Re: + mm-revert-thp-make-madv_hugepage-check-for-mm-def_flags.patch added to -mm tree

[Gerald Schaefer <gerald.schaefer@de.ibm.com>]

On Wed, 26 Feb 2014 16:31:44 +0100
Oleg Nesterov <oleg@redhat.com> wrote:

> On 02/26, Christian Borntraeger wrote:
> >
> > On 26/02/14 15:50, Oleg Nesterov wrote:
> > >
> > > But perhaps qemu can be changed to avoid MADV_HUGEPAGE on s390 ?
> > > Otherwise I'd suggest the change below.
> > >
> > > Oleg.
> > >
> > >
> > > --- x/mm/huge_memory.c
> > > +++ x/mm/huge_memory.c
> > > @@ -1968,8 +1968,6 @@ out:
> > >  int hugepage_madvise(struct vm_area_struct *vma,
> > >  		     unsigned long *vm_flags, int advice)
> > >  {
> > > -	struct mm_struct *mm = vma->vm_mm;
> > > -
> > >  	switch (advice) {
> > >  	case MADV_HUGEPAGE:
> > >  		/*
> > > @@ -1977,8 +1975,16 @@ int hugepage_madvise(struct vm_area_stru
> > >  		 */
> > >  		if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
> > >  			return -EINVAL;
> > > -		if (mm->def_flags & VM_NOHUGEPAGE)
> > > +
> > > +/*
> > > + * MADV_HUGEPAGE after PRCTL_THP_DISABLE is broken on s390 because
> > > + * qemu blindly does madvise(MADV_HUGEPAGE) after s390_enable_sie().
> > > + */
> > > +#ifdef CONFIG_S390
> > > +		if (vma->vm_mm->def_flags & VM_NOHUGEPAGE)
> > >  			return -EINVAL;
> > > +#endif
> > > +
> >
> > Ifdefs are ugly but might be the only way of not breaking existing userspace.
> 
> Yes, agreed. but note that this code is already s390-specific, nobody
> else puts VM_NOHUGEPAGE into ->def_flags (until this series of course).
> 
> > If we come up with a solution for THP in KVM host processes on s390, we can
> > then remove that wart. We could even limit that hack to KVM only processes
> > to retain Alex' prctl capability by checking mm_has_pgste
> 
> Yes, yes, I looked at mm->context.has_pgste too ;) But I wasn't sure
> we can rely on it. Thanks.
> 
> > > +/*
> > > + * MADV_HUGEPAGE after PRCTL_THP_DISABLE is broken on s390 because
> > > + * qemu blindly does madvise(MADV_HUGEPAGE) after s390_enable_sie().
> > > + */
> > > +#ifdef CONFIG_S390
> > > +		if ((vma->vm_mm->def_flags & VM_NOHUGEPAGE) && mm_has_pgste(vma->vm_mm))
> > >  			return -EINVAL;
> 
> Hmm... but why do we need to check VM_NOHUGEPAGE then? Can't the
> change below work?

Yes, good point, that should do the trick.

> 
> Oleg.
> 
> --- x/arch/s390/mm/pgtable.c
> +++ x/arch/s390/mm/pgtable.c
> @@ -1084,7 +1084,6 @@ static inline void thp_split_mm(struct m
>  		vma->vm_flags &= ~VM_HUGEPAGE;
>  		vma->vm_flags |= VM_NOHUGEPAGE;
>  	}
> -	mm->def_flags |= VM_NOHUGEPAGE;
>  }
>  #else
>  static inline void thp_split_mm(struct mm_struct *mm)
> --- x/mm/huge_memory.c
> +++ x/mm/huge_memory.c
> @@ -1968,8 +1968,6 @@ out:
>  int hugepage_madvise(struct vm_area_struct *vma,
>  		     unsigned long *vm_flags, int advice)
>  {
> -	struct mm_struct *mm = vma->vm_mm;
> -
>  	switch (advice) {
>  	case MADV_HUGEPAGE:
>  		/*
> @@ -1977,8 +1975,12 @@ int hugepage_madvise(struct vm_area_stru
>  		 */
>  		if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
>  			return -EINVAL;
> -		if (mm->def_flags & VM_NOHUGEPAGE)
> +
> +#ifdef CONFIG_S390
> +		if (mm_has_pgste(vma->vm_mm))
>  			return -EINVAL;
> +#endif
> +
>  		*vm_flags &= ~VM_NOHUGEPAGE;
>  		*vm_flags |= VM_HUGEPAGE;
>  		/*
>