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_PACKET_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 		if (pgcache)
228 			pgcache->dma_addr = *rbuf + XDP_PACKET_HEADROOM;
229 		nic->rb_page_offset += buf_len;
230 	}
231 
232 	return 0;
233 }
234 
235 /* Build skb around receive buffer */
236 static struct sk_buff *nicvf_rb_ptr_to_skb(struct nicvf *nic,
237 					   u64 rb_ptr, int len)
238 {
239 	void *data;
240 	struct sk_buff *skb;
241 
242 	data = phys_to_virt(rb_ptr);
243 
244 	/* Now build an skb to give to stack */
245 	skb = build_skb(data, RCV_FRAG_LEN);
246 	if (!skb) {
247 		put_page(virt_to_page(data));
248 		return NULL;
249 	}
250 
251 	prefetch(skb->data);
252 	return skb;
253 }
254 
255 /* Allocate RBDR ring and populate receive buffers */
256 static int  nicvf_init_rbdr(struct nicvf *nic, struct rbdr *rbdr,
257 			    int ring_len, int buf_size)
258 {
259 	int idx;
260 	u64 rbuf;
261 	struct rbdr_entry_t *desc;
262 	int err;
263 
264 	err = nicvf_alloc_q_desc_mem(nic, &rbdr->dmem, ring_len,
265 				     sizeof(struct rbdr_entry_t),
266 				     NICVF_RCV_BUF_ALIGN_BYTES);
267 	if (err)
268 		return err;
269 
270 	rbdr->desc = rbdr->dmem.base;
271 	/* Buffer size has to be in multiples of 128 bytes */
272 	rbdr->dma_size = buf_size;
273 	rbdr->enable = true;
274 	rbdr->thresh = RBDR_THRESH;
275 	rbdr->head = 0;
276 	rbdr->tail = 0;
277 
278 	/* Initialize page recycling stuff.
279 	 *
280 	 * Can't use single buffer per page especially with 64K pages.
281 	 * On embedded platforms i.e 81xx/83xx available memory itself
282 	 * is low and minimum ring size of RBDR is 8K, that takes away
283 	 * lots of memory.
284 	 *
285 	 * But for XDP it has to be a single buffer per page.
286 	 */
287 	if (!nic->pnicvf->xdp_prog) {
288 		rbdr->pgcnt = ring_len / (PAGE_SIZE / buf_size);
289 		rbdr->is_xdp = false;
290 	} else {
291 		rbdr->pgcnt = ring_len;
292 		rbdr->is_xdp = true;
293 	}
294 	rbdr->pgcnt = roundup_pow_of_two(rbdr->pgcnt);
295 	rbdr->pgcache = kcalloc(rbdr->pgcnt, sizeof(*rbdr->pgcache),
296 				GFP_KERNEL);
297 	if (!rbdr->pgcache)
298 		return -ENOMEM;
299 	rbdr->pgidx = 0;
300 	rbdr->pgalloc = 0;
301 
302 	nic->rb_page = NULL;
303 	for (idx = 0; idx < ring_len; idx++) {
304 		err = nicvf_alloc_rcv_buffer(nic, rbdr, GFP_KERNEL,
305 					     RCV_FRAG_LEN, &rbuf);
306 		if (err) {
307 			/* To free already allocated and mapped ones */
308 			rbdr->tail = idx - 1;
309 			return err;
310 		}
311 
312 		desc = GET_RBDR_DESC(rbdr, idx);
313 		desc->buf_addr = rbuf & ~(NICVF_RCV_BUF_ALIGN_BYTES - 1);
314 	}
315 
316 	nicvf_get_page(nic);
317 
318 	return 0;
319 }
320 
321 /* Free RBDR ring and its receive buffers */
322 static void nicvf_free_rbdr(struct nicvf *nic, struct rbdr *rbdr)
323 {
324 	int head, tail;
325 	u64 buf_addr, phys_addr;
326 	struct pgcache *pgcache;
327 	struct rbdr_entry_t *desc;
328 
329 	if (!rbdr)
330 		return;
331 
332 	rbdr->enable = false;
333 	if (!rbdr->dmem.base)
334 		return;
335 
336 	head = rbdr->head;
337 	tail = rbdr->tail;
338 
339 	/* Release page references */
340 	while (head != tail) {
341 		desc = GET_RBDR_DESC(rbdr, head);
342 		buf_addr = desc->buf_addr;
343 		phys_addr = nicvf_iova_to_phys(nic, buf_addr);
344 		dma_unmap_page_attrs(&nic->pdev->dev, buf_addr, RCV_FRAG_LEN,
345 				     DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
346 		if (phys_addr)
347 			put_page(virt_to_page(phys_to_virt(phys_addr)));
348 		head++;
349 		head &= (rbdr->dmem.q_len - 1);
350 	}
351 	/* Release buffer of tail desc */
352 	desc = GET_RBDR_DESC(rbdr, tail);
353 	buf_addr = desc->buf_addr;
354 	phys_addr = nicvf_iova_to_phys(nic, buf_addr);
355 	dma_unmap_page_attrs(&nic->pdev->dev, buf_addr, RCV_FRAG_LEN,
356 			     DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
357 	if (phys_addr)
358 		put_page(virt_to_page(phys_to_virt(phys_addr)));
359 
360 	/* Sync page cache info */
361 	smp_rmb();
362 
363 	/* Release additional page references held for recycling */
364 	head = 0;
365 	while (head < rbdr->pgcnt) {
366 		pgcache = &rbdr->pgcache[head];
367 		if (pgcache->page && page_ref_count(pgcache->page) != 0) {
368 			if (!rbdr->is_xdp) {
369 				put_page(pgcache->page);
370 				continue;
371 			}
372 			page_ref_sub(pgcache->page, pgcache->ref_count - 1);
373 			put_page(pgcache->page);
374 		}
375 		head++;
376 	}
377 
378 	/* Free RBDR ring */
379 	nicvf_free_q_desc_mem(nic, &rbdr->dmem);
380 }
381 
382 /* Refill receive buffer descriptors with new buffers.
383  */
384 static void nicvf_refill_rbdr(struct nicvf *nic, gfp_t gfp)
385 {
386 	struct queue_set *qs = nic->qs;
387 	int rbdr_idx = qs->rbdr_cnt;
388 	int tail, qcount;
389 	int refill_rb_cnt;
390 	struct rbdr *rbdr;
391 	struct rbdr_entry_t *desc;
392 	u64 rbuf;
393 	int new_rb = 0;
394 
395 refill:
396 	if (!rbdr_idx)
397 		return;
398 	rbdr_idx--;
399 	rbdr = &qs->rbdr[rbdr_idx];
400 	/* Check if it's enabled */
401 	if (!rbdr->enable)
402 		goto next_rbdr;
403 
404 	/* Get no of desc's to be refilled */
405 	qcount = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, rbdr_idx);
406 	qcount &= 0x7FFFF;
407 	/* Doorbell can be ringed with a max of ring size minus 1 */
408 	if (qcount >= (qs->rbdr_len - 1))
409 		goto next_rbdr;
410 	else
411 		refill_rb_cnt = qs->rbdr_len - qcount - 1;
412 
413 	/* Sync page cache info */
414 	smp_rmb();
415 
416 	/* Start filling descs from tail */
417 	tail = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_TAIL, rbdr_idx) >> 3;
418 	while (refill_rb_cnt) {
419 		tail++;
420 		tail &= (rbdr->dmem.q_len - 1);
421 
422 		if (nicvf_alloc_rcv_buffer(nic, rbdr, gfp, RCV_FRAG_LEN, &rbuf))
423 			break;
424 
425 		desc = GET_RBDR_DESC(rbdr, tail);
426 		desc->buf_addr = rbuf & ~(NICVF_RCV_BUF_ALIGN_BYTES - 1);
427 		refill_rb_cnt--;
428 		new_rb++;
429 	}
430 
431 	nicvf_get_page(nic);
432 
433 	/* make sure all memory stores are done before ringing doorbell */
434 	smp_wmb();
435 
436 	/* Check if buffer allocation failed */
437 	if (refill_rb_cnt)
438 		nic->rb_alloc_fail = true;
439 	else
440 		nic->rb_alloc_fail = false;
441 
442 	/* Notify HW */
443 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR,
444 			      rbdr_idx, new_rb);
445 next_rbdr:
446 	/* Re-enable RBDR interrupts only if buffer allocation is success */
447 	if (!nic->rb_alloc_fail && rbdr->enable &&
448 	    netif_running(nic->pnicvf->netdev))
449 		nicvf_enable_intr(nic, NICVF_INTR_RBDR, rbdr_idx);
450 
451 	if (rbdr_idx)
452 		goto refill;
453 }
454 
455 /* Alloc rcv buffers in non-atomic mode for better success */
456 void nicvf_rbdr_work(struct work_struct *work)
457 {
