Re: [PATCH 2/7] x86/percpu: Clean up percpu_to_op()

From: Nick Desaulniers
Date: Tue May 26 2020 - 13:54:49 EST


On Thu, May 21, 2020 at 6:06 AM Brian Gerst <brgerst@xxxxxxxxx> wrote:
>
> On Wed, May 20, 2020 at 1:26 PM Nick Desaulniers
> <ndesaulniers@xxxxxxxxxx> wrote:
> >
> > On Mon, May 18, 2020 at 8:38 PM Brian Gerst <brgerst@xxxxxxxxx> wrote:
> > >
> > > On Mon, May 18, 2020 at 5:15 PM Nick Desaulniers
> > > <ndesaulniers@xxxxxxxxxx> wrote:
> > > >
> > > > On Sun, May 17, 2020 at 8:29 AM Brian Gerst <brgerst@xxxxxxxxx> wrote:
> > > > >
> > > > > The core percpu macros already have a switch on the data size, so the switch
> > > > > in the x86 code is redundant and produces more dead code.
> > > > >
> > > > > Also use appropriate types for the width of the instructions. This avoids
> > > > > errors when compiling with Clang.
> > > > >
> > > > > Signed-off-by: Brian Gerst <brgerst@xxxxxxxxx>
> > > > > ---
> > > > > arch/x86/include/asm/percpu.h | 90 ++++++++++++++---------------------
> > > > > 1 file changed, 35 insertions(+), 55 deletions(-)
> > > > >
> > > > > diff --git a/arch/x86/include/asm/percpu.h b/arch/x86/include/asm/percpu.h
> > > > > index 89f918a3e99b..233c7a78d1a6 100644
> > > > > --- a/arch/x86/include/asm/percpu.h
> > > > > +++ b/arch/x86/include/asm/percpu.h
> > > > > @@ -117,37 +117,17 @@ extern void __bad_percpu_size(void);
> > > > > #define __pcpu_reg_imm_4(x) "ri" (x)
> > > > > #define __pcpu_reg_imm_8(x) "re" (x)
> > > > >
> > > > > -#define percpu_to_op(qual, op, var, val) \
> > > > > -do { \
> > > > > - typedef typeof(var) pto_T__; \
> > > > > - if (0) { \
> > > > > - pto_T__ pto_tmp__; \
> > > > > - pto_tmp__ = (val); \
> > > > > - (void)pto_tmp__; \
> > > > > - } \
> > > > > - switch (sizeof(var)) { \
> > > > > - case 1: \
> > > > > - asm qual (op "b %1,"__percpu_arg(0) \
> > > > > - : "+m" (var) \
> > > > > - : "qi" ((pto_T__)(val))); \
> > > > > - break; \
> > > > > - case 2: \
> > > > > - asm qual (op "w %1,"__percpu_arg(0) \
> > > > > - : "+m" (var) \
> > > > > - : "ri" ((pto_T__)(val))); \
> > > > > - break; \
> > > > > - case 4: \
> > > > > - asm qual (op "l %1,"__percpu_arg(0) \
> > > > > - : "+m" (var) \
> > > > > - : "ri" ((pto_T__)(val))); \
> > > > > - break; \
> > > > > - case 8: \
> > > > > - asm qual (op "q %1,"__percpu_arg(0) \
> > > > > - : "+m" (var) \
> > > > > - : "re" ((pto_T__)(val))); \
> > > > > - break; \
> > > > > - default: __bad_percpu_size(); \
> > > > > - } \
> > > > > +#define percpu_to_op(size, qual, op, _var, _val) \
> > > > > +do { \
> > > > > + __pcpu_type_##size pto_val__ = __pcpu_cast_##size(_val); \
> > > > > + if (0) { \
> > > > > + typeof(_var) pto_tmp__; \
> > > > > + pto_tmp__ = (_val); \
> > > > > + (void)pto_tmp__; \
> > > > > + } \
> > > >
> > > > Please replace the whole `if (0)` block with:
> > > > ```c
> > > > __same_type(_var, _val);
> > > > ```
> > > > from include/linux/compiler.h.
> > >
> > > The problem with __builtin_types_compatible_p() is that it considers
> > > unsigned long and u64 (aka unsigned long long) as different types even
> > > though they are the same width on x86-64. While this may be a good
> > > cleanup to look at in the future, it's not a simple drop-in
> > > replacement.
> >
> > Does it trigger errors in this case?
>
> Yes, see boot_init_stack_canary(). That code looks a bit sketchy but
> it's not wrong, for x86-64 at least.
>
> It also doesn't seem to like "void *" compared to any other pointer type:
