Re: [PATCH v1 net] net: Remove RTNL dance for SIOCBRADDIF and SIOCBRDELIF.

On Thu, Mar 13, 2025 at 05:59:55PM -0700, Kuniyuki Iwashima wrote:

SIOCBRDELIF is passed to dev_ioctl() first and later forwarded to
br_ioctl_call(), which causes unnecessary RTNL dance and the splat
below [0] under RTNL pressure.

Let's say Thread A is trying to detach a device from a bridge and
Thread B is trying to remove the bridge.

In dev_ioctl(), Thread A bumps the bridge device's refcnt by
netdev_hold() and releases RTNL because the following br_ioctl_call()
also re-acquires RTNL.

In the race window, Thread B could acquire RTNL and try to remove
the bridge device.  Then, rtnl_unlock() by Thread B will release RTNL
and wait for netdev_put() by Thread A.

Thread A, however, must hold RTNL twice after the unlock in dev_ifsioc(),
which may take long under RTNL pressure, resulting in the splat by
Thread B.

  Thread A (SIOCBRDELIF)           Thread B (SIOCBRDELBR)
  ----------------------           ----------------------
  sock_ioctl                       sock_ioctl
  `- sock_do_ioctl                 `- br_ioctl_call
     `- dev_ioctl                     `- br_ioctl_stub
        |- rtnl_lock                     |
        |- dev_ifsioc                    '
        '  |- dev = __dev_get_by_name(...)
           |- netdev_hold(dev, ...)      .
       /   |- rtnl_unlock  ------.       |
       |   |- br_ioctl_call       `--->  |- rtnl_lock
  Race |   |  `- br_ioctl_stub           |- br_del_bridge
  Window   |     |                       |  |- dev = __dev_get_by_name(...)
       |   |     |  May take long        |  `- br_dev_delete(dev, ...)
       |   |     |  under RTNL pressure  |     `- unregister_netdevice_queue(dev, ...)
       |   |     |                |      `- rtnl_unlock
       |   |     |- rtnl_lock  <--|         `- netdev_run_todo
       |   |     |- ...           |            `- netdev_run_todo
       |   |     `- rtnl_unlock   |               |- __rtnl_unlock
       |   |                      |               |- netdev_wait_allrefs_any
       \   |- rtnl_lock  <--------'                  |
           |- netdev_put(dev, ...)  <----------------'  Wait refcnt decrement
                                                        and log splat below

Isn't the race window a bit smaller? dev_ifsioc() does netdev_put()
before rtnl_lock().

To avoid blocking SIOCBRDELBR unnecessarily, let's not call
dev_ioctl() for SIOCBRADDIF and SIOCBRDELIF.

In the dev_ioctl() path, we do the following:

  1. Copy struct ifreq by get_user_ifreq in sock_do_ioctl()
  2. Check CAP_NET_ADMIN in dev_ioctl()
  3. Call dev_load() in dev_ioctl()
  4. Fetch the master dev from ifr.ifr_name in dev_ifsioc()

3. can be done by request_module() in br_ioctl_call(), so we move
1., 2., and 4. to br_ioctl_stub().

Note that 2. is also checked later in add_del_if(), but it's better
performed before RTNL.

SIOCBRADDIF and SIOCBRDELIF have been processed in dev_ioctl() since
the pre-git era, and there seems to be no specific reason to process
them there.

I couldn't find an explanation as well.

Doesn't seem like we have any tests for the IOCTL path, but FWIW I
verified that basic operations using brctl still work after this patch.

[0]:
unregister_netdevice: waiting for wpan3 to become free. Usage count = 2
ref_tracker: wpan3@ffff8880662d8608 has 1/1 users at
     __netdev_tracker_alloc include/linux/netdevice.h:4282 [inline]
     netdev_hold include/linux/netdevice.h:4311 [inline]
     dev_ifsioc+0xc6a/0x1160 net/core/dev_ioctl.c:624
     dev_ioctl+0x255/0x10c0 net/core/dev_ioctl.c:826
     sock_do_ioctl+0x1ca/0x260 net/socket.c:1213
     sock_ioctl+0x23a/0x6c0 net/socket.c:1318
     vfs_ioctl fs/ioctl.c:51 [inline]
     __do_sys_ioctl fs/ioctl.c:906 [inline]
     __se_sys_ioctl fs/ioctl.c:892 [inline]
     __x64_sys_ioctl+0x1a4/0x210 fs/ioctl.c:892
     do_syscall_x64 arch/x86/entry/common.c:52 [inline]
     do_syscall_64+0xcb/0x250 arch/x86/entry/common.c:83
     entry_SYSCALL_64_after_hwframe+0x77/0x7f

Fixes: 893b19587534 ("net: bridge: fix ioctl locking")
Reported-by: syzkaller <redacted>
Reported-by: yan kang <redacted>
Reported-by: yue sun <redacted>
Closes: https://lore.kernel.org/netdev/SY8P300MB0421225D54EB92762AE8F0F2A1D32@SY8P300MB0421.AUSP300.PROD.OUTLOOK.COM/ (local)
Signed-off-by: Kuniyuki Iwashima <redacted>

Thanks for the fix and the detailed commit message. One nit below.

Why not a simple if statement? Unlikely that we will add more commands
to this switch statement.

+		void __user *data;
+		char *colon;
+
+		if (!ns_capable(net->user_ns, CAP_NET_ADMIN))
+			return -EPERM;
+
+		if (get_user_ifreq(&ifr, &data, uarg))
+			return -EFAULT;
+
+		ifr.ifr_name[IFNAMSIZ - 1] = 0;
+		colon = strchr(ifr.ifr_name, ':');
+		if (colon)
+			*colon = 0;
+	}
+	}
 
 	rtnl_lock();