1 // SPDX-License-Identifier: GPL-2.0
2 /* Marvell RVU Ethernet driver
3  *
4  * Copyright (C) 2020 Marvell.
5  *
6  */
7 
8 #include <linux/etherdevice.h>
9 #include <net/ip.h>
10 #include <net/tso.h>
11 #include <linux/bpf.h>
12 #include <linux/bpf_trace.h>
13 
14 #include "otx2_reg.h"
15 #include "otx2_common.h"
16 #include "otx2_struct.h"
17 #include "otx2_txrx.h"
18 #include "otx2_ptp.h"
19 #include "cn10k.h"
20 
21 #define CQE_ADDR(CQ, idx) ((CQ)->cqe_base + ((CQ)->cqe_size * (idx)))
22 #define PTP_PORT	        0x13F
23 /* PTPv2 header Original Timestamp starts at byte offset 34 and
24  * contains 6 byte seconds field and 4 byte nano seconds field.
25  */
26 #define PTP_SYNC_SEC_OFFSET	34
27 
28 static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
29 				     struct bpf_prog *prog,
30 				     struct nix_cqe_rx_s *cqe,
31 				     struct otx2_cq_queue *cq);
32 
33 static int otx2_nix_cq_op_status(struct otx2_nic *pfvf,
34 				 struct otx2_cq_queue *cq)
35 {
36 	u64 incr = (u64)(cq->cq_idx) << 32;
37 	u64 status;
38 
39 	status = otx2_atomic64_fetch_add(incr, pfvf->cq_op_addr);
40 
41 	if (unlikely(status & BIT_ULL(CQ_OP_STAT_OP_ERR) ||
42 		     status & BIT_ULL(CQ_OP_STAT_CQ_ERR))) {
43 		dev_err(pfvf->dev, "CQ stopped due to error");
44 		return -EINVAL;
45 	}
46 
47 	cq->cq_tail = status & 0xFFFFF;
48 	cq->cq_head = (status >> 20) & 0xFFFFF;
49 	if (cq->cq_tail < cq->cq_head)
50 		cq->pend_cqe = (cq->cqe_cnt - cq->cq_head) +
51 				cq->cq_tail;
52 	else
53 		cq->pend_cqe = cq->cq_tail - cq->cq_head;
54 
55 	return 0;
56 }
57 
58 static struct nix_cqe_hdr_s *otx2_get_next_cqe(struct otx2_cq_queue *cq)
59 {
60 	struct nix_cqe_hdr_s *cqe_hdr;
61 
62 	cqe_hdr = (struct nix_cqe_hdr_s *)CQE_ADDR(cq, cq->cq_head);
63 	if (cqe_hdr->cqe_type == NIX_XQE_TYPE_INVALID)
64 		return NULL;
65 
66 	cq->cq_head++;
67 	cq->cq_head &= (cq->cqe_cnt - 1);
68 
69 	return cqe_hdr;
70 }
71 
72 static unsigned int frag_num(unsigned int i)
73 {
74 #ifdef __BIG_ENDIAN
75 	return (i & ~3) + 3 - (i & 3);
76 #else
77 	return i;
78 #endif
79 }
80 
81 static dma_addr_t otx2_dma_map_skb_frag(struct otx2_nic *pfvf,
82 					struct sk_buff *skb, int seg, int *len)
83 {
84 	const skb_frag_t *frag;
85 	struct page *page;
86 	int offset;
87 
88 	/* First segment is always skb->data */
89 	if (!seg) {
90 		page = virt_to_page(skb->data);
91 		offset = offset_in_page(skb->data);
92 		*len = skb_headlen(skb);
93 	} else {
94 		frag = &skb_shinfo(skb)->frags[seg - 1];
95 		page = skb_frag_page(frag);
96 		offset = skb_frag_off(frag);
97 		*len = skb_frag_size(frag);
98 	}
99 	return otx2_dma_map_page(pfvf, page, offset, *len, DMA_TO_DEVICE);
100 }
101 
102 static void otx2_dma_unmap_skb_frags(struct otx2_nic *pfvf, struct sg_list *sg)
103 {
104 	int seg;
105 
106 	for (seg = 0; seg < sg->num_segs; seg++) {
107 		otx2_dma_unmap_page(pfvf, sg->dma_addr[seg],
108 				    sg->size[seg], DMA_TO_DEVICE);
109 	}
110 	sg->num_segs = 0;
111 }
112 
113 static void otx2_xdp_snd_pkt_handler(struct otx2_nic *pfvf,
114 				     struct otx2_snd_queue *sq,
115 				 struct nix_cqe_tx_s *cqe)
116 {
117 	struct nix_send_comp_s *snd_comp = &cqe->comp;
118 	struct sg_list *sg;
119 	struct page *page;
120 	u64 pa;
121 
122 	sg = &sq->sg[snd_comp->sqe_id];
123 
124 	pa = otx2_iova_to_phys(pfvf->iommu_domain, sg->dma_addr[0]);
125 	otx2_dma_unmap_page(pfvf, sg->dma_addr[0],
126 			    sg->size[0], DMA_TO_DEVICE);
127 	page = virt_to_page(phys_to_virt(pa));
128 	put_page(page);
129 }
130 
131 static void otx2_snd_pkt_handler(struct otx2_nic *pfvf,
132 				 struct otx2_cq_queue *cq,
133 				 struct otx2_snd_queue *sq,
134 				 struct nix_cqe_tx_s *cqe,
135 				 int budget, int *tx_pkts, int *tx_bytes)
136 {
137 	struct nix_send_comp_s *snd_comp = &cqe->comp;
138 	struct skb_shared_hwtstamps ts;
139 	struct sk_buff *skb = NULL;
140 	u64 timestamp, tsns;
141 	struct sg_list *sg;
142 	int err;
143 
144 	if (unlikely(snd_comp->status) && netif_msg_tx_err(pfvf))
145 		net_err_ratelimited("%s: TX%d: Error in send CQ status:%x\n",
146 				    pfvf->netdev->name, cq->cint_idx,
147 				    snd_comp->status);
148 
149 	sg = &sq->sg[snd_comp->sqe_id];
150 	skb = (struct sk_buff *)sg->skb;
151 	if (unlikely(!skb))
152 		return;
153 
154 	if (skb_shinfo(skb)->tx_flags & SKBTX_IN_PROGRESS) {
155 		timestamp = ((u64 *)sq->timestamps->base)[snd_comp->sqe_id];
156 		if (timestamp != 1) {
157 			timestamp = pfvf->ptp->convert_tx_ptp_tstmp(timestamp);
158 			err = otx2_ptp_tstamp2time(pfvf, timestamp, &tsns);
159 			if (!err) {
160 				memset(&ts, 0, sizeof(ts));
161 				ts.hwtstamp = ns_to_ktime(tsns);
162 				skb_tstamp_tx(skb, &ts);
163 			}
164 		}
165 	}
166 
167 	*tx_bytes += skb->len;
168 	(*tx_pkts)++;
169 	otx2_dma_unmap_skb_frags(pfvf, sg);
170 	napi_consume_skb(skb, budget);
171 	sg->skb = (u64)NULL;
172 }
173 
174 static void otx2_set_rxtstamp(struct otx2_nic *pfvf,
175 			      struct sk_buff *skb, void *data)
176 {
177 	u64 timestamp, tsns;
178 	int err;
179 
180 	if (!(pfvf->flags & OTX2_FLAG_RX_TSTAMP_ENABLED))
181 		return;
182 
183 	timestamp = pfvf->ptp->convert_rx_ptp_tstmp(*(u64 *)data);
184 	/* The first 8 bytes is the timestamp */
185 	err = otx2_ptp_tstamp2time(pfvf, timestamp, &tsns);
186 	if (err)
187 		return;
188 
189 	skb_hwtstamps(skb)->hwtstamp = ns_to_ktime(tsns);
190 }
191 
192 static bool otx2_skb_add_frag(struct otx2_nic *pfvf, struct sk_buff *skb,
193 			      u64 iova, int len, struct nix_rx_parse_s *parse,
194 			      int qidx)
195 {
196 	struct page *page;
197 	int off = 0;
198 	void *va;
199 
200 	va = phys_to_virt(otx2_iova_to_phys(pfvf->iommu_domain, iova));
201 
202 	if (likely(!skb_shinfo(skb)->nr_frags)) {
203 		/* Check if data starts at some nonzero offset
204 		 * from the start of the buffer.  For now the
205 		 * only possible offset is 8 bytes in the case
206 		 * where packet is prepended by a timestamp.
