xref: /linux/drivers/net/ethernet/cavium/thunder/nicvf_queues.c (revision 02680c23d7b3febe45ea3d4f9818c2b2dc89020a)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2015 Cavium, Inc.
4  */
5 
6 #include <linux/pci.h>
7 #include <linux/netdevice.h>
8 #include <linux/ip.h>
9 #include <linux/etherdevice.h>
10 #include <linux/iommu.h>
11 #include <net/ip.h>
12 #include <net/tso.h>
13 
14 #include "nic_reg.h"
15 #include "nic.h"
16 #include "q_struct.h"
17 #include "nicvf_queues.h"
18 
19 static inline void nicvf_sq_add_gather_subdesc(struct snd_queue *sq, int qentry,
20 					       int size, u64 data);
21 static void nicvf_get_page(struct nicvf *nic)
22 {
23 	if (!nic->rb_pageref || !nic->rb_page)
24 		return;
25 
26 	page_ref_add(nic->rb_page, nic->rb_pageref);
27 	nic->rb_pageref = 0;
28 }
29 
30 /* Poll a register for a specific value */
31 static int nicvf_poll_reg(struct nicvf *nic, int qidx,
32 			  u64 reg, int bit_pos, int bits, int val)
33 {
34 	u64 bit_mask;
35 	u64 reg_val;
36 	int timeout = 10;
37 
38 	bit_mask = (1ULL << bits) - 1;
39 	bit_mask = (bit_mask << bit_pos);
40 
41 	while (timeout) {
42 		reg_val = nicvf_queue_reg_read(nic, reg, qidx);
43 		if (((reg_val & bit_mask) >> bit_pos) == val)
44 			return 0;
45 		usleep_range(1000, 2000);
46 		timeout--;
47 	}
48 	netdev_err(nic->netdev, "Poll on reg 0x%llx failed\n", reg);
49 	return 1;
50 }
51 
52 /* Allocate memory for a queue's descriptors */
53 static int nicvf_alloc_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem,
54 				  int q_len, int desc_size, int align_bytes)
55 {
56 	dmem->q_len = q_len;
57 	dmem->size = (desc_size * q_len) + align_bytes;
58 	/* Save address, need it while freeing */
59 	dmem->unalign_base = dma_alloc_coherent(&nic->pdev->dev, dmem->size,
60 						&dmem->dma, GFP_KERNEL);
61 	if (!dmem->unalign_base)
62 		return -ENOMEM;
63 
64 	/* Align memory address for 'align_bytes' */
65 	dmem->phys_base = NICVF_ALIGNED_ADDR((u64)dmem->dma, align_bytes);
66 	dmem->base = dmem->unalign_base + (dmem->phys_base - dmem->dma);
67 	return 0;
68 }
69 
70 /* Free queue's descriptor memory */
71 static void nicvf_free_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem)
72 {
73 	if (!dmem)
74 		return;
75 
76 	dma_free_coherent(&nic->pdev->dev, dmem->size,
77 			  dmem->unalign_base, dmem->dma);
78 	dmem->unalign_base = NULL;
79 	dmem->base = NULL;
80 }
81 
82 #define XDP_PAGE_REFCNT_REFILL 256
83 
84 /* Allocate a new page or recycle one if possible
85  *
86  * We cannot optimize dma mapping here, since
87  * 1. It's only one RBDR ring for 8 Rx queues.
88  * 2. CQE_RX gives address of the buffer where pkt has been DMA'ed
89  *    and not idx into RBDR ring, so can't refer to saved info.
90  * 3. There are multiple receive buffers per page
91  */
92 static inline struct pgcache *nicvf_alloc_page(struct nicvf *nic,
93 					       struct rbdr *rbdr, gfp_t gfp)
94 {
95 	int ref_count;
96 	struct page *page = NULL;
97 	struct pgcache *pgcache, *next;
98 
99 	/* Check if page is already allocated */
100 	pgcache = &rbdr->pgcache[rbdr->pgidx];
101 	page = pgcache->page;
102 	/* Check if page can be recycled */
103 	if (page) {
104 		ref_count = page_ref_count(page);
105 		/* This page can be recycled if internal ref_count and page's
106 		 * ref_count are equal, indicating that the page has been used
107 		 * once for packet transmission. For non-XDP mode, internal
108 		 * ref_count is always '1'.
109 		 */
110 		if (rbdr->is_xdp) {
111 			if (ref_count == pgcache->ref_count)
112 				pgcache->ref_count--;
113 			else
114 				page = NULL;
115 		} else if (ref_count != 1) {
116 			page = NULL;
117 		}
118 	}
119 
120 	if (!page) {
121 		page = alloc_pages(gfp | __GFP_COMP | __GFP_NOWARN, 0);
122 		if (!page)
123 			return NULL;
124 
125 		this_cpu_inc(nic->pnicvf->drv_stats->page_alloc);
126 
127 		/* Check for space */
128 		if (rbdr->pgalloc >= rbdr->pgcnt) {
129 			/* Page can still be used */
130 			nic->rb_page = page;
131 			return NULL;
132 		}
133 
134 		/* Save the page in page cache */
135 		pgcache->page = page;
136 		pgcache->dma_addr = 0;
137 		pgcache->ref_count = 0;
138 		rbdr->pgalloc++;
139 	}
140 
141 	/* Take additional page references for recycling */
142 	if (rbdr->is_xdp) {
143 		/* Since there is single RBDR (i.e single core doing
144 		 * page recycling) per 8 Rx queues, in XDP mode adjusting
145 		 * page references atomically is the biggest bottleneck, so
146 		 * take bunch of references at a time.
147 		 *
148 		 * So here, below reference counts defer by '1'.
149 		 */
150 		if (!pgcache->ref_count) {
151 			pgcache->ref_count = XDP_PAGE_REFCNT_REFILL;
152 			page_ref_add(page, XDP_PAGE_REFCNT_REFILL);
153 		}
154 	} else {
155 		/* In non-XDP case, single 64K page is divided across multiple
156 		 * receive buffers, so cost of recycling is less anyway.
157 		 * So we can do with just one extra reference.
158 		 */
159 		page_ref_add(page, 1);
160 	}
161 
162 	rbdr->pgidx++;
163 	rbdr->pgidx &= (rbdr->pgcnt - 1);
164 
165 	/* Prefetch refcount of next page in page cache */
166 	next = &rbdr->pgcache[rbdr->pgidx];
167 	page = next->page;
168 	if (page)
169 		prefetch(&page->_refcount);
170 
171 	return pgcache;
172 }
173 
174 /* Allocate buffer for packet reception */
175 static inline int nicvf_alloc_rcv_buffer(struct nicvf *nic, struct rbdr *rbdr,
176 					 gfp_t gfp, u32 buf_len, u64 *rbuf)
177 {
178 	struct pgcache *pgcache = NULL;
179 
180 	/* Check if request can be accomodated in previous allocated page.
181 	 * But in XDP mode only one buffer per page is permitted.
182 	 */
183 	if (!rbdr->is_xdp && nic->rb_page &&
184 	    ((nic->rb_page_offset + buf_len) <= PAGE_SIZE)) {
185 		nic->rb_pageref++;
186 		goto ret;
187 	}
188 
189 	nicvf_get_page(nic);
190 	nic->rb_page = NULL;
191 
192 	/* Get new page, either recycled or new one */
193 	pgcache = nicvf_alloc_page(nic, rbdr, gfp);
194 	if (!pgcache && !nic->rb_page) {
195 		this_cpu_inc(nic->pnicvf->drv_stats->rcv_buffer_alloc_failures);
196 		return -ENOMEM;
197 	}
198 
199 	nic->rb_page_offset = 0;
200 
201 	/* Reserve space for header modifications by BPF program */
202 	if (rbdr->is_xdp)
203 		buf_len += XDP_PACKET_HEADROOM;
204 
205 	/* Check if it's recycled */
206 	if (pgcache)
207 		nic->rb_page = pgcache->page;
208 ret:
209 	if (rbdr->is_xdp && pgcache && pgcache->dma_addr) {
210 		*rbuf = pgcache->dma_addr;
211 	} else {
212 		/* HW will ensure data coherency, CPU sync not required */
213 		*rbuf = (u64)dma_map_page_attrs(&nic->pdev->dev, nic->rb_page,
214 						nic->rb_page_offset, buf_len,
215 						DMA_FROM_DEVICE,
216 						DMA_ATTR_SKIP_CPU_SYNC);
217 		if (dma_mapping_error(&nic->pdev->dev, (dma_addr_t)*rbuf)) {
218 			if (!nic->rb_page_offset)
219 				__free_pages(nic->rb_page, 0);
220 			nic->rb_page = NULL;
221 			return -ENOMEM;
222 		}
223 		if (pgcache)
224 			pgcache->dma_addr = *rbuf + XDP_PACKET_HEADROOM;
225 		nic->rb_page_offset += buf_len;
226 	}
227 
228 	return 0;
229 }
230 
231 /* Build skb around receive buffer */
232 static struct sk_buff *nicvf_rb_ptr_to_skb(struct nicvf *nic,
233 					   u64 rb_ptr, int len)
234 {
235 	void *data;
236 	struct sk_buff *skb;
237 
238 	data = phys_to_virt(rb_ptr);
239 
240 	/* Now build an skb to give to stack */
241 	skb = build_skb(data, RCV_FRAG_LEN);
242 	if (!skb) {
243 		put_page(virt_to_page(data));
244 		return NULL;
245 	}
246 
247 	prefetch(skb->data);
248 	return skb;
249 }
250 
251 /* Allocate RBDR ring and populate receive buffers */
252 static int  nicvf_init_rbdr(struct nicvf *nic, struct rbdr *rbdr,
253 			    int ring_len, int buf_size)
254 {
255 	int idx;
256 	u64 rbuf;
257 	struct rbdr_entry_t *desc;
258 	int err;
259 
260 	err = nicvf_alloc_q_desc_mem(nic, &rbdr->dmem, ring_len,
261 				     sizeof(struct rbdr_entry_t),
262 				     NICVF_RCV_BUF_ALIGN_BYTES);
263 	if (err)
264 		return err;
265 
266 	rbdr->desc = rbdr->dmem.base;
267 	/* Buffer size has to be in multiples of 128 bytes */
268 	rbdr->dma_size = buf_size;
269 	rbdr->enable = true;
270 	rbdr->thresh = RBDR_THRESH;
271 	rbdr->head = 0;
272 	rbdr->tail = 0;
273 
274 	/* Initialize page recycling stuff.
