From: Alexander Sverdlin <[email protected]>
Drop the mutex earlier in the loop so that LOCKDEP is not being provoked:
WARNING: possible circular locking dependency detected
.../9201 is trying to acquire lock:
ffffff85656d4e78 (kn->count#159){++++}, at: kernfs_remove_by_name_ns+0x5c/0xac
but task is already holding lock:
ffffff857b329e80 (&adap->userspace_clients_lock){+.+.}, at: i2c_sysfs_delete_device+0x8c/0x1f0
which lock already depends on the new lock.
the existing dependency chain (in reverse order) is:
-> #1 (&adap->userspace_clients_lock){+.+.}:
__mutex_lock+0xa4/0x924
mutex_lock_nested+0x48/0x5c
i2c_sysfs_new_device+0x108/0x220
dev_attr_store+0x48/0x60
sysfs_kf_write+0x54/0x80
kernfs_fop_write+0x120/0x240
__vfs_write+0x4c/0x90
vfs_write+0xe8/0x1d0
ksys_write+0x7c/0x104
__arm64_sys_write+0x28/0x3c
el0_svc_handler+0x90/0x200
el0_svc+0x8/0x16c
-> #0 (kn->count#159){++++}:
__lock_acquire+0xe8c/0x1a90
lock_acquire+0xd8/0x270
__kernfs_remove+0x2fc/0x360
kernfs_remove_by_name_ns+0x5c/0xac
remove_files+0x48/0x90
sysfs_remove_group+0x50/0xa0
sysfs_remove_groups+0x44/0x60
device_remove_attrs+0x60/0x80
device_del+0x134/0x360
device_unregister+0x28/0x80
i2c_del_adapter+0x210/0x2a4
i2c_mux_del_adapters+0xa4/0xfc
pca9641_remove+0x24/0x34 [i2c_mux_pca9641]
i2c_device_remove+0x60/0xd4
__device_release_driver+0x164/0x1f4
device_release_driver+0x38/0x4c
bus_remove_device+0xe0/0x15c
device_del+0x13c/0x360
device_unregister+0x28/0x80
i2c_unregister_device+0x5c/0x70
i2c_sysfs_delete_device+0x1b8/0x1f0
dev_attr_store+0x48/0x60
sysfs_kf_write+0x54/0x80
kernfs_fop_write+0x120/0x240
__vfs_write+0x4c/0x90
vfs_write+0xe8/0x1d0
ksys_write+0x7c/0x104
__arm64_sys_write+0x28/0x3c
el0_svc_handler+0x90/0x200
el0_svc+0x8/0x16c
other info that might help us debug this:
Possible unsafe locking scenario:
CPU0 CPU1
---- ----
lock(&adap->userspace_clients_lock);
lock(kn->count#159);
lock(&adap->userspace_clients_lock);
lock(kn->count#159);
*** DEADLOCK ***
5 locks held by .../9201:
#0: ffffff83f29c3ce0 (&f->f_pos_lock){+.+.}, at: __fdget_pos+0x74/0xe0
#1: ffffff857a2f0410 (sb_writers#4){.+.+}, at: vfs_write+0x174/0x1d0
#2: ffffff856bbbe680 (&of->mutex){+.+.}, at: kernfs_fop_write+0xf0/0x240
#3: ffffff857b329e80 (&adap->userspace_clients_lock){+.+.}, at: i2c_sysfs_delete_device+0x8c/0x1f0
#4: ffffff83f134ad80 (&dev->mutex){....}, at: device_release_driver+0x2c/0x4c
stack backtrace:
Call trace:
dump_backtrace+0x0/0x180
show_stack+0x28/0x3c
dump_stack+0xf4/0x150
print_circular_bug+0x21c/0x22c
check_noncircular+0x15c/0x1e0
__lock_acquire+0xe8c/0x1a90
lock_acquire+0xd8/0x270
__kernfs_remove+0x2fc/0x360
kernfs_remove_by_name_ns+0x5c/0xac
remove_files+0x48/0x90
sysfs_remove_group+0x50/0xa0
sysfs_remove_groups+0x44/0x60
device_remove_attrs+0x60/0x80
device_del+0x134/0x360
device_unregister+0x28/0x80
i2c_del_adapter+0x210/0x2a4
i2c_mux_del_adapters+0xa4/0xfc
pca9641_remove+0x24/0x34 [i2c_mux_pca9641]
i2c_device_remove+0x60/0xd4
__device_release_driver+0x164/0x1f4
device_release_driver+0x38/0x4c
bus_remove_device+0xe0/0x15c
device_del+0x13c/0x360
device_unregister+0x28/0x80
i2c_unregister_device+0x5c/0x70
i2c_sysfs_delete_device+0x1b8/0x1f0
dev_attr_store+0x48/0x60
sysfs_kf_write+0x54/0x80
kernfs_fop_write+0x120/0x240
__vfs_write+0x4c/0x90
vfs_write+0xe8/0x1d0
ksys_write+0x7c/0x104
__arm64_sys_write+0x28/0x3c
el0_svc_handler+0x90/0x200
el0_svc+0x8/0x16c
Signed-off-by: Alexander Sverdlin <[email protected]>
---
drivers/i2c/i2c-core-base.c | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/drivers/i2c/i2c-core-base.c b/drivers/i2c/i2c-core-base.c
index d43db2c..bd2e3e4 100644
--- a/drivers/i2c/i2c-core-base.c
+++ b/drivers/i2c/i2c-core-base.c
@@ -1272,9 +1272,14 @@ delete_device_store(struct device *dev, struct device_attribute *attr,
"delete_device", client->name, client->addr);
list_del(&client->detected);
+ /*
+ * We drop the mutex here, because device unregister
+ * will take sysfs lock (kn->count) which, as LOCKDEP
+ * would think, depends on this mutex
+ */
+ mutex_unlock(&adap->userspace_clients_lock);
i2c_unregister_device(client);
- res = count;
- break;
+ return count;
}
}
mutex_unlock(&adap->userspace_clients_lock);
--
2.10.2
> + /*
> + * We drop the mutex here, because device unregister
> + * will take sysfs lock (kn->count) which, as LOCKDEP
> + * would think, depends on this mutex
> + */
I have to ask: can't lockdep be fixed or the lock somehow be annotated
to prevent the false positive?
Hello Wolfram,
On 21/05/2022 11:50, Wolfram Sang wrote:
>> + /*
>> + * We drop the mutex here, because device unregister
>> + * will take sysfs lock (kn->count) which, as LOCKDEP
>> + * would think, depends on this mutex
>> + */
> I have to ask: can't lockdep be fixed or the lock somehow be annotated
> to prevent the false positive?
I would be happy to test any other proposal, I personally didn't find other working
approach until now. This particular patch actually has an advantage of reducing the
locked region, which makes sense if only userspace_clients is protected by the mutex.
--
Best regards,
Alexander Sverdlin.