patch-2_6_7-vs1_9_1_12
[linux-2.6.git] / mm / page-writeback.c
index 042ccad..a2aa511 100644 (file)
@@ -92,11 +92,29 @@ int block_dump;
  */
 int laptop_mode;
 
+EXPORT_SYMBOL(laptop_mode);
+
 /* End of sysctl-exported parameters */
 
 
 static void background_writeout(unsigned long _min_pages);
 
+struct writeback_state
+{
+       unsigned long nr_dirty;
+       unsigned long nr_unstable;
+       unsigned long nr_mapped;
+       unsigned long nr_writeback;
+};
+
+static void get_writeback_state(struct writeback_state *wbs)
+{
+       wbs->nr_dirty = read_page_state(nr_dirty);
+       wbs->nr_unstable = read_page_state(nr_unstable);
+       wbs->nr_mapped = read_page_state(nr_mapped);
+       wbs->nr_writeback = read_page_state(nr_writeback);
+}
+
 /*
  * Work out the current dirty-memory clamping and background writeout
  * thresholds.
@@ -115,7 +133,7 @@ static void background_writeout(unsigned long _min_pages);
  * clamping level.
  */
 static void
-get_dirty_limits(struct page_state *ps, long *pbackground, long *pdirty)
+get_dirty_limits(struct writeback_state *wbs, long *pbackground, long *pdirty)
 {
        int background_ratio;           /* Percentages */
        int dirty_ratio;
@@ -124,9 +142,9 @@ get_dirty_limits(struct page_state *ps, long *pbackground, long *pdirty)
        long dirty;
        struct task_struct *tsk;
 
-       get_page_state(ps);
+       get_writeback_state(wbs);
 
-       unmapped_ratio = 100 - (ps->nr_mapped * 100) / total_pages;
+       unmapped_ratio = 100 - (wbs->nr_mapped * 100) / total_pages;
 
        dirty_ratio = vm_dirty_ratio;
        if (dirty_ratio > unmapped_ratio / 2)
@@ -159,7 +177,7 @@ get_dirty_limits(struct page_state *ps, long *pbackground, long *pdirty)
  */
 static void balance_dirty_pages(struct address_space *mapping)
 {
-       struct page_state ps;
+       struct writeback_state wbs;
        long nr_reclaimable;
        long background_thresh;
        long dirty_thresh;
@@ -176,9 +194,9 @@ static void balance_dirty_pages(struct address_space *mapping)
                        .nr_to_write    = write_chunk,
                };
 
-               get_dirty_limits(&ps, &background_thresh, &dirty_thresh);
-               nr_reclaimable = ps.nr_dirty + ps.nr_unstable;
-               if (nr_reclaimable + ps.nr_writeback <= dirty_thresh)
+               get_dirty_limits(&wbs, &background_thresh, &dirty_thresh);
+               nr_reclaimable = wbs.nr_dirty + wbs.nr_unstable;
+               if (nr_reclaimable + wbs.nr_writeback <= dirty_thresh)
                        break;
 
                dirty_exceeded = 1;
@@ -191,10 +209,10 @@ static void balance_dirty_pages(struct address_space *mapping)
                 */
                if (nr_reclaimable) {
                        writeback_inodes(&wbc);
-                       get_dirty_limits(&ps, &background_thresh,
+                       get_dirty_limits(&wbs, &background_thresh,
                                        &dirty_thresh);
-                       nr_reclaimable = ps.nr_dirty + ps.nr_unstable;
-                       if (nr_reclaimable + ps.nr_writeback <= dirty_thresh)
+                       nr_reclaimable = wbs.nr_dirty + wbs.nr_unstable;
+                       if (nr_reclaimable + wbs.nr_writeback <= dirty_thresh)
                                break;
                        pages_written += write_chunk - wbc.nr_to_write;
                        if (pages_written >= write_chunk)
@@ -203,7 +221,7 @@ static void balance_dirty_pages(struct address_space *mapping)
                blk_congestion_wait(WRITE, HZ/10);
        }
 
