Re: [PATCH] arch, mm: introduce arch_tlb_gather_mmu_lazy

[Minchan Kim]

On Wed, Nov 22, 2017 at 07:30:50PM +0000, Will Deacon wrote:
> Hi Michal,
> 
> On Mon, Nov 20, 2017 at 05:04:22PM +0100, Michal Hocko wrote:
> > On Mon 20-11-17 14:24:44, Will Deacon wrote:
> > > On Thu, Nov 16, 2017 at 10:20:42AM +0100, Michal Hocko wrote:
> > > > On Wed 15-11-17 17:33:32, Will Deacon wrote:
> > [...]
> > > > > > diff --git a/arch/arm64/include/asm/tlb.h b/arch/arm64/include/asm/tlb.h
> > > > > > index ffdaea7954bb..7adde19b2bcc 100644
> > > > > > --- a/arch/arm64/include/asm/tlb.h
> > > > > > +++ b/arch/arm64/include/asm/tlb.h
> > > > > > @@ -43,7 +43,7 @@ static inline void tlb_flush(struct mmu_gather *tlb)
> > > > > >  	 * The ASID allocator will either invalidate the ASID or mark
> > > > > >  	 * it as used.
> > > > > >  	 */
> > > > > > -	if (tlb->fullmm)
> > > > > > +	if (tlb->lazy)
> > > > > >  		return;
> > > > > 
> > > > > This looks like the right idea, but I'd rather make this check:
> > > > > 
> > > > > 	if (tlb->fullmm && tlb->lazy)
> > > > > 
> > > > > since the optimisation doesn't work for anything than tearing down the
> > > > > entire address space.
> > > > 
> > > > OK, that makes sense.
> > > > 
> > > > > Alternatively, I could actually go check MMF_UNSTABLE in tlb->mm, which
> > > > > would save you having to add an extra flag in the first place, e.g.:
> > > > > 
> > > > > 	if (tlb->fullmm && !test_bit(MMF_UNSTABLE, &tlb->mm->flags))
> > > > > 
> > > > > which is a nice one-liner.
> > > > 
> > > > But that would make it oom_reaper specific. What about the softdirty
> > > > case Minchan has mentioned earlier?
> > > 
> > > We don't (yet) support that on arm64, so we're ok for now. If we do grow
> > > support for it, then I agree that we want a flag to identify the case where
> > > the address space is going away and only elide the invalidation then.
> > 
> > What do you think about the following patch instead? I have to confess
> > I do not really understand the fullmm semantic so I might introduce some
> > duplication by this flag. If you think this is a good idea, I will post
> > it in a separate thread.
> 
> 
> Please do! My only suggestion would be s/lazy/exit/, since I don't think the
> optimisation works in any other situation than the address space going away
> for good.

Yes, address space going. That's why I wanted to add additional check that
address space going without adding new flags.

http://lkml.kernel.org/r/<20171113002833.GA18301@bbox>

However, if you guys love to add new flag to distinguish, I prefer
"exit" to "lazy". It also would be better to add WARN_ON to catch
future potential wrong use case like OOM reaper.
Anyway, I'm not strong against so it up to you, Michal.

        WARN_ON_ONCE(exit == true && atomic_read(&mm->mm_users) > 0);