xref: /linux/drivers/net/ethernet/cavium/thunder/nicvf_queues.c (revision 6ebe6dbd6886af07b102aca42e44edbee94a22d9)
1 /*
2  * Copyright (C) 2015 Cavium, Inc.
3  *
4  * This program is free software; you can redistribute it and/or modify it
5  * under the terms of version 2 of the GNU General Public License
6  * as published by the Free Software Foundation.
7  */
8 
9 #include <linux/pci.h>
10 #include <linux/netdevice.h>
11 #include <linux/ip.h>
12 #include <linux/etherdevice.h>
13 #include <linux/iommu.h>
14 #include <net/ip.h>
15 #include <net/tso.h>
16 
17 #include "nic_reg.h"
18 #include "nic.h"
19 #include "q_struct.h"
20 #include "nicvf_queues.h"
21 
22 static inline void nicvf_sq_add_gather_subdesc(struct snd_queue *sq, int qentry,
23 					       int size, u64 data);
24 static void nicvf_get_page(struct nicvf *nic)
25 {
26 	if (!nic->rb_pageref || !nic->rb_page)
27 		return;
28 
29 	page_ref_add(nic->rb_page, nic->rb_pageref);
30 	nic->rb_pageref = 0;
31 }
32 
33 /* Poll a register for a specific value */
34 static int nicvf_poll_reg(struct nicvf *nic, int qidx,
35 			  u64 reg, int bit_pos, int bits, int val)
36 {
37 	u64 bit_mask;
38 	u64 reg_val;
39 	int timeout = 10;
40 
41 	bit_mask = (1ULL << bits) - 1;
42 	bit_mask = (bit_mask << bit_pos);
43 
44 	while (timeout) {
45 		reg_val = nicvf_queue_reg_read(nic, reg, qidx);
46 		if (((reg_val & bit_mask) >> bit_pos) == val)
47 			return 0;
48 		usleep_range(1000, 2000);
49 		timeout--;
50 	}
51 	netdev_err(nic->netdev, "Poll on reg 0x%llx failed\n", reg);
52 	return 1;
53 }
54 
55 /* Allocate memory for a queue's descriptors */
56 static int nicvf_alloc_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem,
57 				  int q_len, int desc_size, int align_bytes)
58 {
59 	dmem->q_len = q_len;
60 	dmem->size = (desc_size * q_len) + align_bytes;
61 	/* Save address, need it while freeing */
62 	dmem->unalign_base = dma_zalloc_coherent(&nic->pdev->dev, dmem->size,
63 						&dmem->dma, GFP_KERNEL);
64 	if (!dmem->unalign_base)
65 		return -ENOMEM;
66 
67 	/* Align memory address for 'align_bytes' */
68 	dmem->phys_base = NICVF_ALIGNED_ADDR((u64)dmem->dma, align_bytes);
69 	dmem->base = dmem->unalign_base + (dmem->phys_base - dmem->dma);
70 	return 0;
71 }
72 
73 /* Free queue's descriptor memory */
74 static void nicvf_free_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem)
75 {
76 	if (!dmem)
77 		return;
78 
79 	dma_free_coherent(&nic->pdev->dev, dmem->size,
80 			  dmem->unalign_base, dmem->dma);
81 	dmem->unalign_base = NULL;
82 	dmem->base = NULL;
83 }
84 
85 #define XDP_PAGE_REFCNT_REFILL 256
86 
87 /* Allocate a new page or recycle one if possible
88  *
89  * We cannot optimize dma mapping here, since
90  * 1. It's only one RBDR ring for 8 Rx queues.
91  * 2. CQE_RX gives address of the buffer where pkt has been DMA'ed
92  *    and not idx into RBDR ring, so can't refer to saved info.
93  * 3. There are multiple receive buffers per page
94  */
95 static inline struct pgcache *nicvf_alloc_page(struct nicvf *nic,
96 					       struct rbdr *rbdr, gfp_t gfp)
97 {
98 	int ref_count;
99 	struct page *page = NULL;
100 	struct pgcache *pgcache, *next;
101 
102 	/* Check if page is already allocated */
103 	pgcache = &rbdr->pgcache[rbdr->pgidx];
104 	page = pgcache->page;
105 	/* Check if page can be recycled */
106 	if (page) {
107 		ref_count = page_ref_count(page);
108 		/* Check if this page has been used once i.e 'put_page'
109 		 * called after packet transmission i.e internal ref_count
110 		 * and page's ref_count are equal i.e page can be recycled.
111 		 */
112 		if (rbdr->is_xdp && (ref_count == pgcache->ref_count))
113 			pgcache->ref_count--;
114 		else
115 			page = NULL;
116 
117 		/* In non-XDP mode, page's ref_count needs to be '1' for it
118 		 * to be recycled.
119 		 */
120 		if (!rbdr->is_xdp && (ref_count != 1))
121 			page = NULL;
122 	}
123 
124 	if (!page) {
125 		page = alloc_pages(gfp | __GFP_COMP | __GFP_NOWARN, 0);
126 		if (!page)
127 			return NULL;
128 
129 		this_cpu_inc(nic->pnicvf->drv_stats->page_alloc);
130 
131 		/* Check for space */
132 		if (rbdr->pgalloc >= rbdr->pgcnt) {
133 			/* Page can still be used */
134 			nic->rb_page = page;
135 			return NULL;
136 		}
137 
138 		/* Save the page in page cache */
139 		pgcache->page = page;
140 		pgcache->dma_addr = 0;
141 		pgcache->ref_count = 0;
142 		rbdr->pgalloc++;
143 	}
144 
145 	/* Take additional page references for recycling */
146 	if (rbdr->is_xdp) {
147 		/* Since there is single RBDR (i.e single core doing
148 		 * page recycling) per 8 Rx queues, in XDP mode adjusting
149 		 * page references atomically is the biggest bottleneck, so
150 		 * take bunch of references at a time.
151 		 *
152 		 * So here, below reference counts defer by '1'.
153 		 */
154 		if (!pgcache->ref_count) {
155 			pgcache->ref_count = XDP_PAGE_REFCNT_REFILL;
156 			page_ref_add(page, XDP_PAGE_REFCNT_REFILL);
157 		}
158 	} else {
159 		/* In non-XDP case, single 64K page is divided across multiple
160 		 * receive buffers, so cost of recycling is less anyway.
161 		 * So we can do with just one extra reference.
162 		 */
163 		page_ref_add(page, 1);
164 	}
165 
166 	rbdr->pgidx++;
167 	rbdr->pgidx &= (rbdr->pgcnt - 1);
168 
169 	/* Prefetch refcount of next page in page cache */
170 	next = &rbdr->pgcache[rbdr->pgidx];
171 	page = next->page;
172 	if (page)
173 		prefetch(&page->_refcount);
174 
175 	return pgcache;
176 }
177 
178 /* Allocate buffer for packet reception */
179 static inline int nicvf_alloc_rcv_buffer(struct nicvf *nic, struct rbdr *rbdr,
180 					 gfp_t gfp, u32 buf_len, u64 *rbuf)
181 {
182 	struct pgcache *pgcache = NULL;
183 
184 	/* Check if request can be accomodated in previous allocated page.
185 	 * But in XDP mode only one buffer per page is permitted.
186 	 */
187 	if (!rbdr->is_xdp && nic->rb_page &&
188 	    ((nic->rb_page_offset + buf_len) <= PAGE_SIZE)) {
189 		nic->rb_pageref++;
190 		goto ret;
191 	}
192 
193 	nicvf_get_page(nic);
194 	nic->rb_page = NULL;
195 
196 	/* Get new page, either recycled or new one */
197 	pgcache = nicvf_alloc_page(nic, rbdr, gfp);
198 	if (!pgcache && !nic->rb_page) {
199 		this_cpu_inc(nic->pnicvf->drv_stats->rcv_buffer_alloc_failures);
200 		return -ENOMEM;
201 	}
202 
203 	nic->rb_page_offset = 0;
204 
205 	/* Reserve space for header modifications by BPF program */
206 	if (rbdr->is_xdp)
207 		buf_len += XDP_HEADROOM;
208 
209 	/* Check if it's recycled */
210 	if (pgcache)
211 		nic->rb_page = pgcache->page;
212 ret:
213 	if (rbdr->is_xdp && pgcache && pgcache->dma_addr) {
214 		*rbuf = pgcache->dma_addr;
215 	} else {
216 		/* HW will ensure data coherency, CPU sync not required */
217 		*rbuf = (u64)dma_map_page_attrs(&nic->pdev->dev, nic->rb_page,
218 						nic->rb_page_offset, buf_len,
219 						DMA_FROM_DEVICE,
220 						DMA_ATTR_SKIP_CPU_SYNC);
221 		if (dma_mapping_error(&nic->pdev->dev, (dma_addr_t)*rbuf)) {
222 			if (!nic->rb_page_offset)
223 				__free_pages(nic->rb_page, 0);
224 			nic->rb_page = NULL;
225 			return -ENOMEM;
226 		}
227 
228 		if (pgcache)
229 			pgcache->dma_addr = *rbuf + XDP_HEADROOM;
230 		nic->rb_page_offset += buf_len;
231 	}
232 
233 	return 0;
234 }
235 
236 /* Build skb around receive buffer */
237 static struct sk_buff *nicvf_rb_ptr_to_skb(struct nicvf *nic,
238 					   u64 rb_ptr, int len)
239 {
240 	void *data;
241 	struct sk_buff *skb;
242 
243 	data = phys_to_virt(rb_ptr);
244 
245 	/* Now build an skb to give to stack */
246 	skb = build_skb(data, RCV_FRAG_LEN);
247 	if (!skb) {
248 		put_page(virt_to_page(data));
249 		return NULL;
250 	}
251 
252 	prefetch(skb->data);
253 	return skb;
254 }
255 
256 /* Allocate RBDR ring and populate receive buffers */
257 static int  nicvf_init_rbdr(struct nicvf *nic, struct rbdr *rbdr,
258 			    int ring_len, int buf_size)
259 {
260 	int idx;
261 	u64 rbuf;
262 	struct rbdr_entry_t *desc;
263 	int err;
264 
265 	err = nicvf_alloc_q_desc_mem(nic, &rbdr->dmem, ring_len,
266 				     sizeof(struct rbdr_entry_t),
267 				     NICVF_RCV_BUF_ALIGN_BYTES);
268 	if (err)
269 		return err;
270 
271 	rbdr->desc = rbdr->dmem.base;
272 	/* Buffer size has to be in multiples of 128 bytes */
273 	rbdr->dma_size = buf_size;
274 	rbdr->enable = true;
275 	rbdr->thresh = RBDR_THRESH;
276 	rbdr->head = 0;
277 	rbdr->tail = 0;
278 
279 	/* Initialize page recycling stuff.