275 	 *
276 	 * Can't use single buffer per page especially with 64K pages.
277 	 * On embedded platforms i.e 81xx/83xx available memory itself
278 	 * is low and minimum ring size of RBDR is 8K, that takes away
279 	 * lots of memory.
280 	 *
281 	 * But for XDP it has to be a single buffer per page.
282 	 */
283 	if (!nic->pnicvf->xdp_prog) {
284 		rbdr->pgcnt = ring_len / (PAGE_SIZE / buf_size);
285 		rbdr->is_xdp = false;
286 	} else {
287 		rbdr->pgcnt = ring_len;
288 		rbdr->is_xdp = true;
289 	}
290 	rbdr->pgcnt = roundup_pow_of_two(rbdr->pgcnt);
291 	rbdr->pgcache = kcalloc(rbdr->pgcnt, sizeof(*rbdr->pgcache),
292 				GFP_KERNEL);
293 	if (!rbdr->pgcache)
294 		return -ENOMEM;
295 	rbdr->pgidx = 0;
296 	rbdr->pgalloc = 0;
297 
298 	nic->rb_page = NULL;
299 	for (idx = 0; idx < ring_len; idx++) {
300 		err = nicvf_alloc_rcv_buffer(nic, rbdr, GFP_KERNEL,
301 					     RCV_FRAG_LEN, &rbuf);
302 		if (err) {
303 			/* To free already allocated and mapped ones */
304 			rbdr->tail = idx - 1;
305 			return err;
306 		}
307 
308 		desc = GET_RBDR_DESC(rbdr, idx);
309 		desc->buf_addr = rbuf & ~(NICVF_RCV_BUF_ALIGN_BYTES - 1);
310 	}
311 
312 	nicvf_get_page(nic);
313 
314 	return 0;
315 }
316 
317 /* Free RBDR ring and its receive buffers */
318 static void nicvf_free_rbdr(struct nicvf *nic, struct rbdr *rbdr)
319 {
320 	int head, tail;
321 	u64 buf_addr, phys_addr;
322 	struct pgcache *pgcache;
323 	struct rbdr_entry_t *desc;
324 
325 	if (!rbdr)
326 		return;
327 
328 	rbdr->enable = false;
329 	if (!rbdr->dmem.base)
330 		return;
331 
332 	head = rbdr->head;
333 	tail = rbdr->tail;
334 
335 	/* Release page references */
336 	while (head != tail) {
337 		desc = GET_RBDR_DESC(rbdr, head);
338 		buf_addr = desc->buf_addr;
339 		phys_addr = nicvf_iova_to_phys(nic, buf_addr);
340 		dma_unmap_page_attrs(&nic->pdev->dev, buf_addr, RCV_FRAG_LEN,
341 				     DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
342 		if (phys_addr)
343 			put_page(virt_to_page(phys_to_virt(phys_addr)));
344 		head++;
345 		head &= (rbdr->dmem.q_len - 1);
346 	}
347 	/* Release buffer of tail desc */
348 	desc = GET_RBDR_DESC(rbdr, tail);
349 	buf_addr = desc->buf_addr;
350 	phys_addr = nicvf_iova_to_phys(nic, buf_addr);
351 	dma_unmap_page_attrs(&nic->pdev->dev, buf_addr, RCV_FRAG_LEN,
352 			     DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
353 	if (phys_addr)
354 		put_page(virt_to_page(phys_to_virt(phys_addr)));
355 
356 	/* Sync page cache info */
357 	smp_rmb();
358 
359 	/* Release additional page references held for recycling */
360 	head = 0;
361 	while (head < rbdr->pgcnt) {
362 		pgcache = &rbdr->pgcache[head];
363 		if (pgcache->page && page_ref_count(pgcache->page) != 0) {
364 			if (rbdr->is_xdp) {
365 				page_ref_sub(pgcache->page,
366 					     pgcache->ref_count - 1);
367 			}
368 			put_page(pgcache->page);
369 		}
370 		head++;
371 	}
372 
373 	/* Free RBDR ring */
374 	nicvf_free_q_desc_mem(nic, &rbdr->dmem);
375 }
376 
377 /* Refill receive buffer descriptors with new buffers.
378  */
379 static void nicvf_refill_rbdr(struct nicvf *nic, gfp_t gfp)
380 {
381 	struct queue_set *qs = nic->qs;
382 	int rbdr_idx = qs->rbdr_cnt;
383 	int tail, qcount;
384 	int refill_rb_cnt;
385 	struct rbdr *rbdr;
386 	struct rbdr_entry_t *desc;
387 	u64 rbuf;
388 	int new_rb = 0;
389 
390 refill:
391 	if (!rbdr_idx)
392 		return;
393 	rbdr_idx--;
394 	rbdr = &qs->rbdr[rbdr_idx];
395 	/* Check if it's enabled */
396 	if (!rbdr->enable)
397 		goto next_rbdr;
398 
399 	/* Get no of desc's to be refilled */
400 	qcount = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, rbdr_idx);
401 	qcount &= 0x7FFFF;
402 	/* Doorbell can be ringed with a max of ring size minus 1 */
403 	if (qcount >= (qs->rbdr_len - 1))
404 		goto next_rbdr;
405 	else
406 		refill_rb_cnt = qs->rbdr_len - qcount - 1;
407 
408 	/* Sync page cache info */
409 	smp_rmb();
410 
411 	/* Start filling descs from tail */
412 	tail = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_TAIL, rbdr_idx) >> 3;
413 	while (refill_rb_cnt) {
414 		tail++;
415 		tail &= (rbdr->dmem.q_len - 1);
416 
417 		if (nicvf_alloc_rcv_buffer(nic, rbdr, gfp, RCV_FRAG_LEN, &rbuf))
418 			break;
419 
420 		desc = GET_RBDR_DESC(rbdr, tail);
421 		desc->buf_addr = rbuf & ~(NICVF_RCV_BUF_ALIGN_BYTES - 1);
422 		refill_rb_cnt--;
423 		new_rb++;
424 	}
425 
426 	nicvf_get_page(nic);
427 
428 	/* make sure all memory stores are done before ringing doorbell */
429 	smp_wmb();
430 
431 	/* Check if buffer allocation failed */
432 	if (refill_rb_cnt)
433 		nic->rb_alloc_fail = true;
434 	else
435 		nic->rb_alloc_fail = false;
436 
437 	/* Notify HW */
438 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR,
439 			      rbdr_idx, new_rb);
440 next_rbdr:
441 	/* Re-enable RBDR interrupts only if buffer allocation is success */
442 	if (!nic->rb_alloc_fail && rbdr->enable &&
443 	    netif_running(nic->pnicvf->netdev))
444 		nicvf_enable_intr(nic, NICVF_INTR_RBDR, rbdr_idx);
445 
446 	if (rbdr_idx)
447 		goto refill;
448 }
449 
450 /* Alloc rcv buffers in non-atomic mode for better success */
451 void nicvf_rbdr_work(struct work_struct *work)
452 {
453 	struct nicvf *nic = container_of(work, struct nicvf, rbdr_work.work);
454 
455 	nicvf_refill_rbdr(nic, GFP_KERNEL);
456 	if (nic->rb_alloc_fail)
457 		schedule_delayed_work(&nic->rbdr_work, msecs_to_jiffies(10));
458 	else
459 		nic->rb_work_scheduled = false;
460 }
461 
462 /* In Softirq context, alloc rcv buffers in atomic mode */
463 void nicvf_rbdr_task(struct tasklet_struct *t)
464 {
465 	struct nicvf *nic = from_tasklet(nic, t, rbdr_task);
466 
467 	nicvf_refill_rbdr(nic, GFP_ATOMIC);
468 	if (nic->rb_alloc_fail) {
469 		nic->rb_work_scheduled = true;
470 		schedule_delayed_work(&nic->rbdr_work, msecs_to_jiffies(10));
471 	}
472 }
473 
474 /* Initialize completion queue */
475 static int nicvf_init_cmp_queue(struct nicvf *nic,
