Re: [patch][resend] MAP_HUGETLB munmap fails with size not 2MB aligned

[Vlastimil Babka <vbabka-AlSwsSmVLrQ@public.gmane.org>]

On 03/26/2015 08:39 PM, Davide Libenzi wrote:
> On Thu, 26 Mar 2015, David Rientjes wrote:
> 
>> Yes, this munmap() behavior of lengths <= hugepage_size - PAGE_SIZE for a 
>> hugetlb vma is long standing and there may be applications that break as a 
>> result of changing the behavior: a database that reserves all allocated 
>> hugetlb memory with mmap() so that it always has exclusive access to those 
>> hugepages, whether they are faulted or not, and maintains its own hugepage 
>> pool (which is common), may test the return value of munmap() and depend 
>> on it returning -EINVAL to determine if it is freeing memory that was 
>> either dynamically allocated or mapped from the hugetlb reserved pool.
> 
> You went a long way to create such a case.
> But, in your case, that application will erroneously considering hugepage 
> mmaped memory, as dynamically allocated, since it will always get EINVAL, 
> unless it passes an aligned size. Aligned size, which a fix like the one 
> posted in the patch will still leave as success.
> OTOH, an application, which might be more common than the one you posted,
> which calls munmap() to release a pointer which it validly got from a 
> previous mmap(), will leak huge pages as all the issued munmaps will fail.
> 
> 
>> If we were to go back in time and decide this when the munmap() behavior 
>> for hugetlb vmas was originally introduced, that would be valid.  The 
>> problem is that it could lead to userspace breakage and that's a 
>> non-starter.
>> 
>> What we can do is improve the documentation and man-page to clearly 
>> specify the long-standing behavior so that nobody encounters unexpected 
>> results in the future.
> 
> This way you will leave the mmap API with broken semantics.
> In any case, I am done arguing.
> I will leave to Andrew to sort it out, and to Michael Kerrisk to update 
> the mmap man pages with the new funny behaviour.

+ CC's

You know that people don't always magically CC themselves, or read all of
lkml/linux-mm? :)

> 
> 
> - Davide
> 
> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo-Bw31MaZKKs0EbZ0PF+XxCw@public.gmane.org  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont-Bw31MaZKKs3YtjvyW6yDsg@public.gmane.org"> email-Bw31MaZKKs3YtjvyW6yDsg@public.gmane.org </a>
>