280 	 *
281 	 * Can't use single buffer per page especially with 64K pages.
282 	 * On embedded platforms i.e 81xx/83xx available memory itself
283 	 * is low and minimum ring size of RBDR is 8K, that takes away
284 	 * lots of memory.
285 	 *
286 	 * But for XDP it has to be a single buffer per page.
287 	 */
288 	if (!nic->pnicvf->xdp_prog) {
289 		rbdr->pgcnt = ring_len / (PAGE_SIZE / buf_size);
290 		rbdr->is_xdp = false;
291 	} else {
292 		rbdr->pgcnt = ring_len;
293 		rbdr->is_xdp = true;
294 	}
295 	rbdr->pgcnt = roundup_pow_of_two(rbdr->pgcnt);
296 	rbdr->pgcache = kzalloc(sizeof(*rbdr->pgcache) *
297 				rbdr->pgcnt, GFP_KERNEL);
298 	if (!rbdr->pgcache)
299 		return -ENOMEM;
300 	rbdr->pgidx = 0;
301 	rbdr->pgalloc = 0;
302 
303 	nic->rb_page = NULL;
304 	for (idx = 0; idx < ring_len; idx++) {
305 		err = nicvf_alloc_rcv_buffer(nic, rbdr, GFP_KERNEL,
306 					     RCV_FRAG_LEN, &rbuf);
307 		if (err) {
308 			/* To free already allocated and mapped ones */
309 			rbdr->tail = idx - 1;
310 			return err;
311 		}
312 
313 		desc = GET_RBDR_DESC(rbdr, idx);
314 		desc->buf_addr = rbuf & ~(NICVF_RCV_BUF_ALIGN_BYTES - 1);
315 	}
316 
317 	nicvf_get_page(nic);
318 
319 	return 0;
320 }
321 
322 /* Free RBDR ring and its receive buffers */
323 static void nicvf_free_rbdr(struct nicvf *nic, struct rbdr *rbdr)
324 {
325 	int head, tail;
326 	u64 buf_addr, phys_addr;
327 	struct pgcache *pgcache;
328 	struct rbdr_entry_t *desc;
329 
330 	if (!rbdr)
331 		return;
332 
333 	rbdr->enable = false;
334 	if (!rbdr->dmem.base)
335 		return;
336 
337 	head = rbdr->head;
338 	tail = rbdr->tail;
339 
340 	/* Release page references */
341 	while (head != tail) {
342 		desc = GET_RBDR_DESC(rbdr, head);
343 		buf_addr = desc->buf_addr;
344 		phys_addr = nicvf_iova_to_phys(nic, buf_addr);
345 		dma_unmap_page_attrs(&nic->pdev->dev, buf_addr, RCV_FRAG_LEN,
346 				     DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
347 		if (phys_addr)
348 			put_page(virt_to_page(phys_to_virt(phys_addr)));
349 		head++;
350 		head &= (rbdr->dmem.q_len - 1);
351 	}
352 	/* Release buffer of tail desc */
353 	desc = GET_RBDR_DESC(rbdr, tail);
354 	buf_addr = desc->buf_addr;
355 	phys_addr = nicvf_iova_to_phys(nic, buf_addr);
356 	dma_unmap_page_attrs(&nic->pdev->dev, buf_addr, RCV_FRAG_LEN,
357 			     DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
358 	if (phys_addr)
359 		put_page(virt_to_page(phys_to_virt(phys_addr)));
360 
361 	/* Sync page cache info */
362 	smp_rmb();
363 
364 	/* Release additional page references held for recycling */
365 	head = 0;
366 	while (head < rbdr->pgcnt) {
367 		pgcache = &rbdr->pgcache[head];
368 		if (pgcache->page && page_ref_count(pgcache->page) != 0) {
369 			if (!rbdr->is_xdp) {
370 				put_page(pgcache->page);
371 				continue;
372 			}
373 			page_ref_sub(pgcache->page, pgcache->ref_count - 1);
374 			put_page(pgcache->page);
375 		}
376 		head++;
377 	}
378 
379 	/* Free RBDR ring */
380 	nicvf_free_q_desc_mem(nic, &rbdr->dmem);
381 }
382 
383 /* Refill receive buffer descriptors with new buffers.
384  */
385 static void nicvf_refill_rbdr(struct nicvf *nic, gfp_t gfp)
386 {
387 	struct queue_set *qs = nic->qs;
388 	int rbdr_idx = qs->rbdr_cnt;
389 	int tail, qcount;
390 	int refill_rb_cnt;
391 	struct rbdr *rbdr;
392 	struct rbdr_entry_t *desc;
393 	u64 rbuf;
394 	int new_rb = 0;
395 
396 refill:
397 	if (!rbdr_idx)
398 		return;
399 	rbdr_idx--;
400 	rbdr = &qs->rbdr[rbdr_idx];
401 	/* Check if it's enabled */
402 	if (!rbdr->enable)
403 		goto next_rbdr;
404 
405 	/* Get no of desc's to be refilled */
406 	qcount = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, rbdr_idx);
407 	qcount &= 0x7FFFF;
408 	/* Doorbell can be ringed with a max of ring size minus 1 */
409 	if (qcount >= (qs->rbdr_len - 1))
410 		goto next_rbdr;
411 	else
412 		refill_rb_cnt = qs->rbdr_len - qcount - 1;
413 
414 	/* Sync page cache info */
415 	smp_rmb();
416 
417 	/* Start filling descs from tail */
418 	tail = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_TAIL, rbdr_idx) >> 3;
419 	while (refill_rb_cnt) {
420 		tail++;
421 		tail &= (rbdr->dmem.q_len - 1);
422 
423 		if (nicvf_alloc_rcv_buffer(nic, rbdr, gfp, RCV_FRAG_LEN, &rbuf))
424 			break;
425 
426 		desc = GET_RBDR_DESC(rbdr, tail);
427 		desc->buf_addr = rbuf & ~(NICVF_RCV_BUF_ALIGN_BYTES - 1);
428 		refill_rb_cnt--;
429 		new_rb++;
430 	}
431 
432 	nicvf_get_page(nic);
433 
434 	/* make sure all memory stores are done before ringing doorbell */
435 	smp_wmb();
436 
437 	/* Check if buffer allocation failed */
438 	if (refill_rb_cnt)
439 		nic->rb_alloc_fail = true;
440 	else
441 		nic->rb_alloc_fail = false;
442 
443 	/* Notify HW */
444 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR,
445 			      rbdr_idx, new_rb);
446 next_rbdr:
447 	/* Re-enable RBDR interrupts only if buffer allocation is success */
448 	if (!nic->rb_alloc_fail && rbdr->enable &&
449 	    netif_running(nic->pnicvf->netdev))
450 		nicvf_enable_intr(nic, NICVF_INTR_RBDR, rbdr_idx);
451 
452 	if (rbdr_idx)
453 		goto refill;
454 }
455 
456 /* Alloc rcv buffers in non-atomic mode for better success */
457 void nicvf_rbdr_work(struct work_struct *work)
458 {
459 	struct nicvf *nic = container_of(work, struct nicvf, rbdr_work.work);
460 
461 	nicvf_refill_rbdr(nic, GFP_KERNEL);
462 	if (nic->rb_alloc_fail)
463 		schedule_delayed_work(&nic->rbdr_work, msecs_to_jiffies(10));
464 	else
465 		nic->rb_work_scheduled = false;
466 }
467 
468 /* In Softirq context, alloc rcv buffers in atomic mode */
469 void nicvf_rbdr_task(unsigned long data)
470 {
471 	struct nicvf *nic = (struct nicvf *)data;
472 
473 	nicvf_refill_rbdr(nic, GFP_ATOMIC);
474 	if (nic->rb_alloc_fail) {
475 		nic->rb_work_scheduled = true;
476 		schedule_delayed_work(&nic->rbdr_work, msecs_to_jiffies(10));
477 	}
478 }
479 
480 /* Initialize completion queue */
481 static int nicvf_init_cmp_queue(struct nicvf *nic,
482 				struct cmp_queue *cq, int q_len)
483 {
484 	int err;
485 
486 	err = nicvf_alloc_q_desc_mem(nic, &cq->dmem, q_len, CMP_QUEUE_DESC_SIZE,
487 				     NICVF_CQ_BASE_ALIGN_BYTES);
488 	if (err)
489 		return err;
490 
491 	cq->desc = cq->dmem.base;
492 	cq->thresh = pass1_silicon(nic->pdev) ? 0 : CMP_QUEUE_CQE_THRESH;
493 	nic->cq_coalesce_usecs = (CMP_QUEUE_TIMER_THRESH * 0.05) - 1;
494 
495 	return 0;
496 }
497 
498 static void nicvf_free_cmp_queue(struct nicvf *nic, struct cmp_queue *cq)
499 {
500 	if (!cq)
501 		return;
502 	if (!cq->dmem.base)
503 		return;
504 
505 	nicvf_free_q_desc_mem(nic, &cq->dmem);
506 }
507 
508 /* Initialize transmit queue */
509 static int nicvf_init_snd_queue(struct nicvf *nic,
510 				struct snd_queue *sq, int q_len, int qidx)
511 {
512 	int err;
513 
514 	err = nicvf_alloc_q_desc_mem(nic, &sq->dmem, q_len, SND_QUEUE_DESC_SIZE,
515 				     NICVF_SQ_BASE_ALIGN_BYTES);
516 	if (err)
517 		return err;
518 
519 	sq->desc = sq->dmem.base;
520 	sq->skbuff = kcalloc(q_len, sizeof(u64), GFP_KERNEL);
521 	if (!sq->skbuff)
522 		return -ENOMEM;
523 
524 	sq->head = 0;
525 	sq->tail = 0;
526 	sq->thresh = SND_QUEUE_THRESH;
527 
528 	/* Check if this SQ is a XDP TX queue */
529 	if (nic->sqs_mode)
530 		qidx += ((nic->sqs_id + 1) * MAX_SND_QUEUES_PER_QS);
531 	if (qidx < nic->pnicvf->xdp_tx_queues) {
532 		/* Alloc memory to save page pointers for XDP_TX */
533 		sq->xdp_page = kcalloc(q_len, sizeof(u64), GFP_KERNEL);
534 		if (!sq->xdp_page)
535 			return -ENOMEM;
536 		sq->xdp_desc_cnt = 0;
537 		sq->xdp_free_cnt = q_len - 1;
538 		sq->is_xdp = true;
539 	} else {
540 		sq->xdp_page = NULL;
541 		sq->xdp_desc_cnt = 0;
542 		sq->xdp_free_cnt = 0;
543 		sq->is_xdp = false;
544 
545 		atomic_set(&sq->free_cnt, q_len - 1);
546 
547 		/* Preallocate memory for TSO segment's header */
548 		sq->tso_hdrs = dma_alloc_coherent(&nic->pdev->dev,
549 						  q_len * TSO_HEADER_SIZE,
550 						  &sq->tso_hdrs_phys,
551 						  GFP_KERNEL);
552 		if (!sq->tso_hdrs)
553 			return -ENOMEM;
554 	}
555 
556 	return 0;
