Re: [patch v3] mm: lru_cache_disable: replace work queue synchronization with synchronize_rcu

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, Mar 03, 2022 at 05:49:30PM -0800, Paul E. McKenney wrote:
> On Thu, Mar 03, 2022 at 05:03:23PM -0800, Andrew Morton wrote:
> > (Question for paulmck below, please)
> > 
> > On Tue, 22 Feb 2022 13:07:35 -0300 Marcelo Tosatti <mtosatti@redhat.com> wrote:
> > 
> > > 
> > > On systems that run FIFO:1 applications that busy loop 
> > > on isolated CPUs, executing tasks on such CPUs under
> > > lower priority is undesired (since that will either
> > > hang the system, or cause longer interruption to the
> > > FIFO task due to execution of lower priority task 
> > > with very small sched slices).
> > > 
> > > Commit d479960e44f27e0e52ba31b21740b703c538027c ("mm: disable LRU 
> > > pagevec during the migration temporarily") relies on 
> > > queueing work items on all online CPUs to ensure visibility
> > > of lru_disable_count.
> > > 
> > > However, its possible to use synchronize_rcu which will provide the same
> > > guarantees (see comment this patch modifies on lru_cache_disable).
> > > 
> > > Fixes:
> > > 
> > > [ 1873.243925] INFO: task kworker/u160:0:9 blocked for more than 622 seconds.
> > > [ 1873.243927]       Tainted: G          I      --------- ---  5.14.0-31.rt21.31.el9.x86_64 #1
> > > [ 1873.243929] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> > > [ 1873.243929] task:kworker/u160:0  state:D stack:    0 pid:    9 ppid:     2 flags:0x00004000
> > > [ 1873.243932] Workqueue: cpuset_migrate_mm cpuset_migrate_mm_workfn
> > > [ 1873.243936] Call Trace:
> > > [ 1873.243938]  __schedule+0x21b/0x5b0
> > > [ 1873.243941]  schedule+0x43/0xe0
> > > [ 1873.243943]  schedule_timeout+0x14d/0x190
> > > [ 1873.243946]  ? resched_curr+0x20/0xe0
> > > [ 1873.243953]  ? __prepare_to_swait+0x4b/0x70
> > > [ 1873.243958]  wait_for_completion+0x84/0xe0
> > > [ 1873.243962]  __flush_work.isra.0+0x146/0x200
> > > [ 1873.243966]  ? flush_workqueue_prep_pwqs+0x130/0x130
> > > [ 1873.243971]  __lru_add_drain_all+0x158/0x1f0
> > > [ 1873.243978]  do_migrate_pages+0x3d/0x2d0
> > > [ 1873.243985]  ? pick_next_task_fair+0x39/0x3b0
> > > [ 1873.243989]  ? put_prev_task_fair+0x1e/0x30
> > > [ 1873.243992]  ? pick_next_task+0xb30/0xbd0
> > > [ 1873.243995]  ? __tick_nohz_task_switch+0x1e/0x70
> > > [ 1873.244000]  ? raw_spin_rq_unlock+0x18/0x60
> > > [ 1873.244002]  ? finish_task_switch.isra.0+0xc1/0x2d0
> > > [ 1873.244005]  ? __switch_to+0x12f/0x510
> > > [ 1873.244013]  cpuset_migrate_mm_workfn+0x22/0x40
> > > [ 1873.244016]  process_one_work+0x1e0/0x410
> > > [ 1873.244019]  worker_thread+0x50/0x3b0
> > > [ 1873.244022]  ? process_one_work+0x410/0x410
> > > [ 1873.244024]  kthread+0x173/0x190
> > > [ 1873.244027]  ? set_kthread_struct+0x40/0x40
> > > [ 1873.244031]  ret_from_fork+0x1f/0x30
> > > 
> > > ...
> > >
> > > --- a/mm/swap.c
> > > +++ b/mm/swap.c
> > > @@ -831,8 +831,7 @@ inline void __lru_add_drain_all(bool force_all_cpus)
> > >  	for_each_online_cpu(cpu) {
> > >  		struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
> > >  
> > > -		if (force_all_cpus ||
> > > -		    pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
> > > +		if (pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
> > >  		    data_race(pagevec_count(&per_cpu(lru_rotate.pvec, cpu))) ||
> > >  		    pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
> > >  		    pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
> > 
> > This change appears to be "don't queue work on CPUs which don't have
> > any work to do".  Correct?  This isn't changelogged?

