Re: [dpdk-dev] [PATCH] net/mlx5: implement drop action in hardware classifier

["Nélio Laranjeiro" <nelio.laranjeiro@6wind.com>]

>> From: Nélio Laranjeiro [mailto:nelio.laranjeiro@6wind.com] 
>> Sent: Monday, May 29, 2017 4:08 PM
>> To: Shachar Beiser <shacharbe@mellanox.com>
>> Cc: dev@dpdk.org; Adrien Mazarguil <adrien.mazarguil@6wind.com>
>> Subject: Re: [PATCH] net/mlx5: implement drop action in hardware classifier
>> 
>> On Sun, May 28, 2017 at 06:49:20AM +0000, Shachar Beiser wrote:
>> > The current drop action is implemented as a queue tail drop,
>> > requiring to instantiate multiple WQs to maintain high drop rate.
>> > This commit, implements the drop action in hardware classifier.
>> > This enables to reduce the amount of contexts needed for the drop,
>> > without affecting the drop rate.
>> > 
>> > Signed-off-by: Shachar Beiser <shacharbe@mellanox.com>
>> > ---
>> >  drivers/net/mlx5/Makefile    |  5 +++++
>> >  drivers/net/mlx5/mlx5_flow.c | 43 +++++++++++++++++++++++++++++++++++++++++++
>> >  2 files changed, 48 insertions(+)
>> > 
>> > diff --git a/drivers/net/mlx5/Makefile b/drivers/net/mlx5/Makefile
>> > index c079959..daf8013 100644
>> > --- a/drivers/net/mlx5/Makefile
>> > +++ b/drivers/net/mlx5/Makefile
>> > @@ -101,6 +101,11 @@ mlx5_autoconf.h.new: FORCE
>> >  mlx5_autoconf.h.new: $(RTE_SDK)/buildtools/auto-config-h.sh
>> >  	$Q $(RM) -f -- '$@'
>> >  	$Q sh -- '$<' '$@' \
>> > +		HAVE_VERBS_IBV_EXP_FLOW_SPEC_ACTION_DROP \
>> > +		infiniband/verbs_exp.h \
>> > +		enum IBV_EXP_FLOW_SPEC_ACTION_DROP \
>> > +		$(AUTOCONF_OUTPUT)
>> > +	$Q sh -- '$<' '$@' \
>> >  		HAVE_VERBS_IBV_EXP_CQ_COMPRESSED_CQE \
>> >  		infiniband/verbs_exp.h \
>> >  		enum IBV_EXP_CQ_COMPRESSED_CQE \
>> > diff --git a/drivers/net/mlx5/mlx5_flow.c b/drivers/net/mlx5/mlx5_flow.c
>> > index adcbe3f..e243d39 100644
>> > --- a/drivers/net/mlx5/mlx5_flow.c
>> > +++ b/drivers/net/mlx5/mlx5_flow.c
>> > @@ -994,6 +994,11 @@ struct mlx5_flow_action {
>> >  {
>> >  	struct rte_flow *rte_flow;
>> >  
>> > +#ifdef HAVE_VERBS_IBV_EXP_FLOW_SPEC_ACTION_DROP
>> > +	struct ibv_exp_flow_spec_action_drop *drop;
>> > +	unsigned int size = sizeof(struct ibv_exp_flow_spec_action_drop);
>> > +#endif
>> 
>> Extra empty line before the #ifdef
>> 
>> > +
>> >  	assert(priv->pd);
>> >  	assert(priv->ctx);
>> >  	rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
>> > @@ -1007,6 +1012,15 @@ struct mlx5_flow_action {
>> >  	rte_flow->qp = priv->flow_drop_queue->qp;
>> >  	if (!priv->started)
>> >  		return rte_flow;
>> > +#ifdef HAVE_VERBS_IBV_EXP_FLOW_SPEC_ACTION_DROP
>> > +	drop = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
>> > +	*drop = (struct ibv_exp_flow_spec_action_drop){
>> > +			.type = IBV_EXP_FLOW_SPEC_ACTION_DROP,
>> > +			.size = size,
>> > +	};
>> > +	++flow->ibv_attr->num_of_specs;
>> > +	flow->offset += sizeof(struct ibv_exp_flow_spec_action_drop);
>> > +#endif
>> >  	rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
>> >  						 rte_flow->ibv_attr);
>> >  	if (!rte_flow->ibv_flow) {
>> >[...]
>> 
>> From what I see by just changing the value of MLX5_DROP_WQ_N when
>> HAVE_VERBS_IBV_EXP_FLOW_SPEC_ACTION_DROP (in the same source file)
>> limits the patch to this point.
>> 
>> Am I missing something?
>> 
>> Regards,
> On Mon, May 29, 2017 at 01:35:57PM +0000, Shachar Beiser wrote:
> The current firmware version does not implement queue tail drop
> therefore queue allocation is not required by firmware. Unfortunately,
> the ibverbs APIs requires to allocate WQ & CQ  .In the future this
> allocation will be saved . Right now, you are right. 

What I mean is, in your patch, what you need is the snippet I let above,
and at compilation time you can update the MLX5_DROP_WQ_N to 1 (defined
at 4 in the same file), this way the remaining code will only create a
single CQ/WQ as you expect.

This small snippet above plus 3 lines handles the job.

Regards,

PS: Avoid top answering on this mailing list.

-- 
Nélio Laranjeiro
6WIND