d/numam-dpdk
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity
eab3ca4858
numam-dpdk/drivers/net/mlx5/mlx5_rxtx_vec.c
Alexander Kozyrev 5c68764377 net/mlx5: improve vectorized MPRQ descriptors locality 
There is a performance penalty for the replenish scheme
used in vectorized Rx burst for both MPRQ and SPRQ.
Mbuf elements are being filled at the end of the mbufs
array and being replenished at the beginning. That leads
to an increase in cache misses and the performance drop.
The more Rx descriptors are used the worse the situation.

Change the allocation scheme for vectorized MPRQ Rx burst:
allocate new mbufs only when consumed mbufs are almost
depleted (always have one burst gap between allocated and
consumed indices). Keeping a small number of mbufs allocated
improves cache locality and improves performance a lot.

Unfortunately, this approach cannot be applied to SPRQ Rx
burst routine. In MPRQ Rx burst we simply copy packets from
external MPRQ buffers or attach these buffers to mbufs.
In SPRQ Rx burst we allow the NIC to fill mbufs for us.
Hence keeping a small number of allocated mbufs will limit
NIC ability to fill as many buffers as possible. This fact
offsets the advantage of better cache locality.

Fixes: 0f20acbf5e ("net/mlx5: implement vectorized MPRQ burst")

Signed-off-by: Alexander Kozyrev <akozyrev@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>
2020-11-13 23:36:14 +01:00

578 lines
16 KiB
C
Raw Blame History

/* SPDX-License-Identifier: BSD-3-Clause
* Copyright 2017 6WIND S.A.
* Copyright 2017 Mellanox Technologies, Ltd
*/
#include <stdint.h>
#include <string.h>
#include <stdlib.h>
#include <rte_mbuf.h>
#include <rte_mempool.h>
#include <rte_prefetch.h>
#include <rte_vect.h>
#include <mlx5_glue.h>
#include <mlx5_prm.h>
#include "mlx5_defs.h"
#include "mlx5.h"
#include "mlx5_utils.h"
#include "mlx5_rxtx.h"
#include "mlx5_rxtx_vec.h"
#include "mlx5_autoconf.h"
#if defined RTE_ARCH_X86_64
#include "mlx5_rxtx_vec_sse.h"
#elif defined RTE_ARCH_ARM64
#include "mlx5_rxtx_vec_neon.h"
#elif defined RTE_ARCH_PPC_64
#include "mlx5_rxtx_vec_altivec.h"
#else
#error "This should not be compiled if SIMD instructions are not supported."
#endif
/**
* Skip error packets.
*
* @param rxq
* Pointer to RX queue structure.
* @param[out] pkts
* Array to store received packets.
* @param pkts_n
* Maximum number of packets in array.
*
* @return
* Number of packets successfully received (<= pkts_n).
*/
static uint16_t
rxq_handle_pending_error(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts,
uint16_t pkts_n)
{
uint16_t n = 0;
unsigned int i;
#ifdef MLX5_PMD_SOFT_COUNTERS
uint32_t err_bytes = 0;
#endif
for (i = 0; i < pkts_n; ++i) {
struct rte_mbuf *pkt = pkts[i];
if (pkt->packet_type == RTE_PTYPE_ALL_MASK || rxq->err_state) {
#ifdef MLX5_PMD_SOFT_COUNTERS
err_bytes += PKT_LEN(pkt);
#endif
rte_pktmbuf_free_seg(pkt);
} else {
pkts[n++] = pkt;
}
}
rxq->stats.idropped += (pkts_n - n);
#ifdef MLX5_PMD_SOFT_COUNTERS
/* Correct counters of errored completions. */
rxq->stats.ipackets -= (pkts_n - n);
rxq->stats.ibytes -= err_bytes;
#endif
mlx5_rx_err_handle(rxq, 1);
return n;
}
/**
* Replenish buffers for RX in bulk.
*
* @param rxq
* Pointer to RX queue structure.
