xref: /dpdk/drivers/net/virtio/virtio_rxtx_simple_neon.c (revision 3e0ceb9f17fff027fc6c8f18de35e11719ffa61e)
1 /*
2  *   BSD LICENSE
3  *
4  *   Copyright (C) Cavium, Inc. 2016
5  *
6  *   Redistribution and use in source and binary forms, with or without
7  *   modification, are permitted provided that the following conditions
8  *   are met:
9  *
10  *     * Redistributions of source code must retain the above copyright
11  *       notice, this list of conditions and the following disclaimer.
12  *     * Redistributions in binary form must reproduce the above copyright
13  *       notice, this list of conditions and the following disclaimer in
14  *       the documentation and/or other materials provided with the
15  *       distribution.
16  *     * Neither the name of Cavium, Inc nor the names of its
17  *       contributors may be used to endorse or promote products derived
18  *       from this software without specific prior written permission.
19  *
20  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
23  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
24  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
25  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
26  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
27  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
28  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
29  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
30  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
31 */
32 
33 #include <stdint.h>
34 #include <stdio.h>
35 #include <stdlib.h>
36 #include <string.h>
37 #include <errno.h>
38 
39 #include <rte_byteorder.h>
40 #include <rte_branch_prediction.h>
41 #include <rte_cycles.h>
42 #include <rte_ether.h>
43 #include <rte_ethdev.h>
44 #include <rte_errno.h>
45 #include <rte_memory.h>
46 #include <rte_mempool.h>
47 #include <rte_malloc.h>
48 #include <rte_mbuf.h>
49 #include <rte_prefetch.h>
50 #include <rte_string_fns.h>
51 #include <rte_vect.h>
52 
53 #include "virtio_rxtx_simple.h"
54 
55 #define RTE_VIRTIO_VPMD_RX_BURST 32
56 #define RTE_VIRTIO_DESC_PER_LOOP 8
57 #define RTE_VIRTIO_VPMD_RX_REARM_THRESH RTE_VIRTIO_VPMD_RX_BURST
58 
59 /* virtio vPMD receive routine, only accept(nb_pkts >= RTE_VIRTIO_DESC_PER_LOOP)
60  *
61  * This routine is for non-mergeable RX, one desc for each guest buffer.
62  * This routine is based on the RX ring layout optimization. Each entry in the
63  * avail ring points to the desc with the same index in the desc ring and this
64  * will never be changed in the driver.
65  *
66  * - nb_pkts < RTE_VIRTIO_DESC_PER_LOOP, just return no packet
67  */
68 uint16_t
69 virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
70 	uint16_t nb_pkts)
71 {
72 	struct virtnet_rx *rxvq = rx_queue;
73 	struct virtqueue *vq = rxvq->vq;
74 	struct virtio_hw *hw = vq->hw;
75 	uint16_t nb_used;
76 	uint16_t desc_idx;
77 	struct vring_used_elem *rused;
78 	struct rte_mbuf **sw_ring;
79 	struct rte_mbuf **sw_ring_end;
80 	uint16_t nb_pkts_received = 0;
81 
82 	uint8x16_t shuf_msk1 = {
83 		0xFF, 0xFF, 0xFF, 0xFF, /* packet type */
84 		4, 5, 0xFF, 0xFF,       /* pkt len */
85 		4, 5,                   /* dat len */
86 		0xFF, 0xFF,             /* vlan tci */
87 		0xFF, 0xFF, 0xFF, 0xFF
88 	};
89 
90 	uint8x16_t shuf_msk2 = {
91 		0xFF, 0xFF, 0xFF, 0xFF, /* packet type */
92 		12, 13, 0xFF, 0xFF,     /* pkt len */
93 		12, 13,                 /* dat len */
94 		0xFF, 0xFF,             /* vlan tci */
95 		0xFF, 0xFF, 0xFF, 0xFF
96 	};
97 
98 	/* Subtract the header length.
99 	 *  In which case do we need the header length in used->len ?
