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 	kvfree(rx->data.page_info);
40 	rx->data.page_info = NULL;
41 }
42 
43 static void gve_rx_free_ring(struct gve_priv *priv, int idx)
44 {
45 	struct gve_rx_ring *rx = &priv->rx[idx];
46 	struct device *dev = &priv->pdev->dev;
47 	u32 slots = rx->mask + 1;
48 	size_t bytes;
49 
50 	gve_rx_remove_from_block(priv, idx);
51 
52 	bytes = sizeof(struct gve_rx_desc) * priv->rx_desc_cnt;
53 	dma_free_coherent(dev, bytes, rx->desc.desc_ring, rx->desc.bus);
54 	rx->desc.desc_ring = NULL;
55 
56 	dma_free_coherent(dev, sizeof(*rx->q_resources),
57 			  rx->q_resources, rx->q_resources_bus);
58 	rx->q_resources = NULL;
59 
60 	gve_rx_unfill_pages(priv, rx);
61 
62 	bytes = sizeof(*rx->data.data_ring) * slots;
63 	dma_free_coherent(dev, bytes, rx->data.data_ring,
64 			  rx->data.data_bus);
65 	rx->data.data_ring = NULL;
66 	netif_dbg(priv, drv, priv->dev, "freed rx ring %d\n", idx);
67 }
68 
69 static void gve_setup_rx_buffer(struct gve_rx_slot_page_info *page_info,
70 			     dma_addr_t addr, struct page *page, __be64 *slot_addr)
71 {
72 	page_info->page = page;
73 	page_info->page_offset = 0;
74 	page_info->page_address = page_address(page);
75 	*slot_addr = cpu_to_be64(addr);
76 	/* The page already has 1 ref */
77 	page_ref_add(page, INT_MAX - 1);
78 	page_info->pagecnt_bias = INT_MAX;
79 }
80 
81 static int gve_rx_alloc_buffer(struct gve_priv *priv, struct device *dev,
82 			       struct gve_rx_slot_page_info *page_info,
83 			       union gve_rx_data_slot *data_slot)
84 {
85 	struct page *page;
86 	dma_addr_t dma;
87 	int err;
88 
89 	err = gve_alloc_page(priv, dev, &page, &dma, DMA_FROM_DEVICE);
90 	if (err)
91 		return err;
92 
93 	gve_setup_rx_buffer(page_info, dma, page, &data_slot->addr);
94 	return 0;
95 }
96 
97 static int gve_prefill_rx_pages(struct gve_rx_ring *rx)
98 {
99 	struct gve_priv *priv = rx->gve;
100 	u32 slots;
101 	int err;
102 	int i;
103 
104 	/* Allocate one page per Rx queue slot. Each page is split into two
105 	 * packet buffers, when possible we "page flip" between the two.
