xref: /dpdk/drivers/net/sfc/sfc_ef10_tx.c (revision 8809f78c7dd9f33a44a4f89c58fc91ded34296ed)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  *
3  * Copyright(c) 2019-2020 Xilinx, Inc.
4  * Copyright(c) 2016-2019 Solarflare Communications Inc.
5  *
6  * This software was jointly developed between OKTET Labs (under contract
7  * for Solarflare) and Solarflare Communications, Inc.
8  */
9 
10 #include <stdbool.h>
11 
12 #include <rte_mbuf.h>
13 #include <rte_io.h>
14 #include <rte_ip.h>
15 #include <rte_tcp.h>
16 
17 #include "efx.h"
18 #include "efx_types.h"
19 #include "efx_regs.h"
20 #include "efx_regs_ef10.h"
21 
22 #include "sfc_debug.h"
23 #include "sfc_dp_tx.h"
24 #include "sfc_tweak.h"
25 #include "sfc_kvargs.h"
26 #include "sfc_ef10.h"
27 #include "sfc_tso.h"
28 
29 #define sfc_ef10_tx_err(dpq, ...) \
30 	SFC_DP_LOG(SFC_KVARG_DATAPATH_EF10, ERR, dpq, __VA_ARGS__)
31 
32 #define sfc_ef10_tx_info(dpq, ...) \
33 	SFC_DP_LOG(SFC_KVARG_DATAPATH_EF10, INFO, dpq, __VA_ARGS__)
34 
35 /** Maximum length of the DMA descriptor data */
36 #define SFC_EF10_TX_DMA_DESC_LEN_MAX \
37 	((1u << ESF_DZ_TX_KER_BYTE_CNT_WIDTH) - 1)
38 
39 /**
40  * Maximum number of descriptors/buffers in the Tx ring.
41  * It should guarantee that corresponding event queue never overfill.
42  * EF10 native datapath uses event queue of the same size as Tx queue.
43  * Maximum number of events on datapath can be estimated as number of
44  * Tx queue entries (one event per Tx buffer in the worst case) plus
45  * Tx error and flush events.
46  */
47 #define SFC_EF10_TXQ_LIMIT(_ndesc) \
48 	((_ndesc) - 1 /* head must not step on tail */ - \
49 	 (SFC_EF10_EV_PER_CACHE_LINE - 1) /* max unused EvQ entries */ - \
50 	 1 /* Rx error */ - 1 /* flush */)
51 
52 struct sfc_ef10_tx_sw_desc {
53 	struct rte_mbuf			*mbuf;
54 };
55 
56 struct sfc_ef10_txq {
57 	unsigned int			flags;
58 #define SFC_EF10_TXQ_STARTED		0x1
59 #define SFC_EF10_TXQ_NOT_RUNNING	0x2
60 #define SFC_EF10_TXQ_EXCEPTION		0x4
61 
62 	unsigned int			ptr_mask;
63 	unsigned int			added;
64 	unsigned int			completed;
65 	unsigned int			max_fill_level;
66 	unsigned int			free_thresh;
67 	unsigned int			evq_read_ptr;
68 	struct sfc_ef10_tx_sw_desc	*sw_ring;
69 	efx_qword_t			*txq_hw_ring;
70 	volatile void			*doorbell;
71 	efx_qword_t			*evq_hw_ring;
72 	uint8_t				*tsoh;
73 	rte_iova_t			tsoh_iova;
74 	uint16_t			tso_tcp_header_offset_limit;
75 
76 	/* Datapath transmit queue anchor */
77 	struct sfc_dp_txq		dp;
78 };
79 
80 static inline struct sfc_ef10_txq *
81 sfc_ef10_txq_by_dp_txq(struct sfc_dp_txq *dp_txq)
82 {
83 	return container_of(dp_txq, struct sfc_ef10_txq, dp);
84 }
85 
86 static bool
87 sfc_ef10_tx_get_event(struct sfc_ef10_txq *txq, efx_qword_t *tx_ev)
88 {
89 	volatile efx_qword_t *evq_hw_ring = txq->evq_hw_ring;
90 
91 	/*
92 	 * Exception flag is set when reap is done.
93 	 * It is never done twice per packet burst get and absence of
94 	 * the flag is checked on burst get entry.
95 	 */
96 	SFC_ASSERT((txq->flags & SFC_EF10_TXQ_EXCEPTION) == 0);
97 
98 	*tx_ev = evq_hw_ring[txq->evq_read_ptr & txq->ptr_mask];
99 
100 	if (!sfc_ef10_ev_present(*tx_ev))
101 		return false;
102 
103 	if (unlikely(EFX_QWORD_FIELD(*tx_ev, FSF_AZ_EV_CODE) !=
104 		     FSE_AZ_EV_CODE_TX_EV)) {
105 		/*
106 		 * Do not move read_ptr to keep the event for exception
107 		 * handling by the control path.
108 		 */
109 		txq->flags |= SFC_EF10_TXQ_EXCEPTION;
110 		sfc_ef10_tx_err(&txq->dp.dpq,
111 				"TxQ exception at EvQ read ptr %#x",
112 				txq->evq_read_ptr);
113 		return false;
114 	}
115 
116 	txq->evq_read_ptr++;
117 	return true;
118 }
119 
120 static unsigned int
121 sfc_ef10_tx_process_events(struct sfc_ef10_txq *txq)
122 {
123 	const unsigned int curr_done = txq->completed - 1;
124 	unsigned int anew_done = curr_done;
125 	efx_qword_t tx_ev;
126 
127 	while (sfc_ef10_tx_get_event(txq, &tx_ev)) {
128 		/*
129 		 * DROP_EVENT is an internal to the NIC, software should
130 		 * never see it and, therefore, may ignore it.
