4 * Copyright(c) 2010-2014 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.
35 #include <linux/virtio_net.h>
38 #include <rte_memcpy.h>
39 #include <rte_virtio_net.h>
41 #include "vhost-net-cdev.h"
43 #define MAX_PKT_BURST 32
46 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
47 * be received from the physical port or from another virtio device. A packet
48 * count is returned to indicate the number of packets that are succesfully
49 * added to the RX queue. This function works when mergeable is disabled.
51 static inline uint32_t __attribute__((always_inline))
52 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
53 struct rte_mbuf **pkts, uint32_t count)
55 struct vhost_virtqueue *vq;
56 struct vring_desc *desc;
57 struct rte_mbuf *buff;
58 /* The virtio_hdr is initialised to 0. */
59 struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
60 uint64_t buff_addr = 0;
61 uint64_t buff_hdr_addr = 0;
62 uint32_t head[MAX_PKT_BURST], packet_len = 0;
63 uint32_t head_idx, packet_success = 0;
64 uint16_t avail_idx, res_cur_idx;
65 uint16_t res_base_idx, res_end_idx;
66 uint16_t free_entries;
69 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_rx()\n", dev->device_fh);
70 if (unlikely(queue_id != VIRTIO_RXQ)) {
71 LOG_DEBUG(VHOST_DATA, "mq isn't supported in this version.\n");
75 vq = dev->virtqueue[VIRTIO_RXQ];
76 count = (count > MAX_PKT_BURST) ? MAX_PKT_BURST : count;
79 * As many data cores may want access to available buffers,
80 * they need to be reserved.
83 res_base_idx = vq->last_used_idx_res;
84 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
86 free_entries = (avail_idx - res_base_idx);
87 /*check that we have enough buffers*/
88 if (unlikely(count > free_entries))
94 res_end_idx = res_base_idx + count;
95 /* vq->last_used_idx_res is atomically updated. */
96 /* TODO: Allow to disable cmpset if no concurrency in application. */
97 success = rte_atomic16_cmpset(&vq->last_used_idx_res,
98 res_base_idx, res_end_idx);
99 } while (unlikely(success == 0));
100 res_cur_idx = res_base_idx;
101 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") Current Index %d| End Index %d\n",
102 dev->device_fh, res_cur_idx, res_end_idx);
104 /* Prefetch available ring to retrieve indexes. */
105 rte_prefetch0(&vq->avail->ring[res_cur_idx & (vq->size - 1)]);
107 /* Retrieve all of the head indexes first to avoid caching issues. */
108 for (head_idx = 0; head_idx < count; head_idx++)
109 head[head_idx] = vq->avail->ring[(res_cur_idx + head_idx) &
112 /*Prefetch descriptor index. */
113 rte_prefetch0(&vq->desc[head[packet_success]]);
115 while (res_cur_idx != res_end_idx) {
116 /* Get descriptor from available ring */
117 desc = &vq->desc[head[packet_success]];
119 buff = pkts[packet_success];
121 /* Convert from gpa to vva (guest physical addr -> vhost virtual addr) */
122 buff_addr = gpa_to_vva(dev, desc->addr);
123 /* Prefetch buffer address. */
124 rte_prefetch0((void *)(uintptr_t)buff_addr);
126 /* Copy virtio_hdr to packet and increment buffer address */
127 buff_hdr_addr = buff_addr;
128 packet_len = rte_pktmbuf_data_len(buff) + vq->vhost_hlen;
131 * If the descriptors are chained the header and data are
132 * placed in separate buffers.
