* [RFC PATCH] netfilter: call synchronize_net only once from nf_register_net_hooks
@ 2017-11-22 10:40 Giuseppe Scrivano
2017-11-22 10:41 ` Florian Westphal
2017-11-22 11:06 ` Florian Westphal
0 siblings, 2 replies; 8+ messages in thread
From: Giuseppe Scrivano @ 2017-11-22 10:40 UTC (permalink / raw)
To: netfilter-devel
SELinux, if enabled, registers for each new network namespace 6
netfilter hooks. Avoid to use synchronize_net for each new hook, but do
it once after all the hooks are added. The net benefit on an SMP
machine with two cores is that creating a new network namespace takes
-40% of the original time.
Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
---
net/netfilter/core.c | 15 ++++++++++++---
1 file changed, 12 insertions(+), 3 deletions(-)
diff --git a/net/netfilter/core.c b/net/netfilter/core.c
index 52cd2901a097..beeb0b36f429 100644
--- a/net/netfilter/core.c
+++ b/net/netfilter/core.c
@@ -252,7 +252,7 @@ static struct nf_hook_entries __rcu **nf_hook_entry_head(struct net *net, const
return NULL;
}
-int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
+static int __nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
{
struct nf_hook_entries *p, *new_hooks;
struct nf_hook_entries __rcu **pp;
@@ -291,11 +291,19 @@ int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
#ifdef HAVE_JUMP_LABEL
static_key_slow_inc(&nf_hooks_needed[reg->pf][reg->hooknum]);
#endif
- synchronize_net();
BUG_ON(p == new_hooks);
kvfree(p);
return 0;
}
+
+int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
+{
+ int ret = __nf_register_net_hook(net, reg);
+ if (ret < 0)
+ return ret;
+ synchronize_net();
+ return 0;
+}
EXPORT_SYMBOL(nf_register_net_hook);
/*
@@ -379,10 +387,11 @@ int nf_register_net_hooks(struct net *net, const struct nf_hook_ops *reg,
int err = 0;
for (i = 0; i < n; i++) {
- err = nf_register_net_hook(net, ®[i]);
+ err = __nf_register_net_hook(net, ®[i]);
if (err)
goto err;
}
+ synchronize_net();
return err;
err:
--
2.14.3
^ permalink raw reply related [flat|nested] 8+ messages in thread
* Re: [RFC PATCH] netfilter: call synchronize_net only once from nf_register_net_hooks
2017-11-22 10:40 [RFC PATCH] netfilter: call synchronize_net only once from nf_register_net_hooks Giuseppe Scrivano
@ 2017-11-22 10:41 ` Florian Westphal
2017-11-22 11:06 ` Florian Westphal
1 sibling, 0 replies; 8+ messages in thread
From: Florian Westphal @ 2017-11-22 10:41 UTC (permalink / raw)
To: Giuseppe Scrivano; +Cc: netfilter-devel
Giuseppe Scrivano <gscrivan@redhat.com> wrote:
> SELinux, if enabled, registers for each new network namespace 6
> netfilter hooks. Avoid to use synchronize_net for each new hook, but do
> it once after all the hooks are added. The net benefit on an SMP
> machine with two cores is that creating a new network namespace takes
> -40% of the original time.
>
> Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
> ---
> net/netfilter/core.c | 15 ++++++++++++---
> 1 file changed, 12 insertions(+), 3 deletions(-)
>
> diff --git a/net/netfilter/core.c b/net/netfilter/core.c
> index 52cd2901a097..beeb0b36f429 100644
> --- a/net/netfilter/core.c
> +++ b/net/netfilter/core.c
> @@ -252,7 +252,7 @@ static struct nf_hook_entries __rcu **nf_hook_entry_head(struct net *net, const
> return NULL;
> }
>
> -int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
> +static int __nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
> {
> struct nf_hook_entries *p, *new_hooks;
> struct nf_hook_entries __rcu **pp;
> @@ -291,11 +291,19 @@ int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
> #ifdef HAVE_JUMP_LABEL
> static_key_slow_inc(&nf_hooks_needed[reg->pf][reg->hooknum]);
> #endif
> - synchronize_net();
> BUG_ON(p == new_hooks);
> kvfree(p);
This adds use after free?
