xref: /linux/drivers/net/ethernet/broadcom/bnxt/bnxt_xdp.c (revision eed4edda910fe34dfae8c6bfbcf57f4593a54295)
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 
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) |
52 		((num_frags + 1) << TX_BD_FLAGS_BD_CNT_SHIFT) |
53 		bnxt_lhint_arr[len >> 9];
54 	txbd->tx_bd_len_flags_type = cpu_to_le32(flags);
55 	txbd->tx_bd_opaque = SET_TX_OPAQUE(bp, txr, prod, 1 + num_frags);
56 	txbd->tx_bd_haddr = cpu_to_le64(mapping);
57 
58 	/* now let us fill up the frags into the next buffers */
59 	for (i = 0; i < num_frags ; i++) {
60 		skb_frag_t *frag = &sinfo->frags[i];
61 		struct bnxt_sw_tx_bd *frag_tx_buf;
62 		dma_addr_t frag_mapping;
63 		int frag_len;
64 
65 		prod = NEXT_TX(prod);
66 		WRITE_ONCE(txr->tx_prod, prod);
67 
68 		/* first fill up the first buffer */
69 		frag_tx_buf = &txr->tx_buf_ring[RING_TX(bp, prod)];
70 		frag_tx_buf->page = skb_frag_page(frag);
71 
72 		txbd = &txr->tx_desc_ring[TX_RING(bp, prod)][TX_IDX(prod)];
73 
74 		frag_len = skb_frag_size(frag);
75 		flags = frag_len << TX_BD_LEN_SHIFT;
76 		txbd->tx_bd_len_flags_type = cpu_to_le32(flags);
77 		frag_mapping = page_pool_get_dma_addr(skb_frag_page(frag)) +
78 			       skb_frag_off(frag);
79 		txbd->tx_bd_haddr = cpu_to_le64(frag_mapping);
80 
81 		len = frag_len;
82 	}
83 
84 	flags &= ~TX_BD_LEN;
85 	txbd->tx_bd_len_flags_type = cpu_to_le32(((len) << TX_BD_LEN_SHIFT) | flags |
86 			TX_BD_FLAGS_PACKET_END);
87 	/* Sync TX BD */
88 	wmb();
89 	prod = NEXT_TX(prod);
90 	WRITE_ONCE(txr->tx_prod, prod);
91 
92 	return tx_buf;
93 }
94 
95 static void __bnxt_xmit_xdp(struct bnxt *bp, struct bnxt_tx_ring_info *txr,
96 			    dma_addr_t mapping, u32 len, u16 rx_prod,
97 			    struct xdp_buff *xdp)
98 {
99 	struct bnxt_sw_tx_bd *tx_buf;
100 
101 	tx_buf = bnxt_xmit_bd(bp, txr, mapping, len, xdp);
102 	tx_buf->rx_prod = rx_prod;
103 	tx_buf->action = XDP_TX;
104 
105 }
106 
107 static void __bnxt_xmit_xdp_redirect(struct bnxt *bp,
108 				     struct bnxt_tx_ring_info *txr,
109 				     dma_addr_t mapping, u32 len,
110 				     struct xdp_frame *xdpf)
111 {
112 	struct bnxt_sw_tx_bd *tx_buf;
113 
114 	tx_buf = bnxt_xmit_bd(bp, txr, mapping, len, NULL);
115 	tx_buf->action = XDP_REDIRECT;
116 	tx_buf->xdpf = xdpf;
117 	dma_unmap_addr_set(tx_buf, mapping, mapping);
118 	dma_unmap_len_set(tx_buf, len, 0);
119 }
120 
121 void bnxt_tx_int_xdp(struct bnxt *bp, struct bnxt_napi *bnapi, int budget)
122 {
123 	struct bnxt_tx_ring_info *txr = bnapi->tx_ring[0];
124 	struct bnxt_rx_ring_info *rxr = bnapi->rx_ring;
125 	u16 tx_hw_cons = txr->tx_hw_cons;
126 	bool rx_doorbell_needed = false;
127 	struct bnxt_sw_tx_bd *tx_buf;
128 	u16 tx_cons = txr->tx_cons;
