a32e3229f868243afcc6936b8fc56367daa8d4a9
[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 #include <rte_cpuflags.h>
54 #include <rte_net.h>
55 #include <rte_ip.h>
56 #include <rte_udp.h>
57 #include <rte_tcp.h>
58
59 #include "virtio_logs.h"
60 #include "virtio_ethdev.h"
61 #include "virtio_pci.h"
62 #include "virtqueue.h"
63 #include "virtio_rxtx.h"
64
65 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
66 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
67 #else
68 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
69 #endif
70
71
72 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
73         ETH_TXQ_FLAGS_NOOFFLOADS)
74
75 int
76 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
77 {
78         struct virtnet_rx *rxvq = rxq;
79         struct virtqueue *vq = rxvq->vq;
80
81         return VIRTQUEUE_NUSED(vq) >= offset;
82 }
83
84 static void
85 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
86 {
87         struct vring_desc *dp, *dp_tail;
88         struct vq_desc_extra *dxp;
89         uint16_t desc_idx_last = desc_idx;
90
91         dp  = &vq->vq_ring.desc[desc_idx];
92         dxp = &vq->vq_descx[desc_idx];
93         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
94         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
95                 while (dp->flags & VRING_DESC_F_NEXT) {
96                         desc_idx_last = dp->next;
97                         dp = &vq->vq_ring.desc[dp->next];
98                 }
99         }
100         dxp->ndescs = 0;
101
102         /*
103          * We must append the existing free chain, if any, to the end of
104          * newly freed chain. If the virtqueue was completely used, then
105          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
106          */
107         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
108                 vq->vq_desc_head_idx = desc_idx;
109         } else {
110                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
111                 dp_tail->next = desc_idx;
112         }
113
114         vq->vq_desc_tail_idx = desc_idx_last;
115         dp->next = VQ_RING_DESC_CHAIN_END;
116 }
117
118 static uint16_t
119 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
120                            uint32_t *len, uint16_t num)
121 {
122         struct vring_used_elem *uep;
123         struct rte_mbuf *cookie;
124         uint16_t used_idx, desc_idx;
125         uint16_t i;
126
127         /*  Caller does the check */
128         for (i = 0; i < num ; i++) {
129                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
130                 uep = &vq->vq_ring.used->ring[used_idx];
131                 desc_idx = (uint16_t) uep->id;
132                 len[i] = uep->len;
133                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
134
135                 if (unlikely(cookie == NULL)) {
136                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
137                                 vq->vq_used_cons_idx);
138                         break;
139                 }
140
141                 rte_prefetch0(cookie);
142                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
143                 rx_pkts[i]  = cookie;
144                 vq->vq_used_cons_idx++;
145                 vq_ring_free_chain(vq, desc_idx);
146                 vq->vq_descx[desc_idx].cookie = NULL;
147         }
148
149         return i;
150 }
151
152 #ifndef DEFAULT_TX_FREE_THRESH
153 #define DEFAULT_TX_FREE_THRESH 32
154 #endif
155
156 /* Cleanup from completed transmits. */
157 static void
158 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
159 {
160         uint16_t i, used_idx, desc_idx;
161         for (i = 0; i < num; i++) {
162                 struct vring_used_elem *uep;
163                 struct vq_desc_extra *dxp;
164
165                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
166                 uep = &vq->vq_ring.used->ring[used_idx];
167
168                 desc_idx = (uint16_t) uep->id;
169                 dxp = &vq->vq_descx[desc_idx];
170                 vq->vq_used_cons_idx++;
171                 vq_ring_free_chain(vq, desc_idx);
172
173                 if (dxp->cookie != NULL) {
174                         rte_pktmbuf_free(dxp->cookie);
175                         dxp->cookie = NULL;
176                 }
177         }
178 }
179
180
181 static inline int
182 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
183 {
184         struct vq_desc_extra *dxp;
185         struct virtio_hw *hw = vq->hw;
186         struct vring_desc *start_dp;
187         uint16_t needed = 1;
188         uint16_t head_idx, idx;
189
190         if (unlikely(vq->vq_free_cnt == 0))
191                 return -ENOSPC;
192         if (unlikely(vq->vq_free_cnt < needed))
193                 return -EMSGSIZE;
194
195         head_idx = vq->vq_desc_head_idx;
196         if (unlikely(head_idx >= vq->vq_nentries))
197                 return -EFAULT;
