Linux Ethernet Bridging - Jan 2019 - [Bridge] [PATCH v2] netfilter: account ebt_table_info to kmemcg

The [ip,ip6,arp]_tables use x_tables_info internally and the underlying
memory is already accounted to kmemcg. Do the same for ebtables. The
syzbot, by using setsockopt(EBT_SO_SET_ENTRIES), was able to OOM the
whole system from a restricted memcg, a potential DoS.

By accounting the ebt_table_info, the memory used for ebt_table_info can
be contained within the memcg of the allocating process. However the
lifetime of ebt_table_info is independent of the allocating process and
is tied to the network namespace. So, the oom-killer will not be able to
relieve the memory pressure due to ebt_table_info memory. The memory for
ebt_table_info is allocated through vmalloc. Currently vmalloc does not
handle the oom-killed allocating process correctly and one large
allocation can bypass memcg limit enforcement. So, with this patch,
at least the small allocations will be contained. For large allocations,
we need to fix vmalloc.

Reported-by: syzbot+7713f3aa67be76b1552c at syzkaller.appspotmail.com
Signed-off-by: Shakeel Butt <shakeelb at google.com>
Cc: Florian Westphal <fw at strlen.de>
Cc: Michal Hocko <mhocko at kernel.org>
Cc: Kirill Tkhai <ktkhai at virtuozzo.com>
Cc: Pablo Neira Ayuso <pablo at netfilter.org>
Cc: Jozsef Kadlecsik <kadlec at blackhole.kfki.hu>
Cc: Roopa Prabhu <roopa at cumulusnetworks.com>
Cc: Nikolay Aleksandrov <nikolay at cumulusnetworks.com>
Cc: Andrew Morton <akpm at linux-foundation.org>
Cc: Linux MM <linux-mm at kvack.org>
Cc: netfilter-devel at vger.kernel.org
Cc: coreteam at netfilter.org
Cc: bridge at lists.linux-foundation.org
Cc: LKML <linux-kernel at vger.kernel.org>
---
Changelog since v1:
- More descriptive commit message.

 net/bridge/netfilter/ebtables.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/net/bridge/netfilter/ebtables.c b/net/bridge/netfilter/ebtables.c
index 491828713e0b..5e55cef0cec3 100644
--- a/net/bridge/netfilter/ebtables.c
+++ b/net/bridge/netfilter/ebtables.c
@@ -1137,14 +1137,16 @@ static int do_replace(struct net *net, const void __user
*user,
 	tmp.name[sizeof(tmp.name) - 1] = 0;
 
 	countersize = COUNTER_OFFSET(tmp.nentries) * nr_cpu_ids;
-	newinfo = vmalloc(sizeof(*newinfo) + countersize);
+	newinfo = __vmalloc(sizeof(*newinfo) + countersize, GFP_KERNEL_ACCOUNT,
+			    PAGE_KERNEL);
 	if (!newinfo)
 		return -ENOMEM;
 
 	if (countersize)
 		memset(newinfo->counters, 0, countersize);
 
-	newinfo->entries = vmalloc(tmp.entries_size);
+	newinfo->entries = __vmalloc(tmp.entries_size, GFP_KERNEL_ACCOUNT,
+				     PAGE_KERNEL);
 	if (!newinfo->entries) {
 		ret = -ENOMEM;
 		goto free_newinfo;
-- 
2.20.1.415.g653613c723-goog

> On Jan 2, 2019, at 8:14 PM, Shakeel Butt <shakeelb at google.com>
wrote:
> 
> 	countersize = COUNTER_OFFSET(tmp.nentries) * nr_cpu_ids;
> -	newinfo = vmalloc(sizeof(*newinfo) + countersize);
> +	newinfo = __vmalloc(sizeof(*newinfo) + countersize, GFP_KERNEL_ACCOUNT,
> +			    PAGE_KERNEL);
> 	if (!newinfo)
> 		return -ENOMEM;
> 
> 	if (countersize)
> 		memset(newinfo->counters, 0, countersize);
> 
> -	newinfo->entries = vmalloc(tmp.entries_size);
> +	newinfo->entries = __vmalloc(tmp.entries_size, GFP_KERNEL_ACCOUNT,
> +				     PAGE_KERNEL);
> 	if (!newinfo->entries) {
> 		ret = -ENOMEM;
> 		goto free_newinfo;
> -- 
Just out of curiosity, what are the actual sizes of these areas in typical use
given __vmalloc() will be allocating by the page?

