1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2019, Intel Corporation. */
3 
4 #include <linux/bpf_trace.h>
5 #include <net/xdp_sock_drv.h>
6 #include <net/xdp.h>
7 #include "ice.h"
8 #include "ice_base.h"
9 #include "ice_type.h"
10 #include "ice_xsk.h"
11 #include "ice_txrx.h"
12 #include "ice_txrx_lib.h"
13 #include "ice_lib.h"
14 
15 static struct xdp_buff **ice_xdp_buf(struct ice_rx_ring *rx_ring, u32 idx)
16 {
17 	return &rx_ring->xdp_buf[idx];
18 }
19 
20 /**
21  * ice_qp_reset_stats - Resets all stats for rings of given index
22  * @vsi: VSI that contains rings of interest
23  * @q_idx: ring index in array
24  */
25 static void ice_qp_reset_stats(struct ice_vsi *vsi, u16 q_idx)
26 {
27 	struct ice_vsi_stats *vsi_stat;
28 	struct ice_pf *pf;
29 
30 	pf = vsi->back;
31 	if (!pf->vsi_stats)
32 		return;
33 
34 	vsi_stat = pf->vsi_stats[vsi->idx];
35 	if (!vsi_stat)
36 		return;
37 
38 	memset(&vsi_stat->rx_ring_stats[q_idx]->rx_stats, 0,
39 	       sizeof(vsi_stat->rx_ring_stats[q_idx]->rx_stats));
40 	memset(&vsi_stat->tx_ring_stats[q_idx]->stats, 0,
41 	       sizeof(vsi_stat->tx_ring_stats[q_idx]->stats));
42 	if (ice_is_xdp_ena_vsi(vsi))
43 		memset(&vsi->xdp_rings[q_idx]->ring_stats->stats, 0,
44 		       sizeof(vsi->xdp_rings[q_idx]->ring_stats->stats));
45 }
46 
47 /**
48  * ice_qp_clean_rings - Cleans all the rings of a given index
49  * @vsi: VSI that contains rings of interest
50  * @q_idx: ring index in array
51  */
52 static void ice_qp_clean_rings(struct ice_vsi *vsi, u16 q_idx)
53 {
54 	ice_clean_tx_ring(vsi->tx_rings[q_idx]);
55 	if (ice_is_xdp_ena_vsi(vsi)) {
56 		synchronize_rcu();
57 		ice_clean_tx_ring(vsi->xdp_rings[q_idx]);
58 	}
59 	ice_clean_rx_ring(vsi->rx_rings[q_idx]);
60 }
61 
62 /**
63  * ice_qvec_toggle_napi - Enables/disables NAPI for a given q_vector
64  * @vsi: VSI that has netdev
65  * @q_vector: q_vector that has NAPI context
66  * @enable: true for enable, false for disable
67  */
68 static void
69 ice_qvec_toggle_napi(struct ice_vsi *vsi, struct ice_q_vector *q_vector,
70 		     bool enable)
71 {
72 	if (!vsi->netdev || !q_vector)
73 		return;
74 
75 	if (enable)
76 		napi_enable(&q_vector->napi);
77 	else
78 		napi_disable(&q_vector->napi);
79 }
80 
81 /**
82  * ice_qvec_dis_irq - Mask off queue interrupt generation on given ring
83  * @vsi: the VSI that contains queue vector being un-configured
84  * @rx_ring: Rx ring that will have its IRQ disabled
85  * @q_vector: queue vector
86  */
87 static void
88 ice_qvec_dis_irq(struct ice_vsi *vsi, struct ice_rx_ring *rx_ring,
89 		 struct ice_q_vector *q_vector)
90 {
91 	struct ice_pf *pf = vsi->back;
92 	struct ice_hw *hw = &pf->hw;
93 	u16 reg;
94 	u32 val;
95 
96 	/* QINT_TQCTL is being cleared in ice_vsi_stop_tx_ring, so handle
97 	 * here only QINT_RQCTL
98 	 */
99 	reg = rx_ring->reg_idx;
100 	val = rd32(hw, QINT_RQCTL(reg));
101 	val &= ~QINT_RQCTL_CAUSE_ENA_M;
102 	wr32(hw, QINT_RQCTL(reg), val);
103 
104 	if (q_vector) {
105 		wr32(hw, GLINT_DYN_CTL(q_vector->reg_idx), 0);
106 		ice_flush(hw);
107 		synchronize_irq(q_vector->irq.virq);
108 	}
109 }
110 
111 /**
112  * ice_qvec_cfg_msix - Enable IRQ for given queue vector
113  * @vsi: the VSI that contains queue vector
114  * @q_vector: queue vector
115  */
116 static void
117 ice_qvec_cfg_msix(struct ice_vsi *vsi, struct ice_q_vector *q_vector)
118 {
119 	u16 reg_idx = q_vector->reg_idx;
120 	struct ice_pf *pf = vsi->back;
121 	struct ice_hw *hw = &pf->hw;
122 	struct ice_tx_ring *tx_ring;
123 	struct ice_rx_ring *rx_ring;
124 
125 	ice_cfg_itr(hw, q_vector);
126 
127 	ice_for_each_tx_ring(tx_ring, q_vector->tx)
128 		ice_cfg_txq_interrupt(vsi, tx_ring->reg_idx, reg_idx,
129 				      q_vector->tx.itr_idx);
130 
131 	ice_for_each_rx_ring(rx_ring, q_vector->rx)
132 		ice_cfg_rxq_interrupt(vsi, rx_ring->reg_idx, reg_idx,
133 				      q_vector->rx.itr_idx);
134 
135 	ice_flush(hw);
136 }
137 
138 /**
139  * ice_qvec_ena_irq - Enable IRQ for given queue vector
140  * @vsi: the VSI that contains queue vector
141  * @q_vector: queue vector
142  */
143 static void ice_qvec_ena_irq(struct ice_vsi *vsi, struct ice_q_vector *q_vector)
144 {
145 	struct ice_pf *pf = vsi->back;
146 	struct ice_hw *hw = &pf->hw;
147 
148 	ice_irq_dynamic_ena(hw, vsi, q_vector);
149 
150 	ice_flush(hw);
151 }
152 
153 /**
154  * ice_qp_dis - Disables a queue pair
155  * @vsi: VSI of interest
156  * @q_idx: ring index in array
157  *
158  * Returns 0 on success, negative on failure.
