]> git.proxmox.com Git - ceph.git/blob - ceph/src/dpdk/lib/librte_vhost/virtio_net.c
projects / ceph.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
add subtree-ish sources for 12.0.3
[ceph.git] / ceph / src / dpdk / lib / librte_vhost / virtio_net.c
1 /*-
2 * BSD LICENSE
3 *
4 * Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 *
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
16 * distribution.
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.
20 *
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.
32 */
33
34 #include <stdint.h>
35 #include <stdbool.h>
36 #include <linux/virtio_net.h>
37
38 #include <rte_mbuf.h>
39 #include <rte_memcpy.h>
40 #include <rte_ether.h>
41 #include <rte_ip.h>
42 #include <rte_virtio_net.h>
43 #include <rte_tcp.h>
44 #include <rte_udp.h>
45 #include <rte_sctp.h>
46 #include <rte_arp.h>
47
48 #include "vhost.h"
49
50 #define MAX_PKT_BURST 32
51 #define VHOST_LOG_PAGE 4096
52
53 static inline void __attribute__((always_inline))
54 vhost_log_page(uint8_t *log_base, uint64_t page)
55 {
56 log_base[page / 8] |= 1 << (page % 8);
57 }
58
59 static inline void __attribute__((always_inline))
60 vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
61 {
62 uint64_t page;
63
64 if (likely(((dev->features & (1ULL << VHOST_F_LOG_ALL)) == 0) ||
65 !dev->log_base || !len))
66 return;
67
68 if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
69 return;
70
71 /* To make sure guest memory updates are committed before logging */
72 rte_smp_wmb();
73
74 page = addr / VHOST_LOG_PAGE;
75 while (page * VHOST_LOG_PAGE < addr + len) {
76 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
77 page += 1;
78 }
79 }
80
81 static inline void __attribute__((always_inline))
82 vhost_log_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,
83 uint64_t offset, uint64_t len)
84 {
85 vhost_log_write(dev, vq->log_guest_addr + offset, len);
86 }
87
88 static bool
89 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t qp_nb)
90 {
91 return (is_tx ^ (idx & 1)) == 0 && idx < qp_nb * VIRTIO_QNUM;
92 }
93
94 static inline void __attribute__((always_inline))
95 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
96 uint16_t to, uint16_t from, uint16_t size)
97 {
98 rte_memcpy(&vq->used->ring[to],
99 &vq->shadow_used_ring[from],
100 size * sizeof(struct vring_used_elem));
101 vhost_log_used_vring(dev, vq,
102 offsetof(struct vring_used, ring[to]),
103 size * sizeof(struct vring_used_elem));
104 }
105
106 static inline void __attribute__((always_inline))
107 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
108 {
109 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
110
111 if (used_idx + vq->shadow_used_idx <= vq->size) {
112 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
113 vq->shadow_used_idx);
114 } else {
115 uint16_t size;
116
117 /* update used ring interval [used_idx, vq->size] */
118 size = vq->size - used_idx;
119 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
120
121 /* update the left half used ring interval [0, left_size] */
122 do_flush_shadow_used_ring(dev, vq, 0, size,
123 vq->shadow_used_idx - size);
124 }
125 vq->last_used_idx += vq->shadow_used_idx;
126
127 rte_smp_wmb();
128
129 *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
130 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
131 sizeof(vq->used->idx));
132 }
133
134 static inline void __attribute__((always_inline))
135 update_shadow_used_ring(struct vhost_virtqueue *vq,
136 uint16_t desc_idx, uint16_t len)
137 {
138 uint16_t i = vq->shadow_used_idx++;
139
140 vq->shadow_used_ring[i].id = desc_idx;
141 vq->shadow_used_ring[i].len = len;
142 }
143
144 static void
145 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
146 {
147 if (m_buf->ol_flags & PKT_TX_L4_MASK) {
148 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
