1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * drivers/net/team/team_mode_loadbalance.c - Load-balancing mode for team
4 * Copyright (c) 2012 Jiri Pirko <jpirko@redhat.com>
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>
17 static rx_handler_result_t lb_receive(struct team *team, struct team_port *port,
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;
24 if (is_link_local_ether_addr(dest) && dest[5] == 0x02)
25 return RX_HANDLER_EXACT;
27 return RX_HANDLER_ANOTHER;
32 typedef struct team_port *lb_select_tx_port_func_t(struct team *,
37 #define LB_TX_HASHTABLE_SIZE 256 /* hash is a char */
43 struct lb_pcpu_stats {
44 struct lb_stats hash_stats[LB_TX_HASHTABLE_SIZE];
45 struct u64_stats_sync syncp;
48 struct lb_stats_info {
49 struct lb_stats stats;
50 struct lb_stats last_stats;
51 struct team_option_inst_info *opt_inst_info;
54 struct lb_port_mapping {
55 struct team_port __rcu *port;
56 struct team_option_inst_info *opt_inst_info;
61 struct lb_port_mapping tx_hash_to_port_mapping[LB_TX_HASHTABLE_SIZE];
62 struct sock_fprog_kern *orig_fprog;
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];
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 */
77 static struct lb_priv *get_lb_priv(struct team *team)
79 return (struct lb_priv *) &team->mode_priv;
83 struct lb_stats __percpu *pcpu_stats;
84 struct lb_stats_info stats_info;
87 static struct lb_port_priv *get_lb_port_priv(struct team_port *port)
89 return (struct lb_port_priv *) &port->mode_priv;
92 #define LB_HTPM_PORT_BY_HASH(lp_priv, hash) \
93 (lb_priv)->ex->tx_hash_to_port_mapping[hash].port
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
98 static void lb_tx_hash_to_port_mapping_null_port(struct team *team,
99 struct team_port *port)
101 struct lb_priv *lb_priv = get_lb_priv(team);
102 bool changed = false;
105 for (i = 0; i < LB_TX_HASHTABLE_SIZE; i++) {
106 struct lb_port_mapping *pm;
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);
116 team_options_change_check(team);
119 /* Basic tx selection based solely by hash */
120 static struct team_port *lb_hash_select_tx_port(struct team *team,
121 struct lb_priv *lb_priv,
125 int port_index = team_num_to_port_index(team, hash);
127 return team_get_port_by_index_rcu(team, port_index);
130 /* Hash to port mapping select tx port */
131 static struct team_port *lb_htpm_select_tx_port(struct team *team,
132 struct lb_priv *lb_priv,
136 struct team_port *port;
138 port = rcu_dereference_bh(LB_HTPM_PORT_BY_HASH(lb_priv, hash));
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);
145 struct lb_select_tx_port {
147 lb_select_tx_port_func_t *func;
150 static const struct lb_select_tx_port lb_select_tx_port_list[] = {
153 .func = lb_hash_select_tx_port,
156 .name = "hash_to_port_mapping",
157 .func = lb_htpm_select_tx_port,
160 #define LB_SELECT_TX_PORT_LIST_COUNT ARRAY_SIZE(lb_select_tx_port_list)
162 static char *lb_select_tx_port_get_name(lb_select_tx_port_func_t *func)
166 for (i = 0; i < LB_SELECT_TX_PORT_LIST_COUNT; i++) {
167 const struct lb_select_tx_port *item;
169 item = &lb_select_tx_port_list[i];
170 if (item->func == func)
176 static lb_select_tx_port_func_t *lb_select_tx_port_get_func(const char *name)
180 for (i = 0; i < LB_SELECT_TX_PORT_LIST_COUNT; i++) {
181 const struct lb_select_tx_port *item;
183 item = &lb_select_tx_port_list[i];
184 if (!strcmp(item->name, name))
190 static unsigned int lb_get_skb_hash(struct lb_priv *lb_priv,
197 fp = rcu_dereference_bh(lb_priv->fp);
200 lhash = bpf_prog_run(fp, skb);
202 return c[0] ^ c[1] ^ c[2] ^ c[3];
205 static void lb_update_tx_stats(unsigned int tx_bytes, struct lb_priv *lb_priv,
206 struct lb_port_priv *lb_port_priv,
209 struct lb_pcpu_stats *pcpu_stats;
