1 // SPDX-License-Identifier: (GPL-2.0 OR MIT)
2 /* Google virtual Ethernet (gve) driver
3  *
4  * Copyright (C) 2015-2021 Google, Inc.
5  */
6 
7 #include "gve.h"
8 #include "gve_adminq.h"
9 #include "gve_utils.h"
10 #include <linux/etherdevice.h>
11 
12 static void gve_rx_free_buffer(struct device *dev,
13 			       struct gve_rx_slot_page_info *page_info,
14 			       union gve_rx_data_slot *data_slot)
15 {
16 	dma_addr_t dma = (dma_addr_t)(be64_to_cpu(data_slot->addr) &
17 				      GVE_DATA_SLOT_ADDR_PAGE_MASK);
18 
19 	page_ref_sub(page_info->page, page_info->pagecnt_bias - 1);
20 	gve_free_page(dev, page_info->page, dma, DMA_FROM_DEVICE);
21 }
22 
23 static void gve_rx_unfill_pages(struct gve_priv *priv, struct gve_rx_ring *rx)
24 {
25 	u32 slots = rx->mask + 1;
26 	int i;
27 
28 	if (rx->data.raw_addressing) {
29 		for (i = 0; i < slots; i++)
30 			gve_rx_free_buffer(&priv->pdev->dev, &rx->data.page_info[i],
31 					   &rx->data.data_ring[i]);
32 	} else {
33 		for (i = 0; i < slots; i++)
34 			page_ref_sub(rx->data.page_info[i].page,
35 				     rx->data.page_info[i].pagecnt_bias - 1);
36 		gve_unassign_qpl(priv, rx->data.qpl->id);
37 		rx->data.qpl = NULL;
38 
39 		for (i = 0; i < rx->qpl_copy_pool_mask + 1; i++) {
40 			page_ref_sub(rx->qpl_copy_pool[i].page,
41 				     rx->qpl_copy_pool[i].pagecnt_bias - 1);
42 			put_page(rx->qpl_copy_pool[i].page);
43 		}
44 	}
45 	kvfree(rx->data.page_info);
46 	rx->data.page_info = NULL;
47 }
48 
49 static void gve_rx_free_ring(struct gve_priv *priv, int idx)
50 {
51 	struct gve_rx_ring *rx = &priv->rx[idx];
52 	struct device *dev = &priv->pdev->dev;
53 	u32 slots = rx->mask + 1;
54 	size_t bytes;
55 
56 	gve_rx_remove_from_block(priv, idx);
57 
58 	bytes = sizeof(struct gve_rx_desc) * priv->rx_desc_cnt;
59 	dma_free_coherent(dev, bytes, rx->desc.desc_ring, rx->desc.bus);
60 	rx->desc.desc_ring = NULL;
61 
62 	dma_free_coherent(dev, sizeof(*rx->q_resources),
63 			  rx->q_resources, rx->q_resources_bus);
64 	rx->q_resources = NULL;
65 
66 	gve_rx_unfill_pages(priv, rx);
67 
68 	bytes = sizeof(*rx->data.data_ring) * slots;
69 	dma_free_coherent(dev, bytes, rx->data.data_ring,
70 			  rx->data.data_bus);
71 	rx->data.data_ring = NULL;
72 
73 	kvfree(rx->qpl_copy_pool);
74 	rx->qpl_copy_pool = NULL;
75 
76 	netif_dbg(priv, drv, priv->dev, "freed rx ring %d\n", idx);
77 }
78 
79 static void gve_setup_rx_buffer(struct gve_rx_slot_page_info *page_info,
80 			     dma_addr_t addr, struct page *page, __be64 *slot_addr)
81 {
82 	page_info->page = page;
83 	page_info->page_offset = 0;
84 	page_info->page_address = page_address(page);
85 	*slot_addr = cpu_to_be64(addr);
86 	/* The page already has 1 ref */
87 	page_ref_add(page, INT_MAX - 1);
88 	page_info->pagecnt_bias = INT_MAX;
89 }
90 
91 static int gve_rx_alloc_buffer(struct gve_priv *priv, struct device *dev,
92 			       struct gve_rx_slot_page_info *page_info,
93 			       union gve_rx_data_slot *data_slot)
94 {
95 	struct page *page;
96 	dma_addr_t dma;
97 	int err;
98 
99 	err = gve_alloc_page(priv, dev, &page, &dma, DMA_FROM_DEVICE,
100 			     GFP_ATOMIC);
101 	if (err)
102 		return err;
103 
104 	gve_setup_rx_buffer(page_info, dma, page, &data_slot->addr);
105 	return 0;
106 }
107 
108 static int gve_prefill_rx_pages(struct gve_rx_ring *rx)
109 {
110 	struct gve_priv *priv = rx->gve;
111 	u32 slots;
112 	int err;
113 	int i;
114 	int j;
115 
116 	/* Allocate one page per Rx queue slot. Each page is split into two
117 	 * packet buffers, when possible we "page flip" between the two.
