Re: [PATCH] mm/gup.c: Handling ERR within unpin_user_pages()

[John Hubbard <jhubbard@nvidia.com>]

On 9/14/20 1:52 PM, Souptick Joarder wrote:
> On Mon, Sep 14, 2020 at 7:38 PM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>>
>> On Mon, Sep 14, 2020 at 07:20:34AM +0530, Souptick Joarder wrote:
>>> On Sun, Sep 13, 2020 at 8:25 PM Matthew Wilcox <willy@infradead.org> wrote:
>>>>
>>>> On Sun, Sep 13, 2020 at 08:02:35PM +0530, Souptick Joarder wrote:
>>>>> It is possible that a buggy caller of unpin_user_pages()
>>>>> (specially in error handling path) may end up calling it with
>>>>> npages < 0 which is unnecessary.
>>>>> @@ -328,6 +328,9 @@ void unpin_user_pages(struct page **pages, unsigned long npages)
>>>>>   {
>>>>>        unsigned long index;
>>>>>
>>>>> +     if (WARN_ON_ONCE(npages < 0))
>>>>> +             return;
>>>>
>>>> But npages is unsigned long.  So it can't be less than zero.
>>>
>>> Sorry, I missed it.
>>>
>>> Then, it means if npages is assigned with -ERRNO by caller, unpin_user_pages()
>>> may end up calling a big loop, which is unnecessary.
>>
>> How will a caller allocate memory of the right size and still manage
>> to call with the wrong npages? Do you have an example of a broken caller?
> 
> These are two broken callers which might end up calling unpin_user_pages()
> with -ERRNO.
> drivers/rapidio/devices/rio_mport_cdev.c#L952
> drivers/misc/mic/scif/scif_rma.c#L1399
> 
> They both are in error handling paths, so might not have any serious impact.
> But theoretically possible.
> 

Eventually, I settled on fixing up the callers so that they match the gup/pup
API better. In other words, gup/pup has signed int for both input and return
value, and the callers need to handle that perfectly.

So let's fix up the callers.

thanks,
-- 
John Hubbard
NVIDIA