xref: /linux/drivers/net/ethernet/huawei/hinic/hinic_tx.c (revision b7019ac550eb3916f34d79db583e9b7ea2524afa)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Huawei HiNIC PCI Express Linux driver
4  * Copyright(c) 2017 Huawei Technologies Co., Ltd
5  */
6 
7 #include <linux/kernel.h>
8 #include <linux/netdevice.h>
9 #include <linux/u64_stats_sync.h>
10 #include <linux/errno.h>
11 #include <linux/types.h>
12 #include <linux/pci.h>
13 #include <linux/device.h>
14 #include <linux/dma-mapping.h>
15 #include <linux/slab.h>
16 #include <linux/interrupt.h>
17 #include <linux/skbuff.h>
18 #include <linux/smp.h>
19 #include <asm/byteorder.h>
20 #include <linux/ip.h>
21 #include <linux/tcp.h>
22 #include <linux/sctp.h>
23 #include <linux/ipv6.h>
24 #include <net/ipv6.h>
25 #include <net/checksum.h>
26 #include <net/ip6_checksum.h>
27 
28 #include "hinic_common.h"
29 #include "hinic_hw_if.h"
30 #include "hinic_hw_wqe.h"
31 #include "hinic_hw_wq.h"
32 #include "hinic_hw_qp.h"
33 #include "hinic_hw_dev.h"
34 #include "hinic_dev.h"
35 #include "hinic_tx.h"
36 
37 #define TX_IRQ_NO_PENDING               0
38 #define TX_IRQ_NO_COALESC               0
39 #define TX_IRQ_NO_LLI_TIMER             0
40 #define TX_IRQ_NO_CREDIT                0
41 #define TX_IRQ_NO_RESEND_TIMER          0
42 
43 #define CI_UPDATE_NO_PENDING            0
44 #define CI_UPDATE_NO_COALESC            0
45 
46 #define HW_CONS_IDX(sq)                 be16_to_cpu(*(u16 *)((sq)->hw_ci_addr))
47 
48 #define MIN_SKB_LEN                     17
49 
50 #define	MAX_PAYLOAD_OFFSET	        221
51 #define TRANSPORT_OFFSET(l4_hdr, skb)	((u32)((l4_hdr) - (skb)->data))
52 
53 union hinic_l3 {
54 	struct iphdr *v4;
55 	struct ipv6hdr *v6;
56 	unsigned char *hdr;
57 };
58 
59 union hinic_l4 {
60 	struct tcphdr *tcp;
61 	struct udphdr *udp;
62 	unsigned char *hdr;
63 };
64 
65 enum hinic_offload_type {
66 	TX_OFFLOAD_TSO     = BIT(0),
67 	TX_OFFLOAD_CSUM    = BIT(1),
68 	TX_OFFLOAD_VLAN    = BIT(2),
69 	TX_OFFLOAD_INVALID = BIT(3),
70 };
71 
72 /**
73  * hinic_txq_clean_stats - Clean the statistics of specific queue
74  * @txq: Logical Tx Queue
75  **/
76 void hinic_txq_clean_stats(struct hinic_txq *txq)
77 {
78 	struct hinic_txq_stats *txq_stats = &txq->txq_stats;
79 
80 	u64_stats_update_begin(&txq_stats->syncp);
81 	txq_stats->pkts    = 0;
82 	txq_stats->bytes   = 0;
83 	txq_stats->tx_busy = 0;
84 	txq_stats->tx_wake = 0;
85 	txq_stats->tx_dropped = 0;
86 	u64_stats_update_end(&txq_stats->syncp);
87 }
88 
89 /**
90  * hinic_txq_get_stats - get statistics of Tx Queue
91  * @txq: Logical Tx Queue
92  * @stats: return updated stats here
93  **/
94 void hinic_txq_get_stats(struct hinic_txq *txq, struct hinic_txq_stats *stats)
95 {
96 	struct hinic_txq_stats *txq_stats = &txq->txq_stats;
97 	unsigned int start;
98 
99 	u64_stats_update_begin(&stats->syncp);
100 	do {
101 		start = u64_stats_fetch_begin(&txq_stats->syncp);
102 		stats->pkts    = txq_stats->pkts;
103 		stats->bytes   = txq_stats->bytes;
104 		stats->tx_busy = txq_stats->tx_busy;
105 		stats->tx_wake = txq_stats->tx_wake;
106 		stats->tx_dropped = txq_stats->tx_dropped;