159  */
160 static int ice_qp_dis(struct ice_vsi *vsi, u16 q_idx)
161 {
162 	struct ice_txq_meta txq_meta = { };
163 	struct ice_q_vector *q_vector;
164 	struct ice_tx_ring *tx_ring;
165 	struct ice_rx_ring *rx_ring;
166 	int timeout = 50;
167 	int err;
168 
169 	if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq)
170 		return -EINVAL;
171 
172 	tx_ring = vsi->tx_rings[q_idx];
173 	rx_ring = vsi->rx_rings[q_idx];
174 	q_vector = rx_ring->q_vector;
175 
176 	while (test_and_set_bit(ICE_CFG_BUSY, vsi->state)) {
177 		timeout--;
178 		if (!timeout)
179 			return -EBUSY;
180 		usleep_range(1000, 2000);
181 	}
182 	netif_tx_stop_queue(netdev_get_tx_queue(vsi->netdev, q_idx));
183 
184 	ice_fill_txq_meta(vsi, tx_ring, &txq_meta);
185 	err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, tx_ring, &txq_meta);
186 	if (err)
187 		return err;
188 	if (ice_is_xdp_ena_vsi(vsi)) {
189 		struct ice_tx_ring *xdp_ring = vsi->xdp_rings[q_idx];
190 
191 		memset(&txq_meta, 0, sizeof(txq_meta));
192 		ice_fill_txq_meta(vsi, xdp_ring, &txq_meta);
193 		err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, xdp_ring,
194 					   &txq_meta);
195 		if (err)
196 			return err;
197 	}
198 	ice_qvec_dis_irq(vsi, rx_ring, q_vector);
199 
200 	err = ice_vsi_ctrl_one_rx_ring(vsi, false, q_idx, true);
201 	if (err)
202 		return err;
203 
204 	ice_qvec_toggle_napi(vsi, q_vector, false);
205 	ice_qp_clean_rings(vsi, q_idx);
206 	ice_qp_reset_stats(vsi, q_idx);
207 
208 	return 0;
209 }
210 
211 /**
212  * ice_qp_ena - Enables a queue pair
213  * @vsi: VSI of interest
214  * @q_idx: ring index in array
215  *
216  * Returns 0 on success, negative on failure.
217  */
218 static int ice_qp_ena(struct ice_vsi *vsi, u16 q_idx)
219 {
220 	struct ice_aqc_add_tx_qgrp *qg_buf;
221 	struct ice_q_vector *q_vector;
222 	struct ice_tx_ring *tx_ring;
223 	struct ice_rx_ring *rx_ring;
224 	u16 size;
225 	int err;
226 
227 	if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq)
228 		return -EINVAL;
229 
230 	size = struct_size(qg_buf, txqs, 1);
231 	qg_buf = kzalloc(size, GFP_KERNEL);
232 	if (!qg_buf)
233 		return -ENOMEM;
234 
235 	qg_buf->num_txqs = 1;
236 
237 	tx_ring = vsi->tx_rings[q_idx];
238 	rx_ring = vsi->rx_rings[q_idx];
239 	q_vector = rx_ring->q_vector;
240 
241 	err = ice_vsi_cfg_txq(vsi, tx_ring, qg_buf);
242 	if (err)
243 		goto free_buf;
244 
245 	if (ice_is_xdp_ena_vsi(vsi)) {
246 		struct ice_tx_ring *xdp_ring = vsi->xdp_rings[q_idx];
247 
248 		memset(qg_buf, 0, size);
249 		qg_buf->num_txqs = 1;
250 		err = ice_vsi_cfg_txq(vsi, xdp_ring, qg_buf);
251 		if (err)
252 			goto free_buf;
253 		ice_set_ring_xdp(xdp_ring);
254 		ice_tx_xsk_pool(vsi, q_idx);
255 	}
256 
257 	err = ice_vsi_cfg_rxq(rx_ring);
258 	if (err)
259 		goto free_buf;
260 
261 	ice_qvec_cfg_msix(vsi, q_vector);
262 
263 	err = ice_vsi_ctrl_one_rx_ring(vsi, true, q_idx, true);
264 	if (err)
265 		goto free_buf;
266 
267 	clear_bit(ICE_CFG_BUSY, vsi->state);
268 	ice_qvec_toggle_napi(vsi, q_vector, true);
269 	ice_qvec_ena_irq(vsi, q_vector);
270 
271 	netif_tx_start_queue(netdev_get_tx_queue(vsi->netdev, q_idx));
272 free_buf:
273 	kfree(qg_buf);
274 	return err;
275 }
276 
277 /**
278  * ice_xsk_pool_disable - disable a buffer pool region
279  * @vsi: Current VSI
280  * @qid: queue ID
281  *
282  * Returns 0 on success, negative on failure
283  */
284 static int ice_xsk_pool_disable(struct ice_vsi *vsi, u16 qid)
285 {
286 	struct xsk_buff_pool *pool = xsk_get_pool_from_qid(vsi->netdev, qid);
287 
288 	if (!pool)
289 		return -EINVAL;
290 
291 	clear_bit(qid, vsi->af_xdp_zc_qps);
292 	xsk_pool_dma_unmap(pool, ICE_RX_DMA_ATTR);
293 
294 	return 0;
295 }
296 
297 /**
298  * ice_xsk_pool_enable - enable a buffer pool region
299  * @vsi: Current VSI
300  * @pool: pointer to a requested buffer pool region
301  * @qid: queue ID
302  *
303  * Returns 0 on success, negative on failure
304  */
305 static int
306 ice_xsk_pool_enable(struct ice_vsi *vsi, struct xsk_buff_pool *pool, u16 qid)
307 {
308 	int err;
309 
310 	if (vsi->type != ICE_VSI_PF)
311 		return -EINVAL;
312 
313 	if (qid >= vsi->netdev->real_num_rx_queues ||
314 	    qid >= vsi->netdev->real_num_tx_queues)
315 		return -EINVAL;
316 
317 	err = xsk_pool_dma_map(pool, ice_pf_to_dev(vsi->back),
318 			       ICE_RX_DMA_ATTR);
319 	if (err)
320 		return err;
321 
322 	set_bit(qid, vsi->af_xdp_zc_qps);
323 
324 	return 0;
325 }
326 
327 /**
328  * ice_realloc_rx_xdp_bufs - reallocate for either XSK or normal buffer
329  * @rx_ring: Rx ring
330  * @pool_present: is pool for XSK present
331  *
332  * Try allocating memory and return ENOMEM, if failed to allocate.
