Re: [PATCH V9 21/21] csky: support dword access for get_user_size()

From: Arnd Bergmann
Date: Wed Oct 24 2018 - 03:18:09 EST


On Thu, Oct 18, 2018 at 9:57 AM Guo Ren <ren_guo@xxxxxxxxx> wrote:
>
> On Thu, Oct 18, 2018 at 10:34:00AM +0200, Arnd Bergmann wrote:
> > On Thu, Oct 18, 2018 at 5:41 AM Guo Ren <ren_guo@xxxxxxxxx> wrote:
> > >
> > > On Wed, Oct 17, 2018 at 05:44:17PM +0200, Arnd Bergmann wrote:
> > > > On Tue, Oct 16, 2018 at 5:33 AM Guo Ren <ren_guo@xxxxxxxxx> wrote:
> > > > >
> > > > > Support dword access for get_user_size and redesign put_user_size with
> > > > > the same style of get_user_size. It's Ok to use xxx_user_asm_common for
> > > > > all size of variable with ldb, ldh, ldw, ld.d
> > > > >
> > > > > ld.d rx, (addr, 0) could "rx <= addr" "and r(x+1) <= addr+4" and this also
> > > > > follow abiv2 gcc ABI for dword access.
> > > >
> > > > Are you sure this is correct for this?
> > > >
> > > > static inline u32 get_64_to_32(__u64 __user *p)
> > > > {
> > > > u32 ret;
> > > > get_user(ret, p);
> > > > return ret;
> > > > }
> > > >
> > > > If I read __get_user_asm_common() right, the ld.d would overwrite
> > > > two registers, but the caller only expects one, so it clobbers one
> > > > that might be in use.
> > > Ah... BUG! I only consider the get_user(u64, u64 *) :P
> > >
> > > Change to:
> > > case 8: \
> > > __get_user_asm_dword((x), ptr, "ld.d", retval); \
> > > break;
> > >
> > > #define __get_user_asm_dword(x, ptr, err) \
> > > do { \
> > > u64 tmp; \
> > > __get_user_asm_common(tmp, ptr, "ld.d", err); \
> > > x = typeof(x) tmp; \
> > > } while(0)
> > >
> > > #define __put_user_asm_dword(x, ptr, err) \
> > > do { \
> > > u64 tmp = (u64) x; \
> > > __put_user_asm_common(tmp, ptr, "st.d", err); \
> > > } while(0)
> > >
> >
> > I think this will cause warnings for code that passes a pointer.
> >
> > The 64-bit __get_user() is really hard, and most 32-bit architectures don't
> > implement it at all. If you really want to add it, have a look at what
> > x86 and arm do. IIRC they both use __builtin_choose_expr(),
> Thx for the tips and I'll drop the patch first for the upstream.
>
> I want to implement it because of make allmodconfig and
> drivers/android/binder.c need it. I'll learn __builtin_choose_expr()
> and prepare patch next.

I think we should fix binder.c instead. Many other architectures
have the same problem.

Arnd