[RFC PATCH 0/4] drivers/base: Generic framework for tracking internal interfaces
On 05/01/2014 11:11 AM, Russell King - ARM Linux wrote:
> On Thu, May 01, 2014 at 09:04:19AM +0200, Andrzej Hajda wrote:
>> 2. You replace calls of component_add and component_del with calls
>> to interface_tracker_ifup(dev, INTERFACE_TRACKER_TYPE_COMPONENT,
>> &specific_component_ops),
>> or interface_tracker_ifdown.
>> Thats all for components.
> How does the master get to decide which components are for it?
This is not a part of the framework. The master can construct
the list of its components anyhow. Some examples:
1. List of device tree video interface port nodes pointed by
master's remote-endpoint phandles.
2. List of device nodes pointed by supernode phandles.
3. Nodes pointed by other phandles in master's node.
4. Devices compatible with the list of drivers.
This is for create list of objects. As interface type it should
use the types of interface it expects at given nodes and is able to handle.
Small issue:
If the master creates list of devices and for particular interface
type expects DT node as the object, translation is easy: dev->of_node.
But if the situation is reverse kernel does not provide generic helper
for translating of_node to device, however kernel have everything to
provide such function if necessary.
Other solution is to use only DT nodes for object identification,
it will narrow the framework usage to DT architectures, but seems to be
more flexible anyway - we can have more than one interface of given type per
device, we can distinguish them by port node.
> As
> I see it, all masters see all components of a particular "type".
> What if you have a system with two masters each of which are bound
> to a set of unique components but some of the components are of a
> the same type?
The master receives notifications only about interfaces he has
registered callback for. For example:
interface_tracker_register(node, INTERFACE_TRACKER_TYPE_DRM_PANEL, cb);
means that 'cb' callback will be called only if panel identifed by node
is up or down.
If the driver expect that at the 'node' there could be also component it
can also
listen for components:
interface_tracker_register(node, INTERFACE_TRACKER_TYPE_COMPONENT, cb);
Now 'cb' will be called if component or panel appears/disappears at node
'node'.
so callback function can look like:
void cb_func(struct interface_tracker_block *itb, const void *object,
unsigned long type, bool on,
void *data)
{
struct priv_struct *priv = container_of(itb, struct priv_struct, itb);
switch(type) {
case INTERFACE_TRACKER_TYPE_DRM_PANEL:
handle_drm_panel(priv, data, on);
break;
case INTERFACE_TRACKER_TYPE_DRM_COMPONENT:
handle_drm_component(priv, data, object, on);
break;
}
}
where handlers can look like:
void handle_drm_panel(struct priv_struct *priv, struct drm_panel *panel,
bool on);
void handle_drm_component(struct priv_struct *priv, struct component_ops
*ops, struct device *dev, bool on);
>
> How does the master know what "type" to use?
>
It should use type which it expects at the given node.
Regards
Andrzej
[RFC PATCH 0/4] drivers/base: Generic framework for tracking internal interfaces
On Thu, May 01, 2014 at 09:04:19AM +0200, Andrzej Hajda wrote: > 2. You replace calls of component_add and component_del with calls > to interface_tracker_ifup(dev, INTERFACE_TRACKER_TYPE_COMPONENT, > &specific_component_ops), > or interface_tracker_ifdown. > Thats all for components. How does the master get to decide which components are for it? As I see it, all masters see all components of a particular "type". What if you have a system with two masters each of which are bound to a set of unique components but some of the components are of a the same type? How does the master know what "type" to use? -- FTTC broadband for 0.8mile line: now at 9.7Mbps down 460kbps up... slowly improving, and getting towards what was expected from it.
