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_bitmap.h>
22 #include <rte_branch_prediction.h>
23 #include <rte_ether.h>
24 #include <rte_common.h>
25 #include <rte_errno.h>
26 #include <rte_memory.h>
30 #include <rte_bus_vmbus.h>
31 #include <rte_spinlock.h>
39 #define HN_NVS_SEND_MSG_SIZE \
40 (sizeof(struct vmbus_chanpkt_hdr) + sizeof(struct hn_nvs_rndis))
42 #define HN_TXD_CACHE_SIZE 32 /* per cpu tx_descriptor pool cache */
43 #define HN_TXCOPY_THRESHOLD 512
45 #define HN_RXCOPY_THRESHOLD 256
46 #define HN_RXQ_EVENT_DEFAULT 2048
55 #define HN_RXINFO_VLAN 0x0001
56 #define HN_RXINFO_CSUM 0x0002
57 #define HN_RXINFO_HASHINF 0x0004
58 #define HN_RXINFO_HASHVAL 0x0008
59 #define HN_RXINFO_ALL \
65 #define HN_NDIS_VLAN_INFO_INVALID 0xffffffff
66 #define HN_NDIS_RXCSUM_INFO_INVALID 0
67 #define HN_NDIS_HASH_INFO_INVALID 0
70 * Per-transmit book keeping.
71 * A slot in transmit ring (chim_index) is reserved for each transmit.
73 * There are two types of transmit:
74 * - buffered transmit where chimney buffer is used and RNDIS header
75 * is in the buffer. mbuf == NULL for this case.
77 * - direct transmit where RNDIS header is in the in rndis_pkt
78 * mbuf is freed after transmit.
80 * Descriptors come from per-port pool which is used
81 * to limit number of outstanding requests per device.
92 struct rndis_packet_msg *rndis_pkt;
95 #define HN_RNDIS_PKT_LEN \
96 (sizeof(struct rndis_packet_msg) + \
97 RNDIS_PKTINFO_SIZE(NDIS_HASH_VALUE_SIZE) + \
98 RNDIS_PKTINFO_SIZE(NDIS_VLAN_INFO_SIZE) + \
99 RNDIS_PKTINFO_SIZE(NDIS_LSO2_INFO_SIZE) + \
100 RNDIS_PKTINFO_SIZE(NDIS_TXCSUM_INFO_SIZE))
102 #define HN_RNDIS_PKT_ALIGNED RTE_ALIGN(HN_RNDIS_PKT_LEN, RTE_CACHE_LINE_SIZE)
104 /* Minimum space required for a packet */
105 #define HN_PKTSIZE_MIN(align) \
106 RTE_ALIGN(RTE_ETHER_MIN_LEN + HN_RNDIS_PKT_LEN, align)
108 #define DEFAULT_TX_FREE_THRESH 32
111 hn_update_packet_stats(struct hn_stats *stats, const struct rte_mbuf *m)
113 uint32_t s = m->pkt_len;
114 const struct rte_ether_addr *ea;
117 stats->size_bins[1]++;
118 } else if (s > 64 && s < 1024) {
121 /* count zeros, and offset into correct bin */
122 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
123 stats->size_bins[bin]++;
126 stats->size_bins[0]++;
128 stats->size_bins[6]++;
130 stats->size_bins[7]++;
133 ea = rte_pktmbuf_mtod(m, const struct rte_ether_addr *);
134 if (rte_is_multicast_ether_addr(ea)) {
135 if (rte_is_broadcast_ether_addr(ea))
142 static inline unsigned int hn_rndis_pktlen(const struct rndis_packet_msg *pkt)
144 return pkt->pktinfooffset + pkt->pktinfolen;
147 static inline uint32_t
148 hn_rndis_pktmsg_offset(uint32_t ofs)
150 return ofs - offsetof(struct rndis_packet_msg, dataoffset);
153 static void hn_txd_init(struct rte_mempool *mp __rte_unused,
154 void *opaque, void *obj, unsigned int idx)
156 struct hn_tx_queue *txq = opaque;
157 struct hn_txdesc *txd = obj;
159 memset(txd, 0, sizeof(*txd));
161 txd->queue_id = txq->queue_id;
162 txd->chim_index = NVS_CHIM_IDX_INVALID;
163 txd->rndis_pkt = (struct rndis_packet_msg *)(char *)txq->tx_rndis
164 + idx * HN_RNDIS_PKT_ALIGNED;
168 hn_chim_init(struct rte_eth_dev *dev)
170 struct hn_data *hv = dev->data->dev_private;
171 uint32_t i, chim_bmp_size;
173 rte_spinlock_init(&hv->chim_lock);
174 chim_bmp_size = rte_bitmap_get_memory_footprint(hv->chim_cnt);
175 hv->chim_bmem = rte_zmalloc("hn_chim_bitmap", chim_bmp_size,
176 RTE_CACHE_LINE_SIZE);
177 if (hv->chim_bmem == NULL) {
178 PMD_INIT_LOG(ERR, "failed to allocate bitmap size %u",
183 hv->chim_bmap = rte_bitmap_init(hv->chim_cnt,
184 hv->chim_bmem, chim_bmp_size);
185 if (hv->chim_bmap == NULL) {
186 PMD_INIT_LOG(ERR, "failed to init chim bitmap");
190 for (i = 0; i < hv->chim_cnt; i++)
191 rte_bitmap_set(hv->chim_bmap, i);
