Re: Build regressions/improvements in v5.5-rc2

[Geert Uytterhoeven]

Hi Greg,

On Mon, Dec 16, 2019 at 9:38 AM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
> Below is the list of build error/warning regressions/improvements in
> v5.5-rc2[1] compared to v5.4[2].

> [1] http://kisskb.ellerman.id.au/kisskb/branch/linus/head/d1eef1c619749b2a57e514a3fa67d9a516ffa919/ (all 232 configs)
> [2] http://kisskb.ellerman.id.au/kisskb/branch/linus/head/219d54332a09e8d8741c1e1982f5eae56099de85/ (all 232 configs)

>   + /kisskb/src/arch/m68k/include/asm/string.h: warning: '__builtin_memcpy' forming offset [3, 4] is out of the bounds [0, 2] of object '__gu_val' with type 'short unsigned int' [-Warray-bounds]:  => 72:25

The upgrade from gcc 4.6.3 to 81.0 seems to have revealed a potential
issue in get_user() in arch/m68k/include/asm/uaccess_no.h.

        #define get_user(x, ptr)                                        \
        ({                                                              \
            int __gu_err = 0;                                           \
            typeof(x) __gu_val = 0;                                     \
            ^^^^^^^^^
            This is the type of the destination

            switch (sizeof(*(ptr))) {                                   \
            case 1:                                                     \
                __get_user_asm(__gu_err, __gu_val, ptr, b, "=d");       \
                break;                                                  \
            case 2:                                                     \
                __get_user_asm(__gu_err, __gu_val, ptr, w, "=r");       \
                break;                                                  \
            case 4:                                                     \
                __get_user_asm(__gu_err, __gu_val, ptr, l, "=r");       \
                break;                                                  \
            case 8:                                                     \
                memcpy((void *) &__gu_val, ptr, sizeof (*(ptr)));       \
                break;                                                  \
            default:                                                    \
                __gu_val = 0;                                           \
                __gu_err = __get_user_bad();                            \
                break;                                                  \
            }                                                           \
            (x) = (typeof(*(ptr))) __gu_val;                            \
            __gu_err;                                                   \
        })

ext2_ioctl() calls this like

        unsigned short rsv_window_size;
        if (get_user(rsv_window_size, (int __user *)arg)) { ... }

So a 32-bit value is being copied to an unsigned short value, leading
to the warning (for the memcpy() in the non-taken "case 8" branch).

Fortunately the compiler emits a register move for this, so no real harm
is done:

        | fs/ext2/ioctl.c:123:          if (get_user(rsv_window_size,
(int __user *)arg))
                move.l 48(%sp),%a0      | arg,
        #APP
        | 123 "fs/ext2/ioctl.c" 1
                movel (%a0),%d2 | *arg.32_43, __gu_val

The corresponding code in arch/m68k/include/asm/uaccess_mm.h
uses a temporary __gu_val of the right sized type based on the source
type to avoid this.


Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds