]> git.proxmox.com Git - ceph.git/blob - ceph/src/spdk/dpdk/lib/librte_vhost/virtio_net.c
projects / ceph.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
update source to Ceph Pacific 16.2.2
[ceph.git] / ceph / src / spdk / dpdk / lib / librte_vhost / virtio_net.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2016 Intel Corporation
3 */
4
5 #include <stdint.h>
6 #include <stdbool.h>
7 #include <linux/virtio_net.h>
8
9 #include <rte_mbuf.h>
10 #include <rte_memcpy.h>
11 #include <rte_ether.h>
12 #include <rte_ip.h>
13 #include <rte_vhost.h>
14 #include <rte_tcp.h>
15 #include <rte_udp.h>
16 #include <rte_sctp.h>
17 #include <rte_arp.h>
18 #include <rte_spinlock.h>
19 #include <rte_malloc.h>
20
21 #include "iotlb.h"
22 #include "vhost.h"
23
24 #define MAX_PKT_BURST 32
25
26 #define MAX_BATCH_LEN 256
27
28 static __rte_always_inline bool
29 rxvq_is_mergeable(struct virtio_net *dev)
30 {
31 return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
32 }
33
34 static __rte_always_inline bool
35 virtio_net_is_inorder(struct virtio_net *dev)
36 {
37 return dev->features & (1ULL << VIRTIO_F_IN_ORDER);
38 }
39
40 static bool
41 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
42 {
43 return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
44 }
45
46 static inline void
47 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
48 {
49 struct batch_copy_elem *elem = vq->batch_copy_elems;
50 uint16_t count = vq->batch_copy_nb_elems;
51 int i;
52
53 for (i = 0; i < count; i++) {
54 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
55 vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
56 elem[i].len);
57 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
58 }
59
60 vq->batch_copy_nb_elems = 0;
61 }
62
63 static inline void
64 do_data_copy_dequeue(struct vhost_virtqueue *vq)
65 {
66 struct batch_copy_elem *elem = vq->batch_copy_elems;
67 uint16_t count = vq->batch_copy_nb_elems;
68 int i;
69
70 for (i = 0; i < count; i++)
71 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
72
73 vq->batch_copy_nb_elems = 0;
74 }
75
76 static __rte_always_inline void
77 do_flush_shadow_used_ring_split(struct virtio_net *dev,
78 struct vhost_virtqueue *vq,
79 uint16_t to, uint16_t from, uint16_t size)
80 {
81 rte_memcpy(&vq->used->ring[to],
82 &vq->shadow_used_split[from],
83 size * sizeof(struct vring_used_elem));
84 vhost_log_cache_used_vring(dev, vq,
85 offsetof(struct vring_used, ring[to]),
86 size * sizeof(struct vring_used_elem));
87 }
88
89 static __rte_always_inline void
90 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
91 {
92 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
93
94 if (used_idx + vq->shadow_used_idx <= vq->size) {
95 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
96 vq->shadow_used_idx);
97 } else {
98 uint16_t size;
99
100 /* update used ring interval [used_idx, vq->size] */
101 size = vq->size - used_idx;
102 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
103
104 /* update the left half used ring interval [0, left_size] */
105 do_flush_shadow_used_ring_split(dev, vq, 0, size,
106 vq->shadow_used_idx - size);
107 }
108 vq->last_used_idx += vq->shadow_used_idx;
109
110 vhost_log_cache_sync(dev, vq);
111
112 __atomic_add_fetch(&vq->used->idx, vq->shadow_used_idx,
113 __ATOMIC_RELEASE);
114 vq->shadow_used_idx = 0;
115 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
116 sizeof(vq->used->idx));
117 }
118
119 static __rte_always_inline void
120 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
121 uint16_t desc_idx, uint32_t len)
122 {
123 uint16_t i = vq->shadow_used_idx++;
124
125 vq->shadow_used_split[i].id = desc_idx;
126 vq->shadow_used_split[i].len = len;
127 }
128
129 static __rte_always_inline void
130 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
131 struct vhost_virtqueue *vq)
132 {
133 int i;
134 uint16_t used_idx = vq->last_used_idx;
135 uint16_t head_idx = vq->last_used_idx;
136 uint16_t head_flags = 0;
137
138 /* Split loop in two to save memory barriers */
139 for (i = 0; i < vq->shadow_used_idx; i++) {
140 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
141 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
142
143 used_idx += vq->shadow_used_packed[i].count;
144 if (used_idx >= vq->size)
145 used_idx -= vq->size;
146 }
147
148 rte_smp_wmb();
149
150 for (i = 0; i < vq->shadow_used_idx; i++) {
151 uint16_t flags;
152
153 if (vq->shadow_used_packed[i].len)
154 flags = VRING_DESC_F_WRITE;
155 else
156 flags = 0;
157
158 if (vq->used_wrap_counter) {
159 flags |= VRING_DESC_F_USED;
160 flags |= VRING_DESC_F_AVAIL;
161 } else {
162 flags &= ~VRING_DESC_F_USED;
163 flags &= ~VRING_DESC_F_AVAIL;
164 }
165
166 if (i > 0) {
167 vq->desc_packed[vq->last_used_idx].flags = flags;
168
169 vhost_log_cache_used_vring(dev, vq,
170 vq->last_used_idx *
171 sizeof(struct vring_packed_desc),
172 sizeof(struct vring_packed_desc));
173 } else {
174 head_idx = vq->last_used_idx;
175 head_flags = flags;
176 }
177
178 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
179 }
180
181 vq->desc_packed[head_idx].flags = head_flags;
182
183 vhost_log_cache_used_vring(dev, vq,
184 head_idx *
185 sizeof(struct vring_packed_desc),
186 sizeof(struct vring_packed_desc));
187
188 vq->shadow_used_idx = 0;
189 vhost_log_cache_sync(dev, vq);
190 }
191
192 static __rte_always_inline void
193 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
194 struct vhost_virtqueue *vq)
195 {
196 struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
197
198 vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
199 rte_smp_wmb();
200 vq->desc_packed[vq->shadow_last_used_idx].flags = used_elem->flags;
201
202 vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
203 sizeof(struct vring_packed_desc),
204 sizeof(struct vring_packed_desc));
205 vq->shadow_used_idx = 0;
206 vhost_log_cache_sync(dev, vq);
207 }
208
209 static __rte_always_inline void
210 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
211 struct vhost_virtqueue *vq,
212 uint64_t *lens,
213 uint16_t *ids)
214 {
215 uint16_t i;
216 uint16_t flags;
217
218 if (vq->shadow_used_idx) {
219 do_data_copy_enqueue(dev, vq);
220 vhost_flush_enqueue_shadow_packed(dev, vq);
221 }
222
223 flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
224
225 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
226 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
227 vq->desc_packed[vq->last_used_idx + i].len = lens[i];
228 }
229
230 rte_smp_wmb();
231
232 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
233 vq->desc_packed[vq->last_used_idx + i].flags = flags;
234
235 vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
236 sizeof(struct vring_packed_desc),
237 sizeof(struct vring_packed_desc) *
238 PACKED_BATCH_SIZE);
239 vhost_log_cache_sync(dev, vq);
240
241 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
242 }
243
244 static __rte_always_inline void
245 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
246 uint16_t id)
247 {
248 vq->shadow_used_packed[0].id = id;
249
250 if (!vq->shadow_used_idx) {
251 vq->shadow_last_used_idx = vq->last_used_idx;
252 vq->shadow_used_packed[0].flags =
253 PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
254 vq->shadow_used_packed[0].len = 0;
255 vq->shadow_used_packed[0].count = 1;
256 vq->shadow_used_idx++;
257 }
258
259 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
260 }
261
262 static __rte_always_inline void
263 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
264 struct vhost_virtqueue *vq,
265 uint16_t *ids)
266 {
267 uint16_t flags;
268 uint16_t i;
269 uint16_t begin;
270
271 flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
272
273 if (!vq->shadow_used_idx) {
274 vq->shadow_last_used_idx = vq->last_used_idx;
275 vq->shadow_used_packed[0].id = ids[0];
276 vq->shadow_used_packed[0].len = 0;
277 vq->shadow_used_packed[0].count = 1;
278 vq->shadow_used_packed[0].flags = flags;
279 vq->shadow_used_idx++;
280 begin = 1;
281 } else
282 begin = 0;
283
284 vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
285 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
286 vq->desc_packed[vq->last_used_idx + i].len = 0;
287 }
288
289 rte_smp_wmb();
290 vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
291 vq->desc_packed[vq->last_used_idx + i].flags = flags;
292
293 vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
294 sizeof(struct vring_packed_desc),
