xref: /dpdk/drivers/net/netvsc/hn_rxtx.c (revision e11bdd37745229bf26b557305c07d118c3dbaad7)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2016-2018 Microsoft Corporation
3  * Copyright(c) 2013-2016 Brocade Communications Systems, Inc.
4  * All rights reserved.
5  */
6 
7 #include <stdint.h>
8 #include <string.h>
9 #include <stdio.h>
10 #include <errno.h>
11 #include <unistd.h>
12 #include <strings.h>
13 #include <malloc.h>
14 
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>
27 #include <rte_eal.h>
28 #include <rte_dev.h>
29 #include <rte_net.h>
30 #include <rte_bus_vmbus.h>
31 #include <rte_spinlock.h>
32 
33 #include "hn_logs.h"
34 #include "hn_var.h"
35 #include "hn_rndis.h"
36 #include "hn_nvs.h"
37 #include "ndis.h"
38 
39 #define HN_NVS_SEND_MSG_SIZE \
40 	(sizeof(struct vmbus_chanpkt_hdr) + sizeof(struct hn_nvs_rndis))
41 
42 #define HN_TXD_CACHE_SIZE	32 /* per cpu tx_descriptor pool cache */
43 #define HN_TXCOPY_THRESHOLD	512
44 
45 #define HN_RXCOPY_THRESHOLD	256
46 #define HN_RXQ_EVENT_DEFAULT	2048
47 
48 struct hn_rxinfo {
49 	uint32_t	vlan_info;
50 	uint32_t	csum_info;
51 	uint32_t	hash_info;
52 	uint32_t	hash_value;
53 };
54 
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			\
60 	(HN_RXINFO_VLAN |		\
61 	 HN_RXINFO_CSUM |		\
62 	 HN_RXINFO_HASHINF |		\
63 	 HN_RXINFO_HASHVAL)
64 
65 #define HN_NDIS_VLAN_INFO_INVALID	0xffffffff
66 #define HN_NDIS_RXCSUM_INFO_INVALID	0
67 #define HN_NDIS_HASH_INFO_INVALID	0
68 
69 /*
70  * Per-transmit book keeping.
71  * A slot in transmit ring (chim_index) is reserved for each transmit.
72  *
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.
76  *
77  *   - direct transmit where RNDIS header is in the in  rndis_pkt
78  *     mbuf is freed after transmit.
79  *
80  * Descriptors come from per-port pool which is used
81  * to limit number of outstanding requests per device.
82  */
83 struct hn_txdesc {
84 	struct rte_mbuf *m;
85 
86 	uint16_t	queue_id;
87 	uint32_t	chim_index;
88 	uint32_t	chim_size;
89 	uint32_t	data_size;
90 	uint32_t	packets;
91 
92 	struct rndis_packet_msg *rndis_pkt;
93 };
94 
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))
101 
102 #define HN_RNDIS_PKT_ALIGNED	RTE_ALIGN(HN_RNDIS_PKT_LEN, RTE_CACHE_LINE_SIZE)
103 
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)
107 
108 #define DEFAULT_TX_FREE_THRESH 32
109 
110 static void
111 hn_update_packet_stats(struct hn_stats *stats, const struct rte_mbuf *m)
112 {
113 	uint32_t s = m->pkt_len;
114 	const struct rte_ether_addr *ea;
115 
116 	if (s == 64) {
117 		stats->size_bins[1]++;
118 	} else if (s > 64 && s < 1024) {
119 		uint32_t bin;
120 
121 		/* count zeros, and offset into correct bin */
122 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
123 		stats->size_bins[bin]++;
124 	} else {
125 		if (s < 64)
126 			stats->size_bins[0]++;
127 		else if (s < 1519)
128 			stats->size_bins[6]++;
129 		else
130 			stats->size_bins[7]++;
131 	}
132 
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))
136 			stats->broadcast++;
137 		else
138 			stats->multicast++;
139 	}
140 }
141 
142 static inline unsigned int hn_rndis_pktlen(const struct rndis_packet_msg *pkt)
143 {
144 	return pkt->pktinfooffset + pkt->pktinfolen;
145 }
146 
147 static inline uint32_t
148 hn_rndis_pktmsg_offset(uint32_t ofs)
149 {
150 	return ofs - offsetof(struct rndis_packet_msg, dataoffset);
151 }
152 
153 static void hn_txd_init(struct rte_mempool *mp __rte_unused,
154 			void *opaque, void *obj, unsigned int idx)
155 {
156 	struct hn_tx_queue *txq = opaque;
157 	struct hn_txdesc *txd = obj;
158 
159 	memset(txd, 0, sizeof(*txd));
160 
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;
165 }
166 
167 int
168 hn_chim_init(struct rte_eth_dev *dev)
169 {
170 	struct hn_data *hv = dev->data->dev_private;
171 	uint32_t i, chim_bmp_size;
172 
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",
179 			     chim_bmp_size);
180 		return -1;
181 	}
182 
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");
187 		return -1;
188 	}
189 
190 	for (i = 0; i < hv->chim_cnt; i++)
191 		rte_bitmap_set(hv->chim_bmap, i);
192 
193 	return 0;
194 }
195 
196 void
197 hn_chim_uninit(struct rte_eth_dev *dev)
198 {
199 	struct hn_data *hv = dev->data->dev_private;
200 
201 	rte_bitmap_free(hv->chim_bmap);
202 	rte_free(hv->chim_bmem);
203 	hv->chim_bmem = NULL;
204 }
205 
206 static uint32_t hn_chim_alloc(struct hn_data *hv)
207 {
208 	uint32_t index = NVS_CHIM_IDX_INVALID;
209 	uint64_t slab;
210 
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);
215 
216 	return index;
217 }
218 
219 static void hn_chim_free(struct hn_data *hv, uint32_t chim_idx)
220 {
221 	if (chim_idx >= hv->chim_cnt) {
222 		PMD_DRV_LOG(ERR, "Invalid chimney index %u", chim_idx);
223 	} else {
224 		rte_spinlock_lock(&hv->chim_lock);
225 		rte_bitmap_set(hv->chim_bmap, chim_idx);
226 		rte_spinlock_unlock(&hv->chim_lock);
227 	}
228 }
229 
230 static void hn_reset_txagg(struct hn_tx_queue *txq)
231 {
232 	txq->agg_szleft = txq->agg_szmax;
233 	txq->agg_pktleft = txq->agg_pktmax;
234 	txq->agg_txd = NULL;
235 	txq->agg_prevpkt = NULL;
236 }
237 
238 int
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)
243 
244 {
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;
249 	int err = -ENOMEM;
250 
251 	PMD_INIT_FUNC_TRACE();
252 
253 	txq = rte_zmalloc_socket("HN_TXQ", sizeof(*txq), RTE_CACHE_LINE_SIZE,
254 				 socket_id);
255 	if (!txq)
