{{{#!comment Credit for textual content goes to Petr Mánek. }}} == Basics This section gives details on the basic structure of USB device drivers, their role in the system and components they usually interact with. === Framework USB device drivers use the generic //Device Driver Framework// available in HelenOS. Because all USB drivers have similar initialization routines, a thin layer – specific to USB devices – was added above the generic one. This layer mainly serves as a middleware for easy communication with USB host controller drivers, performing USB specific resource management and enabling device drivers to initialize endpoint pipes and interact with USB devices. For those reasons, USB device drivers are recommended to be linked with libdrv and libusbdev, which contain both aforementioned layers respectively. It is expected that USB device drivers specify in advance not only their relevant match identifiers, which are used by the Device Manager to pair new devices with available drivers, but also all endpoints, which shall be present on the device through a USB driver structure. Later when a new device is found, a specialized device structure is prepared and pipe abstractions are initialized. Device drivers live the same life cycle as any other drivers controlled by the Device Manager. A quick summary follows: 1. The driver is started at the convenience of the Device Manager if and when a compatible device is found. At startup, the driver registers with the USB framework, which in turn registers also with the Device Driver Framework. 2. During its lifetime, the driver receives callbacks from the USB Framework, informing it about relevant device events. On the basis of these events, the driver then communicates with the device and exposes various interfaces to other tasks in the system. At this point, the controlled device usually becomes visible and useful to the user. 3. When there is no more need for the driver to run (i.e. no devices to control), the Device Manager may terminate the driver to save resources. In the Device Driver Framework, drivers are the //consumers// of devices and providers of //functions//. This paradigm allows them to expose an unlimited number of nodes (representing logical or physical units) for every device they control. The same basic principle holds for USB drivers as well. === Device Callbacks As explained in the previous section, USB device drivers are informed about relevant device events by asynchronous callbacks from the USB framework. To simplify usage, these callbacks are identical to those of the Device Driver Framework: Add Device:: This event notifies the driver that a new device has been discovered and matched to it. From this point on, the driver is allowed to communicate with the device in order to configure it and expose its functions to the rest of the system. Further communication with the device will likely depend on remote calls originating from other system tasks utilizing the exposed interface. Remove Device:: This event instructs the driver to immediately disallow new user operations on a device, terminate all currently running operations in a timely manner, and hand device control back to the system, as the device will likely be physically removed from the bus in the foreseeable future. Device Gone:: This event informs the driver that a device has been physically disconnected from the system without a prior Remove Device event. Since the device is no longer reachable, the driver is to force interrupt all user operations, which were running at the time of receiving the event and report failure to the callers. Offline Function:: By receiving this event, the driver is asked by the system to explicitly transition a specific function exposed by one of its controlled devices into the Offline state. The meaning of such transition might depend on the interpretation of the function. For more information, see the [wiki:DeviceDrivers the device drivers guide]. Online Function:: By receiving this event, the driver is asked by the system to explicitly transition a specific function exposed by one of its controlled devices into the Online state. Again, the meaning of such transition might depend on the interpretation of the function. For more information, see the [wiki:DeviceDrivers the device drivers guide]. The listed events are delivered to device drivers through function calls to event handlers specified in the main USB driver operations structure (see the following minimal example). The order of event delivery models the physical lifecycle of the device within the system. For instance, it is not possible for the //Add Device// event to occur more than once for the same device in a row. {{{#!c static int device_add(usb_device_t *dev) { usb_log_info("Device '%s' added.", usb_device_get_name(dev)); return EOK; } static int device_remove(usb_device_t *dev) { usb_log_info("Device '%s' removed.", usb_device_get_name(dev)); return EOK; } static int device_gone(usb_device_t *dev) { usb_log_info("Device '%s' gone.", usb_device_get_name(dev)); return EOK; } static const usb_driver_ops_t driver_ops = { .device_add = device_add, .device_remove = device_remove, .device_gone = device_gone, }; }}} {{{#!comment TODO: Figure For reference, the exact device states and transitions between them (corresponding to the events with respective names) are shown in Figure A.1. }}} Furthermore, since there are no guarantees on the synchronization between calls, the driver is explicitly responsible for synchronizing accesses to its private data structures as well as device communication. == Device Communication In USB, a pipe is an abstraction primitive for a communication channel between a device and a host computer. For the sake of simplicity, the provided framework relies on a mechanism based on the same abstraction in order to facilitate communication between device drivers and devices. == Endpoint Description In order to use pipes, device drivers must first define at least one //endpoint description//. The purpose of such definition is to specify all device endpoints, which will be used for communication by the device driver throughout its lifecycle. For that reason, all descriptions ought to be specified in advance and referenced in the main device driver structure. An example of possible endpoint description definition follows: {{{#!c static const usb_endpoint_description_t bulk_in_ep = { .transfer_type = USB_TRANSFER_BULK, .direction = USB_DIRECTION_IN, .interface_class = USB_CLASS_MASS_STORAGE, .interface_subclass = USB_MASSSTOR_SUBCLASS_SCSI, .interface_protocol = USB_MASSSTOR_PROTOCOL_BBB, .flags = 0 }; static const usb_endpoint_description_t bulk_out_ep = { .transfer_type = USB_TRANSFER_BULK, .direction = USB_DIRECTION_OUT, .interface_class = USB_CLASS_MASS_STORAGE, .interface_subclass = USB_MASSSTOR_SUBCLASS_SCSI, .interface_protocol = USB_MASSSTOR_PROTOCOL_BBB, .flags = 0 }; static const usb_endpoint_description_t *endpoints[] = { &bulk_in_ep, &bulk_out_ep, NULL }; }}} Endpoint description contains information which can be matched to USB endpoint descriptors in very much the same way as match identifiers are used to pair devices with device drivers in HelenOS. Following this scheme, the presented USB framework does most of the heavy lifting for device drivers. When a new device is added, prior to delivery of the //Add Device// event, all endpoint descriptions provided by the driver are matched to device endpoints, resulting in two possible outcomes for every description: 1. A device endpoint matching the provided description is found and //an endpoint mapping// is created. 2. No device endpoint matches the provided description, hence no mapping is created. Device drivers can later query the result of the matching, and if successful, retrieve the created endpoint mapping containing a fully initialized pipe instance. An example of this is shown in the following snippet: {{{#!c static int device_add(usb_device_t *dev) { /* Find mapping for the endpoint description. */ usb_endpoint_mapping_t *mapping = usb_device_get_mapped_ep_desc(dev, &bulk_out_ep); /* Determine if the mapping was successful. */ if (!mapping || !mapping->present) { usb_log_error("Endpoint mapping failed!"); return ENOENT; } usb_pipe_t *pipe = &mapping->pipe; /* Now we can write to `pipe`. */ return write_data(pipe); } }}} === Pipe I/O Provided that an endpoint description has been defined, successfully matched and an endpoint mapping has been retrieved along with a pipe instance, a device driver can schedule data transfers to a device. Pipes offer a very simple synchronous command interface similar to that of an open file descriptor, namely `usb_pipe_read` and `usb_pipe_write` for reading from IN endpoints and writing to OUT endpoints respectively. While semantics of these functions depends on endpoint types (consult USB specification for details), their basic usage is as follows: 1. Device driver allocates a buffer of appropriate size (and fills it with data if writing). 2. Either the `usb_pipe_read` or `usb_pipe_write` function is called, scheduling transfers to the device. The function receives a pipe pointer, pointer to the user-allocated buffer and data size. 3. The calling fibril is blocked until transfers succeed or fail (see error code). If reading, the number of read bytes is returned as well. 4. Device driver recycles or disposes of the allocated buffer. Note that while the write function blocks the calling fibril until the entire buffer is transferred to the device, the read function might return successfully with a lower number of bytes read than the actual data size requested, and may thus have to be called multiple times. In addition, due to the multiplatform nature of HelenOS, it is not advisable to assume endianity of the host system. Instead, the `uint{16|32}_host2usb` and `uint{16|32}_usb2host` macros serve to convert host system endianity to and from the transfer endianity defined in the USB Protocol Specification at driver's convenience. A complete example of pipe write