Re: [PATCH] xfrm: policy: Restructure RCU-read locking in xfrm_sk_policy_lookup

On Tue, Jun 22, 2021 at 01:51:24PM +0200, Frederic Weisbecker wrote:

On Tue, Jun 22, 2021 at 01:21:59PM +0200, Steffen Klassert wrote:

quoted

On Mon, Jun 21, 2021 at 01:05:28PM +0200, Steffen Klassert wrote:

quoted

On Mon, Jun 21, 2021 at 11:11:18AM +0200, Varad Gautam wrote:

quoted

Right, I misread the call chain - security_xfrm_policy_lookup does not reach
xfrm_policy_lookup, making this patch unnecessary. The bug I have is:

T1, holding hash_resize_mutex and sleeping inside synchronize_rcu:

__schedule
schedule
schedule_timeout
wait_for_completion
__wait_rcu_gp
synchronize_rcu
xfrm_hash_resize

And T2 producing RCU-stalls since it blocked on the mutex:

__schedule
schedule
__rt_mutex_slowlock
rt_mutex_slowlock_locked
rt_mutex_slowlock
xfrm_policy_lookup_bytype.constprop.77

Ugh, why does xfrm_policy_lookup_bytype use a mutex? This is called
in the receive path inside a sofirq.

The bug was introduced by: 

commit 77cc278f7b202e4f16f8596837219d02cb090b96
Author: Ahmed S. Darwish [off-list ref]
Date:   Mon Jul 20 17:55:22 2020 +0200

    xfrm: policy: Use sequence counters with associated lock

    A sequence counter write side critical section must be protected by some
    form of locking to serialize writers. If the serialization primitive is
    not disabling preemption implicitly, preemption has to be explicitly
    disabled before entering the sequence counter write side critical
    section.

    A plain seqcount_t does not contain the information of which lock must
    be held when entering a write side critical section.

    Use the new seqcount_spinlock_t and seqcount_mutex_t data types instead,
    which allow to associate a lock with the sequence counter. This enables
    lockdep to verify that the lock used for writer serialization is held
    when the write side critical section is entered.

    If lockdep is disabled this lock association is compiled out and has
    neither storage size nor runtime overhead.

    Signed-off-by: Ahmed S. Darwish [off-list ref]
    Signed-off-by: Peter Zijlstra (Intel) [off-list ref]
    Link: https://lkml.kernel.org/r/20200720155530.1173732-17-a.darwish@linutronix.de

This uses a seqcount_mutex_t for xfrm_policy_hash_generation, that's
wrong.

Varad, can you try to replace the seqcount_mutex_t for xfrm_policy_hash_generation
by a seqcount_spinlock_t? I'm not familiar with that seqcount changes,
but we should not end up with using a mutex in this codepath.

Something like this? (beware, untested, also I don't know if the read side
should then disable bh, doesn't look necessary for PREEMPT_RT, but I may be
missing something...)

Looking a bit deeper into this it seems that the problem is that
xfrm_policy_hash_generation and hash_resize_mutex do not protect
the same thing.

hash_resize_mutex protects user configuration against a worker thread
that rebalances the hash buckets. xfrm_policy_hash_generation protects
user configuration against the data path that runs in softirq.

Finally the following line from xfrm_init() relates these two:

seqcount_mutex_init(&xfrm_policy_hash_generation, &hash_resize_mutex);

That looks a bit odd. This line was also introduced with the above
mentioned patch.