1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2016-2018 Microsoft Corporation
3 * Copyright(c) 2013-2016 Brocade Communications Systems, Inc.
14 #include <rte_ethdev.h>
15 #include <rte_memcpy.h>
16 #include <rte_string_fns.h>
17 #include <rte_memzone.h>
18 #include <rte_malloc.h>
19 #include <rte_atomic.h>
20 #include <rte_branch_prediction.h>
21 #include <rte_ether.h>
22 #include <rte_common.h>
23 #include <rte_errno.h>
24 #include <rte_memory.h>
27 #include <rte_bus_vmbus.h>
28 #include <rte_spinlock.h>
36 #define HN_NVS_SEND_MSG_SIZE \
37 (sizeof(struct vmbus_chanpkt_hdr) + sizeof(struct hn_nvs_rndis))
39 #define HN_TXD_CACHE_SIZE 32 /* per cpu tx_descriptor pool cache */
40 #define HN_TXCOPY_THRESHOLD 512
42 #define HN_RXCOPY_THRESHOLD 256
43 #define HN_RXQ_EVENT_DEFAULT 2048
52 #define HN_RXINFO_VLAN 0x0001
53 #define HN_RXINFO_CSUM 0x0002
54 #define HN_RXINFO_HASHINF 0x0004
55 #define HN_RXINFO_HASHVAL 0x0008
56 #define HN_RXINFO_ALL \
62 #define HN_NDIS_VLAN_INFO_INVALID 0xffffffff
63 #define HN_NDIS_RXCSUM_INFO_INVALID 0
64 #define HN_NDIS_HASH_INFO_INVALID 0
67 * Per-transmit book keeping.
68 * A slot in transmit ring (chim_index) is reserved for each transmit.
70 * There are two types of transmit:
71 * - buffered transmit where chimney buffer is used and RNDIS header
72 * is in the buffer. mbuf == NULL for this case.
74 * - direct transmit where RNDIS header is in the in rndis_pkt
75 * mbuf is freed after transmit.
77 * Descriptors come from per-port pool which is used
78 * to limit number of outstanding requests per device.
89 struct rndis_packet_msg *rndis_pkt;
92 #define HN_RNDIS_PKT_LEN \
93 (sizeof(struct rndis_packet_msg) + \
94 RNDIS_PKTINFO_SIZE(NDIS_HASH_VALUE_SIZE) + \
95 RNDIS_PKTINFO_SIZE(NDIS_VLAN_INFO_SIZE) + \
96 RNDIS_PKTINFO_SIZE(NDIS_LSO2_INFO_SIZE) + \
97 RNDIS_PKTINFO_SIZE(NDIS_TXCSUM_INFO_SIZE))
99 /* Minimum space required for a packet */
100 #define HN_PKTSIZE_MIN(align) \
101 RTE_ALIGN(ETHER_MIN_LEN + HN_RNDIS_PKT_LEN, align)
103 #define DEFAULT_TX_FREE_THRESH 32U
106 hn_update_packet_stats(struct hn_stats *stats, const struct rte_mbuf *m)
108 uint32_t s = m->pkt_len;
109 const struct ether_addr *ea;
112 stats->size_bins[1]++;
113 } else if (s > 64 && s < 1024) {
116 /* count zeros, and offset into correct bin */
117 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
118 stats->size_bins[bin]++;
121 stats->size_bins[0]++;
123 stats->size_bins[6]++;
125 stats->size_bins[7]++;
128 ea = rte_pktmbuf_mtod(m, const struct ether_addr *);
129 if (is_multicast_ether_addr(ea)) {
130 if (is_broadcast_ether_addr(ea))
137 static inline unsigned int hn_rndis_pktlen(const struct rndis_packet_msg *pkt)
139 return pkt->pktinfooffset + pkt->pktinfolen;
142 static inline uint32_t
143 hn_rndis_pktmsg_offset(uint32_t ofs)
145 return ofs - offsetof(struct rndis_packet_msg, dataoffset);
148 static void hn_txd_init(struct rte_mempool *mp __rte_unused,
149 void *opaque, void *obj, unsigned int idx)
151 struct hn_txdesc *txd = obj;
152 struct rte_eth_dev *dev = opaque;
153 struct rndis_packet_msg *pkt;
155 memset(txd, 0, sizeof(*txd));
156 txd->chim_index = idx;
158 pkt = rte_malloc_socket("RNDIS_TX", HN_RNDIS_PKT_LEN,
159 rte_align32pow2(HN_RNDIS_PKT_LEN),
160 dev->device->numa_node);
162 rte_exit(EXIT_FAILURE, "can not allocate RNDIS header");
164 txd->rndis_pkt = pkt;
168 * Unlike Linux and FreeBSD, this driver uses a mempool
169 * to limit outstanding transmits and reserve buffers
172 hn_tx_pool_init(struct rte_eth_dev *dev)
174 struct hn_data *hv = dev->data->dev_private;
175 char name[RTE_MEMPOOL_NAMESIZE];
176 struct rte_mempool *mp;
178 snprintf(name, sizeof(name),
179 "hn_txd_%u", dev->data->port_id);
181 PMD_INIT_LOG(DEBUG, "create a TX send pool %s n=%u size=%zu socket=%d",
182 name, hv->chim_cnt, sizeof(struct hn_txdesc),
183 dev->device->numa_node);
185 mp = rte_mempool_create(name, hv->chim_cnt, sizeof(struct hn_txdesc),
186 HN_TXD_CACHE_SIZE, 0,
189 dev->device->numa_node, 0);
192 "mempool %s create failed: %d", name, rte_errno);
