Re: [PATCH v3] mm/gup: stop leaking pinned pages in low memory conditions

[John Hubbard <jhubbard@nvidia.com>]

On 10/20/24 10:39 PM, Alistair Popple wrote:
> 
> John Hubbard <jhubbard@nvidia.com> writes:
> 
>> On 10/20/24 4:26 PM, Alistair Popple wrote:
>>> John Hubbard <jhubbard@nvidia.com> writes:
>>> [...]
>>>> @@ -2437,8 +2440,10 @@ static long check_and_migrate_movable_pages(unsigned long nr_pages,
>>>>    	long i, ret;
>>>>      	folios = kmalloc_array(nr_pages, sizeof(*folios), GFP_KERNEL);
>>>> -	if (!folios)
>>>> +	if (!folios) {
>>>> +		unpin_user_pages(pages, nr_pages);
>>> ie. Doesn't this unpinning need to happen in
>>> check_and_migrate_movable_folios()?
>>
>> It already does.
>>
>> check_and_migrate_movable_folios() calls
>> migrate_longterm_unpinnable_folios(), which unpins if errors occur.
> 
> Right you are.
> 
> Reviewed-by: Alistair Popple <apopple@nvidia.com>

Thanks for the review!

> 
> As an aside for future clean-ups we could probably get something nicer
> if we reversed the process of pin/migrate to migrate/pin. In other words
> if FOLL_LONGERM try and migrate the entire range first out of
> ZONE_MOVABLE first. Migration invovles walking page tables and getting a
> reference on the pages anyway, so if it turns out there is nothing to
> migrate you haven't lost anything performance wise.
> 

Yes. In fact, I see our emails crossed, and I just suggested the same thing
in reply to your other comment (in the v2 review thread) about short vs.
long term pinning. Great! :)


thanks,
-- 
John Hubbard