Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
From:   Joel Fernandes <joel@joelfernandes.org>
Mime-Version: 1.0 (1.0)
Subject: Re: [RFC 0/2] srcu: Remove pre-flip memory barrier
Date:   Thu, 22 Dec 2022 13:56:17 -0500
Message-Id: <F492F726-00AA-4FC8-A5E5-BBF006CE946C@joelfernandes.org>
References: <20221222185314.GR4001@paulmck-ThinkPad-P17-Gen-1>
Cc:     Frederic Weisbecker <frederic@kernel.org>,
        Mathieu Desnoyers <mathieu.desnoyers@efficios.com>,
        linux-kernel@vger.kernel.org,
        Josh Triplett <josh@joshtriplett.org>,
        Lai Jiangshan <jiangshanlai@gmail.com>, rcu@vger.kernel.org,
        Steven Rostedt <rostedt@goodmis.org>
In-Reply-To: <20221222185314.GR4001@paulmck-ThinkPad-P17-Gen-1>
To:     paulmck@kernel.org
Precedence: bulk


> On Dec 22, 2022, at 1:53 PM, Paul E. McKenney <paulmck@kernel.org> wrote:
>=20
> =EF=BB=BFOn Thu, Dec 22, 2022 at 01:19:06PM -0500, Joel Fernandes wrote:
>>=20
>>=20
>>>> On Dec 22, 2022, at 11:43 AM, Paul E. McKenney <paulmck@kernel.org> wro=
te:
>>>=20
>>> =EF=BB=BFOn Thu, Dec 22, 2022 at 01:40:10PM +0100, Frederic Weisbecker w=
rote:
>>>>> On Wed, Dec 21, 2022 at 12:11:42PM -0500, Mathieu Desnoyers wrote:
>>>>> On 2022-12-21 06:59, Frederic Weisbecker wrote:
>>>>>>> On Tue, Dec 20, 2022 at 10:34:19PM -0500, Mathieu Desnoyers wrote:
>>>>> [...]
>>>>>>>=20
>>>>>>> The memory ordering constraint I am concerned about here is:
>>>>>>>=20
>>>>>>> * [...] In addition,
>>>>>>> * each CPU having an SRCU read-side critical section that extends be=
yond
>>>>>>> * the return from synchronize_srcu() is guaranteed to have executed a=

>>>>>>> * full memory barrier after the beginning of synchronize_srcu() and b=
efore
>>>>>>> * the beginning of that SRCU read-side critical section. [...]
>>>>>>>=20
>>>>>>> So if we have a SRCU read-side critical section that begins after th=
e beginning
>>>>>>> of synchronize_srcu, but before its first memory barrier, it would m=
iss the
>>>>>>> guarantee that the full memory barrier is issued before the beginnin=
g of that
>>>>>>> SRCU read-side critical section. IOW, that memory barrier needs to b=
e at the
>>>>>>> very beginning of the grace period.
>>>>>>=20
>>>>>> I'm confused, what's wrong with this ?
>>>>>>=20
>>>>>> UPDATER                  READER
>>>>>> -------                  ------
>>>>>> STORE X =3D 1              STORE srcu_read_lock++
>>>>>> // rcu_seq_snap()        smp_mb()
>>>>>> smp_mb()                 READ X
>>>>>> // scans
>>>>>> READ srcu_read_lock
>>>>>=20
>>>>> What you refer to here is only memory ordering of the store to X and l=
oad
>>>>> from X wrt loading/increment of srcu_read_lock, which is internal to t=
he
>>>>> srcu implementation. If we really want to model the provided high-leve=
l
>>>>> memory ordering guarantees, we should consider a scenario where SRCU i=
s used
>>>>> for its memory ordering properties to synchronize other variables.
>>>>>=20
>>>>> I'm concerned about the following Dekker scenario, where synchronize_s=
rcu()
>>>>> and srcu_read_lock/unlock would be used instead of memory barriers:
>>>>>=20
>>>>> Initial state: X =3D 0, Y =3D 0
>>>>>=20
>>>>> Thread A                   Thread B
>>>>> ---------------------------------------------
>>>>> STORE X =3D 1                STORE Y =3D 1
>>>>> synchronize_srcu()
>>>>>                          srcu_read_lock()
>>>>>                          r1 =3D LOAD X
>>>>>                          srcu_read_unlock()
>>>>> r0 =3D LOAD Y
>>>>>=20
>>>>> BUG_ON(!r0 && !r1)
>>>>>=20
>>>>> So in the synchronize_srcu implementation, there appears to be two
>>>>> major scenarios: either srcu_gp_start_if_needed starts a gp or expedit=
ed gp,
>>>>> or it uses an already started gp/expedited gp. When snapshotting with
>>>>> rcu_seq_snap, the fact that the memory barrier is after the ssp->srcu_=
gp_seq
>>>>> load means that it does not order prior memory accesses before that lo=
ad.
>>>>> This sequence value is then used to identify which gp_seq to wait for w=
hen
>>>>> piggy-backing on another already-started gp. I worry about reordering
>>>>> between STORE X =3D 1 and load of ssp->srcu_gp_seq, which is then used=
 to
>>>>> piggy-back on an already-started gp.
>>>>>=20
>>>>> I suspect that the implicit barrier in srcu_read_lock() invoked at the=

>>>>> beginning of srcu_gp_start_if_needed() is really the barrier that make=
s
>>>>> all this behave as expected. But without documentation it's rather har=
d to
>>>>> follow.
>>>>=20
>>>> Oh ok I see now. It might be working that way by accident or on forgott=
en
>>>> purpose. In any case, we really want to add a comment above that
>>>> __srcu_read_lock_nmisafe() call.
>>>=20
>>> Another test for the safety (or not) of removing either D or E is
>>> to move that WRITE_ONCE() to follow (or, respectively, precede) the
>>> adjacent scans.
>>=20
>> Good idea, though I believe the MBs that the above talk about are not the=
 flip ones. They are the ones in synchronize_srcu() beginning and end, that o=
rder with respect to grace period start and end.
>>=20
>> So that (flipping MBs) is unrelated, or did I miss something?
>=20
> The thought is to manually similate in the source code the maximum
> memory-reference reordering that a maximally hostile compiler and CPU
> would be permitted to carry out.  So yes, given that there are other
> memory barriers before and after, these other memory barriers limit how
> far the flip may be moved in the source code.
>=20
> Here I am talking about the memory barriers associated with the flip,
> but the same trick can of course be applied to other memory barriers.
> In general, remove a given memory barrier and (in the source code)
> maximally rearrange the memory references that were previously ordered
> by the memory barrier in question.
>=20
> Again, the presence of other memory barriers will limit the permitted
> maximal source-code rearrangement.


Makes sense if the memory barrier is explicit. In this case, the memory barr=
iers are implicit apparently, with a srcu_read_lock() in the beginning of sy=
nchronize_rcu() having the implicit / indirect memory barrier. So I am not s=
ure if that can be implemented without breaking SRCU readers.

Thanks.


>=20
>                            Thanx, Paul