xref: /dpdk/drivers/net/netvsc/hn_rxtx.c (revision cb440babbd45a80c059f8bc80e87c48d09086fd7)
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 void
323 hn_dev_tx_queue_info(struct rte_eth_dev *dev, uint16_t queue_id,
324 		     struct rte_eth_txq_info *qinfo)
325 {
326 	struct hn_tx_queue *txq = dev->data->tx_queues[queue_id];
327 
328 	qinfo->nb_desc = txq->txdesc_pool->size;
329 	qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads;
330 }
331 
332 static struct hn_txdesc *hn_txd_get(struct hn_tx_queue *txq)
333 {
334 	struct hn_txdesc *txd;
335 
336 	if (rte_mempool_get(txq->txdesc_pool, (void **)&txd)) {
337 		++txq->stats.ring_full;
338 		PMD_TX_LOG(DEBUG, "tx pool exhausted!");
339 		return NULL;
340 	}
341 
342 	txd->m = NULL;
343 	txd->packets = 0;
344 	txd->data_size = 0;
345 	txd->chim_size = 0;
346 
347 	return txd;
348 }
349 
350 static void hn_txd_put(struct hn_tx_queue *txq, struct hn_txdesc *txd)
351 {
352 	rte_mempool_put(txq->txdesc_pool, txd);
353 }
354 
355 void
356 hn_dev_tx_queue_release(void *arg)
357 {
358 	struct hn_tx_queue *txq = arg;
359 
360 	PMD_INIT_FUNC_TRACE();
361 
362 	if (!txq)
363 		return;
364 
365 	if (txq->txdesc_pool)
366 		rte_mempool_free(txq->txdesc_pool);
367 
368 	rte_free(txq->tx_rndis);
369 	rte_free(txq);
370 }
371 
372 /*
373  * Check the status of a Tx descriptor in the queue.
374  *
375  * returns:
376  *  - -EINVAL              - offset outside of tx_descriptor pool.
377  *  - RTE_ETH_TX_DESC_FULL - descriptor is not acknowledged by host.
378  *  - RTE_ETH_TX_DESC_DONE - descriptor is available.
379  */
380 int hn_dev_tx_descriptor_status(void *arg, uint16_t offset)
381 {
382 	const struct hn_tx_queue *txq = arg;
383 
384 	hn_process_events(txq->hv, txq->queue_id, 0);
385 
386 	if (offset >= rte_mempool_avail_count(txq->txdesc_pool))
387 		return -EINVAL;
388 
389 	if (offset < rte_mempool_in_use_count(txq->txdesc_pool))
390 		return RTE_ETH_TX_DESC_FULL;
391 	else
392 		return RTE_ETH_TX_DESC_DONE;
393 }
394 
395 static void
396 hn_nvs_send_completed(struct rte_eth_dev *dev, uint16_t queue_id,
397 		      unsigned long xactid, const struct hn_nvs_rndis_ack *ack)
398 {
399 	struct hn_data *hv = dev->data->dev_private;
400 	struct hn_txdesc *txd = (struct hn_txdesc *)xactid;
401 	struct hn_tx_queue *txq;
402 
403 	/* Control packets are sent with xacid == 0 */
404 	if (!txd)
405 		return;
406 
407 	txq = dev->data->tx_queues[queue_id];
408 	if (likely(ack->status == NVS_STATUS_OK)) {
409 		PMD_TX_LOG(DEBUG, "port %u:%u complete tx %u packets %u bytes %u",
410 			   txq->port_id, txq->queue_id, txd->chim_index,
411 			   txd->packets, txd->data_size);
412 		txq->stats.bytes += txd->data_size;
413 		txq->stats.packets += txd->packets;
414 	} else {
415 		PMD_DRV_LOG(NOTICE, "port %u:%u complete tx %u failed status %u",
416 			    txq->port_id, txq->queue_id, txd->chim_index, ack->status);
417 		++txq->stats.errors;
418 	}
419 
420 	if (txd->chim_index != NVS_CHIM_IDX_INVALID)
421 		hn_chim_free(hv, txd->chim_index);
422 
423 	rte_pktmbuf_free(txd->m);
424 	hn_txd_put(txq, txd);
425 }
426 
427 /* Handle transmit completion events */
428 static void
429 hn_nvs_handle_comp(struct rte_eth_dev *dev, uint16_t queue_id,
430 		   const struct vmbus_chanpkt_hdr *pkt,
431 		   const void *data)
432 {
433 	const struct hn_nvs_hdr *hdr = data;
434 
435 	switch (hdr->type) {
436 	case NVS_TYPE_RNDIS_ACK:
437 		hn_nvs_send_completed(dev, queue_id, pkt->xactid, data);
438 		break;
439 
440 	default:
441 		PMD_DRV_LOG(NOTICE, "unexpected send completion type %u",
442 			   hdr->type);
443 	}
444 }
445 
446 /* Parse per-packet info (meta data) */
447 static int
448 hn_rndis_rxinfo(const void *info_data, unsigned int info_dlen,
449 		struct hn_rxinfo *info)
450 {
451 	const struct rndis_pktinfo *pi = info_data;
452 	uint32_t mask = 0;
453 
454 	while (info_dlen != 0) {
455 		const void *data;
456 		uint32_t dlen;
457 
458 		if (unlikely(info_dlen < sizeof(*pi)))
459 			return -EINVAL;
460 
461 		if (unlikely(info_dlen < pi->size))
462 			return -EINVAL;
463 		info_dlen -= pi->size;
464 
465 		if (unlikely(pi->size & RNDIS_PKTINFO_SIZE_ALIGNMASK))
466 			return -EINVAL;
467 		if (unlikely(pi->size < pi->offset))
468 			return -EINVAL;
469 
470 		dlen = pi->size - pi->offset;
471 		data = pi->data;
472 
473 		switch (pi->type) {
474 		case NDIS_PKTINFO_TYPE_VLAN:
475 			if (unlikely(dlen < NDIS_VLAN_INFO_SIZE))
476 				return -EINVAL;
477 			info->vlan_info = *((const uint32_t *)data);
478 			mask |= HN_RXINFO_VLAN;
479 			break;
480 
481 		case NDIS_PKTINFO_TYPE_CSUM:
482 			if (unlikely(dlen < NDIS_RXCSUM_INFO_SIZE))
483 				return -EINVAL;
484 			info->csum_info = *((const uint32_t *)data);
485 			mask |= HN_RXINFO_CSUM;
486 			break;
487 
488 		case NDIS_PKTINFO_TYPE_HASHVAL:
489 			if (unlikely(dlen < NDIS_HASH_VALUE_SIZE))
490 				return -EINVAL;
491 			info->hash_value = *((const uint32_t *)data);
492 			mask |= HN_RXINFO_HASHVAL;
493 			break;
494 
495 		case NDIS_PKTINFO_TYPE_HASHINF:
496 			if (unlikely(dlen < NDIS_HASH_INFO_SIZE))
497 				return -EINVAL;
498 			info->hash_info = *((const uint32_t *)data);
499 			mask |= HN_RXINFO_HASHINF;
500 			break;
501 
502 		default:
503 			goto next;
504 		}
505 
506 		if (mask == HN_RXINFO_ALL)
507 			break; /* All found; done */
508 next:
509 		pi = (const struct rndis_pktinfo *)
510 		    ((const uint8_t *)pi + pi->size);
511 	}
512 
513 	/*
514 	 * Final fixup.
