Loading...
1// SPDX-License-Identifier: GPL-2.0
2/* Copyright(c) 2018 Intel Corporation. */
3
4#include <linux/bpf_trace.h>
5#include <net/xdp_sock_drv.h>
6#include "i40e_txrx_common.h"
7#include "i40e_xsk.h"
8
9void i40e_clear_rx_bi_zc(struct i40e_ring *rx_ring)
10{
11 memset(rx_ring->rx_bi_zc, 0,
12 sizeof(*rx_ring->rx_bi_zc) * rx_ring->count);
13}
14
15static struct xdp_buff **i40e_rx_bi(struct i40e_ring *rx_ring, u32 idx)
16{
17 return &rx_ring->rx_bi_zc[idx];
18}
19
20/**
21 * i40e_realloc_rx_xdp_bi - reallocate SW ring for either XSK or normal buffer
22 * @rx_ring: Current rx ring
23 * @pool_present: is pool for XSK present
24 *
25 * Try allocating memory and return ENOMEM, if failed to allocate.
26 * If allocation was successful, substitute buffer with allocated one.
27 * Returns 0 on success, negative on failure
28 */
29static int i40e_realloc_rx_xdp_bi(struct i40e_ring *rx_ring, bool pool_present)
30{
31 size_t elem_size = pool_present ? sizeof(*rx_ring->rx_bi_zc) :
32 sizeof(*rx_ring->rx_bi);
33 void *sw_ring = kcalloc(rx_ring->count, elem_size, GFP_KERNEL);
34
35 if (!sw_ring)
36 return -ENOMEM;
37
38 if (pool_present) {
39 kfree(rx_ring->rx_bi);
40 rx_ring->rx_bi = NULL;
41 rx_ring->rx_bi_zc = sw_ring;
42 } else {
43 kfree(rx_ring->rx_bi_zc);
44 rx_ring->rx_bi_zc = NULL;
45 rx_ring->rx_bi = sw_ring;
46 }
47 return 0;
48}
49
50/**
51 * i40e_realloc_rx_bi_zc - reallocate rx SW rings
52 * @vsi: Current VSI
53 * @zc: is zero copy set
54 *
55 * Reallocate buffer for rx_rings that might be used by XSK.
56 * XDP requires more memory, than rx_buf provides.
57 * Returns 0 on success, negative on failure
58 */
59int i40e_realloc_rx_bi_zc(struct i40e_vsi *vsi, bool zc)
60{
61 struct i40e_ring *rx_ring;
62 unsigned long q;
63
64 for_each_set_bit(q, vsi->af_xdp_zc_qps, vsi->alloc_queue_pairs) {
65 rx_ring = vsi->rx_rings[q];
66 if (i40e_realloc_rx_xdp_bi(rx_ring, zc))
67 return -ENOMEM;
68 }
69 return 0;
70}
71
72/**
73 * i40e_xsk_pool_enable - Enable/associate an AF_XDP buffer pool to a
74 * certain ring/qid
75 * @vsi: Current VSI
76 * @pool: buffer pool
77 * @qid: Rx ring to associate buffer pool with
78 *
79 * Returns 0 on success, <0 on failure
80 **/
81static int i40e_xsk_pool_enable(struct i40e_vsi *vsi,
82 struct xsk_buff_pool *pool,
83 u16 qid)
84{
85 struct net_device *netdev = vsi->netdev;
86 bool if_running;
87 int err;
88
89 if (vsi->type != I40E_VSI_MAIN)
90 return -EINVAL;
91
92 if (qid >= vsi->num_queue_pairs)
93 return -EINVAL;
94
95 if (qid >= netdev->real_num_rx_queues ||
96 qid >= netdev->real_num_tx_queues)
97 return -EINVAL;
98
99 err = xsk_pool_dma_map(pool, &vsi->back->pdev->dev, I40E_RX_DMA_ATTR);
100 if (err)
101 return err;
102
103 set_bit(qid, vsi->af_xdp_zc_qps);
104
105 if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
106
107 if (if_running) {
108 err = i40e_queue_pair_disable(vsi, qid);
109 if (err)
110 return err;
111
112 err = i40e_realloc_rx_xdp_bi(vsi->rx_rings[qid], true);
113 if (err)
114 return err;
115
116 err = i40e_queue_pair_enable(vsi, qid);
117 if (err)
118 return err;
119
120 /* Kick start the NAPI context so that receiving will start */
121 err = i40e_xsk_wakeup(vsi->netdev, qid, XDP_WAKEUP_RX);
122 if (err)
123 return err;
124 }
125
126 return 0;
127}
128
129/**
130 * i40e_xsk_pool_disable - Disassociate an AF_XDP buffer pool from a
131 * certain ring/qid
132 * @vsi: Current VSI
133 * @qid: Rx ring to associate buffer pool with
134 *
135 * Returns 0 on success, <0 on failure
136 **/
137static int i40e_xsk_pool_disable(struct i40e_vsi *vsi, u16 qid)
138{
139 struct net_device *netdev = vsi->netdev;
140 struct xsk_buff_pool *pool;
141 bool if_running;
142 int err;
143
144 pool = xsk_get_pool_from_qid(netdev, qid);
145 if (!pool)
146 return -EINVAL;
147
148 if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
149
150 if (if_running) {
151 err = i40e_queue_pair_disable(vsi, qid);
152 if (err)
153 return err;
154 }
155
156 clear_bit(qid, vsi->af_xdp_zc_qps);
157 xsk_pool_dma_unmap(pool, I40E_RX_DMA_ATTR);
158
159 if (if_running) {
160 err = i40e_realloc_rx_xdp_bi(vsi->rx_rings[qid], false);
161 if (err)
162 return err;
163 err = i40e_queue_pair_enable(vsi, qid);
164 if (err)
165 return err;
166 }
167
168 return 0;
169}
170
171/**
172 * i40e_xsk_pool_setup - Enable/disassociate an AF_XDP buffer pool to/from
173 * a ring/qid
174 * @vsi: Current VSI
175 * @pool: Buffer pool to enable/associate to a ring, or NULL to disable
176 * @qid: Rx ring to (dis)associate buffer pool (from)to
177 *
178 * This function enables or disables a buffer pool to a certain ring.