106 	 */
107 	slots = rx->mask + 1;
108 
109 	rx->data.page_info = kvzalloc(slots *
110 				      sizeof(*rx->data.page_info), GFP_KERNEL);
111 	if (!rx->data.page_info)
112 		return -ENOMEM;
113 
114 	if (!rx->data.raw_addressing) {
115 		rx->data.qpl = gve_assign_rx_qpl(priv);
116 		if (!rx->data.qpl) {
117 			kvfree(rx->data.page_info);
118 			rx->data.page_info = NULL;
119 			return -ENOMEM;
120 		}
121 	}
122 	for (i = 0; i < slots; i++) {
123 		if (!rx->data.raw_addressing) {
124 			struct page *page = rx->data.qpl->pages[i];
125 			dma_addr_t addr = i * PAGE_SIZE;
126 
127 			gve_setup_rx_buffer(&rx->data.page_info[i], addr, page,
128 					    &rx->data.data_ring[i].qpl_offset);
129 			continue;
130 		}
131 		err = gve_rx_alloc_buffer(priv, &priv->pdev->dev, &rx->data.page_info[i],
132 					  &rx->data.data_ring[i]);
133 		if (err)
134 			goto alloc_err;
135 	}
136 
137 	return slots;
138 alloc_err:
139 	while (i--)
140 		gve_rx_free_buffer(&priv->pdev->dev,
141 				   &rx->data.page_info[i],
142 				   &rx->data.data_ring[i]);
143 	return err;
144 }
145 
146 static void gve_rx_ctx_clear(struct gve_rx_ctx *ctx)
147 {
148 	ctx->curr_frag_cnt = 0;
149 	ctx->total_expected_size = 0;
150 	ctx->expected_frag_cnt = 0;
151 	ctx->skb_head = NULL;
152 	ctx->skb_tail = NULL;
153 	ctx->reuse_frags = false;
154 }
155 
156 static int gve_rx_alloc_ring(struct gve_priv *priv, int idx)
157 {
158 	struct gve_rx_ring *rx = &priv->rx[idx];
159 	struct device *hdev = &priv->pdev->dev;
160 	u32 slots, npages;
161 	int filled_pages;
162 	size_t bytes;
163 	int err;
164 
165 	netif_dbg(priv, drv, priv->dev, "allocating rx ring\n");
166 	/* Make sure everything is zeroed to start with */
167 	memset(rx, 0, sizeof(*rx));
168 
169 	rx->gve = priv;
170 	rx->q_num = idx;
171 
172 	slots = priv->rx_data_slot_cnt;
173 	rx->mask = slots - 1;
174 	rx->data.raw_addressing = priv->queue_format == GVE_GQI_RDA_FORMAT;
175 
176 	/* alloc rx data ring */
177 	bytes = sizeof(*rx->data.data_ring) * slots;
178 	rx->data.data_ring = dma_alloc_coherent(hdev, bytes,
179 						&rx->data.data_bus,
180 						GFP_KERNEL);
181 	if (!rx->data.data_ring)
182 		return -ENOMEM;
183 	filled_pages = gve_prefill_rx_pages(rx);
184 	if (filled_pages < 0) {
185 		err = -ENOMEM;
186 		goto abort_with_slots;
187 	}
188 	rx->fill_cnt = filled_pages;
189 	/* Ensure data ring slots (packet buffers) are visible. */
190 	dma_wmb();
191 
192 	/* Alloc gve_queue_resources */
193 	rx->q_resources =
194 		dma_alloc_coherent(hdev,
195 				   sizeof(*rx->q_resources),
196 				   &rx->q_resources_bus,
197 				   GFP_KERNEL);
198 	if (!rx->q_resources) {
199 		err = -ENOMEM;
200 		goto abort_filled;
201 	}
202 	netif_dbg(priv, drv, priv->dev, "rx[%d]->data.data_bus=%lx\n", idx,
203 		  (unsigned long)rx->data.data_bus);
204 
205 	/* alloc rx desc ring */
206 	bytes = sizeof(struct gve_rx_desc) * priv->rx_desc_cnt;
207 	npages = bytes / PAGE_SIZE;
208 	if (npages * PAGE_SIZE != bytes) {
209 		err = -EIO;
210 		goto abort_with_q_resources;
211 	}
212 
213 	rx->desc.desc_ring = dma_alloc_coherent(hdev, bytes, &rx->desc.bus,
214 						GFP_KERNEL);
215 	if (!rx->desc.desc_ring) {
216 		err = -ENOMEM;
217 		goto abort_with_q_resources;
218 	}
219 	rx->cnt = 0;
220 	rx->db_threshold = priv->rx_desc_cnt / 2;
221 	rx->desc.seqno = 1;
222 
223 	/* Allocating half-page buffers allows page-flipping which is faster
224 	 * than copying or allocating new pages.