100 	 */
101 	uint16x8_t len_adjust = {
102 		0, 0,
103 		(uint16_t)vq->hw->vtnet_hdr_size, 0,
104 		(uint16_t)vq->hw->vtnet_hdr_size,
105 		0,
106 		0, 0
107 	};
108 
109 	if (unlikely(hw->started == 0))
110 		return nb_pkts_received;
111 
112 	if (unlikely(nb_pkts < RTE_VIRTIO_DESC_PER_LOOP))
113 		return 0;
114 
115 	nb_used = VIRTQUEUE_NUSED(vq);
116 
117 	rte_rmb();
118 
119 	if (unlikely(nb_used == 0))
120 		return 0;
121 
122 	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_VIRTIO_DESC_PER_LOOP);
123 	nb_used = RTE_MIN(nb_used, nb_pkts);
124 
125 	desc_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
126 	rused = &vq->vq_ring.used->ring[desc_idx];
127 	sw_ring  = &vq->sw_ring[desc_idx];
128 	sw_ring_end = &vq->sw_ring[vq->vq_nentries];
129 
130 	rte_prefetch_non_temporal(rused);
131 
132 	if (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
133 		virtio_rxq_rearm_vec(rxvq);
134 		if (unlikely(virtqueue_kick_prepare(vq)))
135 			virtqueue_notify(vq);
136 	}
137 
138 	for (nb_pkts_received = 0;
139 		nb_pkts_received < nb_used;) {
140 		uint64x2_t desc[RTE_VIRTIO_DESC_PER_LOOP / 2];
141 		uint64x2_t mbp[RTE_VIRTIO_DESC_PER_LOOP / 2];
142 		uint64x2_t pkt_mb[RTE_VIRTIO_DESC_PER_LOOP];
143 
144 		mbp[0] = vld1q_u64((uint64_t *)(sw_ring + 0));
145 		desc[0] = vld1q_u64((uint64_t *)(rused + 0));
146 		vst1q_u64((uint64_t *)&rx_pkts[0], mbp[0]);
147 
148 		mbp[1] = vld1q_u64((uint64_t *)(sw_ring + 2));
149 		desc[1] = vld1q_u64((uint64_t *)(rused + 2));
150 		vst1q_u64((uint64_t *)&rx_pkts[2], mbp[1]);
151 
152 		mbp[2] = vld1q_u64((uint64_t *)(sw_ring + 4));
153 		desc[2] = vld1q_u64((uint64_t *)(rused + 4));
154 		vst1q_u64((uint64_t *)&rx_pkts[4], mbp[2]);
155 
156 		mbp[3] = vld1q_u64((uint64_t *)(sw_ring + 6));
157 		desc[3] = vld1q_u64((uint64_t *)(rused + 6));
158 		vst1q_u64((uint64_t *)&rx_pkts[6], mbp[3]);
159 
160 		pkt_mb[1] = vreinterpretq_u64_u8(vqtbl1q_u8(
161 				vreinterpretq_u8_u64(desc[0]), shuf_msk2));
162 		pkt_mb[0] = vreinterpretq_u64_u8(vqtbl1q_u8(
163 				vreinterpretq_u8_u64(desc[0]), shuf_msk1));
164 		pkt_mb[1] = vreinterpretq_u64_u16(vsubq_u16(
165 				vreinterpretq_u16_u64(pkt_mb[1]), len_adjust));
166 		pkt_mb[0] = vreinterpretq_u64_u16(vsubq_u16(
167 				vreinterpretq_u16_u64(pkt_mb[0]), len_adjust));
168 		vst1q_u64((void *)&rx_pkts[1]->rx_descriptor_fields1,
169 			pkt_mb[1]);
170 		vst1q_u64((void *)&rx_pkts[0]->rx_descriptor_fields1,
171 			pkt_mb[0]);
172 
173 		pkt_mb[3] = vreinterpretq_u64_u8(vqtbl1q_u8(
