Hello,

On Wed, Dec 13, 2017 at 12:00:04PM +0100, Jan Kara wrote:

OK, but this effectively prevents writeback from sync_inodes_sb() to ever
make inode switch wbs. Cannot that be abused in some way like making sure
writeback of our memcg is "invisible" by forcing it out using sync(2)? It
just looks a bit dangerous to me...

That's true.  There are a couple mitigating factors tho.

* While it can delay switching during sync(2), it'll all still be
  recorded and the switch will happen soon if needed.

* sync(2) is hugely disruptive with or without this and can easily be
  used to DOS the whole system.  People are working on restricting the
  blast radius of sync(2) to mitigate this problem, which most likely
  make this a non-problem too.

If you can think of a better solution, I'm all ears.

Thanks.

-- 
tejun