgbe_dev_cancel_link_thread
> ?
> 
> Note: I looked at it because I am not sure multiplying the interrupt threads is
> a good idea.
> Basically the link status management is too long in this driver.
> Instead of fixing the root cause, you move the annoying workload
> somewhere else. But it is still there...
> 
> Please could you work on a long term fix?
>
  

21/02/2020 09:19, Zhu, TaoX:
> Hi Thomas
>  
> Thank you for your correction spelling error of 'cancel'.
> 
> Indeed it is not the best solution by creating a thread. I refer to the same solution with Linux kernel driver. Linux kernel driver manages link status by using a thread. Maybe we can figure out another better solution to fix this problem but it may take much more time. At this time, 20.02 formal release is coming and this problem affect some basic library. 
> Tks for your understanding.

I understand, that's why I already accepted this patch in mainline (while fixing typo).

Please it would be really appreciated to work on a better solution.



> > -----Original Message-----
> > From: Thomas Monjalon [mailto:thomas@monjalon.net]
> > Sent: Thursday, February 20, 2020 11:37 PM
> > To: Ye, Xiaolong <xiaolong.ye@intel.com>; Zhu, TaoX <taox.zhu@intel.com>;
> > Yigit, Ferruh <ferruh.yigit@intel.com>
> > Cc: dev@dpdk.org; Ananyev, Konstantin <konstantin.ananyev@intel.com>;
> > Lu, Wenzhuo <wenzhuo.lu@intel.com>; stable@dpdk.org
> > Subject: Re: [dpdk-dev] [PATCH v2] net/ixgbe: fix blocking system events
> > 
> > 17/02/2020 14:01, Ferruh Yigit:
> > > On 2/15/2020 3:41 PM, Ye Xiaolong wrote:
> > > > On 01/15, taox.zhu@intel.com wrote:
> > > >> From: Zhu Tao <taox.zhu@intel.com>
> > > >>
> > > >> IXGBE link status task use rte alarm thread in old implementation.
> > > >
> > > > s/use/uses
> > > >
> > > >> Sometime ixgbe link status task takes up to 9 seconds. This will
> > > >> severely affect the rte-alarm-thread-dependent a task in the
> > > >> system, like interrupt or hotplug event. So replace with a
> > > >
> > > > s/a/an
> > > >
> > > >> independent thread which has the same thread affinity settings as
> > > >> rte interrupt.
> > > >>
> > > >> Fixes: 0408f47b ("net/ixgbe: fix busy polling while fiber link
> > > >> update")
> > > >
> > > > Should be:
> > > >
> > > > Fixes: 0408f47ba4d6 ("net/ixgbe: fix busy polling while fiber link
> > > > update")
> > > >
> > > >> Cc: stable@dpdk.org
> > > >>
> > > >
> > > > Applied to dpdk-next-net-intel with Konstantin's ack, Thanks.
> > > >
> > >
> > > Shared build is failing because of missing pthread library, fixing
> > > while merging to next-net:
> > 
> > One more thing looks strange in this patch:
> > 	ixgbe_dev_cannel_link_thread
> > Should it be
> > 	ixgbe_dev_cancel_link_thread
> > ?
> > 
> > Note: I looked at it because I am not sure multiplying the interrupt threads is
> > a good idea.
> > Basically the link status management is too long in this driver.
> > Instead of fixing the root cause, you move the annoying workload
> > somewhere else. But it is still there...
> > 
> > Please could you work on a long term fix?
> > 
> 
>