118 	 */
119 	slots = rx->mask + 1;
120 
121 	rx->data.page_info = kvzalloc(slots *
122 				      sizeof(*rx->data.page_info), GFP_KERNEL);
123 	if (!rx->data.page_info)
124 		return -ENOMEM;
125 
126 	if (!rx->data.raw_addressing) {
127 		rx->data.qpl = gve_assign_rx_qpl(priv);
128 		if (!rx->data.qpl) {
129 			kvfree(rx->data.page_info);
130 			rx->data.page_info = NULL;
131 			return -ENOMEM;
132 		}
133 	}
134 	for (i = 0; i < slots; i++) {
135 		if (!rx->data.raw_addressing) {
136 			struct page *page = rx->data.qpl->pages[i];
137 			dma_addr_t addr = i * PAGE_SIZE;
138 
139 			gve_setup_rx_buffer(&rx->data.page_info[i], addr, page,
140 					    &rx->data.data_ring[i].qpl_offset);
141 			continue;
142 		}
143 		err = gve_rx_alloc_buffer(priv, &priv->pdev->dev, &rx->data.page_info[i],
144 					  &rx->data.data_ring[i]);
145 		if (err)
146 			goto alloc_err;
147 	}
148 
149 	if (!rx->data.raw_addressing) {
150 		for (j = 0; j < rx->qpl_copy_pool_mask + 1; j++) {
151 			struct page *page = alloc_page(GFP_KERNEL);
152 
153 			if (!page) {
154 				err = -ENOMEM;
155 				goto alloc_err_qpl;
156 			}
157 
158 			rx->qpl_copy_pool[j].page = page;
159 			rx->qpl_copy_pool[j].page_offset = 0;
160 			rx->qpl_copy_pool[j].page_address = page_address(page);
161 
162 			/* The page already has 1 ref. */
163 			page_ref_add(page, INT_MAX - 1);
164 			rx->qpl_copy_pool[j].pagecnt_bias = INT_MAX;
165 		}
166 	}
167 
168 	return slots;
169 
170 alloc_err_qpl:
171 	while (j--) {
172 		page_ref_sub(rx->qpl_copy_pool[j].page,
173 			     rx->qpl_copy_pool[j].pagecnt_bias - 1);
174 		put_page(rx->qpl_copy_pool[j].page);
175 	}
176 alloc_err:
177 	while (i--)
178 		gve_rx_free_buffer(&priv->pdev->dev,
179 				   &rx->data.page_info[i],
180 				   &rx->data.data_ring[i]);
181 	return err;
182 }
183 
184 static void gve_rx_ctx_clear(struct gve_rx_ctx *ctx)
185 {
186 	ctx->skb_head = NULL;
187 	ctx->skb_tail = NULL;
188 	ctx->total_size = 0;
189 	ctx->frag_cnt = 0;
190 	ctx->drop_pkt = false;
191 }
192 
193 static int gve_rx_alloc_ring(struct gve_priv *priv, int idx)
194 {
195 	struct gve_rx_ring *rx = &priv->rx[idx];
196 	struct device *hdev = &priv->pdev->dev;
197 	u32 slots, npages;
198 	int filled_pages;
199 	size_t bytes;
200 	int err;
201 
202 	netif_dbg(priv, drv, priv->dev, "allocating rx ring\n");
203 	/* Make sure everything is zeroed to start with */
204 	memset(rx, 0, sizeof(*rx));
205 
206 	rx->gve = priv;
207 	rx->q_num = idx;
208 
209 	slots = priv->rx_data_slot_cnt;
210 	rx->mask = slots - 1;
211 	rx->data.raw_addressing = priv->queue_format == GVE_GQI_RDA_FORMAT;
212 
213 	/* alloc rx data ring */
214 	bytes = sizeof(*rx->data.data_ring) * slots;
215 	rx->data.data_ring = dma_alloc_coherent(hdev, bytes,
216 						&rx->data.data_bus,
217 						GFP_KERNEL);
218 	if (!rx->data.data_ring)
219 		return -ENOMEM;
220 
221 	rx->qpl_copy_pool_mask = min_t(u32, U32_MAX, slots * 2) - 1;
222 	rx->qpl_copy_pool_head = 0;
223 	rx->qpl_copy_pool = kvcalloc(rx->qpl_copy_pool_mask + 1,