515 	 * - If there is no hash value, invalidate the hash info.
516 	 */
517 	if (!(mask & HN_RXINFO_HASHVAL))
518 		info->hash_info = HN_NDIS_HASH_INFO_INVALID;
519 	return 0;
520 }
521 
522 static void hn_rx_buf_free_cb(void *buf __rte_unused, void *opaque)
523 {
524 	struct hn_rx_bufinfo *rxb = opaque;
525 	struct hn_data *hv = rxb->hv;
526 
527 	rte_atomic32_dec(&hv->rxbuf_outstanding);
528 	hn_nvs_ack_rxbuf(rxb->chan, rxb->xactid);
529 }
530 
531 static struct hn_rx_bufinfo *hn_rx_buf_init(const struct hn_rx_queue *rxq,
532 					    const struct vmbus_chanpkt_rxbuf *pkt)
533 {
534 	struct hn_rx_bufinfo *rxb;
535 
536 	rxb = rxq->hv->rxbuf_info + pkt->hdr.xactid;
537 	rxb->chan = rxq->chan;
538 	rxb->xactid = pkt->hdr.xactid;
539 	rxb->hv = rxq->hv;
540 
541 	rxb->shinfo.free_cb = hn_rx_buf_free_cb;
542 	rxb->shinfo.fcb_opaque = rxb;
543 	rte_mbuf_ext_refcnt_set(&rxb->shinfo, 1);
544 	return rxb;
545 }
546 
547 static void hn_rxpkt(struct hn_rx_queue *rxq, struct hn_rx_bufinfo *rxb,
548 		     uint8_t *data, unsigned int headroom, unsigned int dlen,
549 		     const struct hn_rxinfo *info)
550 {
551 	struct hn_data *hv = rxq->hv;
552 	struct rte_mbuf *m;
553 	bool use_extbuf = false;
554 
555 	m = rte_pktmbuf_alloc(rxq->mb_pool);
556 	if (unlikely(!m)) {
557 		struct rte_eth_dev *dev =
558 			&rte_eth_devices[rxq->port_id];
559 
560 		dev->data->rx_mbuf_alloc_failed++;
561 		return;
562 	}
563 
564 	/*
565 	 * For large packets, avoid copy if possible but need to keep
566 	 * some space available in receive area for later packets.
567 	 */
568 	if (dlen >= HN_RXCOPY_THRESHOLD &&
569 	    (uint32_t)rte_atomic32_read(&hv->rxbuf_outstanding) <
570 			hv->rxbuf_section_cnt / 2) {
571 		struct rte_mbuf_ext_shared_info *shinfo;
572 		const void *rxbuf;
573 		rte_iova_t iova;
574 
575 		/*
576 		 * Build an external mbuf that points to recveive area.
577 		 * Use refcount to handle multiple packets in same
578 		 * receive buffer section.
579 		 */
580 		rxbuf = hv->rxbuf_res->addr;
581 		iova = rte_mem_virt2iova(rxbuf) + RTE_PTR_DIFF(data, rxbuf);
582 		shinfo = &rxb->shinfo;
583 
584 		/* shinfo is already set to 1 by the caller */
585 		if (rte_mbuf_ext_refcnt_update(shinfo, 1) == 2)
586 			rte_atomic32_inc(&hv->rxbuf_outstanding);
587 
588 		rte_pktmbuf_attach_extbuf(m, data, iova,
589 					  dlen + headroom, shinfo);
590 		m->data_off = headroom;
591 		use_extbuf = true;
592 	} else {
593 		/* Mbuf's in pool must be large enough to hold small packets */
594 		if (unlikely(rte_pktmbuf_tailroom(m) < dlen)) {
595 			rte_pktmbuf_free_seg(m);
596 			++rxq->stats.errors;
597 			return;
598 		}
599 		rte_memcpy(rte_pktmbuf_mtod(m, void *),
600 			   data + headroom, dlen);
601 	}
602 
603 	m->port = rxq->port_id;
604 	m->pkt_len = dlen;
605 	m->data_len = dlen;
606 	m->packet_type = rte_net_get_ptype(m, NULL,
607 					   RTE_PTYPE_L2_MASK |
608 					   RTE_PTYPE_L3_MASK |
609 					   RTE_PTYPE_L4_MASK);
610 
611 	if (info->vlan_info != HN_NDIS_VLAN_INFO_INVALID) {
612 		m->vlan_tci = info->vlan_info;
613 		m->ol_flags |= PKT_RX_VLAN_STRIPPED | PKT_RX_VLAN;
614 
615 		/* NDIS always strips tag, put it back if necessary */
616 		if (!hv->vlan_strip && rte_vlan_insert(&m)) {
617 			PMD_DRV_LOG(DEBUG, "vlan insert failed");
618 			++rxq->stats.errors;
619 			if (use_extbuf)
620 				rte_pktmbuf_detach_extbuf(m);
621 			rte_pktmbuf_free(m);
622 			return;
623 		}
624 	}
625 
626 	if (info->csum_info != HN_NDIS_RXCSUM_INFO_INVALID) {
627 		if (info->csum_info & NDIS_RXCSUM_INFO_IPCS_OK)
628 			m->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
629 
630 		if (info->csum_info & (NDIS_RXCSUM_INFO_UDPCS_OK
631 				       | NDIS_RXCSUM_INFO_TCPCS_OK))
632 			m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
633 		else if (info->csum_info & (NDIS_RXCSUM_INFO_TCPCS_FAILED
634 					    | NDIS_RXCSUM_INFO_UDPCS_FAILED))
635 			m->ol_flags |= PKT_RX_L4_CKSUM_BAD;
636 	}
637 
638 	if (info->hash_info != HN_NDIS_HASH_INFO_INVALID) {
639 		m->ol_flags |= PKT_RX_RSS_HASH;
640 		m->hash.rss = info->hash_value;
641 	}
642 
643 	PMD_RX_LOG(DEBUG,
644 		   "port %u:%u RX id %"PRIu64" size %u type %#x ol_flags %#"PRIx64,
645 		   rxq->port_id, rxq->queue_id, rxb->xactid,
646 		   m->pkt_len, m->packet_type, m->ol_flags);
647 
648 	++rxq->stats.packets;
649 	rxq->stats.bytes += m->pkt_len;
650 	hn_update_packet_stats(&rxq->stats, m);
651 
652 	if (unlikely(rte_ring_sp_enqueue(rxq->rx_ring, m) != 0)) {
653 		++rxq->stats.ring_full;
654 		PMD_RX_LOG(DEBUG, "rx ring full");
655 		if (use_extbuf)
656 			rte_pktmbuf_detach_extbuf(m);
657 		rte_pktmbuf_free(m);
658 	}
659 }
660 
661 static void hn_rndis_rx_data(struct hn_rx_queue *rxq,
662 			     struct hn_rx_bufinfo *rxb,
663 			     void *data, uint32_t dlen)
664 {
665 	unsigned int data_off, data_len, pktinfo_off, pktinfo_len;