458 	struct nicvf *nic = container_of(work, struct nicvf, rbdr_work.work);
459 
460 	nicvf_refill_rbdr(nic, GFP_KERNEL);
461 	if (nic->rb_alloc_fail)
462 		schedule_delayed_work(&nic->rbdr_work, msecs_to_jiffies(10));
463 	else
464 		nic->rb_work_scheduled = false;
465 }
466 
467 /* In Softirq context, alloc rcv buffers in atomic mode */
468 void nicvf_rbdr_task(unsigned long data)
469 {
470 	struct nicvf *nic = (struct nicvf *)data;
471 
472 	nicvf_refill_rbdr(nic, GFP_ATOMIC);
473 	if (nic->rb_alloc_fail) {
474 		nic->rb_work_scheduled = true;
475 		schedule_delayed_work(&nic->rbdr_work, msecs_to_jiffies(10));
476 	}
477 }
478 
479 /* Initialize completion queue */
480 static int nicvf_init_cmp_queue(struct nicvf *nic,
481 				struct cmp_queue *cq, int q_len)
482 {
483 	int err;
484 
485 	err = nicvf_alloc_q_desc_mem(nic, &cq->dmem, q_len, CMP_QUEUE_DESC_SIZE,
486 				     NICVF_CQ_BASE_ALIGN_BYTES);
487 	if (err)
488 		return err;
489 
490 	cq->desc = cq->dmem.base;
491 	cq->thresh = pass1_silicon(nic->pdev) ? 0 : CMP_QUEUE_CQE_THRESH;
492 	nic->cq_coalesce_usecs = (CMP_QUEUE_TIMER_THRESH * 0.05) - 1;
493 
494 	return 0;
495 }
496 
497 static void nicvf_free_cmp_queue(struct nicvf *nic, struct cmp_queue *cq)
498 {
499 	if (!cq)
500 		return;
501 	if (!cq->dmem.base)
502 		return;
503 
504 	nicvf_free_q_desc_mem(nic, &cq->dmem);
505 }
506 
507 /* Initialize transmit queue */
508 static int nicvf_init_snd_queue(struct nicvf *nic,
509 				struct snd_queue *sq, int q_len, int qidx)
510 {
511 	int err;
512 
513 	err = nicvf_alloc_q_desc_mem(nic, &sq->dmem, q_len, SND_QUEUE_DESC_SIZE,
514 				     NICVF_SQ_BASE_ALIGN_BYTES);
515 	if (err)
516 		return err;
517 
518 	sq->desc = sq->dmem.base;
519 	sq->skbuff = kcalloc(q_len, sizeof(u64), GFP_KERNEL);
520 	if (!sq->skbuff)
521 		return -ENOMEM;
522 
523 	sq->head = 0;
524 	sq->tail = 0;
525 	sq->thresh = SND_QUEUE_THRESH;
526 
527 	/* Check if this SQ is a XDP TX queue */
528 	if (nic->sqs_mode)
529 		qidx += ((nic->sqs_id + 1) * MAX_SND_QUEUES_PER_QS);
530 	if (qidx < nic->pnicvf->xdp_tx_queues) {
531 		/* Alloc memory to save page pointers for XDP_TX */
532 		sq->xdp_page = kcalloc(q_len, sizeof(u64), GFP_KERNEL);
533 		if (!sq->xdp_page)
534 			return -ENOMEM;
535 		sq->xdp_desc_cnt = 0;
536 		sq->xdp_free_cnt = q_len - 1;
537 		sq->is_xdp = true;
538 	} else {
539 		sq->xdp_page = NULL;
540 		sq->xdp_desc_cnt = 0;
541 		sq->xdp_free_cnt = 0;
542 		sq->is_xdp = false;
543 
544 		atomic_set(&sq->free_cnt, q_len - 1);
545 
546 		/* Preallocate memory for TSO segment's header */
547 		sq->tso_hdrs = dma_alloc_coherent(&nic->pdev->dev,
548 						  q_len * TSO_HEADER_SIZE,
549 						  &sq->tso_hdrs_phys,
550 						  GFP_KERNEL);
551 		if (!sq->tso_hdrs)
552 			return -ENOMEM;
553 	}
554 
555 	return 0;
556 }
557 
558 void nicvf_unmap_sndq_buffers(struct nicvf *nic, struct snd_queue *sq,
559 			      int hdr_sqe, u8 subdesc_cnt)
560 {
561 	u8 idx;
562 	struct sq_gather_subdesc *gather;
563 
564 	/* Unmap DMA mapped skb data buffers */
565 	for (idx = 0; idx < subdesc_cnt; idx++) {
566 		hdr_sqe++;
567 		hdr_sqe &= (sq->dmem.q_len - 1);
568 		gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, hdr_sqe);
569 		/* HW will ensure data coherency, CPU sync not required */
570 		dma_unmap_page_attrs(&nic->pdev->dev, gather->addr,
571 				     gather->size, DMA_TO_DEVICE,
572 				     DMA_ATTR_SKIP_CPU_SYNC);
573 	}
574 }
575 
576 static void nicvf_free_snd_queue(struct nicvf *nic, struct snd_queue *sq)
577 {
578 	struct sk_buff *skb;
579 	struct page *page;
580 	struct sq_hdr_subdesc *hdr;
581 	struct sq_hdr_subdesc *tso_sqe;
582 
583 	if (!sq)
584 		return;
585 	if (!sq->dmem.base)
586 		return;
587 
588 	if (sq->tso_hdrs)
589 		dma_free_coherent(&nic->pdev->dev,
590 				  sq->dmem.q_len * TSO_HEADER_SIZE,
591 				  sq->tso_hdrs, sq->tso_hdrs_phys);
592 
593 	/* Free pending skbs in the queue */
594 	smp_rmb();
595 	while (sq->head != sq->tail) {
596 		skb = (struct sk_buff *)sq->skbuff[sq->head];
597 		if (!skb || !sq->xdp_page)
598 			goto next;
599 
600 		page = (struct page *)sq->xdp_page[sq->head];
601 		if (!page)
602 			goto next;
603 		else
604 			put_page(page);
605 
606 		hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, sq->head);
607 		/* Check for dummy descriptor used for HW TSO offload on 88xx */
608 		if (hdr->dont_send) {
609 			/* Get actual TSO descriptors and unmap them */
610 			tso_sqe =
611 			 (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, hdr->rsvd2);
612 			nicvf_unmap_sndq_buffers(nic, sq, hdr->rsvd2,
613 						 tso_sqe->subdesc_cnt);
614 		} else {
615 			nicvf_unmap_sndq_buffers(nic, sq, sq->head,
616 						 hdr->subdesc_cnt);
617 		}
618 		if (skb)
619 			dev_kfree_skb_any(skb);
620 next:
621 		sq->head++;
622 		sq->head &= (sq->dmem.q_len - 1);
623 	}
624 	kfree(sq->skbuff);
625 	kfree(sq->xdp_page);
626 	nicvf_free_q_desc_mem(nic, &sq->dmem);
627 }
628 
629 static void nicvf_reclaim_snd_queue(struct nicvf *nic,
630 				    struct queue_set *qs, int qidx)
631 {
632 	/* Disable send queue */
633 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, 0);
634 	/* Check if SQ is stopped */
635 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_SQ_0_7_STATUS, 21, 1, 0x01))
636 		return;
637 	/* Reset send queue */
638 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET);
639 }
640 
641 static void nicvf_reclaim_rcv_queue(struct nicvf *nic,
642 				    struct queue_set *qs, int qidx)
643 {
644 	union nic_mbx mbx = {};
645 
646 	/* Make sure all packets in the pipeline are written back into mem */
647 	mbx.msg.msg = NIC_MBOX_MSG_RQ_SW_SYNC;
648 	nicvf_send_msg_to_pf(nic, &mbx);
649 }
650 
651 static void nicvf_reclaim_cmp_queue(struct nicvf *nic,
652 				    struct queue_set *qs, int qidx)
653 {
654 	/* Disable timer threshold (doesn't get reset upon CQ reset */
655 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2, qidx, 0);
656 	/* Disable completion queue */
657 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, 0);
658 	/* Reset completion queue */
659 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET);
660 }
661 
662 static void nicvf_reclaim_rbdr(struct nicvf *nic,
663 			       struct rbdr *rbdr, int qidx)
664 {
665 	u64 tmp, fifo_state;
666 	int timeout = 10;
667 
668 	/* Save head and tail pointers for feeing up buffers */
669 	rbdr->head = nicvf_queue_reg_read(nic,
670 					  NIC_QSET_RBDR_0_1_HEAD,
671 					  qidx) >> 3;
672 	rbdr->tail = nicvf_queue_reg_read(nic,
673 					  NIC_QSET_RBDR_0_1_TAIL,
674 					  qidx) >> 3;
675 
676 	/* If RBDR FIFO is in 'FAIL' state then do a reset first
677 	 * before relaiming.
