]> asedeno.scripts.mit.edu Git - linux.git/blob - drivers/net/ethernet/intel/i40e/i40e_xsk.c
Merge branch 'opp/linux-next' of git://git.kernel.org/pub/scm/linux/kernel/git/vireshk/pm
[linux.git] / drivers / net / ethernet / intel / i40e / i40e_xsk.c
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_alloc_xsk_umems - Allocate an array to store per ring UMEMs
14  * @vsi: Current VSI
15  *
16  * Returns 0 on success, <0 on failure
17  **/
18 static int i40e_alloc_xsk_umems(struct i40e_vsi *vsi)
19 {
20         if (vsi->xsk_umems)
21                 return 0;
22
23         vsi->num_xsk_umems_used = 0;
24         vsi->num_xsk_umems = vsi->alloc_queue_pairs;
25         vsi->xsk_umems = kcalloc(vsi->num_xsk_umems, sizeof(*vsi->xsk_umems),
26                                  GFP_KERNEL);
27         if (!vsi->xsk_umems) {
28                 vsi->num_xsk_umems = 0;
29                 return -ENOMEM;
30         }
31
32         return 0;
33 }
34
35 /**
36  * i40e_add_xsk_umem - Store a UMEM for a certain ring/qid
37  * @vsi: Current VSI
38  * @umem: UMEM to store
39  * @qid: Ring/qid to associate with the UMEM
40  *
41  * Returns 0 on success, <0 on failure
42  **/
43 static int i40e_add_xsk_umem(struct i40e_vsi *vsi, struct xdp_umem *umem,
44                              u16 qid)
45 {
46         int err;
47
48         err = i40e_alloc_xsk_umems(vsi);
49         if (err)
50                 return err;
51
52         vsi->xsk_umems[qid] = umem;
53         vsi->num_xsk_umems_used++;
54
55         return 0;
56 }
57
58 /**
59  * i40e_remove_xsk_umem - Remove a UMEM for a certain ring/qid
60  * @vsi: Current VSI
61  * @qid: Ring/qid associated with the UMEM
62  **/
63 static void i40e_remove_xsk_umem(struct i40e_vsi *vsi, u16 qid)
64 {
65         vsi->xsk_umems[qid] = NULL;
66         vsi->num_xsk_umems_used--;
67
68         if (vsi->num_xsk_umems == 0) {
69                 kfree(vsi->xsk_umems);
70                 vsi->xsk_umems = NULL;
71                 vsi->num_xsk_umems = 0;
72         }
73 }
74
75 /**
76  * i40e_xsk_umem_dma_map - DMA maps all UMEM memory for the netdev
77  * @vsi: Current VSI
78  * @umem: UMEM to DMA map
79  *
80  * Returns 0 on success, <0 on failure
81  **/
82 static int i40e_xsk_umem_dma_map(struct i40e_vsi *vsi, struct xdp_umem *umem)
83 {
84         struct i40e_pf *pf = vsi->back;
85         struct device *dev;
86         unsigned int i, j;
87         dma_addr_t dma;
88
89         dev = &pf->pdev->dev;
90         for (i = 0; i < umem->npgs; i++) {
91                 dma = dma_map_page_attrs(dev, umem->pgs[i], 0, PAGE_SIZE,
92                                          DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
93                 if (dma_mapping_error(dev, dma))
94                         goto out_unmap;
95
96                 umem->pages[i].dma = dma;
97         }
98
99         return 0;
100
101 out_unmap:
102         for (j = 0; j < i; j++) {
103                 dma_unmap_page_attrs(dev, umem->pages[i].dma, PAGE_SIZE,
104                                      DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
105                 umem->pages[i].dma = 0;
106         }
107
108         return -1;
109 }
110
111 /**
112  * i40e_xsk_umem_dma_unmap - DMA unmaps all UMEM memory for the netdev
113  * @vsi: Current VSI
114  * @umem: UMEM to DMA map
115  **/
116 static void i40e_xsk_umem_dma_unmap(struct i40e_vsi *vsi, struct xdp_umem *umem)
117 {
118         struct i40e_pf *pf = vsi->back;
119         struct device *dev;
120         unsigned int i;
121
122         dev = &pf->pdev->dev;
123
124         for (i = 0; i < umem->npgs; i++) {
125                 dma_unmap_page_attrs(dev, umem->pages[i].dma, PAGE_SIZE,
126                                      DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
127
128                 umem->pages[i].dma = 0;
129         }
130 }
131
132 /**
133  * i40e_xsk_umem_enable - Enable/associate a UMEM to a certain ring/qid
