virtio: restrict vector Rx/Tx to x86 SSSE3
[dpdk.git] / drivers / net / virtio / virtio_rxtx.c
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2010-2014 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 <stdio.h>
36 #include <stdlib.h>
37 #include <string.h>
38 #include <errno.h>
39
40 #include <rte_cycles.h>
41 #include <rte_memory.h>
42 #include <rte_memzone.h>
43 #include <rte_branch_prediction.h>
44 #include <rte_mempool.h>
45 #include <rte_malloc.h>
46 #include <rte_mbuf.h>
47 #include <rte_ether.h>
48 #include <rte_ethdev.h>
49 #include <rte_prefetch.h>
50 #include <rte_string_fns.h>
51 #include <rte_errno.h>
52 #include <rte_byteorder.h>
53
54 #include "virtio_logs.h"
55 #include "virtio_ethdev.h"
56 #include "virtio_pci.h"
57 #include "virtqueue.h"
58 #include "virtio_rxtx.h"
59
60 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
61 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
62 #else
63 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
64 #endif
65
66
67 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
68         ETH_TXQ_FLAGS_NOOFFLOADS)
69
70 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
71 static int use_simple_rxtx;
72 #endif
73
74 static void
75 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
76 {
77         struct vring_desc *dp, *dp_tail;
78         struct vq_desc_extra *dxp;
79         uint16_t desc_idx_last = desc_idx;
80
81         dp  = &vq->vq_ring.desc[desc_idx];
82         dxp = &vq->vq_descx[desc_idx];
83         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
84         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
85                 while (dp->flags & VRING_DESC_F_NEXT) {
86                         desc_idx_last = dp->next;
87                         dp = &vq->vq_ring.desc[dp->next];
88                 }
89         }
90         dxp->ndescs = 0;
91
92         /*
93          * We must append the existing free chain, if any, to the end of
94          * newly freed chain. If the virtqueue was completely used, then
95          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
96          */
97         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
98                 vq->vq_desc_head_idx = desc_idx;
99         } else {
100                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
101                 dp_tail->next = desc_idx;
102         }
103
104         vq->vq_desc_tail_idx = desc_idx_last;
105         dp->next = VQ_RING_DESC_CHAIN_END;
106 }
107
108 static uint16_t
109 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
110                            uint32_t *len, uint16_t num)
111 {
112         struct vring_used_elem *uep;
113         struct rte_mbuf *cookie;
114         uint16_t used_idx, desc_idx;
115         uint16_t i;
116
117         /*  Caller does the check */
118         for (i = 0; i < num ; i++) {
119                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
120                 uep = &vq->vq_ring.used->ring[used_idx];
121                 desc_idx = (uint16_t) uep->id;
122                 len[i] = uep->len;
123                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
124
125                 if (unlikely(cookie == NULL)) {
126                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u\n",
127                                 vq->vq_used_cons_idx);
128                         break;
129                 }
130
131                 rte_prefetch0(cookie);
132                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
133                 rx_pkts[i]  = cookie;
134                 vq->vq_used_cons_idx++;
135                 vq_ring_free_chain(vq, desc_idx);
136                 vq->vq_descx[desc_idx].cookie = NULL;
137         }
138
139         return i;
140 }
141
142 #ifndef DEFAULT_TX_FREE_THRESH
143 #define DEFAULT_TX_FREE_THRESH 32
144 #endif
145
146 /* Cleanup from completed transmits. */
147 static void
148 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
149 {
150         uint16_t i, used_idx, desc_idx;
151         for (i = 0; i < num; i++) {
152                 struct vring_used_elem *uep;
153                 struct vq_desc_extra *dxp;
154
155                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
156                 uep = &vq->vq_ring.used->ring[used_idx];
157
158                 desc_idx = (uint16_t) uep->id;
159                 dxp = &vq->vq_descx[desc_idx];
160                 vq->vq_used_cons_idx++;
161                 vq_ring_free_chain(vq, desc_idx);
162
163                 if (dxp->cookie != NULL) {
164                         rte_pktmbuf_free(dxp->cookie);
