Yeah I got the clear picture. Eventually MSIX can be routed to worker cores 
which can wake on eventfd, hence unix-epoll-input node required.

Thanks,
Nitin

From: Damjan Marion <dmar...@me.com>
Sent: Wednesday, February 5, 2020 3:08 PM
To: Nitin Saxena <nsax...@marvell.com>
Cc: vpp-dev@lists.fd.io
Subject: Re: [EXT] [vpp-dev] Regarding unix_epoll_input on worker cores


Yes, but you keep interrupts disabled while in polling mode. At the point when 
traffic load goes down, you can switch to interrupt mode by switching MSI-X 
interrupts on.
We already have that functionality today, only issue is that interrupts are 
routed to main thread, so it may happen that worker is sleeping and that will 
cause delay in taking packet from the rx queue handled by that worker...


On 5 Feb 2020, at 10:24, Nitin Saxena 
<nsax...@marvell.com<mailto:nsax...@marvell.com>> wrote:

>> No, there is no need to change default state. State can be changed during 
>> runtime...
That part I understand. I realized that AVF supports MSIX interrupts hence 
native plugin can easily leverage this feature (provided hardware supports MSIX 
per RQ)

Thanks,
Nitin

From: vpp-dev@lists.fd.io<mailto:vpp-dev@lists.fd.io> 
<vpp-dev@lists.fd.io<mailto:vpp-dev@lists.fd.io>> On Behalf Of Damjan Marion 
via Lists.Fd.Io
Sent: Wednesday, February 5, 2020 2:32 PM
To: Nitin Saxena <nsax...@marvell.com<mailto:nsax...@marvell.com>>
Cc: vpp-dev@lists.fd.io<mailto:vpp-dev@lists.fd.io>
Subject: Re: [EXT] Re: [vpp-dev] Regarding unix_epoll_input on worker cores

No, there is no need to change default state. State can be changed during 
runtime...
—
Damjan


On 5 Feb 2020, at 09:54, Nitin Saxena 
<nsax...@marvell.com<mailto:nsax...@marvell.com>> wrote:

Thanks Damjan.

Then I think there will be plan to change default state of INPUT nodes from 
POLLING to INTERRUPTIBLE. I think it is required to enable this feature at some 
point.

Thanks,
Nitin

From: Damjan Marion <dmar...@me.com<mailto:dmar...@me.com>>
Sent: Wednesday, February 5, 2020 2:00 PM
To: Nitin Saxena <nsax...@marvell.com<mailto:nsax...@marvell.com>>
Cc: vpp-dev <vpp-dev@lists.fd.io<mailto:vpp-dev@lists.fd.io>>
Subject: [EXT] Re: [vpp-dev] Regarding unix_epoll_input on worker cores

External Email
________________________________
There are 2 use cases:

- delivering RX queue interrupts directpy to the worker which handles specific 
queue
- waking up worker thread over eventfd

None of them is used today, but it should be at some point...

--
DAmjan




On 5 Feb 2020, at 08:39, Nitin Saxena 
<nsax...@marvell.com<mailto:nsax...@marvell.com>> wrote:

Hi,

I understand the purpose of unix-epoll-input PRE_INPUT node on main thread but 
I am not able to envision why it is required on worker core.
The way I see it all clib_file_xx() operations are on main core.

It will be helpful if any scenario can be explained where presence of 
unix-epoll-input() on worker core solves the purpose

Thanks,
Nitin

--
Damjan



--
Damjan

-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.

View/Reply Online (#15346): https://lists.fd.io/g/vpp-dev/message/15346
Mute This Topic: https://lists.fd.io/mt/70992570/21656
Group Owner: vpp-dev+ow...@lists.fd.io
Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub  [arch...@mail-archive.com]
-=-=-=-=-=-=-=-=-=-=-=-

Reply via email to