Re: net: deadlock between ip_expire/sch_direct_xmit

On Tue, 2017-03-14 at 08:09 -0700, Eric Dumazet wrote:

On Tue, Mar 14, 2017 at 8:03 AM, Paolo Abeni [off-list ref] wrote:

quoted

I'm wondering if we really need to keep the fragment queue lock held
while sending the icmp packet ? we hold a reference to the struct, so
it can't be deleted, and AFAICS after ipq_kill() nobody else could
access/modify that queue.

That lock is there pretty much forever, but perhaps is only a leftover
and we can release it just after ipq_kill() ?

Maybe, but for peace of mind I would make sure this code path owns the
skb (head) before releasing the lock.

Seems something to try for net-next ?

Agreed.

I asked because I was in doubt I missed something obvious.

Thank you,

Paolo