666 	const struct rndis_packet_msg *pkt = data;
667 	struct hn_rxinfo info = {
668 		.vlan_info = HN_NDIS_VLAN_INFO_INVALID,
669 		.csum_info = HN_NDIS_RXCSUM_INFO_INVALID,
670 		.hash_info = HN_NDIS_HASH_INFO_INVALID,
671 	};
672 	int err;
673 
674 	hn_rndis_dump(pkt);
675 
676 	if (unlikely(dlen < sizeof(*pkt)))
677 		goto error;
678 
679 	if (unlikely(dlen < pkt->len))
680 		goto error; /* truncated RNDIS from host */
681 
682 	if (unlikely(pkt->len < pkt->datalen
683 		     + pkt->oobdatalen + pkt->pktinfolen))
684 		goto error;
685 
686 	if (unlikely(pkt->datalen == 0))
687 		goto error;
688 
689 	/* Check offsets. */
690 	if (unlikely(pkt->dataoffset < RNDIS_PACKET_MSG_OFFSET_MIN))
691 		goto error;
692 
693 	if (likely(pkt->pktinfooffset > 0) &&
694 	    unlikely(pkt->pktinfooffset < RNDIS_PACKET_MSG_OFFSET_MIN ||
695 		     (pkt->pktinfooffset & RNDIS_PACKET_MSG_OFFSET_ALIGNMASK)))
696 		goto error;
697 
698 	data_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
699 	data_len = pkt->datalen;
700 	pktinfo_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->pktinfooffset);
701 	pktinfo_len = pkt->pktinfolen;
702 
703 	if (likely(pktinfo_len > 0)) {
704 		err = hn_rndis_rxinfo((const uint8_t *)pkt + pktinfo_off,
705 				      pktinfo_len, &info);
706 		if (err)
707 			goto error;
708 	}
709 
710 	if (unlikely(data_off + data_len > pkt->len))
711 		goto error;
712 
713 	if (unlikely(data_len < RTE_ETHER_HDR_LEN))
714 		goto error;
715 
716 	hn_rxpkt(rxq, rxb, data, data_off, data_len, &info);
717 	return;
718 error:
719 	++rxq->stats.errors;
720 }
721 
722 static void
723 hn_rndis_receive(struct rte_eth_dev *dev, struct hn_rx_queue *rxq,
724 		 struct hn_rx_bufinfo *rxb, void *buf, uint32_t len)
725 {
726 	const struct rndis_msghdr *hdr = buf;
727 
728 	switch (hdr->type) {
729 	case RNDIS_PACKET_MSG:
730 		if (dev->data->dev_started)
731 			hn_rndis_rx_data(rxq, rxb, buf, len);
732 		break;
733 
734 	case RNDIS_INDICATE_STATUS_MSG:
735 		hn_rndis_link_status(dev, buf);
736 		break;
737 
738 	case RNDIS_INITIALIZE_CMPLT:
739 	case RNDIS_QUERY_CMPLT:
740 	case RNDIS_SET_CMPLT:
741 		hn_rndis_receive_response(rxq->hv, buf, len);
742 		break;
743 
744 	default:
745 		PMD_DRV_LOG(NOTICE,
746 			    "unexpected RNDIS message (type %#x len %u)",
747 			    hdr->type, len);
748 		break;
749 	}
750 }
751 
752 static void
753 hn_nvs_handle_rxbuf(struct rte_eth_dev *dev,
754 		    struct hn_data *hv,
755 		    struct hn_rx_queue *rxq,
756 		    const struct vmbus_chanpkt_hdr *hdr,
757 		    const void *buf)
758 {
759 	const struct vmbus_chanpkt_rxbuf *pkt;
760 	const struct hn_nvs_hdr *nvs_hdr = buf;
761 	uint32_t rxbuf_sz = hv->rxbuf_res->len;
762 	char *rxbuf = hv->rxbuf_res->addr;
763 	unsigned int i, hlen, count;
764 	struct hn_rx_bufinfo *rxb;
765 
766 	/* At minimum we need type header */
767 	if (unlikely(vmbus_chanpkt_datalen(hdr) < sizeof(*nvs_hdr))) {
768 		PMD_RX_LOG(ERR, "invalid receive nvs RNDIS");
769 		return;
770 	}
771 
772 	/* Make sure that this is a RNDIS message. */
773 	if (unlikely(nvs_hdr->type != NVS_TYPE_RNDIS)) {
774 		PMD_RX_LOG(ERR, "nvs type %u, not RNDIS",
775 			   nvs_hdr->type);
776 		return;
777 	}
778 
779 	hlen = vmbus_chanpkt_getlen(hdr->hlen);
780 	if (unlikely(hlen < sizeof(*pkt))) {
781 		PMD_RX_LOG(ERR, "invalid rxbuf chanpkt");
782 		return;
783 	}
784 
785 	pkt = container_of(hdr, const struct vmbus_chanpkt_rxbuf, hdr);
786 	if (unlikely(pkt->rxbuf_id != NVS_RXBUF_SIG)) {
787 		PMD_RX_LOG(ERR, "invalid rxbuf_id 0x%08x",
788 			   pkt->rxbuf_id);
789 		return;
790 	}
791 
792 	count = pkt->rxbuf_cnt;
793 	if (unlikely(hlen < offsetof(struct vmbus_chanpkt_rxbuf,
794 				     rxbuf[count]))) {
795 		PMD_RX_LOG(ERR, "invalid rxbuf_cnt %u", count);
796 		return;
797 	}
798 
799 	if (pkt->hdr.xactid > hv->rxbuf_section_cnt) {
800 		PMD_RX_LOG(ERR, "invalid rxbuf section id %" PRIx64,
801 			   pkt->hdr.xactid);
802 		return;
803 	}
804 
805 	/* Setup receive buffer info to allow for callback */
806 	rxb = hn_rx_buf_init(rxq, pkt);
807 
808 	/* Each range represents 1 RNDIS pkt that contains 1 Ethernet frame */
809 	for (i = 0; i < count; ++i) {
810 		unsigned int ofs, len;
811 
812 		ofs = pkt->rxbuf[i].ofs;
813 		len = pkt->rxbuf[i].len;
814 
815 		if (unlikely(ofs + len > rxbuf_sz)) {
816 			PMD_RX_LOG(ERR,
817 				   "%uth RNDIS msg overflow ofs %u, len %u",
818 				   i, ofs, len);
819 			continue;
820 		}
821 
822 		if (unlikely(len == 0)) {
823 			PMD_RX_LOG(ERR, "%uth RNDIS msg len %u", i, len);
824 			continue;
825 		}
826 
827 		hn_rndis_receive(dev, rxq, rxb,
828 				 rxbuf + ofs, len);
829 	}
830 
831 	/* Send ACK now if external mbuf not used */
832 	if (rte_mbuf_ext_refcnt_update(&rxb->shinfo, -1) == 0)
833 		hn_nvs_ack_rxbuf(rxb->chan, rxb->xactid);
834 }
835 
836 /*
837  * Called when NVS inband events are received.