476 				struct cmp_queue *cq, int q_len)
477 {
478 	int err;
479 
480 	err = nicvf_alloc_q_desc_mem(nic, &cq->dmem, q_len, CMP_QUEUE_DESC_SIZE,
481 				     NICVF_CQ_BASE_ALIGN_BYTES);
482 	if (err)
483 		return err;
484 
485 	cq->desc = cq->dmem.base;
486 	cq->thresh = pass1_silicon(nic->pdev) ? 0 : CMP_QUEUE_CQE_THRESH;
487 	nic->cq_coalesce_usecs = (CMP_QUEUE_TIMER_THRESH * 0.05) - 1;
488 
489 	return 0;
490 }
491 
492 static void nicvf_free_cmp_queue(struct nicvf *nic, struct cmp_queue *cq)
493 {
494 	if (!cq)
495 		return;
496 	if (!cq->dmem.base)
497 		return;
498 
499 	nicvf_free_q_desc_mem(nic, &cq->dmem);
500 }
501 
502 /* Initialize transmit queue */
503 static int nicvf_init_snd_queue(struct nicvf *nic,
504 				struct snd_queue *sq, int q_len, int qidx)
505 {
506 	int err;
507 
508 	err = nicvf_alloc_q_desc_mem(nic, &sq->dmem, q_len, SND_QUEUE_DESC_SIZE,
509 				     NICVF_SQ_BASE_ALIGN_BYTES);
510 	if (err)
511 		return err;
512 
513 	sq->desc = sq->dmem.base;
514 	sq->skbuff = kcalloc(q_len, sizeof(u64), GFP_KERNEL);
515 	if (!sq->skbuff)
516 		return -ENOMEM;
517 
518 	sq->head = 0;
519 	sq->tail = 0;
520 	sq->thresh = SND_QUEUE_THRESH;
521 
522 	/* Check if this SQ is a XDP TX queue */
523 	if (nic->sqs_mode)
524 		qidx += ((nic->sqs_id + 1) * MAX_SND_QUEUES_PER_QS);
525 	if (qidx < nic->pnicvf->xdp_tx_queues) {
526 		/* Alloc memory to save page pointers for XDP_TX */
527 		sq->xdp_page = kcalloc(q_len, sizeof(u64), GFP_KERNEL);
528 		if (!sq->xdp_page)
529 			return -ENOMEM;
530 		sq->xdp_desc_cnt = 0;
531 		sq->xdp_free_cnt = q_len - 1;
532 		sq->is_xdp = true;
533 	} else {
534 		sq->xdp_page = NULL;
535 		sq->xdp_desc_cnt = 0;
536 		sq->xdp_free_cnt = 0;
537 		sq->is_xdp = false;
538 
539 		atomic_set(&sq->free_cnt, q_len - 1);
540 
541 		/* Preallocate memory for TSO segment's header */
542 		sq->tso_hdrs = dma_alloc_coherent(&nic->pdev->dev,
543 						  q_len * TSO_HEADER_SIZE,
544 						  &sq->tso_hdrs_phys,
545 						  GFP_KERNEL);
546 		if (!sq->tso_hdrs)
547 			return -ENOMEM;
548 	}
549 
550 	return 0;
551 }
552 
553 void nicvf_unmap_sndq_buffers(struct nicvf *nic, struct snd_queue *sq,
554 			      int hdr_sqe, u8 subdesc_cnt)
555 {
556 	u8 idx;
557 	struct sq_gather_subdesc *gather;
558 
559 	/* Unmap DMA mapped skb data buffers */
560 	for (idx = 0; idx < subdesc_cnt; idx++) {
561 		hdr_sqe++;
562 		hdr_sqe &= (sq->dmem.q_len - 1);
563 		gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, hdr_sqe);
564 		/* HW will ensure data coherency, CPU sync not required */
565 		dma_unmap_page_attrs(&nic->pdev->dev, gather->addr,
566 				     gather->size, DMA_TO_DEVICE,
567 				     DMA_ATTR_SKIP_CPU_SYNC);
568 	}
569 }
570 
571 static void nicvf_free_snd_queue(struct nicvf *nic, struct snd_queue *sq)
572 {
573 	struct sk_buff *skb;
574 	struct page *page;
575 	struct sq_hdr_subdesc *hdr;
576 	struct sq_hdr_subdesc *tso_sqe;
577 
578 	if (!sq)
579 		return;
580 	if (!sq->dmem.base)
581 		return;
582 
583 	if (sq->tso_hdrs) {
584 		dma_free_coherent(&nic->pdev->dev,
585 				  sq->dmem.q_len * TSO_HEADER_SIZE,
586 				  sq->tso_hdrs, sq->tso_hdrs_phys);
587 		sq->tso_hdrs = NULL;
588 	}
589 
590 	/* Free pending skbs in the queue */
591 	smp_rmb();
592 	while (sq->head != sq->tail) {
593 		skb = (struct sk_buff *)sq->skbuff[sq->head];
594 		if (!skb || !sq->xdp_page)
595 			goto next;
596 
597 		page = (struct page *)sq->xdp_page[sq->head];
598 		if (!page)
599 			goto next;
600 		else
601 			put_page(page);
602 
603 		hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, sq->head);
604 		/* Check for dummy descriptor used for HW TSO offload on 88xx */
605 		if (hdr->dont_send) {
606 			/* Get actual TSO descriptors and unmap them */
607 			tso_sqe =
608 			 (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, hdr->rsvd2);
609 			nicvf_unmap_sndq_buffers(nic, sq, hdr->rsvd2,
610 						 tso_sqe->subdesc_cnt);
611 		} else {
612 			nicvf_unmap_sndq_buffers(nic, sq, sq->head,
613 						 hdr->subdesc_cnt);
614 		}
615 		if (skb)
616 			dev_kfree_skb_any(skb);
617 next:
618 		sq->head++;
619 		sq->head &= (sq->dmem.q_len - 1);
620 	}
621 	kfree(sq->skbuff);
622 	kfree(sq->xdp_page);
623 	nicvf_free_q_desc_mem(nic, &sq->dmem);
624 }
625 
626 static void nicvf_reclaim_snd_queue(struct nicvf *nic,
627 				    struct queue_set *qs, int qidx)
628 {
629 	/* Disable send queue */
630 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, 0);
631 	/* Check if SQ is stopped */
632 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_SQ_0_7_STATUS, 21, 1, 0x01))
633 		return;
634 	/* Reset send queue */
635 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET);
636 }
637 
638 static void nicvf_reclaim_rcv_queue(struct nicvf *nic,
639 				    struct queue_set *qs, int qidx)
640 {
641 	union nic_mbx mbx = {};
642 
643 	/* Make sure all packets in the pipeline are written back into mem */
644 	mbx.msg.msg = NIC_MBOX_MSG_RQ_SW_SYNC;
645 	nicvf_send_msg_to_pf(nic, &mbx);
646 }
647 
648 static void nicvf_reclaim_cmp_queue(struct nicvf *nic,
649 				    struct queue_set *qs, int qidx)
650 {
651 	/* Disable timer threshold (doesn't get reset upon CQ reset */
652 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2, qidx, 0);
653 	/* Disable completion queue */
654 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, 0);
655 	/* Reset completion queue */
656 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET);
657 }
658 
659 static void nicvf_reclaim_rbdr(struct nicvf *nic,
660 			       struct rbdr *rbdr, int qidx)
661 {
662 	u64 tmp, fifo_state;
663 	int timeout = 10;
664 
665 	/* Save head and tail pointers for feeing up buffers */
666 	rbdr->head = nicvf_queue_reg_read(nic,
667 					  NIC_QSET_RBDR_0_1_HEAD,
668 					  qidx) >> 3;
669 	rbdr->tail = nicvf_queue_reg_read(nic,
670 					  NIC_QSET_RBDR_0_1_TAIL,
671 					  qidx) >> 3;
672 
673 	/* If RBDR FIFO is in 'FAIL' state then do a reset first
674 	 * before relaiming.
