| Message ID | 20210428215257.500088-2-hdegoede@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | drm + usb-type-c: Add support for out-of-band hotplug notification | expand |
On Wed, Apr 28, 2021 at 11:52:49PM +0200, Hans de Goede wrote: > Userspace could hold open a reference to the connector->kdev device, > through e.g. holding a sysfs-atrtribute open after > drm_sysfs_connector_remove() has been called. In this case the connector > could be free-ed while the connector->kdev device's drvdata is still > pointing to it. > > Give drm_connector devices there own device type, which allows > us to specify our own release function and make drm_sysfs_connector_add() > take a reference on the connector object, and have the new release > function put the reference when the device is released. > > Giving drm_connector devices there own device type, will also allow > checking if a device is a drm_connector device with a > "if (device->type == &drm_sysfs_device_connector)" check. > > Note that the setting of the name member of the device_type struct will > cause udev events for drm_connector-s to now contain DEVTYPE=drm_connector > as extra info. So this extends the uevent part of the userspace API. > > Signed-off-by: Hans de Goede <hdegoede@redhat.com> Are you sure? I thought sysfs is supposed to flush out any pending operations (they complete fast) and handle open fd internally? Also I'd assume this creates a loop since the connector holds a reference on the kdev? -Daniel > --- > drivers/gpu/drm/drm_sysfs.c | 54 +++++++++++++++++++++++++++++++------ > 1 file changed, 46 insertions(+), 8 deletions(-) > > diff --git a/drivers/gpu/drm/drm_sysfs.c b/drivers/gpu/drm/drm_sysfs.c > index f0336c804639..c344c6d5e738 100644 > --- a/drivers/gpu/drm/drm_sysfs.c > +++ b/drivers/gpu/drm/drm_sysfs.c > @@ -50,6 +50,10 @@ static struct device_type drm_sysfs_device_minor = { > .name = "drm_minor" > }; > > +static struct device_type drm_sysfs_device_connector = { > + .name = "drm_connector", > +}; > + > struct class *drm_class; > > static char *drm_devnode(struct device *dev, umode_t *mode) > @@ -271,30 +275,64 @@ static const struct attribute_group *connector_dev_groups[] = { > NULL > }; > > +static void drm_sysfs_connector_release(struct device *dev) > +{ > + struct drm_connector *connector = to_drm_connector(dev); > + > + drm_connector_put(connector); > + kfree(dev); > +} > + > int drm_sysfs_connector_add(struct drm_connector *connector) > { > struct drm_device *dev = connector->dev; > + struct device *kdev; > + int r; > > if (connector->kdev) > return 0; > > - connector->kdev = > - device_create_with_groups(drm_class, dev->primary->kdev, 0, > - connector, connector_dev_groups, > - "card%d-%s", dev->primary->index, > - connector->name); > + kdev = kzalloc(sizeof(*kdev), GFP_KERNEL); > + if (!kdev) > + return -ENOMEM; > + > + device_initialize(kdev); > + kdev->class = drm_class; > + kdev->type = &drm_sysfs_device_connector; > + kdev->parent = dev->primary->kdev; > + kdev->groups = connector_dev_groups; > + kdev->release = drm_sysfs_connector_release; > + dev_set_drvdata(kdev, connector); > + > + r = dev_set_name(kdev, "card%d-%s", dev->primary->index, connector->name); > + if (r) > + goto err_free; > + > DRM_DEBUG("adding \"%s\" to sysfs\n", > connector->name); > > - if (IS_ERR(connector->kdev)) { > - DRM_ERROR("failed to register connector device: %ld\n", PTR_ERR(connector->kdev)); > - return PTR_ERR(connector->kdev); > + r = device_add(kdev); > + if (r) { > + DRM_ERROR("failed to register connector device: %d\n", r); > + goto err_free; > } > > + /* > + * Ensure the connector object does not get free-ed if userspace still has > + * references open to the device through e.g. the connector sysfs-attributes. > + */ > + drm_connector_get(connector); > + > + connector->kdev = kdev; > + > if (connector->ddc) > return sysfs_create_link(&connector->kdev->kobj, > &connector->ddc->dev.kobj, "ddc"); > return 0; > + > +err_free: > + put_device(kdev); > + return r; > } > > void drm_sysfs_connector_remove(struct drm_connector *connector) > -- > 2.31.1 >
On Thu, Apr 29, 2021 at 01:40:28PM +0200, Daniel Vetter wrote: > On Wed, Apr 28, 2021 at 11:52:49PM +0200, Hans de Goede wrote: > > Userspace could hold open a reference to the connector->kdev device, > > through e.g. holding a sysfs-atrtribute open after > > drm_sysfs_connector_remove() has been called. In this case the connector > > could be free-ed while the connector->kdev device's drvdata is still > > pointing to it. > > > > Give drm_connector devices there own device type, which allows > > us to specify our own release function and make drm_sysfs_connector_add() > > take a reference on the connector object, and have the new release > > function put the reference when the device is released. > > > > Giving drm_connector devices there own device type, will also allow > > checking if a device is a drm_connector device with a > > "if (device->type == &drm_sysfs_device_connector)" check. > > > > Note that the setting of the name member of the device_type struct will > > cause udev events for drm_connector-s to now contain DEVTYPE=drm_connector > > as extra info. So this extends the uevent part of the userspace API. > > > > Signed-off-by: Hans de Goede <hdegoede@redhat.com> > > Are you sure? I thought sysfs is supposed to flush out any pending > operations (they complete fast) and handle open fd internally? Yes, it "should" :)
On Thu, Apr 29, 2021 at 01:54:46PM +0200, Greg Kroah-Hartman wrote: > On Thu, Apr 29, 2021 at 01:40:28PM +0200, Daniel Vetter wrote: > > On Wed, Apr 28, 2021 at 11:52:49PM +0200, Hans de Goede wrote: > > > Userspace could hold open a reference to the connector->kdev device, > > > through e.g. holding a sysfs-atrtribute open after > > > drm_sysfs_connector_remove() has been called. In this case the connector > > > could be free-ed while the connector->kdev device's drvdata is still > > > pointing to it. > > > > > > Give drm_connector devices there own device type, which allows > > > us to specify our own release function and make drm_sysfs_connector_add() > > > take a reference on the connector object, and have the new release > > > function put the reference when the device is released. > > > > > > Giving drm_connector devices there own device type, will also allow > > > checking if a device is a drm_connector device with a > > > "if (device->type == &drm_sysfs_device_connector)" check. > > > > > > Note that the setting of the name member of the device_type struct will > > > cause udev events for drm_connector-s to now contain DEVTYPE=drm_connector > > > as extra info. So this extends the uevent part of the userspace API. > > > > > > Signed-off-by: Hans de Goede <hdegoede@redhat.com> > > > > Are you sure? I thought sysfs is supposed to flush out any pending > > operations (they complete fast) and handle open fd internally? > > Yes, it "should" :) Thanks for confirming my vague memories :-) Hans, pls drop this one. -Daniel
