* [PATCH 2/3] net/qede: fix Rx callback
2022-03-04 12:08 [PATCH 1/3] net/qede: fix Tx callback completion routine Devendra Singh Rawat
@ 2022-03-04 12:08 ` Devendra Singh Rawat
2022-03-04 12:08 ` [PATCH 3/3] net/qede: fix max Rx pktlen calculation Devendra Singh Rawat
2022-03-10 7:40 ` [PATCH 1/3] net/qede: fix Tx callback completion routine Jerin Jacob
2 siblings, 0 replies; 6+ messages in thread
From: Devendra Singh Rawat @ 2022-03-04 12:08 UTC (permalink / raw)
To: dev, thomas, jerinj, ferruh.yigit, rmody
Cc: palok, Devendra Singh Rawat, stable
qede_alloc_rx_bulk_mbufs was trimming the no. of requested mbufs count
to QEDE_MAX_BULK_ALLOC_COUNT. The RX callback was ignorant of this
trimming and it was always resetting the no. of empty RX BD ring
slots to 0. This resulted in RX BD ring getting into an inconsistent
state and ultimately the application fails to receive any traffic.
The fix trims the no. of requested mbufs count before making call to
qede_alloc_rx_bulk_mbufs. After qede_alloc_rx_bulk_mbufs returns
successfully, the no. of empty RX BD ring slots are decremented by the
correct count.
Fixes: 8f2312474529 ("net/qede: fix performance bottleneck in Rx path")
Cc: stable@dpdk.org
Signed-off-by: Devendra Singh Rawat <dsinghrawat@marvell.com>
Signed-off-by: Rasesh Mody <rmody@marvell.com>
---
drivers/net/qede/qede_rxtx.c | 68 ++++++++++++++++--------------------
1 file changed, 31 insertions(+), 37 deletions(-)
diff --git a/drivers/net/qede/qede_rxtx.c b/drivers/net/qede/qede_rxtx.c
index 0c52568180..02fa1fcaa1 100644
--- a/drivers/net/qede/qede_rxtx.c
+++ b/drivers/net/qede/qede_rxtx.c
@@ -38,48 +38,40 @@ static inline int qede_alloc_rx_buffer(struct qede_rx_queue *rxq)
static inline int qede_alloc_rx_bulk_mbufs(struct qede_rx_queue *rxq, int count)
{
+ void *obj_p[QEDE_MAX_BULK_ALLOC_COUNT] __rte_cache_aligned;
struct rte_mbuf *mbuf = NULL;
struct eth_rx_bd *rx_bd;
dma_addr_t mapping;
int i, ret = 0;
uint16_t idx;
- uint16_t mask = NUM_RX_BDS(rxq);
-
- if (count > QEDE_MAX_BULK_ALLOC_COUNT)
- count = QEDE_MAX_BULK_ALLOC_COUNT;
idx = rxq->sw_rx_prod & NUM_RX_BDS(rxq);
- if (count > mask - idx + 1)
- count = mask - idx + 1;
-
- ret = rte_mempool_get_bulk(rxq->mb_pool, (void **)&rxq->sw_rx_ring[idx],
- count);
-
+ ret = rte_mempool_get_bulk(rxq->mb_pool, obj_p, count);
if (unlikely(ret)) {
PMD_RX_LOG(ERR, rxq,
"Failed to allocate %d rx buffers "
"sw_rx_prod %u sw_rx_cons %u mp entries %u free %u",
- count,
- rxq->sw_rx_prod & NUM_RX_BDS(rxq),
