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_dqo.h"
9 #include "gve_adminq.h"
10 #include "gve_utils.h"
11 #include <linux/ip.h>
12 #include <linux/ipv6.h>
13 #include <linux/skbuff.h>
14 #include <linux/slab.h>
15 #include <net/ip6_checksum.h>
16 #include <net/ipv6.h>
17 #include <net/tcp.h>
18 
gve_rx_free_hdr_bufs(struct gve_priv * priv,struct gve_rx_ring * rx)19 static void gve_rx_free_hdr_bufs(struct gve_priv *priv, struct gve_rx_ring *rx)
20 {
21 	struct device *hdev = &priv->pdev->dev;
22 	int buf_count = rx->dqo.bufq.mask + 1;
23 
24 	if (rx->dqo.hdr_bufs.data) {
25 		dma_free_coherent(hdev, priv->header_buf_size * buf_count,
26 				  rx->dqo.hdr_bufs.data, rx->dqo.hdr_bufs.addr);
27 		rx->dqo.hdr_bufs.data = NULL;
28 	}
29 }
30 
gve_rx_init_ring_state_dqo(struct gve_rx_ring * rx,const u32 buffer_queue_slots,const u32 completion_queue_slots)31 static void gve_rx_init_ring_state_dqo(struct gve_rx_ring *rx,
32 				       const u32 buffer_queue_slots,
33 				       const u32 completion_queue_slots)
34 {
35 	int i;
36 
37 	/* Set buffer queue state */
38 	rx->dqo.bufq.mask = buffer_queue_slots - 1;
39 	rx->dqo.bufq.head = 0;
40 	rx->dqo.bufq.tail = 0;
41 
42 	/* Set completion queue state */
43 	rx->dqo.complq.num_free_slots = completion_queue_slots;
44 	rx->dqo.complq.mask = completion_queue_slots - 1;
45 	rx->dqo.complq.cur_gen_bit = 0;
46 	rx->dqo.complq.head = 0;
47 
48 	/* Set RX SKB context */
49 	rx->ctx.skb_head = NULL;
50 	rx->ctx.skb_tail = NULL;
51 
52 	/* Set up linked list of buffer IDs */
53 	if (rx->dqo.buf_states) {
54 		for (i = 0; i < rx->dqo.num_buf_states - 1; i++)
55 			rx->dqo.buf_states[i].next = i + 1;
56 		rx->dqo.buf_states[rx->dqo.num_buf_states - 1].next = -1;
57 	}
58 
59 	rx->dqo.free_buf_states = 0;
60 	rx->dqo.recycled_buf_states.head = -1;
61 	rx->dqo.recycled_buf_states.tail = -1;
62 	rx->dqo.used_buf_states.head = -1;
63 	rx->dqo.used_buf_states.tail = -1;
64 }
65 
gve_rx_reset_ring_dqo(struct gve_priv * priv,int idx)66 static void gve_rx_reset_ring_dqo(struct gve_priv *priv, int idx)
67 {
68 	struct gve_rx_ring *rx = &priv->rx[idx];
69 	size_t size;
70 	int i;
71 
72 	const u32 buffer_queue_slots = priv->rx_desc_cnt;
73 	const u32 completion_queue_slots = priv->rx_desc_cnt;
74 
75 	/* Reset buffer queue */
76 	if (rx->dqo.bufq.desc_ring) {
77 		size = sizeof(rx->dqo.bufq.desc_ring[0]) *
78 			buffer_queue_slots;
79 		memset(rx->dqo.bufq.desc_ring, 0, size);
80 	}
81 
82 	/* Reset completion queue */
83 	if (rx->dqo.complq.desc_ring) {
84 		size = sizeof(rx->dqo.complq.desc_ring[0]) *
85 			completion_queue_slots;
86 		memset(rx->dqo.complq.desc_ring, 0, size);
87 	}
88 
89 	/* Reset q_resources */
90 	if (rx->q_resources)
91 		memset(rx->q_resources, 0, sizeof(*rx->q_resources));
92 
93 	/* Reset buf states */
94 	if (rx->dqo.buf_states) {
95 		for (i = 0; i < rx->dqo.num_buf_states; i++) {
96 			struct gve_rx_buf_state_dqo *bs = &rx->dqo.buf_states[i];
97 
98 			if (rx->dqo.page_pool)
99 				gve_free_to_page_pool(rx, bs, false);