557 }
558 
559 void nicvf_unmap_sndq_buffers(struct nicvf *nic, struct snd_queue *sq,
560 			      int hdr_sqe, u8 subdesc_cnt)
561 {
562 	u8 idx;
563 	struct sq_gather_subdesc *gather;
564 
565 	/* Unmap DMA mapped skb data buffers */
566 	for (idx = 0; idx < subdesc_cnt; idx++) {
567 		hdr_sqe++;
568 		hdr_sqe &= (sq->dmem.q_len - 1);
569 		gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, hdr_sqe);
570 		/* HW will ensure data coherency, CPU sync not required */
571 		dma_unmap_page_attrs(&nic->pdev->dev, gather->addr,
572 				     gather->size, DMA_TO_DEVICE,
573 				     DMA_ATTR_SKIP_CPU_SYNC);
574 	}
575 }
576 
577 static void nicvf_free_snd_queue(struct nicvf *nic, struct snd_queue *sq)
578 {
579 	struct sk_buff *skb;
580 	struct page *page;
581 	struct sq_hdr_subdesc *hdr;
582 	struct sq_hdr_subdesc *tso_sqe;
583 
584 	if (!sq)
585 		return;
586 	if (!sq->dmem.base)
587 		return;
588 
589 	if (sq->tso_hdrs)
590 		dma_free_coherent(&nic->pdev->dev,
591 				  sq->dmem.q_len * TSO_HEADER_SIZE,
592 				  sq->tso_hdrs, sq->tso_hdrs_phys);
593 
594 	/* Free pending skbs in the queue */
595 	smp_rmb();
596 	while (sq->head != sq->tail) {
597 		skb = (struct sk_buff *)sq->skbuff[sq->head];
598 		if (!skb || !sq->xdp_page)
599 			goto next;
600 
601 		page = (struct page *)sq->xdp_page[sq->head];
602 		if (!page)
603 			goto next;
604 		else
605 			put_page(page);
606 
607 		hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, sq->head);
608 		/* Check for dummy descriptor used for HW TSO offload on 88xx */
609 		if (hdr->dont_send) {
610 			/* Get actual TSO descriptors and unmap them */
611 			tso_sqe =
612 			 (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, hdr->rsvd2);
613 			nicvf_unmap_sndq_buffers(nic, sq, hdr->rsvd2,
614 						 tso_sqe->subdesc_cnt);
615 		} else {
616 			nicvf_unmap_sndq_buffers(nic, sq, sq->head,
617 						 hdr->subdesc_cnt);
618 		}
619 		if (skb)
620 			dev_kfree_skb_any(skb);
621 next:
622 		sq->head++;
623 		sq->head &= (sq->dmem.q_len - 1);
624 	}
625 	kfree(sq->skbuff);
626 	kfree(sq->xdp_page);
627 	nicvf_free_q_desc_mem(nic, &sq->dmem);
628 }
629 
630 static void nicvf_reclaim_snd_queue(struct nicvf *nic,
631 				    struct queue_set *qs, int qidx)
632 {
633 	/* Disable send queue */
634 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, 0);
635 	/* Check if SQ is stopped */
636 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_SQ_0_7_STATUS, 21, 1, 0x01))
637 		return;
638 	/* Reset send queue */
639 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET);
640 }
641 
642 static void nicvf_reclaim_rcv_queue(struct nicvf *nic,
643 				    struct queue_set *qs, int qidx)
644 {
645 	union nic_mbx mbx = {};
646 
647 	/* Make sure all packets in the pipeline are written back into mem */
648 	mbx.msg.msg = NIC_MBOX_MSG_RQ_SW_SYNC;
649 	nicvf_send_msg_to_pf(nic, &mbx);
650 }
651 
652 static void nicvf_reclaim_cmp_queue(struct nicvf *nic,
653 				    struct queue_set *qs, int qidx)
654 {
655 	/* Disable timer threshold (doesn't get reset upon CQ reset */
656 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2, qidx, 0);
657 	/* Disable completion queue */
658 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, 0);
659 	/* Reset completion queue */
660 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET);
661 }
662 
663 static void nicvf_reclaim_rbdr(struct nicvf *nic,
664 			       struct rbdr *rbdr, int qidx)
665 {
666 	u64 tmp, fifo_state;
667 	int timeout = 10;
668 
669 	/* Save head and tail pointers for feeing up buffers */
670 	rbdr->head = nicvf_queue_reg_read(nic,
671 					  NIC_QSET_RBDR_0_1_HEAD,
672 					  qidx) >> 3;
673 	rbdr->tail = nicvf_queue_reg_read(nic,
674 					  NIC_QSET_RBDR_0_1_TAIL,
675 					  qidx) >> 3;
676 
677 	/* If RBDR FIFO is in 'FAIL' state then do a reset first
678 	 * before relaiming.
679 	 */
680 	fifo_state = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, qidx);
681 	if (((fifo_state >> 62) & 0x03) == 0x3)
682 		nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
683 				      qidx, NICVF_RBDR_RESET);
684 
685 	/* Disable RBDR */
686 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0);
687 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00))
688 		return;
689 	while (1) {
690 		tmp = nicvf_queue_reg_read(nic,
691 					   NIC_QSET_RBDR_0_1_PREFETCH_STATUS,
692 					   qidx);
693 		if ((tmp & 0xFFFFFFFF) == ((tmp >> 32) & 0xFFFFFFFF))
694 			break;
695 		usleep_range(1000, 2000);
696 		timeout--;
697 		if (!timeout) {
698 			netdev_err(nic->netdev,
699 				   "Failed polling on prefetch status\n");
700 			return;
701 		}
702 	}
703 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
704 			      qidx, NICVF_RBDR_RESET);
705 
706 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x02))
707 		return;
708 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0x00);
709 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00))
710 		return;
711 }
712 
713 void nicvf_config_vlan_stripping(struct nicvf *nic, netdev_features_t features)
714 {
715 	u64 rq_cfg;
716 	int sqs;
717 
718 	rq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_RQ_GEN_CFG, 0);
719 
720 	/* Enable first VLAN stripping */
721 	if (features & NETIF_F_HW_VLAN_CTAG_RX)
722 		rq_cfg |= (1ULL << 25);
723 	else
724 		rq_cfg &= ~(1ULL << 25);
725 	nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0, rq_cfg);
726 
727 	/* Configure Secondary Qsets, if any */
728 	for (sqs = 0; sqs < nic->sqs_count; sqs++)
729 		if (nic->snicvf[sqs])
730 			nicvf_queue_reg_write(nic->snicvf[sqs],
731 					      NIC_QSET_RQ_GEN_CFG, 0, rq_cfg);
732 }
733 
734 static void nicvf_reset_rcv_queue_stats(struct nicvf *nic)
735 {
736 	union nic_mbx mbx = {};
737 
738 	/* Reset all RQ/SQ and VF stats */
739 	mbx.reset_stat.msg = NIC_MBOX_MSG_RESET_STAT_COUNTER;
740 	mbx.reset_stat.rx_stat_mask = 0x3FFF;
741 	mbx.reset_stat.tx_stat_mask = 0x1F;
742 	mbx.reset_stat.rq_stat_mask = 0xFFFF;
743 	mbx.reset_stat.sq_stat_mask = 0xFFFF;
744 	nicvf_send_msg_to_pf(nic, &mbx);
745 }
746 
747 /* Configures receive queue */
748 static void nicvf_rcv_queue_config(struct nicvf *nic, struct queue_set *qs,
749 				   int qidx, bool enable)
750 {
751 	union nic_mbx mbx = {};
752 	struct rcv_queue *rq;
753 	struct rq_cfg rq_cfg;
754 
755 	rq = &qs->rq[qidx];
756 	rq->enable = enable;
757 
758 	/* Disable receive queue */
759 	nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, 0);
760 
761 	if (!rq->enable) {
762 		nicvf_reclaim_rcv_queue(nic, qs, qidx);
763 		return;
764 	}
765 
766 	rq->cq_qs = qs->vnic_id;
767 	rq->cq_idx = qidx;
768 	rq->start_rbdr_qs = qs->vnic_id;
769 	rq->start_qs_rbdr_idx = qs->rbdr_cnt - 1;
770 	rq->cont_rbdr_qs = qs->vnic_id;
771 	rq->cont_qs_rbdr_idx = qs->rbdr_cnt - 1;
772 	/* all writes of RBDR data to be loaded into L2 Cache as well*/
773 	rq->caching = 1;
774 
775 	/* Send a mailbox msg to PF to config RQ */
776 	mbx.rq.msg = NIC_MBOX_MSG_RQ_CFG;
777 	mbx.rq.qs_num = qs->vnic_id;
778 	mbx.rq.rq_num = qidx;
779 	mbx.rq.cfg = (rq->caching << 26) | (rq->cq_qs << 19) |
780 			  (rq->cq_idx << 16) | (rq->cont_rbdr_qs << 9) |
781 			  (rq->cont_qs_rbdr_idx << 8) |
782 			  (rq->start_rbdr_qs << 1) | (rq->start_qs_rbdr_idx);
783 	nicvf_send_msg_to_pf(nic, &mbx);
784 
785 	mbx.rq.msg = NIC_MBOX_MSG_RQ_BP_CFG;
786 	mbx.rq.cfg = BIT_ULL(63) | BIT_ULL(62) |
787 		     (RQ_PASS_RBDR_LVL << 16) | (RQ_PASS_CQ_LVL << 8) |
788 		     (qs->vnic_id << 0);
789 	nicvf_send_msg_to_pf(nic, &mbx);
790 
791 	/* RQ drop config
792 	 * Enable CQ drop to reserve sufficient CQEs for all tx packets
793 	 */
794 	mbx.rq.msg = NIC_MBOX_MSG_RQ_DROP_CFG;
795 	mbx.rq.cfg = BIT_ULL(63) | BIT_ULL(62) |
796 		     (RQ_PASS_RBDR_LVL << 40) | (RQ_DROP_RBDR_LVL << 32) |
797 		     (RQ_PASS_CQ_LVL << 16) | (RQ_DROP_CQ_LVL << 8);
798 	nicvf_send_msg_to_pf(nic, &mbx);
799 
800 	if (!nic->sqs_mode && (qidx == 0)) {
801 		/* Enable checking L3/L4 length and TCP/UDP checksums
802 		 * Also allow IPv6 pkts with zero UDP checksum.
