1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3 * Driver for Solarflare network controllers and boards
4 * Copyright 2023, Advanced Micro Devices, Inc.
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation, incorporated herein by reference.
9 */
10
11 #include "tc_conntrack.h"
12 #include "tc.h"
13 #include "mae.h"
14
15 static int efx_tc_flow_block(enum tc_setup_type type, void *type_data,
16 void *cb_priv);
17
18 static const struct rhashtable_params efx_tc_ct_zone_ht_params = {
19 .key_len = offsetof(struct efx_tc_ct_zone, linkage),
20 .key_offset = 0,
21 .head_offset = offsetof(struct efx_tc_ct_zone, linkage),
22 };
23
24 static const struct rhashtable_params efx_tc_ct_ht_params = {
25 .key_len = offsetof(struct efx_tc_ct_entry, linkage),
26 .key_offset = 0,
27 .head_offset = offsetof(struct efx_tc_ct_entry, linkage),
28 };
29
efx_tc_ct_zone_free(void * ptr,void * arg)30 static void efx_tc_ct_zone_free(void *ptr, void *arg)
31 {
32 struct efx_tc_ct_zone *zone = ptr;
33 struct efx_nic *efx = zone->efx;
34
35 netif_err(efx, drv, efx->net_dev,
36 "tc ct_zone %u still present at teardown, removing\n",
37 zone->zone);
38
39 nf_flow_table_offload_del_cb(zone->nf_ft, efx_tc_flow_block, zone);
40 kfree(zone);
41 }
42
efx_tc_ct_free(void * ptr,void * arg)43 static void efx_tc_ct_free(void *ptr, void *arg)
44 {
45 struct efx_tc_ct_entry *conn = ptr;
46 struct efx_nic *efx = arg;
47
48 netif_err(efx, drv, efx->net_dev,
49 "tc ct_entry %lx still present at teardown\n",
50 conn->cookie);
51
52 /* We can release the counter, but we can't remove the CT itself
53 * from hardware because the table meta is already gone.
54 */
55 efx_tc_flower_release_counter(efx, conn->cnt);
56 kfree(conn);
57 }
58
efx_tc_init_conntrack(struct efx_nic * efx)59 int efx_tc_init_conntrack(struct efx_nic *efx)
60 {
61 int rc;
62
63 rc = rhashtable_init(&efx->tc->ct_zone_ht, &efx_tc_ct_zone_ht_params);
64 if (rc < 0)
65 goto fail_ct_zone_ht;
66 rc = rhashtable_init(&efx->tc->ct_ht, &efx_tc_ct_ht_params);
67 if (rc < 0)
68 goto fail_ct_ht;
69 return 0;
70 fail_ct_ht:
71 rhashtable_destroy(&efx->tc->ct_zone_ht);
72 fail_ct_zone_ht:
73 return rc;
74 }
75
76 /* Only call this in init failure teardown.
77 * Normal exit should fini instead as there may be entries in the table.