131 		 */
132 
133 		/* Update the latest done descriptor */
134 		anew_done = EFX_QWORD_FIELD(tx_ev, ESF_DZ_TX_DESCR_INDX);
135 	}
136 	return (anew_done - curr_done) & txq->ptr_mask;
137 }
138 
139 static void
140 sfc_ef10_tx_reap(struct sfc_ef10_txq *txq)
141 {
142 	const unsigned int old_read_ptr = txq->evq_read_ptr;
143 	const unsigned int ptr_mask = txq->ptr_mask;
144 	unsigned int completed = txq->completed;
145 	unsigned int pending = completed;
146 
147 	pending += sfc_ef10_tx_process_events(txq);
148 
149 	if (pending != completed) {
150 		struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
151 		unsigned int nb = 0;
152 
153 		do {
154 			struct sfc_ef10_tx_sw_desc *txd;
155 			struct rte_mbuf *m;
156 
157 			txd = &txq->sw_ring[completed & ptr_mask];
158 			if (txd->mbuf == NULL)
159 				continue;
160 
161 			m = rte_pktmbuf_prefree_seg(txd->mbuf);
162 			txd->mbuf = NULL;
163 			if (m == NULL)
164 				continue;
165 
166 			if ((nb == RTE_DIM(bulk)) ||
167 			    ((nb != 0) && (m->pool != bulk[0]->pool))) {
168 				rte_mempool_put_bulk(bulk[0]->pool,
169 						     (void *)bulk, nb);
170 				nb = 0;
171 			}
172 
173 			bulk[nb++] = m;
174 		} while (++completed != pending);
175 
176 		if (nb != 0)
177 			rte_mempool_put_bulk(bulk[0]->pool, (void *)bulk, nb);
178 
179 		txq->completed = completed;
180 	}
181 
182 	sfc_ef10_ev_qclear(txq->evq_hw_ring, ptr_mask, old_read_ptr,
183 			   txq->evq_read_ptr);
184 }
185 
186 static void
187 sfc_ef10_tx_qdesc_dma_create(rte_iova_t addr, uint16_t size, bool eop,
188 			     efx_qword_t *edp)
189 {
190 	EFX_POPULATE_QWORD_4(*edp,
191 			     ESF_DZ_TX_KER_TYPE, 0,
192 			     ESF_DZ_TX_KER_CONT, !eop,
193 			     ESF_DZ_TX_KER_BYTE_CNT, size,
194 			     ESF_DZ_TX_KER_BUF_ADDR, addr);
195 }
196 
197 static void
198 sfc_ef10_tx_qdesc_tso2_create(struct sfc_ef10_txq * const txq,
199 			      unsigned int added, uint16_t ipv4_id,
200 			      uint16_t outer_ipv4_id, uint32_t tcp_seq,
201 			      uint16_t tcp_mss)
202 {
203 	EFX_POPULATE_QWORD_5(txq->txq_hw_ring[added & txq->ptr_mask],
204 			    ESF_DZ_TX_DESC_IS_OPT, 1,
205 			    ESF_DZ_TX_OPTION_TYPE,
206 			    ESE_DZ_TX_OPTION_DESC_TSO,
207 			    ESF_DZ_TX_TSO_OPTION_TYPE,
208 			    ESE_DZ_TX_TSO_OPTION_DESC_FATSO2A,
209 			    ESF_DZ_TX_TSO_IP_ID, ipv4_id,
210 			    ESF_DZ_TX_TSO_TCP_SEQNO, tcp_seq);
211 	EFX_POPULATE_QWORD_5(txq->txq_hw_ring[(added + 1) & txq->ptr_mask],
212 			    ESF_DZ_TX_DESC_IS_OPT, 1,
213 			    ESF_DZ_TX_OPTION_TYPE,
214 			    ESE_DZ_TX_OPTION_DESC_TSO,
215 			    ESF_DZ_TX_TSO_OPTION_TYPE,
216 			    ESE_DZ_TX_TSO_OPTION_DESC_FATSO2B,
217 			    ESF_DZ_TX_TSO_TCP_MSS, tcp_mss,
218 			    ESF_DZ_TX_TSO_OUTER_IPID, outer_ipv4_id);
219 }
220 
221 static inline void
222 sfc_ef10_tx_qpush(struct sfc_ef10_txq *txq, unsigned int added,
223 		  unsigned int pushed)
224 {
225 	efx_qword_t desc;
226 	efx_oword_t oword;
227 
228 	/*
229 	 * This improves performance by pushing a TX descriptor at the same
230 	 * time as the doorbell. The descriptor must be added to the TXQ,
231 	 * so that can be used if the hardware decides not to use the pushed
232 	 * descriptor.
233 	 */
234 	desc.eq_u64[0] = txq->txq_hw_ring[pushed & txq->ptr_mask].eq_u64[0];
235 	EFX_POPULATE_OWORD_3(oword,
236 		ERF_DZ_TX_DESC_WPTR, added & txq->ptr_mask,
237 		ERF_DZ_TX_DESC_HWORD, EFX_QWORD_FIELD(desc, EFX_DWORD_1),
238 		ERF_DZ_TX_DESC_LWORD, EFX_QWORD_FIELD(desc, EFX_DWORD_0));
239 
240 	/* DMA sync to device is not required */
241 
242 	/*
243 	 * rte_io_wmb() which guarantees that the STORE operations
244 	 * (i.e. Tx and event descriptor updates) that precede
245 	 * the rte_io_wmb() call are visible to NIC before the STORE
246 	 * operations that follow it (i.e. doorbell write).
247 	 */
248 	rte_io_wmb();
249 
250 	*(volatile __m128i *)txq->doorbell = oword.eo_u128[0];
251 }
252 
253 static unsigned int
254 sfc_ef10_tx_pkt_descs_max(const struct rte_mbuf *m)
255 {
256 	unsigned int extra_descs_per_seg;
257 	unsigned int extra_descs_per_pkt;
258 
259 	/*
260 	 * VLAN offload is not supported yet, so no extra descriptors
261 	 * are required for VLAN option descriptor.
262 	 */
263 
264 /** Maximum length of the mbuf segment data */
265 #define SFC_MBUF_SEG_LEN_MAX		UINT16_MAX
266 	RTE_BUILD_BUG_ON(sizeof(m->data_len) != 2);
267 
268 	/*
269 	 * Each segment is already counted once below.  So, calculate
270 	 * how many extra DMA descriptors may be required per segment in
271 	 * the worst case because of maximum DMA descriptor length limit.
272 	 * If maximum segment length is less or equal to maximum DMA
273 	 * descriptor length, no extra DMA descriptors are required.
