Re: Re: Re: Re: Re: [RFC][PATCH 2/5] mips/atomic: Fix loongson_llsc_mb() wreckage

[Peter Zijlstra]

(sorry for the delay, I got sidetracked elsewhere)

On Fri, Apr 26, 2019 at 10:57:20AM +0800, huangpei@xxxxxxxxxxx wrote:
> > -----ååéä-----
> > On Thu, Apr 25, 2019 at 08:51:17PM +0800, huangpei@xxxxxxxxxxx wrote:
> > 
> > > > So basically the initial value of @v is set to 1.
> > > > 
> > > > Then CPU-1 does atomic_add_unless(v, 1, 0)
> > > >      CPU-2 does atomic_set(v, 0)
> > > > 
> > > > If CPU1 goes first, it will see 1, which is not 0 and thus add 1 to 1
> > > > and obtains 2. Then CPU2 goes and writes 0, so the exist clause sees
> > > > v==0 and doesn't observe 2.
> > > > 
> > > > The other way around, CPU-2 goes first, writes a 0, then CPU-1 goes and
> > > > observes the 0, finds it matches 0 and doesn't add.  Again, the exist
> > > > clause will find 0 doesn't match 2.
> > > > 
> > > > This all goes unstuck if interleaved like:
> > > > 
> > > > 
> > > > 	CPU-1			CPU-2
> > > > 
> > > > 				xor	t0, t0
> > > > 1:	ll	t0, v
> > > > 	bez	t0, 2f
> > > > 				sw	t0, v
> > > > 	add	t0, t1
> > > > 	sc	t0, v
> > > > 	beqz t0, 1b
> > > > 
> > > > (sorry if I got the MIPS asm wrong; it's not something I normally write)
> > > > 
> > > > And the store-word from CPU-2 doesn't make the SC from CPU-1 fail.
> > > > 
> > > 
> > > loongson's llsc bug DOES NOT fail this litmus( we will not get V=2)ï
> > > 
> > > only speculative memory access from CPU-1 can "blind" CPU-1(here blind means do ll/sc
> > >  wrongï, this speculative memory access can be observed corrently by CPU2. In this 
> > > case, sw from CPU-2 can get I , which can be observed by CPU-1, and clear llbitïthen 
> > > failed sc. 
> > 
> > I'm not following, suppose CPU-1 happens as a speculation (imagine
> > whatever code is required to make that happen before). CPU-2 sw will
> > cause I on CPU-1's ll but, as in the previous email, CPU-1 will continue
> > as if it still has E and complete the SC.
> > 
> > That is; I'm just not seeing why this case would be different from two
> > competing LL/SCs.
> > 
> 
> I get your point. I kept my eye on the sw from CPU-2, but forgot the speculative
>  mem access from CPU-1. 
> 
> There is no difference bewteen this one and the former case.
> 
> ========================================================================= 
>                        V = 1
> 
>     CPU-1                       CPU-2
> 
>                                 xor  t0, t0
> 1:  ll     t0, V               
>     beqz   t0, 2f
> 
>     /* if speculative mem 
>     access kick cacheline of
>     V out, it can blind CPU-1 
>     and make CPU-1 believe it 
>     still hold E on V, and can
>     NOT see the sw from CPU-2
>     actually invalid V, which 
>     should clear LLBit of CPU-1, 
>     but not */
>                                 sw   t0, V     // just after sw, V = 0
>     addiu  t0, t0, 1            
> 
>     sc     t0, V
>     /* oops, sc write t0(2) 
>     into V with LLBit */
> 
>     /* get V=2 */
>     beqz   t0, 1b
>     nop
> 2:
> ================================================================================    
>                
> if speculative mem access *does not* kick out cache line of V, CPU-1 can see sw
> from CPU-2, and clear LLBit, which cause sc fail and retry, That's OK

OK; so do I understand it correctly that your CPU _can_ in fact fail
that test and result in 2? If so I think I'm (finally) understanding :-)