78 */
efx_tc_destroy_conntrack(struct efx_nic * efx)79 void efx_tc_destroy_conntrack(struct efx_nic *efx)
80 {
81 rhashtable_destroy(&efx->tc->ct_ht);
82 rhashtable_destroy(&efx->tc->ct_zone_ht);
83 }
84
efx_tc_fini_conntrack(struct efx_nic * efx)85 void efx_tc_fini_conntrack(struct efx_nic *efx)
86 {
87 rhashtable_free_and_destroy(&efx->tc->ct_zone_ht, efx_tc_ct_zone_free, NULL);
88 rhashtable_free_and_destroy(&efx->tc->ct_ht, efx_tc_ct_free, efx);
89 }
90
91 #define EFX_NF_TCP_FLAG(flg) cpu_to_be16(be32_to_cpu(TCP_FLAG_##flg) >> 16)
92
efx_tc_ct_parse_match(struct efx_nic * efx,struct flow_rule * fr,struct efx_tc_ct_entry * conn)93 static int efx_tc_ct_parse_match(struct efx_nic *efx, struct flow_rule *fr,
94 struct efx_tc_ct_entry *conn)
95 {
96 struct flow_dissector *dissector = fr->match.dissector;
97 unsigned char ipv = 0;
98 bool tcp = false;
99
100 if (flow_rule_match_key(fr, FLOW_DISSECTOR_KEY_CONTROL)) {
101 struct flow_match_control fm;
102
103 flow_rule_match_control(fr, &fm);
104 if (IS_ALL_ONES(fm.mask->addr_type))
105 switch (fm.key->addr_type) {
106 case FLOW_DISSECTOR_KEY_IPV4_ADDRS:
107 ipv = 4;
108 break;
109 case FLOW_DISSECTOR_KEY_IPV6_ADDRS:
110 ipv = 6;
111 break;
112 default:
113 break;
114 }
115 }
116
117 if (!ipv) {
118 netif_dbg(efx, drv, efx->net_dev,
119 "Conntrack missing ipv specification\n");
120 return -EOPNOTSUPP;
121 }
122
123 if (dissector->used_keys &
124 ~(BIT_ULL(FLOW_DISSECTOR_KEY_CONTROL) |
125 BIT_ULL(FLOW_DISSECTOR_KEY_BASIC) |
126 BIT_ULL(FLOW_DISSECTOR_KEY_IPV4_ADDRS) |
127 BIT_ULL(FLOW_DISSECTOR_KEY_IPV6_ADDRS) |
128 BIT_ULL(FLOW_DISSECTOR_KEY_PORTS) |
129 BIT_ULL(FLOW_DISSECTOR_KEY_TCP) |
130 BIT_ULL(FLOW_DISSECTOR_KEY_META))) {
131 netif_dbg(efx, drv, efx->net_dev,
132 "Unsupported conntrack keys %#llx\n",
133 dissector->used_keys);
134 return -EOPNOTSUPP;
135 }
136
137 if (flow_rule_match_key(fr, FLOW_DISSECTOR_KEY_BASIC)) {
138 struct flow_match_basic fm;
139
140 flow_rule_match_basic(fr, &fm);
141 if (!IS_ALL_ONES(fm.mask->n_proto)) {
142 netif_dbg(efx, drv, efx->net_dev,
143 "Conntrack eth_proto is not exact-match; mask %04x\n",
144 ntohs(fm.mask->n_proto));
145 return -EOPNOTSUPP;
146 }
147 conn->eth_proto = fm.key->n_proto;
148 if (conn->eth_proto != (ipv == 4 ? htons(ETH_P_IP)
149 : htons(ETH_P_IPV6))) {
150 netif_dbg(efx, drv, efx->net_dev,
151 "Conntrack eth_proto is not IPv%u, is %04x\n",
152 ipv, ntohs(conn->eth_proto));
153 return -EOPNOTSUPP;
154 }
155 if (!IS_ALL_ONES(fm.mask->ip_proto)) {
156 netif_dbg(efx, drv, efx->net_dev,
157 "Conntrack ip_proto is not exact-match; mask %02x\n",
158 fm.mask->ip_proto);
159 return -EOPNOTSUPP;
160 }
161 conn->ip_proto = fm.key->ip_proto;
162 switch (conn->ip_proto) {
163 case IPPROTO_TCP:
164 tcp = true;
165 break;
166 case IPPROTO_UDP:
167 break;
168 default:
169 netif_dbg(efx, drv, efx->net_dev,
