[PATCH net] rtnetlink: invoke 'cb->done' destructor before 'cb->args' reset

cb->args[2] can store the pointer to the struct fib6_walker,
allocated in inet6_dump_fib(). On the next loop iteration in
rtnl_dump_all(), 'memset(&cb, 0, sizeof(cb->args))' can reset
that pointer, leaking the memory [1].

Fix it by calling cb->done, if it is set, before filling 'cb->args'
with zeros.

Looks like the recent changes in rtnl_dump_all() contributed to
the appearance of this kmemleak [1], commit c63586dc9b3e ("net:
rtnl_dump_all needs to propagate error from dumpit function")
breaks the loop only on an error now.

[1]:
unreferenced object 0xffff88001322a200 (size 96):
  comm "sshd", pid 1484, jiffies 4296032768 (age 1432.542s)
  hex dump (first 32 bytes):
    00 01 00 00 00 00 ad de 00 02 00 00 00 00 ad de  ................
    18 09 41 36 00 88 ff ff 18 09 41 36 00 88 ff ff  ..A6......A6....
  backtrace:
    [<0000000095846b39>] kmem_cache_alloc_trace+0x151/0x220
    [<000000007d12709f>] inet6_dump_fib+0x68d/0x940
    [<000000002775a316>] rtnl_dump_all+0x1d9/0x2d0
    [<00000000d7cd302b>] netlink_dump+0x945/0x11a0
    [<000000002f43485f>] __netlink_dump_start+0x55d/0x800
    [<00000000f76bbeec>] rtnetlink_rcv_msg+0x4fa/0xa00
    [<000000009b5761f3>] netlink_rcv_skb+0x29c/0x420
    [<0000000087a1dae1>] rtnetlink_rcv+0x15/0x20
    [<00000000691b703b>] netlink_unicast+0x4e3/0x6c0
    [<00000000b5be0204>] netlink_sendmsg+0x7f2/0xba0
    [<0000000096d2aa60>] sock_sendmsg+0xba/0xf0
    [<000000008c1b786f>] __sys_sendto+0x1e4/0x330
    [<0000000019587b3f>] __x64_sys_sendto+0xe1/0x1a0
    [<00000000071f4d56>] do_syscall_64+0x9f/0x300
    [<000000002737577f>] entry_SYSCALL_64_after_hwframe+0x44/0xa9
    [<0000000057587684>] 0xffffffffffffffff

Fixes: 1b43af5480c3 ("[IPV6]: Increase number of possible routing tables to 2^32")
Signed-off-by: Alexey Kodanev <redacted>
---
 net/core/rtnetlink.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/net/core/rtnetlink.c b/net/core/rtnetlink.c
index f679c7a..314c683 100644
--- a/net/core/rtnetlink.c
+++ b/net/core/rtnetlink.c

@@ -3362,6 +3362,8 @@ static int rtnl_dump_all(struct sk_buff *skb, struct netlink_callback *cb)
 			continue;
 
 		if (idx > s_idx) {
+			if (cb->done)
+				cb->done(cb);
 			memset(&cb->args[0], 0, sizeof(cb->args));
 			cb->prev_seq = 0;
 			cb->seq = 0;

-- 
1.8.3.1

On 31.10.2018 09:42, Alexey Kodanev wrote:

cb->args[2] can store the pointer to the struct fib6_walker,
allocated in inet6_dump_fib(). On the next loop iteration in
rtnl_dump_all(), 'memset(&cb, 0, sizeof(cb->args))' can reset
that pointer, leaking the memory [1].

On the second thought we could as well save the state of fib6_walker
in inet6_dump_fib() with cb->data. That should fix the leak too.
Is it sounds reasonable?

Thanks,
Alexey

 

On 10/31/18 12:42 AM, Alexey Kodanev wrote:

cb->args[2] can store the pointer to the struct fib6_walker,
allocated in inet6_dump_fib(). On the next loop iteration in
rtnl_dump_all(), 'memset(&cb, 0, sizeof(cb->args))' can reset
that pointer, leaking the memory [1].

Fix it by calling cb->done, if it is set, before filling 'cb->args'
with zeros.

Looks like the recent changes in rtnl_dump_all() contributed to
the appearance of this kmemleak [1], commit c63586dc9b3e ("net:
rtnl_dump_all needs to propagate error from dumpit function")
breaks the loop only on an error now.

...

It is more efficient to keep going.

I think the simplest fix for 4.20 is to break the loop if ret is non-0 -
restore the previous behavior. For net-next I think the done callback is
not needed for ipv6; I think there is a simpler way to do it.

On 10/31/18 10:55 AM, David Ahern wrote:

I think the simplest fix for 4.20 is to break the loop if ret is non-0 -
restore the previous behavior. 

that is the only recourse. It has to bail if ret is non-0. Do you want
to send a patch with that fix?

On 10/31/2018 08:35 PM, David Ahern wrote:

On 10/31/18 10:55 AM, David Ahern wrote:

quoted

I think the simplest fix for 4.20 is to break the loop if ret is non-0 -
restore the previous behavior. 

that is the only recourse. It has to bail if ret is non-0. Do you want
to send a patch with that fix?

I see, and inet6_dump_fib() cleanups fib6_walker if ret is zero. Will send the fix.

On 11/01/2018 04:11 PM, Alexey Kodanev wrote:

On 10/31/2018 08:35 PM, David Ahern wrote:

quoted

On 10/31/18 10:55 AM, David Ahern wrote:

quoted

I think the simplest fix for 4.20 is to break the loop if ret is non-0 -
restore the previous behavior. 

that is the only recourse. It has to bail if ret is non-0. Do you want
to send a patch with that fix?

I see, and inet6_dump_fib() cleanups fib6_walker if ret is zero. Will send the fix.

Can it happen that inet6_dump_fib() returns skb->len (0) in the below cases?

* 	if (arg.filter.flags & RTM_F_CLONED)
		return skb->len;

        ...

	w = (void *)cb->args[2];
	if (!w) {
		...
		w = kzalloc(...)
                ...

* 	if (arg.filter.table_id) {
                ...
		if (!tb) {
			if (arg.filter.dump_all_families)
				return skb->len;


Would it be safer to add "res = skb->len; goto out;" instead of "return skb->len;"
so that it can call fib6_dump_end() for "res <= 0"? Or use cb->data instead of
cb->args?

On 11/1/18 7:42 AM, Alexey Kodanev wrote:

On 11/01/2018 04:11 PM, Alexey Kodanev wrote:

quoted

On 10/31/2018 08:35 PM, David Ahern wrote:

quoted

On 10/31/18 10:55 AM, David Ahern wrote:

quoted

I think the simplest fix for 4.20 is to break the loop if ret is non-0 -
restore the previous behavior. 

that is the only recourse. It has to bail if ret is non-0. Do you want
to send a patch with that fix?

I see, and inet6_dump_fib() cleanups fib6_walker if ret is zero. Will send the fix.

Can it happen that inet6_dump_fib() returns skb->len (0) in the below cases?

* 	if (arg.filter.flags & RTM_F_CLONED)
		return skb->len;

        ...

	w = (void *)cb->args[2];
	if (!w) {
		...
		w = kzalloc(...)
                ...

* 	if (arg.filter.table_id) {
                ...
		if (!tb) {
			if (arg.filter.dump_all_families)
				return skb->len;


Would it be safer to add "res = skb->len; goto out;" instead of "return skb->len;"
so that it can call fib6_dump_end() for "res <= 0"? Or use cb->data instead of
cb->args?

Since res is initialized to 0, both of those can just be 'goto out;'
The break in dump_all is still needed though.