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/ip.h>
11 #include <linux/tcp.h>
12 #include <linux/vmalloc.h>
13 #include <linux/skbuff.h>
14 
15 static inline void gve_tx_put_doorbell(struct gve_priv *priv,
16 				       struct gve_queue_resources *q_resources,
17 				       u32 val)
18 {
19 	iowrite32be(val, &priv->db_bar2[be32_to_cpu(q_resources->db_index)]);
20 }
21 
22 /* gvnic can only transmit from a Registered Segment.
23  * We copy skb payloads into the registered segment before writing Tx
24  * descriptors and ringing the Tx doorbell.
25  *
26  * gve_tx_fifo_* manages the Registered Segment as a FIFO - clients must
27  * free allocations in the order they were allocated.
28  */
29 
30 static int gve_tx_fifo_init(struct gve_priv *priv, struct gve_tx_fifo *fifo)
31 {
32 	fifo->base = vmap(fifo->qpl->pages, fifo->qpl->num_entries, VM_MAP,
33 			  PAGE_KERNEL);
34 	if (unlikely(!fifo->base)) {
35 		netif_err(priv, drv, priv->dev, "Failed to vmap fifo, qpl_id = %d\n",
36 			  fifo->qpl->id);
37 		return -ENOMEM;
38 	}
39 
40 	fifo->size = fifo->qpl->num_entries * PAGE_SIZE;
41 	atomic_set(&fifo->available, fifo->size);
42 	fifo->head = 0;
43 	return 0;
44 }
45 
46 static void gve_tx_fifo_release(struct gve_priv *priv, struct gve_tx_fifo *fifo)
47 {
48 	WARN(atomic_read(&fifo->available) != fifo->size,
49 	     "Releasing non-empty fifo");
50 
51 	vunmap(fifo->base);
52 }
53 
54 static int gve_tx_fifo_pad_alloc_one_frag(struct gve_tx_fifo *fifo,
55 					  size_t bytes)
56 {
57 	return (fifo->head + bytes < fifo->size) ? 0 : fifo->size - fifo->head;
58 }
59 
60 static bool gve_tx_fifo_can_alloc(struct gve_tx_fifo *fifo, size_t bytes)
61 {
62 	return (atomic_read(&fifo->available) <= bytes) ? false : true;
63 }
64 
65 /* gve_tx_alloc_fifo - Allocate fragment(s) from Tx FIFO
66  * @fifo: FIFO to allocate from
67  * @bytes: Allocation size
68  * @iov: Scatter-gather elements to fill with allocation fragment base/len
69  *
70  * Returns number of valid elements in iov[] or negative on error.
71  *
72  * Allocations from a given FIFO must be externally synchronized but concurrent
73  * allocation and frees are allowed.
74  */
75 static int gve_tx_alloc_fifo(struct gve_tx_fifo *fifo, size_t bytes,
76 			     struct gve_tx_iovec iov[2])
77 {
78 	size_t overflow, padding;
79 	u32 aligned_head;
80 	int nfrags = 0;
81 
82 	if (!bytes)
83 		return 0;
84 
85 	/* This check happens before we know how much padding is needed to
86 	 * align to a cacheline boundary for the payload, but that is fine,
87 	 * because the FIFO head always start aligned, and the FIFO's boundaries
88 	 * are aligned, so if there is space for the data, there is space for
89 	 * the padding to the next alignment.
90 	 */
91 	WARN(!gve_tx_fifo_can_alloc(fifo, bytes),
92 	     "Reached %s when there's not enough space in the fifo", __func__);
93 
94 	nfrags++;
95 
96 	iov[0].iov_offset = fifo->head;
97 	iov[0].iov_len = bytes;
98 	fifo->head += bytes;
99 
100 	if (fifo->head > fifo->size) {
101 		/* If the allocation did not fit in the tail fragment of the
102 		 * FIFO, also use the head fragment.
