mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages...

author Mel Gorman <mgorman@suse.de>

Wed, 3 Jul 2013 22:01:57 +0000 (15:01 -0700)

committer Linus Torvalds <torvalds@linux-foundation.org>

Wed, 3 Jul 2013 23:07:28 +0000 (16:07 -0700)
author Mel Gorman <mgorman@suse.de>
Wed, 3 Jul 2013 22:01:57 +0000 (15:01 -0700)
committer Linus Torvalds <torvalds@linux-foundation.org>
Wed, 3 Jul 2013 23:07:28 +0000 (16:07 -0700)
diff --git a/mm/vmscan.c b/mm/vmscan.c

index 4a43c289b23a21e8bd27c88eae345f8f138465e6..999ef0b9399ae0e31ec579da43e6c9a76c380cc5 100644 (file)
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -669,6 +669,25 @@ static enum page_references page_check_references(struct page *page,
         return PAGEREF_RECLAIM;
  }
  
+/* Check if a page is dirty or under writeback */
+static void page_check_dirty_writeback(struct page *page,
+                                      bool *dirty, bool *writeback)
+{
+       /*
+        * Anonymous pages are not handled by flushers and must be written
+        * from reclaim context. Do not stall reclaim based on them
+        */
+       if (!page_is_file_cache(page)) {
+               *dirty = false;
+               *writeback = false;
+               return;
+       }
+
+       /* By default assume that the page flags are accurate */
+       *dirty = PageDirty(page);
+       *writeback = PageWriteback(page);
+}
+
  /*
   * shrink_page_list() returns the number of reclaimed pages
   */
@@ -697,6 +716,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
                 struct page *page;
                 int may_enter_fs;
                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
+               bool dirty, writeback;
  
                 cond_resched();
  
@@ -724,6 +744,24 @@ static unsigned long shrink_page_list(struct list_head *page_list,
                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
  
+               /*
+                * The number of dirty pages determines if a zone is marked
+                * reclaim_congested which affects wait_iff_congested. kswapd
+                * will stall and start writing pages if the tail of the LRU
+                * is all dirty unqueued pages.
+                */
+               page_check_dirty_writeback(page, &dirty, &writeback);
+               if (dirty || writeback)
+                       nr_dirty++;
+
+               if (dirty && !writeback)
+                       nr_unqueued_dirty++;
+
+               /* Treat this page as congested if underlying BDI is */
+               mapping = page_mapping(page);
+               if (mapping && bdi_write_congested(mapping->backing_dev_info))
+                       nr_congested++;
+
                 /*
                  * If a page at the tail of the LRU is under writeback, there
                  * are three cases to consider.
@@ -819,9 +857,10 @@ static unsigned long shrink_page_list(struct list_head *page_list,
                         if (!add_to_swap(page, page_list))
                                 goto activate_locked;
                         may_enter_fs = 1;
-               }
  
-               mapping = page_mapping(page);
+                       /* Adding to swap updated mapping */
+                       mapping = page_mapping(page);
+               }
  
                 /*
                  * The page is mapped into the page tables of one or more
@@ -841,11 +880,6 @@ static unsigned long shrink_page_list(struct list_head *page_list,
                 }
  
                 if (PageDirty(page)) {
-                       nr_dirty++;
-
-                       if (!PageWriteback(page))
-                               nr_unqueued_dirty++;
-
                         /*
                          * Only kswapd can writeback filesystem pages to
                          * avoid risk of stack overflow but only writeback
@@ -876,7 +910,6 @@ static unsigned long shrink_page_list(struct list_head *page_list,
                         /* Page is dirty, try to write it out here */
                         switch (pageout(page, mapping, sc)) {
                         case PAGE_KEEP:
-                               nr_congested++;
                                 goto keep_locked;
                         case PAGE_ACTIVATE:
                                 goto activate_locked;
@@ -1318,7 +1351,7 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
         unsigned long nr_scanned;
         unsigned long nr_reclaimed = 0;
         unsigned long nr_taken;
-       unsigned long nr_dirty = 0;
+       unsigned long nr_unqueued_dirty = 0;
         unsigned long nr_writeback = 0;
         isolate_mode_t isolate_mode = 0;
         int file = is_file_lru(lru);
@@ -1361,7 +1394,7 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
                 return 0;
  
         nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
-                                       &nr_dirty, &nr_writeback, false);
+                               &nr_unqueued_dirty, &nr_writeback, false);
  
         spin_lock_irq(&zone->lru_lock);
  
@@ -1416,11 +1449,13 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
         /*
          * Similarly, if many dirty pages are encountered that are not
          * currently being written then flag that kswapd should start
-        * writing back pages.
+        * writing back pages and stall to give a chance for flushers
+        * to catch up.
          */
-       if (global_reclaim(sc) && nr_dirty &&
-                       nr_dirty >= (nr_taken >> (DEF_PRIORITY - sc->priority)))
+       if (global_reclaim(sc) && nr_unqueued_dirty == nr_taken) {
+               congestion_wait(BLK_RW_ASYNC, HZ/10);
                 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
+       }
  
         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
                 zone_idx(zone),
author	Mel Gorman <mgorman@suse.de>
	Wed, 3 Jul 2013 22:01:57 +0000 (15:01 -0700)
committer	Linus Torvalds <torvalds@linux-foundation.org>
	Wed, 3 Jul 2013 23:07:28 +0000 (16:07 -0700)