Re: [tpmdd-devel] [RFC PATCH 1/2] tee: generic TEE subsystem

From: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
To: Jens Wiklander <jens.wiklander@linaro.org>
Cc: Russell King - ARM Linux <linux@arm.linux.org.uk>,
	valentin.manea@huawei.com, devicetree@vger.kernel.org,
	javier@javigon.com, emmanuel.michel@st.com,
	Herbert Xu <herbert@gondor.apana.org.au>,
	Arnd Bergmann <arnd@arndb.de>,
	Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
	linux-kernel@vger.kernel.org, jean-michel.delorme@st.com,
	tpmdd-devel@lists.sourceforge.net,
	linux-arm-kernel@lists.infradead.org
Subject: Re: [tpmdd-devel] [RFC PATCH 1/2] tee: generic TEE subsystem
Date: Mon, 20 Apr 2015 11:55:15 -0600	[thread overview]
Message-ID: <20150420175515.GA31958@obsidianresearch.com> (raw)
In-Reply-To: <20150420130203.GB25054@ermac>

On Mon, Apr 20, 2015 at 03:02:03PM +0200, Jens Wiklander wrote:
> > It appeared to me this driver was copying TPM's old architecture,
> > which is very much known to be broken.
> 
> The struct tee_device holds a shared memory pool from which shared
> memory objects are allocated. These shared memory objects can be mapped
> both by user space and secure world. 

So this is a whole other set of problems besides what was already
brought up.

You need to figure out a lifetime model for this shared memory that
works.

> To come around the problem with what should happen when the driver
> is removed I'm increasing the refcount on the driver for each
> allocated shared memory object and created file pointers. As long as
> any resource is in use by either user space or secure world the
> driver can't be unloaded.

This isn't how the kernel works. The module refcount effects module
unload (it protects the .text) - it does not interact with driver
detatch. Userspace can trigger driver detatch (which results in
tee_unregister being called) at any time via sysfs.

If you properly design for that case then module unload sequencing
works properly for free.

Based on what I gather, I would suggest the following sequence in
tee_unregister
 - unregister all sysfs and char dev registrations.
 - Write lock ops and set to null. This will error future cdev ioctls,
   and guarentees no driver ops callbacks are in progress, or will be
   started in future.
 - Wait until all client accesses to shared memory are
   released.
 - Command the driver to release it's side of the
   shared memory and wait for that to complete
 - Free the shared memory
 - deref the tee_device's struct device (match ref in tee_register)

Then in your struct tee_device's release function free the tee_device
memory.

Replace all the module locking code with an active count in struct
tee_device (see something like kernfs_drain for an example).

> * Change to use the pattern (with a struct device etc) as described
>   above.

Yes, I think Greg confirmed you need to use a struct device, and purge
misc_device from the mid layer.

>   I can't protect the ops with just a mutex since tee_ioctl_cmd() needs to
>   be multithreaded.

Then use a sleeping read/write lock - aka an active count.

Jason