Loading...
1/*
2 * drivers/net/team/team_mode_loadbalance.c - Load-balancing mode for team
3 * Copyright (c) 2012 Jiri Pirko <jpirko@redhat.com>
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 */
10
11#include <linux/kernel.h>
12#include <linux/types.h>
13#include <linux/module.h>
14#include <linux/init.h>
15#include <linux/errno.h>
16#include <linux/netdevice.h>
17#include <linux/filter.h>
18#include <linux/if_team.h>
19
20struct lb_priv;
21
22typedef struct team_port *lb_select_tx_port_func_t(struct team *,
23 struct lb_priv *,
24 struct sk_buff *,
25 unsigned char);
26
27#define LB_TX_HASHTABLE_SIZE 256 /* hash is a char */
28
29struct lb_stats {
30 u64 tx_bytes;
31};
32
33struct lb_pcpu_stats {
34 struct lb_stats hash_stats[LB_TX_HASHTABLE_SIZE];
35 struct u64_stats_sync syncp;
36};
37
38struct lb_stats_info {
39 struct lb_stats stats;
40 struct lb_stats last_stats;
41 struct team_option_inst_info *opt_inst_info;
42};
43
44struct lb_port_mapping {
45 struct team_port __rcu *port;
46 struct team_option_inst_info *opt_inst_info;
47};
48
49struct lb_priv_ex {
50 struct team *team;
51 struct lb_port_mapping tx_hash_to_port_mapping[LB_TX_HASHTABLE_SIZE];
52 struct sock_fprog *orig_fprog;
53 struct {
54 unsigned int refresh_interval; /* in tenths of second */
55 struct delayed_work refresh_dw;
56 struct lb_stats_info info[LB_TX_HASHTABLE_SIZE];
57 } stats;
58};
59
60struct lb_priv {
61 struct sk_filter __rcu *fp;
62 lb_select_tx_port_func_t __rcu *select_tx_port_func;
63 struct lb_pcpu_stats __percpu *pcpu_stats;
64 struct lb_priv_ex *ex; /* priv extension */
65};
66
67static struct lb_priv *get_lb_priv(struct team *team)
68{
69 return (struct lb_priv *) &team->mode_priv;
70}
71
72struct lb_port_priv {
73 struct lb_stats __percpu *pcpu_stats;
74 struct lb_stats_info stats_info;
75};
76
77static struct lb_port_priv *get_lb_port_priv(struct team_port *port)
78{
79 return (struct lb_port_priv *) &port->mode_priv;
80}
81
82#define LB_HTPM_PORT_BY_HASH(lp_priv, hash) \
83 (lb_priv)->ex->tx_hash_to_port_mapping[hash].port
84
85#define LB_HTPM_OPT_INST_INFO_BY_HASH(lp_priv, hash) \
86 (lb_priv)->ex->tx_hash_to_port_mapping[hash].opt_inst_info
87
88static void lb_tx_hash_to_port_mapping_null_port(struct team *team,
89 struct team_port *port)
90{
91 struct lb_priv *lb_priv = get_lb_priv(team);
92 bool changed = false;
93 int i;
94
95 for (i = 0; i < LB_TX_HASHTABLE_SIZE; i++) {
96 struct lb_port_mapping *pm;
97
98 pm = &lb_priv->ex->tx_hash_to_port_mapping[i];
99 if (rcu_access_pointer(pm->port) == port) {
100 RCU_INIT_POINTER(pm->port, NULL);
101 team_option_inst_set_change(pm->opt_inst_info);
102 changed = true;
103 }
104 }
105 if (changed)
106 team_options_change_check(team);
107}
108
109/* Basic tx selection based solely by hash */
110static struct team_port *lb_hash_select_tx_port(struct team *team,
111 struct lb_priv *lb_priv,
112 struct sk_buff *skb,
113 unsigned char hash)
114{
115 int port_index = team_num_to_port_index(team, hash);
116
117 return team_get_port_by_index_rcu(team, port_index);
118}
119
120/* Hash to port mapping select tx port */
121static struct team_port *lb_htpm_select_tx_port(struct team *team,
122 struct lb_priv *lb_priv,
123 struct sk_buff *skb,
124 unsigned char hash)
125{
126 return rcu_dereference_bh(LB_HTPM_PORT_BY_HASH(lb_priv, hash));
127}
128
129struct lb_select_tx_port {
130 char *name;
131 lb_select_tx_port_func_t *func;
132};
133
134static const struct lb_select_tx_port lb_select_tx_port_list[] = {
135 {
136 .name = "hash",
137 .func = lb_hash_select_tx_port,
138 },
139 {
140 .name = "hash_to_port_mapping",
141 .func = lb_htpm_select_tx_port,
142 },
143};
144#define LB_SELECT_TX_PORT_LIST_COUNT ARRAY_SIZE(lb_select_tx_port_list)
145
146static char *lb_select_tx_port_get_name(lb_select_tx_port_func_t *func)
147{
148 int i;
149
150 for (i = 0; i < LB_SELECT_TX_PORT_LIST_COUNT; i++) {
151 const struct lb_select_tx_port *item;
152
153 item = &lb_select_tx_port_list[i];
154 if (item->func == func)
155 return item->name;
156 }
157 return NULL;
158}
159
160static lb_select_tx_port_func_t *lb_select_tx_port_get_func(const char *name)
161{
162 int i;
163
164 for (i = 0; i < LB_SELECT_TX_PORT_LIST_COUNT; i++) {
165 const struct lb_select_tx_port *item;
166
167 item = &lb_select_tx_port_list[i];
168 if (!strcmp(item->name, name))
169 return item->func;
170 }
171 return NULL;
172}
173
174static unsigned int lb_get_skb_hash(struct lb_priv *lb_priv,
175 struct sk_buff *skb)
176{
177 struct sk_filter *fp;
178 uint32_t lhash;
179 unsigned char *c;
180
181 fp = rcu_dereference_bh(lb_priv->fp);
182 if (unlikely(!fp))
183 return 0;
184 lhash = SK_RUN_FILTER(fp, skb);
185 c = (char *) &lhash;
186 return c[0] ^ c[1] ^ c[2] ^ c[3];
187}
188
189static void lb_update_tx_stats(unsigned int tx_bytes, struct lb_priv *lb_priv,
