1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
2 /* Copyright (C) 2018 Netronome Systems, Inc */
3 /* Copyright (C) 2021 Corigine, Inc */
4 
5 #include <linux/bpf_trace.h>
6 #include <linux/netdevice.h>
7 
8 #include "../nfp_app.h"
9 #include "../nfp_net.h"
10 #include "../nfp_net_dp.h"
11 #include "../nfp_net_xsk.h"
12 #include "nfd3.h"
13 
14 static bool
15 nfp_nfd3_xsk_tx_xdp(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
16 		    struct nfp_net_rx_ring *rx_ring,
17 		    struct nfp_net_tx_ring *tx_ring,
18 		    struct nfp_net_xsk_rx_buf *xrxbuf, unsigned int pkt_len,
19 		    int pkt_off)
20 {
21 	struct xsk_buff_pool *pool = r_vec->xsk_pool;
22 	struct nfp_nfd3_tx_buf *txbuf;
23 	struct nfp_nfd3_tx_desc *txd;
24 	unsigned int wr_idx;
25 
26 	if (nfp_net_tx_space(tx_ring) < 1)
27 		return false;
28 
29 	xsk_buff_raw_dma_sync_for_device(pool, xrxbuf->dma_addr + pkt_off,
30 					 pkt_len);
31 
32 	wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
33 
34 	txbuf = &tx_ring->txbufs[wr_idx];
35 	txbuf->xdp = xrxbuf->xdp;
36 	txbuf->real_len = pkt_len;
37 	txbuf->is_xsk_tx = true;
38 
39 	/* Build TX descriptor */
40 	txd = &tx_ring->txds[wr_idx];
41 	txd->offset_eop = NFD3_DESC_TX_EOP;
42 	txd->dma_len = cpu_to_le16(pkt_len);
43 	nfp_desc_set_dma_addr_40b(txd, xrxbuf->dma_addr + pkt_off);
44 	txd->data_len = cpu_to_le16(pkt_len);
45 
46 	txd->flags = 0;
47 	txd->mss = 0;
48 	txd->lso_hdrlen = 0;
49 
50 	tx_ring->wr_ptr_add++;
51 	tx_ring->wr_p++;
52 
53 	return true;
54 }
55 
56 static void nfp_nfd3_xsk_rx_skb(struct nfp_net_rx_ring *rx_ring,
57 				const struct nfp_net_rx_desc *rxd,
58 				struct nfp_net_xsk_rx_buf *xrxbuf,
59 				const struct nfp_meta_parsed *meta,
60 				unsigned int pkt_len,
61 				bool meta_xdp,
62 				unsigned int *skbs_polled)
63 {
64 	struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
65 	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
66 	struct net_device *netdev;
67 	struct sk_buff *skb;
68 
69 	if (likely(!meta->portid)) {
70 		netdev = dp->netdev;
71 	} else {
72 		struct nfp_net *nn = netdev_priv(dp->netdev);
73 
74 		netdev = nfp_app_dev_get(nn->app, meta->portid, NULL);
75 		if (unlikely(!netdev)) {
76 			nfp_net_xsk_rx_drop(r_vec, xrxbuf);
77 			return;
78 		}
79 		nfp_repr_inc_rx_stats(netdev, pkt_len);
80 	}
81 
82 	skb = napi_alloc_skb(&r_vec->napi, pkt_len);
83 	if (!skb) {
84 		nfp_net_xsk_rx_drop(r_vec, xrxbuf);
85 		return;
86 	}
87 	skb_put_data(skb, xrxbuf->xdp->data, pkt_len);
88 
89 	skb->mark = meta->mark;
90 	skb_set_hash(skb, meta->hash, meta->hash_type);
91 
92 	skb_record_rx_queue(skb, rx_ring->idx);
93 	skb->protocol = eth_type_trans(skb, netdev);
94 
95 	nfp_nfd3_rx_csum(dp, r_vec, rxd, meta, skb);
96 
97 	if (unlikely(!nfp_net_vlan_strip(skb, rxd, meta))) {
98 		dev_kfree_skb_any(skb);
99 		nfp_net_xsk_rx_drop(r_vec, xrxbuf);
100 		return;
101 	}
102 
103 	if (meta_xdp)
104 		skb_metadata_set(skb,
