* [dpdk-dev] [PATCH] net/mlx5: add vectorized Rx
@ 2017-05-28 15:40 Yongseok Koh
2017-05-30 10:45 ` Thomas Monjalon
0 siblings, 1 reply; 2+ messages in thread
From: Yongseok Koh @ 2017-05-28 15:40 UTC (permalink / raw)
To: ferruh.yigit; +Cc: dev, adrien.mazarguil, nelio.laranjeiro, Yongseok Koh
Vectorized Rx API for x86 is added.
Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
---
drivers/net/mlx5/Makefile | 1 +
drivers/net/mlx5/mlx5_defs.h | 6 +
drivers/net/mlx5/mlx5_ethdev.c | 8 +
drivers/net/mlx5/mlx5_rxq.c | 39 +++-
drivers/net/mlx5/mlx5_rxtx.c | 485 +++++++++++++++++++++++++++++++++++++++++
drivers/net/mlx5/mlx5_rxtx.h | 10 +
6 files changed, 548 insertions(+), 1 deletion(-)
diff --git a/drivers/net/mlx5/Makefile b/drivers/net/mlx5/Makefile
index c0799591b..02774e773 100644
--- a/drivers/net/mlx5/Makefile
+++ b/drivers/net/mlx5/Makefile
@@ -60,6 +60,7 @@ CFLAGS += -D_DEFAULT_SOURCE
CFLAGS += -D_XOPEN_SOURCE=600
CFLAGS += $(WERROR_FLAGS)
CFLAGS += -Wno-strict-prototypes
+CFLAGS += -msse4.1
LDLIBS += -libverbs
# A few warnings cannot be avoided in external headers.
diff --git a/drivers/net/mlx5/mlx5_defs.h b/drivers/net/mlx5/mlx5_defs.h
index 201bb3362..d2752b6a4 100644
--- a/drivers/net/mlx5/mlx5_defs.h
+++ b/drivers/net/mlx5/mlx5_defs.h
@@ -89,4 +89,10 @@
/* Maximum Packet headers size (L2+L3+L4) for TSO. */
#define MLX5_MAX_TSO_HEADER 128
+/* Parameters for Vectorized PMD */
+#undef MLX5_VECTORIZED_RX
+#define MLX5_VPMD_RXQ_RPLNSH_THRESH 32U
+#define MLX5_VPMD_MAX_RX_BURST MLX5_VPMD_RXQ_RPLNSH_THRESH
+#define MLX5_VPMD_DESCS_PER_LOOP 4
+
#endif /* RTE_PMD_MLX5_DEFS_H_ */
diff --git a/drivers/net/mlx5/mlx5_ethdev.c b/drivers/net/mlx5/mlx5_ethdev.c
index 3fd22cb85..8cd73c394 100644
--- a/drivers/net/mlx5/mlx5_ethdev.c
+++ b/drivers/net/mlx5/mlx5_ethdev.c
@@ -723,7 +723,11 @@ mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
};
+#ifdef MLX5_VECTORIZED_RX
+ if (dev->rx_pkt_burst == mlx5_rx_burst_vec)
+#else
if (dev->rx_pkt_burst == mlx5_rx_burst)
+#endif
return ptypes;
return NULL;
}
@@ -1605,5 +1609,9 @@ priv_select_tx_function(struct priv *priv)
void
priv_select_rx_function(struct priv *priv)
{
+#ifdef MLX5_VECTORIZED_RX
+ priv->dev->rx_pkt_burst = mlx5_rx_burst_vec;
+#else
priv->dev->rx_pkt_burst = mlx5_rx_burst;
+#endif
}
diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
index 8b7823360..4b3295607 100644
--- a/drivers/net/mlx5/mlx5_rxq.c
+++ b/drivers/net/mlx5/mlx5_rxq.c
@@ -652,7 +652,20 @@ rxq_alloc_elts(struct rxq_ctrl *rxq_ctrl, unsigned int elts_n,
const unsigned int sges_n = 1 << rxq_ctrl->rxq.sges_n;
unsigned int i;
int ret = 0;
-
+#ifdef MLX5_VECTORIZED_RX
+ uintptr_t p;
+ struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
+
+ /* Initialize default mbuf fields for vPMD */
+ mb_def.nb_segs = 1;
+ rte_pktmbuf_reset_headroom(&mb_def);
+ mb_def.port = rxq_ctrl->rxq.port_id;
+ rte_mbuf_refcnt_set(&mb_def, 1);
+ /* prevent compiler reordering: rearm_data covers previous fields */
+ rte_compiler_barrier();
+ p = (uintptr_t)&mb_def.rearm_data;
+ rxq_ctrl->rxq.mbuf_initializer = *(uint64_t *)p;
+#endif
/* Iterate on segments. */
for (i = 0; (i != elts_n); ++i) {
struct rte_mbuf *buf;