107 	} while (u64_stats_fetch_retry(&txq_stats->syncp, start));
108 	u64_stats_update_end(&stats->syncp);
109 }
110 
111 /**
112  * txq_stats_init - Initialize the statistics of specific queue
113  * @txq: Logical Tx Queue
114  **/
115 static void txq_stats_init(struct hinic_txq *txq)
116 {
117 	struct hinic_txq_stats *txq_stats = &txq->txq_stats;
118 
119 	u64_stats_init(&txq_stats->syncp);
120 	hinic_txq_clean_stats(txq);
121 }
122 
123 /**
124  * tx_map_skb - dma mapping for skb and return sges
125  * @nic_dev: nic device
126  * @skb: the skb
127  * @sges: returned sges
128  *
129  * Return 0 - Success, negative - Failure
130  **/
131 static int tx_map_skb(struct hinic_dev *nic_dev, struct sk_buff *skb,
132 		      struct hinic_sge *sges)
133 {
134 	struct hinic_hwdev *hwdev = nic_dev->hwdev;
135 	struct hinic_hwif *hwif = hwdev->hwif;
136 	struct pci_dev *pdev = hwif->pdev;
137 	struct skb_frag_struct *frag;
138 	dma_addr_t dma_addr;
139 	int i, j;
140 
141 	dma_addr = dma_map_single(&pdev->dev, skb->data, skb_headlen(skb),
142 				  DMA_TO_DEVICE);
143 	if (dma_mapping_error(&pdev->dev, dma_addr)) {
144 		dev_err(&pdev->dev, "Failed to map Tx skb data\n");
145 		return -EFAULT;
146 	}
147 
148 	hinic_set_sge(&sges[0], dma_addr, skb_headlen(skb));
149 
150 	for (i = 0 ; i < skb_shinfo(skb)->nr_frags; i++) {
151 		frag = &skb_shinfo(skb)->frags[i];
152 
153 		dma_addr = skb_frag_dma_map(&pdev->dev, frag, 0,
154 					    skb_frag_size(frag),
155 					    DMA_TO_DEVICE);
156 		if (dma_mapping_error(&pdev->dev, dma_addr)) {
157 			dev_err(&pdev->dev, "Failed to map Tx skb frag\n");
158 			goto err_tx_map;
159 		}
160 
161 		hinic_set_sge(&sges[i + 1], dma_addr, skb_frag_size(frag));
162 	}
163 
164 	return 0;
165 
166 err_tx_map:
167 	for (j = 0; j < i; j++)
168 		dma_unmap_page(&pdev->dev, hinic_sge_to_dma(&sges[j + 1]),
169 			       sges[j + 1].len, DMA_TO_DEVICE);
170 
171 	dma_unmap_single(&pdev->dev, hinic_sge_to_dma(&sges[0]), sges[0].len,
172 			 DMA_TO_DEVICE);
173 	return -EFAULT;
174 }
175 
176 /**
177  * tx_unmap_skb - unmap the dma address of the skb
178  * @nic_dev: nic device
179  * @skb: the skb
180  * @sges: the sges that are connected to the skb
181  **/
182 static void tx_unmap_skb(struct hinic_dev *nic_dev, struct sk_buff *skb,
183 			 struct hinic_sge *sges)
184 {
185 	struct hinic_hwdev *hwdev = nic_dev->hwdev;
186 	struct hinic_hwif *hwif = hwdev->hwif;
187 	struct pci_dev *pdev = hwif->pdev;
188 	int i;
189 
190 	for (i = 0; i < skb_shinfo(skb)->nr_frags ; i++)
191 		dma_unmap_page(&pdev->dev, hinic_sge_to_dma(&sges[i + 1]),
192 			       sges[i + 1].len, DMA_TO_DEVICE);
193 
194 	dma_unmap_single(&pdev->dev, hinic_sge_to_dma(&sges[0]), sges[0].len,
195 			 DMA_TO_DEVICE);
196 }
197 
198 static void get_inner_l3_l4_type(struct sk_buff *skb, union hinic_l3 *ip,
199 				 union hinic_l4 *l4,
200 				 enum hinic_offload_type offload_type,
201 				 enum hinic_l3_offload_type *l3_type,
202 				 u8 *l4_proto)
203 {
204 	u8 *exthdr;
205 
206 	if (ip->v4->version == 4) {
207 		*l3_type = (offload_type == TX_OFFLOAD_CSUM) ?
