1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2018 Intel Corporation. */
3 
4 #include <linux/bpf_trace.h>
5 #include <net/xdp_sock.h>
6 #include <net/xdp.h>
7 
8 #include "i40e.h"
9 #include "i40e_txrx_common.h"
10 #include "i40e_xsk.h"
11 
12 /**
13  * i40e_xsk_umem_dma_map - DMA maps all UMEM memory for the netdev
14  * @vsi: Current VSI
15  * @umem: UMEM to DMA map
16  *
17  * Returns 0 on success, <0 on failure
18  **/
19 static int i40e_xsk_umem_dma_map(struct i40e_vsi *vsi, struct xdp_umem *umem)
20 {
21 	struct i40e_pf *pf = vsi->back;
22 	struct device *dev;
23 	unsigned int i, j;
24 	dma_addr_t dma;
25 
26 	dev = &pf->pdev->dev;
27 	for (i = 0; i < umem->npgs; i++) {
28 		dma = dma_map_page_attrs(dev, umem->pgs[i], 0, PAGE_SIZE,
29 					 DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
30 		if (dma_mapping_error(dev, dma))
31 			goto out_unmap;
32 
33 		umem->pages[i].dma = dma;
34 	}
35 
36 	return 0;
37 
38 out_unmap:
39 	for (j = 0; j < i; j++) {
40 		dma_unmap_page_attrs(dev, umem->pages[i].dma, PAGE_SIZE,
41 				     DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
42 		umem->pages[i].dma = 0;
43 	}
44 
45 	return -1;
46 }
47 
48 /**
49  * i40e_xsk_umem_dma_unmap - DMA unmaps all UMEM memory for the netdev
50  * @vsi: Current VSI
51  * @umem: UMEM to DMA map
52  **/
53 static void i40e_xsk_umem_dma_unmap(struct i40e_vsi *vsi, struct xdp_umem *umem)
54 {
55 	struct i40e_pf *pf = vsi->back;
56 	struct device *dev;
57 	unsigned int i;
58 
59 	dev = &pf->pdev->dev;
60 
61 	for (i = 0; i < umem->npgs; i++) {
62 		dma_unmap_page_attrs(dev, umem->pages[i].dma, PAGE_SIZE,
63 				     DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
64 
65 		umem->pages[i].dma = 0;
66 	}
67 }
68 
69 /**
70  * i40e_xsk_umem_enable - Enable/associate a UMEM to a certain ring/qid
71  * @vsi: Current VSI
72  * @umem: UMEM
73  * @qid: Rx ring to associate UMEM to
74  *
75  * Returns 0 on success, <0 on failure
76  **/
77 static int i40e_xsk_umem_enable(struct i40e_vsi *vsi, struct xdp_umem *umem,
78 				u16 qid)
79 {
80 	struct net_device *netdev = vsi->netdev;
81 	struct xdp_umem_fq_reuse *reuseq;
82 	bool if_running;
83 	int err;
84 
85 	if (vsi->type != I40E_VSI_MAIN)
86 		return -EINVAL;
87 
88 	if (qid >= vsi->num_queue_pairs)
89 		return -EINVAL;
90 
91 	if (qid >= netdev->real_num_rx_queues ||
92 	    qid >= netdev->real_num_tx_queues)
93 		return -EINVAL;
94 
95 	reuseq = xsk_reuseq_prepare(vsi->rx_rings[0]->count);
96 	if (!reuseq)
97 		return -ENOMEM;
98 
99 	xsk_reuseq_free(xsk_reuseq_swap(umem, reuseq));
100 
101 	err = i40e_xsk_umem_dma_map(vsi, umem);
102 	if (err)
103 		return err;
104 
105 	set_bit(qid, vsi->af_xdp_zc_qps);
106 
107 	if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
108 
109 	if (if_running) {
110 		err = i40e_queue_pair_disable(vsi, qid);
111 		if (err)
112 			return err;
113 
114 		err = i40e_queue_pair_enable(vsi, qid);
115 		if (err)
116 			return err;
117 
118 		/* Kick start the NAPI context so that receiving will start */
119 		err = i40e_xsk_async_xmit(vsi->netdev, qid);
120 		if (err)
121 			return err;
122 	}
123 
124 	return 0;
125 }
126 
127 /**
128  * i40e_xsk_umem_disable - Disassociate a UMEM from a certain ring/qid