@@ -854,6 +867,9 @@ rxq_setup(struct rxq_ctrl *tmpl)
tmpl->rxq.cqe_n = log2above(ibcq->cqe);
tmpl->rxq.cq_ci = 0;
tmpl->rxq.rq_ci = 0;
+#ifdef MLX5_VECTORIZED_RX
+ tmpl->rxq.rq_pi = 0;
+#endif
tmpl->rxq.cq_db = cq->dbrec;
tmpl->rxq.wqes =
(volatile struct mlx5_wqe_data_seg (*)[])
@@ -908,6 +924,9 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
unsigned int mb_len = rte_pktmbuf_data_room_size(mp);
unsigned int cqe_n = desc - 1;
struct rte_mbuf *(*elts)[desc] = NULL;
+#ifdef MLX5_VECTORIZED_RX
+ int i;
+#endif
int ret = 0;
(void)conf; /* Thresholds configuration (ignored). */
@@ -987,7 +1006,11 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
if (priv->cqe_comp) {
attr.cq.comp_mask |= IBV_EXP_CQ_INIT_ATTR_FLAGS;
attr.cq.flags |= IBV_EXP_CQ_COMPRESSED_CQE;
+#ifdef MLX5_VECTORIZED_RX
+ cqe_n = desc - 1;
+#else
cqe_n = (desc * 2) - 1; /* Double the number of CQEs. */
+#endif
}
tmpl.cq = ibv_exp_create_cq(priv->ctx, cqe_n, NULL, tmpl.channel, 0,
&attr.cq);
@@ -1117,6 +1140,10 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
rte_free(tmpl.rxq.elts);
tmpl.rxq.elts = elts;
*rxq_ctrl = tmpl;
+#ifdef MLX5_VECTORIZED_RX
+ for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
+ (*elts)[desc + i] = &rxq_ctrl->rxq.fake_mbuf;
+#endif
/* Update doorbell counter. */
rxq_ctrl->rxq.rq_ci = desc >> rxq_ctrl->rxq.sges_n;
rte_wmb();
@@ -1192,7 +1219,12 @@ mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
if (rxq_ctrl->rxq.elts_n != log2above(desc)) {
rxq_ctrl = rte_realloc(rxq_ctrl,
sizeof(*rxq_ctrl) +
+#ifdef MLX5_VECTORIZED_RX
+ (desc + MLX5_VPMD_DESCS_PER_LOOP) *
+ sizeof(struct rte_mbuf *),
+#else
desc * sizeof(struct rte_mbuf *),
+#endif
RTE_CACHE_LINE_SIZE);
if (!rxq_ctrl) {
ERROR("%p: unable to reallocate queue index %u",
@@ -1203,7 +1235,12 @@ mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
}
} else {
rxq_ctrl = rte_calloc_socket("RXQ", 1, sizeof(*rxq_ctrl) +
+#ifdef MLX5_VECTORIZED_RX
+ (desc + MLX5_VPMD_DESCS_PER_LOOP) *
+ sizeof(struct rte_mbuf *),
+#else
desc * sizeof(struct rte_mbuf *),
+#endif
0, socket);
if (rxq_ctrl == NULL) {
ERROR("%p: unable to allocate queue index %u",
diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c
index 6254228a9..1d74ce5e1 100644
--- a/drivers/net/mlx5/mlx5_rxtx.c
+++ b/drivers/net/mlx5/mlx5_rxtx.c
@@ -1948,6 +1948,491 @@ rxq_cq_to_ol_flags(struct rxq *rxq, volatile struct mlx5_cqe *cqe)
return ol_flags;
}
+#ifdef MLX5_VECTORIZED_RX
+#if !defined (RTE_ARCH_X86_64)
+#error Currently supports only 64bit architectures
+#endif
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+static inline void
+rxq_copy_mbuf_v(struct rxq *rxq, struct rte_mbuf **pkts, uint16_t n)
+{
+ const uint16_t q_mask = (1 << rxq->elts_n) - 1;
+ struct rte_mbuf **elts = &(*rxq->elts)[rxq->rq_pi & q_mask];
+ unsigned int pos;
+ uint16_t p = n & -2;
+
+ for (pos = 0; pos < p; pos += 2) {
+ __m128i mbp;
+ mbp = _mm_loadu_si128((__m128i *)&elts[pos]);
+ _mm_storeu_si128((__m128i *)&pkts[pos], mbp);
+ }
+ if (n & 1)
+ pkts[pos] = elts[pos];
+}
+
+static inline void
+rxq_replenish_bulk_mbuf(struct rxq *rxq)
+{
+ const unsigned int q_mask = (1 << rxq->elts_n) - 1;
+ unsigned int elts_idx = rxq->rq_ci & q_mask;
+ struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
+ volatile struct mlx5_wqe_data_seg *wq = &(*rxq->wqes)[elts_idx];
+ unsigned int i;
+
+ if (rte_mempool_get_bulk(rxq->mp, (void *)elts,
+ MLX5_VPMD_RXQ_RPLNSH_THRESH) < 0) {
+ /* TODO: exception handling by fake_mbuf? */
+ rxq->stats.rx_nombuf += MLX5_VPMD_RXQ_RPLNSH_THRESH;
+ return;
+ }
+ for (i = 0; i < MLX5_VPMD_RXQ_RPLNSH_THRESH; i++)
+ wq[i].addr = htonll(rte_pktmbuf_mtod(elts[i], uintptr_t));
+ rxq->rq_ci += MLX5_VPMD_RXQ_RPLNSH_THRESH;
+ rte_wmb();
+ *rxq->rq_db = htonl(rxq->rq_ci);
+}
+
+static inline void
+rxq_cq_uncompress_v(struct rxq *rxq,
+ volatile struct mlx5_cqe *cq,
+ struct rte_mbuf **elts)
+{
+ volatile struct mlx5_mini_cqe8 *mcq = (void *)(cq + 1);
+ struct rte_mbuf *t_pkt = elts[0]; /* Title packet */
+ unsigned int pos;
+ uint16_t mcqe_n;
+ __m128i shuf_mask1, shuf_mask2;
+ __m128i rearm, rxdf;
+ const __m128i crc_adjust = _mm_set_epi16(
+ 0, 0, 0,
+ rxq->crc_present * (-ETHER_CRC_LEN),
+ 0,
+ rxq->crc_present * (-ETHER_CRC_LEN),
+ 0, 0
+ );
+
+ rearm = _mm_loadu_si128((__m128i *)&t_pkt->rearm_data);
+ rxdf = _mm_loadu_si128((__m128i *)&t_pkt->rx_descriptor_fields1);
+ /* mask to shuffle from extracted mini CQE to mbuf */
+ shuf_mask1 = _mm_set_epi8(
+ 0, 1, 2, 3, /* rss, bswap32 */
+ 0xFF, 0xFF, /* skip vlan_tci */
+ 6, 7, /* data_len, bswap16 */
+ 0xFF, 0xFF, 6, 7, /* pkt_len, bswap16 */
+ 0xFF, 0xFF, 0xFF, 0xFF /* skip packet_type */
+ );
+ shuf_mask2 = _mm_set_epi8(
+ 8, 9, 10, 11, /* rss, bswap32 */
+ 0xFF, 0xFF, /* skip vlan_tci */
+ 14, 15, /* data_len, bswap16 */
+ 0xFF, 0xFF, 14, 15, /* pkt_len, bswap16 */
+ 0xFF, 0xFF, 0xFF, 0xFF /* skip packet_type */
+ );
+ /* Restore the compressed count. Must be 16 bits. */
+ mcqe_n = t_pkt->data_len + (rxq->crc_present * ETHER_CRC_LEN);
+ /*
+ * A. load mCQEs into a 128bit register.
+ * B. store rearm data to mbuf.
+ * C. combine data from mCQEs with rx_descriptor_fields1.
+ * D. store rx_descriptor_fields1.
+ */
+ for (pos = 0; pos < mcqe_n; ) {
+ __m128i mcqe1, mcqe2;
+ __m128i rxdf1, rxdf2;
+
+ /* A.1 load mCQEs into a 128bit register. */
+ mcqe1 = _mm_loadu_si128((__m128i *)&mcq[pos % 8]);
+ mcqe2 = _mm_loadu_si128((__m128i *)&mcq[pos % 8 + 2]);
+ /* B.1 store rearm data to mbuf. */
+ _mm_storeu_si128((__m128i *)&elts[pos]->rearm_data, rearm);
+ _mm_storeu_si128((__m128i *)&elts[pos + 1]->rearm_data, rearm);
+ _mm_storeu_si128((__m128i *)&elts[pos + 2]->rearm_data, rearm);
+ _mm_storeu_si128((__m128i *)&elts[pos + 3]->rearm_data, rearm);
+ /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
+ rxdf1 = _mm_shuffle_epi8(mcqe1, shuf_mask1);
+ rxdf2 = _mm_shuffle_epi8(mcqe1, shuf_mask2);
+ rxdf1 = _mm_add_epi16(rxdf1, crc_adjust);
+ rxdf2 = _mm_add_epi16(rxdf2, crc_adjust);
+ rxdf1 = _mm_blend_epi16(rxdf1, rxdf, 0x23);
+ rxdf2 = _mm_blend_epi16(rxdf2, rxdf, 0x23);
+ /* D.1 store rx_descriptor_fields1. */
+ _mm_storeu_si128((__m128i *)&elts[pos]->rx_descriptor_fields1,
+ rxdf1);
+ _mm_storeu_si128((__m128i *)&elts[pos + 1]->rx_descriptor_fields1,
+ rxdf2);
+ /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
+ rxdf1 = _mm_shuffle_epi8(mcqe2, shuf_mask1);
+ rxdf2 = _mm_shuffle_epi8(mcqe2, shuf_mask2);
