d/numam-dpdk
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

net/mlx5: remove Rx queues indexes correlation

Browse Source 
There is a full correlation between the CQE indexes to the WQE indexes
in the vectorized Rx queues management.

When the RQ is inserted to the reset state, the correlation may break
because the HW starts the RQ polling from index 0 while the CQ polling
continues regularly.

As an arrangement to CQE errors handling, when the RQ can be reset,
the correlation dependence should be removed from all the Rx queues
index managements.

Remove the aforementioned dependence from the vectorized Rx burst
functions.

Cc: stable@dpdk.org

Signed-off-by: Matan Azrad <matan@mellanox.com>
Acked-by: Shahaf Shuler <shahafs@mellanox.com>
This commit is contained in:
Matan Azrad 2019-05-30 10:20:31 +00:00 committed by Ferruh Yigit
parent bd282035b3
commit 1c191691a6
4 changed files with 32 additions and 27 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
drivers/net/mlx5/mlx5_rxq.c
View File

@ -1073,6 +1073,7 @@ mlx5_rxq_ibv_new(struct rte_eth_dev *dev, uint16_t idx)
rxq_data->cq_uar = cq_info.cq_uar;
rxq_data->cqn = cq_info.cqn;
rxq_data->cq_arm_sn = 0;
rxq_data->decompressed = 0;
/* Update doorbell counter. */
rxq_data->rq_ci = wqe_n >> rxq_data->sges_n;
rte_wmb();

6
drivers/net/mlx5/mlx5_rxtx.h
View File

@ -101,11 +101,15 @@ struct mlx5_rxq_data {
uint32_t rq_pi;
uint32_t cq_ci;
uint16_t rq_repl_thresh; /* Threshold for buffer replenishment. */
union {
struct rxq_zip zip; /* Compressed context. */
uint16_t decompressed;
/* Number of ready mbufs decompressed from the CQ. */
};
struct mlx5_mr_ctrl mr_ctrl; /* MR control descriptor. */
uint16_t mprq_max_memcpy_len; /* Maximum size of packet to memcpy. */
volatile void *wqes;
volatile struct mlx5_cqe(*cqes)[];
struct rxq_zip zip; /* Compressed context. */
RTE_STD_C11
union {
struct rte_mbuf *(*elts)[];

26
drivers/net/mlx5/mlx5_rxtx_vec_neon.h
View File

@ -352,8 +352,11 @@ rxq_copy_mbuf_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t n)
* @param elts
* Pointer to SW ring to be filled. The first mbuf has to be pre-built from
* the title completion descriptor to be copied to the rest of mbufs.
*
* @return
* Number of mini-CQEs successfully decompressed.
*/
static inline void
static inline uint16_t
rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
struct rte_mbuf **elts)
{
@ -505,6 +508,7 @@ rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
rxq->stats.ibytes += rcvd_byte;
#endif
rxq->cq_ci += mcqe_n;
return mcqe_n;
}
/**
@ -729,24 +733,17 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
rte_prefetch_non_temporal(cq + 2);
rte_prefetch_non_temporal(cq + 3);
pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
/*
* Order of indexes:
* rq_ci >= cq_ci >= rq_pi
* Definition of indexes:
* rq_ci - cq_ci := # of buffers owned by HW (posted).
* cq_ci - rq_pi := # of buffers not returned to app (decompressed).
* N - (rq_ci - rq_pi) := # of buffers consumed (to be replenished).
*/
repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
if (repl_n >= rxq->rq_repl_thresh)
mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
/* See if there're unreturned mbufs from compressed CQE. */
rcvd_pkt = rxq->cq_ci - rxq->rq_pi;
rcvd_pkt = rxq->decompressed;
if (rcvd_pkt > 0) {
rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
rxq->rq_pi += rcvd_pkt;
pkts += rcvd_pkt;
rxq->decompressed -= rcvd_pkt;
}
elts_idx = rxq->rq_pi & q_mask;
elts = &(*rxq->elts)[elts_idx];
@ -754,10 +751,11 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
/* Not to cross queue end. */
pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
if (!pkts_n)
return rcvd_pkt;
/* At this point, there shouldn't be any remained packets. */
assert(rxq->rq_pi == rxq->cq_ci);
assert(rxq->decompressed == 0);
/*
* Note that vectors have reverse order - {v3, v2, v1, v0}, because
* there's no instruction to count trailing zeros. __builtin_clzl() is
@ -1003,15 +1001,17 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
/* Decompress the last CQE if compressed. */
if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
assert(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
rxq->decompressed = rxq_cq_decompress_v(rxq, &cq[nocmp_n],
&elts[nocmp_n]);
/* Return more packets if needed. */
if (nocmp_n < pkts_n) {
uint16_t n = rxq->cq_ci - rxq->rq_pi;
uint16_t n = rxq->decompressed;
n = RTE_MIN(n, pkts_n - nocmp_n);
rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
rxq->rq_pi += n;
rcvd_pkt += n;
rxq->decompressed -= n;
}
}
rte_compiler_barrier();

26
drivers/net/mlx5/mlx5_rxtx_vec_sse.h
View File

@ -349,8 +349,11 @@ rxq_copy_mbuf_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t n)
* @param elts
* Pointer to SW ring to be filled. The first mbuf has to be pre-built from
* the title completion descriptor to be copied to the rest of mbufs.
*
* @return
* Number of mini-CQEs successfully decompressed.
*/
static inline void
static inline uint16_t
rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
struct rte_mbuf **elts)
{
@ -486,6 +489,7 @@ rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
rxq->stats.ibytes += rcvd_byte;
#endif
rxq->cq_ci += mcqe_n;
return mcqe_n;
}
/**
@ -712,23 +716,16 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
rte_prefetch0(cq + 2);
rte_prefetch0(cq + 3);
pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
/*
* Order of indexes:
* rq_ci >= cq_ci >= rq_pi
* Definition of indexes:
* rq_ci - cq_ci := # of buffers owned by HW (posted).
* cq_ci - rq_pi := # of buffers not returned to app (decompressed).
* N - (rq_ci - rq_pi) := # of buffers consumed (to be replenished).
*/
repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
if (repl_n >= rxq->rq_repl_thresh)
mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
/* See if there're unreturned mbufs from compressed CQE. */
rcvd_pkt = rxq->cq_ci - rxq->rq_pi;
rcvd_pkt = rxq->decompressed;
if (rcvd_pkt > 0) {
rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
rxq->rq_pi += rcvd_pkt;
rxq->decompressed -= rcvd_pkt;
pkts += rcvd_pkt;
}
elts_idx = rxq->rq_pi & q_mask;
@ -737,10 +734,11 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
/* Not to cross queue end. */
pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
if (!pkts_n)
return rcvd_pkt;
/* At this point, there shouldn't be any remained packets. */
assert(rxq->rq_pi == rxq->cq_ci);
assert(rxq->decompressed == 0);
/*
* A. load first Qword (8bytes) in one loop.
* B. copy 4 mbuf pointers from elts ring to returing pkts.
@ -953,15 +951,17 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
/* Decompress the last CQE if compressed. */
if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
assert(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
rxq->decompressed = rxq_cq_decompress_v(rxq, &cq[nocmp_n],
&elts[nocmp_n]);
/* Return more packets if needed. */
if (nocmp_n < pkts_n) {
uint16_t n = rxq->cq_ci - rxq->rq_pi;
uint16_t n = rxq->decompressed;
n = RTE_MIN(n, pkts_n - nocmp_n);
rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
rxq->rq_pi += n;
rcvd_pkt += n;
rxq->decompressed -= n;
}
}
rte_compiler_barrier();