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;
506 if (info->csum_info != HN_NDIS_RXCSUM_INFO_INVALID) {
507 if (info->csum_info & NDIS_RXCSUM_INFO_IPCS_OK)
508 m->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
510 if (info->csum_info & (NDIS_RXCSUM_INFO_UDPCS_OK
511 | NDIS_RXCSUM_INFO_TCPCS_OK))
512 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
513 else if (info->csum_info & (NDIS_RXCSUM_INFO_TCPCS_FAILED
514 | NDIS_RXCSUM_INFO_UDPCS_FAILED))
515 m->ol_flags |= PKT_RX_L4_CKSUM_BAD;
518 if (info->hash_info != HN_NDIS_HASH_INFO_INVALID) {
519 m->ol_flags |= PKT_RX_RSS_HASH;
520 m->hash.rss = info->hash_value;
524 "port %u:%u RX id %"PRIu64" size %u type %#x ol_flags %#"PRIx64,
525 rxq->port_id, rxq->queue_id, rxb->xactid,
526 m->pkt_len, m->packet_type, m->ol_flags);
528 ++rxq->stats.packets;
529 rxq->stats.bytes += m->pkt_len;
530 hn_update_packet_stats(&rxq->stats, m);
532 if (unlikely(rte_ring_sp_enqueue(rxq->rx_ring, m) != 0)) {
533 ++rxq->stats.ring_full;
538 static void hn_rndis_rx_data(struct hn_rx_queue *rxq,
539 struct hn_rx_bufinfo *rxb,
540 void *data, uint32_t dlen)
542 unsigned int data_off, data_len, pktinfo_off, pktinfo_len;
543 const struct rndis_packet_msg *pkt = data;
544 struct hn_rxinfo info = {
545 .vlan_info = HN_NDIS_VLAN_INFO_INVALID,
546 .csum_info = HN_NDIS_RXCSUM_INFO_INVALID,
547 .hash_info = HN_NDIS_HASH_INFO_INVALID,
553 if (unlikely(dlen < sizeof(*pkt)))
556 if (unlikely(dlen < pkt->len))
557 goto error; /* truncated RNDIS from host */
559 if (unlikely(pkt->len < pkt->datalen
560 + pkt->oobdatalen + pkt->pktinfolen))
563 if (unlikely(pkt->datalen == 0))
567 if (unlikely(pkt->dataoffset < RNDIS_PACKET_MSG_OFFSET_MIN))
570 if (likely(pkt->pktinfooffset > 0) &&
571 unlikely(pkt->pktinfooffset < RNDIS_PACKET_MSG_OFFSET_MIN ||
572 (pkt->pktinfooffset & RNDIS_PACKET_MSG_OFFSET_ALIGNMASK)))
575 data_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
576 data_len = pkt->datalen;
577 pktinfo_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->pktinfooffset);
578 pktinfo_len = pkt->pktinfolen;
580 if (likely(pktinfo_len > 0)) {
581 err = hn_rndis_rxinfo((const uint8_t *)pkt + pktinfo_off,
587 if (unlikely(data_off + data_len > pkt->len))
590 if (unlikely(data_len < ETHER_HDR_LEN))
593 hn_rxpkt(rxq, rxb, data, data_off, data_len, &info);
600 hn_rndis_receive(struct rte_eth_dev *dev, struct hn_rx_queue *rxq,
