xref: /dpdk/drivers/net/ngbe/ngbe_rxtx.c (revision 62fc35e63d0e61d1db99386dde08c11ef313a85c)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2018-2021 Beijing WangXun Technology Co., Ltd.
3  * Copyright(c) 2010-2017 Intel Corporation
4  */
5 
6 #include <sys/queue.h>
7 
8 #include <stdint.h>
9 #include <rte_ethdev.h>
10 #include <ethdev_driver.h>
11 #include <rte_malloc.h>
12 
13 #include "ngbe_logs.h"
14 #include "base/ngbe.h"
15 #include "ngbe_ethdev.h"
16 #include "ngbe_rxtx.h"
17 
18 /*********************************************************************
19  *
20  *  Queue management functions
21  *
22  **********************************************************************/
23 
24 static void
25 ngbe_tx_queue_release_mbufs(struct ngbe_tx_queue *txq)
26 {
27 	unsigned int i;
28 
29 	if (txq->sw_ring != NULL) {
30 		for (i = 0; i < txq->nb_tx_desc; i++) {
31 			if (txq->sw_ring[i].mbuf != NULL) {
32 				rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
33 				txq->sw_ring[i].mbuf = NULL;
34 			}
35 		}
36 	}
37 }
38 
39 static void
40 ngbe_tx_free_swring(struct ngbe_tx_queue *txq)
41 {
42 	if (txq != NULL)
43 		rte_free(txq->sw_ring);
44 }
45 
46 static void
47 ngbe_tx_queue_release(struct ngbe_tx_queue *txq)
48 {
49 	if (txq != NULL) {
50 		if (txq->ops != NULL) {
51 			txq->ops->release_mbufs(txq);
52 			txq->ops->free_swring(txq);
53 		}
54 		rte_free(txq);
55 	}
56 }
57 
58 void
59 ngbe_dev_tx_queue_release(void *txq)
60 {
61 	ngbe_tx_queue_release(txq);
62 }
63 
64 /* (Re)set dynamic ngbe_tx_queue fields to defaults */
65 static void
66 ngbe_reset_tx_queue(struct ngbe_tx_queue *txq)
67 {
68 	static const struct ngbe_tx_desc zeroed_desc = {0};
69 	struct ngbe_tx_entry *txe = txq->sw_ring;
70 	uint16_t prev, i;
71 
72 	/* Zero out HW ring memory */
73 	for (i = 0; i < txq->nb_tx_desc; i++)
74 		txq->tx_ring[i] = zeroed_desc;
75 
76 	/* Initialize SW ring entries */
77 	prev = (uint16_t)(txq->nb_tx_desc - 1);
78 	for (i = 0; i < txq->nb_tx_desc; i++) {
79 		/* the ring can also be modified by hardware */
80 		volatile struct ngbe_tx_desc *txd = &txq->tx_ring[i];
81 
82 		txd->dw3 = rte_cpu_to_le_32(NGBE_TXD_DD);
83 		txe[i].mbuf = NULL;
84 		txe[i].last_id = i;
85 		txe[prev].next_id = i;
86 		prev = i;
87 	}
88 
89 	txq->tx_next_dd = (uint16_t)(txq->tx_free_thresh - 1);
90 	txq->tx_tail = 0;
91 
92 	/*
93 	 * Always allow 1 descriptor to be un-allocated to avoid
94 	 * a H/W race condition
95 	 */
96 	txq->last_desc_cleaned = (uint16_t)(txq->nb_tx_desc - 1);
97 	txq->nb_tx_free = (uint16_t)(txq->nb_tx_desc - 1);
98 	txq->ctx_curr = 0;
99 	memset((void *)&txq->ctx_cache, 0,
100 		NGBE_CTX_NUM * sizeof(struct ngbe_ctx_info));
101 }
102 
103 static const struct ngbe_txq_ops def_txq_ops = {
104 	.release_mbufs = ngbe_tx_queue_release_mbufs,
105 	.free_swring = ngbe_tx_free_swring,
106 	.reset = ngbe_reset_tx_queue,
107 };
108 
109 int
110 ngbe_dev_tx_queue_setup(struct rte_eth_dev *dev,
111 			 uint16_t queue_idx,
112 			 uint16_t nb_desc,
113 			 unsigned int socket_id,
114 			 const struct rte_eth_txconf *tx_conf)
115 {
116 	const struct rte_memzone *tz;
117 	struct ngbe_tx_queue *txq;
118 	struct ngbe_hw     *hw;
119 	uint16_t tx_free_thresh;
120 
121 	PMD_INIT_FUNC_TRACE();
122 	hw = ngbe_dev_hw(dev);
123 
124 	/*
125 	 * The Tx descriptor ring will be cleaned after txq->tx_free_thresh
126 	 * descriptors are used or if the number of descriptors required
127 	 * to transmit a packet is greater than the number of free Tx
128 	 * descriptors.
129 	 * One descriptor in the Tx ring is used as a sentinel to avoid a
130 	 * H/W race condition, hence the maximum threshold constraints.
