jianhai / rpms / kernel

Forked from rpms/kernel 3 years ago
Clone
e293be
centosplus patch bug#15115
e293be
e293be
e293be
commit 6d8409580bee356ce418dcb94260b24dda639934
e293be
Author: Michal Hocko <mhocko@suse.com>
e293be
Date:   Mon Dec 12 16:42:23 2016 -0800
e293be
e293be
    mm, mempolicy: clean up __GFP_THISNODE confusion in policy_zonelist
e293be
    
e293be
    __GFP_THISNODE is documented to enforce the allocation to be satisified
e293be
    from the requested node with no fallbacks or placement policy
e293be
    enforcements.  policy_zonelist seemingly breaks this semantic if the
e293be
    current policy is MPOL_MBIND and instead of taking the node it will
e293be
    fallback to the first node in the mask if the requested one is not in
e293be
    the mask.  This is confusing to say the least because it fact we
e293be
    shouldn't ever go that path.  First tasks shouldn't be scheduled on CPUs
e293be
    with nodes outside of their mempolicy binding.  And secondly
e293be
    policy_zonelist is called only from 3 places:
e293be
    
e293be
     - huge_zonelist - never should do __GFP_THISNODE when going this path
e293be
    
e293be
     - alloc_pages_vma - which shouldn't depend on __GFP_THISNODE either
e293be
    
e293be
     - alloc_pages_current - which uses default_policy id __GFP_THISNODE is
e293be
       used
e293be
    
e293be
    So we shouldn't even need to care about this possibility and can drop
e293be
    the confusing code.  Let's keep a WARN_ON_ONCE in place to catch
e293be
    potential users and fix them up properly (aka use a different allocation
e293be
    function which ignores mempolicy).
e293be
    
e293be
    [akpm@linux-foundation.org: coding-style fixes]
e293be
    Link: http://lkml.kernel.org/r/20161013125958.32155-1-mhocko@kernel.org
e293be
    Signed-off-by: Michal Hocko <mhocko@suse.com>
e293be
    Acked-by: Vlastimil Babka <vbabka@suse.cz>
e293be
    Cc: Mel Gorman <mgorman@suse.de>
e293be
    Cc: David Rientjes <rientjes@google.com>
e293be
    Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
e293be
    Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
e293be
    Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
e293be
e293be
    Applied-by: Akemi Yagi <toracat@centos.org>
e293be
e293be
diff --git a/mm/mempolicy.c b/mm/mempolicy.c
e293be
index a6a27e5..4d58021 100644
e293be
--- a/mm/mempolicy.c
e293be
+++ b/mm/mempolicy.c
e293be
@@ -1679,25 +1679,17 @@ static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
e293be
 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
e293be
 	int nd)
e293be
 {
e293be
-	switch (policy->mode) {
e293be
-	case MPOL_PREFERRED:
e293be
-		if (!(policy->flags & MPOL_F_LOCAL))
e293be
-			nd = policy->v.preferred_node;
e293be
-		break;
e293be
-	case MPOL_BIND:
e293be
+	if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
e293be
+		nd = policy->v.preferred_node;
e293be
+	else {
e293be
 		/*
e293be
-		 * Normally, MPOL_BIND allocations are node-local within the
e293be
-		 * allowed nodemask.  However, if __GFP_THISNODE is set and the
e293be
-		 * current node isn't part of the mask, we use the zonelist for
e293be
-		 * the first node in the mask instead.
e293be
+		 * __GFP_THISNODE shouldn't even be used with the bind policy
e293be
+		 * because we might easily break the expectation to stay on the
e293be
+		 * requested node and not break the policy.
e293be
 		 */
e293be
-		if (unlikely(gfp & __GFP_THISNODE) &&
e293be
-				unlikely(!node_isset(nd, policy->v.nodes)))
e293be
-			nd = first_node(policy->v.nodes);
e293be
-		break;
e293be
-	default:
e293be
-		BUG();
e293be
+		WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
e293be
 	}
e293be
+
e293be
 	return node_zonelist(nd, gfp);
e293be
 }
e293be