295 sizeof(struct vring_packed_desc) *
296 PACKED_BATCH_SIZE);
297 vhost_log_cache_sync(dev, vq);
298
299 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
300 }
301
302 static __rte_always_inline void
303 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
304 uint16_t buf_id,
305 uint16_t count)
306 {
307 uint16_t flags;
308
309 flags = vq->desc_packed[vq->last_used_idx].flags;
310 if (vq->used_wrap_counter) {
311 flags |= VRING_DESC_F_USED;
312 flags |= VRING_DESC_F_AVAIL;
313 } else {
314 flags &= ~VRING_DESC_F_USED;
315 flags &= ~VRING_DESC_F_AVAIL;
316 }
317
318 if (!vq->shadow_used_idx) {
319 vq->shadow_last_used_idx = vq->last_used_idx;
320
321 vq->shadow_used_packed[0].id = buf_id;
322 vq->shadow_used_packed[0].len = 0;
323 vq->shadow_used_packed[0].flags = flags;
324 vq->shadow_used_idx++;
325 } else {
326 vq->desc_packed[vq->last_used_idx].id = buf_id;
327 vq->desc_packed[vq->last_used_idx].len = 0;
328 vq->desc_packed[vq->last_used_idx].flags = flags;
329 }
330
331 vq_inc_last_used_packed(vq, count);
332 }
333
334 static __rte_always_inline void
335 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
336 uint16_t buf_id,
337 uint16_t count)
338 {
339 uint16_t flags;
340
341 vq->shadow_used_packed[0].id = buf_id;
342
343 flags = vq->desc_packed[vq->last_used_idx].flags;
344 if (vq->used_wrap_counter) {
345 flags |= VRING_DESC_F_USED;
346 flags |= VRING_DESC_F_AVAIL;
347 } else {
348 flags &= ~VRING_DESC_F_USED;
349 flags &= ~VRING_DESC_F_AVAIL;
350 }
351
352 if (!vq->shadow_used_idx) {
353 vq->shadow_last_used_idx = vq->last_used_idx;
354 vq->shadow_used_packed[0].len = 0;
355 vq->shadow_used_packed[0].flags = flags;
356 vq->shadow_used_idx++;
357 }
358
359 vq_inc_last_used_packed(vq, count);
360 }
361
362 static __rte_always_inline void
363 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
364 struct vhost_virtqueue *vq,
365 uint32_t len[],
366 uint16_t id[],
367 uint16_t count[],
368 uint16_t num_buffers)
369 {
370 uint16_t i;
371 for (i = 0; i < num_buffers; i++) {
372 /* enqueue shadow flush action aligned with batch num */
373 if (!vq->shadow_used_idx)
374 vq->shadow_aligned_idx = vq->last_used_idx &
375 PACKED_BATCH_MASK;
376 vq->shadow_used_packed[vq->shadow_used_idx].id = id[i];
377 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
378 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
379 vq->shadow_aligned_idx += count[i];
380 vq->shadow_used_idx++;
381 }
382
383 if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
384 do_data_copy_enqueue(dev, vq);
385 vhost_flush_enqueue_shadow_packed(dev, vq);
386 }
387 }
388
389 /* avoid write operation when necessary, to lessen cache issues */
390 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
391 if ((var) != (val)) \
392 (var) = (val); \
393 } while (0)
394
395 static __rte_always_inline void
396 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
397 {
398 uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
399
400 if (m_buf->ol_flags & PKT_TX_TCP_SEG)
401 csum_l4 |= PKT_TX_TCP_CKSUM;
402
403 if (csum_l4) {
404 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
405 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
406
407 switch (csum_l4) {
408 case PKT_TX_TCP_CKSUM:
409 net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
410 cksum));
411 break;
412 case PKT_TX_UDP_CKSUM:
413 net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
414 dgram_cksum));
415 break;
416 case PKT_TX_SCTP_CKSUM:
417 net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
418 cksum));
419 break;
420 }
421 } else {
422 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
423 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
424 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
425 }
426
427 /* IP cksum verification cannot be bypassed, then calculate here */
428 if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
429 struct rte_ipv4_hdr *ipv4_hdr;
430
431 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
432 m_buf->l2_len);
433 ipv4_hdr->hdr_checksum = 0;
434 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
435 }
436
437 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
438 if (m_buf->ol_flags & PKT_TX_IPV4)
439 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
440 else
441 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
442 net_hdr->gso_size = m_buf->tso_segsz;
443 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
444 + m_buf->l4_len;
445 } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
446 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
447 net_hdr->gso_size = m_buf->tso_segsz;
448 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
449 m_buf->l4_len;
450 } else {
451 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
452 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
453 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
454 }
455 }
456
457 static __rte_always_inline int
458 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
459 struct buf_vector *buf_vec, uint16_t *vec_idx,
460 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
461 {
462 uint16_t vec_id = *vec_idx;
463
464 while (desc_len) {
465 uint64_t desc_addr;
466 uint64_t desc_chunck_len = desc_len;
467
468 if (unlikely(vec_id >= BUF_VECTOR_MAX))
469 return -1;
470
471 desc_addr = vhost_iova_to_vva(dev, vq,
472 desc_iova,
473 &desc_chunck_len,
474 perm);
475 if (unlikely(!desc_addr))
476 return -1;
477
478 rte_prefetch0((void *)(uintptr_t)desc_addr);
479
480 buf_vec[vec_id].buf_iova = desc_iova;
481 buf_vec[vec_id].buf_addr = desc_addr;
482 buf_vec[vec_id].buf_len = desc_chunck_len;
483
484 desc_len -= desc_chunck_len;
485 desc_iova += desc_chunck_len;
486 vec_id++;
487 }
488 *vec_idx = vec_id;
489
490 return 0;
491 }
492
493 static __rte_always_inline int
494 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
495 uint32_t avail_idx, uint16_t *vec_idx,
496 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
497 uint32_t *desc_chain_len, uint8_t perm)
498 {
499 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
500 uint16_t vec_id = *vec_idx;
501 uint32_t len = 0;
502 uint64_t dlen;
503 uint32_t nr_descs = vq->size;
504 uint32_t cnt = 0;
505 struct vring_desc *descs = vq->desc;
506 struct vring_desc *idesc = NULL;
507
508 if (unlikely(idx >= vq->size))
509 return -1;
510
511 *desc_chain_head = idx;
512
513 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
514 dlen = vq->desc[idx].len;
515 nr_descs = dlen / sizeof(struct vring_desc);
516 if (unlikely(nr_descs > vq->size))
517 return -1;
518
519 descs = (struct vring_desc *)(uintptr_t)
520 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
521 &dlen,
522 VHOST_ACCESS_RO);
523 if (unlikely(!descs))
524 return -1;
525
526 if (unlikely(dlen < vq->desc[idx].len)) {
527 /*
528 * The indirect desc table is not contiguous
529 * in process VA space, we have to copy it.
530 */
531 idesc = vhost_alloc_copy_ind_table(dev, vq,
532 vq->desc[idx].addr, vq->desc[idx].len);
533 if (unlikely(!idesc))
534 return -1;
535
536 descs = idesc;
537 }
538
539 idx = 0;
540 }
541
542 while (1) {
543 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
544 free_ind_table(idesc);
545 return -1;
546 }
547
548 len += descs[idx].len;
549
550 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
551 descs[idx].addr, descs[idx].len,
552 perm))) {
553 free_ind_table(idesc);
554 return -1;
555 }
556
557 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
558 break;
559
560 idx = descs[idx].next;
561 }
562
563 *desc_chain_len = len;
564 *vec_idx = vec_id;
565
566 if (unlikely(!!idesc))
567 free_ind_table(idesc);
568
569 return 0;
570 }
571
572 /*
573 * Returns -1 on fail, 0 on success
574 */
575 static inline int
576 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
577 uint32_t size, struct buf_vector *buf_vec,
578 uint16_t *num_buffers, uint16_t avail_head,
579 uint16_t *nr_vec)
580 {
581 uint16_t cur_idx;
582 uint16_t vec_idx = 0;
583 uint16_t max_tries, tries = 0;
584
585 uint16_t head_idx = 0;
586 uint32_t len = 0;
587
588 *num_buffers = 0;
589 cur_idx = vq->last_avail_idx;
590
591 if (rxvq_is_mergeable(dev))
592 max_tries = vq->size - 1;
593 else
594 max_tries = 1;
595
596 while (size > 0) {
597 if (unlikely(cur_idx == avail_head))
598 return -1;
599 /*
600 * if we tried all available ring items, and still
601 * can't get enough buf, it means something abnormal
602 * happened.