207 		 */
208 		if (parse->laptr) {
209 			otx2_set_rxtstamp(pfvf, skb, va);
210 			off = OTX2_HW_TIMESTAMP_LEN;
211 		}
212 	}
213 
214 	page = virt_to_page(va);
215 	if (likely(skb_shinfo(skb)->nr_frags < MAX_SKB_FRAGS)) {
216 		skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page,
217 				va - page_address(page) + off,
218 				len - off, pfvf->rbsize);
219 
220 		otx2_dma_unmap_page(pfvf, iova - OTX2_HEAD_ROOM,
221 				    pfvf->rbsize, DMA_FROM_DEVICE);
222 		return true;
223 	}
224 
225 	/* If more than MAX_SKB_FRAGS fragments are received then
226 	 * give back those buffer pointers to hardware for reuse.
227 	 */
228 	pfvf->hw_ops->aura_freeptr(pfvf, qidx, iova & ~0x07ULL);
229 
230 	return false;
231 }
232 
233 static void otx2_set_rxhash(struct otx2_nic *pfvf,
234 			    struct nix_cqe_rx_s *cqe, struct sk_buff *skb)
235 {
236 	enum pkt_hash_types hash_type = PKT_HASH_TYPE_NONE;
237 	struct otx2_rss_info *rss;
238 	u32 hash = 0;
239 
240 	if (!(pfvf->netdev->features & NETIF_F_RXHASH))
241 		return;
242 
243 	rss = &pfvf->hw.rss_info;
244 	if (rss->flowkey_cfg) {
245 		if (rss->flowkey_cfg &
246 		    ~(NIX_FLOW_KEY_TYPE_IPV4 | NIX_FLOW_KEY_TYPE_IPV6))
247 			hash_type = PKT_HASH_TYPE_L4;
248 		else
249 			hash_type = PKT_HASH_TYPE_L3;
250 		hash = cqe->hdr.flow_tag;
251 	}
252 	skb_set_hash(skb, hash, hash_type);
253 }
254 
255 static void otx2_free_rcv_seg(struct otx2_nic *pfvf, struct nix_cqe_rx_s *cqe,
256 			      int qidx)
257 {
258 	struct nix_rx_sg_s *sg = &cqe->sg;
259 	void *end, *start;
260 	u64 *seg_addr;
261 	int seg;
262 
263 	start = (void *)sg;
264 	end = start + ((cqe->parse.desc_sizem1 + 1) * 16);
265 	while (start < end) {
266 		sg = (struct nix_rx_sg_s *)start;
267 		seg_addr = &sg->seg_addr;
268 		for (seg = 0; seg < sg->segs; seg++, seg_addr++)
269 			pfvf->hw_ops->aura_freeptr(pfvf, qidx,
270 						   *seg_addr & ~0x07ULL);
271 		start += sizeof(*sg);
272 	}
273 }
274 
275 static bool otx2_check_rcv_errors(struct otx2_nic *pfvf,
276 				  struct nix_cqe_rx_s *cqe, int qidx)
277 {
278 	struct otx2_drv_stats *stats = &pfvf->hw.drv_stats;
279 	struct nix_rx_parse_s *parse = &cqe->parse;
280 
281 	if (netif_msg_rx_err(pfvf))
282 		netdev_err(pfvf->netdev,
283 			   "RQ%d: Error pkt with errlev:0x%x errcode:0x%x\n",
284 			   qidx, parse->errlev, parse->errcode);
285 
286 	if (parse->errlev == NPC_ERRLVL_RE) {
287 		switch (parse->errcode) {
288 		case ERRCODE_FCS:
289 		case ERRCODE_FCS_RCV:
290 			atomic_inc(&stats->rx_fcs_errs);
291 			break;
292 		case ERRCODE_UNDERSIZE:
293 			atomic_inc(&stats->rx_undersize_errs);
294 			break;
295 		case ERRCODE_OVERSIZE:
296 			atomic_inc(&stats->rx_oversize_errs);
297 			break;
298 		case ERRCODE_OL2_LEN_MISMATCH:
299 			atomic_inc(&stats->rx_len_errs);
300 			break;
301 		default:
302 			atomic_inc(&stats->rx_other_errs);
303 			break;
304 		}
305 	} else if (parse->errlev == NPC_ERRLVL_NIX) {
306 		switch (parse->errcode) {
307 		case ERRCODE_OL3_LEN:
308 		case ERRCODE_OL4_LEN:
309 		case ERRCODE_IL3_LEN:
310 		case ERRCODE_IL4_LEN:
311 			atomic_inc(&stats->rx_len_errs);
312 			break;
313 		case ERRCODE_OL4_CSUM:
314 		case ERRCODE_IL4_CSUM:
315 			atomic_inc(&stats->rx_csum_errs);
316 			break;
317 		default:
318 			atomic_inc(&stats->rx_other_errs);
319 			break;
320 		}
321 	} else {
322 		atomic_inc(&stats->rx_other_errs);
323 		/* For now ignore all the NPC parser errors and
324 		 * pass the packets to stack.
