xref: /dpdk/drivers/net/netvsc/hn_rxtx.c (revision 3cc6ecfdfe85d2577fef30e1791bb7534e3d60b3)
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 = 0;
210 
211 	rte_spinlock_lock(&hv->chim_lock);
212 	if (rte_bitmap_scan(hv->chim_bmap, &index, &slab)) {
213 		index += rte_bsf64(slab);
214 		rte_bitmap_clear(hv->chim_bmap, index);
215 	}
216 	rte_spinlock_unlock(&hv->chim_lock);
217 
218 	return index;
219 }
220 
221 static void hn_chim_free(struct hn_data *hv, uint32_t chim_idx)
222 {
223 	if (chim_idx >= hv->chim_cnt) {
224 		PMD_DRV_LOG(ERR, "Invalid chimney index %u", chim_idx);
225 	} else {
226 		rte_spinlock_lock(&hv->chim_lock);
227 		rte_bitmap_set(hv->chim_bmap, chim_idx);
228 		rte_spinlock_unlock(&hv->chim_lock);
229 	}
230 }
231 
232 static void hn_reset_txagg(struct hn_tx_queue *txq)
233 {
234 	txq->agg_szleft = txq->agg_szmax;
235 	txq->agg_pktleft = txq->agg_pktmax;
236 	txq->agg_txd = NULL;
237 	txq->agg_prevpkt = NULL;
238 }
239 
240 int
241 hn_dev_tx_queue_setup(struct rte_eth_dev *dev,
242 		      uint16_t queue_idx, uint16_t nb_desc,
243 		      unsigned int socket_id,
244 		      const struct rte_eth_txconf *tx_conf)
245 
246 {
247 	struct hn_data *hv = dev->data->dev_private;
248 	struct hn_tx_queue *txq;
249 	char name[RTE_MEMPOOL_NAMESIZE];
250 	uint32_t tx_free_thresh;
251 	int err = -ENOMEM;
252 
253 	PMD_INIT_FUNC_TRACE();
254 
255 	txq = rte_zmalloc_socket("HN_TXQ", sizeof(*txq), RTE_CACHE_LINE_SIZE,
256 				 socket_id);
257 	if (!txq)
258 		return -ENOMEM;
259 
260 	txq->hv = hv;
261 	txq->chan = hv->channels[queue_idx];
262 	txq->port_id = dev->data->port_id;
263 	txq->queue_id = queue_idx;
264 
265 	tx_free_thresh = tx_conf->tx_free_thresh;
266 	if (tx_free_thresh == 0)
267 		tx_free_thresh = RTE_MIN(nb_desc / 4,
268 					 DEFAULT_TX_FREE_THRESH);
269 
270 	if (tx_free_thresh + 3 >= nb_desc) {
271 		PMD_INIT_LOG(ERR,
272 			     "tx_free_thresh must be less than the number of TX entries minus 3(%u)."
273 			     " (tx_free_thresh=%u port=%u queue=%u)\n",
274 			     nb_desc - 3,
275 			     tx_free_thresh, dev->data->port_id, queue_idx);
276 		return -EINVAL;
277 	}
278 
279 	txq->free_thresh = tx_free_thresh;
280 
281 	snprintf(name, sizeof(name),
282 		 "hn_txd_%u_%u", dev->data->port_id, queue_idx);
283 
284 	PMD_INIT_LOG(DEBUG, "TX descriptor pool %s n=%u size=%zu",
285 		     name, nb_desc, sizeof(struct hn_txdesc));
286 
287 	txq->tx_rndis = rte_calloc("hn_txq_rndis", nb_desc,
288 				   HN_RNDIS_PKT_ALIGNED, RTE_CACHE_LINE_SIZE);
289 	if (txq->tx_rndis == NULL)
290 		goto error;
291 
292 	txq->txdesc_pool = rte_mempool_create(name, nb_desc,
293 					      sizeof(struct hn_txdesc),
294 					      0, 0, NULL, NULL,
295 					      hn_txd_init, txq,
296 					      dev->device->numa_node, 0);
297 	if (txq->txdesc_pool == NULL) {
298 		PMD_DRV_LOG(ERR,
299 			    "mempool %s create failed: %d", name, rte_errno);
300 		goto error;
301 	}
302 
303 	txq->agg_szmax  = RTE_MIN(hv->chim_szmax, hv->rndis_agg_size);
304 	txq->agg_pktmax = hv->rndis_agg_pkts;
305 	txq->agg_align  = hv->rndis_agg_align;
306 
307 	hn_reset_txagg(txq);
308 
309 	err = hn_vf_tx_queue_setup(dev, queue_idx, nb_desc,
310 				     socket_id, tx_conf);
311 	if (err == 0) {
312 		dev->data->tx_queues[queue_idx] = txq;
313 		return 0;
314 	}
315 
316 error:
317 	if (txq->txdesc_pool)
318 		rte_mempool_free(txq->txdesc_pool);
319 	rte_free(txq->tx_rndis);
320 	rte_free(txq);
321 	return err;
322 }
323 
324 void
325 hn_dev_tx_queue_info(struct rte_eth_dev *dev, uint16_t queue_id,
326 		     struct rte_eth_txq_info *qinfo)
327 {
328 	struct hn_tx_queue *txq = dev->data->tx_queues[queue_id];
329 
330 	qinfo->nb_desc = txq->txdesc_pool->size;
331 	qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads;
332 }
333 
334 static struct hn_txdesc *hn_txd_get(struct hn_tx_queue *txq)
335 {
336 	struct hn_txdesc *txd;
337 
338 	if (rte_mempool_get(txq->txdesc_pool, (void **)&txd)) {
339 		++txq->stats.ring_full;
340 		PMD_TX_LOG(DEBUG, "tx pool exhausted!");
341 		return NULL;
342 	}
343 
344 	txd->m = NULL;
345 	txd->packets = 0;
346 	txd->data_size = 0;
347 	txd->chim_size = 0;
348 
349 	return txd;
350 }
351 
352 static void hn_txd_put(struct hn_tx_queue *txq, struct hn_txdesc *txd)
353 {
354 	rte_mempool_put(txq->txdesc_pool, txd);
355 }
356 
357 void
358 hn_dev_tx_queue_release(void *arg)
359 {
360 	struct hn_tx_queue *txq = arg;
361 
362 	PMD_INIT_FUNC_TRACE();
363 
364 	if (!txq)
365 		return;
366 
367 	if (txq->txdesc_pool)
368 		rte_mempool_free(txq->txdesc_pool);
369 
370 	rte_free(txq->tx_rndis);
371 	rte_free(txq);
372 }
373 
374 /*
375  * Check the status of a Tx descriptor in the queue.
376  *
377  * returns:
378  *  - -EINVAL              - offset outside of tx_descriptor pool.
379  *  - RTE_ETH_TX_DESC_FULL - descriptor is not acknowledged by host.
380  *  - RTE_ETH_TX_DESC_DONE - descriptor is available.
