[PATCH] arm64: Fix 32-bit compatible userspace write size overflow error

[Mark Rutland]

On Thu, Nov 16, 2023 at 03:47:05PM +0800, Jinjie Ruan wrote:
> For 32-bit compatible userspace program, write with size = -1 return not
> -1 but unexpected other values, which is due to the __access_ok() check is
> not right.

Can you please explain why you believe that is unexpected?

e.g. Is that documented somewhere? Do you see a real application depending on
that somewhow?

> The specified "addr + size" is greater than 32-bit limit and
> should return -EFAULT, but TASK_SIZE_MAX still defined as UL(1) << VA_BITS
> in U32 mode, which is much greater than "addr + size" and cannot catch the
> overflow error.

The check against TASK_SIZE_MAX is not intended to catch 32-bit addr + size
overflow; it's intended to check that uaccesses never touch kernel memory. The
kernel's uaccess routines use 64-bit (or 65-bit) arithmetic, so these won't
wrap and access memory at the start of the user address space.

> Fix above error by checking 32-bit limit if it is 32-bit compatible
> userspace program.
> 
> How to reproduce:
> 
> The test program is as below:
> 
> cat test.c
> 	#include <unistd.h>
> 	#include <fcntl.h>
> 	#include <stdio.h>
> 	#include <stdint.h>
> 	#include <stdlib.h>
> 	#include <assert.h>
> 
> 	#define pinfo(fmt, args...) \
> 	    fprintf(stderr, "[INFO][%s][%d][%s]:"fmt, \
> 	    __FILE__,__LINE__,__func__,##args)
> 
> 	#undef SIZE_MAX
> 	#define SIZE_MAX -1
> 
> 	int main()
> 	{
> 	    char wbuf[3] = { 'x', 'y', 'z' };
> 	    char *path = "write.tmp";
> 	    int ret;
> 
> 	    int fd = open(path, O_RDWR | O_CREAT);
> 	    if (fd<0)
> 	    {
> 	        pinfo("fd=%d\n", fd);
> 	        exit(-1);
> 	    }
> 
> 	    assert(write(fd, wbuf, 3) == 3);
> 
> 	    ret = write (fd, wbuf, SIZE_MAX);
> 	    pinfo("ret=%d\n", ret);
> 	    pinfo("size_max=%d\n",SIZE_MAX);
> 	    assert(ret==-1);
> 	    close(fd);
> 	    pinfo("INFO: end\n");
> 
> 	    return 0;
> 	}
> 
> aarch64-linux-gnu-gcc --static test.c -o test
> arm-linux-gnueabi-gcc --static test.c -o test32
> 
> Before applying this patch, userspace 32-bit program return 1112 if the
> write size = -1 as below:
> 	/root # ./test
> 	[INFO][test.c][32][main]:ret=-1
> 	[INFO][test.c][33][main]:size_max=-1
> 	[INFO][test.c][36][main]:INFO: end
> 	/root # ./test32
> 	[INFO][test.c][32][main]:ret=1112
> 	[INFO][test.c][33][main]:size_max=-1
> 	test32: test.c:34: main: Assertion `ret==-1' failed.
> 	Aborted
> 
> After applying this patch, userspace 32-bit program return -1 if the write
> size = -1 as expected as below:
> 	/root # ./test
> 	[INFO][test.c][32][main]:ret=-1
> 	[INFO][test.c][33][main]:size_max=-1
> 	[INFO][test.c][36][main]:INFO: end
> 	/root # ./test32
> 	[INFO][test.c][32][main]:ret=-1
> 	[INFO][test.c][33][main]:size_max=-1
> 	[INFO][test.c][36][main]:INFO: end
> 
> Fixes: 967747bbc084 ("uaccess: remove CONFIG_SET_FS")

As above, this is *not* a fix. This is the intended behaviour.

AFAICT, the behaviour didn't change on arm64 in that commit either; we were
unconditionally using TASK_SIZE_MAX many commits earlier, e.g. in commit:

  3d2403fd10a1dbb3 ("arm64: uaccess: remove set_fs()")

... so the fixes tag is bogus on both fronts.

> Signed-off-by: Jinjie Ruan <ruanjinjie at huawei.com>
> ---
>  arch/arm64/include/asm/processor.h | 5 +++++
>  1 file changed, 5 insertions(+)
> 
> diff --git a/arch/arm64/include/asm/processor.h b/arch/arm64/include/asm/processor.h
> index e5bc54522e71..6a087d58a90a 100644
> --- a/arch/arm64/include/asm/processor.h
> +++ b/arch/arm64/include/asm/processor.h
> @@ -52,7 +52,12 @@
>  
>  #define DEFAULT_MAP_WINDOW_64	(UL(1) << VA_BITS_MIN)
>  #define TASK_SIZE_64		(UL(1) << vabits_actual)
> +#ifdef CONFIG_COMPAT
> +#define TASK_SIZE_MAX		(test_thread_flag(TIF_32BIT) ? \
> +				UL(0x100000000) : (UL(1) << VA_BITS))
> +#else
>  #define TASK_SIZE_MAX		(UL(1) << VA_BITS)
> +#endif

This isn't even the same as on 32-bit. On 32-bit arm, the task size split can
be 1G/3G, 2G/2G, or 3G/1G depending on configuration, and 4G/4G isn't currently
an option.

I don't believe that userspace is actually dependent upon this for functional
reasons, and I don't believe that there's a security issue here. Even if
access_ok() allows addr+size to go past 4G, the kernel address calculations are
64-bit and won't wrap.

For all the reasons above, I don't beleive this is correct nor do I believe
this is necesssary. Given that, NAK to this patch.

Thanks,
Mark.

>  
>  #ifdef CONFIG_COMPAT
>  #if defined(CONFIG_ARM64_64K_PAGES) && defined(CONFIG_KUSER_HELPERS)
> -- 
> 2.34.1
>