writeback: ensure that WB_SYNC_NONE writeback with sb pinned is sync
authorJens Axboe <jens.axboe@oracle.com>
Tue, 18 May 2010 12:29:29 +0000 (14:29 +0200)
committerJens Axboe <jens.axboe@oracle.com>
Fri, 21 May 2010 18:00:25 +0000 (20:00 +0200)
Even if the writeout itself isn't a data integrity operation, we need
to ensure that the caller doesn't drop the sb umount sem before we
have actually done the writeback.

This is a fixup for commit e913fc82.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
fs/fs-writeback.c

index 0f629571234f001a9576d95bf102d54769c47c37..76f546d56a646d0eceb6b6147af097e00aaa1e01 100644 (file)
@@ -193,7 +193,8 @@ static void bdi_wait_on_work_clear(struct bdi_work *work)
 }
 
 static void bdi_alloc_queue_work(struct backing_dev_info *bdi,
-                                struct wb_writeback_args *args)
+                                struct wb_writeback_args *args,
+                                int wait)
 {
        struct bdi_work *work;
 
@@ -205,6 +206,8 @@ static void bdi_alloc_queue_work(struct backing_dev_info *bdi,
        if (work) {
                bdi_work_init(work, args);
                bdi_queue_work(bdi, work);
+               if (wait)
+                       bdi_wait_on_work_clear(work);
        } else {
                struct bdi_writeback *wb = &bdi->wb;
 
@@ -279,7 +282,7 @@ void bdi_start_writeback(struct backing_dev_info *bdi, struct super_block *sb,
                args.for_background = 1;
        }
 
-       bdi_alloc_queue_work(bdi, &args);
+       bdi_alloc_queue_work(bdi, &args, sb_locked);
 }
 
 /*
@@ -909,6 +912,7 @@ long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
 
        while ((work = get_next_work_item(bdi, wb)) != NULL) {
                struct wb_writeback_args args = work->args;
+               int post_clear;
 
                /*
                 * Override sync mode, in case we must wait for completion
@@ -916,11 +920,13 @@ long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
                if (force_wait)
                        work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
 
+               post_clear = WB_SYNC_ALL || args.sb_pinned;
+
                /*
                 * If this isn't a data integrity operation, just notify
                 * that we have seen this work and we are now starting it.
                 */
-               if (args.sync_mode == WB_SYNC_NONE)
+               if (!post_clear)
                        wb_clear_pending(wb, work);
 
                wrote += wb_writeback(wb, &args);
@@ -929,7 +935,7 @@ long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
                 * This is a data integrity writeback, so only do the
                 * notification when we have completed the work.
                 */
-               if (args.sync_mode == WB_SYNC_ALL)
+               if (post_clear)
                        wb_clear_pending(wb, work);
        }
 
@@ -1000,7 +1006,7 @@ static void bdi_writeback_all(struct super_block *sb, long nr_pages)
                if (!bdi_has_dirty_io(bdi))
                        continue;
 
-               bdi_alloc_queue_work(bdi, &args);
+               bdi_alloc_queue_work(bdi, &args, 0);
        }
 
        rcu_read_unlock();