net/netfilter/ipvs: Fix data-race in ip_vs_add_service / ip_vs_out_hook
Subsystems (4)
PATCH 1/1 net/netfilter/ipvs: Fix data-race in ip_vs_add_service / ip_vs_out_hook
A data-race was detected by KCSAN between ip_vs_add_service() which acts as a writer, and ip_vs_out_hook() which acts as a reader. This can lead to unpredictable behavior and crashes. One observed symptom is the "no destination available" error when processing packets.
The race occurs on the enable flag within the netns_ipvs struct. This flag was being written in the configuration path without any protection, while concurrently being read in the packet processing path. This lack of synchronization means a reader on one CPU could see a partially initialized service, leading to incorrect behavior.
To fix this, convert the enable flag from a plain integer to an atomic_t. This ensures that all reads and writes to the flag are atomic. More importantly, using atomic_set() and atomic_read() provides the necessary memory barriers to guarantee that changes to other fields of the service are visible to the reader CPU before the service is marked as enabled.
--- include/net/ip_vs.h | 2 +- net/netfilter/ipvs/ip_vs_conn.c | 4 ++-- net/netfilter/ipvs/ip_vs_core.c | 10 +++++----- net/netfilter/ipvs/ip_vs_ctl.c | 6 +++--- net/netfilter/ipvs/ip_vs_est.c | 16 ++++++++-------- 5 files changed, 19 insertions(+), 19 deletions(-)
include/net/ip_vs.h +1 -1
@@ -895,7 +895,7 @@ struct ipvs_sync_daemon_cfg {/* IPVS in network namespace */ struct netns_ipvs {int gen;/* Generation */ -int enable;/* enable like nf_hooks do */ +atomic_t enable;/* enable like nf_hooks do */ /* Hash table: for real service lookups */ #define IP_VS_RTAB_BITS 4#define IP_VS_RTAB_SIZE (1 << IP_VS_RTAB_BITS)
net/netfilter/ipvs/ip_vs_conn.c +2 -2
@@ -885,7 +885,7 @@ static void ip_vs_conn_expire(struct timer_list *t)* conntrack cleanup for the net. */ smp_rmb (); -if (ipvs->enable) +if (atomic_read (&ipvs->enable))ip_vs_conn_drop_conntrack (cp); }
@@ -1439,7 +1439,7 @@ void ip_vs_expire_nodest_conn_flush(struct netns_ipvs *ipvs)cond_resched_rcu ();/* netns clean up started, abort delayed work */ -if (!ipvs->enable) +if (!atomic_read (&ipvs->enable))break ; }rcu_read_unlock ();
net/netfilter/ipvs/ip_vs_core.c +5 -5
@@ -1353,7 +1353,7 @@ ip_vs_out_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statif (unlikely (!skb_dst (skb)))return NF_ACCEPT ; -if (!ipvs->enable) +if (!atomic_read (&ipvs->enable))return NF_ACCEPT ;ip_vs_fill_iph_skb (af, skb, false, &iph);
@@ -1940,7 +1940,7 @@ ip_vs_in_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statereturn NF_ACCEPT ; }/* ipvs enabled in this netns ? */ -if (unlikely (sysctl_backup_only (ipvs) || !ipvs->enable)) +if (unlikely (sysctl_backup_only (ipvs) || !atomic_read (&ipvs->enable)))return NF_ACCEPT ;ip_vs_fill_iph_skb (af, skb, false, &iph);
@@ -2108,7 +2108,7 @@ ip_vs_forward_icmp(void *priv, struct sk_buff *skb,int r;/* ipvs enabled in this netns ? */ -if (unlikely (sysctl_backup_only (ipvs) || !ipvs->enable)) +if (unlikely (sysctl_backup_only (ipvs) || !atomic_read (&ipvs->enable)))return NF_ACCEPT ;if (state->pf ==NFPROTO_IPV4 ) {
@@ -2295,7 +2295,7 @@ static int __net_init __ip_vs_init(struct net *net)return -ENOMEM ;/* Hold the beast until a service is registered */ - ipvs->enable =0 ; +atomic_set (&ipvs->enable,0 ); ipvs->net = net;/* Counters used for creating unique names */ ipvs->gen =atomic_read (&ipvs_netns_cnt);
@@ -2367,7 +2367,7 @@ static void __net_exit __ip_vs_dev_cleanup_batch(struct list_head *net_list) ipvs =net_ipvs (net);ip_vs_unregister_hooks (ipvs,AF_INET );ip_vs_unregister_hooks (ipvs,AF_INET6 ); - ipvs->enable =0 ;/* Disable packet reception */ +atomic_set (&ipvs->enable,0 );/* Disable packet reception */ smp_wmb ();ip_vs_sync_net_cleanup (ipvs); }
net/netfilter/ipvs/ip_vs_ctl.c +3 -3
@@ -256,7 +256,7 @@ static void est_reload_work_handler(struct work_struct *work)struct ip_vs_est_kt_data *kd = ipvs->est_kt_arr[id];/* netns clean up started, abort delayed work */ -if (!ipvs->enable) +if (!atomic_read (&ipvs->enable)) goto unlock;if (!kd)continue ;
@@ -1483,9 +1483,9 @@ ip_vs_add_service(struct netns_ipvs *ipvs, struct ip_vs_service_user_kern *u, *svc_p = svc; -if (!ipvs->enable) { +if (!atomic_read (&ipvs->enable)) {/* Now there is a service - full throttle */ - ipvs->enable =1 ; +atomic_set (&ipvs->enable,1 );/* Start estimation for first time */ ip_vs_est_reload_start (ipvs);
net/netfilter/ipvs/ip_vs_est.c +8 -8
@@ -231,7 +231,7 @@ static int ip_vs_estimation_kthread(void *data)void ip_vs_est_reload_start (struct netns_ipvs *ipvs) {/* Ignore reloads before first service is added */ -if (!ipvs->enable) +if (!atomic_read (&ipvs->enable))return ;ip_vs_est_stopped_recalc (ipvs);/* Bump the kthread configuration genid */
@@ -306,7 +306,7 @@ static int ip_vs_est_add_kthread(struct netns_ipvs *ipvs)int i;if ((unsigned long )ipvs->est_kt_count >= ipvs->est_max_threads && - ipvs->enable && ipvs->est_max_threads) +atomic_read (&ipvs->enable) && ipvs->est_max_threads)return -EINVAL ;mutex_lock (&ipvs->est_mutex);
@@ -343,7 +343,7 @@ static int ip_vs_est_add_kthread(struct netns_ipvs *ipvs) }/* Start kthread tasks only when services are present */ -if (ipvs->enable && !ip_vs_est_stopped (ipvs)) { +if (atomic_read (&ipvs->enable) && !ip_vs_est_stopped (ipvs)) { ret =ip_vs_est_kthread_start (ipvs, kd);if (ret <0 ) goto out;
@@ -486,7 +486,7 @@ int ip_vs_start_estimator(struct netns_ipvs *ipvs, struct ip_vs_stats *stats)struct ip_vs_estimator *est = &stats->est;int ret; -if (!ipvs->est_max_threads && ipvs->enable) +if (!ipvs->est_max_threads &&atomic_read (&ipvs->enable)) ipvs->est_max_threads =ip_vs_est_max_threads (ipvs); est->ktid =-1 ;
@@ -663,7 +663,7 @@ static int ip_vs_est_calc_limits(struct netns_ipvs *ipvs, int *chain_max)/* Wait for cpufreq frequency transition */ wait_event_idle_timeout (wq,kthread_should_stop (),HZ /50 ); -if (!ipvs->enable ||kthread_should_stop ()) +if (!atomic_read (&ipvs->enable) ||kthread_should_stop ()) goto stop; }
@@ -681,7 +681,7 @@ static int ip_vs_est_calc_limits(struct netns_ipvs *ipvs, int *chain_max)rcu_read_unlock ();local_bh_enable (); -if (!ipvs->enable ||kthread_should_stop ()) +if (!atomic_read (&ipvs->enable) ||kthread_should_stop ()) goto stop;cond_resched ();
@@ -757,7 +757,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs)mutex_lock (&ipvs->est_mutex);for (id =1 ; id < ipvs->est_kt_count; id++) {/* netns clean up started, abort */ -if (!ipvs->enable) +if (!atomic_read (&ipvs->enable)) goto unlock2; kd = ipvs->est_kt_arr[id];if (!kd)
@@ -787,7 +787,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs) id = ipvs->est_kt_count; next_kt: -if (!ipvs->enable ||kthread_should_stop ()) +if (!atomic_read (&ipvs->enable) ||kthread_should_stop ()) goto unlock; id--;if (id <0 )
Signature · 1 line
2.47.3
Florian Westphal
maintainerQuoted 3 lines · Zhang Tengfei
A data-race was detected by KCSAN between
ip_vs_add_service()which acts as a writer, andip_vs_out_hook()which acts as a reader. This can lead to unpredictable behavior and crashes.