190 struct lb_port_priv *lb_port_priv,
191 unsigned char hash)
192{
193 struct lb_pcpu_stats *pcpu_stats;
194 struct lb_stats *port_stats;
195 struct lb_stats *hash_stats;
196
197 pcpu_stats = this_cpu_ptr(lb_priv->pcpu_stats);
198 port_stats = this_cpu_ptr(lb_port_priv->pcpu_stats);
199 hash_stats = &pcpu_stats->hash_stats[hash];
200 u64_stats_update_begin(&pcpu_stats->syncp);
201 port_stats->tx_bytes += tx_bytes;
202 hash_stats->tx_bytes += tx_bytes;
203 u64_stats_update_end(&pcpu_stats->syncp);
204}
205
206static bool lb_transmit(struct team *team, struct sk_buff *skb)
207{
208 struct lb_priv *lb_priv = get_lb_priv(team);
209 lb_select_tx_port_func_t *select_tx_port_func;
210 struct team_port *port;
211 unsigned char hash;
212 unsigned int tx_bytes = skb->len;
213
214 hash = lb_get_skb_hash(lb_priv, skb);
215 select_tx_port_func = rcu_dereference_bh(lb_priv->select_tx_port_func);
216 port = select_tx_port_func(team, lb_priv, skb, hash);
217 if (unlikely(!port))
218 goto drop;
219 if (team_dev_queue_xmit(team, port, skb))
220 return false;
221 lb_update_tx_stats(tx_bytes, lb_priv, get_lb_port_priv(port), hash);
222 return true;
223
224drop:
225 dev_kfree_skb_any(skb);
226 return false;
227}
228
229static int lb_bpf_func_get(struct team *team, struct team_gsetter_ctx *ctx)
230{
231 struct lb_priv *lb_priv = get_lb_priv(team);
232
233 if (!lb_priv->ex->orig_fprog) {
234 ctx->data.bin_val.len = 0;
235 ctx->data.bin_val.ptr = NULL;
236 return 0;
237 }
238 ctx->data.bin_val.len = lb_priv->ex->orig_fprog->len *
239 sizeof(struct sock_filter);
240 ctx->data.bin_val.ptr = lb_priv->ex->orig_fprog->filter;
241 return 0;
242}
243
244static int __fprog_create(struct sock_fprog **pfprog, u32 data_len,
245 const void *data)
246{
247 struct sock_fprog *fprog;
248 struct sock_filter *filter = (struct sock_filter *) data;
249
250 if (data_len % sizeof(struct sock_filter))
251 return -EINVAL;
252 fprog = kmalloc(sizeof(struct sock_fprog), GFP_KERNEL);
253 if (!fprog)
254 return -ENOMEM;
255 fprog->filter = kmemdup(filter, data_len, GFP_KERNEL);
256 if (!fprog->filter) {
257 kfree(fprog);
258 return -ENOMEM;
259 }
260 fprog->len = data_len / sizeof(struct sock_filter);
261 *pfprog = fprog;
262 return 0;
263}
264
265static void __fprog_destroy(struct sock_fprog *fprog)
266{
267 kfree(fprog->filter);
268 kfree(fprog);
269}
270
271static int lb_bpf_func_set(struct team *team, struct team_gsetter_ctx *ctx)
272{
273 struct lb_priv *lb_priv = get_lb_priv(team);
274 struct sk_filter *fp = NULL;
275 struct sk_filter *orig_fp;
276 struct sock_fprog *fprog = NULL;
277 int err;
278
279 if (ctx->data.bin_val.len) {
280 err = __fprog_create(&fprog, ctx->data.bin_val.len,
281 ctx->data.bin_val.ptr);
282 if (err)
283 return err;
284 err = sk_unattached_filter_create(&fp, fprog);
285 if (err) {
286 __fprog_destroy(fprog);
287 return err;
288 }
289 }
290
291 if (lb_priv->ex->orig_fprog) {
292 /* Clear old filter data */
293 __fprog_destroy(lb_priv->ex->orig_fprog);
294 orig_fp = rcu_dereference_protected(lb_priv->fp,
295 lockdep_is_held(&team->lock));
296 sk_unattached_filter_destroy(orig_fp);
297 }
298
299 rcu_assign_pointer(lb_priv->fp, fp);
300 lb_priv->ex->orig_fprog = fprog;
301 return 0;
302}
303
304static int lb_tx_method_get(struct team *team, struct team_gsetter_ctx *ctx)
305{
306 struct lb_priv *lb_priv = get_lb_priv(team);
307 lb_select_tx_port_func_t *func;
308 char *name;
309
310 func = rcu_dereference_protected(lb_priv->select_tx_port_func,
311 lockdep_is_held(&team->lock));
312 name = lb_select_tx_port_get_name(func);
313 BUG_ON(!name);
314 ctx->data.str_val = name;
315 return 0;
316}
317
318static int lb_tx_method_set(struct team *team, struct team_gsetter_ctx *ctx)
319{
320 struct lb_priv *lb_priv = get_lb_priv(team);
321 lb_select_tx_port_func_t *func;
322
323 func = lb_select_tx_port_get_func(ctx->data.str_val);
324 if (!func)
325 return -EINVAL;
326 rcu_assign_pointer(lb_priv->select_tx_port_func, func);
327 return 0;
328}
329
330static int lb_tx_hash_to_port_mapping_init(struct team *team,
331 struct team_option_inst_info *info)
332{
333 struct lb_priv *lb_priv = get_lb_priv(team);
334 unsigned char hash = info->array_index;
335
336 LB_HTPM_OPT_INST_INFO_BY_HASH(lb_priv, hash) = info;
337 return 0;
338}
339
340static int lb_tx_hash_to_port_mapping_get(struct team *team,
341 struct team_gsetter_ctx *ctx)
342{
343 struct lb_priv *lb_priv = get_lb_priv(team);
344 struct team_port *port;
345 unsigned char hash = ctx->info->array_index;
346
347 port = LB_HTPM_PORT_BY_HASH(lb_priv, hash);
348 ctx->data.u32_val = port ? port->dev->ifindex : 0;
349 return 0;
350}
351
352static int lb_tx_hash_to_port_mapping_set(struct team *team,
353 struct team_gsetter_ctx *ctx)
354{
355 struct lb_priv *lb_priv = get_lb_priv(team);