103 		 */
104 		nfrags++;
105 		overflow = fifo->head - fifo->size;
106 		iov[0].iov_len -= overflow;
107 		iov[1].iov_offset = 0;	/* Start of fifo*/
108 		iov[1].iov_len = overflow;
109 
110 		fifo->head = overflow;
111 	}
112 
113 	/* Re-align to a cacheline boundary */
114 	aligned_head = L1_CACHE_ALIGN(fifo->head);
115 	padding = aligned_head - fifo->head;
116 	iov[nfrags - 1].iov_padding = padding;
117 	atomic_sub(bytes + padding, &fifo->available);
118 	fifo->head = aligned_head;
119 
120 	if (fifo->head == fifo->size)
121 		fifo->head = 0;
122 
123 	return nfrags;
124 }
125 
126 /* gve_tx_free_fifo - Return space to Tx FIFO
127  * @fifo: FIFO to return fragments to
128  * @bytes: Bytes to free
129  */
130 static void gve_tx_free_fifo(struct gve_tx_fifo *fifo, size_t bytes)
131 {
132 	atomic_add(bytes, &fifo->available);
133 }
134 
135 static int gve_clean_tx_done(struct gve_priv *priv, struct gve_tx_ring *tx,
136 			     u32 to_do, bool try_to_wake);
137 
138 static void gve_tx_free_ring(struct gve_priv *priv, int idx)
139 {
140 	struct gve_tx_ring *tx = &priv->tx[idx];
141 	struct device *hdev = &priv->pdev->dev;
142 	size_t bytes;
143 	u32 slots;
144 
145 	gve_tx_remove_from_block(priv, idx);
146 	slots = tx->mask + 1;
147 	gve_clean_tx_done(priv, tx, priv->tx_desc_cnt, false);
148 	netdev_tx_reset_queue(tx->netdev_txq);
149 
150 	dma_free_coherent(hdev, sizeof(*tx->q_resources),
151 			  tx->q_resources, tx->q_resources_bus);
152 	tx->q_resources = NULL;
153 
154 	if (!tx->raw_addressing) {
155 		gve_tx_fifo_release(priv, &tx->tx_fifo);
156 		gve_unassign_qpl(priv, tx->tx_fifo.qpl->id);
157 		tx->tx_fifo.qpl = NULL;
158 	}
159 
160 	bytes = sizeof(*tx->desc) * slots;
161 	dma_free_coherent(hdev, bytes, tx->desc, tx->bus);
162 	tx->desc = NULL;
163 
164 	vfree(tx->info);
165 	tx->info = NULL;
166 
167 	netif_dbg(priv, drv, priv->dev, "freed tx queue %d\n", idx);
168 }
169 
170 static int gve_tx_alloc_ring(struct gve_priv *priv, int idx)
171 {
172 	struct gve_tx_ring *tx = &priv->tx[idx];
173 	struct device *hdev = &priv->pdev->dev;
174 	u32 slots = priv->tx_desc_cnt;
175 	size_t bytes;
176 
177 	/* Make sure everything is zeroed to start */
178 	memset(tx, 0, sizeof(*tx));
179 	spin_lock_init(&tx->clean_lock);
180 	tx->q_num = idx;
181 
182 	tx->mask = slots - 1;
183 
184 	/* alloc metadata */
185 	tx->info = vzalloc(sizeof(*tx->info) * slots);
186 	if (!tx->info)
187 		return -ENOMEM;
188 
189 	/* alloc tx queue */
190 	bytes = sizeof(*tx->desc) * slots;
191 	tx->desc = dma_alloc_coherent(hdev, bytes, &tx->bus, GFP_KERNEL);
192 	if (!tx->desc)
193 		goto abort_with_info;
194 
195 	tx->raw_addressing = priv->queue_format == GVE_GQI_RDA_FORMAT;
196 	tx->dev = &priv->pdev->dev;
197 	if (!tx->raw_addressing) {
198 		tx->tx_fifo.qpl = gve_assign_tx_qpl(priv);
199 		if (!tx->tx_fifo.qpl)
200 			goto abort_with_desc;
201 		/* map Tx FIFO */
202 		if (gve_tx_fifo_init(priv, &tx->tx_fifo))
203 			goto abort_with_qpl;
204 	}
205 
206 	tx->q_resources =
207 		dma_alloc_coherent(hdev,
208 				   sizeof(*tx->q_resources),
209 				   &tx->q_resources_bus,
210 				   GFP_KERNEL);
211 	if (!tx->q_resources)
212 		goto abort_with_fifo;
213 