100 			else
101 				gve_free_qpl_page_dqo(bs);
102 		}
103 	}
104 
105 	gve_rx_init_ring_state_dqo(rx, buffer_queue_slots,
106 				   completion_queue_slots);
107 }
108 
gve_rx_stop_ring_dqo(struct gve_priv * priv,int idx)109 void gve_rx_stop_ring_dqo(struct gve_priv *priv, int idx)
110 {
111 	int ntfy_idx = gve_rx_idx_to_ntfy(priv, idx);
112 	struct gve_rx_ring *rx = &priv->rx[idx];
113 
114 	if (!gve_rx_was_added_to_block(priv, idx))
115 		return;
116 
117 	if (rx->dqo.page_pool)
118 		page_pool_disable_direct_recycling(rx->dqo.page_pool);
119 	gve_remove_napi(priv, ntfy_idx);
120 	gve_rx_remove_from_block(priv, idx);
121 	gve_rx_reset_ring_dqo(priv, idx);
122 }
123 
gve_rx_free_ring_dqo(struct gve_priv * priv,struct gve_rx_ring * rx,struct gve_rx_alloc_rings_cfg * cfg)124 void gve_rx_free_ring_dqo(struct gve_priv *priv, struct gve_rx_ring *rx,
125 			  struct gve_rx_alloc_rings_cfg *cfg)
126 {
127 	struct device *hdev = &priv->pdev->dev;
128 	size_t completion_queue_slots;
129 	size_t buffer_queue_slots;
130 	int idx = rx->q_num;
131 	size_t size;
132 	u32 qpl_id;
133 	int i;
134 
135 	completion_queue_slots = rx->dqo.complq.mask + 1;
136 	buffer_queue_slots = rx->dqo.bufq.mask + 1;
137 
138 	if (rx->q_resources) {
139 		dma_free_coherent(hdev, sizeof(*rx->q_resources),
140 				  rx->q_resources, rx->q_resources_bus);
141 		rx->q_resources = NULL;
142 	}
143 
144 	for (i = 0; i < rx->dqo.num_buf_states; i++) {
145 		struct gve_rx_buf_state_dqo *bs = &rx->dqo.buf_states[i];
146 
147 		if (rx->dqo.page_pool)
148 			gve_free_to_page_pool(rx, bs, false);
149 		else
150 			gve_free_qpl_page_dqo(bs);
151 	}
152 
153 	if (rx->dqo.qpl) {
154 		qpl_id = gve_get_rx_qpl_id(cfg->qcfg_tx, rx->q_num);
155 		gve_free_queue_page_list(priv, rx->dqo.qpl, qpl_id);
156 		rx->dqo.qpl = NULL;
157 	}
158 
159 	if (rx->dqo.bufq.desc_ring) {
160 		size = sizeof(rx->dqo.bufq.desc_ring[0]) * buffer_queue_slots;
161 		dma_free_coherent(hdev, size, rx->dqo.bufq.desc_ring,
162 				  rx->dqo.bufq.bus);
163 		rx->dqo.bufq.desc_ring = NULL;
164 	}
165 
166 	if (rx->dqo.complq.desc_ring) {
167 		size = sizeof(rx->dqo.complq.desc_ring[0]) *
168 			completion_queue_slots;
169 		dma_free_coherent(hdev, size, rx->dqo.complq.desc_ring,
170 				  rx->dqo.complq.bus);
171 		rx->dqo.complq.desc_ring = NULL;
172 	}
173 
174 	kvfree(rx->dqo.buf_states);
175 	rx->dqo.buf_states = NULL;
176 
177 	if (rx->dqo.page_pool) {
178 		page_pool_destroy(rx->dqo.page_pool);
179 		rx->dqo.page_pool = NULL;
180 	}
181 
182 	gve_rx_free_hdr_bufs(priv, rx);
183 
184 	netif_dbg(priv, drv, priv->dev, "freed rx ring %d\n", idx);
185 }
186 
gve_rx_alloc_hdr_bufs(struct gve_priv * priv,struct gve_rx_ring * rx,const u32 buf_count)187 static int gve_rx_alloc_hdr_bufs(struct gve_priv *priv, struct gve_rx_ring *rx,
188 				 const u32 buf_count)
189 {
190 	struct device *hdev = &priv->pdev->dev;
191 
192 	rx->dqo.hdr_bufs.data = dma_alloc_coherent(hdev, priv->header_buf_size * buf_count,
193 						   &rx->dqo.hdr_bufs.addr, GFP_KERNEL);