356 struct team_port *port;
357 unsigned char hash = ctx->info->array_index;
358
359 list_for_each_entry(port, &team->port_list, list) {
360 if (ctx->data.u32_val == port->dev->ifindex &&
361 team_port_enabled(port)) {
362 rcu_assign_pointer(LB_HTPM_PORT_BY_HASH(lb_priv, hash),
363 port);
364 return 0;
365 }
366 }
367 return -ENODEV;
368}
369
370static int lb_hash_stats_init(struct team *team,
371 struct team_option_inst_info *info)
372{
373 struct lb_priv *lb_priv = get_lb_priv(team);
374 unsigned char hash = info->array_index;
375
376 lb_priv->ex->stats.info[hash].opt_inst_info = info;
377 return 0;
378}
379
380static int lb_hash_stats_get(struct team *team, struct team_gsetter_ctx *ctx)
381{
382 struct lb_priv *lb_priv = get_lb_priv(team);
383 unsigned char hash = ctx->info->array_index;
384
385 ctx->data.bin_val.ptr = &lb_priv->ex->stats.info[hash].stats;
386 ctx->data.bin_val.len = sizeof(struct lb_stats);
387 return 0;
388}
389
390static int lb_port_stats_init(struct team *team,
391 struct team_option_inst_info *info)
392{
393 struct team_port *port = info->port;
394 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
395
396 lb_port_priv->stats_info.opt_inst_info = info;
397 return 0;
398}
399
400static int lb_port_stats_get(struct team *team, struct team_gsetter_ctx *ctx)
401{
402 struct team_port *port = ctx->info->port;
403 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
404
405 ctx->data.bin_val.ptr = &lb_port_priv->stats_info.stats;
406 ctx->data.bin_val.len = sizeof(struct lb_stats);
407 return 0;
408}
409
410static void __lb_stats_info_refresh_prepare(struct lb_stats_info *s_info)
411{
412 memcpy(&s_info->last_stats, &s_info->stats, sizeof(struct lb_stats));
413 memset(&s_info->stats, 0, sizeof(struct lb_stats));
414}
415
416static bool __lb_stats_info_refresh_check(struct lb_stats_info *s_info,
417 struct team *team)
418{
419 if (memcmp(&s_info->last_stats, &s_info->stats,
420 sizeof(struct lb_stats))) {
421 team_option_inst_set_change(s_info->opt_inst_info);
422 return true;
423 }
424 return false;
425}
426
427static void __lb_one_cpu_stats_add(struct lb_stats *acc_stats,
428 struct lb_stats *cpu_stats,
429 struct u64_stats_sync *syncp)
430{
431 unsigned int start;
432 struct lb_stats tmp;
433
434 do {
435 start = u64_stats_fetch_begin_irq(syncp);
436 tmp.tx_bytes = cpu_stats->tx_bytes;
437 } while (u64_stats_fetch_retry_irq(syncp, start));
438 acc_stats->tx_bytes += tmp.tx_bytes;
439}
440
441static void lb_stats_refresh(struct work_struct *work)
442{
443 struct team *team;
444 struct lb_priv *lb_priv;
445 struct lb_priv_ex *lb_priv_ex;
446 struct lb_pcpu_stats *pcpu_stats;
447 struct lb_stats *stats;
448 struct lb_stats_info *s_info;
449 struct team_port *port;
450 bool changed = false;
451 int i;
452 int j;
453
454 lb_priv_ex = container_of(work, struct lb_priv_ex,
455 stats.refresh_dw.work);
456
457 team = lb_priv_ex->team;
458 lb_priv = get_lb_priv(team);
459
460 if (!mutex_trylock(&team->lock)) {
461 schedule_delayed_work(&lb_priv_ex->stats.refresh_dw, 0);
462 return;
463 }
464
465 for (j = 0; j < LB_TX_HASHTABLE_SIZE; j++) {
466 s_info = &lb_priv->ex->stats.info[j];
467 __lb_stats_info_refresh_prepare(s_info);
468 for_each_possible_cpu(i) {
469 pcpu_stats = per_cpu_ptr(lb_priv->pcpu_stats, i);
470 stats = &pcpu_stats->hash_stats[j];
471 __lb_one_cpu_stats_add(&s_info->stats, stats,
472 &pcpu_stats->syncp);
473 }
474 changed |= __lb_stats_info_refresh_check(s_info, team);
475 }
476
477 list_for_each_entry(port, &team->port_list, list) {
478 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
479
480 s_info = &lb_port_priv->stats_info;
481 __lb_stats_info_refresh_prepare(s_info);
482 for_each_possible_cpu(i) {
483 pcpu_stats = per_cpu_ptr(lb_priv->pcpu_stats, i);
484 stats = per_cpu_ptr(lb_port_priv->pcpu_stats, i);
485 __lb_one_cpu_stats_add(&s_info->stats, stats,
486 &pcpu_stats->syncp);
487 }
488 changed |= __lb_stats_info_refresh_check(s_info, team);
489 }
490
491 if (changed)
492 team_options_change_check(team);
493
494 schedule_delayed_work(&lb_priv_ex->stats.refresh_dw,
495 (lb_priv_ex->stats.refresh_interval * HZ) / 10);
496
497 mutex_unlock(&team->lock);
498}
499
500static int lb_stats_refresh_interval_get(struct team *team,
501 struct team_gsetter_ctx *ctx)
502{
503 struct lb_priv *lb_priv = get_lb_priv(team);
504
505 ctx->data.u32_val = lb_priv->ex->stats.refresh_interval;
506 return 0;
507}
508
509static int lb_stats_refresh_interval_set(struct team *team,
510 struct team_gsetter_ctx *ctx)
511{
512 struct lb_priv *lb_priv = get_lb_priv(team);
513 unsigned int interval;
514
515 interval = ctx->data.u32_val;
516 if (lb_priv->ex->stats.refresh_interval == interval)
517 return 0;
518 lb_priv->ex->stats.refresh_interval = interval;
519 if (interval)