134 if (desc->flags & VRING_DESC_F_NEXT) {
135 desc->len = vq->vhost_hlen;
136 desc = &vq->desc[desc->next];
137 /* Buffer address translation. */
138 buff_addr = gpa_to_vva(dev, desc->addr);
139 desc->len = rte_pktmbuf_data_len(buff);
141 buff_addr += vq->vhost_hlen;
142 desc->len = packet_len;
145 /* Update used ring with desc information */
146 vq->used->ring[res_cur_idx & (vq->size - 1)].id =
147 head[packet_success];
148 vq->used->ring[res_cur_idx & (vq->size - 1)].len = packet_len;
150 /* Copy mbuf data to buffer */
151 /* FIXME for sg mbuf and the case that desc couldn't hold the mbuf data */
152 rte_memcpy((void *)(uintptr_t)buff_addr,
153 rte_pktmbuf_mtod(buff, const void *),
154 rte_pktmbuf_data_len(buff));
155 PRINT_PACKET(dev, (uintptr_t)buff_addr,
156 rte_pktmbuf_data_len(buff), 0);
161 rte_memcpy((void *)(uintptr_t)buff_hdr_addr,
162 (const void *)&virtio_hdr, vq->vhost_hlen);
164 PRINT_PACKET(dev, (uintptr_t)buff_hdr_addr, vq->vhost_hlen, 1);
166 if (res_cur_idx < res_end_idx) {
167 /* Prefetch descriptor index. */
168 rte_prefetch0(&vq->desc[head[packet_success]]);
172 rte_compiler_barrier();
174 /* Wait until it's our turn to add our buffer to the used ring. */
175 while (unlikely(vq->last_used_idx != res_base_idx))
178 *(volatile uint16_t *)&vq->used->idx += count;
179 vq->last_used_idx = res_end_idx;
181 /* Kick the guest if necessary. */
182 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
183 eventfd_write((int)vq->kickfd, 1);
187 static inline uint32_t __attribute__((always_inline))
188 copy_from_mbuf_to_vring(struct virtio_net *dev, uint16_t res_base_idx,
189 uint16_t res_end_idx, struct rte_mbuf *pkt)
191 uint32_t vec_idx = 0;
192 uint32_t entry_success = 0;
193 struct vhost_virtqueue *vq;
194 /* The virtio_hdr is initialised to 0. */
195 struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {
196 {0, 0, 0, 0, 0, 0}, 0};
197 uint16_t cur_idx = res_base_idx;
198 uint64_t vb_addr = 0;
199 uint64_t vb_hdr_addr = 0;
200 uint32_t seg_offset = 0;
201 uint32_t vb_offset = 0;
204 uint32_t cpy_len, entry_len;
209 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") Current Index %d| "
211 dev->device_fh, cur_idx, res_end_idx);
214 * Convert from gpa to vva
215 * (guest physical addr -> vhost virtual addr)
217 vq = dev->virtqueue[VIRTIO_RXQ];
219 gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
220 vb_hdr_addr = vb_addr;
222 /* Prefetch buffer address. */
223 rte_prefetch0((void *)(uintptr_t)vb_addr);
225 virtio_hdr.num_buffers = res_end_idx - res_base_idx;
227 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") RX: Num merge buffers %d\n",
228 dev->device_fh, virtio_hdr.num_buffers);
230 rte_memcpy((void *)(uintptr_t)vb_hdr_addr,
231 (const void *)&virtio_hdr, vq->vhost_hlen);
233 PRINT_PACKET(dev, (uintptr_t)vb_hdr_addr, vq->vhost_hlen, 1);
235 seg_avail = rte_pktmbuf_data_len(pkt);
236 vb_offset = vq->vhost_hlen;
238 vq->buf_vec[vec_idx].buf_len - vq->vhost_hlen;
240 entry_len = vq->vhost_hlen;
244 vq->buf_vec[vec_idx].desc_idx;
245 vq->desc[desc_idx].len = vq->vhost_hlen;
247 if ((vq->desc[desc_idx].flags
248 & VRING_DESC_F_NEXT) == 0) {
249 /* Update used ring with desc information */
250 vq->used->ring[cur_idx & (vq->size - 1)].id
251 = vq->buf_vec[vec_idx].desc_idx;
252 vq->used->ring[cur_idx & (vq->size - 1)].len
262 gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
264 /* Prefetch buffer address. */
265 rte_prefetch0((void *)(uintptr_t)vb_addr);
267 vb_avail = vq->buf_vec[vec_idx].buf_len;
270 cpy_len = RTE_MIN(vb_avail, seg_avail);
272 while (cpy_len > 0) {
273 /* Copy mbuf data to vring buffer */
274 rte_memcpy((void *)(uintptr_t)(vb_addr + vb_offset),
275 (const void *)(rte_pktmbuf_mtod(pkt, char*) + seg_offset),
279 (uintptr_t)(vb_addr + vb_offset),
282 seg_offset += cpy_len;
283 vb_offset += cpy_len;
284 seg_avail -= cpy_len;
286 entry_len += cpy_len;
288 if (seg_avail != 0) {
290 * The virtio buffer in this vring
291 * entry reach to its end.
