Linux Audio

Check our new training course

Loading...
v6.2
  1// SPDX-License-Identifier: GPL-2.0
  2/* Copyright(c) 2018 Intel Corporation. */
  3
  4#include <linux/bpf_trace.h>
  5#include <linux/stringify.h>
  6#include <net/xdp_sock_drv.h>
  7#include <net/xdp.h>
  8
  9#include "i40e.h"
 10#include "i40e_txrx_common.h"
 11#include "i40e_xsk.h"
 12
 
 
 
 
 
 
 
 
 13void i40e_clear_rx_bi_zc(struct i40e_ring *rx_ring)
 14{
 15	memset(rx_ring->rx_bi_zc, 0,
 16	       sizeof(*rx_ring->rx_bi_zc) * rx_ring->count);
 17}
 18
 19static struct xdp_buff **i40e_rx_bi(struct i40e_ring *rx_ring, u32 idx)
 20{
 21	return &rx_ring->rx_bi_zc[idx];
 22}
 23
 24/**
 25 * i40e_realloc_rx_xdp_bi - reallocate SW ring for either XSK or normal buffer
 26 * @rx_ring: Current rx ring
 27 * @pool_present: is pool for XSK present
 28 *
 29 * Try allocating memory and return ENOMEM, if failed to allocate.
 30 * If allocation was successful, substitute buffer with allocated one.
 31 * Returns 0 on success, negative on failure
 32 */
 33static int i40e_realloc_rx_xdp_bi(struct i40e_ring *rx_ring, bool pool_present)
 34{
 35	size_t elem_size = pool_present ? sizeof(*rx_ring->rx_bi_zc) :
 36					  sizeof(*rx_ring->rx_bi);
 37	void *sw_ring = kcalloc(rx_ring->count, elem_size, GFP_KERNEL);
 38
 39	if (!sw_ring)
 40		return -ENOMEM;
 41
 42	if (pool_present) {
 43		kfree(rx_ring->rx_bi);
 44		rx_ring->rx_bi = NULL;
 45		rx_ring->rx_bi_zc = sw_ring;
 46	} else {
 47		kfree(rx_ring->rx_bi_zc);
 48		rx_ring->rx_bi_zc = NULL;
 49		rx_ring->rx_bi = sw_ring;
 50	}
 51	return 0;
 52}
 53
 54/**
 55 * i40e_realloc_rx_bi_zc - reallocate rx SW rings
 56 * @vsi: Current VSI
 57 * @zc: is zero copy set
 58 *
 59 * Reallocate buffer for rx_rings that might be used by XSK.
 60 * XDP requires more memory, than rx_buf provides.
 61 * Returns 0 on success, negative on failure
 62 */
 63int i40e_realloc_rx_bi_zc(struct i40e_vsi *vsi, bool zc)
 64{
 65	struct i40e_ring *rx_ring;
 66	unsigned long q;
 67
 68	for_each_set_bit(q, vsi->af_xdp_zc_qps, vsi->alloc_queue_pairs) {
 69		rx_ring = vsi->rx_rings[q];
 70		if (i40e_realloc_rx_xdp_bi(rx_ring, zc))
 71			return -ENOMEM;
 72	}
 73	return 0;
 74}
 75
 76/**
 77 * i40e_xsk_pool_enable - Enable/associate an AF_XDP buffer pool to a
 78 * certain ring/qid
 79 * @vsi: Current VSI
 80 * @pool: buffer pool
 81 * @qid: Rx ring to associate buffer pool with
 82 *
 83 * Returns 0 on success, <0 on failure
 84 **/
 85static int i40e_xsk_pool_enable(struct i40e_vsi *vsi,
 86				struct xsk_buff_pool *pool,
 87				u16 qid)
 88{
 89	struct net_device *netdev = vsi->netdev;
 90	bool if_running;
 91	int err;
 92
 93	if (vsi->type != I40E_VSI_MAIN)
 94		return -EINVAL;
 95
 96	if (qid >= vsi->num_queue_pairs)
 97		return -EINVAL;
 98
 99	if (qid >= netdev->real_num_rx_queues ||
100	    qid >= netdev->real_num_tx_queues)
101		return -EINVAL;
102
103	err = xsk_pool_dma_map(pool, &vsi->back->pdev->dev, I40E_RX_DMA_ATTR);
104	if (err)
105		return err;
106
107	set_bit(qid, vsi->af_xdp_zc_qps);
108
109	if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
110
111	if (if_running) {
112		err = i40e_queue_pair_disable(vsi, qid);
113		if (err)
114			return err;
115
116		err = i40e_realloc_rx_xdp_bi(vsi->rx_rings[qid], true);
117		if (err)
118			return err;
119
120		err = i40e_queue_pair_enable(vsi, qid);
121		if (err)
122			return err;
123
124		/* Kick start the NAPI context so that receiving will start */
125		err = i40e_xsk_wakeup(vsi->netdev, qid, XDP_WAKEUP_RX);
126		if (err)
127			return err;
128	}
129
130	return 0;
131}
132
133/**
134 * i40e_xsk_pool_disable - Disassociate an AF_XDP buffer pool from a
135 * certain ring/qid
136 * @vsi: Current VSI
137 * @qid: Rx ring to associate buffer pool with
138 *
139 * Returns 0 on success, <0 on failure
140 **/
141static int i40e_xsk_pool_disable(struct i40e_vsi *vsi, u16 qid)
142{
143	struct net_device *netdev = vsi->netdev;
144	struct xsk_buff_pool *pool;
145	bool if_running;
146	int err;
147
148	pool = xsk_get_pool_from_qid(netdev, qid);
149	if (!pool)
150		return -EINVAL;
151
152	if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
153
154	if (if_running) {
155		err = i40e_queue_pair_disable(vsi, qid);
156		if (err)
157			return err;
158	}
159
160	clear_bit(qid, vsi->af_xdp_zc_qps);
161	xsk_pool_dma_unmap(pool, I40E_RX_DMA_ATTR);
162
163	if (if_running) {
164		err = i40e_realloc_rx_xdp_bi(vsi->rx_rings[qid], false);
165		if (err)
166			return err;
167		err = i40e_queue_pair_enable(vsi, qid);
168		if (err)
169			return err;
170	}
171
172	return 0;
173}
174
175/**
176 * i40e_xsk_pool_setup - Enable/disassociate an AF_XDP buffer pool to/from
177 * a ring/qid
178 * @vsi: Current VSI
179 * @pool: Buffer pool to enable/associate to a ring, or NULL to disable
180 * @qid: Rx ring to (dis)associate buffer pool (from)to
181 *
182 * This function enables or disables a buffer pool to a certain ring.
183 *
184 * Returns 0 on success, <0 on failure
185 **/
186int i40e_xsk_pool_setup(struct i40e_vsi *vsi, struct xsk_buff_pool *pool,
187			u16 qid)
188{
189	return pool ? i40e_xsk_pool_enable(vsi, pool, qid) :
190		i40e_xsk_pool_disable(vsi, qid);
191}
192
193/**
194 * i40e_run_xdp_zc - Executes an XDP program on an xdp_buff
195 * @rx_ring: Rx ring
196 * @xdp: xdp_buff used as input to the XDP program
197 * @xdp_prog: XDP program to run
198 *
199 * Returns any of I40E_XDP_{PASS, CONSUMED, TX, REDIR}
200 **/
201static int i40e_run_xdp_zc(struct i40e_ring *rx_ring, struct xdp_buff *xdp,
202			   struct bpf_prog *xdp_prog)
203{
204	int err, result = I40E_XDP_PASS;
205	struct i40e_ring *xdp_ring;
 