214 	netif_dbg(priv, drv, priv->dev, "tx[%d]->bus=%lx\n", idx,
215 		  (unsigned long)tx->bus);
216 	tx->netdev_txq = netdev_get_tx_queue(priv->dev, idx);
217 	gve_tx_add_to_block(priv, idx);
218 
219 	return 0;
220 
221 abort_with_fifo:
222 	if (!tx->raw_addressing)
223 		gve_tx_fifo_release(priv, &tx->tx_fifo);
224 abort_with_qpl:
225 	if (!tx->raw_addressing)
226 		gve_unassign_qpl(priv, tx->tx_fifo.qpl->id);
227 abort_with_desc:
228 	dma_free_coherent(hdev, bytes, tx->desc, tx->bus);
229 	tx->desc = NULL;
230 abort_with_info:
231 	vfree(tx->info);
232 	tx->info = NULL;
233 	return -ENOMEM;
234 }
235 
236 int gve_tx_alloc_rings(struct gve_priv *priv)
237 {
238 	int err = 0;
239 	int i;
240 
241 	for (i = 0; i < priv->tx_cfg.num_queues; i++) {
242 		err = gve_tx_alloc_ring(priv, i);
243 		if (err) {
244 			netif_err(priv, drv, priv->dev,
245 				  "Failed to alloc tx ring=%d: err=%d\n",
246 				  i, err);
247 			break;
248 		}
249 	}
250 	/* Unallocate if there was an error */
251 	if (err) {
252 		int j;
253 
254 		for (j = 0; j < i; j++)
255 			gve_tx_free_ring(priv, j);
256 	}
257 	return err;
258 }
259 
260 void gve_tx_free_rings_gqi(struct gve_priv *priv)
261 {
262 	int i;
263 
264 	for (i = 0; i < priv->tx_cfg.num_queues; i++)
265 		gve_tx_free_ring(priv, i);
266 }
267 
268 /* gve_tx_avail - Calculates the number of slots available in the ring
269  * @tx: tx ring to check
270  *
271  * Returns the number of slots available
272  *
273  * The capacity of the queue is mask + 1. We don't need to reserve an entry.
274  **/
275 static inline u32 gve_tx_avail(struct gve_tx_ring *tx)
276 {
277 	return tx->mask + 1 - (tx->req - tx->done);
278 }
279 
280 static inline int gve_skb_fifo_bytes_required(struct gve_tx_ring *tx,
281 					      struct sk_buff *skb)
282 {
283 	int pad_bytes, align_hdr_pad;
284 	int bytes;
285 	int hlen;
286 
287 	hlen = skb_is_gso(skb) ? skb_checksum_start_offset(skb) +
288 				 tcp_hdrlen(skb) : skb_headlen(skb);
289 
290 	pad_bytes = gve_tx_fifo_pad_alloc_one_frag(&tx->tx_fifo,
291 						   hlen);
292 	/* We need to take into account the header alignment padding. */
293 	align_hdr_pad = L1_CACHE_ALIGN(hlen) - hlen;
294 	bytes = align_hdr_pad + pad_bytes + skb->len;
295 
296 	return bytes;
297 }
298 
299 /* The most descriptors we could need is MAX_SKB_FRAGS + 4 :
300  * 1 for each skb frag
301  * 1 for the skb linear portion
302  * 1 for when tcp hdr needs to be in separate descriptor
303  * 1 if the payload wraps to the beginning of the FIFO
304  * 1 for metadata descriptor
305  */
306 #define MAX_TX_DESC_NEEDED	(MAX_SKB_FRAGS + 4)
307 static void gve_tx_unmap_buf(struct device *dev, struct gve_tx_buffer_state *info)
308 {
309 	if (info->skb) {
310 		dma_unmap_single(dev, dma_unmap_addr(info, dma),
311 				 dma_unmap_len(info, len),
312 				 DMA_TO_DEVICE);
313 		dma_unmap_len_set(info, len, 0);
314 	} else {
315 		dma_unmap_page(dev, dma_unmap_addr(info, dma),
316 			       dma_unmap_len(info, len),
317 			       DMA_TO_DEVICE);
318 		dma_unmap_len_set(info, len, 0);
319 	}
320 }
321 
322 /* Check if sufficient resources (descriptor ring space, FIFO space) are
323  * available to transmit the given number of bytes.
