[PATCH] iptables compat code error way & module refcounting fix

[PATCH] iptables compat code error way & module refcounting fix

[Dmitry Mishin]

This patch fixes bug in iptables modules refcounting on compat error way.

As we are getting modules in check_compat_entry_size_and_hooks(), in case of
later error, we should put them all in translate_compat_table(), not  in the
compat_copy_entry_from_user() or compat_copy_match_from_user(), as it is now.

Signed-off-by: Dmitry Mishin <dim@openvz.org>
Acked-by: Vasily Averin <vvs@openvz.org>
Acked-by: Kirill Korotaev <dev@openvz.org>

 ip_tables.c |   36 +++++++++++-------------------------
 1 file changed, 11 insertions(+), 25 deletions(-)

----
diff --git a/net/ipv4/netfilter/ip_tables.c b/net/ipv4/netfilter/ip_tables.c
index 4b90927..128ba23 100644
--- a/net/ipv4/netfilter/ip_tables.c
+++ b/net/ipv4/netfilter/ip_tables.c
@@ -1513,7 +1513,7 @@ cleanup_matches:
 
 static inline int compat_copy_match_from_user(struct ipt_entry_match *m,
 	void **dstptr, compat_uint_t *size, const char *name,
-	const struct ipt_ip *ip, unsigned int hookmask, int *i)
+	const struct ipt_ip *ip, unsigned int hookmask)
 {
 	struct ipt_entry_match *dm;
 	struct ipt_match *match;
@@ -1526,22 +1526,13 @@ static inline int compat_copy_match_from
 	ret = xt_check_match(match, AF_INET, dm->u.match_size - sizeof(*dm),
 			     name, hookmask, ip->proto,
 			     ip->invflags & IPT_INV_PROTO);
-	if (ret)
-		goto err;
-
-	if (m->u.kernel.match->checkentry
+	if (!ret && m->u.kernel.match->checkentry
 	    && !m->u.kernel.match->checkentry(name, ip, match, dm->data,
 					      hookmask)) {
 		duprintf("ip_tables: check failed for `%s'.\n",
 			 m->u.kernel.match->name);
 		ret = -EINVAL;
-		goto err;
 	}
-	(*i)++;
-	return 0;
-
-err:
-	module_put(m->u.kernel.match->me);
 	return ret;
 }
 
@@ -1553,19 +1544,18 @@ static int compat_copy_entry_from_user(s
 	struct ipt_target *target;
 	struct ipt_entry *de;
 	unsigned int origsize;
-	int ret, h, j;
+	int ret, h;
 
 	ret = 0;
 	origsize = *size;
 	de = (struct ipt_entry *)*dstptr;
 	memcpy(de, e, sizeof(struct ipt_entry));
 
-	j = 0;
 	*dstptr += sizeof(struct compat_ipt_entry);
 	ret = IPT_MATCH_ITERATE(e, compat_copy_match_from_user, dstptr, size,
-			name, &de->ip, de->comefrom, &j);
+			name, &de->ip, de->comefrom);
 	if (ret)
-		goto cleanup_matches;
+		goto err;
 	de->target_offset = e->target_offset - (origsize - *size);
 	t = ipt_get_target(e);
 	target = t->u.kernel.target;
@@ -1599,12 +1589,7 @@ static int compat_copy_entry_from_user(s
 		goto err;
 	}
 	ret = 0;
-	return ret;
-
 err:
-	module_put(t->u.kernel.target->me);
-cleanup_matches:
-	IPT_MATCH_ITERATE(e, cleanup_match, &j);
 	return ret;
 }
 
@@ -1618,7 +1603,7 @@ translate_compat_table(const char *name,
 		unsigned int *hook_entries,
 		unsigned int *underflows)
 {
-	unsigned int i;
+	unsigned int i, j;
 	struct xt_table_info *newinfo, *info;
 	void *pos, *entry0, *entry1;
 	unsigned int size;
@@ -1636,21 +1621,21 @@ translate_compat_table(const char *name,
 	}
 
 	duprintf("translate_compat_table: size %u\n", info->size);
-	i = 0;
+	j = 0;
 	xt_compat_lock(AF_INET);
 	/* Walk through entries, checking offsets. */
 	ret = IPT_ENTRY_ITERATE(entry0, total_size,
 				check_compat_entry_size_and_hooks,
 				info, &size, entry0,
 				entry0 + total_size,
-				hook_entries, underflows, &i, name);
+				hook_entries, underflows, &j, name);
 	if (ret != 0)
 		goto out_unlock;
 
 	ret = -EINVAL;
-	if (i != number) {
+	if (j != number) {
 		duprintf("translate_compat_table: %u not %u entries\n",
-			 i, number);
+			 j, number);
 		goto out_unlock;
 	}
 
@@ -1709,6 +1694,7 @@ translate_compat_table(const char *name,
 free_newinfo:
 	xt_free_table_info(newinfo);
 out:
+	IPT_ENTRY_ITERATE(entry0, total_size, cleanup_entry, &j);
 	return ret;
 out_unlock:
 	xt_compat_unlock(AF_INET);

Re: [PATCH] iptables compat code error way & module refcounting fix

[Patrick McHardy]