206	u32 act;
207
 
 
 
 
 
208	act = bpf_prog_run_xdp(xdp_prog, xdp);
209
210	if (likely(act == XDP_REDIRECT)) {
211		err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
212		if (!err)
213			return I40E_XDP_REDIR;
214		if (xsk_uses_need_wakeup(rx_ring->xsk_pool) && err == -ENOBUFS)
215			result = I40E_XDP_EXIT;
216		else
217			result = I40E_XDP_CONSUMED;
218		goto out_failure;
219	}
220
221	switch (act) {
222	case XDP_PASS:
223		break;
224	case XDP_TX:
225		xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->queue_index];
226		result = i40e_xmit_xdp_tx_ring(xdp, xdp_ring);
227		if (result == I40E_XDP_CONSUMED)
228			goto out_failure;
229		break;
230	case XDP_DROP:
231		result = I40E_XDP_CONSUMED;
 
232		break;
233	default:
234		bpf_warn_invalid_xdp_action(rx_ring->netdev, xdp_prog, act);
235		fallthrough;
236	case XDP_ABORTED:
237		result = I40E_XDP_CONSUMED;
238out_failure:
239		trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
 
 
 
 
240	}
 
241	return result;
242}
243
244bool i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count)
245{
246	u16 ntu = rx_ring->next_to_use;
247	union i40e_rx_desc *rx_desc;
248	struct xdp_buff **xdp;
249	u32 nb_buffs, i;
250	dma_addr_t dma;
 
251
252	rx_desc = I40E_RX_DESC(rx_ring, ntu);
253	xdp = i40e_rx_bi(rx_ring, ntu);
254
255	nb_buffs = min_t(u16, count, rx_ring->count - ntu);
256	nb_buffs = xsk_buff_alloc_batch(rx_ring->xsk_pool, xdp, nb_buffs);
257	if (!nb_buffs)
258		return false;
259
260	i = nb_buffs;
261	while (i--) {
262		dma = xsk_buff_xdp_get_dma(*xdp);
263		rx_desc->read.pkt_addr = cpu_to_le64(dma);
264		rx_desc->read.hdr_addr = 0;
265
266		rx_desc++;
267		xdp++;
268	}
269
270	ntu += nb_buffs;
271	if (ntu == rx_ring->count) {
272		rx_desc = I40E_RX_DESC(rx_ring, 0);
273		ntu = 0;
274	}
 
 
 
275
276	/* clear the status bits for the next_to_use descriptor */
277	rx_desc->wb.qword1.status_error_len = 0;
278	i40e_release_rx_desc(rx_ring, ntu);
279
280	return count == nb_buffs;
281}
282
283/**
284 * i40e_construct_skb_zc - Create skbuff from zero-copy Rx buffer
285 * @rx_ring: Rx ring
286 * @xdp: xdp_buff
287 *
288 * This functions allocates a new skb from a zero-copy Rx buffer.
289 *
290 * Returns the skb, or NULL on failure.
291 **/
292static struct sk_buff *i40e_construct_skb_zc(struct i40e_ring *rx_ring,
293					     struct xdp_buff *xdp)
294{
295	unsigned int totalsize = xdp->data_end - xdp->data_meta;
296	unsigned int metasize = xdp->data - xdp->data_meta;
 
297	struct sk_buff *skb;
298
299	net_prefetch(xdp->data_meta);
300
301	/* allocate a skb to store the frags */
302	skb = __napi_alloc_skb(&rx_ring->q_vector->napi, totalsize,
 
303			       GFP_ATOMIC | __GFP_NOWARN);
304	if (unlikely(!skb))
305		goto out;
306
307	memcpy(__skb_put(skb, totalsize), xdp->data_meta,
308	       ALIGN(totalsize, sizeof(long)));
309
310	if (metasize) {
 