520 schedule_delayed_work(&lb_priv->ex->stats.refresh_dw, 0);
521 else
522 cancel_delayed_work(&lb_priv->ex->stats.refresh_dw);
523 return 0;
524}
525
526static const struct team_option lb_options[] = {
527 {
528 .name = "bpf_hash_func",
529 .type = TEAM_OPTION_TYPE_BINARY,
530 .getter = lb_bpf_func_get,
531 .setter = lb_bpf_func_set,
532 },
533 {
534 .name = "lb_tx_method",
535 .type = TEAM_OPTION_TYPE_STRING,
536 .getter = lb_tx_method_get,
537 .setter = lb_tx_method_set,
538 },
539 {
540 .name = "lb_tx_hash_to_port_mapping",
541 .array_size = LB_TX_HASHTABLE_SIZE,
542 .type = TEAM_OPTION_TYPE_U32,
543 .init = lb_tx_hash_to_port_mapping_init,
544 .getter = lb_tx_hash_to_port_mapping_get,
545 .setter = lb_tx_hash_to_port_mapping_set,
546 },
547 {
548 .name = "lb_hash_stats",
549 .array_size = LB_TX_HASHTABLE_SIZE,
550 .type = TEAM_OPTION_TYPE_BINARY,
551 .init = lb_hash_stats_init,
552 .getter = lb_hash_stats_get,
553 },
554 {
555 .name = "lb_port_stats",
556 .per_port = true,
557 .type = TEAM_OPTION_TYPE_BINARY,
558 .init = lb_port_stats_init,
559 .getter = lb_port_stats_get,
560 },
561 {
562 .name = "lb_stats_refresh_interval",
563 .type = TEAM_OPTION_TYPE_U32,
564 .getter = lb_stats_refresh_interval_get,
565 .setter = lb_stats_refresh_interval_set,
566 },
567};
568
569static int lb_init(struct team *team)
570{
571 struct lb_priv *lb_priv = get_lb_priv(team);
572 lb_select_tx_port_func_t *func;
573 int i, err;
574
575 /* set default tx port selector */
576 func = lb_select_tx_port_get_func("hash");
577 BUG_ON(!func);
578 rcu_assign_pointer(lb_priv->select_tx_port_func, func);
579
580 lb_priv->ex = kzalloc(sizeof(*lb_priv->ex), GFP_KERNEL);
581 if (!lb_priv->ex)
582 return -ENOMEM;
583 lb_priv->ex->team = team;
584
585 lb_priv->pcpu_stats = alloc_percpu(struct lb_pcpu_stats);
586 if (!lb_priv->pcpu_stats) {
587 err = -ENOMEM;
588 goto err_alloc_pcpu_stats;
589 }
590
591 for_each_possible_cpu(i) {
592 struct lb_pcpu_stats *team_lb_stats;
593 team_lb_stats = per_cpu_ptr(lb_priv->pcpu_stats, i);
594 u64_stats_init(&team_lb_stats->syncp);
595 }
596
597
598 INIT_DELAYED_WORK(&lb_priv->ex->stats.refresh_dw, lb_stats_refresh);
599
600 err = team_options_register(team, lb_options, ARRAY_SIZE(lb_options));
601 if (err)
602 goto err_options_register;
603 return 0;
604
605err_options_register:
606 free_percpu(lb_priv->pcpu_stats);
607err_alloc_pcpu_stats:
608 kfree(lb_priv->ex);
609 return err;
610}
611
612static void lb_exit(struct team *team)
613{
614 struct lb_priv *lb_priv = get_lb_priv(team);
615
616 team_options_unregister(team, lb_options,
617 ARRAY_SIZE(lb_options));
618 cancel_delayed_work_sync(&lb_priv->ex->stats.refresh_dw);
619 free_percpu(lb_priv->pcpu_stats);
620 kfree(lb_priv->ex);
621}
622
623static int lb_port_enter(struct team *team, struct team_port *port)
624{
625 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
626
627 lb_port_priv->pcpu_stats = alloc_percpu(struct lb_stats);
628 if (!lb_port_priv->pcpu_stats)
629 return -ENOMEM;
630 return 0;
631}
632
633static void lb_port_leave(struct team *team, struct team_port *port)
634{
635 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
636
637 free_percpu(lb_port_priv->pcpu_stats);
638}
639
640static void lb_port_disabled(struct team *team, struct team_port *port)
641{
642 lb_tx_hash_to_port_mapping_null_port(team, port);
643}
644
645static const struct team_mode_ops lb_mode_ops = {
646 .init = lb_init,
647 .exit = lb_exit,
648 .port_enter = lb_port_enter,
649 .port_leave = lb_port_leave,
650 .port_disabled = lb_port_disabled,
651 .transmit = lb_transmit,
652};
653
654static const struct team_mode lb_mode = {
655 .kind = "loadbalance",
656 .owner = THIS_MODULE,
657 .priv_size = sizeof(struct lb_priv),
658 .port_priv_size = sizeof(struct lb_port_priv),
659 .ops = &lb_mode_ops,
660};
661
662static int __init lb_init_module(void)
663{
664 return team_mode_register(&lb_mode);
665}
666
667static void __exit lb_cleanup_module(void)
668{
669 team_mode_unregister(&lb_mode);
670}
671
672module_init(lb_init_module);
673module_exit(lb_cleanup_module);
674
675MODULE_LICENSE("GPL v2");
676MODULE_AUTHOR("Jiri Pirko <jpirko@redhat.com>");
677MODULE_DESCRIPTION("Load-balancing mode for team");
678MODULE_ALIAS("team-mode-loadbalance");
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * drivers/net/team/team_mode_loadbalance.c - Load-balancing mode for team
4 * Copyright (c) 2012 Jiri Pirko <jpirko@redhat.com>
5 */
6
7#include <linux/kernel.h>
8#include <linux/types.h>
9#include <linux/module.h>
10#include <linux/init.h>
11#include <linux/errno.h>
12#include <linux/netdevice.h>
13#include <linux/etherdevice.h>
14#include <linux/filter.h>
15#include <linux/if_team.h>
16
17static rx_handler_result_t lb_receive(struct team *team, struct team_port *port,
18 struct sk_buff *skb)
19{