675 	 */
676 	fifo_state = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, qidx);
677 	if (((fifo_state >> 62) & 0x03) == 0x3)
678 		nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
679 				      qidx, NICVF_RBDR_RESET);
680 
681 	/* Disable RBDR */
682 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0);
683 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00))
684 		return;
685 	while (1) {
686 		tmp = nicvf_queue_reg_read(nic,
687 					   NIC_QSET_RBDR_0_1_PREFETCH_STATUS,
688 					   qidx);
689 		if ((tmp & 0xFFFFFFFF) == ((tmp >> 32) & 0xFFFFFFFF))
690 			break;
691 		usleep_range(1000, 2000);
692 		timeout--;
693 		if (!timeout) {
694 			netdev_err(nic->netdev,
695 				   "Failed polling on prefetch status\n");
696 			return;
697 		}
698 	}
699 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
700 			      qidx, NICVF_RBDR_RESET);
701 
702 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x02))
703 		return;
704 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0x00);
705 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00))
706 		return;
707 }
708 
709 void nicvf_config_vlan_stripping(struct nicvf *nic, netdev_features_t features)
710 {
711 	u64 rq_cfg;
712 	int sqs;
713 
714 	rq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_RQ_GEN_CFG, 0);
715 
716 	/* Enable first VLAN stripping */
717 	if (features & NETIF_F_HW_VLAN_CTAG_RX)
718 		rq_cfg |= (1ULL << 25);
719 	else
720 		rq_cfg &= ~(1ULL << 25);
721 	nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0, rq_cfg);
722 
723 	/* Configure Secondary Qsets, if any */
724 	for (sqs = 0; sqs < nic->sqs_count; sqs++)
725 		if (nic->snicvf[sqs])
726 			nicvf_queue_reg_write(nic->snicvf[sqs],
727 					      NIC_QSET_RQ_GEN_CFG, 0, rq_cfg);
728 }
729 
730 static void nicvf_reset_rcv_queue_stats(struct nicvf *nic)
731 {
732 	union nic_mbx mbx = {};
733 
734 	/* Reset all RQ/SQ and VF stats */
735 	mbx.reset_stat.msg = NIC_MBOX_MSG_RESET_STAT_COUNTER;
736 	mbx.reset_stat.rx_stat_mask = 0x3FFF;
737 	mbx.reset_stat.tx_stat_mask = 0x1F;
738 	mbx.reset_stat.rq_stat_mask = 0xFFFF;
739 	mbx.reset_stat.sq_stat_mask = 0xFFFF;
740 	nicvf_send_msg_to_pf(nic, &mbx);
741 }
742 
743 /* Configures receive queue */
744 static void nicvf_rcv_queue_config(struct nicvf *nic, struct queue_set *qs,
745 				   int qidx, bool enable)
746 {
747 	union nic_mbx mbx = {};
748 	struct rcv_queue *rq;
749 	struct rq_cfg rq_cfg;
750 
751 	rq = &qs->rq[qidx];
752 	rq->enable = enable;
753 
754 	/* Disable receive queue */
755 	nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, 0);
756 
757 	if (!rq->enable) {
758 		nicvf_reclaim_rcv_queue(nic, qs, qidx);
759 		xdp_rxq_info_unreg(&rq->xdp_rxq);
760 		return;
761 	}
762 
763 	rq->cq_qs = qs->vnic_id;
764 	rq->cq_idx = qidx;
765 	rq->start_rbdr_qs = qs->vnic_id;
766 	rq->start_qs_rbdr_idx = qs->rbdr_cnt - 1;
767 	rq->cont_rbdr_qs = qs->vnic_id;
768 	rq->cont_qs_rbdr_idx = qs->rbdr_cnt - 1;
769 	/* all writes of RBDR data to be loaded into L2 Cache as well*/
770 	rq->caching = 1;
771 
772 	/* Driver have no proper error path for failed XDP RX-queue info reg */
773 	WARN_ON(xdp_rxq_info_reg(&rq->xdp_rxq, nic->netdev, qidx, 0) < 0);
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 = ((u64)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 		/* Enable Tx timestamping capability */
982 		if (nic->ptp_clock)
983 			qs_cfg->send_tstmp_ena = 1;
984 	}
985 	nicvf_send_msg_to_pf(nic, &mbx);
986 }
987 
988 static void nicvf_free_resources(struct nicvf *nic)
989 {
990 	int qidx;
991 	struct queue_set *qs = nic->qs;
992 
993 	/* Free receive buffer descriptor ring */
994 	for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
995 		nicvf_free_rbdr(nic, &qs->rbdr[qidx]);
996 
997 	/* Free completion queue */
998 	for (qidx = 0; qidx < qs->cq_cnt; qidx++)
999 		nicvf_free_cmp_queue(nic, &qs->cq[qidx]);
1000 
1001 	/* Free send queue */
1002 	for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1003 		nicvf_free_snd_queue(nic, &qs->sq[qidx]);
1004 }
1005 
1006 static int nicvf_alloc_resources(struct nicvf *nic)
1007 {
1008 	int qidx;
1009 	struct queue_set *qs = nic->qs;
1010 
1011 	/* Alloc receive buffer descriptor ring */
1012 	for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) {
1013 		if (nicvf_init_rbdr(nic, &qs->rbdr[qidx], qs->rbdr_len,
1014 				    DMA_BUFFER_LEN))
1015 			goto alloc_fail;
1016 	}
1017 
1018 	/* Alloc send queue */
1019 	for (qidx = 0; qidx < qs->sq_cnt; qidx++) {
1020 		if (nicvf_init_snd_queue(nic, &qs->sq[qidx], qs->sq_len, qidx))
1021 			goto alloc_fail;
1022 	}
1023 
1024 	/* Alloc completion queue */
1025 	for (qidx = 0; qidx < qs->cq_cnt; qidx++) {
1026 		if (nicvf_init_cmp_queue(nic, &qs->cq[qidx], qs->cq_len))
1027 			goto alloc_fail;
1028 	}
1029 
1030 	return 0;
1031 alloc_fail:
1032 	nicvf_free_resources(nic);
1033 	return -ENOMEM;
1034 }
1035 
1036 int nicvf_set_qset_resources(struct nicvf *nic)
1037 {
1038 	struct queue_set *qs;
1039 
1040 	qs = devm_kzalloc(&nic->pdev->dev, sizeof(*qs), GFP_KERNEL);
1041 	if (!qs)
1042 		return -ENOMEM;
1043 	nic->qs = qs;
1044 
1045 	/* Set count of each queue */
1046 	qs->rbdr_cnt = DEFAULT_RBDR_CNT;
1047 	qs->rq_cnt = min_t(u8, MAX_RCV_QUEUES_PER_QS, num_online_cpus());
1048 	qs->sq_cnt = min_t(u8, MAX_SND_QUEUES_PER_QS, num_online_cpus());
1049 	qs->cq_cnt = max_t(u8, qs->rq_cnt, qs->sq_cnt);
1050 
1051 	/* Set queue lengths */
1052 	qs->rbdr_len = RCV_BUF_COUNT;
1053 	qs->sq_len = SND_QUEUE_LEN;
1054 	qs->cq_len = CMP_QUEUE_LEN;
1055 
1056 	nic->rx_queues = qs->rq_cnt;
1057 	nic->tx_queues = qs->sq_cnt;
1058 	nic->xdp_tx_queues = 0;
1059 
1060 	return 0;
1061 }
1062 
1063 int nicvf_config_data_transfer(struct nicvf *nic, bool enable)
1064 {
1065 	bool disable = false;
1066 	struct queue_set *qs = nic->qs;
1067 	struct queue_set *pqs = nic->pnicvf->qs;
1068 	int qidx;
1069 
1070 	if (!qs)
1071 		return 0;
1072 
1073 	/* Take primary VF's queue lengths.
1074 	 * This is needed to take queue lengths set from ethtool
1075 	 * into consideration.
1076 	 */
1077 	if (nic->sqs_mode && pqs) {
1078 		qs->cq_len = pqs->cq_len;
1079 		qs->sq_len = pqs->sq_len;
1080 	}
1081 
1082 	if (enable) {
1083 		if (nicvf_alloc_resources(nic))
1084 			return -ENOMEM;
1085 
1086 		for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1087 			nicvf_snd_queue_config(nic, qs, qidx, enable);
1088 		for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1089 			nicvf_cmp_queue_config(nic, qs, qidx, enable);
1090 		for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
1091 			nicvf_rbdr_config(nic, qs, qidx, enable);
1092 		for (qidx = 0; qidx < qs->rq_cnt; qidx++)
1093 			nicvf_rcv_queue_config(nic, qs, qidx, enable);
1094 	} else {
1095 		for (qidx = 0; qidx < qs->rq_cnt; qidx++)
1096 			nicvf_rcv_queue_config(nic, qs, qidx, disable);
1097 		for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
1098 			nicvf_rbdr_config(nic, qs, qidx, disable);
1099 		for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1100 			nicvf_snd_queue_config(nic, qs, qidx, disable);
1101 		for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1102 			nicvf_cmp_queue_config(nic, qs, qidx, disable);
1103 
1104 		nicvf_free_resources(nic);
1105 	}
1106 
1107 	/* Reset RXQ's stats.
1108 	 * SQ's stats will get reset automatically once SQ is reset.