Really? How can this cause a crash?
Quoted 13 lines
The race occurs on the
enableflag within thenetns_ipvsstruct. This flag was being written in the configuration path without any protection, while concurrently being read in the packet processing path. This lack of synchronization means a reader on one CPU could see a partially initialized service, leading to incorrect behavior.To fix this, convert the
enableflag from a plain integer to anatomic_t. This ensures that all reads and writes to the flag are atomic. More importantly, usingatomic_set()andatomic_read()provides the necessary memory barriers to guarantee that changes to other fields of the service are visible to the reader CPU before the service is marked as enabled.
include/net/ip_vs.h struct ipvs_sync_daemon_cfg { · line 898
Quoted 2 lines
include/net/ip_vs.h +1
-1-int enable;/* enable like nf_hooks do */ +atomic_t enable;/* enable like nf_hooks do */
Julian, Simon, I will defer to your judgment but I dislike this, because I see no reason for atomic_t. To me is seems better to use READ/WRITE_ONCE as ->enable is only ever set but not modified (increment for instance).
Quoted 7 lines · Florian Westphal
Quoted 3 lines · Zhang Tengfei
A data-race was detected by KCSAN between
ip_vs_add_service()which acts as a writer, andip_vs_out_hook()which acts as a reader. This can lead to unpredictable behavior and crashes.Really? How can this cause a crash?
KCSAN + panic_on_warn=1 : Only in debug environment
Quoted 14 lines
Quoted 13 lines
The race occurs on the
enableflag within thenetns_ipvsstruct. This flag was being written in the configuration path without any protection, while concurrently being read in the packet processing path. This lack of synchronization means a reader on one CPU could see a partially initialized service, leading to incorrect behavior.To fix this, convert the
enableflag from a plain integer to anatomic_t. This ensures that all reads and writes to the flag are atomic. More importantly, usingatomic_set()andatomic_read()provides the necessary memory barriers to guarantee that changes to other fields of the service are visible to the reader CPU before the service is marked as enabled.
Quoted 7 lines
include/net/ip_vs.h struct ipvs_sync_daemon_cfg { · line 898
Quoted 2 lines
include/net/ip_vs.h +1
-1-int enable;/* enable like nf_hooks do */ +atomic_t enable;/* enable like nf_hooks do */ Julian, Simon, I will defer to your judgment but I dislike this, because I see no reason for
atomic_t. To me is seems better to use READ/WRITE_ONCE as ->enable is only ever set but not modified (increment for instance).
+2
Julian Anastasov
maintainerHello,
Quoted 4 lines · Zhang Tengfei
A data-race was detected by KCSAN between
ip_vs_add_service()which acts as a writer, andip_vs_out_hook()which acts as a reader. This can lead to unpredictable behavior and crashes. One observed symptom is the "no destination available" error when processing packets.
You can not solve the "no destination available" race
in IPVS itself. When service is added in ip_vs_add_service() things happen in some order:
- hooks are registered (if first service)
- service is registered
- enable flag is set (if first service)
All this is part of single IP_VS_SO_SET_ADD call.
You can reorder the above actions as you wish but without any gain because the dests (real servers) are added with a following IP_VS_SO_SET_ADDDEST call. There will be always a gap between the both calls where packets can hit service without any dests, with enable=0 or 1.
One can stop the traffic with firewall rules until
all IPVS rules are added. It is decided by user space tools when to route the traffic via IPVS (after or during configuration).
Quoted 6 lines
The race occurs on the
enableflag within thenetns_ipvsstruct. This flag was being written in the configuration path without any protection, while concurrently being read in the packet processing path. This lack of synchronization means a reader on one CPU could see a partially initialized service, leading to incorrect behavior.
No, service is added with hlist_add_head_rcu() inip_vs_svc_hash() which commits all writes before adding svc to list. This is also an answer for the concerns in your paragraph below.
If "partially initialized" refers to the missing
dests, then see above, they are added later with IP_VS_SO_SET_ADDDEST.
Quoted 6 lines
To fix this, convert the
enableflag from a plain integer to anatomic_t. This ensures that all reads and writes to the flag are atomic. More importantly, usingatomic_set()andatomic_read()provides the necessary memory barriers to guarantee that changes to other fields of the service are visible to the reader CPU before the service is marked as enabled.
We use RCU for proper svc registration.
Quoted 29 lines
Reported-by [email protected]Signed-off-by Zhang Tengfei <[email protected]>--- include/net/ip_vs.h | 2 +- net/netfilter/ipvs/ip_vs_conn.c | 4 ++-- net/netfilter/ipvs/ip_vs_core.c | 10 +++++----- net/netfilter/ipvs/ip_vs_ctl.c | 6 +++--- net/netfilter/ipvs/ip_vs_est.c | 16 ++++++++-------- 5 files changed, 19 insertions(+), 19 deletions(-)include/net/ip_vs.h +1
-1@@ -895,7 +895,7 @@ struct ipvs_sync_daemon_cfg {/* IPVS in network namespace */ struct netns_ipvs {int gen;/* Generation */ -int enable;/* enable like nf_hooks do */ +atomic_t enable;/* enable like nf_hooks do */ /* Hash table: for real service lookups */ #define IP_VS_RTAB_BITS 4#define IP_VS_RTAB_SIZE (1 << IP_VS_RTAB_BITS)net/netfilter/ipvs/ip_vs_conn.c +0
-0
net/netfilter/ipvs/ip_vs_conn.c static void ip_vs_conn_expire(struct timer_list *t) · line 888
Quoted 5 lines
net/netfilter/ipvs/ip_vs_conn.c +1
-1@@ -885,7 +885,7 @@ static void ip_vs_conn_expire(struct timer_list *t)* conntrack cleanup for the net. */ smp_rmb (); -if (ipvs->enable) +if (atomic_read (&ipvs->enable))
This place is a valid user of this flag.
This is one of the reasons we should keep such flag.
Quoted 17 lines
net/netfilter/ipvs/ip_vs_conn.c +1
-1ip_vs_conn_drop_conntrack (cp); }@@ -1439,7 +1439,7 @@ void ip_vs_expire_nodest_conn_flush(struct netns_ipvs *ipvs)cond_resched_rcu ();/* netns clean up started, abort delayed work */ -if (!ipvs->enable) +if (!atomic_read (&ipvs->enable))break ; }rcu_read_unlock ();net/netfilter/ipvs/ip_vs_core.c +0
-0
net/netfilter/ipvs/ip_vs_core.c ip_vs_out_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *stat · line 1356
Quoted 4 lines
net/netfilter/ipvs/ip_vs_core.c +0
-1@@ -1353,7 +1353,7 @@ ip_vs_out_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statif (unlikely (!skb_dst (skb)))return NF_ACCEPT ; -if (!ipvs->enable)
The checks in the hooks are not needed anymore after
commit 857ca89711de ("ipvs: register hooks only with services") because the hooks are registered when the first service is added.
So, you can not see enable=0 in hooks anymore which
was possible before first service was added. Or it is possible for small time between adding the hooks and setting the flag but it is irrelevant because there are no dests yet.