174 				vreinterpretq_u8_u64(desc[1]), shuf_msk2));
175 		pkt_mb[2] = vreinterpretq_u64_u8(vqtbl1q_u8(
176 				vreinterpretq_u8_u64(desc[1]), shuf_msk1));
177 		pkt_mb[3] = vreinterpretq_u64_u16(vsubq_u16(
178 				vreinterpretq_u16_u64(pkt_mb[3]), len_adjust));
179 		pkt_mb[2] = vreinterpretq_u64_u16(vsubq_u16(
180 				vreinterpretq_u16_u64(pkt_mb[2]), len_adjust));
181 		vst1q_u64((void *)&rx_pkts[3]->rx_descriptor_fields1,
182 			pkt_mb[3]);
183 		vst1q_u64((void *)&rx_pkts[2]->rx_descriptor_fields1,
184 			pkt_mb[2]);
185 
186 		pkt_mb[5] = vreinterpretq_u64_u8(vqtbl1q_u8(
187 				vreinterpretq_u8_u64(desc[2]), shuf_msk2));
188 		pkt_mb[4] = vreinterpretq_u64_u8(vqtbl1q_u8(
189 				vreinterpretq_u8_u64(desc[2]), shuf_msk1));
190 		pkt_mb[5] = vreinterpretq_u64_u16(vsubq_u16(
191 				vreinterpretq_u16_u64(pkt_mb[5]), len_adjust));
192 		pkt_mb[4] = vreinterpretq_u64_u16(vsubq_u16(
193 				vreinterpretq_u16_u64(pkt_mb[4]), len_adjust));
194 		vst1q_u64((void *)&rx_pkts[5]->rx_descriptor_fields1,
195 			pkt_mb[5]);
196 		vst1q_u64((void *)&rx_pkts[4]->rx_descriptor_fields1,
197 			pkt_mb[4]);
198 
199 		pkt_mb[7] = vreinterpretq_u64_u8(vqtbl1q_u8(
200 				vreinterpretq_u8_u64(desc[3]), shuf_msk2));
201 		pkt_mb[6] = vreinterpretq_u64_u8(vqtbl1q_u8(
202 				vreinterpretq_u8_u64(desc[3]), shuf_msk1));
203 		pkt_mb[7] = vreinterpretq_u64_u16(vsubq_u16(
204 				vreinterpretq_u16_u64(pkt_mb[7]), len_adjust));
205 		pkt_mb[6] = vreinterpretq_u64_u16(vsubq_u16(
206 				vreinterpretq_u16_u64(pkt_mb[6]), len_adjust));
207 		vst1q_u64((void *)&rx_pkts[7]->rx_descriptor_fields1,
208 			pkt_mb[7]);
209 		vst1q_u64((void *)&rx_pkts[6]->rx_descriptor_fields1,
210 			pkt_mb[6]);
211 
212 		if (unlikely(nb_used <= RTE_VIRTIO_DESC_PER_LOOP)) {
213 			if (sw_ring + nb_used <= sw_ring_end)
214 				nb_pkts_received += nb_used;
215 			else
216 				nb_pkts_received += sw_ring_end - sw_ring;
217 			break;
218 		} else {
219 			if (unlikely(sw_ring + RTE_VIRTIO_DESC_PER_LOOP >=
220 				sw_ring_end)) {
221 				nb_pkts_received += sw_ring_end - sw_ring;
222 				break;
223 			} else {
224 				nb_pkts_received += RTE_VIRTIO_DESC_PER_LOOP;
225 
226 				rx_pkts += RTE_VIRTIO_DESC_PER_LOOP;
227 				sw_ring += RTE_VIRTIO_DESC_PER_LOOP;
228 				rused   += RTE_VIRTIO_DESC_PER_LOOP;
229 				nb_used -= RTE_VIRTIO_DESC_PER_LOOP;
230 			}
231 		}
232 	}
233 
234 	vq->vq_used_cons_idx += nb_pkts_received;
235 	vq->vq_free_cnt += nb_pkts_received;
236 	rxvq->stats.packets += nb_pkts_received;
237 	return nb_pkts_received;
238 }
239