Hi, On 4/29/21 1:40 PM, Daniel Vetter wrote: > On Wed, Apr 28, 2021 at 11:52:49PM +0200, Hans de Goede wrote: >> Userspace could hold open a reference to the connector->kdev device, >> through e.g. holding a sysfs-atrtribute open after >> drm_sysfs_connector_remove() has been called. In this case the connector >> could be free-ed while the connector->kdev device's drvdata is still >> pointing to it. >> >> Give drm_connector devices there own device type, which allows >> us to specify our own release function and make drm_sysfs_connector_add() >> take a reference on the connector object, and have the new release >> function put the reference when the device is released. >> >> Giving drm_connector devices there own device type, will also allow >> checking if a device is a drm_connector device with a >> "if (device->type == &drm_sysfs_device_connector)" check. >> >> Note that the setting of the name member of the device_type struct will >> cause udev events for drm_connector-s to now contain DEVTYPE=drm_connector >> as extra info. So this extends the uevent part of the userspace API. >> >> Signed-off-by: Hans de Goede <hdegoede@redhat.com> > > Are you sure? I thought sysfs is supposed to flush out any pending > operations (they complete fast) and handle open fd internally? So I did some digging in fs/kernfs and it looks like you right, once the file has been removed from sysfs any accesses through an open fd will fail with -ENODEV, interesting I did not know this. We still need this change though to make sure that the "drm/connector: Add drm_connector_find_by_fwnode() function" does not end up following a dangling drvdat pointer from one if the drm_connector kdev-s. The class_dev_iter_init() in drm_connector_find_by_fwnode() gets a reference on all devices and between getting that reference and it calling drm_connector_get() - drm_connector_unregister() may run and drop the possibly last reference to the drm_connector object, freeing it and leaving the kdev's drvdata as a dangling pointer. But I obviously need to rewrite the commit message of this commit as it currently is completely wrong. Maybe I should even squash this into the commit adding drm_connector_find_by_fwnode() ? Note sure about that though I personally think this is best kept as a new preparation patch but with a new commit msg. > Also I'd assume this creates a loop since the connector holds a reference > on the kdev? So I was wondering the same thing when working on this code and I checked. the reference on the kdev is dropped from: drm_connector_unregister() -> drm_sysfs_connector_remove() and then happens independent of the reference count on the connector-drm-obj dropping to 0. So what this change does is it keeps a reference to the drm_connector obj as long as someone is keeping a reference to the connnector->kdev device around after drm_connector_unregister() but as soon as that kdev device ref is dropped, so will the drm_connector's obj reference. I also tested this using a dock with DP MST, which dynamically adds/removes connectors on plug-in / plug-out of the dock-cable and I added a printk to the new drm_sysfs_connector_release() this adds and that printk triggered pretty much immediately on unplug as expected, releasing the extra drm_connector obj as soon as drm_connector_unregister() got called. Regards, Hans > -Daniel > >> --- >> drivers/gpu/drm/drm_sysfs.c | 54 +++++++++++++++++++++++++++++++------ >> 1 file changed, 46 insertions(+), 8 deletions(-) >> >> diff --git a/drivers/gpu/drm/drm_sysfs.c b/drivers/gpu/drm/drm_sysfs.c >> index f0336c804639..c344c6d5e738 100644 >> --- a/drivers/gpu/drm/drm_sysfs.c >> +++ b/drivers/gpu/drm/drm_sysfs.c >> @@ -50,6 +50,10 @@ static struct device_type drm_sysfs_device_minor = { >> .name = "drm_minor" >> }; >> >> +static struct device_type drm_sysfs_device_connector = { >> + .name = "drm_connector", >> +}; >> + >> struct class *drm_class; >> >> static char *drm_devnode(struct device *dev, umode_t *mode) >> @@ -271,30 +275,64 @@ static const struct attribute_group *connector_dev_groups[] = { >> NULL >> }; >> >> +static void drm_sysfs_connector_release(struct device *dev) >> +{ >> + struct drm_connector *connector = to_drm_connector(dev); >> + >> + drm_connector_put(connector); >> + kfree(dev); >> +} >> + >> int drm_sysfs_connector_add(struct drm_connector *connector) >> { >> struct drm_device *dev = connector->dev; >> + struct device *kdev; >> + int r; >> >> if (connector->kdev) >> return 0; >> >> - connector->kdev = >> - device_create_with_groups(drm_class, dev->primary->kdev, 0, >> - connector, connector_dev_groups, >> - "card%d-%s", dev->primary->index, >> - connector->name); >> + kdev = kzalloc(sizeof(*kdev), GFP_KERNEL); >> + if (!kdev) >> + return -ENOMEM; >> + >> + device_initialize(kdev); >> + kdev->class = drm_class; >> + kdev->type = &drm_sysfs_device_connector; >> + kdev->parent = dev->primary->kdev; >> + kdev->groups = connector_dev_groups; >> + kdev->release = drm_sysfs_connector_release; >> + dev_set_drvdata(kdev, connector); >> + >> + r = dev_set_name(kdev, "card%d-%s", dev->primary->index, connector->name); >> + if (r) >> + goto err_free; >> + >> DRM_DEBUG("adding \"%s\" to sysfs\n", >> connector->name); >> >> - if (IS_ERR(connector->kdev)) { >> - DRM_ERROR("failed to register connector device: %ld\n", PTR_ERR(connector->kdev)); >> - return PTR_ERR(connector->kdev); >> + r = device_add(kdev); >> + if (r) { >> + DRM_ERROR("failed to register connector device: %d\n", r); >> + goto err_free; >> } >> >> + /* >> + * Ensure the connector object does not get free-ed if userspace still has >> + * references open to the device through e.g. the connector sysfs-attributes. >> + */ >> + drm_connector_get(connector); >> + >> + connector->kdev = kdev; >> + >> if (connector->ddc) >> return sysfs_create_link(&connector->kdev->kobj, >> &connector->ddc->dev.kobj, "ddc"); >> return 0; >> + >> +err_free: >> + put_device(kdev); >> + return r; >> } >> >> void drm_sysfs_connector_remove(struct drm_connector *connector) >> -- >> 2.31.1 >> >
