Hi!

On Mon, Mar 17, 2014 at 04:18:53PM -0700, Stephen Hemminger wrote:

IPv6 duplicate address detection is triggering the following assertion
failure when using macvlan + vif + multicast.
 RTNL: assertion failed at net/core/dev.c (4496)

This happens because the DAD timer is adding a multicast address without
acquiring the RTNL mutex. In order to acquire the RTNL mutex, it must be
done in process context; therefore it must be in a workqueue.

Full backtrace:
[  541.030090] RTNL: assertion failed at net/core/dev.c (4496)
[  541.031143] CPU: 0 PID: 0 Comm: swapper/0 Tainted: G           O 3.10.33-1-amd64-vyatta #1
[  541.031145] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007
[  541.031146]  ffffffff8148a9f0 000000000000002f ffffffff813c98c1 ffff88007c4451f8
[  541.031148]  0000000000000000 0000000000000000 ffffffff813d3540 ffff88007fc03d18
[  541.031150]  0000880000000006 ffff88007c445000 ffffffffa0194160 0000000000000000
[  541.031152] Call Trace:
[  541.031153]  <IRQ>  [<ffffffff8148a9f0>] ? dump_stack+0xd/0x17
[  541.031180]  [<ffffffff813c98c1>] ? __dev_set_promiscuity+0x101/0x180
[  541.031183]  [<ffffffff813d3540>] ? __hw_addr_create_ex+0x60/0xc0
[  541.031185]  [<ffffffff813cfe1a>] ? __dev_set_rx_mode+0xaa/0xc0
[  541.031189]  [<ffffffff813d3a81>] ? __dev_mc_add+0x61/0x90
[  541.031198]  [<ffffffffa01dcf9c>] ? igmp6_group_added+0xfc/0x1a0 [ipv6]
[  541.031208]  [<ffffffff8111237b>] ? kmem_cache_alloc+0xcb/0xd0
[  541.031212]  [<ffffffffa01ddcd7>] ? ipv6_dev_mc_inc+0x267/0x300 [ipv6]
[  541.031216]  [<ffffffffa01c2fae>] ? addrconf_join_solict+0x2e/0x40 [ipv6]
[  541.031219]  [<ffffffffa01ba2e9>] ? ipv6_dev_ac_inc+0x159/0x1f0 [ipv6]
[  541.031223]  [<ffffffffa01c0772>] ? addrconf_join_anycast+0x92/0xa0 [ipv6]
[  541.031226]  [<ffffffffa01c311e>] ? __ipv6_ifa_notify+0x11e/0x1e0 [ipv6]
[  541.031229]  [<ffffffffa01c3213>] ? ipv6_ifa_notify+0x33/0x50 [ipv6]

This is the most often case but I fear there are more of them.

addrconf_verify seems unsafe, too, when removing the last ipv6 address. So
does addrconf_prefix_rcv if adding first address.

I wonder if we should put the whole ipv6_ifa_notify infrastructure in a
workqueue? I don't like that either and it could add subtile races.

Those races also seem possible if we only defer addrconf_join_solict,
addrconf_leave_solict, addrconf_join_anycast and addrconf_leave_anycast to
workqueues.

This change is certainly going into the right direction but I am not sure if
we could generalize it.

Greetings,

  Hannes