- rxq->sw_rx_cons & NUM_RX_BDS(rxq),
+ count, idx, rxq->sw_rx_cons & NUM_RX_BDS(rxq),
rte_mempool_avail_count(rxq->mb_pool),
rte_mempool_in_use_count(rxq->mb_pool));
return -ENOMEM;
}
for (i = 0; i < count; i++) {
- rte_prefetch0(rxq->sw_rx_ring[(idx + 1) & NUM_RX_BDS(rxq)]);
- mbuf = rxq->sw_rx_ring[idx & NUM_RX_BDS(rxq)];
+ mbuf = obj_p[i];
+ if (likely(i < count - 1))
+ rte_prefetch0(obj_p[i + 1]);
+ idx = rxq->sw_rx_prod & NUM_RX_BDS(rxq);
+ rxq->sw_rx_ring[idx] = mbuf;
mapping = rte_mbuf_data_iova_default(mbuf);
rx_bd = (struct eth_rx_bd *)
ecore_chain_produce(&rxq->rx_bd_ring);
rx_bd->addr.hi = rte_cpu_to_le_32(U64_HI(mapping));
rx_bd->addr.lo = rte_cpu_to_le_32(U64_LO(mapping));
- idx++;
+ rxq->sw_rx_prod++;
}
- rxq->sw_rx_prod = idx;
return 0;
}
@@ -1544,25 +1536,26 @@ qede_recv_pkts_regular(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
uint8_t bitfield_val;
#endif
uint8_t offset, flags, bd_num;
-
+ uint16_t count = 0;
/* Allocate buffers that we used in previous loop */
if (rxq->rx_alloc_count) {
- if (unlikely(qede_alloc_rx_bulk_mbufs(rxq,
- rxq->rx_alloc_count))) {
+ count = rxq->rx_alloc_count > QEDE_MAX_BULK_ALLOC_COUNT ?
+ QEDE_MAX_BULK_ALLOC_COUNT : rxq->rx_alloc_count;
+
+ if (unlikely(qede_alloc_rx_bulk_mbufs(rxq, count))) {
struct rte_eth_dev *dev;
PMD_RX_LOG(ERR, rxq,
- "New buffer allocation failed,"
- "dropping incoming packetn");
+ "New buffers allocation failed,"
+ "dropping incoming packets\n");
dev = &rte_eth_devices[rxq->port_id];
- dev->data->rx_mbuf_alloc_failed +=
- rxq->rx_alloc_count;
- rxq->rx_alloc_errors += rxq->rx_alloc_count;
+ dev->data->rx_mbuf_alloc_failed += count;
+ rxq->rx_alloc_errors += count;
return 0;
}
qede_update_rx_prod(qdev, rxq);
- rxq->rx_alloc_count = 0;
+ rxq->rx_alloc_count -= count;
}
hw_comp_cons = rte_le_to_cpu_16(*rxq->hw_cons_ptr);
@@ -1731,7 +1724,7 @@ qede_recv_pkts_regular(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
}
/* Request number of buffers to be allocated in next loop */
- rxq->rx_alloc_count = rx_alloc_count;
+ rxq->rx_alloc_count += rx_alloc_count;
rxq->rcv_pkts += rx_pkt;
rxq->rx_segs += rx_pkt;
@@ -1771,25 +1764,26 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
struct qede_agg_info *tpa_info = NULL;
uint32_t rss_hash;
int rx_alloc_count = 0;
-
+ uint16_t count = 0;
/* Allocate buffers that we used in previous loop */
if (rxq->rx_alloc_count) {
- if (unlikely(qede_alloc_rx_bulk_mbufs(rxq,
- rxq->rx_alloc_count))) {
+ count = rxq->rx_alloc_count > QEDE_MAX_BULK_ALLOC_COUNT ?