208 			   IPV4_PKT_NO_CHKSUM_OFFLOAD :
209 			   IPV4_PKT_WITH_CHKSUM_OFFLOAD;
210 		*l4_proto = ip->v4->protocol;
211 	} else if (ip->v4->version == 6) {
212 		*l3_type = IPV6_PKT;
213 		exthdr = ip->hdr + sizeof(*ip->v6);
214 		*l4_proto = ip->v6->nexthdr;
215 		if (exthdr != l4->hdr) {
216 			int start = exthdr - skb->data;
217 			__be16 frag_off;
218 
219 			ipv6_skip_exthdr(skb, start, l4_proto, &frag_off);
220 		}
221 	} else {
222 		*l3_type = L3TYPE_UNKNOWN;
223 		*l4_proto = 0;
224 	}
225 }
226 
227 static void get_inner_l4_info(struct sk_buff *skb, union hinic_l4 *l4,
228 			      enum hinic_offload_type offload_type, u8 l4_proto,
229 			      enum hinic_l4_offload_type *l4_offload,
230 			      u32 *l4_len, u32 *offset)
231 {
232 	*l4_offload = OFFLOAD_DISABLE;
233 	*offset = 0;
234 	*l4_len = 0;
235 
236 	switch (l4_proto) {
237 	case IPPROTO_TCP:
238 		*l4_offload = TCP_OFFLOAD_ENABLE;
239 		/* doff in unit of 4B */
240 		*l4_len = l4->tcp->doff * 4;
241 		*offset = *l4_len + TRANSPORT_OFFSET(l4->hdr, skb);
242 		break;
243 
244 	case IPPROTO_UDP:
245 		*l4_offload = UDP_OFFLOAD_ENABLE;
246 		*l4_len = sizeof(struct udphdr);
247 		*offset = TRANSPORT_OFFSET(l4->hdr, skb);
248 		break;
249 
250 	case IPPROTO_SCTP:
251 		/* only csum offload support sctp */
252 		if (offload_type != TX_OFFLOAD_CSUM)
253 			break;
254 
255 		*l4_offload = SCTP_OFFLOAD_ENABLE;
256 		*l4_len = sizeof(struct sctphdr);
257 		*offset = TRANSPORT_OFFSET(l4->hdr, skb);
258 		break;
259 
260 	default:
261 		break;
262 	}
263 }
264 
265 static __sum16 csum_magic(union hinic_l3 *ip, unsigned short proto)
266 {
267 	return (ip->v4->version == 4) ?
268 		csum_tcpudp_magic(ip->v4->saddr, ip->v4->daddr, 0, proto, 0) :
269 		csum_ipv6_magic(&ip->v6->saddr, &ip->v6->daddr, 0, proto, 0);
270 }
271 
272 static int offload_tso(struct hinic_sq_task *task, u32 *queue_info,
273 		       struct sk_buff *skb)
274 {
275 	u32 offset, l4_len, ip_identify, network_hdr_len;
276 	enum hinic_l3_offload_type l3_offload;
277 	enum hinic_l4_offload_type l4_offload;
278 	union hinic_l3 ip;
279 	union hinic_l4 l4;
280 	u8 l4_proto;
281 
282 	if (!skb_is_gso(skb))
283 		return 0;
284 
285 	if (skb_cow_head(skb, 0) < 0)
286 		return -EPROTONOSUPPORT;
287 
288 	if (skb->encapsulation) {
289 		u32 gso_type = skb_shinfo(skb)->gso_type;
290 		u32 tunnel_type = 0;
291 		u32 l4_tunnel_len;
292 
293 		ip.hdr = skb_network_header(skb);
294 		l4.hdr = skb_transport_header(skb);
295 		network_hdr_len = skb_inner_network_header_len(skb);
296 
297 		if (ip.v4->version == 4) {
298 			ip.v4->tot_len = 0;
299 			l3_offload = IPV4_PKT_WITH_CHKSUM_OFFLOAD;
300 		} else if (ip.v4->version == 6) {
301 			l3_offload = IPV6_PKT;
