On Thu, Oct 12, 2017 at 05:56:35PM +0200, Peter Zijlstra wrote: > On Thu, Oct 12, 2017 at 08:38:17AM -0700, Tejun Heo wrote: > > > > As long as we have the same level of protection, simpler code is of > > course preferable. That said, I haven't followed the discussion > > closely and don't want to apply it without Peter's ack. Peter? > > I'm really tied up atm; and feel we should be addressing the false > positives generated by the current code before we start doing new stuff > on top.
We can never avoid adding false dependencies as long as we use acquisitions in that way the workqueue code does, even though you successfully replace write acquisitions with recursive-read ones after making them work, as you know. Speaking a bit more about the reason, it's because all write locks used in every work->func() obviously generate false dependencies(links) with 'work' lockdep_map and 'wq' lockdep_map, when they do not involve flush. This is why I used a word, 'speculative', whenever we were talking. At the beginning of this issue, I suggested to use recursive-read acquisitions instead of write ones, which you are working on, since anyway it reduces false ones. But, if it's allowed to add a new primitive that just suits that purpose, I want to propose it to be used instead, which makes false ones reduced more. A read acuisition is a real acquisition used for a read lock. Semantics are similar to but not same as what we need for that speculative purpose.