134  * @vsi: Current VSI
135  * @umem: UMEM
136  * @qid: Rx ring to associate UMEM to
137  *
138  * Returns 0 on success, <0 on failure
139  **/
140 static int i40e_xsk_umem_enable(struct i40e_vsi *vsi, struct xdp_umem *umem,
141                                 u16 qid)
142 {
143         struct xdp_umem_fq_reuse *reuseq;
144         bool if_running;
145         int err;
146
147         if (vsi->type != I40E_VSI_MAIN)
148                 return -EINVAL;
149
150         if (qid >= vsi->num_queue_pairs)
151                 return -EINVAL;
152
153         if (vsi->xsk_umems) {
154                 if (qid >= vsi->num_xsk_umems)
155                         return -EINVAL;
156                 if (vsi->xsk_umems[qid])
157                         return -EBUSY;
158         }
159
160         reuseq = xsk_reuseq_prepare(vsi->rx_rings[0]->count);
161         if (!reuseq)
162                 return -ENOMEM;
163
164         xsk_reuseq_free(xsk_reuseq_swap(umem, reuseq));
165
166         err = i40e_xsk_umem_dma_map(vsi, umem);
167         if (err)
168                 return err;
169
170         if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
171
172         if (if_running) {
173                 err = i40e_queue_pair_disable(vsi, qid);
174                 if (err)
175                         return err;
176         }
177
178         err = i40e_add_xsk_umem(vsi, umem, qid);
179         if (err)
180                 return err;
181
182         if (if_running) {
183                 err = i40e_queue_pair_enable(vsi, qid);
184                 if (err)
185                         return err;
186         }
187
188         return 0;
189 }
190
191 /**
192  * i40e_xsk_umem_disable - Disassociate a UMEM from a certain ring/qid
193  * @vsi: Current VSI
194  * @qid: Rx ring to associate UMEM to
195  *
196  * Returns 0 on success, <0 on failure
197  **/
198 static int i40e_xsk_umem_disable(struct i40e_vsi *vsi, u16 qid)
199 {
200         bool if_running;
201         int err;
202
203         if (!vsi->xsk_umems || qid >= vsi->num_xsk_umems ||
204             !vsi->xsk_umems[qid])
205                 return -EINVAL;
206
207         if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
208
209         if (if_running) {
210                 err = i40e_queue_pair_disable(vsi, qid);
211                 if (err)
212                         return err;
213         }
214
215         i40e_xsk_umem_dma_unmap(vsi, vsi->xsk_umems[qid]);
216         i40e_remove_xsk_umem(vsi, qid);
217
218         if (if_running) {
219                 err = i40e_queue_pair_enable(vsi, qid);
220                 if (err)
221                         return err;
222         }
223
224         return 0;
225 }
226
227 /**
228  * i40e_xsk_umem_query - Queries a certain ring/qid for its UMEM
229  * @vsi: Current VSI
230  * @umem: UMEM associated to the ring, if any
231  * @qid: Rx ring to associate UMEM to
232  *
233  * This function will store, if any, the UMEM associated to certain ring.
234  *
235  * Returns 0 on success, <0 on failure
236  **/
237 int i40e_xsk_umem_query(struct i40e_vsi *vsi, struct xdp_umem **umem,
238                         u16 qid)
239 {
240         if (vsi->type != I40E_VSI_MAIN)
241                 return -EINVAL;
242
243         if (qid >= vsi->num_queue_pairs)
244                 return -EINVAL;
245
246         if (vsi->xsk_umems) {
247                 if (qid >= vsi->num_xsk_umems)
248                         return -EINVAL;
249                 *umem = vsi->xsk_umems[qid];
250                 return 0;
251         }
252
253         *umem = NULL;
254         return 0;
255 }
256
257 /**
258  * i40e_xsk_umem_setup - Enable/disassociate a UMEM to/from a ring/qid
259  * @vsi: Current VSI
260  * @umem: UMEM to enable/associate to a ring, or NULL to disable
261  * @qid: Rx ring to (dis)associate UMEM (from)to
262  *
263  * This function enables or disables a UMEM to a certain ring.