200 static void hn_reset_txagg(struct hn_tx_queue *txq)
202 txq->agg_szleft = txq->agg_szmax;
203 txq->agg_pktleft = txq->agg_pktmax;
205 txq->agg_prevpkt = NULL;
209 hn_dev_tx_queue_setup(struct rte_eth_dev *dev,
210 uint16_t queue_idx, uint16_t nb_desc __rte_unused,
211 unsigned int socket_id,
212 const struct rte_eth_txconf *tx_conf)
215 struct hn_data *hv = dev->data->dev_private;
216 struct hn_tx_queue *txq;
217 uint32_t tx_free_thresh;
219 PMD_INIT_FUNC_TRACE();
221 txq = rte_zmalloc_socket("HN_TXQ", sizeof(*txq), RTE_CACHE_LINE_SIZE,
227 txq->chan = hv->channels[queue_idx];
228 txq->port_id = dev->data->port_id;
229 txq->queue_id = queue_idx;
231 tx_free_thresh = tx_conf->tx_free_thresh;
232 if (tx_free_thresh == 0)
233 tx_free_thresh = RTE_MIN(hv->chim_cnt / 4,
234 DEFAULT_TX_FREE_THRESH);
236 if (tx_free_thresh >= hv->chim_cnt - 3)
237 tx_free_thresh = hv->chim_cnt - 3;
239 txq->free_thresh = tx_free_thresh;
241 txq->agg_szmax = RTE_MIN(hv->chim_szmax, hv->rndis_agg_size);
242 txq->agg_pktmax = hv->rndis_agg_pkts;
243 txq->agg_align = hv->rndis_agg_align;
247 dev->data->tx_queues[queue_idx] = txq;
253 hn_dev_tx_queue_release(void *arg)
255 struct hn_tx_queue *txq = arg;
256 struct hn_txdesc *txd;
258 PMD_INIT_FUNC_TRACE();
263 /* If any pending data is still present just drop it */
266 rte_mempool_put(txq->hv->tx_pool, txd);
272 hn_dev_tx_queue_info(struct rte_eth_dev *dev, uint16_t queue_idx,
273 struct rte_eth_txq_info *qinfo)
275 struct hn_data *hv = dev->data->dev_private;
276 struct hn_tx_queue *txq = dev->data->rx_queues[queue_idx];
278 qinfo->conf.tx_free_thresh = txq->free_thresh;
279 qinfo->nb_desc = hv->tx_pool->size;
283 hn_nvs_send_completed(struct rte_eth_dev *dev, uint16_t queue_id,
284 unsigned long xactid, const struct hn_nvs_rndis_ack *ack)
286 struct hn_txdesc *txd = (struct hn_txdesc *)xactid;
287 struct hn_tx_queue *txq;
289 /* Control packets are sent with xacid == 0 */
293 txq = dev->data->tx_queues[queue_id];
294 if (likely(ack->status == NVS_STATUS_OK)) {
295 PMD_TX_LOG(DEBUG, "port %u:%u complete tx %u packets %u bytes %u",
296 txq->port_id, txq->queue_id, txd->chim_index,
297 txd->packets, txd->data_size);
298 txq->stats.bytes += txd->data_size;
299 txq->stats.packets += txd->packets;
301 PMD_TX_LOG(NOTICE, "port %u:%u complete tx %u failed status %u",
302 txq->port_id, txq->queue_id, txd->chim_index, ack->status);
306 rte_pktmbuf_free(txd->m);
308 rte_mempool_put(txq->hv->tx_pool, txd);
311 /* Handle transmit completion events */
313 hn_nvs_handle_comp(struct rte_eth_dev *dev, uint16_t queue_id,
314 const struct vmbus_chanpkt_hdr *pkt,
317 const struct hn_nvs_hdr *hdr = data;
320 case NVS_TYPE_RNDIS_ACK:
321 hn_nvs_send_completed(dev, queue_id, pkt->xactid, data);
326 "unexpected send completion type %u",
331 /* Parse per-packet info (meta data) */
333 hn_rndis_rxinfo(const void *info_data, unsigned int info_dlen,
334 struct hn_rxinfo *info)
336 const struct rndis_pktinfo *pi = info_data;
339 while (info_dlen != 0) {
343 if (unlikely(info_dlen < sizeof(*pi)))
346 if (unlikely(info_dlen < pi->size))
348 info_dlen -= pi->size;
350 if (unlikely(pi->size & RNDIS_PKTINFO_SIZE_ALIGNMASK))
352 if (unlikely(pi->size < pi->offset))
355 dlen = pi->size - pi->offset;
359 case NDIS_PKTINFO_TYPE_VLAN:
360 if (unlikely(dlen < NDIS_VLAN_INFO_SIZE))
362 info->vlan_info = *((const uint32_t *)data);
363 mask |= HN_RXINFO_VLAN;
366 case NDIS_PKTINFO_TYPE_CSUM:
367 if (unlikely(dlen < NDIS_RXCSUM_INFO_SIZE))
369 info->csum_info = *((const uint32_t *)data);
370 mask |= HN_RXINFO_CSUM;
373 case NDIS_PKTINFO_TYPE_HASHVAL:
374 if (unlikely(dlen < NDIS_HASH_VALUE_SIZE))
376 info->hash_value = *((const uint32_t *)data);
377 mask |= HN_RXINFO_HASHVAL;
380 case NDIS_PKTINFO_TYPE_HASHINF:
381 if (unlikely(dlen < NDIS_HASH_INFO_SIZE))
383 info->hash_info = *((const uint32_t *)data);
384 mask |= HN_RXINFO_HASHINF;
391 if (mask == HN_RXINFO_ALL)
392 break; /* All found; done */
394 pi = (const struct rndis_pktinfo *)
395 ((const uint8_t *)pi + pi->size);
400 * - If there is no hash value, invalidate the hash info.
