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

source code of linux/drivers/net/ethernet/intel/ice/ice_xsk.c