lkml.org 
[lkml]   [2022]   [Aug]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v1 1/2] mm/hugetlb: fix hugetlb not supporting write-notify
On 08/05/22 19:13, Peter Xu wrote:
> On Fri, Aug 05, 2022 at 01:48:35PM -0700, Mike Kravetz wrote:
> > On 08/05/22 20:57, David Hildenbrand wrote:
> > > On 05.08.22 20:33, Mike Kravetz wrote:
> > > > On 08/05/22 20:25, David Hildenbrand wrote:
> > > >> On 05.08.22 20:23, Mike Kravetz wrote:
> > > >>> On 08/05/22 14:14, Peter Xu wrote:
> > > >>>> On Fri, Aug 05, 2022 at 01:03:28PM +0200, David Hildenbrand wrote:
> > > >>>>> diff --git a/mm/mmap.c b/mm/mmap.c
> > > >>>>> index 61e6135c54ef..462a6b0344ac 100644
> > > >>>>> --- a/mm/mmap.c
> > > >>>>> +++ b/mm/mmap.c
> > > >>>>> @@ -1683,6 +1683,13 @@ int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
> > > >>>>> if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
> > > >>>>> return 0;
> > > >>>>>
> > > >>>>> + /*
> > > >>>>> + * Hugetlb does not require/support writenotify; especially, it does not
> > > >>>>> + * support softdirty tracking.
> > > >>>>> + */
> > > >>>>> + if (is_vm_hugetlb_page(vma))
> > > >>>>> + return 0;
> > > >>>>
> > > >>>> I'm kind of confused here.. you seems to be fixing up soft-dirty for
> > > >>>> hugetlb but here it's explicitly forbidden.
> > > >>>>
> > > >>>> Could you explain a bit more on why this patch is needed if (assume
> > > >>>> there'll be a working) patch 2 being provided?
> > > >>>>
> > > >>>
> > > >>> No comments on the patch, but ...
> > > >>>
> > > >>> Since it required little thought, I ran the test program on next-20220802 and
> > > >>> was surprised that the issue did not recreate. Even added a simple printk
> > > >>> to make sure we were getting into vma_wants_writenotify with a hugetlb vma.
> > > >>> We were.
> > > >>
> > > >>
> > > >> ... does your config have CONFIG_MEM_SOFT_DIRTY enabled?
> > > >>
> > > >
> > > > No, Duh!
> > > >
> > > > FYI - Some time back, I started looking at adding soft dirty support for
> > > > hugetlb mappings. I did not finish that work. But, I seem to recall
> > > > places where code was operating on hugetlb mappings when perhaps it should
> > > > not.
> > > >
> > > > Perhaps, it would also be good to just disable soft dirty for hugetlb at
> > > > the source?
> > >
> > > I thought about that as well. But I came to the conclusion that without
> > > patch #2, hugetlb VMAs cannot possibly support write-notify, so there is
> > > no need to bother in vma_wants_writenotify() at all.
> > >
> > > The "root" would be places where we clear VM_SOFTDIRTY. That should only
> > > be fs/proc/task_mmu.c:clear_refs_write() IIRC.
> > >
> > > So I don't particularly care, I consider this patch a bit cleaner and
> > > more generic, but I can adjust clear_refs_write() instead of there is a
> > > preference.
> > >
> >
> > After a closer look, I agree that this may be the simplest/cleanest way to
> > proceed. I was going to suggest that you note hugetlb does not support
> > softdirty, but see you did in the comment.
> >
> > Acked-by: Mike Kravetz <mike.kravetz@oracle.com>
>
> Filtering out hugetlbfs in vma_wants_writenotify() is still a bit hard to
> follow to me, since it's not clear why hugetlbfs never wants writenotify.
>
> If it's only about soft-dirty, we could have added the hugetlbfs check into
> vma_soft_dirty_enabled(), then I think it'll achieve the same thing and
> much clearer - with the soft-dirty check constantly returning false for it,
> hugetlbfs shared vmas should have vma_wants_writenotify() naturally return
> 0 already.
>
> For the long term - shouldn't we just enable soft-dirty for hugetlbfs? I
> remember Mike used to have that in todo. Since we've got patch 2 already,
> I feel like that's really much close (is the only missing piece the clear
> refs write part? or maybe some more that I didn't notice).
>
> Then patch 1 (or IMHO equivalant check in vma_soft_dirty_enabled(), but
> maybe in stable trees we don't have vma_soft_dirty_enabled then it's
> exactly patch 1) can be a stable-only backport just to avoid the bug from
> triggering.

It looks like vma_soft_dirty_enabled is recent and not in any stable
trees (or even 5.19).

Yes, I did start working on hugetlb softdirty support in the past.
https://lore.kernel.org/lkml/20210211000322.159437-1-mike.kravetz@oracle.com/

Unfortunately, it got preempted by other things. I will try to move it up
the priority list.
--
Mike Kravetz

\
 
 \ /
  Last update: 2022-08-06 01:35    [W:0.098 / U:0.216 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site