mm: compaction: fix endless looping over same migrate block

During stress testing, the following situation was observed:

     70 root      39  19       0      0      0 R 100.0   0.0 959:29.92 khugepaged
 310936 root      20   0   84416  25620    512 R  99.7   1.5 642:37.22 hugealloc

Tracing shows isolate_migratepages_block() endlessly looping over the
first block in the DMA zone:

       hugealloc-310936  [001] ..... 237297.415718: mm_compaction_finished: node=0 zone=DMA      order=9 ret=no_suitable_page
       hugealloc-310936  [001] ..... 237297.415718: mm_compaction_isolate_migratepages: range=(0x1 ~ 0x400) nr_scanned=513 nr_taken=0
       hugealloc-310936  [001] ..... 237297.415718: mm_compaction_finished: node=0 zone=DMA      order=9 ret=no_suitable_page
       hugealloc-310936  [001] ..... 237297.415718: mm_compaction_isolate_migratepages: range=(0x1 ~ 0x400) nr_scanned=513 nr_taken=0
       hugealloc-310936  [001] ..... 237297.415718: mm_compaction_finished: node=0 zone=DMA      order=9 ret=no_suitable_page
       hugealloc-310936  [001] ..... 237297.415718: mm_compaction_isolate_migratepages: range=(0x1 ~ 0x400) nr_scanned=513 nr_taken=0
       hugealloc-310936  [001] ..... 237297.415718: mm_compaction_finished: node=0 zone=DMA      order=9 ret=no_suitable_page
       hugealloc-310936  [001] ..... 237297.415718: mm_compaction_isolate_migratepages: range=(0x1 ~ 0x400) nr_scanned=513 nr_taken=0

The problem is that the functions tries to test and set the skip bit once
on the block, to avoid skipping on its own skip-set, using
pageblock_aligned() on the pfn as a test.  But because this is the DMA
zone which starts at pfn 1, this is never true for the first block, and
the skip bit isn't set or tested at all.  As a result,
fast_find_migrateblock() returns the same pageblock over and over.

If the pfn isn't pageblock-aligned, also check if it's the start of the
zone to ensure test-and-set-exactly-once on unaligned ranges.

Thanks to Vlastimil Babka for the help in debugging this.

Link: https://lkml.kernel.org/r/20230731172450.1632195-1-hannes@cmpxchg.org
Fixes: 90ed667 ("Revert "Revert "mm/compaction: fix set skip in fast_find_migrateblock""")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm/compaction.c

@@ -912,11 +912,12 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 
 		/*
 		 * Check if the pageblock has already been marked skipped.
-		 * Only the aligned PFN is checked as the caller isolates
+		 * Only the first PFN is checked as the caller isolates
 		 * COMPACT_CLUSTER_MAX at a time so the second call must
 		 * not falsely conclude that the block should be skipped.
 		 */
-		if (!valid_page && pageblock_aligned(low_pfn)) {
+		if (!valid_page && (pageblock_aligned(low_pfn) ||
+				    low_pfn == cc->zone->zone_start_pfn)) {
 			if (!isolation_suitable(cc, page)) {
 				low_pfn = end_pfn;
 				folio = NULL;
@@ -2002,7 +2003,8 @@ static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
 		 * before making it "skip" so other compaction instances do
 		 * not scan the same block.
 		 */
-		if (pageblock_aligned(low_pfn) &&
+		if ((pageblock_aligned(low_pfn) ||
+		     low_pfn == cc->zone->zone_start_pfn) &&
 		    !fast_find_block && !isolation_suitable(cc, page))
 			continue;