601 struct hn_rx_bufinfo *rxb, void *buf, uint32_t len)
603 const struct rndis_msghdr *hdr = buf;
606 case RNDIS_PACKET_MSG:
607 if (dev->data->dev_started)
608 hn_rndis_rx_data(rxq, rxb, buf, len);
611 case RNDIS_INDICATE_STATUS_MSG:
612 hn_rndis_link_status(dev, buf);
615 case RNDIS_INITIALIZE_CMPLT:
616 case RNDIS_QUERY_CMPLT:
617 case RNDIS_SET_CMPLT:
618 hn_rndis_receive_response(rxq->hv, buf, len);
623 "unexpected RNDIS message (type %#x len %u)",
630 hn_nvs_handle_rxbuf(struct rte_eth_dev *dev,
632 struct hn_rx_queue *rxq,
633 const struct vmbus_chanpkt_hdr *hdr,
636 const struct vmbus_chanpkt_rxbuf *pkt;
637 const struct hn_nvs_hdr *nvs_hdr = buf;
638 uint32_t rxbuf_sz = hv->rxbuf_res->len;
639 char *rxbuf = hv->rxbuf_res->addr;
640 unsigned int i, hlen, count;
641 struct hn_rx_bufinfo *rxb;
643 /* At minimum we need type header */
644 if (unlikely(vmbus_chanpkt_datalen(hdr) < sizeof(*nvs_hdr))) {
645 PMD_RX_LOG(ERR, "invalid receive nvs RNDIS");
649 /* Make sure that this is a RNDIS message. */
650 if (unlikely(nvs_hdr->type != NVS_TYPE_RNDIS)) {
651 PMD_RX_LOG(ERR, "nvs type %u, not RNDIS",
656 hlen = vmbus_chanpkt_getlen(hdr->hlen);
657 if (unlikely(hlen < sizeof(*pkt))) {
658 PMD_RX_LOG(ERR, "invalid rxbuf chanpkt");
662 pkt = container_of(hdr, const struct vmbus_chanpkt_rxbuf, hdr);
663 if (unlikely(pkt->rxbuf_id != NVS_RXBUF_SIG)) {
664 PMD_RX_LOG(ERR, "invalid rxbuf_id 0x%08x",
669 count = pkt->rxbuf_cnt;
670 if (unlikely(hlen < offsetof(struct vmbus_chanpkt_rxbuf,
672 PMD_RX_LOG(ERR, "invalid rxbuf_cnt %u", count);
676 if (pkt->hdr.xactid > hv->rxbuf_section_cnt) {
677 PMD_RX_LOG(ERR, "invalid rxbuf section id %" PRIx64,
682 /* Setup receive buffer info to allow for callback */
683 rxb = hn_rx_buf_init(rxq, pkt);
685 /* Each range represents 1 RNDIS pkt that contains 1 Ethernet frame */
686 for (i = 0; i < count; ++i) {
687 unsigned int ofs, len;
689 ofs = pkt->rxbuf[i].ofs;
690 len = pkt->rxbuf[i].len;
692 if (unlikely(ofs + len > rxbuf_sz)) {
694 "%uth RNDIS msg overflow ofs %u, len %u",
699 if (unlikely(len == 0)) {
700 PMD_RX_LOG(ERR, "%uth RNDIS msg len %u", i, len);
704 hn_rndis_receive(dev, rxq, rxb,
708 /* Send ACK now if external mbuf not used */
709 hn_rx_buf_release(rxb);
713 * Called when NVS inband events are received.
714 * Send up a two part message with port_id and the NVS message
715 * to the pipe to the netvsc-vf-event control thread.