197 hn_chim_uninit(struct rte_eth_dev *dev)
199 struct hn_data *hv = dev->data->dev_private;
201 rte_bitmap_free(hv->chim_bmap);
202 rte_free(hv->chim_bmem);
203 hv->chim_bmem = NULL;
206 static uint32_t hn_chim_alloc(struct hn_data *hv)
208 uint32_t index = NVS_CHIM_IDX_INVALID;
211 rte_spinlock_lock(&hv->chim_lock);
212 if (rte_bitmap_scan(hv->chim_bmap, &index, &slab))
213 rte_bitmap_clear(hv->chim_bmap, index);
214 rte_spinlock_unlock(&hv->chim_lock);
219 static void hn_chim_free(struct hn_data *hv, uint32_t chim_idx)
221 if (chim_idx >= hv->chim_cnt) {
222 PMD_DRV_LOG(ERR, "Invalid chimney index %u", chim_idx);
224 rte_spinlock_lock(&hv->chim_lock);
225 rte_bitmap_set(hv->chim_bmap, chim_idx);
226 rte_spinlock_unlock(&hv->chim_lock);
230 static void hn_reset_txagg(struct hn_tx_queue *txq)
232 txq->agg_szleft = txq->agg_szmax;
233 txq->agg_pktleft = txq->agg_pktmax;
235 txq->agg_prevpkt = NULL;
239 hn_dev_tx_queue_setup(struct rte_eth_dev *dev,
240 uint16_t queue_idx, uint16_t nb_desc,
241 unsigned int socket_id,
242 const struct rte_eth_txconf *tx_conf)
245 struct hn_data *hv = dev->data->dev_private;
246 struct hn_tx_queue *txq;
247 char name[RTE_MEMPOOL_NAMESIZE];
248 uint32_t tx_free_thresh;
251 PMD_INIT_FUNC_TRACE();
253 txq = rte_zmalloc_socket("HN_TXQ", sizeof(*txq), RTE_CACHE_LINE_SIZE,
259 txq->chan = hv->channels[queue_idx];
260 txq->port_id = dev->data->port_id;
261 txq->queue_id = queue_idx;
263 tx_free_thresh = tx_conf->tx_free_thresh;
264 if (tx_free_thresh == 0)
265 tx_free_thresh = RTE_MIN(nb_desc / 4,
266 DEFAULT_TX_FREE_THRESH);
268 if (tx_free_thresh + 3 >= nb_desc) {
270 "tx_free_thresh must be less than the number of TX entries minus 3(%u)."
271 " (tx_free_thresh=%u port=%u queue=%u)\n",
273 tx_free_thresh, dev->data->port_id, queue_idx);
277 txq->free_thresh = tx_free_thresh;
279 snprintf(name, sizeof(name),
280 "hn_txd_%u_%u", dev->data->port_id, queue_idx);
282 PMD_INIT_LOG(DEBUG, "TX descriptor pool %s n=%u size=%zu",
283 name, nb_desc, sizeof(struct hn_txdesc));
285 txq->tx_rndis = rte_calloc("hn_txq_rndis", nb_desc,
286 HN_RNDIS_PKT_ALIGNED, RTE_CACHE_LINE_SIZE);
287 if (txq->tx_rndis == NULL)
290 txq->txdesc_pool = rte_mempool_create(name, nb_desc,
291 sizeof(struct hn_txdesc),
294 dev->device->numa_node, 0);
295 if (txq->txdesc_pool == NULL) {
297 "mempool %s create failed: %d", name, rte_errno);
301 txq->agg_szmax = RTE_MIN(hv->chim_szmax, hv->rndis_agg_size);
302 txq->agg_pktmax = hv->rndis_agg_pkts;
303 txq->agg_align = hv->rndis_agg_align;
307 err = hn_vf_tx_queue_setup(dev, queue_idx, nb_desc,
310 dev->data->tx_queues[queue_idx] = txq;
315 if (txq->txdesc_pool)
316 rte_mempool_free(txq->txdesc_pool);
317 rte_free(txq->tx_rndis);
323 static struct hn_txdesc *hn_txd_get(struct hn_tx_queue *txq)
325 struct hn_txdesc *txd;
327 if (rte_mempool_get(txq->txdesc_pool, (void **)&txd)) {
328 ++txq->stats.ring_full;
329 PMD_TX_LOG(DEBUG, "tx pool exhausted!");
341 static void hn_txd_put(struct hn_tx_queue *txq, struct hn_txdesc *txd)
343 rte_mempool_put(txq->txdesc_pool, txd);
347 hn_dev_tx_queue_release(void *arg)
349 struct hn_tx_queue *txq = arg;
351 PMD_INIT_FUNC_TRACE();
356 if (txq->txdesc_pool)
357 rte_mempool_free(txq->txdesc_pool);
359 rte_free(txq->tx_rndis);
364 hn_nvs_send_completed(struct rte_eth_dev *dev, uint16_t queue_id,
365 unsigned long xactid, const struct hn_nvs_rndis_ack *ack)
367 struct hn_data *hv = dev->data->dev_private;
368 struct hn_txdesc *txd = (struct hn_txdesc *)xactid;
369 struct hn_tx_queue *txq;
371 /* Control packets are sent with xacid == 0 */
375 txq = dev->data->tx_queues[queue_id];
376 if (likely(ack->status == NVS_STATUS_OK)) {
377 PMD_TX_LOG(DEBUG, "port %u:%u complete tx %u packets %u bytes %u",
378 txq->port_id, txq->queue_id, txd->chim_index,
