Re: arc_usr_cmpxchg and preemption

[Alexey Brodkin <Alexey.Brodkin@synopsys.com>]

Hi Peter, Vineet,

On Wed, 2018-03-14 at 18:53 +0100, Peter Zijlstra wrote:
> On Wed, Mar 14, 2018 at 09:58:19AM -0700, Vineet Gupta wrote:
> 
> > Well it is broken wrt the semantics the syscall is supposed to provide.
> > Preemption disabling is what prevents a concurrent thread from coming in and
> > modifying the same location (Imagine a variable which is being cmpxchg
> > concurrently by 2 threads).
> > 
> > One approach is to do it the MIPS way, emulate the llsc flag - set it under
> > preemption disabled section and clear it in switch_to
> 
> *shudder*... just catch the -EFAULT, force the write fault and retry.
> 
> Something like:
> 
> int sys_cmpxchg(u32 __user *user_ptr, u32 old, u32 new)
> {
> 	u32 val;
> 	int ret;
> 
> again:
> 	ret = 0;
> 
> 	preempt_disable();
> 	val = get_user(user_ptr);
> 	if (val == old)
> 		ret = put_user(new, user_ptr);
> 	preempt_enable();
> 
> 	if (ret == -EFAULT) {
> 		struct page *page;
> 		ret = get_user_pages_fast((unsigned long)user_ptr, 1, 1, &page);
> 		if (ret < 0)
> 			return ret;
> 		put_page(page);
> 		goto again;

I guess this jump we need to do only once, right?
If for whatever reason get_user_pages_fast() fails we return immediately
and if it succeeds there's no reason for put_user() to not succeed as
required page is supposed to be prepared for write.

Otherwise if something goes way too bad we may end-up in an infinite loop
which we'd better prevent.

> 	}
> 
> 	return ret;
> }

@Vineet, are you OK with proposed implementation?

-Alexey