[PATCH] ipv4: fib: avoid NULL dereference

In tnode_free() we iterate over a callback_head list with a while loop.
At the start of the loop body we generate the next head pointer, and at
the end of the loop body we generate the tn pointer for the next
iteration of the loop by using container_of() on the head pointer to
find the tnode, and deriving the kv pointer from this.

In the final iteration of the loop, this means that we derive a pointer
from NULL, which is undefined behaviour, which UBSAN detects:

================================================================================
UBSAN: Undefined behaviour in net/ipv4/fib_trie.c:504:6
member access within null pointer of type 'struct tnode'
CPU: 1 PID: 94 Comm: ip Not tainted 4.18.0-rc3+ #23
Hardware name: linux,dummy-virt (DT)
Call trace:
 dump_backtrace+0x0/0x458
 show_stack+0x20/0x30
 dump_stack+0x18c/0x248
 ubsan_epilogue+0x18/0x94
 handle_null_ptr_deref+0x1d4/0x228
 __ubsan_handle_type_mismatch_v1+0x188/0x1e0
 tnode_free+0x16c/0x1d0
 replace+0x1e0/0x5d0
 resize+0xd98/0x2008
 fib_insert_alias+0xb38/0x10c8
 fib_table_insert+0x7d0/0x1108
 fib_magic+0x530/0x780
 fib_add_ifaddr+0x378/0x468
 fib_netdev_event+0x2ac/0x3e8
 notifier_call_chain+0x190/0x2f8
 raw_notifier_call_chain+0x3c/0x68
 call_netdevice_notifiers_info+0x3c/0xc0
 __dev_notify_flags+0x1f8/0x398
 dev_change_flags+0xe8/0x150
 do_setlink+0x924/0x4050
 rtnl_newlink+0x8c4/0x14b0
 rtnetlink_rcv_msg+0x408/0xef8
 netlink_rcv_skb+0x144/0x390
 rtnetlink_rcv+0x24/0x30
 netlink_unicast+0x4e8/0x740
 netlink_sendmsg+0x6d8/0xe78
 sock_sendmsg+0x90/0x168
 ___sys_sendmsg+0x680/0x9b0
 __sys_sendmsg+0xf0/0x230
 sys_sendmsg+0x34/0x48
 el0_svc_naked+0x30/0x34
================================================================================

We can avoid the undefined behaviour by generating tn for the current
iteration of the loop before we advance head, so let's do that.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Alexey Kuznetsov <redacted>
Cc: David S. Miller <davem@davemloft.net>
Cc: Hideaki YOSHIFUJI <redacted>
Cc: netdev@vger.kernel.org
---
 net/ipv4/fib_trie.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c
index 5bc0c89e81e4..8d98c8162554 100644
--- a/net/ipv4/fib_trie.c
+++ b/net/ipv4/fib_trie.c

@@ -497,11 +497,11 @@ static void tnode_free(struct key_vector *tn)
 	struct callback_head *head = &tn_info(tn)->rcu;
 
 	while (head) {
+		tn = container_of(head, struct tnode, rcu)->kv;
+
 		head = head->next;
 		tnode_free_size += TNODE_SIZE(1ul << tn->bits);
 		node_free(tn);
-
-		tn = container_of(head, struct tnode, rcu)->kv;
 	}
 
 	if (tnode_free_size >= PAGE_SIZE * sync_pages) {

-- 
2.11.0

On 07/06/2018 07:28 AM, Mark Rutland wrote:

In tnode_free() we iterate over a callback_head list with a while loop.
At the start of the loop body we generate the next head pointer, and at
the end of the loop body we generate the tn pointer for the next
iteration of the loop by using container_of() on the head pointer to
find the tnode, and deriving the kv pointer from this.

In the final iteration of the loop, this means that we derive a pointer
from NULL, which is undefined behaviour, which UBSAN detects:

There is no dereference, your patch title is misleading.

UBSAN might be fooled, not the C compiler.

