Re: "mm: uninline copy_overflow()" breaks i386 build in Mellanox MLX4

[Jason Gunthorpe]

On Thu, Apr 21, 2022 at 10:47:01PM +0200, Mateusz Jończyk wrote:
> Hello,
> 
> commit ad7489d5262d ("mm: uninline copy_overflow()")
> 
> breaks for me a build for i386 in the Mellanox MLX4 driver:
> 
>         In file included from ./arch/x86/include/asm/preempt.h:7,
>                          from ./include/linux/preempt.h:78,
>                          from ./include/linux/percpu.h:6,
>                          from ./include/linux/context_tracking_state.h:5,
>                          from ./include/linux/hardirq.h:5,
>                          from drivers/net/ethernet/mellanox/mlx4/cq.c:37:
>         In function ‘check_copy_size’,
>             inlined from ‘copy_to_user’ at ./include/linux/uaccess.h:159:6,
>             inlined from ‘mlx4_init_user_cqes’ at drivers/net/ethernet/mellanox/mlx4/cq.c:317:9,
>             inlined from ‘mlx4_cq_alloc’ at drivers/net/ethernet/mellanox/mlx4/cq.c:394:10:
>         ./include/linux/thread_info.h:228:4: error: call to ‘__bad_copy_from’ declared with attribute error: copy source size is too small
>           228 |    __bad_copy_from();
>               |    ^~~~~~~~~~~~~~~~~
>         make[5]: *** [scripts/Makefile.build:288: drivers/net/ethernet/mellanox/mlx4/cq.o] Błąd 1
>         make[4]: *** [scripts/Makefile.build:550: drivers/net/ethernet/mellanox/mlx4] Błąd 2
>         make[3]: *** [scripts/Makefile.build:550: drivers/net/ethernet/mellanox] Błąd 2
>         make[2]: *** [scripts/Makefile.build:550: drivers/net/ethernet] Błąd 2
>         make[1]: *** [scripts/Makefile.build:550: drivers/net] Błąd 2
> 
> Reverting this commit fixes the build. Disabling Mellanox Ethernet drivers
> in Kconfig (tested only with also disabling of all Infiniband support) also fixes the build.
> 
> It appears that uninlining of copy_overflow() causes GCC to analyze the code deeper.

This looks like a compiler bug to me, array_size(entries, cqe_size)
cannot be known at compile time, so the __builtin_constant_p(bytes)
should be compile time false meaning the other two bad branches should
have been eliminated.

Jason