xref: /dpdk/drivers/net/sfc/sfc_ef10_essb_rx.c (revision 089e5ed727a15da2729cfee9b63533dd120bd04c)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  *
3  * Copyright (c) 2017-2018 Solarflare Communications Inc.
4  * All rights reserved.
5  *
6  * This software was jointly developed between OKTET Labs (under contract
7  * for Solarflare) and Solarflare Communications, Inc.
8  */
9 
10 /* EF10 equal stride packed stream receive native datapath implementation */
11 
12 #include <stdbool.h>
13 
14 #include <rte_byteorder.h>
15 #include <rte_mbuf_ptype.h>
16 #include <rte_mbuf.h>
17 #include <rte_io.h>
18 
19 #include "efx.h"
20 #include "efx_types.h"
21 #include "efx_regs.h"
22 #include "efx_regs_ef10.h"
23 
24 #include "sfc_tweak.h"
25 #include "sfc_dp_rx.h"
26 #include "sfc_kvargs.h"
27 #include "sfc_ef10.h"
28 
29 /* Tunnels are not supported */
30 #define SFC_EF10_RX_EV_ENCAP_SUPPORT	0
31 #include "sfc_ef10_rx_ev.h"
32 
33 #define sfc_ef10_essb_rx_err(dpq, ...) \
34 	SFC_DP_LOG(SFC_KVARG_DATAPATH_EF10_ESSB, ERR, dpq, __VA_ARGS__)
35 
36 #define sfc_ef10_essb_rx_info(dpq, ...) \
37 	SFC_DP_LOG(SFC_KVARG_DATAPATH_EF10_ESSB, INFO, dpq, __VA_ARGS__)
38 
39 /*
40  * Fake length for RXQ descriptors in equal stride super-buffer mode
41  * to make hardware happy.
42  */
43 #define SFC_EF10_ESSB_RX_FAKE_BUF_SIZE	32
44 
45 /**
46  * Minimum number of Rx buffers the datapath allows to use.
47  *
48  * Each HW Rx descriptor has many Rx buffers. The number of buffers
49  * in one HW Rx descriptor is equal to size of contiguous block
50  * provided by Rx buffers memory pool. The contiguous block size
51  * depends on CONFIG_RTE_DRIVER_MEMPOOL_BUCKET_SIZE_KB and rte_mbuf
52  * data size specified on the memory pool creation. Typical rte_mbuf
53  * data size is about 2k which makes a bit less than 32 buffers in
54  * contiguous block with default bucket size equal to 64k.
55  * Since HW Rx descriptors are pushed by 8 (see SFC_EF10_RX_WPTR_ALIGN),
56  * it makes about 256 as required minimum. Double it in advertised
57  * minimum to allow for at least 2 refill blocks.
58  */
59 #define SFC_EF10_ESSB_RX_DESCS_MIN	512
60 
61 /**
62  * Number of Rx buffers should be aligned to.
63  *
64  * There are no extra requirements on alignment since actual number of
65  * pushed Rx buffers will be multiple by contiguous block size which
66  * is unknown beforehand.
67  */
68 #define SFC_EF10_ESSB_RX_DESCS_ALIGN	1
69 
70 /**
71  * Maximum number of descriptors/buffers in the Rx ring.
72  * It should guarantee that corresponding event queue never overfill.
