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_alloc_xsk_umems - Allocate an array to store per ring UMEMs
14  * @vsi: Current VSI
15  *
16  * Returns 0 on success, <0 on failure
17  **/
18 static int i40e_alloc_xsk_umems(struct i40e_vsi *vsi)
19 {
20 	if (vsi->xsk_umems)
21 		return 0;
22 
23 	vsi->num_xsk_umems_used = 0;
24 	vsi->num_xsk_umems = vsi->alloc_queue_pairs;
25 	vsi->xsk_umems = kcalloc(vsi->num_xsk_umems, sizeof(*vsi->xsk_umems),
26 				 GFP_KERNEL);
27 	if (!vsi->xsk_umems) {
28 		vsi->num_xsk_umems = 0;
29 		return -ENOMEM;
30 	}
31 
32 	return 0;
33 }
34 
35 /**
36  * i40e_add_xsk_umem - Store a UMEM for a certain ring/qid
37  * @vsi: Current VSI
38  * @umem: UMEM to store
39  * @qid: Ring/qid to associate with the UMEM
40  *
41  * Returns 0 on success, <0 on failure
42  **/
43 static int i40e_add_xsk_umem(struct i40e_vsi *vsi, struct xdp_umem *umem,
44 			     u16 qid)
45 {
46 	int err;
47 
48 	err = i40e_alloc_xsk_umems(vsi);
49 	if (err)
50 		return err;
51 
52 	vsi->xsk_umems[qid] = umem;
53 	vsi->num_xsk_umems_used++;
54 
55 	return 0;
56 }
57 
58 /**
59  * i40e_remove_xsk_umem - Remove a UMEM for a certain ring/qid
60  * @vsi: Current VSI
61  * @qid: Ring/qid associated with the UMEM
62  **/
63 static void i40e_remove_xsk_umem(struct i40e_vsi *vsi, u16 qid)
64 {
65 	vsi->xsk_umems[qid] = NULL;
66 	vsi->num_xsk_umems_used--;
67 
68 	if (vsi->num_xsk_umems == 0) {
69 		kfree(vsi->xsk_umems);
70 		vsi->xsk_umems = NULL;
71 		vsi->num_xsk_umems = 0;
72 	}
73 }
74 
75 /**
76  * i40e_xsk_umem_dma_map - DMA maps all UMEM memory for the netdev
77  * @vsi: Current VSI
78  * @umem: UMEM to DMA map
79  *
80  * Returns 0 on success, <0 on failure
81  **/
82 static int i40e_xsk_umem_dma_map(struct i40e_vsi *vsi, struct xdp_umem *umem)
83 {
84 	struct i40e_pf *pf = vsi->back;
85 	struct device *dev;
86 	unsigned int i, j;
87 	dma_addr_t dma;
88 
89 	dev = &pf->pdev->dev;
90 	for (i = 0; i < umem->npgs; i++) {
91 		dma = dma_map_page_attrs(dev, umem->pgs[i], 0, PAGE_SIZE,
92 					 DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
93 		if (dma_mapping_error(dev, dma))
94 			goto out_unmap;
95 
96 		umem->pages[i].dma = dma;
97 	}
98 
99 	return 0;
100 
101 out_unmap:
102 	for (j = 0; j < i; j++) {
103 		dma_unmap_page_attrs(dev, umem->pages[i].dma, PAGE_SIZE,
104 				     DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
105 		umem->pages[i].dma = 0;
106 	}
107 
108 	return -1;
109 }
110 
111 /**
112  * i40e_xsk_umem_dma_unmap - DMA unmaps all UMEM memory for the netdev
113  * @vsi: Current VSI
114  * @umem: UMEM to DMA map
115  **/
116 static void i40e_xsk_umem_dma_unmap(struct i40e_vsi *vsi, struct xdp_umem *umem)
117 {
118 	struct i40e_pf *pf = vsi->back;
119 	struct device *dev;
120 	unsigned int i;
121 
122 	dev = &pf->pdev->dev;
123 
124 	for (i = 0; i < umem->npgs; i++) {
125 		dma_unmap_page_attrs(dev, umem->pages[i].dma, PAGE_SIZE,
126 				     DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
127 
128 		umem->pages[i].dma = 0;
129 	}
130 }
131 
132 /**
133  * i40e_xsk_umem_enable - Enable/associate a UMEM to a certain ring/qid
134  * @vsi: Current VSI
135  * @umem: UMEM
136  * @qid: Rx ring to associate UMEM to
137  *
