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;
221 PMD_INIT_FUNC_TRACE();
223 txq = rte_zmalloc_socket("HN_TXQ", sizeof(*txq), RTE_CACHE_LINE_SIZE,
229 txq->chan = hv->channels[queue_idx];
230 txq->port_id = dev->data->port_id;
231 txq->queue_id = queue_idx;
233 tx_free_thresh = tx_conf->tx_free_thresh;
234 if (tx_free_thresh == 0)
235 tx_free_thresh = RTE_MIN(hv->chim_cnt / 4,
236 DEFAULT_TX_FREE_THRESH);
238 if (tx_free_thresh >= hv->chim_cnt - 3)
239 tx_free_thresh = hv->chim_cnt - 3;
241 txq->free_thresh = tx_free_thresh;
243 txq->agg_szmax = RTE_MIN(hv->chim_szmax, hv->rndis_agg_size);
244 txq->agg_pktmax = hv->rndis_agg_pkts;
245 txq->agg_align = hv->rndis_agg_align;
249 dev->data->tx_queues[queue_idx] = txq;
255 hn_dev_tx_queue_release(void *arg)
257 struct hn_tx_queue *txq = arg;
258 struct hn_txdesc *txd;
260 PMD_INIT_FUNC_TRACE();
265 /* If any pending data is still present just drop it */
268 rte_mempool_put(txq->hv->tx_pool, txd);
274 hn_dev_tx_queue_info(struct rte_eth_dev *dev, uint16_t queue_idx,
275 struct rte_eth_txq_info *qinfo)
277 struct hn_data *hv = dev->data->dev_private;
278 struct hn_tx_queue *txq = dev->data->rx_queues[queue_idx];
280 qinfo->conf.tx_free_thresh = txq->free_thresh;
281 qinfo->nb_desc = hv->tx_pool->size;
285 hn_nvs_send_completed(struct rte_eth_dev *dev, uint16_t queue_id,
286 unsigned long xactid, const struct hn_nvs_rndis_ack *ack)
288 struct hn_txdesc *txd = (struct hn_txdesc *)xactid;
289 struct hn_tx_queue *txq;
291 /* Control packets are sent with xacid == 0 */
295 txq = dev->data->tx_queues[queue_id];
296 if (likely(ack->status == NVS_STATUS_OK)) {
297 PMD_TX_LOG(DEBUG, "port %u:%u complete tx %u packets %u bytes %u",
298 txq->port_id, txq->queue_id, txd->chim_index,
299 txd->packets, txd->data_size);
300 txq->stats.bytes += txd->data_size;
301 txq->stats.packets += txd->packets;
303 PMD_TX_LOG(NOTICE, "port %u:%u complete tx %u failed status %u",
304 txq->port_id, txq->queue_id, txd->chim_index, ack->status);
308 rte_pktmbuf_free(txd->m);
310 rte_mempool_put(txq->hv->tx_pool, txd);
313 /* Handle transmit completion events */
315 hn_nvs_handle_comp(struct rte_eth_dev *dev, uint16_t queue_id,
316 const struct vmbus_chanpkt_hdr *pkt,
319 const struct hn_nvs_hdr *hdr = data;
322 case NVS_TYPE_RNDIS_ACK:
323 hn_nvs_send_completed(dev, queue_id, pkt->xactid, data);
328 "unexpected send completion type %u",
333 /* Parse per-packet info (meta data) */
335 hn_rndis_rxinfo(const void *info_data, unsigned int info_dlen,
336 struct hn_rxinfo *info)
338 const struct rndis_pktinfo *pi = info_data;
341 while (info_dlen != 0) {
345 if (unlikely(info_dlen < sizeof(*pi)))
348 if (unlikely(info_dlen < pi->size))
350 info_dlen -= pi->size;
352 if (unlikely(pi->size & RNDIS_PKTINFO_SIZE_ALIGNMASK))
354 if (unlikely(pi->size < pi->offset))
357 dlen = pi->size - pi->offset;
361 case NDIS_PKTINFO_TYPE_VLAN:
362 if (unlikely(dlen < NDIS_VLAN_INFO_SIZE))
364 info->vlan_info = *((const uint32_t *)data);
365 mask |= HN_RXINFO_VLAN;
368 case NDIS_PKTINFO_TYPE_CSUM:
369 if (unlikely(dlen < NDIS_RXCSUM_INFO_SIZE))
371 info->csum_info = *((const uint32_t *)data);
372 mask |= HN_RXINFO_CSUM;
375 case NDIS_PKTINFO_TYPE_HASHVAL:
376 if (unlikely(dlen < NDIS_HASH_VALUE_SIZE))
378 info->hash_value = *((const uint32_t *)data);
379 mask |= HN_RXINFO_HASHVAL;
382 case NDIS_PKTINFO_TYPE_HASHINF:
383 if (unlikely(dlen < NDIS_HASH_INFO_SIZE))
385 info->hash_info = *((const uint32_t *)data);
386 mask |= HN_RXINFO_HASHINF;
393 if (mask == HN_RXINFO_ALL)
394 break; /* All found; done */
396 pi = (const struct rndis_pktinfo *)
397 ((const uint8_t *)pi + pi->size);
402 * - If there is no hash value, invalidate the hash info.
