xref: /openbmc/linux/drivers/net/ethernet/sfc/rx.c (revision d9fd5a71)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3  * Driver for Solarflare network controllers and boards
4  * Copyright 2005-2006 Fen Systems Ltd.
5  * Copyright 2005-2013 Solarflare Communications Inc.
6  */
7 
8 #include <linux/socket.h>
9 #include <linux/in.h>
10 #include <linux/slab.h>
11 #include <linux/ip.h>
12 #include <linux/ipv6.h>
13 #include <linux/tcp.h>
14 #include <linux/udp.h>
15 #include <linux/prefetch.h>
16 #include <linux/moduleparam.h>
17 #include <linux/iommu.h>
18 #include <net/ip.h>
19 #include <net/checksum.h>
20 #include <net/xdp.h>
21 #include <linux/bpf_trace.h>
22 #include "net_driver.h"
23 #include "efx.h"
24 #include "rx_common.h"
25 #include "filter.h"
26 #include "nic.h"
27 #include "selftest.h"
28 #include "workarounds.h"
29 
30 /* Preferred number of descriptors to fill at once */
31 #define EFX_RX_PREFERRED_BATCH 8U
32 
33 /* Maximum rx prefix used by any architecture. */
34 #define EFX_MAX_RX_PREFIX_SIZE 16
35 
36 /* Size of buffer allocated for skb header area. */
37 #define EFX_SKB_HEADERS  128u
38 
39 /* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */
40 #define EFX_RX_MAX_FRAGS DIV_ROUND_UP(EFX_MAX_FRAME_LEN(EFX_MAX_MTU), \
41 				      EFX_RX_USR_BUF_SIZE)
42 
43 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
44 				     struct efx_rx_buffer *rx_buf,
45 				     int len)
46 {
47 	struct efx_nic *efx = rx_queue->efx;
48 	unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
49 
50 	if (likely(len <= max_len))
51 		return;
52 
53 	/* The packet must be discarded, but this is only a fatal error
54 	 * if the caller indicated it was
55 	 */
56 	rx_buf->flags |= EFX_RX_PKT_DISCARD;
57 
58 	if (net_ratelimit())
59 		netif_err(efx, rx_err, efx->net_dev,
60 			  "RX queue %d overlength RX event (%#x > %#x)\n",
61 			  efx_rx_queue_index(rx_queue), len, max_len);
62 
63 	efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
64 }
65 
66 /* Allocate and construct an SKB around page fragments */
67 static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
68 				     struct efx_rx_buffer *rx_buf,
69 				     unsigned int n_frags,
70 				     u8 *eh, int hdr_len)
71 {
72 	struct efx_nic *efx = channel->efx;
73 	struct sk_buff *skb;
74 
75 	/* Allocate an SKB to store the headers */
76 	skb = netdev_alloc_skb(efx->net_dev,
77 			       efx->rx_ip_align + efx->rx_prefix_size +
78 			       hdr_len);
79 	if (unlikely(skb == NULL)) {
80 		atomic_inc(&efx->n_rx_noskb_drops);
81 		return NULL;
82 	}
83 
84 	EFX_WARN_ON_ONCE_PARANOID(rx_buf->len < hdr_len);
85 
86 	memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size,
87 	       efx->rx_prefix_size + hdr_len);
88 	skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size);
89 	__skb_put(skb, hdr_len);
90 
91 	/* Append the remaining page(s) onto the frag list */
92 	if (rx_buf->len > hdr_len) {
93 		rx_buf->page_offset += hdr_len;
94 		rx_buf->len -= hdr_len;
95 
96 		for (;;) {
97 			skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
98 					   rx_buf->page, rx_buf->page_offset,
99 					   rx_buf->len);
100 			rx_buf->page = NULL;
101 			skb->len += rx_buf->len;
102 			skb->data_len += rx_buf->len;
103 			if (skb_shinfo(skb)->nr_frags == n_frags)
104 				break;
105 
106 			rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
107 		}
108 	} else {
109 		__free_pages(rx_buf->page, efx->rx_buffer_order);
110 		rx_buf->page = NULL;
111 		n_frags = 0;
112 	}
113 
114 	skb->truesize += n_frags * efx->rx_buffer_truesize;
115 
116 	/* Move past the ethernet header */
117 	skb->protocol = eth_type_trans(skb, efx->net_dev);
118 
119 	skb_mark_napi_id(skb, &channel->napi_str);
120 
121 	return skb;
122 }
123 
124 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
125 		   unsigned int n_frags, unsigned int len, u16 flags)
126 {
127 	struct efx_nic *efx = rx_queue->efx;
128 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
129 	struct efx_rx_buffer *rx_buf;
130 
131 	rx_queue->rx_packets++;
132 
133 	rx_buf = efx_rx_buffer(rx_queue, index);
134 	rx_buf->flags |= flags;
135 
136 	/* Validate the number of fragments and completed length */
137 	if (n_frags == 1) {
138 		if (!(flags & EFX_RX_PKT_PREFIX_LEN))
139 			efx_rx_packet__check_len(rx_queue, rx_buf, len);
140 	} else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
141 		   unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
142 		   unlikely(len > n_frags * efx->rx_dma_len) ||
143 		   unlikely(!efx->rx_scatter)) {
144 		/* If this isn't an explicit discard request, either
145 		 * the hardware or the driver is broken.