138  * Returns 0 on success, <0 on failure
139  **/
140 static int i40e_xsk_umem_enable(struct i40e_vsi *vsi, struct xdp_umem *umem,
141 				u16 qid)
142 {
143 	struct xdp_umem_fq_reuse *reuseq;
144 	bool if_running;
145 	int err;
146 
147 	if (vsi->type != I40E_VSI_MAIN)
148 		return -EINVAL;
149 
150 	if (qid >= vsi->num_queue_pairs)
151 		return -EINVAL;
152 
153 	if (vsi->xsk_umems) {
154 		if (qid >= vsi->num_xsk_umems)
155 			return -EINVAL;
156 		if (vsi->xsk_umems[qid])
157 			return -EBUSY;
158 	}
159 
160 	reuseq = xsk_reuseq_prepare(vsi->rx_rings[0]->count);
161 	if (!reuseq)
162 		return -ENOMEM;
163 
164 	xsk_reuseq_free(xsk_reuseq_swap(umem, reuseq));
165 
166 	err = i40e_xsk_umem_dma_map(vsi, umem);
167 	if (err)
168 		return err;
169 
170 	if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
171 
172 	if (if_running) {
173 		err = i40e_queue_pair_disable(vsi, qid);
174 		if (err)
175 			return err;
176 	}
177 
178 	err = i40e_add_xsk_umem(vsi, umem, qid);
179 	if (err)
180 		return err;
181 
182 	if (if_running) {
183 		err = i40e_queue_pair_enable(vsi, qid);
184 		if (err)
185 			return err;
186 
187 		/* Kick start the NAPI context so that receiving will start */
188 		err = i40e_xsk_async_xmit(vsi->netdev, qid);
189 		if (err)
190 			return err;
191 	}
192 
193 	return 0;
194 }
195 
196 /**
197  * i40e_xsk_umem_disable - Disassociate a UMEM from a certain ring/qid
198  * @vsi: Current VSI
199  * @qid: Rx ring to associate UMEM to
200  *
201  * Returns 0 on success, <0 on failure
202  **/
203 static int i40e_xsk_umem_disable(struct i40e_vsi *vsi, u16 qid)
204 {
205 	bool if_running;
206 	int err;
207 
208 	if (!vsi->xsk_umems || qid >= vsi->num_xsk_umems ||
209 	    !vsi->xsk_umems[qid])
210 		return -EINVAL;
211 
212 	if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
213 
214 	if (if_running) {
215 		err = i40e_queue_pair_disable(vsi, qid);
216 		if (err)
217 			return err;
218 	}
219 
220 	i40e_xsk_umem_dma_unmap(vsi, vsi->xsk_umems[qid]);
221 	i40e_remove_xsk_umem(vsi, qid);
222 
223 	if (if_running) {
224 		err = i40e_queue_pair_enable(vsi, qid);
225 		if (err)
226 			return err;
227 	}
228 
229 	return 0;
230 }
231 
232 /**
233  * i40e_xsk_umem_query - Queries a certain ring/qid for its UMEM
234  * @vsi: Current VSI
235  * @umem: UMEM associated to the ring, if any
236  * @qid: Rx ring to associate UMEM to
237  *
238  * This function will store, if any, the UMEM associated to certain ring.
239  *
240  * Returns 0 on success, <0 on failure
241  **/
242 int i40e_xsk_umem_query(struct i40e_vsi *vsi, struct xdp_umem **umem,
243 			u16 qid)
244 {
245 	if (vsi->type != I40E_VSI_MAIN)
246 		return -EINVAL;
247 
248 	if (qid >= vsi->num_queue_pairs)
249 		return -EINVAL;
250 
251 	if (vsi->xsk_umems) {
252 		if (qid >= vsi->num_xsk_umems)
253 			return -EINVAL;
254 		*umem = vsi->xsk_umems[qid];
255 		return 0;
256 	}
257 
258 	*umem = NULL;
259 	return 0;
260 }
261 
262 /**
263  * i40e_xsk_umem_setup - Enable/disassociate a UMEM to/from a ring/qid
264  * @vsi: Current VSI
265  * @umem: UMEM to enable/associate to a ring, or NULL to disable
266  * @qid: Rx ring to (dis)associate UMEM (from)to
267  *
268  * This function enables or disables a UMEM to a certain ring.