256 		return -ENOMEM;
257 
258 	txq->hv = hv;
259 	txq->chan = hv->channels[queue_idx];
260 	txq->port_id = dev->data->port_id;
261 	txq->queue_id = queue_idx;
262 
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);
267 
268 	if (tx_free_thresh + 3 >= nb_desc) {
269 		PMD_INIT_LOG(ERR,
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",
272 			     nb_desc - 3,
273 			     tx_free_thresh, dev->data->port_id, queue_idx);
274 		return -EINVAL;
275 	}
276 
277 	txq->free_thresh = tx_free_thresh;
278 
279 	snprintf(name, sizeof(name),
280 		 "hn_txd_%u_%u", dev->data->port_id, queue_idx);
281 
282 	PMD_INIT_LOG(DEBUG, "TX descriptor pool %s n=%u size=%zu",
283 		     name, nb_desc, sizeof(struct hn_txdesc));
284 
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)
288 		goto error;
289 
290 	txq->txdesc_pool = rte_mempool_create(name, nb_desc,
291 					      sizeof(struct hn_txdesc),
292 					      0, 0, NULL, NULL,
293 					      hn_txd_init, txq,
294 					      dev->device->numa_node, 0);
295 	if (txq->txdesc_pool == NULL) {
296 		PMD_DRV_LOG(ERR,
297 			    "mempool %s create failed: %d", name, rte_errno);
298 		goto error;
299 	}
300 
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;
304 
305 	hn_reset_txagg(txq);
306 
307 	err = hn_vf_tx_queue_setup(dev, queue_idx, nb_desc,
308 				     socket_id, tx_conf);
309 	if (err == 0) {
310 		dev->data->tx_queues[queue_idx] = txq;
311 		return 0;
312 	}
313 
314 error:
315 	if (txq->txdesc_pool)
316 		rte_mempool_free(txq->txdesc_pool);
317 	rte_free(txq->tx_rndis);
318 	rte_free(txq);
319 	return err;
320 }
321 
322 
323 static struct hn_txdesc *hn_txd_get(struct hn_tx_queue *txq)
324 {
325 	struct hn_txdesc *txd;
326 
327 	if (rte_mempool_get(txq->txdesc_pool, (void **)&txd)) {
328 		++txq->stats.ring_full;
329 		PMD_TX_LOG(DEBUG, "tx pool exhausted!");
330 		return NULL;
331 	}
332 
333 	txd->m = NULL;
334 	txd->packets = 0;
335 	txd->data_size = 0;
336 	txd->chim_size = 0;
337 
338 	return txd;
339 }
340 
341 static void hn_txd_put(struct hn_tx_queue *txq, struct hn_txdesc *txd)
342 {
343 	rte_mempool_put(txq->txdesc_pool, txd);
344 }
345 
346 void
347 hn_dev_tx_queue_release(void *arg)
348 {
349 	struct hn_tx_queue *txq = arg;
350 
351 	PMD_INIT_FUNC_TRACE();
352 
353 	if (!txq)
354 		return;
355 
356 	if (txq->txdesc_pool)
357 		rte_mempool_free(txq->txdesc_pool);
358 
359 	rte_free(txq->tx_rndis);
360 	rte_free(txq);
361 }
362 
363 static void
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)
366 {
367 	struct hn_data *hv = dev->data->dev_private;
368 	struct hn_txdesc *txd = (struct hn_txdesc *)xactid;
369 	struct hn_tx_queue *txq;
370 
371 	/* Control packets are sent with xacid == 0 */
372 	if (!txd)
373 		return;
374 
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;
382 	} else {
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);
385 		++txq->stats.errors;
386 	}
387 
388 	if (txd->chim_index != NVS_CHIM_IDX_INVALID)
389 		hn_chim_free(hv, txd->chim_index);
390 
391 	rte_pktmbuf_free(txd->m);
392 	hn_txd_put(txq, txd);
393 }
394 
395 /* Handle transmit completion events */
396 static void
397 hn_nvs_handle_comp(struct rte_eth_dev *dev, uint16_t queue_id,
398 		   const struct vmbus_chanpkt_hdr *pkt,
399 		   const void *data)
400 {
401 	const struct hn_nvs_hdr *hdr = data;
402 
403 	switch (hdr->type) {
404 	case NVS_TYPE_RNDIS_ACK:
405 		hn_nvs_send_completed(dev, queue_id, pkt->xactid, data);
406 		break;
407 
408 	default:
409 		PMD_TX_LOG(NOTICE,
410 			   "unexpected send completion type %u",
411 			   hdr->type);
412 	}
413 }
414 
415 /* Parse per-packet info (meta data) */
416 static int
417 hn_rndis_rxinfo(const void *info_data, unsigned int info_dlen,
418 		struct hn_rxinfo *info)
419 {
420 	const struct rndis_pktinfo *pi = info_data;
421 	uint32_t mask = 0;
422 
423 	while (info_dlen != 0) {
424 		const void *data;
425 		uint32_t dlen;
426 
427 		if (unlikely(info_dlen < sizeof(*pi)))
428 			return -EINVAL;
429 
430 		if (unlikely(info_dlen < pi->size))
431 			return -EINVAL;
432 		info_dlen -= pi->size;
433 
434 		if (unlikely(pi->size & RNDIS_PKTINFO_SIZE_ALIGNMASK))
435 			return -EINVAL;
436 		if (unlikely(pi->size < pi->offset))
437 			return -EINVAL;
438 
439 		dlen = pi->size - pi->offset;
440 		data = pi->data;
441 
442 		switch (pi->type) {
443 		case NDIS_PKTINFO_TYPE_VLAN:
444 			if (unlikely(dlen < NDIS_VLAN_INFO_SIZE))
445 				return -EINVAL;
446 			info->vlan_info = *((const uint32_t *)data);
447 			mask |= HN_RXINFO_VLAN;
448 			break;
449 
450 		case NDIS_PKTINFO_TYPE_CSUM:
451 			if (unlikely(dlen < NDIS_RXCSUM_INFO_SIZE))
452 				return -EINVAL;
453 			info->csum_info = *((const uint32_t *)data);
454 			mask |= HN_RXINFO_CSUM;
455 			break;
456 
457 		case NDIS_PKTINFO_TYPE_HASHVAL:
458 			if (unlikely(dlen < NDIS_HASH_VALUE_SIZE))
459 				return -EINVAL;
460 			info->hash_value = *((const uint32_t *)data);
461 			mask |= HN_RXINFO_HASHVAL;
462 			break;
463 
464 		case NDIS_PKTINFO_TYPE_HASHINF:
465 			if (unlikely(dlen < NDIS_HASH_INFO_SIZE))
466 				return -EINVAL;
467 			info->hash_info = *((const uint32_t *)data);
468 			mask |= HN_RXINFO_HASHINF;
469 			break;
470 
471 		default:
472 			goto next;
473 		}
474 
475 		if (mask == HN_RXINFO_ALL)
476 			break; /* All found; done */
477 next:
478 		pi = (const struct rndis_pktinfo *)
479 		    ((const uint8_t *)pi + pi->size);
480 	}
481 
482 	/*
483 	 * Final fixup.