20 if (unlikely(skb->protocol == htons(ETH_P_SLOW))) {
21 /* LACPDU packets should go to exact delivery */
22 const unsigned char *dest = eth_hdr(skb)->h_dest;
23
24 if (is_link_local_ether_addr(dest) && dest[5] == 0x02)
25 return RX_HANDLER_EXACT;
26 }
27 return RX_HANDLER_ANOTHER;
28}
29
30struct lb_priv;
31
32typedef struct team_port *lb_select_tx_port_func_t(struct team *,
33 struct lb_priv *,
34 struct sk_buff *,
35 unsigned char);
36
37#define LB_TX_HASHTABLE_SIZE 256 /* hash is a char */
38
39struct lb_stats {
40 u64 tx_bytes;
41};
42
43struct lb_pcpu_stats {
44 struct lb_stats hash_stats[LB_TX_HASHTABLE_SIZE];
45 struct u64_stats_sync syncp;
46};
47
48struct lb_stats_info {
49 struct lb_stats stats;
50 struct lb_stats last_stats;
51 struct team_option_inst_info *opt_inst_info;
52};
53
54struct lb_port_mapping {
55 struct team_port __rcu *port;
56 struct team_option_inst_info *opt_inst_info;
57};
58
59struct lb_priv_ex {
60 struct team *team;
61 struct lb_port_mapping tx_hash_to_port_mapping[LB_TX_HASHTABLE_SIZE];
62 struct sock_fprog_kern *orig_fprog;
63 struct {
64 unsigned int refresh_interval; /* in tenths of second */
65 struct delayed_work refresh_dw;
66 struct lb_stats_info info[LB_TX_HASHTABLE_SIZE];
67 } stats;
68};
69
70struct lb_priv {
71 struct bpf_prog __rcu *fp;
72 lb_select_tx_port_func_t __rcu *select_tx_port_func;
73 struct lb_pcpu_stats __percpu *pcpu_stats;
74 struct lb_priv_ex *ex; /* priv extension */
75};
76
77static struct lb_priv *get_lb_priv(struct team *team)
78{
79 return (struct lb_priv *) &team->mode_priv;
80}
81
82struct lb_port_priv {
83 struct lb_stats __percpu *pcpu_stats;
84 struct lb_stats_info stats_info;
85};
86
87static struct lb_port_priv *get_lb_port_priv(struct team_port *port)
88{
89 return (struct lb_port_priv *) &port->mode_priv;
90}
91
92#define LB_HTPM_PORT_BY_HASH(lp_priv, hash) \
93 (lb_priv)->ex->tx_hash_to_port_mapping[hash].port
94
95#define LB_HTPM_OPT_INST_INFO_BY_HASH(lp_priv, hash) \
96 (lb_priv)->ex->tx_hash_to_port_mapping[hash].opt_inst_info
97
98static void lb_tx_hash_to_port_mapping_null_port(struct team *team,
99 struct team_port *port)
100{
101 struct lb_priv *lb_priv = get_lb_priv(team);
102 bool changed = false;
103 int i;
104
105 for (i = 0; i < LB_TX_HASHTABLE_SIZE; i++) {
106 struct lb_port_mapping *pm;
107
108 pm = &lb_priv->ex->tx_hash_to_port_mapping[i];
109 if (rcu_access_pointer(pm->port) == port) {
110 RCU_INIT_POINTER(pm->port, NULL);
111 team_option_inst_set_change(pm->opt_inst_info);
112 changed = true;
113 }
114 }
115 if (changed)
116 team_options_change_check(team);
117}
118
119/* Basic tx selection based solely by hash */
120static struct team_port *lb_hash_select_tx_port(struct team *team,
121 struct lb_priv *lb_priv,
122 struct sk_buff *skb,
123 unsigned char hash)
124{
125 int port_index = team_num_to_port_index(team, hash);
126
127 return team_get_port_by_index_rcu(team, port_index);
128}
129
130/* Hash to port mapping select tx port */
131static struct team_port *lb_htpm_select_tx_port(struct team *team,
132 struct lb_priv *lb_priv,
133 struct sk_buff *skb,
134 unsigned char hash)
135{
136 struct team_port *port;
137
138 port = rcu_dereference_bh(LB_HTPM_PORT_BY_HASH(lb_priv, hash));
139 if (likely(port))
140 return port;
141 /* If no valid port in the table, fall back to simple hash */
142 return lb_hash_select_tx_port(team, lb_priv, skb, hash);
143}
144
145struct lb_select_tx_port {
146 char *name;
147 lb_select_tx_port_func_t *func;
148};
149
150static const struct lb_select_tx_port lb_select_tx_port_list[] = {
151 {
152 .name = "hash",
153 .func = lb_hash_select_tx_port,
154 },
155 {
156 .name = "hash_to_port_mapping",
157 .func = lb_htpm_select_tx_port,
158 },
159};
160#define LB_SELECT_TX_PORT_LIST_COUNT ARRAY_SIZE(lb_select_tx_port_list)
161
162static char *lb_select_tx_port_get_name(lb_select_tx_port_func_t *func)
163{
164 int i;
165
166 for (i = 0; i < LB_SELECT_TX_PORT_LIST_COUNT; i++) {
167 const struct lb_select_tx_port *item;
168
169 item = &lb_select_tx_port_list[i];
170 if (item->func == func)
171 return item->name;
172 }
173 return NULL;
174}
175
176static lb_select_tx_port_func_t *lb_select_tx_port_get_func(const char *name)
177{
178 int i;
179
180 for (i = 0; i < LB_SELECT_TX_PORT_LIST_COUNT; i++) {
181 const struct lb_select_tx_port *item;
182
183 item = &lb_select_tx_port_list[i];
184 if (!strcmp(item->name, name))
185 return item->func;
186 }
187 return NULL;
188}
189
190static unsigned int lb_get_skb_hash(struct lb_priv *lb_priv,
191 struct sk_buff *skb)
192{
193 struct bpf_prog *fp;
194 uint32_t lhash;
195 unsigned char *c;
196
197 fp = rcu_dereference_bh(lb_priv->fp);
198 if (unlikely(!fp))
199 return 0;
200 lhash = BPF_PROG_RUN(fp, skb);
201 c = (char *) &lhash;
202 return c[0] ^ c[1] ^ c[2] ^ c[3];