is shown in the following code: {{{#!c static int write_data(usb_pipe_t *pipe) { int rc = EOK; /* Allocate memory. */ static const size_t size = 64; char *buffer = (char *) malloc(size); if (!buffer) { rc = ENOMEM; goto err; } /* Fill `buffer` with arithmetic sequence of bytes. */ for (int i = 0; i < size; ++i) buffer[i] = (char) i; /* Write data. */ rc = usb_pipe_write(pipe, buffer, size); if (rc != EOK) { usb_log_error("Write failed with error: %s", str_error(rc)); goto err_buffer; } /* Clean up. */ err_buffer: free(buffer); err: return rc; } }}} === Automated Polling In USB devices which often interact with a human user, it is important to wait for user input by spinning (or polling) on some //Interrupt IN// endpoint. In the abstraction used, this translates to calling `usb_pipe_read` in a perpetual loop and responding to incoming data or errors. Since this behavior is common to a significant class of drivers, a reusable unified implementation is provided by the USB framework. The implementation is represented by the `usb_polling_t` structure and its associated functions. Similarly to device event handlers, device drivers are expected to configure this structure with polling handler functions. Later when polling should be initiated, a call to the `usb_polling_start` function will spawn a separate fibril, spinning on the pipe and calling handlers when data arrive or `usb_pipe_read` fails with error. When polling is to be ceased, a call to the `usb_polling_join` function will spontaneously wake up the polling fibril and block the caller fibril until its termination. If polling is started, a call to this function must //always// occur no later than in the //Remove Device// or //Device Gone// event handler in order to prevent the fibril from outlasting the device lifecycle within driver private data structures.. Furthermore, note that once any device driver starts polling on a pipe, it transfers the pipe ownership to the polling fibril, relinquishing all control over it. That prohibits subsequent calls to `usb_pipe_read` or other relevant functions. It is also not feasible to poll only for a limited time period, then join and reuse the pipe for arbitrary data reads. This is due to the fact that once `usb_polling_join` is called, the pipe is left closed for all communication, destroying the underlying endpoint mapping. An example usage of the pipe polling interface follows: {{{#!c static usb_polling_t polling; static uint8_t buffer[13]; static bool callback(usb_device_t *dev, uint8_t *buffer, size_t size, void *arg) { printf("Have data!/n"); // Return true if we wish to continue polling. return true; } static void demo() { // Initialize. usb_polling_init(&polling); // Configure. polling.device = /* some usb_device_t here */; polling.ep_mapping = /* some interrupt(in) endpoint of the device */; polling.buffer = buffer; polling.request_size = sizeof(buffer); polling.on_data = callback; // Start polling. usb_polling_start(&polling); // Sleep synchronously for a while. async_usleep(10000); // End polling and clean up. usb_polling_join(&polling); usb_polling_fini(&polling); } }}} == Miscellaneous This section features general suggestions and recommendations for USB device driver implementation. === Device-specific Data Structures Device drivers often need to maintain stateful and descriptive information related to their controlled devices. A common practice is to store this kind of information in device-specific data structures. As their name suggests, such structures usually correspond to a single controlled device, and can thus closely follow its lifecycle, being allocated and configured when the device is added and being freed upon its removal. The presented USB framework offers device drivers a straightforward mechanism to associate any of their data structures with USB devices by specifying custom //device data/ in the `usb_device_t` structure, which is present in all device-related events. Device data is allocated at most once for every device and can have any driver-specified, albeit constant size. When the device is later removed from the driver, device data is automatically freed if present. Allocation of device data is performed by a call to the `usb_device_data_alloc` function. This function has the same return value and semantics as the `calloc` function. In addition, the pointer returned by this function is stored within the `usb_device_t` structure and can be retrieved from that point onward until the device is removed by a call to the `usb_device_data_get` function. Example usage of this mechanism can be seen here: {{{#!c typedef struct my_data { /* Any device-specific data here. */ usb_polling_t polling; } my_data_t; static int device_add(usb_device_t *dev) { /* Allocate device data. */ my_data_t *data = (my_data_t *) usb_device_data_alloc(dev, sizeof(my_data_t)); if (!data) { return ENOMEM; } usb_polling_init(&data->polling); /* ... configure polling ... */ usb_polling_start(&data->polling); return EOK; } static int device_remove(usb_device_t *dev) { /* Retrieve device data. */ my_data_t *data = (my_data_t *) usb_device_data_get(dev); /* Stop polling. */ usb_polling_join(&data->polling); usb_polling_fini(&data->polling); /* After returning, device data is freed automatically. */ return EOK; } }}} In this case, the `my_data_t` device-specific data structure is used to keep track of automated pipe polling. Note that some error conditions have been intentionally ignored for the sake of brevity. === Logging Another useful feature of the USB device driver framework is a set of logging macros. These macros copy the scheme common to the HelenOS logging framework, offering various log levels a format string syntax for a variable number of arguments consistent with the \fnc{printf} function syntax. With decreasing severity, the presented logging macros are named as follows: * `usb_log_fatal` for fatal errors, * `usb_log_error` for recoverable errors, * `usb_log_warning` for warnings, * `usb_log_info` for informational messages (produces a log message of the `Note` level), * `usb_log_debug` for debugging messages, * `usb_log_debug2` for verbose debugging messages. Prior to printing the first log message, the `log_init` function must be called. During runtime, the default HelenOS log level can be adjusted by a call to the `logctl_set_log_level` function. Keep in mind that in the current implementation of HelenOS, every call to a logging macro results in an IPC call. For that reason, it is not advisable to perform logging in performance-sensitive parts of the code, even if the log level is sufficiently high. For more information about logging in HelenOS, consult the HelenOS Documentation. === Exposing an Interface In order for controlled hardware to become usable to system tasks, HelenOS device drivers can expose IPC interfaces to abstract hardware-specific features and provide a set of well-defined logical operations for device interaction. In HelenOS IPC, one of the preferred ways to achieve this is by interacting with //services// -- system tasks, which run as daemons, and whose purpose is to connect user tasks with device drivers. Since services usually keep track of devices of the same category, it is necessary to first identify the appropriate service for a driver and consult its documentation. While the specifics of exposing IPC interfaces may vary for different device categories, the general scheme remains the same: 1. For a DDF function, fill a specific structure with configuration and call handlers Then register them with the respective system service. 2. Respond to service-specific calls in compliance with service specification. Handling of these calls will likely result in communication with the device. 3. When the device is removed, cease handling service calls and unregister the device from the service. === Descriptor Parser While they are sometimes stored in a serialized form, USB descriptors can be viewed as a tree. For instance, configuration descriptor is followed by interface and endpoint descriptors, which are not directly accessible. To aid with problems such as this, a simple descriptor parser is provided by the presented framework. The parser has been designed to be as generic as possible and its interface might thus be somewhat terse. It operates on a byte array and offers functions for finding the first nested descriptor and its next sibling (i.e. descriptor on the same depth of nesting). The parser expects that the input is an array of bytes where the descriptors are sorted in //prefix tree traversal// order. Next, it expects that each descriptor has its length stored in the first byte and the descriptor type in the second byte. That corresponds to standard descriptor layout. The parser determines the nesting from a list of parent-child pairs that is given to it during parser initialization. This list is terminated with a pair where both parent and child are set to -1. The parser uses the `usb_dp_parser_t` structure, which contains array with possible descriptor nesting (`usb_dp_descriptor_nesting_t`). The data for the parser are stored in the `usb_dp_parser_data_t`, the arg field is intended for custom data. For processing the actual data, two functions are available. The `usb_dp_get_nested_descriptor` function takes a parser, parser data and a pointer to parent as parameters. For retrieving the sibling descriptor (i.e. descriptor at the same depth) one can use `usb_dp_get_sibling_descriptor`. Whereas this function takes the same arguments, it also requires two extra arguments — pointer to the parent descriptor (i.e. parent of both nested ones) and a pointer to the preceding descriptor. Such pointer must always point to the first byte of the descriptor (i.e. to the length of the descriptor, not to its type). Additionally, there exists a simple iterator over the descriptors. The function `usb_dp_walk_simple` takes a callback as a parameter, which is executed for each found descriptor and receives thee current depth (starting with 0) and a pointer to current descriptor.