xref: /f-stack/dpdk/drivers/net/fm10k/fm10k_rxtx.c (revision 4418919f)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2013-2016 Intel Corporation
3  */
4 
5 #include <inttypes.h>
6 
7 #include <rte_ethdev_driver.h>
8 #include <rte_common.h>
9 #include <rte_net.h>
10 #include "fm10k.h"
11 #include "base/fm10k_type.h"
12 
13 #ifdef RTE_PMD_PACKET_PREFETCH
14 #define rte_packet_prefetch(p)  rte_prefetch1(p)
15 #else
16 #define rte_packet_prefetch(p)  do {} while (0)
17 #endif
18 
19 #ifdef RTE_LIBRTE_FM10K_DEBUG_RX
dump_rxd(union fm10k_rx_desc * rxd)20 static inline void dump_rxd(union fm10k_rx_desc *rxd)
21 {
22 	PMD_RX_LOG(DEBUG, "+----------------|----------------+");
23 	PMD_RX_LOG(DEBUG, "|     GLORT      | PKT HDR & TYPE |");
24 	PMD_RX_LOG(DEBUG, "|   0x%08x   |   0x%08x   |", rxd->d.glort,
25 			rxd->d.data);
26 	PMD_RX_LOG(DEBUG, "+----------------|----------------+");
27 	PMD_RX_LOG(DEBUG, "|   VLAN & LEN   |     STATUS     |");
28 	PMD_RX_LOG(DEBUG, "|   0x%08x   |   0x%08x   |", rxd->d.vlan_len,
29 			rxd->d.staterr);
30 	PMD_RX_LOG(DEBUG, "+----------------|----------------+");
31 	PMD_RX_LOG(DEBUG, "|    RESERVED    |    RSS_HASH    |");
32 	PMD_RX_LOG(DEBUG, "|   0x%08x   |   0x%08x   |", 0, rxd->d.rss);
33 	PMD_RX_LOG(DEBUG, "+----------------|----------------+");
34 	PMD_RX_LOG(DEBUG, "|            TIME TAG             |");
35 	PMD_RX_LOG(DEBUG, "|       0x%016"PRIx64"        |", rxd->q.timestamp);
36 	PMD_RX_LOG(DEBUG, "+----------------|----------------+");
37 }
38 #endif
39 
40 #define FM10K_TX_OFFLOAD_MASK (  \
41 		PKT_TX_VLAN_PKT |        \
42 		PKT_TX_IPV6 |            \
43 		PKT_TX_IPV4 |            \
44 		PKT_TX_IP_CKSUM |        \
45 		PKT_TX_L4_MASK |         \
46 		PKT_TX_TCP_SEG)
47 
48 #define FM10K_TX_OFFLOAD_NOTSUP_MASK \
49 		(PKT_TX_OFFLOAD_MASK ^ FM10K_TX_OFFLOAD_MASK)
50 
51 /* @note: When this function is changed, make corresponding change to
52  * fm10k_dev_supported_ptypes_get()
53  */
54 static inline void
rx_desc_to_ol_flags(struct rte_mbuf * m,const union fm10k_rx_desc * d)55 rx_desc_to_ol_flags(struct rte_mbuf *m, const union fm10k_rx_desc *d)
56 {
57 	static const uint32_t
58 		ptype_table[FM10K_RXD_PKTTYPE_MASK >> FM10K_RXD_PKTTYPE_SHIFT]
59 			__rte_cache_aligned = {
60 		[FM10K_PKTTYPE_OTHER] = RTE_PTYPE_L2_ETHER,
61 		[FM10K_PKTTYPE_IPV4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4,
62 		[FM10K_PKTTYPE_IPV4_EX] = RTE_PTYPE_L2_ETHER |
63 			RTE_PTYPE_L3_IPV4_EXT,
64 		[FM10K_PKTTYPE_IPV6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6,
65 		[FM10K_PKTTYPE_IPV6_EX] = RTE_PTYPE_L2_ETHER |
66 			RTE_PTYPE_L3_IPV6_EXT,
67 		[FM10K_PKTTYPE_IPV4 | FM10K_PKTTYPE_TCP] = RTE_PTYPE_L2_ETHER |