198
199         idx = head_idx;
200         dxp = &vq->vq_descx[idx];
201         dxp->cookie = (void *)cookie;
202         dxp->ndescs = needed;
203
204         start_dp = vq->vq_ring.desc;
205         start_dp[idx].addr =
206                 VIRTIO_MBUF_ADDR(cookie, vq) +
207                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
208         start_dp[idx].len =
209                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
210         start_dp[idx].flags =  VRING_DESC_F_WRITE;
211         idx = start_dp[idx].next;
212         vq->vq_desc_head_idx = idx;
213         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
214                 vq->vq_desc_tail_idx = idx;
215         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
216         vq_update_avail_ring(vq, head_idx);
217
218         return 0;
219 }
220
221 /* When doing TSO, the IP length is not included in the pseudo header
222  * checksum of the packet given to the PMD, but for virtio it is
223  * expected.
224  */
225 static void
226 virtio_tso_fix_cksum(struct rte_mbuf *m)
227 {
228         /* common case: header is not fragmented */
229         if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
230                         m->l4_len)) {
231                 struct ipv4_hdr *iph;
232                 struct ipv6_hdr *ip6h;
233                 struct tcp_hdr *th;
234                 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
235                 uint32_t tmp;
236
237                 iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
238                 th = RTE_PTR_ADD(iph, m->l3_len);
239                 if ((iph->version_ihl >> 4) == 4) {
240                         iph->hdr_checksum = 0;
241                         iph->hdr_checksum = rte_ipv4_cksum(iph);
242                         ip_len = iph->total_length;
243                         ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
244                                 m->l3_len);
245                 } else {
246                         ip6h = (struct ipv6_hdr *)iph;
247                         ip_paylen = ip6h->payload_len;
248                 }
249
250                 /* calculate the new phdr checksum not including ip_paylen */
251                 prev_cksum = th->cksum;
252                 tmp = prev_cksum;
253                 tmp += ip_paylen;
254                 tmp = (tmp & 0xffff) + (tmp >> 16);
255                 new_cksum = tmp;
256
257                 /* replace it in the packet */
258                 th->cksum = new_cksum;
259         }
260 }
261
262 static inline int
263 tx_offload_enabled(struct virtio_hw *hw)
264 {
265         return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM) ||
266                 vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO4) ||
267                 vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO6);
268 }
269
270 /* avoid write operation when necessary, to lessen cache issues */
271 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
272         if ((var) != (val))                     \
273                 (var) = (val);                  \
274 } while (0)
275
276 static inline void
277 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
278                        uint16_t needed, int use_indirect, int can_push)
279 {
280         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
281         struct vq_desc_extra *dxp;
282         struct virtqueue *vq = txvq->vq;
283         struct vring_desc *start_dp;
284         uint16_t seg_num = cookie->nb_segs;
285         uint16_t head_idx, idx;
286         uint16_t head_size = vq->hw->vtnet_hdr_size;
287         struct virtio_net_hdr *hdr;
288         int offload;
289
290         offload = tx_offload_enabled(vq->hw);
291         head_idx = vq->vq_desc_head_idx;
292         idx = head_idx;
293         dxp = &vq->vq_descx[idx];
294         dxp->cookie = (void *)cookie;
295         dxp->ndescs = needed;
296
297         start_dp = vq->vq_ring.desc;
298
299         if (can_push) {
300                 /* prepend cannot fail, checked by caller */
301                 hdr = (struct virtio_net_hdr *)
302                         rte_pktmbuf_prepend(cookie, head_size);
303                 /* if offload disabled, it is not zeroed below, do it now */
304                 if (offload == 0) {
305                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
306                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
307                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
308                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
309                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
310                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
311                 }
312         } else if (use_indirect) {
313                 /* setup tx ring slot to point to indirect
314                  * descriptor list stored in reserved region.