381  */
382 int hn_dev_tx_descriptor_status(void *arg, uint16_t offset)
383 {
384 	const struct hn_tx_queue *txq = arg;
385 
386 	hn_process_events(txq->hv, txq->queue_id, 0);
387 
388 	if (offset >= rte_mempool_avail_count(txq->txdesc_pool))
389 		return -EINVAL;
390 
391 	if (offset < rte_mempool_in_use_count(txq->txdesc_pool))
392 		return RTE_ETH_TX_DESC_FULL;
393 	else
394 		return RTE_ETH_TX_DESC_DONE;
395 }
396 
397 static void
398 hn_nvs_send_completed(struct rte_eth_dev *dev, uint16_t queue_id,
399 		      unsigned long xactid, const struct hn_nvs_rndis_ack *ack)
400 {
401 	struct hn_data *hv = dev->data->dev_private;
402 	struct hn_txdesc *txd = (struct hn_txdesc *)xactid;
403 	struct hn_tx_queue *txq;
404 
405 	/* Control packets are sent with xacid == 0 */
406 	if (!txd)
407 		return;
408 
409 	txq = dev->data->tx_queues[queue_id];
410 	if (likely(ack->status == NVS_STATUS_OK)) {
411 		PMD_TX_LOG(DEBUG, "port %u:%u complete tx %u packets %u bytes %u",
412 			   txq->port_id, txq->queue_id, txd->chim_index,
413 			   txd->packets, txd->data_size);
414 		txq->stats.bytes += txd->data_size;
415 		txq->stats.packets += txd->packets;
416 	} else {
417 		PMD_DRV_LOG(NOTICE, "port %u:%u complete tx %u failed status %u",
418 			    txq->port_id, txq->queue_id, txd->chim_index, ack->status);
419 		++txq->stats.errors;
420 	}
421 
422 	if (txd->chim_index != NVS_CHIM_IDX_INVALID)
423 		hn_chim_free(hv, txd->chim_index);
424 
425 	rte_pktmbuf_free(txd->m);
426 	hn_txd_put(txq, txd);
427 }
428 
429 /* Handle transmit completion events */
430 static void
431 hn_nvs_handle_comp(struct rte_eth_dev *dev, uint16_t queue_id,
432 		   const struct vmbus_chanpkt_hdr *pkt,
433 		   const void *data)
434 {
435 	const struct hn_nvs_hdr *hdr = data;
436 
437 	switch (hdr->type) {
438 	case NVS_TYPE_RNDIS_ACK:
439 		hn_nvs_send_completed(dev, queue_id, pkt->xactid, data);
440 		break;
441 
442 	default:
443 		PMD_DRV_LOG(NOTICE, "unexpected send completion type %u",
444 			   hdr->type);
445 	}
446 }
447 
448 /* Parse per-packet info (meta data) */
449 static int
450 hn_rndis_rxinfo(const void *info_data, unsigned int info_dlen,
451 		struct hn_rxinfo *info)
452 {
453 	const struct rndis_pktinfo *pi = info_data;
454 	uint32_t mask = 0;
455 
456 	while (info_dlen != 0) {
457 		const void *data;
458 		uint32_t dlen;
459 
460 		if (unlikely(info_dlen < sizeof(*pi)))
461 			return -EINVAL;
462 
463 		if (unlikely(info_dlen < pi->size))
464 			return -EINVAL;
465 		info_dlen -= pi->size;
466 
467 		if (unlikely(pi->size & RNDIS_PKTINFO_SIZE_ALIGNMASK))
468 			return -EINVAL;
469 		if (unlikely(pi->size < pi->offset))
470 			return -EINVAL;
471 
472 		dlen = pi->size - pi->offset;
473 		data = pi->data;
474 
475 		switch (pi->type) {
476 		case NDIS_PKTINFO_TYPE_VLAN:
477 			if (unlikely(dlen < NDIS_VLAN_INFO_SIZE))
478 				return -EINVAL;
479 			info->vlan_info = *((const uint32_t *)data);
480 			mask |= HN_RXINFO_VLAN;
481 			break;
482 
483 		case NDIS_PKTINFO_TYPE_CSUM:
484 			if (unlikely(dlen < NDIS_RXCSUM_INFO_SIZE))
485 				return -EINVAL;
486 			info->csum_info = *((const uint32_t *)data);
487 			mask |= HN_RXINFO_CSUM;
488 			break;
489 
490 		case NDIS_PKTINFO_TYPE_HASHVAL:
491 			if (unlikely(dlen < NDIS_HASH_VALUE_SIZE))
492 				return -EINVAL;
493 			info->hash_value = *((const uint32_t *)data);
494 			mask |= HN_RXINFO_HASHVAL;
495 			break;
496 
497 		case NDIS_PKTINFO_TYPE_HASHINF:
498 			if (unlikely(dlen < NDIS_HASH_INFO_SIZE))
499 				return -EINVAL;
500 			info->hash_info = *((const uint32_t *)data);
501 			mask |= HN_RXINFO_HASHINF;
502 			break;
503 
504 		default:
505 			goto next;
506 		}
507 
508 		if (mask == HN_RXINFO_ALL)
509 			break; /* All found; done */
510 next:
511 		pi = (const struct rndis_pktinfo *)
512 		    ((const uint8_t *)pi + pi->size);
513 	}
514 
515 	/*
516 	 * Final fixup.
517 	 * - If there is no hash value, invalidate the hash info.
518 	 */
519 	if (!(mask & HN_RXINFO_HASHVAL))
520 		info->hash_info = HN_NDIS_HASH_INFO_INVALID;
521 	return 0;
522 }
523 
524 static void hn_rx_buf_free_cb(void *buf __rte_unused, void *opaque)
525 {
526 	struct hn_rx_bufinfo *rxb = opaque;
527 	struct hn_data *hv = rxb->hv;
528 
529 	rte_atomic32_dec(&hv->rxbuf_outstanding);
530 	hn_nvs_ack_rxbuf(rxb->chan, rxb->xactid);
531 }
532 
533 static struct hn_rx_bufinfo *hn_rx_buf_init(const struct hn_rx_queue *rxq,
534 					    const struct vmbus_chanpkt_rxbuf *pkt)
535 {
536 	struct hn_rx_bufinfo *rxb;
537 
538 	rxb = rxq->hv->rxbuf_info + pkt->hdr.xactid;
539 	rxb->chan = rxq->chan;
540 	rxb->xactid = pkt->hdr.xactid;
541 	rxb->hv = rxq->hv;
542 
543 	rxb->shinfo.free_cb = hn_rx_buf_free_cb;
544 	rxb->shinfo.fcb_opaque = rxb;
545 	rte_mbuf_ext_refcnt_set(&rxb->shinfo, 1);
546 	return rxb;
547 }
548 
549 static void hn_rxpkt(struct hn_rx_queue *rxq, struct hn_rx_bufinfo *rxb,
550 		     uint8_t *data, unsigned int headroom, unsigned int dlen,
551 		     const struct hn_rxinfo *info)
552 {
553 	struct hn_data *hv = rxq->hv;
554 	struct rte_mbuf *m;
555 	bool use_extbuf = false;
556 
557 	m = rte_pktmbuf_alloc(rxq->mb_pool);
558 	if (unlikely(!m)) {
559 		struct rte_eth_dev *dev =
560 			&rte_eth_devices[rxq->port_id];
561 
562 		dev->data->rx_mbuf_alloc_failed++;
563 		return;
564 	}
565 
566 	/*
567 	 * For large packets, avoid copy if possible but need to keep
568 	 * some space available in receive area for later packets.