603 */
604 if (unlikely(++tries > max_tries))
605 return -1;
606
607 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
608 &vec_idx, buf_vec,
609 &head_idx, &len,
610 VHOST_ACCESS_RW) < 0))
611 return -1;
612 len = RTE_MIN(len, size);
613 update_shadow_used_ring_split(vq, head_idx, len);
614 size -= len;
615
616 cur_idx++;
617 *num_buffers += 1;
618 }
619
620 *nr_vec = vec_idx;
621
622 return 0;
623 }
624
625 static __rte_always_inline int
626 fill_vec_buf_packed_indirect(struct virtio_net *dev,
627 struct vhost_virtqueue *vq,
628 struct vring_packed_desc *desc, uint16_t *vec_idx,
629 struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
630 {
631 uint16_t i;
632 uint32_t nr_descs;
633 uint16_t vec_id = *vec_idx;
634 uint64_t dlen;
635 struct vring_packed_desc *descs, *idescs = NULL;
636
637 dlen = desc->len;
638 descs = (struct vring_packed_desc *)(uintptr_t)
639 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
640 if (unlikely(!descs))
641 return -1;
642
643 if (unlikely(dlen < desc->len)) {
644 /*
645 * The indirect desc table is not contiguous
646 * in process VA space, we have to copy it.
647 */
648 idescs = vhost_alloc_copy_ind_table(dev,
649 vq, desc->addr, desc->len);
650 if (unlikely(!idescs))
651 return -1;
652
653 descs = idescs;
654 }
655
656 nr_descs = desc->len / sizeof(struct vring_packed_desc);
657 if (unlikely(nr_descs >= vq->size)) {
658 free_ind_table(idescs);
659 return -1;
660 }
661
662 for (i = 0; i < nr_descs; i++) {
663 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
664 free_ind_table(idescs);
665 return -1;
666 }
667
668 *len += descs[i].len;
669 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
670 descs[i].addr, descs[i].len,
671 perm)))
672 return -1;
673 }
674 *vec_idx = vec_id;
675
676 if (unlikely(!!idescs))
677 free_ind_table(idescs);
678
679 return 0;
680 }
681
682 static __rte_always_inline int
683 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
684 uint16_t avail_idx, uint16_t *desc_count,
685 struct buf_vector *buf_vec, uint16_t *vec_idx,
686 uint16_t *buf_id, uint32_t *len, uint8_t perm)
687 {
688 bool wrap_counter = vq->avail_wrap_counter;
689 struct vring_packed_desc *descs = vq->desc_packed;
690 uint16_t vec_id = *vec_idx;
691
692 if (avail_idx < vq->last_avail_idx)
693 wrap_counter ^= 1;
694
695 /*
696 * Perform a load-acquire barrier in desc_is_avail to
697 * enforce the ordering between desc flags and desc
698 * content.
699 */
700 if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
701 return -1;
702
703 *desc_count = 0;
704 *len = 0;
705
706 while (1) {
707 if (unlikely(vec_id >= BUF_VECTOR_MAX))
708 return -1;
709
710 if (unlikely(*desc_count >= vq->size))
711 return -1;
712
713 *desc_count += 1;
714 *buf_id = descs[avail_idx].id;
715
716 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
717 if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
718 &descs[avail_idx],
719 &vec_id, buf_vec,
720 len, perm) < 0))
721 return -1;
722 } else {
723 *len += descs[avail_idx].len;
724
725 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
726 descs[avail_idx].addr,
727 descs[avail_idx].len,
728 perm)))
729 return -1;
730 }
731
732 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
733 break;
734
735 if (++avail_idx >= vq->size) {
736 avail_idx -= vq->size;
737 wrap_counter ^= 1;
738 }
739 }
740
741 *vec_idx = vec_id;
742
743 return 0;
744 }
745
746 static __rte_noinline void
747 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
748 struct buf_vector *buf_vec,
749 struct virtio_net_hdr_mrg_rxbuf *hdr)
750 {
751 uint64_t len;
752 uint64_t remain = dev->vhost_hlen;
753 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
754 uint64_t iova = buf_vec->buf_iova;
755
756 while (remain) {
757 len = RTE_MIN(remain,
758 buf_vec->buf_len);
759 dst = buf_vec->buf_addr;
760 rte_memcpy((void *)(uintptr_t)dst,
761 (void *)(uintptr_t)src,
762 len);
763
764 PRINT_PACKET(dev, (uintptr_t)dst,
765 (uint32_t)len, 0);
766 vhost_log_cache_write_iova(dev, vq,
767 iova, len);
768
769 remain -= len;
770 iova += len;
771 src += len;
772 buf_vec++;
773 }
774 }
775
776 static __rte_always_inline int
777 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
778 struct rte_mbuf *m, struct buf_vector *buf_vec,
779 uint16_t nr_vec, uint16_t num_buffers)
780 {
781 uint32_t vec_idx = 0;
782 uint32_t mbuf_offset, mbuf_avail;
783 uint32_t buf_offset, buf_avail;
784 uint64_t buf_addr, buf_iova, buf_len;
785 uint32_t cpy_len;
786 uint64_t hdr_addr;
787 struct rte_mbuf *hdr_mbuf;
788 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
789 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
790 int error = 0;
791
792 if (unlikely(m == NULL)) {
793 error = -1;
794 goto out;
795 }
796
797 buf_addr = buf_vec[vec_idx].buf_addr;
798 buf_iova = buf_vec[vec_idx].buf_iova;
799 buf_len = buf_vec[vec_idx].buf_len;
800
801 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
802 error = -1;
803 goto out;
804 }
805
806 hdr_mbuf = m;
807 hdr_addr = buf_addr;
808 if (unlikely(buf_len < dev->vhost_hlen))
809 hdr = &tmp_hdr;
810 else
811 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
812
813 VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
814 dev->vid, num_buffers);
815
816 if (unlikely(buf_len < dev->vhost_hlen)) {
817 buf_offset = dev->vhost_hlen - buf_len;
818 vec_idx++;
819 buf_addr = buf_vec[vec_idx].buf_addr;
820 buf_iova = buf_vec[vec_idx].buf_iova;
821 buf_len = buf_vec[vec_idx].buf_len;
822 buf_avail = buf_len - buf_offset;
823 } else {
824 buf_offset = dev->vhost_hlen;
825 buf_avail = buf_len - dev->vhost_hlen;
826 }
827
828 mbuf_avail = rte_pktmbuf_data_len(m);
829 mbuf_offset = 0;
830 while (mbuf_avail != 0 || m->next != NULL) {
831 /* done with current buf, get the next one */
832 if (buf_avail == 0) {
833 vec_idx++;
834 if (unlikely(vec_idx >= nr_vec)) {
835 error = -1;
836 goto out;
837 }
838
839 buf_addr = buf_vec[vec_idx].buf_addr;
840 buf_iova = buf_vec[vec_idx].buf_iova;
841 buf_len = buf_vec[vec_idx].buf_len;
842
843 buf_offset = 0;
844 buf_avail = buf_len;
845 }
846
847 /* done with current mbuf, get the next one */
848 if (mbuf_avail == 0) {
849 m = m->next;
850
851 mbuf_offset = 0;
852 mbuf_avail = rte_pktmbuf_data_len(m);
853 }
854
855 if (hdr_addr) {
856 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
857 if (rxvq_is_mergeable(dev))
858 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
859 num_buffers);
860
861 if (unlikely(hdr == &tmp_hdr)) {
862 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
863 } else {
864 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
865 dev->vhost_hlen, 0);
866 vhost_log_cache_write_iova(dev, vq,
867 buf_vec[0].buf_iova,
868 dev->vhost_hlen);
869 }
870
871 hdr_addr = 0;
872 }
873
874 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
875
876 if (likely(cpy_len > MAX_BATCH_LEN ||
877 vq->batch_copy_nb_elems >= vq->size)) {
878 rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
879 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
880 cpy_len);
881 vhost_log_cache_write_iova(dev, vq,
882 buf_iova + buf_offset,
883 cpy_len);
884 PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
885 cpy_len, 0);
886 } else {
887 batch_copy[vq->batch_copy_nb_elems].dst =
888 (void *)((uintptr_t)(buf_addr + buf_offset));
889 batch_copy[vq->batch_copy_nb_elems].src =
890 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
891 batch_copy[vq->batch_copy_nb_elems].log_addr =
892 buf_iova + buf_offset;
893 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
894 vq->batch_copy_nb_elems++;
895 }
896
897 mbuf_avail -= cpy_len;
898 mbuf_offset += cpy_len;
899 buf_avail -= cpy_len;
900 buf_offset += cpy_len;
901 }
902
903 out:
904
905 return error;
906 }
907
908 static __rte_always_inline int
909 vhost_enqueue_single_packed(struct virtio_net *dev,
910 struct vhost_virtqueue *vq,
911 struct rte_mbuf *pkt,
912 struct buf_vector *buf_vec,
913 uint16_t *nr_descs)
914 {
915 uint16_t nr_vec = 0;
916 uint16_t avail_idx = vq->last_avail_idx;
917 uint16_t max_tries, tries = 0;
918 uint16_t buf_id = 0;
919 uint32_t len = 0;
920 uint16_t desc_count;
921 uint32_t size = pkt->pkt_len + dev->vhost_hlen;
922 uint16_t num_buffers = 0;
923 uint32_t buffer_len[vq->size];
924 uint16_t buffer_buf_id[vq->size];
925 uint16_t buffer_desc_count[vq->size];
926
927 if (rxvq_is_mergeable(dev))
928 max_tries = vq->size - 1;
929 else
930 max_tries = 1;
931
932 while (size > 0) {
933 /*
934 * if we tried all available ring items, and still
935 * can't get enough buf, it means something abnormal
936 * happened.