129 	u16 last_tx_cons = tx_cons;
130 	int j, frags;
131 
132 	if (!budget)
133 		return;
134 
135 	while (RING_TX(bp, tx_cons) != tx_hw_cons) {
136 		tx_buf = &txr->tx_buf_ring[RING_TX(bp, tx_cons)];
137 
138 		if (tx_buf->action == XDP_REDIRECT) {
139 			struct pci_dev *pdev = bp->pdev;
140 
141 			dma_unmap_single(&pdev->dev,
142 					 dma_unmap_addr(tx_buf, mapping),
143 					 dma_unmap_len(tx_buf, len),
144 					 DMA_TO_DEVICE);
145 			xdp_return_frame(tx_buf->xdpf);
146 			tx_buf->action = 0;
147 			tx_buf->xdpf = NULL;
148 		} else if (tx_buf->action == XDP_TX) {
149 			tx_buf->action = 0;
150 			rx_doorbell_needed = true;
151 			last_tx_cons = tx_cons;
152 
153 			frags = tx_buf->nr_frags;
154 			for (j = 0; j < frags; j++) {
155 				tx_cons = NEXT_TX(tx_cons);
156 				tx_buf = &txr->tx_buf_ring[RING_TX(bp, tx_cons)];
157 				page_pool_recycle_direct(rxr->page_pool, tx_buf->page);
158 			}
159 		} else {
160 			bnxt_sched_reset_txr(bp, txr, tx_cons);
161 			return;
162 		}
163 		tx_cons = NEXT_TX(tx_cons);
164 	}
165 
166 	bnapi->events &= ~BNXT_TX_CMP_EVENT;
167 	WRITE_ONCE(txr->tx_cons, tx_cons);
168 	if (rx_doorbell_needed) {
169 		tx_buf = &txr->tx_buf_ring[RING_TX(bp, last_tx_cons)];
170 		bnxt_db_write(bp, &rxr->rx_db, tx_buf->rx_prod);
171 
172 	}
173 }
174 
175 bool bnxt_xdp_attached(struct bnxt *bp, struct bnxt_rx_ring_info *rxr)
176 {
177 	struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog);
178 
179 	return !!xdp_prog;
180 }
181 
182 void bnxt_xdp_buff_init(struct bnxt *bp, struct bnxt_rx_ring_info *rxr,
183 			u16 cons, u8 *data_ptr, unsigned int len,
184 			struct xdp_buff *xdp)
185 {
186 	u32 buflen = BNXT_RX_PAGE_SIZE;
187 	struct bnxt_sw_rx_bd *rx_buf;
188 	struct pci_dev *pdev;
189 	dma_addr_t mapping;
190 	u32 offset;
191 
192 	pdev = bp->pdev;
193 	rx_buf = &rxr->rx_buf_ring[cons];
194 	offset = bp->rx_offset;
195 
196 	mapping = rx_buf->mapping - bp->rx_dma_offset;
197 	dma_sync_single_for_cpu(&pdev->dev, mapping + offset, len, bp->rx_dir);
198 
199 	xdp_init_buff(xdp, buflen, &rxr->xdp_rxq);
200 	xdp_prepare_buff(xdp, data_ptr - offset, offset, len, false);
201 }
202 
203 void bnxt_xdp_buff_frags_free(struct bnxt_rx_ring_info *rxr,
204 			      struct xdp_buff *xdp)
205 {
206 	struct skb_shared_info *shinfo;
207 	int i;
208 
209 	if (!xdp || !xdp_buff_has_frags(xdp))
210 		return;
211 	shinfo = xdp_get_shared_info_from_buff(xdp);
212 	for (i = 0; i < shinfo->nr_frags; i++) {
213 		struct page *page = skb_frag_page(&shinfo->frags[i]);
214 
215 		page_pool_recycle_direct(rxr->page_pool, page);
216 	}
217 	shinfo->nr_frags = 0;
218 }
219 
220 /* returns the following:
221  * true    - packet consumed by XDP and new buffer is allocated.
