xref: /f-stack/dpdk/drivers/net/sfc/sfc_ef10_rx.c (revision 031be553)
1 /*-
2  *   BSD LICENSE
3  *
4  * Copyright (c) 2016 Solarflare Communications Inc.
5  * All rights reserved.
6  *
7  * This software was jointly developed between OKTET Labs (under contract
8  * for Solarflare) and Solarflare Communications, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions are met:
12  *
13  * 1. Redistributions of source code must retain the above copyright notice,
14  *    this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright notice,
16  *    this list of conditions and the following disclaimer in the documentation
17  *    and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
20  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
21  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
23  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
24  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
25  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
26  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
27  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
28  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
29  * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /* EF10 native datapath implementation */
33 
34 #include <stdbool.h>
35 
36 #include <rte_byteorder.h>
37 #include <rte_mbuf_ptype.h>
38 #include <rte_mbuf.h>
39 #include <rte_io.h>
40 
41 #include "efx.h"
42 #include "efx_types.h"
43 #include "efx_regs.h"
44 #include "efx_regs_ef10.h"
45 
46 #include "sfc_tweak.h"
47 #include "sfc_dp_rx.h"
48 #include "sfc_kvargs.h"
49 #include "sfc_ef10.h"
50 
51 #define sfc_ef10_rx_err(dpq, ...) \
52 	SFC_DP_LOG(SFC_KVARG_DATAPATH_EF10, ERR, dpq, __VA_ARGS__)
53 
54 /**
55  * Alignment requirement for value written to RX WPTR:
56  * the WPTR must be aligned to an 8 descriptor boundary.
57  */
58 #define SFC_EF10_RX_WPTR_ALIGN	8
59 
60 /**
61  * Maximum number of descriptors/buffers in the Rx ring.
62  * It should guarantee that corresponding event queue never overfill.
63  * EF10 native datapath uses event queue of the same size as Rx queue.
64  * Maximum number of events on datapath can be estimated as number of
65  * Rx queue entries (one event per Rx buffer in the worst case) plus
66  * Rx error and flush events.
67  */
68 #define SFC_EF10_RXQ_LIMIT(_ndesc) \
69 	((_ndesc) - 1 /* head must not step on tail */ - \
70 	 (SFC_EF10_EV_PER_CACHE_LINE - 1) /* max unused EvQ entries */ - \
71 	 1 /* Rx error */ - 1 /* flush */)
72 
73 struct sfc_ef10_rx_sw_desc {
74 	struct rte_mbuf			*mbuf;
75 };
76 
77 struct sfc_ef10_rxq {
78 	/* Used on data path */
79 	unsigned int			flags;
80 #define SFC_EF10_RXQ_STARTED		0x1
81 #define SFC_EF10_RXQ_NOT_RUNNING	0x2
82 #define SFC_EF10_RXQ_EXCEPTION		0x4
83 #define SFC_EF10_RXQ_RSS_HASH		0x8
84 	unsigned int			ptr_mask;
85 	unsigned int			prepared;
86 	unsigned int			completed;
87 	unsigned int			evq_read_ptr;
88 	efx_qword_t			*evq_hw_ring;
89 	struct sfc_ef10_rx_sw_desc	*sw_ring;
90 	uint64_t			rearm_data;
91 	uint16_t			prefix_size;
92 
93 	/* Used on refill */
94 	uint16_t			buf_size;
95 	unsigned int			added;
96 	unsigned int			refill_threshold;
97 	struct rte_mempool		*refill_mb_pool;
98 	efx_qword_t			*rxq_hw_ring;
99 	volatile void			*doorbell;
100 
101 	/* Datapath receive queue anchor */
102 	struct sfc_dp_rxq		dp;
103 };
104 
105 static inline struct sfc_ef10_rxq *
106 sfc_ef10_rxq_by_dp_rxq(struct sfc_dp_rxq *dp_rxq)
107 {
108 	return container_of(dp_rxq, struct sfc_ef10_rxq, dp);
109 }
110 
111 static void
112 sfc_ef10_rx_qpush(struct sfc_ef10_rxq *rxq)
113 {
114 	efx_dword_t dword;
115 
116 	/* Hardware has alignment restriction for WPTR */
117 	RTE_BUILD_BUG_ON(SFC_RX_REFILL_BULK % SFC_EF10_RX_WPTR_ALIGN != 0);
118 	SFC_ASSERT(RTE_ALIGN(rxq->added, SFC_EF10_RX_WPTR_ALIGN) == rxq->added);
119 
120 	EFX_POPULATE_DWORD_1(dword, ERF_DZ_RX_DESC_WPTR,
121 			     rxq->added & rxq->ptr_mask);
122 
123 	/* DMA sync to device is not required */
124 
125 	/*
126 	 * rte_write32() has rte_io_wmb() which guarantees that the STORE
127 	 * operations (i.e. Rx and event descriptor updates) that precede
128 	 * the rte_io_wmb() call are visible to NIC before the STORE
129 	 * operations that follow it (i.e. doorbell write).