179 *
180 * Returns 0 on success, <0 on failure
181 **/
182int i40e_xsk_pool_setup(struct i40e_vsi *vsi, struct xsk_buff_pool *pool,
183 u16 qid)
184{
185 return pool ? i40e_xsk_pool_enable(vsi, pool, qid) :
186 i40e_xsk_pool_disable(vsi, qid);
187}
188
189/**
190 * i40e_run_xdp_zc - Executes an XDP program on an xdp_buff
191 * @rx_ring: Rx ring
192 * @xdp: xdp_buff used as input to the XDP program
193 * @xdp_prog: XDP program to run
194 *
195 * Returns any of I40E_XDP_{PASS, CONSUMED, TX, REDIR}
196 **/
197static int i40e_run_xdp_zc(struct i40e_ring *rx_ring, struct xdp_buff *xdp,
198 struct bpf_prog *xdp_prog)
199{
200 int err, result = I40E_XDP_PASS;
201 struct i40e_ring *xdp_ring;
202 u32 act;
203
204 act = bpf_prog_run_xdp(xdp_prog, xdp);
205
206 if (likely(act == XDP_REDIRECT)) {
207 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
208 if (!err)
209 return I40E_XDP_REDIR;
210 if (xsk_uses_need_wakeup(rx_ring->xsk_pool) && err == -ENOBUFS)
211 result = I40E_XDP_EXIT;
212 else
213 result = I40E_XDP_CONSUMED;
214 goto out_failure;
215 }
216
217 switch (act) {
218 case XDP_PASS:
219 break;
220 case XDP_TX:
221 xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->queue_index];
222 result = i40e_xmit_xdp_tx_ring(xdp, xdp_ring);
223 if (result == I40E_XDP_CONSUMED)
224 goto out_failure;
225 break;
226 case XDP_DROP:
227 result = I40E_XDP_CONSUMED;
228 break;
229 default:
230 bpf_warn_invalid_xdp_action(rx_ring->netdev, xdp_prog, act);
231 fallthrough;
232 case XDP_ABORTED:
233 result = I40E_XDP_CONSUMED;
234out_failure:
235 trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
236 }
237 return result;
238}
239
240bool i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count)
241{
242 u16 ntu = rx_ring->next_to_use;
243 union i40e_rx_desc *rx_desc;
244 struct xdp_buff **xdp;
245 u32 nb_buffs, i;
246 dma_addr_t dma;
247
248 rx_desc = I40E_RX_DESC(rx_ring, ntu);
249 xdp = i40e_rx_bi(rx_ring, ntu);
250
251 nb_buffs = min_t(u16, count, rx_ring->count - ntu);
252 nb_buffs = xsk_buff_alloc_batch(rx_ring->xsk_pool, xdp, nb_buffs);
253 if (!nb_buffs)
254 return false;
255
256 i = nb_buffs;
257 while (i--) {
258 dma = xsk_buff_xdp_get_dma(*xdp);
259 rx_desc->read.pkt_addr = cpu_to_le64(dma);
260 rx_desc->read.hdr_addr = 0;
261
262 rx_desc++;
263 xdp++;
264 }
265
266 ntu += nb_buffs;
267 if (ntu == rx_ring->count) {
268 rx_desc = I40E_RX_DESC(rx_ring, 0);
269 ntu = 0;
270 }
271
272 /* clear the status bits for the next_to_use descriptor */
273 rx_desc->wb.qword1.status_error_len = 0;
274 i40e_release_rx_desc(rx_ring, ntu);
275
276 return count == nb_buffs;
277}
278
279/**
280 * i40e_construct_skb_zc - Create skbuff from zero-copy Rx buffer
281 * @rx_ring: Rx ring
282 * @xdp: xdp_buff
283 *
284 * This functions allocates a new skb from a zero-copy Rx buffer.
285 *
286 * Returns the skb, or NULL on failure.
287 **/
288static struct sk_buff *i40e_construct_skb_zc(struct i40e_ring *rx_ring,
289 struct xdp_buff *xdp)
290{
291 unsigned int totalsize = xdp->data_end - xdp->data_meta;
292 unsigned int metasize = xdp->data - xdp->data_meta;
293 struct skb_shared_info *sinfo = NULL;
294 struct sk_buff *skb;
295 u32 nr_frags = 0;
296
297 if (unlikely(xdp_buff_has_frags(xdp))) {
298 sinfo = xdp_get_shared_info_from_buff(xdp);
299 nr_frags = sinfo->nr_frags;
300 }
301 net_prefetch(xdp->data_meta);
302
303 /* allocate a skb to store the frags */
304 skb = napi_alloc_skb(&rx_ring->q_vector->napi, totalsize);
305 if (unlikely(!skb))
306 goto out;
307
308 memcpy(__skb_put(skb, totalsize), xdp->data_meta,
309 ALIGN(totalsize, sizeof(long)));
310
311 if (metasize) {
312 skb_metadata_set(skb, metasize);
313 __skb_pull(skb, metasize);
314 }
315
316 if (likely(!xdp_buff_has_frags(xdp)))
317 goto out;
318
319 for (int i = 0; i < nr_frags; i++) {
320 struct skb_shared_info *skinfo = skb_shinfo(skb);
321 skb_frag_t *frag = &sinfo->frags[i];
322 struct page *page;
323 void *addr;
324
325 page = dev_alloc_page();
326 if (!page) {
327 dev_kfree_skb(skb);
328 return NULL;
329 }
330 addr = page_to_virt(page);
331
332 memcpy(addr, skb_frag_page(frag), skb_frag_size(frag));
333
334 __skb_fill_page_desc_noacc(skinfo, skinfo->nr_frags++,
335 addr, 0, skb_frag_size(frag));
336 }
337
338out:
339 xsk_buff_free(xdp);
340 return skb;
341}
342
343static void i40e_handle_xdp_result_zc(struct i40e_ring *rx_ring,
344 struct xdp_buff *xdp_buff,
345 union i40e_rx_desc *rx_desc,
346 unsigned int *rx_packets,
347 unsigned int *rx_bytes,
348 unsigned int xdp_res,
349 bool *failure)
350{
351 struct sk_buff *skb;
352
353 *rx_packets = 1;
354 *rx_bytes = xdp_get_buff_len(xdp_buff);
355
356 if (likely(xdp_res == I40E_XDP_REDIR) || xdp_res == I40E_XDP_TX)
357 return;
358
359 if (xdp_res == I40E_XDP_EXIT) {
360 *failure = true;
361 return;
362 }
363
364 if (xdp_res == I40E_XDP_CONSUMED) {
365 xsk_buff_free(xdp_buff);
366 return;
367 }
368 if (xdp_res == I40E_XDP_PASS) {
369 /* NB! We are not checking for errors using
370 * i40e_test_staterr with
371 * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that
372 * SBP is *not* set in PRT_SBPVSI (default not set).
373 */
374 skb = i40e_construct_skb_zc(rx_ring, xdp_buff);
375 if (!skb) {
376 rx_ring->rx_stats.alloc_buff_failed++;
377 *rx_packets = 0;
378 *rx_bytes = 0;
379 return;
380 }
381
382 if (eth_skb_pad(skb)) {
383 *rx_packets = 0;
384 *rx_bytes = 0;
385 return;
386 }
387
388 i40e_process_skb_fields(rx_ring, rx_desc, skb);
389 napi_gro_receive(&rx_ring->q_vector->napi, skb);
390 return;
391 }
392
393 /* Should never get here, as all valid cases have been handled already.
394 */
395 WARN_ON_ONCE(1);
396}
397
398static int
399i40e_add_xsk_frag(struct i40e_ring *rx_ring, struct xdp_buff *first,
400 struct xdp_buff *xdp, const unsigned int size)
401{
402 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(first);
403
404 if (!xdp_buff_has_frags(first)) {
405 sinfo->nr_frags = 0;
406 sinfo->xdp_frags_size = 0;
407 xdp_buff_set_frags_flag(first);
408 }
409
410 if (unlikely(sinfo->nr_frags == MAX_SKB_FRAGS)) {
411 xsk_buff_free(first);
412 return -ENOMEM;
413 }
414
415 __skb_fill_page_desc_noacc(sinfo, sinfo->nr_frags++,
416 virt_to_page(xdp->data_hard_start),
417 XDP_PACKET_HEADROOM, size);
418 sinfo->xdp_frags_size += size;
419 xsk_buff_add_frag(xdp);
420
421 return 0;
422}
423
424/**
425 * i40e_clean_rx_irq_zc - Consumes Rx packets from the hardware ring
426 * @rx_ring: Rx ring
427 * @budget: NAPI budget
428 *
429 * Returns amount of work completed
430 **/
431int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget)
432{
433 unsigned int total_rx_bytes = 0, total_rx_packets = 0;
434 u16 next_to_process = rx_ring->next_to_process;
435 u16 next_to_clean = rx_ring->next_to_clean;
436 unsigned int xdp_res, xdp_xmit = 0;
437 struct xdp_buff *first = NULL;
438 u32 count = rx_ring->count;
439 struct bpf_prog *xdp_prog;
440 u32 entries_to_alloc;
441 bool failure = false;
442
443 if (next_to_process != next_to_clean)
444 first = *i40e_rx_bi(rx_ring, next_to_clean);
445
446 /* NB! xdp_prog will always be !NULL, due to the fact that
447 * this path is enabled by setting an XDP program.
