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

[KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>]

> On Fri, 31 Jul 2009, KOSAKI Motohiro wrote:
> 
> > diff --git a/fs/proc/base.c b/fs/proc/base.c
> > index 3ce5ae9..c64499e 100644
> > --- a/fs/proc/base.c
> > +++ b/fs/proc/base.c
> > @@ -1008,7 +1008,7 @@ static ssize_t oom_adjust_read(struct file *file, char __user *buf,
> >  		return -ESRCH;
> >  	task_lock(task);
> >  	if (task->mm)
> > -		oom_adjust = task->mm->oom_adj;
> > +		oom_adjust = task->signal->oom_adj;
> >  	else
> >  		oom_adjust = OOM_DISABLE;
> >  	task_unlock(task);
> 
> This may display a /proc/pid/oom_adj that is radically different from 
> task->mm->oom_adj_cached without knowledge to userspace and you can't 
> simply display task->mm>oom_adj_cached here because it gets reset on every 
> write to /proc/pid/oom_adj.

Is this necessary?
different value was only happen when vfork process change oom_adj. but
vfork()ed process call exec() soon.
userland monitor program don't confuse.

Plus, if you can post any troublesome testcase, I can fix it.
this is not essential code in this patch.



> > @@ -1046,12 +1046,13 @@ static ssize_t oom_adjust_write(struct file *file, const char __user *buf,
> >  		put_task_struct(task);
> >  		return -EINVAL;
> >  	}
> > -	if (oom_adjust < task->mm->oom_adj && !capable(CAP_SYS_RESOURCE)) {
> > +	if (oom_adjust < task->signal->oom_adj && !capable(CAP_SYS_RESOURCE)) {
> >  		task_unlock(task);
> >  		put_task_struct(task);
> >  		return -EACCES;
> >  	}
> > -	task->mm->oom_adj = oom_adjust;
> > +	task->signal->oom_adj = oom_adjust;
> > +	task->mm->oom_adj_cached = OOM_CACHE_DEFAULT;
> >  	task_unlock(task);
> >  	put_task_struct(task);
> >  	if (end - buffer == 0)
> > diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> > index 7acc843..f93f97f 100644
> > --- a/include/linux/mm_types.h
> > +++ b/include/linux/mm_types.h
> > @@ -240,7 +240,8 @@ struct mm_struct {
> >  
> >  	unsigned long saved_auxv[AT_VECTOR_SIZE]; /* for /proc/PID/auxv */
> >  
> > -	s8 oom_adj;	/* OOM kill score adjustment (bit shift) */
> > +	s8 oom_adj_cached;	/* mirror from signal_struct->oom_adj.
> > +				   in vfork case, multiple processes use the same mm. */
> >  
> >  	cpumask_t cpu_vm_mask;
> >  
> > diff --git a/include/linux/oom.h b/include/linux/oom.h
> > index a7979ba..a219480 100644
> > --- a/include/linux/oom.h
> > +++ b/include/linux/oom.h
> > @@ -3,6 +3,7 @@
> >  
> >  /* /proc/<pid>/oom_adj set to -17 protects from the oom-killer */
> >  #define OOM_DISABLE (-17)
> > +#define OOM_CACHE_DEFAULT (15)
> >  /* inclusive */
> >  #define OOM_ADJUST_MIN (-16)
> >  #define OOM_ADJUST_MAX 15
> > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > index 3ab08e4..e10b12b 100644
> > --- a/include/linux/sched.h
> > +++ b/include/linux/sched.h
> > @@ -629,6 +629,8 @@ struct signal_struct {
> >  	unsigned audit_tty;
> >  	struct tty_audit_buf *tty_audit_buf;
> >  #endif
> > +
> > +	s8 oom_adj;	/* OOM kill score adjustment (bit shift) */
> >  };
> >  
> >  /* Context switch must be unlocked if interrupts are to be enabled */
> 
> I don't believe oom_adj is an appropriate use of signal_struct, sorry.

Real world and life aren't so simple... and linux too.
At least, regression is definitely worst result. any other way
is better than it.


