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.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 /* flush used->idx update before we read avail->flags. */
184 /* Kick the guest if necessary. */
185 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
186 eventfd_write((int)vq->callfd, 1);
190 static inline uint32_t __attribute__((always_inline))
191 copy_from_mbuf_to_vring(struct virtio_net *dev, uint16_t res_base_idx,
192 uint16_t res_end_idx, struct rte_mbuf *pkt)
194 uint32_t vec_idx = 0;
195 uint32_t entry_success = 0;
196 struct vhost_virtqueue *vq;
197 /* The virtio_hdr is initialised to 0. */
198 struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {
199 {0, 0, 0, 0, 0, 0}, 0};
200 uint16_t cur_idx = res_base_idx;
201 uint64_t vb_addr = 0;
202 uint64_t vb_hdr_addr = 0;
203 uint32_t seg_offset = 0;
204 uint32_t vb_offset = 0;
207 uint32_t cpy_len, entry_len;
212 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") Current Index %d| "
214 dev->device_fh, cur_idx, res_end_idx);
217 * Convert from gpa to vva
218 * (guest physical addr -> vhost virtual addr)
220 vq = dev->virtqueue[VIRTIO_RXQ];
222 gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
223 vb_hdr_addr = vb_addr;
225 /* Prefetch buffer address. */
226 rte_prefetch0((void *)(uintptr_t)vb_addr);
228 virtio_hdr.num_buffers = res_end_idx - res_base_idx;
230 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") RX: Num merge buffers %d\n",
231 dev->device_fh, virtio_hdr.num_buffers);
233 rte_memcpy((void *)(uintptr_t)vb_hdr_addr,
234 (const void *)&virtio_hdr, vq->vhost_hlen);
236 PRINT_PACKET(dev, (uintptr_t)vb_hdr_addr, vq->vhost_hlen, 1);
238 seg_avail = rte_pktmbuf_data_len(pkt);
239 vb_offset = vq->vhost_hlen;
241 vq->buf_vec[vec_idx].buf_len - vq->vhost_hlen;
243 entry_len = vq->vhost_hlen;
247 vq->buf_vec[vec_idx].desc_idx;
248 vq->desc[desc_idx].len = vq->vhost_hlen;
250 if ((vq->desc[desc_idx].flags
251 & VRING_DESC_F_NEXT) == 0) {
252 /* Update used ring with desc information */
253 vq->used->ring[cur_idx & (vq->size - 1)].id
254 = vq->buf_vec[vec_idx].desc_idx;
255 vq->used->ring[cur_idx & (vq->size - 1)].len
265 gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
267 /* Prefetch buffer address. */
268 rte_prefetch0((void *)(uintptr_t)vb_addr);
270 vb_avail = vq->buf_vec[vec_idx].buf_len;
273 cpy_len = RTE_MIN(vb_avail, seg_avail);
275 while (cpy_len > 0) {
276 /* Copy mbuf data to vring buffer */
277 rte_memcpy((void *)(uintptr_t)(vb_addr + vb_offset),
278 (const void *)(rte_pktmbuf_mtod(pkt, char*) + seg_offset),
282 (uintptr_t)(vb_addr + vb_offset),
285 seg_offset += cpy_len;
286 vb_offset += cpy_len;
287 seg_avail -= cpy_len;
289 entry_len += cpy_len;
291 if (seg_avail != 0) {
293 * The virtio buffer in this vring
294 * entry reach to its end.
295 * But the segment doesn't complete.
297 if ((vq->desc[vq->buf_vec[vec_idx].desc_idx].flags &
298 VRING_DESC_F_NEXT) == 0) {
299 /* Update used ring with desc information */
300 vq->used->ring[cur_idx & (vq->size - 1)].id
301 = vq->buf_vec[vec_idx].desc_idx;
302 vq->used->ring[cur_idx & (vq->size - 1)].len
310 vb_addr = gpa_to_vva(dev,
311 vq->buf_vec[vec_idx].buf_addr);
313 vb_avail = vq->buf_vec[vec_idx].buf_len;
314 cpy_len = RTE_MIN(vb_avail, seg_avail);
317 * This current segment complete, need continue to
318 * check if the whole packet complete or not.
323 * There are more segments.
327 * This current buffer from vring is
328 * used up, need fetch next buffer
332 vq->buf_vec[vec_idx].desc_idx;
333 vq->desc[desc_idx].len = vb_offset;
335 if ((vq->desc[desc_idx].flags &
336 VRING_DESC_F_NEXT) == 0) {
337 uint16_t wrapped_idx =
338 cur_idx & (vq->size - 1);
340 * Update used ring with the
341 * descriptor information
343 vq->used->ring[wrapped_idx].id
345 vq->used->ring[wrapped_idx].len
352 /* Get next buffer from buf_vec. */
354 vb_addr = gpa_to_vva(dev,
355 vq->buf_vec[vec_idx].buf_addr);
357 vq->buf_vec[vec_idx].buf_len;
362 seg_avail = rte_pktmbuf_data_len(pkt);
363 cpy_len = RTE_MIN(vb_avail, seg_avail);
366 * This whole packet completes.