379 txd->packets, txd->data_size);
380 txq->stats.bytes += txd->data_size;
381 txq->stats.packets += txd->packets;
383 PMD_TX_LOG(NOTICE, "port %u:%u complete tx %u failed status %u",
384 txq->port_id, txq->queue_id, txd->chim_index, ack->status);
388 if (txd->chim_index != NVS_CHIM_IDX_INVALID)
389 hn_chim_free(hv, txd->chim_index);
391 rte_pktmbuf_free(txd->m);
392 hn_txd_put(txq, txd);
395 /* Handle transmit completion events */
397 hn_nvs_handle_comp(struct rte_eth_dev *dev, uint16_t queue_id,
398 const struct vmbus_chanpkt_hdr *pkt,
401 const struct hn_nvs_hdr *hdr = data;
404 case NVS_TYPE_RNDIS_ACK:
405 hn_nvs_send_completed(dev, queue_id, pkt->xactid, data);
410 "unexpected send completion type %u",
415 /* Parse per-packet info (meta data) */
417 hn_rndis_rxinfo(const void *info_data, unsigned int info_dlen,
418 struct hn_rxinfo *info)
420 const struct rndis_pktinfo *pi = info_data;
423 while (info_dlen != 0) {
427 if (unlikely(info_dlen < sizeof(*pi)))
430 if (unlikely(info_dlen < pi->size))
432 info_dlen -= pi->size;
434 if (unlikely(pi->size & RNDIS_PKTINFO_SIZE_ALIGNMASK))
436 if (unlikely(pi->size < pi->offset))
439 dlen = pi->size - pi->offset;
443 case NDIS_PKTINFO_TYPE_VLAN:
444 if (unlikely(dlen < NDIS_VLAN_INFO_SIZE))
446 info->vlan_info = *((const uint32_t *)data);
447 mask |= HN_RXINFO_VLAN;
450 case NDIS_PKTINFO_TYPE_CSUM:
451 if (unlikely(dlen < NDIS_RXCSUM_INFO_SIZE))
453 info->csum_info = *((const uint32_t *)data);
454 mask |= HN_RXINFO_CSUM;
457 case NDIS_PKTINFO_TYPE_HASHVAL:
458 if (unlikely(dlen < NDIS_HASH_VALUE_SIZE))
460 info->hash_value = *((const uint32_t *)data);
461 mask |= HN_RXINFO_HASHVAL;
464 case NDIS_PKTINFO_TYPE_HASHINF:
465 if (unlikely(dlen < NDIS_HASH_INFO_SIZE))
467 info->hash_info = *((const uint32_t *)data);
468 mask |= HN_RXINFO_HASHINF;
475 if (mask == HN_RXINFO_ALL)
476 break; /* All found; done */
478 pi = (const struct rndis_pktinfo *)
479 ((const uint8_t *)pi + pi->size);
484 * - If there is no hash value, invalidate the hash info.
486 if (!(mask & HN_RXINFO_HASHVAL))
487 info->hash_info = HN_NDIS_HASH_INFO_INVALID;
492 * Ack the consumed RXBUF associated w/ this channel packet,
493 * so that this RXBUF can be recycled by the hypervisor.
495 static void hn_rx_buf_release(struct hn_rx_bufinfo *rxb)
497 struct rte_mbuf_ext_shared_info *shinfo = &rxb->shinfo;
498 struct hn_data *hv = rxb->hv;
500 if (rte_mbuf_ext_refcnt_update(shinfo, -1) == 0) {
501 hn_nvs_ack_rxbuf(rxb->chan, rxb->xactid);
502 --hv->rxbuf_outstanding;
506 static void hn_rx_buf_free_cb(void *buf __rte_unused, void *opaque)
508 hn_rx_buf_release(opaque);
511 static struct hn_rx_bufinfo *hn_rx_buf_init(const struct hn_rx_queue *rxq,
512 const struct vmbus_chanpkt_rxbuf *pkt)
514 struct hn_rx_bufinfo *rxb;
516 rxb = rxq->hv->rxbuf_info + pkt->hdr.xactid;
517 rxb->chan = rxq->chan;
518 rxb->xactid = pkt->hdr.xactid;
521 rxb->shinfo.free_cb = hn_rx_buf_free_cb;
522 rxb->shinfo.fcb_opaque = rxb;
523 rte_mbuf_ext_refcnt_set(&rxb->shinfo, 1);
527 static void hn_rxpkt(struct hn_rx_queue *rxq, struct hn_rx_bufinfo *rxb,
528 uint8_t *data, unsigned int headroom, unsigned int dlen,
529 const struct hn_rxinfo *info)
531 struct hn_data *hv = rxq->hv;
534 m = rte_pktmbuf_alloc(rxq->mb_pool);
536 struct rte_eth_dev *dev =
537 &rte_eth_devices[rxq->port_id];
539 dev->data->rx_mbuf_alloc_failed++;
544 * For large packets, avoid copy if possible but need to keep
545 * some space available in receive area for later packets.
547 if (dlen >= HN_RXCOPY_THRESHOLD &&
548 hv->rxbuf_outstanding < hv->rxbuf_section_cnt / 2) {
549 struct rte_mbuf_ext_shared_info *shinfo;
554 * Build an external mbuf that points to recveive area.
555 * Use refcount to handle multiple packets in same
556 * receive buffer section.