165                         dxp->cookie = NULL;
166                 }
167         }
168 }
169
170
171 static inline int
172 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
173 {
174         struct vq_desc_extra *dxp;
175         struct virtio_hw *hw = vq->hw;
176         struct vring_desc *start_dp;
177         uint16_t needed = 1;
178         uint16_t head_idx, idx;
179
180         if (unlikely(vq->vq_free_cnt == 0))
181                 return -ENOSPC;
182         if (unlikely(vq->vq_free_cnt < needed))
183                 return -EMSGSIZE;
184
185         head_idx = vq->vq_desc_head_idx;
186         if (unlikely(head_idx >= vq->vq_nentries))
187                 return -EFAULT;
188
189         idx = head_idx;
190         dxp = &vq->vq_descx[idx];
191         dxp->cookie = (void *)cookie;
192         dxp->ndescs = needed;
193
194         start_dp = vq->vq_ring.desc;
195         start_dp[idx].addr =
196                 (uint64_t)(cookie->buf_physaddr + RTE_PKTMBUF_HEADROOM
197                 - hw->vtnet_hdr_size);
198         start_dp[idx].len =
199                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
200         start_dp[idx].flags =  VRING_DESC_F_WRITE;
201         idx = start_dp[idx].next;
202         vq->vq_desc_head_idx = idx;
203         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
204                 vq->vq_desc_tail_idx = idx;
205         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
206         vq_update_avail_ring(vq, head_idx);
207
208         return 0;
209 }
210
211 static int
212 virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie)
213 {
214         struct vq_desc_extra *dxp;
215         struct vring_desc *start_dp;
216         uint16_t seg_num = cookie->nb_segs;
217         uint16_t needed = 1 + seg_num;
218         uint16_t head_idx, idx;
219         size_t head_size = txvq->hw->vtnet_hdr_size;
220
221         if (unlikely(txvq->vq_free_cnt == 0))
222                 return -ENOSPC;
223         if (unlikely(txvq->vq_free_cnt < needed))
224                 return -EMSGSIZE;
225         head_idx = txvq->vq_desc_head_idx;
226         if (unlikely(head_idx >= txvq->vq_nentries))
227                 return -EFAULT;
228
229         idx = head_idx;
230         dxp = &txvq->vq_descx[idx];
231         dxp->cookie = (void *)cookie;
232         dxp->ndescs = needed;
233
234         start_dp = txvq->vq_ring.desc;
235         start_dp[idx].addr =
236                 txvq->virtio_net_hdr_mem + idx * head_size;
237         start_dp[idx].len = head_size;
238         start_dp[idx].flags = VRING_DESC_F_NEXT;
239
240         for (; ((seg_num > 0) && (cookie != NULL)); seg_num--) {
241                 idx = start_dp[idx].next;
242                 start_dp[idx].addr  = RTE_MBUF_DATA_DMA_ADDR(cookie);
243                 start_dp[idx].len   = cookie->data_len;
244                 start_dp[idx].flags = VRING_DESC_F_NEXT;
245                 cookie = cookie->next;
246         }
247
248         start_dp[idx].flags &= ~VRING_DESC_F_NEXT;
249         idx = start_dp[idx].next;
250         txvq->vq_desc_head_idx = idx;
251         if (txvq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
252                 txvq->vq_desc_tail_idx = idx;
253         txvq->vq_free_cnt = (uint16_t)(txvq->vq_free_cnt - needed);
254         vq_update_avail_ring(txvq, head_idx);
255
256         return 0;
257 }
258
259 static inline struct rte_mbuf *
260 rte_rxmbuf_alloc(struct rte_mempool *mp)
261 {
262         struct rte_mbuf *m;
263
264         m = __rte_mbuf_raw_alloc(mp);
265         __rte_mbuf_sanity_check_raw(m, 0);
266
267         return m;
268 }
269
270 static void
271 virtio_dev_vring_start(struct virtqueue *vq, int queue_type)
272 {
273         struct rte_mbuf *m;
274         int i, nbufs, error, size = vq->vq_nentries;
275         struct vring *vr = &vq->vq_ring;
276         uint8_t *ring_mem = vq->vq_ring_virt_mem;
277
278         PMD_INIT_FUNC_TRACE();
279
280         /*
281          * Reinitialise since virtio port might have been stopped and restarted
282          */
283         memset(vq->vq_ring_virt_mem, 0, vq->vq_ring_size);
284         vring_init(vr, size, ring_mem, VIRTIO_PCI_VRING_ALIGN);
285         vq->vq_used_cons_idx = 0;
286         vq->vq_desc_head_idx = 0;
287         vq->vq_avail_idx = 0;
288         vq->vq_desc_tail_idx = (uint16_t)(vq->vq_nentries - 1);
289         vq->vq_free_cnt = vq->vq_nentries;
290         memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
291
292         /* Chain all the descriptors in the ring with an END */
293         for (i = 0; i < size - 1; i++)