264  *
265  * Returns 0 on success, <0 on failure
266  **/
267 int i40e_xsk_umem_setup(struct i40e_vsi *vsi, struct xdp_umem *umem,
268                         u16 qid)
269 {
270         return umem ? i40e_xsk_umem_enable(vsi, umem, qid) :
271                 i40e_xsk_umem_disable(vsi, qid);
272 }
273
274 /**
275  * i40e_run_xdp_zc - Executes an XDP program on an xdp_buff
276  * @rx_ring: Rx ring
277  * @xdp: xdp_buff used as input to the XDP program
278  *
279  * This function enables or disables a UMEM to a certain ring.
280  *
281  * Returns any of I40E_XDP_{PASS, CONSUMED, TX, REDIR}
282  **/
283 static int i40e_run_xdp_zc(struct i40e_ring *rx_ring, struct xdp_buff *xdp)
284 {
285         int err, result = I40E_XDP_PASS;
286         struct i40e_ring *xdp_ring;
287         struct bpf_prog *xdp_prog;
288         u32 act;
289
290         rcu_read_lock();
291         /* NB! xdp_prog will always be !NULL, due to the fact that
292          * this path is enabled by setting an XDP program.
293          */
294         xdp_prog = READ_ONCE(rx_ring->xdp_prog);
295         act = bpf_prog_run_xdp(xdp_prog, xdp);
296         xdp->handle += xdp->data - xdp->data_hard_start;
297         switch (act) {
298         case XDP_PASS:
299                 break;
300         case XDP_TX:
301                 xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->queue_index];
302                 result = i40e_xmit_xdp_tx_ring(xdp, xdp_ring);
303                 break;
304         case XDP_REDIRECT:
305                 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
306                 result = !err ? I40E_XDP_REDIR : I40E_XDP_CONSUMED;
307                 break;
308         default:
309                 bpf_warn_invalid_xdp_action(act);
310         case XDP_ABORTED:
311                 trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
312                 /* fallthrough -- handle aborts by dropping packet */
313         case XDP_DROP:
314                 result = I40E_XDP_CONSUMED;
315                 break;
316         }
317         rcu_read_unlock();
318         return result;
319 }
320
321 /**
322  * i40e_alloc_buffer_zc - Allocates an i40e_rx_buffer
323  * @rx_ring: Rx ring
324  * @bi: Rx buffer to populate
325  *
326  * This function allocates an Rx buffer. The buffer can come from fill
327  * queue, or via the recycle queue (next_to_alloc).
328  *
329  * Returns true for a successful allocation, false otherwise
330  **/
331 static bool i40e_alloc_buffer_zc(struct i40e_ring *rx_ring,
332                                  struct i40e_rx_buffer *bi)
333 {
334         struct xdp_umem *umem = rx_ring->xsk_umem;
335         void *addr = bi->addr;
336         u64 handle, hr;
337
338         if (addr) {
339                 rx_ring->rx_stats.page_reuse_count++;
340                 return true;
341         }
342
343         if (!xsk_umem_peek_addr(umem, &handle)) {
344                 rx_ring->rx_stats.alloc_page_failed++;
345                 return false;
346         }
347
348         hr = umem->headroom + XDP_PACKET_HEADROOM;
349
350         bi->dma = xdp_umem_get_dma(umem, handle);
351         bi->dma += hr;
352
353         bi->addr = xdp_umem_get_data(umem, handle);
354         bi->addr += hr;
355
356         bi->handle = handle + umem->headroom;
357
358         xsk_umem_discard_addr(umem);
359         return true;
360 }
361
362 /**
363  * i40e_alloc_buffer_slow_zc - Allocates an i40e_rx_buffer
364  * @rx_ring: Rx ring
365  * @bi: Rx buffer to populate
366  *
367  * This function allocates an Rx buffer. The buffer can come from fill
368  * queue, or via the reuse queue.