On Thu, Jan 3, 2019 at 2:15 AM William Kucharski
<william.kucharski at oracle.com> wrote:
>
>
>
> > On Jan 2, 2019, at 8:14 PM, Shakeel Butt <shakeelb at
google.com> wrote:
> >
> >       countersize = COUNTER_OFFSET(tmp.nentries) * nr_cpu_ids;
> > -     newinfo = vmalloc(sizeof(*newinfo) + countersize);
> > +     newinfo = __vmalloc(sizeof(*newinfo) + countersize,
GFP_KERNEL_ACCOUNT,
> > +                         PAGE_KERNEL);
> >       if (!newinfo)
> >               return -ENOMEM;
> >
> >       if (countersize)
> >               memset(newinfo->counters, 0, countersize);
> >
> > -     newinfo->entries = vmalloc(tmp.entries_size);
> > +     newinfo->entries = __vmalloc(tmp.entries_size,
GFP_KERNEL_ACCOUNT,
> > +                                  PAGE_KERNEL);
> >       if (!newinfo->entries) {
> >               ret = -ENOMEM;
> >               goto free_newinfo;
> > --
>
> Just out of curiosity, what are the actual sizes of these areas in typical
use
> given __vmalloc() will be allocating by the page?
>
We don't really use this in production, so, I don't have a good idea
of the size in the typical case. The size depends on the workload. The
motivation behind this patch was the system OOM triggered by a syzbot
running in a restricted memcg.

Shakeel

On 03.01.2019 06:14, Shakeel Butt wrote:
> The [ip,ip6,arp]_tables use x_tables_info internally and the underlying
> memory is already accounted to kmemcg. Do the same for ebtables. The
> syzbot, by using setsockopt(EBT_SO_SET_ENTRIES), was able to OOM the
> whole system from a restricted memcg, a potential DoS.
> 
> By accounting the ebt_table_info, the memory used for ebt_table_info can
> be contained within the memcg of the allocating process. However the
> lifetime of ebt_table_info is independent of the allocating process and
> is tied to the network namespace. So, the oom-killer will not be able to
> relieve the memory pressure due to ebt_table_info memory. The memory for
> ebt_table_info is allocated through vmalloc. Currently vmalloc does not
> handle the oom-killed allocating process correctly and one large
> allocation can bypass memcg limit enforcement. So, with this patch,
> at least the small allocations will be contained. For large allocations,
> we need to fix vmalloc.
> 
> Reported-by: syzbot+7713f3aa67be76b1552c at syzkaller.appspotmail.com
> Signed-off-by: Shakeel Butt <shakeelb at google.com>
> Cc: Florian Westphal <fw at strlen.de>
> Cc: Michal Hocko <mhocko at kernel.org>
> Cc: Kirill Tkhai <ktkhai at virtuozzo.com>
> Cc: Pablo Neira Ayuso <pablo at netfilter.org>
> Cc: Jozsef Kadlecsik <kadlec at blackhole.kfki.hu>
> Cc: Roopa Prabhu <roopa at cumulusnetworks.com>
> Cc: Nikolay Aleksandrov <nikolay at cumulusnetworks.com>
> Cc: Andrew Morton <akpm at linux-foundation.org>
> Cc: Linux MM <linux-mm at kvack.org>
> Cc: netfilter-devel at vger.kernel.org
> Cc: coreteam at netfilter.org
> Cc: bridge at lists.linux-foundation.org
> Cc: LKML <linux-kernel at vger.kernel.org>
> ---
> Changelog since v1:
> - More descriptive commit message.
Reviewed-by: Kirill Tkhai <ktkhai at virtuozzo.com>
> 
>  net/bridge/netfilter/ebtables.c | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
> 
> diff --git a/net/bridge/netfilter/ebtables.c
b/net/bridge/netfilter/ebtables.c
> index 491828713e0b..5e55cef0cec3 100644
> --- a/net/bridge/netfilter/ebtables.c
> +++ b/net/bridge/netfilter/ebtables.c
> @@ -1137,14 +1137,16 @@ static int do_replace(struct net *net, const void
__user *user,
>  	tmp.name[sizeof(tmp.name) - 1] = 0;
>  
>  	countersize = COUNTER_OFFSET(tmp.nentries) * nr_cpu_ids;
> -	newinfo = vmalloc(sizeof(*newinfo) + countersize);
> +	newinfo = __vmalloc(sizeof(*newinfo) + countersize, GFP_KERNEL_ACCOUNT,
> +			    PAGE_KERNEL);
>  	if (!newinfo)
>  		return -ENOMEM;
>  
>  	if (countersize)
>  		memset(newinfo->counters, 0, countersize);
>  
> -	newinfo->entries = vmalloc(tmp.entries_size);
> +	newinfo->entries = __vmalloc(tmp.entries_size, GFP_KERNEL_ACCOUNT,
> +				     PAGE_KERNEL);
>  	if (!newinfo->entries) {
>  		ret = -ENOMEM;
>  		goto free_newinfo;
>

