1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2016-2018 Microsoft Corporation
3 * Copyright(c) 2013-2016 Brocade Communications Systems, Inc.
15 #include <rte_ethdev.h>
16 #include <rte_memcpy.h>
17 #include <rte_string_fns.h>
18 #include <rte_memzone.h>
19 #include <rte_malloc.h>
20 #include <rte_atomic.h>
21 #include <rte_branch_prediction.h>
22 #include <rte_ether.h>
23 #include <rte_common.h>
24 #include <rte_errno.h>
25 #include <rte_memory.h>
29 #include <rte_bus_vmbus.h>
30 #include <rte_spinlock.h>
38 #define HN_NVS_SEND_MSG_SIZE \
39 (sizeof(struct vmbus_chanpkt_hdr) + sizeof(struct hn_nvs_rndis))
41 #define HN_TXD_CACHE_SIZE 32 /* per cpu tx_descriptor pool cache */
42 #define HN_TXCOPY_THRESHOLD 512
44 #define HN_RXCOPY_THRESHOLD 256
45 #define HN_RXQ_EVENT_DEFAULT 2048
54 #define HN_RXINFO_VLAN 0x0001
55 #define HN_RXINFO_CSUM 0x0002
56 #define HN_RXINFO_HASHINF 0x0004
57 #define HN_RXINFO_HASHVAL 0x0008
58 #define HN_RXINFO_ALL \
64 #define HN_NDIS_VLAN_INFO_INVALID 0xffffffff
65 #define HN_NDIS_RXCSUM_INFO_INVALID 0
66 #define HN_NDIS_HASH_INFO_INVALID 0
69 * Per-transmit book keeping.
70 * A slot in transmit ring (chim_index) is reserved for each transmit.
72 * There are two types of transmit:
73 * - buffered transmit where chimney buffer is used and RNDIS header
74 * is in the buffer. mbuf == NULL for this case.
76 * - direct transmit where RNDIS header is in the in rndis_pkt
77 * mbuf is freed after transmit.
79 * Descriptors come from per-port pool which is used
80 * to limit number of outstanding requests per device.
91 struct rndis_packet_msg *rndis_pkt;
94 #define HN_RNDIS_PKT_LEN \
95 (sizeof(struct rndis_packet_msg) + \
96 RNDIS_PKTINFO_SIZE(NDIS_HASH_VALUE_SIZE) + \
97 RNDIS_PKTINFO_SIZE(NDIS_VLAN_INFO_SIZE) + \
98 RNDIS_PKTINFO_SIZE(NDIS_LSO2_INFO_SIZE) + \
99 RNDIS_PKTINFO_SIZE(NDIS_TXCSUM_INFO_SIZE))
101 /* Minimum space required for a packet */
102 #define HN_PKTSIZE_MIN(align) \
103 RTE_ALIGN(ETHER_MIN_LEN + HN_RNDIS_PKT_LEN, align)
105 #define DEFAULT_TX_FREE_THRESH 32U
108 hn_update_packet_stats(struct hn_stats *stats, const struct rte_mbuf *m)
110 uint32_t s = m->pkt_len;
111 const struct ether_addr *ea;
114 stats->size_bins[1]++;
115 } else if (s > 64 && s < 1024) {
118 /* count zeros, and offset into correct bin */
119 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
120 stats->size_bins[bin]++;
123 stats->size_bins[0]++;
125 stats->size_bins[6]++;
127 stats->size_bins[7]++;
130 ea = rte_pktmbuf_mtod(m, const struct ether_addr *);
131 if (is_multicast_ether_addr(ea)) {
132 if (is_broadcast_ether_addr(ea))
139 static inline unsigned int hn_rndis_pktlen(const struct rndis_packet_msg *pkt)
141 return pkt->pktinfooffset + pkt->pktinfolen;
144 static inline uint32_t
145 hn_rndis_pktmsg_offset(uint32_t ofs)
147 return ofs - offsetof(struct rndis_packet_msg, dataoffset);
150 static void hn_txd_init(struct rte_mempool *mp __rte_unused,
151 void *opaque, void *obj, unsigned int idx)
153 struct hn_txdesc *txd = obj;
154 struct rte_eth_dev *dev = opaque;
155 struct rndis_packet_msg *pkt;
157 memset(txd, 0, sizeof(*txd));
158 txd->chim_index = idx;
160 pkt = rte_malloc_socket("RNDIS_TX", HN_RNDIS_PKT_LEN,
161 rte_align32pow2(HN_RNDIS_PKT_LEN),
162 dev->device->numa_node);
164 rte_exit(EXIT_FAILURE, "can not allocate RNDIS header");
166 txd->rndis_pkt = pkt;
170 * Unlike Linux and FreeBSD, this driver uses a mempool
171 * to limit outstanding transmits and reserve buffers
174 hn_tx_pool_init(struct rte_eth_dev *dev)
176 struct hn_data *hv = dev->data->dev_private;
177 char name[RTE_MEMPOOL_NAMESIZE];
178 struct rte_mempool *mp;
180 snprintf(name, sizeof(name),
181 "hn_txd_%u", dev->data->port_id);
183 PMD_INIT_LOG(DEBUG, "create a TX send pool %s n=%u size=%zu socket=%d",
184 name, hv->chim_cnt, sizeof(struct hn_txdesc),
185 dev->device->numa_node);
187 mp = rte_mempool_create(name, hv->chim_cnt, sizeof(struct hn_txdesc),