 
311		skb_metadata_set(skb, metasize);
312		__skb_pull(skb, metasize);
313	}
314
315out:
316	xsk_buff_free(xdp);
317	return skb;
318}
319
320static void i40e_handle_xdp_result_zc(struct i40e_ring *rx_ring,
321				      struct xdp_buff *xdp_buff,
322				      union i40e_rx_desc *rx_desc,
323				      unsigned int *rx_packets,
324				      unsigned int *rx_bytes,
325				      unsigned int size,
326				      unsigned int xdp_res,
327				      bool *failure)
328{
329	struct sk_buff *skb;
330
331	*rx_packets = 1;
332	*rx_bytes = size;
333
334	if (likely(xdp_res == I40E_XDP_REDIR) || xdp_res == I40E_XDP_TX)
335		return;
336
337	if (xdp_res == I40E_XDP_EXIT) {
338		*failure = true;
339		return;
340	}
341
342	if (xdp_res == I40E_XDP_CONSUMED) {
343		xsk_buff_free(xdp_buff);
344		return;
345	}
346	if (xdp_res == I40E_XDP_PASS) {
347		/* NB! We are not checking for errors using
348		 * i40e_test_staterr with
349		 * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that
350		 * SBP is *not* set in PRT_SBPVSI (default not set).
351		 */
352		skb = i40e_construct_skb_zc(rx_ring, xdp_buff);
353		if (!skb) {
354			rx_ring->rx_stats.alloc_buff_failed++;
355			*rx_packets = 0;
356			*rx_bytes = 0;
357			return;
358		}
359
360		if (eth_skb_pad(skb)) {
361			*rx_packets = 0;
362			*rx_bytes = 0;
363			return;
364		}
365
366		*rx_bytes = skb->len;
367		i40e_process_skb_fields(rx_ring, rx_desc, skb);
368		napi_gro_receive(&rx_ring->q_vector->napi, skb);
369		return;
370	}
371
372	/* Should never get here, as all valid cases have been handled already.
373	 */
374	WARN_ON_ONCE(1);
375}
376
377/**
378 * i40e_clean_rx_irq_zc - Consumes Rx packets from the hardware ring
379 * @rx_ring: Rx ring
380 * @budget: NAPI budget
381 *
382 * Returns amount of work completed
383 **/
384int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget)
385{
386	unsigned int total_rx_bytes = 0, total_rx_packets = 0;
387	u16 next_to_clean = rx_ring->next_to_clean;
388	u16 count_mask = rx_ring->count - 1;
389	unsigned int xdp_res, xdp_xmit = 0;
390	struct bpf_prog *xdp_prog;
391	bool failure = false;
392	u16 cleaned_count;
393
394	/* NB! xdp_prog will always be !NULL, due to the fact that
395	 * this path is enabled by setting an XDP program.
396	 */
397	xdp_prog = READ_ONCE(rx_ring->xdp_prog);
398
399	while (likely(total_rx_packets < (unsigned int)budget)) {
400		union i40e_rx_desc *rx_desc;
401		unsigned int rx_packets;
402		unsigned int rx_bytes;
403		struct xdp_buff *bi;
404		unsigned int size;
405		u64 qword;
406
407		rx_desc = I40E_RX_DESC(rx_ring, next_to_clean);
 
 
 
 
 
 
 
408		qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
409
410		/* This memory barrier is needed to keep us from reading
411		 * any other fields out of the rx_desc until we have
412		 * verified the descriptor has been written back.
413		 */
414		dma_rmb();
415
416		if (i40e_rx_is_programming_status(qword)) {
417			i40e_clean_programming_status(rx_ring,
418						      rx_desc->raw.qword[0],
419						      qword);
420			bi = *i40e_rx_bi(rx_ring, next_to_clean);
421			xsk_buff_free(bi);
422			next_to_clean = (next_to_clean + 1) & count_mask;
 
 
423			continue;
424		}
425
 
426		size = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
427		       I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
428		if (!size)
429			break;
430
431		bi = *i40e_rx_bi(rx_ring, next_to_clean);
432		xsk_buff_set_size(bi, size);
433		xsk_buff_dma_sync_for_cpu(bi, rx_ring->xsk_pool);
434
435		xdp_res = i40e_run_xdp_zc(rx_ring, bi, xdp_prog);
436		i40e_handle_xdp_result_zc(rx_ring, bi, rx_desc, &rx_packets,
437					  &rx_bytes, size, xdp_res, &failure);
438		if (failure)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
439			break;
440		total_rx_packets += rx_packets;
441		total_rx_bytes += rx_bytes;
442		xdp_xmit |= xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR);
443		next_to_clean = (next_to_clean + 1) & count_mask;
444	}
 
 
445
446	rx_ring->next_to_clean = next_to_clean;
447	cleaned_count = (next_to_clean - rx_ring->next_to_use - 1) & count_mask;
448
449	if (cleaned_count >= I40E_RX_BUFFER_WRITE)
450		failure |= !i40e_alloc_rx_buffers_zc(rx_ring, cleaned_count);
 
