Commit 865c50e1d279 ("x86/uaccess: utilize CONFIG_CC_HAS_ASM_GOTO_OUTPUT")
added an optimised version of __get_user_asm() for x86 using 'asm goto'.
Like the non-optimised code, the 32-bit implementation of 64-bit get_user()
expands to a pair of 32-bit accesses. Unlike the non-optimised code, the
_original_ pointer is incremented to copy the high word instead of loading
through a new pointer explicitly constructed to point at a 32-bit type.
Consequently, if the pointer points at a 64-bit type then we end up
loading the wrong data for the upper 32-bits.
This was observed as a mount() failure in Android targetting i686 after
b0cfcdd9b967 ("d_path: make 'prepend()' fill up the buffer exactly on
overflow") because the call to copy_from_kernel_nofault() from
prepend_copy() ends up in __get_kernel_nofault() and casts the source
pointer to a 'u64 __user *'. An attempt to mount at "/debug_ramdisk"
therefore ends up failing trying to mount "/debumdismdisk".
Use the existing '__gu_ptr' source pointer to unsigned int for 32-bit
__get_user_asm_u64() instead of the original pointer.
Cc: Nick Desaulniers <[email protected]>
Cc: Bill Wendling <[email protected]>
Cc: Linus Torvalds <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Borislav Petkov <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Reported-by: Greg Kroah-Hartman <[email protected]>
Fixes: 865c50e1d279 ("x86/uaccess: utilize CONFIG_CC_HAS_ASM_GOTO_OUTPUT")
Signed-off-by: Will Deacon <[email protected]>
---
arch/x86/include/asm/uaccess.h | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/arch/x86/include/asm/uaccess.h b/arch/x86/include/asm/uaccess.h
index c9fa7be3df82..5c95d242f38d 100644
--- a/arch/x86/include/asm/uaccess.h
+++ b/arch/x86/include/asm/uaccess.h
@@ -301,8 +301,8 @@ do { \
unsigned int __gu_low, __gu_high; \
const unsigned int __user *__gu_ptr; \
__gu_ptr = (const void __user *)(ptr); \
- __get_user_asm(__gu_low, ptr, "l", "=r", label); \
- __get_user_asm(__gu_high, ptr+1, "l", "=r", label); \
+ __get_user_asm(__gu_low, __gu_ptr, "l", "=r", label); \
+ __get_user_asm(__gu_high, __gu_ptr+1, "l", "=r", label); \
(x) = ((unsigned long long)__gu_high << 32) | __gu_low; \
} while (0)
#else
--
2.33.0.309.g3052b89438-goog
On Mon, Sep 13, 2021 at 9:36 AM Will Deacon <[email protected]> wrote:
>
> Commit 865c50e1d279 ("x86/uaccess: utilize CONFIG_CC_HAS_ASM_GOTO_OUTPUT")
> added an optimised version of __get_user_asm() for x86 using 'asm goto'.
>
> Like the non-optimised code, the 32-bit implementation of 64-bit get_user()
> expands to a pair of 32-bit accesses. Unlike the non-optimised code, the
> _original_ pointer is incremented to copy the high word instead of loading
> through a new pointer explicitly constructed to point at a 32-bit type.
> Consequently, if the pointer points at a 64-bit type then we end up
> loading the wrong data for the upper 32-bits.
>
> This was observed as a mount() failure in Android targetting i686 after
s/targetting/targeting/
> b0cfcdd9b967 ("d_path: make 'prepend()' fill up the buffer exactly on
> overflow") because the call to copy_from_kernel_nofault() from
> prepend_copy() ends up in __get_kernel_nofault() and casts the source
> pointer to a 'u64 __user *'. An attempt to mount at "/debug_ramdisk"
> therefore ends up failing trying to mount "/debumdismdisk".
>
> Use the existing '__gu_ptr' source pointer to unsigned int for 32-bit
> __get_user_asm_u64() instead of the original pointer.