194 	if (!rx->dqo.hdr_bufs.data)
195 		return -ENOMEM;
196 
197 	return 0;
198 }
199 
gve_rx_start_ring_dqo(struct gve_priv * priv,int idx)200 void gve_rx_start_ring_dqo(struct gve_priv *priv, int idx)
201 {
202 	int ntfy_idx = gve_rx_idx_to_ntfy(priv, idx);
203 
204 	gve_rx_add_to_block(priv, idx);
205 	gve_add_napi(priv, ntfy_idx, gve_napi_poll_dqo);
206 }
207 
gve_rx_alloc_ring_dqo(struct gve_priv * priv,struct gve_rx_alloc_rings_cfg * cfg,struct gve_rx_ring * rx,int idx)208 int gve_rx_alloc_ring_dqo(struct gve_priv *priv,
209 			  struct gve_rx_alloc_rings_cfg *cfg,
210 			  struct gve_rx_ring *rx,
211 			  int idx)
212 {
213 	struct device *hdev = &priv->pdev->dev;
214 	struct page_pool *pool;
215 	int qpl_page_cnt;
216 	size_t size;
217 	u32 qpl_id;
218 
219 	const u32 buffer_queue_slots = cfg->ring_size;
220 	const u32 completion_queue_slots = cfg->ring_size;
221 
222 	netif_dbg(priv, drv, priv->dev, "allocating rx ring DQO\n");
223 
224 	memset(rx, 0, sizeof(*rx));
225 	rx->gve = priv;
226 	rx->q_num = idx;
227 
228 	rx->dqo.num_buf_states = cfg->raw_addressing ? buffer_queue_slots :
229 		gve_get_rx_pages_per_qpl_dqo(cfg->ring_size);
230 	rx->dqo.buf_states = kvcalloc(rx->dqo.num_buf_states,
231 				      sizeof(rx->dqo.buf_states[0]),
232 				      GFP_KERNEL);
233 	if (!rx->dqo.buf_states)
234 		return -ENOMEM;
235 
236 	/* Allocate header buffers for header-split */
237 	if (cfg->enable_header_split)
238 		if (gve_rx_alloc_hdr_bufs(priv, rx, buffer_queue_slots))
239 			goto err;
240 
241 	/* Allocate RX completion queue */
242 	size = sizeof(rx->dqo.complq.desc_ring[0]) *
243 		completion_queue_slots;
244 	rx->dqo.complq.desc_ring =
245 		dma_alloc_coherent(hdev, size, &rx->dqo.complq.bus, GFP_KERNEL);
246 	if (!rx->dqo.complq.desc_ring)
247 		goto err;
248 
249 	/* Allocate RX buffer queue */
250 	size = sizeof(rx->dqo.bufq.desc_ring[0]) * buffer_queue_slots;
251 	rx->dqo.bufq.desc_ring =
252 		dma_alloc_coherent(hdev, size, &rx->dqo.bufq.bus, GFP_KERNEL);
253 	if (!rx->dqo.bufq.desc_ring)
254 		goto err;
255 
256 	if (cfg->raw_addressing) {
257 		pool = gve_rx_create_page_pool(priv, rx);
258 		if (IS_ERR(pool))
259 			goto err;
260 
261 		rx->dqo.page_pool = pool;
262 	} else {
263 		qpl_id = gve_get_rx_qpl_id(cfg->qcfg_tx, rx->q_num);
264 		qpl_page_cnt = gve_get_rx_pages_per_qpl_dqo(cfg->ring_size);
265 
266 		rx->dqo.qpl = gve_alloc_queue_page_list(priv, qpl_id,
267 							qpl_page_cnt);
268 		if (!rx->dqo.qpl)
269 			goto err;
270 		rx->dqo.next_qpl_page_idx = 0;
271 	}
272 
273 	rx->q_resources = dma_alloc_coherent(hdev, sizeof(*rx->q_resources),
274 					     &rx->q_resources_bus, GFP_KERNEL);
275 	if (!rx->q_resources)
276 		goto err;
277 
278 	gve_rx_init_ring_state_dqo(rx, buffer_queue_slots,
279 				   completion_queue_slots);
280 
281 	return 0;
282 
283 err:
284 	gve_rx_free_ring_dqo(priv, rx, cfg);
285 	return -ENOMEM;
286 }
287 
gve_rx_write_doorbell_dqo(const struct gve_priv * priv,int queue_idx)288 void gve_rx_write_doorbell_dqo(const struct gve_priv *priv, int queue_idx)