838  * Send up a two part message with port_id and the NVS message
839  * to the pipe to the netvsc-vf-event control thread.
840  */
841 static void hn_nvs_handle_notify(struct rte_eth_dev *dev,
842 				 const struct vmbus_chanpkt_hdr *pkt,
843 				 const void *data)
844 {
845 	const struct hn_nvs_hdr *hdr = data;
846 
847 	switch (hdr->type) {
848 	case NVS_TYPE_TXTBL_NOTE:
849 		/* Transmit indirection table has locking problems
850 		 * in DPDK and therefore not implemented
851 		 */
852 		PMD_DRV_LOG(DEBUG, "host notify of transmit indirection table");
853 		break;
854 
855 	case NVS_TYPE_VFASSOC_NOTE:
856 		hn_nvs_handle_vfassoc(dev, pkt, data);
857 		break;
858 
859 	default:
860 		PMD_DRV_LOG(INFO,
861 			    "got notify, nvs type %u", hdr->type);
862 	}
863 }
864 
865 struct hn_rx_queue *hn_rx_queue_alloc(struct hn_data *hv,
866 				      uint16_t queue_id,
867 				      unsigned int socket_id)
868 {
869 	struct hn_rx_queue *rxq;
870 
871 	rxq = rte_zmalloc_socket("HN_RXQ", sizeof(*rxq),
872 				 RTE_CACHE_LINE_SIZE, socket_id);
873 	if (!rxq)
874 		return NULL;
875 
876 	rxq->hv = hv;
877 	rxq->chan = hv->channels[queue_id];
878 	rte_spinlock_init(&rxq->ring_lock);
879 	rxq->port_id = hv->port_id;
880 	rxq->queue_id = queue_id;
881 	rxq->event_sz = HN_RXQ_EVENT_DEFAULT;
882 	rxq->event_buf = rte_malloc_socket("HN_EVENTS", HN_RXQ_EVENT_DEFAULT,
883 					   RTE_CACHE_LINE_SIZE, socket_id);
884 	if (!rxq->event_buf) {
885 		rte_free(rxq);
886 		return NULL;
887 	}
888 
889 	return rxq;
890 }
891 
892 void
893 hn_dev_rx_queue_info(struct rte_eth_dev *dev, uint16_t queue_id,
894 		     struct rte_eth_rxq_info *qinfo)
895 {
896 	struct hn_rx_queue *rxq = dev->data->rx_queues[queue_id];
897 
898 	qinfo->mp = rxq->mb_pool;
899 	qinfo->nb_desc = rxq->rx_ring->size;
900 	qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
901 }
902 
903 int
904 hn_dev_rx_queue_setup(struct rte_eth_dev *dev,
905 		      uint16_t queue_idx, uint16_t nb_desc,
906 		      unsigned int socket_id,
907 		      const struct rte_eth_rxconf *rx_conf,
908 		      struct rte_mempool *mp)
909 {
910 	struct hn_data *hv = dev->data->dev_private;
911 	char ring_name[RTE_RING_NAMESIZE];
912 	struct hn_rx_queue *rxq;
913 	unsigned int count;
914 	int error = -ENOMEM;
915 
916 	PMD_INIT_FUNC_TRACE();
917 
918 	if (queue_idx == 0) {
919 		rxq = hv->primary;
920 	} else {
921 		rxq = hn_rx_queue_alloc(hv, queue_idx, socket_id);
922 		if (!rxq)
923 			return -ENOMEM;
924 	}
925 
926 	rxq->mb_pool = mp;
927 	count = rte_mempool_avail_count(mp) / dev->data->nb_rx_queues;
928 	if (nb_desc == 0 || nb_desc > count)
929 		nb_desc = count;
930 
931 	/*
932 	 * Staging ring from receive event logic to rx_pkts.
933 	 * rx_pkts assumes caller is handling multi-thread issue.
934 	 * event logic has locking.
935 	 */
936 	snprintf(ring_name, sizeof(ring_name),
937 		 "hn_rx_%u_%u", dev->data->port_id, queue_idx);
938 	rxq->rx_ring = rte_ring_create(ring_name,
939 				       rte_align32pow2(nb_desc),
940 				       socket_id, 0);
941 	if (!rxq->rx_ring)
942 		goto fail;
943 
944 	error = hn_vf_rx_queue_setup(dev, queue_idx, nb_desc,
945 				     socket_id, rx_conf, mp);
946 	if (error)
947 		goto fail;
948 
949 	dev->data->rx_queues[queue_idx] = rxq;
950 	return 0;
951 
952 fail:
953 	rte_ring_free(rxq->rx_ring);
954 	rte_free(rxq->event_buf);
955 	rte_free(rxq);
956 	return error;
957 }
958 
959 static void
960 hn_rx_queue_free(struct hn_rx_queue *rxq, bool keep_primary)
961 {
962 
963 	if (!rxq)
964 		return;
965 
966 	rte_ring_free(rxq->rx_ring);
967 	rxq->rx_ring = NULL;
968 	rxq->mb_pool = NULL;
969 
970 	hn_vf_rx_queue_release(rxq->hv, rxq->queue_id);
971 
972 	/* Keep primary queue to allow for control operations */
973 	if (keep_primary && rxq == rxq->hv->primary)
974 		return;
975 
976 	rte_free(rxq->event_buf);
977 	rte_free(rxq);
978 }
979 
980 void
981 hn_dev_rx_queue_release(void *arg)
982 {
983 	struct hn_rx_queue *rxq = arg;
984 
985 	PMD_INIT_FUNC_TRACE();
986 
987 	hn_rx_queue_free(rxq, true);
988 }
989 
990 /*
991  * Get the number of used descriptor in a rx queue
992  * For this device that means how many packets are pending in the ring.
