162306a36Sopenharmony_ci// SPDX-License-Identifier: GPL-2.0-only 262306a36Sopenharmony_ci/* 362306a36Sopenharmony_ci * net/sched/act_sample.c - Packet sampling tc action 462306a36Sopenharmony_ci * Copyright (c) 2017 Yotam Gigi <yotamg@mellanox.com> 562306a36Sopenharmony_ci */ 662306a36Sopenharmony_ci 762306a36Sopenharmony_ci#include <linux/types.h> 862306a36Sopenharmony_ci#include <linux/kernel.h> 962306a36Sopenharmony_ci#include <linux/string.h> 1062306a36Sopenharmony_ci#include <linux/errno.h> 1162306a36Sopenharmony_ci#include <linux/skbuff.h> 1262306a36Sopenharmony_ci#include <linux/rtnetlink.h> 1362306a36Sopenharmony_ci#include <linux/module.h> 1462306a36Sopenharmony_ci#include <linux/init.h> 1562306a36Sopenharmony_ci#include <linux/gfp.h> 1662306a36Sopenharmony_ci#include <net/net_namespace.h> 1762306a36Sopenharmony_ci#include <net/netlink.h> 1862306a36Sopenharmony_ci#include <net/pkt_sched.h> 1962306a36Sopenharmony_ci#include <linux/tc_act/tc_sample.h> 2062306a36Sopenharmony_ci#include <net/tc_act/tc_sample.h> 2162306a36Sopenharmony_ci#include <net/psample.h> 2262306a36Sopenharmony_ci#include <net/pkt_cls.h> 2362306a36Sopenharmony_ci#include <net/tc_wrapper.h> 2462306a36Sopenharmony_ci 2562306a36Sopenharmony_ci#include <linux/if_arp.h> 2662306a36Sopenharmony_ci 2762306a36Sopenharmony_cistatic struct tc_action_ops act_sample_ops; 2862306a36Sopenharmony_ci 2962306a36Sopenharmony_cistatic const struct nla_policy sample_policy[TCA_SAMPLE_MAX + 1] = { 3062306a36Sopenharmony_ci [TCA_SAMPLE_PARMS] = { .len = sizeof(struct tc_sample) }, 3162306a36Sopenharmony_ci [TCA_SAMPLE_RATE] = { .type = NLA_U32 }, 3262306a36Sopenharmony_ci [TCA_SAMPLE_TRUNC_SIZE] = { .type = NLA_U32 }, 3362306a36Sopenharmony_ci [TCA_SAMPLE_PSAMPLE_GROUP] = { .type = NLA_U32 }, 3462306a36Sopenharmony_ci}; 3562306a36Sopenharmony_ci 3662306a36Sopenharmony_cistatic int tcf_sample_init(struct net *net, struct nlattr *nla, 3762306a36Sopenharmony_ci struct nlattr *est, struct tc_action **a, 3862306a36Sopenharmony_ci struct tcf_proto *tp, 3962306a36Sopenharmony_ci u32 flags, struct netlink_ext_ack *extack) 4062306a36Sopenharmony_ci{ 4162306a36Sopenharmony_ci struct tc_action_net *tn = net_generic(net, act_sample_ops.net_id); 4262306a36Sopenharmony_ci bool bind = flags & TCA_ACT_FLAGS_BIND; 4362306a36Sopenharmony_ci struct nlattr *tb[TCA_SAMPLE_MAX + 1]; 4462306a36Sopenharmony_ci struct psample_group *psample_group; 4562306a36Sopenharmony_ci u32 psample_group_num, rate, index; 4662306a36Sopenharmony_ci struct tcf_chain *goto_ch = NULL; 4762306a36Sopenharmony_ci struct tc_sample *parm; 4862306a36Sopenharmony_ci struct tcf_sample *s; 4962306a36Sopenharmony_ci bool exists = false; 5062306a36Sopenharmony_ci int ret, err; 5162306a36Sopenharmony_ci 5262306a36Sopenharmony_ci if (!nla) 5362306a36Sopenharmony_ci return -EINVAL; 5462306a36Sopenharmony_ci ret = nla_parse_nested_deprecated(tb, TCA_SAMPLE_MAX, nla, 5562306a36Sopenharmony_ci sample_policy, NULL); 5662306a36Sopenharmony_ci if (ret < 0) 