Re: [dpdk-dev] [PATCH v3] net/i40e: fix Rx instability with vector mode

["Zhang, Qi Z" <qi.z.zhang@intel.com>]

> -----Original Message-----
> From: Ananyev, Konstantin
> Sent: Thursday, November 1, 2018 6:48 AM
> To: Xing, Beilei <beilei.xing@intel.com>; Zhang, Qi Z <qi.z.zhang@intel.com>
> Cc: dev@dpdk.org; stable@dpdk.org
> Subject: RE: [dpdk-dev] [PATCH v3] net/i40e: fix Rx instability with vector
> mode
> 
> 
> Hi
> 
> >
> > Previously, there is instability during vector Rx if descriptor number
> > is not power of 2, e.g. process hang and some Rx packets are
> > unexpectedly empty. That's because vector Rx mode assumes Rx
> > descriptor number is power of 2 when doing bit mask.
> > This patch allows vector mode only when the number of Rx descriptor is
> > power of 2.
> >
> > Fixes: 8e109464c022 ("i40e: allow vector Rx and Tx usage")
> > Fixes: a3c83a2527e1 ("net/i40e: enable runtime queue setup")
> > Cc: stable@dpdk.org
> >
> > Signed-off-by: Beilei Xing <beilei.xing@intel.com>
> > ---
> > v3 changes:
> >  - Add branch for non-first queue during runtime queue setup.
> >  - Use function rte_is_power_of_2().
> >  - Configure rx_vec_allowed during setting Rx function.
> > v2 changes:
> >  - rx_vec_allowed is global configuration, avoid overwrite.
> >
> >  doc/guides/nics/i40e.rst     |  7 +++++++
> >  drivers/net/i40e/i40e_rxtx.c | 30 +++++++++++++++++++++++++++---
> >  2 files changed, 34 insertions(+), 3 deletions(-)
> >
> > diff --git a/doc/guides/nics/i40e.rst b/doc/guides/nics/i40e.rst index
> > ab3928a..bfacbd1 100644
> > --- a/doc/guides/nics/i40e.rst
> > +++ b/doc/guides/nics/i40e.rst
> > @@ -172,6 +172,13 @@ Runtime Config Options
> >
> >    -w 84:00.0,use-latest-supported-vec=1
> >
> > +Vector RX Pre-conditions
> > +~~~~~~~~~~~~~~~~~~~~~~~~
> > +For Vector RX it is assumed that the number of descriptor rings will
> > +be a power of 2. With this pre-condition, the ring pointer can easily
> > +scroll back to the head after hitting the tail without a conditional
> > +check. In addition Vector RX can use this assumption to do a bit mask
> using ``ring_size - 1``.
> > +
> >  Driver compilation and testing
> >  ------------------------------
> >
> > diff --git a/drivers/net/i40e/i40e_rxtx.c
> > b/drivers/net/i40e/i40e_rxtx.c index a827456..771193a 100644
> > --- a/drivers/net/i40e/i40e_rxtx.c
> > +++ b/drivers/net/i40e/i40e_rxtx.c
> > @@ -1735,12 +1735,21 @@ i40e_dev_rx_queue_setup_runtime(struct
> rte_eth_dev *dev,
> >  		 * i40e_set_rx_function.
> >  		 */
> >  		ad->rx_bulk_alloc_allowed = true;
> > -		ad->rx_vec_allowed = true;
> >  		dev->data->scattered_rx = use_scattered_rx;
> >  		if (use_def_burst_func)
> >  			ad->rx_bulk_alloc_allowed = false;
> > +		/**
> > +		 * Vector mode is allowed only when number of Rx queue
> > +		 * descriptor is a power of 2.
> > +		 */
> > +		ad->rx_vec_allowed = rte_is_power_of_2(rxq->nb_rx_desc);
> 
> Actually do we need to do it here?
> We call set_rx_function() anyway, it would do that check for us, wouldn't it?
> 
> >  		i40e_set_rx_function(dev);
> >  		return 0;
> > +	} else if (ad->rx_vec_allowed && !rte_is_power_of_2(rxq->nb_rx_desc))
> {
> > +		PMD_DRV_LOG(ERR, "Vector mode is allowed, but descriptor"
> > +			    " number %d of queue %d isn't power of 2",
> > +			    rxq->nb_rx_desc, rxq->queue_id);
> > +		return -EINVAL;
> >  	}
> >
> >  	/* check bulk alloc conflict */
> > @@ -2948,11 +2957,26 @@ i40e_set_rx_function(struct rte_eth_dev *dev)
> > {
> >  	struct i40e_adapter *ad =
> >  		I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
> > +	struct i40e_rx_queue *rxq;
> >  	uint16_t rx_using_sse, i;
> > +	uint16_t desc;
> 
> Please add empty line between var definitions and start of code.
> Helps readability.
> 
> >  	/* In order to allow Vector Rx there are a few configuration
> >  	 * conditions to be met and Rx Bulk Allocation should be allowed.
> >  	 */
> >  	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
> > +		if (!dev->data->dev_started) {
> > +			for (i = 0; i < dev->data->nb_rx_queues; i++) {
> > +				rxq = dev->data->rx_queues[i];
> > +				desc = rxq->nb_rx_desc;
> > +				if (!i)
queue_setup is not necessary start from queue 0, we can setup queue at any order, so we can't assume i=0 is the first queue to setup.
And also we need to check if (dev->data->rx_queues[i] == NULL) to avoid segment fault, since some queue may never be setup yet.

> > +					ad->rx_vec_allowed =
> > +						rte_is_power_of_2(desc);
> > +				else if (ad->rx_vec_allowed &&
> > +					 !rte_is_power_of_2(desc))
> > +					ad->rx_vec_allowed = false;
> 
> Wouldn't be a bit cleaner:
> ad->rx_vec_allowed = (ad->rx_vec_allowed == true) ?
> ad->rte_is_power_of_2(desc)) : ad->rx_vec_allowed;


> 
> > +			}
> > +		}
> > +
> 
> Probably better to move that code into
> i40e_rx_vec_dev_conf_condition_check()
> that will be called on next line?
> 
> >  		if (i40e_rx_vec_dev_conf_condition_check(dev) ||
> >  		    !ad->rx_bulk_alloc_allowed) {
> >  			PMD_INIT_LOG(DEBUG, "Port[%d] doesn't meet"
> > @@ -2961,10 +2985,10 @@ i40e_set_rx_function(struct rte_eth_dev *dev)
> >
> >  			ad->rx_vec_allowed = false;
> >  		}
> > +
> >  		if (ad->rx_vec_allowed) {
> >  			for (i = 0; i < dev->data->nb_rx_queues; i++) {
> > -				struct i40e_rx_queue *rxq =
> > -					dev->data->rx_queues[i];
> > +				rxq = dev->data->rx_queues[i];
> >
> >  				if (rxq && i40e_rxq_vec_setup(rxq)) {
> >  					ad->rx_vec_allowed = false;
> > --
> > 2.5.5