Messages in this thread | | | Subject | Re: workqueue thing | From | Peter Zijlstra <> | Date | Fri, 18 Dec 2009 16:39:48 +0100 |
| |
On Fri, 2009-12-18 at 07:30 -0800, Linus Torvalds wrote: > > On Fri, 18 Dec 2009, Peter Zijlstra wrote: > > > > r1. The first design goal of cmwq is solving the issues the current > > > workqueue implementation has including hard to detect > > > deadlocks, > > > > lockdep is quite proficient at finding these these days. > > I don't think so. > > The reason it is not is that workqueues fundamentally do _different_ > things in the same context, adn lockdep has no clue what-so-ever. > > IOW, if you hold a lock, and then do 'flush_workqueue()', lockdep has no > idea that maybe one of the entries on a workqueue might need the lock that > you are holding. But I don't think lockdep sees the dependency that gets > created by the flush - because it's not a direct code execution > dependency. > > It's not a deadlock _directly_ due to lock ordering, but indirectly due to > waiting for unrelated code that needs locks. > > Now, maybe lockdep could be _taught_ to consider workqueues themselves to > be 'locks', and ordering those pseudo-locks wrt the real locks they take. > So if workqueue Q takes lock A, the fact that it is _taken_ in a workqueue > makes the ordering be Q->A. Then, if somebody does a "flush_workqueue" > while holding lock B, the flush implies a "lock ordering" of B->Q (where > "Q" is the set of all workqueues that could be flushed).
That's exactly what it does..
4e6045f134784f4b158b3c0f7a282b04bd816887 eb13ba873881abd5e15af784756a61af635e665e a67da70dc0955580665f5444f318b92e69a3c272
| |