149 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
150
151 switch (m_buf->ol_flags & PKT_TX_L4_MASK) {
152 case PKT_TX_TCP_CKSUM:
153 net_hdr->csum_offset = (offsetof(struct tcp_hdr,
154 cksum));
155 break;
156 case PKT_TX_UDP_CKSUM:
157 net_hdr->csum_offset = (offsetof(struct udp_hdr,
158 dgram_cksum));
159 break;
160 case PKT_TX_SCTP_CKSUM:
161 net_hdr->csum_offset = (offsetof(struct sctp_hdr,
162 cksum));
163 break;
164 }
165 }
166
167 if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
168 if (m_buf->ol_flags & PKT_TX_IPV4)
169 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
170 else
171 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
172 net_hdr->gso_size = m_buf->tso_segsz;
173 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
174 + m_buf->l4_len;
175 }
176 }
177
178 static inline void
179 copy_virtio_net_hdr(struct virtio_net *dev, uint64_t desc_addr,
180 struct virtio_net_hdr_mrg_rxbuf hdr)
181 {
182 if (dev->vhost_hlen == sizeof(struct virtio_net_hdr_mrg_rxbuf))
183 *(struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr = hdr;
184 else
185 *(struct virtio_net_hdr *)(uintptr_t)desc_addr = hdr.hdr;
186 }
187
188 static inline int __attribute__((always_inline))
189 copy_mbuf_to_desc(struct virtio_net *dev, struct vring_desc *descs,
190 struct rte_mbuf *m, uint16_t desc_idx, uint32_t size)
191 {
192 uint32_t desc_avail, desc_offset;
193 uint32_t mbuf_avail, mbuf_offset;
194 uint32_t cpy_len;
195 struct vring_desc *desc;
196 uint64_t desc_addr;
197 struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
198
199 desc = &descs[desc_idx];
200 desc_addr = gpa_to_vva(dev, desc->addr);
201 /*
202 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
203 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
204 * otherwise stores offset on the stack instead of in a register.
205 */
206 if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr)
207 return -1;
208
209 rte_prefetch0((void *)(uintptr_t)desc_addr);
210
211 virtio_enqueue_offload(m, &virtio_hdr.hdr);
212 copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
213 vhost_log_write(dev, desc->addr, dev->vhost_hlen);
214 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
215
216 desc_offset = dev->vhost_hlen;
217 desc_avail = desc->len - dev->vhost_hlen;
218
219 mbuf_avail = rte_pktmbuf_data_len(m);
220 mbuf_offset = 0;
221 while (mbuf_avail != 0 || m->next != NULL) {
222 /* done with current mbuf, fetch next */
223 if (mbuf_avail == 0) {
224 m = m->next;
225
226 mbuf_offset = 0;
227 mbuf_avail = rte_pktmbuf_data_len(m);
228 }
229
230 /* done with current desc buf, fetch next */
231 if (desc_avail == 0) {
232 if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
233 /* Room in vring buffer is not enough */
234 return -1;
235 }
236 if (unlikely(desc->next >= size))
237 return -1;
238
239 desc = &descs[desc->next];
240 desc_addr = gpa_to_vva(dev, desc->addr);
241 if (unlikely(!desc_addr))
242 return -1;
243
244 desc_offset = 0;
245 desc_avail = desc->len;
246 }
247
248 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
249 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
250 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
251 cpy_len);
252 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
253 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
254 cpy_len, 0);
255
256 mbuf_avail -= cpy_len;
257 mbuf_offset += cpy_len;
258 desc_avail -= cpy_len;
259 desc_offset += cpy_len;
260 }
261
262 return 0;
263 }
264
265 /**
266 * This function adds buffers to the virtio devices RX virtqueue. Buffers can
267 * be received from the physical port or from another virtio device. A packet
268 * count is returned to indicate the number of packets that are succesfully
269 * added to the RX queue. This function works when the mbuf is scattered, but
270 * it doesn't support the mergeable feature.