130 	 */
131 	rte_write32(dword.ed_u32[0], rxq->doorbell);
132 }
133 
134 static void
135 sfc_ef10_rx_qrefill(struct sfc_ef10_rxq *rxq)
136 {
137 	const unsigned int ptr_mask = rxq->ptr_mask;
138 	const uint32_t buf_size = rxq->buf_size;
139 	unsigned int free_space;
140 	unsigned int bulks;
141 	void *objs[SFC_RX_REFILL_BULK];
142 	unsigned int added = rxq->added;
143 
144 	free_space = SFC_EF10_RXQ_LIMIT(ptr_mask + 1) -
145 		(added - rxq->completed);
146 
147 	if (free_space < rxq->refill_threshold)
148 		return;
149 
150 	bulks = free_space / RTE_DIM(objs);
151 	/* refill_threshold guarantees that bulks is positive */
152 	SFC_ASSERT(bulks > 0);
153 
154 	do {
155 		unsigned int id;
156 		unsigned int i;
157 
158 		if (unlikely(rte_mempool_get_bulk(rxq->refill_mb_pool, objs,
159 						  RTE_DIM(objs)) < 0)) {
160 			struct rte_eth_dev_data *dev_data =
161 				rte_eth_devices[rxq->dp.dpq.port_id].data;
162 
163 			/*
164 			 * It is hardly a safe way to increment counter
165 			 * from different contexts, but all PMDs do it.
166 			 */
167 			dev_data->rx_mbuf_alloc_failed += RTE_DIM(objs);
168 			/* Return if we have posted nothing yet */
169 			if (added == rxq->added)
170 				return;
171 			/* Push posted */
172 			break;
173 		}
174 
175 		for (i = 0, id = added & ptr_mask;
176 		     i < RTE_DIM(objs);
177 		     ++i, ++id) {
178 			struct rte_mbuf *m = objs[i];
179 			struct sfc_ef10_rx_sw_desc *rxd;
180 			rte_iova_t phys_addr;
181 
182 			SFC_ASSERT((id & ~ptr_mask) == 0);
183 			rxd = &rxq->sw_ring[id];
184 			rxd->mbuf = m;
185 
186 			/*
187 			 * Avoid writing to mbuf. It is cheaper to do it
188 			 * when we receive packet and fill in nearby
189 			 * structure members.
190 			 */
191 
192 			phys_addr = rte_mbuf_data_iova_default(m);
193 			EFX_POPULATE_QWORD_2(rxq->rxq_hw_ring[id],
194 			    ESF_DZ_RX_KER_BYTE_CNT, buf_size,
195 			    ESF_DZ_RX_KER_BUF_ADDR, phys_addr);
196 		}
197 
198 		added += RTE_DIM(objs);
199 	} while (--bulks > 0);
200 
201 	SFC_ASSERT(rxq->added != added);
202 	rxq->added = added;
203 	sfc_ef10_rx_qpush(rxq);
204 }
205 
206 static void
207 sfc_ef10_rx_prefetch_next(struct sfc_ef10_rxq *rxq, unsigned int next_id)
208 {
209 	struct rte_mbuf *next_mbuf;
210 
211 	/* Prefetch next bunch of software descriptors */
212 	if ((next_id % (RTE_CACHE_LINE_SIZE / sizeof(rxq->sw_ring[0]))) == 0)
213 		rte_prefetch0(&rxq->sw_ring[next_id]);
214 
215 	/*
216 	 * It looks strange to prefetch depending on previous prefetch
217 	 * data, but measurements show that it is really efficient and
218 	 * increases packet rate.
