xref: /openbmc/linux/drivers/net/ethernet/cavium/thunder/nicvf_queues.c (revision 9977a8c3497a8f7f7f951994f298a8e4d961234f)
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 		xdp_rxq_info_unreg(&rq->xdp_rxq);
764 		return;
765 	}
766 
767 	rq->cq_qs = qs->vnic_id;
768 	rq->cq_idx = qidx;
769 	rq->start_rbdr_qs = qs->vnic_id;
770 	rq->start_qs_rbdr_idx = qs->rbdr_cnt - 1;
771 	rq->cont_rbdr_qs = qs->vnic_id;
772 	rq->cont_qs_rbdr_idx = qs->rbdr_cnt - 1;
773 	/* all writes of RBDR data to be loaded into L2 Cache as well*/
774 	rq->caching = 1;
775 
776 	/* Driver have no proper error path for failed XDP RX-queue info reg */
777 	WARN_ON(xdp_rxq_info_reg(&rq->xdp_rxq, nic->netdev, qidx) < 0);
778 
779 	/* Send a mailbox msg to PF to config RQ */
780 	mbx.rq.msg = NIC_MBOX_MSG_RQ_CFG;
781 	mbx.rq.qs_num = qs->vnic_id;
782 	mbx.rq.rq_num = qidx;
783 	mbx.rq.cfg = (rq->caching << 26) | (rq->cq_qs << 19) |
784 			  (rq->cq_idx << 16) | (rq->cont_rbdr_qs << 9) |
785 			  (rq->cont_qs_rbdr_idx << 8) |
786 			  (rq->start_rbdr_qs << 1) | (rq->start_qs_rbdr_idx);
787 	nicvf_send_msg_to_pf(nic, &mbx);
788 
789 	mbx.rq.msg = NIC_MBOX_MSG_RQ_BP_CFG;
790 	mbx.rq.cfg = BIT_ULL(63) | BIT_ULL(62) |
791 		     (RQ_PASS_RBDR_LVL << 16) | (RQ_PASS_CQ_LVL << 8) |
792 		     (qs->vnic_id << 0);
793 	nicvf_send_msg_to_pf(nic, &mbx);
794 
795 	/* RQ drop config
796 	 * Enable CQ drop to reserve sufficient CQEs for all tx packets
797 	 */
798 	mbx.rq.msg = NIC_MBOX_MSG_RQ_DROP_CFG;
799 	mbx.rq.cfg = BIT_ULL(63) | BIT_ULL(62) |
800 		     (RQ_PASS_RBDR_LVL << 40) | (RQ_DROP_RBDR_LVL << 32) |
801 		     (RQ_PASS_CQ_LVL << 16) | (RQ_DROP_CQ_LVL << 8);
802 	nicvf_send_msg_to_pf(nic, &mbx);
803 
804 	if (!nic->sqs_mode && (qidx == 0)) {
805 		/* Enable checking L3/L4 length and TCP/UDP checksums
806 		 * Also allow IPv6 pkts with zero UDP checksum.
807 		 */
808 		nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0,
809 				      (BIT(24) | BIT(23) | BIT(21) | BIT(20)));
810 		nicvf_config_vlan_stripping(nic, nic->netdev->features);
811 	}
812 
813 	/* Enable Receive queue */
814 	memset(&rq_cfg, 0, sizeof(struct rq_cfg));
815 	rq_cfg.ena = 1;
816 	rq_cfg.tcp_ena = 0;
817 	nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, *(u64 *)&rq_cfg);
818 }
819 
820 /* Configures completion queue */
821 void nicvf_cmp_queue_config(struct nicvf *nic, struct queue_set *qs,
822 			    int qidx, bool enable)
823 {
824 	struct cmp_queue *cq;
825 	struct cq_cfg cq_cfg;
826 
827 	cq = &qs->cq[qidx];
828 	cq->enable = enable;
829 
830 	if (!cq->enable) {
831 		nicvf_reclaim_cmp_queue(nic, qs, qidx);
832 		return;
833 	}
834 
835 	/* Reset completion queue */
836 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET);
837 
838 	if (!cq->enable)
839 		return;
840 
841 	spin_lock_init(&cq->lock);
842 	/* Set completion queue base address */
843 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_BASE,
844 			      qidx, (u64)(cq->dmem.phys_base));
845 
846 	/* Enable Completion queue */
847 	memset(&cq_cfg, 0, sizeof(struct cq_cfg));
848 	cq_cfg.ena = 1;
849 	cq_cfg.reset = 0;
850 	cq_cfg.caching = 0;
851 	cq_cfg.qsize = ilog2(qs->cq_len >> 10);
852 	cq_cfg.avg_con = 0;
853 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, *(u64 *)&cq_cfg);
854 
855 	/* Set threshold value for interrupt generation */
856 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_THRESH, qidx, cq->thresh);
857 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2,
858 			      qidx, CMP_QUEUE_TIMER_THRESH);
859 }
860 
861 /* Configures transmit queue */
862 static void nicvf_snd_queue_config(struct nicvf *nic, struct queue_set *qs,
863 				   int qidx, bool enable)
864 {
865 	union nic_mbx mbx = {};
866 	struct snd_queue *sq;
867 	struct sq_cfg sq_cfg;
868 
869 	sq = &qs->sq[qidx];
870 	sq->enable = enable;
871 
872 	if (!sq->enable) {
873 		nicvf_reclaim_snd_queue(nic, qs, qidx);
874 		return;
875 	}
876 
877 	/* Reset send queue */
878 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET);
879 
880 	sq->cq_qs = qs->vnic_id;
881 	sq->cq_idx = qidx;
882 
883 	/* Send a mailbox msg to PF to config SQ */
884 	mbx.sq.msg = NIC_MBOX_MSG_SQ_CFG;
885 	mbx.sq.qs_num = qs->vnic_id;
886 	mbx.sq.sq_num = qidx;
887 	mbx.sq.sqs_mode = nic->sqs_mode;
888 	mbx.sq.cfg = (sq->cq_qs << 3) | sq->cq_idx;
889 	nicvf_send_msg_to_pf(nic, &mbx);
890 
891 	/* Set queue base address */
892 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_BASE,
893 			      qidx, (u64)(sq->dmem.phys_base));
894 
895 	/* Enable send queue  & set queue size */
896 	memset(&sq_cfg, 0, sizeof(struct sq_cfg));
897 	sq_cfg.ena = 1;
898 	sq_cfg.reset = 0;
899 	sq_cfg.ldwb = 0;
900 	sq_cfg.qsize = ilog2(qs->sq_len >> 10);
901 	sq_cfg.tstmp_bgx_intf = 0;
902 	/* CQ's level at which HW will stop processing SQEs to avoid
903 	 * transmitting a pkt with no space in CQ to post CQE_TX.
