releases/6.8.6/rcu-nocb-fix-warn_on_once-in-the-rcu_nocb_bypass_loc.patch

   1 From 9aa0bc914bab8715b7b1dc81a5ec49f81c9cb736 Mon Sep 17 00:00:00 2001
   2 From: Sasha Levin <sashal@kernel.org>
   3 Date: Wed, 10 Jan 2024 16:11:28 +0800
   4 Subject: rcu/nocb: Fix WARN_ON_ONCE() in the rcu_nocb_bypass_lock()
   5
   6 From: Zqiang <qiang.zhang1211@gmail.com>
   7
   8 [ Upstream commit dda98810b552fc6bf650f4270edeebdc2f28bd3f ]
   9
  10 For the kernels built with CONFIG_RCU_NOCB_CPU_DEFAULT_ALL=y and
  11 CONFIG_RCU_LAZY=y, the following scenarios will trigger WARN_ON_ONCE()
  12 in the rcu_nocb_bypass_lock() and rcu_nocb_wait_contended() functions:
  13
  14         CPU2                                               CPU11
  15 kthread
  16 rcu_nocb_cb_kthread                                       ksys_write
  17 rcu_do_batch                                              vfs_write
  18 rcu_torture_timer_cb                                      proc_sys_write
  19 __kmem_cache_free                                         proc_sys_call_handler
  20 kmemleak_free                                             drop_caches_sysctl_handler
  21 delete_object_full                                        drop_slab
  22 __delete_object                                           shrink_slab
  23 put_object                                                lazy_rcu_shrink_scan
  24 call_rcu                                                  rcu_nocb_flush_bypass
  25 __call_rcu_commn                                            rcu_nocb_bypass_lock
  26                                                             raw_spin_trylock(&rdp->nocb_bypass_lock) fail
  27                                                             atomic_inc(&rdp->nocb_lock_contended);
  28 rcu_nocb_wait_contended                                     WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
  29  WARN_ON_ONCE(atomic_read(&rdp->nocb_lock_contended))                                          |
  30                             |_ _ _ _ _ _ _ _ _ _same rdp and rdp->cpu != 11_ _ _ _ _ _ _ _ _ __|
  31
  32 Reproduce this bug with "echo 3 > /proc/sys/vm/drop_caches".
  33
  34 This commit therefore uses rcu_nocb_try_flush_bypass() instead of
  35 rcu_nocb_flush_bypass() in lazy_rcu_shrink_scan().  If the nocb_bypass
  36 queue is being flushed, then rcu_nocb_try_flush_bypass will return
  37 directly.
  38
  39 Signed-off-by: Zqiang <qiang.zhang1211@gmail.com>
  40 Reviewed-by: Joel Fernandes (Google) <joel@joelfernandes.org>
  41 Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
  42 Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
  43 Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
  44 Signed-off-by: Sasha Levin <sashal@kernel.org>
  45 ---
  46  kernel/rcu/tree_nocb.h | 2 +-
  47  1 file changed, 1 insertion(+), 1 deletion(-)
  48
  49 diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
  50 index 4efbf7333d4e1..d430b4656f59e 100644
  51 --- a/kernel/rcu/tree_nocb.h
  52 +++ b/kernel/rcu/tree_nocb.h
  53 @@ -1383,7 +1383,7 @@ lazy_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
  54                         rcu_nocb_unlock_irqrestore(rdp, flags);
  55                         continue;
  56                 }
  57 -               WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, false));
  58 +               rcu_nocb_try_flush_bypass(rdp, jiffies);
  59                 rcu_nocb_unlock_irqrestore(rdp, flags);
  60                 wake_nocb_gp(rdp, false);
  61                 sc->nr_to_scan -= _count;
  62 --
  63 2.43.0
  64