224 				     sizeof(rx->qpl_copy_pool[0]),
225 				     GFP_KERNEL);
226 
227 	if (!rx->qpl_copy_pool) {
228 		err = -ENOMEM;
229 		goto abort_with_slots;
230 	}
231 
232 	filled_pages = gve_prefill_rx_pages(rx);
233 	if (filled_pages < 0) {
234 		err = -ENOMEM;
235 		goto abort_with_copy_pool;
236 	}
237 	rx->fill_cnt = filled_pages;
238 	/* Ensure data ring slots (packet buffers) are visible. */
239 	dma_wmb();
240 
241 	/* Alloc gve_queue_resources */
242 	rx->q_resources =
243 		dma_alloc_coherent(hdev,
244 				   sizeof(*rx->q_resources),
245 				   &rx->q_resources_bus,
246 				   GFP_KERNEL);
247 	if (!rx->q_resources) {
248 		err = -ENOMEM;
249 		goto abort_filled;
250 	}
251 	netif_dbg(priv, drv, priv->dev, "rx[%d]->data.data_bus=%lx\n", idx,
252 		  (unsigned long)rx->data.data_bus);
253 
254 	/* alloc rx desc ring */
255 	bytes = sizeof(struct gve_rx_desc) * priv->rx_desc_cnt;
256 	npages = bytes / PAGE_SIZE;
257 	if (npages * PAGE_SIZE != bytes) {
258 		err = -EIO;
259 		goto abort_with_q_resources;
260 	}
261 
262 	rx->desc.desc_ring = dma_alloc_coherent(hdev, bytes, &rx->desc.bus,
263 						GFP_KERNEL);
264 	if (!rx->desc.desc_ring) {
265 		err = -ENOMEM;
266 		goto abort_with_q_resources;
267 	}
268 	rx->cnt = 0;
269 	rx->db_threshold = priv->rx_desc_cnt / 2;
270 	rx->desc.seqno = 1;
271 
272 	/* Allocating half-page buffers allows page-flipping which is faster
273 	 * than copying or allocating new pages.
274 	 */
275 	rx->packet_buffer_size = PAGE_SIZE / 2;
276 	gve_rx_ctx_clear(&rx->ctx);
277 	gve_rx_add_to_block(priv, idx);
278 
279 	return 0;
280 
281 abort_with_q_resources:
282 	dma_free_coherent(hdev, sizeof(*rx->q_resources),
283 			  rx->q_resources, rx->q_resources_bus);
284 	rx->q_resources = NULL;
285 abort_filled:
286 	gve_rx_unfill_pages(priv, rx);
287 abort_with_copy_pool:
288 	kvfree(rx->qpl_copy_pool);
289 	rx->qpl_copy_pool = NULL;
290 abort_with_slots:
291 	bytes = sizeof(*rx->data.data_ring) * slots;
292 	dma_free_coherent(hdev, bytes, rx->data.data_ring, rx->data.data_bus);
293 	rx->data.data_ring = NULL;
294 
295 	return err;
296 }
297 
298 int gve_rx_alloc_rings(struct gve_priv *priv)
299 {
300 	int err = 0;
301 	int i;
302 
303 	for (i = 0; i < priv->rx_cfg.num_queues; i++) {
304 		err = gve_rx_alloc_ring(priv, i);
305 		if (err) {
306 			netif_err(priv, drv, priv->dev,
307 				  "Failed to alloc rx ring=%d: err=%d\n",
308 				  i, err);
309 			break;
310 		}
311 	}
312 	/* Unallocate if there was an error */
313 	if (err) {
314 		int j;
315 
316 		for (j = 0; j < i; j++)
317 			gve_rx_free_ring(priv, j);
318 	}
319 	return err;
320 }
321 
322 void gve_rx_free_rings_gqi(struct gve_priv *priv)
323 {
324 	int i;
325 
326 	for (i = 0; i < priv->rx_cfg.num_queues; i++)
327 		gve_rx_free_ring(priv, i);
328 }
329 
330 void gve_rx_write_doorbell(struct gve_priv *priv, struct gve_rx_ring *rx)
331 {
332 	u32 db_idx = be32_to_cpu(rx->q_resources->db_index);
333 