333  * If allocation was successful, substitute buffer with allocated one.
334  * Returns 0 on success, negative on failure
335  */
336 static int
337 ice_realloc_rx_xdp_bufs(struct ice_rx_ring *rx_ring, bool pool_present)
338 {
339 	size_t elem_size = pool_present ? sizeof(*rx_ring->xdp_buf) :
340 					  sizeof(*rx_ring->rx_buf);
341 	void *sw_ring = kcalloc(rx_ring->count, elem_size, GFP_KERNEL);
342 
343 	if (!sw_ring)
344 		return -ENOMEM;
345 
346 	if (pool_present) {
347 		kfree(rx_ring->rx_buf);
348 		rx_ring->rx_buf = NULL;
349 		rx_ring->xdp_buf = sw_ring;
350 	} else {
351 		kfree(rx_ring->xdp_buf);
352 		rx_ring->xdp_buf = NULL;
353 		rx_ring->rx_buf = sw_ring;
354 	}
355 
356 	return 0;
357 }
358 
359 /**
360  * ice_realloc_zc_buf - reallocate XDP ZC queue pairs
361  * @vsi: Current VSI
362  * @zc: is zero copy set
363  *
364  * Reallocate buffer for rx_rings that might be used by XSK.
365  * XDP requires more memory, than rx_buf provides.
366  * Returns 0 on success, negative on failure
367  */
368 int ice_realloc_zc_buf(struct ice_vsi *vsi, bool zc)
369 {
370 	struct ice_rx_ring *rx_ring;
371 	unsigned long q;
372 
373 	for_each_set_bit(q, vsi->af_xdp_zc_qps,
374 			 max_t(int, vsi->alloc_txq, vsi->alloc_rxq)) {
375 		rx_ring = vsi->rx_rings[q];
376 		if (ice_realloc_rx_xdp_bufs(rx_ring, zc))
377 			return -ENOMEM;
378 	}
379 
380 	return 0;
381 }
382 
383 /**
384  * ice_xsk_pool_setup - enable/disable a buffer pool region depending on its state
385  * @vsi: Current VSI
386  * @pool: buffer pool to enable/associate to a ring, NULL to disable
387  * @qid: queue ID
388  *
389  * Returns 0 on success, negative on failure
390  */
391 int ice_xsk_pool_setup(struct ice_vsi *vsi, struct xsk_buff_pool *pool, u16 qid)
392 {
393 	bool if_running, pool_present = !!pool;
394 	int ret = 0, pool_failure = 0;
395 
396 	if (qid >= vsi->num_rxq || qid >= vsi->num_txq) {
397 		netdev_err(vsi->netdev, "Please use queue id in scope of combined queues count\n");
398 		pool_failure = -EINVAL;
399 		goto failure;
400 	}
401 
402 	if_running = netif_running(vsi->netdev) && ice_is_xdp_ena_vsi(vsi);
403 
404 	if (if_running) {
405 		struct ice_rx_ring *rx_ring = vsi->rx_rings[qid];
406 
407 		ret = ice_qp_dis(vsi, qid);
408 		if (ret) {
409 			netdev_err(vsi->netdev, "ice_qp_dis error = %d\n", ret);
410 			goto xsk_pool_if_up;
411 		}
412 
413 		ret = ice_realloc_rx_xdp_bufs(rx_ring, pool_present);
414 		if (ret)
415 			goto xsk_pool_if_up;
416 	}
417 
418 	pool_failure = pool_present ? ice_xsk_pool_enable(vsi, pool, qid) :
419 				      ice_xsk_pool_disable(vsi, qid);
420 
421 xsk_pool_if_up:
422 	if (if_running) {
423 		ret = ice_qp_ena(vsi, qid);
424 		if (!ret && pool_present)
425 			napi_schedule(&vsi->rx_rings[qid]->xdp_ring->q_vector->napi);
426 		else if (ret)
427 			netdev_err(vsi->netdev, "ice_qp_ena error = %d\n", ret);
428 	}
429 
430 failure:
431 	if (pool_failure) {
432 		netdev_err(vsi->netdev, "Could not %sable buffer pool, error = %d\n",
433 			   pool_present ? "en" : "dis", pool_failure);
434 		return pool_failure;
435 	}
436 
437 	return ret;
438 }
439 
440 /**
441  * ice_fill_rx_descs - pick buffers from XSK buffer pool and use it
442  * @pool: XSK Buffer pool to pull the buffers from
443  * @xdp: SW ring of xdp_buff that will hold the buffers
444  * @rx_desc: Pointer to Rx descriptors that will be filled
445  * @count: The number of buffers to allocate
446  *
447  * This function allocates a number of Rx buffers from the fill ring
448  * or the internal recycle mechanism and places them on the Rx ring.
449  *
450  * Note that ring wrap should be handled by caller of this function.
451  *
452  * Returns the amount of allocated Rx descriptors
453  */
454 static u16 ice_fill_rx_descs(struct xsk_buff_pool *pool, struct xdp_buff **xdp,
455 			     union ice_32b_rx_flex_desc *rx_desc, u16 count)
456 {
457 	dma_addr_t dma;
458 	u16 buffs;
459 	int i;
460 
461 	buffs = xsk_buff_alloc_batch(pool, xdp, count);
462 	for (i = 0; i < buffs; i++) {
463 		dma = xsk_buff_xdp_get_dma(*xdp);
464 		rx_desc->read.pkt_addr = cpu_to_le64(dma);
465 		rx_desc->wb.status_error0 = 0;
466 
467 		rx_desc++;
468 		xdp++;
469 	}
470 
471 	return buffs;
472 }
473 
474 /**
475  * __ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
476  * @rx_ring: Rx ring
477  * @count: The number of buffers to allocate
478  *
479  * Place the @count of descriptors onto Rx ring. Handle the ring wrap
480  * for case where space from next_to_use up to the end of ring is less
481  * than @count. Finally do a tail bump.