188 HN_TXD_CACHE_SIZE, 0,
191 dev->device->numa_node, 0);
194 "mempool %s create failed: %d", name, rte_errno);
202 static void hn_reset_txagg(struct hn_tx_queue *txq)
204 txq->agg_szleft = txq->agg_szmax;
205 txq->agg_pktleft = txq->agg_pktmax;
207 txq->agg_prevpkt = NULL;
211 hn_dev_tx_queue_setup(struct rte_eth_dev *dev,
212 uint16_t queue_idx, uint16_t nb_desc __rte_unused,
213 unsigned int socket_id,
214 const struct rte_eth_txconf *tx_conf)
217 struct hn_data *hv = dev->data->dev_private;
218 struct hn_tx_queue *txq;
219 uint32_t tx_free_thresh;
222 PMD_INIT_FUNC_TRACE();
224 txq = rte_zmalloc_socket("HN_TXQ", sizeof(*txq), RTE_CACHE_LINE_SIZE,
230 txq->chan = hv->channels[queue_idx];
231 txq->port_id = dev->data->port_id;
232 txq->queue_id = queue_idx;
234 tx_free_thresh = tx_conf->tx_free_thresh;
235 if (tx_free_thresh == 0)
236 tx_free_thresh = RTE_MIN(hv->chim_cnt / 4,
237 DEFAULT_TX_FREE_THRESH);
239 if (tx_free_thresh >= hv->chim_cnt - 3)
240 tx_free_thresh = hv->chim_cnt - 3;
242 txq->free_thresh = tx_free_thresh;
244 txq->agg_szmax = RTE_MIN(hv->chim_szmax, hv->rndis_agg_size);
245 txq->agg_pktmax = hv->rndis_agg_pkts;
246 txq->agg_align = hv->rndis_agg_align;
250 err = hn_vf_tx_queue_setup(dev, queue_idx, nb_desc,
257 dev->data->tx_queues[queue_idx] = txq;
262 hn_dev_tx_queue_release(void *arg)
264 struct hn_tx_queue *txq = arg;
265 struct hn_txdesc *txd;
267 PMD_INIT_FUNC_TRACE();
272 /* If any pending data is still present just drop it */
275 rte_mempool_put(txq->hv->tx_pool, txd);
281 hn_nvs_send_completed(struct rte_eth_dev *dev, uint16_t queue_id,
282 unsigned long xactid, const struct hn_nvs_rndis_ack *ack)
284 struct hn_txdesc *txd = (struct hn_txdesc *)xactid;
285 struct hn_tx_queue *txq;
287 /* Control packets are sent with xacid == 0 */
291 txq = dev->data->tx_queues[queue_id];
292 if (likely(ack->status == NVS_STATUS_OK)) {
293 PMD_TX_LOG(DEBUG, "port %u:%u complete tx %u packets %u bytes %u",
294 txq->port_id, txq->queue_id, txd->chim_index,
295 txd->packets, txd->data_size);
296 txq->stats.bytes += txd->data_size;
297 txq->stats.packets += txd->packets;
299 PMD_TX_LOG(NOTICE, "port %u:%u complete tx %u failed status %u",
300 txq->port_id, txq->queue_id, txd->chim_index, ack->status);
304 rte_pktmbuf_free(txd->m);
306 rte_mempool_put(txq->hv->tx_pool, txd);
309 /* Handle transmit completion events */
311 hn_nvs_handle_comp(struct rte_eth_dev *dev, uint16_t queue_id,
312 const struct vmbus_chanpkt_hdr *pkt,
315 const struct hn_nvs_hdr *hdr = data;
318 case NVS_TYPE_RNDIS_ACK:
319 hn_nvs_send_completed(dev, queue_id, pkt->xactid, data);
324 "unexpected send completion type %u",
329 /* Parse per-packet info (meta data) */
331 hn_rndis_rxinfo(const void *info_data, unsigned int info_dlen,
332 struct hn_rxinfo *info)
334 const struct rndis_pktinfo *pi = info_data;
337 while (info_dlen != 0) {
341 if (unlikely(info_dlen < sizeof(*pi)))
344 if (unlikely(info_dlen < pi->size))
346 info_dlen -= pi->size;
348 if (unlikely(pi->size & RNDIS_PKTINFO_SIZE_ALIGNMASK))
350 if (unlikely(pi->size < pi->offset))
353 dlen = pi->size - pi->offset;
357 case NDIS_PKTINFO_TYPE_VLAN:
358 if (unlikely(dlen < NDIS_VLAN_INFO_SIZE))
360 info->vlan_info = *((const uint32_t *)data);
361 mask |= HN_RXINFO_VLAN;
364 case NDIS_PKTINFO_TYPE_CSUM:
365 if (unlikely(dlen < NDIS_RXCSUM_INFO_SIZE))
367 info->csum_info = *((const uint32_t *)data);
368 mask |= HN_RXINFO_CSUM;
371 case NDIS_PKTINFO_TYPE_HASHVAL:
372 if (unlikely(dlen < NDIS_HASH_VALUE_SIZE))
374 info->hash_value = *((const uint32_t *)data);
375 mask |= HN_RXINFO_HASHVAL;
378 case NDIS_PKTINFO_TYPE_HASHINF:
379 if (unlikely(dlen < NDIS_HASH_INFO_SIZE))
381 info->hash_info = *((const uint32_t *)data);
382 mask |= HN_RXINFO_HASHINF;
389 if (mask == HN_RXINFO_ALL)
390 break; /* All found; done */
392 pi = (const struct rndis_pktinfo *)
393 ((const uint8_t *)pi + pi->size);
398 * - If there is no hash value, invalidate the hash info.