678 	 */
679 	fifo_state = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, qidx);
680 	if (((fifo_state >> 62) & 0x03) == 0x3)
681 		nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
682 				      qidx, NICVF_RBDR_RESET);
683 
684 	/* Disable RBDR */
685 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0);
686 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00))
687 		return;
688 	while (1) {
689 		tmp = nicvf_queue_reg_read(nic,
690 					   NIC_QSET_RBDR_0_1_PREFETCH_STATUS,
691 					   qidx);
692 		if ((tmp & 0xFFFFFFFF) == ((tmp >> 32) & 0xFFFFFFFF))
693 			break;
694 		usleep_range(1000, 2000);
695 		timeout--;
696 		if (!timeout) {
697 			netdev_err(nic->netdev,
698 				   "Failed polling on prefetch status\n");
699 			return;
700 		}
701 	}
702 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
703 			      qidx, NICVF_RBDR_RESET);
704 
705 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x02))
706 		return;
707 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0x00);
708 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00))
709 		return;
710 }
711 
712 void nicvf_config_vlan_stripping(struct nicvf *nic, netdev_features_t features)
713 {
714 	u64 rq_cfg;
715 	int sqs;
716 
717 	rq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_RQ_GEN_CFG, 0);
718 
719 	/* Enable first VLAN stripping */
720 	if (features & NETIF_F_HW_VLAN_CTAG_RX)
721 		rq_cfg |= (1ULL << 25);
722 	else
723 		rq_cfg &= ~(1ULL << 25);
724 	nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0, rq_cfg);
725 
726 	/* Configure Secondary Qsets, if any */
727 	for (sqs = 0; sqs < nic->sqs_count; sqs++)
728 		if (nic->snicvf[sqs])
729 			nicvf_queue_reg_write(nic->snicvf[sqs],
730 					      NIC_QSET_RQ_GEN_CFG, 0, rq_cfg);
731 }
732 
733 static void nicvf_reset_rcv_queue_stats(struct nicvf *nic)
734 {
735 	union nic_mbx mbx = {};
736 
737 	/* Reset all RQ/SQ and VF stats */
738 	mbx.reset_stat.msg = NIC_MBOX_MSG_RESET_STAT_COUNTER;
739 	mbx.reset_stat.rx_stat_mask = 0x3FFF;
740 	mbx.reset_stat.tx_stat_mask = 0x1F;
741 	mbx.reset_stat.rq_stat_mask = 0xFFFF;
742 	mbx.reset_stat.sq_stat_mask = 0xFFFF;
743 	nicvf_send_msg_to_pf(nic, &mbx);
744 }
745 
746 /* Configures receive queue */
747 static void nicvf_rcv_queue_config(struct nicvf *nic, struct queue_set *qs,
748 				   int qidx, bool enable)
749 {
750 	union nic_mbx mbx = {};
751 	struct rcv_queue *rq;
752 	struct rq_cfg rq_cfg;
753 
754 	rq = &qs->rq[qidx];
755 	rq->enable = enable;
756 
757 	/* Disable receive queue */
758 	nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, 0);
759 
760 	if (!rq->enable) {
761 		nicvf_reclaim_rcv_queue(nic, qs, qidx);
762 		xdp_rxq_info_unreg(&rq->xdp_rxq);
763 		return;
764 	}
765 
766 	rq->cq_qs = qs->vnic_id;
767 	rq->cq_idx = qidx;
768 	rq->start_rbdr_qs = qs->vnic_id;
769 	rq->start_qs_rbdr_idx = qs->rbdr_cnt - 1;
770 	rq->cont_rbdr_qs = qs->vnic_id;
771 	rq->cont_qs_rbdr_idx = qs->rbdr_cnt - 1;
772 	/* all writes of RBDR data to be loaded into L2 Cache as well*/
773 	rq->caching = 1;
774 
775 	/* Driver have no proper error path for failed XDP RX-queue info reg */
776 	WARN_ON(xdp_rxq_info_reg(&rq->xdp_rxq, nic->netdev, qidx) < 0);
777 
778 	/* Send a mailbox msg to PF to config RQ */
779 	mbx.rq.msg = NIC_MBOX_MSG_RQ_CFG;
780 	mbx.rq.qs_num = qs->vnic_id;
781 	mbx.rq.rq_num = qidx;
782 	mbx.rq.cfg = (rq->caching << 26) | (rq->cq_qs << 19) |
783 			  (rq->cq_idx << 16) | (rq->cont_rbdr_qs << 9) |
784 			  (rq->cont_qs_rbdr_idx << 8) |
785 			  (rq->start_rbdr_qs << 1) | (rq->start_qs_rbdr_idx);
786 	nicvf_send_msg_to_pf(nic, &mbx);
787 
788 	mbx.rq.msg = NIC_MBOX_MSG_RQ_BP_CFG;
789 	mbx.rq.cfg = BIT_ULL(63) | BIT_ULL(62) |
790 		     (RQ_PASS_RBDR_LVL << 16) | (RQ_PASS_CQ_LVL << 8) |
791 		     (qs->vnic_id << 0);
792 	nicvf_send_msg_to_pf(nic, &mbx);
793 
794 	/* RQ drop config
795 	 * Enable CQ drop to reserve sufficient CQEs for all tx packets
796 	 */
797 	mbx.rq.msg = NIC_MBOX_MSG_RQ_DROP_CFG;
798 	mbx.rq.cfg = BIT_ULL(63) | BIT_ULL(62) |
799 		     (RQ_PASS_RBDR_LVL << 40) | (RQ_DROP_RBDR_LVL << 32) |
800 		     (RQ_PASS_CQ_LVL << 16) | (RQ_DROP_CQ_LVL << 8);
801 	nicvf_send_msg_to_pf(nic, &mbx);
802 
803 	if (!nic->sqs_mode && (qidx == 0)) {
804 		/* Enable checking L3/L4 length and TCP/UDP checksums
805 		 * Also allow IPv6 pkts with zero UDP checksum.
806 		 */
807 		nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0,
808 				      (BIT(24) | BIT(23) | BIT(21) | BIT(20)));
809 		nicvf_config_vlan_stripping(nic, nic->netdev->features);
810 	}
811 
812 	/* Enable Receive queue */
813 	memset(&rq_cfg, 0, sizeof(struct rq_cfg));
814 	rq_cfg.ena = 1;
815 	rq_cfg.tcp_ena = 0;
816 	nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, *(u64 *)&rq_cfg);
817 }
818 
819 /* Configures completion queue */
820 void nicvf_cmp_queue_config(struct nicvf *nic, struct queue_set *qs,
821 			    int qidx, bool enable)
822 {
823 	struct cmp_queue *cq;
824 	struct cq_cfg cq_cfg;
825 
826 	cq = &qs->cq[qidx];
827 	cq->enable = enable;
828 
829 	if (!cq->enable) {
830 		nicvf_reclaim_cmp_queue(nic, qs, qidx);
831 		return;
832 	}
833 
834 	/* Reset completion queue */
835 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET);
836 
837 	if (!cq->enable)
838 		return;
839 
840 	spin_lock_init(&cq->lock);
841 	/* Set completion queue base address */
842 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_BASE,
843 			      qidx, (u64)(cq->dmem.phys_base));
844 
845 	/* Enable Completion queue */
846 	memset(&cq_cfg, 0, sizeof(struct cq_cfg));
847 	cq_cfg.ena = 1;
848 	cq_cfg.reset = 0;
849 	cq_cfg.caching = 0;
850 	cq_cfg.qsize = ilog2(qs->cq_len >> 10);
851 	cq_cfg.avg_con = 0;
852 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, *(u64 *)&cq_cfg);
853 
854 	/* Set threshold value for interrupt generation */
855 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_THRESH, qidx, cq->thresh);
856 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2,
857 			      qidx, CMP_QUEUE_TIMER_THRESH);
858 }
859 
860 /* Configures transmit queue */
861 static void nicvf_snd_queue_config(struct nicvf *nic, struct queue_set *qs,
862 				   int qidx, bool enable)
863 {
864 	union nic_mbx mbx = {};
865 	struct snd_queue *sq;
866 	struct sq_cfg sq_cfg;
867 
868 	sq = &qs->sq[qidx];
869 	sq->enable = enable;
870 
871 	if (!sq->enable) {
872 		nicvf_reclaim_snd_queue(nic, qs, qidx);
873 		return;
874 	}
875 
876 	/* Reset send queue */
877 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET);
878 
879 	sq->cq_qs = qs->vnic_id;
880 	sq->cq_idx = qidx;
881 
882 	/* Send a mailbox msg to PF to config SQ */
883 	mbx.sq.msg = NIC_MBOX_MSG_SQ_CFG;
884 	mbx.sq.qs_num = qs->vnic_id;
885 	mbx.sq.sq_num = qidx;
886 	mbx.sq.sqs_mode = nic->sqs_mode;
887 	mbx.sq.cfg = (sq->cq_qs << 3) | sq->cq_idx;
888 	nicvf_send_msg_to_pf(nic, &mbx);
889 
890 	/* Set queue base address */
891 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_BASE,
892 			      qidx, (u64)(sq->dmem.phys_base));
893 
894 	/* Enable send queue  & set queue size */
895 	memset(&sq_cfg, 0, sizeof(struct sq_cfg));
896 	sq_cfg.ena = 1;
897 	sq_cfg.reset = 0;
898 	sq_cfg.ldwb = 0;
899 	sq_cfg.qsize = ilog2(qs->sq_len >> 10);
900 	sq_cfg.tstmp_bgx_intf = 0;
901 	/* CQ's level at which HW will stop processing SQEs to avoid
902 	 * transmitting a pkt with no space in CQ to post CQE_TX.