803 		 */
804 		nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0,
805 				      (BIT(24) | BIT(23) | BIT(21) | BIT(20)));
806 		nicvf_config_vlan_stripping(nic, nic->netdev->features);
807 	}
808 
809 	/* Enable Receive queue */
810 	memset(&rq_cfg, 0, sizeof(struct rq_cfg));
811 	rq_cfg.ena = 1;
812 	rq_cfg.tcp_ena = 0;
813 	nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, *(u64 *)&rq_cfg);
814 }
815 
816 /* Configures completion queue */
817 void nicvf_cmp_queue_config(struct nicvf *nic, struct queue_set *qs,
818 			    int qidx, bool enable)
819 {
820 	struct cmp_queue *cq;
821 	struct cq_cfg cq_cfg;
822 
823 	cq = &qs->cq[qidx];
824 	cq->enable = enable;
825 
826 	if (!cq->enable) {
827 		nicvf_reclaim_cmp_queue(nic, qs, qidx);
828 		return;
829 	}
830 
831 	/* Reset completion queue */
832 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET);
833 
834 	if (!cq->enable)
835 		return;
836 
837 	spin_lock_init(&cq->lock);
838 	/* Set completion queue base address */
839 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_BASE,
840 			      qidx, (u64)(cq->dmem.phys_base));
841 
842 	/* Enable Completion queue */
843 	memset(&cq_cfg, 0, sizeof(struct cq_cfg));
844 	cq_cfg.ena = 1;
845 	cq_cfg.reset = 0;
846 	cq_cfg.caching = 0;
847 	cq_cfg.qsize = ilog2(qs->cq_len >> 10);
848 	cq_cfg.avg_con = 0;
849 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, *(u64 *)&cq_cfg);
850 
851 	/* Set threshold value for interrupt generation */
852 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_THRESH, qidx, cq->thresh);
853 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2,
854 			      qidx, CMP_QUEUE_TIMER_THRESH);
855 }
856 
857 /* Configures transmit queue */
858 static void nicvf_snd_queue_config(struct nicvf *nic, struct queue_set *qs,
859 				   int qidx, bool enable)
860 {
861 	union nic_mbx mbx = {};
862 	struct snd_queue *sq;
863 	struct sq_cfg sq_cfg;
864 
865 	sq = &qs->sq[qidx];
866 	sq->enable = enable;
867 
868 	if (!sq->enable) {
869 		nicvf_reclaim_snd_queue(nic, qs, qidx);
870 		return;
871 	}
872 
873 	/* Reset send queue */
874 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET);
875 
876 	sq->cq_qs = qs->vnic_id;
877 	sq->cq_idx = qidx;
878 
879 	/* Send a mailbox msg to PF to config SQ */
880 	mbx.sq.msg = NIC_MBOX_MSG_SQ_CFG;
881 	mbx.sq.qs_num = qs->vnic_id;
882 	mbx.sq.sq_num = qidx;
883 	mbx.sq.sqs_mode = nic->sqs_mode;
884 	mbx.sq.cfg = (sq->cq_qs << 3) | sq->cq_idx;
885 	nicvf_send_msg_to_pf(nic, &mbx);
886 
887 	/* Set queue base address */
888 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_BASE,
889 			      qidx, (u64)(sq->dmem.phys_base));
890 
891 	/* Enable send queue  & set queue size */
892 	memset(&sq_cfg, 0, sizeof(struct sq_cfg));
893 	sq_cfg.ena = 1;
894 	sq_cfg.reset = 0;
895 	sq_cfg.ldwb = 0;
896 	sq_cfg.qsize = ilog2(qs->sq_len >> 10);
897 	sq_cfg.tstmp_bgx_intf = 0;
898 	/* CQ's level at which HW will stop processing SQEs to avoid
899 	 * transmitting a pkt with no space in CQ to post CQE_TX.
900 	 */
901 	sq_cfg.cq_limit = (CMP_QUEUE_PIPELINE_RSVD * 256) / qs->cq_len;
902 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, *(u64 *)&sq_cfg);
903 
904 	/* Set threshold value for interrupt generation */
905 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_THRESH, qidx, sq->thresh);
906 
907 	/* Set queue:cpu affinity for better load distribution */
908 	if (cpu_online(qidx)) {
909 		cpumask_set_cpu(qidx, &sq->affinity_mask);
910 		netif_set_xps_queue(nic->netdev,
911 				    &sq->affinity_mask, qidx);
912 	}
913 }
914 
915 /* Configures receive buffer descriptor ring */
916 static void nicvf_rbdr_config(struct nicvf *nic, struct queue_set *qs,
917 			      int qidx, bool enable)
918 {
919 	struct rbdr *rbdr;
920 	struct rbdr_cfg rbdr_cfg;
921 
922 	rbdr = &qs->rbdr[qidx];
923 	nicvf_reclaim_rbdr(nic, rbdr, qidx);
924 	if (!enable)
925 		return;
926 
927 	/* Set descriptor base address */
928 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_BASE,
929 			      qidx, (u64)(rbdr->dmem.phys_base));
930 
931 	/* Enable RBDR  & set queue size */
932 	/* Buffer size should be in multiples of 128 bytes */
933 	memset(&rbdr_cfg, 0, sizeof(struct rbdr_cfg));
934 	rbdr_cfg.ena = 1;
935 	rbdr_cfg.reset = 0;
936 	rbdr_cfg.ldwb = 0;
937 	rbdr_cfg.qsize = RBDR_SIZE;
938 	rbdr_cfg.avg_con = 0;
939 	rbdr_cfg.lines = rbdr->dma_size / 128;
940 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
941 			      qidx, *(u64 *)&rbdr_cfg);
942 
943 	/* Notify HW */
944 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR,
945 			      qidx, qs->rbdr_len - 1);
946 
947 	/* Set threshold value for interrupt generation */
948 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_THRESH,
949 			      qidx, rbdr->thresh - 1);
950 }
951 
952 /* Requests PF to assign and enable Qset */
953 void nicvf_qset_config(struct nicvf *nic, bool enable)
954 {
955 	union nic_mbx mbx = {};
956 	struct queue_set *qs = nic->qs;
957 	struct qs_cfg *qs_cfg;
958 
959 	if (!qs) {
960 		netdev_warn(nic->netdev,
961 			    "Qset is still not allocated, don't init queues\n");
962 		return;
963 	}
964 
965 	qs->enable = enable;
966 	qs->vnic_id = nic->vf_id;
967 
968 	/* Send a mailbox msg to PF to config Qset */
969 	mbx.qs.msg = NIC_MBOX_MSG_QS_CFG;
970 	mbx.qs.num = qs->vnic_id;
971 	mbx.qs.sqs_count = nic->sqs_count;
972 
973 	mbx.qs.cfg = 0;
974 	qs_cfg = (struct qs_cfg *)&mbx.qs.cfg;
975 	if (qs->enable) {
976 		qs_cfg->ena = 1;
977 #ifdef __BIG_ENDIAN
978 		qs_cfg->be = 1;
979 #endif
980 		qs_cfg->vnic = qs->vnic_id;
981 	}
982 	nicvf_send_msg_to_pf(nic, &mbx);
983 }
984 
985 static void nicvf_free_resources(struct nicvf *nic)
986 {
987 	int qidx;
988 	struct queue_set *qs = nic->qs;
989 
990 	/* Free receive buffer descriptor ring */
991 	for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
992 		nicvf_free_rbdr(nic, &qs->rbdr[qidx]);
993 
994 	/* Free completion queue */
995 	for (qidx = 0; qidx < qs->cq_cnt; qidx++)
996 		nicvf_free_cmp_queue(nic, &qs->cq[qidx]);
997 
998 	/* Free send queue */
999 	for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1000 		nicvf_free_snd_queue(nic, &qs->sq[qidx]);
1001 }
1002 
1003 static int nicvf_alloc_resources(struct nicvf *nic)
1004 {
1005 	int qidx;
1006 	struct queue_set *qs = nic->qs;
1007 
1008 	/* Alloc receive buffer descriptor ring */
1009 	for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) {
1010 		if (nicvf_init_rbdr(nic, &qs->rbdr[qidx], qs->rbdr_len,
1011 				    DMA_BUFFER_LEN))
1012 			goto alloc_fail;
1013 	}
1014 
1015 	/* Alloc send queue */
1016 	for (qidx = 0; qidx < qs->sq_cnt; qidx++) {
1017 		if (nicvf_init_snd_queue(nic, &qs->sq[qidx], qs->sq_len, qidx))
1018 			goto alloc_fail;
1019 	}
1020 
1021 	/* Alloc completion queue */
1022 	for (qidx = 0; qidx < qs->cq_cnt; qidx++) {
1023 		if (nicvf_init_cmp_queue(nic, &qs->cq[qidx], qs->cq_len))
1024 			goto alloc_fail;
1025 	}
1026 
1027 	return 0;
1028 alloc_fail:
1029 	nicvf_free_resources(nic);
1030 	return -ENOMEM;
1031 }
1032 
1033 int nicvf_set_qset_resources(struct nicvf *nic)
1034 {
1035 	struct queue_set *qs;
1036 
1037 	qs = devm_kzalloc(&nic->pdev->dev, sizeof(*qs), GFP_KERNEL);
1038 	if (!qs)
1039 		return -ENOMEM;
1040 	nic->qs = qs;
1041 
1042 	/* Set count of each queue */
1043 	qs->rbdr_cnt = DEFAULT_RBDR_CNT;
1044 	qs->rq_cnt = min_t(u8, MAX_RCV_QUEUES_PER_QS, num_online_cpus());
1045 	qs->sq_cnt = min_t(u8, MAX_SND_QUEUES_PER_QS, num_online_cpus());
1046 	qs->cq_cnt = max_t(u8, qs->rq_cnt, qs->sq_cnt);
1047 
1048 	/* Set queue lengths */
1049 	qs->rbdr_len = RCV_BUF_COUNT;
1050 	qs->sq_len = SND_QUEUE_LEN;
1051 	qs->cq_len = CMP_QUEUE_LEN;
1052 
1053 	nic->rx_queues = qs->rq_cnt;
1054 	nic->tx_queues = qs->sq_cnt;
1055 	nic->xdp_tx_queues = 0;
1056 
1057 	return 0;
1058 }
1059 
1060 int nicvf_config_data_transfer(struct nicvf *nic, bool enable)
1061 {
1062 	bool disable = false;
1063 	struct queue_set *qs = nic->qs;
1064 	struct queue_set *pqs = nic->pnicvf->qs;
1065 	int qidx;
1066 
1067 	if (!qs)
1068 		return 0;
1069 
1070 	/* Take primary VF's queue lengths.
