Content-Type: text/plain;
        charset=us-ascii
Mime-Version: 1.0 (Mac OS X Mail 11.5 \(3445.9.1\))
Subject: Re: [PATCH] SUNRPC: Don't allow compiler optimisation of
 svc_xprt_release_slot()
From:   Chuck Lever <chuck.lever@oracle.com>
In-Reply-To: <BB7240BE-3F20-43AD-B709-7F18CC33CC02@oracle.com>
Date:   Fri, 11 Jan 2019 16:54:01 -0500
Cc:     Trond Myklebust <trondmy@hammerspace.com>,
        Linux NFS Mailing List <linux-nfs@vger.kernel.org>
Content-Transfer-Encoding: quoted-printable
Message-Id: <6F5B73B7-E9F8-4FDB-8381-E5C02772C6A5@oracle.com>
References: <20190103141712.24381-1-trond.myklebust@hammerspace.com>
 <20190103224529.GA6907@fieldses.org>
 <a0493a5f022a666779d19fde5cdfbfdf9c7316c4.camel@hammerspace.com>
 <20190104173912.GC11787@fieldses.org> <20190107213218.GD7753@fieldses.org>
 <e14e17167ef98c456e7648ad0536827b7870f0ff.camel@hammerspace.com>
 <20190108150107.GA15921@fieldses.org>
 <4077991d3d3acee4c37c7c8c6dc2b76930c9584e.camel@hammerspace.com>
 <20190109165142.GB32189@fieldses.org>
 <300445038b75d5efafe9391eb4b8e83d9d6e3633.camel@hammerspace.com>
 <20190111211235.GA27206@fieldses.org>
 <BB7240BE-3F20-43AD-B709-7F18CC33CC02@oracle.com>
To:     Bruce Fields <bfields@fieldses.org>
Sender: linux-nfs-owner@vger.kernel.org
Precedence: bulk


> On Jan 11, 2019, at 4:52 PM, Chuck Lever <chuck.lever@oracle.com> =
wrote:
>=20
>=20
>=20
>> On Jan 11, 2019, at 4:12 PM, bfields@fieldses.org wrote:
>>=20
>> On Wed, Jan 09, 2019 at 05:41:36PM +0000, Trond Myklebust wrote:
>>> The above is stating that
>>>=20
>>> smp_rmb();
>>> smp_read_barrier_depends();
>>> if (xprt->xpt_flags & ....)
>>>=20
>>> is redundant and can be replaced with just
>>>=20
>>> smp_rmb();
>>> if (xprt->xpt_flags & ....)
>>>=20
>>> However that's not the case for smp_rmb() followed by READ_ONCE(). =
That
>>> would expand to
>>>=20
>>> smp_rmb();
>>> if (xprt->xpt_flags & ...) {
>>>   smp_read_barrier_depends();
>>> } else
>>>   smp_read_barrier_depends();
>>>=20
>>> which is not redundant. It is ensuring (on alpha only) that the read =
of
>>> xprt->xpt_flags is also not re-ordered w.r.t. other data reads that
>>> follow.
>>>=20
>>> See, for instance, kernel/events/core.c which has several examples, =
or
>>> kernel/exit.c.
>>=20
>> You're right, I was confused.
>>=20
>> So, I think we need your patch plus something like this.
>>=20
>> Chuck, maybe you could help me with the "XXX: Chuck:" parts?
>=20
> I haven't been following. Why do you think those are necessary?
> We've had set_bit and atomic_{inc,dec} in this code for ages,
> and I've never noticed a problem.
>=20
> Rather than adding another CPU pipeline bubble in the RDMA code,
> though, could you simply move the set_bit() call site inside the
> critical sections?

er, inside the preceding critical section. Just reverse the order
of the spin_unlock and the set_bit.


