On 24.11.2016 12:56, Peter Zijlstra wrote:
On Thu, Nov 24, 2016 at 12:52:25PM +0100, Daniel Vetter wrote:
On Thu, Nov 24, 2016 at 12:40 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
I do believe we can win a bit by keeping the wait list sorted, if we also
make sure that waiters don't add themselves in the first place if they see
that a deadlock situation cannot be avoided.
I will probably want to extend struct mutex_waiter with ww_mutex-specific
fields to facilitate this (i.e. ctx pointer, perhaps stamp as well to reduce
pointer-chasing). That should be fine since it lives on the stack.
Right, shouldn't be a problem I think.
The only 'problem' I can see with using that is that its possible to mix
ww and !ww waiters through ww_mutex_lock(.ctx = NULL). This makes the
list order somewhat tricky.
Ideally we'd remove that feature, although I see its actually used quite
a bit :/
I guess we could create a small fake acquire_ctx for single-lock
paths. That way callers still don't need to deal with having an
explicit ctx, but we can assume the timestamp (for ensuring fairness)
is available for all cases. Otherwise there's indeed a problem with
correctly (well fairly) interleaving ctx and non-ctx lockers I think.
Actually tried that, but we need a ww_class to get a stamp from, and
ww_mutex_lock() doesn't have one of those..
The acquire context needs to be live until the unlock anyway, so this is
something that requires modifying the callers of ww_mutex_lock. Those
should all have a ww_class available, or something is very wrong :)
Nicolai
--
To unsubscribe from this list: send the line "unsubscribe stable" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html