mm, compaction: be selective about what pageblocks to clear skip hints
authorMel Gorman <mgorman@techsingularity.net>
Tue, 5 Mar 2019 23:45:38 +0000 (15:45 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 6 Mar 2019 05:07:17 +0000 (21:07 -0800)
Pageblock hints are cleared when compaction restarts or kswapd makes
enough progress that it can sleep but it's over-eager in that the bit is
cleared for migration sources with no LRU pages and migration targets
with no free pages.  As pageblock skip hint flushes are relatively rare
and out-of-band with respect to kswapd, this patch makes a few more
expensive checks to see if it's appropriate to even clear the bit.
Every pageblock that is not cleared will avoid 512 pages being scanned
unnecessarily on x86-64.

The impact is variable with different workloads showing small
differences in latency, success rates and scan rates.  This is expected
as clearing the hints is not that common but doing a small amount of
work out-of-band to avoid a large amount of work in-band later is
generally a good thing.

Link: http://lkml.kernel.org/r/20190118175136.31341-22-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Qian Cai <cai@lca.pw>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Carpenter <dan.carpenter@oracle.com>
Cc: David Rientjes <rientjes@google.com>
Cc: YueHaibing <yuehaibing@huawei.com>
[cai@lca.pw: no stuck in __reset_isolation_pfn()]
Link: http://lkml.kernel.org/r/20190206034732.75687-1-cai@lca.pw
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/mmzone.h
mm/compaction.c

index 842f9189537bb04d51285c3b237d9263273fc559..90c13cdeefb5de7b77d489357e6e096a35a81413 100644 (file)
@@ -480,6 +480,8 @@ struct zone {
        unsigned long           compact_cached_free_pfn;
        /* pfn where async and sync compaction migration scanner should start */
        unsigned long           compact_cached_migrate_pfn[2];
+       unsigned long           compact_init_migrate_pfn;
+       unsigned long           compact_init_free_pfn;
 #endif
 
 #ifdef CONFIG_COMPACTION
index b83cdb42f24969ba8a1ebb1cafbb0bc7e9da4a44..3084cee77fdab2ed856d7b523345e48e7e6cbec7 100644 (file)
@@ -237,6 +237,70 @@ static bool pageblock_skip_persistent(struct page *page)
        return false;
 }
 
+static bool
+__reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
+                                                       bool check_target)
+{
+       struct page *page = pfn_to_online_page(pfn);
+       struct page *end_page;
+       unsigned long block_pfn;
+
+       if (!page)
+               return false;
+       if (zone != page_zone(page))
+               return false;
+       if (pageblock_skip_persistent(page))
+               return false;
+
+       /*
+        * If skip is already cleared do no further checking once the
+        * restart points have been set.
+        */
+       if (check_source && check_target && !get_pageblock_skip(page))
+               return true;
+
+       /*
+        * If clearing skip for the target scanner, do not select a
+        * non-movable pageblock as the starting point.
+        */
+       if (!check_source && check_target &&
+           get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
+               return false;
+
+       /*
+        * Only clear the hint if a sample indicates there is either a
+        * free page or an LRU page in the block. One or other condition
+        * is necessary for the block to be a migration source/target.
+        */
+       block_pfn = pageblock_start_pfn(pfn);
+       pfn = max(block_pfn, zone->zone_start_pfn);
+       page = pfn_to_page(pfn);
+       if (zone != page_zone(page))
+               return false;
+       pfn = block_pfn + pageblock_nr_pages;
+       pfn = min(pfn, zone_end_pfn(zone));
+       end_page = pfn_to_page(pfn);
+
+       do {
+               if (pfn_valid_within(pfn)) {
+                       if (check_source && PageLRU(page)) {
+                               clear_pageblock_skip(page);
+                               return true;
+                       }
+
+                       if (check_target && PageBuddy(page)) {
+                               clear_pageblock_skip(page);
+                               return true;
+                       }
+               }
+
+               page += (1 << PAGE_ALLOC_COSTLY_ORDER);
+               pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
+       } while (page < end_page);
+
+       return false;
+}
+
 /*
  * This function is called to clear all cached information on pageblocks that
  * should be skipped for page isolation when the migrate and free page scanner
@@ -244,30 +308,54 @@ static bool pageblock_skip_persistent(struct page *page)
  */
 static void __reset_isolation_suitable(struct zone *zone)
 {
-       unsigned long start_pfn = zone->zone_start_pfn;
-       unsigned long end_pfn = zone_end_pfn(zone);
-       unsigned long pfn;
+       unsigned long migrate_pfn = zone->zone_start_pfn;
+       unsigned long free_pfn = zone_end_pfn(zone);
+       unsigned long reset_migrate = free_pfn;
+       unsigned long reset_free = migrate_pfn;
+       bool source_set = false;
+       bool free_set = false;
 
-       zone->compact_blockskip_flush = false;
+       if (!zone->compact_blockskip_flush)
+               return;
 
-       /* Walk the zone and mark every pageblock as suitable for isolation */
-       for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
-               struct page *page;
+       zone->compact_blockskip_flush = false;
 
+       /*
+        * Walk the zone and update pageblock skip information. Source looks
+        * for PageLRU while target looks for PageBuddy. When the scanner
+        * is found, both PageBuddy and PageLRU are checked as the pageblock
+        * is suitable as both source and target.
+        */
+       for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
+                                       free_pfn -= pageblock_nr_pages) {
                cond_resched();
 
-               page = pfn_to_online_page(pfn);
-               if (!page)
-                       continue;
-               if (zone != page_zone(page))
-                       continue;
-               if (pageblock_skip_persistent(page))
-                       continue;
+               /* Update the migrate PFN */
+               if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
+                   migrate_pfn < reset_migrate) {
+                       source_set = true;
+                       reset_migrate = migrate_pfn;
+                       zone->compact_init_migrate_pfn = reset_migrate;
+                       zone->compact_cached_migrate_pfn[0] = reset_migrate;
+                       zone->compact_cached_migrate_pfn[1] = reset_migrate;
+               }
 
-               clear_pageblock_skip(page);
+               /* Update the free PFN */
+               if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
+                   free_pfn > reset_free) {
+                       free_set = true;
+                       reset_free = free_pfn;
+                       zone->compact_init_free_pfn = reset_free;
+                       zone->compact_cached_free_pfn = reset_free;
+               }
        }
 
-       reset_cached_positions(zone);
+       /* Leave no distance if no suitable block was reset */
+       if (reset_migrate >= reset_free) {
+               zone->compact_cached_migrate_pfn[0] = migrate_pfn;
+               zone->compact_cached_migrate_pfn[1] = migrate_pfn;
+               zone->compact_cached_free_pfn = free_pfn;
+       }
 }
 
 void reset_isolation_suitable(pg_data_t *pgdat)
@@ -1190,7 +1278,7 @@ fast_isolate_freepages(struct compact_control *cc)
         * If starting the scan, use a deeper search and use the highest
         * PFN found if a suitable one is not found.
         */
-       if (cc->free_pfn == pageblock_start_pfn(zone_end_pfn(cc->zone) - 1)) {
+       if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
                limit = pageblock_nr_pages >> 1;
                scan_start = true;
        }
@@ -2017,7 +2105,7 @@ static enum compact_result compact_zone(struct compact_control *cc)
                        cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
                }
 
-               if (cc->migrate_pfn == start_pfn)
+               if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
                        cc->whole_zone = true;
        }