170 "Conntrack ip_proto not TCP or UDP, is %02x\n",
171 conn->ip_proto);
172 return -EOPNOTSUPP;
173 }
174 } else {
175 netif_dbg(efx, drv, efx->net_dev,
176 "Conntrack missing eth_proto, ip_proto\n");
177 return -EOPNOTSUPP;
178 }
179
180 if (ipv == 4 && flow_rule_match_key(fr, FLOW_DISSECTOR_KEY_IPV4_ADDRS)) {
181 struct flow_match_ipv4_addrs fm;
182
183 flow_rule_match_ipv4_addrs(fr, &fm);
184 if (!IS_ALL_ONES(fm.mask->src)) {
185 netif_dbg(efx, drv, efx->net_dev,
186 "Conntrack ipv4.src is not exact-match; mask %08x\n",
187 ntohl(fm.mask->src));
188 return -EOPNOTSUPP;
189 }
190 conn->src_ip = fm.key->src;
191 if (!IS_ALL_ONES(fm.mask->dst)) {
192 netif_dbg(efx, drv, efx->net_dev,
193 "Conntrack ipv4.dst is not exact-match; mask %08x\n",
194 ntohl(fm.mask->dst));
195 return -EOPNOTSUPP;
196 }
197 conn->dst_ip = fm.key->dst;
198 } else if (ipv == 6 && flow_rule_match_key(fr, FLOW_DISSECTOR_KEY_IPV6_ADDRS)) {
199 struct flow_match_ipv6_addrs fm;
200
201 flow_rule_match_ipv6_addrs(fr, &fm);
202 if (!efx_ipv6_addr_all_ones(&fm.mask->src)) {
203 netif_dbg(efx, drv, efx->net_dev,
204 "Conntrack ipv6.src is not exact-match; mask %pI6\n",
205 &fm.mask->src);
206 return -EOPNOTSUPP;
207 }
208 conn->src_ip6 = fm.key->src;
209 if (!efx_ipv6_addr_all_ones(&fm.mask->dst)) {
210 netif_dbg(efx, drv, efx->net_dev,
211 "Conntrack ipv6.dst is not exact-match; mask %pI6\n",
212 &fm.mask->dst);
213 return -EOPNOTSUPP;
214 }
215 conn->dst_ip6 = fm.key->dst;
216 } else {
217 netif_dbg(efx, drv, efx->net_dev,
218 "Conntrack missing IPv%u addrs\n", ipv);
219 return -EOPNOTSUPP;
220 }
221
222 if (flow_rule_match_key(fr, FLOW_DISSECTOR_KEY_PORTS)) {
223 struct flow_match_ports fm;
224
225 flow_rule_match_ports(fr, &fm);
226 if (!IS_ALL_ONES(fm.mask->src)) {
227 netif_dbg(efx, drv, efx->net_dev,
228 "Conntrack ports.src is not exact-match; mask %04x\n",
229 ntohs(fm.mask->src));
230 return -EOPNOTSUPP;
231 }
232 conn->l4_sport = fm.key->src;
233 if (!IS_ALL_ONES(fm.mask->dst)) {
234 netif_dbg(efx, drv, efx->net_dev,
235 "Conntrack ports.dst is not exact-match; mask %04x\n",
236 ntohs(fm.mask->dst));
237 return -EOPNOTSUPP;
238 }
239 conn->l4_dport = fm.key->dst;
240 } else {
241 netif_dbg(efx, drv, efx->net_dev, "Conntrack missing L4 ports\n");
242 return -EOPNOTSUPP;
243 }
244
245 if (flow_rule_match_key(fr, FLOW_DISSECTOR_KEY_TCP)) {
246 __be16 tcp_interesting_flags;
247 struct flow_match_tcp fm;
248
249 if (!tcp) {
250 netif_dbg(efx, drv, efx->net_dev,
251 "Conntrack matching on TCP keys but ipproto is not tcp\n");
252 return -EOPNOTSUPP;
253 }
254 flow_rule_match_tcp(fr, &fm);
255 tcp_interesting_flags = EFX_NF_TCP_FLAG(SYN) |
256 EFX_NF_TCP_FLAG(RST) |
257 EFX_NF_TCP_FLAG(FIN);
258 /* If any of the tcp_interesting_flags is set, we always
259 * inhibit CT lookup in LHS (so SW can update CT table).