Quoted 32 lines
net/netfilter/ipvs/ip_vs_core.c +4
-3+if (!atomic_read (&ipvs->enable))return NF_ACCEPT ;ip_vs_fill_iph_skb (af, skb, false, &iph);@@ -1940,7 +1940,7 @@ ip_vs_in_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statereturn NF_ACCEPT ; }/* ipvs enabled in this netns ? */ -if (unlikely (sysctl_backup_only (ipvs) || !ipvs->enable)) +if (unlikely (sysctl_backup_only (ipvs) || !atomic_read (&ipvs->enable)))return NF_ACCEPT ;ip_vs_fill_iph_skb (af, skb, false, &iph);@@ -2108,7 +2108,7 @@ ip_vs_forward_icmp(void *priv, struct sk_buff *skb,int r;/* ipvs enabled in this netns ? */ -if (unlikely (sysctl_backup_only (ipvs) || !ipvs->enable)) +if (unlikely (sysctl_backup_only (ipvs) || !atomic_read (&ipvs->enable)))return NF_ACCEPT ;if (state->pf ==NFPROTO_IPV4 ) {@@ -2295,7 +2295,7 @@ static int __net_init __ip_vs_init(struct net *net)return -ENOMEM ;/* Hold the beast until a service is registered */ - ipvs->enable =0 ; +atomic_set (&ipvs->enable,0 ); ipvs->net = net;/* Counters used for creating unique names */ ipvs->gen =atomic_read (&ipvs_netns_cnt);
net/netfilter/ipvs/ip_vs_core.c static void __net_exit __ip_vs_dev_cleanup_batch(struct list_head *net_list) · line 2369
Quoted 3 lines
net/netfilter/ipvs/ip_vs_core.c +0
-0@@ -2367,7 +2367,7 @@ static void __net_exit __ip_vs_dev_cleanup_batch(struct list_head *net_list) ipvs =net_ipvs (net);ip_vs_unregister_hooks (ipvs,AF_INET );ip_vs_unregister_hooks (ipvs,AF_INET6 );
No new/flying packets after this point...
net/netfilter/ipvs/ip_vs_core.c static void __net_exit __ip_vs_dev_cleanup_batch(struct list_head *net_list) · line 2370
Quoted 1 line
net/netfilter/ipvs/ip_vs_core.c +0
-1- ipvs->enable =0 ;/* Disable packet reception */
It was an old method to disable reception but now
the flag helps for other purposes.
net/netfilter/ipvs/ip_vs_ctl.c static void est_reload_work_handler(struct work_struct *work) · line 259
Quoted 13 lines
net/netfilter/ipvs/ip_vs_core.c +1
-0+atomic_set (&ipvs->enable,0 );/* Disable packet reception */ smp_wmb ();ip_vs_sync_net_cleanup (ipvs); }net/netfilter/ipvs/ip_vs_ctl.c +0
-1@@ -256,7 +256,7 @@ static void est_reload_work_handler(struct work_struct *work)struct ip_vs_est_kt_data *kd = ipvs->est_kt_arr[id];/* netns clean up started, abort delayed work */ -if (!ipvs->enable)
Such checks in slow path just speedup the cleanup.
Quoted 85 lines
net/netfilter/ipvs/ip_vs_ctl.c +3
-2+if (!atomic_read (&ipvs->enable)) goto unlock;if (!kd)continue ;@@ -1483,9 +1483,9 @@ ip_vs_add_service(struct netns_ipvs *ipvs, struct ip_vs_service_user_kern *u, *svc_p = svc; -if (!ipvs->enable) { +if (!atomic_read (&ipvs->enable)) {/* Now there is a service - full throttle */ - ipvs->enable =1 ; +atomic_set (&ipvs->enable,1 );/* Start estimation for first time */ ip_vs_est_reload_start (ipvs);net/netfilter/ipvs/ip_vs_est.c +7
-7@@ -231,7 +231,7 @@ static int ip_vs_estimation_kthread(void *data)void ip_vs_est_reload_start (struct netns_ipvs *ipvs) {/* Ignore reloads before first service is added */ -if (!ipvs->enable) +if (!atomic_read (&ipvs->enable))return ;ip_vs_est_stopped_recalc (ipvs);/* Bump the kthread configuration genid */ @@ -306,7 +306,7 @@ static int ip_vs_est_add_kthread(struct netns_ipvs *ipvs)int i;if ((unsigned long )ipvs->est_kt_count >= ipvs->est_max_threads && - ipvs->enable && ipvs->est_max_threads) +atomic_read (&ipvs->enable) && ipvs->est_max_threads)return -EINVAL ;mutex_lock (&ipvs->est_mutex);@@ -343,7 +343,7 @@ static int ip_vs_est_add_kthread(struct netns_ipvs *ipvs) }/* Start kthread tasks only when services are present */ -if (ipvs->enable && !ip_vs_est_stopped (ipvs)) { +if (atomic_read (&ipvs->enable) && !ip_vs_est_stopped (ipvs)) { ret =ip_vs_est_kthread_start (ipvs, kd);if (ret <0 ) goto out;@@ -486,7 +486,7 @@ int ip_vs_start_estimator(struct netns_ipvs *ipvs, struct ip_vs_stats *stats)struct ip_vs_estimator *est = &stats->est;int ret; -if (!ipvs->est_max_threads && ipvs->enable) +if (!ipvs->est_max_threads &&atomic_read (&ipvs->enable)) ipvs->est_max_threads =ip_vs_est_max_threads (ipvs); est->ktid =-1 ;@@ -663,7 +663,7 @@ static int ip_vs_est_calc_limits(struct netns_ipvs *ipvs, int *chain_max)/* Wait for cpufreq frequency transition */ wait_event_idle_timeout (wq,kthread_should_stop (),HZ /50 ); -if (!ipvs->enable ||kthread_should_stop ()) +if (!atomic_read (&ipvs->enable) ||kthread_should_stop ()) goto stop; }@@ -681,7 +681,7 @@ static int ip_vs_est_calc_limits(struct netns_ipvs *ipvs, int *chain_max)rcu_read_unlock ();local_bh_enable (); -if (!ipvs->enable ||kthread_should_stop ()) +if (!atomic_read (&ipvs->enable) ||kthread_should_stop ()) goto stop;cond_resched ();@@ -757,7 +757,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs)mutex_lock (&ipvs->est_mutex);for (id =1 ; id < ipvs->est_kt_count; id++) {/* netns clean up started, abort */ -if (!ipvs->enable) +if (!atomic_read (&ipvs->enable)) goto unlock2; kd = ipvs->est_kt_arr[id];if (!kd)
net/netfilter/ipvs/ip_vs_est.c static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs) · line 793
Quoted 8 lines
net/netfilter/ipvs/ip_vs_est.c +1
-1@@ -787,7 +787,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs) id = ipvs->est_kt_count; next_kt: -if (!ipvs->enable ||kthread_should_stop ()) +if (!atomic_read (&ipvs->enable) ||kthread_should_stop ()) goto unlock; id--;if (id <0 )
In summary, the checks in fast path (in hooks) are
useless/obsolete while the other checks in slow path do not need atomic operations. Such races are normal to happen because service and its dests are not added in same atomic transaction.