325 		 */
326 		return false;
327 	}
328 
329 	/* If RXALL is enabled pass on packets to stack. */
330 	if (pfvf->netdev->features & NETIF_F_RXALL)
331 		return false;
332 
333 	/* Free buffer back to pool */
334 	if (cqe->sg.segs)
335 		otx2_free_rcv_seg(pfvf, cqe, qidx);
336 	return true;
337 }
338 
339 static void otx2_rcv_pkt_handler(struct otx2_nic *pfvf,
340 				 struct napi_struct *napi,
341 				 struct otx2_cq_queue *cq,
342 				 struct nix_cqe_rx_s *cqe)
343 {
344 	struct nix_rx_parse_s *parse = &cqe->parse;
345 	struct nix_rx_sg_s *sg = &cqe->sg;
346 	struct sk_buff *skb = NULL;
347 	void *end, *start;
348 	u64 *seg_addr;
349 	u16 *seg_size;
350 	int seg;
351 
352 	if (unlikely(parse->errlev || parse->errcode)) {
353 		if (otx2_check_rcv_errors(pfvf, cqe, cq->cq_idx))
354 			return;
355 	}
356 
357 	if (pfvf->xdp_prog)
358 		if (otx2_xdp_rcv_pkt_handler(pfvf, pfvf->xdp_prog, cqe, cq))
359 			return;
360 
361 	skb = napi_get_frags(napi);
362 	if (unlikely(!skb))
363 		return;
364 
365 	start = (void *)sg;
366 	end = start + ((cqe->parse.desc_sizem1 + 1) * 16);
367 	while (start < end) {
368 		sg = (struct nix_rx_sg_s *)start;
369 		seg_addr = &sg->seg_addr;
370 		seg_size = (void *)sg;
371 		for (seg = 0; seg < sg->segs; seg++, seg_addr++) {
372 			if (otx2_skb_add_frag(pfvf, skb, *seg_addr,
373 					      seg_size[seg], parse, cq->cq_idx))
374 				cq->pool_ptrs++;
375 		}
376 		start += sizeof(*sg);
377 	}
378 	otx2_set_rxhash(pfvf, cqe, skb);
379 
380 	skb_record_rx_queue(skb, cq->cq_idx);
381 	if (pfvf->netdev->features & NETIF_F_RXCSUM)
382 		skb->ip_summed = CHECKSUM_UNNECESSARY;
383 
384 	napi_gro_frags(napi);
385 }
386 
387 static int otx2_rx_napi_handler(struct otx2_nic *pfvf,
388 				struct napi_struct *napi,
389 				struct otx2_cq_queue *cq, int budget)
390 {
391 	struct nix_cqe_rx_s *cqe;
392 	int processed_cqe = 0;
393 
394 	if (cq->pend_cqe >= budget)
395 		goto process_cqe;
396 
397 	if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe)
398 		return 0;
399 
400 process_cqe:
401 	while (likely(processed_cqe < budget) && cq->pend_cqe) {
402 		cqe = (struct nix_cqe_rx_s *)CQE_ADDR(cq, cq->cq_head);
403 		if (cqe->hdr.cqe_type == NIX_XQE_TYPE_INVALID ||
404 		    !cqe->sg.seg_addr) {
405 			if (!processed_cqe)
406 				return 0;
407 			break;
408 		}
409 		cq->cq_head++;
410 		cq->cq_head &= (cq->cqe_cnt - 1);
411 
412 		otx2_rcv_pkt_handler(pfvf, napi, cq, cqe);
413 
414 		cqe->hdr.cqe_type = NIX_XQE_TYPE_INVALID;
415 		cqe->sg.seg_addr = 0x00;
416 		processed_cqe++;
417 		cq->pend_cqe--;
418 	}
419 
420 	/* Free CQEs to HW */
421 	otx2_write64(pfvf, NIX_LF_CQ_OP_DOOR,
422 		     ((u64)cq->cq_idx << 32) | processed_cqe);
423 
424 	return processed_cqe;
425 }
426 
427 void otx2_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq)
428 {
429 	struct otx2_nic *pfvf = dev;
430 	dma_addr_t bufptr;
431 
432 	while (cq->pool_ptrs) {
433 		if (otx2_alloc_buffer(pfvf, cq, &bufptr))
434 			break;
435 		otx2_aura_freeptr(pfvf, cq->cq_idx, bufptr + OTX2_HEAD_ROOM);
436 		cq->pool_ptrs--;
437 	}
438 }
439 
440 static int otx2_tx_napi_handler(struct otx2_nic *pfvf,
441 				struct otx2_cq_queue *cq, int budget)
442 {
443 	int tx_pkts = 0, tx_bytes = 0, qidx;
444 	struct otx2_snd_queue *sq;
445 	struct nix_cqe_tx_s *cqe;
446 	int processed_cqe = 0;
447 
448 	if (cq->pend_cqe >= budget)
449 		goto process_cqe;
450 
451 	if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe)
452 		return 0;
453 
454 process_cqe:
455 	qidx = cq->cq_idx - pfvf->hw.rx_queues;
456 	sq = &pfvf->qset.sq[qidx];
457 
458 	while (likely(processed_cqe < budget) && cq->pend_cqe) {
459 		cqe = (struct nix_cqe_tx_s *)otx2_get_next_cqe(cq);
460 		if (unlikely(!cqe)) {
461 			if (!processed_cqe)
462 				return 0;
463 			break;
464 		}
465 
466 		if (cq->cq_type == CQ_XDP) {
467 			otx2_xdp_snd_pkt_handler(pfvf, sq, cqe);
468 		} else {
469 			otx2_snd_pkt_handler(pfvf, cq, sq, cqe, budget,
470 					     &tx_pkts, &tx_bytes);
471 		}
472 
473 		cqe->hdr.cqe_type = NIX_XQE_TYPE_INVALID;
474 		processed_cqe++;
475 		cq->pend_cqe--;
476 
477 		sq->cons_head++;
478 		sq->cons_head &= (sq->sqe_cnt - 1);
479 	}
480 
481 	/* Free CQEs to HW */
482 	otx2_write64(pfvf, NIX_LF_CQ_OP_DOOR,
483 		     ((u64)cq->cq_idx << 32) | processed_cqe);
484 
485 	if (likely(tx_pkts)) {
486 		struct netdev_queue *txq;
487 
488 		txq = netdev_get_tx_queue(pfvf->netdev, cq->cint_idx);
489 		netdev_tx_completed_queue(txq, tx_pkts, tx_bytes);
490 		/* Check if queue was stopped earlier due to ring full */
491 		smp_mb();
492 		if (netif_tx_queue_stopped(txq) &&
493 		    netif_carrier_ok(pfvf->netdev))
494 			netif_tx_wake_queue(txq);
495 	}
496 	return 0;
497 }
498 
499 static void otx2_adjust_adaptive_coalese(struct otx2_nic *pfvf, struct otx2_cq_poll *cq_poll)
500 {
501 	struct dim_sample dim_sample;
502 	u64 rx_frames, rx_bytes;
503 
504 	rx_frames = OTX2_GET_RX_STATS(RX_BCAST) + OTX2_GET_RX_STATS(RX_MCAST) +
505 		OTX2_GET_RX_STATS(RX_UCAST);
506 	rx_bytes = OTX2_GET_RX_STATS(RX_OCTS);
507 	dim_update_sample(pfvf->napi_events, rx_frames, rx_bytes, &dim_sample);
508 	net_dim(&cq_poll->dim, dim_sample);
509 }
510 
511 int otx2_napi_handler(struct napi_struct *napi, int budget)
