drivers/base/core: refcount kobject and bus on device attribute read / store
Subsystems (2)
PATCH 1/1 drivers/base/core: refcount kobject and bus on device attribute read / store
It's possible today to have a device attribute read or store race against device removal. When this happens there is a small chance that the derefence for the private data area of the driver is NULL.
Let's consider the zram driver as an example. Its possible to run into a race where a sysfs knob is being used, we get preempted, and a zram device is removed before we complete use of the sysfs knob. This can happen for instance on block devices, where for instance the zram block devices just part of the private data of the block device.
For instance this can happen in the following two situations as examples to illustrate this better:
CPU 1 CPU 2destroy_devices
...
compact_store()
zram = dev_to_zram(dev);idr_for_each(zram_remove_cb
zram_remove
...
kfree(zram)
down_read(&zram->init_lock);
CPU 1 CPU 2hot_remove_store
compact_store()
zram = dev_to_zram(dev);
zram_remove
kfree(zram)
down_read(&zram->init_lock);To ensure the private data pointer is valid we could use bdget() / bdput() in between access, however that would mean doing that in all sysfs reads/stores on the driver. Instead a generic solution for all drivers is to ensure the device kobject is still valid and also the bus, if a bus is present.
This issue does not fix a known crash, however this race was spotted by Minchan Kim through code inspection upon code review of another zram patch.
--- drivers/base/bus.c | 4 ++-- drivers/base/core.c | 45 +++++++++++++++++++++++++++++++++++++++++---- 2 files changed, 43 insertions(+), 6 deletions(-)
drivers/base/bus.c +2 -2
@@ -39,7 +39,7 @@ static struct kset *system_kset;static int __must_check bus_rescan_devices_helper (struct device *dev,void *data); -static struct bus_type *bus_get (struct bus_type *bus) +struct bus_type *bus_get (struct bus_type *bus) {if (bus) {kset_get (&bus->p->subsys);
@@ -48,7 +48,7 @@ static struct bus_type *bus_get(struct bus_type *bus)return NULL ; } -static void bus_put (struct bus_type *bus) +void bus_put (struct bus_type *bus) {if (bus)kset_put (&bus->p->subsys);
drivers/base/core.c +41 -4
72 lines not on this page · raw message
Signature · 1 line
2.27.0
Greg KH
maintainerdrivers/base/core.c EXPORT_SYMBOL(dev_driver_string); · line 2044
Quoted 9 lines · Luis Chamberlain
drivers/base/core.c +3
-0@@ -2039,31 +2039,68 @@ EXPORT_SYMBOL(dev_driver_string);#define to_dev_attr (_attr) container_of(_attr, struct device_attribute, attr) +struct bus_type *bus_get (struct bus_type *bus); +void bus_put (struct bus_type *bus); +
Didn't checkpatch complain about this?
We have a local .h file for stuff like this, can you please use it?
thanks,
greg k-h
Greg KH
maintainerdrivers/base/core.c EXPORT_SYMBOL(dev_driver_string); · line 2055
Quoted 13 lines · Luis Chamberlain
static
ssize_tdev_attr_show(struct kobject *kobj, struct attribute *attr,drivers/base/core.c +6
-2char *buf) { -struct device_attribute *dev_attr =to_dev_attr (attr); -struct device *dev =kobj_to_dev (kobj); +struct device_attribute *dev_attr; +struct device *dev; +struct bus_type *bus =NULL ;ssize_t ret = -EIO ; + dev =get_device (kobj_to_dev (kobj)); +if (!dev) +return ret;
That check is impossible to ever hit, please recognize what things like kobj_to_dev() really are doing when calling it.
thanks,
greg k-h