Re: [CRAZY-RFF] debugfs: track open files and release on remove

[Johannes Berg]

On Sat, 2020-10-10 at 11:38 +0200, Greg KH wrote:
> On Fri, Oct 09, 2020 at 10:48:09AM +0200, Johannes Berg wrote:
> > On Fri, 2020-10-09 at 10:47 +0200, Greg KH wrote:
> > 
> > > > I think adding the .owner everywhere would be good, and perhaps we can
> > > > somehow put a check somewhere like
> > > > 
> > > > 	WARN_ON(is_module_address((unsigned long)fops) && !fops->owner);
> > > > 
> > > > to prevent the issue in the future?
> > > 
> > > That will fail for all of the debugfs_create_* operations, as there is
> > > only one set of file operations for all of the different files created
> > > with these calls.
> > 
> > Why would it fail? Those have their fops in the core debugfs code, which
> > might have a .owner assigned but is probably built-in anyway?
> 
> Bad choice of terms, it would "fail" in that this type of check would
> never actually work because the debugfs code is built into the kernel,
> and there is no module owner for it.  But the value it is referencing is
> an address in a module.

Ahh.

Yes and no. I mean, yes, the check wouldn't really work.

But OTOH, this is exactly what the proxy_fops protects against.

The _only_ thing that proxy_fops *doesn't* proxy is the ->release()
method.

If you have a debugfs file that's say debugfs_create_u32(), then the
code is all built into the kernel, and - if ->release() even exists, I
didn't check now - it would surely not dereference the pointer you gave
to debugfs_create_u32(). So as long as the file is debugfs_remove()d
before the pointer becomes invalid, there's no issue.

The check I'm proposing (and actually wrote in my separate RFC patch
that didn't seem quite as crazy) would basically protect the ->release()
method only, if needed. Everything else is handled by proxy_fops.

> > > Which, now that I remember it, is why we went down the proxy "solution"
> > > in the first place :(
> > 
> > Not sure I understand. That was related more to (arbitrary) files having
> > to be disappeared rather than anything else?
> 
> Isn't this the same issue?

Well, not exactly? The difference is that proxy_fops basically protects
the *value*, read/write/etc., but not ->release(). So it protects more
against bus unbind or the like, where the *device* disappears, rather
than the *code* disappearing.

Now, you still need to be careful that ->release() doesn't actually
access anything related to the device, of course. As long as we don't
have a general revoke() at least.

I guess in that sense this crazy patch actually makes things *better*
than the RFC patch because it *does* call the ->release() during
debugfs_remove() and therefore allows even ->release() to access data of
the device or other data structures that are being removed; whereas the
RFC patch I also sent doesn't protect that, it just protects the code
itself.

johannes