xref: /dpdk/drivers/net/netvsc/hn_rxtx.c (revision 2d0c29a37a9c080c1cccb1ad7941aba2ccf5437e)
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_branch_prediction.h>
22 #include <rte_ether.h>
23 #include <rte_common.h>
24 #include <rte_errno.h>
25 #include <rte_memory.h>
26 #include <rte_eal.h>
27 #include <rte_dev.h>
28 #include <rte_net.h>
29 #include <rte_bus_vmbus.h>
30 #include <rte_spinlock.h>
31 
32 #include "hn_logs.h"
33 #include "hn_var.h"
34 #include "hn_rndis.h"
35 #include "hn_nvs.h"
36 #include "ndis.h"
37 
38 #define HN_NVS_SEND_MSG_SIZE \
39 	(sizeof(struct vmbus_chanpkt_hdr) + sizeof(struct hn_nvs_rndis))
40 
41 #define HN_TXD_CACHE_SIZE	32 /* per cpu tx_descriptor pool cache */
42 #define HN_TXCOPY_THRESHOLD	512
43 
44 #define HN_RXCOPY_THRESHOLD	256
45 #define HN_RXQ_EVENT_DEFAULT	2048
46 
47 struct hn_rxinfo {
48 	uint32_t	vlan_info;
49 	uint32_t	csum_info;
50 	uint32_t	hash_info;
51 	uint32_t	hash_value;
52 };
53 
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			\
59 	(HN_RXINFO_VLAN |		\
60 	 HN_RXINFO_CSUM |		\
61 	 HN_RXINFO_HASHINF |		\
62 	 HN_RXINFO_HASHVAL)
63 
64 #define HN_NDIS_VLAN_INFO_INVALID	0xffffffff
65 #define HN_NDIS_RXCSUM_INFO_INVALID	0
66 #define HN_NDIS_HASH_INFO_INVALID	0
67 
68 /*
69  * Per-transmit book keeping.
70  * A slot in transmit ring (chim_index) is reserved for each transmit.
71  *
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.
75  *
76  *   - direct transmit where RNDIS header is in the in  rndis_pkt
77  *     mbuf is freed after transmit.
78  *
79  * Descriptors come from per-port pool which is used
80  * to limit number of outstanding requests per device.
81  */
82 struct hn_txdesc {
83 	struct rte_mbuf *m;
84 
85 	uint16_t	queue_id;
86 	uint16_t	chim_index;
87 	uint32_t	chim_size;
88 	uint32_t	data_size;
89 	uint32_t	packets;
90 
91 	struct rndis_packet_msg *rndis_pkt;
92 };
93 
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))
100 
101 /* Minimum space required for a packet */
102 #define HN_PKTSIZE_MIN(align) \
103 	RTE_ALIGN(ETHER_MIN_LEN + HN_RNDIS_PKT_LEN, align)
104 
105 #define DEFAULT_TX_FREE_THRESH 32U
106 
107 static void
108 hn_update_packet_stats(struct hn_stats *stats, const struct rte_mbuf *m)
109 {
110 	uint32_t s = m->pkt_len;
111 	const struct ether_addr *ea;
112 
113 	if (s == 64) {
114 		stats->size_bins[1]++;
115 	} else if (s > 64 && s < 1024) {
116 		uint32_t bin;
117 
118 		/* count zeros, and offset into correct bin */
119 		bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
120 		stats->size_bins[bin]++;
121 	} else {
122 		if (s < 64)
123 			stats->size_bins[0]++;
124 		else if (s < 1519)
125 			stats->size_bins[6]++;
126 		else
127 			stats->size_bins[7]++;
128 	}
129 
130 	ea = rte_pktmbuf_mtod(m, const struct ether_addr *);
131 	if (is_multicast_ether_addr(ea)) {
132 		if (is_broadcast_ether_addr(ea))
133 			stats->broadcast++;
134 		else
135 			stats->multicast++;
136 	}
137 }
138 
139 static inline unsigned int hn_rndis_pktlen(const struct rndis_packet_msg *pkt)
140 {
141 	return pkt->pktinfooffset + pkt->pktinfolen;
142 }
143 
144 static inline uint32_t
145 hn_rndis_pktmsg_offset(uint32_t ofs)
146 {
147 	return ofs - offsetof(struct rndis_packet_msg, dataoffset);
148 }
149 
150 static void hn_txd_init(struct rte_mempool *mp __rte_unused,
151 			void *opaque, void *obj, unsigned int idx)
152 {
153 	struct hn_txdesc *txd = obj;
154 	struct rte_eth_dev *dev = opaque;
155 	struct rndis_packet_msg *pkt;
156 
157 	memset(txd, 0, sizeof(*txd));
158 	txd->chim_index = idx;
159 
160 	pkt = rte_malloc_socket("RNDIS_TX", HN_RNDIS_PKT_LEN,
161 				rte_align32pow2(HN_RNDIS_PKT_LEN),
162 				dev->device->numa_node);
163 	if (!pkt)
164 		rte_exit(EXIT_FAILURE, "can not allocate RNDIS header");
165 
166 	txd->rndis_pkt = pkt;
167 }
168 
169 /*
170  * Unlike Linux and FreeBSD, this driver uses a mempool
171  * to limit outstanding transmits and reserve buffers
172  */
173 int
174 hn_tx_pool_init(struct rte_eth_dev *dev)
175 {
176 	struct hn_data *hv = dev->data->dev_private;
177 	char name[RTE_MEMPOOL_NAMESIZE];
178 	struct rte_mempool *mp;
179 
180 	snprintf(name, sizeof(name),
181 		 "hn_txd_%u", dev->data->port_id);
182 
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);
186 
187 	mp = rte_mempool_create(name, hv->chim_cnt, sizeof(struct hn_txdesc),
188 				HN_TXD_CACHE_SIZE, 0,
189 				NULL, NULL,
190 				hn_txd_init, dev,
191 				dev->device->numa_node, 0);
192 	if (!mp) {
193 		PMD_DRV_LOG(ERR,
194 			    "mempool %s create failed: %d", name, rte_errno);
195 		return -rte_errno;
196 	}
197 
198 	hv->tx_pool = mp;
199 	return 0;
200 }
201 
202 void
203 hn_tx_pool_uninit(struct rte_eth_dev *dev)
204 {
205 	struct hn_data *hv = dev->data->dev_private;
206 
207 	if (hv->tx_pool) {
208 		rte_mempool_free(hv->tx_pool);
209 		hv->tx_pool = NULL;
210 	}
211 }
212 
213 static void hn_reset_txagg(struct hn_tx_queue *txq)
214 {
215 	txq->agg_szleft = txq->agg_szmax;
216 	txq->agg_pktleft = txq->agg_pktmax;
217 	txq->agg_txd = NULL;
218 	txq->agg_prevpkt = NULL;
219 }
220 
221 int
222 hn_dev_tx_queue_setup(struct rte_eth_dev *dev,