334 	iowrite32be(rx->fill_cnt, &priv->db_bar2[db_idx]);
335 }
336 
337 static enum pkt_hash_types gve_rss_type(__be16 pkt_flags)
338 {
339 	if (likely(pkt_flags & (GVE_RXF_TCP | GVE_RXF_UDP)))
340 		return PKT_HASH_TYPE_L4;
341 	if (pkt_flags & (GVE_RXF_IPV4 | GVE_RXF_IPV6))
342 		return PKT_HASH_TYPE_L3;
343 	return PKT_HASH_TYPE_L2;
344 }
345 
346 static struct sk_buff *gve_rx_add_frags(struct napi_struct *napi,
347 					struct gve_rx_slot_page_info *page_info,
348 					u16 packet_buffer_size, u16 len,
349 					struct gve_rx_ctx *ctx)
350 {
351 	u32 offset = page_info->page_offset + page_info->pad;
352 	struct sk_buff *skb = ctx->skb_tail;
353 	int num_frags = 0;
354 
355 	if (!skb) {
356 		skb = napi_get_frags(napi);
357 		if (unlikely(!skb))
358 			return NULL;
359 
360 		ctx->skb_head = skb;
361 		ctx->skb_tail = skb;
362 	} else {
363 		num_frags = skb_shinfo(ctx->skb_tail)->nr_frags;
364 		if (num_frags == MAX_SKB_FRAGS) {
365 			skb = napi_alloc_skb(napi, 0);
366 			if (!skb)
367 				return NULL;
368 
369 			// We will never chain more than two SKBs: 2 * 16 * 2k > 64k
370 			// which is why we do not need to chain by using skb->next
371 			skb_shinfo(ctx->skb_tail)->frag_list = skb;
372 
373 			ctx->skb_tail = skb;
374 			num_frags = 0;
375 		}
376 	}
377 
378 	if (skb != ctx->skb_head) {
379 		ctx->skb_head->len += len;
380 		ctx->skb_head->data_len += len;
381 		ctx->skb_head->truesize += packet_buffer_size;
382 	}
383 	skb_add_rx_frag(skb, num_frags, page_info->page,
384 			offset, len, packet_buffer_size);
385 
386 	return ctx->skb_head;
387 }
388 
389 static void gve_rx_flip_buff(struct gve_rx_slot_page_info *page_info, __be64 *slot_addr)
390 {
391 	const __be64 offset = cpu_to_be64(PAGE_SIZE / 2);
392 
393 	/* "flip" to other packet buffer on this page */
394 	page_info->page_offset ^= PAGE_SIZE / 2;
395 	*(slot_addr) ^= offset;
396 }
397 
398 static int gve_rx_can_recycle_buffer(struct gve_rx_slot_page_info *page_info)
399 {
400 	int pagecount = page_count(page_info->page);
401 
402 	/* This page is not being used by any SKBs - reuse */
403 	if (pagecount == page_info->pagecnt_bias)
404 		return 1;
405 	/* This page is still being used by an SKB - we can't reuse */
406 	else if (pagecount > page_info->pagecnt_bias)
407 		return 0;
408 	WARN(pagecount < page_info->pagecnt_bias,
409 	     "Pagecount should never be less than the bias.");
410 	return -1;
411 }
412 
413 static struct sk_buff *
414 gve_rx_raw_addressing(struct device *dev, struct net_device *netdev,
415 		      struct gve_rx_slot_page_info *page_info, u16 len,
416 		      struct napi_struct *napi,
417 		      union gve_rx_data_slot *data_slot,
418 		      u16 packet_buffer_size, struct gve_rx_ctx *ctx)
419 {
420 	struct sk_buff *skb = gve_rx_add_frags(napi, page_info, packet_buffer_size, len, ctx);
421 
422 	if (!skb)
423 		return NULL;
424 
425 	/* Optimistically stop the kernel from freeing the page.
426 	 * We will check again in refill to determine if we need to alloc a
427 	 * new page.