Hi, On 4/29/21 2:04 PM, Daniel Vetter wrote: > On Thu, Apr 29, 2021 at 01:54:46PM +0200, Greg Kroah-Hartman wrote: >> On Thu, Apr 29, 2021 at 01:40:28PM +0200, Daniel Vetter wrote: >>> On Wed, Apr 28, 2021 at 11:52:49PM +0200, Hans de Goede wrote: >>>> Userspace could hold open a reference to the connector->kdev device, >>>> through e.g. holding a sysfs-atrtribute open after >>>> drm_sysfs_connector_remove() has been called. In this case the connector >>>> could be free-ed while the connector->kdev device's drvdata is still >>>> pointing to it. >>>> >>>> Give drm_connector devices there own device type, which allows >>>> us to specify our own release function and make drm_sysfs_connector_add() >>>> take a reference on the connector object, and have the new release >>>> function put the reference when the device is released. >>>> >>>> Giving drm_connector devices there own device type, will also allow >>>> checking if a device is a drm_connector device with a >>>> "if (device->type == &drm_sysfs_device_connector)" check. >>>> >>>> Note that the setting of the name member of the device_type struct will >>>> cause udev events for drm_connector-s to now contain DEVTYPE=drm_connector >>>> as extra info. So this extends the uevent part of the userspace API. >>>> >>>> Signed-off-by: Hans de Goede <hdegoede@redhat.com> >>> >>> Are you sure? I thought sysfs is supposed to flush out any pending >>> operations (they complete fast) and handle open fd internally? >> >> Yes, it "should" :) > > Thanks for confirming my vague memories :-) > > Hans, pls drop this one. Please see my earlier reply to your review of this patch, it is still needed but for a different reason: """ We still need this change though to make sure that the "drm/connector: Add drm_connector_find_by_fwnode() function" does not end up following a dangling drvdat pointer from one if the drm_connector kdev-s. The class_dev_iter_init() in drm_connector_find_by_fwnode() gets a reference on all devices and between getting that reference and it calling drm_connector_get() - drm_connector_unregister() may run and drop the possibly last reference to the drm_connector object, freeing it and leaving the kdev's drvdata as a dangling pointer. """ This is actually why I added it initially, and while adding it I came up with this wrong theory of why it was necessary independently of the drm_connector_find_by_fwnode() addition, sorry about that. Regards, Hans
On Thu, Apr 29, 2021 at 02:33:17PM +0200, Hans de Goede wrote: > Hi, > > On 4/29/21 2:04 PM, Daniel Vetter wrote: > > On Thu, Apr 29, 2021 at 01:54:46PM +0200, Greg Kroah-Hartman wrote: > >> On Thu, Apr 29, 2021 at 01:40:28PM +0200, Daniel Vetter wrote: > >>> On Wed, Apr 28, 2021 at 11:52:49PM +0200, Hans de Goede wrote: > >>>> Userspace could hold open a reference to the connector->kdev device, > >>>> through e.g. holding a sysfs-atrtribute open after > >>>> drm_sysfs_connector_remove() has been called. In this case the connector > >>>> could be free-ed while the connector->kdev device's drvdata is still > >>>> pointing to it. > >>>> > >>>> Give drm_connector devices there own device type, which allows > >>>> us to specify our own release function and make drm_sysfs_connector_add() > >>>> take a reference on the connector object, and have the new release > >>>> function put the reference when the device is released. > >>>> > >>>> Giving drm_connector devices there own device type, will also allow > >>>> checking if a device is a drm_connector device with a > >>>> "if (device->type == &drm_sysfs_device_connector)" check. > >>>> > >>>> Note that the setting of the name member of the device_type struct will > >>>> cause udev events for drm_connector-s to now contain DEVTYPE=drm_connector > >>>> as extra info. So this extends the uevent part of the userspace API. > >>>> > >>>> Signed-off-by: Hans de Goede <hdegoede@redhat.com> > >>> > >>> Are you sure? I thought sysfs is supposed to flush out any pending > >>> operations (they complete fast) and handle open fd internally? > >> > >> Yes, it "should" :) > > > > Thanks for confirming my vague memories :-) > > > > Hans, pls drop this one. > > Please see my earlier reply to your review of this patch, it is > still needed but for a different reason: > > """ > We still need this change though to make sure that the > "drm/connector: Add drm_connector_find_by_fwnode() function" > does not end up following a dangling drvdat pointer from one > if the drm_connector kdev-s. > > The class_dev_iter_init() in drm_connector_find_by_fwnode() gets > a reference on all devices and between getting that reference > and it calling drm_connector_get() - drm_connector_unregister() > may run and drop the possibly last reference to the > drm_connector object, freeing it and leaving the kdev's > drvdata as a dangling pointer. > """ > > This is actually why I added it initially, and while adding it > I came up with this wrong theory of why it was necessary independently > of the drm_connector_find_by_fwnode() addition, sorry about that. Generally that's handled by a kref_get_unless_zero under the protection of the lock which protects the weak reference. Which I think is the right model here (at a glance at least) since this is a lookup function. Lookup tables holding full references tends to lead to all kinds of bad side effects. -Daniel
Hi, On 4/29/21 9:09 PM, Daniel Vetter wrote: > On Thu, Apr 29, 2021 at 02:33:17PM +0200, Hans de Goede wrote: >> Hi, >> >> On 4/29/21 2:04 PM, Daniel Vetter wrote: >>> On Thu, Apr 29, 2021 at 01:54:46PM +0200, Greg Kroah-Hartman wrote: >>>> On Thu, Apr 29, 2021 at 01:40:28PM +0200, Daniel Vetter wrote: >>>>> On Wed, Apr 28, 2021 at 11:52:49PM +0200, Hans de Goede wrote: >>>>>> Userspace could hold open a reference to the connector->kdev device, >>>>>> through e.g. holding a sysfs-atrtribute open after >>>>>> drm_sysfs_connector_remove() has been called. In this case the connector >>>>>> could be free-ed while the connector->kdev device's drvdata is still >>>>>> pointing to it. >>>>>> >>>>>> Give drm_connector devices there own device type, which allows >>>>>> us to specify our own release function and make drm_sysfs_connector_add() >>>>>> take a reference on the connector object, and have the new release >>>>>> function put the reference when the device is released. >>>>>> >>>>>> Giving drm_connector devices there own device type, will also allow >>>>>> checking if a device is a drm_connector device with a >>>>>> "if (device->type == &drm_sysfs_device_connector)" check. >>>>>> >>>>>> Note that the setting of the name member of the device_type struct will >>>>>> cause udev events for drm_connector-s to now contain DEVTYPE=drm_connector >>>>>> as extra info. So this extends the uevent part of the userspace API. >>>>>> >>>>>> Signed-off-by: Hans de Goede <hdegoede@redhat.com> >>>>> >>>>> Are you sure? I thought sysfs is supposed to flush out any pending >>>>> operations (they complete fast) and handle open fd internally? >>>> >>>> Yes, it "should" :) >>> >>> Thanks for confirming my vague memories :-) >>> >>> Hans, pls drop this one. >> >> Please see my earlier reply to your review of this patch, it is >> still needed but for a different reason: >> >> """ >> We still need this change though to make sure that the >> "drm/connector: Add drm_connector_find_by_fwnode() function" >> does not end up following a dangling drvdat pointer from one >> if the drm_connector kdev-s. >> >> The class_dev_iter_init() in drm_connector_find_by_fwnode() gets >> a reference on all devices and between getting that reference >> and it calling drm_connector_get() - drm_connector_unregister() >> may run and drop the possibly last reference to the >> drm_connector object, freeing it and leaving the kdev's >> drvdata as a dangling pointer. >> """ >> >> This is actually why I added it initially, and while adding it >> I came up with this wrong theory of why it was necessary independently >> of the drm_connector_find_by_fwnode() addition, sorry about that. > > Generally that's handled by a kref_get_unless_zero under the protection of > the lock which protects the weak reference. Which I think is the right > model here (at a glance at least) since this is a lookup function. I'm afraid that things are a bit more complicated here. The idea here is that we have a subsystem outside of the DRM subsystem which received a hotplug event for a drm-connector. The only info which this subsystem has is a reference on the fwnode level (either through device-tree or to platform-code instantiating software-fwnode-s + links for this). So in order to deliver the hotplug event to the connector we need to lookup the connector by fwnode. I've chosen to implement this by iterating over all drm_class devices with a dev_type of drm_connector using class_dev_iter_init() and friends. This makes sure that we either get a reference to the device, or that we skip the device if it is being deleted. But this just gives us a reference to the connector->kdev, not to the connector itself. A pointer to the connector itself is stored as drvdata inside the device, but without taking a reference as this patch does, there is no guarantee that that pointer does not point to possibly free-ed mem. We could set drvdata to 0 from drm_sysfs_connector_remove() Before calling device_unregister(connector->kdev) and then do something like this inside drm_connector_find_by_fwnode(): /* * Lock the device to ensure we either see the drvdata == NULL * set by drm_sysfs_connector_remove(); or we block the removal * from continuing until we are done with the device. */ device_lock(dev); connector = dev_get_drvdata(dev); if (connector && connector->fwnode == fwnode) { drm_connector_get(connector); found = connector; } device_unlock(dev); With the device_lock() synchronizing against the device_lock() in device_unregister(connector->kdev). So that we either see drvdata == NULL if we race with unregistering; or we get a reference on the drm_connector obj before its ref-count can drop to 0. There might be places though where we call code take the device_lock while holding a lock necessary for the drm_connector_get() , so this approach might lead to an AB BA deadlock. As such I think my original approach is better (also see below). > Lookup tables holding full references tends to lead to all kinds of bad > side effects. The proposed reference is not part of a lookup list, it is a reference from the kdev on the drm_connector object which gets dropped as soon as the kdev's refcount hits 0, which normally happens directly after drm_connector_unregister() has run. In many other places in the kernel problems like this are solved by embedding the device struct inside the containing data struct (so the drm_connector struct) and using the device_struct's refcounting for all refcounting and using the device struct's release callback as the release callback for the entire object. That is not doable here since the drm_object code has its own refcounting going on. What this patch is in essence doing is simulating having only 1 refcount, by making sure the main-object release callback does not get run until the drm_objects' refcount and the device's refcount have both reached 0 (by keeping the drm_object's refcount at a minimum of 1 as long as there are references to the device). Regards, Hans
On Fri, Apr 30, 2021 at 1:28 PM Hans de Goede <hdegoede@redhat.com> wrote: > > Hi, > > On 4/29/21 9:09 PM, Daniel Vetter wrote: > > On Thu, Apr 29, 2021 at 02:33:17PM +0200, Hans de Goede wrote: > >> Hi, > >> > >> On 4/29/21 2:04 PM, Daniel Vetter wrote: > >>> On Thu, Apr 29, 2021 at 01:54:46PM +0200, Greg Kroah-Hartman wrote: > >>>> On Thu, Apr 29, 2021 at 01:40:28PM +0200, Daniel Vetter wrote: > >>>>> On Wed, Apr 28, 2021 at 11:52:49PM +0200, Hans de Goede wrote: > >>>>>> Userspace could hold open a reference to the connector->kdev device, > >>>>>> through e.g. holding a sysfs-atrtribute open after > >>>>>> drm_sysfs_connector_remove() has been called. In this case the connector > >>>>>> could be free-ed while the connector->kdev device's drvdata is still > >>>>>> pointing to it. > >>>>>> > >>>>>> Give drm_connector devices there own device type, which allows > >>>>>> us to specify our own release function and make drm_sysfs_connector_add() > >>>>>> take a reference on the connector object, and have the new release > >>>>>> function put the reference when the device is released. > >>>>>> > >>>>>> Giving drm_connector devices there own device type, will also allow > >>>>>> checking if a device is a drm_connector device with a > >>>>>> "if (device->type == &drm_sysfs_device_connector)" check. > >>>>>> > >>>>>> Note that the setting of the name member of the device_type struct will > >>>>>> cause udev events for drm_connector-s to now contain DEVTYPE=drm_connector > >>>>>> as extra info. So this extends the uevent part of the userspace API. > >>>>>> > >>>>>> Signed-off-by: Hans de Goede <hdegoede@redhat.com> > >>>>> > >>>>> Are you sure? I thought sysfs is supposed to flush out any pending > >>>>> operations (they complete fast) and handle open fd internally? > >>>> > >>>> Yes, it "should" :) > >>> > >>> Thanks for confirming my vague memories :-) > >>> > >>> Hans, pls drop this one. > >> > >> Please see my earlier reply to your review of this patch, it is > >> still needed but for a different reason: > >> > >> """ > >> We still need this change though to make sure that the > >> "drm/connector: Add drm_connector_find_by_fwnode() function" > >> does not end up following a dangling drvdat pointer from one > >> if the drm_connector kdev-s. > >> > >> The class_dev_iter_init() in drm_connector_find_by_fwnode() gets > >> a reference on all devices and between getting that reference > >> and it calling drm_connector_get() - drm_connector_unregister() > >> may run and drop the possibly last reference to the > >> drm_connector object, freeing it and leaving the kdev's > >> drvdata as a dangling pointer. > >> """ > >> > >> This is actually why I added it initially, and while adding it > >> I came up with this wrong theory of why it was necessary independently > >> of the drm_connector_find_by_fwnode() addition, sorry about that. > > > > Generally that's handled by a kref_get_unless_zero under the protection of > > the lock which protects the weak reference. Which I think is the right > > model here (at a glance at least) since this is a lookup function. > > I'm afraid that things are a bit more complicated here. The idea here > is that we have a subsystem outside of the DRM subsystem which received > a hotplug event for a drm-connector. The only info which this subsystem > has is a reference on the fwnode level (either through device-tree or > to platform-code instantiating software-fwnode-s + links for this). > > So in order to deliver the hotplug event to the connector we need > to lookup the connector by fwnode. > > I've chosen to implement