223 		      uint16_t queue_idx, uint16_t nb_desc __rte_unused,
224 		      unsigned int socket_id,
225 		      const struct rte_eth_txconf *tx_conf)
226 
227 {
228 	struct hn_data *hv = dev->data->dev_private;
229 	struct hn_tx_queue *txq;
230 	uint32_t tx_free_thresh;
231 	int err;
232 
233 	PMD_INIT_FUNC_TRACE();
234 
235 	txq = rte_zmalloc_socket("HN_TXQ", sizeof(*txq), RTE_CACHE_LINE_SIZE,
236 				 socket_id);
237 	if (!txq)
238 		return -ENOMEM;
239 
240 	txq->hv = hv;
241 	txq->chan = hv->channels[queue_idx];
242 	txq->port_id = dev->data->port_id;
243 	txq->queue_id = queue_idx;
244 
245 	tx_free_thresh = tx_conf->tx_free_thresh;
246 	if (tx_free_thresh == 0)
247 		tx_free_thresh = RTE_MIN(hv->chim_cnt / 4,
248 					 DEFAULT_TX_FREE_THRESH);
249 
250 	if (tx_free_thresh >= hv->chim_cnt - 3)
251 		tx_free_thresh = hv->chim_cnt - 3;
252 
253 	txq->free_thresh = tx_free_thresh;
254 
255 	txq->agg_szmax  = RTE_MIN(hv->chim_szmax, hv->rndis_agg_size);
256 	txq->agg_pktmax = hv->rndis_agg_pkts;
257 	txq->agg_align  = hv->rndis_agg_align;
258 
259 	hn_reset_txagg(txq);
260 
261 	err = hn_vf_tx_queue_setup(dev, queue_idx, nb_desc,
262 				     socket_id, tx_conf);
263 	if (err) {
264 		rte_free(txq);
265 		return err;
266 	}
267 
268 	dev->data->tx_queues[queue_idx] = txq;
269 	return 0;
270 }
271 
272 void
273 hn_dev_tx_queue_release(void *arg)
274 {
275 	struct hn_tx_queue *txq = arg;
276 	struct hn_txdesc *txd;
277 
278 	PMD_INIT_FUNC_TRACE();
279 
280 	if (!txq)
281 		return;
282 
283 	/* If any pending data is still present just drop it */
284 	txd = txq->agg_txd;
285 	if (txd)
286 		rte_mempool_put(txq->hv->tx_pool, txd);
287 
288 	rte_free(txq);
289 }
290 
291 static void
292 hn_nvs_send_completed(struct rte_eth_dev *dev, uint16_t queue_id,
293 		      unsigned long xactid, const struct hn_nvs_rndis_ack *ack)
294 {
295 	struct hn_txdesc *txd = (struct hn_txdesc *)xactid;
296 	struct hn_tx_queue *txq;
297 
298 	/* Control packets are sent with xacid == 0 */
299 	if (!txd)
300 		return;
301 
302 	txq = dev->data->tx_queues[queue_id];
303 	if (likely(ack->status == NVS_STATUS_OK)) {
304 		PMD_TX_LOG(DEBUG, "port %u:%u complete tx %u packets %u bytes %u",
305 			   txq->port_id, txq->queue_id, txd->chim_index,
306 			   txd->packets, txd->data_size);
307 		txq->stats.bytes += txd->data_size;
308 		txq->stats.packets += txd->packets;
309 	} else {
310 		PMD_TX_LOG(NOTICE, "port %u:%u complete tx %u failed status %u",
311 			   txq->port_id, txq->queue_id, txd->chim_index, ack->status);
312 		++txq->stats.errors;
313 	}
314 
315 	rte_pktmbuf_free(txd->m);
316 
317 	rte_mempool_put(txq->hv->tx_pool, txd);
318 }
319 
320 /* Handle transmit completion events */
321 static void
322 hn_nvs_handle_comp(struct rte_eth_dev *dev, uint16_t queue_id,
323 		   const struct vmbus_chanpkt_hdr *pkt,
324 		   const void *data)
325 {
326 	const struct hn_nvs_hdr *hdr = data;
327 
328 	switch (hdr->type) {
329 	case NVS_TYPE_RNDIS_ACK:
330 		hn_nvs_send_completed(dev, queue_id, pkt->xactid, data);
331 		break;
332 
333 	default:
334 		PMD_TX_LOG(NOTICE,
335 			   "unexpected send completion type %u",
336 			   hdr->type);
337 	}
338 }
339 
340 /* Parse per-packet info (meta data) */
341 static int
342 hn_rndis_rxinfo(const void *info_data, unsigned int info_dlen,
343 		struct hn_rxinfo *info)
344 {
345 	const struct rndis_pktinfo *pi = info_data;
346 	uint32_t mask = 0;
347 
348 	while (info_dlen != 0) {
349 		const void *data;
350 		uint32_t dlen;
351 
352 		if (unlikely(info_dlen < sizeof(*pi)))
353 			return -EINVAL;
354 
355 		if (unlikely(info_dlen < pi->size))
356 			return -EINVAL;
357 		info_dlen -= pi->size;
358 
359 		if (unlikely(pi->size & RNDIS_PKTINFO_SIZE_ALIGNMASK))
360 			return -EINVAL;
361 		if (unlikely(pi->size < pi->offset))
362 			return -EINVAL;
363 
364 		dlen = pi->size - pi->offset;
365 		data = pi->data;
366 
367 		switch (pi->type) {
368 		case NDIS_PKTINFO_TYPE_VLAN:
369 			if (unlikely(dlen < NDIS_VLAN_INFO_SIZE))
370 				return -EINVAL;
371 			info->vlan_info = *((const uint32_t *)data);
372 			mask |= HN_RXINFO_VLAN;
373 			break;
374 
375 		case NDIS_PKTINFO_TYPE_CSUM:
376 			if (unlikely(dlen < NDIS_RXCSUM_INFO_SIZE))
377 				return -EINVAL;
378 			info->csum_info = *((const uint32_t *)data);
379 			mask |= HN_RXINFO_CSUM;
380 			break;
381 
382 		case NDIS_PKTINFO_TYPE_HASHVAL:
383 			if (unlikely(dlen < NDIS_HASH_VALUE_SIZE))
384 				return -EINVAL;
385 			info->hash_value = *((const uint32_t *)data);
386 			mask |= HN_RXINFO_HASHVAL;
387 			break;
388 
389 		case NDIS_PKTINFO_TYPE_HASHINF:
390 			if (unlikely(dlen < NDIS_HASH_INFO_SIZE))
391 				return -EINVAL;
392 			info->hash_info = *((const uint32_t *)data);
393 			mask |= HN_RXINFO_HASHINF;
394 			break;
395 
396 		default:
397 			goto next;
398 		}
399 
400 		if (mask == HN_RXINFO_ALL)
401 			break; /* All found; done */
402 next:
403 		pi = (const struct rndis_pktinfo *)
404 		    ((const uint8_t *)pi + pi->size);
405 	}
406 
407 	/*
408 	 * Final fixup.
409 	 * - If there is no hash value, invalidate the hash info.