[RFC PATCH 0/4] drivers/base: Generic framework for tracking internal interfaces
Russell King - ARM Linux wrote, On 01.05.2014 00:28: > On Wed, Apr 30, 2014 at 11:42:09PM +0200, Andrzej Hajda wrote: >> The main problem with component framework is that componentization >> significantly changes every driver and changes it in a way which is not >> compatible with traditional drivers, so devices which are intended to >> work with different DRM masters are hard to componentize if some of DRMs >> are componentized and some not. > Many of the problems which the component helpers are designed to solve > are those where you need the drm_device structure (or snd_card, or whatever > subsystem specific card/device representation structure) pre-created in > order to initialise the components. > > In the case of DRM, you can't initialise encoders or connectors without > their drm_device structure pre-existing - because these components are > attached to the drm_device. > > Your solution to that is to delay those calls, but the DRM subsystem is > not designed to cope like that - it's designed such that when the > connector or encoder initialisation functions are called, it is assumed > that the driver is initialising its state. (I've raised this point before > but you've just fobbed it off in the past.) > > Another issue here is that the order of initialisation matters greatly. > Take CRTCs for example. In DRM, the order of attachment of CRTCs defines > their identity, changing the order changes their identity, and changes > how they are bound to their respective connectors. > The two problems you show here are not a real problems in this framework: 1. making real device initialization during drm initialization - decision is left to driver developer what should be done in probe, what should be done in 'bind', I guess this is also true for components, at least the framework allows it. 2. initialization order - if you put initialization into components 'bind' function, master can choose any order of calls to 'bind'. Anyway you can implement the same behaviour as components with interface_tracker. Just simple proof of concept, how to convert componentized drivers to interface_tracker: Components: 1. you can reuse component_ops 2. You replace calls of component_add and component_del with calls to interface_tracker_ifup(dev, INTERFACE_TRACKER_TYPE_COMPONENT, &specific_component_ops), or interface_tracker_ifdown. Thats all for components. Master: 1. you register callback for tracking all components. 2. in the callback you check if all components are up, if yes you do the same as in component framework initialization, to simplify it helper function can be added. I guess it should work the same way, if there is interest in it I can develop the helper next week, I hope. What is the benefit of interface_tracker: 1. interface_tracker is more generic - it can track not only components. 2. you put component initialization code into helper function - sounds like mid-layer removal, developer can choose different helper if it suits better. So from component point of view interface_tracker can be treated as kind of extensions of the component framework. I hope I have answerer all your concerns. I have holidays till Sunday and I am not sure if I will be able to answer next emails before Monday. Regards Andrzej
[RFC PATCH 0/4] drivers/base: Generic framework for tracking internal interfaces
Hi Greg, Thanks for comments. I CC Laurent, I hope it could be interesting for him also. Greg Kroah-Hartman wrote, On 30.04.2014 17:49: > On Wed, Apr 30, 2014 at 04:02:50PM +0200, Andrzej Hajda wrote: >> Generic framework for tracking internal interfaces >> == >> >> Summary >> --- >> >> interface_tracker is a generic framework which allows to track appearance >> and disappearance of different interfaces provided by kernel/driver code >> inside >> the kernel. Examples of such interfaces: clocks, phys, regulators, >> drm_panel... >> Interface is specified by a pair of object pointer and interface type. Object >> type depends on interface type, for example interface type drm_panel >> determines >> that object is a device_node. Object pointer is used to distinguish different >> interfaces of the same type and should identify object the interface is >> bound to. >> For example it could be DT node, struct device,... >> >> The framework provides two functions for interface providers which should be >> called just after interface becomes available or just before interface >> removal. Interface consumers can register callback which will be called >> when requested interface changes its state, if during callback registration >> the interface is already up, notification will be sent also. >> >> The framework allows nesting calls, for example it is possible to signal one >> interface in tracker callback of another interface. It is done by putting >> every request into the queue. If the queue is empty before adding >> the request, the queue will be processed after, if there is already another >> request in the queue it means the queue is already processed by different >> thread so no additional action is required. With this pattern only spinlock >> is necessary to protect the queue. However in case of removal of either >> callback or interface caller should be sure that his request has been >> processed so framework waits until the queue becomes empty, it is done using >> completion mechanism. >> >> The framework functions should not be called in atomic context. Callbacks >> should be aware that they can be called in any time between registration and >> de-registration, so they should use synchronization mechanisms carefully. >> Callbacks should also avoid to perform time consuming tasks, the framework >> serializes them, so it could stall other callbacks. >> >> Use cases >> - >> >> The interface is very generic, there are many situations it can be useful: >> >> 1. Replacement for subsystem specific methods of publishing/unpublishing >> interfaces. Currently many frameworks allows only querying for presence of >> given >> interface. In such cases client can defer probing or implement interface >> polling, which is usually subobtimal. Additionally subsystems often do not >> provide methods to signal to the consumers that they are about to be >> destroyed. >> >> 2. Monitoring for different interfaces provided by the same object. An >> example >> should explain it better. Lets assume in device tree drm crtc device node >> have >> video link to another node, so it knows only that there is something >> connected >> to its RGB output. It can be a RGB panel (drm_panel