302 		} else {
303 			l3_offload = 0;
304 		}
305 
306 		hinic_task_set_outter_l3(task, l3_offload,
307 					 skb_network_header_len(skb));
308 
309 		if (gso_type & SKB_GSO_UDP_TUNNEL_CSUM) {
310 			l4.udp->check = ~csum_magic(&ip, IPPROTO_UDP);
311 			tunnel_type = TUNNEL_UDP_CSUM;
312 		} else if (gso_type & SKB_GSO_UDP_TUNNEL) {
313 			tunnel_type = TUNNEL_UDP_NO_CSUM;
314 		}
315 
316 		l4_tunnel_len = skb_inner_network_offset(skb) -
317 				skb_transport_offset(skb);
318 		hinic_task_set_tunnel_l4(task, tunnel_type, l4_tunnel_len);
319 
320 		ip.hdr = skb_inner_network_header(skb);
321 		l4.hdr = skb_inner_transport_header(skb);
322 	} else {
323 		ip.hdr = skb_network_header(skb);
324 		l4.hdr = skb_transport_header(skb);
325 		network_hdr_len = skb_network_header_len(skb);
326 	}
327 
328 	/* initialize inner IP header fields */
329 	if (ip.v4->version == 4)
330 		ip.v4->tot_len = 0;
331 	else
332 		ip.v6->payload_len = 0;
333 
334 	get_inner_l3_l4_type(skb, &ip, &l4, TX_OFFLOAD_TSO, &l3_offload,
335 			     &l4_proto);
336 
337 	hinic_task_set_inner_l3(task, l3_offload, network_hdr_len);
338 
339 	ip_identify = 0;
340 	if (l4_proto == IPPROTO_TCP)
341 		l4.tcp->check = ~csum_magic(&ip, IPPROTO_TCP);
342 
343 	get_inner_l4_info(skb, &l4, TX_OFFLOAD_TSO, l4_proto, &l4_offload,
344 			  &l4_len, &offset);
345 
346 	hinic_set_tso_inner_l4(task, queue_info, l4_offload, l4_len, offset,
347 			       ip_identify, skb_shinfo(skb)->gso_size);
348 
349 	return 1;
350 }
351 
352 static int offload_csum(struct hinic_sq_task *task, u32 *queue_info,
353 			struct sk_buff *skb)
354 {
355 	enum hinic_l4_offload_type l4_offload;
356 	u32 offset, l4_len, network_hdr_len;
357 	enum hinic_l3_offload_type l3_type;
358 	union hinic_l3 ip;
359 	union hinic_l4 l4;
360 	u8 l4_proto;
361 
362 	if (skb->ip_summed != CHECKSUM_PARTIAL)
363 		return 0;
364 
365 	if (skb->encapsulation) {
366 		u32 l4_tunnel_len;
367 
368 		ip.hdr = skb_network_header(skb);
369 
370 		if (ip.v4->version == 4)
371 			l3_type = IPV4_PKT_NO_CHKSUM_OFFLOAD;
372 		else if (ip.v4->version == 6)
373 			l3_type = IPV6_PKT;
374 		else
375 			l3_type = L3TYPE_UNKNOWN;
376 
377 		hinic_task_set_outter_l3(task, l3_type,
378 					 skb_network_header_len(skb));
379 
380 		l4_tunnel_len = skb_inner_network_offset(skb) -
381 				skb_transport_offset(skb);
382 
383 		hinic_task_set_tunnel_l4(task, TUNNEL_UDP_NO_CSUM,
384 					 l4_tunnel_len);
385 
386 		ip.hdr = skb_inner_network_header(skb);
387 		l4.hdr = skb_inner_transport_header(skb);
388 		network_hdr_len = skb_inner_network_header_len(skb);
389 	} else {
390 		ip.hdr = skb_network_header(skb);
391 		l4.hdr = skb_transport_header(skb);
392 		network_hdr_len = skb_network_header_len(skb);