Regards
--
Julian Anastasov <[email protected]>
Julian Anastasov
maintainerHello,
Quoted 34 lines · Zhang Tengfei
A data-race was detected by KCSAN between
ip_vs_add_service()which acts as a writer, andip_vs_out_hook()which acts as a reader. This can lead to unpredictable behavior and crashes. One observed symptom is the "no destination available" error when processing packets.The race occurs on the
enableflag within thenetns_ipvsstruct. This flag was being written in the configuration path without any protection, while concurrently being read in the packet processing path. This lack of synchronization means a reader on one CPU could see a partially initialized service, leading to incorrect behavior.To fix this, convert the
enableflag from a plain integer to anatomic_t. This ensures that all reads and writes to the flag are atomic. More importantly, usingatomic_set()andatomic_read()provides the necessary memory barriers to guarantee that changes to other fields of the service are visible to the reader CPU before the service is marked as enabled.Reported-by [email protected]Signed-off-by Zhang Tengfei <[email protected]>--- include/net/ip_vs.h | 2 +- net/netfilter/ipvs/ip_vs_conn.c | 4 ++-- net/netfilter/ipvs/ip_vs_core.c | 10 +++++----- net/netfilter/ipvs/ip_vs_ctl.c | 6 +++--- net/netfilter/ipvs/ip_vs_est.c | 16 ++++++++-------- 5 files changed, 19 insertions(+), 19 deletions(-)net/netfilter/ipvs/ip_vs_est.c +0
-0
...
net/netfilter/ipvs/ip_vs_est.c static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs) · line 761
Quoted 7 lines
net/netfilter/ipvs/ip_vs_est.c +1
-1@@ -757,7 +757,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs)mutex_lock (&ipvs->est_mutex);for (id =1 ; id < ipvs->est_kt_count; id++) {/* netns clean up started, abort */ -if (!ipvs->enable) +if (!atomic_read (&ipvs->enable)) goto unlock2;
It is a simple flag but as it is checked in loops
in a few places in ip_vs_est.c, lets use READ_ONCE/WRITE_ONCE as suggested by Florian and Eric. The 3 checks in hooks in ip_vs_core.c can be simply removed: in ip_vs_out_hook, ip_vs_in_hook and ip_vs_forward_icmp. We can see enable=0 in rare cases which is not fatal. It is a flying packet in two possible cases:
1. after hooks are registered but before the flag is set
2. after the hooks are unregistered on cleanup_net
Regards
--
Julian Anastasov <[email protected]>
Hi everyone,
Here is the v2 patch that incorporates the feedback.
Many thanks to Julian for his thorough review and for providing the detailed plan for this new version, and thanks to Florian and Eric for suggestions.
Subject: [PATCH v2] net/netfilter/ipvs: Use READ_ONCE/WRITE_ONCE for
ipvs->enable
KCSAN reported a data-race on the ipvs->enable flag, which is written in the control path and read concurrently from many other contexts.
Following a suggestion by Julian, this patch fixes the race by converting all accesses to use WRITE_ONCE()/READ_ONCE(). This lightweight approach ensures atomic access and acts as a compiler barrier, preventing unsafe optimizations where the flag is checked in loops (e.g., in ip_vs_est.c).
Additionally, the now-obsolete enable checks in the fast path hooks (ip_vs_in_hook, ip_vs_out_hook, ip_vs_forward_icmp) are removed. These are unnecessary since commit 857ca89711de ("ipvs: register hooks only with services").
---
v2:
- Switched from atomic_t to the suggested READ_ONCE()/WRITE_ONCE().
- Removed obsolete checks from the packet processing hooks.
- Polished commit message based on feedback.
--- net/netfilter/ipvs/ip_vs_conn.c | 4 ++-- net/netfilter/ipvs/ip_vs_core.c | 11 ++++------- net/netfilter/ipvs/ip_vs_ctl.c | 6 +++--- net/netfilter/ipvs/ip_vs_est.c | 16 ++++++++-------- 4 files changed, 17 insertions(+), 20 deletions(-)
net/netfilter/ipvs/ip_vs_conn.c +2 -2
@@ -885,7 +885,7 @@ static void ip_vs_conn_expire(struct timer_list *t)* conntrack cleanup for the net. */ smp_rmb (); -if (ipvs->enable) +if (READ_ONCE (ipvs->enable))ip_vs_conn_drop_conntrack (cp); }
@@ -1439,7 +1439,7 @@ void ip_vs_expire_nodest_conn_flush(struct netns_ipvs *ipvs)cond_resched_rcu ();/* netns clean up started, abort delayed work */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable))break ; }rcu_read_unlock ();
net/netfilter/ipvs/ip_vs_core.c +4 -7
@@ -1353,9 +1353,6 @@ ip_vs_out_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statif (unlikely (!skb_dst (skb)))return NF_ACCEPT ; -if (!ipvs->enable) -return NF_ACCEPT ; -ip_vs_fill_iph_skb (af, skb, false, &iph);#ifdef CONFIG_IP_VS_IPV6 if (af ==AF_INET6 ) {
@@ -1940,7 +1937,7 @@ ip_vs_in_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statereturn NF_ACCEPT ; }/* ipvs enabled in this netns ? */ -if (unlikely (sysctl_backup_only (ipvs) || !ipvs->enable)) +if (unlikely (sysctl_backup_only (ipvs)))return NF_ACCEPT ;ip_vs_fill_iph_skb (af, skb, false, &iph);
@@ -2108,7 +2105,7 @@ ip_vs_forward_icmp(void *priv, struct sk_buff *skb,int r;/* ipvs enabled in this netns ? */ -if (unlikely (sysctl_backup_only (ipvs) || !ipvs->enable)) +if (unlikely (sysctl_backup_only (ipvs)))return NF_ACCEPT ;if (state->pf ==NFPROTO_IPV4 ) {
@@ -2295,7 +2292,7 @@ static int __net_init __ip_vs_init(struct net *net)return -ENOMEM ;/* Hold the beast until a service is registered */ - ipvs->enable =0 ; +WRITE_ONCE (ipvs->enable,0 ); ipvs->net = net;/* Counters used for creating unique names */ ipvs->gen =atomic_read (&ipvs_netns_cnt);
@@ -2367,7 +2364,7 @@ static void __net_exit __ip_vs_dev_cleanup_batch(struct list_head *net_list) ipvs =net_ipvs (net);ip_vs_unregister_hooks (ipvs,AF_INET );ip_vs_unregister_hooks (ipvs,AF_INET6 ); - ipvs->enable =0 ;/* Disable packet reception */ +WRITE_ONCE (ipvs->enable,0 );/* Disable packet reception */ smp_wmb ();ip_vs_sync_net_cleanup (ipvs); }
net/netfilter/ipvs/ip_vs_ctl.c +3 -3
@@ -256,7 +256,7 @@ static void est_reload_work_handler(struct work_struct *work)struct ip_vs_est_kt_data *kd = ipvs->est_kt_arr[id];/* netns clean up started, abort delayed work */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable)) goto unlock;if (!kd)continue ;
@@ -1483,9 +1483,9 @@ ip_vs_add_service(struct netns_ipvs *ipvs, struct ip_vs_service_user_kern *u, *svc_p = svc; -if (!ipvs->enable) { +if (!READ_ONCE (ipvs->enable)) {/* Now there is a service - full throttle */ - ipvs->enable =1 ; +WRITE_ONCE (ipvs->enable,1 );/* Start estimation for first time */ ip_vs_est_reload_start (ipvs);
net/netfilter/ipvs/ip_vs_est.c +8 -8
@@ -231,7 +231,7 @@ static int ip_vs_estimation_kthread(void *data)void ip_vs_est_reload_start (struct netns_ipvs *ipvs) {/* Ignore reloads before first service is added */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable))return ;ip_vs_est_stopped_recalc (ipvs);/* Bump the kthread configuration genid */
@@ -306,7 +306,7 @@ static int ip_vs_est_add_kthread(struct netns_ipvs *ipvs)int i;if ((unsigned long )ipvs->est_kt_count >= ipvs->est_max_threads && - ipvs->enable && ipvs->est_max_threads) +READ_ONCE (ipvs->enable) && ipvs->est_max_threads)return -EINVAL ;mutex_lock (&ipvs->est_mutex);
@@ -343,7 +343,7 @@ static int ip_vs_est_add_kthread(struct netns_ipvs *ipvs) }/* Start kthread tasks only when services are present */ -if (ipvs->enable && !ip_vs_est_stopped (ipvs)) { +if (READ_ONCE (ipvs->enable) && !ip_vs_est_stopped (ipvs)) { ret =ip_vs_est_kthread_start (ipvs, kd);if (ret <0 ) goto out;
@@ -486,7 +486,7 @@ int ip_vs_start_estimator(struct netns_ipvs *ipvs, struct ip_vs_stats *stats)struct ip_vs_estimator *est = &stats->est;int ret; -if (!ipvs->est_max_threads && ipvs->enable) +if (!ipvs->est_max_threads &&READ_ONCE (ipvs->enable)) ipvs->est_max_threads =ip_vs_est_max_threads (ipvs); est->ktid =-1 ;
@@ -663,7 +663,7 @@ static int ip_vs_est_calc_limits(struct netns_ipvs *ipvs, int *chain_max)/* Wait for cpufreq frequency transition */ wait_event_idle_timeout (wq,kthread_should_stop (),HZ /50 ); -if (!ipvs->enable ||kthread_should_stop ()) +if (!READ_ONCE (ipvs->enable) ||kthread_should_stop ()) goto stop; }
@@ -681,7 +681,7 @@ static int ip_vs_est_calc_limits(struct netns_ipvs *ipvs, int *chain_max)rcu_read_unlock ();local_bh_enable (); -if (!ipvs->enable ||kthread_should_stop ()) +if (!READ_ONCE (ipvs->enable) ||kthread_should_stop ()) goto stop;cond_resched ();
@@ -757,7 +757,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs)mutex_lock (&ipvs->est_mutex);for (id =1 ; id < ipvs->est_kt_count; id++) {/* netns clean up started, abort */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable)) goto unlock2; kd = ipvs->est_kt_arr[id];if (!kd)
@@ -787,7 +787,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs) id = ipvs->est_kt_count; next_kt: -if (!ipvs->enable ||kthread_should_stop ()) +if (!READ_ONCE (ipvs->enable) ||kthread_should_stop ()) goto unlock; id--;if (id <0 )
Signature · 1 line
2.34.1
Pablo Neira Ayuso
maintainerRe:
Quoted 3 lines · Zhang Tengfei
Hi everyone,
Here is the v2 patch that incorporates the feedback.