>
> Cc: Nick Desaulniers <[email protected]>
> Cc: Bill Wendling <[email protected]>
> Cc: Linus Torvalds <[email protected]>
> Cc: Thomas Gleixner <[email protected]>
> Cc: Ingo Molnar <[email protected]>
> Cc: Borislav Petkov <[email protected]>
> Cc: Peter Zijlstra <[email protected]>
> Reported-by: Greg Kroah-Hartman <[email protected]>
> Fixes: 865c50e1d279 ("x86/uaccess: utilize CONFIG_CC_HAS_ASM_GOTO_OUTPUT")
> Signed-off-by: Will Deacon <[email protected]>
Sorry I missed this; I think -Wunused-variable would have helped here.
Thanks for debugging+fixing.
Reviewed-by: Nick Desaulniers <[email protected]>
> ---
> arch/x86/include/asm/uaccess.h | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/include/asm/uaccess.h b/arch/x86/include/asm/uaccess.h
> index c9fa7be3df82..5c95d242f38d 100644
> --- a/arch/x86/include/asm/uaccess.h
> +++ b/arch/x86/include/asm/uaccess.h
> @@ -301,8 +301,8 @@ do { \
> unsigned int __gu_low, __gu_high; \
> const unsigned int __user *__gu_ptr; \
> __gu_ptr = (const void __user *)(ptr); \
> - __get_user_asm(__gu_low, ptr, "l", "=r", label); \
> - __get_user_asm(__gu_high, ptr+1, "l", "=r", label); \
> + __get_user_asm(__gu_low, __gu_ptr, "l", "=r", label); \
> + __get_user_asm(__gu_high, __gu_ptr+1, "l", "=r", label); \
> (x) = ((unsigned long long)__gu_high << 32) | __gu_low; \
> } while (0)
> #else
> --
> 2.33.0.309.g3052b89438-goog
>
--
Thanks,
~Nick Desaulniers
On Mon, Sep 13, 2021 at 9:36 AM Will Deacon <[email protected]> wrote:
>
> Use the existing '__gu_ptr' source pointer to unsigned int for 32-bit
> __get_user_asm_u64() instead of the original pointer.
Heh. And by "existing" you mean "the one that exists _purely_ for this
exact reason and that wasn't used" ;)
What a silly bug. And it's existed for a year, which I think shows
just how little 32-bit x86 is used these days (*).
Linus
(*) but also probably how few 64-bit user accesses we do - the fact
that the bug was actually found by the "copy_from_kernel_nofault" code
that just shares the infrastructure on x86 rather than any user access
code is kind of interesting.
On Mon, Sep 13, 2021 at 05:35:47PM +0100, Will Deacon wrote:
> Commit 865c50e1d279 ("x86/uaccess: utilize CONFIG_CC_HAS_ASM_GOTO_OUTPUT")
> added an optimised version of __get_user_asm() for x86 using 'asm goto'.
>
> Like the non-optimised code, the 32-bit implementation of 64-bit get_user()
> expands to a pair of 32-bit accesses. Unlike the non-optimised code, the
> _original_ pointer is incremented to copy the high word instead of loading
> through a new pointer explicitly constructed to point at a 32-bit type.
> Consequently, if the pointer points at a 64-bit type then we end up
> loading the wrong data for the upper 32-bits.
>
> This was observed as a mount() failure in Android targetting i686 after
> b0cfcdd9b967 ("d_path: make 'prepend()' fill up the buffer exactly on
> overflow") because the call to copy_from_kernel_nofault() from
> prepend_copy() ends up in __get_kernel_nofault() and casts the source
> pointer to a 'u64 __user *'. An attempt to mount at "/debug_ramdisk"
> therefore ends up failing trying to mount "/debumdismdisk".
>
> Use the existing '__gu_ptr' source pointer to unsigned int for 32-bit
> __get_user_asm_u64() instead of the original pointer.