484 	 * - If there is no hash value, invalidate the hash info.
485 	 */
486 	if (!(mask & HN_RXINFO_HASHVAL))
487 		info->hash_info = HN_NDIS_HASH_INFO_INVALID;
488 	return 0;
489 }
490 
491 /*
492  * Ack the consumed RXBUF associated w/ this channel packet,
493  * so that this RXBUF can be recycled by the hypervisor.
494  */
495 static void hn_rx_buf_release(struct hn_rx_bufinfo *rxb)
496 {
497 	struct rte_mbuf_ext_shared_info *shinfo = &rxb->shinfo;
498 	struct hn_data *hv = rxb->hv;
499 
500 	if (rte_mbuf_ext_refcnt_update(shinfo, -1) == 0) {
501 		hn_nvs_ack_rxbuf(rxb->chan, rxb->xactid);
502 		--hv->rxbuf_outstanding;
503 	}
504 }
505 
506 static void hn_rx_buf_free_cb(void *buf __rte_unused, void *opaque)
507 {
508 	hn_rx_buf_release(opaque);
509 }
510 
511 static struct hn_rx_bufinfo *hn_rx_buf_init(const struct hn_rx_queue *rxq,
512 					    const struct vmbus_chanpkt_rxbuf *pkt)
513 {
514 	struct hn_rx_bufinfo *rxb;
515 
516 	rxb = rxq->hv->rxbuf_info + pkt->hdr.xactid;
517 	rxb->chan = rxq->chan;
518 	rxb->xactid = pkt->hdr.xactid;
519 	rxb->hv = rxq->hv;
520 
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);
524 	return rxb;
525 }
526 
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)
530 {
531 	struct hn_data *hv = rxq->hv;
532 	struct rte_mbuf *m;
533 
534 	m = rte_pktmbuf_alloc(rxq->mb_pool);
535 	if (unlikely(!m)) {
536 		struct rte_eth_dev *dev =
537 			&rte_eth_devices[rxq->port_id];
538 
539 		dev->data->rx_mbuf_alloc_failed++;
540 		return;
541 	}
542 
543 	/*
544 	 * For large packets, avoid copy if possible but need to keep
545 	 * some space available in receive area for later packets.
546 	 */
547 	if (dlen >= HN_RXCOPY_THRESHOLD &&
548 	    hv->rxbuf_outstanding < hv->rxbuf_section_cnt / 2) {
549 		struct rte_mbuf_ext_shared_info *shinfo;
550 		const void *rxbuf;
551 		rte_iova_t iova;
552 
553 		/*
554 		 * Build an external mbuf that points to recveive area.
555 		 * Use refcount to handle multiple packets in same
556 		 * receive buffer section.
557 		 */
558 		rxbuf = hv->rxbuf_res->addr;
559 		iova = rte_mem_virt2iova(rxbuf) + RTE_PTR_DIFF(data, rxbuf);
560 		shinfo = &rxb->shinfo;
561 
562 		if (rte_mbuf_ext_refcnt_update(shinfo, 1) == 1)
563 			++hv->rxbuf_outstanding;
564 
565 		rte_pktmbuf_attach_extbuf(m, data, iova,
566 					  dlen + headroom, shinfo);
567 		m->data_off = headroom;
568 	} else {
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);
572 			++rxq->stats.errors;
573 			return;
574 		}
575 		rte_memcpy(rte_pktmbuf_mtod(m, void *),
576 			   data + headroom, dlen);
577 	}
578 
579 	m->port = rxq->port_id;
580 	m->pkt_len = dlen;
581 	m->data_len = dlen;
582 	m->packet_type = rte_net_get_ptype(m, NULL,
583 					   RTE_PTYPE_L2_MASK |
584 					   RTE_PTYPE_L3_MASK |
585 					   RTE_PTYPE_L4_MASK);
586 
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;
590 
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");
594 			++rxq->stats.errors;
595 			rte_pktmbuf_free(m);
596 			return;
597 		}
598 	}
599 
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;
603 
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;
610 	}
611 
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;
615 	}
616 
617 	PMD_RX_LOG(DEBUG,
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);
621 
622 	++rxq->stats.packets;
623 	rxq->stats.bytes += m->pkt_len;
624 	hn_update_packet_stats(&rxq->stats, m);
625 
626 	if (unlikely(rte_ring_sp_enqueue(rxq->rx_ring, m) != 0)) {
627 		++rxq->stats.ring_full;
628 		rte_pktmbuf_free(m);
629 	}
630 }
631 
632 static void hn_rndis_rx_data(struct hn_rx_queue *rxq,
633 			     struct hn_rx_bufinfo *rxb,
634 			     void *data, uint32_t dlen)
635 {
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,
642 	};
643 	int err;
644 
645 	hn_rndis_dump(pkt);
646 
647 	if (unlikely(dlen < sizeof(*pkt)))
648 		goto error;
649 
650 	if (unlikely(dlen < pkt->len))
651 		goto error; /* truncated RNDIS from host */
652 
653 	if (unlikely(pkt->len < pkt->datalen
654 		     + pkt->oobdatalen + pkt->pktinfolen))
655 		goto error;
656 
657 	if (unlikely(pkt->datalen == 0))
658 		goto error;
659 
660 	/* Check offsets. */
661 	if (unlikely(pkt->dataoffset < RNDIS_PACKET_MSG_OFFSET_MIN))
662 		goto error;
663 
664 	if (likely(pkt->pktinfooffset > 0) &&
665 	    unlikely(pkt->pktinfooffset < RNDIS_PACKET_MSG_OFFSET_MIN ||
666 		     (pkt->pktinfooffset & RNDIS_PACKET_MSG_OFFSET_ALIGNMASK)))
667 		goto error;
668 
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;
673 
674 	if (likely(pktinfo_len > 0)) {
675 		err = hn_rndis_rxinfo((const uint8_t *)pkt + pktinfo_off,
676 				      pktinfo_len, &info);
677 		if (err)
678 			goto error;
679 	}
