[PATCH] bpf: Fix integer overflow in argument calculation for bpf_map_area_alloc

In 32-bit architecture, the result of sizeof() is a 32-bit integer so
the expression becomes the multiplication between 2 32-bit integer which
can potentially leads to integer overflow. As a result,
bpf_map_area_alloc() allocates less memory than needed.

Fix this by casting 1 operand to u64.

Signed-off-by: Bui Quang Minh <redacted>
---
 kernel/bpf/devmap.c | 4 ++--
 net/core/sock_map.c | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/kernel/bpf/devmap.c b/kernel/bpf/devmap.c
index f6e9c68afdd4..e849c3e8a49f 100644
--- a/kernel/bpf/devmap.c
+++ b/kernel/bpf/devmap.c

@@ -92,7 +92,7 @@ static struct hlist_head *dev_map_create_hash(unsigned int entries,
 	int i;
 	struct hlist_head *hash;
 
-	hash = bpf_map_area_alloc(entries * sizeof(*hash), numa_node);
+	hash = bpf_map_area_alloc((u64) entries * sizeof(*hash), numa_node);
 	if (hash != NULL)
 		for (i = 0; i < entries; i++)
 			INIT_HLIST_HEAD(&hash[i]);

@@ -143,7 +143,7 @@ static int dev_map_init_map(struct bpf_dtab *dtab, union bpf_attr *attr)
 
 		spin_lock_init(&dtab->index_lock);
 	} else {
-		dtab->netdev_map = bpf_map_area_alloc(dtab->map.max_entries *
+		dtab->netdev_map = bpf_map_area_alloc((u64) dtab->map.max_entries *
 						      sizeof(struct bpf_dtab_netdev *),
 						      dtab->map.numa_node);
 		if (!dtab->netdev_map)

diff --git a/net/core/sock_map.c b/net/core/sock_map.c
index 64b5ec14ff50..7a42016a981d 100644
--- a/net/core/sock_map.c
+++ b/net/core/sock_map.c

@@ -44,7 +44,7 @@ static struct bpf_map *sock_map_alloc(union bpf_attr *attr)
 	bpf_map_init_from_attr(&stab->map, attr);
 	raw_spin_lock_init(&stab->lock);
 
-	stab->sks = bpf_map_area_alloc(stab->map.max_entries *
+	stab->sks = bpf_map_area_alloc((u64) stab->map.max_entries *
 				       sizeof(struct sock *),
 				       stab->map.numa_node);
 	if (!stab->sks) {

-- 
2.17.1

On Tue, 26 Jan 2021 at 08:26, Bui Quang Minh [off-list ref] wrote:

In 32-bit architecture, the result of sizeof() is a 32-bit integer so
the expression becomes the multiplication between 2 32-bit integer which
can potentially leads to integer overflow. As a result,
bpf_map_area_alloc() allocates less memory than needed.

Fix this by casting 1 operand to u64.

Some quick thoughts:
* Should this have a Fixes tag?
* Seems like there are quite a few similar calls scattered around
(cpumap, etc.). Did you audit these as well?
* I'd prefer a calloc style version of bpf_map_area_alloc although
that might conflict with Fixes tag.

Lorenz

-- 
Lorenz Bauer  |  Systems Engineer
6th Floor, County Hall/The Riverside Building, SE1 7PB, UK

www.cloudflare.com

On Tue, Jan 26, 2021 at 09:36:57AM +0000, Lorenz Bauer wrote:

On Tue, 26 Jan 2021 at 08:26, Bui Quang Minh [off-list ref] wrote:

quoted

In 32-bit architecture, the result of sizeof() is a 32-bit integer so
the expression becomes the multiplication between 2 32-bit integer which
can potentially leads to integer overflow. As a result,
bpf_map_area_alloc() allocates less memory than needed.

Fix this by casting 1 operand to u64.

Some quick thoughts:
* Should this have a Fixes tag?

Ok, I will add Fixes tag in later version patch.

* Seems like there are quite a few similar calls scattered around
(cpumap, etc.). Did you audit these as well?

I spotted another bug after re-auditting. In hashtab, there ares 2 places using
the same calls

	static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
	{
		/* ... snip ... */
		if (htab->n_buckets == 0 ||
		    htab->n_buckets > U32_MAX / sizeof(struct bucket))
			goto free_htab;

		htab->buckets = bpf_map_area_alloc(htab->n_buckets *
						   sizeof(struct bucket),
						   htab->map.numa_node);
	}

This is safe because of the above check.

	static int prealloc_init(struct bpf_htab *htab)
	{
		u32 num_entries = htab->map.max_entries;
		htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
						 htab->map.numa_node);
	}

This is not safe since there is no limit check in elem_size.

In cpumap,

	static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
	{
		cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
						   sizeof(struct bpf_cpu_map_entry *),
						   cmap->map.numa_node);
	}

I think this is safe because max_entries is not permitted to be larger than NR_CPUS.

In stackmap, there is a place that I'm not very sure about

	static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
	{
		u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
		smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
						 smap->map.numa_node);
	}

This is called after another bpf_map_area_alloc in stack_map_alloc(). In the first
bpf_map_area_alloc() the argument is calculated in an u64 variable; so if in the second
one, there is an integer overflow then the first one must be called with size > 4GB. I 
think the first one will probably fail (I am not sure about the actual limit of vmalloc()),
so the second one might not be called.

Overall, I think it is error prone in this pattern, maybe we should use typecasting in all
similar calls or make a comment why we don't use typecasting. As I see typecasting is not so
expensive and we can typecast the sizeof() operand so this change only affect 32-bit
architecture.

* I'd prefer a calloc style version of bpf_map_area_alloc although
that might conflict with Fixes tag.

Yes, I think the calloc style will prevent this kind of integer overflow bug.

