Re: [regression] cpuset,mm: update tasks' mems_allowed in time(58568d2)

[David Rientjes]

On Wed, 24 Feb 2010, Miao Xie wrote:

> >> Sorry, Could you explain what you advised?
> >> I think it is hard to fix this problem by adding a variant, because it is
> >> hard to avoid loading a word of the mask before
> >>
> >> 	nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
> >>
> >> and then loading another word of the mask after
> >>
> >> 	tsk->mems_allowed = *newmems;
> >>
> >> unless we use lock.
> >>
> >> Maybe we need a rw-lock to protect task->mems_allowed.
> >>
> > 
> > I meant that we need to define synchronization only for configurations 
> > that do not do atomic nodemask_t stores, it's otherwise unnecessary.  
> > We'll need to load and store tsk->mems_allowed via a helper function that 
> > is defined to take the rwlock for such configs and only read/write the 
> > nodemask for others.
> > 
> 
> By investigating, we found that it is hard to guarantee the consistent between
> mempolicy and mems_allowed because mempolicy was designed as a self-update function.
> it just can be changed by one's self. Maybe we must change the implement of mempolicy.
> 

Before your change, cpuset nodemask changes were serialized on 
manage_mutex which would, in turn, serialize the rebinding of each 
attached task's mempolicy.  update_nodemask() is now serialized on 
cgroup_lock(), which also protects scan_for_empty_cpusets(), so the cpuset 
code protects it adequately.  If a concurrent mempolicy change from a 
user's set_mempolicy() happens, however, it could introduce an 
inconsistency between them.

If we protect current->mems_allowed with a rwlock or seqlock for configs 
where MAX_NUMNODES > BITS_PER_LONG, then we can always guarantee that we 
get the entire nodemask.  The same problem is present for 
current->cpus_allowed, however, with NR_CPUS > BITS_PER_LONG.  We must be 
able to safely dereference both masks without the chance of returning 
nodes_empty() or cpus_empty().
--
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/