crash on device removal
Ming Lin
mlin at kernel.org
Tue Jul 12 14:47:56 PDT 2016
On Tue, 2016-07-12 at 16:09 -0500, Steve Wise wrote:
> > On Tue, 2016-07-12 at 11:34 -0500, Steve Wise wrote:
> > > Hey Christoph,
> > >
> > > I see a crash when shutting down a nvme host node via 'reboot' that has 1 target
> > > device attached. The shutdown causes iw_cxgb4 to be removed which triggers
> > the
> > > device removal logic in the nvmf rdma transport. The crash is here:
> > >
> > > (gdb) list *nvme_rdma_free_qe+0x18
> > > 0x1e8 is in nvme_rdma_free_qe (drivers/nvme/host/rdma.c:196).
> > > 191 }
> > > 192
> > > 193 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct
> > > nvme_rdma_qe *qe,
> > > 194 size_t capsule_size, enum dma_data_direction dir)
> > > 195 {
> > > 196 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
> > > 197 kfree(qe->data);
> > > 198 }
> > > 199
> > > 200 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct
> > > nvme_rdma_qe *qe,
> > >
> > > Apparently qe is NULL.
> > >
> > > Looking at the device removal path, the logic appears correct (see
> > > nvme_rdma_device_unplug() and the nice function comment :) ). I'm wondering
> > if
> > > concurrently to the host device removal path cleaning up queues, the target is
> > > disconnecting all of its queues due to the first disconnect event from the host
> > > causing some cleanup race on the host side? Although since the removal path
> > > executing in the cma event handler upcall, I don't think another thread would be
> > > handling a disconnect event. Maybe the qp async event handler flow?
> > >
> > > Thoughts?
> >
> > We actually missed a kref_get in nvme_get_ns_from_disk().
> >
> > This should fix it. Could you help to verify?
> >
> > diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
> > index 4babdf0..b146f52 100644
> > --- a/drivers/nvme/host/core.c
> > +++ b/drivers/nvme/host/core.c
> > @@ -183,6 +183,8 @@ static struct nvme_ns *nvme_get_ns_from_disk(struct
> > gendisk *disk)
> > }
> > spin_unlock(&dev_list_lock);
> >
> > + kref_get(&ns->ctrl->kref);
> > +
> > return ns;
> >
> > fail_put_ns:
>
> Hey Ming. This avoids the crash in nvme_rdma_free_qe(), but now I see another crash:
>
> [ 975.633436] nvme nvme0: new ctrl: NQN "nqn.2014-08.org.nvmexpress.discovery", addr 10.0.1.14:4420
> [ 978.463636] nvme nvme0: creating 32 I/O queues.
> [ 979.187826] nvme nvme0: new ctrl: NQN "testnqn", addr 10.0.1.14:4420
> [ 987.778287] nvme nvme0: Got rdma device removal event, deleting ctrl
> [ 987.882202] BUG: unable to handle kernel paging request at ffff880e770e01f8
> [ 987.890024] IP: [<ffffffffa03a1a46>] __ib_process_cq+0x46/0xc0 [ib_core]
>
> This looks like another problem with freeing the tag sets before stopping the QP. I thought we fixed that once and for all, but perhaps there is some other path we missed. :(
Sorry, the previous patch was wrong.
Here is the right one.
diff --git a/drivers/nvme/host/fabrics.c b/drivers/nvme/host/fabrics.c
index 1ad47c5..f13e3a6 100644
--- a/drivers/nvme/host/fabrics.c
+++ b/drivers/nvme/host/fabrics.c
@@ -845,6 +845,7 @@ static ssize_t nvmf_dev_write(struct file *file, const char __user *ubuf,
goto out_unlock;
}
+ kref_get(&ctrl->kref);
seq_file->private = ctrl;
out_unlock:
More information about the Linux-nvme
mailing list