131 	 * When set to zero use default values.
132 	 */
133 	tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
134 			tx_conf->tx_free_thresh : DEFAULT_TX_FREE_THRESH);
135 	if (tx_free_thresh >= (nb_desc - 3)) {
136 		PMD_INIT_LOG(ERR,
137 			     "tx_free_thresh must be less than the number of TX descriptors minus 3. (tx_free_thresh=%u port=%d queue=%d)",
138 			     (unsigned int)tx_free_thresh,
139 			     (int)dev->data->port_id, (int)queue_idx);
140 		return -(EINVAL);
141 	}
142 
143 	if (nb_desc % tx_free_thresh != 0) {
144 		PMD_INIT_LOG(ERR,
145 			     "tx_free_thresh must be a divisor of the number of Tx descriptors. (tx_free_thresh=%u port=%d queue=%d)",
146 			     (unsigned int)tx_free_thresh,
147 			     (int)dev->data->port_id, (int)queue_idx);
148 		return -(EINVAL);
149 	}
150 
151 	/* Free memory prior to re-allocation if needed... */
152 	if (dev->data->tx_queues[queue_idx] != NULL) {
153 		ngbe_tx_queue_release(dev->data->tx_queues[queue_idx]);
154 		dev->data->tx_queues[queue_idx] = NULL;
155 	}
156 
157 	/* First allocate the Tx queue data structure */
158 	txq = rte_zmalloc_socket("ethdev Tx queue",
159 				 sizeof(struct ngbe_tx_queue),
160 				 RTE_CACHE_LINE_SIZE, socket_id);
161 	if (txq == NULL)
162 		return -ENOMEM;
163 
164 	/*
165 	 * Allocate Tx ring hardware descriptors. A memzone large enough to
166 	 * handle the maximum ring size is allocated in order to allow for
167 	 * resizing in later calls to the queue setup function.
168 	 */
169 	tz = rte_eth_dma_zone_reserve(dev, "tx_ring", queue_idx,
170 			sizeof(struct ngbe_tx_desc) * NGBE_RING_DESC_MAX,
171 			NGBE_ALIGN, socket_id);
172 	if (tz == NULL) {
173 		ngbe_tx_queue_release(txq);
174 		return -ENOMEM;
175 	}
176 
177 	txq->nb_tx_desc = nb_desc;
178 	txq->tx_free_thresh = tx_free_thresh;
179 	txq->pthresh = tx_conf->tx_thresh.pthresh;
180 	txq->hthresh = tx_conf->tx_thresh.hthresh;
181 	txq->wthresh = tx_conf->tx_thresh.wthresh;
182 	txq->queue_id = queue_idx;
183 	txq->reg_idx = queue_idx;
184 	txq->port_id = dev->data->port_id;
185 	txq->ops = &def_txq_ops;
186 	txq->tx_deferred_start = tx_conf->tx_deferred_start;
187 
188 	txq->tdt_reg_addr = NGBE_REG_ADDR(hw, NGBE_TXWP(txq->reg_idx));
189 	txq->tdc_reg_addr = NGBE_REG_ADDR(hw, NGBE_TXCFG(txq->reg_idx));
190 
191 	txq->tx_ring_phys_addr = TMZ_PADDR(tz);
192 	txq->tx_ring = (struct ngbe_tx_desc *)TMZ_VADDR(tz);
193 
194 	/* Allocate software ring */
195 	txq->sw_ring = rte_zmalloc_socket("txq->sw_ring",
196 				sizeof(struct ngbe_tx_entry) * nb_desc,
197 				RTE_CACHE_LINE_SIZE, socket_id);
198 	if (txq->sw_ring == NULL) {
199 		ngbe_tx_queue_release(txq);
200 		return -ENOMEM;
201 	}
202 	PMD_INIT_LOG(DEBUG,
203 		     "sw_ring=%p hw_ring=%p dma_addr=0x%" PRIx64,
204 		     txq->sw_ring, txq->tx_ring, txq->tx_ring_phys_addr);
205 
206 	txq->ops->reset(txq);
207 
208 	dev->data->tx_queues[queue_idx] = txq;
209 
210 	return 0;
211 }
212 
213 /**
214  * ngbe_free_sc_cluster - free the not-yet-completed scattered cluster
215  *
216  * The "next" pointer of the last segment of (not-yet-completed) RSC clusters
217  * in the sw_sc_ring is not set to NULL but rather points to the next
218  * mbuf of this RSC aggregation (that has not been completed yet and still
219  * resides on the HW ring). So, instead of calling for rte_pktmbuf_free() we
220  * will just free first "nb_segs" segments of the cluster explicitly by calling
221  * an rte_pktmbuf_free_seg().