225 	 */
226 	rx->packet_buffer_size = PAGE_SIZE / 2;
227 	gve_rx_ctx_clear(&rx->ctx);
228 	gve_rx_add_to_block(priv, idx);
229 
230 	return 0;
231 
232 abort_with_q_resources:
233 	dma_free_coherent(hdev, sizeof(*rx->q_resources),
234 			  rx->q_resources, rx->q_resources_bus);
235 	rx->q_resources = NULL;
236 abort_filled:
237 	gve_rx_unfill_pages(priv, rx);
238 abort_with_slots:
239 	bytes = sizeof(*rx->data.data_ring) * slots;
240 	dma_free_coherent(hdev, bytes, rx->data.data_ring, rx->data.data_bus);
241 	rx->data.data_ring = NULL;
242 
243 	return err;
244 }
245 
246 int gve_rx_alloc_rings(struct gve_priv *priv)
247 {
248 	int err = 0;
249 	int i;
250 
251 	for (i = 0; i < priv->rx_cfg.num_queues; i++) {
252 		err = gve_rx_alloc_ring(priv, i);
253 		if (err) {
254 			netif_err(priv, drv, priv->dev,
255 				  "Failed to alloc rx ring=%d: err=%d\n",
256 				  i, err);
257 			break;
258 		}
259 	}
260 	/* Unallocate if there was an error */
261 	if (err) {
262 		int j;
263 
264 		for (j = 0; j < i; j++)
265 			gve_rx_free_ring(priv, j);
266 	}
267 	return err;
268 }
269 
270 void gve_rx_free_rings_gqi(struct gve_priv *priv)
271 {
272 	int i;
273 
274 	for (i = 0; i < priv->rx_cfg.num_queues; i++)
275 		gve_rx_free_ring(priv, i);
276 }
277 
278 void gve_rx_write_doorbell(struct gve_priv *priv, struct gve_rx_ring *rx)
279 {
280 	u32 db_idx = be32_to_cpu(rx->q_resources->db_index);
281 
282 	iowrite32be(rx->fill_cnt, &priv->db_bar2[db_idx]);
283 }
284 
285 static enum pkt_hash_types gve_rss_type(__be16 pkt_flags)
286 {
287 	if (likely(pkt_flags & (GVE_RXF_TCP | GVE_RXF_UDP)))
288 		return PKT_HASH_TYPE_L4;
289 	if (pkt_flags & (GVE_RXF_IPV4 | GVE_RXF_IPV6))
290 		return PKT_HASH_TYPE_L3;
291 	return PKT_HASH_TYPE_L2;
292 }
293 
294 static u16 gve_rx_ctx_padding(struct gve_rx_ctx *ctx)
295 {
296 	return (ctx->curr_frag_cnt == 0) ? GVE_RX_PAD : 0;
297 }
298 
299 static struct sk_buff *gve_rx_add_frags(struct napi_struct *napi,
300 					struct gve_rx_slot_page_info *page_info,
301 					u16 packet_buffer_size, u16 len,
302 					struct gve_rx_ctx *ctx)
303 {
304 	u32 offset = page_info->page_offset +  gve_rx_ctx_padding(ctx);
305 	struct sk_buff *skb;
306 
307 	if (!ctx->skb_head)
308 		ctx->skb_head = napi_get_frags(napi);
309 
310 	if (unlikely(!ctx->skb_head))
311 		return NULL;
312 
313 	skb = ctx->skb_head;
314 	skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page_info->page,
315 			offset, len, packet_buffer_size);
316 
317 	return skb;
318 }
319 
320 static void gve_rx_flip_buff(struct gve_rx_slot_page_info *page_info, __be64 *slot_addr)
321 {
322 	const __be64 offset = cpu_to_be64(PAGE_SIZE / 2);
323 
324 	/* "flip" to other packet buffer on this page */
325 	page_info->page_offset ^= PAGE_SIZE / 2;
326 	*(slot_addr) ^= offset;
327 }
328 
329 static int gve_rx_can_recycle_buffer(struct gve_rx_slot_page_info *page_info)
330 {
331 	int pagecount = page_count(page_info->page);
332 
333 	/* This page is not being used by any SKBs - reuse */
334 	if (pagecount == page_info->pagecnt_bias)
335 		return 1;
336 	/* This page is still being used by an SKB - we can't reuse */
337 	else if (pagecount > page_info->pagecnt_bias)
338 		return 0;
339 	WARN(pagecount < page_info->pagecnt_bias,
340 	     "Pagecount should never be less than the bias.");
341 	return -1;
342 }
343 
344 static struct sk_buff *
345 gve_rx_raw_addressing(struct device *dev, struct net_device *netdev,
346 		      struct gve_rx_slot_page_info *page_info, u16 len,
347 		      struct napi_struct *napi,
348 		      union gve_rx_data_slot *data_slot,
349 		      u16 packet_buffer_size, struct gve_rx_ctx *ctx)
350 {
351 	struct sk_buff *skb = gve_rx_add_frags(napi, page_info, packet_buffer_size, len, ctx);
352 
353 	if (!skb)
354 		return NULL;
355 
356 	/* Optimistically stop the kernel from freeing the page.
