4 * Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of Intel Corporation nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40 #include <tmmintrin.h>
42 #include <rte_cycles.h>
43 #include <rte_memory.h>
44 #include <rte_memzone.h>
45 #include <rte_branch_prediction.h>
46 #include <rte_mempool.h>
47 #include <rte_malloc.h>
49 #include <rte_ether.h>
50 #include <rte_ethdev.h>
51 #include <rte_prefetch.h>
52 #include <rte_string_fns.h>
53 #include <rte_errno.h>
54 #include <rte_byteorder.h>
56 #include "virtio_logs.h"
57 #include "virtio_ethdev.h"
58 #include "virtqueue.h"
59 #include "virtio_rxtx.h"
61 #define RTE_VIRTIO_VPMD_RX_BURST 32
62 #define RTE_VIRTIO_DESC_PER_LOOP 8
63 #define RTE_VIRTIO_VPMD_RX_REARM_THRESH RTE_VIRTIO_VPMD_RX_BURST
65 #ifndef __INTEL_COMPILER
66 #pragma GCC diagnostic ignored "-Wcast-qual"
69 int __attribute__((cold))
70 virtqueue_enqueue_recv_refill_simple(struct virtqueue *vq,
71 struct rte_mbuf *cookie)
73 struct vq_desc_extra *dxp;
74 struct vring_desc *start_dp;
77 desc_idx = vq->vq_avail_idx & (vq->vq_nentries - 1);
78 dxp = &vq->vq_descx[desc_idx];
79 dxp->cookie = (void *)cookie;
80 vq->sw_ring[desc_idx] = cookie;
82 start_dp = vq->vq_ring.desc;
83 start_dp[desc_idx].addr = (uint64_t)((uintptr_t)cookie->buf_physaddr +
84 RTE_PKTMBUF_HEADROOM - sizeof(struct virtio_net_hdr));
85 start_dp[desc_idx].len = cookie->buf_len -
86 RTE_PKTMBUF_HEADROOM + sizeof(struct virtio_net_hdr);
95 virtio_rxq_rearm_vec(struct virtqueue *rxvq)
99 struct rte_mbuf **sw_ring;
100 struct vring_desc *start_dp;
103 desc_idx = rxvq->vq_avail_idx & (rxvq->vq_nentries - 1);
104 sw_ring = &rxvq->sw_ring[desc_idx];
105 start_dp = &rxvq->vq_ring.desc[desc_idx];
107 ret = rte_mempool_get_bulk(rxvq->mpool, (void **)sw_ring,
108 RTE_VIRTIO_VPMD_RX_REARM_THRESH);
110 rte_eth_devices[rxvq->port_id].data->rx_mbuf_alloc_failed +=
111 RTE_VIRTIO_VPMD_RX_REARM_THRESH;
115 for (i = 0; i < RTE_VIRTIO_VPMD_RX_REARM_THRESH; i++) {
118 p = (uintptr_t)&sw_ring[i]->rearm_data;
119 *(uint64_t *)p = rxvq->mbuf_initializer;
122 (uint64_t)((uintptr_t)sw_ring[i]->buf_physaddr +
123 RTE_PKTMBUF_HEADROOM - sizeof(struct virtio_net_hdr));
124 start_dp[i].len = sw_ring[i]->buf_len -
125 RTE_PKTMBUF_HEADROOM + sizeof(struct virtio_net_hdr);
128 rxvq->vq_avail_idx += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
129 rxvq->vq_free_cnt -= RTE_VIRTIO_VPMD_RX_REARM_THRESH;
130 vq_update_avail_idx(rxvq);
133 /* virtio vPMD receive routine, only accept(nb_pkts >= RTE_VIRTIO_DESC_PER_LOOP)
135 * This routine is for non-mergeable RX, one desc for each guest buffer.
136 * This routine is based on the RX ring layout optimization. Each entry in the
137 * avail ring points to the desc with the same index in the desc ring and this
138 * will never be changed in the driver.
