Driver Model ============ This README contains high-level information about driver model, a unified way of declaring and accessing drivers in U-Boot. The original work was done by: Marek Vasut Pavel Herrmann Viktor Křivák Tomas Hlavacek This has been both simplified and extended into the current implementation by: Simon Glass Terminology ----------- Uclass - a group of devices which operate in the same way. A uclass provides a way of accessing individual devices within the group, but always using the same interface. For example a GPIO uclass provides operations for get/set value. An I2C uclass may have 10 I2C ports, 4 with one driver, and 6 with another. Driver - some code which talks to a peripheral and presents a higher-level interface to it. Device - an instance of a driver, tied to a particular port or peripheral. How to try it ------------- Build U-Boot sandbox and run it: make sandbox_config make ./u-boot (type 'reset' to exit U-Boot) There is a uclass called 'demo'. This uclass handles saying hello, and reporting its status. There are two drivers in this uclass: - simple: Just prints a message for hello, doesn't implement status - shape: Prints shapes and reports number of characters printed as status The demo class is pretty simple, but not trivial. The intention is that it can be used for testing, so it will implement all driver model features and provide good code coverage of them. It does have multiple drivers, it handles parameter data and platdata (data which tells the driver how to operate on a particular platform) and it uses private driver data. To try it, see the example session below: =>demo hello 1 Hello '@' from 07981110: red 4 =>demo status 2 Status: 0 =>demo hello 2 g r@ e@@ e@@@ n@@@@ g@@@@@ =>demo status 2 Status: 21 =>demo hello 4 ^ y^^^ e^^^^^ l^^^^^^^ l^^^^^^^ o^^^^^ w^^^ =>demo status 4 Status: 36 => Running the tests ----------------- The intent with driver model is that the core portion has 100% test coverage in sandbox, and every uclass has its own test. As a move towards this, tests are provided in test/dm. To run them, try: ./test/dm/test-dm.sh You should see something like this: <...U-Boot banner...> Running 12 driver model tests Test: dm_test_autobind Test: dm_test_autoprobe Test: dm_test_children Test: dm_test_fdt Test: dm_test_gpio sandbox_gpio: sb_gpio_get_value: error: offset 4 not reserved Test: dm_test_leak Warning: Please add '#define DEBUG' to the top of common/dlmalloc.c Warning: Please add '#define DEBUG' to the top of common/dlmalloc.c Test: dm_test_lifecycle Test: dm_test_operations Test: dm_test_ordering Test: dm_test_platdata Test: dm_test_remove Test: dm_test_uclass Failures: 0 (You can add '#define DEBUG' as suggested to check for memory leaks) What is going on? ----------------- Let's start at the top. The demo command is in common/cmd_demo.c. It does the usual command processing and then: struct udevice *demo_dev; ret = uclass_get_device(UCLASS_DEMO, devnum, &demo_dev); UCLASS_DEMO means the class of devices which implement 'demo'. Other classes might be MMC, or GPIO, hashing or serial. The idea is that the devices in the class all share a particular way of working. The class presents a unified view of all these devices to U-Boot. This function looks up a device for the demo uclass. Given a device number we can find the device because all devices have registered with the UCLASS_DEMO uclass. The device is automatically activated ready for use by uclass_get_device(). Now that we have the device we can do things like: return demo_hello(demo_dev, ch); This function is in the demo uclass. It takes care of calling the 'hello' method of the relevant driver. Bearing in mind that there are two drivers, this particular device may use one or other of them. The code for demo_hello() is in drivers/demo/demo-uclass.c: int demo_hello(struct udevice *dev, int ch) { const struct demo_ops *ops = device_get_ops(dev); if (!ops->hello) return -ENOSYS; return ops->hello(dev, ch); } As you can see it just calls the relevant driver method. One of these is in drivers/demo/demo-simple.c: static int simple_hello(struct udevice *dev, int ch) { const struct dm_demo_pdata *pdata = dev_get_platdata(dev); printf("Hello from %08x: %s %d\n", map_to_sysmem(dev), pdata->colour, pdata->sides); return 0; } So that is a trip from top (command execution) to bottom (driver action) but it leaves a lot of topics to address. Declaring Drivers ----------------- A driver declaration looks something like this (see drivers/demo/demo-shape.c): static const struct demo_ops shape_ops = { .hello = shape_hello, .status = shape_status, }; U_BOOT_DRIVER(demo_shape_drv) = { .name = "demo_shape_drv", .id = UCLASS_DEMO, .ops = &shape_ops, .priv_data_size = sizeof(struct shape_data), }; This driver has two methods (hello and status) and requires a bit of private data (accessible through dev_get_priv(dev) once the driver has been probed). It is a member of UCLASS_DEMO so will register itself there. In U_BOOT_DRIVER it is also possible to specify special methods for bind and unbind, and these are called at appropriate times. For many drivers it is hoped that only 'probe' and 'remove' will be needed. The U_BOOT_DRIVER macro creates a data structure accessible from C, so driver model can find the drivers that are available. The