73  */
74 #define SFC_EF10_ESSB_RXQ_LIMIT(_nevs) \
75 	((_nevs) - 1 /* head must not step on tail */ - \
76 	 (SFC_EF10_EV_PER_CACHE_LINE - 1) /* max unused EvQ entries */ - \
77 	 1 /* Rx error */ - 1 /* flush */)
78 
79 struct sfc_ef10_essb_rx_sw_desc {
80 	struct rte_mbuf			*first_mbuf;
81 };
82 
83 struct sfc_ef10_essb_rxq {
84 	/* Used on data path */
85 	unsigned int			flags;
86 #define SFC_EF10_ESSB_RXQ_STARTED	0x1
87 #define SFC_EF10_ESSB_RXQ_NOT_RUNNING	0x2
88 #define SFC_EF10_ESSB_RXQ_EXCEPTION	0x4
89 	unsigned int			rxq_ptr_mask;
90 	unsigned int			block_size;
91 	unsigned int			buf_stride;
92 	unsigned int			bufs_ptr;
93 	unsigned int			completed;
94 	unsigned int			pending_id;
95 	unsigned int			bufs_pending;
96 	unsigned int			left_in_completed;
97 	unsigned int			left_in_pending;
98 	unsigned int			evq_read_ptr;
99 	unsigned int			evq_ptr_mask;
100 	efx_qword_t			*evq_hw_ring;
101 	struct sfc_ef10_essb_rx_sw_desc	*sw_ring;
102 	uint16_t			port_id;
103 
104 	/* Used on refill */
105 	unsigned int			added;
106 	unsigned int			max_fill_level;
107 	unsigned int			refill_threshold;
108 	struct rte_mempool		*refill_mb_pool;
109 	efx_qword_t			*rxq_hw_ring;
110 	volatile void			*doorbell;
111 
112 	/* Datapath receive queue anchor */
113 	struct sfc_dp_rxq		dp;
114 };
115 
116 static inline struct sfc_ef10_essb_rxq *
117 sfc_ef10_essb_rxq_by_dp_rxq(struct sfc_dp_rxq *dp_rxq)
118 {
119 	return container_of(dp_rxq, struct sfc_ef10_essb_rxq, dp);
120 }
121 
122 static struct rte_mbuf *
123 sfc_ef10_essb_next_mbuf(const struct sfc_ef10_essb_rxq *rxq,
124 			struct rte_mbuf *mbuf)
125 {
126 	struct rte_mbuf *m;
127 
128 	m = (struct rte_mbuf *)((uintptr_t)mbuf + rxq->buf_stride);
129 	MBUF_RAW_ALLOC_CHECK(m);
130 	return m;
131 }
132 
133 static struct rte_mbuf *
134 sfc_ef10_essb_mbuf_by_index(const struct sfc_ef10_essb_rxq *rxq,
135 			    struct rte_mbuf *mbuf, unsigned int idx)
136 {
137 	struct rte_mbuf *m;
138 
139 	m = (struct rte_mbuf *)((uintptr_t)mbuf + idx * rxq->buf_stride);
140 	MBUF_RAW_ALLOC_CHECK(m);
141 	return m;
142 }
143 
144 static struct rte_mbuf *
145 sfc_ef10_essb_maybe_next_completed(struct sfc_ef10_essb_rxq *rxq)
146 {
147 	const struct sfc_ef10_essb_rx_sw_desc *rxd;
148 
149 	if (rxq->left_in_completed != 0) {
150 		rxd = &rxq->sw_ring[rxq->completed & rxq->rxq_ptr_mask];
151 		return sfc_ef10_essb_mbuf_by_index(rxq, rxd->first_mbuf,
152 				rxq->block_size - rxq->left_in_completed);
153 	} else {
154 		rxq->completed++;
155 		rxd = &rxq->sw_ring[rxq->completed & rxq->rxq_ptr_mask];
156 		rxq->left_in_completed = rxq->block_size;
157 		return rxd->first_mbuf;
158 	}
159 }
160 
161 static void
162 sfc_ef10_essb_rx_qrefill(struct sfc_ef10_essb_rxq *rxq)
163 {
164 	const unsigned int rxq_ptr_mask = rxq->rxq_ptr_mask;
165 	unsigned int free_space;
166 	unsigned int bulks;
167 	void *mbuf_blocks[SFC_EF10_RX_WPTR_ALIGN];
168 	unsigned int added = rxq->added;
169 
170 	free_space = rxq->max_fill_level - (added - rxq->completed);
171 
172 	if (free_space < rxq->refill_threshold)
173 		return;
174 
175 	bulks = free_space / RTE_DIM(mbuf_blocks);
176 	/* refill_threshold guarantees that bulks is positive */
177 	SFC_ASSERT(bulks > 0);
178 
179 	do {
180 		unsigned int id;
181 		unsigned int i;
182 
183 		if (unlikely(rte_mempool_get_contig_blocks(rxq->refill_mb_pool,
184 				mbuf_blocks, RTE_DIM(mbuf_blocks)) < 0)) {
185 			struct rte_eth_dev_data *dev_data =
186 				rte_eth_devices[rxq->port_id].data;
187 
188 			/*
189 			 * It is hardly a safe way to increment counter
190 			 * from different contexts, but all PMDs do it.
