1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Atlantic Network Driver
3 *
4 * Copyright (C) 2014-2019 aQuantia Corporation
5 * Copyright (C) 2019-2020 Marvell International Ltd.
6 */
7
8 /* File aq_ring.c: Definition of functions for Rx/Tx rings. */
9
10 #include "aq_nic.h"
11 #include "aq_hw.h"
12 #include "aq_hw_utils.h"
13 #include "aq_ptp.h"
14 #include "aq_vec.h"
15 #include "aq_main.h"
16
17 #include <net/xdp.h>
18 #include <linux/filter.h>
19 #include <linux/bpf_trace.h>
20 #include <linux/netdevice.h>
21 #include <linux/etherdevice.h>
22
aq_get_rxpages_xdp(struct aq_ring_buff_s * buff,struct xdp_buff * xdp)23 static void aq_get_rxpages_xdp(struct aq_ring_buff_s *buff,
24 struct xdp_buff *xdp)
25 {
26 struct skb_shared_info *sinfo;
27 int i;
28
29 if (xdp_buff_has_frags(xdp)) {
30 sinfo = xdp_get_shared_info_from_buff(xdp);
31
32 for (i = 0; i < sinfo->nr_frags; i++) {
33 skb_frag_t *frag = &sinfo->frags[i];
34
35 page_ref_inc(skb_frag_page(frag));
36 }
37 }
38 page_ref_inc(buff->rxdata.page);
39 }
40
aq_free_rxpage(struct aq_rxpage * rxpage,struct device * dev)41 static inline void aq_free_rxpage(struct aq_rxpage *rxpage, struct device *dev)
42 {
43 unsigned int len = PAGE_SIZE << rxpage->order;
44
45 dma_unmap_page(dev, rxpage->daddr, len, DMA_FROM_DEVICE);
46
47 /* Drop the ref for being in the ring. */
48 __free_pages(rxpage->page, rxpage->order);
49 rxpage->page = NULL;
50 }
51
aq_alloc_rxpages(struct aq_rxpage * rxpage,struct aq_ring_s * rx_ring)52 static int aq_alloc_rxpages(struct aq_rxpage *rxpage, struct aq_ring_s *rx_ring)
53 {
54 struct device *dev = aq_nic_get_dev(rx_ring->aq_nic);
55 unsigned int order = rx_ring->page_order;
56 struct page *page;
57 int ret = -ENOMEM;
58 dma_addr_t daddr;
59
60 page = dev_alloc_pages(order);
61 if (unlikely(!page))
62 goto err_exit;
63
64 daddr = dma_map_page(dev, page, 0, PAGE_SIZE << order,
65 DMA_FROM_DEVICE);
66
67 if (unlikely(dma_mapping_error(dev, daddr)))
68 goto free_page;
69
70 rxpage->page = page;
71 rxpage->daddr = daddr;
72 rxpage->order = order;
73 rxpage->pg_off = rx_ring->page_offset;
74
75 return 0;
76
77 free_page:
78 __free_pages(page, order);
79
80 err_exit:
81 return ret;
82 }
83
aq_get_rxpages(struct aq_ring_s * self,struct aq_ring_buff_s * rxbuf)84 static int aq_get_rxpages(struct aq_ring_s *self, struct aq_ring_buff_s *rxbuf)
85 {
86 unsigned int order = self->page_order;
87 u16 page_offset = self->page_offset;
88 u16 frame_max = self->frame_max;
89 u16 tail_size = self->tail_size;
90 int ret;
91
92 if (rxbuf->rxdata.page) {
93 /* One means ring is the only user and can reuse */
94 if (page_ref_count(rxbuf->rxdata.page) > 1) {
95 /* Try reuse buffer */
96 rxbuf->rxdata.pg_off += frame_max + page_offset +
97 tail_size;
98 if (rxbuf->rxdata.pg_off + frame_max + tail_size <=
99 (PAGE_SIZE << order)) {
100 u64_stats_update_begin(&self->stats.rx.syncp);
101 self->stats.rx.pg_flips++;
102 u64_stats_update_end(&self->stats.rx.syncp);
103
104 } else {
105 /* Buffer exhausted. We have other users and
106 * should release this page and realloc
107 */
108 aq_free_rxpage(&rxbuf->rxdata,
109 aq_nic_get_dev(self->aq_nic));
110 u64_stats_update_begin(&self->stats.rx.syncp);
111 self->stats.rx.pg_losts++;
112 u64_stats_update_end(&self->stats.rx.syncp);
113 }
114 } else {
115 rxbuf->rxdata.pg_off = page_offset;
116 u64_stats_update_begin(&self->stats.rx.syncp);
117 self->stats.rx.pg_reuses++;
118 u64_stats_update_end(&self->stats.rx.syncp);
119 }
120 }
121
122 if (!rxbuf->rxdata.page) {
123 ret = aq_alloc_rxpages(&rxbuf->rxdata, self);
124 if (ret) {
125 u64_stats_update_begin(&self->stats.rx.syncp);
126 self->stats.rx.alloc_fails++;
127 u64_stats_update_end(&self->stats.rx.syncp);
128 }
129 return ret;
130 }
131
132 return 0;
133 }
134
aq_ring_alloc(struct aq_ring_s * self,struct aq_nic_s * aq_nic)135 static int aq_ring_alloc(struct aq_ring_s *self,
136 struct aq_nic_s *aq_nic)
137 {
138 int err = 0;
139
140 self->buff_ring =
141 kzalloc_objs(struct aq_ring_buff_s, self->size);
142
143 if (!self->buff_ring) {
144 err = -ENOMEM;
145 goto err_exit;
146 }
147
148 self->dx_ring = dma_alloc_coherent(aq_nic_get_dev(aq_nic),
149 self->size * self->dx_size,
150 &self->dx_ring_pa, GFP_KERNEL);
151 if (!self->dx_ring) {
152 err = -ENOMEM;
153 goto err_exit;
154 }
155
156 err_exit:
157 if (err < 0) {
158 aq_ring_free(self);
159 }
160
161 return err;
162 }
163
aq_ring_tx_alloc(struct aq_ring_s * self,struct aq_nic_s * aq_nic,unsigned int idx,struct aq_nic_cfg_s * aq_nic_cfg)164 int aq_ring_tx_alloc(struct aq_ring_s *self,
165 struct aq_nic_s *aq_nic,
166 unsigned int idx,
167 struct aq_nic_cfg_s *aq_nic_cfg)
168 {
169 self->aq_nic = aq_nic;
170 self->idx = idx;
171 self->size = aq_nic_cfg->txds;
172 self->dx_size = aq_nic_cfg->aq_hw_caps->txd_size;
173
174 return aq_ring_alloc(self, aq_nic);
175 }
176
aq_ring_rx_alloc(struct aq_ring_s * self,struct aq_nic_s * aq_nic,unsigned int idx,struct aq_nic_cfg_s * aq_nic_cfg)177 int aq_ring_rx_alloc(struct aq_ring_s *self,
178 struct aq_nic_s *aq_nic,
179 unsigned int idx,
180 struct aq_nic_cfg_s *aq_nic_cfg)
181 {
182 self->aq_nic = aq_nic;
183 self->idx = idx;
184 self->size = aq_nic_cfg->rxds;
185 self->dx_size = aq_nic_cfg->aq_hw_caps->rxd_size;
186 self->xdp_prog = aq_nic->xdp_prog;
187 self->frame_max = AQ_CFG_RX_FRAME_MAX;
188
189 /* Only order-2 is allowed if XDP is enabled */
190 if (READ_ONCE(self->xdp_prog)) {
191 self->page_offset = AQ_XDP_HEADROOM;
192 self->page_order = AQ_CFG_XDP_PAGEORDER;
193 self->tail_size = AQ_XDP_TAILROOM;
194 } else {
195 self->page_offset = 0;
196 self->page_order = fls(self->frame_max / PAGE_SIZE +
197 (self->frame_max % PAGE_SIZE ? 1 : 0)) - 1;
198 if (aq_nic_cfg->rxpageorder > self->page_order)
199 self->page_order = aq_nic_cfg->rxpageorder;
200 self->tail_size = 0;
201 }
202
203 return aq_ring_alloc(self, aq_nic);
204 }
205
206 int
aq_ring_hwts_rx_alloc(struct aq_ring_s * self,struct aq_nic_s * aq_nic,unsigned int idx,unsigned int size,unsigned int dx_size)207 aq_ring_hwts_rx_alloc(struct aq_ring_s *self, struct aq_nic_s *aq_nic,
208 unsigned int idx, unsigned int size, unsigned int dx_size)
209 {
210 struct device *dev = aq_nic_get_dev(aq_nic);
211 size_t sz = size * dx_size + AQ_CFG_RXDS_DEF;
212
213 memset(self, 0, sizeof(*self));
214
215 self->aq_nic = aq_nic;
216 self->idx = idx;
217 self->size = size;
218 self->dx_size = dx_size;
219
220 self->dx_ring = dma_alloc_coherent(dev, sz, &self->dx_ring_pa,
221 GFP_KERNEL);
222 if (!self->dx_ring) {
223 aq_ring_free(self);
224 return -ENOMEM;
225 }
226
227 return 0;
228 }
229
aq_ring_init(struct aq_ring_s * self,const enum atl_ring_type ring_type)230 int aq_ring_init(struct aq_ring_s *self, const enum atl_ring_type ring_type)
231 {
232 self->hw_head = 0;
233 self->sw_head = 0;
234 self->sw_tail = 0;
235 self->ring_type = ring_type;
236
237 if (self->ring_type == ATL_RING_RX)
238 u64_stats_init(&self->stats.rx.syncp);
239 else
240 u64_stats_init(&self->stats.tx.syncp);
241
242 return 0;
243 }
244
aq_ring_dx_in_range(unsigned int h,unsigned int i,unsigned int t)245 static inline bool aq_ring_dx_in_range(unsigned int h, unsigned int i,
246 unsigned int t)
247 {
248 return (h < t) ? ((h < i) && (i < t)) : ((h < i) || (i < t));
249 }
250
aq_ring_update_queue_state(struct aq_ring_s * ring)251 void aq_ring_update_queue_state(struct aq_ring_s *ring)
252 {
253 if (aq_ring_avail_dx(ring) <= AQ_CFG_SKB_FRAGS_MAX)
254 aq_ring_queue_stop(ring);
255 else if (aq_ring_avail_dx(ring) > AQ_CFG_RESTART_DESC_THRES)
256 aq_ring_queue_wake(ring);
257 }
258
aq_ring_queue_wake(struct aq_ring_s * ring)259 void aq_ring_queue_wake(struct aq_ring_s *ring)
260 {
261 struct net_device *ndev = aq_nic_get_ndev(ring->aq_nic);
262
263 if (__netif_subqueue_stopped(ndev,