937 */
938 if (unlikely(++tries > max_tries))
939 return -1;
940
941 if (unlikely(fill_vec_buf_packed(dev, vq,
942 avail_idx, &desc_count,
943 buf_vec, &nr_vec,
944 &buf_id, &len,
945 VHOST_ACCESS_RW) < 0))
946 return -1;
947
948 len = RTE_MIN(len, size);
949 size -= len;
950
951 buffer_len[num_buffers] = len;
952 buffer_buf_id[num_buffers] = buf_id;
953 buffer_desc_count[num_buffers] = desc_count;
954 num_buffers += 1;
955
956 *nr_descs += desc_count;
957 avail_idx += desc_count;
958 if (avail_idx >= vq->size)
959 avail_idx -= vq->size;
960 }
961
962 if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
963 return -1;
964
965 vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
966 buffer_desc_count, num_buffers);
967
968 return 0;
969 }
970
971 static __rte_noinline uint32_t
972 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
973 struct rte_mbuf **pkts, uint32_t count)
974 {
975 uint32_t pkt_idx = 0;
976 uint16_t num_buffers;
977 struct buf_vector buf_vec[BUF_VECTOR_MAX];
978 uint16_t avail_head;
979
980 /*
981 * The ordering between avail index and
982 * desc reads needs to be enforced.
983 */
984 avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
985
986 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
987
988 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
989 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
990 uint16_t nr_vec = 0;
991
992 if (unlikely(reserve_avail_buf_split(dev, vq,
993 pkt_len, buf_vec, &num_buffers,
994 avail_head, &nr_vec) < 0)) {
995 VHOST_LOG_DATA(DEBUG,
996 "(%d) failed to get enough desc from vring\n",
997 dev->vid);
998 vq->shadow_used_idx -= num_buffers;
999 break;
1000 }
1001
1002 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1003 dev->vid, vq->last_avail_idx,
1004 vq->last_avail_idx + num_buffers);
1005
1006 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1007 buf_vec, nr_vec,
1008 num_buffers) < 0) {
1009 vq->shadow_used_idx -= num_buffers;
1010 break;
1011 }
1012
1013 vq->last_avail_idx += num_buffers;
1014 }
1015
1016 do_data_copy_enqueue(dev, vq);
1017
1018 if (likely(vq->shadow_used_idx)) {
1019 flush_shadow_used_ring_split(dev, vq);
1020 vhost_vring_call_split(dev, vq);
1021 }
1022
1023 return pkt_idx;
1024 }
1025
1026 static __rte_always_inline int
1027 virtio_dev_rx_batch_packed(struct virtio_net *dev,
1028 struct vhost_virtqueue *vq,
1029 struct rte_mbuf **pkts)
1030 {
1031 bool wrap_counter = vq->avail_wrap_counter;
1032 struct vring_packed_desc *descs = vq->desc_packed;
1033 uint16_t avail_idx = vq->last_avail_idx;
1034 uint64_t desc_addrs[PACKED_BATCH_SIZE];
1035 struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1036 uint32_t buf_offset = dev->vhost_hlen;
1037 uint64_t lens[PACKED_BATCH_SIZE];
1038 uint16_t ids[PACKED_BATCH_SIZE];
1039 uint16_t i;
1040
1041 if (unlikely(avail_idx & PACKED_BATCH_MASK))
1042 return -1;
1043
1044 if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1045 return -1;
1046
1047 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1048 if (unlikely(pkts[i]->next != NULL))
1049 return -1;
1050 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1051 wrap_counter)))
1052 return -1;
1053 }
1054
1055 rte_smp_rmb();
1056
1057 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1058 lens[i] = descs[avail_idx + i].len;
1059
1060 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1061 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1062 return -1;
1063 }
1064
1065 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1066 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1067 descs[avail_idx + i].addr,
1068 &lens[i],
1069 VHOST_ACCESS_RW);
1070
1071 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1072 if (unlikely(!desc_addrs[i]))
1073 return -1;
1074 if (unlikely(lens[i] != descs[avail_idx + i].len))
1075 return -1;
1076 }
1077
1078 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1079 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1080 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1081 (uintptr_t)desc_addrs[i];
1082 lens[i] = pkts[i]->pkt_len + dev->vhost_hlen;
1083 }
1084
1085 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1086 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1087
1088 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1089
1090 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1091 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1092 rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1093 pkts[i]->pkt_len);
1094 }
1095
1096 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1097 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1098 lens[i]);
1099
1100 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1101 ids[i] = descs[avail_idx + i].id;
1102
1103 vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1104
1105 return 0;
1106 }
1107
1108 static __rte_always_inline int16_t
1109 virtio_dev_rx_single_packed(struct virtio_net *dev,
1110 struct vhost_virtqueue *vq,
1111 struct rte_mbuf *pkt)
1112 {
1113 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1114 uint16_t nr_descs = 0;
1115
1116 rte_smp_rmb();
1117 if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1118 &nr_descs) < 0)) {
1119 VHOST_LOG_DATA(DEBUG,
1120 "(%d) failed to get enough desc from vring\n",
1121 dev->vid);
1122 return -1;
1123 }
1124
1125 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1126 dev->vid, vq->last_avail_idx,
1127 vq->last_avail_idx + nr_descs);
1128
1129 vq_inc_last_avail_packed(vq, nr_descs);
1130
1131 return 0;
1132 }
1133
1134 static __rte_noinline uint32_t
1135 virtio_dev_rx_packed(struct virtio_net *dev,
1136 struct vhost_virtqueue *vq,
1137 struct rte_mbuf **pkts,
1138 uint32_t count)
1139 {
1140 uint32_t pkt_idx = 0;
1141 uint32_t remained = count;
1142
1143 do {
1144 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1145
1146 if (remained >= PACKED_BATCH_SIZE) {
1147 if (!virtio_dev_rx_batch_packed(dev, vq,
1148 &pkts[pkt_idx])) {
1149 pkt_idx += PACKED_BATCH_SIZE;
1150 remained -= PACKED_BATCH_SIZE;
1151 continue;
1152 }
1153 }
1154
1155 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1156 break;
1157 pkt_idx++;
1158 remained--;
1159
1160 } while (pkt_idx < count);
1161
1162 if (vq->shadow_used_idx) {
1163 do_data_copy_enqueue(dev, vq);
1164 vhost_flush_enqueue_shadow_packed(dev, vq);
1165 }
1166
1167 if (pkt_idx)
1168 vhost_vring_call_packed(dev, vq);
1169
1170 return pkt_idx;
1171 }
1172
1173 static __rte_always_inline uint32_t
1174 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1175 struct rte_mbuf **pkts, uint32_t count)
1176 {
1177 struct vhost_virtqueue *vq;
1178 uint32_t nb_tx = 0;
1179
1180 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1181 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1182 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1183 dev->vid, __func__, queue_id);
1184 return 0;
1185 }
1186
1187 vq = dev->virtqueue[queue_id];
1188
1189 rte_spinlock_lock(&vq->access_lock);
1190
1191 if (unlikely(vq->enabled == 0))
1192 goto out_access_unlock;
1193
1194 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1195 vhost_user_iotlb_rd_lock(vq);
1196
1197 if (unlikely(vq->access_ok == 0))
1198 if (unlikely(vring_translate(dev, vq) < 0))
1199 goto out;
1200
1201 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1202 if (count == 0)
1203 goto out;
1204
1205 if (vq_is_packed(dev))
1206 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1207 else
1208 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1209