On 7/6/18 8:28 AM, Mark Rutland wrote:

you are changing tn before the above 2 calls are made on the tn passed in

-
-		tn = container_of(head, struct tnode, rcu)->kv;
 	}
 
 	if (tnode_free_size >= PAGE_SIZE * sync_pages) {

On Fri, Jul 06, 2018 at 08:49:14AM -0600, David Ahern wrote:

On 7/6/18 8:28 AM, Mark Rutland wrote:

quoted

diff --git a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c
index 5bc0c89e81e4..8d98c8162554 100644
--- a/net/ipv4/fib_trie.c
+++ b/net/ipv4/fib_trie.c

@@ -497,11 +497,11 @@ static void tnode_free(struct key_vector *tn)
 	struct callback_head *head = &tn_info(tn)->rcu;
 
 	while (head) {
+		tn = container_of(head, struct tnode, rcu)->kv;
+
 		head = head->next;
 		tnode_free_size += TNODE_SIZE(1ul << tn->bits);
 		node_free(tn);

you are changing tn before the above 2 calls are made on the tn passed in

As tn_info(tn) is container_of(kv, struct tnode, tn), the end result is
the same tn value, so this is not a problem.

Thanks,
Mark.

On Fri, Jul 06, 2018 at 07:47:04AM -0700, Eric Dumazet wrote:

On 07/06/2018 07:28 AM, Mark Rutland wrote:

quoted

In tnode_free() we iterate over a callback_head list with a while loop.
At the start of the loop body we generate the next head pointer, and at
the end of the loop body we generate the tn pointer for the next
iteration of the loop by using container_of() on the head pointer to
find the tnode, and deriving the kv pointer from this.

In the final iteration of the loop, this means that we derive a pointer
from NULL, which is undefined behaviour, which UBSAN detects:

There is no dereference, your patch title is misleading.

UBSAN might be fooled, not the C compiler.

I'm happy to change the title to "avoid undefined behaviour".

Thanks,
Mark.

On 07/06/2018 07:57 AM, Mark Rutland wrote:

On Fri, Jul 06, 2018 at 07:47:04AM -0700, Eric Dumazet wrote:

quoted

On 07/06/2018 07:28 AM, Mark Rutland wrote:

quoted

In tnode_free() we iterate over a callback_head list with a while loop.
At the start of the loop body we generate the next head pointer, and at
the end of the loop body we generate the tn pointer for the next
iteration of the loop by using container_of() on the head pointer to
find the tnode, and deriving the kv pointer from this.

In the final iteration of the loop, this means that we derive a pointer
from NULL, which is undefined behaviour, which UBSAN detects:

There is no dereference, your patch title is misleading.

UBSAN might be fooled, not the C compiler.

I'm happy to change the title to "avoid undefined behaviour".

Are you planning to change this as well ?

include/linux/stddef.h:19:#define offsetof(TYPE, MEMBER)	((size_t)&((TYPE *)0)->MEMBER)

(And probably dozens of other locations)

On Fri, Jul 06, 2018 at 08:39:11AM -0700, Eric Dumazet wrote:

On 07/06/2018 07:57 AM, Mark Rutland wrote:

quoted

On Fri, Jul 06, 2018 at 07:47:04AM -0700, Eric Dumazet wrote:

quoted

On 07/06/2018 07:28 AM, Mark Rutland wrote:

quoted

In tnode_free() we iterate over a callback_head list with a while loop.
At the start of the loop body we generate the next head pointer, and at
the end of the loop body we generate the tn pointer for the next
iteration of the loop by using container_of() on the head pointer to
find the tnode, and deriving the kv pointer from this.

In the final iteration of the loop, this means that we derive a pointer
from NULL, which is undefined behaviour, which UBSAN detects:

There is no dereference, your patch title is misleading.

UBSAN might be fooled, not the C compiler.

I'm happy to change the title to "avoid undefined behaviour".

Are you planning to change this as well ?

include/linux/stddef.h:19:#define offsetof(TYPE, MEMBER)	((size_t)&((TYPE *)0)->MEMBER)

No, because __builtin_offsetof() is used these days (since GCC 4),
avoiding the undefined behaviour.

(And probably dozens of other locations)

I do concede that if this is everywhere it's not worth the effort, and
from the looks of things, the gnaliest cases are where we do things
like:

  get_user(var, &struct->field)

... where the user could validly pass a NULL pointer if it wished.

So I guess I'll give up.

Thanks,
Mark.