129  * @vsi: Current VSI
130  * @qid: Rx ring to associate UMEM to
131  *
132  * Returns 0 on success, <0 on failure
133  **/
134 static int i40e_xsk_umem_disable(struct i40e_vsi *vsi, u16 qid)
135 {
136 	struct net_device *netdev = vsi->netdev;
137 	struct xdp_umem *umem;
138 	bool if_running;
139 	int err;
140 
141 	umem = xdp_get_umem_from_qid(netdev, qid);
142 	if (!umem)
143 		return -EINVAL;
144 
145 	if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
146 
147 	if (if_running) {
148 		err = i40e_queue_pair_disable(vsi, qid);
149 		if (err)
150 			return err;
151 	}
152 
153 	clear_bit(qid, vsi->af_xdp_zc_qps);
154 	i40e_xsk_umem_dma_unmap(vsi, umem);
155 
156 	if (if_running) {
157 		err = i40e_queue_pair_enable(vsi, qid);
158 		if (err)
159 			return err;
160 	}
161 
162 	return 0;
163 }
164 
165 /**
166  * i40e_xsk_umem_setup - Enable/disassociate a UMEM to/from a ring/qid
167  * @vsi: Current VSI
168  * @umem: UMEM to enable/associate to a ring, or NULL to disable
169  * @qid: Rx ring to (dis)associate UMEM (from)to
170  *
171  * This function enables or disables a UMEM to a certain ring.
172  *
173  * Returns 0 on success, <0 on failure
174  **/
175 int i40e_xsk_umem_setup(struct i40e_vsi *vsi, struct xdp_umem *umem,
176 			u16 qid)
177 {
178 	return umem ? i40e_xsk_umem_enable(vsi, umem, qid) :
179 		i40e_xsk_umem_disable(vsi, qid);
180 }
181 
182 /**
183  * i40e_run_xdp_zc - Executes an XDP program on an xdp_buff
184  * @rx_ring: Rx ring
185  * @xdp: xdp_buff used as input to the XDP program
186  *
187  * This function enables or disables a UMEM to a certain ring.
188  *
189  * Returns any of I40E_XDP_{PASS, CONSUMED, TX, REDIR}
190  **/
191 static int i40e_run_xdp_zc(struct i40e_ring *rx_ring, struct xdp_buff *xdp)
192 {
193 	int err, result = I40E_XDP_PASS;
194 	struct i40e_ring *xdp_ring;
195 	struct bpf_prog *xdp_prog;
196 	u32 act;
197 
198 	rcu_read_lock();
199 	/* NB! xdp_prog will always be !NULL, due to the fact that
200 	 * this path is enabled by setting an XDP program.
201 	 */
202 	xdp_prog = READ_ONCE(rx_ring->xdp_prog);
203 	act = bpf_prog_run_xdp(xdp_prog, xdp);
204 	xdp->handle += xdp->data - xdp->data_hard_start;
205 	switch (act) {
206 	case XDP_PASS:
207 		break;
208 	case XDP_TX:
209 		xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->queue_index];
210 		result = i40e_xmit_xdp_tx_ring(xdp, xdp_ring);
211 		break;
212 	case XDP_REDIRECT:
213 		err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
214 		result = !err ? I40E_XDP_REDIR : I40E_XDP_CONSUMED;
215 		break;
216 	default:
217 		bpf_warn_invalid_xdp_action(act);
218 		/* fall through */
219 	case XDP_ABORTED:
220 		trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
221 		/* fallthrough -- handle aborts by dropping packet */
222 	case XDP_DROP:
223 		result = I40E_XDP_CONSUMED;
224 		break;
225 	}
226 	rcu_read_unlock();
227 	return result;
228 }
229 
230 /**
231  * i40e_alloc_buffer_zc - Allocates an i40e_rx_buffer
232  * @rx_ring: Rx ring
233  * @bi: Rx buffer to populate
234  *
235  * This function allocates an Rx buffer. The buffer can come from fill
236  * queue, or via the recycle queue (next_to_alloc).