271 */
272 static inline uint32_t __attribute__((always_inline))
273 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
274 struct rte_mbuf **pkts, uint32_t count)
275 {
276 struct vhost_virtqueue *vq;
277 uint16_t avail_idx, free_entries, start_idx;
278 uint16_t desc_indexes[MAX_PKT_BURST];
279 struct vring_desc *descs;
280 uint16_t used_idx;
281 uint32_t i, sz;
282
283 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
284 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
285 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
286 dev->vid, __func__, queue_id);
287 return 0;
288 }
289
290 vq = dev->virtqueue[queue_id];
291 if (unlikely(vq->enabled == 0))
292 return 0;
293
294 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
295 start_idx = vq->last_used_idx;
296 free_entries = avail_idx - start_idx;
297 count = RTE_MIN(count, free_entries);
298 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
299 if (count == 0)
300 return 0;
301
302 LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
303 dev->vid, start_idx, start_idx + count);
304
305 /* Retrieve all of the desc indexes first to avoid caching issues. */
306 rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
307 for (i = 0; i < count; i++) {
308 used_idx = (start_idx + i) & (vq->size - 1);
309 desc_indexes[i] = vq->avail->ring[used_idx];
310 vq->used->ring[used_idx].id = desc_indexes[i];
311 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
312 dev->vhost_hlen;
313 vhost_log_used_vring(dev, vq,
314 offsetof(struct vring_used, ring[used_idx]),
315 sizeof(vq->used->ring[used_idx]));
316 }
317
318 rte_prefetch0(&vq->desc[desc_indexes[0]]);
319 for (i = 0; i < count; i++) {
320 uint16_t desc_idx = desc_indexes[i];
321 int err;
322
323 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
324 descs = (struct vring_desc *)(uintptr_t)gpa_to_vva(dev,
325 vq->desc[desc_idx].addr);
326 if (unlikely(!descs)) {
327 count = i;
328 break;
329 }
330
331 desc_idx = 0;
332 sz = vq->desc[desc_idx].len / sizeof(*descs);
333 } else {
334 descs = vq->desc;
335 sz = vq->size;
336 }
337
338 err = copy_mbuf_to_desc(dev, descs, pkts[i], desc_idx, sz);
339 if (unlikely(err)) {
340 used_idx = (start_idx + i) & (vq->size - 1);
341 vq->used->ring[used_idx].len = dev->vhost_hlen;
342 vhost_log_used_vring(dev, vq,
343 offsetof(struct vring_used, ring[used_idx]),
344 sizeof(vq->used->ring[used_idx]));
345 }
346
347 if (i + 1 < count)
348 rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
349 }
350
351 rte_smp_wmb();
352
353 *(volatile uint16_t *)&vq->used->idx += count;
354 vq->last_used_idx += count;
355 vhost_log_used_vring(dev, vq,
356 offsetof(struct vring_used, idx),
357 sizeof(vq->used->idx));
358
359 /* flush used->idx update before we read avail->flags. */
360 rte_mb();
361
362 /* Kick the guest if necessary. */
363 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
364 && (vq->callfd >= 0))
365 eventfd_write(vq->callfd, (eventfd_t)1);
366 return count;
367 }
368
369 static inline int __attribute__((always_inline))
370 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
371 uint32_t avail_idx, uint32_t *vec_idx,
372 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
373 uint16_t *desc_chain_len)
374 {
375 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
376 uint32_t vec_id = *vec_idx;
377 uint32_t len = 0;
378 struct vring_desc *descs = vq->desc;
379
380 *desc_chain_head = idx;
381
382 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
383 descs = (struct vring_desc *)(uintptr_t)
384 gpa_to_vva(dev, vq->desc[idx].addr);
385 if (unlikely(!descs))
386 return -1;
387
388 idx = 0;
389 }
390
391 while (1) {
392 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
393 return -1;
394
395 len += descs[idx].len;
396 buf_vec[vec_id].buf_addr = descs[idx].addr;
397 buf_vec[vec_id].buf_len = descs[idx].len;
398 buf_vec[vec_id].desc_idx = idx;
399 vec_id++;
400
401 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
402 break;
403
404 idx = descs[idx].next;
405 }
406
407 *desc_chain_len = len;
408 *vec_idx = vec_id;
409
410 return 0;
411 }
412
413 /*
414 * Returns -1 on fail, 0 on success
415 */
416 static inline int
417 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
418 uint32_t size, struct buf_vector *buf_vec,
419 uint16_t *num_buffers, uint16_t avail_head)
420 {
421 uint16_t cur_idx;
422 uint32_t vec_idx = 0;
423 uint16_t tries = 0;
424
425 uint16_t head_idx = 0;
426 uint16_t len = 0;
427
428 *num_buffers = 0;
429 cur_idx = vq->last_avail_idx;
430
431 while (size > 0) {
432 if (unlikely(cur_idx == avail_head))
433 return -1;
434
435 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
436 &head_idx, &len) < 0))
437 return -1;
438 len = RTE_MIN(len, size);
439 update_shadow_used_ring(vq, head_idx, len);
440 size -= len;
441
442 cur_idx++;
443 tries++;
444 *num_buffers += 1;
445
446 /*
447 * if we tried all available ring items, and still
448 * can't get enough buf, it means something abnormal
449 * happened.