219 	 */
220 	next_mbuf = rxq->sw_ring[next_id].mbuf;
221 	if (likely(next_mbuf != NULL)) {
222 		/* Prefetch the next mbuf structure */
223 		rte_mbuf_prefetch_part1(next_mbuf);
224 
225 		/* Prefetch pseudo header of the next packet */
226 		/* data_off is not filled in yet */
227 		/* Yes, data could be not ready yet, but we hope */
228 		rte_prefetch0((uint8_t *)next_mbuf->buf_addr +
229 			      RTE_PKTMBUF_HEADROOM);
230 	}
231 }
232 
233 static uint16_t
234 sfc_ef10_rx_prepared(struct sfc_ef10_rxq *rxq, struct rte_mbuf **rx_pkts,
235 		     uint16_t nb_pkts)
236 {
237 	uint16_t n_rx_pkts = RTE_MIN(nb_pkts, rxq->prepared);
238 	unsigned int completed = rxq->completed;
239 	unsigned int i;
240 
241 	rxq->prepared -= n_rx_pkts;
242 	rxq->completed = completed + n_rx_pkts;
243 
244 	for (i = 0; i < n_rx_pkts; ++i, ++completed)
245 		rx_pkts[i] = rxq->sw_ring[completed & rxq->ptr_mask].mbuf;
246 
247 	return n_rx_pkts;
248 }
249 
250 static void
251 sfc_ef10_rx_ev_to_offloads(struct sfc_ef10_rxq *rxq, const efx_qword_t rx_ev,
252 			   struct rte_mbuf *m)
253 {
254 	uint32_t l2_ptype = 0;
255 	uint32_t l3_ptype = 0;
256 	uint32_t l4_ptype = 0;
257 	uint64_t ol_flags = 0;
258 
259 	if (unlikely(EFX_TEST_QWORD_BIT(rx_ev, ESF_DZ_RX_PARSE_INCOMPLETE_LBN)))
260 		goto done;
261 
262 	switch (EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_ETH_TAG_CLASS)) {
263 	case ESE_DZ_ETH_TAG_CLASS_NONE:
264 		l2_ptype = RTE_PTYPE_L2_ETHER;
265 		break;
266 	case ESE_DZ_ETH_TAG_CLASS_VLAN1:
267 		l2_ptype = RTE_PTYPE_L2_ETHER_VLAN;
268 		break;
269 	case ESE_DZ_ETH_TAG_CLASS_VLAN2:
270 		l2_ptype = RTE_PTYPE_L2_ETHER_QINQ;
271 		break;
272 	default:
273 		/* Unexpected Eth tag class */
274 		SFC_ASSERT(false);
275 	}
276 
277 	switch (EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_L3_CLASS)) {
278 	case ESE_DZ_L3_CLASS_IP4_FRAG:
279 		l4_ptype = RTE_PTYPE_L4_FRAG;
280 		/* FALLTHROUGH */
281 	case ESE_DZ_L3_CLASS_IP4:
282 		l3_ptype = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
283 		ol_flags |= PKT_RX_RSS_HASH |
284 			((EFX_TEST_QWORD_BIT(rx_ev,
285 					     ESF_DZ_RX_IPCKSUM_ERR_LBN)) ?
286 			 PKT_RX_IP_CKSUM_BAD : PKT_RX_IP_CKSUM_GOOD);
287 		break;
288 	case ESE_DZ_L3_CLASS_IP6_FRAG:
289 		l4_ptype = RTE_PTYPE_L4_FRAG;
290 		/* FALLTHROUGH */
291 	case ESE_DZ_L3_CLASS_IP6:
292 		l3_ptype = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
293 		ol_flags |= PKT_RX_RSS_HASH;
294 		break;
295 	case ESE_DZ_L3_CLASS_ARP:
296 		/* Override Layer 2 packet type */
297 		l2_ptype = RTE_PTYPE_L2_ETHER_ARP;
298 		break;
299 	case ESE_DZ_L3_CLASS_UNKNOWN:
300 		break;
301 	default:
302 		/* Unexpected Layer 3 class */
303 		SFC_ASSERT(false);
304 	}
305 
306 	switch (EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_L4_CLASS)) {
307 	case ESE_DZ_L4_CLASS_TCP:
308 		l4_ptype = RTE_PTYPE_L4_TCP;
309 		ol_flags |=
310 			(EFX_TEST_QWORD_BIT(rx_ev,
311 					    ESF_DZ_RX_TCPUDP_CKSUM_ERR_LBN)) ?
312 			PKT_RX_L4_CKSUM_BAD : PKT_RX_L4_CKSUM_GOOD;
313 		break;
314 	case ESE_DZ_L4_CLASS_UDP:
315 		l4_ptype = RTE_PTYPE_L4_UDP;
316 		ol_flags |=
317 			(EFX_TEST_QWORD_BIT(rx_ev,
318 					    ESF_DZ_RX_TCPUDP_CKSUM_ERR_LBN)) ?
319 			PKT_RX_L4_CKSUM_BAD : PKT_RX_L4_CKSUM_GOOD;
320 		break;
321 	case ESE_DZ_L4_CLASS_UNKNOWN:
322 		break;
323 	default:
324 		/* Unexpected Layer 4 class */
325 		SFC_ASSERT(false);
326 	}
327 
328 	/* Remove RSS hash offload flag if RSS is not enabled */
329 	if (~rxq->flags & SFC_EF10_RXQ_RSS_HASH)
330 		ol_flags &= ~PKT_RX_RSS_HASH;
331 
332 done:
333 	m->ol_flags = ol_flags;
334 	m->packet_type = l2_ptype | l3_ptype | l4_ptype;
335 }
336 
337 static uint16_t
338 sfc_ef10_rx_pseudo_hdr_get_len(const uint8_t *pseudo_hdr)
339 {
340 	return rte_le_to_cpu_16(*(const uint16_t *)&pseudo_hdr[8]);
341 }
342 
343 static uint32_t
344 sfc_ef10_rx_pseudo_hdr_get_hash(const uint8_t *pseudo_hdr)
345 {
346 	return rte_le_to_cpu_32(*(const uint32_t *)pseudo_hdr);
347 }
348 
349 static uint16_t
350 sfc_ef10_rx_process_event(struct sfc_ef10_rxq *rxq, efx_qword_t rx_ev,
351 			  struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
352 {
353 	const unsigned int ptr_mask = rxq->ptr_mask;
354 	unsigned int completed = rxq->completed;
355 	unsigned int ready;
356 	struct sfc_ef10_rx_sw_desc *rxd;
357 	struct rte_mbuf *m;
358 	struct rte_mbuf *m0;
359 	uint16_t n_rx_pkts;
360 	const uint8_t *pseudo_hdr;
361 	uint16_t pkt_len;
362 
363 	ready = (EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_DSC_PTR_LBITS) - completed) &
364 		EFX_MASK32(ESF_DZ_RX_DSC_PTR_LBITS);
365 	SFC_ASSERT(ready > 0);
366 
367 	if (rx_ev.eq_u64[0] &
368 	    rte_cpu_to_le_64((1ull << ESF_DZ_RX_ECC_ERR_LBN) |
369 			     (1ull << ESF_DZ_RX_ECRC_ERR_LBN))) {
370 		SFC_ASSERT(rxq->prepared == 0);
371 		rxq->completed += ready;
372 		while (ready-- > 0) {
373 			rxd = &rxq->sw_ring[completed++ & ptr_mask];
