mm, vmscan: begin reclaiming pages on a per-node basis
authorMel Gorman <mgorman@techsingularity.net>
Thu, 28 Jul 2016 22:45:37 +0000 (15:45 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 28 Jul 2016 23:07:41 +0000 (16:07 -0700)
This patch makes reclaim decisions on a per-node basis.  A reclaimer
knows what zone is required by the allocation request and skips pages
from higher zones.  In many cases this will be ok because it's a
GFP_HIGHMEM request of some description.  On 64-bit, ZONE_DMA32 requests
will cause some problems but 32-bit devices on 64-bit platforms are
increasingly rare.  Historically it would have been a major problem on
32-bit with big Highmem:Lowmem ratios but such configurations are also
now rare and even where they exist, they are not encouraged.  If it
really becomes a problem, it'll manifest as very low reclaim
efficiencies.

Link: http://lkml.kernel.org/r/1467970510-21195-6-git-send-email-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@surriel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/vmscan.c

index 86a523a761c9128559e81cc156dad31facb291b1..766b36bec829b85f105df9c54911c4864edee6be 100644 (file)
@@ -84,6 +84,9 @@ struct scan_control {
        /* Scan (total_size >> priority) pages at once */
        int priority;
 
+       /* The highest zone to isolate pages for reclaim from */
+       enum zone_type reclaim_idx;
+
        unsigned int may_writepage:1;
 
        /* Can mapped pages be reclaimed? */
@@ -1392,6 +1395,7 @@ static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
        unsigned long nr_taken = 0;
        unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
        unsigned long scan, nr_pages;
+       LIST_HEAD(pages_skipped);
 
        for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
                                        !list_empty(src); scan++) {
@@ -1402,6 +1406,11 @@ static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
 
                VM_BUG_ON_PAGE(!PageLRU(page), page);
 
+               if (page_zonenum(page) > sc->reclaim_idx) {
+                       list_move(&page->lru, &pages_skipped);
+                       continue;
+               }
+
                switch (__isolate_lru_page(page, mode)) {
                case 0:
                        nr_pages = hpage_nr_pages(page);
@@ -1420,6 +1429,15 @@ static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
                }
        }
 
+       /*
+        * Splice any skipped pages to the start of the LRU list. Note that
+        * this disrupts the LRU order when reclaiming for lower zones but
+        * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
+        * scanning would soon rescan the same pages to skip and put the
+        * system at risk of premature OOM.
+        */
+       if (!list_empty(&pages_skipped))
+               list_splice(&pages_skipped, src);
        *nr_scanned = scan;
        trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
                                    nr_taken, mode, is_file_lru(lru));
@@ -1589,7 +1607,7 @@ static int current_may_throttle(void)
 }
 
 /*
- * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
+ * shrink_inactive_list() is a helper for shrink_node().  It returns the number
  * of reclaimed pages
  */
 static noinline_for_stack unsigned long
@@ -2401,12 +2419,13 @@ static inline bool should_continue_reclaim(struct zone *zone,
        }
 }
 
-static bool shrink_zone(struct zone *zone, struct scan_control *sc,
-                       bool is_classzone)
+static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc,
+                       enum zone_type classzone_idx)
 {
        struct reclaim_state *reclaim_state = current->reclaim_state;
        unsigned long nr_reclaimed, nr_scanned;
        bool reclaimable = false;
+       struct zone *zone = &pgdat->node_zones[classzone_idx];
 
        do {
                struct mem_cgroup *root = sc->target_mem_cgroup;
@@ -2438,7 +2457,7 @@ static bool shrink_zone(struct zone *zone, struct scan_control *sc,
                        shrink_zone_memcg(zone, memcg, sc, &lru_pages);
                        zone_lru_pages += lru_pages;
 
-                       if (memcg && is_classzone)
+                       if (!global_reclaim(sc))
                                shrink_slab(sc->gfp_mask, zone_to_nid(zone),
                                            memcg, sc->nr_scanned - scanned,
                                            lru_pages);
@@ -2469,7 +2488,7 @@ static bool shrink_zone(struct zone *zone, struct scan_control *sc,
                 * Shrink the slab caches in the same proportion that
                 * the eligible LRU pages were scanned.
                 */
-               if (global_reclaim(sc) && is_classzone)
+               if (global_reclaim(sc))
                        shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
                                    sc->nr_scanned - nr_scanned,
                                    zone_lru_pages);
@@ -2553,7 +2572,7 @@ static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
        unsigned long nr_soft_reclaimed;
        unsigned long nr_soft_scanned;
        gfp_t orig_mask;
-       enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
+       enum zone_type classzone_idx;
 
        /*
         * If the number of buffer_heads in the machine exceeds the maximum
@@ -2561,17 +2580,23 @@ static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
         * highmem pages could be pinning lowmem pages storing buffer_heads
         */
        orig_mask = sc->gfp_mask;
-       if (buffer_heads_over_limit)
+       if (buffer_heads_over_limit) {
                sc->gfp_mask |= __GFP_HIGHMEM;
+               sc->reclaim_idx = classzone_idx = gfp_zone(sc->gfp_mask);
+       }
 
        for_each_zone_zonelist_nodemask(zone, z, zonelist,
-                                       gfp_zone(sc->gfp_mask), sc->nodemask) {
-               enum zone_type classzone_idx;
-
+                                       sc->reclaim_idx, sc->nodemask) {
                if (!populated_zone(zone))
                        continue;
 