569 	 */
570 	if (dlen >= HN_RXCOPY_THRESHOLD &&
571 	    (uint32_t)rte_atomic32_read(&hv->rxbuf_outstanding) <
572 			hv->rxbuf_section_cnt / 2) {
573 		struct rte_mbuf_ext_shared_info *shinfo;
574 		const void *rxbuf;
575 		rte_iova_t iova;
576 
577 		/*
578 		 * Build an external mbuf that points to recveive area.
579 		 * Use refcount to handle multiple packets in same
580 		 * receive buffer section.
581 		 */
582 		rxbuf = hv->rxbuf_res->addr;
583 		iova = rte_mem_virt2iova(rxbuf) + RTE_PTR_DIFF(data, rxbuf);
584 		shinfo = &rxb->shinfo;
585 
586 		/* shinfo is already set to 1 by the caller */
587 		if (rte_mbuf_ext_refcnt_update(shinfo, 1) == 2)
588 			rte_atomic32_inc(&hv->rxbuf_outstanding);
589 
590 		rte_pktmbuf_attach_extbuf(m, data, iova,
591 					  dlen + headroom, shinfo);
592 		m->data_off = headroom;
593 		use_extbuf = true;
594 	} else {
595 		/* Mbuf's in pool must be large enough to hold small packets */
596 		if (unlikely(rte_pktmbuf_tailroom(m) < dlen)) {
597 			rte_pktmbuf_free_seg(m);
598 			++rxq->stats.errors;
599 			return;
600 		}
601 		rte_memcpy(rte_pktmbuf_mtod(m, void *),
602 			   data + headroom, dlen);
603 	}
604 
605 	m->port = rxq->port_id;
606 	m->pkt_len = dlen;
607 	m->data_len = dlen;
608 	m->packet_type = rte_net_get_ptype(m, NULL,
609 					   RTE_PTYPE_L2_MASK |
610 					   RTE_PTYPE_L3_MASK |
611 					   RTE_PTYPE_L4_MASK);
612 
613 	if (info->vlan_info != HN_NDIS_VLAN_INFO_INVALID) {
614 		m->vlan_tci = info->vlan_info;
615 		m->ol_flags |= PKT_RX_VLAN_STRIPPED | PKT_RX_VLAN;
616 
617 		/* NDIS always strips tag, put it back if necessary */
618 		if (!hv->vlan_strip && rte_vlan_insert(&m)) {
619 			PMD_DRV_LOG(DEBUG, "vlan insert failed");
620 			++rxq->stats.errors;
621 			if (use_extbuf)
622 				rte_pktmbuf_detach_extbuf(m);
623 			rte_pktmbuf_free(m);
624 			return;
625 		}
626 	}
627 
628 	if (info->csum_info != HN_NDIS_RXCSUM_INFO_INVALID) {
629 		if (info->csum_info & NDIS_RXCSUM_INFO_IPCS_OK)
630 			m->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
631 
632 		if (info->csum_info & (NDIS_RXCSUM_INFO_UDPCS_OK
633 				       | NDIS_RXCSUM_INFO_TCPCS_OK))
634 			m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
635 		else if (info->csum_info & (NDIS_RXCSUM_INFO_TCPCS_FAILED
636 					    | NDIS_RXCSUM_INFO_UDPCS_FAILED))
637 			m->ol_flags |= PKT_RX_L4_CKSUM_BAD;
638 	}
639 
640 	if (info->hash_info != HN_NDIS_HASH_INFO_INVALID) {
641 		m->ol_flags |= PKT_RX_RSS_HASH;
642 		m->hash.rss = info->hash_value;
643 	}
644 
645 	PMD_RX_LOG(DEBUG,
646 		   "port %u:%u RX id %"PRIu64" size %u type %#x ol_flags %#"PRIx64,
647 		   rxq->port_id, rxq->queue_id, rxb->xactid,
648 		   m->pkt_len, m->packet_type, m->ol_flags);
649 
650 	++rxq->stats.packets;
651 	rxq->stats.bytes += m->pkt_len;
652 	hn_update_packet_stats(&rxq->stats, m);
653 
654 	if (unlikely(rte_ring_sp_enqueue(rxq->rx_ring, m) != 0)) {
655 		++rxq->stats.ring_full;
656 		PMD_RX_LOG(DEBUG, "rx ring full");
657 		if (use_extbuf)
658 			rte_pktmbuf_detach_extbuf(m);
659 		rte_pktmbuf_free(m);
660 	}
661 }
662 
663 static void hn_rndis_rx_data(struct hn_rx_queue *rxq,
664 			     struct hn_rx_bufinfo *rxb,
665 			     void *data, uint32_t dlen)
666 {
667 	unsigned int data_off, data_len, pktinfo_off, pktinfo_len;
668 	const struct rndis_packet_msg *pkt = data;
669 	struct hn_rxinfo info = {
670 		.vlan_info = HN_NDIS_VLAN_INFO_INVALID,
671 		.csum_info = HN_NDIS_RXCSUM_INFO_INVALID,
672 		.hash_info = HN_NDIS_HASH_INFO_INVALID,
673 	};
674 	int err;
675 
676 	hn_rndis_dump(pkt);
677 
678 	if (unlikely(dlen < sizeof(*pkt)))
679 		goto error;
680 
681 	if (unlikely(dlen < pkt->len))
682 		goto error; /* truncated RNDIS from host */
683 
684 	if (unlikely(pkt->len < pkt->datalen
685 		     + pkt->oobdatalen + pkt->pktinfolen))
686 		goto error;
687 
688 	if (unlikely(pkt->datalen == 0))
689 		goto error;
690 
691 	/* Check offsets. */
692 	if (unlikely(pkt->dataoffset < RNDIS_PACKET_MSG_OFFSET_MIN))
693 		goto error;
694 
695 	if (likely(pkt->pktinfooffset > 0) &&
696 	    unlikely(pkt->pktinfooffset < RNDIS_PACKET_MSG_OFFSET_MIN ||
697 		     (pkt->pktinfooffset & RNDIS_PACKET_MSG_OFFSET_ALIGNMASK)))
698 		goto error;
699 
700 	data_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
701 	data_len = pkt->datalen;
702 	pktinfo_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->pktinfooffset);
703 	pktinfo_len = pkt->pktinfolen;
704 
705 	if (likely(pktinfo_len > 0)) {
706 		err = hn_rndis_rxinfo((const uint8_t *)pkt + pktinfo_off,
707 				      pktinfo_len, &info);
708 		if (err)
709 			goto error;
710 	}
711 
712 	if (unlikely(data_off + data_len > pkt->len))
713 		goto error;
714 
715 	if (unlikely(data_len < RTE_ETHER_HDR_LEN))
716 		goto error;
717 
718 	hn_rxpkt(rxq, rxb, data, data_off, data_len, &info);
719 	return;
720 error:
721 	++rxq->stats.errors;
722 }
723 
724 static void
725 hn_rndis_receive(struct rte_eth_dev *dev, struct hn_rx_queue *rxq,
726 		 struct hn_rx_bufinfo *rxb, void *buf, uint32_t len)
727 {
728 	const struct rndis_msghdr *hdr = buf;
729 
730 	switch (hdr->type) {
731 	case RNDIS_PACKET_MSG:
732 		if (dev->data->dev_started)
733 			hn_rndis_rx_data(rxq, rxb, buf, len);
734 		break;
735 
736 	case RNDIS_INDICATE_STATUS_MSG:
737 		hn_rndis_link_status(dev, buf);
738 		break;
739 
740 	case RNDIS_INITIALIZE_CMPLT:
741 	case RNDIS_QUERY_CMPLT:
742 	case RNDIS_SET_CMPLT:
743 		hn_rndis_receive_response(rxq->hv, buf, len);
744 		break;
745 
746 	default:
747 		PMD_DRV_LOG(NOTICE,
748 			    "unexpected RNDIS message (type %#x len %u)",
749 			    hdr->type, len);
750 		break;
751 	}
752 }
753 
754 static void
755 hn_nvs_handle_rxbuf(struct rte_eth_dev *dev,
756 		    struct hn_data *hv,
757 		    struct hn_rx_queue *rxq,
758 		    const struct vmbus_chanpkt_hdr *hdr,
759 		    const void *buf)
760 {
761 	const struct vmbus_chanpkt_rxbuf *pkt;
762 	const struct hn_nvs_hdr *nvs_hdr = buf;
763 	uint32_t rxbuf_sz = hv->rxbuf_res->len;
764 	char *rxbuf = hv->rxbuf_res->addr;
765 	unsigned int i, hlen, count;
766 	struct hn_rx_bufinfo *rxb;
767 
768 	/* At minimum we need type header */
769 	if (unlikely(vmbus_chanpkt_datalen(hdr) < sizeof(*nvs_hdr))) {
770 		PMD_RX_LOG(ERR, "invalid receive nvs RNDIS");
771 		return;
772 	}
773 
774 	/* Make sure that this is a RNDIS message. */
775 	if (unlikely(nvs_hdr->type != NVS_TYPE_RNDIS)) {
776 		PMD_RX_LOG(ERR, "nvs type %u, not RNDIS",
777 			   nvs_hdr->type);
778 		return;
779 	}
780 
781 	hlen = vmbus_chanpkt_getlen(hdr->hlen);
782 	if (unlikely(hlen < sizeof(*pkt))) {
783 		PMD_RX_LOG(ERR, "invalid rxbuf chanpkt");
784 		return;
785 	}
786 
787 	pkt = container_of(hdr, const struct vmbus_chanpkt_rxbuf, hdr);
788 	if (unlikely(pkt->rxbuf_id != NVS_RXBUF_SIG)) {
789 		PMD_RX_LOG(ERR, "invalid rxbuf_id 0x%08x",
790 			   pkt->rxbuf_id);
791 		return;
792 	}
793 
794 	count = pkt->rxbuf_cnt;
795 	if (unlikely(hlen < offsetof(struct vmbus_chanpkt_rxbuf,
796 				     rxbuf[count]))) {
797 		PMD_RX_LOG(ERR, "invalid rxbuf_cnt %u", count);
798 		return;
799 	}
800 
801 	if (pkt->hdr.xactid > hv->rxbuf_section_cnt) {
802 		PMD_RX_LOG(ERR, "invalid rxbuf section id %" PRIx64,
803 			   pkt->hdr.xactid);
804 		return;
805 	}
806 
807 	/* Setup receive buffer info to allow for callback */
808 	rxb = hn_rx_buf_init(rxq, pkt);
809 
810 	/* Each range represents 1 RNDIS pkt that contains 1 Ethernet frame */
811 	for (i = 0; i < count; ++i) {
812 		unsigned int ofs, len;
813 
814 		ofs = pkt->rxbuf[i].ofs;
815 		len = pkt->rxbuf[i].len;
816 
817 		if (unlikely(ofs + len > rxbuf_sz)) {
818 			PMD_RX_LOG(ERR,
819 				   "%uth RNDIS msg overflow ofs %u, len %u",
820 				   i, ofs, len);
821 			continue;
822 		}
823 
824 		if (unlikely(len == 0)) {
825 			PMD_RX_LOG(ERR, "%uth RNDIS msg len %u", i, len);
826 			continue;
827 		}
828 
829 		hn_rndis_receive(dev, rxq, rxb,
830 				 rxbuf + ofs, len);
831 	}
832 
833 	/* Send ACK now if external mbuf not used */
834 	if (rte_mbuf_ext_refcnt_update(&rxb->shinfo, -1) == 0)
835 		hn_nvs_ack_rxbuf(rxb->chan, rxb->xactid);
836 }
837 
838 /*
839  * Called when NVS inband events are received.
840  * Send up a two part message with port_id and the NVS message
841  * to the pipe to the netvsc-vf-event control thread.
842  */
843 static void hn_nvs_handle_notify(struct rte_eth_dev *dev,
844 				 const struct vmbus_chanpkt_hdr *pkt,
845 				 const void *data)
846 {
847 	const struct hn_nvs_hdr *hdr = data;
848 
849 	switch (hdr->type) {
850 	case NVS_TYPE_TXTBL_NOTE:
851 		/* Transmit indirection table has locking problems
852 		 * in DPDK and therefore not implemented
853 		 */
854 		PMD_DRV_LOG(DEBUG, "host notify of transmit indirection table");
855 		break;
856 
857 	case NVS_TYPE_VFASSOC_NOTE:
858 		hn_nvs_handle_vfassoc(dev, pkt, data);
859 		break;
860 
861 	default:
862 		PMD_DRV_LOG(INFO,
863 			    "got notify, nvs type %u", hdr->type);
864 	}
865 }
866 
867 struct hn_rx_queue *hn_rx_queue_alloc(struct hn_data *hv,
868 				      uint16_t queue_id,
869 				      unsigned int socket_id)
870 {
871 	struct hn_rx_queue *rxq;
872 
873 	rxq = rte_zmalloc_socket("HN_RXQ", sizeof(*rxq),
874 				 RTE_CACHE_LINE_SIZE, socket_id);
875 	if (!rxq)
876 		return NULL;
877 
878 	rxq->hv = hv;
879 	rxq->chan = hv->channels[queue_id];
880 	rte_spinlock_init(&rxq->ring_lock);
881 	rxq->port_id = hv->port_id;
882 	rxq->queue_id = queue_id;
883 	rxq->event_sz = HN_RXQ_EVENT_DEFAULT;
884 	rxq->event_buf = rte_malloc_socket("HN_EVENTS", HN_RXQ_EVENT_DEFAULT,
885 					   RTE_CACHE_LINE_SIZE, socket_id);
886 	if (!rxq->event_buf) {
887 		rte_free(rxq);
888 		return NULL;
889 	}
890 
891 	return rxq;
892 }
893 
894 void
895 hn_dev_rx_queue_info(struct rte_eth_dev *dev, uint16_t queue_id,
896 		     struct rte_eth_rxq_info *qinfo)
897 {
898 	struct hn_rx_queue *rxq = dev->data->rx_queues[queue_id];
899 
900 	qinfo->mp = rxq->mb_pool;
901 	qinfo->nb_desc = rxq->rx_ring->size;
902 	qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
903 }
904 
905 int
906 hn_dev_rx_queue_setup(struct rte_eth_dev *dev,
907 		      uint16_t queue_idx, uint16_t nb_desc,
908 		      unsigned int socket_id,
909 		      const struct rte_eth_rxconf *rx_conf,
910 		      struct rte_mempool *mp)
911 {
912 	struct hn_data *hv = dev->data->dev_private;
913 	char ring_name[RTE_RING_NAMESIZE];
914 	struct hn_rx_queue *rxq;
915 	unsigned int count;
916 	int error = -ENOMEM;
917 
918 	PMD_INIT_FUNC_TRACE();
919 
920 	if (queue_idx == 0) {
921 		rxq = hv->primary;
922 	} else {
923 		rxq = hn_rx_queue_alloc(hv, queue_idx, socket_id);
924 		if (!rxq)
925 			return -ENOMEM;
926 	}
927 
928 	rxq->mb_pool = mp;
929 	count = rte_mempool_avail_count(mp) / dev->data->nb_rx_queues;
930 	if (nb_desc == 0 || nb_desc > count)
931 		nb_desc = count;
932 
933 	/*
934 	 * Staging ring from receive event logic to rx_pkts.
