On Wed, Jun 05, 2024 at 07:10:55PM +0800, Hillf Danton wrote: > On Tue, 4 Jun 2024 21:58:04 +0300 Leon Romanovsky <leon@xxxxxxxxxx> > > On Tue, Jun 04, 2024 at 06:30:49AM -1000, Tejun Heo wrote: > > > On Tue, Jun 04, 2024 at 02:38:34PM +0300, Leon Romanovsky wrote: > > > > Thanks, it is very rare situation where call to flush/drain queue > > > > (in our case kthread_flush_worker) in the middle of the allocation > > > > flow can be correct. I can't remember any such case. > > > > > > > > So even we don't fully understand the root cause, the reimplementation > > > > is still valid and improves existing code. > > > > > > It's not valid. pwq release is async and while wq free in the error path > > > isn't. The flush is there so that we finish the async part before > > > synchronize error handling. The patch you posted will can lead to double > > > free after a pwq allocation failure. We can make the error path synchronous > > > but the pwq free path should be updated first so that it stays synchronous > > > in the error path. Note that it *needs* to be asynchronous in non-error > > > paths, so it's going to be a bit subtle one way or the other. > > > > But at that point, we didn't add newly created WQ to any list which will execute > > that asynchronous release. Did I miss something? > > > Maybe it is more subtle than thought, but not difficult to make the wq > allocation path sync. See if the patch could survive your test. Thanks, I started to run our tests with Dan's revert. https://lore.kernel.org/all/171711745834.1628941.5259278474013108507.stgit@xxxxxxxxxxxxxxxxxxxxxxxxx/ As premature results, it fixed my lockdep warnings, but it will take time till I get full confidence. If not, I will try your patch. Thanks > > --- x/include/linux/workqueue.h > +++ y/include/linux/workqueue.h > @@ -402,6 +402,7 @@ enum wq_flags { > */ > WQ_POWER_EFFICIENT = 1 << 7, > > + __WQ_INITIALIZING = 1 << 14, /* internal: workqueue is initializing */ > __WQ_DESTROYING = 1 << 15, /* internal: workqueue is destroying */ > __WQ_DRAINING = 1 << 16, /* internal: workqueue is draining */ > __WQ_ORDERED = 1 << 17, /* internal: workqueue is ordered */ > --- x/kernel/workqueue.c > +++ y/kernel/workqueue.c > @@ -5080,6 +5080,8 @@ static void pwq_release_workfn(struct kt > * is gonna access it anymore. Schedule RCU free. > */ > if (is_last) { > + if (wq->flags & __WQ_INITIALIZING) > + return; > wq_unregister_lockdep(wq); > call_rcu(&wq->rcu, rcu_free_wq); > } > @@ -5714,8 +5716,10 @@ struct workqueue_struct *alloc_workqueue > goto err_unreg_lockdep; > } > > + wq->flags |= __WQ_INITIALIZING; > if (alloc_and_link_pwqs(wq) < 0) > goto err_free_node_nr_active; > + wq->flags &= ~__WQ_INITIALIZING; > > if (wq_online && init_rescuer(wq) < 0) > goto err_destroy; > -- >