393 	}
394 
395 	get_inner_l3_l4_type(skb, &ip, &l4, TX_OFFLOAD_CSUM, &l3_type,
396 			     &l4_proto);
397 
398 	hinic_task_set_inner_l3(task, l3_type, network_hdr_len);
399 
400 	get_inner_l4_info(skb, &l4, TX_OFFLOAD_CSUM, l4_proto, &l4_offload,
401 			  &l4_len, &offset);
402 
403 	hinic_set_cs_inner_l4(task, queue_info, l4_offload, l4_len, offset);
404 
405 	return 1;
406 }
407 
408 static int hinic_tx_offload(struct sk_buff *skb, struct hinic_sq_task *task,
409 			    u32 *queue_info)
410 {
411 	enum hinic_offload_type offload = 0;
412 	int enabled;
413 
414 	enabled = offload_tso(task, queue_info, skb);
415 	if (enabled > 0) {
416 		offload |= TX_OFFLOAD_TSO;
417 	} else if (enabled == 0) {
418 		enabled = offload_csum(task, queue_info, skb);
419 		if (enabled)
420 			offload |= TX_OFFLOAD_CSUM;
421 	} else {
422 		return -EPROTONOSUPPORT;
423 	}
424 
425 	if (offload)
426 		hinic_task_set_l2hdr(task, skb_network_offset(skb));
427 
428 	/* payload offset should not more than 221 */
429 	if (HINIC_SQ_CTRL_GET(*queue_info, QUEUE_INFO_PLDOFF) >
430 	    MAX_PAYLOAD_OFFSET) {
431 		return -EPROTONOSUPPORT;
432 	}
433 
434 	/* mss should not less than 80 */
435 	if (HINIC_SQ_CTRL_GET(*queue_info, QUEUE_INFO_MSS) < HINIC_MSS_MIN) {
436 		*queue_info = HINIC_SQ_CTRL_CLEAR(*queue_info, QUEUE_INFO_MSS);
437 		*queue_info |= HINIC_SQ_CTRL_SET(HINIC_MSS_MIN, QUEUE_INFO_MSS);
438 	}
439 
440 	return 0;
441 }
442 
443 netdev_tx_t hinic_xmit_frame(struct sk_buff *skb, struct net_device *netdev)
444 {
445 	struct hinic_dev *nic_dev = netdev_priv(netdev);
446 	u16 prod_idx, q_id = skb->queue_mapping;
447 	struct netdev_queue *netdev_txq;
448 	int nr_sges, err = NETDEV_TX_OK;
449 	struct hinic_sq_wqe *sq_wqe;
450 	unsigned int wqe_size;
451 	struct hinic_txq *txq;
452 	struct hinic_qp *qp;
453 
454 	txq = &nic_dev->txqs[q_id];
455 	qp = container_of(txq->sq, struct hinic_qp, sq);
456 
457 	if (skb->len < MIN_SKB_LEN) {
458 		if (skb_pad(skb, MIN_SKB_LEN - skb->len)) {
459 			netdev_err(netdev, "Failed to pad skb\n");
460 			goto update_error_stats;
461 		}
462 
463 		skb->len = MIN_SKB_LEN;
464 	}
465 
466 	nr_sges = skb_shinfo(skb)->nr_frags + 1;
467 	if (nr_sges > txq->max_sges) {
468 		netdev_err(netdev, "Too many Tx sges\n");
469 		goto skb_error;
470 	}
471 
472 	err = tx_map_skb(nic_dev, skb, txq->sges);
473 	if (err)
474 		goto skb_error;
475 
476 	wqe_size = HINIC_SQ_WQE_SIZE(nr_sges);
477 
478 	sq_wqe = hinic_sq_get_wqe(txq->sq, wqe_size, &prod_idx);
479 	if (!sq_wqe) {
480 		netif_stop_subqueue(netdev, qp->q_id);
481 
482 		/* Check for the case free_tx_poll is called in another cpu
483 		 * and we stopped the subqueue after free_tx_poll check.