269  *
270  * Returns 0 on success, <0 on failure
271  **/
272 int i40e_xsk_umem_setup(struct i40e_vsi *vsi, struct xdp_umem *umem,
273 			u16 qid)
274 {
275 	return umem ? i40e_xsk_umem_enable(vsi, umem, qid) :
276 		i40e_xsk_umem_disable(vsi, qid);
277 }
278 
279 /**
280  * i40e_run_xdp_zc - Executes an XDP program on an xdp_buff
281  * @rx_ring: Rx ring
282  * @xdp: xdp_buff used as input to the XDP program
283  *
284  * This function enables or disables a UMEM to a certain ring.
285  *
286  * Returns any of I40E_XDP_{PASS, CONSUMED, TX, REDIR}
287  **/
288 static int i40e_run_xdp_zc(struct i40e_ring *rx_ring, struct xdp_buff *xdp)
289 {
290 	int err, result = I40E_XDP_PASS;
291 	struct i40e_ring *xdp_ring;
292 	struct bpf_prog *xdp_prog;
293 	u32 act;
294 
295 	rcu_read_lock();
296 	/* NB! xdp_prog will always be !NULL, due to the fact that
297 	 * this path is enabled by setting an XDP program.
298 	 */
299 	xdp_prog = READ_ONCE(rx_ring->xdp_prog);
300 	act = bpf_prog_run_xdp(xdp_prog, xdp);
301 	xdp->handle += xdp->data - xdp->data_hard_start;
302 	switch (act) {
303 	case XDP_PASS:
304 		break;
305 	case XDP_TX:
306 		xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->queue_index];
307 		result = i40e_xmit_xdp_tx_ring(xdp, xdp_ring);
308 		break;
309 	case XDP_REDIRECT:
310 		err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
311 		result = !err ? I40E_XDP_REDIR : I40E_XDP_CONSUMED;
312 		break;
313 	default:
314 		bpf_warn_invalid_xdp_action(act);
315 	case XDP_ABORTED:
316 		trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
317 		/* fallthrough -- handle aborts by dropping packet */
318 	case XDP_DROP:
319 		result = I40E_XDP_CONSUMED;
320 		break;
321 	}
322 	rcu_read_unlock();
323 	return result;
324 }
325 
326 /**
327  * i40e_alloc_buffer_zc - Allocates an i40e_rx_buffer
328  * @rx_ring: Rx ring
329  * @bi: Rx buffer to populate
330  *
331  * This function allocates an Rx buffer. The buffer can come from fill
332  * queue, or via the recycle queue (next_to_alloc).
333  *
334  * Returns true for a successful allocation, false otherwise
335  **/
336 static bool i40e_alloc_buffer_zc(struct i40e_ring *rx_ring,
337 				 struct i40e_rx_buffer *bi)
338 {
339 	struct xdp_umem *umem = rx_ring->xsk_umem;
340 	void *addr = bi->addr;
341 	u64 handle, hr;
342 
343 	if (addr) {
344 		rx_ring->rx_stats.page_reuse_count++;
345 		return true;
346 	}
347 
348 	if (!xsk_umem_peek_addr(umem, &handle)) {
349 		rx_ring->rx_stats.alloc_page_failed++;
350 		return false;
351 	}
352 
353 	hr = umem->headroom + XDP_PACKET_HEADROOM;
354 
355 	bi->dma = xdp_umem_get_dma(umem, handle);
356 	bi->dma += hr;
357 
358 	bi->addr = xdp_umem_get_data(umem, handle);
359 	bi->addr += hr;
360 
361 	bi->handle = handle + umem->headroom;
362 
363 	xsk_umem_discard_addr(umem);
364 	return true;
365 }
366 
367 /**
368  * i40e_alloc_buffer_slow_zc - Allocates an i40e_rx_buffer
369  * @rx_ring: Rx ring
370  * @bi: Rx buffer to populate
371  *
372  * This function allocates an Rx buffer. The buffer can come from fill
373  * queue, or via the reuse queue.
