2015-04-13 12:57:08

by Kirill A. Shutemov

[permalink] [raw]
Subject: Re: mlock() on DAX returns -ENOMEM

On Sun, Apr 12, 2015 at 03:56:33PM +0300, Yigal Korman wrote:
> Hi,
> I've tried to mlock() a range of an ext4-dax file and got "-ENOMEM" in return.

Is it comes from mlock_fixup() or -EFAULT from GUP translated to -ENOMEM
by __mlock_posix_error_return()?

> Looking at the code, it seems that this is related to the fact that
> DAX uses VM_MIXEDMAP and mlock assumes/requires regular page cache.
> To me it seems that DAX should simply return success in mlock() as all
> data is always in memory and no swapping is possible.
> Is this a bug or intentional? Is there a fix planned?

I think it's a bug.

But first we need to define what mlock() means for DAX mappings.

For writable MAP_PRIVATE: we should be able to trigger COW for the range
and mlock resulting pages. It means we should fix kernel to handle
GUP(FOLL_TOUCH | FOLL_POPULATE | FOLL_WRITE | FOLL_FORCE) successfully on
such VMAs.

For MAP_SHARED and non-writable MAP_PRIVATE we should be able to populate
the mapping with PTEs. Not sure if we need to set VM_LOCKED for such VMAs.
We probably should, as we want to re-instantiate PTEs on mremap() and such.
It means we need to get working at least GUP(FOLL_POPULATE | FOLL_FORCE).

In general we need to adjust GUP to avoid going to struct page unless
FOLL_* speficly imply struct page, such as FOLL_GET or FOLL_TOUCH.

Not sure if we need to differentiate DAX mappings from other VM_MIXEDMAP.

Any comments?

> Also, the same code path that is used in mlock is also used for
> MAP_POPULATE (pre-fault pages in mmap) so this flag doesn't work as
> well (doesn't fail but simply doesn't pre-fault anything).
>
> Thanks,
> Yigal
> --
> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html

--
Kirill A. Shutemov


2015-04-14 10:27:47

by Yigal Korman

[permalink] [raw]
Subject: Re: mlock() on DAX returns -ENOMEM

On Mon, Apr 13, 2015 at 3:56 PM, Kirill A. Shutemov
<[email protected]> wrote:
> On Sun, Apr 12, 2015 at 03:56:33PM +0300, Yigal Korman wrote:
>> Hi,
>> I've tried to mlock() a range of an ext4-dax file and got "-ENOMEM" in return.
>
> Is it comes from mlock_fixup() or -EFAULT from GUP translated to -ENOMEM
> by __mlock_posix_error_return()?

It comes from GUP. An associate followed the flow from GUP to
vm_normal_page which returns NULL for VM_MIXEDMAP.

>
>> Looking at the code, it seems that this is related to the fact that
>> DAX uses VM_MIXEDMAP and mlock assumes/requires regular page cache.
>> To me it seems that DAX should simply return success in mlock() as all
>> data is always in memory and no swapping is possible.
>> Is this a bug or intentional? Is there a fix planned?
>
> I think it's a bug.
>
> But first we need to define what mlock() means for DAX mappings.
>
> For writable MAP_PRIVATE: we should be able to trigger COW for the range
> and mlock resulting pages. It means we should fix kernel to handle
> GUP(FOLL_TOUCH | FOLL_POPULATE | FOLL_WRITE | FOLL_FORCE) successfully on
> such VMAs.

writable MAP_PRIVATE seems to work OK already.

>
> For MAP_SHARED and non-writable MAP_PRIVATE we should be able to populate
> the mapping with PTEs. Not sure if we need to set VM_LOCKED for such VMAs.
> We probably should, as we want to re-instantiate PTEs on mremap() and such.
> It means we need to get working at least GUP(FOLL_POPULATE | FOLL_FORCE).
>
> In general we need to adjust GUP to avoid going to struct page unless
> FOLL_* speficly imply struct page, such as FOLL_GET or FOLL_TOUCH.
>
> Not sure if we need to differentiate DAX mappings from other VM_MIXEDMAP.
>
> Any comments?

Indeed, the issue is VM_MIXEDMAP and not DAX specifically.
What will happen for other users of VM_MIXEDMAP that, unlike DAX, do
swap pages and need some special callback to handle mlock?