450 */
451 if (unlikely(tries >= vq->size))
452 return -1;
453 }
454
455 return 0;
456 }
457
458 static inline int __attribute__((always_inline))
459 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct rte_mbuf *m,
460 struct buf_vector *buf_vec, uint16_t num_buffers)
461 {
462 struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
463 uint32_t vec_idx = 0;
464 uint64_t desc_addr;
465 uint32_t mbuf_offset, mbuf_avail;
466 uint32_t desc_offset, desc_avail;
467 uint32_t cpy_len;
468 uint64_t hdr_addr, hdr_phys_addr;
469 struct rte_mbuf *hdr_mbuf;
470
471 if (unlikely(m == NULL))
472 return -1;
473
474 desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
475 if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
476 return -1;
477
478 hdr_mbuf = m;
479 hdr_addr = desc_addr;
480 hdr_phys_addr = buf_vec[vec_idx].buf_addr;
481 rte_prefetch0((void *)(uintptr_t)hdr_addr);
482
483 virtio_hdr.num_buffers = num_buffers;
484 LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
485 dev->vid, num_buffers);
486
487 desc_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
488 desc_offset = dev->vhost_hlen;
489
490 mbuf_avail = rte_pktmbuf_data_len(m);
491 mbuf_offset = 0;
492 while (mbuf_avail != 0 || m->next != NULL) {
493 /* done with current desc buf, get the next one */
494 if (desc_avail == 0) {
495 vec_idx++;
496 desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
497 if (unlikely(!desc_addr))
498 return -1;
499
500 /* Prefetch buffer address. */
501 rte_prefetch0((void *)(uintptr_t)desc_addr);
502 desc_offset = 0;
503 desc_avail = buf_vec[vec_idx].buf_len;
504 }
505
506 /* done with current mbuf, get the next one */
507 if (mbuf_avail == 0) {
508 m = m->next;
509
510 mbuf_offset = 0;
511 mbuf_avail = rte_pktmbuf_data_len(m);
512 }
513
514 if (hdr_addr) {
515 virtio_enqueue_offload(hdr_mbuf, &virtio_hdr.hdr);
516 copy_virtio_net_hdr(dev, hdr_addr, virtio_hdr);
517 vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
518 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
519 dev->vhost_hlen, 0);
520
521 hdr_addr = 0;
522 }
523
524 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
525 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
526 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
527 cpy_len);
528 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
529 cpy_len);
530 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
531 cpy_len, 0);
532
533 mbuf_avail -= cpy_len;
534 mbuf_offset += cpy_len;
535 desc_avail -= cpy_len;
536 desc_offset += cpy_len;
537 }
538
539 return 0;
540 }
541
542 static inline uint32_t __attribute__((always_inline))
543 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
544 struct rte_mbuf **pkts, uint32_t count)
545 {
546 struct vhost_virtqueue *vq;
547 uint32_t pkt_idx = 0;
548 uint16_t num_buffers;
549 struct buf_vector buf_vec[BUF_VECTOR_MAX];
550 uint16_t avail_head;
551
552 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
553 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
554 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
555 dev->vid, __func__, queue_id);
556 return 0;
557 }
558
559 vq = dev->virtqueue[queue_id];
560 if (unlikely(vq->enabled == 0))
561 return 0;
562
563 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
564 if (count == 0)
565 return 0;
566
567 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
568
569 vq->shadow_used_idx = 0;
570 avail_head = *((volatile uint16_t *)&vq->avail->idx);
571 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
572 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
573
574 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
575 pkt_len, buf_vec, &num_buffers,
576 avail_head) < 0)) {
577 LOG_DEBUG(VHOST_DATA,
578 "(%d) failed to get enough desc from vring\n",
579 dev->vid);
580 vq->shadow_used_idx -= num_buffers;
581 break;
582 }
583
584 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
585 dev->vid, vq->last_avail_idx,
586 vq->last_avail_idx + num_buffers);
587
588 if (copy_mbuf_to_desc_mergeable(dev, pkts[pkt_idx],
589 buf_vec, num_buffers) < 0) {
590 vq->shadow_used_idx -= num_buffers;
591 break;
592 }
593
594 vq->last_avail_idx += num_buffers;
595 }
596