374 			rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
375 		}
376 		return 0;
377 	}
378 
379 	n_rx_pkts = RTE_MIN(ready, nb_pkts);
380 	rxq->prepared = ready - n_rx_pkts;
381 	rxq->completed += n_rx_pkts;
382 
383 	rxd = &rxq->sw_ring[completed++ & ptr_mask];
384 
385 	sfc_ef10_rx_prefetch_next(rxq, completed & ptr_mask);
386 
387 	m = rxd->mbuf;
388 
389 	*rx_pkts++ = m;
390 
391 	RTE_BUILD_BUG_ON(sizeof(m->rearm_data[0]) != sizeof(rxq->rearm_data));
392 	m->rearm_data[0] = rxq->rearm_data;
393 
394 	/* Classify packet based on Rx event */
395 	sfc_ef10_rx_ev_to_offloads(rxq, rx_ev, m);
396 
397 	/* data_off already moved past pseudo header */
398 	pseudo_hdr = (uint8_t *)m->buf_addr + RTE_PKTMBUF_HEADROOM;
399 
400 	/*
401 	 * Always get RSS hash from pseudo header to avoid
402 	 * condition/branching. If it is valid or not depends on
403 	 * PKT_RX_RSS_HASH in m->ol_flags.
404 	 */
405 	m->hash.rss = sfc_ef10_rx_pseudo_hdr_get_hash(pseudo_hdr);
406 
407 	if (ready == 1)
408 		pkt_len = EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_BYTES) -
409 			rxq->prefix_size;
410 	else
411 		pkt_len = sfc_ef10_rx_pseudo_hdr_get_len(pseudo_hdr);
412 	SFC_ASSERT(pkt_len > 0);
413 	rte_pktmbuf_data_len(m) = pkt_len;
414 	rte_pktmbuf_pkt_len(m) = pkt_len;
415 
416 	SFC_ASSERT(m->next == NULL);
417 
418 	/* Remember mbuf to copy offload flags and packet type from */
419 	m0 = m;
420 	for (--ready; ready > 0; --ready) {
421 		rxd = &rxq->sw_ring[completed++ & ptr_mask];
422 
423 		sfc_ef10_rx_prefetch_next(rxq, completed & ptr_mask);
424 
425 		m = rxd->mbuf;
426 
427 		if (ready > rxq->prepared)
428 			*rx_pkts++ = m;
429 
430 		RTE_BUILD_BUG_ON(sizeof(m->rearm_data[0]) !=
431 				 sizeof(rxq->rearm_data));
432 		m->rearm_data[0] = rxq->rearm_data;
433 
434 		/* Event-dependent information is the same */
435 		m->ol_flags = m0->ol_flags;
436 		m->packet_type = m0->packet_type;
437 
438 		/* data_off already moved past pseudo header */
439 		pseudo_hdr = (uint8_t *)m->buf_addr + RTE_PKTMBUF_HEADROOM;
440 
441 		/*
442 		 * Always get RSS hash from pseudo header to avoid
443 		 * condition/branching. If it is valid or not depends on
444 		 * PKT_RX_RSS_HASH in m->ol_flags.
445 		 */
446 		m->hash.rss = sfc_ef10_rx_pseudo_hdr_get_hash(pseudo_hdr);
447 
448 		pkt_len = sfc_ef10_rx_pseudo_hdr_get_len(pseudo_hdr);
449 		SFC_ASSERT(pkt_len > 0);
450 		rte_pktmbuf_data_len(m) = pkt_len;
451 		rte_pktmbuf_pkt_len(m) = pkt_len;
452 
453 		SFC_ASSERT(m->next == NULL);
454 	}
455 
456 	return n_rx_pkts;
457 }
458 
459 static bool
460 sfc_ef10_rx_get_event(struct sfc_ef10_rxq *rxq, efx_qword_t *rx_ev)
461 {
462 	*rx_ev = rxq->evq_hw_ring[rxq->evq_read_ptr & rxq->ptr_mask];
463 
464 	if (!sfc_ef10_ev_present(*rx_ev))
465 		return false;
466 
467 	if (unlikely(EFX_QWORD_FIELD(*rx_ev, FSF_AZ_EV_CODE) !=
468 		     FSE_AZ_EV_CODE_RX_EV)) {
469 		/*
470 		 * Do not move read_ptr to keep the event for exception
471 		 * handling by the control path.