237  *
238  * Returns true for a successful allocation, false otherwise
239  **/
240 static bool i40e_alloc_buffer_zc(struct i40e_ring *rx_ring,
241 				 struct i40e_rx_buffer *bi)
242 {
243 	struct xdp_umem *umem = rx_ring->xsk_umem;
244 	void *addr = bi->addr;
245 	u64 handle, hr;
246 
247 	if (addr) {
248 		rx_ring->rx_stats.page_reuse_count++;
249 		return true;
250 	}
251 
252 	if (!xsk_umem_peek_addr(umem, &handle)) {
253 		rx_ring->rx_stats.alloc_page_failed++;
254 		return false;
255 	}
256 
257 	hr = umem->headroom + XDP_PACKET_HEADROOM;
258 
259 	bi->dma = xdp_umem_get_dma(umem, handle);
260 	bi->dma += hr;
261 
262 	bi->addr = xdp_umem_get_data(umem, handle);
263 	bi->addr += hr;
264 
265 	bi->handle = handle + umem->headroom;
266 
267 	xsk_umem_discard_addr(umem);
268 	return true;
269 }
270 
271 /**
272  * i40e_alloc_buffer_slow_zc - Allocates an i40e_rx_buffer
273  * @rx_ring: Rx ring
274  * @bi: Rx buffer to populate
275  *
276  * This function allocates an Rx buffer. The buffer can come from fill
277  * queue, or via the reuse queue.
278  *
279  * Returns true for a successful allocation, false otherwise
280  **/
281 static bool i40e_alloc_buffer_slow_zc(struct i40e_ring *rx_ring,
282 				      struct i40e_rx_buffer *bi)
283 {
284 	struct xdp_umem *umem = rx_ring->xsk_umem;
285 	u64 handle, hr;
286 
287 	if (!xsk_umem_peek_addr_rq(umem, &handle)) {
288 		rx_ring->rx_stats.alloc_page_failed++;
289 		return false;
290 	}
291 
292 	handle &= rx_ring->xsk_umem->chunk_mask;
293 
294 	hr = umem->headroom + XDP_PACKET_HEADROOM;
295 
296 	bi->dma = xdp_umem_get_dma(umem, handle);
297 	bi->dma += hr;
298 
299 	bi->addr = xdp_umem_get_data(umem, handle);
300 	bi->addr += hr;
301 
302 	bi->handle = handle + umem->headroom;
303 
304 	xsk_umem_discard_addr_rq(umem);
305 	return true;
306 }
307 
308 static __always_inline bool
309 __i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count,
310 			   bool alloc(struct i40e_ring *rx_ring,
311 				      struct i40e_rx_buffer *bi))
312 {
313 	u16 ntu = rx_ring->next_to_use;
314 	union i40e_rx_desc *rx_desc;
315 	struct i40e_rx_buffer *bi;
316 	bool ok = true;
317 
318 	rx_desc = I40E_RX_DESC(rx_ring, ntu);
319 	bi = &rx_ring->rx_bi[ntu];
320 	do {
321 		if (!alloc(rx_ring, bi)) {
322 			ok = false;
323 			goto no_buffers;
324 		}
325 
326 		dma_sync_single_range_for_device(rx_ring->dev, bi->dma, 0,
327 						 rx_ring->rx_buf_len,
328 						 DMA_BIDIRECTIONAL);
329 
330 		rx_desc->read.pkt_addr = cpu_to_le64(bi->dma);
331 
332 		rx_desc++;
333 		bi++;
334 		ntu++;
335 
336 		if (unlikely(ntu == rx_ring->count)) {
337 			rx_desc = I40E_RX_DESC(rx_ring, 0);
338 			bi = rx_ring->rx_bi;
339 			ntu = 0;
340 		}
341 
342 		rx_desc->wb.qword1.status_error_len = 0;
343 		count--;
344 	} while (count);
345 
346 no_buffers:
347 	if (rx_ring->next_to_use != ntu)
348 		i40e_release_rx_desc(rx_ring, ntu);
349 
350 	return ok;
351 }
352 
353 /**
354  * i40e_alloc_rx_buffers_zc - Allocates a number of Rx buffers
355  * @rx_ring: Rx ring
356  * @count: The number of buffers to allocate
357  *
358  * This function allocates a number of Rx buffers from the reuse queue
359  * or fill ring and places them on the Rx ring.