>
>> Also, the same code path that is used in mlock is also used for
>> MAP_POPULATE (pre-fault pages in mmap) so this flag doesn't work as
>> well (doesn't fail but simply doesn't pre-fault anything).
>>
>> Thanks,
>> Yigal
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
>> the body of a message to [email protected]
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>
> --
> Kirill A. Shutemov

Thanks,
Yigal

2015-04-14 11:42:33

by Kirill A. Shutemov

[permalink] [raw]
Subject: Re: mlock() on DAX returns -ENOMEM

On Tue, Apr 14, 2015 at 01:27:39PM +0300, Yigal Korman wrote:
> On Mon, Apr 13, 2015 at 3:56 PM, Kirill A. Shutemov
> <[email protected]> wrote:
> > On Sun, Apr 12, 2015 at 03:56:33PM +0300, Yigal Korman wrote:
> >> Hi,
> >> I've tried to mlock() a range of an ext4-dax file and got "-ENOMEM" in return.
> >
> > Is it comes from mlock_fixup() or -EFAULT from GUP translated to -ENOMEM
> > by __mlock_posix_error_return()?
>
> It comes from GUP. An associate followed the flow from GUP to
> vm_normal_page which returns NULL for VM_MIXEDMAP.
>
> >
> >> Looking at the code, it seems that this is related to the fact that
> >> DAX uses VM_MIXEDMAP and mlock assumes/requires regular page cache.
> >> To me it seems that DAX should simply return success in mlock() as all
> >> data is always in memory and no swapping is possible.
> >> Is this a bug or intentional? Is there a fix planned?
> >
> > I think it's a bug.
> >
> > But first we need to define what mlock() means for DAX mappings.
> >
> > For writable MAP_PRIVATE: we should be able to trigger COW for the range
> > and mlock resulting pages. It means we should fix kernel to handle
> > GUP(FOLL_TOUCH | FOLL_POPULATE | FOLL_WRITE | FOLL_FORCE) successfully on
> > such VMAs.
>
> writable MAP_PRIVATE seems to work OK already.

Really? Have you try mlock() on writable MAP_PRIVATE mappings after some
*read* read faults to the VMA?

I suspect mlock() will fail if VMA was populated with PFNs before mlock()
call.

>
> >
> > For MAP_SHARED and non-writable MAP_PRIVATE we should be able to populate
> > the mapping with PTEs. Not sure if we need to set VM_LOCKED for such VMAs.
> > We probably should, as we want to re-instantiate PTEs on mremap() and such.
> > It means we need to get working at least GUP(FOLL_POPULATE | FOLL_FORCE).
> >
> > In general we need to adjust GUP to avoid going to struct page unless
> > FOLL_* speficly imply struct page, such as FOLL_GET or FOLL_TOUCH.
> >
> > Not sure if we need to differentiate DAX mappings from other VM_MIXEDMAP.
> >
> > Any comments?
>
> Indeed, the issue is VM_MIXEDMAP and not DAX specifically.
> What will happen for other users of VM_MIXEDMAP that, unlike DAX, do
> swap pages and need some special callback to handle mlock?

Do far only DRM drivers uses VM_MIXEDMAP. It shouldn't be too hard to
audit this code.

Or we can identify DAX mapping some other way. Or introduce VM_DAX.

> >
> >> Also, the same code path that is used in mlock is also used for
> >> MAP_POPULATE (pre-fault pages in mmap) so this flag doesn't work as
> >> well (doesn't fail but simply doesn't pre-fault anything).
> >>
> >> Thanks,
> >> Yigal
> >> --
> >> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
> >> the body of a message to [email protected]
> >> More majordomo info at http://vger.kernel.org/majordomo-info.html
> >
> > --
> > Kirill A. Shutemov
>
> Thanks,
> Yigal
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/

--
Kirill A. Shutemov

2015-04-14 19:44:27

by Matthew Wilcox

[permalink] [raw]
Subject: Re: mlock() on DAX returns -ENOMEM

On Tue, Apr 14, 2015 at 02:42:12PM +0300, Kirill A. Shutemov wrote:
> Or we can identify DAX mapping some other way. Or introduce VM_DAX.

IS_DAX(vma->vm_file->f_mapping->host)