472 		 */
473 		rxq->flags |= SFC_EF10_RXQ_EXCEPTION;
474 		sfc_ef10_rx_err(&rxq->dp.dpq,
475 				"RxQ exception at EvQ read ptr %#x",
476 				rxq->evq_read_ptr);
477 		return false;
478 	}
479 
480 	rxq->evq_read_ptr++;
481 	return true;
482 }
483 
484 static uint16_t
485 sfc_ef10_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
486 {
487 	struct sfc_ef10_rxq *rxq = sfc_ef10_rxq_by_dp_rxq(rx_queue);
488 	unsigned int evq_old_read_ptr;
489 	uint16_t n_rx_pkts;
490 	efx_qword_t rx_ev;
491 
492 	if (unlikely(rxq->flags &
493 		     (SFC_EF10_RXQ_NOT_RUNNING | SFC_EF10_RXQ_EXCEPTION)))
494 		return 0;
495 
496 	n_rx_pkts = sfc_ef10_rx_prepared(rxq, rx_pkts, nb_pkts);
497 
498 	evq_old_read_ptr = rxq->evq_read_ptr;
499 	while (n_rx_pkts != nb_pkts && sfc_ef10_rx_get_event(rxq, &rx_ev)) {
500 		/*
501 		 * DROP_EVENT is an internal to the NIC, software should
502 		 * never see it and, therefore, may ignore it.
503 		 */
504 
505 		n_rx_pkts += sfc_ef10_rx_process_event(rxq, rx_ev,
506 						       rx_pkts + n_rx_pkts,
507 						       nb_pkts - n_rx_pkts);
508 	}
509 
510 	sfc_ef10_ev_qclear(rxq->evq_hw_ring, rxq->ptr_mask, evq_old_read_ptr,
511 			   rxq->evq_read_ptr);
512 
513 	/* It is not a problem if we refill in the case of exception */
514 	sfc_ef10_rx_qrefill(rxq);
515 
516 	return n_rx_pkts;
517 }
518 
519 static const uint32_t *
520 sfc_ef10_supported_ptypes_get(void)
521 {
522 	static const uint32_t ef10_native_ptypes[] = {
523 		RTE_PTYPE_L2_ETHER,
524 		RTE_PTYPE_L2_ETHER_ARP,
525 		RTE_PTYPE_L2_ETHER_VLAN,
526 		RTE_PTYPE_L2_ETHER_QINQ,
527 		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
528 		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
529 		RTE_PTYPE_L4_FRAG,
530 		RTE_PTYPE_L4_TCP,
531 		RTE_PTYPE_L4_UDP,
532 		RTE_PTYPE_UNKNOWN
533 	};
534 
535 	return ef10_native_ptypes;
536 }
537 
538 static sfc_dp_rx_qdesc_npending_t sfc_ef10_rx_qdesc_npending;
539 static unsigned int
540 sfc_ef10_rx_qdesc_npending(__rte_unused struct sfc_dp_rxq *dp_rxq)
541 {
542 	/*
543 	 * Correct implementation requires EvQ polling and events
544 	 * processing (keeping all ready mbufs in prepared).
