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

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