512 {
513 	struct otx2_cq_queue *rx_cq = NULL;
514 	struct otx2_cq_poll *cq_poll;
515 	int workdone = 0, cq_idx, i;
516 	struct otx2_cq_queue *cq;
517 	struct otx2_qset *qset;
518 	struct otx2_nic *pfvf;
519 
520 	cq_poll = container_of(napi, struct otx2_cq_poll, napi);
521 	pfvf = (struct otx2_nic *)cq_poll->dev;
522 	qset = &pfvf->qset;
523 
524 	for (i = 0; i < CQS_PER_CINT; i++) {
525 		cq_idx = cq_poll->cq_ids[i];
526 		if (unlikely(cq_idx == CINT_INVALID_CQ))
527 			continue;
528 		cq = &qset->cq[cq_idx];
529 		if (cq->cq_type == CQ_RX) {
530 			rx_cq = cq;
531 			workdone += otx2_rx_napi_handler(pfvf, napi,
532 							 cq, budget);
533 		} else {
534 			workdone += otx2_tx_napi_handler(pfvf, cq, budget);
535 		}
536 	}
537 
538 	if (rx_cq && rx_cq->pool_ptrs)
539 		pfvf->hw_ops->refill_pool_ptrs(pfvf, rx_cq);
540 	/* Clear the IRQ */
541 	otx2_write64(pfvf, NIX_LF_CINTX_INT(cq_poll->cint_idx), BIT_ULL(0));
542 
543 	if (workdone < budget && napi_complete_done(napi, workdone)) {
544 		/* If interface is going down, don't re-enable IRQ */
545 		if (pfvf->flags & OTX2_FLAG_INTF_DOWN)
546 			return workdone;
547 
548 		/* Check for adaptive interrupt coalesce */
549 		if (workdone != 0 &&
550 		    ((pfvf->flags & OTX2_FLAG_ADPTV_INT_COAL_ENABLED) ==
551 		     OTX2_FLAG_ADPTV_INT_COAL_ENABLED)) {
552 			/* Adjust irq coalese using net_dim */
553 			otx2_adjust_adaptive_coalese(pfvf, cq_poll);
554 			/* Update irq coalescing */
555 			for (i = 0; i < pfvf->hw.cint_cnt; i++)
556 				otx2_config_irq_coalescing(pfvf, i);
557 		}
558 
559 		/* Re-enable interrupts */
560 		otx2_write64(pfvf, NIX_LF_CINTX_ENA_W1S(cq_poll->cint_idx),
561 			     BIT_ULL(0));
562 	}
563 	return workdone;
564 }
565 
566 void otx2_sqe_flush(void *dev, struct otx2_snd_queue *sq,
567 		    int size, int qidx)
568 {
569 	u64 status;
570 
571 	/* Packet data stores should finish before SQE is flushed to HW */
572 	dma_wmb();
573 
574 	do {
575 		memcpy(sq->lmt_addr, sq->sqe_base, size);
576 		status = otx2_lmt_flush(sq->io_addr);
577 	} while (status == 0);
578 
579 	sq->head++;
580 	sq->head &= (sq->sqe_cnt - 1);
581 }
582 
583 #define MAX_SEGS_PER_SG	3
584 /* Add SQE scatter/gather subdescriptor structure */
585 static bool otx2_sqe_add_sg(struct otx2_nic *pfvf, struct otx2_snd_queue *sq,
586 			    struct sk_buff *skb, int num_segs, int *offset)
587 {
588 	struct nix_sqe_sg_s *sg = NULL;
589 	u64 dma_addr, *iova = NULL;
590 	u16 *sg_lens = NULL;
591 	int seg, len;
592 
593 	sq->sg[sq->head].num_segs = 0;
594 
595 	for (seg = 0; seg < num_segs; seg++) {
596 		if ((seg % MAX_SEGS_PER_SG) == 0) {
597 			sg = (struct nix_sqe_sg_s *)(sq->sqe_base + *offset);
598 			sg->ld_type = NIX_SEND_LDTYPE_LDD;
599 			sg->subdc = NIX_SUBDC_SG;
600 			sg->segs = 0;
601 			sg_lens = (void *)sg;
602 			iova = (void *)sg + sizeof(*sg);
603 			/* Next subdc always starts at a 16byte boundary.
604 			 * So if sg->segs is whether 2 or 3, offset += 16bytes.
605 			 */
606 			if ((num_segs - seg) >= (MAX_SEGS_PER_SG - 1))
607 				*offset += sizeof(*sg) + (3 * sizeof(u64));
608 			else
609 				*offset += sizeof(*sg) + sizeof(u64);
610 		}
611 		dma_addr = otx2_dma_map_skb_frag(pfvf, skb, seg, &len);
612 		if (dma_mapping_error(pfvf->dev, dma_addr))
613 			return false;
614 
615 		sg_lens[frag_num(seg % MAX_SEGS_PER_SG)] = len;
616 		sg->segs++;
617 		*iova++ = dma_addr;
618 
619 		/* Save DMA mapping info for later unmapping */
620 		sq->sg[sq->head].dma_addr[seg] = dma_addr;
621 		sq->sg[sq->head].size[seg] = len;
622 		sq->sg[sq->head].num_segs++;
623 	}
624 
625 	sq->sg[sq->head].skb = (u64)skb;
626 	return true;
627 }
628 
629 /* Add SQE extended header subdescriptor */
630 static void otx2_sqe_add_ext(struct otx2_nic *pfvf, struct otx2_snd_queue *sq,
631 			     struct sk_buff *skb, int *offset)
632 {
633 	struct nix_sqe_ext_s *ext;
634 
635 	ext = (struct nix_sqe_ext_s *)(sq->sqe_base + *offset);
636 	ext->subdc = NIX_SUBDC_EXT;
637 	if (skb_shinfo(skb)->gso_size) {
638 		ext->lso = 1;
639 		ext->lso_sb = skb_tcp_all_headers(skb);
640 		ext->lso_mps = skb_shinfo(skb)->gso_size;
641 
642 		/* Only TSOv4 and TSOv6 GSO offloads are supported */
643 		if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
644 			ext->lso_format = pfvf->hw.lso_tsov4_idx;
645 
646 			/* HW adds payload size to 'ip_hdr->tot_len' while
647 			 * sending TSO segment, hence set payload length
648 			 * in IP header of the packet to just header length.
649 			 */
650 			ip_hdr(skb)->tot_len =
651 				htons(ext->lso_sb - skb_network_offset(skb));
652 		} else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
653 			ext->lso_format = pfvf->hw.lso_tsov6_idx;
654 
655 			ipv6_hdr(skb)->payload_len =
656 				htons(ext->lso_sb - skb_network_offset(skb));
657 		} else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
658 			__be16 l3_proto = vlan_get_protocol(skb);
659 			struct udphdr *udph = udp_hdr(skb);
660 			u16 iplen;
661 
662 			ext->lso_sb = skb_transport_offset(skb) +
663 					sizeof(struct udphdr);
664 
665 			/* HW adds payload size to length fields in IP and
666 			 * UDP headers while segmentation, hence adjust the
667 			 * lengths to just header sizes.
668 			 */
669 			iplen = htons(ext->lso_sb - skb_network_offset(skb));
670 			if (l3_proto == htons(ETH_P_IP)) {
671 				ip_hdr(skb)->tot_len = iplen;
672 				ext->lso_format = pfvf->hw.lso_udpv4_idx;
673 			} else {
674 				ipv6_hdr(skb)->payload_len = iplen;
675 				ext->lso_format = pfvf->hw.lso_udpv6_idx;
676 			}
677 
678 			udph->len = htons(sizeof(struct udphdr));
679 		}
680 	} else if (skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
681 		ext->tstmp = 1;
682 	}
683 
684 #define OTX2_VLAN_PTR_OFFSET     (ETH_HLEN - ETH_TLEN)
685 	if (skb_vlan_tag_present(skb)) {
686 		if (skb->vlan_proto == htons(ETH_P_8021Q)) {
687 			ext->vlan1_ins_ena = 1;
688 			ext->vlan1_ins_ptr = OTX2_VLAN_PTR_OFFSET;
689 			ext->vlan1_ins_tci = skb_vlan_tag_get(skb);
690 		} else if (skb->vlan_proto == htons(ETH_P_8021AD)) {
691 			ext->vlan0_ins_ena = 1;
692 			ext->vlan0_ins_ptr = OTX2_VLAN_PTR_OFFSET;
693 			ext->vlan0_ins_tci = skb_vlan_tag_get(skb);
694 		}
695 	}
696 
697 	*offset += sizeof(*ext);
698 }
699 
700 static void otx2_sqe_add_mem(struct otx2_snd_queue *sq, int *offset,
701 			     int alg, u64 iova, int ptp_offset,
702 			     u64 base_ns, int udp_csum)
703 {
704 	struct nix_sqe_mem_s *mem;
705 
706 	mem = (struct nix_sqe_mem_s *)(sq->sqe_base + *offset);
707 	mem->subdc = NIX_SUBDC_MEM;
708 	mem->alg = alg;
709 	mem->wmem = 1; /* wait for the memory operation */
710 	mem->addr = iova;
711 
712 	if (ptp_offset) {
713 		mem->start_offset = ptp_offset;
714 		mem->udp_csum_crt = udp_csum;
715 		mem->base_ns = base_ns;
716 		mem->step_type = 1;
717 	}
718 
719 	*offset += sizeof(*mem);
720 }
721 
722 /* Add SQE header subdescriptor structure */
723 static void otx2_sqe_add_hdr(struct otx2_nic *pfvf, struct otx2_snd_queue *sq,
724 			     struct nix_sqe_hdr_s *sqe_hdr,
725 			     struct sk_buff *skb, u16 qidx)
726 {
727 	int proto = 0;
728 
729 	/* Check if SQE was framed before, if yes then no need to
730 	 * set these constants again and again.
731 	 */
732 	if (!sqe_hdr->total) {
733 		/* Don't free Tx buffers to Aura */
734 		sqe_hdr->df = 1;
735 		sqe_hdr->aura = sq->aura_id;
736 		/* Post a CQE Tx after pkt transmission */
737 		sqe_hdr->pnc = 1;
738 		sqe_hdr->sq = qidx;
739 	}
740 	sqe_hdr->total = skb->len;
741 	/* Set SQE identifier which will be used later for freeing SKB */
742 	sqe_hdr->sqe_id = sq->head;
743 
744 	/* Offload TCP/UDP checksum to HW */
745 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
746 		sqe_hdr->ol3ptr = skb_network_offset(skb);
747 		sqe_hdr->ol4ptr = skb_transport_offset(skb);
748 		/* get vlan protocol Ethertype */
749 		if (eth_type_vlan(skb->protocol))
750 			skb->protocol = vlan_get_protocol(skb);
751 
752 		if (skb->protocol == htons(ETH_P_IP)) {
753 			proto = ip_hdr(skb)->protocol;
754 			/* In case of TSO, HW needs this to be explicitly set.
755 			 * So set this always, instead of adding a check.
756 			 */
757 			sqe_hdr->ol3type = NIX_SENDL3TYPE_IP4_CKSUM;
758 		} else if (skb->protocol == htons(ETH_P_IPV6)) {
759 			proto = ipv6_hdr(skb)->nexthdr;
760 			sqe_hdr->ol3type = NIX_SENDL3TYPE_IP6;
761 		}
762 
763 		if (proto == IPPROTO_TCP)
764 			sqe_hdr->ol4type = NIX_SENDL4TYPE_TCP_CKSUM;
765 		else if (proto == IPPROTO_UDP)
766 			sqe_hdr->ol4type = NIX_SENDL4TYPE_UDP_CKSUM;
767 	}
768 }
769 
770 static int otx2_dma_map_tso_skb(struct otx2_nic *pfvf,
771 				struct otx2_snd_queue *sq,
772 				struct sk_buff *skb, int sqe, int hdr_len)
773 {
774 	int num_segs = skb_shinfo(skb)->nr_frags + 1;
775 	struct sg_list *sg = &sq->sg[sqe];
776 	u64 dma_addr;
777 	int seg, len;
778 
779 	sg->num_segs = 0;
780 
781 	/* Get payload length at skb->data */
782 	len = skb_headlen(skb) - hdr_len;
783 
784 	for (seg = 0; seg < num_segs; seg++) {
785 		/* Skip skb->data, if there is no payload */
786 		if (!seg && !len)
787 			continue;
788 		dma_addr = otx2_dma_map_skb_frag(pfvf, skb, seg, &len);
789 		if (dma_mapping_error(pfvf->dev, dma_addr))
790 			goto unmap;
791 
792 		/* Save DMA mapping info for later unmapping */
793 		sg->dma_addr[sg->num_segs] = dma_addr;
794 		sg->size[sg->num_segs] = len;
795 		sg->num_segs++;
796 	}
797 	return 0;
798 unmap:
799 	otx2_dma_unmap_skb_frags(pfvf, sg);
800 	return -EINVAL;
801 }
802 
803 static u64 otx2_tso_frag_dma_addr(struct otx2_snd_queue *sq,
804 				  struct sk_buff *skb, int seg,
805 				  u64 seg_addr, int hdr_len, int sqe)
806 {
807 	struct sg_list *sg = &sq->sg[sqe];
808 	const skb_frag_t *frag;
809 	int offset;
810 
811 	if (seg < 0)
812 		return sg->dma_addr[0] + (seg_addr - (u64)skb->data);
813 
814 	frag = &skb_shinfo(skb)->frags[seg];
815 	offset = seg_addr - (u64)skb_frag_address(frag);
816 	if (skb_headlen(skb) - hdr_len)
817 		seg++;
818 	return sg->dma_addr[seg] + offset;
819 }
820 
821 static void otx2_sqe_tso_add_sg(struct otx2_snd_queue *sq,
822 				struct sg_list *list, int *offset)
823 {
824 	struct nix_sqe_sg_s *sg = NULL;
825 	u16 *sg_lens = NULL;
826 	u64 *iova = NULL;
827 	int seg;
828 
829 	/* Add SG descriptors with buffer addresses */
830 	for (seg = 0; seg < list->num_segs; seg++) {
831 		if ((seg % MAX_SEGS_PER_SG) == 0) {
832 			sg = (struct nix_sqe_sg_s *)(sq->sqe_base + *offset);
833 			sg->ld_type = NIX_SEND_LDTYPE_LDD;
834 			sg->subdc = NIX_SUBDC_SG;
835 			sg->segs = 0;
836 			sg_lens = (void *)sg;
837 			iova = (void *)sg + sizeof(*sg);
838 			/* Next subdc always starts at a 16byte boundary.
839 			 * So if sg->segs is whether 2 or 3, offset += 16bytes.