482  *
483  * Returns true if all allocations were successful, false if any fail.
484  */
485 static bool __ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring, u16 count)
486 {
487 	u32 nb_buffs_extra = 0, nb_buffs = 0;
488 	union ice_32b_rx_flex_desc *rx_desc;
489 	u16 ntu = rx_ring->next_to_use;
490 	u16 total_count = count;
491 	struct xdp_buff **xdp;
492 
493 	rx_desc = ICE_RX_DESC(rx_ring, ntu);
494 	xdp = ice_xdp_buf(rx_ring, ntu);
495 
496 	if (ntu + count >= rx_ring->count) {
497 		nb_buffs_extra = ice_fill_rx_descs(rx_ring->xsk_pool, xdp,
498 						   rx_desc,
499 						   rx_ring->count - ntu);
500 		if (nb_buffs_extra != rx_ring->count - ntu) {
501 			ntu += nb_buffs_extra;
502 			goto exit;
503 		}
504 		rx_desc = ICE_RX_DESC(rx_ring, 0);
505 		xdp = ice_xdp_buf(rx_ring, 0);
506 		ntu = 0;
507 		count -= nb_buffs_extra;
508 		ice_release_rx_desc(rx_ring, 0);
509 	}
510 
511 	nb_buffs = ice_fill_rx_descs(rx_ring->xsk_pool, xdp, rx_desc, count);
512 
513 	ntu += nb_buffs;
514 	if (ntu == rx_ring->count)
515 		ntu = 0;
516 
517 exit:
518 	if (rx_ring->next_to_use != ntu)
519 		ice_release_rx_desc(rx_ring, ntu);
520 
521 	return total_count == (nb_buffs_extra + nb_buffs);
522 }
523 
524 /**
525  * ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
526  * @rx_ring: Rx ring
527  * @count: The number of buffers to allocate
528  *
529  * Wrapper for internal allocation routine; figure out how many tail
530  * bumps should take place based on the given threshold
531  *
532  * Returns true if all calls to internal alloc routine succeeded
533  */
534 bool ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring, u16 count)
535 {
536 	u16 rx_thresh = ICE_RING_QUARTER(rx_ring);
537 	u16 leftover, i, tail_bumps;
538 
539 	tail_bumps = count / rx_thresh;
540 	leftover = count - (tail_bumps * rx_thresh);
541 
542 	for (i = 0; i < tail_bumps; i++)
543 		if (!__ice_alloc_rx_bufs_zc(rx_ring, rx_thresh))
544 			return false;
545 	return __ice_alloc_rx_bufs_zc(rx_ring, leftover);
546 }
547 
548 /**
549  * ice_construct_skb_zc - Create an sk_buff from zero-copy buffer
550  * @rx_ring: Rx ring
551  * @xdp: Pointer to XDP buffer
552  *
553  * This function allocates a new skb from a zero-copy Rx buffer.
554  *
555  * Returns the skb on success, NULL on failure.
556  */
557 static struct sk_buff *
558 ice_construct_skb_zc(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp)
559 {
560 	unsigned int totalsize = xdp->data_end - xdp->data_meta;
561 	unsigned int metasize = xdp->data - xdp->data_meta;
562 	struct skb_shared_info *sinfo = NULL;
563 	struct sk_buff *skb;
564 	u32 nr_frags = 0;
565 
566 	if (unlikely(xdp_buff_has_frags(xdp))) {
567 		sinfo = xdp_get_shared_info_from_buff(xdp);
568 		nr_frags = sinfo->nr_frags;
569 	}
570 	net_prefetch(xdp->data_meta);
571 
572 	skb = __napi_alloc_skb(&rx_ring->q_vector->napi, totalsize,
573 			       GFP_ATOMIC | __GFP_NOWARN);
574 	if (unlikely(!skb))
575 		return NULL;
576 
577 	memcpy(__skb_put(skb, totalsize), xdp->data_meta,
578 	       ALIGN(totalsize, sizeof(long)));
579 
580 	if (metasize) {
581 		skb_metadata_set(skb, metasize);
582 		__skb_pull(skb, metasize);
583 	}
584 
585 	if (likely(!xdp_buff_has_frags(xdp)))
586 		goto out;
587 
588 	for (int i = 0; i < nr_frags; i++) {
589 		struct skb_shared_info *skinfo = skb_shinfo(skb);
590 		skb_frag_t *frag = &sinfo->frags[i];
591 		struct page *page;
592 		void *addr;
593 
594 		page = dev_alloc_page();
595 		if (!page) {
596 			dev_kfree_skb(skb);
597 			return NULL;
598 		}
599 		addr = page_to_virt(page);
600 
601 		memcpy(addr, skb_frag_page(frag), skb_frag_size(frag));
602 
603 		__skb_fill_page_desc_noacc(skinfo, skinfo->nr_frags++,
604 					   addr, 0, skb_frag_size(frag));
605 	}
606 
607 out:
608 	xsk_buff_free(xdp);
609 	return skb;
610 }
611 
612 /**
613  * ice_clean_xdp_irq_zc - produce AF_XDP descriptors to CQ
614  * @xdp_ring: XDP Tx ring
615  */