>
> Cc: Nick Desaulniers <[email protected]>
> Cc: Bill Wendling <[email protected]>
> Cc: Linus Torvalds <[email protected]>
> Cc: Thomas Gleixner <[email protected]>
> Cc: Ingo Molnar <[email protected]>
> Cc: Borislav Petkov <[email protected]>
> Cc: Peter Zijlstra <[email protected]>
> Reported-by: Greg Kroah-Hartman <[email protected]>
> Fixes: 865c50e1d279 ("x86/uaccess: utilize CONFIG_CC_HAS_ASM_GOTO_OUTPUT")
> Signed-off-by: Will Deacon <[email protected]>
> ---
> arch/x86/include/asm/uaccess.h | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/include/asm/uaccess.h b/arch/x86/include/asm/uaccess.h
> index c9fa7be3df82..5c95d242f38d 100644
> --- a/arch/x86/include/asm/uaccess.h
> +++ b/arch/x86/include/asm/uaccess.h
> @@ -301,8 +301,8 @@ do { \
> unsigned int __gu_low, __gu_high; \
> const unsigned int __user *__gu_ptr; \
> __gu_ptr = (const void __user *)(ptr); \
> - __get_user_asm(__gu_low, ptr, "l", "=r", label); \
> - __get_user_asm(__gu_high, ptr+1, "l", "=r", label); \
> + __get_user_asm(__gu_low, __gu_ptr, "l", "=r", label); \
> + __get_user_asm(__gu_high, __gu_ptr+1, "l", "=r", label); \
> (x) = ((unsigned long long)__gu_high << 32) | __gu_low; \
> } while (0)
> #else
> --
> 2.33.0.309.g3052b89438-goog
>
Tested-by: Greg Kroah-Hartman <[email protected]>
Thanks for finding and fixing this!
greg k-h
On Mon, 13 Sept 2021 at 22:06, Will Deacon <[email protected]> wrote:
>
> Commit 865c50e1d279 ("x86/uaccess: utilize CONFIG_CC_HAS_ASM_GOTO_OUTPUT")
> added an optimised version of __get_user_asm() for x86 using 'asm goto'.
>
> Like the non-optimised code, the 32-bit implementation of 64-bit get_user()
> expands to a pair of 32-bit accesses. Unlike the non-optimised code, the
> _original_ pointer is incremented to copy the high word instead of loading
> through a new pointer explicitly constructed to point at a 32-bit type.
> Consequently, if the pointer points at a 64-bit type then we end up
> loading the wrong data for the upper 32-bits.
>
> This was observed as a mount() failure in Android targetting i686 after
> b0cfcdd9b967 ("d_path: make 'prepend()' fill up the buffer exactly on
> overflow") because the call to copy_from_kernel_nofault() from
> prepend_copy() ends up in __get_kernel_nofault() and casts the source
> pointer to a 'u64 __user *'. An attempt to mount at "/debug_ramdisk"
> therefore ends up failing trying to mount "/debumdismdisk".
>
> Use the existing '__gu_ptr' source pointer to unsigned int for 32-bit
> __get_user_asm_u64() instead of the original pointer.
>
> Cc: Nick Desaulniers <[email protected]>
> Cc: Bill Wendling <[email protected]>
> Cc: Linus Torvalds <[email protected]>
> Cc: Thomas Gleixner <[email protected]>
> Cc: Ingo Molnar <[email protected]>
> Cc: Borislav Petkov <[email protected]>
> Cc: Peter Zijlstra <[email protected]>
> Reported-by: Greg Kroah-Hartman <[email protected]>
> Fixes: 865c50e1d279 ("x86/uaccess: utilize CONFIG_CC_HAS_ASM_GOTO_OUTPUT")
> Signed-off-by: Will Deacon <[email protected]>
This patch merged into Linux next (next-20210914) and the reported
problem has been fixed.
Tested-by: Linux Kernel Functional Testing <[email protected]>
--
Linaro LKFT
https://lkft.linaro.org