904 	 */
905 	sq_cfg.cq_limit = (CMP_QUEUE_PIPELINE_RSVD * 256) / qs->cq_len;
906 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, *(u64 *)&sq_cfg);
907 
908 	/* Set threshold value for interrupt generation */
909 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_THRESH, qidx, sq->thresh);
910 
911 	/* Set queue:cpu affinity for better load distribution */
912 	if (cpu_online(qidx)) {
913 		cpumask_set_cpu(qidx, &sq->affinity_mask);
914 		netif_set_xps_queue(nic->netdev,
915 				    &sq->affinity_mask, qidx);
916 	}
917 }
918 
919 /* Configures receive buffer descriptor ring */
920 static void nicvf_rbdr_config(struct nicvf *nic, struct queue_set *qs,
921 			      int qidx, bool enable)
922 {
923 	struct rbdr *rbdr;
924 	struct rbdr_cfg rbdr_cfg;
925 
926 	rbdr = &qs->rbdr[qidx];
927 	nicvf_reclaim_rbdr(nic, rbdr, qidx);
928 	if (!enable)
929 		return;
930 
931 	/* Set descriptor base address */
932 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_BASE,
933 			      qidx, (u64)(rbdr->dmem.phys_base));
934 
935 	/* Enable RBDR  & set queue size */
936 	/* Buffer size should be in multiples of 128 bytes */
937 	memset(&rbdr_cfg, 0, sizeof(struct rbdr_cfg));
938 	rbdr_cfg.ena = 1;
939 	rbdr_cfg.reset = 0;
940 	rbdr_cfg.ldwb = 0;
941 	rbdr_cfg.qsize = RBDR_SIZE;
942 	rbdr_cfg.avg_con = 0;
943 	rbdr_cfg.lines = rbdr->dma_size / 128;
944 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
945 			      qidx, *(u64 *)&rbdr_cfg);
946 
947 	/* Notify HW */
948 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR,
949 			      qidx, qs->rbdr_len - 1);
950 
951 	/* Set threshold value for interrupt generation */
952 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_THRESH,
953 			      qidx, rbdr->thresh - 1);
954 }
955 
956 /* Requests PF to assign and enable Qset */
957 void nicvf_qset_config(struct nicvf *nic, bool enable)
958 {
959 	union nic_mbx mbx = {};
960 	struct queue_set *qs = nic->qs;
961 	struct qs_cfg *qs_cfg;
962 
963 	if (!qs) {
964 		netdev_warn(nic->netdev,
965 			    "Qset is still not allocated, don't init queues\n");
966 		return;
967 	}
968 
969 	qs->enable = enable;
970 	qs->vnic_id = nic->vf_id;
971 
972 	/* Send a mailbox msg to PF to config Qset */
973 	mbx.qs.msg = NIC_MBOX_MSG_QS_CFG;
974 	mbx.qs.num = qs->vnic_id;
975 	mbx.qs.sqs_count = nic->sqs_count;
976 
977 	mbx.qs.cfg = 0;
978 	qs_cfg = (struct qs_cfg *)&mbx.qs.cfg;
979 	if (qs->enable) {
980 		qs_cfg->ena = 1;
981 #ifdef __BIG_ENDIAN
982 		qs_cfg->be = 1;
983 #endif
984 		qs_cfg->vnic = qs->vnic_id;
985 		/* Enable Tx timestamping capability */
986 		if (nic->ptp_clock)
987 			qs_cfg->send_tstmp_ena = 1;
988 	}
989 	nicvf_send_msg_to_pf(nic, &mbx);
990 }
991 
992 static void nicvf_free_resources(struct nicvf *nic)
993 {
994 	int qidx;
995 	struct queue_set *qs = nic->qs;
996 
997 	/* Free receive buffer descriptor ring */
998 	for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
999 		nicvf_free_rbdr(nic, &qs->rbdr[qidx]);
1000 
1001 	/* Free completion queue */
1002 	for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1003 		nicvf_free_cmp_queue(nic, &qs->cq[qidx]);
1004 
1005 	/* Free send queue */
1006 	for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1007 		nicvf_free_snd_queue(nic, &qs->sq[qidx]);
1008 }
1009 
1010 static int nicvf_alloc_resources(struct nicvf *nic)
1011 {
1012 	int qidx;
1013 	struct queue_set *qs = nic->qs;
1014 
1015 	/* Alloc receive buffer descriptor ring */
1016 	for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) {
1017 		if (nicvf_init_rbdr(nic, &qs->rbdr[qidx], qs->rbdr_len,
1018 				    DMA_BUFFER_LEN))
1019 			goto alloc_fail;
1020 	}
1021 
1022 	/* Alloc send queue */
1023 	for (qidx = 0; qidx < qs->sq_cnt; qidx++) {
1024 		if (nicvf_init_snd_queue(nic, &qs->sq[qidx], qs->sq_len, qidx))
1025 			goto alloc_fail;
1026 	}
1027 
1028 	/* Alloc completion queue */
1029 	for (qidx = 0; qidx < qs->cq_cnt; qidx++) {
1030 		if (nicvf_init_cmp_queue(nic, &qs->cq[qidx], qs->cq_len))
1031 			goto alloc_fail;
1032 	}
1033 
1034 	return 0;
1035 alloc_fail:
1036 	nicvf_free_resources(nic);
1037 	return -ENOMEM;
1038 }
1039 
1040 int nicvf_set_qset_resources(struct nicvf *nic)
1041 {
1042 	struct queue_set *qs;
1043 
1044 	qs = devm_kzalloc(&nic->pdev->dev, sizeof(*qs), GFP_KERNEL);
1045 	if (!qs)
1046 		return -ENOMEM;
1047 	nic->qs = qs;
1048 
1049 	/* Set count of each queue */
1050 	qs->rbdr_cnt = DEFAULT_RBDR_CNT;
1051 	qs->rq_cnt = min_t(u8, MAX_RCV_QUEUES_PER_QS, num_online_cpus());
1052 	qs->sq_cnt = min_t(u8, MAX_SND_QUEUES_PER_QS, num_online_cpus());
1053 	qs->cq_cnt = max_t(u8, qs->rq_cnt, qs->sq_cnt);
1054 
1055 	/* Set queue lengths */
1056 	qs->rbdr_len = RCV_BUF_COUNT;
1057 	qs->sq_len = SND_QUEUE_LEN;
1058 	qs->cq_len = CMP_QUEUE_LEN;
1059 
1060 	nic->rx_queues = qs->rq_cnt;
1061 	nic->tx_queues = qs->sq_cnt;
1062 	nic->xdp_tx_queues = 0;
1063 
1064 	return 0;
1065 }
1066 
1067 int nicvf_config_data_transfer(struct nicvf *nic, bool enable)
1068 {
1069 	bool disable = false;
1070 	struct queue_set *qs = nic->qs;
1071 	struct queue_set *pqs = nic->pnicvf->qs;
1072 	int qidx;
1073 
1074 	if (!qs)
1075 		return 0;
1076 
1077 	/* Take primary VF's queue lengths.