369  *
370  * Returns true for a successful allocation, false otherwise
371  **/
372 static bool i40e_alloc_buffer_slow_zc(struct i40e_ring *rx_ring,
373                                       struct i40e_rx_buffer *bi)
374 {
375         struct xdp_umem *umem = rx_ring->xsk_umem;
376         u64 handle, hr;
377
378         if (!xsk_umem_peek_addr_rq(umem, &handle)) {
379                 rx_ring->rx_stats.alloc_page_failed++;
380                 return false;
381         }
382
383         handle &= rx_ring->xsk_umem->chunk_mask;
384
385         hr = umem->headroom + XDP_PACKET_HEADROOM;
386
387         bi->dma = xdp_umem_get_dma(umem, handle);
388         bi->dma += hr;
389
390         bi->addr = xdp_umem_get_data(umem, handle);
391         bi->addr += hr;
392
393         bi->handle = handle + umem->headroom;
394
395         xsk_umem_discard_addr_rq(umem);
396         return true;
397 }
398
399 static __always_inline bool
400 __i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count,
401                            bool alloc(struct i40e_ring *rx_ring,
402                                       struct i40e_rx_buffer *bi))
403 {
404         u16 ntu = rx_ring->next_to_use;
405         union i40e_rx_desc *rx_desc;
406         struct i40e_rx_buffer *bi;
407         bool ok = true;
408
409         rx_desc = I40E_RX_DESC(rx_ring, ntu);
410         bi = &rx_ring->rx_bi[ntu];
411         do {
412                 if (!alloc(rx_ring, bi)) {
413                         ok = false;
414                         goto no_buffers;
415                 }
416
417                 dma_sync_single_range_for_device(rx_ring->dev, bi->dma, 0,
418                                                  rx_ring->rx_buf_len,
419                                                  DMA_BIDIRECTIONAL);
420
421                 rx_desc->read.pkt_addr = cpu_to_le64(bi->dma);
422
423                 rx_desc++;
424                 bi++;
425                 ntu++;
426
427                 if (unlikely(ntu == rx_ring->count)) {
428                         rx_desc = I40E_RX_DESC(rx_ring, 0);
429                         bi = rx_ring->rx_bi;
430                         ntu = 0;
431                 }
432
433                 rx_desc->wb.qword1.status_error_len = 0;
434                 count--;
435         } while (count);
436
437 no_buffers:
438         if (rx_ring->next_to_use != ntu)
439                 i40e_release_rx_desc(rx_ring, ntu);
440
441         return ok;
442 }
443
444 /**
445  * i40e_alloc_rx_buffers_zc - Allocates a number of Rx buffers
446  * @rx_ring: Rx ring
447  * @count: The number of buffers to allocate
448  *
449  * This function allocates a number of Rx buffers from the reuse queue
450  * or fill ring and places them on the Rx ring.
451  *
452  * Returns true for a successful allocation, false otherwise
453  **/
454 bool i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count)
455 {
456         return __i40e_alloc_rx_buffers_zc(rx_ring, count,
457                                           i40e_alloc_buffer_slow_zc);
458 }
459
460 /**
461  * i40e_alloc_rx_buffers_fast_zc - Allocates a number of Rx buffers
462  * @rx_ring: Rx ring
463  * @count: The number of buffers to allocate
464  *
465  * This function allocates a number of Rx buffers from the fill ring
466  * or the internal recycle mechanism and places them on the Rx ring.
467  *
468  * Returns true for a successful allocation, false otherwise
469  **/
470 static bool i40e_alloc_rx_buffers_fast_zc(struct i40e_ring *rx_ring, u16 count)
471 {
472         return __i40e_alloc_rx_buffers_zc(rx_ring, count,
473                                           i40e_alloc_buffer_zc);
474 }
475
476 /**
477  * i40e_get_rx_buffer_zc - Return the current Rx buffer
478  * @rx_ring: Rx ring
479  * @size: The size of the rx buffer (read from descriptor)
480  *
481  * This function returns the current, received Rx buffer, and also
482  * does DMA synchronization.  the Rx ring.