404 if (!(mask & HN_RXINFO_HASHVAL))
405 info->hash_info = HN_NDIS_HASH_INFO_INVALID;
410 * Ack the consumed RXBUF associated w/ this channel packet,
411 * so that this RXBUF can be recycled by the hypervisor.
413 static void hn_rx_buf_release(struct hn_rx_bufinfo *rxb)
415 struct rte_mbuf_ext_shared_info *shinfo = &rxb->shinfo;
416 struct hn_data *hv = rxb->hv;
418 if (rte_mbuf_ext_refcnt_update(shinfo, -1) == 0) {
419 hn_nvs_ack_rxbuf(rxb->chan, rxb->xactid);
420 --hv->rxbuf_outstanding;
424 static void hn_rx_buf_free_cb(void *buf __rte_unused, void *opaque)
426 hn_rx_buf_release(opaque);
429 static struct hn_rx_bufinfo *hn_rx_buf_init(const struct hn_rx_queue *rxq,
430 const struct vmbus_chanpkt_rxbuf *pkt)
432 struct hn_rx_bufinfo *rxb;
434 rxb = rxq->hv->rxbuf_info + pkt->hdr.xactid;
435 rxb->chan = rxq->chan;
436 rxb->xactid = pkt->hdr.xactid;
439 rxb->shinfo.free_cb = hn_rx_buf_free_cb;
440 rxb->shinfo.fcb_opaque = rxb;
441 rte_mbuf_ext_refcnt_set(&rxb->shinfo, 1);
445 static void hn_rxpkt(struct hn_rx_queue *rxq, struct hn_rx_bufinfo *rxb,
446 uint8_t *data, unsigned int headroom, unsigned int dlen,
447 const struct hn_rxinfo *info)
449 struct hn_data *hv = rxq->hv;
452 m = rte_pktmbuf_alloc(rxq->mb_pool);
454 struct rte_eth_dev *dev =
455 &rte_eth_devices[rxq->port_id];
457 dev->data->rx_mbuf_alloc_failed++;
462 * For large packets, avoid copy if possible but need to keep
463 * some space available in receive area for later packets.
465 if (dlen >= HN_RXCOPY_THRESHOLD &&
466 hv->rxbuf_outstanding < hv->rxbuf_section_cnt / 2) {
467 struct rte_mbuf_ext_shared_info *shinfo;
472 * Build an external mbuf that points to recveive area.
473 * Use refcount to handle multiple packets in same
474 * receive buffer section.
