Re: [PATCH] ARM: fix __get_user_check() in case uaccess_* calls are not inlined

[Nick Desaulniers]

On Sun, Sep 29, 2019 at 11:00 PM Masahiro Yamada
<yamada.masahiro@xxxxxxxxxxxxx> wrote:
>
> KernelCI reports that bcm2835_defconfig is no longer booting since
> commit ac7c3e4ff401 ("compiler: enable CONFIG_OPTIMIZE_INLINING
> forcibly"):
>
>   https://lkml.org/lkml/2019/9/26/825
>
> I also received a regression report from Nicolas Saenz Julienne:
>
>   https://lkml.org/lkml/2019/9/27/263
>
> This problem has cropped up on arch/arm/config/bcm2835_defconfig
> because it enables CONFIG_CC_OPTIMIZE_FOR_SIZE. The compiler tends
> to prefer not inlining functions with -Os. I was able to reproduce
> it with other boards and defconfig files by manually enabling
> CONFIG_CC_OPTIMIZE_FOR_SIZE.
>
> The __get_user_check() specifically uses r0, r1, r2 registers.

Yep, that part is obvious, but...

> So, uaccess_save_and_enable() and uaccess_restore() must be inlined
> in order to avoid those registers being overwritten in the callees.

Right, r0, r1, r2 are caller saved, meaning that __get_user_check must
save/restore them when making function calls. So
uaccess_save_and_enable() and uaccess_restore() should either be made
into macros (macros and typecheck (see include/linux/typecheck.h) are
peanut butter and chocolate), or occur at different points in the
function when those register variables are no longer in use.

>
> Prior to commit 9012d011660e ("compiler: allow all arches to enable
> CONFIG_OPTIMIZE_INLINING"), the 'inline' marker was always enough for
> inlining functions, except on x86.
>
> Since that commit, all architectures can enable CONFIG_OPTIMIZE_INLINING.
> So, __always_inline is now the only guaranteed way of forcible inlining.
>
> I want to keep as much compiler's freedom as possible about the inlining
> decision. So, I changed the function call order instead of adding
> __always_inline around.
>
> Call uaccess_save_and_enable() before assigning the __p ("r0"), and
> uaccess_restore() after evacuating the __e ("r0").
>
> Fixes: 9012d011660e ("compiler: allow all arches to enable CONFIG_OPTIMIZE_INLINING")
> Reported-by: "kernelci.org bot" <bot@xxxxxxxxxxxx>
> Reported-by: Nicolas Saenz Julienne <nsaenzjulienne@xxxxxxx>
> Signed-off-by: Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx>
> ---
>
>  arch/arm/include/asm/uaccess.h | 8 +++++---
>  1 file changed, 5 insertions(+), 3 deletions(-)
>
> diff --git a/arch/arm/include/asm/uaccess.h b/arch/arm/include/asm/uaccess.h
> index 303248e5b990..559f252d7e3c 100644
> --- a/arch/arm/include/asm/uaccess.h
> +++ b/arch/arm/include/asm/uaccess.h
> @@ -191,11 +191,12 @@ extern int __get_user_64t_4(void *);
>  #define __get_user_check(x, p)                                         \
>         ({                                                              \
>                 unsigned long __limit = current_thread_info()->addr_limit - 1; \
> +               unsigned int __ua_flags = uaccess_save_and_enable();    \
>                 register typeof(*(p)) __user *__p asm("r0") = (p);      \
>                 register __inttype(x) __r2 asm("r2");                   \
>                 register unsigned long __l asm("r1") = __limit;         \
>                 register int __e asm("r0");                             \

What does it mean for there to be two different local variables pinned
to the same register? Ie. it looks like __e and __p are defined to
exist in r0.  Would having one variable and an explicit cast result in
differing storage?

> -               unsigned int __ua_flags = uaccess_save_and_enable();    \
> +               unsigned int __err;                                     \
>                 switch (sizeof(*(__p))) {                               \
>                 case 1:                                                 \
>                         if (sizeof((x)) >= 8)                           \
> @@ -223,9 +224,10 @@ extern int __get_user_64t_4(void *);
>                         break;                                          \
>                 default: __e = __get_user_bad(); break;                 \

^ I think this assignment to __e should be replaced with an assignment
to __err?  We no longer need the register at this point and could skip
the assignment of x.

>                 }                                                       \
> -               uaccess_restore(__ua_flags);                            \
> +               __err = __e;                                            \
>                 x = (typeof(*(p))) __r2;                                \
> -               __e;                                                    \
> +               uaccess_restore(__ua_flags);                            \
> +               __err;                                                  \
>         })
>
>  #define get_user(x, p)                                                 \
> --
> 2.17.1
>


-- 
Thanks,
~Nick Desaulniers