483  *
484  * Returns the received Rx buffer
485  **/
486 static struct i40e_rx_buffer *i40e_get_rx_buffer_zc(struct i40e_ring *rx_ring,
487                                                     const unsigned int size)
488 {
489         struct i40e_rx_buffer *bi;
490
491         bi = &rx_ring->rx_bi[rx_ring->next_to_clean];
492
493         /* we are reusing so sync this buffer for CPU use */
494         dma_sync_single_range_for_cpu(rx_ring->dev,
495                                       bi->dma, 0,
496                                       size,
497                                       DMA_BIDIRECTIONAL);
498
499         return bi;
500 }
501
502 /**
503  * i40e_reuse_rx_buffer_zc - Recycle an Rx buffer
504  * @rx_ring: Rx ring
505  * @old_bi: The Rx buffer to recycle
506  *
507  * This function recycles a finished Rx buffer, and places it on the
508  * recycle queue (next_to_alloc).
509  **/
510 static void i40e_reuse_rx_buffer_zc(struct i40e_ring *rx_ring,
511                                     struct i40e_rx_buffer *old_bi)
512 {
513         struct i40e_rx_buffer *new_bi = &rx_ring->rx_bi[rx_ring->next_to_alloc];
514         unsigned long mask = (unsigned long)rx_ring->xsk_umem->chunk_mask;
515         u64 hr = rx_ring->xsk_umem->headroom + XDP_PACKET_HEADROOM;
516         u16 nta = rx_ring->next_to_alloc;
517
518         /* update, and store next to alloc */
519         nta++;
520         rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
521
522         /* transfer page from old buffer to new buffer */
523         new_bi->dma = old_bi->dma & mask;
524         new_bi->dma += hr;
525
526         new_bi->addr = (void *)((unsigned long)old_bi->addr & mask);
527         new_bi->addr += hr;
528
529         new_bi->handle = old_bi->handle & mask;
530         new_bi->handle += rx_ring->xsk_umem->headroom;
531
532         old_bi->addr = NULL;
533 }
534
535 /**
536  * i40e_zca_free - Free callback for MEM_TYPE_ZERO_COPY allocations
537  * @alloc: Zero-copy allocator
538  * @handle: Buffer handle
539  **/
540 void i40e_zca_free(struct zero_copy_allocator *alloc, unsigned long handle)
541 {
542         struct i40e_rx_buffer *bi;
543         struct i40e_ring *rx_ring;
544         u64 hr, mask;
545         u16 nta;
546
547         rx_ring = container_of(alloc, struct i40e_ring, zca);
548         hr = rx_ring->xsk_umem->headroom + XDP_PACKET_HEADROOM;
549         mask = rx_ring->xsk_umem->chunk_mask;
550
551         nta = rx_ring->next_to_alloc;
552         bi = &rx_ring->rx_bi[nta];
553
554         nta++;
555         rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
556
557         handle &= mask;
558
559         bi->dma = xdp_umem_get_dma(rx_ring->xsk_umem, handle);
560         bi->dma += hr;
561
562         bi->addr = xdp_umem_get_data(rx_ring->xsk_umem, handle);
563         bi->addr += hr;
564
565         bi->handle = (u64)handle + rx_ring->xsk_umem->headroom;
566 }
567
568 /**
569  * i40e_construct_skb_zc - Create skbufff from zero-copy Rx buffer
570  * @rx_ring: Rx ring
571  * @bi: Rx buffer
572  * @xdp: xdp_buff
573  *
574  * This functions allocates a new skb from a zero-copy Rx buffer.
575  *
576  * Returns the skb, or NULL on failure.
