Andrew Morton wrote:
> On Wed, 29 Aug 2007 11:33:06 -0700 (PDT) [email protected] wrote:
>
>> http://bugzilla.kernel.org/show_bug.cgi?id=8957
>>
>> Summary: Exported functions and variables should not be reachable
>> by the outside of the module until module_init finishes
>> Product: Other
>> Version: 2.5
>> KernelVersion: 2.6.23-rc4
>> Platform: All
>> OS/Version: Linux
>> Tree: Mainline
>> Status: NEW
>> Severity: normal
>> Priority: P1
>> Component: Modules
>> AssignedTo: [email protected]
>> ReportedBy: [email protected]
>>
>>
>> Problem Description: a module's exported functions can be called before before
>> they are properly initialized by the module_init function.
>>
>> Steps to reproduce: write a module that exports functions that require
>> initialization by the module_init function to work correctly.
>>
>> E.g. spin lock variables are no longer allowed to be initialized by C
>> initializers of the module but only by spin_lock_init that can be called by the
>> module_init function. If an exported function calls spin_lock before it is
>> initialized, it deadlocks.
>>
>
> ooh, nice bug ;)
Under what circumstances is this actually happening? What are these
functions that are being called?
Normally things are set up such that this isn't a problem, i.e. if
module A depends on module B, module A can't load until module B is
finished loading.
--
Robert Hancock Saskatoon, SK, Canada
To email, remove "nospam" from [email protected]
Home Page: http://www.roberthancock.com/
On Wed, 29 Aug 2007 19:33:48 -0600 Robert Hancock <[email protected]> wrote:
> Andrew Morton wrote:
> > On Wed, 29 Aug 2007 11:33:06 -0700 (PDT) [email protected] wrote:
> >
> >> http://bugzilla.kernel.org/show_bug.cgi?id=8957
> >>
> >> Summary: Exported functions and variables should not be reachable
> >> by the outside of the module until module_init finishes
> >> Product: Other
> >> Version: 2.5
> >> KernelVersion: 2.6.23-rc4
> >> Platform: All
> >> OS/Version: Linux
> >> Tree: Mainline
> >> Status: NEW
> >> Severity: normal
> >> Priority: P1
> >> Component: Modules
> >> AssignedTo: [email protected]
> >> ReportedBy: [email protected]
> >>
> >>
> >> Problem Description: a module's exported functions can be called before before
> >> they are properly initialized by the module_init function.
> >>
> >> Steps to reproduce: write a module that exports functions that require
> >> initialization by the module_init function to work correctly.
> >>
> >> E.g. spin lock variables are no longer allowed to be initialized by C
> >> initializers of the module but only by spin_lock_init that can be called by the
> >> module_init function. If an exported function calls spin_lock before it is
> >> initialized, it deadlocks.
> >>
> >
> > ooh, nice bug ;)
>
> Under what circumstances is this actually happening? What are these
> functions that are being called?
>
> Normally things are set up such that this isn't a problem, i.e. if
> module A depends on module B, module A can't load until module B is
> finished loading.
>
Good point.
This thus-far-undescribed module could make its internals externally
visible via one of the kernel's many register_foo() interfaces, but it
would be a buggy module if it was doing register_foo(my_foo) before
my_foo() was ready to be called.
On Wed, 29 Aug 2007, Robert Hancock wrote:
> Andrew Morton wrote:
> > On Wed, 29 Aug 2007 11:33:06 -0700 (PDT) [email protected]
> > wrote:
> >
> > > http://bugzilla.kernel.org/show_bug.cgi?id=8957
> > >
> > > Summary: Exported functions and variables should not be
> > > reachable
> > > by the outside of the module until module_init
> > > finishes
> > > Product: Other
> > > Version: 2.5
> > > KernelVersion: 2.6.23-rc4
> > > Platform: All
> > > OS/Version: Linux
> > > Tree: Mainline
> > > >> Status: NEW
> > > Severity: normal
> > > Priority: P1
> > > Component: Modules
> > > AssignedTo: [email protected]
> > > ReportedBy: [email protected]
> > >
> > >
> > > Problem Description: a module's exported functions can be called before
> > > before
> > > they are properly initialized by the module_init function.
> > >
> > > Steps to reproduce: write a module that exports functions that require
> > > initialization by the module_init function to work correctly.
> > >
> > > E.g. spin lock variables are no longer allowed to be initialized by C
> > > initializers of the module but only by spin_lock_init that can be called
> > > by the
> > > module_init function. If an exported function calls spin_lock before it is
> > > initialized, it deadlocks.