191 			 */
192 			dev_data->rx_mbuf_alloc_failed += RTE_DIM(mbuf_blocks);
193 			/* Return if we have posted nothing yet */
194 			if (added == rxq->added)
195 				return;
196 			/* Push posted */
197 			break;
198 		}
199 
200 		for (i = 0, id = added & rxq_ptr_mask;
201 		     i < RTE_DIM(mbuf_blocks);
202 		     ++i, ++id) {
203 			struct rte_mbuf *m = mbuf_blocks[i];
204 			struct sfc_ef10_essb_rx_sw_desc *rxd;
205 
206 			SFC_ASSERT((id & ~rxq_ptr_mask) == 0);
207 			rxd = &rxq->sw_ring[id];
208 			rxd->first_mbuf = m;
209 
210 			/* RX_KER_BYTE_CNT is ignored by firmware */
211 			EFX_POPULATE_QWORD_2(rxq->rxq_hw_ring[id],
212 					     ESF_DZ_RX_KER_BYTE_CNT,
213 					     SFC_EF10_ESSB_RX_FAKE_BUF_SIZE,
214 					     ESF_DZ_RX_KER_BUF_ADDR,
215 					     rte_mbuf_data_iova_default(m));
216 		}
217 
218 		added += RTE_DIM(mbuf_blocks);
219 
220 	} while (--bulks > 0);
221 
222 	SFC_ASSERT(rxq->added != added);
223 	rxq->added = added;
224 	sfc_ef10_rx_qpush(rxq->doorbell, added, rxq_ptr_mask);
225 }
226 
227 static bool
228 sfc_ef10_essb_rx_event_get(struct sfc_ef10_essb_rxq *rxq, efx_qword_t *rx_ev)
229 {
230 	*rx_ev = rxq->evq_hw_ring[rxq->evq_read_ptr & rxq->evq_ptr_mask];
231 
232 	if (!sfc_ef10_ev_present(*rx_ev))
233 		return false;
234 
235 	if (unlikely(EFX_QWORD_FIELD(*rx_ev, FSF_AZ_EV_CODE) !=
236 		     FSE_AZ_EV_CODE_RX_EV)) {
237 		/*
238 		 * Do not move read_ptr to keep the event for exception
239 		 * handling
240 		 */
241 		rxq->flags |= SFC_EF10_ESSB_RXQ_EXCEPTION;
242 		sfc_ef10_essb_rx_err(&rxq->dp.dpq,
243 				     "RxQ exception at EvQ read ptr %#x",
244 				     rxq->evq_read_ptr);
245 		return false;
246 	}
247 
248 	rxq->evq_read_ptr++;
249 	return true;
250 }
251 
252 static void
253 sfc_ef10_essb_rx_process_ev(struct sfc_ef10_essb_rxq *rxq, efx_qword_t rx_ev)
254 {
255 	unsigned int ready;
256 
257 	ready = (EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_DSC_PTR_LBITS) -
258 		 rxq->bufs_ptr) &
259 		EFX_MASK32(ESF_DZ_RX_DSC_PTR_LBITS);
260 
261 	rxq->bufs_ptr += ready;
262 	rxq->bufs_pending += ready;
263 
264 	SFC_ASSERT(ready > 0);
265 	do {
266 		const struct sfc_ef10_essb_rx_sw_desc *rxd;
267 		struct rte_mbuf *m;
268 		unsigned int todo_bufs;
269 		struct rte_mbuf *m0;
270 
271 		rxd = &rxq->sw_ring[rxq->pending_id];
272 		m = sfc_ef10_essb_mbuf_by_index(rxq, rxd->first_mbuf,
273 			rxq->block_size - rxq->left_in_pending);
274 
275 		if (ready < rxq->left_in_pending) {
276 			todo_bufs = ready;
277 			ready = 0;
278 			rxq->left_in_pending -= todo_bufs;
279 		} else {
280 			todo_bufs = rxq->left_in_pending;
281 			ready -= todo_bufs;
282 			rxq->left_in_pending = rxq->block_size;
283 			if (rxq->pending_id != rxq->rxq_ptr_mask)
284 				rxq->pending_id++;
285 			else
286 				rxq->pending_id = 0;
287 		}
288 
289 		SFC_ASSERT(todo_bufs > 0);
290 		--todo_bufs;
291 
292 		sfc_ef10_rx_ev_to_offloads(rx_ev, m, ~0ull);
293 
294 		/* Prefetch pseudo-header */
295 		rte_prefetch0((uint8_t *)m->buf_addr + RTE_PKTMBUF_HEADROOM);
296 
297 		m0 = m;
298 		while (todo_bufs-- > 0) {