680 
681 	if (unlikely(data_off + data_len > pkt->len))
682 		goto error;
683 
684 	if (unlikely(data_len < RTE_ETHER_HDR_LEN))
685 		goto error;
686 
687 	hn_rxpkt(rxq, rxb, data, data_off, data_len, &info);
688 	return;
689 error:
690 	++rxq->stats.errors;
691 }
692 
693 static void
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)
696 {
697 	const struct rndis_msghdr *hdr = buf;
698 
699 	switch (hdr->type) {
700 	case RNDIS_PACKET_MSG:
701 		if (dev->data->dev_started)
702 			hn_rndis_rx_data(rxq, rxb, buf, len);
703 		break;
704 
705 	case RNDIS_INDICATE_STATUS_MSG:
706 		hn_rndis_link_status(dev, buf);
707 		break;
708 
709 	case RNDIS_INITIALIZE_CMPLT:
710 	case RNDIS_QUERY_CMPLT:
711 	case RNDIS_SET_CMPLT:
712 		hn_rndis_receive_response(rxq->hv, buf, len);
713 		break;
714 
715 	default:
716 		PMD_DRV_LOG(NOTICE,
717 			    "unexpected RNDIS message (type %#x len %u)",
718 			    hdr->type, len);
719 		break;
720 	}
721 }
722 
723 static void
724 hn_nvs_handle_rxbuf(struct rte_eth_dev *dev,
725 		    struct hn_data *hv,
726 		    struct hn_rx_queue *rxq,
727 		    const struct vmbus_chanpkt_hdr *hdr,
728 		    const void *buf)
729 {
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;
736 
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");
740 		return;
741 	}
742 
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",
746 			   nvs_hdr->type);
747 		return;
748 	}
749 
750 	hlen = vmbus_chanpkt_getlen(hdr->hlen);
751 	if (unlikely(hlen < sizeof(*pkt))) {
752 		PMD_RX_LOG(ERR, "invalid rxbuf chanpkt");
753 		return;
754 	}
755 
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",
759 			   pkt->rxbuf_id);
760 		return;
761 	}
762 
763 	count = pkt->rxbuf_cnt;
764 	if (unlikely(hlen < offsetof(struct vmbus_chanpkt_rxbuf,
765 				     rxbuf[count]))) {
766 		PMD_RX_LOG(ERR, "invalid rxbuf_cnt %u", count);
767 		return;
768 	}
769 
770 	if (pkt->hdr.xactid > hv->rxbuf_section_cnt) {
771 		PMD_RX_LOG(ERR, "invalid rxbuf section id %" PRIx64,
772 			   pkt->hdr.xactid);
773 		return;
774 	}
775 
776 	/* Setup receive buffer info to allow for callback */
777 	rxb = hn_rx_buf_init(rxq, pkt);
778 
779 	/* Each range represents 1 RNDIS pkt that contains 1 Ethernet frame */
780 	for (i = 0; i < count; ++i) {
781 		unsigned int ofs, len;
782 
783 		ofs = pkt->rxbuf[i].ofs;
784 		len = pkt->rxbuf[i].len;
785 
786 		if (unlikely(ofs + len > rxbuf_sz)) {
787 			PMD_RX_LOG(ERR,
788 				   "%uth RNDIS msg overflow ofs %u, len %u",
789 				   i, ofs, len);
790 			continue;
791 		}
792 
793 		if (unlikely(len == 0)) {
794 			PMD_RX_LOG(ERR, "%uth RNDIS msg len %u", i, len);
795 			continue;
796 		}
797 
798 		hn_rndis_receive(dev, rxq, rxb,
799 				 rxbuf + ofs, len);
800 	}
801 
802 	/* Send ACK now if external mbuf not used */
803 	hn_rx_buf_release(rxb);
804 }
805 
806 /*
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.
810  */
811 static void hn_nvs_handle_notify(struct rte_eth_dev *dev,
812 				 const struct vmbus_chanpkt_hdr *pkt,
813 				 const void *data)
814 {
815 	const struct hn_nvs_hdr *hdr = data;
816 
817 	switch (hdr->type) {
818 	case NVS_TYPE_TXTBL_NOTE:
819 		/* Transmit indirection table has locking problems
820 		 * in DPDK and therefore not implemented
821 		 */
822 		PMD_DRV_LOG(DEBUG, "host notify of transmit indirection table");
823 		break;
824 
825 	case NVS_TYPE_VFASSOC_NOTE:
826 		hn_nvs_handle_vfassoc(dev, pkt, data);
827 		break;
828 
829 	default:
830 		PMD_DRV_LOG(INFO,
831 			    "got notify, nvs type %u", hdr->type);
832 	}
833 }
834 
835 struct hn_rx_queue *hn_rx_queue_alloc(struct hn_data *hv,
836 				      uint16_t queue_id,
837 				      unsigned int socket_id)
838 {
839 	struct hn_rx_queue *rxq;
840 
841 	rxq = rte_zmalloc_socket("HN_RXQ", sizeof(*rxq),
842 				 RTE_CACHE_LINE_SIZE, socket_id);
843 	if (!rxq)
844 		return NULL;
845 
846 	rxq->hv = hv;
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) {
855 		rte_free(rxq);
856 		return NULL;
857 	}
858 
859 	return rxq;
860 }
861 
862 int
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)
868 {
869 	struct hn_data *hv = dev->data->dev_private;
870 	char ring_name[RTE_RING_NAMESIZE];
871 	struct hn_rx_queue *rxq;
872 	unsigned int count;
873 	int error = -ENOMEM;
874 
875 	PMD_INIT_FUNC_TRACE();
876 
877 	if (queue_idx == 0) {
878 		rxq = hv->primary;
879 	} else {
880 		rxq = hn_rx_queue_alloc(hv, queue_idx, socket_id);
881 		if (!rxq)
882 			return -ENOMEM;
883 	}
884 
885 	rxq->mb_pool = mp;
886 	count = rte_mempool_avail_count(mp) / dev->data->nb_rx_queues;
887 	if (nb_desc == 0 || nb_desc > count)
888 		nb_desc = count;
889 
890 	/*
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.