Thank you,
Quang Minh.

On Wed, Jan 27, 2021 at 11:23:41AM +0700, Bui Quang Minh wrote:

quoted

* Seems like there are quite a few similar calls scattered around
(cpumap, etc.). Did you audit these as well?

I spotted another bug after re-auditting. In hashtab, there ares 2 places using
the same calls

	static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
	{
		/* ... snip ... */
		if (htab->n_buckets == 0 ||
		    htab->n_buckets > U32_MAX / sizeof(struct bucket))
			goto free_htab;

		htab->buckets = bpf_map_area_alloc(htab->n_buckets *
						   sizeof(struct bucket),
						   htab->map.numa_node);
	}

This is safe because of the above check.

	static int prealloc_init(struct bpf_htab *htab)
	{
		u32 num_entries = htab->map.max_entries;
		htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
						 htab->map.numa_node);
	}

This is not safe since there is no limit check in elem_size.

So sorry but I rechecked and saw this bug in hashtab has been fixed with commit
e1868b9e36d0ca

Thank you,
Quang Minh.

On 1/27/21 5:23 AM, Bui Quang Minh wrote:

On Tue, Jan 26, 2021 at 09:36:57AM +0000, Lorenz Bauer wrote:

quoted

On Tue, 26 Jan 2021 at 08:26, Bui Quang Minh [off-list ref] wrote:

quoted

In 32-bit architecture, the result of sizeof() is a 32-bit integer so
the expression becomes the multiplication between 2 32-bit integer which
can potentially leads to integer overflow. As a result,
bpf_map_area_alloc() allocates less memory than needed.

Fix this by casting 1 operand to u64.

Some quick thoughts:
* Should this have a Fixes tag?

Ok, I will add Fixes tag in later version patch.

quoted

* Seems like there are quite a few similar calls scattered around
(cpumap, etc.). Did you audit these as well?

[...]

In cpumap,

	static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
	{
		cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
						   sizeof(struct bpf_cpu_map_entry *),
						   cmap->map.numa_node);
	}

I think this is safe because max_entries is not permitted to be larger than NR_CPUS.

Yes.

In stackmap, there is a place that I'm not very sure about

	static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
	{
		u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
		smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
						 smap->map.numa_node);
	}

This is called after another bpf_map_area_alloc in stack_map_alloc(). In the first
bpf_map_area_alloc() the argument is calculated in an u64 variable; so if in the second
one, there is an integer overflow then the first one must be called with size > 4GB. I
think the first one will probably fail (I am not sure about the actual limit of vmalloc()),
so the second one might not be called.

I would sanity check this as well. Looks like k*alloc()/v*alloc() call sites typically
use array_size() which returns SIZE_MAX on overflow, 610b15c50e86 ("overflow.h: Add
allocation size calculation helpers").

Thanks,
Daniel

On 1/28/21 7:41 AM, Daniel Borkmann wrote:

On 1/27/21 5:23 AM, Bui Quang Minh wrote:

quoted

On Tue, Jan 26, 2021 at 09:36:57AM +0000, Lorenz Bauer wrote:

quoted

On Tue, 26 Jan 2021 at 08:26, Bui Quang Minh 
[off-list ref] wrote:

quoted

In 32-bit architecture, the result of sizeof() is a 32-bit integer so
the expression becomes the multiplication between 2 32-bit integer 
which
can potentially leads to integer overflow. As a result,
bpf_map_area_alloc() allocates less memory than needed.

Fix this by casting 1 operand to u64.

Some quick thoughts:
* Should this have a Fixes tag?

Ok, I will add Fixes tag in later version patch.

quoted

* Seems like there are quite a few similar calls scattered around
(cpumap, etc.). Did you audit these as well?

[...]

quoted

In cpumap,

    static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
    {
        cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
                           sizeof(struct bpf_cpu_map_entry *),
                           cmap->map.numa_node);
    }

I think this is safe because max_entries is not permitted to be larger 
than NR_CPUS.

Yes.

quoted

In stackmap, there is a place that I'm not very sure about

    static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
    {
        u32 elem_size = sizeof(struct stack_map_bucket) + 
smap->map.value_size;
        smap->elems = bpf_map_area_alloc(elem_size * 
smap->map.max_entries,
                         smap->map.numa_node);
    }

This is called after another bpf_map_area_alloc in stack_map_alloc(). 
In the first
bpf_map_area_alloc() the argument is calculated in an u64 variable; so 
if in the second
one, there is an integer overflow then the first one must be called 
with size > 4GB. I
think the first one will probably fail (I am not sure about the actual 
limit of vmalloc()),
so the second one might not be called.

I would sanity check this as well. Looks like k*alloc()/v*alloc() call 
sites typically
use array_size() which returns SIZE_MAX on overflow, 610b15c50e86 
("overflow.h: Add
allocation size calculation helpers").

Hi,

I almost forget about this patch, I have checked the bpf_map_area_alloc 
in in stackmap.c and I can see that integer overflow cannot happen in 
this stackmap.c case.

In stack_map_alloc(),

	u64 cost;
	...
	cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
	cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
	smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr)); (1)
	...
	prealloc_elems_and_freelist(smap);

In prealloc_elems_and_freelist(),

	u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
	smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries, 
smap->map.numa_node); (2)

Argument calculation at (1) is safe. Argument calculation at (2) can 
potentially result in an integer overflow in 32-bit architecture. 
However, if the integer overflow happens, it means argument at (1) must 
be 2**32, which cannot pass the SIZE_MAX check in __bpf_map_area_alloc()

In __bpf_map_area_alloc()

	if (size >= SIZE_MAX)
		return NULL;

So I think the original patch has fixed instances of this bug pattern.

Thank you,
Quang Minh.