374  *
375  * Returns true for a successful allocation, false otherwise
376  **/
377 static bool i40e_alloc_buffer_slow_zc(struct i40e_ring *rx_ring,
378 				      struct i40e_rx_buffer *bi)
379 {
380 	struct xdp_umem *umem = rx_ring->xsk_umem;
381 	u64 handle, hr;
382 
383 	if (!xsk_umem_peek_addr_rq(umem, &handle)) {
384 		rx_ring->rx_stats.alloc_page_failed++;
385 		return false;
386 	}
387 
388 	handle &= rx_ring->xsk_umem->chunk_mask;
389 
390 	hr = umem->headroom + XDP_PACKET_HEADROOM;
391 
392 	bi->dma = xdp_umem_get_dma(umem, handle);
393 	bi->dma += hr;
394 
395 	bi->addr = xdp_umem_get_data(umem, handle);
396 	bi->addr += hr;
397 
398 	bi->handle = handle + umem->headroom;
399 
400 	xsk_umem_discard_addr_rq(umem);
401 	return true;
402 }
403 
404 static __always_inline bool
405 __i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count,
406 			   bool alloc(struct i40e_ring *rx_ring,
407 				      struct i40e_rx_buffer *bi))
408 {
409 	u16 ntu = rx_ring->next_to_use;
410 	union i40e_rx_desc *rx_desc;
411 	struct i40e_rx_buffer *bi;
412 	bool ok = true;
413 
414 	rx_desc = I40E_RX_DESC(rx_ring, ntu);
415 	bi = &rx_ring->rx_bi[ntu];
416 	do {
417 		if (!alloc(rx_ring, bi)) {
418 			ok = false;
419 			goto no_buffers;
420 		}
421 
422 		dma_sync_single_range_for_device(rx_ring->dev, bi->dma, 0,
423 						 rx_ring->rx_buf_len,
424 						 DMA_BIDIRECTIONAL);
425 
426 		rx_desc->read.pkt_addr = cpu_to_le64(bi->dma);
427 
428 		rx_desc++;
429 		bi++;
430 		ntu++;
431 
432 		if (unlikely(ntu == rx_ring->count)) {
433 			rx_desc = I40E_RX_DESC(rx_ring, 0);
434 			bi = rx_ring->rx_bi;
435 			ntu = 0;
436 		}
437 
438 		rx_desc->wb.qword1.status_error_len = 0;
439 		count--;
440 	} while (count);
441 
442 no_buffers:
443 	if (rx_ring->next_to_use != ntu)
444 		i40e_release_rx_desc(rx_ring, ntu);
445 
446 	return ok;
447 }
448 
449 /**
450  * i40e_alloc_rx_buffers_zc - Allocates a number of Rx buffers
451  * @rx_ring: Rx ring
452  * @count: The number of buffers to allocate
453  *
454  * This function allocates a number of Rx buffers from the reuse queue
455  * or fill ring and places them on the Rx ring.
456  *
457  * Returns true for a successful allocation, false otherwise
458  **/
459 bool i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count)
460 {
461 	return __i40e_alloc_rx_buffers_zc(rx_ring, count,
462 					  i40e_alloc_buffer_slow_zc);
463 }
464 
465 /**
466  * i40e_alloc_rx_buffers_fast_zc - Allocates a number of Rx buffers
467  * @rx_ring: Rx ring
468  * @count: The number of buffers to allocate
469  *
470  * This function allocates a number of Rx buffers from the fill ring
471  * or the internal recycle mechanism and places them on the Rx ring.
472  *
473  * Returns true for a successful allocation, false otherwise
474  **/
475 static bool i40e_alloc_rx_buffers_fast_zc(struct i40e_ring *rx_ring, u16 count)
476 {
477 	return __i40e_alloc_rx_buffers_zc(rx_ring, count,
478 					  i40e_alloc_buffer_zc);
479 }
480 
481 /**
482  * i40e_get_rx_buffer_zc - Return the current Rx buffer
483  * @rx_ring: Rx ring
484  * @size: The size of the rx buffer (read from descriptor)
485  *
486  * This function returns the current, received Rx buffer, and also
487  * does DMA synchronization.  the Rx ring.
