Loading...
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(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 */
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(&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 */
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 int base = vsi->base_vector;
94 u16 reg;
95 u32 val;
96
97 /* QINT_TQCTL is being cleared in ice_vsi_stop_tx_ring, so handle
98 * here only QINT_RQCTL
99 */
100 reg = rx_ring->reg_idx;
101 val = rd32(hw, QINT_RQCTL(reg));
102 val &= ~QINT_RQCTL_CAUSE_ENA_M;
103 wr32(hw, QINT_RQCTL(reg), val);
104
105 if (q_vector) {
106 u16 v_idx = q_vector->v_idx;
107
108 wr32(hw, GLINT_DYN_CTL(q_vector->reg_idx), 0);
109 ice_flush(hw);
110 synchronize_irq(pf->msix_entries[v_idx + base].vector);
111 }
112}
113
114/**
115 * ice_qvec_cfg_msix - Enable IRQ for given queue vector
116 * @vsi: the VSI that contains queue vector
117 * @q_vector: queue vector
118 */
119static void
120ice_qvec_cfg_msix(struct ice_vsi *vsi, struct ice_q_vector *q_vector)
121{
122 u16 reg_idx = q_vector->reg_idx;
123 struct ice_pf *pf = vsi->back;
124 struct ice_hw *hw = &pf->hw;
125 struct ice_tx_ring *tx_ring;
126 struct ice_rx_ring *rx_ring;
127
128 ice_cfg_itr(hw, q_vector);
129
130 ice_for_each_tx_ring(tx_ring, q_vector->tx)
131 ice_cfg_txq_interrupt(vsi, tx_ring->reg_idx, reg_idx,
132 q_vector->tx.itr_idx);
133
134 ice_for_each_rx_ring(rx_ring, q_vector->rx)
135 ice_cfg_rxq_interrupt(vsi, rx_ring->reg_idx, reg_idx,
136 q_vector->rx.itr_idx);
137
138 ice_flush(hw);
139}
140
141/**
142 * ice_qvec_ena_irq - Enable IRQ for given queue vector
143 * @vsi: the VSI that contains queue vector
144 * @q_vector: queue vector
145 */
146static void ice_qvec_ena_irq(struct ice_vsi *vsi, struct ice_q_vector *q_vector)
147{
148 struct ice_pf *pf = vsi->back;
149 struct ice_hw *hw = &pf->hw;
150
151 ice_irq_dynamic_ena(hw, vsi, q_vector);
152
153 ice_flush(hw);
154}
155
156/**
157 * ice_qp_dis - Disables a queue pair
158 * @vsi: VSI of interest
159 * @q_idx: ring index in array
160 *
161 * Returns 0 on success, negative on failure.
162 */
163static int ice_qp_dis(struct ice_vsi *vsi, u16 q_idx)
164{
165 struct ice_txq_meta txq_meta = { };
166 struct ice_q_vector *q_vector;
167 struct ice_tx_ring *tx_ring;
168 struct ice_rx_ring *rx_ring;
169 int timeout = 50;
170 int err;
171
172 if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq)
173 return -EINVAL;
174
175 tx_ring = vsi->tx_rings[q_idx];
176 rx_ring = vsi->rx_rings[q_idx];
177 q_vector = rx_ring->q_vector;
178
179 while (test_and_set_bit(ICE_CFG_BUSY, vsi->state)) {
180 timeout--;
181 if (!timeout)
182 return -EBUSY;
183 usleep_range(1000, 2000);
184 }
185 netif_tx_stop_queue(netdev_get_tx_queue(vsi->netdev, q_idx));
186
187 ice_qvec_dis_irq(vsi, rx_ring, q_vector);
188
189 ice_fill_txq_meta(vsi, tx_ring, &txq_meta);
190 err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, tx_ring, &txq_meta);
191 if (err)
192 return err;
193 if (ice_is_xdp_ena_vsi(vsi)) {
194 struct ice_tx_ring *xdp_ring = vsi->xdp_rings[q_idx];
195
196 memset(&txq_meta, 0, sizeof(txq_meta));
197 ice_fill_txq_meta(vsi, xdp_ring, &txq_meta);
198 err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, xdp_ring,
199 &txq_meta);
200 if (err)
201 return err;
202 }
203 err = ice_vsi_ctrl_one_rx_ring(vsi, false, q_idx, true);
204 if (err)
205 return err;
206 ice_clean_rx_ring(rx_ring);
207
208 ice_qvec_toggle_napi(vsi, q_vector, false);
209 ice_qp_clean_rings(vsi, q_idx);
210 ice_qp_reset_stats(vsi, q_idx);
211
212 return 0;
213}
214
215/**
216 * ice_qp_ena - Enables a queue pair
217 * @vsi: VSI of interest
218 * @q_idx: ring index in array
219 *
220 * Returns 0 on success, negative on failure.
221 */
222static int ice_qp_ena(struct ice_vsi *vsi, u16 q_idx)
223{
224 struct ice_aqc_add_tx_qgrp *qg_buf;
225 struct ice_q_vector *q_vector;
226 struct ice_tx_ring *tx_ring;
227 struct ice_rx_ring *rx_ring;
228 u16 size;
229 int err;
230
231 if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq)
232 return -EINVAL;
233
234 size = struct_size(qg_buf, txqs, 1);
235 qg_buf = kzalloc(size, GFP_KERNEL);
236 if (!qg_buf)
237 return -ENOMEM;
238
239 qg_buf->num_txqs = 1;
240
241 tx_ring = vsi->tx_rings[q_idx];
242 rx_ring = vsi->rx_rings[q_idx];
243 q_vector = rx_ring->q_vector;
244
245 err = ice_vsi_cfg_txq(vsi, tx_ring, qg_buf);
246 if (err)
247 goto free_buf;
248
249 if (ice_is_xdp_ena_vsi(vsi)) {
250 struct ice_tx_ring *xdp_ring = vsi->xdp_rings[q_idx];
251
252 memset(qg_buf, 0, size);
253 qg_buf->num_txqs = 1;
254 err = ice_vsi_cfg_txq(vsi, xdp_ring, qg_buf);
255 if (err)
256 goto free_buf;
257 ice_set_ring_xdp(xdp_ring);
258 ice_tx_xsk_pool(vsi, q_idx);
259 }
260
261 err = ice_vsi_cfg_rxq(rx_ring);
262 if (err)
263 goto free_buf;
264
265 ice_qvec_cfg_msix(vsi, q_vector);
266
267 err = ice_vsi_ctrl_one_rx_ring(vsi, true, q_idx, true);
268 if (err)
269 goto free_buf;
270
271 clear_bit(ICE_CFG_BUSY, vsi->state);
272 ice_qvec_toggle_napi(vsi, q_vector, true);
273 ice_qvec_ena_irq(vsi, q_vector);
274
275 netif_tx_start_queue(netdev_get_tx_queue(vsi->netdev, q_idx));
276free_buf:
277 kfree(qg_buf);
278 return err;
279}
280
281/**
282 * ice_xsk_pool_disable - disable a buffer pool region
283 * @vsi: Current VSI
284 * @qid: queue ID
285 *
286 * Returns 0 on success, negative on failure
287 */
288static int ice_xsk_pool_disable(struct ice_vsi *vsi, u16 qid)
289{
290 struct xsk_buff_pool *pool = xsk_get_pool_from_qid(vsi->netdev, qid);
291
292 if (!pool)
293 return -EINVAL;
294
295 clear_bit(qid, vsi->af_xdp_zc_qps);
296 xsk_pool_dma_unmap(pool, ICE_RX_DMA_ATTR);
297
298 return 0;
299}
300
301/**
302 * ice_xsk_pool_enable - enable a buffer pool region
303 * @vsi: Current VSI
304 * @pool: pointer to a requested buffer pool region
305 * @qid: queue ID
306 *
307 * Returns 0 on success, negative on failure
308 */
309static int
310ice_xsk_pool_enable(struct ice_vsi *vsi, struct xsk_buff_pool *pool, u16 qid)
311{
312 int err;
313
314 if (vsi->type != ICE_VSI_PF)
315 return -EINVAL;
316
317 if (qid >= vsi->netdev->real_num_rx_queues ||
318 qid >= vsi->netdev->real_num_tx_queues)
319 return -EINVAL;
320
321 err = xsk_pool_dma_map(pool, ice_pf_to_dev(vsi->back),
322 ICE_RX_DMA_ATTR);
323 if (err)
324 return err;
325
326 set_bit(qid, vsi->af_xdp_zc_qps);
327
328 return 0;
329}
330
331/**
332 * ice_realloc_rx_xdp_bufs - reallocate for either XSK or normal buffer
333 * @rx_ring: Rx ring
334 * @pool_present: is pool for XSK present
335 *
336 * Try allocating memory and return ENOMEM, if failed to allocate.