this by iterating over all drm_class > devices with a dev_type of drm_connector using class_dev_iter_init() > and friends. This makes sure that we either get a reference to > the device, or that we skip the device if it is being deleted. > > But this just gives us a reference to the connector->kdev, not > to the connector itself. A pointer to the connector itself is stored > as drvdata inside the device, but without taking a reference as > this patch does, there is no guarantee that that pointer does not > point to possibly free-ed mem. > > We could set drvdata to 0 from drm_sysfs_connector_remove() > Before calling device_unregister(connector->kdev) and then do > something like this inside drm_connector_find_by_fwnode(): > > /* > * Lock the device to ensure we either see the drvdata == NULL > * set by drm_sysfs_connector_remove(); or we block the removal > * from continuing until we are done with the device. > */ > device_lock(dev); > connector = dev_get_drvdata(dev); > if (connector && connector->fwnode == fwnode) { > drm_connector_get(connector); > found = connector; > } > device_unlock(dev); Yes this is what I mean. Except not a drm_connector_get, but a kref_get_unless_zero. The connector might already be on it's way out, but the drvdata not yet cleared. > With the device_lock() synchronizing against the device_lock() > in device_unregister(connector->kdev). So that we either see > drvdata == NULL if we race with unregistering; or we get > a reference on the drm_connector obj before its ref-count can > drop to 0. The trouble is that most connectors aren't full drivers on their kdev. So this isn't the right lock. We need another lock which protects the drvdata pointer appropriately for drm connectors. > There might be places though where we call code take the device_lock > while holding a lock necessary for the drm_connector_get() , so > this approach might lead to an AB BA deadlock. As such I think > my original approach is better (also see below). > > > Lookup tables holding full references tends to lead to all kinds of bad > > side effects. > > The proposed reference is not part of a lookup list, it is a > reference from the kdev on the drm_connector object which gets > dropped as soon as the kdev's refcount hits 0, which normally > happens directly after drm_connector_unregister() has run. Yeah but the way you use it is for lookup purposes. What we're implementing is the "get me the drm_connector for this fwnode" functionality, and that _is_ a lookup. How its implemented is an internal detail really, and somehow using full references for lookup functionality isn't great. I'm also not sure why we have to use the kdev stuff here. For other random objects we need to look up we're building that functionality on that object. It means you need to keep another list_head around for that lookup, but that's really not a big cost. E.g. drm_bridge/panel work like that. > In many other places in the kernel problems like this are > solved by embedding the device struct inside the containing > data struct (so the drm_connector struct) and using the > device_struct's refcounting for all refcounting and using > the device struct's release callback as the release callback for > the entire object. > > That is not doable here since the drm_object code has its own > refcounting going on. What this patch is in essence doing is > simulating having only 1 refcount, by making sure the > main-object release callback does not get run until > the drm_objects' refcount and the device's refcount have > both reached 0 (by keeping the drm_object's refcount at > a minimum of 1 as long as there are references to the > device). Uh yeah that sounds bad. If you really need the full refcount we should really only have one. -Daniel
Hi, On 4/30/21 1:38 PM, Daniel Vetter wrote: > On Fri, Apr 30, 2021 at 1:28 PM Hans de Goede <hdegoede@redhat.com> wrote: >> >> Hi, >> >> On 4/29/21 9:09 PM, Daniel Vetter wrote: >>> On Thu, Apr 29, 2021 at 02:33:17PM +0200, Hans de Goede wrote: >>>> Hi, >>>> >>>> On 4/29/21 2:04 PM, Daniel Vetter wrote: >>>>> On Thu, Apr 29, 2021 at 01:54:46PM +0200, Greg Kroah-Hartman wrote: >>>>>> On Thu, Apr 29, 2021 at 01:40:28PM +0200, Daniel Vetter wrote: >>>>>>> On Wed, Apr 28, 2021 at 11:52:49PM +0200, Hans de Goede wrote: >>>>>>>> Userspace could hold open a reference to the connector->kdev device, >>>>>>>> through e.g. holding a sysfs-atrtribute open after >>>>>>>> drm_sysfs_connector_remove() has been called. In this case the connector >>>>>>>> could be free-ed while the connector->kdev device's drvdata is still >>>>>>>> pointing to it. >>>>>>>> >>>>>>>> Give drm_connector devices there own device type, which allows >>>>>>>> us to specify our own release function and make drm_sysfs_connector_add() >>>>>>>> take a reference on the connector object, and have the new release >>>>>>>> function put the reference when the device is released. >>>>>>>> >>>>>>>> Giving drm_connector devices there own device type, will also allow >>>>>>>> checking if a device is a drm_connector device with a >>>>>>>> "if (device->type == &drm_sysfs_device_connector)" check. >>>>>>>> >>>>>>>> Note that the setting of the name member of the device_type struct will >>>>>>>> cause udev events for drm_connector-s to now contain DEVTYPE=drm_connector >>>>>>>> as extra info. So this extends the uevent part of the userspace API. >>>>>>>> >>>>>>>> Signed-off-by: Hans de Goede <hdegoede@redhat.com> >>>>>>> >>>>>>> Are you sure? I thought sysfs is supposed to flush out any pending >>>>>>> operations (they complete fast) and handle open fd internally? >>>>>> >>>>>> Yes, it "should" :) >>>>> >>>>> Thanks for confirming my vague memories :-) >>>>> >>>>> Hans, pls drop this one. >>>> >>>> Please see my earlier reply to your review of this patch, it is >>>> still needed but for a different reason: >>>> >>>> """ >>>> We still need this change though to make sure that the >>>> "drm/connector: Add drm_connector_find_by_fwnode() function" >>>> does not end up following a dangling drvdat pointer from one >>>> if the drm_connector kdev-s. >>>> >>>> The class_dev_iter_init() in drm_connector_find_by_fwnode() gets >>>> a reference on all devices and between getting that reference >>>> and it calling drm_connector_get() - drm_connector_unregister() >>>> may run and drop the possibly last reference to the >>>> drm_connector object, freeing it and leaving the kdev's >>>> drvdata as a dangling pointer. >>>> """ >>>> >>>> This is actually why I added it initially, and while adding it >>>> I came up with this wrong theory of why it was necessary independently >>>> of the drm_connector_find_by_fwnode() addition, sorry about that. >>> >>> Generally that's handled by a kref_get_unless_zero under the protection of >>> the lock which protects the weak reference. Which I think is the right >>> model here (at a glance at least) since this is a lookup function. >> >> I'm afraid that things are a bit more complicated here. The idea here >> is