[PATCH v2] mm/slab: don't use kfree_rcu sheaves on PREEMPT_RT in kvfree_call_rcu()

[ThangNN99 <ngocthang2710.1999@gmail.com>]

syzbot reports a possible circular locking dependency between
&p->pi_lock and the per-CPU kfree_rcu sheaf lock (_T->lock) on
PREEMPT_RT:

  __balance_push_cpu_stop()               [holds p->pi_lock, raw]
    select_fallback_rq()
      cpuset_cpus_allowed_fallback()
        set_cpus_allowed_force()
          kfree_rcu(ac.user_mask)
            kvfree_call_rcu()
              kfree_rcu_sheaf()
                __kfree_rcu_sheaf()
                  local_trylock(&s->cpu_sheaves->lock)   <- _T->lock

set_cpus_allowed_force() uses kfree_rcu() instead of kfree() here
specifically because it can be called with p->pi_lock (a raw
spinlock) held, and plain kfree() may sleep under PREEMPT_RT.

Commit 2a8bb29ec9b2 ("mm/slab: allow kfree_rcu_sheaf() on
PREEMPT_RT") made kvfree_call_rcu() try the sheaves fast path on
PREEMPT_RT too, since __kfree_rcu_sheaf() only trylocks there and
so cannot itself block. True, but the sheaf/barn locks it trylocks
are also taken as regular, blocking locks elsewhere, so lockdep
still records a lock-class ordering cycle against any raw spinlock
already held by the caller, which is what syzbot caught.

The plain kfree_rcu()/kvfree_rcu() API gives kvfree_call_rcu() no
way to know the caller is in such a context, so keep it
conservative on PREEMPT_RT and skip the sheaves layer there,
falling back to the existing raw_spinlock_t-protected krcp list,
which is always safe to nest under another raw spinlock. This
restores the pre-2a8bb29ec9b2 behavior of kvfree_call_rcu().

kfree_call_rcu_nolock(), added later in commit 3bc999d944b3
("mm/slab: introduce kfree_rcu_nolock()"), is untouched by this
patch. Note it would not be a safe substitute here either: it still
reaches __kfree_rcu_sheaf()'s local_trylock() on &s->cpu_sheaves->lock
unconditionally, so a caller already holding a raw spinlock would hit
the same lockdep ordering cycle through that path too.

Reported-by: syzbot+acf142088e0182172e58@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=acf142088e0182172e58
Fixes: 2a8bb29ec9b2 ("mm/slab: allow kfree_rcu_sheaf() on PREEMPT_RT")
Signed-off-by: ThangNN99 <ngocthang2710.1999@gmail.com>
---
v2 (per automated review on v1):
 - Corrected commit message / comments: kfree_call_rcu_nolock() is not
   a safe alternative here either, since it still trylocks the same
   &s->cpu_sheaves->lock unconditionally.
 - Removed the now-unreachable CONFIG_PREEMPT_RT branch inside
   kfree_rcu_sheaf() left over by this fix (it can no longer run,
   since kvfree_call_rcu() already skips calling it on PREEMPT_RT).

 mm/slab_common.c | 20 ++++++++++----------
 mm/slub.c        |  6 +++---
 2 files changed, 13 insertions(+), 13 deletions(-)

diff --git a/mm/slab_common.c b/mm/slab_common.c
index b19ba1b31484..3de1eabe6c77 100644
--- a/mm/slab_common.c
+++ b/mm/slab_common.c
@@ -1667,15 +1667,8 @@ static bool kfree_rcu_sheaf(void *obj)
 {
 	struct kmem_cache *s;
 	struct slab *slab;
-	unsigned int free_flags = SLAB_FREE_DEFAULT;
-
-	/*
-	 * It is not safe to spin on PREEMPT_RT because the kernel might be
-	 * holding a raw spinlock and slab acquires sleeping locks.
-	 */
-	if (IS_ENABLED(CONFIG_PREEMPT_RT))
-		free_flags = SLAB_FREE_NOLOCK;
 
+	/* Callers on PREEMPT_RT never reach here, see kvfree_call_rcu(). */
 	if (is_vmalloc_addr(obj))
 		return false;
 
@@ -1685,7 +1678,7 @@ static bool kfree_rcu_sheaf(void *obj)
 
 	s = slab->slab_cache;
 	if (likely(!IS_ENABLED(CONFIG_NUMA) || slab_nid(slab) == numa_mem_id()))
-		return __kfree_rcu_sheaf(s, obj, free_flags);
+		return __kfree_rcu_sheaf(s, obj, SLAB_FREE_DEFAULT);
 
 	return false;
 }
@@ -2034,7 +2027,14 @@ void kvfree_call_rcu(struct kvfree_rcu_head *head, void *ptr)
 	if (!head)
 		might_sleep();
 
-	if (kfree_rcu_sheaf(ptr))
+	/*
+	 * Callers may hold a raw spinlock here on PREEMPT_RT (e.g.
+	 * set_cpus_allowed_force() with p->pi_lock held), and the sheaf/barn
+	 * locks are also taken as blocking locks elsewhere, so trying them
+	 * here creates a lockdep-visible ordering conflict. Skip sheaves on
+	 * PREEMPT_RT.
+	 */
+	if (!IS_ENABLED(CONFIG_PREEMPT_RT) && kfree_rcu_sheaf(ptr))
 		return;
 
 	// Queue the object but don't yet schedule the batch.
diff --git a/mm/slub.c b/mm/slub.c
index f9b56cb439e7..1e8bad7a018e 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -6088,10 +6088,10 @@ static void rcu_free_sheaf(struct rcu_head *head)
 /*
  * kvfree_call_rcu() can be called while holding a raw_spinlock_t. Since
  * __kfree_rcu_sheaf() may acquire a spinlock_t (sleeping lock on PREEMPT_RT),
- * this would violate lock nesting rules. Therefore, kvfree_call_rcu() avoids
- * this problem by passing SLAB_FREE_NOLOCK on PREEMPT_RT.
+ * this would violate lock nesting rules. kvfree_call_rcu() avoids this by
+ * bypassing the sheaves layer on PREEMPT_RT.
  *
- * However, lockdep still complains that it is invalid to acquire spinlock_t
+ * lockdep still complains that it is invalid to acquire spinlock_t
  * while holding raw_spinlock_t, even on !PREEMPT_RT where spinlock_t is a
  * spinning lock. Tell lockdep that acquiring spinlock_t is valid here
  * by temporarily raising the wait-type to LD_WAIT_CONFIG. Skip the lockdep map
-- 
2.43.0