476 rxbuf = hv->rxbuf_res->addr;
477 iova = rte_mem_virt2iova(rxbuf) + RTE_PTR_DIFF(data, rxbuf);
478 shinfo = &rxb->shinfo;
480 if (rte_mbuf_ext_refcnt_update(shinfo, 1) == 1)
481 ++hv->rxbuf_outstanding;
483 rte_pktmbuf_attach_extbuf(m, data, iova,
484 dlen + headroom, shinfo);
485 m->data_off = headroom;
487 /* Mbuf's in pool must be large enough to hold small packets */
488 if (unlikely(rte_pktmbuf_tailroom(m) < dlen)) {
489 rte_pktmbuf_free_seg(m);
493 rte_memcpy(rte_pktmbuf_mtod(m, void *),
494 data + headroom, dlen);
497 m->port = rxq->port_id;
500 m->packet_type = rte_net_get_ptype(m, NULL,
505 if (info->vlan_info != HN_NDIS_VLAN_INFO_INVALID) {
506 m->vlan_tci = info->vlan_info;
507 m->ol_flags |= PKT_RX_VLAN_STRIPPED | PKT_RX_VLAN;
510 if (info->csum_info != HN_NDIS_RXCSUM_INFO_INVALID) {
511 if (info->csum_info & NDIS_RXCSUM_INFO_IPCS_OK)
512 m->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
514 if (info->csum_info & (NDIS_RXCSUM_INFO_UDPCS_OK
515 | NDIS_RXCSUM_INFO_TCPCS_OK))
516 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
517 else if (info->csum_info & (NDIS_RXCSUM_INFO_TCPCS_FAILED
518 | NDIS_RXCSUM_INFO_UDPCS_FAILED))
519 m->ol_flags |= PKT_RX_L4_CKSUM_BAD;
522 if (info->hash_info != HN_NDIS_HASH_INFO_INVALID) {
523 m->ol_flags |= PKT_RX_RSS_HASH;
524 m->hash.rss = info->hash_value;
528 "port %u:%u RX id %"PRIu64" size %u type %#x ol_flags %#"PRIx64,
529 rxq->port_id, rxq->queue_id, rxb->xactid,
530 m->pkt_len, m->packet_type, m->ol_flags);
532 ++rxq->stats.packets;
533 rxq->stats.bytes += m->pkt_len;
534 hn_update_packet_stats(&rxq->stats, m);
536 if (unlikely(rte_ring_sp_enqueue(rxq->rx_ring, m) != 0)) {
537 ++rxq->stats.ring_full;
542 static void hn_rndis_rx_data(struct hn_rx_queue *rxq,
543 struct hn_rx_bufinfo *rxb,
544 void *data, uint32_t dlen)
546 unsigned int data_off, data_len, pktinfo_off, pktinfo_len;
547 const struct rndis_packet_msg *pkt = data;
548 struct hn_rxinfo info = {
549 .vlan_info = HN_NDIS_VLAN_INFO_INVALID,
550 .csum_info = HN_NDIS_RXCSUM_INFO_INVALID,
551 .hash_info = HN_NDIS_HASH_INFO_INVALID,
557 if (unlikely(dlen < sizeof(*pkt)))
560 if (unlikely(dlen < pkt->len))
561 goto error; /* truncated RNDIS from host */
563 if (unlikely(pkt->len < pkt->datalen
564 + pkt->oobdatalen + pkt->pktinfolen))
567 if (unlikely(pkt->datalen == 0))
571 if (unlikely(pkt->dataoffset < RNDIS_PACKET_MSG_OFFSET_MIN))
574 if (likely(pkt->pktinfooffset > 0) &&
575 unlikely(pkt->pktinfooffset < RNDIS_PACKET_MSG_OFFSET_MIN ||
576 (pkt->pktinfooffset & RNDIS_PACKET_MSG_OFFSET_ALIGNMASK)))
579 data_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
580 data_len = pkt->datalen;
581 pktinfo_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->pktinfooffset);
582 pktinfo_len = pkt->pktinfolen;
584 if (likely(pktinfo_len > 0)) {
585 err = hn_rndis_rxinfo((const uint8_t *)pkt + pktinfo_off,
591 if (unlikely(data_off + data_len > pkt->len))
594 if (unlikely(data_len < ETHER_HDR_LEN))
597 hn_rxpkt(rxq, rxb, data, data_off, data_len, &info);
604 hn_rndis_receive(const struct rte_eth_dev *dev, struct hn_rx_queue *rxq,
605 struct hn_rx_bufinfo *rxb, void *buf, uint32_t len)
607 const struct rndis_msghdr *hdr = buf;
610 case RNDIS_PACKET_MSG:
611 if (dev->data->dev_started)
612 hn_rndis_rx_data(rxq, rxb, buf, len);
