1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2020 Inspur Corporation
3 */
4
5 #include <rte_malloc.h>
6 #include <rte_mbuf.h>
7 #include <rte_cycles.h>
8 #include <rte_ethdev.h>
9 #include <rte_udp.h>
10
11 #include "gro_vxlan_udp4.h"
12
13 void *
gro_vxlan_udp4_tbl_create(uint16_t socket_id,uint16_t max_flow_num,uint16_t max_item_per_flow)14 gro_vxlan_udp4_tbl_create(uint16_t socket_id,
15 uint16_t max_flow_num,
16 uint16_t max_item_per_flow)
17 {
18 struct gro_vxlan_udp4_tbl *tbl;
19 size_t size;
20 uint32_t entries_num, i;
21
22 entries_num = max_flow_num * max_item_per_flow;
23 entries_num = RTE_MIN(entries_num, GRO_VXLAN_UDP4_TBL_MAX_ITEM_NUM);
24
25 if (entries_num == 0)
26 return NULL;
27
28 tbl = rte_zmalloc_socket(__func__,
29 sizeof(struct gro_vxlan_udp4_tbl),
30 RTE_CACHE_LINE_SIZE,
31 socket_id);
32 if (tbl == NULL)
33 return NULL;
34
35 size = sizeof(struct gro_vxlan_udp4_item) * entries_num;
36 tbl->items = rte_zmalloc_socket(__func__,
37 size,
38 RTE_CACHE_LINE_SIZE,
39 socket_id);
40 if (tbl->items == NULL) {
41 rte_free(tbl);
42 return NULL;
43 }
44 tbl->max_item_num = entries_num;
45
46 size = sizeof(struct gro_vxlan_udp4_flow) * entries_num;
47 tbl->flows = rte_zmalloc_socket(__func__,
48 size,
49 RTE_CACHE_LINE_SIZE,
50 socket_id);
51 if (tbl->flows == NULL) {
52 rte_free(tbl->items);
53 rte_free(tbl);
54 return NULL;
55 }
56
57 for (i = 0; i < entries_num; i++)
58 tbl->flows[i].start_index = INVALID_ARRAY_INDEX;
59 tbl->max_flow_num = entries_num;
60
61 return tbl;
62 }
63
64 void
gro_vxlan_udp4_tbl_destroy(void * tbl)65 gro_vxlan_udp4_tbl_destroy(void *tbl)
66 {
67 struct gro_vxlan_udp4_tbl *vxlan_tbl = tbl;
68
69 if (vxlan_tbl) {
70 rte_free(vxlan_tbl->items);
71 rte_free(vxlan_tbl->flows);
72 }
73 rte_free(vxlan_tbl);
74 }
75
76 static inline uint32_t
find_an_empty_item(struct gro_vxlan_udp4_tbl * tbl)77 find_an_empty_item(struct gro_vxlan_udp4_tbl *tbl)
78 {
79 uint32_t max_item_num = tbl->max_item_num, i;
80
81 for (i = 0; i < max_item_num; i++)
82 if (tbl->items[i].inner_item.firstseg == NULL)
83 return i;
84 return INVALID_ARRAY_INDEX;
85 }
86
87 static inline uint32_t
find_an_empty_flow(struct gro_vxlan_udp4_tbl * tbl)88 find_an_empty_flow(struct gro_vxlan_udp4_tbl *tbl)
89 {
90 uint32_t max_flow_num = tbl->max_flow_num, i;
91
92 for (i = 0; i < max_flow_num; i++)
93 if (tbl->flows[i].start_index == INVALID_ARRAY_INDEX)
94 return i;
95 return INVALID_ARRAY_INDEX;
96 }
97
98 static inline uint32_t
insert_new_item(struct gro_vxlan_udp4_tbl * tbl,struct rte_mbuf * pkt,uint64_t start_time,uint32_t prev_idx,uint16_t frag_offset,uint8_t is_last_frag)99 insert_new_item(struct gro_vxlan_udp4_tbl *tbl,
100 struct rte_mbuf *pkt,
101 uint64_t start_time,
102 uint32_t prev_idx,
103 uint16_t frag_offset,
104 uint8_t is_last_frag)
105 {
106 uint32_t item_idx;
107
108 item_idx = find_an_empty_item(tbl);
109 if (unlikely(item_idx == INVALID_ARRAY_INDEX))
110 return INVALID_ARRAY_INDEX;
111
112 tbl->items[item_idx].inner_item.firstseg = pkt;
113 tbl->items[item_idx].inner_item.lastseg = rte_pktmbuf_lastseg(pkt);
