[PATCH driver-core-linus] kernfs: Fix spurious lockdep warning in kernfs_find_and_get_node_by_id()

From: Tejun Heo
Date: Mon Oct 10 2022 - 19:54:24 EST

c25491747b21 ("kernfs: Add KERNFS_REMOVING flags") made
kernfs_find_and_get_node_by_id() test kernfs_active() instead of
KERNFS_ACTIVATED. kernfs_find_and_get_by_id() is called without holding the
kernfs_rwsem triggering the following lockdep warning.

WARNING: CPU: 1 PID: 6191 at fs/kernfs/dir.c:36 kernfs_active+0xe8/0x120 fs/kernfs/dir.c:38
Modules linked in:
CPU: 1 PID: 6191 Comm: syz-executor.1 Not tainted 6.0.0-syzkaller-09413-g4899a36f91a9 #0
Hardware name: linux,dummy-virt (DT)
pstate: 10000005 (nzcV daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
pc : kernfs_active+0xe8/0x120 fs/kernfs/dir.c:36
lr : lock_is_held include/linux/lockdep.h:283 [inline]
lr : kernfs_active+0x94/0x120 fs/kernfs/dir.c:36
sp : ffff8000182c7a00
x29: ffff8000182c7a00 x28: 0000000000000002 x27: 0000000000000001
x26: ffff00000ee1f6a8 x25: 1fffe00001dc3ed5 x24: 0000000000000000
x23: ffff80000ca1fba0 x22: ffff8000089efcb0 x21: 0000000000000001
x20: ffff0000091181d0 x19: ffff0000091181d0 x18: ffff00006a9e6b88
x17: 0000000000000000 x16: 0000000000000000 x15: ffff00006a9e6bc4
x14: 1ffff00003058f0e x13: 1fffe0000258c816 x12: ffff700003058f39
x11: 1ffff00003058f38 x10: ffff700003058f38 x9 : dfff800000000000
x8 : ffff80000e482f20 x7 : ffff0000091d8058 x6 : ffff80000e482c60
x5 : ffff000009402ee8 x4 : 1ffff00001bd1f46 x3 : 1fffe0000258c6d1
x2 : 0000000000000003 x1 : 00000000000000c0 x0 : 0000000000000000
Call trace:
kernfs_active+0xe8/0x120 fs/kernfs/dir.c:38
kernfs_find_and_get_node_by_id+0x6c/0x140 fs/kernfs/dir.c:708
__kernfs_fh_to_dentry fs/kernfs/mount.c:102 [inline]
kernfs_fh_to_dentry+0x88/0x1fc fs/kernfs/mount.c:128
exportfs_decode_fh_raw+0x104/0x560 fs/exportfs/expfs.c:435
exportfs_decode_fh+0x10/0x5c fs/exportfs/expfs.c:575
do_handle_to_path fs/fhandle.c:152 [inline]
handle_to_path fs/fhandle.c:207 [inline]
do_handle_open+0x2a4/0x7b0 fs/fhandle.c:223
__do_compat_sys_open_by_handle_at fs/fhandle.c:277 [inline]
__se_compat_sys_open_by_handle_at fs/fhandle.c:274 [inline]
__arm64_compat_sys_open_by_handle_at+0x6c/0x9c fs/fhandle.c:274
__invoke_syscall arch/arm64/kernel/syscall.c:38 [inline]
invoke_syscall+0x6c/0x260 arch/arm64/kernel/syscall.c:52
el0_svc_common.constprop.0+0xc4/0x254 arch/arm64/kernel/syscall.c:142
do_el0_svc_compat+0x40/0x70 arch/arm64/kernel/syscall.c:212
el0_svc_compat+0x54/0x140 arch/arm64/kernel/entry-common.c:772
el0t_32_sync_handler+0x90/0x140 arch/arm64/kernel/entry-common.c:782
el0t_32_sync+0x190/0x194 arch/arm64/kernel/entry.S:586
irq event stamp: 232
hardirqs last enabled at (231): [<ffff8000081edf70>] raw_spin_rq_unlock_irq kernel/sched/sched.h:1367 [inline]
hardirqs last enabled at (231): [<ffff8000081edf70>] finish_lock_switch kernel/sched/core.c:4943 [inline]
hardirqs last enabled at (231): [<ffff8000081edf70>] finish_task_switch.isra.0+0x200/0x880 kernel/sched/core.c:5061
hardirqs last disabled at (232): [<ffff80000c888bb4>] el1_dbg+0x24/0x80 arch/arm64/kernel/entry-common.c:404
softirqs last enabled at (228): [<ffff800008010938>] _stext+0x938/0xf58
softirqs last disabled at (207): [<ffff800008019380>] ____do_softirq+0x10/0x20 arch/arm64/kernel/irq.c:79
---[ end trace 0000000000000000 ]---

The lockdep warning in kernfs_active() is there to ensure that the activated
state stays stable for the caller. For kernfs_find_and_get_node_by_id(), all
that's needed is ensuring that a node which has never been activated can't
be looked up and guaranteeing lookup success when the caller knows the node
to be active, both of which can be achieved by testing the active count
without holding the kernfs_rwsem.

Fix the spurious warning by introducing __kernfs_active() which doesn't have
the lockdep annotation.

Signed-off-by: Tejun Heo <tj@xxxxxxxxxx>
Reported-by: syzbot+590ce62b128e79cf0a35@xxxxxxxxxxxxxxxxxxxxxxxxx
Fixes: c25491747b21 ("kernfs: Add KERNFS_REMOVING flags")
Cc: Amir Goldstein <amir73il@xxxxxxxxx>
Cc: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
fs/kernfs/dir.c | 14 ++++++++++++--
1 file changed, 12 insertions(+), 2 deletions(-)

diff --git a/fs/kernfs/dir.c b/fs/kernfs/dir.c
index 3990f3e270cb6..f33b3baad07cb 100644
--- a/fs/kernfs/dir.c
+++ b/fs/kernfs/dir.c
@@ -31,10 +31,15 @@ static DEFINE_SPINLOCK(kernfs_idr_lock); /* root->ino_idr */

#define rb_to_kn(X) rb_entry((X), struct kernfs_node, rb)

+static bool __kernfs_active(struct kernfs_node *kn)
+ return atomic_read(&kn->active) >= 0;
static bool kernfs_active(struct kernfs_node *kn)
- return atomic_read(&kn->active) >= 0;
+ return __kernfs_active(kn);

static bool kernfs_lockdep(struct kernfs_node *kn)
@@ -705,7 +710,12 @@ struct kernfs_node *kernfs_find_and_get_node_by_id(struct kernfs_root *root,
goto err_unlock;

- if (unlikely(!kernfs_active(kn) || !atomic_inc_not_zero(&kn->count)))
+ /*
+ * We should fail if @kn has never been activated and guarantee success
+ * if the caller knows that @kn is active. Both can be achieved by
+ * __kernfs_active() which tests @kn->active without kernfs_rwsem.
+ */
+ if (unlikely(!__kernfs_active(kn) || !atomic_inc_not_zero(&kn->count)))
goto err_unlock;