451
452	i40e_finalize_xdp_rx(rx_ring, xdp_xmit);
453	i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets);
454
455	if (xsk_uses_need_wakeup(rx_ring->xsk_pool)) {
456		if (failure || next_to_clean == rx_ring->next_to_use)
457			xsk_set_rx_need_wakeup(rx_ring->xsk_pool);
458		else
459			xsk_clear_rx_need_wakeup(rx_ring->xsk_pool);
460
461		return (int)total_rx_packets;
462	}
463	return failure ? budget : (int)total_rx_packets;
464}
465
466static void i40e_xmit_pkt(struct i40e_ring *xdp_ring, struct xdp_desc *desc,
467			  unsigned int *total_bytes)
468{
469	struct i40e_tx_desc *tx_desc;
470	dma_addr_t dma;
471
472	dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc->addr);
473	xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc->len);
474
475	tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use++);
476	tx_desc->buffer_addr = cpu_to_le64(dma);
477	tx_desc->cmd_type_offset_bsz = build_ctob(I40E_TX_DESC_CMD_ICRC | I40E_TX_DESC_CMD_EOP,
478						  0, desc->len, 0);
479
480	*total_bytes += desc->len;
481}
482
483static void i40e_xmit_pkt_batch(struct i40e_ring *xdp_ring, struct xdp_desc *desc,
484				unsigned int *total_bytes)
485{
486	u16 ntu = xdp_ring->next_to_use;
487	struct i40e_tx_desc *tx_desc;
488	dma_addr_t dma;
489	u32 i;
490
491	loop_unrolled_for(i = 0; i < PKTS_PER_BATCH; i++) {
492		dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc[i].addr);
493		xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc[i].len);
494
495		tx_desc = I40E_TX_DESC(xdp_ring, ntu++);
496		tx_desc->buffer_addr = cpu_to_le64(dma);
497		tx_desc->cmd_type_offset_bsz = build_ctob(I40E_TX_DESC_CMD_ICRC |
498							  I40E_TX_DESC_CMD_EOP,
499							  0, desc[i].len, 0);
500
501		*total_bytes += desc[i].len;
502	}
503
504	xdp_ring->next_to_use = ntu;
505}
506
507static void i40e_fill_tx_hw_ring(struct i40e_ring *xdp_ring, struct xdp_desc *descs, u32 nb_pkts,
508				 unsigned int *total_bytes)
509{
510	u32 batched, leftover, i;
511
512	batched = nb_pkts & ~(PKTS_PER_BATCH - 1);
513	leftover = nb_pkts & (PKTS_PER_BATCH - 1);
514	for (i = 0; i < batched; i += PKTS_PER_BATCH)
515		i40e_xmit_pkt_batch(xdp_ring, &descs[i], total_bytes);
516	for (i = batched; i < batched + leftover; i++)
517		i40e_xmit_pkt(xdp_ring, &descs[i], total_bytes);
518}
519
520static void i40e_set_rs_bit(struct i40e_ring *xdp_ring)
521{
522	u16 ntu = xdp_ring->next_to_use ? xdp_ring->next_to_use - 1 : xdp_ring->count - 1;
523	struct i40e_tx_desc *tx_desc;
524
525	tx_desc = I40E_TX_DESC(xdp_ring, ntu);
526	tx_desc->cmd_type_offset_bsz |= cpu_to_le64(I40E_TX_DESC_CMD_RS << I40E_TXD_QW1_CMD_SHIFT);
527}
528
529/**
530 * i40e_xmit_zc - Performs zero-copy Tx AF_XDP
531 * @xdp_ring: XDP Tx ring
532 * @budget: NAPI budget
533 *
534 * Returns true if the work is finished.
535 **/
536static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
537{
538	struct xdp_desc *descs = xdp_ring->xsk_pool->tx_descs;
539	u32 nb_pkts, nb_processed = 0;
540	unsigned int total_bytes = 0;
 
 
541
542	nb_pkts = xsk_tx_peek_release_desc_batch(xdp_ring->xsk_pool, budget);
543	if (!nb_pkts)
544		return true;
545
546	if (xdp_ring->next_to_use + nb_pkts >= xdp_ring->count) {
547		nb_processed = xdp_ring->count - xdp_ring->next_to_use;
548		i40e_fill_tx_hw_ring(xdp_ring, descs, nb_processed, &total_bytes);
549		xdp_ring->next_to_use = 0;
550	}
551
552	i40e_fill_tx_hw_ring(xdp_ring, &descs[nb_processed], nb_pkts - nb_processed,
553			     &total_bytes);
554
555	/* Request an interrupt for the last frame and bump tail ptr. */
556	i40e_set_rs_bit(xdp_ring);
557	i40e_xdp_ring_update_tail(xdp_ring);
 
 
 
 
 
 
 
 
 
 
 
558
559	i40e_update_tx_stats(xdp_ring, nb_pkts, total_bytes);
 
 
 
 
560
561	return nb_pkts < budget;
 
 
 
 
562}
563
564/**
565 * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry
566 * @tx_ring: XDP Tx ring
567 * @tx_bi: Tx buffer info to clean
568 **/
569static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring,
570				     struct i40e_tx_buffer *tx_bi)
571{
572	xdp_return_frame(tx_bi->xdpf);
573	tx_ring->xdp_tx_active--;
574	dma_unmap_single(tx_ring->dev,
575			 dma_unmap_addr(tx_bi, dma),
576			 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE);
577	dma_unmap_len_set(tx_bi, len, 0);
578}
579
580/**
581 * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries
582 * @vsi: Current VSI
583 * @tx_ring: XDP Tx ring
584 *
585 * Returns true if cleanup/tranmission is done.
586 **/
587bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi, struct i40e_ring *tx_ring)
588{
589	struct xsk_buff_pool *bp = tx_ring->xsk_pool;
590	u32 i, completed_frames, xsk_frames = 0;
591	u32 head_idx = i40e_get_head(tx_ring);
592	struct i40e_tx_buffer *tx_bi;
593	unsigned int ntc;
594
595	if (head_idx < tx_ring->next_to_clean)
596		head_idx += tx_ring->count;
597	completed_frames = head_idx - tx_ring->next_to_clean;
598
599	if (completed_frames == 0)
600		goto out_xmit;
601
602	if (likely(!tx_ring->xdp_tx_active)) {
603		xsk_frames = completed_frames;
604		goto skip;
605	}
606
607	ntc = tx_ring->next_to_clean;
608
609	for (i = 0; i < completed_frames; i++) {
610		tx_bi = &tx_ring->tx_bi[ntc];
611
612		if (tx_bi->xdpf) {
613			i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
614			tx_bi->xdpf = NULL;
615		} else {
616			xsk_frames++;
617		}
618
619		if (++ntc >= tx_ring->count)
620			ntc = 0;
621	}
622
623skip:
624	tx_ring->next_to_clean += completed_frames;
625	if (unlikely(tx_ring->next_to_clean >= tx_ring->count))
626		tx_ring->next_to_clean -= tx_ring->count;
627
628	if (xsk_frames)
629		xsk_tx_completed(bp, xsk_frames);
630
631	i40e_arm_wb(tx_ring, vsi, completed_frames);
632
633out_xmit:
634	if (xsk_uses_need_wakeup(tx_ring->xsk_pool))
635		xsk_set_tx_need_wakeup(tx_ring->xsk_pool);
636
637	return i40e_xmit_zc(tx_ring, I40E_DESC_UNUSED(tx_ring));
638}
639
640/**
641 * i40e_xsk_wakeup - Implements the ndo_xsk_wakeup
642 * @dev: the netdevice
643 * @queue_id: queue id to wake up
644 * @flags: ignored in our case since we have Rx and Tx in the same NAPI.
645 *
646 * Returns <0 for errors, 0 otherwise.
647 **/
648int i40e_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
649{
650	struct i40e_netdev_priv *np = netdev_priv(dev);
651	struct i40e_vsi *vsi = np->vsi;
652	struct i40e_pf *pf = vsi->back;
653	struct i40e_ring *ring;
654
655	if (test_bit(__I40E_CONFIG_BUSY, pf->state))
656		return -EAGAIN;
657
658	if (test_bit(__I40E_VSI_DOWN, vsi->state))
659		return -ENETDOWN;
660
661	if (!i40e_enabled_xdp_vsi(vsi))
662		return -EINVAL;
663
664	if (queue_id >= vsi->num_queue_pairs)
665		return -EINVAL;
666
667	if (!vsi->xdp_rings[queue_id]->xsk_pool)
668		return -EINVAL;
669
670	ring = vsi->xdp_rings[queue_id];
671
672	/* The idea here is that if NAPI is running, mark a miss, so
673	 * it will run again. If not, trigger an interrupt and
674	 * schedule the NAPI from interrupt context. If NAPI would be
675	 * scheduled here, the interrupt affinity would not be
676	 * honored.
677	 */
678	if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi))
679		i40e_force_wb(vsi, ring->q_vector);
680
681	return 0;
682}
683
684void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring)
685{
686	u16 count_mask = rx_ring->count - 1;
687	u16 ntc = rx_ring->next_to_clean;
688	u16 ntu = rx_ring->next_to_use;
689
690	for ( ; ntc != ntu; ntc = (ntc + 1)  & count_mask) {
691		struct xdp_buff *rx_bi = *i40e_rx_bi(rx_ring, ntc);
 
