xref: /dpdk/drivers/net/virtio/virtio_rxtx_simple_neon.c (revision 4e30ead5e7ca886535e2b30632b2948d2aac1681)
1 /*
2  *   BSD LICENSE
3  *
4  *   Copyright (C) Cavium networks Ltd. 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 networks 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_memzone.h>
47 #include <rte_mempool.h>
48 #include <rte_malloc.h>
49 #include <rte_mbuf.h>
50 #include <rte_prefetch.h>
51 #include <rte_string_fns.h>
52 #include <rte_vect.h>
53 
54 #include "virtio_rxtx_simple.h"
55 
56 #define RTE_VIRTIO_VPMD_RX_BURST 32
57 #define RTE_VIRTIO_DESC_PER_LOOP 8
58 #define RTE_VIRTIO_VPMD_RX_REARM_THRESH RTE_VIRTIO_VPMD_RX_BURST
59 
60 /* virtio vPMD receive routine, only accept(nb_pkts >= RTE_VIRTIO_DESC_PER_LOOP)
61  *
62  * This routine is for non-mergeable RX, one desc for each guest buffer.
63  * This routine is based on the RX ring layout optimization. Each entry in the
64  * avail ring points to the desc with the same index in the desc ring and this
65  * will never be changed in the driver.
66  *
67  * - nb_pkts < RTE_VIRTIO_DESC_PER_LOOP, just return no packet
68  */
69 uint16_t
70 virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
71 	uint16_t nb_pkts)
72 {
73 	struct virtnet_rx *rxvq = rx_queue;
74 	struct virtqueue *vq = rxvq->vq;
75 	struct virtio_hw *hw = vq->hw;
76 	uint16_t nb_used;
77 	uint16_t desc_idx;
78 	struct vring_used_elem *rused;
79 	struct rte_mbuf **sw_ring;
80 	struct rte_mbuf **sw_ring_end;
81 	uint16_t nb_pkts_received = 0;
82 
83 	uint8x16_t shuf_msk1 = {
84 		0xFF, 0xFF, 0xFF, 0xFF, /* packet type */
85 		4, 5, 0xFF, 0xFF,       /* pkt len */
86 		4, 5,                   /* dat len */
87 		0xFF, 0xFF,             /* vlan tci */
88 		0xFF, 0xFF, 0xFF, 0xFF
89 	};
90 
91 	uint8x16_t shuf_msk2 = {
92 		0xFF, 0xFF, 0xFF, 0xFF, /* packet type */
93 		12, 13, 0xFF, 0xFF,     /* pkt len */
94 		12, 13,                 /* dat len */
95 		0xFF, 0xFF,             /* vlan tci */
96 		0xFF, 0xFF, 0xFF, 0xFF
97 	};
98 
99 	/* Subtract the header length.
100 	 *  In which case do we need the header length in used->len ?
101 	 */
102 	uint16x8_t len_adjust = {
103 		0, 0,
104 		(uint16_t)vq->hw->vtnet_hdr_size, 0,
105 		(uint16_t)vq->hw->vtnet_hdr_size,
106 		0,
107 		0, 0
108 	};
109 
110 	if (unlikely(hw->started == 0))
111 		return nb_pkts_received;
112 
113 	if (unlikely(nb_pkts < RTE_VIRTIO_DESC_PER_LOOP))
114 		return 0;
115 
116 	nb_used = VIRTQUEUE_NUSED(vq);
117 
118 	rte_rmb();
119 
120 	if (unlikely(nb_used == 0))
121 		return 0;
122 
123 	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_VIRTIO_DESC_PER_LOOP);
124 	nb_used = RTE_MIN(nb_used, nb_pkts);
125 
126 	desc_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
127 	rused = &vq->vq_ring.used->ring[desc_idx];
128 	sw_ring  = &vq->sw_ring[desc_idx];
129 	sw_ring_end = &vq->sw_ring[vq->vq_nentries];
130 
131 	rte_prefetch_non_temporal(rused);
132 
133 	if (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
134 		virtio_rxq_rearm_vec(rxvq);
135 		if (unlikely(virtqueue_kick_prepare(vq)))
136 			virtqueue_notify(vq);
137 	}
138 
139 	for (nb_pkts_received = 0;
140 		nb_pkts_received < nb_used;) {
141 		uint64x2_t desc[RTE_VIRTIO_DESC_PER_LOOP / 2];
142 		uint64x2_t mbp[RTE_VIRTIO_DESC_PER_LOOP / 2];
143 		uint64x2_t pkt_mb[RTE_VIRTIO_DESC_PER_LOOP];
144 
145 		mbp[0] = vld1q_u64((uint64_t *)(sw_ring + 0));
146 		desc[0] = vld1q_u64((uint64_t *)(rused + 0));
147 		vst1q_u64((uint64_t *)&rx_pkts[0], mbp[0]);
148 
149 		mbp[1] = vld1q_u64((uint64_t *)(sw_ring + 2));