274 	 */
275 	extra_descs_per_seg =
276 		(SFC_MBUF_SEG_LEN_MAX - 1) / SFC_EF10_TX_DMA_DESC_LEN_MAX;
277 
278 /** Maximum length of the packet */
279 #define SFC_MBUF_PKT_LEN_MAX		UINT32_MAX
280 	RTE_BUILD_BUG_ON(sizeof(m->pkt_len) != 4);
281 
282 	/*
283 	 * One more limitation on maximum number of extra DMA descriptors
284 	 * comes from slicing entire packet because of DMA descriptor length
285 	 * limit taking into account that there is at least one segment
286 	 * which is already counted below (so division of the maximum
287 	 * packet length minus one with round down).
288 	 * TSO is not supported yet, so packet length is limited by
289 	 * maximum PDU size.
290 	 */
291 	extra_descs_per_pkt =
292 		(RTE_MIN((unsigned int)EFX_MAC_PDU_MAX,
293 			 SFC_MBUF_PKT_LEN_MAX) - 1) /
294 		SFC_EF10_TX_DMA_DESC_LEN_MAX;
295 
296 	return m->nb_segs + RTE_MIN(m->nb_segs * extra_descs_per_seg,
297 				    extra_descs_per_pkt);
298 }
299 
300 static bool
301 sfc_ef10_try_reap(struct sfc_ef10_txq * const txq, unsigned int added,
302 		  unsigned int needed_desc, unsigned int *dma_desc_space,
303 		  bool *reap_done)
304 {
305 	if (*reap_done)
306 		return false;
307 
308 	if (added != txq->added) {
309 		sfc_ef10_tx_qpush(txq, added, txq->added);
310 		txq->added = added;
311 	}
312 
313 	sfc_ef10_tx_reap(txq);
314 	*reap_done = true;
315 
316 	/*
317 	 * Recalculate DMA descriptor space since Tx reap may change
318 	 * the number of completed descriptors
319 	 */
320 	*dma_desc_space = txq->max_fill_level -
321 		(added - txq->completed);
322 
323 	return (needed_desc <= *dma_desc_space);
324 }
325 
326 static uint16_t
327 sfc_ef10_prepare_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
328 		      uint16_t nb_pkts)
329 {
330 	struct sfc_ef10_txq * const txq = sfc_ef10_txq_by_dp_txq(tx_queue);
331 	uint16_t i;
332 
333 	for (i = 0; i < nb_pkts; i++) {
334 		struct rte_mbuf *m = tx_pkts[i];
335 		int ret;
336 
337 #ifdef RTE_LIBRTE_SFC_EFX_DEBUG
338 		/*
339 		 * In non-TSO case, check that a packet segments do not exceed
340 		 * the size limit. Perform the check in debug mode since MTU
341 		 * more than 9k is not supported, but the limit here is 16k-1.
342 		 */
343 		if (!(m->ol_flags & PKT_TX_TCP_SEG)) {
344 			struct rte_mbuf *m_seg;
345 
346 			for (m_seg = m; m_seg != NULL; m_seg = m_seg->next) {
347 				if (m_seg->data_len >
348 				    SFC_EF10_TX_DMA_DESC_LEN_MAX) {
349 					rte_errno = EINVAL;
350 					break;
351 				}
352 			}
353 		}
354 #endif
355 		ret = sfc_dp_tx_prepare_pkt(m, 0, SFC_TSOH_STD_LEN,
356 				txq->tso_tcp_header_offset_limit,
357 				txq->max_fill_level,
358 				SFC_EF10_TSO_OPT_DESCS_NUM, 0);
359 		if (unlikely(ret != 0)) {
360 			rte_errno = ret;
361 			break;
362 		}
363 	}
364 
365 	return i;
366 }
367 
368 static int
369 sfc_ef10_xmit_tso_pkt(struct sfc_ef10_txq * const txq, struct rte_mbuf *m_seg,
370 		      unsigned int *added, unsigned int *dma_desc_space,
371 		      bool *reap_done)
372 {
373 	size_t iph_off = ((m_seg->ol_flags & PKT_TX_TUNNEL_MASK) ?
374 			  m_seg->outer_l2_len + m_seg->outer_l3_len : 0) +
375 			 m_seg->l2_len;
376 	size_t tcph_off = iph_off + m_seg->l3_len;
377 	size_t header_len = tcph_off + m_seg->l4_len;
378 	/* Offset of the payload in the last segment that contains the header */
379 	size_t in_off = 0;
380 	const struct rte_tcp_hdr *th;
381 	uint16_t packet_id = 0;
382 	uint16_t outer_packet_id = 0;
383 	uint32_t sent_seq;
384 	uint8_t *hdr_addr;
385 	rte_iova_t hdr_iova;
386 	struct rte_mbuf *first_m_seg = m_seg;
387 	unsigned int pkt_start = *added;
388 	unsigned int needed_desc;
389 	struct rte_mbuf *m_seg_to_free_up_to = first_m_seg;
390 	bool eop;
391 
392 	/*
393 	 * Preliminary estimation of required DMA descriptors, including extra
394 	 * descriptor for TSO header that is needed when the header is
395 	 * separated from payload in one segment. It does not include
396 	 * extra descriptors that may appear when a big segment is split across
397 	 * several descriptors.
398 	 */
399 	needed_desc = m_seg->nb_segs +
400 			(unsigned int)SFC_EF10_TSO_OPT_DESCS_NUM +
401 			(unsigned int)SFC_EF10_TSO_HDR_DESCS_NUM;
402 
403 	if (needed_desc > *dma_desc_space &&
404 	    !sfc_ef10_try_reap(txq, pkt_start, needed_desc,
405 			       dma_desc_space, reap_done)) {
406 		/*
407 		 * If a future Tx reap may increase available DMA descriptor
408 		 * space, do not try to send the packet.
409 		 */
410 		if (txq->completed != pkt_start)
411 			return ENOSPC;
412 		/*
413 		 * Do not allow to send packet if the maximum DMA
414 		 * descriptor space is not sufficient to hold TSO
415 		 * descriptors, header descriptor and at least 1
416 		 * segment descriptor.
