Re: [PATCH] mm: disable `vm.max_map_count' sysctl limit

[John Hubbard]

On 11/27/2017 11:52 AM, Michal Hocko wrote:
> On Mon 27-11-17 20:18:00, Mikael Pettersson wrote:
>> On Mon, Nov 27, 2017 at 11:12 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>>>> I've kept the kernel tunable to not break the API towards user-space,
>>>> but it's a no-op now.  Also the distinction between split_vma() and
>>>> __split_vma() disappears, so they are merged.
>>>
>>> Could you be more explicit about _why_ we need to remove this tunable?
>>> I am not saying I disagree, the removal simplifies the code but I do not
>>> really see any justification here.
>>
>> In principle you don't "need" to, as those that know about it can bump it
>> to some insanely high value and get on with life.  Meanwhile those that don't
>> (and I was one of them until fairly recently, and I'm no newcomer to Unix or
>> Linux) get to scratch their heads and wonder why the kernel says ENOMEM
>> when one has loads of free RAM.
> 
> I agree that our error reporting is more than suboptimal in this regard.
> These are all historical mistakes and we have much more of those. The
> thing is that we have means to debug these issues (check
> /proc/<pid>/maps e.g.).
> 
>> But what _is_ the justification for having this arbitrary limit?
>> There might have been historical reasons, but at least ELF core dumps
>> are no longer a problem.
> 
> Andi has already mentioned the the resource consumption. You can create
> a lot of unreclaimable memory and there should be some cap. Whether our
> default is good is questionable. Whether we can remove it altogether is
> a different thing.
> 
> As I've said I am not a great fan of the limit but "I've just notice it
> breaks on me" doesn't sound like a very good justification. You still
> have an option to increase it. Considering we do not have too many
> reports suggests that this is not such a big deal for most users.
> 

Let me add a belated report, then: we ran into this limit while implementing 
an early version of Unified Memory[1], back in 2013. The implementation
at the time depended on tracking that assumed "one allocation == one vma".
So, with only 64K vmas, we quickly ran out, and changed the design to work
around that. (And later, the design was *completely* changed to use a separate
tracking system altogether). 

The existing limit seems rather too low, at least from my perspective. Maybe
it would be better, if expressed as a function of RAM size?


[1] https://devblogs.nvidia.com/parallelforall/unified-memory-in-cuda-6/

    This is a way to automatically (via page faulting) migrate memory
    between CPUs and devices (GPUs, here). This is before HMM, of course.

thanks,
John Hubbard