1078 	 * This is needed to take queue lengths set from ethtool
1079 	 * into consideration.
1080 	 */
1081 	if (nic->sqs_mode && pqs) {
1082 		qs->cq_len = pqs->cq_len;
1083 		qs->sq_len = pqs->sq_len;
1084 	}
1085 
1086 	if (enable) {
1087 		if (nicvf_alloc_resources(nic))
1088 			return -ENOMEM;
1089 
1090 		for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1091 			nicvf_snd_queue_config(nic, qs, qidx, enable);
1092 		for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1093 			nicvf_cmp_queue_config(nic, qs, qidx, enable);
1094 		for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
1095 			nicvf_rbdr_config(nic, qs, qidx, enable);
1096 		for (qidx = 0; qidx < qs->rq_cnt; qidx++)
1097 			nicvf_rcv_queue_config(nic, qs, qidx, enable);
1098 	} else {
1099 		for (qidx = 0; qidx < qs->rq_cnt; qidx++)
1100 			nicvf_rcv_queue_config(nic, qs, qidx, disable);
1101 		for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
1102 			nicvf_rbdr_config(nic, qs, qidx, disable);
1103 		for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1104 			nicvf_snd_queue_config(nic, qs, qidx, disable);
1105 		for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1106 			nicvf_cmp_queue_config(nic, qs, qidx, disable);
1107 
1108 		nicvf_free_resources(nic);
1109 	}
1110 
1111 	/* Reset RXQ's stats.
1112 	 * SQ's stats will get reset automatically once SQ is reset.
1113 	 */
1114 	nicvf_reset_rcv_queue_stats(nic);
1115 
1116 	return 0;
1117 }
1118 
1119 /* Get a free desc from SQ
1120  * returns descriptor ponter & descriptor number
1121  */
1122 static inline int nicvf_get_sq_desc(struct snd_queue *sq, int desc_cnt)
1123 {
1124 	int qentry;
1125 
1126 	qentry = sq->tail;
1127 	if (!sq->is_xdp)
1128 		atomic_sub(desc_cnt, &sq->free_cnt);
1129 	else
1130 		sq->xdp_free_cnt -= desc_cnt;
1131 	sq->tail += desc_cnt;
1132 	sq->tail &= (sq->dmem.q_len - 1);
1133 
1134 	return qentry;
1135 }
1136 
1137 /* Rollback to previous tail pointer when descriptors not used */
1138 static inline void nicvf_rollback_sq_desc(struct snd_queue *sq,
1139 					  int qentry, int desc_cnt)
1140 {
1141 	sq->tail = qentry;
1142 	atomic_add(desc_cnt, &sq->free_cnt);
1143 }
1144 
1145 /* Free descriptor back to SQ for future use */
1146 void nicvf_put_sq_desc(struct snd_queue *sq, int desc_cnt)
1147 {
1148 	if (!sq->is_xdp)
1149 		atomic_add(desc_cnt, &sq->free_cnt);
1150 	else
1151 		sq->xdp_free_cnt += desc_cnt;
1152 	sq->head += desc_cnt;
1153 	sq->head &= (sq->dmem.q_len - 1);
1154 }
1155 
1156 static inline int nicvf_get_nxt_sqentry(struct snd_queue *sq, int qentry)
1157 {
1158 	qentry++;
1159 	qentry &= (sq->dmem.q_len - 1);
1160 	return qentry;
1161 }
1162 
1163 void nicvf_sq_enable(struct nicvf *nic, struct snd_queue *sq, int qidx)
1164 {
1165 	u64 sq_cfg;
1166 
1167 	sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx);
1168 	sq_cfg |= NICVF_SQ_EN;
1169 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg);
1170 	/* Ring doorbell so that H/W restarts processing SQEs */
1171 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR, qidx, 0);
1172 }
1173 
1174 void nicvf_sq_disable(struct nicvf *nic, int qidx)
1175 {
1176 	u64 sq_cfg;
1177 
1178 	sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx);
1179 	sq_cfg &= ~NICVF_SQ_EN;
1180 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg);
1181 }
1182 
1183 void nicvf_sq_free_used_descs(struct net_device *netdev, struct snd_queue *sq,
1184 			      int qidx)
1185 {
1186 	u64 head, tail;
1187 	struct sk_buff *skb;
1188 	struct nicvf *nic = netdev_priv(netdev);
1189 	struct sq_hdr_subdesc *hdr;
1190 
1191 	head = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_HEAD, qidx) >> 4;
1192 	tail = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_TAIL, qidx) >> 4;
1193 	while (sq->head != head) {
1194 		hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, sq->head);
1195 		if (hdr->subdesc_type != SQ_DESC_TYPE_HEADER) {
1196 			nicvf_put_sq_desc(sq, 1);
1197 			continue;
1198 		}
1199 		skb = (struct sk_buff *)sq->skbuff[sq->head];
1200 		if (skb)
1201 			dev_kfree_skb_any(skb);
1202 		atomic64_add(1, (atomic64_t *)&netdev->stats.tx_packets);
1203 		atomic64_add(hdr->tot_len,
1204 			     (atomic64_t *)&netdev->stats.tx_bytes);
1205 		nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1);
1206 	}
1207 }
1208 
1209 /* XDP Transmit APIs */
1210 void nicvf_xdp_sq_doorbell(struct nicvf *nic,
1211 			   struct snd_queue *sq, int sq_num)
1212 {
1213 	if (!sq->xdp_desc_cnt)
1214 		return;
1215 
1216 	/* make sure all memory stores are done before ringing doorbell */
1217 	wmb();
1218 
1219 	/* Inform HW to xmit all TSO segments */
1220 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR,
1221 			      sq_num, sq->xdp_desc_cnt);
1222 	sq->xdp_desc_cnt = 0;
1223 }
1224 
1225 static inline void
1226 nicvf_xdp_sq_add_hdr_subdesc(struct snd_queue *sq, int qentry,
1227 			     int subdesc_cnt, u64 data, int len)
1228 {
1229 	struct sq_hdr_subdesc *hdr;
1230 
1231 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1232 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1233 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1234 	hdr->subdesc_cnt = subdesc_cnt;
1235 	hdr->tot_len = len;
1236 	hdr->post_cqe = 1;
1237 	sq->xdp_page[qentry] = (u64)virt_to_page((void *)data);
1238 }
1239 
1240 int nicvf_xdp_sq_append_pkt(struct nicvf *nic, struct snd_queue *sq,
1241 			    u64 bufaddr, u64 dma_addr, u16 len)
1242 {
1243 	int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT;
1244 	int qentry;
1245 
1246 	if (subdesc_cnt > sq->xdp_free_cnt)
1247 		return -1;
1248 
1249 	qentry = nicvf_get_sq_desc(sq, subdesc_cnt);
1250 
1251 	nicvf_xdp_sq_add_hdr_subdesc(sq, qentry, subdesc_cnt - 1, bufaddr, len);
1252 
1253 	qentry = nicvf_get_nxt_sqentry(sq, qentry);
1254 	nicvf_sq_add_gather_subdesc(sq, qentry, len, dma_addr);
1255 
1256 	sq->xdp_desc_cnt += subdesc_cnt;
1257 
1258 	return 0;
1259 }
1260 
1261 /* Calculate no of SQ subdescriptors needed to transmit all
1262  * segments of this TSO packet.