488  *
489  * Returns the received Rx buffer
490  **/
491 static struct i40e_rx_buffer *i40e_get_rx_buffer_zc(struct i40e_ring *rx_ring,
492 						    const unsigned int size)
493 {
494 	struct i40e_rx_buffer *bi;
495 
496 	bi = &rx_ring->rx_bi[rx_ring->next_to_clean];
497 
498 	/* we are reusing so sync this buffer for CPU use */
499 	dma_sync_single_range_for_cpu(rx_ring->dev,
500 				      bi->dma, 0,
501 				      size,
502 				      DMA_BIDIRECTIONAL);
503 
504 	return bi;
505 }
506 
507 /**
508  * i40e_reuse_rx_buffer_zc - Recycle an Rx buffer
509  * @rx_ring: Rx ring
510  * @old_bi: The Rx buffer to recycle
511  *
512  * This function recycles a finished Rx buffer, and places it on the
513  * recycle queue (next_to_alloc).
514  **/
515 static void i40e_reuse_rx_buffer_zc(struct i40e_ring *rx_ring,
516 				    struct i40e_rx_buffer *old_bi)
517 {
518 	struct i40e_rx_buffer *new_bi = &rx_ring->rx_bi[rx_ring->next_to_alloc];
519 	unsigned long mask = (unsigned long)rx_ring->xsk_umem->chunk_mask;
520 	u64 hr = rx_ring->xsk_umem->headroom + XDP_PACKET_HEADROOM;
521 	u16 nta = rx_ring->next_to_alloc;
522 
523 	/* update, and store next to alloc */
524 	nta++;
525 	rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
526 
527 	/* transfer page from old buffer to new buffer */
528 	new_bi->dma = old_bi->dma & mask;
529 	new_bi->dma += hr;
530 
531 	new_bi->addr = (void *)((unsigned long)old_bi->addr & mask);
532 	new_bi->addr += hr;
533 
534 	new_bi->handle = old_bi->handle & mask;
535 	new_bi->handle += rx_ring->xsk_umem->headroom;
536 
537 	old_bi->addr = NULL;
538 }
539 
540 /**
541  * i40e_zca_free - Free callback for MEM_TYPE_ZERO_COPY allocations
542  * @alloc: Zero-copy allocator
543  * @handle: Buffer handle
544  **/
545 void i40e_zca_free(struct zero_copy_allocator *alloc, unsigned long handle)
546 {
547 	struct i40e_rx_buffer *bi;
548 	struct i40e_ring *rx_ring;
549 	u64 hr, mask;
550 	u16 nta;
551 
552 	rx_ring = container_of(alloc, struct i40e_ring, zca);
553 	hr = rx_ring->xsk_umem->headroom + XDP_PACKET_HEADROOM;
554 	mask = rx_ring->xsk_umem->chunk_mask;
555 
556 	nta = rx_ring->next_to_alloc;
557 	bi = &rx_ring->rx_bi[nta];
558 
559 	nta++;
560 	rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
561 
562 	handle &= mask;
563 
564 	bi->dma = xdp_umem_get_dma(rx_ring->xsk_umem, handle);
565 	bi->dma += hr;
566 
567 	bi->addr = xdp_umem_get_data(rx_ring->xsk_umem, handle);
568 	bi->addr += hr;
569 
570 	bi->handle = (u64)handle + rx_ring->xsk_umem->headroom;
571 }
572 
573 /**
574  * i40e_construct_skb_zc - Create skbufff from zero-copy Rx buffer
575  * @rx_ring: Rx ring
576  * @bi: Rx buffer
577  * @xdp: xdp_buff
578  *
579  * This functions allocates a new skb from a zero-copy Rx buffer.
580  *
581  * Returns the skb, or NULL on failure.
