Re: [dpdk-dev] [PATCH] [RFC] ethdev: support flow aging

[Matan Azrad <matan@mellanox.com>]

Hi Jerin

From: Jerin Jacob
> Sent: Thursday, June 6, 2019 3:16 PM
> To: Matan Azrad <matan@mellanox.com>; Adrien Mazarguil
> <adrien.mazarguil@6wind.com>; dev@dpdk.org
> Subject: RE: [PATCH] [RFC] ethdev: support flow aging
> 
> > -----Original Message-----
> > From: Matan Azrad <matan@mellanox.com>
> > Sent: Thursday, June 6, 2019 4:22 PM
> > To: Jerin Jacob Kollanukkaran <jerinj@marvell.com>; Adrien Mazarguil
> > <adrien.mazarguil@6wind.com>; dev@dpdk.org
> > Subject: [EXT] RE: [PATCH] [RFC] ethdev: support flow aging
> >
> > Hi Jerin
> 
> Hi Matan,
> 
> >
> > From: Jerin Jacob
> > > > -----Original Message-----
> > > > From: dev <dev-bounces@dpdk.org> On Behalf Of Matan Azrad
> > > > Sent: Sunday, May 26, 2019 3:48 PM
> > > > To: Adrien Mazarguil <adrien.mazarguil@6wind.com>; dev@dpdk.org
> > > > Subject: [dpdk-dev] [PATCH] [RFC] ethdev: support flow aging
> > > >
> > > > One of the reasons to destroy a flow is the fact that no packet
> > > > matches the flow for "timeout" time.
> > > > For example, when TCP\UDP sessions are suddenly closed.
> > > >
> > > > Currently, there is no any dpdk mechanism for flow aging and the
> > > > applications use there own ways to detect and destroy aged-out flows.
> > > >
> > > > This RFC introduces flow aging APIs to offload the flow aging task
> > > > from the application to the port.
> > > >
> > > > Design:
> > > > - A new rte_flow action: RTE_FLOW_ACTION_TYPE_AGE to set the
> > timeout
> > > > and
> > > >   the application flow context for each flow.
> > > > - A new ethdev event: RTE_ETH_EVENT_FLOW_AGED for the driver to
> > > report
> > > >   that there are new aged-out flows.
> > > > - A new rte_flow API: rte_flow_get_aged_flows to get the aged-out
> > flows
> > > >   contexts from the port.
> > > >
> > > > By this design each PMD can use its best way to do the aging with
> > > > the device offloads supported by its HW.
> > > >
> > > > Signed-off-by: Matan Azrad <matan@mellanox.com>
> > > > ---
> > > >  lib/librte_ethdev/rte_ethdev.h |  1 +
> > > >  lib/librte_ethdev/rte_flow.h   | 56
> > > > ++++++++++++++++++++++++++++++++++++++++++
> > > >  2 files changed, 57 insertions(+)
> > > >
> > > > diff --git a/lib/librte_ethdev/rte_ethdev.h
> > > > b/lib/librte_ethdev/rte_ethdev.h index 1f35e1d..6fc1531 100644
> > > > --- a/lib/librte_ethdev/rte_ethdev.h
> > > > +++ b/lib/librte_ethdev/rte_ethdev.h
> > > > @@ -2771,6 +2771,7 @@ enum rte_eth_event_type {
> > > >  	RTE_ETH_EVENT_NEW,      /**< port is probed */
> > > >  	RTE_ETH_EVENT_DESTROY,  /**< port is released */
> > > >  	RTE_ETH_EVENT_IPSEC,    /**< IPsec offload related event */
> > > > +	RTE_ETH_EVENT_FLOW_AGED,/**< New aged-out flows detected in
> > > > the port
> > > Does this event supported in HW?
> > It depends in the PMD implementation and HW capability.
> >
> > > Or Are planning to implement with alarm or timer.
> > Again, depends in the PMD implementation.
> >
> > > Just asking because, if none of the HW supports the interrupt then
> > > only rte_flow_get_aged_flows sync API be enough()
> > Why?
> 
> If none of the HW supports it then application/common code can periodically
> polls it.
> If mlx5 hw supports it then it fine to have interrupt.

Actually MLX5 doesn't support aging fully by HW but the HW can help to do it better.
Look, the PMD is the best one to know what is the best way to do aging by its HW even if aging is not fully supported by it.
And it may add a meaningful efficiency to the application. 

> But I think, we need to have means to express a HW/Implementation does
> not support its As there may following reasons why drivers choose to not
> take timer/alarm path
> 1) Some EAL port does not support timer/alarm example: FreeBSD DPDK port
	OK, but why not to support it for other cases (no FreeBSD port)?

> 2) If we need to support a few killo rules then timer/alarm implementation
> will be heavy

Not sure, Depend in the HW ability.

> So an option to express un supported event would be fine.

Can you explain more what is your intension here (2)?

> > According to the above design this is the way for the PMD to notify
> > the application when it has some aged flows ASAP.
> > So, if the PMD uses an alarm\timer or any other way to support aging
> > action it is better in part of the cases to notify the user
> > asynchronically instead of doing polling by the application.
> > The idea is to let the application to decide what is better for its usage.
> >
> > For mlx5 case,
> > The plan is to raise this event from an HW interrupt handling(same as
> > link event).
> 
> Good to know.

The MLX5 plan is still to use timer/alarm and interrupt mechanism to support aging:
 The HW help here is the ability to query batch of flows counters asynchronically, so getting the response of the new counters values by an interrupt.

The timer\alarm will call to devX operation to read batch of counters asynchronically - fast command.
The interrupt handler to catch the response and to check timeout for each flow
(no need to copy the counters from the HW memory - the values are in the PMD memory) - if there is a new aged flow - raise the event.