1109 	 */
1110 	nicvf_reset_rcv_queue_stats(nic);
1111 
1112 	return 0;
1113 }
1114 
1115 /* Get a free desc from SQ
1116  * returns descriptor ponter & descriptor number
1117  */
1118 static inline int nicvf_get_sq_desc(struct snd_queue *sq, int desc_cnt)
1119 {
1120 	int qentry;
1121 
1122 	qentry = sq->tail;
1123 	if (!sq->is_xdp)
1124 		atomic_sub(desc_cnt, &sq->free_cnt);
1125 	else
1126 		sq->xdp_free_cnt -= desc_cnt;
1127 	sq->tail += desc_cnt;
1128 	sq->tail &= (sq->dmem.q_len - 1);
1129 
1130 	return qentry;
1131 }
1132 
1133 /* Rollback to previous tail pointer when descriptors not used */
1134 static inline void nicvf_rollback_sq_desc(struct snd_queue *sq,
1135 					  int qentry, int desc_cnt)
1136 {
1137 	sq->tail = qentry;
1138 	atomic_add(desc_cnt, &sq->free_cnt);
1139 }
1140 
1141 /* Free descriptor back to SQ for future use */
1142 void nicvf_put_sq_desc(struct snd_queue *sq, int desc_cnt)
1143 {
1144 	if (!sq->is_xdp)
1145 		atomic_add(desc_cnt, &sq->free_cnt);
1146 	else
1147 		sq->xdp_free_cnt += desc_cnt;
1148 	sq->head += desc_cnt;
1149 	sq->head &= (sq->dmem.q_len - 1);
1150 }
1151 
1152 static inline int nicvf_get_nxt_sqentry(struct snd_queue *sq, int qentry)
1153 {
1154 	qentry++;
1155 	qentry &= (sq->dmem.q_len - 1);
1156 	return qentry;
1157 }
1158 
1159 void nicvf_sq_enable(struct nicvf *nic, struct snd_queue *sq, int qidx)
1160 {
1161 	u64 sq_cfg;
1162 
1163 	sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx);
1164 	sq_cfg |= NICVF_SQ_EN;
1165 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg);
1166 	/* Ring doorbell so that H/W restarts processing SQEs */
1167 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR, qidx, 0);
1168 }
1169 
1170 void nicvf_sq_disable(struct nicvf *nic, int qidx)
1171 {
1172 	u64 sq_cfg;
1173 
1174 	sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx);
1175 	sq_cfg &= ~NICVF_SQ_EN;
1176 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg);
1177 }
1178 
1179 void nicvf_sq_free_used_descs(struct net_device *netdev, struct snd_queue *sq,
1180 			      int qidx)
1181 {
1182 	u64 head;
1183 	struct sk_buff *skb;
1184 	struct nicvf *nic = netdev_priv(netdev);
1185 	struct sq_hdr_subdesc *hdr;
1186 
1187 	head = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_HEAD, qidx) >> 4;
1188 	while (sq->head != head) {
1189 		hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, sq->head);
1190 		if (hdr->subdesc_type != SQ_DESC_TYPE_HEADER) {
1191 			nicvf_put_sq_desc(sq, 1);
1192 			continue;
1193 		}
1194 		skb = (struct sk_buff *)sq->skbuff[sq->head];
1195 		if (skb)
1196 			dev_kfree_skb_any(skb);
1197 		atomic64_add(1, (atomic64_t *)&netdev->stats.tx_packets);
1198 		atomic64_add(hdr->tot_len,
1199 			     (atomic64_t *)&netdev->stats.tx_bytes);
1200 		nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1);
1201 	}
1202 }
1203 
1204 /* XDP Transmit APIs */
1205 void nicvf_xdp_sq_doorbell(struct nicvf *nic,
1206 			   struct snd_queue *sq, int sq_num)
1207 {
1208 	if (!sq->xdp_desc_cnt)
1209 		return;
1210 
1211 	/* make sure all memory stores are done before ringing doorbell */
1212 	wmb();
1213 
1214 	/* Inform HW to xmit all TSO segments */
1215 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR,
1216 			      sq_num, sq->xdp_desc_cnt);
1217 	sq->xdp_desc_cnt = 0;
1218 }
1219 
1220 static inline void
1221 nicvf_xdp_sq_add_hdr_subdesc(struct snd_queue *sq, int qentry,
1222 			     int subdesc_cnt, u64 data, int len)
1223 {
1224 	struct sq_hdr_subdesc *hdr;
1225 
1226 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1227 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1228 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1229 	hdr->subdesc_cnt = subdesc_cnt;
1230 	hdr->tot_len = len;
1231 	hdr->post_cqe = 1;
1232 	sq->xdp_page[qentry] = (u64)virt_to_page((void *)data);
1233 }
1234 
1235 int nicvf_xdp_sq_append_pkt(struct nicvf *nic, struct snd_queue *sq,
1236 			    u64 bufaddr, u64 dma_addr, u16 len)
1237 {
1238 	int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT;
1239 	int qentry;
1240 
1241 	if (subdesc_cnt > sq->xdp_free_cnt)
1242 		return 0;
1243 
1244 	qentry = nicvf_get_sq_desc(sq, subdesc_cnt);
1245 
1246 	nicvf_xdp_sq_add_hdr_subdesc(sq, qentry, subdesc_cnt - 1, bufaddr, len);
1247 
1248 	qentry = nicvf_get_nxt_sqentry(sq, qentry);
1249 	nicvf_sq_add_gather_subdesc(sq, qentry, len, dma_addr);
1250 
1251 	sq->xdp_desc_cnt += subdesc_cnt;
1252 
1253 	return 1;
1254 }
1255 
1256 /* Calculate no of SQ subdescriptors needed to transmit all
1257  * segments of this TSO packet.
1258  * Taken from 'Tilera network driver' with a minor modification.
1259  */
1260 static int nicvf_tso_count_subdescs(struct sk_buff *skb)
1261 {
1262 	struct skb_shared_info *sh = skb_shinfo(skb);
1263 	unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1264 	unsigned int data_len = skb->len - sh_len;
1265 	unsigned int p_len = sh->gso_size;
1266 	long f_id = -1;    /* id of the current fragment */
1267 	long f_size = skb_headlen(skb) - sh_len;  /* current fragment size */
1268 	long f_used = 0;  /* bytes used from the current fragment */
1269 	long n;            /* size of the current piece of payload */
1270 	int num_edescs = 0;
1271 	int segment;
1272 
1273 	for (segment = 0; segment < sh->gso_segs; segment++) {
1274 		unsigned int p_used = 0;
1275 
1276 		/* One edesc for header and for each piece of the payload. */
1277 		for (num_edescs++; p_used < p_len; num_edescs++) {
1278 			/* Advance as needed. */
1279 			while (f_used >= f_size) {
1280 				f_id++;
1281 				f_size = skb_frag_size(&sh->frags[f_id]);
1282 				f_used = 0;
1283 			}
1284 
1285 			/* Use bytes from the current fragment. */
1286 			n = p_len - p_used;
1287 			if (n > f_size - f_used)
1288 				n = f_size - f_used;
1289 			f_used += n;
1290 			p_used += n;
1291 		}
1292 
1293 		/* The last segment may be less than gso_size. */
1294 		data_len -= p_len;
1295 		if (data_len < p_len)
1296 			p_len = data_len;
1297 	}
1298 
1299 	/* '+ gso_segs' for SQ_HDR_SUDESCs for each segment */
1300 	return num_edescs + sh->gso_segs;
1301 }
1302 
1303 #define POST_CQE_DESC_COUNT 2
1304 
1305 /* Get the number of SQ descriptors needed to xmit this skb */
1306 static int nicvf_sq_subdesc_required(struct nicvf *nic, struct sk_buff *skb)
1307 {
1308 	int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT;
1309 
1310 	if (skb_shinfo(skb)->gso_size && !nic->hw_tso) {
1311 		subdesc_cnt = nicvf_tso_count_subdescs(skb);
1312 		return subdesc_cnt;
1313 	}
1314 
1315 	/* Dummy descriptors to get TSO pkt completion notification */
1316 	if (nic->t88 && nic->hw_tso && skb_shinfo(skb)->gso_size)
1317 		subdesc_cnt += POST_CQE_DESC_COUNT;
1318 
1319 	if (skb_shinfo(skb)->nr_frags)
1320 		subdesc_cnt += skb_shinfo(skb)->nr_frags;
1321 
1322 	return subdesc_cnt;
1323 }
1324 
1325 /* Add SQ HEADER subdescriptor.
1326  * First subdescriptor for every send descriptor.
1327  */
1328 static inline void
1329 nicvf_sq_add_hdr_subdesc(struct nicvf *nic, struct snd_queue *sq, int qentry,
1330 			 int subdesc_cnt, struct sk_buff *skb, int len)
1331 {
1332 	int proto;
1333 	struct sq_hdr_subdesc *hdr;
1334 	union {
1335 		struct iphdr *v4;
1336 		struct ipv6hdr *v6;
1337 		unsigned char *hdr;
1338 	} ip;
1339 
1340 	ip.hdr = skb_network_header(skb);
1341 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1342 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1343 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1344 
1345 	if (nic->t88 && nic->hw_tso && skb_shinfo(skb)->gso_size) {
1346 		/* post_cqe = 0, to avoid HW posting a CQE for every TSO
1347 		 * segment transmitted on 88xx.