210 struct lb_stats *port_stats;
211 struct lb_stats *hash_stats;
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);
222 static bool lb_transmit(struct team *team, struct sk_buff *skb)
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;
228 unsigned int tx_bytes = skb->len;
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);
235 if (team_dev_queue_xmit(team, port, skb))
237 lb_update_tx_stats(tx_bytes, lb_priv, get_lb_port_priv(port), hash);
241 dev_kfree_skb_any(skb);
245 static int lb_bpf_func_get(struct team *team, struct team_gsetter_ctx *ctx)
247 struct lb_priv *lb_priv = get_lb_priv(team);
249 if (!lb_priv->ex->orig_fprog) {
250 ctx->data.bin_val.len = 0;
251 ctx->data.bin_val.ptr = NULL;
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;
260 static int __fprog_create(struct sock_fprog_kern **pfprog, u32 data_len,
263 struct sock_fprog_kern *fprog;
264 struct sock_filter *filter = (struct sock_filter *) data;
266 if (data_len % sizeof(struct sock_filter))
268 fprog = kmalloc(sizeof(*fprog), GFP_KERNEL);
271 fprog->filter = kmemdup(filter, data_len, GFP_KERNEL);
272 if (!fprog->filter) {
276 fprog->len = data_len / sizeof(struct sock_filter);
281 static void __fprog_destroy(struct sock_fprog_kern *fprog)
283 kfree(fprog->filter);
287 static int lb_bpf_func_set(struct team *team, struct team_gsetter_ctx *ctx)
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;
295 if (ctx->data.bin_val.len) {
296 err = __fprog_create(&fprog, ctx->data.bin_val.len,
297 ctx->data.bin_val.ptr);
300 err = bpf_prog_create(&fp, fprog);
302 __fprog_destroy(fprog);
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));
314 rcu_assign_pointer(lb_priv->fp, fp);
315 lb_priv->ex->orig_fprog = fprog;
319 bpf_prog_destroy(orig_fp);
324 static void lb_bpf_func_free(struct team *team)
326 struct lb_priv *lb_priv = get_lb_priv(team);
329 if (!lb_priv->ex->orig_fprog)
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);
338 static int lb_tx_method_get(struct team *team, struct team_gsetter_ctx *ctx)
340 struct lb_priv *lb_priv = get_lb_priv(team);
341 lb_select_tx_port_func_t *func;
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);
348 ctx->data.str_val = name;
352 static int lb_tx_method_set(struct team *team, struct team_gsetter_ctx *ctx)
354 struct lb_priv *lb_priv = get_lb_priv(team);
355 lb_select_tx_port_func_t *func;
357 func = lb_select_tx_port_get_func(ctx->data.str_val);
360 rcu_assign_pointer(lb_priv->select_tx_port_func, func);
364 static void lb_tx_hash_to_port_mapping_init(struct team *team,
365 struct team_option_inst_info *info)
367 struct lb_priv *lb_priv = get_lb_priv(team);
368 unsigned char hash = info->array_index;
370 LB_HTPM_OPT_INST_INFO_BY_HASH(lb_priv, hash) = info;
373 static int lb_tx_hash_to_port_mapping_get(struct team *team,
374 struct team_gsetter_ctx *ctx)
376 struct lb_priv *lb_priv = get_lb_priv(team);
377 struct team_port *port;
378 unsigned char hash = ctx->info->array_index;
380 port = LB_HTPM_PORT_BY_HASH(lb_priv, hash);
381 ctx->data.u32_val = port ? port->dev->ifindex : 0;
385 static int lb_tx_hash_to_port_mapping_set(struct team *team,
386 struct team_gsetter_ctx *ctx)
388 struct lb_priv *lb_priv = get_lb_priv(team);
389 struct team_port *port;
390 unsigned char hash = ctx->info->array_index;
392 list_for_each_entry(port, &team->port_list, list) {
393 if (ctx->data.u32_val == port->dev->ifindex &&
394 team_port_enabled(port)) {
395 rcu_assign_pointer(LB_HTPM_PORT_BY_HASH(lb_priv, hash),
403 static void lb_hash_stats_init(struct team *team,
404 struct team_option_inst_info *info)
406 struct lb_priv *lb_priv = get_lb_priv(team);
407 unsigned char hash = info->array_index;