-       if (nr_reclaimable + ps.nr_writeback <= dirty_thresh)
+       if (nr_reclaimable + wbs.nr_writeback <= dirty_thresh)
                dirty_exceeded = 0;
 
        if (writeback_in_progress(bdi))
@@ -230,10 +248,10 @@ static void balance_dirty_pages(struct address_space *mapping)
  * which was newly dirtied.  The function will periodically check the system's
  * dirty state and will initiate writeback if needed.
  *
- * On really big machines, get_page_state is expensive, so try to avoid calling
- * it too often (ratelimiting).  But once we're over the dirty memory limit we
- * decrease the ratelimiting by a lot, to prevent individual processes from
- * overshooting the limit by (ratelimit_pages) each.
+ * On really big machines, get_writeback_state is expensive, so try to avoid
+ * calling it too often (ratelimiting).  But once we're over the dirty memory
+ * limit we decrease the ratelimiting by a lot, to prevent individual processes
+ * from overshooting the limit by (ratelimit_pages) each.
  */
 void balance_dirty_pages_ratelimited(struct address_space *mapping)
 {
@@ -274,12 +292,12 @@ static void background_writeout(unsigned long _min_pages)
        };
 
        for ( ; ; ) {
-               struct page_state ps;
+               struct writeback_state wbs;
                long background_thresh;
                long dirty_thresh;
 
-               get_dirty_limits(&ps, &background_thresh, &dirty_thresh);
-               if (ps.nr_dirty + ps.nr_unstable < background_thresh
+               get_dirty_limits(&wbs, &background_thresh, &dirty_thresh);
+               if (wbs.nr_dirty + wbs.nr_unstable < background_thresh
                                && min_pages <= 0)
                        break;
                wbc.encountered_congestion = 0;
@@ -304,10 +322,10 @@ static void background_writeout(unsigned long _min_pages)
 int wakeup_bdflush(long nr_pages)
 {
        if (nr_pages == 0) {
-               struct page_state ps;
+               struct writeback_state wbs;
 
-               get_page_state(&ps);
-               nr_pages = ps.nr_dirty + ps.nr_unstable;
+               get_writeback_state(&wbs);
+               nr_pages = wbs.nr_dirty + wbs.nr_unstable;
        }
        return pdflush_operation(background_writeout, nr_pages);
 }
@@ -341,7 +359,7 @@ static void wb_kupdate(unsigned long arg)
        unsigned long start_jif;
        unsigned long next_jif;
        long nr_to_write;
-       struct page_state ps;
+       struct writeback_state wbs;
        struct writeback_control wbc = {
                .bdi            = NULL,
                .sync_mode      = WB_SYNC_NONE,
@@ -353,11 +371,11 @@ static void wb_kupdate(unsigned long arg)
 
        sync_supers();
 
-       get_page_state(&ps);
+       get_writeback_state(&wbs);
        oldest_jif = jiffies - (dirty_expire_centisecs * HZ) / 100;
        start_jif = jiffies;
        next_jif = start_jif + (dirty_writeback_centisecs * HZ) / 100;
-       nr_to_write = ps.nr_dirty + ps.nr_unstable +
+       nr_to_write = wbs.nr_dirty + wbs.nr_unstable +
                        (inodes_stat.nr_inodes - inodes_stat.nr_unused);
        while (nr_to_write > 0) {
                wbc.encountered_congestion = 0;
@@ -432,8 +450,8 @@ void laptop_sync_completion(void)
 /*
  * If ratelimit_pages is too high then we can get into dirty-data overload
  * if a large number of processes all perform writes at the same time.
- * If it is too low then SMP machines will call the (expensive) get_page_state
- * too often.
+ * If it is too low then SMP machines will call the (expensive)
+ * get_writeback_state too often.
  *
  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
@@ -545,34 +563,40 @@ EXPORT_SYMBOL(write_one_page);
  * For address_spaces which do not use buffers.  Just tag the page as dirty in
  * its radix tree.
  *
- * __set_page_dirty_nobuffers() may return -ENOSPC.  But if it does, the page
- * is still safe, as long as it actually manages to find some blocks at
- * writeback time.
- *
  * This is also used when a single buffer is being dirtied: we want to set the
  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
+ *
+ * Most callers have locked the page, which pins the address_space in memory.
+ * But zap_pte_range() does not lock the page, however in that case the
+ * mapping is pinned by the vma's ->vm_file reference.
+ *
+ * We take care to handle the case where the page was truncated from the
+ * mapping by re-checking page_mapping() insode tree_lock.
  */
 int __set_page_dirty_nobuffers(struct page *page)
 {
        int ret = 0;
 