203}
204
205static void lb_update_tx_stats(unsigned int tx_bytes, struct lb_priv *lb_priv,
206 struct lb_port_priv *lb_port_priv,
207 unsigned char hash)
208{
209 struct lb_pcpu_stats *pcpu_stats;
210 struct lb_stats *port_stats;
211 struct lb_stats *hash_stats;
212
213 pcpu_stats = this_cpu_ptr(lb_priv->pcpu_stats);
214 port_stats = this_cpu_ptr(lb_port_priv->pcpu_stats);
215 hash_stats = &pcpu_stats->hash_stats[hash];
216 u64_stats_update_begin(&pcpu_stats->syncp);
217 port_stats->tx_bytes += tx_bytes;
218 hash_stats->tx_bytes += tx_bytes;
219 u64_stats_update_end(&pcpu_stats->syncp);
220}
221
222static bool lb_transmit(struct team *team, struct sk_buff *skb)
223{
224 struct lb_priv *lb_priv = get_lb_priv(team);
225 lb_select_tx_port_func_t *select_tx_port_func;
226 struct team_port *port;
227 unsigned char hash;
228 unsigned int tx_bytes = skb->len;
229
230 hash = lb_get_skb_hash(lb_priv, skb);
231 select_tx_port_func = rcu_dereference_bh(lb_priv->select_tx_port_func);
232 port = select_tx_port_func(team, lb_priv, skb, hash);
233 if (unlikely(!port))
234 goto drop;
235 if (team_dev_queue_xmit(team, port, skb))
236 return false;
237 lb_update_tx_stats(tx_bytes, lb_priv, get_lb_port_priv(port), hash);
238 return true;
239
240drop:
241 dev_kfree_skb_any(skb);
242 return false;
243}
244
245static int lb_bpf_func_get(struct team *team, struct team_gsetter_ctx *ctx)
246{
247 struct lb_priv *lb_priv = get_lb_priv(team);
248
249 if (!lb_priv->ex->orig_fprog) {
250 ctx->data.bin_val.len = 0;
251 ctx->data.bin_val.ptr = NULL;
252 return 0;
253 }
254 ctx->data.bin_val.len = lb_priv->ex->orig_fprog->len *
255 sizeof(struct sock_filter);
256 ctx->data.bin_val.ptr = lb_priv->ex->orig_fprog->filter;
257 return 0;
258}
259
260static int __fprog_create(struct sock_fprog_kern **pfprog, u32 data_len,
261 const void *data)
262{
263 struct sock_fprog_kern *fprog;
264 struct sock_filter *filter = (struct sock_filter *) data;
265
266 if (data_len % sizeof(struct sock_filter))
267 return -EINVAL;
268 fprog = kmalloc(sizeof(*fprog), GFP_KERNEL);
269 if (!fprog)
270 return -ENOMEM;
271 fprog->filter = kmemdup(filter, data_len, GFP_KERNEL);
272 if (!fprog->filter) {
273 kfree(fprog);
274 return -ENOMEM;
275 }
276 fprog->len = data_len / sizeof(struct sock_filter);
277 *pfprog = fprog;
278 return 0;
279}
280
281static void __fprog_destroy(struct sock_fprog_kern *fprog)
282{
283 kfree(fprog->filter);
284 kfree(fprog);
285}
286
287static int lb_bpf_func_set(struct team *team, struct team_gsetter_ctx *ctx)
288{
289 struct lb_priv *lb_priv = get_lb_priv(team);
290 struct bpf_prog *fp = NULL;
291 struct bpf_prog *orig_fp = NULL;
292 struct sock_fprog_kern *fprog = NULL;
293 int err;
294
295 if (ctx->data.bin_val.len) {
296 err = __fprog_create(&fprog, ctx->data.bin_val.len,
297 ctx->data.bin_val.ptr);
298 if (err)
299 return err;
300 err = bpf_prog_create(&fp, fprog);
301 if (err) {
302 __fprog_destroy(fprog);
303 return err;
304 }
305 }
306
307 if (lb_priv->ex->orig_fprog) {
308 /* Clear old filter data */
309 __fprog_destroy(lb_priv->ex->orig_fprog);
310 orig_fp = rcu_dereference_protected(lb_priv->fp,
311 lockdep_is_held(&team->lock));
312 }
313
314 rcu_assign_pointer(lb_priv->fp, fp);
315 lb_priv->ex->orig_fprog = fprog;
316
317 if (orig_fp) {
318 synchronize_rcu();
319 bpf_prog_destroy(orig_fp);
320 }
321 return 0;
322}
323
324static void lb_bpf_func_free(struct team *team)
325{
326 struct lb_priv *lb_priv = get_lb_priv(team);
327 struct bpf_prog *fp;
328
329 if (!lb_priv->ex->orig_fprog)
330 return;
331
332 __fprog_destroy(lb_priv->ex->orig_fprog);
333 fp = rcu_dereference_protected(lb_priv->fp,
334 lockdep_is_held(&team->lock));
335 bpf_prog_destroy(fp);
336}
337
338static int lb_tx_method_get(struct team *team, struct team_gsetter_ctx *ctx)
339{
340 struct lb_priv *lb_priv = get_lb_priv(team);
341 lb_select_tx_port_func_t *func;
342 char *name;
343
344 func = rcu_dereference_protected(lb_priv->select_tx_port_func,
345 lockdep_is_held(&team->lock));
346 name = lb_select_tx_port_get_name(func);
347 BUG_ON(!name);
348 ctx->data.str_val = name;
349 return 0;
350}
351
352static int lb_tx_method_set(struct team *team, struct team_gsetter_ctx *ctx)
353{
354 struct lb_priv *lb_priv = get_lb_priv(team);
355 lb_select_tx_port_func_t *func;
356
357 func = lb_select_tx_port_get_func(ctx->data.str_val);
358 if (!func)
359 return -EINVAL;
360 rcu_assign_pointer(lb_priv->select_tx_port_func, func);
361 return 0;
362}
363
364static int lb_tx_hash_to_port_mapping_init(struct team *team,
365 struct team_option_inst_info *info)
366{
367 struct lb_priv *lb_priv = get_lb_priv(team);
368 unsigned char hash = info->array_index;
369
370 LB_HTPM_OPT_INST_INFO_BY_HASH(lb_priv, hash) = info;
371 return 0;
372}
373
374static int lb_tx_hash_to_port_mapping_get(struct team *team,