68 			RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_TCP,
69 		[FM10K_PKTTYPE_IPV6 | FM10K_PKTTYPE_TCP] = RTE_PTYPE_L2_ETHER |
70 			RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_TCP,
71 		[FM10K_PKTTYPE_IPV4 | FM10K_PKTTYPE_UDP] = RTE_PTYPE_L2_ETHER |
72 			RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_UDP,
73 		[FM10K_PKTTYPE_IPV6 | FM10K_PKTTYPE_UDP] = RTE_PTYPE_L2_ETHER |
74 			RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_UDP,
75 	};
76 
77 	m->packet_type = ptype_table[(d->w.pkt_info & FM10K_RXD_PKTTYPE_MASK)
78 						>> FM10K_RXD_PKTTYPE_SHIFT];
79 
80 	if (d->w.pkt_info & FM10K_RXD_RSSTYPE_MASK)
81 		m->ol_flags |= PKT_RX_RSS_HASH;
82 
83 	if (unlikely((d->d.staterr &
84 		(FM10K_RXD_STATUS_IPCS | FM10K_RXD_STATUS_IPE)) ==
85 		(FM10K_RXD_STATUS_IPCS | FM10K_RXD_STATUS_IPE)))
86 		m->ol_flags |= PKT_RX_IP_CKSUM_BAD;
87 	else
88 		m->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
89 
90 	if (unlikely((d->d.staterr &
91 		(FM10K_RXD_STATUS_L4CS | FM10K_RXD_STATUS_L4E)) ==
92 		(FM10K_RXD_STATUS_L4CS | FM10K_RXD_STATUS_L4E)))
93 		m->ol_flags |= PKT_RX_L4_CKSUM_BAD;
94 	else
95 		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
96 }
97 
98 uint16_t
fm10k_recv_pkts(void * rx_queue,struct rte_mbuf ** rx_pkts,uint16_t nb_pkts)99 fm10k_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
100 	uint16_t nb_pkts)
101 {
102 	struct rte_mbuf *mbuf;
103 	union fm10k_rx_desc desc;
104 	struct fm10k_rx_queue *q = rx_queue;
105 	uint16_t count = 0;
106 	int alloc = 0;
107 	uint16_t next_dd;
108 	int ret;
109 
110 	next_dd = q->next_dd;
111 
112 	nb_pkts = RTE_MIN(nb_pkts, q->alloc_thresh);
113 	for (count = 0; count < nb_pkts; ++count) {
114 		if (!(q->hw_ring[next_dd].d.staterr & FM10K_RXD_STATUS_DD))
115 			break;
116 		mbuf = q->sw_ring[next_dd];
117 		desc = q->hw_ring[next_dd];
118 #ifdef RTE_LIBRTE_FM10K_DEBUG_RX
119 		dump_rxd(&desc);
120 #endif
121 		rte_pktmbuf_pkt_len(mbuf) = desc.w.length;
122 		rte_pktmbuf_data_len(mbuf) = desc.w.length;
123 
124 		mbuf->ol_flags = 0;
125 #ifdef RTE_LIBRTE_FM10K_RX_OLFLAGS_ENABLE
126 		rx_desc_to_ol_flags(mbuf, &desc);
127 #endif
128 
129 		mbuf->hash.rss = desc.d.rss;
130 		/**
131 		 * Packets in fm10k device always carry at least one VLAN tag.
132 		 * For those packets coming in without VLAN tag,
133 		 * the port default VLAN tag will be used.
134 		 * So, always PKT_RX_VLAN flag is set and vlan_tci
135 		 * is valid for each RX packet's mbuf.
136 		 */
137 		mbuf->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
138 		mbuf->vlan_tci = desc.w.vlan;
139 		/**
140 		 * mbuf->vlan_tci_outer is an idle field in fm10k driver,
141 		 * so it can be selected to store sglort value.