402 if (!(mask & HN_RXINFO_HASHVAL))
403 info->hash_info = HN_NDIS_HASH_INFO_INVALID;
408 * Ack the consumed RXBUF associated w/ this channel packet,
409 * so that this RXBUF can be recycled by the hypervisor.
411 static void hn_rx_buf_release(struct hn_rx_bufinfo *rxb)
413 struct rte_mbuf_ext_shared_info *shinfo = &rxb->shinfo;
414 struct hn_data *hv = rxb->hv;
416 if (rte_mbuf_ext_refcnt_update(shinfo, -1) == 0) {
417 hn_nvs_ack_rxbuf(rxb->chan, rxb->xactid);
418 --hv->rxbuf_outstanding;
422 static void hn_rx_buf_free_cb(void *buf __rte_unused, void *opaque)
424 hn_rx_buf_release(opaque);
427 static struct hn_rx_bufinfo *hn_rx_buf_init(const struct hn_rx_queue *rxq,
428 const struct vmbus_chanpkt_rxbuf *pkt)
430 struct hn_rx_bufinfo *rxb;
432 rxb = rxq->hv->rxbuf_info + pkt->hdr.xactid;
433 rxb->chan = rxq->chan;
434 rxb->xactid = pkt->hdr.xactid;
437 rxb->shinfo.free_cb = hn_rx_buf_free_cb;
438 rxb->shinfo.fcb_opaque = rxb;
439 rte_mbuf_ext_refcnt_set(&rxb->shinfo, 1);
443 static void hn_rxpkt(struct hn_rx_queue *rxq, struct hn_rx_bufinfo *rxb,
444 uint8_t *data, unsigned int headroom, unsigned int dlen,
445 const struct hn_rxinfo *info)
447 struct hn_data *hv = rxq->hv;
450 m = rte_pktmbuf_alloc(rxq->mb_pool);
452 struct rte_eth_dev *dev =
453 &rte_eth_devices[rxq->port_id];
455 dev->data->rx_mbuf_alloc_failed++;
460 * For large packets, avoid copy if possible but need to keep
461 * some space available in receive area for later packets.
463 if (dlen >= HN_RXCOPY_THRESHOLD &&
464 hv->rxbuf_outstanding < hv->rxbuf_section_cnt / 2) {
465 struct rte_mbuf_ext_shared_info *shinfo;
470 * Build an external mbuf that points to recveive area.
471 * Use refcount to handle multiple packets in same
472 * receive buffer section.