840 			 */
841 			if ((list->num_segs - seg) >= (MAX_SEGS_PER_SG - 1))
842 				*offset += sizeof(*sg) + (3 * sizeof(u64));
843 			else
844 				*offset += sizeof(*sg) + sizeof(u64);
845 		}
846 		sg_lens[frag_num(seg % MAX_SEGS_PER_SG)] = list->size[seg];
847 		*iova++ = list->dma_addr[seg];
848 		sg->segs++;
849 	}
850 }
851 
852 static void otx2_sq_append_tso(struct otx2_nic *pfvf, struct otx2_snd_queue *sq,
853 			       struct sk_buff *skb, u16 qidx)
854 {
855 	struct netdev_queue *txq = netdev_get_tx_queue(pfvf->netdev, qidx);
856 	int hdr_len, tcp_data, seg_len, pkt_len, offset;
857 	struct nix_sqe_hdr_s *sqe_hdr;
858 	int first_sqe = sq->head;
859 	struct sg_list list;
860 	struct tso_t tso;
861 
862 	hdr_len = tso_start(skb, &tso);
863 
864 	/* Map SKB's fragments to DMA.
865 	 * It's done here to avoid mapping for every TSO segment's packet.
866 	 */
867 	if (otx2_dma_map_tso_skb(pfvf, sq, skb, first_sqe, hdr_len)) {
868 		dev_kfree_skb_any(skb);
869 		return;
870 	}
871 
872 	netdev_tx_sent_queue(txq, skb->len);
873 
874 	tcp_data = skb->len - hdr_len;
875 	while (tcp_data > 0) {
876 		char *hdr;
877 
878 		seg_len = min_t(int, skb_shinfo(skb)->gso_size, tcp_data);
879 		tcp_data -= seg_len;
880 
881 		/* Set SQE's SEND_HDR */
882 		memset(sq->sqe_base, 0, sq->sqe_size);
883 		sqe_hdr = (struct nix_sqe_hdr_s *)(sq->sqe_base);
884 		otx2_sqe_add_hdr(pfvf, sq, sqe_hdr, skb, qidx);
885 		offset = sizeof(*sqe_hdr);
886 
887 		/* Add TSO segment's pkt header */
888 		hdr = sq->tso_hdrs->base + (sq->head * TSO_HEADER_SIZE);
889 		tso_build_hdr(skb, hdr, &tso, seg_len, tcp_data == 0);
890 		list.dma_addr[0] =
891 			sq->tso_hdrs->iova + (sq->head * TSO_HEADER_SIZE);
892 		list.size[0] = hdr_len;
893 		list.num_segs = 1;
894 
895 		/* Add TSO segment's payload data fragments */
896 		pkt_len = hdr_len;
897 		while (seg_len > 0) {
898 			int size;
899 
900 			size = min_t(int, tso.size, seg_len);
901 
902 			list.size[list.num_segs] = size;
903 			list.dma_addr[list.num_segs] =
904 				otx2_tso_frag_dma_addr(sq, skb,
905 						       tso.next_frag_idx - 1,
906 						       (u64)tso.data, hdr_len,
907 						       first_sqe);
908 			list.num_segs++;
909 			pkt_len += size;
910 			seg_len -= size;
911 			tso_build_data(skb, &tso, size);
912 		}
913 		sqe_hdr->total = pkt_len;
914 		otx2_sqe_tso_add_sg(sq, &list, &offset);
915 
916 		/* DMA mappings and skb needs to be freed only after last
917 		 * TSO segment is transmitted out. So set 'PNC' only for
918 		 * last segment. Also point last segment's sqe_id to first
919 		 * segment's SQE index where skb address and DMA mappings
920 		 * are saved.
921 		 */
922 		if (!tcp_data) {
923 			sqe_hdr->pnc = 1;
924 			sqe_hdr->sqe_id = first_sqe;
925 			sq->sg[first_sqe].skb = (u64)skb;
926 		} else {
927 			sqe_hdr->pnc = 0;
928 		}
929 
930 		sqe_hdr->sizem1 = (offset / 16) - 1;
931 
932 		/* Flush SQE to HW */
933 		pfvf->hw_ops->sqe_flush(pfvf, sq, offset, qidx);
934 	}
935 }
936 
937 static bool is_hw_tso_supported(struct otx2_nic *pfvf,
938 				struct sk_buff *skb)
939 {
940 	int payload_len, last_seg_size;
941 
942 	if (test_bit(HW_TSO, &pfvf->hw.cap_flag))
943 		return true;
944 
945 	/* On 96xx A0, HW TSO not supported */
946 	if (!is_96xx_B0(pfvf->pdev))
947 		return false;
948 
949 	/* HW has an issue due to which when the payload of the last LSO
950 	 * segment is shorter than 16 bytes, some header fields may not
951 	 * be correctly modified, hence don't offload such TSO segments.
952 	 */
953 
954 	payload_len = skb->len - skb_tcp_all_headers(skb);
955 	last_seg_size = payload_len % skb_shinfo(skb)->gso_size;
956 	if (last_seg_size && last_seg_size < 16)
957 		return false;
958 
959 	return true;
960 }
961 
962 static int otx2_get_sqe_count(struct otx2_nic *pfvf, struct sk_buff *skb)
963 {
964 	if (!skb_shinfo(skb)->gso_size)
965 		return 1;
966 
967 	/* HW TSO */
968 	if (is_hw_tso_supported(pfvf, skb))
969 		return 1;
970 
971 	/* SW TSO */
972 	return skb_shinfo(skb)->gso_segs;
973 }
974 
975 static bool otx2_validate_network_transport(struct sk_buff *skb)
976 {
977 	if ((ip_hdr(skb)->protocol == IPPROTO_UDP) ||
978 	    (ipv6_hdr(skb)->nexthdr == IPPROTO_UDP)) {
979 		struct udphdr *udph = udp_hdr(skb);
980 
981 		if (udph->source == htons(PTP_PORT) &&
982 		    udph->dest == htons(PTP_PORT))
983 			return true;
984 	}
985 
986 	return false;
987 }
988 
989 static bool otx2_ptp_is_sync(struct sk_buff *skb, int *offset, int *udp_csum)
990 {
991 	struct ethhdr *eth = (struct ethhdr *)(skb->data);
992 	u16 nix_offload_hlen = 0, inner_vhlen = 0;
993 	u8 *data = skb->data, *msgtype;
994 	__be16 proto = eth->h_proto;
995 	int network_depth = 0;
996 
997 	/* NIX is programmed to offload outer  VLAN header
998 	 * in case of single vlan protocol field holds Network header ETH_IP/V6
999 	 * in case of stacked vlan protocol field holds Inner vlan (8100)
1000 	 */
1001 	if (skb->dev->features & NETIF_F_HW_VLAN_CTAG_TX &&
1002 	    skb->dev->features & NETIF_F_HW_VLAN_STAG_TX) {
1003 		if (skb->vlan_proto == htons(ETH_P_8021AD)) {
1004 			/* Get vlan protocol */
1005 			proto = __vlan_get_protocol(skb, eth->h_proto, NULL);
