Re: [RFC PATCH v5 03/12] __wr_after_init: Core and default arch

[Igor Stoppa]

On 14/02/2019 13:28, Peter Zijlstra wrote:
> On Thu, Feb 14, 2019 at 12:41:32AM +0200, Igor Stoppa wrote:

[...]

> > +#define wr_rcu_assign_pointer(p, v) ({	\
> > +	smp_mb();			\
> > +	wr_assign(p, v);		\
> > +	p;				\
> > +})
>
> This requires that wr_memcpy() (through wr_assign) is single-copy-atomic
> for native types. There is not a comment in sight that states this.

Right, I kinda expected native-aligned <-> atomic, but it's not 
necessarily true. It should be confirmed when enabling write rare on a 
new architecture. I'll add the comment.

> Also, is this true of x86/arm64 memcpy ?


For x86_64:
https://elixir.bootlin.com/linux/v5.0-rc6/source/arch/x86/include/asm/uaccess.h#L462 
 the mov"itype"  part should deal with atomic copy of native, aligned 
types.


For arm64:
https://elixir.bootlin.com/linux/v5.0-rc6/source/arch/arm64/lib/copy_template.S#L110 
.Ltiny15 deals with copying less than 16 bytes, which includes pointers. 
When the data is aligned, the copy of a pointer should be atomic.


--
igor