+ rxdf1 = _mm_add_epi16(rxdf1, crc_adjust);
+ rxdf2 = _mm_add_epi16(rxdf2, crc_adjust);
+ rxdf1 = _mm_blend_epi16(rxdf1, rxdf, 0x23);
+ rxdf2 = _mm_blend_epi16(rxdf2, rxdf, 0x23);
+ /* D.1 store rx_descriptor_fields1. */
+ _mm_storeu_si128((__m128i *)&elts[pos + 2]->rx_descriptor_fields1,
+ rxdf1);
+ _mm_storeu_si128((__m128i *)&elts[pos + 3]->rx_descriptor_fields1,
+ rxdf2);
+ /* TODO: need to copy hash.fdir.hi ?? */
+ pos += MLX5_VPMD_DESCS_PER_LOOP;
+ /* Move to next CQE */
+ if (!(pos & 0x7))
+ mcq = (void *)(cq + pos);
+ }
+ /* Invalidate consumed CQEs. */
+ for (pos = 1; pos < mcqe_n; ++pos)
+ cq[pos].op_own = MLX5_CQE_INVALIDATE;
+ rxq->cq_ci += mcqe_n;
+}
+
+static uint32_t mlx5_ptype_table[16] = {
+ RTE_PTYPE_UNKNOWN,
+ RTE_PTYPE_L3_IPV6_EXT_UNKNOWN, /* b0001 */
+ RTE_PTYPE_L3_IPV4_EXT_UNKNOWN, /* b0010 */
+ RTE_PTYPE_UNKNOWN, RTE_PTYPE_UNKNOWN,
+ RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
+ RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN, /* b0101 */
+ RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
+ RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN, /* b0110 */
+ RTE_PTYPE_UNKNOWN, RTE_PTYPE_UNKNOWN,
+ RTE_PTYPE_L3_IPV6_EXT_UNKNOWN, /* b1001 */
+ RTE_PTYPE_L3_IPV4_EXT_UNKNOWN, /* b1010 */
+ RTE_PTYPE_UNKNOWN, RTE_PTYPE_UNKNOWN,
+ RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
+ RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN, /* b1101 */
+ RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
+ RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN, /* b1110 */
+ RTE_PTYPE_UNKNOWN
+};
+
+
+static inline void
+rxq_cq_to_ptype_oflags_v(struct rxq *rxq, __m128i cqes[4], struct rte_mbuf **pkts)
+{
+ __m128i pinfo0;
+ __m128i pinfo1;
+ __m128i pinfo, ptype;
+ const __m128i ptype_mask = _mm_set_epi32(
+ 0xd07, 0xd07, 0xd07, 0xd07);
+ const __m128i pinfo_mask = _mm_set_epi32(
+ 0x3, 0x3, 0x3, 0x3);
+ const __m128i mbuf_init = _mm_loadl_epi64((__m128i *)&rxq->mbuf_initializer);
+ __m128i rearm0, rearm1, rearm2, rearm3;
+
+ /* Extract pkt_info field */
+ pinfo0 = _mm_unpacklo_epi32(cqes[0], cqes[1]);
+ pinfo1 = _mm_unpacklo_epi32(cqes[2], cqes[3]);
+ pinfo = _mm_unpacklo_epi64(pinfo0, pinfo1);
+ /* Extract hdr_type_etc field */
+ pinfo0 = _mm_unpackhi_epi32(cqes[0], cqes[1]);
+ pinfo1 = _mm_unpackhi_epi32(cqes[2], cqes[3]);
+ ptype = _mm_unpacklo_epi64(pinfo0, pinfo1);
+ /*
+ * Merge the two fields to generate the following
+ * bit[0] = l2_ok, bit[1] = l3_ok
+ * bit[2] = l4_ok,
+ * bit[8] = cv, bit[11:10] = l3_hdr_type
+ * bit[12] = tunneled, bit[13] = outer_l3_type
+ */
+ ptype = _mm_and_si128(ptype, ptype_mask);
+ pinfo = _mm_and_si128(pinfo, pinfo_mask);
+ pinfo = _mm_slli_epi32(pinfo, 12);
+ ptype = _mm_or_si128(ptype, pinfo);
+ pinfo = _mm_srli_epi32(ptype, 10);
+ pkts[0]->packet_type = mlx5_ptype_table[_mm_extract_epi8(ptype, 0)];
+ pkts[1]->packet_type = mlx5_ptype_table[_mm_extract_epi8(ptype, 4)];
+ pkts[2]->packet_type = mlx5_ptype_table[_mm_extract_epi8(ptype, 8)];
+ pkts[3]->packet_type = mlx5_ptype_table[_mm_extract_epi8(ptype, 12)];
+ /* TODO: Fill ol_flags */
+ rearm0 = _mm_blend_epi16(mbuf_init, mbuf_init, 0x01);
+ rearm1 = _mm_blend_epi16(mbuf_init, mbuf_init, 0x02);