1006 			/* SKB APIs like skb_transport_offset does not include
1007 			 * offloaded vlan header length. Need to explicitly add
1008 			 * the length
1009 			 */
1010 			nix_offload_hlen = VLAN_HLEN;
1011 			inner_vhlen = VLAN_HLEN;
1012 		} else if (skb->vlan_proto == htons(ETH_P_8021Q)) {
1013 			nix_offload_hlen = VLAN_HLEN;
1014 		}
1015 	} else if (eth_type_vlan(eth->h_proto)) {
1016 		proto = __vlan_get_protocol(skb, eth->h_proto, &network_depth);
1017 	}
1018 
1019 	switch (ntohs(proto)) {
1020 	case ETH_P_1588:
1021 		if (network_depth)
1022 			*offset = network_depth;
1023 		else
1024 			*offset = ETH_HLEN + nix_offload_hlen +
1025 				  inner_vhlen;
1026 		break;
1027 	case ETH_P_IP:
1028 	case ETH_P_IPV6:
1029 		if (!otx2_validate_network_transport(skb))
1030 			return false;
1031 
1032 		*udp_csum = 1;
1033 		*offset = nix_offload_hlen + skb_transport_offset(skb) +
1034 			  sizeof(struct udphdr);
1035 	}
1036 
1037 	msgtype = data + *offset;
1038 
1039 	/* Check PTP messageId is SYNC or not */
1040 	return (*msgtype & 0xf) == 0;
1041 }
1042 
1043 static void otx2_set_txtstamp(struct otx2_nic *pfvf, struct sk_buff *skb,
1044 			      struct otx2_snd_queue *sq, int *offset)
1045 {
1046 	struct ptpv2_tstamp *origin_tstamp;
1047 	int ptp_offset = 0, udp_csum = 0;
1048 	struct timespec64 ts;
1049 	u64 iova;
1050 
1051 	if (unlikely(!skb_shinfo(skb)->gso_size &&
1052 		     (skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP))) {
1053 		if (unlikely(pfvf->flags & OTX2_FLAG_PTP_ONESTEP_SYNC)) {
1054 			if (otx2_ptp_is_sync(skb, &ptp_offset, &udp_csum)) {
1055 				origin_tstamp = (struct ptpv2_tstamp *)
1056 						((u8 *)skb->data + ptp_offset +
1057 						 PTP_SYNC_SEC_OFFSET);
1058 				ts = ns_to_timespec64(pfvf->ptp->tstamp);
1059 				origin_tstamp->seconds_msb = htons((ts.tv_sec >> 32) & 0xffff);
1060 				origin_tstamp->seconds_lsb = htonl(ts.tv_sec & 0xffffffff);
1061 				origin_tstamp->nanoseconds = htonl(ts.tv_nsec);
1062 				/* Point to correction field in PTP packet */
1063 				ptp_offset += 8;
1064 			}
1065 		} else {
1066 			skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
1067 		}
1068 		iova = sq->timestamps->iova + (sq->head * sizeof(u64));
1069 		otx2_sqe_add_mem(sq, offset, NIX_SENDMEMALG_E_SETTSTMP, iova,
1070 				 ptp_offset, pfvf->ptp->base_ns, udp_csum);
1071 	} else {
1072 		skb_tx_timestamp(skb);
1073 	}
1074 }
1075 
1076 bool otx2_sq_append_skb(struct net_device *netdev, struct otx2_snd_queue *sq,
1077 			struct sk_buff *skb, u16 qidx)
1078 {
1079 	struct netdev_queue *txq = netdev_get_tx_queue(netdev, qidx);
1080 	struct otx2_nic *pfvf = netdev_priv(netdev);
1081 	int offset, num_segs, free_desc;
1082 	struct nix_sqe_hdr_s *sqe_hdr;
1083 
1084 	/* Check if there is enough room between producer
1085 	 * and consumer index.
1086 	 */
1087 	free_desc = (sq->cons_head - sq->head - 1 + sq->sqe_cnt) & (sq->sqe_cnt - 1);
1088 	if (free_desc < sq->sqe_thresh)
1089 		return false;
1090 
1091 	if (free_desc < otx2_get_sqe_count(pfvf, skb))
1092 		return false;
1093 
1094 	num_segs = skb_shinfo(skb)->nr_frags + 1;
1095 
1096 	/* If SKB doesn't fit in a single SQE, linearize it.
1097 	 * TODO: Consider adding JUMP descriptor instead.
1098 	 */
1099 	if (unlikely(num_segs > OTX2_MAX_FRAGS_IN_SQE)) {
1100 		if (__skb_linearize(skb)) {
1101 			dev_kfree_skb_any(skb);
1102 			return true;
1103 		}
1104 		num_segs = skb_shinfo(skb)->nr_frags + 1;
1105 	}
1106 
1107 	if (skb_shinfo(skb)->gso_size && !is_hw_tso_supported(pfvf, skb)) {
1108 		/* Insert vlan tag before giving pkt to tso */
1109 		if (skb_vlan_tag_present(skb))
1110 			skb = __vlan_hwaccel_push_inside(skb);
1111 		otx2_sq_append_tso(pfvf, sq, skb, qidx);
1112 		return true;
1113 	}
1114 
1115 	/* Set SQE's SEND_HDR.
1116 	 * Do not clear the first 64bit as it contains constant info.
1117 	 */
1118 	memset(sq->sqe_base + 8, 0, sq->sqe_size - 8);
1119 	sqe_hdr = (struct nix_sqe_hdr_s *)(sq->sqe_base);
1120 	otx2_sqe_add_hdr(pfvf, sq, sqe_hdr, skb, qidx);
1121 	offset = sizeof(*sqe_hdr);
1122 
1123 	/* Add extended header if needed */
1124 	otx2_sqe_add_ext(pfvf, sq, skb, &offset);
1125 
1126 	/* Add SG subdesc with data frags */
1127 	if (!otx2_sqe_add_sg(pfvf, sq, skb, num_segs, &offset)) {
1128 		otx2_dma_unmap_skb_frags(pfvf, &sq->sg[sq->head]);
1129 		return false;
1130 	}
1131 
1132 	otx2_set_txtstamp(pfvf, skb, sq, &offset);
1133 
1134 	sqe_hdr->sizem1 = (offset / 16) - 1;
1135 
1136 	netdev_tx_sent_queue(txq, skb->len);
1137 
1138 	/* Flush SQE to HW */
1139 	pfvf->hw_ops->sqe_flush(pfvf, sq, offset, qidx);
1140 
1141 	return true;
1142 }
1143 EXPORT_SYMBOL(otx2_sq_append_skb);
1144 
1145 void otx2_cleanup_rx_cqes(struct otx2_nic *pfvf, struct otx2_cq_queue *cq)
1146 {
1147 	struct nix_cqe_rx_s *cqe;
1148 	int processed_cqe = 0;
1149 	u64 iova, pa;
1150 
1151 	if (pfvf->xdp_prog)
1152 		xdp_rxq_info_unreg(&cq->xdp_rxq);
1153 
1154 	if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe)
1155 		return;
1156 
1157 	while (cq->pend_cqe) {
1158 		cqe = (struct nix_cqe_rx_s *)otx2_get_next_cqe(cq);
1159 		processed_cqe++;
1160 		cq->pend_cqe--;
1161 
1162 		if (!cqe)
1163 			continue;
1164 		if (cqe->sg.segs > 1) {