894 	 */
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),
899 				       socket_id, 0);
900 	if (!rxq->rx_ring)
901 		goto fail;
902 
903 	error = hn_vf_rx_queue_setup(dev, queue_idx, nb_desc,
904 				     socket_id, rx_conf, mp);
905 	if (error)
906 		goto fail;
907 
908 	dev->data->rx_queues[queue_idx] = rxq;
909 	return 0;
910 
911 fail:
912 	rte_ring_free(rxq->rx_ring);
913 	rte_free(rxq->event_buf);
914 	rte_free(rxq);
915 	return error;
916 }
917 
918 static void
919 hn_rx_queue_free(struct hn_rx_queue *rxq, bool keep_primary)
920 {
921 
922 	if (!rxq)
923 		return;
924 
925 	rte_ring_free(rxq->rx_ring);
926 	rxq->rx_ring = NULL;
927 	rxq->mb_pool = NULL;
928 
929 	hn_vf_rx_queue_release(rxq->hv, rxq->queue_id);
930 
931 	/* Keep primary queue to allow for control operations */
932 	if (keep_primary && rxq == rxq->hv->primary)
933 		return;
934 
935 	rte_free(rxq->event_buf);
936 	rte_free(rxq);
937 }
938 
939 void
940 hn_dev_rx_queue_release(void *arg)
941 {
942 	struct hn_rx_queue *rxq = arg;
943 
944 	PMD_INIT_FUNC_TRACE();
945 
946 	hn_rx_queue_free(rxq, true);
947 }
948 
949 int
950 hn_dev_tx_done_cleanup(void *arg, uint32_t free_cnt)
951 {
952 	struct hn_tx_queue *txq = arg;
953 
954 	return hn_process_events(txq->hv, txq->queue_id, free_cnt);
955 }
956 
957 /*
958  * Process pending events on the channel.
959  * Called from both Rx queue poll and Tx cleanup
960  */
961 uint32_t hn_process_events(struct hn_data *hv, uint16_t queue_id,
962 			   uint32_t tx_limit)
963 {
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;
968 	int ret = 0;
969 
970 	rxq = queue_id == 0 ? hv->primary : dev->data->rx_queues[queue_id];
971 
972 	/*
973 	 * Since channel is shared between Rx and TX queue need to have a lock
974 	 * since DPDK does not force same CPU to be used for Rx/Tx.
975 	 */
976 	if (unlikely(!rte_spinlock_trylock(&rxq->ring_lock)))
977 		return 0;
978 
979 	for (;;) {
980 		const struct vmbus_chanpkt_hdr *pkt;
981 		uint32_t len = rxq->event_sz;
982 		const void *data;
983 
984 retry:
985 		ret = rte_vmbus_chan_recv_raw(rxq->chan, rxq->event_buf, &len);
986 		if (ret == -EAGAIN)
987 			break;	/* ring is empty */
988 
989 		if (unlikely(ret == -ENOBUFS)) {
990 			/* event buffer not large enough to read ring */
991 
992 			PMD_DRV_LOG(DEBUG,
993 				    "event buffer expansion (need %u)", len);
994 			rxq->event_sz = len + len / 4;
995 			rxq->event_buf = rte_realloc(rxq->event_buf, rxq->event_sz,
996 						     RTE_CACHE_LINE_SIZE);
997 			if (rxq->event_buf)
998 				goto retry;
999 			/* out of memory, no more events now */
1000 			rxq->event_sz = 0;
1001 			break;
1002 		}
1003 
1004 		if (unlikely(ret <= 0)) {
1005 			/* This indicates a failure to communicate (or worse) */
1006 			rte_exit(EXIT_FAILURE,
1007 				 "vmbus ring buffer error: %d", ret);
1008 		}
1009 
1010 		bytes_read += ret;
1011 		pkt = (const struct vmbus_chanpkt_hdr *)rxq->event_buf;
1012 		data = (char *)rxq->event_buf + vmbus_chanpkt_getlen(pkt->hlen);
1013 
1014 		switch (pkt->type) {
1015 		case VMBUS_CHANPKT_TYPE_COMP:
1016 			++tx_done;
1017 			hn_nvs_handle_comp(dev, queue_id, pkt, data);
1018 			break;
1019 
1020 		case VMBUS_CHANPKT_TYPE_RXBUF:
1021 			hn_nvs_handle_rxbuf(dev, hv, rxq, pkt, data);
1022 			break;
1023 
1024 		case VMBUS_CHANPKT_TYPE_INBAND:
1025 			hn_nvs_handle_notify(dev, pkt, data);
1026 			break;
1027 
1028 		default:
1029 			PMD_DRV_LOG(ERR, "unknown chan pkt %u", pkt->type);
1030 			break;
1031 		}
1032 
1033 		if (tx_limit && tx_done >= tx_limit)
1034 			break;
1035 	}
1036 
1037 	if (bytes_read > 0)
1038 		rte_vmbus_chan_signal_read(rxq->chan, bytes_read);
1039 
1040 	rte_spinlock_unlock(&rxq->ring_lock);
1041 
1042 	return tx_done;
1043 }
1044 
1045 static void hn_append_to_chim(struct hn_tx_queue *txq,
1046 			      struct rndis_packet_msg *pkt,
1047 			      const struct rte_mbuf *m)
1048 {
1049 	struct hn_txdesc *txd = txq->agg_txd;
1050 	uint8_t *buf = (uint8_t *)pkt;
1051 	unsigned int data_offs;
1052 
1053 	hn_rndis_dump(pkt);
1054 
1055 	data_offs = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
1056 	txd->chim_size += pkt->len;
1057 	txd->data_size += m->pkt_len;
1058 	++txd->packets;
1059 	hn_update_packet_stats(&txq->stats, m);
1060 
1061 	for (; m; m = m->next) {
1062 		uint16_t len = rte_pktmbuf_data_len(m);
1063 
1064 		rte_memcpy(buf + data_offs,
1065 			   rte_pktmbuf_mtod(m, const char *), len);
1066 		data_offs += len;
1067 	}
1068 }
1069 
1070 /*
1071  * Send pending aggregated data in chimney buffer (if any).