369 vq->buf_vec[vec_idx].desc_idx;
370 vq->desc[desc_idx].len = vb_offset;
372 while (vq->desc[desc_idx].flags &
374 desc_idx = vq->desc[desc_idx].next;
375 vq->desc[desc_idx].len = 0;
378 /* Update used ring with desc information */
379 vq->used->ring[cur_idx & (vq->size - 1)].id
380 = vq->buf_vec[vec_idx].desc_idx;
381 vq->used->ring[cur_idx & (vq->size - 1)].len
387 cpy_len = RTE_MIN(vb_avail, seg_avail);
392 return entry_success;
396 * This function works for mergeable RX.
398 static inline uint32_t __attribute__((always_inline))
399 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
400 struct rte_mbuf **pkts, uint32_t count)
402 struct vhost_virtqueue *vq;
403 uint32_t pkt_idx = 0, entry_success = 0;
404 uint16_t avail_idx, res_cur_idx;
405 uint16_t res_base_idx, res_end_idx;
408 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_merge_rx()\n",
410 if (unlikely(queue_id != VIRTIO_RXQ)) {
411 LOG_DEBUG(VHOST_DATA, "mq isn't supported in this version.\n");
414 vq = dev->virtqueue[VIRTIO_RXQ];
415 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
420 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
421 uint32_t secure_len = 0;
423 uint32_t vec_idx = 0;
424 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + vq->vhost_hlen;
429 * As many data cores may want access to available
430 * buffers, they need to be reserved.
432 res_base_idx = vq->last_used_idx_res;
433 res_cur_idx = res_base_idx;
436 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
437 if (unlikely(res_cur_idx == avail_idx)) {
438 LOG_DEBUG(VHOST_DATA,
439 "(%"PRIu64") Failed "
440 "to get enough desc from "
445 uint16_t wrapped_idx =
446 (res_cur_idx) & (vq->size - 1);
448 vq->avail->ring[wrapped_idx];
453 secure_len += vq->desc[idx].len;
454 if (vq->desc[idx].flags &
456 idx = vq->desc[idx].next;
463 } while (pkt_len > secure_len);
465 /* vq->last_used_idx_res is atomically updated. */
466 success = rte_atomic16_cmpset(&vq->last_used_idx_res,
469 } while (success == 0);
472 need_cnt = res_cur_idx - res_base_idx;
474 for (i = 0; i < need_cnt; i++, id++) {
475 uint16_t wrapped_idx = id & (vq->size - 1);
476 uint32_t idx = vq->avail->ring[wrapped_idx];
480 vq->buf_vec[vec_idx].buf_addr =
482 vq->buf_vec[vec_idx].buf_len =
484 vq->buf_vec[vec_idx].desc_idx = idx;
487 if (vq->desc[idx].flags & VRING_DESC_F_NEXT) {
488 idx = vq->desc[idx].next;
494 res_end_idx = res_cur_idx;
496 entry_success = copy_from_mbuf_to_vring(dev, res_base_idx,
497 res_end_idx, pkts[pkt_idx]);
499 rte_compiler_barrier();
502 * Wait until it's our turn to add our buffer
505 while (unlikely(vq->last_used_idx != res_base_idx))
508 *(volatile uint16_t *)&vq->used->idx += entry_success;
509 vq->last_used_idx = res_end_idx;
511 /* flush used->idx update before we read avail->flags. */
514 /* Kick the guest if necessary. */
515 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
516 eventfd_write((int)vq->callfd, 1);
523 rte_vhost_enqueue_burst(struct virtio_net *dev, uint16_t queue_id,
524 struct rte_mbuf **pkts, uint16_t count)
526 if (unlikely(dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)))
527 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
529 return virtio_dev_rx(dev, queue_id, pkts, count);
533 rte_vhost_dequeue_burst(struct virtio_net *dev, uint16_t queue_id,
534 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
536 struct rte_mbuf *m, *prev;
537 struct vhost_virtqueue *vq;
538 struct vring_desc *desc;
539 uint64_t vb_addr = 0;
540 uint32_t head[MAX_PKT_BURST];
543 uint16_t free_entries, entry_success = 0;
546 if (unlikely(queue_id != VIRTIO_TXQ)) {
547 LOG_DEBUG(VHOST_DATA, "mq isn't supported in this version.\n");
551 vq = dev->virtqueue[VIRTIO_TXQ];
552 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
554 /* If there are no available buffers then return. */
555 if (vq->last_used_idx == avail_idx)
558 LOG_DEBUG(VHOST_DATA, "%s (%"PRIu64")\n", __func__,