+ rearm2 = _mm_blend_epi16(mbuf_init, mbuf_init, 0x04);
+ rearm3 = _mm_blend_epi16(mbuf_init, mbuf_init, 0x08);
+ _mm_store_si128((__m128i *)&pkts[0]->rearm_data, rearm0);
+ _mm_store_si128((__m128i *)&pkts[1]->rearm_data, rearm1);
+ _mm_store_si128((__m128i *)&pkts[2]->rearm_data, rearm2);
+ _mm_store_si128((__m128i *)&pkts[3]->rearm_data, rearm3);
+}
+
+/**
+ * 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 rxq *rxq = dpdk_rxq;
+ const uint16_t q_n = 1 << rxq->cqe_n;
+ const uint16_t q_mask = q_n - 1;
+ volatile struct mlx5_cqe *cq;
+ struct rte_mbuf **elts;
+ unsigned int pos;
+ 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;
+ unsigned int ownership = !!(rxq->cq_ci & (q_mask + 1));
+ __m128i owner_check, opcode_check, format_check;
+ __m128i shuf_mask, blend_mask;
+ const __m128i zero = _mm_setzero_si128();
+ const __m128i ones = _mm_cmpeq_epi32(zero, zero);
+ const __m128i crc_adjust = _mm_set_epi16(
+ 0, 0, 0, 0, 0,
+ rxq->crc_present * (-ETHER_CRC_LEN),
+ 0,
+ rxq->crc_present * (-ETHER_CRC_LEN)
+ );
+
+ /* TODO: Consider striding for multi-seg packet? */
+ assert(rxq->sges_n == 0);
+ 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_MAX_RX_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 (uncompressed).
+ * N - (rq_ci - rq_pi) := # of buffers consumed (to be replenished).
+ */
+ if ((uint16_t)(q_n - (rxq->rq_ci - rxq->rq_pi)) >=
+ MLX5_VPMD_RXQ_RPLNSH_THRESH)
+ rxq_replenish_bulk_mbuf(rxq);
+ /* See if there're unreturned mbufs from compressed CQE. */
+ rcvd_pkt = rxq->cq_ci - rxq->rq_pi;
+ 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;
+ }
+ elts_idx = rxq->rq_pi & q_mask;
+ elts = &(*rxq->elts)[elts_idx];
+ pkts_n = RTE_MIN((uint16_t)(pkts_n - rcvd_pkt), q_n - elts_idx);
+ /*
+ * TODO; Need to make pkts not overflow. May need to reserve a few NULL
+ * CQEs at the end of CQ.
+ *
+ * pkts_n = RTE_ALIGN_FLOOR(pkts_n, MLX5_VPMD_DESCS_PER_LOOP);
+ */
+ if (!pkts_n) {
+#ifdef MLX5_PMD_SOFT_COUNTERS
+ /* Increment packets counter. */
+ rxq->stats.ipackets += rcvd_pkt;
+#endif
+ return rcvd_pkt;
+ }
+ /* At this point, there shouldn't be any remained packets */
+ assert(rxq->rq_pi == rxq->cq_ci);
+ owner_check =
+ _mm_set_epi64x(0x0100000001000000LL, 0x0100000001000000LL);
+ opcode_check =
+ _mm_set_epi64x(0xf0000000f0000000LL, 0xf0000000f0000000LL);
+ format_check =
+ _mm_set_epi64x(0x0c0000000c000000LL, 0x0c0000000c000000LL);
+ /* Mask to blend from the last Qword to the first DQword */
+ blend_mask = _mm_set_epi8(
+ 0x80, 0x80, 0x80, 0x80,
+ 0x80, 0x80, 0x80, 0x80,
+ 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x80
+ );
+ /* Mask to shuffle from extracted CQE to mbuf */
+ shuf_mask = _mm_set_epi8(
+ 0xFF, 1, 2, 3, /* fdir.hi, bswap32 */
+ 12, 13, 14, 15, /* rss, bswap32 */
+ 10, 11, /* vlan_tci, bswap16 */
+ 4, 5, /* data_len, bswap16 */
+ 0xFF, 0xFF, /* zero out 2nd half of pkt_len */
+ 4, 5 /* pkt_len, bswap16 */
+ );
+ /*
+ * A. load first Qword (8bytes) in one loop.
+ * B. copy 4 mbuf pointers from elts ring to returing pkts.
+ * C. load remained CQE data and extract necessary fields.
+ * D. fill in mbuf.
+ * E. get vaild CQEs.
+ * F. find compressed CQE.