222  *
223  * @m scattered cluster head
224  */
225 static void
226 ngbe_free_sc_cluster(struct rte_mbuf *m)
227 {
228 	uint16_t i, nb_segs = m->nb_segs;
229 	struct rte_mbuf *next_seg;
230 
231 	for (i = 0; i < nb_segs; i++) {
232 		next_seg = m->next;
233 		rte_pktmbuf_free_seg(m);
234 		m = next_seg;
235 	}
236 }
237 
238 static void
239 ngbe_rx_queue_release_mbufs(struct ngbe_rx_queue *rxq)
240 {
241 	unsigned int i;
242 
243 	if (rxq->sw_ring != NULL) {
244 		for (i = 0; i < rxq->nb_rx_desc; i++) {
245 			if (rxq->sw_ring[i].mbuf != NULL) {
246 				rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
247 				rxq->sw_ring[i].mbuf = NULL;
248 			}
249 		}
250 		for (i = 0; i < rxq->rx_nb_avail; ++i) {
251 			struct rte_mbuf *mb;
252 
253 			mb = rxq->rx_stage[rxq->rx_next_avail + i];
254 			rte_pktmbuf_free_seg(mb);
255 		}
256 		rxq->rx_nb_avail = 0;
257 	}
258 
259 	if (rxq->sw_sc_ring != NULL)
260 		for (i = 0; i < rxq->nb_rx_desc; i++)
261 			if (rxq->sw_sc_ring[i].fbuf != NULL) {
262 				ngbe_free_sc_cluster(rxq->sw_sc_ring[i].fbuf);
263 				rxq->sw_sc_ring[i].fbuf = NULL;
264 			}
265 }
266 
267 static void
268 ngbe_rx_queue_release(struct ngbe_rx_queue *rxq)
269 {
270 	if (rxq != NULL) {
271 		ngbe_rx_queue_release_mbufs(rxq);
272 		rte_free(rxq->sw_ring);
273 		rte_free(rxq->sw_sc_ring);
274 		rte_free(rxq);
275 	}
276 }
277 
278 void
279 ngbe_dev_rx_queue_release(void *rxq)
280 {
281 	ngbe_rx_queue_release(rxq);
282 }
283 
284 /*
285  * Check if Rx Burst Bulk Alloc function can be used.
286  * Return
287  *        0: the preconditions are satisfied and the bulk allocation function
288  *           can be used.
289  *  -EINVAL: the preconditions are NOT satisfied and the default Rx burst
290  *           function must be used.
291  */
292 static inline int
293 check_rx_burst_bulk_alloc_preconditions(struct ngbe_rx_queue *rxq)
294 {
295 	int ret = 0;
296 
297 	/*
298 	 * Make sure the following pre-conditions are satisfied:
299 	 *   rxq->rx_free_thresh >= RTE_PMD_NGBE_RX_MAX_BURST
300 	 *   rxq->rx_free_thresh < rxq->nb_rx_desc
301 	 *   (rxq->nb_rx_desc % rxq->rx_free_thresh) == 0
302 	 * Scattered packets are not supported.  This should be checked
303 	 * outside of this function.
304 	 */
305 	if (rxq->rx_free_thresh < RTE_PMD_NGBE_RX_MAX_BURST) {
306 		PMD_INIT_LOG(DEBUG,
307 			     "Rx Burst Bulk Alloc Preconditions: rxq->rx_free_thresh=%d, RTE_PMD_NGBE_RX_MAX_BURST=%d",
308 			     rxq->rx_free_thresh, RTE_PMD_NGBE_RX_MAX_BURST);
309 		ret = -EINVAL;
310 	} else if (rxq->rx_free_thresh >= rxq->nb_rx_desc) {
311 		PMD_INIT_LOG(DEBUG,
312 			     "Rx Burst Bulk Alloc Preconditions: rxq->rx_free_thresh=%d, rxq->nb_rx_desc=%d",
313 			     rxq->rx_free_thresh, rxq->nb_rx_desc);
314 		ret = -EINVAL;
315 	} else if ((rxq->nb_rx_desc % rxq->rx_free_thresh) != 0) {
316 		PMD_INIT_LOG(DEBUG,
317 			     "Rx Burst Bulk Alloc Preconditions: rxq->nb_rx_desc=%d, rxq->rx_free_thresh=%d",
318 			     rxq->nb_rx_desc, rxq->rx_free_thresh);
319 		ret = -EINVAL;
320 	}
321 
322 	return ret;
323 }
324 
325 /* Reset dynamic ngbe_rx_queue fields back to defaults */
326 static void
327 ngbe_reset_rx_queue(struct ngbe_adapter *adapter, struct ngbe_rx_queue *rxq)
328 {
329 	static const struct ngbe_rx_desc zeroed_desc = {
330 						{{0}, {0} }, {{0}, {0} } };
331 	unsigned int i;
332 	uint16_t len = rxq->nb_rx_desc;
333 
334 	/*
335 	 * By default, the Rx queue setup function allocates enough memory for
336 	 * NGBE_RING_DESC_MAX.  The Rx Burst bulk allocation function requires
337 	 * extra memory at the end of the descriptor ring to be zero'd out.