360  *
361  * Returns true for a successful allocation, false otherwise
362  **/
363 bool i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count)
364 {
365 	return __i40e_alloc_rx_buffers_zc(rx_ring, count,
366 					  i40e_alloc_buffer_slow_zc);
367 }
368 
369 /**
370  * i40e_alloc_rx_buffers_fast_zc - Allocates a number of Rx buffers
371  * @rx_ring: Rx ring
372  * @count: The number of buffers to allocate
373  *
374  * This function allocates a number of Rx buffers from the fill ring
375  * or the internal recycle mechanism and places them on the Rx ring.
376  *
377  * Returns true for a successful allocation, false otherwise
378  **/
379 static bool i40e_alloc_rx_buffers_fast_zc(struct i40e_ring *rx_ring, u16 count)
380 {
381 	return __i40e_alloc_rx_buffers_zc(rx_ring, count,
382 					  i40e_alloc_buffer_zc);
383 }
384 
385 /**
386  * i40e_get_rx_buffer_zc - Return the current Rx buffer
387  * @rx_ring: Rx ring
388  * @size: The size of the rx buffer (read from descriptor)
389  *
390  * This function returns the current, received Rx buffer, and also
391  * does DMA synchronization.  the Rx ring.
392  *
393  * Returns the received Rx buffer
394  **/
395 static struct i40e_rx_buffer *i40e_get_rx_buffer_zc(struct i40e_ring *rx_ring,
396 						    const unsigned int size)
397 {
398 	struct i40e_rx_buffer *bi;
399 
400 	bi = &rx_ring->rx_bi[rx_ring->next_to_clean];
401 
402 	/* we are reusing so sync this buffer for CPU use */
403 	dma_sync_single_range_for_cpu(rx_ring->dev,
404 				      bi->dma, 0,
405 				      size,
406 				      DMA_BIDIRECTIONAL);
407 
408 	return bi;
409 }
410 
411 /**
412  * i40e_reuse_rx_buffer_zc - Recycle an Rx buffer
413  * @rx_ring: Rx ring
414  * @old_bi: The Rx buffer to recycle
415  *
416  * This function recycles a finished Rx buffer, and places it on the
417  * recycle queue (next_to_alloc).