292 * But the segment doesn't complete.
294 if ((vq->desc[vq->buf_vec[vec_idx].desc_idx].flags &
295 VRING_DESC_F_NEXT) == 0) {
296 /* Update used ring with desc information */
297 vq->used->ring[cur_idx & (vq->size - 1)].id
298 = vq->buf_vec[vec_idx].desc_idx;
299 vq->used->ring[cur_idx & (vq->size - 1)].len
307 vb_addr = gpa_to_vva(dev,
308 vq->buf_vec[vec_idx].buf_addr);
310 vb_avail = vq->buf_vec[vec_idx].buf_len;
311 cpy_len = RTE_MIN(vb_avail, seg_avail);
314 * This current segment complete, need continue to
315 * check if the whole packet complete or not.
320 * There are more segments.
324 * This current buffer from vring is
325 * used up, need fetch next buffer
329 vq->buf_vec[vec_idx].desc_idx;
330 vq->desc[desc_idx].len = vb_offset;
332 if ((vq->desc[desc_idx].flags &
333 VRING_DESC_F_NEXT) == 0) {
334 uint16_t wrapped_idx =
335 cur_idx & (vq->size - 1);
337 * Update used ring with the
338 * descriptor information
340 vq->used->ring[wrapped_idx].id
342 vq->used->ring[wrapped_idx].len
349 /* Get next buffer from buf_vec. */
351 vb_addr = gpa_to_vva(dev,
352 vq->buf_vec[vec_idx].buf_addr);
354 vq->buf_vec[vec_idx].buf_len;
359 seg_avail = rte_pktmbuf_data_len(pkt);
360 cpy_len = RTE_MIN(vb_avail, seg_avail);
363 * This whole packet completes.
366 vq->buf_vec[vec_idx].desc_idx;
367 vq->desc[desc_idx].len = vb_offset;
369 while (vq->desc[desc_idx].flags &
371 desc_idx = vq->desc[desc_idx].next;
372 vq->desc[desc_idx].len = 0;
375 /* Update used ring with desc information */
376 vq->used->ring[cur_idx & (vq->size - 1)].id
377 = vq->buf_vec[vec_idx].desc_idx;
378 vq->used->ring[cur_idx & (vq->size - 1)].len
384 cpy_len = RTE_MIN(vb_avail, seg_avail);
389 return entry_success;
393 * This function works for mergeable RX.
395 static inline uint32_t __attribute__((always_inline))
396 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
397 struct rte_mbuf **pkts, uint32_t count)
399 struct vhost_virtqueue *vq;
400 uint32_t pkt_idx = 0, entry_success = 0;
401 uint16_t avail_idx, res_cur_idx;
402 uint16_t res_base_idx, res_end_idx;
405 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_merge_rx()\n",
407 if (unlikely(queue_id != VIRTIO_RXQ)) {
408 LOG_DEBUG(VHOST_DATA, "mq isn't supported in this version.\n");
411 vq = dev->virtqueue[VIRTIO_RXQ];
412 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
417 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
418 uint32_t secure_len = 0;
420 uint32_t vec_idx = 0;
421 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + vq->vhost_hlen;
426 * As many data cores may want access to available
427 * buffers, they need to be reserved.