1071 	 * This is needed to take queue lengths set from ethtool
1072 	 * into consideration.
1073 	 */
1074 	if (nic->sqs_mode && pqs) {
1075 		qs->cq_len = pqs->cq_len;
1076 		qs->sq_len = pqs->sq_len;
1077 	}
1078 
1079 	if (enable) {
1080 		if (nicvf_alloc_resources(nic))
1081 			return -ENOMEM;
1082 
1083 		for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1084 			nicvf_snd_queue_config(nic, qs, qidx, enable);
1085 		for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1086 			nicvf_cmp_queue_config(nic, qs, qidx, enable);
1087 		for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
1088 			nicvf_rbdr_config(nic, qs, qidx, enable);
1089 		for (qidx = 0; qidx < qs->rq_cnt; qidx++)
1090 			nicvf_rcv_queue_config(nic, qs, qidx, enable);
1091 	} else {
1092 		for (qidx = 0; qidx < qs->rq_cnt; qidx++)
1093 			nicvf_rcv_queue_config(nic, qs, qidx, disable);
1094 		for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
1095 			nicvf_rbdr_config(nic, qs, qidx, disable);
1096 		for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1097 			nicvf_snd_queue_config(nic, qs, qidx, disable);
1098 		for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1099 			nicvf_cmp_queue_config(nic, qs, qidx, disable);
1100 
1101 		nicvf_free_resources(nic);
1102 	}
1103 
1104 	/* Reset RXQ's stats.
1105 	 * SQ's stats will get reset automatically once SQ is reset.
1106 	 */
1107 	nicvf_reset_rcv_queue_stats(nic);
1108 
1109 	return 0;
1110 }
1111 
1112 /* Get a free desc from SQ
1113  * returns descriptor ponter & descriptor number
1114  */
1115 static inline int nicvf_get_sq_desc(struct snd_queue *sq, int desc_cnt)
1116 {
1117 	int qentry;
1118 
1119 	qentry = sq->tail;
1120 	if (!sq->is_xdp)
1121 		atomic_sub(desc_cnt, &sq->free_cnt);
1122 	else
1123 		sq->xdp_free_cnt -= desc_cnt;
1124 	sq->tail += desc_cnt;
1125 	sq->tail &= (sq->dmem.q_len - 1);
1126 
1127 	return qentry;
1128 }
1129 
1130 /* Rollback to previous tail pointer when descriptors not used */
1131 static inline void nicvf_rollback_sq_desc(struct snd_queue *sq,
1132 					  int qentry, int desc_cnt)
1133 {
1134 	sq->tail = qentry;
1135 	atomic_add(desc_cnt, &sq->free_cnt);
1136 }
1137 
1138 /* Free descriptor back to SQ for future use */
1139 void nicvf_put_sq_desc(struct snd_queue *sq, int desc_cnt)
1140 {
1141 	if (!sq->is_xdp)
1142 		atomic_add(desc_cnt, &sq->free_cnt);
1143 	else
1144 		sq->xdp_free_cnt += desc_cnt;
1145 	sq->head += desc_cnt;
1146 	sq->head &= (sq->dmem.q_len - 1);
1147 }
1148 
1149 static inline int nicvf_get_nxt_sqentry(struct snd_queue *sq, int qentry)
1150 {
1151 	qentry++;
1152 	qentry &= (sq->dmem.q_len - 1);
1153 	return qentry;
1154 }
1155 
1156 void nicvf_sq_enable(struct nicvf *nic, struct snd_queue *sq, int qidx)
1157 {
1158 	u64 sq_cfg;
1159 
1160 	sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx);
1161 	sq_cfg |= NICVF_SQ_EN;
1162 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg);
1163 	/* Ring doorbell so that H/W restarts processing SQEs */
1164 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR, qidx, 0);
1165 }
1166 
1167 void nicvf_sq_disable(struct nicvf *nic, int qidx)
1168 {
1169 	u64 sq_cfg;
1170 
1171 	sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx);
1172 	sq_cfg &= ~NICVF_SQ_EN;
1173 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg);
1174 }
1175 
1176 void nicvf_sq_free_used_descs(struct net_device *netdev, struct snd_queue *sq,
1177 			      int qidx)
1178 {
1179 	u64 head, tail;
1180 	struct sk_buff *skb;
1181 	struct nicvf *nic = netdev_priv(netdev);
1182 	struct sq_hdr_subdesc *hdr;
1183 
1184 	head = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_HEAD, qidx) >> 4;
1185 	tail = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_TAIL, qidx) >> 4;
1186 	while (sq->head != head) {
1187 		hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, sq->head);
1188 		if (hdr->subdesc_type != SQ_DESC_TYPE_HEADER) {
1189 			nicvf_put_sq_desc(sq, 1);
1190 			continue;
1191 		}
1192 		skb = (struct sk_buff *)sq->skbuff[sq->head];
1193 		if (skb)
1194 			dev_kfree_skb_any(skb);
1195 		atomic64_add(1, (atomic64_t *)&netdev->stats.tx_packets);
1196 		atomic64_add(hdr->tot_len,
1197 			     (atomic64_t *)&netdev->stats.tx_bytes);
1198 		nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1);
1199 	}
1200 }
1201 
1202 /* XDP Transmit APIs */
1203 void nicvf_xdp_sq_doorbell(struct nicvf *nic,
1204 			   struct snd_queue *sq, int sq_num)
1205 {
1206 	if (!sq->xdp_desc_cnt)
1207 		return;
1208 
1209 	/* make sure all memory stores are done before ringing doorbell */
1210 	wmb();
1211 
1212 	/* Inform HW to xmit all TSO segments */
1213 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR,
1214 			      sq_num, sq->xdp_desc_cnt);
1215 	sq->xdp_desc_cnt = 0;
1216 }
1217 
1218 static inline void
1219 nicvf_xdp_sq_add_hdr_subdesc(struct snd_queue *sq, int qentry,
1220 			     int subdesc_cnt, u64 data, int len)
1221 {
1222 	struct sq_hdr_subdesc *hdr;
1223 
1224 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1225 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1226 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1227 	hdr->subdesc_cnt = subdesc_cnt;
1228 	hdr->tot_len = len;
1229 	hdr->post_cqe = 1;
1230 	sq->xdp_page[qentry] = (u64)virt_to_page((void *)data);
1231 }
1232 
1233 int nicvf_xdp_sq_append_pkt(struct nicvf *nic, struct snd_queue *sq,
1234 			    u64 bufaddr, u64 dma_addr, u16 len)
1235 {
1236 	int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT;
1237 	int qentry;
1238 
1239 	if (subdesc_cnt > sq->xdp_free_cnt)
1240 		return -1;
1241 
1242 	qentry = nicvf_get_sq_desc(sq, subdesc_cnt);
1243 
1244 	nicvf_xdp_sq_add_hdr_subdesc(sq, qentry, subdesc_cnt - 1, bufaddr, len);
1245 
1246 	qentry = nicvf_get_nxt_sqentry(sq, qentry);
1247 	nicvf_sq_add_gather_subdesc(sq, qentry, len, dma_addr);
1248 
1249 	sq->xdp_desc_cnt += subdesc_cnt;
1250 
1251 	return 0;
1252 }
1253 
1254 /* Calculate no of SQ subdescriptors needed to transmit all
1255  * segments of this TSO packet.
1256  * Taken from 'Tilera network driver' with a minor modification.
