Commit 74e72f894d56 ("lib/percpu_counter.c: fix __percpu_counter_add()")
looked very plausible, but its arithmetic was badly wrong: obvious once
you see the fix, but maddening to get there from the weird tmpfs ENOSPCs
That would explain why there is ENOSPC reported without du agreeing when counting the usage.
There is one commit that is not yet in 3.13.0-4 but has a very suspicious comment in the commit:
commit d1969a84dd6a44d 375aa82bba7d6c3 8713a429c3
Author: Hugh Dickins <email address hidden>
Date: Thu Jan 16 15:26:48 2014 -0800
percpu_counter: unbreak __percpu_ counter_ add()
Commit 74e72f894d56 ("lib/percpu_ counter. c: fix __percpu_ counter_ add()")
looked very plausible, but its arithmetic was badly wrong: obvious once
you see the fix, but maddening to get there from the weird tmpfs ENOSPCs
That would explain why there is ENOSPC reported without du agreeing when counting the usage.