417 		 */
418 		if (*dma_desc_space < SFC_EF10_TSO_OPT_DESCS_NUM +
419 				SFC_EF10_TSO_HDR_DESCS_NUM + 1)
420 			return EMSGSIZE;
421 	}
422 
423 	/* Check if the header is not fragmented */
424 	if (rte_pktmbuf_data_len(m_seg) >= header_len) {
425 		hdr_addr = rte_pktmbuf_mtod(m_seg, uint8_t *);
426 		hdr_iova = rte_mbuf_data_iova(m_seg);
427 		if (rte_pktmbuf_data_len(m_seg) == header_len) {
428 			/* Cannot send a packet that consists only of header */
429 			if (unlikely(m_seg->next == NULL))
430 				return EMSGSIZE;
431 			/*
432 			 * Associate header mbuf with header descriptor
433 			 * which is located after TSO descriptors.
434 			 */
435 			txq->sw_ring[(pkt_start + SFC_EF10_TSO_OPT_DESCS_NUM) &
436 				     txq->ptr_mask].mbuf = m_seg;
437 			m_seg = m_seg->next;
438 			in_off = 0;
439 
440 			/*
441 			 * If there is no payload offset (payload starts at the
442 			 * beginning of a segment) then an extra descriptor for
443 			 * separated header is not needed.
444 			 */
445 			needed_desc--;
446 		} else {
447 			in_off = header_len;
448 		}
449 	} else {
450 		unsigned int copied_segs;
451 		unsigned int hdr_addr_off = (*added & txq->ptr_mask) *
452 				SFC_TSOH_STD_LEN;
453 
454 		/*
455 		 * Discard a packet if header linearization is needed but
456 		 * the header is too big.
457 		 * Duplicate Tx prepare check here to avoid spoil of
458 		 * memory if Tx prepare is skipped.
459 		 */
460 		if (unlikely(header_len > SFC_TSOH_STD_LEN))
461 			return EMSGSIZE;
462 
463 		hdr_addr = txq->tsoh + hdr_addr_off;
464 		hdr_iova = txq->tsoh_iova + hdr_addr_off;
465 		copied_segs = sfc_tso_prepare_header(hdr_addr, header_len,
466 						     &m_seg, &in_off);
467 
468 		/* Cannot send a packet that consists only of header */
469 		if (unlikely(m_seg == NULL))
470 			return EMSGSIZE;
471 
472 		m_seg_to_free_up_to = m_seg;
473 		/*
474 		 * Reduce the number of needed descriptors by the number of
475 		 * segments that entirely consist of header data.
476 		 */
477 		needed_desc -= copied_segs;
478 
479 		/* Extra descriptor for separated header is not needed */
480 		if (in_off == 0)
481 			needed_desc--;
482 	}
483 
484 	/*
485 	 * Tx prepare has debug-only checks that offload flags are correctly
486 	 * filled in in TSO mbuf. Use zero IPID if there is no IPv4 flag.
487 	 * If the packet is still IPv4, HW will simply start from zero IPID.
488 	 */
489 	if (first_m_seg->ol_flags & PKT_TX_IPV4)
490 		packet_id = sfc_tso_ip4_get_ipid(hdr_addr, iph_off);
491 
492 	if (first_m_seg->ol_flags & PKT_TX_OUTER_IPV4)
493 		outer_packet_id = sfc_tso_ip4_get_ipid(hdr_addr,
494 						first_m_seg->outer_l2_len);
495 
496 	th = (const struct rte_tcp_hdr *)(hdr_addr + tcph_off);
497 	rte_memcpy(&sent_seq, &th->sent_seq, sizeof(uint32_t));
498 	sent_seq = rte_be_to_cpu_32(sent_seq);
499 
500 	sfc_ef10_tx_qdesc_tso2_create(txq, *added, packet_id, outer_packet_id,
501 			sent_seq, first_m_seg->tso_segsz);
502 	(*added) += SFC_EF10_TSO_OPT_DESCS_NUM;
503 
504 	sfc_ef10_tx_qdesc_dma_create(hdr_iova, header_len, false,
505 			&txq->txq_hw_ring[(*added) & txq->ptr_mask]);
506 	(*added)++;
507 
508 	do {
509 		rte_iova_t next_frag = rte_mbuf_data_iova(m_seg);
510 		unsigned int seg_len = rte_pktmbuf_data_len(m_seg);
511 		unsigned int id;
512 
513 		next_frag += in_off;
514 		seg_len -= in_off;
515 		in_off = 0;
516 
517 		do {
518 			rte_iova_t frag_addr = next_frag;
519 			size_t frag_len;
520 
521 			frag_len = RTE_MIN(seg_len,
522 					   SFC_EF10_TX_DMA_DESC_LEN_MAX);
523 
524 			next_frag += frag_len;
525 			seg_len -= frag_len;
526 
527 			eop = (seg_len == 0 && m_seg->next == NULL);
528 
529 			id = (*added) & txq->ptr_mask;
530 			(*added)++;
531 
532 			/*
533 			 * Initially we assume that one DMA descriptor is needed
534 			 * for every segment. When the segment is split across
535 			 * several DMA descriptors, increase the estimation.
536 			 */
537 			needed_desc += (seg_len != 0);
538 
539 			/*
540 			 * When no more descriptors can be added, but not all
541 			 * segments are processed.
542 			 */
543 			if (*added - pkt_start == *dma_desc_space &&
544 			    !eop &&
545 			    !sfc_ef10_try_reap(txq, pkt_start, needed_desc,
546 						dma_desc_space, reap_done)) {
547 				struct rte_mbuf *m;
548 				struct rte_mbuf *m_next;
549 
550 				if (txq->completed != pkt_start) {
551 					unsigned int i;
552 
553 					/*
554 					 * Reset mbuf associations with added
555 					 * descriptors.