409 lb_priv->ex->stats.info[hash].opt_inst_info = info;
412 static int lb_hash_stats_get(struct team *team, struct team_gsetter_ctx *ctx)
414 struct lb_priv *lb_priv = get_lb_priv(team);
415 unsigned char hash = ctx->info->array_index;
417 ctx->data.bin_val.ptr = &lb_priv->ex->stats.info[hash].stats;
418 ctx->data.bin_val.len = sizeof(struct lb_stats);
422 static void lb_port_stats_init(struct team *team,
423 struct team_option_inst_info *info)
425 struct team_port *port = info->port;
426 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
428 lb_port_priv->stats_info.opt_inst_info = info;
431 static int lb_port_stats_get(struct team *team, struct team_gsetter_ctx *ctx)
433 struct team_port *port = ctx->info->port;
434 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
436 ctx->data.bin_val.ptr = &lb_port_priv->stats_info.stats;
437 ctx->data.bin_val.len = sizeof(struct lb_stats);
441 static void __lb_stats_info_refresh_prepare(struct lb_stats_info *s_info)
443 memcpy(&s_info->last_stats, &s_info->stats, sizeof(struct lb_stats));
444 memset(&s_info->stats, 0, sizeof(struct lb_stats));
447 static bool __lb_stats_info_refresh_check(struct lb_stats_info *s_info,
450 if (memcmp(&s_info->last_stats, &s_info->stats,
451 sizeof(struct lb_stats))) {
452 team_option_inst_set_change(s_info->opt_inst_info);
458 static void __lb_one_cpu_stats_add(struct lb_stats *acc_stats,
459 struct lb_stats *cpu_stats,
460 struct u64_stats_sync *syncp)
466 start = u64_stats_fetch_begin(syncp);
467 tmp.tx_bytes = cpu_stats->tx_bytes;
468 } while (u64_stats_fetch_retry(syncp, start));
469 acc_stats->tx_bytes += tmp.tx_bytes;
472 static void lb_stats_refresh(struct work_struct *work)
475 struct lb_priv *lb_priv;
476 struct lb_priv_ex *lb_priv_ex;
477 struct lb_pcpu_stats *pcpu_stats;
478 struct lb_stats *stats;
479 struct lb_stats_info *s_info;
480 struct team_port *port;
481 bool changed = false;
485 lb_priv_ex = container_of(work, struct lb_priv_ex,
486 stats.refresh_dw.work);
488 team = lb_priv_ex->team;
489 lb_priv = get_lb_priv(team);
491 if (!mutex_trylock(&team->lock)) {
492 schedule_delayed_work(&lb_priv_ex->stats.refresh_dw, 0);
496 for (j = 0; j < LB_TX_HASHTABLE_SIZE; j++) {
497 s_info = &lb_priv->ex->stats.info[j];
498 __lb_stats_info_refresh_prepare(s_info);
499 for_each_possible_cpu(i) {
500 pcpu_stats = per_cpu_ptr(lb_priv->pcpu_stats, i);
501 stats = &pcpu_stats->hash_stats[j];
502 __lb_one_cpu_stats_add(&s_info->stats, stats,
505 changed |= __lb_stats_info_refresh_check(s_info, team);
508 list_for_each_entry(port, &team->port_list, list) {
509 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
511 s_info = &lb_port_priv->stats_info;
512 __lb_stats_info_refresh_prepare(s_info);
513 for_each_possible_cpu(i) {
514 pcpu_stats = per_cpu_ptr(lb_priv->pcpu_stats, i);
515 stats = per_cpu_ptr(lb_port_priv->pcpu_stats, i);
516 __lb_one_cpu_stats_add(&s_info->stats, stats,
519 changed |= __lb_stats_info_refresh_check(s_info, team);
523 team_options_change_check(team);
525 schedule_delayed_work(&lb_priv_ex->stats.refresh_dw,
526 (lb_priv_ex->stats.refresh_interval * HZ) / 10);
528 mutex_unlock(&team->lock);
531 static int lb_stats_refresh_interval_get(struct team *team,
532 struct team_gsetter_ctx *ctx)
534 struct lb_priv *lb_priv = get_lb_priv(team);
536 ctx->data.u32_val = lb_priv->ex->stats.refresh_interval;
540 static int lb_stats_refresh_interval_set(struct team *team,
541 struct team_gsetter_ctx *ctx)
543 struct lb_priv *lb_priv = get_lb_priv(team);
544 unsigned int interval;
546 interval = ctx->data.u32_val;
547 if (lb_priv->ex->stats.refresh_interval == interval)
549 lb_priv->ex->stats.refresh_interval = interval;
551 schedule_delayed_work(&lb_priv->ex->stats.refresh_dw, 0);
