ipv6: anycast: insert aca into global hash under idev->lock

ipv6: anycast: insert aca into global hash under idev->lock

syzbot reported a splat [1]: a slab-use-after-free in
ipv6_chk_acast_addr(), which walks the global inet6_acaddr_lst[] hash
under RCU and dereferences a struct ifacaddr6 that has already been
freed while still linked in the hash, so a later reader walks into a
dangling node.

In __ipv6_dev_ac_inc() the aca is allocated with refcount 1, then
aca_get() bumps it to 2 to keep it alive across the unlocked region.
It is published to idev->ac_list under idev->lock, but
ipv6_add_acaddr_hash() runs after write_unlock_bh(). A concurrent
teardown (ipv6_ac_destroy_dev() from addrconf_ifdown(), under RTNL)
can slip into that window:

  CPU0 __ipv6_dev_ac_inc           CPU1 ipv6_ac_destroy_dev (RTNL)
  ------------------------------   ------------------------------------
  aca_alloc()              refcnt 1
  aca_get()               refcnt 2
  write_lock_bh(idev->lock)
    add aca to ac_list
  write_unlock_bh(idev->lock)
                                   write_lock_bh(idev->lock)
                                     pull aca off ac_list
                                   write_unlock_bh(idev->lock)
                                   ipv6_del_acaddr_hash(aca)
                                     hlist_del_init_rcu() is a no-op,
                                     aca is not in the hash yet
                                   aca_put()           refcnt 2->1
  ipv6_add_acaddr_hash(aca)
    aca now inserted into the hash
  aca_put()                refcnt 1->0
    call_rcu(aca_free_rcu) -> kfree(aca)

The hash removal becomes a no-op because the insertion has not
happened yet, so once CPU0 inserts and drops the last reference, the
aca is freed while still linked in inet6_acaddr_lst[], and readers
dereference freed memory after the slab slot is reused.

This window opened once RTNL stopped serializing the join path against
device teardown. Move ipv6_add_acaddr_hash() inside the idev->lock
section so the ac_list and hash insertions are atomic with respect to
teardown: a racing remover now either misses the aca entirely or finds
it in both lists.

acaddr_hash_lock is now nested under idev->lock, which is acquired in
softirq context, so switch all acaddr_hash_lock sites to spin_lock_bh()
to avoid the irq lock inversion reported in [2].

[1] https://syzkaller.appspot.com/bug?extid=a01df04303c131efbf3a
[2] https://lore.kernel.org/netdev/6a194ef7.ba3b1513.1890b4.0000.GAE@google.com/

Reported-by: syzbot+819eb928d120d2bdad0e@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/all/6a191f87.ce022c6e.138e56.0003.GAE@google.com/T/
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Fixes: eb1ac9ff6c4a ("ipv6: anycast: Don't hold RTNL for IPV6_JOIN_ANYCAST.")
Signed-off-by: Jiayuan Chen <jiayuan.chen@linux.dev>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Link: https://patch.msgid.link/20260529152219.235475-1-jiayuan.chen@linux.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>