114 tbl->items[item_idx].inner_item.start_time = start_time;
115 tbl->items[item_idx].inner_item.next_pkt_idx = INVALID_ARRAY_INDEX;
116 tbl->items[item_idx].inner_item.frag_offset = frag_offset;
117 tbl->items[item_idx].inner_item.is_last_frag = is_last_frag;
118 tbl->items[item_idx].inner_item.nb_merged = 1;
119 tbl->item_num++;
120
121 /* If the previous packet exists, chain the new one with it. */
122 if (prev_idx != INVALID_ARRAY_INDEX) {
123 tbl->items[item_idx].inner_item.next_pkt_idx =
124 tbl->items[prev_idx].inner_item.next_pkt_idx;
125 tbl->items[prev_idx].inner_item.next_pkt_idx = item_idx;
126 }
127
128 return item_idx;
129 }
130
131 static inline uint32_t
delete_item(struct gro_vxlan_udp4_tbl * tbl,uint32_t item_idx,uint32_t prev_item_idx)132 delete_item(struct gro_vxlan_udp4_tbl *tbl,
133 uint32_t item_idx,
134 uint32_t prev_item_idx)
135 {
136 uint32_t next_idx = tbl->items[item_idx].inner_item.next_pkt_idx;
137
138 /* NULL indicates an empty item. */
139 tbl->items[item_idx].inner_item.firstseg = NULL;
140 tbl->item_num--;
141 if (prev_item_idx != INVALID_ARRAY_INDEX)
142 tbl->items[prev_item_idx].inner_item.next_pkt_idx = next_idx;
143
144 return next_idx;
145 }
146
147 static inline uint32_t
insert_new_flow(struct gro_vxlan_udp4_tbl * tbl,struct vxlan_udp4_flow_key * src,uint32_t item_idx)148 insert_new_flow(struct gro_vxlan_udp4_tbl *tbl,
149 struct vxlan_udp4_flow_key *src,
150 uint32_t item_idx)
151 {
152 struct vxlan_udp4_flow_key *dst;
153 uint32_t flow_idx;
154
155 flow_idx = find_an_empty_flow(tbl);
156 if (unlikely(flow_idx == INVALID_ARRAY_INDEX))
157 return INVALID_ARRAY_INDEX;
158
159 dst = &(tbl->flows[flow_idx].key);
160
161 rte_ether_addr_copy(&(src->inner_key.eth_saddr),
162 &(dst->inner_key.eth_saddr));
163 rte_ether_addr_copy(&(src->inner_key.eth_daddr),
164 &(dst->inner_key.eth_daddr));
165 dst->inner_key.ip_src_addr = src->inner_key.ip_src_addr;
166 dst->inner_key.ip_dst_addr = src->inner_key.ip_dst_addr;
167 dst->inner_key.ip_id = src->inner_key.ip_id;
168
169 dst->vxlan_hdr.vx_flags = src->vxlan_hdr.vx_flags;
170 dst->vxlan_hdr.vx_vni = src->vxlan_hdr.vx_vni;
171 rte_ether_addr_copy(&(src->outer_eth_saddr), &(dst->outer_eth_saddr));
172 rte_ether_addr_copy(&(src->outer_eth_daddr), &(dst->outer_eth_daddr));
173 dst->outer_ip_src_addr = src->outer_ip_src_addr;
174 dst->outer_ip_dst_addr = src->outer_ip_dst_addr;
175 dst->outer_dst_port = src->outer_dst_port;
176
177 tbl->flows[flow_idx].start_index = item_idx;
178 tbl->flow_num++;
179
180 return flow_idx;
181 }
182
183 static inline int
is_same_vxlan_udp4_flow(struct vxlan_udp4_flow_key k1,struct vxlan_udp4_flow_key k2)184 is_same_vxlan_udp4_flow(struct vxlan_udp4_flow_key k1,
185 struct vxlan_udp4_flow_key k2)
186 {
187 /* For VxLAN packet, outer udp src port is calculated from
188 * inner packet RSS hash, udp src port of the first UDP
189 * fragment is different from one of other UDP fragments
190 * even if they are same flow, so we have to skip outer udp
191 * src port comparison here.
