xref: /linux/net/core/tso.c (revision 05e352444b2430de4b183b4a988085381e5fd6ad)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/export.h>
3 #include <linux/if_vlan.h>
4 #include <net/ip.h>
5 #include <net/tso.h>
6 #include <linux/dma-mapping.h>
7 #include <linux/unaligned.h>
8 
9 void tso_build_hdr(const struct sk_buff *skb, char *hdr, struct tso_t *tso,
10 		   int size, bool is_last)
11 {
12 	int hdr_len = skb_transport_offset(skb) + tso->tlen;
13 	int mac_hdr_len = skb_network_offset(skb);
14 
15 	memcpy(hdr, skb->data, hdr_len);
16 	if (!tso->ipv6) {
17 		struct iphdr *iph = (void *)(hdr + mac_hdr_len);
18 
19 		iph->id = htons(tso->ip_id);
20 		iph->tot_len = htons(size + hdr_len - mac_hdr_len);
21 		tso->ip_id++;
22 	} else {
23 		struct ipv6hdr *iph = (void *)(hdr + mac_hdr_len);
24 
25 		iph->payload_len = htons(size + tso->tlen);
26 	}
27 	hdr += skb_transport_offset(skb);
28 	if (tso->tlen != sizeof(struct udphdr)) {
29 		struct tcphdr *tcph = (struct tcphdr *)hdr;
30 
31 		put_unaligned_be32(tso->tcp_seq, &tcph->seq);
32 
33 		if (!is_last) {
34 			/* Clear all special flags for not last packet */
35 			tcph->psh = 0;
36 			tcph->fin = 0;
37 			tcph->rst = 0;
38 		}
39 	} else {
40 		struct udphdr *uh = (struct udphdr *)hdr;
41 
42 		uh->len = htons(sizeof(*uh) + size);
43 	}
44 }
45 EXPORT_SYMBOL(tso_build_hdr);
46 
47 void tso_build_data(const struct sk_buff *skb, struct tso_t *tso, int size)
48 {
49 	tso->tcp_seq += size; /* not worth avoiding this operation for UDP */
50 	tso->size -= size;
51 	tso->data += size;
52 
53 	if ((tso->size == 0) &&
54 	    (tso->next_frag_idx < skb_shinfo(skb)->nr_frags)) {
55 		skb_frag_t *frag = &skb_shinfo(skb)->frags[tso->next_frag_idx];
56 
57 		/* Move to next segment */
58 		tso->size = skb_frag_size(frag);
59 		tso->data = skb_frag_address(frag);
60 		tso->next_frag_idx++;
61 	}
62 }
63 EXPORT_SYMBOL(tso_build_data);
64 
65 int tso_start(struct sk_buff *skb, struct tso_t *tso)
66 {
67 	int tlen = skb_is_gso_tcp(skb) ? tcp_hdrlen(skb) : sizeof(struct udphdr);
68 	int hdr_len = skb_transport_offset(skb) + tlen;
69 
70 	tso->tlen = tlen;
71 	tso->ip_id = ntohs(ip_hdr(skb)->id);
72 	tso->tcp_seq = (tlen != sizeof(struct udphdr)) ? ntohl(tcp_hdr(skb)->seq) : 0;
73 	tso->next_frag_idx = 0;
74 	tso->ipv6 = vlan_get_protocol(skb) == htons(ETH_P_IPV6);
75 
76 	/* Build first data */
77 	tso->size = skb_headlen(skb) - hdr_len;
78 	tso->data = skb->data + hdr_len;
79 	if ((tso->size == 0) &&
80 	    (tso->next_frag_idx < skb_shinfo(skb)->nr_frags)) {
81 		skb_frag_t *frag = &skb_shinfo(skb)->frags[tso->next_frag_idx];
82 
83 		/* Move to next segment */
84 		tso->size = skb_frag_size(frag);
85 		tso->data = skb_frag_address(frag);
86 		tso->next_frag_idx++;
87 	}
88 	return hdr_len;
89 }
90 EXPORT_SYMBOL(tso_start);
91 
92 static int tso_dma_iova_try(struct device *dev, struct tso_dma_map *map,
93 			    phys_addr_t phys, size_t linear_len,
94 			    size_t total_len, size_t *offset)
95 {
96 	const struct sk_buff *skb;
97 	unsigned int nr_frags;
98 	int i;
99 
100 	if (!dma_iova_try_alloc(dev, &map->iova_state, phys, total_len))
101 		return 1;
102 
103 	skb = map->skb;
104 	nr_frags = skb_shinfo(skb)->nr_frags;
105 
106 	if (linear_len) {
107 		if (dma_iova_link(dev, &map->iova_state,
108 				  phys, *offset, linear_len,
109 				  DMA_TO_DEVICE, 0))
110 			goto iova_fail;
111 		map->linear_len = linear_len;
112 		*offset += linear_len;
113 	}
114 
115 	for (i = 0; i < nr_frags; i++) {
116 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