558 rxbuf = hv->rxbuf_res->addr;
559 iova = rte_mem_virt2iova(rxbuf) + RTE_PTR_DIFF(data, rxbuf);
560 shinfo = &rxb->shinfo;
562 if (rte_mbuf_ext_refcnt_update(shinfo, 1) == 1)
563 ++hv->rxbuf_outstanding;
565 rte_pktmbuf_attach_extbuf(m, data, iova,
566 dlen + headroom, shinfo);
567 m->data_off = headroom;
569 /* Mbuf's in pool must be large enough to hold small packets */
570 if (unlikely(rte_pktmbuf_tailroom(m) < dlen)) {
571 rte_pktmbuf_free_seg(m);
575 rte_memcpy(rte_pktmbuf_mtod(m, void *),
576 data + headroom, dlen);
579 m->port = rxq->port_id;
582 m->packet_type = rte_net_get_ptype(m, NULL,
587 if (info->vlan_info != HN_NDIS_VLAN_INFO_INVALID) {
588 m->vlan_tci = info->vlan_info;
589 m->ol_flags |= PKT_RX_VLAN_STRIPPED | PKT_RX_VLAN;
591 /* NDIS always strips tag, put it back if necessary */
592 if (!hv->vlan_strip && rte_vlan_insert(&m)) {
593 PMD_DRV_LOG(DEBUG, "vlan insert failed");
600 if (info->csum_info != HN_NDIS_RXCSUM_INFO_INVALID) {
601 if (info->csum_info & NDIS_RXCSUM_INFO_IPCS_OK)
602 m->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
604 if (info->csum_info & (NDIS_RXCSUM_INFO_UDPCS_OK
605 | NDIS_RXCSUM_INFO_TCPCS_OK))
606 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
607 else if (info->csum_info & (NDIS_RXCSUM_INFO_TCPCS_FAILED
608 | NDIS_RXCSUM_INFO_UDPCS_FAILED))
609 m->ol_flags |= PKT_RX_L4_CKSUM_BAD;
612 if (info->hash_info != HN_NDIS_HASH_INFO_INVALID) {
613 m->ol_flags |= PKT_RX_RSS_HASH;
614 m->hash.rss = info->hash_value;
618 "port %u:%u RX id %"PRIu64" size %u type %#x ol_flags %#"PRIx64,
619 rxq->port_id, rxq->queue_id, rxb->xactid,
620 m->pkt_len, m->packet_type, m->ol_flags);
622 ++rxq->stats.packets;
623 rxq->stats.bytes += m->pkt_len;
624 hn_update_packet_stats(&rxq->stats, m);
626 if (unlikely(rte_ring_sp_enqueue(rxq->rx_ring, m) != 0)) {
627 ++rxq->stats.ring_full;
632 static void hn_rndis_rx_data(struct hn_rx_queue *rxq,
633 struct hn_rx_bufinfo *rxb,
634 void *data, uint32_t dlen)
636 unsigned int data_off, data_len, pktinfo_off, pktinfo_len;
637 const struct rndis_packet_msg *pkt = data;
638 struct hn_rxinfo info = {
639 .vlan_info = HN_NDIS_VLAN_INFO_INVALID,
640 .csum_info = HN_NDIS_RXCSUM_INFO_INVALID,
641 .hash_info = HN_NDIS_HASH_INFO_INVALID,
647 if (unlikely(dlen < sizeof(*pkt)))
650 if (unlikely(dlen < pkt->len))
651 goto error; /* truncated RNDIS from host */
653 if (unlikely(pkt->len < pkt->datalen
654 + pkt->oobdatalen + pkt->pktinfolen))
657 if (unlikely(pkt->datalen == 0))
661 if (unlikely(pkt->dataoffset < RNDIS_PACKET_MSG_OFFSET_MIN))
664 if (likely(pkt->pktinfooffset > 0) &&
665 unlikely(pkt->pktinfooffset < RNDIS_PACKET_MSG_OFFSET_MIN ||
666 (pkt->pktinfooffset & RNDIS_PACKET_MSG_OFFSET_ALIGNMASK)))
669 data_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
670 data_len = pkt->datalen;
671 pktinfo_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->pktinfooffset);
672 pktinfo_len = pkt->pktinfolen;
674 if (likely(pktinfo_len > 0)) {
675 err = hn_rndis_rxinfo((const uint8_t *)pkt + pktinfo_off,
681 if (unlikely(data_off + data_len > pkt->len))
684 if (unlikely(data_len < RTE_ETHER_HDR_LEN))
687 hn_rxpkt(rxq, rxb, data, data_off, data_len, &info);
694 hn_rndis_receive(struct rte_eth_dev *dev, struct hn_rx_queue *rxq,
695 struct hn_rx_bufinfo *rxb, void *buf, uint32_t len)
697 const struct rndis_msghdr *hdr = buf;
700 case RNDIS_PACKET_MSG:
701 if (dev->data->dev_started)
702 hn_rndis_rx_data(rxq, rxb, buf, len);
705 case RNDIS_INDICATE_STATUS_MSG:
706 hn_rndis_link_status(dev, buf);
709 case RNDIS_INITIALIZE_CMPLT:
710 case RNDIS_QUERY_CMPLT:
711 case RNDIS_SET_CMPLT:
712 hn_rndis_receive_response(rxq->hv, buf, len);
717 "unexpected RNDIS message (type %#x len %u)",
724 hn_nvs_handle_rxbuf(struct rte_eth_dev *dev,
726 struct hn_rx_queue *rxq,
727 const struct vmbus_chanpkt_hdr *hdr,
730 const struct vmbus_chanpkt_rxbuf *pkt;
731 const struct hn_nvs_hdr *nvs_hdr = buf;
732 uint32_t rxbuf_sz = hv->rxbuf_res->len;
733 char *rxbuf = hv->rxbuf_res->addr;
734 unsigned int i, hlen, count;
735 struct hn_rx_bufinfo *rxb;
737 /* At minimum we need type header */
738 if (unlikely(vmbus_chanpkt_datalen(hdr) < sizeof(*nvs_hdr))) {
739 PMD_RX_LOG(ERR, "invalid receive nvs RNDIS");
743 /* Make sure that this is a RNDIS message. */
744 if (unlikely(nvs_hdr->type != NVS_TYPE_RNDIS)) {
745 PMD_RX_LOG(ERR, "nvs type %u, not RNDIS",
750 hlen = vmbus_chanpkt_getlen(hdr->hlen);
751 if (unlikely(hlen < sizeof(*pkt))) {
752 PMD_RX_LOG(ERR, "invalid rxbuf chanpkt");
756 pkt = container_of(hdr, const struct vmbus_chanpkt_rxbuf, hdr);
757 if (unlikely(pkt->rxbuf_id != NVS_RXBUF_SIG)) {
758 PMD_RX_LOG(ERR, "invalid rxbuf_id 0x%08x",
763 count = pkt->rxbuf_cnt;
764 if (unlikely(hlen < offsetof(struct vmbus_chanpkt_rxbuf,
766 PMD_RX_LOG(ERR, "invalid rxbuf_cnt %u", count);
770 if (pkt->hdr.xactid > hv->rxbuf_section_cnt) {
771 PMD_RX_LOG(ERR, "invalid rxbuf section id %" PRIx64,
776 /* Setup receive buffer info to allow for callback */
777 rxb = hn_rx_buf_init(rxq, pkt);
779 /* Each range represents 1 RNDIS pkt that contains 1 Ethernet frame */
780 for (i = 0; i < count; ++i) {
781 unsigned int ofs, len;
783 ofs = pkt->rxbuf[i].ofs;
784 len = pkt->rxbuf[i].len;
786 if (unlikely(ofs + len > rxbuf_sz)) {
788 "%uth RNDIS msg overflow ofs %u, len %u",
793 if (unlikely(len == 0)) {
794 PMD_RX_LOG(ERR, "%uth RNDIS msg len %u", i, len);
798 hn_rndis_receive(dev, rxq, rxb,
802 /* Send ACK now if external mbuf not used */
803 hn_rx_buf_release(rxb);
807 * Called when NVS inband events are received.