192 */
193 return (rte_is_same_ether_addr(&k1.outer_eth_saddr,
194 &k2.outer_eth_saddr) &&
195 rte_is_same_ether_addr(&k1.outer_eth_daddr,
196 &k2.outer_eth_daddr) &&
197 (k1.outer_ip_src_addr == k2.outer_ip_src_addr) &&
198 (k1.outer_ip_dst_addr == k2.outer_ip_dst_addr) &&
199 (k1.outer_dst_port == k2.outer_dst_port) &&
200 (k1.vxlan_hdr.vx_flags == k2.vxlan_hdr.vx_flags) &&
201 (k1.vxlan_hdr.vx_vni == k2.vxlan_hdr.vx_vni) &&
202 is_same_udp4_flow(k1.inner_key, k2.inner_key));
203 }
204
205 static inline int
udp4_check_vxlan_neighbor(struct gro_vxlan_udp4_item * item,uint16_t frag_offset,uint16_t ip_dl)206 udp4_check_vxlan_neighbor(struct gro_vxlan_udp4_item *item,
207 uint16_t frag_offset,
208 uint16_t ip_dl)
209 {
210 struct rte_mbuf *pkt = item->inner_item.firstseg;
211 int cmp;
212 uint16_t l2_offset;
213 int ret = 0;
214
215 /* Note: if outer DF bit is set, i.e outer_is_atomic is 0,
216 * we needn't compare outer_ip_id because they are same,
217 * for the case outer_is_atomic is 1, we also have no way
218 * to compare outer_ip_id because the difference between
219 * outer_ip_ids of two received packets isn't always +/-1.
220 * So skip outer_ip_id comparison here.
221 */
222
223 l2_offset = pkt->outer_l2_len + pkt->outer_l3_len;
224 cmp = udp4_check_neighbor(&item->inner_item, frag_offset, ip_dl,
225 l2_offset);
226 if (cmp > 0)
227 /* Append the new packet. */
228 ret = 1;
229 else if (cmp < 0)
230 /* Prepend the new packet. */
231 ret = -1;
232
233 return ret;
234 }
235
236 static inline int
merge_two_vxlan_udp4_packets(struct gro_vxlan_udp4_item * item,struct rte_mbuf * pkt,int cmp,uint16_t frag_offset,uint8_t is_last_frag)237 merge_two_vxlan_udp4_packets(struct gro_vxlan_udp4_item *item,
238 struct rte_mbuf *pkt,
239 int cmp,
240 uint16_t frag_offset,
241 uint8_t is_last_frag)
242 {
243 if (merge_two_udp4_packets(&item->inner_item, pkt, cmp, frag_offset,
244 is_last_frag,
245 pkt->outer_l2_len + pkt->outer_l3_len)) {
246 return 1;
247 }
248
249 return 0;
250 }
251
252 static inline void
update_vxlan_header(struct gro_vxlan_udp4_item * item)253 update_vxlan_header(struct gro_vxlan_udp4_item *item)
254 {
255 struct rte_ipv4_hdr *ipv4_hdr;
256 struct rte_udp_hdr *udp_hdr;
257 struct rte_mbuf *pkt = item->inner_item.firstseg;
258 uint16_t len;
259 uint16_t frag_offset;
260
261 /* Update the outer IPv4 header. */
262 len = pkt->pkt_len - pkt->outer_l2_len;