484 		 */
485 		sq_wqe = hinic_sq_get_wqe(txq->sq, wqe_size, &prod_idx);
486 		if (sq_wqe) {
487 			netif_wake_subqueue(nic_dev->netdev, qp->q_id);
488 			goto process_sq_wqe;
489 		}
490 
491 		tx_unmap_skb(nic_dev, skb, txq->sges);
492 
493 		u64_stats_update_begin(&txq->txq_stats.syncp);
494 		txq->txq_stats.tx_busy++;
495 		u64_stats_update_end(&txq->txq_stats.syncp);
496 		err = NETDEV_TX_BUSY;
497 		wqe_size = 0;
498 		goto flush_skbs;
499 	}
500 
501 process_sq_wqe:
502 	hinic_sq_prepare_wqe(txq->sq, prod_idx, sq_wqe, txq->sges, nr_sges);
503 
504 	err = hinic_tx_offload(skb, &sq_wqe->task, &sq_wqe->ctrl.queue_info);
505 	if (err)
506 		goto offload_error;
507 
508 	hinic_sq_write_wqe(txq->sq, prod_idx, sq_wqe, skb, wqe_size);
509 
510 flush_skbs:
511 	netdev_txq = netdev_get_tx_queue(netdev, q_id);
512 	if ((!netdev_xmit_more()) || (netif_xmit_stopped(netdev_txq)))
513 		hinic_sq_write_db(txq->sq, prod_idx, wqe_size, 0);
514 
515 	return err;
516 
517 offload_error:
518 	hinic_sq_return_wqe(txq->sq, wqe_size);
519 	tx_unmap_skb(nic_dev, skb, txq->sges);
520 
521 skb_error:
522 	dev_kfree_skb_any(skb);
523 
524 update_error_stats:
525 	u64_stats_update_begin(&txq->txq_stats.syncp);
526 	txq->txq_stats.tx_dropped++;
527 	u64_stats_update_end(&txq->txq_stats.syncp);
528 
529 	return NETDEV_TX_OK;
530 }
531 
532 /**
533  * tx_free_skb - unmap and free skb
534  * @nic_dev: nic device
535  * @skb: the skb
536  * @sges: the sges that are connected to the skb
537  **/
538 static void tx_free_skb(struct hinic_dev *nic_dev, struct sk_buff *skb,
539 			struct hinic_sge *sges)
540 {
541 	tx_unmap_skb(nic_dev, skb, sges);
542 
543 	dev_kfree_skb_any(skb);
544 }
545 
546 /**
547  * free_all_rx_skbs - free all skbs in tx queue
548  * @txq: tx queue
549  **/
550 static void free_all_tx_skbs(struct hinic_txq *txq)
551 {
552 	struct hinic_dev *nic_dev = netdev_priv(txq->netdev);
553 	struct hinic_sq *sq = txq->sq;
554 	struct hinic_sq_wqe *sq_wqe;
555 	unsigned int wqe_size;
556 	struct sk_buff *skb;
557 	int nr_sges;
558 	u16 ci;
559 
560 	while ((sq_wqe = hinic_sq_read_wqebb(sq, &skb, &wqe_size, &ci))) {
561 		sq_wqe = hinic_sq_read_wqe(sq, &skb, wqe_size, &ci);
562 		if (!sq_wqe)
563 			break;
564 
565 		nr_sges = skb_shinfo(skb)->nr_frags + 1;
566 
567 		hinic_sq_get_sges(sq_wqe, txq->free_sges, nr_sges);
568 
569 		hinic_sq_put_wqe(sq, wqe_size);
570 
571 		tx_free_skb(nic_dev, skb, txq->free_sges);
572 	}
573 }
574 
575 /**
576  * free_tx_poll - free finished tx skbs in tx queue that connected to napi
577  * @napi: napi
578  * @budget: number of tx
579  *
580  * Return 0 - Success, negative - Failure
581  **/
582 static int free_tx_poll(struct napi_struct *napi, int budget)
583 {
584 	struct hinic_txq *txq = container_of(napi, struct hinic_txq, napi);
585 	struct hinic_qp *qp = container_of(txq->sq, struct hinic_qp, sq);
586 	struct hinic_dev *nic_dev = netdev_priv(txq->netdev);
587 	struct netdev_queue *netdev_txq;
588 	struct hinic_sq *sq = txq->sq;
589 	struct hinic_wq *wq = sq->wq;
590 	struct hinic_sq_wqe *sq_wqe;
591 	unsigned int wqe_size;
592 	int nr_sges, pkts = 0;
593 	struct sk_buff *skb;
594 	u64 tx_bytes = 0;
595 	u16 hw_ci, sw_ci;
596 
597 	do {
598 		hw_ci = HW_CONS_IDX(sq) & wq->mask;
599 
600 		/* Reading a WQEBB to get real WQE size and consumer index. */
601 		sq_wqe = hinic_sq_read_wqebb(sq, &skb, &wqe_size, &sw_ci);
602 		if ((!sq_wqe) ||
603 		    (((hw_ci - sw_ci) & wq->mask) * wq->wqebb_size < wqe_size))
604 			break;
605 
606 		/* If this WQE have multiple WQEBBs, we will read again to get
607 		 * full size WQE.