545 	 */
546 	return -ENOTSUP;
547 }
548 
549 static sfc_dp_rx_qdesc_status_t sfc_ef10_rx_qdesc_status;
550 static int
551 sfc_ef10_rx_qdesc_status(__rte_unused struct sfc_dp_rxq *dp_rxq,
552 			 __rte_unused uint16_t offset)
553 {
554 	return -ENOTSUP;
555 }
556 
557 
558 static uint64_t
559 sfc_ef10_mk_mbuf_rearm_data(uint16_t port_id, uint16_t prefix_size)
560 {
561 	struct rte_mbuf m;
562 
563 	memset(&m, 0, sizeof(m));
564 
565 	rte_mbuf_refcnt_set(&m, 1);
566 	m.data_off = RTE_PKTMBUF_HEADROOM + prefix_size;
567 	m.nb_segs = 1;
568 	m.port = port_id;
569 
570 	/* rearm_data covers structure members filled in above */
571 	rte_compiler_barrier();
572 	RTE_BUILD_BUG_ON(sizeof(m.rearm_data[0]) != sizeof(uint64_t));
573 	return m.rearm_data[0];
574 }
575 
576 static sfc_dp_rx_qcreate_t sfc_ef10_rx_qcreate;
577 static int
578 sfc_ef10_rx_qcreate(uint16_t port_id, uint16_t queue_id,
579 		    const struct rte_pci_addr *pci_addr, int socket_id,
580 		    const struct sfc_dp_rx_qcreate_info *info,
581 		    struct sfc_dp_rxq **dp_rxqp)
582 {
583 	struct sfc_ef10_rxq *rxq;
584 	int rc;
585 
586 	rc = EINVAL;
587 	if (info->rxq_entries != info->evq_entries)
588 		goto fail_rxq_args;
589 
590 	rc = ENOMEM;
591 	rxq = rte_zmalloc_socket("sfc-ef10-rxq", sizeof(*rxq),
592 				 RTE_CACHE_LINE_SIZE, socket_id);
593 	if (rxq == NULL)
594 		goto fail_rxq_alloc;
595 
596 	sfc_dp_queue_init(&rxq->dp.dpq, port_id, queue_id, pci_addr);
597 
598 	rc = ENOMEM;
599 	rxq->sw_ring = rte_calloc_socket("sfc-ef10-rxq-sw_ring",
600 					 info->rxq_entries,
601 					 sizeof(*rxq->sw_ring),
602 					 RTE_CACHE_LINE_SIZE, socket_id);
603 	if (rxq->sw_ring == NULL)
604 		goto fail_desc_alloc;
605 
606 	rxq->flags |= SFC_EF10_RXQ_NOT_RUNNING;
607 	if (info->flags & SFC_RXQ_FLAG_RSS_HASH)
608 		rxq->flags |= SFC_EF10_RXQ_RSS_HASH;
609 	rxq->ptr_mask = info->rxq_entries - 1;
610 	rxq->evq_hw_ring = info->evq_hw_ring;
611 	rxq->refill_threshold = info->refill_threshold;
612 	rxq->rearm_data =
613 		sfc_ef10_mk_mbuf_rearm_data(port_id, info->prefix_size);
614 	rxq->prefix_size = info->prefix_size;
615 	rxq->buf_size = info->buf_size;
616 	rxq->refill_mb_pool = info->refill_mb_pool;
617 	rxq->rxq_hw_ring = info->rxq_hw_ring;
618 	rxq->doorbell = (volatile uint8_t *)info->mem_bar +
619 			ER_DZ_RX_DESC_UPD_REG_OFST +
620 			info->hw_index * ER_DZ_RX_DESC_UPD_REG_STEP;
621 
622 	*dp_rxqp = &rxq->dp;
623 	return 0;
624 
625 fail_desc_alloc:
626 	rte_free(rxq);
627 
628 fail_rxq_alloc:
629 fail_rxq_args:
630 	return rc;
631 }
632 
633 static sfc_dp_rx_qdestroy_t sfc_ef10_rx_qdestroy;
634 static void