615 case RNDIS_INDICATE_STATUS_MSG:
616 hn_rndis_link_status(rxq->hv, buf);
619 case RNDIS_INITIALIZE_CMPLT:
620 case RNDIS_QUERY_CMPLT:
621 case RNDIS_SET_CMPLT:
622 hn_rndis_receive_response(rxq->hv, buf, len);
627 "unexpected RNDIS message (type %#x len %u)",
634 hn_nvs_handle_rxbuf(struct rte_eth_dev *dev,
636 struct hn_rx_queue *rxq,
637 const struct vmbus_chanpkt_hdr *hdr,
640 const struct vmbus_chanpkt_rxbuf *pkt;
641 const struct hn_nvs_hdr *nvs_hdr = buf;
642 uint32_t rxbuf_sz = hv->rxbuf_res->len;
643 char *rxbuf = hv->rxbuf_res->addr;
644 unsigned int i, hlen, count;
645 struct hn_rx_bufinfo *rxb;
647 /* At minimum we need type header */
648 if (unlikely(vmbus_chanpkt_datalen(hdr) < sizeof(*nvs_hdr))) {
649 PMD_RX_LOG(ERR, "invalid receive nvs RNDIS");
653 /* Make sure that this is a RNDIS message. */
654 if (unlikely(nvs_hdr->type != NVS_TYPE_RNDIS)) {
655 PMD_RX_LOG(ERR, "nvs type %u, not RNDIS",
660 hlen = vmbus_chanpkt_getlen(hdr->hlen);
661 if (unlikely(hlen < sizeof(*pkt))) {
662 PMD_RX_LOG(ERR, "invalid rxbuf chanpkt");
666 pkt = container_of(hdr, const struct vmbus_chanpkt_rxbuf, hdr);
667 if (unlikely(pkt->rxbuf_id != NVS_RXBUF_SIG)) {
668 PMD_RX_LOG(ERR, "invalid rxbuf_id 0x%08x",
673 count = pkt->rxbuf_cnt;
674 if (unlikely(hlen < offsetof(struct vmbus_chanpkt_rxbuf,
676 PMD_RX_LOG(ERR, "invalid rxbuf_cnt %u", count);
680 if (pkt->hdr.xactid > hv->rxbuf_section_cnt) {
681 PMD_RX_LOG(ERR, "invalid rxbuf section id %" PRIx64,
686 /* Setup receive buffer info to allow for callback */
687 rxb = hn_rx_buf_init(rxq, pkt);
689 /* Each range represents 1 RNDIS pkt that contains 1 Ethernet frame */
690 for (i = 0; i < count; ++i) {
691 unsigned int ofs, len;
693 ofs = pkt->rxbuf[i].ofs;
694 len = pkt->rxbuf[i].len;
696 if (unlikely(ofs + len > rxbuf_sz)) {
698 "%uth RNDIS msg overflow ofs %u, len %u",
703 if (unlikely(len == 0)) {
704 PMD_RX_LOG(ERR, "%uth RNDIS msg len %u", i, len);
708 hn_rndis_receive(dev, rxq, rxb,
712 /* Send ACK now if external mbuf not used */
713 hn_rx_buf_release(rxb);
716 struct hn_rx_queue *hn_rx_queue_alloc(struct hn_data *hv,
718 unsigned int socket_id)
720 struct hn_rx_queue *rxq;
722 rxq = rte_zmalloc_socket("HN_RXQ", sizeof(*rxq),
723 RTE_CACHE_LINE_SIZE, socket_id);
728 rxq->chan = hv->channels[queue_id];
729 rte_spinlock_init(&rxq->ring_lock);
730 rxq->port_id = hv->port_id;
731 rxq->queue_id = queue_id;
732 rxq->event_sz = HN_RXQ_EVENT_DEFAULT;
733 rxq->event_buf = rte_malloc_socket("HN_EVENTS", HN_RXQ_EVENT_DEFAULT,
734 RTE_CACHE_LINE_SIZE, socket_id);
735 if (!rxq->event_buf) {
744 hn_dev_rx_queue_setup(struct rte_eth_dev *dev,
745 uint16_t queue_idx, uint16_t nb_desc,
746 unsigned int socket_id,
747 const struct rte_eth_rxconf *rx_conf __rte_unused,
748 struct rte_mempool *mp)
750 struct hn_data *hv = dev->data->dev_private;
751 char ring_name[RTE_RING_NAMESIZE];
752 struct hn_rx_queue *rxq;
755 PMD_INIT_FUNC_TRACE();
757 if (queue_idx == 0) {
760 rxq = hn_rx_queue_alloc(hv, queue_idx, socket_id);
766 count = rte_mempool_avail_count(mp) / dev->data->nb_rx_queues;
767 if (nb_desc == 0 || nb_desc > count)
771 * Staging ring from receive event logic to rx_pkts.
772 * rx_pkts assumes caller is handling multi-thread issue.
773 * event logic has locking.