400 if (!(mask & HN_RXINFO_HASHVAL))
401 info->hash_info = HN_NDIS_HASH_INFO_INVALID;
406 * Ack the consumed RXBUF associated w/ this channel packet,
407 * so that this RXBUF can be recycled by the hypervisor.
409 static void hn_rx_buf_release(struct hn_rx_bufinfo *rxb)
411 struct rte_mbuf_ext_shared_info *shinfo = &rxb->shinfo;
412 struct hn_data *hv = rxb->hv;
414 if (rte_mbuf_ext_refcnt_update(shinfo, -1) == 0) {
415 hn_nvs_ack_rxbuf(rxb->chan, rxb->xactid);
416 --hv->rxbuf_outstanding;
420 static void hn_rx_buf_free_cb(void *buf __rte_unused, void *opaque)
422 hn_rx_buf_release(opaque);
425 static struct hn_rx_bufinfo *hn_rx_buf_init(const struct hn_rx_queue *rxq,
426 const struct vmbus_chanpkt_rxbuf *pkt)
428 struct hn_rx_bufinfo *rxb;
430 rxb = rxq->hv->rxbuf_info + pkt->hdr.xactid;
431 rxb->chan = rxq->chan;
432 rxb->xactid = pkt->hdr.xactid;
435 rxb->shinfo.free_cb = hn_rx_buf_free_cb;
436 rxb->shinfo.fcb_opaque = rxb;
437 rte_mbuf_ext_refcnt_set(&rxb->shinfo, 1);
441 static void hn_rxpkt(struct hn_rx_queue *rxq, struct hn_rx_bufinfo *rxb,
442 uint8_t *data, unsigned int headroom, unsigned int dlen,
443 const struct hn_rxinfo *info)
445 struct hn_data *hv = rxq->hv;
448 m = rte_pktmbuf_alloc(rxq->mb_pool);
450 struct rte_eth_dev *dev =
451 &rte_eth_devices[rxq->port_id];
453 dev->data->rx_mbuf_alloc_failed++;
458 * For large packets, avoid copy if possible but need to keep
459 * some space available in receive area for later packets.
461 if (dlen >= HN_RXCOPY_THRESHOLD &&
462 hv->rxbuf_outstanding < hv->rxbuf_section_cnt / 2) {
463 struct rte_mbuf_ext_shared_info *shinfo;
468 * Build an external mbuf that points to recveive area.
469 * Use refcount to handle multiple packets in same
470 * receive buffer section.
472 rxbuf = hv->rxbuf_res->addr;
473 iova = rte_mem_virt2iova(rxbuf) + RTE_PTR_DIFF(data, rxbuf);
474 shinfo = &rxb->shinfo;
476 if (rte_mbuf_ext_refcnt_update(shinfo, 1) == 1)
477 ++hv->rxbuf_outstanding;
479 rte_pktmbuf_attach_extbuf(m, data, iova,
480 dlen + headroom, shinfo);
481 m->data_off = headroom;
483 /* Mbuf's in pool must be large enough to hold small packets */
484 if (unlikely(rte_pktmbuf_tailroom(m) < dlen)) {
485 rte_pktmbuf_free_seg(m);
489 rte_memcpy(rte_pktmbuf_mtod(m, void *),
490 data + headroom, dlen);
493 m->port = rxq->port_id;
496 m->packet_type = rte_net_get_ptype(m, NULL,
501 if (info->vlan_info != HN_NDIS_VLAN_INFO_INVALID) {
502 m->vlan_tci = info->vlan_info;
503 m->ol_flags |= PKT_RX_VLAN_STRIPPED | PKT_RX_VLAN;
505 /* NDIS always strips tag, put it back if necessary */
506 if (!hv->vlan_strip && rte_vlan_insert(&m)) {
507 PMD_DRV_LOG(DEBUG, "vlan insert failed");
514 if (info->csum_info != HN_NDIS_RXCSUM_INFO_INVALID) {
515 if (info->csum_info & NDIS_RXCSUM_INFO_IPCS_OK)
516 m->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
518 if (info->csum_info & (NDIS_RXCSUM_INFO_UDPCS_OK
519 | NDIS_RXCSUM_INFO_TCPCS_OK))
520 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
521 else if (info->csum_info & (NDIS_RXCSUM_INFO_TCPCS_FAILED
522 | NDIS_RXCSUM_INFO_UDPCS_FAILED))
523 m->ol_flags |= PKT_RX_L4_CKSUM_BAD;
526 if (info->hash_info != HN_NDIS_HASH_INFO_INVALID) {
527 m->ol_flags |= PKT_RX_RSS_HASH;
528 m->hash.rss = info->hash_value;
532 "port %u:%u RX id %"PRIu64" size %u type %#x ol_flags %#"PRIx64,
533 rxq->port_id, rxq->queue_id, rxb->xactid,
534 m->pkt_len, m->packet_type, m->ol_flags);
536 ++rxq->stats.packets;
537 rxq->stats.bytes += m->pkt_len;
538 hn_update_packet_stats(&rxq->stats, m);
540 if (unlikely(rte_ring_sp_enqueue(rxq->rx_ring, m) != 0)) {
541 ++rxq->stats.ring_full;
546 static void hn_rndis_rx_data(struct hn_rx_queue *rxq,