635 sfc_ef10_rx_qdestroy(struct sfc_dp_rxq *dp_rxq)
636 {
637 	struct sfc_ef10_rxq *rxq = sfc_ef10_rxq_by_dp_rxq(dp_rxq);
638 
639 	rte_free(rxq->sw_ring);
640 	rte_free(rxq);
641 }
642 
643 static sfc_dp_rx_qstart_t sfc_ef10_rx_qstart;
644 static int
645 sfc_ef10_rx_qstart(struct sfc_dp_rxq *dp_rxq, unsigned int evq_read_ptr)
646 {
647 	struct sfc_ef10_rxq *rxq = sfc_ef10_rxq_by_dp_rxq(dp_rxq);
648 
649 	rxq->prepared = 0;
650 	rxq->completed = rxq->added = 0;
651 
652 	sfc_ef10_rx_qrefill(rxq);
653 
654 	rxq->evq_read_ptr = evq_read_ptr;
655 
656 	rxq->flags |= SFC_EF10_RXQ_STARTED;
657 	rxq->flags &= ~(SFC_EF10_RXQ_NOT_RUNNING | SFC_EF10_RXQ_EXCEPTION);
658 
659 	return 0;
660 }
661 
662 static sfc_dp_rx_qstop_t sfc_ef10_rx_qstop;
663 static void
664 sfc_ef10_rx_qstop(struct sfc_dp_rxq *dp_rxq, unsigned int *evq_read_ptr)
665 {
666 	struct sfc_ef10_rxq *rxq = sfc_ef10_rxq_by_dp_rxq(dp_rxq);
667 
668 	rxq->flags |= SFC_EF10_RXQ_NOT_RUNNING;
669 
670 	*evq_read_ptr = rxq->evq_read_ptr;
671 }
672 
673 static sfc_dp_rx_qrx_ev_t sfc_ef10_rx_qrx_ev;
674 static bool
675 sfc_ef10_rx_qrx_ev(struct sfc_dp_rxq *dp_rxq, __rte_unused unsigned int id)
676 {
677 	__rte_unused struct sfc_ef10_rxq *rxq = sfc_ef10_rxq_by_dp_rxq(dp_rxq);
678 
679 	SFC_ASSERT(rxq->flags & SFC_EF10_RXQ_NOT_RUNNING);
680 
681 	/*
682 	 * It is safe to ignore Rx event since we free all mbufs on
683 	 * queue purge anyway.
684 	 */
685 
686 	return false;
687 }
688 
689 static sfc_dp_rx_qpurge_t sfc_ef10_rx_qpurge;
690 static void
691 sfc_ef10_rx_qpurge(struct sfc_dp_rxq *dp_rxq)
692 {
693 	struct sfc_ef10_rxq *rxq = sfc_ef10_rxq_by_dp_rxq(dp_rxq);
694 	unsigned int i;
695 	struct sfc_ef10_rx_sw_desc *rxd;
696 
697 	for (i = rxq->completed; i != rxq->added; ++i) {
698 		rxd = &rxq->sw_ring[i & rxq->ptr_mask];
699 		rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
700 		rxd->mbuf = NULL;
701 	}
702 
703 	rxq->flags &= ~SFC_EF10_RXQ_STARTED;
704 }
705 
706 struct sfc_dp_rx sfc_ef10_rx = {
707 	.dp = {
708 		.name		= SFC_KVARG_DATAPATH_EF10,
709 		.type		= SFC_DP_RX,
710 		.hw_fw_caps	= SFC_DP_HW_FW_CAP_EF10,
711 	},
712 	.features		= SFC_DP_RX_FEAT_MULTI_PROCESS,
713 	.qcreate		= sfc_ef10_rx_qcreate,
714 	.qdestroy		= sfc_ef10_rx_qdestroy,
715 	.qstart			= sfc_ef10_rx_qstart,
716 	.qstop			= sfc_ef10_rx_qstop,
717 	.qrx_ev			= sfc_ef10_rx_qrx_ev,
718 	.qpurge			= sfc_ef10_rx_qpurge,
719 	.supported_ptypes_get	= sfc_ef10_supported_ptypes_get,
720 	.qdesc_npending		= sfc_ef10_rx_qdesc_npending,
721 	.qdesc_status		= sfc_ef10_rx_qdesc_status,
722 	.pkt_burst		= sfc_ef10_recv_pkts,
723 };
724