717 static void hn_nvs_handle_notify(struct rte_eth_dev *dev,
718 const struct vmbus_chanpkt_hdr *pkt,
721 const struct hn_nvs_hdr *hdr = data;
724 case NVS_TYPE_TXTBL_NOTE:
725 /* Transmit indirection table has locking problems
726 * in DPDK and therefore not implemented
728 PMD_DRV_LOG(DEBUG, "host notify of transmit indirection table");
731 case NVS_TYPE_VFASSOC_NOTE:
732 hn_nvs_handle_vfassoc(dev, pkt, data);
737 "got notify, nvs type %u", hdr->type);
741 struct hn_rx_queue *hn_rx_queue_alloc(struct hn_data *hv,
743 unsigned int socket_id)
745 struct hn_rx_queue *rxq;
747 rxq = rte_zmalloc_socket("HN_RXQ", sizeof(*rxq),
748 RTE_CACHE_LINE_SIZE, socket_id);
753 rxq->chan = hv->channels[queue_id];
754 rte_spinlock_init(&rxq->ring_lock);
755 rxq->port_id = hv->port_id;
756 rxq->queue_id = queue_id;
757 rxq->event_sz = HN_RXQ_EVENT_DEFAULT;
758 rxq->event_buf = rte_malloc_socket("HN_EVENTS", HN_RXQ_EVENT_DEFAULT,
759 RTE_CACHE_LINE_SIZE, socket_id);
760 if (!rxq->event_buf) {
769 hn_dev_rx_queue_setup(struct rte_eth_dev *dev,
770 uint16_t queue_idx, uint16_t nb_desc,
771 unsigned int socket_id,
772 const struct rte_eth_rxconf *rx_conf,
773 struct rte_mempool *mp)
775 struct hn_data *hv = dev->data->dev_private;
776 char ring_name[RTE_RING_NAMESIZE];
777 struct hn_rx_queue *rxq;
781 PMD_INIT_FUNC_TRACE();
783 if (queue_idx == 0) {
786 rxq = hn_rx_queue_alloc(hv, queue_idx, socket_id);
792 count = rte_mempool_avail_count(mp) / dev->data->nb_rx_queues;
793 if (nb_desc == 0 || nb_desc > count)
797 * Staging ring from receive event logic to rx_pkts.
798 * rx_pkts assumes caller is handling multi-thread issue.
799 * event logic has locking.
801 snprintf(ring_name, sizeof(ring_name),
802 "hn_rx_%u_%u", dev->data->port_id, queue_idx);
803 rxq->rx_ring = rte_ring_create(ring_name,
804 rte_align32pow2(nb_desc),
809 error = hn_vf_rx_queue_setup(dev, queue_idx, nb_desc,
810 socket_id, rx_conf, mp);
814 dev->data->rx_queues[queue_idx] = rxq;
818 rte_ring_free(rxq->rx_ring);
819 rte_free(rxq->event_buf);
825 hn_dev_rx_queue_release(void *arg)
827 struct hn_rx_queue *rxq = arg;
829 PMD_INIT_FUNC_TRACE();
834 rte_ring_free(rxq->rx_ring);
838 hn_vf_rx_queue_release(rxq->hv, rxq->queue_id);
840 /* Keep primary queue to allow for control operations */
841 if (rxq != rxq->hv->primary) {
842 rte_free(rxq->event_buf);
848 hn_dev_tx_done_cleanup(void *arg, uint32_t free_cnt)
850 struct hn_tx_queue *txq = arg;
852 return hn_process_events(txq->hv, txq->queue_id, free_cnt);
856 * Process pending events on the channel.
857 * Called from both Rx queue poll and Tx cleanup
859 uint32_t hn_process_events(struct hn_data *hv, uint16_t queue_id,
862 struct rte_eth_dev *dev = &rte_eth_devices[hv->port_id];
863 struct hn_rx_queue *rxq;
864 uint32_t bytes_read = 0;
865 uint32_t tx_done = 0;
868 rxq = queue_id == 0 ? hv->primary : dev->data->rx_queues[queue_id];
870 /* If no pending data then nothing to do */
871 if (rte_vmbus_chan_rx_empty(rxq->chan))
875 * Since channel is shared between Rx and TX queue need to have a lock
876 * since DPDK does not force same CPU to be used for Rx/Tx.