that we have a subsystem outside of the DRM subsystem which received >> a hotplug event for a drm-connector. The only info which this subsystem >> has is a reference on the fwnode level (either through device-tree or >> to platform-code instantiating software-fwnode-s + links for this). >> >> So in order to deliver the hotplug event to the connector we need >> to lookup the connector by fwnode. >> >> I've chosen to implement this by iterating over all drm_class >> devices with a dev_type of drm_connector using class_dev_iter_init() >> and friends. This makes sure that we either get a reference to >> the device, or that we skip the device if it is being deleted. >> >> But this just gives us a reference to the connector->kdev, not >> to the connector itself. A pointer to the connector itself is stored >> as drvdata inside the device, but without taking a reference as >> this patch does, there is no guarantee that that pointer does not >> point to possibly free-ed mem. >> >> We could set drvdata to 0 from drm_sysfs_connector_remove() >> Before calling device_unregister(connector->kdev) and then do >> something like this inside drm_connector_find_by_fwnode(): >> >> /* >> * Lock the device to ensure we either see the drvdata == NULL >> * set by drm_sysfs_connector_remove(); or we block the removal >> * from continuing until we are done with the device. >> */ >> device_lock(dev); >> connector = dev_get_drvdata(dev); >> if (connector && connector->fwnode == fwnode) { >> drm_connector_get(connector); >> found = connector; >> } >> device_unlock(dev); > > Yes this is what I mean. Except not a drm_connector_get, but a > kref_get_unless_zero. The connector might already be on it's way out, > but the drvdata not yet cleared. The function we race with is drm_sysfs_connector_remove() and either: 1. The lookup wins the race in which case drm_sysfs_connector_remove() can only complete after the drm_connector_get(); and the connector kref won't drop to 0 before drm_sysfs_connector_remove() completes; or 2. drm_sysfs_connector_remove() wins the race in which case drvdata will be 0. So using kref_get_unless_zero here will not make a difference and requires poking inside the drm_connector internals. Note I will probably go with your suggestion below, so whether or not to use kref_get_unless_zero here is likely no longer relevant. >> With the device_lock() synchronizing against the device_lock() >> in device_unregister(connector->kdev). So that we either see >> drvdata == NULL if we race with unregistering; or we get >> a reference on the drm_connector obj before its ref-count can >> drop to 0. > > The trouble is that most connectors aren't full drivers on their kdev. > So this isn't the right lock. We need another lock which protects the > drvdata pointer appropriately for drm connectors. > >> There might be places though where we call code take the device_lock >> while holding a lock necessary for the drm_connector_get() , so >> this approach might lead to an AB BA deadlock. As such I think >> my original approach is better (also see below). >> >>> Lookup tables holding full references tends to lead to all kinds of bad >>> side effects. >> >> The proposed reference is not part of a lookup list, it is a >> reference from the kdev on the drm_connector object which gets >> dropped as soon as the kdev's refcount hits 0, which normally >> happens directly after drm_connector_unregister() has run. > > Yeah but the way you use it is for lookup purposes. What we're > implementing is the "get me the drm_connector for this fwnode" > functionality, and that _is_ a lookup. Ack. > How its implemented is an > internal detail really, and somehow using full references for lookup > functionality isn't great. Ok, note that the caller of this only needs the reference for a short while, what the caller does is: connector = drm_connector_find_by_fwnode(dp->connector_fwnode); if (connector) { drm_connector_oob_hotplug_event(connector, &data); drm_connector_put(connector); } As a result of out discussion I have been thinking about enforcing this short-lifetime of the reference by changing: void drm_connector_oob_hotplug_event(struct drm_connector *connector, struct drm_connector_oob_hotplug_event_data *data); to: void drm_connector_oob_hotplug_event(struct fwnode_handle connector_fwnode, struct drm_connector_oob_hotplug_event_data *data); And making that do the lookup (+ almost immediate put) internally, making the connector-lookup a purely drm-subsys internal thing and enforcing code outside of the drm-subsys not holding a long-time reference to the connector this way. Please let me know if you prefer the variant where the connector lookup details are hidden from the callers ? Then I can change this for for v2 of this patch/series. > I'm also not sure why we have to use the kdev stuff here. For other > random objects we need to look up we're building that functionality on > that object. It means you need to keep another list_head around for > that lookup, but that's really not a big cost. E.g. drm_bridge/panel > work like that. Using class_for_each_dev seemed like a good way to iterate over all the connectors. But given the discussion this has caused, just adding a new static list + mutex for this to drivers/gpu/drm/drm_connector.c sounds like it might be a better approach indeed. So shall I change thing over to this approach for v2 of this patch/series? Regards, Hans
p.s. On 4/30/21 1:38 PM, Daniel Vetter wrote: Offtopic: > I'm also not sure why we have to use the kdev stuff here. For other > random objects we need to look up we're building that functionality on > that object. It means you need to keep another list_head around for > that lookup, but that's really not a big cost. E.g. drm_bridge/panel > work like that. So I took a peek at the bridge/panel code and that actually seems to have an issue with removal vs lookup. It is not even just a race, it seems a lookup does not take a reference and there is nothing stopping a user from doing an unbind or rmmod causing the panel to be removed while other code which got a pointer to the panel through of_drm_find_panel() will not be prepared to deal with that pointer all of a sudden no longer being valid. Now this would be a case of the user shooting his-self in the foot (where as connectors can actually dynamically disappear under normal circumstances), but ideally we really should do better here. Is there a TODO list somewhere for issues like this ? Or shall I submit a patch adding a FIXME comment, or is this considered not worth the trouble of fixing it? Regards, Hans