142 		 */
143 		if (q->rx_ftag_en)
144 			mbuf->vlan_tci_outer = rte_le_to_cpu_16(desc.w.sglort);
145 
146 		rx_pkts[count] = mbuf;
147 		if (++next_dd == q->nb_desc) {
148 			next_dd = 0;
149 			alloc = 1;
150 		}
151 
152 		/* Prefetch next mbuf while processing current one. */
153 		rte_prefetch0(q->sw_ring[next_dd]);
154 
155 		/*
156 		 * When next RX descriptor is on a cache-line boundary,
157 		 * prefetch the next 4 RX descriptors and the next 8 pointers
158 		 * to mbufs.
159 		 */
160 		if ((next_dd & 0x3) == 0) {
161 			rte_prefetch0(&q->hw_ring[next_dd]);
162 			rte_prefetch0(&q->sw_ring[next_dd]);
163 		}
164 	}
165 
166 	q->next_dd = next_dd;
167 
168 	if ((q->next_dd > q->next_trigger) || (alloc == 1)) {
169 		ret = rte_mempool_get_bulk(q->mp,
170 					(void **)&q->sw_ring[q->next_alloc],
171 					q->alloc_thresh);
172 
173 		if (unlikely(ret != 0)) {
174 			uint16_t port = q->port_id;
175 			PMD_RX_LOG(ERR, "Failed to alloc mbuf");
176 			/*
177 			 * Need to restore next_dd if we cannot allocate new
178 			 * buffers to replenish the old ones.
179 			 */
180 			q->next_dd = (q->next_dd + q->nb_desc - count) %
181 								q->nb_desc;
182 			rte_eth_devices[port].data->rx_mbuf_alloc_failed++;
183 			return 0;
184 		}
185 
186 		for (; q->next_alloc <= q->next_trigger; ++q->next_alloc) {
187 			mbuf = q->sw_ring[q->next_alloc];
188 
189 			/* setup static mbuf fields */
190 			fm10k_pktmbuf_reset(mbuf, q->port_id);
191 
192 			/* write descriptor */
193 			desc.q.pkt_addr = MBUF_DMA_ADDR_DEFAULT(mbuf);
194 			desc.q.hdr_addr = MBUF_DMA_ADDR_DEFAULT(mbuf);
195 			q->hw_ring[q->next_alloc] = desc;
196 		}
197 		FM10K_PCI_REG_WRITE(q->tail_ptr, q->next_trigger);
198 		q->next_trigger += q->alloc_thresh;
199 		if (q->next_trigger >= q->nb_desc) {
200 			q->next_trigger = q->alloc_thresh - 1;
201 			q->next_alloc = 0;
202 		}
203 	}
204 
205 	return count;
206 }
207 
208 uint16_t
fm10k_recv_scattered_pkts(void * rx_queue,struct rte_mbuf ** rx_pkts,uint16_t nb_pkts)209 fm10k_recv_scattered_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
210 				uint16_t nb_pkts)
211 {
212 	struct rte_mbuf *mbuf;
213 	union fm10k_rx_desc desc;
214 	struct fm10k_rx_queue *q = rx_queue;
215 	uint16_t count = 0;
216 	uint16_t nb_rcv, nb_seg;
217 	int alloc = 0;
218 	uint16_t next_dd;
219 	struct rte_mbuf *first_seg = q->pkt_first_seg;
220 	struct rte_mbuf *last_seg = q->pkt_last_seg;
221 	int ret;
222 
223 	next_dd = q->next_dd;
224 	nb_rcv = 0;
225 
226 	nb_seg = RTE_MIN(nb_pkts, q->alloc_thresh);
227 	for (count = 0; count < nb_seg; count++) {
228 		if (!(q->hw_ring[next_dd].d.staterr & FM10K_RXD_STATUS_DD))
229 			break;
230 		mbuf = q->sw_ring[next_dd];
231 		desc = q->hw_ring[next_dd];
232 #ifdef RTE_LIBRTE_FM10K_DEBUG_RX
233 		dump_rxd(&desc);
234 #endif
235 
236 		if (++next_dd == q->nb_desc) {
237 			next_dd = 0;
238 			alloc = 1;
239 		}
240 
241 		/* Prefetch next mbuf while processing current one. */
242 		rte_prefetch0(q->sw_ring[next_dd]);
243 
244 		/*
245 		 * When next RX descriptor is on a cache-line boundary,
246 		 * prefetch the next 4 RX descriptors and the next 8 pointers
247 		 * to mbufs.
