netlink: lock nl_cb_mutex in netlink_release
Subsystems (2)
RFC 1/1 netlink: lock nl_cb_mutex in netlink_release
While investigating the following bug report on a 6.6 kernel (similar to [1]), I believe I have identified a race condition between __netlink_dump_start(), netlink_release(), and netlink_recvmsg() (which calls netlink_dump()).
My understanding is that __netlink_dump_start() locks nl_cb_mutex, sets cb_running and calls netlink_dump(), but this call into netlink_dump() unlocks nl_cb_mutex and returns before clearing cb_running. Then, netlink_recvmsg() is called, which calls back into netlink_dump(), which then locks nl_cb_mutex, checks to make sure cb_running is set, then proceeds with the dump.
It is at this point after the netlink_dump() checks for cb_running that netlink_release() is called, which tears everything down. While the change in [2] clears cb_running, it does so without holding the lock. This causes the NULL pointer dereference in netlink_dump().
I think this should be resolved by locking the nl_cb_mutex in netlink_release() as seen in the change below.
By locking this mutex early in netlink_release(), then performing the cleanup and setting nlk->cb_running to false prior to releasing the lock, netlink_dump() should be blocked waiting for the lock. Once netlink_dump() acquires the lock, it will fail the check for nlk->cb_running, and therefore jump to the errout_skb label. Since *skb is NULL at this point, the call to kfree_skb() should be harmless.
Additionally, I believe [2] should be backported to stable trees, as it does not exist on 6.6 where this bug was seen. It should be noted that the fixes tag was removed for this patch per the request in [3], as it was considered cleanup since 'code can't be run once we're in release', however in both [1] and the trace collected below, the USB NCM gadget is running, and I am suspicious that a cable unplug event asyncronously causes netlink_release to be called during the netlink_dump(). I will send a separate request to stable for this to be backported.
Note, I was unable to reproduce this bug, and therefore don't have a good way to test this patch, but it seems to be a likely solution. That being said, I do not have extensive background with netlink, and may have overlooked something. Thanks for the thorough review.
Abbreviated trace below:
_______________________________________________________________________
Unable to handle kernel NULL pointer dereference at virtual address 00 ... pstate: 63400005 (nZCv daif +PAN -UAO +TCO +DIT -SSBS BTYPE=--) pc : __pi_strlen (arch/arm64/lib/strlen.S:87) lr : nla_put_string (include/net/netlink.h:1523) sp : ffffffc09e59b7d0 x29: ffffffc09e59b7d0 x28: 0000000000000043 x27: ffffff8976deb000 x26: 00000000c71085df x25: 0000000000000010 x24: 00000000ffffffff x23: 0000000000000000 x22: 0000000000000000 x21: ffffff89dfb97300 x20: 0000000000000038 x19: 0000000000000000 x18: ffffffd1032dc5c0 x17: 00000000d6618229 x16: 0000000000000018 x15: 0000000000000000 x14: 0000000000000000 x13: 0000000000000000 x12: 00000000000029b8 x11: 0000000000003e80 x10: 0000000000000b38 x9 : 000000000a370000 x8 : 0101010101010101 x7 : 0000000000000000 x6 : 0000000000000000 x5 : ffffff800a3729b8 x4 : 0000000000000000 x3 : 0000000000000000 x2 : 0000000000000000 x1 : 0000000000000038 x0 : 0000000000000000 Call trace: __pi_strlen (arch/arm64/lib/strlen.S:87) rtnl_fill_ifinfo (net/core/rtnetlink.c:1935) rtnl_dump_ifinfo (net/core/rtnetlink.c:2276) netlink_dump (net/netlink/af_netlink.c:2257) netlink_recvmsg (net/netlink/af_netlink.c:1978) __sys_recvfrom (net/socket.c:1047 net/socket.c:1069 net/socket.c:2256) __arm64_sys_recvfrom (net/socket.c:2274 net/socket.c:2270 net/socket.c:2270) invoke_syscall (arch/arm64/kernel/syscall.c:0 arch/arm64/kernel/syscall.c:51) el0_svc_common (arch/arm64/kernel/syscall.c:0) do_el0_svc (arch/arm64/kernel/syscall.c:154) el0_svc (arch/arm64/kernel/entry-common.c:136 arch/arm64/kernel/entry-common.c:147 arch/arm64/kernel/entry-common.c:684) el0t_64_sync_handler (arch/arm64/kernel/entry-common.c:0) el0t_64_sync (arch/arm64/kernel/entry.S:599) Code: 92400c04 b200c3e8 f100009f 5400088c (a9400c02)
_______________________________________________________________________
[1] https://lore.kernel.org/all/[email protected]%2F/
[2] https://lore.kernel.org/all/[email protected]/
[3] https://lore.kernel.org/all/[email protected]/
--- net/netlink/af_netlink.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)
net/netlink/af_netlink.c +4 -1
@@ -724,9 +724,11 @@ static int netlink_release(struct socket *sock)if (!sk)return 0 ; + nlk =nlk_sk (sk); +mutex_lock (&nlk->nl_cb_mutex); +netlink_remove (sk);sock_orphan (sk); - nlk =nlk_sk (sk);/* * OK. Socket is unlinked, any packets that arrive now
@@ -773,6 +775,7 @@ static int netlink_release(struct socket *sock)WRITE_ONCE (nlk->cb_running, false); } +mutex_unlock (&nlk->nl_cb_mutex);module_put (nlk->module);if (netlink_is_kernel (sk)) {
Signature · 1 line
2.50.0.727.gbf7dc18ff4-goog
Jakub Kicinski
maintainerQuoted 4 lines · On Tue, 8 Jul 2025 23:04:52 +0000 Neill Kapron wrote:
It is at this point after the
netlink_dump()checks forcb_runningthatnetlink_release()is called, which tears everything down. While the change in [2] clearscb_running, it does so without holding the lock. This causes the NULL pointer dereference innetlink_dump().
Closing a socket while it's being read is pretty unlikely for real life processes. I don't think its even possible, but I'm not a VFS expert.
AFAICT the crash is on:
if (dev->dev.parent && nla_put_string(skb, IFLA_PARENT_DEV_NAME, dev_name(dev->dev.parent))) goto nla_put_failure;
So my first suspicion would be dev.parent getting broken under our feet for the "special" device you're dealing with..