1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2017 Intel Corporation
7 #include <linux/virtio_net.h>
10 #include <rte_memcpy.h>
11 #include <rte_vhost.h>
16 * A very simple vhost-user net driver implementation, without
17 * any extra features being enabled, such as TSO and mrg-Rx.
21 vs_vhost_net_setup(struct vhost_dev *dev)
25 struct vhost_queue *queue;
27 RTE_LOG(INFO, VHOST_CONFIG,
28 "setting builtin vhost-user net driver\n");
30 rte_vhost_get_negotiated_features(vid, &dev->features);
31 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
32 dev->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
34 dev->hdr_len = sizeof(struct virtio_net_hdr);
36 rte_vhost_get_mem_table(vid, &dev->mem);
38 dev->nr_vrings = rte_vhost_get_vring_num(vid);
39 for (i = 0; i < dev->nr_vrings; i++) {
40 queue = &dev->queues[i];
42 queue->last_used_idx = 0;
43 queue->last_avail_idx = 0;
44 rte_vhost_get_vhost_vring(vid, i, &queue->vr);
49 vs_vhost_net_remove(struct vhost_dev *dev)
54 static __rte_always_inline int
55 enqueue_pkt(struct vhost_dev *dev, struct rte_vhost_vring *vr,
56 struct rte_mbuf *m, uint16_t desc_idx)
58 uint32_t desc_avail, desc_offset;
59 uint32_t mbuf_avail, mbuf_offset;
61 struct vring_desc *desc;
63 struct virtio_net_hdr virtio_hdr = {0, 0, 0, 0, 0, 0};
64 /* A counter to avoid desc dead loop chain */
67 desc = &vr->desc[desc_idx];
68 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
70 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
71 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
72 * otherwise stores offset on the stack instead of in a register.
74 if (unlikely(desc->len < dev->hdr_len) || !desc_addr)
77 rte_prefetch0((void *)(uintptr_t)desc_addr);
79 /* write virtio-net header */
80 *(struct virtio_net_hdr *)(uintptr_t)desc_addr = virtio_hdr;
82 desc_offset = dev->hdr_len;
83 desc_avail = desc->len - dev->hdr_len;
85 mbuf_avail = rte_pktmbuf_data_len(m);
87 while (mbuf_avail != 0 || m->next != NULL) {
88 /* done with current mbuf, fetch next */
89 if (mbuf_avail == 0) {
93 mbuf_avail = rte_pktmbuf_data_len(m);
96 /* done with current desc buf, fetch next */
97 if (desc_avail == 0) {
98 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
99 /* Room in vring buffer is not enough */
102 if (unlikely(desc->next >= vr->size ||
103 ++nr_desc > vr->size))
106 desc = &vr->desc[desc->next];
107 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
108 if (unlikely(!desc_addr))
112 desc_avail = desc->len;
115 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
116 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
117 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
120 mbuf_avail -= cpy_len;
121 mbuf_offset += cpy_len;
122 desc_avail -= cpy_len;
123 desc_offset += cpy_len;
130 vs_enqueue_pkts(struct vhost_dev *dev, uint16_t queue_id,
131 struct rte_mbuf **pkts, uint32_t count)
133 struct vhost_queue *queue;
134 struct rte_vhost_vring *vr;
135 uint16_t avail_idx, free_entries, start_idx;
136 uint16_t desc_indexes[MAX_PKT_BURST];
140 queue = &dev->queues[queue_id];
143 avail_idx = *((volatile uint16_t *)&vr->avail->idx);
144 start_idx = queue->last_used_idx;
145 free_entries = avail_idx - start_idx;
146 count = RTE_MIN(count, free_entries);
147 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
151 /* Retrieve all of the desc indexes first to avoid caching issues. */
152 rte_prefetch0(&vr->avail->ring[start_idx & (vr->size - 1)]);
153 for (i = 0; i < count; i++) {
154 used_idx = (start_idx + i) & (vr->size - 1);
155 desc_indexes[i] = vr->avail->ring[used_idx];
156 vr->used->ring[used_idx].id = desc_indexes[i];
157 vr->used->ring[used_idx].len = pkts[i]->pkt_len +
161 rte_prefetch0(&vr->desc[desc_indexes[0]]);
162 for (i = 0; i < count; i++) {
163 uint16_t desc_idx = desc_indexes[i];
166 err = enqueue_pkt(dev, vr, pkts[i], desc_idx);
168 used_idx = (start_idx + i) & (vr->size - 1);
169 vr->used->ring[used_idx].len = dev->hdr_len;
173 rte_prefetch0(&vr->desc[desc_indexes[i+1]]);
178 *(volatile uint16_t *)&vr->used->idx += count;