577  **/
578 static struct sk_buff *i40e_construct_skb_zc(struct i40e_ring *rx_ring,
579                                              struct i40e_rx_buffer *bi,
580                                              struct xdp_buff *xdp)
581 {
582         unsigned int metasize = xdp->data - xdp->data_meta;
583         unsigned int datasize = xdp->data_end - xdp->data;
584         struct sk_buff *skb;
585
586         /* allocate a skb to store the frags */
587         skb = __napi_alloc_skb(&rx_ring->q_vector->napi,
588                                xdp->data_end - xdp->data_hard_start,
589                                GFP_ATOMIC | __GFP_NOWARN);
590         if (unlikely(!skb))
591                 return NULL;
592
593         skb_reserve(skb, xdp->data - xdp->data_hard_start);
594         memcpy(__skb_put(skb, datasize), xdp->data, datasize);
595         if (metasize)
596                 skb_metadata_set(skb, metasize);
597
598         i40e_reuse_rx_buffer_zc(rx_ring, bi);
599         return skb;
600 }
601
602 /**
603  * i40e_inc_ntc: Advance the next_to_clean index
604  * @rx_ring: Rx ring
605  **/
606 static void i40e_inc_ntc(struct i40e_ring *rx_ring)
607 {
608         u32 ntc = rx_ring->next_to_clean + 1;
609
610         ntc = (ntc < rx_ring->count) ? ntc : 0;
611         rx_ring->next_to_clean = ntc;
612         prefetch(I40E_RX_DESC(rx_ring, ntc));
613 }
614
615 /**
616  * i40e_clean_rx_irq_zc - Consumes Rx packets from the hardware ring
617  * @rx_ring: Rx ring
618  * @budget: NAPI budget
619  *
620  * Returns amount of work completed
621  **/
622 int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget)
623 {
624         unsigned int total_rx_bytes = 0, total_rx_packets = 0;
625         u16 cleaned_count = I40E_DESC_UNUSED(rx_ring);
626         unsigned int xdp_res, xdp_xmit = 0;
627         bool failure = false;
628         struct sk_buff *skb;
629         struct xdp_buff xdp;
630
631         xdp.rxq = &rx_ring->xdp_rxq;
632
633         while (likely(total_rx_packets < (unsigned int)budget)) {
634                 struct i40e_rx_buffer *bi;
635                 union i40e_rx_desc *rx_desc;
636                 unsigned int size;
637                 u16 vlan_tag;
638                 u8 rx_ptype;
639                 u64 qword;
640
641                 if (cleaned_count >= I40E_RX_BUFFER_WRITE) {
642                         failure = failure ||
643                                   !i40e_alloc_rx_buffers_fast_zc(rx_ring,
644                                                                  cleaned_count);
645                         cleaned_count = 0;
646                 }
647
648                 rx_desc = I40E_RX_DESC(rx_ring, rx_ring->next_to_clean);
649                 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
650
651                 /* This memory barrier is needed to keep us from reading
652                  * any other fields out of the rx_desc until we have
653                  * verified the descriptor has been written back.
654                  */
655                 dma_rmb();
656
657                 bi = i40e_clean_programming_status(rx_ring, rx_desc,
658                                                    qword);
659                 if (unlikely(bi)) {
660                         i40e_reuse_rx_buffer_zc(rx_ring, bi);
661                         cleaned_count++;
662                         continue;
663                 }
664
665                 size = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
666                        I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
667                 if (!size)
668                         break;
669
670                 bi = i40e_get_rx_buffer_zc(rx_ring, size);
671                 xdp.data = bi->addr;
672                 xdp.data_meta = xdp.data;
673                 xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
674                 xdp.data_end = xdp.data + size;
675                 xdp.handle = bi->handle;
676
677                 xdp_res = i40e_run_xdp_zc(rx_ring, &xdp);
678                 if (xdp_res) {
679                         if (xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR)) {
680                                 xdp_xmit |= xdp_res;
681                                 bi->addr = NULL;
682                         } else {
683                                 i40e_reuse_rx_buffer_zc(rx_ring, bi);
684                         }
685
686                         total_rx_bytes += size;
687                         total_rx_packets++;
688
689                         cleaned_count++;
690                         i40e_inc_ntc(rx_ring);
691                         continue;
692                 }
693
694                 /* XDP_PASS path */
695
696                 /* NB! We are not checking for errors using
697                  * i40e_test_staterr with
698                  * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that
699                  * SBP is *not* set in PRT_SBPVSI (default not set).
700                  */
701                 skb = i40e_construct_skb_zc(rx_ring, bi, &xdp);
702                 if (!skb) {
703                         rx_ring->rx_stats.alloc_buff_failed++;
704                         break;
705                 }
706
707                 cleaned_count++;
708                 i40e_inc_ntc(rx_ring);
709
710                 if (eth_skb_pad(skb))
711                         continue;
712
713                 total_rx_bytes += skb->len;
714                 total_rx_packets++;
715
716                 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
717                 rx_ptype = (qword & I40E_RXD_QW1_PTYPE_MASK) >>
718                            I40E_RXD_QW1_PTYPE_SHIFT;
719                 i40e_process_skb_fields(rx_ring, rx_desc, skb, rx_ptype);
720
721                 vlan_tag = (qword & BIT(I40E_RX_DESC_STATUS_L2TAG1P_SHIFT)) ?