294                 vr->desc[i].next = (uint16_t)(i + 1);
295         vr->desc[i].next = VQ_RING_DESC_CHAIN_END;
296
297         /*
298          * Disable device(host) interrupting guest
299          */
300         virtqueue_disable_intr(vq);
301
302         /* Only rx virtqueue needs mbufs to be allocated at initialization */
303         if (queue_type == VTNET_RQ) {
304                 if (vq->mpool == NULL)
305                         rte_exit(EXIT_FAILURE,
306                         "Cannot allocate initial mbufs for rx virtqueue");
307
308                 /* Allocate blank mbufs for the each rx descriptor */
309                 nbufs = 0;
310                 error = ENOSPC;
311
312 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
313                 if (use_simple_rxtx)
314                         for (i = 0; i < vq->vq_nentries; i++) {
315                                 vq->vq_ring.avail->ring[i] = i;
316                                 vq->vq_ring.desc[i].flags = VRING_DESC_F_WRITE;
317                         }
318 #endif
319                 memset(&vq->fake_mbuf, 0, sizeof(vq->fake_mbuf));
320                 for (i = 0; i < RTE_PMD_VIRTIO_RX_MAX_BURST; i++)
321                         vq->sw_ring[vq->vq_nentries + i] = &vq->fake_mbuf;
322
323                 while (!virtqueue_full(vq)) {
324                         m = rte_rxmbuf_alloc(vq->mpool);
325                         if (m == NULL)
326                                 break;
327
328                         /******************************************
329                         *         Enqueue allocated buffers        *
330                         *******************************************/
331 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
332                         if (use_simple_rxtx)
333                                 error = virtqueue_enqueue_recv_refill_simple(vq, m);
334                         else
335 #endif
336                                 error = virtqueue_enqueue_recv_refill(vq, m);
337                         if (error) {
338                                 rte_pktmbuf_free(m);
339                                 break;
340                         }
341                         nbufs++;
342                 }
343
344                 vq_update_avail_idx(vq);
345
346                 PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
347         } else if (queue_type == VTNET_TQ) {
348 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
349                 if (use_simple_rxtx) {
350                         int mid_idx  = vq->vq_nentries >> 1;
351                         for (i = 0; i < mid_idx; i++) {
352                                 vq->vq_ring.avail->ring[i] = i + mid_idx;
353                                 vq->vq_ring.desc[i + mid_idx].next = i;
354                                 vq->vq_ring.desc[i + mid_idx].addr =
355                                         vq->virtio_net_hdr_mem +
356                                                 mid_idx * vq->hw->vtnet_hdr_size;
357                                 vq->vq_ring.desc[i + mid_idx].len =
358                                         vq->hw->vtnet_hdr_size;
359                                 vq->vq_ring.desc[i + mid_idx].flags =
360                                         VRING_DESC_F_NEXT;
361                                 vq->vq_ring.desc[i].flags = 0;
362                         }
363                         for (i = mid_idx; i < vq->vq_nentries; i++)
364                                 vq->vq_ring.avail->ring[i] = i;
365                 }
366 #endif
367         }
368 }
369
370 void
371 virtio_dev_cq_start(struct rte_eth_dev *dev)
372 {
373         struct virtio_hw *hw = dev->data->dev_private;
374
375         if (hw->cvq) {
376                 virtio_dev_vring_start(hw->cvq, VTNET_CQ);
377                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq);
378         }
379 }
380
381 void
382 virtio_dev_rxtx_start(struct rte_eth_dev *dev)
383 {
384         /*
385          * Start receive and transmit vrings
386          * -    Setup vring structure for all queues
387          * -    Initialize descriptor for the rx vring
388          * -    Allocate blank mbufs for the each rx descriptor
389          *
390          */
391         int i;
392
393         PMD_INIT_FUNC_TRACE();
394
395         /* Start rx vring. */
396         for (i = 0; i < dev->data->nb_rx_queues; i++) {
397                 virtio_dev_vring_start(dev->data->rx_queues[i], VTNET_RQ);
398                 VIRTQUEUE_DUMP((struct virtqueue *)dev->data->rx_queues[i]);
