[PATCH] nvme/mpath: fix hang when disk goes live over reconnect

[Anton Eidelman]

Looking into making nvme_start_ctrl capable of returning an error.
This is a bit more complicated IMHO, because the in most transports
the ctrl is already in LIVE state when nvme_start_ctrl is invoked,
so we need to bail out carefully.

On Tue, Oct 19, 2021 at 05:37:30PM +0300, Sagi Grimberg wrote:
> 
> 
> On 10/5/21 3:38 PM, Sagi Grimberg wrote:
> > 
> > > > How do we proceed with this fix?
> > > 
> > > Please resend with the suggested updates.
> > > 
> > > > I believe error propagation here is not wanted because:
> > > > 1) A failure to fetch or parse the ANA log should not be considered
> > > >     an error in ctrl initialization.
> > > 
> > > We must handle error that are due to a controller failing to initialize.
> > > 
> > > > 2) Such error will not cause problems in teardown.
> > > > 3) The same failure is possible in ANA work and we do not take
> > > >     any action in such case.
> > > 
> > > Failing in a workqueue is different from failing in the initialization
> > > path.
> > > 
> > > > 4) Adding support for failure to nvme_start_ctrl() adds complexity
> > > >     and does not look useful due to the above 1-3.
> > 
> > The feedback here is that this patch is changing the functionality
> > as before we failed initialization and now we don't.
> > 
> > What is the issue in propagating the error and then modify
> > the call-sites? Shouldn't it be simple enough to do?
> > -- 
> > diff --git a/drivers/nvme/host/fc.c b/drivers/nvme/host/fc.c
> > index aa14ad963d91..d86623b90eea 100644
> > --- a/drivers/nvme/host/fc.c
> > +++ b/drivers/nvme/host/fc.c
> > @@ -3158,8 +3158,11 @@ nvme_fc_create_association(struct nvme_fc_ctrl
> > *ctrl)
> > 
> >          ctrl->ctrl.nr_reconnects = 0;
> > 
> > -       if (changed)
> > -               nvme_start_ctrl(&ctrl->ctrl);
> > +       if (changed) {
> > +               ret = nvme_start_ctrl(&ctrl->ctrl);
> > +               if (ret)
> > +                       goto out_term_aen_ops;
> > +       }
> > 
> >          return 0;       /* Success */
> > 
> > diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
> > index b82492cd7503..59bfdd72a51a 100644
> > --- a/drivers/nvme/host/pci.c
> > +++ b/drivers/nvme/host/pci.c
> > @@ -2827,7 +2827,10 @@ static void nvme_reset_work(struct work_struct
> > *work)
> >                          &nvme_pci_attr_group))
> >                  dev->attrs_added = true;
> > 
> > -       nvme_start_ctrl(&dev->ctrl);
> > +       ret = nvme_start_ctrl(&dev->ctrl);
> > +       if (ret)
> > +               goto out_unlock;
> > +
> >          return;
> > 
> >    out_unlock:
> > diff --git a/drivers/nvme/host/rdma.c b/drivers/nvme/host/rdma.c
> > index 042c594bc57e..9e9e34a012e7 100644
> > --- a/drivers/nvme/host/rdma.c
> > +++ b/drivers/nvme/host/rdma.c
> > @@ -1141,7 +1141,10 @@ static int nvme_rdma_setup_ctrl(struct
> > nvme_rdma_ctrl *ctrl, bool new)
> >                  goto destroy_io;
> >          }
> > 
> > -       nvme_start_ctrl(&ctrl->ctrl);
> > +       ret = nvme_start_ctrl(&ctrl->ctrl);
> > +       if (ret)
> > +               goto destroy_io;
> > +
> >          return 0;
> > 
> >   destroy_io:
> > diff --git a/drivers/nvme/host/tcp.c b/drivers/nvme/host/tcp.c
> > index 10b164f02b5d..22199281ad17 100644
> > --- a/drivers/nvme/host/tcp.c
> > +++ b/drivers/nvme/host/tcp.c
> > @@ -923,6 +923,11 @@ static void nvme_tcp_fail_request(struct
> > nvme_tcp_request *req)
> >          nvme_tcp_end_request(blk_mq_rq_from_pdu(req),
> > NVME_SC_HOST_PATH_ERROR);
> >   }
> > @@ -2045,7 +2052,10 @@ static int nvme_tcp_setup_ctrl(struct nvme_ctrl
> > *ctrl, bool new)
> >                  goto destroy_io;
> >          }
> > 
> > -       nvme_start_ctrl(ctrl);
> > +       ret = nvme_start_ctrl(ctrl);
> > +       if (ret)
> > +               goto destroy_io;
> > +
> >          return 0;
> > 
> >   destroy_io:
> > diff --git a/drivers/nvme/target/loop.c b/drivers/nvme/target/loop.c
> > index 0285ccc7541f..3fa89650c0d1 100644
> > --- a/drivers/nvme/target/loop.c
> > +++ b/drivers/nvme/target/loop.c
> > @@ -490,7 +490,9 @@ static void nvme_loop_reset_ctrl_work(struct
> > work_struct *work)
> >          if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
> >                  WARN_ON_ONCE(1);
> > 
> > -       nvme_start_ctrl(&ctrl->ctrl);
> > +       ret = nvme_start_ctrl(&ctrl->ctrl);
> > +       if (ret)
> > +               goto out_destroy_io;
> > 
> >          return;
> > -- 
> 
> So Anton, can this suggestion work? I think Hannes reported the same
> issue that this patch addresses.