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.h>
19 #include "bnxt_hsi.h"
20 #include "bnxt.h"
21 #include "bnxt_xdp.h"
22 
23 struct bnxt_sw_tx_bd *bnxt_xmit_bd(struct bnxt *bp,
24 				   struct bnxt_tx_ring_info *txr,
25 				   dma_addr_t mapping, u32 len)
26 {
27 	struct bnxt_sw_tx_bd *tx_buf;
28 	struct tx_bd *txbd;
29 	u32 flags;
30 	u16 prod;
31 
32 	prod = txr->tx_prod;
33 	tx_buf = &txr->tx_buf_ring[prod];
34 
35 	txbd = &txr->tx_desc_ring[TX_RING(prod)][TX_IDX(prod)];
36 	flags = (len << TX_BD_LEN_SHIFT) | (1 << TX_BD_FLAGS_BD_CNT_SHIFT) |
37 		TX_BD_FLAGS_PACKET_END | bnxt_lhint_arr[len >> 9];
38 	txbd->tx_bd_len_flags_type = cpu_to_le32(flags);
39 	txbd->tx_bd_opaque = prod;
40 	txbd->tx_bd_haddr = cpu_to_le64(mapping);
41 
42 	prod = NEXT_TX(prod);
43 	txr->tx_prod = prod;
44 	return tx_buf;
45 }
46 
47 static void __bnxt_xmit_xdp(struct bnxt *bp, struct bnxt_tx_ring_info *txr,
48 			    dma_addr_t mapping, u32 len, u16 rx_prod)
49 {
50 	struct bnxt_sw_tx_bd *tx_buf;
51 
52 	tx_buf = bnxt_xmit_bd(bp, txr, mapping, len);
53 	tx_buf->rx_prod = rx_prod;
54 	tx_buf->action = XDP_TX;
55 }
56 
57 static void __bnxt_xmit_xdp_redirect(struct bnxt *bp,
58 				     struct bnxt_tx_ring_info *txr,
59 				     dma_addr_t mapping, u32 len,
60 				     struct xdp_frame *xdpf)
61 {
62 	struct bnxt_sw_tx_bd *tx_buf;
63 
64 	tx_buf = bnxt_xmit_bd(bp, txr, mapping, len);
65 	tx_buf->action = XDP_REDIRECT;
66 	tx_buf->xdpf = xdpf;
67 	dma_unmap_addr_set(tx_buf, mapping, mapping);
68 	dma_unmap_len_set(tx_buf, len, 0);
69 }
70 
71 void bnxt_tx_int_xdp(struct bnxt *bp, struct bnxt_napi *bnapi, int nr_pkts)
72 {
73 	struct bnxt_tx_ring_info *txr = bnapi->tx_ring;
74 	struct bnxt_rx_ring_info *rxr = bnapi->rx_ring;
75 	bool rx_doorbell_needed = false;
76 	struct bnxt_sw_tx_bd *tx_buf;
77 	u16 tx_cons = txr->tx_cons;
78 	u16 last_tx_cons = tx_cons;
79 	int i;
80 
81 	for (i = 0; i < nr_pkts; i++) {
82 		tx_buf = &txr->tx_buf_ring[tx_cons];
83 
84 		if (tx_buf->action == XDP_REDIRECT) {
85 			struct pci_dev *pdev = bp->pdev;
86 
87 			dma_unmap_single(&pdev->dev,
88 					 dma_unmap_addr(tx_buf, mapping),
89 					 dma_unmap_len(tx_buf, len),
90 					 PCI_DMA_TODEVICE);
91 			xdp_return_frame(tx_buf->xdpf);
92 			tx_buf->action = 0;
93 			tx_buf->xdpf = NULL;
94 		} else if (tx_buf->action == XDP_TX) {
95 			rx_doorbell_needed = true;
96 			last_tx_cons = tx_cons;
97 		}
98 		tx_cons = NEXT_TX(tx_cons);
99 	}
100 	txr->tx_cons = tx_cons;
101 	if (rx_doorbell_needed) {
102 		tx_buf = &txr->tx_buf_ring[last_tx_cons];
103 		bnxt_db_write(bp, &rxr->rx_db, tx_buf->rx_prod);
104 	}
105 }
106 
107 /* returns the following:
108  * true    - packet consumed by XDP and new buffer is allocated.
109  * false   - packet should be passed to the stack.