448 */
449 xdp_prog = READ_ONCE(rx_ring->xdp_prog);
450
451 while (likely(total_rx_packets < (unsigned int)budget)) {
452 union i40e_rx_desc *rx_desc;
453 unsigned int rx_packets;
454 unsigned int rx_bytes;
455 struct xdp_buff *bi;
456 unsigned int size;
457 u64 qword;
458
459 rx_desc = I40E_RX_DESC(rx_ring, next_to_process);
460 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
461
462 /* This memory barrier is needed to keep us from reading
463 * any other fields out of the rx_desc until we have
464 * verified the descriptor has been written back.
465 */
466 dma_rmb();
467
468 if (i40e_rx_is_programming_status(qword)) {
469 i40e_clean_programming_status(rx_ring,
470 rx_desc->raw.qword[0],
471 qword);
472 bi = *i40e_rx_bi(rx_ring, next_to_process);
473 xsk_buff_free(bi);
474 if (++next_to_process == count)
475 next_to_process = 0;
476 continue;
477 }
478
479 size = FIELD_GET(I40E_RXD_QW1_LENGTH_PBUF_MASK, qword);
480 if (!size)
481 break;
482
483 bi = *i40e_rx_bi(rx_ring, next_to_process);
484 xsk_buff_set_size(bi, size);
485 xsk_buff_dma_sync_for_cpu(bi);
486
487 if (!first)
488 first = bi;
489 else if (i40e_add_xsk_frag(rx_ring, first, bi, size))
490 break;
491
492 if (++next_to_process == count)
493 next_to_process = 0;
494
495 if (i40e_is_non_eop(rx_ring, rx_desc))
496 continue;
497
498 xdp_res = i40e_run_xdp_zc(rx_ring, first, xdp_prog);
499 i40e_handle_xdp_result_zc(rx_ring, first, rx_desc, &rx_packets,
500 &rx_bytes, xdp_res, &failure);
501 next_to_clean = next_to_process;
502 if (failure)
503 break;
504 total_rx_packets += rx_packets;
505 total_rx_bytes += rx_bytes;
506 xdp_xmit |= xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR);
507 first = NULL;
508 }
509
510 rx_ring->next_to_clean = next_to_clean;
511 rx_ring->next_to_process = next_to_process;
512
513 entries_to_alloc = I40E_DESC_UNUSED(rx_ring);
514 if (entries_to_alloc >= I40E_RX_BUFFER_WRITE)
515 failure |= !i40e_alloc_rx_buffers_zc(rx_ring, entries_to_alloc);
516
517 i40e_finalize_xdp_rx(rx_ring, xdp_xmit);
518 i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets);
519
520 if (xsk_uses_need_wakeup(rx_ring->xsk_pool)) {
521 if (failure || next_to_clean == rx_ring->next_to_use)
522 xsk_set_rx_need_wakeup(rx_ring->xsk_pool);
523 else
524 xsk_clear_rx_need_wakeup(rx_ring->xsk_pool);
525
526 return (int)total_rx_packets;
527 }
528 return failure ? budget : (int)total_rx_packets;
529}
530
531static void i40e_xmit_pkt(struct i40e_ring *xdp_ring, struct xdp_desc *desc,
532 unsigned int *total_bytes)
533{
534 u32 cmd = I40E_TX_DESC_CMD_ICRC | xsk_is_eop_desc(desc);
535 struct i40e_tx_desc *tx_desc;
536 dma_addr_t dma;
537
538 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc->addr);
539 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc->len);
540
541 tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use++);
542 tx_desc->buffer_addr = cpu_to_le64(dma);
543 tx_desc->cmd_type_offset_bsz = build_ctob(cmd, 0, desc->len, 0);
544
545 *total_bytes += desc->len;
546}
547
548static void i40e_xmit_pkt_batch(struct i40e_ring *xdp_ring, struct xdp_desc *desc,
549 unsigned int *total_bytes)
550{
551 u16 ntu = xdp_ring->next_to_use;
552 struct i40e_tx_desc *tx_desc;
553 dma_addr_t dma;
554 u32 i;
555
556 loop_unrolled_for(i = 0; i < PKTS_PER_BATCH; i++) {
557 u32 cmd = I40E_TX_DESC_CMD_ICRC | xsk_is_eop_desc(&desc[i]);
558
559 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc[i].addr);
560 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc[i].len);
561
562 tx_desc = I40E_TX_DESC(xdp_ring, ntu++);
563 tx_desc->buffer_addr = cpu_to_le64(dma);
564 tx_desc->cmd_type_offset_bsz = build_ctob(cmd, 0, desc[i].len, 0);
565
566 *total_bytes += desc[i].len;
567 }
568
569 xdp_ring->next_to_use = ntu;
570}
571
572static void i40e_fill_tx_hw_ring(struct i40e_ring *xdp_ring, struct xdp_desc *descs, u32 nb_pkts,
573 unsigned int *total_bytes)
574{
575 u32 batched, leftover, i;
576
577 batched = nb_pkts & ~(PKTS_PER_BATCH - 1);
578 leftover = nb_pkts & (PKTS_PER_BATCH - 1);
579 for (i = 0; i < batched; i += PKTS_PER_BATCH)
580 i40e_xmit_pkt_batch(xdp_ring, &descs[i], total_bytes);
581 for (i = batched; i < batched + leftover; i++)
582 i40e_xmit_pkt(xdp_ring, &descs[i], total_bytes);
583}
584
585static void i40e_set_rs_bit(struct i40e_ring *xdp_ring)
586{
587 u16 ntu = xdp_ring->next_to_use ? xdp_ring->next_to_use - 1 : xdp_ring->count - 1;
588 struct i40e_tx_desc *tx_desc;
589
590 tx_desc = I40E_TX_DESC(xdp_ring, ntu);
591 tx_desc->cmd_type_offset_bsz |= cpu_to_le64(I40E_TX_DESC_CMD_RS << I40E_TXD_QW1_CMD_SHIFT);
592}
593
594/**
595 * i40e_xmit_zc - Performs zero-copy Tx AF_XDP
596 * @xdp_ring: XDP Tx ring
597 * @budget: NAPI budget
598 *
599 * Returns true if the work is finished.