418  **/
419 static void i40e_reuse_rx_buffer_zc(struct i40e_ring *rx_ring,
420 				    struct i40e_rx_buffer *old_bi)
421 {
422 	struct i40e_rx_buffer *new_bi = &rx_ring->rx_bi[rx_ring->next_to_alloc];
423 	unsigned long mask = (unsigned long)rx_ring->xsk_umem->chunk_mask;
424 	u64 hr = rx_ring->xsk_umem->headroom + XDP_PACKET_HEADROOM;
425 	u16 nta = rx_ring->next_to_alloc;
426 
427 	/* update, and store next to alloc */
428 	nta++;
429 	rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
430 
431 	/* transfer page from old buffer to new buffer */
432 	new_bi->dma = old_bi->dma & mask;
433 	new_bi->dma += hr;
434 
435 	new_bi->addr = (void *)((unsigned long)old_bi->addr & mask);
436 	new_bi->addr += hr;
437 
438 	new_bi->handle = old_bi->handle & mask;
439 	new_bi->handle += rx_ring->xsk_umem->headroom;
440 
441 	old_bi->addr = NULL;
442 }
443 
444 /**
445  * i40e_zca_free - Free callback for MEM_TYPE_ZERO_COPY allocations
446  * @alloc: Zero-copy allocator
447  * @handle: Buffer handle
448  **/
449 void i40e_zca_free(struct zero_copy_allocator *alloc, unsigned long handle)
450 {
451 	struct i40e_rx_buffer *bi;
452 	struct i40e_ring *rx_ring;
453 	u64 hr, mask;
454 	u16 nta;
455 
456 	rx_ring = container_of(alloc, struct i40e_ring, zca);
457 	hr = rx_ring->xsk_umem->headroom + XDP_PACKET_HEADROOM;
458 	mask = rx_ring->xsk_umem->chunk_mask;
459 
460 	nta = rx_ring->next_to_alloc;
461 	bi = &rx_ring->rx_bi[nta];
462 
463 	nta++;
464 	rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
465 
466 	handle &= mask;
467 
468 	bi->dma = xdp_umem_get_dma(rx_ring->xsk_umem, handle);
469 	bi->dma += hr;
470 
471 	bi->addr = xdp_umem_get_data(rx_ring->xsk_umem, handle);
472 	bi->addr += hr;
473 
474 	bi->handle = (u64)handle + rx_ring->xsk_umem->headroom;
475 }
476 
477 /**
478  * i40e_construct_skb_zc - Create skbufff from zero-copy Rx buffer
479  * @rx_ring: Rx ring
480  * @bi: Rx buffer
481  * @xdp: xdp_buff
482  *
483  * This functions allocates a new skb from a zero-copy Rx buffer.
484  *
485  * Returns the skb, or NULL on failure.
486  **/
487 static struct sk_buff *i40e_construct_skb_zc(struct i40e_ring *rx_ring,
488 					     struct i40e_rx_buffer *bi,
489 					     struct xdp_buff *xdp)
490 {
491 	unsigned int metasize = xdp->data - xdp->data_meta;
492 	unsigned int datasize = xdp->data_end - xdp->data;
493 	struct sk_buff *skb;
494 
495 	/* allocate a skb to store the frags */
496 	skb = __napi_alloc_skb(&rx_ring->q_vector->napi,
497 			       xdp->data_end - xdp->data_hard_start,
498 			       GFP_ATOMIC | __GFP_NOWARN);
499 	if (unlikely(!skb))
500 		return NULL;
501 
502 	skb_reserve(skb, xdp->data - xdp->data_hard_start);
503 	memcpy(__skb_put(skb, datasize), xdp->data, datasize);
504 	if (metasize)
505 		skb_metadata_set(skb, metasize);
506 
507 	i40e_reuse_rx_buffer_zc(rx_ring, bi);
508 	return skb;
509 }
510 
511 /**
512  * i40e_inc_ntc: Advance the next_to_clean index
513  * @rx_ring: Rx ring
514  **/
515 static void i40e_inc_ntc(struct i40e_ring *rx_ring)
516 {
517 	u32 ntc = rx_ring->next_to_clean + 1;
518 
519 	ntc = (ntc < rx_ring->count) ? ntc : 0;
520 	rx_ring->next_to_clean = ntc;
521 	prefetch(I40E_RX_DESC(rx_ring, ntc));
522 }
523 
524 /**
525  * i40e_clean_rx_irq_zc - Consumes Rx packets from the hardware ring
526  * @rx_ring: Rx ring
527  * @budget: NAPI budget
528  *
529  * Returns amount of work completed
530  **/
531 int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget)
532 {
533 	unsigned int total_rx_bytes = 0, total_rx_packets = 0;
534 	u16 cleaned_count = I40E_DESC_UNUSED(rx_ring);
535 	unsigned int xdp_res, xdp_xmit = 0;
536 	bool failure = false;
537 	struct sk_buff *skb;
538 	struct xdp_buff xdp;
539 
540 	xdp.rxq = &rx_ring->xdp_rxq;
541 
542 	while (likely(total_rx_packets < (unsigned int)budget)) {
543 		struct i40e_rx_buffer *bi;
544 		union i40e_rx_desc *rx_desc;
545 		unsigned int size;
546 		u64 qword;
547 
548 		if (cleaned_count >= I40E_RX_BUFFER_WRITE) {
549 			failure = failure ||
550 				  !i40e_alloc_rx_buffers_fast_zc(rx_ring,
551 								 cleaned_count);
552 			cleaned_count = 0;
553 		}
554 
555 		rx_desc = I40E_RX_DESC(rx_ring, rx_ring->next_to_clean);
556 		qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
557 
558 		/* This memory barrier is needed to keep us from reading
559 		 * any other fields out of the rx_desc until we have
560 		 * verified the descriptor has been written back.