399         }
400
401         /* Start tx vring. */
402         for (i = 0; i < dev->data->nb_tx_queues; i++) {
403                 virtio_dev_vring_start(dev->data->tx_queues[i], VTNET_TQ);
404                 VIRTQUEUE_DUMP((struct virtqueue *)dev->data->tx_queues[i]);
405         }
406 }
407
408 int
409 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
410                         uint16_t queue_idx,
411                         uint16_t nb_desc,
412                         unsigned int socket_id,
413                         __rte_unused const struct rte_eth_rxconf *rx_conf,
414                         struct rte_mempool *mp)
415 {
416         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
417         struct virtqueue *vq;
418         int ret;
419
420         PMD_INIT_FUNC_TRACE();
421         ret = virtio_dev_queue_setup(dev, VTNET_RQ, queue_idx, vtpci_queue_idx,
422                         nb_desc, socket_id, &vq);
423         if (ret < 0) {
424                 PMD_INIT_LOG(ERR, "rvq initialization failed");
425                 return ret;
426         }
427
428         /* Create mempool for rx mbuf allocation */
429         vq->mpool = mp;
430
431         dev->data->rx_queues[queue_idx] = vq;
432
433 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
434         virtio_rxq_vec_setup(vq);
435 #endif
436
437         return 0;
438 }
439
440 void
441 virtio_dev_rx_queue_release(void *rxq)
442 {
443         virtio_dev_queue_release(rxq);
444 }
445
446 /*
447  * struct rte_eth_dev *dev: Used to update dev
448  * uint16_t nb_desc: Defaults to values read from config space
449  * unsigned int socket_id: Used to allocate memzone
450  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
451  * uint16_t queue_idx: Just used as an index in dev txq list
452  */
453 int
454 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
455                         uint16_t queue_idx,
456                         uint16_t nb_desc,
457                         unsigned int socket_id,
458                         const struct rte_eth_txconf *tx_conf)
459 {
460         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
461
462 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
463         struct virtio_hw *hw = dev->data->dev_private;
464 #endif
465         struct virtqueue *vq;
466         uint16_t tx_free_thresh;
467         int ret;
468
469         PMD_INIT_FUNC_TRACE();
470
471         if ((tx_conf->txq_flags & ETH_TXQ_FLAGS_NOXSUMS)
472             != ETH_TXQ_FLAGS_NOXSUMS) {
473                 PMD_INIT_LOG(ERR, "TX checksum offload not supported\n");
474                 return -EINVAL;
475         }
476
477 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
478         /* Use simple rx/tx func if single segment and no offloads */
479         if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
480              !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
481                 PMD_INIT_LOG(INFO, "Using simple rx/tx path");
482                 dev->tx_pkt_burst = virtio_xmit_pkts_simple;
483                 dev->rx_pkt_burst = virtio_recv_pkts_vec;
484                 use_simple_rxtx = 1;
485         }
486 #endif
487
488         ret = virtio_dev_queue_setup(dev, VTNET_TQ, queue_idx, vtpci_queue_idx,
489                         nb_desc, socket_id, &vq);
490         if (ret < 0) {
491                 PMD_INIT_LOG(ERR, "rvq initialization failed");
492                 return ret;
493         }
494
495         tx_free_thresh = tx_conf->tx_free_thresh;
496         if (tx_free_thresh == 0)
497                 tx_free_thresh =
498                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
499
500         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
501                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
502                         "number of TX entries minus 3 (%u)."
503                         " (tx_free_thresh=%u port=%u queue=%u)\n",
504                         vq->vq_nentries - 3,
505                         tx_free_thresh, dev->data->port_id, queue_idx);
506                 return -EINVAL;
507         }
508
509         vq->vq_free_thresh = tx_free_thresh;
510
511         dev->data->tx_queues[queue_idx] = vq;
512         return 0;
513 }
514
515 void
516 virtio_dev_tx_queue_release(void *txq)
517 {
518         virtio_dev_queue_release(txq);
519 }
520
521 static void
522 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
523 {
524         int error;
525         /*
526          * Requeue the discarded mbuf. This should always be
527          * successful since it was just dequeued.