428 	 */
429 	gve_dec_pagecnt_bias(page_info);
430 
431 	return skb;
432 }
433 
434 static struct sk_buff *gve_rx_copy_to_pool(struct gve_rx_ring *rx,
435 					   struct gve_rx_slot_page_info *page_info,
436 					   u16 len, struct napi_struct *napi)
437 {
438 	u32 pool_idx = rx->qpl_copy_pool_head & rx->qpl_copy_pool_mask;
439 	void *src = page_info->page_address + page_info->page_offset;
440 	struct gve_rx_slot_page_info *copy_page_info;
441 	struct gve_rx_ctx *ctx = &rx->ctx;
442 	bool alloc_page = false;
443 	struct sk_buff *skb;
444 	void *dst;
445 
446 	copy_page_info = &rx->qpl_copy_pool[pool_idx];
447 	if (!copy_page_info->can_flip) {
448 		int recycle = gve_rx_can_recycle_buffer(copy_page_info);
449 
450 		if (unlikely(recycle < 0)) {
451 			gve_schedule_reset(rx->gve);
452 			return NULL;
453 		}
454 		alloc_page = !recycle;
455 	}
456 
457 	if (alloc_page) {
458 		struct gve_rx_slot_page_info alloc_page_info;
459 		struct page *page;
460 
461 		/* The least recently used page turned out to be
462 		 * still in use by the kernel. Ignoring it and moving
463 		 * on alleviates head-of-line blocking.
464 		 */
465 		rx->qpl_copy_pool_head++;
466 
467 		page = alloc_page(GFP_ATOMIC);
468 		if (!page)
469 			return NULL;
470 
471 		alloc_page_info.page = page;
472 		alloc_page_info.page_offset = 0;
473 		alloc_page_info.page_address = page_address(page);
474 		alloc_page_info.pad = page_info->pad;
475 
476 		memcpy(alloc_page_info.page_address, src, page_info->pad + len);
477 		skb = gve_rx_add_frags(napi, &alloc_page_info,
478 				       rx->packet_buffer_size,
479 				       len, ctx);
480 
481 		u64_stats_update_begin(&rx->statss);
482 		rx->rx_frag_copy_cnt++;
483 		rx->rx_frag_alloc_cnt++;
484 		u64_stats_update_end(&rx->statss);
485 
486 		return skb;
487 	}
488 
489 	dst = copy_page_info->page_address + copy_page_info->page_offset;
490 	memcpy(dst, src, page_info->pad + len);
491 	copy_page_info->pad = page_info->pad;
492 
493 	skb = gve_rx_add_frags(napi, copy_page_info,
494 			       rx->packet_buffer_size, len, ctx);
495 	if (unlikely(!skb))
496 		return NULL;
497 
498 	gve_dec_pagecnt_bias(copy_page_info);
499 	copy_page_info->page_offset += rx->packet_buffer_size;
500 	copy_page_info->page_offset &= (PAGE_SIZE - 1);
501 
502 	if (copy_page_info->can_flip) {
503 		/* We have used both halves of this copy page, it
504 		 * is time for it to go to the back of the queue.
505 		 */
506 		copy_page_info->can_flip = false;
507 		rx->qpl_copy_pool_head++;
508 		prefetch(rx->qpl_copy_pool[rx->qpl_copy_pool_head & rx->qpl_copy_pool_mask].page);
509 	} else {
510 		copy_page_info->can_flip = true;
511 	}
512 
513 	u64_stats_update_begin(&rx->statss);
514 	rx->rx_frag_copy_cnt++;
515 	u64_stats_update_end(&rx->statss);
516 
517 	return skb;
518 }
519 
520 static struct sk_buff *
521 gve_rx_qpl(struct device *dev, struct net_device *netdev,
522 	   struct gve_rx_ring *rx, struct gve_rx_slot_page_info *page_info,
523 	   u16 len, struct napi_struct *napi,
524 	   union gve_rx_data_slot *data_slot)
525 {
526 	struct gve_rx_ctx *ctx = &rx->ctx;
527 	struct sk_buff *skb;
528 
529 	/* if raw_addressing mode is not enabled gvnic can only receive into
530 	 * registered segments. If the buffer can't be recycled, our only
531 	 * choice is to copy the data out of it so that we can return it to the
532 	 * device.