561 		 */
562 		dma_rmb();
563 
564 		bi = i40e_clean_programming_status(rx_ring, rx_desc,
565 						   qword);
566 		if (unlikely(bi)) {
567 			i40e_reuse_rx_buffer_zc(rx_ring, bi);
568 			cleaned_count++;
569 			continue;
570 		}
571 
572 		size = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
573 		       I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
574 		if (!size)
575 			break;
576 
577 		bi = i40e_get_rx_buffer_zc(rx_ring, size);
578 		xdp.data = bi->addr;
579 		xdp.data_meta = xdp.data;
580 		xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
581 		xdp.data_end = xdp.data + size;
582 		xdp.handle = bi->handle;
583 
584 		xdp_res = i40e_run_xdp_zc(rx_ring, &xdp);
585 		if (xdp_res) {
586 			if (xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR)) {
587 				xdp_xmit |= xdp_res;
588 				bi->addr = NULL;
589 			} else {
590 				i40e_reuse_rx_buffer_zc(rx_ring, bi);
591 			}
592 
593 			total_rx_bytes += size;
594 			total_rx_packets++;
595 
596 			cleaned_count++;
597 			i40e_inc_ntc(rx_ring);
598 			continue;
599 		}
600 
601 		/* XDP_PASS path */
602 
603 		/* NB! We are not checking for errors using
604 		 * i40e_test_staterr with
605 		 * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that
606 		 * SBP is *not* set in PRT_SBPVSI (default not set).
607 		 */
608 		skb = i40e_construct_skb_zc(rx_ring, bi, &xdp);
609 		if (!skb) {
610 			rx_ring->rx_stats.alloc_buff_failed++;
611 			break;
612 		}
613 
614 		cleaned_count++;
615 		i40e_inc_ntc(rx_ring);
616 
617 		if (eth_skb_pad(skb))
618 			continue;
619 
620 		total_rx_bytes += skb->len;
621 		total_rx_packets++;
622 
623 		i40e_process_skb_fields(rx_ring, rx_desc, skb);
624 		napi_gro_receive(&rx_ring->q_vector->napi, skb);
625 	}
626 
627 	i40e_finalize_xdp_rx(rx_ring, xdp_xmit);
628 	i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets);
629 	return failure ? budget : (int)total_rx_packets;
630 }
631 
632 /**
633  * i40e_xmit_zc - Performs zero-copy Tx AF_XDP
634  * @xdp_ring: XDP Tx ring
635  * @budget: NAPI budget
636  *
637  * Returns true if the work is finished.