 
 
692
693		xsk_buff_free(rx_bi);
 
694	}
695}
696
697/**
698 * i40e_xsk_clean_tx_ring - Clean the XDP Tx ring on shutdown
699 * @tx_ring: XDP Tx ring
700 **/
701void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring)
702{
703	u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
704	struct xsk_buff_pool *bp = tx_ring->xsk_pool;
705	struct i40e_tx_buffer *tx_bi;
706	u32 xsk_frames = 0;
707
708	while (ntc != ntu) {
709		tx_bi = &tx_ring->tx_bi[ntc];
710
711		if (tx_bi->xdpf)
712			i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
713		else
714			xsk_frames++;
715
716		tx_bi->xdpf = NULL;
717
718		ntc++;
719		if (ntc >= tx_ring->count)
720			ntc = 0;
721	}
722
723	if (xsk_frames)
724		xsk_tx_completed(bp, xsk_frames);
725}
726
727/**
728 * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have an AF_XDP
729 * buffer pool attached
730 * @vsi: vsi
731 *
732 * Returns true if any of the Rx rings has an AF_XDP buffer pool attached
733 **/
734bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi)
735{
736	struct net_device *netdev = vsi->netdev;
737	int i;
738
739	for (i = 0; i < vsi->num_queue_pairs; i++) {
740		if (xsk_get_pool_from_qid(netdev, i))
741			return true;
742	}
743
744	return false;
745}
v5.9
  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_drv.h>
  6#include <net/xdp.h>
  7
  8#include "i40e.h"
  9#include "i40e_txrx_common.h"
 10#include "i40e_xsk.h"
 11
 12int i40e_alloc_rx_bi_zc(struct i40e_ring *rx_ring)
 13{
 14	unsigned long sz = sizeof(*rx_ring->rx_bi_zc) * rx_ring->count;
 15
 16	rx_ring->rx_bi_zc = kzalloc(sz, GFP_KERNEL);
 17	return rx_ring->rx_bi_zc ? 0 : -ENOMEM;
 18}
 19
 20void i40e_clear_rx_bi_zc(struct i40e_ring *rx_ring)
 21{
 22	memset(rx_ring->rx_bi_zc, 0,
 23	       sizeof(*rx_ring->rx_bi_zc) * rx_ring->count);
 24}
 25
 26static struct xdp_buff **i40e_rx_bi(struct i40e_ring *rx_ring, u32 idx)
 27{
 28	return &rx_ring->rx_bi_zc[idx];
 29}
 30
 31/**
 32 * i40e_xsk_umem_enable - Enable/associate a UMEM to a certain ring/qid
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 33 * @vsi: Current VSI
 34 * @umem: UMEM
 35 * @qid: Rx ring to associate UMEM to
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 36 *
 37 * Returns 0 on success, <0 on failure
 38 **/
 39static int i40e_xsk_umem_enable(struct i40e_vsi *vsi, struct xdp_umem *umem,
 
 40				u16 qid)
 41{
 42	struct net_device *netdev = vsi->netdev;
 43	bool if_running;
 44	int err;
 45
 46	if (vsi->type != I40E_VSI_MAIN)
 47		return -EINVAL;
 48
 49	if (qid >= vsi->num_queue_pairs)
 50		return -EINVAL;
 51
 52	if (qid >= netdev->real_num_rx_queues ||
 53	    qid >= netdev->real_num_tx_queues)
 54		return -EINVAL;
 55
 56	err = xsk_buff_dma_map(umem, &vsi->back->pdev->dev, I40E_RX_DMA_ATTR);
 57	if (err)
 58		return err;
 59
 60	set_bit(qid, vsi->af_xdp_zc_qps);
 61
 62	if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
 63
 64	if (if_running) {
 65		err = i40e_queue_pair_disable(vsi, qid);
 66		if (err)
 67			return err;
 68
 
 
 
 
 69		err = i40e_queue_pair_enable(vsi, qid);
 70		if (err)
 71			return err;
 72
 73		/* Kick start the NAPI context so that receiving will start */
 74		err = i40e_xsk_wakeup(vsi->netdev, qid, XDP_WAKEUP_RX);
 75		if (err)
 76			return err;
 77	}
 78
 79	return 0;
 80}
 81
 82/**
 83 * i40e_xsk_umem_disable - Disassociate a UMEM from a certain ring/qid
 