248 		 */
249 		if ((next_dd & 0x3) == 0) {
250 			rte_prefetch0(&q->hw_ring[next_dd]);
251 			rte_prefetch0(&q->sw_ring[next_dd]);
252 		}
253 
254 		/* Fill data length */
255 		rte_pktmbuf_data_len(mbuf) = desc.w.length;
256 
257 		/*
258 		 * If this is the first buffer of the received packet,
259 		 * set the pointer to the first mbuf of the packet and
260 		 * initialize its context.
261 		 * Otherwise, update the total length and the number of segments
262 		 * of the current scattered packet, and update the pointer to
263 		 * the last mbuf of the current packet.
264 		 */
265 		if (!first_seg) {
266 			first_seg = mbuf;
267 			first_seg->pkt_len = desc.w.length;
268 		} else {
269 			first_seg->pkt_len =
270 					(uint16_t)(first_seg->pkt_len +
271 					rte_pktmbuf_data_len(mbuf));
272 			first_seg->nb_segs++;
273 			last_seg->next = mbuf;
274 		}
275 
276 		/*
277 		 * If this is not the last buffer of the received packet,
278 		 * update the pointer to the last mbuf of the current scattered
279 		 * packet and continue to parse the RX ring.
280 		 */
281 		if (!(desc.d.staterr & FM10K_RXD_STATUS_EOP)) {
282 			last_seg = mbuf;
283 			continue;
284 		}
285 
286 		first_seg->ol_flags = 0;
287 #ifdef RTE_LIBRTE_FM10K_RX_OLFLAGS_ENABLE
288 		rx_desc_to_ol_flags(first_seg, &desc);
289 #endif
290 		first_seg->hash.rss = desc.d.rss;
291 		/**
292 		 * Packets in fm10k device always carry at least one VLAN tag.
293 		 * For those packets coming in without VLAN tag,
294 		 * the port default VLAN tag will be used.
295 		 * So, always PKT_RX_VLAN flag is set and vlan_tci
296 		 * is valid for each RX packet's mbuf.
297 		 */
298 		first_seg->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
299 		first_seg->vlan_tci = desc.w.vlan;
300 		/**
301 		 * mbuf->vlan_tci_outer is an idle field in fm10k driver,
302 		 * so it can be selected to store sglort value.
303 		 */
304 		if (q->rx_ftag_en)
305 			first_seg->vlan_tci_outer =
306 				rte_le_to_cpu_16(desc.w.sglort);
307 
308 		/* Prefetch data of first segment, if configured to do so. */
309 		rte_packet_prefetch((char *)first_seg->buf_addr +
310 			first_seg->data_off);
311 
312 		/*
313 		 * Store the mbuf address into the next entry of the array
314 		 * of returned packets.
315 		 */
316 		rx_pkts[nb_rcv++] = first_seg;
317 
318 		/*
319 		 * Setup receipt context for a new packet.
320 		 */
321 		first_seg = NULL;
322 	}
323 
324 	q->next_dd = next_dd;
325 
326 	if ((q->next_dd > q->next_trigger) || (alloc == 1)) {
327 		ret = rte_mempool_get_bulk(q->mp,
328 					(void **)&q->sw_ring[q->next_alloc],
329 					q->alloc_thresh);
330 
331 		if (unlikely(ret != 0)) {
332 			uint16_t port = q->port_id;
333 			PMD_RX_LOG(ERR, "Failed to alloc mbuf");
334 			/*
335 			 * Need to restore next_dd if we cannot allocate new
336 			 * buffers to replenish the old ones.