222  * false   - packet should be passed to the stack.
223  */
224 bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons,
225 		 struct xdp_buff xdp, struct page *page, u8 **data_ptr,
226 		 unsigned int *len, u8 *event)
227 {
228 	struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog);
229 	struct bnxt_tx_ring_info *txr;
230 	struct bnxt_sw_rx_bd *rx_buf;
231 	struct pci_dev *pdev;
232 	dma_addr_t mapping;
233 	u32 tx_needed = 1;
234 	void *orig_data;
235 	u32 tx_avail;
236 	u32 offset;
237 	u32 act;
238 
239 	if (!xdp_prog)
240 		return false;
241 
242 	pdev = bp->pdev;
243 	offset = bp->rx_offset;
244 
245 	txr = rxr->bnapi->tx_ring[0];
246 	/* BNXT_RX_PAGE_MODE(bp) when XDP enabled */
247 	orig_data = xdp.data;
248 
249 	act = bpf_prog_run_xdp(xdp_prog, &xdp);
250 
251 	tx_avail = bnxt_tx_avail(bp, txr);
252 	/* If the tx ring is not full, we must not update the rx producer yet
253 	 * because we may still be transmitting on some BDs.
254 	 */
255 	if (tx_avail != bp->tx_ring_size)
256 		*event &= ~BNXT_RX_EVENT;
257 
258 	*len = xdp.data_end - xdp.data;
259 	if (orig_data != xdp.data) {
260 		offset = xdp.data - xdp.data_hard_start;
261 		*data_ptr = xdp.data_hard_start + offset;
262 	}
263 
264 	switch (act) {
265 	case XDP_PASS:
266 		return false;
267 
268 	case XDP_TX:
269 		rx_buf = &rxr->rx_buf_ring[cons];
270 		mapping = rx_buf->mapping - bp->rx_dma_offset;
271 		*event &= BNXT_TX_CMP_EVENT;
272 
273 		if (unlikely(xdp_buff_has_frags(&xdp))) {
274 			struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(&xdp);
275 
276 			tx_needed += sinfo->nr_frags;
277 			*event = BNXT_AGG_EVENT;
278 		}
279 
280 		if (tx_avail < tx_needed) {
281 			trace_xdp_exception(bp->dev, xdp_prog, act);
282 			bnxt_xdp_buff_frags_free(rxr, &xdp);
283 			bnxt_reuse_rx_data(rxr, cons, page);
284 			return true;
285 		}
286 
287 		dma_sync_single_for_device(&pdev->dev, mapping + offset, *len,
288 					   bp->rx_dir);
289 
290 		*event |= BNXT_TX_EVENT;
291 		__bnxt_xmit_xdp(bp, txr, mapping + offset, *len,
292 				NEXT_RX(rxr->rx_prod), &xdp);
293 		bnxt_reuse_rx_data(rxr, cons, page);
294 		return true;
295 	case XDP_REDIRECT:
296 		/* if we are calling this here then we know that the
297 		 * redirect is coming from a frame received by the
298 		 * bnxt_en driver.