264 AQ_NIC_RING2QMAP(ring->aq_nic,
265 ring->idx))) {
266 netif_wake_subqueue(ndev,
267 AQ_NIC_RING2QMAP(ring->aq_nic, ring->idx));
268 u64_stats_update_begin(&ring->stats.tx.syncp);
269 ring->stats.tx.queue_restarts++;
270 u64_stats_update_end(&ring->stats.tx.syncp);
271 }
272 }
273
aq_ring_queue_stop(struct aq_ring_s * ring)274 void aq_ring_queue_stop(struct aq_ring_s *ring)
275 {
276 struct net_device *ndev = aq_nic_get_ndev(ring->aq_nic);
277
278 if (!__netif_subqueue_stopped(ndev,
279 AQ_NIC_RING2QMAP(ring->aq_nic,
280 ring->idx)))
281 netif_stop_subqueue(ndev,
282 AQ_NIC_RING2QMAP(ring->aq_nic, ring->idx));
283 }
284
aq_ring_tx_clean(struct aq_ring_s * self)285 bool aq_ring_tx_clean(struct aq_ring_s *self)
286 {
287 struct device *dev = aq_nic_get_dev(self->aq_nic);
288 unsigned int budget;
289
290 for (budget = AQ_CFG_TX_CLEAN_BUDGET;
291 budget && self->sw_head != self->hw_head; budget--) {
292 struct aq_ring_buff_s *buff = &self->buff_ring[self->sw_head];
293
294 if (likely(buff->is_mapped)) {
295 if (unlikely(buff->is_sop)) {
296 if (!buff->is_eop &&
297 buff->eop_index != 0xffffU &&
298 (!aq_ring_dx_in_range(self->sw_head,
299 buff->eop_index,
300 self->hw_head)))
301 break;
302
303 dma_unmap_single(dev, buff->pa, buff->len,
304 DMA_TO_DEVICE);
305 } else {
306 dma_unmap_page(dev, buff->pa, buff->len,
307 DMA_TO_DEVICE);
308 }
309 }
310
311 if (likely(!buff->is_eop))
312 goto out;
313
314 if (unlikely(buff->request_ts) &&
315 self->aq_nic->aq_hw_ops->hw_ring_tx_ptp_get_ts) {
316 u64 ts = self->aq_nic->aq_hw_ops->hw_ring_tx_ptp_get_ts(self);
317
318 if (!ts) {
319 if (time_after(jiffies,
320 self->ptp_ts_deadline)) {
321 /* Timeout: drain skb_ring head to
322 * keep in sync with buff_ring
323 */
324 aq_ptp_tx_skb_drop_head(self->aq_nic);
325 buff->request_ts = 0;
326 dev_kfree_skb_any(buff->skb);
327 buff->skb = NULL;
328 goto out;
329 } else {
330 buff->is_mapped = 0;
331 buff->pa = 0U;
332 break;
333 }
334 }
335
336 aq_ptp_tx_hwtstamp(self->aq_nic, ts);
337 }
338 if (buff->skb) {
339 u64_stats_update_begin(&self->stats.tx.syncp);
340 ++self->stats.tx.packets;
341 self->stats.tx.bytes += buff->skb->len;
342 u64_stats_update_end(&self->stats.tx.syncp);
343 dev_kfree_skb_any(buff->skb);
344 } else if (buff->xdpf) {
345 u64_stats_update_begin(&self->stats.tx.syncp);
346 ++self->stats.tx.packets;
347 self->stats.tx.bytes += xdp_get_frame_len(buff->xdpf);
348 u64_stats_update_end(&self->stats.tx.syncp);
349 xdp_return_frame_rx_napi(buff->xdpf);
350 }
351
352 out:
353 buff->skb = NULL;
354 buff->xdpf = NULL;
355 buff->pa = 0U;
356 buff->eop_index = 0xffffU;
357 self->sw_head = aq_ring_next_dx(self, self->sw_head);
358 }
359
360 return !!budget;
361 }
362
aq_ring_tx_deinit(struct aq_ring_s * self)363 void aq_ring_tx_deinit(struct aq_ring_s *self)
364 {
365 if (!self)
366 return;
367
368 for (; self->sw_head != self->sw_tail;
369 self->sw_head = aq_ring_next_dx(self, self->sw_head)) {
370 struct aq_ring_buff_s *buff = &self->buff_ring[self->sw_head];
371 struct device *ndev = aq_nic_get_dev(self->aq_nic);
372
373 if (buff->is_mapped) {
374 if (buff->is_sop) {
375 dma_unmap_single(ndev, buff->pa, buff->len,
376 DMA_TO_DEVICE);
377 } else {
378 dma_unmap_page(ndev, buff->pa, buff->len,
379 DMA_TO_DEVICE);
380 }
381 }
382
383 if (buff->is_eop) {
384 if (buff->skb)
385 dev_kfree_skb_any(buff->skb);
386 else if (buff->xdpf)
387 xdp_return_frame(buff->xdpf);
388 }
389 }
390 }
391
aq_rx_checksum(struct aq_ring_s * self,struct aq_ring_buff_s * buff,struct sk_buff * skb)392 static void aq_rx_checksum(struct aq_ring_s *self,
393 struct aq_ring_buff_s *buff,
394 struct sk_buff *skb)
395 {
396 if (!(self->aq_nic->ndev->features & NETIF_F_RXCSUM))
397 return;
398
399 if (unlikely(buff->is_cso_err)) {
400 u64_stats_update_begin(&self->stats.rx.syncp);
401 ++self->stats.rx.errors;
402 u64_stats_update_end(&self->stats.rx.syncp);
403 skb->ip_summed = CHECKSUM_NONE;
