On 10/28/18 10:58 AM, Peter Zijlstra wrote:
On Sat, Oct 27, 2018 at 04:37:45PM +1100, Dave Chinner wrote:
As an API, however, this needs to be consistent with
down_read_non_owner()/up_read_non_owner() which are for exactly this
same issue issue (different acquire/release contexts) on the read
side of the rwsem. Indeed, it can probably use the same
infrastructure...
Also, from a quick look, this has been broken for donkeys years,
right?
Hi Peter,
If you are referring to the direct I/O code: the lockdep complaint shown
in my patch description is new. I think the following commit from two
months ago (which seems fine to me) made that lockdep complaint appear:
87915adc3f0a ("workqueue: re-add lockdep dependencies for flushing").
Bart.