140 * - nb_pkts < RTE_VIRTIO_DESC_PER_LOOP, just return no packet
143 virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
146 struct virtqueue *rxvq = rx_queue;
149 struct vring_used_elem *rused;
150 struct rte_mbuf **sw_ring;
151 struct rte_mbuf **sw_ring_end;
152 uint16_t nb_pkts_received;
153 __m128i shuf_msk1, shuf_msk2, len_adjust;
155 shuf_msk1 = _mm_set_epi8(
156 0xFF, 0xFF, 0xFF, 0xFF,
157 0xFF, 0xFF, /* vlan tci */
159 0xFF, 0xFF, 5, 4, /* pkt len */
160 0xFF, 0xFF, 0xFF, 0xFF /* packet type */
164 shuf_msk2 = _mm_set_epi8(
165 0xFF, 0xFF, 0xFF, 0xFF,
166 0xFF, 0xFF, /* vlan tci */
167 13, 12, /* dat len */
168 0xFF, 0xFF, 13, 12, /* pkt len */
169 0xFF, 0xFF, 0xFF, 0xFF /* packet type */
172 /* Subtract the header length.
173 * In which case do we need the header length in used->len ?
175 len_adjust = _mm_set_epi16(
178 (uint16_t) -sizeof(struct virtio_net_hdr),
179 0, (uint16_t) -sizeof(struct virtio_net_hdr),
182 if (unlikely(nb_pkts < RTE_VIRTIO_DESC_PER_LOOP))
185 nb_used = *(volatile uint16_t *)&rxvq->vq_ring.used->idx -
186 rxvq->vq_used_cons_idx;
188 rte_compiler_barrier();
190 if (unlikely(nb_used == 0))
193 nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_VIRTIO_DESC_PER_LOOP);
194 nb_used = RTE_MIN(nb_used, nb_pkts);
196 desc_idx = (uint16_t)(rxvq->vq_used_cons_idx & (rxvq->vq_nentries - 1));
197 rused = &rxvq->vq_ring.used->ring[desc_idx];
198 sw_ring = &rxvq->sw_ring[desc_idx];
199 sw_ring_end = &rxvq->sw_ring[rxvq->vq_nentries];
201 _mm_prefetch((const void *)rused, _MM_HINT_T0);
203 if (rxvq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
204 virtio_rxq_rearm_vec(rxvq);
205 if (unlikely(virtqueue_kick_prepare(rxvq)))
206 virtqueue_notify(rxvq);
209 for (nb_pkts_received = 0;
210 nb_pkts_received < nb_used;) {
211 __m128i desc[RTE_VIRTIO_DESC_PER_LOOP / 2];
212 __m128i mbp[RTE_VIRTIO_DESC_PER_LOOP / 2];
213 __m128i pkt_mb[RTE_VIRTIO_DESC_PER_LOOP];
215 mbp[0] = _mm_loadu_si128((__m128i *)(sw_ring + 0));
216 desc[0] = _mm_loadu_si128((__m128i *)(rused + 0));
217 _mm_storeu_si128((__m128i *)&rx_pkts[0], mbp[0]);
219 mbp[1] = _mm_loadu_si128((__m128i *)(sw_ring + 2));
220 desc[1] = _mm_loadu_si128((__m128i *)(rused + 2));
221 _mm_storeu_si128((__m128i *)&rx_pkts[2], mbp[1]);
223 mbp[2] = _mm_loadu_si128((__m128i *)(sw_ring + 4));
224 desc[2] = _mm_loadu_si128((__m128i *)(rused + 4));
225 _mm_storeu_si128((__m128i *)&rx_pkts[4], mbp[2]);
227 mbp[3] = _mm_loadu_si128((__m128i *)(sw_ring + 6));
228 desc[3] = _mm_loadu_si128((__m128i *)(rused + 6));