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [RFC PATCH] netfilter: call synchronize_net only once from nf_register_net_hooks
2017-11-22 10:40 [RFC PATCH] netfilter: call synchronize_net only once from nf_register_net_hooks Giuseppe Scrivano
2017-11-22 10:41 ` Florian Westphal
@ 2017-11-22 11:06 ` Florian Westphal
2017-11-22 11:10 ` Florian Westphal
2017-11-22 11:25 ` Giuseppe Scrivano
1 sibling, 2 replies; 8+ messages in thread
From: Florian Westphal @ 2017-11-22 11:06 UTC (permalink / raw)
To: Giuseppe Scrivano; +Cc: netfilter-devel
Giuseppe Scrivano <gscrivan@redhat.com> wrote:
> SELinux, if enabled, registers for each new network namespace 6
> netfilter hooks. Avoid to use synchronize_net for each new hook, but do
> it once after all the hooks are added. The net benefit on an SMP
> machine with two cores is that creating a new network namespace takes
> -40% of the original time.
but this needs more work.
> Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
> ---
> net/netfilter/core.c | 15 ++++++++++++---
> 1 file changed, 12 insertions(+), 3 deletions(-)
>
> diff --git a/net/netfilter/core.c b/net/netfilter/core.c
> index 52cd2901a097..beeb0b36f429 100644
> --- a/net/netfilter/core.c
> +++ b/net/netfilter/core.c
> @@ -252,7 +252,7 @@ static struct nf_hook_entries __rcu **nf_hook_entry_head(struct net *net, const
> return NULL;
> }
>
> -int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
> +static int __nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
Change this to return struct nf_hook_entries *
> {
> struct nf_hook_entries *p, *new_hooks;
> struct nf_hook_entries __rcu **pp;
> @@ -291,11 +291,19 @@ int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
> #ifdef HAVE_JUMP_LABEL
> static_key_slow_inc(&nf_hooks_needed[reg->pf][reg->hooknum]);
> #endif
> - synchronize_net();
> BUG_ON(p == new_hooks);
> kvfree(p);
remove kvfree()
> return 0;
return p;
> +int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
> +{
> + int ret = __nf_register_net_hook(net, reg);
> + if (ret < 0)
> + return ret;
> + synchronize_net();
then free p here.
> + return 0;
> +}
> EXPORT_SYMBOL(nf_register_net_hook);
>
> /*
> @@ -379,10 +387,11 @@ int nf_register_net_hooks(struct net *net, const struct nf_hook_ops *reg,
> int err = 0;
>
> for (i = 0; i < n; i++) {
> - err = nf_register_net_hook(net, ®[i]);
> + err = __nf_register_net_hook(net, ®[i]);
here, store the old 'p' values and then
free them after synchronize_net().
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [RFC PATCH] netfilter: call synchronize_net only once from nf_register_net_hooks
2017-11-22 11:06 ` Florian Westphal
@ 2017-11-22 11:10 ` Florian Westphal
2017-11-22 11:25 ` Giuseppe Scrivano
1 sibling, 0 replies; 8+ messages in thread
From: Florian Westphal @ 2017-11-22 11:10 UTC (permalink / raw)
To: Florian Westphal; +Cc: Giuseppe Scrivano, netfilter-devel
Florian Westphal <fw@strlen.de> wrote:
> Giuseppe Scrivano <gscrivan@redhat.com> wrote:
> > SELinux, if enabled, registers for each new network namespace 6
> > netfilter hooks. Avoid to use synchronize_net for each new hook, but do
> > it once after all the hooks are added. The net benefit on an SMP
> > machine with two cores is that creating a new network namespace takes
> > -40% of the original time.
>
> but this needs more work.
>
> > Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
> > ---
> > net/netfilter/core.c | 15 ++++++++++++---
> > 1 file changed, 12 insertions(+), 3 deletions(-)
> >
> > diff --git a/net/netfilter/core.c b/net/netfilter/core.c
> > index 52cd2901a097..beeb0b36f429 100644
> > --- a/net/netfilter/core.c
> > +++ b/net/netfilter/core.c
> > @@ -252,7 +252,7 @@ static struct nf_hook_entries __rcu **nf_hook_entry_head(struct net *net, const
> > return NULL;
> > }
> >
> > -int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
> > +static int __nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
>
> Change this to return struct nf_hook_entries *
>
> > {
> > struct nf_hook_entries *p, *new_hooks;
> > struct nf_hook_entries __rcu **pp;
> > @@ -291,11 +291,19 @@ int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
> > #ifdef HAVE_JUMP_LABEL
> > static_key_slow_inc(&nf_hooks_needed[reg->pf][reg->hooknum]);
> > #endif
> > - synchronize_net();
> > BUG_ON(p == new_hooks);
> > kvfree(p);
>
> remove kvfree()
>
> > return 0;
>
> return p;
>
> > +int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
> > +{
> > + int ret = __nf_register_net_hook(net, reg);
> > + if (ret < 0)
> > + return ret;
> > + synchronize_net();
>
> then free p here.