338 	 */
339 	if (adapter->rx_bulk_alloc_allowed)
340 		/* zero out extra memory */
341 		len += RTE_PMD_NGBE_RX_MAX_BURST;
342 
343 	/*
344 	 * Zero out HW ring memory. Zero out extra memory at the end of
345 	 * the H/W ring so look-ahead logic in Rx Burst bulk alloc function
346 	 * reads extra memory as zeros.
347 	 */
348 	for (i = 0; i < len; i++)
349 		rxq->rx_ring[i] = zeroed_desc;
350 
351 	/*
352 	 * initialize extra software ring entries. Space for these extra
353 	 * entries is always allocated
354 	 */
355 	memset(&rxq->fake_mbuf, 0x0, sizeof(rxq->fake_mbuf));
356 	for (i = rxq->nb_rx_desc; i < len; ++i)
357 		rxq->sw_ring[i].mbuf = &rxq->fake_mbuf;
358 
359 	rxq->rx_nb_avail = 0;
360 	rxq->rx_next_avail = 0;
361 	rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
362 	rxq->rx_tail = 0;
363 	rxq->nb_rx_hold = 0;
364 	rxq->pkt_first_seg = NULL;
365 	rxq->pkt_last_seg = NULL;
366 }
367 
368 int
369 ngbe_dev_rx_queue_setup(struct rte_eth_dev *dev,
370 			 uint16_t queue_idx,
371 			 uint16_t nb_desc,
372 			 unsigned int socket_id,
373 			 const struct rte_eth_rxconf *rx_conf,
374 			 struct rte_mempool *mp)
375 {
376 	const struct rte_memzone *rz;
377 	struct ngbe_rx_queue *rxq;
378 	struct ngbe_hw     *hw;
379 	uint16_t len;
380 	struct ngbe_adapter *adapter = ngbe_dev_adapter(dev);
381 
382 	PMD_INIT_FUNC_TRACE();
383 	hw = ngbe_dev_hw(dev);
384 
385 	/* Free memory prior to re-allocation if needed... */
386 	if (dev->data->rx_queues[queue_idx] != NULL) {
387 		ngbe_rx_queue_release(dev->data->rx_queues[queue_idx]);
388 		dev->data->rx_queues[queue_idx] = NULL;
389 	}
390 
391 	/* First allocate the Rx queue data structure */
392 	rxq = rte_zmalloc_socket("ethdev RX queue",
393 				 sizeof(struct ngbe_rx_queue),
394 				 RTE_CACHE_LINE_SIZE, socket_id);
395 	if (rxq == NULL)
396 		return -ENOMEM;
397 	rxq->mb_pool = mp;
398 	rxq->nb_rx_desc = nb_desc;
399 	rxq->rx_free_thresh = rx_conf->rx_free_thresh;
400 	rxq->queue_id = queue_idx;
401 	rxq->reg_idx = queue_idx;
402 	rxq->port_id = dev->data->port_id;
403 	rxq->drop_en = rx_conf->rx_drop_en;
404 	rxq->rx_deferred_start = rx_conf->rx_deferred_start;
405 
406 	/*
407 	 * Allocate Rx ring hardware descriptors. A memzone large enough to
408 	 * handle the maximum ring size is allocated in order to allow for
409 	 * resizing in later calls to the queue setup function.
410 	 */
411 	rz = rte_eth_dma_zone_reserve(dev, "rx_ring", queue_idx,
412 				      RX_RING_SZ, NGBE_ALIGN, socket_id);
413 	if (rz == NULL) {
414 		ngbe_rx_queue_release(rxq);
415 		return -ENOMEM;
416 	}
417 
418 	/*
419 	 * Zero init all the descriptors in the ring.
420 	 */
421 	memset(rz->addr, 0, RX_RING_SZ);
422 
423 	rxq->rdt_reg_addr = NGBE_REG_ADDR(hw, NGBE_RXWP(rxq->reg_idx));
424 	rxq->rdh_reg_addr = NGBE_REG_ADDR(hw, NGBE_RXRP(rxq->reg_idx));
425 
426 	rxq->rx_ring_phys_addr = TMZ_PADDR(rz);
427 	rxq->rx_ring = (struct ngbe_rx_desc *)TMZ_VADDR(rz);
428 
429 	/*
430 	 * Certain constraints must be met in order to use the bulk buffer
431 	 * allocation Rx burst function. If any of Rx queues doesn't meet them
432 	 * the feature should be disabled for the whole port.
433 	 */
434 	if (check_rx_burst_bulk_alloc_preconditions(rxq)) {
435 		PMD_INIT_LOG(DEBUG,
436 			     "queue[%d] doesn't meet Rx Bulk Alloc preconditions - canceling the feature for the whole port[%d]",
437 			     rxq->queue_id, rxq->port_id);
438 		adapter->rx_bulk_alloc_allowed = false;
439 	}
440 
441 	/*
442 	 * Allocate software ring. Allow for space at the end of the
443 	 * S/W ring to make sure look-ahead logic in bulk alloc Rx burst
444 	 * function does not access an invalid memory region.