179 queue->last_used_idx += count;
181 rte_vhost_vring_call(dev->vid, queue_id);
186 static __rte_always_inline int
187 dequeue_pkt(struct vhost_dev *dev, struct rte_vhost_vring *vr,
188 struct rte_mbuf *m, uint16_t desc_idx,
189 struct rte_mempool *mbuf_pool)
191 struct vring_desc *desc;
193 uint32_t desc_avail, desc_offset;
194 uint32_t mbuf_avail, mbuf_offset;
196 struct rte_mbuf *cur = m, *prev = m;
197 /* A counter to avoid desc dead loop chain */
198 uint32_t nr_desc = 1;
200 desc = &vr->desc[desc_idx];
201 if (unlikely((desc->len < dev->hdr_len)) ||
202 (desc->flags & VRING_DESC_F_INDIRECT))
205 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
206 if (unlikely(!desc_addr))
210 * We don't support ANY_LAYOUT, neither VERSION_1, meaning
211 * a Tx packet from guest must have 2 desc buffers at least:
212 * the first for storing the header and the others for
215 * And since we don't support TSO, we could simply skip the
218 desc = &vr->desc[desc->next];
219 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
220 if (unlikely(!desc_addr))
222 rte_prefetch0((void *)(uintptr_t)desc_addr);
225 desc_avail = desc->len;
229 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
231 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
232 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
234 (void *)((uintptr_t)(desc_addr + desc_offset)),
237 mbuf_avail -= cpy_len;
238 mbuf_offset += cpy_len;
239 desc_avail -= cpy_len;
240 desc_offset += cpy_len;
242 /* This desc reaches to its end, get the next one */
243 if (desc_avail == 0) {
244 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
247 if (unlikely(desc->next >= vr->size ||
248 ++nr_desc > vr->size))
250 desc = &vr->desc[desc->next];
252 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
253 if (unlikely(!desc_addr))
255 rte_prefetch0((void *)(uintptr_t)desc_addr);
258 desc_avail = desc->len;
262 * This mbuf reaches to its end, get a new one
265 if (mbuf_avail == 0) {
266 cur = rte_pktmbuf_alloc(mbuf_pool);
267 if (unlikely(cur == NULL)) {
268 RTE_LOG(ERR, VHOST_DATA, "Failed to "
269 "allocate memory for mbuf.\n");
274 prev->data_len = mbuf_offset;
276 m->pkt_len += mbuf_offset;
280 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
284 prev->data_len = mbuf_offset;
285 m->pkt_len += mbuf_offset;
291 vs_dequeue_pkts(struct vhost_dev *dev, uint16_t queue_id,
292 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
294 struct vhost_queue *queue;
295 struct rte_vhost_vring *vr;
296 uint32_t desc_indexes[MAX_PKT_BURST];
299 uint16_t free_entries;
302 queue = &dev->queues[queue_id];
305 free_entries = *((volatile uint16_t *)&vr->avail->idx) -
306 queue->last_avail_idx;
307 if (free_entries == 0)
310 /* Prefetch available and used ring */
311 avail_idx = queue->last_avail_idx & (vr->size - 1);
312 used_idx = queue->last_used_idx & (vr->size - 1);
313 rte_prefetch0(&vr->avail->ring[avail_idx]);
314 rte_prefetch0(&vr->used->ring[used_idx]);
316 count = RTE_MIN(count, MAX_PKT_BURST);
317 count = RTE_MIN(count, free_entries);
319 if (unlikely(count == 0))
323 * Retrieve all of the head indexes first and pre-update used entries
324 * to avoid caching issues.
326 for (i = 0; i < count; i++) {
327 avail_idx = (queue->last_avail_idx + i) & (vr->size - 1);
328 used_idx = (queue->last_used_idx + i) & (vr->size - 1);
329 desc_indexes[i] = vr->avail->ring[avail_idx];
331 vr->used->ring[used_idx].id = desc_indexes[i];
332 vr->used->ring[used_idx].len = 0;
335 /* Prefetch descriptor index. */
336 rte_prefetch0(&vr->desc[desc_indexes[0]]);
337 for (i = 0; i < count; i++) {
340 if (likely(i + 1 < count))
341 rte_prefetch0(&vr->desc[desc_indexes[i + 1]]);
343 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
344 if (unlikely(pkts[i] == NULL)) {
345 RTE_LOG(ERR, VHOST_DATA,
346 "Failed to allocate memory for mbuf.\n");
350 err = dequeue_pkt(dev, vr, pkts[i], desc_indexes[i], mbuf_pool);
352 rte_pktmbuf_free(pkts[i]);
358 queue->last_avail_idx += i;
359 queue->last_used_idx += i;
365 rte_vhost_vring_call(dev->vid, queue_id);