On Mon, Sep 27, 2021 at 09:38:02AM -0700, Luis Chamberlain wrote: > When driver sysfs attributes use a lock also used on module removal we > can race to deadlock. This happens when for instance a sysfs file on > a driver is used, then at the same time we have module removal call > trigger. The module removal call code holds a lock, and then the > driver's sysfs file entry waits for the same lock. While holding the > lock the module removal tries to remove the sysfs entries, but these > cannot be removed yet as one is waiting for a lock. This won't complete > as the lock is already held. Likewise module removal cannot complete, > and so we deadlock. > > This can now be easily reproducible with our sysfs selftest as follows: > > ./tools/testing/selftests/sysfs/sysfs.sh -t 0027 > > This uses a local driver lock. Test 0028 can also be used, that uses > the rtnl_lock(): > > ./tools/testing/selftests/sysfs/sysfs.sh -t 0028 > > To fix this we extend the struct kernfs_node with a module reference > and use the try_module_get() after kernfs_get_active() is called. As > documented in the prior patch, we now know that once kernfs_get_active() > is called the module is implicitly guarded to exist and cannot be removed. > This is because the module is the one in charge of removing the same > sysfs file it created, and removal of sysfs files on module exit will wait > until they don't have any active references. By using a try_module_get() > after kernfs_get_active() we yield to let module removal trump calls to > process a sysfs operation, while also preventing module removal if a sysfs > operation is in already progress. This prevents the deadlock. > > This deadlock was first reported with the zram driver, however the live Looks not see the lock pattern you mentioned in zram driver, can you share the related zram code? > patching folks have acknowledged they have observed this as well with > live patching, when a live patch is removed. I was then able to > reproduce easily by creating a dedicated selftest for it. > > A sketch of how this can happen follows, consider foo a local mutex > part of a driver, and used on the driver's module exit routine and > on one of its sysfs ops: > > foo.c: > static DEFINE_MUTEX(foo); > static ssize_t foo_store(struct device *dev, > struct device_attribute *attr, > const char *buf, size_t count) > { > ... > mutex_lock(&foo); > ... > mutex_lock(&foo); > ... > } > static DEVICE_ATTR_RW(foo); > ... > void foo_exit(void) > { > mutex_lock(&foo); > ... > mutex_unlock(&foo); > } > module_exit(foo_exit); > > And this can lead to this condition: > > CPU A CPU B > foo_store() > foo_exit() > mutex_lock(&foo) > mutex_lock(&foo) > del_gendisk(some_struct->disk); > device_del() > device_remove_groups() I guess the deadlock exists if foo_exit() is called anywhere. If yes, look the issue may not be related with removing module directly, right? Thanks, Ming