324  */
325 static inline bool gve_can_tx(struct gve_tx_ring *tx, int bytes_required)
326 {
327 	bool can_alloc = true;
328 
329 	if (!tx->raw_addressing)
330 		can_alloc = gve_tx_fifo_can_alloc(&tx->tx_fifo, bytes_required);
331 
332 	return (gve_tx_avail(tx) >= MAX_TX_DESC_NEEDED && can_alloc);
333 }
334 
335 static_assert(NAPI_POLL_WEIGHT >= MAX_TX_DESC_NEEDED);
336 
337 /* Stops the queue if the skb cannot be transmitted. */
338 static int gve_maybe_stop_tx(struct gve_priv *priv, struct gve_tx_ring *tx,
339 			     struct sk_buff *skb)
340 {
341 	int bytes_required = 0;
342 	u32 nic_done;
343 	u32 to_do;
344 	int ret;
345 
346 	if (!tx->raw_addressing)
347 		bytes_required = gve_skb_fifo_bytes_required(tx, skb);
348 
349 	if (likely(gve_can_tx(tx, bytes_required)))
350 		return 0;
351 
352 	ret = -EBUSY;
353 	spin_lock(&tx->clean_lock);
354 	nic_done = gve_tx_load_event_counter(priv, tx);
355 	to_do = nic_done - tx->done;
356 
357 	/* Only try to clean if there is hope for TX */
358 	if (to_do + gve_tx_avail(tx) >= MAX_TX_DESC_NEEDED) {
359 		if (to_do > 0) {
360 			to_do = min_t(u32, to_do, NAPI_POLL_WEIGHT);
361 			gve_clean_tx_done(priv, tx, to_do, false);
362 		}
363 		if (likely(gve_can_tx(tx, bytes_required)))
364 			ret = 0;
365 	}
366 	if (ret) {
367 		/* No space, so stop the queue */
368 		tx->stop_queue++;
369 		netif_tx_stop_queue(tx->netdev_txq);
370 	}
371 	spin_unlock(&tx->clean_lock);
372 
373 	return ret;
374 }
375 
376 static void gve_tx_fill_pkt_desc(union gve_tx_desc *pkt_desc,
377 				 struct sk_buff *skb, bool is_gso,
378 				 int l4_hdr_offset, u32 desc_cnt,
379 				 u16 hlen, u64 addr)
380 {
381 	/* l4_hdr_offset and csum_offset are in units of 16-bit words */
382 	if (is_gso) {
383 		pkt_desc->pkt.type_flags = GVE_TXD_TSO | GVE_TXF_L4CSUM;
384 		pkt_desc->pkt.l4_csum_offset = skb->csum_offset >> 1;
385 		pkt_desc->pkt.l4_hdr_offset = l4_hdr_offset >> 1;
386 	} else if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
387 		pkt_desc->pkt.type_flags = GVE_TXD_STD | GVE_TXF_L4CSUM;
388 		pkt_desc->pkt.l4_csum_offset = skb->csum_offset >> 1;
389 		pkt_desc->pkt.l4_hdr_offset = l4_hdr_offset >> 1;
390 	} else {
391 		pkt_desc->pkt.type_flags = GVE_TXD_STD;
392 		pkt_desc->pkt.l4_csum_offset = 0;
393 		pkt_desc->pkt.l4_hdr_offset = 0;
394 	}
395 	pkt_desc->pkt.desc_cnt = desc_cnt;
396 	pkt_desc->pkt.len = cpu_to_be16(skb->len);
397 	pkt_desc->pkt.seg_len = cpu_to_be16(hlen);
398 	pkt_desc->pkt.seg_addr = cpu_to_be64(addr);
399 }
400 