528          */
529         error = virtqueue_enqueue_recv_refill(vq, m);
530         if (unlikely(error)) {
531                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
532                 rte_pktmbuf_free(m);
533         }
534 }
535
536 static void
537 virtio_update_packet_stats(struct virtqueue *vq, struct rte_mbuf *mbuf)
538 {
539         uint32_t s = mbuf->pkt_len;
540         struct ether_addr *ea;
541
542         if (s == 64) {
543                 vq->size_bins[1]++;
544         } else if (s > 64 && s < 1024) {
545                 uint32_t bin;
546
547                 /* count zeros, and offset into correct bin */
548                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
549                 vq->size_bins[bin]++;
550         } else {
551                 if (s < 64)
552                         vq->size_bins[0]++;
553                 else if (s < 1519)
554                         vq->size_bins[6]++;
555                 else if (s >= 1519)
556                         vq->size_bins[7]++;
557         }
558
559         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
560         vq->multicast += is_multicast_ether_addr(ea);
561         vq->broadcast += is_broadcast_ether_addr(ea);
562 }
563
564 #define VIRTIO_MBUF_BURST_SZ 64
565 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
566 uint16_t
567 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
568 {
569         struct virtqueue *rxvq = rx_queue;
570         struct virtio_hw *hw;
571         struct rte_mbuf *rxm, *new_mbuf;
572         uint16_t nb_used, num, nb_rx;
573         uint32_t len[VIRTIO_MBUF_BURST_SZ];
574         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
575         int error;
576         uint32_t i, nb_enqueued;
577         uint32_t hdr_size;
578
579         nb_used = VIRTQUEUE_NUSED(rxvq);
580
581         virtio_rmb();
582
583         num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
584         num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
585         if (likely(num > DESC_PER_CACHELINE))
586                 num = num - ((rxvq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
587
588         if (num == 0)
589                 return 0;
590
591         num = virtqueue_dequeue_burst_rx(rxvq, rcv_pkts, len, num);
592         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
593
594         hw = rxvq->hw;
595         nb_rx = 0;
596         nb_enqueued = 0;
597         hdr_size = hw->vtnet_hdr_size;
598
599         for (i = 0; i < num ; i++) {
600                 rxm = rcv_pkts[i];
601
602                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
603
604                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
605                         PMD_RX_LOG(ERR, "Packet drop");
606                         nb_enqueued++;
607                         virtio_discard_rxbuf(rxvq, rxm);
608                         rxvq->errors++;
609                         continue;
610                 }
611
612                 rxm->port = rxvq->port_id;
613                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
614                 rxm->ol_flags = 0;
615                 rxm->vlan_tci = 0;
616
617                 rxm->nb_segs = 1;
618                 rxm->next = NULL;
619                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
620                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
621
622                 if (hw->vlan_strip)
623                         rte_vlan_strip(rxm);
624
625                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
626
627                 rx_pkts[nb_rx++] = rxm;
628
629                 rxvq->bytes += rx_pkts[nb_rx - 1]->pkt_len;
630                 virtio_update_packet_stats(rxvq, rxm);
631         }
632
633         rxvq->packets += nb_rx;
634
635         /* Allocate new mbuf for the used descriptor */
636         error = ENOSPC;
637         while (likely(!virtqueue_full(rxvq))) {
