mm: vmscan: check if reclaim should really abort even if compaction_ready() is true...
[firefly-linux-kernel-4.4.55.git] / mm / compaction.c
index 021a2960ef9e18d061c972590f7651a35e7652ce..8ea7308601bc4e5807acce8be8870f4f0aa151fc 100644 (file)
@@ -35,10 +35,6 @@ struct compact_control {
        unsigned long migrate_pfn;      /* isolate_migratepages search base */
        bool sync;                      /* Synchronous migration */
 
-       /* Account for isolated anon and file pages */
-       unsigned long nr_anon;
-       unsigned long nr_file;
-
        unsigned int order;             /* order a direct compactor needs */
        int migratetype;                /* MOVABLE, RECLAIMABLE etc */
        struct zone *zone;
@@ -144,9 +140,20 @@ static void isolate_freepages(struct zone *zone,
        int nr_freepages = cc->nr_freepages;
        struct list_head *freelist = &cc->freepages;
 
+       /*
+        * Initialise the free scanner. The starting point is where we last
+        * scanned from (or the end of the zone if starting). The low point
+        * is the end of the pageblock the migration scanner is using.
+        */
        pfn = cc->free_pfn;
        low_pfn = cc->migrate_pfn + pageblock_nr_pages;
-       high_pfn = low_pfn;
+
+       /*
+        * Take care that if the migration scanner is at the end of the zone
+        * that the free scanner does not accidentally move to the next zone
+        * in the next isolation cycle.
+        */
+       high_pfn = min(low_pfn, pfn);
 
        /*
         * Isolate free pages until enough are available to migrate the
@@ -212,17 +219,13 @@ static void isolate_freepages(struct zone *zone,
 static void acct_isolated(struct zone *zone, struct compact_control *cc)
 {
        struct page *page;
-       unsigned int count[NR_LRU_LISTS] = { 0, };
+       unsigned int count[2] = { 0, };
 
-       list_for_each_entry(page, &cc->migratepages, lru) {
-               int lru = page_lru_base_type(page);
-               count[lru]++;
-       }
+       list_for_each_entry(page, &cc->migratepages, lru)
+               count[!!page_is_file_cache(page)]++;
 
-       cc->nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
-       cc->nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
-       __mod_zone_page_state(zone, NR_ISOLATED_ANON, cc->nr_anon);
-       __mod_zone_page_state(zone, NR_ISOLATED_FILE, cc->nr_file);
+       __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
+       __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
 }
 
 /* Similar to reclaim, but different enough that they don't share logic */
@@ -240,17 +243,25 @@ static bool too_many_isolated(struct zone *zone)
        return isolated > (inactive + active) / 2;
 }
 
+/* possible outcome of isolate_migratepages */
+typedef enum {
+       ISOLATE_ABORT,          /* Abort compaction now */
+       ISOLATE_NONE,           /* No pages isolated, continue scanning */
+       ISOLATE_SUCCESS,        /* Pages isolated, migrate */
+} isolate_migrate_t;
+
 /*
  * Isolate all pages that can be migrated from the block pointed to by
  * the migrate scanner within compact_control.
  */
-static unsigned long isolate_migratepages(struct zone *zone,
+static isolate_migrate_t isolate_migratepages(struct zone *zone,
                                        struct compact_control *cc)
 {
        unsigned long low_pfn, end_pfn;
        unsigned long last_pageblock_nr = 0, pageblock_nr;
        unsigned long nr_scanned = 0, nr_isolated = 0;
        struct list_head *migratelist = &cc->migratepages;
+       isolate_mode_t mode = ISOLATE_ACTIVE|ISOLATE_INACTIVE;
 
        /* Do not scan outside zone boundaries */
        low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
@@ -261,7 +272,7 @@ static unsigned long isolate_migratepages(struct zone *zone,
        /* Do not cross the free scanner or scan within a memory hole */
        if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
                cc->migrate_pfn = end_pfn;
-               return 0;
+               return ISOLATE_NONE;
        }
 
        /*
@@ -270,10 +281,14 @@ static unsigned long isolate_migratepages(struct zone *zone,
         * delay for some time until fewer pages are isolated
         */
        while (unlikely(too_many_isolated(zone))) {
+               /* async migration should just abort */
+               if (!cc->sync)
+                       return ISOLATE_ABORT;
+
                congestion_wait(BLK_RW_ASYNC, HZ/10);
 
                if (fatal_signal_pending(current))
-                       return 0;
+                       return ISOLATE_ABORT;
        }
 
        /* Time to isolate some pages for migration */
@@ -298,12 +313,34 @@ static unsigned long isolate_migratepages(struct zone *zone,
                } else if (!locked)
                        spin_lock_irq(&zone->lru_lock);
 
+               /*
+                * migrate_pfn does not necessarily start aligned to a
+                * pageblock. Ensure that pfn_valid is called when moving
+                * into a new MAX_ORDER_NR_PAGES range in case of large
+                * memory holes within the zone
+                */
+               if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
+                       if (!pfn_valid(low_pfn)) {
+                               low_pfn += MAX_ORDER_NR_PAGES - 1;
+                               continue;
+                       }
+               }
+
                if (!pfn_valid_within(low_pfn))
                        continue;
                nr_scanned++;
 