401 static void gve_tx_fill_mtd_desc(union gve_tx_desc *mtd_desc,
402 				 struct sk_buff *skb)
403 {
404 	BUILD_BUG_ON(sizeof(mtd_desc->mtd) != sizeof(mtd_desc->pkt));
405 
406 	mtd_desc->mtd.type_flags = GVE_TXD_MTD | GVE_MTD_SUBTYPE_PATH;
407 	mtd_desc->mtd.path_state = GVE_MTD_PATH_STATE_DEFAULT |
408 				   GVE_MTD_PATH_HASH_L4;
409 	mtd_desc->mtd.path_hash = cpu_to_be32(skb->hash);
410 	mtd_desc->mtd.reserved0 = 0;
411 	mtd_desc->mtd.reserved1 = 0;
412 }
413 
414 static void gve_tx_fill_seg_desc(union gve_tx_desc *seg_desc,
415 				 struct sk_buff *skb, bool is_gso,
416 				 u16 len, u64 addr)
417 {
418 	seg_desc->seg.type_flags = GVE_TXD_SEG;
419 	if (is_gso) {
420 		if (skb_is_gso_v6(skb))
421 			seg_desc->seg.type_flags |= GVE_TXSF_IPV6;
422 		seg_desc->seg.l3_offset = skb_network_offset(skb) >> 1;
423 		seg_desc->seg.mss = cpu_to_be16(skb_shinfo(skb)->gso_size);
424 	}
425 	seg_desc->seg.seg_len = cpu_to_be16(len);
426 	seg_desc->seg.seg_addr = cpu_to_be64(addr);
427 }
428 
429 static void gve_dma_sync_for_device(struct device *dev, dma_addr_t *page_buses,
430 				    u64 iov_offset, u64 iov_len)
431 {
432 	u64 last_page = (iov_offset + iov_len - 1) / PAGE_SIZE;
433 	u64 first_page = iov_offset / PAGE_SIZE;
434 	u64 page;
435 
436 	for (page = first_page; page <= last_page; page++)
437 		dma_sync_single_for_device(dev, page_buses[page], PAGE_SIZE, DMA_TO_DEVICE);
438 }
439 
440 static int gve_tx_add_skb_copy(struct gve_priv *priv, struct gve_tx_ring *tx, struct sk_buff *skb)
441 {
442 	int pad_bytes, hlen, hdr_nfrags, payload_nfrags, l4_hdr_offset;
443 	union gve_tx_desc *pkt_desc, *seg_desc;
444 	struct gve_tx_buffer_state *info;
445 	int mtd_desc_nr = !!skb->l4_hash;
446 	bool is_gso = skb_is_gso(skb);
447 	u32 idx = tx->req & tx->mask;
448 	int payload_iov = 2;
449 	int copy_offset;
450 	u32 next_idx;
451 	int i;
452 
453 	info = &tx->info[idx];
454 	pkt_desc = &tx->desc[idx];
455 
456 	l4_hdr_offset = skb_checksum_start_offset(skb);
457 	/* If the skb is gso, then we want the tcp header in the first segment
458 	 * otherwise we want the linear portion of the skb (which will contain
459 	 * the checksum because skb->csum_start and skb->csum_offset are given
460 	 * relative to skb->head) in the first segment.
461 	 */
462 	hlen = is_gso ? l4_hdr_offset + tcp_hdrlen(skb) :
463 			skb_headlen(skb);
464 
465 	info->skb =  skb;
466 	/* We don't want to split the header, so if necessary, pad to the end
467 	 * of the fifo and then put the header at the beginning of the fifo.
