Re: [PATCH] UBIFS: Implement ->migratepage()

[Richard Weinberger <richard@nod.at>]

Am 17.03.2016 um 10:57 schrieb Vlastimil Babka:
> +CC Hugh, Mel
> 
> On 03/16/2016 11:55 PM, Richard Weinberger wrote:
>> From: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
>>
>> When using CMA during page migrations UBIFS might get confused
> 
> It shouldn't be CMA specific, the same code runs from compaction, autonuma balancing...
> 
>> and the following assert triggers:
>> UBIFS assert failed in ubifs_set_page_dirty at 1451 (pid 436)
>>
>> UBIFS is using PagePrivate() which can have different meanings across
>> filesystems. Therefore the generic page migration code cannot handle this
>> case correctly.
>> We have to implement our own migration function which basically does a
>> plain copy but also duplicates the page private flag.
> 
> Lack of PagePrivate() migration is surely a bug, but at a glance of how UBIFS uses the flag, it's more about accounting, it shouldn't prevent a page from being marked PageDirty()?
> I suspect your initial bug (which is IIUC the fact that there's a dirty pte, but PageDirty(page) is false) comes from the generic fallback_migrate_page() which does:
> 
>         if (PageDirty(page)) {
>                 /* Only writeback pages in full synchronous migration */
>                 if (mode != MIGRATE_SYNC)
>                         return -EBUSY;
>                 return writeout(mapping, page);
>         }
> 
> And writeout() seems to Clear PageDirty() through clear_page_dirty_for_io() but I'm not so sure about the pte (or pte's in all rmaps). But this comment in the latter function:
> 
>                  * Yes, Virginia, this is indeed insane.
> 
> scared me enough to not investigate further. Hopefully the people I CC'd understand more about page migration than me. I'm just an user :)
> 
> In any case, this patch would solve both lack of PageDirty() transfer, and avoid the path leading from fallback_migrate_page() to writeout(). But I'm not confident enough here to
> ack it.

Hugh? Mel? Anyone? :-)

It is still not clear to me whether this needs fixing in MM or UBIFS.

Thanks,
//richard