775 snprintf(ring_name, sizeof(ring_name),
776 "hn_rx_%u_%u", dev->data->port_id, queue_idx);
777 rxq->rx_ring = rte_ring_create(ring_name,
778 rte_align32pow2(nb_desc),
783 dev->data->rx_queues[queue_idx] = rxq;
787 rte_ring_free(rxq->rx_ring);
788 rte_free(rxq->event_buf);
794 hn_dev_rx_queue_release(void *arg)
796 struct hn_rx_queue *rxq = arg;
798 PMD_INIT_FUNC_TRACE();
803 rte_ring_free(rxq->rx_ring);
807 if (rxq != rxq->hv->primary) {
808 rte_free(rxq->event_buf);
814 hn_dev_tx_done_cleanup(void *arg, uint32_t free_cnt)
816 struct hn_tx_queue *txq = arg;
818 return hn_process_events(txq->hv, txq->queue_id, free_cnt);
822 hn_dev_rx_queue_info(struct rte_eth_dev *dev, uint16_t queue_idx,
823 struct rte_eth_rxq_info *qinfo)
825 struct hn_rx_queue *rxq = dev->data->rx_queues[queue_idx];
827 qinfo->mp = rxq->mb_pool;
828 qinfo->scattered_rx = 1;
829 qinfo->nb_desc = rte_ring_get_capacity(rxq->rx_ring);
833 hn_nvs_handle_notify(const struct vmbus_chanpkt_hdr *pkthdr,
836 const struct hn_nvs_hdr *hdr = data;
838 if (unlikely(vmbus_chanpkt_datalen(pkthdr) < sizeof(*hdr))) {
839 PMD_DRV_LOG(ERR, "invalid nvs notify");
844 "got notify, nvs type %u", hdr->type);
848 * Process pending events on the channel.
849 * Called from both Rx queue poll and Tx cleanup
851 uint32_t hn_process_events(struct hn_data *hv, uint16_t queue_id,
854 struct rte_eth_dev *dev = &rte_eth_devices[hv->port_id];
855 struct hn_rx_queue *rxq;
856 uint32_t bytes_read = 0;
857 uint32_t tx_done = 0;
860 rxq = queue_id == 0 ? hv->primary : dev->data->rx_queues[queue_id];
862 /* If no pending data then nothing to do */
863 if (rte_vmbus_chan_rx_empty(rxq->chan))
867 * Since channel is shared between Rx and TX queue need to have a lock
868 * since DPDK does not force same CPU to be used for Rx/Tx.
870 if (unlikely(!rte_spinlock_trylock(&rxq->ring_lock)))
874 const struct vmbus_chanpkt_hdr *pkt;
875 uint32_t len = rxq->event_sz;
879 ret = rte_vmbus_chan_recv_raw(rxq->chan, rxq->event_buf, &len);
881 break; /* ring is empty */
883 if (unlikely(ret == -ENOBUFS)) {
884 /* event buffer not large enough to read ring */
887 "event buffer expansion (need %u)", len);
888 rxq->event_sz = len + len / 4;
889 rxq->event_buf = rte_realloc(rxq->event_buf, rxq->event_sz,
890 RTE_CACHE_LINE_SIZE);
893 /* out of memory, no more events now */
898 if (unlikely(ret <= 0)) {
899 /* This indicates a failure to communicate (or worse) */
900 rte_exit(EXIT_FAILURE,
901 "vmbus ring buffer error: %d", ret);
905 pkt = (const struct vmbus_chanpkt_hdr *)rxq->event_buf;
906 data = (char *)rxq->event_buf + vmbus_chanpkt_getlen(pkt->hlen);
909 case VMBUS_CHANPKT_TYPE_COMP:
911 hn_nvs_handle_comp(dev, queue_id, pkt, data);
914 case VMBUS_CHANPKT_TYPE_RXBUF:
915 hn_nvs_handle_rxbuf(dev, hv, rxq, pkt, data);
918 case VMBUS_CHANPKT_TYPE_INBAND:
919 hn_nvs_handle_notify(pkt, data);
923 PMD_DRV_LOG(ERR, "unknown chan pkt %u", pkt->type);
927 if (tx_limit && tx_done >= tx_limit)
930 if (rxq->rx_ring && rte_ring_full(rxq->rx_ring))
935 rte_vmbus_chan_signal_read(rxq->chan, bytes_read);
937 rte_spinlock_unlock(&rxq->ring_lock);
942 static void hn_append_to_chim(struct hn_tx_queue *txq,
943 struct rndis_packet_msg *pkt,
944 const struct rte_mbuf *m)
946 struct hn_txdesc *txd = txq->agg_txd;
947 uint8_t *buf = (uint8_t *)pkt;
948 unsigned int data_offs;
952 data_offs = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
953 txd->chim_size += pkt->len;
954 txd->data_size += m->pkt_len;
956 hn_update_packet_stats(&txq->stats, m);
958 for (; m; m = m->next) {
959 uint16_t len = rte_pktmbuf_data_len(m);
961 rte_memcpy(buf + data_offs,
962 rte_pktmbuf_mtod(m, const char *), len);
968 * Send pending aggregated data in chimney buffer (if any).