878 if (unlikely(!rte_spinlock_trylock(&rxq->ring_lock)))
882 const struct vmbus_chanpkt_hdr *pkt;
883 uint32_t len = rxq->event_sz;
887 ret = rte_vmbus_chan_recv_raw(rxq->chan, rxq->event_buf, &len);
889 break; /* ring is empty */
891 if (unlikely(ret == -ENOBUFS)) {
892 /* event buffer not large enough to read ring */
895 "event buffer expansion (need %u)", len);
896 rxq->event_sz = len + len / 4;
897 rxq->event_buf = rte_realloc(rxq->event_buf, rxq->event_sz,
898 RTE_CACHE_LINE_SIZE);
901 /* out of memory, no more events now */
906 if (unlikely(ret <= 0)) {
907 /* This indicates a failure to communicate (or worse) */
908 rte_exit(EXIT_FAILURE,
909 "vmbus ring buffer error: %d", ret);
913 pkt = (const struct vmbus_chanpkt_hdr *)rxq->event_buf;
914 data = (char *)rxq->event_buf + vmbus_chanpkt_getlen(pkt->hlen);
917 case VMBUS_CHANPKT_TYPE_COMP:
919 hn_nvs_handle_comp(dev, queue_id, pkt, data);
922 case VMBUS_CHANPKT_TYPE_RXBUF:
923 hn_nvs_handle_rxbuf(dev, hv, rxq, pkt, data);
926 case VMBUS_CHANPKT_TYPE_INBAND:
927 hn_nvs_handle_notify(dev, pkt, data);
931 PMD_DRV_LOG(ERR, "unknown chan pkt %u", pkt->type);
935 if (tx_limit && tx_done >= tx_limit)
938 if (rxq->rx_ring && rte_ring_full(rxq->rx_ring))
943 rte_vmbus_chan_signal_read(rxq->chan, bytes_read);
945 rte_spinlock_unlock(&rxq->ring_lock);
950 static void hn_append_to_chim(struct hn_tx_queue *txq,
951 struct rndis_packet_msg *pkt,
952 const struct rte_mbuf *m)
954 struct hn_txdesc *txd = txq->agg_txd;
955 uint8_t *buf = (uint8_t *)pkt;
956 unsigned int data_offs;
960 data_offs = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
961 txd->chim_size += pkt->len;
962 txd->data_size += m->pkt_len;
964 hn_update_packet_stats(&txq->stats, m);
966 for (; m; m = m->next) {
967 uint16_t len = rte_pktmbuf_data_len(m);
969 rte_memcpy(buf + data_offs,
970 rte_pktmbuf_mtod(m, const char *), len);
976 * Send pending aggregated data in chimney buffer (if any).
977 * Returns error if send was unsuccessful because channel ring buffer
980 static int hn_flush_txagg(struct hn_tx_queue *txq, bool *need_sig)
983 struct hn_txdesc *txd = txq->agg_txd;
984 struct hn_nvs_rndis rndis;
990 rndis = (struct hn_nvs_rndis) {
991 .type = NVS_TYPE_RNDIS,
992 .rndis_mtype = NVS_RNDIS_MTYPE_DATA,
993 .chim_idx = txd->chim_index,
994 .chim_sz = txd->chim_size,
997 PMD_TX_LOG(DEBUG, "port %u:%u tx %u size %u",
998 txq->port_id, txq->queue_id, txd->chim_index, txd->chim_size);
1000 ret = hn_nvs_send(txq->chan, VMBUS_CHANPKT_FLAG_RC,
1001 &rndis, sizeof(rndis), (uintptr_t)txd, need_sig);
1003 if (likely(ret == 0))
1004 hn_reset_txagg(txq);
1006 PMD_TX_LOG(NOTICE, "port %u:%u send failed: %d",
1007 txq->port_id, txq->queue_id, ret);
1012 static struct hn_txdesc *hn_new_txd(struct hn_data *hv,
1013 struct hn_tx_queue *txq)
1015 struct hn_txdesc *txd;
1017 if (rte_mempool_get(hv->tx_pool, (void **)&txd)) {
1018 ++txq->stats.ring_full;
1019 PMD_TX_LOG(DEBUG, "tx pool exhausted!");
1024 txd->queue_id = txq->queue_id;
1033 hn_try_txagg(struct hn_data *hv, struct hn_tx_queue *txq, uint32_t pktsize)
1035 struct hn_txdesc *agg_txd = txq->agg_txd;
1036 struct rndis_packet_msg *pkt;
1040 unsigned int padding, olen;
1043 * Update the previous RNDIS packet's total length,
1044 * it can be increased due to the mandatory alignment
1045 * padding for this RNDIS packet. And update the
1046 * aggregating txdesc's chimney sending buffer size
1049 * Zero-out the padding, as required by the RNDIS spec.