1263  * Taken from 'Tilera network driver' with a minor modification.
1264  */
1265 static int nicvf_tso_count_subdescs(struct sk_buff *skb)
1266 {
1267 	struct skb_shared_info *sh = skb_shinfo(skb);
1268 	unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1269 	unsigned int data_len = skb->len - sh_len;
1270 	unsigned int p_len = sh->gso_size;
1271 	long f_id = -1;    /* id of the current fragment */
1272 	long f_size = skb_headlen(skb) - sh_len;  /* current fragment size */
1273 	long f_used = 0;  /* bytes used from the current fragment */
1274 	long n;            /* size of the current piece of payload */
1275 	int num_edescs = 0;
1276 	int segment;
1277 
1278 	for (segment = 0; segment < sh->gso_segs; segment++) {
1279 		unsigned int p_used = 0;
1280 
1281 		/* One edesc for header and for each piece of the payload. */
1282 		for (num_edescs++; p_used < p_len; num_edescs++) {
1283 			/* Advance as needed. */
1284 			while (f_used >= f_size) {
1285 				f_id++;
1286 				f_size = skb_frag_size(&sh->frags[f_id]);
1287 				f_used = 0;
1288 			}
1289 
1290 			/* Use bytes from the current fragment. */
1291 			n = p_len - p_used;
1292 			if (n > f_size - f_used)
1293 				n = f_size - f_used;
1294 			f_used += n;
1295 			p_used += n;
1296 		}
1297 
1298 		/* The last segment may be less than gso_size. */
1299 		data_len -= p_len;
1300 		if (data_len < p_len)
1301 			p_len = data_len;
1302 	}
1303 
1304 	/* '+ gso_segs' for SQ_HDR_SUDESCs for each segment */
1305 	return num_edescs + sh->gso_segs;
1306 }
1307 
1308 #define POST_CQE_DESC_COUNT 2
1309 
1310 /* Get the number of SQ descriptors needed to xmit this skb */
1311 static int nicvf_sq_subdesc_required(struct nicvf *nic, struct sk_buff *skb)
1312 {
1313 	int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT;
1314 
1315 	if (skb_shinfo(skb)->gso_size && !nic->hw_tso) {
1316 		subdesc_cnt = nicvf_tso_count_subdescs(skb);
1317 		return subdesc_cnt;
1318 	}
1319 
1320 	/* Dummy descriptors to get TSO pkt completion notification */
1321 	if (nic->t88 && nic->hw_tso && skb_shinfo(skb)->gso_size)
1322 		subdesc_cnt += POST_CQE_DESC_COUNT;
1323 
1324 	if (skb_shinfo(skb)->nr_frags)
1325 		subdesc_cnt += skb_shinfo(skb)->nr_frags;
1326 
1327 	return subdesc_cnt;
1328 }
1329 
1330 /* Add SQ HEADER subdescriptor.
1331  * First subdescriptor for every send descriptor.
1332  */
1333 static inline void
1334 nicvf_sq_add_hdr_subdesc(struct nicvf *nic, struct snd_queue *sq, int qentry,
1335 			 int subdesc_cnt, struct sk_buff *skb, int len)
1336 {
1337 	int proto;
1338 	struct sq_hdr_subdesc *hdr;
1339 	union {
1340 		struct iphdr *v4;
1341 		struct ipv6hdr *v6;
1342 		unsigned char *hdr;
1343 	} ip;
1344 
1345 	ip.hdr = skb_network_header(skb);
1346 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1347 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1348 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1349 
1350 	if (nic->t88 && nic->hw_tso && skb_shinfo(skb)->gso_size) {
1351 		/* post_cqe = 0, to avoid HW posting a CQE for every TSO
1352 		 * segment transmitted on 88xx.
1353 		 */
1354 		hdr->subdesc_cnt = subdesc_cnt - POST_CQE_DESC_COUNT;
1355 	} else {
1356 		sq->skbuff[qentry] = (u64)skb;
1357 		/* Enable notification via CQE after processing SQE */
1358 		hdr->post_cqe = 1;
1359 		/* No of subdescriptors following this */
1360 		hdr->subdesc_cnt = subdesc_cnt;
1361 	}
1362 	hdr->tot_len = len;
1363 
1364 	/* Offload checksum calculation to HW */
1365 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
1366 		if (ip.v4->version == 4)
1367 			hdr->csum_l3 = 1; /* Enable IP csum calculation */
1368 		hdr->l3_offset = skb_network_offset(skb);
1369 		hdr->l4_offset = skb_transport_offset(skb);
1370 
1371 		proto = (ip.v4->version == 4) ? ip.v4->protocol :
1372 			ip.v6->nexthdr;
1373 
1374 		switch (proto) {
1375 		case IPPROTO_TCP:
1376 			hdr->csum_l4 = SEND_L4_CSUM_TCP;
1377 			break;
1378 		case IPPROTO_UDP:
1379 			hdr->csum_l4 = SEND_L4_CSUM_UDP;
1380 			break;
1381 		case IPPROTO_SCTP:
1382 			hdr->csum_l4 = SEND_L4_CSUM_SCTP;
1383 			break;
1384 		}
1385 	}
1386 
1387 	if (nic->hw_tso && skb_shinfo(skb)->gso_size) {
1388 		hdr->tso = 1;
1389 		hdr->tso_start = skb_transport_offset(skb) + tcp_hdrlen(skb);
1390 		hdr->tso_max_paysize = skb_shinfo(skb)->gso_size;
1391 		/* For non-tunneled pkts, point this to L2 ethertype */
1392 		hdr->inner_l3_offset = skb_network_offset(skb) - 2;
1393 		this_cpu_inc(nic->pnicvf->drv_stats->tx_tso);
1394 	}
1395 
1396 	/* Check if timestamp is requested */
1397 	if (!(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)) {
1398 		skb_tx_timestamp(skb);
1399 		return;
1400 	}
1401 
1402 	/* Tx timestamping not supported along with TSO, so ignore request */
1403 	if (skb_shinfo(skb)->gso_size)
1404 		return;
1405 
1406 	/* HW supports only a single outstanding packet to timestamp */
1407 	if (!atomic_add_unless(&nic->pnicvf->tx_ptp_skbs, 1, 1))
1408 		return;
1409 
1410 	/* Mark the SKB for later reference */
1411 	skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
1412 
1413 	/* Finally enable timestamp generation
1414 	 * Since 'post_cqe' is also set, two CQEs will be posted
1415 	 * for this packet i.e CQE_TYPE_SEND and CQE_TYPE_SEND_PTP.