+ QEDE_MAX_BULK_ALLOC_COUNT : rxq->rx_alloc_count;
+
+ if (unlikely(qede_alloc_rx_bulk_mbufs(rxq, count))) {
struct rte_eth_dev *dev;
PMD_RX_LOG(ERR, rxq,
- "New buffer allocation failed,"
- "dropping incoming packetn");
+ "New buffers allocation failed,"
+ "dropping incoming packets\n");
dev = &rte_eth_devices[rxq->port_id];
- dev->data->rx_mbuf_alloc_failed +=
- rxq->rx_alloc_count;
- rxq->rx_alloc_errors += rxq->rx_alloc_count;
+ dev->data->rx_mbuf_alloc_failed += count;
+ rxq->rx_alloc_errors += count;
return 0;
}
qede_update_rx_prod(qdev, rxq);
- rxq->rx_alloc_count = 0;
+ rxq->rx_alloc_count -= count;
}
hw_comp_cons = rte_le_to_cpu_16(*rxq->hw_cons_ptr);
@@ -2028,7 +2022,7 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
}
/* Request number of buffers to be allocated in next loop */
- rxq->rx_alloc_count = rx_alloc_count;
+ rxq->rx_alloc_count += rx_alloc_count;
rxq->rcv_pkts += rx_pkt;
--
2.18.2
^ permalink raw reply [flat|nested] 6+ messages in thread* [PATCH 3/3] net/qede: fix max Rx pktlen calculation
2022-03-04 12:08 [PATCH 1/3] net/qede: fix Tx callback completion routine Devendra Singh Rawat
2022-03-04 12:08 ` [PATCH 2/3] net/qede: fix Rx callback Devendra Singh Rawat
@ 2022-03-04 12:08 ` Devendra Singh Rawat
2022-03-10 7:40 ` [PATCH 1/3] net/qede: fix Tx callback completion routine Jerin Jacob
2 siblings, 0 replies; 6+ messages in thread
From: Devendra Singh Rawat @ 2022-03-04 12:08 UTC (permalink / raw)
To: dev, thomas, jerinj, ferruh.yigit, rmody
Cc: palok, Devendra Singh Rawat, stable
size of CRC is not added to max_rx_pktlen, due to this bigger sized
packets(size 1480, 1490 1500) are being dropped.
This fix adds RTE_ETHER_CRC_LEN to max_rx_pktlen.
Fixes: 1bb4a528c41f ("ethdev: fix max Rx packet length")
Cc: stable@dpdk.org
Signed-off-by: Devendra Singh Rawat <dsinghrawat@marvell.com>
Signed-off-by: Rasesh Mody <rmody@marvell.com>
---
drivers/net/qede/qede_rxtx.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/net/qede/qede_rxtx.c b/drivers/net/qede/qede_rxtx.c
index 02fa1fcaa1..c35585f5fd 100644
--- a/drivers/net/qede/qede_rxtx.c
+++ b/drivers/net/qede/qede_rxtx.c
@@ -235,7 +235,7 @@ qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t qid,
dev->data->rx_queues[qid] = NULL;
}
- max_rx_pktlen = dev->data->mtu + RTE_ETHER_HDR_LEN;
+ max_rx_pktlen = dev->data->mtu + RTE_ETHER_HDR_LEN + RTE_ETHER_CRC_LEN;
/* Fix up RX buffer size */
bufsz = (uint16_t)rte_pktmbuf_data_room_size(mp) - RTE_PKTMBUF_HEADROOM;
--
2.18.2
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: [PATCH 1/3] net/qede: fix Tx callback completion routine
2022-03-04 12:08 [PATCH 1/3] net/qede: fix Tx callback completion routine Devendra Singh Rawat
2022-03-04 12:08 ` [PATCH 2/3] net/qede: fix Rx callback Devendra Singh Rawat
2022-03-04 12:08 ` [PATCH 3/3] net/qede: fix max Rx pktlen calculation Devendra Singh Rawat
@ 2022-03-10 7:40 ` Jerin Jacob
2022-03-10 8:56 ` Thomas Monjalon
2 siblings, 1 reply; 6+ messages in thread
From: Jerin Jacob @ 2022-03-10 7:40 UTC (permalink / raw)
To: Devendra Singh Rawat
Cc: dpdk-dev, Thomas Monjalon, Jerin Jacob, Ferruh Yigit,
Rasesh Mody, palok, dpdk stable
On Fri, Mar 4, 2022 at 5:38 PM Devendra Singh Rawat
<dsinghrawat@marvell.com> wrote:
>
> Tx completion routine was first incrementing no. of free slots in Tx
> ring and then freeing corresponding mbufs in bulk. In some situations
> no. of mbufs freed were less than no. of Tx ring slots freed. This
> caused TX ring to get into an inconsistent state and ultimately
> application fails to transmit further traffic.