1051 pkt = txq->agg_prevpkt;
1053 padding = RTE_ALIGN(olen, txq->agg_align) - olen;
1055 agg_txd->chim_size += padding;
1056 pkt->len += padding;
1057 memset((uint8_t *)pkt + olen, 0, padding);
1060 chim = (uint8_t *)pkt + pkt->len;
1063 txq->agg_szleft -= pktsize;
1064 if (txq->agg_szleft < HN_PKTSIZE_MIN(txq->agg_align)) {
1066 * Probably can't aggregate more packets,
1067 * flush this aggregating txdesc proactively.
1069 txq->agg_pktleft = 0;
1072 agg_txd = hn_new_txd(hv, txq);
1076 chim = (uint8_t *)hv->chim_res->addr
1077 + agg_txd->chim_index * hv->chim_szmax;
1079 txq->agg_txd = agg_txd;
1080 txq->agg_pktleft = txq->agg_pktmax - 1;
1081 txq->agg_szleft = txq->agg_szmax - pktsize;
1083 txq->agg_prevpkt = chim;
1088 static inline void *
1089 hn_rndis_pktinfo_append(struct rndis_packet_msg *pkt,
1090 uint32_t pi_dlen, uint32_t pi_type)
1092 const uint32_t pi_size = RNDIS_PKTINFO_SIZE(pi_dlen);
1093 struct rndis_pktinfo *pi;
1096 * Per-packet-info does not move; it only grows.
1099 * pktinfooffset in this phase counts from the beginning
1100 * of rndis_packet_msg.
1102 pi = (struct rndis_pktinfo *)((uint8_t *)pkt + hn_rndis_pktlen(pkt));
1104 pkt->pktinfolen += pi_size;
1108 pi->offset = RNDIS_PKTINFO_OFFSET;
1113 /* Put RNDIS header and packet info on packet */
1114 static void hn_encap(struct rndis_packet_msg *pkt,
1116 const struct rte_mbuf *m)
1118 unsigned int hlen = m->l2_len + m->l3_len;
1122 pkt->type = RNDIS_PACKET_MSG;
1123 pkt->len = m->pkt_len;
1124 pkt->dataoffset = 0;
1125 pkt->datalen = m->pkt_len;
1126 pkt->oobdataoffset = 0;
1127 pkt->oobdatalen = 0;
1128 pkt->oobdataelements = 0;
1129 pkt->pktinfooffset = sizeof(*pkt);
1130 pkt->pktinfolen = 0;
1135 * Set the hash value for this packet, to the queue_id to cause
1136 * TX done event for this packet on the right channel.
1138 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_HASH_VALUE_SIZE,
1139 NDIS_PKTINFO_TYPE_HASHVAL);
1140 *pi_data = queue_id;
1142 if (m->ol_flags & PKT_TX_VLAN_PKT) {
1143 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_VLAN_INFO_SIZE,
1144 NDIS_PKTINFO_TYPE_VLAN);
1145 *pi_data = m->vlan_tci;
1148 if (m->ol_flags & PKT_TX_TCP_SEG) {
1149 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_LSO2_INFO_SIZE,
1150 NDIS_PKTINFO_TYPE_LSO);
1152 if (m->ol_flags & PKT_TX_IPV6) {
1153 *pi_data = NDIS_LSO2_INFO_MAKEIPV6(hlen,
1156 *pi_data = NDIS_LSO2_INFO_MAKEIPV4(hlen,
1159 } else if (m->ol_flags &
1160 (PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM | PKT_TX_IP_CKSUM)) {
1161 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_TXCSUM_INFO_SIZE,
1162 NDIS_PKTINFO_TYPE_CSUM);
1165 if (m->ol_flags & PKT_TX_IPV6)
1166 *pi_data |= NDIS_TXCSUM_INFO_IPV6;
1167 if (m->ol_flags & PKT_TX_IPV4) {
1168 *pi_data |= NDIS_TXCSUM_INFO_IPV4;