*/
static inline void
mlx5_rx_replenish_bulk_mbuf(struct mlx5_rxq_data *rxq)
{
const uint16_t q_n = 1 << rxq->elts_n;
const uint16_t q_mask = q_n - 1;
uint16_t n = q_n - (rxq->rq_ci - rxq->rq_pi);
uint16_t elts_idx = rxq->rq_ci & q_mask;
struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
volatile struct mlx5_wqe_data_seg *wq =
&((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[elts_idx];
unsigned int i;
if (n >= rxq->rq_repl_thresh) {
MLX5_ASSERT(n >= MLX5_VPMD_RXQ_RPLNSH_THRESH(q_n));
MLX5_ASSERT(MLX5_VPMD_RXQ_RPLNSH_THRESH(q_n) >
MLX5_VPMD_DESCS_PER_LOOP);
/* Not to cross queue end. */
n = RTE_MIN(n - MLX5_VPMD_DESCS_PER_LOOP, q_n - elts_idx);
if (rte_mempool_get_bulk(rxq->mp, (void *)elts, n) < 0) {
rxq->stats.rx_nombuf += n;
return;
}
for (i = 0; i < n; ++i) {
void *buf_addr;
/*
* In order to support the mbufs with external attached
* data buffer we should use the buf_addr pointer
* instead of rte_mbuf_buf_addr(). It touches the mbuf
* itself and may impact the performance.
*/
buf_addr = elts[i]->buf_addr;
wq[i].addr = rte_cpu_to_be_64((uintptr_t)buf_addr +
RTE_PKTMBUF_HEADROOM);
/* If there's a single MR, no need to replace LKey. */
if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh)
> 1))
wq[i].lkey = mlx5_rx_mb2mr(rxq, elts[i]);
}
rxq->rq_ci += n;
/* Prevent overflowing into consumed mbufs. */
elts_idx = rxq->rq_ci & q_mask;
for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
(*rxq->elts)[elts_idx + i] = &rxq->fake_mbuf;
rte_io_wmb();
*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
}
}
/**
* Replenish buffers for MPRQ RX in bulk.
*
* @param rxq
* Pointer to RX queue structure.
*/
static inline void
mlx5_rx_mprq_replenish_bulk_mbuf(struct mlx5_rxq_data *rxq)
{
const uint16_t wqe_n = 1 << rxq->elts_n;
const uint32_t strd_n = 1 << rxq->strd_num_n;
const uint32_t elts_n = wqe_n * strd_n;
const uint32_t wqe_mask = elts_n - 1;
uint32_t n = rxq->elts_ci - rxq->rq_pi;
uint32_t elts_idx = rxq->elts_ci & wqe_mask;
struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
if (n <= rxq->rq_repl_thresh) {
MLX5_ASSERT(n + MLX5_VPMD_RX_MAX_BURST >=
MLX5_VPMD_RXQ_RPLNSH_THRESH(elts_n));
MLX5_ASSERT(MLX5_VPMD_RXQ_RPLNSH_THRESH(elts_n) >
MLX5_VPMD_DESCS_PER_LOOP);
/* Not to cross queue end. */
n = RTE_MIN(n + MLX5_VPMD_RX_MAX_BURST, elts_n - elts_idx);
if (rte_mempool_get_bulk(rxq->mp, (void *)elts, n) < 0) {
rxq->stats.rx_nombuf += n;
return;
}
rxq->elts_ci += n;
}
}
/**
* Copy or attach MPRQ buffers to RX SW ring.
*
* @param rxq
* Pointer to RX queue structure.
* @param pkts
* Pointer to array of packets to be stored.
* @param pkts_n
* Number of packets to be stored.
*
* @return
* Number of packets successfully copied/attached (<= pkts_n).
