Re: [PATCH V2 11/11] x86/rcu: Add THUNK rcu_read_unlock_special_thunk

From: Lai Jiangshan
Date: Tue Apr 23 2024 - 22:43:29 EST


On Wed, Apr 24, 2024 at 1:26 AM Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
>
> Hello Lai,
>
> On Sun, Apr 7, 2024 at 5:07 AM Lai Jiangshan <jiangshanlai@xxxxxxxxx> wrote:
> >
> > From: Lai Jiangshan <jiangshan.ljs@xxxxxxxxxxxx>
> >
> > Add rcu_read_unlock_special_thunk(), so that the inlined rcu_read_unlock()
> > doesn't need any code to save the caller-saved registers.
> >
> > Make rcu_read_unlock() only two instructions in the slow path at the
> > caller site.
> >
> > Cc: "Paul E. McKenney" <paulmck@xxxxxxxxxx>
> > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > Cc: Frederic Weisbecker <frederic@xxxxxxxxxx>
> > Signed-off-by: Lai Jiangshan <jiangshan.ljs@xxxxxxxxxxxx>
> > ---
> > arch/x86/entry/thunk.S | 5 +++++
> > arch/x86/include/asm/rcu_preempt.h | 4 +++-
> > 2 files changed, 8 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/x86/entry/thunk.S b/arch/x86/entry/thunk.S
> > index 119ebdc3d362..10c60369a67c 100644
> > --- a/arch/x86/entry/thunk.S
> > +++ b/arch/x86/entry/thunk.S
> > @@ -13,3 +13,8 @@ THUNK preempt_schedule_thunk, preempt_schedule
> > THUNK preempt_schedule_notrace_thunk, preempt_schedule_notrace
> > EXPORT_SYMBOL(preempt_schedule_thunk)
> > EXPORT_SYMBOL(preempt_schedule_notrace_thunk)
> > +
> > +#ifdef CONFIG_PCPU_RCU_PREEMPT_COUNT
> > +THUNK rcu_read_unlock_special_thunk, rcu_read_unlock_special
> > +EXPORT_SYMBOL_GPL(rcu_read_unlock_special_thunk)
> > +#endif /* #ifdef CONFIG_PCPU_RCU_PREEMPT_COUNT */
> > diff --git a/arch/x86/include/asm/rcu_preempt.h b/arch/x86/include/asm/rcu_preempt.h
> > index cb25ebe038a5..acdd73b74c05 100644
> > --- a/arch/x86/include/asm/rcu_preempt.h
> > +++ b/arch/x86/include/asm/rcu_preempt.h
> > @@ -97,9 +97,11 @@ static __always_inline bool pcpu_rcu_preempt_count_dec_and_test(void)
> > __percpu_arg([var]));
> > }
> >
> > +extern asmlinkage void rcu_read_unlock_special_thunk(void);
> > +
> > #define pcpu_rcu_read_unlock_special() \
> > do { \
> > - rcu_read_unlock_special();
>
> Instead, can you not use __no_caller_saved_registers attribute for
> definition of rcu_read_unlock_special() or does that not work for what
> you're trying to do here?
>

I think it is paramount to make it the same as preempt_schedule[_thunk]()
when it comes to avoiding the caller-saving-registers-procedures in the
unlock paths.

I had no idea of 'no_caller_saved_registers' before, so I haven't tried it.

And there are limits to 'no_caller_saved_registers' either:

https://clang.llvm.org/docs/AttributeReference.html#no-caller-saved-registers:

Functions specified with the ‘no_caller_saved_registers’ attribute
should only call other functions with the ‘no_caller_saved_registers’
attribute, or should be compiled with the ‘-mgeneral-regs-only’ flag
to avoid saving unused non-GPR registers.

https://gcc.gnu.org/onlinedocs/gcc/x86-Function-Attributes.html#index-no_005fcaller_005fsaved_005fregisters-function-attribute_002c-x86:

Since GCC doesn’t preserve SSE, MMX nor x87 states, the GCC option
-mgeneral-regs-only should be used to compile functions with
no_caller_saved_registers attribute.

And I don't think ‘-mgeneral-regs-only’ is being used in the kernel for x86.

Thanks
Lai