Re: system gets stuck in a lock during boot

[Jason Baron]

On Tue, Oct 06, 2009 at 10:02:01PM -0400, Steven Rostedt wrote:
> > So the problem I'm seeing is an oops on boot caused by the call->system pointer
> > deference in event_create_dir(). The 'call' variable is of type 'struct
> > ftrace_event_call'. 
> > 
> > What's going on is that the 'struct ftrace_event_call' is of size 168 bytes
> > (sizeof(struct ftrace_event_call)) = 168 = 0xA8. However, in memory the
> > structures are 16-byte aligned. Thus, the stride for walking through the
> > pointers needs to be 176 (0xB0), but instead its 168 causing the oops.
> > 
> > I've only seen this issue while using gcc (GCC) 4.5.0 20090916, on a
> > vanilla 2.6.31 kernel.
> > 
> > That said, I'm not sure the compiler is doing the wrong thing here. The
> > 'struct ftrace_event_call' contains an embedded 'struct list_head' which
> > is 16 bytes. According to the gcc docs, the aligned attribute, 'specifies a
> > minimum alignment for the variable or structure field, measured in bytes'.
> > Thus, at least according to the docs, gcc can increase the alignment of the
> > 'struct ftrace_event_call', from its original specification of 4, to 16. Even
> > in the case where we are working corectly the structures are 8-byte aligned.
> > 
> > Thus, I would reccommend the patch below as a preventive measure. Its
> > the minimal patch I've found to resolve this issue. In general, if we
> > are going to walk data structures embedded in a special elf section, I
> > think the general rules needs to be to set the alignment to the power of
> > two which is greater than or equal to the largest item in the structure.
> > 
> > thanks,
> > 
> > -Jason
> > 
> > Signed-off-by: Jason Baron <jbaron@xxxxxxxxxx>
> > 
> > 
> > diff --git a/include/linux/ftrace_event.h b/include/linux/ftrace_event.h
> > index a81170d..7182f03 100644
> > --- a/include/linux/ftrace_event.h
> > +++ b/include/linux/ftrace_event.h
> > @@ -124,7 +124,10 @@ struct ftrace_event_call {
> >  	atomic_t		profile_count;
> >  	int			(*profile_enable)(struct ftrace_event_call *);
> >  	void			(*profile_disable)(struct ftrace_event_call *);
> > -};
> > +} __attribute__((aligned(16)));
> > +
> > +/* Align to the largest field in the data structure:
> > + * sizeof(struct list_head) = 16 */
> 
> Is this true for i386?
> 
> I just tried this patch and it seems to work. Can you give it a try.
> 
> Signed-off-by: Steven Rostedt <rostedt@xxxxxxxxxxx>
> 
> 
> diff --git a/include/linux/ftrace_event.h b/include/linux/ftrace_event.h
> index 4ec5e67..044b70d 100644
> --- a/include/linux/ftrace_event.h
> +++ b/include/linux/ftrace_event.h
> @@ -133,7 +133,7 @@ struct ftrace_event_call {
>  	atomic_t		profile_count;
>  	int			(*profile_enable)(void);
>  	void			(*profile_disable)(void);
> -};
> +} __attribute__((aligned(sizeof(struct list_head))));
>  
>  #define FTRACE_MAX_PROFILE_SIZE	2048
>  
> diff --git a/include/trace/ftrace.h b/include/trace/ftrace.h
> index cc0d966..31e7637 100644
> --- a/include/trace/ftrace.h
> +++ b/include/trace/ftrace.h
> @@ -501,7 +501,6 @@ static void ftrace_profile_disable_##call(void)				\
>   * }
>   *
>   * static struct ftrace_event_call __used
> - * __attribute__((__aligned__(4)))
>   * __attribute__((section("_ftrace_events"))) event_<call> = {
>   *	.name			= "<call>",
>   *	.system			= "<system>",
> @@ -619,7 +618,6 @@ static int ftrace_raw_init_event_##call(void)				\
>  }									\
>  									\
>  static struct ftrace_event_call __used					\
> -__attribute__((__aligned__(4)))						\
>  __attribute__((section("_ftrace_events"))) event_##call = {		\
>  	.name			= #call,				\
>  	.system			= __stringify(TRACE_SYSTEM),		\
> 
> 

indeed your patch works as well for me, its much cleaner! 

However, I want to make sure this fix is sufficient and is the best way to
address this type of issue in general. For example, I know tracepoints are
using the aligned attribute in all 3 places -> definition, usage, and linker
alignment. (adding Mathieu to 'cc list). Is just the definition 'aligned'
sufficient? Also, once we find a method for solving these issues in general,
we need to review all users of this kind of technique to make sure they are
consistent. I also think your patch above needs to add a comment to say what
its doing.

thanks,

-Jason


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/