638                 new_mbuf = rte_rxmbuf_alloc(rxvq->mpool);
639                 if (unlikely(new_mbuf == NULL)) {
640                         struct rte_eth_dev *dev
641                                 = &rte_eth_devices[rxvq->port_id];
642                         dev->data->rx_mbuf_alloc_failed++;
643                         break;
644                 }
645                 error = virtqueue_enqueue_recv_refill(rxvq, new_mbuf);
646                 if (unlikely(error)) {
647                         rte_pktmbuf_free(new_mbuf);
648                         break;
649                 }
650                 nb_enqueued++;
651         }
652
653         if (likely(nb_enqueued)) {
654                 vq_update_avail_idx(rxvq);
655
656                 if (unlikely(virtqueue_kick_prepare(rxvq))) {
657                         virtqueue_notify(rxvq);
658                         PMD_RX_LOG(DEBUG, "Notified\n");
659                 }
660         }
661
662         return nb_rx;
663 }
664
665 uint16_t
666 virtio_recv_mergeable_pkts(void *rx_queue,
667                         struct rte_mbuf **rx_pkts,
668                         uint16_t nb_pkts)
669 {
670         struct virtqueue *rxvq = rx_queue;
671         struct virtio_hw *hw;
672         struct rte_mbuf *rxm, *new_mbuf;
673         uint16_t nb_used, num, nb_rx;
674         uint32_t len[VIRTIO_MBUF_BURST_SZ];
675         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
676         struct rte_mbuf *prev;
677         int error;
678         uint32_t i, nb_enqueued;
679         uint32_t seg_num;
680         uint16_t extra_idx;
681         uint32_t seg_res;
682         uint32_t hdr_size;
683
684         nb_used = VIRTQUEUE_NUSED(rxvq);
685
686         virtio_rmb();
687
688         if (nb_used == 0)
689                 return 0;
690
691         PMD_RX_LOG(DEBUG, "used:%d\n", nb_used);
692
693         hw = rxvq->hw;
694         nb_rx = 0;
695         i = 0;
696         nb_enqueued = 0;
697         seg_num = 0;
698         extra_idx = 0;
699         seg_res = 0;
700         hdr_size = hw->vtnet_hdr_size;
701
702         while (i < nb_used) {
703                 struct virtio_net_hdr_mrg_rxbuf *header;
704
705                 if (nb_rx == nb_pkts)
706                         break;
707
708                 num = virtqueue_dequeue_burst_rx(rxvq, rcv_pkts, len, 1);
709                 if (num != 1)
710                         continue;
711
712                 i++;
713
714                 PMD_RX_LOG(DEBUG, "dequeue:%d\n", num);
715                 PMD_RX_LOG(DEBUG, "packet len:%d\n", len[0]);
716
717                 rxm = rcv_pkts[0];
718
719                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
720                         PMD_RX_LOG(ERR, "Packet drop\n");
721                         nb_enqueued++;
722                         virtio_discard_rxbuf(rxvq, rxm);
723                         rxvq->errors++;
724                         continue;
725                 }
726
727                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
728                         RTE_PKTMBUF_HEADROOM - hdr_size);
729                 seg_num = header->num_buffers;
730
731                 if (seg_num == 0)
732                         seg_num = 1;
733
734                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
735                 rxm->nb_segs = seg_num;
736                 rxm->next = NULL;
737                 rxm->ol_flags = 0;
738                 rxm->vlan_tci = 0;
739                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
740                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
741
742                 rxm->port = rxvq->port_id;
743                 rx_pkts[nb_rx] = rxm;
744                 prev = rxm;
745
746                 seg_res = seg_num - 1;
747
748                 while (seg_res != 0) {
749                         /*
750                          * Get extra segments for current uncompleted packet.