410 	 */
411 	if (!(mask & HN_RXINFO_HASHVAL))
412 		info->hash_info = HN_NDIS_HASH_INFO_INVALID;
413 	return 0;
414 }
415 
416 /*
417  * Ack the consumed RXBUF associated w/ this channel packet,
418  * so that this RXBUF can be recycled by the hypervisor.
419  */
420 static void hn_rx_buf_release(struct hn_rx_bufinfo *rxb)
421 {
422 	struct rte_mbuf_ext_shared_info *shinfo = &rxb->shinfo;
423 	struct hn_data *hv = rxb->hv;
424 
425 	if (rte_mbuf_ext_refcnt_update(shinfo, -1) == 0) {
426 		hn_nvs_ack_rxbuf(rxb->chan, rxb->xactid);
427 		--hv->rxbuf_outstanding;
428 	}
429 }
430 
431 static void hn_rx_buf_free_cb(void *buf __rte_unused, void *opaque)
432 {
433 	hn_rx_buf_release(opaque);
434 }
435 
436 static struct hn_rx_bufinfo *hn_rx_buf_init(const struct hn_rx_queue *rxq,
437 					    const struct vmbus_chanpkt_rxbuf *pkt)
438 {
439 	struct hn_rx_bufinfo *rxb;
440 
441 	rxb = rxq->hv->rxbuf_info + pkt->hdr.xactid;
442 	rxb->chan = rxq->chan;
443 	rxb->xactid = pkt->hdr.xactid;
444 	rxb->hv = rxq->hv;
445 
446 	rxb->shinfo.free_cb = hn_rx_buf_free_cb;
447 	rxb->shinfo.fcb_opaque = rxb;
448 	rte_mbuf_ext_refcnt_set(&rxb->shinfo, 1);
449 	return rxb;
450 }
451 
452 static void hn_rxpkt(struct hn_rx_queue *rxq, struct hn_rx_bufinfo *rxb,
453 		     uint8_t *data, unsigned int headroom, unsigned int dlen,
454 		     const struct hn_rxinfo *info)
455 {
456 	struct hn_data *hv = rxq->hv;
457 	struct rte_mbuf *m;
458 
459 	m = rte_pktmbuf_alloc(rxq->mb_pool);
460 	if (unlikely(!m)) {
461 		struct rte_eth_dev *dev =
462 			&rte_eth_devices[rxq->port_id];
463 
464 		dev->data->rx_mbuf_alloc_failed++;
465 		return;
466 	}
467 
468 	/*
469 	 * For large packets, avoid copy if possible but need to keep
470 	 * some space available in receive area for later packets.
471 	 */
472 	if (dlen >= HN_RXCOPY_THRESHOLD &&
473 	    hv->rxbuf_outstanding < hv->rxbuf_section_cnt / 2) {
474 		struct rte_mbuf_ext_shared_info *shinfo;
475 		const void *rxbuf;
476 		rte_iova_t iova;
477 
478 		/*
479 		 * Build an external mbuf that points to recveive area.
480 		 * Use refcount to handle multiple packets in same
481 		 * receive buffer section.
482 		 */
483 		rxbuf = hv->rxbuf_res->addr;
484 		iova = rte_mem_virt2iova(rxbuf) + RTE_PTR_DIFF(data, rxbuf);
485 		shinfo = &rxb->shinfo;
486 
487 		if (rte_mbuf_ext_refcnt_update(shinfo, 1) == 1)
488 			++hv->rxbuf_outstanding;
489 
490 		rte_pktmbuf_attach_extbuf(m, data, iova,
491 					  dlen + headroom, shinfo);
492 		m->data_off = headroom;
493 	} else {
494 		/* Mbuf's in pool must be large enough to hold small packets */
495 		if (unlikely(rte_pktmbuf_tailroom(m) < dlen)) {
496 			rte_pktmbuf_free_seg(m);
497 			++rxq->stats.errors;
498 			return;
499 		}
500 		rte_memcpy(rte_pktmbuf_mtod(m, void *),
501 			   data + headroom, dlen);
502 	}
503 
504 	m->port = rxq->port_id;
505 	m->pkt_len = dlen;
506 	m->data_len = dlen;
507 	m->packet_type = rte_net_get_ptype(m, NULL,
508 					   RTE_PTYPE_L2_MASK |
509 					   RTE_PTYPE_L3_MASK |
510 					   RTE_PTYPE_L4_MASK);
511 
512 	if (info->vlan_info != HN_NDIS_VLAN_INFO_INVALID) {
513 		m->vlan_tci = info->vlan_info;
514 		m->ol_flags |= PKT_RX_VLAN_STRIPPED | PKT_RX_VLAN;
515 
516 		/* NDIS always strips tag, put it back if necessary */
517 		if (!hv->vlan_strip && rte_vlan_insert(&m)) {
518 			PMD_DRV_LOG(DEBUG, "vlan insert failed");
519 			++rxq->stats.errors;
520 			rte_pktmbuf_free(m);
521 			return;
522 		}
523 	}
524 
525 	if (info->csum_info != HN_NDIS_RXCSUM_INFO_INVALID) {
526 		if (info->csum_info & NDIS_RXCSUM_INFO_IPCS_OK)
527 			m->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
528 
529 		if (info->csum_info & (NDIS_RXCSUM_INFO_UDPCS_OK
530 				       | NDIS_RXCSUM_INFO_TCPCS_OK))
531 			m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
532 		else if (info->csum_info & (NDIS_RXCSUM_INFO_TCPCS_FAILED
533 					    | NDIS_RXCSUM_INFO_UDPCS_FAILED))
534 			m->ol_flags |= PKT_RX_L4_CKSUM_BAD;
535 	}
536 
537 	if (info->hash_info != HN_NDIS_HASH_INFO_INVALID) {
538 		m->ol_flags |= PKT_RX_RSS_HASH;
539 		m->hash.rss = info->hash_value;
540 	}
541 
542 	PMD_RX_LOG(DEBUG,
543 		   "port %u:%u RX id %"PRIu64" size %u type %#x ol_flags %#"PRIx64,
544 		   rxq->port_id, rxq->queue_id, rxb->xactid,
545 		   m->pkt_len, m->packet_type, m->ol_flags);
546 
547 	++rxq->stats.packets;
548 	rxq->stats.bytes += m->pkt_len;
549 	hn_update_packet_stats(&rxq->stats, m);
550 
551 	if (unlikely(rte_ring_sp_enqueue(rxq->rx_ring, m) != 0)) {
552 		++rxq->stats.ring_full;
553 		rte_pktmbuf_free(m);
554 	}
555 }
556 
557 static void hn_rndis_rx_data(struct hn_rx_queue *rxq,
558 			     struct hn_rx_bufinfo *rxb,
559 			     void *data, uint32_t dlen)
560 {
561 	unsigned int data_off, data_len, pktinfo_off, pktinfo_len;
562 	const struct rndis_packet_msg *pkt = data;
