1 /* Broadcom NetXtreme-C/E network driver.
2 *
3 * Copyright (c) 2016-2017 Broadcom Limited
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation.
8 */
9 #include <linux/kernel.h>
10 #include <linux/errno.h>
11 #include <linux/pci.h>
12 #include <linux/netdevice.h>
13 #include <linux/etherdevice.h>
14 #include <linux/if_vlan.h>
15 #include <linux/bpf.h>
16 #include <linux/bpf_trace.h>
17 #include <linux/filter.h>
18 #include <net/page_pool/helpers.h>
19 #include "bnxt_hsi.h"
20 #include "bnxt.h"
21 #include "bnxt_xdp.h"
22
23 DEFINE_STATIC_KEY_FALSE(bnxt_xdp_locking_key);
24
bnxt_xmit_bd(struct bnxt * bp,struct bnxt_tx_ring_info * txr,dma_addr_t mapping,u32 len,struct xdp_buff * xdp)25 struct bnxt_sw_tx_bd *bnxt_xmit_bd(struct bnxt *bp,
26 struct bnxt_tx_ring_info *txr,
27 dma_addr_t mapping, u32 len,
28 struct xdp_buff *xdp)
29 {
30 struct skb_shared_info *sinfo;
31 struct bnxt_sw_tx_bd *tx_buf;
32 struct tx_bd *txbd;
33 int num_frags = 0;
34 u32 flags;
35 u16 prod;
36 int i;
37
38 if (xdp && xdp_buff_has_frags(xdp)) {
39 sinfo = xdp_get_shared_info_from_buff(xdp);
40 num_frags = sinfo->nr_frags;
41 }
42
43 /* fill up the first buffer */
44 prod = txr->tx_prod;
45 tx_buf = &txr->tx_buf_ring[RING_TX(bp, prod)];
46 tx_buf->nr_frags = num_frags;
47 if (xdp)
48 tx_buf->page = virt_to_head_page(xdp->data);
49
50 txbd = &txr->tx_desc_ring[TX_RING(bp, prod)][TX_IDX(prod)];
51 flags = (len << TX_BD_LEN_SHIFT) | TX_BD_CNT(num_frags + 1) |
52 bnxt_lhint_arr[len >> 9];
53 txbd->tx_bd_len_flags_type = cpu_to_le32(flags);
54 txbd->tx_bd_opaque = SET_TX_OPAQUE(bp, txr, prod, 1 + num_frags);
55 txbd->tx_bd_haddr = cpu_to_le64(mapping);
56
57 /* now let us fill up the frags into the next buffers */
58 for (i = 0; i < num_frags ; i++) {
59 skb_frag_t *frag = &sinfo->frags[i];
60 struct bnxt_sw_tx_bd *frag_tx_buf;
61 dma_addr_t frag_mapping;
62 int frag_len;
63
64 prod = NEXT_TX(prod);
65 WRITE_ONCE(txr->tx_prod, prod);
66
67 /* first fill up the first buffer */
68 frag_tx_buf = &txr->tx_buf_ring[RING_TX(bp, prod)];
69 frag_tx_buf->page = skb_frag_page(frag);
70
71 txbd = &txr->tx_desc_ring[TX_RING(bp, prod)][TX_IDX(prod)];
72
73 frag_len = skb_frag_size(frag);
74 flags = frag_len << TX_BD_LEN_SHIFT;
75 txbd->tx_bd_len_flags_type = cpu_to_le32(flags);
76 frag_mapping = page_pool_get_dma_addr(skb_frag_page(frag)) +
77 skb_frag_off(frag);
78 txbd->tx_bd_haddr = cpu_to_le64(frag_mapping);
79
80 len = frag_len;
81 }
82
83 flags &= ~TX_BD_LEN;
84 txbd->tx_bd_len_flags_type = cpu_to_le32(((len) << TX_BD_LEN_SHIFT) | flags |
85 TX_BD_FLAGS_PACKET_END);
86 /* Sync TX BD */
87 wmb();
88 prod = NEXT_TX(prod);
89 WRITE_ONCE(txr->tx_prod, prod);