429 res_base_idx = vq->last_used_idx_res;
430 res_cur_idx = res_base_idx;
433 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
434 if (unlikely(res_cur_idx == avail_idx)) {
435 LOG_DEBUG(VHOST_DATA,
436 "(%"PRIu64") Failed "
437 "to get enough desc from "
442 uint16_t wrapped_idx =
443 (res_cur_idx) & (vq->size - 1);
445 vq->avail->ring[wrapped_idx];
450 secure_len += vq->desc[idx].len;
451 if (vq->desc[idx].flags &
453 idx = vq->desc[idx].next;
460 } while (pkt_len > secure_len);
462 /* vq->last_used_idx_res is atomically updated. */
463 success = rte_atomic16_cmpset(&vq->last_used_idx_res,
466 } while (success == 0);
469 need_cnt = res_cur_idx - res_base_idx;
471 for (i = 0; i < need_cnt; i++, id++) {
472 uint16_t wrapped_idx = id & (vq->size - 1);
473 uint32_t idx = vq->avail->ring[wrapped_idx];
477 vq->buf_vec[vec_idx].buf_addr =
479 vq->buf_vec[vec_idx].buf_len =
481 vq->buf_vec[vec_idx].desc_idx = idx;
484 if (vq->desc[idx].flags & VRING_DESC_F_NEXT) {
485 idx = vq->desc[idx].next;
491 res_end_idx = res_cur_idx;
493 entry_success = copy_from_mbuf_to_vring(dev, res_base_idx,
494 res_end_idx, pkts[pkt_idx]);
496 rte_compiler_barrier();
499 * Wait until it's our turn to add our buffer
502 while (unlikely(vq->last_used_idx != res_base_idx))
505 *(volatile uint16_t *)&vq->used->idx += entry_success;
506 vq->last_used_idx = res_end_idx;
508 /* Kick the guest if necessary. */
509 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
510 eventfd_write((int)vq->kickfd, 1);
517 rte_vhost_enqueue_burst(struct virtio_net *dev, uint16_t queue_id,
518 struct rte_mbuf **pkts, uint16_t count)
520 if (unlikely(dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)))
521 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
523 return virtio_dev_rx(dev, queue_id, pkts, count);
527 rte_vhost_dequeue_burst(struct virtio_net *dev, uint16_t queue_id,
528 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
530 struct rte_mbuf *m, *prev;
531 struct vhost_virtqueue *vq;
532 struct vring_desc *desc;
533 uint64_t vb_addr = 0;
534 uint32_t head[MAX_PKT_BURST];
537 uint16_t free_entries, entry_success = 0;
540 if (unlikely(queue_id != VIRTIO_TXQ)) {
541 LOG_DEBUG(VHOST_DATA, "mq isn't supported in this version.\n");
545 vq = dev->virtqueue[VIRTIO_TXQ];
546 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
548 /* If there are no available buffers then return. */
549 if (vq->last_used_idx == avail_idx)
552 LOG_DEBUG(VHOST_DATA, "%s (%"PRIu64")\n", __func__,
555 /* Prefetch available ring to retrieve head indexes. */
556 rte_prefetch0(&vq->avail->ring[vq->last_used_idx & (vq->size - 1)]);
558 /*get the number of free entries in the ring*/
559 free_entries = (avail_idx - vq->last_used_idx);
561 free_entries = RTE_MIN(free_entries, count);
562 /* Limit to MAX_PKT_BURST. */
563 free_entries = RTE_MIN(free_entries, MAX_PKT_BURST);
565 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") Buffers available %d\n",
566 dev->device_fh, free_entries);
567 /* Retrieve all of the head indexes first to avoid caching issues. */
568 for (i = 0; i < free_entries; i++)
569 head[i] = vq->avail->ring[(vq->last_used_idx + i) & (vq->size - 1)];
571 /* Prefetch descriptor index. */
572 rte_prefetch0(&vq->desc[head[entry_success]]);