263 ipv4_hdr = (struct rte_ipv4_hdr *)(rte_pktmbuf_mtod(pkt, char *) +
264 pkt->outer_l2_len);
265 ipv4_hdr->total_length = rte_cpu_to_be_16(len);
266
267 /* Update the outer UDP header. */
268 len -= pkt->outer_l3_len;
269 udp_hdr = (struct rte_udp_hdr *)((char *)ipv4_hdr + pkt->outer_l3_len);
270 udp_hdr->dgram_len = rte_cpu_to_be_16(len);
271
272 /* Update the inner IPv4 header. */
273 len -= pkt->l2_len;
274 ipv4_hdr = (struct rte_ipv4_hdr *)((char *)udp_hdr + pkt->l2_len);
275 ipv4_hdr->total_length = rte_cpu_to_be_16(len);
276
277 /* Clear MF bit if it is last fragment */
278 if (item->inner_item.is_last_frag) {
279 frag_offset = rte_be_to_cpu_16(ipv4_hdr->fragment_offset);
280 ipv4_hdr->fragment_offset =
281 rte_cpu_to_be_16(frag_offset & ~RTE_IPV4_HDR_MF_FLAG);
282 }
283 }
284
285 int32_t
gro_vxlan_udp4_reassemble(struct rte_mbuf * pkt,struct gro_vxlan_udp4_tbl * tbl,uint64_t start_time)286 gro_vxlan_udp4_reassemble(struct rte_mbuf *pkt,
287 struct gro_vxlan_udp4_tbl *tbl,
288 uint64_t start_time)
289 {
290 struct rte_ether_hdr *outer_eth_hdr, *eth_hdr;
291 struct rte_ipv4_hdr *outer_ipv4_hdr, *ipv4_hdr;
292 struct rte_udp_hdr *udp_hdr;
293 struct rte_vxlan_hdr *vxlan_hdr;
294 uint16_t frag_offset;
295 uint8_t is_last_frag;
296 int16_t ip_dl;
297 uint16_t ip_id;
298
299 struct vxlan_udp4_flow_key key;
300 uint32_t cur_idx, prev_idx, item_idx;
301 uint32_t i, max_flow_num, remaining_flow_num;
302 int cmp;
303 uint16_t hdr_len;
304 uint8_t find;
305
306 outer_eth_hdr = rte_pktmbuf_mtod(pkt, struct rte_ether_hdr *);
307 outer_ipv4_hdr = (struct rte_ipv4_hdr *)((char *)outer_eth_hdr +
308 pkt->outer_l2_len);
309
310 udp_hdr = (struct rte_udp_hdr *)((char *)outer_ipv4_hdr +
311 pkt->outer_l3_len);
312 vxlan_hdr = (struct rte_vxlan_hdr *)((char *)udp_hdr +
313 sizeof(struct rte_udp_hdr));
314 eth_hdr = (struct rte_ether_hdr *)((char *)vxlan_hdr +
315 sizeof(struct rte_vxlan_hdr));
316 /* l2_len = outer udp hdr len + vxlan hdr len + inner l2 len */
317 ipv4_hdr = (struct rte_ipv4_hdr *)((char *)udp_hdr + pkt->l2_len);
318
319 /*
320 * Don't process the packet which has non-fragment inner IP.
321 */
322 if (!is_ipv4_fragment(ipv4_hdr))
323 return -1;
324
325 hdr_len = pkt->outer_l2_len + pkt->outer_l3_len + pkt->l2_len +
326 pkt->l3_len;
327 /*
328 * Don't process the packet whose payload length is less than or
329 * equal to 0.