600 **/
601static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
602{
603 struct xdp_desc *descs = xdp_ring->xsk_pool->tx_descs;
604 u32 nb_pkts, nb_processed = 0;
605 unsigned int total_bytes = 0;
606
607 nb_pkts = xsk_tx_peek_release_desc_batch(xdp_ring->xsk_pool, budget);
608 if (!nb_pkts)
609 return true;
610
611 if (xdp_ring->next_to_use + nb_pkts >= xdp_ring->count) {
612 nb_processed = xdp_ring->count - xdp_ring->next_to_use;
613 i40e_fill_tx_hw_ring(xdp_ring, descs, nb_processed, &total_bytes);
614 xdp_ring->next_to_use = 0;
615 }
616
617 i40e_fill_tx_hw_ring(xdp_ring, &descs[nb_processed], nb_pkts - nb_processed,
618 &total_bytes);
619
620 /* Request an interrupt for the last frame and bump tail ptr. */
621 i40e_set_rs_bit(xdp_ring);
622 i40e_xdp_ring_update_tail(xdp_ring);
623
624 i40e_update_tx_stats(xdp_ring, nb_pkts, total_bytes);
625
626 return nb_pkts < budget;
627}
628
629/**
630 * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry
631 * @tx_ring: XDP Tx ring
632 * @tx_bi: Tx buffer info to clean
633 **/
634static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring,
635 struct i40e_tx_buffer *tx_bi)
636{
637 xdp_return_frame(tx_bi->xdpf);
638 tx_ring->xdp_tx_active--;
639 dma_unmap_single(tx_ring->dev,
640 dma_unmap_addr(tx_bi, dma),
641 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE);
642 dma_unmap_len_set(tx_bi, len, 0);
643}
644
645/**
646 * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries
647 * @vsi: Current VSI
648 * @tx_ring: XDP Tx ring
649 *
650 * Returns true if cleanup/transmission is done.
651 **/
652bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi, struct i40e_ring *tx_ring)
653{
654 struct xsk_buff_pool *bp = tx_ring->xsk_pool;
655 u32 i, completed_frames, xsk_frames = 0;
656 u32 head_idx = i40e_get_head(tx_ring);
657 struct i40e_tx_buffer *tx_bi;
658 unsigned int ntc;
659
660 if (head_idx < tx_ring->next_to_clean)
661 head_idx += tx_ring->count;
662 completed_frames = head_idx - tx_ring->next_to_clean;
663
664 if (completed_frames == 0)
665 goto out_xmit;
666
667 if (likely(!tx_ring->xdp_tx_active)) {
668 xsk_frames = completed_frames;
669 goto skip;
670 }
671
672 ntc = tx_ring->next_to_clean;
673
674 for (i = 0; i < completed_frames; i++) {
675 tx_bi = &tx_ring->tx_bi[ntc];
676
677 if (tx_bi->xdpf) {
678 i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
679 tx_bi->xdpf = NULL;
680 } else {
681 xsk_frames++;
682 }
683
684 if (++ntc >= tx_ring->count)
685 ntc = 0;
686 }
687
688skip:
689 tx_ring->next_to_clean += completed_frames;
690 if (unlikely(tx_ring->next_to_clean >= tx_ring->count))
691 tx_ring->next_to_clean -= tx_ring->count;
692
693 if (xsk_frames)
694 xsk_tx_completed(bp, xsk_frames);
695
696 i40e_arm_wb(tx_ring, vsi, completed_frames);
697
698out_xmit:
699 if (xsk_uses_need_wakeup(tx_ring->xsk_pool))
700 xsk_set_tx_need_wakeup(tx_ring->xsk_pool);
701
702 return i40e_xmit_zc(tx_ring, I40E_DESC_UNUSED(tx_ring));
703}
704
705/**
706 * i40e_xsk_wakeup - Implements the ndo_xsk_wakeup
707 * @dev: the netdevice
708 * @queue_id: queue id to wake up
709 * @flags: ignored in our case since we have Rx and Tx in the same NAPI.
710 *
711 * Returns <0 for errors, 0 otherwise.
712 **/
713int i40e_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
714{
715 struct i40e_netdev_priv *np = netdev_priv(dev);
716 struct i40e_vsi *vsi = np->vsi;
717 struct i40e_pf *pf = vsi->back;
718 struct i40e_ring *ring;
719
720 if (test_bit(__I40E_CONFIG_BUSY, pf->state))
721 return -EAGAIN;
722
723 if (test_bit(__I40E_VSI_DOWN, vsi->state))
724 return -ENETDOWN;
725
726 if (!i40e_enabled_xdp_vsi(vsi))
727 return -EINVAL;
728
729 if (queue_id >= vsi->num_queue_pairs)
730 return -EINVAL;
731
732 if (!vsi->xdp_rings[queue_id]->xsk_pool)
733 return -EINVAL;
734
735 ring = vsi->xdp_rings[queue_id];
736
737 /* The idea here is that if NAPI is running, mark a miss, so
738 * it will run again. If not, trigger an interrupt and
739 * schedule the NAPI from interrupt context. If NAPI would be
740 * scheduled here, the interrupt affinity would not be
741 * honored.
742 */
743 if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi))
744 i40e_force_wb(vsi, ring->q_vector);
745
746 return 0;
747}
748
749void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring)
750{
751 u16 ntc = rx_ring->next_to_clean;
752 u16 ntu = rx_ring->next_to_use;
753
754 while (ntc != ntu) {
755 struct xdp_buff *rx_bi = *i40e_rx_bi(rx_ring, ntc);
756
757 xsk_buff_free(rx_bi);
758 ntc++;
759 if (ntc >= rx_ring->count)
760 ntc = 0;
761 }
762}
763
764/**
765 * i40e_xsk_clean_tx_ring - Clean the XDP Tx ring on shutdown
766 * @tx_ring: XDP Tx ring
767 **/
768void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring)
769{
770 u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
771 struct xsk_buff_pool *bp = tx_ring->xsk_pool;
772 struct i40e_tx_buffer *tx_bi;
773 u32 xsk_frames = 0;
774
775 while (ntc != ntu) {
776 tx_bi = &tx_ring->tx_bi[ntc];
777
778 if (tx_bi->xdpf)
779 i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
780 else
781 xsk_frames++;
782
783 tx_bi->xdpf = NULL;
784
785 ntc++;
786 if (ntc >= tx_ring->count)
787 ntc = 0;
788 }
789
790 if (xsk_frames)
791 xsk_tx_completed(bp, xsk_frames);
792}
793
794/**
795 * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have an AF_XDP
796 * buffer pool attached
797 * @vsi: vsi
798 *
799 * Returns true if any of the Rx rings has an AF_XDP buffer pool attached
800 **/
801bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi)
802{
803 struct net_device *netdev = vsi->netdev;
804 int i;
805
806 for (i = 0; i < vsi->num_queue_pairs; i++) {
807 if (xsk_get_pool_from_qid(netdev, i))
808 return true;
809 }
810
811 return false;
812}
1// SPDX-License-Identifier: GPL-2.0
2/* Copyright(c) 2018 Intel Corporation. */
3
4#include <linux/bpf_trace.h>
5#include <net/xdp_sock.h>
6#include <net/xdp.h>
7
8#include "i40e.h"
9#include "i40e_txrx_common.h"
10#include "i40e_xsk.h"
11
12/**