framework), it can be an >> image enhancer (SoC specific framework) or it can be some signal converter >> (drm_encoder, drm_bridge, drm_encoder_slave...). Driver have only phandle to >> another node. Currently it is difficult to handle such situations in a >> generic >> way. interface_tracker should make it simple: crtc should monitor all >> supported >> interface types that appears at the device_node pointed by the phandle. >> >> Potential use cases >> --- >> >> Points mentioned above were the reasons for writing this framework. During >> development I have realized that this framework can be also useful for other >> tasks. >> >> 3. Replacement for deferred probing - if for some reason driver do not wants >> to >> defer but it requires given resources it can use interface_tracker. It >> should be >> possible to create an helper which will wait for appearance of all interfaces >> from a given list, and 'continue' probe only when all resources becomes >> available. >> >> 4. Replacement or helper for subsystem specific solutions: >> - V4L2 subdev async registration, >> - component framework. >> Both frameworks solves a problem of tracking sub-components (un-)registration >> by master device, it should be possible to do the same with interface_tracker >> framework. Some additional helpers can be convienent to make the >> implementation >> easier. >> >> 5. Cure the situation with sysfs 'unbind' property. Currently many drivers >> are >> providers of different resources/interfaces: regulators, clocks, phys, >> V4L2 subdevs, ... They are usually protected from module unloading by getting >> module reference, but there is no protection from driver unbinding using >> sysfs >>
[RFC PATCH 0/4] drivers/base: Generic framework for tracking internal interfaces
On Wed, Apr 30, 2014 at 11:42:09PM +0200, Andrzej Hajda wrote: > The main problem with component framework is that componentization > significantly changes every driver and changes it in a way which is not > compatible with traditional drivers, so devices which are intended to > work with different DRM masters are hard to componentize if some of DRMs > are componentized and some not. Many of the problems which the component helpers are designed to solve are those where you need the drm_device structure (or snd_card, or whatever subsystem specific card/device representation structure) pre-created in order to initialise the components. In the case of DRM, you can't initialise encoders or connectors without their drm_device structure pre-existing - because these components are attached to the drm_device. Your solution to that is to delay those calls, but the DRM subsystem is not designed to cope like that - it's designed such that when the connector or encoder initialisation functions are called, it is assumed that the driver is initialising its state. (I've raised this point before but you've just fobbed it off in the past.) Another issue here is that the order of initialisation matters greatly. Take CRTCs for example. In DRM, the order of attachment of CRTCs defines their identity, changing the order changes their identity, and changes how they are bound to their respective connectors. -- FTTC broadband for 0.8mile line: now at 9.7Mbps down 460kbps up... slowly improving, and getting towards what was expected from it.
[RFC PATCH 0/4] drivers/base: Generic framework for tracking internal interfaces
Generic framework for tracking internal interfaces == Summary --- interface_tracker is a generic framework which allows to track appearance and disappearance of different interfaces provided by kernel/driver code inside the kernel. Examples of such interfaces: clocks, phys, regulators, drm_panel... Interface is specified by a pair of object pointer and interface type. Object type depends on interface type, for example interface type drm_panel determines that object is a device_node. Object pointer is used to distinguish different interfaces of the same type and should identify object the interface is bound to. For example it could be DT node, struct device,... The framework provides two functions for interface providers which should be called just after interface becomes available or just before interface removal. Interface consumers can register callback which will be called when requested interface changes its state, if during callback registration the interface is already up, notification will be sent also. The framework allows nesting calls, for example it is possible to signal one interface in tracker callback of another interface. It is done by putting every request into the queue. If the queue is empty before adding the request, the queue will be processed after, if there is already another request in the queue it means the queue is already processed by different thread so no additional action is required. With this pattern only spinlock is necessary to protect the queue. However in case of removal of either callback or interface caller should be sure that his request has been processed so framework waits until the queue becomes empty, it is done using completion mechanism. The framework functions should not be called in atomic context. Callbacks should be aware that they can be called in any time between registration and de-registration, so they should use synchronization mechanisms carefully. Callbacks should also avoid to perform time consuming tasks, the framework serializes them, so it could stall other callbacks. Use cases - The interface is very generic, there are many situations it can be useful: 1. Replacement for subsystem specific methods of publishing/unpublishing interfaces. Currently many frameworks allows only querying for presence of given interface. In such cases client can defer probing or implement interface polling, which is usually subobtimal. Additionally subsystems often do not provide methods to signal to the consumers that they are about to be destroyed. 2. Monitoring for different interfaces provided by the same object. An example should explain it better. Lets assume in device tree drm crtc device node have video link to another node, so it knows only that there is something connected to its RGB output. It can be a RGB panel (drm_panel framework), it can be an image enhancer (SoC specific framework) or it can be some signal converter (drm_encoder, drm_bridge, drm_encoder_slave...). Driver have only phandle to another node. Currently it is difficult to handle such situations in a generic way. interface_tracker should make it simple: crtc should monitor all supported interface types that appears at the device_node pointed by the phandle. Potential use cases --- Points mentioned above were the reasons for writing this framework. During development I have realized that this framework can be also useful for other tasks. 