337 * If allocation was successful, substitute buffer with allocated one.
338 * Returns 0 on success, negative on failure
339 */
340static int
341ice_realloc_rx_xdp_bufs(struct ice_rx_ring *rx_ring, bool pool_present)
342{
343 size_t elem_size = pool_present ? sizeof(*rx_ring->xdp_buf) :
344 sizeof(*rx_ring->rx_buf);
345 void *sw_ring = kcalloc(rx_ring->count, elem_size, GFP_KERNEL);
346
347 if (!sw_ring)
348 return -ENOMEM;
349
350 if (pool_present) {
351 kfree(rx_ring->rx_buf);
352 rx_ring->rx_buf = NULL;
353 rx_ring->xdp_buf = sw_ring;
354 } else {
355 kfree(rx_ring->xdp_buf);
356 rx_ring->xdp_buf = NULL;
357 rx_ring->rx_buf = sw_ring;
358 }
359
360 return 0;
361}
362
363/**
364 * ice_realloc_zc_buf - reallocate XDP ZC queue pairs
365 * @vsi: Current VSI
366 * @zc: is zero copy set
367 *
368 * Reallocate buffer for rx_rings that might be used by XSK.
369 * XDP requires more memory, than rx_buf provides.
370 * Returns 0 on success, negative on failure
371 */
372int ice_realloc_zc_buf(struct ice_vsi *vsi, bool zc)
373{
374 struct ice_rx_ring *rx_ring;
375 unsigned long q;
376
377 for_each_set_bit(q, vsi->af_xdp_zc_qps,
378 max_t(int, vsi->alloc_txq, vsi->alloc_rxq)) {
379 rx_ring = vsi->rx_rings[q];
380 if (ice_realloc_rx_xdp_bufs(rx_ring, zc))
381 return -ENOMEM;
382 }
383
384 return 0;
385}
386
387/**
388 * ice_xsk_pool_setup - enable/disable a buffer pool region depending on its state
389 * @vsi: Current VSI
390 * @pool: buffer pool to enable/associate to a ring, NULL to disable
391 * @qid: queue ID
392 *
393 * Returns 0 on success, negative on failure
394 */
395int ice_xsk_pool_setup(struct ice_vsi *vsi, struct xsk_buff_pool *pool, u16 qid)
396{
397 bool if_running, pool_present = !!pool;
398 int ret = 0, pool_failure = 0;
399
400 if (qid >= vsi->num_rxq || qid >= vsi->num_txq) {
401 netdev_err(vsi->netdev, "Please use queue id in scope of combined queues count\n");
402 pool_failure = -EINVAL;
403 goto failure;
404 }
405
406 if_running = netif_running(vsi->netdev) && ice_is_xdp_ena_vsi(vsi);
407
408 if (if_running) {
409 struct ice_rx_ring *rx_ring = vsi->rx_rings[qid];
410
411 ret = ice_qp_dis(vsi, qid);
412 if (ret) {
413 netdev_err(vsi->netdev, "ice_qp_dis error = %d\n", ret);
414 goto xsk_pool_if_up;
415 }
416
417 ret = ice_realloc_rx_xdp_bufs(rx_ring, pool_present);
418 if (ret)
419 goto xsk_pool_if_up;
420 }
421
422 pool_failure = pool_present ? ice_xsk_pool_enable(vsi, pool, qid) :
423 ice_xsk_pool_disable(vsi, qid);
424
425xsk_pool_if_up:
426 if (if_running) {
427 ret = ice_qp_ena(vsi, qid);
428 if (!ret && pool_present)
429 napi_schedule(&vsi->rx_rings[qid]->xdp_ring->q_vector->napi);
430 else if (ret)
431 netdev_err(vsi->netdev, "ice_qp_ena error = %d\n", ret);
432 }
433
434failure:
435 if (pool_failure) {
436 netdev_err(vsi->netdev, "Could not %sable buffer pool, error = %d\n",
437 pool_present ? "en" : "dis", pool_failure);
438 return pool_failure;
439 }
440
441 return ret;
442}
443
444/**
445 * ice_fill_rx_descs - pick buffers from XSK buffer pool and use it
446 * @pool: XSK Buffer pool to pull the buffers from
447 * @xdp: SW ring of xdp_buff that will hold the buffers
448 * @rx_desc: Pointer to Rx descriptors that will be filled
449 * @count: The number of buffers to allocate
450 *
451 * This function allocates a number of Rx buffers from the fill ring
452 * or the internal recycle mechanism and places them on the Rx ring.
453 *
454 * Note that ring wrap should be handled by caller of this function.
455 *
456 * Returns the amount of allocated Rx descriptors
457 */
458static u16 ice_fill_rx_descs(struct xsk_buff_pool *pool, struct xdp_buff **xdp,
459 union ice_32b_rx_flex_desc *rx_desc, u16 count)
460{
461 dma_addr_t dma;
462 u16 buffs;
463 int i;
464
465 buffs = xsk_buff_alloc_batch(pool, xdp, count);
466 for (i = 0; i < buffs; i++) {
467 dma = xsk_buff_xdp_get_dma(*xdp);
468 rx_desc->read.pkt_addr = cpu_to_le64(dma);
469 rx_desc->wb.status_error0 = 0;
470
471 rx_desc++;
472 xdp++;
473 }
474
475 return buffs;
476}
477
478/**
479 * __ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
480 * @rx_ring: Rx ring
481 * @count: The number of buffers to allocate
482 *
483 * Place the @count of descriptors onto Rx ring. Handle the ring wrap
484 * for case where space from next_to_use up to the end of ring is less
485 * than @count. Finally do a tail bump.
486 *
487 * Returns true if all allocations were successful, false if any fail.
488 */
489static bool __ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring, u16 count)
490{
491 u32 nb_buffs_extra = 0, nb_buffs = 0;
492 union ice_32b_rx_flex_desc *rx_desc;
493 u16 ntu = rx_ring->next_to_use;
494 u16 total_count = count;
495 struct xdp_buff **xdp;
496
497 rx_desc = ICE_RX_DESC(rx_ring, ntu);
498 xdp = ice_xdp_buf(rx_ring, ntu);
499
500 if (ntu + count >= rx_ring->count) {
501 nb_buffs_extra = ice_fill_rx_descs(rx_ring->xsk_pool, xdp,
502 rx_desc,
503 rx_ring->count - ntu);
504 if (nb_buffs_extra != rx_ring->count - ntu) {
505 ntu += nb_buffs_extra;
506 goto exit;
507 }
508 rx_desc = ICE_RX_DESC(rx_ring, 0);
509 xdp = ice_xdp_buf(rx_ring, 0);
510 ntu = 0;
511 count -= nb_buffs_extra;
512 ice_release_rx_desc(rx_ring, 0);
513 }
514
515 nb_buffs = ice_fill_rx_descs(rx_ring->xsk_pool, xdp, rx_desc, count);
516
517 ntu += nb_buffs;
518 if (ntu == rx_ring->count)
519 ntu = 0;
520
521exit:
522 if (rx_ring->next_to_use != ntu)
523 ice_release_rx_desc(rx_ring, ntu);
524
525 return total_count == (nb_buffs_extra + nb_buffs);
526}
527
528/**
529 * ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
530 * @rx_ring: Rx ring
531 * @count: The number of buffers to allocate
532 *
533 * Wrapper for internal allocation routine; figure out how many tail
534 * bumps should take place based on the given threshold
535 *
536 * Returns true if all calls to internal alloc routine succeeded
537 */
538bool ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring, u16 count)
539{
540 u16 rx_thresh = ICE_RING_QUARTER(rx_ring);
541 u16 leftover, i, tail_bumps;
542
543 tail_bumps = count / rx_thresh;
544 leftover = count - (tail_bumps * rx_thresh);
545
546 for (i = 0; i < tail_bumps; i++)
547 if (!__ice_alloc_rx_bufs_zc(rx_ring, rx_thresh))
548 return false;
549 return __ice_alloc_rx_bufs_zc(rx_ring, leftover);
550}
551
552/**
553 * ice_bump_ntc - Bump the next_to_clean counter of an Rx ring
554 * @rx_ring: Rx ring
555 */
556static void ice_bump_ntc(struct ice_rx_ring *rx_ring)
557{
558 int ntc = rx_ring->next_to_clean + 1;
559
560 ntc = (ntc < rx_ring->count) ? ntc : 0;
561 rx_ring->next_to_clean = ntc;
562 prefetch(ICE_RX_DESC(rx_ring, ntc));
563}
564
565/**
566 * ice_construct_skb_zc - Create an sk_buff from zero-copy buffer
567 * @rx_ring: Rx ring
568 * @xdp: Pointer to XDP buffer
569 *
570 * This function allocates a new skb from a zero-copy Rx buffer.