315                  *
316                  * the first slot in indirect ring is already preset
317                  * to point to the header in reserved region
318                  */
319                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
320                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
321                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
322                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
323                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
324
325                 /* loop below will fill in rest of the indirect elements */
326                 start_dp = txr[idx].tx_indir;
327                 idx = 1;
328         } else {
329                 /* setup first tx ring slot to point to header
330                  * stored in reserved region.
331                  */
332                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
333                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
334                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
335                 start_dp[idx].flags = VRING_DESC_F_NEXT;
336                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
337
338                 idx = start_dp[idx].next;
339         }
340
341         /* Checksum Offload / TSO */
342         if (offload) {
343                 if (cookie->ol_flags & PKT_TX_TCP_SEG)
344                         cookie->ol_flags |= PKT_TX_TCP_CKSUM;
345
346                 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
347                 case PKT_TX_UDP_CKSUM:
348                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
349                         hdr->csum_offset = offsetof(struct udp_hdr,
350                                 dgram_cksum);
351                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
352                         break;
353
354                 case PKT_TX_TCP_CKSUM:
355                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
356                         hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
357                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
358                         break;
359
360                 default:
361                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
362                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
363                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
364                         break;
365                 }
366
367                 /* TCP Segmentation Offload */
368                 if (cookie->ol_flags & PKT_TX_TCP_SEG) {
369                         virtio_tso_fix_cksum(cookie);
370                         hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
371                                 VIRTIO_NET_HDR_GSO_TCPV6 :
372                                 VIRTIO_NET_HDR_GSO_TCPV4;
373                         hdr->gso_size = cookie->tso_segsz;
374                         hdr->hdr_len =
375                                 cookie->l2_len +
376                                 cookie->l3_len +
377                                 cookie->l4_len;
378                 } else {
379                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
380                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
381                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
382                 }
383         }
384
385         do {
386                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
387                 start_dp[idx].len   = cookie->data_len;
388                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
389                 idx = start_dp[idx].next;
390         } while ((cookie = cookie->next) != NULL);
391
392         if (use_indirect)
393                 idx = vq->vq_ring.desc[head_idx].next;
394
395         vq->vq_desc_head_idx = idx;
396         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
397                 vq->vq_desc_tail_idx = idx;
398         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
399         vq_update_avail_ring(vq, head_idx);
400 }
401
402 void
403 virtio_dev_cq_start(struct rte_eth_dev *dev)
404 {
405         struct virtio_hw *hw = dev->data->dev_private;
406
407         if (hw->cvq && hw->cvq->vq) {
408                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
409         }
410 }
411
412 int
413 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
414                         uint16_t queue_idx,
415                         uint16_t nb_desc,
416                         unsigned int socket_id __rte_unused,
417                         __rte_unused const struct rte_eth_rxconf *rx_conf,
418                         struct rte_mempool *mp)
419 {
420         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
421         struct virtio_hw *hw = dev->data->dev_private;
422         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
423         struct virtnet_rx *rxvq;
424
425         PMD_INIT_FUNC_TRACE();
426
427         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
428                 nb_desc = vq->vq_nentries;
429         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