-               /* Get the page and skip if free */
+               /*
+                * Get the page and ensure the page is within the same zone.
+                * See the comment in isolate_freepages about overlapping
+                * nodes. It is deliberate that the new zone lock is not taken
+                * as memory compaction should not move pages between nodes.
+                */
                page = pfn_to_page(low_pfn);
+               if (page_zone(page) != zone)
+                       continue;
+
+               /* Skip if free */
                if (PageBuddy(page))
                        continue;
 
@@ -334,8 +371,11 @@ static unsigned long isolate_migratepages(struct zone *zone,
                        continue;
                }
 
+               if (!cc->sync)
+                       mode |= ISOLATE_ASYNC_MIGRATE;
+
                /* Try isolate the page */
-               if (__isolate_lru_page(page, ISOLATE_BOTH, 0) != 0)
+               if (__isolate_lru_page(page, mode, 0) != 0)
                        continue;
 
                VM_BUG_ON(PageTransCompound(page));
@@ -358,7 +398,7 @@ static unsigned long isolate_migratepages(struct zone *zone,
 
        trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
 
-       return cc->nr_migratepages;
+       return ISOLATE_SUCCESS;
 }
 
 /*
@@ -420,13 +460,6 @@ static int compact_finished(struct zone *zone,
        if (cc->free_pfn <= cc->migrate_pfn)
                return COMPACT_COMPLETE;
 
-       /* Compaction run is not finished if the watermark is not met */
-       watermark = low_wmark_pages(zone);
-       watermark += (1 << cc->order);
-
-       if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
-               return COMPACT_CONTINUE;
-
        /*
         * order == -1 is expected when compacting via
         * /proc/sys/vm/compact_memory
@@ -434,6 +467,13 @@ static int compact_finished(struct zone *zone,
        if (cc->order == -1)
                return COMPACT_CONTINUE;
 
+       /* Compaction run is not finished if the watermark is not met */
+       watermark = low_wmark_pages(zone);
+       watermark += (1 << cc->order);
+
+       if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
+               return COMPACT_CONTINUE;
+
        /* Direct compactor: Is a suitable page free? */
        for (order = cc->order; order < MAX_ORDER; order++) {
                /* Job done if page is free of the right migratetype */
@@ -460,6 +500,13 @@ unsigned long compaction_suitable(struct zone *zone, int order)
        int fragindex;
        unsigned long watermark;
 
+       /*
+        * order == -1 is expected when compacting via
+        * /proc/sys/vm/compact_memory
+        */
+       if (order == -1)
+               return COMPACT_CONTINUE;
+
        /*
         * Watermarks for order-0 must be met for compaction. Note the 2UL.
         * This is because during migration, copies of pages need to be
@@ -469,18 +516,12 @@ unsigned long compaction_suitable(struct zone *zone, int order)
        if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
                return COMPACT_SKIPPED;
 
-       /*
-        * order == -1 is expected when compacting via
-        * /proc/sys/vm/compact_memory
-        */
-       if (order == -1)
-               return COMPACT_CONTINUE;
-
        /*
         * fragmentation index determines if allocation failures are due to
         * low memory or external fragmentation
         *
-        * index of -1 implies allocations might succeed dependingon watermarks
+        * index of -1000 implies allocations might succeed depending on
+        * watermarks
         * index towards 0 implies failure is due to lack of memory
         * index towards 1000 implies failure is due to fragmentation
         *
@@ -490,7 +531,8 @@ unsigned long compaction_suitable(struct zone *zone, int order)
        if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
                return COMPACT_SKIPPED;
 
-       if (fragindex == -1 && zone_watermark_ok(zone, order, watermark, 0, 0))
+       if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
+           0, 0))
                return COMPACT_PARTIAL;
 
        return COMPACT_CONTINUE;
@@ -522,13 +564,20 @@ static int compact_zone(struct zone *zone, struct compact_control *cc)
                unsigned long nr_migrate, nr_remaining;
                int err;
 
-               if (!isolate_migratepages(zone, cc))
+               switch (isolate_migratepages(zone, cc)) {
+               case ISOLATE_ABORT:
+                       ret = COMPACT_PARTIAL;
+                       goto out;
+               case ISOLATE_NONE:
                        continue;
+               case ISOLATE_SUCCESS:
+                       ;
+               }
 
                nr_migrate = cc->nr_migratepages;
                err = migrate_pages(&cc->migratepages, compaction_alloc,
                                (unsigned long)cc, false,
-                               cc->sync);
+                               cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
                update_nr_listpages(cc);
                nr_remaining = cc->nr_migratepages;
 
@@ -543,10 +592,14 @@ static int compact_zone(struct zone *zone, struct compact_control *cc)
                if (err) {
                        putback_lru_pages(&cc->migratepages);
                        cc->nr_migratepages = 0;
+                       if (err == -ENOMEM) {
+                               ret = COMPACT_PARTIAL;
+                               goto out;
+                       }
                }
-
        }
 
+out:
        /* Release free pages and check accounting */
        cc->nr_freepages -= release_freepages(&cc->freepages);
        VM_BUG_ON(cc->nr_freepages != 0);