On 07/06/2018 08:54 AM, Mark Rutland wrote:

On Fri, Jul 06, 2018 at 08:39:11AM -0700, Eric Dumazet wrote:

quoted

On 07/06/2018 07:57 AM, Mark Rutland wrote:

quoted

On Fri, Jul 06, 2018 at 07:47:04AM -0700, Eric Dumazet wrote:

quoted

On 07/06/2018 07:28 AM, Mark Rutland wrote:

quoted

In tnode_free() we iterate over a callback_head list with a while loop.
At the start of the loop body we generate the next head pointer, and at
the end of the loop body we generate the tn pointer for the next
iteration of the loop by using container_of() on the head pointer to
find the tnode, and deriving the kv pointer from this.

In the final iteration of the loop, this means that we derive a pointer
from NULL, which is undefined behaviour, which UBSAN detects:

There is no dereference, your patch title is misleading.

UBSAN might be fooled, not the C compiler.

I'm happy to change the title to "avoid undefined behaviour".

Are you planning to change this as well ?

include/linux/stddef.h:19:#define offsetof(TYPE, MEMBER)	((size_t)&((TYPE *)0)->MEMBER)

No, because __builtin_offsetof() is used these days (since GCC 4),
avoiding the undefined behaviour.

Ah... should we remove the line and declare linux must be compiled
with GCC 4 at least ?

quoted

(And probably dozens of other locations)

I do concede that if this is everywhere it's not worth the effort, and
from the looks of things, the gnaliest cases are where we do things
like:

  get_user(var, &struct->field)

... where the user could validly pass a NULL pointer if it wished.

So I guess I'll give up.

There is value to your patch, since it makes UBSAN happy.

But please change the title and changelog accordingly.

Thanks.

On Fri, Jul 06, 2018 at 09:20:44AM -0700, Eric Dumazet wrote:

On 07/06/2018 08:54 AM, Mark Rutland wrote:

quoted

On Fri, Jul 06, 2018 at 08:39:11AM -0700, Eric Dumazet wrote:

quoted

On 07/06/2018 07:57 AM, Mark Rutland wrote:

quoted

On Fri, Jul 06, 2018 at 07:47:04AM -0700, Eric Dumazet wrote:

quoted

On 07/06/2018 07:28 AM, Mark Rutland wrote:

quoted

In tnode_free() we iterate over a callback_head list with a while loop.
At the start of the loop body we generate the next head pointer, and at
the end of the loop body we generate the tn pointer for the next
iteration of the loop by using container_of() on the head pointer to
find the tnode, and deriving the kv pointer from this.

In the final iteration of the loop, this means that we derive a pointer
from NULL, which is undefined behaviour, which UBSAN detects:

There is no dereference, your patch title is misleading.

UBSAN might be fooled, not the C compiler.

I'm happy to change the title to "avoid undefined behaviour".

Are you planning to change this as well ?

include/linux/stddef.h:19:#define offsetof(TYPE, MEMBER)	((size_t)&((TYPE *)0)->MEMBER)

No, because __builtin_offsetof() is used these days (since GCC 4),
avoiding the undefined behaviour.

Ah... should we remove the line and declare linux must be compiled
with GCC 4 at least ?

Good question -- IIUC x86 already mandates at least GCC 4.6, as this is
necessary for jump labels.

There was a push to mandate a more recent compiler generally, but I'm
not sure of the current state of things.

Arnd, were we planning to mandate at least GCC 4.x?

quoted

I do concede that if this is everywhere it's not worth the effort, and
from the looks of things, the gnaliest cases are where we do things
like:

  get_user(var, &struct->field)

... where the user could validly pass a NULL pointer if it wished.

So I guess I'll give up.

There is value to your patch, since it makes UBSAN happy.

But please change the title and changelog accordingly.

Sure thing. Are you happy with the "deriving a pointer from NULL"
wording in the explanation? That wasn't intended to mean a dereference.

I'll change the title to:

  ipv4: fib: avoid undefined behaviour

Thanks,
Mark.

On 07/06/2018 09:33 AM, Mark Rutland wrote:

Sure thing. Are you happy with the "deriving a pointer from NULL"
wording in the explanation? That wasn't intended to mean a dereference.

I'll change the title to:

  ipv4: fib: avoid undefined behaviour

This would be fine for me, thanks.