*/
static inline uint16_t
rxq_copy_mprq_mbuf_v(struct mlx5_rxq_data *rxq,
struct rte_mbuf **pkts, uint16_t pkts_n)
{
const uint16_t wqe_n = 1 << rxq->elts_n;
const uint16_t wqe_mask = wqe_n - 1;
const uint16_t strd_sz = 1 << rxq->strd_sz_n;
const uint32_t strd_n = 1 << rxq->strd_num_n;
const uint32_t elts_n = wqe_n * strd_n;
const uint32_t elts_mask = elts_n - 1;
uint32_t elts_idx = rxq->rq_pi & elts_mask;
struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
uint32_t rq_ci = rxq->rq_ci;
struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wqe_mask];
uint16_t copied = 0;
uint16_t i = 0;
for (i = 0; i < pkts_n; ++i) {
uint16_t strd_cnt;
enum mlx5_rqx_code rxq_code;
if (rxq->consumed_strd == strd_n) {
/* Replace WQE if the buffer is still in use. */
mprq_buf_replace(rxq, rq_ci & wqe_mask);
/* Advance to the next WQE. */
rxq->consumed_strd = 0;
rq_ci++;
buf = (*rxq->mprq_bufs)[rq_ci & wqe_mask];
}
if (!elts[i]->pkt_len) {
rxq->consumed_strd = strd_n;
rte_pktmbuf_free_seg(elts[i]);
#ifdef MLX5_PMD_SOFT_COUNTERS
rxq->stats.ipackets -= 1;
#endif
continue;
}
strd_cnt = (elts[i]->pkt_len / strd_sz) +
((elts[i]->pkt_len % strd_sz) ? 1 : 0);
rxq_code = mprq_buf_to_pkt(rxq, elts[i], elts[i]->pkt_len,
buf, rxq->consumed_strd, strd_cnt);
rxq->consumed_strd += strd_cnt;
if (unlikely(rxq_code != MLX5_RXQ_CODE_EXIT)) {
rte_pktmbuf_free_seg(elts[i]);
#ifdef MLX5_PMD_SOFT_COUNTERS
rxq->stats.ipackets -= 1;
rxq->stats.ibytes -= elts[i]->pkt_len;
#endif
if (rxq_code == MLX5_RXQ_CODE_NOMBUF) {
++rxq->stats.rx_nombuf;
break;
}
if (rxq_code == MLX5_RXQ_CODE_DROPPED) {
++rxq->stats.idropped;
continue;
}
}
pkts[copied++] = elts[i];
}
rxq->rq_pi += i;
rxq->cq_ci += i;
rte_io_wmb();
*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
if (rq_ci != rxq->rq_ci) {
rxq->rq_ci = rq_ci;
rte_io_wmb();
*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
}
return copied;
}
/**
* Receive burst of packets. An errored completion also consumes a mbuf, but the
* packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
* before returning to application.
*
* @param rxq
* Pointer to RX queue structure.
* @param[out] pkts
* Array to store received packets.
* @param pkts_n
* Maximum number of packets in array.
* @param[out] err
* Pointer to a flag. Set non-zero value if pkts array has at least one error
* packet to handle.
* @param[out] no_cq
* Pointer to a boolean. Set true if no new CQE seen.
*
* @return
* Number of packets received including errors (<= pkts_n).
*/
static inline uint16_t
rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts,
uint16_t pkts_n, uint64_t *err, bool *no_cq)
{
const uint16_t q_n = 1 << rxq->cqe_n;
const uint16_t q_mask = q_n - 1;
const uint16_t e_n = 1 << rxq->elts_n;
const uint16_t e_mask = e_n - 1;
volatile struct mlx5_cqe *cq;
struct rte_mbuf **elts;
uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
uint16_t nocmp_n = 0;
uint16_t rcvd_pkt = 0;
unsigned int cq_idx = rxq->cq_ci & q_mask;
unsigned int elts_idx;
MLX5_ASSERT(rxq->sges_n == 0);
MLX5_ASSERT(rxq->cqe_n == rxq->elts_n);
cq = &(*rxq->cqes)[cq_idx];
rte_prefetch0(cq);
rte_prefetch0(cq + 1);
rte_prefetch0(cq + 2);
rte_prefetch0(cq + 3);
pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
mlx5_rx_replenish_bulk_mbuf(rxq);
/* See if there're unreturned mbufs from compressed CQE. */
rcvd_pkt = rxq->decompressed;
if (rcvd_pkt > 0) {
rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
rxq_copy_mbuf_v(&(*rxq->elts)[rxq->rq_pi & e_mask],
pkts, rcvd_pkt);
rxq->rq_pi += rcvd_pkt;