150 		desc[1] = vld1q_u64((uint64_t *)(rused + 2));
151 		vst1q_u64((uint64_t *)&rx_pkts[2], mbp[1]);
152 
153 		mbp[2] = vld1q_u64((uint64_t *)(sw_ring + 4));
154 		desc[2] = vld1q_u64((uint64_t *)(rused + 4));
155 		vst1q_u64((uint64_t *)&rx_pkts[4], mbp[2]);
156 
157 		mbp[3] = vld1q_u64((uint64_t *)(sw_ring + 6));
158 		desc[3] = vld1q_u64((uint64_t *)(rused + 6));
159 		vst1q_u64((uint64_t *)&rx_pkts[6], mbp[3]);
160 
161 		pkt_mb[1] = vreinterpretq_u64_u8(vqtbl1q_u8(
162 				vreinterpretq_u8_u64(desc[0]), shuf_msk2));
163 		pkt_mb[0] = vreinterpretq_u64_u8(vqtbl1q_u8(
164 				vreinterpretq_u8_u64(desc[0]), shuf_msk1));
165 		pkt_mb[1] = vreinterpretq_u64_u16(vsubq_u16(
166 				vreinterpretq_u16_u64(pkt_mb[1]), len_adjust));
167 		pkt_mb[0] = vreinterpretq_u64_u16(vsubq_u16(
168 				vreinterpretq_u16_u64(pkt_mb[0]), len_adjust));
169 		vst1q_u64((void *)&rx_pkts[1]->rx_descriptor_fields1,
170 			pkt_mb[1]);
171 		vst1q_u64((void *)&rx_pkts[0]->rx_descriptor_fields1,
172 			pkt_mb[0]);
173 
174 		pkt_mb[3] = vreinterpretq_u64_u8(vqtbl1q_u8(
175 				vreinterpretq_u8_u64(desc[1]), shuf_msk2));
176 		pkt_mb[2] = vreinterpretq_u64_u8(vqtbl1q_u8(
177 				vreinterpretq_u8_u64(desc[1]), shuf_msk1));
178 		pkt_mb[3] = vreinterpretq_u64_u16(vsubq_u16(
179 				vreinterpretq_u16_u64(pkt_mb[3]), len_adjust));
180 		pkt_mb[2] = vreinterpretq_u64_u16(vsubq_u16(
181 				vreinterpretq_u16_u64(pkt_mb[2]), len_adjust));
182 		vst1q_u64((void *)&rx_pkts[3]->rx_descriptor_fields1,
183 			pkt_mb[3]);
184 		vst1q_u64((void *)&rx_pkts[2]->rx_descriptor_fields1,
185 			pkt_mb[2]);
186 
187 		pkt_mb[5] = vreinterpretq_u64_u8(vqtbl1q_u8(
188 				vreinterpretq_u8_u64(desc[2]), shuf_msk2));
189 		pkt_mb[4] = vreinterpretq_u64_u8(vqtbl1q_u8(
190 				vreinterpretq_u8_u64(desc[2]), shuf_msk1));
191 		pkt_mb[5] = vreinterpretq_u64_u16(vsubq_u16(
192 				vreinterpretq_u16_u64(pkt_mb[5]), len_adjust));
193 		pkt_mb[4] = vreinterpretq_u64_u16(vsubq_u16(
194 				vreinterpretq_u16_u64(pkt_mb[4]), len_adjust));
195 		vst1q_u64((void *)&rx_pkts[5]->rx_descriptor_fields1,
196 			pkt_mb[5]);
197 		vst1q_u64((void *)&rx_pkts[4]->rx_descriptor_fields1,
198 			pkt_mb[4]);
199 
200 		pkt_mb[7] = vreinterpretq_u64_u8(vqtbl1q_u8(
201 				vreinterpretq_u8_u64(desc[3]), shuf_msk2));
202 		pkt_mb[6] = vreinterpretq_u64_u8(vqtbl1q_u8(
203 				vreinterpretq_u8_u64(desc[3]), shuf_msk1));
204 		pkt_mb[7] = vreinterpretq_u64_u16(vsubq_u16(
205 				vreinterpretq_u16_u64(pkt_mb[7]), len_adjust));
206 		pkt_mb[6] = vreinterpretq_u64_u16(vsubq_u16(
207 				vreinterpretq_u16_u64(pkt_mb[6]), len_adjust));
208 		vst1q_u64((void *)&rx_pkts[7]->rx_descriptor_fields1,
209 			pkt_mb[7]);
210 		vst1q_u64((void *)&rx_pkts[6]->rx_descriptor_fields1,
211 			pkt_mb[6]);
212 
213 		if (unlikely(nb_used <= RTE_VIRTIO_DESC_PER_LOOP)) {
214 			if (sw_ring + nb_used <= sw_ring_end)
215 				nb_pkts_received += nb_used;
216 			else
217 				nb_pkts_received += sw_ring_end - sw_ring;
218 			break;
219 		} else {
220 			if (unlikely(sw_ring + RTE_VIRTIO_DESC_PER_LOOP >=
221 				sw_ring_end)) {
222 				nb_pkts_received += sw_ring_end - sw_ring;
223 				break;
224 			} else {
225 				nb_pkts_received += RTE_VIRTIO_DESC_PER_LOOP;
226 
227 				rx_pkts += RTE_VIRTIO_DESC_PER_LOOP;
228 				sw_ring += RTE_VIRTIO_DESC_PER_LOOP;
229 				rused   += RTE_VIRTIO_DESC_PER_LOOP;
230 				nb_used -= RTE_VIRTIO_DESC_PER_LOOP;
231 			}
232 		}
233 	}
234 
235 	vq->vq_used_cons_idx += nb_pkts_received;
236 	vq->vq_free_cnt += nb_pkts_received;
237 	rxvq->stats.packets += nb_pkts_received;
238 	return nb_pkts_received;
239 }
240