563 	struct hn_rxinfo info = {
564 		.vlan_info = HN_NDIS_VLAN_INFO_INVALID,
565 		.csum_info = HN_NDIS_RXCSUM_INFO_INVALID,
566 		.hash_info = HN_NDIS_HASH_INFO_INVALID,
567 	};
568 	int err;
569 
570 	hn_rndis_dump(pkt);
571 
572 	if (unlikely(dlen < sizeof(*pkt)))
573 		goto error;
574 
575 	if (unlikely(dlen < pkt->len))
576 		goto error; /* truncated RNDIS from host */
577 
578 	if (unlikely(pkt->len < pkt->datalen
579 		     + pkt->oobdatalen + pkt->pktinfolen))
580 		goto error;
581 
582 	if (unlikely(pkt->datalen == 0))
583 		goto error;
584 
585 	/* Check offsets. */
586 	if (unlikely(pkt->dataoffset < RNDIS_PACKET_MSG_OFFSET_MIN))
587 		goto error;
588 
589 	if (likely(pkt->pktinfooffset > 0) &&
590 	    unlikely(pkt->pktinfooffset < RNDIS_PACKET_MSG_OFFSET_MIN ||
591 		     (pkt->pktinfooffset & RNDIS_PACKET_MSG_OFFSET_ALIGNMASK)))
592 		goto error;
593 
594 	data_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
595 	data_len = pkt->datalen;
596 	pktinfo_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->pktinfooffset);
597 	pktinfo_len = pkt->pktinfolen;
598 
599 	if (likely(pktinfo_len > 0)) {
600 		err = hn_rndis_rxinfo((const uint8_t *)pkt + pktinfo_off,
601 				      pktinfo_len, &info);
602 		if (err)
603 			goto error;
604 	}
605 
606 	if (unlikely(data_off + data_len > pkt->len))
607 		goto error;
608 
609 	if (unlikely(data_len < ETHER_HDR_LEN))
610 		goto error;
611 
612 	hn_rxpkt(rxq, rxb, data, data_off, data_len, &info);
613 	return;
614 error:
615 	++rxq->stats.errors;
616 }
617 
618 static void
619 hn_rndis_receive(struct rte_eth_dev *dev, struct hn_rx_queue *rxq,
620 		 struct hn_rx_bufinfo *rxb, void *buf, uint32_t len)
621 {
622 	const struct rndis_msghdr *hdr = buf;
623 
624 	switch (hdr->type) {
625 	case RNDIS_PACKET_MSG:
626 		if (dev->data->dev_started)
627 			hn_rndis_rx_data(rxq, rxb, buf, len);
628 		break;
629 
630 	case RNDIS_INDICATE_STATUS_MSG:
631 		hn_rndis_link_status(dev, buf);
632 		break;
633 
634 	case RNDIS_INITIALIZE_CMPLT:
635 	case RNDIS_QUERY_CMPLT:
636 	case RNDIS_SET_CMPLT:
637 		hn_rndis_receive_response(rxq->hv, buf, len);
638 		break;
639 
640 	default:
641 		PMD_DRV_LOG(NOTICE,
642 			    "unexpected RNDIS message (type %#x len %u)",
643 			    hdr->type, len);
644 		break;
645 	}
646 }
647 
648 static void
649 hn_nvs_handle_rxbuf(struct rte_eth_dev *dev,
650 		    struct hn_data *hv,
651 		    struct hn_rx_queue *rxq,
652 		    const struct vmbus_chanpkt_hdr *hdr,
653 		    const void *buf)
654 {
655 	const struct vmbus_chanpkt_rxbuf *pkt;
656 	const struct hn_nvs_hdr *nvs_hdr = buf;
657 	uint32_t rxbuf_sz = hv->rxbuf_res->len;
658 	char *rxbuf = hv->rxbuf_res->addr;
659 	unsigned int i, hlen, count;
660 	struct hn_rx_bufinfo *rxb;
661 
662 	/* At minimum we need type header */
663 	if (unlikely(vmbus_chanpkt_datalen(hdr) < sizeof(*nvs_hdr))) {
664 		PMD_RX_LOG(ERR, "invalid receive nvs RNDIS");
665 		return;
666 	}
667 
668 	/* Make sure that this is a RNDIS message. */
669 	if (unlikely(nvs_hdr->type != NVS_TYPE_RNDIS)) {
670 		PMD_RX_LOG(ERR, "nvs type %u, not RNDIS",
671 			   nvs_hdr->type);
672 		return;
673 	}
674 
675 	hlen = vmbus_chanpkt_getlen(hdr->hlen);
676 	if (unlikely(hlen < sizeof(*pkt))) {
677 		PMD_RX_LOG(ERR, "invalid rxbuf chanpkt");
678 		return;
679 	}
680 
681 	pkt = container_of(hdr, const struct vmbus_chanpkt_rxbuf, hdr);
682 	if (unlikely(pkt->rxbuf_id != NVS_RXBUF_SIG)) {
683 		PMD_RX_LOG(ERR, "invalid rxbuf_id 0x%08x",
684 			   pkt->rxbuf_id);
685 		return;
686 	}
687 
688 	count = pkt->rxbuf_cnt;
689 	if (unlikely(hlen < offsetof(struct vmbus_chanpkt_rxbuf,
690 				     rxbuf[count]))) {
691 		PMD_RX_LOG(ERR, "invalid rxbuf_cnt %u", count);
692 		return;
693 	}
694 
695 	if (pkt->hdr.xactid > hv->rxbuf_section_cnt) {
696 		PMD_RX_LOG(ERR, "invalid rxbuf section id %" PRIx64,
697 			   pkt->hdr.xactid);
698 		return;
699 	}
700 
701 	/* Setup receive buffer info to allow for callback */
702 	rxb = hn_rx_buf_init(rxq, pkt);
703 
704 	/* Each range represents 1 RNDIS pkt that contains 1 Ethernet frame */
705 	for (i = 0; i < count; ++i) {
706 		unsigned int ofs, len;
707 
708 		ofs = pkt->rxbuf[i].ofs;
709 		len = pkt->rxbuf[i].len;
710 
711 		if (unlikely(ofs + len > rxbuf_sz)) {
712 			PMD_RX_LOG(ERR,
713 				   "%uth RNDIS msg overflow ofs %u, len %u",
714 				   i, ofs, len);
715 			continue;
716 		}
717 
718 		if (unlikely(len == 0)) {
719 			PMD_RX_LOG(ERR, "%uth RNDIS msg len %u", i, len);
720 			continue;
721 		}
722 
723 		hn_rndis_receive(dev, rxq, rxb,
724 				 rxbuf + ofs, len);
725 	}
726 
727 	/* Send ACK now if external mbuf not used */
728 	hn_rx_buf_release(rxb);
729 }
730 
731 /*
732  * Called when NVS inband events are received.
733  * Send up a two part message with port_id and the NVS message
734  * to the pipe to the netvsc-vf-event control thread.