547 struct hn_rx_bufinfo *rxb,
548 void *data, uint32_t dlen)
550 unsigned int data_off, data_len, pktinfo_off, pktinfo_len;
551 const struct rndis_packet_msg *pkt = data;
552 struct hn_rxinfo info = {
553 .vlan_info = HN_NDIS_VLAN_INFO_INVALID,
554 .csum_info = HN_NDIS_RXCSUM_INFO_INVALID,
555 .hash_info = HN_NDIS_HASH_INFO_INVALID,
561 if (unlikely(dlen < sizeof(*pkt)))
564 if (unlikely(dlen < pkt->len))
565 goto error; /* truncated RNDIS from host */
567 if (unlikely(pkt->len < pkt->datalen
568 + pkt->oobdatalen + pkt->pktinfolen))
571 if (unlikely(pkt->datalen == 0))
575 if (unlikely(pkt->dataoffset < RNDIS_PACKET_MSG_OFFSET_MIN))
578 if (likely(pkt->pktinfooffset > 0) &&
579 unlikely(pkt->pktinfooffset < RNDIS_PACKET_MSG_OFFSET_MIN ||
580 (pkt->pktinfooffset & RNDIS_PACKET_MSG_OFFSET_ALIGNMASK)))
583 data_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
584 data_len = pkt->datalen;
585 pktinfo_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->pktinfooffset);
586 pktinfo_len = pkt->pktinfolen;
588 if (likely(pktinfo_len > 0)) {
589 err = hn_rndis_rxinfo((const uint8_t *)pkt + pktinfo_off,
595 if (unlikely(data_off + data_len > pkt->len))
598 if (unlikely(data_len < ETHER_HDR_LEN))
601 hn_rxpkt(rxq, rxb, data, data_off, data_len, &info);
608 hn_rndis_receive(struct rte_eth_dev *dev, struct hn_rx_queue *rxq,
609 struct hn_rx_bufinfo *rxb, void *buf, uint32_t len)
611 const struct rndis_msghdr *hdr = buf;
614 case RNDIS_PACKET_MSG:
615 if (dev->data->dev_started)
616 hn_rndis_rx_data(rxq, rxb, buf, len);
619 case RNDIS_INDICATE_STATUS_MSG:
620 hn_rndis_link_status(dev, buf);
623 case RNDIS_INITIALIZE_CMPLT:
624 case RNDIS_QUERY_CMPLT:
625 case RNDIS_SET_CMPLT:
626 hn_rndis_receive_response(rxq->hv, buf, len);
631 "unexpected RNDIS message (type %#x len %u)",
638 hn_nvs_handle_rxbuf(struct rte_eth_dev *dev,
640 struct hn_rx_queue *rxq,
641 const struct vmbus_chanpkt_hdr *hdr,
644 const struct vmbus_chanpkt_rxbuf *pkt;
645 const struct hn_nvs_hdr *nvs_hdr = buf;
646 uint32_t rxbuf_sz = hv->rxbuf_res->len;
647 char *rxbuf = hv->rxbuf_res->addr;
648 unsigned int i, hlen, count;
649 struct hn_rx_bufinfo *rxb;
651 /* At minimum we need type header */
652 if (unlikely(vmbus_chanpkt_datalen(hdr) < sizeof(*nvs_hdr))) {
653 PMD_RX_LOG(ERR, "invalid receive nvs RNDIS");
657 /* Make sure that this is a RNDIS message. */
658 if (unlikely(nvs_hdr->type != NVS_TYPE_RNDIS)) {
659 PMD_RX_LOG(ERR, "nvs type %u, not RNDIS",
664 hlen = vmbus_chanpkt_getlen(hdr->hlen);
665 if (unlikely(hlen < sizeof(*pkt))) {
666 PMD_RX_LOG(ERR, "invalid rxbuf chanpkt");
670 pkt = container_of(hdr, const struct vmbus_chanpkt_rxbuf, hdr);
671 if (unlikely(pkt->rxbuf_id != NVS_RXBUF_SIG)) {
672 PMD_RX_LOG(ERR, "invalid rxbuf_id 0x%08x",
677 count = pkt->rxbuf_cnt;
678 if (unlikely(hlen < offsetof(struct vmbus_chanpkt_rxbuf,
680 PMD_RX_LOG(ERR, "invalid rxbuf_cnt %u", count);
684 if (pkt->hdr.xactid > hv->rxbuf_section_cnt) {
685 PMD_RX_LOG(ERR, "invalid rxbuf section id %" PRIx64,
690 /* Setup receive buffer info to allow for callback */
691 rxb = hn_rx_buf_init(rxq, pkt);
693 /* Each range represents 1 RNDIS pkt that contains 1 Ethernet frame */
694 for (i = 0; i < count; ++i) {
695 unsigned int ofs, len;
697 ofs = pkt->rxbuf[i].ofs;
698 len = pkt->rxbuf[i].len;
700 if (unlikely(ofs + len > rxbuf_sz)) {
702 "%uth RNDIS msg overflow ofs %u, len %u",
707 if (unlikely(len == 0)) {
708 PMD_RX_LOG(ERR, "%uth RNDIS msg len %u", i, len);
712 hn_rndis_receive(dev, rxq, rxb,
716 /* Send ACK now if external mbuf not used */
717 hn_rx_buf_release(rxb);
721 * Called when NVS inband events are received.