468 	 */
469 	pad_bytes = gve_tx_fifo_pad_alloc_one_frag(&tx->tx_fifo, hlen);
470 	hdr_nfrags = gve_tx_alloc_fifo(&tx->tx_fifo, hlen + pad_bytes,
471 				       &info->iov[0]);
472 	WARN(!hdr_nfrags, "hdr_nfrags should never be 0!");
473 	payload_nfrags = gve_tx_alloc_fifo(&tx->tx_fifo, skb->len - hlen,
474 					   &info->iov[payload_iov]);
475 
476 	gve_tx_fill_pkt_desc(pkt_desc, skb, is_gso, l4_hdr_offset,
477 			     1 + mtd_desc_nr + payload_nfrags, hlen,
478 			     info->iov[hdr_nfrags - 1].iov_offset);
479 
480 	skb_copy_bits(skb, 0,
481 		      tx->tx_fifo.base + info->iov[hdr_nfrags - 1].iov_offset,
482 		      hlen);
483 	gve_dma_sync_for_device(&priv->pdev->dev, tx->tx_fifo.qpl->page_buses,
484 				info->iov[hdr_nfrags - 1].iov_offset,
485 				info->iov[hdr_nfrags - 1].iov_len);
486 	copy_offset = hlen;
487 
488 	if (mtd_desc_nr) {
489 		next_idx = (tx->req + 1) & tx->mask;
490 		gve_tx_fill_mtd_desc(&tx->desc[next_idx], skb);
491 	}
492 
493 	for (i = payload_iov; i < payload_nfrags + payload_iov; i++) {
494 		next_idx = (tx->req + 1 + mtd_desc_nr + i - payload_iov) & tx->mask;
495 		seg_desc = &tx->desc[next_idx];
496 
497 		gve_tx_fill_seg_desc(seg_desc, skb, is_gso,
498 				     info->iov[i].iov_len,
499 				     info->iov[i].iov_offset);
500 
501 		skb_copy_bits(skb, copy_offset,
502 			      tx->tx_fifo.base + info->iov[i].iov_offset,
503 			      info->iov[i].iov_len);
504 		gve_dma_sync_for_device(&priv->pdev->dev, tx->tx_fifo.qpl->page_buses,
505 					info->iov[i].iov_offset,
506 					info->iov[i].iov_len);
507 		copy_offset += info->iov[i].iov_len;
508 	}
509 
510 	return 1 + mtd_desc_nr + payload_nfrags;
511 }
512 
513 static int gve_tx_add_skb_no_copy(struct gve_priv *priv, struct gve_tx_ring *tx,
514 				  struct sk_buff *skb)
515 {
516 	const struct skb_shared_info *shinfo = skb_shinfo(skb);
517 	int hlen, num_descriptors, l4_hdr_offset;
518 	union gve_tx_desc *pkt_desc, *mtd_desc, *seg_desc;
519 	struct gve_tx_buffer_state *info;
520 	int mtd_desc_nr = !!skb->l4_hash;
521 	bool is_gso = skb_is_gso(skb);
522 	u32 idx = tx->req & tx->mask;
523 	u64 addr;
524 	u32 len;
525 	int i;
526 
527 	info = &tx->info[idx];
528 	pkt_desc = &tx->desc[idx];
529 
530 	l4_hdr_offset = skb_checksum_start_offset(skb);
531 	/* If the skb is gso, then we want only up to the tcp header in the first segment
532 	 * to efficiently replicate on each segment otherwise we want the linear portion
533 	 * of the skb (which will contain the checksum because skb->csum_start and
534 	 * skb->csum_offset are given relative to skb->head) in the first segment.
535 	 */
536 	hlen = is_gso ? l4_hdr_offset + tcp_hdrlen(skb) : skb_headlen(skb);
537 	len = skb_headlen(skb);
538 
539 	info->skb =  skb;
540 
541 	addr = dma_map_single(tx->dev, skb->data, len, DMA_TO_DEVICE);
542 	if (unlikely(dma_mapping_error(tx->dev, addr))) {
543 		tx->dma_mapping_error++;
544 		goto drop;
545 	}
546 	dma_unmap_len_set(info, len, len);
547 	dma_unmap_addr_set(info, dma, addr);
548 
549 	num_descriptors = 1 + shinfo->nr_frags;
550 	if (hlen < len)
551 		num_descriptors++;
552 	if (mtd_desc_nr)
553 		num_descriptors++;
554 
555 	gve_tx_fill_pkt_desc(pkt_desc, skb, is_gso, l4_hdr_offset,
556 			     num_descriptors, hlen, addr);
557 
558 	if (mtd_desc_nr) {
559 		idx = (idx + 1) & tx->mask;
560 		mtd_desc = &tx->desc[idx];
561 		gve_tx_fill_mtd_desc(mtd_desc, skb);
562 	}
563 
564 	if (hlen < len) {
565 		/* For gso the rest of the linear portion of the skb needs to
566 		 * be in its own descriptor.