Dmitry Mishin wrote:
> This patch fixes bug in iptables modules refcounting on compat error way.
> 
> As we are getting modules in check_compat_entry_size_and_hooks(), in case of
> later error, we should put them all in translate_compat_table(), not  in the
> compat_copy_entry_from_user() or compat_copy_match_from_user(), as it is now.
> 
> Signed-off-by: Dmitry Mishin <dim@openvz.org>
> Acked-by: Vasily Averin <vvs@openvz.org>
> Acked-by: Kirill Korotaev <dev@openvz.org>
> 
>  ip_tables.c |   36 +++++++++++-------------------------
>  1 file changed, 11 insertions(+), 25 deletions(-)
> 
> ----
> @@ -1618,7 +1603,7 @@ translate_compat_table(const char *name,
>  		unsigned int *hook_entries,
>  		unsigned int *underflows)
>  {
> -	unsigned int i;
> +	unsigned int i, j;
>  	struct xt_table_info *newinfo, *info;
>  	void *pos, *entry0, *entry1;
>  	unsigned int size;
> @@ -1636,21 +1621,21 @@ translate_compat_table(const char *name,
>  	}
>  
>  	duprintf("translate_compat_table: size %u\n", info->size);
> -	i = 0;
> +	j = 0;
>  	xt_compat_lock(AF_INET);
>  	/* Walk through entries, checking offsets. */
>  	ret = IPT_ENTRY_ITERATE(entry0, total_size,
>  				check_compat_entry_size_and_hooks,
>  				info, &size, entry0,
>  				entry0 + total_size,
> -				hook_entries, underflows, &i, name);
> +				hook_entries, underflows, &j, name);
>  	if (ret != 0)
>  		goto out_unlock;
>  
>  	ret = -EINVAL;
> -	if (i != number) {
> +	if (j != number) {
>  		duprintf("translate_compat_table: %u not %u entries\n",
> -			 i, number);
> +			 j, number);
>  		goto out_unlock;
>  	}
>  
> @@ -1709,6 +1694,7 @@ translate_compat_table(const char *name,
>  free_newinfo:
>  	xt_free_table_info(newinfo);
>  out:
> +	IPT_ENTRY_ITERATE(entry0, total_size, cleanup_entry, &j);
>  	return ret;
>  out_unlock:
>  	xt_compat_unlock(AF_INET);
> 

This doesn't look right, if we fail at xt_alloc_table_info for
example the module references are not released. What case
exactly is this patch supposed to fix?

Re: [PATCH] iptables compat code error way & module refcounting fix

[Dmitry Mishin]

On Monday 30 October 2006 18:45, Patrick McHardy wrote:
> Dmitry Mishin wrote:
> > This patch fixes bug in iptables modules refcounting on compat error way.
> >
> > As we are getting modules in check_compat_entry_size_and_hooks(), in case
> > of later error, we should put them all in translate_compat_table(), not 
> > in the compat_copy_entry_from_user() or compat_copy_match_from_user(), as
> > it is now.
> >
> > Signed-off-by: Dmitry Mishin <dim@openvz.org>
> > Acked-by: Vasily Averin <vvs@openvz.org>
> > Acked-by: Kirill Korotaev <dev@openvz.org>
> >
> >  ip_tables.c |   36 +++++++++++-------------------------
> >  1 file changed, 11 insertions(+), 25 deletions(-)
> >
> > ----
> > @@ -1618,7 +1603,7 @@ translate_compat_table(const char *name,
> >  		unsigned int *hook_entries,
> >  		unsigned int *underflows)
> >  {
> > -	unsigned int i;
> > +	unsigned int i, j;
> >  	struct xt_table_info *newinfo, *info;
> >  	void *pos, *entry0, *entry1;
> >  	unsigned int size;
> > @@ -1636,21 +1621,21 @@ translate_compat_table(const char *name,
> >  	}
> >
> >  	duprintf("translate_compat_table: size %u\n", info->size);
> > -	i = 0;
> > +	j = 0;
> >  	xt_compat_lock(AF_INET);
> >  	/* Walk through entries, checking offsets. */
> >  	ret = IPT_ENTRY_ITERATE(entry0, total_size,
> >  				check_compat_entry_size_and_hooks,
> >  				info, &size, entry0,
> >  				entry0 + total_size,
> > -				hook_entries, underflows, &i, name);
> > +				hook_entries, underflows, &j, name);
> >  	if (ret != 0)
> >  		goto out_unlock;
> >
> >  	ret = -EINVAL;
> > -	if (i != number) {
> > +	if (j != number) {
> >  		duprintf("translate_compat_table: %u not %u entries\n",
> > -			 i, number);
> > +			 j, number);
> >  		goto out_unlock;
> >  	}
> >
> > @@ -1709,6 +1694,7 @@ translate_compat_table(const char *name,
> >  free_newinfo:
> >  	xt_free_table_info(newinfo);
> >  out:
> > +	IPT_ENTRY_ITERATE(entry0, total_size, cleanup_entry, &j);
> >  	return ret;
> >  out_unlock:
> >  	xt_compat_unlock(AF_INET);
>
> This doesn't look right, if we fail at xt_alloc_table_info for
> example the module references are not released.
module references will be released because of 'goto out;' line just after 
xt_compat_unlock() - it is missed because of standard +-3 lines patch 
context, sorry.

-- 
Thanks,
Dmitry.

Re: [PATCH] iptables compat code error way & module refcounting fix

[Patrick McHardy]

Dmitry Mishin wrote:
> On Monday 30 October 2006 18:45, Patrick McHardy wrote:
> 
>>This doesn't look right, if we fail at xt_alloc_table_info for
>>example the module references are not released.
> 
> module references will be released because of 'goto out;' line just after 
> xt_compat_unlock() - it is missed because of standard +-3 lines patch 
> context, sorry.

You're right, I misread the code. Applied, thanks.