357 	 * We will check again in refill to determine if we need to alloc a
358 	 * new page.
359 	 */
360 	gve_dec_pagecnt_bias(page_info);
361 
362 	return skb;
363 }
364 
365 static struct sk_buff *
366 gve_rx_qpl(struct device *dev, struct net_device *netdev,
367 	   struct gve_rx_ring *rx, struct gve_rx_slot_page_info *page_info,
368 	   u16 len, struct napi_struct *napi,
369 	   union gve_rx_data_slot *data_slot)
370 {
371 	struct gve_rx_ctx *ctx = &rx->ctx;
372 	struct sk_buff *skb;
373 
374 	/* if raw_addressing mode is not enabled gvnic can only receive into
375 	 * registered segments. If the buffer can't be recycled, our only
376 	 * choice is to copy the data out of it so that we can return it to the
377 	 * device.
378 	 */
379 	if (ctx->reuse_frags) {
380 		skb = gve_rx_add_frags(napi, page_info, rx->packet_buffer_size, len, ctx);
381 		/* No point in recycling if we didn't get the skb */
382 		if (skb) {
383 			/* Make sure that the page isn't freed. */
384 			gve_dec_pagecnt_bias(page_info);
385 			gve_rx_flip_buff(page_info, &data_slot->qpl_offset);
386 		}
387 	} else {
388 		const u16 padding = gve_rx_ctx_padding(ctx);
389 
390 		skb = gve_rx_copy(netdev, napi, page_info, len, padding, ctx);
391 		if (skb) {
392 			u64_stats_update_begin(&rx->statss);
393 			rx->rx_frag_copy_cnt++;
394 			u64_stats_update_end(&rx->statss);
395 		}
396 	}
397 	return skb;
398 }
399 
400 #define GVE_PKTCONT_BIT_IS_SET(x) (GVE_RXF_PKT_CONT & (x))
401 static u16 gve_rx_get_fragment_size(struct gve_rx_ctx *ctx, struct gve_rx_desc *desc)
402 {
403 	return be16_to_cpu(desc->len) - gve_rx_ctx_padding(ctx);
404 }
405 
406 static bool gve_rx_ctx_init(struct gve_rx_ctx *ctx, struct gve_rx_ring *rx)
407 {
408 	bool qpl_mode = !rx->data.raw_addressing, packet_size_error = false;
409 	bool buffer_error = false, desc_error = false, seqno_error = false;
410 	struct gve_rx_slot_page_info *page_info;
411 	struct gve_priv *priv = rx->gve;
412 	u32 idx = rx->cnt & rx->mask;
413 	bool reuse_frags, can_flip;
414 	struct gve_rx_desc *desc;
415 	u16 packet_size = 0;
416 	u16 n_frags = 0;
417 	int recycle;
418 
419 	/** In QPL mode, we only flip buffers when all buffers containing the packet
420 	 * can be flipped. RDA can_flip decisions will be made later, per frag.