        if (!TestSetPageDirty(page)) {
-               struct address_space *mapping = page->mapping;
+               struct address_space *mapping = page_mapping(page);
 
                if (mapping) {
                        spin_lock_irq(&mapping->tree_lock);
-                       if (page->mapping) {    /* Race with truncate? */
-                               BUG_ON(page->mapping != mapping);
+                       mapping = page_mapping(page);
+                       if (page_mapping(page)) { /* Race with truncate? */
+                               BUG_ON(page_mapping(page) != mapping);
                                if (!mapping->backing_dev_info->memory_backed)
                                        inc_page_state(nr_dirty);
                                radix_tree_tag_set(&mapping->page_tree,
-                                       page->index, PAGECACHE_TAG_DIRTY);
+                                       page_index(page), PAGECACHE_TAG_DIRTY);
                        }
                        spin_unlock_irq(&mapping->tree_lock);
-                       if (!PageSwapCache(page))
+                       if (mapping->host) {
+                               /* !PageAnon && !swapper_space */
                                __mark_inode_dirty(mapping->host,
                                                        I_DIRTY_PAGES);
+                       }
                }
        }
        return ret;
@@ -598,14 +622,16 @@ EXPORT_SYMBOL(redirty_page_for_writepage);
 int fastcall set_page_dirty(struct page *page)
 {
        struct address_space *mapping = page_mapping(page);
-       int (*spd)(struct page *);
 
-       if (!mapping) {
-               SetPageDirty(page);
-               return 0;
+       if (likely(mapping)) {
+               int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
+               if (spd)
+                       return (*spd)(page);
+               return __set_page_dirty_buffers(page);
        }
-       spd = mapping->a_ops->set_page_dirty;
-       return spd? (*spd)(page): __set_page_dirty_buffers(page);
+       if (!PageDirty(page))
+               SetPageDirty(page);
+       return 0;
 }
 EXPORT_SYMBOL(set_page_dirty);
 
@@ -642,7 +668,8 @@ int test_clear_page_dirty(struct page *page)
        if (mapping) {
                spin_lock_irqsave(&mapping->tree_lock, flags);
                if (TestClearPageDirty(page)) {
-                       radix_tree_tag_clear(&mapping->page_tree, page->index,
+                       radix_tree_tag_clear(&mapping->page_tree,
+                                               page_index(page),
                                                PAGECACHE_TAG_DIRTY);
                        spin_unlock_irqrestore(&mapping->tree_lock, flags);
                        if (!mapping->backing_dev_info->memory_backed)
@@ -698,7 +725,8 @@ int __clear_page_dirty(struct page *page)
 
                spin_lock_irqsave(&mapping->tree_lock, flags);
                if (TestClearPageDirty(page)) {
-                       radix_tree_tag_clear(&mapping->page_tree, page->index,
+                       radix_tree_tag_clear(&mapping->page_tree,
+                                               page_index(page),
                                                PAGECACHE_TAG_DIRTY);
                        spin_unlock_irqrestore(&mapping->tree_lock, flags);
                        return 1;
@@ -720,7 +748,8 @@ int test_clear_page_writeback(struct page *page)
                spin_lock_irqsave(&mapping->tree_lock, flags);
                ret = TestClearPageWriteback(page);
                if (ret)
-                       radix_tree_tag_clear(&mapping->page_tree, page->index,
+                       radix_tree_tag_clear(&mapping->page_tree,
+                                               page_index(page),
                                                PAGECACHE_TAG_WRITEBACK);
                spin_unlock_irqrestore(&mapping->tree_lock, flags);
        } else {
@@ -740,10 +769,12 @@ int test_set_page_writeback(struct page *page)
                spin_lock_irqsave(&mapping->tree_lock, flags);
                ret = TestSetPageWriteback(page);
                if (!ret)
-                       radix_tree_tag_set(&mapping->page_tree, page->index,
+                       radix_tree_tag_set(&mapping->page_tree,
+                                               page_index(page),
                                                PAGECACHE_TAG_WRITEBACK);
                if (!PageDirty(page))
-                       radix_tree_tag_clear(&mapping->page_tree, page->index,
+                       radix_tree_tag_clear(&mapping->page_tree,
+                                               page_index(page),
                                                PAGECACHE_TAG_DIRTY);
                spin_unlock_irqrestore(&mapping->tree_lock, flags);
        } else {