rxq->decompressed -= rcvd_pkt;
pkts += rcvd_pkt;
}
elts_idx = rxq->rq_pi & e_mask;
elts = &(*rxq->elts)[elts_idx];
/* Not to overflow pkts array. */
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) {
*no_cq = !rcvd_pkt;
return rcvd_pkt;
}
/* At this point, there shouldn't be any remaining packets. */
MLX5_ASSERT(rxq->decompressed == 0);
/* Process all the CQEs */
nocmp_n = rxq_cq_process_v(rxq, cq, elts, pkts, pkts_n, err, &comp_idx);
/* If no new CQE seen, return without updating cq_db. */
if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP)) {
*no_cq = true;
return rcvd_pkt;
}
/* Update the consumer indexes for non-compressed CQEs. */
MLX5_ASSERT(nocmp_n <= pkts_n);
rxq->cq_ci += nocmp_n;
rxq->rq_pi += nocmp_n;
rcvd_pkt += nocmp_n;
/* Decompress the last CQE if compressed. */
if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP) {
MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
rxq->decompressed = rxq_cq_decompress_v(rxq, &cq[nocmp_n],
&elts[nocmp_n]);
rxq->cq_ci += rxq->decompressed;
/* Return more packets if needed. */
if (nocmp_n < pkts_n) {
uint16_t n = rxq->decompressed;
n = RTE_MIN(n, pkts_n - nocmp_n);
rxq_copy_mbuf_v(&(*rxq->elts)[rxq->rq_pi & e_mask],
&pkts[nocmp_n], n);
rxq->rq_pi += n;
rcvd_pkt += n;
rxq->decompressed -= n;
}
}
rte_io_wmb();
*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
*no_cq = !rcvd_pkt;
return rcvd_pkt;
}
/**
* DPDK callback for vectorized RX.
*
* @param dpdk_rxq
* Generic pointer to RX queue structure.
* @param[out] pkts
* Array to store received packets.
* @param pkts_n
* Maximum number of packets in array.
*
* @return
* Number of packets successfully received (<= pkts_n).
*/
uint16_t
mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
{
struct mlx5_rxq_data *rxq = dpdk_rxq;
uint16_t nb_rx = 0;
uint16_t tn = 0;
uint64_t err = 0;
bool no_cq = false;
do {
nb_rx = rxq_burst_v(rxq, pkts + tn, pkts_n - tn,
&err, &no_cq);
if (unlikely(err | rxq->err_state))
nb_rx = rxq_handle_pending_error(rxq, pkts + tn, nb_rx);
tn += nb_rx;
if (unlikely(no_cq))
break;
} while (tn != pkts_n);
return tn;
}
/**
* Receive burst of packets. An errored completion also consumes a mbuf, but the
* packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
* before returning to application.
*
* @param rxq
* Pointer to RX queue structure.
* @param[out] pkts
* Array to store received packets.
* @param pkts_n
* Maximum number of packets in array.
* @param[out] err
* Pointer to a flag. Set non-zero value if pkts array has at least one error
* packet to handle.
* @param[out] no_cq
* Pointer to a boolean. Set true if no new CQE seen.
*
* @return
* Number of packets received including errors (<= pkts_n).
*/
static inline uint16_t
rxq_burst_mprq_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts,
uint16_t pkts_n, uint64_t *err, bool *no_cq)
{
const uint16_t q_n = 1 << rxq->cqe_n;
const uint16_t q_mask = q_n - 1;
const uint16_t wqe_n = 1 << rxq->elts_n;
const uint32_t strd_n = 1 << rxq->strd_num_n;
const uint32_t elts_n = wqe_n * strd_n;
const uint32_t elts_mask = elts_n - 1;
volatile struct mlx5_cqe *cq;
struct rte_mbuf **elts;
uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
uint16_t nocmp_n = 0;
uint16_t rcvd_pkt = 0;
uint16_t cp_pkt = 0;
unsigned int cq_idx = rxq->cq_ci & q_mask;
unsigned int elts_idx;
MLX5_ASSERT(rxq->sges_n == 0);
cq = &(*rxq->cqes)[cq_idx];
rte_prefetch0(cq);
rte_prefetch0(cq + 1);
rte_prefetch0(cq + 2);
rte_prefetch0(cq + 3);
pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
