Re: [PATCH 1/2] thunderbolt: Fix a leak in tb_retimer_add()

[Mika Westerberg]

Hi,

On Mon, Mar 29, 2021 at 10:02:20AM -0300, Jason Gunthorpe wrote:
> On Mon, Mar 29, 2021 at 09:07:18AM +0300, Dan Carpenter wrote:
> > After the device_register() succeeds, then the correct way to clean up
> > is to call device_unregister().  The unregister calls both device_del()
> > and device_put().  Since this code was only device_del() it results in
> > a memory leak.
> > 
> > Fixes: dacb12877d92 ("thunderbolt: Add support for on-board retimers")
> > Signed-off-by: Dan Carpenter <dan.carpenter@xxxxxxxxxx>
> > ---
> > This is from a new static checker warning.  Not tested.  With new
> > warnings it's also possible that I have misunderstood something
> > fundamental so review carefully etc.
> 
> It looks OK to me

I agree too.

> Reviewed-by: Jason Gunthorpe <jgg@xxxxxxxxxx>

Thanks for the review!

> This also highlights the code has an ordering issue too, it calls
> device_register() then goes to do tb_retimer_nvm_add() however
> device_register() makes sysfs attributes visible before the rt->nvm is
> initialized and this:
> 
> static ssize_t nvm_authenticate_store(struct device *dev,
> 	struct device_attribute *attr, const char *buf, size_t count)
> {
> 	if (!rt->nvm) {
> 
> Isn't strong enough to close the potential racing. The nvm should be
> setup before device_register and all the above tests in the sysfs
> deleted so we can rely on the CPU barriers built into
> device_register() for correctness.
> 
> [which is a general tip, be very suspicious if device_register() is
> being error unwound]

The nvm is a separate (physical Linux) device that gets added under this
one. It cannot be added before AFAICT.

The code you refer actually looks like this:

static ssize_t nvm_authenticate_store(struct device *dev,
 	struct device_attribute *attr, const char *buf, size_t count)
{
	...
        if (!mutex_trylock(&rt->tb->lock)) {
                ret = restart_syscall();
                goto exit_rpm;
        }

        if (!rt->nvm) {
                ret = -EAGAIN;
                goto exit_unlock;
        }


Idea here is that if the NVMem (nvm) is not yet registered the attribute is
there but we return -EAGAIN to the userspace.