337 			 */
338 			q->next_dd = (q->next_dd + q->nb_desc - count) %
339 								q->nb_desc;
340 			rte_eth_devices[port].data->rx_mbuf_alloc_failed++;
341 			return 0;
342 		}
343 
344 		for (; q->next_alloc <= q->next_trigger; ++q->next_alloc) {
345 			mbuf = q->sw_ring[q->next_alloc];
346 
347 			/* setup static mbuf fields */
348 			fm10k_pktmbuf_reset(mbuf, q->port_id);
349 
350 			/* write descriptor */
351 			desc.q.pkt_addr = MBUF_DMA_ADDR_DEFAULT(mbuf);
352 			desc.q.hdr_addr = MBUF_DMA_ADDR_DEFAULT(mbuf);
353 			q->hw_ring[q->next_alloc] = desc;
354 		}
355 		FM10K_PCI_REG_WRITE(q->tail_ptr, q->next_trigger);
356 		q->next_trigger += q->alloc_thresh;
357 		if (q->next_trigger >= q->nb_desc) {
358 			q->next_trigger = q->alloc_thresh - 1;
359 			q->next_alloc = 0;
360 		}
361 	}
362 
363 	q->pkt_first_seg = first_seg;
364 	q->pkt_last_seg = last_seg;
365 
366 	return nb_rcv;
367 }
368 
369 uint32_t
fm10k_dev_rx_queue_count(struct rte_eth_dev * dev,uint16_t rx_queue_id)370 fm10k_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
371 {
372 #define FM10K_RXQ_SCAN_INTERVAL 4
373 	volatile union fm10k_rx_desc *rxdp;
374 	struct fm10k_rx_queue *rxq;
375 	uint16_t desc = 0;
376 
377 	rxq = dev->data->rx_queues[rx_queue_id];
378 	rxdp = &rxq->hw_ring[rxq->next_dd];
379 	while ((desc < rxq->nb_desc) &&
380 		rxdp->w.status & rte_cpu_to_le_16(FM10K_RXD_STATUS_DD)) {
381 		/**
382 		 * Check the DD bit of a rx descriptor of each group of 4 desc,
383 		 * to avoid checking too frequently and downgrading performance
384 		 * too much.
385 		 */
386 		desc += FM10K_RXQ_SCAN_INTERVAL;
387 		rxdp += FM10K_RXQ_SCAN_INTERVAL;
388 		if (rxq->next_dd + desc >= rxq->nb_desc)
389 			rxdp = &rxq->hw_ring[rxq->next_dd + desc -
390 				rxq->nb_desc];
391 	}
392 
393 	return desc;
394 }
395 
396 int
fm10k_dev_rx_descriptor_done(void * rx_queue,uint16_t offset)397 fm10k_dev_rx_descriptor_done(void *rx_queue, uint16_t offset)
398 {
399 	volatile union fm10k_rx_desc *rxdp;
400 	struct fm10k_rx_queue *rxq = rx_queue;
401 	uint16_t desc;
402 	int ret;
403 
404 	if (unlikely(offset >= rxq->nb_desc)) {
405 		PMD_DRV_LOG(ERR, "Invalid RX descriptor offset %u", offset);
406 		return 0;
407 	}
408 
409 	desc = rxq->next_dd + offset;
410 	if (desc >= rxq->nb_desc)
411 		desc -= rxq->nb_desc;
412 
413 	rxdp = &rxq->hw_ring[desc];
414 
415 	ret = !!(rxdp->w.status &
416 			rte_cpu_to_le_16(FM10K_RXD_STATUS_DD));
417 
418 	return ret;
419 }
420 
421 int
fm10k_dev_rx_descriptor_status(void * rx_queue,uint16_t offset)422 fm10k_dev_rx_descriptor_status(void *rx_queue, uint16_t offset)
423 {
424 	volatile union fm10k_rx_desc *rxdp;
425 	struct fm10k_rx_queue *rxq = rx_queue;
426 	uint16_t nb_hold, trigger_last;
427 	uint16_t desc;
428 	int ret;
429 
430 	if (unlikely(offset >= rxq->nb_desc)) {
431 		PMD_DRV_LOG(ERR, "Invalid RX descriptor offset %u", offset);
432 		return 0;
433 	}
434 
435 	if (rxq->next_trigger < rxq->alloc_thresh)
436 		trigger_last = rxq->next_trigger +