1072  * Returns error if send was unsuccessful because channel ring buffer
1073  * was full.
1074  */
1075 static int hn_flush_txagg(struct hn_tx_queue *txq, bool *need_sig)
1076 
1077 {
1078 	struct hn_txdesc *txd = txq->agg_txd;
1079 	struct hn_nvs_rndis rndis;
1080 	int ret;
1081 
1082 	if (!txd)
1083 		return 0;
1084 
1085 	rndis = (struct hn_nvs_rndis) {
1086 		.type = NVS_TYPE_RNDIS,
1087 		.rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1088 		.chim_idx = txd->chim_index,
1089 		.chim_sz = txd->chim_size,
1090 	};
1091 
1092 	PMD_TX_LOG(DEBUG, "port %u:%u tx %u size %u",
1093 		   txq->port_id, txq->queue_id, txd->chim_index, txd->chim_size);
1094 
1095 	ret = hn_nvs_send(txq->chan, VMBUS_CHANPKT_FLAG_RC,
1096 			  &rndis, sizeof(rndis), (uintptr_t)txd, need_sig);
1097 
1098 	if (likely(ret == 0))
1099 		hn_reset_txagg(txq);
1100 	else
1101 		PMD_TX_LOG(NOTICE, "port %u:%u send failed: %d",
1102 			   txq->port_id, txq->queue_id, ret);
1103 
1104 	return ret;
1105 }
1106 
1107 /*
1108  * Try and find a place in a send chimney buffer to put
1109  * the small packet. If space is available, this routine
1110  * returns a pointer of where to place the data.
1111  * If no space, caller should try direct transmit.
1112  */
1113 static void *
1114 hn_try_txagg(struct hn_data *hv, struct hn_tx_queue *txq,
1115 	     struct hn_txdesc *txd, uint32_t pktsize)
1116 {
1117 	struct hn_txdesc *agg_txd = txq->agg_txd;
1118 	struct rndis_packet_msg *pkt;
1119 	void *chim;
1120 
1121 	if (agg_txd) {
1122 		unsigned int padding, olen;
1123 
1124 		/*
1125 		 * Update the previous RNDIS packet's total length,
1126 		 * it can be increased due to the mandatory alignment
1127 		 * padding for this RNDIS packet.  And update the
1128 		 * aggregating txdesc's chimney sending buffer size
1129 		 * accordingly.
1130 		 *
1131 		 * Zero-out the padding, as required by the RNDIS spec.
1132 		 */
1133 		pkt = txq->agg_prevpkt;
1134 		olen = pkt->len;
1135 		padding = RTE_ALIGN(olen, txq->agg_align) - olen;
1136 		if (padding > 0) {
1137 			agg_txd->chim_size += padding;
1138 			pkt->len += padding;
1139 			memset((uint8_t *)pkt + olen, 0, padding);
1140 		}
1141 
1142 		chim = (uint8_t *)pkt + pkt->len;
1143 		txq->agg_prevpkt = chim;
1144 		txq->agg_pktleft--;
1145 		txq->agg_szleft -= pktsize;
1146 		if (txq->agg_szleft < HN_PKTSIZE_MIN(txq->agg_align)) {
1147 			/*
1148 			 * Probably can't aggregate more packets,
1149 			 * flush this aggregating txdesc proactively.
1150 			 */
1151 			txq->agg_pktleft = 0;
1152 		}
1153 
1154 		hn_txd_put(txq, txd);
1155 		return chim;
1156 	}
1157 
1158 	txd->chim_index = hn_chim_alloc(hv);
1159 	if (txd->chim_index == NVS_CHIM_IDX_INVALID)
1160 		return NULL;
1161 
1162 	chim = (uint8_t *)hv->chim_res->addr
1163 			+ txd->chim_index * hv->chim_szmax;
1164 
1165 	txq->agg_txd = txd;
1166 	txq->agg_pktleft = txq->agg_pktmax - 1;
1167 	txq->agg_szleft = txq->agg_szmax - pktsize;
1168 	txq->agg_prevpkt = chim;
1169 
1170 	return chim;
1171 }
1172 
1173 static inline void *
1174 hn_rndis_pktinfo_append(struct rndis_packet_msg *pkt,
1175 			uint32_t pi_dlen, uint32_t pi_type)
1176 {
1177 	const uint32_t pi_size = RNDIS_PKTINFO_SIZE(pi_dlen);
1178 	struct rndis_pktinfo *pi;
1179 
1180 	/*
1181 	 * Per-packet-info does not move; it only grows.
1182 	 *
1183 	 * NOTE:
1184 	 * pktinfooffset in this phase counts from the beginning
1185 	 * of rndis_packet_msg.
1186 	 */
1187 	pi = (struct rndis_pktinfo *)((uint8_t *)pkt + hn_rndis_pktlen(pkt));
1188 
1189 	pkt->pktinfolen += pi_size;
1190 
1191 	pi->size = pi_size;
1192 	pi->type = pi_type;
1193 	pi->offset = RNDIS_PKTINFO_OFFSET;
1194 
1195 	return pi->data;
1196 }
1197 
1198 /* Put RNDIS header and packet info on packet */
1199 static void hn_encap(struct rndis_packet_msg *pkt,
1200 		     uint16_t queue_id,
1201 		     const struct rte_mbuf *m)
1202 {
1203 	unsigned int hlen = m->l2_len + m->l3_len;
1204 	uint32_t *pi_data;
1205 	uint32_t pkt_hlen;
1206 
1207 	pkt->type = RNDIS_PACKET_MSG;
1208 	pkt->len = m->pkt_len;
1209 	pkt->dataoffset = 0;
1210 	pkt->datalen = m->pkt_len;
1211 	pkt->oobdataoffset = 0;
1212 	pkt->oobdatalen = 0;
1213 	pkt->oobdataelements = 0;
1214 	pkt->pktinfooffset = sizeof(*pkt);
1215 	pkt->pktinfolen = 0;
1216 	pkt->vchandle = 0;
1217 	pkt->reserved = 0;
1218 
1219 	/*
1220 	 * Set the hash value for this packet, to the queue_id to cause
1221 	 * TX done event for this packet on the right channel.