299 			m = sfc_ef10_essb_next_mbuf(rxq, m);
300 			m->ol_flags = m0->ol_flags;
301 			m->packet_type = m0->packet_type;
302 			/* Prefetch pseudo-header */
303 			rte_prefetch0((uint8_t *)m->buf_addr +
304 				      RTE_PKTMBUF_HEADROOM);
305 		}
306 	} while (ready > 0);
307 }
308 
309 static unsigned int
310 sfc_ef10_essb_rx_get_pending(struct sfc_ef10_essb_rxq *rxq,
311 			     struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
312 {
313 	unsigned int n_rx_pkts = 0;
314 	unsigned int todo_bufs;
315 	struct rte_mbuf *m;
316 
317 	while ((todo_bufs = RTE_MIN(nb_pkts - n_rx_pkts,
318 				    rxq->bufs_pending)) > 0) {
319 		m = sfc_ef10_essb_maybe_next_completed(rxq);
320 
321 		todo_bufs = RTE_MIN(todo_bufs, rxq->left_in_completed);
322 
323 		rxq->bufs_pending -= todo_bufs;
324 		rxq->left_in_completed -= todo_bufs;
325 
326 		SFC_ASSERT(todo_bufs > 0);
327 		todo_bufs--;
328 
329 		do {
330 			const efx_qword_t *qwordp;
331 			uint16_t pkt_len;
332 
333 			/* Buffers to be discarded have 0 in packet type */
334 			if (unlikely(m->packet_type == 0)) {
335 				rte_mbuf_raw_free(m);
336 				goto next_buf;
337 			}
338 
339 			rx_pkts[n_rx_pkts++] = m;
340 
341 			/* Parse pseudo-header */
342 			qwordp = (const efx_qword_t *)
343 				((uint8_t *)m->buf_addr + RTE_PKTMBUF_HEADROOM);
344 			pkt_len =
345 				EFX_QWORD_FIELD(*qwordp,
346 						ES_EZ_ESSB_RX_PREFIX_DATA_LEN);
347 
348 			m->data_off = RTE_PKTMBUF_HEADROOM +
349 				ES_EZ_ESSB_RX_PREFIX_LEN;
350 			m->port = rxq->port_id;
351 
352 			rte_pktmbuf_pkt_len(m) = pkt_len;
353 			rte_pktmbuf_data_len(m) = pkt_len;
354 
355 			m->ol_flags |=
356 				(PKT_RX_RSS_HASH *
357 				 !!EFX_TEST_QWORD_BIT(*qwordp,
358 					ES_EZ_ESSB_RX_PREFIX_HASH_VALID_LBN)) |
359 				(PKT_RX_FDIR_ID *
360 				 !!EFX_TEST_QWORD_BIT(*qwordp,
361 					ES_EZ_ESSB_RX_PREFIX_MARK_VALID_LBN)) |
362 				(PKT_RX_FDIR *
363 				 !!EFX_TEST_QWORD_BIT(*qwordp,
364 					ES_EZ_ESSB_RX_PREFIX_MATCH_FLAG_LBN));
365 
366 			/* EFX_QWORD_FIELD converts little-endian to CPU */
367 			m->hash.rss =
368 				EFX_QWORD_FIELD(*qwordp,
369 						ES_EZ_ESSB_RX_PREFIX_HASH);
370 			m->hash.fdir.hi =
371 				EFX_QWORD_FIELD(*qwordp,
372 						ES_EZ_ESSB_RX_PREFIX_MARK);
373 
374 next_buf:
375 			m = sfc_ef10_essb_next_mbuf(rxq, m);
376 		} while (todo_bufs-- > 0);
377 	}
378 
379 	return n_rx_pkts;
380 }
381 
382 
383 static uint16_t
384 sfc_ef10_essb_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
385 			uint16_t nb_pkts)
386 {
387 	struct sfc_ef10_essb_rxq *rxq = sfc_ef10_essb_rxq_by_dp_rxq(rx_queue);
388 	const unsigned int evq_old_read_ptr = rxq->evq_read_ptr;
389 	uint16_t n_rx_pkts;
390 	efx_qword_t rx_ev;
391 
392 	if (unlikely(rxq->flags & (SFC_EF10_ESSB_RXQ_NOT_RUNNING |
393 				   SFC_EF10_ESSB_RXQ_EXCEPTION)))
394 		return 0;
395 
396 	n_rx_pkts = sfc_ef10_essb_rx_get_pending(rxq, rx_pkts, nb_pkts);
397 
398 	while (n_rx_pkts != nb_pkts &&
399 	       sfc_ef10_essb_rx_event_get(rxq, &rx_ev)) {
400 		/*
401 		 * DROP_EVENT is an internal to the NIC, software should
402 		 * never see it and, therefore, may ignore it.