561 /* Prefetch available ring to retrieve head indexes. */
562 rte_prefetch0(&vq->avail->ring[vq->last_used_idx & (vq->size - 1)]);
564 /*get the number of free entries in the ring*/
565 free_entries = (avail_idx - vq->last_used_idx);
567 free_entries = RTE_MIN(free_entries, count);
568 /* Limit to MAX_PKT_BURST. */
569 free_entries = RTE_MIN(free_entries, MAX_PKT_BURST);
571 LOG_DEBUG(VHOST_DATA, "(%"PRIu64") Buffers available %d\n",
572 dev->device_fh, free_entries);
573 /* Retrieve all of the head indexes first to avoid caching issues. */
574 for (i = 0; i < free_entries; i++)
575 head[i] = vq->avail->ring[(vq->last_used_idx + i) & (vq->size - 1)];
577 /* Prefetch descriptor index. */
578 rte_prefetch0(&vq->desc[head[entry_success]]);
579 rte_prefetch0(&vq->used->ring[vq->last_used_idx & (vq->size - 1)]);
581 while (entry_success < free_entries) {
582 uint32_t vb_avail, vb_offset;
583 uint32_t seg_avail, seg_offset;
585 uint32_t seg_num = 0;
586 struct rte_mbuf *cur;
587 uint8_t alloc_err = 0;
589 desc = &vq->desc[head[entry_success]];
591 /* Discard first buffer as it is the virtio header */
592 desc = &vq->desc[desc->next];
594 /* Buffer address translation. */
595 vb_addr = gpa_to_vva(dev, desc->addr);
596 /* Prefetch buffer address. */
597 rte_prefetch0((void *)(uintptr_t)vb_addr);
599 used_idx = vq->last_used_idx & (vq->size - 1);
601 if (entry_success < (free_entries - 1)) {
602 /* Prefetch descriptor index. */
603 rte_prefetch0(&vq->desc[head[entry_success+1]]);
604 rte_prefetch0(&vq->used->ring[(used_idx + 1) & (vq->size - 1)]);
607 /* Update used index buffer information. */
608 vq->used->ring[used_idx].id = head[entry_success];
609 vq->used->ring[used_idx].len = 0;
612 vb_avail = desc->len;
613 /* Allocate an mbuf and populate the structure. */
614 m = rte_pktmbuf_alloc(mbuf_pool);
615 if (unlikely(m == NULL)) {
616 RTE_LOG(ERR, VHOST_DATA,
617 "Failed to allocate memory for mbuf.\n");
621 seg_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
622 cpy_len = RTE_MIN(vb_avail, seg_avail);
624 PRINT_PACKET(dev, (uintptr_t)vb_addr, desc->len, 0);
629 while (cpy_len != 0) {
630 rte_memcpy((void *)(rte_pktmbuf_mtod(cur, char *) + seg_offset),
631 (void *)((uintptr_t)(vb_addr + vb_offset)),
634 seg_offset += cpy_len;
635 vb_offset += cpy_len;
637 seg_avail -= cpy_len;
641 * The segment reachs to its end,
642 * while the virtio buffer in TX vring has
643 * more data to be copied.
645 cur->data_len = seg_offset;
646 m->pkt_len += seg_offset;
647 /* Allocate mbuf and populate the structure. */
648 cur = rte_pktmbuf_alloc(mbuf_pool);
649 if (unlikely(cur == NULL)) {
650 RTE_LOG(ERR, VHOST_DATA, "Failed to "
651 "allocate memory for mbuf.\n");
661 seg_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
663 if (desc->flags & VRING_DESC_F_NEXT) {
665 * There are more virtio buffers in
666 * same vring entry need to be copied.
668 if (seg_avail == 0) {
670 * The current segment hasn't
671 * room to accomodate more
674 cur->data_len = seg_offset;
675 m->pkt_len += seg_offset;
677 * Allocate an mbuf and
678 * populate the structure.
680 cur = rte_pktmbuf_alloc(mbuf_pool);
681 if (unlikely(cur == NULL)) {
695 seg_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
698 desc = &vq->desc[desc->next];
700 /* Buffer address translation. */
701 vb_addr = gpa_to_vva(dev, desc->addr);
702 /* Prefetch buffer address. */
703 rte_prefetch0((void *)(uintptr_t)vb_addr);
705 vb_avail = desc->len;
707 PRINT_PACKET(dev, (uintptr_t)vb_addr,
710 /* The whole packet completes. */
711 cur->data_len = seg_offset;
712 m->pkt_len += seg_offset;
717 cpy_len = RTE_MIN(vb_avail, seg_avail);
720 if (unlikely(alloc_err == 1))
723 m->nb_segs = seg_num;
725 pkts[entry_success] = m;
730 rte_compiler_barrier();
731 vq->used->idx += entry_success;
732 /* Kick guest if required. */
733 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
734 eventfd_write((int)vq->callfd, 1);
735 return entry_success;