299 		 */
300 		rx_buf = &rxr->rx_buf_ring[cons];
301 		mapping = rx_buf->mapping - bp->rx_dma_offset;
302 		dma_unmap_page_attrs(&pdev->dev, mapping,
303 				     BNXT_RX_PAGE_SIZE, bp->rx_dir,
304 				     DMA_ATTR_WEAK_ORDERING);
305 
306 		/* if we are unable to allocate a new buffer, abort and reuse */
307 		if (bnxt_alloc_rx_data(bp, rxr, rxr->rx_prod, GFP_ATOMIC)) {
308 			trace_xdp_exception(bp->dev, xdp_prog, act);
309 			bnxt_xdp_buff_frags_free(rxr, &xdp);
310 			bnxt_reuse_rx_data(rxr, cons, page);
311 			return true;
312 		}
313 
314 		if (xdp_do_redirect(bp->dev, &xdp, xdp_prog)) {
315 			trace_xdp_exception(bp->dev, xdp_prog, act);
316 			page_pool_recycle_direct(rxr->page_pool, page);
317 			return true;
318 		}
319 
320 		*event |= BNXT_REDIRECT_EVENT;
321 		break;
322 	default:
323 		bpf_warn_invalid_xdp_action(bp->dev, xdp_prog, act);
324 		fallthrough;
325 	case XDP_ABORTED:
326 		trace_xdp_exception(bp->dev, xdp_prog, act);
327 		fallthrough;
328 	case XDP_DROP:
329 		bnxt_xdp_buff_frags_free(rxr, &xdp);
330 		bnxt_reuse_rx_data(rxr, cons, page);
331 		break;
332 	}
333 	return true;
334 }
335 
336 int bnxt_xdp_xmit(struct net_device *dev, int num_frames,
337 		  struct xdp_frame **frames, u32 flags)
338 {
339 	struct bnxt *bp = netdev_priv(dev);
340 	struct bpf_prog *xdp_prog = READ_ONCE(bp->xdp_prog);
341 	struct pci_dev *pdev = bp->pdev;
342 	struct bnxt_tx_ring_info *txr;
343 	dma_addr_t mapping;
344 	int nxmit = 0;
345 	int ring;
346 	int i;
347 
348 	if (!test_bit(BNXT_STATE_OPEN, &bp->state) ||
349 	    !bp->tx_nr_rings_xdp ||
350 	    !xdp_prog)
351 		return -EINVAL;
352 
353 	ring = smp_processor_id() % bp->tx_nr_rings_xdp;
354 	txr = &bp->tx_ring[ring];
355 
356 	if (READ_ONCE(txr->dev_state) == BNXT_DEV_STATE_CLOSING)
357 		return -EINVAL;
358 
359 	if (static_branch_unlikely(&bnxt_xdp_locking_key))
360 		spin_lock(&txr->xdp_tx_lock);
361 
362 	for (i = 0; i < num_frames; i++) {
363 		struct xdp_frame *xdp = frames[i];
364 
365 		if (!bnxt_tx_avail(bp, txr))
366 			break;
367 
368 		mapping = dma_map_single(&pdev->dev, xdp->data, xdp->len,
369 					 DMA_TO_DEVICE);
370 
371 		if (dma_mapping_error(&pdev->dev, mapping))
372 			break;
373 
374 		__bnxt_xmit_xdp_redirect(bp, txr, mapping, xdp->len, xdp);
375 		nxmit++;
376 	}
377 
378 	if (flags & XDP_XMIT_FLUSH) {
379 		/* Sync BD data before updating doorbell */
380 		wmb();
381 		bnxt_db_write(bp, &txr->tx_db, txr->tx_prod);
382 	}
383 
384 	if (static_branch_unlikely(&bnxt_xdp_locking_key))
385 		spin_unlock(&txr->xdp_tx_lock);
386 
387 	return nxmit;
388 }
389 
390 /* Under rtnl_lock */
391 static int bnxt_xdp_set(struct bnxt *bp, struct bpf_prog *prog)
392 {
393 	struct net_device *dev = bp->dev;
394 	int tx_xdp = 0, tx_cp, rc, tc;