105 				 xrxbuf->xdp->data - xrxbuf->xdp->data_meta);
106 
107 	napi_gro_receive(&rx_ring->r_vec->napi, skb);
108 
109 	nfp_net_xsk_rx_free(xrxbuf);
110 
111 	(*skbs_polled)++;
112 }
113 
114 static unsigned int
115 nfp_nfd3_xsk_rx(struct nfp_net_rx_ring *rx_ring, int budget,
116 		unsigned int *skbs_polled)
117 {
118 	struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
119 	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
120 	struct nfp_net_tx_ring *tx_ring;
121 	struct bpf_prog *xdp_prog;
122 	bool xdp_redir = false;
123 	int pkts_polled = 0;
124 
125 	xdp_prog = READ_ONCE(dp->xdp_prog);
126 	tx_ring = r_vec->xdp_ring;
127 
128 	while (pkts_polled < budget) {
129 		unsigned int meta_len, data_len, pkt_len, pkt_off;
130 		struct nfp_net_xsk_rx_buf *xrxbuf;
131 		struct nfp_net_rx_desc *rxd;
132 		struct nfp_meta_parsed meta;
133 		int idx, act;
134 
135 		idx = D_IDX(rx_ring, rx_ring->rd_p);
136 
137 		rxd = &rx_ring->rxds[idx];
138 		if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
139 			break;
140 
141 		rx_ring->rd_p++;
142 		pkts_polled++;
143 
144 		xrxbuf = &rx_ring->xsk_rxbufs[idx];
145 
146 		/* If starved of buffers "drop" it and scream. */
147 		if (rx_ring->rd_p >= rx_ring->wr_p) {
148 			nn_dp_warn(dp, "Starved of RX buffers\n");
149 			nfp_net_xsk_rx_drop(r_vec, xrxbuf);
150 			break;
151 		}
152 
153 		/* Memory barrier to ensure that we won't do other reads
154 		 * before the DD bit.
155 		 */
156 		dma_rmb();
157 
158 		memset(&meta, 0, sizeof(meta));
159 
160 		/* Only supporting AF_XDP with dynamic metadata so buffer layout
161 		 * is always:
162 		 *
163 		 *  ---------------------------------------------------------
164 		 * |  off | metadata  |             packet           | XXXX  |
165 		 *  ---------------------------------------------------------
166 		 */
167 		meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
168 		data_len = le16_to_cpu(rxd->rxd.data_len);
169 		pkt_len = data_len - meta_len;
170 
171 		if (unlikely(meta_len > NFP_NET_MAX_PREPEND)) {
172 			nn_dp_warn(dp, "Oversized RX packet metadata %u\n",
173 				   meta_len);
174 			nfp_net_xsk_rx_drop(r_vec, xrxbuf);
175 			continue;
176 		}
177 
178 		/* Stats update. */
179 		u64_stats_update_begin(&r_vec->rx_sync);
180 		r_vec->rx_pkts++;
181 		r_vec->rx_bytes += pkt_len;
182 		u64_stats_update_end(&r_vec->rx_sync);
183 
184 		xrxbuf->xdp->data += meta_len;
185 		xrxbuf->xdp->data_end = xrxbuf->xdp->data + pkt_len;
186 		xdp_set_data_meta_invalid(xrxbuf->xdp);
187 		xsk_buff_dma_sync_for_cpu(xrxbuf->xdp, r_vec->xsk_pool);
188 		net_prefetch(xrxbuf->xdp->data);
189 
190 		if (meta_len) {
191 			if (unlikely(nfp_nfd3_parse_meta(dp->netdev, &meta,
192 							 xrxbuf->xdp->data -
193 							 meta_len,
194 							 xrxbuf->xdp->data,
195 							 pkt_len, meta_len))) {
196 				nn_dp_warn(dp, "Invalid RX packet metadata\n");
197 				nfp_net_xsk_rx_drop(r_vec, xrxbuf);
198 				continue;
199 			}