903 	 */
904 	sq_cfg.cq_limit = (CMP_QUEUE_PIPELINE_RSVD * 256) / qs->cq_len;
905 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, *(u64 *)&sq_cfg);
906 
907 	/* Set threshold value for interrupt generation */
908 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_THRESH, qidx, sq->thresh);
909 
910 	/* Set queue:cpu affinity for better load distribution */
911 	if (cpu_online(qidx)) {
912 		cpumask_set_cpu(qidx, &sq->affinity_mask);
913 		netif_set_xps_queue(nic->netdev,
914 				    &sq->affinity_mask, qidx);
915 	}
916 }
917 
918 /* Configures receive buffer descriptor ring */
919 static void nicvf_rbdr_config(struct nicvf *nic, struct queue_set *qs,
920 			      int qidx, bool enable)
921 {
922 	struct rbdr *rbdr;
923 	struct rbdr_cfg rbdr_cfg;
924 
925 	rbdr = &qs->rbdr[qidx];
926 	nicvf_reclaim_rbdr(nic, rbdr, qidx);
927 	if (!enable)
928 		return;
929 
930 	/* Set descriptor base address */
931 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_BASE,
932 			      qidx, (u64)(rbdr->dmem.phys_base));
933 
934 	/* Enable RBDR  & set queue size */
935 	/* Buffer size should be in multiples of 128 bytes */
936 	memset(&rbdr_cfg, 0, sizeof(struct rbdr_cfg));
937 	rbdr_cfg.ena = 1;
938 	rbdr_cfg.reset = 0;
939 	rbdr_cfg.ldwb = 0;
940 	rbdr_cfg.qsize = RBDR_SIZE;
941 	rbdr_cfg.avg_con = 0;
942 	rbdr_cfg.lines = rbdr->dma_size / 128;
943 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
944 			      qidx, *(u64 *)&rbdr_cfg);
945 
946 	/* Notify HW */
947 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR,
948 			      qidx, qs->rbdr_len - 1);
949 
950 	/* Set threshold value for interrupt generation */
951 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_THRESH,
952 			      qidx, rbdr->thresh - 1);
953 }
954 
955 /* Requests PF to assign and enable Qset */
956 void nicvf_qset_config(struct nicvf *nic, bool enable)
957 {
958 	union nic_mbx mbx = {};
959 	struct queue_set *qs = nic->qs;
960 	struct qs_cfg *qs_cfg;
961 
962 	if (!qs) {
963 		netdev_warn(nic->netdev,
964 			    "Qset is still not allocated, don't init queues\n");
965 		return;
966 	}
967 
968 	qs->enable = enable;
969 	qs->vnic_id = nic->vf_id;
970 
971 	/* Send a mailbox msg to PF to config Qset */
972 	mbx.qs.msg = NIC_MBOX_MSG_QS_CFG;
973 	mbx.qs.num = qs->vnic_id;
974 	mbx.qs.sqs_count = nic->sqs_count;
975 
976 	mbx.qs.cfg = 0;
977 	qs_cfg = (struct qs_cfg *)&mbx.qs.cfg;
978 	if (qs->enable) {
979 		qs_cfg->ena = 1;
980 #ifdef __BIG_ENDIAN
981 		qs_cfg->be = 1;
982 #endif
983 		qs_cfg->vnic = qs->vnic_id;
984 		/* Enable Tx timestamping capability */
985 		if (nic->ptp_clock)
986 			qs_cfg->send_tstmp_ena = 1;
987 	}
988 	nicvf_send_msg_to_pf(nic, &mbx);
989 }
990 
991 static void nicvf_free_resources(struct nicvf *nic)
992 {
993 	int qidx;
994 	struct queue_set *qs = nic->qs;
995 
996 	/* Free receive buffer descriptor ring */
997 	for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
998 		nicvf_free_rbdr(nic, &qs->rbdr[qidx]);
999 
1000 	/* Free completion queue */
1001 	for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1002 		nicvf_free_cmp_queue(nic, &qs->cq[qidx]);
1003 
1004 	/* Free send queue */
1005 	for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1006 		nicvf_free_snd_queue(nic, &qs->sq[qidx]);
1007 }
1008 
1009 static int nicvf_alloc_resources(struct nicvf *nic)
1010 {
1011 	int qidx;
1012 	struct queue_set *qs = nic->qs;
1013 
1014 	/* Alloc receive buffer descriptor ring */
1015 	for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) {
1016 		if (nicvf_init_rbdr(nic, &qs->rbdr[qidx], qs->rbdr_len,
1017 				    DMA_BUFFER_LEN))
1018 			goto alloc_fail;
1019 	}
1020 
1021 	/* Alloc send queue */
1022 	for (qidx = 0; qidx < qs->sq_cnt; qidx++) {
1023 		if (nicvf_init_snd_queue(nic, &qs->sq[qidx], qs->sq_len, qidx))
1024 			goto alloc_fail;
1025 	}
1026 
1027 	/* Alloc completion queue */
1028 	for (qidx = 0; qidx < qs->cq_cnt; qidx++) {
1029 		if (nicvf_init_cmp_queue(nic, &qs->cq[qidx], qs->cq_len))
1030 			goto alloc_fail;
1031 	}
1032 
1033 	return 0;
1034 alloc_fail:
1035 	nicvf_free_resources(nic);
1036 	return -ENOMEM;
1037 }
1038 
1039 int nicvf_set_qset_resources(struct nicvf *nic)
1040 {
1041 	struct queue_set *qs;
1042 
1043 	qs = devm_kzalloc(&nic->pdev->dev, sizeof(*qs), GFP_KERNEL);
1044 	if (!qs)
1045 		return -ENOMEM;
1046 	nic->qs = qs;
1047 
1048 	/* Set count of each queue */
1049 	qs->rbdr_cnt = DEFAULT_RBDR_CNT;
1050 	qs->rq_cnt = min_t(u8, MAX_RCV_QUEUES_PER_QS, num_online_cpus());
1051 	qs->sq_cnt = min_t(u8, MAX_SND_QUEUES_PER_QS, num_online_cpus());
1052 	qs->cq_cnt = max_t(u8, qs->rq_cnt, qs->sq_cnt);
1053 
1054 	/* Set queue lengths */
1055 	qs->rbdr_len = RCV_BUF_COUNT;
1056 	qs->sq_len = SND_QUEUE_LEN;
1057 	qs->cq_len = CMP_QUEUE_LEN;
1058 
1059 	nic->rx_queues = qs->rq_cnt;
1060 	nic->tx_queues = qs->sq_cnt;
1061 	nic->xdp_tx_queues = 0;
1062 
1063 	return 0;
1064 }
1065 
1066 int nicvf_config_data_transfer(struct nicvf *nic, bool enable)
1067 {
1068 	bool disable = false;
1069 	struct queue_set *qs = nic->qs;
1070 	struct queue_set *pqs = nic->pnicvf->qs;
1071 	int qidx;
1072 
1073 	if (!qs)
1074 		return 0;
1075 
1076 	/* Take primary VF's queue lengths.
1077 	 * This is needed to take queue lengths set from ethtool
1078 	 * into consideration.
1079 	 */
1080 	if (nic->sqs_mode && pqs) {
1081 		qs->cq_len = pqs->cq_len;
1082 		qs->sq_len = pqs->sq_len;
1083 	}
1084 
1085 	if (enable) {
1086 		if (nicvf_alloc_resources(nic))
1087 			return -ENOMEM;
1088 
1089 		for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1090 			nicvf_snd_queue_config(nic, qs, qidx, enable);
1091 		for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1092 			nicvf_cmp_queue_config(nic, qs, qidx, enable);
1093 		for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
1094 			nicvf_rbdr_config(nic, qs, qidx, enable);
1095 		for (qidx = 0; qidx < qs->rq_cnt; qidx++)
1096 			nicvf_rcv_queue_config(nic, qs, qidx, enable);
1097 	} else {
1098 		for (qidx = 0; qidx < qs->rq_cnt; qidx++)
1099 			nicvf_rcv_queue_config(nic, qs, qidx, disable);
1100 		for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
1101 			nicvf_rbdr_config(nic, qs, qidx, disable);
1102 		for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1103 			nicvf_snd_queue_config(nic, qs, qidx, disable);
1104 		for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1105 			nicvf_cmp_queue_config(nic, qs, qidx, disable);
1106 
1107 		nicvf_free_resources(nic);
1108 	}
1109 
1110 	/* Reset RXQ's stats.
1111 	 * SQ's stats will get reset automatically once SQ is reset.