616 static u32 ice_clean_xdp_irq_zc(struct ice_tx_ring *xdp_ring)
617 {
618 	u16 ntc = xdp_ring->next_to_clean;
619 	struct ice_tx_desc *tx_desc;
620 	u16 cnt = xdp_ring->count;
621 	struct ice_tx_buf *tx_buf;
622 	u16 completed_frames = 0;
623 	u16 xsk_frames = 0;
624 	u16 last_rs;
625 	int i;
626 
627 	last_rs = xdp_ring->next_to_use ? xdp_ring->next_to_use - 1 : cnt - 1;
628 	tx_desc = ICE_TX_DESC(xdp_ring, last_rs);
629 	if (tx_desc->cmd_type_offset_bsz &
630 	    cpu_to_le64(ICE_TX_DESC_DTYPE_DESC_DONE)) {
631 		if (last_rs >= ntc)
632 			completed_frames = last_rs - ntc + 1;
633 		else
634 			completed_frames = last_rs + cnt - ntc + 1;
635 	}
636 
637 	if (!completed_frames)
638 		return 0;
639 
640 	if (likely(!xdp_ring->xdp_tx_active)) {
641 		xsk_frames = completed_frames;
642 		goto skip;
643 	}
644 
645 	ntc = xdp_ring->next_to_clean;
646 	for (i = 0; i < completed_frames; i++) {
647 		tx_buf = &xdp_ring->tx_buf[ntc];
648 
649 		if (tx_buf->type == ICE_TX_BUF_XSK_TX) {
650 			tx_buf->type = ICE_TX_BUF_EMPTY;
651 			xsk_buff_free(tx_buf->xdp);
652 			xdp_ring->xdp_tx_active--;
653 		} else {
654 			xsk_frames++;
655 		}
656 
657 		ntc++;
658 		if (ntc >= xdp_ring->count)
659 			ntc = 0;
660 	}
661 skip:
662 	tx_desc->cmd_type_offset_bsz = 0;
663 	xdp_ring->next_to_clean += completed_frames;
664 	if (xdp_ring->next_to_clean >= cnt)
665 		xdp_ring->next_to_clean -= cnt;
666 	if (xsk_frames)
667 		xsk_tx_completed(xdp_ring->xsk_pool, xsk_frames);
668 
669 	return completed_frames;
670 }
671 
672 /**
673  * ice_xmit_xdp_tx_zc - AF_XDP ZC handler for XDP_TX
674  * @xdp: XDP buffer to xmit
675  * @xdp_ring: XDP ring to produce descriptor onto
676  *
677  * note that this function works directly on xdp_buff, no need to convert
678  * it to xdp_frame. xdp_buff pointer is stored to ice_tx_buf so that cleaning
679  * side will be able to xsk_buff_free() it.
680  *
681  * Returns ICE_XDP_TX for successfully produced desc, ICE_XDP_CONSUMED if there
682  * was not enough space on XDP ring
683  */
684 static int ice_xmit_xdp_tx_zc(struct xdp_buff *xdp,
685 			      struct ice_tx_ring *xdp_ring)
686 {
687 	struct skb_shared_info *sinfo = NULL;
688 	u32 size = xdp->data_end - xdp->data;
689 	u32 ntu = xdp_ring->next_to_use;
690 	struct ice_tx_desc *tx_desc;
691 	struct ice_tx_buf *tx_buf;
692 	struct xdp_buff *head;
693 	u32 nr_frags = 0;
694 	u32 free_space;
695 	u32 frag = 0;
696 
697 	free_space = ICE_DESC_UNUSED(xdp_ring);
698 	if (free_space < ICE_RING_QUARTER(xdp_ring))
699 		free_space += ice_clean_xdp_irq_zc(xdp_ring);
700 
701 	if (unlikely(!free_space))
702 		goto busy;
703 
704 	if (unlikely(xdp_buff_has_frags(xdp))) {
705 		sinfo = xdp_get_shared_info_from_buff(xdp);
706 		nr_frags = sinfo->nr_frags;
707 		if (free_space < nr_frags + 1)
708 			goto busy;
709 	}
710 
711 	tx_desc = ICE_TX_DESC(xdp_ring, ntu);
712 	tx_buf = &xdp_ring->tx_buf[ntu];
713 	head = xdp;
714 
715 	for (;;) {
716 		dma_addr_t dma;
717 
718 		dma = xsk_buff_xdp_get_dma(xdp);
719 		xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, size);
720 
721 		tx_buf->xdp = xdp;
722 		tx_buf->type = ICE_TX_BUF_XSK_TX;
723 		tx_desc->buf_addr = cpu_to_le64(dma);
724 		tx_desc->cmd_type_offset_bsz = ice_build_ctob(0, 0, size, 0);
725 		/* account for each xdp_buff from xsk_buff_pool */
726 		xdp_ring->xdp_tx_active++;
727 
728 		if (++ntu == xdp_ring->count)
729 			ntu = 0;
730 
731 		if (frag == nr_frags)
732 			break;
733 
734 		tx_desc = ICE_TX_DESC(xdp_ring, ntu);
735 		tx_buf = &xdp_ring->tx_buf[ntu];
736 
737 		xdp = xsk_buff_get_frag(head);
738 		size = skb_frag_size(&sinfo->frags[frag]);
739 		frag++;
740 	}
741 
742 	xdp_ring->next_to_use = ntu;
743 	/* update last descriptor from a frame with EOP */
744 	tx_desc->cmd_type_offset_bsz |=
745 		cpu_to_le64(ICE_TX_DESC_CMD_EOP << ICE_TXD_QW1_CMD_S);
746 
747 	return ICE_XDP_TX;
748 
749 busy:
750 	xdp_ring->ring_stats->tx_stats.tx_busy++;
