Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <161117153248.2853729.2452425259045172318.stgit@dwillia2-desk3.amr.corp.intel.com>
 <161117153776.2853729.6944617921517514510.stgit@dwillia2-desk3.amr.corp.intel.com>
 <809823fb-6eb6-8ce9-c49a-d85b03897fc7@deltatee.com>
In-Reply-To: <809823fb-6eb6-8ce9-c49a-d85b03897fc7@deltatee.com>
From:   Dan Williams <dan.j.williams@intel.com>
Date:   Wed, 20 Jan 2021 12:38:27 -0800
Message-ID: <CAPcyv4j8nfFT5=0T5U7rpNnD29XpTaAmNU5fkaA1--CrmWfRUw@mail.gmail.com>
Subject: Re: [PATCH 1/3] cdev: Finish the cdev api with queued mode support
To:     Logan Gunthorpe <logang@deltatee.com>
Cc:     Greg KH <gregkh@linuxfoundation.org>,
        Hans Verkuil <hans.verkuil@cisco.com>,
        Alexandre Belloni <alexandre.belloni@free-electrons.com>,
        Alexander Viro <viro@zeniv.linux.org.uk>,
        Dave Jiang <dave.jiang@intel.com>,
        Vishal L Verma <vishal.l.verma@intel.com>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        linux-nvdimm <linux-nvdimm@lists.01.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, Jan 20, 2021 at 11:51 AM Logan Gunthorpe <logang@deltatee.com> wrote:
>
>
>
>
> On 2021-01-20 12:38 p.m., Dan Williams wrote:
> > ...common reference count handling scenarios were addressed, but the
> > shutdown-synchronization problem was only mentioned as something driver
> > developers need to be aware in the following note:
> >
> >     NOTE: This guarantees that associated sysfs callbacks are not running
> >     or runnable, however any cdevs already open will remain and their fops
> >     will still be callable even after this function returns.
> >
> > Remove that responsibility from driver developers with the concept of a
> > 'queued' mode for cdevs.
>
> I find the queued name confusing. What's being queued?

Yeah, as I mentioned to Christoph, a bit too much inspiration from
q_usage_count. Perhaps "managed" makes more sense.

>
> > +static const struct file_operations cdev_queued_fops = {
> > +     .owner = THIS_MODULE,
> > +     .open = cdev_queued_open,
> > +     .unlocked_ioctl = cdev_queued_ioctl,
> > +     .compat_ioctl = compat_ptr_ioctl,
> > +     .llseek = noop_llseek,
> > +};
>
> Why do we only protect these fops? I'd find it a bit confusing to have
> ioctl protected from use after del, but not write/read/etc.

More ops can certainly be added over time, I didn't want to go do the
work to wrap all file_operations before getting consensus on the idea
that the cdev core should provide managed ops at all.

The other question I'm posing with cdev_operations is whether the cdev
core should take away some of the flexibility from end drivers in
favor of adding more type safety. For example, mandate that all ioctls
take a pointer argument not an integer argument? The question of
whether wrapping cdev file_operations around a new cdev_operations is
a good idea can be deferred after finalizing a mechanism for managed
cdev file_operations.