430
431         rxvq = &vq->rxq;
432         rxvq->queue_id = queue_idx;
433         rxvq->mpool = mp;
434         if (rxvq->mpool == NULL) {
435                 rte_exit(EXIT_FAILURE,
436                         "Cannot allocate mbufs for rx virtqueue");
437         }
438         dev->data->rx_queues[queue_idx] = rxvq;
439
440         return 0;
441 }
442
443 int
444 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
445 {
446         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
447         struct virtio_hw *hw = dev->data->dev_private;
448         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
449         struct virtnet_rx *rxvq = &vq->rxq;
450         struct rte_mbuf *m;
451         uint16_t desc_idx;
452         int error, nbufs;
453
454         PMD_INIT_FUNC_TRACE();
455
456         /* Allocate blank mbufs for the each rx descriptor */
457         nbufs = 0;
458
459         if (hw->use_simple_rxtx) {
460                 for (desc_idx = 0; desc_idx < vq->vq_nentries;
461                      desc_idx++) {
462                         vq->vq_ring.avail->ring[desc_idx] = desc_idx;
463                         vq->vq_ring.desc[desc_idx].flags =
464                                 VRING_DESC_F_WRITE;
465                 }
466         }
467
468         memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
469         for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
470              desc_idx++) {
471                 vq->sw_ring[vq->vq_nentries + desc_idx] =
472                         &rxvq->fake_mbuf;
473         }
474
475         while (!virtqueue_full(vq)) {
476                 m = rte_mbuf_raw_alloc(rxvq->mpool);
477                 if (m == NULL)
478                         break;
479
480                 /* Enqueue allocated buffers */
481                 if (hw->use_simple_rxtx)
482                         error = virtqueue_enqueue_recv_refill_simple(vq, m);
483                 else
484                         error = virtqueue_enqueue_recv_refill(vq, m);
485
486                 if (error) {
487                         rte_pktmbuf_free(m);
488                         break;
489                 }
490                 nbufs++;
491         }
492
493         vq_update_avail_idx(vq);
494
495         PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
496
497         virtio_rxq_vec_setup(rxvq);
498
499         VIRTQUEUE_DUMP(vq);
500
501         return 0;
502 }
503
504 static void
505 virtio_update_rxtx_handler(struct rte_eth_dev *dev,
506                            const struct rte_eth_txconf *tx_conf)
507 {
508         uint8_t use_simple_rxtx = 0;
509         struct virtio_hw *hw = dev->data->dev_private;
510
511 #if defined RTE_ARCH_X86
512         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE3))
513                 use_simple_rxtx = 1;
514 #elif defined RTE_ARCH_ARM64 || defined CONFIG_RTE_ARCH_ARM
515         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_NEON))
516                 use_simple_rxtx = 1;
517 #endif
518         /* Use simple rx/tx func if single segment and no offloads */
519         if (use_simple_rxtx &&
520             (tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
521             !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
522                 PMD_INIT_LOG(INFO, "Using simple rx/tx path");
523                 dev->tx_pkt_burst = virtio_xmit_pkts_simple;
524                 dev->rx_pkt_burst = virtio_recv_pkts_vec;
525                 hw->use_simple_rxtx = use_simple_rxtx;
526         }
527 }
528
529 /*
530  * struct rte_eth_dev *dev: Used to update dev
531  * uint16_t nb_desc: Defaults to values read from config space
532  * unsigned int socket_id: Used to allocate memzone
533  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
534  * uint16_t queue_idx: Just used as an index in dev txq list
535  */
536 int
537 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
538                         uint16_t queue_idx,
539                         uint16_t nb_desc,
540                         unsigned int socket_id __rte_unused,
541                         const struct rte_eth_txconf *tx_conf)
542 {
543         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
544         struct virtio_hw *hw = dev->data->dev_private;
545         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
546         struct virtnet_tx *txvq;
547         uint16_t tx_free_thresh;
548
549         PMD_INIT_FUNC_TRACE();
550
551         virtio_update_rxtx_handler(dev, tx_conf);
552
553         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
554                 nb_desc = vq->vq_nentries;
555         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
556
557         txvq = &vq->txq;
558         txvq->queue_id = queue_idx;
559
560         tx_free_thresh = tx_conf->tx_free_thresh;
561         if (tx_free_thresh == 0)
562                 tx_free_thresh =
563                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
564
565         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
566                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
567                         "number of TX entries minus 3 (%u)."