Patch without subject will not fly too far, I'm afraid you will have to resubmit. One more comment below.
Quoted 39 lines
Many thanks to Julian for his thorough review and for providing the detailed plan for this new version, and thanks to Florian and Eric for suggestions.
Subject: [PATCH v2]
net/netfilter/ipvs: Use READ_ONCE/WRITE_ONCE for
ipvs->enableKCSAN reported a data-race on the
ipvs->enableflag, which is written in the control path and read concurrently from many other contexts.Following a suggestion by Julian, this patch fixes the race by converting all accesses to use
WRITE_ONCE()/READ_ONCE(). This lightweight approach ensures atomic access and acts as a compiler barrier, preventing unsafe optimizations where the flag is checked in loops (e.g., inip_vs_est.c).Additionally, the now-obsolete
enablechecks in the fast path hooks (ip_vs_in_hook,ip_vs_out_hook,ip_vs_forward_icmp) are removed. These are unnecessary since commit857ca89711de("ipvs: register hooks only with services").Reported-by [email protected]Suggested-by Julian Anastasov <[email protected]>Signed-off-by Zhang Tengfei <[email protected]>---
v2:
- Switched fromatomic_tto the suggested READ_ONCE()/WRITE_ONCE().
- Removed obsolete checks from the packet processing hooks.
- Polished commit message based on feedback.--- net/netfilter/ipvs/ip_vs_conn.c | 4 ++-- net/netfilter/ipvs/ip_vs_core.c | 11 ++++------- net/netfilter/ipvs/ip_vs_ctl.c | 6 +++--- net/netfilter/ipvs/ip_vs_est.c | 16 ++++++++-------- 4 files changed, 17 insertions(+), 20 deletions(-)
[...]
net/netfilter/ipvs/ip_vs_core.c ip_vs_out_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *stat · line 1357
Quoted 10 lines
net/netfilter/ipvs/ip_vs_core.c +0
-2@@ -1353,9 +1353,6 @@ ip_vs_out_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statif (unlikely (!skb_dst (skb)))return NF_ACCEPT ; -if (!ipvs->enable) -return NF_ACCEPT ;
Patch does say why is this going away? If you think this is not necessary, then make a separated patch and example why this is needed?
Thanks
Quoted 3 lines
net/netfilter/ipvs/ip_vs_core.c +0
-0ip_vs_fill_iph_skb (af, skb, false, &iph);#ifdef CONFIG_IP_VS_IPV6 if (af ==AF_INET6 ) {
PATCH v2 1/1 net/netfilter/ipvs: Use READ_ONCE/WRITE_ONCE for ipvs->enable
In reply to Julian AnastasovKCSAN reported a data-race on the ipvs->enable flag, which is written in the control path and read concurrently from many other contexts.
Following a suggestion by Julian, this patch fixes the race by converting all accesses to use WRITE_ONCE()/READ_ONCE(). This lightweight approach ensures atomic access and acts as a compiler barrier, preventing unsafe optimizations where the flag is checked in loops (e.g., in ip_vs_est.c).
Additionally, the enable checks in the fast-path hooks (ip_vs_in_hook, ip_vs_out_hook, ip_vs_forward_icmp) are removed. They are considered unnecessary because the enable=0 condition they check for can only occur in two rare and non-fatal
scenarios: 1) after hooks are registered but before the flag is set, and 2) after hooks are unregistered on cleanup_net. In the worst case, a single packet might be mishandled (e.g., dropped), which does not lead to a system crash or data corruption. Adding a check in the performance-critical fast-path to handle this harmless condition is not a worthwhile trade-off.
---
v2:
- Switched from atomic_t to the suggested READ_ONCE()/WRITE_ONCE().
- Removed obsolete checks from the packet processing hooks.
- Polished commit message based on feedback from maintainers.
--- net/netfilter/ipvs/ip_vs_conn.c | 4 ++-- net/netfilter/ipvs/ip_vs_core.c | 11 ++++------- net/netfilter/ipvs/ip_vs_ctl.c | 6 +++--- net/netfilter/ipvs/ip_vs_est.c | 16 ++++++++-------- 4 files changed, 17 insertions(+), 20 deletions(-)
net/netfilter/ipvs/ip_vs_conn.c +2 -2
@@ -885,7 +885,7 @@ static void ip_vs_conn_expire(struct timer_list *t)* conntrack cleanup for the net. */ smp_rmb (); -if (ipvs->enable) +if (READ_ONCE (ipvs->enable))ip_vs_conn_drop_conntrack (cp); }
@@ -1439,7 +1439,7 @@ void ip_vs_expire_nodest_conn_flush(struct netns_ipvs *ipvs)cond_resched_rcu ();/* netns clean up started, abort delayed work */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable))break ; }rcu_read_unlock ();
net/netfilter/ipvs/ip_vs_core.c +4 -7
@@ -1353,9 +1353,6 @@ ip_vs_out_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statif (unlikely (!skb_dst (skb)))return NF_ACCEPT ; -if (!ipvs->enable) -return NF_ACCEPT ; -ip_vs_fill_iph_skb (af, skb, false, &iph);#ifdef CONFIG_IP_VS_IPV6 if (af ==AF_INET6 ) {
@@ -1940,7 +1937,7 @@ ip_vs_in_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statereturn NF_ACCEPT ; }/* ipvs enabled in this netns ? */ -if (unlikely (sysctl_backup_only (ipvs) || !ipvs->enable)) +if (unlikely (sysctl_backup_only (ipvs)))return NF_ACCEPT ;ip_vs_fill_iph_skb (af, skb, false, &iph);
@@ -2108,7 +2105,7 @@ ip_vs_forward_icmp(void *priv, struct sk_buff *skb,int r;/* ipvs enabled in this netns ? */ -if (unlikely (sysctl_backup_only (ipvs) || !ipvs->enable)) +if (unlikely (sysctl_backup_only (ipvs)))return NF_ACCEPT ;if (state->pf ==NFPROTO_IPV4 ) {
@@ -2295,7 +2292,7 @@ static int __net_init __ip_vs_init(struct net *net)return -ENOMEM ;/* Hold the beast until a service is registered */ - ipvs->enable =0 ; +WRITE_ONCE (ipvs->enable,0 ); ipvs->net = net;/* Counters used for creating unique names */ ipvs->gen =atomic_read (&ipvs_netns_cnt);
@@ -2367,7 +2364,7 @@ static void __net_exit __ip_vs_dev_cleanup_batch(struct list_head *net_list) ipvs =net_ipvs (net);ip_vs_unregister_hooks (ipvs,AF_INET );ip_vs_unregister_hooks (ipvs,AF_INET6 ); - ipvs->enable =0 ;/* Disable packet reception */ +WRITE_ONCE (ipvs->enable,0 );/* Disable packet reception */ smp_wmb ();ip_vs_sync_net_cleanup (ipvs); }
net/netfilter/ipvs/ip_vs_ctl.c +3 -3
@@ -256,7 +256,7 @@ static void est_reload_work_handler(struct work_struct *work)struct ip_vs_est_kt_data *kd = ipvs->est_kt_arr[id];/* netns clean up started, abort delayed work */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable)) goto unlock;if (!kd)continue ;