5762306a36Sopenharmony_ci return ret; 5862306a36Sopenharmony_ci 5962306a36Sopenharmony_ci if (!tb[TCA_SAMPLE_PARMS]) 6062306a36Sopenharmony_ci return -EINVAL; 6162306a36Sopenharmony_ci 6262306a36Sopenharmony_ci parm = nla_data(tb[TCA_SAMPLE_PARMS]); 6362306a36Sopenharmony_ci index = parm->index; 6462306a36Sopenharmony_ci err = tcf_idr_check_alloc(tn, &index, a, bind); 6562306a36Sopenharmony_ci if (err < 0) 6662306a36Sopenharmony_ci return err; 6762306a36Sopenharmony_ci exists = err; 6862306a36Sopenharmony_ci if (exists && bind) 6962306a36Sopenharmony_ci return 0; 7062306a36Sopenharmony_ci 7162306a36Sopenharmony_ci if (!exists) { 7262306a36Sopenharmony_ci ret = tcf_idr_create(tn, index, est, a, 7362306a36Sopenharmony_ci &act_sample_ops, bind, true, flags); 7462306a36Sopenharmony_ci if (ret) { 7562306a36Sopenharmony_ci tcf_idr_cleanup(tn, index); 7662306a36Sopenharmony_ci return ret; 7762306a36Sopenharmony_ci } 7862306a36Sopenharmony_ci ret = ACT_P_CREATED; 7962306a36Sopenharmony_ci } else if (!(flags & TCA_ACT_FLAGS_REPLACE)) { 8062306a36Sopenharmony_ci tcf_idr_release(*a, bind); 8162306a36Sopenharmony_ci return -EEXIST; 8262306a36Sopenharmony_ci } 8362306a36Sopenharmony_ci 8462306a36Sopenharmony_ci if (!tb[TCA_SAMPLE_RATE] || !tb[TCA_SAMPLE_PSAMPLE_GROUP]) { 8562306a36Sopenharmony_ci NL_SET_ERR_MSG(extack, "sample rate and group are required"); 8662306a36Sopenharmony_ci err = -EINVAL; 8762306a36Sopenharmony_ci goto release_idr; 8862306a36Sopenharmony_ci } 8962306a36Sopenharmony_ci 9062306a36Sopenharmony_ci err = tcf_action_check_ctrlact(parm->action, tp, &goto_ch, extack); 9162306a36Sopenharmony_ci if (err < 0) 9262306a36Sopenharmony_ci goto release_idr; 9362306a36Sopenharmony_ci 9462306a36Sopenharmony_ci rate = nla_get_u32(tb[TCA_SAMPLE_RATE]); 9562306a36Sopenharmony_ci if (!rate) { 9662306a36Sopenharmony_ci NL_SET_ERR_MSG(extack, "invalid sample rate"); 9762306a36Sopenharmony_ci err = -EINVAL; 9862306a36Sopenharmony_ci goto put_chain; 9962306a36Sopenharmony_ci } 10062306a36Sopenharmony_ci psample_group_num = nla_get_u32(tb[TCA_SAMPLE_PSAMPLE_GROUP]); 10162306a36Sopenharmony_ci psample_group = psample_group_get(net, psample_group_num); 10262306a36Sopenharmony_ci if (!psample_group) { 10362306a36Sopenharmony_ci err = -ENOMEM; 10462306a36Sopenharmony_ci goto put_chain; 10562306a36Sopenharmony_ci } 10662306a36Sopenharmony_ci 10762306a36Sopenharmony_ci s = to_sample(*a); 10862306a36Sopenharmony_ci 10962306a36Sopenharmony_ci spin_lock_bh(&s->tcf_lock); 11062306a36Sopenharmony_ci goto_ch = tcf_action_set_ctrlact(*a, parm->action, goto_ch); 11162306a36Sopenharmony_ci s->rate = rate; 11262306a36Sopenharmony_ci s->psample_group_num = psample_group_num; 11362306a36Sopenharmony_ci psample_group = rcu_replace_pointer(s->psample_group, psample_group, 11462306a36Sopenharmony_ci lockdep_is_held(&s->tcf_lock)); 11562306a36Sopenharmony_ci 11662306a36Sopenharmony_ci if (tb[TCA_SAMPLE_TRUNC_SIZE]) { 11762306a36Sopenharmony_ci s->truncate = true; 11862306a36Sopenharmony_ci s->trunc_size = nla_get_u32(tb[TCA_SAMPLE_TRUNC_SIZE]); 11962306a36Sopenharmony_ci } 12062306a36Sopenharmony_ci spin_unlock_bh(&s->tcf_lock); 12162306a36Sopenharmony_ci 12262306a36Sopenharmony_ci if (psample_group) 12362306a36Sopenharmony_ci psample_group_put(psample_group); 12462306a36Sopenharmony_ci if (goto_ch) 12562306a36Sopenharmony_ci tcf_chain_put_by_act(goto_ch); 12662306a36Sopenharmony_ci 12762306a36Sopenharmony_ci return ret; 12862306a36Sopenharmony_ciput_chain: 12962306a36Sopenharmony_ci if (goto_ch) 13062306a36Sopenharmony_ci tcf_chain_put_by_act(goto_ch); 13162306a36Sopenharmony_cirelease_idr: 13262306a36Sopenharmony_ci tcf_idr_release(*a, bind); 13362306a36Sopenharmony_ci return err; 13462306a36Sopenharmony_ci} 13562306a36Sopenharmony_ci 13662306a36Sopenharmony_cistatic void tcf_sample_cleanup(struct tc_action *a) 13762306a36Sopenharmony_ci{ 13862306a36Sopenharmony_ci struct tcf_sample *s = to_sample(a); 13962306a36Sopenharmony_ci struct psample_group *psample_group; 14062306a36Sopenharmony_ci 14162306a36Sopenharmony_ci /* last reference to action, no need to lock */ 14262306a36Sopenharmony_ci psample_group = rcu_dereference_protected(s->psample_group, 1); 14362306a36Sopenharmony_ci RCU_INIT_POINTER(s->psample_group, NULL); 14462306a36Sopenharmony_ci if (psample_group) 14562306a36Sopenharmony_ci psample_group_put(psample_group); 14662306a36Sopenharmony_ci} 14762306a36Sopenharmony_ci 14862306a36Sopenharmony_cistatic bool tcf_sample_dev_ok_push(struct net_device *dev) 14962306a36Sopenharmony_ci{ 15062306a36Sopenharmony_ci switch (dev->type) { 15162306a36Sopenharmony_ci case ARPHRD_TUNNEL: 15262306a36Sopenharmony_ci case ARPHRD_TUNNEL6: 15362306a36Sopenharmony_ci case ARPHRD_SIT: 15462306a36Sopenharmony_ci case ARPHRD_IPGRE: 15562306a36Sopenharmony_ci case ARPHRD_IP6GRE: 15662306a36Sopenharmony_ci case ARPHRD_VOID: 15762306a36Sopenharmony_ci case ARPHRD_NONE: 15862306a36Sopenharmony_ci return false; 15962306a36Sopenharmony_ci default: 16062306a36Sopenharmony_ci return true; 16162306a36Sopenharmony_ci } 16262306a36Sopenharmony_ci} 16362306a36Sopenharmony_ci 16462306a36Sopenharmony_ciTC_INDIRECT_SCOPE int tcf_sample_act(struct sk_buff *skb, 16562306a36Sopenharmony_ci const struct tc_action *a, 16662306a36Sopenharmony_ci struct tcf_result *res) 16762306a36Sopenharmony_ci{ 16862306a36Sopenharmony_ci struct tcf_sample *s = to_sample(a); 16962306a36Sopenharmony_ci struct psample_group *psample_group; 17062306a36Sopenharmony_ci struct psample_metadata md = {}; 17162306a36Sopenharmony_ci int retval; 17262306a36Sopenharmony_ci 17362306a36Sopenharmony_ci tcf_lastuse_update(&s->tcf_tm); 17462306a36Sopenharmony_ci bstats_update(this_cpu_ptr(s->common.cpu_bstats), skb); 17562306a36Sopenharmony_ci retval = READ_ONCE(s->tcf_action); 17662306a36Sopenharmony_ci 17762306a36Sopenharmony_ci psample_group = rcu_dereference_bh(s->psample_group); 17862306a36Sopenharmony_ci 17962306a36Sopenharmony_ci /* randomly sample packets according to rate */ 18062306a36Sopenharmony_ci if (psample_group && (get_random_u32_below(s->rate) == 0)) { 18162306a36Sopenharmony_ci if (!skb_at_tc_ingress(skb)) { 18262306a36Sopenharmony_ci md.in_ifindex = skb->skb_iif; 18362306a36Sopenharmony_ci md.out_ifindex = skb->dev->ifindex; 18462306a36Sopenharmony_ci } else { 18562306a36Sopenharmony_ci md.in_ifindex = skb->dev->ifindex; 18662306a36Sopenharmony_ci } 18762306a36Sopenharmony_ci 18862306a36Sopenharmony_ci /* on ingress, the mac header gets popped, so push it back */ 18962306a36Sopenharmony_ci if (skb_at_tc_ingress(skb) && tcf_sample_dev_ok_push(skb->dev)) 19062306a36Sopenharmony_ci skb_push(skb, skb->mac_len); 19162306a36Sopenharmony_ci 19262306a36Sopenharmony_ci md.trunc_size = s->truncate ? s->trunc_size : skb->len; 19362306a36Sopenharmony_ci psample_sample_packet(psample_group, skb, s->rate, &md); 19462306a36Sopenharmony_ci 19562306a36Sopenharmony_ci if (skb_at_tc_ingress(skb) && tcf_sample_dev_ok_push(skb->dev)) 19662306a36Sopenharmony_ci skb_pull(skb, skb->mac_len); 19762306a36Sopenharmony_ci } 19862306a36Sopenharmony_ci 19962306a36Sopenharmony_ci return retval; 20062306a36Sopenharmony_ci} 20162306a36Sopenharmony_ci 20262306a36Sopenharmony_cistatic void tcf_sample_stats_update(struct tc_action *a, u64 bytes, u64 packets, 20362306a36Sopenharmony_ci u64 drops, u64 lastuse, bool hw) 20462306a36Sopenharmony_ci{ 20562306a36Sopenharmony_ci struct tcf_sample *s = to_sample(a); 20662306a36Sopenharmony_ci struct tcf_t *tm = &s->tcf_tm; 20762306a36Sopenharmony_ci 20862306a36Sopenharmony_ci tcf_action_update_stats(a, bytes, packets, drops, hw); 20962306a36Sopenharmony_ci tm->lastuse = max_t(u64, tm->lastuse, lastuse); 21062306a36Sopenharmony_ci} 21162306a36Sopenharmony_ci 21262306a36Sopenharmony_cistatic int tcf_sample_dump(struct sk_buff *skb, struct tc_action *a, 21362306a36Sopenharmony_ci int bind, int ref) 21462306a36Sopenharmony_ci{ 21562306a36Sopenharmony_ci unsigned char *b = skb_tail_pointer(skb); 21662306a36Sopenharmony_ci struct tcf_sample *s = to_sample(a); 21762306a36Sopenharmony_ci struct tc_sample opt = { 21862306a36Sopenharmony_ci .index = s->tcf_index, 21962306a36Sopenharmony_ci .refcnt = refcount_read(&s->tcf_refcnt) - ref, 22062306a36Sopenharmony_ci .bindcnt = atomic_read(&s->tcf_bindcnt) - bind, 22162306a36Sopenharmony_ci }; 22262306a36Sopenharmony_ci struct tcf_t t; 22362306a36Sopenharmony_ci 22462306a36Sopenharmony_ci spin_lock_bh(&s->tcf_lock); 22562306a36Sopenharmony_ci opt.action = s->tcf_action; 22662306a36Sopenharmony_ci if (nla_put(skb, TCA_SAMPLE_PARMS, sizeof(opt), &opt)) 22762306a36Sopenharmony_ci goto nla_put_failure; 22862306a36Sopenharmony_ci 22962306a36Sopenharmony_ci tcf_tm_dump(&t, &s->tcf_tm); 23062306a36Sopenharmony_ci if (nla_put_64bit(skb, TCA_SAMPLE_TM, sizeof(t), &t, TCA_SAMPLE_PAD)) 23162306a36Sopenharmony_ci goto nla_put_failure; 23262306a36Sopenharmony_ci 23362306a36Sopenharmony_ci if (nla_put_u32(skb, TCA_SAMPLE_RATE, s->rate)) 23462306a36Sopenharmony_ci goto nla_put_failure; 23562306a36Sopenharmony_ci 23662306a36Sopenharmony_ci