568                         " (tx_free_thresh=%u port=%u queue=%u)\n",
569                         vq->vq_nentries - 3,
570                         tx_free_thresh, dev->data->port_id, queue_idx);
571                 return -EINVAL;
572         }
573
574         vq->vq_free_thresh = tx_free_thresh;
575
576         dev->data->tx_queues[queue_idx] = txvq;
577         return 0;
578 }
579
580 int
581 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
582                                 uint16_t queue_idx)
583 {
584         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
585         struct virtio_hw *hw = dev->data->dev_private;
586         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
587         uint16_t mid_idx = vq->vq_nentries >> 1;
588         struct virtnet_tx *txvq = &vq->txq;
589         uint16_t desc_idx;
590
591         PMD_INIT_FUNC_TRACE();
592
593         if (hw->use_simple_rxtx) {
594                 for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
595                         vq->vq_ring.avail->ring[desc_idx] =
596                                 desc_idx + mid_idx;
597                         vq->vq_ring.desc[desc_idx + mid_idx].next =
598                                 desc_idx;
599                         vq->vq_ring.desc[desc_idx + mid_idx].addr =
600                                 txvq->virtio_net_hdr_mem +
601                                 offsetof(struct virtio_tx_region, tx_hdr);
602                         vq->vq_ring.desc[desc_idx + mid_idx].len =
603                                 vq->hw->vtnet_hdr_size;
604                         vq->vq_ring.desc[desc_idx + mid_idx].flags =
605                                 VRING_DESC_F_NEXT;
606                         vq->vq_ring.desc[desc_idx].flags = 0;
607                 }
608                 for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
609                      desc_idx++)
610                         vq->vq_ring.avail->ring[desc_idx] = desc_idx;
611         }
612
613         VIRTQUEUE_DUMP(vq);
614
615         return 0;
616 }
617
618 static void
619 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
620 {
621         int error;
622         /*
623          * Requeue the discarded mbuf. This should always be
624          * successful since it was just dequeued.
625          */
626         error = virtqueue_enqueue_recv_refill(vq, m);
627         if (unlikely(error)) {
628                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
629                 rte_pktmbuf_free(m);
630         }
631 }
632
633 static void
634 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
635 {
636         uint32_t s = mbuf->pkt_len;
637         struct ether_addr *ea;
638
639         if (s == 64) {
640                 stats->size_bins[1]++;
641         } else if (s > 64 && s < 1024) {
642                 uint32_t bin;
643
644                 /* count zeros, and offset into correct bin */
645                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
646                 stats->size_bins[bin]++;
647         } else {
648                 if (s < 64)
649                         stats->size_bins[0]++;
650                 else if (s < 1519)
651                         stats->size_bins[6]++;
652                 else if (s >= 1519)
653                         stats->size_bins[7]++;
654         }
655
656         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
657         if (is_multicast_ether_addr(ea)) {
658                 if (is_broadcast_ether_addr(ea))
659                         stats->broadcast++;
660                 else
661                         stats->multicast++;
662         }
663 }
664
665 /* Optionally fill offload information in structure */
666 static int
667 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
668 {
669         struct rte_net_hdr_lens hdr_lens;
670         uint32_t hdrlen, ptype;
671         int l4_supported = 0;
672
673         /* nothing to do */
674         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
675                 return 0;
676
677         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
678
679         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
680         m->packet_type = ptype;
681         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
682             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
683             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
684                 l4_supported = 1;
685
686         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
687                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
688                 if (hdr->csum_start <= hdrlen && l4_supported) {
689                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
690                 } else {
691                         /* Unknown proto or tunnel, do sw cksum. We can assume
692                          * the cksum field is in the first segment since the
693                          * buffers we provided to the host are large enough.
694                          * In case of SCTP, this will be wrong since it's a CRC
695                          * but there's nothing we can do.