1257  */
1258 static int nicvf_tso_count_subdescs(struct sk_buff *skb)
1259 {
1260 	struct skb_shared_info *sh = skb_shinfo(skb);
1261 	unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1262 	unsigned int data_len = skb->len - sh_len;
1263 	unsigned int p_len = sh->gso_size;
1264 	long f_id = -1;    /* id of the current fragment */
1265 	long f_size = skb_headlen(skb) - sh_len;  /* current fragment size */
1266 	long f_used = 0;  /* bytes used from the current fragment */
1267 	long n;            /* size of the current piece of payload */
1268 	int num_edescs = 0;
1269 	int segment;
1270 
1271 	for (segment = 0; segment < sh->gso_segs; segment++) {
1272 		unsigned int p_used = 0;
1273 
1274 		/* One edesc for header and for each piece of the payload. */
1275 		for (num_edescs++; p_used < p_len; num_edescs++) {
1276 			/* Advance as needed. */
1277 			while (f_used >= f_size) {
1278 				f_id++;
1279 				f_size = skb_frag_size(&sh->frags[f_id]);
1280 				f_used = 0;
1281 			}
1282 
1283 			/* Use bytes from the current fragment. */
1284 			n = p_len - p_used;
1285 			if (n > f_size - f_used)
1286 				n = f_size - f_used;
1287 			f_used += n;
1288 			p_used += n;
1289 		}
1290 
1291 		/* The last segment may be less than gso_size. */
1292 		data_len -= p_len;
1293 		if (data_len < p_len)
1294 			p_len = data_len;
1295 	}
1296 
1297 	/* '+ gso_segs' for SQ_HDR_SUDESCs for each segment */
1298 	return num_edescs + sh->gso_segs;
1299 }
1300 
1301 #define POST_CQE_DESC_COUNT 2
1302 
1303 /* Get the number of SQ descriptors needed to xmit this skb */
1304 static int nicvf_sq_subdesc_required(struct nicvf *nic, struct sk_buff *skb)
1305 {
1306 	int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT;
1307 
1308 	if (skb_shinfo(skb)->gso_size && !nic->hw_tso) {
1309 		subdesc_cnt = nicvf_tso_count_subdescs(skb);
1310 		return subdesc_cnt;
1311 	}
1312 
1313 	/* Dummy descriptors to get TSO pkt completion notification */
1314 	if (nic->t88 && nic->hw_tso && skb_shinfo(skb)->gso_size)
1315 		subdesc_cnt += POST_CQE_DESC_COUNT;
1316 
1317 	if (skb_shinfo(skb)->nr_frags)
1318 		subdesc_cnt += skb_shinfo(skb)->nr_frags;
1319 
1320 	return subdesc_cnt;
1321 }
1322 
1323 /* Add SQ HEADER subdescriptor.
1324  * First subdescriptor for every send descriptor.
1325  */
1326 static inline void
1327 nicvf_sq_add_hdr_subdesc(struct nicvf *nic, struct snd_queue *sq, int qentry,
1328 			 int subdesc_cnt, struct sk_buff *skb, int len)
1329 {
1330 	int proto;
1331 	struct sq_hdr_subdesc *hdr;
1332 	union {
1333 		struct iphdr *v4;
1334 		struct ipv6hdr *v6;
1335 		unsigned char *hdr;
1336 	} ip;
1337 
1338 	ip.hdr = skb_network_header(skb);
1339 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1340 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1341 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1342 
1343 	if (nic->t88 && nic->hw_tso && skb_shinfo(skb)->gso_size) {
1344 		/* post_cqe = 0, to avoid HW posting a CQE for every TSO
1345 		 * segment transmitted on 88xx.
1346 		 */
1347 		hdr->subdesc_cnt = subdesc_cnt - POST_CQE_DESC_COUNT;
1348 	} else {
1349 		sq->skbuff[qentry] = (u64)skb;
1350 		/* Enable notification via CQE after processing SQE */
1351 		hdr->post_cqe = 1;
1352 		/* No of subdescriptors following this */
1353 		hdr->subdesc_cnt = subdesc_cnt;
1354 	}
1355 	hdr->tot_len = len;
1356 
1357 	/* Offload checksum calculation to HW */
1358 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
1359 		if (ip.v4->version == 4)
1360 			hdr->csum_l3 = 1; /* Enable IP csum calculation */
1361 		hdr->l3_offset = skb_network_offset(skb);
1362 		hdr->l4_offset = skb_transport_offset(skb);
1363 
1364 		proto = (ip.v4->version == 4) ? ip.v4->protocol :
1365 			ip.v6->nexthdr;
1366 
1367 		switch (proto) {
1368 		case IPPROTO_TCP:
1369 			hdr->csum_l4 = SEND_L4_CSUM_TCP;
1370 			break;
1371 		case IPPROTO_UDP:
1372 			hdr->csum_l4 = SEND_L4_CSUM_UDP;
1373 			break;
1374 		case IPPROTO_SCTP:
1375 			hdr->csum_l4 = SEND_L4_CSUM_SCTP;
1376 			break;
1377 		}
1378 	}
1379 
1380 	if (nic->hw_tso && skb_shinfo(skb)->gso_size) {
1381 		hdr->tso = 1;
1382 		hdr->tso_start = skb_transport_offset(skb) + tcp_hdrlen(skb);
1383 		hdr->tso_max_paysize = skb_shinfo(skb)->gso_size;
1384 		/* For non-tunneled pkts, point this to L2 ethertype */
1385 		hdr->inner_l3_offset = skb_network_offset(skb) - 2;
1386 		this_cpu_inc(nic->pnicvf->drv_stats->tx_tso);
1387 	}
1388 }
1389 
1390 /* SQ GATHER subdescriptor
1391  * Must follow HDR descriptor
1392  */
1393 static inline void nicvf_sq_add_gather_subdesc(struct snd_queue *sq, int qentry,
1394 					       int size, u64 data)
1395 {
1396 	struct sq_gather_subdesc *gather;
1397 
1398 	qentry &= (sq->dmem.q_len - 1);
1399 	gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, qentry);
1400 
1401 	memset(gather, 0, SND_QUEUE_DESC_SIZE);
1402 	gather->subdesc_type = SQ_DESC_TYPE_GATHER;
1403 	gather->ld_type = NIC_SEND_LD_TYPE_E_LDD;
1404 	gather->size = size;
1405 	gather->addr = data;
1406 }
1407 
1408 /* Add HDR + IMMEDIATE subdescriptors right after descriptors of a TSO
1409  * packet so that a CQE is posted as a notifation for transmission of
1410  * TSO packet.
1411  */
1412 static inline void nicvf_sq_add_cqe_subdesc(struct snd_queue *sq, int qentry,
1413 					    int tso_sqe, struct sk_buff *skb)
1414 {
1415 	struct sq_imm_subdesc *imm;
1416 	struct sq_hdr_subdesc *hdr;
1417 
1418 	sq->skbuff[qentry] = (u64)skb;
1419 
1420 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1421 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1422 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1423 	/* Enable notification via CQE after processing SQE */
1424 	hdr->post_cqe = 1;
1425 	/* There is no packet to transmit here */
1426 	hdr->dont_send = 1;
1427 	hdr->subdesc_cnt = POST_CQE_DESC_COUNT - 1;
1428 	hdr->tot_len = 1;
1429 	/* Actual TSO header SQE index, needed for cleanup */
1430 	hdr->rsvd2 = tso_sqe;
1431 
1432 	qentry = nicvf_get_nxt_sqentry(sq, qentry);
1433 	imm = (struct sq_imm_subdesc *)GET_SQ_DESC(sq, qentry);
1434 	memset(imm, 0, SND_QUEUE_DESC_SIZE);
1435 	imm->subdesc_type = SQ_DESC_TYPE_IMMEDIATE;
1436 	imm->len = 1;
1437 }
1438 
1439 static inline void nicvf_sq_doorbell(struct nicvf *nic, struct sk_buff *skb,
1440 				     int sq_num, int desc_cnt)
1441 {
1442 	struct netdev_queue *txq;
1443 
1444 	txq = netdev_get_tx_queue(nic->pnicvf->netdev,
1445 				  skb_get_queue_mapping(skb));
1446 
1447 	netdev_tx_sent_queue(txq, skb->len);
1448 
1449 	/* make sure all memory stores are done before ringing doorbell */
1450 	smp_wmb();
1451 
1452 	/* Inform HW to xmit all TSO segments */
1453 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR,
1454 			      sq_num, desc_cnt);
1455 }
1456 
1457 /* Segment a TSO packet into 'gso_size' segments and append
1458  * them to SQ for transfer
1459  */
1460 static int nicvf_sq_append_tso(struct nicvf *nic, struct snd_queue *sq,
1461 			       int sq_num, int qentry, struct sk_buff *skb)
1462 {
1463 	struct tso_t tso;
1464 	int seg_subdescs = 0, desc_cnt = 0;
1465 	int seg_len, total_len, data_left;
1466 	int hdr_qentry = qentry;
1467 	int hdr_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1468 
1469 	tso_start(skb, &tso);
1470 	total_len = skb->len - hdr_len;
1471 	while (total_len > 0) {
1472 		char *hdr;
1473 
1474 		/* Save Qentry for adding HDR_SUBDESC at the end */
1475 		hdr_qentry = qentry;
1476 
1477 		data_left = min_t(int, skb_shinfo(skb)->gso_size, total_len);