229 _mm_storeu_si128((__m128i *)&rx_pkts[6], mbp[3]);
231 pkt_mb[1] = _mm_shuffle_epi8(desc[0], shuf_msk2);
232 pkt_mb[0] = _mm_shuffle_epi8(desc[0], shuf_msk1);
233 pkt_mb[1] = _mm_add_epi16(pkt_mb[1], len_adjust);
234 pkt_mb[0] = _mm_add_epi16(pkt_mb[0], len_adjust);
235 _mm_storeu_si128((void *)&rx_pkts[1]->rx_descriptor_fields1,
237 _mm_storeu_si128((void *)&rx_pkts[0]->rx_descriptor_fields1,
240 pkt_mb[3] = _mm_shuffle_epi8(desc[1], shuf_msk2);
241 pkt_mb[2] = _mm_shuffle_epi8(desc[1], shuf_msk1);
242 pkt_mb[3] = _mm_add_epi16(pkt_mb[3], len_adjust);
243 pkt_mb[2] = _mm_add_epi16(pkt_mb[2], len_adjust);
244 _mm_storeu_si128((void *)&rx_pkts[3]->rx_descriptor_fields1,
246 _mm_storeu_si128((void *)&rx_pkts[2]->rx_descriptor_fields1,
249 pkt_mb[5] = _mm_shuffle_epi8(desc[2], shuf_msk2);
250 pkt_mb[4] = _mm_shuffle_epi8(desc[2], shuf_msk1);
251 pkt_mb[5] = _mm_add_epi16(pkt_mb[5], len_adjust);
252 pkt_mb[4] = _mm_add_epi16(pkt_mb[4], len_adjust);
253 _mm_storeu_si128((void *)&rx_pkts[5]->rx_descriptor_fields1,
255 _mm_storeu_si128((void *)&rx_pkts[4]->rx_descriptor_fields1,
258 pkt_mb[7] = _mm_shuffle_epi8(desc[3], shuf_msk2);
259 pkt_mb[6] = _mm_shuffle_epi8(desc[3], shuf_msk1);
260 pkt_mb[7] = _mm_add_epi16(pkt_mb[7], len_adjust);
261 pkt_mb[6] = _mm_add_epi16(pkt_mb[6], len_adjust);
262 _mm_storeu_si128((void *)&rx_pkts[7]->rx_descriptor_fields1,
264 _mm_storeu_si128((void *)&rx_pkts[6]->rx_descriptor_fields1,
267 if (unlikely(nb_used <= RTE_VIRTIO_DESC_PER_LOOP)) {
268 if (sw_ring + nb_used <= sw_ring_end)
269 nb_pkts_received += nb_used;
271 nb_pkts_received += sw_ring_end - sw_ring;
274 if (unlikely(sw_ring + RTE_VIRTIO_DESC_PER_LOOP >=
276 nb_pkts_received += sw_ring_end - sw_ring;
279 nb_pkts_received += RTE_VIRTIO_DESC_PER_LOOP;
281 rx_pkts += RTE_VIRTIO_DESC_PER_LOOP;
282 sw_ring += RTE_VIRTIO_DESC_PER_LOOP;
283 rused += RTE_VIRTIO_DESC_PER_LOOP;
284 nb_used -= RTE_VIRTIO_DESC_PER_LOOP;
289 rxvq->vq_used_cons_idx += nb_pkts_received;
290 rxvq->vq_free_cnt += nb_pkts_received;
291 rxvq->packets += nb_pkts_received;
292 return nb_pkts_received;
295 int __attribute__((cold))
296 virtio_rxq_vec_setup(struct virtqueue *rxq)
299 struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
302 mb_def.data_off = RTE_PKTMBUF_HEADROOM;
303 mb_def.port = rxq->port_id;
304 rte_mbuf_refcnt_set(&mb_def, 1);
306 /* prevent compiler reordering: rearm_data covers previous fields */
307 rte_compiler_barrier();
308 p = (uintptr_t)&mb_def.rearm_data;
309 rxq->mbuf_initializer = *(uint64_t *)p;