13 * i40e_xsk_umem_dma_map - DMA maps all UMEM memory for the netdev
14 * @vsi: Current VSI
15 * @umem: UMEM to DMA map
16 *
17 * Returns 0 on success, <0 on failure
18 **/
19static int i40e_xsk_umem_dma_map(struct i40e_vsi *vsi, struct xdp_umem *umem)
20{
21 struct i40e_pf *pf = vsi->back;
22 struct device *dev;
23 unsigned int i, j;
24 dma_addr_t dma;
25
26 dev = &pf->pdev->dev;
27 for (i = 0; i < umem->npgs; i++) {
28 dma = dma_map_page_attrs(dev, umem->pgs[i], 0, PAGE_SIZE,
29 DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
30 if (dma_mapping_error(dev, dma))
31 goto out_unmap;
32
33 umem->pages[i].dma = dma;
34 }
35
36 return 0;
37
38out_unmap:
39 for (j = 0; j < i; j++) {
40 dma_unmap_page_attrs(dev, umem->pages[i].dma, PAGE_SIZE,
41 DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
42 umem->pages[i].dma = 0;
43 }
44
45 return -1;
46}
47
48/**
49 * i40e_xsk_umem_dma_unmap - DMA unmaps all UMEM memory for the netdev
50 * @vsi: Current VSI
51 * @umem: UMEM to DMA map
52 **/
53static void i40e_xsk_umem_dma_unmap(struct i40e_vsi *vsi, struct xdp_umem *umem)
54{
55 struct i40e_pf *pf = vsi->back;
56 struct device *dev;
57 unsigned int i;
58
59 dev = &pf->pdev->dev;
60
61 for (i = 0; i < umem->npgs; i++) {
62 dma_unmap_page_attrs(dev, umem->pages[i].dma, PAGE_SIZE,
63 DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
64
65 umem->pages[i].dma = 0;
66 }
67}
68
69/**
70 * i40e_xsk_umem_enable - Enable/associate a UMEM to a certain ring/qid
71 * @vsi: Current VSI
72 * @umem: UMEM
73 * @qid: Rx ring to associate UMEM to
74 *
75 * Returns 0 on success, <0 on failure
76 **/
77static int i40e_xsk_umem_enable(struct i40e_vsi *vsi, struct xdp_umem *umem,
78 u16 qid)
79{
80 struct net_device *netdev = vsi->netdev;
81 struct xdp_umem_fq_reuse *reuseq;
82 bool if_running;
83 int err;
84
85 if (vsi->type != I40E_VSI_MAIN)
86 return -EINVAL;
87
88 if (qid >= vsi->num_queue_pairs)
89 return -EINVAL;
90
91 if (qid >= netdev->real_num_rx_queues ||
92 qid >= netdev->real_num_tx_queues)
93 return -EINVAL;
94
95 reuseq = xsk_reuseq_prepare(vsi->rx_rings[0]->count);
96 if (!reuseq)
97 return -ENOMEM;
98
99 xsk_reuseq_free(xsk_reuseq_swap(umem, reuseq));
100
101 err = i40e_xsk_umem_dma_map(vsi, umem);
102 if (err)
103 return err;
104
105 set_bit(qid, vsi->af_xdp_zc_qps);
106
107 if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
108
109 if (if_running) {
110 err = i40e_queue_pair_disable(vsi, qid);
111 if (err)
112 return err;
113
114 err = i40e_queue_pair_enable(vsi, qid);
115 if (err)
116 return err;
117
118 /* Kick start the NAPI context so that receiving will start */
119 err = i40e_xsk_wakeup(vsi->netdev, qid, XDP_WAKEUP_RX);
120 if (err)
121 return err;
122 }
123
124 return 0;
125}
126
127/**
128 * i40e_xsk_umem_disable - Disassociate a UMEM from a certain ring/qid
129 * @vsi: Current VSI
130 * @qid: Rx ring to associate UMEM to
131 *
132 * Returns 0 on success, <0 on failure
133 **/
134static int i40e_xsk_umem_disable(struct i40e_vsi *vsi, u16 qid)
135{
136 struct net_device *netdev = vsi->netdev;
137 struct xdp_umem *umem;
138 bool if_running;
139 int err;
140
141 umem = xdp_get_umem_from_qid(netdev, qid);
142 if (!umem)
143 return -EINVAL;
144
145 if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
146
147 if (if_running) {
148 err = i40e_queue_pair_disable(vsi, qid);
149 if (err)
150 return err;
151 }
152
153 clear_bit(qid, vsi->af_xdp_zc_qps);
154 i40e_xsk_umem_dma_unmap(vsi, umem);
155
156 if (if_running) {
157 err = i40e_queue_pair_enable(vsi, qid);
158 if (err)
159 return err;
160 }
161
162 return 0;
163}
164
165/**
166 * i40e_xsk_umem_setup - Enable/disassociate a UMEM to/from a ring/qid
167 * @vsi: Current VSI
168 * @umem: UMEM to enable/associate to a ring, or NULL to disable
169 * @qid: Rx ring to (dis)associate UMEM (from)to
170 *
171 * This function enables or disables a UMEM to a certain ring.
172 *
173 * Returns 0 on success, <0 on failure
174 **/
175int i40e_xsk_umem_setup(struct i40e_vsi *vsi, struct xdp_umem *umem,
176 u16 qid)
177{
178 return umem ? i40e_xsk_umem_enable(vsi, umem, qid) :
179 i40e_xsk_umem_disable(vsi, qid);
180}
181
182/**
183 * i40e_run_xdp_zc - Executes an XDP program on an xdp_buff
184 * @rx_ring: Rx ring
185 * @xdp: xdp_buff used as input to the XDP program
186 *
187 * This function enables or disables a UMEM to a certain ring.
188 *
189 * Returns any of I40E_XDP_{PASS, CONSUMED, TX, REDIR}
190 **/
191static int i40e_run_xdp_zc(struct i40e_ring *rx_ring, struct xdp_buff *xdp)
192{
193 struct xdp_umem *umem = rx_ring->xsk_umem;
194 int err, result = I40E_XDP_PASS;
195 struct i40e_ring *xdp_ring;
196 struct bpf_prog *xdp_prog;
197 u64 offset;
198 u32 act;
199
200 rcu_read_lock();
201 /* NB! xdp_prog will always be !NULL, due to the fact that
202 * this path is enabled by setting an XDP program.
203 */
204 xdp_prog = READ_ONCE(rx_ring->xdp_prog);
205 act = bpf_prog_run_xdp(xdp_prog, xdp);
206 offset = xdp->data - xdp->data_hard_start;
207
208 xdp->handle = xsk_umem_adjust_offset(umem, xdp->handle, offset);
209
210 switch (act) {
211 case XDP_PASS:
212 break;
213 case XDP_TX:
214 xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->queue_index];
215 result = i40e_xmit_xdp_tx_ring(xdp, xdp_ring);
216 break;
217 case XDP_REDIRECT:
218 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
219 result = !err ? I40E_XDP_REDIR : I40E_XDP_CONSUMED;
220 break;
221 default:
222 bpf_warn_invalid_xdp_action(act);
223 /* fall through */
224 case XDP_ABORTED:
225 trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
226 /* fallthrough -- handle aborts by dropping packet */
227 case XDP_DROP:
228 result = I40E_XDP_CONSUMED;
229 break;
230 }
231 rcu_read_unlock();
232 return result;
233}
234
235/**
236 * i40e_alloc_buffer_zc - Allocates an i40e_rx_buffer
237 * @rx_ring: Rx ring
238 * @bi: Rx buffer to populate
239 *
240 * This function allocates an Rx buffer. The buffer can come from fill
241 * queue, or via the recycle queue (next_to_alloc).