403 		 */
404 
405 		sfc_ef10_essb_rx_process_ev(rxq, rx_ev);
406 		n_rx_pkts += sfc_ef10_essb_rx_get_pending(rxq,
407 							  rx_pkts + n_rx_pkts,
408 							  nb_pkts - n_rx_pkts);
409 	}
410 
411 	sfc_ef10_ev_qclear(rxq->evq_hw_ring, rxq->evq_ptr_mask,
412 			   evq_old_read_ptr, rxq->evq_read_ptr);
413 
414 	/* It is not a problem if we refill in the case of exception */
415 	sfc_ef10_essb_rx_qrefill(rxq);
416 
417 	return n_rx_pkts;
418 }
419 
420 static sfc_dp_rx_qdesc_npending_t sfc_ef10_essb_rx_qdesc_npending;
421 static unsigned int
422 sfc_ef10_essb_rx_qdesc_npending(struct sfc_dp_rxq *dp_rxq)
423 {
424 	struct sfc_ef10_essb_rxq *rxq = sfc_ef10_essb_rxq_by_dp_rxq(dp_rxq);
425 	const unsigned int evq_old_read_ptr = rxq->evq_read_ptr;
426 	efx_qword_t rx_ev;
427 
428 	if (unlikely(rxq->flags & (SFC_EF10_ESSB_RXQ_NOT_RUNNING |
429 				   SFC_EF10_ESSB_RXQ_EXCEPTION)))
430 		return rxq->bufs_pending;
431 
432 	while (sfc_ef10_essb_rx_event_get(rxq, &rx_ev)) {
433 		/*
434 		 * DROP_EVENT is an internal to the NIC, software should
435 		 * never see it and, therefore, may ignore it.
436 		 */
437 		sfc_ef10_essb_rx_process_ev(rxq, rx_ev);
438 	}
439 
440 	sfc_ef10_ev_qclear(rxq->evq_hw_ring, rxq->evq_ptr_mask,
441 			   evq_old_read_ptr, rxq->evq_read_ptr);
442 
443 	return rxq->bufs_pending;
444 }
445 
446 static sfc_dp_rx_qdesc_status_t sfc_ef10_essb_rx_qdesc_status;
447 static int
448 sfc_ef10_essb_rx_qdesc_status(struct sfc_dp_rxq *dp_rxq, uint16_t offset)
449 {
450 	struct sfc_ef10_essb_rxq *rxq = sfc_ef10_essb_rxq_by_dp_rxq(dp_rxq);
451 	unsigned int pending = sfc_ef10_essb_rx_qdesc_npending(dp_rxq);
452 
453 	if (offset < pending)
454 		return RTE_ETH_RX_DESC_DONE;
455 
456 	if (offset < (rxq->added - rxq->completed) * rxq->block_size +
457 		     rxq->left_in_completed - rxq->block_size)
458 		return RTE_ETH_RX_DESC_AVAIL;
459 
460 	return RTE_ETH_RX_DESC_UNAVAIL;
461 }
462 
463 static sfc_dp_rx_get_dev_info_t sfc_ef10_essb_rx_get_dev_info;
464 static void
465 sfc_ef10_essb_rx_get_dev_info(struct rte_eth_dev_info *dev_info)
466 {
467 	/*
468 	 * Number of descriptors just defines maximum number of pushed
469 	 * descriptors (fill level).