289 {
290 	const struct gve_rx_ring *rx = &priv->rx[queue_idx];
291 	u64 index = be32_to_cpu(rx->q_resources->db_index);
292 
293 	iowrite32(rx->dqo.bufq.tail, &priv->db_bar2[index]);
294 }
295 
gve_rx_alloc_rings_dqo(struct gve_priv * priv,struct gve_rx_alloc_rings_cfg * cfg)296 int gve_rx_alloc_rings_dqo(struct gve_priv *priv,
297 			   struct gve_rx_alloc_rings_cfg *cfg)
298 {
299 	struct gve_rx_ring *rx;
300 	int err;
301 	int i;
302 
303 	rx = kvcalloc(cfg->qcfg->max_queues, sizeof(struct gve_rx_ring),
304 		      GFP_KERNEL);
305 	if (!rx)
306 		return -ENOMEM;
307 
308 	for (i = 0; i < cfg->qcfg->num_queues; i++) {
309 		err = gve_rx_alloc_ring_dqo(priv, cfg, &rx[i], i);
310 		if (err) {
311 			netif_err(priv, drv, priv->dev,
312 				  "Failed to alloc rx ring=%d: err=%d\n",
313 				  i, err);
314 			goto err;
315 		}
316 	}
317 
318 	cfg->rx = rx;
319 	return 0;
320 
321 err:
322 	for (i--; i >= 0; i--)
323 		gve_rx_free_ring_dqo(priv, &rx[i], cfg);
324 	kvfree(rx);
325 	return err;
326 }
327 
gve_rx_free_rings_dqo(struct gve_priv * priv,struct gve_rx_alloc_rings_cfg * cfg)328 void gve_rx_free_rings_dqo(struct gve_priv *priv,
329 			   struct gve_rx_alloc_rings_cfg *cfg)
330 {
331 	struct gve_rx_ring *rx = cfg->rx;
332 	int i;
333 
334 	if (!rx)
335 		return;
336 
337 	for (i = 0; i < cfg->qcfg->num_queues;  i++)
338 		gve_rx_free_ring_dqo(priv, &rx[i], cfg);
339 
340 	kvfree(rx);
341 	cfg->rx = NULL;
342 }
343 
gve_rx_post_buffers_dqo(struct gve_rx_ring * rx)344 void gve_rx_post_buffers_dqo(struct gve_rx_ring *rx)
345 {
346 	struct gve_rx_compl_queue_dqo *complq = &rx->dqo.complq;
347 	struct gve_rx_buf_queue_dqo *bufq = &rx->dqo.bufq;
348 	struct gve_priv *priv = rx->gve;
349 	u32 num_avail_slots;
350 	u32 num_full_slots;
351 	u32 num_posted = 0;
352 
353 	num_full_slots = (bufq->tail - bufq->head) & bufq->mask;
354 	num_avail_slots = bufq->mask - num_full_slots;
355 
356 	num_avail_slots = min_t(u32, num_avail_slots, complq->num_free_slots);
357 	while (num_posted < num_avail_slots) {
358 		struct gve_rx_desc_dqo *desc = &bufq->desc_ring[bufq->tail];
359 
360 		if (unlikely(gve_alloc_buffer(rx, desc))) {
361 			u64_stats_update_begin(&rx->statss);
362 			rx->rx_buf_alloc_fail++;
363 			u64_stats_update_end(&rx->statss);
364 			break;
365 		}
366 
367 		if (rx->dqo.hdr_bufs.data)
368 			desc->header_buf_addr =
369 				cpu_to_le64(rx->dqo.hdr_bufs.addr +
370 					    priv->header_buf_size * bufq->tail);
371 
372 		bufq->tail = (bufq->tail + 1) & bufq->mask;
373 		complq->num_free_slots--;
374 		num_posted++;
375 
376 		if ((bufq->tail & (GVE_RX_BUF_THRESH_DQO - 1)) == 0)
377 			gve_rx_write_doorbell_dqo(priv, rx->q_num);
378 	}
379 
380 	rx->fill_cnt += num_posted;
381 }
382 
gve_rx_skb_csum(struct sk_buff * skb,const struct gve_rx_compl_desc_dqo * desc,struct gve_ptype ptype)383 static void gve_rx_skb_csum(struct sk_buff *skb,
384 			    const struct gve_rx_compl_desc_dqo *desc,
385 			    struct gve_ptype ptype)
386 {
387 	skb->ip_summed = CHECKSUM_NONE;
388 