445 	 */
446 	len = nb_desc;
447 	if (adapter->rx_bulk_alloc_allowed)
448 		len += RTE_PMD_NGBE_RX_MAX_BURST;
449 
450 	rxq->sw_ring = rte_zmalloc_socket("rxq->sw_ring",
451 					  sizeof(struct ngbe_rx_entry) * len,
452 					  RTE_CACHE_LINE_SIZE, socket_id);
453 	if (rxq->sw_ring == NULL) {
454 		ngbe_rx_queue_release(rxq);
455 		return -ENOMEM;
456 	}
457 
458 	/*
459 	 * Always allocate even if it's not going to be needed in order to
460 	 * simplify the code.
461 	 *
462 	 * This ring is used in Scattered Rx cases and Scattered Rx may
463 	 * be requested in ngbe_dev_rx_init(), which is called later from
464 	 * dev_start() flow.
465 	 */
466 	rxq->sw_sc_ring =
467 		rte_zmalloc_socket("rxq->sw_sc_ring",
468 				  sizeof(struct ngbe_scattered_rx_entry) * len,
469 				  RTE_CACHE_LINE_SIZE, socket_id);
470 	if (rxq->sw_sc_ring == NULL) {
471 		ngbe_rx_queue_release(rxq);
472 		return -ENOMEM;
473 	}
474 
475 	PMD_INIT_LOG(DEBUG,
476 		     "sw_ring=%p sw_sc_ring=%p hw_ring=%p dma_addr=0x%" PRIx64,
477 		     rxq->sw_ring, rxq->sw_sc_ring, rxq->rx_ring,
478 		     rxq->rx_ring_phys_addr);
479 
480 	dev->data->rx_queues[queue_idx] = rxq;
481 
482 	ngbe_reset_rx_queue(adapter, rxq);
483 
484 	return 0;
485 }
486 
487 void
488 ngbe_dev_clear_queues(struct rte_eth_dev *dev)
489 {
490 	unsigned int i;
491 	struct ngbe_adapter *adapter = ngbe_dev_adapter(dev);
492 
493 	PMD_INIT_FUNC_TRACE();
494 
495 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
496 		struct ngbe_tx_queue *txq = dev->data->tx_queues[i];
497 
498 		if (txq != NULL) {
499 			txq->ops->release_mbufs(txq);
500 			txq->ops->reset(txq);
501 		}
502 	}
503 
504 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
505 		struct ngbe_rx_queue *rxq = dev->data->rx_queues[i];
506 
507 		if (rxq != NULL) {
508 			ngbe_rx_queue_release_mbufs(rxq);
509 			ngbe_reset_rx_queue(adapter, rxq);
510 		}
511 	}
512 }
513 
514 static int
515 ngbe_alloc_rx_queue_mbufs(struct ngbe_rx_queue *rxq)
516 {
517 	struct ngbe_rx_entry *rxe = rxq->sw_ring;
518 	uint64_t dma_addr;
519 	unsigned int i;
520 
521 	/* Initialize software ring entries */
522 	for (i = 0; i < rxq->nb_rx_desc; i++) {
523 		/* the ring can also be modified by hardware */
524 		volatile struct ngbe_rx_desc *rxd;
525 		struct rte_mbuf *mbuf = rte_mbuf_raw_alloc(rxq->mb_pool);
526 
527 		if (mbuf == NULL) {
528 			PMD_INIT_LOG(ERR, "Rx mbuf alloc failed queue_id=%u port_id=%u",
529 				     (unsigned int)rxq->queue_id,
530 				     (unsigned int)rxq->port_id);
531 			return -ENOMEM;
532 		}
533 
534 		mbuf->data_off = RTE_PKTMBUF_HEADROOM;
535 		mbuf->port = rxq->port_id;
536 
537 		dma_addr =
538 			rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
539 		rxd = &rxq->rx_ring[i];
540 		NGBE_RXD_HDRADDR(rxd, 0);
541 		NGBE_RXD_PKTADDR(rxd, dma_addr);
542 		rxe[i].mbuf = mbuf;
543 	}
544 
545 	return 0;
546 }
547 
548 /*
549  * Initializes Receive Unit.
550  */
551 int
552 ngbe_dev_rx_init(struct rte_eth_dev *dev)
553 {
554 	struct ngbe_hw *hw;
555 	struct ngbe_rx_queue *rxq;
556 	uint64_t bus_addr;
557 	uint32_t fctrl;
558 	uint32_t hlreg0;
559 	uint32_t srrctl;
560 	uint16_t buf_size;
561 	uint16_t i;
562 
563 	PMD_INIT_FUNC_TRACE();
564 	hw = ngbe_dev_hw(dev);
565 
566 	/*
567 	 * Make sure receives are disabled while setting
568 	 * up the Rx context (registers, descriptor rings, etc.).
