Date: Tue, 14 May 2024 07:22:47 -0700
From: "Paul E. McKenney" <paulmck@...nel.org>
To: Boqun Feng <boqun.feng@...il.com>
Cc: linux-arch@...r.kernel.org, linux-kernel@...r.kernel.org,
	elver@...gle.com, akpm@...ux-foundation.org, tglx@...utronix.de,
	peterz@...radead.org, dianders@...omium.org, pmladek@...e.com,
	arnd@...db.de, torvalds@...ux-foundation.org, kernel-team@...a.com,
	Mark Rutland <mark.rutland@....com>
Subject: Re: [PATCH v2 cmpxchg 09/13] lib: Add one-byte emulation function

On Mon, May 13, 2024 at 02:19:37PM -0700, Boqun Feng wrote:
> On Mon, May 13, 2024 at 08:57:16AM -0700, Boqun Feng wrote:
> > On Mon, May 13, 2024 at 08:41:27AM -0700, Paul E. McKenney wrote:
> > [...]
> > > > > +#include <linux/types.h>
> > > > > +#include <linux/export.h>
> > > > > +#include <linux/instrumented.h>
> > > > > +#include <linux/atomic.h>
> > > > > +#include <linux/panic.h>
> > > > > +#include <linux/bug.h>
> > > > > +#include <asm-generic/rwonce.h>
> > > > > +#include <linux/cmpxchg-emu.h>
> > > > > +
> > > > > +union u8_32 {
> > > > > +	u8 b[4];
> > > > > +	u32 w;
> > > > > +};
> > > > > +
> > > > > +/* Emulate one-byte cmpxchg() in terms of 4-byte cmpxchg. */
> > > > > +uintptr_t cmpxchg_emu_u8(volatile u8 *p, uintptr_t old, uintptr_t new)
> > > > > +{
> > > > > +	u32 *p32 = (u32 *)(((uintptr_t)p) & ~0x3);
> > > > > +	int i = ((uintptr_t)p) & 0x3;
> > > > > +	union u8_32 old32;
> > > > > +	union u8_32 new32;
> > > > > +	u32 ret;
> > > > > +
> > > > > +	ret = READ_ONCE(*p32);
> > > > > +	do {
> > > > > +		old32.w = ret;
> > > > > +		if (old32.b[i] != old)
> > > > > +			return old32.b[i];
> > > > > +		new32.w = old32.w;
> > > > > +		new32.b[i] = new;
> > > > > +		instrument_atomic_read_write(p, 1);
> > > > > +		ret = data_race(cmpxchg(p32, old32.w, new32.w)); // Overridden above.
> > > > 
> > > > Just out of curiosity, why is this `data_race` needed? cmpxchg is atomic
> > > > so there should be no chance for a data race?
> > > 
> > > That is what I thought, too.  ;-)
> > > 
> > > The problem is that the cmpxchg() covers 32 bits, and so without that
> > > data_race(), KCSAN would complain about data races with perfectly
> > > legitimate concurrent accesses to the other three bytes.
> > > 
> > > The instrument_atomic_read_write(p, 1) beforehand tells KCSAN to complain
> > > about concurrent accesses, but only to that one byte.
> > > 
> > 
> > Oh, I see. For that purpose, maybe we can just use raw_cmpxchg() here,
> > i.e. a cmpxchg() without any instrument in it. Cc Mark in case I'm
> > missing something.
> > 
> 
> I just realized that the KCSAN instrumentation is already done in
> cmpxchg() layer:
> 
> 	#define cmpxchg(ptr, ...) \
> 	({ \
> 		typeof(ptr) __ai_ptr = (ptr); \
> 		kcsan_mb(); \
> 		instrument_atomic_read_write(__ai_ptr, sizeof(*__ai_ptr)); \
> 		raw_cmpxchg(__ai_ptr, __VA_ARGS__); \
> 	})
> 
> and, this function is lower in the layer, so it shouldn't have the
> instrumentation itself. How about the following (based on today's RCU
> dev branch)?

The raw_cmpxchg() looks nicer than the added data_race()!

One question below, though.