389 	/* HW did not identify and process L3 and L4 headers. */
390 	if (unlikely(!desc->l3_l4_processed))
391 		return;
392 
393 	if (ptype.l3_type == GVE_L3_TYPE_IPV4) {
394 		if (unlikely(desc->csum_ip_err || desc->csum_external_ip_err))
395 			return;
396 	} else if (ptype.l3_type == GVE_L3_TYPE_IPV6) {
397 		/* Checksum should be skipped if this flag is set. */
398 		if (unlikely(desc->ipv6_ex_add))
399 			return;
400 	}
401 
402 	if (unlikely(desc->csum_l4_err))
403 		return;
404 
405 	switch (ptype.l4_type) {
406 	case GVE_L4_TYPE_TCP:
407 	case GVE_L4_TYPE_UDP:
408 	case GVE_L4_TYPE_ICMP:
409 	case GVE_L4_TYPE_SCTP:
410 		skb->ip_summed = CHECKSUM_UNNECESSARY;
411 		break;
412 	default:
413 		break;
414 	}
415 }
416 
gve_rx_skb_hash(struct sk_buff * skb,const struct gve_rx_compl_desc_dqo * compl_desc,struct gve_ptype ptype)417 static void gve_rx_skb_hash(struct sk_buff *skb,
418 			    const struct gve_rx_compl_desc_dqo *compl_desc,
419 			    struct gve_ptype ptype)
420 {
421 	enum pkt_hash_types hash_type = PKT_HASH_TYPE_L2;
422 
423 	if (ptype.l4_type != GVE_L4_TYPE_UNKNOWN)
424 		hash_type = PKT_HASH_TYPE_L4;
425 	else if (ptype.l3_type != GVE_L3_TYPE_UNKNOWN)
426 		hash_type = PKT_HASH_TYPE_L3;
427 
428 	skb_set_hash(skb, le32_to_cpu(compl_desc->hash), hash_type);
429 }
430 
gve_rx_free_skb(struct napi_struct * napi,struct gve_rx_ring * rx)431 static void gve_rx_free_skb(struct napi_struct *napi, struct gve_rx_ring *rx)
432 {
433 	if (!rx->ctx.skb_head)
434 		return;
435 
436 	if (rx->ctx.skb_head == napi->skb)
437 		napi->skb = NULL;
438 	dev_kfree_skb_any(rx->ctx.skb_head);
439 	rx->ctx.skb_head = NULL;
440 	rx->ctx.skb_tail = NULL;
441 }
442 
gve_rx_should_trigger_copy_ondemand(struct gve_rx_ring * rx)443 static bool gve_rx_should_trigger_copy_ondemand(struct gve_rx_ring *rx)
444 {
445 	if (!rx->dqo.qpl)
446 		return false;
447 	if (rx->dqo.used_buf_states_cnt <
448 		     (rx->dqo.num_buf_states -
449 		     GVE_DQO_QPL_ONDEMAND_ALLOC_THRESHOLD))
450 		return false;
451 	return true;
452 }
453 
gve_rx_copy_ondemand(struct gve_rx_ring * rx,struct gve_rx_buf_state_dqo * buf_state,u16 buf_len)454 static int gve_rx_copy_ondemand(struct gve_rx_ring *rx,
455 				struct gve_rx_buf_state_dqo *buf_state,
456 				u16 buf_len)
457 {
458 	struct page *page = alloc_page(GFP_ATOMIC);
459 	int num_frags;
460 
461 	if (!page)
462 		return -ENOMEM;
463 
464 	memcpy(page_address(page),
465 	       buf_state->page_info.page_address +
466 	       buf_state->page_info.page_offset,
467 	       buf_len);
468 	num_frags = skb_shinfo(rx->ctx.skb_tail)->nr_frags;
469 	skb_add_rx_frag(rx->ctx.skb_tail, num_frags, page,
470 			0, buf_len, PAGE_SIZE);
471 
472 	u64_stats_update_begin(&rx->statss);
473 	rx->rx_frag_alloc_cnt++;
474 	u64_stats_update_end(&rx->statss);
475 	/* Return unused buffer. */
476 	gve_enqueue_buf_state(rx, &rx->dqo.recycled_buf_states, buf_state);
477 	return 0;
478 }
479 
480 /* Chains multi skbs for single rx packet.
481  * Returns 0 if buffer is appended, -1 otherwise.