1348 		 */
1349 		hdr->subdesc_cnt = subdesc_cnt - POST_CQE_DESC_COUNT;
1350 	} else {
1351 		sq->skbuff[qentry] = (u64)skb;
1352 		/* Enable notification via CQE after processing SQE */
1353 		hdr->post_cqe = 1;
1354 		/* No of subdescriptors following this */
1355 		hdr->subdesc_cnt = subdesc_cnt;
1356 	}
1357 	hdr->tot_len = len;
1358 
1359 	/* Offload checksum calculation to HW */
1360 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
1361 		if (ip.v4->version == 4)
1362 			hdr->csum_l3 = 1; /* Enable IP csum calculation */
1363 		hdr->l3_offset = skb_network_offset(skb);
1364 		hdr->l4_offset = skb_transport_offset(skb);
1365 
1366 		proto = (ip.v4->version == 4) ? ip.v4->protocol :
1367 			ip.v6->nexthdr;
1368 
1369 		switch (proto) {
1370 		case IPPROTO_TCP:
1371 			hdr->csum_l4 = SEND_L4_CSUM_TCP;
1372 			break;
1373 		case IPPROTO_UDP:
1374 			hdr->csum_l4 = SEND_L4_CSUM_UDP;
1375 			break;
1376 		case IPPROTO_SCTP:
1377 			hdr->csum_l4 = SEND_L4_CSUM_SCTP;
1378 			break;
1379 		}
1380 	}
1381 
1382 	if (nic->hw_tso && skb_shinfo(skb)->gso_size) {
1383 		hdr->tso = 1;
1384 		hdr->tso_start = skb_transport_offset(skb) + tcp_hdrlen(skb);
1385 		hdr->tso_max_paysize = skb_shinfo(skb)->gso_size;
1386 		/* For non-tunneled pkts, point this to L2 ethertype */
1387 		hdr->inner_l3_offset = skb_network_offset(skb) - 2;
1388 		this_cpu_inc(nic->pnicvf->drv_stats->tx_tso);
1389 	}
1390 
1391 	/* Check if timestamp is requested */
1392 	if (!(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)) {
1393 		skb_tx_timestamp(skb);
1394 		return;
1395 	}
1396 
1397 	/* Tx timestamping not supported along with TSO, so ignore request */
1398 	if (skb_shinfo(skb)->gso_size)
1399 		return;
1400 
1401 	/* HW supports only a single outstanding packet to timestamp */
1402 	if (!atomic_add_unless(&nic->pnicvf->tx_ptp_skbs, 1, 1))
1403 		return;
1404 
1405 	/* Mark the SKB for later reference */
1406 	skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
1407 
1408 	/* Finally enable timestamp generation
1409 	 * Since 'post_cqe' is also set, two CQEs will be posted
1410 	 * for this packet i.e CQE_TYPE_SEND and CQE_TYPE_SEND_PTP.
1411 	 */
1412 	hdr->tstmp = 1;
1413 }
1414 
1415 /* SQ GATHER subdescriptor
1416  * Must follow HDR descriptor
1417  */
1418 static inline void nicvf_sq_add_gather_subdesc(struct snd_queue *sq, int qentry,
1419 					       int size, u64 data)
1420 {
1421 	struct sq_gather_subdesc *gather;
1422 
1423 	qentry &= (sq->dmem.q_len - 1);
1424 	gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, qentry);
1425 
1426 	memset(gather, 0, SND_QUEUE_DESC_SIZE);
1427 	gather->subdesc_type = SQ_DESC_TYPE_GATHER;
1428 	gather->ld_type = NIC_SEND_LD_TYPE_E_LDD;
1429 	gather->size = size;
1430 	gather->addr = data;
1431 }
1432 
1433 /* Add HDR + IMMEDIATE subdescriptors right after descriptors of a TSO
1434  * packet so that a CQE is posted as a notifation for transmission of
1435  * TSO packet.
1436  */
1437 static inline void nicvf_sq_add_cqe_subdesc(struct snd_queue *sq, int qentry,
1438 					    int tso_sqe, struct sk_buff *skb)
1439 {
1440 	struct sq_imm_subdesc *imm;
1441 	struct sq_hdr_subdesc *hdr;
1442 
1443 	sq->skbuff[qentry] = (u64)skb;
1444 
1445 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1446 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1447 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1448 	/* Enable notification via CQE after processing SQE */
1449 	hdr->post_cqe = 1;
1450 	/* There is no packet to transmit here */
1451 	hdr->dont_send = 1;
1452 	hdr->subdesc_cnt = POST_CQE_DESC_COUNT - 1;
1453 	hdr->tot_len = 1;
1454 	/* Actual TSO header SQE index, needed for cleanup */
1455 	hdr->rsvd2 = tso_sqe;
1456 
1457 	qentry = nicvf_get_nxt_sqentry(sq, qentry);
1458 	imm = (struct sq_imm_subdesc *)GET_SQ_DESC(sq, qentry);
1459 	memset(imm, 0, SND_QUEUE_DESC_SIZE);
1460 	imm->subdesc_type = SQ_DESC_TYPE_IMMEDIATE;
1461 	imm->len = 1;
1462 }
1463 
1464 static inline void nicvf_sq_doorbell(struct nicvf *nic, struct sk_buff *skb,
1465 				     int sq_num, int desc_cnt)
1466 {
1467 	struct netdev_queue *txq;
1468 
1469 	txq = netdev_get_tx_queue(nic->pnicvf->netdev,
1470 				  skb_get_queue_mapping(skb));
1471 
1472 	netdev_tx_sent_queue(txq, skb->len);
1473 
1474 	/* make sure all memory stores are done before ringing doorbell */
1475 	smp_wmb();
1476 
1477 	/* Inform HW to xmit all TSO segments */
1478 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR,
1479 			      sq_num, desc_cnt);
1480 }
1481 
1482 /* Segment a TSO packet into 'gso_size' segments and append
1483  * them to SQ for transfer
1484  */
1485 static int nicvf_sq_append_tso(struct nicvf *nic, struct snd_queue *sq,
1486 			       int sq_num, int qentry, struct sk_buff *skb)
1487 {
1488 	struct tso_t tso;
1489 	int seg_subdescs = 0, desc_cnt = 0;
1490 	int seg_len, total_len, data_left;
1491 	int hdr_qentry = qentry;
1492 	int hdr_len;
1493 
1494 	hdr_len = tso_start(skb, &tso);
1495 
1496 	total_len = skb->len - hdr_len;
1497 	while (total_len > 0) {
1498 		char *hdr;
1499 
1500 		/* Save Qentry for adding HDR_SUBDESC at the end */
1501 		hdr_qentry = qentry;
1502 
1503 		data_left = min_t(int, skb_shinfo(skb)->gso_size, total_len);
1504 		total_len -= data_left;
1505 
1506 		/* Add segment's header */
1507 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1508 		hdr = sq->tso_hdrs + qentry * TSO_HEADER_SIZE;
1509 		tso_build_hdr(skb, hdr, &tso, data_left, total_len == 0);
1510 		nicvf_sq_add_gather_subdesc(sq, qentry, hdr_len,
1511 					    sq->tso_hdrs_phys +
1512 					    qentry * TSO_HEADER_SIZE);
1513 		/* HDR_SUDESC + GATHER */
1514 		seg_subdescs = 2;
1515 		seg_len = hdr_len;
1516 
1517 		/* Add segment's payload fragments */
1518 		while (data_left > 0) {
1519 			int size;
1520 
1521 			size = min_t(int, tso.size, data_left);
1522 
1523 			qentry = nicvf_get_nxt_sqentry(sq, qentry);
1524 			nicvf_sq_add_gather_subdesc(sq, qentry, size,
1525 						    virt_to_phys(tso.data));
1526 			seg_subdescs++;
1527 			seg_len += size;
1528 
1529 			data_left -= size;
1530 			tso_build_data(skb, &tso, size);
1531 		}
1532 		nicvf_sq_add_hdr_subdesc(nic, sq, hdr_qentry,
1533 					 seg_subdescs - 1, skb, seg_len);
1534 		sq->skbuff[hdr_qentry] = (u64)NULL;
1535 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1536 
1537 		desc_cnt += seg_subdescs;
1538 	}
1539 	/* Save SKB in the last segment for freeing */
1540 	sq->skbuff[hdr_qentry] = (u64)skb;
1541 
1542 	nicvf_sq_doorbell(nic, skb, sq_num, desc_cnt);
1543 
1544 	this_cpu_inc(nic->pnicvf->drv_stats->tx_tso);
1545 	return 1;
1546 }
1547 
1548 /* Append an skb to a SQ for packet transfer. */
1549 int nicvf_sq_append_skb(struct nicvf *nic, struct snd_queue *sq,
