On Tue, Nov 15, 2016 at 11:31:06AM +0530, Aneesh Kumar K.V wrote:

Huang Shijie [off-list ref] writes:

quoted

 #ifdef CONFIG_HUGETLBFS
 /*
+ * huge_nodemask(@vma, @addr)
+ * @vma: virtual memory area whose policy is sought
+ * @addr: address in @vma for shared policy lookup and interleave policy
+ *
+ * If the effective policy is BIND, returns a pointer to the mempolicy's
+ * @nodemask.
+ */
+nodemask_t *huge_nodemask(struct vm_area_struct *vma, unsigned long addr)
+{
+	nodemask_t *nodes_mask = NULL;
+	struct mempolicy *mpol = get_vma_policy(vma, addr);
+
+	if (mpol->mode == MPOL_BIND)
+		nodes_mask = &mpol->v.nodes;
+	mpol_cond_put(mpol);

What if it is MPOL_PREFERED or MPOL_INTERLEAVE ? we don't honor node
mask in that case ?

I wrote this code by following the logic in the huge_zonelist().
So I ignored the support for MPOL_PREFERED/MPOL_INTERLEAVE.

IMHO, it is okay to allocate a gigantic page with MPOL_PREFERED/MPOL_BIND.
But I am not sure if we can allocate a gigantic page with MPOL_INTERLEAVE, since
since the gigantic page's order is bigger then MAX_ORDER.

Could you give me some advice about this?

Thanks
Huang Shijie