735  */
736 static void hn_nvs_handle_notify(struct rte_eth_dev *dev,
737 				 const struct vmbus_chanpkt_hdr *pkt,
738 				 const void *data)
739 {
740 	const struct hn_nvs_hdr *hdr = data;
741 
742 	switch (hdr->type) {
743 	case NVS_TYPE_TXTBL_NOTE:
744 		/* Transmit indirection table has locking problems
745 		 * in DPDK and therefore not implemented
746 		 */
747 		PMD_DRV_LOG(DEBUG, "host notify of transmit indirection table");
748 		break;
749 
750 	case NVS_TYPE_VFASSOC_NOTE:
751 		hn_nvs_handle_vfassoc(dev, pkt, data);
752 		break;
753 
754 	default:
755 		PMD_DRV_LOG(INFO,
756 			    "got notify, nvs type %u", hdr->type);
757 	}
758 }
759 
760 struct hn_rx_queue *hn_rx_queue_alloc(struct hn_data *hv,
761 				      uint16_t queue_id,
762 				      unsigned int socket_id)
763 {
764 	struct hn_rx_queue *rxq;
765 
766 	rxq = rte_zmalloc_socket("HN_RXQ", sizeof(*rxq),
767 				 RTE_CACHE_LINE_SIZE, socket_id);
768 	if (!rxq)
769 		return NULL;
770 
771 	rxq->hv = hv;
772 	rxq->chan = hv->channels[queue_id];
773 	rte_spinlock_init(&rxq->ring_lock);
774 	rxq->port_id = hv->port_id;
775 	rxq->queue_id = queue_id;
776 	rxq->event_sz = HN_RXQ_EVENT_DEFAULT;
777 	rxq->event_buf = rte_malloc_socket("HN_EVENTS", HN_RXQ_EVENT_DEFAULT,
778 					   RTE_CACHE_LINE_SIZE, socket_id);
779 	if (!rxq->event_buf) {
780 		rte_free(rxq);
781 		return NULL;
782 	}
783 
784 	return rxq;
785 }
786 
787 int
788 hn_dev_rx_queue_setup(struct rte_eth_dev *dev,
789 		      uint16_t queue_idx, uint16_t nb_desc,
790 		      unsigned int socket_id,
791 		      const struct rte_eth_rxconf *rx_conf,
792 		      struct rte_mempool *mp)
793 {
794 	struct hn_data *hv = dev->data->dev_private;
795 	char ring_name[RTE_RING_NAMESIZE];
796 	struct hn_rx_queue *rxq;
797 	unsigned int count;
798 	int error = -ENOMEM;
799 
800 	PMD_INIT_FUNC_TRACE();
801 
802 	if (queue_idx == 0) {
803 		rxq = hv->primary;
804 	} else {
805 		rxq = hn_rx_queue_alloc(hv, queue_idx, socket_id);
806 		if (!rxq)
807 			return -ENOMEM;
808 	}
809 
810 	rxq->mb_pool = mp;
811 	count = rte_mempool_avail_count(mp) / dev->data->nb_rx_queues;
812 	if (nb_desc == 0 || nb_desc > count)
813 		nb_desc = count;
814 
815 	/*
816 	 * Staging ring from receive event logic to rx_pkts.
817 	 * rx_pkts assumes caller is handling multi-thread issue.
818 	 * event logic has locking.
819 	 */
820 	snprintf(ring_name, sizeof(ring_name),
821 		 "hn_rx_%u_%u", dev->data->port_id, queue_idx);
822 	rxq->rx_ring = rte_ring_create(ring_name,
823 				       rte_align32pow2(nb_desc),
824 				       socket_id, 0);
825 	if (!rxq->rx_ring)
826 		goto fail;
827 
828 	error = hn_vf_rx_queue_setup(dev, queue_idx, nb_desc,
829 				     socket_id, rx_conf, mp);
830 	if (error)
831 		goto fail;
832 
833 	dev->data->rx_queues[queue_idx] = rxq;
834 	return 0;
835 
836 fail:
837 	rte_ring_free(rxq->rx_ring);
838 	rte_free(rxq->event_buf);
839 	rte_free(rxq);
840 	return error;
841 }
842 
843 void
844 hn_dev_rx_queue_release(void *arg)
845 {
846 	struct hn_rx_queue *rxq = arg;
847 
848 	PMD_INIT_FUNC_TRACE();
849 
850 	if (!rxq)
851 		return;
852 
853 	rte_ring_free(rxq->rx_ring);
854 	rxq->rx_ring = NULL;
855 	rxq->mb_pool = NULL;
856 
857 	hn_vf_rx_queue_release(rxq->hv, rxq->queue_id);
858 
859 	/* Keep primary queue to allow for control operations */
860 	if (rxq != rxq->hv->primary) {
861 		rte_free(rxq->event_buf);
862 		rte_free(rxq);
863 	}
864 }
865 
866 int
867 hn_dev_tx_done_cleanup(void *arg, uint32_t free_cnt)
868 {
869 	struct hn_tx_queue *txq = arg;
870 
871 	return hn_process_events(txq->hv, txq->queue_id, free_cnt);
872 }
873 
874 /*
875  * Process pending events on the channel.
876  * Called from both Rx queue poll and Tx cleanup
877  */
878 uint32_t hn_process_events(struct hn_data *hv, uint16_t queue_id,
879 			   uint32_t tx_limit)
880 {
881 	struct rte_eth_dev *dev = &rte_eth_devices[hv->port_id];
882 	struct hn_rx_queue *rxq;
883 	uint32_t bytes_read = 0;
884 	uint32_t tx_done = 0;
885 	int ret = 0;
886 
887 	rxq = queue_id == 0 ? hv->primary : dev->data->rx_queues[queue_id];
888 
889 	/* If no pending data then nothing to do */
890 	if (rte_vmbus_chan_rx_empty(rxq->chan))
891 		return 0;
892 
893 	/*
894 	 * Since channel is shared between Rx and TX queue need to have a lock
895 	 * since DPDK does not force same CPU to be used for Rx/Tx.