404 return;
405 }
406 if (buff->is_ip_cso) {
407 __skb_incr_checksum_unnecessary(skb);
408 } else {
409 skb->ip_summed = CHECKSUM_NONE;
410 }
411
412 if (buff->is_udp_cso || buff->is_tcp_cso)
413 __skb_incr_checksum_unnecessary(skb);
414 }
415
aq_xdp_xmit(struct net_device * dev,int num_frames,struct xdp_frame ** frames,u32 flags)416 int aq_xdp_xmit(struct net_device *dev, int num_frames,
417 struct xdp_frame **frames, u32 flags)
418 {
419 struct aq_nic_s *aq_nic = netdev_priv(dev);
420 unsigned int vec, i, drop = 0;
421 int cpu = smp_processor_id();
422 struct aq_nic_cfg_s *aq_cfg;
423 struct aq_ring_s *ring;
424
425 aq_cfg = aq_nic_get_cfg(aq_nic);
426 vec = cpu % aq_cfg->vecs;
427 ring = aq_nic->aq_ring_tx[AQ_NIC_CFG_TCVEC2RING(aq_cfg, 0, vec)];
428
429 for (i = 0; i < num_frames; i++) {
430 struct xdp_frame *xdpf = frames[i];
431
432 if (aq_nic_xmit_xdpf(aq_nic, ring, xdpf) == NETDEV_TX_BUSY)
433 drop++;
434 }
435
436 return num_frames - drop;
437 }
438
aq_xdp_build_skb(struct xdp_buff * xdp,struct net_device * dev,struct aq_ring_buff_s * buff)439 static struct sk_buff *aq_xdp_build_skb(struct xdp_buff *xdp,
440 struct net_device *dev,
441 struct aq_ring_buff_s *buff)
442 {
443 struct xdp_frame *xdpf;
444 struct sk_buff *skb;
445
446 xdpf = xdp_convert_buff_to_frame(xdp);
447 if (unlikely(!xdpf))
448 return NULL;
449
450 skb = xdp_build_skb_from_frame(xdpf, dev);
451 if (!skb)
452 return NULL;
453
454 aq_get_rxpages_xdp(buff, xdp);
455 return skb;
456 }
457
aq_xdp_run_prog(struct aq_nic_s * aq_nic,struct xdp_buff * xdp,struct aq_ring_s * rx_ring,struct aq_ring_buff_s * buff)458 static struct sk_buff *aq_xdp_run_prog(struct aq_nic_s *aq_nic,
459 struct xdp_buff *xdp,
460 struct aq_ring_s *rx_ring,
461 struct aq_ring_buff_s *buff)
462 {
463 int result = NETDEV_TX_BUSY;
464 struct aq_ring_s *tx_ring;
465 struct xdp_frame *xdpf;
466 struct bpf_prog *prog;
467 u32 act = XDP_ABORTED;
468 struct sk_buff *skb;
469
470 u64_stats_update_begin(&rx_ring->stats.rx.syncp);
471 ++rx_ring->stats.rx.packets;
472 rx_ring->stats.rx.bytes += xdp_get_buff_len(xdp);
473 u64_stats_update_end(&rx_ring->stats.rx.syncp);
474
475 prog = READ_ONCE(rx_ring->xdp_prog);
476 if (!prog)
477 return aq_xdp_build_skb(xdp, aq_nic->ndev, buff);
478
479 prefetchw(xdp->data_hard_start); /* xdp_frame write */
480
481 /* single buffer XDP program, but packet is multi buffer, aborted */
482 if (xdp_buff_has_frags(xdp) && !prog->aux->xdp_has_frags)
483 goto out_aborted;
484
485 act = bpf_prog_run_xdp(prog, xdp);
486 switch (act) {
487 case XDP_PASS:
488 skb = aq_xdp_build_skb(xdp, aq_nic->ndev, buff);
489 if (!skb)
490 goto out_aborted;
491 u64_stats_update_begin(&rx_ring->stats.rx.syncp);
492 ++rx_ring->stats.rx.xdp_pass;
493 u64_stats_update_end(&rx_ring->stats.rx.syncp);
494 return skb;
495 case XDP_TX:
496 xdpf = xdp_convert_buff_to_frame(xdp);
497 if (unlikely(!xdpf))
498 goto out_aborted;
499 tx_ring = aq_nic->aq_ring_tx[rx_ring->idx];
500 result = aq_nic_xmit_xdpf(aq_nic, tx_ring, xdpf);
501 if (result == NETDEV_TX_BUSY)
502 goto out_aborted;
503 u64_stats_update_begin(&rx_ring->stats.rx.syncp);
504 ++rx_ring->stats.rx.xdp_tx;
505 u64_stats_update_end(&rx_ring->stats.rx.syncp);
506 aq_get_rxpages_xdp(buff, xdp);
507 break;
508 case XDP_REDIRECT:
509 if (xdp_do_redirect(aq_nic->ndev, xdp, prog) < 0)
510 goto out_aborted;
511 xdp_do_flush();
512 u64_stats_update_begin(&rx_ring->stats.rx.syncp);
513 ++rx_ring->stats.rx.xdp_redirect;
514 u64_stats_update_end(&rx_ring->stats.rx.syncp);
515 aq_get_rxpages_xdp(buff, xdp);
516 break;
517 default:
518 fallthrough;
519 case XDP_ABORTED:
520 out_aborted:
521 u64_stats_update_begin(&rx_ring->stats.rx.syncp);
522 ++rx_ring->stats.rx.xdp_aborted;
523 u64_stats_update_end(&rx_ring->stats.rx.syncp);
524 trace_xdp_exception(aq_nic->ndev, prog, act);
525 bpf_warn_invalid_xdp_action(aq_nic->ndev, prog, act);