808 * Send up a two part message with port_id and the NVS message
809 * to the pipe to the netvsc-vf-event control thread.
811 static void hn_nvs_handle_notify(struct rte_eth_dev *dev,
812 const struct vmbus_chanpkt_hdr *pkt,
815 const struct hn_nvs_hdr *hdr = data;
818 case NVS_TYPE_TXTBL_NOTE:
819 /* Transmit indirection table has locking problems
820 * in DPDK and therefore not implemented
822 PMD_DRV_LOG(DEBUG, "host notify of transmit indirection table");
825 case NVS_TYPE_VFASSOC_NOTE:
826 hn_nvs_handle_vfassoc(dev, pkt, data);
831 "got notify, nvs type %u", hdr->type);
835 struct hn_rx_queue *hn_rx_queue_alloc(struct hn_data *hv,
837 unsigned int socket_id)
839 struct hn_rx_queue *rxq;
841 rxq = rte_zmalloc_socket("HN_RXQ", sizeof(*rxq),
842 RTE_CACHE_LINE_SIZE, socket_id);
847 rxq->chan = hv->channels[queue_id];
848 rte_spinlock_init(&rxq->ring_lock);
849 rxq->port_id = hv->port_id;
850 rxq->queue_id = queue_id;
851 rxq->event_sz = HN_RXQ_EVENT_DEFAULT;
852 rxq->event_buf = rte_malloc_socket("HN_EVENTS", HN_RXQ_EVENT_DEFAULT,
853 RTE_CACHE_LINE_SIZE, socket_id);
854 if (!rxq->event_buf) {
863 hn_dev_rx_queue_setup(struct rte_eth_dev *dev,
864 uint16_t queue_idx, uint16_t nb_desc,
865 unsigned int socket_id,
866 const struct rte_eth_rxconf *rx_conf,
867 struct rte_mempool *mp)
869 struct hn_data *hv = dev->data->dev_private;
870 char ring_name[RTE_RING_NAMESIZE];
871 struct hn_rx_queue *rxq;
875 PMD_INIT_FUNC_TRACE();
877 if (queue_idx == 0) {
880 rxq = hn_rx_queue_alloc(hv, queue_idx, socket_id);
886 count = rte_mempool_avail_count(mp) / dev->data->nb_rx_queues;
887 if (nb_desc == 0 || nb_desc > count)
891 * Staging ring from receive event logic to rx_pkts.
892 * rx_pkts assumes caller is handling multi-thread issue.
893 * event logic has locking.
895 snprintf(ring_name, sizeof(ring_name),
896 "hn_rx_%u_%u", dev->data->port_id, queue_idx);
897 rxq->rx_ring = rte_ring_create(ring_name,
898 rte_align32pow2(nb_desc),
903 error = hn_vf_rx_queue_setup(dev, queue_idx, nb_desc,
904 socket_id, rx_conf, mp);
908 dev->data->rx_queues[queue_idx] = rxq;
912 rte_ring_free(rxq->rx_ring);
913 rte_free(rxq->event_buf);
919 hn_rx_queue_free(struct hn_rx_queue *rxq, bool keep_primary)
925 rte_ring_free(rxq->rx_ring);
929 hn_vf_rx_queue_release(rxq->hv, rxq->queue_id);
931 /* Keep primary queue to allow for control operations */
932 if (keep_primary && rxq == rxq->hv->primary)
935 rte_free(rxq->event_buf);
940 hn_dev_rx_queue_release(void *arg)
942 struct hn_rx_queue *rxq = arg;
944 PMD_INIT_FUNC_TRACE();
946 hn_rx_queue_free(rxq, true);
950 hn_dev_tx_done_cleanup(void *arg, uint32_t free_cnt)
952 struct hn_tx_queue *txq = arg;
954 return hn_process_events(txq->hv, txq->queue_id, free_cnt);
958 * Process pending events on the channel.
959 * Called from both Rx queue poll and Tx cleanup
961 uint32_t hn_process_events(struct hn_data *hv, uint16_t queue_id,
964 struct rte_eth_dev *dev = &rte_eth_devices[hv->port_id];
965 struct hn_rx_queue *rxq;
966 uint32_t bytes_read = 0;
967 uint32_t tx_done = 0;
970 rxq = queue_id == 0 ? hv->primary : dev->data->rx_queues[queue_id];
972 /* If no pending data then nothing to do */
973 if (rte_vmbus_chan_rx_empty(rxq->chan))
977 * Since channel is shared between Rx and TX queue need to have a lock
978 * since DPDK does not force same CPU to be used for Rx/Tx.