567 		 */
568 		len -= hlen;
569 		addr += hlen;
570 		idx = (idx + 1) & tx->mask;
571 		seg_desc = &tx->desc[idx];
572 		gve_tx_fill_seg_desc(seg_desc, skb, is_gso, len, addr);
573 	}
574 
575 	for (i = 0; i < shinfo->nr_frags; i++) {
576 		const skb_frag_t *frag = &shinfo->frags[i];
577 
578 		idx = (idx + 1) & tx->mask;
579 		seg_desc = &tx->desc[idx];
580 		len = skb_frag_size(frag);
581 		addr = skb_frag_dma_map(tx->dev, frag, 0, len, DMA_TO_DEVICE);
582 		if (unlikely(dma_mapping_error(tx->dev, addr))) {
583 			tx->dma_mapping_error++;
584 			goto unmap_drop;
585 		}
586 		tx->info[idx].skb = NULL;
587 		dma_unmap_len_set(&tx->info[idx], len, len);
588 		dma_unmap_addr_set(&tx->info[idx], dma, addr);
589 
590 		gve_tx_fill_seg_desc(seg_desc, skb, is_gso, len, addr);
591 	}
592 
593 	return num_descriptors;
594 
595 unmap_drop:
596 	i += num_descriptors - shinfo->nr_frags;
597 	while (i--) {
598 		/* Skip metadata descriptor, if set */
599 		if (i == 1 && mtd_desc_nr == 1)
600 			continue;
601 		idx--;
602 		gve_tx_unmap_buf(tx->dev, &tx->info[idx & tx->mask]);
603 	}
604 drop:
605 	tx->dropped_pkt++;
606 	return 0;
607 }
608 
609 netdev_tx_t gve_tx(struct sk_buff *skb, struct net_device *dev)
610 {
611 	struct gve_priv *priv = netdev_priv(dev);
612 	struct gve_tx_ring *tx;
613 	int nsegs;
614 
615 	WARN(skb_get_queue_mapping(skb) >= priv->tx_cfg.num_queues,
616 	     "skb queue index out of range");
617 	tx = &priv->tx[skb_get_queue_mapping(skb)];
618 	if (unlikely(gve_maybe_stop_tx(priv, tx, skb))) {
619 		/* We need to ring the txq doorbell -- we have stopped the Tx
620 		 * queue for want of resources, but prior calls to gve_tx()
621 		 * may have added descriptors without ringing the doorbell.
622 		 */
623 
624 		gve_tx_put_doorbell(priv, tx->q_resources, tx->req);
625 		return NETDEV_TX_BUSY;
626 	}
627 	if (tx->raw_addressing)
628 		nsegs = gve_tx_add_skb_no_copy(priv, tx, skb);
629 	else
630 		nsegs = gve_tx_add_skb_copy(priv, tx, skb);
631 
632 	/* If the packet is getting sent, we need to update the skb */
633 	if (nsegs) {
634 		netdev_tx_sent_queue(tx->netdev_txq, skb->len);
635 		skb_tx_timestamp(skb);
636 		tx->req += nsegs;
637 	} else {
638 		dev_kfree_skb_any(skb);
639 	}
640 
641 	if (!netif_xmit_stopped(tx->netdev_txq) && netdev_xmit_more())
642 		return NETDEV_TX_OK;
643 
644 	/* Give packets to NIC. Even if this packet failed to send the doorbell
645 	 * might need to be rung because of xmit_more.