> 
> > diff --git a/kernel/exit.c b/kernel/exit.c
> > index 869dc22..c741a45 100644
> > --- a/kernel/exit.c
> > +++ b/kernel/exit.c
> > @@ -688,6 +689,7 @@ static void exit_mm(struct task_struct * tsk)
> >  	enter_lazy_tlb(mm, current);
> >  	/* We don't want this task to be frozen prematurely */
> >  	clear_freeze_flag(tsk);
> > +	mm->oom_adj_cached = OOM_CACHE_DEFAULT;
> >  	task_unlock(tsk);
> >  	mm_update_next_owner(mm);
> >  	mmput(mm);
> 
> This is similiar to an early proposal that wanted to keep an array of 
> oom_adj values for tasks attached to the mm in mm_struct.  The problem is 
> that you're obviously losing information about all threads attached to the 
> mm any time one of the threads exits or writes to /proc/pid/oom_adj.  That 
> information can only be regenerated with a tasklist scan.

I agree this. but I can't imazine this darkside. Can you please explain this?
maybe, we can improve this implementaion obviously. reset at each thread exitting
is caused by my lazy and small patch preference.
a bit line adding can change this.

> 
> > diff --git a/kernel/fork.c b/kernel/fork.c
> > index 9b42695..b7cb474 100644
> > --- a/kernel/fork.c
> > +++ b/kernel/fork.c
> > @@ -426,6 +427,7 @@ static struct mm_struct * mm_init(struct mm_struct * mm, struct task_struct *p)
> >  	init_rwsem(&mm->mmap_sem);
> >  	INIT_LIST_HEAD(&mm->mmlist);
> >  	mm->flags = (current->mm) ? current->mm->flags : default_dump_filter;
> > +	mm->oom_adj_cached = OOM_CACHE_DEFAULT;
> >  	mm->core_state = NULL;
> >  	mm->nr_ptes = 0;
> >  	set_mm_counter(mm, file_rss, 0);
> > diff --git a/mm/oom_kill.c b/mm/oom_kill.c
> > index 175a67a..eae2d78 100644
> > --- a/mm/oom_kill.c
> > +++ b/mm/oom_kill.c
> > @@ -58,7 +58,7 @@ unsigned long badness(struct task_struct *p, unsigned long uptime)
> >  	unsigned long points, cpu_time, run_time;
> >  	struct mm_struct *mm;
> >  	struct task_struct *child;
> > -	int oom_adj;
> > +	s8 oom_adj;
> >  
> >  	task_lock(p);
> >  	mm = p->mm;
> > @@ -66,7 +66,10 @@ unsigned long badness(struct task_struct *p, unsigned long uptime)
> >  		task_unlock(p);
> >  		return 0;
> >  	}
> > -	oom_adj = mm->oom_adj;
> > +
> > +	if (mm->oom_adj_cached < p->signal->oom_adj)
> > +		mm->oom_adj_cached = p->signal->oom_adj;
> 
> This conditional will never be true since mm->oom_adj_cached is 
> initialized to 15, which is the upper bound on which p->signal->oom_adj 
> may ever be, so mm->oom_adj_cached never gets changed from 
> OOM_CACHE_DEFAULT.

Ah, good catch. this condition shold be reserve. thanks.


> Thus, this patch doesn't even work, and you probably would have noticed 
> that if you'd checked /proc/pid/oom_score for any pid.
> 
> Even if mm->oom_adj_cached _was_ properly updated here, 
> /proc/pid/oom_score would be out of sync with more negative oom_adj values 
> for threads sharing the mm_struct since it calls badness() for only a 
> single thread.

Hm, I don't think this is serious problem. but I can fix it easily
because badness() isn't fast-path.



> 
> > +	oom_adj = mm->oom_adj_cached;
> >  	if (oom_adj == OOM_DISABLE) {
> >  		task_unlock(p);
> >  		return 0;
> > @@ -350,7 +354,7 @@ static int oom_kill_task(struct task_struct *p)
> >  
> >  	task_lock(p);
> >  	mm = p->mm;
> > -	if (!mm || mm->oom_adj == OOM_DISABLE) {
> > +	if (!mm || p->signal->oom_adj == OOM_DISABLE) {
> >  		task_unlock(p);
> >  		return 1;
> >  	}