1222 	 */
1223 	pi_data = hn_rndis_pktinfo_append(pkt, NDIS_HASH_VALUE_SIZE,
1224 					  NDIS_PKTINFO_TYPE_HASHVAL);
1225 	*pi_data = queue_id;
1226 
1227 	if (m->ol_flags & PKT_TX_VLAN_PKT) {
1228 		pi_data = hn_rndis_pktinfo_append(pkt, NDIS_VLAN_INFO_SIZE,
1229 						  NDIS_PKTINFO_TYPE_VLAN);
1230 		*pi_data = m->vlan_tci;
1231 	}
1232 
1233 	if (m->ol_flags & PKT_TX_TCP_SEG) {
1234 		pi_data = hn_rndis_pktinfo_append(pkt, NDIS_LSO2_INFO_SIZE,
1235 						  NDIS_PKTINFO_TYPE_LSO);
1236 
1237 		if (m->ol_flags & PKT_TX_IPV6) {
1238 			*pi_data = NDIS_LSO2_INFO_MAKEIPV6(hlen,
1239 							   m->tso_segsz);
1240 		} else {
1241 			*pi_data = NDIS_LSO2_INFO_MAKEIPV4(hlen,
1242 							   m->tso_segsz);
1243 		}
1244 	} else if (m->ol_flags &
1245 		   (PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM | PKT_TX_IP_CKSUM)) {
1246 		pi_data = hn_rndis_pktinfo_append(pkt, NDIS_TXCSUM_INFO_SIZE,
1247 						  NDIS_PKTINFO_TYPE_CSUM);
1248 		*pi_data = 0;
1249 
1250 		if (m->ol_flags & PKT_TX_IPV6)
1251 			*pi_data |= NDIS_TXCSUM_INFO_IPV6;
1252 		if (m->ol_flags & PKT_TX_IPV4) {
1253 			*pi_data |= NDIS_TXCSUM_INFO_IPV4;
1254 
1255 			if (m->ol_flags & PKT_TX_IP_CKSUM)
1256 				*pi_data |= NDIS_TXCSUM_INFO_IPCS;
1257 		}
1258 
1259 		if (m->ol_flags & PKT_TX_TCP_CKSUM)
1260 			*pi_data |= NDIS_TXCSUM_INFO_MKTCPCS(hlen);
1261 		else if (m->ol_flags & PKT_TX_UDP_CKSUM)
1262 			*pi_data |= NDIS_TXCSUM_INFO_MKUDPCS(hlen);
1263 	}
1264 
1265 	pkt_hlen = pkt->pktinfooffset + pkt->pktinfolen;
1266 	/* Fixup RNDIS packet message total length */
1267 	pkt->len += pkt_hlen;
1268 
1269 	/* Convert RNDIS packet message offsets */
1270 	pkt->dataoffset = hn_rndis_pktmsg_offset(pkt_hlen);
1271 	pkt->pktinfooffset = hn_rndis_pktmsg_offset(pkt->pktinfooffset);
1272 }
1273 
1274 /* How many scatter gather list elements ar needed */
1275 static unsigned int hn_get_slots(const struct rte_mbuf *m)
1276 {
1277 	unsigned int slots = 1; /* for RNDIS header */
1278 
1279 	while (m) {
1280 		unsigned int size = rte_pktmbuf_data_len(m);
1281 		unsigned int offs = rte_mbuf_data_iova(m) & PAGE_MASK;
1282 
1283 		slots += (offs + size + PAGE_SIZE - 1) / PAGE_SIZE;
1284 		m = m->next;
1285 	}
1286 
1287 	return slots;
1288 }
1289 
1290 /* Build scatter gather list from chained mbuf */
1291 static unsigned int hn_fill_sg(struct vmbus_gpa *sg,
1292 			       const struct rte_mbuf *m)
1293 {
1294 	unsigned int segs = 0;
1295 
1296 	while (m) {
1297 		rte_iova_t addr = rte_mbuf_data_iova(m);
1298 		unsigned int page = addr / PAGE_SIZE;
1299 		unsigned int offset = addr & PAGE_MASK;
1300 		unsigned int len = rte_pktmbuf_data_len(m);
1301 
1302 		while (len > 0) {
1303 			unsigned int bytes = RTE_MIN(len, PAGE_SIZE - offset);
1304 
1305 			sg[segs].page = page;
1306 			sg[segs].ofs = offset;
1307 			sg[segs].len = bytes;
1308 			segs++;
1309 
1310 			++page;
1311 			offset = 0;
1312 			len -= bytes;
1313 		}
1314 		m = m->next;
1315 	}
1316 
1317 	return segs;
1318 }
1319 
1320 /* Transmit directly from mbuf */
1321 static int hn_xmit_sg(struct hn_tx_queue *txq,
1322 		      const struct hn_txdesc *txd, const struct rte_mbuf *m,
1323 		      bool *need_sig)
1324 {
1325 	struct vmbus_gpa sg[hn_get_slots(m)];
1326 	struct hn_nvs_rndis nvs_rndis = {
1327 		.type = NVS_TYPE_RNDIS,
1328 		.rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1329 		.chim_sz = txd->chim_size,
1330 	};
1331 	rte_iova_t addr;
1332 	unsigned int segs;
1333 
1334 	/* attach aggregation data if present */
1335 	if (txd->chim_size > 0)
1336 		nvs_rndis.chim_idx = txd->chim_index;
1337 	else
1338 		nvs_rndis.chim_idx = NVS_CHIM_IDX_INVALID;
1339 
1340 	hn_rndis_dump(txd->rndis_pkt);
1341 
1342 	/* pass IOVA of rndis header in first segment */
1343 	addr = rte_malloc_virt2iova(txd->rndis_pkt);
1344 	if (unlikely(addr == RTE_BAD_IOVA)) {
1345 		PMD_DRV_LOG(ERR, "RNDIS transmit can not get iova");
1346 		return -EINVAL;
1347 	}
1348 
1349 	sg[0].page = addr / PAGE_SIZE;
1350 	sg[0].ofs = addr & PAGE_MASK;
1351 	sg[0].len = RNDIS_PACKET_MSG_OFFSET_ABS(hn_rndis_pktlen(txd->rndis_pkt));
1352 	segs = 1;
1353 
1354 	hn_update_packet_stats(&txq->stats, m);
1355 
1356 	segs += hn_fill_sg(sg + 1, m);
1357 
1358 	PMD_TX_LOG(DEBUG, "port %u:%u tx %u segs %u size %u",
1359 		   txq->port_id, txq->queue_id, txd->chim_index,
1360 		   segs, nvs_rndis.chim_sz);
1361 
1362 	return hn_nvs_send_sglist(txq->chan, sg, segs,
1363 				  &nvs_rndis, sizeof(nvs_rndis),
1364 				  (uintptr_t)txd, need_sig);
1365 }
1366 
1367 uint16_t
1368 hn_xmit_pkts(void *ptxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1369 {
