Re: [PATCH] KVM: s390: remove delayed reallocation of page tables for KVM

[Alexander Graf <agraf@suse.de>]

On 04/27/2015 03:57 PM, Martin Schwidefsky wrote:
> On Mon, 27 Apr 2015 15:48:42 +0200
> Alexander Graf <agraf@suse.de> wrote:
>
>> On 04/23/2015 02:13 PM, Martin Schwidefsky wrote:
>>> On Thu, 23 Apr 2015 14:01:23 +0200
>>> Alexander Graf <agraf@suse.de> wrote:
>>>
>>>> As far as alternative approaches go, I don't have a great idea otoh.
>>>> We could have an elf flag indicating that this process needs 4k page
>>>> tables to limit the impact to a single process. In fact, could we
>>>> maybe still limit the scope to non-global? A personality may work
>>>> as well. Or ulimit?
>>> I tried the ELF flag approach, does not work. The trouble is that
>>> allocate_mm() has to create the page tables with 4K tables if you
>>> want to change the page table layout later on. We have learned the
>>> hard way that the direction 2K to 4K does not work due to races
>>> in the mm.
>>>
>>> Now there are two major cases: 1) fork + execve and 2) fork only.
>>> The ELF flag can be used to reduce from 4K to 2K for 1) but not 2).
>>> 2) is required for apps that use lots of forking, e.g. database or
>>> web servers. Same goes for the approach with a personality flag or
>>> ulimit.
>>>
>>> We would have to distinguish the two cases for allocate_mm(),
>>> if the new mm is allocated for a fork the current mm decides
>>> 2K vs. 4K. If the new mm is allocated by binfmt_elf, then start
>>> with 4K and do the downgrade after the ELF flag has been evaluated.
>> Well, you could also make it a personality flag for example, no? Then
>> every new process below a certain one always gets 4k page tables until
>> they drop the personality, at which point each child would only get 2k
>> page tables again.
>>
>> I'm mostly concerned that people will end up mixing VMs and other
>> workloads on the same LPAR, so I don't think there's a one-shoe-fits-all
>> solution.
> If I add an argument to mm_init() to indicate if this context
> is for fork() or execve() then the ELF header flag approach works.

So you don't need the sysctl?


Alex