@@ -1483,9 +1483,9 @@ ip_vs_add_service(struct netns_ipvs *ipvs, struct ip_vs_service_user_kern *u, *svc_p = svc; -if (!ipvs->enable) { +if (!READ_ONCE (ipvs->enable)) {/* Now there is a service - full throttle */ - ipvs->enable =1 ; +WRITE_ONCE (ipvs->enable,1 );/* Start estimation for first time */ ip_vs_est_reload_start (ipvs);
net/netfilter/ipvs/ip_vs_est.c +8 -8
@@ -231,7 +231,7 @@ static int ip_vs_estimation_kthread(void *data)void ip_vs_est_reload_start (struct netns_ipvs *ipvs) {/* Ignore reloads before first service is added */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable))return ;ip_vs_est_stopped_recalc (ipvs);/* Bump the kthread configuration genid */
@@ -306,7 +306,7 @@ static int ip_vs_est_add_kthread(struct netns_ipvs *ipvs)int i;if ((unsigned long )ipvs->est_kt_count >= ipvs->est_max_threads && - ipvs->enable && ipvs->est_max_threads) +READ_ONCE (ipvs->enable) && ipvs->est_max_threads)return -EINVAL ;mutex_lock (&ipvs->est_mutex);
@@ -343,7 +343,7 @@ static int ip_vs_est_add_kthread(struct netns_ipvs *ipvs) }/* Start kthread tasks only when services are present */ -if (ipvs->enable && !ip_vs_est_stopped (ipvs)) { +if (READ_ONCE (ipvs->enable) && !ip_vs_est_stopped (ipvs)) { ret =ip_vs_est_kthread_start (ipvs, kd);if (ret <0 ) goto out;
@@ -486,7 +486,7 @@ int ip_vs_start_estimator(struct netns_ipvs *ipvs, struct ip_vs_stats *stats)struct ip_vs_estimator *est = &stats->est;int ret; -if (!ipvs->est_max_threads && ipvs->enable) +if (!ipvs->est_max_threads &&READ_ONCE (ipvs->enable)) ipvs->est_max_threads =ip_vs_est_max_threads (ipvs); est->ktid =-1 ;
@@ -663,7 +663,7 @@ static int ip_vs_est_calc_limits(struct netns_ipvs *ipvs, int *chain_max)/* Wait for cpufreq frequency transition */ wait_event_idle_timeout (wq,kthread_should_stop (),HZ /50 ); -if (!ipvs->enable ||kthread_should_stop ()) +if (!READ_ONCE (ipvs->enable) ||kthread_should_stop ()) goto stop; }
@@ -681,7 +681,7 @@ static int ip_vs_est_calc_limits(struct netns_ipvs *ipvs, int *chain_max)rcu_read_unlock ();local_bh_enable (); -if (!ipvs->enable ||kthread_should_stop ()) +if (!READ_ONCE (ipvs->enable) ||kthread_should_stop ()) goto stop;cond_resched ();
@@ -757,7 +757,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs)mutex_lock (&ipvs->est_mutex);for (id =1 ; id < ipvs->est_kt_count; id++) {/* netns clean up started, abort */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable)) goto unlock2; kd = ipvs->est_kt_arr[id];if (!kd)
@@ -787,7 +787,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs) id = ipvs->est_kt_count; next_kt: -if (!ipvs->enable ||kthread_should_stop ()) +if (!READ_ONCE (ipvs->enable) ||kthread_should_stop ()) goto unlock; id--;if (id <0 )
Signature · 1 line
2.34.1
Re: [PATCH v2] net/netfilter/ipvs: Use READ_ONCE/WRITE_ONCE for ipvs->enable
Hi everyone,
My apologies for the noise and confusion with the multiple v2 submissions. I sent a malformed reply earlier with no subject.
This patch in *this* thread is the correct and final version for review. Please disregard the previous malformed email.
Many thanks to Pablo Neira Ayuso for pointing out the formatting error.
Julian Anastasov
maintainerRe: [PATCH v2] net/netfilter/ipvs: Use READ_ONCE/WRITE_ONCE for ipvs->enable
In reply to Zhang TengfeiHello,
Quoted 10 lines · Zhang Tengfei
KCSAN reported a data-race on the
ipvs->enableflag, which is written in the control path and read concurrently from many other contexts.Following a suggestion by Julian, this patch fixes the race by converting all accesses to use
WRITE_ONCE()/READ_ONCE(). This lightweight approach ensures atomic access and acts as a compiler barrier, preventing unsafe optimizations where the flag is checked in loops (e.g., inip_vs_est.c).
Quoted 3 lines
Additionally, the
enablechecks in the fast-path hooks (ip_vs_in_hook,ip_vs_out_hook,ip_vs_forward_icmp) are removed. They are considered unnecessary because theenable=0
It was good idea to mention about the 857ca89711de
commit here as in the previous v2 version. You can even add it as Fixes tag as suggested here:
scripts/checkpatch.pl --strict /tmp/file.patch
As for the Subject line, you probably can use
[PATCH v3 nf-next] ipvs: Use READ_ONCE/WRITE_ONCE for ipvs->enable to specify the desired target tree ('nf-next' or 'nf' if such data-race needs it).
As for the patch code, it looks ok.
Quoted 14 lines
condition they check for can only occur in two rare and non-fatal
scenarios: 1) after hooks are registered but before the flag is set, and 2) after hooks are unregistered oncleanup_net. In the worst case, a single packet might be mishandled (e.g., dropped), which does not lead to a system crash or data corruption. Adding a check in the performance-critical fast-path to handle this harmless condition is not a worthwhile trade-off.Reported-by [email protected]Suggested-by Julian Anastasov <[email protected]>Signed-off-by Zhang Tengfei <[email protected]>
Quoted 5 lines
---
v2:
- Switched fromatomic_tto the suggested READ_ONCE()/WRITE_ONCE().
- Removed obsolete checks from the packet processing hooks.
- Polished commit message based on feedback from maintainers.
...
Regards
--
Julian Anastasov <[email protected]>
PATCH v3 1/1 ipvs: Use READ_ONCE/WRITE_ONCE for ipvs->enable
KCSAN reported a data-race on the ipvs->enable flag, which is written in the control path and read concurrently from many other contexts.
Following a suggestion by Julian, this patch fixes the race by converting all accesses to use WRITE_ONCE()/READ_ONCE(). This lightweight approach ensures atomic access and acts as a compiler barrier, preventing unsafe optimizations where the flag is checked in loops (e.g., in ip_vs_est.c).
Additionally, the enable checks in the fast-path hooks (ip_vs_in_hook, ip_vs_out_hook, ip_vs_forward_icmp) are removed. These are unnecessary since commit 857ca89711de ("ipvs: register hooks only with services"). The enable=0 condition they check for can only occur in two rare and non-fatal
scenarios: 1) after hooks are registered but before the flag is set, and 2) after hooks are unregistered on cleanup_net. In the worst case, a single packet might be mishandled (e.g., dropped), which does not lead to a system crash or data corruption. Adding a check in the performance-critical fast-path to handle this harmless condition is not a worthwhile trade-off.