556 					 */
557 					for (i = pkt_start; i != *added; i++) {
558 						id = i & txq->ptr_mask;
559 						txq->sw_ring[id].mbuf = NULL;
560 					}
561 					return ENOSPC;
562 				}
563 
564 				/* Free the segments that cannot be sent */
565 				for (m = m_seg->next; m != NULL; m = m_next) {
566 					m_next = m->next;
567 					rte_pktmbuf_free_seg(m);
568 				}
569 				eop = true;
570 				/* Ignore the rest of the segment */
571 				seg_len = 0;
572 			}
573 
574 			sfc_ef10_tx_qdesc_dma_create(frag_addr, frag_len,
575 					eop, &txq->txq_hw_ring[id]);
576 
577 		} while (seg_len != 0);
578 
579 		txq->sw_ring[id].mbuf = m_seg;
580 
581 		m_seg = m_seg->next;
582 	} while (!eop);
583 
584 	/*
585 	 * Free segments which content was entirely copied to the TSO header
586 	 * memory space of Tx queue
587 	 */
588 	for (m_seg = first_m_seg; m_seg != m_seg_to_free_up_to;) {
589 		struct rte_mbuf *seg_to_free = m_seg;
590 
591 		m_seg = m_seg->next;
592 		rte_pktmbuf_free_seg(seg_to_free);
593 	}
594 
595 	return 0;
596 }
597 
598 static uint16_t
599 sfc_ef10_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
600 {
601 	struct sfc_ef10_txq * const txq = sfc_ef10_txq_by_dp_txq(tx_queue);
602 	unsigned int added;
603 	unsigned int dma_desc_space;
604 	bool reap_done;
605 	struct rte_mbuf **pktp;
606 	struct rte_mbuf **pktp_end;
607 
608 	if (unlikely(txq->flags &
609 		     (SFC_EF10_TXQ_NOT_RUNNING | SFC_EF10_TXQ_EXCEPTION)))
610 		return 0;
611 
612 	added = txq->added;
613 	dma_desc_space = txq->max_fill_level - (added - txq->completed);
614 
615 	reap_done = (dma_desc_space < txq->free_thresh);
616 	if (reap_done) {
617 		sfc_ef10_tx_reap(txq);
618 		dma_desc_space = txq->max_fill_level - (added - txq->completed);
619 	}
620 
621 	for (pktp = &tx_pkts[0], pktp_end = &tx_pkts[nb_pkts];
622 	     pktp != pktp_end;
623 	     ++pktp) {
624 		struct rte_mbuf *m_seg = *pktp;
625 		unsigned int pkt_start = added;
626 		uint32_t pkt_len;
627 
628 		if (likely(pktp + 1 != pktp_end))
629 			rte_mbuf_prefetch_part1(pktp[1]);
630 
631 		if (m_seg->ol_flags & PKT_TX_TCP_SEG) {
632 			int rc;
633 
634 			rc = sfc_ef10_xmit_tso_pkt(txq, m_seg, &added,
635 					&dma_desc_space, &reap_done);
636 			if (rc != 0) {
637 				added = pkt_start;
638 
639 				/* Packet can be sent in following xmit calls */
640 				if (likely(rc == ENOSPC))
641 					break;
642 
643 				/*
644 				 * Packet cannot be sent, tell RTE that
645 				 * it is sent, but actually drop it and
646 				 * continue with another packet
647 				 */
648 				rte_pktmbuf_free(*pktp);
649 				continue;
650 			}
651 
652 			goto dma_desc_space_update;
653 		}
654 
655 		if (sfc_ef10_tx_pkt_descs_max(m_seg) > dma_desc_space) {
656 			if (reap_done)
657 				break;
658 
659 			/* Push already prepared descriptors before polling */
660 			if (added != txq->added) {
661 				sfc_ef10_tx_qpush(txq, added, txq->added);
662 				txq->added = added;
663 			}
664 
665 			sfc_ef10_tx_reap(txq);
666 			reap_done = true;
667 			dma_desc_space = txq->max_fill_level -
668 				(added - txq->completed);
669 			if (sfc_ef10_tx_pkt_descs_max(m_seg) > dma_desc_space)
670 				break;
671 		}
672 
673 		pkt_len = m_seg->pkt_len;
674 		do {
675 			rte_iova_t seg_addr = rte_mbuf_data_iova(m_seg);
676 			unsigned int seg_len = rte_pktmbuf_data_len(m_seg);
677 			unsigned int id = added & txq->ptr_mask;
678 
679 			SFC_ASSERT(seg_len <= SFC_EF10_TX_DMA_DESC_LEN_MAX);
680 
681 			pkt_len -= seg_len;
682 
683 			sfc_ef10_tx_qdesc_dma_create(seg_addr,
684 				seg_len, (pkt_len == 0),
685 				&txq->txq_hw_ring[id]);
686 
687 			/*
688 			 * rte_pktmbuf_free() is commonly used in DPDK for
689 			 * recycling packets - the function checks every
690 			 * segment's reference counter and returns the
691 			 * buffer to its pool whenever possible;
692 			 * nevertheless, freeing mbuf segments one by one
693 			 * may entail some performance decline;
694 			 * from this point, sfc_efx_tx_reap() does the same job
695 			 * on its own and frees buffers in bulks (all mbufs
696 			 * within a bulk belong to the same pool);
697 			 * from this perspective, individual segment pointers
698 			 * must be associated with the corresponding SW
699 			 * descriptors independently so that only one loop
700 			 * is sufficient on reap to inspect all the buffers
701 			 */
702 			txq->sw_ring[id].mbuf = m_seg;
703 
704 			++added;
705 
706 		} while ((m_seg = m_seg->next) != 0);
707 
708 dma_desc_space_update:
709 		dma_desc_space -= (added - pkt_start);
710 	}
711 
712 	if (likely(added != txq->added)) {
713 		sfc_ef10_tx_qpush(txq, added, txq->added);
714 		txq->added = added;
715 	}
716 
717 #if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
718 	if (!reap_done)
719 		sfc_ef10_tx_reap(txq);
720 #endif
721 
722 	return pktp - &tx_pkts[0];
723 }
724 
725 static void
726 sfc_ef10_simple_tx_reap(struct sfc_ef10_txq *txq)
727 {
728 	const unsigned int old_read_ptr = txq->evq_read_ptr;
729 	const unsigned int ptr_mask = txq->ptr_mask;
730 	unsigned int completed = txq->completed;
731 	unsigned int pending = completed;
732 
733 	pending += sfc_ef10_tx_process_events(txq);
734 
735 	if (pending != completed) {
736 		struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
737 		unsigned int nb = 0;
738 
739 		do {
740 			struct sfc_ef10_tx_sw_desc *txd;
741 
742 			txd = &txq->sw_ring[completed & ptr_mask];
743 
744 			if (nb == RTE_DIM(bulk)) {
745 				rte_mempool_put_bulk(bulk[0]->pool,
746 						     (void *)bulk, nb);
747 				nb = 0;
748 			}
749 
750 			bulk[nb++] = txd->mbuf;
751 		} while (++completed != pending);
752 
753 		rte_mempool_put_bulk(bulk[0]->pool, (void *)bulk, nb);
754 
755 		txq->completed = completed;
756 	}
757 
758 	sfc_ef10_ev_qclear(txq->evq_hw_ring, ptr_mask, old_read_ptr,
759 			   txq->evq_read_ptr);
760 }
761 
762 #ifdef RTE_LIBRTE_SFC_EFX_DEBUG
763 static uint16_t
764 sfc_ef10_simple_prepare_pkts(__rte_unused void *tx_queue,
765 			     struct rte_mbuf **tx_pkts,
766 			     uint16_t nb_pkts)
767 {
768 	uint16_t i;
769 
770 	for (i = 0; i < nb_pkts; i++) {
771 		struct rte_mbuf *m = tx_pkts[i];
772 		int ret;
773 
774 		ret = rte_validate_tx_offload(m);
775 		if (unlikely(ret != 0)) {
776 			/*
777 			 * Negative error code is returned by
778 			 * rte_validate_tx_offload(), but positive are used
779 			 * inside net/sfc PMD.
