On 6/15/2021 10:58 PM, Greg KH wrote:
On Tue, Jun 15, 2021 at 12:03:26PM -0700, Siddharth Gupta wrote:
On 6/14/2021 9:56 PM, Greg KH wrote:
On Mon, Jun 14, 2021 at 07:21:08PM -0700, Siddharth Gupta wrote:
When cdev_add is called after device_add has been called there is no way for the userspace to know about the addition of a cdev as cdev_add itself doesn't trigger a uevent notification, or for the kernel to know about the change to devt. This results in two problems:
- mknod is never called for the cdev and hence no cdev appears on devtmpfs.
- sysfs links to the new cdev are not established.
The cdev needs to be added and devt assigned before device_add() is called in order for the relevant sysfs and devtmpfs entries to be created and the uevent to be properly populated.
So this means no one ever ran this code on a system that used devtmpfs?
How was it ever tested?
My testing was done with toybox + Android's ueventd ramdisk. As I mentioned in the discussion, the race became evident recently. I will make sure to test all such changes without systemd/ueventd in the future.
It isn't an issue of systemd/ueventd, those do not control /dev on a normal system, that is what devtmpfs is for.
I am not fully aware of when devtmpfs is enabled or not, but in case it is not - systemd/ueventd will create these files with mknod, right? I was even manually able to call mknod from the terminal when some of the remoteproc character device entries showed up (using major number from there, and minor number being the remoteproc id), and that allowed me to boot up the remoteprocs as well.
And devtmpfs nodes are only created if you create a struct device somewhere with a proper major/minor, which you were not doing here, so you must have had a static /dev on your test systems, right?
I am not sure of what you mean by a static /dev? Could you explain? In case you mean the character device would be non-functional, that is not the case. They have been working for us since the beginning.
Thanks, Sid
thanks,
greg k-h