Le 02/09/2016 à 19:24, Cong Wang a écrit :

On Fri, Sep 2, 2016 at 9:39 AM, Cong Wang [off-list ref] wrote:

quoted

On Fri, Sep 2, 2016 at 1:12 AM, Nicolas Dichtel
[off-list ref] wrote:

quoted

Le 02/09/2016 à 06:53, Cong Wang a écrit :

quoted

We never read or change netns id in hardirq context,
the only place we read netns id in softirq context
is in vxlan_xmit(). So, it should be enough to just
disable BH.

Are you sure? Did you audit all part of the code?

I did audit all the callers, and I didn't find any of them in IRQ context.

quoted

peernet2id() is called from netlink core system (do_one_broadcast()). Are you
sure that no driver call this function from an hard irq context?

I audit all callers of netlink_broadcast(), and I don't see any of
them in hardirq context.

Note, you can rule out most of them by checking GFP_KERNEL,
which indicates a process context. ;) For GFP_ATOMIC cases,
I don't see any of them in hardirq context either, but I am definitely
not familiar with drivers like infiniband.

Yes, I was thinking to that one. But I'm also not familiar with it ;-)