582  **/
583 static struct sk_buff *i40e_construct_skb_zc(struct i40e_ring *rx_ring,
584 					     struct i40e_rx_buffer *bi,
585 					     struct xdp_buff *xdp)
586 {
587 	unsigned int metasize = xdp->data - xdp->data_meta;
588 	unsigned int datasize = xdp->data_end - xdp->data;
589 	struct sk_buff *skb;
590 
591 	/* allocate a skb to store the frags */
592 	skb = __napi_alloc_skb(&rx_ring->q_vector->napi,
593 			       xdp->data_end - xdp->data_hard_start,
594 			       GFP_ATOMIC | __GFP_NOWARN);
595 	if (unlikely(!skb))
596 		return NULL;
597 
598 	skb_reserve(skb, xdp->data - xdp->data_hard_start);
599 	memcpy(__skb_put(skb, datasize), xdp->data, datasize);
600 	if (metasize)
601 		skb_metadata_set(skb, metasize);
602 
603 	i40e_reuse_rx_buffer_zc(rx_ring, bi);
604 	return skb;
605 }
606 
607 /**
608  * i40e_inc_ntc: Advance the next_to_clean index
609  * @rx_ring: Rx ring
610  **/
611 static void i40e_inc_ntc(struct i40e_ring *rx_ring)
612 {
613 	u32 ntc = rx_ring->next_to_clean + 1;
614 
615 	ntc = (ntc < rx_ring->count) ? ntc : 0;
616 	rx_ring->next_to_clean = ntc;
617 	prefetch(I40E_RX_DESC(rx_ring, ntc));
618 }
619 
620 /**
621  * i40e_clean_rx_irq_zc - Consumes Rx packets from the hardware ring
622  * @rx_ring: Rx ring
623  * @budget: NAPI budget
624  *
625  * Returns amount of work completed
626  **/
627 int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget)
628 {
629 	unsigned int total_rx_bytes = 0, total_rx_packets = 0;
630 	u16 cleaned_count = I40E_DESC_UNUSED(rx_ring);
631 	unsigned int xdp_res, xdp_xmit = 0;
632 	bool failure = false;
633 	struct sk_buff *skb;
634 	struct xdp_buff xdp;
635 
636 	xdp.rxq = &rx_ring->xdp_rxq;
637 
638 	while (likely(total_rx_packets < (unsigned int)budget)) {
639 		struct i40e_rx_buffer *bi;
640 		union i40e_rx_desc *rx_desc;
641 		unsigned int size;
642 		u64 qword;
643 
644 		if (cleaned_count >= I40E_RX_BUFFER_WRITE) {
645 			failure = failure ||
646 				  !i40e_alloc_rx_buffers_fast_zc(rx_ring,
647 								 cleaned_count);
648 			cleaned_count = 0;
649 		}
650 
651 		rx_desc = I40E_RX_DESC(rx_ring, rx_ring->next_to_clean);
652 		qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
653 
654 		/* This memory barrier is needed to keep us from reading
655 		 * any other fields out of the rx_desc until we have
656 		 * verified the descriptor has been written back.
657 		 */
658 		dma_rmb();
659 
660 		bi = i40e_clean_programming_status(rx_ring, rx_desc,
661 						   qword);
662 		if (unlikely(bi)) {
663 			i40e_reuse_rx_buffer_zc(rx_ring, bi);
664 			cleaned_count++;
665 			continue;
666 		}
667 
668 		size = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
669 		       I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
670 		if (!size)
671 			break;
672 
673 		bi = i40e_get_rx_buffer_zc(rx_ring, size);
674 		xdp.data = bi->addr;
675 		xdp.data_meta = xdp.data;
676 		xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
677 		xdp.data_end = xdp.data + size;
678 		xdp.handle = bi->handle;
679 
680 		xdp_res = i40e_run_xdp_zc(rx_ring, &xdp);
681 		if (xdp_res) {
682 			if (xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR)) {
683 				xdp_xmit |= xdp_res;
684 				bi->addr = NULL;
685 			} else {
686 				i40e_reuse_rx_buffer_zc(rx_ring, bi);
687 			}
688 
689 			total_rx_bytes += size;
690 			total_rx_packets++;
691 
692 			cleaned_count++;
693 			i40e_inc_ntc(rx_ring);
694 			continue;
695 		}
696 
697 		/* XDP_PASS path */
698 
699 		/* NB! We are not checking for errors using
700 		 * i40e_test_staterr with
701 		 * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that
702 		 * SBP is *not* set in PRT_SBPVSI (default not set).
703 		 */
704 		skb = i40e_construct_skb_zc(rx_ring, bi, &xdp);
705 		if (!skb) {
706 			rx_ring->rx_stats.alloc_buff_failed++;
707 			break;
708 		}
709 
710 		cleaned_count++;
711 		i40e_inc_ntc(rx_ring);
712 
713 		if (eth_skb_pad(skb))
714 			continue;
715 
716 		total_rx_bytes += skb->len;
717 		total_rx_packets++;
718 
719 		i40e_process_skb_fields(rx_ring, rx_desc, skb);
720 		napi_gro_receive(&rx_ring->q_vector->napi, skb);
721 	}
722 
723 	i40e_finalize_xdp_rx(rx_ring, xdp_xmit);
724 	i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets);
725 	return failure ? budget : (int)total_rx_packets;
726 }
727 
728 /**
729  * i40e_xmit_zc - Performs zero-copy Tx AF_XDP
730  * @xdp_ring: XDP Tx ring
731  * @budget: NAPI budget
732  *
733  * Returns true if the work is finished.