1165 			otx2_free_rcv_seg(pfvf, cqe, cq->cq_idx);
1166 			continue;
1167 		}
1168 		iova = cqe->sg.seg_addr - OTX2_HEAD_ROOM;
1169 		pa = otx2_iova_to_phys(pfvf->iommu_domain, iova);
1170 		otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize, DMA_FROM_DEVICE);
1171 		put_page(virt_to_page(phys_to_virt(pa)));
1172 	}
1173 
1174 	/* Free CQEs to HW */
1175 	otx2_write64(pfvf, NIX_LF_CQ_OP_DOOR,
1176 		     ((u64)cq->cq_idx << 32) | processed_cqe);
1177 }
1178 
1179 void otx2_cleanup_tx_cqes(struct otx2_nic *pfvf, struct otx2_cq_queue *cq)
1180 {
1181 	struct sk_buff *skb = NULL;
1182 	struct otx2_snd_queue *sq;
1183 	struct nix_cqe_tx_s *cqe;
1184 	int processed_cqe = 0;
1185 	struct sg_list *sg;
1186 
1187 	sq = &pfvf->qset.sq[cq->cint_idx];
1188 
1189 	if (otx2_nix_cq_op_status(pfvf, cq) || !cq->pend_cqe)
1190 		return;
1191 
1192 	while (cq->pend_cqe) {
1193 		cqe = (struct nix_cqe_tx_s *)otx2_get_next_cqe(cq);
1194 		processed_cqe++;
1195 		cq->pend_cqe--;
1196 
1197 		if (!cqe)
1198 			continue;
1199 		sg = &sq->sg[cqe->comp.sqe_id];
1200 		skb = (struct sk_buff *)sg->skb;
1201 		if (skb) {
1202 			otx2_dma_unmap_skb_frags(pfvf, sg);
1203 			dev_kfree_skb_any(skb);
1204 			sg->skb = (u64)NULL;
1205 		}
1206 	}
1207 
1208 	/* Free CQEs to HW */
1209 	otx2_write64(pfvf, NIX_LF_CQ_OP_DOOR,
1210 		     ((u64)cq->cq_idx << 32) | processed_cqe);
1211 }
1212 
1213 int otx2_rxtx_enable(struct otx2_nic *pfvf, bool enable)
1214 {
1215 	struct msg_req *msg;
1216 	int err;
1217 
1218 	mutex_lock(&pfvf->mbox.lock);
1219 	if (enable)
1220 		msg = otx2_mbox_alloc_msg_nix_lf_start_rx(&pfvf->mbox);
1221 	else
1222 		msg = otx2_mbox_alloc_msg_nix_lf_stop_rx(&pfvf->mbox);
1223 
1224 	if (!msg) {
1225 		mutex_unlock(&pfvf->mbox.lock);
1226 		return -ENOMEM;
1227 	}
1228 
1229 	err = otx2_sync_mbox_msg(&pfvf->mbox);
1230 	mutex_unlock(&pfvf->mbox.lock);
1231 	return err;
1232 }
1233 
1234 static void otx2_xdp_sqe_add_sg(struct otx2_snd_queue *sq, u64 dma_addr,
1235 				int len, int *offset)
1236 {
1237 	struct nix_sqe_sg_s *sg = NULL;
1238 	u64 *iova = NULL;
1239 
1240 	sg = (struct nix_sqe_sg_s *)(sq->sqe_base + *offset);
1241 	sg->ld_type = NIX_SEND_LDTYPE_LDD;
1242 	sg->subdc = NIX_SUBDC_SG;
1243 	sg->segs = 1;
1244 	sg->seg1_size = len;
1245 	iova = (void *)sg + sizeof(*sg);
1246 	*iova = dma_addr;
1247 	*offset += sizeof(*sg) + sizeof(u64);
1248 
1249 	sq->sg[sq->head].dma_addr[0] = dma_addr;
1250 	sq->sg[sq->head].size[0] = len;
1251 	sq->sg[sq->head].num_segs = 1;
1252 }
1253 
1254 bool otx2_xdp_sq_append_pkt(struct otx2_nic *pfvf, u64 iova, int len, u16 qidx)
1255 {
1256 	struct nix_sqe_hdr_s *sqe_hdr;
1257 	struct otx2_snd_queue *sq;
1258 	int offset, free_sqe;
1259 
1260 	sq = &pfvf->qset.sq[qidx];
1261 	free_sqe = (sq->num_sqbs - *sq->aura_fc_addr) * sq->sqe_per_sqb;
1262 	if (free_sqe < sq->sqe_thresh)
1263 		return false;
1264 
1265 	memset(sq->sqe_base + 8, 0, sq->sqe_size - 8);
1266 
1267 	sqe_hdr = (struct nix_sqe_hdr_s *)(sq->sqe_base);
1268 
1269 	if (!sqe_hdr->total) {
1270 		sqe_hdr->aura = sq->aura_id;
1271 		sqe_hdr->df = 1;
1272 		sqe_hdr->sq = qidx;
1273 		sqe_hdr->pnc = 1;
1274 	}
1275 	sqe_hdr->total = len;
1276 	sqe_hdr->sqe_id = sq->head;
1277 
1278 	offset = sizeof(*sqe_hdr);
1279 
1280 	otx2_xdp_sqe_add_sg(sq, iova, len, &offset);
1281 	sqe_hdr->sizem1 = (offset / 16) - 1;
1282 	pfvf->hw_ops->sqe_flush(pfvf, sq, offset, qidx);
1283 
1284 	return true;
1285 }
1286 
1287 static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
1288 				     struct bpf_prog *prog,
1289 				     struct nix_cqe_rx_s *cqe,
1290 				     struct otx2_cq_queue *cq)
1291 {
1292 	unsigned char *hard_start, *data;
1293 	int qidx = cq->cq_idx;
1294 	struct xdp_buff xdp;
1295 	struct page *page;
1296 	u64 iova, pa;
1297 	u32 act;
1298 	int err;
1299 
1300 	iova = cqe->sg.seg_addr - OTX2_HEAD_ROOM;
1301 	pa = otx2_iova_to_phys(pfvf->iommu_domain, iova);
1302 	page = virt_to_page(phys_to_virt(pa));
1303 
1304 	xdp_init_buff(&xdp, pfvf->rbsize, &cq->xdp_rxq);
1305 
1306 	data = (unsigned char *)phys_to_virt(pa);
1307 	hard_start = page_address(page);
1308 	xdp_prepare_buff(&xdp, hard_start, data - hard_start,
1309 			 cqe->sg.seg_size, false);
1310 
1311 	act = bpf_prog_run_xdp(prog, &xdp);
1312 
1313 	switch (act) {
1314 	case XDP_PASS:
1315 		break;
1316 	case XDP_TX:
1317 		qidx += pfvf->hw.tx_queues;
1318 		cq->pool_ptrs++;
1319 		return otx2_xdp_sq_append_pkt(pfvf, iova,
1320 					      cqe->sg.seg_size, qidx);
1321 	case XDP_REDIRECT:
1322 		cq->pool_ptrs++;
1323 		err = xdp_do_redirect(pfvf->netdev, &xdp, prog);
1324 
1325 		otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize,
1326 				    DMA_FROM_DEVICE);
1327 		if (!err)
1328 			return true;
1329 		put_page(page);
1330 		break;
1331 	default:
1332 		bpf_warn_invalid_xdp_action(pfvf->netdev, prog, act);
1333 		break;
1334 	case XDP_ABORTED:
1335 		trace_xdp_exception(pfvf->netdev, prog, act);
1336 		break;
1337 	case XDP_DROP:
1338 		otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize,
1339 				    DMA_FROM_DEVICE);
1340 		put_page(page);
1341 		cq->pool_ptrs++;
1342 		return true;
1343 	}
1344 	return false;
1345 }
1346