110  */
111 bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons,
112 		 struct page *page, u8 **data_ptr, unsigned int *len, u8 *event)
113 {
114 	struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog);
115 	struct bnxt_tx_ring_info *txr;
116 	struct bnxt_sw_rx_bd *rx_buf;
117 	struct pci_dev *pdev;
118 	struct xdp_buff xdp;
119 	dma_addr_t mapping;
120 	void *orig_data;
121 	u32 tx_avail;
122 	u32 offset;
123 	u32 act;
124 
125 	if (!xdp_prog)
126 		return false;
127 
128 	pdev = bp->pdev;
129 	rx_buf = &rxr->rx_buf_ring[cons];
130 	offset = bp->rx_offset;
131 
132 	mapping = rx_buf->mapping - bp->rx_dma_offset;
133 	dma_sync_single_for_cpu(&pdev->dev, mapping + offset, *len, bp->rx_dir);
134 
135 	txr = rxr->bnapi->tx_ring;
136 	xdp.data_hard_start = *data_ptr - offset;
137 	xdp.data = *data_ptr;
138 	xdp_set_data_meta_invalid(&xdp);
139 	xdp.data_end = *data_ptr + *len;
140 	xdp.rxq = &rxr->xdp_rxq;
141 	xdp.frame_sz = PAGE_SIZE; /* BNXT_RX_PAGE_MODE(bp) when XDP enabled */
142 	orig_data = xdp.data;
143 
144 	rcu_read_lock();
145 	act = bpf_prog_run_xdp(xdp_prog, &xdp);
146 	rcu_read_unlock();
147 
148 	tx_avail = bnxt_tx_avail(bp, txr);
149 	/* If the tx ring is not full, we must not update the rx producer yet
150 	 * because we may still be transmitting on some BDs.
151 	 */
152 	if (tx_avail != bp->tx_ring_size)
153 		*event &= ~BNXT_RX_EVENT;
154 
155 	*len = xdp.data_end - xdp.data;
156 	if (orig_data != xdp.data) {
157 		offset = xdp.data - xdp.data_hard_start;
158 		*data_ptr = xdp.data_hard_start + offset;
159 	}
160 	switch (act) {
161 	case XDP_PASS:
162 		return false;
163 
164 	case XDP_TX:
165 		if (tx_avail < 1) {
166 			trace_xdp_exception(bp->dev, xdp_prog, act);
167 			bnxt_reuse_rx_data(rxr, cons, page);
168 			return true;
169 		}
170 
171 		*event = BNXT_TX_EVENT;
172 		dma_sync_single_for_device(&pdev->dev, mapping + offset, *len,
173 					   bp->rx_dir);
174 		__bnxt_xmit_xdp(bp, txr, mapping + offset, *len,
175 				NEXT_RX(rxr->rx_prod));
176 		bnxt_reuse_rx_data(rxr, cons, page);
177 		return true;
178 	case XDP_REDIRECT:
179 		/* if we are calling this here then we know that the
180 		 * redirect is coming from a frame received by the
181 		 * bnxt_en driver.
182 		 */
183 		dma_unmap_page_attrs(&pdev->dev, mapping,
184 				     PAGE_SIZE, bp->rx_dir,
185 				     DMA_ATTR_WEAK_ORDERING);
186 
187 		/* if we are unable to allocate a new buffer, abort and reuse */
188 		if (bnxt_alloc_rx_data(bp, rxr, rxr->rx_prod, GFP_ATOMIC)) {
189 			trace_xdp_exception(bp->dev, xdp_prog, act);
190 			bnxt_reuse_rx_data(rxr, cons, page);
191 			return true;
192 		}
193 
194 		if (xdp_do_redirect(bp->dev, &xdp, xdp_prog)) {
195 			trace_xdp_exception(bp->dev, xdp_prog, act);
196 			page_pool_recycle_direct(rxr->page_pool, page);
197 			return true;
198 		}
199 
200 		*event |= BNXT_REDIRECT_EVENT;
201 		break;
202 	default:
203 		bpf_warn_invalid_xdp_action(act);
204 		/* Fall thru */
205 	case XDP_ABORTED:
206 		trace_xdp_exception(bp->dev, xdp_prog, act);
207 		/* Fall thru */
208 	case XDP_DROP:
209 		bnxt_reuse_rx_data(rxr, cons, page);
210 		break;
211 	}
212 	return true;
213 }
214 
215 int bnxt_xdp_xmit(struct net_device *dev, int num_frames,
216 		  struct xdp_frame **frames, u32 flags)
217 {