571 *
572 * Returns the skb on success, NULL on failure.
573 */
574static struct sk_buff *
575ice_construct_skb_zc(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp)
576{
577 unsigned int totalsize = xdp->data_end - xdp->data_meta;
578 unsigned int metasize = xdp->data - xdp->data_meta;
579 struct sk_buff *skb;
580
581 net_prefetch(xdp->data_meta);
582
583 skb = __napi_alloc_skb(&rx_ring->q_vector->napi, totalsize,
584 GFP_ATOMIC | __GFP_NOWARN);
585 if (unlikely(!skb))
586 return NULL;
587
588 memcpy(__skb_put(skb, totalsize), xdp->data_meta,
589 ALIGN(totalsize, sizeof(long)));
590
591 if (metasize) {
592 skb_metadata_set(skb, metasize);
593 __skb_pull(skb, metasize);
594 }
595
596 xsk_buff_free(xdp);
597 return skb;
598}
599
600/**
601 * ice_run_xdp_zc - Executes an XDP program in zero-copy path
602 * @rx_ring: Rx ring
603 * @xdp: xdp_buff used as input to the XDP program
604 * @xdp_prog: XDP program to run
605 * @xdp_ring: ring to be used for XDP_TX action
606 *
607 * Returns any of ICE_XDP_{PASS, CONSUMED, TX, REDIR}
608 */
609static int
610ice_run_xdp_zc(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp,
611 struct bpf_prog *xdp_prog, struct ice_tx_ring *xdp_ring)
612{
613 int err, result = ICE_XDP_PASS;
614 u32 act;
615
616 act = bpf_prog_run_xdp(xdp_prog, xdp);
617
618 if (likely(act == XDP_REDIRECT)) {
619 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
620 if (!err)
621 return ICE_XDP_REDIR;
622 if (xsk_uses_need_wakeup(rx_ring->xsk_pool) && err == -ENOBUFS)
623 result = ICE_XDP_EXIT;
624 else
625 result = ICE_XDP_CONSUMED;
626 goto out_failure;
627 }
628
629 switch (act) {
630 case XDP_PASS:
631 break;
632 case XDP_TX:
633 result = ice_xmit_xdp_buff(xdp, xdp_ring);
634 if (result == ICE_XDP_CONSUMED)
635 goto out_failure;
636 break;
637 case XDP_DROP:
638 result = ICE_XDP_CONSUMED;
639 break;
640 default:
641 bpf_warn_invalid_xdp_action(rx_ring->netdev, xdp_prog, act);
642 fallthrough;
643 case XDP_ABORTED:
644 result = ICE_XDP_CONSUMED;
645out_failure:
646 trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
647 break;
648 }
649
650 return result;
651}
652
653/**
654 * ice_clean_rx_irq_zc - consumes packets from the hardware ring
655 * @rx_ring: AF_XDP Rx ring
656 * @budget: NAPI budget
657 *
658 * Returns number of processed packets on success, remaining budget on failure.
659 */
660int ice_clean_rx_irq_zc(struct ice_rx_ring *rx_ring, int budget)
661{
662 unsigned int total_rx_bytes = 0, total_rx_packets = 0;
663 struct ice_tx_ring *xdp_ring;
664 unsigned int xdp_xmit = 0;
665 struct bpf_prog *xdp_prog;
666 bool failure = false;
667 int entries_to_alloc;
668
669 /* ZC patch is enabled only when XDP program is set,
670 * so here it can not be NULL
671 */
672 xdp_prog = READ_ONCE(rx_ring->xdp_prog);
673 xdp_ring = rx_ring->xdp_ring;
674
675 while (likely(total_rx_packets < (unsigned int)budget)) {
676 union ice_32b_rx_flex_desc *rx_desc;
677 unsigned int size, xdp_res = 0;
678 struct xdp_buff *xdp;
679 struct sk_buff *skb;
680 u16 stat_err_bits;
681 u16 vlan_tag = 0;
682 u16 rx_ptype;
683
684 rx_desc = ICE_RX_DESC(rx_ring, rx_ring->next_to_clean);
685
686 stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_DD_S);
687 if (!ice_test_staterr(rx_desc->wb.status_error0, stat_err_bits))
688 break;
689
690 /* This memory barrier is needed to keep us from reading
691 * any other fields out of the rx_desc until we have
692 * verified the descriptor has been written back.
693 */
694 dma_rmb();
695
696 if (unlikely(rx_ring->next_to_clean == rx_ring->next_to_use))
697 break;
698
699 xdp = *ice_xdp_buf(rx_ring, rx_ring->next_to_clean);
700
701 size = le16_to_cpu(rx_desc->wb.pkt_len) &
702 ICE_RX_FLX_DESC_PKT_LEN_M;
703 if (!size) {
704 xdp->data = NULL;
705 xdp->data_end = NULL;
706 xdp->data_hard_start = NULL;
707 xdp->data_meta = NULL;
708 goto construct_skb;
709 }
710
711 xsk_buff_set_size(xdp, size);
712 xsk_buff_dma_sync_for_cpu(xdp, rx_ring->xsk_pool);
713
714 xdp_res = ice_run_xdp_zc(rx_ring, xdp, xdp_prog, xdp_ring);
715 if (likely(xdp_res & (ICE_XDP_TX | ICE_XDP_REDIR))) {
716 xdp_xmit |= xdp_res;
717 } else if (xdp_res == ICE_XDP_EXIT) {
718 failure = true;
719 break;
720 } else if (xdp_res == ICE_XDP_CONSUMED) {
721 xsk_buff_free(xdp);
722 } else if (xdp_res == ICE_XDP_PASS) {
723 goto construct_skb;
724 }
725
726 total_rx_bytes += size;
727 total_rx_packets++;
728
729 ice_bump_ntc(rx_ring);
730 continue;
731
732construct_skb:
733 /* XDP_PASS path */
734 skb = ice_construct_skb_zc(rx_ring, xdp);
735 if (!skb) {
736 rx_ring->ring_stats->rx_stats.alloc_buf_failed++;
737 break;
738 }
739
740 ice_bump_ntc(rx_ring);
741
742 if (eth_skb_pad(skb)) {
743 skb = NULL;
744 continue;
745 }
746
747 total_rx_bytes += skb->len;
748 total_rx_packets++;
749
750 vlan_tag = ice_get_vlan_tag_from_rx_desc(rx_desc);
751
752 rx_ptype = le16_to_cpu(rx_desc->wb.ptype_flex_flags0) &
753 ICE_RX_FLEX_DESC_PTYPE_M;
754
755 ice_process_skb_fields(rx_ring, rx_desc, skb, rx_ptype);
756 ice_receive_skb(rx_ring, skb, vlan_tag);
757 }
758
759 entries_to_alloc = ICE_DESC_UNUSED(rx_ring);
760 if (entries_to_alloc > ICE_RING_QUARTER(rx_ring))
761 failure |= !ice_alloc_rx_bufs_zc(rx_ring, entries_to_alloc);
762
763 ice_finalize_xdp_rx(xdp_ring, xdp_xmit);
764 ice_update_rx_ring_stats(rx_ring, total_rx_packets, total_rx_bytes);
765
766 if (xsk_uses_need_wakeup(rx_ring->xsk_pool)) {
767 if (failure || rx_ring->next_to_clean == rx_ring->next_to_use)
768 xsk_set_rx_need_wakeup(rx_ring->xsk_pool);
769 else
770 xsk_clear_rx_need_wakeup(rx_ring->xsk_pool);
771
772 return (int)total_rx_packets;
773 }
774
775 return failure ? budget : (int)total_rx_packets;
776}
777
778/**
779 * ice_clean_xdp_tx_buf - Free and unmap XDP Tx buffer
780 * @xdp_ring: XDP Tx ring
781 * @tx_buf: Tx buffer to clean
782 */
783static void
784ice_clean_xdp_tx_buf(struct ice_tx_ring *xdp_ring, struct ice_tx_buf *tx_buf)
785{
786 page_frag_free(tx_buf->raw_buf);
787 xdp_ring->xdp_tx_active--;
788 dma_unmap_single(xdp_ring->dev, dma_unmap_addr(tx_buf, dma),
789 dma_unmap_len(tx_buf, len), DMA_TO_DEVICE);
790 dma_unmap_len_set(tx_buf, len, 0);
791}
792
793/**
794 * ice_clean_xdp_irq_zc - produce AF_XDP descriptors to CQ
