Re: [PATCH] mm: make PR_SET_THP_DISABLE immediately active

[Mike Rapoport]

On Sat, Jun 03, 2017 at 01:34:52PM +0300, Mike Rapoprt wrote:
> 
> 
> On June 2, 2017 11:55:12 PM GMT+03:00, Vlastimil Babka <vbabka@xxxxxxx> wrote:
> >On 06/02/2017 10:40 PM, Andrew Morton wrote:
> >> On Fri, 2 Jun 2017 22:31:47 +0200 Vlastimil Babka <vbabka@xxxxxxx>
> >wrote:
> >>>> Perhaps we should be adding new prctl modes to select this new
> >>>> behaviour and leave the existing PR_SET_THP_DISABLE behaviour
> >as-is?
> >>>
> >>> I think we can reasonably assume that most users of the prctl do
> >just
> >>> the fork() & exec() thing, so they will be unaffected.
> >> 
> >> That sounds optimistic.  Perhaps people are using the current
> >behaviour
> >> to set on particular mapping to MMF_DISABLE_THP, with
> >> 
> >> 	prctl(PR_SET_THP_DISABLE)
> >> 	mmap()
> >> 	prctl(PR_CLR_THP_DISABLE)
> >> 
> >> ?
> >> 
> >> Seems a reasonable thing to do.
> >
> >Using madvise(MADV_NOHUGEPAGE) seems reasonabler to me, with the same
> >effect. And it's older (2.6.38).
> >
> >> But who knows - people do all sorts of
> >> inventive things.
> >
> >Yeah :( but we can hope they don't even know that the prctl currently
> >behaves they way it does - man page doesn't suggest it would, and most
> >of us in this thread found it surprising.
> >
> >>> And as usual, if
> >>> somebody does complain in the end, we revert and try the other way?
> >> 
> >> But by then it's too late - the new behaviour will be out in the
> >field.
> >
> >Revert in stable then?
> >But I don't think this patch should go to stable. I understand right
> >that CRIU will switch to the UFFDIO_COPY approach and doesn't need the
> >prctl change/new madvise anymore?
> 
> Yes, we are going to use UFFDIO_COPY. We still might want to have control
> over THP in the future without changing per-VMA flags, though.

Unfortunately, I was over optimistic about ability of CRIU to use
UFFDIO_COPY for pre-copy part :(
I was too concentrated on the simplified flow and overlooked some important
details. After I've spent some time trying to actually implement usage of
UFFDIO_COPY, I realized that registering memory with userfault at that
point of the restore flow quite contradicts CRIU architecture :(

That said, we would really want to have the interface this patch proposes.
 
-- 
Sincerely yours,
Mike.