Re: [patch -mm v2] mm: introduce oom_adj_child

[David Rientjes]

On Sat, 1 Aug 2009, KAMEZAWA Hiroyuki wrote:

> Summarizing I think now .....
>   - rename mm->oom_adj as mm->effective_oom_adj
>   - re-add per-thread oom_adj
>   - update mm->effective_oom_adj based on per-thread oom_adj
>   - if necessary, plz add read-only /proc/pid/effective_oom_adj file.
>     or show 2 values in /proc/pid/oom_adj
>   - rewrite documentation about oom_score.
>    " it's calclulated from  _process's_ memory usage and oom_adj of
>     all threads which shares a memor  context".
>    This behavior is not changed from old implemtation, anyway.
>  - If necessary, rewrite oom_kill itself to scan only thread group
>    leader. It's a way to go regardless of  vfork problem.
> 

Ok, so you've abandoned the signal_struct proposal and now want to add it 
back to task_struct with an effective member in mm_struct by changing the 
documentation.  Hmm.

This solves the livelock problem by adding additional tunables, but 
doesn't match how the documentation describes the use case for 
/proc/pid/oom_adj.  Your argument is that the behavior of that value can't 
change: that it must be per-thread.  And that allowance leads to one of 
two inconsistent scenarios:

 - /proc/pid/oom_score is inconsistent when tuning /proc/pid/oom_adj if it
   relies on the per-thread oom_adj; it now really represents nothing but
   an incorrect value if other threads share that memory and misleads the
   user on how the oom killer chooses victims, or

 - /proc/pid/oom_score is inconsistent when the thread that set the
   effective per-mm oom_adj exits and it is now obsolete since you have
   no way to determine what the next effective oom_adj value shall be.

Determining the next effective per-mm oom_adj isn't possible when the only 
threads sharing the mm remaining have different per-thread oom_adj values.  
That's a horribly inconsistent state to be getting into because it allows 
oom_score to change when a thread exits, which is completely unknown to 
userspace, OR is allows the effective per-mm oom_adj to be different from 
all threads sharing the same memory (and, thus, /proc/pid/oom_score not 
being representative of any thread's /proc/pid/oom_adj).

> I think documentation is wrong. It should say "you should think of
> multi-thread effect to oom_adj/oom_score".
> 

It's more likely than not that applications were probably written to the 
way the documentation described the two files: that is, adjust 
/proc/pid/oom_score by tuning /proc/pid/oom_adj instead of relying on an 
undocumented implementation detail concerning the tuning of oom_adj for a 
vfork'd child prior to exec().  The user is probably unaware of the oom 
killer's implementation and simply interprets a higher oom_score as a more 
likely candidate for oom kill.  My patches preserve that in all scenarios 
without altering the documentation or adding additional files that would 
be required to leave the oom_adj value itself in an inconsistent state as 
you propose.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/