526 break;
527 case XDP_DROP:
528 u64_stats_update_begin(&rx_ring->stats.rx.syncp);
529 ++rx_ring->stats.rx.xdp_drop;
530 u64_stats_update_end(&rx_ring->stats.rx.syncp);
531 break;
532 }
533
534 return ERR_PTR(-result);
535 }
536
aq_add_rx_fragment(struct device * dev,struct aq_ring_s * ring,struct aq_ring_buff_s * buff,struct xdp_buff * xdp)537 static bool aq_add_rx_fragment(struct device *dev,
538 struct aq_ring_s *ring,
539 struct aq_ring_buff_s *buff,
540 struct xdp_buff *xdp)
541 {
542 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
543 struct aq_ring_buff_s *buff_ = buff;
544
545 memset(sinfo, 0, sizeof(*sinfo));
546 do {
547 skb_frag_t *frag;
548
549 if (unlikely(sinfo->nr_frags >= MAX_SKB_FRAGS))
550 return true;
551
552 frag = &sinfo->frags[sinfo->nr_frags++];
553 buff_ = &ring->buff_ring[buff_->next];
554 dma_sync_single_range_for_cpu(dev,
555 buff_->rxdata.daddr,
556 buff_->rxdata.pg_off,
557 buff_->len,
558 DMA_FROM_DEVICE);
559 sinfo->xdp_frags_size += buff_->len;
560 skb_frag_fill_page_desc(frag, buff_->rxdata.page,
561 buff_->rxdata.pg_off,
562 buff_->len);
563
564 buff_->is_cleaned = 1;
565
566 buff->is_ip_cso &= buff_->is_ip_cso;
567 buff->is_udp_cso &= buff_->is_udp_cso;
568 buff->is_tcp_cso &= buff_->is_tcp_cso;
569 buff->is_cso_err |= buff_->is_cso_err;
570
571 if (page_is_pfmemalloc(buff_->rxdata.page))
572 xdp_buff_set_frag_pfmemalloc(xdp);
573
574 } while (!buff_->is_eop);
575
576 xdp_buff_set_frags_flag(xdp);
577
578 return false;
579 }
580
__aq_ring_rx_clean(struct aq_ring_s * self,struct napi_struct * napi,int * work_done,int budget)581 static int __aq_ring_rx_clean(struct aq_ring_s *self, struct napi_struct *napi,
582 int *work_done, int budget)
583 {
584 struct net_device *ndev = aq_nic_get_ndev(self->aq_nic);
585 int err = 0;
586
587 for (; (self->sw_head != self->hw_head) && budget;
588 self->sw_head = aq_ring_next_dx(self, self->sw_head),
589 --budget, ++(*work_done)) {
590 struct aq_ring_buff_s *buff = &self->buff_ring[self->sw_head];
591 bool is_ptp_ring = aq_ptp_ring(self->aq_nic, self);
592 struct aq_ring_buff_s *buff_ = NULL;
593 struct sk_buff *skb = NULL;
594 unsigned int next_ = 0U;
595 unsigned int i = 0U;
596 u16 hdr_len;
597
598 if (buff->is_cleaned)
599 continue;
600
601 if (!buff->is_eop) {
602 unsigned int frag_cnt = 0U;
603
604 /* There will be an extra fragment */
605 if (buff->len > AQ_CFG_RX_HDR_SIZE)
606 frag_cnt++;
607
608 buff_ = buff;
609 do {
610 bool is_rsc_completed = true;
611
612 if (buff_->next >= self->size) {
613 err = -EIO;
614 goto err_exit;
615 }
616
617 frag_cnt++;
618 next_ = buff_->next;
619 buff_ = &self->buff_ring[next_];
620 is_rsc_completed =
621 aq_ring_dx_in_range(self->sw_head,
622 next_,
623 self->hw_head);
624
625 if (unlikely(!is_rsc_completed) ||
626 frag_cnt > MAX_SKB_FRAGS) {
627 err = 0;
628 goto err_exit;
629 }
630
631 buff->is_error |= buff_->is_error;
632 buff->is_cso_err |= buff_->is_cso_err;
633
634 } while (!buff_->is_eop);
635
636 if (buff->is_error ||
637 (buff->is_lro && buff->is_cso_err)) {
638 buff_ = buff;
639 do {
640 if (buff_->next >= self->size) {
641 err = -EIO;
642 goto err_exit;
643 }
644 next_ = buff_->next;
645 buff_ = &self->buff_ring[next_];
646
647 buff_->is_cleaned = true;
648 } while (!buff_->is_eop);
649
650 u64_stats_update_begin(&self->stats.rx.syncp);
651 ++self->stats.rx.errors;
652 u64_stats_update_end(&self->stats.rx.syncp);
653 continue;
654 }
655 }
656
657 if (buff->is_error) {
658 u64_stats_update_begin(&self->stats.rx.syncp);
659 ++self->stats.rx.errors;
660 u64_stats_update_end(&self->stats.rx.syncp);
661 continue;
662 }
663
664 dma_sync_single_range_for_cpu(aq_nic_get_dev(self->aq_nic),
665 buff->rxdata.daddr,
666 buff->rxdata.pg_off,
667 buff->len, DMA_FROM_DEVICE);
668
669 skb = napi_alloc_skb(napi, AQ_CFG_RX_HDR_SIZE);
670 if (unlikely(!skb)) {
671 u64_stats_update_begin(&self->stats.rx.syncp);