260 */
261 if (fm.key->flags & tcp_interesting_flags) {
262 netif_dbg(efx, drv, efx->net_dev,
263 "Unsupported conntrack tcp.flags %04x/%04x\n",
264 ntohs(fm.key->flags), ntohs(fm.mask->flags));
265 return -EOPNOTSUPP;
266 }
267 /* Other TCP flags cannot be filtered at CT */
268 if (fm.mask->flags & ~tcp_interesting_flags) {
269 netif_dbg(efx, drv, efx->net_dev,
270 "Unsupported conntrack tcp.flags %04x/%04x\n",
271 ntohs(fm.key->flags), ntohs(fm.mask->flags));
272 return -EOPNOTSUPP;
273 }
274 }
275
276 return 0;
277 }
278
efx_tc_ct_replace(struct efx_tc_ct_zone * ct_zone,struct flow_cls_offload * tc)279 static int efx_tc_ct_replace(struct efx_tc_ct_zone *ct_zone,
280 struct flow_cls_offload *tc)
281 {
282 struct flow_rule *fr = flow_cls_offload_flow_rule(tc);
283 struct efx_tc_ct_entry *conn, *old;
284 struct efx_nic *efx = ct_zone->efx;
285 const struct flow_action_entry *fa;
286 struct efx_tc_counter *cnt;
287 int rc, i;
288
289 if (WARN_ON(!efx->tc))
290 return -ENETDOWN;
291 if (WARN_ON(!efx->tc->up))
292 return -ENETDOWN;
293
294 conn = kzalloc(sizeof(*conn), GFP_USER);
295 if (!conn)
296 return -ENOMEM;
297 conn->cookie = tc->cookie;
298 old = rhashtable_lookup_get_insert_fast(&efx->tc->ct_ht,
299 &conn->linkage,
300 efx_tc_ct_ht_params);
301 if (IS_ERR(old)) {
302 rc = PTR_ERR(old);
303 goto release;
304 } else if (old) {
305 netif_dbg(efx, drv, efx->net_dev,
306 "Already offloaded conntrack (cookie %lx)\n", tc->cookie);
307 rc = -EEXIST;
308 goto release;
309 }
310
311 /* Parse match */
312 conn->zone = ct_zone;
313 rc = efx_tc_ct_parse_match(efx, fr, conn);
314 if (rc)
315 goto release;
316
317 /* Parse actions */
318 flow_action_for_each(i, fa, &fr->action) {
319 switch (fa->id) {
320 case FLOW_ACTION_CT_METADATA:
321 conn->mark = fa->ct_metadata.mark;
322 if (memchr_inv(fa->ct_metadata.labels, 0, sizeof(fa->ct_metadata.labels))) {
323 netif_dbg(efx, drv, efx->net_dev,
324 "Setting CT label not supported\n");
325 rc = -EOPNOTSUPP;
326 goto release;
327 }
328 break;
329 default:
330 netif_dbg(efx, drv, efx->net_dev,
331 "Unhandled action %u for conntrack\n", fa->id);
332 rc = -EOPNOTSUPP;
333 goto release;
334 }
335 }
336
337 /* fill in defaults for unmangled values */
338 conn->nat_ip = conn->dnat ? conn->dst_ip : conn->src_ip;
339 conn->l4_natport = conn->dnat ? conn->l4_dport : conn->l4_sport;
340
341 cnt = efx_tc_flower_allocate_counter(efx, EFX_TC_COUNTER_TYPE_CT);
342 if (IS_ERR(cnt)) {
343 rc = PTR_ERR(cnt);
344 goto release;
345 }
346 conn->cnt = cnt;
347
348 rc = efx_mae_insert_ct(efx, conn);
349 if (rc) {
350 netif_dbg(efx, drv, efx->net_dev,
351 "Failed to insert conntrack, %d\n", rc);
352 goto release;
353 }
354 mutex_lock(&ct_zone->mutex);
355 list_add_tail(&conn->list, &ct_zone->cts);
356 mutex_unlock(&ct_zone->mutex);
357 return 0;
358 release:
359 if (conn->cnt)
360 efx_tc_flower_release_counter(efx, conn->cnt);
361 if (!old)