795 * @xdp_ring: XDP Tx ring
796 */
797static void ice_clean_xdp_irq_zc(struct ice_tx_ring *xdp_ring)
798{
799 u16 ntc = xdp_ring->next_to_clean;
800 struct ice_tx_desc *tx_desc;
801 u16 cnt = xdp_ring->count;
802 struct ice_tx_buf *tx_buf;
803 u16 completed_frames = 0;
804 u16 xsk_frames = 0;
805 u16 last_rs;
806 int i;
807
808 last_rs = xdp_ring->next_to_use ? xdp_ring->next_to_use - 1 : cnt - 1;
809 tx_desc = ICE_TX_DESC(xdp_ring, last_rs);
810 if ((tx_desc->cmd_type_offset_bsz &
811 cpu_to_le64(ICE_TX_DESC_DTYPE_DESC_DONE))) {
812 if (last_rs >= ntc)
813 completed_frames = last_rs - ntc + 1;
814 else
815 completed_frames = last_rs + cnt - ntc + 1;
816 }
817
818 if (!completed_frames)
819 return;
820
821 if (likely(!xdp_ring->xdp_tx_active)) {
822 xsk_frames = completed_frames;
823 goto skip;
824 }
825
826 ntc = xdp_ring->next_to_clean;
827 for (i = 0; i < completed_frames; i++) {
828 tx_buf = &xdp_ring->tx_buf[ntc];
829
830 if (tx_buf->raw_buf) {
831 ice_clean_xdp_tx_buf(xdp_ring, tx_buf);
832 tx_buf->raw_buf = NULL;
833 } else {
834 xsk_frames++;
835 }
836
837 ntc++;
838 if (ntc >= xdp_ring->count)
839 ntc = 0;
840 }
841skip:
842 tx_desc->cmd_type_offset_bsz = 0;
843 xdp_ring->next_to_clean += completed_frames;
844 if (xdp_ring->next_to_clean >= cnt)
845 xdp_ring->next_to_clean -= cnt;
846 if (xsk_frames)
847 xsk_tx_completed(xdp_ring->xsk_pool, xsk_frames);
848}
849
850/**
851 * ice_xmit_pkt - produce a single HW Tx descriptor out of AF_XDP descriptor
852 * @xdp_ring: XDP ring to produce the HW Tx descriptor on
853 * @desc: AF_XDP descriptor to pull the DMA address and length from
854 * @total_bytes: bytes accumulator that will be used for stats update
855 */
856static void ice_xmit_pkt(struct ice_tx_ring *xdp_ring, struct xdp_desc *desc,
857 unsigned int *total_bytes)
858{
859 struct ice_tx_desc *tx_desc;
860 dma_addr_t dma;
861
862 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc->addr);
863 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc->len);
864
865 tx_desc = ICE_TX_DESC(xdp_ring, xdp_ring->next_to_use++);
866 tx_desc->buf_addr = cpu_to_le64(dma);
867 tx_desc->cmd_type_offset_bsz = ice_build_ctob(ICE_TX_DESC_CMD_EOP,
868 0, desc->len, 0);
869
870 *total_bytes += desc->len;
871}
872
873/**
874 * ice_xmit_pkt_batch - produce a batch of HW Tx descriptors out of AF_XDP descriptors
875 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
876 * @descs: AF_XDP descriptors to pull the DMA addresses and lengths from
877 * @total_bytes: bytes accumulator that will be used for stats update
878 */
879static void ice_xmit_pkt_batch(struct ice_tx_ring *xdp_ring, struct xdp_desc *descs,
880 unsigned int *total_bytes)
881{
882 u16 ntu = xdp_ring->next_to_use;
883 struct ice_tx_desc *tx_desc;
884 u32 i;
885
886 loop_unrolled_for(i = 0; i < PKTS_PER_BATCH; i++) {
887 dma_addr_t dma;
888
889 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, descs[i].addr);
890 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, descs[i].len);
891
892 tx_desc = ICE_TX_DESC(xdp_ring, ntu++);
893 tx_desc->buf_addr = cpu_to_le64(dma);
894 tx_desc->cmd_type_offset_bsz = ice_build_ctob(ICE_TX_DESC_CMD_EOP,
895 0, descs[i].len, 0);
896
897 *total_bytes += descs[i].len;
898 }
899
900 xdp_ring->next_to_use = ntu;
901}
902
903/**
904 * ice_fill_tx_hw_ring - produce the number of Tx descriptors onto ring
905 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
906 * @descs: AF_XDP descriptors to pull the DMA addresses and lengths from
907 * @nb_pkts: count of packets to be send
908 * @total_bytes: bytes accumulator that will be used for stats update
909 */
910static void ice_fill_tx_hw_ring(struct ice_tx_ring *xdp_ring, struct xdp_desc *descs,
911 u32 nb_pkts, unsigned int *total_bytes)
912{
913 u32 batched, leftover, i;
914
915 batched = ALIGN_DOWN(nb_pkts, PKTS_PER_BATCH);
916 leftover = nb_pkts & (PKTS_PER_BATCH - 1);
917 for (i = 0; i < batched; i += PKTS_PER_BATCH)
918 ice_xmit_pkt_batch(xdp_ring, &descs[i], total_bytes);
919 for (; i < batched + leftover; i++)
920 ice_xmit_pkt(xdp_ring, &descs[i], total_bytes);
921}
922
923/**
924 * ice_set_rs_bit - set RS bit on last produced descriptor (one behind current NTU)
925 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
926 */
927static void ice_set_rs_bit(struct ice_tx_ring *xdp_ring)
928{
929 u16 ntu = xdp_ring->next_to_use ? xdp_ring->next_to_use - 1 : xdp_ring->count - 1;
930 struct ice_tx_desc *tx_desc;
931
932 tx_desc = ICE_TX_DESC(xdp_ring, ntu);
933 tx_desc->cmd_type_offset_bsz |=
934 cpu_to_le64(ICE_TX_DESC_CMD_RS << ICE_TXD_QW1_CMD_S);
935}
936
937/**
938 * ice_xmit_zc - take entries from XSK Tx ring and place them onto HW Tx ring
939 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
940 *
941 * Returns true if there is no more work that needs to be done, false otherwise
942 */
943bool ice_xmit_zc(struct ice_tx_ring *xdp_ring)
944{
945 struct xdp_desc *descs = xdp_ring->xsk_pool->tx_descs;
946 u32 nb_pkts, nb_processed = 0;
947 unsigned int total_bytes = 0;
948 int budget;
949
950 ice_clean_xdp_irq_zc(xdp_ring);
951
952 budget = ICE_DESC_UNUSED(xdp_ring);
953 budget = min_t(u16, budget, ICE_RING_QUARTER(xdp_ring));
954
955 nb_pkts = xsk_tx_peek_release_desc_batch(xdp_ring->xsk_pool, budget);
956 if (!nb_pkts)
957 return true;
958
959 if (xdp_ring->next_to_use + nb_pkts >= xdp_ring->count) {
960 nb_processed = xdp_ring->count - xdp_ring->next_to_use;
961 ice_fill_tx_hw_ring(xdp_ring, descs, nb_processed, &total_bytes);
962 xdp_ring->next_to_use = 0;
963 }
964
965 ice_fill_tx_hw_ring(xdp_ring, &descs[nb_processed], nb_pkts - nb_processed,
966 &total_bytes);
967
968 ice_set_rs_bit(xdp_ring);
969 ice_xdp_ring_update_tail(xdp_ring);
970 ice_update_tx_ring_stats(xdp_ring, nb_pkts, total_bytes);
971
972 if (xsk_uses_need_wakeup(xdp_ring->xsk_pool))
973 xsk_set_tx_need_wakeup(xdp_ring->xsk_pool);
974
975 return nb_pkts < budget;
976}
977
978/**
979 * ice_xsk_wakeup - Implements ndo_xsk_wakeup
980 * @netdev: net_device
981 * @queue_id: queue to wake up
982 * @flags: ignored in our case, since we have Rx and Tx in the same NAPI
983 *
984 * Returns negative on error, zero otherwise.