533 	 */
534 	if (page_info->can_flip) {
535 		skb = gve_rx_add_frags(napi, page_info, rx->packet_buffer_size, len, ctx);
536 		/* No point in recycling if we didn't get the skb */
537 		if (skb) {
538 			/* Make sure that the page isn't freed. */
539 			gve_dec_pagecnt_bias(page_info);
540 			gve_rx_flip_buff(page_info, &data_slot->qpl_offset);
541 		}
542 	} else {
543 		skb = gve_rx_copy_to_pool(rx, page_info, len, napi);
544 	}
545 	return skb;
546 }
547 
548 static struct sk_buff *gve_rx_skb(struct gve_priv *priv, struct gve_rx_ring *rx,
549 				  struct gve_rx_slot_page_info *page_info, struct napi_struct *napi,
550 				  u16 len, union gve_rx_data_slot *data_slot,
551 				  bool is_only_frag)
552 {
553 	struct net_device *netdev = priv->dev;
554 	struct gve_rx_ctx *ctx = &rx->ctx;
555 	struct sk_buff *skb = NULL;
556 
557 	if (len <= priv->rx_copybreak && is_only_frag)  {
558 		/* Just copy small packets */
559 		skb = gve_rx_copy(netdev, napi, page_info, len, GVE_RX_PAD);
560 		if (skb) {
561 			u64_stats_update_begin(&rx->statss);
562 			rx->rx_copied_pkt++;
563 			rx->rx_frag_copy_cnt++;
564 			rx->rx_copybreak_pkt++;
565 			u64_stats_update_end(&rx->statss);
566 		}
567 	} else {
568 		int recycle = gve_rx_can_recycle_buffer(page_info);
569 
570 		if (unlikely(recycle < 0)) {
571 			gve_schedule_reset(priv);
572 			return NULL;
573 		}
574 		page_info->can_flip = recycle;
575 		if (page_info->can_flip) {
576 			u64_stats_update_begin(&rx->statss);
577 			rx->rx_frag_flip_cnt++;
578 			u64_stats_update_end(&rx->statss);
579 		}
580 
581 		if (rx->data.raw_addressing) {
582 			skb = gve_rx_raw_addressing(&priv->pdev->dev, netdev,
583 						    page_info, len, napi,
584 						    data_slot,
585 						    rx->packet_buffer_size, ctx);
586 		} else {
587 			skb = gve_rx_qpl(&priv->pdev->dev, netdev, rx,
588 					 page_info, len, napi, data_slot);
589 		}
590 	}
591 	return skb;
592 }
593 
594 #define GVE_PKTCONT_BIT_IS_SET(x) (GVE_RXF_PKT_CONT & (x))
595 static void gve_rx(struct gve_rx_ring *rx, netdev_features_t feat,
596 		   struct gve_rx_desc *desc, u32 idx,
597 		   struct gve_rx_cnts *cnts)
598 {
599 	bool is_last_frag = !GVE_PKTCONT_BIT_IS_SET(desc->flags_seq);
600 	struct gve_rx_slot_page_info *page_info;
601 	u16 frag_size = be16_to_cpu(desc->len);
602 	struct gve_rx_ctx *ctx = &rx->ctx;
603 	union gve_rx_data_slot *data_slot;
604 	struct gve_priv *priv = rx->gve;
605 	struct sk_buff *skb = NULL;
606 	dma_addr_t page_bus;
607 	void *va;
608 
609 	struct napi_struct *napi = &priv->ntfy_blocks[rx->ntfy_id].napi;
610 	bool is_first_frag = ctx->frag_cnt == 0;
611 
612 	bool is_only_frag = is_first_frag && is_last_frag;
613 
614 	if (unlikely(ctx->drop_pkt))
615 		goto finish_frag;
616 
617 	if (desc->flags_seq & GVE_RXF_ERR) {
618 		ctx->drop_pkt = true;
619 		cnts->desc_err_pkt_cnt++;
620 		napi_free_frags(napi);
621 		goto finish_frag;
622 	}
623 
624 	if (unlikely(frag_size > rx->packet_buffer_size)) {
625 		netdev_warn(priv->dev, "Unexpected frag size %d, can't exceed %d, scheduling reset",
626 			    frag_size, rx->packet_buffer_size);
627 		ctx->drop_pkt = true;
628 		napi_free_frags(napi);
629 		gve_schedule_reset(rx->gve);
630 		goto finish_frag;
631 	}
632 
633 	/* Prefetch two packet buffers ahead, we will need it soon. */
634 	page_info = &rx->data.page_info[(idx + 2) & rx->mask];
635 	va = page_info->page_address + page_info->page_offset;
636 	prefetch(page_info->page); /* Kernel page struct. */
637 	prefetch(va);              /* Packet header. */
638 	prefetch(va + 64);         /* Next cacheline too. */
639 
640 	page_info = &rx->data.page_info[idx];
641 	data_slot = &rx->data.data_ring[idx];
642 	page_bus = (rx->data.raw_addressing) ?