1170 if (m->ol_flags & PKT_TX_IP_CKSUM)
1171 *pi_data |= NDIS_TXCSUM_INFO_IPCS;
1174 if (m->ol_flags & PKT_TX_TCP_CKSUM)
1175 *pi_data |= NDIS_TXCSUM_INFO_MKTCPCS(hlen);
1176 else if (m->ol_flags & PKT_TX_UDP_CKSUM)
1177 *pi_data |= NDIS_TXCSUM_INFO_MKUDPCS(hlen);
1180 pkt_hlen = pkt->pktinfooffset + pkt->pktinfolen;
1181 /* Fixup RNDIS packet message total length */
1182 pkt->len += pkt_hlen;
1184 /* Convert RNDIS packet message offsets */
1185 pkt->dataoffset = hn_rndis_pktmsg_offset(pkt_hlen);
1186 pkt->pktinfooffset = hn_rndis_pktmsg_offset(pkt->pktinfooffset);
1189 /* How many scatter gather list elements ar needed */
1190 static unsigned int hn_get_slots(const struct rte_mbuf *m)
1192 unsigned int slots = 1; /* for RNDIS header */
1195 unsigned int size = rte_pktmbuf_data_len(m);
1196 unsigned int offs = rte_mbuf_data_iova(m) & PAGE_MASK;
1198 slots += (offs + size + PAGE_SIZE - 1) / PAGE_SIZE;
1205 /* Build scatter gather list from chained mbuf */
1206 static unsigned int hn_fill_sg(struct vmbus_gpa *sg,
1207 const struct rte_mbuf *m)
1209 unsigned int segs = 0;
1212 rte_iova_t addr = rte_mbuf_data_iova(m);
1213 unsigned int page = addr / PAGE_SIZE;
1214 unsigned int offset = addr & PAGE_MASK;
1215 unsigned int len = rte_pktmbuf_data_len(m);
1218 unsigned int bytes = RTE_MIN(len, PAGE_SIZE - offset);
1220 sg[segs].page = page;
1221 sg[segs].ofs = offset;
1222 sg[segs].len = bytes;
1235 /* Transmit directly from mbuf */
1236 static int hn_xmit_sg(struct hn_tx_queue *txq,
1237 const struct hn_txdesc *txd, const struct rte_mbuf *m,
1240 struct vmbus_gpa sg[hn_get_slots(m)];
1241 struct hn_nvs_rndis nvs_rndis = {
1242 .type = NVS_TYPE_RNDIS,
1243 .rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1244 .chim_sz = txd->chim_size,
1249 /* attach aggregation data if present */
1250 if (txd->chim_size > 0)
1251 nvs_rndis.chim_idx = txd->chim_index;
1253 nvs_rndis.chim_idx = NVS_CHIM_IDX_INVALID;
1255 hn_rndis_dump(txd->rndis_pkt);
1257 /* pass IOVA of rndis header in first segment */
1258 addr = rte_malloc_virt2iova(txd->rndis_pkt);
1259 if (unlikely(addr == RTE_BAD_IOVA)) {
1260 PMD_DRV_LOG(ERR, "RNDIS transmit can not get iova");
1264 sg[0].page = addr / PAGE_SIZE;
1265 sg[0].ofs = addr & PAGE_MASK;
1266 sg[0].len = RNDIS_PACKET_MSG_OFFSET_ABS(hn_rndis_pktlen(txd->rndis_pkt));
1269 hn_update_packet_stats(&txq->stats, m);
1271 segs += hn_fill_sg(sg + 1, m);
1273 PMD_TX_LOG(DEBUG, "port %u:%u tx %u segs %u size %u",
1274 txq->port_id, txq->queue_id, txd->chim_index,
1275 segs, nvs_rndis.chim_sz);
1277 return hn_nvs_send_sglist(txq->chan, sg, segs,
1278 &nvs_rndis, sizeof(nvs_rndis),
1279 (uintptr_t)txd, need_sig);
1283 hn_xmit_pkts(void *ptxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1285 struct hn_tx_queue *txq = ptxq;
