[PATCH v2] nvme: avoid "(efault)" from nvme_sysfs_show_subsysnqn()

[Keith Busch]

On Wed, Feb 03, 2021 at 02:19:44PM -0800, Sagi Grimberg wrote:
> 
> 
> On 2/3/21 8:52 AM, Keith Busch wrote:
> > On Wed, Feb 03, 2021 at 01:31:13PM +0100, Martin Wilck wrote:
> > > On Tue, 2021-02-02 at 08:58 +0100, Christoph Hellwig wrote:
> > > > On Tue, Feb 02, 2021 at 01:59:51AM +0100, mwilck at suse.com wrote:
> > > > > From: Martin Wilck <mwilck at suse.com>
> > > > > 
> > > > > While a controller is still connecting, ctrl->subsys is NULL and
> > > > > thus reading the controller's "subsysnqn" sysfs attribute returns
> > > > > "(efault)". This happens all the time when user space processes
> > > > > "add" events for NVMe controller devices.
> > > > > 
> > > > > Fix it by returning the nvmf_ctrl_options' subsysnqn attribute
> > > > > if ctrl->subsys isn't initialized yet.
> > > > 
> > > > This is just papering over symptoms.  We should not be sending
> > > > events or display sysfs files until the controlle is live.
> > > 
> > > I don't follow. What's wrong with a "connecting" controller device in
> > > sysfs?
> > > 
> > > Controllers can loose connectivity any time. What's the difference
> > > between a controller that is reconnecting after a connection loss and a
> > > freshly created controller that's just not live yet? You wouldn't
> > > delete the former from sysfs, I suppose.
> > > 
> > >  From the user space point of view, it's the same, a controller device
> > > that is not in usable state. I don't see a general problem with that.
> > > Seeing a controller appear in connecting state might actually be useful
> > > for user space.
> > > 
> > > The subsysnqn attribute is correct for controllers that were once
> > > connected and lost connectivity. My patch was only intended to fix the
> > > "just created and never connected" case, where it is not, nothing more.
> > 
> > I think Christoph is suggesting just wait until we have all the
> > properties the driver exports rather than specifically exporting only
> > during a 'live' state. Something like this (untested) should do the
> > trick:
> > 
> > ---
> > diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
> > index 1a3cdc6b1036..4c0258d7eb91 100644
> > --- a/drivers/nvme/host/core.c
> > +++ b/drivers/nvme/host/core.c
> > @@ -3069,6 +3069,9 @@ int nvme_init_identify(struct nvme_ctrl *ctrl)
> >   		if (ret)
> >   			goto out_free;
> > +		ret = cdev_device_add(&ctrl->cdev, ctrl->device);
> > +		if (ret)
> > +			goto out_free;
> 
> This happens on every reset, we should probably just do this once?

Right, and while the function is invoked on each reset, the above code
sequence happens only once. You can't immediately tell from looking at
the diff, but it's in the 'if (!ctrl->identified)' code block.