On Wed, Jun 30, 2021 at 08:57:53AM +0200, Steffen Klassert wrote:

On Mon, Jun 28, 2021 at 03:34:28PM +0200, Frederic Weisbecker wrote:

quoted

xfrm_bydst_resize() calls synchronize_rcu() while holding
hash_resize_mutex. But then on PREEMPT_RT configurations,
xfrm_policy_lookup_bytype() may acquire that mutex while running in an
RCU read side critical section. This results in a deadlock.

In fact the scope of hash_resize_mutex is way beyond the purpose of
xfrm_policy_lookup_bytype() to just fetch a coherent and stable policy
for a given destination/direction, along with other details.

The lower level net->xfrm.xfrm_policy_lock, which among other things
protects per destination/direction references to policy entries, is
enough to serialize and benefit from priority inheritance against the
write side. As a bonus, it makes it officially a per network namespace
synchronization business where a policy table resize on namespace A
shouldn't block a policy lookup on namespace B.

Fixes: 77cc278f7b20 (xfrm: policy: Use sequence counters with associated lock)
Cc: stable@vger.kernel.org
Cc: Ahmed S. Darwish <redacted>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Varad Gautam <redacted>
Cc: Steffen Klassert <steffen.klassert@secunet.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: David S. Miller <davem@davemloft.net>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>

Your patch has a conflicht with ("commit d7b0408934c7 xfrm: policy: Read
seqcount outside of rcu-read side in xfrm_policy_lookup_bytype")
from Varad. Can you please rebase onto the ipsec tree?

This patch is now applied to the ipsec tree (on top of the
revert of commit d7b0408934c7).

Thanks everyone!