Re: [PATCH 2/2 V2] livepatch: handle kzalloc failure properly

[Josh Poimboeuf]

On Thu, Dec 13, 2018 at 03:39:20PM -0500, Joe Lawrence wrote:
> Ahh, I understand the question now.  Yeah, by making those routines local 
> static, the compiler applied optimizations that renamed the symbols:
> 
>   noinline static
>   % readelf --syms samples/livepatch/livepatch-shadow-mod.o | grep dummy_                                          
>        5: 0000000000000000    20 FUNC    LOCAL  DEFAULT    1 dummy_check.isra.0
>        7: 0000000000000020    52 FUNC    LOCAL  DEFAULT    1 dummy_free.constprop.1
>       12: 00000000000000c0    32 OBJECT  LOCAL  DEFAULT    3 dummy_list_mutex
>       13: 00000000000000e0    16 OBJECT  LOCAL  DEFAULT    3 dummy_list
>       15: 0000000000000160   115 FUNC    LOCAL  DEFAULT    1 dummy_alloc
> 
> 
> I can avoid that optimization (and successfully load all the modules) 
> by using either:
> 
>   __attribute__((optimize("O0"))) noinline static
>   % readelf --syms samples/livepatch/livepatch-shadow-mod.o | grep dummy_
>        6: 0000000000000000  6016 FUNC    LOCAL  DEFAULT    1 dummy_alloc
>       11: 00000000000000c0    32 OBJECT  LOCAL  DEFAULT    3 dummy_list_mutex
>       12: 00000000000000e0    16 OBJECT  LOCAL  DEFAULT    3 dummy_list
>       14: 0000000000001810    73 FUNC    LOCAL  DEFAULT    1 dummy_free
>       16: 0000000000001860   108 FUNC    LOCAL  DEFAULT    1 dummy_check
> 
> or:
> 
>   __noclone noinline static
>   % readelf --syms samples/livepatch/livepatch-shadow-mod.o | grep dummy_
>        5: 0000000000000000    22 FUNC    LOCAL  DEFAULT    1 dummy_check
>        7: 0000000000000020    51 FUNC    LOCAL  DEFAULT    1 dummy_free
>       12: 00000000000000c0    32 OBJECT  LOCAL  DEFAULT    3 dummy_list_mutex
>       13: 00000000000000e0    16 OBJECT  LOCAL  DEFAULT    3 dummy_list
>       15: 0000000000000160   115 FUNC    LOCAL  DEFAULT    1 dummy_alloc
> 
> but I'm not sure if either is the definitive way to avoid such
> optimization.  Anyone know for sure?

Yeah, for now I think "static __noclone" is the way to go.  Soon we'll
have a GCC flag which disables such optimizations for all functions.

And the dummy_list* variables can just be "static".

-- 
Josh