1286 uint16_t queue_id = txq->queue_id;
1287 struct hn_data *hv = txq->hv;
1288 struct rte_eth_dev *vf_dev;
1289 bool need_sig = false;
1293 if (unlikely(hv->closed))
1296 /* Transmit over VF if present and up */
1297 vf_dev = hv->vf_dev;
1298 rte_compiler_barrier();
1299 if (vf_dev && vf_dev->data->dev_started) {
1300 void *sub_q = vf_dev->data->tx_queues[queue_id];
1302 return (*vf_dev->tx_pkt_burst)(sub_q, tx_pkts, nb_pkts);
1305 if (rte_mempool_avail_count(hv->tx_pool) <= txq->free_thresh)
1306 hn_process_events(hv, txq->queue_id, 0);
1308 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1309 struct rte_mbuf *m = tx_pkts[nb_tx];
1310 uint32_t pkt_size = m->pkt_len + HN_RNDIS_PKT_LEN;
1311 struct rndis_packet_msg *pkt;
1313 /* For small packets aggregate them in chimney buffer */
1314 if (m->pkt_len < HN_TXCOPY_THRESHOLD && pkt_size <= txq->agg_szmax) {
1315 /* If this packet will not fit, then flush */
1316 if (txq->agg_pktleft == 0 ||
1317 RTE_ALIGN(pkt_size, txq->agg_align) > txq->agg_szleft) {
1318 if (hn_flush_txagg(txq, &need_sig))
1322 pkt = hn_try_txagg(hv, txq, pkt_size);
1326 hn_encap(pkt, queue_id, m);
1327 hn_append_to_chim(txq, pkt, m);
1329 rte_pktmbuf_free(m);
1331 /* if buffer is full, flush */
1332 if (txq->agg_pktleft == 0 &&
1333 hn_flush_txagg(txq, &need_sig))
1336 struct hn_txdesc *txd;
1338 /* can send chimney data and large packet at once */
1341 hn_reset_txagg(txq);
1343 txd = hn_new_txd(hv, txq);
1348 pkt = txd->rndis_pkt;
1350 txd->data_size += m->pkt_len;
1353 hn_encap(pkt, queue_id, m);
1355 ret = hn_xmit_sg(txq, txd, m, &need_sig);
1356 if (unlikely(ret != 0)) {
1357 PMD_TX_LOG(NOTICE, "sg send failed: %d", ret);
1358 ++txq->stats.errors;
1359 rte_mempool_put(hv->tx_pool, txd);
1365 /* If partial buffer left, then try and send it.
1366 * if that fails, then reuse it on next send.
1368 hn_flush_txagg(txq, &need_sig);
1372 rte_vmbus_chan_signal_tx(txq->chan);
1378 hn_recv_pkts(void *prxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1380 struct hn_rx_queue *rxq = prxq;
1381 struct hn_data *hv = rxq->hv;
1382 struct rte_eth_dev *vf_dev;
1385 if (unlikely(hv->closed))
1388 vf_dev = hv->vf_dev;
1389 rte_compiler_barrier();
1391 if (vf_dev && vf_dev->data->dev_started) {
1392 /* Normally, with SR-IOV the ring buffer will be empty */
1393 hn_process_events(hv, rxq->queue_id, 0);
1395 /* Get mbufs some bufs off of staging ring */
1396 nb_rcv = rte_ring_sc_dequeue_burst(rxq->rx_ring,
1399 /* And rest off of VF */
1400 nb_rcv += rte_eth_rx_burst(vf_dev->data->port_id,
1402 rx_pkts + nb_rcv, nb_pkts - nb_rcv);
1404 /* If receive ring is not full then get more */
1405 if (rte_ring_count(rxq->rx_ring) < nb_pkts)
1406 hn_process_events(hv, rxq->queue_id, 0);
1408 nb_rcv = rte_ring_sc_dequeue_burst(rxq->rx_ring,