672 self->stats.rx.skb_alloc_fails++;
673 u64_stats_update_end(&self->stats.rx.syncp);
674 err = -ENOMEM;
675 goto err_exit;
676 }
677 if (is_ptp_ring)
678 buff->len -=
679 aq_ptp_extract_ts(self->aq_nic, skb_hwtstamps(skb),
680 aq_buf_vaddr(&buff->rxdata),
681 buff->len);
682
683 hdr_len = buff->len;
684 if (hdr_len > AQ_CFG_RX_HDR_SIZE)
685 hdr_len = eth_get_headlen(skb->dev,
686 aq_buf_vaddr(&buff->rxdata),
687 AQ_CFG_RX_HDR_SIZE);
688
689 memcpy(__skb_put(skb, hdr_len), aq_buf_vaddr(&buff->rxdata),
690 ALIGN(hdr_len, sizeof(long)));
691
692 if (buff->len - hdr_len > 0) {
693 skb_add_rx_frag(skb, i++, buff->rxdata.page,
694 buff->rxdata.pg_off + hdr_len,
695 buff->len - hdr_len,
696 self->frame_max);
697 page_ref_inc(buff->rxdata.page);
698 }
699
700 if (!buff->is_eop) {
701 buff_ = buff;
702 do {
703 next_ = buff_->next;
704 buff_ = &self->buff_ring[next_];
705
706 dma_sync_single_range_for_cpu(aq_nic_get_dev(self->aq_nic),
707 buff_->rxdata.daddr,
708 buff_->rxdata.pg_off,
709 buff_->len,
710 DMA_FROM_DEVICE);
711 skb_add_rx_frag(skb, i++,
712 buff_->rxdata.page,
713 buff_->rxdata.pg_off,
714 buff_->len,
715 self->frame_max);
716 page_ref_inc(buff_->rxdata.page);
717 buff_->is_cleaned = 1;
718
719 buff->is_ip_cso &= buff_->is_ip_cso;
720 buff->is_udp_cso &= buff_->is_udp_cso;
721 buff->is_tcp_cso &= buff_->is_tcp_cso;
722 buff->is_cso_err |= buff_->is_cso_err;
723
724 } while (!buff_->is_eop);
725 }
726
727 if (buff->is_vlan)
728 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
729 buff->vlan_rx_tag);
730
731 skb->protocol = eth_type_trans(skb, ndev);
732
733 aq_rx_checksum(self, buff, skb);
734
735 skb_set_hash(skb, buff->rss_hash,
736 buff->is_hash_l4 ? PKT_HASH_TYPE_L4 :
737 PKT_HASH_TYPE_NONE);
738 /* Send all PTP traffic to 0 queue */
739 skb_record_rx_queue(skb,
740 is_ptp_ring ? 0
741 : AQ_NIC_RING2QMAP(self->aq_nic,
742 self->idx));
743
744 u64_stats_update_begin(&self->stats.rx.syncp);
745 ++self->stats.rx.packets;
746 self->stats.rx.bytes += skb->len;
747 u64_stats_update_end(&self->stats.rx.syncp);
748
749 napi_gro_receive(napi, skb);
750 }
751
752 err_exit:
753 return err;
754 }
755
__aq_ring_xdp_clean(struct aq_ring_s * rx_ring,struct napi_struct * napi,int * work_done,int budget)756 static int __aq_ring_xdp_clean(struct aq_ring_s *rx_ring,
757 struct napi_struct *napi, int *work_done,
758 int budget)
759 {
760 int frame_sz = rx_ring->page_offset + rx_ring->frame_max +
761 rx_ring->tail_size;
762 struct aq_nic_s *aq_nic = rx_ring->aq_nic;
763 bool is_rsc_completed = true;
764 struct device *dev;
765 int err = 0;
766
767 dev = aq_nic_get_dev(aq_nic);
768 for (; (rx_ring->sw_head != rx_ring->hw_head) && budget;
769 rx_ring->sw_head = aq_ring_next_dx(rx_ring, rx_ring->sw_head),
770 --budget, ++(*work_done)) {
771 struct aq_ring_buff_s *buff = &rx_ring->buff_ring[rx_ring->sw_head];
772 bool is_ptp_ring = aq_ptp_ring(rx_ring->aq_nic, rx_ring);
773 struct aq_ring_buff_s *buff_ = NULL;
774 u16 ptp_hwtstamp_len = 0;
775 struct skb_shared_hwtstamps shhwtstamps;
776 struct sk_buff *skb = NULL;
777 unsigned int next_ = 0U;
778 struct xdp_buff xdp;
779 void *hard_start;
780
781 if (buff->is_cleaned)
782 continue;
783
784 if (!buff->is_eop) {
785 buff_ = buff;
786 do {
787 if (buff_->next >= rx_ring->size) {
788 err = -EIO;
789 goto err_exit;
790 }
791 next_ = buff_->next;
792 buff_ = &rx_ring->buff_ring[next_];
793 is_rsc_completed =
794 aq_ring_dx_in_range(rx_ring->sw_head,
795 next_,
796 rx_ring->hw_head);
797
798 if (unlikely(!is_rsc_completed))
799 break;
800
801 buff->is_error |= buff_->is_error;
802 buff->is_cso_err |= buff_->is_cso_err;
803 } while (!buff_->is_eop);
804
805 if (!is_rsc_completed) {
806 err = 0;
807 goto err_exit;
808 }
809 if (buff->is_error ||
810 (buff->is_lro && buff->is_cso_err)) {
811 buff_ = buff;
812 do {
813 if (buff_->next >= rx_ring->size) {
814 err = -EIO;
815 goto err_exit;
816 }