Hmm, can you post some sample code / sample module to reproduce this?
I don't think exported symbols can be resolved till our module finishes
loading + initializing. There's a whole lot of dancing in the libusual
module precisely to cope with this behaviour.
> > ooh, nice bug ;)
>
> Under what circumstances is this actually happening? What are these functions
> that are being called?
>
> Normally things are set up such that this isn't a problem, i.e. if module A
> depends on module B, module A can't load until module B is finished loading.
See drivers/usb/storage/libusual.c -- pretty unusual goings on there :-)
It needs to request_module() another module (that will reference our
exported symbols). To cope with the fact that our exported modules
_cannot_ be resolved till we finish loading, it uses semaphore-used-as-
completion-handler kludge to let another "probe" kthread know when our
module_init() function is done, so that it can proceed to request_module()
the other module.
Interestingly, the kthread that request_module()s the other module is
spawned from the struct usb_driver ->probe() function (not an exported
function) and the claim there is that (1) usb_driver ->probe() can be
called out without the module_init() of libusual having finished, and,
(2) the newly requested module's loading will fail because it cannot
resolve libusual's exported symbols till we have finished module_init().
Satyam
Hi Andrew,
On Wed, 29 Aug 2007, Andrew Morton wrote:
> On Wed, 29 Aug 2007 19:33:48 -0600 Robert Hancock <[email protected]> wrote:
>
> > Andrew Morton wrote:
> > > On Wed, 29 Aug 2007 11:33:06 -0700 (PDT) [email protected] wrote:
> > >
> > >> http://bugzilla.kernel.org/show_bug.cgi?id=8957
> > >>
> > >> Summary: Exported functions and variables should not be reachable
> > >> by the outside of the module until module_init finishes
> > >> Product: Other
> > >> Version: 2.5
> > >> KernelVersion: 2.6.23-rc4
> > >> Platform: All
> > >> OS/Version: Linux
> > >> Tree: Mainline
> > >> Status: NEW
> > >> Severity: normal
> > >> Priority: P1
> > >> Component: Modules
> > >> AssignedTo: [email protected]
> > >> ReportedBy: [email protected]
> > >>
> > >>
> > >> Problem Description: a module's exported functions can be called before before
> > >> they are properly initialized by the module_init function.
> > >>
> > >> Steps to reproduce: write a module that exports functions that require
> > >> initialization by the module_init function to work correctly.
> > >>
> > >> E.g. spin lock variables are no longer allowed to be initialized by C
> > >> initializers of the module but only by spin_lock_init that can be called by the
> > >> module_init function. If an exported function calls spin_lock before it is
> > >> initialized, it deadlocks.
> > >
> > > ooh, nice bug ;)
> >
> > Under what circumstances is this actually happening? What are these
> > functions that are being called?
> >
> > Normally things are set up such that this isn't a problem, i.e. if
> > module A depends on module B, module A can't load until module B is
> > finished loading.
>
> Good point.
>
> This thus-far-undescribed module could make its internals externally
> visible via one of the kernel's many register_foo() interfaces,
What you're saying is a plausible problem, but note that it is quite a
completely different issue to what Matti Linnanvuori suggested in the
original bug report.
The report was about module B (which depends on module A, because it
references symbol exported by module A) being able to call a function
(or access data) /exported/ by module A _without_ the module_init()
function of module A having finished completely (and hence the possibility
of accessing uninitialized data etc). But this is not possible -- see the
last reply to Matti.
You're referring to is a module implementing an (possibly un-exported)
function that refers to module-local data, and registering that function
(say through a notifier_block) _before_ initializing_ the data used by
that function. But ...
> but it
> would be a buggy module if it was doing register_foo(my_foo) before
> my_foo() was ready to be called.
... exactly. That module is the buggy culprit here, nothing wrong with
the kernel's core module code.
[ BTW I suspect there /are/ modules out there that get this register_foo()
ordering wrong in their module_init functions.
Even more widespread (as I have noticed) is the sad habit of modules
to not unregister_foo() their stuff (in the module_exit function) in
the exact reverse order of the register_foo() calls made during
module_init. This can clearly lead to oopsen, but the only reason why
we don't see them frequently is because the module_init and module_exit
codepaths are rarely ever executed at runtime, and even more rarely
concurrently with other stuff that's using the module. ]
Satyam