421 	 */
422 	can_flip = qpl_mode;
423 	reuse_frags = can_flip;
424 	do {
425 		u16 frag_size;
426 
427 		n_frags++;
428 		desc = &rx->desc.desc_ring[idx];
429 		desc_error = unlikely(desc->flags_seq & GVE_RXF_ERR) || desc_error;
430 		if (GVE_SEQNO(desc->flags_seq) != rx->desc.seqno) {
431 			seqno_error = true;
432 			netdev_warn(priv->dev,
433 				    "RX seqno error: want=%d, got=%d, dropping packet and scheduling reset.",
434 				    rx->desc.seqno, GVE_SEQNO(desc->flags_seq));
435 		}
436 		frag_size = be16_to_cpu(desc->len);
437 		packet_size += frag_size;
438 		if (frag_size > rx->packet_buffer_size) {
439 			packet_size_error = true;
440 			netdev_warn(priv->dev,
441 				    "RX fragment error: packet_buffer_size=%d, frag_size=%d, droping packet.",
442 				    rx->packet_buffer_size, be16_to_cpu(desc->len));
443 		}
444 		page_info = &rx->data.page_info[idx];
445 		if (can_flip) {
446 			recycle = gve_rx_can_recycle_buffer(page_info);
447 			reuse_frags = reuse_frags && recycle > 0;
448 			buffer_error = buffer_error || unlikely(recycle < 0);
449 		}
450 		idx = (idx + 1) & rx->mask;
451 		rx->desc.seqno = gve_next_seqno(rx->desc.seqno);
452 	} while (GVE_PKTCONT_BIT_IS_SET(desc->flags_seq));
453 
454 	prefetch(rx->desc.desc_ring + idx);
455 
456 	ctx->curr_frag_cnt = 0;
457 	ctx->total_expected_size = packet_size - GVE_RX_PAD;
458 	ctx->expected_frag_cnt = n_frags;
459 	ctx->skb_head = NULL;
460 	ctx->reuse_frags = reuse_frags;
461 
462 	if (ctx->expected_frag_cnt > 1) {
463 		u64_stats_update_begin(&rx->statss);
464 		rx->rx_cont_packet_cnt++;
465 		u64_stats_update_end(&rx->statss);
466 	}
467 	if (ctx->total_expected_size > priv->rx_copybreak && !ctx->reuse_frags && qpl_mode) {
468 		u64_stats_update_begin(&rx->statss);
469 		rx->rx_copied_pkt++;
470 		u64_stats_update_end(&rx->statss);
471 	}
472 
473 	if (unlikely(buffer_error || seqno_error || packet_size_error)) {
474 		gve_schedule_reset(priv);
475 		return false;
476 	}
477 
478 	if (unlikely(desc_error)) {
479 		u64_stats_update_begin(&rx->statss);
480 		rx->rx_desc_err_dropped_pkt++;
481 		u64_stats_update_end(&rx->statss);
482 		return false;
483 	}
484 	return true;
485 }
486 
487 static struct sk_buff *gve_rx_skb(struct gve_priv *priv, struct gve_rx_ring *rx,
488 				  struct gve_rx_slot_page_info *page_info, struct napi_struct *napi,
489 				  u16 len, union gve_rx_data_slot *data_slot)
490 {
491 	struct net_device *netdev = priv->dev;
492 	struct gve_rx_ctx *ctx = &rx->ctx;
493 	struct sk_buff *skb = NULL;
494 
495 	if (len <= priv->rx_copybreak && ctx->expected_frag_cnt == 1) {