969 * Returns error if send was unsuccessful because channel ring buffer
972 static int hn_flush_txagg(struct hn_tx_queue *txq, bool *need_sig)
975 struct hn_txdesc *txd = txq->agg_txd;
976 struct hn_nvs_rndis rndis;
982 rndis = (struct hn_nvs_rndis) {
983 .type = NVS_TYPE_RNDIS,
984 .rndis_mtype = NVS_RNDIS_MTYPE_DATA,
985 .chim_idx = txd->chim_index,
986 .chim_sz = txd->chim_size,
989 PMD_TX_LOG(DEBUG, "port %u:%u tx %u size %u",
990 txq->port_id, txq->queue_id, txd->chim_index, txd->chim_size);
992 ret = hn_nvs_send(txq->chan, VMBUS_CHANPKT_FLAG_RC,
993 &rndis, sizeof(rndis), (uintptr_t)txd, need_sig);
995 if (likely(ret == 0))
998 PMD_TX_LOG(NOTICE, "port %u:%u send failed: %d",
999 txq->port_id, txq->queue_id, ret);
1004 static struct hn_txdesc *hn_new_txd(struct hn_data *hv,
1005 struct hn_tx_queue *txq)
1007 struct hn_txdesc *txd;
1009 if (rte_mempool_get(hv->tx_pool, (void **)&txd)) {
1010 ++txq->stats.ring_full;
1011 PMD_TX_LOG(DEBUG, "tx pool exhausted!");
1016 txd->queue_id = txq->queue_id;
1025 hn_try_txagg(struct hn_data *hv, struct hn_tx_queue *txq, uint32_t pktsize)
1027 struct hn_txdesc *agg_txd = txq->agg_txd;
1028 struct rndis_packet_msg *pkt;
1032 unsigned int padding, olen;
1035 * Update the previous RNDIS packet's total length,
1036 * it can be increased due to the mandatory alignment
1037 * padding for this RNDIS packet. And update the
1038 * aggregating txdesc's chimney sending buffer size
1041 * Zero-out the padding, as required by the RNDIS spec.
1043 pkt = txq->agg_prevpkt;
1045 padding = RTE_ALIGN(olen, txq->agg_align) - olen;
1047 agg_txd->chim_size += padding;
1048 pkt->len += padding;
1049 memset((uint8_t *)pkt + olen, 0, padding);
1052 chim = (uint8_t *)pkt + pkt->len;
1055 txq->agg_szleft -= pktsize;
1056 if (txq->agg_szleft < HN_PKTSIZE_MIN(txq->agg_align)) {
1058 * Probably can't aggregate more packets,
1059 * flush this aggregating txdesc proactively.
1061 txq->agg_pktleft = 0;
1064 agg_txd = hn_new_txd(hv, txq);
1068 chim = (uint8_t *)hv->chim_res->addr
1069 + agg_txd->chim_index * hv->chim_szmax;
1071 txq->agg_txd = agg_txd;
1072 txq->agg_pktleft = txq->agg_pktmax - 1;
1073 txq->agg_szleft = txq->agg_szmax - pktsize;
1075 txq->agg_prevpkt = chim;
1080 static inline void *
1081 hn_rndis_pktinfo_append(struct rndis_packet_msg *pkt,
1082 uint32_t pi_dlen, uint32_t pi_type)
1084 const uint32_t pi_size = RNDIS_PKTINFO_SIZE(pi_dlen);
1085 struct rndis_pktinfo *pi;
1088 * Per-packet-info does not move; it only grows.
1091 * pktinfooffset in this phase counts from the beginning
1092 * of rndis_packet_msg.