242 *
243 * Returns true for a successful allocation, false otherwise
244 **/
245static bool i40e_alloc_buffer_zc(struct i40e_ring *rx_ring,
246 struct i40e_rx_buffer *bi)
247{
248 struct xdp_umem *umem = rx_ring->xsk_umem;
249 void *addr = bi->addr;
250 u64 handle, hr;
251
252 if (addr) {
253 rx_ring->rx_stats.page_reuse_count++;
254 return true;
255 }
256
257 if (!xsk_umem_peek_addr(umem, &handle)) {
258 rx_ring->rx_stats.alloc_page_failed++;
259 return false;
260 }
261
262 hr = umem->headroom + XDP_PACKET_HEADROOM;
263
264 bi->dma = xdp_umem_get_dma(umem, handle);
265 bi->dma += hr;
266
267 bi->addr = xdp_umem_get_data(umem, handle);
268 bi->addr += hr;
269
270 bi->handle = xsk_umem_adjust_offset(umem, handle, umem->headroom);
271
272 xsk_umem_discard_addr(umem);
273 return true;
274}
275
276/**
277 * i40e_alloc_buffer_slow_zc - Allocates an i40e_rx_buffer
278 * @rx_ring: Rx ring
279 * @bi: Rx buffer to populate
280 *
281 * This function allocates an Rx buffer. The buffer can come from fill
282 * queue, or via the reuse queue.
283 *
284 * Returns true for a successful allocation, false otherwise
285 **/
286static bool i40e_alloc_buffer_slow_zc(struct i40e_ring *rx_ring,
287 struct i40e_rx_buffer *bi)
288{
289 struct xdp_umem *umem = rx_ring->xsk_umem;
290 u64 handle, hr;
291
292 if (!xsk_umem_peek_addr_rq(umem, &handle)) {
293 rx_ring->rx_stats.alloc_page_failed++;
294 return false;
295 }
296
297 handle &= rx_ring->xsk_umem->chunk_mask;
298
299 hr = umem->headroom + XDP_PACKET_HEADROOM;
300
301 bi->dma = xdp_umem_get_dma(umem, handle);
302 bi->dma += hr;
303
304 bi->addr = xdp_umem_get_data(umem, handle);
305 bi->addr += hr;
306
307 bi->handle = xsk_umem_adjust_offset(umem, handle, umem->headroom);
308
309 xsk_umem_discard_addr_rq(umem);
310 return true;
311}
312
313static __always_inline bool
314__i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count,
315 bool alloc(struct i40e_ring *rx_ring,
316 struct i40e_rx_buffer *bi))
317{
318 u16 ntu = rx_ring->next_to_use;
319 union i40e_rx_desc *rx_desc;
320 struct i40e_rx_buffer *bi;
321 bool ok = true;
322
323 rx_desc = I40E_RX_DESC(rx_ring, ntu);
324 bi = &rx_ring->rx_bi[ntu];
325 do {
326 if (!alloc(rx_ring, bi)) {
327 ok = false;
328 goto no_buffers;
329 }
330
331 dma_sync_single_range_for_device(rx_ring->dev, bi->dma, 0,
332 rx_ring->rx_buf_len,
333 DMA_BIDIRECTIONAL);
334
335 rx_desc->read.pkt_addr = cpu_to_le64(bi->dma);
336
337 rx_desc++;
338 bi++;
339 ntu++;
340
341 if (unlikely(ntu == rx_ring->count)) {
342 rx_desc = I40E_RX_DESC(rx_ring, 0);
343 bi = rx_ring->rx_bi;
344 ntu = 0;
345 }
346
347 rx_desc->wb.qword1.status_error_len = 0;
348 count--;
349 } while (count);
350
351no_buffers:
352 if (rx_ring->next_to_use != ntu)
353 i40e_release_rx_desc(rx_ring, ntu);
354
355 return ok;
356}
357
358/**
359 * i40e_alloc_rx_buffers_zc - Allocates a number of Rx buffers
360 * @rx_ring: Rx ring
361 * @count: The number of buffers to allocate
362 *
363 * This function allocates a number of Rx buffers from the reuse queue
364 * or fill ring and places them on the Rx ring.
365 *
366 * Returns true for a successful allocation, false otherwise
367 **/
368bool i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count)
369{
370 return __i40e_alloc_rx_buffers_zc(rx_ring, count,
371 i40e_alloc_buffer_slow_zc);
372}
373
374/**
375 * i40e_alloc_rx_buffers_fast_zc - Allocates a number of Rx buffers
376 * @rx_ring: Rx ring
377 * @count: The number of buffers to allocate
378 *
379 * This function allocates a number of Rx buffers from the fill ring
380 * or the internal recycle mechanism and places them on the Rx ring.
381 *
382 * Returns true for a successful allocation, false otherwise
383 **/
384static bool i40e_alloc_rx_buffers_fast_zc(struct i40e_ring *rx_ring, u16 count)
385{
386 return __i40e_alloc_rx_buffers_zc(rx_ring, count,
387 i40e_alloc_buffer_zc);
388}
389
390/**
391 * i40e_get_rx_buffer_zc - Return the current Rx buffer
392 * @rx_ring: Rx ring
393 * @size: The size of the rx buffer (read from descriptor)
394 *
395 * This function returns the current, received Rx buffer, and also
396 * does DMA synchronization. the Rx ring.
397 *
398 * Returns the received Rx buffer
399 **/
400static struct i40e_rx_buffer *i40e_get_rx_buffer_zc(struct i40e_ring *rx_ring,
401 const unsigned int size)
402{
403 struct i40e_rx_buffer *bi;
404
405 bi = &rx_ring->rx_bi[rx_ring->next_to_clean];
406
407 /* we are reusing so sync this buffer for CPU use */
408 dma_sync_single_range_for_cpu(rx_ring->dev,
409 bi->dma, 0,
410 size,
411 DMA_BIDIRECTIONAL);
412
413 return bi;
414}
415
416/**
417 * i40e_reuse_rx_buffer_zc - Recycle an Rx buffer
418 * @rx_ring: Rx ring
419 * @old_bi: The Rx buffer to recycle
420 *
421 * This function recycles a finished Rx buffer, and places it on the
422 * recycle queue (next_to_alloc).