if (s->truncate) 23762306a36Sopenharmony_ci if (nla_put_u32(skb, TCA_SAMPLE_TRUNC_SIZE, s->trunc_size)) 23862306a36Sopenharmony_ci goto nla_put_failure; 23962306a36Sopenharmony_ci 24062306a36Sopenharmony_ci if (nla_put_u32(skb, TCA_SAMPLE_PSAMPLE_GROUP, s->psample_group_num)) 24162306a36Sopenharmony_ci goto nla_put_failure; 24262306a36Sopenharmony_ci spin_unlock_bh(&s->tcf_lock); 24362306a36Sopenharmony_ci 24462306a36Sopenharmony_ci return skb->len; 24562306a36Sopenharmony_ci 24662306a36Sopenharmony_cinla_put_failure: 24762306a36Sopenharmony_ci spin_unlock_bh(&s->tcf_lock); 24862306a36Sopenharmony_ci nlmsg_trim(skb, b); 24962306a36Sopenharmony_ci return -1; 25062306a36Sopenharmony_ci} 25162306a36Sopenharmony_ci 25262306a36Sopenharmony_cistatic void tcf_psample_group_put(void *priv) 25362306a36Sopenharmony_ci{ 25462306a36Sopenharmony_ci struct psample_group *group = priv; 25562306a36Sopenharmony_ci 25662306a36Sopenharmony_ci psample_group_put(group); 25762306a36Sopenharmony_ci} 25862306a36Sopenharmony_ci 25962306a36Sopenharmony_cistatic struct psample_group * 26062306a36Sopenharmony_citcf_sample_get_group(const struct tc_action *a, 26162306a36Sopenharmony_ci tc_action_priv_destructor *destructor) 26262306a36Sopenharmony_ci{ 26362306a36Sopenharmony_ci struct tcf_sample *s = to_sample(a); 26462306a36Sopenharmony_ci struct psample_group *group; 26562306a36Sopenharmony_ci 26662306a36Sopenharmony_ci group = rcu_dereference_protected(s->psample_group, 26762306a36Sopenharmony_ci lockdep_is_held(&s->tcf_lock)); 26862306a36Sopenharmony_ci if (group) { 26962306a36Sopenharmony_ci psample_group_take(group); 27062306a36Sopenharmony_ci *destructor = tcf_psample_group_put; 27162306a36Sopenharmony_ci } 27262306a36Sopenharmony_ci 27362306a36Sopenharmony_ci return group; 27462306a36Sopenharmony_ci} 27562306a36Sopenharmony_ci 27662306a36Sopenharmony_cistatic void tcf_offload_sample_get_group(struct flow_action_entry *entry, 27762306a36Sopenharmony_ci const struct tc_action *act) 27862306a36Sopenharmony_ci{ 27962306a36Sopenharmony_ci entry->sample.psample_group = 28062306a36Sopenharmony_ci act->ops->get_psample_group(act, &entry->destructor); 28162306a36Sopenharmony_ci entry->destructor_priv = entry->sample.psample_group; 28262306a36Sopenharmony_ci} 28362306a36Sopenharmony_ci 28462306a36Sopenharmony_cistatic int tcf_sample_offload_act_setup(struct tc_action *act, void *entry_data, 28562306a36Sopenharmony_ci u32 *index_inc, bool bind, 28662306a36Sopenharmony_ci struct netlink_ext_ack *extack) 28762306a36Sopenharmony_ci{ 28862306a36Sopenharmony_ci if (bind) { 28962306a36Sopenharmony_ci struct flow_action_entry *entry = entry_data; 29062306a36Sopenharmony_ci 29162306a36Sopenharmony_ci entry->id = FLOW_ACTION_SAMPLE; 29262306a36Sopenharmony_ci entry->sample.trunc_size = tcf_sample_trunc_size(act); 29362306a36Sopenharmony_ci entry->sample.truncate = tcf_sample_truncate(act); 29462306a36Sopenharmony_ci entry->sample.rate = tcf_sample_rate(act); 29562306a36Sopenharmony_ci