-               classzone_idx = requested_highidx;
+               /*
+                * Note that reclaim_idx does not change as it is the highest
+                * zone reclaimed from which for empty zones is a no-op but
+                * classzone_idx is used by shrink_node to test if the slabs
+                * should be shrunk on a given node.
+                */
+               classzone_idx = sc->reclaim_idx;
                while (!populated_zone(zone->zone_pgdat->node_zones +
                                                        classzone_idx))
                        classzone_idx--;
@@ -2600,8 +2625,8 @@ static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
                         */
                        if (IS_ENABLED(CONFIG_COMPACTION) &&
                            sc->order > PAGE_ALLOC_COSTLY_ORDER &&
-                           zonelist_zone_idx(z) <= requested_highidx &&
-                           compaction_ready(zone, sc->order, requested_highidx)) {
+                           zonelist_zone_idx(z) <= classzone_idx &&
+                           compaction_ready(zone, sc->order, classzone_idx)) {
                                sc->compaction_ready = true;
                                continue;
                        }
@@ -2621,7 +2646,7 @@ static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
                        /* need some check for avoid more shrink_zone() */
                }
 
-               shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
+               shrink_node(zone->zone_pgdat, sc, classzone_idx);
        }
 
        /*
@@ -2847,6 +2872,7 @@ unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
        struct scan_control sc = {
                .nr_to_reclaim = SWAP_CLUSTER_MAX,
                .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
+               .reclaim_idx = gfp_zone(gfp_mask),
                .order = order,
                .nodemask = nodemask,
                .priority = DEF_PRIORITY,
@@ -2886,6 +2912,7 @@ unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
                .target_mem_cgroup = memcg,
                .may_writepage = !laptop_mode,
                .may_unmap = 1,
+               .reclaim_idx = MAX_NR_ZONES - 1,
                .may_swap = !noswap,
        };
        unsigned long lru_pages;
@@ -2924,6 +2951,7 @@ unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
                .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
                .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
                                (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
+               .reclaim_idx = MAX_NR_ZONES - 1,
                .target_mem_cgroup = memcg,
                .priority = DEF_PRIORITY,
                .may_writepage = !laptop_mode,
@@ -3118,7 +3146,7 @@ static bool kswapd_shrink_zone(struct zone *zone,
                                                balance_gap, classzone_idx))
                return true;
 
-       shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
+       shrink_node(zone->zone_pgdat, sc, classzone_idx);
 
        /* TODO: ANOMALY */
        clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
@@ -3167,6 +3195,7 @@ static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
        unsigned long nr_soft_scanned;
        struct scan_control sc = {
                .gfp_mask = GFP_KERNEL,
+               .reclaim_idx = MAX_NR_ZONES - 1,
                .order = order,
                .priority = DEF_PRIORITY,
                .may_writepage = !laptop_mode,
@@ -3237,15 +3266,14 @@ static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
                        sc.may_writepage = 1;
 
                /*
-                * Now scan the zone in the dma->highmem direction, stopping
-                * at the last zone which needs scanning.
-                *
-                * We do this because the page allocator works in the opposite
-                * direction.  This prevents the page allocator from allocating
-                * pages behind kswapd's direction of progress, which would
-                * cause too much scanning of the lower zones.
+                * Continue scanning in the highmem->dma direction stopping at
+                * the last zone which needs scanning. This may reclaim lowmem
+                * pages that are not necessary for zone balancing but it
+                * preserves LRU ordering. It is assumed that the bulk of
+                * allocation requests can use arbitrary zones with the
+                * possible exception of big highmem:lowmem configurations.
                 */
-               for (i = 0; i <= end_zone; i++) {
+               for (i = end_zone; i >= 0; i--) {
                        struct zone *zone = pgdat->node_zones + i;
 
                        if (!populated_zone(zone))
@@ -3256,6 +3284,7 @@ static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
                                continue;
 
                        sc.nr_scanned = 0;
+                       sc.reclaim_idx = i;
 
                        nr_soft_scanned = 0;
                        /*
@@ -3513,6 +3542,7 @@ unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
        struct scan_control sc = {
                .nr_to_reclaim = nr_to_reclaim,
                .gfp_mask = GFP_HIGHUSER_MOVABLE,
+               .reclaim_idx = MAX_NR_ZONES - 1,
                .priority = DEF_PRIORITY,
                .may_writepage = 1,
                .may_unmap = 1,
@@ -3704,6 +3734,7 @@ static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
                .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
                .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
                .may_swap = 1,
+               .reclaim_idx = zone_idx(zone),
        };
 
        cond_resched();
@@ -3723,7 +3754,7 @@ static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
                 * priorities until we have enough memory freed.
                 */
                do {
-                       shrink_zone(zone, &sc, true);
+                       shrink_node(zone->zone_pgdat, &sc, zone_idx(zone));
                } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
        }