Re: [syzbot] unregister_netdevice: waiting for DEV to become free (7)

[wangyufen]

在 2022/11/22 10:13, Jason Gunthorpe 写道:
> On Fri, Nov 18, 2022 at 02:28:53PM +0100, Dmitry Vyukov wrote:
> > On Fri, 18 Nov 2022 at 12:39, syzbot
> > <syzbot+5e70d01ee8985ae62a3b@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
> > > Hello,
> > >
> > > syzbot found the following issue on:
> > >
> > > HEAD commit:    9c8774e629a1 net: eql: Use kzalloc instead of kmalloc/memset
> > > git tree:       net-next
> > > console output: https://syzkaller.appspot.com/x/log.txt?x=17bf6cc8f00000
> > > kernel config:  https://syzkaller.appspot.com/x/.config?x=9eb259db6b1893cf
> > > dashboard link: https://syzkaller.appspot.com/bug?extid=5e70d01ee8985ae62a3b
> > > compiler:       gcc (Debian 10.2.1-6) 10.2.1 20210110, GNU ld (GNU Binutils for Debian) 2.35.2
> > > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=1136d592f00000
> > > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=1193ae64f00000
> > >
> > > Bisection is inconclusive: the issue happens on the oldest tested release.
> > >
> > > bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=167c33a2f00000
> > > final oops:     https://syzkaller.appspot.com/x/report.txt?x=157c33a2f00000
> > > console output: https://syzkaller.appspot.com/x/log.txt?x=117c33a2f00000
> > >
> > > IMPORTANT: if you fix the issue, please add the following tag to the commit:
> > > Reported-by: syzbot+5e70d01ee8985ae62a3b@xxxxxxxxxxxxxxxxxxxxxxxxx
> > >
> > > iwpm_register_pid: Unable to send a nlmsg (client = 2)
> > > infiniband syj1: RDMA CMA: cma_listen_on_dev, error -98
> > > unregister_netdevice: waiting for vlan0 to become free. Usage count = 2
> >
> > +RDMA maintainers
> >
> > There are 4 reproducers and all contain:
> >
> > r0 = socket$nl_rdma(0x10, 0x3, 0x14)
> > sendmsg$RDMA_NLDEV_CMD_NEWLINK(...)
> >
> > Also the preceding print looks related (a bug in the error handling
> > path there?):
> >
> > infiniband syj1: RDMA CMA: cma_listen_on_dev, error -98
>
> I'm pretty sure it is an rxe bug
>
> ib_device_set_netdev() will hold the netdev until the caller destroys
> the ib_device
>
> rxe calls it during rxe_register_device() because the user asked for a
> stacked ib_device on top of the netdev
>
> Presumably rxe needs to have a notifier to also self destroy the rxe
> device if the underlying net device is to be destroyed?
>
> Can someone from rxe check into this?

The following patch may fix the issue：

--- a/drivers/infiniband/core/cma.c
+++ b/drivers/infiniband/core/cma.c
@@ -4049,6 +4049,9 @@ int rdma_listen(struct rdma_cm_id *id, int backlog)
        return 0;
 err:
        id_priv->backlog = 0;
+       if (id_priv->cma_dev)
+               cma_release_dev(id_priv);
+
        /*
         * All the failure paths that lead here will not allow the 
req_handler's
         * to have run.



The causes are as follows:

rdma_listen()
  rdma_bind_addr()
    cma_acquire_dev_by_src_ip()
      cma_attach_to_dev()
        _cma_attach_to_dev()
	      cma_dev_get()
		
  cma_check_port()
  <--The return value is not zero， goto err

err：
<-- The error handling here is missing the operation of cma_release_dev.

> Jason