696                          */
697                         uint16_t csum, off;
698
699                         rte_raw_cksum_mbuf(m, hdr->csum_start,
700                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
701                                 &csum);
702                         if (likely(csum != 0xffff))
703                                 csum = ~csum;
704                         off = hdr->csum_offset + hdr->csum_start;
705                         if (rte_pktmbuf_data_len(m) >= off + 1)
706                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
707                                         off) = csum;
708                 }
709         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
710                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
711         }
712
713         /* GSO request, save required information in mbuf */
714         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
715                 /* Check unsupported modes */
716                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
717                     (hdr->gso_size == 0)) {
718                         return -EINVAL;
719                 }
720
721                 /* Update mss lengthes in mbuf */
722                 m->tso_segsz = hdr->gso_size;
723                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
724                         case VIRTIO_NET_HDR_GSO_TCPV4:
725                         case VIRTIO_NET_HDR_GSO_TCPV6:
726                                 m->ol_flags |= PKT_RX_LRO | \
727                                         PKT_RX_L4_CKSUM_NONE;
728                                 break;
729                         default:
730                                 return -EINVAL;
731                 }
732         }
733
734         return 0;
735 }
736
737 static inline int
738 rx_offload_enabled(struct virtio_hw *hw)
739 {
740         return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM) ||
741                 vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO4) ||
742                 vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO6);
743 }
744
745 #define VIRTIO_MBUF_BURST_SZ 64
746 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
747 uint16_t
748 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
749 {
750         struct virtnet_rx *rxvq = rx_queue;
751         struct virtqueue *vq = rxvq->vq;
752         struct virtio_hw *hw = vq->hw;
753         struct rte_mbuf *rxm, *new_mbuf;
754         uint16_t nb_used, num, nb_rx;
755         uint32_t len[VIRTIO_MBUF_BURST_SZ];
756         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
757         int error;
758         uint32_t i, nb_enqueued;
759         uint32_t hdr_size;
760         int offload;
761         struct virtio_net_hdr *hdr;
762
763         nb_rx = 0;
764         if (unlikely(hw->started == 0))
765                 return nb_rx;
766
767         nb_used = VIRTQUEUE_NUSED(vq);
768
769         virtio_rmb();
770
771         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
772         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
773                 num = VIRTIO_MBUF_BURST_SZ;
774         if (likely(num > DESC_PER_CACHELINE))
775                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
776
777         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
778         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
779
780         nb_enqueued = 0;
781         hdr_size = hw->vtnet_hdr_size;
782         offload = rx_offload_enabled(hw);
783
784         for (i = 0; i < num ; i++) {
785                 rxm = rcv_pkts[i];
786
787                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
788
789                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
790                         PMD_RX_LOG(ERR, "Packet drop");
791                         nb_enqueued++;
792                         virtio_discard_rxbuf(vq, rxm);
793                         rxvq->stats.errors++;
794                         continue;
795                 }
796
797                 rxm->port = rxvq->port_id;
798                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
799                 rxm->ol_flags = 0;
800                 rxm->vlan_tci = 0;
801
802                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
803                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
804
805                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
806                         RTE_PKTMBUF_HEADROOM - hdr_size);
807
808                 if (hw->vlan_strip)
809                         rte_vlan_strip(rxm);
810
811                 if (offload && virtio_rx_offload(rxm, hdr) < 0) {
812                         virtio_discard_rxbuf(vq, rxm);
813                         rxvq->stats.errors++;
814                         continue;
815                 }
816
817                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
818
819                 rx_pkts[nb_rx++] = rxm;
820
821                 rxvq->stats.bytes += rxm->pkt_len;
822                 virtio_update_packet_stats(&rxvq->stats, rxm);
823         }
824
825         rxvq->stats.packets += nb_rx;
826
827         /* Allocate new mbuf for the used descriptor */
828         error = ENOSPC;
829         while (likely(!virtqueue_full(vq))) {
830                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
831                 if (unlikely(new_mbuf == NULL)) {
832                         struct rte_eth_dev *dev
833                                 = &rte_eth_devices[rxvq->port_id];
834                         dev->data->rx_mbuf_alloc_failed++;