1094 pi = (struct rndis_pktinfo *)((uint8_t *)pkt + hn_rndis_pktlen(pkt));
1096 pkt->pktinfolen += pi_size;
1100 pi->offset = RNDIS_PKTINFO_OFFSET;
1105 /* Put RNDIS header and packet info on packet */
1106 static void hn_encap(struct rndis_packet_msg *pkt,
1108 const struct rte_mbuf *m)
1110 unsigned int hlen = m->l2_len + m->l3_len;
1114 pkt->type = RNDIS_PACKET_MSG;
1115 pkt->len = m->pkt_len;
1116 pkt->dataoffset = 0;
1117 pkt->datalen = m->pkt_len;
1118 pkt->oobdataoffset = 0;
1119 pkt->oobdatalen = 0;
1120 pkt->oobdataelements = 0;
1121 pkt->pktinfooffset = sizeof(*pkt);
1122 pkt->pktinfolen = 0;
1127 * Set the hash value for this packet, to the queue_id to cause
1128 * TX done event for this packet on the right channel.
1130 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_HASH_VALUE_SIZE,
1131 NDIS_PKTINFO_TYPE_HASHVAL);
1132 *pi_data = queue_id;
1134 if (m->ol_flags & PKT_TX_VLAN_PKT) {
1135 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_VLAN_INFO_SIZE,
1136 NDIS_PKTINFO_TYPE_VLAN);
1137 *pi_data = m->vlan_tci;
1140 if (m->ol_flags & PKT_TX_TCP_SEG) {
1141 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_LSO2_INFO_SIZE,
1142 NDIS_PKTINFO_TYPE_LSO);
1144 if (m->ol_flags & PKT_TX_IPV6) {
1145 *pi_data = NDIS_LSO2_INFO_MAKEIPV6(hlen,
1148 *pi_data = NDIS_LSO2_INFO_MAKEIPV4(hlen,
1151 } else if (m->ol_flags &
1152 (PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM | PKT_TX_IP_CKSUM)) {
1153 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_TXCSUM_INFO_SIZE,
1154 NDIS_PKTINFO_TYPE_CSUM);
1157 if (m->ol_flags & PKT_TX_IPV6)
1158 *pi_data |= NDIS_TXCSUM_INFO_IPV6;
1159 if (m->ol_flags & PKT_TX_IPV4) {
1160 *pi_data |= NDIS_TXCSUM_INFO_IPV4;
1162 if (m->ol_flags & PKT_TX_IP_CKSUM)
1163 *pi_data |= NDIS_TXCSUM_INFO_IPCS;
1166 if (m->ol_flags & PKT_TX_TCP_CKSUM)
1167 *pi_data |= NDIS_TXCSUM_INFO_MKTCPCS(hlen);
1168 else if (m->ol_flags & PKT_TX_UDP_CKSUM)
1169 *pi_data |= NDIS_TXCSUM_INFO_MKUDPCS(hlen);
1172 pkt_hlen = pkt->pktinfooffset + pkt->pktinfolen;
1173 /* Fixup RNDIS packet message total length */
1174 pkt->len += pkt_hlen;
1176 /* Convert RNDIS packet message offsets */
1177 pkt->dataoffset = hn_rndis_pktmsg_offset(pkt_hlen);
1178 pkt->pktinfooffset = hn_rndis_pktmsg_offset(pkt->pktinfooffset);
1181 /* How many scatter gather list elements ar needed */
1182 static unsigned int hn_get_slots(const struct rte_mbuf *m)
1184 unsigned int slots = 1; /* for RNDIS header */
1187 unsigned int size = rte_pktmbuf_data_len(m);
1188 unsigned int offs = rte_mbuf_data_iova(m) & PAGE_MASK;
1190 slots += (offs + size + PAGE_SIZE - 1) / PAGE_SIZE;
1197 /* Build scatter gather list from chained mbuf */
1198 static unsigned int hn_fill_sg(struct vmbus_gpa *sg,
1199 const struct rte_mbuf *m)
1201 unsigned int segs = 0;
1204 rte_iova_t addr = rte_mbuf_data_iova(m);
1205 unsigned int page = addr / PAGE_SIZE;
1206 unsigned int offset = addr & PAGE_MASK;
1207 unsigned int len = rte_pktmbuf_data_len(m);
1210 unsigned int bytes = RTE_MIN(len, PAGE_SIZE - offset);
1212 sg[segs].page = page;
1213 sg[segs].ofs = offset;
1214 sg[segs].len = bytes;
1227 /* Transmit directly from mbuf */
1228 static int hn_xmit_sg(struct hn_tx_queue *txq,
1229 const struct hn_txdesc *txd, const struct rte_mbuf *m,