496 		/* Just copy small packets */
497 		skb = gve_rx_copy(netdev, napi, page_info, len, GVE_RX_PAD, ctx);
498 		if (skb) {
499 			u64_stats_update_begin(&rx->statss);
500 			rx->rx_copied_pkt++;
501 			rx->rx_frag_copy_cnt++;
502 			rx->rx_copybreak_pkt++;
503 			u64_stats_update_end(&rx->statss);
504 		}
505 	} else {
506 		if (rx->data.raw_addressing) {
507 			int recycle = gve_rx_can_recycle_buffer(page_info);
508 
509 			if (unlikely(recycle < 0)) {
510 				gve_schedule_reset(priv);
511 				return NULL;
512 			}
513 			page_info->can_flip = recycle;
514 			if (page_info->can_flip) {
515 				u64_stats_update_begin(&rx->statss);
516 				rx->rx_frag_flip_cnt++;
517 				u64_stats_update_end(&rx->statss);
518 			}
519 			skb = gve_rx_raw_addressing(&priv->pdev->dev, netdev,
520 						    page_info, len, napi,
521 						    data_slot,
522 						    rx->packet_buffer_size, ctx);
523 		} else {
524 			if (ctx->reuse_frags) {
525 				u64_stats_update_begin(&rx->statss);
526 				rx->rx_frag_flip_cnt++;
527 				u64_stats_update_end(&rx->statss);
528 			}
529 			skb = gve_rx_qpl(&priv->pdev->dev, netdev, rx,
530 					 page_info, len, napi, data_slot);
531 		}
532 	}
533 	return skb;
534 }
535 
536 static bool gve_rx(struct gve_rx_ring *rx, netdev_features_t feat,
537 		   u64 *packet_size_bytes, u32 *work_done)
538 {
539 	struct gve_rx_slot_page_info *page_info;
540 	struct gve_rx_ctx *ctx = &rx->ctx;
541 	union gve_rx_data_slot *data_slot;
542 	struct gve_priv *priv = rx->gve;
543 	struct gve_rx_desc *first_desc;
544 	struct sk_buff *skb = NULL;
545 	struct gve_rx_desc *desc;
546 	struct napi_struct *napi;
547 	dma_addr_t page_bus;
548 	u32 work_cnt = 0;
549 	void *va;
550 	u32 idx;
551 	u16 len;
552 
553 	idx = rx->cnt & rx->mask;
554 	first_desc = &rx->desc.desc_ring[idx];
555 	desc = first_desc;
556 	napi = &priv->ntfy_blocks[rx->ntfy_id].napi;
557 
558 	if (unlikely(!gve_rx_ctx_init(ctx, rx)))
559 		goto skb_alloc_fail;
560 
561 	while (ctx->curr_frag_cnt < ctx->expected_frag_cnt) {
562 		/* Prefetch two packet buffers ahead, we will need it soon. */
563 		page_info = &rx->data.page_info[(idx + 2) & rx->mask];
564 		va = page_info->page_address + page_info->page_offset;
565 
566 		prefetch(page_info->page); /* Kernel page struct. */
567 		prefetch(va);              /* Packet header. */
568 		prefetch(va + 64);         /* Next cacheline too. */
569 
570 		len = gve_rx_get_fragment_size(ctx, desc);
571 
572 		page_info = &rx->data.page_info[idx];
573 		data_slot = &rx->data.data_ring[idx];
574 		page_bus = rx->data.raw_addressing ?