1478 		total_len -= data_left;
1479 
1480 		/* Add segment's header */
1481 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1482 		hdr = sq->tso_hdrs + qentry * TSO_HEADER_SIZE;
1483 		tso_build_hdr(skb, hdr, &tso, data_left, total_len == 0);
1484 		nicvf_sq_add_gather_subdesc(sq, qentry, hdr_len,
1485 					    sq->tso_hdrs_phys +
1486 					    qentry * TSO_HEADER_SIZE);
1487 		/* HDR_SUDESC + GATHER */
1488 		seg_subdescs = 2;
1489 		seg_len = hdr_len;
1490 
1491 		/* Add segment's payload fragments */
1492 		while (data_left > 0) {
1493 			int size;
1494 
1495 			size = min_t(int, tso.size, data_left);
1496 
1497 			qentry = nicvf_get_nxt_sqentry(sq, qentry);
1498 			nicvf_sq_add_gather_subdesc(sq, qentry, size,
1499 						    virt_to_phys(tso.data));
1500 			seg_subdescs++;
1501 			seg_len += size;
1502 
1503 			data_left -= size;
1504 			tso_build_data(skb, &tso, size);
1505 		}
1506 		nicvf_sq_add_hdr_subdesc(nic, sq, hdr_qentry,
1507 					 seg_subdescs - 1, skb, seg_len);
1508 		sq->skbuff[hdr_qentry] = (u64)NULL;
1509 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1510 
1511 		desc_cnt += seg_subdescs;
1512 	}
1513 	/* Save SKB in the last segment for freeing */
1514 	sq->skbuff[hdr_qentry] = (u64)skb;
1515 
1516 	nicvf_sq_doorbell(nic, skb, sq_num, desc_cnt);
1517 
1518 	this_cpu_inc(nic->pnicvf->drv_stats->tx_tso);
1519 	return 1;
1520 }
1521 
1522 /* Append an skb to a SQ for packet transfer. */
1523 int nicvf_sq_append_skb(struct nicvf *nic, struct snd_queue *sq,
1524 			struct sk_buff *skb, u8 sq_num)
1525 {
1526 	int i, size;
1527 	int subdesc_cnt, hdr_sqe = 0;
1528 	int qentry;
1529 	u64 dma_addr;
1530 
1531 	subdesc_cnt = nicvf_sq_subdesc_required(nic, skb);
1532 	if (subdesc_cnt > atomic_read(&sq->free_cnt))
1533 		goto append_fail;
1534 
1535 	qentry = nicvf_get_sq_desc(sq, subdesc_cnt);
1536 
1537 	/* Check if its a TSO packet */
1538 	if (skb_shinfo(skb)->gso_size && !nic->hw_tso)
1539 		return nicvf_sq_append_tso(nic, sq, sq_num, qentry, skb);
1540 
1541 	/* Add SQ header subdesc */
1542 	nicvf_sq_add_hdr_subdesc(nic, sq, qentry, subdesc_cnt - 1,
1543 				 skb, skb->len);
1544 	hdr_sqe = qentry;
1545 
1546 	/* Add SQ gather subdescs */
1547 	qentry = nicvf_get_nxt_sqentry(sq, qentry);
1548 	size = skb_is_nonlinear(skb) ? skb_headlen(skb) : skb->len;
1549 	/* HW will ensure data coherency, CPU sync not required */
1550 	dma_addr = dma_map_page_attrs(&nic->pdev->dev, virt_to_page(skb->data),
1551 				      offset_in_page(skb->data), size,
1552 				      DMA_TO_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
1553 	if (dma_mapping_error(&nic->pdev->dev, dma_addr)) {
1554 		nicvf_rollback_sq_desc(sq, qentry, subdesc_cnt);
1555 		return 0;
1556 	}
1557 
1558 	nicvf_sq_add_gather_subdesc(sq, qentry, size, dma_addr);
1559 
1560 	/* Check for scattered buffer */
1561 	if (!skb_is_nonlinear(skb))
1562 		goto doorbell;
1563 
1564 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1565 		const struct skb_frag_struct *frag;
1566 
1567 		frag = &skb_shinfo(skb)->frags[i];
1568 
1569 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1570 		size = skb_frag_size(frag);
1571 		dma_addr = dma_map_page_attrs(&nic->pdev->dev,
1572 					      skb_frag_page(frag),
1573 					      frag->page_offset, size,
1574 					      DMA_TO_DEVICE,
1575 					      DMA_ATTR_SKIP_CPU_SYNC);
1576 		if (dma_mapping_error(&nic->pdev->dev, dma_addr)) {
1577 			/* Free entire chain of mapped buffers
1578 			 * here 'i' = frags mapped + above mapped skb->data
1579 			 */
1580 			nicvf_unmap_sndq_buffers(nic, sq, hdr_sqe, i);
1581 			nicvf_rollback_sq_desc(sq, qentry, subdesc_cnt);
1582 			return 0;
1583 		}
1584 		nicvf_sq_add_gather_subdesc(sq, qentry, size, dma_addr);
1585 	}
1586 
1587 doorbell:
1588 	if (nic->t88 && skb_shinfo(skb)->gso_size) {
1589 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1590 		nicvf_sq_add_cqe_subdesc(sq, qentry, hdr_sqe, skb);
1591 	}
1592 
1593 	nicvf_sq_doorbell(nic, skb, sq_num, subdesc_cnt);
1594 
1595 	return 1;
1596 
1597 append_fail:
1598 	/* Use original PCI dev for debug log */
1599 	nic = nic->pnicvf;
1600 	netdev_dbg(nic->netdev, "Not enough SQ descriptors to xmit pkt\n");
1601 	return 0;
1602 }
1603 
1604 static inline unsigned frag_num(unsigned i)
1605 {
1606 #ifdef __BIG_ENDIAN
1607 	return (i & ~3) + 3 - (i & 3);
1608 #else
1609 	return i;
1610 #endif
1611 }
1612 
1613 static void nicvf_unmap_rcv_buffer(struct nicvf *nic, u64 dma_addr,
1614 				   u64 buf_addr, bool xdp)
1615 {
1616 	struct page *page = NULL;
1617 	int len = RCV_FRAG_LEN;
1618 
1619 	if (xdp) {
1620 		page = virt_to_page(phys_to_virt(buf_addr));
1621 		/* Check if it's a recycled page, if not
1622 		 * unmap the DMA mapping.
1623 		 *
1624 		 * Recycled page holds an extra reference.
1625 		 */
1626 		if (page_ref_count(page) != 1)
1627 			return;
1628 
1629 		len += XDP_HEADROOM;
1630 		/* Receive buffers in XDP mode are mapped from page start */
1631 		dma_addr &= PAGE_MASK;
1632 	}
1633 	dma_unmap_page_attrs(&nic->pdev->dev, dma_addr, len,
1634 			     DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
1635 }
1636 
1637 /* Returns SKB for a received packet */
1638 struct sk_buff *nicvf_get_rcv_skb(struct nicvf *nic,
1639 				  struct cqe_rx_t *cqe_rx, bool xdp)
1640 {
1641 	int frag;
1642 	int payload_len = 0;
1643 	struct sk_buff *skb = NULL;
1644 	struct page *page;
1645 	int offset;
1646 	u16 *rb_lens = NULL;
1647 	u64 *rb_ptrs = NULL;
1648 	u64 phys_addr;
1649 
1650 	rb_lens = (void *)cqe_rx + (3 * sizeof(u64));
1651 	/* Except 88xx pass1 on all other chips CQE_RX2_S is added to
1652 	 * CQE_RX at word6, hence buffer pointers move by word
1653 	 *
1654 	 * Use existing 'hw_tso' flag which will be set for all chips
1655 	 * except 88xx pass1 instead of a additional cache line
1656 	 * access (or miss) by using pci dev's revision.
1657 	 */
1658 	if (!nic->hw_tso)
1659 		rb_ptrs = (void *)cqe_rx + (6 * sizeof(u64));
1660 	else
1661 		rb_ptrs = (void *)cqe_rx + (7 * sizeof(u64));
1662 
1663 	for (frag = 0; frag < cqe_rx->rb_cnt; frag++) {
1664 		payload_len = rb_lens[frag_num(frag)];
1665 		phys_addr = nicvf_iova_to_phys(nic, *rb_ptrs);
1666 		if (!phys_addr) {
1667 			if (skb)
1668 				dev_kfree_skb_any(skb);
1669 			return NULL;
1670 		}
1671 
1672 		if (!frag) {
1673 			/* First fragment */
1674 			nicvf_unmap_rcv_buffer(nic,
1675 					       *rb_ptrs - cqe_rx->align_pad,
1676 					       phys_addr, xdp);
1677 			skb = nicvf_rb_ptr_to_skb(nic,
1678 						  phys_addr - cqe_rx->align_pad,
1679 						  payload_len);
1680 			if (!skb)
1681 				return NULL;
1682 			skb_reserve(skb, cqe_rx->align_pad);
1683 			skb_put(skb, payload_len);
1684 		} else {
1685 			/* Add fragments */
1686 			nicvf_unmap_rcv_buffer(nic, *rb_ptrs, phys_addr, xdp);
1687 			page = virt_to_page(phys_to_virt(phys_addr));
1688 			offset = phys_to_virt(phys_addr) - page_address(page);
1689 			skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page,
1690 					offset, payload_len, RCV_FRAG_LEN);
1691 		}
1692 		/* Next buffer pointer */
1693 		rb_ptrs++;
1694 	}
1695 	return skb;
1696 }
1697 
1698 static u64 nicvf_int_type_to_mask(int int_type, int q_idx)
1699 {
1700 	u64 reg_val;
1701 
1702 	switch (int_type) {
1703 	case NICVF_INTR_CQ:
1704 		reg_val = ((1ULL << q_idx) << NICVF_INTR_CQ_SHIFT);
1705 		break;
1706 	case NICVF_INTR_SQ:
1707 		reg_val = ((1ULL << q_idx) << NICVF_INTR_SQ_SHIFT);
1708 		break;
