Messages in this thread | | | Date | Thu, 11 Jun 2015 13:37:10 +0200 | From | Peter Zijlstra <> | Subject | Re: [PATCH 13/14] lockdep: Implement lock pinning |
| |
On Fri, Jun 05, 2015 at 11:55:52AM +0200, Ingo Molnar wrote: > > * Peter Zijlstra <peterz@infradead.org> wrote: > > > RFC: a possible alternative API would be something like: > > > > int cookie = lockdep_pin_lock(&foo); > > ... > > lockdep_unpin_lock(&foo, cookie); > > Yeah, this would be even nicer. > > > Where we pick a random number for the pin_count; this makes it > > impossible to sneak a lock break in without also passing the right > > cookie along. > > > > I've not done this because it ends up generating code for !LOCKDEP, > > esp. if you need to pass the cookie around for some reason. > > The cookie could be a zero-size structure, which can be 'passed around' > syntactically but creates no overhead in the code. > > But I'd expect cookie-passing to be a sign of badness in most cases: the lock > should generally be unpinned at the same level of abstraction...
I have tried to make this work, but so far I've failed at making the !LOCKDEP case generate 'similar' code.
Esp, things like:
rq = task_rq_lock(p, flags);
...
task_rq_unlock(rq, p, flags);
Need to somehow pass the cookie, and all pure stack based approaches I've tried ended up being ugly and generating weird code.
So I'll keep the non-cookie approach for now.
| |