>
> The fix first updates Tx ring SW consumer index, then increments Tx ring
> free slot no. and finally frees the mbuf, this is done in a single
> iteration of loop.
>
> Fixes: 2c41740bf19e ("net/qede: get consumer index once")
> Fixes: 4996b959cde6 ("net/qede: free packets in bulk")
> Cc: stable@dpdk.org
>
> Signed-off-by: Devendra Singh Rawat <dsinghrawat@marvell.com>
> Signed-off-by: Rasesh Mody <rmody@marvell.com>
Updated the git commits as follows and applied series to
dpdk-next-net-mrvl/for-next-net. Thanks
--
net/qede: fix max Rx packet length calculation
Size of CRC is not added to max_rx_pktlen, due to this bigger sized
packets(size 1480, 1490 1500) are being dropped.
This fix adds RTE_ETHER_CRC_LEN to max_rx_pktlen.
Fixes: 1bb4a528c41f ("ethdev: fix max Rx packet length")
Cc: stable@dpdk.org
Signed-off-by: Devendra Singh Rawat <dsinghrawat@marvell.com>
Signed-off-by: Rasesh Mody <rmody@marvell.com>
---
net/qede: fix Rx callback
qede_alloc_rx_bulk_mbufs() was trimming the number of requested
mbufs count to QEDE_MAX_BULK_ALLOC_COUNT.
The RX callback was ignorant of this trimming and it was always
resetting the number of empty Rx BD ring slots to 0.
This resulted in Rx BD ring getting into an inconsistent
state and ultimately the application fails to receive any traffic.
The fix trims the number of requested mbufs count before
making call to qede_alloc_rx_bulk_mbufs().
After qede_alloc_rx_bulk_mbufs() returns successfully, the
number of empty Rx BD ring slots are decremented by the
correct count.
Fixes: 8f2312474529 ("net/qede: fix performance bottleneck in Rx path")
Cc: stable@dpdk.org
Signed-off-by: Devendra Singh Rawat <dsinghrawat@marvell.com>
Signed-off-by: Rasesh Mody <rmody@marvell.com>
--
net/qede: fix Tx callback completion routine
Tx completion routine was first incrementing the number of free slots in
Tx ring and then freeing corresponding mbufs in bulk.
In some situations, the number of mbufs freed were less than
number of Tx ring slots freed. This caused Tx ring to get into an
inconsistent state and ultimately application fails to transmit
further traffic.
The fix first updates Tx ring SW consumer index, then increments
Tx ring free slot number and finally frees the mbuf,
this is done in a single iteration of loop.
Fixes: 2c41740bf19e ("net/qede: get consumer index once")
Fixes: 4996b959cde6 ("net/qede: free packets in bulk")
Cc: stable@dpdk.org
Signed-off-by: Devendra Singh Rawat <dsinghrawat@marvell.com>
Signed-off-by: Rasesh Mody <rmody@marvell.com>
---
> ---
> drivers/net/qede/qede_rxtx.c | 79 +++++++++++++++---------------------
> 1 file changed, 33 insertions(+), 46 deletions(-)
>
> diff --git a/drivers/net/qede/qede_rxtx.c b/drivers/net/qede/qede_rxtx.c
> index 911bb1a260..0c52568180 100644
> --- a/drivers/net/qede/qede_rxtx.c
> +++ b/drivers/net/qede/qede_rxtx.c
> @@ -885,68 +885,55 @@ qede_tx_queue_start(struct rte_eth_dev *eth_dev, uint16_t tx_queue_id)