1550 			struct sk_buff *skb, u8 sq_num)
1551 {
1552 	int i, size;
1553 	int subdesc_cnt, hdr_sqe = 0;
1554 	int qentry;
1555 	u64 dma_addr;
1556 
1557 	subdesc_cnt = nicvf_sq_subdesc_required(nic, skb);
1558 	if (subdesc_cnt > atomic_read(&sq->free_cnt))
1559 		goto append_fail;
1560 
1561 	qentry = nicvf_get_sq_desc(sq, subdesc_cnt);
1562 
1563 	/* Check if its a TSO packet */
1564 	if (skb_shinfo(skb)->gso_size && !nic->hw_tso)
1565 		return nicvf_sq_append_tso(nic, sq, sq_num, qentry, skb);
1566 
1567 	/* Add SQ header subdesc */
1568 	nicvf_sq_add_hdr_subdesc(nic, sq, qentry, subdesc_cnt - 1,
1569 				 skb, skb->len);
1570 	hdr_sqe = qentry;
1571 
1572 	/* Add SQ gather subdescs */
1573 	qentry = nicvf_get_nxt_sqentry(sq, qentry);
1574 	size = skb_is_nonlinear(skb) ? skb_headlen(skb) : skb->len;
1575 	/* HW will ensure data coherency, CPU sync not required */
1576 	dma_addr = dma_map_page_attrs(&nic->pdev->dev, virt_to_page(skb->data),
1577 				      offset_in_page(skb->data), size,
1578 				      DMA_TO_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
1579 	if (dma_mapping_error(&nic->pdev->dev, dma_addr)) {
1580 		nicvf_rollback_sq_desc(sq, qentry, subdesc_cnt);
1581 		return 0;
1582 	}
1583 
1584 	nicvf_sq_add_gather_subdesc(sq, qentry, size, dma_addr);
1585 
1586 	/* Check for scattered buffer */
1587 	if (!skb_is_nonlinear(skb))
1588 		goto doorbell;
1589 
1590 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1591 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1592 
1593 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1594 		size = skb_frag_size(frag);
1595 		dma_addr = dma_map_page_attrs(&nic->pdev->dev,
1596 					      skb_frag_page(frag),
1597 					      skb_frag_off(frag), size,
1598 					      DMA_TO_DEVICE,
1599 					      DMA_ATTR_SKIP_CPU_SYNC);
1600 		if (dma_mapping_error(&nic->pdev->dev, dma_addr)) {
1601 			/* Free entire chain of mapped buffers
1602 			 * here 'i' = frags mapped + above mapped skb->data
1603 			 */
1604 			nicvf_unmap_sndq_buffers(nic, sq, hdr_sqe, i);
1605 			nicvf_rollback_sq_desc(sq, qentry, subdesc_cnt);
1606 			return 0;
1607 		}
1608 		nicvf_sq_add_gather_subdesc(sq, qentry, size, dma_addr);
1609 	}
1610 
1611 doorbell:
1612 	if (nic->t88 && skb_shinfo(skb)->gso_size) {
1613 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1614 		nicvf_sq_add_cqe_subdesc(sq, qentry, hdr_sqe, skb);
1615 	}
1616 
1617 	nicvf_sq_doorbell(nic, skb, sq_num, subdesc_cnt);
1618 
1619 	return 1;
1620 
1621 append_fail:
1622 	/* Use original PCI dev for debug log */
1623 	nic = nic->pnicvf;
1624 	netdev_dbg(nic->netdev, "Not enough SQ descriptors to xmit pkt\n");
1625 	return 0;
1626 }
1627 
1628 static inline unsigned frag_num(unsigned i)
1629 {
1630 #ifdef __BIG_ENDIAN
1631 	return (i & ~3) + 3 - (i & 3);
1632 #else
1633 	return i;
1634 #endif
1635 }
1636 
1637 static void nicvf_unmap_rcv_buffer(struct nicvf *nic, u64 dma_addr,
1638 				   u64 buf_addr, bool xdp)
1639 {
1640 	struct page *page = NULL;
1641 	int len = RCV_FRAG_LEN;
1642 
1643 	if (xdp) {
1644 		page = virt_to_page(phys_to_virt(buf_addr));
1645 		/* Check if it's a recycled page, if not
1646 		 * unmap the DMA mapping.
1647 		 *
1648 		 * Recycled page holds an extra reference.
1649 		 */
1650 		if (page_ref_count(page) != 1)
1651 			return;
1652 
1653 		len += XDP_PACKET_HEADROOM;
1654 		/* Receive buffers in XDP mode are mapped from page start */
1655 		dma_addr &= PAGE_MASK;
1656 	}
1657 	dma_unmap_page_attrs(&nic->pdev->dev, dma_addr, len,
1658 			     DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
1659 }
1660 
1661 /* Returns SKB for a received packet */
1662 struct sk_buff *nicvf_get_rcv_skb(struct nicvf *nic,
1663 				  struct cqe_rx_t *cqe_rx, bool xdp)
1664 {
1665 	int frag;
1666 	int payload_len = 0;
1667 	struct sk_buff *skb = NULL;
1668 	struct page *page;
1669 	int offset;
1670 	u16 *rb_lens = NULL;
1671 	u64 *rb_ptrs = NULL;
1672 	u64 phys_addr;
1673 
1674 	rb_lens = (void *)cqe_rx + (3 * sizeof(u64));
1675 	/* Except 88xx pass1 on all other chips CQE_RX2_S is added to
1676 	 * CQE_RX at word6, hence buffer pointers move by word
1677 	 *
1678 	 * Use existing 'hw_tso' flag which will be set for all chips
1679 	 * except 88xx pass1 instead of a additional cache line
1680 	 * access (or miss) by using pci dev's revision.
1681 	 */
1682 	if (!nic->hw_tso)
1683 		rb_ptrs = (void *)cqe_rx + (6 * sizeof(u64));
1684 	else
1685 		rb_ptrs = (void *)cqe_rx + (7 * sizeof(u64));
1686 
1687 	for (frag = 0; frag < cqe_rx->rb_cnt; frag++) {
1688 		payload_len = rb_lens[frag_num(frag)];
1689 		phys_addr = nicvf_iova_to_phys(nic, *rb_ptrs);
1690 		if (!phys_addr) {
1691 			if (skb)
1692 				dev_kfree_skb_any(skb);
1693 			return NULL;
1694 		}
1695 
1696 		if (!frag) {
1697 			/* First fragment */
1698 			nicvf_unmap_rcv_buffer(nic,
1699 					       *rb_ptrs - cqe_rx->align_pad,
1700 					       phys_addr, xdp);
1701 			skb = nicvf_rb_ptr_to_skb(nic,
1702 						  phys_addr - cqe_rx->align_pad,
1703 						  payload_len);
1704 			if (!skb)
1705 				return NULL;
1706 			skb_reserve(skb, cqe_rx->align_pad);
1707 			skb_put(skb, payload_len);
1708 		} else {
1709 			/* Add fragments */
1710 			nicvf_unmap_rcv_buffer(nic, *rb_ptrs, phys_addr, xdp);
1711 			page = virt_to_page(phys_to_virt(phys_addr));
1712 			offset = phys_to_virt(phys_addr) - page_address(page);
1713 			skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page,
1714 					offset, payload_len, RCV_FRAG_LEN);
1715 		}
1716 		/* Next buffer pointer */
1717 		rb_ptrs++;
1718 	}
1719 	return skb;
1720 }
1721 
1722 static u64 nicvf_int_type_to_mask(int int_type, int q_idx)
1723 {
1724 	u64 reg_val;
1725 
1726 	switch (int_type) {
1727 	case NICVF_INTR_CQ:
1728 		reg_val = ((1ULL << q_idx) << NICVF_INTR_CQ_SHIFT);
1729 		break;
1730 	case NICVF_INTR_SQ:
1731 		reg_val = ((1ULL << q_idx) << NICVF_INTR_SQ_SHIFT);
1732 		break;
1733 	case NICVF_INTR_RBDR:
1734 		reg_val = ((1ULL << q_idx) << NICVF_INTR_RBDR_SHIFT);
1735 		break;
1736 	case NICVF_INTR_PKT_DROP:
1737 		reg_val = (1ULL << NICVF_INTR_PKT_DROP_SHIFT);
1738 		break;
1739 	case NICVF_INTR_TCP_TIMER:
1740 		reg_val = (1ULL << NICVF_INTR_TCP_TIMER_SHIFT);
1741 		break;
1742 	case NICVF_INTR_MBOX:
1743 		reg_val = (1ULL << NICVF_INTR_MBOX_SHIFT);
1744 		break;
1745 	case NICVF_INTR_QS_ERR:
1746 		reg_val = (1ULL << NICVF_INTR_QS_ERR_SHIFT);
1747 		break;
1748 	default:
1749 		reg_val = 0;
1750 	}
1751 
1752 	return reg_val;
1753 }
1754 
1755 /* Enable interrupt */
1756 void nicvf_enable_intr(struct nicvf *nic, int int_type, int q_idx)
1757 {
1758 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1759 
1760 	if (!mask) {
