Received-SPF: pass (google.com: domain of linux-kernel+bounces-181348-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
From: Valentin Schneider <vschneid@redhat.com>
To: Frederic Weisbecker <frederic@kernel.org>
Cc: LKML <linux-kernel@vger.kernel.org>, "Paul E . McKenney"
 <paulmck@kernel.org>, Boqun Feng <boqun.feng@gmail.com>, Joel Fernandes
 <joel@joelfernandes.org>, Neeraj Upadhyay <neeraj.upadhyay@amd.com>,
 Uladzislau Rezki <urezki@gmail.com>, Zqiang <qiang.zhang1211@gmail.com>,
 rcu <rcu@vger.kernel.org>
Subject: Re: [PATCH 2/6] rcu: Remove superfluous full memory barrier upon
 first EQS snapshot
In-Reply-To: <ZkYvemdrEOVFNtVu@lothringen>
References: <20240515125332.9306-1-frederic@kernel.org>
 <20240515125332.9306-3-frederic@kernel.org>
 <xhsmhfruhhixv.mognet@vschneid-thinkpadt14sgen2i.remote.csb>
 <ZkYvemdrEOVFNtVu@lothringen>
Date: Thu, 16 May 2024 19:08:57 +0200
Message-ID: <xhsmha5kphefq.mognet@vschneid-thinkpadt14sgen2i.remote.csb>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain

On 16/05/24 18:08, Frederic Weisbecker wrote:
> On Thu, May 16, 2024 at 05:31:40PM +0200, Valentin Schneider wrote:
>> On 15/05/24 14:53, Frederic Weisbecker wrote:
>> > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
>> > index 58415cdc54f8..f5354de5644b 100644
>> > --- a/kernel/rcu/tree.c
>> > +++ b/kernel/rcu/tree.c
>> > @@ -773,7 +773,12 @@ static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
>> >   */
>> >  static int dyntick_save_progress_counter(struct rcu_data *rdp)
>> >  {
>> > -	rdp->dynticks_snap = rcu_dynticks_snap(rdp->cpu);
>>
>> So for PPC, which gets the smp_mb() at the lock acquisition, this is an
>> "obvious" redundant smp_mb().
>>
>> For the other archs, per the definition of smp_mb__after_unlock_lock() it
>> seems implied that UNLOCK+LOCK is a full memory barrier, but I wanted to
>> see it explicitly stated somewhere. From a bit of spelunking below I still
>> think it's the case, but is there a "better" source of truth?
>>
>>   01352fb81658 ("locking: Add an smp_mb__after_unlock_lock() for UNLOCK+BLOCK barrier")
>>   """
>>   The Linux kernel has traditionally required that an UNLOCK+LOCK pair act as a
>>   full memory barrier when either (1) that UNLOCK+LOCK pair was executed by the
>>   same CPU or task, or (2) the same lock variable was used for the UNLOCK and
>>   LOCK.
>>   """
>>
>> and
>>
>>   https://lore.kernel.org/all/1436789704-10086-1-git-send-email-will.deacon@arm.com/
>>   """
>>   This ordering guarantee is already provided without the barrier on
>>   all architectures apart from PowerPC
>>   """
>
> You seem to have found the accurate informations! But I must admit
> they are hard to find and it would be welcome to document that properly, for example
> in Documentation/memory-barriers.txt
>
> I think the reason is that it's not supposed to be used outside RCU, perhaps
> because its semantics are too fragile to use for general purpose? Even that
> could be stated along in Documentation/memory-barriers.txt
>

That's also what I suspected when I stumbled on

  12d560f4ea87 ("rcu,locking: Privatize smp_mb__after_unlock_lock()")

which removed the references to it from Documentation/memory-barriers.txt

> Another thing is that its semantics are similar to smp_mb__after_spinlock()
> (itself badly documented), although slightly different. I'm not even completely
> sure how. I assume that smp_mb__after_spinlock() can be just used once to
> produce the required ordering and subsequent lock on that spinlock don't need
> to repeat the barrier to propagate the ordering against what is before the
> smp_mb__after_spinlock. However IUUC smp_mb__after_unlock_lock() has to be
> chained/repeated on all subsequent locking of the spinlock...

IIUC (big if) the chaining is a requirement of RCU itself, per:

  2a67e741bbbc ("rcu: Create transitive rnp->lock acquisition functions")

   * Because the rcu_nodes form a tree, the tree traversal locking will observe
   * different lock values, this in turn means that an UNLOCK of one level
   * followed by a LOCK of another level does not imply a full memory barrier;
   * and most importantly transitivity is lost.
   *
   * In order to restore full ordering between tree levels, augment the regular
   * lock acquire functions with smp_mb__after_unlock_lock().