474 rxbuf = hv->rxbuf_res->addr;
475 iova = rte_mem_virt2iova(rxbuf) + RTE_PTR_DIFF(data, rxbuf);
476 shinfo = &rxb->shinfo;
478 if (rte_mbuf_ext_refcnt_update(shinfo, 1) == 1)
479 ++hv->rxbuf_outstanding;
481 rte_pktmbuf_attach_extbuf(m, data, iova,
482 dlen + headroom, shinfo);
483 m->data_off = headroom;
485 /* Mbuf's in pool must be large enough to hold small packets */
486 if (unlikely(rte_pktmbuf_tailroom(m) < dlen)) {
487 rte_pktmbuf_free_seg(m);
491 rte_memcpy(rte_pktmbuf_mtod(m, void *),
492 data + headroom, dlen);
495 m->port = rxq->port_id;
499 if (info->vlan_info != HN_NDIS_VLAN_INFO_INVALID) {
500 m->vlan_tci = info->vlan_info;
501 m->ol_flags |= PKT_RX_VLAN_STRIPPED | PKT_RX_VLAN;
504 if (info->csum_info != HN_NDIS_RXCSUM_INFO_INVALID) {
505 if (info->csum_info & NDIS_RXCSUM_INFO_IPCS_OK)
506 m->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
508 if (info->csum_info & (NDIS_RXCSUM_INFO_UDPCS_OK
509 | NDIS_RXCSUM_INFO_TCPCS_OK))
510 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
513 if (info->hash_info != HN_NDIS_HASH_INFO_INVALID) {
514 m->ol_flags |= PKT_RX_RSS_HASH;
515 m->hash.rss = info->hash_value;
518 PMD_RX_LOG(DEBUG, "port %u:%u RX id %" PRIu64 " size %u ol_flags %#" PRIx64,
519 rxq->port_id, rxq->queue_id, rxb->xactid,
520 m->pkt_len, m->ol_flags);
522 ++rxq->stats.packets;
523 rxq->stats.bytes += m->pkt_len;
524 hn_update_packet_stats(&rxq->stats, m);
526 if (unlikely(rte_ring_sp_enqueue(rxq->rx_ring, m) != 0)) {
532 static void hn_rndis_rx_data(struct hn_rx_queue *rxq,
533 struct hn_rx_bufinfo *rxb,
534 void *data, uint32_t dlen)
536 unsigned int data_off, data_len, pktinfo_off, pktinfo_len;
537 const struct rndis_packet_msg *pkt = data;
538 struct hn_rxinfo info = {
539 .vlan_info = HN_NDIS_VLAN_INFO_INVALID,
540 .csum_info = HN_NDIS_RXCSUM_INFO_INVALID,
541 .hash_info = HN_NDIS_HASH_INFO_INVALID,
547 if (unlikely(dlen < sizeof(*pkt)))
550 if (unlikely(dlen < pkt->len))
551 goto error; /* truncated RNDIS from host */
553 if (unlikely(pkt->len < pkt->datalen
554 + pkt->oobdatalen + pkt->pktinfolen))
557 if (unlikely(pkt->datalen == 0))
561 if (unlikely(pkt->dataoffset < RNDIS_PACKET_MSG_OFFSET_MIN))
564 if (likely(pkt->pktinfooffset > 0) &&
565 unlikely(pkt->pktinfooffset < RNDIS_PACKET_MSG_OFFSET_MIN ||
566 (pkt->pktinfooffset & RNDIS_PACKET_MSG_OFFSET_ALIGNMASK)))
569 data_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
570 data_len = pkt->datalen;
571 pktinfo_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->pktinfooffset);
572 pktinfo_len = pkt->pktinfolen;
574 if (likely(pktinfo_len > 0)) {
575 err = hn_rndis_rxinfo((const uint8_t *)pkt + pktinfo_off,
581 if (unlikely(data_off + data_len > pkt->len))
584 if (unlikely(data_len < ETHER_HDR_LEN))
587 hn_rxpkt(rxq, rxb, data, data_off, data_len, &info);
594 hn_rndis_receive(const struct rte_eth_dev *dev, struct hn_rx_queue *rxq,
595 struct hn_rx_bufinfo *rxb, void *buf, uint32_t len)
597 const struct rndis_msghdr *hdr = buf;
600 case RNDIS_PACKET_MSG:
601 if (dev->data->dev_started)
602 hn_rndis_rx_data(rxq, rxb, buf, len);
605 case RNDIS_INDICATE_STATUS_MSG:
606 hn_rndis_link_status(rxq->hv, buf);
609 case RNDIS_INITIALIZE_CMPLT:
610 case RNDIS_QUERY_CMPLT:
611 case RNDIS_SET_CMPLT:
612 hn_rndis_receive_response(rxq->hv, buf, len);
617 "unexpected RNDIS message (type %#x len %u)",
624 hn_nvs_handle_rxbuf(struct rte_eth_dev *dev,
626 struct hn_rx_queue *rxq,
627 const struct vmbus_chanpkt_hdr *hdr,
630 const struct vmbus_chanpkt_rxbuf *pkt;
631 const struct hn_nvs_hdr *nvs_hdr = buf;
632 uint32_t rxbuf_sz = hv->rxbuf_res->len;
633 char *rxbuf = hv->rxbuf_res->addr;
634 unsigned int i, hlen, count;
635 struct hn_rx_bufinfo *rxb;
637 /* At minimum we need type header */
638 if (unlikely(vmbus_chanpkt_datalen(hdr) < sizeof(*nvs_hdr))) {
639 PMD_RX_LOG(ERR, "invalid receive nvs RNDIS");
643 /* Make sure that this is a RNDIS message. */
