Re: [PATCH] mm: Unsigned 'nr_pages' always larger than zero

[Vlastimil Babka]

On 9/4/19 9:01 PM, Andrew Morton wrote:
> On Wed, 4 Sep 2019 13:24:58 +0200 Vlastimil Babka <vbabka@xxxxxxx> wrote:
> 
>> On 9/4/19 12:26 PM, zhong jiang wrote:
>>> With the help of unsigned_lesser_than_zero.cocci. Unsigned 'nr_pages"'
>>> compare with zero. And __get_user_pages_locked will return an long value.
>>> Hence, Convert the long to compare with zero is feasible.
>>
>> It would be nicer if the parameter nr_pages was long again instead of unsigned
>> long (note there are two variants of the function, so both should be changed).
>>
>>> Signed-off-by: zhong jiang <zhongjiang@xxxxxxxxxx>
>>
>> Fixes: 932f4a630a69 ("mm/gup: replace get_user_pages_longterm() with FOLL_LONGTERM")
>>
>> (which changed long to unsigned long)
>>
>> AFAICS... stable shouldn't be needed as the only "risk" is that we goto
>> check_again even when we fail, which should be harmless.
>>
> 
> Really?  If nr_pages gets a value of -EFAULT from the
> __get_user_pages_locked() call, check_and_migrate_cma_pages() will go
> berzerk?

Hmm it should only reach that goto when it migrated something, which
means it won't have to be migrated again, so eventually it should
terminate. But it's very subtle, so I might be wrong.

> And does __get_user_pages_locked() correctly handle a -ve errno
> returned by __get_user_pages()?  It's hard to see how...

I think it does, but agree.