On Mon, Dec 09, 2024 at 01:00:05PM +0100, Alice Ryhl wrote: > On Mon, Dec 9, 2024 at 12:53 PM Greg Kroah-Hartman > <gregkh@xxxxxxxxxxxxxxxxxxx> wrote: > > > > On Mon, Dec 09, 2024 at 12:38:32PM +0100, Alice Ryhl wrote: > > > On Mon, Dec 9, 2024 at 12:10 PM Greg Kroah-Hartman > > > <gregkh@xxxxxxxxxxxxxxxxxxx> wrote: > > > > > > > > On Mon, Dec 09, 2024 at 11:50:57AM +0100, Alice Ryhl wrote: > > > > > On Mon, Dec 9, 2024 at 9:48 AM Greg Kroah-Hartman > > > > > <gregkh@xxxxxxxxxxxxxxxxxxx> wrote: > > > > > > > > > > > > On Mon, Dec 09, 2024 at 07:27:47AM +0000, Alice Ryhl wrote: > > > > > > > Providing access to the underlying `struct miscdevice` is useful for > > > > > > > various reasons. For example, this allows you access the miscdevice's > > > > > > > internal `struct device` for use with the `dev_*` printing macros. > > > > > > > > > > > > > > Note that since the underlying `struct miscdevice` could get freed at > > > > > > > any point after the fops->open() call, only the open call is given > > > > > > > access to it. To print from other calls, they should take a refcount on > > > > > > > the device to keep it alive. > > > > > > > > > > > > The lifespan of the miscdevice is at least from open until close, so > > > > > > it's safe for at least then (i.e. read/write/ioctl/etc.) > > > > > > > > > > How is that enforced? What happens if I call misc_deregister while > > > > > there are open fds? > > > > > > > > You shouldn't be able to do that as the code that would be calling > > > > misc_deregister() (i.e. in a module unload path) would not work because > > > > the module reference count is incremented at this point in time due to > > > > the file operation module reference. > > > > > > Oh .. so misc_deregister must only be called when the module is being unloaded? > > > > Traditionally yes, that's when it is called. Do you see it happening in > > any other place in the kernel today? > > I had not looked, but I know that Binder allows dynamically creating > and removing its devices at runtime. It happens to be the case that > this is only supported when binderfs is used, which is when it doesn't > use miscdevice, so technically Binder does not call misc_deregister() > outside of module unload, but following its example it's not hard to > imagine that such removals could happen. That's why those are files and not misc devices :) > > > > Yeah, it's a horrid hack, and one day we will put "real" revoke logic in > > > > here to detach the misc device from the file operations if this were to > > > > happen. It's a very very common anti-pattern that many subsystems have > > > > that is a bug that we all have been talking about for a very very long > > > > time. Wolfram even has a plan for how to fix it all up (see his Japan > > > > LinuxCon talk from 2 years ago), but I don't think anyone is doing the > > > > work on it :( > > > > > > > > The media and drm layers have internal hacks/work-arounds to try to > > > > handle this issue, but luckily for us, the odds of a misc device being > > > > dynamically removed from the system is pretty low. > > > > > > > > Once / if ever, we get the revoke type logic implemented, then we can > > > > apply that to the misc device code and follow it through to the rust > > > > side if needed. > > > > > > If dynamically deregistering is not safe, then we need to change the > > > Rust abstractions to prevent it. > > > > Dynamically deregistering is not unsafe, it's just that I don't think > > you will physically ever have the misc_deregister() path called if a > > file handle is open. Same should be the case for rust code, it should > > "just work" without any extra code to do so. > > Well, if I give files access to the struct miscdevice in all fops > hooks, then deregistering does become unsafe since accessing it in an > ioctl after deregistering would be a UAF. I'd like to prevent the user > from doing that. I don't think that the deregister would succeed in the vfs layer if an open file reference was currently held, but I haven't tried that in a long time. If you can come up with a way to prevent that, wonderful, but I wouldn't worry too much as again, this "should not" happen due to the file reference count, and if it does, it's a major logic error on the driver's part, just like we have today in C. thanks, greg k-h