3. Replacement for deferred probing - if for some reason driver do not wants to defer but it requires given resources it can use interface_tracker. It should be possible to create an helper which will wait for appearance of all interfaces from a given list, and 'continue' probe only when all resources becomes available. 4. Replacement or helper for subsystem specific solutions: - V4L2 subdev async registration, - component framework. Both frameworks solves a problem of tracking sub-components (un-)registration by master device, it should be possible to do the same with interface_tracker framework. Some additional helpers can be convienent to make the implementation easier. 5. Cure the situation with sysfs 'unbind' property. Currently many drivers are providers of different resources/interfaces: regulators, clocks, phys, V4L2 subdevs, ... They are usually protected from module unloading by getting module reference, but there is no protection from driver unbinding using sysfs method: echo 'device' >/sys/bus/.../drivers/.../unbind. After unbind consumer stays with the pointer to non-existing object, next time it tries to use it system usually crashes. interface_tracker do not add any protection, but it adds a way to signal to the consumer that given resource will disappear. It allows to handle such situations more gently. Potential issues/extensions --- 1. Scalability - the framework serializes all tasks and callbacks. In case there are no many users
[RFC PATCH 0/4] drivers/base: Generic framework for tracking internal interfaces
On Wed, Apr 30, 2014 at 04:02:50PM +0200, Andrzej Hajda wrote: > Generic framework for tracking internal interfaces > == > > Summary > --- > > interface_tracker is a generic framework which allows to track appearance > and disappearance of different interfaces provided by kernel/driver code > inside > the kernel. Examples of such interfaces: clocks, phys, regulators, > drm_panel... > Interface is specified by a pair of object pointer and interface type. Object > type depends on interface type, for example interface type drm_panel > determines > that object is a device_node. Object pointer is used to distinguish different > interfaces of the same type and should identify object the interface is bound > to. > For example it could be DT node, struct device,... > > The framework provides two functions for interface providers which should be > called just after interface becomes available or just before interface > removal. Interface consumers can register callback which will be called > when requested interface changes its state, if during callback registration > the interface is already up, notification will be sent also. > > The framework allows nesting calls, for example it is possible to signal one > interface in tracker callback of another interface. It is done by putting > every request into the queue. If the queue is empty before adding > the request, the queue will be processed after, if there is already another > request in the queue it means the queue is already processed by different > thread so no additional action is required. With this pattern only spinlock > is necessary to protect the queue. However in case of removal of either > callback or interface caller should be sure that his request has been > processed so framework waits until the queue becomes empty, it is done using > completion mechanism. > > The framework functions should not be called in atomic context. Callbacks > should be aware that they can be called in any time between registration and > de-registration, so they should use synchronization mechanisms carefully. > Callbacks should also avoid to perform time consuming tasks, the framework > serializes them, so it could stall other callbacks. > > Use cases > - > > The interface is very generic, there are many situations it can be useful: > > 1. Replacement for subsystem specific methods of publishing/unpublishing > interfaces. Currently many frameworks allows only querying for presence of > given > interface. In such cases client can defer probing or implement interface > polling, which is usually subobtimal. Additionally subsystems often do not > provide methods to signal to the consumers that they are about to be > destroyed. > > 2. Monitoring for different interfaces provided by the same object. An example > should explain it better. Lets assume in device tree drm crtc device node have > video link to another node, so it knows only that there is something connected > to its RGB output. It can be a RGB panel (drm_panel framework), it can be an > image enhancer (SoC specific framework) or it can be some signal converter > (drm_encoder, drm_bridge, drm_encoder_slave...). Driver have only phandle to > another node. Currently it is difficult to handle such situations in a generic > way. interface_tracker should make it simple: crtc should monitor all > supported > interface types that appears at the device_node pointed by the phandle. > > Potential use cases > --- > > Points mentioned above were the reasons for writing this framework. During > development I have realized that this framework can be also useful for other > tasks. > > 3. Replacement for deferred probing - if for some reason driver do not wants > to > defer but it requires given resources it can use interface_tracker. It should > be > possible to create an helper which will wait for appearance of all interfaces > from a given list, and 'continue' probe only when all resources becomes > available. > > 4. Replacement or helper for subsystem specific solutions: > - V4L2 subdev async registration, > - component framework. > Both frameworks solves a problem of tracking sub-components (un-)registration > by master device, it should be possible to do the same with interface_tracker > framework. Some additional helpers can be convienent to make the > implementation > easier. > > 5. Cure the situation with sysfs 'unbind' property. Currently many drivers are > providers of different resources/interfaces: regulators, clocks, phys, > V4L2 subdevs, ... They are usually protected from module unloading by getting > module reference, but there is no protection from driver unbinding using sysfs > method: echo 'device' >/sys/bus/.../drivers/.../unbind. After unbind consumer > stays with the pointer to non-existing object, next time it tries to use it > system usually crashes. interface_tracker do not add any protection, but it > adds > a