>
> In function âfpregs_deactivateâ,
> inlined from âfpu__dropâ at arch/x86/kernel/fpu/core.c:285:3:
> ./include/linux/compiler.h:379:38: error: call to
> â__compiletime_assert_317â declared with attribute error: BUILD_BUG_ON
> failed: !__same_type((fpu_fpregs_owner_ctx), ((void *)0))
> 379 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__)
> | ^
> ./include/linux/compiler.h:360:4: note: in definition of macro
> â__compiletime_assertâ
> 360 | prefix ## suffix(); \
> | ^~~~~~
> ./include/linux/compiler.h:379:2: note: in expansion of macro
> â_compiletime_assertâ
> 379 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__)
> | ^~~~~~~~~~~~~~~~~~~
> ./include/linux/build_bug.h:39:37: note: in expansion of macro
> âcompiletime_assertâ
> 39 | #define BUILD_BUG_ON_MSG(cond, msg) compiletime_assert(!(cond), msg)
> | ^~~~~~~~~~~~~~~~~~
> ./include/linux/build_bug.h:50:2: note: in expansion of macro âBUILD_BUG_ON_MSGâ
> 50 | BUILD_BUG_ON_MSG(condition, "BUILD_BUG_ON failed: " #condition)
> | ^~~~~~~~~~~~~~~~
> ./arch/x86/include/asm/percpu.h:105:2: note: in expansion of macro
> âBUILD_BUG_ONâ
> 105 | BUILD_BUG_ON(!__same_type(_var, _val)); \
> | ^~~~~~~~~~~~
> ./arch/x86/include/asm/percpu.h:338:37: note: in expansion of macro
> âpercpu_to_opâ
> 338 | #define this_cpu_write_8(pcp, val) percpu_to_op(8, volatile,
> "mov", (pcp), val)
> | ^~~~~~~~~~~~
> ./include/linux/percpu-defs.h:380:11: note: in expansion of macro
> âthis_cpu_write_8â
> 380 | case 8: stem##8(variable, __VA_ARGS__);break; \
> | ^~~~
> ./include/linux/percpu-defs.h:508:34: note: in expansion of macro
> â__pcpu_size_callâ
> 508 | #define this_cpu_write(pcp, val)
> __pcpu_size_call(this_cpu_write_, pcp, val)
> | ^~~~~~~~~~~~~~~~
> ./arch/x86/include/asm/fpu/internal.h:525:2: note: in expansion of
> macro âthis_cpu_writeâ
> 525 | this_cpu_write(fpu_fpregs_owner_ctx, NULL);
> | ^~~~~~~~~~~~~~
>
> >
> > It's interesting to know how this trick differs from
> > __builtin_types_compatible_p(). Might even be helpful to wrap this
> > pattern in a macro with a comment with the pros/cons of this approach
> > vs __same_type.
>
> I think the original code is more to catch a mismatch between pointers
> and integers. It doesn't seem to care about truncation
>
> > On the other hand, the use of `long` seems tricky in x86 code as x86
> > (32b) is ILP32 but x86_64 (64b) is LP64. So the use of `long` is
> > ambiguous in the sense that it's a different size depending on the
> > target ABI. Wouldn't it potentially be a bug for x86 kernel code to
> > use `long` percpu variables (or rather mix `long` and `long long` in
> > the same operation) in that case, since the sizes of the two would be
> > different for i386?
>
> Not necessarily. Some things like registers are naturally 32-bit on a
> 32-bit kernel and 64-bit on a 64-bit kernel, so 'long' is appropriate
> there.

Sorry for not getting back to this sooner, amazing how fast emails get
buried in an inbox.

Interesting findings. Feels almost like a _Static_assert that the
sizeof these types match might be more straightforward, but I don't
need to nitpick pre-existing code that this patch simply carries
forward. I realized I never signed off on this.

Reviewed-by: Nick Desaulniers <ndesaulniers@xxxxxxxxxx>

It looks like there's still an outstanding issue with patch 4/7?
https://lore.kernel.org/lkml/CAKwvOdnVU3kZnGzkYjEFJWMPuVjOmAHuRSB8FJ-Ks+FWzX2M_Q@xxxxxxxxxxxxxx/
--
Thanks,
~Nick Desaulniers