On Wed, Jan 02, 2019 at 07:14:31PM -0800, Shakeel Butt
wrote:
> The [ip,ip6,arp]_tables use x_tables_info internally and the underlying
> memory is already accounted to kmemcg. Do the same for ebtables. The
> syzbot, by using setsockopt(EBT_SO_SET_ENTRIES), was able to OOM the
> whole system from a restricted memcg, a potential DoS.
> 
> By accounting the ebt_table_info, the memory used for ebt_table_info can
> be contained within the memcg of the allocating process. However the
> lifetime of ebt_table_info is independent of the allocating process and
> is tied to the network namespace. So, the oom-killer will not be able to
> relieve the memory pressure due to ebt_table_info memory. The memory for
> ebt_table_info is allocated through vmalloc. Currently vmalloc does not
> handle the oom-killed allocating process correctly and one large
> allocation can bypass memcg limit enforcement. So, with this patch,
> at least the small allocations will be contained. For large allocations,
> we need to fix vmalloc.
Fine with this -mm?

If no objections, I'll apply this to the netfilter tree. Thanks.
> Reported-by: syzbot+7713f3aa67be76b1552c at syzkaller.appspotmail.com
> Signed-off-by: Shakeel Butt <shakeelb at google.com>
> Cc: Florian Westphal <fw at strlen.de>
> Cc: Michal Hocko <mhocko at kernel.org>
> Cc: Kirill Tkhai <ktkhai at virtuozzo.com>
> Cc: Pablo Neira Ayuso <pablo at netfilter.org>
> Cc: Jozsef Kadlecsik <kadlec at blackhole.kfki.hu>
> Cc: Roopa Prabhu <roopa at cumulusnetworks.com>
> Cc: Nikolay Aleksandrov <nikolay at cumulusnetworks.com>
> Cc: Andrew Morton <akpm at linux-foundation.org>
> Cc: Linux MM <linux-mm at kvack.org>
> Cc: netfilter-devel at vger.kernel.org
> Cc: coreteam at netfilter.org
> Cc: bridge at lists.linux-foundation.org
> Cc: LKML <linux-kernel at vger.kernel.org>
> ---
> Changelog since v1:
> - More descriptive commit message.
> 
>  net/bridge/netfilter/ebtables.c | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
> 
> diff --git a/net/bridge/netfilter/ebtables.c
b/net/bridge/netfilter/ebtables.c
> index 491828713e0b..5e55cef0cec3 100644
> --- a/net/bridge/netfilter/ebtables.c
> +++ b/net/bridge/netfilter/ebtables.c
> @@ -1137,14 +1137,16 @@ static int do_replace(struct net *net, const void
__user *user,
>  	tmp.name[sizeof(tmp.name) - 1] = 0;
>  
>  	countersize = COUNTER_OFFSET(tmp.nentries) * nr_cpu_ids;
> -	newinfo = vmalloc(sizeof(*newinfo) + countersize);
> +	newinfo = __vmalloc(sizeof(*newinfo) + countersize, GFP_KERNEL_ACCOUNT,
> +			    PAGE_KERNEL);
>  	if (!newinfo)
>  		return -ENOMEM;
>  
>  	if (countersize)
>  		memset(newinfo->counters, 0, countersize);
>  
> -	newinfo->entries = vmalloc(tmp.entries_size);
> +	newinfo->entries = __vmalloc(tmp.entries_size, GFP_KERNEL_ACCOUNT,
> +				     PAGE_KERNEL);
>  	if (!newinfo->entries) {
>  		ret = -ENOMEM;
>  		goto free_newinfo;
> -- 
> 2.20.1.415.g653613c723-goog
>

On Wed, Jan 02, 2019 at 07:14:31PM -0800, Shakeel Butt
wrote:
> The [ip,ip6,arp]_tables use x_tables_info internally and the underlying
> memory is already accounted to kmemcg. Do the same for ebtables. The
> syzbot, by using setsockopt(EBT_SO_SET_ENTRIES), was able to OOM the
> whole system from a restricted memcg, a potential DoS.
> 
> By accounting the ebt_table_info, the memory used for ebt_table_info can
> be contained within the memcg of the allocating process. However the
> lifetime of ebt_table_info is independent of the allocating process and
> is tied to the network namespace. So, the oom-killer will not be able to
> relieve the memory pressure due to ebt_table_info memory. The memory for
> ebt_table_info is allocated through vmalloc. Currently vmalloc does not
> handle the oom-killed allocating process correctly and one large
> allocation can bypass memcg limit enforcement. So, with this patch,
> at least the small allocations will be contained. For large allocations,
> we need to fix vmalloc.
OK, patch is applied, thanks.