896 	 */
897 	if (unlikely(!rte_spinlock_trylock(&rxq->ring_lock)))
898 		return 0;
899 
900 	for (;;) {
901 		const struct vmbus_chanpkt_hdr *pkt;
902 		uint32_t len = rxq->event_sz;
903 		const void *data;
904 
905 retry:
906 		ret = rte_vmbus_chan_recv_raw(rxq->chan, rxq->event_buf, &len);
907 		if (ret == -EAGAIN)
908 			break;	/* ring is empty */
909 
910 		if (unlikely(ret == -ENOBUFS)) {
911 			/* event buffer not large enough to read ring */
912 
913 			PMD_DRV_LOG(DEBUG,
914 				    "event buffer expansion (need %u)", len);
915 			rxq->event_sz = len + len / 4;
916 			rxq->event_buf = rte_realloc(rxq->event_buf, rxq->event_sz,
917 						     RTE_CACHE_LINE_SIZE);
918 			if (rxq->event_buf)
919 				goto retry;
920 			/* out of memory, no more events now */
921 			rxq->event_sz = 0;
922 			break;
923 		}
924 
925 		if (unlikely(ret <= 0)) {
926 			/* This indicates a failure to communicate (or worse) */
927 			rte_exit(EXIT_FAILURE,
928 				 "vmbus ring buffer error: %d", ret);
929 		}
930 
931 		bytes_read += ret;
932 		pkt = (const struct vmbus_chanpkt_hdr *)rxq->event_buf;
933 		data = (char *)rxq->event_buf + vmbus_chanpkt_getlen(pkt->hlen);
934 
935 		switch (pkt->type) {
936 		case VMBUS_CHANPKT_TYPE_COMP:
937 			++tx_done;
938 			hn_nvs_handle_comp(dev, queue_id, pkt, data);
939 			break;
940 
941 		case VMBUS_CHANPKT_TYPE_RXBUF:
942 			hn_nvs_handle_rxbuf(dev, hv, rxq, pkt, data);
943 			break;
944 
945 		case VMBUS_CHANPKT_TYPE_INBAND:
946 			hn_nvs_handle_notify(dev, pkt, data);
947 			break;
948 
949 		default:
950 			PMD_DRV_LOG(ERR, "unknown chan pkt %u", pkt->type);
951 			break;
952 		}
953 
954 		if (tx_limit && tx_done >= tx_limit)
955 			break;
956 
957 		if (rxq->rx_ring && rte_ring_full(rxq->rx_ring))
958 			break;
959 	}
960 
961 	if (bytes_read > 0)
962 		rte_vmbus_chan_signal_read(rxq->chan, bytes_read);
963 
964 	rte_spinlock_unlock(&rxq->ring_lock);
965 
966 	return tx_done;
967 }
968 
969 static void hn_append_to_chim(struct hn_tx_queue *txq,
970 			      struct rndis_packet_msg *pkt,
971 			      const struct rte_mbuf *m)
972 {
973 	struct hn_txdesc *txd = txq->agg_txd;
974 	uint8_t *buf = (uint8_t *)pkt;
975 	unsigned int data_offs;
976 
977 	hn_rndis_dump(pkt);
978 
979 	data_offs = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
980 	txd->chim_size += pkt->len;
981 	txd->data_size += m->pkt_len;
982 	++txd->packets;
983 	hn_update_packet_stats(&txq->stats, m);
984 
985 	for (; m; m = m->next) {
986 		uint16_t len = rte_pktmbuf_data_len(m);
987 
988 		rte_memcpy(buf + data_offs,
989 			   rte_pktmbuf_mtod(m, const char *), len);
990 		data_offs += len;
991 	}
992 }
993 
994 /*
995  * Send pending aggregated data in chimney buffer (if any).
996  * Returns error if send was unsuccessful because channel ring buffer
997  * was full.
998  */
999 static int hn_flush_txagg(struct hn_tx_queue *txq, bool *need_sig)
1000 
1001 {
1002 	struct hn_txdesc *txd = txq->agg_txd;
1003 	struct hn_nvs_rndis rndis;
1004 	int ret;
1005 
1006 	if (!txd)
1007 		return 0;
1008 
1009 	rndis = (struct hn_nvs_rndis) {
1010 		.type = NVS_TYPE_RNDIS,
1011 		.rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1012 		.chim_idx = txd->chim_index,
1013 		.chim_sz = txd->chim_size,
1014 	};
1015 
1016 	PMD_TX_LOG(DEBUG, "port %u:%u tx %u size %u",
1017 		   txq->port_id, txq->queue_id, txd->chim_index, txd->chim_size);
1018 
1019 	ret = hn_nvs_send(txq->chan, VMBUS_CHANPKT_FLAG_RC,
1020 			  &rndis, sizeof(rndis), (uintptr_t)txd, need_sig);
1021 
1022 	if (likely(ret == 0))
1023 		hn_reset_txagg(txq);
1024 	else
1025 		PMD_TX_LOG(NOTICE, "port %u:%u send failed: %d",
1026 			   txq->port_id, txq->queue_id, ret);
1027 
1028 	return ret;
1029 }
1030 
1031 static struct hn_txdesc *hn_new_txd(struct hn_data *hv,
1032 				    struct hn_tx_queue *txq)
1033 {
1034 	struct hn_txdesc *txd;
1035 
1036 	if (rte_mempool_get(hv->tx_pool, (void **)&txd)) {
1037 		++txq->stats.ring_full;
1038 		PMD_TX_LOG(DEBUG, "tx pool exhausted!");
1039 		return NULL;
1040 	}
1041 
1042 	txd->m = NULL;
1043 	txd->queue_id = txq->queue_id;
1044 	txd->packets = 0;
1045 	txd->data_size = 0;
1046 	txd->chim_size = 0;
1047 
1048 	return txd;
1049 }
1050 
1051 static void *
1052 hn_try_txagg(struct hn_data *hv, struct hn_tx_queue *txq, uint32_t pktsize)
1053 {
1054 	struct hn_txdesc *agg_txd = txq->agg_txd;
1055 	struct rndis_packet_msg *pkt;
1056 	void *chim;
1057 
1058 	if (agg_txd) {
1059 		unsigned int padding, olen;
1060 
1061 		/*
1062 		 * Update the previous RNDIS packet's total length,
1063 		 * it can be increased due to the mandatory alignment
1064 		 * padding for this RNDIS packet.  And update the
1065 		 * aggregating txdesc's chimney sending buffer size
1066 		 * accordingly.
1067 		 *
1068 		 * Zero-out the padding, as required by the RNDIS spec.
1069 		 */
1070 		pkt = txq->agg_prevpkt;
1071 		olen = pkt->len;
1072 		padding = RTE_ALIGN(olen, txq->agg_align) - olen;
1073 		if (padding > 0) {
1074 			agg_txd->chim_size += padding;
1075 			pkt->len += padding;
1076 			memset((uint8_t *)pkt + olen, 0, padding);
1077 		}
1078 
1079 		chim = (uint8_t *)pkt + pkt->len;
1080 
1081 		txq->agg_pktleft--;
1082 		txq->agg_szleft -= pktsize;
1083 		if (txq->agg_szleft < HN_PKTSIZE_MIN(txq->agg_align)) {
1084 			/*
1085 			 * Probably can't aggregate more packets,
1086 			 * flush this aggregating txdesc proactively.
1087 			 */
1088 			txq->agg_pktleft = 0;
1089 		}
1090 	} else {
1091 		agg_txd = hn_new_txd(hv, txq);
1092 		if (!agg_txd)
1093 			return NULL;
1094 
1095 		chim = (uint8_t *)hv->chim_res->addr
1096 			+ agg_txd->chim_index * hv->chim_szmax;
1097 
1098 		txq->agg_txd = agg_txd;
1099 		txq->agg_pktleft = txq->agg_pktmax - 1;
1100 		txq->agg_szleft = txq->agg_szmax - pktsize;
1101 	}
1102 	txq->agg_prevpkt = chim;
1103 
1104 	return chim;
1105 }
1106 
1107 static inline void *
1108 hn_rndis_pktinfo_append(struct rndis_packet_msg *pkt,
1109 			uint32_t pi_dlen, uint32_t pi_type)
1110 {
1111 	const uint32_t pi_size = RNDIS_PKTINFO_SIZE(pi_dlen);
1112 	struct rndis_pktinfo *pi;
1113 
1114 	/*
1115 	 * Per-packet-info does not move; it only grows.
