Re: [PATCH] kernel/module: add documentation for try_module_get()

From: gregkh@xxxxxxxxxxxxxxxxxxx
Date: Tue Jul 27 2021 - 13:46:40 EST


On Tue, Jul 27, 2021 at 10:30:36AM -0700, Luis Chamberlain wrote:
> On Sat, Jul 24, 2021 at 12:15:10PM +0000, David Laight wrote:
> > From: Luis Chamberlain
> > > Sent: 22 July 2021 23:19
> > >
> > > There is quite a bit of tribal knowledge around proper use of
> > > try_module_get() and that it must be used only in a context which
> > > can ensure the module won't be gone during the operation. Document
> > > this little bit of tribal knowledge.
> > >
> > ...
> >
> > Some typos.
> >
> > > +/**
> > > + * try_module_get - yields to module removal and bumps reference count otherwise
> > > + * @module: the module we should check for
> > > + *
> > > + * This can be used to check if userspace has requested to remove a module,
> > a module be removed
> > > + * and if so let the caller give up. Otherwise it takes a reference count to
> > > + * ensure a request from userspace to remove the module cannot happen.
> > > + *
> > > + * Care must be taken to ensure the module cannot be removed during
> > > + * try_module_get(). This can be done by having another entity other than the
> > > + * module itself increment the module reference count, or through some other
> > > + * means which gaurantees the module could not be removed during an operation.
> > guarantees
> > > + * An example of this later case is using this call in a sysfs file which the
> > > + * module created. The sysfs store / read file operation is ensured to exist
> > ^^^^^^^^^^^^^^^^^^^
> > Not sure what that is supposed to mean.
>
> I'll clarify further. How about:
>
> The sysfs store / read file operations are gauranteed to exist using
> kernfs's active reference (see kernfs_active()).

But that has nothing to do with module reference counts. kernfs knows
nothing about modules.

> > So there is a potentially horrid race:
> > The module unload is going to do:
> > driver_data->module_ref = 0;
> > and elsewhere there'll be:
> > ref = driver_data->module_ref;
> > if (!ref || !try_module_get(ref))
> > return -error;
> >
> > You have to have try_module_get() to allow the module unload
> > function to sleep.
> > But the above code still needs a driver lock to ensure the
> > unload code doesn't race with the try_module_get() and the
> > 'ref' be invalidated before try_module_get() looks at it.
> > (eg if an interrupt defers processing.)
> >
> > So there can be no 'yielding'.
>
> Oh but there is. Consider access to a random sysfs file 'add_new_device'
> which takes as input a name, for driver foo, and so foo's
> add_new_foobar_device(name="bar") is called. Unless sysfs file
> "yields" by using try_module_get() before trying to add a new
> foo device called "bar", it will essentially be racing with the
> exit routine of module foo, and depending on how locking is implemented
> (most drivers get it wrong), this easily leads to crashes.
>
> In fact, this documentation patch was motivated by my own solution to a
> possible deadlock when sysfs is used. Using the same example above, if
> the same sysfs file uses *any* lock, which is *also* used on the exit
> routine, you can easily trigger a deadlock. This can happen for example
> by the lock being obtained by the removal routine, then the sysfs file
> gets called, waits for the lock to complete, then the module's exit
> routine starts cleaning up and removing sysfs files, but we won't be
> able to remove the sysfs file (due to kernefs active reference) until
> the sysfs file complets, but it cannot complete because the lock is
> already held.
>
> Yes, this is a generic problem. Yes I have proof [0]. Yes, a generic
> solution has been proposed [1], and because Greg is not convinced and I
> need to move on with life, I am suggesting a temporary driver specific
> solution (to which Greg is still NACK'ing, without even proposing any
> alternatives) [2].
>
> [0] https://lkml.kernel.org/r/20210703004632.621662-5-mcgrof@xxxxxxxxxx
> [1] https://lkml.kernel.org/r/20210401235925.GR4332@xxxxxxxxxxxxxxxxxxx
> [2] https://lkml.kernel.org/r/20210723174919.ka3tzyre432uilf7@garbanzo

My problem with your proposed solution is that it is still racy, you can
not increment your own module reference count from 0 -> 1 and expect it
to work properly. You need external code to do that somewhere.

Now trying to tie sysfs files to the modules that own them would be
nice, but as we have seen, that way lies way too many kernel changes,
right?

Hm, maybe. Did we think about this from the kobj_attribute level? If
we use the "wrapper" logic there and the use of the macros we already
have for attributes, we might be able to get the module pointer directly
"for free".

Did we try that? this thread has been going on for so long I can't
remember anymore...

> > I'm pretty much certain try_module_get(THIS_MODULE) is pretty
> > much never going to fail.
>
> It would have to take something very asynchronous and detached from
> the module to run. But the only thing I can think now, is something
> takes a module pointer right before after try_stop_module() and then
> a piece of code in between try_stop_module() and free_module()
> asynchronously tries to run something with that pointer.
>
> In the end I can only think of buggy code. Perhaps the more type of
> common issue could be code which purposely leave module pointers around
> with the intent of cleaning up using a module removal notifier event and
> that for some stupid reason runs something asynchronously with that
> pointer.
>
> > (It is mostly needed to give a worker thread a reference.)
>
> Greg, do you have a real world example which demonstrates the race
> better? Or perhaps a selftest? Or a kunit test?

Hah, nice try :)

greg k-h