722 * Send up a two part message with port_id and the NVS message
723 * to the pipe to the netvsc-vf-event control thread.
725 static void hn_nvs_handle_notify(struct rte_eth_dev *dev,
726 const struct vmbus_chanpkt_hdr *pkt,
729 const struct hn_nvs_hdr *hdr = data;
732 case NVS_TYPE_TXTBL_NOTE:
733 /* Transmit indirection table has locking problems
734 * in DPDK and therefore not implemented
736 PMD_DRV_LOG(DEBUG, "host notify of transmit indirection table");
739 case NVS_TYPE_VFASSOC_NOTE:
740 hn_nvs_handle_vfassoc(dev, pkt, data);
745 "got notify, nvs type %u", hdr->type);
749 struct hn_rx_queue *hn_rx_queue_alloc(struct hn_data *hv,
751 unsigned int socket_id)
753 struct hn_rx_queue *rxq;
755 rxq = rte_zmalloc_socket("HN_RXQ", sizeof(*rxq),
756 RTE_CACHE_LINE_SIZE, socket_id);
761 rxq->chan = hv->channels[queue_id];
762 rte_spinlock_init(&rxq->ring_lock);
763 rxq->port_id = hv->port_id;
764 rxq->queue_id = queue_id;
765 rxq->event_sz = HN_RXQ_EVENT_DEFAULT;
766 rxq->event_buf = rte_malloc_socket("HN_EVENTS", HN_RXQ_EVENT_DEFAULT,
767 RTE_CACHE_LINE_SIZE, socket_id);
768 if (!rxq->event_buf) {
777 hn_dev_rx_queue_setup(struct rte_eth_dev *dev,
778 uint16_t queue_idx, uint16_t nb_desc,
779 unsigned int socket_id,
780 const struct rte_eth_rxconf *rx_conf,
781 struct rte_mempool *mp)
783 struct hn_data *hv = dev->data->dev_private;
784 char ring_name[RTE_RING_NAMESIZE];
785 struct hn_rx_queue *rxq;
789 PMD_INIT_FUNC_TRACE();
791 if (queue_idx == 0) {
794 rxq = hn_rx_queue_alloc(hv, queue_idx, socket_id);
800 count = rte_mempool_avail_count(mp) / dev->data->nb_rx_queues;
801 if (nb_desc == 0 || nb_desc > count)
805 * Staging ring from receive event logic to rx_pkts.
806 * rx_pkts assumes caller is handling multi-thread issue.
807 * event logic has locking.
809 snprintf(ring_name, sizeof(ring_name),
810 "hn_rx_%u_%u", dev->data->port_id, queue_idx);
811 rxq->rx_ring = rte_ring_create(ring_name,
812 rte_align32pow2(nb_desc),
817 error = hn_vf_rx_queue_setup(dev, queue_idx, nb_desc,
818 socket_id, rx_conf, mp);
822 dev->data->rx_queues[queue_idx] = rxq;
826 rte_ring_free(rxq->rx_ring);
827 rte_free(rxq->event_buf);
833 hn_dev_rx_queue_release(void *arg)
835 struct hn_rx_queue *rxq = arg;
837 PMD_INIT_FUNC_TRACE();
842 rte_ring_free(rxq->rx_ring);
846 hn_vf_rx_queue_release(rxq->hv, rxq->queue_id);
848 /* Keep primary queue to allow for control operations */
849 if (rxq != rxq->hv->primary) {
850 rte_free(rxq->event_buf);
856 hn_dev_tx_done_cleanup(void *arg, uint32_t free_cnt)
858 struct hn_tx_queue *txq = arg;
860 return hn_process_events(txq->hv, txq->queue_id, free_cnt);
864 * Process pending events on the channel.
865 * Called from both Rx queue poll and Tx cleanup
867 uint32_t hn_process_events(struct hn_data *hv, uint16_t queue_id,
870 struct rte_eth_dev *dev = &rte_eth_devices[hv->port_id];
871 struct hn_rx_queue *rxq;
872 uint32_t bytes_read = 0;
873 uint32_t tx_done = 0;
876 rxq = queue_id == 0 ? hv->primary : dev->data->rx_queues[queue_id];
878 /* If no pending data then nothing to do */
879 if (rte_vmbus_chan_rx_empty(rxq->chan))
883 * Since channel is shared between Rx and TX queue need to have a lock
884 * since DPDK does not force same CPU to be used for Rx/Tx.