On Fri, Apr 30, 2021 at 3:32 PM Hans de Goede <hdegoede@redhat.com> wrote: > > Hi, > > On 4/30/21 1:38 PM, Daniel Vetter wrote: > > On Fri, Apr 30, 2021 at 1:28 PM Hans de Goede <hdegoede@redhat.com> wrote: > >> > >> Hi, > >> > >> On 4/29/21 9:09 PM, Daniel Vetter wrote: > >>> On Thu, Apr 29, 2021 at 02:33:17PM +0200, Hans de Goede wrote: > >>>> Hi, > >>>> > >>>> On 4/29/21 2:04 PM, Daniel Vetter wrote: > >>>>> On Thu, Apr 29, 2021 at 01:54:46PM +0200, Greg Kroah-Hartman wrote: > >>>>>> On Thu, Apr 29, 2021 at 01:40:28PM +0200, Daniel Vetter wrote: > >>>>>>> On Wed, Apr 28, 2021 at 11:52:49PM +0200, Hans de Goede wrote: > >>>>>>>> Userspace could hold open a reference to the connector->kdev device, > >>>>>>>> through e.g. holding a sysfs-atrtribute open after > >>>>>>>> drm_sysfs_connector_remove() has been called. In this case the connector > >>>>>>>> could be free-ed while the connector->kdev device's drvdata is still > >>>>>>>> pointing to it. > >>>>>>>> > >>>>>>>> Give drm_connector devices there own device type, which allows > >>>>>>>> us to specify our own release function and make drm_sysfs_connector_add() > >>>>>>>> take a reference on the connector object, and have the new release > >>>>>>>> function put the reference when the device is released. > >>>>>>>> > >>>>>>>> Giving drm_connector devices there own device type, will also allow > >>>>>>>> checking if a device is a drm_connector device with a > >>>>>>>> "if (device->type == &drm_sysfs_device_connector)" check. > >>>>>>>> > >>>>>>>> Note that the setting of the name member of the device_type struct will > >>>>>>>> cause udev events for drm_connector-s to now contain DEVTYPE=drm_connector > >>>>>>>> as extra info. So this extends the uevent part of the userspace API. > >>>>>>>> > >>>>>>>> Signed-off-by: Hans de Goede <hdegoede@redhat.com> > >>>>>>> > >>>>>>> Are you sure? I thought sysfs is supposed to flush out any pending > >>>>>>> operations (they complete fast) and handle open fd internally? > >>>>>> > >>>>>> Yes, it "should" :) > >>>>> > >>>>> Thanks for confirming my vague memories :-) > >>>>> > >>>>> Hans, pls drop this one. > >>>> > >>>> Please see my earlier reply to your review of this patch, it is > >>>> still needed but for a different reason: > >>>> > >>>> """ > >>>> We still need this change though to make sure that the > >>>> "drm/connector: Add drm_connector_find_by_fwnode() function" > >>>> does not end up following a dangling drvdat pointer from one > >>>> if the drm_connector kdev-s. > >>>> > >>>> The class_dev_iter_init() in drm_connector_find_by_fwnode() gets > >>>> a reference on all devices and between getting that reference > >>>> and it calling drm_connector_get() - drm_connector_unregister() > >>>> may run and drop the possibly last reference to the > >>>> drm_connector object, freeing it and leaving the kdev's > >>>> drvdata as a dangling pointer. > >>>> """ > >>>> > >>>> This is actually why I added it initially, and while adding it > >>>> I came up with this wrong theory of why it was necessary independently > >>>> of the drm_connector_find_by_fwnode() addition, sorry about that. > >>> > >>> Generally that's handled by a kref_get_unless_zero under the protection of > >>> the lock which protects the weak reference. Which I think is the right > >>> model here (at a glance at least) since this is a lookup function. > >> > >> I'm afraid that things are a bit more complicated here. The idea here > >> is that we have a subsystem outside of the DRM subsystem which received > >> a hotplug event for a drm-connector. The only info which this subsystem > >> has is a reference on the fwnode level (either through device-tree or > >> to platform-code instantiating software-fwnode-s + links for this). > >> > >> So in order to deliver the hotplug event to the connector we need > >> to lookup the connector by fwnode. > >> > >> I've chosen to implement this by iterating over all drm_class > >> devices with a dev_type of drm_connector using class_dev_iter_init() > >> and friends. This makes sure that we either get a reference to > >> the device, or that we skip the device if it is being deleted. > >> > >> But this just gives us a reference to the connector->kdev, not > >> to the connector itself. A pointer to the connector itself is stored > >> as drvdata inside the device, but without taking a reference as > >> this patch does, there is no guarantee that that pointer does not > >> point to possibly free-ed mem. > >> > >> We could set drvdata to 0 from drm_sysfs_connector_remove() > >> Before calling device_unregister(connector->kdev) and then do > >> something like this inside drm_connector_find_by_fwnode(): > >> > >> /* > >> * Lock the device to ensure we either see the drvdata == NULL > >> * set by drm_sysfs_connector_remove(); or we block the removal > >> * from continuing until we are done with the device. > >> */ > >> device_lock(dev); > >> connector = dev_get_drvdata(dev); > >> if (connector && connector->fwnode == fwnode) { > >> drm_connector_get(connector); > >> found = connector; > >> } > >> device_unlock(dev); > > > > Yes this is what I mean. Except not a drm_connector_get, but a > > kref_get_unless_zero. The connector might already be on it's way out, > > but the drvdata not yet cleared. > > The function we race with is drm_sysfs_connector_remove() and either: > > 1. The lookup wins the race in which case drm_sysfs_connector_remove() > can only complete after the drm_connector_get(); and the connector > kref won't drop to 0 before drm_sysfs_connector_remove() completes; or > 2. drm_sysfs_connector_remove() wins the race in which case drvdata will > be 0. > > So using kref_get_unless_zero here will not make a difference and > requires poking inside the drm_connector internals. > > Note I will probably go with your suggestion below, so whether or > not to use kref_get_unless_zero here is likely no longer relevant. Ah I missed that nuance, it's what I get for not reading the patchset :-/ I was assuming that this is a lookup function which races rather freely. The way you explain it's wired up here it's clear that we remove the lookup entry as part of hotunplug, not as part of final drm_connector cleanup. So in that special case the additional refcount due to the lookup entry isn't a problem, but it's still better to aim for something where we only deal with a single refcount for drm_connector. > >> With the device_lock() synchronizing against the device_lock() > >> in device_unregister(connector->kdev). So that we either see > >> drvdata == NULL if we race with unregistering; or we get > >> a reference on the drm_connector obj before its ref-count can > >> drop to 0. > > > > The trouble is that most connectors aren't full drivers on their kdev. > > So this isn't the right lock. We need another lock which protects the > > drvdata pointer appropriately for drm connectors. > > > >> There might be places though where we call code take the device_lock > >> while holding a lock necessary for the drm_connector_get() , so > >> this approach might lead to an AB BA deadlock. As such I think > >> my original approach is better (also see below). > >> > >>> Lookup tables holding full references tends to lead to all kinds of bad > >>> side effects. > >> > >> The proposed reference is not part of a lookup list, it is a > >> reference from the kdev on the drm_connector object which gets > >> dropped