 84 * @vsi: Current VSI
 85 * @qid: Rx ring to associate UMEM to
 86 *
 87 * Returns 0 on success, <0 on failure
 88 **/
 89static int i40e_xsk_umem_disable(struct i40e_vsi *vsi, u16 qid)
 90{
 91	struct net_device *netdev = vsi->netdev;
 92	struct xdp_umem *umem;
 93	bool if_running;
 94	int err;
 95
 96	umem = xdp_get_umem_from_qid(netdev, qid);
 97	if (!umem)
 98		return -EINVAL;
 99
100	if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
101
102	if (if_running) {
103		err = i40e_queue_pair_disable(vsi, qid);
104		if (err)
105			return err;
106	}
107
108	clear_bit(qid, vsi->af_xdp_zc_qps);
109	xsk_buff_dma_unmap(umem, I40E_RX_DMA_ATTR);
110
111	if (if_running) {
 
 
 
112		err = i40e_queue_pair_enable(vsi, qid);
113		if (err)
114			return err;
115	}
116
117	return 0;
118}
119
120/**
121 * i40e_xsk_umem_setup - Enable/disassociate a UMEM to/from a ring/qid
 
122 * @vsi: Current VSI
123 * @umem: UMEM to enable/associate to a ring, or NULL to disable
124 * @qid: Rx ring to (dis)associate UMEM (from)to
125 *
126 * This function enables or disables a UMEM to a certain ring.
127 *
128 * Returns 0 on success, <0 on failure
129 **/
130int i40e_xsk_umem_setup(struct i40e_vsi *vsi, struct xdp_umem *umem,
131			u16 qid)
132{
133	return umem ? i40e_xsk_umem_enable(vsi, umem, qid) :
134		i40e_xsk_umem_disable(vsi, qid);
135}
136
137/**
138 * i40e_run_xdp_zc - Executes an XDP program on an xdp_buff
139 * @rx_ring: Rx ring
140 * @xdp: xdp_buff used as input to the XDP program
 
141 *
142 * Returns any of I40E_XDP_{PASS, CONSUMED, TX, REDIR}
143 **/
144static int i40e_run_xdp_zc(struct i40e_ring *rx_ring, struct xdp_buff *xdp)
 
145{
146	int err, result = I40E_XDP_PASS;
147	struct i40e_ring *xdp_ring;
148	struct bpf_prog *xdp_prog;
149	u32 act;
150
151	rcu_read_lock();
152	/* NB! xdp_prog will always be !NULL, due to the fact that
153	 * this path is enabled by setting an XDP program.
154	 */
155	xdp_prog = READ_ONCE(rx_ring->xdp_prog);
156	act = bpf_prog_run_xdp(xdp_prog, xdp);
157
 
 
 
 
 
 
 
 
 
 
 
158	switch (act) {
159	case XDP_PASS:
160		break;
161	case XDP_TX:
162		xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->queue_index];
163		result = i40e_xmit_xdp_tx_ring(xdp, xdp_ring);
 
 
164		break;
165	case XDP_REDIRECT:
166		err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
167		result = !err ? I40E_XDP_REDIR : I40E_XDP_CONSUMED;
168		break;
169	default:
170		bpf_warn_invalid_xdp_action(act);
171		fallthrough;
172	case XDP_ABORTED:
 
 
173		trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
174		fallthrough; /* handle aborts by dropping packet */
175	case XDP_DROP:
176		result = I40E_XDP_CONSUMED;
177		break;
178	}
179	rcu_read_unlock();
180	return result;
181}
182
183bool i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count)
184{
185	u16 ntu = rx_ring->next_to_use;
186	union i40e_rx_desc *rx_desc;
187	struct xdp_buff **bi, *xdp;
 
188	dma_addr_t dma;
189	bool ok = true;
190
191	rx_desc = I40E_RX_DESC(rx_ring, ntu);
192	bi = i40e_rx_bi(rx_ring, ntu);
193	do {
194		xdp = xsk_buff_alloc(rx_ring->xsk_umem);
195		if (!xdp) {
196			ok = false;
197			goto no_buffers;
198		}
199		*bi = xdp;
200		dma = xsk_buff_xdp_get_dma(xdp);
 
201		rx_desc->read.pkt_addr = cpu_to_le64(dma);
202		rx_desc->read.hdr_addr = 0;
203
204		rx_desc++;
205		bi++;
206		ntu++;
207
208		if (unlikely(ntu == rx_ring->count)) {
209			rx_desc = I40E_RX_DESC(rx_ring, 0);
210			bi = i40e_rx_bi(rx_ring, 0);
211			ntu = 0;
212		}
213
214		count--;
215	} while (count);
216
217no_buffers:
218	if (rx_ring->next_to_use != ntu)
219		i40e_release_rx_desc(rx_ring, ntu);
220
221	return ok;
222}
223
224/**
225 * i40e_construct_skb_zc - Create skbuff from zero-copy Rx buffer
226 * @rx_ring: Rx ring
227 * @xdp: xdp_buff
228 *
229 * This functions allocates a new skb from a zero-copy Rx buffer.
230 *
231 * Returns the skb, or NULL on failure.
232 **/
233static struct sk_buff *i40e_construct_skb_zc(struct i40e_ring *rx_ring,
234					     struct xdp_buff *xdp)
235{
 
236	unsigned int metasize = xdp->data - xdp->data_meta;
237	unsigned int datasize = xdp->data_end - xdp->data;
238	struct sk_buff *skb;
239
 
 
240	/* allocate a skb to store the frags */
241	skb = __napi_alloc_skb(&rx_ring->q_vector->napi,
242			       xdp->data_end - xdp->data_hard_start,
243			       GFP_ATOMIC | __GFP_NOWARN);
244	if (unlikely(!skb))
245		return NULL;
 
 
 
246
247	skb_reserve(skb, xdp->data - xdp->data_hard_start);
248	memcpy(__skb_put(skb, datasize), xdp->data, datasize);
249	if (metasize)
250		skb_metadata_set(skb, metasize);
 