437 					rxq->nb_desc - rxq->alloc_thresh;
438 	else
439 		trigger_last = rxq->next_trigger - rxq->alloc_thresh;
440 
441 	if (rxq->next_dd < trigger_last)
442 		nb_hold = rxq->next_dd + rxq->nb_desc - trigger_last;
443 	else
444 		nb_hold = rxq->next_dd - trigger_last;
445 
446 	if (offset >= rxq->nb_desc - nb_hold)
447 		return RTE_ETH_RX_DESC_UNAVAIL;
448 
449 	desc = rxq->next_dd + offset;
450 	if (desc >= rxq->nb_desc)
451 		desc -= rxq->nb_desc;
452 
453 	rxdp = &rxq->hw_ring[desc];
454 
455 	ret = !!(rxdp->w.status &
456 			rte_cpu_to_le_16(FM10K_RXD_STATUS_DD));
457 
458 	return ret;
459 }
460 
461 int
fm10k_dev_tx_descriptor_status(void * tx_queue,uint16_t offset)462 fm10k_dev_tx_descriptor_status(void *tx_queue, uint16_t offset)
463 {
464 	volatile struct fm10k_tx_desc *txdp;
465 	struct fm10k_tx_queue *txq = tx_queue;
466 	uint16_t desc;
467 	uint16_t next_rs = txq->nb_desc;
468 	struct fifo rs_tracker = txq->rs_tracker;
469 	struct fifo *r = &rs_tracker;
470 
471 	if (unlikely(offset >= txq->nb_desc))
472 		return -EINVAL;
473 
474 	desc = txq->next_free + offset;
475 	/* go to next desc that has the RS bit */
476 	desc = (desc / txq->rs_thresh + 1) *
477 		txq->rs_thresh - 1;
478 
479 	if (desc >= txq->nb_desc) {
480 		desc -= txq->nb_desc;
481 		if (desc >= txq->nb_desc)
482 			desc -= txq->nb_desc;
483 	}
484 
485 	r->head = r->list;
486 	for ( ; r->head != r->endp; ) {
487 		if (*r->head >= desc && *r->head < next_rs)
488 			next_rs = *r->head;
489 		++r->head;
490 	}
491 
492 	txdp = &txq->hw_ring[next_rs];
493 	if (txdp->flags & FM10K_TXD_FLAG_DONE)
494 		return RTE_ETH_TX_DESC_DONE;
495 
496 	return RTE_ETH_TX_DESC_FULL;
497 }
498 
499 /*
500  * Free multiple TX mbuf at a time if they are in the same pool
501  *
502  * @txep: software desc ring index that starts to free
503  * @num: number of descs to free
504  *
505  */
tx_free_bulk_mbuf(struct rte_mbuf ** txep,int num)506 static inline void tx_free_bulk_mbuf(struct rte_mbuf **txep, int num)
507 {
508 	struct rte_mbuf *m, *free[RTE_FM10K_TX_MAX_FREE_BUF_SZ];
509 	int i;
510 	int nb_free = 0;
511 
512 	if (unlikely(num == 0))
513 		return;
514 
515 	m = rte_pktmbuf_prefree_seg(txep[0]);
516 	if (likely(m != NULL)) {
517 		free[0] = m;
518 		nb_free = 1;
519 		for (i = 1; i < num; i++) {
520 			m = rte_pktmbuf_prefree_seg(txep[i]);
521 			if (likely(m != NULL)) {
522 				if (likely(m->pool == free[0]->pool))
523 					free[nb_free++] = m;
524 				else {
525 					rte_mempool_put_bulk(free[0]->pool,
526 							(void *)free, nb_free);
527 					free[0] = m;
528 					nb_free = 1;
529 				}
530 			}
531 			txep[i] = NULL;
532 		}
533 		rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
534 	} else {
535 		for (i = 1; i < num; i++) {
536 			m = rte_pktmbuf_prefree_seg(txep[i]);
537 			if (m != NULL)
538 				rte_mempool_put(m->pool, m);
539 			txep[i] = NULL;
540 		}
541 	}
542 }
543 