1370 	struct hn_tx_queue *txq = ptxq;
1371 	uint16_t queue_id = txq->queue_id;
1372 	struct hn_data *hv = txq->hv;
1373 	struct rte_eth_dev *vf_dev;
1374 	bool need_sig = false;
1375 	uint16_t nb_tx, avail;
1376 	int ret;
1377 
1378 	if (unlikely(hv->closed))
1379 		return 0;
1380 
1381 	/* Transmit over VF if present and up */
1382 	vf_dev = hn_get_vf_dev(hv);
1383 
1384 	if (vf_dev && vf_dev->data->dev_started) {
1385 		void *sub_q = vf_dev->data->tx_queues[queue_id];
1386 
1387 		return (*vf_dev->tx_pkt_burst)(sub_q, tx_pkts, nb_pkts);
1388 	}
1389 
1390 	avail = rte_mempool_avail_count(txq->txdesc_pool);
1391 	if (nb_pkts > avail || avail <= txq->free_thresh)
1392 		hn_process_events(hv, txq->queue_id, 0);
1393 
1394 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1395 		struct rte_mbuf *m = tx_pkts[nb_tx];
1396 		uint32_t pkt_size = m->pkt_len + HN_RNDIS_PKT_LEN;
1397 		struct rndis_packet_msg *pkt;
1398 		struct hn_txdesc *txd;
1399 
1400 		txd = hn_txd_get(txq);
1401 		if (txd == NULL)
1402 			break;
1403 
1404 		/* For small packets aggregate them in chimney buffer */
1405 		if (m->pkt_len < HN_TXCOPY_THRESHOLD && pkt_size <= txq->agg_szmax) {
1406 			/* If this packet will not fit, then flush  */
1407 			if (txq->agg_pktleft == 0 ||
1408 			    RTE_ALIGN(pkt_size, txq->agg_align) > txq->agg_szleft) {
1409 				if (hn_flush_txagg(txq, &need_sig))
1410 					goto fail;
1411 			}
1412 
1413 
1414 			pkt = hn_try_txagg(hv, txq, txd, pkt_size);
1415 			if (unlikely(!pkt))
1416 				break;
1417 
1418 			hn_encap(pkt, queue_id, m);
1419 			hn_append_to_chim(txq, pkt, m);
1420 
1421 			rte_pktmbuf_free(m);
1422 
1423 			/* if buffer is full, flush */
1424 			if (txq->agg_pktleft == 0 &&
1425 			    hn_flush_txagg(txq, &need_sig))
1426 				goto fail;
1427 		} else {
1428 			/* Send any outstanding packets in buffer */
1429 			if (txq->agg_txd && hn_flush_txagg(txq, &need_sig))
1430 				goto fail;
1431 
1432 			pkt = txd->rndis_pkt;
1433 			txd->m = m;
1434 			txd->data_size = m->pkt_len;
1435 			++txd->packets;
1436 
1437 			hn_encap(pkt, queue_id, m);
1438 
1439 			ret = hn_xmit_sg(txq, txd, m, &need_sig);
1440 			if (unlikely(ret != 0)) {
1441 				PMD_TX_LOG(NOTICE, "sg send failed: %d", ret);
1442 				++txq->stats.errors;
1443 				hn_txd_put(txq, txd);
1444 				goto fail;
1445 			}
1446 		}
1447 	}
1448 
1449 	/* If partial buffer left, then try and send it.
1450 	 * if that fails, then reuse it on next send.
1451 	 */
1452 	hn_flush_txagg(txq, &need_sig);
1453 
1454 fail:
1455 	if (need_sig)
1456 		rte_vmbus_chan_signal_tx(txq->chan);
1457 
1458 	return nb_tx;
1459 }
1460 
1461 static uint16_t
1462 hn_recv_vf(uint16_t vf_port, const struct hn_rx_queue *rxq,
1463 	   struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1464 {
1465 	uint16_t i, n;
1466 
1467 	if (unlikely(nb_pkts == 0))
1468 		return 0;
1469 
1470 	n = rte_eth_rx_burst(vf_port, rxq->queue_id, rx_pkts, nb_pkts);
1471 
1472 	/* relabel the received mbufs */
1473 	for (i = 0; i < n; i++)
1474 		rx_pkts[i]->port = rxq->port_id;
1475 
1476 	return n;
1477 }
1478 
1479 uint16_t
1480 hn_recv_pkts(void *prxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1481 {
1482 	struct hn_rx_queue *rxq = prxq;
1483 	struct hn_data *hv = rxq->hv;
1484 	struct rte_eth_dev *vf_dev;
1485 	uint16_t nb_rcv;
1486 
1487 	if (unlikely(hv->closed))
1488 		return 0;
1489 
1490 	/* Receive from VF if present and up */
1491 	vf_dev = hn_get_vf_dev(hv);
1492 
1493 	/* Check for new completions */
1494 	if (likely(rte_ring_count(rxq->rx_ring) < nb_pkts))
1495 		hn_process_events(hv, rxq->queue_id, 0);
1496 
1497 	/* Always check the vmbus path for multicast and new flows */
1498 	nb_rcv = rte_ring_sc_dequeue_burst(rxq->rx_ring,
1499 					   (void **)rx_pkts, nb_pkts, NULL);
1500 
1501 	/* If VF is available, check that as well */
1502 	if (vf_dev && vf_dev->data->dev_started)
1503 		nb_rcv += hn_recv_vf(vf_dev->data->port_id, rxq,
1504 				     rx_pkts + nb_rcv, nb_pkts - nb_rcv);
1505 
1506 	return nb_rcv;
1507 }
1508 
1509 void
1510 hn_dev_free_queues(struct rte_eth_dev *dev)
1511 {
1512 	unsigned int i;
1513 
1514 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
1515 		struct hn_rx_queue *rxq = dev->data->rx_queues[i];
1516 
1517 		hn_rx_queue_free(rxq, false);
1518 		dev->data->rx_queues[i] = NULL;
1519 	}
1520 	dev->data->nb_rx_queues = 0;
1521 
1522 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
1523 		hn_dev_tx_queue_release(dev->data->tx_queues[i]);
1524 		dev->data->tx_queues[i] = NULL;
1525 	}
1526 	dev->data->nb_tx_queues = 0;
1527 }
1528