993  */
994 uint32_t
995 hn_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t queue_id)
996 {
997 	struct hn_rx_queue *rxq = dev->data->rx_queues[queue_id];
998 
999 	return rte_ring_count(rxq->rx_ring);
1000 }
1001 
1002 /*
1003  * Check the status of a Rx descriptor in the queue
1004  *
1005  * returns:
1006  *  - -EINVAL               - offset outside of ring
1007  *  - RTE_ETH_RX_DESC_AVAIL - no data available yet
1008  *  - RTE_ETH_RX_DESC_DONE  - data is waiting in stagin ring
1009  */
1010 int hn_dev_rx_queue_status(void *arg, uint16_t offset)
1011 {
1012 	const struct hn_rx_queue *rxq = arg;
1013 
1014 	hn_process_events(rxq->hv, rxq->queue_id, 0);
1015 	if (offset >= rxq->rx_ring->capacity)
1016 		return -EINVAL;
1017 
1018 	if (offset < rte_ring_count(rxq->rx_ring))
1019 		return RTE_ETH_RX_DESC_DONE;
1020 	else
1021 		return RTE_ETH_RX_DESC_AVAIL;
1022 }
1023 
1024 int
1025 hn_dev_tx_done_cleanup(void *arg, uint32_t free_cnt)
1026 {
1027 	struct hn_tx_queue *txq = arg;
1028 
1029 	return hn_process_events(txq->hv, txq->queue_id, free_cnt);
1030 }
1031 
1032 /*
1033  * Process pending events on the channel.
1034  * Called from both Rx queue poll and Tx cleanup
1035  */
1036 uint32_t hn_process_events(struct hn_data *hv, uint16_t queue_id,
1037 			   uint32_t tx_limit)
1038 {
1039 	struct rte_eth_dev *dev = &rte_eth_devices[hv->port_id];
1040 	struct hn_rx_queue *rxq;
1041 	uint32_t bytes_read = 0;
1042 	uint32_t tx_done = 0;
1043 	int ret = 0;
1044 
1045 	rxq = queue_id == 0 ? hv->primary : dev->data->rx_queues[queue_id];
1046 
1047 	/*
1048 	 * Since channel is shared between Rx and TX queue need to have a lock
1049 	 * since DPDK does not force same CPU to be used for Rx/Tx.
1050 	 */
1051 	if (unlikely(!rte_spinlock_trylock(&rxq->ring_lock)))
1052 		return 0;
1053 
1054 	for (;;) {
1055 		const struct vmbus_chanpkt_hdr *pkt;
1056 		uint32_t len = rxq->event_sz;
1057 		const void *data;
1058 
1059 retry:
1060 		ret = rte_vmbus_chan_recv_raw(rxq->chan, rxq->event_buf, &len);
1061 		if (ret == -EAGAIN)
1062 			break;	/* ring is empty */
1063 
1064 		if (unlikely(ret == -ENOBUFS)) {
1065 			/* event buffer not large enough to read ring */
1066 
1067 			PMD_DRV_LOG(DEBUG,
1068 				    "event buffer expansion (need %u)", len);
1069 			rxq->event_sz = len + len / 4;
1070 			rxq->event_buf = rte_realloc(rxq->event_buf, rxq->event_sz,
1071 						     RTE_CACHE_LINE_SIZE);
1072 			if (rxq->event_buf)
1073 				goto retry;
1074 			/* out of memory, no more events now */
1075 			rxq->event_sz = 0;
1076 			break;
1077 		}
1078 
1079 		if (unlikely(ret <= 0)) {
1080 			/* This indicates a failure to communicate (or worse) */
1081 			rte_exit(EXIT_FAILURE,
1082 				 "vmbus ring buffer error: %d", ret);
1083 		}
1084 
1085 		bytes_read += ret;
1086 		pkt = (const struct vmbus_chanpkt_hdr *)rxq->event_buf;
1087 		data = (char *)rxq->event_buf + vmbus_chanpkt_getlen(pkt->hlen);
1088 
1089 		switch (pkt->type) {
1090 		case VMBUS_CHANPKT_TYPE_COMP:
1091 			++tx_done;
1092 			hn_nvs_handle_comp(dev, queue_id, pkt, data);
1093 			break;
1094 
1095 		case VMBUS_CHANPKT_TYPE_RXBUF:
1096 			hn_nvs_handle_rxbuf(dev, hv, rxq, pkt, data);
1097 			break;
1098 
1099 		case VMBUS_CHANPKT_TYPE_INBAND:
1100 			hn_nvs_handle_notify(dev, pkt, data);
1101 			break;
1102 
1103 		default:
1104 			PMD_DRV_LOG(ERR, "unknown chan pkt %u", pkt->type);
1105 			break;
1106 		}
1107 
1108 		if (tx_limit && tx_done >= tx_limit)
1109 			break;
1110 	}
1111 
1112 	if (bytes_read > 0)
1113 		rte_vmbus_chan_signal_read(rxq->chan, bytes_read);
1114 
1115 	rte_spinlock_unlock(&rxq->ring_lock);
1116 
1117 	return tx_done;
1118 }
1119 
1120 static void hn_append_to_chim(struct hn_tx_queue *txq,
1121 			      struct rndis_packet_msg *pkt,
1122 			      const struct rte_mbuf *m)
1123 {
1124 	struct hn_txdesc *txd = txq->agg_txd;
1125 	uint8_t *buf = (uint8_t *)pkt;
1126 	unsigned int data_offs;
1127 
1128 	hn_rndis_dump(pkt);
1129 
1130 	data_offs = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset);
1131 	txd->chim_size += pkt->len;
1132 	txd->data_size += m->pkt_len;
1133 	++txd->packets;
1134 	hn_update_packet_stats(&txq->stats, m);
1135 
1136 	for (; m; m = m->next) {
1137 		uint16_t len = rte_pktmbuf_data_len(m);
1138 
1139 		rte_memcpy(buf + data_offs,
1140 			   rte_pktmbuf_mtod(m, const char *), len);
1141 		data_offs += len;
1142 	}
1143 }
1144 
1145 /*
1146  * Send pending aggregated data in chimney buffer (if any).
1147  * Returns error if send was unsuccessful because channel ring buffer
1148  * was full.