375 struct team_gsetter_ctx *ctx)
376{
377 struct lb_priv *lb_priv = get_lb_priv(team);
378 struct team_port *port;
379 unsigned char hash = ctx->info->array_index;
380
381 port = LB_HTPM_PORT_BY_HASH(lb_priv, hash);
382 ctx->data.u32_val = port ? port->dev->ifindex : 0;
383 return 0;
384}
385
386static int lb_tx_hash_to_port_mapping_set(struct team *team,
387 struct team_gsetter_ctx *ctx)
388{
389 struct lb_priv *lb_priv = get_lb_priv(team);
390 struct team_port *port;
391 unsigned char hash = ctx->info->array_index;
392
393 list_for_each_entry(port, &team->port_list, list) {
394 if (ctx->data.u32_val == port->dev->ifindex &&
395 team_port_enabled(port)) {
396 rcu_assign_pointer(LB_HTPM_PORT_BY_HASH(lb_priv, hash),
397 port);
398 return 0;
399 }
400 }
401 return -ENODEV;
402}
403
404static int lb_hash_stats_init(struct team *team,
405 struct team_option_inst_info *info)
406{
407 struct lb_priv *lb_priv = get_lb_priv(team);
408 unsigned char hash = info->array_index;
409
410 lb_priv->ex->stats.info[hash].opt_inst_info = info;
411 return 0;
412}
413
414static int lb_hash_stats_get(struct team *team, struct team_gsetter_ctx *ctx)
415{
416 struct lb_priv *lb_priv = get_lb_priv(team);
417 unsigned char hash = ctx->info->array_index;
418
419 ctx->data.bin_val.ptr = &lb_priv->ex->stats.info[hash].stats;
420 ctx->data.bin_val.len = sizeof(struct lb_stats);
421 return 0;
422}
423
424static int lb_port_stats_init(struct team *team,
425 struct team_option_inst_info *info)
426{
427 struct team_port *port = info->port;
428 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
429
430 lb_port_priv->stats_info.opt_inst_info = info;
431 return 0;
432}
433
434static int lb_port_stats_get(struct team *team, struct team_gsetter_ctx *ctx)
435{
436 struct team_port *port = ctx->info->port;
437 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
438
439 ctx->data.bin_val.ptr = &lb_port_priv->stats_info.stats;
440 ctx->data.bin_val.len = sizeof(struct lb_stats);
441 return 0;
442}
443
444static void __lb_stats_info_refresh_prepare(struct lb_stats_info *s_info)
445{
446 memcpy(&s_info->last_stats, &s_info->stats, sizeof(struct lb_stats));
447 memset(&s_info->stats, 0, sizeof(struct lb_stats));
448}
449
450static bool __lb_stats_info_refresh_check(struct lb_stats_info *s_info,
451 struct team *team)
452{
453 if (memcmp(&s_info->last_stats, &s_info->stats,
454 sizeof(struct lb_stats))) {
455 team_option_inst_set_change(s_info->opt_inst_info);
456 return true;
457 }
458 return false;
459}
460
461static void __lb_one_cpu_stats_add(struct lb_stats *acc_stats,
462 struct lb_stats *cpu_stats,
463 struct u64_stats_sync *syncp)
464{
465 unsigned int start;
466 struct lb_stats tmp;
467
468 do {
469 start = u64_stats_fetch_begin_irq(syncp);
470 tmp.tx_bytes = cpu_stats->tx_bytes;
471 } while (u64_stats_fetch_retry_irq(syncp, start));
472 acc_stats->tx_bytes += tmp.tx_bytes;
473}
474
475static void lb_stats_refresh(struct work_struct *work)
476{
477 struct team *team;
478 struct lb_priv *lb_priv;
479 struct lb_priv_ex *lb_priv_ex;
480 struct lb_pcpu_stats *pcpu_stats;
481 struct lb_stats *stats;
482 struct lb_stats_info *s_info;
483 struct team_port *port;
484 bool changed = false;
485 int i;
486 int j;
487
488 lb_priv_ex = container_of(work, struct lb_priv_ex,
489 stats.refresh_dw.work);
490
491 team = lb_priv_ex->team;
492 lb_priv = get_lb_priv(team);
493
494 if (!mutex_trylock(&team->lock)) {
495 schedule_delayed_work(&lb_priv_ex->stats.refresh_dw, 0);
496 return;
497 }
498
499 for (j = 0; j < LB_TX_HASHTABLE_SIZE; j++) {
500 s_info = &lb_priv->ex->stats.info[j];
501 __lb_stats_info_refresh_prepare(s_info);
502 for_each_possible_cpu(i) {
503 pcpu_stats = per_cpu_ptr(lb_priv->pcpu_stats, i);
504 stats = &pcpu_stats->hash_stats[j];
505 __lb_one_cpu_stats_add(&s_info->stats, stats,
506 &pcpu_stats->syncp);
507 }
508 changed |= __lb_stats_info_refresh_check(s_info, team);
509 }
510
511 list_for_each_entry(port, &team->port_list, list) {
512 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
513
514 s_info = &lb_port_priv->stats_info;
515 __lb_stats_info_refresh_prepare(s_info);
516 for_each_possible_cpu(i) {
517 pcpu_stats = per_cpu_ptr(lb_priv->pcpu_stats, i);
518 stats = per_cpu_ptr(lb_port_priv->pcpu_stats, i);
519 __lb_one_cpu_stats_add(&s_info->stats, stats,
520 &pcpu_stats->syncp);
521 }
522 changed |= __lb_stats_info_refresh_check(s_info, team);
523 }
524
525 if (changed)
526 team_options_change_check(team);
527
528 schedule_delayed_work(&lb_priv_ex->stats.refresh_dw,
529 (lb_priv_ex->stats.refresh_interval * HZ) / 10);
530
531 mutex_unlock(&team->lock);
532}
533
534static int lb_stats_refresh_interval_get(struct team *team,
535 struct team_gsetter_ctx *ctx)
536{
537 struct lb_priv *lb_priv = get_lb_priv(team);
538