1112 	 */
1113 	nicvf_reset_rcv_queue_stats(nic);
1114 
1115 	return 0;
1116 }
1117 
1118 /* Get a free desc from SQ
1119  * returns descriptor ponter & descriptor number
1120  */
1121 static inline int nicvf_get_sq_desc(struct snd_queue *sq, int desc_cnt)
1122 {
1123 	int qentry;
1124 
1125 	qentry = sq->tail;
1126 	if (!sq->is_xdp)
1127 		atomic_sub(desc_cnt, &sq->free_cnt);
1128 	else
1129 		sq->xdp_free_cnt -= desc_cnt;
1130 	sq->tail += desc_cnt;
1131 	sq->tail &= (sq->dmem.q_len - 1);
1132 
1133 	return qentry;
1134 }
1135 
1136 /* Rollback to previous tail pointer when descriptors not used */
1137 static inline void nicvf_rollback_sq_desc(struct snd_queue *sq,
1138 					  int qentry, int desc_cnt)
1139 {
1140 	sq->tail = qentry;
1141 	atomic_add(desc_cnt, &sq->free_cnt);
1142 }
1143 
1144 /* Free descriptor back to SQ for future use */
1145 void nicvf_put_sq_desc(struct snd_queue *sq, int desc_cnt)
1146 {
1147 	if (!sq->is_xdp)
1148 		atomic_add(desc_cnt, &sq->free_cnt);
1149 	else
1150 		sq->xdp_free_cnt += desc_cnt;
1151 	sq->head += desc_cnt;
1152 	sq->head &= (sq->dmem.q_len - 1);
1153 }
1154 
1155 static inline int nicvf_get_nxt_sqentry(struct snd_queue *sq, int qentry)
1156 {
1157 	qentry++;
1158 	qentry &= (sq->dmem.q_len - 1);
1159 	return qentry;
1160 }
1161 
1162 void nicvf_sq_enable(struct nicvf *nic, struct snd_queue *sq, int qidx)
1163 {
1164 	u64 sq_cfg;
1165 
1166 	sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx);
1167 	sq_cfg |= NICVF_SQ_EN;
1168 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg);
1169 	/* Ring doorbell so that H/W restarts processing SQEs */
1170 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR, qidx, 0);
1171 }
1172 
1173 void nicvf_sq_disable(struct nicvf *nic, int qidx)
1174 {
1175 	u64 sq_cfg;
1176 
1177 	sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx);
1178 	sq_cfg &= ~NICVF_SQ_EN;
1179 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg);
1180 }
1181 
1182 void nicvf_sq_free_used_descs(struct net_device *netdev, struct snd_queue *sq,
1183 			      int qidx)
1184 {
1185 	u64 head, tail;
1186 	struct sk_buff *skb;
1187 	struct nicvf *nic = netdev_priv(netdev);
1188 	struct sq_hdr_subdesc *hdr;
1189 
1190 	head = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_HEAD, qidx) >> 4;
1191 	tail = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_TAIL, qidx) >> 4;
1192 	while (sq->head != head) {
1193 		hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, sq->head);
1194 		if (hdr->subdesc_type != SQ_DESC_TYPE_HEADER) {
1195 			nicvf_put_sq_desc(sq, 1);
1196 			continue;
1197 		}
1198 		skb = (struct sk_buff *)sq->skbuff[sq->head];
1199 		if (skb)
1200 			dev_kfree_skb_any(skb);
1201 		atomic64_add(1, (atomic64_t *)&netdev->stats.tx_packets);
1202 		atomic64_add(hdr->tot_len,
1203 			     (atomic64_t *)&netdev->stats.tx_bytes);
1204 		nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1);
1205 	}
1206 }
1207 
1208 /* XDP Transmit APIs */
1209 void nicvf_xdp_sq_doorbell(struct nicvf *nic,
1210 			   struct snd_queue *sq, int sq_num)
1211 {
1212 	if (!sq->xdp_desc_cnt)
1213 		return;
1214 
1215 	/* make sure all memory stores are done before ringing doorbell */
1216 	wmb();
1217 
1218 	/* Inform HW to xmit all TSO segments */
1219 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR,
1220 			      sq_num, sq->xdp_desc_cnt);
1221 	sq->xdp_desc_cnt = 0;
1222 }
1223 
1224 static inline void
1225 nicvf_xdp_sq_add_hdr_subdesc(struct snd_queue *sq, int qentry,
1226 			     int subdesc_cnt, u64 data, int len)
1227 {
1228 	struct sq_hdr_subdesc *hdr;
1229 
1230 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1231 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1232 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1233 	hdr->subdesc_cnt = subdesc_cnt;
1234 	hdr->tot_len = len;
1235 	hdr->post_cqe = 1;
1236 	sq->xdp_page[qentry] = (u64)virt_to_page((void *)data);
1237 }
1238 
1239 int nicvf_xdp_sq_append_pkt(struct nicvf *nic, struct snd_queue *sq,
1240 			    u64 bufaddr, u64 dma_addr, u16 len)
1241 {
1242 	int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT;
1243 	int qentry;
1244 
1245 	if (subdesc_cnt > sq->xdp_free_cnt)
1246 		return 0;
1247 
1248 	qentry = nicvf_get_sq_desc(sq, subdesc_cnt);
1249 
1250 	nicvf_xdp_sq_add_hdr_subdesc(sq, qentry, subdesc_cnt - 1, bufaddr, len);
1251 
1252 	qentry = nicvf_get_nxt_sqentry(sq, qentry);
1253 	nicvf_sq_add_gather_subdesc(sq, qentry, len, dma_addr);
1254 
1255 	sq->xdp_desc_cnt += subdesc_cnt;
1256 
1257 	return 1;
1258 }
1259 
1260 /* Calculate no of SQ subdescriptors needed to transmit all
1261  * segments of this TSO packet.
1262  * Taken from 'Tilera network driver' with a minor modification.
1263  */
1264 static int nicvf_tso_count_subdescs(struct sk_buff *skb)
1265 {
1266 	struct skb_shared_info *sh = skb_shinfo(skb);
1267 	unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1268 	unsigned int data_len = skb->len - sh_len;
1269 	unsigned int p_len = sh->gso_size;
1270 	long f_id = -1;    /* id of the current fragment */
1271 	long f_size = skb_headlen(skb) - sh_len;  /* current fragment size */
1272 	long f_used = 0;  /* bytes used from the current fragment */
1273 	long n;            /* size of the current piece of payload */
1274 	int num_edescs = 0;
1275 	int segment;
1276 
1277 	for (segment = 0; segment < sh->gso_segs; segment++) {
1278 		unsigned int p_used = 0;
1279 
1280 		/* One edesc for header and for each piece of the payload. */
1281 		for (num_edescs++; p_used < p_len; num_edescs++) {
1282 			/* Advance as needed. */
1283 			while (f_used >= f_size) {
1284 				f_id++;
1285 				f_size = skb_frag_size(&sh->frags[f_id]);
1286 				f_used = 0;
1287 			}
1288 
1289 			/* Use bytes from the current fragment. */
1290 			n = p_len - p_used;
1291 			if (n > f_size - f_used)
1292 				n = f_size - f_used;
1293 			f_used += n;
1294 			p_used += n;
1295 		}
1296 
1297 		/* The last segment may be less than gso_size. */
1298 		data_len -= p_len;
1299 		if (data_len < p_len)
1300 			p_len = data_len;
1301 	}
1302 
1303 	/* '+ gso_segs' for SQ_HDR_SUDESCs for each segment */
1304 	return num_edescs + sh->gso_segs;
1305 }
1306 
1307 #define POST_CQE_DESC_COUNT 2
1308 
1309 /* Get the number of SQ descriptors needed to xmit this skb */
1310 static int nicvf_sq_subdesc_required(struct nicvf *nic, struct sk_buff *skb)
1311 {
1312 	int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT;
1313 
1314 	if (skb_shinfo(skb)->gso_size && !nic->hw_tso) {
1315 		subdesc_cnt = nicvf_tso_count_subdescs(skb);
1316 		return subdesc_cnt;
1317 	}
1318 
1319 	/* Dummy descriptors to get TSO pkt completion notification */
1320 	if (nic->t88 && nic->hw_tso && skb_shinfo(skb)->gso_size)
1321 		subdesc_cnt += POST_CQE_DESC_COUNT;
1322 
1323 	if (skb_shinfo(skb)->nr_frags)
1324 		subdesc_cnt += skb_shinfo(skb)->nr_frags;
1325 
1326 	return subdesc_cnt;
1327 }
1328 
1329 /* Add SQ HEADER subdescriptor.
1330  * First subdescriptor for every send descriptor.
1331  */
1332 static inline void
1333 nicvf_sq_add_hdr_subdesc(struct nicvf *nic, struct snd_queue *sq, int qentry,
1334 			 int subdesc_cnt, struct sk_buff *skb, int len)
1335 {
1336 	int proto;
1337 	struct sq_hdr_subdesc *hdr;
1338 	union {
1339 		struct iphdr *v4;
1340 		struct ipv6hdr *v6;
1341 		unsigned char *hdr;
1342 	} ip;
1343 
1344 	ip.hdr = skb_network_header(skb);
1345 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1346 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1347 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1348 
1349 	if (nic->t88 && nic->hw_tso && skb_shinfo(skb)->gso_size) {
1350 		/* post_cqe = 0, to avoid HW posting a CQE for every TSO
1351 		 * segment transmitted on 88xx.