638  **/
639 static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
640 {
641 	struct i40e_tx_desc *tx_desc = NULL;
642 	struct i40e_tx_buffer *tx_bi;
643 	bool work_done = true;
644 	struct xdp_desc desc;
645 	dma_addr_t dma;
646 
647 	while (budget-- > 0) {
648 		if (!unlikely(I40E_DESC_UNUSED(xdp_ring))) {
649 			xdp_ring->tx_stats.tx_busy++;
650 			work_done = false;
651 			break;
652 		}
653 
654 		if (!xsk_umem_consume_tx(xdp_ring->xsk_umem, &desc))
655 			break;
656 
657 		dma = xdp_umem_get_dma(xdp_ring->xsk_umem, desc.addr);
658 
659 		dma_sync_single_for_device(xdp_ring->dev, dma, desc.len,
660 					   DMA_BIDIRECTIONAL);
661 
662 		tx_bi = &xdp_ring->tx_bi[xdp_ring->next_to_use];
663 		tx_bi->bytecount = desc.len;
664 
665 		tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use);
666 		tx_desc->buffer_addr = cpu_to_le64(dma);
667 		tx_desc->cmd_type_offset_bsz =
668 			build_ctob(I40E_TX_DESC_CMD_ICRC
669 				   | I40E_TX_DESC_CMD_EOP,
670 				   0, desc.len, 0);
671 
672 		xdp_ring->next_to_use++;
673 		if (xdp_ring->next_to_use == xdp_ring->count)
674 			xdp_ring->next_to_use = 0;
675 	}
676 
677 	if (tx_desc) {
678 		/* Request an interrupt for the last frame and bump tail ptr. */
679 		tx_desc->cmd_type_offset_bsz |= (I40E_TX_DESC_CMD_RS <<
680 						 I40E_TXD_QW1_CMD_SHIFT);
681 		i40e_xdp_ring_update_tail(xdp_ring);
682 
683 		xsk_umem_consume_tx_done(xdp_ring->xsk_umem);
684 	}
685 
686 	return !!budget && work_done;
687 }
688 
689 /**
690  * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry
691  * @tx_ring: XDP Tx ring
692  * @tx_bi: Tx buffer info to clean
693  **/
694 static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring,
695 				     struct i40e_tx_buffer *tx_bi)
696 {
697 	xdp_return_frame(tx_bi->xdpf);
698 	dma_unmap_single(tx_ring->dev,
699 			 dma_unmap_addr(tx_bi, dma),
700 			 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE);
701 	dma_unmap_len_set(tx_bi, len, 0);
702 }
703 
704 /**
705  * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries
706  * @tx_ring: XDP Tx ring
707  * @tx_bi: Tx buffer info to clean
708  *
709  * Returns true if cleanup/tranmission is done.
710  **/
711 bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi,
712 			   struct i40e_ring *tx_ring, int napi_budget)
713 {
714 	unsigned int ntc, total_bytes = 0, budget = vsi->work_limit;
715 	u32 i, completed_frames, frames_ready, xsk_frames = 0;
716 	struct xdp_umem *umem = tx_ring->xsk_umem;
717 	u32 head_idx = i40e_get_head(tx_ring);
718 	bool work_done = true, xmit_done;
719 	struct i40e_tx_buffer *tx_bi;
720 
721 	if (head_idx < tx_ring->next_to_clean)
722 		head_idx += tx_ring->count;
723 	frames_ready = head_idx - tx_ring->next_to_clean;
724 
725 	if (frames_ready == 0) {
726 		goto out_xmit;
727 	} else if (frames_ready > budget) {
728 		completed_frames = budget;
729 		work_done = false;
730 	} else {
731 		completed_frames = frames_ready;
732 	}
733 
734 	ntc = tx_ring->next_to_clean;
735 
736 	for (i = 0; i < completed_frames; i++) {
737 		tx_bi = &tx_ring->tx_bi[ntc];
738 
739 		if (tx_bi->xdpf)
740 			i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
741 		else
742 			xsk_frames++;
743 
744 		tx_bi->xdpf = NULL;
745 		total_bytes += tx_bi->bytecount;
746 
747 		if (++ntc >= tx_ring->count)
748 			ntc = 0;
749 	}
750 
751 	tx_ring->next_to_clean += completed_frames;
752 	if (unlikely(tx_ring->next_to_clean >= tx_ring->count))
753 		tx_ring->next_to_clean -= tx_ring->count;
754 
755 	if (xsk_frames)
756 		xsk_umem_complete_tx(umem, xsk_frames);
757 
758 	i40e_arm_wb(tx_ring, vsi, budget);
759 	i40e_update_tx_stats(tx_ring, completed_frames, total_bytes);
760 
761 out_xmit:
762 	xmit_done = i40e_xmit_zc(tx_ring, budget);
763 
764 	return work_done && xmit_done;
765 }
766 
767 /**
768  * i40e_xsk_async_xmit - Implements the ndo_xsk_async_xmit
769  * @dev: the netdevice
770  * @queue_id: queue id to wake up
771  *
772  * Returns <0 for errors, 0 otherwise.