1149  */
1150 static int hn_flush_txagg(struct hn_tx_queue *txq, bool *need_sig)
1151 
1152 {
1153 	struct hn_txdesc *txd = txq->agg_txd;
1154 	struct hn_nvs_rndis rndis;
1155 	int ret;
1156 
1157 	if (!txd)
1158 		return 0;
1159 
1160 	rndis = (struct hn_nvs_rndis) {
1161 		.type = NVS_TYPE_RNDIS,
1162 		.rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1163 		.chim_idx = txd->chim_index,
1164 		.chim_sz = txd->chim_size,
1165 	};
1166 
1167 	PMD_TX_LOG(DEBUG, "port %u:%u tx %u size %u",
1168 		   txq->port_id, txq->queue_id, txd->chim_index, txd->chim_size);
1169 
1170 	ret = hn_nvs_send(txq->chan, VMBUS_CHANPKT_FLAG_RC,
1171 			  &rndis, sizeof(rndis), (uintptr_t)txd, need_sig);
1172 
1173 	if (likely(ret == 0))
1174 		hn_reset_txagg(txq);
1175 	else if (ret == -EAGAIN) {
1176 		PMD_TX_LOG(DEBUG, "port %u:%u channel full",
1177 			   txq->port_id, txq->queue_id);
1178 		++txq->stats.channel_full;
1179 	} else {
1180 		++txq->stats.errors;
1181 
1182 		PMD_DRV_LOG(NOTICE, "port %u:%u send failed: %d",
1183 			   txq->port_id, txq->queue_id, ret);
1184 	}
1185 	return ret;
1186 }
1187 
1188 /*
1189  * Try and find a place in a send chimney buffer to put
1190  * the small packet. If space is available, this routine
1191  * returns a pointer of where to place the data.
1192  * If no space, caller should try direct transmit.
1193  */
1194 static void *
1195 hn_try_txagg(struct hn_data *hv, struct hn_tx_queue *txq,
1196 	     struct hn_txdesc *txd, uint32_t pktsize)
1197 {
1198 	struct hn_txdesc *agg_txd = txq->agg_txd;
1199 	struct rndis_packet_msg *pkt;
1200 	void *chim;
1201 
1202 	if (agg_txd) {
1203 		unsigned int padding, olen;
1204 
1205 		/*
1206 		 * Update the previous RNDIS packet's total length,
1207 		 * it can be increased due to the mandatory alignment
1208 		 * padding for this RNDIS packet.  And update the
1209 		 * aggregating txdesc's chimney sending buffer size
1210 		 * accordingly.
1211 		 *
1212 		 * Zero-out the padding, as required by the RNDIS spec.
1213 		 */
1214 		pkt = txq->agg_prevpkt;
1215 		olen = pkt->len;
1216 		padding = RTE_ALIGN(olen, txq->agg_align) - olen;
1217 		if (padding > 0) {
1218 			agg_txd->chim_size += padding;
1219 			pkt->len += padding;
1220 			memset((uint8_t *)pkt + olen, 0, padding);
1221 		}
1222 
1223 		chim = (uint8_t *)pkt + pkt->len;
1224 		txq->agg_prevpkt = chim;
1225 		txq->agg_pktleft--;
1226 		txq->agg_szleft -= pktsize;
1227 		if (txq->agg_szleft < HN_PKTSIZE_MIN(txq->agg_align)) {
1228 			/*
1229 			 * Probably can't aggregate more packets,
1230 			 * flush this aggregating txdesc proactively.
1231 			 */
1232 			txq->agg_pktleft = 0;
1233 		}
1234 
1235 		hn_txd_put(txq, txd);
1236 		return chim;
1237 	}
1238 
1239 	txd->chim_index = hn_chim_alloc(hv);
1240 	if (txd->chim_index == NVS_CHIM_IDX_INVALID)
1241 		return NULL;
1242 
1243 	chim = (uint8_t *)hv->chim_res->addr
1244 			+ txd->chim_index * hv->chim_szmax;
1245 
1246 	txq->agg_txd = txd;
1247 	txq->agg_pktleft = txq->agg_pktmax - 1;
1248 	txq->agg_szleft = txq->agg_szmax - pktsize;
1249 	txq->agg_prevpkt = chim;
1250 
1251 	return chim;
1252 }
1253 
1254 static inline void *
1255 hn_rndis_pktinfo_append(struct rndis_packet_msg *pkt,
1256 			uint32_t pi_dlen, uint32_t pi_type)
1257 {
1258 	const uint32_t pi_size = RNDIS_PKTINFO_SIZE(pi_dlen);
1259 	struct rndis_pktinfo *pi;
1260 
1261 	/*
1262 	 * Per-packet-info does not move; it only grows.
1263 	 *
1264 	 * NOTE:
1265 	 * pktinfooffset in this phase counts from the beginning
1266 	 * of rndis_packet_msg.
1267 	 */
1268 	pi = (struct rndis_pktinfo *)((uint8_t *)pkt + hn_rndis_pktlen(pkt));
1269 
1270 	pkt->pktinfolen += pi_size;
1271 
1272 	pi->size = pi_size;
1273 	pi->type = pi_type;
1274 	pi->offset = RNDIS_PKTINFO_OFFSET;
1275 
1276 	return pi->data;
1277 }
1278 
1279 /* Put RNDIS header and packet info on packet */
1280 static void hn_encap(struct rndis_packet_msg *pkt,
1281 		     uint16_t queue_id,
1282 		     const struct rte_mbuf *m)
1283 {
1284 	unsigned int hlen = m->l2_len + m->l3_len;
1285 	uint32_t *pi_data;
1286 	uint32_t pkt_hlen;
1287 
1288 	pkt->type = RNDIS_PACKET_MSG;
1289 	pkt->len = m->pkt_len;
1290 	pkt->dataoffset = 0;
1291 	pkt->datalen = m->pkt_len;
1292 	pkt->oobdataoffset = 0;
1293 	pkt->oobdatalen = 0;
1294 	pkt->oobdataelements = 0;
1295 	pkt->pktinfooffset = sizeof(*pkt);
1296 	pkt->pktinfolen = 0;
1297 	pkt->vchandle = 0;
1298 	pkt->reserved = 0;
1299 
1300 	/*
1301 	 * Set the hash value for this packet, to the queue_id to cause
1302 	 * TX done event for this packet on the right channel.