597 if (likely(vq->shadow_used_idx)) {
598 flush_shadow_used_ring(dev, vq);
599
600 /* flush used->idx update before we read avail->flags. */
601 rte_mb();
602
603 /* Kick the guest if necessary. */
604 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
605 && (vq->callfd >= 0))
606 eventfd_write(vq->callfd, (eventfd_t)1);
607 }
608
609 return pkt_idx;
610 }
611
612 uint16_t
613 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
614 struct rte_mbuf **pkts, uint16_t count)
615 {
616 struct virtio_net *dev = get_device(vid);
617
618 if (!dev)
619 return 0;
620
621 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
622 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
623 else
624 return virtio_dev_rx(dev, queue_id, pkts, count);
625 }
626
627 static inline bool
628 virtio_net_with_host_offload(struct virtio_net *dev)
629 {
630 if (dev->features &
631 (VIRTIO_NET_F_CSUM | VIRTIO_NET_F_HOST_ECN |
632 VIRTIO_NET_F_HOST_TSO4 | VIRTIO_NET_F_HOST_TSO6 |
633 VIRTIO_NET_F_HOST_UFO))
634 return true;
635
636 return false;
637 }
638
639 static void
640 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
641 {
642 struct ipv4_hdr *ipv4_hdr;
643 struct ipv6_hdr *ipv6_hdr;
644 void *l3_hdr = NULL;
645 struct ether_hdr *eth_hdr;
646 uint16_t ethertype;
647
648 eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
649
650 m->l2_len = sizeof(struct ether_hdr);
651 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
652
653 if (ethertype == ETHER_TYPE_VLAN) {
654 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
655
656 m->l2_len += sizeof(struct vlan_hdr);
657 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
658 }
659
660 l3_hdr = (char *)eth_hdr + m->l2_len;
661
662 switch (ethertype) {
663 case ETHER_TYPE_IPv4:
664 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
665 *l4_proto = ipv4_hdr->next_proto_id;
666 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
667 *l4_hdr = (char *)l3_hdr + m->l3_len;
668 m->ol_flags |= PKT_TX_IPV4;
669 break;
670 case ETHER_TYPE_IPv6:
671 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
672 *l4_proto = ipv6_hdr->proto;
673 m->l3_len = sizeof(struct ipv6_hdr);
674 *l4_hdr = (char *)l3_hdr + m->l3_len;
675 m->ol_flags |= PKT_TX_IPV6;
676 break;
677 default:
678 m->l3_len = 0;
679 *l4_proto = 0;
680 break;
681 }
682 }
683
684 static inline void __attribute__((always_inline))
685 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
686 {
687 uint16_t l4_proto = 0;
688 void *l4_hdr = NULL;
689 struct tcp_hdr *tcp_hdr = NULL;
690
691 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
692 return;
693
694 parse_ethernet(m, &l4_proto, &l4_hdr);
695 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
696 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
697 switch (hdr->csum_offset) {
698 case (offsetof(struct tcp_hdr, cksum)):
699 if (l4_proto == IPPROTO_TCP)
700 m->ol_flags |= PKT_TX_TCP_CKSUM;
701 break;
702 case (offsetof(struct udp_hdr, dgram_cksum)):
703 if (l4_proto == IPPROTO_UDP)
704 m->ol_flags |= PKT_TX_UDP_CKSUM;
705 break;
706 case (offsetof(struct sctp_hdr, cksum)):
707 if (l4_proto == IPPROTO_SCTP)
708 m->ol_flags |= PKT_TX_SCTP_CKSUM;
709 break;
710 default:
711 break;
712 }
713 }
714 }
715
716 if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
717 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
718 case VIRTIO_NET_HDR_GSO_TCPV4:
719 case VIRTIO_NET_HDR_GSO_TCPV6:
720 tcp_hdr = (struct tcp_hdr *)l4_hdr;
721 m->ol_flags |= PKT_TX_TCP_SEG;
722 m->tso_segsz = hdr->gso_size;
723 m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
724 break;
725 default:
726 RTE_LOG(WARNING, VHOST_DATA,
727 "unsupported gso type %u.\n", hdr->gso_type);
728 break;
729 }
730 }
731 }
732
733 #define RARP_PKT_SIZE 64
734
735 static int
736 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
737 {
738 struct ether_hdr *eth_hdr;
739 struct arp_hdr *rarp;
740
741 if (rarp_mbuf->buf_len < 64) {
742 RTE_LOG(WARNING, VHOST_DATA,
743 "failed to make RARP; mbuf size too small %u (< %d)\n",
744 rarp_mbuf->buf_len, RARP_PKT_SIZE);
745 return -1;
746 }
747
748 /* Ethernet header. */
749 eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
750 memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