330 */
331 if (pkt->pkt_len <= hdr_len)
332 return -1;
333
334 ip_dl = pkt->pkt_len - hdr_len;
335
336 ip_id = rte_be_to_cpu_16(ipv4_hdr->packet_id);
337 frag_offset = rte_be_to_cpu_16(ipv4_hdr->fragment_offset);
338 is_last_frag = ((frag_offset & RTE_IPV4_HDR_MF_FLAG) == 0) ? 1 : 0;
339 frag_offset = (uint16_t)(frag_offset & RTE_IPV4_HDR_OFFSET_MASK) << 3;
340
341 rte_ether_addr_copy(&(eth_hdr->s_addr), &(key.inner_key.eth_saddr));
342 rte_ether_addr_copy(&(eth_hdr->d_addr), &(key.inner_key.eth_daddr));
343 key.inner_key.ip_src_addr = ipv4_hdr->src_addr;
344 key.inner_key.ip_dst_addr = ipv4_hdr->dst_addr;
345 key.inner_key.ip_id = ip_id;
346
347 key.vxlan_hdr.vx_flags = vxlan_hdr->vx_flags;
348 key.vxlan_hdr.vx_vni = vxlan_hdr->vx_vni;
349 rte_ether_addr_copy(&(outer_eth_hdr->s_addr), &(key.outer_eth_saddr));
350 rte_ether_addr_copy(&(outer_eth_hdr->d_addr), &(key.outer_eth_daddr));
351 key.outer_ip_src_addr = outer_ipv4_hdr->src_addr;
352 key.outer_ip_dst_addr = outer_ipv4_hdr->dst_addr;
353 /* Note: It is unnecessary to save outer_src_port here because it can
354 * be different for VxLAN UDP fragments from the same flow.
355 */
356 key.outer_dst_port = udp_hdr->dst_port;
357
358 /* Search for a matched flow. */
359 max_flow_num = tbl->max_flow_num;
360 remaining_flow_num = tbl->flow_num;
361 find = 0;
362 for (i = 0; i < max_flow_num && remaining_flow_num; i++) {
363 if (tbl->flows[i].start_index != INVALID_ARRAY_INDEX) {
364 if (is_same_vxlan_udp4_flow(tbl->flows[i].key, key)) {
365 find = 1;
366 break;
367 }
368 remaining_flow_num--;
369 }
370 }
371
372 /*
373 * Can't find a matched flow. Insert a new flow and store the
374 * packet into the flow.
375 */
376 if (find == 0) {
377 item_idx = insert_new_item(tbl, pkt, start_time,
378 INVALID_ARRAY_INDEX, frag_offset,
379 is_last_frag);
380 if (unlikely(item_idx == INVALID_ARRAY_INDEX))
381 return -1;
382 if (insert_new_flow(tbl, &key, item_idx) ==
383 INVALID_ARRAY_INDEX) {
384 /*
385 * Fail to insert a new flow, so
386 * delete the inserted packet.
387 */
388 delete_item(tbl, item_idx, INVALID_ARRAY_INDEX);
389 return -1;
390 }
391 return 0;
392 }
393
394 /* Check all packets in the flow and try to find a neighbor. */
395 cur_idx = tbl->flows[i].start_index;
396 prev_idx = cur_idx;
397 do {
398 cmp = udp4_check_vxlan_neighbor(&(tbl->items[cur_idx]),
399 frag_offset, ip_dl);
400 if (cmp) {
401 if (merge_two_vxlan_udp4_packets(
402 &(tbl->items[cur_idx]),
403 pkt, cmp, frag_offset,
404 is_last_frag)) {
405 return 1;
406 }
407 /*
408 * Can't merge two packets, as the packet
409 * length will be greater than the max value.
410 * Insert the packet into the flow.
411 */
412 if (insert_new_item(tbl, pkt, start_time, prev_idx,
413 frag_offset, is_last_frag) ==
414 INVALID_ARRAY_INDEX)
415 return -1;
416 return 0;
417 }
418
419 /* Ensure inserted items are ordered by frag_offset */
420 if (frag_offset
421 < tbl->items[cur_idx].inner_item.frag_offset) {
422 break;
423 }
424
425 prev_idx = cur_idx;
426 cur_idx = tbl->items[cur_idx].inner_item.next_pkt_idx;
427 } while (cur_idx != INVALID_ARRAY_INDEX);
428
429 /* Can't find neighbor. Insert the packet into the flow. */
430 if (cur_idx == tbl->flows[i].start_index) {
431 /* Insert it before the first packet of the flow */