1210 out:
1211 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1212 vhost_user_iotlb_rd_unlock(vq);
1213
1214 out_access_unlock:
1215 rte_spinlock_unlock(&vq->access_lock);
1216
1217 return nb_tx;
1218 }
1219
1220 uint16_t
1221 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1222 struct rte_mbuf **pkts, uint16_t count)
1223 {
1224 struct virtio_net *dev = get_device(vid);
1225
1226 if (!dev)
1227 return 0;
1228
1229 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1230 VHOST_LOG_DATA(ERR,
1231 "(%d) %s: built-in vhost net backend is disabled.\n",
1232 dev->vid, __func__);
1233 return 0;
1234 }
1235
1236 return virtio_dev_rx(dev, queue_id, pkts, count);
1237 }
1238
1239 static inline bool
1240 virtio_net_with_host_offload(struct virtio_net *dev)
1241 {
1242 if (dev->features &
1243 ((1ULL << VIRTIO_NET_F_CSUM) |
1244 (1ULL << VIRTIO_NET_F_HOST_ECN) |
1245 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
1246 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
1247 (1ULL << VIRTIO_NET_F_HOST_UFO)))
1248 return true;
1249
1250 return false;
1251 }
1252
1253 static void
1254 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
1255 {
1256 struct rte_ipv4_hdr *ipv4_hdr;
1257 struct rte_ipv6_hdr *ipv6_hdr;
1258 void *l3_hdr = NULL;
1259 struct rte_ether_hdr *eth_hdr;
1260 uint16_t ethertype;
1261
1262 eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1263
1264 m->l2_len = sizeof(struct rte_ether_hdr);
1265 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
1266
1267 if (ethertype == RTE_ETHER_TYPE_VLAN) {
1268 struct rte_vlan_hdr *vlan_hdr =
1269 (struct rte_vlan_hdr *)(eth_hdr + 1);
1270
1271 m->l2_len += sizeof(struct rte_vlan_hdr);
1272 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
1273 }
1274
1275 l3_hdr = (char *)eth_hdr + m->l2_len;
1276
1277 switch (ethertype) {
1278 case RTE_ETHER_TYPE_IPV4:
1279 ipv4_hdr = l3_hdr;
1280 *l4_proto = ipv4_hdr->next_proto_id;
1281 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
1282 *l4_hdr = (char *)l3_hdr + m->l3_len;
1283 m->ol_flags |= PKT_TX_IPV4;
1284 break;
1285 case RTE_ETHER_TYPE_IPV6:
1286 ipv6_hdr = l3_hdr;
1287 *l4_proto = ipv6_hdr->proto;
1288 m->l3_len = sizeof(struct rte_ipv6_hdr);
1289 *l4_hdr = (char *)l3_hdr + m->l3_len;
1290 m->ol_flags |= PKT_TX_IPV6;
1291 break;
1292 default:
1293 m->l3_len = 0;
1294 *l4_proto = 0;
1295 *l4_hdr = NULL;
1296 break;
1297 }
1298 }
1299
1300 static __rte_always_inline void
1301 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1302 {
1303 uint16_t l4_proto = 0;
1304 void *l4_hdr = NULL;
1305 struct rte_tcp_hdr *tcp_hdr = NULL;
1306
1307 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1308 return;
1309
1310 parse_ethernet(m, &l4_proto, &l4_hdr);
1311 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1312 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1313 switch (hdr->csum_offset) {
1314 case (offsetof(struct rte_tcp_hdr, cksum)):
1315 if (l4_proto == IPPROTO_TCP)
1316 m->ol_flags |= PKT_TX_TCP_CKSUM;
1317 break;
1318 case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1319 if (l4_proto == IPPROTO_UDP)
1320 m->ol_flags |= PKT_TX_UDP_CKSUM;
1321 break;
1322 case (offsetof(struct rte_sctp_hdr, cksum)):
1323 if (l4_proto == IPPROTO_SCTP)
1324 m->ol_flags |= PKT_TX_SCTP_CKSUM;
1325 break;
1326 default:
1327 break;
1328 }
1329 }
1330 }
1331
1332 if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1333 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1334 case VIRTIO_NET_HDR_GSO_TCPV4:
1335 case VIRTIO_NET_HDR_GSO_TCPV6:
1336 tcp_hdr = l4_hdr;
1337 m->ol_flags |= PKT_TX_TCP_SEG;
1338 m->tso_segsz = hdr->gso_size;
1339 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1340 break;
1341 case VIRTIO_NET_HDR_GSO_UDP:
1342 m->ol_flags |= PKT_TX_UDP_SEG;
1343 m->tso_segsz = hdr->gso_size;
1344 m->l4_len = sizeof(struct rte_udp_hdr);
1345 break;
1346 default:
1347 VHOST_LOG_DATA(WARNING,
1348 "unsupported gso type %u.\n", hdr->gso_type);
1349 break;
1350 }
1351 }
1352 }
1353
1354 static __rte_noinline void
1355 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1356 struct buf_vector *buf_vec)
1357 {
1358 uint64_t len;
1359 uint64_t remain = sizeof(struct virtio_net_hdr);
1360 uint64_t src;
1361 uint64_t dst = (uint64_t)(uintptr_t)hdr;
1362
1363 while (remain) {
1364 len = RTE_MIN(remain, buf_vec->buf_len);
1365 src = buf_vec->buf_addr;
1366 rte_memcpy((void *)(uintptr_t)dst,
1367 (void *)(uintptr_t)src, len);
1368
1369 remain -= len;
1370 dst += len;
1371 buf_vec++;
1372 }
1373 }
1374
1375 static __rte_always_inline int
1376 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1377 struct buf_vector *buf_vec, uint16_t nr_vec,
1378 struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1379 {
1380 uint32_t buf_avail, buf_offset;
1381 uint64_t buf_addr, buf_iova, buf_len;
1382 uint32_t mbuf_avail, mbuf_offset;
1383 uint32_t cpy_len;
1384 struct rte_mbuf *cur = m, *prev = m;
1385 struct virtio_net_hdr tmp_hdr;
1386 struct virtio_net_hdr *hdr = NULL;
1387 /* A counter to avoid desc dead loop chain */
1388 uint16_t vec_idx = 0;
1389 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1390 int error = 0;
1391
1392 buf_addr = buf_vec[vec_idx].buf_addr;
1393 buf_iova = buf_vec[vec_idx].buf_iova;
1394 buf_len = buf_vec[vec_idx].buf_len;
1395
1396 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1397 error = -1;
1398 goto out;
1399 }
1400
1401 if (virtio_net_with_host_offload(dev)) {
1402 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1403 /*
1404 * No luck, the virtio-net header doesn't fit
1405 * in a contiguous virtual area.
1406 */
1407 copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1408 hdr = &tmp_hdr;
1409 } else {
1410 hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1411 }
1412 }
1413
1414 /*
1415 * A virtio driver normally uses at least 2 desc buffers
1416 * for Tx: the first for storing the header, and others
1417 * for storing the data.
1418 */
1419 if (unlikely(buf_len < dev->vhost_hlen)) {
1420 buf_offset = dev->vhost_hlen - buf_len;
1421 vec_idx++;
1422 buf_addr = buf_vec[vec_idx].buf_addr;
1423 buf_iova = buf_vec[vec_idx].buf_iova;
1424 buf_len = buf_vec[vec_idx].buf_len;
1425 buf_avail = buf_len - buf_offset;
1426 } else if (buf_len == dev->vhost_hlen) {
1427 if (unlikely(++vec_idx >= nr_vec))
1428 goto out;
1429 buf_addr = buf_vec[vec_idx].buf_addr;
1430 buf_iova = buf_vec[vec_idx].buf_iova;
1431 buf_len = buf_vec[vec_idx].buf_len;
1432
1433 buf_offset = 0;
1434 buf_avail = buf_len;
1435 } else {
1436 buf_offset = dev->vhost_hlen;
1437 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1438 }
1439
1440 PRINT_PACKET(dev,
1441 (uintptr_t)(buf_addr + buf_offset),
1442 (uint32_t)buf_avail, 0);
1443
1444 mbuf_offset = 0;
1445 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
1446 while (1) {
1447 uint64_t hpa;
1448
1449 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1450
1451 /*
1452 * A desc buf might across two host physical pages that are
1453 * not continuous. In such case (gpa_to_hpa returns 0), data
1454 * will be copied even though zero copy is enabled.
1455 */
1456 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1457 buf_iova + buf_offset, cpy_len)))) {
1458 cur->data_len = cpy_len;
1459 cur->data_off = 0;
1460 cur->buf_addr =
1461 (void *)(uintptr_t)(buf_addr + buf_offset);
1462 cur->buf_iova = hpa;
1463
1464 /*
1465 * In zero copy mode, one mbuf can only reference data
1466 * for one or partial of one desc buff.