722                            le16_to_cpu(rx_desc->wb.qword0.lo_dword.l2tag1) : 0;
723                 i40e_receive_skb(rx_ring, skb, vlan_tag);
724         }
725
726         i40e_finalize_xdp_rx(rx_ring, xdp_xmit);
727         i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets);
728         return failure ? budget : (int)total_rx_packets;
729 }
730
731 /**
732  * i40e_xmit_zc - Performs zero-copy Tx AF_XDP
733  * @xdp_ring: XDP Tx ring
734  * @budget: NAPI budget
735  *
736  * Returns true if the work is finished.
737  **/
738 static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
739 {
740         struct i40e_tx_desc *tx_desc = NULL;
741         struct i40e_tx_buffer *tx_bi;
742         bool work_done = true;
743         dma_addr_t dma;
744         u32 len;
745
746         while (budget-- > 0) {
747                 if (!unlikely(I40E_DESC_UNUSED(xdp_ring))) {
748                         xdp_ring->tx_stats.tx_busy++;
749                         work_done = false;
750                         break;
751                 }
752
753                 if (!xsk_umem_consume_tx(xdp_ring->xsk_umem, &dma, &len))
754                         break;
755
756                 dma_sync_single_for_device(xdp_ring->dev, dma, len,
757                                            DMA_BIDIRECTIONAL);
758
759                 tx_bi = &xdp_ring->tx_bi[xdp_ring->next_to_use];
760                 tx_bi->bytecount = len;
761
762                 tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use);
763                 tx_desc->buffer_addr = cpu_to_le64(dma);
764                 tx_desc->cmd_type_offset_bsz =
765                         build_ctob(I40E_TX_DESC_CMD_ICRC
766                                    | I40E_TX_DESC_CMD_EOP,
767                                    0, len, 0);
768
769                 xdp_ring->next_to_use++;
770                 if (xdp_ring->next_to_use == xdp_ring->count)
771                         xdp_ring->next_to_use = 0;
772         }
773
774         if (tx_desc) {
775                 /* Request an interrupt for the last frame and bump tail ptr. */
776                 tx_desc->cmd_type_offset_bsz |= (I40E_TX_DESC_CMD_RS <<
777                                                  I40E_TXD_QW1_CMD_SHIFT);
778                 i40e_xdp_ring_update_tail(xdp_ring);
779
780                 xsk_umem_consume_tx_done(xdp_ring->xsk_umem);
781         }
782
783         return !!budget && work_done;
784 }
785
786 /**
787  * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry
788  * @tx_ring: XDP Tx ring
789  * @tx_bi: Tx buffer info to clean
790  **/
791 static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring,
792                                      struct i40e_tx_buffer *tx_bi)
793 {
794         xdp_return_frame(tx_bi->xdpf);
795         dma_unmap_single(tx_ring->dev,
796                          dma_unmap_addr(tx_bi, dma),
797                          dma_unmap_len(tx_bi, len), DMA_TO_DEVICE);
798         dma_unmap_len_set(tx_bi, len, 0);
799 }
800
801 /**
802  * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries
803  * @tx_ring: XDP Tx ring
804  * @tx_bi: Tx buffer info to clean
805  *
806  * Returns true if cleanup/tranmission is done.
