Re: [PATCH v2 09/10] x86/percpu: Clean up percpu_stable_op()

From: Peter Zijlstra <peterz@infradead.org>
To: Brian Gerst <brgerst@gmail.com>
Cc: Nick Desaulniers <ndesaulniers@google.com>,
	LKML <linux-kernel@vger.kernel.org>,
	"maintainer:X86 ARCHITECTURE (32-BIT AND 64-BIT)"
	<x86@kernel.org>, Thomas Gleixner <tglx@linutronix.de>,
	Ingo Molnar <mingo@kernel.org>, Borislav Petkov <bp@alien8.de>,
	"H . Peter Anvin" <hpa@zytor.com>,
	Andy Lutomirski <luto@kernel.org>
Subject: Re: [PATCH v2 09/10] x86/percpu: Clean up percpu_stable_op()
Date: Thu, 9 Jul 2020 12:25:50 +0200	[thread overview]
Message-ID: <20200709102550.GN597537@hirez.programming.kicks-ass.net> (raw)
In-Reply-To: <CAMzpN2hHSeN9Z3BO_Uo1-yCfh9YdEbkK0wb9E3D8rq6onQDZdg@mail.gmail.com>

On Tue, Jun 02, 2020 at 10:19:51AM -0400, Brian Gerst wrote:
> On Mon, Jun 1, 2020 at 4:43 PM Nick Desaulniers <ndesaulniers@google.com> wrote:
> >
> > On Sat, May 30, 2020 at 3:11 PM Brian Gerst <brgerst@gmail.com> wrote:
> > >
> > > Use __pcpu_size_call_return() to simplify this_cpu_read_stable().
> >
> > Clever! As in this_cpu_read() in include/linux/percpu-defs.h.  Could
> > be its own patch before this, but it's fine.
> > Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>
> >
> > > Also remove __bad_percpu_size() which is now unused.
> > >
> > > Signed-off-by: Brian Gerst <brgerst@gmail.com>
> > > ---
> > >  arch/x86/include/asm/percpu.h | 41 ++++++++++-------------------------
> > >  1 file changed, 12 insertions(+), 29 deletions(-)
> > >
> > > diff --git a/arch/x86/include/asm/percpu.h b/arch/x86/include/asm/percpu.h
> > > index 7efc0b5c4ff0..cf2b9c2a241e 100644
> > > --- a/arch/x86/include/asm/percpu.h
> > > +++ b/arch/x86/include/asm/percpu.h
> > > @@ -85,7 +85,6 @@
> > >
> > >  /* For arch-specific code, we can use direct single-insn ops (they
> > >   * don't give an lvalue though). */
> > > -extern void __bad_percpu_size(void);
> > >
> > >  #define __pcpu_type_1 u8
> > >  #define __pcpu_type_2 u16
> > > @@ -167,33 +166,13 @@ do {                                                                      \
> > >         (typeof(_var))(unsigned long) pfo_val__;                        \
> > >  })
> > >
> > > -#define percpu_stable_op(op, var)                      \
> > > -({                                                     \
> > > -       typeof(var) pfo_ret__;                          \
> > > -       switch (sizeof(var)) {                          \
> > > -       case 1:                                         \
> > > -               asm(op "b "__percpu_arg(P1)",%0"        \
> >
> > What does the `P` do here?
> > https://gcc.gnu.org/onlinedocs/gcc/Simple-Constraints.html#Simple-Constraints
> > says can be machine dependent integral literal in a certain range.
> > https://gcc.gnu.org/onlinedocs/gcc/Machine-Constraints.html#Machine-Constraints
> > doesn't document `P` for x86 though...
> 
> https://gcc.gnu.org/onlinedocs/gcc-10.1.0/gcc/Extended-Asm.html#x86-Operand-Modifiers
> 
> Removing the 'P' modifier results in this:
>         movq %gs:$current_task, %rdx    #, pfo_val__
> 
> This is because the 'p' constraint treats a memory address as a
> constant.  I tried replacing it with __this_cpu_read(), which since
> commit 0b9ccc0a should have similar non-volatile semantics.  But the
> compiler still reloaded it on every use, so I left the asm template
> as-is for now until that can be resolved.

Right, I can into that same issue a while back and gave up staring at
compiler innards. __this_cpu_read() *should* allow re-loads, and it does
in places (we've had bugs because of it), but this_cpu_read_stable() is
somehow far more 'effective'.

It would be good if someone can update the comment with that thing, to
explain matters better.