751                          */
752                         uint16_t  rcv_cnt =
753                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
754                         if (likely(VIRTQUEUE_NUSED(rxvq) >= rcv_cnt)) {
755                                 uint32_t rx_num =
756                                         virtqueue_dequeue_burst_rx(rxvq,
757                                         rcv_pkts, len, rcv_cnt);
758                                 i += rx_num;
759                                 rcv_cnt = rx_num;
760                         } else {
761                                 PMD_RX_LOG(ERR,
762                                         "No enough segments for packet.\n");
763                                 nb_enqueued++;
764                                 virtio_discard_rxbuf(rxvq, rxm);
765                                 rxvq->errors++;
766                                 break;
767                         }
768
769                         extra_idx = 0;
770
771                         while (extra_idx < rcv_cnt) {
772                                 rxm = rcv_pkts[extra_idx];
773
774                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
775                                 rxm->next = NULL;
776                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
777                                 rxm->data_len = (uint16_t)(len[extra_idx]);
778
779                                 if (prev)
780                                         prev->next = rxm;
781
782                                 prev = rxm;
783                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
784                                 extra_idx++;
785                         };
786                         seg_res -= rcv_cnt;
787                 }
788
789                 if (hw->vlan_strip)
790                         rte_vlan_strip(rx_pkts[nb_rx]);
791
792                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
793                         rx_pkts[nb_rx]->data_len);
794
795                 rxvq->bytes += rx_pkts[nb_rx]->pkt_len;
796                 virtio_update_packet_stats(rxvq, rx_pkts[nb_rx]);
797                 nb_rx++;
798         }
799
800         rxvq->packets += nb_rx;
801
802         /* Allocate new mbuf for the used descriptor */
803         error = ENOSPC;
804         while (likely(!virtqueue_full(rxvq))) {
805                 new_mbuf = rte_rxmbuf_alloc(rxvq->mpool);
806                 if (unlikely(new_mbuf == NULL)) {
807                         struct rte_eth_dev *dev
808                                 = &rte_eth_devices[rxvq->port_id];
809                         dev->data->rx_mbuf_alloc_failed++;
810                         break;
811                 }
812                 error = virtqueue_enqueue_recv_refill(rxvq, new_mbuf);
813                 if (unlikely(error)) {
814                         rte_pktmbuf_free(new_mbuf);
815                         break;
816                 }
817                 nb_enqueued++;
818         }
819
820         if (likely(nb_enqueued)) {
821                 vq_update_avail_idx(rxvq);
822
823                 if (unlikely(virtqueue_kick_prepare(rxvq))) {
824                         virtqueue_notify(rxvq);
825                         PMD_RX_LOG(DEBUG, "Notified");
826                 }
827         }
828
829         return nb_rx;
830 }
831
832 uint16_t
833 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
834 {
835         struct virtqueue *txvq = tx_queue;
836         uint16_t nb_used, nb_tx;
837         int error;
838
839         if (unlikely(nb_pkts < 1))
840                 return nb_pkts;
841
842         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
843         nb_used = VIRTQUEUE_NUSED(txvq);
844
845         virtio_rmb();
846         if (likely(nb_used > txvq->vq_nentries - txvq->vq_free_thresh))
847                 virtio_xmit_cleanup(txvq, nb_used);
848
849         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
850                 struct rte_mbuf *txm = tx_pkts[nb_tx];
851                 /* Need one more descriptor for virtio header. */
852                 int need = txm->nb_segs - txvq->vq_free_cnt + 1;
853
854                 /* Positive value indicates it need free vring descriptors */
855                 if (unlikely(need > 0)) {
856                         nb_used = VIRTQUEUE_NUSED(txvq);
857                         virtio_rmb();
858                         need = RTE_MIN(need, (int)nb_used);
859
860                         virtio_xmit_cleanup(txvq, need);
861                         need = txm->nb_segs - txvq->vq_free_cnt + 1;
862                         if (unlikely(need > 0)) {
863                                 PMD_TX_LOG(ERR,
864                                            "No free tx descriptors to transmit");
865                                 break;
866                         }
867                 }
868
869                 /* Do VLAN tag insertion */
870                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
871                         error = rte_vlan_insert(&txm);
872                         if (unlikely(error)) {
873                                 rte_pktmbuf_free(txm);
874                                 continue;
875                         }
876                 }
877
878                 /* Enqueue Packet buffers */
879                 error = virtqueue_enqueue_xmit(txvq, txm);
880                 if (unlikely(error)) {
881                         if (error == ENOSPC)
882                                 PMD_TX_LOG(ERR, "virtqueue_enqueue Free count = 0");
883                         else if (error == EMSGSIZE)
884                                 PMD_TX_LOG(ERR, "virtqueue_enqueue Free count < 1");
885                         else
886                                 PMD_TX_LOG(ERR, "virtqueue_enqueue error: %d", error);
887                         break;
888                 }
889
890                 txvq->bytes += txm->pkt_len;
891                 virtio_update_packet_stats(txvq, txm);
892         }
893
894         txvq->packets += nb_tx;
895
896         if (likely(nb_tx)) {
897                 vq_update_avail_idx(txvq);
898
899                 if (unlikely(virtqueue_kick_prepare(txvq))) {
900                         virtqueue_notify(txvq);
901                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
902                 }
903         }
904
905         return nb_tx;
906 }