On Wed, Apr 17, 2019 at 07:22:59PM +0000, Parav Pandit wrote:
> Hi Mike,
>
> > From: linux-rdma-owner@xxxxxxxxxxxxxxx <linux-rdma-
> > owner@xxxxxxxxxxxxxxx> On Behalf Of Marciniszyn, Mike
> > Sent: Wednesday, April 17, 2019 1:56 PM
> > To: Leon Romanovsky <leonro@xxxxxxxxxxxx>
> > Cc: linux-rdma@xxxxxxxxxxxxxxx
> > Subject: new trace in RDMA next
> >
> > Leon,
> >
> > We seem to be getting a new trace with the RDMA for-next patches.
> >
> > Here is the trace:
> >
> > systemd: Starting Hostname Service...
> > kobject (00000000b8a5bae6): tried to init an initialized object, something is
> > seriously wrong.
> > CPU: 68 PID: 2098 Comm: (ostnamed) Not tainted 5.1.0-rc4 #1
> >
> > Call Trace:
> > dump_stack+0x5a/0x73
> > kobject_init+0x74/0x80
> > kobject_init_and_add+0x35/0xb0
> > hfi1_create_port_files+0x6e/0x3c0 [hfi1]
> > ib_setup_port_attrs+0x43b/0x560 [ib_core]
> > add_one_compat_dev+0x16a/0x230 [ib_core]
> > rdma_dev_init_net+0x110/0x160 [ib_core]
> > ops_init+0x38/0xf0
> > setup_net+0xcf/0x1e0
> > copy_net_ns+0xb7/0x130
> > create_new_namespaces+0x11a/0x1b0
> > unshare_nsproxy_namespaces+0x55/0xa0
> > ksys_unshare+0x1a7/0x340
> > __x64_sys_unshare+0xe/0x20
> > do_syscall_64+0x5b/0x180
> > entry_SYSCALL_64_after_hwframe+0x44/0xa9
> >
> > It seems the core is calling init_port(hfi1_create_port_files) twice, the first
> > during ib_register_device() and the second from the stack above.
> >
> > There is a comment in add_one_compat_dev() noting the race:
> >
> > /*
> > * The first of init_net() or ib_register_device() to take the
> > * compat_devs_mutex wins and gets to add the device. Others will wait
> > * for completion here.
> > */
> > mutex_lock(&device->compat_devs_mutex);
> > cdev = xa_load(&device->compat_devs, rnet->id); if (cdev) {
> > ret = 0;
> > goto done;
> > }
> >
> > I don't see any xa_store() to compat_devs or use of the mutex in the
> > ib_register_device() path?
> >
> The issue is not with xa_store or lock.
> Issue is that hfi1_create_port_files() is trying to create sysfs files out ppd->sc2vl_kobj, which is already initialized during ib_register_device().
> For every net ns, my patches are adding compat devices which probably shouldn't invoke init_port for those compat devices as it not meaningful.
>
> Attached patch should fix it.
> Can you please test it? I do not have hfi hw.
> I will inline the patch, in case attachment doesn't work.
convert the double bool arguments to a flag please before you send it.
Jason