980 if (unlikely(!rte_spinlock_trylock(&rxq->ring_lock)))
984 const struct vmbus_chanpkt_hdr *pkt;
985 uint32_t len = rxq->event_sz;
989 ret = rte_vmbus_chan_recv_raw(rxq->chan, rxq->event_buf, &len);
991 break; /* ring is empty */
993 if (unlikely(ret == -ENOBUFS)) {
994 /* event buffer not large enough to read ring */
997 "event buffer expansion (need %u)", len);
998 rxq->event_sz = len + len / 4;
999 rxq->event_buf = rte_realloc(rxq->event_buf, rxq->event_sz,
1000 RTE_CACHE_LINE_SIZE);
1003 /* out of memory, no more events now */
1008 if (unlikely(ret <= 0)) {
1009 /* This indicates a failure to communicate (or worse) */
1010 rte_exit(EXIT_FAILURE,
1011 "vmbus ring buffer error: %d", ret);
1015 pkt = (const struct vmbus_chanpkt_hdr *)rxq->event_buf;
1016 data = (char *)rxq->event_buf + vmbus_chanpkt_getlen(pkt->hlen);
1018 switch (pkt->type) {
1019 case VMBUS_CHANPKT_TYPE_COMP:
1021 hn_nvs_handle_comp(dev, queue_id, pkt, data);
1024 case VMBUS_CHANPKT_TYPE_RXBUF:
1025 hn_nvs_handle_rxbuf(dev, hv, rxq, pkt, data);
1028 case VMBUS_CHANPKT_TYPE_INBAND:
1029 hn_nvs_handle_notify(dev, pkt, data);
1033 PMD_DRV_LOG(ERR, "unknown chan pkt %u", pkt->type);
1037 if (tx_limit && tx_done >= tx_limit)
1040 if (rxq->rx_ring && rte_ring_full(rxq->rx_ring))
1045 rte_vmbus_chan_signal_read(rxq->chan, bytes_read);
1047 rte_spinlock_unlock(&rxq->ring_lock);
1052 static void hn_append_to_chim(struct hn_tx_queue *txq,
1053 struct rndis_packet_msg *pkt,
1054 const struct rte_mbuf *m)
1056 struct hn_txdesc *txd = txq->agg_txd;
1057 uint8_t *buf = (uint8_t *)pkt;
1058 unsigned int data_offs;
1062 data_offs = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
1063 txd->chim_size += pkt->len;
1064 txd->data_size += m->pkt_len;
1066 hn_update_packet_stats(&txq->stats, m);
1068 for (; m; m = m->next) {
1069 uint16_t len = rte_pktmbuf_data_len(m);
1071 rte_memcpy(buf + data_offs,
1072 rte_pktmbuf_mtod(m, const char *), len);
1078 * Send pending aggregated data in chimney buffer (if any).
1079 * Returns error if send was unsuccessful because channel ring buffer
1082 static int hn_flush_txagg(struct hn_tx_queue *txq, bool *need_sig)
1085 struct hn_txdesc *txd = txq->agg_txd;
1086 struct hn_nvs_rndis rndis;
1092 rndis = (struct hn_nvs_rndis) {
1093 .type = NVS_TYPE_RNDIS,
1094 .rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1095 .chim_idx = txd->chim_index,
1096 .chim_sz = txd->chim_size,
1099 PMD_TX_LOG(DEBUG, "port %u:%u tx %u size %u",
1100 txq->port_id, txq->queue_id, txd->chim_index, txd->chim_size);
1102 ret = hn_nvs_send(txq->chan, VMBUS_CHANPKT_FLAG_RC,
1103 &rndis, sizeof(rndis), (uintptr_t)txd, need_sig);
1105 if (likely(ret == 0))
1106 hn_reset_txagg(txq);
1108 PMD_TX_LOG(NOTICE, "port %u:%u send failed: %d",
1109 txq->port_id, txq->queue_id, ret);
1115 * Try and find a place in a send chimney buffer to put
1116 * the small packet. If space is available, this routine
1117 * returns a pointer of where to place the data.
1118 * If no space, caller should try direct transmit.
1121 hn_try_txagg(struct hn_data *hv, struct hn_tx_queue *txq,
1122 struct hn_txdesc *txd, uint32_t pktsize)
1124 struct hn_txdesc *agg_txd = txq->agg_txd;
1125 struct rndis_packet_msg *pkt;
1129 unsigned int padding, olen;
1132 * Update the previous RNDIS packet's total length,
1133 * it can be increased due to the mandatory alignment
1134 * padding for this RNDIS packet. And update the
1135 * aggregating txdesc's chimney sending buffer size
1138 * Zero-out the padding, as required by the RNDIS spec.
1140 pkt = txq->agg_prevpkt;
1142 padding = RTE_ALIGN(olen, txq->agg_align) - olen;
1144 agg_txd->chim_size += padding;
1145 pkt->len += padding;
1146 memset((uint8_t *)pkt + olen, 0, padding);
1149 chim = (uint8_t *)pkt + pkt->len;
1150 txq->agg_prevpkt = chim;
1152 txq->agg_szleft -= pktsize;
1153 if (txq->agg_szleft < HN_PKTSIZE_MIN(txq->agg_align)) {
1155 * Probably can't aggregate more packets,
1156 * flush this aggregating txdesc proactively.
1158 txq->agg_pktleft = 0;
1161 hn_txd_put(txq, txd);
1165 txd->chim_index = hn_chim_alloc(hv);
1166 if (txd->chim_index == NVS_CHIM_IDX_INVALID)
1169 chim = (uint8_t *)hv->chim_res->addr
1170 + txd->chim_index * hv->chim_szmax;
1173 txq->agg_pktleft = txq->agg_pktmax - 1;
1174 txq->agg_szleft = txq->agg_szmax - pktsize;
1175 txq->agg_prevpkt = chim;
1180 static inline void *
1181 hn_rndis_pktinfo_append(struct rndis_packet_msg *pkt,
1182 uint32_t pi_dlen, uint32_t pi_type)
1184 const uint32_t pi_size = RNDIS_PKTINFO_SIZE(pi_dlen);
1185 struct rndis_pktinfo *pi;
1188 * Per-packet-info does not move; it only grows.