90
91 return tx_buf;
92 }
93
__bnxt_xmit_xdp(struct bnxt * bp,struct bnxt_tx_ring_info * txr,dma_addr_t mapping,u32 len,u16 rx_prod,struct xdp_buff * xdp)94 static void __bnxt_xmit_xdp(struct bnxt *bp, struct bnxt_tx_ring_info *txr,
95 dma_addr_t mapping, u32 len, u16 rx_prod,
96 struct xdp_buff *xdp)
97 {
98 struct bnxt_sw_tx_bd *tx_buf;
99
100 tx_buf = bnxt_xmit_bd(bp, txr, mapping, len, xdp);
101 tx_buf->rx_prod = rx_prod;
102 tx_buf->action = XDP_TX;
103
104 }
105
__bnxt_xmit_xdp_redirect(struct bnxt * bp,struct bnxt_tx_ring_info * txr,dma_addr_t mapping,u32 len,struct xdp_frame * xdpf)106 static void __bnxt_xmit_xdp_redirect(struct bnxt *bp,
107 struct bnxt_tx_ring_info *txr,
108 dma_addr_t mapping, u32 len,
109 struct xdp_frame *xdpf)
110 {
111 struct bnxt_sw_tx_bd *tx_buf;
112
113 tx_buf = bnxt_xmit_bd(bp, txr, mapping, len, NULL);
114 tx_buf->action = XDP_REDIRECT;
115 tx_buf->xdpf = xdpf;
116 dma_unmap_addr_set(tx_buf, mapping, mapping);
117 dma_unmap_len_set(tx_buf, len, 0);
118 }
119
bnxt_tx_int_xdp(struct bnxt * bp,struct bnxt_napi * bnapi,int budget)120 void bnxt_tx_int_xdp(struct bnxt *bp, struct bnxt_napi *bnapi, int budget)
121 {
122 struct bnxt_tx_ring_info *txr = bnapi->tx_ring[0];
123 struct bnxt_rx_ring_info *rxr = bnapi->rx_ring;
124 u16 tx_hw_cons = txr->tx_hw_cons;
125 bool rx_doorbell_needed = false;
126 struct bnxt_sw_tx_bd *tx_buf;
127 u16 tx_cons = txr->tx_cons;
128 u16 last_tx_cons = tx_cons;
129 int j, frags;
130
131 if (!budget)
132 return;
133
134 while (RING_TX(bp, tx_cons) != tx_hw_cons) {
135 tx_buf = &txr->tx_buf_ring[RING_TX(bp, tx_cons)];
136
137 if (tx_buf->action == XDP_REDIRECT) {
138 struct pci_dev *pdev = bp->pdev;
139
140 dma_unmap_single(&pdev->dev,
141 dma_unmap_addr(tx_buf, mapping),
142 dma_unmap_len(tx_buf, len),
143 DMA_TO_DEVICE);
144 xdp_return_frame(tx_buf->xdpf);
145 tx_buf->action = 0;
146 tx_buf->xdpf = NULL;
147 } else if (tx_buf->action == XDP_TX) {
148 tx_buf->action = 0;
149 rx_doorbell_needed = true;
150 last_tx_cons = tx_cons;
151
152 frags = tx_buf->nr_frags;
153 for (j = 0; j < frags; j++) {
154 tx_cons = NEXT_TX(tx_cons);
155 tx_buf = &txr->tx_buf_ring[RING_TX(bp, tx_cons)];
156 page_pool_recycle_direct(rxr->page_pool, tx_buf->page);
157 }
158 } else {
159 bnxt_sched_reset_txr(bp, txr, tx_cons);
160 return;
161 }
162 tx_cons = NEXT_TX(tx_cons);
163 }
164
165 bnapi->events &= ~BNXT_TX_CMP_EVENT;
166 WRITE_ONCE(txr->tx_cons, tx_cons);
167 if (rx_doorbell_needed) {
168 tx_buf = &txr->tx_buf_ring[RING_TX(bp, last_tx_cons)];
169 bnxt_db_write(bp, &rxr->rx_db, tx_buf->rx_prod);
170
171 }
172 }
173
