On Tue, Feb 25, 2020 at 6:46 PM Jason Gunthorpe <jgg@xxxxxxxxxxxx> wrote:
>
> On Tue, Feb 25, 2020 at 09:40:42AM +0530, Selvin Xavier wrote:
> > On Mon, Feb 24, 2020 at 7:13 PM Jason Gunthorpe <jgg@xxxxxxxxxxxx> wrote:
> > >
> > > On Mon, Feb 24, 2020 at 02:49:55AM -0800, Selvin Xavier wrote:
> > > >
> > > > @@ -1785,32 +1777,23 @@ static int __init bnxt_re_mod_init(void)
> > > >
> > > > static void __exit bnxt_re_mod_exit(void)
> > > > {
> > > > - struct bnxt_re_dev *rdev, *next;
> > > > - LIST_HEAD(to_be_deleted);
> > > > + struct bnxt_re_dev *rdev;
> > > >
> > > > + flush_workqueue(bnxt_re_wq);
> > >
> > > What is this for? Usually flushing a work queue before preventing new
> > > work from queueing (ie via unregister) is racy.
> > To flush out any netdev events scheduled to be handled by
> > bnxt_re_task. Mainly to wait for
> > case where we are in the middle of NETDEV_REGISTER event handled from
> > bnxt_re_task.
>
> Then the netdev notifer should be unregistered before doing the flush.
>
Agree that the netdev notifier should be unregistered before invoking
ib_unregister_driver
so that new device additions are prevented. I will move the unregister
code to the beginning of the mod_exit.
But still we have to flush the bnxt_re_wq work queue. bnxt_re_task
work is scheduled from netdev notifier. We need to make sure that
there are no pending work before proceeding with unregister.
I will modify the code to unregister notifier and then delete
bnxt_re_wq workqueue before calling
ib_unregister_driver.
> Jason