+ */
+ for (pos = 0; pos < pkts_n;
+ pos += MLX5_VPMD_DESCS_PER_LOOP) {
+ __m128i cqes[MLX5_VPMD_DESCS_PER_LOOP];
+ __m128i cqe_tmp1, cqe_tmp2;
+ __m128i pkt_mb0, pkt_mb1, pkt_mb2, pkt_mb3;
+ __m128i op_own, op_own_tmp1, op_own_tmp2;
+ __m128i owner_mask, invalid_mask, comp_mask;
+ __m128i mbp1, mbp2;
+ uint64_t n;
+
+ /* Prefetch next 4 CQEs */
+ rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP]);
+ rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 1]);
+ rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 2]);
+ rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 3]);
+ /* A.1 load cqes. */
+ cqes[3] = _mm_loadl_epi64((__m128i *)&cq[pos + 3].sop_drop_qpn);
+ rte_compiler_barrier();
+ cqes[2] = _mm_loadl_epi64((__m128i *)&cq[pos + 2].sop_drop_qpn);
+ rte_compiler_barrier();
+ /* B.1 load mbuf pointers. */
+ mbp1 = _mm_loadu_si128((__m128i *)&elts[pos]);
+ mbp2 = _mm_loadu_si128((__m128i *)&elts[pos + 2]);
+ /* A.1 load cqes. */
+ cqes[1] = _mm_loadl_epi64((__m128i *)&cq[pos + 1].sop_drop_qpn);
+ rte_compiler_barrier();
+ cqes[0] = _mm_loadl_epi64((__m128i *)&cq[pos].sop_drop_qpn);
+ /* B.2 copy mbuf pointers. */
+ _mm_storeu_si128((__m128i *)&pkts[pos], mbp1);
+ _mm_storeu_si128((__m128i *)&pkts[pos + 2], mbp2);
+ rte_compiler_barrier();
+ /* C.1 load remained CQE data and extract necessary fields. */
+ cqe_tmp2 = _mm_loadu_si128((__m128i *)&cq[pos + 3]);
+ cqe_tmp1 = _mm_loadu_si128((__m128i *)&cq[pos + 2]);
+ cqes[3] = _mm_blendv_epi8(cqes[3], cqe_tmp2, blend_mask);
+ cqes[2] = _mm_blendv_epi8(cqes[2], cqe_tmp1, blend_mask);
+ cqe_tmp2 = _mm_loadu_si128((__m128i *)&cq[pos + 3].rsvd1[3]);
+ cqe_tmp1 = _mm_loadu_si128((__m128i *)&cq[pos + 2].rsvd1[3]);
+ cqes[3] = _mm_blend_epi16(cqes[3], cqe_tmp2, 0x30);
+ cqes[2] = _mm_blend_epi16(cqes[2], cqe_tmp1, 0x30);
+ cqe_tmp2 = _mm_loadl_epi64((__m128i *)&cq[pos + 3].rsvd2[10]);
+ cqe_tmp1 = _mm_loadl_epi64((__m128i *)&cq[pos + 2].rsvd2[10]);
+ cqes[3] = _mm_blend_epi16(cqes[3], cqe_tmp2, 0x04);
+ cqes[2] = _mm_blend_epi16(cqes[2], cqe_tmp1, 0x04);
+ /* C.2 generate final structure for mbuf with swapping bytes */
+ pkt_mb3 = _mm_shuffle_epi8(cqes[3], shuf_mask);
+ pkt_mb2 = _mm_shuffle_epi8(cqes[2], shuf_mask);
+ /* C.3 adjust CRC length */
+ pkt_mb3 = _mm_add_epi16(pkt_mb3, crc_adjust);
+ pkt_mb2 = _mm_add_epi16(pkt_mb2, crc_adjust);
+ /* D.1 fill in mbuf - rx_descriptor_fields1. */
+ _mm_storeu_si128((void *)&pkts[pos + 3]->pkt_len, pkt_mb3);
+ _mm_storeu_si128((void *)&pkts[pos + 2]->pkt_len, pkt_mb2);
+ /* E.1 extract op_own field. */
+ op_own_tmp2 = _mm_unpacklo_epi32(cqes[2], cqes[3]);
+ /* C.1 load remained CQE data and extract necessary fields. */
+ cqe_tmp2 = _mm_loadu_si128((__m128i *)&cq[pos + 1]);
+ cqe_tmp1 = _mm_loadu_si128((__m128i *)&cq[pos]);
+ cqes[1] = _mm_blendv_epi8(cqes[1], cqe_tmp2, blend_mask);
+ cqes[0] = _mm_blendv_epi8(cqes[0], cqe_tmp1, blend_mask);
+ cqe_tmp2 = _mm_loadu_si128((__m128i *)&cq[pos + 1].rsvd1[3]);
+ cqe_tmp1 = _mm_loadu_si128((__m128i *)&cq[pos].rsvd1[3]);
+ cqes[1] = _mm_blend_epi16(cqes[1], cqe_tmp2, 0x30);
+ cqes[0] = _mm_blend_epi16(cqes[0], cqe_tmp1, 0x30);
+ cqe_tmp2 = _mm_loadl_epi64((__m128i *)&cq[pos + 1].rsvd2[10]);
+ cqe_tmp1 = _mm_loadl_epi64((__m128i *)&cq[pos].rsvd2[10]);
+ cqes[1] = _mm_blend_epi16(cqes[1], cqe_tmp2, 0x04);
+ cqes[0] = _mm_blend_epi16(cqes[0], cqe_tmp1, 0x04);