---
v3:
- Restore reference to commit 857ca89711de in commit message.
- Add corresponding Fixes tag.
v2:
- Switched from atomic_t to the suggested READ_ONCE()/WRITE_ONCE().
- Removed obsolete checks from the packet processing hooks.
- Polished commit message based on feedback from maintainers.
--- net/netfilter/ipvs/ip_vs_conn.c | 4 ++-- net/netfilter/ipvs/ip_vs_core.c | 11 ++++------- net/netfilter/ipvs/ip_vs_ctl.c | 6 +++--- net/netfilter/ipvs/ip_vs_est.c | 16 ++++++++-------- 4 files changed, 17 insertions(+), 20 deletions(-)
net/netfilter/ipvs/ip_vs_conn.c +2 -2
@@ -885,7 +885,7 @@ static void ip_vs_conn_expire(struct timer_list *t)* conntrack cleanup for the net. */ smp_rmb (); -if (ipvs->enable) +if (READ_ONCE (ipvs->enable))ip_vs_conn_drop_conntrack (cp); }
@@ -1439,7 +1439,7 @@ void ip_vs_expire_nodest_conn_flush(struct netns_ipvs *ipvs)cond_resched_rcu ();/* netns clean up started, abort delayed work */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable))break ; }rcu_read_unlock ();
net/netfilter/ipvs/ip_vs_core.c +4 -7
@@ -1353,9 +1353,6 @@ ip_vs_out_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statif (unlikely (!skb_dst (skb)))return NF_ACCEPT ; -if (!ipvs->enable) -return NF_ACCEPT ; -ip_vs_fill_iph_skb (af, skb, false, &iph);#ifdef CONFIG_IP_VS_IPV6 if (af ==AF_INET6 ) {
@@ -1940,7 +1937,7 @@ ip_vs_in_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statereturn NF_ACCEPT ; }/* ipvs enabled in this netns ? */ -if (unlikely (sysctl_backup_only (ipvs) || !ipvs->enable)) +if (unlikely (sysctl_backup_only (ipvs)))return NF_ACCEPT ;ip_vs_fill_iph_skb (af, skb, false, &iph);
@@ -2108,7 +2105,7 @@ ip_vs_forward_icmp(void *priv, struct sk_buff *skb,int r;/* ipvs enabled in this netns ? */ -if (unlikely (sysctl_backup_only (ipvs) || !ipvs->enable)) +if (unlikely (sysctl_backup_only (ipvs)))return NF_ACCEPT ;if (state->pf ==NFPROTO_IPV4 ) {
@@ -2295,7 +2292,7 @@ static int __net_init __ip_vs_init(struct net *net)return -ENOMEM ;/* Hold the beast until a service is registered */ - ipvs->enable =0 ; +WRITE_ONCE (ipvs->enable,0 ); ipvs->net = net;/* Counters used for creating unique names */ ipvs->gen =atomic_read (&ipvs_netns_cnt);
@@ -2367,7 +2364,7 @@ static void __net_exit __ip_vs_dev_cleanup_batch(struct list_head *net_list) ipvs =net_ipvs (net);ip_vs_unregister_hooks (ipvs,AF_INET );ip_vs_unregister_hooks (ipvs,AF_INET6 ); - ipvs->enable =0 ;/* Disable packet reception */ +WRITE_ONCE (ipvs->enable,0 );/* Disable packet reception */ smp_wmb ();ip_vs_sync_net_cleanup (ipvs); }
net/netfilter/ipvs/ip_vs_ctl.c +3 -3
@@ -256,7 +256,7 @@ static void est_reload_work_handler(struct work_struct *work)struct ip_vs_est_kt_data *kd = ipvs->est_kt_arr[id];/* netns clean up started, abort delayed work */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable)) goto unlock;if (!kd)continue ;
@@ -1483,9 +1483,9 @@ ip_vs_add_service(struct netns_ipvs *ipvs, struct ip_vs_service_user_kern *u, *svc_p = svc; -if (!ipvs->enable) { +if (!READ_ONCE (ipvs->enable)) {/* Now there is a service - full throttle */ - ipvs->enable =1 ; +WRITE_ONCE (ipvs->enable,1 );/* Start estimation for first time */ ip_vs_est_reload_start (ipvs);
net/netfilter/ipvs/ip_vs_est.c +8 -8
@@ -231,7 +231,7 @@ static int ip_vs_estimation_kthread(void *data)void ip_vs_est_reload_start (struct netns_ipvs *ipvs) {/* Ignore reloads before first service is added */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable))return ;ip_vs_est_stopped_recalc (ipvs);/* Bump the kthread configuration genid */
@@ -306,7 +306,7 @@ static int ip_vs_est_add_kthread(struct netns_ipvs *ipvs)int i;if ((unsigned long )ipvs->est_kt_count >= ipvs->est_max_threads && - ipvs->enable && ipvs->est_max_threads) +READ_ONCE (ipvs->enable) && ipvs->est_max_threads)return -EINVAL ;mutex_lock (&ipvs->est_mutex);
@@ -343,7 +343,7 @@ static int ip_vs_est_add_kthread(struct netns_ipvs *ipvs) }/* Start kthread tasks only when services are present */ -if (ipvs->enable && !ip_vs_est_stopped (ipvs)) { +if (READ_ONCE (ipvs->enable) && !ip_vs_est_stopped (ipvs)) { ret =ip_vs_est_kthread_start (ipvs, kd);if (ret <0 ) goto out;
@@ -486,7 +486,7 @@ int ip_vs_start_estimator(struct netns_ipvs *ipvs, struct ip_vs_stats *stats)struct ip_vs_estimator *est = &stats->est;int ret; -if (!ipvs->est_max_threads && ipvs->enable) +if (!ipvs->est_max_threads &&READ_ONCE (ipvs->enable)) ipvs->est_max_threads =ip_vs_est_max_threads (ipvs); est->ktid =-1 ;
@@ -663,7 +663,7 @@ static int ip_vs_est_calc_limits(struct netns_ipvs *ipvs, int *chain_max)/* Wait for cpufreq frequency transition */ wait_event_idle_timeout (wq,kthread_should_stop (),HZ /50 ); -if (!ipvs->enable ||kthread_should_stop ()) +if (!READ_ONCE (ipvs->enable) ||kthread_should_stop ()) goto stop; }
@@ -681,7 +681,7 @@ static int ip_vs_est_calc_limits(struct netns_ipvs *ipvs, int *chain_max)rcu_read_unlock ();local_bh_enable (); -if (!ipvs->enable ||kthread_should_stop ()) +if (!READ_ONCE (ipvs->enable) ||kthread_should_stop ()) goto stop;cond_resched ();
@@ -757,7 +757,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs)mutex_lock (&ipvs->est_mutex);for (id =1 ; id < ipvs->est_kt_count; id++) {/* netns clean up started, abort */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable)) goto unlock2; kd = ipvs->est_kt_arr[id];if (!kd)
@@ -787,7 +787,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs) id = ipvs->est_kt_count; next_kt: -if (!ipvs->enable ||kthread_should_stop ()) +if (!READ_ONCE (ipvs->enable) ||kthread_should_stop ()) goto unlock; id--;if (id <0 )
Signature · 1 line
2.34.1
Re: [PATCH v3 nf-next] ipvs: Use READ_ONCE/WRITE_ONCE for ipvs->enable
Hello,
Quoted 22 lines · Zhang Tengfei
KCSAN reported a data-race on the
ipvs->enableflag, which is written in the control path and read concurrently from many other contexts.Following a suggestion by Julian, this patch fixes the race by converting all accesses to use
WRITE_ONCE()/READ_ONCE(). This lightweight approach ensures atomic access and acts as a compiler barrier, preventing unsafe optimizations where the flag is checked in loops (e.g., inip_vs_est.c).Additionally, the
enablechecks in the fast-path hooks (ip_vs_in_hook,ip_vs_out_hook,ip_vs_forward_icmp) are removed. These are unnecessary since commit857ca89711de("ipvs: register hooks only with services"). Theenable=0condition they check for can only occur in two rare and non-fatal
scenarios: 1) after hooks are registered but before the flag is set, and 2) after hooks are unregistered oncleanup_net. In the worst case, a single packet might be mishandled (e.g., dropped), which does not lead to a system crash or data corruption. Adding a check in the performance-critical fast-path to handle this harmless condition is not a worthwhile trade-off.