985 */
986int
987ice_xsk_wakeup(struct net_device *netdev, u32 queue_id,
988 u32 __always_unused flags)
989{
990 struct ice_netdev_priv *np = netdev_priv(netdev);
991 struct ice_q_vector *q_vector;
992 struct ice_vsi *vsi = np->vsi;
993 struct ice_tx_ring *ring;
994
995 if (test_bit(ICE_VSI_DOWN, vsi->state))
996 return -ENETDOWN;
997
998 if (!ice_is_xdp_ena_vsi(vsi))
999 return -EINVAL;
1000
1001 if (queue_id >= vsi->num_txq || queue_id >= vsi->num_rxq)
1002 return -EINVAL;
1003
1004 ring = vsi->rx_rings[queue_id]->xdp_ring;
1005
1006 if (!ring->xsk_pool)
1007 return -EINVAL;
1008
1009 /* The idea here is that if NAPI is running, mark a miss, so
1010 * it will run again. If not, trigger an interrupt and
1011 * schedule the NAPI from interrupt context. If NAPI would be
1012 * scheduled here, the interrupt affinity would not be
1013 * honored.
1014 */
1015 q_vector = ring->q_vector;
1016 if (!napi_if_scheduled_mark_missed(&q_vector->napi))
1017 ice_trigger_sw_intr(&vsi->back->hw, q_vector);
1018
1019 return 0;
1020}
1021
1022/**
1023 * ice_xsk_any_rx_ring_ena - Checks if Rx rings have AF_XDP buff pool attached
1024 * @vsi: VSI to be checked
1025 *
1026 * Returns true if any of the Rx rings has an AF_XDP buff pool attached
1027 */
1028bool ice_xsk_any_rx_ring_ena(struct ice_vsi *vsi)
1029{
1030 int i;
1031
1032 ice_for_each_rxq(vsi, i) {
1033 if (xsk_get_pool_from_qid(vsi->netdev, i))
1034 return true;
1035 }
1036
1037 return false;
1038}
1039
1040/**
1041 * ice_xsk_clean_rx_ring - clean buffer pool queues connected to a given Rx ring
1042 * @rx_ring: ring to be cleaned
1043 */
1044void ice_xsk_clean_rx_ring(struct ice_rx_ring *rx_ring)
1045{
1046 u16 ntc = rx_ring->next_to_clean;
1047 u16 ntu = rx_ring->next_to_use;
1048
1049 while (ntc != ntu) {
1050 struct xdp_buff *xdp = *ice_xdp_buf(rx_ring, ntc);
1051
1052 xsk_buff_free(xdp);
1053 ntc++;
1054 if (ntc >= rx_ring->count)
1055 ntc = 0;
1056 }
1057}
1058
1059/**
1060 * ice_xsk_clean_xdp_ring - Clean the XDP Tx ring and its buffer pool queues
1061 * @xdp_ring: XDP_Tx ring
1062 */
1063void ice_xsk_clean_xdp_ring(struct ice_tx_ring *xdp_ring)
1064{
1065 u16 ntc = xdp_ring->next_to_clean, ntu = xdp_ring->next_to_use;
1066 u32 xsk_frames = 0;
1067
1068 while (ntc != ntu) {
1069 struct ice_tx_buf *tx_buf = &xdp_ring->tx_buf[ntc];
1070
1071 if (tx_buf->raw_buf)
1072 ice_clean_xdp_tx_buf(xdp_ring, tx_buf);
1073 else
1074 xsk_frames++;
1075
1076 tx_buf->raw_buf = NULL;
1077
1078 ntc++;
1079 if (ntc >= xdp_ring->count)
1080 ntc = 0;
1081 }
1082
1083 if (xsk_frames)
1084 xsk_tx_completed(xdp_ring->xsk_pool, xsk_frames);
1085}
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/**
16 * ice_qp_reset_stats - Resets all stats for rings of given index
17 * @vsi: VSI that contains rings of interest
18 * @q_idx: ring index in array
19 */
20static void ice_qp_reset_stats(struct ice_vsi *vsi, u16 q_idx)
21{
22 memset(&vsi->rx_rings[q_idx]->rx_stats, 0,
23 sizeof(vsi->rx_rings[q_idx]->rx_stats));
24 memset(&vsi->tx_rings[q_idx]->stats, 0,
25 sizeof(vsi->tx_rings[q_idx]->stats));
26 if (ice_is_xdp_ena_vsi(vsi))
27 memset(&vsi->xdp_rings[q_idx]->stats, 0,
28 sizeof(vsi->xdp_rings[q_idx]->stats));
29}
30
31/**
32 * ice_qp_clean_rings - Cleans all the rings of a given index
33 * @vsi: VSI that contains rings of interest
34 * @q_idx: ring index in array
35 */
36static void ice_qp_clean_rings(struct ice_vsi *vsi, u16 q_idx)
37{
38 ice_clean_tx_ring(vsi->tx_rings[q_idx]);
39 if (ice_is_xdp_ena_vsi(vsi))
40 ice_clean_tx_ring(vsi->xdp_rings[q_idx]);
41 ice_clean_rx_ring(vsi->rx_rings[q_idx]);
42}
43
44/**
45 * ice_qvec_toggle_napi - Enables/disables NAPI for a given q_vector
46 * @vsi: VSI that has netdev
47 * @q_vector: q_vector that has NAPI context
48 * @enable: true for enable, false for disable
49 */
50static void
51ice_qvec_toggle_napi(struct ice_vsi *vsi, struct ice_q_vector *q_vector,
52 bool enable)
53{
54 if (!vsi->netdev || !q_vector)
55 return;
56
57 if (enable)
58 napi_enable(&q_vector->napi);
59 else
60 napi_disable(&q_vector->napi);
61}
62
63/**
64 * ice_qvec_dis_irq - Mask off queue interrupt generation on given ring
65 * @vsi: the VSI that contains queue vector being un-configured
66 * @rx_ring: Rx ring that will have its IRQ disabled
67 * @q_vector: queue vector
68 */
69static void
70ice_qvec_dis_irq(struct ice_vsi *vsi, struct ice_ring *rx_ring,
71 struct ice_q_vector *q_vector)
72{
73 struct ice_pf *pf = vsi->back;
74 struct ice_hw *hw = &pf->hw;
75 int base = vsi->base_vector;
76 u16 reg;
77 u32 val;
78
79 /* QINT_TQCTL is being cleared in ice_vsi_stop_tx_ring, so handle
80 * here only QINT_RQCTL
81 */
82 reg = rx_ring->reg_idx;
83 val = rd32(hw, QINT_RQCTL(reg));
84 val &= ~QINT_RQCTL_CAUSE_ENA_M;
85 wr32(hw, QINT_RQCTL(reg), val);
86
87 if (q_vector) {
88 u16 v_idx = q_vector->v_idx;
89
90 wr32(hw, GLINT_DYN_CTL(q_vector->reg_idx), 0);
91 ice_flush(hw);
92 synchronize_irq(pf->msix_entries[v_idx + base].vector);
93 }
94}
95
96/**
97 * ice_qvec_cfg_msix - Enable IRQ for given queue vector
98 * @vsi: the VSI that contains queue vector
99 * @q_vector: queue vector
100 */
101static void
102ice_qvec_cfg_msix(struct ice_vsi *vsi, struct ice_q_vector *q_vector)
103{
104 u16 reg_idx = q_vector->reg_idx;
105 struct ice_pf *pf = vsi->back;
106 struct ice_hw *hw = &pf->hw;
107 struct ice_ring *ring;
108
109 ice_cfg_itr(hw, q_vector);
110
111 ice_for_each_ring(ring, q_vector->tx)
112 ice_cfg_txq_interrupt(vsi, ring->reg_idx, reg_idx,
113 q_vector->tx.itr_idx);
114
115 ice_for_each_ring(ring, q_vector->rx)
116 ice_cfg_rxq_interrupt(vsi, ring->reg_idx, reg_idx,
117 q_vector->rx.itr_idx);
118
119 ice_flush(hw);
120}
121
122/**
123 * ice_qvec_ena_irq - Enable IRQ for given queue vector
124 * @vsi: the VSI that contains queue vector
125 * @q_vector: queue vector
126 */
127static void ice_qvec_ena_irq(struct ice_vsi *vsi, struct ice_q_vector *q_vector)
128{
129 struct ice_pf *pf = vsi->back;
130 struct ice_hw *hw = &pf->hw;
131
132 ice_irq_dynamic_ena(hw, vsi, q_vector);
133
134 ice_flush(hw);
135}
136
137/**
138 * ice_qp_dis - Disables a queue pair
139 * @vsi: VSI of interest
140 * @q_idx: ring index in array
141 *
142 * Returns 0 on success, negative on failure.