1191 * pktinfooffset in this phase counts from the beginning
1192 * of rndis_packet_msg.
1194 pi = (struct rndis_pktinfo *)((uint8_t *)pkt + hn_rndis_pktlen(pkt));
1196 pkt->pktinfolen += pi_size;
1200 pi->offset = RNDIS_PKTINFO_OFFSET;
1205 /* Put RNDIS header and packet info on packet */
1206 static void hn_encap(struct rndis_packet_msg *pkt,
1208 const struct rte_mbuf *m)
1210 unsigned int hlen = m->l2_len + m->l3_len;
1214 pkt->type = RNDIS_PACKET_MSG;
1215 pkt->len = m->pkt_len;
1216 pkt->dataoffset = 0;
1217 pkt->datalen = m->pkt_len;
1218 pkt->oobdataoffset = 0;
1219 pkt->oobdatalen = 0;
1220 pkt->oobdataelements = 0;
1221 pkt->pktinfooffset = sizeof(*pkt);
1222 pkt->pktinfolen = 0;
1227 * Set the hash value for this packet, to the queue_id to cause
1228 * TX done event for this packet on the right channel.
1230 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_HASH_VALUE_SIZE,
1231 NDIS_PKTINFO_TYPE_HASHVAL);
1232 *pi_data = queue_id;
1234 if (m->ol_flags & PKT_TX_VLAN_PKT) {
1235 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_VLAN_INFO_SIZE,
1236 NDIS_PKTINFO_TYPE_VLAN);
1237 *pi_data = m->vlan_tci;
1240 if (m->ol_flags & PKT_TX_TCP_SEG) {
1241 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_LSO2_INFO_SIZE,
1242 NDIS_PKTINFO_TYPE_LSO);
1244 if (m->ol_flags & PKT_TX_IPV6) {
1245 *pi_data = NDIS_LSO2_INFO_MAKEIPV6(hlen,
1248 *pi_data = NDIS_LSO2_INFO_MAKEIPV4(hlen,
1251 } else if (m->ol_flags &
1252 (PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM | PKT_TX_IP_CKSUM)) {
1253 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_TXCSUM_INFO_SIZE,
1254 NDIS_PKTINFO_TYPE_CSUM);
1257 if (m->ol_flags & PKT_TX_IPV6)
1258 *pi_data |= NDIS_TXCSUM_INFO_IPV6;
1259 if (m->ol_flags & PKT_TX_IPV4) {
1260 *pi_data |= NDIS_TXCSUM_INFO_IPV4;
1262 if (m->ol_flags & PKT_TX_IP_CKSUM)
1263 *pi_data |= NDIS_TXCSUM_INFO_IPCS;
1266 if (m->ol_flags & PKT_TX_TCP_CKSUM)
1267 *pi_data |= NDIS_TXCSUM_INFO_MKTCPCS(hlen);
1268 else if (m->ol_flags & PKT_TX_UDP_CKSUM)
1269 *pi_data |= NDIS_TXCSUM_INFO_MKUDPCS(hlen);
1272 pkt_hlen = pkt->pktinfooffset + pkt->pktinfolen;
1273 /* Fixup RNDIS packet message total length */
1274 pkt->len += pkt_hlen;
1276 /* Convert RNDIS packet message offsets */
1277 pkt->dataoffset = hn_rndis_pktmsg_offset(pkt_hlen);
1278 pkt->pktinfooffset = hn_rndis_pktmsg_offset(pkt->pktinfooffset);
1281 /* How many scatter gather list elements ar needed */
1282 static unsigned int hn_get_slots(const struct rte_mbuf *m)
1284 unsigned int slots = 1; /* for RNDIS header */
1287 unsigned int size = rte_pktmbuf_data_len(m);
1288 unsigned int offs = rte_mbuf_data_iova(m) & PAGE_MASK;
1290 slots += (offs + size + PAGE_SIZE - 1) / PAGE_SIZE;
1297 /* Build scatter gather list from chained mbuf */
1298 static unsigned int hn_fill_sg(struct vmbus_gpa *sg,
1299 const struct rte_mbuf *m)
1301 unsigned int segs = 0;
1304 rte_iova_t addr = rte_mbuf_data_iova(m);
1305 unsigned int page = addr / PAGE_SIZE;
1306 unsigned int offset = addr & PAGE_MASK;
1307 unsigned int len = rte_pktmbuf_data_len(m);
1310 unsigned int bytes = RTE_MIN(len, PAGE_SIZE - offset);
1312 sg[segs].page = page;
1313 sg[segs].ofs = offset;
1314 sg[segs].len = bytes;
1327 /* Transmit directly from mbuf */
1328 static int hn_xmit_sg(struct hn_tx_queue *txq,
1329 const struct hn_txdesc *txd, const struct rte_mbuf *m,
1332 struct vmbus_gpa sg[hn_get_slots(m)];
1333 struct hn_nvs_rndis nvs_rndis = {
1334 .type = NVS_TYPE_RNDIS,
1335 .rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1336 .chim_sz = txd->chim_size,
1341 /* attach aggregation data if present */
1342 if (txd->chim_size > 0)
1343 nvs_rndis.chim_idx = txd->chim_index;
1345 nvs_rndis.chim_idx = NVS_CHIM_IDX_INVALID;
1347 hn_rndis_dump(txd->rndis_pkt);
1349 /* pass IOVA of rndis header in first segment */
1350 addr = rte_malloc_virt2iova(txd->rndis_pkt);
1351 if (unlikely(addr == RTE_BAD_IOVA)) {
1352 PMD_DRV_LOG(ERR, "RNDIS transmit can not get iova");
1356 sg[0].page = addr / PAGE_SIZE;
1357 sg[0].ofs = addr & PAGE_MASK;
