tmpfs: copy truncate_inode_pages_range
[linux-2.6.git] / mm / compaction.c
index 47fca10..6cc604b 100644 (file)
@@ -144,16 +144,26 @@ 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
         * pages on cc->migratepages. We stop searching if the migrate
         * and free page scanners meet or enough free pages are isolated.
         */
-       spin_lock_irqsave(&zone->lock, flags);
        for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
                                        pfn -= pageblock_nr_pages) {
                unsigned long isolated;
@@ -176,9 +186,19 @@ static void isolate_freepages(struct zone *zone,
                if (!suitable_migration_target(page))
                        continue;
 
-               /* Found a block suitable for isolating free pages from */
-               isolated = isolate_freepages_block(zone, pfn, freelist);
-               nr_freepages += isolated;
+               /*
+                * Found a block suitable for isolating free pages from. Now
+                * we disabled interrupts, double check things are ok and
+                * isolate the pages. This is to minimise the time IRQs
+                * are disabled
+                */
+               isolated = 0;
+               spin_lock_irqsave(&zone->lock, flags);
+               if (suitable_migration_target(page)) {
+                       isolated = isolate_freepages_block(zone, pfn, freelist);
+                       nr_freepages += isolated;
+               }
+               spin_unlock_irqrestore(&zone->lock, flags);
 
                /*
                 * Record the highest PFN we isolated pages from. When next
@@ -188,7 +208,6 @@ static void isolate_freepages(struct zone *zone,
                if (isolated)
                        high_pfn = max(high_pfn, pfn);
        }
-       spin_unlock_irqrestore(&zone->lock, flags);
 
        /* split_free_page does not map the pages */
        list_for_each_entry(page, freelist, lru) {
@@ -232,14 +251,22 @@ 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;
 
@@ -252,7 +279,7 @@ 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;
        }
 
        /*
@@ -261,16 +288,38 @@ 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 */
+       cond_resched();
        spin_lock_irq(&zone->lru_lock);
        for (; low_pfn < end_pfn; low_pfn++) {
                struct page *page;
+               bool locked = true;
+
+               /* give a chance to irqs before checking need_resched() */
+               if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) {
+                       spin_unlock_irq(&zone->lru_lock);
+                       locked = false;
+               }
+               if (need_resched() || spin_is_contended(&zone->lru_lock)) {
+                       if (locked)
+                               spin_unlock_irq(&zone->lru_lock);
+                       cond_resched();
+                       spin_lock_irq(&zone->lru_lock);
+                       if (fatal_signal_pending(current))
+                               break;
+               } else if (!locked)
+                       spin_lock_irq(&zone->lru_lock);
+
                if (!pfn_valid_within(low_pfn))
                        continue;
                nr_scanned++;
@@ -280,10 +329,39 @@ static unsigned long isolate_migratepages(struct zone *zone,
                if (PageBuddy(page))
                        continue;
 
+               /*
+                * For async migration, also only scan in MOVABLE blocks. Async
+                * migration is optimistic to see if the minimum amount of work
+                * satisfies the allocation
+                */
+               pageblock_nr = low_pfn >> pageblock_order;
+               if (!cc->sync && last_pageblock_nr != pageblock_nr &&
+                               get_pageblock_migratetype(page) != MIGRATE_MOVABLE) {
+                       low_pfn += pageblock_nr_pages;
+                       low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
+                       last_pageblock_nr = pageblock_nr;
+                       continue;
+               }
+
+               if (!PageLRU(page))
+                       continue;
+
+               /*
+                * PageLRU is set, and lru_lock excludes isolation,
+                * splitting and collapsing (collapsing has already
+                * happened if PageLRU is set).
+                */
+               if (PageTransHuge(page)) {
+                       low_pfn += (1 << compound_order(page)) - 1;
+                       continue;
+               }
+
                /* Try isolate the page */
                if (__isolate_lru_page(page, ISOLATE_BOTH, 0) != 0)
                        continue;
 
+               VM_BUG_ON(PageTransCompound(page));
+
                /* Successfully isolated */
                del_page_from_lru_list(zone, page, page_lru(page));
                list_add(&page->lru, migratelist);
@@ -302,7 +380,7 @@ static unsigned long isolate_migratepages(struct zone *zone,
 
        trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
 
-       return cc->nr_migratepages;
+       return ISOLATE_SUCCESS;
 }
 
 /*
@@ -352,10 +430,10 @@ static void update_nr_listpages(struct compact_control *cc)
 }
 
 static int compact_finished(struct zone *zone,
-                                               struct compact_control *cc)
+                           struct compact_control *cc)
 {
        unsigned int order;
-       unsigned long watermark = low_wmark_pages(zone) + (1 << cc->order);
+       unsigned long watermark;
 
        if (fatal_signal_pending(current))
                return COMPACT_PARTIAL;
@@ -364,11 +442,18 @@ 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 */
-       if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
+       /*
+        * order == -1 is expected when compacting via
+        * /proc/sys/vm/compact_memory
+        */
+       if (cc->order == -1)
                return COMPACT_CONTINUE;
 
-       if (cc->order == -1)
+       /* 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? */
@@ -398,6 +483,13 @@ unsigned long compaction_suitable(struct zone *zone, int order)
        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
         * allocated and for a short time, the footprint is higher
@@ -410,7 +502,8 @@ unsigned long compaction_suitable(struct zone *zone, int order)
         * 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
         *
@@ -420,7 +513,8 @@ 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;
@@ -450,13 +544,21 @@ static int compact_zone(struct zone *zone, struct compact_control *cc)
 
        while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
                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;
-               migrate_pages(&cc->migratepages, compaction_alloc,
-                               (unsigned long)cc, 0,
+               err = migrate_pages(&cc->migratepages, compaction_alloc,
+                               (unsigned long)cc, false,
                                cc->sync);
                update_nr_listpages(cc);
                nr_remaining = cc->nr_migratepages;
@@ -469,13 +571,14 @@ static int compact_zone(struct zone *zone, struct compact_control *cc)
                                                nr_remaining);
 
                /* Release LRU pages not migrated */
-               if (!list_empty(&cc->migratepages)) {
+               if (err) {
                        putback_lru_pages(&cc->migratepages);
                        cc->nr_migratepages = 0;
                }
 
        }
 
+out:
        /* Release free pages and check accounting */
        cc->nr_freepages -= release_freepages(&cc->freepages);
        VM_BUG_ON(cc->nr_freepages != 0);
@@ -484,8 +587,8 @@ static int compact_zone(struct zone *zone, struct compact_control *cc)
 }
 
 unsigned long compact_zone_order(struct zone *zone,
-                                               int order, gfp_t gfp_mask,
-                                               bool sync)
+                                int order, gfp_t gfp_mask,
+                                bool sync)
 {
        struct compact_control cc = {
                .nr_freepages = 0,
@@ -529,7 +632,7 @@ unsigned long try_to_compact_pages(struct zonelist *zonelist,
         * made because an assumption is made that the page allocator can satisfy
         * the "cheaper" orders without taking special steps
         */
-       if (order <= PAGE_ALLOC_COSTLY_ORDER || !may_enter_fs || !may_perform_io)
+       if (!order || !may_enter_fs || !may_perform_io)
                return rc;
 
        count_vm_event(COMPACTSTALL);