835                         break;
836                 }
837                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
838                 if (unlikely(error)) {
839                         rte_pktmbuf_free(new_mbuf);
840                         break;
841                 }
842                 nb_enqueued++;
843         }
844
845         if (likely(nb_enqueued)) {
846                 vq_update_avail_idx(vq);
847
848                 if (unlikely(virtqueue_kick_prepare(vq))) {
849                         virtqueue_notify(vq);
850                         PMD_RX_LOG(DEBUG, "Notified");
851                 }
852         }
853
854         return nb_rx;
855 }
856
857 uint16_t
858 virtio_recv_mergeable_pkts(void *rx_queue,
859                         struct rte_mbuf **rx_pkts,
860                         uint16_t nb_pkts)
861 {
862         struct virtnet_rx *rxvq = rx_queue;
863         struct virtqueue *vq = rxvq->vq;
864         struct virtio_hw *hw = vq->hw;
865         struct rte_mbuf *rxm, *new_mbuf;
866         uint16_t nb_used, num, nb_rx;
867         uint32_t len[VIRTIO_MBUF_BURST_SZ];
868         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
869         struct rte_mbuf *prev;
870         int error;
871         uint32_t i, nb_enqueued;
872         uint32_t seg_num;
873         uint16_t extra_idx;
874         uint32_t seg_res;
875         uint32_t hdr_size;
876         int offload;
877
878         nb_rx = 0;
879         if (unlikely(hw->started == 0))
880                 return nb_rx;
881
882         nb_used = VIRTQUEUE_NUSED(vq);
883
884         virtio_rmb();
885
886         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
887
888         i = 0;
889         nb_enqueued = 0;
890         seg_num = 0;
891         extra_idx = 0;
892         seg_res = 0;
893         hdr_size = hw->vtnet_hdr_size;
894         offload = rx_offload_enabled(hw);
895
896         while (i < nb_used) {
897                 struct virtio_net_hdr_mrg_rxbuf *header;
898
899                 if (nb_rx == nb_pkts)
900                         break;
901
902                 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
903                 if (num != 1)
904                         continue;
905
906                 i++;
907
908                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
909                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
910
911                 rxm = rcv_pkts[0];
912
913                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
914                         PMD_RX_LOG(ERR, "Packet drop");
915                         nb_enqueued++;
916                         virtio_discard_rxbuf(vq, rxm);
917                         rxvq->stats.errors++;
918                         continue;
919                 }
920
921                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
922                         RTE_PKTMBUF_HEADROOM - hdr_size);
923                 seg_num = header->num_buffers;
924
925                 if (seg_num == 0)
926                         seg_num = 1;
927
928                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
929                 rxm->nb_segs = seg_num;
930                 rxm->ol_flags = 0;
931                 rxm->vlan_tci = 0;
932                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
933                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
934
935                 rxm->port = rxvq->port_id;
936                 rx_pkts[nb_rx] = rxm;
937                 prev = rxm;
938
939                 if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
940                         virtio_discard_rxbuf(vq, rxm);
941                         rxvq->stats.errors++;
942                         continue;
943                 }
944
945                 seg_res = seg_num - 1;
946
947                 while (seg_res != 0) {
948                         /*
949                          * Get extra segments for current uncompleted packet.
950                          */
951                         uint16_t  rcv_cnt =
952                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
953                         if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
954                                 uint32_t rx_num =
955                                         virtqueue_dequeue_burst_rx(vq,
956                                         rcv_pkts, len, rcv_cnt);
957                                 i += rx_num;
958                                 rcv_cnt = rx_num;
959                         } else {
960                                 PMD_RX_LOG(ERR,
961                                            "No enough segments for packet.");
962                                 nb_enqueued++;
963                                 virtio_discard_rxbuf(vq, rxm);
964                                 rxvq->stats.errors++;
965                                 break;
966                         }
967
968                         extra_idx = 0;
969
970                         while (extra_idx < rcv_cnt) {
971                                 rxm = rcv_pkts[extra_idx];
972
973                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
974                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
975                                 rxm->data_len = (uint16_t)(len[extra_idx]);
976
977                                 if (prev)
978                                         prev->next = rxm;
979
980                                 prev = rxm;
981                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
982                                 extra_idx++;
983                         };