362 rhashtable_remove_fast(&efx->tc->ct_ht, &conn->linkage,
363 efx_tc_ct_ht_params);
364 kfree(conn);
365 return rc;
366 }
367
368 /* Caller must follow with efx_tc_ct_remove_finish() after RCU grace period! */
efx_tc_ct_remove(struct efx_nic * efx,struct efx_tc_ct_entry * conn)369 static void efx_tc_ct_remove(struct efx_nic *efx, struct efx_tc_ct_entry *conn)
370 {
371 int rc;
372
373 /* Remove it from HW */
374 rc = efx_mae_remove_ct(efx, conn);
375 /* Delete it from SW */
376 rhashtable_remove_fast(&efx->tc->ct_ht, &conn->linkage,
377 efx_tc_ct_ht_params);
378 if (rc) {
379 netif_err(efx, drv, efx->net_dev,
380 "Failed to remove conntrack %lx from hw, rc %d\n",
381 conn->cookie, rc);
382 } else {
383 netif_dbg(efx, drv, efx->net_dev, "Removed conntrack %lx\n",
384 conn->cookie);
385 }
386 }
387
efx_tc_ct_remove_finish(struct efx_nic * efx,struct efx_tc_ct_entry * conn)388 static void efx_tc_ct_remove_finish(struct efx_nic *efx, struct efx_tc_ct_entry *conn)
389 {
390 /* Remove related CT counter. This is delayed after the conn object we
391 * are working with has been successfully removed. This protects the
392 * counter from being used-after-free inside efx_tc_ct_stats.
393 */
394 efx_tc_flower_release_counter(efx, conn->cnt);
395 kfree(conn);
396 }
397
efx_tc_ct_destroy(struct efx_tc_ct_zone * ct_zone,struct flow_cls_offload * tc)398 static int efx_tc_ct_destroy(struct efx_tc_ct_zone *ct_zone,
399 struct flow_cls_offload *tc)
400 {
401 struct efx_nic *efx = ct_zone->efx;
402 struct efx_tc_ct_entry *conn;
403
404 conn = rhashtable_lookup_fast(&efx->tc->ct_ht, &tc->cookie,
405 efx_tc_ct_ht_params);
406 if (!conn) {
407 netif_warn(efx, drv, efx->net_dev,
408 "Conntrack %lx not found to remove\n", tc->cookie);
409 return -ENOENT;
410 }
411
412 mutex_lock(&ct_zone->mutex);
413 list_del(&conn->list);
414 efx_tc_ct_remove(efx, conn);
415 mutex_unlock(&ct_zone->mutex);
416 synchronize_rcu();
417 efx_tc_ct_remove_finish(efx, conn);
418 return 0;
419 }
420
efx_tc_ct_stats(struct efx_tc_ct_zone * ct_zone,struct flow_cls_offload * tc)421 static int efx_tc_ct_stats(struct efx_tc_ct_zone *ct_zone,
422 struct flow_cls_offload *tc)
423 {
424 struct efx_nic *efx = ct_zone->efx;
425 struct efx_tc_ct_entry *conn;
426 struct efx_tc_counter *cnt;
427
428 rcu_read_lock();
429 conn = rhashtable_lookup_fast(&efx->tc->ct_ht, &tc->cookie,
430 efx_tc_ct_ht_params);
431 if (!conn) {
432 netif_warn(efx, drv, efx->net_dev,
433 "Conntrack %lx not found for stats\n", tc->cookie);
434 rcu_read_unlock();
435 return -ENOENT;
436 }
437
438 cnt = conn->cnt;
439 spin_lock_bh(&cnt->lock);
440 /* Report only last use */
441 flow_stats_update(&tc->stats, 0, 0, 0, cnt->touched,
442 FLOW_ACTION_HW_STATS_DELAYED);
443 spin_unlock_bh(&cnt->lock);
444 rcu_read_unlock();
445
446 return 0;
447 }
448
efx_tc_flow_block(enum tc_setup_type type,void * type_data,void * cb_priv)449 static int efx_tc_flow_block(enum tc_setup_type type, void *type_data,