tx_free_descriptors(struct fm10k_tx_queue * q)544 static inline void tx_free_descriptors(struct fm10k_tx_queue *q)
545 {
546 	uint16_t next_rs, count = 0;
547 
548 	next_rs = fifo_peek(&q->rs_tracker);
549 	if (!(q->hw_ring[next_rs].flags & FM10K_TXD_FLAG_DONE))
550 		return;
551 
552 	/* the DONE flag is set on this descriptor so remove the ID
553 	 * from the RS bit tracker and free the buffers */
554 	fifo_remove(&q->rs_tracker);
555 
556 	/* wrap around? if so, free buffers from last_free up to but NOT
557 	 * including nb_desc */
558 	if (q->last_free > next_rs) {
559 		count = q->nb_desc - q->last_free;
560 		tx_free_bulk_mbuf(&q->sw_ring[q->last_free], count);
561 		q->last_free = 0;
562 	}
563 
564 	/* adjust free descriptor count before the next loop */
565 	q->nb_free += count + (next_rs + 1 - q->last_free);
566 
567 	/* free buffers from last_free, up to and including next_rs */
568 	if (q->last_free <= next_rs) {
569 		count = next_rs - q->last_free + 1;
570 		tx_free_bulk_mbuf(&q->sw_ring[q->last_free], count);
571 		q->last_free += count;
572 	}
573 
574 	if (q->last_free == q->nb_desc)
575 		q->last_free = 0;
576 }
577 
tx_xmit_pkt(struct fm10k_tx_queue * q,struct rte_mbuf * mb)578 static inline void tx_xmit_pkt(struct fm10k_tx_queue *q, struct rte_mbuf *mb)
579 {
580 	uint16_t last_id;
581 	uint8_t flags, hdrlen;
582 
583 	/* always set the LAST flag on the last descriptor used to
584 	 * transmit the packet */
585 	flags = FM10K_TXD_FLAG_LAST;
586 	last_id = q->next_free + mb->nb_segs - 1;
587 	if (last_id >= q->nb_desc)
588 		last_id = last_id - q->nb_desc;
589 
590 	/* but only set the RS flag on the last descriptor if rs_thresh
591 	 * descriptors will be used since the RS flag was last set */
592 	if ((q->nb_used + mb->nb_segs) >= q->rs_thresh) {
593 		flags |= FM10K_TXD_FLAG_RS;
594 		fifo_insert(&q->rs_tracker, last_id);
595 		q->nb_used = 0;
596 	} else {
597 		q->nb_used = q->nb_used + mb->nb_segs;
598 	}
599 
600 	q->nb_free -= mb->nb_segs;
601 
602 	q->hw_ring[q->next_free].flags = 0;
603 	if (q->tx_ftag_en)
604 		q->hw_ring[q->next_free].flags |= FM10K_TXD_FLAG_FTAG;
605 	/* set checksum flags on first descriptor of packet. SCTP checksum
606 	 * offload is not supported, but we do not explicitly check for this
607 	 * case in favor of greatly simplified processing. */
608 	if (mb->ol_flags & (PKT_TX_IP_CKSUM | PKT_TX_L4_MASK | PKT_TX_TCP_SEG))
609 		q->hw_ring[q->next_free].flags |= FM10K_TXD_FLAG_CSUM;
610 
611 	/* set vlan if requested */
612 	if (mb->ol_flags & PKT_TX_VLAN_PKT)
613 		q->hw_ring[q->next_free].vlan = mb->vlan_tci;
614 	else
615 		q->hw_ring[q->next_free].vlan = 0;
616 
617 	q->sw_ring[q->next_free] = mb;
618 	q->hw_ring[q->next_free].buffer_addr =
619 			rte_cpu_to_le_64(MBUF_DMA_ADDR(mb));
620 	q->hw_ring[q->next_free].buflen =
621 			rte_cpu_to_le_16(rte_pktmbuf_data_len(mb));
622 
623 	if (mb->ol_flags & PKT_TX_TCP_SEG) {
624 		hdrlen = mb->l2_len + mb->l3_len + mb->l4_len;
625 		hdrlen += (mb->ol_flags & PKT_TX_TUNNEL_MASK) ?