1467 */
1468 mbuf_avail = cpy_len;
1469 } else {
1470 if (likely(cpy_len > MAX_BATCH_LEN ||
1471 vq->batch_copy_nb_elems >= vq->size ||
1472 (hdr && cur == m))) {
1473 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1474 mbuf_offset),
1475 (void *)((uintptr_t)(buf_addr +
1476 buf_offset)),
1477 cpy_len);
1478 } else {
1479 batch_copy[vq->batch_copy_nb_elems].dst =
1480 rte_pktmbuf_mtod_offset(cur, void *,
1481 mbuf_offset);
1482 batch_copy[vq->batch_copy_nb_elems].src =
1483 (void *)((uintptr_t)(buf_addr +
1484 buf_offset));
1485 batch_copy[vq->batch_copy_nb_elems].len =
1486 cpy_len;
1487 vq->batch_copy_nb_elems++;
1488 }
1489 }
1490
1491 mbuf_avail -= cpy_len;
1492 mbuf_offset += cpy_len;
1493 buf_avail -= cpy_len;
1494 buf_offset += cpy_len;
1495
1496 /* This buf reaches to its end, get the next one */
1497 if (buf_avail == 0) {
1498 if (++vec_idx >= nr_vec)
1499 break;
1500
1501 buf_addr = buf_vec[vec_idx].buf_addr;
1502 buf_iova = buf_vec[vec_idx].buf_iova;
1503 buf_len = buf_vec[vec_idx].buf_len;
1504
1505 buf_offset = 0;
1506 buf_avail = buf_len;
1507
1508 PRINT_PACKET(dev, (uintptr_t)buf_addr,
1509 (uint32_t)buf_avail, 0);
1510 }
1511
1512 /*
1513 * This mbuf reaches to its end, get a new one
1514 * to hold more data.
1515 */
1516 if (mbuf_avail == 0) {
1517 cur = rte_pktmbuf_alloc(mbuf_pool);
1518 if (unlikely(cur == NULL)) {
1519 VHOST_LOG_DATA(ERR, "Failed to "
1520 "allocate memory for mbuf.\n");
1521 error = -1;
1522 goto out;
1523 }
1524 if (unlikely(dev->dequeue_zero_copy))
1525 rte_mbuf_refcnt_update(cur, 1);
1526
1527 prev->next = cur;
1528 prev->data_len = mbuf_offset;
1529 m->nb_segs += 1;
1530 m->pkt_len += mbuf_offset;
1531 prev = cur;
1532
1533 mbuf_offset = 0;
1534 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1535 }
1536 }
1537
1538 prev->data_len = mbuf_offset;
1539 m->pkt_len += mbuf_offset;
1540
1541 if (hdr)
1542 vhost_dequeue_offload(hdr, m);
1543
1544 out:
1545
1546 return error;
1547 }
1548
1549 static __rte_always_inline struct zcopy_mbuf *
1550 get_zmbuf(struct vhost_virtqueue *vq)
1551 {
1552 uint16_t i;
1553 uint16_t last;
1554 int tries = 0;
1555
1556 /* search [last_zmbuf_idx, zmbuf_size) */
1557 i = vq->last_zmbuf_idx;
1558 last = vq->zmbuf_size;
1559
1560 again:
1561 for (; i < last; i++) {
1562 if (vq->zmbufs[i].in_use == 0) {
1563 vq->last_zmbuf_idx = i + 1;
1564 vq->zmbufs[i].in_use = 1;
1565 return &vq->zmbufs[i];
1566 }
1567 }
1568
1569 tries++;
1570 if (tries == 1) {
1571 /* search [0, last_zmbuf_idx) */
1572 i = 0;
1573 last = vq->last_zmbuf_idx;
1574 goto again;
1575 }
1576
1577 return NULL;
1578 }
1579
1580 static void
1581 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
1582 {
1583 rte_free(opaque);
1584 }
1585
1586 static int
1587 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
1588 {
1589 struct rte_mbuf_ext_shared_info *shinfo = NULL;
1590 uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
1591 uint16_t buf_len;
1592 rte_iova_t iova;
1593 void *buf;
1594
1595 /* Try to use pkt buffer to store shinfo to reduce the amount of memory
1596 * required, otherwise store shinfo in the new buffer.
1597 */
1598 if (rte_pktmbuf_tailroom(pkt) >= sizeof(*shinfo))
1599 shinfo = rte_pktmbuf_mtod(pkt,
1600 struct rte_mbuf_ext_shared_info *);
1601 else {
1602 total_len += sizeof(*shinfo) + sizeof(uintptr_t);
1603 total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
1604 }
1605
1606 if (unlikely(total_len > UINT16_MAX))
1607 return -ENOSPC;
1608
1609 buf_len = total_len;
1610 buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
1611 if (unlikely(buf == NULL))
1612 return -ENOMEM;
1613
1614 /* Initialize shinfo */
1615 if (shinfo) {
1616 shinfo->free_cb = virtio_dev_extbuf_free;
1617 shinfo->fcb_opaque = buf;
1618 rte_mbuf_ext_refcnt_set(shinfo, 1);
1619 } else {
1620 shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
1621 virtio_dev_extbuf_free, buf);
1622 if (unlikely(shinfo == NULL)) {
1623 rte_free(buf);
1624 VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
1625 return -1;
1626 }
1627 }
1628
1629 iova = rte_malloc_virt2iova(buf);
1630 rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
1631 rte_pktmbuf_reset_headroom(pkt);
1632
1633 return 0;
1634 }
1635
1636 /*
1637 * Allocate a host supported pktmbuf.
1638 */
1639 static __rte_always_inline struct rte_mbuf *
1640 virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
1641 uint32_t data_len)
1642 {
1643 struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
1644
1645 if (unlikely(pkt == NULL)) {
1646 VHOST_LOG_DATA(ERR,
1647 "Failed to allocate memory for mbuf.\n");
1648 return NULL;
1649 }
1650
1651 if (rte_pktmbuf_tailroom(pkt) >= data_len)
1652 return pkt;
1653
1654 /* attach an external buffer if supported */
1655 if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
1656 return pkt;
1657
1658 /* check if chained buffers are allowed */
1659 if (!dev->linearbuf)
1660 return pkt;
1661
1662 /* Data doesn't fit into the buffer and the host supports
1663 * only linear buffers
1664 */
1665 rte_pktmbuf_free(pkt);
1666
1667 return NULL;
1668 }
1669
1670 static __rte_noinline uint16_t
1671 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1672 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1673 {
1674 uint16_t i;
1675 uint16_t free_entries;
1676 uint16_t dropped = 0;
1677 static bool allocerr_warned;
1678
1679 if (unlikely(dev->dequeue_zero_copy)) {
1680 struct zcopy_mbuf *zmbuf, *next;
1681
1682 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1683 zmbuf != NULL; zmbuf = next) {
1684 next = TAILQ_NEXT(zmbuf, next);
1685
1686 if (mbuf_is_consumed(zmbuf->mbuf)) {
1687 update_shadow_used_ring_split(vq,
1688 zmbuf->desc_idx, 0);
1689 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1690 restore_mbuf(zmbuf->mbuf);
1691 rte_pktmbuf_free(zmbuf->mbuf);
1692 put_zmbuf(zmbuf);
1693 vq->nr_zmbuf -= 1;
1694 }
1695 }
1696
1697 if (likely(vq->shadow_used_idx)) {
1698 flush_shadow_used_ring_split(dev, vq);
1699 vhost_vring_call_split(dev, vq);
1700 }
1701 }
1702
1703 /*
1704 * The ordering between avail index and
1705 * desc reads needs to be enforced.
1706 */
1707 free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
1708 vq->last_avail_idx;
1709 if (free_entries == 0)
1710 return 0;
1711
1712 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1713
1714 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1715
1716 count = RTE_MIN(count, MAX_PKT_BURST);
1717 count = RTE_MIN(count, free_entries);
1718 VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
1719 dev->vid, count);
1720
1721 for (i = 0; i < count; i++) {
1722 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1723 uint16_t head_idx;
1724 uint32_t buf_len;
1725 uint16_t nr_vec = 0;
1726 int err;
1727
1728 if (unlikely(fill_vec_buf_split(dev, vq,
1729 vq->last_avail_idx + i,
1730 &nr_vec, buf_vec,
1731 &head_idx, &buf_len,
1732 VHOST_ACCESS_RO) < 0))
1733 break;
1734
1735 if (likely(dev->dequeue_zero_copy == 0))
1736 update_shadow_used_ring_split(vq, head_idx, 0);
1737
1738 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
1739 if (unlikely(pkts[i] == NULL)) {
1740 /*
1741 * mbuf allocation fails for jumbo packets when external
1742 * buffer allocation is not allowed and linear buffer
1743 * is required. Drop this packet.
