Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Wed, 4 Sep 2019 12:54:25 -0700 (PDT)
From:   David Rientjes <rientjes@google.com>
To:     Linus Torvalds <torvalds@linux-foundation.org>,
        Andrew Morton <akpm@linux-foundation.org>
cc:     Andrea Arcangeli <aarcange@redhat.com>,
        Michal Hocko <mhocko@suse.com>, Mel Gorman <mgorman@suse.de>,
        Vlastimil Babka <vbabka@suse.cz>,
        "Kirill A. Shutemov" <kirill@shutemov.name>,
        linux-kernel@vger.kernel.org, linux-mm@kvack.org
Subject: [rfc 4/4] mm, page_alloc: allow hugepage fallback to remote nodes
 when madvised
Message-ID: <alpine.DEB.2.21.1909041253560.94813@chino.kir.corp.google.com>
User-Agent: Alpine 2.21 (DEB 202 2017-01-01)
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

For systems configured to always try hard to allocate transparent
hugepages (thp defrag setting of "always") or for memory that has been
explicitly madvised to MADV_HUGEPAGE, it is often better to fallback to
remote memory to allocate the hugepage if the local allocation fails
first.

The point is to allow the initial call to __alloc_pages_node() to attempt
to defragment local memory to make a hugepage available, if possible,
rather than immediately fallback to remote memory.  Local hugepages will
always have a better access latency than remote (huge)pages, so an attempt
to make a hugepage available locally is always preferred.

If memory compaction cannot be successful locally, however, it is likely
better to fallback to remote memory.  This could take on two forms: either
allow immediate fallback to remote memory or do per-zone watermark checks.
It would be possible to fallback only when per-zone watermarks fail for
order-0 memory, since that would require local reclaim for all subsequent
faults so remote huge allocation is likely better than thrashing the local
zone for large workloads.

In this case, it is assumed that because the system is configured to try
hard to allocate hugepages or the vma is advised to explicitly want to try
hard for hugepages that remote allocation is better when local allocation
and memory compaction have both failed.

Signed-off-by: David Rientjes <rientjes@google.com>
---
 mm/mempolicy.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/mm/mempolicy.c b/mm/mempolicy.c
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -2133,6 +2133,17 @@ alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
 			mpol_cond_put(pol);
 			page = __alloc_pages_node(hpage_node,
 						gfp | __GFP_THISNODE, order);
+
+			/*
+			 * If hugepage allocations are configured to always
+			 * synchronous compact or the vma has been madvised
+			 * to prefer hugepage backing, retry allowing remote
+			 * memory as well.
+			 */
+			if (!page && (gfp & __GFP_DIRECT_RECLAIM))
+				page = __alloc_pages_node(hpage_node,
+						gfp | __GFP_NORETRY, order);
+
 			goto out;
 		}
 	}