 
251
 
252	xsk_buff_free(xdp);
253	return skb;
254}
255
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
256/**
257 * i40e_clean_rx_irq_zc - Consumes Rx packets from the hardware ring
258 * @rx_ring: Rx ring
259 * @budget: NAPI budget
260 *
261 * Returns amount of work completed
262 **/
263int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget)
264{
265	unsigned int total_rx_bytes = 0, total_rx_packets = 0;
266	u16 cleaned_count = I40E_DESC_UNUSED(rx_ring);
 
267	unsigned int xdp_res, xdp_xmit = 0;
 
268	bool failure = false;
269	struct sk_buff *skb;
 
 
 
 
 
270
271	while (likely(total_rx_packets < (unsigned int)budget)) {
272		union i40e_rx_desc *rx_desc;
273		struct xdp_buff **bi;
 
 
274		unsigned int size;
275		u64 qword;
276
277		if (cleaned_count >= I40E_RX_BUFFER_WRITE) {
278			failure = failure ||
279				  !i40e_alloc_rx_buffers_zc(rx_ring,
280							    cleaned_count);
281			cleaned_count = 0;
282		}
283
284		rx_desc = I40E_RX_DESC(rx_ring, rx_ring->next_to_clean);
285		qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
286
287		/* This memory barrier is needed to keep us from reading
288		 * any other fields out of the rx_desc until we have
289		 * verified the descriptor has been written back.
290		 */
291		dma_rmb();
292
293		if (i40e_rx_is_programming_status(qword)) {
294			i40e_clean_programming_status(rx_ring,
295						      rx_desc->raw.qword[0],
296						      qword);
297			bi = i40e_rx_bi(rx_ring, rx_ring->next_to_clean);
298			xsk_buff_free(*bi);
299			*bi = NULL;
300			cleaned_count++;
301			i40e_inc_ntc(rx_ring);
302			continue;
303		}
304
305		bi = i40e_rx_bi(rx_ring, rx_ring->next_to_clean);
306		size = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
307		       I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
308		if (!size)
309			break;
310
311		bi = i40e_rx_bi(rx_ring, rx_ring->next_to_clean);
312		(*bi)->data_end = (*bi)->data + size;
313		xsk_buff_dma_sync_for_cpu(*bi);
314
315		xdp_res = i40e_run_xdp_zc(rx_ring, *bi);
316		if (xdp_res) {
317			if (xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR))
318				xdp_xmit |= xdp_res;
319			else
320				xsk_buff_free(*bi);
321
322			*bi = NULL;
323			total_rx_bytes += size;
324			total_rx_packets++;
325
326			cleaned_count++;
327			i40e_inc_ntc(rx_ring);
328			continue;
329		}
330
331		/* XDP_PASS path */
332
333		/* NB! We are not checking for errors using
334		 * i40e_test_staterr with
335		 * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that
336		 * SBP is *not* set in PRT_SBPVSI (default not set).
337		 */
338		skb = i40e_construct_skb_zc(rx_ring, *bi);
339		*bi = NULL;
340		if (!skb) {
341			rx_ring->rx_stats.alloc_buff_failed++;
342			break;
343		}
344
345		cleaned_count++;
346		i40e_inc_ntc(rx_ring);
347
348		if (eth_skb_pad(skb))
349			continue;
350
351		total_rx_bytes += skb->len;
352		total_rx_packets++;
353
354		i40e_process_skb_fields(rx_ring, rx_desc, skb);
355		napi_gro_receive(&rx_ring->q_vector->napi, skb);
356	}
357
358	i40e_finalize_xdp_rx(rx_ring, xdp_xmit);
359	i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets);
360
361	if (xsk_umem_uses_need_wakeup(rx_ring->xsk_umem)) {
362		if (failure || rx_ring->next_to_clean == rx_ring->next_to_use)
363			xsk_set_rx_need_wakeup(rx_ring->xsk_umem);
364		else
365			xsk_clear_rx_need_wakeup(rx_ring->xsk_umem);
366
367		return (int)total_rx_packets;
368	}
369	return failure ? budget : (int)total_rx_packets;
370}
371
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
372/**
373 * i40e_xmit_zc - Performs zero-copy Tx AF_XDP
374 * @xdp_ring: XDP Tx ring
375 * @budget: NAPI budget
376 *
377 * Returns true if the work is finished.
378 **/
379static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
380{
381	unsigned int sent_frames = 0, total_bytes = 0;
382	struct i40e_tx_desc *tx_desc = NULL;
383	struct i40e_tx_buffer *tx_bi;
384	struct xdp_desc desc;
385	dma_addr_t dma;
386
387	while (budget-- > 0) {
388		if (!xsk_umem_consume_tx(xdp_ring->xsk_umem, &desc))
389			break;
390
391		dma = xsk_buff_raw_get_dma(xdp_ring->xsk_umem, desc.addr);
392		xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_umem, dma,
393						 desc.len);
 