573 rte_prefetch0(&vq->used->ring[vq->last_used_idx & (vq->size - 1)]);
575 while (entry_success < free_entries) {
576 uint32_t vb_avail, vb_offset;
577 uint32_t seg_avail, seg_offset;
579 uint32_t seg_num = 0;
580 struct rte_mbuf *cur;
581 uint8_t alloc_err = 0;
583 desc = &vq->desc[head[entry_success]];
585 /* Discard first buffer as it is the virtio header */
586 desc = &vq->desc[desc->next];
588 /* Buffer address translation. */
589 vb_addr = gpa_to_vva(dev, desc->addr);
590 /* Prefetch buffer address. */
591 rte_prefetch0((void *)(uintptr_t)vb_addr);
593 used_idx = vq->last_used_idx & (vq->size - 1);
595 if (entry_success < (free_entries - 1)) {
596 /* Prefetch descriptor index. */
597 rte_prefetch0(&vq->desc[head[entry_success+1]]);
598 rte_prefetch0(&vq->used->ring[(used_idx + 1) & (vq->size - 1)]);
601 /* Update used index buffer information. */
602 vq->used->ring[used_idx].id = head[entry_success];
603 vq->used->ring[used_idx].len = 0;
606 vb_avail = desc->len;
607 /* Allocate an mbuf and populate the structure. */
608 m = rte_pktmbuf_alloc(mbuf_pool);
609 if (unlikely(m == NULL)) {
610 RTE_LOG(ERR, VHOST_DATA,
611 "Failed to allocate memory for mbuf.\n");
612 return entry_success;
615 seg_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
616 cpy_len = RTE_MIN(vb_avail, seg_avail);
618 PRINT_PACKET(dev, (uintptr_t)vb_addr, desc->len, 0);
623 while (cpy_len != 0) {
624 rte_memcpy((void *)(rte_pktmbuf_mtod(cur, char *) + seg_offset),
625 (void *)((uintptr_t)(vb_addr + vb_offset)),
628 seg_offset += cpy_len;
629 vb_offset += cpy_len;
631 seg_avail -= cpy_len;
635 * The segment reachs to its end,
636 * while the virtio buffer in TX vring has
637 * more data to be copied.
639 cur->data_len = seg_offset;
640 m->pkt_len += seg_offset;
641 /* Allocate mbuf and populate the structure. */
642 cur = rte_pktmbuf_alloc(mbuf_pool);
643 if (unlikely(cur == NULL)) {
644 RTE_LOG(ERR, VHOST_DATA, "Failed to "
645 "allocate memory for mbuf.\n");
655 seg_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
657 if (desc->flags & VRING_DESC_F_NEXT) {
659 * There are more virtio buffers in
660 * same vring entry need to be copied.
662 if (seg_avail == 0) {
664 * The current segment hasn't
665 * room to accomodate more
668 cur->data_len = seg_offset;
669 m->pkt_len += seg_offset;
671 * Allocate an mbuf and
672 * populate the structure.
674 cur = rte_pktmbuf_alloc(mbuf_pool);
675 if (unlikely(cur == NULL)) {
689 seg_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
692 desc = &vq->desc[desc->next];
694 /* Buffer address translation. */
695 vb_addr = gpa_to_vva(dev, desc->addr);
696 /* Prefetch buffer address. */
697 rte_prefetch0((void *)(uintptr_t)vb_addr);
699 vb_avail = desc->len;
701 PRINT_PACKET(dev, (uintptr_t)vb_addr,
704 /* The whole packet completes. */
705 cur->data_len = seg_offset;
706 m->pkt_len += seg_offset;
711 cpy_len = RTE_MIN(vb_avail, seg_avail);
714 if (unlikely(alloc_err == 1))
717 m->nb_segs = seg_num;
719 pkts[entry_success] = m;
724 rte_compiler_barrier();
725 vq->used->idx += entry_success;
726 /* Kick guest if required. */
727 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
728 eventfd_write((int)vq->kickfd, 1);
729 return entry_success;