432 item_idx = insert_new_item(tbl, pkt, start_time,
433 INVALID_ARRAY_INDEX, frag_offset,
434 is_last_frag);
435 if (unlikely(item_idx == INVALID_ARRAY_INDEX))
436 return -1;
437 tbl->items[item_idx].inner_item.next_pkt_idx = cur_idx;
438 tbl->flows[i].start_index = item_idx;
439 } else {
440 if (insert_new_item(tbl, pkt, start_time, prev_idx,
441 frag_offset, is_last_frag
442 ) == INVALID_ARRAY_INDEX)
443 return -1;
444 }
445
446 return 0;
447 }
448
449 static int
gro_vxlan_udp4_merge_items(struct gro_vxlan_udp4_tbl * tbl,uint32_t start_idx)450 gro_vxlan_udp4_merge_items(struct gro_vxlan_udp4_tbl *tbl,
451 uint32_t start_idx)
452 {
453 uint16_t frag_offset;
454 uint8_t is_last_frag;
455 int16_t ip_dl;
456 struct rte_mbuf *pkt;
457 int cmp;
458 uint32_t item_idx;
459 uint16_t hdr_len;
460
461 item_idx = tbl->items[start_idx].inner_item.next_pkt_idx;
462 while (item_idx != INVALID_ARRAY_INDEX) {
463 pkt = tbl->items[item_idx].inner_item.firstseg;
464 hdr_len = pkt->outer_l2_len + pkt->outer_l3_len + pkt->l2_len +
465 pkt->l3_len;
466 ip_dl = pkt->pkt_len - hdr_len;
467 frag_offset = tbl->items[item_idx].inner_item.frag_offset;
468 is_last_frag = tbl->items[item_idx].inner_item.is_last_frag;
469 cmp = udp4_check_vxlan_neighbor(&(tbl->items[start_idx]),
470 frag_offset, ip_dl);
471 if (cmp) {
472 if (merge_two_vxlan_udp4_packets(
473 &(tbl->items[start_idx]),
474 pkt, cmp, frag_offset,
475 is_last_frag)) {
476 item_idx = delete_item(tbl, item_idx,
477 INVALID_ARRAY_INDEX);
478 tbl->items[start_idx].inner_item.next_pkt_idx
479 = item_idx;
480 } else
481 return 0;
482 } else
483 return 0;
484 }
485
486 return 0;
487 }
488
489 uint16_t
gro_vxlan_udp4_tbl_timeout_flush(struct gro_vxlan_udp4_tbl * tbl,uint64_t flush_timestamp,struct rte_mbuf ** out,uint16_t nb_out)490 gro_vxlan_udp4_tbl_timeout_flush(struct gro_vxlan_udp4_tbl *tbl,
491 uint64_t flush_timestamp,
492 struct rte_mbuf **out,
493 uint16_t nb_out)
494 {
495 uint16_t k = 0;
496 uint32_t i, j;
497 uint32_t max_flow_num = tbl->max_flow_num;
498
499 for (i = 0; i < max_flow_num; i++) {
500 if (unlikely(tbl->flow_num == 0))
501 return k;
502
503 j = tbl->flows[i].start_index;
504 while (j != INVALID_ARRAY_INDEX) {
505 if (tbl->items[j].inner_item.start_time <=
506 flush_timestamp) {
507 gro_vxlan_udp4_merge_items(tbl, j);
508 out[k++] = tbl->items[j].inner_item.firstseg;
509 if (tbl->items[j].inner_item.nb_merged > 1)
510 update_vxlan_header(&(tbl->items[j]));
511 /*
512 * Delete the item and get the next packet
513 * index.
514 */
515 j = delete_item(tbl, j, INVALID_ARRAY_INDEX);
516 tbl->flows[i].start_index = j;
517 if (j == INVALID_ARRAY_INDEX)
518 tbl->flow_num--;
519
520 if (unlikely(k == nb_out))
521 return k;
522 } else
523 /*
524 * Flushing packets does not strictly follow
525 * timestamp. It does not flush left packets of
526 * the flow this time once it finds one item
527 * whose start_time is greater than
528 * flush_timestamp. So go to check other flows.
529 */
530 break;
531 }
532 }
533 return k;
534 }
535
536 uint32_t
gro_vxlan_udp4_tbl_pkt_count(void * tbl)537 gro_vxlan_udp4_tbl_pkt_count(void *tbl)
538 {
539 struct gro_vxlan_udp4_tbl *gro_tbl = tbl;
540
541 if (gro_tbl)
542 return gro_tbl->item_num;
543
544 return 0;
545 }
546