751 ether_addr_copy(mac, &eth_hdr->s_addr);
752 eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
753
754 /* RARP header. */
755 rarp = (struct arp_hdr *)(eth_hdr + 1);
756 rarp->arp_hrd = htons(ARP_HRD_ETHER);
757 rarp->arp_pro = htons(ETHER_TYPE_IPv4);
758 rarp->arp_hln = ETHER_ADDR_LEN;
759 rarp->arp_pln = 4;
760 rarp->arp_op = htons(ARP_OP_REVREQUEST);
761
762 ether_addr_copy(mac, &rarp->arp_data.arp_sha);
763 ether_addr_copy(mac, &rarp->arp_data.arp_tha);
764 memset(&rarp->arp_data.arp_sip, 0x00, 4);
765 memset(&rarp->arp_data.arp_tip, 0x00, 4);
766
767 rarp_mbuf->pkt_len = rarp_mbuf->data_len = RARP_PKT_SIZE;
768
769 return 0;
770 }
771
772 static inline void __attribute__((always_inline))
773 put_zmbuf(struct zcopy_mbuf *zmbuf)
774 {
775 zmbuf->in_use = 0;
776 }
777
778 static inline int __attribute__((always_inline))
779 copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
780 uint16_t max_desc, struct rte_mbuf *m, uint16_t desc_idx,
781 struct rte_mempool *mbuf_pool)
782 {
783 struct vring_desc *desc;
784 uint64_t desc_addr;
785 uint32_t desc_avail, desc_offset;
786 uint32_t mbuf_avail, mbuf_offset;
787 uint32_t cpy_len;
788 struct rte_mbuf *cur = m, *prev = m;
789 struct virtio_net_hdr *hdr = NULL;
790 /* A counter to avoid desc dead loop chain */
791 uint32_t nr_desc = 1;
792
793 desc = &descs[desc_idx];
794 if (unlikely((desc->len < dev->vhost_hlen)) ||
795 (desc->flags & VRING_DESC_F_INDIRECT))
796 return -1;
797
798 desc_addr = gpa_to_vva(dev, desc->addr);
799 if (unlikely(!desc_addr))
800 return -1;
801
802 if (virtio_net_with_host_offload(dev)) {
803 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
804 rte_prefetch0(hdr);
805 }
806
807 /*
808 * A virtio driver normally uses at least 2 desc buffers
809 * for Tx: the first for storing the header, and others
810 * for storing the data.
811 */
812 if (likely((desc->len == dev->vhost_hlen) &&
813 (desc->flags & VRING_DESC_F_NEXT) != 0)) {
814 desc = &descs[desc->next];
815 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
816 return -1;
817
818 desc_addr = gpa_to_vva(dev, desc->addr);
819 if (unlikely(!desc_addr))
820 return -1;
821
822 desc_offset = 0;
823 desc_avail = desc->len;
824 nr_desc += 1;
825 } else {
826 desc_avail = desc->len - dev->vhost_hlen;
827 desc_offset = dev->vhost_hlen;
828 }
829
830 rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
831
832 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
833
834 mbuf_offset = 0;
835 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
836 while (1) {
837 uint64_t hpa;
838
839 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
840
841 /*
842 * A desc buf might across two host physical pages that are
843 * not continuous. In such case (gpa_to_hpa returns 0), data
844 * will be copied even though zero copy is enabled.
845 */
846 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
847 desc->addr + desc_offset, cpy_len)))) {
848 cur->data_len = cpy_len;
849 cur->data_off = 0;
850 cur->buf_addr = (void *)(uintptr_t)desc_addr;
851 cur->buf_physaddr = hpa;
852
853 /*
854 * In zero copy mode, one mbuf can only reference data
855 * for one or partial of one desc buff.
856 */
857 mbuf_avail = cpy_len;
858 } else {
859 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
860 mbuf_offset),
861 (void *)((uintptr_t)(desc_addr + desc_offset)),
862 cpy_len);
863 }
864
865 mbuf_avail -= cpy_len;
866 mbuf_offset += cpy_len;
867 desc_avail -= cpy_len;
868 desc_offset += cpy_len;
869
870 /* This desc reaches to its end, get the next one */
871 if (desc_avail == 0) {
872 if ((desc->flags & VRING_DESC_F_NEXT) == 0)
873 break;
874
875 if (unlikely(desc->next >= max_desc ||
876 ++nr_desc > max_desc))
877 return -1;
878 desc = &descs[desc->next];
879 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
880 return -1;
881
882 desc_addr = gpa_to_vva(dev, desc->addr);
883 if (unlikely(!desc_addr))
884 return -1;
885
886 rte_prefetch0((void *)(uintptr_t)desc_addr);
887
888 desc_offset = 0;
889 desc_avail = desc->len;
890
891 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
892 }
893
894 /*
895 * This mbuf reaches to its end, get a new one
896 * to hold more data.
