[dpdk-dev] [PATCH] eventdev: remove experimental label
jerin.jacob at caviumnetworks.com
Mon Oct 30 18:38:05 CET 2017
> Date: Mon, 23 Oct 2017 18:27:52 +0000
> From: "Eads, Gage" <gage.eads at intel.com>
> To: Jerin Jacob <jerin.jacob at caviumnetworks.com>, "dev at dpdk.org"
> <dev at dpdk.org>
> CC: "Richardson, Bruce" <bruce.richardson at intel.com>, "Van Haaren, Harry"
> <harry.van.haaren at intel.com>, Hemant Agrawal <hemant.agrawal at nxp.com>,
> Nipun Gupta <nipun.gupta at nxp.com>, "Rao, Nikhil" <nikhil.rao at intel.com>,
> Pavan Nikhilesh <pbhagavatula at caviumnetworks.com>, Thomas Monjalon
> <thomas at monjalon.net>
> Subject: RE: [dpdk-dev] [PATCH] eventdev: remove experimental label
> Hi Jerin,
> I have one concern with the API that may delay changing the label.
> The implicit release that in rte_event_dequeue_burst() is a problem when using asynchronous/look-aside hardware, like a cryptodev. For instance, let's say in pipeline stage N the worker takes the event's mbuf and places it in a per-worker crypto request queue. When the worker next calls rte_event_dequeue_burst(), that function will release the previous event which could cause the flow to migrate to another worker, and this could result in packet reordering.
> To prevent this, the worker can't call dequeue until the look-aside operation completes...in effect treating the asynchronous/look-aside hardware as synchronous. Another option is to feed stage N's queue to a single port to avoid the flow migration, but that port may become a bottleneck.
> We could remove the implicit release functionality or add a port configuration flag to disable it, so the default behavior is unchanged. Removing it will completely will likely require changes in existing code, but it simplifies the usage model (all dequeued events must be either forwarded or released) and the PMD's dequeue code. This functionality could be removed from the software eventdev fairly easily, but I haven't looked into the hardware PMDs.
The HW implementations, I know, it does the implicit release. Otherwise it
will result in deadlock because it cannot hold reordering metadata for
the longtime(SRAM is limited etc)
Coming back to cryptodev use case, if I understand it correctly, before
application enqueues to crypto queue, the application will change the tag and
submit to ATOMIC queue. So as long as crypto queue competes for the
crypto work in order then the order will be maintained.
In typical outbound IPSec use case,
- Stage 1 will be processed in ORDERED where application does the SA
- Once SA found, application enqueue to ATOMIC stage with SA as flow_id.
- When the event comes from the ATOMIC queue, it in ingress order and
then it submits to the crypto queue
- Crypto queue maintains the FIFO order.
- On IPSec crypto work competition, packets will come in Stage 3.
- So at Stage 3, packets are in ingress order for the given SA flow id.
Having said that, If SW implementation needs to do differently for
performance reasons then we will end up in capability as HW
implementation works in the implicit release. May we can sort out
through capability or separate adapter for crypto case. But I think, those will
be new additions to the API.So removing the experimental tags may be OK.
But if you have strong opinion on keeping the experimental tag till we address
the crypto use case then I am fine with that.
> > -----Original Message-----
> > From: Jerin Jacob [mailto:jerin.jacob at caviumnetworks.com]
> > Sent: Monday, October 16, 2017 5:33 AM
> > To: dev at dpdk.org
> > Cc: Jerin Jacob <jerin.jacob at caviumnetworks.com>; Richardson, Bruce
> > <bruce.richardson at intel.com>; Van Haaren, Harry
> > <harry.van.haaren at intel.com>; Eads, Gage <gage.eads at intel.com>; Hemant
> > Agrawal <hemant.agrawal at nxp.com>; Nipun Gupta <nipun.gupta at nxp.com>;
> > Rao, Nikhil <nikhil.rao at intel.com>; Pavan Nikhilesh
> > <pbhagavatula at caviumnetworks.com>; Thomas Monjalon
> > <thomas at monjalon.net>
> > Subject: [dpdk-dev] [PATCH] eventdev: remove experimental label
> > The eventdev API was introduced in DPDK 17.05 release.
> > Since then it
> > - has been reviewed and iterated for 17.08, 17.11 releases
> > - three drivers were implemented using the API.
> > - introduced another subsystem like service core and ethdev-eventdev Rx
> > adapter APIs to abstract the difference between HW and SW eventdev
> > implementations in a transparent way.
> > - had extensive use by the app/test-eventdev/ and
> > examples/eventdev_pipeline_sw_pmd/
> > I believe the API is now stable and the EXPERIMENTAL label should be removed.
> > CC: Bruce Richardson <bruce.richardson at intel.com>
> > CC: Harry van Haaren <harry.van.haaren at intel.com>
> > CC: Gage Eads <gage.eads at intel.com>
> > CC: Hemant Agrawal <hemant.agrawal at nxp.com>
> > CC: Nipun Gupta <nipun.gupta at nxp.com>
> > CC: Nikhil Rao <nikhil.rao at intel.com>
> > CC: Pavan Nikhilesh <pbhagavatula at caviumnetworks.com>
> > CC: Thomas Monjalon <thomas at monjalon.net>
> > Signed-off-by: Jerin Jacob <jerin.jacob at caviumnetworks.com>
> > ---
> > There are two more outstanding eventdev API changes. Please find below.
> > Please express if you have any concern in changing those APIs. I would
> > like to fix this API issue and remove experimental tag in v17.11,
> > if we all agree.
> > - evendev: fix inconsistency in event queue config
> > http://dpdk.org/dev/patchwork/patch/30293/
> > - remove rte_event_schedule() API and use service core infrastructure
> > http://dpdk.org/dev/patchwork/patch/30375/
> > ---
> > MAINTAINERS | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> > diff --git a/MAINTAINERS b/MAINTAINERS
> > index 2a58378b7..4a4be3a54 100644
> > --- a/MAINTAINERS
> > +++ b/MAINTAINERS
> > @@ -262,7 +262,7 @@ F: lib/librte_cryptodev/
> > F: test/test/test_cryptodev*
> > F: examples/l2fwd-crypto/
> > -Eventdev API - EXPERIMENTAL
> > +Eventdev API
> > M: Jerin Jacob <jerin.jacob at caviumnetworks.com>
> > T: git://dpdk.org/next/dpdk-next-eventdev
> > F: lib/librte_eventdev/
> > --
> > 2.14.2
More information about the dev