Yet another alternative is to place an rcu_head at the end of
struct nf_hook_entries.
This is a bit tricky because the location is not fixed due to
dynamically sized array members.
But if you do this you can use call_rcu to do the freeing which removes
need for all but the nfqueue related synchronize_net() calls.
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [RFC PATCH] netfilter: call synchronize_net only once from nf_register_net_hooks
2017-11-22 11:06 ` Florian Westphal
2017-11-22 11:10 ` Florian Westphal
@ 2017-11-22 11:25 ` Giuseppe Scrivano
2017-11-22 11:56 ` Florian Westphal
1 sibling, 1 reply; 8+ messages in thread
From: Giuseppe Scrivano @ 2017-11-22 11:25 UTC (permalink / raw)
To: Florian Westphal; +Cc: netfilter-devel
Florian Westphal <fw@strlen.de> writes:
> Giuseppe Scrivano <gscrivan@redhat.com> wrote:
>> SELinux, if enabled, registers for each new network namespace 6
>> netfilter hooks. Avoid to use synchronize_net for each new hook, but do
>> it once after all the hooks are added. The net benefit on an SMP
>> machine with two cores is that creating a new network namespace takes
>> -40% of the original time.
>
> but this needs more work.
>
>> Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
>> ---
>> net/netfilter/core.c | 15 ++++++++++++---
>> 1 file changed, 12 insertions(+), 3 deletions(-)
>>
>> diff --git a/net/netfilter/core.c b/net/netfilter/core.c
>> index 52cd2901a097..beeb0b36f429 100644
>> --- a/net/netfilter/core.c
>> +++ b/net/netfilter/core.c
>> @@ -252,7 +252,7 @@ static struct nf_hook_entries __rcu **nf_hook_entry_head(struct net *net, const
>> return NULL;
>> }
>>
>> -int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
>> +static int __nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
>
> Change this to return struct nf_hook_entries *
thanks for the quick review. Are you fine if I change it to:
static int __nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg, struct nf_hook_entries **old)
so that it is still possible to return other errors as the current code does?
Giuseppe
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [RFC PATCH] netfilter: call synchronize_net only once from nf_register_net_hooks
2017-11-22 11:25 ` Giuseppe Scrivano
@ 2017-11-22 11:56 ` Florian Westphal
2017-11-22 14:30 ` Giuseppe Scrivano
0 siblings, 1 reply; 8+ messages in thread
From: Florian Westphal @ 2017-11-22 11:56 UTC (permalink / raw)
To: Giuseppe Scrivano; +Cc: Florian Westphal, netfilter-devel
Giuseppe Scrivano <gscrivan@redhat.com> wrote:
> Florian Westphal <fw@strlen.de> writes:
>
> > Giuseppe Scrivano <gscrivan@redhat.com> wrote:
> >> SELinux, if enabled, registers for each new network namespace 6
> >> netfilter hooks. Avoid to use synchronize_net for each new hook, but do
> >> it once after all the hooks are added. The net benefit on an SMP
> >> machine with two cores is that creating a new network namespace takes
> >> -40% of the original time.
> >
> > but this needs more work.
> >
> >> Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
> >> ---
> >> net/netfilter/core.c | 15 ++++++++++++---
> >> 1 file changed, 12 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/net/netfilter/core.c b/net/netfilter/core.c
> >> index 52cd2901a097..beeb0b36f429 100644
> >> --- a/net/netfilter/core.c
> >> +++ b/net/netfilter/core.c
> >> @@ -252,7 +252,7 @@ static struct nf_hook_entries __rcu **nf_hook_entry_head(struct net *net, const
> >> return NULL;
> >> }
> >>
> >> -int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
> >> +static int __nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
> >
> > Change this to return struct nf_hook_entries *
>
> thanks for the quick review. Are you fine if I change it to:
>
> static int __nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg, struct nf_hook_entries **old)
You can use ERR_PTR/PTR_ERR instead.