984                         seg_res -= rcv_cnt;
985                 }
986
987                 if (hw->vlan_strip)
988                         rte_vlan_strip(rx_pkts[nb_rx]);
989
990                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
991                         rx_pkts[nb_rx]->data_len);
992
993                 rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
994                 virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
995                 nb_rx++;
996         }
997
998         rxvq->stats.packets += nb_rx;
999
1000         /* Allocate new mbuf for the used descriptor */
1001         error = ENOSPC;
1002         while (likely(!virtqueue_full(vq))) {
1003                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
1004                 if (unlikely(new_mbuf == NULL)) {
1005                         struct rte_eth_dev *dev
1006                                 = &rte_eth_devices[rxvq->port_id];
1007                         dev->data->rx_mbuf_alloc_failed++;
1008                         break;
1009                 }
1010                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
1011                 if (unlikely(error)) {
1012                         rte_pktmbuf_free(new_mbuf);
1013                         break;
1014                 }
1015                 nb_enqueued++;
1016         }
1017
1018         if (likely(nb_enqueued)) {
1019                 vq_update_avail_idx(vq);
1020
1021                 if (unlikely(virtqueue_kick_prepare(vq))) {
1022                         virtqueue_notify(vq);
1023                         PMD_RX_LOG(DEBUG, "Notified");
1024                 }
1025         }
1026
1027         return nb_rx;
1028 }
1029
1030 uint16_t
1031 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1032 {
1033         struct virtnet_tx *txvq = tx_queue;
1034         struct virtqueue *vq = txvq->vq;
1035         struct virtio_hw *hw = vq->hw;
1036         uint16_t hdr_size = hw->vtnet_hdr_size;
1037         uint16_t nb_used, nb_tx = 0;
1038         int error;
1039
1040         if (unlikely(hw->started == 0))
1041                 return nb_tx;
1042
1043         if (unlikely(nb_pkts < 1))
1044                 return nb_pkts;
1045
1046         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1047         nb_used = VIRTQUEUE_NUSED(vq);
1048
1049         virtio_rmb();
1050         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1051                 virtio_xmit_cleanup(vq, nb_used);
1052
1053         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1054                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1055                 int can_push = 0, use_indirect = 0, slots, need;
1056
1057                 /* Do VLAN tag insertion */
1058                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1059                         error = rte_vlan_insert(&txm);
1060                         if (unlikely(error)) {
1061                                 rte_pktmbuf_free(txm);
1062                                 continue;
1063                         }
1064                 }
1065
1066                 /* optimize ring usage */
1067                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1068                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1069                     rte_mbuf_refcnt_read(txm) == 1 &&
1070                     RTE_MBUF_DIRECT(txm) &&
1071                     txm->nb_segs == 1 &&
1072                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1073                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1074                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1075                         can_push = 1;
1076                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1077                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1078                         use_indirect = 1;
1079
1080                 /* How many main ring entries are needed to this Tx?
1081                  * any_layout => number of segments
1082                  * indirect   => 1
1083                  * default    => number of segments + 1
1084                  */
1085                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1086                 need = slots - vq->vq_free_cnt;
1087
1088                 /* Positive value indicates it need free vring descriptors */
1089                 if (unlikely(need > 0)) {
1090                         nb_used = VIRTQUEUE_NUSED(vq);
1091                         virtio_rmb();
1092                         need = RTE_MIN(need, (int)nb_used);
1093
1094                         virtio_xmit_cleanup(vq, need);
1095                         need = slots - vq->vq_free_cnt;
1096                         if (unlikely(need > 0)) {
1097                                 PMD_TX_LOG(ERR,
1098                                            "No free tx descriptors to transmit");
1099                                 break;
1100                         }
1101                 }
1102
1103                 /* Enqueue Packet buffers */
1104                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
1105
1106                 txvq->stats.bytes += txm->pkt_len;
1107                 virtio_update_packet_stats(&txvq->stats, txm);
1108         }
1109
1110         txvq->stats.packets += nb_tx;
1111
1112         if (likely(nb_tx)) {
1113                 vq_update_avail_idx(vq);
1114
1115                 if (unlikely(virtqueue_kick_prepare(vq))) {
1116                         virtqueue_notify(vq);
1117                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1118                 }
1119         }
1120
1121         return nb_tx;
1122 }