writeback: fix WB_SYNC_NONE writeback from umount
[pandora-kernel.git] / fs / fs-writeback.c
index 781a322..67db897 100644 (file)
@@ -45,6 +45,7 @@ struct wb_writeback_args {
        int for_kupdate:1;
        int range_cyclic:1;
        int for_background:1;
+       int sb_pinned:1;
 };
 
 /*
@@ -230,6 +231,11 @@ static void bdi_sync_writeback(struct backing_dev_info *bdi,
                .sync_mode      = WB_SYNC_ALL,
                .nr_pages       = LONG_MAX,
                .range_cyclic   = 0,
+               /*
+                * Setting sb_pinned is not necessary for WB_SYNC_ALL, but
+                * lets make it explicitly clear.
+                */
+               .sb_pinned      = 1,
        };
        struct bdi_work work;
 
@@ -245,21 +251,23 @@ static void bdi_sync_writeback(struct backing_dev_info *bdi,
  * @bdi: the backing device to write from
  * @sb: write inodes from this super_block
  * @nr_pages: the number of pages to write
+ * @sb_locked: caller already holds sb umount sem.
  *
  * Description:
  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
  *   started when this function returns, we make no guarentees on
- *   completion. Caller need not hold sb s_umount semaphore.
+ *   completion. Caller specifies whether sb umount sem is held already or not.
  *
  */
 void bdi_start_writeback(struct backing_dev_info *bdi, struct super_block *sb,
-                        long nr_pages)
+                        long nr_pages, int sb_locked)
 {
        struct wb_writeback_args args = {
                .sb             = sb,
                .sync_mode      = WB_SYNC_NONE,
                .nr_pages       = nr_pages,
                .range_cyclic   = 1,
+               .sb_pinned      = sb_locked,
        };
 
        /*
@@ -554,108 +562,85 @@ select_queue:
        return ret;
 }
 
-static void unpin_sb_for_writeback(struct super_block **psb)
+static void unpin_sb_for_writeback(struct super_block *sb)
 {
-       struct super_block *sb = *psb;
-
-       if (sb) {
-               up_read(&sb->s_umount);
-               put_super(sb);
-               *psb = NULL;
-       }
+       up_read(&sb->s_umount);
+       put_super(sb);
 }
 
+enum sb_pin_state {
+       SB_PINNED,
+       SB_NOT_PINNED,
+       SB_PIN_FAILED
+};
+
 /*
  * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
  * before calling writeback. So make sure that we do pin it, so it doesn't
  * go away while we are writing inodes from it.
- *
- * Returns 0 if the super was successfully pinned (or pinning wasn't needed),
- * 1 if we failed.
  */
-static int pin_sb_for_writeback(struct writeback_control *wbc,
-                               struct inode *inode, struct super_block **psb)
+static enum sb_pin_state pin_sb_for_writeback(struct writeback_control *wbc,
+                                             struct super_block *sb)
 {
-       struct super_block *sb = inode->i_sb;
-
-       /*
-        * If this sb is already pinned, nothing more to do. If not and
-        * *psb is non-NULL, unpin the old one first
-        */
-       if (sb == *psb)
-               return 0;
-       else if (*psb)
-               unpin_sb_for_writeback(psb);
-
        /*
         * Caller must already hold the ref for this
         */
-       if (wbc->sync_mode == WB_SYNC_ALL) {
+       if (wbc->sync_mode == WB_SYNC_ALL || wbc->sb_pinned) {
                WARN_ON(!rwsem_is_locked(&sb->s_umount));
-               return 0;
+               return SB_NOT_PINNED;
        }
-
        spin_lock(&sb_lock);
        sb->s_count++;
        if (down_read_trylock(&sb->s_umount)) {
                if (sb->s_root) {
                        spin_unlock(&sb_lock);
-                       goto pinned;
+                       return SB_PINNED;
                }
                /*
                 * umounted, drop rwsem again and fall through to failure
                 */
                up_read(&sb->s_umount);
        }
-
        sb->s_count--;
        spin_unlock(&sb_lock);
-       return 1;
-pinned:
-       *psb = sb;
-       return 0;
+       return SB_PIN_FAILED;
 }
 
-static void writeback_inodes_wb(struct bdi_writeback *wb,
-                               struct writeback_control *wbc)
+/*
+ * Write a portion of b_io inodes which belong to @sb.
+ * If @wbc->sb != NULL, then find and write all such
+ * inodes. Otherwise write only ones which go sequentially
+ * in reverse order.
+ * Return 1, if the caller writeback routine should be
+ * interrupted. Otherwise return 0.
+ */
+static int writeback_sb_inodes(struct super_block *sb,
+                              struct bdi_writeback *wb,
+                              struct writeback_control *wbc)
 {
-       struct super_block *sb = wbc->sb, *pin_sb = NULL;
-       const unsigned long start = jiffies;    /* livelock avoidance */
-
-       spin_lock(&inode_lock);
-
-       if (!wbc->for_kupdate || list_empty(&wb->b_io))
-               queue_io(wb, wbc->older_than_this);
-
        while (!list_empty(&wb->b_io)) {
-               struct inode *inode = list_entry(wb->b_io.prev,
-                                               struct inode, i_list);
                long pages_skipped;
-
-               /*
-                * super block given and doesn't match, skip this inode
-                */
-               if (sb && sb != inode->i_sb) {
+               struct inode *inode = list_entry(wb->b_io.prev,
+                                                struct inode, i_list);
+               if (wbc->sb && sb != inode->i_sb) {
+                       /* super block given and doesn't
+                          match, skip this inode */
                        redirty_tail(inode);
                        continue;
                }
-
+               if (sb != inode->i_sb)
+                       /* finish with this superblock */
+                       return 0;
                if (inode->i_state & (I_NEW | I_WILL_FREE)) {
                        requeue_io(inode);
                        continue;
                }
-
                /*
                 * Was this inode dirtied after sync_sb_inodes was called?
                 * This keeps sync from extra jobs and livelock.
                 */
-               if (inode_dirtied_after(inode, start))
-                       break;
-
-               if (pin_sb_for_writeback(wbc, inode, &pin_sb)) {
-                       requeue_io(inode);
-                       continue;
-               }
+               if (inode_dirtied_after(inode, wbc->wb_start))
+                       return 1;
 
                BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
                __iget(inode);
@@ -674,14 +659,50 @@ static void writeback_inodes_wb(struct bdi_writeback *wb,
                spin_lock(&inode_lock);
                if (wbc->nr_to_write <= 0) {
                        wbc->more_io = 1;
-                       break;
+                       return 1;
                }
                if (!list_empty(&wb->b_more_io))
                        wbc->more_io = 1;
        }
+       /* b_io is empty */
+       return 1;
+}
+
+static void writeback_inodes_wb(struct bdi_writeback *wb,
+                               struct writeback_control *wbc)
+{
+       int ret = 0;
+
+       wbc->wb_start = jiffies; /* livelock avoidance */
+       spin_lock(&inode_lock);
+       if (!wbc->for_kupdate || list_empty(&wb->b_io))
+               queue_io(wb, wbc->older_than_this);
+
+       while (!list_empty(&wb->b_io)) {
+               struct inode *inode = list_entry(wb->b_io.prev,
+                                                struct inode, i_list);
+               struct super_block *sb = inode->i_sb;
+               enum sb_pin_state state;
+
+               if (wbc->sb && sb != wbc->sb) {
+                       /* super block given and doesn't
+                          match, skip this inode */
+                       redirty_tail(inode);
+                       continue;
+               }
+               state = pin_sb_for_writeback(wbc, sb);
 
-       unpin_sb_for_writeback(&pin_sb);
+               if (state == SB_PIN_FAILED) {
+                       requeue_io(inode);
+                       continue;
+               }
+               ret = writeback_sb_inodes(sb, wb, wbc);
 
+               if (state == SB_PINNED)
+                       unpin_sb_for_writeback(sb);
+               if (ret)
+                       break;
+       }
        spin_unlock(&inode_lock);
        /* Leave any unwritten inodes on b_io */
 }
@@ -738,6 +759,7 @@ static long wb_writeback(struct bdi_writeback *wb,
                .for_kupdate            = args->for_kupdate,
                .for_background         = args->for_background,
                .range_cyclic           = args->range_cyclic,
+               .sb_pinned              = args->sb_pinned,
        };
        unsigned long oldest_jif;
        long wrote = 0;
@@ -839,6 +861,12 @@ static long wb_check_old_data_flush(struct bdi_writeback *wb)
        unsigned long expired;
        long nr_pages;
 
+       /*
+        * When set to zero, disable periodic writeback
+        */
+       if (!dirty_writeback_interval)
+               return 0;
+
        expired = wb->last_old_flush +
                        msecs_to_jiffies(dirty_writeback_interval * 10);
        if (time_before(jiffies, expired))
@@ -934,8 +962,12 @@ int bdi_writeback_task(struct bdi_writeback *wb)
                                break;
                }
 
-               wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
-               schedule_timeout_interruptible(wait_jiffies);
+               if (dirty_writeback_interval) {
+                       wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
+                       schedule_timeout_interruptible(wait_jiffies);
+               } else
+                       schedule();
+
                try_to_freeze();
        }
 
@@ -1170,6 +1202,18 @@ static void wait_sb_inodes(struct super_block *sb)
        iput(old_inode);
 }
 
+static void __writeback_inodes_sb(struct super_block *sb, int sb_locked)
+{
+       unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
+       unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
+       long nr_to_write;
+
+       nr_to_write = nr_dirty + nr_unstable +
+                       (inodes_stat.nr_inodes - inodes_stat.nr_unused);
+
+       bdi_start_writeback(sb->s_bdi, sb, nr_to_write, sb_locked);
+}
+
 /**
  * writeback_inodes_sb -       writeback dirty inodes from given super_block
  * @sb: the superblock
@@ -1181,17 +1225,22 @@ static void wait_sb_inodes(struct super_block *sb)
  */
 void writeback_inodes_sb(struct super_block *sb)
 {
-       unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
-       unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
-       long nr_to_write;
-
-       nr_to_write = nr_dirty + nr_unstable +
-                       (inodes_stat.nr_inodes - inodes_stat.nr_unused);
-
-       bdi_start_writeback(sb->s_bdi, sb, nr_to_write);
+       __writeback_inodes_sb(sb, 0);
 }
 EXPORT_SYMBOL(writeback_inodes_sb);
 
+/**
+ * writeback_inodes_sb_locked  - writeback dirty inodes from given super_block
+ * @sb: the superblock
+ *
+ * Like writeback_inodes_sb(), except the caller already holds the
+ * sb umount sem.
+ */
+void writeback_inodes_sb_locked(struct super_block *sb)
+{
+       __writeback_inodes_sb(sb, 1);
+}
+
 /**
  * writeback_inodes_sb_if_idle -       start writeback if none underway
  * @sb: the superblock