395 	struct bpf_prog *old;
396 
397 	if (prog && !prog->aux->xdp_has_frags &&
398 	    bp->dev->mtu > BNXT_MAX_PAGE_MODE_MTU) {
399 		netdev_warn(dev, "MTU %d larger than %d without XDP frag support.\n",
400 			    bp->dev->mtu, BNXT_MAX_PAGE_MODE_MTU);
401 		return -EOPNOTSUPP;
402 	}
403 	if (!(bp->flags & BNXT_FLAG_SHARED_RINGS)) {
404 		netdev_warn(dev, "ethtool rx/tx channels must be combined to support XDP.\n");
405 		return -EOPNOTSUPP;
406 	}
407 	if (prog)
408 		tx_xdp = bp->rx_nr_rings;
409 
410 	tc = bp->num_tc;
411 	if (!tc)
412 		tc = 1;
413 	rc = bnxt_check_rings(bp, bp->tx_nr_rings_per_tc, bp->rx_nr_rings,
414 			      true, tc, tx_xdp);
415 	if (rc) {
416 		netdev_warn(dev, "Unable to reserve enough TX rings to support XDP.\n");
417 		return rc;
418 	}
419 	if (netif_running(dev))
420 		bnxt_close_nic(bp, true, false);
421 
422 	old = xchg(&bp->xdp_prog, prog);
423 	if (old)
424 		bpf_prog_put(old);
425 
426 	if (prog) {
427 		bnxt_set_rx_skb_mode(bp, true);
428 		xdp_features_set_redirect_target(dev, true);
429 	} else {
430 		int rx, tx;
431 
432 		xdp_features_clear_redirect_target(dev);
433 		bnxt_set_rx_skb_mode(bp, false);
434 		bnxt_get_max_rings(bp, &rx, &tx, true);
435 		if (rx > 1) {
436 			bp->flags &= ~BNXT_FLAG_NO_AGG_RINGS;
437 			bp->dev->hw_features |= NETIF_F_LRO;
438 		}
439 	}
440 	bp->tx_nr_rings_xdp = tx_xdp;
441 	bp->tx_nr_rings = bp->tx_nr_rings_per_tc * tc + tx_xdp;
442 	tx_cp = bnxt_num_tx_to_cp(bp, bp->tx_nr_rings);
443 	bp->cp_nr_rings = max_t(int, tx_cp, bp->rx_nr_rings);
444 	bnxt_set_tpa_flags(bp);
445 	bnxt_set_ring_params(bp);
446 
447 	if (netif_running(dev))
448 		return bnxt_open_nic(bp, true, false);
449 
450 	return 0;
451 }
452 
453 int bnxt_xdp(struct net_device *dev, struct netdev_bpf *xdp)
454 {
455 	struct bnxt *bp = netdev_priv(dev);
456 	int rc;
457 
458 	switch (xdp->command) {
459 	case XDP_SETUP_PROG:
460 		rc = bnxt_xdp_set(bp, xdp->prog);
461 		break;
462 	default:
463 		rc = -EINVAL;
464 		break;
465 	}
466 	return rc;
467 }
468 
469 struct sk_buff *
470 bnxt_xdp_build_skb(struct bnxt *bp, struct sk_buff *skb, u8 num_frags,
471 		   struct page_pool *pool, struct xdp_buff *xdp,
472 		   struct rx_cmp_ext *rxcmp1)
473 {
474 	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
475 
476 	if (!skb)
477 		return NULL;
478 	skb_checksum_none_assert(skb);
479 	if (RX_CMP_L4_CS_OK(rxcmp1)) {
480 		if (bp->dev->features & NETIF_F_RXCSUM) {
481 			skb->ip_summed = CHECKSUM_UNNECESSARY;
482 			skb->csum_level = RX_CMP_ENCAP(rxcmp1);
483 		}
484 	}
485 	xdp_update_skb_shared_info(skb, num_frags,
486 				   sinfo->xdp_frags_size,
487 				   BNXT_RX_PAGE_SIZE * sinfo->nr_frags,
488 				   xdp_buff_is_frag_pfmemalloc(xdp));
489 	return skb;
490 }
491