482  */
gve_rx_append_frags(struct napi_struct * napi,struct gve_rx_buf_state_dqo * buf_state,u16 buf_len,struct gve_rx_ring * rx,struct gve_priv * priv)483 static int gve_rx_append_frags(struct napi_struct *napi,
484 			       struct gve_rx_buf_state_dqo *buf_state,
485 			       u16 buf_len, struct gve_rx_ring *rx,
486 			       struct gve_priv *priv)
487 {
488 	int num_frags = skb_shinfo(rx->ctx.skb_tail)->nr_frags;
489 
490 	if (unlikely(num_frags == MAX_SKB_FRAGS)) {
491 		struct sk_buff *skb;
492 
493 		skb = napi_alloc_skb(napi, 0);
494 		if (!skb)
495 			return -1;
496 
497 		if (rx->dqo.page_pool)
498 			skb_mark_for_recycle(skb);
499 
500 		if (rx->ctx.skb_tail == rx->ctx.skb_head)
501 			skb_shinfo(rx->ctx.skb_head)->frag_list = skb;
502 		else
503 			rx->ctx.skb_tail->next = skb;
504 		rx->ctx.skb_tail = skb;
505 		num_frags = 0;
506 	}
507 	if (rx->ctx.skb_tail != rx->ctx.skb_head) {
508 		rx->ctx.skb_head->len += buf_len;
509 		rx->ctx.skb_head->data_len += buf_len;
510 		rx->ctx.skb_head->truesize += buf_state->page_info.buf_size;
511 	}
512 
513 	/* Trigger ondemand page allocation if we are running low on buffers */
514 	if (gve_rx_should_trigger_copy_ondemand(rx))
515 		return gve_rx_copy_ondemand(rx, buf_state, buf_len);
516 
517 	skb_add_rx_frag(rx->ctx.skb_tail, num_frags,
518 			buf_state->page_info.page,
519 			buf_state->page_info.page_offset,
520 			buf_len, buf_state->page_info.buf_size);
521 	gve_reuse_buffer(rx, buf_state);
522 	return 0;
523 }
524 
525 /* Returns 0 if descriptor is completed successfully.
526  * Returns -EINVAL if descriptor is invalid.
527  * Returns -ENOMEM if data cannot be copied to skb.
528  */
gve_rx_dqo(struct napi_struct * napi,struct gve_rx_ring * rx,const struct gve_rx_compl_desc_dqo * compl_desc,u32 desc_idx,int queue_idx)529 static int gve_rx_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
530 		      const struct gve_rx_compl_desc_dqo *compl_desc,
531 		      u32 desc_idx, int queue_idx)
532 {
533 	const u16 buffer_id = le16_to_cpu(compl_desc->buf_id);
534 	const bool hbo = compl_desc->header_buffer_overflow;
535 	const bool eop = compl_desc->end_of_packet != 0;
536 	const bool hsplit = compl_desc->split_header;
537 	struct gve_rx_buf_state_dqo *buf_state;
538 	struct gve_priv *priv = rx->gve;
539 	u16 buf_len;
540 	u16 hdr_len;
541 
542 	if (unlikely(buffer_id >= rx->dqo.num_buf_states)) {
543 		net_err_ratelimited("%s: Invalid RX buffer_id=%u\n",
544 				    priv->dev->name, buffer_id);
545 		return -EINVAL;
546 	}
547 	buf_state = &rx->dqo.buf_states[buffer_id];
548 	if (unlikely(!gve_buf_state_is_allocated(rx, buf_state))) {
549 		net_err_ratelimited("%s: RX buffer_id is not allocated: %u\n",
550 				    priv->dev->name, buffer_id);
551 		return -EINVAL;
552 	}
553 
554 	if (unlikely(compl_desc->rx_error)) {
555 		gve_free_buffer(rx, buf_state);
556 		return -EINVAL;
557 	}
558 
559 	buf_len = compl_desc->packet_len;
560 	hdr_len = compl_desc->header_len;
561 
562 	/* Page might have not been used for awhile and was likely last written
563 	 * by a different thread.
