Hello, Vivek.

On Fri, Mar 27, 2015 at 02:06:26PM -0400, Vivek Goyal wrote:
...

quoted

+int mapping_congested(struct address_space *mapping,
+		      struct task_struct *task, int cong_bits)
+{
+	struct inode *inode = mapping->host;
+	struct backing_dev_info *bdi = inode_to_bdi(inode);
+	struct bdi_writeback *wb;
+	int ret = 0;
+
+	if (!inode || !inode_cgwb_enabled(inode))
+		return wb_congested(&bdi->wb, cong_bits);
+
+	rcu_read_lock();
+	wb = wb_find_current(bdi);

I am wondering that why do we lookup bdi_writeback using blkcg of
task and why not use the bdi_writeback associated with inode?

IIUC, whole idea is to attach an inode to bdi_writeback (and
change it later if need be) and that writeback is used for
controlling IO to that inode. And blkcg associated with the
writeback will be put in bio which in turn will be used
by block layer.

IOW, blkcg of a bio gets decided by the bdi_writeback
attached to inode and current writer does not seem to
matter. So I am not sure why mapping_congested() should
take task's blkcg into consideration instead of just
taking bdi_writeback from inode and see if it is congested
or not.

Yeap, I agree that attributing to the inode's blkcg makes more sense.
I need to think more about it but will prolly change it to use
inode->i_wb instead.

Thanks.

-- 
tejun