At Sat, 16 Aug 2014 04:50:07 +0200,
Luis R. Rodriguez wrote:
> 
> On Fri, Aug 15, 2014 at 04:39:02PM +0200, Oleg Nesterov wrote:
> > On 08/15, Luis R. Rodriguez wrote:
> > >
> > > On Wed, Aug 13, 2014 at 07:51:01PM +0200, Oleg Nesterov wrote:
> > > > On 08/12, Luis R. Rodriguez wrote:
> > > > >
> > > > > +/* To be used by modules which can take over 30 seconds at probe */
> > > >
> > > > Probably the comment should explain that this hack should only be
> > > > used if the driver is buggy and is wating for "real fix".
> > > >
> > > > > +#define module_long_probe_init(initfn)                               
> > > > > \
> > > > > +     static struct task_struct *__init_thread;               \
> > > > > +     static int _long_probe_##initfn(void *arg)              \
> > > > > +     {                                                       \
> > > > > +             return initfn();                                \
> > > > > +     }                                                       \
> > > > > +     static inline __init int __long_probe_##initfn(void)    \
> > > > > +     {                                                       \
> > > > > +             __init_thread = kthread_run(_long_probe_##initfn,\
> > > > > +                                         NULL,               \
> > > > > +                                         #initfn);           \
> > > > > +             if (IS_ERR(__init_thread))                      \
> > > > > +                     return PTR_ERR(__init_thread);          \
> > > > > +             return 0;                                       \
> > > > > +     }                                                       \
> > > > > +     module_init(__long_probe_##initfn);
> > > > > +/* To be used by modules that require module_long_probe_init() */
> > > > > +#define module_long_probe_exit(exitfn)                               
> > > > > \
> > > > > +     static inline void __long_probe_##exitfn(void)          \
> > > > > +     {                                                       \
> > > > > +             exitfn();                                       \
> > > > > +             if (__init_thread)                              \
> > > > > +                     kthread_stop(__init_thread);            \
> > > > > +     }                                                       \
> > > >
> > > > exitfn() should be called after kthread_stop(), and only if initfn()
> > > > returns 0. So it should probably do
> > > >
> > > >         int err = kthread_stop(__init_thread);
> > > >         if (!err)
> > > >                 exitfn();
> > >
> > > Thanks! With the check for __init_thread as well as it can be
> > > ERR_PTR(-ENOMEM), ERR_PTR(-EINTR), or NULL (for whatever other
> > > reason).
> > 
> > Do you mean __long_probe_##exitfn() should also check 
> > ERR_PTR(__init_thread)?
> > I don't think so. If kthread_run() above fails, module_init() should return
> > the error (it does), so module_exit() won't be called.
> 
> Good point.
> 
> > > > But there is an additional complication, you can't use __init_thread
> > > > without get_task_struct(),
> > >
> > > Can you elaborate why ? kthread_stop() uses get_task_struct(),
> > 
> > This is too late. This task_struct can be already freed/reused. See below.
> > 
> > > wake_up_process() and finally put_task_struct(), and we're the
> > > only user of this thread. Also kthread_run() ensures wake_up_process()
> > > gets called on startup, so not sure where the race would be provided
> > > all users here and with the respective helpers on buggy drivers.
> > >
> > > > so  __long_probe_##initfn() can't use
> > > > kthread_run(). It needs kthread_create() + get_task_struct() + wakeup.
> > >
> > > I fail to see why we'd need to add get_task_struct() on
> > > module_long_probe_init(), can you clarify?
> > 
> > kthread_stop(kthread_run(callback)) is only safe if callback() can not exit
> > on its own, without checking kthread_should_stop(). And btw that is why
> > kthread_stop() does get_task_struct()).
> > 
> > If callback() can exit (if it calls do_exit() or simply returns), then 
> > nothing
> > protects this task_struct, it will be freed.
> 
> OK thanks, yeah I see the issue now, and I was able to create a null
> pointer dereference by simply calling schedule() quite a bit, will
> roll in the required fixes, but come to think of it if there are
> other uses (I haven't SmPLd grep'd for grammar uses yet) perhaps
> generic helpers would be good? kthread_run_alloc() kthread_run_free().

How about just increasing/decreasing the module count for blocking the
exit call?  For example:

#define module_long_probe_init(initfn)                          \
        static int _long_probe_##initfn(void *arg)              \
        {                                                       \
                int ret = initfn();                             \
                module_put(THIS_MODULE);                        \
                return ret;                                     \
        }                                                       \
        static inline __init int __long_probe_##initfn(void)    \
        {                                                       \
                struct task_struct *__init_thread;              \
                __module_get(THIS_MODULE);                      \
                __init_thread = kthread_run(_long_probe_##initfn,\
                                            NULL,               \
                                            #initfn);           \
                if (IS_ERR(__init_thread)) {                    \
                        module_put(THIS_MODULE);                \
                        return PTR_ERR(__init_thread);          \
                }                                               \
                return 0;                                       \
        }                                                       \
        module_init(__long_probe_##initfn);
/* To be used by modules that require module_long_probe_init() */
#define module_long_probe_exit(exitfn)                          \
        module_exit(exitfn);



Takashi
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Reply via email to