Re: [PATCH v1 2/6] mm/page_isolation: don't dump_page(NULL) in set_migratetype_isolate()

[Mike Kravetz <mike.kravetz@oracle.com>]

On 6/30/20 7:26 AM, David Hildenbrand wrote:
> Right now, if we have two isolations racing, we might trigger the
> WARN_ON_ONCE() and to dump_page(NULL), dereferencing NULL. Let's just
> return directly.

Just curious, what call path has the WARN_ON_ONCE()/dump_page(NULL)?

> 
> In the future, we might want to report -EAGAIN to the caller instead, as
> this could indicate a temporary isolation failure only.
> 
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: Michal Hocko <mhocko@suse.com>
> Cc: Michael S. Tsirkin <mst@redhat.com>
> Signed-off-by: David Hildenbrand <david@redhat.com>

Hi David,

That 'return -EAGAIN' was added as a sort of synchronization mechanism.
See commit message for 2c7452a075d4d.  Before adding the 'return -EAGAIN',
I could create races which would abandon isolated pageblocks.  Repeating
those races over and over would result in a good chunk of system memory
being isolated and unusable.

Admittedly, these races are rare and I had to work really hard to produce
them.  I'll try to find my testing mechanism.  My concern is reintroducing
this abandoning of pageblocks.  I have not looked further in your series
to see if this potentially addressed later.  If not, then we should not
remove the return code.

-- 
Mike Kravetz

> ---
>  mm/page_isolation.c | 8 +++++---
>  1 file changed, 5 insertions(+), 3 deletions(-)
> 
> diff --git a/mm/page_isolation.c b/mm/page_isolation.c
> index f6d07c5f0d34d..553b49a34cf71 100644
> --- a/mm/page_isolation.c
> +++ b/mm/page_isolation.c
> @@ -29,10 +29,12 @@ static int set_migratetype_isolate(struct page *page, int migratetype, int isol_
>  	/*
>  	 * We assume the caller intended to SET migrate type to isolate.
>  	 * If it is already set, then someone else must have raced and
> -	 * set it before us.  Return -EBUSY
> +	 * set it before us.
>  	 */
> -	if (is_migrate_isolate_page(page))
> -		goto out;
> +	if (is_migrate_isolate_page(page)) {
> +		spin_unlock_irqrestore(&zone->lock, flags);
> +		return -EBUSY;
> +	}
>  
>  	/*
>  	 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
>