218 	struct bnxt *bp = netdev_priv(dev);
219 	struct bpf_prog *xdp_prog = READ_ONCE(bp->xdp_prog);
220 	struct pci_dev *pdev = bp->pdev;
221 	struct bnxt_tx_ring_info *txr;
222 	dma_addr_t mapping;
223 	int drops = 0;
224 	int ring;
225 	int i;
226 
227 	if (!test_bit(BNXT_STATE_OPEN, &bp->state) ||
228 	    !bp->tx_nr_rings_xdp ||
229 	    !xdp_prog)
230 		return -EINVAL;
231 
232 	ring = smp_processor_id() % bp->tx_nr_rings_xdp;
233 	txr = &bp->tx_ring[ring];
234 
235 	for (i = 0; i < num_frames; i++) {
236 		struct xdp_frame *xdp = frames[i];
237 
238 		if (!txr || !bnxt_tx_avail(bp, txr) ||
239 		    !(bp->bnapi[ring]->flags & BNXT_NAPI_FLAG_XDP)) {
240 			xdp_return_frame_rx_napi(xdp);
241 			drops++;
242 			continue;
243 		}
244 
245 		mapping = dma_map_single(&pdev->dev, xdp->data, xdp->len,
246 					 DMA_TO_DEVICE);
247 
248 		if (dma_mapping_error(&pdev->dev, mapping)) {
249 			xdp_return_frame_rx_napi(xdp);
250 			drops++;
251 			continue;
252 		}
253 		__bnxt_xmit_xdp_redirect(bp, txr, mapping, xdp->len, xdp);
254 	}
255 
256 	if (flags & XDP_XMIT_FLUSH) {
257 		/* Sync BD data before updating doorbell */
258 		wmb();
259 		bnxt_db_write(bp, &txr->tx_db, txr->tx_prod);
260 	}
261 
262 	return num_frames - drops;
263 }
264 
265 /* Under rtnl_lock */
266 static int bnxt_xdp_set(struct bnxt *bp, struct bpf_prog *prog)
267 {
268 	struct net_device *dev = bp->dev;
269 	int tx_xdp = 0, rc, tc;
270 	struct bpf_prog *old;
271 
272 	if (prog && bp->dev->mtu > BNXT_MAX_PAGE_MODE_MTU) {
273 		netdev_warn(dev, "MTU %d larger than largest XDP supported MTU %d.\n",
274 			    bp->dev->mtu, BNXT_MAX_PAGE_MODE_MTU);
275 		return -EOPNOTSUPP;
276 	}
277 	if (!(bp->flags & BNXT_FLAG_SHARED_RINGS)) {
278 		netdev_warn(dev, "ethtool rx/tx channels must be combined to support XDP.\n");
279 		return -EOPNOTSUPP;
280 	}
281 	if (prog)
282 		tx_xdp = bp->rx_nr_rings;
283 
284 	tc = netdev_get_num_tc(dev);
285 	if (!tc)
286 		tc = 1;
287 	rc = bnxt_check_rings(bp, bp->tx_nr_rings_per_tc, bp->rx_nr_rings,
288 			      true, tc, tx_xdp);
289 	if (rc) {
290 		netdev_warn(dev, "Unable to reserve enough TX rings to support XDP.\n");
291 		return rc;
292 	}
293 	if (netif_running(dev))
294 		bnxt_close_nic(bp, true, false);
295 
296 	old = xchg(&bp->xdp_prog, prog);
297 	if (old)
298 		bpf_prog_put(old);
299 
300 	if (prog) {
301 		bnxt_set_rx_skb_mode(bp, true);
302 	} else {
303 		int rx, tx;
304 
305 		bnxt_set_rx_skb_mode(bp, false);
306 		bnxt_get_max_rings(bp, &rx, &tx, true);
307 		if (rx > 1) {
308 			bp->flags &= ~BNXT_FLAG_NO_AGG_RINGS;
309 			bp->dev->hw_features |= NETIF_F_LRO;
310 		}
311 	}
312 	bp->tx_nr_rings_xdp = tx_xdp;
313 	bp->tx_nr_rings = bp->tx_nr_rings_per_tc * tc + tx_xdp;
314 	bp->cp_nr_rings = max_t(int, bp->tx_nr_rings, bp->rx_nr_rings);
315 	bnxt_set_tpa_flags(bp);
316 	bnxt_set_ring_params(bp);
317 
318 	if (netif_running(dev))
319 		return bnxt_open_nic(bp, true, false);
320 
321 	return 0;
322 }
323 
324 int bnxt_xdp(struct net_device *dev, struct netdev_bpf *xdp)
325 {
326 	struct bnxt *bp = netdev_priv(dev);
327 	int rc;
328 
329 	switch (xdp->command) {
330 	case XDP_SETUP_PROG:
331 		rc = bnxt_xdp_set(bp, xdp->prog);
332 		break;
333 	default:
334 		rc = -EINVAL;
335 		break;
336 	}
337 	return rc;
338 }
339