On Fri, Feb 14, 2020 at 03:43:41PM -0500, Greg KH wrote:
> On Fri, Feb 14, 2020 at 04:34:55PM -0400, Jason Gunthorpe wrote:
> > On Fri, Feb 14, 2020 at 09:02:40AM -0800, Greg KH wrote:
> > > > +/**
> > > > + * virtbus_dev_register - add a virtual bus device
> > > > + * @vdev: virtual bus device to add
> > > > + */
> > > > +int virtbus_dev_register(struct virtbus_device *vdev)
> > > > +{
> > > > + int ret;
> > > > +
> > > > + if (!vdev->release) {
> > > > + dev_err(&vdev->dev, "virtbus_device .release callback NULL\n");
> > >
> > > "virtbus_device MUST have a .release callback that does something!\n"
> > >
> > > > + return -EINVAL;
> > > > + }
> > > > +
> > > > + device_initialize(&vdev->dev);
> > > > +
> > > > + vdev->dev.bus = &virtual_bus_type;
> > > > + vdev->dev.release = virtbus_dev_release;
> > > > + /* All device IDs are automatically allocated */
> > > > + ret = ida_simple_get(&virtbus_dev_ida, 0, 0, GFP_KERNEL);
> > > > + if (ret < 0) {
> > > > + dev_err(&vdev->dev, "get IDA idx for virtbus device failed!\n");
> > > > + put_device(&vdev->dev);
> > >
> > > If you allocate the number before device_initialize(), no need to call
> > > put_device(). Just a minor thing, no big deal.
> >
> > If *_regster does put_device on error then it must always do
> > put_device on any error, for instance the above return -EINVAL with
> > no put_device leaks memory.
>
> That's why I said to move the ida_simple_get() call to before
> device_initialize() is called. Once device_initialize() is called, you
> HAVE to call put_device().
Yes put_device() becomes mandatory, but if the ida is moved up then
the caller doesn't know how to handle an error:
if (ida_simple_get() < 0)
return -EINVAL; // caller must do kfree
device_initialize();
if (device_register())
return -EINVAL // caller must do put_device
If the device_initialize is bundled in the function the best answer is
to always do device_initialize() and never do put_device(). The caller
must realize the unwind switches from kfree to put_device (tricky and
uglyifies the goto unwind!).
This is the pattern something like platform_device_register() uses,
and with a random survey I found only __ipmi_bmc_register() getting it
right. Even then it seems to have a bug related to bmc_reg_mutex due
to the ugly split goto unwind..
I prefer to see device_initialize done shortly after allocation, that
seems to be the most likely to end up correct..
Jason