tcf_offload_sample_get_group(entry, act); 29662306a36Sopenharmony_ci *index_inc = 1; 29762306a36Sopenharmony_ci } else { 29862306a36Sopenharmony_ci struct flow_offload_action *fl_action = entry_data; 29962306a36Sopenharmony_ci 30062306a36Sopenharmony_ci fl_action->id = FLOW_ACTION_SAMPLE; 30162306a36Sopenharmony_ci } 30262306a36Sopenharmony_ci 30362306a36Sopenharmony_ci return 0; 30462306a36Sopenharmony_ci} 30562306a36Sopenharmony_ci 30662306a36Sopenharmony_cistatic struct tc_action_ops act_sample_ops = { 30762306a36Sopenharmony_ci .kind = "sample", 30862306a36Sopenharmony_ci .id = TCA_ID_SAMPLE, 30962306a36Sopenharmony_ci .owner = THIS_MODULE, 31062306a36Sopenharmony_ci .act = tcf_sample_act, 31162306a36Sopenharmony_ci .stats_update = tcf_sample_stats_update, 31262306a36Sopenharmony_ci .dump = tcf_sample_dump, 31362306a36Sopenharmony_ci .init = tcf_sample_init, 31462306a36Sopenharmony_ci .cleanup = tcf_sample_cleanup, 31562306a36Sopenharmony_ci .get_psample_group = tcf_sample_get_group, 31662306a36Sopenharmony_ci .offload_act_setup = tcf_sample_offload_act_setup, 31762306a36Sopenharmony_ci .size = sizeof(struct tcf_sample), 31862306a36Sopenharmony_ci}; 31962306a36Sopenharmony_ci 32062306a36Sopenharmony_cistatic __net_init int sample_init_net(struct net *net) 32162306a36Sopenharmony_ci{ 32262306a36Sopenharmony_ci struct tc_action_net *tn = net_generic(net, act_sample_ops.net_id); 32362306a36Sopenharmony_ci 32462306a36Sopenharmony_ci return tc_action_net_init(net, tn, &act_sample_ops); 32562306a36Sopenharmony_ci} 32662306a36Sopenharmony_ci 32762306a36Sopenharmony_cistatic void __net_exit sample_exit_net(struct list_head *net_list) 32862306a36Sopenharmony_ci{ 32962306a36Sopenharmony_ci tc_action_net_exit(net_list, act_sample_ops.net_id); 33062306a36Sopenharmony_ci} 33162306a36Sopenharmony_ci 33262306a36Sopenharmony_cistatic struct pernet_operations sample_net_ops = { 33362306a36Sopenharmony_ci .init = sample_init_net, 33462306a36Sopenharmony_ci .exit_batch = sample_exit_net, 33562306a36Sopenharmony_ci .id = &act_sample_ops.net_id, 33662306a36Sopenharmony_ci .size = sizeof(struct tc_action_net), 33762306a36Sopenharmony_ci}; 33862306a36Sopenharmony_ci 33962306a36Sopenharmony_cistatic int __init sample_init_module(void) 34062306a36Sopenharmony_ci{ 34162306a36Sopenharmony_ci return tcf_register_action(&act_sample_ops, &sample_net_ops); 34262306a36Sopenharmony_ci} 34362306a36Sopenharmony_ci 34462306a36Sopenharmony_cistatic void __exit sample_cleanup_module(void) 34562306a36Sopenharmony_ci{ 34662306a36Sopenharmony_ci tcf_unregister_action(&act_sample_ops, &sample_net_ops); 34762306a36Sopenharmony_ci} 34862306a36Sopenharmony_ci 34962306a36Sopenharmony_cimodule_init(sample_init_module); 35062306a36Sopenharmony_cimodule_exit(sample_cleanup_module); 35162306a36Sopenharmony_ci 35262306a36Sopenharmony_ciMODULE_AUTHOR("Yotam Gigi <yotam.gi@gmail.com>"); 35362306a36Sopenharmony_ciMODULE_DESCRIPTION("Packet sampling action"); 35462306a36Sopenharmony_ciMODULE_LICENSE("GPL v2"); 355