564 	 */
565 	prefetch(buf_state->page_info.page);
566 
567 	/* Copy the header into the skb in the case of header split */
568 	if (hsplit) {
569 		int unsplit = 0;
570 
571 		if (hdr_len && !hbo) {
572 			rx->ctx.skb_head = gve_rx_copy_data(priv->dev, napi,
573 							    rx->dqo.hdr_bufs.data +
574 							    desc_idx * priv->header_buf_size,
575 							    hdr_len);
576 			if (unlikely(!rx->ctx.skb_head))
577 				goto error;
578 			rx->ctx.skb_tail = rx->ctx.skb_head;
579 
580 			if (rx->dqo.page_pool)
581 				skb_mark_for_recycle(rx->ctx.skb_head);
582 		} else {
583 			unsplit = 1;
584 		}
585 		u64_stats_update_begin(&rx->statss);
586 		rx->rx_hsplit_pkt++;
587 		rx->rx_hsplit_unsplit_pkt += unsplit;
588 		rx->rx_hsplit_bytes += hdr_len;
589 		u64_stats_update_end(&rx->statss);
590 	}
591 
592 	/* Sync the portion of dma buffer for CPU to read. */
593 	dma_sync_single_range_for_cpu(&priv->pdev->dev, buf_state->addr,
594 				      buf_state->page_info.page_offset,
595 				      buf_len, DMA_FROM_DEVICE);
596 
597 	/* Append to current skb if one exists. */
598 	if (rx->ctx.skb_head) {
599 		if (unlikely(gve_rx_append_frags(napi, buf_state, buf_len, rx,
600 						 priv)) != 0) {
601 			goto error;
602 		}
603 		return 0;
604 	}
605 
606 	if (eop && buf_len <= priv->rx_copybreak) {
607 		rx->ctx.skb_head = gve_rx_copy(priv->dev, napi,
608 					       &buf_state->page_info, buf_len);
609 		if (unlikely(!rx->ctx.skb_head))
610 			goto error;
611 		rx->ctx.skb_tail = rx->ctx.skb_head;
612 
613 		u64_stats_update_begin(&rx->statss);
614 		rx->rx_copied_pkt++;
615 		rx->rx_copybreak_pkt++;
616 		u64_stats_update_end(&rx->statss);
617 
618 		gve_free_buffer(rx, buf_state);
619 		return 0;
620 	}
621 
622 	rx->ctx.skb_head = napi_get_frags(napi);
623 	if (unlikely(!rx->ctx.skb_head))
624 		goto error;
625 	rx->ctx.skb_tail = rx->ctx.skb_head;
626 
627 	if (gve_rx_should_trigger_copy_ondemand(rx)) {
628 		if (gve_rx_copy_ondemand(rx, buf_state, buf_len) < 0)
629 			goto error;
630 		return 0;
631 	}
632 
633 	if (rx->dqo.page_pool)
634 		skb_mark_for_recycle(rx->ctx.skb_head);
635 
636 	skb_add_rx_frag(rx->ctx.skb_head, 0, buf_state->page_info.page,
637 			buf_state->page_info.page_offset, buf_len,
638 			buf_state->page_info.buf_size);
639 	gve_reuse_buffer(rx, buf_state);
640 	return 0;
641 
642 error:
643 	gve_free_buffer(rx, buf_state);
644 	return -ENOMEM;
645 }
646 
gve_rx_complete_rsc(struct sk_buff * skb,const struct gve_rx_compl_desc_dqo * desc,struct gve_ptype ptype)647 static int gve_rx_complete_rsc(struct sk_buff *skb,
648 			       const struct gve_rx_compl_desc_dqo *desc,
649 			       struct gve_ptype ptype)
650 {
651 	struct skb_shared_info *shinfo = skb_shinfo(skb);
652 
653 	/* Only TCP is supported right now. */
654 	if (ptype.l4_type != GVE_L4_TYPE_TCP)
655 		return -EINVAL;
656 
657 	switch (ptype.l3_type) {
658 	case GVE_L3_TYPE_IPV4:
659 		shinfo->gso_type = SKB_GSO_TCPV4;
660 		break;
661 	case GVE_L3_TYPE_IPV6:
662 		shinfo->gso_type = SKB_GSO_TCPV6;
663 		break;
664 	default:
665 		return -EINVAL;
666 	}
667 
668 	shinfo->gso_size = le16_to_cpu(desc->rsc_seg_len);
669 	return 0;
670 }
671 
672 /* Returns 0 if skb is completed successfully, -1 otherwise. */
gve_rx_complete_skb(struct gve_rx_ring * rx,struct napi_struct * napi,const struct gve_rx_compl_desc_dqo * desc,netdev_features_t feat)673 static int gve_rx_complete_skb(struct gve_rx_ring *rx, struct napi_struct *napi,
674 			       const struct gve_rx_compl_desc_dqo *desc,
675 			       netdev_features_t feat)
676 {
677 	struct gve_ptype ptype =
678 		rx->gve->ptype_lut_dqo->ptypes[desc->packet_type];
679 	int err;
680 
681 	skb_record_rx_queue(rx->ctx.skb_head, rx->q_num);
682 
683 	if (feat & NETIF_F_RXHASH)
684 		gve_rx_skb_hash(rx->ctx.skb_head, desc, ptype);
685 
686 	if (feat & NETIF_F_RXCSUM)
687 		gve_rx_skb_csum(rx->ctx.skb_head, desc, ptype);
688 
689 	/* RSC packets must set gso_size otherwise the TCP stack will complain
690 	 * that packets are larger than MTU.