817 next_ = buff_->next;
818 buff_ = &rx_ring->buff_ring[next_];
819
820 buff_->is_cleaned = true;
821 } while (!buff_->is_eop);
822
823 u64_stats_update_begin(&rx_ring->stats.rx.syncp);
824 ++rx_ring->stats.rx.errors;
825 u64_stats_update_end(&rx_ring->stats.rx.syncp);
826 continue;
827 }
828 }
829
830 if (buff->is_error) {
831 u64_stats_update_begin(&rx_ring->stats.rx.syncp);
832 ++rx_ring->stats.rx.errors;
833 u64_stats_update_end(&rx_ring->stats.rx.syncp);
834 continue;
835 }
836
837 dma_sync_single_range_for_cpu(dev,
838 buff->rxdata.daddr,
839 buff->rxdata.pg_off,
840 buff->len, DMA_FROM_DEVICE);
841 hard_start = page_address(buff->rxdata.page) +
842 buff->rxdata.pg_off - rx_ring->page_offset;
843
844 if (is_ptp_ring) {
845 ptp_hwtstamp_len = aq_ptp_extract_ts(rx_ring->aq_nic, &shhwtstamps,
846 aq_buf_vaddr(&buff->rxdata),
847 buff->len);
848 buff->len -= ptp_hwtstamp_len;
849 }
850
851 xdp_init_buff(&xdp, frame_sz, &rx_ring->xdp_rxq);
852 xdp_prepare_buff(&xdp, hard_start, rx_ring->page_offset,
853 buff->len, false);
854 if (!buff->is_eop) {
855 if (aq_add_rx_fragment(dev, rx_ring, buff, &xdp)) {
856 u64_stats_update_begin(&rx_ring->stats.rx.syncp);
857 ++rx_ring->stats.rx.packets;
858 rx_ring->stats.rx.bytes += xdp_get_buff_len(&xdp);
859 ++rx_ring->stats.rx.xdp_aborted;
860 u64_stats_update_end(&rx_ring->stats.rx.syncp);
861 continue;
862 }
863 }
864
865 skb = aq_xdp_run_prog(aq_nic, &xdp, rx_ring, buff);
866 if (IS_ERR(skb) || !skb)
867 continue;
868
869 if (ptp_hwtstamp_len > 0)
870 *skb_hwtstamps(skb) = shhwtstamps;
871
872 if (buff->is_vlan)
873 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
874 buff->vlan_rx_tag);
875
876 aq_rx_checksum(rx_ring, buff, skb);
877
878 skb_set_hash(skb, buff->rss_hash,
879 buff->is_hash_l4 ? PKT_HASH_TYPE_L4 :
880 PKT_HASH_TYPE_NONE);
881 /* Send all PTP traffic to 0 queue */
882 skb_record_rx_queue(skb,
883 is_ptp_ring ? 0
884 : AQ_NIC_RING2QMAP(rx_ring->aq_nic,
885 rx_ring->idx));
886
887 napi_gro_receive(napi, skb);
888 }
889
890 err_exit:
891 return err;
892 }
893
aq_ring_rx_clean(struct aq_ring_s * self,struct napi_struct * napi,int * work_done,int budget)894 int aq_ring_rx_clean(struct aq_ring_s *self,
895 struct napi_struct *napi,
896 int *work_done,
897 int budget)
898 {
899 if (static_branch_unlikely(&aq_xdp_locking_key))
900 return __aq_ring_xdp_clean(self, napi, work_done, budget);
901 else
902 return __aq_ring_rx_clean(self, napi, work_done, budget);
903 }
904
aq_ring_hwts_rx_clean(struct aq_ring_s * self,struct aq_nic_s * aq_nic)905 void aq_ring_hwts_rx_clean(struct aq_ring_s *self, struct aq_nic_s *aq_nic)
906 {
907 #if IS_REACHABLE(CONFIG_PTP_1588_CLOCK)
908 while (self->sw_head != self->hw_head) {
909 u64 ns;
910
911 aq_nic->aq_hw_ops->extract_hwts(aq_nic->aq_hw,
912 self->dx_ring +
913 (self->sw_head * self->dx_size),
914 self->dx_size, &ns);
915 aq_ptp_tx_hwtstamp(aq_nic, ns);
916
917 self->sw_head = aq_ring_next_dx(self, self->sw_head);
918 }
919 #endif
920 }
921
aq_ring_rx_fill(struct aq_ring_s * self)922 int aq_ring_rx_fill(struct aq_ring_s *self)
923 {
924 struct aq_ring_buff_s *buff = NULL;
925 int err = 0;
926 int i = 0;
927
928 if (aq_ring_avail_dx(self) < min_t(unsigned int, AQ_CFG_RX_REFILL_THRES,
929 self->size / 2))
930 return err;
931
932 for (i = aq_ring_avail_dx(self); i--;
933 self->sw_tail = aq_ring_next_dx(self, self->sw_tail)) {
934 buff = &self->buff_ring[self->sw_tail];
935
936 buff->flags = 0U;
937 buff->len = self->frame_max;
938
939 err = aq_get_rxpages(self, buff);
940 if (err)
941 goto err_exit;
942
943 buff->pa = aq_buf_daddr(&buff->rxdata);
944 buff = NULL;
945 }
946
947 err_exit:
948 return err;
949 }
950
aq_ring_rx_deinit(struct aq_ring_s * self)951 void aq_ring_rx_deinit(struct aq_ring_s *self)
952 {
953 unsigned int i;
954
955 if (!self || !self->buff_ring)
956 return;
957
958 /* Release every page still owned by the ring.