553 cancel_delayed_work(&lb_priv->ex->stats.refresh_dw);
557 static const struct team_option lb_options[] = {
559 .name = "bpf_hash_func",
560 .type = TEAM_OPTION_TYPE_BINARY,
561 .getter = lb_bpf_func_get,
562 .setter = lb_bpf_func_set,
565 .name = "lb_tx_method",
566 .type = TEAM_OPTION_TYPE_STRING,
567 .getter = lb_tx_method_get,
568 .setter = lb_tx_method_set,
571 .name = "lb_tx_hash_to_port_mapping",
572 .array_size = LB_TX_HASHTABLE_SIZE,
573 .type = TEAM_OPTION_TYPE_U32,
574 .init = lb_tx_hash_to_port_mapping_init,
575 .getter = lb_tx_hash_to_port_mapping_get,
576 .setter = lb_tx_hash_to_port_mapping_set,
579 .name = "lb_hash_stats",
580 .array_size = LB_TX_HASHTABLE_SIZE,
581 .type = TEAM_OPTION_TYPE_BINARY,
582 .init = lb_hash_stats_init,
583 .getter = lb_hash_stats_get,
586 .name = "lb_port_stats",
588 .type = TEAM_OPTION_TYPE_BINARY,
589 .init = lb_port_stats_init,
590 .getter = lb_port_stats_get,
593 .name = "lb_stats_refresh_interval",
594 .type = TEAM_OPTION_TYPE_U32,
595 .getter = lb_stats_refresh_interval_get,
596 .setter = lb_stats_refresh_interval_set,
600 static int lb_init(struct team *team)
602 struct lb_priv *lb_priv = get_lb_priv(team);
603 lb_select_tx_port_func_t *func;
606 /* set default tx port selector */
607 func = lb_select_tx_port_get_func("hash");
609 rcu_assign_pointer(lb_priv->select_tx_port_func, func);
611 lb_priv->ex = kzalloc(sizeof(*lb_priv->ex), GFP_KERNEL);
614 lb_priv->ex->team = team;
616 lb_priv->pcpu_stats = alloc_percpu(struct lb_pcpu_stats);
617 if (!lb_priv->pcpu_stats) {
619 goto err_alloc_pcpu_stats;
622 for_each_possible_cpu(i) {
623 struct lb_pcpu_stats *team_lb_stats;
624 team_lb_stats = per_cpu_ptr(lb_priv->pcpu_stats, i);
625 u64_stats_init(&team_lb_stats->syncp);
629 INIT_DELAYED_WORK(&lb_priv->ex->stats.refresh_dw, lb_stats_refresh);
631 err = team_options_register(team, lb_options, ARRAY_SIZE(lb_options));
633 goto err_options_register;
636 err_options_register:
637 free_percpu(lb_priv->pcpu_stats);
638 err_alloc_pcpu_stats:
643 static void lb_exit(struct team *team)
645 struct lb_priv *lb_priv = get_lb_priv(team);
647 team_options_unregister(team, lb_options,
648 ARRAY_SIZE(lb_options));
649 lb_bpf_func_free(team);
650 cancel_delayed_work_sync(&lb_priv->ex->stats.refresh_dw);
651 free_percpu(lb_priv->pcpu_stats);
655 static int lb_port_enter(struct team *team, struct team_port *port)
657 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
659 lb_port_priv->pcpu_stats = alloc_percpu(struct lb_stats);
660 if (!lb_port_priv->pcpu_stats)
665 static void lb_port_leave(struct team *team, struct team_port *port)
667 struct lb_port_priv *lb_port_priv = get_lb_port_priv(port);
669 free_percpu(lb_port_priv->pcpu_stats);
672 static void lb_port_disabled(struct team *team, struct team_port *port)
674 lb_tx_hash_to_port_mapping_null_port(team, port);
677 static const struct team_mode_ops lb_mode_ops = {
680 .port_enter = lb_port_enter,
681 .port_leave = lb_port_leave,
682 .port_disabled = lb_port_disabled,
683 .receive = lb_receive,
684 .transmit = lb_transmit,
687 static const struct team_mode lb_mode = {
688 .kind = "loadbalance",
689 .owner = THIS_MODULE,
690 .priv_size = sizeof(struct lb_priv),
691 .port_priv_size = sizeof(struct lb_port_priv),
693 .lag_tx_type = NETDEV_LAG_TX_TYPE_HASH,
696 static int __init lb_init_module(void)
698 return team_mode_register(&lb_mode);
701 static void __exit lb_cleanup_module(void)
703 team_mode_unregister(&lb_mode);
706 module_init(lb_init_module);
707 module_exit(lb_cleanup_module);
709 MODULE_LICENSE("GPL v2");
710 MODULE_AUTHOR("Jiri Pirko <jpirko@redhat.com>");
711 MODULE_DESCRIPTION("Load-balancing mode for team");
712 MODULE_ALIAS_TEAM_MODE("loadbalance");