143 */
144static int ice_qp_dis(struct ice_vsi *vsi, u16 q_idx)
145{
146 struct ice_txq_meta txq_meta = { };
147 struct ice_ring *tx_ring, *rx_ring;
148 struct ice_q_vector *q_vector;
149 int timeout = 50;
150 int err;
151
152 if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq)
153 return -EINVAL;
154
155 tx_ring = vsi->tx_rings[q_idx];
156 rx_ring = vsi->rx_rings[q_idx];
157 q_vector = rx_ring->q_vector;
158
159 while (test_and_set_bit(ICE_CFG_BUSY, vsi->state)) {
160 timeout--;
161 if (!timeout)
162 return -EBUSY;
163 usleep_range(1000, 2000);
164 }
165 netif_tx_stop_queue(netdev_get_tx_queue(vsi->netdev, q_idx));
166
167 ice_qvec_dis_irq(vsi, rx_ring, q_vector);
168
169 ice_fill_txq_meta(vsi, tx_ring, &txq_meta);
170 err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, tx_ring, &txq_meta);
171 if (err)
172 return err;
173 if (ice_is_xdp_ena_vsi(vsi)) {
174 struct ice_ring *xdp_ring = vsi->xdp_rings[q_idx];
175
176 memset(&txq_meta, 0, sizeof(txq_meta));
177 ice_fill_txq_meta(vsi, xdp_ring, &txq_meta);
178 err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, xdp_ring,
179 &txq_meta);
180 if (err)
181 return err;
182 }
183 err = ice_vsi_ctrl_one_rx_ring(vsi, false, q_idx, true);
184 if (err)
185 return err;
186
187 ice_qvec_toggle_napi(vsi, q_vector, false);
188 ice_qp_clean_rings(vsi, q_idx);
189 ice_qp_reset_stats(vsi, q_idx);
190
191 return 0;
192}
193
194/**
195 * ice_qp_ena - Enables a queue pair
196 * @vsi: VSI of interest
197 * @q_idx: ring index in array
198 *
199 * Returns 0 on success, negative on failure.
200 */
201static int ice_qp_ena(struct ice_vsi *vsi, u16 q_idx)
202{
203 struct ice_aqc_add_tx_qgrp *qg_buf;
204 struct ice_ring *tx_ring, *rx_ring;
205 struct ice_q_vector *q_vector;
206 u16 size;
207 int err;
208
209 if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq)
210 return -EINVAL;
211
212 size = struct_size(qg_buf, txqs, 1);
213 qg_buf = kzalloc(size, GFP_KERNEL);
214 if (!qg_buf)
215 return -ENOMEM;
216
217 qg_buf->num_txqs = 1;
218
219 tx_ring = vsi->tx_rings[q_idx];
220 rx_ring = vsi->rx_rings[q_idx];
221 q_vector = rx_ring->q_vector;
222
223 err = ice_vsi_cfg_txq(vsi, tx_ring, qg_buf);
224 if (err)
225 goto free_buf;
226
227 if (ice_is_xdp_ena_vsi(vsi)) {
228 struct ice_ring *xdp_ring = vsi->xdp_rings[q_idx];
229
230 memset(qg_buf, 0, size);
231 qg_buf->num_txqs = 1;
232 err = ice_vsi_cfg_txq(vsi, xdp_ring, qg_buf);
233 if (err)
234 goto free_buf;
235 ice_set_ring_xdp(xdp_ring);
236 xdp_ring->xsk_pool = ice_xsk_pool(xdp_ring);
237 }
238
239 err = ice_vsi_cfg_rxq(rx_ring);
240 if (err)
241 goto free_buf;
242
243 ice_qvec_cfg_msix(vsi, q_vector);
244
245 err = ice_vsi_ctrl_one_rx_ring(vsi, true, q_idx, true);
246 if (err)
247 goto free_buf;
248
249 clear_bit(ICE_CFG_BUSY, vsi->state);
250 ice_qvec_toggle_napi(vsi, q_vector, true);
251 ice_qvec_ena_irq(vsi, q_vector);
252
253 netif_tx_start_queue(netdev_get_tx_queue(vsi->netdev, q_idx));
254free_buf:
255 kfree(qg_buf);
256 return err;
257}
258
259/**
260 * ice_xsk_pool_disable - disable a buffer pool region
261 * @vsi: Current VSI
262 * @qid: queue ID
263 *
264 * Returns 0 on success, negative on failure
265 */
266static int ice_xsk_pool_disable(struct ice_vsi *vsi, u16 qid)
267{
268 struct xsk_buff_pool *pool = xsk_get_pool_from_qid(vsi->netdev, qid);
269
270 if (!pool)
271 return -EINVAL;
272
273 clear_bit(qid, vsi->af_xdp_zc_qps);
274 xsk_pool_dma_unmap(pool, ICE_RX_DMA_ATTR);
275
276 return 0;
277}
278
279/**
280 * ice_xsk_pool_enable - enable a buffer pool region
281 * @vsi: Current VSI
282 * @pool: pointer to a requested buffer pool region
283 * @qid: queue ID
284 *
285 * Returns 0 on success, negative on failure
286 */
287static int
288ice_xsk_pool_enable(struct ice_vsi *vsi, struct xsk_buff_pool *pool, u16 qid)
289{
290 int err;
291
292 if (vsi->type != ICE_VSI_PF)
293 return -EINVAL;
294
295 if (qid >= vsi->netdev->real_num_rx_queues ||
296 qid >= vsi->netdev->real_num_tx_queues)
297 return -EINVAL;
298
299 err = xsk_pool_dma_map(pool, ice_pf_to_dev(vsi->back),
300 ICE_RX_DMA_ATTR);
301 if (err)
302 return err;
303
304 set_bit(qid, vsi->af_xdp_zc_qps);
305
306 return 0;
307}
308
309/**
310 * ice_xsk_pool_setup - enable/disable a buffer pool region depending on its state
311 * @vsi: Current VSI
312 * @pool: buffer pool to enable/associate to a ring, NULL to disable
313 * @qid: queue ID
314 *
315 * Returns 0 on success, negative on failure
316 */
317int ice_xsk_pool_setup(struct ice_vsi *vsi, struct xsk_buff_pool *pool, u16 qid)
318{
319 bool if_running, pool_present = !!pool;
320 int ret = 0, pool_failure = 0;
321
322 if_running = netif_running(vsi->netdev) && ice_is_xdp_ena_vsi(vsi);
323
324 if (if_running) {
325 ret = ice_qp_dis(vsi, qid);
326 if (ret) {
327 netdev_err(vsi->netdev, "ice_qp_dis error = %d\n", ret);
328 goto xsk_pool_if_up;
329 }
330 }
331
332 pool_failure = pool_present ? ice_xsk_pool_enable(vsi, pool, qid) :
333 ice_xsk_pool_disable(vsi, qid);
334
335xsk_pool_if_up:
336 if (if_running) {
337 ret = ice_qp_ena(vsi, qid);
338 if (!ret && pool_present)
339 napi_schedule(&vsi->xdp_rings[qid]->q_vector->napi);
340 else if (ret)
341 netdev_err(vsi->netdev, "ice_qp_ena error = %d\n", ret);
342 }
343
344 if (pool_failure) {
345 netdev_err(vsi->netdev, "Could not %sable buffer pool, error = %d\n",
346 pool_present ? "en" : "dis", pool_failure);
347 return pool_failure;
348 }
349
350 return ret;
351}
352
353/**
354 * ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
355 * @rx_ring: Rx ring
356 * @count: The number of buffers to allocate
357 *
358 * This function allocates a number of Rx buffers from the fill ring
359 * or the internal recycle mechanism and places them on the Rx ring.
360 *
361 * Returns true if all allocations were successful, false if any fail.