146 		 */
147 		WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
148 		rx_buf->flags |= EFX_RX_PKT_DISCARD;
149 	}
150 
151 	netif_vdbg(efx, rx_status, efx->net_dev,
152 		   "RX queue %d received ids %x-%x len %d %s%s\n",
153 		   efx_rx_queue_index(rx_queue), index,
154 		   (index + n_frags - 1) & rx_queue->ptr_mask, len,
155 		   (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
156 		   (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
157 
158 	/* Discard packet, if instructed to do so.  Process the
159 	 * previous receive first.
160 	 */
161 	if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
162 		efx_rx_flush_packet(channel);
163 		efx_discard_rx_packet(channel, rx_buf, n_frags);
164 		return;
165 	}
166 
167 	if (n_frags == 1 && !(flags & EFX_RX_PKT_PREFIX_LEN))
168 		rx_buf->len = len;
169 
170 	/* Release and/or sync the DMA mapping - assumes all RX buffers
171 	 * consumed in-order per RX queue.
172 	 */
173 	efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
174 
175 	/* Prefetch nice and early so data will (hopefully) be in cache by
176 	 * the time we look at it.
177 	 */
178 	prefetch(efx_rx_buf_va(rx_buf));
179 
180 	rx_buf->page_offset += efx->rx_prefix_size;
181 	rx_buf->len -= efx->rx_prefix_size;
182 
183 	if (n_frags > 1) {
184 		/* Release/sync DMA mapping for additional fragments.
185 		 * Fix length for last fragment.
186 		 */
187 		unsigned int tail_frags = n_frags - 1;
188 
189 		for (;;) {
190 			rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
191 			if (--tail_frags == 0)
192 				break;
193 			efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
194 		}
195 		rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
196 		efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
197 	}
198 
199 	/* All fragments have been DMA-synced, so recycle pages. */
200 	rx_buf = efx_rx_buffer(rx_queue, index);
201 	efx_recycle_rx_pages(channel, rx_buf, n_frags);
202 
203 	/* Pipeline receives so that we give time for packet headers to be
204 	 * prefetched into cache.
205 	 */
206 	efx_rx_flush_packet(channel);
207 	channel->rx_pkt_n_frags = n_frags;
208 	channel->rx_pkt_index = index;
209 }
210 
211 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
212 			   struct efx_rx_buffer *rx_buf,
213 			   unsigned int n_frags)
214 {
215 	struct sk_buff *skb;
216 	u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
217 
218 	skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
219 	if (unlikely(skb == NULL)) {
220 		struct efx_rx_queue *rx_queue;
221 
222 		rx_queue = efx_channel_get_rx_queue(channel);
223 		efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
224 		return;
225 	}
226 	skb_record_rx_queue(skb, channel->rx_queue.core_index);
227 
228 	/* Set the SKB flags */
229 	skb_checksum_none_assert(skb);
230 	if (likely(rx_buf->flags & EFX_RX_PKT_CSUMMED)) {
231 		skb->ip_summed = CHECKSUM_UNNECESSARY;
232 		skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
233 	}
234 
235 	efx_rx_skb_attach_timestamp(channel, skb);
236 
237 	if (channel->type->receive_skb)
238 		if (channel->type->receive_skb(channel, skb))
239 			return;
240 
241 	/* Pass the packet up */
242 	if (channel->rx_list != NULL)
243 		/* Add to list, will pass up later */
244 		list_add_tail(&skb->list, channel->rx_list);
245 	else
246 		/* No list, so pass it up now */
247 		netif_receive_skb(skb);
248 }
249 
250 /** efx_do_xdp: perform XDP processing on a received packet
251  *
252  * Returns true if packet should still be delivered.
