Re: [RFC PATCH] rcu: back off on allocation failure in fill_page_cache_func

From: Paul E. McKenney
Date: Fri Jun 24 2022 - 09:02:58 EST


On Fri, Jun 24, 2022 at 09:17:47AM +0200, Michal Hocko wrote:
> On Thu 23-06-22 20:39:50, Paul E. McKenney wrote:
> > On Wed, Jun 22, 2022 at 08:54:22PM +0200, Uladzislau Rezki wrote:
> > > On Wed, Jun 22, 2022 at 01:47:11PM +0200, Michal Hocko wrote:
> > > > From: Michal Hocko <mhocko@xxxxxxxx>
> > > >
> > > > fill_page_cache_func allocates couple of pages to store
> > > > kvfree_rcu_bulk_data. This is a lightweight (GFP_NORETRY) allocation
> > > > which can fail under memory pressure. The function will, however keep
> > > > retrying even when the previous attempt has failed.
> > > >
> > > > While this is not really incorrect there is one thing to consider. This
> > > > allocation is invoked from the WQ context and that means that if the
> > > > memory reclaim gets stuck it can hog the worker for quite some time.
> > > > WQ concurrency is only triggered when the worker context sleeps and that
> > > > is not guaranteed for __GFP_NORETRY allocation attempts (see
> > > > should_reclaim_retry).
> > > >
> > > > We have seen WQ lockups
> > > > kernel: BUG: workqueue lockup - pool cpus=93 node=1 flags=0x1 nice=0 stuck for 32s!
> > > > [...]
> > > > kernel: pool 74: cpus=37 node=0 flags=0x1 nice=0 hung=32s workers=2 manager: 2146
> > > > kernel: pwq 498: cpus=249 node=1 flags=0x1 nice=0 active=4/256 refcnt=5
> > > > kernel: in-flight: 1917:fill_page_cache_func
> > > > kernel: pending: dbs_work_handler, free_work, kfree_rcu_monitor
> > > >
> > > > Originaly, we thought that several retries with direct reclaim being
> > > > stuck is the underlying reason but we couldn't have confirmed that and
> > > > have seen a similar lockups detected even without any heavy memory
> > > > pressure so there is likely something else/more going on. On the other
> > > > hand failing the allocation shouldn't have a big impact and from the
> > > > code it is not really obvious why retrying is desirable so back off
> > > > after the allocation failure.
> > > >
> > > > Cc: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx>
> > > > Cc: "Paul E. McKenney" <paulmck@xxxxxxxxxx>
> > > > Cc: Frederic Weisbecker <frederic@xxxxxxxxxx>
> > > > Cc: Neeraj Upadhyay <quic_neeraju@xxxxxxxxxxx>
> > > > Cc: Josh Triplett <josh@xxxxxxxxxxxxxxxx>
> > > > Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
> > > > Cc: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
> > > > Cc: Lai Jiangshan <jiangshanlai@xxxxxxxxx>
> > > > Cc: Joel Fernandes <joel@xxxxxxxxxxxxxxxxx>
> > > > Signed-off-by: Michal Hocko <mhocko@xxxxxxxx>
> > > > ---
> > > >
> > > > Hi,
> > > > I am sending this as an RFC because I couldn't prove that the WQ
> > > > concurency issue as a result from the allocation retry is really a
> > > > problem. On the other hand I couldn't see a good reason to retry after a
> > > > previous failure. While the kswapd running in the background could have
> > > > released some memory this is a not really guaranteed and mostly a
> > > > wishful thinking.
> > > >
> > > > I do not understand the code well enough so I could be easily missing
> > > > something. If the patch is a wrong thing to do then it would be really
> > > > nice to add a comment why the retry is desirable and a good thing to do.
> > > >
> > > > The retry loop should be bound to rcu_min_cached_objs which is quite
> > > > small but configurable so this can get large in some setups.
> > > >
> > > > Thanks
> > > >
> > > > kernel/rcu/tree.c | 17 +++++++++--------
> > > > 1 file changed, 9 insertions(+), 8 deletions(-)
> > > >
> > > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > > index c25ba442044a..54a3a19c4c0b 100644
> > > > --- a/kernel/rcu/tree.c
> > > > +++ b/kernel/rcu/tree.c
> > > > @@ -3508,15 +3508,16 @@ static void fill_page_cache_func(struct work_struct *work)
> > > > bnode = (struct kvfree_rcu_bulk_data *)
> > > > __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
> > > >
> > > > - if (bnode) {
> > > > - raw_spin_lock_irqsave(&krcp->lock, flags);
> > > > - pushed = put_cached_bnode(krcp, bnode);
> > > > - raw_spin_unlock_irqrestore(&krcp->lock, flags);
> > > > + if (!bnode)
> > > > + break;
> > > >
> > > > - if (!pushed) {
> > > > - free_page((unsigned long) bnode);
> > > > - break;
> > > > - }
> > > > + raw_spin_lock_irqsave(&krcp->lock, flags);
> > > > + pushed = put_cached_bnode(krcp, bnode);
> > > > + raw_spin_unlock_irqrestore(&krcp->lock, flags);
> > > > +
> > > > + if (!pushed) {
> > > > + free_page((unsigned long) bnode);
> > > > + break;
> > > > }
> > > > }
> > > >
> > > > --
> > > > 2.30.2
> > > >
> > > OK. You would like to break the loop once an allocation does not succeed.
> > > To me it also makes sense, i mean there is no reason to repeat it several
> > > times that can lead to worqueue hogging.
> > >
> > > Reviewed-by: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx>
>
> Thanks for the review!
>
> > Queued for testing and further review, thank you both!
> >
> > My normal process would send this not into the upcoming merge window
> > (v5.20?) but instead into the next one (v6.0?). Please let me know if
> > this is more urgent then that, and I will see what I can do.
>
> No, it is not really urgent. As I've said, it turned out that there are
> other contributors to the WQ stalls so this is mostly a cosmetic change.