1352 		 */
1353 		hdr->subdesc_cnt = subdesc_cnt - POST_CQE_DESC_COUNT;
1354 	} else {
1355 		sq->skbuff[qentry] = (u64)skb;
1356 		/* Enable notification via CQE after processing SQE */
1357 		hdr->post_cqe = 1;
1358 		/* No of subdescriptors following this */
1359 		hdr->subdesc_cnt = subdesc_cnt;
1360 	}
1361 	hdr->tot_len = len;
1362 
1363 	/* Offload checksum calculation to HW */
1364 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
1365 		if (ip.v4->version == 4)
1366 			hdr->csum_l3 = 1; /* Enable IP csum calculation */
1367 		hdr->l3_offset = skb_network_offset(skb);
1368 		hdr->l4_offset = skb_transport_offset(skb);
1369 
1370 		proto = (ip.v4->version == 4) ? ip.v4->protocol :
1371 			ip.v6->nexthdr;
1372 
1373 		switch (proto) {
1374 		case IPPROTO_TCP:
1375 			hdr->csum_l4 = SEND_L4_CSUM_TCP;
1376 			break;
1377 		case IPPROTO_UDP:
1378 			hdr->csum_l4 = SEND_L4_CSUM_UDP;
1379 			break;
1380 		case IPPROTO_SCTP:
1381 			hdr->csum_l4 = SEND_L4_CSUM_SCTP;
1382 			break;
1383 		}
1384 	}
1385 
1386 	if (nic->hw_tso && skb_shinfo(skb)->gso_size) {
1387 		hdr->tso = 1;
1388 		hdr->tso_start = skb_transport_offset(skb) + tcp_hdrlen(skb);
1389 		hdr->tso_max_paysize = skb_shinfo(skb)->gso_size;
1390 		/* For non-tunneled pkts, point this to L2 ethertype */
1391 		hdr->inner_l3_offset = skb_network_offset(skb) - 2;
1392 		this_cpu_inc(nic->pnicvf->drv_stats->tx_tso);
1393 	}
1394 
1395 	/* Check if timestamp is requested */
1396 	if (!(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)) {
1397 		skb_tx_timestamp(skb);
1398 		return;
1399 	}
1400 
1401 	/* Tx timestamping not supported along with TSO, so ignore request */
1402 	if (skb_shinfo(skb)->gso_size)
1403 		return;
1404 
1405 	/* HW supports only a single outstanding packet to timestamp */
1406 	if (!atomic_add_unless(&nic->pnicvf->tx_ptp_skbs, 1, 1))
1407 		return;
1408 
1409 	/* Mark the SKB for later reference */
1410 	skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
1411 
1412 	/* Finally enable timestamp generation
1413 	 * Since 'post_cqe' is also set, two CQEs will be posted
1414 	 * for this packet i.e CQE_TYPE_SEND and CQE_TYPE_SEND_PTP.
1415 	 */
1416 	hdr->tstmp = 1;
1417 }
1418 
1419 /* SQ GATHER subdescriptor
1420  * Must follow HDR descriptor
1421  */
1422 static inline void nicvf_sq_add_gather_subdesc(struct snd_queue *sq, int qentry,
1423 					       int size, u64 data)
1424 {
1425 	struct sq_gather_subdesc *gather;
1426 
1427 	qentry &= (sq->dmem.q_len - 1);
1428 	gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, qentry);
1429 
1430 	memset(gather, 0, SND_QUEUE_DESC_SIZE);
1431 	gather->subdesc_type = SQ_DESC_TYPE_GATHER;
1432 	gather->ld_type = NIC_SEND_LD_TYPE_E_LDD;
1433 	gather->size = size;
1434 	gather->addr = data;
1435 }
1436 
1437 /* Add HDR + IMMEDIATE subdescriptors right after descriptors of a TSO
1438  * packet so that a CQE is posted as a notifation for transmission of
1439  * TSO packet.
1440  */
1441 static inline void nicvf_sq_add_cqe_subdesc(struct snd_queue *sq, int qentry,
1442 					    int tso_sqe, struct sk_buff *skb)
1443 {
1444 	struct sq_imm_subdesc *imm;
1445 	struct sq_hdr_subdesc *hdr;
1446 
1447 	sq->skbuff[qentry] = (u64)skb;
1448 
1449 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1450 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1451 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1452 	/* Enable notification via CQE after processing SQE */
1453 	hdr->post_cqe = 1;
1454 	/* There is no packet to transmit here */
1455 	hdr->dont_send = 1;
1456 	hdr->subdesc_cnt = POST_CQE_DESC_COUNT - 1;
1457 	hdr->tot_len = 1;
1458 	/* Actual TSO header SQE index, needed for cleanup */
1459 	hdr->rsvd2 = tso_sqe;
1460 
1461 	qentry = nicvf_get_nxt_sqentry(sq, qentry);
1462 	imm = (struct sq_imm_subdesc *)GET_SQ_DESC(sq, qentry);
1463 	memset(imm, 0, SND_QUEUE_DESC_SIZE);
1464 	imm->subdesc_type = SQ_DESC_TYPE_IMMEDIATE;
1465 	imm->len = 1;
1466 }
1467 
1468 static inline void nicvf_sq_doorbell(struct nicvf *nic, struct sk_buff *skb,
1469 				     int sq_num, int desc_cnt)
1470 {
1471 	struct netdev_queue *txq;
1472 
1473 	txq = netdev_get_tx_queue(nic->pnicvf->netdev,
1474 				  skb_get_queue_mapping(skb));
1475 
1476 	netdev_tx_sent_queue(txq, skb->len);
1477 
1478 	/* make sure all memory stores are done before ringing doorbell */
1479 	smp_wmb();
1480 
1481 	/* Inform HW to xmit all TSO segments */
1482 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR,
1483 			      sq_num, desc_cnt);
1484 }
1485 
1486 /* Segment a TSO packet into 'gso_size' segments and append
1487  * them to SQ for transfer
1488  */
1489 static int nicvf_sq_append_tso(struct nicvf *nic, struct snd_queue *sq,
1490 			       int sq_num, int qentry, struct sk_buff *skb)
1491 {
1492 	struct tso_t tso;
1493 	int seg_subdescs = 0, desc_cnt = 0;
1494 	int seg_len, total_len, data_left;
1495 	int hdr_qentry = qentry;
1496 	int hdr_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1497 
1498 	tso_start(skb, &tso);
1499 	total_len = skb->len - hdr_len;
1500 	while (total_len > 0) {
1501 		char *hdr;
1502 
1503 		/* Save Qentry for adding HDR_SUBDESC at the end */
1504 		hdr_qentry = qentry;
1505 
1506 		data_left = min_t(int, skb_shinfo(skb)->gso_size, total_len);
1507 		total_len -= data_left;
1508 
1509 		/* Add segment's header */
1510 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1511 		hdr = sq->tso_hdrs + qentry * TSO_HEADER_SIZE;
1512 		tso_build_hdr(skb, hdr, &tso, data_left, total_len == 0);
1513 		nicvf_sq_add_gather_subdesc(sq, qentry, hdr_len,
1514 					    sq->tso_hdrs_phys +
1515 					    qentry * TSO_HEADER_SIZE);
1516 		/* HDR_SUDESC + GATHER */
1517 		seg_subdescs = 2;
1518 		seg_len = hdr_len;
1519 
1520 		/* Add segment's payload fragments */
1521 		while (data_left > 0) {
1522 			int size;
1523 
1524 			size = min_t(int, tso.size, data_left);
1525 
1526 			qentry = nicvf_get_nxt_sqentry(sq, qentry);
1527 			nicvf_sq_add_gather_subdesc(sq, qentry, size,
1528 						    virt_to_phys(tso.data));
1529 			seg_subdescs++;
1530 			seg_len += size;
1531 
1532 			data_left -= size;
1533 			tso_build_data(skb, &tso, size);
1534 		}
1535 		nicvf_sq_add_hdr_subdesc(nic, sq, hdr_qentry,
1536 					 seg_subdescs - 1, skb, seg_len);
1537 		sq->skbuff[hdr_qentry] = (u64)NULL;
1538 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1539 
1540 		desc_cnt += seg_subdescs;
1541 	}
1542 	/* Save SKB in the last segment for freeing */
1543 	sq->skbuff[hdr_qentry] = (u64)skb;
1544 
1545 	nicvf_sq_doorbell(nic, skb, sq_num, desc_cnt);
1546 
1547 	this_cpu_inc(nic->pnicvf->drv_stats->tx_tso);
1548 	return 1;
1549 }
1550 
1551 /* Append an skb to a SQ for packet transfer. */