608 		 */
609 		if (wqe_size > wq->wqebb_size) {
610 			sq_wqe = hinic_sq_read_wqe(sq, &skb, wqe_size, &sw_ci);
611 			if (unlikely(!sq_wqe))
612 				break;
613 		}
614 
615 		tx_bytes += skb->len;
616 		pkts++;
617 
618 		nr_sges = skb_shinfo(skb)->nr_frags + 1;
619 
620 		hinic_sq_get_sges(sq_wqe, txq->free_sges, nr_sges);
621 
622 		hinic_sq_put_wqe(sq, wqe_size);
623 
624 		tx_free_skb(nic_dev, skb, txq->free_sges);
625 	} while (pkts < budget);
626 
627 	if (__netif_subqueue_stopped(nic_dev->netdev, qp->q_id) &&
628 	    hinic_get_sq_free_wqebbs(sq) >= HINIC_MIN_TX_NUM_WQEBBS(sq)) {
629 		netdev_txq = netdev_get_tx_queue(txq->netdev, qp->q_id);
630 
631 		__netif_tx_lock(netdev_txq, smp_processor_id());
632 
633 		netif_wake_subqueue(nic_dev->netdev, qp->q_id);
634 
635 		__netif_tx_unlock(netdev_txq);
636 
637 		u64_stats_update_begin(&txq->txq_stats.syncp);
638 		txq->txq_stats.tx_wake++;
639 		u64_stats_update_end(&txq->txq_stats.syncp);
640 	}
641 
642 	u64_stats_update_begin(&txq->txq_stats.syncp);
643 	txq->txq_stats.bytes += tx_bytes;
644 	txq->txq_stats.pkts += pkts;
645 	u64_stats_update_end(&txq->txq_stats.syncp);
646 
647 	if (pkts < budget) {
648 		napi_complete(napi);
649 		hinic_hwdev_set_msix_state(nic_dev->hwdev,
650 					   sq->msix_entry,
651 					   HINIC_MSIX_ENABLE);
652 		return pkts;
653 	}
654 
655 	return budget;
656 }
657 
658 static void tx_napi_add(struct hinic_txq *txq, int weight)
659 {
660 	netif_napi_add(txq->netdev, &txq->napi, free_tx_poll, weight);
661 	napi_enable(&txq->napi);
662 }
663 
664 static void tx_napi_del(struct hinic_txq *txq)
665 {
666 	napi_disable(&txq->napi);
667 	netif_napi_del(&txq->napi);
668 }
669 
670 static irqreturn_t tx_irq(int irq, void *data)
671 {
672 	struct hinic_txq *txq = data;
673 	struct hinic_dev *nic_dev;
674 
675 	nic_dev = netdev_priv(txq->netdev);
676 
677 	/* Disable the interrupt until napi will be completed */
678 	hinic_hwdev_set_msix_state(nic_dev->hwdev,
679 				   txq->sq->msix_entry,
680 				   HINIC_MSIX_DISABLE);
681 
682 	hinic_hwdev_msix_cnt_set(nic_dev->hwdev, txq->sq->msix_entry);
683 
684 	napi_schedule(&txq->napi);
685 	return IRQ_HANDLED;
686 }
687 
688 static int tx_request_irq(struct hinic_txq *txq)
689 {
690 	struct hinic_dev *nic_dev = netdev_priv(txq->netdev);
691 	struct hinic_hwdev *hwdev = nic_dev->hwdev;
692 	struct hinic_hwif *hwif = hwdev->hwif;
693 	struct pci_dev *pdev = hwif->pdev;
694 	struct hinic_sq *sq = txq->sq;
695 	int err;
696 
697 	tx_napi_add(txq, nic_dev->tx_weight);
698 
699 	hinic_hwdev_msix_set(nic_dev->hwdev, sq->msix_entry,
700 			     TX_IRQ_NO_PENDING, TX_IRQ_NO_COALESC,
701 			     TX_IRQ_NO_LLI_TIMER, TX_IRQ_NO_CREDIT,
702 			     TX_IRQ_NO_RESEND_TIMER);
703 
