Message-ID: <20200901150609.GA1506230@google.com>
Date:   Tue, 1 Sep 2020 11:06:09 -0400
From:   Joel Fernandes <joel@...lfernandes.org>
To:     "Paul E. McKenney" <paulmck@...nel.org>
Cc:     linux-kernel@...r.kernel.org, boqun.feng@...il.com,
        dave@...olabs.net, Ingo Molnar <mingo@...hat.com>,
        Josh Triplett <josh@...htriplett.org>,
        Lai Jiangshan <jiangshanlai@...il.com>,
        Madhuparna Bhowmik <madhuparnabhowmik10@...il.com>,
        Mathieu Desnoyers <mathieu.desnoyers@...icios.com>,
        neeraj.iitr10@...il.com, rcu@...r.kernel.org,
        Steven Rostedt <rostedt@...dmis.org>,
        "Uladzislau Rezki (Sony)" <urezki@...il.com>, vineethrp@...il.com
Subject: Re: [PATCH v4 -rcu 1/4] rcu/segcblist: Do not depend on rcl->len to
 store the segcb len during merge

Hi,
Resuming regular activities here, post-LPC :)

On Fri, Aug 28, 2020 at 07:18:55AM -0700, Paul E. McKenney wrote:
> On Thu, Aug 27, 2020 at 06:55:18PM -0400, Joel Fernandes wrote:
> > On Wed, Aug 26, 2020 at 07:20:28AM -0700, Paul E. McKenney wrote:
> > [...]
> > > > > Or better yet, please see below, which should allow getting rid of both
> > > > > of them.
> > > > > 
> > > > > >  	rcu_segcblist_extract_done_cbs(src_rsclp, &donecbs);
> > > > > >  	rcu_segcblist_extract_pend_cbs(src_rsclp, &pendcbs);
> > > > > > -	rcu_segcblist_insert_count(dst_rsclp, &donecbs);
> > > > > > +
> > > > > > +	rcu_segcblist_add_len(dst_rsclp, src_len);
> > > > > >  	rcu_segcblist_insert_done_cbs(dst_rsclp, &donecbs);
> > > > > >  	rcu_segcblist_insert_pend_cbs(dst_rsclp, &pendcbs);
> > > > > 
> > > > > Rather than adding the blank lines, why not have the rcu_cblist structures
> > > > > carry the lengths?  You are already adjusting one of the two call sites
> > > > > that care (rcu_do_batch()), and the other is srcu_invoke_callbacks().
> > > > > That should shorten this function a bit more.  And make callback handling
> > > > > much more approachable, I suspect.
> > > > 
> > > > Sorry, I did not understand. The rcu_cblist structure already has a length
> > > > field. I do modify rcu_segcblist_extract_done_cbs() and
> > > > rcu_segcblist_extract_pend_cbs() to carry the length already, in a later
> > > > patch.
> > > > 
> > > > Just to emphasize, this patch is just a small refactor to avoid an issue in
> > > > later patches. It aims to keep current functionality unchanged.
> > > 
> > > True enough.  I am just suggesting that an equally small refactor in
> > > a slightly different direction should get to a better place.  The key
> > > point enabling this slightly different direction is that this code is
> > > an exception to the "preserve ->cblist.len" rule because it is invoked
> > > only from the CPU hotplug code.
> > > 
> > > So you could use the rcu_cblist .len field to update the ->cblist.len
> > > field, thus combining the _cbs and _count updates.  One thing that helps
> > > is that setting th e rcu_cblist .len field doesn't hurt the other use
> > > cases that require careful handling of ->cblist.len.
> > 
> > Thank you for the ideas. I am trying something like this on top of this
> > series based on the ideas. One thing I concerned a bit is if getting rid of
> > the rcu_segcblist_xchg_len() function (which has memory barriers in them)
> > causes issues in the hotplug path. I am now directly updating the length
> > without additional memory barriers. I will test it more and try to reason
> > more about it as well.
> 
> In this particular case, the CPU-hotplug locks prevent rcu_barrier()
> from running concurrently, so it should be OK.  Is there an easy way
> to make lockdep help us check this?  Does lockdep_assert_cpus_held()
> suffice, or is it too easily satisfied?

Just to clarify, the race you mention is rcu_barrier() should not run
concurrently with CPU hotplug operation.

The reason is:
rcu_barrier() checks whether the segmented list's length is 0 before sending
IPIs for entraining a callback onto the destination CPU. But if it
misunderstands the length's value due to a lack of memory barriers, it may
not missing sending an IPI causing the barrier to fail. Please correct me if
I'm wrong though.