200 
201 			if (unlikely(meta.portid)) {
202 				struct nfp_net *nn = netdev_priv(dp->netdev);
203 
204 				if (meta.portid != NFP_META_PORT_ID_CTRL) {
205 					nfp_nfd3_xsk_rx_skb(rx_ring, rxd,
206 							    xrxbuf, &meta,
207 							    pkt_len, false,
208 							    skbs_polled);
209 					continue;
210 				}
211 
212 				nfp_app_ctrl_rx_raw(nn->app, xrxbuf->xdp->data,
213 						    pkt_len);
214 				nfp_net_xsk_rx_free(xrxbuf);
215 				continue;
216 			}
217 		}
218 
219 		act = bpf_prog_run_xdp(xdp_prog, xrxbuf->xdp);
220 
221 		pkt_len = xrxbuf->xdp->data_end - xrxbuf->xdp->data;
222 		pkt_off = xrxbuf->xdp->data - xrxbuf->xdp->data_hard_start;
223 
224 		switch (act) {
225 		case XDP_PASS:
226 			nfp_nfd3_xsk_rx_skb(rx_ring, rxd, xrxbuf, &meta, pkt_len,
227 					    true, skbs_polled);
228 			break;
229 		case XDP_TX:
230 			if (!nfp_nfd3_xsk_tx_xdp(dp, r_vec, rx_ring, tx_ring,
231 						 xrxbuf, pkt_len, pkt_off))
232 				nfp_net_xsk_rx_drop(r_vec, xrxbuf);
233 			else
234 				nfp_net_xsk_rx_unstash(xrxbuf);
235 			break;
236 		case XDP_REDIRECT:
237 			if (xdp_do_redirect(dp->netdev, xrxbuf->xdp, xdp_prog)) {
238 				nfp_net_xsk_rx_drop(r_vec, xrxbuf);
239 			} else {
240 				nfp_net_xsk_rx_unstash(xrxbuf);
241 				xdp_redir = true;
242 			}
243 			break;
244 		default:
245 			bpf_warn_invalid_xdp_action(dp->netdev, xdp_prog, act);
246 			fallthrough;
247 		case XDP_ABORTED:
248 			trace_xdp_exception(dp->netdev, xdp_prog, act);
249 			fallthrough;
250 		case XDP_DROP:
251 			nfp_net_xsk_rx_drop(r_vec, xrxbuf);
252 			break;
253 		}
254 	}
255 
256 	nfp_net_xsk_rx_ring_fill_freelist(r_vec->rx_ring);
257 
258 	if (xdp_redir)
259 		xdp_do_flush_map();
260 
261 	if (tx_ring->wr_ptr_add)
262 		nfp_net_tx_xmit_more_flush(tx_ring);
263 
264 	return pkts_polled;
265 }
266 
267 void nfp_nfd3_xsk_tx_free(struct nfp_nfd3_tx_buf *txbuf)
268 {
269 	xsk_buff_free(txbuf->xdp);
270 
271 	txbuf->dma_addr = 0;
272 	txbuf->xdp = NULL;
273 }
274 
275 static bool nfp_nfd3_xsk_complete(struct nfp_net_tx_ring *tx_ring)
276 {
277 	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
278 	u32 done_pkts = 0, done_bytes = 0, reused = 0;
279 	bool done_all;
280 	int idx, todo;
281 	u32 qcp_rd_p;
282 
283 	if (tx_ring->wr_p == tx_ring->rd_p)
284 		return true;
285 
286 	/* Work out how many descriptors have been transmitted. */
287 	qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q);
288 
289 	if (qcp_rd_p == tx_ring->qcp_rd_p)
290 		return true;
291 
292 	todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
293 
294 	done_all = todo <= NFP_NET_XDP_MAX_COMPLETE;
295 	todo = min(todo, NFP_NET_XDP_MAX_COMPLETE);
296 
297 	tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo);
298 
299 	done_pkts = todo;
300 	while (todo--) {
301 		struct nfp_nfd3_tx_buf *txbuf;
302 
303 		idx = D_IDX(tx_ring, tx_ring->rd_p);
304 		tx_ring->rd_p++;
305 
306 		txbuf = &tx_ring->txbufs[idx];
307 		if (unlikely(!txbuf->real_len))