646 	 */
647 	gve_tx_put_doorbell(priv, tx->q_resources, tx->req);
648 	return NETDEV_TX_OK;
649 }
650 
651 #define GVE_TX_START_THRESH	PAGE_SIZE
652 
653 static int gve_clean_tx_done(struct gve_priv *priv, struct gve_tx_ring *tx,
654 			     u32 to_do, bool try_to_wake)
655 {
656 	struct gve_tx_buffer_state *info;
657 	u64 pkts = 0, bytes = 0;
658 	size_t space_freed = 0;
659 	struct sk_buff *skb;
660 	int i, j;
661 	u32 idx;
662 
663 	for (j = 0; j < to_do; j++) {
664 		idx = tx->done & tx->mask;
665 		netif_info(priv, tx_done, priv->dev,
666 			   "[%d] %s: idx=%d (req=%u done=%u)\n",
667 			   tx->q_num, __func__, idx, tx->req, tx->done);
668 		info = &tx->info[idx];
669 		skb = info->skb;
670 
671 		/* Unmap the buffer */
672 		if (tx->raw_addressing)
673 			gve_tx_unmap_buf(tx->dev, info);
674 		tx->done++;
675 		/* Mark as free */
676 		if (skb) {
677 			info->skb = NULL;
678 			bytes += skb->len;
679 			pkts++;
680 			dev_consume_skb_any(skb);
681 			if (tx->raw_addressing)
682 				continue;
683 			/* FIFO free */
684 			for (i = 0; i < ARRAY_SIZE(info->iov); i++) {
685 				space_freed += info->iov[i].iov_len + info->iov[i].iov_padding;
686 				info->iov[i].iov_len = 0;
687 				info->iov[i].iov_padding = 0;
688 			}
689 		}
690 	}
691 
692 	if (!tx->raw_addressing)
693 		gve_tx_free_fifo(&tx->tx_fifo, space_freed);
694 	u64_stats_update_begin(&tx->statss);
695 	tx->bytes_done += bytes;
696 	tx->pkt_done += pkts;
697 	u64_stats_update_end(&tx->statss);
698 	netdev_tx_completed_queue(tx->netdev_txq, pkts, bytes);
699 
700 	/* start the queue if we've stopped it */
701 #ifndef CONFIG_BQL
702 	/* Make sure that the doorbells are synced */
703 	smp_mb();
704 #endif
705 	if (try_to_wake && netif_tx_queue_stopped(tx->netdev_txq) &&
706 	    likely(gve_can_tx(tx, GVE_TX_START_THRESH))) {
707 		tx->wake_queue++;
708 		netif_tx_wake_queue(tx->netdev_txq);
709 	}
710 
711 	return pkts;
712 }
713 
714 u32 gve_tx_load_event_counter(struct gve_priv *priv,
715 			      struct gve_tx_ring *tx)
716 {
717 	u32 counter_index = be32_to_cpu(tx->q_resources->counter_index);
718 	__be32 counter = READ_ONCE(priv->counter_array[counter_index]);
719 
720 	return be32_to_cpu(counter);
721 }
722 
723 bool gve_tx_poll(struct gve_notify_block *block, int budget)
724 {
725 	struct gve_priv *priv = block->priv;
726 	struct gve_tx_ring *tx = block->tx;
727 	u32 nic_done;
728 	u32 to_do;
729 
730 	/* If budget is 0, do all the work */
731 	if (budget == 0)
732 		budget = INT_MAX;
733 
734 	/* In TX path, it may try to clean completed pkts in order to xmit,
735 	 * to avoid cleaning conflict, use spin_lock(), it yields better
736 	 * concurrency between xmit/clean than netif's lock.
737 	 */
738 	spin_lock(&tx->clean_lock);
739 	/* Find out how much work there is to be done */
740 	nic_done = gve_tx_load_event_counter(priv, tx);
741 	to_do = min_t(u32, (nic_done - tx->done), budget);
742 	gve_clean_tx_done(priv, tx, to_do, true);
743 	spin_unlock(&tx->clean_lock);
744 	/* If we still have work we want to repoll */
745 	return nic_done != tx->done;
746 }
747 
748 bool gve_tx_clean_pending(struct gve_priv *priv, struct gve_tx_ring *tx)
749 {
750 	u32 nic_done = gve_tx_load_event_counter(priv, tx);
751 
752 	return nic_done != tx->done;
753 }
754