734  **/
735 static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
736 {
737 	struct i40e_tx_desc *tx_desc = NULL;
738 	struct i40e_tx_buffer *tx_bi;
739 	bool work_done = true;
740 	dma_addr_t dma;
741 	u32 len;
742 
743 	while (budget-- > 0) {
744 		if (!unlikely(I40E_DESC_UNUSED(xdp_ring))) {
745 			xdp_ring->tx_stats.tx_busy++;
746 			work_done = false;
747 			break;
748 		}
749 
750 		if (!xsk_umem_consume_tx(xdp_ring->xsk_umem, &dma, &len))
751 			break;
752 
753 		dma_sync_single_for_device(xdp_ring->dev, dma, len,
754 					   DMA_BIDIRECTIONAL);
755 
756 		tx_bi = &xdp_ring->tx_bi[xdp_ring->next_to_use];
757 		tx_bi->bytecount = len;
758 
759 		tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use);
760 		tx_desc->buffer_addr = cpu_to_le64(dma);
761 		tx_desc->cmd_type_offset_bsz =
762 			build_ctob(I40E_TX_DESC_CMD_ICRC
763 				   | I40E_TX_DESC_CMD_EOP,
764 				   0, len, 0);
765 
766 		xdp_ring->next_to_use++;
767 		if (xdp_ring->next_to_use == xdp_ring->count)
768 			xdp_ring->next_to_use = 0;
769 	}
770 
771 	if (tx_desc) {
772 		/* Request an interrupt for the last frame and bump tail ptr. */
773 		tx_desc->cmd_type_offset_bsz |= (I40E_TX_DESC_CMD_RS <<
774 						 I40E_TXD_QW1_CMD_SHIFT);
775 		i40e_xdp_ring_update_tail(xdp_ring);
776 
777 		xsk_umem_consume_tx_done(xdp_ring->xsk_umem);
778 	}
779 
780 	return !!budget && work_done;
781 }
782 
783 /**
784  * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry
785  * @tx_ring: XDP Tx ring
786  * @tx_bi: Tx buffer info to clean
787  **/
788 static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring,
789 				     struct i40e_tx_buffer *tx_bi)
790 {
791 	xdp_return_frame(tx_bi->xdpf);
792 	dma_unmap_single(tx_ring->dev,
793 			 dma_unmap_addr(tx_bi, dma),
794 			 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE);
795 	dma_unmap_len_set(tx_bi, len, 0);
796 }
797 
798 /**
799  * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries
800  * @tx_ring: XDP Tx ring
801  * @tx_bi: Tx buffer info to clean
802  *
803  * Returns true if cleanup/tranmission is done.