1416 	 */
1417 	hdr->tstmp = 1;
1418 }
1419 
1420 /* SQ GATHER subdescriptor
1421  * Must follow HDR descriptor
1422  */
1423 static inline void nicvf_sq_add_gather_subdesc(struct snd_queue *sq, int qentry,
1424 					       int size, u64 data)
1425 {
1426 	struct sq_gather_subdesc *gather;
1427 
1428 	qentry &= (sq->dmem.q_len - 1);
1429 	gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, qentry);
1430 
1431 	memset(gather, 0, SND_QUEUE_DESC_SIZE);
1432 	gather->subdesc_type = SQ_DESC_TYPE_GATHER;
1433 	gather->ld_type = NIC_SEND_LD_TYPE_E_LDD;
1434 	gather->size = size;
1435 	gather->addr = data;
1436 }
1437 
1438 /* Add HDR + IMMEDIATE subdescriptors right after descriptors of a TSO
1439  * packet so that a CQE is posted as a notifation for transmission of
1440  * TSO packet.
1441  */
1442 static inline void nicvf_sq_add_cqe_subdesc(struct snd_queue *sq, int qentry,
1443 					    int tso_sqe, struct sk_buff *skb)
1444 {
1445 	struct sq_imm_subdesc *imm;
1446 	struct sq_hdr_subdesc *hdr;
1447 
1448 	sq->skbuff[qentry] = (u64)skb;
1449 
1450 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1451 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1452 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1453 	/* Enable notification via CQE after processing SQE */
1454 	hdr->post_cqe = 1;
1455 	/* There is no packet to transmit here */
1456 	hdr->dont_send = 1;
1457 	hdr->subdesc_cnt = POST_CQE_DESC_COUNT - 1;
1458 	hdr->tot_len = 1;
1459 	/* Actual TSO header SQE index, needed for cleanup */
1460 	hdr->rsvd2 = tso_sqe;
1461 
1462 	qentry = nicvf_get_nxt_sqentry(sq, qentry);
1463 	imm = (struct sq_imm_subdesc *)GET_SQ_DESC(sq, qentry);
1464 	memset(imm, 0, SND_QUEUE_DESC_SIZE);
1465 	imm->subdesc_type = SQ_DESC_TYPE_IMMEDIATE;
1466 	imm->len = 1;
1467 }
1468 
1469 static inline void nicvf_sq_doorbell(struct nicvf *nic, struct sk_buff *skb,
1470 				     int sq_num, int desc_cnt)
1471 {
1472 	struct netdev_queue *txq;
1473 
1474 	txq = netdev_get_tx_queue(nic->pnicvf->netdev,
1475 				  skb_get_queue_mapping(skb));
1476 
1477 	netdev_tx_sent_queue(txq, skb->len);
1478 
1479 	/* make sure all memory stores are done before ringing doorbell */
1480 	smp_wmb();
1481 
1482 	/* Inform HW to xmit all TSO segments */
1483 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR,
1484 			      sq_num, desc_cnt);
1485 }
1486 
1487 /* Segment a TSO packet into 'gso_size' segments and append
1488  * them to SQ for transfer
1489  */
1490 static int nicvf_sq_append_tso(struct nicvf *nic, struct snd_queue *sq,
1491 			       int sq_num, int qentry, struct sk_buff *skb)
1492 {
1493 	struct tso_t tso;
1494 	int seg_subdescs = 0, desc_cnt = 0;
1495 	int seg_len, total_len, data_left;
1496 	int hdr_qentry = qentry;
1497 	int hdr_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1498 
1499 	tso_start(skb, &tso);
1500 	total_len = skb->len - hdr_len;
1501 	while (total_len > 0) {
1502 		char *hdr;
1503 
1504 		/* Save Qentry for adding HDR_SUBDESC at the end */
1505 		hdr_qentry = qentry;
1506 
1507 		data_left = min_t(int, skb_shinfo(skb)->gso_size, total_len);
1508 		total_len -= data_left;
1509 
1510 		/* Add segment's header */
1511 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1512 		hdr = sq->tso_hdrs + qentry * TSO_HEADER_SIZE;
1513 		tso_build_hdr(skb, hdr, &tso, data_left, total_len == 0);
1514 		nicvf_sq_add_gather_subdesc(sq, qentry, hdr_len,
1515 					    sq->tso_hdrs_phys +
1516 					    qentry * TSO_HEADER_SIZE);
1517 		/* HDR_SUDESC + GATHER */
1518 		seg_subdescs = 2;
1519 		seg_len = hdr_len;
1520 
1521 		/* Add segment's payload fragments */
1522 		while (data_left > 0) {
1523 			int size;
1524 
1525 			size = min_t(int, tso.size, data_left);
1526 
1527 			qentry = nicvf_get_nxt_sqentry(sq, qentry);
1528 			nicvf_sq_add_gather_subdesc(sq, qentry, size,
1529 						    virt_to_phys(tso.data));
1530 			seg_subdescs++;
1531 			seg_len += size;
1532 
1533 			data_left -= size;
1534 			tso_build_data(skb, &tso, size);
1535 		}
1536 		nicvf_sq_add_hdr_subdesc(nic, sq, hdr_qentry,
1537 					 seg_subdescs - 1, skb, seg_len);
1538 		sq->skbuff[hdr_qentry] = (u64)NULL;
1539 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1540 
1541 		desc_cnt += seg_subdescs;
1542 	}
1543 	/* Save SKB in the last segment for freeing */
1544 	sq->skbuff[hdr_qentry] = (u64)skb;
1545 
1546 	nicvf_sq_doorbell(nic, skb, sq_num, desc_cnt);
1547 
1548 	this_cpu_inc(nic->pnicvf->drv_stats->tx_tso);
1549 	return 1;
1550 }
