On Tue, 14 Jan 2025, Kurt Borja wrote:
> On Tue, Jan 14, 2025 at 06:55:34PM +0200, Ilpo Järvinen wrote:
> > On Tue, 14 Jan 2025, Kurt Borja wrote:
> >
> > > In order to protect the platform_profile_handler from API consumers,
> > > allocate it in platform_profile_register() and modify it's signature
> > > accordingly.
> > >
> > > Remove the platform_profile_handler from all consumer drivers and
> > > replace them with a pointer to the class device, which is
> > > now returned from platform_profile_register().
> > >
> > > Replace *pprof with a pointer to the class device in the rest of
> > > exported symbols.
> > >
> > > Signed-off-by: Kurt Borja <kuurtb@xxxxxxxxx>
> > > ---
> > > -int platform_profile_register(struct platform_profile_handler *pprof, void *drvdata)
> > > +struct device *platform_profile_register(struct device *dev, const char *name,
> > > + void *drvdata,
> > > + const struct platform_profile_ops *ops)
> > > {
> > > + int minor;
> > > int err;
> > >
> > > - /* Sanity check the profile handler */
> > > - if (!pprof || !pprof->ops->profile_set || !pprof->ops->profile_get ||
> > > - !pprof->ops->probe) {
> > > + /* Sanity check */
> > > + if (!dev || !name || !ops || !ops->profile_get ||
> > > + !ops->profile_set || !ops->probe) {
> > > pr_err("platform_profile: handler is invalid\n");
> > > - return -EINVAL;
> > > + return ERR_PTR(-EINVAL);
> > > }
> > >
> > > - err = pprof->ops->probe(drvdata, pprof->choices);
> > > + struct platform_profile_handler *pprof __free(kfree) = kzalloc(
> > > + sizeof(*pprof), GFP_KERNEL);
> > > + if (!pprof)
> > > + return ERR_PTR(-ENOMEM);
> > > +
> > > + err = ops->probe(drvdata, pprof->choices);
> > > if (err < 0)
> > > - return err;
> > > + return ERR_PTR(err);
> > >
> > > if (bitmap_empty(pprof->choices, PLATFORM_PROFILE_LAST)) {
> > > pr_err("platform_profile: no available profiles\n");
> > > - return -EINVAL;
> > > + return ERR_PTR(-EINVAL);
> > > }
> > >
> > > guard(mutex)(&profile_lock);
> > >
> > > /* create class interface for individual handler */
> > > - pprof->minor = ida_alloc(&platform_profile_ida, GFP_KERNEL);
> > > - if (pprof->minor < 0)
> > > - return pprof->minor;
> > > + minor = ida_alloc(&platform_profile_ida, GFP_KERNEL);
> > > + if (minor < 0)
> > > + return ERR_PTR(minor);
> > >
> > > + pprof->name = name;
> > > + pprof->ops = ops;
> > > + pprof->minor = minor;
> > > pprof->class_dev.class = &platform_profile_class;
> > > - pprof->class_dev.parent = pprof->dev;
> > > + pprof->class_dev.parent = dev;
> > > dev_set_drvdata(&pprof->class_dev, drvdata);
> > > dev_set_name(&pprof->class_dev, "platform-profile-%d", pprof->minor);
> > > err = device_register(&pprof->class_dev);
> > > if (err) {
> > > - put_device(&pprof->class_dev);
> > > + put_device(&no_free_ptr(pprof)->class_dev);
> > > goto cleanup_ida;
> > > }
> > >
> > > @@ -504,20 +524,21 @@ int platform_profile_register(struct platform_profile_handler *pprof, void *drvd
> > > if (err)
> > > goto cleanup_cur;
> > >
> > > - return 0;
> > > + return &no_free_ptr(pprof)->class_dev;
> > >
> > > cleanup_cur:
> > > - device_unregister(&pprof->class_dev);
> > > + device_unregister(&no_free_ptr(pprof)->class_dev);
> >
> > I don't like how this is architected.
> >
> > IMO, no_free_ptr() should not appear on error/rollback paths. The pointer
> > is going to be freed despite the code just told it's not going to be
> > freed, which sends conflicting signals. Obviously, it is because this
> > function has relinquished its ownership of the pointer but as is it seems
> > a dangerous/confusing pattern.
>
> Makes sense.
>
> Quick fix would be to replace `goto cleanup_cur` with
>
> device_unregister(&no_free_ptr(pprof)->class_dev);
> goto cleanup_ida;
>
> and add a comment about ownership. Similar to the put_device() call
> above. Is this ok? If not I will think of a better way of writing this.
I think it would still be on the error path which is undesirable. While a
comment would make it understandable, it would be more logical to call
no_free_ptr() near device_register() which is when the ownership
gets transferred.
The trouble with that approach then is that no_free_ptr(pprof) will set
the pprof to NULL because of how the internal cleanup magic prevents
automatic freeing of pprof (Don't ask me how I know about that trap :-D).
I suppose you could take pointer of the pprof->class_dev into a local
variable before making the device_register() call since that is all you
need after that point?
So my suggestion is along the lines of:
/* device_register() takes the ownership of the pointer */
class_dev = &no_free_ptr(pprof)->class_dev;
err = device_register(class_dev);
...
> > > cleanup_ida:
> > > - ida_free(&platform_profile_ida, pprof->minor);
> > > + ida_free(&platform_profile_ida, minor);
> > >
> > > - return err;
> > > + return ERR_PTR(err);
> > > }
> > > EXPORT_SYMBOL_GPL(platform_profile_register);
--
i.