751 
752 	return ICE_XDP_CONSUMED;
753 }
754 
755 /**
756  * ice_run_xdp_zc - Executes an XDP program in zero-copy path
757  * @rx_ring: Rx ring
758  * @xdp: xdp_buff used as input to the XDP program
759  * @xdp_prog: XDP program to run
760  * @xdp_ring: ring to be used for XDP_TX action
761  *
762  * Returns any of ICE_XDP_{PASS, CONSUMED, TX, REDIR}
763  */
764 static int
765 ice_run_xdp_zc(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp,
766 	       struct bpf_prog *xdp_prog, struct ice_tx_ring *xdp_ring)
767 {
768 	int err, result = ICE_XDP_PASS;
769 	u32 act;
770 
771 	act = bpf_prog_run_xdp(xdp_prog, xdp);
772 
773 	if (likely(act == XDP_REDIRECT)) {
774 		err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
775 		if (!err)
776 			return ICE_XDP_REDIR;
777 		if (xsk_uses_need_wakeup(rx_ring->xsk_pool) && err == -ENOBUFS)
778 			result = ICE_XDP_EXIT;
779 		else
780 			result = ICE_XDP_CONSUMED;
781 		goto out_failure;
782 	}
783 
784 	switch (act) {
785 	case XDP_PASS:
786 		break;
787 	case XDP_TX:
788 		result = ice_xmit_xdp_tx_zc(xdp, xdp_ring);
789 		if (result == ICE_XDP_CONSUMED)
790 			goto out_failure;
791 		break;
792 	case XDP_DROP:
793 		result = ICE_XDP_CONSUMED;
794 		break;
795 	default:
796 		bpf_warn_invalid_xdp_action(rx_ring->netdev, xdp_prog, act);
797 		fallthrough;
798 	case XDP_ABORTED:
799 		result = ICE_XDP_CONSUMED;
800 out_failure:
801 		trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
802 		break;
803 	}
804 
805 	return result;
806 }
807 
808 static int
809 ice_add_xsk_frag(struct ice_rx_ring *rx_ring, struct xdp_buff *first,
810 		 struct xdp_buff *xdp, const unsigned int size)
811 {
812 	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(first);
813 
814 	if (!size)
815 		return 0;
816 
817 	if (!xdp_buff_has_frags(first)) {
818 		sinfo->nr_frags = 0;
819 		sinfo->xdp_frags_size = 0;
820 		xdp_buff_set_frags_flag(first);
821 	}
822 
823 	if (unlikely(sinfo->nr_frags == MAX_SKB_FRAGS)) {
824 		xsk_buff_free(first);
825 		return -ENOMEM;
826 	}
827 
828 	__skb_fill_page_desc_noacc(sinfo, sinfo->nr_frags++,
829 				   virt_to_page(xdp->data_hard_start), 0, size);
830 	sinfo->xdp_frags_size += size;
831 	xsk_buff_add_frag(xdp);
832 
833 	return 0;
834 }
835 
836 /**
837  * ice_clean_rx_irq_zc - consumes packets from the hardware ring
838  * @rx_ring: AF_XDP Rx ring
839  * @budget: NAPI budget
840  *
841  * Returns number of processed packets on success, remaining budget on failure.
842  */
843 int ice_clean_rx_irq_zc(struct ice_rx_ring *rx_ring, int budget)
844 {
845 	unsigned int total_rx_bytes = 0, total_rx_packets = 0;
846 	struct xsk_buff_pool *xsk_pool = rx_ring->xsk_pool;
847 	u32 ntc = rx_ring->next_to_clean;
848 	u32 ntu = rx_ring->next_to_use;
849 	struct xdp_buff *first = NULL;
850 	struct ice_tx_ring *xdp_ring;
851 	unsigned int xdp_xmit = 0;
852 	struct bpf_prog *xdp_prog;
853 	u32 cnt = rx_ring->count;
854 	bool failure = false;
855 	int entries_to_alloc;
856 
857 	/* ZC patch is enabled only when XDP program is set,
858 	 * so here it can not be NULL
859 	 */
860 	xdp_prog = READ_ONCE(rx_ring->xdp_prog);
861 	xdp_ring = rx_ring->xdp_ring;
862 
863 	if (ntc != rx_ring->first_desc)
864 		first = *ice_xdp_buf(rx_ring, rx_ring->first_desc);
865 
866 	while (likely(total_rx_packets < (unsigned int)budget)) {
867 		union ice_32b_rx_flex_desc *rx_desc;
868 		unsigned int size, xdp_res = 0;
869 		struct xdp_buff *xdp;
870 		struct sk_buff *skb;
871 		u16 stat_err_bits;
872 		u16 vlan_tag = 0;
873 		u16 rx_ptype;
874 
875 		rx_desc = ICE_RX_DESC(rx_ring, ntc);
876 
877 		stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_DD_S);
878 		if (!ice_test_staterr(rx_desc->wb.status_error0, stat_err_bits))
879 			break;
880 
881 		/* This memory barrier is needed to keep us from reading
882 		 * any other fields out of the rx_desc until we have
883 		 * verified the descriptor has been written back.