							Thanx, Paul

> Regards,
> Boqun
> 
> -------------------------------------------->8
> Subject: [PATCH] lib: cmpxchg-emu: Make cmpxchg_emu_u8() noinstr
> 
> Currently, cmpxchg_emu_u8() is called via cmpxchg() or raw_cmpxchg()
> which already makes the instrumentation decision:
> 
> * cmpxchg() case:
> 
> 	cmpxchg():
> 	  kcsan_mb();
> 	  instrument_atomic_read_write(...);
> 	  raw_cmpxchg():
> 	    arch_cmpxchg():
> 	      cmpxchg_emu_u8();
> 
> ... should have KCSAN instrumentation.
> 
> * raw_cmpxchg() case:
> 
> 	raw_cmpxchg():
> 	  arch_cmpxchg():
> 	    cmpxchg_emu_u8();
> 
> ... shouldn't have KCSAN instrumentation.
> 
> Therefore it's redundant to put KCSAN instrumentation in
> cmpxchg_emu_u8() (along with the data_race() to get away the
> instrumentation).
> 
> So make cmpxchg_emu_u8() a noinstr function, and remove the KCSAN
> instrumentation inside it.
> 
> Signed-off-by: Boqun Feng <boqun.feng@...il.com>
> ---
>  include/linux/cmpxchg-emu.h |  4 +++-
>  lib/cmpxchg-emu.c           | 14 ++++++++++----
>  2 files changed, 13 insertions(+), 5 deletions(-)
> 
> diff --git a/include/linux/cmpxchg-emu.h b/include/linux/cmpxchg-emu.h
> index 998deec67740..c4c85f41d9f4 100644
> --- a/include/linux/cmpxchg-emu.h
> +++ b/include/linux/cmpxchg-emu.h
> @@ -10,6 +10,8 @@
>  #ifndef __LINUX_CMPXCHG_EMU_H
>  #define __LINUX_CMPXCHG_EMU_H
>  
> -uintptr_t cmpxchg_emu_u8(volatile u8 *p, uintptr_t old, uintptr_t new);
> +#include <linux/compiler.h>
> +
> +noinstr uintptr_t cmpxchg_emu_u8(volatile u8 *p, uintptr_t old, uintptr_t new);
>  
>  #endif /* __LINUX_CMPXCHG_EMU_H */
> diff --git a/lib/cmpxchg-emu.c b/lib/cmpxchg-emu.c
> index 27f6f97cb60d..788c22cd4462 100644
> --- a/lib/cmpxchg-emu.c
> +++ b/lib/cmpxchg-emu.c
> @@ -21,8 +21,13 @@ union u8_32 {
>  	u32 w;
>  };
>  
> -/* Emulate one-byte cmpxchg() in terms of 4-byte cmpxchg. */
> -uintptr_t cmpxchg_emu_u8(volatile u8 *p, uintptr_t old, uintptr_t new)
> +/*
> + * Emulate one-byte cmpxchg() in terms of 4-byte cmpxchg.
> + *
> + * This function is marked as 'noinstr' as the instrumentation should be done at
> + * outer layer.
> + */
> +noinstr uintptr_t cmpxchg_emu_u8(volatile u8 *p, uintptr_t old, uintptr_t new)
>  {
>  	u32 *p32 = (u32 *)(((uintptr_t)p) & ~0x3);
>  	int i = ((uintptr_t)p) & 0x3;
> @@ -37,8 +42,9 @@ uintptr_t cmpxchg_emu_u8(volatile u8 *p, uintptr_t old, uintptr_t new)
>  			return old32.b[i];
>  		new32.w = old32.w;
>  		new32.b[i] = new;
> -		instrument_atomic_read_write(p, 1);

Don't we need to keep that instrument_atomic_read_write() in order
to allow KCSAN to detect data races with plain C-language reads?
Or is that being handled some other way?

> -		ret = data_race(cmpxchg(p32, old32.w, new32.w)); // Overridden above.
> +
> +		// raw_cmpxchg() is used here to avoid instrumentation.
> +		ret = raw_cmpxchg(p32, old32.w, new32.w); // Overridden above.
>  	} while (ret != old32.w);
>  	return old;
>  }
> -- 
> 2.44.0
> 

Powered by blists - more mailing lists