644 if (unlikely(nvs_hdr->type != NVS_TYPE_RNDIS)) {
645 PMD_RX_LOG(ERR, "nvs type %u, not RNDIS",
650 hlen = vmbus_chanpkt_getlen(hdr->hlen);
651 if (unlikely(hlen < sizeof(*pkt))) {
652 PMD_RX_LOG(ERR, "invalid rxbuf chanpkt");
656 pkt = container_of(hdr, const struct vmbus_chanpkt_rxbuf, hdr);
657 if (unlikely(pkt->rxbuf_id != NVS_RXBUF_SIG)) {
658 PMD_RX_LOG(ERR, "invalid rxbuf_id 0x%08x",
663 count = pkt->rxbuf_cnt;
664 if (unlikely(hlen < offsetof(struct vmbus_chanpkt_rxbuf,
666 PMD_RX_LOG(ERR, "invalid rxbuf_cnt %u", count);
670 if (pkt->hdr.xactid > hv->rxbuf_section_cnt) {
671 PMD_RX_LOG(ERR, "invalid rxbuf section id %" PRIx64,
676 /* Setup receive buffer info to allow for callback */
677 rxb = hn_rx_buf_init(rxq, pkt);
679 /* Each range represents 1 RNDIS pkt that contains 1 Ethernet frame */
680 for (i = 0; i < count; ++i) {
681 unsigned int ofs, len;
683 ofs = pkt->rxbuf[i].ofs;
684 len = pkt->rxbuf[i].len;
686 if (unlikely(ofs + len > rxbuf_sz)) {
688 "%uth RNDIS msg overflow ofs %u, len %u",
693 if (unlikely(len == 0)) {
694 PMD_RX_LOG(ERR, "%uth RNDIS msg len %u", i, len);
698 hn_rndis_receive(dev, rxq, rxb,
702 /* Send ACK now if external mbuf not used */
703 hn_rx_buf_release(rxb);
706 struct hn_rx_queue *hn_rx_queue_alloc(struct hn_data *hv,
708 unsigned int socket_id)
710 struct hn_rx_queue *rxq;
712 rxq = rte_zmalloc_socket("HN_RXQ",
713 sizeof(*rxq) + HN_RXQ_EVENT_DEFAULT,
714 RTE_CACHE_LINE_SIZE, socket_id);
717 rxq->chan = hv->channels[queue_id];
718 rte_spinlock_init(&rxq->ring_lock);
719 rxq->port_id = hv->port_id;
720 rxq->queue_id = queue_id;
726 hn_dev_rx_queue_setup(struct rte_eth_dev *dev,
727 uint16_t queue_idx, uint16_t nb_desc,
728 unsigned int socket_id,
729 const struct rte_eth_rxconf *rx_conf __rte_unused,
730 struct rte_mempool *mp)
732 struct hn_data *hv = dev->data->dev_private;
733 char ring_name[RTE_RING_NAMESIZE];
734 struct hn_rx_queue *rxq;
737 PMD_INIT_FUNC_TRACE();
739 if (queue_idx == 0) {
742 rxq = hn_rx_queue_alloc(hv, queue_idx, socket_id);
748 count = rte_mempool_avail_count(mp) / dev->data->nb_rx_queues;
749 if (nb_desc == 0 || nb_desc > count)
753 * Staging ring from receive event logic to rx_pkts.
754 * rx_pkts assumes caller is handling multi-thread issue.
755 * event logic has locking.
757 snprintf(ring_name, sizeof(ring_name),
758 "hn_rx_%u_%u", dev->data->port_id, queue_idx);
759 rxq->rx_ring = rte_ring_create(ring_name,
760 rte_align32pow2(nb_desc),
765 dev->data->rx_queues[queue_idx] = rxq;
769 rte_ring_free(rxq->rx_ring);
770 rte_free(rxq->event_buf);
776 hn_dev_rx_queue_release(void *arg)
778 struct hn_rx_queue *rxq = arg;
780 PMD_INIT_FUNC_TRACE();
785 rte_ring_free(rxq->rx_ring);
789 if (rxq != rxq->hv->primary) {
790 rte_free(rxq->event_buf);
796 hn_dev_rx_queue_info(struct rte_eth_dev *dev, uint16_t queue_idx,
797 struct rte_eth_rxq_info *qinfo)
799 struct hn_rx_queue *rxq = dev->data->rx_queues[queue_idx];
801 qinfo->mp = rxq->mb_pool;
802 qinfo->scattered_rx = 1;
803 qinfo->nb_desc = rte_ring_get_capacity(rxq->rx_ring);
807 hn_nvs_handle_notify(const struct vmbus_chanpkt_hdr *pkthdr,
810 const struct hn_nvs_hdr *hdr = data;
812 if (unlikely(vmbus_chanpkt_datalen(pkthdr) < sizeof(*hdr))) {
813 PMD_DRV_LOG(ERR, "invalid nvs notify");
818 "got notify, nvs type %u", hdr->type);
822 * Process pending events on the channel.
823 * Called from both Rx queue poll and Tx cleanup
825 void hn_process_events(struct hn_data *hv, uint16_t queue_id)
827 struct rte_eth_dev *dev = &rte_eth_devices[hv->port_id];
828 struct hn_rx_queue *rxq;
829 uint32_t bytes_read = 0;
832 rxq = queue_id == 0 ? hv->primary : dev->data->rx_queues[queue_id];
834 /* If no pending data then nothing to do */
835 if (rte_vmbus_chan_rx_empty(rxq->chan))
839 * Since channel is shared between Rx and TX queue need to have a lock
840 * since DPDK does not force same CPU to be used for Rx/Tx.