1552 int nicvf_sq_append_skb(struct nicvf *nic, struct snd_queue *sq,
1553 			struct sk_buff *skb, u8 sq_num)
1554 {
1555 	int i, size;
1556 	int subdesc_cnt, hdr_sqe = 0;
1557 	int qentry;
1558 	u64 dma_addr;
1559 
1560 	subdesc_cnt = nicvf_sq_subdesc_required(nic, skb);
1561 	if (subdesc_cnt > atomic_read(&sq->free_cnt))
1562 		goto append_fail;
1563 
1564 	qentry = nicvf_get_sq_desc(sq, subdesc_cnt);
1565 
1566 	/* Check if its a TSO packet */
1567 	if (skb_shinfo(skb)->gso_size && !nic->hw_tso)
1568 		return nicvf_sq_append_tso(nic, sq, sq_num, qentry, skb);
1569 
1570 	/* Add SQ header subdesc */
1571 	nicvf_sq_add_hdr_subdesc(nic, sq, qentry, subdesc_cnt - 1,
1572 				 skb, skb->len);
1573 	hdr_sqe = qentry;
1574 
1575 	/* Add SQ gather subdescs */
1576 	qentry = nicvf_get_nxt_sqentry(sq, qentry);
1577 	size = skb_is_nonlinear(skb) ? skb_headlen(skb) : skb->len;
1578 	/* HW will ensure data coherency, CPU sync not required */
1579 	dma_addr = dma_map_page_attrs(&nic->pdev->dev, virt_to_page(skb->data),
1580 				      offset_in_page(skb->data), size,
1581 				      DMA_TO_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
1582 	if (dma_mapping_error(&nic->pdev->dev, dma_addr)) {
1583 		nicvf_rollback_sq_desc(sq, qentry, subdesc_cnt);
1584 		return 0;
1585 	}
1586 
1587 	nicvf_sq_add_gather_subdesc(sq, qentry, size, dma_addr);
1588 
1589 	/* Check for scattered buffer */
1590 	if (!skb_is_nonlinear(skb))
1591 		goto doorbell;
1592 
1593 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1594 		const struct skb_frag_struct *frag;
1595 
1596 		frag = &skb_shinfo(skb)->frags[i];
1597 
1598 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1599 		size = skb_frag_size(frag);
1600 		dma_addr = dma_map_page_attrs(&nic->pdev->dev,
1601 					      skb_frag_page(frag),
1602 					      frag->page_offset, size,
1603 					      DMA_TO_DEVICE,
1604 					      DMA_ATTR_SKIP_CPU_SYNC);
1605 		if (dma_mapping_error(&nic->pdev->dev, dma_addr)) {
1606 			/* Free entire chain of mapped buffers
1607 			 * here 'i' = frags mapped + above mapped skb->data
1608 			 */
1609 			nicvf_unmap_sndq_buffers(nic, sq, hdr_sqe, i);
1610 			nicvf_rollback_sq_desc(sq, qentry, subdesc_cnt);
1611 			return 0;
1612 		}
1613 		nicvf_sq_add_gather_subdesc(sq, qentry, size, dma_addr);
1614 	}
1615 
1616 doorbell:
1617 	if (nic->t88 && skb_shinfo(skb)->gso_size) {
1618 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1619 		nicvf_sq_add_cqe_subdesc(sq, qentry, hdr_sqe, skb);
1620 	}
1621 
1622 	nicvf_sq_doorbell(nic, skb, sq_num, subdesc_cnt);
1623 
1624 	return 1;
1625 
1626 append_fail:
1627 	/* Use original PCI dev for debug log */
1628 	nic = nic->pnicvf;
1629 	netdev_dbg(nic->netdev, "Not enough SQ descriptors to xmit pkt\n");
1630 	return 0;
1631 }
1632 
1633 static inline unsigned frag_num(unsigned i)
1634 {
1635 #ifdef __BIG_ENDIAN
1636 	return (i & ~3) + 3 - (i & 3);
1637 #else
1638 	return i;
1639 #endif
1640 }
1641 
1642 static void nicvf_unmap_rcv_buffer(struct nicvf *nic, u64 dma_addr,
1643 				   u64 buf_addr, bool xdp)
1644 {
1645 	struct page *page = NULL;
1646 	int len = RCV_FRAG_LEN;
1647 
1648 	if (xdp) {
1649 		page = virt_to_page(phys_to_virt(buf_addr));
1650 		/* Check if it's a recycled page, if not
1651 		 * unmap the DMA mapping.
1652 		 *
1653 		 * Recycled page holds an extra reference.
1654 		 */
1655 		if (page_ref_count(page) != 1)
1656 			return;
1657 
1658 		len += XDP_PACKET_HEADROOM;
1659 		/* Receive buffers in XDP mode are mapped from page start */
1660 		dma_addr &= PAGE_MASK;
1661 	}
1662 	dma_unmap_page_attrs(&nic->pdev->dev, dma_addr, len,
1663 			     DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
1664 }
1665 
1666 /* Returns SKB for a received packet */
1667 struct sk_buff *nicvf_get_rcv_skb(struct nicvf *nic,
1668 				  struct cqe_rx_t *cqe_rx, bool xdp)
1669 {
1670 	int frag;
1671 	int payload_len = 0;
1672 	struct sk_buff *skb = NULL;
1673 	struct page *page;
1674 	int offset;
1675 	u16 *rb_lens = NULL;
1676 	u64 *rb_ptrs = NULL;
1677 	u64 phys_addr;
1678 
1679 	rb_lens = (void *)cqe_rx + (3 * sizeof(u64));
1680 	/* Except 88xx pass1 on all other chips CQE_RX2_S is added to
1681 	 * CQE_RX at word6, hence buffer pointers move by word
1682 	 *
1683 	 * Use existing 'hw_tso' flag which will be set for all chips
1684 	 * except 88xx pass1 instead of a additional cache line
1685 	 * access (or miss) by using pci dev's revision.
1686 	 */
1687 	if (!nic->hw_tso)
1688 		rb_ptrs = (void *)cqe_rx + (6 * sizeof(u64));
1689 	else
1690 		rb_ptrs = (void *)cqe_rx + (7 * sizeof(u64));
1691 
1692 	for (frag = 0; frag < cqe_rx->rb_cnt; frag++) {
1693 		payload_len = rb_lens[frag_num(frag)];
1694 		phys_addr = nicvf_iova_to_phys(nic, *rb_ptrs);
1695 		if (!phys_addr) {
1696 			if (skb)
1697 				dev_kfree_skb_any(skb);
1698 			return NULL;
1699 		}
1700 
1701 		if (!frag) {
1702 			/* First fragment */
1703 			nicvf_unmap_rcv_buffer(nic,
1704 					       *rb_ptrs - cqe_rx->align_pad,
1705 					       phys_addr, xdp);
1706 			skb = nicvf_rb_ptr_to_skb(nic,
1707 						  phys_addr - cqe_rx->align_pad,
1708 						  payload_len);
1709 			if (!skb)
1710 				return NULL;
1711 			skb_reserve(skb, cqe_rx->align_pad);
1712 			skb_put(skb, payload_len);
1713 		} else {
1714 			/* Add fragments */
1715 			nicvf_unmap_rcv_buffer(nic, *rb_ptrs, phys_addr, xdp);
1716 			page = virt_to_page(phys_to_virt(phys_addr));
1717 			offset = phys_to_virt(phys_addr) - page_address(page);
1718 			skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page,
1719 					offset, payload_len, RCV_FRAG_LEN);
1720 		}
1721 		/* Next buffer pointer */
1722 		rb_ptrs++;
1723 	}
1724 	return skb;
1725 }
1726 
1727 static u64 nicvf_int_type_to_mask(int int_type, int q_idx)
1728 {
1729 	u64 reg_val;
1730 
1731 	switch (int_type) {
1732 	case NICVF_INTR_CQ:
1733 		reg_val = ((1ULL << q_idx) << NICVF_INTR_CQ_SHIFT);
1734 		break;
1735 	case NICVF_INTR_SQ:
1736 		reg_val = ((1ULL << q_idx) << NICVF_INTR_SQ_SHIFT);
1737 		break;
1738 	case NICVF_INTR_RBDR:
1739 		reg_val = ((1ULL << q_idx) << NICVF_INTR_RBDR_SHIFT);
1740 		break;
1741 	case NICVF_INTR_PKT_DROP:
1742 		reg_val = (1ULL << NICVF_INTR_PKT_DROP_SHIFT);
1743 		break;
1744 	case NICVF_INTR_TCP_TIMER:
1745 		reg_val = (1ULL << NICVF_INTR_TCP_TIMER_SHIFT);
1746 		break;
1747 	case NICVF_INTR_MBOX:
1748 		reg_val = (1ULL << NICVF_INTR_MBOX_SHIFT);
1749 		break;
1750 	case NICVF_INTR_QS_ERR:
1751 		reg_val = (1ULL << NICVF_INTR_QS_ERR_SHIFT);
1752 		break;
1753 	default:
1754 		reg_val = 0;
1755 	}
1756 
1757 	return reg_val;
1758 }
1759 
1760 /* Enable interrupt */
1761 void nicvf_enable_intr(struct nicvf *nic, int int_type, int q_idx)
1762 {
1763 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1764 