470 	 */
471 	dev_info->rx_desc_lim.nb_min = SFC_EF10_ESSB_RX_DESCS_MIN;
472 	dev_info->rx_desc_lim.nb_align = SFC_EF10_ESSB_RX_DESCS_ALIGN;
473 }
474 
475 static sfc_dp_rx_pool_ops_supported_t sfc_ef10_essb_rx_pool_ops_supported;
476 static int
477 sfc_ef10_essb_rx_pool_ops_supported(const char *pool)
478 {
479 	SFC_ASSERT(pool != NULL);
480 
481 	if (strcmp(pool, "bucket") == 0)
482 		return 0;
483 
484 	return -ENOTSUP;
485 }
486 
487 static sfc_dp_rx_qsize_up_rings_t sfc_ef10_essb_rx_qsize_up_rings;
488 static int
489 sfc_ef10_essb_rx_qsize_up_rings(uint16_t nb_rx_desc,
490 				struct sfc_dp_rx_hw_limits *limits,
491 				struct rte_mempool *mb_pool,
492 				unsigned int *rxq_entries,
493 				unsigned int *evq_entries,
494 				unsigned int *rxq_max_fill_level)
495 {
496 	int rc;
497 	struct rte_mempool_info mp_info;
498 	unsigned int nb_hw_rx_desc;
499 	unsigned int max_events;
500 
501 	rc = rte_mempool_ops_get_info(mb_pool, &mp_info);
502 	if (rc != 0)
503 		return -rc;
504 	if (mp_info.contig_block_size == 0)
505 		return EINVAL;
506 
507 	/*
508 	 * Calculate required number of hardware Rx descriptors each
509 	 * carrying contig block size Rx buffers.
510 	 * It cannot be less than Rx write pointer alignment plus 1
511 	 * in order to avoid cases when the ring is guaranteed to be
512 	 * empty.
513 	 */
514 	nb_hw_rx_desc = RTE_MAX(SFC_DIV_ROUND_UP(nb_rx_desc,
515 						 mp_info.contig_block_size),
516 				SFC_EF10_RX_WPTR_ALIGN + 1);
517 	if (nb_hw_rx_desc <= limits->rxq_min_entries) {
518 		*rxq_entries = limits->rxq_min_entries;
519 	} else {
520 		*rxq_entries = rte_align32pow2(nb_hw_rx_desc);
521 		if (*rxq_entries > limits->rxq_max_entries)
522 			return EINVAL;
523 	}
524 
525 	max_events = RTE_ALIGN_FLOOR(nb_hw_rx_desc, SFC_EF10_RX_WPTR_ALIGN) *
526 		mp_info.contig_block_size +
527 		(SFC_EF10_EV_PER_CACHE_LINE - 1) /* max unused EvQ entries */ +
528 		1 /* Rx error */ + 1 /* flush */ + 1 /* head-tail space */;
529 
530 	*evq_entries = rte_align32pow2(max_events);
531 	*evq_entries = RTE_MAX(*evq_entries, limits->evq_min_entries);
532 	*evq_entries = RTE_MIN(*evq_entries, limits->evq_max_entries);
533 
534 	/*
535 	 * May be even maximum event queue size is insufficient to handle
536 	 * so many Rx descriptors. If so, we should limit Rx queue fill level.
537 	 */
538 	*rxq_max_fill_level = RTE_MIN(nb_rx_desc,
539 				      SFC_EF10_ESSB_RXQ_LIMIT(*evq_entries));
540 	return 0;
541 }
542 
543 static sfc_dp_rx_qcreate_t sfc_ef10_essb_rx_qcreate;
544 static int
545 sfc_ef10_essb_rx_qcreate(uint16_t port_id, uint16_t queue_id,
546 			 const struct rte_pci_addr *pci_addr, int socket_id,
547 			 const struct sfc_dp_rx_qcreate_info *info,
548 			 struct sfc_dp_rxq **dp_rxqp)
549 {
550 	struct rte_mempool * const mp = info->refill_mb_pool;
551 	struct rte_mempool_info mp_info;
552 	struct sfc_ef10_essb_rxq *rxq;
553 	int rc;
554 
555 	rc = rte_mempool_ops_get_info(mp, &mp_info);
556 	if (rc != 0) {
557 		/* Positive errno is used in the driver */
558 		rc = -rc;
559 		goto fail_get_contig_block_size;
560 	}
561 
562 	/* Check if the mempool provides block dequeue */
563 	rc = EINVAL;
564 	if (mp_info.contig_block_size == 0)
565 		goto fail_no_block_dequeue;
566 
567 	rc = ENOMEM;
568 	rxq = rte_zmalloc_socket("sfc-ef10-rxq", sizeof(*rxq),
569 				 RTE_CACHE_LINE_SIZE, socket_id);
570 	if (rxq == NULL)
571 		goto fail_rxq_alloc;
572 
573 	sfc_dp_queue_init(&rxq->dp.dpq, port_id, queue_id, pci_addr);