780 			 */
781 			SFC_ASSERT(ret < 0);
782 			rte_errno = -ret;
783 			break;
784 		}
785 
786 		/* ef10_simple does not support TSO and VLAN insertion */
787 		if (unlikely(m->ol_flags &
788 			     (PKT_TX_TCP_SEG | PKT_TX_VLAN_PKT))) {
789 			rte_errno = ENOTSUP;
790 			break;
791 		}
792 
793 		/* ef10_simple does not support scattered packets */
794 		if (unlikely(m->nb_segs != 1)) {
795 			rte_errno = ENOTSUP;
796 			break;
797 		}
798 
799 		/*
800 		 * ef10_simple requires fast-free which ignores reference
801 		 * counters
802 		 */
803 		if (unlikely(rte_mbuf_refcnt_read(m) != 1)) {
804 			rte_errno = ENOTSUP;
805 			break;
806 		}
807 
808 		/* ef10_simple requires single pool for all packets */
809 		if (unlikely(m->pool != tx_pkts[0]->pool)) {
810 			rte_errno = ENOTSUP;
811 			break;
812 		}
813 	}
814 
815 	return i;
816 }
817 #endif
818 
819 static uint16_t
820 sfc_ef10_simple_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
821 			  uint16_t nb_pkts)
822 {
823 	struct sfc_ef10_txq * const txq = sfc_ef10_txq_by_dp_txq(tx_queue);
824 	unsigned int ptr_mask;
825 	unsigned int added;
826 	unsigned int dma_desc_space;
827 	bool reap_done;
828 	struct rte_mbuf **pktp;
829 	struct rte_mbuf **pktp_end;
830 
831 	if (unlikely(txq->flags &
832 		     (SFC_EF10_TXQ_NOT_RUNNING | SFC_EF10_TXQ_EXCEPTION)))
833 		return 0;
834 
835 	ptr_mask = txq->ptr_mask;
836 	added = txq->added;
837 	dma_desc_space = txq->max_fill_level - (added - txq->completed);
838 
839 	reap_done = (dma_desc_space < RTE_MAX(txq->free_thresh, nb_pkts));
840 	if (reap_done) {
841 		sfc_ef10_simple_tx_reap(txq);
842 		dma_desc_space = txq->max_fill_level - (added - txq->completed);
843 	}
844 
845 	pktp_end = &tx_pkts[MIN(nb_pkts, dma_desc_space)];
846 	for (pktp = &tx_pkts[0]; pktp != pktp_end; ++pktp) {
847 		struct rte_mbuf *pkt = *pktp;
848 		unsigned int id = added & ptr_mask;
849 
850 		SFC_ASSERT(rte_pktmbuf_data_len(pkt) <=
851 			   SFC_EF10_TX_DMA_DESC_LEN_MAX);
852 
853 		sfc_ef10_tx_qdesc_dma_create(rte_mbuf_data_iova(pkt),
854 					     rte_pktmbuf_data_len(pkt),
855 					     true, &txq->txq_hw_ring[id]);
856 
857 		txq->sw_ring[id].mbuf = pkt;
858 
859 		++added;
860 	}
861 
862 	if (likely(added != txq->added)) {
863 		sfc_ef10_tx_qpush(txq, added, txq->added);
864 		txq->added = added;
865 	}
866 
867 #if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
868 	if (!reap_done)
869 		sfc_ef10_simple_tx_reap(txq);
870 #endif
871 
872 	return pktp - &tx_pkts[0];
873 }
874 
875 static sfc_dp_tx_get_dev_info_t sfc_ef10_get_dev_info;
876 static void
877 sfc_ef10_get_dev_info(struct rte_eth_dev_info *dev_info)
878 {
879 	/*
880 	 * Number of descriptors just defines maximum number of pushed
881 	 * descriptors (fill level).
882 	 */
883 	dev_info->tx_desc_lim.nb_min = 1;
884 	dev_info->tx_desc_lim.nb_align = 1;
885 }
886 
887 static sfc_dp_tx_qsize_up_rings_t sfc_ef10_tx_qsize_up_rings;
888 static int
889 sfc_ef10_tx_qsize_up_rings(uint16_t nb_tx_desc,
890 			   struct sfc_dp_tx_hw_limits *limits,
891 			   unsigned int *txq_entries,
892 			   unsigned int *evq_entries,
893 			   unsigned int *txq_max_fill_level)
894 {
895 	/*
896 	 * rte_ethdev API guarantees that the number meets min, max and
897 	 * alignment requirements.