423 **/
424static void i40e_reuse_rx_buffer_zc(struct i40e_ring *rx_ring,
425 struct i40e_rx_buffer *old_bi)
426{
427 struct i40e_rx_buffer *new_bi = &rx_ring->rx_bi[rx_ring->next_to_alloc];
428 u16 nta = rx_ring->next_to_alloc;
429
430 /* update, and store next to alloc */
431 nta++;
432 rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
433
434 /* transfer page from old buffer to new buffer */
435 new_bi->dma = old_bi->dma;
436 new_bi->addr = old_bi->addr;
437 new_bi->handle = old_bi->handle;
438
439 old_bi->addr = NULL;
440}
441
442/**
443 * i40e_zca_free - Free callback for MEM_TYPE_ZERO_COPY allocations
444 * @alloc: Zero-copy allocator
445 * @handle: Buffer handle
446 **/
447void i40e_zca_free(struct zero_copy_allocator *alloc, unsigned long handle)
448{
449 struct i40e_rx_buffer *bi;
450 struct i40e_ring *rx_ring;
451 u64 hr, mask;
452 u16 nta;
453
454 rx_ring = container_of(alloc, struct i40e_ring, zca);
455 hr = rx_ring->xsk_umem->headroom + XDP_PACKET_HEADROOM;
456 mask = rx_ring->xsk_umem->chunk_mask;
457
458 nta = rx_ring->next_to_alloc;
459 bi = &rx_ring->rx_bi[nta];
460
461 nta++;
462 rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
463
464 handle &= mask;
465
466 bi->dma = xdp_umem_get_dma(rx_ring->xsk_umem, handle);
467 bi->dma += hr;
468
469 bi->addr = xdp_umem_get_data(rx_ring->xsk_umem, handle);
470 bi->addr += hr;
471
472 bi->handle = xsk_umem_adjust_offset(rx_ring->xsk_umem, (u64)handle,
473 rx_ring->xsk_umem->headroom);
474}
475
476/**
477 * i40e_construct_skb_zc - Create skbufff from zero-copy Rx buffer
478 * @rx_ring: Rx ring
479 * @bi: Rx buffer
480 * @xdp: xdp_buff
481 *
482 * This functions allocates a new skb from a zero-copy Rx buffer.
483 *
484 * Returns the skb, or NULL on failure.
485 **/
486static struct sk_buff *i40e_construct_skb_zc(struct i40e_ring *rx_ring,
487 struct i40e_rx_buffer *bi,
488 struct xdp_buff *xdp)
489{
490 unsigned int metasize = xdp->data - xdp->data_meta;
491 unsigned int datasize = xdp->data_end - xdp->data;
492 struct sk_buff *skb;
493
494 /* allocate a skb to store the frags */
495 skb = __napi_alloc_skb(&rx_ring->q_vector->napi,
496 xdp->data_end - xdp->data_hard_start,
497 GFP_ATOMIC | __GFP_NOWARN);
498 if (unlikely(!skb))
499 return NULL;
500
501 skb_reserve(skb, xdp->data - xdp->data_hard_start);
502 memcpy(__skb_put(skb, datasize), xdp->data, datasize);
503 if (metasize)
504 skb_metadata_set(skb, metasize);
505
506 i40e_reuse_rx_buffer_zc(rx_ring, bi);
507 return skb;
508}
509
510/**
511 * i40e_inc_ntc: Advance the next_to_clean index
512 * @rx_ring: Rx ring
513 **/
514static void i40e_inc_ntc(struct i40e_ring *rx_ring)
515{
516 u32 ntc = rx_ring->next_to_clean + 1;
517
518 ntc = (ntc < rx_ring->count) ? ntc : 0;
519 rx_ring->next_to_clean = ntc;
520 prefetch(I40E_RX_DESC(rx_ring, ntc));
521}
522
523/**
524 * i40e_clean_rx_irq_zc - Consumes Rx packets from the hardware ring
525 * @rx_ring: Rx ring
526 * @budget: NAPI budget
527 *
528 * Returns amount of work completed
529 **/
530int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget)
531{
532 unsigned int total_rx_bytes = 0, total_rx_packets = 0;
533 u16 cleaned_count = I40E_DESC_UNUSED(rx_ring);
534 unsigned int xdp_res, xdp_xmit = 0;
535 bool failure = false;
536 struct sk_buff *skb;
537 struct xdp_buff xdp;
538
539 xdp.rxq = &rx_ring->xdp_rxq;
540
541 while (likely(total_rx_packets < (unsigned int)budget)) {
542 struct i40e_rx_buffer *bi;
543 union i40e_rx_desc *rx_desc;
544 unsigned int size;
545 u64 qword;
546
547 if (cleaned_count >= I40E_RX_BUFFER_WRITE) {
548 failure = failure ||
549 !i40e_alloc_rx_buffers_fast_zc(rx_ring,
550 cleaned_count);
551 cleaned_count = 0;
552 }
553
554 rx_desc = I40E_RX_DESC(rx_ring, rx_ring->next_to_clean);
555 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
556
557 /* This memory barrier is needed to keep us from reading
558 * any other fields out of the rx_desc until we have
559 * verified the descriptor has been written back.
560 */
561 dma_rmb();
562
563 bi = i40e_clean_programming_status(rx_ring, rx_desc,
564 qword);
565 if (unlikely(bi)) {
566 i40e_reuse_rx_buffer_zc(rx_ring, bi);
567 cleaned_count++;
568 continue;
569 }
570
571 size = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
572 I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
573 if (!size)
574 break;
575
576 bi = i40e_get_rx_buffer_zc(rx_ring, size);
577 xdp.data = bi->addr;
578 xdp.data_meta = xdp.data;
579 xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
580 xdp.data_end = xdp.data + size;
581 xdp.handle = bi->handle;
582
583 xdp_res = i40e_run_xdp_zc(rx_ring, &xdp);
584 if (xdp_res) {
585 if (xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR)) {
586 xdp_xmit |= xdp_res;
587 bi->addr = NULL;
588 } else {
589 i40e_reuse_rx_buffer_zc(rx_ring, bi);
590 }
591
592 total_rx_bytes += size;
593 total_rx_packets++;
594
595 cleaned_count++;
596 i40e_inc_ntc(rx_ring);
597 continue;
598 }
599
600 /* XDP_PASS path */
601
602 /* NB! We are not checking for errors using
603 * i40e_test_staterr with
604 * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that
605 * SBP is *not* set in PRT_SBPVSI (default not set).
606 */
607 skb = i40e_construct_skb_zc(rx_ring, bi, &xdp);
608 if (!skb) {
609 rx_ring->rx_stats.alloc_buff_failed++;
610 break;
611 }
612
613 cleaned_count++;
614 i40e_inc_ntc(rx_ring);
615
616 if (eth_skb_pad(skb))
617 continue;
618
619 total_rx_bytes += skb->len;
620 total_rx_packets++;
621
622 i40e_process_skb_fields(rx_ring, rx_desc, skb);
623 napi_gro_receive(&rx_ring->q_vector->napi, skb);
624 }
625
626 i40e_finalize_xdp_rx(rx_ring, xdp_xmit);
627 i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets);
628
629 if (xsk_umem_uses_need_wakeup(rx_ring->xsk_umem)) {
630 if (failure || rx_ring->next_to_clean == rx_ring->next_to_use)
631 xsk_set_rx_need_wakeup(rx_ring->xsk_umem);
632 else
633 xsk_clear_rx_need_wakeup(rx_ring->xsk_umem);
634
635 return (int)total_rx_packets;
636 }
637 return failure ? budget : (int)total_rx_packets;
638}
639
640/**
641 * i40e_xmit_zc - Performs zero-copy Tx AF_XDP
642 * @xdp_ring: XDP Tx ring
643 * @budget: NAPI budget
644 *
645 * Returns true if the work is finished.