935 	 * rx_pkts assumes caller is handling multi-thread issue.
936 	 * event logic has locking.
937 	 */
938 	snprintf(ring_name, sizeof(ring_name),
939 		 "hn_rx_%u_%u", dev->data->port_id, queue_idx);
940 	rxq->rx_ring = rte_ring_create(ring_name,
941 				       rte_align32pow2(nb_desc),
942 				       socket_id, 0);
943 	if (!rxq->rx_ring)
944 		goto fail;
945 
946 	error = hn_vf_rx_queue_setup(dev, queue_idx, nb_desc,
947 				     socket_id, rx_conf, mp);
948 	if (error)
949 		goto fail;
950 
951 	dev->data->rx_queues[queue_idx] = rxq;
952 	return 0;
953 
954 fail:
955 	rte_ring_free(rxq->rx_ring);
956 	rte_free(rxq->event_buf);
957 	rte_free(rxq);
958 	return error;
959 }
960 
961 static void
962 hn_rx_queue_free(struct hn_rx_queue *rxq, bool keep_primary)
963 {
964 
965 	if (!rxq)
966 		return;
967 
968 	rte_ring_free(rxq->rx_ring);
969 	rxq->rx_ring = NULL;
970 	rxq->mb_pool = NULL;
971 
972 	hn_vf_rx_queue_release(rxq->hv, rxq->queue_id);
973 
974 	/* Keep primary queue to allow for control operations */
975 	if (keep_primary && rxq == rxq->hv->primary)
976 		return;
977 
978 	rte_free(rxq->event_buf);
979 	rte_free(rxq);
980 }
981 
982 void
983 hn_dev_rx_queue_release(void *arg)
984 {
985 	struct hn_rx_queue *rxq = arg;
986 
987 	PMD_INIT_FUNC_TRACE();
988 
989 	hn_rx_queue_free(rxq, true);
990 }
991 
992 /*
993  * Get the number of used descriptor in a rx queue
994  * For this device that means how many packets are pending in the ring.
995  */
996 uint32_t
997 hn_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t queue_id)
998 {
999 	struct hn_rx_queue *rxq = dev->data->rx_queues[queue_id];
1000 
1001 	return rte_ring_count(rxq->rx_ring);
1002 }
1003 
1004 /*
1005  * Check the status of a Rx descriptor in the queue
1006  *
1007  * returns:
1008  *  - -EINVAL               - offset outside of ring
1009  *  - RTE_ETH_RX_DESC_AVAIL - no data available yet
1010  *  - RTE_ETH_RX_DESC_DONE  - data is waiting in stagin ring
1011  */
1012 int hn_dev_rx_queue_status(void *arg, uint16_t offset)
1013 {
1014 	const struct hn_rx_queue *rxq = arg;
1015 
1016 	hn_process_events(rxq->hv, rxq->queue_id, 0);
1017 	if (offset >= rxq->rx_ring->capacity)
1018 		return -EINVAL;
1019 
1020 	if (offset < rte_ring_count(rxq->rx_ring))
1021 		return RTE_ETH_RX_DESC_DONE;
1022 	else
1023 		return RTE_ETH_RX_DESC_AVAIL;
1024 }
1025 
1026 int
1027 hn_dev_tx_done_cleanup(void *arg, uint32_t free_cnt)
1028 {
1029 	struct hn_tx_queue *txq = arg;
1030 
1031 	return hn_process_events(txq->hv, txq->queue_id, free_cnt);
1032 }
1033 
1034 /*
1035  * Process pending events on the channel.
1036  * Called from both Rx queue poll and Tx cleanup
1037  */
1038 uint32_t hn_process_events(struct hn_data *hv, uint16_t queue_id,
1039 			   uint32_t tx_limit)
1040 {
1041 	struct rte_eth_dev *dev = &rte_eth_devices[hv->port_id];
1042 	struct hn_rx_queue *rxq;
1043 	uint32_t bytes_read = 0;
1044 	uint32_t tx_done = 0;
1045 	int ret = 0;
1046 
1047 	rxq = queue_id == 0 ? hv->primary : dev->data->rx_queues[queue_id];
1048 
1049 	/*
1050 	 * Since channel is shared between Rx and TX queue need to have a lock
1051 	 * since DPDK does not force same CPU to be used for Rx/Tx.
1052 	 */
1053 	if (unlikely(!rte_spinlock_trylock(&rxq->ring_lock)))
1054 		return 0;
1055 
1056 	for (;;) {
1057 		const struct vmbus_chanpkt_hdr *pkt;
1058 		uint32_t len = rxq->event_sz;
1059 		const void *data;
1060 
1061 retry:
1062 		ret = rte_vmbus_chan_recv_raw(rxq->chan, rxq->event_buf, &len);
1063 		if (ret == -EAGAIN)
1064 			break;	/* ring is empty */
1065 
1066 		if (unlikely(ret == -ENOBUFS)) {
1067 			/* event buffer not large enough to read ring */
1068 
1069 			PMD_DRV_LOG(DEBUG,
1070 				    "event buffer expansion (need %u)", len);
1071 			rxq->event_sz = len + len / 4;
1072 			rxq->event_buf = rte_realloc(rxq->event_buf, rxq->event_sz,
1073 						     RTE_CACHE_LINE_SIZE);
1074 			if (rxq->event_buf)
1075 				goto retry;
1076 			/* out of memory, no more events now */
1077 			rxq->event_sz = 0;
1078 			break;
1079 		}
1080 
1081 		if (unlikely(ret <= 0)) {
1082 			/* This indicates a failure to communicate (or worse) */
1083 			rte_exit(EXIT_FAILURE,
1084 				 "vmbus ring buffer error: %d", ret);
1085 		}
1086 
1087 		bytes_read += ret;
1088 		pkt = (const struct vmbus_chanpkt_hdr *)rxq->event_buf;
1089 		data = (char *)rxq->event_buf + vmbus_chanpkt_getlen(pkt->hlen);
1090 
1091 		switch (pkt->type) {
1092 		case VMBUS_CHANPKT_TYPE_COMP:
1093 			++tx_done;
1094 			hn_nvs_handle_comp(dev, queue_id, pkt, data);
1095 			break;
1096 
1097 		case VMBUS_CHANPKT_TYPE_RXBUF:
1098 			hn_nvs_handle_rxbuf(dev, hv, rxq, pkt, data);
1099 			break;
1100 
1101 		case VMBUS_CHANPKT_TYPE_INBAND:
1102 			hn_nvs_handle_notify(dev, pkt, data);
1103 			break;
1104 
1105 		default:
1106 			PMD_DRV_LOG(ERR, "unknown chan pkt %u", pkt->type);
1107 			break;
1108 		}
1109 
1110 		if (tx_limit && tx_done >= tx_limit)
1111 			break;
1112 	}
1113 
1114 	if (bytes_read > 0)
1115 		rte_vmbus_chan_signal_read(rxq->chan, bytes_read);
1116 
1117 	rte_spinlock_unlock(&rxq->ring_lock);
1118 
1119 	return tx_done;
1120 }
1121 
1122 static void hn_append_to_chim(struct hn_tx_queue *txq,
1123 			      struct rndis_packet_msg *pkt,
1124 			      const struct rte_mbuf *m)
1125 {
1126 	struct hn_txdesc *txd = txq->agg_txd;
1127 	uint8_t *buf = (uint8_t *)pkt;
1128 	unsigned int data_offs;
1129 
1130 	hn_rndis_dump(pkt);
1131 
1132 	data_offs = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
1133 	txd->chim_size += pkt->len;
1134 	txd->data_size += m->pkt_len;
1135 	++txd->packets;
1136 	hn_update_packet_stats(&txq->stats, m);
1137 
1138 	for (; m; m = m->next) {
1139 		uint16_t len = rte_pktmbuf_data_len(m);
1140 
1141 		rte_memcpy(buf + data_offs,
1142 			   rte_pktmbuf_mtod(m, const char *), len);
1143 		data_offs += len;
1144 	}
1145 }
1146 
1147 /*
1148  * Send pending aggregated data in chimney buffer (if any).
