[RFC 4/4] net/ice: remove use of VLAs

[Konstantin Ananyev]

> 
> On 5/23/2024 5:26 PM, Konstantin Ananyev wrote:
> > From: Konstantin Ananyev <konstantin.ananyev at huawei.com>
> >
> > ../drivers/net/ice/ice_rxtx.c:1871:29: warning: variable length array used [-Wvla]
> >
> > Here VLA is used as a temp array for mbufs that will be used as a split
> > RX data buffers.
> > As at any given time only one thread can do RX from particular queue,
> > at rx_queue_setup() we can allocate extra space for that array, and then
> > safely use it at RX fast-path.
> >
> 
> Is there a reason to allocate extra space in sw_ring and used some part
> of it for split buffer, instead of allocating a new buffer for it?

Less allocations - less points to fail, less checks to do.
Again, having it close to sw_ring is probably a good thing too,
possibly less pressure on MMU, etc. - even though I don't think it is really critical.
But yes,  it could be a separate rte_zmalloc(), even though
I don't see good reason for that.
 
> 
> > Signed-off-by: Konstantin Ananyev <konstantin.ananyev at huawei.com>
> > ---
> >  drivers/net/ice/ice_rxtx.c | 18 ++++++++++++------
> >  drivers/net/ice/ice_rxtx.h |  2 ++
> >  2 files changed, 14 insertions(+), 6 deletions(-)
> >
> > diff --git a/drivers/net/ice/ice_rxtx.c b/drivers/net/ice/ice_rxtx.c
> > index 95a2db3432..6395a3b50a 100644
> > --- a/drivers/net/ice/ice_rxtx.c
> > +++ b/drivers/net/ice/ice_rxtx.c
> > @@ -1171,7 +1171,7 @@ ice_rx_queue_setup(struct rte_eth_dev *dev,
> >  	struct ice_vsi *vsi = pf->main_vsi;
> >  	struct ice_rx_queue *rxq;
> >  	const struct rte_memzone *rz;
> > -	uint32_t ring_size;
> > +	uint32_t ring_size, tlen;
> >  	uint16_t len;
> >  	int use_def_burst_func = 1;
> >  	uint64_t offloads;
> > @@ -1279,9 +1279,14 @@ ice_rx_queue_setup(struct rte_eth_dev *dev,
> >  	/* always reserve more for bulk alloc */
> >  	len = (uint16_t)(nb_desc + ICE_RX_MAX_BURST);
> >
> > +	/* allocate extra entries for SW split buffer */
> > +	tlen = ((rxq->offloads & RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT) != 0) ?
> > +		rxq->rx_free_thresh : 0;
> > +	tlen += len;
> > +
> >  	/* Allocate the software ring. */
> >  	rxq->sw_ring = rte_zmalloc_socket(NULL,
> > -					  sizeof(struct ice_rx_entry) * len,
> > +					  sizeof(struct ice_rx_entry) * tlen,
> >  					  RTE_CACHE_LINE_SIZE,
> >  					  socket_id);
> >  	if (!rxq->sw_ring) {
> > @@ -1290,6 +1295,8 @@ ice_rx_queue_setup(struct rte_eth_dev *dev,
> >  		return -ENOMEM;
> >  	}
> >
> > +	rxq->sw_split_buf = (tlen == len) ? NULL : rxq->sw_ring + len;
> > +
> >  	ice_reset_rx_queue(rxq);
> >  	rxq->q_set = true;
> >  	dev->data->rx_queues[queue_idx] = rxq;
> > @@ -1868,7 +1875,6 @@ ice_rx_alloc_bufs(struct ice_rx_queue *rxq)
> >  	uint64_t dma_addr;
> >  	int diag, diag_pay;
> >  	uint64_t pay_addr;
> > -	struct rte_mbuf *mbufs_pay[rxq->rx_free_thresh];
> >
> >  	/* Allocate buffers in bulk */
> >  	alloc_idx = (uint16_t)(rxq->rx_free_trigger -
> > @@ -1883,7 +1889,7 @@ ice_rx_alloc_bufs(struct ice_rx_queue *rxq)
> >
> >  	if (rxq->offloads & RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT) {
> >  		diag_pay = rte_mempool_get_bulk(rxq->rxseg[1].mp,
> > -				(void *)mbufs_pay, rxq->rx_free_thresh);
> > +				(void *)rxq->sw_split_buf, rxq->rx_free_thresh);
> >
> 
> Are we allowed to call 'rte_mempool_get_bulk()' with NULL object_table,

Nope.

> as 'rxq->sw_split_buf' can be NULL?
> Perhaps can allocate 'rxq->sw_split_buf' even buffer split offload is
> not enabled?

No,  if (rxq->offloads & RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT) !=0, then
rxq->sw_split_buf should not be NULL.
If it is, then there is a bug in my changes, though right now I don't see
how it can happen: as in ice_rx_queue_setup() we always allocate space
rxq->sw_split_buf when RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT is set.