886 if (unlikely(!rte_spinlock_trylock(&rxq->ring_lock)))
890 const struct vmbus_chanpkt_hdr *pkt;
891 uint32_t len = rxq->event_sz;
895 ret = rte_vmbus_chan_recv_raw(rxq->chan, rxq->event_buf, &len);
897 break; /* ring is empty */
899 if (unlikely(ret == -ENOBUFS)) {
900 /* event buffer not large enough to read ring */
903 "event buffer expansion (need %u)", len);
904 rxq->event_sz = len + len / 4;
905 rxq->event_buf = rte_realloc(rxq->event_buf, rxq->event_sz,
906 RTE_CACHE_LINE_SIZE);
909 /* out of memory, no more events now */
914 if (unlikely(ret <= 0)) {
915 /* This indicates a failure to communicate (or worse) */
916 rte_exit(EXIT_FAILURE,
917 "vmbus ring buffer error: %d", ret);
921 pkt = (const struct vmbus_chanpkt_hdr *)rxq->event_buf;
922 data = (char *)rxq->event_buf + vmbus_chanpkt_getlen(pkt->hlen);
925 case VMBUS_CHANPKT_TYPE_COMP:
927 hn_nvs_handle_comp(dev, queue_id, pkt, data);
930 case VMBUS_CHANPKT_TYPE_RXBUF:
931 hn_nvs_handle_rxbuf(dev, hv, rxq, pkt, data);
934 case VMBUS_CHANPKT_TYPE_INBAND:
935 hn_nvs_handle_notify(dev, pkt, data);
939 PMD_DRV_LOG(ERR, "unknown chan pkt %u", pkt->type);
943 if (tx_limit && tx_done >= tx_limit)
946 if (rxq->rx_ring && rte_ring_full(rxq->rx_ring))
951 rte_vmbus_chan_signal_read(rxq->chan, bytes_read);
953 rte_spinlock_unlock(&rxq->ring_lock);
958 static void hn_append_to_chim(struct hn_tx_queue *txq,
959 struct rndis_packet_msg *pkt,
960 const struct rte_mbuf *m)
962 struct hn_txdesc *txd = txq->agg_txd;
963 uint8_t *buf = (uint8_t *)pkt;
964 unsigned int data_offs;
968 data_offs = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
969 txd->chim_size += pkt->len;
970 txd->data_size += m->pkt_len;
972 hn_update_packet_stats(&txq->stats, m);
974 for (; m; m = m->next) {
975 uint16_t len = rte_pktmbuf_data_len(m);
977 rte_memcpy(buf + data_offs,
978 rte_pktmbuf_mtod(m, const char *), len);
984 * Send pending aggregated data in chimney buffer (if any).
985 * Returns error if send was unsuccessful because channel ring buffer
988 static int hn_flush_txagg(struct hn_tx_queue *txq, bool *need_sig)
991 struct hn_txdesc *txd = txq->agg_txd;
992 struct hn_nvs_rndis rndis;
998 rndis = (struct hn_nvs_rndis) {
999 .type = NVS_TYPE_RNDIS,
1000 .rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1001 .chim_idx = txd->chim_index,
1002 .chim_sz = txd->chim_size,
1005 PMD_TX_LOG(DEBUG, "port %u:%u tx %u size %u",
1006 txq->port_id, txq->queue_id, txd->chim_index, txd->chim_size);
1008 ret = hn_nvs_send(txq->chan, VMBUS_CHANPKT_FLAG_RC,
1009 &rndis, sizeof(rndis), (uintptr_t)txd, need_sig);
1011 if (likely(ret == 0))
1012 hn_reset_txagg(txq);
1014 PMD_TX_LOG(NOTICE, "port %u:%u send failed: %d",
1015 txq->port_id, txq->queue_id, ret);
1020 static struct hn_txdesc *hn_new_txd(struct hn_data *hv,
1021 struct hn_tx_queue *txq)
1023 struct hn_txdesc *txd;
1025 if (rte_mempool_get(hv->tx_pool, (void **)&txd)) {
1026 ++txq->stats.ring_full;
1027 PMD_TX_LOG(DEBUG, "tx pool exhausted!");
1032 txd->queue_id = txq->queue_id;
1041 hn_try_txagg(struct hn_data *hv, struct hn_tx_queue *txq, uint32_t pktsize)
1043 struct hn_txdesc *agg_txd = txq->agg_txd;
1044 struct rndis_packet_msg *pkt;
1048 unsigned int padding, olen;
1051 * Update the previous RNDIS packet's total length,
1052 * it can be increased due to the mandatory alignment
1053 * padding for this RNDIS packet. And update the
1054 * aggregating txdesc's chimney sending buffer size
1057 * Zero-out the padding, as required by the RNDIS spec.
1059 pkt = txq->agg_prevpkt;
1061 padding = RTE_ALIGN(olen, txq->agg_align) - olen;
1063 agg_txd->chim_size += padding;
1064 pkt->len += padding;
1065 memset((uint8_t *)pkt + olen, 0, padding);
1068 chim = (uint8_t *)pkt + pkt->len;
1071 txq->agg_szleft -= pktsize;
1072 if (txq->agg_szleft < HN_PKTSIZE_MIN(txq->agg_align)) {
1074 * Probably can't aggregate more packets,
1075 * flush this aggregating txdesc proactively.
1077 txq->agg_pktleft = 0;
1080 agg_txd = hn_new_txd(hv, txq);
1084 chim = (uint8_t *)hv->chim_res->addr
1085 + agg_txd->chim_index * hv->chim_szmax;
1087 txq->agg_txd = agg_txd;
1088 txq->agg_pktleft = txq->agg_pktmax - 1;
1089 txq->agg_szleft = txq->agg_szmax - pktsize;
1091 txq->agg_prevpkt = chim;
1096 static inline void *
1097 hn_rndis_pktinfo_append(struct rndis_packet_msg *pkt,
1098 uint32_t pi_dlen, uint32_t pi_type)
1100 const uint32_t pi_size = RNDIS_PKTINFO_SIZE(pi_dlen);
1101 struct rndis_pktinfo *pi;
1104 * Per-packet-info does not move; it only grows.