bnxt_xdp_attached(struct bnxt * bp,struct bnxt_rx_ring_info * rxr)174 bool bnxt_xdp_attached(struct bnxt *bp, struct bnxt_rx_ring_info *rxr)
175 {
176 struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog);
177
178 return !!xdp_prog;
179 }
180
bnxt_xdp_buff_init(struct bnxt * bp,struct bnxt_rx_ring_info * rxr,u16 cons,u8 * data_ptr,unsigned int len,struct xdp_buff * xdp)181 void bnxt_xdp_buff_init(struct bnxt *bp, struct bnxt_rx_ring_info *rxr,
182 u16 cons, u8 *data_ptr, unsigned int len,
183 struct xdp_buff *xdp)
184 {
185 u32 buflen = BNXT_RX_PAGE_SIZE;
186 struct bnxt_sw_rx_bd *rx_buf;
187 struct pci_dev *pdev;
188 dma_addr_t mapping;
189 u32 offset;
190
191 pdev = bp->pdev;
192 rx_buf = &rxr->rx_buf_ring[cons];
193 offset = bp->rx_offset;
194
195 mapping = rx_buf->mapping - bp->rx_dma_offset;
196 dma_sync_single_for_cpu(&pdev->dev, mapping + offset, len, bp->rx_dir);
197
198 xdp_init_buff(xdp, buflen, &rxr->xdp_rxq);
199 xdp_prepare_buff(xdp, data_ptr - offset, offset, len, true);
200 }
201
bnxt_xdp_buff_frags_free(struct bnxt_rx_ring_info * rxr,struct xdp_buff * xdp)202 void bnxt_xdp_buff_frags_free(struct bnxt_rx_ring_info *rxr,
203 struct xdp_buff *xdp)
204 {
205 struct skb_shared_info *shinfo;
206 int i;
207
208 if (!xdp || !xdp_buff_has_frags(xdp))
209 return;
210 shinfo = xdp_get_shared_info_from_buff(xdp);
211 for (i = 0; i < shinfo->nr_frags; i++) {
212 struct page *page = skb_frag_page(&shinfo->frags[i]);
213
214 page_pool_recycle_direct(rxr->page_pool, page);
215 }
216 shinfo->nr_frags = 0;
217 }
218
219 /* returns the following:
220 * true - packet consumed by XDP and new buffer is allocated.
221 * false - packet should be passed to the stack.
222 */
bnxt_rx_xdp(struct bnxt * bp,struct bnxt_rx_ring_info * rxr,u16 cons,struct xdp_buff * xdp,struct page * page,u8 ** data_ptr,unsigned int * len,u8 * event)223 bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons,
224 struct xdp_buff *xdp, struct page *page, u8 **data_ptr,
225 unsigned int *len, u8 *event)
226 {
227 struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog);
228 struct bnxt_tx_ring_info *txr;
229 struct bnxt_sw_rx_bd *rx_buf;
230 struct pci_dev *pdev;
231 dma_addr_t mapping;
232 u32 tx_needed = 1;
233 void *orig_data;
234 u32 tx_avail;
235 u32 offset;
236 u32 act;
237
238 if (!xdp_prog)
239 return false;
240
241 pdev = bp->pdev;
242 offset = bp->rx_offset;
243
244 txr = rxr->bnapi->tx_ring[0];
245 /* BNXT_RX_PAGE_MODE(bp) when XDP enabled */
246 orig_data = xdp->data;
247
248 act = bpf_prog_run_xdp(xdp_prog, xdp);
249
250 tx_avail = bnxt_tx_avail(bp, txr);
251 /* If the tx ring is not full, we must not update the rx producer yet
252 * because we may still be transmitting on some BDs.