897 */
898 if (mbuf_avail == 0) {
899 cur = rte_pktmbuf_alloc(mbuf_pool);
900 if (unlikely(cur == NULL)) {
901 RTE_LOG(ERR, VHOST_DATA, "Failed to "
902 "allocate memory for mbuf.\n");
903 return -1;
904 }
905
906 prev->next = cur;
907 prev->data_len = mbuf_offset;
908 m->nb_segs += 1;
909 m->pkt_len += mbuf_offset;
910 prev = cur;
911
912 mbuf_offset = 0;
913 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
914 }
915 }
916
917 prev->data_len = mbuf_offset;
918 m->pkt_len += mbuf_offset;
919
920 if (hdr)
921 vhost_dequeue_offload(hdr, m);
922
923 return 0;
924 }
925
926 static inline void __attribute__((always_inline))
927 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
928 uint32_t used_idx, uint32_t desc_idx)
929 {
930 vq->used->ring[used_idx].id = desc_idx;
931 vq->used->ring[used_idx].len = 0;
932 vhost_log_used_vring(dev, vq,
933 offsetof(struct vring_used, ring[used_idx]),
934 sizeof(vq->used->ring[used_idx]));
935 }
936
937 static inline void __attribute__((always_inline))
938 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
939 uint32_t count)
940 {
941 if (unlikely(count == 0))
942 return;
943
944 rte_smp_wmb();
945 rte_smp_rmb();
946
947 vq->used->idx += count;
948 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
949 sizeof(vq->used->idx));
950
951 /* Kick guest if required. */
952 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
953 && (vq->callfd >= 0))
954 eventfd_write(vq->callfd, (eventfd_t)1);
955 }
956
957 static inline struct zcopy_mbuf *__attribute__((always_inline))
958 get_zmbuf(struct vhost_virtqueue *vq)
959 {
960 uint16_t i;
961 uint16_t last;
962 int tries = 0;
963
964 /* search [last_zmbuf_idx, zmbuf_size) */
965 i = vq->last_zmbuf_idx;
966 last = vq->zmbuf_size;
967
968 again:
969 for (; i < last; i++) {
970 if (vq->zmbufs[i].in_use == 0) {
971 vq->last_zmbuf_idx = i + 1;
972 vq->zmbufs[i].in_use = 1;
973 return &vq->zmbufs[i];
974 }
975 }
976
977 tries++;
978 if (tries == 1) {
979 /* search [0, last_zmbuf_idx) */
980 i = 0;
981 last = vq->last_zmbuf_idx;
982 goto again;
983 }
984
985 return NULL;
986 }
987
988 static inline bool __attribute__((always_inline))
989 mbuf_is_consumed(struct rte_mbuf *m)
990 {
991 while (m) {
992 if (rte_mbuf_refcnt_read(m) > 1)
993 return false;
994 m = m->next;
995 }
996
997 return true;
998 }
999
1000 uint16_t
1001 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1002 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1003 {
1004 struct virtio_net *dev;
1005 struct rte_mbuf *rarp_mbuf = NULL;
1006 struct vhost_virtqueue *vq;
1007 uint32_t desc_indexes[MAX_PKT_BURST];
1008 uint32_t used_idx;
1009 uint32_t i = 0;
1010 uint16_t free_entries;
1011 uint16_t avail_idx;
1012
1013 dev = get_device(vid);
1014 if (!dev)
1015 return 0;
1016
1017 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
1018 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1019 dev->vid, __func__, queue_id);
1020 return 0;
1021 }
1022
1023 vq = dev->virtqueue[queue_id];
1024 if (unlikely(vq->enabled == 0))
1025 return 0;
1026
1027 if (unlikely(dev->dequeue_zero_copy)) {
1028 struct zcopy_mbuf *zmbuf, *next;
1029 int nr_updated = 0;
1030
1031 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1032 zmbuf != NULL; zmbuf = next) {
1033 next = TAILQ_NEXT(zmbuf, next);
1034
1035 if (mbuf_is_consumed(zmbuf->mbuf)) {
1036 used_idx = vq->last_used_idx++ & (vq->size - 1);
1037 update_used_ring(dev, vq, used_idx,
1038 zmbuf->desc_idx);
1039 nr_updated += 1;
1040
1041 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1042 rte_pktmbuf_free(zmbuf->mbuf);
1043 put_zmbuf(zmbuf);
1044 vq->nr_zmbuf -= 1;
1045 }
1046 }
1047
1048 update_used_idx(dev, vq, nr_updated);
1049 }
1050
1051 /*
1052 * Construct a RARP broadcast packet, and inject it to the "pkts"
1053 * array, to looks like that guest actually send such packet.
