Re: [PATCH] linux-user/syscall: Silence warning from the undefined behavior sanitizer

[Thomas Huth]

On 12/02/2021 01.35, Laurent Vivier wrote:
> Le 11/02/2021 à 22:28, Laurent Vivier a écrit :
> > Le 11/02/2021 à 14:29, Thomas Huth a écrit :
> > > When compiling QEMU with -fsanitize=undefined, there is a warning when
> > > running "make check-tcg":
> > >
> > >    TEST    linux-test on m68k
> > > ../linux-user/syscall.c:10499:34: runtime error: member access within
> > >   misaligned address 0x00008006df3c for type 'struct linux_dirent64',
> > >   which requires 8 byte alignment
> > > 0x00008006df3c: note: pointer points here
> > >    00 00 00 00 68 03 28 00  00 00 00 00 5b 96 3e e4  61 4b 05 26 18 00 04 2e  
> > > 00 00 00 00 da 3f 18 00
> > >                ^
> > >
> > > It's likely not an issue in reality, since I assume that on hosts where
> > > the alignment really matters (like sparc64), the Linux kernel likely
> > > adds the right padding. Anyway, let's use the stw_p() / stq_p() accessor
> > > helpers here to silence the warning and thus to allow to compile the code
> > > with -fsanitize=undefined, too.
>
> Wait... if the alignment differs between m68k and the host, I guess the size of 
> the structure differs?

No, as far as I understand, the size and layout of the structure are the 
same on all architecture. The problem is that the "dirp = 
lock_user(VERIFY_WRITE, arg2, count, 0)" already ends up in a pointer that 
is only aligned to a 4-byte boundary. Since the m68k code is only restricted 
to a 4-byte alignment, arg2 is only aligned to 4 bytes. But if the host 
needs 8-byte alignment for the struct, we've certainly lost here.

Having said that, I think my patch is still wrong. It silences the ubsan 
warnings, but it won't fix the problem when the code is e.g. running on a 
sparc64 host. There we likely need to make sure that the buffer on the host 
is already aligned to an 8-byte boundary when doing the sys_getdents64() 
call to the host kernel.
So I guess we need a bounce buffer here anyway?

 Thomas