362 */
363bool ice_alloc_rx_bufs_zc(struct ice_ring *rx_ring, u16 count)
364{
365 union ice_32b_rx_flex_desc *rx_desc;
366 u16 ntu = rx_ring->next_to_use;
367 struct ice_rx_buf *rx_buf;
368 bool ok = true;
369 dma_addr_t dma;
370
371 if (!count)
372 return true;
373
374 rx_desc = ICE_RX_DESC(rx_ring, ntu);
375 rx_buf = &rx_ring->rx_buf[ntu];
376
377 do {
378 rx_buf->xdp = xsk_buff_alloc(rx_ring->xsk_pool);
379 if (!rx_buf->xdp) {
380 ok = false;
381 break;
382 }
383
384 dma = xsk_buff_xdp_get_dma(rx_buf->xdp);
385 rx_desc->read.pkt_addr = cpu_to_le64(dma);
386 rx_desc->wb.status_error0 = 0;
387
388 rx_desc++;
389 rx_buf++;
390 ntu++;
391
392 if (unlikely(ntu == rx_ring->count)) {
393 rx_desc = ICE_RX_DESC(rx_ring, 0);
394 rx_buf = rx_ring->rx_buf;
395 ntu = 0;
396 }
397 } while (--count);
398
399 if (rx_ring->next_to_use != ntu) {
400 /* clear the status bits for the next_to_use descriptor */
401 rx_desc->wb.status_error0 = 0;
402 ice_release_rx_desc(rx_ring, ntu);
403 }
404
405 return ok;
406}
407
408/**
409 * ice_bump_ntc - Bump the next_to_clean counter of an Rx ring
410 * @rx_ring: Rx ring
411 */
412static void ice_bump_ntc(struct ice_ring *rx_ring)
413{
414 int ntc = rx_ring->next_to_clean + 1;
415
416 ntc = (ntc < rx_ring->count) ? ntc : 0;
417 rx_ring->next_to_clean = ntc;
418 prefetch(ICE_RX_DESC(rx_ring, ntc));
419}
420
421/**
422 * ice_construct_skb_zc - Create an sk_buff from zero-copy buffer
423 * @rx_ring: Rx ring
424 * @rx_buf: zero-copy Rx buffer
425 *
426 * This function allocates a new skb from a zero-copy Rx buffer.
427 *
428 * Returns the skb on success, NULL on failure.
429 */
430static struct sk_buff *
431ice_construct_skb_zc(struct ice_ring *rx_ring, struct ice_rx_buf *rx_buf)
432{
433 unsigned int metasize = rx_buf->xdp->data - rx_buf->xdp->data_meta;
434 unsigned int datasize = rx_buf->xdp->data_end - rx_buf->xdp->data;
435 unsigned int datasize_hard = rx_buf->xdp->data_end -
436 rx_buf->xdp->data_hard_start;
437 struct sk_buff *skb;
438
439 skb = __napi_alloc_skb(&rx_ring->q_vector->napi, datasize_hard,
440 GFP_ATOMIC | __GFP_NOWARN);
441 if (unlikely(!skb))
442 return NULL;
443
444 skb_reserve(skb, rx_buf->xdp->data - rx_buf->xdp->data_hard_start);
445 memcpy(__skb_put(skb, datasize), rx_buf->xdp->data, datasize);
446 if (metasize)
447 skb_metadata_set(skb, metasize);
448
449 xsk_buff_free(rx_buf->xdp);
450 rx_buf->xdp = NULL;
451 return skb;
452}
453
454/**
455 * ice_run_xdp_zc - Executes an XDP program in zero-copy path
456 * @rx_ring: Rx ring
457 * @xdp: xdp_buff used as input to the XDP program
458 *
459 * Returns any of ICE_XDP_{PASS, CONSUMED, TX, REDIR}
460 */
461static int
462ice_run_xdp_zc(struct ice_ring *rx_ring, struct xdp_buff *xdp)
463{
464 int err, result = ICE_XDP_PASS;
465 struct bpf_prog *xdp_prog;
466 struct ice_ring *xdp_ring;
467 u32 act;
468
469 /* ZC patch is enabled only when XDP program is set,
470 * so here it can not be NULL
471 */
472 xdp_prog = READ_ONCE(rx_ring->xdp_prog);
473
474 act = bpf_prog_run_xdp(xdp_prog, xdp);
475
476 if (likely(act == XDP_REDIRECT)) {
477 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
478 if (err)
479 goto out_failure;
480 return ICE_XDP_REDIR;
481 }
482
483 switch (act) {
484 case XDP_PASS:
485 break;
486 case XDP_TX:
487 xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->q_index];
488 result = ice_xmit_xdp_buff(xdp, xdp_ring);
489 if (result == ICE_XDP_CONSUMED)
490 goto out_failure;
491 break;
492 default:
493 bpf_warn_invalid_xdp_action(act);
494 fallthrough;
495 case XDP_ABORTED:
496out_failure:
497 trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
498 fallthrough;
499 case XDP_DROP:
500 result = ICE_XDP_CONSUMED;
501 break;
502 }
503
504 return result;
505}
506
507/**
508 * ice_clean_rx_irq_zc - consumes packets from the hardware ring
509 * @rx_ring: AF_XDP Rx ring
510 * @budget: NAPI budget
511 *
512 * Returns number of processed packets on success, remaining budget on failure.
513 */
514int ice_clean_rx_irq_zc(struct ice_ring *rx_ring, int budget)
515{
516 unsigned int total_rx_bytes = 0, total_rx_packets = 0;
517 u16 cleaned_count = ICE_DESC_UNUSED(rx_ring);
518 unsigned int xdp_xmit = 0;
519 bool failure = false;
520
521 while (likely(total_rx_packets < (unsigned int)budget)) {
522 union ice_32b_rx_flex_desc *rx_desc;
523 unsigned int size, xdp_res = 0;
524 struct ice_rx_buf *rx_buf;
525 struct sk_buff *skb;
526 u16 stat_err_bits;
527 u16 vlan_tag = 0;
528 u16 rx_ptype;
529
530 rx_desc = ICE_RX_DESC(rx_ring, rx_ring->next_to_clean);
531
532 stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_DD_S);
533 if (!ice_test_staterr(rx_desc, stat_err_bits))
534 break;
535
536 /* This memory barrier is needed to keep us from reading
537 * any other fields out of the rx_desc until we have
538 * verified the descriptor has been written back.
539 */
540 dma_rmb();
541
542 size = le16_to_cpu(rx_desc->wb.pkt_len) &
543 ICE_RX_FLX_DESC_PKT_LEN_M;
544 if (!size)
545 break;
546
547 rx_buf = &rx_ring->rx_buf[rx_ring->next_to_clean];
548 rx_buf->xdp->data_end = rx_buf->xdp->data + size;
549 xsk_buff_dma_sync_for_cpu(rx_buf->xdp, rx_ring->xsk_pool);
550
551 xdp_res = ice_run_xdp_zc(rx_ring, rx_buf->xdp);
552 if (xdp_res) {
553 if (xdp_res & (ICE_XDP_TX | ICE_XDP_REDIR))
554 xdp_xmit |= xdp_res;
555 else
556 xsk_buff_free(rx_buf->xdp);
557
558 rx_buf->xdp = NULL;
559 total_rx_bytes += size;
560 total_rx_packets++;
561 cleaned_count++;
562
563 ice_bump_ntc(rx_ring);
564 continue;
565 }
566
567 /* XDP_PASS path */
568 skb = ice_construct_skb_zc(rx_ring, rx_buf);
569 if (!skb) {
570 rx_ring->rx_stats.alloc_buf_failed++;
571 break;
572 }
573
574 cleaned_count++;
575 ice_bump_ntc(rx_ring);
576
577 if (eth_skb_pad(skb)) {
578 skb = NULL;
579 continue;
580 }
581
582 total_rx_bytes += skb->len;
583 total_rx_packets++;
584
585 stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_L2TAG1P_S);
586 if (ice_test_staterr(rx_desc, stat_err_bits))
587 vlan_tag = le16_to_cpu(rx_desc->wb.l2tag1);
588
589 rx_ptype = le16_to_cpu(rx_desc->wb.ptype_flex_flags0) &
590 ICE_RX_FLEX_DESC_PTYPE_M;
591
592 ice_process_skb_fields(rx_ring, rx_desc, skb, rx_ptype);
593 ice_receive_skb(rx_ring, skb, vlan_tag);
594 }
595
596 if (cleaned_count >= ICE_RX_BUF_WRITE)
597 failure = !ice_alloc_rx_bufs_zc(rx_ring, cleaned_count);
598
599 ice_finalize_xdp_rx(rx_ring, xdp_xmit);
600 ice_update_rx_ring_stats(rx_ring, total_rx_packets, total_rx_bytes);
601
602 if (xsk_uses_need_wakeup(rx_ring->xsk_pool)) {
603 if (failure || rx_ring->next_to_clean == rx_ring->next_to_use)
604 xsk_set_rx_need_wakeup(rx_ring->xsk_pool);
605 else
606 xsk_clear_rx_need_wakeup(rx_ring->xsk_pool);
607
608 return (int)total_rx_packets;
609 }
610
611 return failure ? budget : (int)total_rx_packets;
612}
613
614/**
615 * ice_xmit_zc - Completes AF_XDP entries, and cleans XDP entries
616 * @xdp_ring: XDP Tx ring
617 * @budget: max number of frames to xmit
618 *
619 * Returns true if cleanup/transmission is done.