 
394
395		tx_bi = &xdp_ring->tx_bi[xdp_ring->next_to_use];
396		tx_bi->bytecount = desc.len;
397
398		tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use);
399		tx_desc->buffer_addr = cpu_to_le64(dma);
400		tx_desc->cmd_type_offset_bsz =
401			build_ctob(I40E_TX_DESC_CMD_ICRC
402				   | I40E_TX_DESC_CMD_EOP,
403				   0, desc.len, 0);
404
405		sent_frames++;
406		total_bytes += tx_bi->bytecount;
407
408		xdp_ring->next_to_use++;
409		if (xdp_ring->next_to_use == xdp_ring->count)
410			xdp_ring->next_to_use = 0;
411	}
412
413	if (tx_desc) {
414		/* Request an interrupt for the last frame and bump tail ptr. */
415		tx_desc->cmd_type_offset_bsz |= (I40E_TX_DESC_CMD_RS <<
416						 I40E_TXD_QW1_CMD_SHIFT);
417		i40e_xdp_ring_update_tail(xdp_ring);
418
419		xsk_umem_consume_tx_done(xdp_ring->xsk_umem);
420		i40e_update_tx_stats(xdp_ring, sent_frames, total_bytes);
421	}
422
423	return !!budget;
424}
425
426/**
427 * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry
428 * @tx_ring: XDP Tx ring
429 * @tx_bi: Tx buffer info to clean
430 **/
431static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring,
432				     struct i40e_tx_buffer *tx_bi)
433{
434	xdp_return_frame(tx_bi->xdpf);
435	tx_ring->xdp_tx_active--;
436	dma_unmap_single(tx_ring->dev,
437			 dma_unmap_addr(tx_bi, dma),
438			 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE);
439	dma_unmap_len_set(tx_bi, len, 0);
440}
441
442/**
443 * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries
444 * @vsi: Current VSI
445 * @tx_ring: XDP Tx ring
446 *
447 * Returns true if cleanup/tranmission is done.
448 **/
449bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi, struct i40e_ring *tx_ring)
450{
451	struct xdp_umem *umem = tx_ring->xsk_umem;
452	u32 i, completed_frames, xsk_frames = 0;
453	u32 head_idx = i40e_get_head(tx_ring);
454	struct i40e_tx_buffer *tx_bi;
455	unsigned int ntc;
456
457	if (head_idx < tx_ring->next_to_clean)
458		head_idx += tx_ring->count;
459	completed_frames = head_idx - tx_ring->next_to_clean;
460
461	if (completed_frames == 0)
462		goto out_xmit;
463
464	if (likely(!tx_ring->xdp_tx_active)) {
465		xsk_frames = completed_frames;
466		goto skip;
467	}
468
469	ntc = tx_ring->next_to_clean;
470
471	for (i = 0; i < completed_frames; i++) {
472		tx_bi = &tx_ring->tx_bi[ntc];
473
474		if (tx_bi->xdpf) {
475			i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
476			tx_bi->xdpf = NULL;
477		} else {
478			xsk_frames++;
479		}
480
481		if (++ntc >= tx_ring->count)
482			ntc = 0;
483	}
484
485skip:
486	tx_ring->next_to_clean += completed_frames;
487	if (unlikely(tx_ring->next_to_clean >= tx_ring->count))
488		tx_ring->next_to_clean -= tx_ring->count;
489
490	if (xsk_frames)
491		xsk_umem_complete_tx(umem, xsk_frames);
492
493	i40e_arm_wb(tx_ring, vsi, completed_frames);
494
495out_xmit:
496	if (xsk_umem_uses_need_wakeup(tx_ring->xsk_umem))
497		xsk_set_tx_need_wakeup(tx_ring->xsk_umem);
498
499	return i40e_xmit_zc(tx_ring, I40E_DESC_UNUSED(tx_ring));
500}
501
502/**
503 * i40e_xsk_wakeup - Implements the ndo_xsk_wakeup
504 * @dev: the netdevice
505 * @queue_id: queue id to wake up
506 * @flags: ignored in our case since we have Rx and Tx in the same NAPI.
507 *
508 * Returns <0 for errors, 0 otherwise.
509 **/
510int i40e_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
511{
512	struct i40e_netdev_priv *np = netdev_priv(dev);
513	struct i40e_vsi *vsi = np->vsi;
514	struct i40e_pf *pf = vsi->back;
515	struct i40e_ring *ring;
516
517	if (test_bit(__I40E_CONFIG_BUSY, pf->state))
518		return -EAGAIN;
519
520	if (test_bit(__I40E_VSI_DOWN, vsi->state))
521		return -ENETDOWN;
522
523	if (!i40e_enabled_xdp_vsi(vsi))
524		return -ENXIO;
525
526	if (queue_id >= vsi->num_queue_pairs)
527		return -ENXIO;
528
529	if (!vsi->xdp_rings[queue_id]->xsk_umem)
530		return -ENXIO;
531
532	ring = vsi->xdp_rings[queue_id];
533
534	/* The idea here is that if NAPI is running, mark a miss, so
535	 * it will run again. If not, trigger an interrupt and
536	 * schedule the NAPI from interrupt context. If NAPI would be
537	 * scheduled here, the interrupt affinity would not be
538	 * honored.
539	 */
540	if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi))
541		i40e_force_wb(vsi, ring->q_vector);
542
543	return 0;
544}
545
546void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring)
547{
548	u16 i;
 
 
549
550	for (i = 0; i < rx_ring->count; i++) {
551		struct xdp_buff *rx_bi = *i40e_rx_bi(rx_ring, i);
552
553		if (!rx_bi)
554			continue;
555
556		xsk_buff_free(rx_bi);
557		rx_bi = NULL;
558	}
559}
560
561/**
562 * i40e_xsk_clean_xdp_ring - Clean the XDP Tx ring on shutdown
563 * @tx_ring: XDP Tx ring
564 **/
565void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring)
566{
567	u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
568	struct xdp_umem *umem = tx_ring->xsk_umem;
569	struct i40e_tx_buffer *tx_bi;
570	u32 xsk_frames = 0;
571
572	while (ntc != ntu) {
573		tx_bi = &tx_ring->tx_bi[ntc];
574
575		if (tx_bi->xdpf)
576			i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
577		else
578			xsk_frames++;
579
580		tx_bi->xdpf = NULL;
581
582		ntc++;
583		if (ntc >= tx_ring->count)
584			ntc = 0;
585	}
586
587	if (xsk_frames)
588		xsk_umem_complete_tx(umem, xsk_frames);
589}
590
591/**
592 * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have AF_XDP UMEM attached
 
593 * @vsi: vsi
594 *
595 * Returns true if any of the Rx rings has an AF_XDP UMEM attached
596 **/
597bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi)
598{
599	struct net_device *netdev = vsi->netdev;
600	int i;
601
602	for (i = 0; i < vsi->num_queue_pairs; i++) {
603		if (xdp_get_umem_from_qid(netdev, i))
604			return true;
605	}
606
607	return false;
608}