117 		unsigned int frag_len = skb_frag_size(frag);
118 
119 		if (dma_iova_link(dev, &map->iova_state,
120 				  skb_frag_phys(frag), *offset,
121 				  frag_len, DMA_TO_DEVICE, 0)) {
122 			map->nr_frags = i;
123 			goto iova_fail;
124 		}
125 		map->frags[i].len = frag_len;
126 		*offset += frag_len;
127 		map->nr_frags = i + 1;
128 	}
129 
130 	if (dma_iova_sync(dev, &map->iova_state, 0, total_len))
131 		goto iova_fail;
132 
133 	return 0;
134 
135 iova_fail:
136 	dma_iova_destroy(dev, &map->iova_state, *offset,
137 			 DMA_TO_DEVICE, 0);
138 	memset(&map->iova_state, 0, sizeof(map->iova_state));
139 
140 	/* reset map state */
141 	map->frag_idx = -1;
142 	map->offset = 0;
143 	map->linear_len = 0;
144 	map->nr_frags = 0;
145 
146 	return 1;
147 }
148 
149 /**
150  * tso_dma_map_init - DMA-map GSO payload regions
151  * @map: map struct to initialize
152  * @dev: device for DMA mapping
153  * @skb: the GSO skb
154  * @hdr_len: per-segment header length in bytes
155  *
156  * DMA-maps the linear payload (after headers) and all frags.
157  * Prefers the DMA IOVA API (one contiguous mapping, one IOTLB sync);
158  * falls back to per-region dma_map_phys() when IOVA is not available.
159  * Positions the iterator at byte 0 of the payload.
160  *
161  * Return: 0 on success, -ENOMEM on DMA mapping failure (partial mappings
162  * are cleaned up internally).
163  */
164 int tso_dma_map_init(struct tso_dma_map *map, struct device *dev,
165 		     const struct sk_buff *skb, unsigned int hdr_len)
166 {
167 	unsigned int linear_len = skb_headlen(skb) - hdr_len;
168 	unsigned int nr_frags = skb_shinfo(skb)->nr_frags;
169 	size_t total_len = skb->len - hdr_len;
170 	size_t offset = 0;
171 	phys_addr_t phys;
172 	int i;
173 
174 	map->dev = dev;
175 	map->skb = skb;
176 	map->hdr_len = hdr_len;
177 	map->frag_idx = -1;
178 	map->offset = 0;
179 	map->iova_offset = 0;
180 	map->total_len = total_len;
181 	map->linear_len = 0;
182 	map->nr_frags = 0;
183 	memset(&map->iova_state, 0, sizeof(map->iova_state));
184 
185 	if (!total_len)
186 		return 0;
187 
188 	if (linear_len)
189 		phys = virt_to_phys(skb->data + hdr_len);
190 	else
191 		phys = skb_frag_phys(&skb_shinfo(skb)->frags[0]);
192 
193 	if (tso_dma_iova_try(dev, map, phys, linear_len, total_len, &offset)) {
194 		/* IOVA path failed, map state was reset. Fallback to
195 		 * per-region dma_map_phys()
196 		 */
197 		if (linear_len) {
198 			map->linear_dma = dma_map_phys(dev, phys, linear_len,
199 						       DMA_TO_DEVICE, 0);
200 			if (dma_mapping_error(dev, map->linear_dma))
201 				return -ENOMEM;
202 			map->linear_len = linear_len;
203 		}
204 
205 		for (i = 0; i < nr_frags; i++) {
206 			skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
207 			unsigned int frag_len = skb_frag_size(frag);
208 
209 			map->frags[i].len = frag_len;
210 			map->frags[i].dma = dma_map_phys(dev, skb_frag_phys(frag),
211 							 frag_len, DMA_TO_DEVICE, 0);
212 			if (dma_mapping_error(dev, map->frags[i].dma)) {
213 				tso_dma_map_cleanup(map);
214 				return -ENOMEM;
215 			}
216 			map->nr_frags = i + 1;
217 		}
218 	}
219 
220 	if (linear_len == 0 && nr_frags > 0)
221 		map->frag_idx = 0;
222 
223 	return 0;
224 }
225 EXPORT_SYMBOL(tso_dma_map_init);
226 
227 /**
228  * tso_dma_map_cleanup - unmap all DMA regions in a tso_dma_map
229  * @map: the map to clean up
230  *
231  * Handles both IOVA and fallback paths. For IOVA, calls
232  * dma_iova_destroy(). For fallback, unmaps each region individually.