>=20
>=20
>> (This applies on top of your patch plus another that just renames the
>> stupidly long svc_xprt_has_something_to_do() to svc_xprt_read().)
>>=20
>> --b.
>>=20
>> commit d7356c3250d4
>> Author: J. Bruce Fields <bfields@redhat.com>
>> Date:   Fri Jan 11 15:36:40 2019 -0500
>>=20
>>   svcrpc: fix unlikely races preventing queueing of sockets
>>=20
>>   In the rpc server, When something happens that might be reason to =
wake
>>   up a thread to do something, what we do is
>>=20
>>           - modify xpt_flags, sk_sock->flags, xpt_reserved, or
>>             xpt_nr_rqsts to indicate the new situation
>>           - call svc_xprt_enqueue() to decide whether to wake up a =
thread.
>>=20
>>   svc_xprt_enqueue may require multiple conditions to be true before
>>   queueing up a thread to handle the xprt.  In the SMP case, one of =
the
>>   other CPU's may have set another required condition, and in that =
case,
>>   although both CPUs run svc_xprt_enqueue(), it's possible that =
neither
>>   call sees the writes done by the other CPU in time, and neither one
>>   recognizes that all the required conditions have been set.  A =
socket
>>   could therefore be ignored indefinitely.
>>=20
>>   Add memory barries to ensure that any svc_xprt_enqueue() call will
>>   always see the conditions changed by other CPUs before deciding to
>>   ignore a socket.
>>=20
>>   I've never seen this race reported.  In the unlikely event it =
happens,
>>   another event will usually come along and the problem will fix =
itself.
>>   So I don't think this is worth backporting to stable.
>>=20
>>   Signed-off-by: J. Bruce Fields <bfields@redhat.com>
>>=20
>> diff --git a/net/sunrpc/svc_xprt.c b/net/sunrpc/svc_xprt.c
>> index d410ae512b02..2af21b84b3b6 100644
>> --- a/net/sunrpc/svc_xprt.c
>> +++ b/net/sunrpc/svc_xprt.c
>> @@ -357,6 +357,7 @@ static void svc_xprt_release_slot(struct svc_rqst =
*rqstp)
>> 	struct svc_xprt	*xprt =3D rqstp->rq_xprt;
>> 	if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
>> 		atomic_dec(&xprt->xpt_nr_rqsts);
>> +		smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
>> 		svc_xprt_enqueue(xprt);
>> 	}
>> }
>> @@ -365,6 +366,15 @@ static bool svc_xprt_ready(struct svc_xprt =
*xprt)
>> {
>> 	unsigned long xpt_flags;
>>=20
>> +	/*
>> +	 * If another cpu has recently updated xpt_flags,
>> +	 * sk_sock->flags, xpt_reserved, or xpt_nr_rqsts, we need to
>> +	 * know about it; otherwise it's possible that both that cpu and
>> +	 * this one could call svc_xprt_enqueue() without either
>> +	 * svc_xprt_enqueue() recognizing that the conditions below
>> +	 * are satisfied, and we could stall indefinitely:
>> +	 */
>> +	smp_rmb();
>> 	READ_ONCE(xprt->xpt_flags);
>>=20
>> 	if (xpt_flags & (BIT(XPT_CONN) | BIT(XPT_CLOSE)))
>> @@ -479,7 +489,7 @@ void svc_reserve(struct svc_rqst *rqstp, int =
space)
>> 	if (xprt && space < rqstp->rq_reserved) {
>> 		atomic_sub((rqstp->rq_reserved - space), =
&xprt->xpt_reserved);
>> 		rqstp->rq_reserved =3D space;
>> -
>> +		smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
>> 		svc_xprt_enqueue(xprt);
>> 	}
>> }
>> diff --git a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c =
b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
>> index 828b149eaaef..377244992ae8 100644
>> --- a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
>> +++ b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
>> @@ -316,6 +316,7 @@ static void svc_rdma_wc_receive(struct ib_cq *cq, =
struct ib_wc *wc)
>> 	list_add_tail(&ctxt->rc_list, &rdma->sc_rq_dto_q);
>> 	spin_unlock(&rdma->sc_rq_dto_lock);
>> 	set_bit(XPT_DATA, &rdma->sc_xprt.xpt_flags);
>> +	/* XXX: Chuck: do we need an smp_mb__after_atomic() here? */
>> 	if (!test_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags))
>> 		svc_xprt_enqueue(&rdma->sc_xprt);
>> 	goto out;
>> diff --git a/net/sunrpc/xprtrdma/svc_rdma_rw.c =
b/net/sunrpc/xprtrdma/svc_rdma_rw.c
>> index dc1951759a8e..e1a790487d69 100644
>> --- a/net/sunrpc/xprtrdma/svc_rdma_rw.c
>> +++ b/net/sunrpc/xprtrdma/svc_rdma_rw.c
>> @@ -290,6 +290,7 @@ static void svc_rdma_wc_read_done(struct ib_cq =
*cq, struct ib_wc *wc)
>> 		spin_unlock(&rdma->sc_rq_dto_lock);
>>=20
>> 		set_bit(XPT_DATA, &rdma->sc_xprt.xpt_flags);
>> +		/* XXX: Chuck: do we need a smp_mb__after_atomic() here? =
*/
>> 		svc_xprt_enqueue(&rdma->sc_xprt);
>> 	}
>>=20
>=20
> --
> Chuck Lever

--
Chuck Lever