On Sun, Mar 08, 2020 at 08:59:33PM +0100, Karsten Graul wrote:
> On 08/03/2020 16:01, Leon Romanovsky wrote:
> > On Fri, Mar 06, 2020 at 02:45:18PM +0100, Karsten Graul wrote:
> >> During IB device removal, cancel the event worker before the device
> >> structure is freed. In the worker, check if the device is being
> >> terminated and do not proceed with the event work in that case.
> >>
> >> Fixes: a4cf0443c414 ("smc: introduce SMC as an IB-client")
> >> Reported-by: syzbot+b297c6825752e7a07272@xxxxxxxxxxxxxxxxxxxxxxxxx
> >> Signed-off-by: Karsten Graul <kgraul@xxxxxxxxxxxxx>
> >> Reviewed-by: Ursula Braun <ubraun@xxxxxxxxxxxxx>
> >> ---
> >> net/smc/smc_ib.c | 4 ++++
> >> 1 file changed, 4 insertions(+)
> >>
> >> diff --git a/net/smc/smc_ib.c b/net/smc/smc_ib.c
> >> index d6ba186f67e2..5e4e64a9aa4b 100644
> >> --- a/net/smc/smc_ib.c
> >> +++ b/net/smc/smc_ib.c
> >> @@ -240,6 +240,9 @@ static void smc_ib_port_event_work(struct work_struct *work)
> >> work, struct smc_ib_device, port_event_work);
> >> u8 port_idx;
> >>
> >> + if (list_empty(&smcibdev->list))
> >> + return;
> >> +
> >
> > How can it be true if you are not holding "smc_ib_devices.lock" during
> > execution of smc_ib_port_event_work()?
> >
>
> It is true when smc_ib_remove_dev() runs before the work actually started.
> Other than that its only a shortcut to return earlier, when the item is
> removed from the list after the check then the processing just takes a
> little bit longer...its still save.
The check itself maybe safe, but it can't fix syzkaller bug reported above.
As you said, the smc_ib_remove_dev() can be called immediately after
your list_empty() check and we return to original behavior.
The correct design will be to ensure that smc_ib_port_event_work() is
executed only smcibdev->list is not empty.
Thanks
>
> >> for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
> >> smc_ib_remember_port_attr(smcibdev, port_idx + 1);
> >> clear_bit(port_idx, &smcibdev->port_event_mask);
> >> @@ -582,6 +585,7 @@ static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
> >> smc_smcr_terminate_all(smcibdev);
> >> smc_ib_cleanup_per_ibdev(smcibdev);
> >> ib_unregister_event_handler(&smcibdev->event_handler);
> >> + cancel_work_sync(&smcibdev->port_event_work);
> >> kfree(smcibdev);
> >> }
> >>
> >> --
> >> 2.17.1
> >>
>
> --
> Karsten
>
> (I'm a dude)
>