1551 
1552 /* Append an skb to a SQ for packet transfer. */
1553 int nicvf_sq_append_skb(struct nicvf *nic, struct snd_queue *sq,
1554 			struct sk_buff *skb, u8 sq_num)
1555 {
1556 	int i, size;
1557 	int subdesc_cnt, hdr_sqe = 0;
1558 	int qentry;
1559 	u64 dma_addr;
1560 
1561 	subdesc_cnt = nicvf_sq_subdesc_required(nic, skb);
1562 	if (subdesc_cnt > atomic_read(&sq->free_cnt))
1563 		goto append_fail;
1564 
1565 	qentry = nicvf_get_sq_desc(sq, subdesc_cnt);
1566 
1567 	/* Check if its a TSO packet */
1568 	if (skb_shinfo(skb)->gso_size && !nic->hw_tso)
1569 		return nicvf_sq_append_tso(nic, sq, sq_num, qentry, skb);
1570 
1571 	/* Add SQ header subdesc */
1572 	nicvf_sq_add_hdr_subdesc(nic, sq, qentry, subdesc_cnt - 1,
1573 				 skb, skb->len);
1574 	hdr_sqe = qentry;
1575 
1576 	/* Add SQ gather subdescs */
1577 	qentry = nicvf_get_nxt_sqentry(sq, qentry);
1578 	size = skb_is_nonlinear(skb) ? skb_headlen(skb) : skb->len;
1579 	/* HW will ensure data coherency, CPU sync not required */
1580 	dma_addr = dma_map_page_attrs(&nic->pdev->dev, virt_to_page(skb->data),
1581 				      offset_in_page(skb->data), size,
1582 				      DMA_TO_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
1583 	if (dma_mapping_error(&nic->pdev->dev, dma_addr)) {
1584 		nicvf_rollback_sq_desc(sq, qentry, subdesc_cnt);
1585 		return 0;
1586 	}
1587 
1588 	nicvf_sq_add_gather_subdesc(sq, qentry, size, dma_addr);
1589 
1590 	/* Check for scattered buffer */
1591 	if (!skb_is_nonlinear(skb))
1592 		goto doorbell;
1593 
1594 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1595 		const struct skb_frag_struct *frag;
1596 
1597 		frag = &skb_shinfo(skb)->frags[i];
1598 
1599 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1600 		size = skb_frag_size(frag);
1601 		dma_addr = dma_map_page_attrs(&nic->pdev->dev,
1602 					      skb_frag_page(frag),
1603 					      frag->page_offset, size,
1604 					      DMA_TO_DEVICE,
1605 					      DMA_ATTR_SKIP_CPU_SYNC);
1606 		if (dma_mapping_error(&nic->pdev->dev, dma_addr)) {
1607 			/* Free entire chain of mapped buffers
1608 			 * here 'i' = frags mapped + above mapped skb->data
1609 			 */
1610 			nicvf_unmap_sndq_buffers(nic, sq, hdr_sqe, i);
1611 			nicvf_rollback_sq_desc(sq, qentry, subdesc_cnt);
1612 			return 0;
1613 		}
1614 		nicvf_sq_add_gather_subdesc(sq, qentry, size, dma_addr);
1615 	}
1616 
1617 doorbell:
1618 	if (nic->t88 && skb_shinfo(skb)->gso_size) {
1619 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1620 		nicvf_sq_add_cqe_subdesc(sq, qentry, hdr_sqe, skb);
1621 	}
1622 
1623 	nicvf_sq_doorbell(nic, skb, sq_num, subdesc_cnt);
1624 
1625 	return 1;
1626 
1627 append_fail:
1628 	/* Use original PCI dev for debug log */
1629 	nic = nic->pnicvf;
1630 	netdev_dbg(nic->netdev, "Not enough SQ descriptors to xmit pkt\n");
1631 	return 0;
1632 }
1633 
1634 static inline unsigned frag_num(unsigned i)
1635 {
1636 #ifdef __BIG_ENDIAN
1637 	return (i & ~3) + 3 - (i & 3);
1638 #else
1639 	return i;
1640 #endif
1641 }
1642 
1643 static void nicvf_unmap_rcv_buffer(struct nicvf *nic, u64 dma_addr,
1644 				   u64 buf_addr, bool xdp)
1645 {
1646 	struct page *page = NULL;
1647 	int len = RCV_FRAG_LEN;
1648 
1649 	if (xdp) {
1650 		page = virt_to_page(phys_to_virt(buf_addr));
1651 		/* Check if it's a recycled page, if not
1652 		 * unmap the DMA mapping.
1653 		 *
1654 		 * Recycled page holds an extra reference.
1655 		 */
1656 		if (page_ref_count(page) != 1)
1657 			return;
1658 
1659 		len += XDP_HEADROOM;
1660 		/* Receive buffers in XDP mode are mapped from page start */
1661 		dma_addr &= PAGE_MASK;
1662 	}
1663 	dma_unmap_page_attrs(&nic->pdev->dev, dma_addr, len,
1664 			     DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
1665 }
1666 
1667 /* Returns SKB for a received packet */
1668 struct sk_buff *nicvf_get_rcv_skb(struct nicvf *nic,
1669 				  struct cqe_rx_t *cqe_rx, bool xdp)
1670 {
1671 	int frag;
1672 	int payload_len = 0;
1673 	struct sk_buff *skb = NULL;
1674 	struct page *page;
1675 	int offset;
1676 	u16 *rb_lens = NULL;
1677 	u64 *rb_ptrs = NULL;
1678 	u64 phys_addr;
1679 
1680 	rb_lens = (void *)cqe_rx + (3 * sizeof(u64));
1681 	/* Except 88xx pass1 on all other chips CQE_RX2_S is added to
1682 	 * CQE_RX at word6, hence buffer pointers move by word
1683 	 *
1684 	 * Use existing 'hw_tso' flag which will be set for all chips
1685 	 * except 88xx pass1 instead of a additional cache line
1686 	 * access (or miss) by using pci dev's revision.