1107 * pktinfooffset in this phase counts from the beginning
1108 * of rndis_packet_msg.
1110 pi = (struct rndis_pktinfo *)((uint8_t *)pkt + hn_rndis_pktlen(pkt));
1112 pkt->pktinfolen += pi_size;
1116 pi->offset = RNDIS_PKTINFO_OFFSET;
1121 /* Put RNDIS header and packet info on packet */
1122 static void hn_encap(struct rndis_packet_msg *pkt,
1124 const struct rte_mbuf *m)
1126 unsigned int hlen = m->l2_len + m->l3_len;
1130 pkt->type = RNDIS_PACKET_MSG;
1131 pkt->len = m->pkt_len;
1132 pkt->dataoffset = 0;
1133 pkt->datalen = m->pkt_len;
1134 pkt->oobdataoffset = 0;
1135 pkt->oobdatalen = 0;
1136 pkt->oobdataelements = 0;
1137 pkt->pktinfooffset = sizeof(*pkt);
1138 pkt->pktinfolen = 0;
1143 * Set the hash value for this packet, to the queue_id to cause
1144 * TX done event for this packet on the right channel.
1146 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_HASH_VALUE_SIZE,
1147 NDIS_PKTINFO_TYPE_HASHVAL);
1148 *pi_data = queue_id;
1150 if (m->ol_flags & PKT_TX_VLAN_PKT) {
1151 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_VLAN_INFO_SIZE,
1152 NDIS_PKTINFO_TYPE_VLAN);
1153 *pi_data = m->vlan_tci;
1156 if (m->ol_flags & PKT_TX_TCP_SEG) {
1157 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_LSO2_INFO_SIZE,
1158 NDIS_PKTINFO_TYPE_LSO);
1160 if (m->ol_flags & PKT_TX_IPV6) {
1161 *pi_data = NDIS_LSO2_INFO_MAKEIPV6(hlen,
1164 *pi_data = NDIS_LSO2_INFO_MAKEIPV4(hlen,
1167 } else if (m->ol_flags &
1168 (PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM | PKT_TX_IP_CKSUM)) {
1169 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_TXCSUM_INFO_SIZE,
1170 NDIS_PKTINFO_TYPE_CSUM);
1173 if (m->ol_flags & PKT_TX_IPV6)
1174 *pi_data |= NDIS_TXCSUM_INFO_IPV6;
1175 if (m->ol_flags & PKT_TX_IPV4) {
1176 *pi_data |= NDIS_TXCSUM_INFO_IPV4;
1178 if (m->ol_flags & PKT_TX_IP_CKSUM)
1179 *pi_data |= NDIS_TXCSUM_INFO_IPCS;
1182 if (m->ol_flags & PKT_TX_TCP_CKSUM)
1183 *pi_data |= NDIS_TXCSUM_INFO_MKTCPCS(hlen);
1184 else if (m->ol_flags & PKT_TX_UDP_CKSUM)
1185 *pi_data |= NDIS_TXCSUM_INFO_MKUDPCS(hlen);
1188 pkt_hlen = pkt->pktinfooffset + pkt->pktinfolen;
1189 /* Fixup RNDIS packet message total length */
1190 pkt->len += pkt_hlen;
1192 /* Convert RNDIS packet message offsets */
1193 pkt->dataoffset = hn_rndis_pktmsg_offset(pkt_hlen);
1194 pkt->pktinfooffset = hn_rndis_pktmsg_offset(pkt->pktinfooffset);
1197 /* How many scatter gather list elements ar needed */
1198 static unsigned int hn_get_slots(const struct rte_mbuf *m)
1200 unsigned int slots = 1; /* for RNDIS header */
1203 unsigned int size = rte_pktmbuf_data_len(m);
1204 unsigned int offs = rte_mbuf_data_iova(m) & PAGE_MASK;
1206 slots += (offs + size + PAGE_SIZE - 1) / PAGE_SIZE;
1213 /* Build scatter gather list from chained mbuf */
1214 static unsigned int hn_fill_sg(struct vmbus_gpa *sg,
1215 const struct rte_mbuf *m)
1217 unsigned int segs = 0;
1220 rte_iova_t addr = rte_mbuf_data_iova(m);
1221 unsigned int page = addr / PAGE_SIZE;
1222 unsigned int offset = addr & PAGE_MASK;
1223 unsigned int len = rte_pktmbuf_data_len(m);
1226 unsigned int bytes = RTE_MIN(len, PAGE_SIZE - offset);
1228 sg[segs].page = page;
1229 sg[segs].ofs = offset;
1230 sg[segs].len = bytes;
1243 /* Transmit directly from mbuf */
1244 static int hn_xmit_sg(struct hn_tx_queue *txq,
1245 const struct hn_txdesc *txd, const struct rte_mbuf *m,
1248 struct vmbus_gpa sg[hn_get_slots(m)];
1249 struct hn_nvs_rndis nvs_rndis = {
1250 .type = NVS_TYPE_RNDIS,
1251 .rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1252 .chim_sz = txd->chim_size,
1257 /* attach aggregation data if present */
1258 if (txd->chim_size > 0)
1259 nvs_rndis.chim_idx = txd->chim_index;
