Re: [PATCH] net: dsa: Fix lockdep false positive splat

[Florian Fainelli <f.fainelli@gmail.com>]

Le 2/2/19 à 9:53 AM, Marc Zyngier a écrit :
> Creating a macvtap on a DSA-backed interface results in the following
> splat when lockdep is enabled:
> 
> [   19.638080] IPv6: ADDRCONF(NETDEV_CHANGE): lan0: link becomes ready
> [   23.041198] device lan0 entered promiscuous mode
> [   23.043445] device eth0 entered promiscuous mode
> [   23.049255]
> [   23.049557] ============================================
> [   23.055021] WARNING: possible recursive locking detected
> [   23.060490] 5.0.0-rc3-00013-g56c857a1b8d3 #118 Not tainted
> [   23.066132] --------------------------------------------
> [   23.071598] ip/2861 is trying to acquire lock:
> [   23.076171] 00000000f61990cb (_xmit_ETHER){+...}, at: dev_set_rx_mode+0x1c/0x38
> [   23.083693]
> [   23.083693] but task is already holding lock:
> [   23.089696] 00000000ecf0c3b4 (_xmit_ETHER){+...}, at: dev_uc_add+0x24/0x70
> [   23.096774]
> [   23.096774] other info that might help us debug this:
> [   23.103494]  Possible unsafe locking scenario:
> [   23.103494]
> [   23.109584]        CPU0
> [   23.112093]        ----
> [   23.114601]   lock(_xmit_ETHER);
> [   23.117917]   lock(_xmit_ETHER);
> [   23.121233]
> [   23.121233]  *** DEADLOCK ***
> [   23.121233]
> [   23.127325]  May be due to missing lock nesting notation
> [   23.127325]
> [   23.134315] 2 locks held by ip/2861:
> [   23.137987]  #0: 000000003b766c72 (rtnl_mutex){+.+.}, at: rtnetlink_rcv_msg+0x338/0x4e0
> [   23.146231]  #1: 00000000ecf0c3b4 (_xmit_ETHER){+...}, at: dev_uc_add+0x24/0x70
> [   23.153757]
> [   23.153757] stack backtrace:
> [   23.158243] CPU: 0 PID: 2861 Comm: ip Not tainted 5.0.0-rc3-00013-g56c857a1b8d3 #118
> [   23.166212] Hardware name: Globalscale Marvell ESPRESSOBin Board (DT)
> [   23.172843] Call trace:
> [   23.175358]  dump_backtrace+0x0/0x188
> [   23.179116]  show_stack+0x14/0x20
> [   23.182524]  dump_stack+0xb4/0xec
> [   23.185928]  __lock_acquire+0x123c/0x1860
> [   23.190048]  lock_acquire+0xc8/0x248
> [   23.193724]  _raw_spin_lock_bh+0x40/0x58
> [   23.197755]  dev_set_rx_mode+0x1c/0x38
> [   23.201607]  dev_set_promiscuity+0x3c/0x50
> [   23.205820]  dsa_slave_change_rx_flags+0x5c/0x70
> [   23.210567]  __dev_set_promiscuity+0x148/0x1e0
> [   23.215136]  __dev_set_rx_mode+0x74/0x98
> [   23.219167]  dev_uc_add+0x54/0x70
> [   23.222575]  macvlan_open+0x170/0x1d0
> [   23.226336]  __dev_open+0xe0/0x160
> [   23.229830]  __dev_change_flags+0x16c/0x1b8
> [   23.234132]  dev_change_flags+0x20/0x60
> [   23.238074]  do_setlink+0x2d0/0xc50
> [   23.241658]  __rtnl_newlink+0x5f8/0x6e8
> [   23.245601]  rtnl_newlink+0x50/0x78
> [   23.249184]  rtnetlink_rcv_msg+0x360/0x4e0
> [   23.253397]  netlink_rcv_skb+0xe8/0x130
> [   23.257338]  rtnetlink_rcv+0x14/0x20
> [   23.261012]  netlink_unicast+0x190/0x210
> [   23.265043]  netlink_sendmsg+0x288/0x350
> [   23.269075]  sock_sendmsg+0x18/0x30
> [   23.272659]  ___sys_sendmsg+0x29c/0x2c8
> [   23.276602]  __sys_sendmsg+0x60/0xb8
> [   23.280276]  __arm64_sys_sendmsg+0x1c/0x28
> [   23.284488]  el0_svc_common+0xd8/0x138
> [   23.288340]  el0_svc_handler+0x24/0x80
> [   23.292192]  el0_svc+0x8/0xc
> 
> This looks fairly harmless (no actual deadlock occurs), and is
> fixed in a similar way to c6894dec8ea9 ("bridge: fix lockdep
> addr_list_lock false positive splat") by putting the addr_list_lock
> in its own lockdep class.

Great timing, I was just looking at this after solving another one seen
with the bridge code on net-next. AFAIR you can also trigger this with
VLAN and pretty much anything that tries to push UC/MC address list down
to the master device.

Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
-- 
Florian