773  **/
774 int i40e_xsk_async_xmit(struct net_device *dev, u32 queue_id)
775 {
776 	struct i40e_netdev_priv *np = netdev_priv(dev);
777 	struct i40e_vsi *vsi = np->vsi;
778 	struct i40e_ring *ring;
779 
780 	if (test_bit(__I40E_VSI_DOWN, vsi->state))
781 		return -ENETDOWN;
782 
783 	if (!i40e_enabled_xdp_vsi(vsi))
784 		return -ENXIO;
785 
786 	if (queue_id >= vsi->num_queue_pairs)
787 		return -ENXIO;
788 
789 	if (!vsi->xdp_rings[queue_id]->xsk_umem)
790 		return -ENXIO;
791 
792 	ring = vsi->xdp_rings[queue_id];
793 
794 	/* The idea here is that if NAPI is running, mark a miss, so
795 	 * it will run again. If not, trigger an interrupt and
796 	 * schedule the NAPI from interrupt context. If NAPI would be
797 	 * scheduled here, the interrupt affinity would not be
798 	 * honored.
799 	 */
800 	if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi))
801 		i40e_force_wb(vsi, ring->q_vector);
802 
803 	return 0;
804 }
805 
806 void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring)
807 {
808 	u16 i;
809 
810 	for (i = 0; i < rx_ring->count; i++) {
811 		struct i40e_rx_buffer *rx_bi = &rx_ring->rx_bi[i];
812 
813 		if (!rx_bi->addr)
814 			continue;
815 
816 		xsk_umem_fq_reuse(rx_ring->xsk_umem, rx_bi->handle);
817 		rx_bi->addr = NULL;
818 	}
819 }
820 
821 /**
822  * i40e_xsk_clean_xdp_ring - Clean the XDP Tx ring on shutdown
823  * @xdp_ring: XDP Tx ring
824  **/
825 void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring)
826 {
827 	u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
828 	struct xdp_umem *umem = tx_ring->xsk_umem;
829 	struct i40e_tx_buffer *tx_bi;
830 	u32 xsk_frames = 0;
831 
832 	while (ntc != ntu) {
833 		tx_bi = &tx_ring->tx_bi[ntc];
834 
835 		if (tx_bi->xdpf)
836 			i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
837 		else
838 			xsk_frames++;
839 
840 		tx_bi->xdpf = NULL;
841 
842 		ntc++;
843 		if (ntc >= tx_ring->count)
844 			ntc = 0;
845 	}
846 
847 	if (xsk_frames)
848 		xsk_umem_complete_tx(umem, xsk_frames);
849 }
850 
851 /**
852  * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have AF_XDP UMEM attached
853  * @vsi: vsi
854  *
855  * Returns true if any of the Rx rings has an AF_XDP UMEM attached
856  **/
857 bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi)
858 {
859 	struct net_device *netdev = vsi->netdev;
860 	int i;
861 
862 	for (i = 0; i < vsi->num_queue_pairs; i++) {
863 		if (xdp_get_umem_from_qid(netdev, i))
864 			return true;
865 	}
866 
867 	return false;
868 }
869