626 			  mb->outer_l2_len + mb->outer_l3_len : 0;
627 		if (q->hw_ring[q->next_free].flags & FM10K_TXD_FLAG_FTAG)
628 			hdrlen += sizeof(struct fm10k_ftag);
629 
630 		if (likely((hdrlen >= FM10K_TSO_MIN_HEADERLEN) &&
631 				(hdrlen <= FM10K_TSO_MAX_HEADERLEN) &&
632 				(mb->tso_segsz >= FM10K_TSO_MINMSS))) {
633 			q->hw_ring[q->next_free].mss = mb->tso_segsz;
634 			q->hw_ring[q->next_free].hdrlen = hdrlen;
635 		}
636 	}
637 
638 	if (++q->next_free == q->nb_desc)
639 		q->next_free = 0;
640 
641 	/* fill up the rings */
642 	for (mb = mb->next; mb != NULL; mb = mb->next) {
643 		q->sw_ring[q->next_free] = mb;
644 		q->hw_ring[q->next_free].buffer_addr =
645 				rte_cpu_to_le_64(MBUF_DMA_ADDR(mb));
646 		q->hw_ring[q->next_free].buflen =
647 				rte_cpu_to_le_16(rte_pktmbuf_data_len(mb));
648 		q->hw_ring[q->next_free].flags = 0;
649 		if (++q->next_free == q->nb_desc)
650 			q->next_free = 0;
651 	}
652 
653 	q->hw_ring[last_id].flags |= flags;
654 }
655 
656 uint16_t
fm10k_xmit_pkts(void * tx_queue,struct rte_mbuf ** tx_pkts,uint16_t nb_pkts)657 fm10k_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
658 	uint16_t nb_pkts)
659 {
660 	struct fm10k_tx_queue *q = tx_queue;
661 	struct rte_mbuf *mb;
662 	uint16_t count;
663 
664 	for (count = 0; count < nb_pkts; ++count) {
665 		mb = tx_pkts[count];
666 
667 		/* running low on descriptors? try to free some... */
668 		if (q->nb_free < q->free_thresh)
669 			tx_free_descriptors(q);
670 
671 		/* make sure there are enough free descriptors to transmit the
672 		 * entire packet before doing anything */
673 		if (q->nb_free < mb->nb_segs)
674 			break;
675 
676 		/* sanity check to make sure the mbuf is valid */
677 		if ((mb->nb_segs == 0) ||
678 		    ((mb->nb_segs > 1) && (mb->next == NULL)))
679 			break;
680 
681 		/* process the packet */
682 		tx_xmit_pkt(q, mb);
683 	}
684 
685 	/* update the tail pointer if any packets were processed */
686 	if (likely(count > 0))
687 		FM10K_PCI_REG_WRITE(q->tail_ptr, q->next_free);
688 
689 	return count;
690 }
691 
692 uint16_t
fm10k_prep_pkts(__rte_unused void * tx_queue,struct rte_mbuf ** tx_pkts,uint16_t nb_pkts)693 fm10k_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
694 		uint16_t nb_pkts)
695 {
696 	int i, ret;
697 	struct rte_mbuf *m;
698 
699 	for (i = 0; i < nb_pkts; i++) {
700 		m = tx_pkts[i];
701 
702 		if ((m->ol_flags & PKT_TX_TCP_SEG) &&
703 				(m->tso_segsz < FM10K_TSO_MINMSS)) {
704 			rte_errno = EINVAL;
705 			return i;
706 		}
707 
708 		if (m->ol_flags & FM10K_TX_OFFLOAD_NOTSUP_MASK) {
709 			rte_errno = ENOTSUP;
710 			return i;
711 		}
712 
713 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
714 		ret = rte_validate_tx_offload(m);
715 		if (ret != 0) {
716 			rte_errno = -ret;
717 			return i;
718 		}
719 #endif
720 		ret = rte_net_intel_cksum_prepare(m);
721 		if (ret != 0) {
722 			rte_errno = -ret;
723 			return i;
724 		}
725 	}
726 
727 	return i;
728 }
729