[QUESTION] Is MPOL_F_MOF user visible?

[QUESTION] Is MPOL_F_MOF user visible?

[Yang Shi]

Hi folks,


When reading the mempolicy code, I got confused by MPOL_F_MOF flag. It 
is defined in include/uapi/linux/mempolicy.h, so it looks visible to the 
users. But, man page doesn't mention it at all. And, the code in 
do_set_mempolicy() -> mpol_new() doesn't set it. It looks it is just set 
by two places:

     - NUMA default policy (preferred_node_policy)

     - When MPOL_MF_LAZY is passed in. But, it is not configurable from 
user since it is not valid MF


So, actually it can't be set by user with set_mempolicy()/mbind() APIs, 
right? As long as the process' or vmas' policy is changed to non-default 
one (i.e. MPOL_BIND), those processes or vmas are *not* eligible for 
migrating with NUMA balancing anymore?


Thanks,

Yang

Re: [QUESTION] Is MPOL_F_MOF user visible?

[Mel Gorman]

On Fri, Mar 08, 2019 at 10:22:59PM -0800, Yang Shi wrote:
> Hi folks,
> 
> 
> When reading the mempolicy code, I got confused by MPOL_F_MOF flag. It is
> defined in include/uapi/linux/mempolicy.h, so it looks visible to the users.
> But, man page doesn't mention it at all. And, the code in do_set_mempolicy()
> -> mpol_new() doesn't set it. It looks it is just set by two places:
> 
>     - NUMA default policy (preferred_node_policy)
> 
>     - When MPOL_MF_LAZY is passed in. But, it is not configurable from user
> since it is not valid MF
> 

It was never exported to userspace because it was not clear how the
policy would be used sensibly outside the context of the default policy.

> So, actually it can't be set by user with set_mempolicy()/mbind() APIs,
> right? As long as the process' or vmas' policy is changed to non-default one
> (i.e. MPOL_BIND), those processes or vmas are *not* eligible for migrating
> with NUMA balancing anymore?
> 

Correct because if the policy is MPOL_BIND, it's not defined how lazy
migration should behave.

-- 
Mel Gorman
SUSE Labs