mlx5_rx_mprq_replenish_bulk_mbuf(rxq);
/* See if there're unreturned mbufs from compressed CQE. */
rcvd_pkt = rxq->decompressed;
if (rcvd_pkt > 0) {
rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
cp_pkt = rxq_copy_mprq_mbuf_v(rxq, pkts, rcvd_pkt);
rxq->decompressed -= rcvd_pkt;
pkts += cp_pkt;
}
elts_idx = rxq->rq_pi & elts_mask;
elts = &(*rxq->elts)[elts_idx];
/* Not to overflow pkts array. */
pkts_n = RTE_ALIGN_FLOOR(pkts_n - cp_pkt, MLX5_VPMD_DESCS_PER_LOOP);
/* Not to cross queue end. */
pkts_n = RTE_MIN(pkts_n, elts_n - elts_idx);
pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
/* Not to move past the allocated mbufs. */
pkts_n = RTE_MIN(pkts_n, rxq->elts_ci - rxq->rq_pi);
if (!pkts_n) {
*no_cq = !cp_pkt;
return cp_pkt;
}
/* At this point, there shouldn't be any remaining packets. */
MLX5_ASSERT(rxq->decompressed == 0);
/* Process all the CQEs */
nocmp_n = rxq_cq_process_v(rxq, cq, elts, pkts, pkts_n, err, &comp_idx);
/* If no new CQE seen, return without updating cq_db. */
if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP)) {
*no_cq = true;
return cp_pkt;
}
/* Update the consumer indexes for non-compressed CQEs. */
MLX5_ASSERT(nocmp_n <= pkts_n);
cp_pkt = rxq_copy_mprq_mbuf_v(rxq, pkts, nocmp_n);
rcvd_pkt += cp_pkt;
/* Decompress the last CQE if compressed. */
if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP) {
MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
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->decompressed;
n = RTE_MIN(n, pkts_n - nocmp_n);
cp_pkt = rxq_copy_mprq_mbuf_v(rxq, &pkts[cp_pkt], n);
rcvd_pkt += cp_pkt;
rxq->decompressed -= n;
}
}
*no_cq = !rcvd_pkt;
return rcvd_pkt;
}
/**
* DPDK callback for vectorized MPRQ RX.
*
* @param dpdk_rxq
* Generic pointer to RX queue structure.
* @param[out] pkts
* Array to store received packets.
* @param pkts_n
* Maximum number of packets in array.
*
* @return
* Number of packets successfully received (<= pkts_n).
*/
uint16_t
mlx5_rx_burst_mprq_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
{
struct mlx5_rxq_data *rxq = dpdk_rxq;
uint16_t nb_rx = 0;
uint16_t tn = 0;
uint64_t err = 0;
bool no_cq = false;
do {
nb_rx = rxq_burst_mprq_v(rxq, pkts + tn, pkts_n - tn,
&err, &no_cq);
if (unlikely(err | rxq->err_state))
nb_rx = rxq_handle_pending_error(rxq, pkts + tn, nb_rx);
tn += nb_rx;
if (unlikely(no_cq))
break;
} while (tn != pkts_n);
return tn;
}
/**
* Check a RX queue can support vectorized RX.
*
* @param rxq
* Pointer to RX queue.
*
* @return
* 1 if supported, negative errno value if not.
*/
int __rte_cold
mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq)
{
struct mlx5_rxq_ctrl *ctrl =
container_of(rxq, struct mlx5_rxq_ctrl, rxq);
if (!ctrl->priv->config.rx_vec_en || rxq->sges_n != 0)
return -ENOTSUP;
if (rxq->lro)
return -ENOTSUP;
return 1;
}
/**
* Check a device can support vectorized RX.
*
* @param dev
* Pointer to Ethernet device.
*
* @return
* 1 if supported, negative errno value if not.
*/
int __rte_cold
mlx5_check_vec_rx_support(struct rte_eth_dev *dev)
{
struct mlx5_priv *priv = dev->data->dev_private;
uint32_t i;
if (rte_vect_get_max_simd_bitwidth() < RTE_VECT_SIMD_128)
return -ENOTSUP;
if (!priv->config.rx_vec_en)
return -ENOTSUP;
/* All the configured queues should support. */
for (i = 0; i < priv->rxqs_n; ++i) {
struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
if (!rxq)
continue;
if (mlx5_rxq_check_vec_support(rxq) < 0)
break;
}
if (i != priv->rxqs_n)
return -ENOTSUP;
return 1;
}
Reference in New Issue View Git Blame Copy Permalink