959 *
960 * Walking [sw_head, sw_tail) is not enough: refill is batched
961 * (aq_ring_rx_fill() waits for AQ_CFG_RX_REFILL_THRES free slots),
962 * so slots that were cleaned but not yet reposted accumulate in the
963 * [sw_tail, sw_head) gap, and they keep their page for reuse. Walk
964 * the whole ring and release whatever is left.
965 */
966 for (i = 0; i < self->size; i++) {
967 struct aq_ring_buff_s *buff = &self->buff_ring[i];
968
969 if (!buff->rxdata.page)
970 continue;
971
972 aq_free_rxpage(&buff->rxdata, aq_nic_get_dev(self->aq_nic));
973 }
974
975 self->sw_head = self->sw_tail;
976 }
977
aq_ring_free(struct aq_ring_s * self)978 void aq_ring_free(struct aq_ring_s *self)
979 {
980 if (!self)
981 return;
982
983 kfree(self->buff_ring);
984 self->buff_ring = NULL;
985
986 if (self->dx_ring) {
987 dma_free_coherent(aq_nic_get_dev(self->aq_nic),
988 self->size * self->dx_size, self->dx_ring,
989 self->dx_ring_pa);
990 self->dx_ring = NULL;
991 }
992 }
993
aq_ring_hwts_rx_free(struct aq_ring_s * self)994 void aq_ring_hwts_rx_free(struct aq_ring_s *self)
995 {
996 if (!self)
997 return;
998
999 if (self->dx_ring) {
1000 dma_free_coherent(aq_nic_get_dev(self->aq_nic),
1001 self->size * self->dx_size + AQ_CFG_RXDS_DEF,
1002 self->dx_ring, self->dx_ring_pa);
1003 self->dx_ring = NULL;
1004 }
1005 }
1006
aq_ring_fill_stats_data(struct aq_ring_s * self,u64 * data)1007 unsigned int aq_ring_fill_stats_data(struct aq_ring_s *self, u64 *data)
1008 {
1009 unsigned int count;
1010 unsigned int start;
1011
1012 if (self->ring_type == ATL_RING_RX) {
1013 /* This data should mimic aq_ethtool_queue_rx_stat_names structure */
1014 do {
1015 count = 0;
1016 start = u64_stats_fetch_begin(&self->stats.rx.syncp);
1017 data[count] = self->stats.rx.packets;
1018 data[++count] = self->stats.rx.jumbo_packets;
1019 data[++count] = self->stats.rx.lro_packets;
1020 data[++count] = self->stats.rx.errors;
1021 data[++count] = self->stats.rx.alloc_fails;
1022 data[++count] = self->stats.rx.skb_alloc_fails;
1023 data[++count] = self->stats.rx.polls;
1024 data[++count] = self->stats.rx.pg_flips;
1025 data[++count] = self->stats.rx.pg_reuses;
1026 data[++count] = self->stats.rx.pg_losts;
1027 data[++count] = self->stats.rx.xdp_aborted;
1028 data[++count] = self->stats.rx.xdp_drop;
1029 data[++count] = self->stats.rx.xdp_pass;
1030 data[++count] = self->stats.rx.xdp_tx;
1031 data[++count] = self->stats.rx.xdp_invalid;
1032 data[++count] = self->stats.rx.xdp_redirect;
1033 } while (u64_stats_fetch_retry(&self->stats.rx.syncp, start));
1034 } else {
1035 /* This data should mimic aq_ethtool_queue_tx_stat_names structure */
1036 do {
1037 count = 0;
1038 start = u64_stats_fetch_begin(&self->stats.tx.syncp);
1039 data[count] = self->stats.tx.packets;
1040 data[++count] = self->stats.tx.queue_restarts;
1041 } while (u64_stats_fetch_retry(&self->stats.tx.syncp, start));
1042 }
1043
1044 return ++count;
1045 }
1046