1358 sg[0].len = RNDIS_PACKET_MSG_OFFSET_ABS(hn_rndis_pktlen(txd->rndis_pkt));
1361 hn_update_packet_stats(&txq->stats, m);
1363 segs += hn_fill_sg(sg + 1, m);
1365 PMD_TX_LOG(DEBUG, "port %u:%u tx %u segs %u size %u",
1366 txq->port_id, txq->queue_id, txd->chim_index,
1367 segs, nvs_rndis.chim_sz);
1369 return hn_nvs_send_sglist(txq->chan, sg, segs,
1370 &nvs_rndis, sizeof(nvs_rndis),
1371 (uintptr_t)txd, need_sig);
1375 hn_xmit_pkts(void *ptxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1377 struct hn_tx_queue *txq = ptxq;
1378 uint16_t queue_id = txq->queue_id;
1379 struct hn_data *hv = txq->hv;
1380 struct rte_eth_dev *vf_dev;
1381 bool need_sig = false;
1385 if (unlikely(hv->closed))
1388 /* Transmit over VF if present and up */
1389 vf_dev = hn_get_vf_dev(hv);
1391 if (vf_dev && vf_dev->data->dev_started) {
1392 void *sub_q = vf_dev->data->tx_queues[queue_id];
1394 return (*vf_dev->tx_pkt_burst)(sub_q, tx_pkts, nb_pkts);
1397 if (rte_mempool_avail_count(txq->txdesc_pool) <= txq->free_thresh)
1398 hn_process_events(hv, txq->queue_id, 0);
1400 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1401 struct rte_mbuf *m = tx_pkts[nb_tx];
1402 uint32_t pkt_size = m->pkt_len + HN_RNDIS_PKT_LEN;
1403 struct rndis_packet_msg *pkt;
1404 struct hn_txdesc *txd;
1406 txd = hn_txd_get(txq);
1410 /* For small packets aggregate them in chimney buffer */
1411 if (m->pkt_len < HN_TXCOPY_THRESHOLD && pkt_size <= txq->agg_szmax) {
1412 /* If this packet will not fit, then flush */
1413 if (txq->agg_pktleft == 0 ||
1414 RTE_ALIGN(pkt_size, txq->agg_align) > txq->agg_szleft) {
1415 if (hn_flush_txagg(txq, &need_sig))
1420 pkt = hn_try_txagg(hv, txq, txd, pkt_size);
1424 hn_encap(pkt, queue_id, m);
1425 hn_append_to_chim(txq, pkt, m);
1427 rte_pktmbuf_free(m);
1429 /* if buffer is full, flush */
1430 if (txq->agg_pktleft == 0 &&
1431 hn_flush_txagg(txq, &need_sig))
1434 /* Send any outstanding packets in buffer */
1435 if (txq->agg_txd && hn_flush_txagg(txq, &need_sig))
1438 pkt = txd->rndis_pkt;
1440 txd->data_size = m->pkt_len;
1443 hn_encap(pkt, queue_id, m);
1445 ret = hn_xmit_sg(txq, txd, m, &need_sig);
1446 if (unlikely(ret != 0)) {
1447 PMD_TX_LOG(NOTICE, "sg send failed: %d", ret);
1448 ++txq->stats.errors;
1449 hn_txd_put(txq, txd);
1455 /* If partial buffer left, then try and send it.
1456 * if that fails, then reuse it on next send.
1458 hn_flush_txagg(txq, &need_sig);
1462 rte_vmbus_chan_signal_tx(txq->chan);
1468 hn_recv_vf(uint16_t vf_port, const struct hn_rx_queue *rxq,
1469 struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1473 if (unlikely(nb_pkts == 0))
1476 n = rte_eth_rx_burst(vf_port, rxq->queue_id, rx_pkts, nb_pkts);
1478 /* relabel the received mbufs */
1479 for (i = 0; i < n; i++)
1480 rx_pkts[i]->port = rxq->port_id;
1486 hn_recv_pkts(void *prxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1488 struct hn_rx_queue *rxq = prxq;
1489 struct hn_data *hv = rxq->hv;
1490 struct rte_eth_dev *vf_dev;
1493 if (unlikely(hv->closed))
1496 /* Receive from VF if present and up */
1497 vf_dev = hn_get_vf_dev(hv);
1499 /* Check for new completions */
1500 if (likely(rte_ring_count(rxq->rx_ring) < nb_pkts))
1501 hn_process_events(hv, rxq->queue_id, 0);
1503 /* Always check the vmbus path for multicast and new flows */
1504 nb_rcv = rte_ring_sc_dequeue_burst(rxq->rx_ring,
1505 (void **)rx_pkts, nb_pkts, NULL);
1507 /* If VF is available, check that as well */
1508 if (vf_dev && vf_dev->data->dev_started)
1509 nb_rcv += hn_recv_vf(vf_dev->data->port_id, rxq,
1510 rx_pkts + nb_rcv, nb_pkts - nb_rcv);
1516 hn_dev_free_queues(struct rte_eth_dev *dev)
1520 for (i = 0; i < dev->data->nb_rx_queues; i++) {
1521 struct hn_rx_queue *rxq = dev->data->rx_queues[i];
1523 hn_rx_queue_free(rxq, false);
1524 dev->data->rx_queues[i] = NULL;
1526 dev->data->nb_rx_queues = 0;
1528 for (i = 0; i < dev->data->nb_tx_queues; i++) {
1529 hn_dev_tx_queue_release(dev->data->tx_queues[i]);
1530 dev->data->tx_queues[i] = NULL;
1532 dev->data->nb_tx_queues = 0;