Re: netconsole deadlock with virtnet

[Leon Romanovsky <leon@kernel.org>]

On Tue, Nov 24, 2020 at 04:57:23PM +0800, Jason Wang wrote:
>
> On 2020/11/24 下午4:01, Leon Romanovsky wrote:
> > On Tue, Nov 24, 2020 at 11:22:03AM +0800, Jason Wang wrote:
> > > On 2020/11/24 上午3:21, Jakub Kicinski wrote:
> > > > On Mon, 23 Nov 2020 14:09:34 -0500 Steven Rostedt wrote:
> > > > > On Mon, 23 Nov 2020 10:52:52 -0800
> > > > > Jakub Kicinski <kuba@kernel.org> wrote:
> > > > >
> > > > > > On Mon, 23 Nov 2020 09:31:28 -0500 Steven Rostedt wrote:
> > > > > > > On Mon, 23 Nov 2020 13:08:55 +0200
> > > > > > > Leon Romanovsky <leon@kernel.org> wrote:
> > > > > > >
> > > > > > > >    [   10.028024] Chain exists of:
> > > > > > > >    [   10.028025]   console_owner --> target_list_lock --> _xmit_ETHER#2
> > > > > > > Note, the problem is that we have a location that grabs the xmit_lock while
> > > > > > > holding target_list_lock (and possibly console_owner).
> > > > > > Well, it try_locks the xmit_lock. Does lockdep understand try-locks?
> > > > > >
> > > > > > (not that I condone the shenanigans that are going on here)
> > > > > Does it?
> > > > >
> > > > > 	virtnet_poll_tx() {
> > > > > 		__netif_tx_lock() {
> > > > > 			spin_lock(&txq->_xmit_lock);
> > > > Umpf. Right. I was looking at virtnet_poll_cleantx()
> > > >
> > > > > That looks like we can have:
> > > > >
> > > > >
> > > > > 	CPU0		CPU1
> > > > > 	----		----
> > > > >      lock(xmit_lock)
> > > > >
> > > > > 		    lock(console)
> > > > > 		    lock(target_list_lock)
> > > > > 		    __netif_tx_lock()
> > > > > 		        lock(xmit_lock);
> > > > >
> > > > > 			[BLOCKED]
> > > > >
> > > > >      <interrupt>
> > > > >      lock(console)
> > > > >
> > > > >      [BLOCKED]
> > > > >
> > > > >
> > > > >
> > > > >    DEADLOCK.
> > > > >
> > > > >
> > > > > So where is the trylock here?
> > > > >
> > > > > Perhaps you need the trylock in virtnet_poll_tx()?
> > > > That could work. Best if we used normal lock if !!budget, and trylock
> > > > when budget is 0. But maybe that's too hairy.
> > >
> > > If we use trylock, we probably lose(or delay) tx notification that may have
> > > side effects to the stack.
> > >
> > >
> > > > I'm assuming all this trickiness comes from virtqueue_get_buf() needing
> > > > locking vs the TX path? It's pretty unusual for the completion path to
> > > > need locking vs xmit path.
> > >
> > > Two reasons for doing this:
> > >
> > > 1) For some historical reason, we try to free transmitted tx packets in xmit
> > > (see free_old_xmit_skbs() in start_xmit()), we can probably remove this if
> > > we remove the non tx interrupt mode.
> > > 2) virtio core requires virtqueue_get_buf() to be synchronized with
> > > virtqueue_add(), we probably can solve this but it requires some non trivial
> > > refactoring in the virtio core
> > So how will we solve our lockdep issues?
> >
> > Thanks
>
>
> It's not clear to me that whether it's a virtio-net specific issue. E.g the
> above deadlock looks like a generic issue so workaround it via virtio-net
> may not help for other drivers.

It is hard to say, no one else complained except me who is using virtio :).

Thanks

>
> Thanks
>
>
> >
> > > Btw, have a quick search, there are several other drivers that uses tx lock
> > > in the tx NAPI.
> > >
> > > Thanks
> > >
>