However, I suggest you try to go with call_rcu to get rid of all of the
synchronize_net() calls. I don't even see why its still needed for
nfqueue case provided we invoke the nfqueue drop handler first.
PoC example, untested:
diff --git a/include/linux/netfilter.h b/include/linux/netfilter.h
--- a/include/linux/netfilter.h
+++ b/include/linux/netfilter.h
@@ -77,6 +77,11 @@ struct nf_hook_entry {
void *priv;
};
+struct nf_hook_entries_rcu_head {
+ struct rcu_head head;
+ void *allocation;
+};
+
struct nf_hook_entries {
u16 num_hook_entries;
/* padding */
@@ -87,10 +92,13 @@ struct nf_hook_entries {
* This is not part of struct nf_hook_entry since its only
* needed in slow path (hook register/unregister).
*
- * const struct nf_hook_ops *orig_ops[]
+ * const struct nf_hook_ops *orig_ops[];
+ * struct nf_hook_entries_rcu_head;
*/
};
+void nf_hook_entries_free_rcu(struct nf_hook_entries *e);
+
static inline struct nf_hook_ops **nf_hook_entries_get_hook_ops(const struct nf_hook_entries *e)
{
unsigned int n = e->num_hook_entries;
diff --git a/net/netfilter/core.c b/net/netfilter/core.c
--- a/net/netfilter/core.c
+++ b/net/netfilter/core.c
@@ -69,22 +69,61 @@ static DEFINE_MUTEX(nf_hook_mutex);
#define nf_entry_dereference(e) \
rcu_dereference_protected(e, lockdep_is_held(&nf_hook_mutex))
+static struct nf_hook_entries_rcu_head *
+nf_hook_entries_rcu_head_get(struct nf_hook_entries *e)
+{
+ unsigned int num = e->num_hook_entries;
+ struct nf_hook_entries_rcu_head *head;
+ static struct nf_hook_ops **ops;
+
+ ops = nf_hook_entries_get_hook_ops(e);
+ head = (void *)&ops[num];
+
+ return head;
+}
+
static struct nf_hook_entries *allocate_hook_entries_size(u16 num)
{
struct nf_hook_entries *e;
size_t alloc = sizeof(*e) +
sizeof(struct nf_hook_entry) * num +
- sizeof(struct nf_hook_ops *) * num;
+ sizeof(struct nf_hook_ops *) * num +
+ sizeof(struct nf_hook_entries_rcu_head);
if (num == 0)
return NULL;
e = kvzalloc(alloc, GFP_KERNEL);
- if (e)
+ if (e) {
+ struct nf_hook_entries_rcu_head *head;
+
e->num_hook_entries = num;
+
+ head = nf_hook_entries_rcu_head_get(e);
+ head->allocation = e;
+ }
return e;
}
+void __nf_hook_entries_free(struct rcu_head *h)
+{
+ struct nf_hook_entries_rcu_head *head;
+
+ head = container_of(h, struct nf_hook_entries_rcu_head, head);
+ kvfree(head->allocation);
+}
+
+void nf_hook_entries_free_rcu(struct nf_hook_entries *e)
+{
+ struct nf_hook_entries_rcu_head *head;
+
+ head = nf_hook_entries_rcu_head_get(e);
+ if (WARN_ON(e != head->allocation))
+ return;
+
+ call_rcu(&head->head, __nf_hook_entries_free);
+}
+
static unsigned int accept_all(void *priv,
struct sk_buff *skb,
const struct nf_hook_state *state)
@@ -291,9 +330,9 @@ int nf_register_net_hook(struct net *net, const struct nf_hook_ops *reg)
#ifdef HAVE_JUMP_LABEL
static_key_slow_inc(&nf_hooks_needed[reg->pf][reg->hooknum]);
#endif
- synchronize_net();
BUG_ON(p == new_hooks);
- kvfree(p);
+ if (p)