Its replaced by synchronize_rcu, and its mentioned in the changelog:

"However, its possible to use synchronize_rcu which will provide the same
 guarantees (see comment this patch modifies on lru_cache_disable)."

Will resend -v4 with a more verbose changelog.


> > > @@ -876,14 +875,19 @@ atomic_t lru_disable_count = ATOMIC_INIT(0);
> > >  void lru_cache_disable(void)
> > >  {
> > >  	atomic_inc(&lru_disable_count);
> > > +	synchronize_rcu();
> > >  #ifdef CONFIG_SMP
> > >  	/*
> > > -	 * lru_add_drain_all in the force mode will schedule draining on
> > > -	 * all online CPUs so any calls of lru_cache_disabled wrapped by
> > > -	 * local_lock or preemption disabled would be ordered by that.
> > > -	 * The atomic operation doesn't need to have stronger ordering
> > > -	 * requirements because that is enforced by the scheduling
> > > -	 * guarantees.
> > > +	 * synchronize_rcu() waits for preemption disabled
> > > +	 * and RCU read side critical sections.
> > > +	 * For the users of lru_disable_count:
> > > +	 *
> > > +	 * preempt_disable, local_irq_disable  [bh_lru_lock()]
> > > +	 * rcu_read_lock		       [rt_spin_lock CONFIG_PREEMPT_RT]
> > > +	 * preempt_disable		       [local_lock !CONFIG_PREEMPT_RT]
> > > +	 *
> > > +	 * so any calls of lru_cache_disabled wrapped by local_lock or
> > > +	 * preemption disabled would be ordered by that.
> > >  	 */
> > >  	__lru_add_drain_all(true);
> > >  #else
> > 
> > Does this also work with CONFIG_TINY_RCU?
> > 
> > This seems abusive of synchronize_rcu().  None of this code uses RCU,
> > but it so happens that synchronize_rcu() happily provides the desired
> > effects.  Changes in RCU's happy side-effects might break this. 
> > Perhaps a formal API function which does whatever-you-want-it-to-do
> > would be better.
> 
> I don't claim to understand the full lru_cache_disable() use case, but
> since v5.1 synchronize_rcu() is guaranteed to wait on preempt_disable()
> regions of code.  In contrast, back in the old days, you had to use
> synchronize_sched() to wait on preempt_disable() regions, even in
> CONFIG_PREEMPT=y kernels.  So if the comment is accurate, it is OK.

OK, will add an additional comment regarding v5.1.

> Just be careful what you backport past v5.1...
> 
> > And...  I really don't understand the fix.  What is it about
> > synchronize_rcu() which guarantees that a work function which is queued
> > on CPU N will now get executed even if CPU N is spinning in SCHED_FIFO
> > userspace?
> 
> I don't understand this part, either.


All CPUs should see lru_disable_count (and therefore not add pages
to per-CPU LRU pvecs, otherwise the page migration bug fixed
by d479960e44f27e0e52ba31b21740b703c538027c can occur.

To do this, the commit above ("mm: disable LRU 
pagevec during the migration temporarily") relies on 
queueing work items on all online CPUs to ensure visibility
of lru_disable_count:

 */
+void lru_cache_disable(void)
+{
+       atomic_inc(&lru_disable_count);
+#ifdef CONFIG_SMP
+       /*
+        * lru_add_drain_all in the force mode will schedule draining on
+        * all online CPUs so any calls of lru_cache_disabled wrapped by
+        * local_lock or preemption disabled would be ordered by that.
+        * The atomic operation doesn't need to have stronger ordering
+        * requirements because that is enforeced by the scheduling
+        * guarantees.
+        */
+       __lru_add_drain_all(true);
+#else


CPU-0					CPU-1

					
					local_lock(&lru_pvecs.lock);
                			pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
					add page to per-CPU LRU pvec
					if atomic_read(lru_disable_count) != 0
						flush per-CPU LRU pvec
atomic_inc(lru_disable_count)

					local_unlock(&lru_pvec.lock)
lru_add_drain_all(force_all_cpus=true)

However queueing the work items disturbs isolated CPUs. To avoid it, its
possible to use synchronize_rcu instead:

CPU-0                                   CPU-1


                                        local_lock(&lru_pvecs.lock);
                                        pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
                                        add page to per-CPU LRU pvec
                                        if atomic_read(lru_disable_count) != 0
                                                flush per-CPU LRU pvec
atomic_inc(lru_disable_count)

                                        local_unlock(&lru_pvec.lock)
synchronize_rcu()

Which will wait for all preemption (or IRQ disabled) sections to
complete, therefore ensuring visibilily of lru_disable_count.