1761 		netdev_dbg(nic->netdev,
1762 			   "Failed to enable interrupt: unknown type\n");
1763 		return;
1764 	}
1765 	nicvf_reg_write(nic, NIC_VF_ENA_W1S,
1766 			nicvf_reg_read(nic, NIC_VF_ENA_W1S) | mask);
1767 }
1768 
1769 /* Disable interrupt */
1770 void nicvf_disable_intr(struct nicvf *nic, int int_type, int q_idx)
1771 {
1772 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1773 
1774 	if (!mask) {
1775 		netdev_dbg(nic->netdev,
1776 			   "Failed to disable interrupt: unknown type\n");
1777 		return;
1778 	}
1779 
1780 	nicvf_reg_write(nic, NIC_VF_ENA_W1C, mask);
1781 }
1782 
1783 /* Clear interrupt */
1784 void nicvf_clear_intr(struct nicvf *nic, int int_type, int q_idx)
1785 {
1786 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1787 
1788 	if (!mask) {
1789 		netdev_dbg(nic->netdev,
1790 			   "Failed to clear interrupt: unknown type\n");
1791 		return;
1792 	}
1793 
1794 	nicvf_reg_write(nic, NIC_VF_INT, mask);
1795 }
1796 
1797 /* Check if interrupt is enabled */
1798 int nicvf_is_intr_enabled(struct nicvf *nic, int int_type, int q_idx)
1799 {
1800 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1801 	/* If interrupt type is unknown, we treat it disabled. */
1802 	if (!mask) {
1803 		netdev_dbg(nic->netdev,
1804 			   "Failed to check interrupt enable: unknown type\n");
1805 		return 0;
1806 	}
1807 
1808 	return mask & nicvf_reg_read(nic, NIC_VF_ENA_W1S);
1809 }
1810 
1811 void nicvf_update_rq_stats(struct nicvf *nic, int rq_idx)
1812 {
1813 	struct rcv_queue *rq;
1814 
1815 #define GET_RQ_STATS(reg) \
1816 	nicvf_reg_read(nic, NIC_QSET_RQ_0_7_STAT_0_1 |\
1817 			    (rq_idx << NIC_Q_NUM_SHIFT) | (reg << 3))
1818 
1819 	rq = &nic->qs->rq[rq_idx];
1820 	rq->stats.bytes = GET_RQ_STATS(RQ_SQ_STATS_OCTS);
1821 	rq->stats.pkts = GET_RQ_STATS(RQ_SQ_STATS_PKTS);
1822 }
1823 
1824 void nicvf_update_sq_stats(struct nicvf *nic, int sq_idx)
1825 {
1826 	struct snd_queue *sq;
1827 
1828 #define GET_SQ_STATS(reg) \
1829 	nicvf_reg_read(nic, NIC_QSET_SQ_0_7_STAT_0_1 |\
1830 			    (sq_idx << NIC_Q_NUM_SHIFT) | (reg << 3))
1831 
1832 	sq = &nic->qs->sq[sq_idx];
1833 	sq->stats.bytes = GET_SQ_STATS(RQ_SQ_STATS_OCTS);
1834 	sq->stats.pkts = GET_SQ_STATS(RQ_SQ_STATS_PKTS);
1835 }
1836 
1837 /* Check for errors in the receive cmp.queue entry */
1838 int nicvf_check_cqe_rx_errs(struct nicvf *nic, struct cqe_rx_t *cqe_rx)
1839 {
1840 	netif_err(nic, rx_err, nic->netdev,
1841 		  "RX error CQE err_level 0x%x err_opcode 0x%x\n",
1842 		  cqe_rx->err_level, cqe_rx->err_opcode);
1843 
1844 	switch (cqe_rx->err_opcode) {
1845 	case CQ_RX_ERROP_RE_PARTIAL:
1846 		this_cpu_inc(nic->drv_stats->rx_bgx_truncated_pkts);
1847 		break;
1848 	case CQ_RX_ERROP_RE_JABBER:
1849 		this_cpu_inc(nic->drv_stats->rx_jabber_errs);
1850 		break;
1851 	case CQ_RX_ERROP_RE_FCS:
1852 		this_cpu_inc(nic->drv_stats->rx_fcs_errs);
1853 		break;
1854 	case CQ_RX_ERROP_RE_RX_CTL:
1855 		this_cpu_inc(nic->drv_stats->rx_bgx_errs);
1856 		break;
1857 	case CQ_RX_ERROP_PREL2_ERR:
1858 		this_cpu_inc(nic->drv_stats->rx_prel2_errs);
1859 		break;
1860 	case CQ_RX_ERROP_L2_MAL:
1861 		this_cpu_inc(nic->drv_stats->rx_l2_hdr_malformed);
1862 		break;
1863 	case CQ_RX_ERROP_L2_OVERSIZE:
1864 		this_cpu_inc(nic->drv_stats->rx_oversize);
1865 		break;
1866 	case CQ_RX_ERROP_L2_UNDERSIZE:
1867 		this_cpu_inc(nic->drv_stats->rx_undersize);
1868 		break;
1869 	case CQ_RX_ERROP_L2_LENMISM:
1870 		this_cpu_inc(nic->drv_stats->rx_l2_len_mismatch);
1871 		break;
1872 	case CQ_RX_ERROP_L2_PCLP:
1873 		this_cpu_inc(nic->drv_stats->rx_l2_pclp);
1874 		break;
1875 	case CQ_RX_ERROP_IP_NOT:
1876 		this_cpu_inc(nic->drv_stats->rx_ip_ver_errs);
1877 		break;
1878 	case CQ_RX_ERROP_IP_CSUM_ERR:
1879 		this_cpu_inc(nic->drv_stats->rx_ip_csum_errs);
1880 		break;
1881 	case CQ_RX_ERROP_IP_MAL:
1882 		this_cpu_inc(nic->drv_stats->rx_ip_hdr_malformed);
1883 		break;
1884 	case CQ_RX_ERROP_IP_MALD:
1885 		this_cpu_inc(nic->drv_stats->rx_ip_payload_malformed);
1886 		break;
1887 	case CQ_RX_ERROP_IP_HOP:
1888 		this_cpu_inc(nic->drv_stats->rx_ip_ttl_errs);
1889 		break;
1890 	case CQ_RX_ERROP_L3_PCLP:
1891 		this_cpu_inc(nic->drv_stats->rx_l3_pclp);
1892 		break;
1893 	case CQ_RX_ERROP_L4_MAL:
1894 		this_cpu_inc(nic->drv_stats->rx_l4_malformed);
1895 		break;
1896 	case CQ_RX_ERROP_L4_CHK:
1897 		this_cpu_inc(nic->drv_stats->rx_l4_csum_errs);
1898 		break;
1899 	case CQ_RX_ERROP_UDP_LEN:
1900 		this_cpu_inc(nic->drv_stats->rx_udp_len_errs);
1901 		break;
1902 	case CQ_RX_ERROP_L4_PORT:
1903 		this_cpu_inc(nic->drv_stats->rx_l4_port_errs);
1904 		break;
1905 	case CQ_RX_ERROP_TCP_FLAG:
1906 		this_cpu_inc(nic->drv_stats->rx_tcp_flag_errs);
1907 		break;
1908 	case CQ_RX_ERROP_TCP_OFFSET:
1909 		this_cpu_inc(nic->drv_stats->rx_tcp_offset_errs);
1910 		break;
1911 	case CQ_RX_ERROP_L4_PCLP:
1912 		this_cpu_inc(nic->drv_stats->rx_l4_pclp);
1913 		break;
1914 	case CQ_RX_ERROP_RBDR_TRUNC:
1915 		this_cpu_inc(nic->drv_stats->rx_truncated_pkts);
1916 		break;
1917 	}
1918 
1919 	return 1;
1920 }
1921 
1922 /* Check for errors in the send cmp.queue entry */
1923 int nicvf_check_cqe_tx_errs(struct nicvf *nic, struct cqe_send_t *cqe_tx)
1924 {
1925 	switch (cqe_tx->send_status) {
1926 	case CQ_TX_ERROP_DESC_FAULT:
1927 		this_cpu_inc(nic->drv_stats->tx_desc_fault);
1928 		break;
1929 	case CQ_TX_ERROP_HDR_CONS_ERR:
1930 		this_cpu_inc(nic->drv_stats->tx_hdr_cons_err);
1931 		break;
1932 	case CQ_TX_ERROP_SUBDC_ERR:
1933 		this_cpu_inc(nic->drv_stats->tx_subdesc_err);
1934 		break;
1935 	case CQ_TX_ERROP_MAX_SIZE_VIOL:
1936 		this_cpu_inc(nic->drv_stats->tx_max_size_exceeded);
1937 		break;
1938 	case CQ_TX_ERROP_IMM_SIZE_OFLOW:
1939 		this_cpu_inc(nic->drv_stats->tx_imm_size_oflow);
1940 		break;
1941 	case CQ_TX_ERROP_DATA_SEQUENCE_ERR:
1942 		this_cpu_inc(nic->drv_stats->tx_data_seq_err);
1943 		break;
1944 	case CQ_TX_ERROP_MEM_SEQUENCE_ERR:
1945 		this_cpu_inc(nic->drv_stats->tx_mem_seq_err);
1946 		break;
1947 	case CQ_TX_ERROP_LOCK_VIOL:
1948 		this_cpu_inc(nic->drv_stats->tx_lock_viol);
1949 		break;
1950 	case CQ_TX_ERROP_DATA_FAULT:
1951 		this_cpu_inc(nic->drv_stats->tx_data_fault);
1952 		break;
1953 	case CQ_TX_ERROP_TSTMP_CONFLICT:
1954 		this_cpu_inc(nic->drv_stats->tx_tstmp_conflict);
1955 		break;
1956 	case CQ_TX_ERROP_TSTMP_TIMEOUT:
1957 		this_cpu_inc(nic->drv_stats->tx_tstmp_timeout);
1958 		break;
1959 	case CQ_TX_ERROP_MEM_FAULT:
1960 		this_cpu_inc(nic->drv_stats->tx_mem_fault);
1961 		break;
1962 	case CQ_TX_ERROP_CK_OVERLAP:
1963 		this_cpu_inc(nic->drv_stats->tx_csum_overlap);
1964 		break;
1965 	case CQ_TX_ERROP_CK_OFLOW:
1966 		this_cpu_inc(nic->drv_stats->tx_csum_overflow);
1967 		break;
1968 	}
1969 
1970 	return 1;
1971 }
1972