1744 */
1745 if (!allocerr_warned) {
1746 VHOST_LOG_DATA(ERR,
1747 "Failed mbuf alloc of size %d from %s on %s.\n",
1748 buf_len, mbuf_pool->name, dev->ifname);
1749 allocerr_warned = true;
1750 }
1751 dropped += 1;
1752 i++;
1753 break;
1754 }
1755
1756 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1757 mbuf_pool);
1758 if (unlikely(err)) {
1759 rte_pktmbuf_free(pkts[i]);
1760 if (!allocerr_warned) {
1761 VHOST_LOG_DATA(ERR,
1762 "Failed to copy desc to mbuf on %s.\n",
1763 dev->ifname);
1764 allocerr_warned = true;
1765 }
1766 dropped += 1;
1767 i++;
1768 break;
1769 }
1770
1771 if (unlikely(dev->dequeue_zero_copy)) {
1772 struct zcopy_mbuf *zmbuf;
1773
1774 zmbuf = get_zmbuf(vq);
1775 if (!zmbuf) {
1776 rte_pktmbuf_free(pkts[i]);
1777 dropped += 1;
1778 i++;
1779 break;
1780 }
1781 zmbuf->mbuf = pkts[i];
1782 zmbuf->desc_idx = head_idx;
1783
1784 /*
1785 * Pin lock the mbuf; we will check later to see
1786 * whether the mbuf is freed (when we are the last
1787 * user) or not. If that's the case, we then could
1788 * update the used ring safely.
1789 */
1790 rte_mbuf_refcnt_update(pkts[i], 1);
1791
1792 vq->nr_zmbuf += 1;
1793 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1794 }
1795 }
1796 vq->last_avail_idx += i;
1797
1798 if (likely(dev->dequeue_zero_copy == 0)) {
1799 do_data_copy_dequeue(vq);
1800 if (unlikely(i < count))
1801 vq->shadow_used_idx = i;
1802 if (likely(vq->shadow_used_idx)) {
1803 flush_shadow_used_ring_split(dev, vq);
1804 vhost_vring_call_split(dev, vq);
1805 }
1806 }
1807
1808 return (i - dropped);
1809 }
1810
1811 static __rte_always_inline int
1812 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
1813 struct vhost_virtqueue *vq,
1814 struct rte_mempool *mbuf_pool,
1815 struct rte_mbuf **pkts,
1816 uint16_t avail_idx,
1817 uintptr_t *desc_addrs,
1818 uint16_t *ids)
1819 {
1820 bool wrap = vq->avail_wrap_counter;
1821 struct vring_packed_desc *descs = vq->desc_packed;
1822 struct virtio_net_hdr *hdr;
1823 uint64_t lens[PACKED_BATCH_SIZE];
1824 uint64_t buf_lens[PACKED_BATCH_SIZE];
1825 uint32_t buf_offset = dev->vhost_hlen;
1826 uint16_t flags, i;
1827
1828 if (unlikely(avail_idx & PACKED_BATCH_MASK))
1829 return -1;
1830 if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1831 return -1;
1832
1833 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1834 flags = descs[avail_idx + i].flags;
1835 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
1836 (wrap == !!(flags & VRING_DESC_F_USED)) ||
1837 (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
1838 return -1;
1839 }
1840
1841 rte_smp_rmb();
1842
1843 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1844 lens[i] = descs[avail_idx + i].len;
1845
1846 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1847 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1848 descs[avail_idx + i].addr,
1849 &lens[i], VHOST_ACCESS_RW);
1850 }
1851
1852 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1853 if (unlikely(!desc_addrs[i]))
1854 return -1;
1855 if (unlikely((lens[i] != descs[avail_idx + i].len)))
1856 return -1;
1857 }
1858
1859 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1860 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, lens[i]);
1861 if (!pkts[i])
1862 goto free_buf;
1863 }
1864
1865 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1866 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
1867
1868 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1869 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
1870 goto free_buf;
1871 }
1872
1873 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1874 pkts[i]->pkt_len = descs[avail_idx + i].len - buf_offset;
1875 pkts[i]->data_len = pkts[i]->pkt_len;
1876 ids[i] = descs[avail_idx + i].id;
1877 }
1878
1879 if (virtio_net_with_host_offload(dev)) {
1880 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1881 hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
1882 vhost_dequeue_offload(hdr, pkts[i]);
1883 }
1884 }
1885
1886 return 0;
1887
1888 free_buf:
1889 for (i = 0; i < PACKED_BATCH_SIZE; i++)
1890 rte_pktmbuf_free(pkts[i]);
1891
1892 return -1;
1893 }
1894
1895 static __rte_always_inline int
1896 virtio_dev_tx_batch_packed(struct virtio_net *dev,
1897 struct vhost_virtqueue *vq,
1898 struct rte_mempool *mbuf_pool,
1899 struct rte_mbuf **pkts)
1900 {
1901 uint16_t avail_idx = vq->last_avail_idx;
1902 uint32_t buf_offset = dev->vhost_hlen;
1903 uintptr_t desc_addrs[PACKED_BATCH_SIZE];
1904 uint16_t ids[PACKED_BATCH_SIZE];
1905 uint16_t i;
1906
1907 if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
1908 avail_idx, desc_addrs, ids))
1909 return -1;
1910
1911 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1912 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1913
1914 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1915 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1916 (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1917 pkts[i]->pkt_len);
1918
1919 if (virtio_net_is_inorder(dev))
1920 vhost_shadow_dequeue_batch_packed_inorder(vq,
1921 ids[PACKED_BATCH_SIZE - 1]);
1922 else
1923 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
1924
1925 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1926
1927 return 0;
1928 }
1929
1930 static __rte_always_inline int
1931 vhost_dequeue_single_packed(struct virtio_net *dev,
1932 struct vhost_virtqueue *vq,
1933 struct rte_mempool *mbuf_pool,
1934 struct rte_mbuf **pkts,
1935 uint16_t *buf_id,
1936 uint16_t *desc_count)
1937 {
1938 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1939 uint32_t buf_len;
1940 uint16_t nr_vec = 0;
1941 int err;
1942 static bool allocerr_warned;
1943
1944 if (unlikely(fill_vec_buf_packed(dev, vq,
1945 vq->last_avail_idx, desc_count,
1946 buf_vec, &nr_vec,
1947 buf_id, &buf_len,
1948 VHOST_ACCESS_RO) < 0))
1949 return -1;
1950
1951 *pkts = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
1952 if (unlikely(*pkts == NULL)) {
1953 if (!allocerr_warned) {
1954 VHOST_LOG_DATA(ERR,
1955 "Failed mbuf alloc of size %d from %s on %s.\n",
1956 buf_len, mbuf_pool->name, dev->ifname);
1957 allocerr_warned = true;
1958 }
1959 return -1;
1960 }
1961
1962 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, *pkts,
1963 mbuf_pool);
1964 if (unlikely(err)) {
1965 if (!allocerr_warned) {
1966 VHOST_LOG_DATA(ERR,
1967 "Failed to copy desc to mbuf on %s.\n",
1968 dev->ifname);
1969 allocerr_warned = true;
1970 }
1971 rte_pktmbuf_free(*pkts);
1972 return -1;
1973 }
1974
1975 return 0;
1976 }
1977
1978 static __rte_always_inline int
1979 virtio_dev_tx_single_packed(struct virtio_net *dev,
1980 struct vhost_virtqueue *vq,
1981 struct rte_mempool *mbuf_pool,
1982 struct rte_mbuf **pkts)
1983 {
1984
1985 uint16_t buf_id, desc_count = 0;
1986 int ret;
1987
1988 ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
1989 &desc_count);
1990
1991 if (likely(desc_count > 0)) {
1992 if (virtio_net_is_inorder(dev))
1993 vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
1994 desc_count);
1995 else
1996 vhost_shadow_dequeue_single_packed(vq, buf_id,
1997 desc_count);
1998
1999 vq_inc_last_avail_packed(vq, desc_count);
2000 }
2001
2002 return ret;
2003 }
2004
2005 static __rte_always_inline int
2006 virtio_dev_tx_batch_packed_zmbuf(struct virtio_net *dev,
2007 struct vhost_virtqueue *vq,
2008 struct rte_mempool *mbuf_pool,
2009 struct rte_mbuf **pkts)
2010 {
2011 struct zcopy_mbuf *zmbufs[PACKED_BATCH_SIZE];
2012 uintptr_t desc_addrs[PACKED_BATCH_SIZE];
2013 uint16_t ids[PACKED_BATCH_SIZE];
2014 uint16_t i;
2015
2016 uint16_t avail_idx = vq->last_avail_idx;
2017
2018 if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