methods a device can provide are documented in the device.h header. Briefly, they are: bind - make the driver model aware of a device (bind it to its driver) unbind - make the driver model forget the device ofdata_to_platdata - convert device tree data to platdata - see later probe - make a device ready for use remove - remove a device so it cannot be used until probed again The sequence to get a device to work is bind, ofdata_to_platdata (if using device tree) and probe. Platform Data ------------- Platform data is like Linux platform data, if you are familiar with that. It provides the board-specific information to start up a device. Why is this information not just stored in the device driver itself? The idea is that the device driver is generic, and can in principle operate on any board that has that type of device. For example, with modern highly-complex SoCs it is common for the IP to come from an IP vendor, and therefore (for example) the MMC controller may be the same on chips from different vendors. It makes no sense to write independent drivers for the MMC controller on each vendor's SoC, when they are all almost the same. Similarly, we may have 6 UARTs in an SoC, all of which are mostly the same, but lie at different addresses in the address space. Using the UART example, we have a single driver and it is instantiated 6 times by supplying 6 lots of platform data. Each lot of platform data gives the driver name and a pointer to a structure containing information about this instance - e.g. the address of the register space. It may be that one of the UARTS supports RS-485 operation - this can be added as a flag in the platform data, which is set for this one port and clear for the rest. Think of your driver as a generic piece of code which knows how to talk to a device, but needs to know where it is, any variant/option information and so on. Platform data provides this link between the generic piece of code and the specific way it is bound on a particular board. Examples of platform data include: - The base address of the IP block's register space - Configuration options, like: - the SPI polarity and maximum speed for a SPI controller - the I2C speed to use for an I2C device - the number of GPIOs available in a GPIO device Where does the platform data come from? It is either held in a structure which is compiled into U-Boot, or it can be parsed from the Device Tree (see 'Device Tree' below). For an example of how it can be compiled in, see demo-pdata.c which sets up a table of driver names and their associated platform data. The data can be interpreted by the drivers however they like - it is basically a communication scheme between the board-specific code and the generic drivers, which are intended to work on any board. Drivers can access their data via dev->info->platdata. Here is the declaration for the platform data, which would normally appear in the board file. static const struct dm_demo_cdata red_square = { .colour = "red", .sides = 4. }; static const struct driver_info info[] = { { .name = "demo_shape_drv", .platdata = &red_square, }, }; demo1 = driver_bind(root, &info[0]); Device Tree ----------- While platdata is useful, a more flexible way of providing device data is by using device tree. With device tree we replace the above code with the following device tree fragment: red-square { compatible = "demo-shape"; colour = "red"; sides = <4>; }; This means that instead of having lots of U_BOOT_DEVICE() declarations in the board file, we put these in the device tree. This approach allows a lot more generality, since the same board file can support many types of boards (e,g. with the same SoC) just by using different device trees. An added benefit is that the Linux device tree can be used, thus further simplifying the task of board-bring up either for U-Boot or Linux devs (whoever gets to the board first!). The easiest way to make this work it to add a few members to the driver: .platdata_auto_alloc_size = sizeof(struct dm_test_pdata), .ofdata_to_platdata = testfdt_ofdata_to_platdata, The 'auto_alloc' feature allowed space for the platdata to be allocated and zeroed before the driver's ofdata_to_platdata() method is called. The ofdata_to_platdata() method, which the driver write supplies, should parse the device tree node for this device and place it in dev->platdata. Thus when the probe method is called later (to set up the device ready for use) the platform data will be present. Note that both methods are optional. If you provide an ofdata_to_platdata method then it will be called first (during activation). If you provide a probe method it will be called next. See Driver Lifecycle below for more details. If you don't want to have the platdata automatically allocated then you can leave out platdata_auto_alloc_size. In this case you can use malloc in your ofdata_to_platdata (or probe) method to allocate the required memory, and you should free it in the remove method. Declaring Uclasses ------------------ The demo uclass is declared like this: U_BOOT_CLASS(demo) = { .id = UCLASS_DEMO, }; It is also possible to specify special methods for probe, etc. The uclass numbering comes from include/dm/uclass.h. To add a new uclass, add to the end of the enum there, then declare your uclass as above. Driver Lifecycle ---------------- Here are the stages that a device goes through in driver model. Note that all methods mentioned here are optional - e.g. if there is no probe() method for a device then it will not be called. A simple device may have very few methods actually defined. 1. Bind stage A device and its driver are bound using one of these two methods: - Scan the U_BOOT_DEVICE() definitions. U-Boot It looks up the name specified by each, to find the appropriate driver. It then calls device_bind() to create a new device and bind' it to its driver. This will call the device's bind() method. - Scan through the device tree definitions. U-Boot looks at top-level nodes in the the device tree. It looks at the compatible string in each node and uses the of_match part of the U_BOOT_DRIVER() structure to find the right driver for each node. It then calls device_bind() to bind the newly-created device to its driver (thereby creating a device structure). This will also call the device's bind() method. At this point all the devices are known, and bound to their drivers. There is a 'struct udevice' allocated for all devices. However, nothing has been activated (except for the root device). Each bound device that was created from a U_BOOT_DEVICE() declaration will hold the platdata pointer specified in that declaration. For a bound device created from the device tree, platdata will be NULL, but of_offset will be the offset of the device tree node that caused the device to be created. The uclass is set correctly for the device. The device's bind() method is permitted to perform simple actions, but should not scan the device tree node, not initialise hardware, nor set up structures or allocate memory. All of these tasks should be left for the probe() method. Note that compared to Linux, U-Boot's driver model has a separate step of probe/remove which is independent of bind/unbind. This is partly because in U-Boot it may be expensive to probe devices and we don't want to do it until they are needed, or perhaps until after relocation. 2. Activation/probe When a device needs to be used, U-Boot activates it, by following these steps (see device_probe()): a. If priv_auto_alloc_size is non-zero, then the device-private space is allocated for the device and zeroed. It will be accessible as dev->priv. The driver can put anything it likes in there, but should use it for run-time information, not platform data (which should be static and known before the device is probed). b. If platdata_auto_alloc_size is non-zero, then the platform data space is allocated. This is only useful for device tree operation, since otherwise you would have to specific the platform data in the U_BOOT_DEVICE() declaration. The space is allocated for the device and zeroed. It will be accessible as dev->platdata. c. If the device's uclass specifies a non-zero per_device_auto_alloc_size, then this space is allocated and zeroed also. It is allocated for and stored in the device, but it is uclass data. owned by the uclass driver. It is possible for the device to access it. d. All parent devices are probed. It is not possible to activate a device unless its predecessors (all the way up to the root device) are activated. This means (for example) that an I2C driver will require that its bus be activated. e. If the driver provides an ofdata_to_platdata() method, then this is called to convert the device tree data into platform data. This should do various calls like fdtdec_get_int(gd->fdt_blob, dev->of_offset, ...) to access the node and store the resulting information into dev->platdata. After this point, the device works the same way whether it was bound using a device tree node or U_BOOT_DEVICE() structure. In either case, the platform data is now stored in the platdata structure. Typically you will use the platdata_auto_alloc_size feature to specify the size of the platform data structure, and U-Boot will automatically allocate and zero it for you before entry to ofdata_to_platdata(). But if not, you can allocate it yourself in ofdata_to_platdata(). Note that it is preferable to do all the device tree decoding in ofdata_to_platdata() rather than in probe(). (Apart from the ugliness of mixing configuration and run-time data, one day it is possible that U-Boot will cache platformat data for devices which are regularly de/activated). f. The device's probe() method is called. This should do anything that is required by the device to get it going. This could include checking that the hardware is actually present, setting up clocks for the hardware and setting up hardware registers to initial values. The code in probe() can access: - platform data in dev->platdata (for configuration) - private data in dev->priv (for run-time state) - uclass data in dev->uclass_priv (for things the uclass stores about this device) Note: If you don't use priv_auto_alloc_size then you will need to allocate the priv space here yourself. The same applies also to platdata_auto_alloc_size. Remember to free them in the remove() method. g. The device is marked 'activated' h. The uclass's post_probe() method is called, if one exists. This may cause the uclass to do some housekeeping to record the device as activated and 'known' by the uclass. 