writeback: make writeback initiation functions handle multiple bdi_writeback's
authorTejun Heo <tj@kernel.org>
Fri, 22 May 2015 21:14:01 +0000 (17:14 -0400)
committerJens Axboe <axboe@fb.com>
Tue, 2 Jun 2015 14:33:37 +0000 (08:33 -0600)
[try_]writeback_inodes_sb[_nr]() and sync_inodes_sb() currently only
handle dirty inodes on the root wb (bdi_writeback) of the target bdi.
This patch implements bdi_split_work_to_wbs() and use it to make these
functions handle multiple wb's.

bdi_split_work_to_wbs() takes a base wb_writeback_work and create
clones of it and issue them to the wb's of the target bdi.  The base
work's nr_pages is distributed using wb_split_bdi_pages() -
ie. according to each wb's write bandwidth's proportion in the bdi.

Cloning a bdi involves memory allocation which may fail.  In such
cases, bdi_split_work_to_wbs() issues the base work directly and waits
for its completion before proceeding to the next wb to guarantee
forward progress and correctness under memory pressure.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
fs/fs-writeback.c

index 0039c5839cdd6e66444ea1cbc135858735cf4c4c..59d76f6c9d8dbf9ae2f4a0f1969b2fb0886e8014 100644 (file)
@@ -292,6 +292,80 @@ static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
                return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
 }
 
+/**
+ * wb_clone_and_queue_work - clone a wb_writeback_work and issue it to a wb
+ * @wb: target bdi_writeback
+ * @base_work: source wb_writeback_work
+ *
+ * Try to make a clone of @base_work and issue it to @wb.  If cloning
+ * succeeds, %true is returned; otherwise, @base_work is issued directly
+ * and %false is returned.  In the latter case, the caller is required to
+ * wait for @base_work's completion using wb_wait_for_single_work().
+ *
+ * A clone is auto-freed on completion.  @base_work never is.
+ */
+static bool wb_clone_and_queue_work(struct bdi_writeback *wb,
+                                   struct wb_writeback_work *base_work)
+{
+       struct wb_writeback_work *work;
+
+       work = kmalloc(sizeof(*work), GFP_ATOMIC);
+       if (work) {
+               *work = *base_work;
+               work->auto_free = 1;
+               work->single_wait = 0;
+       } else {
+               work = base_work;
+               work->auto_free = 0;
+               work->single_wait = 1;
+       }
+       work->single_done = 0;
+       wb_queue_work(wb, work);
+       return work != base_work;
+}
+
+/**
+ * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
+ * @bdi: target backing_dev_info
+ * @base_work: wb_writeback_work to issue
+ * @skip_if_busy: skip wb's which already have writeback in progress
+ *
+ * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
+ * have dirty inodes.  If @base_work->nr_page isn't %LONG_MAX, it's
+ * distributed to the busy wbs according to each wb's proportion in the
+ * total active write bandwidth of @bdi.
+ */
+static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
+                                 struct wb_writeback_work *base_work,
+                                 bool skip_if_busy)
+{
+       long nr_pages = base_work->nr_pages;
+       int next_blkcg_id = 0;
+       struct bdi_writeback *wb;
+       struct wb_iter iter;
+
+       might_sleep();
+
+       if (!bdi_has_dirty_io(bdi))
+               return;
+restart:
+       rcu_read_lock();
+       bdi_for_each_wb(wb, bdi, &iter, next_blkcg_id) {
+               if (!wb_has_dirty_io(wb) ||
+                   (skip_if_busy && writeback_in_progress(wb)))
+                       continue;
+
+               base_work->nr_pages = wb_split_bdi_pages(wb, nr_pages);
+               if (!wb_clone_and_queue_work(wb, base_work)) {
+                       next_blkcg_id = wb->blkcg_css->id + 1;
+                       rcu_read_unlock();
+                       wb_wait_for_single_work(bdi, base_work);
+                       goto restart;
+               }
+       }
+       rcu_read_unlock();
+}
+
 #else  /* CONFIG_CGROUP_WRITEBACK */
 
 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
@@ -299,6 +373,21 @@ static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
        return nr_pages;
 }
 
+static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
+                                 struct wb_writeback_work *base_work,
+                                 bool skip_if_busy)
+{
+       might_sleep();
+
+       if (bdi_has_dirty_io(bdi) &&
+           (!skip_if_busy || !writeback_in_progress(&bdi->wb))) {
+               base_work->auto_free = 0;
+               base_work->single_wait = 0;
+               base_work->single_done = 0;
+               wb_queue_work(&bdi->wb, base_work);
+       }
+}
+
 #endif /* CONFIG_CGROUP_WRITEBACK */
 
 void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
@@ -1599,10 +1688,7 @@ static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
                return;
        WARN_ON(!rwsem_is_locked(&sb->s_umount));
 
-       if (skip_if_busy && writeback_in_progress(&bdi->wb))
-               return;
-
-       wb_queue_work(&bdi->wb, &work);
+       bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
        wb_wait_for_completion(bdi, &done);
 }
 
@@ -1700,7 +1786,7 @@ void sync_inodes_sb(struct super_block *sb)
                return;
        WARN_ON(!rwsem_is_locked(&sb->s_umount));
 
-       wb_queue_work(&bdi->wb, &work);
+       bdi_split_work_to_wbs(bdi, &work, false);
        wb_wait_for_completion(bdi, &done);
 
        wait_sb_inodes(sb);