575 			   be64_to_cpu(data_slot->addr) - page_info->page_offset :
576 			   rx->data.qpl->page_buses[idx];
577 		dma_sync_single_for_cpu(&priv->pdev->dev, page_bus, PAGE_SIZE, DMA_FROM_DEVICE);
578 
579 		skb = gve_rx_skb(priv, rx, page_info, napi, len, data_slot);
580 		if (!skb) {
581 			u64_stats_update_begin(&rx->statss);
582 			rx->rx_skb_alloc_fail++;
583 			u64_stats_update_end(&rx->statss);
584 			goto skb_alloc_fail;
585 		}
586 
587 		ctx->curr_frag_cnt++;
588 		rx->cnt++;
589 		idx = rx->cnt & rx->mask;
590 		work_cnt++;
591 		desc = &rx->desc.desc_ring[idx];
592 	}
593 
594 	if (likely(feat & NETIF_F_RXCSUM)) {
595 		/* NIC passes up the partial sum */
596 		if (first_desc->csum)
597 			skb->ip_summed = CHECKSUM_COMPLETE;
598 		else
599 			skb->ip_summed = CHECKSUM_NONE;
600 		skb->csum = csum_unfold(first_desc->csum);
601 	}
602 
603 	/* parse flags & pass relevant info up */
604 	if (likely(feat & NETIF_F_RXHASH) &&
605 	    gve_needs_rss(first_desc->flags_seq))
606 		skb_set_hash(skb, be32_to_cpu(first_desc->rss_hash),
607 			     gve_rss_type(first_desc->flags_seq));
608 
609 	*packet_size_bytes = skb->len + (skb->protocol ? ETH_HLEN : 0);
610 	*work_done = work_cnt;
611 	if (skb_is_nonlinear(skb))
612 		napi_gro_frags(napi);
613 	else
614 		napi_gro_receive(napi, skb);
615 
616 	gve_rx_ctx_clear(ctx);
617 	return true;
618 
619 skb_alloc_fail:
620 	if (napi->skb)
621 		napi_free_frags(napi);
622 	*packet_size_bytes = 0;
623 	*work_done = ctx->expected_frag_cnt;
624 	while (ctx->curr_frag_cnt < ctx->expected_frag_cnt) {
625 		rx->cnt++;
626 		ctx->curr_frag_cnt++;
627 	}
628 	gve_rx_ctx_clear(ctx);
629 	return false;
630 }
631 
632 bool gve_rx_work_pending(struct gve_rx_ring *rx)
633 {
634 	struct gve_rx_desc *desc;
635 	__be16 flags_seq;
636 	u32 next_idx;
637 
638 	next_idx = rx->cnt & rx->mask;
639 	desc = rx->desc.desc_ring + next_idx;
640 
641 	flags_seq = desc->flags_seq;
642 	/* Make sure we have synchronized the seq no with the device */
643 	smp_rmb();
644 
645 	return (GVE_SEQNO(flags_seq) == rx->desc.seqno);
646 }
647 
648 static bool gve_rx_refill_buffers(struct gve_priv *priv, struct gve_rx_ring *rx)
649 {
650 	int refill_target = rx->mask + 1;
651 	u32 fill_cnt = rx->fill_cnt;
652 
653 	while (fill_cnt - rx->cnt < refill_target) {
654 		struct gve_rx_slot_page_info *page_info;
655 		u32 idx = fill_cnt & rx->mask;
656 
657 		page_info = &rx->data.page_info[idx];
658 		if (page_info->can_flip) {
659 			/* The other half of the page is free because it was
660 			 * free when we processed the descriptor. Flip to it.
661 			 */
662 			union gve_rx_data_slot *data_slot =
663 						&rx->data.data_ring[idx];
664 
665 			gve_rx_flip_buff(page_info, &data_slot->addr);
666 			page_info->can_flip = 0;
667 		} else {
668 			/* It is possible that the networking stack has already
669 			 * finished processing all outstanding packets in the buffer
670 			 * and it can be reused.
671 			 * Flipping is unnecessary here - if the networking stack still
672 			 * owns half the page it is impossible to tell which half. Either
673 			 * the whole page is free or it needs to be replaced.
