lkml.org 
[lkml]   [2010]   [Apr]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 1/2] mm: fix bugs of mpol_rebind_nodemask()
On Thu, 29 Apr 2010, Miao Xie wrote:

> > That's been the behavior for at least three years so changing it from
> > under the applications isn't acceptable, see
> > Documentation/vm/numa_memory_policy.txt regarding mempolicy rebinds and
> > the two flags that are defined that can be used to adjust the behavior.
>
> Is the flags what you said MPOL_F_STATIC_NODES and MPOL_F_RELATIVE_NODES?
> But the codes that I changed isn't under MPOL_F_STATIC_NODES or MPOL_F_RELATIVE_NODES.
> The documentation doesn't say what we should do if either of these two flags is not set.
>

MPOL_F_STATIC_NODES and MPOL_F_RELATIVE_NODES allow you to adjust the
behavior of the rebind: the former requires specific nodes to be assigned
to the mempolicy and could suppress the rebind completely, if necessary;
the latter ensures the mempolicy nodemask has a certain weight as nodes
are assigned in a round-robin manner. The behavior that you're referring
to is provided via MPOL_F_RELATIVE_NODES, which guarantees whatever weight
is passed via set_mempolicy() will be preserved when mems are added to a
cpuset.
Regardless of whether the behavior is documented when either flag is
passed, we can't change the long-standing default behavior that people use
when their cpuset mems are rebound: we can only extend the functionality
and the behavior you're seeking is already available with a
MPOL_F_RELATIVE_NODES flag modifier.

> Furthermore, in order to fix no node to alloc memory, when we want to update mempolicy
> and mems_allowed, we expand the set of nodes first (set all the newly nodes) and
> shrink the set of nodes lazily(clean disallowed nodes).

That's a cpuset implementation choice, not a mempolicy one; mempolicies
have nothing to do with an empty current->mems_allowed.

> But remap() breaks the expanding, so if we don't remove remap(), the problem can't be
> fixed. Otherwise, cpuset has to do the rebinding by itself and the code is ugly.
> Like this:
>
> static void cpuset_change_task_nodemask(struct task_struct *tsk, nodemask_t *newmems)
> {
> nodemask_t tmp;
> ...
> /* expand the set of nodes */
> if (!mpol_store_user_nodemask(tsk->mempolicy)) {
> nodes_remap(tmp, ...);
> nodes_or(tsk->mempolicy->v.nodes, tsk->mempolicy->v.nodes, tmp);
> }
> ...
>
> /* shrink the set of nodes */
> if (!mpol_store_user_nodemask(tsk->mempolicy))
> tsk->mempolicy->v.nodes = tmp;
> }
>

I don't see why this is even necessary, the mempolicy code could simply
return numa_node_id() when nodes_empty(current->mempolicy->v.nodes) to
close the race.

[ Your pseudo-code is also lacking task_lock(tsk), which is required to
safely dereference tsk->mempolicy, and this is only available so far in
-mm since the oom killer rewrite. ]


\
 
 \ /
  Last update: 2010-04-30 20:53    [from the cache]
©2003-2011 Jasper Spaans