Re: [PATCH] mm: Free per cpu pages async to shorten program exit time

[Claudio Imbrenda]

On Fri, 8 Oct 2021 11:15:25 +0200
David Hildenbrand <david@xxxxxxxxxx> wrote:

> On 08.10.21 10:52, Claudio Imbrenda wrote:
> > On Fri, 8 Oct 2021 10:17:50 +0200
> > David Hildenbrand <david@xxxxxxxxxx> wrote:
> >   
> >> On 08.10.21 08:39, ultrachin@xxxxxxx wrote:  
> >>> From: chen xiaoguang <xiaoggchen@xxxxxxxxxxx>
> >>>
> >>> The exit time is long when program allocated big memory and
> >>> the most time consuming part is free memory which takes 99.9%
> >>> of the total exit time. By using async free we can save 25% of
> >>> exit time.
> >>>
> >>> Signed-off-by: chen xiaoguang <xiaoggchen@xxxxxxxxxxx>
> >>> Signed-off-by: zeng jingxiang <linuszeng@xxxxxxxxxxx>
> >>> Signed-off-by: lu yihui <yihuilu@xxxxxxxxxxx>  
> >>
> >> I recently discussed with Claudio if it would be possible to tear down
> >> the process MM deferred, because for some use cases (secure/encrypted
> >> virtualization, very large mmaps) tearing down the page tables is
> >> already the much more expensive operation.
> >>
> >> There is mmdrop_async(), and I wondered if one could reuse that concept
> >> when tearing down a process -- I didn't look into feasibility, however,
> >> so it's just some very rough idea.  
> > 
> > I have done some experiments by unconditionally replacing mmdrop with
> > mmdrop_async in exit.c and nothing broke, and exit time of large
> > processes was almost instant (with the actual cleanup being performed in
> > background)
> > 
> > my approach is probably simpler/cleaner:
> > 
> > diff --git a/include/asm-generic/mmu_context.h b/include/asm-generic/mmu_context.h
> > index 91727065bacb..900931a6a105 100644
> > --- a/include/asm-generic/mmu_context.h
> > +++ b/include/asm-generic/mmu_context.h
> > @@ -73,4 +73,8 @@ static inline void deactivate_mm(struct task_struct *tsk,
> >   }
> >   #endif
> >   
> > +#ifndef arch_exit_mm_mmput
> > +#define arch_exit_mm_mmput mmput
> > +#endif
> > +
> >   #endif /* __ASM_GENERIC_MMU_CONTEXT_H */
> > diff --git a/kernel/exit.c b/kernel/exit.c
> > index 9a89e7f36acb..604cb9c740fa 100644
> > --- a/kernel/exit.c
> > +++ b/kernel/exit.c
> > @@ -498,7 +498,7 @@ static void exit_mm(void)
> >          task_unlock(current);
> >          mmap_read_unlock(mm);
> >          mm_update_next_owner(mm);
> > -       mmput(mm);
> > +       arch_exit_mm_mmput(mm);
> >          if (test_thread_flag(TIF_MEMDIE))
> >                  exit_oom_victim();
> >   }
> > 
> > these are the minimal changes to common code, then each architecture can
> > define their own arch_exit_mm_mmput as they see fit (for example, to free
> > asynchronously only for certain classes of mm, like big ones, VMs, or so).
> > 
> > Another option is to simply always replace mmput with mmput_async, which I
> > expect will raise more eyebrows.  
> 
> Thanks Claudio.
> 
> I guess we'd use some heuristic to keep the eyebrows down. Having 
> something like
> 
> if (should_mput_async_on_exit(mm))
> 	mmput_async(mm);
> else
> 	mmput(mm);
> 
> whereby the heuristic can optionally consult the arch/config-knobs/... 
> doesn't sound too wrong to me if it works.
> 

yes, that is one of the possible solutions I had thought of.

although probably the small patch I posted above is even less intrusive
and should hopefully raise even fewer eyebrows, while also leaving the
door open to arch-specific code to do more than just mmput_async, if
needed.

in the end I really do not have any preference, I simply want something
everybody can agree on :)