Re: linux-next: Tree for Nov 7

[Michael Ellerman]

Michal Hocko <mhocko@xxxxxxxxxx> writes:

> On Mon 13-11-17 22:34:50, Michael Ellerman wrote:
>> Hi Michal,
>> 
>> Michal Hocko <mhocko@xxxxxxxxxx> writes:
>> > On Mon 13-11-17 10:20:06, Michal Hocko wrote:
>> >> [Cc arm and ppc maintainers]
>> >
>> > Hmm, it turned out to be a problem on other architectures as well.
>> > CCing more maintainers. For your reference, we are talking about
>> > http://lkml.kernel.org/r/20171023082608.6167-1-mhocko@xxxxxxxxxx
>> > which has broken architectures which do apply aligning on the mmap
>> > address hint without MAP_FIXED applied. See below my proposed way
>> > around this issue because I belive that the above patch is quite
>> > valuable on its own to be dropped for all archs.
>> 
>> I don't really like your solution sorry :)  The fact that you've had to
>> patch seven arches seems like a red flag.
>> 
>> I think this is a generic problem with MAP_FIXED, which I've heard
>> userspace folks complain about in the past.
>
> The thing is that we canno  change MAP_FIXED behavior as it is carved in
> stone

Yes obviously. I didn't mean to imply we would change MAP_FIXED, rather
we would add a new flag with the new semantics.

>> Currently MAP_FIXED does two things:
>>   1. makes addr not a hint but the required address
>>   2. blasts any existing mapping
>> 
>> You want 1) but not 2).
>
> + fail if there is a clashing range

Yep. I thought that was implied :)

>> So the right solution IMHO would be to add a new mmap flag to request
>> that behaviour, ie. a fixed address but iff there is nothing already
>> mapped there.
>> 
>> I don't know the mm code well enough to know if that's hard for some
>> reason, but it *seems* like it should be doable.
>
> Yes, I have mentioned that in the previous email but the amount of code
> would be even larger. Basically every arch which reimplements
> arch_get_unmapped_area would have to special case new MAP_FIXED flag to
> do vma lookup.

I'd have to look, but my memory of the arch code is that it doesn't deal
with the vma so it wouldn't need any change.

> So this was the most simple solution I could come up
> with. If there was a general interest for MAP_FIXED_SAFE then we can
> introduce it later of course. I would just like the hardening merged
> sooner rather than later.

Sure. But in the scheme of things one more kernel release is not that
big a deal to get it right. Given that the simple approach of dropping
MAP_FIXED turns out to not be simple at all.

cheers