253 */
254 if (tx_avail != bp->tx_ring_size)
255 *event &= ~BNXT_RX_EVENT;
256
257 *len = xdp->data_end - xdp->data;
258 if (orig_data != xdp->data) {
259 offset = xdp->data - xdp->data_hard_start;
260 *data_ptr = xdp->data_hard_start + offset;
261 }
262
263 switch (act) {
264 case XDP_PASS:
265 return false;
266
267 case XDP_TX:
268 rx_buf = &rxr->rx_buf_ring[cons];
269 mapping = rx_buf->mapping - bp->rx_dma_offset;
270 *event &= BNXT_TX_CMP_EVENT;
271
272 if (unlikely(xdp_buff_has_frags(xdp))) {
273 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
274
275 tx_needed += sinfo->nr_frags;
276 *event = BNXT_AGG_EVENT;
277 }
278
279 if (tx_avail < tx_needed) {
280 trace_xdp_exception(bp->dev, xdp_prog, act);
281 bnxt_xdp_buff_frags_free(rxr, xdp);
282 bnxt_reuse_rx_data(rxr, cons, page);
283 return true;
284 }
285
286 dma_sync_single_for_device(&pdev->dev, mapping + offset, *len,
287 bp->rx_dir);
288
289 *event |= BNXT_TX_EVENT;
290 __bnxt_xmit_xdp(bp, txr, mapping + offset, *len,
291 NEXT_RX(rxr->rx_prod), xdp);
292 bnxt_reuse_rx_data(rxr, cons, page);
293 return true;
294 case XDP_REDIRECT:
295 /* if we are calling this here then we know that the
296 * redirect is coming from a frame received by the
297 * bnxt_en driver.
298 */
299
300 /* if we are unable to allocate a new buffer, abort and reuse */
301 if (bnxt_alloc_rx_data(bp, rxr, rxr->rx_prod, GFP_ATOMIC)) {
302 trace_xdp_exception(bp->dev, xdp_prog, act);
303 bnxt_xdp_buff_frags_free(rxr, xdp);
304 bnxt_reuse_rx_data(rxr, cons, page);
305 return true;
306 }
307
308 if (xdp_do_redirect(bp->dev, xdp, xdp_prog)) {
309 trace_xdp_exception(bp->dev, xdp_prog, act);
310 page_pool_recycle_direct(rxr->page_pool, page);
311 return true;
312 }
313
314 *event |= BNXT_REDIRECT_EVENT;
315 break;
316 default:
317 bpf_warn_invalid_xdp_action(bp->dev, xdp_prog, act);
318 fallthrough;
319 case XDP_ABORTED:
320 trace_xdp_exception(bp->dev, xdp_prog, act);
321 fallthrough;
322 case XDP_DROP:
323 bnxt_xdp_buff_frags_free(rxr, xdp);
324 bnxt_reuse_rx_data(rxr, cons, page);
325 break;
326 }
327 return true;
328 }
329
bnxt_xdp_xmit(struct net_device * dev,int num_frames,struct xdp_frame ** frames,u32 flags)330 int bnxt_xdp_xmit(struct net_device *dev, int num_frames,
331 struct xdp_frame **frames, u32 flags)
332 {
333 struct bnxt *bp = netdev_priv(dev);
334 struct bpf_prog *xdp_prog = READ_ONCE(bp->xdp_prog);
335 struct pci_dev *pdev = bp->pdev;
336 struct bnxt_tx_ring_info *txr;
337 dma_addr_t mapping;
338 int nxmit = 0;
339 int ring;
340 int i;
341
342 if (!test_bit(BNXT_STATE_OPEN, &bp->state) ||
343 !bp->tx_nr_rings_xdp ||
344 !xdp_prog)
345 return -EINVAL;
346
347 ring = smp_processor_id() % bp->tx_nr_rings_xdp;
348 txr = &bp->tx_ring[ring];
349
350 if (READ_ONCE(txr->dev_state) == BNXT_DEV_STATE_CLOSING)
351 return -EINVAL;
352
353 if (static_branch_unlikely(&bnxt_xdp_locking_key))
354 spin_lock(&txr->xdp_tx_lock);
355
356 for (i = 0; i < num_frames; i++) {
357 struct xdp_frame *xdp = frames[i];
358
359 if (!bnxt_tx_avail(bp, txr))
360 break;
361
362 mapping = dma_map_single(&pdev->dev, xdp->data, xdp->len,
363 DMA_TO_DEVICE);
364