1303 	 */
1304 	pi_data = hn_rndis_pktinfo_append(pkt, NDIS_HASH_VALUE_SIZE,
1305 					  NDIS_PKTINFO_TYPE_HASHVAL);
1306 	*pi_data = queue_id;
1307 
1308 	if (m->ol_flags & PKT_TX_VLAN_PKT) {
1309 		pi_data = hn_rndis_pktinfo_append(pkt, NDIS_VLAN_INFO_SIZE,
1310 						  NDIS_PKTINFO_TYPE_VLAN);
1311 		*pi_data = m->vlan_tci;
1312 	}
1313 
1314 	if (m->ol_flags & PKT_TX_TCP_SEG) {
1315 		pi_data = hn_rndis_pktinfo_append(pkt, NDIS_LSO2_INFO_SIZE,
1316 						  NDIS_PKTINFO_TYPE_LSO);
1317 
1318 		if (m->ol_flags & PKT_TX_IPV6) {
1319 			*pi_data = NDIS_LSO2_INFO_MAKEIPV6(hlen,
1320 							   m->tso_segsz);
1321 		} else {
1322 			*pi_data = NDIS_LSO2_INFO_MAKEIPV4(hlen,
1323 							   m->tso_segsz);
1324 		}
1325 	} else if (m->ol_flags &
1326 		   (PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM | PKT_TX_IP_CKSUM)) {
1327 		pi_data = hn_rndis_pktinfo_append(pkt, NDIS_TXCSUM_INFO_SIZE,
1328 						  NDIS_PKTINFO_TYPE_CSUM);
1329 		*pi_data = 0;
1330 
1331 		if (m->ol_flags & PKT_TX_IPV6)
1332 			*pi_data |= NDIS_TXCSUM_INFO_IPV6;
1333 		if (m->ol_flags & PKT_TX_IPV4) {
1334 			*pi_data |= NDIS_TXCSUM_INFO_IPV4;
1335 
1336 			if (m->ol_flags & PKT_TX_IP_CKSUM)
1337 				*pi_data |= NDIS_TXCSUM_INFO_IPCS;
1338 		}
1339 
1340 		if (m->ol_flags & PKT_TX_TCP_CKSUM)
1341 			*pi_data |= NDIS_TXCSUM_INFO_MKTCPCS(hlen);
1342 		else if (m->ol_flags & PKT_TX_UDP_CKSUM)
1343 			*pi_data |= NDIS_TXCSUM_INFO_MKUDPCS(hlen);
1344 	}
1345 
1346 	pkt_hlen = pkt->pktinfooffset + pkt->pktinfolen;
1347 	/* Fixup RNDIS packet message total length */
1348 	pkt->len += pkt_hlen;
1349 
1350 	/* Convert RNDIS packet message offsets */
1351 	pkt->dataoffset = hn_rndis_pktmsg_offset(pkt_hlen);
1352 	pkt->pktinfooffset = hn_rndis_pktmsg_offset(pkt->pktinfooffset);
1353 }
1354 
1355 /* How many scatter gather list elements ar needed */
1356 static unsigned int hn_get_slots(const struct rte_mbuf *m)
1357 {
1358 	unsigned int slots = 1; /* for RNDIS header */
1359 
1360 	while (m) {
1361 		unsigned int size = rte_pktmbuf_data_len(m);
1362 		unsigned int offs = rte_mbuf_data_iova(m) & PAGE_MASK;
1363 
1364 		slots += (offs + size + PAGE_SIZE - 1) / PAGE_SIZE;
1365 		m = m->next;
1366 	}
1367 
1368 	return slots;
1369 }
1370 
1371 /* Build scatter gather list from chained mbuf */
1372 static unsigned int hn_fill_sg(struct vmbus_gpa *sg,
1373 			       const struct rte_mbuf *m)
1374 {
1375 	unsigned int segs = 0;
1376 
1377 	while (m) {
1378 		rte_iova_t addr = rte_mbuf_data_iova(m);
1379 		unsigned int page = addr / PAGE_SIZE;
1380 		unsigned int offset = addr & PAGE_MASK;
1381 		unsigned int len = rte_pktmbuf_data_len(m);
1382 
1383 		while (len > 0) {
1384 			unsigned int bytes = RTE_MIN(len, PAGE_SIZE - offset);
1385 
1386 			sg[segs].page = page;
1387 			sg[segs].ofs = offset;
1388 			sg[segs].len = bytes;
1389 			segs++;
1390 
1391 			++page;
1392 			offset = 0;
1393 			len -= bytes;
1394 		}
1395 		m = m->next;
1396 	}
1397 
1398 	return segs;
1399 }
1400 
1401 /* Transmit directly from mbuf */
1402 static int hn_xmit_sg(struct hn_tx_queue *txq,
1403 		      const struct hn_txdesc *txd, const struct rte_mbuf *m,
1404 		      bool *need_sig)
1405 {
1406 	struct vmbus_gpa sg[hn_get_slots(m)];
1407 	struct hn_nvs_rndis nvs_rndis = {
1408 		.type = NVS_TYPE_RNDIS,
1409 		.rndis_mtype = NVS_RNDIS_MTYPE_DATA,
1410 		.chim_sz = txd->chim_size,
1411 	};
1412 	rte_iova_t addr;
1413 	unsigned int segs;
1414 
1415 	/* attach aggregation data if present */
1416 	if (txd->chim_size > 0)
1417 		nvs_rndis.chim_idx = txd->chim_index;
1418 	else
1419 		nvs_rndis.chim_idx = NVS_CHIM_IDX_INVALID;
1420 
1421 	hn_rndis_dump(txd->rndis_pkt);
1422 
1423 	/* pass IOVA of rndis header in first segment */
1424 	addr = rte_malloc_virt2iova(txd->rndis_pkt);
1425 	if (unlikely(addr == RTE_BAD_IOVA)) {
1426 		PMD_DRV_LOG(ERR, "RNDIS transmit can not get iova");
1427 		return -EINVAL;
1428 	}
1429 
1430 	sg[0].page = addr / PAGE_SIZE;
1431 	sg[0].ofs = addr & PAGE_MASK;
1432 	sg[0].len = RNDIS_PACKET_MSG_OFFSET_ABS(hn_rndis_pktlen(txd->rndis_pkt));
1433 	segs = 1;
1434 
1435 	hn_update_packet_stats(&txq->stats, m);
1436 
1437 	segs += hn_fill_sg(sg + 1, m);
1438 
1439 	PMD_TX_LOG(DEBUG, "port %u:%u tx %u segs %u size %u",
1440 		   txq->port_id, txq->queue_id, txd->chim_index,
1441 		   segs, nvs_rndis.chim_sz);
1442 
1443 	return hn_nvs_send_sglist(txq->chan, sg, segs,
1444 				  &nvs_rndis, sizeof(nvs_rndis),
1445 				  (uintptr_t)txd, need_sig);
1446 }
1447 
1448 uint16_t
1449 hn_xmit_pkts(void *ptxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1450 {
1451 	struct hn_tx_queue *txq = ptxq;
1452 	uint16_t queue_id = txq->queue_id;
1453 	struct hn_data *hv = txq->hv;
1454 	struct rte_eth_dev *vf_dev;
1455 	bool need_sig = false;
1456 	uint16_t nb_tx, tx_thresh;
1457 	int ret;
1458 
1459 	if (unlikely(hv->closed))
1460 		return 0;
1461 
1462 	/*
1463 	 * Always check for events on the primary channel
1464 	 * because that is where hotplug notifications occur.
