lkml.org 
[lkml]   [2015]   [Apr]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: mlock() on DAX returns -ENOMEM
From
On Mon, Apr 13, 2015 at 3:56 PM, Kirill A. Shutemov
<kirill@shutemov.name> 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 majordomo@vger.kernel.org
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>
> --
> Kirill A. Shutemov

Thanks,
Yigal


\
 
 \ /
  Last update: 2015-04-14 12:41    [W:0.073 / U:2.744 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site