The merge window after this coming one, then!

Ah, and I forgot to send the wordsmithed patch, please see below. Just
in case I messed something up.

Thanx, Paul

------------------------------------------------------------------------

commit 29dc4a37468641e2a9fca7f8c41c4d30f3c659a9
Author: Michal Hocko <mhocko@xxxxxxxx>
Date: Wed Jun 22 13:47:11 2022 +0200

rcu: Back off upon fill_page_cache_func() allocation failure

The fill_page_cache_func() function allocates couple of pages to store
kvfree_rcu_bulk_data structures. This is a lightweight (GFP_NORETRY)
allocation which can fail under memory pressure. The function will,
however keep retrying even when the previous attempt has failed.

This retrying is in theory correct, but in practice the allocation is
invoked from workqueue context, which means that if the memory reclaim
gets stuck, these retries can hog the worker for quite some time.
Although the workqueues subsystem automatically adjusts concurrency, such
adjustment is not guaranteed to happen until the worker context sleeps.
And the fill_page_cache_func() function's retry loop is not guaranteed
to sleep (see the should_reclaim_retry() function).

And we have seen this function cause workqueue lockups:

kernel: BUG: workqueue lockup - pool cpus=93 node=1 flags=0x1 nice=0 stuck for 32s!
[...]
kernel: pool 74: cpus=37 node=0 flags=0x1 nice=0 hung=32s workers=2 manager: 2146
kernel: pwq 498: cpus=249 node=1 flags=0x1 nice=0 active=4/256 refcnt=5
kernel: in-flight: 1917:fill_page_cache_func
kernel: pending: dbs_work_handler, free_work, kfree_rcu_monitor

Originally, we thought that the root cause of this lockup was several
retries with direct reclaim, but this is not yet confirmed. Furthermore,
we have seen similar lockups without any heavy memory pressure. This
suggests that there are other factors contributing to these lockups.
However, it is not really clear that endless retries are desireable.

So let's make the fill_page_cache_func() function back off after
allocation failure.

Cc: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx>
Cc: "Paul E. McKenney" <paulmck@xxxxxxxxxx>
Cc: Frederic Weisbecker <frederic@xxxxxxxxxx>
Cc: Neeraj Upadhyay <quic_neeraju@xxxxxxxxxxx>
Cc: Josh Triplett <josh@xxxxxxxxxxxxxxxx>
Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
Cc: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
Cc: Lai Jiangshan <jiangshanlai@xxxxxxxxx>
Cc: Joel Fernandes <joel@xxxxxxxxxxxxxxxxx>
Signed-off-by: Michal Hocko <mhocko@xxxxxxxx>
Reviewed-by: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx>
Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxx>

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 8240c4d631eef..89cb173a1a952 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -3183,15 +3183,16 @@ static void fill_page_cache_func(struct work_struct *work)
bnode = (struct kvfree_rcu_bulk_data *)
__get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);

- if (bnode) {
- raw_spin_lock_irqsave(&krcp->lock, flags);
- pushed = put_cached_bnode(krcp, bnode);
- raw_spin_unlock_irqrestore(&krcp->lock, flags);
+ if (!bnode)
+ break;

- if (!pushed) {
- free_page((unsigned long) bnode);
- break;
- }
+ raw_spin_lock_irqsave(&krcp->lock, flags);
+ pushed = put_cached_bnode(krcp, bnode);
+ raw_spin_unlock_irqrestore(&krcp->lock, flags);
+
+ if (!pushed) {
+ free_page((unsigned long) bnode);
+ break;
}
}