+ nf_hook_entries_free_rcu(p);
return 0;
}
EXPORT_SYMBOL(nf_register_net_hook);
@@ -341,7 +380,6 @@ void nf_unregister_net_hook(struct net *net, const struct nf_hook_ops *reg)
{
struct nf_hook_entries __rcu **pp;
struct nf_hook_entries *p;
- unsigned int nfq;
pp = nf_hook_entry_head(net, reg);
if (!pp)
@@ -362,13 +400,8 @@ void nf_unregister_net_hook(struct net *net, const struct nf_hook_ops *reg)
if (!p)
return;
- synchronize_net();
-
- /* other cpu might still process nfqueue verdict that used reg */
- nfq = nf_queue_nf_hook_drop(net);
- if (nfq)
- synchronize_net();
- kvfree(p);
+ nf_queue_nf_hook_drop(net);
+ nf_hook_entries_free_rcu(p);
}
EXPORT_SYMBOL(nf_unregister_net_hook);
@@ -395,63 +428,10 @@ EXPORT_SYMBOL(nf_register_net_hooks);
void nf_unregister_net_hooks(struct net *net, const struct nf_hook_ops *reg,
unsigned int hookcount)
{
- struct nf_hook_entries *to_free[16], *p;
- struct nf_hook_entries __rcu **pp;
- unsigned int i, j, n;
-
- mutex_lock(&nf_hook_mutex);
- for (i = 0; i < hookcount; i++) {
- pp = nf_hook_entry_head(net, ®[i]);
- if (!pp)
- continue;
-
- p = nf_entry_dereference(*pp);
- if (WARN_ON_ONCE(!p))
- continue;
- __nf_unregister_net_hook(p, ®[i]);
- }
- mutex_unlock(&nf_hook_mutex);
-
- do {
- n = min_t(unsigned int, hookcount, ARRAY_SIZE(to_free));
-
- mutex_lock(&nf_hook_mutex);
-
- for (i = 0, j = 0; i < hookcount && j < n; i++) {
- pp = nf_hook_entry_head(net, ®[i]);
- if (!pp)
- continue;
-
- p = nf_entry_dereference(*pp);
- if (!p)
- continue;
-
- to_free[j] = __nf_hook_entries_try_shrink(pp);
- if (to_free[j])
- ++j;
- }
-
- mutex_unlock(&nf_hook_mutex);
-
- if (j) {
- unsigned int nfq;
-
- synchronize_net();
-
- /* need 2nd synchronize_net() if nfqueue is used, skb
- * can get reinjected right before nf_queue_hook_drop()
- */
- nfq = nf_queue_nf_hook_drop(net);
- if (nfq)
- synchronize_net();
-
- for (i = 0; i < j; i++)
- kvfree(to_free[i]);
- }
+ unsigned int i;
- reg += n;
- hookcount -= n;
- } while (hookcount > 0);
+ for (i = 0; i < hookcount; i++)
+ nf_unregister_net_hook(net, ®[i]);
}
EXPORT_SYMBOL(nf_unregister_net_hooks);
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [RFC PATCH] netfilter: call synchronize_net only once from nf_register_net_hooks
2017-11-22 11:56 ` Florian Westphal
@ 2017-11-22 14:30 ` Giuseppe Scrivano
2017-11-22 14:54 ` Florian Westphal
0 siblings, 1 reply; 8+ messages in thread
From: Giuseppe Scrivano @ 2017-11-22 14:30 UTC (permalink / raw)
To: Florian Westphal; +Cc: netfilter-devel
Florian Westphal <fw@strlen.de> writes:
> However, I suggest you try to go with call_rcu to get rid of all of the
> synchronize_net() calls. I don't even see why its still needed for
> nfqueue case provided we invoke the nfqueue drop handler first.
>
> PoC example, untested:
thanks, the patch seems to work for me, at least the problem I was
seeing before is now fixed.
Is there any reason for defining "struct nf_hook_entries_rcu_head"?
