Re: [PATCH v3] mm: page_alloc: make defrag_mode retries follow the promoted order

[Kiryl Shutsemau]

On Fri, Oct 23, 2026 at 12:09:07PM +0530, Dev Jain wrote:
> >> I ran a synthetic benchmark faulting in lots of 64k mTHP with swap enabled,
> >> defrag mode caused me to enter a warn_alloc (and 8 times more compact_stalls),
> >> which didn't happen without defrag.

Thanks for the benchmark. It reproduces on x86 in the same shape.

I think I found the problem.

Under pressure, the free order-4+ pages left are mostly in partly used
unmovable and reclaimable pageblocks. A movable request with
ALLOC_NOFRAGMENT cannot take them. But compaction_suit_allocation_order()
asks __zone_watermark_ok(), which counts free pages of every
migratetype, and returns COMPACT_SUCCESS without compacting anything.
The allocation then fails, and since compaction "succeeded", it is never
deferred.

With the diff below, compaction behaves the same with and without
defrag_mode. Medians of 15 runs of "compact-bench 99 64 16 20 0":

                             defrag_mode=0  defrag_mode=1  defrag_mode=1
                                                           + diff
    compact_stall                    2135          32476          2290
    compact_fail                      668          32330           606
    64K anon_fault_alloc            20689           6780         11804
    64K anon_fault_fallback           776          14502          9687

64K allocation is still lower with defrag_mode. I will look into this
further.

Could you try the diff on your setup?

diff --git a/mm/compaction.c b/mm/compaction.c
index a049415512c6..09be677fff36 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -2365,6 +2365,10 @@ static enum compact_result __compact_finished(struct compact_control *cc)
 			!free_area_empty(area, MIGRATE_CMA))
 			return COMPACT_SUCCESS;
 #endif
+		/* Under ALLOC_NOFRAGMENT it may only claim a whole block */
+		if ((cc->alloc_flags & ALLOC_NOFRAGMENT) && order < pageblock_order)
+			continue;
+
 		/*
 		 * Job done if allocation would steal freepages from
 		 * other migratetype buddy lists.
@@ -2510,6 +2514,43 @@ bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
 	return false;
 }
 
+/*
+ * The watermark check counts free pages of every migratetype.  Under
+ * ALLOC_NOFRAGMENT, a request below pageblock_order cannot take a page from a
+ * block of another migratetype unless it claims the whole block, so look for
+ * a free page the request can actually use.
+ */
+static bool compaction_free_page_usable(struct zone *zone, unsigned int order,
+					int migratetype, unsigned int alloc_flags)
+{
+	unsigned int o;
+	int mt;
+
+	if (!(alloc_flags & ALLOC_NOFRAGMENT) || order >= pageblock_order)
+		return true;
+
+	for (o = order; o < NR_PAGE_ORDERS; o++) {
+		struct free_area *area = &zone->free_area[o];
+
+		if (!free_area_empty(area, migratetype))
+			return true;
+#ifdef CONFIG_CMA
+		if ((alloc_flags & ALLOC_CMA) &&
+		    !free_area_empty(area, MIGRATE_CMA))
+			return true;
+#endif
+		/* From pageblock_order up, a free block can be claimed */
+		if (o < pageblock_order)
+			continue;
+		for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
+			if (!free_area_empty(area, mt))
+				return true;
+		}
+	}
+
+	return false;
+}
+
 /*
  * Should we do compaction for target allocation order.
  * Return COMPACT_SUCCESS if allocation for target order can be already
@@ -2520,7 +2561,7 @@ bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
 static enum compact_result
 compaction_suit_allocation_order(struct zone *zone, unsigned int order,
 				 int highest_zoneidx, unsigned int alloc_flags,
-				 bool async, bool kcompactd)
+				 int migratetype, bool async, bool kcompactd)
 {
 	unsigned long free_pages;
 	unsigned long watermark;
@@ -2532,7 +2573,8 @@ compaction_suit_allocation_order(struct zone *zone, unsigned int order,
 
 	watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
 	if (__zone_watermark_ok(zone, order, watermark, highest_zoneidx,
-				alloc_flags, free_pages))
+				alloc_flags, free_pages) &&
+	    compaction_free_page_usable(zone, order, migratetype, alloc_flags))
 		return COMPACT_SUCCESS;
 
 	/*
@@ -2588,6 +2630,7 @@ compact_zone(struct compact_control *cc, struct capture_control *capc)
 		ret = compaction_suit_allocation_order(cc->zone, cc->order,
 						       cc->highest_zoneidx,
 						       cc->alloc_flags,
+						       cc->migratetype,
 						       cc->mode == MIGRATE_ASYNC,
 						       !cc->direct_compaction);
 		if (ret != COMPACT_CONTINUE)
@@ -3073,7 +3116,7 @@ static bool kcompactd_node_suitable(pg_data_t *pgdat)
 		ret = compaction_suit_allocation_order(zone,
 				pgdat->kcompactd_max_order,
 				highest_zoneidx, alloc_flags,
-				false, true);
+				MIGRATE_MOVABLE, false, true);
 		if (ret == COMPACT_CONTINUE)
 			return true;
 	}
@@ -3116,7 +3159,7 @@ static void kcompactd_do_work(pg_data_t *pgdat)
 
 		ret = compaction_suit_allocation_order(zone,
 				cc.order, zoneid, cc.alloc_flags,
-				false, true);
+				MIGRATE_MOVABLE, false, true);
 		if (ret != COMPACT_CONTINUE)
 			continue;
 
-- 
  Kiryl Shutsemau / Kirill A. Shutemov