> }
>
> static inline void
> -qede_process_tx_compl(__rte_unused struct ecore_dev *edev,
> - struct qede_tx_queue *txq)
> +qede_free_tx_pkt(struct qede_tx_queue *txq)
> {
> - uint16_t hw_bd_cons;
> - uint16_t sw_tx_cons;
> - uint16_t remaining;
> - uint16_t mask;
> struct rte_mbuf *mbuf;
> uint16_t nb_segs;
> uint16_t idx;
> - uint16_t first_idx;
> -
> - rte_compiler_barrier();
> - rte_prefetch0(txq->hw_cons_ptr);
> - sw_tx_cons = ecore_chain_get_cons_idx(&txq->tx_pbl);
> - hw_bd_cons = rte_le_to_cpu_16(*txq->hw_cons_ptr);
> -#ifdef RTE_LIBRTE_QEDE_DEBUG_TX
> - PMD_TX_LOG(DEBUG, txq, "Tx Completions = %u\n",
> - abs(hw_bd_cons - sw_tx_cons));
> -#endif
> -
> - mask = NUM_TX_BDS(txq);
> - idx = txq->sw_tx_cons & mask;
>
> - remaining = hw_bd_cons - sw_tx_cons;
> - txq->nb_tx_avail += remaining;
> - first_idx = idx;
> -
> - while (remaining) {
> - mbuf = txq->sw_tx_ring[idx];
> - RTE_ASSERT(mbuf);
> + idx = TX_CONS(txq);
> + mbuf = txq->sw_tx_ring[idx];
> + if (mbuf) {
> nb_segs = mbuf->nb_segs;
> - remaining -= nb_segs;
> -
> - /* Prefetch the next mbuf. Note that at least the last 4 mbufs
> - * that are prefetched will not be used in the current call.
> - */
> - rte_mbuf_prefetch_part1(txq->sw_tx_ring[(idx + 4) & mask]);
> - rte_mbuf_prefetch_part2(txq->sw_tx_ring[(idx + 4) & mask]);
> -
> PMD_TX_LOG(DEBUG, txq, "nb_segs to free %u\n", nb_segs);
> -
> while (nb_segs) {
> + /* It's like consuming rxbuf in recv() */
> ecore_chain_consume(&txq->tx_pbl);
> + txq->nb_tx_avail++;
> nb_segs--;
> }
> -
> - idx = (idx + 1) & mask;
> + rte_pktmbuf_free(mbuf);
> + txq->sw_tx_ring[idx] = NULL;
> + txq->sw_tx_cons++;
> PMD_TX_LOG(DEBUG, txq, "Freed tx packet\n");
> - }
> - txq->sw_tx_cons = idx;
> -
> - if (first_idx > idx) {
> - rte_pktmbuf_free_bulk(&txq->sw_tx_ring[first_idx],
> - mask - first_idx + 1);
> - rte_pktmbuf_free_bulk(&txq->sw_tx_ring[0], idx);
> } else {
> - rte_pktmbuf_free_bulk(&txq->sw_tx_ring[first_idx],
> - idx - first_idx);
> + ecore_chain_consume(&txq->tx_pbl);
> + txq->nb_tx_avail++;
> }
> }
>
> +static inline void
> +qede_process_tx_compl(__rte_unused struct ecore_dev *edev,
> + struct qede_tx_queue *txq)
> +{
> + uint16_t hw_bd_cons;
> +#ifdef RTE_LIBRTE_QEDE_DEBUG_TX
> + uint16_t sw_tx_cons;
> +#endif
> +
> + hw_bd_cons = rte_le_to_cpu_16(*txq->hw_cons_ptr);
> + /* read barrier prevents speculative execution on stale data */
> + rte_rmb();
> +
> +#ifdef RTE_LIBRTE_QEDE_DEBUG_TX
> + sw_tx_cons = ecore_chain_get_cons_idx(&txq->tx_pbl);
> + PMD_TX_LOG(DEBUG, txq, "Tx Completions = %u\n",
> + abs(hw_bd_cons - sw_tx_cons));
> +#endif
> + while (hw_bd_cons != ecore_chain_get_cons_idx(&txq->tx_pbl))
> + qede_free_tx_pkt(txq);
> +}
> +
> static int qede_drain_txq(struct qede_dev *qdev,
> struct qede_tx_queue *txq, bool allow_drain)
> {
> --
> 2.18.2
>
^ permalink raw reply [flat|nested] 6+ messages in thread