Due to CPU hotplug locking, such race is impossible because both
rcu_barrier() and the CPU migrating the callbacks holds it.

I think the lockdep_assert_cpus_held() may suffice. I can add it to the
rcu_segcblist_merge() function.

BTW, I think we can simplify rcu_barrier() a bit and combine the tracepoint
and rid one outer if clause (diff at end of email).

> >  }
> >  
> >  /*
> > @@ -448,6 +419,7 @@ void rcu_segcblist_insert_done_cbs(struct rcu_segcblist *rsclp,
> >  			break;
> >  	rclp->head = NULL;
> >  	rclp->tail = &rclp->head;
> > +	rcu_segcblist_add_len(rsclp, rclp->len);
> 
> Does there need to be a compensating action in rcu_do_batch(), or is
> this the point of the "rcu_segcblist_add_len(rsclp, -(rclp->len));"
> added to rcu_segcblist_extract_done_cbs() above?
> 
> If so, a comment would be good.

I think external compensation by rcu_do_batch is not needed, its best to
handle all cblist.len modifications internally in the segcblist API
where possible.

> >  /*
> > @@ -463,6 +435,7 @@ void rcu_segcblist_insert_pend_cbs(struct rcu_segcblist *rsclp,
> >  	rcu_segcblist_add_seglen(rsclp, RCU_NEXT_TAIL, rclp->len);
> >  	WRITE_ONCE(*rsclp->tails[RCU_NEXT_TAIL], rclp->head);
> >  	WRITE_ONCE(rsclp->tails[RCU_NEXT_TAIL], rclp->tail);
> > +	rcu_segcblist_add_len(rsclp, rclp->len);
> >  }
> >  
> >  /*
> > @@ -601,16 +574,13 @@ void rcu_segcblist_merge(struct rcu_segcblist *dst_rsclp,
> >  {
> >  	struct rcu_cblist donecbs;
> >  	struct rcu_cblist pendcbs;
> > -	long src_len;
> >  
> >  	rcu_cblist_init(&donecbs);
> >  	rcu_cblist_init(&pendcbs);
> >  
> > -	src_len = rcu_segcblist_xchg_len(src_rsclp, 0);
> >  	rcu_segcblist_extract_done_cbs(src_rsclp, &donecbs);
> >  	rcu_segcblist_extract_pend_cbs(src_rsclp, &pendcbs);
> >  
> > -	rcu_segcblist_add_len(dst_rsclp, src_len);
> >  	rcu_segcblist_insert_done_cbs(dst_rsclp, &donecbs);
> >  	rcu_segcblist_insert_pend_cbs(dst_rsclp, &pendcbs);
> 
> Can we now pair the corresponding _extract_ and _insert_ calls, thus
> requiring only one rcu_cblist structure?  This would drop two more lines
> of code.  Or would that break something?

That could work as well, yes. But may not be worth adding another function to
combine extract/insert operation, since the extract and insert operations are
needed by rcutree and srcutree as well (other than hotplug).

thanks,

 - Joel

---8<-----------------------

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 16ad99a9ebba..274a1845ad38 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -3856,17 +3856,16 @@ void rcu_barrier(void)
 		if (cpu_is_offline(cpu) &&
 		    !rcu_segcblist_is_offloaded(&rdp->cblist))
 			continue;
-		if (rcu_segcblist_n_cbs(&rdp->cblist) && cpu_online(cpu)) {
-			rcu_barrier_trace(TPS("OnlineQ"), cpu,
-					  rcu_state.barrier_sequence);
-			smp_call_function_single(cpu, rcu_barrier_func, (void *)cpu, 1);
-		} else if (rcu_segcblist_n_cbs(&rdp->cblist) &&
-			   cpu_is_offline(cpu)) {
-			rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu,
-					  rcu_state.barrier_sequence);
-			local_irq_disable();
-			rcu_barrier_func((void *)cpu);
-			local_irq_enable();
+		if (rcu_segcblist_n_cbs(&rdp->cblist)) {
+			rcu_barrier_trace(cpu_online(cpu) ? TPS("OnlineQ") : TPS("OfflineNoCBQ"),
+					  cpu, rcu_state.barrier_sequence);
+			if (cpu_online(cpu)) {
+				smp_call_function_single(cpu, rcu_barrier_func, (void *)cpu, 1);
+			} else {
+				local_irq_disable();
+				rcu_barrier_func((void *)cpu);
+				local_irq_enable();
+			}
 		} else if (cpu_is_offline(cpu)) {
 			rcu_barrier_trace(TPS("OfflineNoCBNoQ"), cpu,
 					  rcu_state.barrier_sequence);

Powered by blists - more mailing lists