884 		 */
885 		dma_rmb();
886 
887 		if (unlikely(ntc == ntu))
888 			break;
889 
890 		xdp = *ice_xdp_buf(rx_ring, ntc);
891 
892 		size = le16_to_cpu(rx_desc->wb.pkt_len) &
893 				   ICE_RX_FLX_DESC_PKT_LEN_M;
894 
895 		xsk_buff_set_size(xdp, size);
896 		xsk_buff_dma_sync_for_cpu(xdp, xsk_pool);
897 
898 		if (!first) {
899 			first = xdp;
900 			xdp_buff_clear_frags_flag(first);
901 		} else if (ice_add_xsk_frag(rx_ring, first, xdp, size)) {
902 			break;
903 		}
904 
905 		if (++ntc == cnt)
906 			ntc = 0;
907 
908 		if (ice_is_non_eop(rx_ring, rx_desc))
909 			continue;
910 
911 		xdp_res = ice_run_xdp_zc(rx_ring, first, xdp_prog, xdp_ring);
912 		if (likely(xdp_res & (ICE_XDP_TX | ICE_XDP_REDIR))) {
913 			xdp_xmit |= xdp_res;
914 		} else if (xdp_res == ICE_XDP_EXIT) {
915 			failure = true;
916 			first = NULL;
917 			rx_ring->first_desc = ntc;
918 			break;
919 		} else if (xdp_res == ICE_XDP_CONSUMED) {
920 			xsk_buff_free(first);
921 		} else if (xdp_res == ICE_XDP_PASS) {
922 			goto construct_skb;
923 		}
924 
925 		total_rx_bytes += xdp_get_buff_len(first);
926 		total_rx_packets++;
927 
928 		first = NULL;
929 		rx_ring->first_desc = ntc;
930 		continue;
931 
932 construct_skb:
933 		/* XDP_PASS path */
934 		skb = ice_construct_skb_zc(rx_ring, first);
935 		if (!skb) {
936 			rx_ring->ring_stats->rx_stats.alloc_buf_failed++;
937 			break;
938 		}
939 
940 		first = NULL;
941 		rx_ring->first_desc = ntc;
942 
943 		if (eth_skb_pad(skb)) {
944 			skb = NULL;
945 			continue;
946 		}
947 
948 		total_rx_bytes += skb->len;
949 		total_rx_packets++;
950 
951 		vlan_tag = ice_get_vlan_tag_from_rx_desc(rx_desc);
952 
953 		rx_ptype = le16_to_cpu(rx_desc->wb.ptype_flex_flags0) &
954 				       ICE_RX_FLEX_DESC_PTYPE_M;
955 
956 		ice_process_skb_fields(rx_ring, rx_desc, skb, rx_ptype);
957 		ice_receive_skb(rx_ring, skb, vlan_tag);
958 	}
959 
960 	rx_ring->next_to_clean = ntc;
961 	entries_to_alloc = ICE_RX_DESC_UNUSED(rx_ring);
962 	if (entries_to_alloc > ICE_RING_QUARTER(rx_ring))
963 		failure |= !ice_alloc_rx_bufs_zc(rx_ring, entries_to_alloc);
964 
965 	ice_finalize_xdp_rx(xdp_ring, xdp_xmit, 0);
966 	ice_update_rx_ring_stats(rx_ring, total_rx_packets, total_rx_bytes);
967 
968 	if (xsk_uses_need_wakeup(xsk_pool)) {
969 		/* ntu could have changed when allocating entries above, so
970 		 * use rx_ring value instead of stack based one
971 		 */
972 		if (failure || ntc == rx_ring->next_to_use)
973 			xsk_set_rx_need_wakeup(xsk_pool);
974 		else
975 			xsk_clear_rx_need_wakeup(xsk_pool);
976 
977 		return (int)total_rx_packets;
978 	}
979 
980 	return failure ? budget : (int)total_rx_packets;
981 }
982 
983 /**
984  * ice_xmit_pkt - produce a single HW Tx descriptor out of AF_XDP descriptor
985  * @xdp_ring: XDP ring to produce the HW Tx descriptor on
986  * @desc: AF_XDP descriptor to pull the DMA address and length from
987  * @total_bytes: bytes accumulator that will be used for stats update
988  */
989 static void ice_xmit_pkt(struct ice_tx_ring *xdp_ring, struct xdp_desc *desc,
990 			 unsigned int *total_bytes)
991 {
992 	struct ice_tx_desc *tx_desc;
993 	dma_addr_t dma;
994 
995 	dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc->addr);
996 	xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc->len);
997 
998 	tx_desc = ICE_TX_DESC(xdp_ring, xdp_ring->next_to_use++);
999 	tx_desc->buf_addr = cpu_to_le64(dma);
1000 	tx_desc->cmd_type_offset_bsz = ice_build_ctob(xsk_is_eop_desc(desc),
1001 						      0, desc->len, 0);
1002 
1003 	*total_bytes += desc->len;
1004 }
1005 
1006 /**
1007  * ice_xmit_pkt_batch - produce a batch of HW Tx descriptors out of AF_XDP descriptors
1008  * @xdp_ring: XDP ring to produce the HW Tx descriptors on
1009  * @descs: AF_XDP descriptors to pull the DMA addresses and lengths from
1010  * @total_bytes: bytes accumulator that will be used for stats update
1011  */
1012 static void ice_xmit_pkt_batch(struct ice_tx_ring *xdp_ring, struct xdp_desc *descs,
1013 			       unsigned int *total_bytes)
1014 {
1015 	u16 ntu = xdp_ring->next_to_use;
1016 	struct ice_tx_desc *tx_desc;
1017 	u32 i;
1018 
1019 	loop_unrolled_for(i = 0; i < PKTS_PER_BATCH; i++) {
1020 		dma_addr_t dma;
1021 
1022 		dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, descs[i].addr);
1023 		xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, descs[i].len);
1024 
1025 		tx_desc = ICE_TX_DESC(xdp_ring, ntu++);
1026 		tx_desc->buf_addr = cpu_to_le64(dma);
1027 		tx_desc->cmd_type_offset_bsz = ice_build_ctob(xsk_is_eop_desc(&descs[i]),
1028 							      0, descs[i].len, 0);
1029 
1030 		*total_bytes += descs[i].len;
1031 	}
1032 
1033 	xdp_ring->next_to_use = ntu;
1034 }
1035 
1036 /**
1037  * ice_fill_tx_hw_ring - produce the number of Tx descriptors onto ring
1038  * @xdp_ring: XDP ring to produce the HW Tx descriptors on
1039  * @descs: AF_XDP descriptors to pull the DMA addresses and lengths from
1040  * @nb_pkts: count of packets to be send
1041  * @total_bytes: bytes accumulator that will be used for stats update