807  **/
808 bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi,
809                            struct i40e_ring *tx_ring, int napi_budget)
810 {
811         unsigned int ntc, total_bytes = 0, budget = vsi->work_limit;
812         u32 i, completed_frames, frames_ready, xsk_frames = 0;
813         struct xdp_umem *umem = tx_ring->xsk_umem;
814         u32 head_idx = i40e_get_head(tx_ring);
815         bool work_done = true, xmit_done;
816         struct i40e_tx_buffer *tx_bi;
817
818         if (head_idx < tx_ring->next_to_clean)
819                 head_idx += tx_ring->count;
820         frames_ready = head_idx - tx_ring->next_to_clean;
821
822         if (frames_ready == 0) {
823                 goto out_xmit;
824         } else if (frames_ready > budget) {
825                 completed_frames = budget;
826                 work_done = false;
827         } else {
828                 completed_frames = frames_ready;
829         }
830
831         ntc = tx_ring->next_to_clean;
832
833         for (i = 0; i < completed_frames; i++) {
834                 tx_bi = &tx_ring->tx_bi[ntc];
835
836                 if (tx_bi->xdpf)
837                         i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
838                 else
839                         xsk_frames++;
840
841                 tx_bi->xdpf = NULL;
842                 total_bytes += tx_bi->bytecount;
843
844                 if (++ntc >= tx_ring->count)
845                         ntc = 0;
846         }
847
848         tx_ring->next_to_clean += completed_frames;
849         if (unlikely(tx_ring->next_to_clean >= tx_ring->count))
850                 tx_ring->next_to_clean -= tx_ring->count;
851
852         if (xsk_frames)
853                 xsk_umem_complete_tx(umem, xsk_frames);
854
855         i40e_arm_wb(tx_ring, vsi, budget);
856         i40e_update_tx_stats(tx_ring, completed_frames, total_bytes);
857
858 out_xmit:
859         xmit_done = i40e_xmit_zc(tx_ring, budget);
860
861         return work_done && xmit_done;
862 }
863
864 /**
865  * i40e_xsk_async_xmit - Implements the ndo_xsk_async_xmit
866  * @dev: the netdevice
867  * @queue_id: queue id to wake up
868  *
869  * Returns <0 for errors, 0 otherwise.
870  **/
871 int i40e_xsk_async_xmit(struct net_device *dev, u32 queue_id)
872 {
873         struct i40e_netdev_priv *np = netdev_priv(dev);
874         struct i40e_vsi *vsi = np->vsi;
875         struct i40e_ring *ring;
876
877         if (test_bit(__I40E_VSI_DOWN, vsi->state))
878                 return -ENETDOWN;
879
880         if (!i40e_enabled_xdp_vsi(vsi))
881                 return -ENXIO;
882
883         if (queue_id >= vsi->num_queue_pairs)
884                 return -ENXIO;
885
886         if (!vsi->xdp_rings[queue_id]->xsk_umem)
887                 return -ENXIO;
888
889         ring = vsi->xdp_rings[queue_id];
890
891         /* The idea here is that if NAPI is running, mark a miss, so
892          * it will run again. If not, trigger an interrupt and
893          * schedule the NAPI from interrupt context. If NAPI would be
894          * scheduled here, the interrupt affinity would not be
895          * honored.
896          */
897         if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi))
898                 i40e_force_wb(vsi, ring->q_vector);
899
900         return 0;
901 }
902
903 void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring)
904 {
905         u16 i;
906
907         for (i = 0; i < rx_ring->count; i++) {
908                 struct i40e_rx_buffer *rx_bi = &rx_ring->rx_bi[i];
909
910                 if (!rx_bi->addr)
911                         continue;
912
913                 xsk_umem_fq_reuse(rx_ring->xsk_umem, rx_bi->handle);
914                 rx_bi->addr = NULL;
915         }
916 }
917
918 /**
919  * i40e_xsk_clean_xdp_ring - Clean the XDP Tx ring on shutdown
920  * @xdp_ring: XDP Tx ring
921  **/
922 void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring)
923 {
924         u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
925         struct xdp_umem *umem = tx_ring->xsk_umem;
926         struct i40e_tx_buffer *tx_bi;
927         u32 xsk_frames = 0;
928
929         while (ntc != ntu) {
930                 tx_bi = &tx_ring->tx_bi[ntc];
931
932                 if (tx_bi->xdpf)
933                         i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
934                 else
935                         xsk_frames++;
936
937                 tx_bi->xdpf = NULL;
938
939                 ntc++;
940                 if (ntc >= tx_ring->count)
941                         ntc = 0;
942         }
943
944         if (xsk_frames)
945                 xsk_umem_complete_tx(umem, xsk_frames);
946 }
947
948 /**
949  * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have AF_XDP UMEM attached
950  * @vsi: vsi
951  *
952  * Returns true if any of the Rx rings has an AF_XDP UMEM attached
953  **/
954 bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi)
955 {
956         int i;
957
958         if (!vsi->xsk_umems)
959                 return false;
960
961         for (i = 0; i < vsi->num_queue_pairs; i++) {
962                 if (vsi->xsk_umems[i])
963                         return true;
964         }
965
966         return false;
967 }