On Thu, Dec 28, 2017 at 10:42:21PM +0100, Bartosz Golaszewski wrote:
> 2017-12-28 12:28 GMT+01:00 Johan Hovold <johan@xxxxxxxxxx>:
> > On Wed, Dec 27, 2017 at 03:10:38PM +0100, Bartosz Golaszewski wrote:
> >> This function can fail with -EBUSY, but we don't check its return
> >> value in at24_remove(). Bail-out of remove() if nvmem_unregister()
> >> doesn't succeed.
> >>
> >> Signed-off-by: Bartosz Golaszewski <brgl@xxxxxxxx>
> >> ---
> >> drivers/misc/eeprom/at24.c | 6 ++++--
> >> 1 file changed, 4 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/drivers/misc/eeprom/at24.c b/drivers/misc/eeprom/at24.c
> >> index e79833d62284..fb21e1c45115 100644
> >> --- a/drivers/misc/eeprom/at24.c
> >> +++ b/drivers/misc/eeprom/at24.c
> >> @@ -684,11 +684,13 @@ static int at24_probe(struct i2c_client *client, const struct i2c_device_id *id)
> >> static int at24_remove(struct i2c_client *client)
> >> {
> >> struct at24_data *at24;
> >> - int i;
> >> + int i, ret;
> >>
> >> at24 = i2c_get_clientdata(client);
> >>
> >> - nvmem_unregister(at24->nvmem);
> >> + ret = nvmem_unregister(at24->nvmem);
> >> + if (ret)
> >> + return ret;
> >
> > I don't this makes much sense as a driver cannot refuse an unbind by
> > returning an errno from remove(). The return value is simply ignored,
> > remove() will never be called again, and you'd leave everything in an
> > inconsistent state.
> >
>
> Cc: Srinivas
>
> Hi Johan,
>
> I blindly assumed that if there's a return value in remove() then
> someone cares about it. In that case all users of nvmem_unregister()
> that check the return value and bail-out of remove() on failure are
> wrong and in the (very unlikely) event that this routine fails, we
> leak all resources.
I see only one other driver that bails out on deregistration errors
(lpc18xx_eeprom.c), even if other drivers do indeed propagate errors.
> > It looks like the nvmem code grabs a reference to the owning module
> > in __nvmem_device_get() which would at least prevent a module unload
> > while another driver is using the device. And the (sysfs) userspace
> > interface should be fine as device removal is handled by the kernfs
> > code.
>
> Indeed. I believe we should remove the -EBUSY return case from
> nvmem_register() and just do what gpiolib does - scream loud
> (dev_crit()) when someone forces a module unload or otherwise
> unregisters the device if some cells are still requested. This would
> also allow us to eventually add a devres variant for nvmem_register().
I really don't like using devres for deregistration since typically
you'd need a follow-on deallocation step or you end up with a weird
asymmetric interface, but that's another story.
And again, the module unload case would not be a problem, at least when
the device is looked up from device tree, as nvmem then grabs a module
reference.
Johan
