Does iommu_init_table need to use GFP_ATOMIC allocations?

[Nishanth Aravamudan <nacc@linux.vnet.ibm.com>]

Under heavy (DLPAR?) stress, we tripped this panic() in
arch/powerpc/kernel/iommu.c::iommu_init_table():

        page = alloc_pages_node(nid, GFP_ATOMIC, get_order(sz));
        if (!page)
                panic("iommu_init_table: Can't allocate %ld bytes\n",
sz);

Before the panic() we got a page allocation failure for an order-2
allocation. There appears to be memory free, but perhaps not in the
ATOMIC context. I looked through all the call-sites of
iommu_init_table() and didn't see any obvious reason to need an ATOMIC
allocation. Most call-sites in fact have an explicit GFP_KERNEL
allocation shortly before the call to iommu_init_table(), indicating we
are not in an atomic context. There is some indirection for some paths,
but I didn't see any locks indicating that GFP_KERNEL is inappropriate.
Does anyone know if/why ATOMIC allocations are necessary here?

Thanks,
Nish