643 		be64_to_cpu(data_slot->addr) - page_info->page_offset :
644 		rx->data.qpl->page_buses[idx];
645 	dma_sync_single_for_cpu(&priv->pdev->dev, page_bus,
646 				PAGE_SIZE, DMA_FROM_DEVICE);
647 	page_info->pad = is_first_frag ? GVE_RX_PAD : 0;
648 	frag_size -= page_info->pad;
649 
650 	skb = gve_rx_skb(priv, rx, page_info, napi, frag_size,
651 			 data_slot, is_only_frag);
652 	if (!skb) {
653 		u64_stats_update_begin(&rx->statss);
654 		rx->rx_skb_alloc_fail++;
655 		u64_stats_update_end(&rx->statss);
656 
657 		napi_free_frags(napi);
658 		ctx->drop_pkt = true;
659 		goto finish_frag;
660 	}
661 	ctx->total_size += frag_size;
662 
663 	if (is_first_frag) {
664 		if (likely(feat & NETIF_F_RXCSUM)) {
665 			/* NIC passes up the partial sum */
666 			if (desc->csum)
667 				skb->ip_summed = CHECKSUM_COMPLETE;
668 			else
669 				skb->ip_summed = CHECKSUM_NONE;
670 			skb->csum = csum_unfold(desc->csum);
671 		}
672 
673 		/* parse flags & pass relevant info up */
674 		if (likely(feat & NETIF_F_RXHASH) &&
675 		    gve_needs_rss(desc->flags_seq))
676 			skb_set_hash(skb, be32_to_cpu(desc->rss_hash),
677 				     gve_rss_type(desc->flags_seq));
678 	}
679 
680 	if (is_last_frag) {
681 		skb_record_rx_queue(skb, rx->q_num);
682 		if (skb_is_nonlinear(skb))
683 			napi_gro_frags(napi);
684 		else
685 			napi_gro_receive(napi, skb);
686 		goto finish_ok_pkt;
687 	}
688 
689 	goto finish_frag;
690 
691 finish_ok_pkt:
692 	cnts->ok_pkt_bytes += ctx->total_size;
693 	cnts->ok_pkt_cnt++;
694 finish_frag:
695 	ctx->frag_cnt++;
696 	if (is_last_frag) {
697 		cnts->total_pkt_cnt++;
698 		cnts->cont_pkt_cnt += (ctx->frag_cnt > 1);
699 		gve_rx_ctx_clear(ctx);
700 	}
701 }
702 
703 bool gve_rx_work_pending(struct gve_rx_ring *rx)
704 {
705 	struct gve_rx_desc *desc;
706 	__be16 flags_seq;
707 	u32 next_idx;
708 
709 	next_idx = rx->cnt & rx->mask;
710 	desc = rx->desc.desc_ring + next_idx;
711 
712 	flags_seq = desc->flags_seq;
713 
714 	return (GVE_SEQNO(flags_seq) == rx->desc.seqno);
715 }
716 
717 static bool gve_rx_refill_buffers(struct gve_priv *priv, struct gve_rx_ring *rx)
718 {
719 	int refill_target = rx->mask + 1;
720 	u32 fill_cnt = rx->fill_cnt;
721 
722 	while (fill_cnt - rx->cnt < refill_target) {
723 		struct gve_rx_slot_page_info *page_info;
724 		u32 idx = fill_cnt & rx->mask;
725 
726 		page_info = &rx->data.page_info[idx];
727 		if (page_info->can_flip) {
728 			/* The other half of the page is free because it was
729 			 * free when we processed the descriptor. Flip to it.
730 			 */
731 			union gve_rx_data_slot *data_slot =
732 						&rx->data.data_ring[idx];
733 
734 			gve_rx_flip_buff(page_info, &data_slot->addr);
735 			page_info->can_flip = 0;
736 		} else {
737 			/* It is possible that the networking stack has already
738 			 * finished processing all outstanding packets in the buffer
739 			 * and it can be reused.
740 			 * Flipping is unnecessary here - if the networking stack still
741 			 * owns half the page it is impossible to tell which half. Either
742 			 * the whole page is free or it needs to be replaced.