898 	 */
899 	if (nb_tx_desc <= limits->txq_min_entries)
900 		*txq_entries = limits->txq_min_entries;
901 	else
902 		*txq_entries = rte_align32pow2(nb_tx_desc);
903 
904 	*evq_entries = *txq_entries;
905 
906 	*txq_max_fill_level = RTE_MIN(nb_tx_desc,
907 				      SFC_EF10_TXQ_LIMIT(*evq_entries));
908 	return 0;
909 }
910 
911 static sfc_dp_tx_qcreate_t sfc_ef10_tx_qcreate;
912 static int
913 sfc_ef10_tx_qcreate(uint16_t port_id, uint16_t queue_id,
914 		    const struct rte_pci_addr *pci_addr, int socket_id,
915 		    const struct sfc_dp_tx_qcreate_info *info,
916 		    struct sfc_dp_txq **dp_txqp)
917 {
918 	struct sfc_ef10_txq *txq;
919 	int rc;
920 
921 	rc = EINVAL;
922 	if (info->txq_entries != info->evq_entries)
923 		goto fail_bad_args;
924 
925 	rc = ENOMEM;
926 	txq = rte_zmalloc_socket("sfc-ef10-txq", sizeof(*txq),
927 				 RTE_CACHE_LINE_SIZE, socket_id);
928 	if (txq == NULL)
929 		goto fail_txq_alloc;
930 
931 	sfc_dp_queue_init(&txq->dp.dpq, port_id, queue_id, pci_addr);
932 
933 	rc = ENOMEM;
934 	txq->sw_ring = rte_calloc_socket("sfc-ef10-txq-sw_ring",
935 					 info->txq_entries,
936 					 sizeof(*txq->sw_ring),
937 					 RTE_CACHE_LINE_SIZE, socket_id);
938 	if (txq->sw_ring == NULL)
939 		goto fail_sw_ring_alloc;
940 
941 	if (info->offloads & (DEV_TX_OFFLOAD_TCP_TSO |
942 			      DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
943 			      DEV_TX_OFFLOAD_GENEVE_TNL_TSO)) {
944 		txq->tsoh = rte_calloc_socket("sfc-ef10-txq-tsoh",
945 					      info->txq_entries,
946 					      SFC_TSOH_STD_LEN,
947 					      RTE_CACHE_LINE_SIZE,
948 					      socket_id);
949 		if (txq->tsoh == NULL)
950 			goto fail_tsoh_alloc;
951 
952 		txq->tsoh_iova = rte_malloc_virt2iova(txq->tsoh);
953 	}
954 
955 	txq->flags = SFC_EF10_TXQ_NOT_RUNNING;
956 	txq->ptr_mask = info->txq_entries - 1;
957 	txq->max_fill_level = info->max_fill_level;
958 	txq->free_thresh = info->free_thresh;
959 	txq->txq_hw_ring = info->txq_hw_ring;
960 	txq->doorbell = (volatile uint8_t *)info->mem_bar +
961 			ER_DZ_TX_DESC_UPD_REG_OFST +
962 			(info->hw_index << info->vi_window_shift);
963 	txq->evq_hw_ring = info->evq_hw_ring;
964 	txq->tso_tcp_header_offset_limit = info->tso_tcp_header_offset_limit;
965 
966 	sfc_ef10_tx_info(&txq->dp.dpq, "TxQ doorbell is %p", txq->doorbell);
967 
968 	*dp_txqp = &txq->dp;
969 	return 0;
970 
971 fail_tsoh_alloc:
972 	rte_free(txq->sw_ring);
973 
974 fail_sw_ring_alloc:
975 	rte_free(txq);
976 
977 fail_txq_alloc:
978 fail_bad_args:
979 	return rc;
980 }
981 
982 static sfc_dp_tx_qdestroy_t sfc_ef10_tx_qdestroy;
983 static void
984 sfc_ef10_tx_qdestroy(struct sfc_dp_txq *dp_txq)
985 {
986 	struct sfc_ef10_txq *txq = sfc_ef10_txq_by_dp_txq(dp_txq);
987 
988 	rte_free(txq->tsoh);
989 	rte_free(txq->sw_ring);
990 	rte_free(txq);
991 }
992 
993 static sfc_dp_tx_qstart_t sfc_ef10_tx_qstart;
994 static int
995 sfc_ef10_tx_qstart(struct sfc_dp_txq *dp_txq, unsigned int evq_read_ptr,
996 		   unsigned int txq_desc_index)
997 {
998 	struct sfc_ef10_txq *txq = sfc_ef10_txq_by_dp_txq(dp_txq);
999 
1000 	txq->evq_read_ptr = evq_read_ptr;
1001 	txq->added = txq->completed = txq_desc_index;
1002 
1003 	txq->flags |= SFC_EF10_TXQ_STARTED;
1004 	txq->flags &= ~(SFC_EF10_TXQ_NOT_RUNNING | SFC_EF10_TXQ_EXCEPTION);
1005 
1006 	return 0;
1007 }
1008 
1009 static sfc_dp_tx_qstop_t sfc_ef10_tx_qstop;
1010 static void
1011 sfc_ef10_tx_qstop(struct sfc_dp_txq *dp_txq, unsigned int *evq_read_ptr)
1012 {
1013 	struct sfc_ef10_txq *txq = sfc_ef10_txq_by_dp_txq(dp_txq);
1014 
1015 	txq->flags |= SFC_EF10_TXQ_NOT_RUNNING;
1016 
1017 	*evq_read_ptr = txq->evq_read_ptr;
1018 }
1019 
1020 static sfc_dp_tx_qtx_ev_t sfc_ef10_tx_qtx_ev;
1021 static bool
1022 sfc_ef10_tx_qtx_ev(struct sfc_dp_txq *dp_txq, __rte_unused unsigned int id)
1023 {
1024 	__rte_unused struct sfc_ef10_txq *txq = sfc_ef10_txq_by_dp_txq(dp_txq);
1025 
1026 	SFC_ASSERT(txq->flags & SFC_EF10_TXQ_NOT_RUNNING);
1027 
1028 	/*
1029 	 * It is safe to ignore Tx event since we reap all mbufs on
1030 	 * queue purge anyway.