233  */
234 void tso_dma_map_cleanup(struct tso_dma_map *map)
235 {
236 	int i;
237 
238 	if (dma_use_iova(&map->iova_state)) {
239 		dma_iova_destroy(map->dev, &map->iova_state, map->total_len,
240 				 DMA_TO_DEVICE, 0);
241 		memset(&map->iova_state, 0, sizeof(map->iova_state));
242 	} else {
243 		if (map->linear_len)
244 			dma_unmap_phys(map->dev, map->linear_dma,
245 				       map->linear_len, DMA_TO_DEVICE, 0);
246 
247 		for (i = 0; i < map->nr_frags; i++)
248 			dma_unmap_phys(map->dev, map->frags[i].dma,
249 				       map->frags[i].len, DMA_TO_DEVICE, 0);
250 	}
251 
252 	map->linear_len = 0;
253 	map->nr_frags = 0;
254 }
255 EXPORT_SYMBOL(tso_dma_map_cleanup);
256 
257 /**
258  * tso_dma_map_count - count descriptors for a payload range
259  * @map: the payload map
260  * @len: number of payload bytes in this segment
261  *
262  * Counts how many contiguous DMA region chunks the next @len bytes
263  * will span, without advancing the iterator. On the IOVA path this
264  * is always 1 (contiguous). On the fallback path, uses region sizes
265  * from the current position.
266  *
267  * Return: the number of descriptors needed for @len bytes of payload.
268  */
269 unsigned int tso_dma_map_count(struct tso_dma_map *map, unsigned int len)
270 {
271 	unsigned int offset = map->offset;
272 	int idx = map->frag_idx;
273 	unsigned int count = 0;
274 
275 	if (!len)
276 		return 0;
277 
278 	if (dma_use_iova(&map->iova_state))
279 		return 1;
280 
281 	while (len > 0) {
282 		unsigned int region_len, chunk;
283 
284 		if (idx == -1)
285 			region_len = map->linear_len;
286 		else
287 			region_len = map->frags[idx].len;
288 
289 		chunk = min(len, region_len - offset);
290 		len -= chunk;
291 		count++;
292 		offset = 0;
293 		idx++;
294 	}
295 
296 	return count;
297 }
298 EXPORT_SYMBOL(tso_dma_map_count);
299 
300 /**
301  * tso_dma_map_next - yield the next DMA address range
302  * @map: the payload map
303  * @addr: output DMA address
304  * @chunk_len: output chunk length
305  * @mapping_len: full DMA mapping length when this chunk starts a new
306  *               mapping region, or 0 when continuing a previous one.
307  *               On the IOVA path this is always 0 (driver must not
308  *               do per-region unmaps; use tso_dma_map_cleanup instead).
309  * @seg_remaining: bytes left in current segment
310  *
311  * Yields the next (dma_addr, chunk_len) pair and advances the iterator.
312  * On the IOVA path, the entire payload is contiguous so each segment
313  * is always a single chunk.
314  *
315  * Return: true if a chunk was yielded, false when @seg_remaining is 0.
316  */
317 bool tso_dma_map_next(struct tso_dma_map *map, dma_addr_t *addr,
318 		      unsigned int *chunk_len, unsigned int *mapping_len,
319 		      unsigned int seg_remaining)
320 {
321 	unsigned int region_len, chunk;
322 
323 	if (!seg_remaining)
324 		return false;
325 
326 	/* IOVA path: contiguous DMA range, no region boundaries */
327 	if (dma_use_iova(&map->iova_state)) {
328 		*addr = map->iova_state.addr + map->iova_offset;
329 		*chunk_len = seg_remaining;
330 		*mapping_len = 0;
331 		map->iova_offset += seg_remaining;
332 		return true;
333 	}
334 
335 	/* Fallback path: per-region iteration */
336 
337 	if (map->frag_idx == -1) {
338 		region_len = map->linear_len;
339 		chunk = min(seg_remaining, region_len - map->offset);
340 		*addr = map->linear_dma + map->offset;
341 	} else {
342 		region_len = map->frags[map->frag_idx].len;
343 		chunk = min(seg_remaining, region_len - map->offset);
344 		*addr = map->frags[map->frag_idx].dma + map->offset;
345 	}
346 
347 	*mapping_len = (map->offset == 0) ? region_len : 0;
348 	*chunk_len = chunk;
349 	map->offset += chunk;
350 
351 	if (map->offset >= region_len) {
352 		map->frag_idx++;
353 		map->offset = 0;
354 	}
355 
356 	return true;
357 }
358 EXPORT_SYMBOL(tso_dma_map_next);
359