Re: [RFC PATCH 01/20] asm-generic/mmiowb: Add generic implementation of mmiowb() tracking

From: Will Deacon <will.deacon@arm.com>
To: Linus Torvalds <torvalds@linux-foundation.org>
Cc: linux-arch <linux-arch@vger.kernel.org>,
	Linux List Kernel Mailing <linux-kernel@vger.kernel.org>,
	"Paul E. McKenney" <paulmck@linux.ibm.com>,
	Benjamin Herrenschmidt <benh@kernel.crashing.org>,
	Michael Ellerman <mpe@ellerman.id.au>,
	Arnd Bergmann <arnd@arndb.de>,
	Peter Zijlstra <peterz@infradead.org>,
	Andrea Parri <andrea.parri@amarulasolutions.com>,
	Palmer Dabbelt <palmer@sifive.com>,
	Daniel Lustig <dlustig@nvidia.com>,
	David Howells <dhowells@redhat.com>,
	Alan Stern <stern@rowland.harvard.edu>,
	"Maciej W. Rozycki" <macro@linux-mips.org>,
	Paul Burton <paul.burton@mips.com>,
	Ingo Molnar <mingo@kernel.org>,
	Yoshinori Sato <ysato@users.sourceforge.jp>,
	Rich Felker <dalias@libc.org>, Tony Luck <tony.luck@intel.com>
Subject: Re: [RFC PATCH 01/20] asm-generic/mmiowb: Add generic implementation of mmiowb() tracking
Date: Tue, 26 Feb 2019 18:26:24 +0000	[thread overview]
Message-ID: <20190226182624.GC28709@fuggles.cambridge.arm.com> (raw)
In-Reply-To: <CAHk-=wgamK0=rNsCfDfDzmNXUF_MqUHb0okzqN1Tir9vm65pNg@mail.gmail.com>

Hi Linus,

Thanks for having a look.

On Fri, Feb 22, 2019 at 01:49:32PM -0800, Linus Torvalds wrote:
> On Fri, Feb 22, 2019 at 10:50 AM Will Deacon <will.deacon@arm.com> wrote:
> >
> > +#ifndef mmiowb_set_pending
> > +static inline void mmiowb_set_pending(void)
> > +{
> > +       __this_cpu_write(__mmiowb_state.mmiowb_pending, 1);
> > +}
> > +#endif
> > +
> > +#ifndef mmiowb_spin_lock
> > +static inline void mmiowb_spin_lock(void)
> > +{
> > +       if (__this_cpu_inc_return(__mmiowb_state.nesting_count) == 1)
> > +               __this_cpu_write(__mmiowb_state.mmiowb_pending, 0);
> > +}
> > +#endif
> 
> The case we want to go fast is the spin-lock and unlock case, not the
> "set pending" case.
> 
> And the way you implemented this, it's exactly the wrong way around.
> 
> So I'd suggest instead doing
> 
>   static inline void mmiowb_set_pending(void)
>   {
>       __this_cpu_write(__mmiowb_state.mmiowb_pending,
> __mmiowb_state.nesting_count);
>   }
> 
> and
> 
>   static inline void mmiowb_spin_lock(void)
>   {
>       __this_cpu_inc(__mmiowb_state.nesting_count);
>   }
> 
> which makes that spin-lock code much simpler and avoids the conditional there.

Makes sense; I'll hook that up for the next version.

> Then the unlock case could be something like
> 
>   static inline void mmiowb_spin_unlock(void)
>   {
>       if (unlikely(__this_cpu_read(__mmiowb_state.mmiowb_pending))) {
>           __this_cpu_write(__mmiowb_state.mmiowb_pending, 0);
>           mmiowb();
>       }
>       __this_cpu_dec(__mmiowb_state.nesting_count);
>   }
> 
> or something (xchg is generally much more expensive than read, and the
> common case for spinlocks is that nobody did IO inside of it).

So I *am* using __this_cpu_xchg() here, which means the architecture can
get away with plain old loads and stores (which is what RISC-V does, for
example), but I see that's not the case on e.g. x86 so I'll rework using
read() and write() because it doesn't hurt.

Will