1149  * Returns error if send was unsuccessful because channel ring buffer
1150  * was full.
1151  */
1152 static int hn_flush_txagg(struct hn_tx_queue *txq, bool *need_sig)
1153 
1154 {
1155 	struct hn_txdesc *txd = txq->agg_txd;
1156 	struct hn_nvs_rndis rndis;
1157 	int ret;
1158 
1159 	if (!txd)
1160 		return 0;
1161 
1162 	rndis = (struct hn_nvs_rndis) {
1163 		.type = NVS_TYPE_RNDIS,
1164 		.rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1165 		.chim_idx = txd->chim_index,
1166 		.chim_sz = txd->chim_size,
1167 	};
1168 
1169 	PMD_TX_LOG(DEBUG, "port %u:%u tx %u size %u",
1170 		   txq->port_id, txq->queue_id, txd->chim_index, txd->chim_size);
1171 
1172 	ret = hn_nvs_send(txq->chan, VMBUS_CHANPKT_FLAG_RC,
1173 			  &rndis, sizeof(rndis), (uintptr_t)txd, need_sig);
1174 
1175 	if (likely(ret == 0))
1176 		hn_reset_txagg(txq);
1177 	else if (ret == -EAGAIN) {
1178 		PMD_TX_LOG(DEBUG, "port %u:%u channel full",
1179 			   txq->port_id, txq->queue_id);
1180 		++txq->stats.channel_full;
1181 	} else {
1182 		++txq->stats.errors;
1183 
1184 		PMD_DRV_LOG(NOTICE, "port %u:%u send failed: %d",
1185 			   txq->port_id, txq->queue_id, ret);
1186 	}
1187 	return ret;
1188 }
1189 
1190 /*
1191  * Try and find a place in a send chimney buffer to put
1192  * the small packet. If space is available, this routine
1193  * returns a pointer of where to place the data.
1194  * If no space, caller should try direct transmit.
1195  */
1196 static void *
1197 hn_try_txagg(struct hn_data *hv, struct hn_tx_queue *txq,
1198 	     struct hn_txdesc *txd, uint32_t pktsize)
1199 {
1200 	struct hn_txdesc *agg_txd = txq->agg_txd;
1201 	struct rndis_packet_msg *pkt;
1202 	void *chim;
1203 
1204 	if (agg_txd) {
1205 		unsigned int padding, olen;
1206 
1207 		/*
1208 		 * Update the previous RNDIS packet's total length,
1209 		 * it can be increased due to the mandatory alignment
1210 		 * padding for this RNDIS packet.  And update the
1211 		 * aggregating txdesc's chimney sending buffer size
1212 		 * accordingly.
1213 		 *
1214 		 * Zero-out the padding, as required by the RNDIS spec.
1215 		 */
1216 		pkt = txq->agg_prevpkt;
1217 		olen = pkt->len;
1218 		padding = RTE_ALIGN(olen, txq->agg_align) - olen;
1219 		if (padding > 0) {
1220 			agg_txd->chim_size += padding;
1221 			pkt->len += padding;
1222 			memset((uint8_t *)pkt + olen, 0, padding);
1223 		}
1224 
1225 		chim = (uint8_t *)pkt + pkt->len;
1226 		txq->agg_prevpkt = chim;
1227 		txq->agg_pktleft--;
1228 		txq->agg_szleft -= pktsize;
1229 		if (txq->agg_szleft < HN_PKTSIZE_MIN(txq->agg_align)) {
1230 			/*
1231 			 * Probably can't aggregate more packets,
1232 			 * flush this aggregating txdesc proactively.
1233 			 */
1234 			txq->agg_pktleft = 0;
1235 		}
1236 
1237 		hn_txd_put(txq, txd);
1238 		return chim;
1239 	}
1240 
1241 	txd->chim_index = hn_chim_alloc(hv);
1242 	if (txd->chim_index == NVS_CHIM_IDX_INVALID)
1243 		return NULL;
1244 
1245 	chim = (uint8_t *)hv->chim_res->addr
1246 			+ txd->chim_index * hv->chim_szmax;
1247 
1248 	txq->agg_txd = txd;
1249 	txq->agg_pktleft = txq->agg_pktmax - 1;
1250 	txq->agg_szleft = txq->agg_szmax - pktsize;
1251 	txq->agg_prevpkt = chim;
1252 
1253 	return chim;
1254 }
1255 
1256 static inline void *
1257 hn_rndis_pktinfo_append(struct rndis_packet_msg *pkt,
1258 			uint32_t pi_dlen, uint32_t pi_type)
1259 {
1260 	const uint32_t pi_size = RNDIS_PKTINFO_SIZE(pi_dlen);
1261 	struct rndis_pktinfo *pi;
1262 
1263 	/*
1264 	 * Per-packet-info does not move; it only grows.
1265 	 *
1266 	 * NOTE:
1267 	 * pktinfooffset in this phase counts from the beginning
1268 	 * of rndis_packet_msg.
1269 	 */
1270 	pi = (struct rndis_pktinfo *)((uint8_t *)pkt + hn_rndis_pktlen(pkt));
1271 
1272 	pkt->pktinfolen += pi_size;
1273 
1274 	pi->size = pi_size;
1275 	pi->type = pi_type;
1276 	pi->offset = RNDIS_PKTINFO_OFFSET;
1277 
1278 	return pi->data;
1279 }
1280 
1281 /* Put RNDIS header and packet info on packet */
1282 static void hn_encap(struct rndis_packet_msg *pkt,
1283 		     uint16_t queue_id,
1284 		     const struct rte_mbuf *m)
1285 {
1286 	unsigned int hlen = m->l2_len + m->l3_len;
1287 	uint32_t *pi_data;
1288 	uint32_t pkt_hlen;
1289 
1290 	pkt->type = RNDIS_PACKET_MSG;
1291 	pkt->len = m->pkt_len;
1292 	pkt->dataoffset = 0;
1293 	pkt->datalen = m->pkt_len;
1294 	pkt->oobdataoffset = 0;
1295 	pkt->oobdatalen = 0;
1296 	pkt->oobdataelements = 0;
1297 	pkt->pktinfooffset = sizeof(*pkt);
1298 	pkt->pktinfolen = 0;
1299 	pkt->vchandle = 0;
1300 	pkt->reserved = 0;
1301 
1302 	/*
1303 	 * Set the hash value for this packet, to the queue_id to cause
1304 	 * TX done event for this packet on the right channel.