1465 	 */
1466 	tx_thresh = RTE_MAX(txq->free_thresh, nb_pkts);
1467 	if (txq->queue_id == 0 ||
1468 	    rte_mempool_avail_count(txq->txdesc_pool) < tx_thresh)
1469 		hn_process_events(hv, txq->queue_id, 0);
1470 
1471 	/* Transmit over VF if present and up */
1472 	rte_rwlock_read_lock(&hv->vf_lock);
1473 	vf_dev = hn_get_vf_dev(hv);
1474 	if (vf_dev && vf_dev->data->dev_started) {
1475 		void *sub_q = vf_dev->data->tx_queues[queue_id];
1476 
1477 		nb_tx = (*vf_dev->tx_pkt_burst)(sub_q, tx_pkts, nb_pkts);
1478 		rte_rwlock_read_unlock(&hv->vf_lock);
1479 		return nb_tx;
1480 	}
1481 	rte_rwlock_read_unlock(&hv->vf_lock);
1482 
1483 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1484 		struct rte_mbuf *m = tx_pkts[nb_tx];
1485 		uint32_t pkt_size = m->pkt_len + HN_RNDIS_PKT_LEN;
1486 		struct rndis_packet_msg *pkt;
1487 		struct hn_txdesc *txd;
1488 
1489 		txd = hn_txd_get(txq);
1490 		if (txd == NULL)
1491 			break;
1492 
1493 		/* For small packets aggregate them in chimney buffer */
1494 		if (m->pkt_len < HN_TXCOPY_THRESHOLD && pkt_size <= txq->agg_szmax) {
1495 			/* If this packet will not fit, then flush  */
1496 			if (txq->agg_pktleft == 0 ||
1497 			    RTE_ALIGN(pkt_size, txq->agg_align) > txq->agg_szleft) {
1498 				if (hn_flush_txagg(txq, &need_sig))
1499 					goto fail;
1500 			}
1501 
1502 
1503 			pkt = hn_try_txagg(hv, txq, txd, pkt_size);
1504 			if (unlikely(!pkt))
1505 				break;
1506 
1507 			hn_encap(pkt, queue_id, m);
1508 			hn_append_to_chim(txq, pkt, m);
1509 
1510 			rte_pktmbuf_free(m);
1511 
1512 			/* if buffer is full, flush */
1513 			if (txq->agg_pktleft == 0 &&
1514 			    hn_flush_txagg(txq, &need_sig))
1515 				goto fail;
1516 		} else {
1517 			/* Send any outstanding packets in buffer */
1518 			if (txq->agg_txd && hn_flush_txagg(txq, &need_sig))
1519 				goto fail;
1520 
1521 			pkt = txd->rndis_pkt;
1522 			txd->m = m;
1523 			txd->data_size = m->pkt_len;
1524 			++txd->packets;
1525 
1526 			hn_encap(pkt, queue_id, m);
1527 
1528 			ret = hn_xmit_sg(txq, txd, m, &need_sig);
1529 			if (unlikely(ret != 0)) {
1530 				if (ret == -EAGAIN) {
1531 					PMD_TX_LOG(DEBUG, "sg channel full");
1532 					++txq->stats.channel_full;
1533 				} else {
1534 					PMD_DRV_LOG(NOTICE, "sg send failed: %d", ret);
1535 					++txq->stats.errors;
1536 				}
1537 				hn_txd_put(txq, txd);
1538 				goto fail;
1539 			}
1540 		}
1541 	}
1542 
1543 	/* If partial buffer left, then try and send it.
1544 	 * if that fails, then reuse it on next send.
1545 	 */
1546 	hn_flush_txagg(txq, &need_sig);
1547 
1548 fail:
1549 	if (need_sig)
1550 		rte_vmbus_chan_signal_tx(txq->chan);
1551 
1552 	return nb_tx;
1553 }
1554 
1555 static uint16_t
1556 hn_recv_vf(uint16_t vf_port, const struct hn_rx_queue *rxq,
1557 	   struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1558 {
1559 	uint16_t i, n;
1560 
1561 	if (unlikely(nb_pkts == 0))
1562 		return 0;
1563 
1564 	n = rte_eth_rx_burst(vf_port, rxq->queue_id, rx_pkts, nb_pkts);
1565 
1566 	/* relabel the received mbufs */
1567 	for (i = 0; i < n; i++)
1568 		rx_pkts[i]->port = rxq->port_id;
1569 
1570 	return n;
1571 }
1572 
1573 uint16_t
1574 hn_recv_pkts(void *prxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1575 {
1576 	struct hn_rx_queue *rxq = prxq;
1577 	struct hn_data *hv = rxq->hv;
1578 	struct rte_eth_dev *vf_dev;
1579 	uint16_t nb_rcv;
1580 
1581 	if (unlikely(hv->closed))
1582 		return 0;
1583 
1584 	/* Check for new completions (and hotplug) */
1585 	if (likely(rte_ring_count(rxq->rx_ring) < nb_pkts))
1586 		hn_process_events(hv, rxq->queue_id, 0);
1587 
1588 	/* Always check the vmbus path for multicast and new flows */
1589 	nb_rcv = rte_ring_sc_dequeue_burst(rxq->rx_ring,
1590 					   (void **)rx_pkts, nb_pkts, NULL);
1591 
1592 	/* If VF is available, check that as well */
1593 	rte_rwlock_read_lock(&hv->vf_lock);
1594 	vf_dev = hn_get_vf_dev(hv);
1595 	if (vf_dev && vf_dev->data->dev_started)
1596 		nb_rcv += hn_recv_vf(vf_dev->data->port_id, rxq,
1597 				     rx_pkts + nb_rcv, nb_pkts - nb_rcv);
1598 
1599 	rte_rwlock_read_unlock(&hv->vf_lock);
1600 	return nb_rcv;
1601 }
1602 
1603 void
1604 hn_dev_free_queues(struct rte_eth_dev *dev)
1605 {
1606 	unsigned int i;
1607 
1608 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
1609 		struct hn_rx_queue *rxq = dev->data->rx_queues[i];
1610 
1611 		hn_rx_queue_free(rxq, false);
1612 		dev->data->rx_queues[i] = NULL;
1613 	}
1614 	dev->data->nb_rx_queues = 0;
1615 
1616 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
1617 		hn_dev_tx_queue_release(dev->data->tx_queues[i]);
1618 		dev->data->tx_queues[i] = NULL;
1619 	}
1620 	dev->data->nb_tx_queues = 0;
1621 }
1622