842 if (unlikely(!rte_spinlock_trylock(&rxq->ring_lock)))
846 const struct vmbus_chanpkt_hdr *pkt;
847 uint32_t len = HN_RXQ_EVENT_DEFAULT;
850 ret = rte_vmbus_chan_recv_raw(rxq->chan, rxq->event_buf, &len);
852 break; /* ring is empty */
854 else if (ret == -ENOBUFS)
855 rte_exit(EXIT_FAILURE, "event buffer not big enough (%u < %u)",
856 HN_RXQ_EVENT_DEFAULT, len);
858 rte_exit(EXIT_FAILURE,
859 "vmbus ring buffer error: %d", ret);
862 pkt = (const struct vmbus_chanpkt_hdr *)rxq->event_buf;
863 data = (char *)rxq->event_buf + vmbus_chanpkt_getlen(pkt->hlen);
866 case VMBUS_CHANPKT_TYPE_COMP:
867 hn_nvs_handle_comp(dev, queue_id, pkt, data);
870 case VMBUS_CHANPKT_TYPE_RXBUF:
871 hn_nvs_handle_rxbuf(dev, hv, rxq, pkt, data);
874 case VMBUS_CHANPKT_TYPE_INBAND:
875 hn_nvs_handle_notify(pkt, data);
879 PMD_DRV_LOG(ERR, "unknown chan pkt %u", pkt->type);
883 if (rxq->rx_ring && rte_ring_full(rxq->rx_ring))
888 rte_vmbus_chan_signal_read(rxq->chan, bytes_read);
890 rte_spinlock_unlock(&rxq->ring_lock);
893 static void hn_append_to_chim(struct hn_tx_queue *txq,
894 struct rndis_packet_msg *pkt,
895 const struct rte_mbuf *m)
897 struct hn_txdesc *txd = txq->agg_txd;
898 uint8_t *buf = (uint8_t *)pkt;
899 unsigned int data_offs;
903 data_offs = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
904 txd->chim_size += pkt->len;
905 txd->data_size += m->pkt_len;
907 hn_update_packet_stats(&txq->stats, m);
909 for (; m; m = m->next) {
910 uint16_t len = rte_pktmbuf_data_len(m);
912 rte_memcpy(buf + data_offs,
913 rte_pktmbuf_mtod(m, const char *), len);
919 * Send pending aggregated data in chimney buffer (if any).
920 * Returns error if send was unsuccessful because channel ring buffer
923 static int hn_flush_txagg(struct hn_tx_queue *txq, bool *need_sig)
926 struct hn_txdesc *txd = txq->agg_txd;
927 struct hn_nvs_rndis rndis;
933 rndis = (struct hn_nvs_rndis) {
934 .type = NVS_TYPE_RNDIS,
935 .rndis_mtype = NVS_RNDIS_MTYPE_DATA,
936 .chim_idx = txd->chim_index,
937 .chim_sz = txd->chim_size,
940 PMD_TX_LOG(DEBUG, "port %u:%u tx %u size %u",
941 txq->port_id, txq->queue_id, txd->chim_index, txd->chim_size);
943 ret = hn_nvs_send(txq->chan, VMBUS_CHANPKT_FLAG_RC,
944 &rndis, sizeof(rndis), (uintptr_t)txd, need_sig);
946 if (likely(ret == 0))
949 PMD_TX_LOG(NOTICE, "port %u:%u send failed: %d",
950 txq->port_id, txq->queue_id, ret);
955 static struct hn_txdesc *hn_new_txd(struct hn_data *hv,
956 struct hn_tx_queue *txq)
958 struct hn_txdesc *txd;
960 if (rte_mempool_get(hv->tx_pool, (void **)&txd)) {
961 ++txq->stats.nomemory;
962 PMD_TX_LOG(DEBUG, "tx pool exhausted!");
967 txd->queue_id = txq->queue_id;
976 hn_try_txagg(struct hn_data *hv, struct hn_tx_queue *txq, uint32_t pktsize)
978 struct hn_txdesc *agg_txd = txq->agg_txd;
979 struct rndis_packet_msg *pkt;
983 unsigned int padding, olen;
986 * Update the previous RNDIS packet's total length,
987 * it can be increased due to the mandatory alignment
988 * padding for this RNDIS packet. And update the
989 * aggregating txdesc's chimney sending buffer size
992 * Zero-out the padding, as required by the RNDIS spec.
994 pkt = txq->agg_prevpkt;
996 padding = RTE_ALIGN(olen, txq->agg_align) - olen;
998 agg_txd->chim_size += padding;
1000 memset((uint8_t *)pkt + olen, 0, padding);
1003 chim = (uint8_t *)pkt + pkt->len;
1006 txq->agg_szleft -= pktsize;
1007 if (txq->agg_szleft < HN_PKTSIZE_MIN(txq->agg_align)) {
1009 * Probably can't aggregate more packets,
1010 * flush this aggregating txdesc proactively.