1031 	 */
1032 
1033 	return false;
1034 }
1035 
1036 static sfc_dp_tx_qreap_t sfc_ef10_tx_qreap;
1037 static void
1038 sfc_ef10_tx_qreap(struct sfc_dp_txq *dp_txq)
1039 {
1040 	struct sfc_ef10_txq *txq = sfc_ef10_txq_by_dp_txq(dp_txq);
1041 	unsigned int completed;
1042 
1043 	for (completed = txq->completed; completed != txq->added; ++completed) {
1044 		struct sfc_ef10_tx_sw_desc *txd;
1045 
1046 		txd = &txq->sw_ring[completed & txq->ptr_mask];
1047 		if (txd->mbuf != NULL) {
1048 			rte_pktmbuf_free_seg(txd->mbuf);
1049 			txd->mbuf = NULL;
1050 		}
1051 	}
1052 
1053 	txq->flags &= ~SFC_EF10_TXQ_STARTED;
1054 }
1055 
1056 static unsigned int
1057 sfc_ef10_tx_qdesc_npending(struct sfc_ef10_txq *txq)
1058 {
1059 	const unsigned int curr_done = txq->completed - 1;
1060 	unsigned int anew_done = curr_done;
1061 	efx_qword_t tx_ev;
1062 	const unsigned int evq_old_read_ptr = txq->evq_read_ptr;
1063 
1064 	if (unlikely(txq->flags &
1065 		     (SFC_EF10_TXQ_NOT_RUNNING | SFC_EF10_TXQ_EXCEPTION)))
1066 		return 0;
1067 
1068 	while (sfc_ef10_tx_get_event(txq, &tx_ev))
1069 		anew_done = EFX_QWORD_FIELD(tx_ev, ESF_DZ_TX_DESCR_INDX);
1070 
1071 	/*
1072 	 * The function does not process events, so return event queue read
1073 	 * pointer to the original position to allow the events that were
1074 	 * read to be processed later
1075 	 */
1076 	txq->evq_read_ptr = evq_old_read_ptr;
1077 
1078 	return (anew_done - curr_done) & txq->ptr_mask;
1079 }
1080 
1081 static sfc_dp_tx_qdesc_status_t sfc_ef10_tx_qdesc_status;
1082 static int
1083 sfc_ef10_tx_qdesc_status(struct sfc_dp_txq *dp_txq,
1084 			 uint16_t offset)
1085 {
1086 	struct sfc_ef10_txq *txq = sfc_ef10_txq_by_dp_txq(dp_txq);
1087 	unsigned int npending = sfc_ef10_tx_qdesc_npending(txq);
1088 
1089 	if (unlikely(offset > txq->ptr_mask))
1090 		return -EINVAL;
1091 
1092 	if (unlikely(offset >= txq->max_fill_level))
1093 		return RTE_ETH_TX_DESC_UNAVAIL;
1094 
1095 	if (unlikely(offset < npending))
1096 		return RTE_ETH_TX_DESC_FULL;
1097 
1098 	return RTE_ETH_TX_DESC_DONE;
1099 }
1100 
1101 struct sfc_dp_tx sfc_ef10_tx = {
1102 	.dp = {
1103 		.name		= SFC_KVARG_DATAPATH_EF10,
1104 		.type		= SFC_DP_TX,
1105 		.hw_fw_caps	= SFC_DP_HW_FW_CAP_EF10,
1106 	},
1107 	.features		= SFC_DP_TX_FEAT_MULTI_PROCESS,
1108 	.dev_offload_capa	= DEV_TX_OFFLOAD_MULTI_SEGS,
1109 	.queue_offload_capa	= DEV_TX_OFFLOAD_IPV4_CKSUM |
1110 				  DEV_TX_OFFLOAD_UDP_CKSUM |
1111 				  DEV_TX_OFFLOAD_TCP_CKSUM |
1112 				  DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM |
1113 				  DEV_TX_OFFLOAD_TCP_TSO |
1114 				  DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1115 				  DEV_TX_OFFLOAD_GENEVE_TNL_TSO,
1116 	.get_dev_info		= sfc_ef10_get_dev_info,
1117 	.qsize_up_rings		= sfc_ef10_tx_qsize_up_rings,
1118 	.qcreate		= sfc_ef10_tx_qcreate,
1119 	.qdestroy		= sfc_ef10_tx_qdestroy,
1120 	.qstart			= sfc_ef10_tx_qstart,
1121 	.qtx_ev			= sfc_ef10_tx_qtx_ev,
1122 	.qstop			= sfc_ef10_tx_qstop,
1123 	.qreap			= sfc_ef10_tx_qreap,
1124 	.qdesc_status		= sfc_ef10_tx_qdesc_status,
1125 	.pkt_prepare		= sfc_ef10_prepare_pkts,
1126 	.pkt_burst		= sfc_ef10_xmit_pkts,
1127 };
1128 
1129 struct sfc_dp_tx sfc_ef10_simple_tx = {
1130 	.dp = {
1131 		.name		= SFC_KVARG_DATAPATH_EF10_SIMPLE,
1132 		.type		= SFC_DP_TX,
1133 	},
1134 	.features		= SFC_DP_TX_FEAT_MULTI_PROCESS,
1135 	.dev_offload_capa	= DEV_TX_OFFLOAD_MBUF_FAST_FREE,
1136 	.queue_offload_capa	= DEV_TX_OFFLOAD_IPV4_CKSUM |
1137 				  DEV_TX_OFFLOAD_UDP_CKSUM |
1138 				  DEV_TX_OFFLOAD_TCP_CKSUM |
1139 				  DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM,
1140 	.get_dev_info		= sfc_ef10_get_dev_info,
1141 	.qsize_up_rings		= sfc_ef10_tx_qsize_up_rings,
1142 	.qcreate		= sfc_ef10_tx_qcreate,
1143 	.qdestroy		= sfc_ef10_tx_qdestroy,
1144 	.qstart			= sfc_ef10_tx_qstart,
1145 	.qtx_ev			= sfc_ef10_tx_qtx_ev,
1146 	.qstop			= sfc_ef10_tx_qstop,
1147 	.qreap			= sfc_ef10_tx_qreap,
1148 	.qdesc_status		= sfc_ef10_tx_qdesc_status,
1149 #ifdef RTE_LIBRTE_SFC_EFX_DEBUG
1150 	.pkt_prepare		= sfc_ef10_simple_prepare_pkts,
1151 #endif
1152 	.pkt_burst		= sfc_ef10_simple_xmit_pkts,
1153 };
1154