804  **/
805 bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi,
806 			   struct i40e_ring *tx_ring, int napi_budget)
807 {
808 	unsigned int ntc, total_bytes = 0, budget = vsi->work_limit;
809 	u32 i, completed_frames, frames_ready, xsk_frames = 0;
810 	struct xdp_umem *umem = tx_ring->xsk_umem;
811 	u32 head_idx = i40e_get_head(tx_ring);
812 	bool work_done = true, xmit_done;
813 	struct i40e_tx_buffer *tx_bi;
814 
815 	if (head_idx < tx_ring->next_to_clean)
816 		head_idx += tx_ring->count;
817 	frames_ready = head_idx - tx_ring->next_to_clean;
818 
819 	if (frames_ready == 0) {
820 		goto out_xmit;
821 	} else if (frames_ready > budget) {
822 		completed_frames = budget;
823 		work_done = false;
824 	} else {
825 		completed_frames = frames_ready;
826 	}
827 
828 	ntc = tx_ring->next_to_clean;
829 
830 	for (i = 0; i < completed_frames; i++) {
831 		tx_bi = &tx_ring->tx_bi[ntc];
832 
833 		if (tx_bi->xdpf)
834 			i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
835 		else
836 			xsk_frames++;
837 
838 		tx_bi->xdpf = NULL;
839 		total_bytes += tx_bi->bytecount;
840 
841 		if (++ntc >= tx_ring->count)
842 			ntc = 0;
843 	}
844 
845 	tx_ring->next_to_clean += completed_frames;
846 	if (unlikely(tx_ring->next_to_clean >= tx_ring->count))
847 		tx_ring->next_to_clean -= tx_ring->count;
848 
849 	if (xsk_frames)
850 		xsk_umem_complete_tx(umem, xsk_frames);
851 
852 	i40e_arm_wb(tx_ring, vsi, budget);
853 	i40e_update_tx_stats(tx_ring, completed_frames, total_bytes);
854 
855 out_xmit:
856 	xmit_done = i40e_xmit_zc(tx_ring, budget);
857 
858 	return work_done && xmit_done;
859 }
860 
861 /**
862  * i40e_xsk_async_xmit - Implements the ndo_xsk_async_xmit
863  * @dev: the netdevice
864  * @queue_id: queue id to wake up
865  *
866  * Returns <0 for errors, 0 otherwise.
867  **/
868 int i40e_xsk_async_xmit(struct net_device *dev, u32 queue_id)
869 {
870 	struct i40e_netdev_priv *np = netdev_priv(dev);
871 	struct i40e_vsi *vsi = np->vsi;
872 	struct i40e_ring *ring;
873 
874 	if (test_bit(__I40E_VSI_DOWN, vsi->state))
875 		return -ENETDOWN;
876 
877 	if (!i40e_enabled_xdp_vsi(vsi))
878 		return -ENXIO;
879 
880 	if (queue_id >= vsi->num_queue_pairs)
881 		return -ENXIO;
882 
883 	if (!vsi->xdp_rings[queue_id]->xsk_umem)
884 		return -ENXIO;
885 
886 	ring = vsi->xdp_rings[queue_id];
887 
888 	/* The idea here is that if NAPI is running, mark a miss, so
889 	 * it will run again. If not, trigger an interrupt and
890 	 * schedule the NAPI from interrupt context. If NAPI would be
891 	 * scheduled here, the interrupt affinity would not be
892 	 * honored.
893 	 */
894 	if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi))
895 		i40e_force_wb(vsi, ring->q_vector);
896 
897 	return 0;
898 }
899 
900 void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring)
901 {
902 	u16 i;
903 
904 	for (i = 0; i < rx_ring->count; i++) {
905 		struct i40e_rx_buffer *rx_bi = &rx_ring->rx_bi[i];
906 
907 		if (!rx_bi->addr)
908 			continue;
909 
910 		xsk_umem_fq_reuse(rx_ring->xsk_umem, rx_bi->handle);
911 		rx_bi->addr = NULL;
912 	}
913 }
914 
915 /**
916  * i40e_xsk_clean_xdp_ring - Clean the XDP Tx ring on shutdown
917  * @xdp_ring: XDP Tx ring
918  **/
919 void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring)
920 {
921 	u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
922 	struct xdp_umem *umem = tx_ring->xsk_umem;
923 	struct i40e_tx_buffer *tx_bi;
924 	u32 xsk_frames = 0;
925 
926 	while (ntc != ntu) {
927 		tx_bi = &tx_ring->tx_bi[ntc];
928 
929 		if (tx_bi->xdpf)
930 			i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
931 		else
932 			xsk_frames++;
933 
934 		tx_bi->xdpf = NULL;
935 
936 		ntc++;
937 		if (ntc >= tx_ring->count)
938 			ntc = 0;
939 	}
940 
941 	if (xsk_frames)
942 		xsk_umem_complete_tx(umem, xsk_frames);
943 }
944 
945 /**
946  * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have AF_XDP UMEM attached
947  * @vsi: vsi
948  *
949  * Returns true if any of the Rx rings has an AF_XDP UMEM attached
950  **/
951 bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi)
952 {
953 	int i;
954 
955 	if (!vsi->xsk_umems)
956 		return false;
957 
958 	for (i = 0; i < vsi->num_queue_pairs; i++) {
959 		if (vsi->xsk_umems[i])
960 			return true;
961 	}
962 
963 	return false;
964 }
965