1012 txq->agg_pktleft = 0;
1015 agg_txd = hn_new_txd(hv, txq);
1019 chim = (uint8_t *)hv->chim_res->addr
1020 + agg_txd->chim_index * hv->chim_szmax;
1022 txq->agg_txd = agg_txd;
1023 txq->agg_pktleft = txq->agg_pktmax - 1;
1024 txq->agg_szleft = txq->agg_szmax - pktsize;
1026 txq->agg_prevpkt = chim;
1031 static inline void *
1032 hn_rndis_pktinfo_append(struct rndis_packet_msg *pkt,
1033 uint32_t pi_dlen, uint32_t pi_type)
1035 const uint32_t pi_size = RNDIS_PKTINFO_SIZE(pi_dlen);
1036 struct rndis_pktinfo *pi;
1039 * Per-packet-info does not move; it only grows.
1042 * pktinfooffset in this phase counts from the beginning
1043 * of rndis_packet_msg.
1045 pi = (struct rndis_pktinfo *)((uint8_t *)pkt + hn_rndis_pktlen(pkt));
1047 pkt->pktinfolen += pi_size;
1051 pi->offset = RNDIS_PKTINFO_OFFSET;
1056 /* Put RNDIS header and packet info on packet */
1057 static void hn_encap(struct rndis_packet_msg *pkt,
1059 const struct rte_mbuf *m)
1061 unsigned int hlen = m->l2_len + m->l3_len;
1065 pkt->type = RNDIS_PACKET_MSG;
1066 pkt->len = m->pkt_len;
1067 pkt->dataoffset = 0;
1068 pkt->datalen = m->pkt_len;
1069 pkt->oobdataoffset = 0;
1070 pkt->oobdatalen = 0;
1071 pkt->oobdataelements = 0;
1072 pkt->pktinfooffset = sizeof(*pkt);
1073 pkt->pktinfolen = 0;
1078 * Set the hash value for this packet, to the queue_id to cause
1079 * TX done event for this packet on the right channel.
1081 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_HASH_VALUE_SIZE,
1082 NDIS_PKTINFO_TYPE_HASHVAL);
1083 *pi_data = queue_id;
1085 if (m->ol_flags & PKT_TX_VLAN_PKT) {
1086 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_VLAN_INFO_SIZE,
1087 NDIS_PKTINFO_TYPE_VLAN);
1088 *pi_data = m->vlan_tci;
1091 if (m->ol_flags & PKT_TX_TCP_SEG) {
1092 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_LSO2_INFO_SIZE,
1093 NDIS_PKTINFO_TYPE_LSO);
1095 if (m->ol_flags & PKT_TX_IPV6) {
1096 *pi_data = NDIS_LSO2_INFO_MAKEIPV6(hlen,
1099 *pi_data = NDIS_LSO2_INFO_MAKEIPV4(hlen,
1102 } else if (m->ol_flags &
1103 (PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM | PKT_TX_IP_CKSUM)) {
1104 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_TXCSUM_INFO_SIZE,
1105 NDIS_PKTINFO_TYPE_CSUM);
1108 if (m->ol_flags & PKT_TX_IPV6)
1109 *pi_data |= NDIS_TXCSUM_INFO_IPV6;
1110 if (m->ol_flags & PKT_TX_IPV4) {
1111 *pi_data |= NDIS_TXCSUM_INFO_IPV4;
1113 if (m->ol_flags & PKT_TX_IP_CKSUM)
1114 *pi_data |= NDIS_TXCSUM_INFO_IPCS;
1117 if (m->ol_flags & PKT_TX_TCP_CKSUM)
1118 *pi_data |= NDIS_TXCSUM_INFO_MKTCPCS(hlen);
1119 else if (m->ol_flags & PKT_TX_UDP_CKSUM)
1120 *pi_data |= NDIS_TXCSUM_INFO_MKUDPCS(hlen);
1123 pkt_hlen = pkt->pktinfooffset + pkt->pktinfolen;
1124 /* Fixup RNDIS packet message total length */
1125 pkt->len += pkt_hlen;
1127 /* Convert RNDIS packet message offsets */
1128 pkt->dataoffset = hn_rndis_pktmsg_offset(pkt_hlen);
1129 pkt->pktinfooffset = hn_rndis_pktmsg_offset(pkt->pktinfooffset);
1132 /* How many scatter gather list elements ar needed */
1133 static unsigned int hn_get_slots(const struct rte_mbuf *m)
1135 unsigned int slots = 1; /* for RNDIS header */
1138 unsigned int size = rte_pktmbuf_data_len(m);
1139 unsigned int offs = rte_mbuf_data_iova(m) & PAGE_MASK;
1141 slots += (offs + size + PAGE_SIZE - 1) / PAGE_SIZE;
1148 /* Build scatter gather list from chained mbuf */
1149 static unsigned int hn_fill_sg(struct vmbus_gpa *sg,
1150 const struct rte_mbuf *m)
1152 unsigned int segs = 0;
1155 rte_iova_t addr = rte_mbuf_data_iova(m);
1156 unsigned int page = addr / PAGE_SIZE;
1157 unsigned int offset = addr & PAGE_MASK;