365 if (dma_mapping_error(&pdev->dev, mapping))
366 break;
367
368 __bnxt_xmit_xdp_redirect(bp, txr, mapping, xdp->len, xdp);
369 nxmit++;
370 }
371
372 if (flags & XDP_XMIT_FLUSH) {
373 /* Sync BD data before updating doorbell */
374 wmb();
375 bnxt_db_write(bp, &txr->tx_db, txr->tx_prod);
376 }
377
378 if (static_branch_unlikely(&bnxt_xdp_locking_key))
379 spin_unlock(&txr->xdp_tx_lock);
380
381 return nxmit;
382 }
383
384 /* Under rtnl_lock */
bnxt_xdp_set(struct bnxt * bp,struct bpf_prog * prog)385 static int bnxt_xdp_set(struct bnxt *bp, struct bpf_prog *prog)
386 {
387 struct net_device *dev = bp->dev;
388 int tx_xdp = 0, tx_cp, rc, tc;
389 struct bpf_prog *old;
390
391 if (prog && !prog->aux->xdp_has_frags &&
392 bp->dev->mtu > BNXT_MAX_PAGE_MODE_MTU) {
393 netdev_warn(dev, "MTU %d larger than %d without XDP frag support.\n",
394 bp->dev->mtu, BNXT_MAX_PAGE_MODE_MTU);
395 return -EOPNOTSUPP;
396 }
397 if (prog && bp->flags & BNXT_FLAG_HDS) {
398 netdev_warn(dev, "XDP is disallowed when HDS is enabled.\n");
399 return -EOPNOTSUPP;
400 }
401 if (!(bp->flags & BNXT_FLAG_SHARED_RINGS)) {
402 netdev_warn(dev, "ethtool rx/tx channels must be combined to support XDP.\n");
403 return -EOPNOTSUPP;
404 }
405 if (prog)
406 tx_xdp = bp->rx_nr_rings;
407
408 tc = bp->num_tc;
409 if (!tc)
410 tc = 1;
411 rc = bnxt_check_rings(bp, bp->tx_nr_rings_per_tc, bp->rx_nr_rings,
412 true, tc, tx_xdp);
413 if (rc) {
414 netdev_warn(dev, "Unable to reserve enough TX rings to support XDP.\n");
415 return rc;
416 }
417 if (netif_running(dev))
418 bnxt_close_nic(bp, true, false);
419
420 old = xchg(&bp->xdp_prog, prog);
421 if (old)
422 bpf_prog_put(old);
423
424 if (prog) {
425 bnxt_set_rx_skb_mode(bp, true);
426 xdp_features_set_redirect_target(dev, true);
427 } else {
428 xdp_features_clear_redirect_target(dev);
429 bnxt_set_rx_skb_mode(bp, false);
430 }
431 bp->tx_nr_rings_xdp = tx_xdp;
432 bp->tx_nr_rings = bp->tx_nr_rings_per_tc * tc + tx_xdp;
433 tx_cp = bnxt_num_tx_to_cp(bp, bp->tx_nr_rings);
434 bp->cp_nr_rings = max_t(int, tx_cp, bp->rx_nr_rings);
435 bnxt_set_tpa_flags(bp);
436 bnxt_set_ring_params(bp);
437
438 if (netif_running(dev))
439 return bnxt_open_nic(bp, true, false);
440
441 return 0;
442 }
443
bnxt_xdp(struct net_device * dev,struct netdev_bpf * xdp)444 int bnxt_xdp(struct net_device *dev, struct netdev_bpf *xdp)
445 {
446 struct bnxt *bp = netdev_priv(dev);
447 int rc;
448
449 switch (xdp->command) {
450 case XDP_SETUP_PROG:
451 rc = bnxt_xdp_set(bp, xdp->prog);
452 break;
453 default:
454 rc = -EINVAL;
455 break;
456 }
457 return rc;
458 }
459
460 struct sk_buff *
bnxt_xdp_build_skb(struct bnxt * bp,struct sk_buff * skb,u8 num_frags,struct page_pool * pool,struct xdp_buff * xdp)461 bnxt_xdp_build_skb(struct bnxt *bp, struct sk_buff *skb, u8 num_frags,
462 struct page_pool *pool, struct xdp_buff *xdp)
463 {
464 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
465
466 if (!skb)
467 return NULL;
468
469 xdp_update_skb_shared_info(skb, num_frags,
470 sinfo->xdp_frags_size,
471 BNXT_RX_PAGE_SIZE * num_frags,
472 xdp_buff_is_frag_pfmemalloc(xdp));
473 return skb;
474 }
475