On 17/09/2023 23:09, David Laight wrote:
> From: Clément Léger
>> Sent: 14 September 2023 14:11
>>
>> enabler->uaddr can be aligned on 32 or 64 bits. If aligned on 32 bits,
>> this will result in a misaligned access on 64 bits architectures since
>> set_bit()/clear_bit() are expecting an unsigned long (aligned) pointer.
>> On architecture that do not support misaligned access, this will crash
>> the kernel. Align uaddr on unsigned long size to avoid such behavior.
>> This bug was found while running kselftests on RISC-V.
>
> You don't want to do it on x86-64 either.
> A locked accesses that crosses a cache line boundary is horrid.
> So horrid that recent cpu's can be made to fault.
Hi David,
Thanks for the additional information.
>
> I'd also doubt that other cpu that can do misaligned transfers
> can even do locked ones.
>
> For x86 (and LE) the long[] bitmap can be treated as char[]
> avoiding all the problems.
>
> Perhaps there ought to be bit a bit-array based on char[]
> (not long[]) that would be endianness independent and
> use byte-sized atomics.
That would work for a few architectures but I don't think all of them
have byte "grain" atomics. So I guess Masami solution (long aligned
set/clear_bit()) remains the best out there.
Clément
> (IIRC that is still an issue on sparc32...)
>
> David
>
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1
> 1PT, UK
> Registration No: 1397386 (Wales)