I've used this fixup patch and it works as well. As someone new to the
code, I find it a bit easier to follow:
diff --git a/include/linux/netfilter.h b/include/linux/netfilter.h
index c08fd706d9c7..663046776def 100644
--- a/include/linux/netfilter.h
+++ b/include/linux/netfilter.h
@@ -77,12 +77,8 @@ struct nf_hook_entry {
void *priv;
};
-struct nf_hook_entries_rcu_head {
- struct rcu_head head;
- void *allocation;
-};
-
struct nf_hook_entries {
+ struct rcu_head head;
u16 num_hook_entries;
/* padding */
struct nf_hook_entry hooks[];
diff --git a/net/netfilter/core.c b/net/netfilter/core.c
index 1f25c914f6e9..6bcf90948810 100644
--- a/net/netfilter/core.c
+++ b/net/netfilter/core.c
@@ -69,59 +69,34 @@ static DEFINE_MUTEX(nf_hook_mutex);
#define nf_entry_dereference(e) \
rcu_dereference_protected(e, lockdep_is_held(&nf_hook_mutex))
-static struct nf_hook_entries_rcu_head *
-nf_hook_entries_rcu_head_get(struct nf_hook_entries *e)
-{
- unsigned int num = e->num_hook_entries;
- struct nf_hook_entries_rcu_head *head;
- static struct nf_hook_ops **ops;
-
- ops = nf_hook_entries_get_hook_ops(e);
- head = (void *)&ops[num];
-
- return head;
-}
-
static struct nf_hook_entries *allocate_hook_entries_size(u16 num)
{
struct nf_hook_entries *e;
size_t alloc = sizeof(*e) +
sizeof(struct nf_hook_entry) * num +
- sizeof(struct nf_hook_ops *) * num +
- sizeof(struct nf_hook_entries_rcu_head);
+ sizeof(struct nf_hook_ops *) * num;
if (num == 0)
return NULL;
e = kvzalloc(alloc, GFP_KERNEL);
if (e) {
- struct nf_hook_entries_rcu_head *head;
-
e->num_hook_entries = num;
-
- head = nf_hook_entries_rcu_head_get(e);
- head->allocation = e;
}
return e;
}
void __nf_hook_entries_free(struct rcu_head *h)
{
- struct nf_hook_entries_rcu_head *head;
+ struct nf_hook_entries *entries;
- head = container_of(h, struct nf_hook_entries_rcu_head, head);
- kvfree(head->allocation);
+ entries = container_of(h, struct nf_hook_entries, head);
+ kvfree(entries);
}
void nf_hook_entries_free_rcu(struct nf_hook_entries *e)
{
- struct nf_hook_entries_rcu_head *head;
-
- head = nf_hook_entries_rcu_head_get(e);
- if (WARN_ON(e != head->allocation))
- return;
-
- call_rcu(&head->head, __nf_hook_entries_free);
+ call_rcu(&e->head, __nf_hook_entries_free);
}
^ permalink raw reply related [flat|nested] 8+ messages in thread
* Re: [RFC PATCH] netfilter: call synchronize_net only once from nf_register_net_hooks
2017-11-22 14:30 ` Giuseppe Scrivano
@ 2017-11-22 14:54 ` Florian Westphal
0 siblings, 0 replies; 8+ messages in thread
From: Florian Westphal @ 2017-11-22 14:54 UTC (permalink / raw)
To: Giuseppe Scrivano; +Cc: Florian Westphal, netfilter-devel
Giuseppe Scrivano <gscrivan@redhat.com> wrote:
> > However, I suggest you try to go with call_rcu to get rid of all of the
> > synchronize_net() calls. I don't even see why its still needed for
> > nfqueue case provided we invoke the nfqueue drop handler first.
> >
> > PoC example, untested:
>
> thanks, the patch seems to work for me, at least the problem I was
> seeing before is now fixed.
>
> Is there any reason for defining "struct nf_hook_entries_rcu_head"?
Yes, idea is to only place elements we need in packet path at the
beginning of structure.
> I've used this fixup patch and it works as well. As someone new to the
> code, I find it a bit easier to follow:
>
>
> diff --git a/include/linux/netfilter.h b/include/linux/netfilter.h
> index c08fd706d9c7..663046776def 100644
> --- a/include/linux/netfilter.h
> +++ b/include/linux/netfilter.h
> @@ -77,12 +77,8 @@ struct nf_hook_entry {
> void *priv;
> };
>
> -struct nf_hook_entries_rcu_head {
> - struct rcu_head head;
> - void *allocation;
> -};
> -
> struct nf_hook_entries {
> + struct rcu_head head;
Right, this works too.
I wanted to keep the rcu head out of the fastpath part
as we only need it during destruction.
^ permalink raw reply [flat|nested] 8+ messages in thread
end of thread, other threads:[~2017-11-22 14:55 UTC | newest]
Thread overview: 8+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2017-11-22 10:40 [RFC PATCH] netfilter: call synchronize_net only once from nf_register_net_hooks Giuseppe Scrivano
2017-11-22 10:41 ` Florian Westphal
2017-11-22 11:06 ` Florian Westphal
2017-11-22 11:10 ` Florian Westphal
2017-11-22 11:25 ` Giuseppe Scrivano
2017-11-22 11:56 ` Florian Westphal
2017-11-22 14:30 ` Giuseppe Scrivano
2017-11-22 14:54 ` Florian Westphal
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.