On 2024-03-12 9:53 PM, Charlie Jenkins wrote:
> On Mon, Mar 11, 2024 at 07:19:13PM -0700, Samuel Holland wrote:
>> These macros did not initialize __kr_err, so they could fail even if
>> the access did not fault.
>>
>> Cc: stable@xxxxxxxxxxxxxxx
>> Fixes: d464118cdc41 ("riscv: implement __get_kernel_nofault and __put_user_nofault")
>> Signed-off-by: Samuel Holland <samuel.holland@xxxxxxxxxx>
>> ---
>> Found while testing the unaligned access speed series[1]. The observed
>> behavior was that with RISCV_EFFICIENT_UNALIGNED_ACCESS=y, the
>> copy_from_kernel_nofault() in prepend_copy() failed every time when
>> filling out /proc/self/mounts, so all of the mount points were "xxx".
>>
>> I'm surprised this hasn't been seen before. For reference, I'm compiling
>> with clang 18.
>>
>> [1]: https://lore.kernel.org/linux-riscv/20240308-disable_misaligned_probe_config-v9-0-a388770ba0ce@xxxxxxxxxxxx/
>>
>> arch/riscv/include/asm/uaccess.h | 4 ++--
>> 1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/arch/riscv/include/asm/uaccess.h b/arch/riscv/include/asm/uaccess.h
>> index ec0cab9fbddd..72ec1d9bd3f3 100644
>> --- a/arch/riscv/include/asm/uaccess.h
>> +++ b/arch/riscv/include/asm/uaccess.h
>> @@ -319,7 +319,7 @@ unsigned long __must_check clear_user(void __user *to, unsigned long n)
>>
>> #define __get_kernel_nofault(dst, src, type, err_label) \
>> do { \
>> - long __kr_err; \
>> + long __kr_err = 0; \
>> \
>> __get_user_nocheck(*((type *)(dst)), (type *)(src), __kr_err); \
>> if (unlikely(__kr_err)) \
>> @@ -328,7 +328,7 @@ do { \
>>
>> #define __put_kernel_nofault(dst, src, type, err_label) \
>> do { \
>> - long __kr_err; \
>> + long __kr_err = 0; \
>> \
>> __put_user_nocheck(*((type *)(src)), (type *)(dst), __kr_err); \
>> if (unlikely(__kr_err)) \
>> --
>> 2.43.1
>>
>>
>> _______________________________________________
>> linux-riscv mailing list
>> linux-riscv@xxxxxxxxxxxxxxxxxxx
>> http://lists.infradead.org/mailman/listinfo/linux-riscv
>
> I am not able to reproduce this using Clang 18 with
> RISCV_EFFICIENT_UNALIGNED_ACCESS=y on 6.8. However I can see how this
> could be an issue.
>
> Going down the rabbit hold of macros here, I end up at
> arch/riscv/include/asm/asm-extable.h where the register that hold 'err'
> is written into the __ex_table section:
>
> #define EX_DATA_REG(reg, gpr) \
> "((.L__gpr_num_" #gpr ") << " __stringify(EX_DATA_REG_##reg##_SHIFT) ")"
>
> #define _ASM_EXTABLE_UACCESS_ERR_ZERO(insn, fixup, err, zero) \
> __DEFINE_ASM_GPR_NUMS \
> __ASM_EXTABLE_RAW(#insn, #fixup, \
> __stringify(EX_TYPE_UACCESS_ERR_ZERO), \
> "(" \
> EX_DATA_REG(ERR, err) " | " \
> EX_DATA_REG(ZERO, zero) \
> ")")
>
> I am wondering if setting this value to zero solves the problem by
> hiding another issue. It seems like this shouldn't need to be
> initialized to zero, however I am lost as to how this extable setup
> works so perhaps this is the proper solution.
extable works by running the handler (selected by EX_TYPE_*) if some exception
occurs while executing that instruction -- see the calls to fixup_exception() in
fault.c and traps.c. If there is no exception, then the handler does not run,
and the err register is not written by ex_handler_uaccess_err_zero().
If you look at __get_user_asm(), you can see that the err register is not
touched by the assembly code at all -- the only reference to %0 is in the
extable entry. So if the macro that declares the error variable doesn't
initialize it, nothing will.
Compare __get_user() and __put_user() which do initialize their error variable.
Regards,
Samuel