1305 	 */
1306 	pi_data = hn_rndis_pktinfo_append(pkt, NDIS_HASH_VALUE_SIZE,
1307 					  NDIS_PKTINFO_TYPE_HASHVAL);
1308 	*pi_data = queue_id;
1309 
1310 	if (m->ol_flags & PKT_TX_VLAN_PKT) {
1311 		pi_data = hn_rndis_pktinfo_append(pkt, NDIS_VLAN_INFO_SIZE,
1312 						  NDIS_PKTINFO_TYPE_VLAN);
1313 		*pi_data = m->vlan_tci;
1314 	}
1315 
1316 	if (m->ol_flags & PKT_TX_TCP_SEG) {
1317 		pi_data = hn_rndis_pktinfo_append(pkt, NDIS_LSO2_INFO_SIZE,
1318 						  NDIS_PKTINFO_TYPE_LSO);
1319 
1320 		if (m->ol_flags & PKT_TX_IPV6) {
1321 			*pi_data = NDIS_LSO2_INFO_MAKEIPV6(hlen,
1322 							   m->tso_segsz);
1323 		} else {
1324 			*pi_data = NDIS_LSO2_INFO_MAKEIPV4(hlen,
1325 							   m->tso_segsz);
1326 		}
1327 	} else if (m->ol_flags &
1328 		   (PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM | PKT_TX_IP_CKSUM)) {
1329 		pi_data = hn_rndis_pktinfo_append(pkt, NDIS_TXCSUM_INFO_SIZE,
1330 						  NDIS_PKTINFO_TYPE_CSUM);
1331 		*pi_data = 0;
1332 
1333 		if (m->ol_flags & PKT_TX_IPV6)
1334 			*pi_data |= NDIS_TXCSUM_INFO_IPV6;
1335 		if (m->ol_flags & PKT_TX_IPV4) {
1336 			*pi_data |= NDIS_TXCSUM_INFO_IPV4;
1337 
1338 			if (m->ol_flags & PKT_TX_IP_CKSUM)
1339 				*pi_data |= NDIS_TXCSUM_INFO_IPCS;
1340 		}
1341 
1342 		if (m->ol_flags & PKT_TX_TCP_CKSUM)
1343 			*pi_data |= NDIS_TXCSUM_INFO_MKTCPCS(hlen);
1344 		else if (m->ol_flags & PKT_TX_UDP_CKSUM)
1345 			*pi_data |= NDIS_TXCSUM_INFO_MKUDPCS(hlen);
1346 	}
1347 
1348 	pkt_hlen = pkt->pktinfooffset + pkt->pktinfolen;
1349 	/* Fixup RNDIS packet message total length */
1350 	pkt->len += pkt_hlen;
1351 
1352 	/* Convert RNDIS packet message offsets */
1353 	pkt->dataoffset = hn_rndis_pktmsg_offset(pkt_hlen);
1354 	pkt->pktinfooffset = hn_rndis_pktmsg_offset(pkt->pktinfooffset);
1355 }
1356 
1357 /* How many scatter gather list elements ar needed */
1358 static unsigned int hn_get_slots(const struct rte_mbuf *m)
1359 {
1360 	unsigned int slots = 1; /* for RNDIS header */
1361 
1362 	while (m) {
1363 		unsigned int size = rte_pktmbuf_data_len(m);
1364 		unsigned int offs = rte_mbuf_data_iova(m) & PAGE_MASK;
1365 
1366 		slots += (offs + size + PAGE_SIZE - 1) / PAGE_SIZE;
1367 		m = m->next;
1368 	}
1369 
1370 	return slots;
1371 }
1372 
1373 /* Build scatter gather list from chained mbuf */
1374 static unsigned int hn_fill_sg(struct vmbus_gpa *sg,
1375 			       const struct rte_mbuf *m)
1376 {
1377 	unsigned int segs = 0;
1378 
1379 	while (m) {
1380 		rte_iova_t addr = rte_mbuf_data_iova(m);
1381 		unsigned int page = addr / PAGE_SIZE;
1382 		unsigned int offset = addr & PAGE_MASK;
1383 		unsigned int len = rte_pktmbuf_data_len(m);
1384 
1385 		while (len > 0) {
1386 			unsigned int bytes = RTE_MIN(len, PAGE_SIZE - offset);
1387 
1388 			sg[segs].page = page;
1389 			sg[segs].ofs = offset;
1390 			sg[segs].len = bytes;
1391 			segs++;
1392 
1393 			++page;
1394 			offset = 0;
1395 			len -= bytes;
1396 		}
1397 		m = m->next;
1398 	}
1399 
1400 	return segs;
1401 }
1402 
1403 /* Transmit directly from mbuf */
1404 static int hn_xmit_sg(struct hn_tx_queue *txq,
1405 		      const struct hn_txdesc *txd, const struct rte_mbuf *m,
1406 		      bool *need_sig)
1407 {
1408 	struct vmbus_gpa sg[hn_get_slots(m)];
1409 	struct hn_nvs_rndis nvs_rndis = {
1410 		.type = NVS_TYPE_RNDIS,
1411 		.rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1412 		.chim_sz = txd->chim_size,
1413 	};
1414 	rte_iova_t addr;
1415 	unsigned int segs;
1416 
1417 	/* attach aggregation data if present */
1418 	if (txd->chim_size > 0)
1419 		nvs_rndis.chim_idx = txd->chim_index;
1420 	else
1421 		nvs_rndis.chim_idx = NVS_CHIM_IDX_INVALID;
1422 
1423 	hn_rndis_dump(txd->rndis_pkt);
1424 
1425 	/* pass IOVA of rndis header in first segment */
1426 	addr = rte_malloc_virt2iova(txq->tx_rndis);
1427 	if (unlikely(addr == RTE_BAD_IOVA)) {
1428 		PMD_DRV_LOG(ERR, "RNDIS transmit can not get iova");
1429 		return -EINVAL;
1430 	}
1431 	addr = addr + ((char *)txd->rndis_pkt - (char *)txq->tx_rndis);
1432 
1433 	sg[0].page = addr / PAGE_SIZE;
1434 	sg[0].ofs = addr & PAGE_MASK;
1435 	sg[0].len = RNDIS_PACKET_MSG_OFFSET_ABS(hn_rndis_pktlen(txd->rndis_pkt));
1436 	segs = 1;
1437 
1438 	hn_update_packet_stats(&txq->stats, m);
1439 
1440 	segs += hn_fill_sg(sg + 1, m);
1441 
1442 	PMD_TX_LOG(DEBUG, "port %u:%u tx %u segs %u size %u",
1443 		   txq->port_id, txq->queue_id, txd->chim_index,
1444 		   segs, nvs_rndis.chim_sz);
1445 
1446 	return hn_nvs_send_sglist(txq->chan, sg, segs,
1447 				  &nvs_rndis, sizeof(nvs_rndis),
1448 				  (uintptr_t)txd, need_sig);
1449 }
1450 
1451 uint16_t
1452 hn_xmit_pkts(void *ptxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1453 {
1454 	struct hn_tx_queue *txq = ptxq;
1455 	uint16_t queue_id = txq->queue_id;
1456 	struct hn_data *hv = txq->hv;
1457 	struct rte_eth_dev *vf_dev;
1458 	bool need_sig = false;
1459 	uint16_t nb_tx, tx_thresh;
1460 	int ret;
1461 
1462 	if (unlikely(hv->closed))
1463 		return 0;
1464 
1465 	/*
1466 	 * Always check for events on the primary channel
1467 	 * because that is where hotplug notifications occur.
1468 	 */
1469 	tx_thresh = RTE_MAX(txq->free_thresh, nb_pkts);
1470 	if (txq->queue_id == 0 ||
1471 	    rte_mempool_avail_count(txq->txdesc_pool) < tx_thresh)
1472 		hn_process_events(hv, txq->queue_id, 0);
1473 
1474 	/* Transmit over VF if present and up */
1475 	rte_rwlock_read_lock(&hv->vf_lock);
1476 	vf_dev = hn_get_vf_dev(hv);
1477 	if (vf_dev && vf_dev->data->dev_started) {
1478 		void *sub_q = vf_dev->data->tx_queues[queue_id];
1479 
1480 		nb_tx = (*vf_dev->tx_pkt_burst)(sub_q, tx_pkts, nb_pkts);
1481 		rte_rwlock_read_unlock(&hv->vf_lock);
1482 		return nb_tx;
1483 	}
1484 	rte_rwlock_read_unlock(&hv->vf_lock);
1485 
1486 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1487 		struct rte_mbuf *m = tx_pkts[nb_tx];
1488 		uint32_t pkt_size = m->pkt_len + HN_RNDIS_PKT_LEN;
1489 		struct rndis_packet_msg *pkt;
1490 		struct hn_txdesc *txd;
1491 
1492 		txd = hn_txd_get(txq);
1493 		if (txd == NULL)
1494 			break;
1495 
1496 		/* For small packets aggregate them in chimney buffer */
1497 		if (m->pkt_len < HN_TXCOPY_THRESHOLD && pkt_size <= txq->agg_szmax) {
1498 			/* If this packet will not fit, then flush  */
1499 			if (txq->agg_pktleft == 0 ||
1500 			    RTE_ALIGN(pkt_size, txq->agg_align) > txq->agg_szleft) {
1501 				if (hn_flush_txagg(txq, &need_sig))
1502 					goto fail;
1503 			}
1504 
1505 
1506 			pkt = hn_try_txagg(hv, txq, txd, pkt_size);
1507 			if (unlikely(!pkt))
1508 				break;
1509 
1510 			hn_encap(pkt, queue_id, m);
1511 			hn_append_to_chim(txq, pkt, m);
1512 
1513 			rte_pktmbuf_free(m);
1514 
1515 			/* if buffer is full, flush */
1516 			if (txq->agg_pktleft == 0 &&
1517 			    hn_flush_txagg(txq, &need_sig))
1518 				goto fail;
1519 		} else {
1520 			/* Send any outstanding packets in buffer */
1521 			if (txq->agg_txd && hn_flush_txagg(txq, &need_sig))
1522 				goto fail;
1523 
1524 			pkt = txd->rndis_pkt;
1525 			txd->m = m;
1526 			txd->data_size = m->pkt_len;
1527 			++txd->packets;
1528 
1529 			hn_encap(pkt, queue_id, m);
1530 
1531 			ret = hn_xmit_sg(txq, txd, m, &need_sig);
1532 			if (unlikely(ret != 0)) {
1533 				if (ret == -EAGAIN) {
1534 					PMD_TX_LOG(DEBUG, "sg channel full");
1535 					++txq->stats.channel_full;
1536 				} else {
1537 					PMD_DRV_LOG(NOTICE, "sg send failed: %d", ret);
1538 					++txq->stats.errors;
1539 				}
1540 				hn_txd_put(txq, txd);
1541 				goto fail;
1542 			}
1543 		}
1544 	}
1545 
1546 	/* If partial buffer left, then try and send it.
1547 	 * if that fails, then reuse it on next send.
1548 	 */
1549 	hn_flush_txagg(txq, &need_sig);
1550 
1551 fail:
1552 	if (need_sig)
1553 		rte_vmbus_chan_signal_tx(txq->chan);
1554 
1555 	return nb_tx;
1556 }
1557 
1558 static uint16_t
1559 hn_recv_vf(uint16_t vf_port, const struct hn_rx_queue *rxq,
1560 	   struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1561 {
1562 	uint16_t i, n;
1563 
1564 	if (unlikely(nb_pkts == 0))
1565 		return 0;
1566 
1567 	n = rte_eth_rx_burst(vf_port, rxq->queue_id, rx_pkts, nb_pkts);
1568 
1569 	/* relabel the received mbufs */
1570 	for (i = 0; i < n; i++)
1571 		rx_pkts[i]->port = rxq->port_id;
1572 
1573 	return n;
1574 }
1575 
1576 uint16_t
1577 hn_recv_pkts(void *prxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1578 {
1579 	struct hn_rx_queue *rxq = prxq;
1580 	struct hn_data *hv = rxq->hv;
1581 	struct rte_eth_dev *vf_dev;
1582 	uint16_t nb_rcv;
1583 
1584 	if (unlikely(hv->closed))
1585 		return 0;
1586 
1587 	/* Check for new completions (and hotplug) */
1588 	if (likely(rte_ring_count(rxq->rx_ring) < nb_pkts))
1589 		hn_process_events(hv, rxq->queue_id, 0);
1590 
1591 	/* Always check the vmbus path for multicast and new flows */
1592 	nb_rcv = rte_ring_sc_dequeue_burst(rxq->rx_ring,
1593 					   (void **)rx_pkts, nb_pkts, NULL);
1594 
1595 	/* If VF is available, check that as well */
1596 	rte_rwlock_read_lock(&hv->vf_lock);
1597 	vf_dev = hn_get_vf_dev(hv);
1598 	if (vf_dev && vf_dev->data->dev_started)
1599 		nb_rcv += hn_recv_vf(vf_dev->data->port_id, rxq,
1600 				     rx_pkts + nb_rcv, nb_pkts - nb_rcv);
1601 
1602 	rte_rwlock_read_unlock(&hv->vf_lock);
1603 	return nb_rcv;
1604 }
1605 
1606 void
1607 hn_dev_free_queues(struct rte_eth_dev *dev)
1608 {
1609 	unsigned int i;
1610 
1611 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
1612 		struct hn_rx_queue *rxq = dev->data->rx_queues[i];
1613 
1614 		hn_rx_queue_free(rxq, false);
1615 		dev->data->rx_queues[i] = NULL;
1616 	}
1617 	dev->data->nb_rx_queues = 0;
1618 
1619 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
1620 		hn_dev_tx_queue_release(dev->data->tx_queues[i]);
1621 		dev->data->tx_queues[i] = NULL;
1622 	}
1623 	dev->data->nb_tx_queues = 0;
1624 }
1625