1158 unsigned int len = rte_pktmbuf_data_len(m);
1161 unsigned int bytes = RTE_MIN(len, PAGE_SIZE - offset);
1163 sg[segs].page = page;
1164 sg[segs].ofs = offset;
1165 sg[segs].len = bytes;
1178 /* Transmit directly from mbuf */
1179 static int hn_xmit_sg(struct hn_tx_queue *txq,
1180 const struct hn_txdesc *txd, const struct rte_mbuf *m,
1183 struct vmbus_gpa sg[hn_get_slots(m)];
1184 struct hn_nvs_rndis nvs_rndis = {
1185 .type = NVS_TYPE_RNDIS,
1186 .rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1187 .chim_sz = txd->chim_size,
1192 /* attach aggregation data if present */
1193 if (txd->chim_size > 0)
1194 nvs_rndis.chim_idx = txd->chim_index;
1196 nvs_rndis.chim_idx = NVS_CHIM_IDX_INVALID;
1198 hn_rndis_dump(txd->rndis_pkt);
1200 /* pass IOVA of rndis header in first segment */
1201 addr = rte_malloc_virt2iova(txd->rndis_pkt);
1202 if (unlikely(addr == RTE_BAD_IOVA)) {
1203 PMD_DRV_LOG(ERR, "RNDIS transmit can not get iova");
1207 sg[0].page = addr / PAGE_SIZE;
1208 sg[0].ofs = addr & PAGE_MASK;
1209 sg[0].len = RNDIS_PACKET_MSG_OFFSET_ABS(hn_rndis_pktlen(txd->rndis_pkt));
1212 hn_update_packet_stats(&txq->stats, m);
1214 segs += hn_fill_sg(sg + 1, m);
1216 PMD_TX_LOG(DEBUG, "port %u:%u tx %u segs %u size %u",
1217 txq->port_id, txq->queue_id, txd->chim_index,
1218 segs, nvs_rndis.chim_sz);
1220 return hn_nvs_send_sglist(txq->chan, sg, segs,
1221 &nvs_rndis, sizeof(nvs_rndis),
1222 (uintptr_t)txd, need_sig);
1226 hn_xmit_pkts(void *ptxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1228 struct hn_tx_queue *txq = ptxq;
1229 struct hn_data *hv = txq->hv;
1230 bool need_sig = false;
1234 if (unlikely(hv->closed))
1237 if (rte_mempool_avail_count(hv->tx_pool) <= txq->free_thresh)
1238 hn_process_events(hv, txq->queue_id);
1240 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1241 struct rte_mbuf *m = tx_pkts[nb_tx];
1242 uint32_t pkt_size = m->pkt_len + HN_RNDIS_PKT_LEN;
1243 struct rndis_packet_msg *pkt;
1245 /* For small packets aggregate them in chimney buffer */
1246 if (m->pkt_len < HN_TXCOPY_THRESHOLD && pkt_size <= txq->agg_szmax) {
1247 /* If this packet will not fit, then flush */
1248 if (txq->agg_pktleft == 0 ||
1249 RTE_ALIGN(pkt_size, txq->agg_align) > txq->agg_szleft) {
1250 if (hn_flush_txagg(txq, &need_sig))
1254 pkt = hn_try_txagg(hv, txq, pkt_size);
1258 hn_encap(pkt, txq->queue_id, m);
1259 hn_append_to_chim(txq, pkt, m);
1261 rte_pktmbuf_free(m);
1263 /* if buffer is full, flush */
1264 if (txq->agg_pktleft == 0 &&
1265 hn_flush_txagg(txq, &need_sig))
1268 struct hn_txdesc *txd;
1270 /* can send chimney data and large packet at once */
1273 hn_reset_txagg(txq);
1275 txd = hn_new_txd(hv, txq);
1280 pkt = txd->rndis_pkt;
1282 txd->data_size += m->pkt_len;
1285 hn_encap(pkt, txq->queue_id, m);
1287 ret = hn_xmit_sg(txq, txd, m, &need_sig);
1288 if (unlikely(ret != 0)) {
1289 PMD_TX_LOG(NOTICE, "sg send failed: %d", ret);
1290 ++txq->stats.errors;
1291 rte_mempool_put(hv->tx_pool, txd);
1297 /* If partial buffer left, then try and send it.
1298 * if that fails, then reuse it on next send.
1300 hn_flush_txagg(txq, &need_sig);
1304 rte_vmbus_chan_signal_tx(txq->chan);
1310 hn_recv_pkts(void *prxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1312 struct hn_rx_queue *rxq = prxq;
1313 struct hn_data *hv = rxq->hv;
1315 if (unlikely(hv->closed))
1318 /* If ring is empty then process more */
1319 if (rte_ring_count(rxq->rx_ring) < nb_pkts)
1320 hn_process_events(hv, rxq->queue_id);
1322 /* Get mbufs off staging ring */
1323 return rte_ring_sc_dequeue_burst(rxq->rx_ring, (void **)rx_pkts,