691 	 */
692 	if (desc->rsc) {
693 		err = gve_rx_complete_rsc(rx->ctx.skb_head, desc, ptype);
694 		if (err < 0)
695 			return err;
696 	}
697 
698 	if (skb_headlen(rx->ctx.skb_head) == 0)
699 		napi_gro_frags(napi);
700 	else
701 		napi_gro_receive(napi, rx->ctx.skb_head);
702 
703 	return 0;
704 }
705 
gve_rx_poll_dqo(struct gve_notify_block * block,int budget)706 int gve_rx_poll_dqo(struct gve_notify_block *block, int budget)
707 {
708 	struct napi_struct *napi = &block->napi;
709 	netdev_features_t feat = napi->dev->features;
710 
711 	struct gve_rx_ring *rx = block->rx;
712 	struct gve_rx_compl_queue_dqo *complq = &rx->dqo.complq;
713 
714 	u32 work_done = 0;
715 	u64 bytes = 0;
716 	int err;
717 
718 	while (work_done < budget) {
719 		struct gve_rx_compl_desc_dqo *compl_desc =
720 			&complq->desc_ring[complq->head];
721 		u32 pkt_bytes;
722 
723 		/* No more new packets */
724 		if (compl_desc->generation == complq->cur_gen_bit)
725 			break;
726 
727 		/* Prefetch the next two descriptors. */
728 		prefetch(&complq->desc_ring[(complq->head + 1) & complq->mask]);
729 		prefetch(&complq->desc_ring[(complq->head + 2) & complq->mask]);
730 
731 		/* Do not read data until we own the descriptor */
732 		dma_rmb();
733 
734 		err = gve_rx_dqo(napi, rx, compl_desc, complq->head, rx->q_num);
735 		if (err < 0) {
736 			gve_rx_free_skb(napi, rx);
737 			u64_stats_update_begin(&rx->statss);
738 			if (err == -ENOMEM)
739 				rx->rx_skb_alloc_fail++;
740 			else if (err == -EINVAL)
741 				rx->rx_desc_err_dropped_pkt++;
742 			u64_stats_update_end(&rx->statss);
743 		}
744 
745 		complq->head = (complq->head + 1) & complq->mask;
746 		complq->num_free_slots++;
747 
748 		/* When the ring wraps, the generation bit is flipped. */
749 		complq->cur_gen_bit ^= (complq->head == 0);
750 
751 		/* Receiving a completion means we have space to post another
752 		 * buffer on the buffer queue.
753 		 */
754 		{
755 			struct gve_rx_buf_queue_dqo *bufq = &rx->dqo.bufq;
756 
757 			bufq->head = (bufq->head + 1) & bufq->mask;
758 		}
759 
760 		/* Free running counter of completed descriptors */
761 		rx->cnt++;
762 
763 		if (!rx->ctx.skb_head)
764 			continue;
765 
766 		if (!compl_desc->end_of_packet)
767 			continue;
768 
769 		work_done++;
770 		pkt_bytes = rx->ctx.skb_head->len;
771 		/* The ethernet header (first ETH_HLEN bytes) is snipped off
772 		 * by eth_type_trans.
773 		 */
774 		if (skb_headlen(rx->ctx.skb_head))
775 			pkt_bytes += ETH_HLEN;
776 
777 		/* gve_rx_complete_skb() will consume skb if successful */
778 		if (gve_rx_complete_skb(rx, napi, compl_desc, feat) != 0) {
779 			gve_rx_free_skb(napi, rx);
780 			u64_stats_update_begin(&rx->statss);
781 			rx->rx_desc_err_dropped_pkt++;
782 			u64_stats_update_end(&rx->statss);
783 			continue;
784 		}
785 
786 		bytes += pkt_bytes;
787 		rx->ctx.skb_head = NULL;
788 		rx->ctx.skb_tail = NULL;
789 	}
790 
791 	gve_rx_post_buffers_dqo(rx);
792 
793 	u64_stats_update_begin(&rx->statss);
794 	rx->rpackets += work_done;
795 	rx->rbytes += bytes;
796 	u64_stats_update_end(&rx->statss);
797 
798 	return work_done;
799 }
800