as soon as the kdev's refcount hits 0, which normally > >> happens directly after drm_connector_unregister() has run. > > > > Yeah but the way you use it is for lookup purposes. What we're > > implementing is the "get me the drm_connector for this fwnode" > > functionality, and that _is_ a lookup. > > Ack. > > > How its implemented is an > > internal detail really, and somehow using full references for lookup > > functionality isn't great. > > Ok, note that the caller of this only needs the reference for a > short while, what the caller does is: > > connector = drm_connector_find_by_fwnode(dp->connector_fwnode); > if (connector) { > drm_connector_oob_hotplug_event(connector, &data); > drm_connector_put(connector); > } > > As a result of out discussion I have been thinking about enforcing this > short-lifetime of the reference by changing: > > void drm_connector_oob_hotplug_event(struct drm_connector *connector, > struct drm_connector_oob_hotplug_event_data *data); > > to: > > void drm_connector_oob_hotplug_event(struct fwnode_handle connector_fwnode, > struct drm_connector_oob_hotplug_event_data *data); > > And making that do the lookup (+ almost immediate put) internally, making > the connector-lookup a purely drm-subsys internal thing and enforcing code > outside of the drm-subsys not holding a long-time reference to the connector > this way. > > Please let me know if you prefer the variant where the connector lookup > details are hidden from the callers ? Yeah I think that's a very nice idea. The kref_get_unless_zero is for when the lookup entry is really decoupled from the lifetime of the object, and you want to be able to get a long term access. If we can outright hide the refcounting, even better. Also this preps us for a world where v4l would also want to get these oob hotplug event, so that's a nice bonus. It's just the function name that would need to loose the drm_ prefix for that case maybe. > Then I can change this for for v2 of this patch/series. > > > I'm also not sure why we have to use the kdev stuff here. For other > > random objects we need to look up we're building that functionality on > > that object. It means you need to keep another list_head around for > > that lookup, but that's really not a big cost. E.g. drm_bridge/panel > > work like that. > > Using class_for_each_dev seemed like a good way to iterate over all > the connectors. But given the discussion this has caused, just adding > a new static list + mutex for this to drivers/gpu/drm/drm_connector.c > sounds like it might be a better approach indeed. Yeah I think kobject lifetimes are complex enough already that if there is no need to tie ourselves to the driver model lifetime rules, our own notifier list with our own locking is better. Also before Greg rolls in: This imo only applies for notifiers like we're talking about here. When we want actual references to drivers/devices and all that, then it's better to deal with all the kobject model and use it correctly. Because a lot of thought has been put into handling all the corner cases there. E.g. drm_bridge is hitting a lot of the problems which the driver model has solutions for, similarly component.c is also a bit awkward. Anyway, we still need the refcount dance because calling the oob hotplug with the lock held could deadlock with hotunplug processing and removal of the drm_connector, which needs the same lock. So it's still tricky. > So shall I change thing over to this approach for v2 of this patch/series? Yeah I think we've explored all the options and found something that fits neatly now for v2. -Daniel
diff --git a/drivers/gpu/drm/drm_sysfs.c b/drivers/gpu/drm/drm_sysfs.c index f0336c804639..c344c6d5e738 100644 --- a/drivers/gpu/drm/drm_sysfs.c +++ b/drivers/gpu/drm/drm_sysfs.c @@ -50,6 +50,10 @@ static struct device_type drm_sysfs_device_minor = { .name = "drm_minor" }; +static struct device_type drm_sysfs_device_connector = { + .name = "drm_connector", +}; + struct class *drm_class; static char *drm_devnode(struct device *dev, umode_t *mode) @@ -271,30 +275,64 @@ static const struct attribute_group *connector_dev_groups[] = { NULL }; +static void drm_sysfs_connector_release(struct device *dev) +{ + struct drm_connector *connector = to_drm_connector(dev); + + drm_connector_put(connector); + kfree(dev); +} + int drm_sysfs_connector_add(struct drm_connector *connector) { struct drm_device *dev = connector->dev; + struct device *kdev; + int r; if (connector->kdev) return 0; - connector->kdev = - device_create_with_groups(drm_class, dev->primary->kdev, 0, - connector, connector_dev_groups, - "card%d-%s", dev->primary->index, - connector->name); + kdev = kzalloc(sizeof(*kdev), GFP_KERNEL); + if (!kdev) + return -ENOMEM; + + device_initialize(kdev); + kdev->class = drm_class; + kdev->type = &drm_sysfs_device_connector; + kdev->parent = dev->primary->kdev; + kdev->groups = connector_dev_groups; + kdev->release = drm_sysfs_connector_release; + dev_set_drvdata(kdev, connector); + + r = dev_set_name(kdev, "card%d-%s", dev->primary->index, connector->name); + if (r) + goto err_free; + DRM_DEBUG("adding \"%s\" to sysfs\n", connector->name); - if (IS_ERR(connector->kdev)) { - DRM_ERROR("failed to register connector device: %ld\n", PTR_ERR(connector->kdev)); - return PTR_ERR(connector->kdev); + r = device_add(kdev); + if (r) { + DRM_ERROR("failed to register connector device: %d\n", r); + goto err_free; } + /* + * Ensure the connector object does not get free-ed if userspace still has + * references open to the device through e.g. the connector sysfs-attributes. + */ + drm_connector_get(connector); + + connector->kdev = kdev; + if (connector->ddc) return sysfs_create_link(&connector->kdev->kobj, &connector->ddc->dev.kobj, "ddc"); return 0; + +err_free: + put_device(kdev); + return r; } void drm_sysfs_connector_remove(struct drm_connector *connector)
Userspace could hold open a reference to the connector->kdev device, through e.g. holding a sysfs-atrtribute open after drm_sysfs_connector_remove() has been called. In this case the connector could be free-ed while the connector->kdev device's drvdata is still pointing to it. Give drm_connector devices there own device type, which allows us to specify our own release function and make drm_sysfs_connector_add() take a reference on the connector object, and have the new release function put the reference when the device is released. Giving drm_connector devices there own device type, will also allow checking if a device is a drm_connector device with a "if (device->type == &drm_sysfs_device_connector)" check. Note that the setting of the name member of the device_type struct will cause udev events for drm_connector-s to now contain DEVTYPE=drm_connector as extra info. So this extends the uevent part of the userspace API. Signed-off-by: Hans de Goede <hdegoede@redhat.com> --- drivers/gpu/drm/drm_sysfs.c | 54 +++++++++++++++++++++++++++++++------ 1 file changed, 46 insertions(+), 8 deletions(-)