674 			 */
675 			int recycle = gve_rx_can_recycle_buffer(page_info);
676 
677 			if (recycle < 0) {
678 				if (!rx->data.raw_addressing)
679 					gve_schedule_reset(priv);
680 				return false;
681 			}
682 			if (!recycle) {
683 				/* We can't reuse the buffer - alloc a new one*/
684 				union gve_rx_data_slot *data_slot =
685 						&rx->data.data_ring[idx];
686 				struct device *dev = &priv->pdev->dev;
687 				gve_rx_free_buffer(dev, page_info, data_slot);
688 				page_info->page = NULL;
689 				if (gve_rx_alloc_buffer(priv, dev, page_info,
690 							data_slot)) {
691 					u64_stats_update_begin(&rx->statss);
692 					rx->rx_buf_alloc_fail++;
693 					u64_stats_update_end(&rx->statss);
694 					break;
695 				}
696 			}
697 		}
698 		fill_cnt++;
699 	}
700 	rx->fill_cnt = fill_cnt;
701 	return true;
702 }
703 
704 static int gve_clean_rx_done(struct gve_rx_ring *rx, int budget,
705 			     netdev_features_t feat)
706 {
707 	u32 work_done = 0, total_packet_cnt = 0, ok_packet_cnt = 0;
708 	struct gve_priv *priv = rx->gve;
709 	u32 idx = rx->cnt & rx->mask;
710 	struct gve_rx_desc *desc;
711 	u64 bytes = 0;
712 
713 	desc = &rx->desc.desc_ring[idx];
714 	while ((GVE_SEQNO(desc->flags_seq) == rx->desc.seqno) &&
715 	       work_done < budget) {
716 		u64 packet_size_bytes = 0;
717 		u32 work_cnt = 0;
718 		bool dropped;
719 
720 		netif_info(priv, rx_status, priv->dev,
721 			   "[%d] idx=%d desc=%p desc->flags_seq=0x%x\n",
722 			   rx->q_num, idx, desc, desc->flags_seq);
723 		netif_info(priv, rx_status, priv->dev,
724 			   "[%d] seqno=%d rx->desc.seqno=%d\n",
725 			   rx->q_num, GVE_SEQNO(desc->flags_seq),
726 			   rx->desc.seqno);
727 
728 		dropped = !gve_rx(rx, feat, &packet_size_bytes, &work_cnt);
729 		if (!dropped) {
730 			bytes += packet_size_bytes;
731 			ok_packet_cnt++;
732 		}
733 		total_packet_cnt++;
734 		idx = rx->cnt & rx->mask;
735 		desc = &rx->desc.desc_ring[idx];
736 		work_done += work_cnt;
737 	}
738 
739 	if (!work_done && rx->fill_cnt - rx->cnt > rx->db_threshold)
740 		return 0;
741 
742 	if (work_done) {
743 		u64_stats_update_begin(&rx->statss);
744 		rx->rpackets += ok_packet_cnt;
745 		rx->rbytes += bytes;
746 		u64_stats_update_end(&rx->statss);
747 	}
748 
749 	/* restock ring slots */
750 	if (!rx->data.raw_addressing) {
751 		/* In QPL mode buffs are refilled as the desc are processed */
752 		rx->fill_cnt += work_done;
753 	} else if (rx->fill_cnt - rx->cnt <= rx->db_threshold) {
754 		/* In raw addressing mode buffs are only refilled if the avail
755 		 * falls below a threshold.
756 		 */
757 		if (!gve_rx_refill_buffers(priv, rx))
758 			return 0;
759 
760 		/* If we were not able to completely refill buffers, we'll want
761 		 * to schedule this queue for work again to refill buffers.
762 		 */
763 		if (rx->fill_cnt - rx->cnt <= rx->db_threshold) {
764 			gve_rx_write_doorbell(priv, rx);
765 			return budget;
766 		}
767 	}
768 
769 	gve_rx_write_doorbell(priv, rx);
770 	return total_packet_cnt;
771 }
772 
773 int gve_rx_poll(struct gve_notify_block *block, int budget)
774 {
775 	struct gve_rx_ring *rx = block->rx;
776 	netdev_features_t feat;
777 	int work_done = 0;
778 
779 	feat = block->napi.dev->features;
780 
781 	/* If budget is 0, do all the work */
782 	if (budget == 0)
783 		budget = INT_MAX;
784 
785 	if (budget > 0)
786 		work_done = gve_clean_rx_done(rx, budget, feat);
787 
788 	return work_done;
789 }
790