1116 	 *
1117 	 * NOTE:
1118 	 * pktinfooffset in this phase counts from the beginning
1119 	 * of rndis_packet_msg.
1120 	 */
1121 	pi = (struct rndis_pktinfo *)((uint8_t *)pkt + hn_rndis_pktlen(pkt));
1122 
1123 	pkt->pktinfolen += pi_size;
1124 
1125 	pi->size = pi_size;
1126 	pi->type = pi_type;
1127 	pi->offset = RNDIS_PKTINFO_OFFSET;
1128 
1129 	return pi->data;
1130 }
1131 
1132 /* Put RNDIS header and packet info on packet */
1133 static void hn_encap(struct rndis_packet_msg *pkt,
1134 		     uint16_t queue_id,
1135 		     const struct rte_mbuf *m)
1136 {
1137 	unsigned int hlen = m->l2_len + m->l3_len;
1138 	uint32_t *pi_data;
1139 	uint32_t pkt_hlen;
1140 
1141 	pkt->type = RNDIS_PACKET_MSG;
1142 	pkt->len = m->pkt_len;
1143 	pkt->dataoffset = 0;
1144 	pkt->datalen = m->pkt_len;
1145 	pkt->oobdataoffset = 0;
1146 	pkt->oobdatalen = 0;
1147 	pkt->oobdataelements = 0;
1148 	pkt->pktinfooffset = sizeof(*pkt);
1149 	pkt->pktinfolen = 0;
1150 	pkt->vchandle = 0;
1151 	pkt->reserved = 0;
1152 
1153 	/*
1154 	 * Set the hash value for this packet, to the queue_id to cause
1155 	 * TX done event for this packet on the right channel.
1156 	 */
1157 	pi_data = hn_rndis_pktinfo_append(pkt, NDIS_HASH_VALUE_SIZE,
1158 					  NDIS_PKTINFO_TYPE_HASHVAL);
1159 	*pi_data = queue_id;
1160 
1161 	if (m->ol_flags & PKT_TX_VLAN_PKT) {
1162 		pi_data = hn_rndis_pktinfo_append(pkt, NDIS_VLAN_INFO_SIZE,
1163 						  NDIS_PKTINFO_TYPE_VLAN);
1164 		*pi_data = m->vlan_tci;
1165 	}
1166 
1167 	if (m->ol_flags & PKT_TX_TCP_SEG) {
1168 		pi_data = hn_rndis_pktinfo_append(pkt, NDIS_LSO2_INFO_SIZE,
1169 						  NDIS_PKTINFO_TYPE_LSO);
1170 
1171 		if (m->ol_flags & PKT_TX_IPV6) {
1172 			*pi_data = NDIS_LSO2_INFO_MAKEIPV6(hlen,
1173 							   m->tso_segsz);
1174 		} else {
1175 			*pi_data = NDIS_LSO2_INFO_MAKEIPV4(hlen,
1176 							   m->tso_segsz);
1177 		}
1178 	} else if (m->ol_flags &
1179 		   (PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM | PKT_TX_IP_CKSUM)) {
1180 		pi_data = hn_rndis_pktinfo_append(pkt, NDIS_TXCSUM_INFO_SIZE,
1181 						  NDIS_PKTINFO_TYPE_CSUM);
1182 		*pi_data = 0;
1183 
1184 		if (m->ol_flags & PKT_TX_IPV6)
1185 			*pi_data |= NDIS_TXCSUM_INFO_IPV6;
1186 		if (m->ol_flags & PKT_TX_IPV4) {
1187 			*pi_data |= NDIS_TXCSUM_INFO_IPV4;
1188 
1189 			if (m->ol_flags & PKT_TX_IP_CKSUM)
1190 				*pi_data |= NDIS_TXCSUM_INFO_IPCS;
1191 		}
1192 
1193 		if (m->ol_flags & PKT_TX_TCP_CKSUM)
1194 			*pi_data |= NDIS_TXCSUM_INFO_MKTCPCS(hlen);
1195 		else if (m->ol_flags & PKT_TX_UDP_CKSUM)
1196 			*pi_data |= NDIS_TXCSUM_INFO_MKUDPCS(hlen);
1197 	}
1198 
1199 	pkt_hlen = pkt->pktinfooffset + pkt->pktinfolen;
1200 	/* Fixup RNDIS packet message total length */
1201 	pkt->len += pkt_hlen;
1202 
1203 	/* Convert RNDIS packet message offsets */
1204 	pkt->dataoffset = hn_rndis_pktmsg_offset(pkt_hlen);
1205 	pkt->pktinfooffset = hn_rndis_pktmsg_offset(pkt->pktinfooffset);
1206 }
1207 
1208 /* How many scatter gather list elements ar needed */
1209 static unsigned int hn_get_slots(const struct rte_mbuf *m)
1210 {
1211 	unsigned int slots = 1; /* for RNDIS header */
1212 
1213 	while (m) {
1214 		unsigned int size = rte_pktmbuf_data_len(m);
1215 		unsigned int offs = rte_mbuf_data_iova(m) & PAGE_MASK;
1216 
1217 		slots += (offs + size + PAGE_SIZE - 1) / PAGE_SIZE;
1218 		m = m->next;
1219 	}
1220 
1221 	return slots;
1222 }
1223 
1224 /* Build scatter gather list from chained mbuf */
1225 static unsigned int hn_fill_sg(struct vmbus_gpa *sg,
1226 			       const struct rte_mbuf *m)
1227 {
1228 	unsigned int segs = 0;
1229 
1230 	while (m) {
1231 		rte_iova_t addr = rte_mbuf_data_iova(m);
1232 		unsigned int page = addr / PAGE_SIZE;
1233 		unsigned int offset = addr & PAGE_MASK;
1234 		unsigned int len = rte_pktmbuf_data_len(m);
1235 
1236 		while (len > 0) {
1237 			unsigned int bytes = RTE_MIN(len, PAGE_SIZE - offset);
1238 
1239 			sg[segs].page = page;
1240 			sg[segs].ofs = offset;
1241 			sg[segs].len = bytes;
1242 			segs++;
1243 
1244 			++page;
1245 			offset = 0;
1246 			len -= bytes;
1247 		}
1248 		m = m->next;
1249 	}
1250 
1251 	return segs;
1252 }
1253 
1254 /* Transmit directly from mbuf */
1255 static int hn_xmit_sg(struct hn_tx_queue *txq,
1256 		      const struct hn_txdesc *txd, const struct rte_mbuf *m,
1257 		      bool *need_sig)
1258 {
1259 	struct vmbus_gpa sg[hn_get_slots(m)];
1260 	struct hn_nvs_rndis nvs_rndis = {
1261 		.type = NVS_TYPE_RNDIS,
1262 		.rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1263 		.chim_sz = txd->chim_size,
1264 	};
1265 	rte_iova_t addr;
1266 	unsigned int segs;
1267 
1268 	/* attach aggregation data if present */
1269 	if (txd->chim_size > 0)
1270 		nvs_rndis.chim_idx = txd->chim_index;
1271 	else
1272 		nvs_rndis.chim_idx = NVS_CHIM_IDX_INVALID;
1273 