Quoted 6 lines
Fixes 857ca89711de ("ipvs: register hooks only with services")Reported-by [email protected]Suggested-by Julian Anastasov <[email protected]>Signed-off-by Zhang Tengfei <[email protected]>
Looks good to me, thanks!
Quoted 190 lines
---
v3:
- Restore reference to commit857ca89711dein commit message.
- Add corresponding Fixes tag.
v2:
- Switched fromatomic_tto the suggested READ_ONCE()/WRITE_ONCE().
- Removed obsolete checks from the packet processing hooks.
- Polished commit message based on feedback from maintainers.--- net/netfilter/ipvs/ip_vs_conn.c | 4 ++-- net/netfilter/ipvs/ip_vs_core.c | 11 ++++------- net/netfilter/ipvs/ip_vs_ctl.c | 6 +++--- net/netfilter/ipvs/ip_vs_est.c | 16 ++++++++-------- 4 files changed, 17 insertions(+), 20 deletions(-)net/netfilter/ipvs/ip_vs_conn.c +2
-2@@ -885,7 +885,7 @@ static void ip_vs_conn_expire(struct timer_list *t)* conntrack cleanup for the net. */ smp_rmb (); -if (ipvs->enable) +if (READ_ONCE (ipvs->enable))ip_vs_conn_drop_conntrack (cp); }@@ -1439,7 +1439,7 @@ void ip_vs_expire_nodest_conn_flush(struct netns_ipvs *ipvs)cond_resched_rcu ();/* netns clean up started, abort delayed work */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable))break ; }rcu_read_unlock ();net/netfilter/ipvs/ip_vs_core.c +4
-7@@ -1353,9 +1353,6 @@ ip_vs_out_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statif (unlikely (!skb_dst (skb)))return NF_ACCEPT ; -if (!ipvs->enable) -return NF_ACCEPT ; -ip_vs_fill_iph_skb (af, skb, false, &iph);#ifdef CONFIG_IP_VS_IPV6 if (af ==AF_INET6 ) {@@ -1940,7 +1937,7 @@ ip_vs_in_hook(void *priv, struct sk_buff *skb, const struct nf_hook_state *statereturn NF_ACCEPT ; }/* ipvs enabled in this netns ? */ -if (unlikely (sysctl_backup_only (ipvs) || !ipvs->enable)) +if (unlikely (sysctl_backup_only (ipvs)))return NF_ACCEPT ;ip_vs_fill_iph_skb (af, skb, false, &iph);@@ -2108,7 +2105,7 @@ ip_vs_forward_icmp(void *priv, struct sk_buff *skb,int r;/* ipvs enabled in this netns ? */ -if (unlikely (sysctl_backup_only (ipvs) || !ipvs->enable)) +if (unlikely (sysctl_backup_only (ipvs)))return NF_ACCEPT ;if (state->pf ==NFPROTO_IPV4 ) {@@ -2295,7 +2292,7 @@ static int __net_init __ip_vs_init(struct net *net)return -ENOMEM ;/* Hold the beast until a service is registered */ - ipvs->enable =0 ; +WRITE_ONCE (ipvs->enable,0 ); ipvs->net = net;/* Counters used for creating unique names */ ipvs->gen =atomic_read (&ipvs_netns_cnt);@@ -2367,7 +2364,7 @@ static void __net_exit __ip_vs_dev_cleanup_batch(struct list_head *net_list) ipvs =net_ipvs (net);ip_vs_unregister_hooks (ipvs,AF_INET );ip_vs_unregister_hooks (ipvs,AF_INET6 ); - ipvs->enable =0 ;/* Disable packet reception */ +WRITE_ONCE (ipvs->enable,0 );/* Disable packet reception */ smp_wmb ();ip_vs_sync_net_cleanup (ipvs); }net/netfilter/ipvs/ip_vs_ctl.c +3
-3@@ -256,7 +256,7 @@ static void est_reload_work_handler(struct work_struct *work)struct ip_vs_est_kt_data *kd = ipvs->est_kt_arr[id];/* netns clean up started, abort delayed work */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable)) goto unlock;if (!kd)continue ;@@ -1483,9 +1483,9 @@ ip_vs_add_service(struct netns_ipvs *ipvs, struct ip_vs_service_user_kern *u, *svc_p = svc; -if (!ipvs->enable) { +if (!READ_ONCE (ipvs->enable)) {/* Now there is a service - full throttle */ - ipvs->enable =1 ; +WRITE_ONCE (ipvs->enable,1 );/* Start estimation for first time */ ip_vs_est_reload_start (ipvs);net/netfilter/ipvs/ip_vs_est.c +8
-8@@ -231,7 +231,7 @@ static int ip_vs_estimation_kthread(void *data)void ip_vs_est_reload_start (struct netns_ipvs *ipvs) {/* Ignore reloads before first service is added */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable))return ;ip_vs_est_stopped_recalc (ipvs);/* Bump the kthread configuration genid */ @@ -306,7 +306,7 @@ static int ip_vs_est_add_kthread(struct netns_ipvs *ipvs)int i;if ((unsigned long )ipvs->est_kt_count >= ipvs->est_max_threads && - ipvs->enable && ipvs->est_max_threads) +READ_ONCE (ipvs->enable) && ipvs->est_max_threads)return -EINVAL ;mutex_lock (&ipvs->est_mutex);@@ -343,7 +343,7 @@ static int ip_vs_est_add_kthread(struct netns_ipvs *ipvs) }/* Start kthread tasks only when services are present */ -if (ipvs->enable && !ip_vs_est_stopped (ipvs)) { +if (READ_ONCE (ipvs->enable) && !ip_vs_est_stopped (ipvs)) { ret =ip_vs_est_kthread_start (ipvs, kd);if (ret <0 ) goto out;@@ -486,7 +486,7 @@ int ip_vs_start_estimator(struct netns_ipvs *ipvs, struct ip_vs_stats *stats)struct ip_vs_estimator *est = &stats->est;int ret; -if (!ipvs->est_max_threads && ipvs->enable) +if (!ipvs->est_max_threads &&READ_ONCE (ipvs->enable)) ipvs->est_max_threads =ip_vs_est_max_threads (ipvs); est->ktid =-1 ;@@ -663,7 +663,7 @@ static int ip_vs_est_calc_limits(struct netns_ipvs *ipvs, int *chain_max)/* Wait for cpufreq frequency transition */ wait_event_idle_timeout (wq,kthread_should_stop (),HZ /50 ); -if (!ipvs->enable ||kthread_should_stop ()) +if (!READ_ONCE (ipvs->enable) ||kthread_should_stop ()) goto stop; }@@ -681,7 +681,7 @@ static int ip_vs_est_calc_limits(struct netns_ipvs *ipvs, int *chain_max)rcu_read_unlock ();local_bh_enable (); -if (!ipvs->enable ||kthread_should_stop ()) +if (!READ_ONCE (ipvs->enable) ||kthread_should_stop ()) goto stop;cond_resched ();@@ -757,7 +757,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs)mutex_lock (&ipvs->est_mutex);for (id =1 ; id < ipvs->est_kt_count; id++) {/* netns clean up started, abort */ -if (!ipvs->enable) +if (!READ_ONCE (ipvs->enable)) goto unlock2; kd = ipvs->est_kt_arr[id];if (!kd)@@ -787,7 +787,7 @@ static void ip_vs_est_calc_phase(struct netns_ipvs *ipvs) id = ipvs->est_kt_count; next_kt: -if (!ipvs->enable ||kthread_should_stop ()) +if (!READ_ONCE (ipvs->enable) ||kthread_should_stop ()) goto unlock; id--;if (id <0 )Signature · 1 line
2.34.1
Regards
--
Julian Anastasov <[email protected]>