743 			 */
744 			int recycle = gve_rx_can_recycle_buffer(page_info);
745 
746 			if (recycle < 0) {
747 				if (!rx->data.raw_addressing)
748 					gve_schedule_reset(priv);
749 				return false;
750 			}
751 			if (!recycle) {
752 				/* We can't reuse the buffer - alloc a new one*/
753 				union gve_rx_data_slot *data_slot =
754 						&rx->data.data_ring[idx];
755 				struct device *dev = &priv->pdev->dev;
756 				gve_rx_free_buffer(dev, page_info, data_slot);
757 				page_info->page = NULL;
758 				if (gve_rx_alloc_buffer(priv, dev, page_info,
759 							data_slot)) {
760 					u64_stats_update_begin(&rx->statss);
761 					rx->rx_buf_alloc_fail++;
762 					u64_stats_update_end(&rx->statss);
763 					break;
764 				}
765 			}
766 		}
767 		fill_cnt++;
768 	}
769 	rx->fill_cnt = fill_cnt;
770 	return true;
771 }
772 
773 static int gve_clean_rx_done(struct gve_rx_ring *rx, int budget,
774 			     netdev_features_t feat)
775 {
776 	struct gve_rx_ctx *ctx = &rx->ctx;
777 	struct gve_priv *priv = rx->gve;
778 	struct gve_rx_cnts cnts = {0};
779 	struct gve_rx_desc *next_desc;
780 	u32 idx = rx->cnt & rx->mask;
781 	u32 work_done = 0;
782 
783 	struct gve_rx_desc *desc = &rx->desc.desc_ring[idx];
784 
785 	// Exceed budget only if (and till) the inflight packet is consumed.
786 	while ((GVE_SEQNO(desc->flags_seq) == rx->desc.seqno) &&
787 	       (work_done < budget || ctx->frag_cnt)) {
788 		next_desc = &rx->desc.desc_ring[(idx + 1) & rx->mask];
789 		prefetch(next_desc);
790 
791 		gve_rx(rx, feat, desc, idx, &cnts);
792 
793 		rx->cnt++;
794 		idx = rx->cnt & rx->mask;
795 		desc = &rx->desc.desc_ring[idx];
796 		rx->desc.seqno = gve_next_seqno(rx->desc.seqno);
797 		work_done++;
798 	}
799 
800 	// The device will only send whole packets.
801 	if (unlikely(ctx->frag_cnt)) {
802 		struct napi_struct *napi = &priv->ntfy_blocks[rx->ntfy_id].napi;
803 
804 		napi_free_frags(napi);
805 		gve_rx_ctx_clear(&rx->ctx);
806 		netdev_warn(priv->dev, "Unexpected seq number %d with incomplete packet, expected %d, scheduling reset",
807 			    GVE_SEQNO(desc->flags_seq), rx->desc.seqno);
808 		gve_schedule_reset(rx->gve);
809 	}
810 
811 	if (!work_done && rx->fill_cnt - rx->cnt > rx->db_threshold)
812 		return 0;
813 
814 	if (work_done) {
815 		u64_stats_update_begin(&rx->statss);
816 		rx->rpackets += cnts.ok_pkt_cnt;
817 		rx->rbytes += cnts.ok_pkt_bytes;
818 		rx->rx_cont_packet_cnt += cnts.cont_pkt_cnt;
819 		rx->rx_desc_err_dropped_pkt += cnts.desc_err_pkt_cnt;
820 		u64_stats_update_end(&rx->statss);
821 	}
822 
823 	/* restock ring slots */
824 	if (!rx->data.raw_addressing) {
825 		/* In QPL mode buffs are refilled as the desc are processed */
826 		rx->fill_cnt += work_done;
827 	} else if (rx->fill_cnt - rx->cnt <= rx->db_threshold) {
828 		/* In raw addressing mode buffs are only refilled if the avail
829 		 * falls below a threshold.
830 		 */
831 		if (!gve_rx_refill_buffers(priv, rx))
832 			return 0;
833 
834 		/* If we were not able to completely refill buffers, we'll want
835 		 * to schedule this queue for work again to refill buffers.
836 		 */
837 		if (rx->fill_cnt - rx->cnt <= rx->db_threshold) {
838 			gve_rx_write_doorbell(priv, rx);
839 			return budget;
840 		}
841 	}
842 
843 	gve_rx_write_doorbell(priv, rx);
844 	return cnts.total_pkt_cnt;
845 }
846 
847 int gve_rx_poll(struct gve_notify_block *block, int budget)
848 {
849 	struct gve_rx_ring *rx = block->rx;
850 	netdev_features_t feat;
851 	int work_done = 0;
852 
853 	feat = block->napi.dev->features;
854 
855 	/* If budget is 0, do all the work */
856 	if (budget == 0)
857 		budget = INT_MAX;
858 
859 	if (budget > 0)
860 		work_done = gve_clean_rx_done(rx, budget, feat);
861 
862 	return work_done;
863 }
864