450 void *cb_priv)
451 {
452 struct flow_cls_offload *tcb = type_data;
453 struct efx_tc_ct_zone *ct_zone = cb_priv;
454
455 if (type != TC_SETUP_CLSFLOWER)
456 return -EOPNOTSUPP;
457
458 switch (tcb->command) {
459 case FLOW_CLS_REPLACE:
460 return efx_tc_ct_replace(ct_zone, tcb);
461 case FLOW_CLS_DESTROY:
462 return efx_tc_ct_destroy(ct_zone, tcb);
463 case FLOW_CLS_STATS:
464 return efx_tc_ct_stats(ct_zone, tcb);
465 default:
466 break;
467 }
468
469 return -EOPNOTSUPP;
470 }
471
efx_tc_ct_register_zone(struct efx_nic * efx,u16 zone,struct nf_flowtable * ct_ft)472 struct efx_tc_ct_zone *efx_tc_ct_register_zone(struct efx_nic *efx, u16 zone,
473 struct nf_flowtable *ct_ft)
474 {
475 struct efx_tc_ct_zone *ct_zone, *old;
476 int rc;
477
478 ct_zone = kzalloc(sizeof(*ct_zone), GFP_USER);
479 if (!ct_zone)
480 return ERR_PTR(-ENOMEM);
481 ct_zone->zone = zone;
482 old = rhashtable_lookup_get_insert_fast(&efx->tc->ct_zone_ht,
483 &ct_zone->linkage,
484 efx_tc_ct_zone_ht_params);
485 if (old) {
486 /* don't need our new entry */
487 kfree(ct_zone);
488 if (IS_ERR(old)) /* oh dear, it's actually an error */
489 return ERR_CAST(old);
490 if (!refcount_inc_not_zero(&old->ref))
491 return ERR_PTR(-EAGAIN);
492 /* existing entry found */
493 WARN_ON_ONCE(old->nf_ft != ct_ft);
494 netif_dbg(efx, drv, efx->net_dev,
495 "Found existing ct_zone for %u\n", zone);
496 return old;
497 }
498 ct_zone->nf_ft = ct_ft;
499 ct_zone->efx = efx;
500 INIT_LIST_HEAD(&ct_zone->cts);
501 mutex_init(&ct_zone->mutex);
502 rc = nf_flow_table_offload_add_cb(ct_ft, efx_tc_flow_block, ct_zone);
503 netif_dbg(efx, drv, efx->net_dev, "Adding new ct_zone for %u, rc %d\n",
504 zone, rc);
505 if (rc < 0)
506 goto fail;
507 refcount_set(&ct_zone->ref, 1);
508 return ct_zone;
509 fail:
510 rhashtable_remove_fast(&efx->tc->ct_zone_ht, &ct_zone->linkage,
511 efx_tc_ct_zone_ht_params);
512 kfree(ct_zone);
513 return ERR_PTR(rc);
514 }
515
efx_tc_ct_unregister_zone(struct efx_nic * efx,struct efx_tc_ct_zone * ct_zone)516 void efx_tc_ct_unregister_zone(struct efx_nic *efx,
517 struct efx_tc_ct_zone *ct_zone)
518 {
519 struct efx_tc_ct_entry *conn, *next;
520
521 if (!refcount_dec_and_test(&ct_zone->ref))
522 return; /* still in use */
523 nf_flow_table_offload_del_cb(ct_zone->nf_ft, efx_tc_flow_block, ct_zone);
524 rhashtable_remove_fast(&efx->tc->ct_zone_ht, &ct_zone->linkage,
525 efx_tc_ct_zone_ht_params);
526 mutex_lock(&ct_zone->mutex);
527 list_for_each_entry(conn, &ct_zone->cts, list)
528 efx_tc_ct_remove(efx, conn);
529 synchronize_rcu();
530 /* need to use _safe because efx_tc_ct_remove_finish() frees conn */
531 list_for_each_entry_safe(conn, next, &ct_zone->cts, list)
532 efx_tc_ct_remove_finish(efx, conn);
533 mutex_unlock(&ct_zone->mutex);
534 mutex_destroy(&ct_zone->mutex);
535 netif_dbg(efx, drv, efx->net_dev, "Removed ct_zone for %u\n",
536 ct_zone->zone);
537 kfree(ct_zone);
538 }
539