569 	 */
570 	wr32m(hw, NGBE_MACRXCFG, NGBE_MACRXCFG_ENA, 0);
571 	wr32m(hw, NGBE_PBRXCTL, NGBE_PBRXCTL_ENA, 0);
572 
573 	/* Enable receipt of broadcasted frames */
574 	fctrl = rd32(hw, NGBE_PSRCTL);
575 	fctrl |= NGBE_PSRCTL_BCA;
576 	wr32(hw, NGBE_PSRCTL, fctrl);
577 
578 	hlreg0 = rd32(hw, NGBE_SECRXCTL);
579 	hlreg0 &= ~NGBE_SECRXCTL_XDSA;
580 	wr32(hw, NGBE_SECRXCTL, hlreg0);
581 
582 	wr32m(hw, NGBE_FRMSZ, NGBE_FRMSZ_MAX_MASK,
583 			NGBE_FRMSZ_MAX(NGBE_FRAME_SIZE_DFT));
584 
585 	/* Setup Rx queues */
586 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
587 		rxq = dev->data->rx_queues[i];
588 
589 		/* Setup the Base and Length of the Rx Descriptor Rings */
590 		bus_addr = rxq->rx_ring_phys_addr;
591 		wr32(hw, NGBE_RXBAL(rxq->reg_idx),
592 				(uint32_t)(bus_addr & BIT_MASK32));
593 		wr32(hw, NGBE_RXBAH(rxq->reg_idx),
594 				(uint32_t)(bus_addr >> 32));
595 		wr32(hw, NGBE_RXRP(rxq->reg_idx), 0);
596 		wr32(hw, NGBE_RXWP(rxq->reg_idx), 0);
597 
598 		srrctl = NGBE_RXCFG_RNGLEN(rxq->nb_rx_desc);
599 
600 		/* Set if packets are dropped when no descriptors available */
601 		if (rxq->drop_en)
602 			srrctl |= NGBE_RXCFG_DROP;
603 
604 		/*
605 		 * Configure the Rx buffer size in the PKTLEN field of
606 		 * the RXCFG register of the queue.
607 		 * The value is in 1 KB resolution. Valid values can be from
608 		 * 1 KB to 16 KB.
609 		 */
610 		buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mb_pool) -
611 			RTE_PKTMBUF_HEADROOM);
612 		buf_size = ROUND_DOWN(buf_size, 0x1 << 10);
613 		srrctl |= NGBE_RXCFG_PKTLEN(buf_size);
614 
615 		wr32(hw, NGBE_RXCFG(rxq->reg_idx), srrctl);
616 	}
617 
618 	return 0;
619 }
620 
621 /*
622  * Initializes Transmit Unit.
623  */
624 void
625 ngbe_dev_tx_init(struct rte_eth_dev *dev)
626 {
627 	struct ngbe_hw     *hw;
628 	struct ngbe_tx_queue *txq;
629 	uint64_t bus_addr;
630 	uint16_t i;
631 
632 	PMD_INIT_FUNC_TRACE();
633 	hw = ngbe_dev_hw(dev);
634 
635 	wr32m(hw, NGBE_SECTXCTL, NGBE_SECTXCTL_ODSA, NGBE_SECTXCTL_ODSA);
636 	wr32m(hw, NGBE_SECTXCTL, NGBE_SECTXCTL_XDSA, 0);
637 
638 	/* Setup the Base and Length of the Tx Descriptor Rings */
639 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
640 		txq = dev->data->tx_queues[i];
641 
642 		bus_addr = txq->tx_ring_phys_addr;
643 		wr32(hw, NGBE_TXBAL(txq->reg_idx),
644 				(uint32_t)(bus_addr & BIT_MASK32));
645 		wr32(hw, NGBE_TXBAH(txq->reg_idx),
646 				(uint32_t)(bus_addr >> 32));
647 		wr32m(hw, NGBE_TXCFG(txq->reg_idx), NGBE_TXCFG_BUFLEN_MASK,
648 			NGBE_TXCFG_BUFLEN(txq->nb_tx_desc));
649 		/* Setup the HW Tx Head and TX Tail descriptor pointers */
650 		wr32(hw, NGBE_TXRP(txq->reg_idx), 0);
651 		wr32(hw, NGBE_TXWP(txq->reg_idx), 0);
652 	}
653 }
654 
655 /*
656  * Start Transmit and Receive Units.