2019 avail_idx, desc_addrs, ids))
2020 return -1;
2021
2022 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2023 zmbufs[i] = get_zmbuf(vq);
2024
2025 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2026 if (!zmbufs[i])
2027 goto free_pkt;
2028 }
2029
2030 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2031 zmbufs[i]->mbuf = pkts[i];
2032 zmbufs[i]->desc_idx = ids[i];
2033 zmbufs[i]->desc_count = 1;
2034 }
2035
2036 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2037 rte_mbuf_refcnt_update(pkts[i], 1);
2038
2039 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2040 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbufs[i], next);
2041
2042 vq->nr_zmbuf += PACKED_BATCH_SIZE;
2043 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2044
2045 return 0;
2046
2047 free_pkt:
2048 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2049 rte_pktmbuf_free(pkts[i]);
2050
2051 return -1;
2052 }
2053
2054 static __rte_always_inline int
2055 virtio_dev_tx_single_packed_zmbuf(struct virtio_net *dev,
2056 struct vhost_virtqueue *vq,
2057 struct rte_mempool *mbuf_pool,
2058 struct rte_mbuf **pkts)
2059 {
2060 uint16_t buf_id, desc_count;
2061 struct zcopy_mbuf *zmbuf;
2062
2063 if (vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2064 &desc_count))
2065 return -1;
2066
2067 zmbuf = get_zmbuf(vq);
2068 if (!zmbuf) {
2069 rte_pktmbuf_free(*pkts);
2070 return -1;
2071 }
2072 zmbuf->mbuf = *pkts;
2073 zmbuf->desc_idx = buf_id;
2074 zmbuf->desc_count = desc_count;
2075
2076 rte_mbuf_refcnt_update(*pkts, 1);
2077
2078 vq->nr_zmbuf += 1;
2079 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
2080
2081 vq_inc_last_avail_packed(vq, desc_count);
2082 return 0;
2083 }
2084
2085 static __rte_always_inline void
2086 free_zmbuf(struct vhost_virtqueue *vq)
2087 {
2088 struct zcopy_mbuf *next = NULL;
2089 struct zcopy_mbuf *zmbuf;
2090
2091 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
2092 zmbuf != NULL; zmbuf = next) {
2093 next = TAILQ_NEXT(zmbuf, next);
2094
2095 uint16_t last_used_idx = vq->last_used_idx;
2096
2097 if (mbuf_is_consumed(zmbuf->mbuf)) {
2098 uint16_t flags;
2099 flags = vq->desc_packed[last_used_idx].flags;
2100 if (vq->used_wrap_counter) {
2101 flags |= VRING_DESC_F_USED;
2102 flags |= VRING_DESC_F_AVAIL;
2103 } else {
2104 flags &= ~VRING_DESC_F_USED;
2105 flags &= ~VRING_DESC_F_AVAIL;
2106 }
2107
2108 vq->desc_packed[last_used_idx].id = zmbuf->desc_idx;
2109 vq->desc_packed[last_used_idx].len = 0;
2110
2111 rte_smp_wmb();
2112 vq->desc_packed[last_used_idx].flags = flags;
2113
2114 vq_inc_last_used_packed(vq, zmbuf->desc_count);
2115
2116 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
2117 restore_mbuf(zmbuf->mbuf);
2118 rte_pktmbuf_free(zmbuf->mbuf);
2119 put_zmbuf(zmbuf);
2120 vq->nr_zmbuf -= 1;
2121 }
2122 }
2123 }
2124
2125 static __rte_noinline uint16_t
2126 virtio_dev_tx_packed_zmbuf(struct virtio_net *dev,
2127 struct vhost_virtqueue *vq,
2128 struct rte_mempool *mbuf_pool,
2129 struct rte_mbuf **pkts,
2130 uint32_t count)
2131 {
2132 uint32_t pkt_idx = 0;
2133 uint32_t remained = count;
2134
2135 free_zmbuf(vq);
2136
2137 do {
2138 if (remained >= PACKED_BATCH_SIZE) {
2139 if (!virtio_dev_tx_batch_packed_zmbuf(dev, vq,
2140 mbuf_pool, &pkts[pkt_idx])) {
2141 pkt_idx += PACKED_BATCH_SIZE;
2142 remained -= PACKED_BATCH_SIZE;
2143 continue;
2144 }
2145 }
2146
2147 if (virtio_dev_tx_single_packed_zmbuf(dev, vq, mbuf_pool,
2148 &pkts[pkt_idx]))
2149 break;
2150 pkt_idx++;
2151 remained--;
2152
2153 } while (remained);
2154
2155 if (pkt_idx)
2156 vhost_vring_call_packed(dev, vq);
2157
2158 return pkt_idx;
2159 }
2160
2161 static __rte_noinline uint16_t
2162 virtio_dev_tx_packed(struct virtio_net *dev,
2163 struct vhost_virtqueue *vq,
2164 struct rte_mempool *mbuf_pool,
2165 struct rte_mbuf **pkts,
2166 uint32_t count)
2167 {
2168 uint32_t pkt_idx = 0;
2169 uint32_t remained = count;
2170
2171 do {
2172 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
2173
2174 if (remained >= PACKED_BATCH_SIZE) {
2175 if (!virtio_dev_tx_batch_packed(dev, vq, mbuf_pool,
2176 &pkts[pkt_idx])) {
2177 pkt_idx += PACKED_BATCH_SIZE;
2178 remained -= PACKED_BATCH_SIZE;
2179 continue;
2180 }
2181 }
2182
2183 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
2184 &pkts[pkt_idx]))
2185 break;
2186 pkt_idx++;
2187 remained--;
2188
2189 } while (remained);
2190
2191 if (vq->shadow_used_idx) {
2192 do_data_copy_dequeue(vq);
2193
2194 vhost_flush_dequeue_shadow_packed(dev, vq);
2195 vhost_vring_call_packed(dev, vq);
2196 }
2197
2198 return pkt_idx;
2199 }
2200
2201 uint16_t
2202 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
2203 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2204 {
2205 struct virtio_net *dev;
2206 struct rte_mbuf *rarp_mbuf = NULL;
2207 struct vhost_virtqueue *vq;
2208 int16_t success = 1;
2209
2210 dev = get_device(vid);
2211 if (!dev)
2212 return 0;
2213
2214 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2215 VHOST_LOG_DATA(ERR,
2216 "(%d) %s: built-in vhost net backend is disabled.\n",
2217 dev->vid, __func__);
2218 return 0;
2219 }
2220
2221 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
2222 VHOST_LOG_DATA(ERR,
2223 "(%d) %s: invalid virtqueue idx %d.\n",
2224 dev->vid, __func__, queue_id);
2225 return 0;
2226 }
2227
2228 vq = dev->virtqueue[queue_id];
2229
2230 if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
2231 return 0;
2232
2233 if (unlikely(vq->enabled == 0)) {
2234 count = 0;
2235 goto out_access_unlock;
2236 }
2237
2238 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2239 vhost_user_iotlb_rd_lock(vq);
2240
2241 if (unlikely(vq->access_ok == 0))
2242 if (unlikely(vring_translate(dev, vq) < 0)) {
2243 count = 0;
2244 goto out;
2245 }
2246
2247 /*
2248 * Construct a RARP broadcast packet, and inject it to the "pkts"
2249 * array, to looks like that guest actually send such packet.
2250 *
2251 * Check user_send_rarp() for more information.
2252 *
2253 * broadcast_rarp shares a cacheline in the virtio_net structure
2254 * with some fields that are accessed during enqueue and
2255 * __atomic_compare_exchange_n causes a write if performed compare
2256 * and exchange. This could result in false sharing between enqueue
2257 * and dequeue.
2258 *
2259 * Prevent unnecessary false sharing by reading broadcast_rarp first
2260 * and only performing compare and exchange if the read indicates it
2261 * is likely to be set.
2262 */
2263 if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
2264 __atomic_compare_exchange_n(&dev->broadcast_rarp,
2265 &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
2266
2267 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
2268 if (rarp_mbuf == NULL) {
2269 VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
2270 count = 0;
2271 goto out;
2272 }
2273 count -= 1;
2274 }
2275
2276 if (vq_is_packed(dev)) {
2277 if (unlikely(dev->dequeue_zero_copy))
2278 count = virtio_dev_tx_packed_zmbuf(dev, vq, mbuf_pool,
2279 pkts, count);
2280 else
2281 count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts,
2282 count);
2283 } else
2284 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
2285
2286 out:
2287 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2288 vhost_user_iotlb_rd_unlock(vq);
2289
2290 out_access_unlock:
2291 rte_spinlock_unlock(&vq->access_lock);
2292
2293 if (unlikely(rarp_mbuf != NULL)) {
2294 /*
2295 * Inject it to the head of "pkts" array, so that switch's mac
2296 * learning table will get updated first.
2297 */
2298 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
2299 pkts[0] = rarp_mbuf;
2300 count += 1;
2301 }
2302
2303 return count;
2304 }
Ceph