+ /* C.2 generate final structure for mbuf with swapping bytes */
+ pkt_mb1 = _mm_shuffle_epi8(cqes[1], shuf_mask);
+ pkt_mb0 = _mm_shuffle_epi8(cqes[0], shuf_mask);
+ /* C.3 adjust CRC length */
+ pkt_mb1 = _mm_add_epi16(pkt_mb1, crc_adjust);
+ pkt_mb0 = _mm_add_epi16(pkt_mb0, crc_adjust);
+ /* E.1 extract op_own byte. */
+ op_own_tmp1 = _mm_unpacklo_epi32(cqes[0], cqes[1]);
+ op_own = _mm_unpackhi_epi64(op_own_tmp1, op_own_tmp2);
+ /* D.1 fill in mbuf - rx_descriptor_fields1. */
+ _mm_storeu_si128((void *)&pkts[pos + 1]->pkt_len, pkt_mb1);
+ _mm_storeu_si128((void *)&pkts[pos]->pkt_len, pkt_mb0);
+ /* D.1 fill in mbuf - rearm_data and packet_type. */
+ rxq_cq_to_ptype_oflags_v(rxq, cqes, &pkts[pos]);
+ /* E.2 flip owner bit to mark invalid CQEs. */
+ owner_mask = _mm_and_si128(op_own, owner_check);
+ if (ownership)
+ owner_mask = _mm_xor_si128(owner_mask, owner_check);
+ owner_mask = _mm_cmpeq_epi32(owner_mask, owner_check);
+ owner_mask = _mm_packs_epi32(owner_mask, zero);
+ /* E.3 get mask for invalidated CQEs */
+ invalid_mask = _mm_cmpeq_epi32(opcode_check,
+ _mm_and_si128(op_own, opcode_check));
+ invalid_mask = _mm_packs_epi32(invalid_mask, zero);
+ /* E.4 mask out beyond boundary */
+ invalid_mask = _mm_or_si128(
+ invalid_mask,
+ ones <<
+ ((pkts_n - pos) * sizeof(uint16_t) * 8));
+ /* E.5 merge invalid_mask with invalid owner */
+ invalid_mask = _mm_or_si128(invalid_mask, owner_mask);
+ /* F.1 find compressed CQE format. */
+ comp_mask = _mm_and_si128(op_own, format_check);
+ comp_mask = _mm_cmpeq_epi32(comp_mask, format_check);
+ comp_mask = _mm_packs_epi32(comp_mask, zero);
+ /* F.2 mask out invalid entries */
+ comp_mask = _mm_andnot_si128(invalid_mask, comp_mask);
+ comp_idx = _mm_cvtsi128_si64(comp_mask);
+ /* F.3 get the first compressed CQE */
+ comp_idx = comp_idx ?
+ __builtin_ctzll(comp_idx) /
+ (sizeof(uint16_t) * 8) :
+ MLX5_VPMD_DESCS_PER_LOOP;
+ /* E.6 mask out entries after the compressed CQE */
+ invalid_mask = _mm_or_si128(
+ invalid_mask,
+ ones <<
+ (comp_idx * sizeof(uint16_t) * 8));
+ /* E.7 count non-compressed valid CQEs. */
+ n = _mm_cvtsi128_si64(invalid_mask);
+ n = n ? __builtin_ctzll(n) / (sizeof(uint16_t) * 8) :
+ MLX5_VPMD_DESCS_PER_LOOP;
+ nocmp_n += n;
+ if (likely(n != MLX5_VPMD_DESCS_PER_LOOP))
+ break;
+ }
+ /* If no new CQE seen, return without updating cq_db. */
+ if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
+ return rcvd_pkt;
+ /* Update the consumer indexes for non-compressed CQEs. */
+ assert(nocmp_n <= pkts_n);
+ rxq->cq_ci += nocmp_n;
+ rxq->rq_pi += nocmp_n;
+ rcvd_pkt += nocmp_n;
+ /* Uncompress the last CQE if compressed. */
+ if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP) {
+ assert(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
+ rxq_cq_uncompress_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;
+
+ n = RTE_MIN(n, pkts_n - nocmp_n);
+ rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
+ rxq->rq_pi += n;
+ rcvd_pkt += n;
+ }
+ }
+ rte_wmb();
+ *rxq->cq_db = htonl(rxq->cq_ci);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+ /* Increment packets counter. */
+ rxq->stats.ipackets += rcvd_pkt;
+#endif
+ return rcvd_pkt;
+}
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic pop
+#endif
+#endif /* MLX5_VECTORIZED_RX */
+
/**
* DPDK callback for RX.