657  */
658 int
659 ngbe_dev_rxtx_start(struct rte_eth_dev *dev)
660 {
661 	struct ngbe_hw     *hw;
662 	struct ngbe_tx_queue *txq;
663 	struct ngbe_rx_queue *rxq;
664 	uint32_t dmatxctl;
665 	uint32_t rxctrl;
666 	uint16_t i;
667 	int ret = 0;
668 
669 	PMD_INIT_FUNC_TRACE();
670 	hw = ngbe_dev_hw(dev);
671 
672 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
673 		txq = dev->data->tx_queues[i];
674 		/* Setup Transmit Threshold Registers */
675 		wr32m(hw, NGBE_TXCFG(txq->reg_idx),
676 		      NGBE_TXCFG_HTHRESH_MASK |
677 		      NGBE_TXCFG_WTHRESH_MASK,
678 		      NGBE_TXCFG_HTHRESH(txq->hthresh) |
679 		      NGBE_TXCFG_WTHRESH(txq->wthresh));
680 	}
681 
682 	dmatxctl = rd32(hw, NGBE_DMATXCTRL);
683 	dmatxctl |= NGBE_DMATXCTRL_ENA;
684 	wr32(hw, NGBE_DMATXCTRL, dmatxctl);
685 
686 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
687 		txq = dev->data->tx_queues[i];
688 		if (txq->tx_deferred_start == 0) {
689 			ret = ngbe_dev_tx_queue_start(dev, i);
690 			if (ret < 0)
691 				return ret;
692 		}
693 	}
694 
695 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
696 		rxq = dev->data->rx_queues[i];
697 		if (rxq->rx_deferred_start == 0) {
698 			ret = ngbe_dev_rx_queue_start(dev, i);
699 			if (ret < 0)
700 				return ret;
701 		}
702 	}
703 
704 	/* Enable Receive engine */
705 	rxctrl = rd32(hw, NGBE_PBRXCTL);
706 	rxctrl |= NGBE_PBRXCTL_ENA;
707 	hw->mac.enable_rx_dma(hw, rxctrl);
708 
709 	return 0;
710 }
711 
712 void
713 ngbe_dev_save_rx_queue(struct ngbe_hw *hw, uint16_t rx_queue_id)
714 {
715 	u32 *reg = &hw->q_rx_regs[rx_queue_id * 8];
716 	*(reg++) = rd32(hw, NGBE_RXBAL(rx_queue_id));
717 	*(reg++) = rd32(hw, NGBE_RXBAH(rx_queue_id));
718 	*(reg++) = rd32(hw, NGBE_RXCFG(rx_queue_id));
719 }
720 
721 void
722 ngbe_dev_store_rx_queue(struct ngbe_hw *hw, uint16_t rx_queue_id)
723 {
724 	u32 *reg = &hw->q_rx_regs[rx_queue_id * 8];
725 	wr32(hw, NGBE_RXBAL(rx_queue_id), *(reg++));
726 	wr32(hw, NGBE_RXBAH(rx_queue_id), *(reg++));
727 	wr32(hw, NGBE_RXCFG(rx_queue_id), *(reg++) & ~NGBE_RXCFG_ENA);
728 }
729 
730 void
731 ngbe_dev_save_tx_queue(struct ngbe_hw *hw, uint16_t tx_queue_id)
732 {
733 	u32 *reg = &hw->q_tx_regs[tx_queue_id * 8];
734 	*(reg++) = rd32(hw, NGBE_TXBAL(tx_queue_id));
735 	*(reg++) = rd32(hw, NGBE_TXBAH(tx_queue_id));
736 	*(reg++) = rd32(hw, NGBE_TXCFG(tx_queue_id));
737 }
738 
739 void
740 ngbe_dev_store_tx_queue(struct ngbe_hw *hw, uint16_t tx_queue_id)
741 {
742 	u32 *reg = &hw->q_tx_regs[tx_queue_id * 8];
743 	wr32(hw, NGBE_TXBAL(tx_queue_id), *(reg++));
744 	wr32(hw, NGBE_TXBAH(tx_queue_id), *(reg++));
745 	wr32(hw, NGBE_TXCFG(tx_queue_id), *(reg++) & ~NGBE_TXCFG_ENA);
746 }
747 
748 /*
749  * Start Receive Units for specified queue.
750  */
751 int
752 ngbe_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
753 {
754 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
755 	struct ngbe_rx_queue *rxq;
756 	uint32_t rxdctl;
757 	int poll_ms;
758 
759 	PMD_INIT_FUNC_TRACE();
760 
761 	rxq = dev->data->rx_queues[rx_queue_id];
762 
763 	/* Allocate buffers for descriptor rings */
764 	if (ngbe_alloc_rx_queue_mbufs(rxq) != 0) {
765 		PMD_INIT_LOG(ERR, "Could not alloc mbuf for queue:%d",
766 			     rx_queue_id);
767 		return -1;
768 	}
769 	rxdctl = rd32(hw, NGBE_RXCFG(rxq->reg_idx));
770 	rxdctl |= NGBE_RXCFG_ENA;
771 	wr32(hw, NGBE_RXCFG(rxq->reg_idx), rxdctl);
772 
773 	/* Wait until Rx Enable ready */
774 	poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS;
775 	do {
776 		rte_delay_ms(1);
777 		rxdctl = rd32(hw, NGBE_RXCFG(rxq->reg_idx));
778 	} while (--poll_ms && !(rxdctl & NGBE_RXCFG_ENA));
779 	if (poll_ms == 0)
780 		PMD_INIT_LOG(ERR, "Could not enable Rx Queue %d", rx_queue_id);
781 	rte_wmb();
782 	wr32(hw, NGBE_RXRP(rxq->reg_idx), 0);
783 	wr32(hw, NGBE_RXWP(rxq->reg_idx), rxq->nb_rx_desc - 1);
784 	dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
785 
786 	return 0;
787 }
788 
789 /*
790  * Stop Receive Units for specified queue.