1042  */
1043 static void ice_fill_tx_hw_ring(struct ice_tx_ring *xdp_ring, struct xdp_desc *descs,
1044 				u32 nb_pkts, unsigned int *total_bytes)
1045 {
1046 	u32 batched, leftover, i;
1047 
1048 	batched = ALIGN_DOWN(nb_pkts, PKTS_PER_BATCH);
1049 	leftover = nb_pkts & (PKTS_PER_BATCH - 1);
1050 	for (i = 0; i < batched; i += PKTS_PER_BATCH)
1051 		ice_xmit_pkt_batch(xdp_ring, &descs[i], total_bytes);
1052 	for (; i < batched + leftover; i++)
1053 		ice_xmit_pkt(xdp_ring, &descs[i], total_bytes);
1054 }
1055 
1056 /**
1057  * ice_xmit_zc - take entries from XSK Tx ring and place them onto HW Tx ring
1058  * @xdp_ring: XDP ring to produce the HW Tx descriptors on
1059  *
1060  * Returns true if there is no more work that needs to be done, false otherwise
1061  */
1062 bool ice_xmit_zc(struct ice_tx_ring *xdp_ring)
1063 {
1064 	struct xdp_desc *descs = xdp_ring->xsk_pool->tx_descs;
1065 	u32 nb_pkts, nb_processed = 0;
1066 	unsigned int total_bytes = 0;
1067 	int budget;
1068 
1069 	ice_clean_xdp_irq_zc(xdp_ring);
1070 
1071 	budget = ICE_DESC_UNUSED(xdp_ring);
1072 	budget = min_t(u16, budget, ICE_RING_QUARTER(xdp_ring));
1073 
1074 	nb_pkts = xsk_tx_peek_release_desc_batch(xdp_ring->xsk_pool, budget);
1075 	if (!nb_pkts)
1076 		return true;
1077 
1078 	if (xdp_ring->next_to_use + nb_pkts >= xdp_ring->count) {
1079 		nb_processed = xdp_ring->count - xdp_ring->next_to_use;
1080 		ice_fill_tx_hw_ring(xdp_ring, descs, nb_processed, &total_bytes);
1081 		xdp_ring->next_to_use = 0;
1082 	}
1083 
1084 	ice_fill_tx_hw_ring(xdp_ring, &descs[nb_processed], nb_pkts - nb_processed,
1085 			    &total_bytes);
1086 
1087 	ice_set_rs_bit(xdp_ring);
1088 	ice_xdp_ring_update_tail(xdp_ring);
1089 	ice_update_tx_ring_stats(xdp_ring, nb_pkts, total_bytes);
1090 
1091 	if (xsk_uses_need_wakeup(xdp_ring->xsk_pool))
1092 		xsk_set_tx_need_wakeup(xdp_ring->xsk_pool);
1093 
1094 	return nb_pkts < budget;
1095 }
1096 
1097 /**
1098  * ice_xsk_wakeup - Implements ndo_xsk_wakeup
1099  * @netdev: net_device
1100  * @queue_id: queue to wake up
1101  * @flags: ignored in our case, since we have Rx and Tx in the same NAPI
1102  *
1103  * Returns negative on error, zero otherwise.
1104  */
1105 int
1106 ice_xsk_wakeup(struct net_device *netdev, u32 queue_id,
1107 	       u32 __always_unused flags)
1108 {
1109 	struct ice_netdev_priv *np = netdev_priv(netdev);
1110 	struct ice_q_vector *q_vector;
1111 	struct ice_vsi *vsi = np->vsi;
1112 	struct ice_tx_ring *ring;
1113 
1114 	if (test_bit(ICE_VSI_DOWN, vsi->state))
1115 		return -ENETDOWN;
1116 
1117 	if (!ice_is_xdp_ena_vsi(vsi))
1118 		return -EINVAL;
1119 
1120 	if (queue_id >= vsi->num_txq || queue_id >= vsi->num_rxq)
1121 		return -EINVAL;
1122 
1123 	ring = vsi->rx_rings[queue_id]->xdp_ring;
1124 
1125 	if (!ring->xsk_pool)
1126 		return -EINVAL;
1127 
1128 	/* The idea here is that if NAPI is running, mark a miss, so
1129 	 * it will run again. If not, trigger an interrupt and
1130 	 * schedule the NAPI from interrupt context. If NAPI would be
1131 	 * scheduled here, the interrupt affinity would not be
1132 	 * honored.
1133 	 */
1134 	q_vector = ring->q_vector;
1135 	if (!napi_if_scheduled_mark_missed(&q_vector->napi))
1136 		ice_trigger_sw_intr(&vsi->back->hw, q_vector);
1137 
1138 	return 0;
1139 }
1140 
1141 /**
1142  * ice_xsk_any_rx_ring_ena - Checks if Rx rings have AF_XDP buff pool attached
1143  * @vsi: VSI to be checked
1144  *
1145  * Returns true if any of the Rx rings has an AF_XDP buff pool attached
1146  */
1147 bool ice_xsk_any_rx_ring_ena(struct ice_vsi *vsi)
1148 {
1149 	int i;
1150 
1151 	ice_for_each_rxq(vsi, i) {
1152 		if (xsk_get_pool_from_qid(vsi->netdev, i))
1153 			return true;
1154 	}
1155 
1156 	return false;
1157 }
1158 
1159 /**
1160  * ice_xsk_clean_rx_ring - clean buffer pool queues connected to a given Rx ring
1161  * @rx_ring: ring to be cleaned
1162  */
1163 void ice_xsk_clean_rx_ring(struct ice_rx_ring *rx_ring)
1164 {
1165 	u16 ntc = rx_ring->next_to_clean;
1166 	u16 ntu = rx_ring->next_to_use;
1167 
1168 	while (ntc != ntu) {
1169 		struct xdp_buff *xdp = *ice_xdp_buf(rx_ring, ntc);
1170 
1171 		xsk_buff_free(xdp);
1172 		ntc++;
1173 		if (ntc >= rx_ring->count)
1174 			ntc = 0;
1175 	}
1176 }
1177 
1178 /**
1179  * ice_xsk_clean_xdp_ring - Clean the XDP Tx ring and its buffer pool queues
1180  * @xdp_ring: XDP_Tx ring
1181  */
1182 void ice_xsk_clean_xdp_ring(struct ice_tx_ring *xdp_ring)
1183 {
1184 	u16 ntc = xdp_ring->next_to_clean, ntu = xdp_ring->next_to_use;
1185 	u32 xsk_frames = 0;
1186 
1187 	while (ntc != ntu) {
1188 		struct ice_tx_buf *tx_buf = &xdp_ring->tx_buf[ntc];
1189 
1190 		if (tx_buf->type == ICE_TX_BUF_XSK_TX) {
1191 			tx_buf->type = ICE_TX_BUF_EMPTY;
1192 			xsk_buff_free(tx_buf->xdp);
1193 		} else {
1194 			xsk_frames++;
1195 		}
1196 
1197 		ntc++;
1198 		if (ntc >= xdp_ring->count)
1199 			ntc = 0;
1200 	}
1201 
1202 	if (xsk_frames)
1203 		xsk_tx_completed(xdp_ring->xsk_pool, xsk_frames);
1204 }
1205