1765 	if (!mask) {
1766 		netdev_dbg(nic->netdev,
1767 			   "Failed to enable interrupt: unknown type\n");
1768 		return;
1769 	}
1770 	nicvf_reg_write(nic, NIC_VF_ENA_W1S,
1771 			nicvf_reg_read(nic, NIC_VF_ENA_W1S) | mask);
1772 }
1773 
1774 /* Disable interrupt */
1775 void nicvf_disable_intr(struct nicvf *nic, int int_type, int q_idx)
1776 {
1777 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1778 
1779 	if (!mask) {
1780 		netdev_dbg(nic->netdev,
1781 			   "Failed to disable interrupt: unknown type\n");
1782 		return;
1783 	}
1784 
1785 	nicvf_reg_write(nic, NIC_VF_ENA_W1C, mask);
1786 }
1787 
1788 /* Clear interrupt */
1789 void nicvf_clear_intr(struct nicvf *nic, int int_type, int q_idx)
1790 {
1791 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1792 
1793 	if (!mask) {
1794 		netdev_dbg(nic->netdev,
1795 			   "Failed to clear interrupt: unknown type\n");
1796 		return;
1797 	}
1798 
1799 	nicvf_reg_write(nic, NIC_VF_INT, mask);
1800 }
1801 
1802 /* Check if interrupt is enabled */
1803 int nicvf_is_intr_enabled(struct nicvf *nic, int int_type, int q_idx)
1804 {
1805 	u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1806 	/* If interrupt type is unknown, we treat it disabled. */
1807 	if (!mask) {
1808 		netdev_dbg(nic->netdev,
1809 			   "Failed to check interrupt enable: unknown type\n");
1810 		return 0;
1811 	}
1812 
1813 	return mask & nicvf_reg_read(nic, NIC_VF_ENA_W1S);
1814 }
1815 
1816 void nicvf_update_rq_stats(struct nicvf *nic, int rq_idx)
1817 {
1818 	struct rcv_queue *rq;
1819 
1820 #define GET_RQ_STATS(reg) \
1821 	nicvf_reg_read(nic, NIC_QSET_RQ_0_7_STAT_0_1 |\
1822 			    (rq_idx << NIC_Q_NUM_SHIFT) | (reg << 3))
1823 
1824 	rq = &nic->qs->rq[rq_idx];
1825 	rq->stats.bytes = GET_RQ_STATS(RQ_SQ_STATS_OCTS);
1826 	rq->stats.pkts = GET_RQ_STATS(RQ_SQ_STATS_PKTS);
1827 }
1828 
1829 void nicvf_update_sq_stats(struct nicvf *nic, int sq_idx)
1830 {
1831 	struct snd_queue *sq;
1832 
1833 #define GET_SQ_STATS(reg) \
1834 	nicvf_reg_read(nic, NIC_QSET_SQ_0_7_STAT_0_1 |\
1835 			    (sq_idx << NIC_Q_NUM_SHIFT) | (reg << 3))
1836 
1837 	sq = &nic->qs->sq[sq_idx];
1838 	sq->stats.bytes = GET_SQ_STATS(RQ_SQ_STATS_OCTS);
1839 	sq->stats.pkts = GET_SQ_STATS(RQ_SQ_STATS_PKTS);
1840 }
1841 
1842 /* Check for errors in the receive cmp.queue entry */
1843 int nicvf_check_cqe_rx_errs(struct nicvf *nic, struct cqe_rx_t *cqe_rx)
1844 {
1845 	netif_err(nic, rx_err, nic->netdev,
1846 		  "RX error CQE err_level 0x%x err_opcode 0x%x\n",
1847 		  cqe_rx->err_level, cqe_rx->err_opcode);
1848 
1849 	switch (cqe_rx->err_opcode) {
1850 	case CQ_RX_ERROP_RE_PARTIAL:
1851 		this_cpu_inc(nic->drv_stats->rx_bgx_truncated_pkts);
1852 		break;
1853 	case CQ_RX_ERROP_RE_JABBER:
1854 		this_cpu_inc(nic->drv_stats->rx_jabber_errs);
1855 		break;
1856 	case CQ_RX_ERROP_RE_FCS:
1857 		this_cpu_inc(nic->drv_stats->rx_fcs_errs);
1858 		break;
1859 	case CQ_RX_ERROP_RE_RX_CTL:
1860 		this_cpu_inc(nic->drv_stats->rx_bgx_errs);
1861 		break;
1862 	case CQ_RX_ERROP_PREL2_ERR:
1863 		this_cpu_inc(nic->drv_stats->rx_prel2_errs);
1864 		break;
1865 	case CQ_RX_ERROP_L2_MAL:
1866 		this_cpu_inc(nic->drv_stats->rx_l2_hdr_malformed);
1867 		break;
1868 	case CQ_RX_ERROP_L2_OVERSIZE:
1869 		this_cpu_inc(nic->drv_stats->rx_oversize);
1870 		break;
1871 	case CQ_RX_ERROP_L2_UNDERSIZE:
1872 		this_cpu_inc(nic->drv_stats->rx_undersize);
1873 		break;
1874 	case CQ_RX_ERROP_L2_LENMISM:
1875 		this_cpu_inc(nic->drv_stats->rx_l2_len_mismatch);
1876 		break;
1877 	case CQ_RX_ERROP_L2_PCLP:
1878 		this_cpu_inc(nic->drv_stats->rx_l2_pclp);
1879 		break;
1880 	case CQ_RX_ERROP_IP_NOT:
1881 		this_cpu_inc(nic->drv_stats->rx_ip_ver_errs);
1882 		break;
1883 	case CQ_RX_ERROP_IP_CSUM_ERR:
1884 		this_cpu_inc(nic->drv_stats->rx_ip_csum_errs);
1885 		break;
1886 	case CQ_RX_ERROP_IP_MAL:
1887 		this_cpu_inc(nic->drv_stats->rx_ip_hdr_malformed);
1888 		break;
1889 	case CQ_RX_ERROP_IP_MALD:
1890 		this_cpu_inc(nic->drv_stats->rx_ip_payload_malformed);
1891 		break;
1892 	case CQ_RX_ERROP_IP_HOP:
1893 		this_cpu_inc(nic->drv_stats->rx_ip_ttl_errs);
1894 		break;
1895 	case CQ_RX_ERROP_L3_PCLP:
1896 		this_cpu_inc(nic->drv_stats->rx_l3_pclp);
1897 		break;
1898 	case CQ_RX_ERROP_L4_MAL:
1899 		this_cpu_inc(nic->drv_stats->rx_l4_malformed);
1900 		break;
1901 	case CQ_RX_ERROP_L4_CHK:
1902 		this_cpu_inc(nic->drv_stats->rx_l4_csum_errs);
1903 		break;
1904 	case CQ_RX_ERROP_UDP_LEN:
1905 		this_cpu_inc(nic->drv_stats->rx_udp_len_errs);
1906 		break;
1907 	case CQ_RX_ERROP_L4_PORT:
1908 		this_cpu_inc(nic->drv_stats->rx_l4_port_errs);
1909 		break;
1910 	case CQ_RX_ERROP_TCP_FLAG:
1911 		this_cpu_inc(nic->drv_stats->rx_tcp_flag_errs);
1912 		break;
1913 	case CQ_RX_ERROP_TCP_OFFSET:
1914 		this_cpu_inc(nic->drv_stats->rx_tcp_offset_errs);
1915 		break;
1916 	case CQ_RX_ERROP_L4_PCLP:
1917 		this_cpu_inc(nic->drv_stats->rx_l4_pclp);
1918 		break;
1919 	case CQ_RX_ERROP_RBDR_TRUNC:
1920 		this_cpu_inc(nic->drv_stats->rx_truncated_pkts);
1921 		break;
1922 	}
1923 
1924 	return 1;
1925 }
1926 
1927 /* Check for errors in the send cmp.queue entry */
1928 int nicvf_check_cqe_tx_errs(struct nicvf *nic, struct cqe_send_t *cqe_tx)
1929 {
1930 	switch (cqe_tx->send_status) {
1931 	case CQ_TX_ERROP_DESC_FAULT:
1932 		this_cpu_inc(nic->drv_stats->tx_desc_fault);
1933 		break;
1934 	case CQ_TX_ERROP_HDR_CONS_ERR:
1935 		this_cpu_inc(nic->drv_stats->tx_hdr_cons_err);
1936 		break;
1937 	case CQ_TX_ERROP_SUBDC_ERR:
1938 		this_cpu_inc(nic->drv_stats->tx_subdesc_err);
1939 		break;
1940 	case CQ_TX_ERROP_MAX_SIZE_VIOL:
1941 		this_cpu_inc(nic->drv_stats->tx_max_size_exceeded);
1942 		break;
1943 	case CQ_TX_ERROP_IMM_SIZE_OFLOW:
1944 		this_cpu_inc(nic->drv_stats->tx_imm_size_oflow);
1945 		break;
1946 	case CQ_TX_ERROP_DATA_SEQUENCE_ERR:
1947 		this_cpu_inc(nic->drv_stats->tx_data_seq_err);
1948 		break;
1949 	case CQ_TX_ERROP_MEM_SEQUENCE_ERR:
1950 		this_cpu_inc(nic->drv_stats->tx_mem_seq_err);
1951 		break;
1952 	case CQ_TX_ERROP_LOCK_VIOL:
1953 		this_cpu_inc(nic->drv_stats->tx_lock_viol);
1954 		break;
1955 	case CQ_TX_ERROP_DATA_FAULT:
1956 		this_cpu_inc(nic->drv_stats->tx_data_fault);
1957 		break;
1958 	case CQ_TX_ERROP_TSTMP_CONFLICT:
1959 		this_cpu_inc(nic->drv_stats->tx_tstmp_conflict);
1960 		break;
1961 	case CQ_TX_ERROP_TSTMP_TIMEOUT:
1962 		this_cpu_inc(nic->drv_stats->tx_tstmp_timeout);
1963 		break;
1964 	case CQ_TX_ERROP_MEM_FAULT:
1965 		this_cpu_inc(nic->drv_stats->tx_mem_fault);
1966 		break;
1967 	case CQ_TX_ERROP_CK_OVERLAP:
1968 		this_cpu_inc(nic->drv_stats->tx_csum_overlap);
1969 		break;
1970 	case CQ_TX_ERROP_CK_OFLOW:
1971 		this_cpu_inc(nic->drv_stats->tx_csum_overflow);
1972 		break;
1973 	}
1974 
1975 	return 1;
1976 }
1977