539 ctx->data.u32_val = lb_priv->ex->stats.refresh_interval;
540 return 0;
541}
542
543static int lb_stats_refresh_interval_set(struct team *team,
544 struct team_gsetter_ctx *ctx)
545{
546 struct lb_priv *lb_priv = get_lb_priv(team);
547 unsigned int interval;
548
549 interval = ctx->data.u32_val;
550 if (lb_priv->ex->stats.refresh_interval == interval)
551 return 0;
552 lb_priv->ex->stats.refresh_interval = interval;
553 if (interval)
554 schedule_delayed_work(&lb_priv->ex->stats.refresh_dw, 0);
555 else
556 cancel_delayed_work(&lb_priv->ex->stats.refresh_dw);
557 return 0;
558}
559
560static const struct team_option lb_options[] = {
561 {
562 .name = "bpf_hash_func",
563 .type = TEAM_OPTION_TYPE_BINARY,
564 .getter = lb_bpf_func_get,
565 .setter = lb_bpf_func_set,
566 },
567 {
568 .name = "lb_tx_method",
569 .type = TEAM_OPTION_TYPE_STRING,
570 .getter = lb_tx_method_get,
571 .setter = lb_tx_method_set,
572 },
573 {
574 .name = "lb_tx_hash_to_port_mapping",
575 .array_size = LB_TX_HASHTABLE_SIZE,
576 .type = TEAM_OPTION_TYPE_U32,
577 .init = lb_tx_hash_to_port_mapping_init,
578 .getter = lb_tx_hash_to_port_mapping_get,
579 .setter = lb_tx_hash_to_port_mapping_set,
580 },
581 {
582 .name = "lb_hash_stats",
583 .array_size = LB_TX_HASHTABLE_SIZE,
584 .type = TEAM_OPTION_TYPE_BINARY,
585 .init = lb_hash_stats_init,
586 .getter = lb_hash_stats_get,
587 },
588 {
589 .name = "lb_port_stats",
590 .per_port = true,
591 .type = TEAM_OPTION_TYPE_BINARY,
592 .init = lb_port_stats_init,
593 .getter = lb_port_stats_get,
594 },
595 {
596 .name = "lb_stats_refresh_interval",
597 .type = TEAM_OPTION_TYPE_U32,
598 .getter = lb_stats_refresh_interval_get,
599 .setter = lb_stats_refresh_interval_set,
600 },
601};
602
603static int lb_init(struct team *team)
604{
605 struct lb_priv *lb_priv = get_lb_priv(team);
606 lb_select_tx_port_func_t *func;
607 int i, err;
608
609 /* set default tx port selector */
610 func = lb_select_tx_port_get_func("hash");
611 BUG_ON(!func);
612 rcu_assign_pointer(lb_priv->select_tx_port_func, func);
613
614 lb_priv->ex = kzalloc(sizeof(*lb_priv->ex), GFP_KERNEL);
615 if (!lb_priv->ex)
616 return -ENOMEM;
617 lb_priv->ex->team = team;
618
619 lb_priv->pcpu_stats = alloc_percpu(struct lb_pcpu_stats);
620 if (!lb_priv->pcpu_stats) {
621 err = -ENOMEM;
622 goto err_alloc_pcpu_stats;
623 }
624
625 for_each_possible_cpu(i) {
626 struct lb_pcpu_stats *team_lb_stats;
627 team_lb_stats = per_cpu_ptr(lb_priv->pcpu_stats, i);
628 u64_stats_init(&team_lb_stats->syncp);
629 }
630
631
632 INIT_DELAYED_WORK(&lb_priv->ex->stats.refresh_dw, lb_stats_refresh);
633
634 err = team_options_register(team, lb_options, ARRAY_SIZE(lb_options));
635 if (err)
636 goto err_options_register;
637 return 0;
638
639err_options_register:
640 free_percpu(lb_priv->pcpu_stats);
641err_alloc_pcpu_stats:
642 kfree(lb_priv->ex);
643 return err;
644}
645
646static void lb_exit(struct team *team)
647{
648 struct lb_priv *lb_priv = get_lb_priv(team);
649
650 team_options_unregister(team, lb_options,
651 ARRAY_SIZE(lb_options));
652 lb_bpf_func_free(team);
653 cancel_delayed_work_sync(&lb_priv->ex->stats.refresh_dw);
654 free_percpu(lb_priv->pcpu_stats);
655 kfree(lb_priv->ex);
656}
657
658static int lb_port_enter(struct team *team, struct team_port *port)
659{
660 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
661
662 lb_port_priv->pcpu_stats = alloc_percpu(struct lb_stats);
663 if (!lb_port_priv->pcpu_stats)
664 return -ENOMEM;
665 return 0;
666}
667
668static void lb_port_leave(struct team *team, struct team_port *port)
669{
670 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
671
672 free_percpu(lb_port_priv->pcpu_stats);
673}
674
675static void lb_port_disabled(struct team *team, struct team_port *port)
676{
677 lb_tx_hash_to_port_mapping_null_port(team, port);
678}
679
680static const struct team_mode_ops lb_mode_ops = {
681 .init = lb_init,
682 .exit = lb_exit,
683 .port_enter = lb_port_enter,
684 .port_leave = lb_port_leave,
685 .port_disabled = lb_port_disabled,
686 .receive = lb_receive,
687 .transmit = lb_transmit,
688};
689
690static const struct team_mode lb_mode = {
691 .kind = "loadbalance",
692 .owner = THIS_MODULE,
693 .priv_size = sizeof(struct lb_priv),
694 .port_priv_size = sizeof(struct lb_port_priv),
695 .ops = &lb_mode_ops,
696 .lag_tx_type = NETDEV_LAG_TX_TYPE_HASH,
697};
698
699static int __init lb_init_module(void)
700{
701 return team_mode_register(&lb_mode);
702}
703
704static void __exit lb_cleanup_module(void)
705{
706 team_mode_unregister(&lb_mode);
707}
708
709module_init(lb_init_module);
710module_exit(lb_cleanup_module);
711
712MODULE_LICENSE("GPL v2");
713MODULE_AUTHOR("Jiri Pirko <jpirko@redhat.com>");
714MODULE_DESCRIPTION("Load-balancing mode for team");
715MODULE_ALIAS_TEAM_MODE("loadbalance");