Now these functions are page based, this is not good and if lots of pages are dirtied at the same time, only the first dirty page is doing real work and others just repeat useless work. Furthermore I've already seen huge contention on lock lli_lock so we should revise it to be inode based.

One solution is to invent an atomic count to remember how many dirty pages there are for the inode, and if this count is dropped to zero we'll send done writing RPC to MDS.