1687 	 */
1688 	if (!nic->hw_tso)
1689 		rb_ptrs = (void *)cqe_rx + (6 * sizeof(u64));
1690 	else
1691 		rb_ptrs = (void *)cqe_rx + (7 * sizeof(u64));
1692 
1693 	for (frag = 0; frag < cqe_rx->rb_cnt; frag++) {
1694 		payload_len = rb_lens[frag_num(frag)];
1695 		phys_addr = nicvf_iova_to_phys(nic, *rb_ptrs);
1696 		if (!phys_addr) {
1697 			if (skb)
1698 				dev_kfree_skb_any(skb);
1699 			return NULL;
1700 		}
1701 
1702 		if (!frag) {
1703 			/* First fragment */
1704 			nicvf_unmap_rcv_buffer(nic,
1705 					       *rb_ptrs - cqe_rx->align_pad,
1706 					       phys_addr, xdp);
1707 			skb = nicvf_rb_ptr_to_skb(nic,
1708 						  phys_addr - cqe_rx->align_pad,
1709 						  payload_len);
1710 			if (!skb)
1711 				return NULL;
1712 			skb_reserve(skb, cqe_rx->align_pad);
1713 			skb_put(skb, payload_len);
1714 		} else {
1715 			/* Add fragments */
1716 			nicvf_unmap_rcv_buffer(nic, *rb_ptrs, phys_addr, xdp);
1717 			page = virt_to_page(phys_to_virt(phys_addr));
1718 			offset = phys_to_virt(phys_addr) - page_address(page);
1719 			skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page,
1720 					offset, payload_len, RCV_FRAG_LEN);
1721 		}
1722 		/* Next buffer pointer */
1723 		rb_ptrs++;
1724 	}
1725 	return skb;
1726 }
1727 
1728 static u64 nicvf_int_type_to_mask(int int_type, int q_idx)
1729 {
1730 	u64 reg_val;
1731 
1732 	switch (int_type) {
1733 	case NICVF_INTR_CQ:
1734 		reg_val = ((1ULL << q_idx) << NICVF_INTR_CQ_SHIFT);
1735 		break;
1736 	case NICVF_INTR_SQ:
1737 		reg_val = ((1ULL << q_idx) << NICVF_INTR_SQ_SHIFT);
1738 		break;
1739 	case NICVF_INTR_RBDR:
1740 		reg_val = ((1ULL << q_idx) << NICVF_INTR_RBDR_SHIFT);
1741 		break;
1742 	case NICVF_INTR_PKT_DROP:
1743 		reg_val = (1ULL << NICVF_INTR_PKT_DROP_SHIFT);
1744 		break;
1745 	case NICVF_INTR_TCP_TIMER:
1746 		reg_val = (1ULL << NICVF_INTR_TCP_TIMER_SHIFT);
1747 		break;
1748 	case NICVF_INTR_MBOX:
1749 		reg_val = (1ULL << NICVF_INTR_MBOX_SHIFT);
1750 		break;
1751 	case NICVF_INTR_QS_ERR:
1752 		reg_val = (1ULL << NICVF_INTR_QS_ERR_SHIFT);
1753 		break;
1754 	default:
1755 		reg_val = 0;
1756 	}
1757 
1758 	return reg_val;
1759 }
1760 
1761 /* Enable interrupt */
1762 void nicvf_enable_intr(struct nicvf *nic, int int_type, int q_idx)
1763 {
1764 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1765 
1766 	if (!mask) {
1767 		netdev_dbg(nic->netdev,
1768 			   "Failed to enable interrupt: unknown type\n");
1769 		return;
1770 	}
1771 	nicvf_reg_write(nic, NIC_VF_ENA_W1S,
1772 			nicvf_reg_read(nic, NIC_VF_ENA_W1S) | mask);
1773 }
1774 
1775 /* Disable interrupt */
1776 void nicvf_disable_intr(struct nicvf *nic, int int_type, int q_idx)
1777 {
1778 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1779 
1780 	if (!mask) {
1781 		netdev_dbg(nic->netdev,
1782 			   "Failed to disable interrupt: unknown type\n");
1783 		return;
1784 	}
1785 
1786 	nicvf_reg_write(nic, NIC_VF_ENA_W1C, mask);
1787 }
1788 
1789 /* Clear interrupt */
1790 void nicvf_clear_intr(struct nicvf *nic, int int_type, int q_idx)
1791 {
1792 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1793 
1794 	if (!mask) {
1795 		netdev_dbg(nic->netdev,
1796 			   "Failed to clear interrupt: unknown type\n");
1797 		return;
1798 	}
1799 
1800 	nicvf_reg_write(nic, NIC_VF_INT, mask);
1801 }
1802 
1803 /* Check if interrupt is enabled */
1804 int nicvf_is_intr_enabled(struct nicvf *nic, int int_type, int q_idx)
1805 {
1806 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1807 	/* If interrupt type is unknown, we treat it disabled. */
1808 	if (!mask) {
1809 		netdev_dbg(nic->netdev,
1810 			   "Failed to check interrupt enable: unknown type\n");
1811 		return 0;
1812 	}
1813 
1814 	return mask & nicvf_reg_read(nic, NIC_VF_ENA_W1S);
1815 }
1816 
1817 void nicvf_update_rq_stats(struct nicvf *nic, int rq_idx)
1818 {
1819 	struct rcv_queue *rq;
1820 
1821 #define GET_RQ_STATS(reg) \
1822 	nicvf_reg_read(nic, NIC_QSET_RQ_0_7_STAT_0_1 |\
1823 			    (rq_idx << NIC_Q_NUM_SHIFT) | (reg << 3))
1824 
1825 	rq = &nic->qs->rq[rq_idx];
1826 	rq->stats.bytes = GET_RQ_STATS(RQ_SQ_STATS_OCTS);
1827 	rq->stats.pkts = GET_RQ_STATS(RQ_SQ_STATS_PKTS);
1828 }
1829 
1830 void nicvf_update_sq_stats(struct nicvf *nic, int sq_idx)
1831 {
1832 	struct snd_queue *sq;
1833 
1834 #define GET_SQ_STATS(reg) \
1835 	nicvf_reg_read(nic, NIC_QSET_SQ_0_7_STAT_0_1 |\
1836 			    (sq_idx << NIC_Q_NUM_SHIFT) | (reg << 3))
1837 
1838 	sq = &nic->qs->sq[sq_idx];
1839 	sq->stats.bytes = GET_SQ_STATS(RQ_SQ_STATS_OCTS);
