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