3. Running stage The device is now activated and can be used. From now until it is removed all of the above structures are accessible. The device appears in the uclass's list of devices (so if the device is in UCLASS_GPIO it will appear as a device in the GPIO uclass). This is the 'running' state of the device. 4. Removal stage When the device is no-longer required, you can call device_remove() to remove it. This performs the probe steps in reverse: a. The uclass's pre_remove() method is called, if one exists. This may cause the uclass to do some housekeeping to record the device as deactivated and no-longer 'known' by the uclass. b. All the device's children are removed. It is not permitted to have an active child device with a non-active parent. This means that device_remove() is called for all the children recursively at this point. c. The device's remove() method is called. At this stage nothing has been deallocated so platform data, private data and the uclass data will all still be present. This is where the hardware can be shut down. It is intended that the device be completely inactive at this point, For U-Boot to be sure that no hardware is running, it should be enough to remove all devices. d. The device memory is freed (platform data, private data, uclass data). Note: Because the platform data for a U_BOOT_DEVICE() is defined with a static pointer, it is not de-allocated during the remove() method. For a device instantiated using the device tree data, the platform data will be dynamically allocated, and thus needs to be deallocated during the remove() method, either: 1. if the platdata_auto_alloc_size is non-zero, the deallocation happens automatically within the driver model core; or 2. when platdata_auto_alloc_size is 0, both the allocation (in probe() or preferably ofdata_to_platdata()) and the deallocation in remove() are the responsibility of the driver author. e. The device is marked inactive. Note that it is still bound, so the device structure itself is not freed at this point. Should the device be activated again, then the cycle starts again at step 2 above. 5. Unbind stage The device is unbound. This is the step that actually destroys the device. If a parent has children these will be destroyed first. After this point the device does not exist and its memory has be deallocated. Data Structures --------------- Driver model uses a doubly-linked list as the basic data structure. Some nodes have several lists running through them. Creating a more efficient data structure might be worthwhile in some rare cases, once we understand what the bottlenecks are. Changes since v1 ---------------- For the record, this implementation uses a very similar approach to the original patches, but makes at least the following changes: - Tried to aggressively remove boilerplate, so that for most drivers there is little or no 'driver model' code to write. - Moved some data from code into data structure - e.g. store a pointer to the driver operations structure in the driver, rather than passing it to the driver bind function. - Rename some structures to make them more similar to Linux (struct udevice instead of struct instance, struct platdata, etc.) - Change the name 'core' to 'uclass', meaning U-Boot class. It seems that this concept relates to a class of drivers (or a subsystem). We shouldn't use 'class' since it is a C++ reserved word, so U-Boot class (uclass) seems better than 'core'. - Remove 'struct driver_instance' and just use a single 'struct udevice'. This removes a level of indirection that doesn't seem necessary. - Built in device tree support, to avoid the need for platdata - Removed the concept of driver relocation, and just make it possible for the new driver (created after relocation) to access the old driver data. I feel that relocation is a very special case and will only apply to a few drivers, many of which can/will just re-init anyway. So the overhead of dealing with this might not be worth it. - Implemented a GPIO system, trying to keep it simple Things to punt for later ------------------------ - SPL support - this will have to be present before many drivers can be converted, but it seems like we can add it once we are happy with the core implementation. - Pre-relocation support - similar story That is not to say that no thinking has gone into these - in fact there is quite a lot there. However, getting these right is non-trivial and there is a high cost associated with going down the wrong path. For SPL, it may be possible to fit in a simplified driver model with only bind and probe methods, to reduce size. For pre-relocation we can simply call the driver model init function. Then post relocation we throw that away and re-init driver model again. For drivers which require some sort of continuity between pre- and post-relocation devices, we can provide access to the pre-relocation device pointers. Uclasses are statically numbered at compile time. It would be possible to change this to dynamic numbering, but then we would require some sort of lookup service, perhaps searching by name. This is slightly less efficient so has been left out for now. One small advantage of dynamic numbering might be fewer merge conflicts in uclass-id.h. Simon Glass sjg@chromium.org April 2013 Updated 7-May-13 Updated 14-Jun-13 Updated 18-Oct-13 Updated 5-Nov-13