704 	err = request_irq(sq->irq, tx_irq, 0, txq->irq_name, txq);
705 	if (err) {
706 		dev_err(&pdev->dev, "Failed to request Tx irq\n");
707 		tx_napi_del(txq);
708 		return err;
709 	}
710 
711 	return 0;
712 }
713 
714 static void tx_free_irq(struct hinic_txq *txq)
715 {
716 	struct hinic_sq *sq = txq->sq;
717 
718 	free_irq(sq->irq, txq);
719 	tx_napi_del(txq);
720 }
721 
722 /**
723  * hinic_init_txq - Initialize the Tx Queue
724  * @txq: Logical Tx Queue
725  * @sq: Hardware Tx Queue to connect the Logical queue with
726  * @netdev: network device to connect the Logical queue with
727  *
728  * Return 0 - Success, negative - Failure
729  **/
730 int hinic_init_txq(struct hinic_txq *txq, struct hinic_sq *sq,
731 		   struct net_device *netdev)
732 {
733 	struct hinic_qp *qp = container_of(sq, struct hinic_qp, sq);
734 	struct hinic_dev *nic_dev = netdev_priv(netdev);
735 	struct hinic_hwdev *hwdev = nic_dev->hwdev;
736 	int err, irqname_len;
737 	size_t sges_size;
738 
739 	txq->netdev = netdev;
740 	txq->sq = sq;
741 
742 	txq_stats_init(txq);
743 
744 	txq->max_sges = HINIC_MAX_SQ_BUFDESCS;
745 
746 	sges_size = txq->max_sges * sizeof(*txq->sges);
747 	txq->sges = devm_kzalloc(&netdev->dev, sges_size, GFP_KERNEL);
748 	if (!txq->sges)
749 		return -ENOMEM;
750 
751 	sges_size = txq->max_sges * sizeof(*txq->free_sges);
752 	txq->free_sges = devm_kzalloc(&netdev->dev, sges_size, GFP_KERNEL);
753 	if (!txq->free_sges) {
754 		err = -ENOMEM;
755 		goto err_alloc_free_sges;
756 	}
757 
758 	irqname_len = snprintf(NULL, 0, "hinic_txq%d", qp->q_id) + 1;
759 	txq->irq_name = devm_kzalloc(&netdev->dev, irqname_len, GFP_KERNEL);
760 	if (!txq->irq_name) {
761 		err = -ENOMEM;
762 		goto err_alloc_irqname;
763 	}
764 
765 	sprintf(txq->irq_name, "hinic_txq%d", qp->q_id);
766 
767 	err = hinic_hwdev_hw_ci_addr_set(hwdev, sq, CI_UPDATE_NO_PENDING,
768 					 CI_UPDATE_NO_COALESC);
769 	if (err)
770 		goto err_hw_ci;
771 
772 	err = tx_request_irq(txq);
773 	if (err) {
774 		netdev_err(netdev, "Failed to request Tx irq\n");
775 		goto err_req_tx_irq;
776 	}
777 
778 	return 0;
779 
780 err_req_tx_irq:
781 err_hw_ci:
782 	devm_kfree(&netdev->dev, txq->irq_name);
783 
784 err_alloc_irqname:
785 	devm_kfree(&netdev->dev, txq->free_sges);
786 
787 err_alloc_free_sges:
788 	devm_kfree(&netdev->dev, txq->sges);
789 	return err;
790 }
791 
792 /**
793  * hinic_clean_txq - Clean the Tx Queue
794  * @txq: Logical Tx Queue
795  **/
796 void hinic_clean_txq(struct hinic_txq *txq)
797 {
798 	struct net_device *netdev = txq->netdev;
799 
800 	tx_free_irq(txq);
801 
802 	free_all_tx_skbs(txq);
803 
804 	devm_kfree(&netdev->dev, txq->irq_name);
805 	devm_kfree(&netdev->dev, txq->free_sges);
806 	devm_kfree(&netdev->dev, txq->sges);
807 }
808