574 
575 	rc = ENOMEM;
576 	rxq->sw_ring = rte_calloc_socket("sfc-ef10-rxq-sw_ring",
577 					 info->rxq_entries,
578 					 sizeof(*rxq->sw_ring),
579 					 RTE_CACHE_LINE_SIZE, socket_id);
580 	if (rxq->sw_ring == NULL)
581 		goto fail_desc_alloc;
582 
583 	rxq->block_size = mp_info.contig_block_size;
584 	rxq->buf_stride = mp->header_size + mp->elt_size + mp->trailer_size;
585 	rxq->rxq_ptr_mask = info->rxq_entries - 1;
586 	rxq->evq_ptr_mask = info->evq_entries - 1;
587 	rxq->evq_hw_ring = info->evq_hw_ring;
588 	rxq->port_id = port_id;
589 
590 	rxq->max_fill_level = info->max_fill_level / mp_info.contig_block_size;
591 	rxq->refill_threshold =
592 		RTE_MAX(info->refill_threshold / mp_info.contig_block_size,
593 			SFC_EF10_RX_WPTR_ALIGN);
594 	rxq->refill_mb_pool = mp;
595 	rxq->rxq_hw_ring = info->rxq_hw_ring;
596 
597 	rxq->doorbell = (volatile uint8_t *)info->mem_bar +
598 			ER_DZ_RX_DESC_UPD_REG_OFST +
599 			(info->hw_index << info->vi_window_shift);
600 
601 	sfc_ef10_essb_rx_info(&rxq->dp.dpq,
602 			      "block size is %u, buf stride is %u",
603 			      rxq->block_size, rxq->buf_stride);
604 	sfc_ef10_essb_rx_info(&rxq->dp.dpq,
605 			      "max fill level is %u descs (%u bufs), "
606 			      "refill threashold %u descs (%u bufs)",
607 			      rxq->max_fill_level,
608 			      rxq->max_fill_level * rxq->block_size,
609 			      rxq->refill_threshold,
610 			      rxq->refill_threshold * rxq->block_size);
611 
612 	*dp_rxqp = &rxq->dp;
613 	return 0;
614 
615 fail_desc_alloc:
616 	rte_free(rxq);
617 
618 fail_rxq_alloc:
619 fail_no_block_dequeue:
620 fail_get_contig_block_size:
621 	return rc;
622 }
623 
624 static sfc_dp_rx_qdestroy_t sfc_ef10_essb_rx_qdestroy;
625 static void
626 sfc_ef10_essb_rx_qdestroy(struct sfc_dp_rxq *dp_rxq)
627 {
628 	struct sfc_ef10_essb_rxq *rxq = sfc_ef10_essb_rxq_by_dp_rxq(dp_rxq);
629 
630 	rte_free(rxq->sw_ring);
631 	rte_free(rxq);
632 }
633 
634 static sfc_dp_rx_qstart_t sfc_ef10_essb_rx_qstart;
635 static int
636 sfc_ef10_essb_rx_qstart(struct sfc_dp_rxq *dp_rxq, unsigned int evq_read_ptr)
637 {
638 	struct sfc_ef10_essb_rxq *rxq = sfc_ef10_essb_rxq_by_dp_rxq(dp_rxq);
639 
640 	rxq->evq_read_ptr = evq_read_ptr;
641 
642 	/* Initialize before refill */
643 	rxq->completed = rxq->pending_id = rxq->added = 0;
644 	rxq->left_in_completed = rxq->left_in_pending = rxq->block_size;
645 	rxq->bufs_ptr = UINT_MAX;
646 	rxq->bufs_pending = 0;
647 
648 	sfc_ef10_essb_rx_qrefill(rxq);
649 
650 	rxq->flags |= SFC_EF10_ESSB_RXQ_STARTED;
651 	rxq->flags &=
652 		~(SFC_EF10_ESSB_RXQ_NOT_RUNNING | SFC_EF10_ESSB_RXQ_EXCEPTION);
653 
654 	return 0;
655 }
656 
657 static sfc_dp_rx_qstop_t sfc_ef10_essb_rx_qstop;
658 static void
659 sfc_ef10_essb_rx_qstop(struct sfc_dp_rxq *dp_rxq, unsigned int *evq_read_ptr)
660 {
661 	struct sfc_ef10_essb_rxq *rxq = sfc_ef10_essb_rxq_by_dp_rxq(dp_rxq);
662 
663 	rxq->flags |= SFC_EF10_ESSB_RXQ_NOT_RUNNING;
664 
665 	*evq_read_ptr = rxq->evq_read_ptr;
666 }
667 
668 static sfc_dp_rx_qrx_ev_t sfc_ef10_essb_rx_qrx_ev;
669 static bool
670 sfc_ef10_essb_rx_qrx_ev(struct sfc_dp_rxq *dp_rxq, __rte_unused unsigned int id)
671 {
672 	__rte_unused struct sfc_ef10_essb_rxq *rxq;
673 
674 	rxq = sfc_ef10_essb_rxq_by_dp_rxq(dp_rxq);
675 	SFC_ASSERT(rxq->flags & SFC_EF10_ESSB_RXQ_NOT_RUNNING);
676 
677 	/*
678 	 * It is safe to ignore Rx event since we free all mbufs on
679 	 * queue purge anyway.