308 			continue;
309 
310 		done_bytes += txbuf->real_len;
311 		txbuf->real_len = 0;
312 
313 		if (txbuf->is_xsk_tx) {
314 			nfp_nfd3_xsk_tx_free(txbuf);
315 			reused++;
316 		}
317 	}
318 
319 	u64_stats_update_begin(&r_vec->tx_sync);
320 	r_vec->tx_bytes += done_bytes;
321 	r_vec->tx_pkts += done_pkts;
322 	u64_stats_update_end(&r_vec->tx_sync);
323 
324 	xsk_tx_completed(r_vec->xsk_pool, done_pkts - reused);
325 
326 	WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
327 		  "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
328 		  tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
329 
330 	return done_all;
331 }
332 
333 static void nfp_nfd3_xsk_tx(struct nfp_net_tx_ring *tx_ring)
334 {
335 	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
336 	struct xdp_desc desc[NFP_NET_XSK_TX_BATCH];
337 	struct xsk_buff_pool *xsk_pool;
338 	struct nfp_nfd3_tx_desc *txd;
339 	u32 pkts = 0, wr_idx;
340 	u32 i, got;
341 
342 	xsk_pool = r_vec->xsk_pool;
343 
344 	while (nfp_net_tx_space(tx_ring) >= NFP_NET_XSK_TX_BATCH) {
345 		for (i = 0; i < NFP_NET_XSK_TX_BATCH; i++)
346 			if (!xsk_tx_peek_desc(xsk_pool, &desc[i]))
347 				break;
348 		got = i;
349 		if (!got)
350 			break;
351 
352 		wr_idx = D_IDX(tx_ring, tx_ring->wr_p + i);
353 		prefetchw(&tx_ring->txds[wr_idx]);
354 
355 		for (i = 0; i < got; i++)
356 			xsk_buff_raw_dma_sync_for_device(xsk_pool, desc[i].addr,
357 							 desc[i].len);
358 
359 		for (i = 0; i < got; i++) {
360 			wr_idx = D_IDX(tx_ring, tx_ring->wr_p + i);
361 
362 			tx_ring->txbufs[wr_idx].real_len = desc[i].len;
363 			tx_ring->txbufs[wr_idx].is_xsk_tx = false;
364 
365 			/* Build TX descriptor. */
366 			txd = &tx_ring->txds[wr_idx];
367 			nfp_desc_set_dma_addr_40b(txd,
368 						  xsk_buff_raw_get_dma(xsk_pool, desc[i].addr));
369 			txd->offset_eop = NFD3_DESC_TX_EOP;
370 			txd->dma_len = cpu_to_le16(desc[i].len);
371 			txd->data_len = cpu_to_le16(desc[i].len);
372 		}
373 
374 		tx_ring->wr_p += got;
375 		pkts += got;
376 	}
377 
378 	if (!pkts)
379 		return;
380 
381 	xsk_tx_release(xsk_pool);
382 	/* Ensure all records are visible before incrementing write counter. */
383 	wmb();
384 	nfp_qcp_wr_ptr_add(tx_ring->qcp_q, pkts);
385 }
386 
387 int nfp_nfd3_xsk_poll(struct napi_struct *napi, int budget)
388 {
389 	struct nfp_net_r_vector *r_vec =
390 		container_of(napi, struct nfp_net_r_vector, napi);
391 	unsigned int pkts_polled, skbs = 0;
392 
393 	pkts_polled = nfp_nfd3_xsk_rx(r_vec->rx_ring, budget, &skbs);
394 
395 	if (pkts_polled < budget) {
396 		if (r_vec->tx_ring)
397 			nfp_nfd3_tx_complete(r_vec->tx_ring, budget);
398 
399 		if (!nfp_nfd3_xsk_complete(r_vec->xdp_ring))
400 			pkts_polled = budget;
401 
402 		nfp_nfd3_xsk_tx(r_vec->xdp_ring);
403 
404 		if (pkts_polled < budget && napi_complete_done(napi, skbs))
405 			nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
406 	}
407 
408 	return pkts_polled;
409 }
410