Re: [PATCH v2] ubifs: fix page_count in ->ubifs_migrate_page()

Re: [PATCH v2] ubifs: fix page_count in ->ubifs_migrate_page()

[Richard Weinberger]

Am Donnerstag, 13. Dezember 2018, 20:20:17 CET schrieb zhangjun:
> Because the PagePrivate() in UBIFS is different meanings,

...has different meanings...

I'll fix up that myself after applying your patch. No need to send a v3.

> alloc_cma() will fail when one dirty page cache located in
> the type of MIGRATE_CMA
> 
> If not adjust the 'extra_count' for dirty page,
> ubifs_migrate_page() -> migrate_page_move_mapping() will
> always return -EAGAIN for:
> 	expected_count += page_has_private(page)
> This causes the migration to fail until the page cache is cleaned
> 
> In general, PagePrivate() indicates that buff_head is already bound
> to this page, and at the same time page_count() will also increase.
> But UBIFS set private flag when the cache is dirty, and page_count()
> not increase.
> Therefore, the expected_count of UBIFS is different from the general
> case.
> 
> Signed-off-by: zhangjun <openzhangj@gmail.com>

Fixes: 4ac1c17b2044 ("UBIFS: Implement ->migratepage()")

> ---
>  fs/ubifs/file.c | 11 ++++++++++-
>  1 file changed, 10 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/ubifs/file.c b/fs/ubifs/file.c
> index 1b78f2e..890dfce 100644
> --- a/fs/ubifs/file.c
> +++ b/fs/ubifs/file.c
> @@ -1480,8 +1480,17 @@ static int ubifs_migrate_page(struct address_space *mapping,
>  		struct page *newpage, struct page *page, enum migrate_mode mode)
>  {
>  	int rc;
> +	int extra_count = 0;
>  
> -	rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
> +	/*
> +	 * UBIFS uses PG_private as marker and does not raise the page counter.
> +	 * migrate_page_move_mapping() expects a incremented counter if
> +	 * PG_private is set. Therefore pass -1 as extra_count for this case.
> +	 */
> +	if (page_has_private(page))
> +		extra_count = -1;
> +	rc = migrate_page_move_mapping(mapping, newpage, page,
> +			NULL, mode, extra_count);
>  	if (rc != MIGRATEPAGE_SUCCESS)
>  		return rc;
>  

Let's wait a few days to give Kirill a chance to review, then I'll apply the patch.

Thanks,
//richard 

Re: [PATCH v2] ubifs: fix page_count in ->ubifs_migrate_page()

[Kirill A. Shutemov]

On Thu, Dec 13, 2018 at 10:36:47PM +0100, Richard Weinberger wrote:
> Am Donnerstag, 13. Dezember 2018, 20:20:17 CET schrieb zhangjun:
> > Because the PagePrivate() in UBIFS is different meanings,
> 
> ...has different meanings...
> 
> I'll fix up that myself after applying your patch. No need to send a v3.
> 
> > alloc_cma() will fail when one dirty page cache located in
> > the type of MIGRATE_CMA
> > 
> > If not adjust the 'extra_count' for dirty page,
> > ubifs_migrate_page() -> migrate_page_move_mapping() will
> > always return -EAGAIN for:
> > 	expected_count += page_has_private(page)
> > This causes the migration to fail until the page cache is cleaned
> > 
> > In general, PagePrivate() indicates that buff_head is already bound
> > to this page, and at the same time page_count() will also increase.
> > But UBIFS set private flag when the cache is dirty, and page_count()
> > not increase.
> > Therefore, the expected_count of UBIFS is different from the general
> > case.
> > 
> > Signed-off-by: zhangjun <openzhangj@gmail.com>
> 
> Fixes: 4ac1c17b2044 ("UBIFS: Implement ->migratepage()")
> 
> > ---
> >  fs/ubifs/file.c | 11 ++++++++++-
> >  1 file changed, 10 insertions(+), 1 deletion(-)
> > 
> > diff --git a/fs/ubifs/file.c b/fs/ubifs/file.c
> > index 1b78f2e..890dfce 100644
> > --- a/fs/ubifs/file.c
> > +++ b/fs/ubifs/file.c
> > @@ -1480,8 +1480,17 @@ static int ubifs_migrate_page(struct address_space *mapping,
> >  		struct page *newpage, struct page *page, enum migrate_mode mode)
> >  {
> >  	int rc;
> > +	int extra_count = 0;
> >  
> > -	rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
> > +	/*
> > +	 * UBIFS uses PG_private as marker and does not raise the page counter.
> > +	 * migrate_page_move_mapping() expects a incremented counter if
> > +	 * PG_private is set. Therefore pass -1 as extra_count for this case.
> > +	 */
> > +	if (page_has_private(page))
> > +		extra_count = -1;
> > +	rc = migrate_page_move_mapping(mapping, newpage, page,
> > +			NULL, mode, extra_count);
> >  	if (rc != MIGRATEPAGE_SUCCESS)
> >  		return rc;
> >  
> 
> Let's wait a few days to give Kirill a chance to review, then I'll apply the patch.

I don't remmeber much context now...

Could you remind me why ubifs doesn't take additional pin when sets
PG_private?

Migration is not the only place where the additional pin is implied.
See all users of page_has_private() helper. Notably reclaim path.

-- 
 Kirill A. Shutemov

Re: [PATCH v2] ubifs: fix page_count in ->ubifs_migrate_page()

[Richard Weinberger]

Am Donnerstag, 13. Dezember 2018, 23:00:00 CET schrieb Kirill A. Shutemov:
> > Let's wait a few days to give Kirill a chance to review, then I'll apply the patch.
> 
> I don't remmeber much context now...
> 
> Could you remind me why ubifs doesn't take additional pin when sets
> PG_private?

Because it does not use set_page_private(), it just sets the flag for
internal accounting purposes.

>From UBIFS docs:
 * UBIFS uses 2 page flags: @PG_private and @PG_checked. @PG_private is set if
 * the page is dirty and is used for optimization purposes - dirty pages are
 * not budgeted so the flag shows that 'ubifs_write_end()' should not release
 * the budget for this page. The @PG_checked flag is set if full budgeting is
 * required for the page e.g., when it corresponds to a file hole or it is
 * beyond the file size. The budgeting is done in 'ubifs_write_begin()', because
 * it is OK to fail in this function, and the budget is released in
 * 'ubifs_write_end()'. So the @PG_private and @PG_checked flags carry
 * information about how the page was budgeted, to make it possible to release
 * the budget properly.

> Migration is not the only place where the additional pin is implied.
> See all users of page_has_private() helper. Notably reclaim path.

Hmmm, I need to dig into that.
I this is a problem then f2fs suffers from it too.
At least from what I can tell from reading f2fs_migrate_page().

Thanks,
//richard