1840 	sq->stats.pkts = GET_SQ_STATS(RQ_SQ_STATS_PKTS);
1841 }
1842 
1843 /* Check for errors in the receive cmp.queue entry */
1844 int nicvf_check_cqe_rx_errs(struct nicvf *nic, struct cqe_rx_t *cqe_rx)
1845 {
1846 	netif_err(nic, rx_err, nic->netdev,
1847 		  "RX error CQE err_level 0x%x err_opcode 0x%x\n",
1848 		  cqe_rx->err_level, cqe_rx->err_opcode);
1849 
1850 	switch (cqe_rx->err_opcode) {
1851 	case CQ_RX_ERROP_RE_PARTIAL:
1852 		this_cpu_inc(nic->drv_stats->rx_bgx_truncated_pkts);
1853 		break;
1854 	case CQ_RX_ERROP_RE_JABBER:
1855 		this_cpu_inc(nic->drv_stats->rx_jabber_errs);
1856 		break;
1857 	case CQ_RX_ERROP_RE_FCS:
1858 		this_cpu_inc(nic->drv_stats->rx_fcs_errs);
1859 		break;
1860 	case CQ_RX_ERROP_RE_RX_CTL:
1861 		this_cpu_inc(nic->drv_stats->rx_bgx_errs);
1862 		break;
1863 	case CQ_RX_ERROP_PREL2_ERR:
1864 		this_cpu_inc(nic->drv_stats->rx_prel2_errs);
1865 		break;
1866 	case CQ_RX_ERROP_L2_MAL:
1867 		this_cpu_inc(nic->drv_stats->rx_l2_hdr_malformed);
1868 		break;
1869 	case CQ_RX_ERROP_L2_OVERSIZE:
1870 		this_cpu_inc(nic->drv_stats->rx_oversize);
1871 		break;
1872 	case CQ_RX_ERROP_L2_UNDERSIZE:
1873 		this_cpu_inc(nic->drv_stats->rx_undersize);
1874 		break;
1875 	case CQ_RX_ERROP_L2_LENMISM:
1876 		this_cpu_inc(nic->drv_stats->rx_l2_len_mismatch);
1877 		break;
1878 	case CQ_RX_ERROP_L2_PCLP:
1879 		this_cpu_inc(nic->drv_stats->rx_l2_pclp);
1880 		break;
1881 	case CQ_RX_ERROP_IP_NOT:
1882 		this_cpu_inc(nic->drv_stats->rx_ip_ver_errs);
1883 		break;
1884 	case CQ_RX_ERROP_IP_CSUM_ERR:
1885 		this_cpu_inc(nic->drv_stats->rx_ip_csum_errs);
1886 		break;
1887 	case CQ_RX_ERROP_IP_MAL:
1888 		this_cpu_inc(nic->drv_stats->rx_ip_hdr_malformed);
1889 		break;
1890 	case CQ_RX_ERROP_IP_MALD:
1891 		this_cpu_inc(nic->drv_stats->rx_ip_payload_malformed);
1892 		break;
1893 	case CQ_RX_ERROP_IP_HOP:
1894 		this_cpu_inc(nic->drv_stats->rx_ip_ttl_errs);
1895 		break;
1896 	case CQ_RX_ERROP_L3_PCLP:
1897 		this_cpu_inc(nic->drv_stats->rx_l3_pclp);
1898 		break;
1899 	case CQ_RX_ERROP_L4_MAL:
1900 		this_cpu_inc(nic->drv_stats->rx_l4_malformed);
1901 		break;
1902 	case CQ_RX_ERROP_L4_CHK:
1903 		this_cpu_inc(nic->drv_stats->rx_l4_csum_errs);
1904 		break;
1905 	case CQ_RX_ERROP_UDP_LEN:
1906 		this_cpu_inc(nic->drv_stats->rx_udp_len_errs);
1907 		break;
1908 	case CQ_RX_ERROP_L4_PORT:
1909 		this_cpu_inc(nic->drv_stats->rx_l4_port_errs);
1910 		break;
1911 	case CQ_RX_ERROP_TCP_FLAG:
1912 		this_cpu_inc(nic->drv_stats->rx_tcp_flag_errs);
1913 		break;
1914 	case CQ_RX_ERROP_TCP_OFFSET:
1915 		this_cpu_inc(nic->drv_stats->rx_tcp_offset_errs);
1916 		break;
1917 	case CQ_RX_ERROP_L4_PCLP:
1918 		this_cpu_inc(nic->drv_stats->rx_l4_pclp);
1919 		break;
1920 	case CQ_RX_ERROP_RBDR_TRUNC:
1921 		this_cpu_inc(nic->drv_stats->rx_truncated_pkts);
1922 		break;
1923 	}
1924 
1925 	return 1;
1926 }
1927 
1928 /* Check for errors in the send cmp.queue entry */
1929 int nicvf_check_cqe_tx_errs(struct nicvf *nic, struct cqe_send_t *cqe_tx)
1930 {
1931 	switch (cqe_tx->send_status) {
1932 	case CQ_TX_ERROP_DESC_FAULT:
1933 		this_cpu_inc(nic->drv_stats->tx_desc_fault);
1934 		break;
1935 	case CQ_TX_ERROP_HDR_CONS_ERR:
1936 		this_cpu_inc(nic->drv_stats->tx_hdr_cons_err);
1937 		break;
1938 	case CQ_TX_ERROP_SUBDC_ERR:
1939 		this_cpu_inc(nic->drv_stats->tx_subdesc_err);
1940 		break;
1941 	case CQ_TX_ERROP_MAX_SIZE_VIOL:
1942 		this_cpu_inc(nic->drv_stats->tx_max_size_exceeded);
1943 		break;
1944 	case CQ_TX_ERROP_IMM_SIZE_OFLOW:
1945 		this_cpu_inc(nic->drv_stats->tx_imm_size_oflow);
1946 		break;
1947 	case CQ_TX_ERROP_DATA_SEQUENCE_ERR:
1948 		this_cpu_inc(nic->drv_stats->tx_data_seq_err);
1949 		break;
1950 	case CQ_TX_ERROP_MEM_SEQUENCE_ERR:
1951 		this_cpu_inc(nic->drv_stats->tx_mem_seq_err);
1952 		break;
1953 	case CQ_TX_ERROP_LOCK_VIOL:
1954 		this_cpu_inc(nic->drv_stats->tx_lock_viol);
1955 		break;
1956 	case CQ_TX_ERROP_DATA_FAULT:
1957 		this_cpu_inc(nic->drv_stats->tx_data_fault);
1958 		break;
1959 	case CQ_TX_ERROP_TSTMP_CONFLICT:
1960 		this_cpu_inc(nic->drv_stats->tx_tstmp_conflict);
1961 		break;
1962 	case CQ_TX_ERROP_TSTMP_TIMEOUT:
1963 		this_cpu_inc(nic->drv_stats->tx_tstmp_timeout);
1964 		break;
1965 	case CQ_TX_ERROP_MEM_FAULT:
1966 		this_cpu_inc(nic->drv_stats->tx_mem_fault);
1967 		break;
1968 	case CQ_TX_ERROP_CK_OVERLAP:
1969 		this_cpu_inc(nic->drv_stats->tx_csum_overlap);
1970 		break;
1971 	case CQ_TX_ERROP_CK_OFLOW:
1972 		this_cpu_inc(nic->drv_stats->tx_csum_overflow);
1973 		break;
1974 	}
1975 
1976 	return 1;
1977 }
1978