1274 	hn_rndis_dump(txd->rndis_pkt);
1275 
1276 	/* pass IOVA of rndis header in first segment */
1277 	addr = rte_malloc_virt2iova(txd->rndis_pkt);
1278 	if (unlikely(addr == RTE_BAD_IOVA)) {
1279 		PMD_DRV_LOG(ERR, "RNDIS transmit can not get iova");
1280 		return -EINVAL;
1281 	}
1282 
1283 	sg[0].page = addr / PAGE_SIZE;
1284 	sg[0].ofs = addr & PAGE_MASK;
1285 	sg[0].len = RNDIS_PACKET_MSG_OFFSET_ABS(hn_rndis_pktlen(txd->rndis_pkt));
1286 	segs = 1;
1287 
1288 	hn_update_packet_stats(&txq->stats, m);
1289 
1290 	segs += hn_fill_sg(sg + 1, m);
1291 
1292 	PMD_TX_LOG(DEBUG, "port %u:%u tx %u segs %u size %u",
1293 		   txq->port_id, txq->queue_id, txd->chim_index,
1294 		   segs, nvs_rndis.chim_sz);
1295 
1296 	return hn_nvs_send_sglist(txq->chan, sg, segs,
1297 				  &nvs_rndis, sizeof(nvs_rndis),
1298 				  (uintptr_t)txd, need_sig);
1299 }
1300 
1301 uint16_t
1302 hn_xmit_pkts(void *ptxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1303 {
1304 	struct hn_tx_queue *txq = ptxq;
1305 	uint16_t queue_id = txq->queue_id;
1306 	struct hn_data *hv = txq->hv;
1307 	struct rte_eth_dev *vf_dev;
1308 	bool need_sig = false;
1309 	uint16_t nb_tx;
1310 	int ret;
1311 
1312 	if (unlikely(hv->closed))
1313 		return 0;
1314 
1315 	/* Transmit over VF if present and up */
1316 	vf_dev = hn_get_vf_dev(hv);
1317 
1318 	if (vf_dev && vf_dev->data->dev_started) {
1319 		void *sub_q = vf_dev->data->tx_queues[queue_id];
1320 
1321 		return (*vf_dev->tx_pkt_burst)(sub_q, tx_pkts, nb_pkts);
1322 	}
1323 
1324 	if (rte_mempool_avail_count(hv->tx_pool) <= txq->free_thresh)
1325 		hn_process_events(hv, txq->queue_id, 0);
1326 
1327 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1328 		struct rte_mbuf *m = tx_pkts[nb_tx];
1329 		uint32_t pkt_size = m->pkt_len + HN_RNDIS_PKT_LEN;
1330 		struct rndis_packet_msg *pkt;
1331 
1332 		/* For small packets aggregate them in chimney buffer */
1333 		if (m->pkt_len < HN_TXCOPY_THRESHOLD && pkt_size <= txq->agg_szmax) {
1334 			/* If this packet will not fit, then flush  */
1335 			if (txq->agg_pktleft == 0 ||
1336 			    RTE_ALIGN(pkt_size, txq->agg_align) > txq->agg_szleft) {
1337 				if (hn_flush_txagg(txq, &need_sig))
1338 					goto fail;
1339 			}
1340 
1341 			pkt = hn_try_txagg(hv, txq, pkt_size);
1342 			if (unlikely(!pkt))
1343 				break;
1344 
1345 			hn_encap(pkt, queue_id, m);
1346 			hn_append_to_chim(txq, pkt, m);
1347 
1348 			rte_pktmbuf_free(m);
1349 
1350 			/* if buffer is full, flush */
1351 			if (txq->agg_pktleft == 0 &&
1352 			    hn_flush_txagg(txq, &need_sig))
1353 				goto fail;
1354 		} else {
1355 			struct hn_txdesc *txd;
1356 
1357 			/* can send chimney data and large packet at once */
1358 			txd = txq->agg_txd;
1359 			if (txd) {
1360 				hn_reset_txagg(txq);
1361 			} else {
1362 				txd = hn_new_txd(hv, txq);
1363 				if (unlikely(!txd))
1364 					break;
1365 			}
1366 
1367 			pkt = txd->rndis_pkt;
1368 			txd->m = m;
1369 			txd->data_size += m->pkt_len;
1370 			++txd->packets;
1371 
1372 			hn_encap(pkt, queue_id, m);
1373 
1374 			ret = hn_xmit_sg(txq, txd, m, &need_sig);
1375 			if (unlikely(ret != 0)) {
1376 				PMD_TX_LOG(NOTICE, "sg send failed: %d", ret);
1377 				++txq->stats.errors;
1378 				rte_mempool_put(hv->tx_pool, txd);
1379 				goto fail;
1380 			}
1381 		}
1382 	}
1383 
1384 	/* If partial buffer left, then try and send it.
1385 	 * if that fails, then reuse it on next send.
1386 	 */
1387 	hn_flush_txagg(txq, &need_sig);
1388 
1389 fail:
1390 	if (need_sig)
1391 		rte_vmbus_chan_signal_tx(txq->chan);
1392 
1393 	return nb_tx;
1394 }
1395 
1396 uint16_t
1397 hn_recv_pkts(void *prxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1398 {
1399 	struct hn_rx_queue *rxq = prxq;
1400 	struct hn_data *hv = rxq->hv;
1401 	struct rte_eth_dev *vf_dev;
1402 	uint16_t nb_rcv;
1403 
1404 	if (unlikely(hv->closed))
1405 		return 0;
1406 
1407 	/* Transmit over VF if present and up */
1408 	vf_dev = hn_get_vf_dev(hv);
1409 
1410 	if (vf_dev && vf_dev->data->dev_started) {
1411 		/* Normally, with SR-IOV the ring buffer will be empty */
1412 		hn_process_events(hv, rxq->queue_id, 0);
1413 
1414 		/* Get mbufs some bufs off of staging ring */
1415 		nb_rcv = rte_ring_sc_dequeue_burst(rxq->rx_ring,
1416 						   (void **)rx_pkts,
1417 						   nb_pkts / 2, NULL);
1418 		/* And rest off of VF */
1419 		nb_rcv += rte_eth_rx_burst(vf_dev->data->port_id,
1420 					   rxq->queue_id,
1421 					   rx_pkts + nb_rcv, nb_pkts - nb_rcv);
1422 	} else {
1423 		/* If receive ring is not full then get more */
1424 		if (rte_ring_count(rxq->rx_ring) < nb_pkts)
1425 			hn_process_events(hv, rxq->queue_id, 0);
1426 
1427 		nb_rcv = rte_ring_sc_dequeue_burst(rxq->rx_ring,
1428 						   (void **)rx_pkts,
1429 						   nb_pkts, NULL);
1430 	}
1431 
1432 	return nb_rcv;
1433 }
1434