680 	 */
681 
682 	return false;
683 }
684 
685 static sfc_dp_rx_qpurge_t sfc_ef10_essb_rx_qpurge;
686 static void
687 sfc_ef10_essb_rx_qpurge(struct sfc_dp_rxq *dp_rxq)
688 {
689 	struct sfc_ef10_essb_rxq *rxq = sfc_ef10_essb_rxq_by_dp_rxq(dp_rxq);
690 	unsigned int i;
691 	const struct sfc_ef10_essb_rx_sw_desc *rxd;
692 	struct rte_mbuf *m;
693 
694 	for (i = rxq->completed; i != rxq->added; ++i) {
695 		rxd = &rxq->sw_ring[i & rxq->rxq_ptr_mask];
696 		m = sfc_ef10_essb_mbuf_by_index(rxq, rxd->first_mbuf,
697 				rxq->block_size - rxq->left_in_completed);
698 		while (rxq->left_in_completed > 0) {
699 			rte_mbuf_raw_free(m);
700 			m = sfc_ef10_essb_next_mbuf(rxq, m);
701 			rxq->left_in_completed--;
702 		}
703 		rxq->left_in_completed = rxq->block_size;
704 	}
705 
706 	rxq->flags &= ~SFC_EF10_ESSB_RXQ_STARTED;
707 }
708 
709 struct sfc_dp_rx sfc_ef10_essb_rx = {
710 	.dp = {
711 		.name		= SFC_KVARG_DATAPATH_EF10_ESSB,
712 		.type		= SFC_DP_RX,
713 		.hw_fw_caps	= SFC_DP_HW_FW_CAP_EF10 |
714 				  SFC_DP_HW_FW_CAP_RX_ES_SUPER_BUFFER,
715 	},
716 	.features		= SFC_DP_RX_FEAT_FLOW_FLAG |
717 				  SFC_DP_RX_FEAT_FLOW_MARK,
718 	.dev_offload_capa	= DEV_RX_OFFLOAD_CHECKSUM,
719 	.queue_offload_capa	= 0,
720 	.get_dev_info		= sfc_ef10_essb_rx_get_dev_info,
721 	.pool_ops_supported	= sfc_ef10_essb_rx_pool_ops_supported,
722 	.qsize_up_rings		= sfc_ef10_essb_rx_qsize_up_rings,
723 	.qcreate		= sfc_ef10_essb_rx_qcreate,
724 	.qdestroy		= sfc_ef10_essb_rx_qdestroy,
725 	.qstart			= sfc_ef10_essb_rx_qstart,
726 	.qstop			= sfc_ef10_essb_rx_qstop,
727 	.qrx_ev			= sfc_ef10_essb_rx_qrx_ev,
728 	.qpurge			= sfc_ef10_essb_rx_qpurge,
729 	.supported_ptypes_get	= sfc_ef10_supported_ptypes_get,
730 	.qdesc_npending		= sfc_ef10_essb_rx_qdesc_npending,
731 	.qdesc_status		= sfc_ef10_essb_rx_qdesc_status,
732 	.pkt_burst		= sfc_ef10_essb_recv_pkts,
733 };
734