1232 struct vmbus_gpa sg[hn_get_slots(m)];
1233 struct hn_nvs_rndis nvs_rndis = {
1234 .type = NVS_TYPE_RNDIS,
1235 .rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1236 .chim_sz = txd->chim_size,
1241 /* attach aggregation data if present */
1242 if (txd->chim_size > 0)
1243 nvs_rndis.chim_idx = txd->chim_index;
1245 nvs_rndis.chim_idx = NVS_CHIM_IDX_INVALID;
1247 hn_rndis_dump(txd->rndis_pkt);
1249 /* pass IOVA of rndis header in first segment */
1250 addr = rte_malloc_virt2iova(txd->rndis_pkt);
1251 if (unlikely(addr == RTE_BAD_IOVA)) {
1252 PMD_DRV_LOG(ERR, "RNDIS transmit can not get iova");
1256 sg[0].page = addr / PAGE_SIZE;
1257 sg[0].ofs = addr & PAGE_MASK;
1258 sg[0].len = RNDIS_PACKET_MSG_OFFSET_ABS(hn_rndis_pktlen(txd->rndis_pkt));
1261 hn_update_packet_stats(&txq->stats, m);
1263 segs += hn_fill_sg(sg + 1, m);
1265 PMD_TX_LOG(DEBUG, "port %u:%u tx %u segs %u size %u",
1266 txq->port_id, txq->queue_id, txd->chim_index,
1267 segs, nvs_rndis.chim_sz);
1269 return hn_nvs_send_sglist(txq->chan, sg, segs,
1270 &nvs_rndis, sizeof(nvs_rndis),
1271 (uintptr_t)txd, need_sig);
1275 hn_xmit_pkts(void *ptxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1277 struct hn_tx_queue *txq = ptxq;
1278 struct hn_data *hv = txq->hv;
1279 bool need_sig = false;
1283 if (unlikely(hv->closed))
1286 if (rte_mempool_avail_count(hv->tx_pool) <= txq->free_thresh)
1287 hn_process_events(hv, txq->queue_id, 0);
1289 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1290 struct rte_mbuf *m = tx_pkts[nb_tx];
1291 uint32_t pkt_size = m->pkt_len + HN_RNDIS_PKT_LEN;
1292 struct rndis_packet_msg *pkt;
1294 /* For small packets aggregate them in chimney buffer */
1295 if (m->pkt_len < HN_TXCOPY_THRESHOLD && pkt_size <= txq->agg_szmax) {
1296 /* If this packet will not fit, then flush */
1297 if (txq->agg_pktleft == 0 ||
1298 RTE_ALIGN(pkt_size, txq->agg_align) > txq->agg_szleft) {
1299 if (hn_flush_txagg(txq, &need_sig))
1303 pkt = hn_try_txagg(hv, txq, pkt_size);
1307 hn_encap(pkt, txq->queue_id, m);
1308 hn_append_to_chim(txq, pkt, m);
1310 rte_pktmbuf_free(m);
1312 /* if buffer is full, flush */
1313 if (txq->agg_pktleft == 0 &&
1314 hn_flush_txagg(txq, &need_sig))
1317 struct hn_txdesc *txd;
1319 /* can send chimney data and large packet at once */
1322 hn_reset_txagg(txq);
1324 txd = hn_new_txd(hv, txq);
1329 pkt = txd->rndis_pkt;
1331 txd->data_size += m->pkt_len;
1334 hn_encap(pkt, txq->queue_id, m);
1336 ret = hn_xmit_sg(txq, txd, m, &need_sig);
1337 if (unlikely(ret != 0)) {
1338 PMD_TX_LOG(NOTICE, "sg send failed: %d", ret);
1339 ++txq->stats.errors;
1340 rte_mempool_put(hv->tx_pool, txd);
1346 /* If partial buffer left, then try and send it.
1347 * if that fails, then reuse it on next send.
1349 hn_flush_txagg(txq, &need_sig);
1353 rte_vmbus_chan_signal_tx(txq->chan);
1359 hn_recv_pkts(void *prxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1361 struct hn_rx_queue *rxq = prxq;
1362 struct hn_data *hv = rxq->hv;
1364 if (unlikely(hv->closed))
1367 /* If ring is empty then process more */
1368 if (rte_ring_count(rxq->rx_ring) < nb_pkts)
1369 hn_process_events(hv, rxq->queue_id, 0);
1371 /* Get mbufs off staging ring */
1372 return rte_ring_sc_dequeue_burst(rxq->rx_ring, (void **)rx_pkts,