Re: rtnl_mutex deadlock?

On 08/06/2015 04:50 PM, Daniel Borkmann wrote:

On 08/06/2015 02:30 AM, Herbert Xu wrote:

quoted

On Wed, Aug 05, 2015 at 08:59:07PM +0200, Daniel Borkmann wrote:

quoted

Here's a theory and patch below. Herbert, Thomas, does this make any
sense to you resp. sound plausible? ;)

It's certainly possible.  Whether it's plausible I'm not so sure.
The netlink hashtable is unlimited in size.  So it should always
be expanding, not rehashing.  The bug you found should only affect
rehashing.

quoted

I'm not quite sure what's best to return from here, i.e. whether we
propagate -ENOMEM or instead retry over and over again hoping that the
rehashing completed (and no new rehashing started in the mean time) ...../net/ipv4/af_inet.c:172:static

Please use something other than ENOMEM as it is already heavily
used in this context.  Perhaps EOVERFLOW?

Okay, I'll do that.

quoted

We should probably add a WARN_ON_ONCE in rhashtable_insert_rehash
since two concurrent rehashings indicates something is going
seriously wrong.

So, if I didn't miss anything, it looks like the following could have
happened: the worker thread, that is rht_deferred_worker(), itself could
trigger the first rehashing, e.g. after shrinking or expanding (or also
in case none of both happen).

Then, in __rhashtable_insert_fast(), I could trigger an -EBUSY when I'm
really unlucky and exceed the ht->elasticity limit of 16. I would then
end up in rhashtable_insert_rehash() to find out there's already one
ongoing and thus, I'm getting -EBUSY via __netlink_insert().

Perhaps that is what could have happened? Seems rare though, but it was
also only seen rarely so far ...

Experimenting a bit more, letting __netlink_insert() return -EBUSY so far,
I only managed when either artificially reducing ht->elasticity limit a bit
or biasing the hash function, that means, it would require some specific
knowledge at what slot we end up to overcome the elasticity limit and thus
trigger rehashing. Pretty unlikely though if you ask me. The other thing
I could observe, when I used the bind stress test from Thomas' repo and
reduced the amount of bind()'s, so that we very frequently fluctuate in the
ranges of 4 to 256 of the hashtable size, I could observe that we from time
to time enter rhashtable_insert_rehash() on insertions, but probably the
window was too small to trigger an error. I think in any case, remapping
seems okay.