646 **/
647static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
648{
649 struct i40e_tx_desc *tx_desc = NULL;
650 struct i40e_tx_buffer *tx_bi;
651 bool work_done = true;
652 struct xdp_desc desc;
653 dma_addr_t dma;
654
655 while (budget-- > 0) {
656 if (!unlikely(I40E_DESC_UNUSED(xdp_ring))) {
657 xdp_ring->tx_stats.tx_busy++;
658 work_done = false;
659 break;
660 }
661
662 if (!xsk_umem_consume_tx(xdp_ring->xsk_umem, &desc))
663 break;
664
665 dma = xdp_umem_get_dma(xdp_ring->xsk_umem, desc.addr);
666
667 dma_sync_single_for_device(xdp_ring->dev, dma, desc.len,
668 DMA_BIDIRECTIONAL);
669
670 tx_bi = &xdp_ring->tx_bi[xdp_ring->next_to_use];
671 tx_bi->bytecount = desc.len;
672
673 tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use);
674 tx_desc->buffer_addr = cpu_to_le64(dma);
675 tx_desc->cmd_type_offset_bsz =
676 build_ctob(I40E_TX_DESC_CMD_ICRC
677 | I40E_TX_DESC_CMD_EOP,
678 0, desc.len, 0);
679
680 xdp_ring->next_to_use++;
681 if (xdp_ring->next_to_use == xdp_ring->count)
682 xdp_ring->next_to_use = 0;
683 }
684
685 if (tx_desc) {
686 /* Request an interrupt for the last frame and bump tail ptr. */
687 tx_desc->cmd_type_offset_bsz |= (I40E_TX_DESC_CMD_RS <<
688 I40E_TXD_QW1_CMD_SHIFT);
689 i40e_xdp_ring_update_tail(xdp_ring);
690
691 xsk_umem_consume_tx_done(xdp_ring->xsk_umem);
692 }
693
694 return !!budget && work_done;
695}
696
697/**
698 * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry
699 * @tx_ring: XDP Tx ring
700 * @tx_bi: Tx buffer info to clean
701 **/
702static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring,
703 struct i40e_tx_buffer *tx_bi)
704{
705 xdp_return_frame(tx_bi->xdpf);
706 dma_unmap_single(tx_ring->dev,
707 dma_unmap_addr(tx_bi, dma),
708 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE);
709 dma_unmap_len_set(tx_bi, len, 0);
710}
711
712/**
713 * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries
714 * @tx_ring: XDP Tx ring
715 * @tx_bi: Tx buffer info to clean
716 *
717 * Returns true if cleanup/tranmission is done.
718 **/
719bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi,
720 struct i40e_ring *tx_ring, int napi_budget)
721{
722 unsigned int ntc, total_bytes = 0, budget = vsi->work_limit;
723 u32 i, completed_frames, frames_ready, xsk_frames = 0;
724 struct xdp_umem *umem = tx_ring->xsk_umem;
725 u32 head_idx = i40e_get_head(tx_ring);
726 bool work_done = true, xmit_done;
727 struct i40e_tx_buffer *tx_bi;
728
729 if (head_idx < tx_ring->next_to_clean)
730 head_idx += tx_ring->count;
731 frames_ready = head_idx - tx_ring->next_to_clean;
732
733 if (frames_ready == 0) {
734 goto out_xmit;
735 } else if (frames_ready > budget) {
736 completed_frames = budget;
737 work_done = false;
738 } else {
739 completed_frames = frames_ready;
740 }
741
742 ntc = tx_ring->next_to_clean;
743
744 for (i = 0; i < completed_frames; i++) {
745 tx_bi = &tx_ring->tx_bi[ntc];
746
747 if (tx_bi->xdpf)
748 i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
749 else
750 xsk_frames++;
751
752 tx_bi->xdpf = NULL;
753 total_bytes += tx_bi->bytecount;
754
755 if (++ntc >= tx_ring->count)
756 ntc = 0;
757 }
758
759 tx_ring->next_to_clean += completed_frames;
760 if (unlikely(tx_ring->next_to_clean >= tx_ring->count))
761 tx_ring->next_to_clean -= tx_ring->count;
762
763 if (xsk_frames)
764 xsk_umem_complete_tx(umem, xsk_frames);
765
766 i40e_arm_wb(tx_ring, vsi, budget);
767 i40e_update_tx_stats(tx_ring, completed_frames, total_bytes);
768
769out_xmit:
770 if (xsk_umem_uses_need_wakeup(tx_ring->xsk_umem))
771 xsk_set_tx_need_wakeup(tx_ring->xsk_umem);
772
773 xmit_done = i40e_xmit_zc(tx_ring, budget);
774
775 return work_done && xmit_done;
776}
777
778/**
779 * i40e_xsk_wakeup - Implements the ndo_xsk_wakeup
780 * @dev: the netdevice
781 * @queue_id: queue id to wake up
782 * @flags: ignored in our case since we have Rx and Tx in the same NAPI.
783 *
784 * Returns <0 for errors, 0 otherwise.
785 **/
786int i40e_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
787{
788 struct i40e_netdev_priv *np = netdev_priv(dev);
789 struct i40e_vsi *vsi = np->vsi;
790 struct i40e_ring *ring;
791
792 if (test_bit(__I40E_VSI_DOWN, vsi->state))
793 return -ENETDOWN;
794
795 if (!i40e_enabled_xdp_vsi(vsi))
796 return -ENXIO;
797
798 if (queue_id >= vsi->num_queue_pairs)
799 return -ENXIO;
800
801 if (!vsi->xdp_rings[queue_id]->xsk_umem)
802 return -ENXIO;
803
804 ring = vsi->xdp_rings[queue_id];
805
806 /* The idea here is that if NAPI is running, mark a miss, so
807 * it will run again. If not, trigger an interrupt and
808 * schedule the NAPI from interrupt context. If NAPI would be
809 * scheduled here, the interrupt affinity would not be
810 * honored.
811 */
812 if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi))
813 i40e_force_wb(vsi, ring->q_vector);
814
815 return 0;
816}
817
818void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring)
819{
820 u16 i;
821
822 for (i = 0; i < rx_ring->count; i++) {
823 struct i40e_rx_buffer *rx_bi = &rx_ring->rx_bi[i];
824
825 if (!rx_bi->addr)
826 continue;
827
828 xsk_umem_fq_reuse(rx_ring->xsk_umem, rx_bi->handle);
829 rx_bi->addr = NULL;
830 }
831}
832
833/**
834 * i40e_xsk_clean_xdp_ring - Clean the XDP Tx ring on shutdown
835 * @xdp_ring: XDP Tx ring
836 **/
837void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring)
838{
839 u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
840 struct xdp_umem *umem = tx_ring->xsk_umem;
841 struct i40e_tx_buffer *tx_bi;
842 u32 xsk_frames = 0;
843
844 while (ntc != ntu) {
845 tx_bi = &tx_ring->tx_bi[ntc];
846
847 if (tx_bi->xdpf)
848 i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
849 else
850 xsk_frames++;
851
852 tx_bi->xdpf = NULL;
853
854 ntc++;
855 if (ntc >= tx_ring->count)
856 ntc = 0;
857 }
858
859 if (xsk_frames)
860 xsk_umem_complete_tx(umem, xsk_frames);
861}
862
863/**
864 * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have AF_XDP UMEM attached
865 * @vsi: vsi
866 *
867 * Returns true if any of the Rx rings has an AF_XDP UMEM attached
868 **/
869bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi)
870{
871 struct net_device *netdev = vsi->netdev;
872 int i;
873
874 for (i = 0; i < vsi->num_queue_pairs; i++) {
875 if (xdp_get_umem_from_qid(netdev, i))
876 return true;
877 }
878
879 return false;
880}