253  */
254 static bool efx_do_xdp(struct efx_nic *efx, struct efx_channel *channel,
255 		       struct efx_rx_buffer *rx_buf, u8 **ehp)
256 {
257 	u8 rx_prefix[EFX_MAX_RX_PREFIX_SIZE];
258 	struct efx_rx_queue *rx_queue;
259 	struct bpf_prog *xdp_prog;
260 	struct xdp_frame *xdpf;
261 	struct xdp_buff xdp;
262 	u32 xdp_act;
263 	s16 offset;
264 	int err;
265 
266 	rcu_read_lock();
267 	xdp_prog = rcu_dereference(efx->xdp_prog);
268 	if (!xdp_prog) {
269 		rcu_read_unlock();
270 		return true;
271 	}
272 
273 	rx_queue = efx_channel_get_rx_queue(channel);
274 
275 	if (unlikely(channel->rx_pkt_n_frags > 1)) {
276 		/* We can't do XDP on fragmented packets - drop. */
277 		rcu_read_unlock();
278 		efx_free_rx_buffers(rx_queue, rx_buf,
279 				    channel->rx_pkt_n_frags);
280 		if (net_ratelimit())
281 			netif_err(efx, rx_err, efx->net_dev,
282 				  "XDP is not possible with multiple receive fragments (%d)\n",
283 				  channel->rx_pkt_n_frags);
284 		channel->n_rx_xdp_bad_drops++;
285 		return false;
286 	}
287 
288 	dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr,
289 				rx_buf->len, DMA_FROM_DEVICE);
290 
291 	/* Save the rx prefix. */
292 	EFX_WARN_ON_PARANOID(efx->rx_prefix_size > EFX_MAX_RX_PREFIX_SIZE);
293 	memcpy(rx_prefix, *ehp - efx->rx_prefix_size,
294 	       efx->rx_prefix_size);
295 
296 	xdp_init_buff(&xdp, efx->rx_page_buf_step, &rx_queue->xdp_rxq_info);
297 	/* No support yet for XDP metadata */
298 	xdp_prepare_buff(&xdp, *ehp - EFX_XDP_HEADROOM, EFX_XDP_HEADROOM,
299 			 rx_buf->len, false);
300 
301 	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
302 	rcu_read_unlock();
303 
304 	offset = (u8 *)xdp.data - *ehp;
305 
306 	switch (xdp_act) {
307 	case XDP_PASS:
308 		/* Fix up rx prefix. */
309 		if (offset) {
310 			*ehp += offset;
311 			rx_buf->page_offset += offset;
312 			rx_buf->len -= offset;
313 			memcpy(*ehp - efx->rx_prefix_size, rx_prefix,
314 			       efx->rx_prefix_size);
315 		}
316 		break;
317 
318 	case XDP_TX:
319 		/* Buffer ownership passes to tx on success. */
320 		xdpf = xdp_convert_buff_to_frame(&xdp);
321 		err = efx_xdp_tx_buffers(efx, 1, &xdpf, true);
322 		if (unlikely(err != 1)) {
323 			efx_free_rx_buffers(rx_queue, rx_buf, 1);
324 			if (net_ratelimit())
325 				netif_err(efx, rx_err, efx->net_dev,
326 					  "XDP TX failed (%d)\n", err);
327 			channel->n_rx_xdp_bad_drops++;
328 			trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
329 		} else {
330 			channel->n_rx_xdp_tx++;
331 		}
332 		break;
333 
334 	case XDP_REDIRECT:
335 		err = xdp_do_redirect(efx->net_dev, &xdp, xdp_prog);
336 		if (unlikely(err)) {
337 			efx_free_rx_buffers(rx_queue, rx_buf, 1);
338 			if (net_ratelimit())
339 				netif_err(efx, rx_err, efx->net_dev,
340 					  "XDP redirect failed (%d)\n", err);
341 			channel->n_rx_xdp_bad_drops++;
342 			trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
343 		} else {
344 			channel->n_rx_xdp_redirect++;
345 		}
346 		break;
347 
348 	default:
349 		bpf_warn_invalid_xdp_action(xdp_act);
350 		efx_free_rx_buffers(rx_queue, rx_buf, 1);
351 		channel->n_rx_xdp_bad_drops++;
352 		trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
353 		break;
354 
355 	case XDP_ABORTED:
356 		trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
357 		fallthrough;
358 	case XDP_DROP:
359 		efx_free_rx_buffers(rx_queue, rx_buf, 1);
360 		channel->n_rx_xdp_drops++;
361 		break;
362 	}
363 
364 	return xdp_act == XDP_PASS;
365 }
366 
367 /* Handle a received packet.  Second half: Touches packet payload. */
368 void __efx_rx_packet(struct efx_channel *channel)
369 {
370 	struct efx_nic *efx = channel->efx;
371 	struct efx_rx_buffer *rx_buf =
372 		efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
373 	u8 *eh = efx_rx_buf_va(rx_buf);
374 
375 	/* Read length from the prefix if necessary.  This already
376 	 * excludes the length of the prefix itself.
377 	 */
378 	if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN)
379 		rx_buf->len = le16_to_cpup((__le16 *)
380 					   (eh + efx->rx_packet_len_offset));
381 
382 	/* If we're in loopback test, then pass the packet directly to the
383 	 * loopback layer, and free the rx_buf here
384 	 */
385 	if (unlikely(efx->loopback_selftest)) {
386 		struct efx_rx_queue *rx_queue;
387 
388 		efx_loopback_rx_packet(efx, eh, rx_buf->len);
389 		rx_queue = efx_channel_get_rx_queue(channel);
390 		efx_free_rx_buffers(rx_queue, rx_buf,
391 				    channel->rx_pkt_n_frags);
392 		goto out;
393 	}
394 
395 	if (!efx_do_xdp(efx, channel, rx_buf, &eh))
396 		goto out;
397 
398 	if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
399 		rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
400 
401 	if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb)
402 		efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh, 0);
403 	else
404 		efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
405 out:
406 	channel->rx_pkt_n_frags = 0;
407 }
408