1261 nvs_rndis.chim_idx = NVS_CHIM_IDX_INVALID;
1263 hn_rndis_dump(txd->rndis_pkt);
1265 /* pass IOVA of rndis header in first segment */
1266 addr = rte_malloc_virt2iova(txd->rndis_pkt);
1267 if (unlikely(addr == RTE_BAD_IOVA)) {
1268 PMD_DRV_LOG(ERR, "RNDIS transmit can not get iova");
1272 sg[0].page = addr / PAGE_SIZE;
1273 sg[0].ofs = addr & PAGE_MASK;
1274 sg[0].len = RNDIS_PACKET_MSG_OFFSET_ABS(hn_rndis_pktlen(txd->rndis_pkt));
1277 hn_update_packet_stats(&txq->stats, m);
1279 segs += hn_fill_sg(sg + 1, m);
1281 PMD_TX_LOG(DEBUG, "port %u:%u tx %u segs %u size %u",
1282 txq->port_id, txq->queue_id, txd->chim_index,
1283 segs, nvs_rndis.chim_sz);
1285 return hn_nvs_send_sglist(txq->chan, sg, segs,
1286 &nvs_rndis, sizeof(nvs_rndis),
1287 (uintptr_t)txd, need_sig);
1291 hn_xmit_pkts(void *ptxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1293 struct hn_tx_queue *txq = ptxq;
1294 uint16_t queue_id = txq->queue_id;
1295 struct hn_data *hv = txq->hv;
1296 struct rte_eth_dev *vf_dev;
1297 bool need_sig = false;
1301 if (unlikely(hv->closed))
1304 /* Transmit over VF if present and up */
1305 vf_dev = hv->vf_dev;
1306 rte_compiler_barrier();
1307 if (vf_dev && vf_dev->data->dev_started) {
1308 void *sub_q = vf_dev->data->tx_queues[queue_id];
1310 return (*vf_dev->tx_pkt_burst)(sub_q, tx_pkts, nb_pkts);
1313 if (rte_mempool_avail_count(hv->tx_pool) <= txq->free_thresh)
1314 hn_process_events(hv, txq->queue_id, 0);
1316 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1317 struct rte_mbuf *m = tx_pkts[nb_tx];
1318 uint32_t pkt_size = m->pkt_len + HN_RNDIS_PKT_LEN;
1319 struct rndis_packet_msg *pkt;
1321 /* For small packets aggregate them in chimney buffer */
1322 if (m->pkt_len < HN_TXCOPY_THRESHOLD && pkt_size <= txq->agg_szmax) {
1323 /* If this packet will not fit, then flush */
1324 if (txq->agg_pktleft == 0 ||
1325 RTE_ALIGN(pkt_size, txq->agg_align) > txq->agg_szleft) {
1326 if (hn_flush_txagg(txq, &need_sig))
1330 pkt = hn_try_txagg(hv, txq, pkt_size);
1334 hn_encap(pkt, queue_id, m);
1335 hn_append_to_chim(txq, pkt, m);
1337 rte_pktmbuf_free(m);
1339 /* if buffer is full, flush */
1340 if (txq->agg_pktleft == 0 &&
1341 hn_flush_txagg(txq, &need_sig))
1344 struct hn_txdesc *txd;
1346 /* can send chimney data and large packet at once */
1349 hn_reset_txagg(txq);
1351 txd = hn_new_txd(hv, txq);
1356 pkt = txd->rndis_pkt;
1358 txd->data_size += m->pkt_len;
1361 hn_encap(pkt, queue_id, m);
1363 ret = hn_xmit_sg(txq, txd, m, &need_sig);
1364 if (unlikely(ret != 0)) {
1365 PMD_TX_LOG(NOTICE, "sg send failed: %d", ret);
1366 ++txq->stats.errors;
1367 rte_mempool_put(hv->tx_pool, txd);
1373 /* If partial buffer left, then try and send it.
1374 * if that fails, then reuse it on next send.
1376 hn_flush_txagg(txq, &need_sig);
1380 rte_vmbus_chan_signal_tx(txq->chan);
1386 hn_recv_pkts(void *prxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1388 struct hn_rx_queue *rxq = prxq;
1389 struct hn_data *hv = rxq->hv;
1390 struct rte_eth_dev *vf_dev;
1393 if (unlikely(hv->closed))
1396 vf_dev = hv->vf_dev;
1397 rte_compiler_barrier();
1399 if (vf_dev && vf_dev->data->dev_started) {
1400 /* Normally, with SR-IOV the ring buffer will be empty */
1401 hn_process_events(hv, rxq->queue_id, 0);
1403 /* Get mbufs some bufs off of staging ring */
1404 nb_rcv = rte_ring_sc_dequeue_burst(rxq->rx_ring,
1407 /* And rest off of VF */
1408 nb_rcv += rte_eth_rx_burst(vf_dev->data->port_id,
1410 rx_pkts + nb_rcv, nb_pkts - nb_rcv);
1412 /* If receive ring is not full then get more */
1413 if (rte_ring_count(rxq->rx_ring) < nb_pkts)
1414 hn_process_events(hv, rxq->queue_id, 0);
1416 nb_rcv = rte_ring_sc_dequeue_burst(rxq->rx_ring,