mm/slab_common: fix slab_caches list corruption after kmem_cache_dest…

…roy()

commit 46a9ea6 upstream.

After the commit in Fixes:, if a module that created a slab cache does not
release all of its allocated objects before destroying the cache (at rmmod
time), we might end up releasing the kmem_cache object without removing it
from the slab_caches list thus corrupting the list as kmem_cache_destroy()
ignores the return value from shutdown_cache(), which in turn never removes
the kmem_cache object from slabs_list in case __kmem_cache_shutdown() fails
to release all of the cache's slabs.

This is easily observable on a kernel built with CONFIG_DEBUG_LIST=y
as after that ill release the system will immediately trip on list_add,
or list_del, assertions similar to the one shown below as soon as another
kmem_cache gets created, or destroyed:

  [ 1041.213632] list_del corruption. next->prev should be ffff89f596fb5768, but was 52f1e5016aeee75d. (next=ffff89f595a1b268)
  [ 1041.219165] ------------[ cut here ]------------
  [ 1041.221517] kernel BUG at lib/list_debug.c:62!
  [ 1041.223452] invalid opcode: 0000 [openbmc#1] PREEMPT SMP PTI
  [ 1041.225408] CPU: 2 PID: 1852 Comm: rmmod Kdump: loaded Tainted: G    B   W  OE      6.5.0 openbmc#15
  [ 1041.228244] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS edk2-20230524-3.fc37 05/24/2023
  [ 1041.231212] RIP: 0010:__list_del_entry_valid+0xae/0xb0

Another quick way to trigger this issue, in a kernel with CONFIG_SLUB=y,
is to set slub_debug to poison the released objects and then just run
cat /proc/slabinfo after removing the module that leaks slab objects,
in which case the kernel will panic:

  [   50.954843] general protection fault, probably for non-canonical address 0xa56b6b6b6b6b6b8b: 0000 [openbmc#1] PREEMPT SMP PTI
  [   50.961545] CPU: 2 PID: 1495 Comm: cat Kdump: loaded Tainted: G    B   W  OE      6.5.0 openbmc#15
  [   50.966808] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS edk2-20230524-3.fc37 05/24/2023
  [   50.972663] RIP: 0010:get_slabinfo+0x42/0xf0

This patch fixes this issue by properly checking shutdown_cache()'s
return value before taking the kmem_cache_release() branch.

Fixes: 0495e33 ("mm/slab_common: Deleting kobject in kmem_cache_destroy() without holding slab_mutex/cpu_hotplug_lock")
Signed-off-by: Rafael Aquini <aquini@redhat.com>
Cc: stable@vger.kernel.org
Reviewed-by: Waiman Long <longman@redhat.com>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

mm/slab_common.c

@@ -474,7 +474,7 @@ void slab_kmem_cache_release(struct kmem_cache *s)
 
 void kmem_cache_destroy(struct kmem_cache *s)
 {
-	int refcnt;
+	int err = -EBUSY;
 	bool rcu_set;
 
 	if (unlikely(!s) || !kasan_check_byte(s))
@@ -485,17 +485,17 @@ void kmem_cache_destroy(struct kmem_cache *s)
 
 	rcu_set = s->flags & SLAB_TYPESAFE_BY_RCU;
 
-	refcnt = --s->refcount;
-	if (refcnt)
+	s->refcount--;
+	if (s->refcount)
 		goto out_unlock;
 
-	WARN(shutdown_cache(s),
-	     "%s %s: Slab cache still has objects when called from %pS",
+	err = shutdown_cache(s);
+	WARN(err, "%s %s: Slab cache still has objects when called from %pS",
 	     __func__, s->name, (void *)_RET_IP_);
 out_unlock:
 	mutex_unlock(&slab_mutex);
 	cpus_read_unlock();
-	if (!refcnt && !rcu_set)
+	if (!err && !rcu_set)
 		kmem_cache_release(s);
 }
 EXPORT_SYMBOL(kmem_cache_destroy);