1054 *
1055 * Check user_send_rarp() for more information.
1056 */
1057 if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
1058 &dev->broadcast_rarp.cnt, 1, 0))) {
1059 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1060 if (rarp_mbuf == NULL) {
1061 RTE_LOG(ERR, VHOST_DATA,
1062 "Failed to allocate memory for mbuf.\n");
1063 return 0;
1064 }
1065
1066 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1067 rte_pktmbuf_free(rarp_mbuf);
1068 rarp_mbuf = NULL;
1069 } else {
1070 count -= 1;
1071 }
1072 }
1073
1074 free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1075 vq->last_avail_idx;
1076 if (free_entries == 0)
1077 goto out;
1078
1079 LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1080
1081 /* Prefetch available and used ring */
1082 avail_idx = vq->last_avail_idx & (vq->size - 1);
1083 used_idx = vq->last_used_idx & (vq->size - 1);
1084 rte_prefetch0(&vq->avail->ring[avail_idx]);
1085 rte_prefetch0(&vq->used->ring[used_idx]);
1086
1087 count = RTE_MIN(count, MAX_PKT_BURST);
1088 count = RTE_MIN(count, free_entries);
1089 LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1090 dev->vid, count);
1091
1092 /* Retrieve all of the head indexes first to avoid caching issues. */
1093 for (i = 0; i < count; i++) {
1094 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1095 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
1096 desc_indexes[i] = vq->avail->ring[avail_idx];
1097
1098 if (likely(dev->dequeue_zero_copy == 0))
1099 update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1100 }
1101
1102 /* Prefetch descriptor index. */
1103 rte_prefetch0(&vq->desc[desc_indexes[0]]);
1104 for (i = 0; i < count; i++) {
1105 struct vring_desc *desc;
1106 uint16_t sz, idx;
1107 int err;
1108
1109 if (likely(i + 1 < count))
1110 rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1111
1112 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1113 desc = (struct vring_desc *)(uintptr_t)gpa_to_vva(dev,
1114 vq->desc[desc_indexes[i]].addr);
1115 if (unlikely(!desc))
1116 break;
1117
1118 rte_prefetch0(desc);
1119 sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1120 idx = 0;
1121 } else {
1122 desc = vq->desc;
1123 sz = vq->size;
1124 idx = desc_indexes[i];
1125 }
1126
1127 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1128 if (unlikely(pkts[i] == NULL)) {
1129 RTE_LOG(ERR, VHOST_DATA,
1130 "Failed to allocate memory for mbuf.\n");
1131 break;
1132 }
1133
1134 err = copy_desc_to_mbuf(dev, desc, sz, pkts[i], idx, mbuf_pool);
1135 if (unlikely(err)) {
1136 rte_pktmbuf_free(pkts[i]);
1137 break;
1138 }
1139
1140 if (unlikely(dev->dequeue_zero_copy)) {
1141 struct zcopy_mbuf *zmbuf;
1142
1143 zmbuf = get_zmbuf(vq);
1144 if (!zmbuf) {
1145 rte_pktmbuf_free(pkts[i]);
1146 break;
1147 }
1148 zmbuf->mbuf = pkts[i];
1149 zmbuf->desc_idx = desc_indexes[i];
1150
1151 /*
1152 * Pin lock the mbuf; we will check later to see
1153 * whether the mbuf is freed (when we are the last
1154 * user) or not. If that's the case, we then could
1155 * update the used ring safely.
1156 */
1157 rte_mbuf_refcnt_update(pkts[i], 1);
1158
1159 vq->nr_zmbuf += 1;
1160 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1161 }
1162 }
1163 vq->last_avail_idx += i;
1164
1165 if (likely(dev->dequeue_zero_copy == 0)) {
1166 vq->last_used_idx += i;
1167 update_used_idx(dev, vq, i);
1168 }
1169
1170 out:
1171 if (unlikely(rarp_mbuf != NULL)) {
1172 /*
1173 * Inject it to the head of "pkts" array, so that switch's mac
1174 * learning table will get updated first.
1175 */
1176 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1177 pkts[0] = rarp_mbuf;
1178 i += 1;
1179 }
1180
1181 return i;
1182 }
Ceph