*
diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
index 8db8eb144..424213503 100644
--- a/drivers/net/mlx5/mlx5_rxtx.h
+++ b/drivers/net/mlx5/mlx5_rxtx.h
@@ -119,6 +119,9 @@ struct rxq {
volatile uint32_t *rq_db;
volatile uint32_t *cq_db;
uint16_t rq_ci;
+#ifdef MLX5_VECTORIZED_RX
+ uint16_t rq_pi;
+#endif
uint16_t cq_ci;
volatile struct mlx5_wqe_data_seg(*wqes)[];
volatile struct mlx5_cqe(*cqes)[];
@@ -126,6 +129,10 @@ struct rxq {
struct rte_mbuf *(*elts)[];
struct rte_mempool *mp;
struct mlx5_rxq_stats stats;
+#ifdef MLX5_VECTORIZED_RX
+ uint64_t mbuf_initializer;
+ struct rte_mbuf fake_mbuf;
+#endif
} __rte_cache_aligned;
/* RX queue control descriptor. */
@@ -328,6 +335,9 @@ uint16_t mlx5_tx_burst(void *, struct rte_mbuf **, uint16_t);
uint16_t mlx5_tx_burst_mpw(void *, struct rte_mbuf **, uint16_t);
uint16_t mlx5_tx_burst_mpw_inline(void *, struct rte_mbuf **, uint16_t);
uint16_t mlx5_tx_burst_empw(void *, struct rte_mbuf **, uint16_t);
+#ifdef MLX5_VECTORIZED_RX
+uint16_t mlx5_rx_burst_vec(void *, struct rte_mbuf **, uint16_t);
+#endif
uint16_t mlx5_rx_burst(void *, struct rte_mbuf **, uint16_t);
uint16_t removed_tx_burst(void *, struct rte_mbuf **, uint16_t);
uint16_t removed_rx_burst(void *, struct rte_mbuf **, uint16_t);
--
2.11.0
^ permalink raw reply [flat|nested] 2+ messages in thread* Re: [dpdk-dev] [PATCH] net/mlx5: add vectorized Rx
2017-05-28 15:40 [dpdk-dev] [PATCH] net/mlx5: add vectorized Rx Yongseok Koh
@ 2017-05-30 10:45 ` Thomas Monjalon
0 siblings, 0 replies; 2+ messages in thread
From: Thomas Monjalon @ 2017-05-30 10:45 UTC (permalink / raw)
To: Yongseok Koh; +Cc: dev, ferruh.yigit, adrien.mazarguil, nelio.laranjeiro
28/05/2017 17:40, Yongseok Koh:
> Vectorized Rx API for x86 is added.
>
> Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
> ---
> drivers/net/mlx5/Makefile | 1 +
> drivers/net/mlx5/mlx5_defs.h | 6 +
> drivers/net/mlx5/mlx5_ethdev.c | 8 +
> drivers/net/mlx5/mlx5_rxq.c | 39 +++-
> drivers/net/mlx5/mlx5_rxtx.c | 485 +++++++++++++++++++++++++++++++++++++++++
> drivers/net/mlx5/mlx5_rxtx.h | 10 +
> 6 files changed, 548 insertions(+), 1 deletion(-)
It would be probably clearer to have a specific file for
Rx/Tx on SSE. So it would be compiled only for x86.
> --- a/drivers/net/mlx5/Makefile
> +++ b/drivers/net/mlx5/Makefile
> @@ -60,6 +60,7 @@ CFLAGS += -D_DEFAULT_SOURCE
> CFLAGS += -D_XOPEN_SOURCE=600
> CFLAGS += $(WERROR_FLAGS)
> CFLAGS += -Wno-strict-prototypes
> +CFLAGS += -msse4.1
> LDLIBS += -libverbs
[...]
> --- a/drivers/net/mlx5/mlx5_ethdev.c
> +++ b/drivers/net/mlx5/mlx5_ethdev.c
> @@ -723,7 +723,11 @@ mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
>
> };
>
> +#ifdef MLX5_VECTORIZED_RX
> + if (dev->rx_pkt_burst == mlx5_rx_burst_vec)
> +#else
> if (dev->rx_pkt_burst == mlx5_rx_burst)
> +#endif
> return ptypes;
> return NULL;
> }
Can we avoid an #ifdef here?
We should check SSE support at runtime with rte_cpu_get_flag_enabled().
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2017-05-30 10:45 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2017-05-28 15:40 [dpdk-dev] [PATCH] net/mlx5: add vectorized Rx Yongseok Koh
2017-05-30 10:45 ` Thomas Monjalon
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).