1709 	case NICVF_INTR_RBDR:
1710 		reg_val = ((1ULL << q_idx) << NICVF_INTR_RBDR_SHIFT);
1711 		break;
1712 	case NICVF_INTR_PKT_DROP:
1713 		reg_val = (1ULL << NICVF_INTR_PKT_DROP_SHIFT);
1714 		break;
1715 	case NICVF_INTR_TCP_TIMER:
1716 		reg_val = (1ULL << NICVF_INTR_TCP_TIMER_SHIFT);
1717 		break;
1718 	case NICVF_INTR_MBOX:
1719 		reg_val = (1ULL << NICVF_INTR_MBOX_SHIFT);
1720 		break;
1721 	case NICVF_INTR_QS_ERR:
1722 		reg_val = (1ULL << NICVF_INTR_QS_ERR_SHIFT);
1723 		break;
1724 	default:
1725 		reg_val = 0;
1726 	}
1727 
1728 	return reg_val;
1729 }
1730 
1731 /* Enable interrupt */
1732 void nicvf_enable_intr(struct nicvf *nic, int int_type, int q_idx)
1733 {
1734 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1735 
1736 	if (!mask) {
1737 		netdev_dbg(nic->netdev,
1738 			   "Failed to enable interrupt: unknown type\n");
1739 		return;
1740 	}
1741 	nicvf_reg_write(nic, NIC_VF_ENA_W1S,
1742 			nicvf_reg_read(nic, NIC_VF_ENA_W1S) | mask);
1743 }
1744 
1745 /* Disable interrupt */
1746 void nicvf_disable_intr(struct nicvf *nic, int int_type, int q_idx)
1747 {
1748 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1749 
1750 	if (!mask) {
1751 		netdev_dbg(nic->netdev,
1752 			   "Failed to disable interrupt: unknown type\n");
1753 		return;
1754 	}
1755 
1756 	nicvf_reg_write(nic, NIC_VF_ENA_W1C, mask);
1757 }
1758 
1759 /* Clear interrupt */
1760 void nicvf_clear_intr(struct nicvf *nic, int int_type, int q_idx)
1761 {
1762 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1763 
1764 	if (!mask) {
1765 		netdev_dbg(nic->netdev,
1766 			   "Failed to clear interrupt: unknown type\n");
1767 		return;
1768 	}
1769 
1770 	nicvf_reg_write(nic, NIC_VF_INT, mask);
1771 }
1772 
1773 /* Check if interrupt is enabled */
1774 int nicvf_is_intr_enabled(struct nicvf *nic, int int_type, int q_idx)
1775 {
1776 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1777 	/* If interrupt type is unknown, we treat it disabled. */
1778 	if (!mask) {
1779 		netdev_dbg(nic->netdev,
1780 			   "Failed to check interrupt enable: unknown type\n");
1781 		return 0;
1782 	}
1783 
1784 	return mask & nicvf_reg_read(nic, NIC_VF_ENA_W1S);
1785 }
1786 
1787 void nicvf_update_rq_stats(struct nicvf *nic, int rq_idx)
1788 {
1789 	struct rcv_queue *rq;
1790 
1791 #define GET_RQ_STATS(reg) \
1792 	nicvf_reg_read(nic, NIC_QSET_RQ_0_7_STAT_0_1 |\
1793 			    (rq_idx << NIC_Q_NUM_SHIFT) | (reg << 3))
1794 
1795 	rq = &nic->qs->rq[rq_idx];
1796 	rq->stats.bytes = GET_RQ_STATS(RQ_SQ_STATS_OCTS);
1797 	rq->stats.pkts = GET_RQ_STATS(RQ_SQ_STATS_PKTS);
1798 }
1799 
1800 void nicvf_update_sq_stats(struct nicvf *nic, int sq_idx)
1801 {
1802 	struct snd_queue *sq;
1803 
1804 #define GET_SQ_STATS(reg) \
1805 	nicvf_reg_read(nic, NIC_QSET_SQ_0_7_STAT_0_1 |\
1806 			    (sq_idx << NIC_Q_NUM_SHIFT) | (reg << 3))
1807 
1808 	sq = &nic->qs->sq[sq_idx];
1809 	sq->stats.bytes = GET_SQ_STATS(RQ_SQ_STATS_OCTS);
1810 	sq->stats.pkts = GET_SQ_STATS(RQ_SQ_STATS_PKTS);
1811 }
1812 
1813 /* Check for errors in the receive cmp.queue entry */
1814 int nicvf_check_cqe_rx_errs(struct nicvf *nic, struct cqe_rx_t *cqe_rx)
1815 {
1816 	netif_err(nic, rx_err, nic->netdev,
1817 		  "RX error CQE err_level 0x%x err_opcode 0x%x\n",
1818 		  cqe_rx->err_level, cqe_rx->err_opcode);
1819 
1820 	switch (cqe_rx->err_opcode) {
1821 	case CQ_RX_ERROP_RE_PARTIAL:
1822 		this_cpu_inc(nic->drv_stats->rx_bgx_truncated_pkts);
1823 		break;
1824 	case CQ_RX_ERROP_RE_JABBER:
1825 		this_cpu_inc(nic->drv_stats->rx_jabber_errs);
1826 		break;
1827 	case CQ_RX_ERROP_RE_FCS:
1828 		this_cpu_inc(nic->drv_stats->rx_fcs_errs);
1829 		break;
1830 	case CQ_RX_ERROP_RE_RX_CTL:
1831 		this_cpu_inc(nic->drv_stats->rx_bgx_errs);
1832 		break;
1833 	case CQ_RX_ERROP_PREL2_ERR:
1834 		this_cpu_inc(nic->drv_stats->rx_prel2_errs);
1835 		break;
1836 	case CQ_RX_ERROP_L2_MAL:
1837 		this_cpu_inc(nic->drv_stats->rx_l2_hdr_malformed);
1838 		break;
1839 	case CQ_RX_ERROP_L2_OVERSIZE:
1840 		this_cpu_inc(nic->drv_stats->rx_oversize);
1841 		break;
1842 	case CQ_RX_ERROP_L2_UNDERSIZE:
1843 		this_cpu_inc(nic->drv_stats->rx_undersize);
1844 		break;
1845 	case CQ_RX_ERROP_L2_LENMISM:
1846 		this_cpu_inc(nic->drv_stats->rx_l2_len_mismatch);
1847 		break;
1848 	case CQ_RX_ERROP_L2_PCLP:
1849 		this_cpu_inc(nic->drv_stats->rx_l2_pclp);
1850 		break;
1851 	case CQ_RX_ERROP_IP_NOT:
1852 		this_cpu_inc(nic->drv_stats->rx_ip_ver_errs);
1853 		break;
1854 	case CQ_RX_ERROP_IP_CSUM_ERR:
1855 		this_cpu_inc(nic->drv_stats->rx_ip_csum_errs);
1856 		break;
1857 	case CQ_RX_ERROP_IP_MAL:
1858 		this_cpu_inc(nic->drv_stats->rx_ip_hdr_malformed);
1859 		break;
1860 	case CQ_RX_ERROP_IP_MALD:
1861 		this_cpu_inc(nic->drv_stats->rx_ip_payload_malformed);
1862 		break;
1863 	case CQ_RX_ERROP_IP_HOP:
1864 		this_cpu_inc(nic->drv_stats->rx_ip_ttl_errs);
1865 		break;
1866 	case CQ_RX_ERROP_L3_PCLP:
1867 		this_cpu_inc(nic->drv_stats->rx_l3_pclp);
1868 		break;
1869 	case CQ_RX_ERROP_L4_MAL:
1870 		this_cpu_inc(nic->drv_stats->rx_l4_malformed);
1871 		break;
1872 	case CQ_RX_ERROP_L4_CHK:
1873 		this_cpu_inc(nic->drv_stats->rx_l4_csum_errs);
1874 		break;
1875 	case CQ_RX_ERROP_UDP_LEN:
1876 		this_cpu_inc(nic->drv_stats->rx_udp_len_errs);
1877 		break;
1878 	case CQ_RX_ERROP_L4_PORT:
1879 		this_cpu_inc(nic->drv_stats->rx_l4_port_errs);
1880 		break;
1881 	case CQ_RX_ERROP_TCP_FLAG:
1882 		this_cpu_inc(nic->drv_stats->rx_tcp_flag_errs);
1883 		break;
1884 	case CQ_RX_ERROP_TCP_OFFSET:
1885 		this_cpu_inc(nic->drv_stats->rx_tcp_offset_errs);
1886 		break;
1887 	case CQ_RX_ERROP_L4_PCLP:
1888 		this_cpu_inc(nic->drv_stats->rx_l4_pclp);
1889 		break;
1890 	case CQ_RX_ERROP_RBDR_TRUNC:
1891 		this_cpu_inc(nic->drv_stats->rx_truncated_pkts);
1892 		break;
1893 	}
1894 
1895 	return 1;
1896 }
1897 
1898 /* Check for errors in the send cmp.queue entry */
1899 int nicvf_check_cqe_tx_errs(struct nicvf *nic, struct cqe_send_t *cqe_tx)
1900 {
1901 	switch (cqe_tx->send_status) {
1902 	case CQ_TX_ERROP_DESC_FAULT:
1903 		this_cpu_inc(nic->drv_stats->tx_desc_fault);
1904 		break;
1905 	case CQ_TX_ERROP_HDR_CONS_ERR:
1906 		this_cpu_inc(nic->drv_stats->tx_hdr_cons_err);
1907 		break;
1908 	case CQ_TX_ERROP_SUBDC_ERR:
1909 		this_cpu_inc(nic->drv_stats->tx_subdesc_err);
1910 		break;
1911 	case CQ_TX_ERROP_MAX_SIZE_VIOL:
1912 		this_cpu_inc(nic->drv_stats->tx_max_size_exceeded);
1913 		break;
1914 	case CQ_TX_ERROP_IMM_SIZE_OFLOW:
1915 		this_cpu_inc(nic->drv_stats->tx_imm_size_oflow);
1916 		break;
1917 	case CQ_TX_ERROP_DATA_SEQUENCE_ERR:
1918 		this_cpu_inc(nic->drv_stats->tx_data_seq_err);
1919 		break;
1920 	case CQ_TX_ERROP_MEM_SEQUENCE_ERR:
1921 		this_cpu_inc(nic->drv_stats->tx_mem_seq_err);
1922 		break;
1923 	case CQ_TX_ERROP_LOCK_VIOL:
1924 		this_cpu_inc(nic->drv_stats->tx_lock_viol);
1925 		break;
1926 	case CQ_TX_ERROP_DATA_FAULT:
1927 		this_cpu_inc(nic->drv_stats->tx_data_fault);
1928 		break;
1929 	case CQ_TX_ERROP_TSTMP_CONFLICT:
1930 		this_cpu_inc(nic->drv_stats->tx_tstmp_conflict);
1931 		break;
1932 	case CQ_TX_ERROP_TSTMP_TIMEOUT:
1933 		this_cpu_inc(nic->drv_stats->tx_tstmp_timeout);
1934 		break;
1935 	case CQ_TX_ERROP_MEM_FAULT:
1936 		this_cpu_inc(nic->drv_stats->tx_mem_fault);
1937 		break;
1938 	case CQ_TX_ERROP_CK_OVERLAP:
1939 		this_cpu_inc(nic->drv_stats->tx_csum_overlap);
1940 		break;
1941 	case CQ_TX_ERROP_CK_OFLOW:
1942 		this_cpu_inc(nic->drv_stats->tx_csum_overflow);
1943 		break;
1944 	}
1945 
1946 	return 1;
1947 }
1948