791  */
792 int
793 ngbe_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
794 {
795 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
796 	struct ngbe_adapter *adapter = ngbe_dev_adapter(dev);
797 	struct ngbe_rx_queue *rxq;
798 	uint32_t rxdctl;
799 	int poll_ms;
800 
801 	PMD_INIT_FUNC_TRACE();
802 
803 	rxq = dev->data->rx_queues[rx_queue_id];
804 
805 	ngbe_dev_save_rx_queue(hw, rxq->reg_idx);
806 	wr32m(hw, NGBE_RXCFG(rxq->reg_idx), NGBE_RXCFG_ENA, 0);
807 
808 	/* Wait until Rx Enable bit clear */
809 	poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS;
810 	do {
811 		rte_delay_ms(1);
812 		rxdctl = rd32(hw, NGBE_RXCFG(rxq->reg_idx));
813 	} while (--poll_ms && (rxdctl & NGBE_RXCFG_ENA));
814 	if (poll_ms == 0)
815 		PMD_INIT_LOG(ERR, "Could not disable Rx Queue %d", rx_queue_id);
816 
817 	rte_delay_us(RTE_NGBE_WAIT_100_US);
818 	ngbe_dev_store_rx_queue(hw, rxq->reg_idx);
819 
820 	ngbe_rx_queue_release_mbufs(rxq);
821 	ngbe_reset_rx_queue(adapter, rxq);
822 	dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
823 
824 	return 0;
825 }
826 
827 /*
828  * Start Transmit Units for specified queue.
829  */
830 int
831 ngbe_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
832 {
833 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
834 	struct ngbe_tx_queue *txq;
835 	uint32_t txdctl;
836 	int poll_ms;
837 
838 	PMD_INIT_FUNC_TRACE();
839 
840 	txq = dev->data->tx_queues[tx_queue_id];
841 	wr32m(hw, NGBE_TXCFG(txq->reg_idx), NGBE_TXCFG_ENA, NGBE_TXCFG_ENA);
842 
843 	/* Wait until Tx Enable ready */
844 	poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS;
845 	do {
846 		rte_delay_ms(1);
847 		txdctl = rd32(hw, NGBE_TXCFG(txq->reg_idx));
848 	} while (--poll_ms && !(txdctl & NGBE_TXCFG_ENA));
849 	if (poll_ms == 0)
850 		PMD_INIT_LOG(ERR, "Could not enable Tx Queue %d",
851 			     tx_queue_id);
852 
853 	rte_wmb();
854 	wr32(hw, NGBE_TXWP(txq->reg_idx), txq->tx_tail);
855 	dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
856 
857 	return 0;
858 }
859 
860 /*
861  * Stop Transmit Units for specified queue.
862  */
863 int
864 ngbe_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
865 {
866 	struct ngbe_hw *hw = ngbe_dev_hw(dev);
867 	struct ngbe_tx_queue *txq;
868 	uint32_t txdctl;
869 	uint32_t txtdh, txtdt;
870 	int poll_ms;
871 
872 	PMD_INIT_FUNC_TRACE();
873 
874 	txq = dev->data->tx_queues[tx_queue_id];
875 
876 	/* Wait until Tx queue is empty */
877 	poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS;
878 	do {
879 		rte_delay_us(RTE_NGBE_WAIT_100_US);
880 		txtdh = rd32(hw, NGBE_TXRP(txq->reg_idx));
881 		txtdt = rd32(hw, NGBE_TXWP(txq->reg_idx));
882 	} while (--poll_ms && (txtdh != txtdt));
883 	if (poll_ms == 0)
884 		PMD_INIT_LOG(ERR, "Tx Queue %d is not empty when stopping.",
885 			     tx_queue_id);
886 
887 	ngbe_dev_save_tx_queue(hw, txq->reg_idx);
888 	wr32m(hw, NGBE_TXCFG(txq->reg_idx), NGBE_TXCFG_ENA, 0);
889 
890 	/* Wait until Tx Enable bit clear */
891 	poll_ms = RTE_NGBE_REGISTER_POLL_WAIT_10_MS;
892 	do {
893 		rte_delay_ms(1);
894 		txdctl = rd32(hw, NGBE_TXCFG(txq->reg_idx));
895 	} while (--poll_ms && (txdctl & NGBE_TXCFG_ENA));
896 	if (poll_ms == 0)
897 		PMD_INIT_LOG(ERR, "Could not disable Tx Queue %d",
898 			     tx_queue_id);
899 
900 	rte_delay_us(RTE_NGBE_WAIT_100_US);
901 	ngbe_dev_store_tx_queue(hw, txq->reg_idx);
902 
903 	if (txq->ops != NULL) {
904 		txq->ops->release_mbufs(txq);
905 		txq->ops->reset(txq);
906 	}
907 	dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
908 
909 	return 0;
910 }
911