620 */
621static bool ice_xmit_zc(struct ice_ring *xdp_ring, int budget)
622{
623 struct ice_tx_desc *tx_desc = NULL;
624 bool work_done = true;
625 struct xdp_desc desc;
626 dma_addr_t dma;
627
628 while (likely(budget-- > 0)) {
629 struct ice_tx_buf *tx_buf;
630
631 if (unlikely(!ICE_DESC_UNUSED(xdp_ring))) {
632 xdp_ring->tx_stats.tx_busy++;
633 work_done = false;
634 break;
635 }
636
637 tx_buf = &xdp_ring->tx_buf[xdp_ring->next_to_use];
638
639 if (!xsk_tx_peek_desc(xdp_ring->xsk_pool, &desc))
640 break;
641
642 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc.addr);
643 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma,
644 desc.len);
645
646 tx_buf->bytecount = desc.len;
647
648 tx_desc = ICE_TX_DESC(xdp_ring, xdp_ring->next_to_use);
649 tx_desc->buf_addr = cpu_to_le64(dma);
650 tx_desc->cmd_type_offset_bsz =
651 ice_build_ctob(ICE_TXD_LAST_DESC_CMD, 0, desc.len, 0);
652
653 xdp_ring->next_to_use++;
654 if (xdp_ring->next_to_use == xdp_ring->count)
655 xdp_ring->next_to_use = 0;
656 }
657
658 if (tx_desc) {
659 ice_xdp_ring_update_tail(xdp_ring);
660 xsk_tx_release(xdp_ring->xsk_pool);
661 }
662
663 return budget > 0 && work_done;
664}
665
666/**
667 * ice_clean_xdp_tx_buf - Free and unmap XDP Tx buffer
668 * @xdp_ring: XDP Tx ring
669 * @tx_buf: Tx buffer to clean
670 */
671static void
672ice_clean_xdp_tx_buf(struct ice_ring *xdp_ring, struct ice_tx_buf *tx_buf)
673{
674 xdp_return_frame((struct xdp_frame *)tx_buf->raw_buf);
675 dma_unmap_single(xdp_ring->dev, dma_unmap_addr(tx_buf, dma),
676 dma_unmap_len(tx_buf, len), DMA_TO_DEVICE);
677 dma_unmap_len_set(tx_buf, len, 0);
678}
679
680/**
681 * ice_clean_tx_irq_zc - Completes AF_XDP entries, and cleans XDP entries
682 * @xdp_ring: XDP Tx ring
683 * @budget: NAPI budget
684 *
685 * Returns true if cleanup/tranmission is done.
686 */
687bool ice_clean_tx_irq_zc(struct ice_ring *xdp_ring, int budget)
688{
689 int total_packets = 0, total_bytes = 0;
690 s16 ntc = xdp_ring->next_to_clean;
691 struct ice_tx_desc *tx_desc;
692 struct ice_tx_buf *tx_buf;
693 u32 xsk_frames = 0;
694 bool xmit_done;
695
696 tx_desc = ICE_TX_DESC(xdp_ring, ntc);
697 tx_buf = &xdp_ring->tx_buf[ntc];
698 ntc -= xdp_ring->count;
699
700 do {
701 if (!(tx_desc->cmd_type_offset_bsz &
702 cpu_to_le64(ICE_TX_DESC_DTYPE_DESC_DONE)))
703 break;
704
705 total_bytes += tx_buf->bytecount;
706 total_packets++;
707
708 if (tx_buf->raw_buf) {
709 ice_clean_xdp_tx_buf(xdp_ring, tx_buf);
710 tx_buf->raw_buf = NULL;
711 } else {
712 xsk_frames++;
713 }
714
715 tx_desc->cmd_type_offset_bsz = 0;
716 tx_buf++;
717 tx_desc++;
718 ntc++;
719
720 if (unlikely(!ntc)) {
721 ntc -= xdp_ring->count;
722 tx_buf = xdp_ring->tx_buf;
723 tx_desc = ICE_TX_DESC(xdp_ring, 0);
724 }
725
726 prefetch(tx_desc);
727
728 } while (likely(--budget));
729
730 ntc += xdp_ring->count;
731 xdp_ring->next_to_clean = ntc;
732
733 if (xsk_frames)
734 xsk_tx_completed(xdp_ring->xsk_pool, xsk_frames);
735
736 if (xsk_uses_need_wakeup(xdp_ring->xsk_pool))
737 xsk_set_tx_need_wakeup(xdp_ring->xsk_pool);
738
739 ice_update_tx_ring_stats(xdp_ring, total_packets, total_bytes);
740 xmit_done = ice_xmit_zc(xdp_ring, ICE_DFLT_IRQ_WORK);
741
742 return budget > 0 && xmit_done;
743}
744
745/**
746 * ice_xsk_wakeup - Implements ndo_xsk_wakeup
747 * @netdev: net_device
748 * @queue_id: queue to wake up
749 * @flags: ignored in our case, since we have Rx and Tx in the same NAPI
750 *
751 * Returns negative on error, zero otherwise.
752 */
753int
754ice_xsk_wakeup(struct net_device *netdev, u32 queue_id,
755 u32 __always_unused flags)
756{
757 struct ice_netdev_priv *np = netdev_priv(netdev);
758 struct ice_q_vector *q_vector;
759 struct ice_vsi *vsi = np->vsi;
760 struct ice_ring *ring;
761
762 if (test_bit(ICE_DOWN, vsi->state))
763 return -ENETDOWN;
764
765 if (!ice_is_xdp_ena_vsi(vsi))
766 return -ENXIO;
767
768 if (queue_id >= vsi->num_txq)
769 return -ENXIO;
770
771 if (!vsi->xdp_rings[queue_id]->xsk_pool)
772 return -ENXIO;
773
774 ring = vsi->xdp_rings[queue_id];
775
776 /* The idea here is that if NAPI is running, mark a miss, so
777 * it will run again. If not, trigger an interrupt and
778 * schedule the NAPI from interrupt context. If NAPI would be
779 * scheduled here, the interrupt affinity would not be
780 * honored.
781 */
782 q_vector = ring->q_vector;
783 if (!napi_if_scheduled_mark_missed(&q_vector->napi))
784 ice_trigger_sw_intr(&vsi->back->hw, q_vector);
785
786 return 0;
787}
788
789/**
790 * ice_xsk_any_rx_ring_ena - Checks if Rx rings have AF_XDP buff pool attached
791 * @vsi: VSI to be checked
792 *
793 * Returns true if any of the Rx rings has an AF_XDP buff pool attached
794 */
795bool ice_xsk_any_rx_ring_ena(struct ice_vsi *vsi)
796{
797 int i;
798
799 ice_for_each_rxq(vsi, i) {
800 if (xsk_get_pool_from_qid(vsi->netdev, i))
801 return true;
802 }
803
804 return false;
805}
806
807/**
808 * ice_xsk_clean_rx_ring - clean buffer pool queues connected to a given Rx ring
809 * @rx_ring: ring to be cleaned
810 */
811void ice_xsk_clean_rx_ring(struct ice_ring *rx_ring)
812{
813 u16 i;
814
815 for (i = 0; i < rx_ring->count; i++) {
816 struct ice_rx_buf *rx_buf = &rx_ring->rx_buf[i];
817
818 if (!rx_buf->xdp)
819 continue;
820
821 rx_buf->xdp = NULL;
822 }
823}
824
825/**
826 * ice_xsk_clean_xdp_ring - Clean the XDP Tx ring and its buffer pool queues
827 * @xdp_ring: XDP_Tx ring
828 */
829void ice_xsk_clean_xdp_ring(struct ice_ring *xdp_ring)
830{
831 u16 ntc = xdp_ring->next_to_clean, ntu = xdp_ring->next_to_use;
832 u32 xsk_frames = 0;
833
834 while (ntc != ntu) {
835 struct ice_tx_buf *tx_buf = &xdp_ring->tx_buf[ntc];
836
837 if (tx_buf->raw_buf)
838 ice_clean_xdp_tx_buf(xdp_ring, tx_buf);
839 else
840 xsk_frames++;
841
842 tx_buf->raw_buf = NULL;
843
844 ntc++;
845 if (ntc >= xdp_ring->count)
846 ntc = 0;
847 }
848
849 if (xsk_frames)
850 xsk_tx_completed(xdp_ring->xsk_pool, xsk_frames);
851}